fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 struct btrfs_dio_data {
  70         u64 outstanding_extents;
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74 };
  75
  76 static const struct inode_operations btrfs_dir_inode_operations;
  77 static const struct inode_operations btrfs_symlink_inode_operations;
  78 static const struct inode_operations btrfs_dir_ro_inode_operations;
  79 static const struct inode_operations btrfs_special_inode_operations;
  80 static const struct inode_operations btrfs_file_inode_operations;
  81 static const struct address_space_operations btrfs_aops;
  82 static const struct address_space_operations btrfs_symlink_aops;
  83 static const struct file_operations btrfs_dir_file_operations;
  84 static const struct extent_io_ops btrfs_extent_io_ops;
  85
  86 static struct kmem_cache *btrfs_inode_cachep;
  87 struct kmem_cache *btrfs_trans_handle_cachep;
  88 struct kmem_cache *btrfs_transaction_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, int *page_started,
 109                                    unsigned long *nr_written, int unlock);
 110 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 111                                            u64 len, u64 orig_start,
 112                                            u64 block_start, u64 block_len,
 113                                            u64 orig_block_len, u64 ram_bytes,
 114                                            int type);
 115
 116 static int btrfs_dirty_inode(struct inode *inode);
 117
 118 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 119 void btrfs_test_inode_set_ops(struct inode *inode)
 120 {
 121         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 122 }
 123 #endif
 124
 125 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 126                                      struct inode *inode,  struct inode *dir,
 127                                      const struct qstr *qstr)
 128 {
 129         int err;
 130
 131         err = btrfs_init_acl(trans, inode, dir);
 132         if (!err)
 133                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 134         return err;
 135 }
 136
 137 /*
 138  * this does all the hard work for inserting an inline extent into
 139  * the btree.  The caller should have done a btrfs_drop_extents so that
 140  * no overlapping inline items exist in the btree
 141  */
 142 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 143                                 struct btrfs_path *path, int extent_inserted,
 144                                 struct btrfs_root *root, struct inode *inode,
 145                                 u64 start, size_t size, size_t compressed_size,
 146                                 int compress_type,
 147                                 struct page **compressed_pages)
 148 {
 149         struct extent_buffer *leaf;
 150         struct page *page = NULL;
 151         char *kaddr;
 152         unsigned long ptr;
 153         struct btrfs_file_extent_item *ei;
 154         int err = 0;
 155         int ret;
 156         size_t cur_size = size;
 157         unsigned long offset;
 158
 159         if (compressed_size && compressed_pages)
 160                 cur_size = compressed_size;
 161
 162         inode_add_bytes(inode, size);
 163
 164         if (!extent_inserted) {
 165                 struct btrfs_key key;
 166                 size_t datasize;
 167
 168                 key.objectid = btrfs_ino(inode);
 169                 key.offset = start;
 170                 key.type = BTRFS_EXTENT_DATA_KEY;
 171
 172                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 173                 path->leave_spinning = 1;
 174                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                               datasize);
 176                 if (ret) {
 177                         err = ret;
 178                         goto fail;
 179                 }
 180         }
 181         leaf = path->nodes[0];
 182         ei = btrfs_item_ptr(leaf, path->slots[0],
 183                             struct btrfs_file_extent_item);
 184         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 185         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 186         btrfs_set_file_extent_encryption(leaf, ei, 0);
 187         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 188         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 189         ptr = btrfs_file_extent_inline_start(ei);
 190
 191         if (compress_type != BTRFS_COMPRESS_NONE) {
 192                 struct page *cpage;
 193                 int i = 0;
 194                 while (compressed_size > 0) {
 195                         cpage = compressed_pages[i];
 196                         cur_size = min_t(unsigned long, compressed_size,
 197                                        PAGE_SIZE);
 198
 199                         kaddr = kmap_atomic(cpage);
 200                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 201                         kunmap_atomic(kaddr);
 202
 203                         i++;
 204                         ptr += cur_size;
 205                         compressed_size -= cur_size;
 206                 }
 207                 btrfs_set_file_extent_compression(leaf, ei,
 208                                                   compress_type);
 209         } else {
 210                 page = find_get_page(inode->i_mapping,
 211                                      start >> PAGE_SHIFT);
 212                 btrfs_set_file_extent_compression(leaf, ei, 0);
 213                 kaddr = kmap_atomic(page);
 214                 offset = start & (PAGE_SIZE - 1);
 215                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 216                 kunmap_atomic(kaddr);
 217                 put_page(page);
 218         }
 219         btrfs_mark_buffer_dirty(leaf);
 220         btrfs_release_path(path);
 221
 222         /*
 223          * we're an inline extent, so nobody can
 224          * extend the file past i_size without locking
 225          * a page we already have locked.
 226          *
 227          * We must do any isize and inode updates
 228          * before we unlock the pages.  Otherwise we
 229          * could end up racing with unlink.
 230          */
 231         BTRFS_I(inode)->disk_i_size = inode->i_size;
 232         ret = btrfs_update_inode(trans, root, inode);
 233
 234         return ret;
 235 fail:
 236         return err;
 237 }
 238
 239
 240 /*
 241  * conditionally insert an inline extent into the file.  This
 242  * does the checks required to make sure the data is small enough
 243  * to fit as an inline extent.
 244  */
 245 static noinline int cow_file_range_inline(struct btrfs_root *root,
 246                                           struct inode *inode, u64 start,
 247                                           u64 end, size_t compressed_size,
 248                                           int compress_type,
 249                                           struct page **compressed_pages)
 250 {
 251         struct btrfs_trans_handle *trans;
 252         u64 isize = i_size_read(inode);
 253         u64 actual_end = min(end + 1, isize);
 254         u64 inline_len = actual_end - start;
 255         u64 aligned_end = ALIGN(end, root->sectorsize);
 256         u64 data_len = inline_len;
 257         int ret;
 258         struct btrfs_path *path;
 259         int extent_inserted = 0;
 260         u32 extent_item_size;
 261
 262         if (compressed_size)
 263                 data_len = compressed_size;
 264
 265         if (start > 0 ||
 266             actual_end > root->sectorsize ||
 267             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 268             (!compressed_size &&
 269             (actual_end & (root->sectorsize - 1)) == 0) ||
 270             end + 1 < isize ||
 271             data_len > root->fs_info->max_inline) {
 272                 return 1;
 273         }
 274
 275         path = btrfs_alloc_path();
 276         if (!path)
 277                 return -ENOMEM;
 278
 279         trans = btrfs_join_transaction(root);
 280         if (IS_ERR(trans)) {
 281                 btrfs_free_path(path);
 282                 return PTR_ERR(trans);
 283         }
 284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 285
 286         if (compressed_size && compressed_pages)
 287                 extent_item_size = btrfs_file_extent_calc_inline_size(
 288                    compressed_size);
 289         else
 290                 extent_item_size = btrfs_file_extent_calc_inline_size(
 291                     inline_len);
 292
 293         ret = __btrfs_drop_extents(trans, root, inode, path,
 294                                    start, aligned_end, NULL,
 295                                    1, 1, extent_item_size, &extent_inserted);
 296         if (ret) {
 297                 btrfs_abort_transaction(trans, root, ret);
 298                 goto out;
 299         }
 300
 301         if (isize > actual_end)
 302                 inline_len = min_t(u64, isize, actual_end);
 303         ret = insert_inline_extent(trans, path, extent_inserted,
 304                                    root, inode, start,
 305                                    inline_len, compressed_size,
 306                                    compress_type, compressed_pages);
 307         if (ret && ret != -ENOSPC) {
 308                 btrfs_abort_transaction(trans, root, ret);
 309                 goto out;
 310         } else if (ret == -ENOSPC) {
 311                 ret = 1;
 312                 goto out;
 313         }
 314
 315         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 316         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 317         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 318 out:
 319         /*
 320          * Don't forget to free the reserved space, as for inlined extent
 321          * it won't count as data extent, free them directly here.
 322          * And at reserve time, it's always aligned to page size, so
 323          * just free one page here.
 324          */
 325         btrfs_qgroup_free_data(inode, 0, PAGE_SIZE);
 326         btrfs_free_path(path);
 327         btrfs_end_transaction(trans, root);
 328         return ret;
 329 }
 330
 331 struct async_extent {
 332         u64 start;
 333         u64 ram_size;
 334         u64 compressed_size;
 335         struct page **pages;
 336         unsigned long nr_pages;
 337         int compress_type;
 338         struct list_head list;
 339 };
 340
 341 struct async_cow {
 342         struct inode *inode;
 343         struct btrfs_root *root;
 344         struct page *locked_page;
 345         u64 start;
 346         u64 end;
 347         struct list_head extents;
 348         struct btrfs_work work;
 349 };
 350
 351 static noinline int add_async_extent(struct async_cow *cow,
 352                                      u64 start, u64 ram_size,
 353                                      u64 compressed_size,
 354                                      struct page **pages,
 355                                      unsigned long nr_pages,
 356                                      int compress_type)
 357 {
 358         struct async_extent *async_extent;
 359
 360         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 361         BUG_ON(!async_extent); /* -ENOMEM */
 362         async_extent->start = start;
 363         async_extent->ram_size = ram_size;
 364         async_extent->compressed_size = compressed_size;
 365         async_extent->pages = pages;
 366         async_extent->nr_pages = nr_pages;
 367         async_extent->compress_type = compress_type;
 368         list_add_tail(&async_extent->list, &cow->extents);
 369         return 0;
 370 }
 371
 372 static inline int inode_need_compress(struct inode *inode)
 373 {
 374         struct btrfs_root *root = BTRFS_I(inode)->root;
 375
 376         /* force compress */
 377         if (btrfs_test_opt(root, FORCE_COMPRESS))
 378                 return 1;
 379         /* bad compression ratios */
 380         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 381                 return 0;
 382         if (btrfs_test_opt(root, COMPRESS) ||
 383             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 384             BTRFS_I(inode)->force_compress)
 385                 return 1;
 386         return 0;
 387 }
 388
 389 /*
 390  * we create compressed extents in two phases.  The first
 391  * phase compresses a range of pages that have already been
 392  * locked (both pages and state bits are locked).
 393  *
 394  * This is done inside an ordered work queue, and the compression
 395  * is spread across many cpus.  The actual IO submission is step
 396  * two, and the ordered work queue takes care of making sure that
 397  * happens in the same order things were put onto the queue by
 398  * writepages and friends.
 399  *
 400  * If this code finds it can't get good compression, it puts an
 401  * entry onto the work queue to write the uncompressed bytes.  This
 402  * makes sure that both compressed inodes and uncompressed inodes
 403  * are written in the same order that the flusher thread sent them
 404  * down.
 405  */
 406 static noinline void compress_file_range(struct inode *inode,
 407                                         struct page *locked_page,
 408                                         u64 start, u64 end,
 409                                         struct async_cow *async_cow,
 410                                         int *num_added)
 411 {
 412         struct btrfs_root *root = BTRFS_I(inode)->root;
 413         u64 num_bytes;
 414         u64 blocksize = root->sectorsize;
 415         u64 actual_end;
 416         u64 isize = i_size_read(inode);
 417         int ret = 0;
 418         struct page **pages = NULL;
 419         unsigned long nr_pages;
 420         unsigned long nr_pages_ret = 0;
 421         unsigned long total_compressed = 0;
 422         unsigned long total_in = 0;
 423         unsigned long max_compressed = SZ_128K;
 424         unsigned long max_uncompressed = SZ_128K;
 425         int i;
 426         int will_compress;
 427         int compress_type = root->fs_info->compress_type;
 428         int redirty = 0;
 429
 430         /* if this is a small write inside eof, kick off a defrag */
 431         if ((end - start + 1) < SZ_16K &&
 432             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434
 435         actual_end = min_t(u64, isize, end + 1);
 436 again:
 437         will_compress = 0;
 438         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 439         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_SIZE);
 440
 441         /*
 442          * we don't want to send crud past the end of i_size through
 443          * compression, that's just a waste of CPU time.  So, if the
 444          * end of the file is before the start of our current
 445          * requested range of bytes, we bail out to the uncompressed
 446          * cleanup code that can deal with all of this.
 447          *
 448          * It isn't really the fastest way to fix things, but this is a
 449          * very uncommon corner.
 450          */
 451         if (actual_end <= start)
 452                 goto cleanup_and_bail_uncompressed;
 453
 454         total_compressed = actual_end - start;
 455
 456         /*
 457          * skip compression for a small file range(<=blocksize) that
 458          * isn't an inline extent, since it doesn't save disk space at all.
 459          */
 460         if (total_compressed <= blocksize &&
 461            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 462                 goto cleanup_and_bail_uncompressed;
 463
 464         /* we want to make sure that amount of ram required to uncompress
 465          * an extent is reasonable, so we limit the total size in ram
 466          * of a compressed extent to 128k.  This is a crucial number
 467          * because it also controls how easily we can spread reads across
 468          * cpus for decompression.
 469          *
 470          * We also want to make sure the amount of IO required to do
 471          * a random read is reasonably small, so we limit the size of
 472          * a compressed extent to 128k.
 473          */
 474         total_compressed = min(total_compressed, max_uncompressed);
 475         num_bytes = ALIGN(end - start + 1, blocksize);
 476         num_bytes = max(blocksize,  num_bytes);
 477         total_in = 0;
 478         ret = 0;
 479
 480         /*
 481          * we do compression for mount -o compress and when the
 482          * inode has not been flagged as nocompress.  This flag can
 483          * change at any time if we discover bad compression ratios.
 484          */
 485         if (inode_need_compress(inode)) {
 486                 WARN_ON(pages);
 487                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 488                 if (!pages) {
 489                         /* just bail out to the uncompressed code */
 490                         goto cont;
 491                 }
 492
 493                 if (BTRFS_I(inode)->force_compress)
 494                         compress_type = BTRFS_I(inode)->force_compress;
 495
 496                 /*
 497                  * we need to call clear_page_dirty_for_io on each
 498                  * page in the range.  Otherwise applications with the file
 499                  * mmap'd can wander in and change the page contents while
 500                  * we are compressing them.
 501                  *
 502                  * If the compression fails for any reason, we set the pages
 503                  * dirty again later on.
 504                  */
 505                 extent_range_clear_dirty_for_io(inode, start, end);
 506                 redirty = 1;
 507                 ret = btrfs_compress_pages(compress_type,
 508                                            inode->i_mapping, start,
 509                                            total_compressed, pages,
 510                                            nr_pages, &nr_pages_ret,
 511                                            &total_in,
 512                                            &total_compressed,
 513                                            max_compressed);
 514
 515                 if (!ret) {
 516                         unsigned long offset = total_compressed &
 517                                 (PAGE_SIZE - 1);
 518                         struct page *page = pages[nr_pages_ret - 1];
 519                         char *kaddr;
 520
 521                         /* zero the tail end of the last page, we might be
 522                          * sending it down to disk
 523                          */
 524                         if (offset) {
 525                                 kaddr = kmap_atomic(page);
 526                                 memset(kaddr + offset, 0,
 527                                        PAGE_SIZE - offset);
 528                                 kunmap_atomic(kaddr);
 529                         }
 530                         will_compress = 1;
 531                 }
 532         }
 533 cont:
 534         if (start == 0) {
 535                 /* lets try to make an inline extent */
 536                 if (ret || total_in < (actual_end - start)) {
 537                         /* we didn't compress the entire range, try
 538                          * to make an uncompressed inline extent.
 539                          */
 540                         ret = cow_file_range_inline(root, inode, start, end,
 541                                                     0, 0, NULL);
 542                 } else {
 543                         /* try making a compressed inline extent */
 544                         ret = cow_file_range_inline(root, inode, start, end,
 545                                                     total_compressed,
 546                                                     compress_type, pages);
 547                 }
 548                 if (ret <= 0) {
 549                         unsigned long clear_flags = EXTENT_DELALLOC |
 550                                 EXTENT_DEFRAG;
 551                         unsigned long page_error_op;
 552
 553                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 554                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 555
 556                         /*
 557                          * inline extent creation worked or returned error,
 558                          * we don't need to create any more async work items.
 559                          * Unlock and free up our temp pages.
 560                          */
 561                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 562                                                      clear_flags, PAGE_UNLOCK |
 563                                                      PAGE_CLEAR_DIRTY |
 564                                                      PAGE_SET_WRITEBACK |
 565                                                      page_error_op |
 566                                                      PAGE_END_WRITEBACK);
 567                         goto free_pages_out;
 568                 }
 569         }
 570
 571         if (will_compress) {
 572                 /*
 573                  * we aren't doing an inline extent round the compressed size
 574                  * up to a block size boundary so the allocator does sane
 575                  * things
 576                  */
 577                 total_compressed = ALIGN(total_compressed, blocksize);
 578
 579                 /*
 580                  * one last check to make sure the compression is really a
 581                  * win, compare the page count read with the blocks on disk
 582                  */
 583                 total_in = ALIGN(total_in, PAGE_SIZE);
 584                 if (total_compressed >= total_in) {
 585                         will_compress = 0;
 586                 } else {
 587                         num_bytes = total_in;
 588                         *num_added += 1;
 589
 590                         /*
 591                          * The async work queues will take care of doing actual
 592                          * allocation on disk for these compressed pages, and
 593                          * will submit them to the elevator.
 594                          */
 595                         add_async_extent(async_cow, start, num_bytes,
 596                                         total_compressed, pages, nr_pages_ret,
 597                                         compress_type);
 598
 599                         if (start + num_bytes < end) {
 600                                 start += num_bytes;
 601                                 pages = NULL;
 602                                 cond_resched();
 603                                 goto again;
 604                         }
 605                         return;
 606                 }
 607         }
 608         if (pages) {
 609                 /*
 610                  * the compression code ran but failed to make things smaller,
 611                  * free any pages it allocated and our page pointer array
 612                  */
 613                 for (i = 0; i < nr_pages_ret; i++) {
 614                         WARN_ON(pages[i]->mapping);
 615                         put_page(pages[i]);
 616                 }
 617                 kfree(pages);
 618                 pages = NULL;
 619                 total_compressed = 0;
 620                 nr_pages_ret = 0;
 621
 622                 /* flag the file so we don't compress in the future */
 623                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 624                     !(BTRFS_I(inode)->force_compress)) {
 625                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 626                 }
 627         }
 628 cleanup_and_bail_uncompressed:
 629         /*
 630          * No compression, but we still need to write the pages in the file
 631          * we've been given so far.  redirty the locked page if it corresponds
 632          * to our extent and set things up for the async work queue to run
 633          * cow_file_range to do the normal delalloc dance.
 634          */
 635         if (page_offset(locked_page) >= start &&
 636             page_offset(locked_page) <= end)
 637                 __set_page_dirty_nobuffers(locked_page);
 638                 /* unlocked later on in the async handlers */
 639
 640         if (redirty)
 641                 extent_range_redirty_for_io(inode, start, end);
 642         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 643                          BTRFS_COMPRESS_NONE);
 644         *num_added += 1;
 645
 646         return;
 647
 648 free_pages_out:
 649         for (i = 0; i < nr_pages_ret; i++) {
 650                 WARN_ON(pages[i]->mapping);
 651                 put_page(pages[i]);
 652         }
 653         kfree(pages);
 654 }
 655
 656 static void free_async_extent_pages(struct async_extent *async_extent)
 657 {
 658         int i;
 659
 660         if (!async_extent->pages)
 661                 return;
 662
 663         for (i = 0; i < async_extent->nr_pages; i++) {
 664                 WARN_ON(async_extent->pages[i]->mapping);
 665                 put_page(async_extent->pages[i]);
 666         }
 667         kfree(async_extent->pages);
 668         async_extent->nr_pages = 0;
 669         async_extent->pages = NULL;
 670 }
 671
 672 /*
 673  * phase two of compressed writeback.  This is the ordered portion
 674  * of the code, which only gets called in the order the work was
 675  * queued.  We walk all the async extents created by compress_file_range
 676  * and send them down to the disk.
 677  */
 678 static noinline void submit_compressed_extents(struct inode *inode,
 679                                               struct async_cow *async_cow)
 680 {
 681         struct async_extent *async_extent;
 682         u64 alloc_hint = 0;
 683         struct btrfs_key ins;
 684         struct extent_map *em;
 685         struct btrfs_root *root = BTRFS_I(inode)->root;
 686         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 687         struct extent_io_tree *io_tree;
 688         int ret = 0;
 689
 690 again:
 691         while (!list_empty(&async_cow->extents)) {
 692                 async_extent = list_entry(async_cow->extents.next,
 693                                           struct async_extent, list);
 694                 list_del(&async_extent->list);
 695
 696                 io_tree = &BTRFS_I(inode)->io_tree;
 697
 698 retry:
 699                 /* did the compression code fall back to uncompressed IO? */
 700                 if (!async_extent->pages) {
 701                         int page_started = 0;
 702                         unsigned long nr_written = 0;
 703
 704                         lock_extent(io_tree, async_extent->start,
 705                                          async_extent->start +
 706                                          async_extent->ram_size - 1);
 707
 708                         /* allocate blocks */
 709                         ret = cow_file_range(inode, async_cow->locked_page,
 710                                              async_extent->start,
 711                                              async_extent->start +
 712                                              async_extent->ram_size - 1,
 713                                              &page_started, &nr_written, 0);
 714
 715                         /* JDM XXX */
 716
 717                         /*
 718                          * if page_started, cow_file_range inserted an
 719                          * inline extent and took care of all the unlocking
 720                          * and IO for us.  Otherwise, we need to submit
 721                          * all those pages down to the drive.
 722                          */
 723                         if (!page_started && !ret)
 724                                 extent_write_locked_range(io_tree,
 725                                                   inode, async_extent->start,
 726                                                   async_extent->start +
 727                                                   async_extent->ram_size - 1,
 728                                                   btrfs_get_extent,
 729                                                   WB_SYNC_ALL);
 730                         else if (ret)
 731                                 unlock_page(async_cow->locked_page);
 732                         kfree(async_extent);
 733                         cond_resched();
 734                         continue;
 735                 }
 736
 737                 lock_extent(io_tree, async_extent->start,
 738                             async_extent->start + async_extent->ram_size - 1);
 739
 740                 ret = btrfs_reserve_extent(root,
 741                                            async_extent->compressed_size,
 742                                            async_extent->compressed_size,
 743                                            0, alloc_hint, &ins, 1, 1);
 744                 if (ret) {
 745                         free_async_extent_pages(async_extent);
 746
 747                         if (ret == -ENOSPC) {
 748                                 unlock_extent(io_tree, async_extent->start,
 749                                               async_extent->start +
 750                                               async_extent->ram_size - 1);
 751
 752                                 /*
 753                                  * we need to redirty the pages if we decide to
 754                                  * fallback to uncompressed IO, otherwise we
 755                                  * will not submit these pages down to lower
 756                                  * layers.
 757                                  */
 758                                 extent_range_redirty_for_io(inode,
 759                                                 async_extent->start,
 760                                                 async_extent->start +
 761                                                 async_extent->ram_size - 1);
 762
 763                                 goto retry;
 764                         }
 765                         goto out_free;
 766                 }
 767                 /*
 768                  * here we're doing allocation and writeback of the
 769                  * compressed pages
 770                  */
 771                 btrfs_drop_extent_cache(inode, async_extent->start,
 772                                         async_extent->start +
 773                                         async_extent->ram_size - 1, 0);
 774
 775                 em = alloc_extent_map();
 776                 if (!em) {
 777                         ret = -ENOMEM;
 778                         goto out_free_reserve;
 779                 }
 780                 em->start = async_extent->start;
 781                 em->len = async_extent->ram_size;
 782                 em->orig_start = em->start;
 783                 em->mod_start = em->start;
 784                 em->mod_len = em->len;
 785
 786                 em->block_start = ins.objectid;
 787                 em->block_len = ins.offset;
 788                 em->orig_block_len = ins.offset;
 789                 em->ram_bytes = async_extent->ram_size;
 790                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 791                 em->compress_type = async_extent->compress_type;
 792                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 793                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 794                 em->generation = -1;
 795
 796                 while (1) {
 797                         write_lock(&em_tree->lock);
 798                         ret = add_extent_mapping(em_tree, em, 1);
 799                         write_unlock(&em_tree->lock);
 800                         if (ret != -EEXIST) {
 801                                 free_extent_map(em);
 802                                 break;
 803                         }
 804                         btrfs_drop_extent_cache(inode, async_extent->start,
 805                                                 async_extent->start +
 806                                                 async_extent->ram_size - 1, 0);
 807                 }
 808
 809                 if (ret)
 810                         goto out_free_reserve;
 811
 812                 ret = btrfs_add_ordered_extent_compress(inode,
 813                                                 async_extent->start,
 814                                                 ins.objectid,
 815                                                 async_extent->ram_size,
 816                                                 ins.offset,
 817                                                 BTRFS_ORDERED_COMPRESSED,
 818                                                 async_extent->compress_type);
 819                 if (ret) {
 820                         btrfs_drop_extent_cache(inode, async_extent->start,
 821                                                 async_extent->start +
 822                                                 async_extent->ram_size - 1, 0);
 823                         goto out_free_reserve;
 824                 }
 825                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 826
 827                 /*
 828                  * clear dirty, set writeback and unlock the pages.
 829                  */
 830                 extent_clear_unlock_delalloc(inode, async_extent->start,
 831                                 async_extent->start +
 832                                 async_extent->ram_size - 1,
 833                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 834                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 835                                 PAGE_SET_WRITEBACK);
 836                 ret = btrfs_submit_compressed_write(inode,
 837                                     async_extent->start,
 838                                     async_extent->ram_size,
 839                                     ins.objectid,
 840                                     ins.offset, async_extent->pages,
 841                                     async_extent->nr_pages);
 842                 if (ret) {
 843                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 844                         struct page *p = async_extent->pages[0];
 845                         const u64 start = async_extent->start;
 846                         const u64 end = start + async_extent->ram_size - 1;
 847
 848                         p->mapping = inode->i_mapping;
 849                         tree->ops->writepage_end_io_hook(p, start, end,
 850                                                          NULL, 0);
 851                         p->mapping = NULL;
 852                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 853                                                      PAGE_END_WRITEBACK |
 854                                                      PAGE_SET_ERROR);
 855                         free_async_extent_pages(async_extent);
 856                 }
 857                 alloc_hint = ins.objectid + ins.offset;
 858                 kfree(async_extent);
 859                 cond_resched();
 860         }
 861         return;
 862 out_free_reserve:
 863         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 864         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 865 out_free:
 866         extent_clear_unlock_delalloc(inode, async_extent->start,
 867                                      async_extent->start +
 868                                      async_extent->ram_size - 1,
 869                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 870                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 871                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 872                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 873                                      PAGE_SET_ERROR);
 874         free_async_extent_pages(async_extent);
 875         kfree(async_extent);
 876         goto again;
 877 }
 878
 879 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 880                                       u64 num_bytes)
 881 {
 882         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 883         struct extent_map *em;
 884         u64 alloc_hint = 0;
 885
 886         read_lock(&em_tree->lock);
 887         em = search_extent_mapping(em_tree, start, num_bytes);
 888         if (em) {
 889                 /*
 890                  * if block start isn't an actual block number then find the
 891                  * first block in this inode and use that as a hint.  If that
 892                  * block is also bogus then just don't worry about it.
 893                  */
 894                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 895                         free_extent_map(em);
 896                         em = search_extent_mapping(em_tree, 0, 0);
 897                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 898                                 alloc_hint = em->block_start;
 899                         if (em)
 900                                 free_extent_map(em);
 901                 } else {
 902                         alloc_hint = em->block_start;
 903                         free_extent_map(em);
 904                 }
 905         }
 906         read_unlock(&em_tree->lock);
 907
 908         return alloc_hint;
 909 }
 910
 911 /*
 912  * when extent_io.c finds a delayed allocation range in the file,
 913  * the call backs end up in this code.  The basic idea is to
 914  * allocate extents on disk for the range, and create ordered data structs
 915  * in ram to track those extents.
 916  *
 917  * locked_page is the page that writepage had locked already.  We use
 918  * it to make sure we don't do extra locks or unlocks.
 919  *
 920  * *page_started is set to one if we unlock locked_page and do everything
 921  * required to start IO on it.  It may be clean and already done with
 922  * IO when we return.
 923  */
 924 static noinline int cow_file_range(struct inode *inode,
 925                                    struct page *locked_page,
 926                                    u64 start, u64 end, int *page_started,
 927                                    unsigned long *nr_written,
 928                                    int unlock)
 929 {
 930         struct btrfs_root *root = BTRFS_I(inode)->root;
 931         u64 alloc_hint = 0;
 932         u64 num_bytes;
 933         unsigned long ram_size;
 934         u64 disk_num_bytes;
 935         u64 cur_alloc_size;
 936         u64 blocksize = root->sectorsize;
 937         struct btrfs_key ins;
 938         struct extent_map *em;
 939         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 940         int ret = 0;
 941
 942         if (btrfs_is_free_space_inode(inode)) {
 943                 WARN_ON_ONCE(1);
 944                 ret = -EINVAL;
 945                 goto out_unlock;
 946         }
 947
 948         num_bytes = ALIGN(end - start + 1, blocksize);
 949         num_bytes = max(blocksize,  num_bytes);
 950         disk_num_bytes = num_bytes;
 951
 952         /* if this is a small write inside eof, kick off defrag */
 953         if (num_bytes < SZ_64K &&
 954             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 955                 btrfs_add_inode_defrag(NULL, inode);
 956
 957         if (start == 0) {
 958                 /* lets try to make an inline extent */
 959                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 960                                             NULL);
 961                 if (ret == 0) {
 962                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 963                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 964                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 965                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 966                                      PAGE_END_WRITEBACK);
 967
 968                         *nr_written = *nr_written +
 969                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 970                         *page_started = 1;
 971                         goto out;
 972                 } else if (ret < 0) {
 973                         goto out_unlock;
 974                 }
 975         }
 976
 977         BUG_ON(disk_num_bytes >
 978                btrfs_super_total_bytes(root->fs_info->super_copy));
 979
 980         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 981         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 982
 983         while (disk_num_bytes > 0) {
 984                 unsigned long op;
 985
 986                 cur_alloc_size = disk_num_bytes;
 987                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 988                                            root->sectorsize, 0, alloc_hint,
 989                                            &ins, 1, 1);
 990                 if (ret < 0)
 991                         goto out_unlock;
 992
 993                 em = alloc_extent_map();
 994                 if (!em) {
 995                         ret = -ENOMEM;
 996                         goto out_reserve;
 997                 }
 998                 em->start = start;
 999                 em->orig_start = em->start;
1000                 ram_size = ins.offset;
1001                 em->len = ins.offset;
1002                 em->mod_start = em->start;
1003                 em->mod_len = em->len;
1004
1005                 em->block_start = ins.objectid;
1006                 em->block_len = ins.offset;
1007                 em->orig_block_len = ins.offset;
1008                 em->ram_bytes = ram_size;
1009                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1010                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1011                 em->generation = -1;
1012
1013                 while (1) {
1014                         write_lock(&em_tree->lock);
1015                         ret = add_extent_mapping(em_tree, em, 1);
1016                         write_unlock(&em_tree->lock);
1017                         if (ret != -EEXIST) {
1018                                 free_extent_map(em);
1019                                 break;
1020                         }
1021                         btrfs_drop_extent_cache(inode, start,
1022                                                 start + ram_size - 1, 0);
1023                 }
1024                 if (ret)
1025                         goto out_reserve;
1026
1027                 cur_alloc_size = ins.offset;
1028                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1029                                                ram_size, cur_alloc_size, 0);
1030                 if (ret)
1031                         goto out_drop_extent_cache;
1032
1033                 if (root->root_key.objectid ==
1034                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1035                         ret = btrfs_reloc_clone_csums(inode, start,
1036                                                       cur_alloc_size);
1037                         if (ret)
1038                                 goto out_drop_extent_cache;
1039                 }
1040
1041                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1042
1043                 if (disk_num_bytes < cur_alloc_size)
1044                         break;
1045
1046                 /* we're not doing compressed IO, don't unlock the first
1047                  * page (which the caller expects to stay locked), don't
1048                  * clear any dirty bits and don't set any writeback bits
1049                  *
1050                  * Do set the Private2 bit so we know this page was properly
1051                  * setup for writepage
1052                  */
1053                 op = unlock ? PAGE_UNLOCK : 0;
1054                 op |= PAGE_SET_PRIVATE2;
1055
1056                 extent_clear_unlock_delalloc(inode, start,
1057                                              start + ram_size - 1, locked_page,
1058                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1059                                              op);
1060                 disk_num_bytes -= cur_alloc_size;
1061                 num_bytes -= cur_alloc_size;
1062                 alloc_hint = ins.objectid + ins.offset;
1063                 start += cur_alloc_size;
1064         }
1065 out:
1066         return ret;
1067
1068 out_drop_extent_cache:
1069         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1070 out_reserve:
1071         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1072         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1073 out_unlock:
1074         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1075                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1076                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1077                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1078                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1079         goto out;
1080 }
1081
1082 /*
1083  * work queue call back to started compression on a file and pages
1084  */
1085 static noinline void async_cow_start(struct btrfs_work *work)
1086 {
1087         struct async_cow *async_cow;
1088         int num_added = 0;
1089         async_cow = container_of(work, struct async_cow, work);
1090
1091         compress_file_range(async_cow->inode, async_cow->locked_page,
1092                             async_cow->start, async_cow->end, async_cow,
1093                             &num_added);
1094         if (num_added == 0) {
1095                 btrfs_add_delayed_iput(async_cow->inode);
1096                 async_cow->inode = NULL;
1097         }
1098 }
1099
1100 /*
1101  * work queue call back to submit previously compressed pages
1102  */
1103 static noinline void async_cow_submit(struct btrfs_work *work)
1104 {
1105         struct async_cow *async_cow;
1106         struct btrfs_root *root;
1107         unsigned long nr_pages;
1108
1109         async_cow = container_of(work, struct async_cow, work);
1110
1111         root = async_cow->root;
1112         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1113                 PAGE_SHIFT;
1114
1115         /*
1116          * atomic_sub_return implies a barrier for waitqueue_active
1117          */
1118         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1119             5 * SZ_1M &&
1120             waitqueue_active(&root->fs_info->async_submit_wait))
1121                 wake_up(&root->fs_info->async_submit_wait);
1122
1123         if (async_cow->inode)
1124                 submit_compressed_extents(async_cow->inode, async_cow);
1125 }
1126
1127 static noinline void async_cow_free(struct btrfs_work *work)
1128 {
1129         struct async_cow *async_cow;
1130         async_cow = container_of(work, struct async_cow, work);
1131         if (async_cow->inode)
1132                 btrfs_add_delayed_iput(async_cow->inode);
1133         kfree(async_cow);
1134 }
1135
1136 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1137                                 u64 start, u64 end, int *page_started,
1138                                 unsigned long *nr_written)
1139 {
1140         struct async_cow *async_cow;
1141         struct btrfs_root *root = BTRFS_I(inode)->root;
1142         unsigned long nr_pages;
1143         u64 cur_end;
1144         int limit = 10 * SZ_1M;
1145
1146         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1147                          1, 0, NULL, GFP_NOFS);
1148         while (start < end) {
1149                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1150                 BUG_ON(!async_cow); /* -ENOMEM */
1151                 async_cow->inode = igrab(inode);
1152                 async_cow->root = root;
1153                 async_cow->locked_page = locked_page;
1154                 async_cow->start = start;
1155
1156                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1157                     !btrfs_test_opt(root, FORCE_COMPRESS))
1158                         cur_end = end;
1159                 else
1160                         cur_end = min(end, start + SZ_512K - 1);
1161
1162                 async_cow->end = cur_end;
1163                 INIT_LIST_HEAD(&async_cow->extents);
1164
1165                 btrfs_init_work(&async_cow->work,
1166                                 btrfs_delalloc_helper,
1167                                 async_cow_start, async_cow_submit,
1168                                 async_cow_free);
1169
1170                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1171                         PAGE_SHIFT;
1172                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1173
1174                 btrfs_queue_work(root->fs_info->delalloc_workers,
1175                                  &async_cow->work);
1176
1177                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1178                         wait_event(root->fs_info->async_submit_wait,
1179                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1180                             limit));
1181                 }
1182
1183                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1184                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1185                         wait_event(root->fs_info->async_submit_wait,
1186                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1187                            0));
1188                 }
1189
1190                 *nr_written += nr_pages;
1191                 start = cur_end + 1;
1192         }
1193         *page_started = 1;
1194         return 0;
1195 }
1196
1197 static noinline int csum_exist_in_range(struct btrfs_root *root,
1198                                         u64 bytenr, u64 num_bytes)
1199 {
1200         int ret;
1201         struct btrfs_ordered_sum *sums;
1202         LIST_HEAD(list);
1203
1204         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1205                                        bytenr + num_bytes - 1, &list, 0);
1206         if (ret == 0 && list_empty(&list))
1207                 return 0;
1208
1209         while (!list_empty(&list)) {
1210                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1211                 list_del(&sums->list);
1212                 kfree(sums);
1213         }
1214         return 1;
1215 }
1216
1217 /*
1218  * when nowcow writeback call back.  This checks for snapshots or COW copies
1219  * of the extents that exist in the file, and COWs the file as required.
1220  *
1221  * If no cow copies or snapshots exist, we write directly to the existing
1222  * blocks on disk
1223  */
1224 static noinline int run_delalloc_nocow(struct inode *inode,
1225                                        struct page *locked_page,
1226                               u64 start, u64 end, int *page_started, int force,
1227                               unsigned long *nr_written)
1228 {
1229         struct btrfs_root *root = BTRFS_I(inode)->root;
1230         struct btrfs_trans_handle *trans;
1231         struct extent_buffer *leaf;
1232         struct btrfs_path *path;
1233         struct btrfs_file_extent_item *fi;
1234         struct btrfs_key found_key;
1235         u64 cow_start;
1236         u64 cur_offset;
1237         u64 extent_end;
1238         u64 extent_offset;
1239         u64 disk_bytenr;
1240         u64 num_bytes;
1241         u64 disk_num_bytes;
1242         u64 ram_bytes;
1243         int extent_type;
1244         int ret, err;
1245         int type;
1246         int nocow;
1247         int check_prev = 1;
1248         bool nolock;
1249         u64 ino = btrfs_ino(inode);
1250
1251         path = btrfs_alloc_path();
1252         if (!path) {
1253                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1254                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1255                                              EXTENT_DO_ACCOUNTING |
1256                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1257                                              PAGE_CLEAR_DIRTY |
1258                                              PAGE_SET_WRITEBACK |
1259                                              PAGE_END_WRITEBACK);
1260                 return -ENOMEM;
1261         }
1262
1263         nolock = btrfs_is_free_space_inode(inode);
1264
1265         if (nolock)
1266                 trans = btrfs_join_transaction_nolock(root);
1267         else
1268                 trans = btrfs_join_transaction(root);
1269
1270         if (IS_ERR(trans)) {
1271                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1272                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1273                                              EXTENT_DO_ACCOUNTING |
1274                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1275                                              PAGE_CLEAR_DIRTY |
1276                                              PAGE_SET_WRITEBACK |
1277                                              PAGE_END_WRITEBACK);
1278                 btrfs_free_path(path);
1279                 return PTR_ERR(trans);
1280         }
1281
1282         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1283
1284         cow_start = (u64)-1;
1285         cur_offset = start;
1286         while (1) {
1287                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1288                                                cur_offset, 0);
1289                 if (ret < 0)
1290                         goto error;
1291                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1292                         leaf = path->nodes[0];
1293                         btrfs_item_key_to_cpu(leaf, &found_key,
1294                                               path->slots[0] - 1);
1295                         if (found_key.objectid == ino &&
1296                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1297                                 path->slots[0]--;
1298                 }
1299                 check_prev = 0;
1300 next_slot:
1301                 leaf = path->nodes[0];
1302                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1303                         ret = btrfs_next_leaf(root, path);
1304                         if (ret < 0)
1305                                 goto error;
1306                         if (ret > 0)
1307                                 break;
1308                         leaf = path->nodes[0];
1309                 }
1310
1311                 nocow = 0;
1312                 disk_bytenr = 0;
1313                 num_bytes = 0;
1314                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1315
1316                 if (found_key.objectid > ino)
1317                         break;
1318                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1319                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1320                         path->slots[0]++;
1321                         goto next_slot;
1322                 }
1323                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1324                     found_key.offset > end)
1325                         break;
1326
1327                 if (found_key.offset > cur_offset) {
1328                         extent_end = found_key.offset;
1329                         extent_type = 0;
1330                         goto out_check;
1331                 }
1332
1333                 fi = btrfs_item_ptr(leaf, path->slots[0],
1334                                     struct btrfs_file_extent_item);
1335                 extent_type = btrfs_file_extent_type(leaf, fi);
1336
1337                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1338                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1339                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1340                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1341                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1342                         extent_end = found_key.offset +
1343                                 btrfs_file_extent_num_bytes(leaf, fi);
1344                         disk_num_bytes =
1345                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1346                         if (extent_end <= start) {
1347                                 path->slots[0]++;
1348                                 goto next_slot;
1349                         }
1350                         if (disk_bytenr == 0)
1351                                 goto out_check;
1352                         if (btrfs_file_extent_compression(leaf, fi) ||
1353                             btrfs_file_extent_encryption(leaf, fi) ||
1354                             btrfs_file_extent_other_encoding(leaf, fi))
1355                                 goto out_check;
1356                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1357                                 goto out_check;
1358                         if (btrfs_extent_readonly(root, disk_bytenr))
1359                                 goto out_check;
1360                         if (btrfs_cross_ref_exist(trans, root, ino,
1361                                                   found_key.offset -
1362                                                   extent_offset, disk_bytenr))
1363                                 goto out_check;
1364                         disk_bytenr += extent_offset;
1365                         disk_bytenr += cur_offset - found_key.offset;
1366                         num_bytes = min(end + 1, extent_end) - cur_offset;
1367                         /*
1368                          * if there are pending snapshots for this root,
1369                          * we fall into common COW way.
1370                          */
1371                         if (!nolock) {
1372                                 err = btrfs_start_write_no_snapshoting(root);
1373                                 if (!err)
1374                                         goto out_check;
1375                         }
1376                         /*
1377                          * force cow if csum exists in the range.
1378                          * this ensure that csum for a given extent are
1379                          * either valid or do not exist.
1380                          */
1381                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1382                                 goto out_check;
1383                         if (!btrfs_inc_nocow_writers(root->fs_info,
1384                                                      disk_bytenr))
1385                                 goto out_check;
1386                         nocow = 1;
1387                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1388                         extent_end = found_key.offset +
1389                                 btrfs_file_extent_inline_len(leaf,
1390                                                      path->slots[0], fi);
1391                         extent_end = ALIGN(extent_end, root->sectorsize);
1392                 } else {
1393                         BUG_ON(1);
1394                 }
1395 out_check:
1396                 if (extent_end <= start) {
1397                         path->slots[0]++;
1398                         if (!nolock && nocow)
1399                                 btrfs_end_write_no_snapshoting(root);
1400                         if (nocow)
1401                                 btrfs_dec_nocow_writers(root->fs_info,
1402                                                         disk_bytenr);
1403                         goto next_slot;
1404                 }
1405                 if (!nocow) {
1406                         if (cow_start == (u64)-1)
1407                                 cow_start = cur_offset;
1408                         cur_offset = extent_end;
1409                         if (cur_offset > end)
1410                                 break;
1411                         path->slots[0]++;
1412                         goto next_slot;
1413                 }
1414
1415                 btrfs_release_path(path);
1416                 if (cow_start != (u64)-1) {
1417                         ret = cow_file_range(inode, locked_page,
1418                                              cow_start, found_key.offset - 1,
1419                                              page_started, nr_written, 1);
1420                         if (ret) {
1421                                 if (!nolock && nocow)
1422                                         btrfs_end_write_no_snapshoting(root);
1423                                 if (nocow)
1424                                         btrfs_dec_nocow_writers(root->fs_info,
1425                                                                 disk_bytenr);
1426                                 goto error;
1427                         }
1428                         cow_start = (u64)-1;
1429                 }
1430
1431                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1432                         struct extent_map *em;
1433                         struct extent_map_tree *em_tree;
1434                         em_tree = &BTRFS_I(inode)->extent_tree;
1435                         em = alloc_extent_map();
1436                         BUG_ON(!em); /* -ENOMEM */
1437                         em->start = cur_offset;
1438                         em->orig_start = found_key.offset - extent_offset;
1439                         em->len = num_bytes;
1440                         em->block_len = num_bytes;
1441                         em->block_start = disk_bytenr;
1442                         em->orig_block_len = disk_num_bytes;
1443                         em->ram_bytes = ram_bytes;
1444                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1445                         em->mod_start = em->start;
1446                         em->mod_len = em->len;
1447                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1448                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1449                         em->generation = -1;
1450                         while (1) {
1451                                 write_lock(&em_tree->lock);
1452                                 ret = add_extent_mapping(em_tree, em, 1);
1453                                 write_unlock(&em_tree->lock);
1454                                 if (ret != -EEXIST) {
1455                                         free_extent_map(em);
1456                                         break;
1457                                 }
1458                                 btrfs_drop_extent_cache(inode, em->start,
1459                                                 em->start + em->len - 1, 0);
1460                         }
1461                         type = BTRFS_ORDERED_PREALLOC;
1462                 } else {
1463                         type = BTRFS_ORDERED_NOCOW;
1464                 }
1465
1466                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1467                                                num_bytes, num_bytes, type);
1468                 if (nocow)
1469                         btrfs_dec_nocow_writers(root->fs_info, disk_bytenr);
1470                 BUG_ON(ret); /* -ENOMEM */
1471
1472                 if (root->root_key.objectid ==
1473                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1474                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1475                                                       num_bytes);
1476                         if (ret) {
1477                                 if (!nolock && nocow)
1478                                         btrfs_end_write_no_snapshoting(root);
1479                                 goto error;
1480                         }
1481                 }
1482
1483                 extent_clear_unlock_delalloc(inode, cur_offset,
1484                                              cur_offset + num_bytes - 1,
1485                                              locked_page, EXTENT_LOCKED |
1486                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1487                                              PAGE_SET_PRIVATE2);
1488                 if (!nolock && nocow)
1489                         btrfs_end_write_no_snapshoting(root);
1490                 cur_offset = extent_end;
1491                 if (cur_offset > end)
1492                         break;
1493         }
1494         btrfs_release_path(path);
1495
1496         if (cur_offset <= end && cow_start == (u64)-1) {
1497                 cow_start = cur_offset;
1498                 cur_offset = end;
1499         }
1500
1501         if (cow_start != (u64)-1) {
1502                 ret = cow_file_range(inode, locked_page, cow_start, end,
1503                                      page_started, nr_written, 1);
1504                 if (ret)
1505                         goto error;
1506         }
1507
1508 error:
1509         err = btrfs_end_transaction(trans, root);
1510         if (!ret)
1511                 ret = err;
1512
1513         if (ret && cur_offset < end)
1514                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1515                                              locked_page, EXTENT_LOCKED |
1516                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1517                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1518                                              PAGE_CLEAR_DIRTY |
1519                                              PAGE_SET_WRITEBACK |
1520                                              PAGE_END_WRITEBACK);
1521         btrfs_free_path(path);
1522         return ret;
1523 }
1524
1525 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1526 {
1527
1528         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1529             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1530                 return 0;
1531
1532         /*
1533          * @defrag_bytes is a hint value, no spinlock held here,
1534          * if is not zero, it means the file is defragging.
1535          * Force cow if given extent needs to be defragged.
1536          */
1537         if (BTRFS_I(inode)->defrag_bytes &&
1538             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1539                            EXTENT_DEFRAG, 0, NULL))
1540                 return 1;
1541
1542         return 0;
1543 }
1544
1545 /*
1546  * extent_io.c call back to do delayed allocation processing
1547  */
1548 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1549                               u64 start, u64 end, int *page_started,
1550                               unsigned long *nr_written)
1551 {
1552         int ret;
1553         int force_cow = need_force_cow(inode, start, end);
1554
1555         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1556                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1557                                          page_started, 1, nr_written);
1558         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1559                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1560                                          page_started, 0, nr_written);
1561         } else if (!inode_need_compress(inode)) {
1562                 ret = cow_file_range(inode, locked_page, start, end,
1563                                       page_started, nr_written, 1);
1564         } else {
1565                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1566                         &BTRFS_I(inode)->runtime_flags);
1567                 ret = cow_file_range_async(inode, locked_page, start, end,
1568                                            page_started, nr_written);
1569         }
1570         return ret;
1571 }
1572
1573 static void btrfs_split_extent_hook(struct inode *inode,
1574                                     struct extent_state *orig, u64 split)
1575 {
1576         u64 size;
1577
1578         /* not delalloc, ignore it */
1579         if (!(orig->state & EXTENT_DELALLOC))
1580                 return;
1581
1582         size = orig->end - orig->start + 1;
1583         if (size > BTRFS_MAX_EXTENT_SIZE) {
1584                 u64 num_extents;
1585                 u64 new_size;
1586
1587                 /*
1588                  * See the explanation in btrfs_merge_extent_hook, the same
1589                  * applies here, just in reverse.
1590                  */
1591                 new_size = orig->end - split + 1;
1592                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1593                                         BTRFS_MAX_EXTENT_SIZE);
1594                 new_size = split - orig->start;
1595                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1596                                         BTRFS_MAX_EXTENT_SIZE);
1597                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1598                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1599                         return;
1600         }
1601
1602         spin_lock(&BTRFS_I(inode)->lock);
1603         BTRFS_I(inode)->outstanding_extents++;
1604         spin_unlock(&BTRFS_I(inode)->lock);
1605 }
1606
1607 /*
1608  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1609  * extents so we can keep track of new extents that are just merged onto old
1610  * extents, such as when we are doing sequential writes, so we can properly
1611  * account for the metadata space we'll need.
1612  */
1613 static void btrfs_merge_extent_hook(struct inode *inode,
1614                                     struct extent_state *new,
1615                                     struct extent_state *other)
1616 {
1617         u64 new_size, old_size;
1618         u64 num_extents;
1619
1620         /* not delalloc, ignore it */
1621         if (!(other->state & EXTENT_DELALLOC))
1622                 return;
1623
1624         if (new->start > other->start)
1625                 new_size = new->end - other->start + 1;
1626         else
1627                 new_size = other->end - new->start + 1;
1628
1629         /* we're not bigger than the max, unreserve the space and go */
1630         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1631                 spin_lock(&BTRFS_I(inode)->lock);
1632                 BTRFS_I(inode)->outstanding_extents--;
1633                 spin_unlock(&BTRFS_I(inode)->lock);
1634                 return;
1635         }
1636
1637         /*
1638          * We have to add up either side to figure out how many extents were
1639          * accounted for before we merged into one big extent.  If the number of
1640          * extents we accounted for is <= the amount we need for the new range
1641          * then we can return, otherwise drop.  Think of it like this
1642          *
1643          * [ 4k][MAX_SIZE]
1644          *
1645          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1646          * need 2 outstanding extents, on one side we have 1 and the other side
1647          * we have 1 so they are == and we can return.  But in this case
1648          *
1649          * [MAX_SIZE+4k][MAX_SIZE+4k]
1650          *
1651          * Each range on their own accounts for 2 extents, but merged together
1652          * they are only 3 extents worth of accounting, so we need to drop in
1653          * this case.
1654          */
1655         old_size = other->end - other->start + 1;
1656         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1657                                 BTRFS_MAX_EXTENT_SIZE);
1658         old_size = new->end - new->start + 1;
1659         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1660                                  BTRFS_MAX_EXTENT_SIZE);
1661
1662         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1663                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1664                 return;
1665
1666         spin_lock(&BTRFS_I(inode)->lock);
1667         BTRFS_I(inode)->outstanding_extents--;
1668         spin_unlock(&BTRFS_I(inode)->lock);
1669 }
1670
1671 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1672                                       struct inode *inode)
1673 {
1674         spin_lock(&root->delalloc_lock);
1675         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1676                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1677                               &root->delalloc_inodes);
1678                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1679                         &BTRFS_I(inode)->runtime_flags);
1680                 root->nr_delalloc_inodes++;
1681                 if (root->nr_delalloc_inodes == 1) {
1682                         spin_lock(&root->fs_info->delalloc_root_lock);
1683                         BUG_ON(!list_empty(&root->delalloc_root));
1684                         list_add_tail(&root->delalloc_root,
1685                                       &root->fs_info->delalloc_roots);
1686                         spin_unlock(&root->fs_info->delalloc_root_lock);
1687                 }
1688         }
1689         spin_unlock(&root->delalloc_lock);
1690 }
1691
1692 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1693                                      struct inode *inode)
1694 {
1695         spin_lock(&root->delalloc_lock);
1696         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1697                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1698                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1699                           &BTRFS_I(inode)->runtime_flags);
1700                 root->nr_delalloc_inodes--;
1701                 if (!root->nr_delalloc_inodes) {
1702                         spin_lock(&root->fs_info->delalloc_root_lock);
1703                         BUG_ON(list_empty(&root->delalloc_root));
1704                         list_del_init(&root->delalloc_root);
1705                         spin_unlock(&root->fs_info->delalloc_root_lock);
1706                 }
1707         }
1708         spin_unlock(&root->delalloc_lock);
1709 }
1710
1711 /*
1712  * extent_io.c set_bit_hook, used to track delayed allocation
1713  * bytes in this file, and to maintain the list of inodes that
1714  * have pending delalloc work to be done.
1715  */
1716 static void btrfs_set_bit_hook(struct inode *inode,
1717                                struct extent_state *state, unsigned *bits)
1718 {
1719
1720         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1721                 WARN_ON(1);
1722         /*
1723          * set_bit and clear bit hooks normally require _irqsave/restore
1724          * but in this case, we are only testing for the DELALLOC
1725          * bit, which is only set or cleared with irqs on
1726          */
1727         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1728                 struct btrfs_root *root = BTRFS_I(inode)->root;
1729                 u64 len = state->end + 1 - state->start;
1730                 bool do_list = !btrfs_is_free_space_inode(inode);
1731
1732                 if (*bits & EXTENT_FIRST_DELALLOC) {
1733                         *bits &= ~EXTENT_FIRST_DELALLOC;
1734                 } else {
1735                         spin_lock(&BTRFS_I(inode)->lock);
1736                         BTRFS_I(inode)->outstanding_extents++;
1737                         spin_unlock(&BTRFS_I(inode)->lock);
1738                 }
1739
1740                 /* For sanity tests */
1741                 if (btrfs_test_is_dummy_root(root))
1742                         return;
1743
1744                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1745                                      root->fs_info->delalloc_batch);
1746                 spin_lock(&BTRFS_I(inode)->lock);
1747                 BTRFS_I(inode)->delalloc_bytes += len;
1748                 if (*bits & EXTENT_DEFRAG)
1749                         BTRFS_I(inode)->defrag_bytes += len;
1750                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1751                                          &BTRFS_I(inode)->runtime_flags))
1752                         btrfs_add_delalloc_inodes(root, inode);
1753                 spin_unlock(&BTRFS_I(inode)->lock);
1754         }
1755 }
1756
1757 /*
1758  * extent_io.c clear_bit_hook, see set_bit_hook for why
1759  */
1760 static void btrfs_clear_bit_hook(struct inode *inode,
1761                                  struct extent_state *state,
1762                                  unsigned *bits)
1763 {
1764         u64 len = state->end + 1 - state->start;
1765         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1766                                     BTRFS_MAX_EXTENT_SIZE);
1767
1768         spin_lock(&BTRFS_I(inode)->lock);
1769         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1770                 BTRFS_I(inode)->defrag_bytes -= len;
1771         spin_unlock(&BTRFS_I(inode)->lock);
1772
1773         /*
1774          * set_bit and clear bit hooks normally require _irqsave/restore
1775          * but in this case, we are only testing for the DELALLOC
1776          * bit, which is only set or cleared with irqs on
1777          */
1778         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1779                 struct btrfs_root *root = BTRFS_I(inode)->root;
1780                 bool do_list = !btrfs_is_free_space_inode(inode);
1781
1782                 if (*bits & EXTENT_FIRST_DELALLOC) {
1783                         *bits &= ~EXTENT_FIRST_DELALLOC;
1784                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1785                         spin_lock(&BTRFS_I(inode)->lock);
1786                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1787                         spin_unlock(&BTRFS_I(inode)->lock);
1788                 }
1789
1790                 /*
1791                  * We don't reserve metadata space for space cache inodes so we
1792                  * don't need to call dellalloc_release_metadata if there is an
1793                  * error.
1794                  */
1795                 if (*bits & EXTENT_DO_ACCOUNTING &&
1796                     root != root->fs_info->tree_root)
1797                         btrfs_delalloc_release_metadata(inode, len);
1798
1799                 /* For sanity tests. */
1800                 if (btrfs_test_is_dummy_root(root))
1801                         return;
1802
1803                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1804                     && do_list && !(state->state & EXTENT_NORESERVE))
1805                         btrfs_free_reserved_data_space_noquota(inode,
1806                                         state->start, len);
1807
1808                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1809                                      root->fs_info->delalloc_batch);
1810                 spin_lock(&BTRFS_I(inode)->lock);
1811                 BTRFS_I(inode)->delalloc_bytes -= len;
1812                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1813                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1814                              &BTRFS_I(inode)->runtime_flags))
1815                         btrfs_del_delalloc_inode(root, inode);
1816                 spin_unlock(&BTRFS_I(inode)->lock);
1817         }
1818 }
1819
1820 /*
1821  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1822  * we don't create bios that span stripes or chunks
1823  *
1824  * return 1 if page cannot be merged to bio
1825  * return 0 if page can be merged to bio
1826  * return error otherwise
1827  */
1828 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1829                          size_t size, struct bio *bio,
1830                          unsigned long bio_flags)
1831 {
1832         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1833         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1834         u64 length = 0;
1835         u64 map_length;
1836         int ret;
1837
1838         if (bio_flags & EXTENT_BIO_COMPRESSED)
1839                 return 0;
1840
1841         length = bio->bi_iter.bi_size;
1842         map_length = length;
1843         ret = btrfs_map_block(root->fs_info, rw, logical,
1844                               &map_length, NULL, 0);
1845         if (ret < 0)
1846                 return ret;
1847         if (map_length < length + size)
1848                 return 1;
1849         return 0;
1850 }
1851
1852 /*
1853  * in order to insert checksums into the metadata in large chunks,
1854  * we wait until bio submission time.   All the pages in the bio are
1855  * checksummed and sums are attached onto the ordered extent record.
1856  *
1857  * At IO completion time the cums attached on the ordered extent record
1858  * are inserted into the btree
1859  */
1860 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1861                                     struct bio *bio, int mirror_num,
1862                                     unsigned long bio_flags,
1863                                     u64 bio_offset)
1864 {
1865         struct btrfs_root *root = BTRFS_I(inode)->root;
1866         int ret = 0;
1867
1868         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1869         BUG_ON(ret); /* -ENOMEM */
1870         return 0;
1871 }
1872
1873 /*
1874  * in order to insert checksums into the metadata in large chunks,
1875  * we wait until bio submission time.   All the pages in the bio are
1876  * checksummed and sums are attached onto the ordered extent record.
1877  *
1878  * At IO completion time the cums attached on the ordered extent record
1879  * are inserted into the btree
1880  */
1881 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1882                           int mirror_num, unsigned long bio_flags,
1883                           u64 bio_offset)
1884 {
1885         struct btrfs_root *root = BTRFS_I(inode)->root;
1886         int ret;
1887
1888         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1889         if (ret) {
1890                 bio->bi_error = ret;
1891                 bio_endio(bio);
1892         }
1893         return ret;
1894 }
1895
1896 /*
1897  * extent_io.c submission hook. This does the right thing for csum calculation
1898  * on write, or reading the csums from the tree before a read
1899  */
1900 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1901                           int mirror_num, unsigned long bio_flags,
1902                           u64 bio_offset)
1903 {
1904         struct btrfs_root *root = BTRFS_I(inode)->root;
1905         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1906         int ret = 0;
1907         int skip_sum;
1908         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1909
1910         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1911
1912         if (btrfs_is_free_space_inode(inode))
1913                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1914
1915         if (!(rw & REQ_WRITE)) {
1916                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1917                 if (ret)
1918                         goto out;
1919
1920                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1921                         ret = btrfs_submit_compressed_read(inode, bio,
1922                                                            mirror_num,
1923                                                            bio_flags);
1924                         goto out;
1925                 } else if (!skip_sum) {
1926                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1927                         if (ret)
1928                                 goto out;
1929                 }
1930                 goto mapit;
1931         } else if (async && !skip_sum) {
1932                 /* csum items have already been cloned */
1933                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1934                         goto mapit;
1935                 /* we're doing a write, do the async checksumming */
1936                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1937                                    inode, rw, bio, mirror_num,
1938                                    bio_flags, bio_offset,
1939                                    __btrfs_submit_bio_start,
1940                                    __btrfs_submit_bio_done);
1941                 goto out;
1942         } else if (!skip_sum) {
1943                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1944                 if (ret)
1945                         goto out;
1946         }
1947
1948 mapit:
1949         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1950
1951 out:
1952         if (ret < 0) {
1953                 bio->bi_error = ret;
1954                 bio_endio(bio);
1955         }
1956         return ret;
1957 }
1958
1959 /*
1960  * given a list of ordered sums record them in the inode.  This happens
1961  * at IO completion time based on sums calculated at bio submission time.
1962  */
1963 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1964                              struct inode *inode, u64 file_offset,
1965                              struct list_head *list)
1966 {
1967         struct btrfs_ordered_sum *sum;
1968
1969         list_for_each_entry(sum, list, list) {
1970                 trans->adding_csums = 1;
1971                 btrfs_csum_file_blocks(trans,
1972                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1973                 trans->adding_csums = 0;
1974         }
1975         return 0;
1976 }
1977
1978 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1979                               struct extent_state **cached_state)
1980 {
1981         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
1982         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1983                                    cached_state);
1984 }
1985
1986 /* see btrfs_writepage_start_hook for details on why this is required */
1987 struct btrfs_writepage_fixup {
1988         struct page *page;
1989         struct btrfs_work work;
1990 };
1991
1992 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1993 {
1994         struct btrfs_writepage_fixup *fixup;
1995         struct btrfs_ordered_extent *ordered;
1996         struct extent_state *cached_state = NULL;
1997         struct page *page;
1998         struct inode *inode;
1999         u64 page_start;
2000         u64 page_end;
2001         int ret;
2002
2003         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2004         page = fixup->page;
2005 again:
2006         lock_page(page);
2007         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2008                 ClearPageChecked(page);
2009                 goto out_page;
2010         }
2011
2012         inode = page->mapping->host;
2013         page_start = page_offset(page);
2014         page_end = page_offset(page) + PAGE_SIZE - 1;
2015
2016         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2017                          &cached_state);
2018
2019         /* already ordered? We're done */
2020         if (PagePrivate2(page))
2021                 goto out;
2022
2023         ordered = btrfs_lookup_ordered_range(inode, page_start,
2024                                         PAGE_SIZE);
2025         if (ordered) {
2026                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2027                                      page_end, &cached_state, GFP_NOFS);
2028                 unlock_page(page);
2029                 btrfs_start_ordered_extent(inode, ordered, 1);
2030                 btrfs_put_ordered_extent(ordered);
2031                 goto again;
2032         }
2033
2034         ret = btrfs_delalloc_reserve_space(inode, page_start,
2035                                            PAGE_SIZE);
2036         if (ret) {
2037                 mapping_set_error(page->mapping, ret);
2038                 end_extent_writepage(page, ret, page_start, page_end);
2039                 ClearPageChecked(page);
2040                 goto out;
2041          }
2042
2043         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2044         ClearPageChecked(page);
2045         set_page_dirty(page);
2046 out:
2047         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2048                              &cached_state, GFP_NOFS);
2049 out_page:
2050         unlock_page(page);
2051         put_page(page);
2052         kfree(fixup);
2053 }
2054
2055 /*
2056  * There are a few paths in the higher layers of the kernel that directly
2057  * set the page dirty bit without asking the filesystem if it is a
2058  * good idea.  This causes problems because we want to make sure COW
2059  * properly happens and the data=ordered rules are followed.
2060  *
2061  * In our case any range that doesn't have the ORDERED bit set
2062  * hasn't been properly setup for IO.  We kick off an async process
2063  * to fix it up.  The async helper will wait for ordered extents, set
2064  * the delalloc bit and make it safe to write the page.
2065  */
2066 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2067 {
2068         struct inode *inode = page->mapping->host;
2069         struct btrfs_writepage_fixup *fixup;
2070         struct btrfs_root *root = BTRFS_I(inode)->root;
2071
2072         /* this page is properly in the ordered list */
2073         if (TestClearPagePrivate2(page))
2074                 return 0;
2075
2076         if (PageChecked(page))
2077                 return -EAGAIN;
2078
2079         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2080         if (!fixup)
2081                 return -EAGAIN;
2082
2083         SetPageChecked(page);
2084         get_page(page);
2085         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2086                         btrfs_writepage_fixup_worker, NULL, NULL);
2087         fixup->page = page;
2088         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2089         return -EBUSY;
2090 }
2091
2092 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2093                                        struct inode *inode, u64 file_pos,
2094                                        u64 disk_bytenr, u64 disk_num_bytes,
2095                                        u64 num_bytes, u64 ram_bytes,
2096                                        u8 compression, u8 encryption,
2097                                        u16 other_encoding, int extent_type)
2098 {
2099         struct btrfs_root *root = BTRFS_I(inode)->root;
2100         struct btrfs_file_extent_item *fi;
2101         struct btrfs_path *path;
2102         struct extent_buffer *leaf;
2103         struct btrfs_key ins;
2104         int extent_inserted = 0;
2105         int ret;
2106
2107         path = btrfs_alloc_path();
2108         if (!path)
2109                 return -ENOMEM;
2110
2111         /*
2112          * we may be replacing one extent in the tree with another.
2113          * The new extent is pinned in the extent map, and we don't want
2114          * to drop it from the cache until it is completely in the btree.
2115          *
2116          * So, tell btrfs_drop_extents to leave this extent in the cache.
2117          * the caller is expected to unpin it and allow it to be merged
2118          * with the others.
2119          */
2120         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2121                                    file_pos + num_bytes, NULL, 0,
2122                                    1, sizeof(*fi), &extent_inserted);
2123         if (ret)
2124                 goto out;
2125
2126         if (!extent_inserted) {
2127                 ins.objectid = btrfs_ino(inode);
2128                 ins.offset = file_pos;
2129                 ins.type = BTRFS_EXTENT_DATA_KEY;
2130
2131                 path->leave_spinning = 1;
2132                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2133                                               sizeof(*fi));
2134                 if (ret)
2135                         goto out;
2136         }
2137         leaf = path->nodes[0];
2138         fi = btrfs_item_ptr(leaf, path->slots[0],
2139                             struct btrfs_file_extent_item);
2140         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2141         btrfs_set_file_extent_type(leaf, fi, extent_type);
2142         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2143         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2144         btrfs_set_file_extent_offset(leaf, fi, 0);
2145         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2146         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2147         btrfs_set_file_extent_compression(leaf, fi, compression);
2148         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2149         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2150
2151         btrfs_mark_buffer_dirty(leaf);
2152         btrfs_release_path(path);
2153
2154         inode_add_bytes(inode, num_bytes);
2155
2156         ins.objectid = disk_bytenr;
2157         ins.offset = disk_num_bytes;
2158         ins.type = BTRFS_EXTENT_ITEM_KEY;
2159         ret = btrfs_alloc_reserved_file_extent(trans, root,
2160                                         root->root_key.objectid,
2161                                         btrfs_ino(inode), file_pos,
2162                                         ram_bytes, &ins);
2163         /*
2164          * Release the reserved range from inode dirty range map, as it is
2165          * already moved into delayed_ref_head
2166          */
2167         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2168 out:
2169         btrfs_free_path(path);
2170
2171         return ret;
2172 }
2173
2174 /* snapshot-aware defrag */
2175 struct sa_defrag_extent_backref {
2176         struct rb_node node;
2177         struct old_sa_defrag_extent *old;
2178         u64 root_id;
2179         u64 inum;
2180         u64 file_pos;
2181         u64 extent_offset;
2182         u64 num_bytes;
2183         u64 generation;
2184 };
2185
2186 struct old_sa_defrag_extent {
2187         struct list_head list;
2188         struct new_sa_defrag_extent *new;
2189
2190         u64 extent_offset;
2191         u64 bytenr;
2192         u64 offset;
2193         u64 len;
2194         int count;
2195 };
2196
2197 struct new_sa_defrag_extent {
2198         struct rb_root root;
2199         struct list_head head;
2200         struct btrfs_path *path;
2201         struct inode *inode;
2202         u64 file_pos;
2203         u64 len;
2204         u64 bytenr;
2205         u64 disk_len;
2206         u8 compress_type;
2207 };
2208
2209 static int backref_comp(struct sa_defrag_extent_backref *b1,
2210                         struct sa_defrag_extent_backref *b2)
2211 {
2212         if (b1->root_id < b2->root_id)
2213                 return -1;
2214         else if (b1->root_id > b2->root_id)
2215                 return 1;
2216
2217         if (b1->inum < b2->inum)
2218                 return -1;
2219         else if (b1->inum > b2->inum)
2220                 return 1;
2221
2222         if (b1->file_pos < b2->file_pos)
2223                 return -1;
2224         else if (b1->file_pos > b2->file_pos)
2225                 return 1;
2226
2227         /*
2228          * [------------------------------] ===> (a range of space)
2229          *     |<--->|   |<---->| =============> (fs/file tree A)
2230          * |<---------------------------->| ===> (fs/file tree B)
2231          *
2232          * A range of space can refer to two file extents in one tree while
2233          * refer to only one file extent in another tree.
2234          *
2235          * So we may process a disk offset more than one time(two extents in A)
2236          * and locate at the same extent(one extent in B), then insert two same
2237          * backrefs(both refer to the extent in B).
2238          */
2239         return 0;
2240 }
2241
2242 static void backref_insert(struct rb_root *root,
2243                            struct sa_defrag_extent_backref *backref)
2244 {
2245         struct rb_node **p = &root->rb_node;
2246         struct rb_node *parent = NULL;
2247         struct sa_defrag_extent_backref *entry;
2248         int ret;
2249
2250         while (*p) {
2251                 parent = *p;
2252                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2253
2254                 ret = backref_comp(backref, entry);
2255                 if (ret < 0)
2256                         p = &(*p)->rb_left;
2257                 else
2258                         p = &(*p)->rb_right;
2259         }
2260
2261         rb_link_node(&backref->node, parent, p);
2262         rb_insert_color(&backref->node, root);
2263 }
2264
2265 /*
2266  * Note the backref might has changed, and in this case we just return 0.
2267  */
2268 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2269                                        void *ctx)
2270 {
2271         struct btrfs_file_extent_item *extent;
2272         struct btrfs_fs_info *fs_info;
2273         struct old_sa_defrag_extent *old = ctx;
2274         struct new_sa_defrag_extent *new = old->new;
2275         struct btrfs_path *path = new->path;
2276         struct btrfs_key key;
2277         struct btrfs_root *root;
2278         struct sa_defrag_extent_backref *backref;
2279         struct extent_buffer *leaf;
2280         struct inode *inode = new->inode;
2281         int slot;
2282         int ret;
2283         u64 extent_offset;
2284         u64 num_bytes;
2285
2286         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2287             inum == btrfs_ino(inode))
2288                 return 0;
2289
2290         key.objectid = root_id;
2291         key.type = BTRFS_ROOT_ITEM_KEY;
2292         key.offset = (u64)-1;
2293
2294         fs_info = BTRFS_I(inode)->root->fs_info;
2295         root = btrfs_read_fs_root_no_name(fs_info, &key);
2296         if (IS_ERR(root)) {
2297                 if (PTR_ERR(root) == -ENOENT)
2298                         return 0;
2299                 WARN_ON(1);
2300                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2301                          inum, offset, root_id);
2302                 return PTR_ERR(root);
2303         }
2304
2305         key.objectid = inum;
2306         key.type = BTRFS_EXTENT_DATA_KEY;
2307         if (offset > (u64)-1 << 32)
2308                 key.offset = 0;
2309         else
2310                 key.offset = offset;
2311
2312         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2313         if (WARN_ON(ret < 0))
2314                 return ret;
2315         ret = 0;
2316
2317         while (1) {
2318                 cond_resched();
2319
2320                 leaf = path->nodes[0];
2321                 slot = path->slots[0];
2322
2323                 if (slot >= btrfs_header_nritems(leaf)) {
2324                         ret = btrfs_next_leaf(root, path);
2325                         if (ret < 0) {
2326                                 goto out;
2327                         } else if (ret > 0) {
2328                                 ret = 0;
2329                                 goto out;
2330                         }
2331                         continue;
2332                 }
2333
2334                 path->slots[0]++;
2335
2336                 btrfs_item_key_to_cpu(leaf, &key, slot);
2337
2338                 if (key.objectid > inum)
2339                         goto out;
2340
2341                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2342                         continue;
2343
2344                 extent = btrfs_item_ptr(leaf, slot,
2345                                         struct btrfs_file_extent_item);
2346
2347                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2348                         continue;
2349
2350                 /*
2351                  * 'offset' refers to the exact key.offset,
2352                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2353                  * (key.offset - extent_offset).
2354                  */
2355                 if (key.offset != offset)
2356                         continue;
2357
2358                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2359                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2360
2361                 if (extent_offset >= old->extent_offset + old->offset +
2362                     old->len || extent_offset + num_bytes <=
2363                     old->extent_offset + old->offset)
2364                         continue;
2365                 break;
2366         }
2367
2368         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2369         if (!backref) {
2370                 ret = -ENOENT;
2371                 goto out;
2372         }
2373
2374         backref->root_id = root_id;
2375         backref->inum = inum;
2376         backref->file_pos = offset;
2377         backref->num_bytes = num_bytes;
2378         backref->extent_offset = extent_offset;
2379         backref->generation = btrfs_file_extent_generation(leaf, extent);
2380         backref->old = old;
2381         backref_insert(&new->root, backref);
2382         old->count++;
2383 out:
2384         btrfs_release_path(path);
2385         WARN_ON(ret);
2386         return ret;
2387 }
2388
2389 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2390                                    struct new_sa_defrag_extent *new)
2391 {
2392         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2393         struct old_sa_defrag_extent *old, *tmp;
2394         int ret;
2395
2396         new->path = path;
2397
2398         list_for_each_entry_safe(old, tmp, &new->head, list) {
2399                 ret = iterate_inodes_from_logical(old->bytenr +
2400                                                   old->extent_offset, fs_info,
2401                                                   path, record_one_backref,
2402                                                   old);
2403                 if (ret < 0 && ret != -ENOENT)
2404                         return false;
2405
2406                 /* no backref to be processed for this extent */
2407                 if (!old->count) {
2408                         list_del(&old->list);
2409                         kfree(old);
2410                 }
2411         }
2412
2413         if (list_empty(&new->head))
2414                 return false;
2415
2416         return true;
2417 }
2418
2419 static int relink_is_mergable(struct extent_buffer *leaf,
2420                               struct btrfs_file_extent_item *fi,
2421                               struct new_sa_defrag_extent *new)
2422 {
2423         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2424                 return 0;
2425
2426         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2427                 return 0;
2428
2429         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2430                 return 0;
2431
2432         if (btrfs_file_extent_encryption(leaf, fi) ||
2433             btrfs_file_extent_other_encoding(leaf, fi))
2434                 return 0;
2435
2436         return 1;
2437 }
2438
2439 /*
2440  * Note the backref might has changed, and in this case we just return 0.
2441  */
2442 static noinline int relink_extent_backref(struct btrfs_path *path,
2443                                  struct sa_defrag_extent_backref *prev,
2444                                  struct sa_defrag_extent_backref *backref)
2445 {
2446         struct btrfs_file_extent_item *extent;
2447         struct btrfs_file_extent_item *item;
2448         struct btrfs_ordered_extent *ordered;
2449         struct btrfs_trans_handle *trans;
2450         struct btrfs_fs_info *fs_info;
2451         struct btrfs_root *root;
2452         struct btrfs_key key;
2453         struct extent_buffer *leaf;
2454         struct old_sa_defrag_extent *old = backref->old;
2455         struct new_sa_defrag_extent *new = old->new;
2456         struct inode *src_inode = new->inode;
2457         struct inode *inode;
2458         struct extent_state *cached = NULL;
2459         int ret = 0;
2460         u64 start;
2461         u64 len;
2462         u64 lock_start;
2463         u64 lock_end;
2464         bool merge = false;
2465         int index;
2466
2467         if (prev && prev->root_id == backref->root_id &&
2468             prev->inum == backref->inum &&
2469             prev->file_pos + prev->num_bytes == backref->file_pos)
2470                 merge = true;
2471
2472         /* step 1: get root */
2473         key.objectid = backref->root_id;
2474         key.type = BTRFS_ROOT_ITEM_KEY;
2475         key.offset = (u64)-1;
2476
2477         fs_info = BTRFS_I(src_inode)->root->fs_info;
2478         index = srcu_read_lock(&fs_info->subvol_srcu);
2479
2480         root = btrfs_read_fs_root_no_name(fs_info, &key);
2481         if (IS_ERR(root)) {
2482                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2483                 if (PTR_ERR(root) == -ENOENT)
2484                         return 0;
2485                 return PTR_ERR(root);
2486         }
2487
2488         if (btrfs_root_readonly(root)) {
2489                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2490                 return 0;
2491         }
2492
2493         /* step 2: get inode */
2494         key.objectid = backref->inum;
2495         key.type = BTRFS_INODE_ITEM_KEY;
2496         key.offset = 0;
2497
2498         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2499         if (IS_ERR(inode)) {
2500                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2501                 return 0;
2502         }
2503
2504         srcu_read_unlock(&fs_info->subvol_srcu, index);
2505
2506         /* step 3: relink backref */
2507         lock_start = backref->file_pos;
2508         lock_end = backref->file_pos + backref->num_bytes - 1;
2509         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2510                          &cached);
2511
2512         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2513         if (ordered) {
2514                 btrfs_put_ordered_extent(ordered);
2515                 goto out_unlock;
2516         }
2517
2518         trans = btrfs_join_transaction(root);
2519         if (IS_ERR(trans)) {
2520                 ret = PTR_ERR(trans);
2521                 goto out_unlock;
2522         }
2523
2524         key.objectid = backref->inum;
2525         key.type = BTRFS_EXTENT_DATA_KEY;
2526         key.offset = backref->file_pos;
2527
2528         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2529         if (ret < 0) {
2530                 goto out_free_path;
2531         } else if (ret > 0) {
2532                 ret = 0;
2533                 goto out_free_path;
2534         }
2535
2536         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2537                                 struct btrfs_file_extent_item);
2538
2539         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2540             backref->generation)
2541                 goto out_free_path;
2542
2543         btrfs_release_path(path);
2544
2545         start = backref->file_pos;
2546         if (backref->extent_offset < old->extent_offset + old->offset)
2547                 start += old->extent_offset + old->offset -
2548                          backref->extent_offset;
2549
2550         len = min(backref->extent_offset + backref->num_bytes,
2551                   old->extent_offset + old->offset + old->len);
2552         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2553
2554         ret = btrfs_drop_extents(trans, root, inode, start,
2555                                  start + len, 1);
2556         if (ret)
2557                 goto out_free_path;
2558 again:
2559         key.objectid = btrfs_ino(inode);
2560         key.type = BTRFS_EXTENT_DATA_KEY;
2561         key.offset = start;
2562
2563         path->leave_spinning = 1;
2564         if (merge) {
2565                 struct btrfs_file_extent_item *fi;
2566                 u64 extent_len;
2567                 struct btrfs_key found_key;
2568
2569                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2570                 if (ret < 0)
2571                         goto out_free_path;
2572
2573                 path->slots[0]--;
2574                 leaf = path->nodes[0];
2575                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2576
2577                 fi = btrfs_item_ptr(leaf, path->slots[0],
2578                                     struct btrfs_file_extent_item);
2579                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2580
2581                 if (extent_len + found_key.offset == start &&
2582                     relink_is_mergable(leaf, fi, new)) {
2583                         btrfs_set_file_extent_num_bytes(leaf, fi,
2584                                                         extent_len + len);
2585                         btrfs_mark_buffer_dirty(leaf);
2586                         inode_add_bytes(inode, len);
2587
2588                         ret = 1;
2589                         goto out_free_path;
2590                 } else {
2591                         merge = false;
2592                         btrfs_release_path(path);
2593                         goto again;
2594                 }
2595         }
2596
2597         ret = btrfs_insert_empty_item(trans, root, path, &key,
2598                                         sizeof(*extent));
2599         if (ret) {
2600                 btrfs_abort_transaction(trans, root, ret);
2601                 goto out_free_path;
2602         }
2603
2604         leaf = path->nodes[0];
2605         item = btrfs_item_ptr(leaf, path->slots[0],
2606                                 struct btrfs_file_extent_item);
2607         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2608         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2609         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2610         btrfs_set_file_extent_num_bytes(leaf, item, len);
2611         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2612         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2613         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2614         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2615         btrfs_set_file_extent_encryption(leaf, item, 0);
2616         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2617
2618         btrfs_mark_buffer_dirty(leaf);
2619         inode_add_bytes(inode, len);
2620         btrfs_release_path(path);
2621
2622         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2623                         new->disk_len, 0,
2624                         backref->root_id, backref->inum,
2625                         new->file_pos); /* start - extent_offset */
2626         if (ret) {
2627                 btrfs_abort_transaction(trans, root, ret);
2628                 goto out_free_path;
2629         }
2630
2631         ret = 1;
2632 out_free_path:
2633         btrfs_release_path(path);
2634         path->leave_spinning = 0;
2635         btrfs_end_transaction(trans, root);
2636 out_unlock:
2637         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2638                              &cached, GFP_NOFS);
2639         iput(inode);
2640         return ret;
2641 }
2642
2643 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2644 {
2645         struct old_sa_defrag_extent *old, *tmp;
2646
2647         if (!new)
2648                 return;
2649
2650         list_for_each_entry_safe(old, tmp, &new->head, list) {
2651                 kfree(old);
2652         }
2653         kfree(new);
2654 }
2655
2656 static void relink_file_extents(struct new_sa_defrag_extent *new)
2657 {
2658         struct btrfs_path *path;
2659         struct sa_defrag_extent_backref *backref;
2660         struct sa_defrag_extent_backref *prev = NULL;
2661         struct inode *inode;
2662         struct btrfs_root *root;
2663         struct rb_node *node;
2664         int ret;
2665
2666         inode = new->inode;
2667         root = BTRFS_I(inode)->root;
2668
2669         path = btrfs_alloc_path();
2670         if (!path)
2671                 return;
2672
2673         if (!record_extent_backrefs(path, new)) {
2674                 btrfs_free_path(path);
2675                 goto out;
2676         }
2677         btrfs_release_path(path);
2678
2679         while (1) {
2680                 node = rb_first(&new->root);
2681                 if (!node)
2682                         break;
2683                 rb_erase(node, &new->root);
2684
2685                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2686
2687                 ret = relink_extent_backref(path, prev, backref);
2688                 WARN_ON(ret < 0);
2689
2690                 kfree(prev);
2691
2692                 if (ret == 1)
2693                         prev = backref;
2694                 else
2695                         prev = NULL;
2696                 cond_resched();
2697         }
2698         kfree(prev);
2699
2700         btrfs_free_path(path);
2701 out:
2702         free_sa_defrag_extent(new);
2703
2704         atomic_dec(&root->fs_info->defrag_running);
2705         wake_up(&root->fs_info->transaction_wait);
2706 }
2707
2708 static struct new_sa_defrag_extent *
2709 record_old_file_extents(struct inode *inode,
2710                         struct btrfs_ordered_extent *ordered)
2711 {
2712         struct btrfs_root *root = BTRFS_I(inode)->root;
2713         struct btrfs_path *path;
2714         struct btrfs_key key;
2715         struct old_sa_defrag_extent *old;
2716         struct new_sa_defrag_extent *new;
2717         int ret;
2718
2719         new = kmalloc(sizeof(*new), GFP_NOFS);
2720         if (!new)
2721                 return NULL;
2722
2723         new->inode = inode;
2724         new->file_pos = ordered->file_offset;
2725         new->len = ordered->len;
2726         new->bytenr = ordered->start;
2727         new->disk_len = ordered->disk_len;
2728         new->compress_type = ordered->compress_type;
2729         new->root = RB_ROOT;
2730         INIT_LIST_HEAD(&new->head);
2731
2732         path = btrfs_alloc_path();
2733         if (!path)
2734                 goto out_kfree;
2735
2736         key.objectid = btrfs_ino(inode);
2737         key.type = BTRFS_EXTENT_DATA_KEY;
2738         key.offset = new->file_pos;
2739
2740         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2741         if (ret < 0)
2742                 goto out_free_path;
2743         if (ret > 0 && path->slots[0] > 0)
2744                 path->slots[0]--;
2745
2746         /* find out all the old extents for the file range */
2747         while (1) {
2748                 struct btrfs_file_extent_item *extent;
2749                 struct extent_buffer *l;
2750                 int slot;
2751                 u64 num_bytes;
2752                 u64 offset;
2753                 u64 end;
2754                 u64 disk_bytenr;
2755                 u64 extent_offset;
2756
2757                 l = path->nodes[0];
2758                 slot = path->slots[0];
2759
2760                 if (slot >= btrfs_header_nritems(l)) {
2761                         ret = btrfs_next_leaf(root, path);
2762                         if (ret < 0)
2763                                 goto out_free_path;
2764                         else if (ret > 0)
2765                                 break;
2766                         continue;
2767                 }
2768
2769                 btrfs_item_key_to_cpu(l, &key, slot);
2770
2771                 if (key.objectid != btrfs_ino(inode))
2772                         break;
2773                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2774                         break;
2775                 if (key.offset >= new->file_pos + new->len)
2776                         break;
2777
2778                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2779
2780                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2781                 if (key.offset + num_bytes < new->file_pos)
2782                         goto next;
2783
2784                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2785                 if (!disk_bytenr)
2786                         goto next;
2787
2788                 extent_offset = btrfs_file_extent_offset(l, extent);
2789
2790                 old = kmalloc(sizeof(*old), GFP_NOFS);
2791                 if (!old)
2792                         goto out_free_path;
2793
2794                 offset = max(new->file_pos, key.offset);
2795                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2796
2797                 old->bytenr = disk_bytenr;
2798                 old->extent_offset = extent_offset;
2799                 old->offset = offset - key.offset;
2800                 old->len = end - offset;
2801                 old->new = new;
2802                 old->count = 0;
2803                 list_add_tail(&old->list, &new->head);
2804 next:
2805                 path->slots[0]++;
2806                 cond_resched();
2807         }
2808
2809         btrfs_free_path(path);
2810         atomic_inc(&root->fs_info->defrag_running);
2811
2812         return new;
2813
2814 out_free_path:
2815         btrfs_free_path(path);
2816 out_kfree:
2817         free_sa_defrag_extent(new);
2818         return NULL;
2819 }
2820
2821 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2822                                          u64 start, u64 len)
2823 {
2824         struct btrfs_block_group_cache *cache;
2825
2826         cache = btrfs_lookup_block_group(root->fs_info, start);
2827         ASSERT(cache);
2828
2829         spin_lock(&cache->lock);
2830         cache->delalloc_bytes -= len;
2831         spin_unlock(&cache->lock);
2832
2833         btrfs_put_block_group(cache);
2834 }
2835
2836 /* as ordered data IO finishes, this gets called so we can finish
2837  * an ordered extent if the range of bytes in the file it covers are
2838  * fully written.
2839  */
2840 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2841 {
2842         struct inode *inode = ordered_extent->inode;
2843         struct btrfs_root *root = BTRFS_I(inode)->root;
2844         struct btrfs_trans_handle *trans = NULL;
2845         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2846         struct extent_state *cached_state = NULL;
2847         struct new_sa_defrag_extent *new = NULL;
2848         int compress_type = 0;
2849         int ret = 0;
2850         u64 logical_len = ordered_extent->len;
2851         bool nolock;
2852         bool truncated = false;
2853
2854         nolock = btrfs_is_free_space_inode(inode);
2855
2856         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2857                 ret = -EIO;
2858                 goto out;
2859         }
2860
2861         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2862                                      ordered_extent->file_offset +
2863                                      ordered_extent->len - 1);
2864
2865         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2866                 truncated = true;
2867                 logical_len = ordered_extent->truncated_len;
2868                 /* Truncated the entire extent, don't bother adding */
2869                 if (!logical_len)
2870                         goto out;
2871         }
2872
2873         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2874                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2875
2876                 /*
2877                  * For mwrite(mmap + memset to write) case, we still reserve
2878                  * space for NOCOW range.
2879                  * As NOCOW won't cause a new delayed ref, just free the space
2880                  */
2881                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2882                                        ordered_extent->len);
2883                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2884                 if (nolock)
2885                         trans = btrfs_join_transaction_nolock(root);
2886                 else
2887                         trans = btrfs_join_transaction(root);
2888                 if (IS_ERR(trans)) {
2889                         ret = PTR_ERR(trans);
2890                         trans = NULL;
2891                         goto out;
2892                 }
2893                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2894                 ret = btrfs_update_inode_fallback(trans, root, inode);
2895                 if (ret) /* -ENOMEM or corruption */
2896                         btrfs_abort_transaction(trans, root, ret);
2897                 goto out;
2898         }
2899
2900         lock_extent_bits(io_tree, ordered_extent->file_offset,
2901                          ordered_extent->file_offset + ordered_extent->len - 1,
2902                          &cached_state);
2903
2904         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2905                         ordered_extent->file_offset + ordered_extent->len - 1,
2906                         EXTENT_DEFRAG, 1, cached_state);
2907         if (ret) {
2908                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2909                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2910                         /* the inode is shared */
2911                         new = record_old_file_extents(inode, ordered_extent);
2912
2913                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2914                         ordered_extent->file_offset + ordered_extent->len - 1,
2915                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2916         }
2917
2918         if (nolock)
2919                 trans = btrfs_join_transaction_nolock(root);
2920         else
2921                 trans = btrfs_join_transaction(root);
2922         if (IS_ERR(trans)) {
2923                 ret = PTR_ERR(trans);
2924                 trans = NULL;
2925                 goto out_unlock;
2926         }
2927
2928         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2929
2930         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2931                 compress_type = ordered_extent->compress_type;
2932         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2933                 BUG_ON(compress_type);
2934                 ret = btrfs_mark_extent_written(trans, inode,
2935                                                 ordered_extent->file_offset,
2936                                                 ordered_extent->file_offset +
2937                                                 logical_len);
2938         } else {
2939                 BUG_ON(root == root->fs_info->tree_root);
2940                 ret = insert_reserved_file_extent(trans, inode,
2941                                                 ordered_extent->file_offset,
2942                                                 ordered_extent->start,
2943                                                 ordered_extent->disk_len,
2944                                                 logical_len, logical_len,
2945                                                 compress_type, 0, 0,
2946                                                 BTRFS_FILE_EXTENT_REG);
2947                 if (!ret)
2948                         btrfs_release_delalloc_bytes(root,
2949                                                      ordered_extent->start,
2950                                                      ordered_extent->disk_len);
2951         }
2952         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2953                            ordered_extent->file_offset, ordered_extent->len,
2954                            trans->transid);
2955         if (ret < 0) {
2956                 btrfs_abort_transaction(trans, root, ret);
2957                 goto out_unlock;
2958         }
2959
2960         add_pending_csums(trans, inode, ordered_extent->file_offset,
2961                           &ordered_extent->list);
2962
2963         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2964         ret = btrfs_update_inode_fallback(trans, root, inode);
2965         if (ret) { /* -ENOMEM or corruption */
2966                 btrfs_abort_transaction(trans, root, ret);
2967                 goto out_unlock;
2968         }
2969         ret = 0;
2970 out_unlock:
2971         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2972                              ordered_extent->file_offset +
2973                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2974 out:
2975         if (root != root->fs_info->tree_root)
2976                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2977         if (trans)
2978                 btrfs_end_transaction(trans, root);
2979
2980         if (ret || truncated) {
2981                 u64 start, end;
2982
2983                 if (truncated)
2984                         start = ordered_extent->file_offset + logical_len;
2985                 else
2986                         start = ordered_extent->file_offset;
2987                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2988                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2989
2990                 /* Drop the cache for the part of the extent we didn't write. */
2991                 btrfs_drop_extent_cache(inode, start, end, 0);
2992
2993                 /*
2994                  * If the ordered extent had an IOERR or something else went
2995                  * wrong we need to return the space for this ordered extent
2996                  * back to the allocator.  We only free the extent in the
2997                  * truncated case if we didn't write out the extent at all.
2998                  */
2999                 if ((ret || !logical_len) &&
3000                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3001                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3002                         btrfs_free_reserved_extent(root, ordered_extent->start,
3003                                                    ordered_extent->disk_len, 1);
3004         }
3005
3006
3007         /*
3008          * This needs to be done to make sure anybody waiting knows we are done
3009          * updating everything for this ordered extent.
3010          */
3011         btrfs_remove_ordered_extent(inode, ordered_extent);
3012
3013         /* for snapshot-aware defrag */
3014         if (new) {
3015                 if (ret) {
3016                         free_sa_defrag_extent(new);
3017                         atomic_dec(&root->fs_info->defrag_running);
3018                 } else {
3019                         relink_file_extents(new);
3020                 }
3021         }
3022
3023         /* once for us */
3024         btrfs_put_ordered_extent(ordered_extent);
3025         /* once for the tree */
3026         btrfs_put_ordered_extent(ordered_extent);
3027
3028         return ret;
3029 }
3030
3031 static void finish_ordered_fn(struct btrfs_work *work)
3032 {
3033         struct btrfs_ordered_extent *ordered_extent;
3034         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3035         btrfs_finish_ordered_io(ordered_extent);
3036 }
3037
3038 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3039                                 struct extent_state *state, int uptodate)
3040 {
3041         struct inode *inode = page->mapping->host;
3042         struct btrfs_root *root = BTRFS_I(inode)->root;
3043         struct btrfs_ordered_extent *ordered_extent = NULL;
3044         struct btrfs_workqueue *wq;
3045         btrfs_work_func_t func;
3046
3047         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3048
3049         ClearPagePrivate2(page);
3050         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3051                                             end - start + 1, uptodate))
3052                 return 0;
3053
3054         if (btrfs_is_free_space_inode(inode)) {
3055                 wq = root->fs_info->endio_freespace_worker;
3056                 func = btrfs_freespace_write_helper;
3057         } else {
3058                 wq = root->fs_info->endio_write_workers;
3059                 func = btrfs_endio_write_helper;
3060         }
3061
3062         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3063                         NULL);
3064         btrfs_queue_work(wq, &ordered_extent->work);
3065
3066         return 0;
3067 }
3068
3069 static int __readpage_endio_check(struct inode *inode,
3070                                   struct btrfs_io_bio *io_bio,
3071                                   int icsum, struct page *page,
3072                                   int pgoff, u64 start, size_t len)
3073 {
3074         char *kaddr;
3075         u32 csum_expected;
3076         u32 csum = ~(u32)0;
3077
3078         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3079
3080         kaddr = kmap_atomic(page);
3081         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3082         btrfs_csum_final(csum, (char *)&csum);
3083         if (csum != csum_expected)
3084                 goto zeroit;
3085
3086         kunmap_atomic(kaddr);
3087         return 0;
3088 zeroit:
3089         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3090                 "csum failed ino %llu off %llu csum %u expected csum %u",
3091                            btrfs_ino(inode), start, csum, csum_expected);
3092         memset(kaddr + pgoff, 1, len);
3093         flush_dcache_page(page);
3094         kunmap_atomic(kaddr);
3095         if (csum_expected == 0)
3096                 return 0;
3097         return -EIO;
3098 }
3099
3100 /*
3101  * when reads are done, we need to check csums to verify the data is correct
3102  * if there's a match, we allow the bio to finish.  If not, the code in
3103  * extent_io.c will try to find good copies for us.
3104  */
3105 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3106                                       u64 phy_offset, struct page *page,
3107                                       u64 start, u64 end, int mirror)
3108 {
3109         size_t offset = start - page_offset(page);
3110         struct inode *inode = page->mapping->host;
3111         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3112         struct btrfs_root *root = BTRFS_I(inode)->root;
3113
3114         if (PageChecked(page)) {
3115                 ClearPageChecked(page);
3116                 return 0;
3117         }
3118
3119         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3120                 return 0;
3121
3122         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3123             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3124                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3125                 return 0;
3126         }
3127
3128         phy_offset >>= inode->i_sb->s_blocksize_bits;
3129         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3130                                       start, (size_t)(end - start + 1));
3131 }
3132
3133 void btrfs_add_delayed_iput(struct inode *inode)
3134 {
3135         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3136         struct btrfs_inode *binode = BTRFS_I(inode);
3137
3138         if (atomic_add_unless(&inode->i_count, -1, 1))
3139                 return;
3140
3141         spin_lock(&fs_info->delayed_iput_lock);
3142         if (binode->delayed_iput_count == 0) {
3143                 ASSERT(list_empty(&binode->delayed_iput));
3144                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3145         } else {
3146                 binode->delayed_iput_count++;
3147         }
3148         spin_unlock(&fs_info->delayed_iput_lock);
3149 }
3150
3151 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3152 {
3153         struct btrfs_fs_info *fs_info = root->fs_info;
3154
3155         spin_lock(&fs_info->delayed_iput_lock);
3156         while (!list_empty(&fs_info->delayed_iputs)) {
3157                 struct btrfs_inode *inode;
3158
3159                 inode = list_first_entry(&fs_info->delayed_iputs,
3160                                 struct btrfs_inode, delayed_iput);
3161                 if (inode->delayed_iput_count) {
3162                         inode->delayed_iput_count--;
3163                         list_move_tail(&inode->delayed_iput,
3164                                         &fs_info->delayed_iputs);
3165                 } else {
3166                         list_del_init(&inode->delayed_iput);
3167                 }
3168                 spin_unlock(&fs_info->delayed_iput_lock);
3169                 iput(&inode->vfs_inode);
3170                 spin_lock(&fs_info->delayed_iput_lock);
3171         }
3172         spin_unlock(&fs_info->delayed_iput_lock);
3173 }
3174
3175 /*
3176  * This is called in transaction commit time. If there are no orphan
3177  * files in the subvolume, it removes orphan item and frees block_rsv
3178  * structure.
3179  */
3180 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3181                               struct btrfs_root *root)
3182 {
3183         struct btrfs_block_rsv *block_rsv;
3184         int ret;
3185
3186         if (atomic_read(&root->orphan_inodes) ||
3187             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3188                 return;
3189
3190         spin_lock(&root->orphan_lock);
3191         if (atomic_read(&root->orphan_inodes)) {
3192                 spin_unlock(&root->orphan_lock);
3193                 return;
3194         }
3195
3196         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3197                 spin_unlock(&root->orphan_lock);
3198                 return;
3199         }
3200
3201         block_rsv = root->orphan_block_rsv;
3202         root->orphan_block_rsv = NULL;
3203         spin_unlock(&root->orphan_lock);
3204
3205         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3206             btrfs_root_refs(&root->root_item) > 0) {
3207                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3208                                             root->root_key.objectid);
3209                 if (ret)
3210                         btrfs_abort_transaction(trans, root, ret);
3211                 else
3212                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3213                                   &root->state);
3214         }
3215
3216         if (block_rsv) {
3217                 WARN_ON(block_rsv->size > 0);
3218                 btrfs_free_block_rsv(root, block_rsv);
3219         }
3220 }
3221
3222 /*
3223  * This creates an orphan entry for the given inode in case something goes
3224  * wrong in the middle of an unlink/truncate.
3225  *
3226  * NOTE: caller of this function should reserve 5 units of metadata for
3227  *       this function.
3228  */
3229 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3230 {
3231         struct btrfs_root *root = BTRFS_I(inode)->root;
3232         struct btrfs_block_rsv *block_rsv = NULL;
3233         int reserve = 0;
3234         int insert = 0;
3235         int ret;
3236
3237         if (!root->orphan_block_rsv) {
3238                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3239                 if (!block_rsv)
3240                         return -ENOMEM;
3241         }
3242
3243         spin_lock(&root->orphan_lock);
3244         if (!root->orphan_block_rsv) {
3245                 root->orphan_block_rsv = block_rsv;
3246         } else if (block_rsv) {
3247                 btrfs_free_block_rsv(root, block_rsv);
3248                 block_rsv = NULL;
3249         }
3250
3251         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3252                               &BTRFS_I(inode)->runtime_flags)) {
3253 #if 0
3254                 /*
3255                  * For proper ENOSPC handling, we should do orphan
3256                  * cleanup when mounting. But this introduces backward
3257                  * compatibility issue.
3258                  */
3259                 if (!xchg(&root->orphan_item_inserted, 1))
3260                         insert = 2;
3261                 else
3262                         insert = 1;
3263 #endif
3264                 insert = 1;
3265                 atomic_inc(&root->orphan_inodes);
3266         }
3267
3268         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3269                               &BTRFS_I(inode)->runtime_flags))
3270                 reserve = 1;
3271         spin_unlock(&root->orphan_lock);
3272
3273         /* grab metadata reservation from transaction handle */
3274         if (reserve) {
3275                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3276                 ASSERT(!ret);
3277                 if (ret) {
3278                         atomic_dec(&root->orphan_inodes);
3279                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3280                                   &BTRFS_I(inode)->runtime_flags);
3281                         if (insert)
3282                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3283                                           &BTRFS_I(inode)->runtime_flags);
3284                         return ret;
3285                 }
3286         }
3287
3288         /* insert an orphan item to track this unlinked/truncated file */
3289         if (insert >= 1) {
3290                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3291                 if (ret) {
3292                         atomic_dec(&root->orphan_inodes);
3293                         if (reserve) {
3294                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3295                                           &BTRFS_I(inode)->runtime_flags);
3296                                 btrfs_orphan_release_metadata(inode);
3297                         }
3298                         if (ret != -EEXIST) {
3299                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3300                                           &BTRFS_I(inode)->runtime_flags);
3301                                 btrfs_abort_transaction(trans, root, ret);
3302                                 return ret;
3303                         }
3304                 }
3305                 ret = 0;
3306         }
3307
3308         /* insert an orphan item to track subvolume contains orphan files */
3309         if (insert >= 2) {
3310                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3311                                                root->root_key.objectid);
3312                 if (ret && ret != -EEXIST) {
3313                         btrfs_abort_transaction(trans, root, ret);
3314                         return ret;
3315                 }
3316         }
3317         return 0;
3318 }
3319
3320 /*
3321  * We have done the truncate/delete so we can go ahead and remove the orphan
3322  * item for this particular inode.
3323  */
3324 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3325                             struct inode *inode)
3326 {
3327         struct btrfs_root *root = BTRFS_I(inode)->root;
3328         int delete_item = 0;
3329         int release_rsv = 0;
3330         int ret = 0;
3331
3332         spin_lock(&root->orphan_lock);
3333         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3334                                &BTRFS_I(inode)->runtime_flags))
3335                 delete_item = 1;
3336
3337         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3338                                &BTRFS_I(inode)->runtime_flags))
3339                 release_rsv = 1;
3340         spin_unlock(&root->orphan_lock);
3341
3342         if (delete_item) {
3343                 atomic_dec(&root->orphan_inodes);
3344                 if (trans)
3345                         ret = btrfs_del_orphan_item(trans, root,
3346                                                     btrfs_ino(inode));
3347         }
3348
3349         if (release_rsv)
3350                 btrfs_orphan_release_metadata(inode);
3351
3352         return ret;
3353 }
3354
3355 /*
3356  * this cleans up any orphans that may be left on the list from the last use
3357  * of this root.
3358  */
3359 int btrfs_orphan_cleanup(struct btrfs_root *root)
3360 {
3361         struct btrfs_path *path;
3362         struct extent_buffer *leaf;
3363         struct btrfs_key key, found_key;
3364         struct btrfs_trans_handle *trans;
3365         struct inode *inode;
3366         u64 last_objectid = 0;
3367         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3368
3369         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3370                 return 0;
3371
3372         path = btrfs_alloc_path();
3373         if (!path) {
3374                 ret = -ENOMEM;
3375                 goto out;
3376         }
3377         path->reada = READA_BACK;
3378
3379         key.objectid = BTRFS_ORPHAN_OBJECTID;
3380         key.type = BTRFS_ORPHAN_ITEM_KEY;
3381         key.offset = (u64)-1;
3382
3383         while (1) {
3384                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3385                 if (ret < 0)
3386                         goto out;
3387
3388                 /*
3389                  * if ret == 0 means we found what we were searching for, which
3390                  * is weird, but possible, so only screw with path if we didn't
3391                  * find the key and see if we have stuff that matches
3392                  */
3393                 if (ret > 0) {
3394                         ret = 0;
3395                         if (path->slots[0] == 0)
3396                                 break;
3397                         path->slots[0]--;
3398                 }
3399
3400                 /* pull out the item */
3401                 leaf = path->nodes[0];
3402                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3403
3404                 /* make sure the item matches what we want */
3405                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3406                         break;
3407                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3408                         break;
3409
3410                 /* release the path since we're done with it */
3411                 btrfs_release_path(path);
3412
3413                 /*
3414                  * this is where we are basically btrfs_lookup, without the
3415                  * crossing root thing.  we store the inode number in the
3416                  * offset of the orphan item.
3417                  */
3418
3419                 if (found_key.offset == last_objectid) {
3420                         btrfs_err(root->fs_info,
3421                                 "Error removing orphan entry, stopping orphan cleanup");
3422                         ret = -EINVAL;
3423                         goto out;
3424                 }
3425
3426                 last_objectid = found_key.offset;
3427
3428                 found_key.objectid = found_key.offset;
3429                 found_key.type = BTRFS_INODE_ITEM_KEY;
3430                 found_key.offset = 0;
3431                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3432                 ret = PTR_ERR_OR_ZERO(inode);
3433                 if (ret && ret != -ESTALE)
3434                         goto out;
3435
3436                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3437                         struct btrfs_root *dead_root;
3438                         struct btrfs_fs_info *fs_info = root->fs_info;
3439                         int is_dead_root = 0;
3440
3441                         /*
3442                          * this is an orphan in the tree root. Currently these
3443                          * could come from 2 sources:
3444                          *  a) a snapshot deletion in progress
3445                          *  b) a free space cache inode
3446                          * We need to distinguish those two, as the snapshot
3447                          * orphan must not get deleted.
3448                          * find_dead_roots already ran before us, so if this
3449                          * is a snapshot deletion, we should find the root
3450                          * in the dead_roots list
3451                          */
3452                         spin_lock(&fs_info->trans_lock);
3453                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3454                                             root_list) {
3455                                 if (dead_root->root_key.objectid ==
3456                                     found_key.objectid) {
3457                                         is_dead_root = 1;
3458                                         break;
3459                                 }
3460                         }
3461                         spin_unlock(&fs_info->trans_lock);
3462                         if (is_dead_root) {
3463                                 /* prevent this orphan from being found again */
3464                                 key.offset = found_key.objectid - 1;
3465                                 continue;
3466                         }
3467                 }
3468                 /*
3469                  * Inode is already gone but the orphan item is still there,
3470                  * kill the orphan item.
3471                  */
3472                 if (ret == -ESTALE) {
3473                         trans = btrfs_start_transaction(root, 1);
3474                         if (IS_ERR(trans)) {
3475                                 ret = PTR_ERR(trans);
3476                                 goto out;
3477                         }
3478                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3479                                 found_key.objectid);
3480                         ret = btrfs_del_orphan_item(trans, root,
3481                                                     found_key.objectid);
3482                         btrfs_end_transaction(trans, root);
3483                         if (ret)
3484                                 goto out;
3485                         continue;
3486                 }
3487
3488                 /*
3489                  * add this inode to the orphan list so btrfs_orphan_del does
3490                  * the proper thing when we hit it
3491                  */
3492                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3493                         &BTRFS_I(inode)->runtime_flags);
3494                 atomic_inc(&root->orphan_inodes);
3495
3496                 /* if we have links, this was a truncate, lets do that */
3497                 if (inode->i_nlink) {
3498                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3499                                 iput(inode);
3500                                 continue;
3501                         }
3502                         nr_truncate++;
3503
3504                         /* 1 for the orphan item deletion. */
3505                         trans = btrfs_start_transaction(root, 1);
3506                         if (IS_ERR(trans)) {
3507                                 iput(inode);
3508                                 ret = PTR_ERR(trans);
3509                                 goto out;
3510                         }
3511                         ret = btrfs_orphan_add(trans, inode);
3512                         btrfs_end_transaction(trans, root);
3513                         if (ret) {
3514                                 iput(inode);
3515                                 goto out;
3516                         }
3517
3518                         ret = btrfs_truncate(inode);
3519                         if (ret)
3520                                 btrfs_orphan_del(NULL, inode);
3521                 } else {
3522                         nr_unlink++;
3523                 }
3524
3525                 /* this will do delete_inode and everything for us */
3526                 iput(inode);
3527                 if (ret)
3528                         goto out;
3529         }
3530         /* release the path since we're done with it */
3531         btrfs_release_path(path);
3532
3533         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3534
3535         if (root->orphan_block_rsv)
3536                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3537                                         (u64)-1);
3538
3539         if (root->orphan_block_rsv ||
3540             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3541                 trans = btrfs_join_transaction(root);
3542                 if (!IS_ERR(trans))
3543                         btrfs_end_transaction(trans, root);
3544         }
3545
3546         if (nr_unlink)
3547                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3548         if (nr_truncate)
3549                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3550
3551 out:
3552         if (ret)
3553                 btrfs_err(root->fs_info,
3554                         "could not do orphan cleanup %d", ret);
3555         btrfs_free_path(path);
3556         return ret;
3557 }
3558
3559 /*
3560  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3561  * don't find any xattrs, we know there can't be any acls.
3562  *
3563  * slot is the slot the inode is in, objectid is the objectid of the inode
3564  */
3565 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3566                                           int slot, u64 objectid,
3567                                           int *first_xattr_slot)
3568 {
3569         u32 nritems = btrfs_header_nritems(leaf);
3570         struct btrfs_key found_key;
3571         static u64 xattr_access = 0;
3572         static u64 xattr_default = 0;
3573         int scanned = 0;
3574
3575         if (!xattr_access) {
3576                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3577                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3578                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3579                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3580         }
3581
3582         slot++;
3583         *first_xattr_slot = -1;
3584         while (slot < nritems) {
3585                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3586
3587                 /* we found a different objectid, there must not be acls */
3588                 if (found_key.objectid != objectid)
3589                         return 0;
3590
3591                 /* we found an xattr, assume we've got an acl */
3592                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3593                         if (*first_xattr_slot == -1)
3594                                 *first_xattr_slot = slot;
3595                         if (found_key.offset == xattr_access ||
3596                             found_key.offset == xattr_default)
3597                                 return 1;
3598                 }
3599
3600                 /*
3601                  * we found a key greater than an xattr key, there can't
3602                  * be any acls later on
3603                  */
3604                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3605                         return 0;
3606
3607                 slot++;
3608                 scanned++;
3609
3610                 /*
3611                  * it goes inode, inode backrefs, xattrs, extents,
3612                  * so if there are a ton of hard links to an inode there can
3613                  * be a lot of backrefs.  Don't waste time searching too hard,
3614                  * this is just an optimization
3615                  */
3616                 if (scanned >= 8)
3617                         break;
3618         }
3619         /* we hit the end of the leaf before we found an xattr or
3620          * something larger than an xattr.  We have to assume the inode
3621          * has acls
3622          */
3623         if (*first_xattr_slot == -1)
3624                 *first_xattr_slot = slot;
3625         return 1;
3626 }
3627
3628 /*
3629  * read an inode from the btree into the in-memory inode
3630  */
3631 static void btrfs_read_locked_inode(struct inode *inode)
3632 {
3633         struct btrfs_path *path;
3634         struct extent_buffer *leaf;
3635         struct btrfs_inode_item *inode_item;
3636         struct btrfs_root *root = BTRFS_I(inode)->root;
3637         struct btrfs_key location;
3638         unsigned long ptr;
3639         int maybe_acls;
3640         u32 rdev;
3641         int ret;
3642         bool filled = false;
3643         int first_xattr_slot;
3644
3645         ret = btrfs_fill_inode(inode, &rdev);
3646         if (!ret)
3647                 filled = true;
3648
3649         path = btrfs_alloc_path();
3650         if (!path)
3651                 goto make_bad;
3652
3653         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3654
3655         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3656         if (ret)
3657                 goto make_bad;
3658
3659         leaf = path->nodes[0];
3660
3661         if (filled)
3662                 goto cache_index;
3663
3664         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3665                                     struct btrfs_inode_item);
3666         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3667         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3668         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3669         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3670         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3671
3672         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3673         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3674
3675         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3676         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3677
3678         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3679         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3680
3681         BTRFS_I(inode)->i_otime.tv_sec =
3682                 btrfs_timespec_sec(leaf, &inode_item->otime);
3683         BTRFS_I(inode)->i_otime.tv_nsec =
3684                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3685
3686         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3687         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3688         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3689
3690         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3691         inode->i_generation = BTRFS_I(inode)->generation;
3692         inode->i_rdev = 0;
3693         rdev = btrfs_inode_rdev(leaf, inode_item);
3694
3695         BTRFS_I(inode)->index_cnt = (u64)-1;
3696         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3697
3698 cache_index:
3699         /*
3700          * If we were modified in the current generation and evicted from memory
3701          * and then re-read we need to do a full sync since we don't have any
3702          * idea about which extents were modified before we were evicted from
3703          * cache.
3704          *
3705          * This is required for both inode re-read from disk and delayed inode
3706          * in delayed_nodes_tree.
3707          */
3708         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3709                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3710                         &BTRFS_I(inode)->runtime_flags);
3711
3712         /*
3713          * We don't persist the id of the transaction where an unlink operation
3714          * against the inode was last made. So here we assume the inode might
3715          * have been evicted, and therefore the exact value of last_unlink_trans
3716          * lost, and set it to last_trans to avoid metadata inconsistencies
3717          * between the inode and its parent if the inode is fsync'ed and the log
3718          * replayed. For example, in the scenario:
3719          *
3720          * touch mydir/foo
3721          * ln mydir/foo mydir/bar
3722          * sync
3723          * unlink mydir/bar
3724          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3725          * xfs_io -c fsync mydir/foo
3726          * <power failure>
3727          * mount fs, triggers fsync log replay
3728          *
3729          * We must make sure that when we fsync our inode foo we also log its
3730          * parent inode, otherwise after log replay the parent still has the
3731          * dentry with the "bar" name but our inode foo has a link count of 1
3732          * and doesn't have an inode ref with the name "bar" anymore.
3733          *
3734          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3735          * but it guarantees correctness at the expense of occasional full
3736          * transaction commits on fsync if our inode is a directory, or if our
3737          * inode is not a directory, logging its parent unnecessarily.
3738          */
3739         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3740
3741         path->slots[0]++;
3742         if (inode->i_nlink != 1 ||
3743             path->slots[0] >= btrfs_header_nritems(leaf))
3744                 goto cache_acl;
3745
3746         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3747         if (location.objectid != btrfs_ino(inode))
3748                 goto cache_acl;
3749
3750         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3751         if (location.type == BTRFS_INODE_REF_KEY) {
3752                 struct btrfs_inode_ref *ref;
3753
3754                 ref = (struct btrfs_inode_ref *)ptr;
3755                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3756         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3757                 struct btrfs_inode_extref *extref;
3758
3759                 extref = (struct btrfs_inode_extref *)ptr;
3760                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3761                                                                      extref);
3762         }
3763 cache_acl:
3764         /*
3765          * try to precache a NULL acl entry for files that don't have
3766          * any xattrs or acls
3767          */
3768         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3769                                            btrfs_ino(inode), &first_xattr_slot);
3770         if (first_xattr_slot != -1) {
3771                 path->slots[0] = first_xattr_slot;
3772                 ret = btrfs_load_inode_props(inode, path);
3773                 if (ret)
3774                         btrfs_err(root->fs_info,
3775                                   "error loading props for ino %llu (root %llu): %d",
3776                                   btrfs_ino(inode),
3777                                   root->root_key.objectid, ret);
3778         }
3779         btrfs_free_path(path);
3780
3781         if (!maybe_acls)
3782                 cache_no_acl(inode);
3783
3784         switch (inode->i_mode & S_IFMT) {
3785         case S_IFREG:
3786                 inode->i_mapping->a_ops = &btrfs_aops;
3787                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3788                 inode->i_fop = &btrfs_file_operations;
3789                 inode->i_op = &btrfs_file_inode_operations;
3790                 break;
3791         case S_IFDIR:
3792                 inode->i_fop = &btrfs_dir_file_operations;
3793                 if (root == root->fs_info->tree_root)
3794                         inode->i_op = &btrfs_dir_ro_inode_operations;
3795                 else
3796                         inode->i_op = &btrfs_dir_inode_operations;
3797                 break;
3798         case S_IFLNK:
3799                 inode->i_op = &btrfs_symlink_inode_operations;
3800                 inode_nohighmem(inode);
3801                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3802                 break;
3803         default:
3804                 inode->i_op = &btrfs_special_inode_operations;
3805                 init_special_inode(inode, inode->i_mode, rdev);
3806                 break;
3807         }
3808
3809         btrfs_update_iflags(inode);
3810         return;
3811
3812 make_bad:
3813         btrfs_free_path(path);
3814         make_bad_inode(inode);
3815 }
3816
3817 /*
3818  * given a leaf and an inode, copy the inode fields into the leaf
3819  */
3820 static void fill_inode_item(struct btrfs_trans_handle *trans,
3821                             struct extent_buffer *leaf,
3822                             struct btrfs_inode_item *item,
3823                             struct inode *inode)
3824 {
3825         struct btrfs_map_token token;
3826
3827         btrfs_init_map_token(&token);
3828
3829         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3830         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3831         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3832                                    &token);
3833         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3834         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3835
3836         btrfs_set_token_timespec_sec(leaf, &item->atime,
3837                                      inode->i_atime.tv_sec, &token);
3838         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3839                                       inode->i_atime.tv_nsec, &token);
3840
3841         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3842                                      inode->i_mtime.tv_sec, &token);
3843         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3844                                       inode->i_mtime.tv_nsec, &token);
3845
3846         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3847                                      inode->i_ctime.tv_sec, &token);
3848         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3849                                       inode->i_ctime.tv_nsec, &token);
3850
3851         btrfs_set_token_timespec_sec(leaf, &item->otime,
3852                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3853         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3854                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3855
3856         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3857                                      &token);
3858         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3859                                          &token);
3860         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3861         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3862         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3863         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3864         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3865 }
3866
3867 /*
3868  * copy everything in the in-memory inode into the btree.
3869  */
3870 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3871                                 struct btrfs_root *root, struct inode *inode)
3872 {
3873         struct btrfs_inode_item *inode_item;
3874         struct btrfs_path *path;
3875         struct extent_buffer *leaf;
3876         int ret;
3877
3878         path = btrfs_alloc_path();
3879         if (!path)
3880                 return -ENOMEM;
3881
3882         path->leave_spinning = 1;
3883         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3884                                  1);
3885         if (ret) {
3886                 if (ret > 0)
3887                         ret = -ENOENT;
3888                 goto failed;
3889         }
3890
3891         leaf = path->nodes[0];
3892         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3893                                     struct btrfs_inode_item);
3894
3895         fill_inode_item(trans, leaf, inode_item, inode);
3896         btrfs_mark_buffer_dirty(leaf);
3897         btrfs_set_inode_last_trans(trans, inode);
3898         ret = 0;
3899 failed:
3900         btrfs_free_path(path);
3901         return ret;
3902 }
3903
3904 /*
3905  * copy everything in the in-memory inode into the btree.
3906  */
3907 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3908                                 struct btrfs_root *root, struct inode *inode)
3909 {
3910         int ret;
3911
3912         /*
3913          * If the inode is a free space inode, we can deadlock during commit
3914          * if we put it into the delayed code.
3915          *
3916          * The data relocation inode should also be directly updated
3917          * without delay
3918          */
3919         if (!btrfs_is_free_space_inode(inode)
3920             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3921             && !root->fs_info->log_root_recovering) {
3922                 btrfs_update_root_times(trans, root);
3923
3924                 ret = btrfs_delayed_update_inode(trans, root, inode);
3925                 if (!ret)
3926                         btrfs_set_inode_last_trans(trans, inode);
3927                 return ret;
3928         }
3929
3930         return btrfs_update_inode_item(trans, root, inode);
3931 }
3932
3933 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3934                                          struct btrfs_root *root,
3935                                          struct inode *inode)
3936 {
3937         int ret;
3938
3939         ret = btrfs_update_inode(trans, root, inode);
3940         if (ret == -ENOSPC)
3941                 return btrfs_update_inode_item(trans, root, inode);
3942         return ret;
3943 }
3944
3945 /*
3946  * unlink helper that gets used here in inode.c and in the tree logging
3947  * recovery code.  It remove a link in a directory with a given name, and
3948  * also drops the back refs in the inode to the directory
3949  */
3950 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3951                                 struct btrfs_root *root,
3952                                 struct inode *dir, struct inode *inode,
3953                                 const char *name, int name_len)
3954 {
3955         struct btrfs_path *path;
3956         int ret = 0;
3957         struct extent_buffer *leaf;
3958         struct btrfs_dir_item *di;
3959         struct btrfs_key key;
3960         u64 index;
3961         u64 ino = btrfs_ino(inode);
3962         u64 dir_ino = btrfs_ino(dir);
3963
3964         path = btrfs_alloc_path();
3965         if (!path) {
3966                 ret = -ENOMEM;
3967                 goto out;
3968         }
3969
3970         path->leave_spinning = 1;
3971         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3972                                     name, name_len, -1);
3973         if (IS_ERR(di)) {
3974                 ret = PTR_ERR(di);
3975                 goto err;
3976         }
3977         if (!di) {
3978                 ret = -ENOENT;
3979                 goto err;
3980         }
3981         leaf = path->nodes[0];
3982         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3983         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3984         if (ret)
3985                 goto err;
3986         btrfs_release_path(path);
3987
3988         /*
3989          * If we don't have dir index, we have to get it by looking up
3990          * the inode ref, since we get the inode ref, remove it directly,
3991          * it is unnecessary to do delayed deletion.
3992          *
3993          * But if we have dir index, needn't search inode ref to get it.
3994          * Since the inode ref is close to the inode item, it is better
3995          * that we delay to delete it, and just do this deletion when
3996          * we update the inode item.
3997          */
3998         if (BTRFS_I(inode)->dir_index) {
3999                 ret = btrfs_delayed_delete_inode_ref(inode);
4000                 if (!ret) {
4001                         index = BTRFS_I(inode)->dir_index;
4002                         goto skip_backref;
4003                 }
4004         }
4005
4006         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4007                                   dir_ino, &index);
4008         if (ret) {
4009                 btrfs_info(root->fs_info,
4010                         "failed to delete reference to %.*s, inode %llu parent %llu",
4011                         name_len, name, ino, dir_ino);
4012                 btrfs_abort_transaction(trans, root, ret);
4013                 goto err;
4014         }
4015 skip_backref:
4016         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4017         if (ret) {
4018                 btrfs_abort_transaction(trans, root, ret);
4019                 goto err;
4020         }
4021
4022         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
4023                                          inode, dir_ino);
4024         if (ret != 0 && ret != -ENOENT) {
4025                 btrfs_abort_transaction(trans, root, ret);
4026                 goto err;
4027         }
4028
4029         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4030                                            dir, index);
4031         if (ret == -ENOENT)
4032                 ret = 0;
4033         else if (ret)
4034                 btrfs_abort_transaction(trans, root, ret);
4035 err:
4036         btrfs_free_path(path);
4037         if (ret)
4038                 goto out;
4039
4040         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4041         inode_inc_iversion(inode);
4042         inode_inc_iversion(dir);
4043         inode->i_ctime = dir->i_mtime =
4044                 dir->i_ctime = current_fs_time(inode->i_sb);
4045         ret = btrfs_update_inode(trans, root, dir);
4046 out:
4047         return ret;
4048 }
4049
4050 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4051                        struct btrfs_root *root,
4052                        struct inode *dir, struct inode *inode,
4053                        const char *name, int name_len)
4054 {
4055         int ret;
4056         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4057         if (!ret) {
4058                 drop_nlink(inode);
4059                 ret = btrfs_update_inode(trans, root, inode);
4060         }
4061         return ret;
4062 }
4063
4064 /*
4065  * helper to start transaction for unlink and rmdir.
4066  *
4067  * unlink and rmdir are special in btrfs, they do not always free space, so
4068  * if we cannot make our reservations the normal way try and see if there is
4069  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4070  * allow the unlink to occur.
4071  */
4072 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4073 {
4074         struct btrfs_root *root = BTRFS_I(dir)->root;
4075
4076         /*
4077          * 1 for the possible orphan item
4078          * 1 for the dir item
4079          * 1 for the dir index
4080          * 1 for the inode ref
4081          * 1 for the inode
4082          */
4083         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4084 }
4085
4086 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4087 {
4088         struct btrfs_root *root = BTRFS_I(dir)->root;
4089         struct btrfs_trans_handle *trans;
4090         struct inode *inode = d_inode(dentry);
4091         int ret;
4092
4093         trans = __unlink_start_trans(dir);
4094         if (IS_ERR(trans))
4095                 return PTR_ERR(trans);
4096
4097         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4098
4099         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4100                                  dentry->d_name.name, dentry->d_name.len);
4101         if (ret)
4102                 goto out;
4103
4104         if (inode->i_nlink == 0) {
4105                 ret = btrfs_orphan_add(trans, inode);
4106                 if (ret)
4107                         goto out;
4108         }
4109
4110 out:
4111         btrfs_end_transaction(trans, root);
4112         btrfs_btree_balance_dirty(root);
4113         return ret;
4114 }
4115
4116 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4117                         struct btrfs_root *root,
4118                         struct inode *dir, u64 objectid,
4119                         const char *name, int name_len)
4120 {
4121         struct btrfs_path *path;
4122         struct extent_buffer *leaf;
4123         struct btrfs_dir_item *di;
4124         struct btrfs_key key;
4125         u64 index;
4126         int ret;
4127         u64 dir_ino = btrfs_ino(dir);
4128
4129         path = btrfs_alloc_path();
4130         if (!path)
4131                 return -ENOMEM;
4132
4133         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4134                                    name, name_len, -1);
4135         if (IS_ERR_OR_NULL(di)) {
4136                 if (!di)
4137                         ret = -ENOENT;
4138                 else
4139                         ret = PTR_ERR(di);
4140                 goto out;
4141         }
4142
4143         leaf = path->nodes[0];
4144         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4145         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4146         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4147         if (ret) {
4148                 btrfs_abort_transaction(trans, root, ret);
4149                 goto out;
4150         }
4151         btrfs_release_path(path);
4152
4153         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4154                                  objectid, root->root_key.objectid,
4155                                  dir_ino, &index, name, name_len);
4156         if (ret < 0) {
4157                 if (ret != -ENOENT) {
4158                         btrfs_abort_transaction(trans, root, ret);
4159                         goto out;
4160                 }
4161                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4162                                                  name, name_len);
4163                 if (IS_ERR_OR_NULL(di)) {
4164                         if (!di)
4165                                 ret = -ENOENT;
4166                         else
4167                                 ret = PTR_ERR(di);
4168                         btrfs_abort_transaction(trans, root, ret);
4169                         goto out;
4170                 }
4171
4172                 leaf = path->nodes[0];
4173                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4174                 btrfs_release_path(path);
4175                 index = key.offset;
4176         }
4177         btrfs_release_path(path);
4178
4179         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4180         if (ret) {
4181                 btrfs_abort_transaction(trans, root, ret);
4182                 goto out;
4183         }
4184
4185         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4186         inode_inc_iversion(dir);
4187         dir->i_mtime = dir->i_ctime = current_fs_time(dir->i_sb);
4188         ret = btrfs_update_inode_fallback(trans, root, dir);
4189         if (ret)
4190                 btrfs_abort_transaction(trans, root, ret);
4191 out:
4192         btrfs_free_path(path);
4193         return ret;
4194 }
4195
4196 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4197 {
4198         struct inode *inode = d_inode(dentry);
4199         int err = 0;
4200         struct btrfs_root *root = BTRFS_I(dir)->root;
4201         struct btrfs_trans_handle *trans;
4202
4203         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4204                 return -ENOTEMPTY;
4205         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4206                 return -EPERM;
4207
4208         trans = __unlink_start_trans(dir);
4209         if (IS_ERR(trans))
4210                 return PTR_ERR(trans);
4211
4212         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4213                 err = btrfs_unlink_subvol(trans, root, dir,
4214                                           BTRFS_I(inode)->location.objectid,
4215                                           dentry->d_name.name,
4216                                           dentry->d_name.len);
4217                 goto out;
4218         }
4219
4220         err = btrfs_orphan_add(trans, inode);
4221         if (err)
4222                 goto out;
4223
4224         /* now the directory is empty */
4225         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4226                                  dentry->d_name.name, dentry->d_name.len);
4227         if (!err)
4228                 btrfs_i_size_write(inode, 0);
4229 out:
4230         btrfs_end_transaction(trans, root);
4231         btrfs_btree_balance_dirty(root);
4232
4233         return err;
4234 }
4235
4236 static int truncate_space_check(struct btrfs_trans_handle *trans,
4237                                 struct btrfs_root *root,
4238                                 u64 bytes_deleted)
4239 {
4240         int ret;
4241
4242         /*
4243          * This is only used to apply pressure to the enospc system, we don't
4244          * intend to use this reservation at all.
4245          */
4246         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4247         bytes_deleted *= root->nodesize;
4248         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4249                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4250         if (!ret) {
4251                 trace_btrfs_space_reservation(root->fs_info, "transaction",
4252                                               trans->transid,
4253                                               bytes_deleted, 1);
4254                 trans->bytes_reserved += bytes_deleted;
4255         }
4256         return ret;
4257
4258 }
4259
4260 static int truncate_inline_extent(struct inode *inode,
4261                                   struct btrfs_path *path,
4262                                   struct btrfs_key *found_key,
4263                                   const u64 item_end,
4264                                   const u64 new_size)
4265 {
4266         struct extent_buffer *leaf = path->nodes[0];
4267         int slot = path->slots[0];
4268         struct btrfs_file_extent_item *fi;
4269         u32 size = (u32)(new_size - found_key->offset);
4270         struct btrfs_root *root = BTRFS_I(inode)->root;
4271
4272         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4273
4274         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4275                 loff_t offset = new_size;
4276                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4277
4278                 /*
4279                  * Zero out the remaining of the last page of our inline extent,
4280                  * instead of directly truncating our inline extent here - that
4281                  * would be much more complex (decompressing all the data, then
4282                  * compressing the truncated data, which might be bigger than
4283                  * the size of the inline extent, resize the extent, etc).
4284                  * We release the path because to get the page we might need to
4285                  * read the extent item from disk (data not in the page cache).
4286                  */
4287                 btrfs_release_path(path);
4288                 return btrfs_truncate_block(inode, offset, page_end - offset,
4289                                         0);
4290         }
4291
4292         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4293         size = btrfs_file_extent_calc_inline_size(size);
4294         btrfs_truncate_item(root, path, size, 1);
4295
4296         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4297                 inode_sub_bytes(inode, item_end + 1 - new_size);
4298
4299         return 0;
4300 }
4301
4302 /*
4303  * this can truncate away extent items, csum items and directory items.
4304  * It starts at a high offset and removes keys until it can't find
4305  * any higher than new_size
4306  *
4307  * csum items that cross the new i_size are truncated to the new size
4308  * as well.
4309  *
4310  * min_type is the minimum key type to truncate down to.  If set to 0, this
4311  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4312  */
4313 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4314                                struct btrfs_root *root,
4315                                struct inode *inode,
4316                                u64 new_size, u32 min_type)
4317 {
4318         struct btrfs_path *path;
4319         struct extent_buffer *leaf;
4320         struct btrfs_file_extent_item *fi;
4321         struct btrfs_key key;
4322         struct btrfs_key found_key;
4323         u64 extent_start = 0;
4324         u64 extent_num_bytes = 0;
4325         u64 extent_offset = 0;
4326         u64 item_end = 0;
4327         u64 last_size = new_size;
4328         u32 found_type = (u8)-1;
4329         int found_extent;
4330         int del_item;
4331         int pending_del_nr = 0;
4332         int pending_del_slot = 0;
4333         int extent_type = -1;
4334         int ret;
4335         int err = 0;
4336         u64 ino = btrfs_ino(inode);
4337         u64 bytes_deleted = 0;
4338         bool be_nice = 0;
4339         bool should_throttle = 0;
4340         bool should_end = 0;
4341
4342         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4343
4344         /*
4345          * for non-free space inodes and ref cows, we want to back off from
4346          * time to time
4347          */
4348         if (!btrfs_is_free_space_inode(inode) &&
4349             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4350                 be_nice = 1;
4351
4352         path = btrfs_alloc_path();
4353         if (!path)
4354                 return -ENOMEM;
4355         path->reada = READA_BACK;
4356
4357         /*
4358          * We want to drop from the next block forward in case this new size is
4359          * not block aligned since we will be keeping the last block of the
4360          * extent just the way it is.
4361          */
4362         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4363             root == root->fs_info->tree_root)
4364                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4365                                         root->sectorsize), (u64)-1, 0);
4366
4367         /*
4368          * This function is also used to drop the items in the log tree before
4369          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4370          * it is used to drop the loged items. So we shouldn't kill the delayed
4371          * items.
4372          */
4373         if (min_type == 0 && root == BTRFS_I(inode)->root)
4374                 btrfs_kill_delayed_inode_items(inode);
4375
4376         key.objectid = ino;
4377         key.offset = (u64)-1;
4378         key.type = (u8)-1;
4379
4380 search_again:
4381         /*
4382          * with a 16K leaf size and 128MB extents, you can actually queue
4383          * up a huge file in a single leaf.  Most of the time that
4384          * bytes_deleted is > 0, it will be huge by the time we get here
4385          */
4386         if (be_nice && bytes_deleted > SZ_32M) {
4387                 if (btrfs_should_end_transaction(trans, root)) {
4388                         err = -EAGAIN;
4389                         goto error;
4390                 }
4391         }
4392
4393
4394         path->leave_spinning = 1;
4395         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4396         if (ret < 0) {
4397                 err = ret;
4398                 goto out;
4399         }
4400
4401         if (ret > 0) {
4402                 /* there are no items in the tree for us to truncate, we're
4403                  * done
4404                  */
4405                 if (path->slots[0] == 0)
4406                         goto out;
4407                 path->slots[0]--;
4408         }
4409
4410         while (1) {
4411                 fi = NULL;
4412                 leaf = path->nodes[0];
4413                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4414                 found_type = found_key.type;
4415
4416                 if (found_key.objectid != ino)
4417                         break;
4418
4419                 if (found_type < min_type)
4420                         break;
4421
4422                 item_end = found_key.offset;
4423                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4424                         fi = btrfs_item_ptr(leaf, path->slots[0],
4425                                             struct btrfs_file_extent_item);
4426                         extent_type = btrfs_file_extent_type(leaf, fi);
4427                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4428                                 item_end +=
4429                                     btrfs_file_extent_num_bytes(leaf, fi);
4430                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4431                                 item_end += btrfs_file_extent_inline_len(leaf,
4432                                                          path->slots[0], fi);
4433                         }
4434                         item_end--;
4435                 }
4436                 if (found_type > min_type) {
4437                         del_item = 1;
4438                 } else {
4439                         if (item_end < new_size)
4440                                 break;
4441                         if (found_key.offset >= new_size)
4442                                 del_item = 1;
4443                         else
4444                                 del_item = 0;
4445                 }
4446                 found_extent = 0;
4447                 /* FIXME, shrink the extent if the ref count is only 1 */
4448                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4449                         goto delete;
4450
4451                 if (del_item)
4452                         last_size = found_key.offset;
4453                 else
4454                         last_size = new_size;
4455
4456                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4457                         u64 num_dec;
4458                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4459                         if (!del_item) {
4460                                 u64 orig_num_bytes =
4461                                         btrfs_file_extent_num_bytes(leaf, fi);
4462                                 extent_num_bytes = ALIGN(new_size -
4463                                                 found_key.offset,
4464                                                 root->sectorsize);
4465                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4466                                                          extent_num_bytes);
4467                                 num_dec = (orig_num_bytes -
4468                                            extent_num_bytes);
4469                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4470                                              &root->state) &&
4471                                     extent_start != 0)
4472                                         inode_sub_bytes(inode, num_dec);
4473                                 btrfs_mark_buffer_dirty(leaf);
4474                         } else {
4475                                 extent_num_bytes =
4476                                         btrfs_file_extent_disk_num_bytes(leaf,
4477                                                                          fi);
4478                                 extent_offset = found_key.offset -
4479                                         btrfs_file_extent_offset(leaf, fi);
4480
4481                                 /* FIXME blocksize != 4096 */
4482                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4483                                 if (extent_start != 0) {
4484                                         found_extent = 1;
4485                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4486                                                      &root->state))
4487                                                 inode_sub_bytes(inode, num_dec);
4488                                 }
4489                         }
4490                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4491                         /*
4492                          * we can't truncate inline items that have had
4493                          * special encodings
4494                          */
4495                         if (!del_item &&
4496                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4497                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4498
4499                                 /*
4500                                  * Need to release path in order to truncate a
4501                                  * compressed extent. So delete any accumulated
4502                                  * extent items so far.
4503                                  */
4504                                 if (btrfs_file_extent_compression(leaf, fi) !=
4505                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4506                                         err = btrfs_del_items(trans, root, path,
4507                                                               pending_del_slot,
4508                                                               pending_del_nr);
4509                                         if (err) {
4510                                                 btrfs_abort_transaction(trans,
4511                                                                         root,
4512                                                                         err);
4513                                                 goto error;
4514                                         }
4515                                         pending_del_nr = 0;
4516                                 }
4517
4518                                 err = truncate_inline_extent(inode, path,
4519                                                              &found_key,
4520                                                              item_end,
4521                                                              new_size);
4522                                 if (err) {
4523                                         btrfs_abort_transaction(trans,
4524                                                                 root, err);
4525                                         goto error;
4526                                 }
4527                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4528                                             &root->state)) {
4529                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4530                         }
4531                 }
4532 delete:
4533                 if (del_item) {
4534                         if (!pending_del_nr) {
4535                                 /* no pending yet, add ourselves */
4536                                 pending_del_slot = path->slots[0];
4537                                 pending_del_nr = 1;
4538                         } else if (pending_del_nr &&
4539                                    path->slots[0] + 1 == pending_del_slot) {
4540                                 /* hop on the pending chunk */
4541                                 pending_del_nr++;
4542                                 pending_del_slot = path->slots[0];
4543                         } else {
4544                                 BUG();
4545                         }
4546                 } else {
4547                         break;
4548                 }
4549                 should_throttle = 0;
4550
4551                 if (found_extent &&
4552                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4553                      root == root->fs_info->tree_root)) {
4554                         btrfs_set_path_blocking(path);
4555                         bytes_deleted += extent_num_bytes;
4556                         ret = btrfs_free_extent(trans, root, extent_start,
4557                                                 extent_num_bytes, 0,
4558                                                 btrfs_header_owner(leaf),
4559                                                 ino, extent_offset);
4560                         BUG_ON(ret);
4561                         if (btrfs_should_throttle_delayed_refs(trans, root))
4562                                 btrfs_async_run_delayed_refs(root,
4563                                                              trans->transid,
4564                                         trans->delayed_ref_updates * 2, 0);
4565                         if (be_nice) {
4566                                 if (truncate_space_check(trans, root,
4567                                                          extent_num_bytes)) {
4568                                         should_end = 1;
4569                                 }
4570                                 if (btrfs_should_throttle_delayed_refs(trans,
4571                                                                        root)) {
4572                                         should_throttle = 1;
4573                                 }
4574                         }
4575                 }
4576
4577                 if (found_type == BTRFS_INODE_ITEM_KEY)
4578                         break;
4579
4580                 if (path->slots[0] == 0 ||
4581                     path->slots[0] != pending_del_slot ||
4582                     should_throttle || should_end) {
4583                         if (pending_del_nr) {
4584                                 ret = btrfs_del_items(trans, root, path,
4585                                                 pending_del_slot,
4586                                                 pending_del_nr);
4587                                 if (ret) {
4588                                         btrfs_abort_transaction(trans,
4589                                                                 root, ret);
4590                                         goto error;
4591                                 }
4592                                 pending_del_nr = 0;
4593                         }
4594                         btrfs_release_path(path);
4595                         if (should_throttle) {
4596                                 unsigned long updates = trans->delayed_ref_updates;
4597                                 if (updates) {
4598                                         trans->delayed_ref_updates = 0;
4599                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4600                                         if (ret && !err)
4601                                                 err = ret;
4602                                 }
4603                         }
4604                         /*
4605                          * if we failed to refill our space rsv, bail out
4606                          * and let the transaction restart
4607                          */
4608                         if (should_end) {
4609                                 err = -EAGAIN;
4610                                 goto error;
4611                         }
4612                         goto search_again;
4613                 } else {
4614                         path->slots[0]--;
4615                 }
4616         }
4617 out:
4618         if (pending_del_nr) {
4619                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4620                                       pending_del_nr);
4621                 if (ret)
4622                         btrfs_abort_transaction(trans, root, ret);
4623         }
4624 error:
4625         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4626                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4627
4628         btrfs_free_path(path);
4629
4630         if (be_nice && bytes_deleted > SZ_32M) {
4631                 unsigned long updates = trans->delayed_ref_updates;
4632                 if (updates) {
4633                         trans->delayed_ref_updates = 0;
4634                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4635                         if (ret && !err)
4636                                 err = ret;
4637                 }
4638         }
4639         return err;
4640 }
4641
4642 /*
4643  * btrfs_truncate_block - read, zero a chunk and write a block
4644  * @inode - inode that we're zeroing
4645  * @from - the offset to start zeroing
4646  * @len - the length to zero, 0 to zero the entire range respective to the
4647  *      offset
4648  * @front - zero up to the offset instead of from the offset on
4649  *
4650  * This will find the block for the "from" offset and cow the block and zero the
4651  * part we want to zero.  This is used with truncate and hole punching.
4652  */
4653 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4654                         int front)
4655 {
4656         struct address_space *mapping = inode->i_mapping;
4657         struct btrfs_root *root = BTRFS_I(inode)->root;
4658         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4659         struct btrfs_ordered_extent *ordered;
4660         struct extent_state *cached_state = NULL;
4661         char *kaddr;
4662         u32 blocksize = root->sectorsize;
4663         pgoff_t index = from >> PAGE_SHIFT;
4664         unsigned offset = from & (blocksize - 1);
4665         struct page *page;
4666         gfp_t mask = btrfs_alloc_write_mask(mapping);
4667         int ret = 0;
4668         u64 block_start;
4669         u64 block_end;
4670
4671         if ((offset & (blocksize - 1)) == 0 &&
4672             (!len || ((len & (blocksize - 1)) == 0)))
4673                 goto out;
4674
4675         ret = btrfs_delalloc_reserve_space(inode,
4676                         round_down(from, blocksize), blocksize);
4677         if (ret)
4678                 goto out;
4679
4680 again:
4681         page = find_or_create_page(mapping, index, mask);
4682         if (!page) {
4683                 btrfs_delalloc_release_space(inode,
4684                                 round_down(from, blocksize),
4685                                 blocksize);
4686                 ret = -ENOMEM;
4687                 goto out;
4688         }
4689
4690         block_start = round_down(from, blocksize);
4691         block_end = block_start + blocksize - 1;
4692
4693         if (!PageUptodate(page)) {
4694                 ret = btrfs_readpage(NULL, page);
4695                 lock_page(page);
4696                 if (page->mapping != mapping) {
4697                         unlock_page(page);
4698                         put_page(page);
4699                         goto again;
4700                 }
4701                 if (!PageUptodate(page)) {
4702                         ret = -EIO;
4703                         goto out_unlock;
4704                 }
4705         }
4706         wait_on_page_writeback(page);
4707
4708         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4709         set_page_extent_mapped(page);
4710
4711         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4712         if (ordered) {
4713                 unlock_extent_cached(io_tree, block_start, block_end,
4714                                      &cached_state, GFP_NOFS);
4715                 unlock_page(page);
4716                 put_page(page);
4717                 btrfs_start_ordered_extent(inode, ordered, 1);
4718                 btrfs_put_ordered_extent(ordered);
4719                 goto again;
4720         }
4721
4722         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4723                           EXTENT_DIRTY | EXTENT_DELALLOC |
4724                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4725                           0, 0, &cached_state, GFP_NOFS);
4726
4727         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4728                                         &cached_state);
4729         if (ret) {
4730                 unlock_extent_cached(io_tree, block_start, block_end,
4731                                      &cached_state, GFP_NOFS);
4732                 goto out_unlock;
4733         }
4734
4735         if (offset != blocksize) {
4736                 if (!len)
4737                         len = blocksize - offset;
4738                 kaddr = kmap(page);
4739                 if (front)
4740                         memset(kaddr + (block_start - page_offset(page)),
4741                                 0, offset);
4742                 else
4743                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4744                                 0, len);
4745                 flush_dcache_page(page);
4746                 kunmap(page);
4747         }
4748         ClearPageChecked(page);
4749         set_page_dirty(page);
4750         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4751                              GFP_NOFS);
4752
4753 out_unlock:
4754         if (ret)
4755                 btrfs_delalloc_release_space(inode, block_start,
4756                                              blocksize);
4757         unlock_page(page);
4758         put_page(page);
4759 out:
4760         return ret;
4761 }
4762
4763 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4764                              u64 offset, u64 len)
4765 {
4766         struct btrfs_trans_handle *trans;
4767         int ret;
4768
4769         /*
4770          * Still need to make sure the inode looks like it's been updated so
4771          * that any holes get logged if we fsync.
4772          */
4773         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4774                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4775                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4776                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4777                 return 0;
4778         }
4779
4780         /*
4781          * 1 - for the one we're dropping
4782          * 1 - for the one we're adding
4783          * 1 - for updating the inode.
4784          */
4785         trans = btrfs_start_transaction(root, 3);
4786         if (IS_ERR(trans))
4787                 return PTR_ERR(trans);
4788
4789         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4790         if (ret) {
4791                 btrfs_abort_transaction(trans, root, ret);
4792                 btrfs_end_transaction(trans, root);
4793                 return ret;
4794         }
4795
4796         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4797                                        0, 0, len, 0, len, 0, 0, 0);
4798         if (ret)
4799                 btrfs_abort_transaction(trans, root, ret);
4800         else
4801                 btrfs_update_inode(trans, root, inode);
4802         btrfs_end_transaction(trans, root);
4803         return ret;
4804 }
4805
4806 /*
4807  * This function puts in dummy file extents for the area we're creating a hole
4808  * for.  So if we are truncating this file to a larger size we need to insert
4809  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4810  * the range between oldsize and size
4811  */
4812 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4813 {
4814         struct btrfs_root *root = BTRFS_I(inode)->root;
4815         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4816         struct extent_map *em = NULL;
4817         struct extent_state *cached_state = NULL;
4818         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4819         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4820         u64 block_end = ALIGN(size, root->sectorsize);
4821         u64 last_byte;
4822         u64 cur_offset;
4823         u64 hole_size;
4824         int err = 0;
4825
4826         /*
4827          * If our size started in the middle of a block we need to zero out the
4828          * rest of the block before we expand the i_size, otherwise we could
4829          * expose stale data.
4830          */
4831         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4832         if (err)
4833                 return err;
4834
4835         if (size <= hole_start)
4836                 return 0;
4837
4838         while (1) {
4839                 struct btrfs_ordered_extent *ordered;
4840
4841                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4842                                  &cached_state);
4843                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4844                                                      block_end - hole_start);
4845                 if (!ordered)
4846                         break;
4847                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4848                                      &cached_state, GFP_NOFS);
4849                 btrfs_start_ordered_extent(inode, ordered, 1);
4850                 btrfs_put_ordered_extent(ordered);
4851         }
4852
4853         cur_offset = hole_start;
4854         while (1) {
4855                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4856                                 block_end - cur_offset, 0);
4857                 if (IS_ERR(em)) {
4858                         err = PTR_ERR(em);
4859                         em = NULL;
4860                         break;
4861                 }
4862                 last_byte = min(extent_map_end(em), block_end);
4863                 last_byte = ALIGN(last_byte , root->sectorsize);
4864                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4865                         struct extent_map *hole_em;
4866                         hole_size = last_byte - cur_offset;
4867
4868                         err = maybe_insert_hole(root, inode, cur_offset,
4869                                                 hole_size);
4870                         if (err)
4871                                 break;
4872                         btrfs_drop_extent_cache(inode, cur_offset,
4873                                                 cur_offset + hole_size - 1, 0);
4874                         hole_em = alloc_extent_map();
4875                         if (!hole_em) {
4876                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4877                                         &BTRFS_I(inode)->runtime_flags);
4878                                 goto next;
4879                         }
4880                         hole_em->start = cur_offset;
4881                         hole_em->len = hole_size;
4882                         hole_em->orig_start = cur_offset;
4883
4884                         hole_em->block_start = EXTENT_MAP_HOLE;
4885                         hole_em->block_len = 0;
4886                         hole_em->orig_block_len = 0;
4887                         hole_em->ram_bytes = hole_size;
4888                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4889                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4890                         hole_em->generation = root->fs_info->generation;
4891
4892                         while (1) {
4893                                 write_lock(&em_tree->lock);
4894                                 err = add_extent_mapping(em_tree, hole_em, 1);
4895                                 write_unlock(&em_tree->lock);
4896                                 if (err != -EEXIST)
4897                                         break;
4898                                 btrfs_drop_extent_cache(inode, cur_offset,
4899                                                         cur_offset +
4900                                                         hole_size - 1, 0);
4901                         }
4902                         free_extent_map(hole_em);
4903                 }
4904 next:
4905                 free_extent_map(em);
4906                 em = NULL;
4907                 cur_offset = last_byte;
4908                 if (cur_offset >= block_end)
4909                         break;
4910         }
4911         free_extent_map(em);
4912         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4913                              GFP_NOFS);
4914         return err;
4915 }
4916
4917 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4918 {
4919         struct btrfs_root *root = BTRFS_I(inode)->root;
4920         struct btrfs_trans_handle *trans;
4921         loff_t oldsize = i_size_read(inode);
4922         loff_t newsize = attr->ia_size;
4923         int mask = attr->ia_valid;
4924         int ret;
4925
4926         /*
4927          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4928          * special case where we need to update the times despite not having
4929          * these flags set.  For all other operations the VFS set these flags
4930          * explicitly if it wants a timestamp update.
4931          */
4932         if (newsize != oldsize) {
4933                 inode_inc_iversion(inode);
4934                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4935                         inode->i_ctime = inode->i_mtime =
4936                                 current_fs_time(inode->i_sb);
4937         }
4938
4939         if (newsize > oldsize) {
4940                 /*
4941                  * Don't do an expanding truncate while snapshoting is ongoing.
4942                  * This is to ensure the snapshot captures a fully consistent
4943                  * state of this file - if the snapshot captures this expanding
4944                  * truncation, it must capture all writes that happened before
4945                  * this truncation.
4946                  */
4947                 btrfs_wait_for_snapshot_creation(root);
4948                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4949                 if (ret) {
4950                         btrfs_end_write_no_snapshoting(root);
4951                         return ret;
4952                 }
4953
4954                 trans = btrfs_start_transaction(root, 1);
4955                 if (IS_ERR(trans)) {
4956                         btrfs_end_write_no_snapshoting(root);
4957                         return PTR_ERR(trans);
4958                 }
4959
4960                 i_size_write(inode, newsize);
4961                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4962                 pagecache_isize_extended(inode, oldsize, newsize);
4963                 ret = btrfs_update_inode(trans, root, inode);
4964                 btrfs_end_write_no_snapshoting(root);
4965                 btrfs_end_transaction(trans, root);
4966         } else {
4967
4968                 /*
4969                  * We're truncating a file that used to have good data down to
4970                  * zero. Make sure it gets into the ordered flush list so that
4971                  * any new writes get down to disk quickly.
4972                  */
4973                 if (newsize == 0)
4974                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4975                                 &BTRFS_I(inode)->runtime_flags);
4976
4977                 /*
4978                  * 1 for the orphan item we're going to add
4979                  * 1 for the orphan item deletion.
4980                  */
4981                 trans = btrfs_start_transaction(root, 2);
4982                 if (IS_ERR(trans))
4983                         return PTR_ERR(trans);
4984
4985                 /*
4986                  * We need to do this in case we fail at _any_ point during the
4987                  * actual truncate.  Once we do the truncate_setsize we could
4988                  * invalidate pages which forces any outstanding ordered io to
4989                  * be instantly completed which will give us extents that need
4990                  * to be truncated.  If we fail to get an orphan inode down we
4991                  * could have left over extents that were never meant to live,
4992                  * so we need to guarantee from this point on that everything
4993                  * will be consistent.
4994                  */
4995                 ret = btrfs_orphan_add(trans, inode);
4996                 btrfs_end_transaction(trans, root);
4997                 if (ret)
4998                         return ret;
4999
5000                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5001                 truncate_setsize(inode, newsize);
5002
5003                 /* Disable nonlocked read DIO to avoid the end less truncate */
5004                 btrfs_inode_block_unlocked_dio(inode);
5005                 inode_dio_wait(inode);
5006                 btrfs_inode_resume_unlocked_dio(inode);
5007
5008                 ret = btrfs_truncate(inode);
5009                 if (ret && inode->i_nlink) {
5010                         int err;
5011
5012                         /*
5013                          * failed to truncate, disk_i_size is only adjusted down
5014                          * as we remove extents, so it should represent the true
5015                          * size of the inode, so reset the in memory size and
5016                          * delete our orphan entry.
5017                          */
5018                         trans = btrfs_join_transaction(root);
5019                         if (IS_ERR(trans)) {
5020                                 btrfs_orphan_del(NULL, inode);
5021                                 return ret;
5022                         }
5023                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5024                         err = btrfs_orphan_del(trans, inode);
5025                         if (err)
5026                                 btrfs_abort_transaction(trans, root, err);
5027                         btrfs_end_transaction(trans, root);
5028                 }
5029         }
5030
5031         return ret;
5032 }
5033
5034 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5035 {
5036         struct inode *inode = d_inode(dentry);
5037         struct btrfs_root *root = BTRFS_I(inode)->root;
5038         int err;
5039
5040         if (btrfs_root_readonly(root))
5041                 return -EROFS;
5042
5043         err = inode_change_ok(inode, attr);
5044         if (err)
5045                 return err;
5046
5047         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5048                 err = btrfs_setsize(inode, attr);
5049                 if (err)
5050                         return err;
5051         }
5052
5053         if (attr->ia_valid) {
5054                 setattr_copy(inode, attr);
5055                 inode_inc_iversion(inode);
5056                 err = btrfs_dirty_inode(inode);
5057
5058                 if (!err && attr->ia_valid & ATTR_MODE)
5059                         err = posix_acl_chmod(inode, inode->i_mode);
5060         }
5061
5062         return err;
5063 }
5064
5065 /*
5066  * While truncating the inode pages during eviction, we get the VFS calling
5067  * btrfs_invalidatepage() against each page of the inode. This is slow because
5068  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5069  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5070  * extent_state structures over and over, wasting lots of time.
5071  *
5072  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5073  * those expensive operations on a per page basis and do only the ordered io
5074  * finishing, while we release here the extent_map and extent_state structures,
5075  * without the excessive merging and splitting.
5076  */
5077 static void evict_inode_truncate_pages(struct inode *inode)
5078 {
5079         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5080         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5081         struct rb_node *node;
5082
5083         ASSERT(inode->i_state & I_FREEING);
5084         truncate_inode_pages_final(&inode->i_data);
5085
5086         write_lock(&map_tree->lock);
5087         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5088                 struct extent_map *em;
5089
5090                 node = rb_first(&map_tree->map);
5091                 em = rb_entry(node, struct extent_map, rb_node);
5092                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5093                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5094                 remove_extent_mapping(map_tree, em);
5095                 free_extent_map(em);
5096                 if (need_resched()) {
5097                         write_unlock(&map_tree->lock);
5098                         cond_resched();
5099                         write_lock(&map_tree->lock);
5100                 }
5101         }
5102         write_unlock(&map_tree->lock);
5103
5104         /*
5105          * Keep looping until we have no more ranges in the io tree.
5106          * We can have ongoing bios started by readpages (called from readahead)
5107          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5108          * still in progress (unlocked the pages in the bio but did not yet
5109          * unlocked the ranges in the io tree). Therefore this means some
5110          * ranges can still be locked and eviction started because before
5111          * submitting those bios, which are executed by a separate task (work
5112          * queue kthread), inode references (inode->i_count) were not taken
5113          * (which would be dropped in the end io callback of each bio).
5114          * Therefore here we effectively end up waiting for those bios and
5115          * anyone else holding locked ranges without having bumped the inode's
5116          * reference count - if we don't do it, when they access the inode's
5117          * io_tree to unlock a range it may be too late, leading to an
5118          * use-after-free issue.
5119          */
5120         spin_lock(&io_tree->lock);
5121         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5122                 struct extent_state *state;
5123                 struct extent_state *cached_state = NULL;
5124                 u64 start;
5125                 u64 end;
5126
5127                 node = rb_first(&io_tree->state);
5128                 state = rb_entry(node, struct extent_state, rb_node);
5129                 start = state->start;
5130                 end = state->end;
5131                 spin_unlock(&io_tree->lock);
5132
5133                 lock_extent_bits(io_tree, start, end, &cached_state);
5134
5135                 /*
5136                  * If still has DELALLOC flag, the extent didn't reach disk,
5137                  * and its reserved space won't be freed by delayed_ref.
5138                  * So we need to free its reserved space here.
5139                  * (Refer to comment in btrfs_invalidatepage, case 2)
5140                  *
5141                  * Note, end is the bytenr of last byte, so we need + 1 here.
5142                  */
5143                 if (state->state & EXTENT_DELALLOC)
5144                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5145
5146                 clear_extent_bit(io_tree, start, end,
5147                                  EXTENT_LOCKED | EXTENT_DIRTY |
5148                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5149                                  EXTENT_DEFRAG, 1, 1,
5150                                  &cached_state, GFP_NOFS);
5151
5152                 cond_resched();
5153                 spin_lock(&io_tree->lock);
5154         }
5155         spin_unlock(&io_tree->lock);
5156 }
5157
5158 void btrfs_evict_inode(struct inode *inode)
5159 {
5160         struct btrfs_trans_handle *trans;
5161         struct btrfs_root *root = BTRFS_I(inode)->root;
5162         struct btrfs_block_rsv *rsv, *global_rsv;
5163         int steal_from_global = 0;
5164         u64 min_size;
5165         int ret;
5166
5167         trace_btrfs_inode_evict(inode);
5168
5169         if (!root) {
5170                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5171                 return;
5172         }
5173
5174         min_size = btrfs_calc_trunc_metadata_size(root, 1);
5175
5176         evict_inode_truncate_pages(inode);
5177
5178         if (inode->i_nlink &&
5179             ((btrfs_root_refs(&root->root_item) != 0 &&
5180               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5181              btrfs_is_free_space_inode(inode)))
5182                 goto no_delete;
5183
5184         if (is_bad_inode(inode)) {
5185                 btrfs_orphan_del(NULL, inode);
5186                 goto no_delete;
5187         }
5188         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5189         if (!special_file(inode->i_mode))
5190                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5191
5192         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5193
5194         if (root->fs_info->log_root_recovering) {
5195                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5196                                  &BTRFS_I(inode)->runtime_flags));
5197                 goto no_delete;
5198         }
5199
5200         if (inode->i_nlink > 0) {
5201                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5202                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5203                 goto no_delete;
5204         }
5205
5206         ret = btrfs_commit_inode_delayed_inode(inode);
5207         if (ret) {
5208                 btrfs_orphan_del(NULL, inode);
5209                 goto no_delete;
5210         }
5211
5212         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5213         if (!rsv) {
5214                 btrfs_orphan_del(NULL, inode);
5215                 goto no_delete;
5216         }
5217         rsv->size = min_size;
5218         rsv->failfast = 1;
5219         global_rsv = &root->fs_info->global_block_rsv;
5220
5221         btrfs_i_size_write(inode, 0);
5222
5223         /*
5224          * This is a bit simpler than btrfs_truncate since we've already
5225          * reserved our space for our orphan item in the unlink, so we just
5226          * need to reserve some slack space in case we add bytes and update
5227          * inode item when doing the truncate.
5228          */
5229         while (1) {
5230                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5231                                              BTRFS_RESERVE_FLUSH_LIMIT);
5232
5233                 /*
5234                  * Try and steal from the global reserve since we will
5235                  * likely not use this space anyway, we want to try as
5236                  * hard as possible to get this to work.
5237                  */
5238                 if (ret)
5239                         steal_from_global++;
5240                 else
5241                         steal_from_global = 0;
5242                 ret = 0;
5243
5244                 /*
5245                  * steal_from_global == 0: we reserved stuff, hooray!
5246                  * steal_from_global == 1: we didn't reserve stuff, boo!
5247                  * steal_from_global == 2: we've committed, still not a lot of
5248                  * room but maybe we'll have room in the global reserve this
5249                  * time.
5250                  * steal_from_global == 3: abandon all hope!
5251                  */
5252                 if (steal_from_global > 2) {
5253                         btrfs_warn(root->fs_info,
5254                                 "Could not get space for a delete, will truncate on mount %d",
5255                                 ret);
5256                         btrfs_orphan_del(NULL, inode);
5257                         btrfs_free_block_rsv(root, rsv);
5258                         goto no_delete;
5259                 }
5260
5261                 trans = btrfs_join_transaction(root);
5262                 if (IS_ERR(trans)) {
5263                         btrfs_orphan_del(NULL, inode);
5264                         btrfs_free_block_rsv(root, rsv);
5265                         goto no_delete;
5266                 }
5267
5268                 /*
5269                  * We can't just steal from the global reserve, we need to make
5270                  * sure there is room to do it, if not we need to commit and try
5271                  * again.
5272                  */
5273                 if (steal_from_global) {
5274                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5275                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5276                                                               min_size, 0);
5277                         else
5278                                 ret = -ENOSPC;
5279                 }
5280
5281                 /*
5282                  * Couldn't steal from the global reserve, we have too much
5283                  * pending stuff built up, commit the transaction and try it
5284                  * again.
5285                  */
5286                 if (ret) {
5287                         ret = btrfs_commit_transaction(trans, root);
5288                         if (ret) {
5289                                 btrfs_orphan_del(NULL, inode);
5290                                 btrfs_free_block_rsv(root, rsv);
5291                                 goto no_delete;
5292                         }
5293                         continue;
5294                 } else {
5295                         steal_from_global = 0;
5296                 }
5297
5298                 trans->block_rsv = rsv;
5299
5300                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5301                 if (ret != -ENOSPC && ret != -EAGAIN)
5302                         break;
5303
5304                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5305                 btrfs_end_transaction(trans, root);
5306                 trans = NULL;
5307                 btrfs_btree_balance_dirty(root);
5308         }
5309
5310         btrfs_free_block_rsv(root, rsv);
5311
5312         /*
5313          * Errors here aren't a big deal, it just means we leave orphan items
5314          * in the tree.  They will be cleaned up on the next mount.
5315          */
5316         if (ret == 0) {
5317                 trans->block_rsv = root->orphan_block_rsv;
5318                 btrfs_orphan_del(trans, inode);
5319         } else {
5320                 btrfs_orphan_del(NULL, inode);
5321         }
5322
5323         trans->block_rsv = &root->fs_info->trans_block_rsv;
5324         if (!(root == root->fs_info->tree_root ||
5325               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5326                 btrfs_return_ino(root, btrfs_ino(inode));
5327
5328         btrfs_end_transaction(trans, root);
5329         btrfs_btree_balance_dirty(root);
5330 no_delete:
5331         btrfs_remove_delayed_node(inode);
5332         clear_inode(inode);
5333 }
5334
5335 /*
5336  * this returns the key found in the dir entry in the location pointer.
5337  * If no dir entries were found, location->objectid is 0.
5338  */
5339 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5340                                struct btrfs_key *location)
5341 {
5342         const char *name = dentry->d_name.name;
5343         int namelen = dentry->d_name.len;
5344         struct btrfs_dir_item *di;
5345         struct btrfs_path *path;
5346         struct btrfs_root *root = BTRFS_I(dir)->root;
5347         int ret = 0;
5348
5349         path = btrfs_alloc_path();
5350         if (!path)
5351                 return -ENOMEM;
5352
5353         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5354                                     namelen, 0);
5355         if (IS_ERR(di))
5356                 ret = PTR_ERR(di);
5357
5358         if (IS_ERR_OR_NULL(di))
5359                 goto out_err;
5360
5361         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5362 out:
5363         btrfs_free_path(path);
5364         return ret;
5365 out_err:
5366         location->objectid = 0;
5367         goto out;
5368 }
5369
5370 /*
5371  * when we hit a tree root in a directory, the btrfs part of the inode
5372  * needs to be changed to reflect the root directory of the tree root.  This
5373  * is kind of like crossing a mount point.
5374  */
5375 static int fixup_tree_root_location(struct btrfs_root *root,
5376                                     struct inode *dir,
5377                                     struct dentry *dentry,
5378                                     struct btrfs_key *location,
5379                                     struct btrfs_root **sub_root)
5380 {
5381         struct btrfs_path *path;
5382         struct btrfs_root *new_root;
5383         struct btrfs_root_ref *ref;
5384         struct extent_buffer *leaf;
5385         struct btrfs_key key;
5386         int ret;
5387         int err = 0;
5388
5389         path = btrfs_alloc_path();
5390         if (!path) {
5391                 err = -ENOMEM;
5392                 goto out;
5393         }
5394
5395         err = -ENOENT;
5396         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5397         key.type = BTRFS_ROOT_REF_KEY;
5398         key.offset = location->objectid;
5399
5400         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5401                                 0, 0);
5402         if (ret) {
5403                 if (ret < 0)
5404                         err = ret;
5405                 goto out;
5406         }
5407
5408         leaf = path->nodes[0];
5409         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5410         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5411             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5412                 goto out;
5413
5414         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5415                                    (unsigned long)(ref + 1),
5416                                    dentry->d_name.len);
5417         if (ret)
5418                 goto out;
5419
5420         btrfs_release_path(path);
5421
5422         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5423         if (IS_ERR(new_root)) {
5424                 err = PTR_ERR(new_root);
5425                 goto out;
5426         }
5427
5428         *sub_root = new_root;
5429         location->objectid = btrfs_root_dirid(&new_root->root_item);
5430         location->type = BTRFS_INODE_ITEM_KEY;
5431         location->offset = 0;
5432         err = 0;
5433 out:
5434         btrfs_free_path(path);
5435         return err;
5436 }
5437
5438 static void inode_tree_add(struct inode *inode)
5439 {
5440         struct btrfs_root *root = BTRFS_I(inode)->root;
5441         struct btrfs_inode *entry;
5442         struct rb_node **p;
5443         struct rb_node *parent;
5444         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5445         u64 ino = btrfs_ino(inode);
5446
5447         if (inode_unhashed(inode))
5448                 return;
5449         parent = NULL;
5450         spin_lock(&root->inode_lock);
5451         p = &root->inode_tree.rb_node;
5452         while (*p) {
5453                 parent = *p;
5454                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5455
5456                 if (ino < btrfs_ino(&entry->vfs_inode))
5457                         p = &parent->rb_left;
5458                 else if (ino > btrfs_ino(&entry->vfs_inode))
5459                         p = &parent->rb_right;
5460                 else {
5461                         WARN_ON(!(entry->vfs_inode.i_state &
5462                                   (I_WILL_FREE | I_FREEING)));
5463                         rb_replace_node(parent, new, &root->inode_tree);
5464                         RB_CLEAR_NODE(parent);
5465                         spin_unlock(&root->inode_lock);
5466                         return;
5467                 }
5468         }
5469         rb_link_node(new, parent, p);
5470         rb_insert_color(new, &root->inode_tree);
5471         spin_unlock(&root->inode_lock);
5472 }
5473
5474 static void inode_tree_del(struct inode *inode)
5475 {
5476         struct btrfs_root *root = BTRFS_I(inode)->root;
5477         int empty = 0;
5478
5479         spin_lock(&root->inode_lock);
5480         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5481                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5482                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5483                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5484         }
5485         spin_unlock(&root->inode_lock);
5486
5487         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5488                 synchronize_srcu(&root->fs_info->subvol_srcu);
5489                 spin_lock(&root->inode_lock);
5490                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5491                 spin_unlock(&root->inode_lock);
5492                 if (empty)
5493                         btrfs_add_dead_root(root);
5494         }
5495 }
5496
5497 void btrfs_invalidate_inodes(struct btrfs_root *root)
5498 {
5499         struct rb_node *node;
5500         struct rb_node *prev;
5501         struct btrfs_inode *entry;
5502         struct inode *inode;
5503         u64 objectid = 0;
5504
5505         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5506                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5507
5508         spin_lock(&root->inode_lock);
5509 again:
5510         node = root->inode_tree.rb_node;
5511         prev = NULL;
5512         while (node) {
5513                 prev = node;
5514                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5515
5516                 if (objectid < btrfs_ino(&entry->vfs_inode))
5517                         node = node->rb_left;
5518                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5519                         node = node->rb_right;
5520                 else
5521                         break;
5522         }
5523         if (!node) {
5524                 while (prev) {
5525                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5526                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5527                                 node = prev;
5528                                 break;
5529                         }
5530                         prev = rb_next(prev);
5531                 }
5532         }
5533         while (node) {
5534                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5535                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5536                 inode = igrab(&entry->vfs_inode);
5537                 if (inode) {
5538                         spin_unlock(&root->inode_lock);
5539                         if (atomic_read(&inode->i_count) > 1)
5540                                 d_prune_aliases(inode);
5541                         /*
5542                          * btrfs_drop_inode will have it removed from
5543                          * the inode cache when its usage count
5544                          * hits zero.
5545                          */
5546                         iput(inode);
5547                         cond_resched();
5548                         spin_lock(&root->inode_lock);
5549                         goto again;
5550                 }
5551
5552                 if (cond_resched_lock(&root->inode_lock))
5553                         goto again;
5554
5555                 node = rb_next(node);
5556         }
5557         spin_unlock(&root->inode_lock);
5558 }
5559
5560 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5561 {
5562         struct btrfs_iget_args *args = p;
5563         inode->i_ino = args->location->objectid;
5564         memcpy(&BTRFS_I(inode)->location, args->location,
5565                sizeof(*args->location));
5566         BTRFS_I(inode)->root = args->root;
5567         return 0;
5568 }
5569
5570 static int btrfs_find_actor(struct inode *inode, void *opaque)
5571 {
5572         struct btrfs_iget_args *args = opaque;
5573         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5574                 args->root == BTRFS_I(inode)->root;
5575 }
5576
5577 static struct inode *btrfs_iget_locked(struct super_block *s,
5578                                        struct btrfs_key *location,
5579                                        struct btrfs_root *root)
5580 {
5581         struct inode *inode;
5582         struct btrfs_iget_args args;
5583         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5584
5585         args.location = location;
5586         args.root = root;
5587
5588         inode = iget5_locked(s, hashval, btrfs_find_actor,
5589                              btrfs_init_locked_inode,
5590                              (void *)&args);
5591         return inode;
5592 }
5593
5594 /* Get an inode object given its location and corresponding root.
5595  * Returns in *is_new if the inode was read from disk
5596  */
5597 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5598                          struct btrfs_root *root, int *new)
5599 {
5600         struct inode *inode;
5601
5602         inode = btrfs_iget_locked(s, location, root);
5603         if (!inode)
5604                 return ERR_PTR(-ENOMEM);
5605
5606         if (inode->i_state & I_NEW) {
5607                 btrfs_read_locked_inode(inode);
5608                 if (!is_bad_inode(inode)) {
5609                         inode_tree_add(inode);
5610                         unlock_new_inode(inode);
5611                         if (new)
5612                                 *new = 1;
5613                 } else {
5614                         unlock_new_inode(inode);
5615                         iput(inode);
5616                         inode = ERR_PTR(-ESTALE);
5617                 }
5618         }
5619
5620         return inode;
5621 }
5622
5623 static struct inode *new_simple_dir(struct super_block *s,
5624                                     struct btrfs_key *key,
5625                                     struct btrfs_root *root)
5626 {
5627         struct inode *inode = new_inode(s);
5628
5629         if (!inode)
5630                 return ERR_PTR(-ENOMEM);
5631
5632         BTRFS_I(inode)->root = root;
5633         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5634         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5635
5636         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5637         inode->i_op = &btrfs_dir_ro_inode_operations;
5638         inode->i_fop = &simple_dir_operations;
5639         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5640         inode->i_mtime = current_fs_time(inode->i_sb);
5641         inode->i_atime = inode->i_mtime;
5642         inode->i_ctime = inode->i_mtime;
5643         BTRFS_I(inode)->i_otime = inode->i_mtime;
5644
5645         return inode;
5646 }
5647
5648 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5649 {
5650         struct inode *inode;
5651         struct btrfs_root *root = BTRFS_I(dir)->root;
5652         struct btrfs_root *sub_root = root;
5653         struct btrfs_key location;
5654         int index;
5655         int ret = 0;
5656
5657         if (dentry->d_name.len > BTRFS_NAME_LEN)
5658                 return ERR_PTR(-ENAMETOOLONG);
5659
5660         ret = btrfs_inode_by_name(dir, dentry, &location);
5661         if (ret < 0)
5662                 return ERR_PTR(ret);
5663
5664         if (location.objectid == 0)
5665                 return ERR_PTR(-ENOENT);
5666
5667         if (location.type == BTRFS_INODE_ITEM_KEY) {
5668                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5669                 return inode;
5670         }
5671
5672         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5673
5674         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5675         ret = fixup_tree_root_location(root, dir, dentry,
5676                                        &location, &sub_root);
5677         if (ret < 0) {
5678                 if (ret != -ENOENT)
5679                         inode = ERR_PTR(ret);
5680                 else
5681                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5682         } else {
5683                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5684         }
5685         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5686
5687         if (!IS_ERR(inode) && root != sub_root) {
5688                 down_read(&root->fs_info->cleanup_work_sem);
5689                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5690                         ret = btrfs_orphan_cleanup(sub_root);
5691                 up_read(&root->fs_info->cleanup_work_sem);
5692                 if (ret) {
5693                         iput(inode);
5694                         inode = ERR_PTR(ret);
5695                 }
5696         }
5697
5698         return inode;
5699 }
5700
5701 static int btrfs_dentry_delete(const struct dentry *dentry)
5702 {
5703         struct btrfs_root *root;
5704         struct inode *inode = d_inode(dentry);
5705
5706         if (!inode && !IS_ROOT(dentry))
5707                 inode = d_inode(dentry->d_parent);
5708
5709         if (inode) {
5710                 root = BTRFS_I(inode)->root;
5711                 if (btrfs_root_refs(&root->root_item) == 0)
5712                         return 1;
5713
5714                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5715                         return 1;
5716         }
5717         return 0;
5718 }
5719
5720 static void btrfs_dentry_release(struct dentry *dentry)
5721 {
5722         kfree(dentry->d_fsdata);
5723 }
5724
5725 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5726                                    unsigned int flags)
5727 {
5728         struct inode *inode;
5729
5730         inode = btrfs_lookup_dentry(dir, dentry);
5731         if (IS_ERR(inode)) {
5732                 if (PTR_ERR(inode) == -ENOENT)
5733                         inode = NULL;
5734                 else
5735                         return ERR_CAST(inode);
5736         }
5737
5738         return d_splice_alias(inode, dentry);
5739 }
5740
5741 unsigned char btrfs_filetype_table[] = {
5742         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5743 };
5744
5745 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5746 {
5747         struct inode *inode = file_inode(file);
5748         struct btrfs_root *root = BTRFS_I(inode)->root;
5749         struct btrfs_item *item;
5750         struct btrfs_dir_item *di;
5751         struct btrfs_key key;
5752         struct btrfs_key found_key;
5753         struct btrfs_path *path;
5754         struct list_head ins_list;
5755         struct list_head del_list;
5756         int ret;
5757         struct extent_buffer *leaf;
5758         int slot;
5759         unsigned char d_type;
5760         int over = 0;
5761         u32 di_cur;
5762         u32 di_total;
5763         u32 di_len;
5764         int key_type = BTRFS_DIR_INDEX_KEY;
5765         char tmp_name[32];
5766         char *name_ptr;
5767         int name_len;
5768         int is_curr = 0;        /* ctx->pos points to the current index? */
5769         bool emitted;
5770         bool put = false;
5771
5772         /* FIXME, use a real flag for deciding about the key type */
5773         if (root->fs_info->tree_root == root)
5774                 key_type = BTRFS_DIR_ITEM_KEY;
5775
5776         if (!dir_emit_dots(file, ctx))
5777                 return 0;
5778
5779         path = btrfs_alloc_path();
5780         if (!path)
5781                 return -ENOMEM;
5782
5783         path->reada = READA_FORWARD;
5784
5785         if (key_type == BTRFS_DIR_INDEX_KEY) {
5786                 INIT_LIST_HEAD(&ins_list);
5787                 INIT_LIST_HEAD(&del_list);
5788                 put = btrfs_readdir_get_delayed_items(inode, &ins_list,
5789                                                       &del_list);
5790         }
5791
5792         key.type = key_type;
5793         key.offset = ctx->pos;
5794         key.objectid = btrfs_ino(inode);
5795
5796         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5797         if (ret < 0)
5798                 goto err;
5799
5800         emitted = false;
5801         while (1) {
5802                 leaf = path->nodes[0];
5803                 slot = path->slots[0];
5804                 if (slot >= btrfs_header_nritems(leaf)) {
5805                         ret = btrfs_next_leaf(root, path);
5806                         if (ret < 0)
5807                                 goto err;
5808                         else if (ret > 0)
5809                                 break;
5810                         continue;
5811                 }
5812
5813                 item = btrfs_item_nr(slot);
5814                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5815
5816                 if (found_key.objectid != key.objectid)
5817                         break;
5818                 if (found_key.type != key_type)
5819                         break;
5820                 if (found_key.offset < ctx->pos)
5821                         goto next;
5822                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5823                     btrfs_should_delete_dir_index(&del_list,
5824                                                   found_key.offset))
5825                         goto next;
5826
5827                 ctx->pos = found_key.offset;
5828                 is_curr = 1;
5829
5830                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5831                 di_cur = 0;
5832                 di_total = btrfs_item_size(leaf, item);
5833
5834                 while (di_cur < di_total) {
5835                         struct btrfs_key location;
5836
5837                         if (verify_dir_item(root, leaf, di))
5838                                 break;
5839
5840                         name_len = btrfs_dir_name_len(leaf, di);
5841                         if (name_len <= sizeof(tmp_name)) {
5842                                 name_ptr = tmp_name;
5843                         } else {
5844                                 name_ptr = kmalloc(name_len, GFP_KERNEL);
5845                                 if (!name_ptr) {
5846                                         ret = -ENOMEM;
5847                                         goto err;
5848                                 }
5849                         }
5850                         read_extent_buffer(leaf, name_ptr,
5851                                            (unsigned long)(di + 1), name_len);
5852
5853                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5854                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5855
5856
5857                         /* is this a reference to our own snapshot? If so
5858                          * skip it.
5859                          *
5860                          * In contrast to old kernels, we insert the snapshot's
5861                          * dir item and dir index after it has been created, so
5862                          * we won't find a reference to our own snapshot. We
5863                          * still keep the following code for backward
5864                          * compatibility.
5865                          */
5866                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5867                             location.objectid == root->root_key.objectid) {
5868                                 over = 0;
5869                                 goto skip;
5870                         }
5871                         over = !dir_emit(ctx, name_ptr, name_len,
5872                                        location.objectid, d_type);
5873
5874 skip:
5875                         if (name_ptr != tmp_name)
5876                                 kfree(name_ptr);
5877
5878                         if (over)
5879                                 goto nopos;
5880                         emitted = true;
5881                         di_len = btrfs_dir_name_len(leaf, di) +
5882                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5883                         di_cur += di_len;
5884                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5885                 }
5886 next:
5887                 path->slots[0]++;
5888         }
5889
5890         if (key_type == BTRFS_DIR_INDEX_KEY) {
5891                 if (is_curr)
5892                         ctx->pos++;
5893                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list, &emitted);
5894                 if (ret)
5895                         goto nopos;
5896         }
5897
5898         /*
5899          * If we haven't emitted any dir entry, we must not touch ctx->pos as
5900          * it was was set to the termination value in previous call. We assume
5901          * that "." and ".." were emitted if we reach this point and set the
5902          * termination value as well for an empty directory.
5903          */
5904         if (ctx->pos > 2 && !emitted)
5905                 goto nopos;
5906
5907         /* Reached end of directory/root. Bump pos past the last item. */
5908         ctx->pos++;
5909
5910         /*
5911          * Stop new entries from being returned after we return the last
5912          * entry.
5913          *
5914          * New directory entries are assigned a strictly increasing
5915          * offset.  This means that new entries created during readdir
5916          * are *guaranteed* to be seen in the future by that readdir.
5917          * This has broken buggy programs which operate on names as
5918          * they're returned by readdir.  Until we re-use freed offsets
5919          * we have this hack to stop new entries from being returned
5920          * under the assumption that they'll never reach this huge
5921          * offset.
5922          *
5923          * This is being careful not to overflow 32bit loff_t unless the
5924          * last entry requires it because doing so has broken 32bit apps
5925          * in the past.
5926          */
5927         if (key_type == BTRFS_DIR_INDEX_KEY) {
5928                 if (ctx->pos >= INT_MAX)
5929                         ctx->pos = LLONG_MAX;
5930                 else
5931                         ctx->pos = INT_MAX;
5932         }
5933 nopos:
5934         ret = 0;
5935 err:
5936         if (put)
5937                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5938         btrfs_free_path(path);
5939         return ret;
5940 }
5941
5942 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5943 {
5944         struct btrfs_root *root = BTRFS_I(inode)->root;
5945         struct btrfs_trans_handle *trans;
5946         int ret = 0;
5947         bool nolock = false;
5948
5949         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5950                 return 0;
5951
5952         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5953                 nolock = true;
5954
5955         if (wbc->sync_mode == WB_SYNC_ALL) {
5956                 if (nolock)
5957                         trans = btrfs_join_transaction_nolock(root);
5958                 else
5959                         trans = btrfs_join_transaction(root);
5960                 if (IS_ERR(trans))
5961                         return PTR_ERR(trans);
5962                 ret = btrfs_commit_transaction(trans, root);
5963         }
5964         return ret;
5965 }
5966
5967 /*
5968  * This is somewhat expensive, updating the tree every time the
5969  * inode changes.  But, it is most likely to find the inode in cache.
5970  * FIXME, needs more benchmarking...there are no reasons other than performance
5971  * to keep or drop this code.
5972  */
5973 static int btrfs_dirty_inode(struct inode *inode)
5974 {
5975         struct btrfs_root *root = BTRFS_I(inode)->root;
5976         struct btrfs_trans_handle *trans;
5977         int ret;
5978
5979         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5980                 return 0;
5981
5982         trans = btrfs_join_transaction(root);
5983         if (IS_ERR(trans))
5984                 return PTR_ERR(trans);
5985
5986         ret = btrfs_update_inode(trans, root, inode);
5987         if (ret && ret == -ENOSPC) {
5988                 /* whoops, lets try again with the full transaction */
5989                 btrfs_end_transaction(trans, root);
5990                 trans = btrfs_start_transaction(root, 1);
5991                 if (IS_ERR(trans))
5992                         return PTR_ERR(trans);
5993
5994                 ret = btrfs_update_inode(trans, root, inode);
5995         }
5996         btrfs_end_transaction(trans, root);
5997         if (BTRFS_I(inode)->delayed_node)
5998                 btrfs_balance_delayed_items(root);
5999
6000         return ret;
6001 }
6002
6003 /*
6004  * This is a copy of file_update_time.  We need this so we can return error on
6005  * ENOSPC for updating the inode in the case of file write and mmap writes.
6006  */
6007 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6008                              int flags)
6009 {
6010         struct btrfs_root *root = BTRFS_I(inode)->root;
6011
6012         if (btrfs_root_readonly(root))
6013                 return -EROFS;
6014
6015         if (flags & S_VERSION)
6016                 inode_inc_iversion(inode);
6017         if (flags & S_CTIME)
6018                 inode->i_ctime = *now;
6019         if (flags & S_MTIME)
6020                 inode->i_mtime = *now;
6021         if (flags & S_ATIME)
6022                 inode->i_atime = *now;
6023         return btrfs_dirty_inode(inode);
6024 }
6025
6026 /*
6027  * find the highest existing sequence number in a directory
6028  * and then set the in-memory index_cnt variable to reflect
6029  * free sequence numbers
6030  */
6031 static int btrfs_set_inode_index_count(struct inode *inode)
6032 {
6033         struct btrfs_root *root = BTRFS_I(inode)->root;
6034         struct btrfs_key key, found_key;
6035         struct btrfs_path *path;
6036         struct extent_buffer *leaf;
6037         int ret;
6038
6039         key.objectid = btrfs_ino(inode);
6040         key.type = BTRFS_DIR_INDEX_KEY;
6041         key.offset = (u64)-1;
6042
6043         path = btrfs_alloc_path();
6044         if (!path)
6045                 return -ENOMEM;
6046
6047         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6048         if (ret < 0)
6049                 goto out;
6050         /* FIXME: we should be able to handle this */
6051         if (ret == 0)
6052                 goto out;
6053         ret = 0;
6054
6055         /*
6056          * MAGIC NUMBER EXPLANATION:
6057          * since we search a directory based on f_pos we have to start at 2
6058          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6059          * else has to start at 2
6060          */
6061         if (path->slots[0] == 0) {
6062                 BTRFS_I(inode)->index_cnt = 2;
6063                 goto out;
6064         }
6065
6066         path->slots[0]--;
6067
6068         leaf = path->nodes[0];
6069         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6070
6071         if (found_key.objectid != btrfs_ino(inode) ||
6072             found_key.type != BTRFS_DIR_INDEX_KEY) {
6073                 BTRFS_I(inode)->index_cnt = 2;
6074                 goto out;
6075         }
6076
6077         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6078 out:
6079         btrfs_free_path(path);
6080         return ret;
6081 }
6082
6083 /*
6084  * helper to find a free sequence number in a given directory.  This current
6085  * code is very simple, later versions will do smarter things in the btree
6086  */
6087 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6088 {
6089         int ret = 0;
6090
6091         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6092                 ret = btrfs_inode_delayed_dir_index_count(dir);
6093                 if (ret) {
6094                         ret = btrfs_set_inode_index_count(dir);
6095                         if (ret)
6096                                 return ret;
6097                 }
6098         }
6099
6100         *index = BTRFS_I(dir)->index_cnt;
6101         BTRFS_I(dir)->index_cnt++;
6102
6103         return ret;
6104 }
6105
6106 static int btrfs_insert_inode_locked(struct inode *inode)
6107 {
6108         struct btrfs_iget_args args;
6109         args.location = &BTRFS_I(inode)->location;
6110         args.root = BTRFS_I(inode)->root;
6111
6112         return insert_inode_locked4(inode,
6113                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6114                    btrfs_find_actor, &args);
6115 }
6116
6117 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6118                                      struct btrfs_root *root,
6119                                      struct inode *dir,
6120                                      const char *name, int name_len,
6121                                      u64 ref_objectid, u64 objectid,
6122                                      umode_t mode, u64 *index)
6123 {
6124         struct inode *inode;
6125         struct btrfs_inode_item *inode_item;
6126         struct btrfs_key *location;
6127         struct btrfs_path *path;
6128         struct btrfs_inode_ref *ref;
6129         struct btrfs_key key[2];
6130         u32 sizes[2];
6131         int nitems = name ? 2 : 1;
6132         unsigned long ptr;
6133         int ret;
6134
6135         path = btrfs_alloc_path();
6136         if (!path)
6137                 return ERR_PTR(-ENOMEM);
6138
6139         inode = new_inode(root->fs_info->sb);
6140         if (!inode) {
6141                 btrfs_free_path(path);
6142                 return ERR_PTR(-ENOMEM);
6143         }
6144
6145         /*
6146          * O_TMPFILE, set link count to 0, so that after this point,
6147          * we fill in an inode item with the correct link count.
6148          */
6149         if (!name)
6150                 set_nlink(inode, 0);
6151
6152         /*
6153          * we have to initialize this early, so we can reclaim the inode
6154          * number if we fail afterwards in this function.
6155          */
6156         inode->i_ino = objectid;
6157
6158         if (dir && name) {
6159                 trace_btrfs_inode_request(dir);
6160
6161                 ret = btrfs_set_inode_index(dir, index);
6162                 if (ret) {
6163                         btrfs_free_path(path);
6164                         iput(inode);
6165                         return ERR_PTR(ret);
6166                 }
6167         } else if (dir) {
6168                 *index = 0;
6169         }
6170         /*
6171          * index_cnt is ignored for everything but a dir,
6172          * btrfs_get_inode_index_count has an explanation for the magic
6173          * number
6174          */
6175         BTRFS_I(inode)->index_cnt = 2;
6176         BTRFS_I(inode)->dir_index = *index;
6177         BTRFS_I(inode)->root = root;
6178         BTRFS_I(inode)->generation = trans->transid;
6179         inode->i_generation = BTRFS_I(inode)->generation;
6180
6181         /*
6182          * We could have gotten an inode number from somebody who was fsynced
6183          * and then removed in this same transaction, so let's just set full
6184          * sync since it will be a full sync anyway and this will blow away the
6185          * old info in the log.
6186          */
6187         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6188
6189         key[0].objectid = objectid;
6190         key[0].type = BTRFS_INODE_ITEM_KEY;
6191         key[0].offset = 0;
6192
6193         sizes[0] = sizeof(struct btrfs_inode_item);
6194
6195         if (name) {
6196                 /*
6197                  * Start new inodes with an inode_ref. This is slightly more
6198                  * efficient for small numbers of hard links since they will
6199                  * be packed into one item. Extended refs will kick in if we
6200                  * add more hard links than can fit in the ref item.
6201                  */
6202                 key[1].objectid = objectid;
6203                 key[1].type = BTRFS_INODE_REF_KEY;
6204                 key[1].offset = ref_objectid;
6205
6206                 sizes[1] = name_len + sizeof(*ref);
6207         }
6208
6209         location = &BTRFS_I(inode)->location;
6210         location->objectid = objectid;
6211         location->offset = 0;
6212         location->type = BTRFS_INODE_ITEM_KEY;
6213
6214         ret = btrfs_insert_inode_locked(inode);
6215         if (ret < 0)
6216                 goto fail;
6217
6218         path->leave_spinning = 1;
6219         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6220         if (ret != 0)
6221                 goto fail_unlock;
6222
6223         inode_init_owner(inode, dir, mode);
6224         inode_set_bytes(inode, 0);
6225
6226         inode->i_mtime = current_fs_time(inode->i_sb);
6227         inode->i_atime = inode->i_mtime;
6228         inode->i_ctime = inode->i_mtime;
6229         BTRFS_I(inode)->i_otime = inode->i_mtime;
6230
6231         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6232                                   struct btrfs_inode_item);
6233         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6234                              sizeof(*inode_item));
6235         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6236
6237         if (name) {
6238                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6239                                      struct btrfs_inode_ref);
6240                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6241                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6242                 ptr = (unsigned long)(ref + 1);
6243                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6244         }
6245
6246         btrfs_mark_buffer_dirty(path->nodes[0]);
6247         btrfs_free_path(path);
6248
6249         btrfs_inherit_iflags(inode, dir);
6250
6251         if (S_ISREG(mode)) {
6252                 if (btrfs_test_opt(root, NODATASUM))
6253                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6254                 if (btrfs_test_opt(root, NODATACOW))
6255                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6256                                 BTRFS_INODE_NODATASUM;
6257         }
6258
6259         inode_tree_add(inode);
6260
6261         trace_btrfs_inode_new(inode);
6262         btrfs_set_inode_last_trans(trans, inode);
6263
6264         btrfs_update_root_times(trans, root);
6265
6266         ret = btrfs_inode_inherit_props(trans, inode, dir);
6267         if (ret)
6268                 btrfs_err(root->fs_info,
6269                           "error inheriting props for ino %llu (root %llu): %d",
6270                           btrfs_ino(inode), root->root_key.objectid, ret);
6271
6272         return inode;
6273
6274 fail_unlock:
6275         unlock_new_inode(inode);
6276 fail:
6277         if (dir && name)
6278                 BTRFS_I(dir)->index_cnt--;
6279         btrfs_free_path(path);
6280         iput(inode);
6281         return ERR_PTR(ret);
6282 }
6283
6284 static inline u8 btrfs_inode_type(struct inode *inode)
6285 {
6286         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6287 }
6288
6289 /*
6290  * utility function to add 'inode' into 'parent_inode' with
6291  * a give name and a given sequence number.
6292  * if 'add_backref' is true, also insert a backref from the
6293  * inode to the parent directory.
6294  */
6295 int btrfs_add_link(struct btrfs_trans_handle *trans,
6296                    struct inode *parent_inode, struct inode *inode,
6297                    const char *name, int name_len, int add_backref, u64 index)
6298 {
6299         int ret = 0;
6300         struct btrfs_key key;
6301         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6302         u64 ino = btrfs_ino(inode);
6303         u64 parent_ino = btrfs_ino(parent_inode);
6304
6305         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6306                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6307         } else {
6308                 key.objectid = ino;
6309                 key.type = BTRFS_INODE_ITEM_KEY;
6310                 key.offset = 0;
6311         }
6312
6313         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6314                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6315                                          key.objectid, root->root_key.objectid,
6316                                          parent_ino, index, name, name_len);
6317         } else if (add_backref) {
6318                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6319                                              parent_ino, index);
6320         }
6321
6322         /* Nothing to clean up yet */
6323         if (ret)
6324                 return ret;
6325
6326         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6327                                     parent_inode, &key,
6328                                     btrfs_inode_type(inode), index);
6329         if (ret == -EEXIST || ret == -EOVERFLOW)
6330                 goto fail_dir_item;
6331         else if (ret) {
6332                 btrfs_abort_transaction(trans, root, ret);
6333                 return ret;
6334         }
6335
6336         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6337                            name_len * 2);
6338         inode_inc_iversion(parent_inode);
6339         parent_inode->i_mtime = parent_inode->i_ctime =
6340                 current_fs_time(parent_inode->i_sb);
6341         ret = btrfs_update_inode(trans, root, parent_inode);
6342         if (ret)
6343                 btrfs_abort_transaction(trans, root, ret);
6344         return ret;
6345
6346 fail_dir_item:
6347         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6348                 u64 local_index;
6349                 int err;
6350                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6351                                  key.objectid, root->root_key.objectid,
6352                                  parent_ino, &local_index, name, name_len);
6353
6354         } else if (add_backref) {
6355                 u64 local_index;
6356                 int err;
6357
6358                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6359                                           ino, parent_ino, &local_index);
6360         }
6361         return ret;
6362 }
6363
6364 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6365                             struct inode *dir, struct dentry *dentry,
6366                             struct inode *inode, int backref, u64 index)
6367 {
6368         int err = btrfs_add_link(trans, dir, inode,
6369                                  dentry->d_name.name, dentry->d_name.len,
6370                                  backref, index);
6371         if (err > 0)
6372                 err = -EEXIST;
6373         return err;
6374 }
6375
6376 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6377                         umode_t mode, dev_t rdev)
6378 {
6379         struct btrfs_trans_handle *trans;
6380         struct btrfs_root *root = BTRFS_I(dir)->root;
6381         struct inode *inode = NULL;
6382         int err;
6383         int drop_inode = 0;
6384         u64 objectid;
6385         u64 index = 0;
6386
6387         /*
6388          * 2 for inode item and ref
6389          * 2 for dir items
6390          * 1 for xattr if selinux is on
6391          */
6392         trans = btrfs_start_transaction(root, 5);
6393         if (IS_ERR(trans))
6394                 return PTR_ERR(trans);
6395
6396         err = btrfs_find_free_ino(root, &objectid);
6397         if (err)
6398                 goto out_unlock;
6399
6400         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6401                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6402                                 mode, &index);
6403         if (IS_ERR(inode)) {
6404                 err = PTR_ERR(inode);
6405                 goto out_unlock;
6406         }
6407
6408         /*
6409         * If the active LSM wants to access the inode during
6410         * d_instantiate it needs these. Smack checks to see
6411         * if the filesystem supports xattrs by looking at the
6412         * ops vector.
6413         */
6414         inode->i_op = &btrfs_special_inode_operations;
6415         init_special_inode(inode, inode->i_mode, rdev);
6416
6417         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6418         if (err)
6419                 goto out_unlock_inode;
6420
6421         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6422         if (err) {
6423                 goto out_unlock_inode;
6424         } else {
6425                 btrfs_update_inode(trans, root, inode);
6426                 unlock_new_inode(inode);
6427                 d_instantiate(dentry, inode);
6428         }
6429
6430 out_unlock:
6431         btrfs_end_transaction(trans, root);
6432         btrfs_balance_delayed_items(root);
6433         btrfs_btree_balance_dirty(root);
6434         if (drop_inode) {
6435                 inode_dec_link_count(inode);
6436                 iput(inode);
6437         }
6438         return err;
6439
6440 out_unlock_inode:
6441         drop_inode = 1;
6442         unlock_new_inode(inode);
6443         goto out_unlock;
6444
6445 }
6446
6447 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6448                         umode_t mode, bool excl)
6449 {
6450         struct btrfs_trans_handle *trans;
6451         struct btrfs_root *root = BTRFS_I(dir)->root;
6452         struct inode *inode = NULL;
6453         int drop_inode_on_err = 0;
6454         int err;
6455         u64 objectid;
6456         u64 index = 0;
6457
6458         /*
6459          * 2 for inode item and ref
6460          * 2 for dir items
6461          * 1 for xattr if selinux is on
6462          */
6463         trans = btrfs_start_transaction(root, 5);
6464         if (IS_ERR(trans))
6465                 return PTR_ERR(trans);
6466
6467         err = btrfs_find_free_ino(root, &objectid);
6468         if (err)
6469                 goto out_unlock;
6470
6471         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6472                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6473                                 mode, &index);
6474         if (IS_ERR(inode)) {
6475                 err = PTR_ERR(inode);
6476                 goto out_unlock;
6477         }
6478         drop_inode_on_err = 1;
6479         /*
6480         * If the active LSM wants to access the inode during
6481         * d_instantiate it needs these. Smack checks to see
6482         * if the filesystem supports xattrs by looking at the
6483         * ops vector.
6484         */
6485         inode->i_fop = &btrfs_file_operations;
6486         inode->i_op = &btrfs_file_inode_operations;
6487         inode->i_mapping->a_ops = &btrfs_aops;
6488
6489         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6490         if (err)
6491                 goto out_unlock_inode;
6492
6493         err = btrfs_update_inode(trans, root, inode);
6494         if (err)
6495                 goto out_unlock_inode;
6496
6497         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6498         if (err)
6499                 goto out_unlock_inode;
6500
6501         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6502         unlock_new_inode(inode);
6503         d_instantiate(dentry, inode);
6504
6505 out_unlock:
6506         btrfs_end_transaction(trans, root);
6507         if (err && drop_inode_on_err) {
6508                 inode_dec_link_count(inode);
6509                 iput(inode);
6510         }
6511         btrfs_balance_delayed_items(root);
6512         btrfs_btree_balance_dirty(root);
6513         return err;
6514
6515 out_unlock_inode:
6516         unlock_new_inode(inode);
6517         goto out_unlock;
6518
6519 }
6520
6521 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6522                       struct dentry *dentry)
6523 {
6524         struct btrfs_trans_handle *trans = NULL;
6525         struct btrfs_root *root = BTRFS_I(dir)->root;
6526         struct inode *inode = d_inode(old_dentry);
6527         u64 index;
6528         int err;
6529         int drop_inode = 0;
6530
6531         /* do not allow sys_link's with other subvols of the same device */
6532         if (root->objectid != BTRFS_I(inode)->root->objectid)
6533                 return -EXDEV;
6534
6535         if (inode->i_nlink >= BTRFS_LINK_MAX)
6536                 return -EMLINK;
6537
6538         err = btrfs_set_inode_index(dir, &index);
6539         if (err)
6540                 goto fail;
6541
6542         /*
6543          * 2 items for inode and inode ref
6544          * 2 items for dir items
6545          * 1 item for parent inode
6546          */
6547         trans = btrfs_start_transaction(root, 5);
6548         if (IS_ERR(trans)) {
6549                 err = PTR_ERR(trans);
6550                 trans = NULL;
6551                 goto fail;
6552         }
6553
6554         /* There are several dir indexes for this inode, clear the cache. */
6555         BTRFS_I(inode)->dir_index = 0ULL;
6556         inc_nlink(inode);
6557         inode_inc_iversion(inode);
6558         inode->i_ctime = current_fs_time(inode->i_sb);
6559         ihold(inode);
6560         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6561
6562         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6563
6564         if (err) {
6565                 drop_inode = 1;
6566         } else {
6567                 struct dentry *parent = dentry->d_parent;
6568                 err = btrfs_update_inode(trans, root, inode);
6569                 if (err)
6570                         goto fail;
6571                 if (inode->i_nlink == 1) {
6572                         /*
6573                          * If new hard link count is 1, it's a file created
6574                          * with open(2) O_TMPFILE flag.
6575                          */
6576                         err = btrfs_orphan_del(trans, inode);
6577                         if (err)
6578                                 goto fail;
6579                 }
6580                 d_instantiate(dentry, inode);
6581                 btrfs_log_new_name(trans, inode, NULL, parent);
6582         }
6583
6584         btrfs_balance_delayed_items(root);
6585 fail:
6586         if (trans)
6587                 btrfs_end_transaction(trans, root);
6588         if (drop_inode) {
6589                 inode_dec_link_count(inode);
6590                 iput(inode);
6591         }
6592         btrfs_btree_balance_dirty(root);
6593         return err;
6594 }
6595
6596 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6597 {
6598         struct inode *inode = NULL;
6599         struct btrfs_trans_handle *trans;
6600         struct btrfs_root *root = BTRFS_I(dir)->root;
6601         int err = 0;
6602         int drop_on_err = 0;
6603         u64 objectid = 0;
6604         u64 index = 0;
6605
6606         /*
6607          * 2 items for inode and ref
6608          * 2 items for dir items
6609          * 1 for xattr if selinux is on
6610          */
6611         trans = btrfs_start_transaction(root, 5);
6612         if (IS_ERR(trans))
6613                 return PTR_ERR(trans);
6614
6615         err = btrfs_find_free_ino(root, &objectid);
6616         if (err)
6617                 goto out_fail;
6618
6619         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6620                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6621                                 S_IFDIR | mode, &index);
6622         if (IS_ERR(inode)) {
6623                 err = PTR_ERR(inode);
6624                 goto out_fail;
6625         }
6626
6627         drop_on_err = 1;
6628         /* these must be set before we unlock the inode */
6629         inode->i_op = &btrfs_dir_inode_operations;
6630         inode->i_fop = &btrfs_dir_file_operations;
6631
6632         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6633         if (err)
6634                 goto out_fail_inode;
6635
6636         btrfs_i_size_write(inode, 0);
6637         err = btrfs_update_inode(trans, root, inode);
6638         if (err)
6639                 goto out_fail_inode;
6640
6641         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6642                              dentry->d_name.len, 0, index);
6643         if (err)
6644                 goto out_fail_inode;
6645
6646         d_instantiate(dentry, inode);
6647         /*
6648          * mkdir is special.  We're unlocking after we call d_instantiate
6649          * to avoid a race with nfsd calling d_instantiate.
6650          */
6651         unlock_new_inode(inode);
6652         drop_on_err = 0;
6653
6654 out_fail:
6655         btrfs_end_transaction(trans, root);
6656         if (drop_on_err) {
6657                 inode_dec_link_count(inode);
6658                 iput(inode);
6659         }
6660         btrfs_balance_delayed_items(root);
6661         btrfs_btree_balance_dirty(root);
6662         return err;
6663
6664 out_fail_inode:
6665         unlock_new_inode(inode);
6666         goto out_fail;
6667 }
6668
6669 /* Find next extent map of a given extent map, caller needs to ensure locks */
6670 static struct extent_map *next_extent_map(struct extent_map *em)
6671 {
6672         struct rb_node *next;
6673
6674         next = rb_next(&em->rb_node);
6675         if (!next)
6676                 return NULL;
6677         return container_of(next, struct extent_map, rb_node);
6678 }
6679
6680 static struct extent_map *prev_extent_map(struct extent_map *em)
6681 {
6682         struct rb_node *prev;
6683
6684         prev = rb_prev(&em->rb_node);
6685         if (!prev)
6686                 return NULL;
6687         return container_of(prev, struct extent_map, rb_node);
6688 }
6689
6690 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6691  * the existing extent is the nearest extent to map_start,
6692  * and an extent that you want to insert, deal with overlap and insert
6693  * the best fitted new extent into the tree.
6694  */
6695 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6696                                 struct extent_map *existing,
6697                                 struct extent_map *em,
6698                                 u64 map_start)
6699 {
6700         struct extent_map *prev;
6701         struct extent_map *next;
6702         u64 start;
6703         u64 end;
6704         u64 start_diff;
6705
6706         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6707
6708         if (existing->start > map_start) {
6709                 next = existing;
6710                 prev = prev_extent_map(next);
6711         } else {
6712                 prev = existing;
6713                 next = next_extent_map(prev);
6714         }
6715
6716         start = prev ? extent_map_end(prev) : em->start;
6717         start = max_t(u64, start, em->start);
6718         end = next ? next->start : extent_map_end(em);
6719         end = min_t(u64, end, extent_map_end(em));
6720         start_diff = start - em->start;
6721         em->start = start;
6722         em->len = end - start;
6723         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6724             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6725                 em->block_start += start_diff;
6726                 em->block_len -= start_diff;
6727         }
6728         return add_extent_mapping(em_tree, em, 0);
6729 }
6730
6731 static noinline int uncompress_inline(struct btrfs_path *path,
6732                                       struct page *page,
6733                                       size_t pg_offset, u64 extent_offset,
6734                                       struct btrfs_file_extent_item *item)
6735 {
6736         int ret;
6737         struct extent_buffer *leaf = path->nodes[0];
6738         char *tmp;
6739         size_t max_size;
6740         unsigned long inline_size;
6741         unsigned long ptr;
6742         int compress_type;
6743
6744         WARN_ON(pg_offset != 0);
6745         compress_type = btrfs_file_extent_compression(leaf, item);
6746         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6747         inline_size = btrfs_file_extent_inline_item_len(leaf,
6748                                         btrfs_item_nr(path->slots[0]));
6749         tmp = kmalloc(inline_size, GFP_NOFS);
6750         if (!tmp)
6751                 return -ENOMEM;
6752         ptr = btrfs_file_extent_inline_start(item);
6753
6754         read_extent_buffer(leaf, tmp, ptr, inline_size);
6755
6756         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6757         ret = btrfs_decompress(compress_type, tmp, page,
6758                                extent_offset, inline_size, max_size);
6759         kfree(tmp);
6760         return ret;
6761 }
6762
6763 /*
6764  * a bit scary, this does extent mapping from logical file offset to the disk.
6765  * the ugly parts come from merging extents from the disk with the in-ram
6766  * representation.  This gets more complex because of the data=ordered code,
6767  * where the in-ram extents might be locked pending data=ordered completion.
6768  *
6769  * This also copies inline extents directly into the page.
6770  */
6771
6772 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6773                                     size_t pg_offset, u64 start, u64 len,
6774                                     int create)
6775 {
6776         int ret;
6777         int err = 0;
6778         u64 extent_start = 0;
6779         u64 extent_end = 0;
6780         u64 objectid = btrfs_ino(inode);
6781         u32 found_type;
6782         struct btrfs_path *path = NULL;
6783         struct btrfs_root *root = BTRFS_I(inode)->root;
6784         struct btrfs_file_extent_item *item;
6785         struct extent_buffer *leaf;
6786         struct btrfs_key found_key;
6787         struct extent_map *em = NULL;
6788         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6789         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6790         struct btrfs_trans_handle *trans = NULL;
6791         const bool new_inline = !page || create;
6792
6793 again:
6794         read_lock(&em_tree->lock);
6795         em = lookup_extent_mapping(em_tree, start, len);
6796         if (em)
6797                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6798         read_unlock(&em_tree->lock);
6799
6800         if (em) {
6801                 if (em->start > start || em->start + em->len <= start)
6802                         free_extent_map(em);
6803                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6804                         free_extent_map(em);
6805                 else
6806                         goto out;
6807         }
6808         em = alloc_extent_map();
6809         if (!em) {
6810                 err = -ENOMEM;
6811                 goto out;
6812         }
6813         em->bdev = root->fs_info->fs_devices->latest_bdev;
6814         em->start = EXTENT_MAP_HOLE;
6815         em->orig_start = EXTENT_MAP_HOLE;
6816         em->len = (u64)-1;
6817         em->block_len = (u64)-1;
6818
6819         if (!path) {
6820                 path = btrfs_alloc_path();
6821                 if (!path) {
6822                         err = -ENOMEM;
6823                         goto out;
6824                 }
6825                 /*
6826                  * Chances are we'll be called again, so go ahead and do
6827                  * readahead
6828                  */
6829                 path->reada = READA_FORWARD;
6830         }
6831
6832         ret = btrfs_lookup_file_extent(trans, root, path,
6833                                        objectid, start, trans != NULL);
6834         if (ret < 0) {
6835                 err = ret;
6836                 goto out;
6837         }
6838
6839         if (ret != 0) {
6840                 if (path->slots[0] == 0)
6841                         goto not_found;
6842                 path->slots[0]--;
6843         }
6844
6845         leaf = path->nodes[0];
6846         item = btrfs_item_ptr(leaf, path->slots[0],
6847                               struct btrfs_file_extent_item);
6848         /* are we inside the extent that was found? */
6849         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6850         found_type = found_key.type;
6851         if (found_key.objectid != objectid ||
6852             found_type != BTRFS_EXTENT_DATA_KEY) {
6853                 /*
6854                  * If we backup past the first extent we want to move forward
6855                  * and see if there is an extent in front of us, otherwise we'll
6856                  * say there is a hole for our whole search range which can
6857                  * cause problems.
6858                  */
6859                 extent_end = start;
6860                 goto next;
6861         }
6862
6863         found_type = btrfs_file_extent_type(leaf, item);
6864         extent_start = found_key.offset;
6865         if (found_type == BTRFS_FILE_EXTENT_REG ||
6866             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6867                 extent_end = extent_start +
6868                        btrfs_file_extent_num_bytes(leaf, item);
6869         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6870                 size_t size;
6871                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6872                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6873         }
6874 next:
6875         if (start >= extent_end) {
6876                 path->slots[0]++;
6877                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6878                         ret = btrfs_next_leaf(root, path);
6879                         if (ret < 0) {
6880                                 err = ret;
6881                                 goto out;
6882                         }
6883                         if (ret > 0)
6884                                 goto not_found;
6885                         leaf = path->nodes[0];
6886                 }
6887                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6888                 if (found_key.objectid != objectid ||
6889                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6890                         goto not_found;
6891                 if (start + len <= found_key.offset)
6892                         goto not_found;
6893                 if (start > found_key.offset)
6894                         goto next;
6895                 em->start = start;
6896                 em->orig_start = start;
6897                 em->len = found_key.offset - start;
6898                 goto not_found_em;
6899         }
6900
6901         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6902
6903         if (found_type == BTRFS_FILE_EXTENT_REG ||
6904             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6905                 goto insert;
6906         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6907                 unsigned long ptr;
6908                 char *map;
6909                 size_t size;
6910                 size_t extent_offset;
6911                 size_t copy_size;
6912
6913                 if (new_inline)
6914                         goto out;
6915
6916                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6917                 extent_offset = page_offset(page) + pg_offset - extent_start;
6918                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6919                                   size - extent_offset);
6920                 em->start = extent_start + extent_offset;
6921                 em->len = ALIGN(copy_size, root->sectorsize);
6922                 em->orig_block_len = em->len;
6923                 em->orig_start = em->start;
6924                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6925                 if (create == 0 && !PageUptodate(page)) {
6926                         if (btrfs_file_extent_compression(leaf, item) !=
6927                             BTRFS_COMPRESS_NONE) {
6928                                 ret = uncompress_inline(path, page, pg_offset,
6929                                                         extent_offset, item);
6930                                 if (ret) {
6931                                         err = ret;
6932                                         goto out;
6933                                 }
6934                         } else {
6935                                 map = kmap(page);
6936                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6937                                                    copy_size);
6938                                 if (pg_offset + copy_size < PAGE_SIZE) {
6939                                         memset(map + pg_offset + copy_size, 0,
6940                                                PAGE_SIZE - pg_offset -
6941                                                copy_size);
6942                                 }
6943                                 kunmap(page);
6944                         }
6945                         flush_dcache_page(page);
6946                 } else if (create && PageUptodate(page)) {
6947                         BUG();
6948                         if (!trans) {
6949                                 kunmap(page);
6950                                 free_extent_map(em);
6951                                 em = NULL;
6952
6953                                 btrfs_release_path(path);
6954                                 trans = btrfs_join_transaction(root);
6955
6956                                 if (IS_ERR(trans))
6957                                         return ERR_CAST(trans);
6958                                 goto again;
6959                         }
6960                         map = kmap(page);
6961                         write_extent_buffer(leaf, map + pg_offset, ptr,
6962                                             copy_size);
6963                         kunmap(page);
6964                         btrfs_mark_buffer_dirty(leaf);
6965                 }
6966                 set_extent_uptodate(io_tree, em->start,
6967                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6968                 goto insert;
6969         }
6970 not_found:
6971         em->start = start;
6972         em->orig_start = start;
6973         em->len = len;
6974 not_found_em:
6975         em->block_start = EXTENT_MAP_HOLE;
6976         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6977 insert:
6978         btrfs_release_path(path);
6979         if (em->start > start || extent_map_end(em) <= start) {
6980                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6981                         em->start, em->len, start, len);
6982                 err = -EIO;
6983                 goto out;
6984         }
6985
6986         err = 0;
6987         write_lock(&em_tree->lock);
6988         ret = add_extent_mapping(em_tree, em, 0);
6989         /* it is possible that someone inserted the extent into the tree
6990          * while we had the lock dropped.  It is also possible that
6991          * an overlapping map exists in the tree
6992          */
6993         if (ret == -EEXIST) {
6994                 struct extent_map *existing;
6995
6996                 ret = 0;
6997
6998                 existing = search_extent_mapping(em_tree, start, len);
6999                 /*
7000                  * existing will always be non-NULL, since there must be
7001                  * extent causing the -EEXIST.
7002                  */
7003                 if (existing->start == em->start &&
7004                     extent_map_end(existing) == extent_map_end(em) &&
7005                     em->block_start == existing->block_start) {
7006                         /*
7007                          * these two extents are the same, it happens
7008                          * with inlines especially
7009                          */
7010                         free_extent_map(em);
7011                         em = existing;
7012                         err = 0;
7013
7014                 } else if (start >= extent_map_end(existing) ||
7015                     start <= existing->start) {
7016                         /*
7017                          * The existing extent map is the one nearest to
7018                          * the [start, start + len) range which overlaps
7019                          */
7020                         err = merge_extent_mapping(em_tree, existing,
7021                                                    em, start);
7022                         free_extent_map(existing);
7023                         if (err) {
7024                                 free_extent_map(em);
7025                                 em = NULL;
7026                         }
7027                 } else {
7028                         free_extent_map(em);
7029                         em = existing;
7030                         err = 0;
7031                 }
7032         }
7033         write_unlock(&em_tree->lock);
7034 out:
7035
7036         trace_btrfs_get_extent(root, em);
7037
7038         btrfs_free_path(path);
7039         if (trans) {
7040                 ret = btrfs_end_transaction(trans, root);
7041                 if (!err)
7042                         err = ret;
7043         }
7044         if (err) {
7045                 free_extent_map(em);
7046                 return ERR_PTR(err);
7047         }
7048         BUG_ON(!em); /* Error is always set */
7049         return em;
7050 }
7051
7052 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7053                                            size_t pg_offset, u64 start, u64 len,
7054                                            int create)
7055 {
7056         struct extent_map *em;
7057         struct extent_map *hole_em = NULL;
7058         u64 range_start = start;
7059         u64 end;
7060         u64 found;
7061         u64 found_end;
7062         int err = 0;
7063
7064         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7065         if (IS_ERR(em))
7066                 return em;
7067         if (em) {
7068                 /*
7069                  * if our em maps to
7070                  * -  a hole or
7071                  * -  a pre-alloc extent,
7072                  * there might actually be delalloc bytes behind it.
7073                  */
7074                 if (em->block_start != EXTENT_MAP_HOLE &&
7075                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7076                         return em;
7077                 else
7078                         hole_em = em;
7079         }
7080
7081         /* check to see if we've wrapped (len == -1 or similar) */
7082         end = start + len;
7083         if (end < start)
7084                 end = (u64)-1;
7085         else
7086                 end -= 1;
7087
7088         em = NULL;
7089
7090         /* ok, we didn't find anything, lets look for delalloc */
7091         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7092                                  end, len, EXTENT_DELALLOC, 1);
7093         found_end = range_start + found;
7094         if (found_end < range_start)
7095                 found_end = (u64)-1;
7096
7097         /*
7098          * we didn't find anything useful, return
7099          * the original results from get_extent()
7100          */
7101         if (range_start > end || found_end <= start) {
7102                 em = hole_em;
7103                 hole_em = NULL;
7104                 goto out;
7105         }
7106
7107         /* adjust the range_start to make sure it doesn't
7108          * go backwards from the start they passed in
7109          */
7110         range_start = max(start, range_start);
7111         found = found_end - range_start;
7112
7113         if (found > 0) {
7114                 u64 hole_start = start;
7115                 u64 hole_len = len;
7116
7117                 em = alloc_extent_map();
7118                 if (!em) {
7119                         err = -ENOMEM;
7120                         goto out;
7121                 }
7122                 /*
7123                  * when btrfs_get_extent can't find anything it
7124                  * returns one huge hole
7125                  *
7126                  * make sure what it found really fits our range, and
7127                  * adjust to make sure it is based on the start from
7128                  * the caller
7129                  */
7130                 if (hole_em) {
7131                         u64 calc_end = extent_map_end(hole_em);
7132
7133                         if (calc_end <= start || (hole_em->start > end)) {
7134                                 free_extent_map(hole_em);
7135                                 hole_em = NULL;
7136                         } else {
7137                                 hole_start = max(hole_em->start, start);
7138                                 hole_len = calc_end - hole_start;
7139                         }
7140                 }
7141                 em->bdev = NULL;
7142                 if (hole_em && range_start > hole_start) {
7143                         /* our hole starts before our delalloc, so we
7144                          * have to return just the parts of the hole
7145                          * that go until  the delalloc starts
7146                          */
7147                         em->len = min(hole_len,
7148                                       range_start - hole_start);
7149                         em->start = hole_start;
7150                         em->orig_start = hole_start;
7151                         /*
7152                          * don't adjust block start at all,
7153                          * it is fixed at EXTENT_MAP_HOLE
7154                          */
7155                         em->block_start = hole_em->block_start;
7156                         em->block_len = hole_len;
7157                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7158                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7159                 } else {
7160                         em->start = range_start;
7161                         em->len = found;
7162                         em->orig_start = range_start;
7163                         em->block_start = EXTENT_MAP_DELALLOC;
7164                         em->block_len = found;
7165                 }
7166         } else if (hole_em) {
7167                 return hole_em;
7168         }
7169 out:
7170
7171         free_extent_map(hole_em);
7172         if (err) {
7173                 free_extent_map(em);
7174                 return ERR_PTR(err);
7175         }
7176         return em;
7177 }
7178
7179 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7180                                                   const u64 start,
7181                                                   const u64 len,
7182                                                   const u64 orig_start,
7183                                                   const u64 block_start,
7184                                                   const u64 block_len,
7185                                                   const u64 orig_block_len,
7186                                                   const u64 ram_bytes,
7187                                                   const int type)
7188 {
7189         struct extent_map *em = NULL;
7190         int ret;
7191
7192         down_read(&BTRFS_I(inode)->dio_sem);
7193         if (type != BTRFS_ORDERED_NOCOW) {
7194                 em = create_pinned_em(inode, start, len, orig_start,
7195                                       block_start, block_len, orig_block_len,
7196                                       ram_bytes, type);
7197                 if (IS_ERR(em))
7198                         goto out;
7199         }
7200         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7201                                            len, block_len, type);
7202         if (ret) {
7203                 if (em) {
7204                         free_extent_map(em);
7205                         btrfs_drop_extent_cache(inode, start,
7206                                                 start + len - 1, 0);
7207                 }
7208                 em = ERR_PTR(ret);
7209         }
7210  out:
7211         up_read(&BTRFS_I(inode)->dio_sem);
7212
7213         return em;
7214 }
7215
7216 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7217                                                   u64 start, u64 len)
7218 {
7219         struct btrfs_root *root = BTRFS_I(inode)->root;
7220         struct extent_map *em;
7221         struct btrfs_key ins;
7222         u64 alloc_hint;
7223         int ret;
7224
7225         alloc_hint = get_extent_allocation_hint(inode, start, len);
7226         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7227                                    alloc_hint, &ins, 1, 1);
7228         if (ret)
7229                 return ERR_PTR(ret);
7230
7231         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7232                                      ins.objectid, ins.offset, ins.offset,
7233                                      ins.offset, 0);
7234         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
7235         if (IS_ERR(em))
7236                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7237
7238         return em;
7239 }
7240
7241 /*
7242  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7243  * block must be cow'd
7244  */
7245 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7246                               u64 *orig_start, u64 *orig_block_len,
7247                               u64 *ram_bytes)
7248 {
7249         struct btrfs_trans_handle *trans;
7250         struct btrfs_path *path;
7251         int ret;
7252         struct extent_buffer *leaf;
7253         struct btrfs_root *root = BTRFS_I(inode)->root;
7254         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7255         struct btrfs_file_extent_item *fi;
7256         struct btrfs_key key;
7257         u64 disk_bytenr;
7258         u64 backref_offset;
7259         u64 extent_end;
7260         u64 num_bytes;
7261         int slot;
7262         int found_type;
7263         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7264
7265         path = btrfs_alloc_path();
7266         if (!path)
7267                 return -ENOMEM;
7268
7269         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7270                                        offset, 0);
7271         if (ret < 0)
7272                 goto out;
7273
7274         slot = path->slots[0];
7275         if (ret == 1) {
7276                 if (slot == 0) {
7277                         /* can't find the item, must cow */
7278                         ret = 0;
7279                         goto out;
7280                 }
7281                 slot--;
7282         }
7283         ret = 0;
7284         leaf = path->nodes[0];
7285         btrfs_item_key_to_cpu(leaf, &key, slot);
7286         if (key.objectid != btrfs_ino(inode) ||
7287             key.type != BTRFS_EXTENT_DATA_KEY) {
7288                 /* not our file or wrong item type, must cow */
7289                 goto out;
7290         }
7291
7292         if (key.offset > offset) {
7293                 /* Wrong offset, must cow */
7294                 goto out;
7295         }
7296
7297         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7298         found_type = btrfs_file_extent_type(leaf, fi);
7299         if (found_type != BTRFS_FILE_EXTENT_REG &&
7300             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7301                 /* not a regular extent, must cow */
7302                 goto out;
7303         }
7304
7305         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7306                 goto out;
7307
7308         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7309         if (extent_end <= offset)
7310                 goto out;
7311
7312         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7313         if (disk_bytenr == 0)
7314                 goto out;
7315
7316         if (btrfs_file_extent_compression(leaf, fi) ||
7317             btrfs_file_extent_encryption(leaf, fi) ||
7318             btrfs_file_extent_other_encoding(leaf, fi))
7319                 goto out;
7320
7321         backref_offset = btrfs_file_extent_offset(leaf, fi);
7322
7323         if (orig_start) {
7324                 *orig_start = key.offset - backref_offset;
7325                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7326                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7327         }
7328
7329         if (btrfs_extent_readonly(root, disk_bytenr))
7330                 goto out;
7331
7332         num_bytes = min(offset + *len, extent_end) - offset;
7333         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7334                 u64 range_end;
7335
7336                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7337                 ret = test_range_bit(io_tree, offset, range_end,
7338                                      EXTENT_DELALLOC, 0, NULL);
7339                 if (ret) {
7340                         ret = -EAGAIN;
7341                         goto out;
7342                 }
7343         }
7344
7345         btrfs_release_path(path);
7346
7347         /*
7348          * look for other files referencing this extent, if we
7349          * find any we must cow
7350          */
7351         trans = btrfs_join_transaction(root);
7352         if (IS_ERR(trans)) {
7353                 ret = 0;
7354                 goto out;
7355         }
7356
7357         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7358                                     key.offset - backref_offset, disk_bytenr);
7359         btrfs_end_transaction(trans, root);
7360         if (ret) {
7361                 ret = 0;
7362                 goto out;
7363         }
7364
7365         /*
7366          * adjust disk_bytenr and num_bytes to cover just the bytes
7367          * in this extent we are about to write.  If there
7368          * are any csums in that range we have to cow in order
7369          * to keep the csums correct
7370          */
7371         disk_bytenr += backref_offset;
7372         disk_bytenr += offset - key.offset;
7373         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7374                                 goto out;
7375         /*
7376          * all of the above have passed, it is safe to overwrite this extent
7377          * without cow
7378          */
7379         *len = num_bytes;
7380         ret = 1;
7381 out:
7382         btrfs_free_path(path);
7383         return ret;
7384 }
7385
7386 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7387 {
7388         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7389         int found = false;
7390         void **pagep = NULL;
7391         struct page *page = NULL;
7392         int start_idx;
7393         int end_idx;
7394
7395         start_idx = start >> PAGE_SHIFT;
7396
7397         /*
7398          * end is the last byte in the last page.  end == start is legal
7399          */
7400         end_idx = end >> PAGE_SHIFT;
7401
7402         rcu_read_lock();
7403
7404         /* Most of the code in this while loop is lifted from
7405          * find_get_page.  It's been modified to begin searching from a
7406          * page and return just the first page found in that range.  If the
7407          * found idx is less than or equal to the end idx then we know that
7408          * a page exists.  If no pages are found or if those pages are
7409          * outside of the range then we're fine (yay!) */
7410         while (page == NULL &&
7411                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7412                 page = radix_tree_deref_slot(pagep);
7413                 if (unlikely(!page))
7414                         break;
7415
7416                 if (radix_tree_exception(page)) {
7417                         if (radix_tree_deref_retry(page)) {
7418                                 page = NULL;
7419                                 continue;
7420                         }
7421                         /*
7422                          * Otherwise, shmem/tmpfs must be storing a swap entry
7423                          * here as an exceptional entry: so return it without
7424                          * attempting to raise page count.
7425                          */
7426                         page = NULL;
7427                         break; /* TODO: Is this relevant for this use case? */
7428                 }
7429
7430                 if (!page_cache_get_speculative(page)) {
7431                         page = NULL;
7432                         continue;
7433                 }
7434
7435                 /*
7436                  * Has the page moved?
7437                  * This is part of the lockless pagecache protocol. See
7438                  * include/linux/pagemap.h for details.
7439                  */
7440                 if (unlikely(page != *pagep)) {
7441                         put_page(page);
7442                         page = NULL;
7443                 }
7444         }
7445
7446         if (page) {
7447                 if (page->index <= end_idx)
7448                         found = true;
7449                 put_page(page);
7450         }
7451
7452         rcu_read_unlock();
7453         return found;
7454 }
7455
7456 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7457                               struct extent_state **cached_state, int writing)
7458 {
7459         struct btrfs_ordered_extent *ordered;
7460         int ret = 0;
7461
7462         while (1) {
7463                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7464                                  cached_state);
7465                 /*
7466                  * We're concerned with the entire range that we're going to be
7467                  * doing DIO to, so we need to make sure there's no ordered
7468                  * extents in this range.
7469                  */
7470                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7471                                                      lockend - lockstart + 1);
7472
7473                 /*
7474                  * We need to make sure there are no buffered pages in this
7475                  * range either, we could have raced between the invalidate in
7476                  * generic_file_direct_write and locking the extent.  The
7477                  * invalidate needs to happen so that reads after a write do not
7478                  * get stale data.
7479                  */
7480                 if (!ordered &&
7481                     (!writing ||
7482                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7483                         break;
7484
7485                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7486                                      cached_state, GFP_NOFS);
7487
7488                 if (ordered) {
7489                         /*
7490                          * If we are doing a DIO read and the ordered extent we
7491                          * found is for a buffered write, we can not wait for it
7492                          * to complete and retry, because if we do so we can
7493                          * deadlock with concurrent buffered writes on page
7494                          * locks. This happens only if our DIO read covers more
7495                          * than one extent map, if at this point has already
7496                          * created an ordered extent for a previous extent map
7497                          * and locked its range in the inode's io tree, and a
7498                          * concurrent write against that previous extent map's
7499                          * range and this range started (we unlock the ranges
7500                          * in the io tree only when the bios complete and
7501                          * buffered writes always lock pages before attempting
7502                          * to lock range in the io tree).
7503                          */
7504                         if (writing ||
7505                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7506                                 btrfs_start_ordered_extent(inode, ordered, 1);
7507                         else
7508                                 ret = -ENOTBLK;
7509                         btrfs_put_ordered_extent(ordered);
7510                 } else {
7511                         /*
7512                          * We could trigger writeback for this range (and wait
7513                          * for it to complete) and then invalidate the pages for
7514                          * this range (through invalidate_inode_pages2_range()),
7515                          * but that can lead us to a deadlock with a concurrent
7516                          * call to readpages() (a buffered read or a defrag call
7517                          * triggered a readahead) on a page lock due to an
7518                          * ordered dio extent we created before but did not have
7519                          * yet a corresponding bio submitted (whence it can not
7520                          * complete), which makes readpages() wait for that
7521                          * ordered extent to complete while holding a lock on
7522                          * that page.
7523                          */
7524                         ret = -ENOTBLK;
7525                 }
7526
7527                 if (ret)
7528                         break;
7529
7530                 cond_resched();
7531         }
7532
7533         return ret;
7534 }
7535
7536 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7537                                            u64 len, u64 orig_start,
7538                                            u64 block_start, u64 block_len,
7539                                            u64 orig_block_len, u64 ram_bytes,
7540                                            int type)
7541 {
7542         struct extent_map_tree *em_tree;
7543         struct extent_map *em;
7544         struct btrfs_root *root = BTRFS_I(inode)->root;
7545         int ret;
7546
7547         em_tree = &BTRFS_I(inode)->extent_tree;
7548         em = alloc_extent_map();
7549         if (!em)
7550                 return ERR_PTR(-ENOMEM);
7551
7552         em->start = start;
7553         em->orig_start = orig_start;
7554         em->mod_start = start;
7555         em->mod_len = len;
7556         em->len = len;
7557         em->block_len = block_len;
7558         em->block_start = block_start;
7559         em->bdev = root->fs_info->fs_devices->latest_bdev;
7560         em->orig_block_len = orig_block_len;
7561         em->ram_bytes = ram_bytes;
7562         em->generation = -1;
7563         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7564         if (type == BTRFS_ORDERED_PREALLOC)
7565                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7566
7567         do {
7568                 btrfs_drop_extent_cache(inode, em->start,
7569                                 em->start + em->len - 1, 0);
7570                 write_lock(&em_tree->lock);
7571                 ret = add_extent_mapping(em_tree, em, 1);
7572                 write_unlock(&em_tree->lock);
7573         } while (ret == -EEXIST);
7574
7575         if (ret) {
7576                 free_extent_map(em);
7577                 return ERR_PTR(ret);
7578         }
7579
7580         return em;
7581 }
7582
7583 static void adjust_dio_outstanding_extents(struct inode *inode,
7584                                            struct btrfs_dio_data *dio_data,
7585                                            const u64 len)
7586 {
7587         unsigned num_extents;
7588
7589         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7590                                            BTRFS_MAX_EXTENT_SIZE);
7591         /*
7592          * If we have an outstanding_extents count still set then we're
7593          * within our reservation, otherwise we need to adjust our inode
7594          * counter appropriately.
7595          */
7596         if (dio_data->outstanding_extents) {
7597                 dio_data->outstanding_extents -= num_extents;
7598         } else {
7599                 spin_lock(&BTRFS_I(inode)->lock);
7600                 BTRFS_I(inode)->outstanding_extents += num_extents;
7601                 spin_unlock(&BTRFS_I(inode)->lock);
7602         }
7603 }
7604
7605 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7606                                    struct buffer_head *bh_result, int create)
7607 {
7608         struct extent_map *em;
7609         struct btrfs_root *root = BTRFS_I(inode)->root;
7610         struct extent_state *cached_state = NULL;
7611         struct btrfs_dio_data *dio_data = NULL;
7612         u64 start = iblock << inode->i_blkbits;
7613         u64 lockstart, lockend;
7614         u64 len = bh_result->b_size;
7615         int unlock_bits = EXTENT_LOCKED;
7616         int ret = 0;
7617
7618         if (create)
7619                 unlock_bits |= EXTENT_DIRTY;
7620         else
7621                 len = min_t(u64, len, root->sectorsize);
7622
7623         lockstart = start;
7624         lockend = start + len - 1;
7625
7626         if (current->journal_info) {
7627                 /*
7628                  * Need to pull our outstanding extents and set journal_info to NULL so
7629                  * that anything that needs to check if there's a transaction doesn't get
7630                  * confused.
7631                  */
7632                 dio_data = current->journal_info;
7633                 current->journal_info = NULL;
7634         }
7635
7636         /*
7637          * If this errors out it's because we couldn't invalidate pagecache for
7638          * this range and we need to fallback to buffered.
7639          */
7640         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7641                                create)) {
7642                 ret = -ENOTBLK;
7643                 goto err;
7644         }
7645
7646         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7647         if (IS_ERR(em)) {
7648                 ret = PTR_ERR(em);
7649                 goto unlock_err;
7650         }
7651
7652         /*
7653          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7654          * io.  INLINE is special, and we could probably kludge it in here, but
7655          * it's still buffered so for safety lets just fall back to the generic
7656          * buffered path.
7657          *
7658          * For COMPRESSED we _have_ to read the entire extent in so we can
7659          * decompress it, so there will be buffering required no matter what we
7660          * do, so go ahead and fallback to buffered.
7661          *
7662          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7663          * to buffered IO.  Don't blame me, this is the price we pay for using
7664          * the generic code.
7665          */
7666         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7667             em->block_start == EXTENT_MAP_INLINE) {
7668                 free_extent_map(em);
7669                 ret = -ENOTBLK;
7670                 goto unlock_err;
7671         }
7672
7673         /* Just a good old fashioned hole, return */
7674         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7675                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7676                 free_extent_map(em);
7677                 goto unlock_err;
7678         }
7679
7680         /*
7681          * We don't allocate a new extent in the following cases
7682          *
7683          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7684          * existing extent.
7685          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7686          * just use the extent.
7687          *
7688          */
7689         if (!create) {
7690                 len = min(len, em->len - (start - em->start));
7691                 lockstart = start + len;
7692                 goto unlock;
7693         }
7694
7695         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7696             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7697              em->block_start != EXTENT_MAP_HOLE)) {
7698                 int type;
7699                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7700
7701                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7702                         type = BTRFS_ORDERED_PREALLOC;
7703                 else
7704                         type = BTRFS_ORDERED_NOCOW;
7705                 len = min(len, em->len - (start - em->start));
7706                 block_start = em->block_start + (start - em->start);
7707
7708                 if (can_nocow_extent(inode, start, &len, &orig_start,
7709                                      &orig_block_len, &ram_bytes) == 1 &&
7710                     btrfs_inc_nocow_writers(root->fs_info, block_start)) {
7711                         struct extent_map *em2;
7712
7713                         em2 = btrfs_create_dio_extent(inode, start, len,
7714                                                       orig_start, block_start,
7715                                                       len, orig_block_len,
7716                                                       ram_bytes, type);
7717                         btrfs_dec_nocow_writers(root->fs_info, block_start);
7718                         if (type == BTRFS_ORDERED_PREALLOC) {
7719                                 free_extent_map(em);
7720                                 em = em2;
7721                         }
7722                         if (em2 && IS_ERR(em2)) {
7723                                 ret = PTR_ERR(em2);
7724                                 goto unlock_err;
7725                         }
7726                         goto unlock;
7727                 }
7728         }
7729
7730         /*
7731          * this will cow the extent, reset the len in case we changed
7732          * it above
7733          */
7734         len = bh_result->b_size;
7735         free_extent_map(em);
7736         em = btrfs_new_extent_direct(inode, start, len);
7737         if (IS_ERR(em)) {
7738                 ret = PTR_ERR(em);
7739                 goto unlock_err;
7740         }
7741         len = min(len, em->len - (start - em->start));
7742 unlock:
7743         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7744                 inode->i_blkbits;
7745         bh_result->b_size = len;
7746         bh_result->b_bdev = em->bdev;
7747         set_buffer_mapped(bh_result);
7748         if (create) {
7749                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7750                         set_buffer_new(bh_result);
7751
7752                 /*
7753                  * Need to update the i_size under the extent lock so buffered
7754                  * readers will get the updated i_size when we unlock.
7755                  */
7756                 if (start + len > i_size_read(inode))
7757                         i_size_write(inode, start + len);
7758
7759                 adjust_dio_outstanding_extents(inode, dio_data, len);
7760                 btrfs_free_reserved_data_space(inode, start, len);
7761                 WARN_ON(dio_data->reserve < len);
7762                 dio_data->reserve -= len;
7763                 dio_data->unsubmitted_oe_range_end = start + len;
7764                 current->journal_info = dio_data;
7765         }
7766
7767         /*
7768          * In the case of write we need to clear and unlock the entire range,
7769          * in the case of read we need to unlock only the end area that we
7770          * aren't using if there is any left over space.
7771          */
7772         if (lockstart < lockend) {
7773                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7774                                  lockend, unlock_bits, 1, 0,
7775                                  &cached_state, GFP_NOFS);
7776         } else {
7777                 free_extent_state(cached_state);
7778         }
7779
7780         free_extent_map(em);
7781
7782         return 0;
7783
7784 unlock_err:
7785         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7786                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7787 err:
7788         if (dio_data)
7789                 current->journal_info = dio_data;
7790         /*
7791          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7792          * write less data then expected, so that we don't underflow our inode's
7793          * outstanding extents counter.
7794          */
7795         if (create && dio_data)
7796                 adjust_dio_outstanding_extents(inode, dio_data, len);
7797
7798         return ret;
7799 }
7800
7801 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7802                                         int rw, int mirror_num)
7803 {
7804         struct btrfs_root *root = BTRFS_I(inode)->root;
7805         int ret;
7806
7807         BUG_ON(rw & REQ_WRITE);
7808
7809         bio_get(bio);
7810
7811         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7812                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7813         if (ret)
7814                 goto err;
7815
7816         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7817 err:
7818         bio_put(bio);
7819         return ret;
7820 }
7821
7822 static int btrfs_check_dio_repairable(struct inode *inode,
7823                                       struct bio *failed_bio,
7824                                       struct io_failure_record *failrec,
7825                                       int failed_mirror)
7826 {
7827         int num_copies;
7828
7829         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7830                                       failrec->logical, failrec->len);
7831         if (num_copies == 1) {
7832                 /*
7833                  * we only have a single copy of the data, so don't bother with
7834                  * all the retry and error correction code that follows. no
7835                  * matter what the error is, it is very likely to persist.
7836                  */
7837                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7838                          num_copies, failrec->this_mirror, failed_mirror);
7839                 return 0;
7840         }
7841
7842         failrec->failed_mirror = failed_mirror;
7843         failrec->this_mirror++;
7844         if (failrec->this_mirror == failed_mirror)
7845                 failrec->this_mirror++;
7846
7847         if (failrec->this_mirror > num_copies) {
7848                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7849                          num_copies, failrec->this_mirror, failed_mirror);
7850                 return 0;
7851         }
7852
7853         return 1;
7854 }
7855
7856 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7857                         struct page *page, unsigned int pgoff,
7858                         u64 start, u64 end, int failed_mirror,
7859                         bio_end_io_t *repair_endio, void *repair_arg)
7860 {
7861         struct io_failure_record *failrec;
7862         struct bio *bio;
7863         int isector;
7864         int read_mode;
7865         int ret;
7866
7867         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7868
7869         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7870         if (ret)
7871                 return ret;
7872
7873         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7874                                          failed_mirror);
7875         if (!ret) {
7876                 free_io_failure(inode, failrec);
7877                 return -EIO;
7878         }
7879
7880         if ((failed_bio->bi_vcnt > 1)
7881                 || (failed_bio->bi_io_vec->bv_len
7882                         > BTRFS_I(inode)->root->sectorsize))
7883                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7884         else
7885                 read_mode = READ_SYNC;
7886
7887         isector = start - btrfs_io_bio(failed_bio)->logical;
7888         isector >>= inode->i_sb->s_blocksize_bits;
7889         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7890                                 pgoff, isector, repair_endio, repair_arg);
7891         if (!bio) {
7892                 free_io_failure(inode, failrec);
7893                 return -EIO;
7894         }
7895
7896         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7897                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7898                     read_mode, failrec->this_mirror, failrec->in_validation);
7899
7900         ret = submit_dio_repair_bio(inode, bio, read_mode,
7901                                     failrec->this_mirror);
7902         if (ret) {
7903                 free_io_failure(inode, failrec);
7904                 bio_put(bio);
7905         }
7906
7907         return ret;
7908 }
7909
7910 struct btrfs_retry_complete {
7911         struct completion done;
7912         struct inode *inode;
7913         u64 start;
7914         int uptodate;
7915 };
7916
7917 static void btrfs_retry_endio_nocsum(struct bio *bio)
7918 {
7919         struct btrfs_retry_complete *done = bio->bi_private;
7920         struct inode *inode;
7921         struct bio_vec *bvec;
7922         int i;
7923
7924         if (bio->bi_error)
7925                 goto end;
7926
7927         ASSERT(bio->bi_vcnt == 1);
7928         inode = bio->bi_io_vec->bv_page->mapping->host;
7929         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7930
7931         done->uptodate = 1;
7932         bio_for_each_segment_all(bvec, bio, i)
7933                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7934 end:
7935         complete(&done->done);
7936         bio_put(bio);
7937 }
7938
7939 static int __btrfs_correct_data_nocsum(struct inode *inode,
7940                                        struct btrfs_io_bio *io_bio)
7941 {
7942         struct btrfs_fs_info *fs_info;
7943         struct bio_vec *bvec;
7944         struct btrfs_retry_complete done;
7945         u64 start;
7946         unsigned int pgoff;
7947         u32 sectorsize;
7948         int nr_sectors;
7949         int i;
7950         int ret;
7951
7952         fs_info = BTRFS_I(inode)->root->fs_info;
7953         sectorsize = BTRFS_I(inode)->root->sectorsize;
7954
7955         start = io_bio->logical;
7956         done.inode = inode;
7957
7958         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7959                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
7960                 pgoff = bvec->bv_offset;
7961
7962 next_block_or_try_again:
7963                 done.uptodate = 0;
7964                 done.start = start;
7965                 init_completion(&done.done);
7966
7967                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
7968                                 pgoff, start, start + sectorsize - 1,
7969                                 io_bio->mirror_num,
7970                                 btrfs_retry_endio_nocsum, &done);
7971                 if (ret)
7972                         return ret;
7973
7974                 wait_for_completion(&done.done);
7975
7976                 if (!done.uptodate) {
7977                         /* We might have another mirror, so try again */
7978                         goto next_block_or_try_again;
7979                 }
7980
7981                 start += sectorsize;
7982
7983                 if (nr_sectors--) {
7984                         pgoff += sectorsize;
7985                         goto next_block_or_try_again;
7986                 }
7987         }
7988
7989         return 0;
7990 }
7991
7992 static void btrfs_retry_endio(struct bio *bio)
7993 {
7994         struct btrfs_retry_complete *done = bio->bi_private;
7995         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7996         struct inode *inode;
7997         struct bio_vec *bvec;
7998         u64 start;
7999         int uptodate;
8000         int ret;
8001         int i;
8002
8003         if (bio->bi_error)
8004                 goto end;
8005
8006         uptodate = 1;
8007
8008         start = done->start;
8009
8010         ASSERT(bio->bi_vcnt == 1);
8011         inode = bio->bi_io_vec->bv_page->mapping->host;
8012         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
8013
8014         bio_for_each_segment_all(bvec, bio, i) {
8015                 ret = __readpage_endio_check(done->inode, io_bio, i,
8016                                         bvec->bv_page, bvec->bv_offset,
8017                                         done->start, bvec->bv_len);
8018                 if (!ret)
8019                         clean_io_failure(done->inode, done->start,
8020                                         bvec->bv_page, bvec->bv_offset);
8021                 else
8022                         uptodate = 0;
8023         }
8024
8025         done->uptodate = uptodate;
8026 end:
8027         complete(&done->done);
8028         bio_put(bio);
8029 }
8030
8031 static int __btrfs_subio_endio_read(struct inode *inode,
8032                                     struct btrfs_io_bio *io_bio, int err)
8033 {
8034         struct btrfs_fs_info *fs_info;
8035         struct bio_vec *bvec;
8036         struct btrfs_retry_complete done;
8037         u64 start;
8038         u64 offset = 0;
8039         u32 sectorsize;
8040         int nr_sectors;
8041         unsigned int pgoff;
8042         int csum_pos;
8043         int i;
8044         int ret;
8045
8046         fs_info = BTRFS_I(inode)->root->fs_info;
8047         sectorsize = BTRFS_I(inode)->root->sectorsize;
8048
8049         err = 0;
8050         start = io_bio->logical;
8051         done.inode = inode;
8052
8053         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8054                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8055
8056                 pgoff = bvec->bv_offset;
8057 next_block:
8058                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8059                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8060                                         bvec->bv_page, pgoff, start,
8061                                         sectorsize);
8062                 if (likely(!ret))
8063                         goto next;
8064 try_again:
8065                 done.uptodate = 0;
8066                 done.start = start;
8067                 init_completion(&done.done);
8068
8069                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8070                                 pgoff, start, start + sectorsize - 1,
8071                                 io_bio->mirror_num,
8072                                 btrfs_retry_endio, &done);
8073                 if (ret) {
8074                         err = ret;
8075                         goto next;
8076                 }
8077
8078                 wait_for_completion(&done.done);
8079
8080                 if (!done.uptodate) {
8081                         /* We might have another mirror, so try again */
8082                         goto try_again;
8083                 }
8084 next:
8085                 offset += sectorsize;
8086                 start += sectorsize;
8087
8088                 ASSERT(nr_sectors);
8089
8090                 if (--nr_sectors) {
8091                         pgoff += sectorsize;
8092                         goto next_block;
8093                 }
8094         }
8095
8096         return err;
8097 }
8098
8099 static int btrfs_subio_endio_read(struct inode *inode,
8100                                   struct btrfs_io_bio *io_bio, int err)
8101 {
8102         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8103
8104         if (skip_csum) {
8105                 if (unlikely(err))
8106                         return __btrfs_correct_data_nocsum(inode, io_bio);
8107                 else
8108                         return 0;
8109         } else {
8110                 return __btrfs_subio_endio_read(inode, io_bio, err);
8111         }
8112 }
8113
8114 static void btrfs_endio_direct_read(struct bio *bio)
8115 {
8116         struct btrfs_dio_private *dip = bio->bi_private;
8117         struct inode *inode = dip->inode;
8118         struct bio *dio_bio;
8119         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8120         int err = bio->bi_error;
8121
8122         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8123                 err = btrfs_subio_endio_read(inode, io_bio, err);
8124
8125         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8126                       dip->logical_offset + dip->bytes - 1);
8127         dio_bio = dip->dio_bio;
8128
8129         kfree(dip);
8130
8131         dio_bio->bi_error = bio->bi_error;
8132         dio_end_io(dio_bio, bio->bi_error);
8133
8134         if (io_bio->end_io)
8135                 io_bio->end_io(io_bio, err);
8136         bio_put(bio);
8137 }
8138
8139 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8140                                                     const u64 offset,
8141                                                     const u64 bytes,
8142                                                     const int uptodate)
8143 {
8144         struct btrfs_root *root = BTRFS_I(inode)->root;
8145         struct btrfs_ordered_extent *ordered = NULL;
8146         u64 ordered_offset = offset;
8147         u64 ordered_bytes = bytes;
8148         int ret;
8149
8150 again:
8151         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8152                                                    &ordered_offset,
8153                                                    ordered_bytes,
8154                                                    uptodate);
8155         if (!ret)
8156                 goto out_test;
8157
8158         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8159                         finish_ordered_fn, NULL, NULL);
8160         btrfs_queue_work(root->fs_info->endio_write_workers,
8161                          &ordered->work);
8162 out_test:
8163         /*
8164          * our bio might span multiple ordered extents.  If we haven't
8165          * completed the accounting for the whole dio, go back and try again
8166          */
8167         if (ordered_offset < offset + bytes) {
8168                 ordered_bytes = offset + bytes - ordered_offset;
8169                 ordered = NULL;
8170                 goto again;
8171         }
8172 }
8173
8174 static void btrfs_endio_direct_write(struct bio *bio)
8175 {
8176         struct btrfs_dio_private *dip = bio->bi_private;
8177         struct bio *dio_bio = dip->dio_bio;
8178
8179         btrfs_endio_direct_write_update_ordered(dip->inode,
8180                                                 dip->logical_offset,
8181                                                 dip->bytes,
8182                                                 !bio->bi_error);
8183
8184         kfree(dip);
8185
8186         dio_bio->bi_error = bio->bi_error;
8187         dio_end_io(dio_bio, bio->bi_error);
8188         bio_put(bio);
8189 }
8190
8191 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8192                                     struct bio *bio, int mirror_num,
8193                                     unsigned long bio_flags, u64 offset)
8194 {
8195         int ret;
8196         struct btrfs_root *root = BTRFS_I(inode)->root;
8197         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8198         BUG_ON(ret); /* -ENOMEM */
8199         return 0;
8200 }
8201
8202 static void btrfs_end_dio_bio(struct bio *bio)
8203 {
8204         struct btrfs_dio_private *dip = bio->bi_private;
8205         int err = bio->bi_error;
8206
8207         if (err)
8208                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8209                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8210                            btrfs_ino(dip->inode), bio->bi_rw,
8211                            (unsigned long long)bio->bi_iter.bi_sector,
8212                            bio->bi_iter.bi_size, err);
8213
8214         if (dip->subio_endio)
8215                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8216
8217         if (err) {
8218                 dip->errors = 1;
8219
8220                 /*
8221                  * before atomic variable goto zero, we must make sure
8222                  * dip->errors is perceived to be set.
8223                  */
8224                 smp_mb__before_atomic();
8225         }
8226
8227         /* if there are more bios still pending for this dio, just exit */
8228         if (!atomic_dec_and_test(&dip->pending_bios))
8229                 goto out;
8230
8231         if (dip->errors) {
8232                 bio_io_error(dip->orig_bio);
8233         } else {
8234                 dip->dio_bio->bi_error = 0;
8235                 bio_endio(dip->orig_bio);
8236         }
8237 out:
8238         bio_put(bio);
8239 }
8240
8241 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8242                                        u64 first_sector, gfp_t gfp_flags)
8243 {
8244         struct bio *bio;
8245         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8246         if (bio)
8247                 bio_associate_current(bio);
8248         return bio;
8249 }
8250
8251 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8252                                                  struct inode *inode,
8253                                                  struct btrfs_dio_private *dip,
8254                                                  struct bio *bio,
8255                                                  u64 file_offset)
8256 {
8257         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8258         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8259         int ret;
8260
8261         /*
8262          * We load all the csum data we need when we submit
8263          * the first bio to reduce the csum tree search and
8264          * contention.
8265          */
8266         if (dip->logical_offset == file_offset) {
8267                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8268                                                 file_offset);
8269                 if (ret)
8270                         return ret;
8271         }
8272
8273         if (bio == dip->orig_bio)
8274                 return 0;
8275
8276         file_offset -= dip->logical_offset;
8277         file_offset >>= inode->i_sb->s_blocksize_bits;
8278         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8279
8280         return 0;
8281 }
8282
8283 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8284                                          int rw, u64 file_offset, int skip_sum,
8285                                          int async_submit)
8286 {
8287         struct btrfs_dio_private *dip = bio->bi_private;
8288         int write = rw & REQ_WRITE;
8289         struct btrfs_root *root = BTRFS_I(inode)->root;
8290         int ret;
8291
8292         if (async_submit)
8293                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8294
8295         bio_get(bio);
8296
8297         if (!write) {
8298                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8299                                 BTRFS_WQ_ENDIO_DATA);
8300                 if (ret)
8301                         goto err;
8302         }
8303
8304         if (skip_sum)
8305                 goto map;
8306
8307         if (write && async_submit) {
8308                 ret = btrfs_wq_submit_bio(root->fs_info,
8309                                    inode, rw, bio, 0, 0,
8310                                    file_offset,
8311                                    __btrfs_submit_bio_start_direct_io,
8312                                    __btrfs_submit_bio_done);
8313                 goto err;
8314         } else if (write) {
8315                 /*
8316                  * If we aren't doing async submit, calculate the csum of the
8317                  * bio now.
8318                  */
8319                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8320                 if (ret)
8321                         goto err;
8322         } else {
8323                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8324                                                      file_offset);
8325                 if (ret)
8326                         goto err;
8327         }
8328 map:
8329         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8330 err:
8331         bio_put(bio);
8332         return ret;
8333 }
8334
8335 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8336                                     int skip_sum)
8337 {
8338         struct inode *inode = dip->inode;
8339         struct btrfs_root *root = BTRFS_I(inode)->root;
8340         struct bio *bio;
8341         struct bio *orig_bio = dip->orig_bio;
8342         struct bio_vec *bvec = orig_bio->bi_io_vec;
8343         u64 start_sector = orig_bio->bi_iter.bi_sector;
8344         u64 file_offset = dip->logical_offset;
8345         u64 submit_len = 0;
8346         u64 map_length;
8347         u32 blocksize = root->sectorsize;
8348         int async_submit = 0;
8349         int nr_sectors;
8350         int ret;
8351         int i;
8352
8353         map_length = orig_bio->bi_iter.bi_size;
8354         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8355                               &map_length, NULL, 0);
8356         if (ret)
8357                 return -EIO;
8358
8359         if (map_length >= orig_bio->bi_iter.bi_size) {
8360                 bio = orig_bio;
8361                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8362                 goto submit;
8363         }
8364
8365         /* async crcs make it difficult to collect full stripe writes. */
8366         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8367                 async_submit = 0;
8368         else
8369                 async_submit = 1;
8370
8371         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8372         if (!bio)
8373                 return -ENOMEM;
8374
8375         bio->bi_private = dip;
8376         bio->bi_end_io = btrfs_end_dio_bio;
8377         btrfs_io_bio(bio)->logical = file_offset;
8378         atomic_inc(&dip->pending_bios);
8379
8380         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8381                 nr_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info, bvec->bv_len);
8382                 i = 0;
8383 next_block:
8384                 if (unlikely(map_length < submit_len + blocksize ||
8385                     bio_add_page(bio, bvec->bv_page, blocksize,
8386                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8387                         /*
8388                          * inc the count before we submit the bio so
8389                          * we know the end IO handler won't happen before
8390                          * we inc the count. Otherwise, the dip might get freed
8391                          * before we're done setting it up
8392                          */
8393                         atomic_inc(&dip->pending_bios);
8394                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8395                                                      file_offset, skip_sum,
8396                                                      async_submit);
8397                         if (ret) {
8398                                 bio_put(bio);
8399                                 atomic_dec(&dip->pending_bios);
8400                                 goto out_err;
8401                         }
8402
8403                         start_sector += submit_len >> 9;
8404                         file_offset += submit_len;
8405
8406                         submit_len = 0;
8407
8408                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8409                                                   start_sector, GFP_NOFS);
8410                         if (!bio)
8411                                 goto out_err;
8412                         bio->bi_private = dip;
8413                         bio->bi_end_io = btrfs_end_dio_bio;
8414                         btrfs_io_bio(bio)->logical = file_offset;
8415
8416                         map_length = orig_bio->bi_iter.bi_size;
8417                         ret = btrfs_map_block(root->fs_info, rw,
8418                                               start_sector << 9,
8419                                               &map_length, NULL, 0);
8420                         if (ret) {
8421                                 bio_put(bio);
8422                                 goto out_err;
8423                         }
8424
8425                         goto next_block;
8426                 } else {
8427                         submit_len += blocksize;
8428                         if (--nr_sectors) {
8429                                 i++;
8430                                 goto next_block;
8431                         }
8432                         bvec++;
8433                 }
8434         }
8435
8436 submit:
8437         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8438                                      async_submit);
8439         if (!ret)
8440                 return 0;
8441
8442         bio_put(bio);
8443 out_err:
8444         dip->errors = 1;
8445         /*
8446          * before atomic variable goto zero, we must
8447          * make sure dip->errors is perceived to be set.
8448          */
8449         smp_mb__before_atomic();
8450         if (atomic_dec_and_test(&dip->pending_bios))
8451                 bio_io_error(dip->orig_bio);
8452
8453         /* bio_end_io() will handle error, so we needn't return it */
8454         return 0;
8455 }
8456
8457 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8458                                 struct inode *inode, loff_t file_offset)
8459 {
8460         struct btrfs_dio_private *dip = NULL;
8461         struct bio *io_bio = NULL;
8462         struct btrfs_io_bio *btrfs_bio;
8463         int skip_sum;
8464         int write = rw & REQ_WRITE;
8465         int ret = 0;
8466
8467         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8468
8469         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8470         if (!io_bio) {
8471                 ret = -ENOMEM;
8472                 goto free_ordered;
8473         }
8474
8475         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8476         if (!dip) {
8477                 ret = -ENOMEM;
8478                 goto free_ordered;
8479         }
8480
8481         dip->private = dio_bio->bi_private;
8482         dip->inode = inode;
8483         dip->logical_offset = file_offset;
8484         dip->bytes = dio_bio->bi_iter.bi_size;
8485         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8486         io_bio->bi_private = dip;
8487         dip->orig_bio = io_bio;
8488         dip->dio_bio = dio_bio;
8489         atomic_set(&dip->pending_bios, 0);
8490         btrfs_bio = btrfs_io_bio(io_bio);
8491         btrfs_bio->logical = file_offset;
8492
8493         if (write) {
8494                 io_bio->bi_end_io = btrfs_endio_direct_write;
8495         } else {
8496                 io_bio->bi_end_io = btrfs_endio_direct_read;
8497                 dip->subio_endio = btrfs_subio_endio_read;
8498         }
8499
8500         /*
8501          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8502          * even if we fail to submit a bio, because in such case we do the
8503          * corresponding error handling below and it must not be done a second
8504          * time by btrfs_direct_IO().
8505          */
8506         if (write) {
8507                 struct btrfs_dio_data *dio_data = current->journal_info;
8508
8509                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8510                         dip->bytes;
8511                 dio_data->unsubmitted_oe_range_start =
8512                         dio_data->unsubmitted_oe_range_end;
8513         }
8514
8515         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8516         if (!ret)
8517                 return;
8518
8519         if (btrfs_bio->end_io)
8520                 btrfs_bio->end_io(btrfs_bio, ret);
8521
8522 free_ordered:
8523         /*
8524          * If we arrived here it means either we failed to submit the dip
8525          * or we either failed to clone the dio_bio or failed to allocate the
8526          * dip. If we cloned the dio_bio and allocated the dip, we can just
8527          * call bio_endio against our io_bio so that we get proper resource
8528          * cleanup if we fail to submit the dip, otherwise, we must do the
8529          * same as btrfs_endio_direct_[write|read] because we can't call these
8530          * callbacks - they require an allocated dip and a clone of dio_bio.
8531          */
8532         if (io_bio && dip) {
8533                 io_bio->bi_error = -EIO;
8534                 bio_endio(io_bio);
8535                 /*
8536                  * The end io callbacks free our dip, do the final put on io_bio
8537                  * and all the cleanup and final put for dio_bio (through
8538                  * dio_end_io()).
8539                  */
8540                 dip = NULL;
8541                 io_bio = NULL;
8542         } else {
8543                 if (write)
8544                         btrfs_endio_direct_write_update_ordered(inode,
8545                                                 file_offset,
8546                                                 dio_bio->bi_iter.bi_size,
8547                                                 0);
8548                 else
8549                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8550                               file_offset + dio_bio->bi_iter.bi_size - 1);
8551
8552                 dio_bio->bi_error = -EIO;
8553                 /*
8554                  * Releases and cleans up our dio_bio, no need to bio_put()
8555                  * nor bio_endio()/bio_io_error() against dio_bio.
8556                  */
8557                 dio_end_io(dio_bio, ret);
8558         }
8559         if (io_bio)
8560                 bio_put(io_bio);
8561         kfree(dip);
8562 }
8563
8564 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8565                         const struct iov_iter *iter, loff_t offset)
8566 {
8567         int seg;
8568         int i;
8569         unsigned blocksize_mask = root->sectorsize - 1;
8570         ssize_t retval = -EINVAL;
8571
8572         if (offset & blocksize_mask)
8573                 goto out;
8574
8575         if (iov_iter_alignment(iter) & blocksize_mask)
8576                 goto out;
8577
8578         /* If this is a write we don't need to check anymore */
8579         if (iov_iter_rw(iter) == WRITE)
8580                 return 0;
8581         /*
8582          * Check to make sure we don't have duplicate iov_base's in this
8583          * iovec, if so return EINVAL, otherwise we'll get csum errors
8584          * when reading back.
8585          */
8586         for (seg = 0; seg < iter->nr_segs; seg++) {
8587                 for (i = seg + 1; i < iter->nr_segs; i++) {
8588                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8589                                 goto out;
8590                 }
8591         }
8592         retval = 0;
8593 out:
8594         return retval;
8595 }
8596
8597 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8598 {
8599         struct file *file = iocb->ki_filp;
8600         struct inode *inode = file->f_mapping->host;
8601         struct btrfs_root *root = BTRFS_I(inode)->root;
8602         struct btrfs_dio_data dio_data = { 0 };
8603         loff_t offset = iocb->ki_pos;
8604         size_t count = 0;
8605         int flags = 0;
8606         bool wakeup = true;
8607         bool relock = false;
8608         ssize_t ret;
8609
8610         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8611                 return 0;
8612
8613         inode_dio_begin(inode);
8614         smp_mb__after_atomic();
8615
8616         /*
8617          * The generic stuff only does filemap_write_and_wait_range, which
8618          * isn't enough if we've written compressed pages to this area, so
8619          * we need to flush the dirty pages again to make absolutely sure
8620          * that any outstanding dirty pages are on disk.
8621          */
8622         count = iov_iter_count(iter);
8623         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8624                      &BTRFS_I(inode)->runtime_flags))
8625                 filemap_fdatawrite_range(inode->i_mapping, offset,
8626                                          offset + count - 1);
8627
8628         if (iov_iter_rw(iter) == WRITE) {
8629                 /*
8630                  * If the write DIO is beyond the EOF, we need update
8631                  * the isize, but it is protected by i_mutex. So we can
8632                  * not unlock the i_mutex at this case.
8633                  */
8634                 if (offset + count <= inode->i_size) {
8635                         inode_unlock(inode);
8636                         relock = true;
8637                 }
8638                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8639                 if (ret)
8640                         goto out;
8641                 dio_data.outstanding_extents = div64_u64(count +
8642                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8643                                                 BTRFS_MAX_EXTENT_SIZE);
8644
8645                 /*
8646                  * We need to know how many extents we reserved so that we can
8647                  * do the accounting properly if we go over the number we
8648                  * originally calculated.  Abuse current->journal_info for this.
8649                  */
8650                 dio_data.reserve = round_up(count, root->sectorsize);
8651                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8652                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8653                 current->journal_info = &dio_data;
8654         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8655                                      &BTRFS_I(inode)->runtime_flags)) {
8656                 inode_dio_end(inode);
8657                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8658                 wakeup = false;
8659         }
8660
8661         ret = __blockdev_direct_IO(iocb, inode,
8662                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8663                                    iter, btrfs_get_blocks_direct, NULL,
8664                                    btrfs_submit_direct, flags);
8665         if (iov_iter_rw(iter) == WRITE) {
8666                 current->journal_info = NULL;
8667                 if (ret < 0 && ret != -EIOCBQUEUED) {
8668                         if (dio_data.reserve)
8669                                 btrfs_delalloc_release_space(inode, offset,
8670                                                              dio_data.reserve);
8671                         /*
8672                          * On error we might have left some ordered extents
8673                          * without submitting corresponding bios for them, so
8674                          * cleanup them up to avoid other tasks getting them
8675                          * and waiting for them to complete forever.
8676                          */
8677                         if (dio_data.unsubmitted_oe_range_start <
8678                             dio_data.unsubmitted_oe_range_end)
8679                                 btrfs_endio_direct_write_update_ordered(inode,
8680                                         dio_data.unsubmitted_oe_range_start,
8681                                         dio_data.unsubmitted_oe_range_end -
8682                                         dio_data.unsubmitted_oe_range_start,
8683                                         0);
8684                 } else if (ret >= 0 && (size_t)ret < count)
8685                         btrfs_delalloc_release_space(inode, offset,
8686                                                      count - (size_t)ret);
8687         }
8688 out:
8689         if (wakeup)
8690                 inode_dio_end(inode);
8691         if (relock)
8692                 inode_lock(inode);
8693
8694         return ret;
8695 }
8696
8697 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8698
8699 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8700                 __u64 start, __u64 len)
8701 {
8702         int     ret;
8703
8704         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8705         if (ret)
8706                 return ret;
8707
8708         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8709 }
8710
8711 int btrfs_readpage(struct file *file, struct page *page)
8712 {
8713         struct extent_io_tree *tree;
8714         tree = &BTRFS_I(page->mapping->host)->io_tree;
8715         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8716 }
8717
8718 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8719 {
8720         struct extent_io_tree *tree;
8721         struct inode *inode = page->mapping->host;
8722         int ret;
8723
8724         if (current->flags & PF_MEMALLOC) {
8725                 redirty_page_for_writepage(wbc, page);
8726                 unlock_page(page);
8727                 return 0;
8728         }
8729
8730         /*
8731          * If we are under memory pressure we will call this directly from the
8732          * VM, we need to make sure we have the inode referenced for the ordered
8733          * extent.  If not just return like we didn't do anything.
8734          */
8735         if (!igrab(inode)) {
8736                 redirty_page_for_writepage(wbc, page);
8737                 return AOP_WRITEPAGE_ACTIVATE;
8738         }
8739         tree = &BTRFS_I(page->mapping->host)->io_tree;
8740         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8741         btrfs_add_delayed_iput(inode);
8742         return ret;
8743 }
8744
8745 static int btrfs_writepages(struct address_space *mapping,
8746                             struct writeback_control *wbc)
8747 {
8748         struct extent_io_tree *tree;
8749
8750         tree = &BTRFS_I(mapping->host)->io_tree;
8751         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8752 }
8753
8754 static int
8755 btrfs_readpages(struct file *file, struct address_space *mapping,
8756                 struct list_head *pages, unsigned nr_pages)
8757 {
8758         struct extent_io_tree *tree;
8759         tree = &BTRFS_I(mapping->host)->io_tree;
8760         return extent_readpages(tree, mapping, pages, nr_pages,
8761                                 btrfs_get_extent);
8762 }
8763 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8764 {
8765         struct extent_io_tree *tree;
8766         struct extent_map_tree *map;
8767         int ret;
8768
8769         tree = &BTRFS_I(page->mapping->host)->io_tree;
8770         map = &BTRFS_I(page->mapping->host)->extent_tree;
8771         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8772         if (ret == 1) {
8773                 ClearPagePrivate(page);
8774                 set_page_private(page, 0);
8775                 put_page(page);
8776         }
8777         return ret;
8778 }
8779
8780 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8781 {
8782         if (PageWriteback(page) || PageDirty(page))
8783                 return 0;
8784         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8785 }
8786
8787 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8788                                  unsigned int length)
8789 {
8790         struct inode *inode = page->mapping->host;
8791         struct extent_io_tree *tree;
8792         struct btrfs_ordered_extent *ordered;
8793         struct extent_state *cached_state = NULL;
8794         u64 page_start = page_offset(page);
8795         u64 page_end = page_start + PAGE_SIZE - 1;
8796         u64 start;
8797         u64 end;
8798         int inode_evicting = inode->i_state & I_FREEING;
8799
8800         /*
8801          * we have the page locked, so new writeback can't start,
8802          * and the dirty bit won't be cleared while we are here.
8803          *
8804          * Wait for IO on this page so that we can safely clear
8805          * the PagePrivate2 bit and do ordered accounting
8806          */
8807         wait_on_page_writeback(page);
8808
8809         tree = &BTRFS_I(inode)->io_tree;
8810         if (offset) {
8811                 btrfs_releasepage(page, GFP_NOFS);
8812                 return;
8813         }
8814
8815         if (!inode_evicting)
8816                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8817 again:
8818         start = page_start;
8819         ordered = btrfs_lookup_ordered_range(inode, start,
8820                                         page_end - start + 1);
8821         if (ordered) {
8822                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8823                 /*
8824                  * IO on this page will never be started, so we need
8825                  * to account for any ordered extents now
8826                  */
8827                 if (!inode_evicting)
8828                         clear_extent_bit(tree, start, end,
8829                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8830                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8831                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8832                                          GFP_NOFS);
8833                 /*
8834                  * whoever cleared the private bit is responsible
8835                  * for the finish_ordered_io
8836                  */
8837                 if (TestClearPagePrivate2(page)) {
8838                         struct btrfs_ordered_inode_tree *tree;
8839                         u64 new_len;
8840
8841                         tree = &BTRFS_I(inode)->ordered_tree;
8842
8843                         spin_lock_irq(&tree->lock);
8844                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8845                         new_len = start - ordered->file_offset;
8846                         if (new_len < ordered->truncated_len)
8847                                 ordered->truncated_len = new_len;
8848                         spin_unlock_irq(&tree->lock);
8849
8850                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8851                                                            start,
8852                                                            end - start + 1, 1))
8853                                 btrfs_finish_ordered_io(ordered);
8854                 }
8855                 btrfs_put_ordered_extent(ordered);
8856                 if (!inode_evicting) {
8857                         cached_state = NULL;
8858                         lock_extent_bits(tree, start, end,
8859                                          &cached_state);
8860                 }
8861
8862                 start = end + 1;
8863                 if (start < page_end)
8864                         goto again;
8865         }
8866
8867         /*
8868          * Qgroup reserved space handler
8869          * Page here will be either
8870          * 1) Already written to disk
8871          *    In this case, its reserved space is released from data rsv map
8872          *    and will be freed by delayed_ref handler finally.
8873          *    So even we call qgroup_free_data(), it won't decrease reserved
8874          *    space.
8875          * 2) Not written to disk
8876          *    This means the reserved space should be freed here.
8877          */
8878         btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
8879         if (!inode_evicting) {
8880                 clear_extent_bit(tree, page_start, page_end,
8881                                  EXTENT_LOCKED | EXTENT_DIRTY |
8882                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8883                                  EXTENT_DEFRAG, 1, 1,
8884                                  &cached_state, GFP_NOFS);
8885
8886                 __btrfs_releasepage(page, GFP_NOFS);
8887         }
8888
8889         ClearPageChecked(page);
8890         if (PagePrivate(page)) {
8891                 ClearPagePrivate(page);
8892                 set_page_private(page, 0);
8893                 put_page(page);
8894         }
8895 }
8896
8897 /*
8898  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8899  * called from a page fault handler when a page is first dirtied. Hence we must
8900  * be careful to check for EOF conditions here. We set the page up correctly
8901  * for a written page which means we get ENOSPC checking when writing into
8902  * holes and correct delalloc and unwritten extent mapping on filesystems that
8903  * support these features.
8904  *
8905  * We are not allowed to take the i_mutex here so we have to play games to
8906  * protect against truncate races as the page could now be beyond EOF.  Because
8907  * vmtruncate() writes the inode size before removing pages, once we have the
8908  * page lock we can determine safely if the page is beyond EOF. If it is not
8909  * beyond EOF, then the page is guaranteed safe against truncation until we
8910  * unlock the page.
8911  */
8912 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8913 {
8914         struct page *page = vmf->page;
8915         struct inode *inode = file_inode(vma->vm_file);
8916         struct btrfs_root *root = BTRFS_I(inode)->root;
8917         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8918         struct btrfs_ordered_extent *ordered;
8919         struct extent_state *cached_state = NULL;
8920         char *kaddr;
8921         unsigned long zero_start;
8922         loff_t size;
8923         int ret;
8924         int reserved = 0;
8925         u64 reserved_space;
8926         u64 page_start;
8927         u64 page_end;
8928         u64 end;
8929
8930         reserved_space = PAGE_SIZE;
8931
8932         sb_start_pagefault(inode->i_sb);
8933         page_start = page_offset(page);
8934         page_end = page_start + PAGE_SIZE - 1;
8935         end = page_end;
8936
8937         /*
8938          * Reserving delalloc space after obtaining the page lock can lead to
8939          * deadlock. For example, if a dirty page is locked by this function
8940          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8941          * dirty page write out, then the btrfs_writepage() function could
8942          * end up waiting indefinitely to get a lock on the page currently
8943          * being processed by btrfs_page_mkwrite() function.
8944          */
8945         ret = btrfs_delalloc_reserve_space(inode, page_start,
8946                                            reserved_space);
8947         if (!ret) {
8948                 ret = file_update_time(vma->vm_file);
8949                 reserved = 1;
8950         }
8951         if (ret) {
8952                 if (ret == -ENOMEM)
8953                         ret = VM_FAULT_OOM;
8954                 else /* -ENOSPC, -EIO, etc */
8955                         ret = VM_FAULT_SIGBUS;
8956                 if (reserved)
8957                         goto out;
8958                 goto out_noreserve;
8959         }
8960
8961         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8962 again:
8963         lock_page(page);
8964         size = i_size_read(inode);
8965
8966         if ((page->mapping != inode->i_mapping) ||
8967             (page_start >= size)) {
8968                 /* page got truncated out from underneath us */
8969                 goto out_unlock;
8970         }
8971         wait_on_page_writeback(page);
8972
8973         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8974         set_page_extent_mapped(page);
8975
8976         /*
8977          * we can't set the delalloc bits if there are pending ordered
8978          * extents.  Drop our locks and wait for them to finish
8979          */
8980         ordered = btrfs_lookup_ordered_range(inode, page_start, page_end);
8981         if (ordered) {
8982                 unlock_extent_cached(io_tree, page_start, page_end,
8983                                      &cached_state, GFP_NOFS);
8984                 unlock_page(page);
8985                 btrfs_start_ordered_extent(inode, ordered, 1);
8986                 btrfs_put_ordered_extent(ordered);
8987                 goto again;
8988         }
8989
8990         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8991                 reserved_space = round_up(size - page_start, root->sectorsize);
8992                 if (reserved_space < PAGE_SIZE) {
8993                         end = page_start + reserved_space - 1;
8994                         spin_lock(&BTRFS_I(inode)->lock);
8995                         BTRFS_I(inode)->outstanding_extents++;
8996                         spin_unlock(&BTRFS_I(inode)->lock);
8997                         btrfs_delalloc_release_space(inode, page_start,
8998                                                 PAGE_SIZE - reserved_space);
8999                 }
9000         }
9001
9002         /*
9003          * XXX - page_mkwrite gets called every time the page is dirtied, even
9004          * if it was already dirty, so for space accounting reasons we need to
9005          * clear any delalloc bits for the range we are fixing to save.  There
9006          * is probably a better way to do this, but for now keep consistent with
9007          * prepare_pages in the normal write path.
9008          */
9009         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9010                           EXTENT_DIRTY | EXTENT_DELALLOC |
9011                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9012                           0, 0, &cached_state, GFP_NOFS);
9013
9014         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9015                                         &cached_state);
9016         if (ret) {
9017                 unlock_extent_cached(io_tree, page_start, page_end,
9018                                      &cached_state, GFP_NOFS);
9019                 ret = VM_FAULT_SIGBUS;
9020                 goto out_unlock;
9021         }
9022         ret = 0;
9023
9024         /* page is wholly or partially inside EOF */
9025         if (page_start + PAGE_SIZE > size)
9026                 zero_start = size & ~PAGE_MASK;
9027         else
9028                 zero_start = PAGE_SIZE;
9029
9030         if (zero_start != PAGE_SIZE) {
9031                 kaddr = kmap(page);
9032                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9033                 flush_dcache_page(page);
9034                 kunmap(page);
9035         }
9036         ClearPageChecked(page);
9037         set_page_dirty(page);
9038         SetPageUptodate(page);
9039
9040         BTRFS_I(inode)->last_trans = root->fs_info->generation;
9041         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9042         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9043
9044         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9045
9046 out_unlock:
9047         if (!ret) {
9048                 sb_end_pagefault(inode->i_sb);
9049                 return VM_FAULT_LOCKED;
9050         }
9051         unlock_page(page);
9052 out:
9053         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9054 out_noreserve:
9055         sb_end_pagefault(inode->i_sb);
9056         return ret;
9057 }
9058
9059 static int btrfs_truncate(struct inode *inode)
9060 {
9061         struct btrfs_root *root = BTRFS_I(inode)->root;
9062         struct btrfs_block_rsv *rsv;
9063         int ret = 0;
9064         int err = 0;
9065         struct btrfs_trans_handle *trans;
9066         u64 mask = root->sectorsize - 1;
9067         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
9068
9069         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9070                                        (u64)-1);
9071         if (ret)
9072                 return ret;
9073
9074         /*
9075          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9076          * 3 things going on here
9077          *
9078          * 1) We need to reserve space for our orphan item and the space to
9079          * delete our orphan item.  Lord knows we don't want to have a dangling
9080          * orphan item because we didn't reserve space to remove it.
9081          *
9082          * 2) We need to reserve space to update our inode.
9083          *
9084          * 3) We need to have something to cache all the space that is going to
9085          * be free'd up by the truncate operation, but also have some slack
9086          * space reserved in case it uses space during the truncate (thank you
9087          * very much snapshotting).
9088          *
9089          * And we need these to all be separate.  The fact is we can use a lot of
9090          * space doing the truncate, and we have no earthly idea how much space
9091          * we will use, so we need the truncate reservation to be separate so it
9092          * doesn't end up using space reserved for updating the inode or
9093          * removing the orphan item.  We also need to be able to stop the
9094          * transaction and start a new one, which means we need to be able to
9095          * update the inode several times, and we have no idea of knowing how
9096          * many times that will be, so we can't just reserve 1 item for the
9097          * entirety of the operation, so that has to be done separately as well.
9098          * Then there is the orphan item, which does indeed need to be held on
9099          * to for the whole operation, and we need nobody to touch this reserved
9100          * space except the orphan code.
9101          *
9102          * So that leaves us with
9103          *
9104          * 1) root->orphan_block_rsv - for the orphan deletion.
9105          * 2) rsv - for the truncate reservation, which we will steal from the
9106          * transaction reservation.
9107          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9108          * updating the inode.
9109          */
9110         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
9111         if (!rsv)
9112                 return -ENOMEM;
9113         rsv->size = min_size;
9114         rsv->failfast = 1;
9115
9116         /*
9117          * 1 for the truncate slack space
9118          * 1 for updating the inode.
9119          */
9120         trans = btrfs_start_transaction(root, 2);
9121         if (IS_ERR(trans)) {
9122                 err = PTR_ERR(trans);
9123                 goto out;
9124         }
9125
9126         /* Migrate the slack space for the truncate to our reserve */
9127         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
9128                                       min_size, 0);
9129         BUG_ON(ret);
9130
9131         /*
9132          * So if we truncate and then write and fsync we normally would just
9133          * write the extents that changed, which is a problem if we need to
9134          * first truncate that entire inode.  So set this flag so we write out
9135          * all of the extents in the inode to the sync log so we're completely
9136          * safe.
9137          */
9138         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9139         trans->block_rsv = rsv;
9140
9141         while (1) {
9142                 ret = btrfs_truncate_inode_items(trans, root, inode,
9143                                                  inode->i_size,
9144                                                  BTRFS_EXTENT_DATA_KEY);
9145                 if (ret != -ENOSPC && ret != -EAGAIN) {
9146                         err = ret;
9147                         break;
9148                 }
9149
9150                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9151                 ret = btrfs_update_inode(trans, root, inode);
9152                 if (ret) {
9153                         err = ret;
9154                         break;
9155                 }
9156
9157                 btrfs_end_transaction(trans, root);
9158                 btrfs_btree_balance_dirty(root);
9159
9160                 trans = btrfs_start_transaction(root, 2);
9161                 if (IS_ERR(trans)) {
9162                         ret = err = PTR_ERR(trans);
9163                         trans = NULL;
9164                         break;
9165                 }
9166
9167                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
9168                                               rsv, min_size, 0);
9169                 BUG_ON(ret);    /* shouldn't happen */
9170                 trans->block_rsv = rsv;
9171         }
9172
9173         if (ret == 0 && inode->i_nlink > 0) {
9174                 trans->block_rsv = root->orphan_block_rsv;
9175                 ret = btrfs_orphan_del(trans, inode);
9176                 if (ret)
9177                         err = ret;
9178         }
9179
9180         if (trans) {
9181                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9182                 ret = btrfs_update_inode(trans, root, inode);
9183                 if (ret && !err)
9184                         err = ret;
9185
9186                 ret = btrfs_end_transaction(trans, root);
9187                 btrfs_btree_balance_dirty(root);
9188         }
9189 out:
9190         btrfs_free_block_rsv(root, rsv);
9191
9192         if (ret && !err)
9193                 err = ret;
9194
9195         return err;
9196 }
9197
9198 /*
9199  * create a new subvolume directory/inode (helper for the ioctl).
9200  */
9201 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9202                              struct btrfs_root *new_root,
9203                              struct btrfs_root *parent_root,
9204                              u64 new_dirid)
9205 {
9206         struct inode *inode;
9207         int err;
9208         u64 index = 0;
9209
9210         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9211                                 new_dirid, new_dirid,
9212                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9213                                 &index);
9214         if (IS_ERR(inode))
9215                 return PTR_ERR(inode);
9216         inode->i_op = &btrfs_dir_inode_operations;
9217         inode->i_fop = &btrfs_dir_file_operations;
9218
9219         set_nlink(inode, 1);
9220         btrfs_i_size_write(inode, 0);
9221         unlock_new_inode(inode);
9222
9223         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9224         if (err)
9225                 btrfs_err(new_root->fs_info,
9226                           "error inheriting subvolume %llu properties: %d",
9227                           new_root->root_key.objectid, err);
9228
9229         err = btrfs_update_inode(trans, new_root, inode);
9230
9231         iput(inode);
9232         return err;
9233 }
9234
9235 struct inode *btrfs_alloc_inode(struct super_block *sb)
9236 {
9237         struct btrfs_inode *ei;
9238         struct inode *inode;
9239
9240         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9241         if (!ei)
9242                 return NULL;
9243
9244         ei->root = NULL;
9245         ei->generation = 0;
9246         ei->last_trans = 0;
9247         ei->last_sub_trans = 0;
9248         ei->logged_trans = 0;
9249         ei->delalloc_bytes = 0;
9250         ei->defrag_bytes = 0;
9251         ei->disk_i_size = 0;
9252         ei->flags = 0;
9253         ei->csum_bytes = 0;
9254         ei->index_cnt = (u64)-1;
9255         ei->dir_index = 0;
9256         ei->last_unlink_trans = 0;
9257         ei->last_log_commit = 0;
9258         ei->delayed_iput_count = 0;
9259
9260         spin_lock_init(&ei->lock);
9261         ei->outstanding_extents = 0;
9262         ei->reserved_extents = 0;
9263
9264         ei->runtime_flags = 0;
9265         ei->force_compress = BTRFS_COMPRESS_NONE;
9266
9267         ei->delayed_node = NULL;
9268
9269         ei->i_otime.tv_sec = 0;
9270         ei->i_otime.tv_nsec = 0;
9271
9272         inode = &ei->vfs_inode;
9273         extent_map_tree_init(&ei->extent_tree);
9274         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9275         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9276         ei->io_tree.track_uptodate = 1;
9277         ei->io_failure_tree.track_uptodate = 1;
9278         atomic_set(&ei->sync_writers, 0);
9279         mutex_init(&ei->log_mutex);
9280         mutex_init(&ei->delalloc_mutex);
9281         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9282         INIT_LIST_HEAD(&ei->delalloc_inodes);
9283         INIT_LIST_HEAD(&ei->delayed_iput);
9284         RB_CLEAR_NODE(&ei->rb_node);
9285         init_rwsem(&ei->dio_sem);
9286
9287         return inode;
9288 }
9289
9290 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9291 void btrfs_test_destroy_inode(struct inode *inode)
9292 {
9293         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9294         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9295 }
9296 #endif
9297
9298 static void btrfs_i_callback(struct rcu_head *head)
9299 {
9300         struct inode *inode = container_of(head, struct inode, i_rcu);
9301         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9302 }
9303
9304 void btrfs_destroy_inode(struct inode *inode)
9305 {
9306         struct btrfs_ordered_extent *ordered;
9307         struct btrfs_root *root = BTRFS_I(inode)->root;
9308
9309         WARN_ON(!hlist_empty(&inode->i_dentry));
9310         WARN_ON(inode->i_data.nrpages);
9311         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9312         WARN_ON(BTRFS_I(inode)->reserved_extents);
9313         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9314         WARN_ON(BTRFS_I(inode)->csum_bytes);
9315         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9316
9317         /*
9318          * This can happen where we create an inode, but somebody else also
9319          * created the same inode and we need to destroy the one we already
9320          * created.
9321          */
9322         if (!root)
9323                 goto free;
9324
9325         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9326                      &BTRFS_I(inode)->runtime_flags)) {
9327                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9328                         btrfs_ino(inode));
9329                 atomic_dec(&root->orphan_inodes);
9330         }
9331
9332         while (1) {
9333                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9334                 if (!ordered)
9335                         break;
9336                 else {
9337                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9338                                 ordered->file_offset, ordered->len);
9339                         btrfs_remove_ordered_extent(inode, ordered);
9340                         btrfs_put_ordered_extent(ordered);
9341                         btrfs_put_ordered_extent(ordered);
9342                 }
9343         }
9344         btrfs_qgroup_check_reserved_leak(inode);
9345         inode_tree_del(inode);
9346         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9347 free:
9348         call_rcu(&inode->i_rcu, btrfs_i_callback);
9349 }
9350
9351 int btrfs_drop_inode(struct inode *inode)
9352 {
9353         struct btrfs_root *root = BTRFS_I(inode)->root;
9354
9355         if (root == NULL)
9356                 return 1;
9357
9358         /* the snap/subvol tree is on deleting */
9359         if (btrfs_root_refs(&root->root_item) == 0)
9360                 return 1;
9361         else
9362                 return generic_drop_inode(inode);
9363 }
9364
9365 static void init_once(void *foo)
9366 {
9367         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9368
9369         inode_init_once(&ei->vfs_inode);
9370 }
9371
9372 void btrfs_destroy_cachep(void)
9373 {
9374         /*
9375          * Make sure all delayed rcu free inodes are flushed before we
9376          * destroy cache.
9377          */
9378         rcu_barrier();
9379         kmem_cache_destroy(btrfs_inode_cachep);
9380         kmem_cache_destroy(btrfs_trans_handle_cachep);
9381         kmem_cache_destroy(btrfs_transaction_cachep);
9382         kmem_cache_destroy(btrfs_path_cachep);
9383         kmem_cache_destroy(btrfs_free_space_cachep);
9384 }
9385
9386 int btrfs_init_cachep(void)
9387 {
9388         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9389                         sizeof(struct btrfs_inode), 0,
9390                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9391                         init_once);
9392         if (!btrfs_inode_cachep)
9393                 goto fail;
9394
9395         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9396                         sizeof(struct btrfs_trans_handle), 0,
9397                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9398         if (!btrfs_trans_handle_cachep)
9399                 goto fail;
9400
9401         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9402                         sizeof(struct btrfs_transaction), 0,
9403                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9404         if (!btrfs_transaction_cachep)
9405                 goto fail;
9406
9407         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9408                         sizeof(struct btrfs_path), 0,
9409                         SLAB_MEM_SPREAD, NULL);
9410         if (!btrfs_path_cachep)
9411                 goto fail;
9412
9413         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9414                         sizeof(struct btrfs_free_space), 0,
9415                         SLAB_MEM_SPREAD, NULL);
9416         if (!btrfs_free_space_cachep)
9417                 goto fail;
9418
9419         return 0;
9420 fail:
9421         btrfs_destroy_cachep();
9422         return -ENOMEM;
9423 }
9424
9425 static int btrfs_getattr(struct vfsmount *mnt,
9426                          struct dentry *dentry, struct kstat *stat)
9427 {
9428         u64 delalloc_bytes;
9429         struct inode *inode = d_inode(dentry);
9430         u32 blocksize = inode->i_sb->s_blocksize;
9431
9432         generic_fillattr(inode, stat);
9433         stat->dev = BTRFS_I(inode)->root->anon_dev;
9434
9435         spin_lock(&BTRFS_I(inode)->lock);
9436         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9437         spin_unlock(&BTRFS_I(inode)->lock);
9438         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9439                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9440         return 0;
9441 }
9442
9443 static int btrfs_rename_exchange(struct inode *old_dir,
9444                               struct dentry *old_dentry,
9445                               struct inode *new_dir,
9446                               struct dentry *new_dentry)
9447 {
9448         struct btrfs_trans_handle *trans;
9449         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9450         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9451         struct inode *new_inode = new_dentry->d_inode;
9452         struct inode *old_inode = old_dentry->d_inode;
9453         struct timespec ctime = CURRENT_TIME;
9454         struct dentry *parent;
9455         u64 old_ino = btrfs_ino(old_inode);
9456         u64 new_ino = btrfs_ino(new_inode);
9457         u64 old_idx = 0;
9458         u64 new_idx = 0;
9459         u64 root_objectid;
9460         int ret;
9461         bool root_log_pinned = false;
9462         bool dest_log_pinned = false;
9463
9464         /* we only allow rename subvolume link between subvolumes */
9465         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9466                 return -EXDEV;
9467
9468         /* close the race window with snapshot create/destroy ioctl */
9469         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9470                 down_read(&root->fs_info->subvol_sem);
9471         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9472                 down_read(&dest->fs_info->subvol_sem);
9473
9474         /*
9475          * We want to reserve the absolute worst case amount of items.  So if
9476          * both inodes are subvols and we need to unlink them then that would
9477          * require 4 item modifications, but if they are both normal inodes it
9478          * would require 5 item modifications, so we'll assume their normal
9479          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9480          * should cover the worst case number of items we'll modify.
9481          */
9482         trans = btrfs_start_transaction(root, 12);
9483         if (IS_ERR(trans)) {
9484                 ret = PTR_ERR(trans);
9485                 goto out_notrans;
9486         }
9487
9488         /*
9489          * We need to find a free sequence number both in the source and
9490          * in the destination directory for the exchange.
9491          */
9492         ret = btrfs_set_inode_index(new_dir, &old_idx);
9493         if (ret)
9494                 goto out_fail;
9495         ret = btrfs_set_inode_index(old_dir, &new_idx);
9496         if (ret)
9497                 goto out_fail;
9498
9499         BTRFS_I(old_inode)->dir_index = 0ULL;
9500         BTRFS_I(new_inode)->dir_index = 0ULL;
9501
9502         /* Reference for the source. */
9503         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9504                 /* force full log commit if subvolume involved. */
9505                 btrfs_set_log_full_commit(root->fs_info, trans);
9506         } else {
9507                 btrfs_pin_log_trans(root);
9508                 root_log_pinned = true;
9509                 ret = btrfs_insert_inode_ref(trans, dest,
9510                                              new_dentry->d_name.name,
9511                                              new_dentry->d_name.len,
9512                                              old_ino,
9513                                              btrfs_ino(new_dir), old_idx);
9514                 if (ret)
9515                         goto out_fail;
9516         }
9517
9518         /* And now for the dest. */
9519         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9520                 /* force full log commit if subvolume involved. */
9521                 btrfs_set_log_full_commit(dest->fs_info, trans);
9522         } else {
9523                 btrfs_pin_log_trans(dest);
9524                 dest_log_pinned = true;
9525                 ret = btrfs_insert_inode_ref(trans, root,
9526                                              old_dentry->d_name.name,
9527                                              old_dentry->d_name.len,
9528                                              new_ino,
9529                                              btrfs_ino(old_dir), new_idx);
9530                 if (ret)
9531                         goto out_fail;
9532         }
9533
9534         /* Update inode version and ctime/mtime. */
9535         inode_inc_iversion(old_dir);
9536         inode_inc_iversion(new_dir);
9537         inode_inc_iversion(old_inode);
9538         inode_inc_iversion(new_inode);
9539         old_dir->i_ctime = old_dir->i_mtime = ctime;
9540         new_dir->i_ctime = new_dir->i_mtime = ctime;
9541         old_inode->i_ctime = ctime;
9542         new_inode->i_ctime = ctime;
9543
9544         if (old_dentry->d_parent != new_dentry->d_parent) {
9545                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9546                 btrfs_record_unlink_dir(trans, new_dir, new_inode, 1);
9547         }
9548
9549         /* src is a subvolume */
9550         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9551                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9552                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9553                                           root_objectid,
9554                                           old_dentry->d_name.name,
9555                                           old_dentry->d_name.len);
9556         } else { /* src is an inode */
9557                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9558                                            old_dentry->d_inode,
9559                                            old_dentry->d_name.name,
9560                                            old_dentry->d_name.len);
9561                 if (!ret)
9562                         ret = btrfs_update_inode(trans, root, old_inode);
9563         }
9564         if (ret) {
9565                 btrfs_abort_transaction(trans, root, ret);
9566                 goto out_fail;
9567         }
9568
9569         /* dest is a subvolume */
9570         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9571                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9572                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9573                                           root_objectid,
9574                                           new_dentry->d_name.name,
9575                                           new_dentry->d_name.len);
9576         } else { /* dest is an inode */
9577                 ret = __btrfs_unlink_inode(trans, dest, new_dir,
9578                                            new_dentry->d_inode,
9579                                            new_dentry->d_name.name,
9580                                            new_dentry->d_name.len);
9581                 if (!ret)
9582                         ret = btrfs_update_inode(trans, dest, new_inode);
9583         }
9584         if (ret) {
9585                 btrfs_abort_transaction(trans, root, ret);
9586                 goto out_fail;
9587         }
9588
9589         ret = btrfs_add_link(trans, new_dir, old_inode,
9590                              new_dentry->d_name.name,
9591                              new_dentry->d_name.len, 0, old_idx);
9592         if (ret) {
9593                 btrfs_abort_transaction(trans, root, ret);
9594                 goto out_fail;
9595         }
9596
9597         ret = btrfs_add_link(trans, old_dir, new_inode,
9598                              old_dentry->d_name.name,
9599                              old_dentry->d_name.len, 0, new_idx);
9600         if (ret) {
9601                 btrfs_abort_transaction(trans, root, ret);
9602                 goto out_fail;
9603         }
9604
9605         if (old_inode->i_nlink == 1)
9606                 BTRFS_I(old_inode)->dir_index = old_idx;
9607         if (new_inode->i_nlink == 1)
9608                 BTRFS_I(new_inode)->dir_index = new_idx;
9609
9610         if (root_log_pinned) {
9611                 parent = new_dentry->d_parent;
9612                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9613                 btrfs_end_log_trans(root);
9614                 root_log_pinned = false;
9615         }
9616         if (dest_log_pinned) {
9617                 parent = old_dentry->d_parent;
9618                 btrfs_log_new_name(trans, new_inode, new_dir, parent);
9619                 btrfs_end_log_trans(dest);
9620                 dest_log_pinned = false;
9621         }
9622 out_fail:
9623         /*
9624          * If we have pinned a log and an error happened, we unpin tasks
9625          * trying to sync the log and force them to fallback to a transaction
9626          * commit if the log currently contains any of the inodes involved in
9627          * this rename operation (to ensure we do not persist a log with an
9628          * inconsistent state for any of these inodes or leading to any
9629          * inconsistencies when replayed). If the transaction was aborted, the
9630          * abortion reason is propagated to userspace when attempting to commit
9631          * the transaction. If the log does not contain any of these inodes, we
9632          * allow the tasks to sync it.
9633          */
9634         if (ret && (root_log_pinned || dest_log_pinned)) {
9635                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9636                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9637                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9638                     (new_inode &&
9639                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9640                     btrfs_set_log_full_commit(root->fs_info, trans);
9641
9642                 if (root_log_pinned) {
9643                         btrfs_end_log_trans(root);
9644                         root_log_pinned = false;
9645                 }
9646                 if (dest_log_pinned) {
9647                         btrfs_end_log_trans(dest);
9648                         dest_log_pinned = false;
9649                 }
9650         }
9651         ret = btrfs_end_transaction(trans, root);
9652 out_notrans:
9653         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9654                 up_read(&dest->fs_info->subvol_sem);
9655         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9656                 up_read(&root->fs_info->subvol_sem);
9657
9658         return ret;
9659 }
9660
9661 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9662                                      struct btrfs_root *root,
9663                                      struct inode *dir,
9664                                      struct dentry *dentry)
9665 {
9666         int ret;
9667         struct inode *inode;
9668         u64 objectid;
9669         u64 index;
9670
9671         ret = btrfs_find_free_ino(root, &objectid);
9672         if (ret)
9673                 return ret;
9674
9675         inode = btrfs_new_inode(trans, root, dir,
9676                                 dentry->d_name.name,
9677                                 dentry->d_name.len,
9678                                 btrfs_ino(dir),
9679                                 objectid,
9680                                 S_IFCHR | WHITEOUT_MODE,
9681                                 &index);
9682
9683         if (IS_ERR(inode)) {
9684                 ret = PTR_ERR(inode);
9685                 return ret;
9686         }
9687
9688         inode->i_op = &btrfs_special_inode_operations;
9689         init_special_inode(inode, inode->i_mode,
9690                 WHITEOUT_DEV);
9691
9692         ret = btrfs_init_inode_security(trans, inode, dir,
9693                                 &dentry->d_name);
9694         if (ret)
9695                 goto out;
9696
9697         ret = btrfs_add_nondir(trans, dir, dentry,
9698                                 inode, 0, index);
9699         if (ret)
9700                 goto out;
9701
9702         ret = btrfs_update_inode(trans, root, inode);
9703 out:
9704         unlock_new_inode(inode);
9705         if (ret)
9706                 inode_dec_link_count(inode);
9707         iput(inode);
9708
9709         return ret;
9710 }
9711
9712 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9713                            struct inode *new_dir, struct dentry *new_dentry,
9714                            unsigned int flags)
9715 {
9716         struct btrfs_trans_handle *trans;
9717         unsigned int trans_num_items;
9718         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9719         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9720         struct inode *new_inode = d_inode(new_dentry);
9721         struct inode *old_inode = d_inode(old_dentry);
9722         u64 index = 0;
9723         u64 root_objectid;
9724         int ret;
9725         u64 old_ino = btrfs_ino(old_inode);
9726         bool log_pinned = false;
9727
9728         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9729                 return -EPERM;
9730
9731         /* we only allow rename subvolume link between subvolumes */
9732         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9733                 return -EXDEV;
9734
9735         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9736             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9737                 return -ENOTEMPTY;
9738
9739         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9740             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9741                 return -ENOTEMPTY;
9742
9743
9744         /* check for collisions, even if the  name isn't there */
9745         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9746                              new_dentry->d_name.name,
9747                              new_dentry->d_name.len);
9748
9749         if (ret) {
9750                 if (ret == -EEXIST) {
9751                         /* we shouldn't get
9752                          * eexist without a new_inode */
9753                         if (WARN_ON(!new_inode)) {
9754                                 return ret;
9755                         }
9756                 } else {
9757                         /* maybe -EOVERFLOW */
9758                         return ret;
9759                 }
9760         }
9761         ret = 0;
9762
9763         /*
9764          * we're using rename to replace one file with another.  Start IO on it
9765          * now so  we don't add too much work to the end of the transaction
9766          */
9767         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9768                 filemap_flush(old_inode->i_mapping);
9769
9770         /* close the racy window with snapshot create/destroy ioctl */
9771         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9772                 down_read(&root->fs_info->subvol_sem);
9773         /*
9774          * We want to reserve the absolute worst case amount of items.  So if
9775          * both inodes are subvols and we need to unlink them then that would
9776          * require 4 item modifications, but if they are both normal inodes it
9777          * would require 5 item modifications, so we'll assume they are normal
9778          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9779          * should cover the worst case number of items we'll modify.
9780          * If our rename has the whiteout flag, we need more 5 units for the
9781          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9782          * when selinux is enabled).
9783          */
9784         trans_num_items = 11;
9785         if (flags & RENAME_WHITEOUT)
9786                 trans_num_items += 5;
9787         trans = btrfs_start_transaction(root, trans_num_items);
9788         if (IS_ERR(trans)) {
9789                 ret = PTR_ERR(trans);
9790                 goto out_notrans;
9791         }
9792
9793         if (dest != root)
9794                 btrfs_record_root_in_trans(trans, dest);
9795
9796         ret = btrfs_set_inode_index(new_dir, &index);
9797         if (ret)
9798                 goto out_fail;
9799
9800         BTRFS_I(old_inode)->dir_index = 0ULL;
9801         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9802                 /* force full log commit if subvolume involved. */
9803                 btrfs_set_log_full_commit(root->fs_info, trans);
9804         } else {
9805                 btrfs_pin_log_trans(root);
9806                 log_pinned = true;
9807                 ret = btrfs_insert_inode_ref(trans, dest,
9808                                              new_dentry->d_name.name,
9809                                              new_dentry->d_name.len,
9810                                              old_ino,
9811                                              btrfs_ino(new_dir), index);
9812                 if (ret)
9813                         goto out_fail;
9814         }
9815
9816         inode_inc_iversion(old_dir);
9817         inode_inc_iversion(new_dir);
9818         inode_inc_iversion(old_inode);
9819         old_dir->i_ctime = old_dir->i_mtime =
9820         new_dir->i_ctime = new_dir->i_mtime =
9821         old_inode->i_ctime = current_fs_time(old_dir->i_sb);
9822
9823         if (old_dentry->d_parent != new_dentry->d_parent)
9824                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9825
9826         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9827                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9828                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9829                                         old_dentry->d_name.name,
9830                                         old_dentry->d_name.len);
9831         } else {
9832                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9833                                         d_inode(old_dentry),
9834                                         old_dentry->d_name.name,
9835                                         old_dentry->d_name.len);
9836                 if (!ret)
9837                         ret = btrfs_update_inode(trans, root, old_inode);
9838         }
9839         if (ret) {
9840                 btrfs_abort_transaction(trans, root, ret);
9841                 goto out_fail;
9842         }
9843
9844         if (new_inode) {
9845                 inode_inc_iversion(new_inode);
9846                 new_inode->i_ctime = current_fs_time(new_inode->i_sb);
9847                 if (unlikely(btrfs_ino(new_inode) ==
9848                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9849                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9850                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9851                                                 root_objectid,
9852                                                 new_dentry->d_name.name,
9853                                                 new_dentry->d_name.len);
9854                         BUG_ON(new_inode->i_nlink == 0);
9855                 } else {
9856                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9857                                                  d_inode(new_dentry),
9858                                                  new_dentry->d_name.name,
9859                                                  new_dentry->d_name.len);
9860                 }
9861                 if (!ret && new_inode->i_nlink == 0)
9862                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9863                 if (ret) {
9864                         btrfs_abort_transaction(trans, root, ret);
9865                         goto out_fail;
9866                 }
9867         }
9868
9869         ret = btrfs_add_link(trans, new_dir, old_inode,
9870                              new_dentry->d_name.name,
9871                              new_dentry->d_name.len, 0, index);
9872         if (ret) {
9873                 btrfs_abort_transaction(trans, root, ret);
9874                 goto out_fail;
9875         }
9876
9877         if (old_inode->i_nlink == 1)
9878                 BTRFS_I(old_inode)->dir_index = index;
9879
9880         if (log_pinned) {
9881                 struct dentry *parent = new_dentry->d_parent;
9882
9883                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9884                 btrfs_end_log_trans(root);
9885                 log_pinned = false;
9886         }
9887
9888         if (flags & RENAME_WHITEOUT) {
9889                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9890                                                 old_dentry);
9891
9892                 if (ret) {
9893                         btrfs_abort_transaction(trans, root, ret);
9894                         goto out_fail;
9895                 }
9896         }
9897 out_fail:
9898         /*
9899          * If we have pinned the log and an error happened, we unpin tasks
9900          * trying to sync the log and force them to fallback to a transaction
9901          * commit if the log currently contains any of the inodes involved in
9902          * this rename operation (to ensure we do not persist a log with an
9903          * inconsistent state for any of these inodes or leading to any
9904          * inconsistencies when replayed). If the transaction was aborted, the
9905          * abortion reason is propagated to userspace when attempting to commit
9906          * the transaction. If the log does not contain any of these inodes, we
9907          * allow the tasks to sync it.
9908          */
9909         if (ret && log_pinned) {
9910                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9911                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9912                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9913                     (new_inode &&
9914                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9915                     btrfs_set_log_full_commit(root->fs_info, trans);
9916
9917                 btrfs_end_log_trans(root);
9918                 log_pinned = false;
9919         }
9920         btrfs_end_transaction(trans, root);
9921 out_notrans:
9922         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9923                 up_read(&root->fs_info->subvol_sem);
9924
9925         return ret;
9926 }
9927
9928 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9929                          struct inode *new_dir, struct dentry *new_dentry,
9930                          unsigned int flags)
9931 {
9932         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9933                 return -EINVAL;
9934
9935         if (flags & RENAME_EXCHANGE)
9936                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9937                                           new_dentry);
9938
9939         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9940 }
9941
9942 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9943 {
9944         struct btrfs_delalloc_work *delalloc_work;
9945         struct inode *inode;
9946
9947         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9948                                      work);
9949         inode = delalloc_work->inode;
9950         filemap_flush(inode->i_mapping);
9951         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9952                                 &BTRFS_I(inode)->runtime_flags))
9953                 filemap_flush(inode->i_mapping);
9954
9955         if (delalloc_work->delay_iput)
9956                 btrfs_add_delayed_iput(inode);
9957         else
9958                 iput(inode);
9959         complete(&delalloc_work->completion);
9960 }
9961
9962 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9963                                                     int delay_iput)
9964 {
9965         struct btrfs_delalloc_work *work;
9966
9967         work = kmalloc(sizeof(*work), GFP_NOFS);
9968         if (!work)
9969                 return NULL;
9970
9971         init_completion(&work->completion);
9972         INIT_LIST_HEAD(&work->list);
9973         work->inode = inode;
9974         work->delay_iput = delay_iput;
9975         WARN_ON_ONCE(!inode);
9976         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9977                         btrfs_run_delalloc_work, NULL, NULL);
9978
9979         return work;
9980 }
9981
9982 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9983 {
9984         wait_for_completion(&work->completion);
9985         kfree(work);
9986 }
9987
9988 /*
9989  * some fairly slow code that needs optimization. This walks the list
9990  * of all the inodes with pending delalloc and forces them to disk.
9991  */
9992 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9993                                    int nr)
9994 {
9995         struct btrfs_inode *binode;
9996         struct inode *inode;
9997         struct btrfs_delalloc_work *work, *next;
9998         struct list_head works;
9999         struct list_head splice;
10000         int ret = 0;
10001
10002         INIT_LIST_HEAD(&works);
10003         INIT_LIST_HEAD(&splice);
10004
10005         mutex_lock(&root->delalloc_mutex);
10006         spin_lock(&root->delalloc_lock);
10007         list_splice_init(&root->delalloc_inodes, &splice);
10008         while (!list_empty(&splice)) {
10009                 binode = list_entry(splice.next, struct btrfs_inode,
10010                                     delalloc_inodes);
10011
10012                 list_move_tail(&binode->delalloc_inodes,
10013                                &root->delalloc_inodes);
10014                 inode = igrab(&binode->vfs_inode);
10015                 if (!inode) {
10016                         cond_resched_lock(&root->delalloc_lock);
10017                         continue;
10018                 }
10019                 spin_unlock(&root->delalloc_lock);
10020
10021                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10022                 if (!work) {
10023                         if (delay_iput)
10024                                 btrfs_add_delayed_iput(inode);
10025                         else
10026                                 iput(inode);
10027                         ret = -ENOMEM;
10028                         goto out;
10029                 }
10030                 list_add_tail(&work->list, &works);
10031                 btrfs_queue_work(root->fs_info->flush_workers,
10032                                  &work->work);
10033                 ret++;
10034                 if (nr != -1 && ret >= nr)
10035                         goto out;
10036                 cond_resched();
10037                 spin_lock(&root->delalloc_lock);
10038         }
10039         spin_unlock(&root->delalloc_lock);
10040
10041 out:
10042         list_for_each_entry_safe(work, next, &works, list) {
10043                 list_del_init(&work->list);
10044                 btrfs_wait_and_free_delalloc_work(work);
10045         }
10046
10047         if (!list_empty_careful(&splice)) {
10048                 spin_lock(&root->delalloc_lock);
10049                 list_splice_tail(&splice, &root->delalloc_inodes);
10050                 spin_unlock(&root->delalloc_lock);
10051         }
10052         mutex_unlock(&root->delalloc_mutex);
10053         return ret;
10054 }
10055
10056 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10057 {
10058         int ret;
10059
10060         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
10061                 return -EROFS;
10062
10063         ret = __start_delalloc_inodes(root, delay_iput, -1);
10064         if (ret > 0)
10065                 ret = 0;
10066         /*
10067          * the filemap_flush will queue IO into the worker threads, but
10068          * we have to make sure the IO is actually started and that
10069          * ordered extents get created before we return
10070          */
10071         atomic_inc(&root->fs_info->async_submit_draining);
10072         while (atomic_read(&root->fs_info->nr_async_submits) ||
10073               atomic_read(&root->fs_info->async_delalloc_pages)) {
10074                 wait_event(root->fs_info->async_submit_wait,
10075                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
10076                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
10077         }
10078         atomic_dec(&root->fs_info->async_submit_draining);
10079         return ret;
10080 }
10081
10082 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10083                                int nr)
10084 {
10085         struct btrfs_root *root;
10086         struct list_head splice;
10087         int ret;
10088
10089         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10090                 return -EROFS;
10091
10092         INIT_LIST_HEAD(&splice);
10093
10094         mutex_lock(&fs_info->delalloc_root_mutex);
10095         spin_lock(&fs_info->delalloc_root_lock);
10096         list_splice_init(&fs_info->delalloc_roots, &splice);
10097         while (!list_empty(&splice) && nr) {
10098                 root = list_first_entry(&splice, struct btrfs_root,
10099                                         delalloc_root);
10100                 root = btrfs_grab_fs_root(root);
10101                 BUG_ON(!root);
10102                 list_move_tail(&root->delalloc_root,
10103                                &fs_info->delalloc_roots);
10104                 spin_unlock(&fs_info->delalloc_root_lock);
10105
10106                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10107                 btrfs_put_fs_root(root);
10108                 if (ret < 0)
10109                         goto out;
10110
10111                 if (nr != -1) {
10112                         nr -= ret;
10113                         WARN_ON(nr < 0);
10114                 }
10115                 spin_lock(&fs_info->delalloc_root_lock);
10116         }
10117         spin_unlock(&fs_info->delalloc_root_lock);
10118
10119         ret = 0;
10120         atomic_inc(&fs_info->async_submit_draining);
10121         while (atomic_read(&fs_info->nr_async_submits) ||
10122               atomic_read(&fs_info->async_delalloc_pages)) {
10123                 wait_event(fs_info->async_submit_wait,
10124                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10125                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10126         }
10127         atomic_dec(&fs_info->async_submit_draining);
10128 out:
10129         if (!list_empty_careful(&splice)) {
10130                 spin_lock(&fs_info->delalloc_root_lock);
10131                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10132                 spin_unlock(&fs_info->delalloc_root_lock);
10133         }
10134         mutex_unlock(&fs_info->delalloc_root_mutex);
10135         return ret;
10136 }
10137
10138 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10139                          const char *symname)
10140 {
10141         struct btrfs_trans_handle *trans;
10142         struct btrfs_root *root = BTRFS_I(dir)->root;
10143         struct btrfs_path *path;
10144         struct btrfs_key key;
10145         struct inode *inode = NULL;
10146         int err;
10147         int drop_inode = 0;
10148         u64 objectid;
10149         u64 index = 0;
10150         int name_len;
10151         int datasize;
10152         unsigned long ptr;
10153         struct btrfs_file_extent_item *ei;
10154         struct extent_buffer *leaf;
10155
10156         name_len = strlen(symname);
10157         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
10158                 return -ENAMETOOLONG;
10159
10160         /*
10161          * 2 items for inode item and ref
10162          * 2 items for dir items
10163          * 1 item for updating parent inode item
10164          * 1 item for the inline extent item
10165          * 1 item for xattr if selinux is on
10166          */
10167         trans = btrfs_start_transaction(root, 7);
10168         if (IS_ERR(trans))
10169                 return PTR_ERR(trans);
10170
10171         err = btrfs_find_free_ino(root, &objectid);
10172         if (err)
10173                 goto out_unlock;
10174
10175         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10176                                 dentry->d_name.len, btrfs_ino(dir), objectid,
10177                                 S_IFLNK|S_IRWXUGO, &index);
10178         if (IS_ERR(inode)) {
10179                 err = PTR_ERR(inode);
10180                 goto out_unlock;
10181         }
10182
10183         /*
10184         * If the active LSM wants to access the inode during
10185         * d_instantiate it needs these. Smack checks to see
10186         * if the filesystem supports xattrs by looking at the
10187         * ops vector.
10188         */
10189         inode->i_fop = &btrfs_file_operations;
10190         inode->i_op = &btrfs_file_inode_operations;
10191         inode->i_mapping->a_ops = &btrfs_aops;
10192         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10193
10194         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10195         if (err)
10196                 goto out_unlock_inode;
10197
10198         path = btrfs_alloc_path();
10199         if (!path) {
10200                 err = -ENOMEM;
10201                 goto out_unlock_inode;
10202         }
10203         key.objectid = btrfs_ino(inode);
10204         key.offset = 0;
10205         key.type = BTRFS_EXTENT_DATA_KEY;
10206         datasize = btrfs_file_extent_calc_inline_size(name_len);
10207         err = btrfs_insert_empty_item(trans, root, path, &key,
10208                                       datasize);
10209         if (err) {
10210                 btrfs_free_path(path);
10211                 goto out_unlock_inode;
10212         }
10213         leaf = path->nodes[0];
10214         ei = btrfs_item_ptr(leaf, path->slots[0],
10215                             struct btrfs_file_extent_item);
10216         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10217         btrfs_set_file_extent_type(leaf, ei,
10218                                    BTRFS_FILE_EXTENT_INLINE);
10219         btrfs_set_file_extent_encryption(leaf, ei, 0);
10220         btrfs_set_file_extent_compression(leaf, ei, 0);
10221         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10222         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10223
10224         ptr = btrfs_file_extent_inline_start(ei);
10225         write_extent_buffer(leaf, symname, ptr, name_len);
10226         btrfs_mark_buffer_dirty(leaf);
10227         btrfs_free_path(path);
10228
10229         inode->i_op = &btrfs_symlink_inode_operations;
10230         inode_nohighmem(inode);
10231         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10232         inode_set_bytes(inode, name_len);
10233         btrfs_i_size_write(inode, name_len);
10234         err = btrfs_update_inode(trans, root, inode);
10235         /*
10236          * Last step, add directory indexes for our symlink inode. This is the
10237          * last step to avoid extra cleanup of these indexes if an error happens
10238          * elsewhere above.
10239          */
10240         if (!err)
10241                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
10242         if (err) {
10243                 drop_inode = 1;
10244                 goto out_unlock_inode;
10245         }
10246
10247         unlock_new_inode(inode);
10248         d_instantiate(dentry, inode);
10249
10250 out_unlock:
10251         btrfs_end_transaction(trans, root);
10252         if (drop_inode) {
10253                 inode_dec_link_count(inode);
10254                 iput(inode);
10255         }
10256         btrfs_btree_balance_dirty(root);
10257         return err;
10258
10259 out_unlock_inode:
10260         drop_inode = 1;
10261         unlock_new_inode(inode);
10262         goto out_unlock;
10263 }
10264
10265 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10266                                        u64 start, u64 num_bytes, u64 min_size,
10267                                        loff_t actual_len, u64 *alloc_hint,
10268                                        struct btrfs_trans_handle *trans)
10269 {
10270         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10271         struct extent_map *em;
10272         struct btrfs_root *root = BTRFS_I(inode)->root;
10273         struct btrfs_key ins;
10274         u64 cur_offset = start;
10275         u64 i_size;
10276         u64 cur_bytes;
10277         u64 last_alloc = (u64)-1;
10278         int ret = 0;
10279         bool own_trans = true;
10280
10281         if (trans)
10282                 own_trans = false;
10283         while (num_bytes > 0) {
10284                 if (own_trans) {
10285                         trans = btrfs_start_transaction(root, 3);
10286                         if (IS_ERR(trans)) {
10287                                 ret = PTR_ERR(trans);
10288                                 break;
10289                         }
10290                 }
10291
10292                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10293                 cur_bytes = max(cur_bytes, min_size);
10294                 /*
10295                  * If we are severely fragmented we could end up with really
10296                  * small allocations, so if the allocator is returning small
10297                  * chunks lets make its job easier by only searching for those
10298                  * sized chunks.
10299                  */
10300                 cur_bytes = min(cur_bytes, last_alloc);
10301                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
10302                                            *alloc_hint, &ins, 1, 0);
10303                 if (ret) {
10304                         if (own_trans)
10305                                 btrfs_end_transaction(trans, root);
10306                         break;
10307                 }
10308                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
10309
10310                 last_alloc = ins.offset;
10311                 ret = insert_reserved_file_extent(trans, inode,
10312                                                   cur_offset, ins.objectid,
10313                                                   ins.offset, ins.offset,
10314                                                   ins.offset, 0, 0, 0,
10315                                                   BTRFS_FILE_EXTENT_PREALLOC);
10316                 if (ret) {
10317                         btrfs_free_reserved_extent(root, ins.objectid,
10318                                                    ins.offset, 0);
10319                         btrfs_abort_transaction(trans, root, ret);
10320                         if (own_trans)
10321                                 btrfs_end_transaction(trans, root);
10322                         break;
10323                 }
10324
10325                 btrfs_drop_extent_cache(inode, cur_offset,
10326                                         cur_offset + ins.offset -1, 0);
10327
10328                 em = alloc_extent_map();
10329                 if (!em) {
10330                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10331                                 &BTRFS_I(inode)->runtime_flags);
10332                         goto next;
10333                 }
10334
10335                 em->start = cur_offset;
10336                 em->orig_start = cur_offset;
10337                 em->len = ins.offset;
10338                 em->block_start = ins.objectid;
10339                 em->block_len = ins.offset;
10340                 em->orig_block_len = ins.offset;
10341                 em->ram_bytes = ins.offset;
10342                 em->bdev = root->fs_info->fs_devices->latest_bdev;
10343                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10344                 em->generation = trans->transid;
10345
10346                 while (1) {
10347                         write_lock(&em_tree->lock);
10348                         ret = add_extent_mapping(em_tree, em, 1);
10349                         write_unlock(&em_tree->lock);
10350                         if (ret != -EEXIST)
10351                                 break;
10352                         btrfs_drop_extent_cache(inode, cur_offset,
10353                                                 cur_offset + ins.offset - 1,
10354                                                 0);
10355                 }
10356                 free_extent_map(em);
10357 next:
10358                 num_bytes -= ins.offset;
10359                 cur_offset += ins.offset;
10360                 *alloc_hint = ins.objectid + ins.offset;
10361
10362                 inode_inc_iversion(inode);
10363                 inode->i_ctime = current_fs_time(inode->i_sb);
10364                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10365                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10366                     (actual_len > inode->i_size) &&
10367                     (cur_offset > inode->i_size)) {
10368                         if (cur_offset > actual_len)
10369                                 i_size = actual_len;
10370                         else
10371                                 i_size = cur_offset;
10372                         i_size_write(inode, i_size);
10373                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10374                 }
10375
10376                 ret = btrfs_update_inode(trans, root, inode);
10377
10378                 if (ret) {
10379                         btrfs_abort_transaction(trans, root, ret);
10380                         if (own_trans)
10381                                 btrfs_end_transaction(trans, root);
10382                         break;
10383                 }
10384
10385                 if (own_trans)
10386                         btrfs_end_transaction(trans, root);
10387         }
10388         return ret;
10389 }
10390
10391 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10392                               u64 start, u64 num_bytes, u64 min_size,
10393                               loff_t actual_len, u64 *alloc_hint)
10394 {
10395         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10396                                            min_size, actual_len, alloc_hint,
10397                                            NULL);
10398 }
10399
10400 int btrfs_prealloc_file_range_trans(struct inode *inode,
10401                                     struct btrfs_trans_handle *trans, int mode,
10402                                     u64 start, u64 num_bytes, u64 min_size,
10403                                     loff_t actual_len, u64 *alloc_hint)
10404 {
10405         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10406                                            min_size, actual_len, alloc_hint, trans);
10407 }
10408
10409 static int btrfs_set_page_dirty(struct page *page)
10410 {
10411         return __set_page_dirty_nobuffers(page);
10412 }
10413
10414 static int btrfs_permission(struct inode *inode, int mask)
10415 {
10416         struct btrfs_root *root = BTRFS_I(inode)->root;
10417         umode_t mode = inode->i_mode;
10418
10419         if (mask & MAY_WRITE &&
10420             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10421                 if (btrfs_root_readonly(root))
10422                         return -EROFS;
10423                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10424                         return -EACCES;
10425         }
10426         return generic_permission(inode, mask);
10427 }
10428
10429 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10430 {
10431         struct btrfs_trans_handle *trans;
10432         struct btrfs_root *root = BTRFS_I(dir)->root;
10433         struct inode *inode = NULL;
10434         u64 objectid;
10435         u64 index;
10436         int ret = 0;
10437
10438         /*
10439          * 5 units required for adding orphan entry
10440          */
10441         trans = btrfs_start_transaction(root, 5);
10442         if (IS_ERR(trans))
10443                 return PTR_ERR(trans);
10444
10445         ret = btrfs_find_free_ino(root, &objectid);
10446         if (ret)
10447                 goto out;
10448
10449         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10450                                 btrfs_ino(dir), objectid, mode, &index);
10451         if (IS_ERR(inode)) {
10452                 ret = PTR_ERR(inode);
10453                 inode = NULL;
10454                 goto out;
10455         }
10456
10457         inode->i_fop = &btrfs_file_operations;
10458         inode->i_op = &btrfs_file_inode_operations;
10459
10460         inode->i_mapping->a_ops = &btrfs_aops;
10461         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10462
10463         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10464         if (ret)
10465                 goto out_inode;
10466
10467         ret = btrfs_update_inode(trans, root, inode);
10468         if (ret)
10469                 goto out_inode;
10470         ret = btrfs_orphan_add(trans, inode);
10471         if (ret)
10472                 goto out_inode;
10473
10474         /*
10475          * We set number of links to 0 in btrfs_new_inode(), and here we set
10476          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10477          * through:
10478          *
10479          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10480          */
10481         set_nlink(inode, 1);
10482         unlock_new_inode(inode);
10483         d_tmpfile(dentry, inode);
10484         mark_inode_dirty(inode);
10485
10486 out:
10487         btrfs_end_transaction(trans, root);
10488         if (ret)
10489                 iput(inode);
10490         btrfs_balance_delayed_items(root);
10491         btrfs_btree_balance_dirty(root);
10492         return ret;
10493
10494 out_inode:
10495         unlock_new_inode(inode);
10496         goto out;
10497
10498 }
10499
10500 /* Inspired by filemap_check_errors() */
10501 int btrfs_inode_check_errors(struct inode *inode)
10502 {
10503         int ret = 0;
10504
10505         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10506             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10507                 ret = -ENOSPC;
10508         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10509             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10510                 ret = -EIO;
10511
10512         return ret;
10513 }
10514
10515 static const struct inode_operations btrfs_dir_inode_operations = {
10516         .getattr        = btrfs_getattr,
10517         .lookup         = btrfs_lookup,
10518         .create         = btrfs_create,
10519         .unlink         = btrfs_unlink,
10520         .link           = btrfs_link,
10521         .mkdir          = btrfs_mkdir,
10522         .rmdir          = btrfs_rmdir,
10523         .rename2        = btrfs_rename2,
10524         .symlink        = btrfs_symlink,
10525         .setattr        = btrfs_setattr,
10526         .mknod          = btrfs_mknod,
10527         .setxattr       = generic_setxattr,
10528         .getxattr       = generic_getxattr,
10529         .listxattr      = btrfs_listxattr,
10530         .removexattr    = generic_removexattr,
10531         .permission     = btrfs_permission,
10532         .get_acl        = btrfs_get_acl,
10533         .set_acl        = btrfs_set_acl,
10534         .update_time    = btrfs_update_time,
10535         .tmpfile        = btrfs_tmpfile,
10536 };
10537 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10538         .lookup         = btrfs_lookup,
10539         .permission     = btrfs_permission,
10540         .get_acl        = btrfs_get_acl,
10541         .set_acl        = btrfs_set_acl,
10542         .update_time    = btrfs_update_time,
10543 };
10544
10545 static const struct file_operations btrfs_dir_file_operations = {
10546         .llseek         = generic_file_llseek,
10547         .read           = generic_read_dir,
10548         .iterate_shared = btrfs_real_readdir,
10549         .unlocked_ioctl = btrfs_ioctl,
10550 #ifdef CONFIG_COMPAT
10551         .compat_ioctl   = btrfs_compat_ioctl,
10552 #endif
10553         .release        = btrfs_release_file,
10554         .fsync          = btrfs_sync_file,
10555 };
10556
10557 static const struct extent_io_ops btrfs_extent_io_ops = {
10558         .fill_delalloc = run_delalloc_range,
10559         .submit_bio_hook = btrfs_submit_bio_hook,
10560         .merge_bio_hook = btrfs_merge_bio_hook,
10561         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10562         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10563         .writepage_start_hook = btrfs_writepage_start_hook,
10564         .set_bit_hook = btrfs_set_bit_hook,
10565         .clear_bit_hook = btrfs_clear_bit_hook,
10566         .merge_extent_hook = btrfs_merge_extent_hook,
10567         .split_extent_hook = btrfs_split_extent_hook,
10568 };
10569
10570 /*
10571  * btrfs doesn't support the bmap operation because swapfiles
10572  * use bmap to make a mapping of extents in the file.  They assume
10573  * these extents won't change over the life of the file and they
10574  * use the bmap result to do IO directly to the drive.
10575  *
10576  * the btrfs bmap call would return logical addresses that aren't
10577  * suitable for IO and they also will change frequently as COW
10578  * operations happen.  So, swapfile + btrfs == corruption.
10579  *
10580  * For now we're avoiding this by dropping bmap.
10581  */
10582 static const struct address_space_operations btrfs_aops = {
10583         .readpage       = btrfs_readpage,
10584         .writepage      = btrfs_writepage,
10585         .writepages     = btrfs_writepages,
10586         .readpages      = btrfs_readpages,
10587         .direct_IO      = btrfs_direct_IO,
10588         .invalidatepage = btrfs_invalidatepage,
10589         .releasepage    = btrfs_releasepage,
10590         .set_page_dirty = btrfs_set_page_dirty,
10591         .error_remove_page = generic_error_remove_page,
10592 };
10593
10594 static const struct address_space_operations btrfs_symlink_aops = {
10595         .readpage       = btrfs_readpage,
10596         .writepage      = btrfs_writepage,
10597         .invalidatepage = btrfs_invalidatepage,
10598         .releasepage    = btrfs_releasepage,
10599 };
10600
10601 static const struct inode_operations btrfs_file_inode_operations = {
10602         .getattr        = btrfs_getattr,
10603         .setattr        = btrfs_setattr,
10604         .setxattr       = generic_setxattr,
10605         .getxattr       = generic_getxattr,
10606         .listxattr      = btrfs_listxattr,
10607         .removexattr    = generic_removexattr,
10608         .permission     = btrfs_permission,
10609         .fiemap         = btrfs_fiemap,
10610         .get_acl        = btrfs_get_acl,
10611         .set_acl        = btrfs_set_acl,
10612         .update_time    = btrfs_update_time,
10613 };
10614 static const struct inode_operations btrfs_special_inode_operations = {
10615         .getattr        = btrfs_getattr,
10616         .setattr        = btrfs_setattr,
10617         .permission     = btrfs_permission,
10618         .setxattr       = generic_setxattr,
10619         .getxattr       = generic_getxattr,
10620         .listxattr      = btrfs_listxattr,
10621         .removexattr    = generic_removexattr,
10622         .get_acl        = btrfs_get_acl,
10623         .set_acl        = btrfs_set_acl,
10624         .update_time    = btrfs_update_time,
10625 };
10626 static const struct inode_operations btrfs_symlink_inode_operations = {
10627         .readlink       = generic_readlink,
10628         .get_link       = page_get_link,
10629         .getattr        = btrfs_getattr,
10630         .setattr        = btrfs_setattr,
10631         .permission     = btrfs_permission,
10632         .setxattr       = generic_setxattr,
10633         .getxattr       = generic_getxattr,
10634         .listxattr      = btrfs_listxattr,
10635         .removexattr    = generic_removexattr,
10636         .update_time    = btrfs_update_time,
10637 };
10638
10639 const struct dentry_operations btrfs_dentry_operations = {
10640         .d_delete       = btrfs_dentry_delete,
10641         .d_release      = btrfs_dentry_release,
10642 };