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