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