fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/buffer_head.h>
   9 #include <linux/file.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/highmem.h>
  13 #include <linux/time.h>
  14 #include <linux/init.h>
  15 #include <linux/string.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/writeback.h>
  18 #include <linux/compat.h>
  19 #include <linux/xattr.h>
  20 #include <linux/posix_acl.h>
  21 #include <linux/falloc.h>
  22 #include <linux/slab.h>
  23 #include <linux/ratelimit.h>
  24 #include <linux/btrfs.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/posix_acl_xattr.h>
  27 #include <linux/uio.h>
  28 #include <linux/magic.h>
  29 #include <linux/iversion.h>
  30 #include <asm/unaligned.h>
  31 #include "ctree.h"
  32 #include "disk-io.h"
  33 #include "transaction.h"
  34 #include "btrfs_inode.h"
  35 #include "print-tree.h"
  36 #include "ordered-data.h"
  37 #include "xattr.h"
  38 #include "tree-log.h"
  39 #include "volumes.h"
  40 #include "compression.h"
  41 #include "locking.h"
  42 #include "free-space-cache.h"
  43 #include "inode-map.h"
  44 #include "backref.h"
  45 #include "props.h"
  46 #include "qgroup.h"
  47 #include "dedupe.h"
  48
  49 struct btrfs_iget_args {
  50         struct btrfs_key *location;
  51         struct btrfs_root *root;
  52 };
  53
  54 struct btrfs_dio_data {
  55         u64 reserve;
  56         u64 unsubmitted_oe_range_start;
  57         u64 unsubmitted_oe_range_end;
  58         int overwrite;
  59 };
  60
  61 static const struct inode_operations btrfs_dir_inode_operations;
  62 static const struct inode_operations btrfs_symlink_inode_operations;
  63 static const struct inode_operations btrfs_dir_ro_inode_operations;
  64 static const struct inode_operations btrfs_special_inode_operations;
  65 static const struct inode_operations btrfs_file_inode_operations;
  66 static const struct address_space_operations btrfs_aops;
  67 static const struct address_space_operations btrfs_symlink_aops;
  68 static const struct file_operations btrfs_dir_file_operations;
  69 static const struct extent_io_ops btrfs_extent_io_ops;
  70
  71 static struct kmem_cache *btrfs_inode_cachep;
  72 struct kmem_cache *btrfs_trans_handle_cachep;
  73 struct kmem_cache *btrfs_path_cachep;
  74 struct kmem_cache *btrfs_free_space_cachep;
  75
  76 #define S_SHIFT 12
  77 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  78         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  79         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  80         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  81         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  82         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  83         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  84         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  85 };
  86
  87 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  88 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  89 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  90 static noinline int cow_file_range(struct inode *inode,
  91                                    struct page *locked_page,
  92                                    u64 start, u64 end, u64 delalloc_end,
  93                                    int *page_started, unsigned long *nr_written,
  94                                    int unlock, struct btrfs_dedupe_hash *hash);
  95 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  96                                        u64 orig_start, u64 block_start,
  97                                        u64 block_len, u64 orig_block_len,
  98                                        u64 ram_bytes, int compress_type,
  99                                        int type);
 100
 101 static void __endio_write_update_ordered(struct inode *inode,
 102                                          const u64 offset, const u64 bytes,
 103                                          const bool uptodate);
 104
 105 /*
 106  * Cleanup all submitted ordered extents in specified range to handle errors
 107  * from the fill_dellaloc() callback.
 108  *
 109  * NOTE: caller must ensure that when an error happens, it can not call
 110  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 111  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 112  * to be released, which we want to happen only when finishing the ordered
 113  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 114  * fill_delalloc() callback already does proper cleanup for the first page of
 115  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 116  * range of the first page.
 117  */
 118 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 119                                                  const u64 offset,
 120                                                  const u64 bytes)
 121 {
 122         unsigned long index = offset >> PAGE_SHIFT;
 123         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 124         struct page *page;
 125
 126         while (index <= end_index) {
 127                 page = find_get_page(inode->i_mapping, index);
 128                 index++;
 129                 if (!page)
 130                         continue;
 131                 ClearPagePrivate2(page);
 132                 put_page(page);
 133         }
 134         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 135                                             bytes - PAGE_SIZE, false);
 136 }
 137
 138 static int btrfs_dirty_inode(struct inode *inode);
 139
 140 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 141 void btrfs_test_inode_set_ops(struct inode *inode)
 142 {
 143         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 144 }
 145 #endif
 146
 147 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 148                                      struct inode *inode,  struct inode *dir,
 149                                      const struct qstr *qstr)
 150 {
 151         int err;
 152
 153         err = btrfs_init_acl(trans, inode, dir);
 154         if (!err)
 155                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 156         return err;
 157 }
 158
 159 /*
 160  * this does all the hard work for inserting an inline extent into
 161  * the btree.  The caller should have done a btrfs_drop_extents so that
 162  * no overlapping inline items exist in the btree
 163  */
 164 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 165                                 struct btrfs_path *path, int extent_inserted,
 166                                 struct btrfs_root *root, struct inode *inode,
 167                                 u64 start, size_t size, size_t compressed_size,
 168                                 int compress_type,
 169                                 struct page **compressed_pages)
 170 {
 171         struct extent_buffer *leaf;
 172         struct page *page = NULL;
 173         char *kaddr;
 174         unsigned long ptr;
 175         struct btrfs_file_extent_item *ei;
 176         int ret;
 177         size_t cur_size = size;
 178         unsigned long offset;
 179
 180         if (compressed_size && compressed_pages)
 181                 cur_size = compressed_size;
 182
 183         inode_add_bytes(inode, size);
 184
 185         if (!extent_inserted) {
 186                 struct btrfs_key key;
 187                 size_t datasize;
 188
 189                 key.objectid = btrfs_ino(BTRFS_I(inode));
 190                 key.offset = start;
 191                 key.type = BTRFS_EXTENT_DATA_KEY;
 192
 193                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 194                 path->leave_spinning = 1;
 195                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 196                                               datasize);
 197                 if (ret)
 198                         goto fail;
 199         }
 200         leaf = path->nodes[0];
 201         ei = btrfs_item_ptr(leaf, path->slots[0],
 202                             struct btrfs_file_extent_item);
 203         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 204         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 205         btrfs_set_file_extent_encryption(leaf, ei, 0);
 206         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 207         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 208         ptr = btrfs_file_extent_inline_start(ei);
 209
 210         if (compress_type != BTRFS_COMPRESS_NONE) {
 211                 struct page *cpage;
 212                 int i = 0;
 213                 while (compressed_size > 0) {
 214                         cpage = compressed_pages[i];
 215                         cur_size = min_t(unsigned long, compressed_size,
 216                                        PAGE_SIZE);
 217
 218                         kaddr = kmap_atomic(cpage);
 219                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 220                         kunmap_atomic(kaddr);
 221
 222                         i++;
 223                         ptr += cur_size;
 224                         compressed_size -= cur_size;
 225                 }
 226                 btrfs_set_file_extent_compression(leaf, ei,
 227                                                   compress_type);
 228         } else {
 229                 page = find_get_page(inode->i_mapping,
 230                                      start >> PAGE_SHIFT);
 231                 btrfs_set_file_extent_compression(leaf, ei, 0);
 232                 kaddr = kmap_atomic(page);
 233                 offset = start & (PAGE_SIZE - 1);
 234                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 235                 kunmap_atomic(kaddr);
 236                 put_page(page);
 237         }
 238         btrfs_mark_buffer_dirty(leaf);
 239         btrfs_release_path(path);
 240
 241         /*
 242          * we're an inline extent, so nobody can
 243          * extend the file past i_size without locking
 244          * a page we already have locked.
 245          *
 246          * We must do any isize and inode updates
 247          * before we unlock the pages.  Otherwise we
 248          * could end up racing with unlink.
 249          */
 250         BTRFS_I(inode)->disk_i_size = inode->i_size;
 251         ret = btrfs_update_inode(trans, root, inode);
 252
 253 fail:
 254         return ret;
 255 }
 256
 257
 258 /*
 259  * conditionally insert an inline extent into the file.  This
 260  * does the checks required to make sure the data is small enough
 261  * to fit as an inline extent.
 262  */
 263 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 264                                           u64 end, size_t compressed_size,
 265                                           int compress_type,
 266                                           struct page **compressed_pages)
 267 {
 268         struct btrfs_root *root = BTRFS_I(inode)->root;
 269         struct btrfs_fs_info *fs_info = root->fs_info;
 270         struct btrfs_trans_handle *trans;
 271         u64 isize = i_size_read(inode);
 272         u64 actual_end = min(end + 1, isize);
 273         u64 inline_len = actual_end - start;
 274         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 275         u64 data_len = inline_len;
 276         int ret;
 277         struct btrfs_path *path;
 278         int extent_inserted = 0;
 279         u32 extent_item_size;
 280
 281         if (compressed_size)
 282                 data_len = compressed_size;
 283
 284         if (start > 0 ||
 285             actual_end > fs_info->sectorsize ||
 286             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 287             (!compressed_size &&
 288             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 289             end + 1 < isize ||
 290             data_len > fs_info->max_inline) {
 291                 return 1;
 292         }
 293
 294         path = btrfs_alloc_path();
 295         if (!path)
 296                 return -ENOMEM;
 297
 298         trans = btrfs_join_transaction(root);
 299         if (IS_ERR(trans)) {
 300                 btrfs_free_path(path);
 301                 return PTR_ERR(trans);
 302         }
 303         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 304
 305         if (compressed_size && compressed_pages)
 306                 extent_item_size = btrfs_file_extent_calc_inline_size(
 307                    compressed_size);
 308         else
 309                 extent_item_size = btrfs_file_extent_calc_inline_size(
 310                     inline_len);
 311
 312         ret = __btrfs_drop_extents(trans, root, inode, path,
 313                                    start, aligned_end, NULL,
 314                                    1, 1, extent_item_size, &extent_inserted);
 315         if (ret) {
 316                 btrfs_abort_transaction(trans, ret);
 317                 goto out;
 318         }
 319
 320         if (isize > actual_end)
 321                 inline_len = min_t(u64, isize, actual_end);
 322         ret = insert_inline_extent(trans, path, extent_inserted,
 323                                    root, inode, start,
 324                                    inline_len, compressed_size,
 325                                    compress_type, compressed_pages);
 326         if (ret && ret != -ENOSPC) {
 327                 btrfs_abort_transaction(trans, ret);
 328                 goto out;
 329         } else if (ret == -ENOSPC) {
 330                 ret = 1;
 331                 goto out;
 332         }
 333
 334         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 335         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 336 out:
 337         /*
 338          * Don't forget to free the reserved space, as for inlined extent
 339          * it won't count as data extent, free them directly here.
 340          * And at reserve time, it's always aligned to page size, so
 341          * just free one page here.
 342          */
 343         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 344         btrfs_free_path(path);
 345         btrfs_end_transaction(trans);
 346         return ret;
 347 }
 348
 349 struct async_extent {
 350         u64 start;
 351         u64 ram_size;
 352         u64 compressed_size;
 353         struct page **pages;
 354         unsigned long nr_pages;
 355         int compress_type;
 356         struct list_head list;
 357 };
 358
 359 struct async_cow {
 360         struct inode *inode;
 361         struct btrfs_root *root;
 362         struct page *locked_page;
 363         u64 start;
 364         u64 end;
 365         unsigned int write_flags;
 366         struct list_head extents;
 367         struct btrfs_work work;
 368 };
 369
 370 static noinline int add_async_extent(struct async_cow *cow,
 371                                      u64 start, u64 ram_size,
 372                                      u64 compressed_size,
 373                                      struct page **pages,
 374                                      unsigned long nr_pages,
 375                                      int compress_type)
 376 {
 377         struct async_extent *async_extent;
 378
 379         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 380         BUG_ON(!async_extent); /* -ENOMEM */
 381         async_extent->start = start;
 382         async_extent->ram_size = ram_size;
 383         async_extent->compressed_size = compressed_size;
 384         async_extent->pages = pages;
 385         async_extent->nr_pages = nr_pages;
 386         async_extent->compress_type = compress_type;
 387         list_add_tail(&async_extent->list, &cow->extents);
 388         return 0;
 389 }
 390
 391 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 392 {
 393         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 394
 395         /* force compress */
 396         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 397                 return 1;
 398         /* defrag ioctl */
 399         if (BTRFS_I(inode)->defrag_compress)
 400                 return 1;
 401         /* bad compression ratios */
 402         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 403                 return 0;
 404         if (btrfs_test_opt(fs_info, COMPRESS) ||
 405             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 406             BTRFS_I(inode)->prop_compress)
 407                 return btrfs_compress_heuristic(inode, start, end);
 408         return 0;
 409 }
 410
 411 static inline void inode_should_defrag(struct btrfs_inode *inode,
 412                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 413 {
 414         /* If this is a small write inside eof, kick off a defrag */
 415         if (num_bytes < small_write &&
 416             (start > 0 || end + 1 < inode->disk_i_size))
 417                 btrfs_add_inode_defrag(NULL, inode);
 418 }
 419
 420 /*
 421  * we create compressed extents in two phases.  The first
 422  * phase compresses a range of pages that have already been
 423  * locked (both pages and state bits are locked).
 424  *
 425  * This is done inside an ordered work queue, and the compression
 426  * is spread across many cpus.  The actual IO submission is step
 427  * two, and the ordered work queue takes care of making sure that
 428  * happens in the same order things were put onto the queue by
 429  * writepages and friends.
 430  *
 431  * If this code finds it can't get good compression, it puts an
 432  * entry onto the work queue to write the uncompressed bytes.  This
 433  * makes sure that both compressed inodes and uncompressed inodes
 434  * are written in the same order that the flusher thread sent them
 435  * down.
 436  */
 437 static noinline void compress_file_range(struct inode *inode,
 438                                         struct page *locked_page,
 439                                         u64 start, u64 end,
 440                                         struct async_cow *async_cow,
 441                                         int *num_added)
 442 {
 443         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 444         u64 blocksize = fs_info->sectorsize;
 445         u64 actual_end;
 446         u64 isize = i_size_read(inode);
 447         int ret = 0;
 448         struct page **pages = NULL;
 449         unsigned long nr_pages;
 450         unsigned long total_compressed = 0;
 451         unsigned long total_in = 0;
 452         int i;
 453         int will_compress;
 454         int compress_type = fs_info->compress_type;
 455         int redirty = 0;
 456
 457         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 458                         SZ_16K);
 459
 460         actual_end = min_t(u64, isize, end + 1);
 461 again:
 462         will_compress = 0;
 463         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 464         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 465         nr_pages = min_t(unsigned long, nr_pages,
 466                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 467
 468         /*
 469          * we don't want to send crud past the end of i_size through
 470          * compression, that's just a waste of CPU time.  So, if the
 471          * end of the file is before the start of our current
 472          * requested range of bytes, we bail out to the uncompressed
 473          * cleanup code that can deal with all of this.
 474          *
 475          * It isn't really the fastest way to fix things, but this is a
 476          * very uncommon corner.
 477          */
 478         if (actual_end <= start)
 479                 goto cleanup_and_bail_uncompressed;
 480
 481         total_compressed = actual_end - start;
 482
 483         /*
 484          * skip compression for a small file range(<=blocksize) that
 485          * isn't an inline extent, since it doesn't save disk space at all.
 486          */
 487         if (total_compressed <= blocksize &&
 488            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 489                 goto cleanup_and_bail_uncompressed;
 490
 491         total_compressed = min_t(unsigned long, total_compressed,
 492                         BTRFS_MAX_UNCOMPRESSED);
 493         total_in = 0;
 494         ret = 0;
 495
 496         /*
 497          * we do compression for mount -o compress and when the
 498          * inode has not been flagged as nocompress.  This flag can
 499          * change at any time if we discover bad compression ratios.
 500          */
 501         if (inode_need_compress(inode, start, end)) {
 502                 WARN_ON(pages);
 503                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 504                 if (!pages) {
 505                         /* just bail out to the uncompressed code */
 506                         goto cont;
 507                 }
 508
 509                 if (BTRFS_I(inode)->defrag_compress)
 510                         compress_type = BTRFS_I(inode)->defrag_compress;
 511                 else if (BTRFS_I(inode)->prop_compress)
 512                         compress_type = BTRFS_I(inode)->prop_compress;
 513
 514                 /*
 515                  * we need to call clear_page_dirty_for_io on each
 516                  * page in the range.  Otherwise applications with the file
 517                  * mmap'd can wander in and change the page contents while
 518                  * we are compressing them.
 519                  *
 520                  * If the compression fails for any reason, we set the pages
 521                  * dirty again later on.
 522                  *
 523                  * Note that the remaining part is redirtied, the start pointer
 524                  * has moved, the end is the original one.
 525                  */
 526                 if (!redirty) {
 527                         extent_range_clear_dirty_for_io(inode, start, end);
 528                         redirty = 1;
 529                 }
 530
 531                 /* Compression level is applied here and only here */
 532                 ret = btrfs_compress_pages(
 533                         compress_type | (fs_info->compress_level << 4),
 534                                            inode->i_mapping, start,
 535                                            pages,
 536                                            &nr_pages,
 537                                            &total_in,
 538                                            &total_compressed);
 539
 540                 if (!ret) {
 541                         unsigned long offset = total_compressed &
 542                                 (PAGE_SIZE - 1);
 543                         struct page *page = pages[nr_pages - 1];
 544                         char *kaddr;
 545
 546                         /* zero the tail end of the last page, we might be
 547                          * sending it down to disk
 548                          */
 549                         if (offset) {
 550                                 kaddr = kmap_atomic(page);
 551                                 memset(kaddr + offset, 0,
 552                                        PAGE_SIZE - offset);
 553                                 kunmap_atomic(kaddr);
 554                         }
 555                         will_compress = 1;
 556                 }
 557         }
 558 cont:
 559         if (start == 0) {
 560                 /* lets try to make an inline extent */
 561                 if (ret || total_in < actual_end) {
 562                         /* we didn't compress the entire range, try
 563                          * to make an uncompressed inline extent.
 564                          */
 565                         ret = cow_file_range_inline(inode, start, end, 0,
 566                                                     BTRFS_COMPRESS_NONE, NULL);
 567                 } else {
 568                         /* try making a compressed inline extent */
 569                         ret = cow_file_range_inline(inode, start, end,
 570                                                     total_compressed,
 571                                                     compress_type, pages);
 572                 }
 573                 if (ret <= 0) {
 574                         unsigned long clear_flags = EXTENT_DELALLOC |
 575                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 576                                 EXTENT_DO_ACCOUNTING;
 577                         unsigned long page_error_op;
 578
 579                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 580
 581                         /*
 582                          * inline extent creation worked or returned error,
 583                          * we don't need to create any more async work items.
 584                          * Unlock and free up our temp pages.
 585                          *
 586                          * We use DO_ACCOUNTING here because we need the
 587                          * delalloc_release_metadata to be done _after_ we drop
 588                          * our outstanding extent for clearing delalloc for this
 589                          * range.
 590                          */
 591                         extent_clear_unlock_delalloc(inode, start, end, end,
 592                                                      NULL, clear_flags,
 593                                                      PAGE_UNLOCK |
 594                                                      PAGE_CLEAR_DIRTY |
 595                                                      PAGE_SET_WRITEBACK |
 596                                                      page_error_op |
 597                                                      PAGE_END_WRITEBACK);
 598                         goto free_pages_out;
 599                 }
 600         }
 601
 602         if (will_compress) {
 603                 /*
 604                  * we aren't doing an inline extent round the compressed size
 605                  * up to a block size boundary so the allocator does sane
 606                  * things
 607                  */
 608                 total_compressed = ALIGN(total_compressed, blocksize);
 609
 610                 /*
 611                  * one last check to make sure the compression is really a
 612                  * win, compare the page count read with the blocks on disk,
 613                  * compression must free at least one sector size
 614                  */
 615                 total_in = ALIGN(total_in, PAGE_SIZE);
 616                 if (total_compressed + blocksize <= total_in) {
 617                         *num_added += 1;
 618
 619                         /*
 620                          * The async work queues will take care of doing actual
 621                          * allocation on disk for these compressed pages, and
 622                          * will submit them to the elevator.
 623                          */
 624                         add_async_extent(async_cow, start, total_in,
 625                                         total_compressed, pages, nr_pages,
 626                                         compress_type);
 627
 628                         if (start + total_in < end) {
 629                                 start += total_in;
 630                                 pages = NULL;
 631                                 cond_resched();
 632                                 goto again;
 633                         }
 634                         return;
 635                 }
 636         }
 637         if (pages) {
 638                 /*
 639                  * the compression code ran but failed to make things smaller,
 640                  * free any pages it allocated and our page pointer array
 641                  */
 642                 for (i = 0; i < nr_pages; i++) {
 643                         WARN_ON(pages[i]->mapping);
 644                         put_page(pages[i]);
 645                 }
 646                 kfree(pages);
 647                 pages = NULL;
 648                 total_compressed = 0;
 649                 nr_pages = 0;
 650
 651                 /* flag the file so we don't compress in the future */
 652                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 653                     !(BTRFS_I(inode)->prop_compress)) {
 654                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 655                 }
 656         }
 657 cleanup_and_bail_uncompressed:
 658         /*
 659          * No compression, but we still need to write the pages in the file
 660          * we've been given so far.  redirty the locked page if it corresponds
 661          * to our extent and set things up for the async work queue to run
 662          * cow_file_range to do the normal delalloc dance.
 663          */
 664         if (page_offset(locked_page) >= start &&
 665             page_offset(locked_page) <= end)
 666                 __set_page_dirty_nobuffers(locked_page);
 667                 /* unlocked later on in the async handlers */
 668
 669         if (redirty)
 670                 extent_range_redirty_for_io(inode, start, end);
 671         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 672                          BTRFS_COMPRESS_NONE);
 673         *num_added += 1;
 674
 675         return;
 676
 677 free_pages_out:
 678         for (i = 0; i < nr_pages; i++) {
 679                 WARN_ON(pages[i]->mapping);
 680                 put_page(pages[i]);
 681         }
 682         kfree(pages);
 683 }
 684
 685 static void free_async_extent_pages(struct async_extent *async_extent)
 686 {
 687         int i;
 688
 689         if (!async_extent->pages)
 690                 return;
 691
 692         for (i = 0; i < async_extent->nr_pages; i++) {
 693                 WARN_ON(async_extent->pages[i]->mapping);
 694                 put_page(async_extent->pages[i]);
 695         }
 696         kfree(async_extent->pages);
 697         async_extent->nr_pages = 0;
 698         async_extent->pages = NULL;
 699 }
 700
 701 /*
 702  * phase two of compressed writeback.  This is the ordered portion
 703  * of the code, which only gets called in the order the work was
 704  * queued.  We walk all the async extents created by compress_file_range
 705  * and send them down to the disk.
 706  */
 707 static noinline void submit_compressed_extents(struct inode *inode,
 708                                               struct async_cow *async_cow)
 709 {
 710         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 711         struct async_extent *async_extent;
 712         u64 alloc_hint = 0;
 713         struct btrfs_key ins;
 714         struct extent_map *em;
 715         struct btrfs_root *root = BTRFS_I(inode)->root;
 716         struct extent_io_tree *io_tree;
 717         int ret = 0;
 718
 719 again:
 720         while (!list_empty(&async_cow->extents)) {
 721                 async_extent = list_entry(async_cow->extents.next,
 722                                           struct async_extent, list);
 723                 list_del(&async_extent->list);
 724
 725                 io_tree = &BTRFS_I(inode)->io_tree;
 726
 727 retry:
 728                 /* did the compression code fall back to uncompressed IO? */
 729                 if (!async_extent->pages) {
 730                         int page_started = 0;
 731                         unsigned long nr_written = 0;
 732
 733                         lock_extent(io_tree, async_extent->start,
 734                                          async_extent->start +
 735                                          async_extent->ram_size - 1);
 736
 737                         /* allocate blocks */
 738                         ret = cow_file_range(inode, async_cow->locked_page,
 739                                              async_extent->start,
 740                                              async_extent->start +
 741                                              async_extent->ram_size - 1,
 742                                              async_extent->start +
 743                                              async_extent->ram_size - 1,
 744                                              &page_started, &nr_written, 0,
 745                                              NULL);
 746
 747                         /* JDM XXX */
 748
 749                         /*
 750                          * if page_started, cow_file_range inserted an
 751                          * inline extent and took care of all the unlocking
 752                          * and IO for us.  Otherwise, we need to submit
 753                          * all those pages down to the drive.
 754                          */
 755                         if (!page_started && !ret)
 756                                 extent_write_locked_range(inode,
 757                                                   async_extent->start,
 758                                                   async_extent->start +
 759                                                   async_extent->ram_size - 1,
 760                                                   WB_SYNC_ALL);
 761                         else if (ret)
 762                                 unlock_page(async_cow->locked_page);
 763                         kfree(async_extent);
 764                         cond_resched();
 765                         continue;
 766                 }
 767
 768                 lock_extent(io_tree, async_extent->start,
 769                             async_extent->start + async_extent->ram_size - 1);
 770
 771                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 772                                            async_extent->compressed_size,
 773                                            async_extent->compressed_size,
 774                                            0, alloc_hint, &ins, 1, 1);
 775                 if (ret) {
 776                         free_async_extent_pages(async_extent);
 777
 778                         if (ret == -ENOSPC) {
 779                                 unlock_extent(io_tree, async_extent->start,
 780                                               async_extent->start +
 781                                               async_extent->ram_size - 1);
 782
 783                                 /*
 784                                  * we need to redirty the pages if we decide to
 785                                  * fallback to uncompressed IO, otherwise we
 786                                  * will not submit these pages down to lower
 787                                  * layers.
 788                                  */
 789                                 extent_range_redirty_for_io(inode,
 790                                                 async_extent->start,
 791                                                 async_extent->start +
 792                                                 async_extent->ram_size - 1);
 793
 794                                 goto retry;
 795                         }
 796                         goto out_free;
 797                 }
 798                 /*
 799                  * here we're doing allocation and writeback of the
 800                  * compressed pages
 801                  */
 802                 em = create_io_em(inode, async_extent->start,
 803                                   async_extent->ram_size, /* len */
 804                                   async_extent->start, /* orig_start */
 805                                   ins.objectid, /* block_start */
 806                                   ins.offset, /* block_len */
 807                                   ins.offset, /* orig_block_len */
 808                                   async_extent->ram_size, /* ram_bytes */
 809                                   async_extent->compress_type,
 810                                   BTRFS_ORDERED_COMPRESSED);
 811                 if (IS_ERR(em))
 812                         /* ret value is not necessary due to void function */
 813                         goto out_free_reserve;
 814                 free_extent_map(em);
 815
 816                 ret = btrfs_add_ordered_extent_compress(inode,
 817                                                 async_extent->start,
 818                                                 ins.objectid,
 819                                                 async_extent->ram_size,
 820                                                 ins.offset,
 821                                                 BTRFS_ORDERED_COMPRESSED,
 822                                                 async_extent->compress_type);
 823                 if (ret) {
 824                         btrfs_drop_extent_cache(BTRFS_I(inode),
 825                                                 async_extent->start,
 826                                                 async_extent->start +
 827                                                 async_extent->ram_size - 1, 0);
 828                         goto out_free_reserve;
 829                 }
 830                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 831
 832                 /*
 833                  * clear dirty, set writeback and unlock the pages.
 834                  */
 835                 extent_clear_unlock_delalloc(inode, async_extent->start,
 836                                 async_extent->start +
 837                                 async_extent->ram_size - 1,
 838                                 async_extent->start +
 839                                 async_extent->ram_size - 1,
 840                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 841                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 842                                 PAGE_SET_WRITEBACK);
 843                 if (btrfs_submit_compressed_write(inode,
 844                                     async_extent->start,
 845                                     async_extent->ram_size,
 846                                     ins.objectid,
 847                                     ins.offset, async_extent->pages,
 848                                     async_extent->nr_pages,
 849                                     async_cow->write_flags)) {
 850                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 851                         struct page *p = async_extent->pages[0];
 852                         const u64 start = async_extent->start;
 853                         const u64 end = start + async_extent->ram_size - 1;
 854
 855                         p->mapping = inode->i_mapping;
 856                         tree->ops->writepage_end_io_hook(p, start, end,
 857                                                          NULL, 0);
 858                         p->mapping = NULL;
 859                         extent_clear_unlock_delalloc(inode, start, end, end,
 860                                                      NULL, 0,
 861                                                      PAGE_END_WRITEBACK |
 862                                                      PAGE_SET_ERROR);
 863                         free_async_extent_pages(async_extent);
 864                 }
 865                 alloc_hint = ins.objectid + ins.offset;
 866                 kfree(async_extent);
 867                 cond_resched();
 868         }
 869         return;
 870 out_free_reserve:
 871         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 872         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 873 out_free:
 874         extent_clear_unlock_delalloc(inode, async_extent->start,
 875                                      async_extent->start +
 876                                      async_extent->ram_size - 1,
 877                                      async_extent->start +
 878                                      async_extent->ram_size - 1,
 879                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 880                                      EXTENT_DELALLOC_NEW |
 881                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 882                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 883                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 884                                      PAGE_SET_ERROR);
 885         free_async_extent_pages(async_extent);
 886         kfree(async_extent);
 887         goto again;
 888 }
 889
 890 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 891                                       u64 num_bytes)
 892 {
 893         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 894         struct extent_map *em;
 895         u64 alloc_hint = 0;
 896
 897         read_lock(&em_tree->lock);
 898         em = search_extent_mapping(em_tree, start, num_bytes);
 899         if (em) {
 900                 /*
 901                  * if block start isn't an actual block number then find the
 902                  * first block in this inode and use that as a hint.  If that
 903                  * block is also bogus then just don't worry about it.
 904                  */
 905                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 906                         free_extent_map(em);
 907                         em = search_extent_mapping(em_tree, 0, 0);
 908                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 909                                 alloc_hint = em->block_start;
 910                         if (em)
 911                                 free_extent_map(em);
 912                 } else {
 913                         alloc_hint = em->block_start;
 914                         free_extent_map(em);
 915                 }
 916         }
 917         read_unlock(&em_tree->lock);
 918
 919         return alloc_hint;
 920 }
 921
 922 /*
 923  * when extent_io.c finds a delayed allocation range in the file,
 924  * the call backs end up in this code.  The basic idea is to
 925  * allocate extents on disk for the range, and create ordered data structs
 926  * in ram to track those extents.
 927  *
 928  * locked_page is the page that writepage had locked already.  We use
 929  * it to make sure we don't do extra locks or unlocks.
 930  *
 931  * *page_started is set to one if we unlock locked_page and do everything
 932  * required to start IO on it.  It may be clean and already done with
 933  * IO when we return.
 934  */
 935 static noinline int cow_file_range(struct inode *inode,
 936                                    struct page *locked_page,
 937                                    u64 start, u64 end, u64 delalloc_end,
 938                                    int *page_started, unsigned long *nr_written,
 939                                    int unlock, struct btrfs_dedupe_hash *hash)
 940 {
 941         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 942         struct btrfs_root *root = BTRFS_I(inode)->root;
 943         u64 alloc_hint = 0;
 944         u64 num_bytes;
 945         unsigned long ram_size;
 946         u64 cur_alloc_size = 0;
 947         u64 blocksize = fs_info->sectorsize;
 948         struct btrfs_key ins;
 949         struct extent_map *em;
 950         unsigned clear_bits;
 951         unsigned long page_ops;
 952         bool extent_reserved = false;
 953         int ret = 0;
 954
 955         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 956                 WARN_ON_ONCE(1);
 957                 ret = -EINVAL;
 958                 goto out_unlock;
 959         }
 960
 961         num_bytes = ALIGN(end - start + 1, blocksize);
 962         num_bytes = max(blocksize,  num_bytes);
 963         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 964
 965         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 966
 967         if (start == 0) {
 968                 /* lets try to make an inline extent */
 969                 ret = cow_file_range_inline(inode, start, end, 0,
 970                                             BTRFS_COMPRESS_NONE, NULL);
 971                 if (ret == 0) {
 972                         /*
 973                          * We use DO_ACCOUNTING here because we need the
 974                          * delalloc_release_metadata to be run _after_ we drop
 975                          * our outstanding extent for clearing delalloc for this
 976                          * range.
 977                          */
 978                         extent_clear_unlock_delalloc(inode, start, end,
 979                                      delalloc_end, NULL,
 980                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 981                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 982                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 983                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 984                                      PAGE_END_WRITEBACK);
 985                         *nr_written = *nr_written +
 986                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 987                         *page_started = 1;
 988                         goto out;
 989                 } else if (ret < 0) {
 990                         goto out_unlock;
 991                 }
 992         }
 993
 994         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 995         btrfs_drop_extent_cache(BTRFS_I(inode), start,
 996                         start + num_bytes - 1, 0);
 997
 998         while (num_bytes > 0) {
 999                 cur_alloc_size = num_bytes;
1000                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1001                                            fs_info->sectorsize, 0, alloc_hint,
1002                                            &ins, 1, 1);
1003                 if (ret < 0)
1004                         goto out_unlock;
1005                 cur_alloc_size = ins.offset;
1006                 extent_reserved = true;
1007
1008                 ram_size = ins.offset;
1009                 em = create_io_em(inode, start, ins.offset, /* len */
1010                                   start, /* orig_start */
1011                                   ins.objectid, /* block_start */
1012                                   ins.offset, /* block_len */
1013                                   ins.offset, /* orig_block_len */
1014                                   ram_size, /* ram_bytes */
1015                                   BTRFS_COMPRESS_NONE, /* compress_type */
1016                                   BTRFS_ORDERED_REGULAR /* type */);
1017                 if (IS_ERR(em)) {
1018                         ret = PTR_ERR(em);
1019                         goto out_reserve;
1020                 }
1021                 free_extent_map(em);
1022
1023                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1024                                                ram_size, cur_alloc_size, 0);
1025                 if (ret)
1026                         goto out_drop_extent_cache;
1027
1028                 if (root->root_key.objectid ==
1029                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1030                         ret = btrfs_reloc_clone_csums(inode, start,
1031                                                       cur_alloc_size);
1032                         /*
1033                          * Only drop cache here, and process as normal.
1034                          *
1035                          * We must not allow extent_clear_unlock_delalloc()
1036                          * at out_unlock label to free meta of this ordered
1037                          * extent, as its meta should be freed by
1038                          * btrfs_finish_ordered_io().
1039                          *
1040                          * So we must continue until @start is increased to
1041                          * skip current ordered extent.
1042                          */
1043                         if (ret)
1044                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1045                                                 start + ram_size - 1, 0);
1046                 }
1047
1048                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1049
1050                 /* we're not doing compressed IO, don't unlock the first
1051                  * page (which the caller expects to stay locked), don't
1052                  * clear any dirty bits and don't set any writeback bits
1053                  *
1054                  * Do set the Private2 bit so we know this page was properly
1055                  * setup for writepage
1056                  */
1057                 page_ops = unlock ? PAGE_UNLOCK : 0;
1058                 page_ops |= PAGE_SET_PRIVATE2;
1059
1060                 extent_clear_unlock_delalloc(inode, start,
1061                                              start + ram_size - 1,
1062                                              delalloc_end, locked_page,
1063                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1064                                              page_ops);
1065                 if (num_bytes < cur_alloc_size)
1066                         num_bytes = 0;
1067                 else
1068                         num_bytes -= cur_alloc_size;
1069                 alloc_hint = ins.objectid + ins.offset;
1070                 start += cur_alloc_size;
1071                 extent_reserved = false;
1072
1073                 /*
1074                  * btrfs_reloc_clone_csums() error, since start is increased
1075                  * extent_clear_unlock_delalloc() at out_unlock label won't
1076                  * free metadata of current ordered extent, we're OK to exit.
1077                  */
1078                 if (ret)
1079                         goto out_unlock;
1080         }
1081 out:
1082         return ret;
1083
1084 out_drop_extent_cache:
1085         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1086 out_reserve:
1087         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1088         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1089 out_unlock:
1090         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1091                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1092         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1093                 PAGE_END_WRITEBACK;
1094         /*
1095          * If we reserved an extent for our delalloc range (or a subrange) and
1096          * failed to create the respective ordered extent, then it means that
1097          * when we reserved the extent we decremented the extent's size from
1098          * the data space_info's bytes_may_use counter and incremented the
1099          * space_info's bytes_reserved counter by the same amount. We must make
1100          * sure extent_clear_unlock_delalloc() does not try to decrement again
1101          * the data space_info's bytes_may_use counter, therefore we do not pass
1102          * it the flag EXTENT_CLEAR_DATA_RESV.
1103          */
1104         if (extent_reserved) {
1105                 extent_clear_unlock_delalloc(inode, start,
1106                                              start + cur_alloc_size,
1107                                              start + cur_alloc_size,
1108                                              locked_page,
1109                                              clear_bits,
1110                                              page_ops);
1111                 start += cur_alloc_size;
1112                 if (start >= end)
1113                         goto out;
1114         }
1115         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1116                                      locked_page,
1117                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1118                                      page_ops);
1119         goto out;
1120 }
1121
1122 /*
1123  * work queue call back to started compression on a file and pages
1124  */
1125 static noinline void async_cow_start(struct btrfs_work *work)
1126 {
1127         struct async_cow *async_cow;
1128         int num_added = 0;
1129         async_cow = container_of(work, struct async_cow, work);
1130
1131         compress_file_range(async_cow->inode, async_cow->locked_page,
1132                             async_cow->start, async_cow->end, async_cow,
1133                             &num_added);
1134         if (num_added == 0) {
1135                 btrfs_add_delayed_iput(async_cow->inode);
1136                 async_cow->inode = NULL;
1137         }
1138 }
1139
1140 /*
1141  * work queue call back to submit previously compressed pages
1142  */
1143 static noinline void async_cow_submit(struct btrfs_work *work)
1144 {
1145         struct btrfs_fs_info *fs_info;
1146         struct async_cow *async_cow;
1147         struct btrfs_root *root;
1148         unsigned long nr_pages;
1149
1150         async_cow = container_of(work, struct async_cow, work);
1151
1152         root = async_cow->root;
1153         fs_info = root->fs_info;
1154         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1155                 PAGE_SHIFT;
1156
1157         /* atomic_sub_return implies a barrier */
1158         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1159             5 * SZ_1M)
1160                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1161
1162         if (async_cow->inode)
1163                 submit_compressed_extents(async_cow->inode, async_cow);
1164 }
1165
1166 static noinline void async_cow_free(struct btrfs_work *work)
1167 {
1168         struct async_cow *async_cow;
1169         async_cow = container_of(work, struct async_cow, work);
1170         if (async_cow->inode)
1171                 btrfs_add_delayed_iput(async_cow->inode);
1172         kfree(async_cow);
1173 }
1174
1175 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1176                                 u64 start, u64 end, int *page_started,
1177                                 unsigned long *nr_written,
1178                                 unsigned int write_flags)
1179 {
1180         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1181         struct async_cow *async_cow;
1182         struct btrfs_root *root = BTRFS_I(inode)->root;
1183         unsigned long nr_pages;
1184         u64 cur_end;
1185
1186         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1187                          1, 0, NULL);
1188         while (start < end) {
1189                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1190                 BUG_ON(!async_cow); /* -ENOMEM */
1191                 async_cow->inode = igrab(inode);
1192                 async_cow->root = root;
1193                 async_cow->locked_page = locked_page;
1194                 async_cow->start = start;
1195                 async_cow->write_flags = write_flags;
1196
1197                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1198                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1199                         cur_end = end;
1200                 else
1201                         cur_end = min(end, start + SZ_512K - 1);
1202
1203                 async_cow->end = cur_end;
1204                 INIT_LIST_HEAD(&async_cow->extents);
1205
1206                 btrfs_init_work(&async_cow->work,
1207                                 btrfs_delalloc_helper,
1208                                 async_cow_start, async_cow_submit,
1209                                 async_cow_free);
1210
1211                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1212                         PAGE_SHIFT;
1213                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1214
1215                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1216
1217                 *nr_written += nr_pages;
1218                 start = cur_end + 1;
1219         }
1220         *page_started = 1;
1221         return 0;
1222 }
1223
1224 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1225                                         u64 bytenr, u64 num_bytes)
1226 {
1227         int ret;
1228         struct btrfs_ordered_sum *sums;
1229         LIST_HEAD(list);
1230
1231         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1232                                        bytenr + num_bytes - 1, &list, 0);
1233         if (ret == 0 && list_empty(&list))
1234                 return 0;
1235
1236         while (!list_empty(&list)) {
1237                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1238                 list_del(&sums->list);
1239                 kfree(sums);
1240         }
1241         if (ret < 0)
1242                 return ret;
1243         return 1;
1244 }
1245
1246 /*
1247  * when nowcow writeback call back.  This checks for snapshots or COW copies
1248  * of the extents that exist in the file, and COWs the file as required.
1249  *
1250  * If no cow copies or snapshots exist, we write directly to the existing
1251  * blocks on disk
1252  */
1253 static noinline int run_delalloc_nocow(struct inode *inode,
1254                                        struct page *locked_page,
1255                               u64 start, u64 end, int *page_started, int force,
1256                               unsigned long *nr_written)
1257 {
1258         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1259         struct btrfs_root *root = BTRFS_I(inode)->root;
1260         struct extent_buffer *leaf;
1261         struct btrfs_path *path;
1262         struct btrfs_file_extent_item *fi;
1263         struct btrfs_key found_key;
1264         struct extent_map *em;
1265         u64 cow_start;
1266         u64 cur_offset;
1267         u64 extent_end;
1268         u64 extent_offset;
1269         u64 disk_bytenr;
1270         u64 num_bytes;
1271         u64 disk_num_bytes;
1272         u64 ram_bytes;
1273         int extent_type;
1274         int ret, err;
1275         int type;
1276         int nocow;
1277         int check_prev = 1;
1278         bool nolock;
1279         u64 ino = btrfs_ino(BTRFS_I(inode));
1280
1281         path = btrfs_alloc_path();
1282         if (!path) {
1283                 extent_clear_unlock_delalloc(inode, start, end, end,
1284                                              locked_page,
1285                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1286                                              EXTENT_DO_ACCOUNTING |
1287                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1288                                              PAGE_CLEAR_DIRTY |
1289                                              PAGE_SET_WRITEBACK |
1290                                              PAGE_END_WRITEBACK);
1291                 return -ENOMEM;
1292         }
1293
1294         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1295
1296         cow_start = (u64)-1;
1297         cur_offset = start;
1298         while (1) {
1299                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1300                                                cur_offset, 0);
1301                 if (ret < 0)
1302                         goto error;
1303                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1304                         leaf = path->nodes[0];
1305                         btrfs_item_key_to_cpu(leaf, &found_key,
1306                                               path->slots[0] - 1);
1307                         if (found_key.objectid == ino &&
1308                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1309                                 path->slots[0]--;
1310                 }
1311                 check_prev = 0;
1312 next_slot:
1313                 leaf = path->nodes[0];
1314                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1315                         ret = btrfs_next_leaf(root, path);
1316                         if (ret < 0) {
1317                                 if (cow_start != (u64)-1)
1318                                         cur_offset = cow_start;
1319                                 goto error;
1320                         }
1321                         if (ret > 0)
1322                                 break;
1323                         leaf = path->nodes[0];
1324                 }
1325
1326                 nocow = 0;
1327                 disk_bytenr = 0;
1328                 num_bytes = 0;
1329                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1330
1331                 if (found_key.objectid > ino)
1332                         break;
1333                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1334                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1335                         path->slots[0]++;
1336                         goto next_slot;
1337                 }
1338                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1339                     found_key.offset > end)
1340                         break;
1341
1342                 if (found_key.offset > cur_offset) {
1343                         extent_end = found_key.offset;
1344                         extent_type = 0;
1345                         goto out_check;
1346                 }
1347
1348                 fi = btrfs_item_ptr(leaf, path->slots[0],
1349                                     struct btrfs_file_extent_item);
1350                 extent_type = btrfs_file_extent_type(leaf, fi);
1351
1352                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1353                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1354                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1355                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1356                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1357                         extent_end = found_key.offset +
1358                                 btrfs_file_extent_num_bytes(leaf, fi);
1359                         disk_num_bytes =
1360                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1361                         if (extent_end <= start) {
1362                                 path->slots[0]++;
1363                                 goto next_slot;
1364                         }
1365                         if (disk_bytenr == 0)
1366                                 goto out_check;
1367                         if (btrfs_file_extent_compression(leaf, fi) ||
1368                             btrfs_file_extent_encryption(leaf, fi) ||
1369                             btrfs_file_extent_other_encoding(leaf, fi))
1370                                 goto out_check;
1371                         /*
1372                          * Do the same check as in btrfs_cross_ref_exist but
1373                          * without the unnecessary search.
1374                          */
1375                         if (btrfs_file_extent_generation(leaf, fi) <=
1376                             btrfs_root_last_snapshot(&root->root_item))
1377                                 goto out_check;
1378                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1379                                 goto out_check;
1380                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1381                                 goto out_check;
1382                         ret = btrfs_cross_ref_exist(root, ino,
1383                                                     found_key.offset -
1384                                                     extent_offset, disk_bytenr);
1385                         if (ret) {
1386                                 /*
1387                                  * ret could be -EIO if the above fails to read
1388                                  * metadata.
1389                                  */
1390                                 if (ret < 0) {
1391                                         if (cow_start != (u64)-1)
1392                                                 cur_offset = cow_start;
1393                                         goto error;
1394                                 }
1395
1396                                 WARN_ON_ONCE(nolock);
1397                                 goto out_check;
1398                         }
1399                         disk_bytenr += extent_offset;
1400                         disk_bytenr += cur_offset - found_key.offset;
1401                         num_bytes = min(end + 1, extent_end) - cur_offset;
1402                         /*
1403                          * if there are pending snapshots for this root,
1404                          * we fall into common COW way.
1405                          */
1406                         if (!nolock) {
1407                                 err = btrfs_start_write_no_snapshotting(root);
1408                                 if (!err)
1409                                         goto out_check;
1410                         }
1411                         /*
1412                          * force cow if csum exists in the range.
1413                          * this ensure that csum for a given extent are
1414                          * either valid or do not exist.
1415                          */
1416                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1417                                                   num_bytes);
1418                         if (ret) {
1419                                 if (!nolock)
1420                                         btrfs_end_write_no_snapshotting(root);
1421
1422                                 /*
1423                                  * ret could be -EIO if the above fails to read
1424                                  * metadata.
1425                                  */
1426                                 if (ret < 0) {
1427                                         if (cow_start != (u64)-1)
1428                                                 cur_offset = cow_start;
1429                                         goto error;
1430                                 }
1431                                 WARN_ON_ONCE(nolock);
1432                                 goto out_check;
1433                         }
1434                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1435                                 if (!nolock)
1436                                         btrfs_end_write_no_snapshotting(root);
1437                                 goto out_check;
1438                         }
1439                         nocow = 1;
1440                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1441                         extent_end = found_key.offset +
1442                                 btrfs_file_extent_ram_bytes(leaf, fi);
1443                         extent_end = ALIGN(extent_end,
1444                                            fs_info->sectorsize);
1445                 } else {
1446                         BUG_ON(1);
1447                 }
1448 out_check:
1449                 if (extent_end <= start) {
1450                         path->slots[0]++;
1451                         if (!nolock && nocow)
1452                                 btrfs_end_write_no_snapshotting(root);
1453                         if (nocow)
1454                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1455                         goto next_slot;
1456                 }
1457                 if (!nocow) {
1458                         if (cow_start == (u64)-1)
1459                                 cow_start = cur_offset;
1460                         cur_offset = extent_end;
1461                         if (cur_offset > end)
1462                                 break;
1463                         path->slots[0]++;
1464                         goto next_slot;
1465                 }
1466
1467                 btrfs_release_path(path);
1468                 if (cow_start != (u64)-1) {
1469                         ret = cow_file_range(inode, locked_page,
1470                                              cow_start, found_key.offset - 1,
1471                                              end, page_started, nr_written, 1,
1472                                              NULL);
1473                         if (ret) {
1474                                 if (!nolock && nocow)
1475                                         btrfs_end_write_no_snapshotting(root);
1476                                 if (nocow)
1477                                         btrfs_dec_nocow_writers(fs_info,
1478                                                                 disk_bytenr);
1479                                 goto error;
1480                         }
1481                         cow_start = (u64)-1;
1482                 }
1483
1484                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1485                         u64 orig_start = found_key.offset - extent_offset;
1486
1487                         em = create_io_em(inode, cur_offset, num_bytes,
1488                                           orig_start,
1489                                           disk_bytenr, /* block_start */
1490                                           num_bytes, /* block_len */
1491                                           disk_num_bytes, /* orig_block_len */
1492                                           ram_bytes, BTRFS_COMPRESS_NONE,
1493                                           BTRFS_ORDERED_PREALLOC);
1494                         if (IS_ERR(em)) {
1495                                 if (!nolock && nocow)
1496                                         btrfs_end_write_no_snapshotting(root);
1497                                 if (nocow)
1498                                         btrfs_dec_nocow_writers(fs_info,
1499                                                                 disk_bytenr);
1500                                 ret = PTR_ERR(em);
1501                                 goto error;
1502                         }
1503                         free_extent_map(em);
1504                 }
1505
1506                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1507                         type = BTRFS_ORDERED_PREALLOC;
1508                 } else {
1509                         type = BTRFS_ORDERED_NOCOW;
1510                 }
1511
1512                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1513                                                num_bytes, num_bytes, type);
1514                 if (nocow)
1515                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1516                 BUG_ON(ret); /* -ENOMEM */
1517
1518                 if (root->root_key.objectid ==
1519                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1520                         /*
1521                          * Error handled later, as we must prevent
1522                          * extent_clear_unlock_delalloc() in error handler
1523                          * from freeing metadata of created ordered extent.
1524                          */
1525                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1526                                                       num_bytes);
1527
1528                 extent_clear_unlock_delalloc(inode, cur_offset,
1529                                              cur_offset + num_bytes - 1, end,
1530                                              locked_page, EXTENT_LOCKED |
1531                                              EXTENT_DELALLOC |
1532                                              EXTENT_CLEAR_DATA_RESV,
1533                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1534
1535                 if (!nolock && nocow)
1536                         btrfs_end_write_no_snapshotting(root);
1537                 cur_offset = extent_end;
1538
1539                 /*
1540                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1541                  * handler, as metadata for created ordered extent will only
1542                  * be freed by btrfs_finish_ordered_io().
1543                  */
1544                 if (ret)
1545                         goto error;
1546                 if (cur_offset > end)
1547                         break;
1548         }
1549         btrfs_release_path(path);
1550
1551         if (cur_offset <= end && cow_start == (u64)-1) {
1552                 cow_start = cur_offset;
1553                 cur_offset = end;
1554         }
1555
1556         if (cow_start != (u64)-1) {
1557                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1558                                      page_started, nr_written, 1, NULL);
1559                 if (ret)
1560                         goto error;
1561         }
1562
1563 error:
1564         if (ret && cur_offset < end)
1565                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1566                                              locked_page, EXTENT_LOCKED |
1567                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1568                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1569                                              PAGE_CLEAR_DIRTY |
1570                                              PAGE_SET_WRITEBACK |
1571                                              PAGE_END_WRITEBACK);
1572         btrfs_free_path(path);
1573         return ret;
1574 }
1575
1576 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1577 {
1578
1579         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1580             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1581                 return 0;
1582
1583         /*
1584          * @defrag_bytes is a hint value, no spinlock held here,
1585          * if is not zero, it means the file is defragging.
1586          * Force cow if given extent needs to be defragged.
1587          */
1588         if (BTRFS_I(inode)->defrag_bytes &&
1589             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1590                            EXTENT_DEFRAG, 0, NULL))
1591                 return 1;
1592
1593         return 0;
1594 }
1595
1596 /*
1597  * extent_io.c call back to do delayed allocation processing
1598  */
1599 static int run_delalloc_range(void *private_data, struct page *locked_page,
1600                               u64 start, u64 end, int *page_started,
1601                               unsigned long *nr_written,
1602                               struct writeback_control *wbc)
1603 {
1604         struct inode *inode = private_data;
1605         int ret;
1606         int force_cow = need_force_cow(inode, start, end);
1607         unsigned int write_flags = wbc_to_write_flags(wbc);
1608
1609         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1610                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1611                                          page_started, 1, nr_written);
1612         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1613                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1614                                          page_started, 0, nr_written);
1615         } else if (!inode_need_compress(inode, start, end)) {
1616                 ret = cow_file_range(inode, locked_page, start, end, end,
1617                                       page_started, nr_written, 1, NULL);
1618         } else {
1619                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1620                         &BTRFS_I(inode)->runtime_flags);
1621                 ret = cow_file_range_async(inode, locked_page, start, end,
1622                                            page_started, nr_written,
1623                                            write_flags);
1624         }
1625         if (ret)
1626                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1627         return ret;
1628 }
1629
1630 static void btrfs_split_extent_hook(void *private_data,
1631                                     struct extent_state *orig, u64 split)
1632 {
1633         struct inode *inode = private_data;
1634         u64 size;
1635
1636         /* not delalloc, ignore it */
1637         if (!(orig->state & EXTENT_DELALLOC))
1638                 return;
1639
1640         size = orig->end - orig->start + 1;
1641         if (size > BTRFS_MAX_EXTENT_SIZE) {
1642                 u32 num_extents;
1643                 u64 new_size;
1644
1645                 /*
1646                  * See the explanation in btrfs_merge_extent_hook, the same
1647                  * applies here, just in reverse.
1648                  */
1649                 new_size = orig->end - split + 1;
1650                 num_extents = count_max_extents(new_size);
1651                 new_size = split - orig->start;
1652                 num_extents += count_max_extents(new_size);
1653                 if (count_max_extents(size) >= num_extents)
1654                         return;
1655         }
1656
1657         spin_lock(&BTRFS_I(inode)->lock);
1658         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1659         spin_unlock(&BTRFS_I(inode)->lock);
1660 }
1661
1662 /*
1663  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1664  * extents so we can keep track of new extents that are just merged onto old
1665  * extents, such as when we are doing sequential writes, so we can properly
1666  * account for the metadata space we'll need.
1667  */
1668 static void btrfs_merge_extent_hook(void *private_data,
1669                                     struct extent_state *new,
1670                                     struct extent_state *other)
1671 {
1672         struct inode *inode = private_data;
1673         u64 new_size, old_size;
1674         u32 num_extents;
1675
1676         /* not delalloc, ignore it */
1677         if (!(other->state & EXTENT_DELALLOC))
1678                 return;
1679
1680         if (new->start > other->start)
1681                 new_size = new->end - other->start + 1;
1682         else
1683                 new_size = other->end - new->start + 1;
1684
1685         /* we're not bigger than the max, unreserve the space and go */
1686         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1687                 spin_lock(&BTRFS_I(inode)->lock);
1688                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1689                 spin_unlock(&BTRFS_I(inode)->lock);
1690                 return;
1691         }
1692
1693         /*
1694          * We have to add up either side to figure out how many extents were
1695          * accounted for before we merged into one big extent.  If the number of
1696          * extents we accounted for is <= the amount we need for the new range
1697          * then we can return, otherwise drop.  Think of it like this
1698          *
1699          * [ 4k][MAX_SIZE]
1700          *
1701          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1702          * need 2 outstanding extents, on one side we have 1 and the other side
1703          * we have 1 so they are == and we can return.  But in this case
1704          *
1705          * [MAX_SIZE+4k][MAX_SIZE+4k]
1706          *
1707          * Each range on their own accounts for 2 extents, but merged together
1708          * they are only 3 extents worth of accounting, so we need to drop in
1709          * this case.
1710          */
1711         old_size = other->end - other->start + 1;
1712         num_extents = count_max_extents(old_size);
1713         old_size = new->end - new->start + 1;
1714         num_extents += count_max_extents(old_size);
1715         if (count_max_extents(new_size) >= num_extents)
1716                 return;
1717
1718         spin_lock(&BTRFS_I(inode)->lock);
1719         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1720         spin_unlock(&BTRFS_I(inode)->lock);
1721 }
1722
1723 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1724                                       struct inode *inode)
1725 {
1726         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1727
1728         spin_lock(&root->delalloc_lock);
1729         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1730                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1731                               &root->delalloc_inodes);
1732                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1733                         &BTRFS_I(inode)->runtime_flags);
1734                 root->nr_delalloc_inodes++;
1735                 if (root->nr_delalloc_inodes == 1) {
1736                         spin_lock(&fs_info->delalloc_root_lock);
1737                         BUG_ON(!list_empty(&root->delalloc_root));
1738                         list_add_tail(&root->delalloc_root,
1739                                       &fs_info->delalloc_roots);
1740                         spin_unlock(&fs_info->delalloc_root_lock);
1741                 }
1742         }
1743         spin_unlock(&root->delalloc_lock);
1744 }
1745
1746
1747 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1748                                 struct btrfs_inode *inode)
1749 {
1750         struct btrfs_fs_info *fs_info = root->fs_info;
1751
1752         if (!list_empty(&inode->delalloc_inodes)) {
1753                 list_del_init(&inode->delalloc_inodes);
1754                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1755                           &inode->runtime_flags);
1756                 root->nr_delalloc_inodes--;
1757                 if (!root->nr_delalloc_inodes) {
1758                         ASSERT(list_empty(&root->delalloc_inodes));
1759                         spin_lock(&fs_info->delalloc_root_lock);
1760                         BUG_ON(list_empty(&root->delalloc_root));
1761                         list_del_init(&root->delalloc_root);
1762                         spin_unlock(&fs_info->delalloc_root_lock);
1763                 }
1764         }
1765 }
1766
1767 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1768                                      struct btrfs_inode *inode)
1769 {
1770         spin_lock(&root->delalloc_lock);
1771         __btrfs_del_delalloc_inode(root, inode);
1772         spin_unlock(&root->delalloc_lock);
1773 }
1774
1775 /*
1776  * extent_io.c set_bit_hook, used to track delayed allocation
1777  * bytes in this file, and to maintain the list of inodes that
1778  * have pending delalloc work to be done.
1779  */
1780 static void btrfs_set_bit_hook(void *private_data,
1781                                struct extent_state *state, unsigned *bits)
1782 {
1783         struct inode *inode = private_data;
1784
1785         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1786
1787         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1788                 WARN_ON(1);
1789         /*
1790          * set_bit and clear bit hooks normally require _irqsave/restore
1791          * but in this case, we are only testing for the DELALLOC
1792          * bit, which is only set or cleared with irqs on
1793          */
1794         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1795                 struct btrfs_root *root = BTRFS_I(inode)->root;
1796                 u64 len = state->end + 1 - state->start;
1797                 u32 num_extents = count_max_extents(len);
1798                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1799
1800                 spin_lock(&BTRFS_I(inode)->lock);
1801                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1802                 spin_unlock(&BTRFS_I(inode)->lock);
1803
1804                 /* For sanity tests */
1805                 if (btrfs_is_testing(fs_info))
1806                         return;
1807
1808                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1809                                          fs_info->delalloc_batch);
1810                 spin_lock(&BTRFS_I(inode)->lock);
1811                 BTRFS_I(inode)->delalloc_bytes += len;
1812                 if (*bits & EXTENT_DEFRAG)
1813                         BTRFS_I(inode)->defrag_bytes += len;
1814                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1815                                          &BTRFS_I(inode)->runtime_flags))
1816                         btrfs_add_delalloc_inodes(root, inode);
1817                 spin_unlock(&BTRFS_I(inode)->lock);
1818         }
1819
1820         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1821             (*bits & EXTENT_DELALLOC_NEW)) {
1822                 spin_lock(&BTRFS_I(inode)->lock);
1823                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1824                         state->start;
1825                 spin_unlock(&BTRFS_I(inode)->lock);
1826         }
1827 }
1828
1829 /*
1830  * extent_io.c clear_bit_hook, see set_bit_hook for why
1831  */
1832 static void btrfs_clear_bit_hook(void *private_data,
1833                                  struct extent_state *state,
1834                                  unsigned *bits)
1835 {
1836         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1837         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1838         u64 len = state->end + 1 - state->start;
1839         u32 num_extents = count_max_extents(len);
1840
1841         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1842                 spin_lock(&inode->lock);
1843                 inode->defrag_bytes -= len;
1844                 spin_unlock(&inode->lock);
1845         }
1846
1847         /*
1848          * set_bit and clear bit hooks normally require _irqsave/restore
1849          * but in this case, we are only testing for the DELALLOC
1850          * bit, which is only set or cleared with irqs on
1851          */
1852         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1853                 struct btrfs_root *root = inode->root;
1854                 bool do_list = !btrfs_is_free_space_inode(inode);
1855
1856                 spin_lock(&inode->lock);
1857                 btrfs_mod_outstanding_extents(inode, -num_extents);
1858                 spin_unlock(&inode->lock);
1859
1860                 /*
1861                  * We don't reserve metadata space for space cache inodes so we
1862                  * don't need to call dellalloc_release_metadata if there is an
1863                  * error.
1864                  */
1865                 if (*bits & EXTENT_CLEAR_META_RESV &&
1866                     root != fs_info->tree_root)
1867                         btrfs_delalloc_release_metadata(inode, len, false);
1868
1869                 /* For sanity tests. */
1870                 if (btrfs_is_testing(fs_info))
1871                         return;
1872
1873                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1874                     do_list && !(state->state & EXTENT_NORESERVE) &&
1875                     (*bits & EXTENT_CLEAR_DATA_RESV))
1876                         btrfs_free_reserved_data_space_noquota(
1877                                         &inode->vfs_inode,
1878                                         state->start, len);
1879
1880                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1881                                          fs_info->delalloc_batch);
1882                 spin_lock(&inode->lock);
1883                 inode->delalloc_bytes -= len;
1884                 if (do_list && inode->delalloc_bytes == 0 &&
1885                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1886                                         &inode->runtime_flags))
1887                         btrfs_del_delalloc_inode(root, inode);
1888                 spin_unlock(&inode->lock);
1889         }
1890
1891         if ((state->state & EXTENT_DELALLOC_NEW) &&
1892             (*bits & EXTENT_DELALLOC_NEW)) {
1893                 spin_lock(&inode->lock);
1894                 ASSERT(inode->new_delalloc_bytes >= len);
1895                 inode->new_delalloc_bytes -= len;
1896                 spin_unlock(&inode->lock);
1897         }
1898 }
1899
1900 /*
1901  * Merge bio hook, this must check the chunk tree to make sure we don't create
1902  * bios that span stripes or chunks
1903  *
1904  * return 1 if page cannot be merged to bio
1905  * return 0 if page can be merged to bio
1906  * return error otherwise
1907  */
1908 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1909                          size_t size, struct bio *bio,
1910                          unsigned long bio_flags)
1911 {
1912         struct inode *inode = page->mapping->host;
1913         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1914         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1915         u64 length = 0;
1916         u64 map_length;
1917         int ret;
1918
1919         if (bio_flags & EXTENT_BIO_COMPRESSED)
1920                 return 0;
1921
1922         length = bio->bi_iter.bi_size;
1923         map_length = length;
1924         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1925                               NULL, 0);
1926         if (ret < 0)
1927                 return ret;
1928         if (map_length < length + size)
1929                 return 1;
1930         return 0;
1931 }
1932
1933 /*
1934  * in order to insert checksums into the metadata in large chunks,
1935  * we wait until bio submission time.   All the pages in the bio are
1936  * checksummed and sums are attached onto the ordered extent record.
1937  *
1938  * At IO completion time the cums attached on the ordered extent record
1939  * are inserted into the btree
1940  */
1941 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
1942                                     u64 bio_offset)
1943 {
1944         struct inode *inode = private_data;
1945         blk_status_t ret = 0;
1946
1947         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1948         BUG_ON(ret); /* -ENOMEM */
1949         return 0;
1950 }
1951
1952 /*
1953  * in order to insert checksums into the metadata in large chunks,
1954  * we wait until bio submission time.   All the pages in the bio are
1955  * checksummed and sums are attached onto the ordered extent record.
1956  *
1957  * At IO completion time the cums attached on the ordered extent record
1958  * are inserted into the btree
1959  */
1960 blk_status_t btrfs_submit_bio_done(void *private_data, struct bio *bio,
1961                           int mirror_num)
1962 {
1963         struct inode *inode = private_data;
1964         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1965         blk_status_t ret;
1966
1967         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1968         if (ret) {
1969                 bio->bi_status = ret;
1970                 bio_endio(bio);
1971         }
1972         return ret;
1973 }
1974
1975 /*
1976  * extent_io.c submission hook. This does the right thing for csum calculation
1977  * on write, or reading the csums from the tree before a read.
1978  *
1979  * Rules about async/sync submit,
1980  * a) read:                             sync submit
1981  *
1982  * b) write without checksum:           sync submit
1983  *
1984  * c) write with checksum:
1985  *    c-1) if bio is issued by fsync:   sync submit
1986  *         (sync_writers != 0)
1987  *
1988  *    c-2) if root is reloc root:       sync submit
1989  *         (only in case of buffered IO)
1990  *
1991  *    c-3) otherwise:                   async submit
1992  */
1993 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1994                                  int mirror_num, unsigned long bio_flags,
1995                                  u64 bio_offset)
1996 {
1997         struct inode *inode = private_data;
1998         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1999         struct btrfs_root *root = BTRFS_I(inode)->root;
2000         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2001         blk_status_t ret = 0;
2002         int skip_sum;
2003         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2004
2005         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2006
2007         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2008                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2009
2010         if (bio_op(bio) != REQ_OP_WRITE) {
2011                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2012                 if (ret)
2013                         goto out;
2014
2015                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2016                         ret = btrfs_submit_compressed_read(inode, bio,
2017                                                            mirror_num,
2018                                                            bio_flags);
2019                         goto out;
2020                 } else if (!skip_sum) {
2021                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2022                         if (ret)
2023                                 goto out;
2024                 }
2025                 goto mapit;
2026         } else if (async && !skip_sum) {
2027                 /* csum items have already been cloned */
2028                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2029                         goto mapit;
2030                 /* we're doing a write, do the async checksumming */
2031                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2032                                           bio_offset, inode,
2033                                           btrfs_submit_bio_start);
2034                 goto out;
2035         } else if (!skip_sum) {
2036                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2037                 if (ret)
2038                         goto out;
2039         }
2040
2041 mapit:
2042         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2043
2044 out:
2045         if (ret) {
2046                 bio->bi_status = ret;
2047                 bio_endio(bio);
2048         }
2049         return ret;
2050 }
2051
2052 /*
2053  * given a list of ordered sums record them in the inode.  This happens
2054  * at IO completion time based on sums calculated at bio submission time.
2055  */
2056 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2057                              struct inode *inode, struct list_head *list)
2058 {
2059         struct btrfs_ordered_sum *sum;
2060         int ret;
2061
2062         list_for_each_entry(sum, list, list) {
2063                 trans->adding_csums = true;
2064                 ret = btrfs_csum_file_blocks(trans,
2065                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2066                 trans->adding_csums = false;
2067                 if (ret)
2068                         return ret;
2069         }
2070         return 0;
2071 }
2072
2073 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2074                               unsigned int extra_bits,
2075                               struct extent_state **cached_state, int dedupe)
2076 {
2077         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2078         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2079                                    extra_bits, cached_state);
2080 }
2081
2082 /* see btrfs_writepage_start_hook for details on why this is required */
2083 struct btrfs_writepage_fixup {
2084         struct page *page;
2085         struct btrfs_work work;
2086 };
2087
2088 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2089 {
2090         struct btrfs_writepage_fixup *fixup;
2091         struct btrfs_ordered_extent *ordered;
2092         struct extent_state *cached_state = NULL;
2093         struct extent_changeset *data_reserved = NULL;
2094         struct page *page;
2095         struct inode *inode;
2096         u64 page_start;
2097         u64 page_end;
2098         int ret;
2099
2100         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2101         page = fixup->page;
2102 again:
2103         lock_page(page);
2104         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2105                 ClearPageChecked(page);
2106                 goto out_page;
2107         }
2108
2109         inode = page->mapping->host;
2110         page_start = page_offset(page);
2111         page_end = page_offset(page) + PAGE_SIZE - 1;
2112
2113         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2114                          &cached_state);
2115
2116         /* already ordered? We're done */
2117         if (PagePrivate2(page))
2118                 goto out;
2119
2120         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2121                                         PAGE_SIZE);
2122         if (ordered) {
2123                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2124                                      page_end, &cached_state);
2125                 unlock_page(page);
2126                 btrfs_start_ordered_extent(inode, ordered, 1);
2127                 btrfs_put_ordered_extent(ordered);
2128                 goto again;
2129         }
2130
2131         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2132                                            PAGE_SIZE);
2133         if (ret) {
2134                 mapping_set_error(page->mapping, ret);
2135                 end_extent_writepage(page, ret, page_start, page_end);
2136                 ClearPageChecked(page);
2137                 goto out;
2138          }
2139
2140         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2141                                         &cached_state, 0);
2142         if (ret) {
2143                 mapping_set_error(page->mapping, ret);
2144                 end_extent_writepage(page, ret, page_start, page_end);
2145                 ClearPageChecked(page);
2146                 goto out;
2147         }
2148
2149         ClearPageChecked(page);
2150         set_page_dirty(page);
2151         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, false);
2152 out:
2153         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2154                              &cached_state);
2155 out_page:
2156         unlock_page(page);
2157         put_page(page);
2158         kfree(fixup);
2159         extent_changeset_free(data_reserved);
2160 }
2161
2162 /*
2163  * There are a few paths in the higher layers of the kernel that directly
2164  * set the page dirty bit without asking the filesystem if it is a
2165  * good idea.  This causes problems because we want to make sure COW
2166  * properly happens and the data=ordered rules are followed.
2167  *
2168  * In our case any range that doesn't have the ORDERED bit set
2169  * hasn't been properly setup for IO.  We kick off an async process
2170  * to fix it up.  The async helper will wait for ordered extents, set
2171  * the delalloc bit and make it safe to write the page.
2172  */
2173 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2174 {
2175         struct inode *inode = page->mapping->host;
2176         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2177         struct btrfs_writepage_fixup *fixup;
2178
2179         /* this page is properly in the ordered list */
2180         if (TestClearPagePrivate2(page))
2181                 return 0;
2182
2183         if (PageChecked(page))
2184                 return -EAGAIN;
2185
2186         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2187         if (!fixup)
2188                 return -EAGAIN;
2189
2190         SetPageChecked(page);
2191         get_page(page);
2192         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2193                         btrfs_writepage_fixup_worker, NULL, NULL);
2194         fixup->page = page;
2195         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2196         return -EBUSY;
2197 }
2198
2199 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2200                                        struct inode *inode, u64 file_pos,
2201                                        u64 disk_bytenr, u64 disk_num_bytes,
2202                                        u64 num_bytes, u64 ram_bytes,
2203                                        u8 compression, u8 encryption,
2204                                        u16 other_encoding, int extent_type)
2205 {
2206         struct btrfs_root *root = BTRFS_I(inode)->root;
2207         struct btrfs_file_extent_item *fi;
2208         struct btrfs_path *path;
2209         struct extent_buffer *leaf;
2210         struct btrfs_key ins;
2211         u64 qg_released;
2212         int extent_inserted = 0;
2213         int ret;
2214
2215         path = btrfs_alloc_path();
2216         if (!path)
2217                 return -ENOMEM;
2218
2219         /*
2220          * we may be replacing one extent in the tree with another.
2221          * The new extent is pinned in the extent map, and we don't want
2222          * to drop it from the cache until it is completely in the btree.
2223          *
2224          * So, tell btrfs_drop_extents to leave this extent in the cache.
2225          * the caller is expected to unpin it and allow it to be merged
2226          * with the others.
2227          */
2228         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2229                                    file_pos + num_bytes, NULL, 0,
2230                                    1, sizeof(*fi), &extent_inserted);
2231         if (ret)
2232                 goto out;
2233
2234         if (!extent_inserted) {
2235                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2236                 ins.offset = file_pos;
2237                 ins.type = BTRFS_EXTENT_DATA_KEY;
2238
2239                 path->leave_spinning = 1;
2240                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2241                                               sizeof(*fi));
2242                 if (ret)
2243                         goto out;
2244         }
2245         leaf = path->nodes[0];
2246         fi = btrfs_item_ptr(leaf, path->slots[0],
2247                             struct btrfs_file_extent_item);
2248         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2249         btrfs_set_file_extent_type(leaf, fi, extent_type);
2250         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2251         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2252         btrfs_set_file_extent_offset(leaf, fi, 0);
2253         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2254         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2255         btrfs_set_file_extent_compression(leaf, fi, compression);
2256         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2257         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2258
2259         btrfs_mark_buffer_dirty(leaf);
2260         btrfs_release_path(path);
2261
2262         inode_add_bytes(inode, num_bytes);
2263
2264         ins.objectid = disk_bytenr;
2265         ins.offset = disk_num_bytes;
2266         ins.type = BTRFS_EXTENT_ITEM_KEY;
2267
2268         /*
2269          * Release the reserved range from inode dirty range map, as it is
2270          * already moved into delayed_ref_head
2271          */
2272         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2273         if (ret < 0)
2274                 goto out;
2275         qg_released = ret;
2276         ret = btrfs_alloc_reserved_file_extent(trans, root,
2277                                                btrfs_ino(BTRFS_I(inode)),
2278                                                file_pos, qg_released, &ins);
2279 out:
2280         btrfs_free_path(path);
2281
2282         return ret;
2283 }
2284
2285 /* snapshot-aware defrag */
2286 struct sa_defrag_extent_backref {
2287         struct rb_node node;
2288         struct old_sa_defrag_extent *old;
2289         u64 root_id;
2290         u64 inum;
2291         u64 file_pos;
2292         u64 extent_offset;
2293         u64 num_bytes;
2294         u64 generation;
2295 };
2296
2297 struct old_sa_defrag_extent {
2298         struct list_head list;
2299         struct new_sa_defrag_extent *new;
2300
2301         u64 extent_offset;
2302         u64 bytenr;
2303         u64 offset;
2304         u64 len;
2305         int count;
2306 };
2307
2308 struct new_sa_defrag_extent {
2309         struct rb_root root;
2310         struct list_head head;
2311         struct btrfs_path *path;
2312         struct inode *inode;
2313         u64 file_pos;
2314         u64 len;
2315         u64 bytenr;
2316         u64 disk_len;
2317         u8 compress_type;
2318 };
2319
2320 static int backref_comp(struct sa_defrag_extent_backref *b1,
2321                         struct sa_defrag_extent_backref *b2)
2322 {
2323         if (b1->root_id < b2->root_id)
2324                 return -1;
2325         else if (b1->root_id > b2->root_id)
2326                 return 1;
2327
2328         if (b1->inum < b2->inum)
2329                 return -1;
2330         else if (b1->inum > b2->inum)
2331                 return 1;
2332
2333         if (b1->file_pos < b2->file_pos)
2334                 return -1;
2335         else if (b1->file_pos > b2->file_pos)
2336                 return 1;
2337
2338         /*
2339          * [------------------------------] ===> (a range of space)
2340          *     |<--->|   |<---->| =============> (fs/file tree A)
2341          * |<---------------------------->| ===> (fs/file tree B)
2342          *
2343          * A range of space can refer to two file extents in one tree while
2344          * refer to only one file extent in another tree.
2345          *
2346          * So we may process a disk offset more than one time(two extents in A)
2347          * and locate at the same extent(one extent in B), then insert two same
2348          * backrefs(both refer to the extent in B).
2349          */
2350         return 0;
2351 }
2352
2353 static void backref_insert(struct rb_root *root,
2354                            struct sa_defrag_extent_backref *backref)
2355 {
2356         struct rb_node **p = &root->rb_node;
2357         struct rb_node *parent = NULL;
2358         struct sa_defrag_extent_backref *entry;
2359         int ret;
2360
2361         while (*p) {
2362                 parent = *p;
2363                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2364
2365                 ret = backref_comp(backref, entry);
2366                 if (ret < 0)
2367                         p = &(*p)->rb_left;
2368                 else
2369                         p = &(*p)->rb_right;
2370         }
2371
2372         rb_link_node(&backref->node, parent, p);
2373         rb_insert_color(&backref->node, root);
2374 }
2375
2376 /*
2377  * Note the backref might has changed, and in this case we just return 0.
2378  */
2379 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2380                                        void *ctx)
2381 {
2382         struct btrfs_file_extent_item *extent;
2383         struct old_sa_defrag_extent *old = ctx;
2384         struct new_sa_defrag_extent *new = old->new;
2385         struct btrfs_path *path = new->path;
2386         struct btrfs_key key;
2387         struct btrfs_root *root;
2388         struct sa_defrag_extent_backref *backref;
2389         struct extent_buffer *leaf;
2390         struct inode *inode = new->inode;
2391         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2392         int slot;
2393         int ret;
2394         u64 extent_offset;
2395         u64 num_bytes;
2396
2397         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2398             inum == btrfs_ino(BTRFS_I(inode)))
2399                 return 0;
2400
2401         key.objectid = root_id;
2402         key.type = BTRFS_ROOT_ITEM_KEY;
2403         key.offset = (u64)-1;
2404
2405         root = btrfs_read_fs_root_no_name(fs_info, &key);
2406         if (IS_ERR(root)) {
2407                 if (PTR_ERR(root) == -ENOENT)
2408                         return 0;
2409                 WARN_ON(1);
2410                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2411                          inum, offset, root_id);
2412                 return PTR_ERR(root);
2413         }
2414
2415         key.objectid = inum;
2416         key.type = BTRFS_EXTENT_DATA_KEY;
2417         if (offset > (u64)-1 << 32)
2418                 key.offset = 0;
2419         else
2420                 key.offset = offset;
2421
2422         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2423         if (WARN_ON(ret < 0))
2424                 return ret;
2425         ret = 0;
2426
2427         while (1) {
2428                 cond_resched();
2429
2430                 leaf = path->nodes[0];
2431                 slot = path->slots[0];
2432
2433                 if (slot >= btrfs_header_nritems(leaf)) {
2434                         ret = btrfs_next_leaf(root, path);
2435                         if (ret < 0) {
2436                                 goto out;
2437                         } else if (ret > 0) {
2438                                 ret = 0;
2439                                 goto out;
2440                         }
2441                         continue;
2442                 }
2443
2444                 path->slots[0]++;
2445
2446                 btrfs_item_key_to_cpu(leaf, &key, slot);
2447
2448                 if (key.objectid > inum)
2449                         goto out;
2450
2451                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2452                         continue;
2453
2454                 extent = btrfs_item_ptr(leaf, slot,
2455                                         struct btrfs_file_extent_item);
2456
2457                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2458                         continue;
2459
2460                 /*
2461                  * 'offset' refers to the exact key.offset,
2462                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2463                  * (key.offset - extent_offset).
2464                  */
2465                 if (key.offset != offset)
2466                         continue;
2467
2468                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2469                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2470
2471                 if (extent_offset >= old->extent_offset + old->offset +
2472                     old->len || extent_offset + num_bytes <=
2473                     old->extent_offset + old->offset)
2474                         continue;
2475                 break;
2476         }
2477
2478         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2479         if (!backref) {
2480                 ret = -ENOENT;
2481                 goto out;
2482         }
2483
2484         backref->root_id = root_id;
2485         backref->inum = inum;
2486         backref->file_pos = offset;
2487         backref->num_bytes = num_bytes;
2488         backref->extent_offset = extent_offset;
2489         backref->generation = btrfs_file_extent_generation(leaf, extent);
2490         backref->old = old;
2491         backref_insert(&new->root, backref);
2492         old->count++;
2493 out:
2494         btrfs_release_path(path);
2495         WARN_ON(ret);
2496         return ret;
2497 }
2498
2499 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2500                                    struct new_sa_defrag_extent *new)
2501 {
2502         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2503         struct old_sa_defrag_extent *old, *tmp;
2504         int ret;
2505
2506         new->path = path;
2507
2508         list_for_each_entry_safe(old, tmp, &new->head, list) {
2509                 ret = iterate_inodes_from_logical(old->bytenr +
2510                                                   old->extent_offset, fs_info,
2511                                                   path, record_one_backref,
2512                                                   old, false);
2513                 if (ret < 0 && ret != -ENOENT)
2514                         return false;
2515
2516                 /* no backref to be processed for this extent */
2517                 if (!old->count) {
2518                         list_del(&old->list);
2519                         kfree(old);
2520                 }
2521         }
2522
2523         if (list_empty(&new->head))
2524                 return false;
2525
2526         return true;
2527 }
2528
2529 static int relink_is_mergable(struct extent_buffer *leaf,
2530                               struct btrfs_file_extent_item *fi,
2531                               struct new_sa_defrag_extent *new)
2532 {
2533         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2534                 return 0;
2535
2536         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2537                 return 0;
2538
2539         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2540                 return 0;
2541
2542         if (btrfs_file_extent_encryption(leaf, fi) ||
2543             btrfs_file_extent_other_encoding(leaf, fi))
2544                 return 0;
2545
2546         return 1;
2547 }
2548
2549 /*
2550  * Note the backref might has changed, and in this case we just return 0.
2551  */
2552 static noinline int relink_extent_backref(struct btrfs_path *path,
2553                                  struct sa_defrag_extent_backref *prev,
2554                                  struct sa_defrag_extent_backref *backref)
2555 {
2556         struct btrfs_file_extent_item *extent;
2557         struct btrfs_file_extent_item *item;
2558         struct btrfs_ordered_extent *ordered;
2559         struct btrfs_trans_handle *trans;
2560         struct btrfs_root *root;
2561         struct btrfs_key key;
2562         struct extent_buffer *leaf;
2563         struct old_sa_defrag_extent *old = backref->old;
2564         struct new_sa_defrag_extent *new = old->new;
2565         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2566         struct inode *inode;
2567         struct extent_state *cached = NULL;
2568         int ret = 0;
2569         u64 start;
2570         u64 len;
2571         u64 lock_start;
2572         u64 lock_end;
2573         bool merge = false;
2574         int index;
2575
2576         if (prev && prev->root_id == backref->root_id &&
2577             prev->inum == backref->inum &&
2578             prev->file_pos + prev->num_bytes == backref->file_pos)
2579                 merge = true;
2580
2581         /* step 1: get root */
2582         key.objectid = backref->root_id;
2583         key.type = BTRFS_ROOT_ITEM_KEY;
2584         key.offset = (u64)-1;
2585
2586         index = srcu_read_lock(&fs_info->subvol_srcu);
2587
2588         root = btrfs_read_fs_root_no_name(fs_info, &key);
2589         if (IS_ERR(root)) {
2590                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2591                 if (PTR_ERR(root) == -ENOENT)
2592                         return 0;
2593                 return PTR_ERR(root);
2594         }
2595
2596         if (btrfs_root_readonly(root)) {
2597                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2598                 return 0;
2599         }
2600
2601         /* step 2: get inode */
2602         key.objectid = backref->inum;
2603         key.type = BTRFS_INODE_ITEM_KEY;
2604         key.offset = 0;
2605
2606         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2607         if (IS_ERR(inode)) {
2608                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2609                 return 0;
2610         }
2611
2612         srcu_read_unlock(&fs_info->subvol_srcu, index);
2613
2614         /* step 3: relink backref */
2615         lock_start = backref->file_pos;
2616         lock_end = backref->file_pos + backref->num_bytes - 1;
2617         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2618                          &cached);
2619
2620         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2621         if (ordered) {
2622                 btrfs_put_ordered_extent(ordered);
2623                 goto out_unlock;
2624         }
2625
2626         trans = btrfs_join_transaction(root);
2627         if (IS_ERR(trans)) {
2628                 ret = PTR_ERR(trans);
2629                 goto out_unlock;
2630         }
2631
2632         key.objectid = backref->inum;
2633         key.type = BTRFS_EXTENT_DATA_KEY;
2634         key.offset = backref->file_pos;
2635
2636         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2637         if (ret < 0) {
2638                 goto out_free_path;
2639         } else if (ret > 0) {
2640                 ret = 0;
2641                 goto out_free_path;
2642         }
2643
2644         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2645                                 struct btrfs_file_extent_item);
2646
2647         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2648             backref->generation)
2649                 goto out_free_path;
2650
2651         btrfs_release_path(path);
2652
2653         start = backref->file_pos;
2654         if (backref->extent_offset < old->extent_offset + old->offset)
2655                 start += old->extent_offset + old->offset -
2656                          backref->extent_offset;
2657
2658         len = min(backref->extent_offset + backref->num_bytes,
2659                   old->extent_offset + old->offset + old->len);
2660         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2661
2662         ret = btrfs_drop_extents(trans, root, inode, start,
2663                                  start + len, 1);
2664         if (ret)
2665                 goto out_free_path;
2666 again:
2667         key.objectid = btrfs_ino(BTRFS_I(inode));
2668         key.type = BTRFS_EXTENT_DATA_KEY;
2669         key.offset = start;
2670
2671         path->leave_spinning = 1;
2672         if (merge) {
2673                 struct btrfs_file_extent_item *fi;
2674                 u64 extent_len;
2675                 struct btrfs_key found_key;
2676
2677                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2678                 if (ret < 0)
2679                         goto out_free_path;
2680
2681                 path->slots[0]--;
2682                 leaf = path->nodes[0];
2683                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2684
2685                 fi = btrfs_item_ptr(leaf, path->slots[0],
2686                                     struct btrfs_file_extent_item);
2687                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2688
2689                 if (extent_len + found_key.offset == start &&
2690                     relink_is_mergable(leaf, fi, new)) {
2691                         btrfs_set_file_extent_num_bytes(leaf, fi,
2692                                                         extent_len + len);
2693                         btrfs_mark_buffer_dirty(leaf);
2694                         inode_add_bytes(inode, len);
2695
2696                         ret = 1;
2697                         goto out_free_path;
2698                 } else {
2699                         merge = false;
2700                         btrfs_release_path(path);
2701                         goto again;
2702                 }
2703         }
2704
2705         ret = btrfs_insert_empty_item(trans, root, path, &key,
2706                                         sizeof(*extent));
2707         if (ret) {
2708                 btrfs_abort_transaction(trans, ret);
2709                 goto out_free_path;
2710         }
2711
2712         leaf = path->nodes[0];
2713         item = btrfs_item_ptr(leaf, path->slots[0],
2714                                 struct btrfs_file_extent_item);
2715         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2716         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2717         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2718         btrfs_set_file_extent_num_bytes(leaf, item, len);
2719         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2720         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2721         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2722         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2723         btrfs_set_file_extent_encryption(leaf, item, 0);
2724         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2725
2726         btrfs_mark_buffer_dirty(leaf);
2727         inode_add_bytes(inode, len);
2728         btrfs_release_path(path);
2729
2730         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2731                         new->disk_len, 0,
2732                         backref->root_id, backref->inum,
2733                         new->file_pos); /* start - extent_offset */
2734         if (ret) {
2735                 btrfs_abort_transaction(trans, ret);
2736                 goto out_free_path;
2737         }
2738
2739         ret = 1;
2740 out_free_path:
2741         btrfs_release_path(path);
2742         path->leave_spinning = 0;
2743         btrfs_end_transaction(trans);
2744 out_unlock:
2745         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2746                              &cached);
2747         iput(inode);
2748         return ret;
2749 }
2750
2751 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2752 {
2753         struct old_sa_defrag_extent *old, *tmp;
2754
2755         if (!new)
2756                 return;
2757
2758         list_for_each_entry_safe(old, tmp, &new->head, list) {
2759                 kfree(old);
2760         }
2761         kfree(new);
2762 }
2763
2764 static void relink_file_extents(struct new_sa_defrag_extent *new)
2765 {
2766         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2767         struct btrfs_path *path;
2768         struct sa_defrag_extent_backref *backref;
2769         struct sa_defrag_extent_backref *prev = NULL;
2770         struct inode *inode;
2771         struct rb_node *node;
2772         int ret;
2773
2774         inode = new->inode;
2775
2776         path = btrfs_alloc_path();
2777         if (!path)
2778                 return;
2779
2780         if (!record_extent_backrefs(path, new)) {
2781                 btrfs_free_path(path);
2782                 goto out;
2783         }
2784         btrfs_release_path(path);
2785
2786         while (1) {
2787                 node = rb_first(&new->root);
2788                 if (!node)
2789                         break;
2790                 rb_erase(node, &new->root);
2791
2792                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2793
2794                 ret = relink_extent_backref(path, prev, backref);
2795                 WARN_ON(ret < 0);
2796
2797                 kfree(prev);
2798
2799                 if (ret == 1)
2800                         prev = backref;
2801                 else
2802                         prev = NULL;
2803                 cond_resched();
2804         }
2805         kfree(prev);
2806
2807         btrfs_free_path(path);
2808 out:
2809         free_sa_defrag_extent(new);
2810
2811         atomic_dec(&fs_info->defrag_running);
2812         wake_up(&fs_info->transaction_wait);
2813 }
2814
2815 static struct new_sa_defrag_extent *
2816 record_old_file_extents(struct inode *inode,
2817                         struct btrfs_ordered_extent *ordered)
2818 {
2819         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2820         struct btrfs_root *root = BTRFS_I(inode)->root;
2821         struct btrfs_path *path;
2822         struct btrfs_key key;
2823         struct old_sa_defrag_extent *old;
2824         struct new_sa_defrag_extent *new;
2825         int ret;
2826
2827         new = kmalloc(sizeof(*new), GFP_NOFS);
2828         if (!new)
2829                 return NULL;
2830
2831         new->inode = inode;
2832         new->file_pos = ordered->file_offset;
2833         new->len = ordered->len;
2834         new->bytenr = ordered->start;
2835         new->disk_len = ordered->disk_len;
2836         new->compress_type = ordered->compress_type;
2837         new->root = RB_ROOT;
2838         INIT_LIST_HEAD(&new->head);
2839
2840         path = btrfs_alloc_path();
2841         if (!path)
2842                 goto out_kfree;
2843
2844         key.objectid = btrfs_ino(BTRFS_I(inode));
2845         key.type = BTRFS_EXTENT_DATA_KEY;
2846         key.offset = new->file_pos;
2847
2848         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2849         if (ret < 0)
2850                 goto out_free_path;
2851         if (ret > 0 && path->slots[0] > 0)
2852                 path->slots[0]--;
2853
2854         /* find out all the old extents for the file range */
2855         while (1) {
2856                 struct btrfs_file_extent_item *extent;
2857                 struct extent_buffer *l;
2858                 int slot;
2859                 u64 num_bytes;
2860                 u64 offset;
2861                 u64 end;
2862                 u64 disk_bytenr;
2863                 u64 extent_offset;
2864
2865                 l = path->nodes[0];
2866                 slot = path->slots[0];
2867
2868                 if (slot >= btrfs_header_nritems(l)) {
2869                         ret = btrfs_next_leaf(root, path);
2870                         if (ret < 0)
2871                                 goto out_free_path;
2872                         else if (ret > 0)
2873                                 break;
2874                         continue;
2875                 }
2876
2877                 btrfs_item_key_to_cpu(l, &key, slot);
2878
2879                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2880                         break;
2881                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2882                         break;
2883                 if (key.offset >= new->file_pos + new->len)
2884                         break;
2885
2886                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2887
2888                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2889                 if (key.offset + num_bytes < new->file_pos)
2890                         goto next;
2891
2892                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2893                 if (!disk_bytenr)
2894                         goto next;
2895
2896                 extent_offset = btrfs_file_extent_offset(l, extent);
2897
2898                 old = kmalloc(sizeof(*old), GFP_NOFS);
2899                 if (!old)
2900                         goto out_free_path;
2901
2902                 offset = max(new->file_pos, key.offset);
2903                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2904
2905                 old->bytenr = disk_bytenr;
2906                 old->extent_offset = extent_offset;
2907                 old->offset = offset - key.offset;
2908                 old->len = end - offset;
2909                 old->new = new;
2910                 old->count = 0;
2911                 list_add_tail(&old->list, &new->head);
2912 next:
2913                 path->slots[0]++;
2914                 cond_resched();
2915         }
2916
2917         btrfs_free_path(path);
2918         atomic_inc(&fs_info->defrag_running);
2919
2920         return new;
2921
2922 out_free_path:
2923         btrfs_free_path(path);
2924 out_kfree:
2925         free_sa_defrag_extent(new);
2926         return NULL;
2927 }
2928
2929 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2930                                          u64 start, u64 len)
2931 {
2932         struct btrfs_block_group_cache *cache;
2933
2934         cache = btrfs_lookup_block_group(fs_info, start);
2935         ASSERT(cache);
2936
2937         spin_lock(&cache->lock);
2938         cache->delalloc_bytes -= len;
2939         spin_unlock(&cache->lock);
2940
2941         btrfs_put_block_group(cache);
2942 }
2943
2944 /* as ordered data IO finishes, this gets called so we can finish
2945  * an ordered extent if the range of bytes in the file it covers are
2946  * fully written.
2947  */
2948 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2949 {
2950         struct inode *inode = ordered_extent->inode;
2951         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2952         struct btrfs_root *root = BTRFS_I(inode)->root;
2953         struct btrfs_trans_handle *trans = NULL;
2954         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2955         struct extent_state *cached_state = NULL;
2956         struct new_sa_defrag_extent *new = NULL;
2957         int compress_type = 0;
2958         int ret = 0;
2959         u64 logical_len = ordered_extent->len;
2960         bool nolock;
2961         bool truncated = false;
2962         bool range_locked = false;
2963         bool clear_new_delalloc_bytes = false;
2964
2965         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2966             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2967             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2968                 clear_new_delalloc_bytes = true;
2969
2970         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2971
2972         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2973                 ret = -EIO;
2974                 goto out;
2975         }
2976
2977         btrfs_free_io_failure_record(BTRFS_I(inode),
2978                         ordered_extent->file_offset,
2979                         ordered_extent->file_offset +
2980                         ordered_extent->len - 1);
2981
2982         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2983                 truncated = true;
2984                 logical_len = ordered_extent->truncated_len;
2985                 /* Truncated the entire extent, don't bother adding */
2986                 if (!logical_len)
2987                         goto out;
2988         }
2989
2990         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2991                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2992
2993                 /*
2994                  * For mwrite(mmap + memset to write) case, we still reserve
2995                  * space for NOCOW range.
2996                  * As NOCOW won't cause a new delayed ref, just free the space
2997                  */
2998                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2999                                        ordered_extent->len);
3000                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3001                 if (nolock)
3002                         trans = btrfs_join_transaction_nolock(root);
3003                 else
3004                         trans = btrfs_join_transaction(root);
3005                 if (IS_ERR(trans)) {
3006                         ret = PTR_ERR(trans);
3007                         trans = NULL;
3008                         goto out;
3009                 }
3010                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3011                 ret = btrfs_update_inode_fallback(trans, root, inode);
3012                 if (ret) /* -ENOMEM or corruption */
3013                         btrfs_abort_transaction(trans, ret);
3014                 goto out;
3015         }
3016
3017         range_locked = true;
3018         lock_extent_bits(io_tree, ordered_extent->file_offset,
3019                          ordered_extent->file_offset + ordered_extent->len - 1,
3020                          &cached_state);
3021
3022         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3023                         ordered_extent->file_offset + ordered_extent->len - 1,
3024                         EXTENT_DEFRAG, 0, cached_state);
3025         if (ret) {
3026                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3027                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3028                         /* the inode is shared */
3029                         new = record_old_file_extents(inode, ordered_extent);
3030
3031                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3032                         ordered_extent->file_offset + ordered_extent->len - 1,
3033                         EXTENT_DEFRAG, 0, 0, &cached_state);
3034         }
3035
3036         if (nolock)
3037                 trans = btrfs_join_transaction_nolock(root);
3038         else
3039                 trans = btrfs_join_transaction(root);
3040         if (IS_ERR(trans)) {
3041                 ret = PTR_ERR(trans);
3042                 trans = NULL;
3043                 goto out;
3044         }
3045
3046         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3047
3048         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3049                 compress_type = ordered_extent->compress_type;
3050         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3051                 BUG_ON(compress_type);
3052                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3053                                        ordered_extent->len);
3054                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3055                                                 ordered_extent->file_offset,
3056                                                 ordered_extent->file_offset +
3057                                                 logical_len);
3058         } else {
3059                 BUG_ON(root == fs_info->tree_root);
3060                 ret = insert_reserved_file_extent(trans, inode,
3061                                                 ordered_extent->file_offset,
3062                                                 ordered_extent->start,
3063                                                 ordered_extent->disk_len,
3064                                                 logical_len, logical_len,
3065                                                 compress_type, 0, 0,
3066                                                 BTRFS_FILE_EXTENT_REG);
3067                 if (!ret)
3068                         btrfs_release_delalloc_bytes(fs_info,
3069                                                      ordered_extent->start,
3070                                                      ordered_extent->disk_len);
3071         }
3072         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3073                            ordered_extent->file_offset, ordered_extent->len,
3074                            trans->transid);
3075         if (ret < 0) {
3076                 btrfs_abort_transaction(trans, ret);
3077                 goto out;
3078         }
3079
3080         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3081         if (ret) {
3082                 btrfs_abort_transaction(trans, ret);
3083                 goto out;
3084         }
3085
3086         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3087         ret = btrfs_update_inode_fallback(trans, root, inode);
3088         if (ret) { /* -ENOMEM or corruption */
3089                 btrfs_abort_transaction(trans, ret);
3090                 goto out;
3091         }
3092         ret = 0;
3093 out:
3094         if (range_locked || clear_new_delalloc_bytes) {
3095                 unsigned int clear_bits = 0;
3096
3097                 if (range_locked)
3098                         clear_bits |= EXTENT_LOCKED;
3099                 if (clear_new_delalloc_bytes)
3100                         clear_bits |= EXTENT_DELALLOC_NEW;
3101                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3102                                  ordered_extent->file_offset,
3103                                  ordered_extent->file_offset +
3104                                  ordered_extent->len - 1,
3105                                  clear_bits,
3106                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3107                                  0, &cached_state);
3108         }
3109
3110         if (trans)
3111                 btrfs_end_transaction(trans);
3112
3113         if (ret || truncated) {
3114                 u64 start, end;
3115
3116                 if (truncated)
3117                         start = ordered_extent->file_offset + logical_len;
3118                 else
3119                         start = ordered_extent->file_offset;
3120                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3121                 clear_extent_uptodate(io_tree, start, end, NULL);
3122
3123                 /* Drop the cache for the part of the extent we didn't write. */
3124                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3125
3126                 /*
3127                  * If the ordered extent had an IOERR or something else went
3128                  * wrong we need to return the space for this ordered extent
3129                  * back to the allocator.  We only free the extent in the
3130                  * truncated case if we didn't write out the extent at all.
3131                  */
3132                 if ((ret || !logical_len) &&
3133                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3134                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3135                         btrfs_free_reserved_extent(fs_info,
3136                                                    ordered_extent->start,
3137                                                    ordered_extent->disk_len, 1);
3138         }
3139
3140
3141         /*
3142          * This needs to be done to make sure anybody waiting knows we are done
3143          * updating everything for this ordered extent.
3144          */
3145         btrfs_remove_ordered_extent(inode, ordered_extent);
3146
3147         /* for snapshot-aware defrag */
3148         if (new) {
3149                 if (ret) {
3150                         free_sa_defrag_extent(new);
3151                         atomic_dec(&fs_info->defrag_running);
3152                 } else {
3153                         relink_file_extents(new);
3154                 }
3155         }
3156
3157         /* once for us */
3158         btrfs_put_ordered_extent(ordered_extent);
3159         /* once for the tree */
3160         btrfs_put_ordered_extent(ordered_extent);
3161
3162         /* Try to release some metadata so we don't get an OOM but don't wait */
3163         btrfs_btree_balance_dirty_nodelay(fs_info);
3164
3165         return ret;
3166 }
3167
3168 static void finish_ordered_fn(struct btrfs_work *work)
3169 {
3170         struct btrfs_ordered_extent *ordered_extent;
3171         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3172         btrfs_finish_ordered_io(ordered_extent);
3173 }
3174
3175 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3176                                 struct extent_state *state, int uptodate)
3177 {
3178         struct inode *inode = page->mapping->host;
3179         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3180         struct btrfs_ordered_extent *ordered_extent = NULL;
3181         struct btrfs_workqueue *wq;
3182         btrfs_work_func_t func;
3183
3184         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3185
3186         ClearPagePrivate2(page);
3187         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3188                                             end - start + 1, uptodate))
3189                 return;
3190
3191         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3192                 wq = fs_info->endio_freespace_worker;
3193                 func = btrfs_freespace_write_helper;
3194         } else {
3195                 wq = fs_info->endio_write_workers;
3196                 func = btrfs_endio_write_helper;
3197         }
3198
3199         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3200                         NULL);
3201         btrfs_queue_work(wq, &ordered_extent->work);
3202 }
3203
3204 static int __readpage_endio_check(struct inode *inode,
3205                                   struct btrfs_io_bio *io_bio,
3206                                   int icsum, struct page *page,
3207                                   int pgoff, u64 start, size_t len)
3208 {
3209         char *kaddr;
3210         u32 csum_expected;
3211         u32 csum = ~(u32)0;
3212
3213         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3214
3215         kaddr = kmap_atomic(page);
3216         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3217         btrfs_csum_final(csum, (u8 *)&csum);
3218         if (csum != csum_expected)
3219                 goto zeroit;
3220
3221         kunmap_atomic(kaddr);
3222         return 0;
3223 zeroit:
3224         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3225                                     io_bio->mirror_num);
3226         memset(kaddr + pgoff, 1, len);
3227         flush_dcache_page(page);
3228         kunmap_atomic(kaddr);
3229         return -EIO;
3230 }
3231
3232 /*
3233  * when reads are done, we need to check csums to verify the data is correct
3234  * if there's a match, we allow the bio to finish.  If not, the code in
3235  * extent_io.c will try to find good copies for us.
3236  */
3237 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3238                                       u64 phy_offset, struct page *page,
3239                                       u64 start, u64 end, int mirror)
3240 {
3241         size_t offset = start - page_offset(page);
3242         struct inode *inode = page->mapping->host;
3243         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3244         struct btrfs_root *root = BTRFS_I(inode)->root;
3245
3246         if (PageChecked(page)) {
3247                 ClearPageChecked(page);
3248                 return 0;
3249         }
3250
3251         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3252                 return 0;
3253
3254         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3255             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3256                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3257                 return 0;
3258         }
3259
3260         phy_offset >>= inode->i_sb->s_blocksize_bits;
3261         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3262                                       start, (size_t)(end - start + 1));
3263 }
3264
3265 /*
3266  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3267  *
3268  * @inode: The inode we want to perform iput on
3269  *
3270  * This function uses the generic vfs_inode::i_count to track whether we should
3271  * just decrement it (in case it's > 1) or if this is the last iput then link
3272  * the inode to the delayed iput machinery. Delayed iputs are processed at
3273  * transaction commit time/superblock commit/cleaner kthread.
3274  */
3275 void btrfs_add_delayed_iput(struct inode *inode)
3276 {
3277         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3278         struct btrfs_inode *binode = BTRFS_I(inode);
3279
3280         if (atomic_add_unless(&inode->i_count, -1, 1))
3281                 return;
3282
3283         spin_lock(&fs_info->delayed_iput_lock);
3284         ASSERT(list_empty(&binode->delayed_iput));
3285         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3286         spin_unlock(&fs_info->delayed_iput_lock);
3287 }
3288
3289 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3290 {
3291
3292         spin_lock(&fs_info->delayed_iput_lock);
3293         while (!list_empty(&fs_info->delayed_iputs)) {
3294                 struct btrfs_inode *inode;
3295
3296                 inode = list_first_entry(&fs_info->delayed_iputs,
3297                                 struct btrfs_inode, delayed_iput);
3298                 list_del_init(&inode->delayed_iput);
3299                 spin_unlock(&fs_info->delayed_iput_lock);
3300                 iput(&inode->vfs_inode);
3301                 spin_lock(&fs_info->delayed_iput_lock);
3302         }
3303         spin_unlock(&fs_info->delayed_iput_lock);
3304 }
3305
3306 /*
3307  * This creates an orphan entry for the given inode in case something goes wrong
3308  * in the middle of an unlink.
3309  */
3310 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3311                      struct btrfs_inode *inode)
3312 {
3313         int ret;
3314
3315         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
3316         if (ret && ret != -EEXIST) {
3317                 btrfs_abort_transaction(trans, ret);
3318                 return ret;
3319         }
3320
3321         return 0;
3322 }
3323
3324 /*
3325  * We have done the delete so we can go ahead and remove the orphan item for
3326  * this particular inode.
3327  */
3328 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3329                             struct btrfs_inode *inode)
3330 {
3331         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
3332 }
3333
3334 /*
3335  * this cleans up any orphans that may be left on the list from the last use
3336  * of this root.
3337  */
3338 int btrfs_orphan_cleanup(struct btrfs_root *root)
3339 {
3340         struct btrfs_fs_info *fs_info = root->fs_info;
3341         struct btrfs_path *path;
3342         struct extent_buffer *leaf;
3343         struct btrfs_key key, found_key;
3344         struct btrfs_trans_handle *trans;
3345         struct inode *inode;
3346         u64 last_objectid = 0;
3347         int ret = 0, nr_unlink = 0;
3348
3349         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3350                 return 0;
3351
3352         path = btrfs_alloc_path();
3353         if (!path) {
3354                 ret = -ENOMEM;
3355                 goto out;
3356         }
3357         path->reada = READA_BACK;
3358
3359         key.objectid = BTRFS_ORPHAN_OBJECTID;
3360         key.type = BTRFS_ORPHAN_ITEM_KEY;
3361         key.offset = (u64)-1;
3362
3363         while (1) {
3364                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3365                 if (ret < 0)
3366                         goto out;
3367
3368                 /*
3369                  * if ret == 0 means we found what we were searching for, which
3370                  * is weird, but possible, so only screw with path if we didn't
3371                  * find the key and see if we have stuff that matches
3372                  */
3373                 if (ret > 0) {
3374                         ret = 0;
3375                         if (path->slots[0] == 0)
3376                                 break;
3377                         path->slots[0]--;
3378                 }
3379
3380                 /* pull out the item */
3381                 leaf = path->nodes[0];
3382                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3383
3384                 /* make sure the item matches what we want */
3385                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3386                         break;
3387                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3388                         break;
3389
3390                 /* release the path since we're done with it */
3391                 btrfs_release_path(path);
3392
3393                 /*
3394                  * this is where we are basically btrfs_lookup, without the
3395                  * crossing root thing.  we store the inode number in the
3396                  * offset of the orphan item.
3397                  */
3398
3399                 if (found_key.offset == last_objectid) {
3400                         btrfs_err(fs_info,
3401                                   "Error removing orphan entry, stopping orphan cleanup");
3402                         ret = -EINVAL;
3403                         goto out;
3404                 }
3405
3406                 last_objectid = found_key.offset;
3407
3408                 found_key.objectid = found_key.offset;
3409                 found_key.type = BTRFS_INODE_ITEM_KEY;
3410                 found_key.offset = 0;
3411                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3412                 ret = PTR_ERR_OR_ZERO(inode);
3413                 if (ret && ret != -ENOENT)
3414                         goto out;
3415
3416                 if (ret == -ENOENT && root == fs_info->tree_root) {
3417                         struct btrfs_root *dead_root;
3418                         struct btrfs_fs_info *fs_info = root->fs_info;
3419                         int is_dead_root = 0;
3420
3421                         /*
3422                          * this is an orphan in the tree root. Currently these
3423                          * could come from 2 sources:
3424                          *  a) a snapshot deletion in progress
3425                          *  b) a free space cache inode
3426                          * We need to distinguish those two, as the snapshot
3427                          * orphan must not get deleted.
3428                          * find_dead_roots already ran before us, so if this
3429                          * is a snapshot deletion, we should find the root
3430                          * in the dead_roots list
3431                          */
3432                         spin_lock(&fs_info->trans_lock);
3433                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3434                                             root_list) {
3435                                 if (dead_root->root_key.objectid ==
3436                                     found_key.objectid) {
3437                                         is_dead_root = 1;
3438                                         break;
3439                                 }
3440                         }
3441                         spin_unlock(&fs_info->trans_lock);
3442                         if (is_dead_root) {
3443                                 /* prevent this orphan from being found again */
3444                                 key.offset = found_key.objectid - 1;
3445                                 continue;
3446                         }
3447
3448                 }
3449
3450                 /*
3451                  * If we have an inode with links, there are a couple of
3452                  * possibilities. Old kernels (before v3.12) used to create an
3453                  * orphan item for truncate indicating that there were possibly
3454                  * extent items past i_size that needed to be deleted. In v3.12,
3455                  * truncate was changed to update i_size in sync with the extent
3456                  * items, but the (useless) orphan item was still created. Since
3457                  * v4.18, we don't create the orphan item for truncate at all.
3458                  *
3459                  * So, this item could mean that we need to do a truncate, but
3460                  * only if this filesystem was last used on a pre-v3.12 kernel
3461                  * and was not cleanly unmounted. The odds of that are quite
3462                  * slim, and it's a pain to do the truncate now, so just delete
3463                  * the orphan item.
3464                  *
3465                  * It's also possible that this orphan item was supposed to be
3466                  * deleted but wasn't. The inode number may have been reused,
3467                  * but either way, we can delete the orphan item.
3468                  */
3469                 if (ret == -ENOENT || inode->i_nlink) {
3470                         if (!ret)
3471                                 iput(inode);
3472                         trans = btrfs_start_transaction(root, 1);
3473                         if (IS_ERR(trans)) {
3474                                 ret = PTR_ERR(trans);
3475                                 goto out;
3476                         }
3477                         btrfs_debug(fs_info, "auto deleting %Lu",
3478                                     found_key.objectid);
3479                         ret = btrfs_del_orphan_item(trans, root,
3480                                                     found_key.objectid);
3481                         btrfs_end_transaction(trans);
3482                         if (ret)
3483                                 goto out;
3484                         continue;
3485                 }
3486
3487                 nr_unlink++;
3488
3489                 /* this will do delete_inode and everything for us */
3490                 iput(inode);
3491                 if (ret)
3492                         goto out;
3493         }
3494         /* release the path since we're done with it */
3495         btrfs_release_path(path);
3496
3497         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3498
3499         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3500                 trans = btrfs_join_transaction(root);
3501                 if (!IS_ERR(trans))
3502                         btrfs_end_transaction(trans);
3503         }
3504
3505         if (nr_unlink)
3506                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3507
3508 out:
3509         if (ret)
3510                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3511         btrfs_free_path(path);
3512         return ret;
3513 }
3514
3515 /*
3516  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3517  * don't find any xattrs, we know there can't be any acls.
3518  *
3519  * slot is the slot the inode is in, objectid is the objectid of the inode
3520  */
3521 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3522                                           int slot, u64 objectid,
3523                                           int *first_xattr_slot)
3524 {
3525         u32 nritems = btrfs_header_nritems(leaf);
3526         struct btrfs_key found_key;
3527         static u64 xattr_access = 0;
3528         static u64 xattr_default = 0;
3529         int scanned = 0;
3530
3531         if (!xattr_access) {
3532                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3533                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3534                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3535                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3536         }
3537
3538         slot++;
3539         *first_xattr_slot = -1;
3540         while (slot < nritems) {
3541                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3542
3543                 /* we found a different objectid, there must not be acls */
3544                 if (found_key.objectid != objectid)
3545                         return 0;
3546
3547                 /* we found an xattr, assume we've got an acl */
3548                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3549                         if (*first_xattr_slot == -1)
3550                                 *first_xattr_slot = slot;
3551                         if (found_key.offset == xattr_access ||
3552                             found_key.offset == xattr_default)
3553                                 return 1;
3554                 }
3555
3556                 /*
3557                  * we found a key greater than an xattr key, there can't
3558                  * be any acls later on
3559                  */
3560                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3561                         return 0;
3562
3563                 slot++;
3564                 scanned++;
3565
3566                 /*
3567                  * it goes inode, inode backrefs, xattrs, extents,
3568                  * so if there are a ton of hard links to an inode there can
3569                  * be a lot of backrefs.  Don't waste time searching too hard,
3570                  * this is just an optimization
3571                  */
3572                 if (scanned >= 8)
3573                         break;
3574         }
3575         /* we hit the end of the leaf before we found an xattr or
3576          * something larger than an xattr.  We have to assume the inode
3577          * has acls
3578          */
3579         if (*first_xattr_slot == -1)
3580                 *first_xattr_slot = slot;
3581         return 1;
3582 }
3583
3584 /*
3585  * read an inode from the btree into the in-memory inode
3586  */
3587 static int btrfs_read_locked_inode(struct inode *inode)
3588 {
3589         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3590         struct btrfs_path *path;
3591         struct extent_buffer *leaf;
3592         struct btrfs_inode_item *inode_item;
3593         struct btrfs_root *root = BTRFS_I(inode)->root;
3594         struct btrfs_key location;
3595         unsigned long ptr;
3596         int maybe_acls;
3597         u32 rdev;
3598         int ret;
3599         bool filled = false;
3600         int first_xattr_slot;
3601
3602         ret = btrfs_fill_inode(inode, &rdev);
3603         if (!ret)
3604                 filled = true;
3605
3606         path = btrfs_alloc_path();
3607         if (!path)
3608                 return -ENOMEM;
3609
3610         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3611
3612         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3613         if (ret) {
3614                 btrfs_free_path(path);
3615                 return ret;
3616         }
3617
3618         leaf = path->nodes[0];
3619
3620         if (filled)
3621                 goto cache_index;
3622
3623         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3624                                     struct btrfs_inode_item);
3625         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3626         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3627         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3628         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3629         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3630
3631         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3632         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3633
3634         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3635         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3636
3637         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3638         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3639
3640         BTRFS_I(inode)->i_otime.tv_sec =
3641                 btrfs_timespec_sec(leaf, &inode_item->otime);
3642         BTRFS_I(inode)->i_otime.tv_nsec =
3643                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3644
3645         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3646         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3647         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3648
3649         inode_set_iversion_queried(inode,
3650                                    btrfs_inode_sequence(leaf, inode_item));
3651         inode->i_generation = BTRFS_I(inode)->generation;
3652         inode->i_rdev = 0;
3653         rdev = btrfs_inode_rdev(leaf, inode_item);
3654
3655         BTRFS_I(inode)->index_cnt = (u64)-1;
3656         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3657
3658 cache_index:
3659         /*
3660          * If we were modified in the current generation and evicted from memory
3661          * and then re-read we need to do a full sync since we don't have any
3662          * idea about which extents were modified before we were evicted from
3663          * cache.
3664          *
3665          * This is required for both inode re-read from disk and delayed inode
3666          * in delayed_nodes_tree.
3667          */
3668         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3669                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3670                         &BTRFS_I(inode)->runtime_flags);
3671
3672         /*
3673          * We don't persist the id of the transaction where an unlink operation
3674          * against the inode was last made. So here we assume the inode might
3675          * have been evicted, and therefore the exact value of last_unlink_trans
3676          * lost, and set it to last_trans to avoid metadata inconsistencies
3677          * between the inode and its parent if the inode is fsync'ed and the log
3678          * replayed. For example, in the scenario:
3679          *
3680          * touch mydir/foo
3681          * ln mydir/foo mydir/bar
3682          * sync
3683          * unlink mydir/bar
3684          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3685          * xfs_io -c fsync mydir/foo
3686          * <power failure>
3687          * mount fs, triggers fsync log replay
3688          *
3689          * We must make sure that when we fsync our inode foo we also log its
3690          * parent inode, otherwise after log replay the parent still has the
3691          * dentry with the "bar" name but our inode foo has a link count of 1
3692          * and doesn't have an inode ref with the name "bar" anymore.
3693          *
3694          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3695          * but it guarantees correctness at the expense of occasional full
3696          * transaction commits on fsync if our inode is a directory, or if our
3697          * inode is not a directory, logging its parent unnecessarily.
3698          */
3699         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3700
3701         path->slots[0]++;
3702         if (inode->i_nlink != 1 ||
3703             path->slots[0] >= btrfs_header_nritems(leaf))
3704                 goto cache_acl;
3705
3706         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3707         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3708                 goto cache_acl;
3709
3710         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3711         if (location.type == BTRFS_INODE_REF_KEY) {
3712                 struct btrfs_inode_ref *ref;
3713
3714                 ref = (struct btrfs_inode_ref *)ptr;
3715                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3716         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3717                 struct btrfs_inode_extref *extref;
3718
3719                 extref = (struct btrfs_inode_extref *)ptr;
3720                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3721                                                                      extref);
3722         }
3723 cache_acl:
3724         /*
3725          * try to precache a NULL acl entry for files that don't have
3726          * any xattrs or acls
3727          */
3728         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3729                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3730         if (first_xattr_slot != -1) {
3731                 path->slots[0] = first_xattr_slot;
3732                 ret = btrfs_load_inode_props(inode, path);
3733                 if (ret)
3734                         btrfs_err(fs_info,
3735                                   "error loading props for ino %llu (root %llu): %d",
3736                                   btrfs_ino(BTRFS_I(inode)),
3737                                   root->root_key.objectid, ret);
3738         }
3739         btrfs_free_path(path);
3740
3741         if (!maybe_acls)
3742                 cache_no_acl(inode);
3743
3744         switch (inode->i_mode & S_IFMT) {
3745         case S_IFREG:
3746                 inode->i_mapping->a_ops = &btrfs_aops;
3747                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3748                 inode->i_fop = &btrfs_file_operations;
3749                 inode->i_op = &btrfs_file_inode_operations;
3750                 break;
3751         case S_IFDIR:
3752                 inode->i_fop = &btrfs_dir_file_operations;
3753                 inode->i_op = &btrfs_dir_inode_operations;
3754                 break;
3755         case S_IFLNK:
3756                 inode->i_op = &btrfs_symlink_inode_operations;
3757                 inode_nohighmem(inode);
3758                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3759                 break;
3760         default:
3761                 inode->i_op = &btrfs_special_inode_operations;
3762                 init_special_inode(inode, inode->i_mode, rdev);
3763                 break;
3764         }
3765
3766         btrfs_sync_inode_flags_to_i_flags(inode);
3767         return 0;
3768 }
3769
3770 /*
3771  * given a leaf and an inode, copy the inode fields into the leaf
3772  */
3773 static void fill_inode_item(struct btrfs_trans_handle *trans,
3774                             struct extent_buffer *leaf,
3775                             struct btrfs_inode_item *item,
3776                             struct inode *inode)
3777 {
3778         struct btrfs_map_token token;
3779
3780         btrfs_init_map_token(&token);
3781
3782         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3783         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3784         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3785                                    &token);
3786         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3787         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3788
3789         btrfs_set_token_timespec_sec(leaf, &item->atime,
3790                                      inode->i_atime.tv_sec, &token);
3791         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3792                                       inode->i_atime.tv_nsec, &token);
3793
3794         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3795                                      inode->i_mtime.tv_sec, &token);
3796         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3797                                       inode->i_mtime.tv_nsec, &token);
3798
3799         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3800                                      inode->i_ctime.tv_sec, &token);
3801         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3802                                       inode->i_ctime.tv_nsec, &token);
3803
3804         btrfs_set_token_timespec_sec(leaf, &item->otime,
3805                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3806         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3807                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3808
3809         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3810                                      &token);
3811         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3812                                          &token);
3813         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3814                                        &token);
3815         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3816         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3817         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3818         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3819 }
3820
3821 /*
3822  * copy everything in the in-memory inode into the btree.
3823  */
3824 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3825                                 struct btrfs_root *root, struct inode *inode)
3826 {
3827         struct btrfs_inode_item *inode_item;
3828         struct btrfs_path *path;
3829         struct extent_buffer *leaf;
3830         int ret;
3831
3832         path = btrfs_alloc_path();
3833         if (!path)
3834                 return -ENOMEM;
3835
3836         path->leave_spinning = 1;
3837         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3838                                  1);
3839         if (ret) {
3840                 if (ret > 0)
3841                         ret = -ENOENT;
3842                 goto failed;
3843         }
3844
3845         leaf = path->nodes[0];
3846         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3847                                     struct btrfs_inode_item);
3848
3849         fill_inode_item(trans, leaf, inode_item, inode);
3850         btrfs_mark_buffer_dirty(leaf);
3851         btrfs_set_inode_last_trans(trans, inode);
3852         ret = 0;
3853 failed:
3854         btrfs_free_path(path);
3855         return ret;
3856 }
3857
3858 /*
3859  * copy everything in the in-memory inode into the btree.
3860  */
3861 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3862                                 struct btrfs_root *root, struct inode *inode)
3863 {
3864         struct btrfs_fs_info *fs_info = root->fs_info;
3865         int ret;
3866
3867         /*
3868          * If the inode is a free space inode, we can deadlock during commit
3869          * if we put it into the delayed code.
3870          *
3871          * The data relocation inode should also be directly updated
3872          * without delay
3873          */
3874         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3875             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3876             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3877                 btrfs_update_root_times(trans, root);
3878
3879                 ret = btrfs_delayed_update_inode(trans, root, inode);
3880                 if (!ret)
3881                         btrfs_set_inode_last_trans(trans, inode);
3882                 return ret;
3883         }
3884
3885         return btrfs_update_inode_item(trans, root, inode);
3886 }
3887
3888 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3889                                          struct btrfs_root *root,
3890                                          struct inode *inode)
3891 {
3892         int ret;
3893
3894         ret = btrfs_update_inode(trans, root, inode);
3895         if (ret == -ENOSPC)
3896                 return btrfs_update_inode_item(trans, root, inode);
3897         return ret;
3898 }
3899
3900 /*
3901  * unlink helper that gets used here in inode.c and in the tree logging
3902  * recovery code.  It remove a link in a directory with a given name, and
3903  * also drops the back refs in the inode to the directory
3904  */
3905 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3906                                 struct btrfs_root *root,
3907                                 struct btrfs_inode *dir,
3908                                 struct btrfs_inode *inode,
3909                                 const char *name, int name_len)
3910 {
3911         struct btrfs_fs_info *fs_info = root->fs_info;
3912         struct btrfs_path *path;
3913         int ret = 0;
3914         struct extent_buffer *leaf;
3915         struct btrfs_dir_item *di;
3916         struct btrfs_key key;
3917         u64 index;
3918         u64 ino = btrfs_ino(inode);
3919         u64 dir_ino = btrfs_ino(dir);
3920
3921         path = btrfs_alloc_path();
3922         if (!path) {
3923                 ret = -ENOMEM;
3924                 goto out;
3925         }
3926
3927         path->leave_spinning = 1;
3928         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3929                                     name, name_len, -1);
3930         if (IS_ERR(di)) {
3931                 ret = PTR_ERR(di);
3932                 goto err;
3933         }
3934         if (!di) {
3935                 ret = -ENOENT;
3936                 goto err;
3937         }
3938         leaf = path->nodes[0];
3939         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3940         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3941         if (ret)
3942                 goto err;
3943         btrfs_release_path(path);
3944
3945         /*
3946          * If we don't have dir index, we have to get it by looking up
3947          * the inode ref, since we get the inode ref, remove it directly,
3948          * it is unnecessary to do delayed deletion.
3949          *
3950          * But if we have dir index, needn't search inode ref to get it.
3951          * Since the inode ref is close to the inode item, it is better
3952          * that we delay to delete it, and just do this deletion when
3953          * we update the inode item.
3954          */
3955         if (inode->dir_index) {
3956                 ret = btrfs_delayed_delete_inode_ref(inode);
3957                 if (!ret) {
3958                         index = inode->dir_index;
3959                         goto skip_backref;
3960                 }
3961         }
3962
3963         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3964                                   dir_ino, &index);
3965         if (ret) {
3966                 btrfs_info(fs_info,
3967                         "failed to delete reference to %.*s, inode %llu parent %llu",
3968                         name_len, name, ino, dir_ino);
3969                 btrfs_abort_transaction(trans, ret);
3970                 goto err;
3971         }
3972 skip_backref:
3973         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
3974         if (ret) {
3975                 btrfs_abort_transaction(trans, ret);
3976                 goto err;
3977         }
3978
3979         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
3980                         dir_ino);
3981         if (ret != 0 && ret != -ENOENT) {
3982                 btrfs_abort_transaction(trans, ret);
3983                 goto err;
3984         }
3985
3986         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
3987                         index);
3988         if (ret == -ENOENT)
3989                 ret = 0;
3990         else if (ret)
3991                 btrfs_abort_transaction(trans, ret);
3992 err:
3993         btrfs_free_path(path);
3994         if (ret)
3995                 goto out;
3996
3997         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
3998         inode_inc_iversion(&inode->vfs_inode);
3999         inode_inc_iversion(&dir->vfs_inode);
4000         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4001                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4002         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4003 out:
4004         return ret;
4005 }
4006
4007 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4008                        struct btrfs_root *root,
4009                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4010                        const char *name, int name_len)
4011 {
4012         int ret;
4013         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4014         if (!ret) {
4015                 drop_nlink(&inode->vfs_inode);
4016                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4017         }
4018         return ret;
4019 }
4020
4021 /*
4022  * helper to start transaction for unlink and rmdir.
4023  *
4024  * unlink and rmdir are special in btrfs, they do not always free space, so
4025  * if we cannot make our reservations the normal way try and see if there is
4026  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4027  * allow the unlink to occur.
4028  */
4029 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4030 {
4031         struct btrfs_root *root = BTRFS_I(dir)->root;
4032
4033         /*
4034          * 1 for the possible orphan item
4035          * 1 for the dir item
4036          * 1 for the dir index
4037          * 1 for the inode ref
4038          * 1 for the inode
4039          */
4040         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4041 }
4042
4043 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4044 {
4045         struct btrfs_root *root = BTRFS_I(dir)->root;
4046         struct btrfs_trans_handle *trans;
4047         struct inode *inode = d_inode(dentry);
4048         int ret;
4049
4050         trans = __unlink_start_trans(dir);
4051         if (IS_ERR(trans))
4052                 return PTR_ERR(trans);
4053
4054         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4055                         0);
4056
4057         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4058                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4059                         dentry->d_name.len);
4060         if (ret)
4061                 goto out;
4062
4063         if (inode->i_nlink == 0) {
4064                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4065                 if (ret)
4066                         goto out;
4067         }
4068
4069 out:
4070         btrfs_end_transaction(trans);
4071         btrfs_btree_balance_dirty(root->fs_info);
4072         return ret;
4073 }
4074
4075 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4076                                struct inode *dir, u64 objectid,
4077                                const char *name, int name_len)
4078 {
4079         struct btrfs_root *root = BTRFS_I(dir)->root;
4080         struct btrfs_path *path;
4081         struct extent_buffer *leaf;
4082         struct btrfs_dir_item *di;
4083         struct btrfs_key key;
4084         u64 index;
4085         int ret;
4086         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4087
4088         path = btrfs_alloc_path();
4089         if (!path)
4090                 return -ENOMEM;
4091
4092         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4093                                    name, name_len, -1);
4094         if (IS_ERR_OR_NULL(di)) {
4095                 if (!di)
4096                         ret = -ENOENT;
4097                 else
4098                         ret = PTR_ERR(di);
4099                 goto out;
4100         }
4101
4102         leaf = path->nodes[0];
4103         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4104         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4105         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4106         if (ret) {
4107                 btrfs_abort_transaction(trans, ret);
4108                 goto out;
4109         }
4110         btrfs_release_path(path);
4111
4112         ret = btrfs_del_root_ref(trans, objectid, root->root_key.objectid,
4113                                  dir_ino, &index, name, name_len);
4114         if (ret < 0) {
4115                 if (ret != -ENOENT) {
4116                         btrfs_abort_transaction(trans, ret);
4117                         goto out;
4118                 }
4119                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4120                                                  name, name_len);
4121                 if (IS_ERR_OR_NULL(di)) {
4122                         if (!di)
4123                                 ret = -ENOENT;
4124                         else
4125                                 ret = PTR_ERR(di);
4126                         btrfs_abort_transaction(trans, ret);
4127                         goto out;
4128                 }
4129
4130                 leaf = path->nodes[0];
4131                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4132                 index = key.offset;
4133         }
4134         btrfs_release_path(path);
4135
4136         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
4137         if (ret) {
4138                 btrfs_abort_transaction(trans, ret);
4139                 goto out;
4140         }
4141
4142         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4143         inode_inc_iversion(dir);
4144         dir->i_mtime = dir->i_ctime = current_time(dir);
4145         ret = btrfs_update_inode_fallback(trans, root, dir);
4146         if (ret)
4147                 btrfs_abort_transaction(trans, ret);
4148 out:
4149         btrfs_free_path(path);
4150         return ret;
4151 }
4152
4153 /*
4154  * Helper to check if the subvolume references other subvolumes or if it's
4155  * default.
4156  */
4157 static noinline int may_destroy_subvol(struct btrfs_root *root)
4158 {
4159         struct btrfs_fs_info *fs_info = root->fs_info;
4160         struct btrfs_path *path;
4161         struct btrfs_dir_item *di;
4162         struct btrfs_key key;
4163         u64 dir_id;
4164         int ret;
4165
4166         path = btrfs_alloc_path();
4167         if (!path)
4168                 return -ENOMEM;
4169
4170         /* Make sure this root isn't set as the default subvol */
4171         dir_id = btrfs_super_root_dir(fs_info->super_copy);
4172         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
4173                                    dir_id, "default", 7, 0);
4174         if (di && !IS_ERR(di)) {
4175                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
4176                 if (key.objectid == root->root_key.objectid) {
4177                         ret = -EPERM;
4178                         btrfs_err(fs_info,
4179                                   "deleting default subvolume %llu is not allowed",
4180                                   key.objectid);
4181                         goto out;
4182                 }
4183                 btrfs_release_path(path);
4184         }
4185
4186         key.objectid = root->root_key.objectid;
4187         key.type = BTRFS_ROOT_REF_KEY;
4188         key.offset = (u64)-1;
4189
4190         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
4191         if (ret < 0)
4192                 goto out;
4193         BUG_ON(ret == 0);
4194
4195         ret = 0;
4196         if (path->slots[0] > 0) {
4197                 path->slots[0]--;
4198                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4199                 if (key.objectid == root->root_key.objectid &&
4200                     key.type == BTRFS_ROOT_REF_KEY)
4201                         ret = -ENOTEMPTY;
4202         }
4203 out:
4204         btrfs_free_path(path);
4205         return ret;
4206 }
4207
4208 /* Delete all dentries for inodes belonging to the root */
4209 static void btrfs_prune_dentries(struct btrfs_root *root)
4210 {
4211         struct btrfs_fs_info *fs_info = root->fs_info;
4212         struct rb_node *node;
4213         struct rb_node *prev;
4214         struct btrfs_inode *entry;
4215         struct inode *inode;
4216         u64 objectid = 0;
4217
4218         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4219                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4220
4221         spin_lock(&root->inode_lock);
4222 again:
4223         node = root->inode_tree.rb_node;
4224         prev = NULL;
4225         while (node) {
4226                 prev = node;
4227                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4228
4229                 if (objectid < btrfs_ino(entry))
4230                         node = node->rb_left;
4231                 else if (objectid > btrfs_ino(entry))
4232                         node = node->rb_right;
4233                 else
4234                         break;
4235         }
4236         if (!node) {
4237                 while (prev) {
4238                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4239                         if (objectid <= btrfs_ino(entry)) {
4240                                 node = prev;
4241                                 break;
4242                         }
4243                         prev = rb_next(prev);
4244                 }
4245         }
4246         while (node) {
4247                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4248                 objectid = btrfs_ino(entry) + 1;
4249                 inode = igrab(&entry->vfs_inode);
4250                 if (inode) {
4251                         spin_unlock(&root->inode_lock);
4252                         if (atomic_read(&inode->i_count) > 1)
4253                                 d_prune_aliases(inode);
4254                         /*
4255                          * btrfs_drop_inode will have it removed from the inode
4256                          * cache when its usage count hits zero.
4257                          */
4258                         iput(inode);
4259                         cond_resched();
4260                         spin_lock(&root->inode_lock);
4261                         goto again;
4262                 }
4263
4264                 if (cond_resched_lock(&root->inode_lock))
4265                         goto again;
4266
4267                 node = rb_next(node);
4268         }
4269         spin_unlock(&root->inode_lock);
4270 }
4271
4272 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
4273 {
4274         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
4275         struct btrfs_root *root = BTRFS_I(dir)->root;
4276         struct inode *inode = d_inode(dentry);
4277         struct btrfs_root *dest = BTRFS_I(inode)->root;
4278         struct btrfs_trans_handle *trans;
4279         struct btrfs_block_rsv block_rsv;
4280         u64 root_flags;
4281         int ret;
4282         int err;
4283
4284         /*
4285          * Don't allow to delete a subvolume with send in progress. This is
4286          * inside the inode lock so the error handling that has to drop the bit
4287          * again is not run concurrently.
4288          */
4289         spin_lock(&dest->root_item_lock);
4290         root_flags = btrfs_root_flags(&dest->root_item);
4291         if (dest->send_in_progress == 0) {
4292                 btrfs_set_root_flags(&dest->root_item,
4293                                 root_flags | BTRFS_ROOT_SUBVOL_DEAD);
4294                 spin_unlock(&dest->root_item_lock);
4295         } else {
4296                 spin_unlock(&dest->root_item_lock);
4297                 btrfs_warn(fs_info,
4298                            "attempt to delete subvolume %llu during send",
4299                            dest->root_key.objectid);
4300                 return -EPERM;
4301         }
4302
4303         down_write(&fs_info->subvol_sem);
4304
4305         err = may_destroy_subvol(dest);
4306         if (err)
4307                 goto out_up_write;
4308
4309         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
4310         /*
4311          * One for dir inode,
4312          * two for dir entries,
4313          * two for root ref/backref.
4314          */
4315         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
4316         if (err)
4317                 goto out_up_write;
4318
4319         trans = btrfs_start_transaction(root, 0);
4320         if (IS_ERR(trans)) {
4321                 err = PTR_ERR(trans);
4322                 goto out_release;
4323         }
4324         trans->block_rsv = &block_rsv;
4325         trans->bytes_reserved = block_rsv.size;
4326
4327         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
4328
4329         ret = btrfs_unlink_subvol(trans, dir, dest->root_key.objectid,
4330                                   dentry->d_name.name, dentry->d_name.len);
4331         if (ret) {
4332                 err = ret;
4333                 btrfs_abort_transaction(trans, ret);
4334                 goto out_end_trans;
4335         }
4336
4337         btrfs_record_root_in_trans(trans, dest);
4338
4339         memset(&dest->root_item.drop_progress, 0,
4340                 sizeof(dest->root_item.drop_progress));
4341         dest->root_item.drop_level = 0;
4342         btrfs_set_root_refs(&dest->root_item, 0);
4343
4344         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4345                 ret = btrfs_insert_orphan_item(trans,
4346                                         fs_info->tree_root,
4347                                         dest->root_key.objectid);
4348                 if (ret) {
4349                         btrfs_abort_transaction(trans, ret);
4350                         err = ret;
4351                         goto out_end_trans;
4352                 }
4353         }
4354
4355         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4356                                   BTRFS_UUID_KEY_SUBVOL,
4357                                   dest->root_key.objectid);
4358         if (ret && ret != -ENOENT) {
4359                 btrfs_abort_transaction(trans, ret);
4360                 err = ret;
4361                 goto out_end_trans;
4362         }
4363         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4364                 ret = btrfs_uuid_tree_remove(trans,
4365                                           dest->root_item.received_uuid,
4366                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4367                                           dest->root_key.objectid);
4368                 if (ret && ret != -ENOENT) {
4369                         btrfs_abort_transaction(trans, ret);
4370                         err = ret;
4371                         goto out_end_trans;
4372                 }
4373         }
4374
4375 out_end_trans:
4376         trans->block_rsv = NULL;
4377         trans->bytes_reserved = 0;
4378         ret = btrfs_end_transaction(trans);
4379         if (ret && !err)
4380                 err = ret;
4381         inode->i_flags |= S_DEAD;
4382 out_release:
4383         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4384 out_up_write:
4385         up_write(&fs_info->subvol_sem);
4386         if (err) {
4387                 spin_lock(&dest->root_item_lock);
4388                 root_flags = btrfs_root_flags(&dest->root_item);
4389                 btrfs_set_root_flags(&dest->root_item,
4390                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4391                 spin_unlock(&dest->root_item_lock);
4392         } else {
4393                 d_invalidate(dentry);
4394                 btrfs_prune_dentries(dest);
4395                 ASSERT(dest->send_in_progress == 0);
4396
4397                 /* the last ref */
4398                 if (dest->ino_cache_inode) {
4399                         iput(dest->ino_cache_inode);
4400                         dest->ino_cache_inode = NULL;
4401                 }
4402         }
4403
4404         return err;
4405 }
4406
4407 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4408 {
4409         struct inode *inode = d_inode(dentry);
4410         int err = 0;
4411         struct btrfs_root *root = BTRFS_I(dir)->root;
4412         struct btrfs_trans_handle *trans;
4413         u64 last_unlink_trans;
4414
4415         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4416                 return -ENOTEMPTY;
4417         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4418                 return btrfs_delete_subvolume(dir, dentry);
4419
4420         trans = __unlink_start_trans(dir);
4421         if (IS_ERR(trans))
4422                 return PTR_ERR(trans);
4423
4424         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4425                 err = btrfs_unlink_subvol(trans, dir,
4426                                           BTRFS_I(inode)->location.objectid,
4427                                           dentry->d_name.name,
4428                                           dentry->d_name.len);
4429                 goto out;
4430         }
4431
4432         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4433         if (err)
4434                 goto out;
4435
4436         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4437
4438         /* now the directory is empty */
4439         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4440                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4441                         dentry->d_name.len);
4442         if (!err) {
4443                 btrfs_i_size_write(BTRFS_I(inode), 0);
4444                 /*
4445                  * Propagate the last_unlink_trans value of the deleted dir to
4446                  * its parent directory. This is to prevent an unrecoverable
4447                  * log tree in the case we do something like this:
4448                  * 1) create dir foo
4449                  * 2) create snapshot under dir foo
4450                  * 3) delete the snapshot
4451                  * 4) rmdir foo
4452                  * 5) mkdir foo
4453                  * 6) fsync foo or some file inside foo
4454                  */
4455                 if (last_unlink_trans >= trans->transid)
4456                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4457         }
4458 out:
4459         btrfs_end_transaction(trans);
4460         btrfs_btree_balance_dirty(root->fs_info);
4461
4462         return err;
4463 }
4464
4465 static int truncate_space_check(struct btrfs_trans_handle *trans,
4466                                 struct btrfs_root *root,
4467                                 u64 bytes_deleted)
4468 {
4469         struct btrfs_fs_info *fs_info = root->fs_info;
4470         int ret;
4471
4472         /*
4473          * This is only used to apply pressure to the enospc system, we don't
4474          * intend to use this reservation at all.
4475          */
4476         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4477         bytes_deleted *= fs_info->nodesize;
4478         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4479                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4480         if (!ret) {
4481                 trace_btrfs_space_reservation(fs_info, "transaction",
4482                                               trans->transid,
4483                                               bytes_deleted, 1);
4484                 trans->bytes_reserved += bytes_deleted;
4485         }
4486         return ret;
4487
4488 }
4489
4490 /*
4491  * Return this if we need to call truncate_block for the last bit of the
4492  * truncate.
4493  */
4494 #define NEED_TRUNCATE_BLOCK 1
4495
4496 /*
4497  * this can truncate away extent items, csum items and directory items.
4498  * It starts at a high offset and removes keys until it can't find
4499  * any higher than new_size
4500  *
4501  * csum items that cross the new i_size are truncated to the new size
4502  * as well.
4503  *
4504  * min_type is the minimum key type to truncate down to.  If set to 0, this
4505  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4506  */
4507 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4508                                struct btrfs_root *root,
4509                                struct inode *inode,
4510                                u64 new_size, u32 min_type)
4511 {
4512         struct btrfs_fs_info *fs_info = root->fs_info;
4513         struct btrfs_path *path;
4514         struct extent_buffer *leaf;
4515         struct btrfs_file_extent_item *fi;
4516         struct btrfs_key key;
4517         struct btrfs_key found_key;
4518         u64 extent_start = 0;
4519         u64 extent_num_bytes = 0;
4520         u64 extent_offset = 0;
4521         u64 item_end = 0;
4522         u64 last_size = new_size;
4523         u32 found_type = (u8)-1;
4524         int found_extent;
4525         int del_item;
4526         int pending_del_nr = 0;
4527         int pending_del_slot = 0;
4528         int extent_type = -1;
4529         int ret;
4530         u64 ino = btrfs_ino(BTRFS_I(inode));
4531         u64 bytes_deleted = 0;
4532         bool be_nice = false;
4533         bool should_throttle = false;
4534         bool should_end = false;
4535
4536         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4537
4538         /*
4539          * for non-free space inodes and ref cows, we want to back off from
4540          * time to time
4541          */
4542         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4543             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4544                 be_nice = true;
4545
4546         path = btrfs_alloc_path();
4547         if (!path)
4548                 return -ENOMEM;
4549         path->reada = READA_BACK;
4550
4551         /*
4552          * We want to drop from the next block forward in case this new size is
4553          * not block aligned since we will be keeping the last block of the
4554          * extent just the way it is.
4555          */
4556         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4557             root == fs_info->tree_root)
4558                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4559                                         fs_info->sectorsize),
4560                                         (u64)-1, 0);
4561
4562         /*
4563          * This function is also used to drop the items in the log tree before
4564          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4565          * it is used to drop the loged items. So we shouldn't kill the delayed
4566          * items.
4567          */
4568         if (min_type == 0 && root == BTRFS_I(inode)->root)
4569                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4570
4571         key.objectid = ino;
4572         key.offset = (u64)-1;
4573         key.type = (u8)-1;
4574
4575 search_again:
4576         /*
4577          * with a 16K leaf size and 128MB extents, you can actually queue
4578          * up a huge file in a single leaf.  Most of the time that
4579          * bytes_deleted is > 0, it will be huge by the time we get here
4580          */
4581         if (be_nice && bytes_deleted > SZ_32M &&
4582             btrfs_should_end_transaction(trans)) {
4583                 ret = -EAGAIN;
4584                 goto out;
4585         }
4586
4587         path->leave_spinning = 1;
4588         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4589         if (ret < 0)
4590                 goto out;
4591
4592         if (ret > 0) {
4593                 ret = 0;
4594                 /* there are no items in the tree for us to truncate, we're
4595                  * done
4596                  */
4597                 if (path->slots[0] == 0)
4598                         goto out;
4599                 path->slots[0]--;
4600         }
4601
4602         while (1) {
4603                 fi = NULL;
4604                 leaf = path->nodes[0];
4605                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4606                 found_type = found_key.type;
4607
4608                 if (found_key.objectid != ino)
4609                         break;
4610
4611                 if (found_type < min_type)
4612                         break;
4613
4614                 item_end = found_key.offset;
4615                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4616                         fi = btrfs_item_ptr(leaf, path->slots[0],
4617                                             struct btrfs_file_extent_item);
4618                         extent_type = btrfs_file_extent_type(leaf, fi);
4619                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4620                                 item_end +=
4621                                     btrfs_file_extent_num_bytes(leaf, fi);
4622
4623                                 trace_btrfs_truncate_show_fi_regular(
4624                                         BTRFS_I(inode), leaf, fi,
4625                                         found_key.offset);
4626                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4627                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4628                                                                         fi);
4629
4630                                 trace_btrfs_truncate_show_fi_inline(
4631                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4632                                         found_key.offset);
4633                         }
4634                         item_end--;
4635                 }
4636                 if (found_type > min_type) {
4637                         del_item = 1;
4638                 } else {
4639                         if (item_end < new_size)
4640                                 break;
4641                         if (found_key.offset >= new_size)
4642                                 del_item = 1;
4643                         else
4644                                 del_item = 0;
4645                 }
4646                 found_extent = 0;
4647                 /* FIXME, shrink the extent if the ref count is only 1 */
4648                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4649                         goto delete;
4650
4651                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4652                         u64 num_dec;
4653                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4654                         if (!del_item) {
4655                                 u64 orig_num_bytes =
4656                                         btrfs_file_extent_num_bytes(leaf, fi);
4657                                 extent_num_bytes = ALIGN(new_size -
4658                                                 found_key.offset,
4659                                                 fs_info->sectorsize);
4660                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4661                                                          extent_num_bytes);
4662                                 num_dec = (orig_num_bytes -
4663                                            extent_num_bytes);
4664                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4665                                              &root->state) &&
4666                                     extent_start != 0)
4667                                         inode_sub_bytes(inode, num_dec);
4668                                 btrfs_mark_buffer_dirty(leaf);
4669                         } else {
4670                                 extent_num_bytes =
4671                                         btrfs_file_extent_disk_num_bytes(leaf,
4672                                                                          fi);
4673                                 extent_offset = found_key.offset -
4674                                         btrfs_file_extent_offset(leaf, fi);
4675
4676                                 /* FIXME blocksize != 4096 */
4677                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4678                                 if (extent_start != 0) {
4679                                         found_extent = 1;
4680                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4681                                                      &root->state))
4682                                                 inode_sub_bytes(inode, num_dec);
4683                                 }
4684                         }
4685                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4686                         /*
4687                          * we can't truncate inline items that have had
4688                          * special encodings
4689                          */
4690                         if (!del_item &&
4691                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4692                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4693                             btrfs_file_extent_compression(leaf, fi) == 0) {
4694                                 u32 size = (u32)(new_size - found_key.offset);
4695
4696                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4697                                 size = btrfs_file_extent_calc_inline_size(size);
4698                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4699                         } else if (!del_item) {
4700                                 /*
4701                                  * We have to bail so the last_size is set to
4702                                  * just before this extent.
4703                                  */
4704                                 ret = NEED_TRUNCATE_BLOCK;
4705                                 break;
4706                         }
4707
4708                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4709                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4710                 }
4711 delete:
4712                 if (del_item)
4713                         last_size = found_key.offset;
4714                 else
4715                         last_size = new_size;
4716                 if (del_item) {
4717                         if (!pending_del_nr) {
4718                                 /* no pending yet, add ourselves */
4719                                 pending_del_slot = path->slots[0];
4720                                 pending_del_nr = 1;
4721                         } else if (pending_del_nr &&
4722                                    path->slots[0] + 1 == pending_del_slot) {
4723                                 /* hop on the pending chunk */
4724                                 pending_del_nr++;
4725                                 pending_del_slot = path->slots[0];
4726                         } else {
4727                                 BUG();
4728                         }
4729                 } else {
4730                         break;
4731                 }
4732                 should_throttle = false;
4733
4734                 if (found_extent &&
4735                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4736                      root == fs_info->tree_root)) {
4737                         btrfs_set_path_blocking(path);
4738                         bytes_deleted += extent_num_bytes;
4739                         ret = btrfs_free_extent(trans, root, extent_start,
4740                                                 extent_num_bytes, 0,
4741                                                 btrfs_header_owner(leaf),
4742                                                 ino, extent_offset);
4743                         if (ret) {
4744                                 btrfs_abort_transaction(trans, ret);
4745                                 break;
4746                         }
4747                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4748                                 btrfs_async_run_delayed_refs(fs_info,
4749                                         trans->delayed_ref_updates * 2,
4750                                         trans->transid, 0);
4751                         if (be_nice) {
4752                                 if (truncate_space_check(trans, root,
4753                                                          extent_num_bytes)) {
4754                                         should_end = true;
4755                                 }
4756                                 if (btrfs_should_throttle_delayed_refs(trans,
4757                                                                        fs_info))
4758                                         should_throttle = true;
4759                         }
4760                 }
4761
4762                 if (found_type == BTRFS_INODE_ITEM_KEY)
4763                         break;
4764
4765                 if (path->slots[0] == 0 ||
4766                     path->slots[0] != pending_del_slot ||
4767                     should_throttle || should_end) {
4768                         if (pending_del_nr) {
4769                                 ret = btrfs_del_items(trans, root, path,
4770                                                 pending_del_slot,
4771                                                 pending_del_nr);
4772                                 if (ret) {
4773                                         btrfs_abort_transaction(trans, ret);
4774                                         break;
4775                                 }
4776                                 pending_del_nr = 0;
4777                         }
4778                         btrfs_release_path(path);
4779                         if (should_throttle) {
4780                                 unsigned long updates = trans->delayed_ref_updates;
4781                                 if (updates) {
4782                                         trans->delayed_ref_updates = 0;
4783                                         ret = btrfs_run_delayed_refs(trans,
4784                                                                    updates * 2);
4785                                         if (ret)
4786                                                 break;
4787                                 }
4788                         }
4789                         /*
4790                          * if we failed to refill our space rsv, bail out
4791                          * and let the transaction restart
4792                          */
4793                         if (should_end) {
4794                                 ret = -EAGAIN;
4795                                 break;
4796                         }
4797                         goto search_again;
4798                 } else {
4799                         path->slots[0]--;
4800                 }
4801         }
4802 out:
4803         if (ret >= 0 && pending_del_nr) {
4804                 int err;
4805
4806                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4807                                       pending_del_nr);
4808                 if (err) {
4809                         btrfs_abort_transaction(trans, err);
4810                         ret = err;
4811                 }
4812         }
4813         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4814                 ASSERT(last_size >= new_size);
4815                 if (!ret && last_size > new_size)
4816                         last_size = new_size;
4817                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4818         }
4819
4820         btrfs_free_path(path);
4821
4822         if (be_nice && bytes_deleted > SZ_32M && (ret >= 0 || ret == -EAGAIN)) {
4823                 unsigned long updates = trans->delayed_ref_updates;
4824                 int err;
4825
4826                 if (updates) {
4827                         trans->delayed_ref_updates = 0;
4828                         err = btrfs_run_delayed_refs(trans, updates * 2);
4829                         if (err)
4830                                 ret = err;
4831                 }
4832         }
4833         return ret;
4834 }
4835
4836 /*
4837  * btrfs_truncate_block - read, zero a chunk and write a block
4838  * @inode - inode that we're zeroing
4839  * @from - the offset to start zeroing
4840  * @len - the length to zero, 0 to zero the entire range respective to the
4841  *      offset
4842  * @front - zero up to the offset instead of from the offset on
4843  *
4844  * This will find the block for the "from" offset and cow the block and zero the
4845  * part we want to zero.  This is used with truncate and hole punching.
4846  */
4847 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4848                         int front)
4849 {
4850         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4851         struct address_space *mapping = inode->i_mapping;
4852         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4853         struct btrfs_ordered_extent *ordered;
4854         struct extent_state *cached_state = NULL;
4855         struct extent_changeset *data_reserved = NULL;
4856         char *kaddr;
4857         u32 blocksize = fs_info->sectorsize;
4858         pgoff_t index = from >> PAGE_SHIFT;
4859         unsigned offset = from & (blocksize - 1);
4860         struct page *page;
4861         gfp_t mask = btrfs_alloc_write_mask(mapping);
4862         int ret = 0;
4863         u64 block_start;
4864         u64 block_end;
4865
4866         if (IS_ALIGNED(offset, blocksize) &&
4867             (!len || IS_ALIGNED(len, blocksize)))
4868                 goto out;
4869
4870         block_start = round_down(from, blocksize);
4871         block_end = block_start + blocksize - 1;
4872
4873         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4874                                            block_start, blocksize);
4875         if (ret)
4876                 goto out;
4877
4878 again:
4879         page = find_or_create_page(mapping, index, mask);
4880         if (!page) {
4881                 btrfs_delalloc_release_space(inode, data_reserved,
4882                                              block_start, blocksize, true);
4883                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, true);
4884                 ret = -ENOMEM;
4885                 goto out;
4886         }
4887
4888         if (!PageUptodate(page)) {
4889                 ret = btrfs_readpage(NULL, page);
4890                 lock_page(page);
4891                 if (page->mapping != mapping) {
4892                         unlock_page(page);
4893                         put_page(page);
4894                         goto again;
4895                 }
4896                 if (!PageUptodate(page)) {
4897                         ret = -EIO;
4898                         goto out_unlock;
4899                 }
4900         }
4901         wait_on_page_writeback(page);
4902
4903         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4904         set_page_extent_mapped(page);
4905
4906         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4907         if (ordered) {
4908                 unlock_extent_cached(io_tree, block_start, block_end,
4909                                      &cached_state);
4910                 unlock_page(page);
4911                 put_page(page);
4912                 btrfs_start_ordered_extent(inode, ordered, 1);
4913                 btrfs_put_ordered_extent(ordered);
4914                 goto again;
4915         }
4916
4917         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4918                           EXTENT_DIRTY | EXTENT_DELALLOC |
4919                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4920                           0, 0, &cached_state);
4921
4922         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4923                                         &cached_state, 0);
4924         if (ret) {
4925                 unlock_extent_cached(io_tree, block_start, block_end,
4926                                      &cached_state);
4927                 goto out_unlock;
4928         }
4929
4930         if (offset != blocksize) {
4931                 if (!len)
4932                         len = blocksize - offset;
4933                 kaddr = kmap(page);
4934                 if (front)
4935                         memset(kaddr + (block_start - page_offset(page)),
4936                                 0, offset);
4937                 else
4938                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4939                                 0, len);
4940                 flush_dcache_page(page);
4941                 kunmap(page);
4942         }
4943         ClearPageChecked(page);
4944         set_page_dirty(page);
4945         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4946
4947 out_unlock:
4948         if (ret)
4949                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4950                                              blocksize, true);
4951         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, (ret != 0));
4952         unlock_page(page);
4953         put_page(page);
4954 out:
4955         extent_changeset_free(data_reserved);
4956         return ret;
4957 }
4958
4959 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4960                              u64 offset, u64 len)
4961 {
4962         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4963         struct btrfs_trans_handle *trans;
4964         int ret;
4965
4966         /*
4967          * Still need to make sure the inode looks like it's been updated so
4968          * that any holes get logged if we fsync.
4969          */
4970         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4971                 BTRFS_I(inode)->last_trans = fs_info->generation;
4972                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4973                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4974                 return 0;
4975         }
4976
4977         /*
4978          * 1 - for the one we're dropping
4979          * 1 - for the one we're adding
4980          * 1 - for updating the inode.
4981          */
4982         trans = btrfs_start_transaction(root, 3);
4983         if (IS_ERR(trans))
4984                 return PTR_ERR(trans);
4985
4986         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4987         if (ret) {
4988                 btrfs_abort_transaction(trans, ret);
4989                 btrfs_end_transaction(trans);
4990                 return ret;
4991         }
4992
4993         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4994                         offset, 0, 0, len, 0, len, 0, 0, 0);
4995         if (ret)
4996                 btrfs_abort_transaction(trans, ret);
4997         else
4998                 btrfs_update_inode(trans, root, inode);
4999         btrfs_end_transaction(trans);
5000         return ret;
5001 }
5002
5003 /*
5004  * This function puts in dummy file extents for the area we're creating a hole
5005  * for.  So if we are truncating this file to a larger size we need to insert
5006  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
5007  * the range between oldsize and size
5008  */
5009 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
5010 {
5011         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5012         struct btrfs_root *root = BTRFS_I(inode)->root;
5013         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5014         struct extent_map *em = NULL;
5015         struct extent_state *cached_state = NULL;
5016         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5017         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
5018         u64 block_end = ALIGN(size, fs_info->sectorsize);
5019         u64 last_byte;
5020         u64 cur_offset;
5021         u64 hole_size;
5022         int err = 0;
5023
5024         /*
5025          * If our size started in the middle of a block we need to zero out the
5026          * rest of the block before we expand the i_size, otherwise we could
5027          * expose stale data.
5028          */
5029         err = btrfs_truncate_block(inode, oldsize, 0, 0);
5030         if (err)
5031                 return err;
5032
5033         if (size <= hole_start)
5034                 return 0;
5035
5036         while (1) {
5037                 struct btrfs_ordered_extent *ordered;
5038
5039                 lock_extent_bits(io_tree, hole_start, block_end - 1,
5040                                  &cached_state);
5041                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
5042                                                      block_end - hole_start);
5043                 if (!ordered)
5044                         break;
5045                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
5046                                      &cached_state);
5047                 btrfs_start_ordered_extent(inode, ordered, 1);
5048                 btrfs_put_ordered_extent(ordered);
5049         }
5050
5051         cur_offset = hole_start;
5052         while (1) {
5053                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5054                                 block_end - cur_offset, 0);
5055                 if (IS_ERR(em)) {
5056                         err = PTR_ERR(em);
5057                         em = NULL;
5058                         break;
5059                 }
5060                 last_byte = min(extent_map_end(em), block_end);
5061                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5062                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5063                         struct extent_map *hole_em;
5064                         hole_size = last_byte - cur_offset;
5065
5066                         err = maybe_insert_hole(root, inode, cur_offset,
5067                                                 hole_size);
5068                         if (err)
5069                                 break;
5070                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5071                                                 cur_offset + hole_size - 1, 0);
5072                         hole_em = alloc_extent_map();
5073                         if (!hole_em) {
5074                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5075                                         &BTRFS_I(inode)->runtime_flags);
5076                                 goto next;
5077                         }
5078                         hole_em->start = cur_offset;
5079                         hole_em->len = hole_size;
5080                         hole_em->orig_start = cur_offset;
5081
5082                         hole_em->block_start = EXTENT_MAP_HOLE;
5083                         hole_em->block_len = 0;
5084                         hole_em->orig_block_len = 0;
5085                         hole_em->ram_bytes = hole_size;
5086                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5087                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5088                         hole_em->generation = fs_info->generation;
5089
5090                         while (1) {
5091                                 write_lock(&em_tree->lock);
5092                                 err = add_extent_mapping(em_tree, hole_em, 1);
5093                                 write_unlock(&em_tree->lock);
5094                                 if (err != -EEXIST)
5095                                         break;
5096                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5097                                                         cur_offset,
5098                                                         cur_offset +
5099                                                         hole_size - 1, 0);
5100                         }
5101                         free_extent_map(hole_em);
5102                 }
5103 next:
5104                 free_extent_map(em);
5105                 em = NULL;
5106                 cur_offset = last_byte;
5107                 if (cur_offset >= block_end)
5108                         break;
5109         }
5110         free_extent_map(em);
5111         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5112         return err;
5113 }
5114
5115 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5116 {
5117         struct btrfs_root *root = BTRFS_I(inode)->root;
5118         struct btrfs_trans_handle *trans;
5119         loff_t oldsize = i_size_read(inode);
5120         loff_t newsize = attr->ia_size;
5121         int mask = attr->ia_valid;
5122         int ret;
5123
5124         /*
5125          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5126          * special case where we need to update the times despite not having
5127          * these flags set.  For all other operations the VFS set these flags
5128          * explicitly if it wants a timestamp update.
5129          */
5130         if (newsize != oldsize) {
5131                 inode_inc_iversion(inode);
5132                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5133                         inode->i_ctime = inode->i_mtime =
5134                                 current_time(inode);
5135         }
5136
5137         if (newsize > oldsize) {
5138                 /*
5139                  * Don't do an expanding truncate while snapshotting is ongoing.
5140                  * This is to ensure the snapshot captures a fully consistent
5141                  * state of this file - if the snapshot captures this expanding
5142                  * truncation, it must capture all writes that happened before
5143                  * this truncation.
5144                  */
5145                 btrfs_wait_for_snapshot_creation(root);
5146                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5147                 if (ret) {
5148                         btrfs_end_write_no_snapshotting(root);
5149                         return ret;
5150                 }
5151
5152                 trans = btrfs_start_transaction(root, 1);
5153                 if (IS_ERR(trans)) {
5154                         btrfs_end_write_no_snapshotting(root);
5155                         return PTR_ERR(trans);
5156                 }
5157
5158                 i_size_write(inode, newsize);
5159                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5160                 pagecache_isize_extended(inode, oldsize, newsize);
5161                 ret = btrfs_update_inode(trans, root, inode);
5162                 btrfs_end_write_no_snapshotting(root);
5163                 btrfs_end_transaction(trans);
5164         } else {
5165
5166                 /*
5167                  * We're truncating a file that used to have good data down to
5168                  * zero. Make sure it gets into the ordered flush list so that
5169                  * any new writes get down to disk quickly.
5170                  */
5171                 if (newsize == 0)
5172                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5173                                 &BTRFS_I(inode)->runtime_flags);
5174
5175                 truncate_setsize(inode, newsize);
5176
5177                 /* Disable nonlocked read DIO to avoid the end less truncate */
5178                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5179                 inode_dio_wait(inode);
5180                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5181
5182                 ret = btrfs_truncate(inode, newsize == oldsize);
5183                 if (ret && inode->i_nlink) {
5184                         int err;
5185
5186                         /*
5187                          * Truncate failed, so fix up the in-memory size. We
5188                          * adjusted disk_i_size down as we removed extents, so
5189                          * wait for disk_i_size to be stable and then update the
5190                          * in-memory size to match.
5191                          */
5192                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5193                         if (err)
5194                                 return err;
5195                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5196                 }
5197         }
5198
5199         return ret;
5200 }
5201
5202 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5203 {
5204         struct inode *inode = d_inode(dentry);
5205         struct btrfs_root *root = BTRFS_I(inode)->root;
5206         int err;
5207
5208         if (btrfs_root_readonly(root))
5209                 return -EROFS;
5210
5211         err = setattr_prepare(dentry, attr);
5212         if (err)
5213                 return err;
5214
5215         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5216                 err = btrfs_setsize(inode, attr);
5217                 if (err)
5218                         return err;
5219         }
5220
5221         if (attr->ia_valid) {
5222                 setattr_copy(inode, attr);
5223                 inode_inc_iversion(inode);
5224                 err = btrfs_dirty_inode(inode);
5225
5226                 if (!err && attr->ia_valid & ATTR_MODE)
5227                         err = posix_acl_chmod(inode, inode->i_mode);
5228         }
5229
5230         return err;
5231 }
5232
5233 /*
5234  * While truncating the inode pages during eviction, we get the VFS calling
5235  * btrfs_invalidatepage() against each page of the inode. This is slow because
5236  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5237  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5238  * extent_state structures over and over, wasting lots of time.
5239  *
5240  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5241  * those expensive operations on a per page basis and do only the ordered io
5242  * finishing, while we release here the extent_map and extent_state structures,
5243  * without the excessive merging and splitting.
5244  */
5245 static void evict_inode_truncate_pages(struct inode *inode)
5246 {
5247         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5248         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5249         struct rb_node *node;
5250
5251         ASSERT(inode->i_state & I_FREEING);
5252         truncate_inode_pages_final(&inode->i_data);
5253
5254         write_lock(&map_tree->lock);
5255         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5256                 struct extent_map *em;
5257
5258                 node = rb_first(&map_tree->map);
5259                 em = rb_entry(node, struct extent_map, rb_node);
5260                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5261                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5262                 remove_extent_mapping(map_tree, em);
5263                 free_extent_map(em);
5264                 if (need_resched()) {
5265                         write_unlock(&map_tree->lock);
5266                         cond_resched();
5267                         write_lock(&map_tree->lock);
5268                 }
5269         }
5270         write_unlock(&map_tree->lock);
5271
5272         /*
5273          * Keep looping until we have no more ranges in the io tree.
5274          * We can have ongoing bios started by readpages (called from readahead)
5275          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5276          * still in progress (unlocked the pages in the bio but did not yet
5277          * unlocked the ranges in the io tree). Therefore this means some
5278          * ranges can still be locked and eviction started because before
5279          * submitting those bios, which are executed by a separate task (work
5280          * queue kthread), inode references (inode->i_count) were not taken
5281          * (which would be dropped in the end io callback of each bio).
5282          * Therefore here we effectively end up waiting for those bios and
5283          * anyone else holding locked ranges without having bumped the inode's
5284          * reference count - if we don't do it, when they access the inode's
5285          * io_tree to unlock a range it may be too late, leading to an
5286          * use-after-free issue.
5287          */
5288         spin_lock(&io_tree->lock);
5289         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5290                 struct extent_state *state;
5291                 struct extent_state *cached_state = NULL;
5292                 u64 start;
5293                 u64 end;
5294
5295                 node = rb_first(&io_tree->state);
5296                 state = rb_entry(node, struct extent_state, rb_node);
5297                 start = state->start;
5298                 end = state->end;
5299                 spin_unlock(&io_tree->lock);
5300
5301                 lock_extent_bits(io_tree, start, end, &cached_state);
5302
5303                 /*
5304                  * If still has DELALLOC flag, the extent didn't reach disk,
5305                  * and its reserved space won't be freed by delayed_ref.
5306                  * So we need to free its reserved space here.
5307                  * (Refer to comment in btrfs_invalidatepage, case 2)
5308                  *
5309                  * Note, end is the bytenr of last byte, so we need + 1 here.
5310                  */
5311                 if (state->state & EXTENT_DELALLOC)
5312                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5313
5314                 clear_extent_bit(io_tree, start, end,
5315                                  EXTENT_LOCKED | EXTENT_DIRTY |
5316                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5317                                  EXTENT_DEFRAG, 1, 1, &cached_state);
5318
5319                 cond_resched();
5320                 spin_lock(&io_tree->lock);
5321         }
5322         spin_unlock(&io_tree->lock);
5323 }
5324
5325 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
5326                                                         struct btrfs_block_rsv *rsv,
5327                                                         u64 min_size)
5328 {
5329         struct btrfs_fs_info *fs_info = root->fs_info;
5330         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5331         int failures = 0;
5332
5333         for (;;) {
5334                 struct btrfs_trans_handle *trans;
5335                 int ret;
5336
5337                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5338                                              BTRFS_RESERVE_FLUSH_LIMIT);
5339
5340                 if (ret && ++failures > 2) {
5341                         btrfs_warn(fs_info,
5342                                    "could not allocate space for a delete; will truncate on mount");
5343                         return ERR_PTR(-ENOSPC);
5344                 }
5345
5346                 trans = btrfs_join_transaction(root);
5347                 if (IS_ERR(trans) || !ret)
5348                         return trans;
5349
5350                 /*
5351                  * Try to steal from the global reserve if there is space for
5352                  * it.
5353                  */
5354                 if (!btrfs_check_space_for_delayed_refs(trans, fs_info) &&
5355                     !btrfs_block_rsv_migrate(global_rsv, rsv, min_size, 0))
5356                         return trans;
5357
5358                 /* If not, commit and try again. */
5359                 ret = btrfs_commit_transaction(trans);
5360                 if (ret)
5361                         return ERR_PTR(ret);
5362         }
5363 }
5364
5365 void btrfs_evict_inode(struct inode *inode)
5366 {
5367         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5368         struct btrfs_trans_handle *trans;
5369         struct btrfs_root *root = BTRFS_I(inode)->root;
5370         struct btrfs_block_rsv *rsv;
5371         u64 min_size;
5372         int ret;
5373
5374         trace_btrfs_inode_evict(inode);
5375
5376         if (!root) {
5377                 clear_inode(inode);
5378                 return;
5379         }
5380
5381         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5382
5383         evict_inode_truncate_pages(inode);
5384
5385         if (inode->i_nlink &&
5386             ((btrfs_root_refs(&root->root_item) != 0 &&
5387               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5388              btrfs_is_free_space_inode(BTRFS_I(inode))))
5389                 goto no_delete;
5390
5391         if (is_bad_inode(inode))
5392                 goto no_delete;
5393         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5394         if (!special_file(inode->i_mode))
5395                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5396
5397         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5398
5399         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5400                 goto no_delete;
5401
5402         if (inode->i_nlink > 0) {
5403                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5404                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5405                 goto no_delete;
5406         }
5407
5408         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5409         if (ret)
5410                 goto no_delete;
5411
5412         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5413         if (!rsv)
5414                 goto no_delete;
5415         rsv->size = min_size;
5416         rsv->failfast = 1;
5417
5418         btrfs_i_size_write(BTRFS_I(inode), 0);
5419
5420         while (1) {
5421                 trans = evict_refill_and_join(root, rsv, min_size);
5422                 if (IS_ERR(trans))
5423                         goto free_rsv;
5424
5425                 trans->block_rsv = rsv;
5426
5427                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5428                 trans->block_rsv = &fs_info->trans_block_rsv;
5429                 btrfs_end_transaction(trans);
5430                 btrfs_btree_balance_dirty(fs_info);
5431                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5432                         goto free_rsv;
5433                 else if (!ret)
5434                         break;
5435         }
5436
5437         /*
5438          * Errors here aren't a big deal, it just means we leave orphan items in
5439          * the tree. They will be cleaned up on the next mount. If the inode
5440          * number gets reused, cleanup deletes the orphan item without doing
5441          * anything, and unlink reuses the existing orphan item.
5442          *
5443          * If it turns out that we are dropping too many of these, we might want
5444          * to add a mechanism for retrying these after a commit.
5445          */
5446         trans = evict_refill_and_join(root, rsv, min_size);
5447         if (!IS_ERR(trans)) {
5448                 trans->block_rsv = rsv;
5449                 btrfs_orphan_del(trans, BTRFS_I(inode));
5450                 trans->block_rsv = &fs_info->trans_block_rsv;
5451                 btrfs_end_transaction(trans);
5452         }
5453
5454         if (!(root == fs_info->tree_root ||
5455               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5456                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5457
5458 free_rsv:
5459         btrfs_free_block_rsv(fs_info, rsv);
5460 no_delete:
5461         /*
5462          * If we didn't successfully delete, the orphan item will still be in
5463          * the tree and we'll retry on the next mount. Again, we might also want
5464          * to retry these periodically in the future.
5465          */
5466         btrfs_remove_delayed_node(BTRFS_I(inode));
5467         clear_inode(inode);
5468 }
5469
5470 /*
5471  * this returns the key found in the dir entry in the location pointer.
5472  * If no dir entries were found, returns -ENOENT.
5473  * If found a corrupted location in dir entry, returns -EUCLEAN.
5474  */
5475 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5476                                struct btrfs_key *location)
5477 {
5478         const char *name = dentry->d_name.name;
5479         int namelen = dentry->d_name.len;
5480         struct btrfs_dir_item *di;
5481         struct btrfs_path *path;
5482         struct btrfs_root *root = BTRFS_I(dir)->root;
5483         int ret = 0;
5484
5485         path = btrfs_alloc_path();
5486         if (!path)
5487                 return -ENOMEM;
5488
5489         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5490                         name, namelen, 0);
5491         if (!di) {
5492                 ret = -ENOENT;
5493                 goto out;
5494         }
5495         if (IS_ERR(di)) {
5496                 ret = PTR_ERR(di);
5497                 goto out;
5498         }
5499
5500         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5501         if (location->type != BTRFS_INODE_ITEM_KEY &&
5502             location->type != BTRFS_ROOT_ITEM_KEY) {
5503                 ret = -EUCLEAN;
5504                 btrfs_warn(root->fs_info,
5505 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5506                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5507                            location->objectid, location->type, location->offset);
5508         }
5509 out:
5510         btrfs_free_path(path);
5511         return ret;
5512 }
5513
5514 /*
5515  * when we hit a tree root in a directory, the btrfs part of the inode
5516  * needs to be changed to reflect the root directory of the tree root.  This
5517  * is kind of like crossing a mount point.
5518  */
5519 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5520                                     struct inode *dir,
5521                                     struct dentry *dentry,
5522                                     struct btrfs_key *location,
5523                                     struct btrfs_root **sub_root)
5524 {
5525         struct btrfs_path *path;
5526         struct btrfs_root *new_root;
5527         struct btrfs_root_ref *ref;
5528         struct extent_buffer *leaf;
5529         struct btrfs_key key;
5530         int ret;
5531         int err = 0;
5532
5533         path = btrfs_alloc_path();
5534         if (!path) {
5535                 err = -ENOMEM;
5536                 goto out;
5537         }
5538
5539         err = -ENOENT;
5540         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5541         key.type = BTRFS_ROOT_REF_KEY;
5542         key.offset = location->objectid;
5543
5544         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5545         if (ret) {
5546                 if (ret < 0)
5547                         err = ret;
5548                 goto out;
5549         }
5550
5551         leaf = path->nodes[0];
5552         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5553         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5554             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5555                 goto out;
5556
5557         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5558                                    (unsigned long)(ref + 1),
5559                                    dentry->d_name.len);
5560         if (ret)
5561                 goto out;
5562
5563         btrfs_release_path(path);
5564
5565         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5566         if (IS_ERR(new_root)) {
5567                 err = PTR_ERR(new_root);
5568                 goto out;
5569         }
5570
5571         *sub_root = new_root;
5572         location->objectid = btrfs_root_dirid(&new_root->root_item);
5573         location->type = BTRFS_INODE_ITEM_KEY;
5574         location->offset = 0;
5575         err = 0;
5576 out:
5577         btrfs_free_path(path);
5578         return err;
5579 }
5580
5581 static void inode_tree_add(struct inode *inode)
5582 {
5583         struct btrfs_root *root = BTRFS_I(inode)->root;
5584         struct btrfs_inode *entry;
5585         struct rb_node **p;
5586         struct rb_node *parent;
5587         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5588         u64 ino = btrfs_ino(BTRFS_I(inode));
5589
5590         if (inode_unhashed(inode))
5591                 return;
5592         parent = NULL;
5593         spin_lock(&root->inode_lock);
5594         p = &root->inode_tree.rb_node;
5595         while (*p) {
5596                 parent = *p;
5597                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5598
5599                 if (ino < btrfs_ino(entry))
5600                         p = &parent->rb_left;
5601                 else if (ino > btrfs_ino(entry))
5602                         p = &parent->rb_right;
5603                 else {
5604                         WARN_ON(!(entry->vfs_inode.i_state &
5605                                   (I_WILL_FREE | I_FREEING)));
5606                         rb_replace_node(parent, new, &root->inode_tree);
5607                         RB_CLEAR_NODE(parent);
5608                         spin_unlock(&root->inode_lock);
5609                         return;
5610                 }
5611         }
5612         rb_link_node(new, parent, p);
5613         rb_insert_color(new, &root->inode_tree);
5614         spin_unlock(&root->inode_lock);
5615 }
5616
5617 static void inode_tree_del(struct inode *inode)
5618 {
5619         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5620         struct btrfs_root *root = BTRFS_I(inode)->root;
5621         int empty = 0;
5622
5623         spin_lock(&root->inode_lock);
5624         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5625                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5626                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5627                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5628         }
5629         spin_unlock(&root->inode_lock);
5630
5631         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5632                 synchronize_srcu(&fs_info->subvol_srcu);
5633                 spin_lock(&root->inode_lock);
5634                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5635                 spin_unlock(&root->inode_lock);
5636                 if (empty)
5637                         btrfs_add_dead_root(root);
5638         }
5639 }
5640
5641
5642 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5643 {
5644         struct btrfs_iget_args *args = p;
5645         inode->i_ino = args->location->objectid;
5646         memcpy(&BTRFS_I(inode)->location, args->location,
5647                sizeof(*args->location));
5648         BTRFS_I(inode)->root = args->root;
5649         return 0;
5650 }
5651
5652 static int btrfs_find_actor(struct inode *inode, void *opaque)
5653 {
5654         struct btrfs_iget_args *args = opaque;
5655         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5656                 args->root == BTRFS_I(inode)->root;
5657 }
5658
5659 static struct inode *btrfs_iget_locked(struct super_block *s,
5660                                        struct btrfs_key *location,
5661                                        struct btrfs_root *root)
5662 {
5663         struct inode *inode;
5664         struct btrfs_iget_args args;
5665         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5666
5667         args.location = location;
5668         args.root = root;
5669
5670         inode = iget5_locked(s, hashval, btrfs_find_actor,
5671                              btrfs_init_locked_inode,
5672                              (void *)&args);
5673         return inode;
5674 }
5675
5676 /* Get an inode object given its location and corresponding root.
5677  * Returns in *is_new if the inode was read from disk
5678  */
5679 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5680                          struct btrfs_root *root, int *new)
5681 {
5682         struct inode *inode;
5683
5684         inode = btrfs_iget_locked(s, location, root);
5685         if (!inode)
5686                 return ERR_PTR(-ENOMEM);
5687
5688         if (inode->i_state & I_NEW) {
5689                 int ret;
5690
5691                 ret = btrfs_read_locked_inode(inode);
5692                 if (!ret) {
5693                         inode_tree_add(inode);
5694                         unlock_new_inode(inode);
5695                         if (new)
5696                                 *new = 1;
5697                 } else {
5698                         iget_failed(inode);
5699                         /*
5700                          * ret > 0 can come from btrfs_search_slot called by
5701                          * btrfs_read_locked_inode, this means the inode item
5702                          * was not found.
5703                          */
5704                         if (ret > 0)
5705                                 ret = -ENOENT;
5706                         inode = ERR_PTR(ret);
5707                 }
5708         }
5709
5710         return inode;
5711 }
5712
5713 static struct inode *new_simple_dir(struct super_block *s,
5714                                     struct btrfs_key *key,
5715                                     struct btrfs_root *root)
5716 {
5717         struct inode *inode = new_inode(s);
5718
5719         if (!inode)
5720                 return ERR_PTR(-ENOMEM);
5721
5722         BTRFS_I(inode)->root = root;
5723         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5724         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5725
5726         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5727         inode->i_op = &btrfs_dir_ro_inode_operations;
5728         inode->i_opflags &= ~IOP_XATTR;
5729         inode->i_fop = &simple_dir_operations;
5730         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5731         inode->i_mtime = current_time(inode);
5732         inode->i_atime = inode->i_mtime;
5733         inode->i_ctime = inode->i_mtime;
5734         BTRFS_I(inode)->i_otime = inode->i_mtime;
5735
5736         return inode;
5737 }
5738
5739 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5740 {
5741         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5742         struct inode *inode;
5743         struct btrfs_root *root = BTRFS_I(dir)->root;
5744         struct btrfs_root *sub_root = root;
5745         struct btrfs_key location;
5746         int index;
5747         int ret = 0;
5748
5749         if (dentry->d_name.len > BTRFS_NAME_LEN)
5750                 return ERR_PTR(-ENAMETOOLONG);
5751
5752         ret = btrfs_inode_by_name(dir, dentry, &location);
5753         if (ret < 0)
5754                 return ERR_PTR(ret);
5755
5756         if (location.type == BTRFS_INODE_ITEM_KEY) {
5757                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5758                 return inode;
5759         }
5760
5761         index = srcu_read_lock(&fs_info->subvol_srcu);
5762         ret = fixup_tree_root_location(fs_info, dir, dentry,
5763                                        &location, &sub_root);
5764         if (ret < 0) {
5765                 if (ret != -ENOENT)
5766                         inode = ERR_PTR(ret);
5767                 else
5768                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5769         } else {
5770                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5771         }
5772         srcu_read_unlock(&fs_info->subvol_srcu, index);
5773
5774         if (!IS_ERR(inode) && root != sub_root) {
5775                 down_read(&fs_info->cleanup_work_sem);
5776                 if (!sb_rdonly(inode->i_sb))
5777                         ret = btrfs_orphan_cleanup(sub_root);
5778                 up_read(&fs_info->cleanup_work_sem);
5779                 if (ret) {
5780                         iput(inode);
5781                         inode = ERR_PTR(ret);
5782                 }
5783         }
5784
5785         return inode;
5786 }
5787
5788 static int btrfs_dentry_delete(const struct dentry *dentry)
5789 {
5790         struct btrfs_root *root;
5791         struct inode *inode = d_inode(dentry);
5792
5793         if (!inode && !IS_ROOT(dentry))
5794                 inode = d_inode(dentry->d_parent);
5795
5796         if (inode) {
5797                 root = BTRFS_I(inode)->root;
5798                 if (btrfs_root_refs(&root->root_item) == 0)
5799                         return 1;
5800
5801                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5802                         return 1;
5803         }
5804         return 0;
5805 }
5806
5807 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5808                                    unsigned int flags)
5809 {
5810         struct inode *inode;
5811
5812         inode = btrfs_lookup_dentry(dir, dentry);
5813         if (IS_ERR(inode)) {
5814                 if (PTR_ERR(inode) == -ENOENT)
5815                         inode = NULL;
5816                 else
5817                         return ERR_CAST(inode);
5818         }
5819
5820         return d_splice_alias(inode, dentry);
5821 }
5822
5823 unsigned char btrfs_filetype_table[] = {
5824         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5825 };
5826
5827 /*
5828  * All this infrastructure exists because dir_emit can fault, and we are holding
5829  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5830  * our information into that, and then dir_emit from the buffer.  This is
5831  * similar to what NFS does, only we don't keep the buffer around in pagecache
5832  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5833  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5834  * tree lock.
5835  */
5836 static int btrfs_opendir(struct inode *inode, struct file *file)
5837 {
5838         struct btrfs_file_private *private;
5839
5840         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5841         if (!private)
5842                 return -ENOMEM;
5843         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5844         if (!private->filldir_buf) {
5845                 kfree(private);
5846                 return -ENOMEM;
5847         }
5848         file->private_data = private;
5849         return 0;
5850 }
5851
5852 struct dir_entry {
5853         u64 ino;
5854         u64 offset;
5855         unsigned type;
5856         int name_len;
5857 };
5858
5859 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5860 {
5861         while (entries--) {
5862                 struct dir_entry *entry = addr;
5863                 char *name = (char *)(entry + 1);
5864
5865                 ctx->pos = get_unaligned(&entry->offset);
5866                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
5867                                          get_unaligned(&entry->ino),
5868                                          get_unaligned(&entry->type)))
5869                         return 1;
5870                 addr += sizeof(struct dir_entry) +
5871                         get_unaligned(&entry->name_len);
5872                 ctx->pos++;
5873         }
5874         return 0;
5875 }
5876
5877 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5878 {
5879         struct inode *inode = file_inode(file);
5880         struct btrfs_root *root = BTRFS_I(inode)->root;
5881         struct btrfs_file_private *private = file->private_data;
5882         struct btrfs_dir_item *di;
5883         struct btrfs_key key;
5884         struct btrfs_key found_key;
5885         struct btrfs_path *path;
5886         void *addr;
5887         struct list_head ins_list;
5888         struct list_head del_list;
5889         int ret;
5890         struct extent_buffer *leaf;
5891         int slot;
5892         char *name_ptr;
5893         int name_len;
5894         int entries = 0;
5895         int total_len = 0;
5896         bool put = false;
5897         struct btrfs_key location;
5898
5899         if (!dir_emit_dots(file, ctx))
5900                 return 0;
5901
5902         path = btrfs_alloc_path();
5903         if (!path)
5904                 return -ENOMEM;
5905
5906         addr = private->filldir_buf;
5907         path->reada = READA_FORWARD;
5908
5909         INIT_LIST_HEAD(&ins_list);
5910         INIT_LIST_HEAD(&del_list);
5911         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5912
5913 again:
5914         key.type = BTRFS_DIR_INDEX_KEY;
5915         key.offset = ctx->pos;
5916         key.objectid = btrfs_ino(BTRFS_I(inode));
5917
5918         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5919         if (ret < 0)
5920                 goto err;
5921
5922         while (1) {
5923                 struct dir_entry *entry;
5924
5925                 leaf = path->nodes[0];
5926                 slot = path->slots[0];
5927                 if (slot >= btrfs_header_nritems(leaf)) {
5928                         ret = btrfs_next_leaf(root, path);
5929                         if (ret < 0)
5930                                 goto err;
5931                         else if (ret > 0)
5932                                 break;
5933                         continue;
5934                 }
5935
5936                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5937
5938                 if (found_key.objectid != key.objectid)
5939                         break;
5940                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5941                         break;
5942                 if (found_key.offset < ctx->pos)
5943                         goto next;
5944                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5945                         goto next;
5946                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5947                 name_len = btrfs_dir_name_len(leaf, di);
5948                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5949                     PAGE_SIZE) {
5950                         btrfs_release_path(path);
5951                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5952                         if (ret)
5953                                 goto nopos;
5954                         addr = private->filldir_buf;
5955                         entries = 0;
5956                         total_len = 0;
5957                         goto again;
5958                 }
5959
5960                 entry = addr;
5961                 put_unaligned(name_len, &entry->name_len);
5962                 name_ptr = (char *)(entry + 1);
5963                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5964                                    name_len);
5965                 put_unaligned(btrfs_filetype_table[btrfs_dir_type(leaf, di)],
5966                                 &entry->type);
5967                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5968                 put_unaligned(location.objectid, &entry->ino);
5969                 put_unaligned(found_key.offset, &entry->offset);
5970                 entries++;
5971                 addr += sizeof(struct dir_entry) + name_len;
5972                 total_len += sizeof(struct dir_entry) + name_len;
5973 next:
5974                 path->slots[0]++;
5975         }
5976         btrfs_release_path(path);
5977
5978         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5979         if (ret)
5980                 goto nopos;
5981
5982         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5983         if (ret)
5984                 goto nopos;
5985
5986         /*
5987          * Stop new entries from being returned after we return the last
5988          * entry.
5989          *
5990          * New directory entries are assigned a strictly increasing
5991          * offset.  This means that new entries created during readdir
5992          * are *guaranteed* to be seen in the future by that readdir.
5993          * This has broken buggy programs which operate on names as
5994          * they're returned by readdir.  Until we re-use freed offsets
5995          * we have this hack to stop new entries from being returned
5996          * under the assumption that they'll never reach this huge
5997          * offset.
5998          *
5999          * This is being careful not to overflow 32bit loff_t unless the
6000          * last entry requires it because doing so has broken 32bit apps
6001          * in the past.
6002          */
6003         if (ctx->pos >= INT_MAX)
6004                 ctx->pos = LLONG_MAX;
6005         else
6006                 ctx->pos = INT_MAX;
6007 nopos:
6008         ret = 0;
6009 err:
6010         if (put)
6011                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6012         btrfs_free_path(path);
6013         return ret;
6014 }
6015
6016 /*
6017  * This is somewhat expensive, updating the tree every time the
6018  * inode changes.  But, it is most likely to find the inode in cache.
6019  * FIXME, needs more benchmarking...there are no reasons other than performance
6020  * to keep or drop this code.
6021  */
6022 static int btrfs_dirty_inode(struct inode *inode)
6023 {
6024         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6025         struct btrfs_root *root = BTRFS_I(inode)->root;
6026         struct btrfs_trans_handle *trans;
6027         int ret;
6028
6029         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6030                 return 0;
6031
6032         trans = btrfs_join_transaction(root);
6033         if (IS_ERR(trans))
6034                 return PTR_ERR(trans);
6035
6036         ret = btrfs_update_inode(trans, root, inode);
6037         if (ret && ret == -ENOSPC) {
6038                 /* whoops, lets try again with the full transaction */
6039                 btrfs_end_transaction(trans);
6040                 trans = btrfs_start_transaction(root, 1);
6041                 if (IS_ERR(trans))
6042                         return PTR_ERR(trans);
6043
6044                 ret = btrfs_update_inode(trans, root, inode);
6045         }
6046         btrfs_end_transaction(trans);
6047         if (BTRFS_I(inode)->delayed_node)
6048                 btrfs_balance_delayed_items(fs_info);
6049
6050         return ret;
6051 }
6052
6053 /*
6054  * This is a copy of file_update_time.  We need this so we can return error on
6055  * ENOSPC for updating the inode in the case of file write and mmap writes.
6056  */
6057 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
6058                              int flags)
6059 {
6060         struct btrfs_root *root = BTRFS_I(inode)->root;
6061         bool dirty = flags & ~S_VERSION;
6062
6063         if (btrfs_root_readonly(root))
6064                 return -EROFS;
6065
6066         if (flags & S_VERSION)
6067                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6068         if (flags & S_CTIME)
6069                 inode->i_ctime = *now;
6070         if (flags & S_MTIME)
6071                 inode->i_mtime = *now;
6072         if (flags & S_ATIME)
6073                 inode->i_atime = *now;
6074         return dirty ? btrfs_dirty_inode(inode) : 0;
6075 }
6076
6077 /*
6078  * find the highest existing sequence number in a directory
6079  * and then set the in-memory index_cnt variable to reflect
6080  * free sequence numbers
6081  */
6082 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6083 {
6084         struct btrfs_root *root = inode->root;
6085         struct btrfs_key key, found_key;
6086         struct btrfs_path *path;
6087         struct extent_buffer *leaf;
6088         int ret;
6089
6090         key.objectid = btrfs_ino(inode);
6091         key.type = BTRFS_DIR_INDEX_KEY;
6092         key.offset = (u64)-1;
6093
6094         path = btrfs_alloc_path();
6095         if (!path)
6096                 return -ENOMEM;
6097
6098         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6099         if (ret < 0)
6100                 goto out;
6101         /* FIXME: we should be able to handle this */
6102         if (ret == 0)
6103                 goto out;
6104         ret = 0;
6105
6106         /*
6107          * MAGIC NUMBER EXPLANATION:
6108          * since we search a directory based on f_pos we have to start at 2
6109          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6110          * else has to start at 2
6111          */
6112         if (path->slots[0] == 0) {
6113                 inode->index_cnt = 2;
6114                 goto out;
6115         }
6116
6117         path->slots[0]--;
6118
6119         leaf = path->nodes[0];
6120         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6121
6122         if (found_key.objectid != btrfs_ino(inode) ||
6123             found_key.type != BTRFS_DIR_INDEX_KEY) {
6124                 inode->index_cnt = 2;
6125                 goto out;
6126         }
6127
6128         inode->index_cnt = found_key.offset + 1;
6129 out:
6130         btrfs_free_path(path);
6131         return ret;
6132 }
6133
6134 /*
6135  * helper to find a free sequence number in a given directory.  This current
6136  * code is very simple, later versions will do smarter things in the btree
6137  */
6138 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6139 {
6140         int ret = 0;
6141
6142         if (dir->index_cnt == (u64)-1) {
6143                 ret = btrfs_inode_delayed_dir_index_count(dir);
6144                 if (ret) {
6145                         ret = btrfs_set_inode_index_count(dir);
6146                         if (ret)
6147                                 return ret;
6148                 }
6149         }
6150
6151         *index = dir->index_cnt;
6152         dir->index_cnt++;
6153
6154         return ret;
6155 }
6156
6157 static int btrfs_insert_inode_locked(struct inode *inode)
6158 {
6159         struct btrfs_iget_args args;
6160         args.location = &BTRFS_I(inode)->location;
6161         args.root = BTRFS_I(inode)->root;
6162
6163         return insert_inode_locked4(inode,
6164                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6165                    btrfs_find_actor, &args);
6166 }
6167
6168 /*
6169  * Inherit flags from the parent inode.
6170  *
6171  * Currently only the compression flags and the cow flags are inherited.
6172  */
6173 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6174 {
6175         unsigned int flags;
6176
6177         if (!dir)
6178                 return;
6179
6180         flags = BTRFS_I(dir)->flags;
6181
6182         if (flags & BTRFS_INODE_NOCOMPRESS) {
6183                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6184                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6185         } else if (flags & BTRFS_INODE_COMPRESS) {
6186                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6187                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6188         }
6189
6190         if (flags & BTRFS_INODE_NODATACOW) {
6191                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6192                 if (S_ISREG(inode->i_mode))
6193                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6194         }
6195
6196         btrfs_sync_inode_flags_to_i_flags(inode);
6197 }
6198
6199 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6200                                      struct btrfs_root *root,
6201                                      struct inode *dir,
6202                                      const char *name, int name_len,
6203                                      u64 ref_objectid, u64 objectid,
6204                                      umode_t mode, u64 *index)
6205 {
6206         struct btrfs_fs_info *fs_info = root->fs_info;
6207         struct inode *inode;
6208         struct btrfs_inode_item *inode_item;
6209         struct btrfs_key *location;
6210         struct btrfs_path *path;
6211         struct btrfs_inode_ref *ref;
6212         struct btrfs_key key[2];
6213         u32 sizes[2];
6214         int nitems = name ? 2 : 1;
6215         unsigned long ptr;
6216         int ret;
6217
6218         path = btrfs_alloc_path();
6219         if (!path)
6220                 return ERR_PTR(-ENOMEM);
6221
6222         inode = new_inode(fs_info->sb);
6223         if (!inode) {
6224                 btrfs_free_path(path);
6225                 return ERR_PTR(-ENOMEM);
6226         }
6227
6228         /*
6229          * O_TMPFILE, set link count to 0, so that after this point,
6230          * we fill in an inode item with the correct link count.
6231          */
6232         if (!name)
6233                 set_nlink(inode, 0);
6234
6235         /*
6236          * we have to initialize this early, so we can reclaim the inode
6237          * number if we fail afterwards in this function.
6238          */
6239         inode->i_ino = objectid;
6240
6241         if (dir && name) {
6242                 trace_btrfs_inode_request(dir);
6243
6244                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6245                 if (ret) {
6246                         btrfs_free_path(path);
6247                         iput(inode);
6248                         return ERR_PTR(ret);
6249                 }
6250         } else if (dir) {
6251                 *index = 0;
6252         }
6253         /*
6254          * index_cnt is ignored for everything but a dir,
6255          * btrfs_set_inode_index_count has an explanation for the magic
6256          * number
6257          */
6258         BTRFS_I(inode)->index_cnt = 2;
6259         BTRFS_I(inode)->dir_index = *index;
6260         BTRFS_I(inode)->root = root;
6261         BTRFS_I(inode)->generation = trans->transid;
6262         inode->i_generation = BTRFS_I(inode)->generation;
6263
6264         /*
6265          * We could have gotten an inode number from somebody who was fsynced
6266          * and then removed in this same transaction, so let's just set full
6267          * sync since it will be a full sync anyway and this will blow away the
6268          * old info in the log.
6269          */
6270         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6271
6272         key[0].objectid = objectid;
6273         key[0].type = BTRFS_INODE_ITEM_KEY;
6274         key[0].offset = 0;
6275
6276         sizes[0] = sizeof(struct btrfs_inode_item);
6277
6278         if (name) {
6279                 /*
6280                  * Start new inodes with an inode_ref. This is slightly more
6281                  * efficient for small numbers of hard links since they will
6282                  * be packed into one item. Extended refs will kick in if we
6283                  * add more hard links than can fit in the ref item.
6284                  */
6285                 key[1].objectid = objectid;
6286                 key[1].type = BTRFS_INODE_REF_KEY;
6287                 key[1].offset = ref_objectid;
6288
6289                 sizes[1] = name_len + sizeof(*ref);
6290         }
6291
6292         location = &BTRFS_I(inode)->location;
6293         location->objectid = objectid;
6294         location->offset = 0;
6295         location->type = BTRFS_INODE_ITEM_KEY;
6296
6297         ret = btrfs_insert_inode_locked(inode);
6298         if (ret < 0) {
6299                 iput(inode);
6300                 goto fail;
6301         }
6302
6303         path->leave_spinning = 1;
6304         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6305         if (ret != 0)
6306                 goto fail_unlock;
6307
6308         inode_init_owner(inode, dir, mode);
6309         inode_set_bytes(inode, 0);
6310
6311         inode->i_mtime = current_time(inode);
6312         inode->i_atime = inode->i_mtime;
6313         inode->i_ctime = inode->i_mtime;
6314         BTRFS_I(inode)->i_otime = inode->i_mtime;
6315
6316         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6317                                   struct btrfs_inode_item);
6318         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6319                              sizeof(*inode_item));
6320         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6321
6322         if (name) {
6323                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6324                                      struct btrfs_inode_ref);
6325                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6326                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6327                 ptr = (unsigned long)(ref + 1);
6328                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6329         }
6330
6331         btrfs_mark_buffer_dirty(path->nodes[0]);
6332         btrfs_free_path(path);
6333
6334         btrfs_inherit_iflags(inode, dir);
6335
6336         if (S_ISREG(mode)) {
6337                 if (btrfs_test_opt(fs_info, NODATASUM))
6338                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6339                 if (btrfs_test_opt(fs_info, NODATACOW))
6340                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6341                                 BTRFS_INODE_NODATASUM;
6342         }
6343
6344         inode_tree_add(inode);
6345
6346         trace_btrfs_inode_new(inode);
6347         btrfs_set_inode_last_trans(trans, inode);
6348
6349         btrfs_update_root_times(trans, root);
6350
6351         ret = btrfs_inode_inherit_props(trans, inode, dir);
6352         if (ret)
6353                 btrfs_err(fs_info,
6354                           "error inheriting props for ino %llu (root %llu): %d",
6355                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6356
6357         return inode;
6358
6359 fail_unlock:
6360         discard_new_inode(inode);
6361 fail:
6362         if (dir && name)
6363                 BTRFS_I(dir)->index_cnt--;
6364         btrfs_free_path(path);
6365         return ERR_PTR(ret);
6366 }
6367
6368 static inline u8 btrfs_inode_type(struct inode *inode)
6369 {
6370         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6371 }
6372
6373 /*
6374  * utility function to add 'inode' into 'parent_inode' with
6375  * a give name and a given sequence number.
6376  * if 'add_backref' is true, also insert a backref from the
6377  * inode to the parent directory.
6378  */
6379 int btrfs_add_link(struct btrfs_trans_handle *trans,
6380                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6381                    const char *name, int name_len, int add_backref, u64 index)
6382 {
6383         int ret = 0;
6384         struct btrfs_key key;
6385         struct btrfs_root *root = parent_inode->root;
6386         u64 ino = btrfs_ino(inode);
6387         u64 parent_ino = btrfs_ino(parent_inode);
6388
6389         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6390                 memcpy(&key, &inode->root->root_key, sizeof(key));
6391         } else {
6392                 key.objectid = ino;
6393                 key.type = BTRFS_INODE_ITEM_KEY;
6394                 key.offset = 0;
6395         }
6396
6397         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6398                 ret = btrfs_add_root_ref(trans, key.objectid,
6399                                          root->root_key.objectid, parent_ino,
6400                                          index, name, name_len);
6401         } else if (add_backref) {
6402                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6403                                              parent_ino, index);
6404         }
6405
6406         /* Nothing to clean up yet */
6407         if (ret)
6408                 return ret;
6409
6410         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6411                                     parent_inode, &key,
6412                                     btrfs_inode_type(&inode->vfs_inode), index);
6413         if (ret == -EEXIST || ret == -EOVERFLOW)
6414                 goto fail_dir_item;
6415         else if (ret) {
6416                 btrfs_abort_transaction(trans, ret);
6417                 return ret;
6418         }
6419
6420         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6421                            name_len * 2);
6422         inode_inc_iversion(&parent_inode->vfs_inode);
6423         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6424                 current_time(&parent_inode->vfs_inode);
6425         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6426         if (ret)
6427                 btrfs_abort_transaction(trans, ret);
6428         return ret;
6429
6430 fail_dir_item:
6431         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6432                 u64 local_index;
6433                 int err;
6434                 err = btrfs_del_root_ref(trans, key.objectid,
6435                                          root->root_key.objectid, parent_ino,
6436                                          &local_index, name, name_len);
6437
6438         } else if (add_backref) {
6439                 u64 local_index;
6440                 int err;
6441
6442                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6443                                           ino, parent_ino, &local_index);
6444         }
6445         return ret;
6446 }
6447
6448 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6449                             struct btrfs_inode *dir, struct dentry *dentry,
6450                             struct btrfs_inode *inode, int backref, u64 index)
6451 {
6452         int err = btrfs_add_link(trans, dir, inode,
6453                                  dentry->d_name.name, dentry->d_name.len,
6454                                  backref, index);
6455         if (err > 0)
6456                 err = -EEXIST;
6457         return err;
6458 }
6459
6460 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6461                         umode_t mode, dev_t rdev)
6462 {
6463         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6464         struct btrfs_trans_handle *trans;
6465         struct btrfs_root *root = BTRFS_I(dir)->root;
6466         struct inode *inode = NULL;
6467         int err;
6468         u64 objectid;
6469         u64 index = 0;
6470
6471         /*
6472          * 2 for inode item and ref
6473          * 2 for dir items
6474          * 1 for xattr if selinux is on
6475          */
6476         trans = btrfs_start_transaction(root, 5);
6477         if (IS_ERR(trans))
6478                 return PTR_ERR(trans);
6479
6480         err = btrfs_find_free_ino(root, &objectid);
6481         if (err)
6482                 goto out_unlock;
6483
6484         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6485                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6486                         mode, &index);
6487         if (IS_ERR(inode)) {
6488                 err = PTR_ERR(inode);
6489                 inode = NULL;
6490                 goto out_unlock;
6491         }
6492
6493         /*
6494         * If the active LSM wants to access the inode during
6495         * d_instantiate it needs these. Smack checks to see
6496         * if the filesystem supports xattrs by looking at the
6497         * ops vector.
6498         */
6499         inode->i_op = &btrfs_special_inode_operations;
6500         init_special_inode(inode, inode->i_mode, rdev);
6501
6502         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6503         if (err)
6504                 goto out_unlock;
6505
6506         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6507                         0, index);
6508         if (err)
6509                 goto out_unlock;
6510
6511         btrfs_update_inode(trans, root, inode);
6512         d_instantiate_new(dentry, inode);
6513
6514 out_unlock:
6515         btrfs_end_transaction(trans);
6516         btrfs_btree_balance_dirty(fs_info);
6517         if (err && inode) {
6518                 inode_dec_link_count(inode);
6519                 discard_new_inode(inode);
6520         }
6521         return err;
6522 }
6523
6524 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6525                         umode_t mode, bool excl)
6526 {
6527         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6528         struct btrfs_trans_handle *trans;
6529         struct btrfs_root *root = BTRFS_I(dir)->root;
6530         struct inode *inode = NULL;
6531         int err;
6532         u64 objectid;
6533         u64 index = 0;
6534
6535         /*
6536          * 2 for inode item and ref
6537          * 2 for dir items
6538          * 1 for xattr if selinux is on
6539          */
6540         trans = btrfs_start_transaction(root, 5);
6541         if (IS_ERR(trans))
6542                 return PTR_ERR(trans);
6543
6544         err = btrfs_find_free_ino(root, &objectid);
6545         if (err)
6546                 goto out_unlock;
6547
6548         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6549                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6550                         mode, &index);
6551         if (IS_ERR(inode)) {
6552                 err = PTR_ERR(inode);
6553                 inode = NULL;
6554                 goto out_unlock;
6555         }
6556         /*
6557         * If the active LSM wants to access the inode during
6558         * d_instantiate it needs these. Smack checks to see
6559         * if the filesystem supports xattrs by looking at the
6560         * ops vector.
6561         */
6562         inode->i_fop = &btrfs_file_operations;
6563         inode->i_op = &btrfs_file_inode_operations;
6564         inode->i_mapping->a_ops = &btrfs_aops;
6565
6566         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6567         if (err)
6568                 goto out_unlock;
6569
6570         err = btrfs_update_inode(trans, root, inode);
6571         if (err)
6572                 goto out_unlock;
6573
6574         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6575                         0, index);
6576         if (err)
6577                 goto out_unlock;
6578
6579         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6580         d_instantiate_new(dentry, inode);
6581
6582 out_unlock:
6583         btrfs_end_transaction(trans);
6584         if (err && inode) {
6585                 inode_dec_link_count(inode);
6586                 discard_new_inode(inode);
6587         }
6588         btrfs_btree_balance_dirty(fs_info);
6589         return err;
6590 }
6591
6592 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6593                       struct dentry *dentry)
6594 {
6595         struct btrfs_trans_handle *trans = NULL;
6596         struct btrfs_root *root = BTRFS_I(dir)->root;
6597         struct inode *inode = d_inode(old_dentry);
6598         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6599         u64 index;
6600         int err;
6601         int drop_inode = 0;
6602
6603         /* do not allow sys_link's with other subvols of the same device */
6604         if (root->objectid != BTRFS_I(inode)->root->objectid)
6605                 return -EXDEV;
6606
6607         if (inode->i_nlink >= BTRFS_LINK_MAX)
6608                 return -EMLINK;
6609
6610         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6611         if (err)
6612                 goto fail;
6613
6614         /*
6615          * 2 items for inode and inode ref
6616          * 2 items for dir items
6617          * 1 item for parent inode
6618          * 1 item for orphan item deletion if O_TMPFILE
6619          */
6620         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6621         if (IS_ERR(trans)) {
6622                 err = PTR_ERR(trans);
6623                 trans = NULL;
6624                 goto fail;
6625         }
6626
6627         /* There are several dir indexes for this inode, clear the cache. */
6628         BTRFS_I(inode)->dir_index = 0ULL;
6629         inc_nlink(inode);
6630         inode_inc_iversion(inode);
6631         inode->i_ctime = current_time(inode);
6632         ihold(inode);
6633         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6634
6635         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6636                         1, index);
6637
6638         if (err) {
6639                 drop_inode = 1;
6640         } else {
6641                 struct dentry *parent = dentry->d_parent;
6642                 err = btrfs_update_inode(trans, root, inode);
6643                 if (err)
6644                         goto fail;
6645                 if (inode->i_nlink == 1) {
6646                         /*
6647                          * If new hard link count is 1, it's a file created
6648                          * with open(2) O_TMPFILE flag.
6649                          */
6650                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6651                         if (err)
6652                                 goto fail;
6653                 }
6654                 d_instantiate(dentry, inode);
6655                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6656         }
6657
6658 fail:
6659         if (trans)
6660                 btrfs_end_transaction(trans);
6661         if (drop_inode) {
6662                 inode_dec_link_count(inode);
6663                 iput(inode);
6664         }
6665         btrfs_btree_balance_dirty(fs_info);
6666         return err;
6667 }
6668
6669 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6670 {
6671         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6672         struct inode *inode = NULL;
6673         struct btrfs_trans_handle *trans;
6674         struct btrfs_root *root = BTRFS_I(dir)->root;
6675         int err = 0;
6676         int drop_on_err = 0;
6677         u64 objectid = 0;
6678         u64 index = 0;
6679
6680         /*
6681          * 2 items for inode and ref
6682          * 2 items for dir items
6683          * 1 for xattr if selinux is on
6684          */
6685         trans = btrfs_start_transaction(root, 5);
6686         if (IS_ERR(trans))
6687                 return PTR_ERR(trans);
6688
6689         err = btrfs_find_free_ino(root, &objectid);
6690         if (err)
6691                 goto out_fail;
6692
6693         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6694                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6695                         S_IFDIR | mode, &index);
6696         if (IS_ERR(inode)) {
6697                 err = PTR_ERR(inode);
6698                 inode = NULL;
6699                 goto out_fail;
6700         }
6701
6702         drop_on_err = 1;
6703         /* these must be set before we unlock the inode */
6704         inode->i_op = &btrfs_dir_inode_operations;
6705         inode->i_fop = &btrfs_dir_file_operations;
6706
6707         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6708         if (err)
6709                 goto out_fail;
6710
6711         btrfs_i_size_write(BTRFS_I(inode), 0);
6712         err = btrfs_update_inode(trans, root, inode);
6713         if (err)
6714                 goto out_fail;
6715
6716         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6717                         dentry->d_name.name,
6718                         dentry->d_name.len, 0, index);
6719         if (err)
6720                 goto out_fail;
6721
6722         d_instantiate_new(dentry, inode);
6723         drop_on_err = 0;
6724
6725 out_fail:
6726         btrfs_end_transaction(trans);
6727         if (err && inode) {
6728                 inode_dec_link_count(inode);
6729                 discard_new_inode(inode);
6730         }
6731         btrfs_btree_balance_dirty(fs_info);
6732         return err;
6733 }
6734
6735 static noinline int uncompress_inline(struct btrfs_path *path,
6736                                       struct page *page,
6737                                       size_t pg_offset, u64 extent_offset,
6738                                       struct btrfs_file_extent_item *item)
6739 {
6740         int ret;
6741         struct extent_buffer *leaf = path->nodes[0];
6742         char *tmp;
6743         size_t max_size;
6744         unsigned long inline_size;
6745         unsigned long ptr;
6746         int compress_type;
6747
6748         WARN_ON(pg_offset != 0);
6749         compress_type = btrfs_file_extent_compression(leaf, item);
6750         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6751         inline_size = btrfs_file_extent_inline_item_len(leaf,
6752                                         btrfs_item_nr(path->slots[0]));
6753         tmp = kmalloc(inline_size, GFP_NOFS);
6754         if (!tmp)
6755                 return -ENOMEM;
6756         ptr = btrfs_file_extent_inline_start(item);
6757
6758         read_extent_buffer(leaf, tmp, ptr, inline_size);
6759
6760         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6761         ret = btrfs_decompress(compress_type, tmp, page,
6762                                extent_offset, inline_size, max_size);
6763
6764         /*
6765          * decompression code contains a memset to fill in any space between the end
6766          * of the uncompressed data and the end of max_size in case the decompressed
6767          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6768          * the end of an inline extent and the beginning of the next block, so we
6769          * cover that region here.
6770          */
6771
6772         if (max_size + pg_offset < PAGE_SIZE) {
6773                 char *map = kmap(page);
6774                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6775                 kunmap(page);
6776         }
6777         kfree(tmp);
6778         return ret;
6779 }
6780
6781 /*
6782  * a bit scary, this does extent mapping from logical file offset to the disk.
6783  * the ugly parts come from merging extents from the disk with the in-ram
6784  * representation.  This gets more complex because of the data=ordered code,
6785  * where the in-ram extents might be locked pending data=ordered completion.
6786  *
6787  * This also copies inline extents directly into the page.
6788  */
6789 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6790                 struct page *page,
6791             size_t pg_offset, u64 start, u64 len,
6792                 int create)
6793 {
6794         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6795         int ret;
6796         int err = 0;
6797         u64 extent_start = 0;
6798         u64 extent_end = 0;
6799         u64 objectid = btrfs_ino(inode);
6800         u32 found_type;
6801         struct btrfs_path *path = NULL;
6802         struct btrfs_root *root = inode->root;
6803         struct btrfs_file_extent_item *item;
6804         struct extent_buffer *leaf;
6805         struct btrfs_key found_key;
6806         struct extent_map *em = NULL;
6807         struct extent_map_tree *em_tree = &inode->extent_tree;
6808         struct extent_io_tree *io_tree = &inode->io_tree;
6809         const bool new_inline = !page || create;
6810
6811         read_lock(&em_tree->lock);
6812         em = lookup_extent_mapping(em_tree, start, len);
6813         if (em)
6814                 em->bdev = fs_info->fs_devices->latest_bdev;
6815         read_unlock(&em_tree->lock);
6816
6817         if (em) {
6818                 if (em->start > start || em->start + em->len <= start)
6819                         free_extent_map(em);
6820                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6821                         free_extent_map(em);
6822                 else
6823                         goto out;
6824         }
6825         em = alloc_extent_map();
6826         if (!em) {
6827                 err = -ENOMEM;
6828                 goto out;
6829         }
6830         em->bdev = fs_info->fs_devices->latest_bdev;
6831         em->start = EXTENT_MAP_HOLE;
6832         em->orig_start = EXTENT_MAP_HOLE;
6833         em->len = (u64)-1;
6834         em->block_len = (u64)-1;
6835
6836         if (!path) {
6837                 path = btrfs_alloc_path();
6838                 if (!path) {
6839                         err = -ENOMEM;
6840                         goto out;
6841                 }
6842                 /*
6843                  * Chances are we'll be called again, so go ahead and do
6844                  * readahead
6845                  */
6846                 path->reada = READA_FORWARD;
6847         }
6848
6849         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6850         if (ret < 0) {
6851                 err = ret;
6852                 goto out;
6853         }
6854
6855         if (ret != 0) {
6856                 if (path->slots[0] == 0)
6857                         goto not_found;
6858                 path->slots[0]--;
6859         }
6860
6861         leaf = path->nodes[0];
6862         item = btrfs_item_ptr(leaf, path->slots[0],
6863                               struct btrfs_file_extent_item);
6864         /* are we inside the extent that was found? */
6865         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6866         found_type = found_key.type;
6867         if (found_key.objectid != objectid ||
6868             found_type != BTRFS_EXTENT_DATA_KEY) {
6869                 /*
6870                  * If we backup past the first extent we want to move forward
6871                  * and see if there is an extent in front of us, otherwise we'll
6872                  * say there is a hole for our whole search range which can
6873                  * cause problems.
6874                  */
6875                 extent_end = start;
6876                 goto next;
6877         }
6878
6879         found_type = btrfs_file_extent_type(leaf, item);
6880         extent_start = found_key.offset;
6881         if (found_type == BTRFS_FILE_EXTENT_REG ||
6882             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6883                 extent_end = extent_start +
6884                        btrfs_file_extent_num_bytes(leaf, item);
6885
6886                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6887                                                        extent_start);
6888         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6889                 size_t size;
6890
6891                 size = btrfs_file_extent_ram_bytes(leaf, item);
6892                 extent_end = ALIGN(extent_start + size,
6893                                    fs_info->sectorsize);
6894
6895                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6896                                                       path->slots[0],
6897                                                       extent_start);
6898         }
6899 next:
6900         if (start >= extent_end) {
6901                 path->slots[0]++;
6902                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6903                         ret = btrfs_next_leaf(root, path);
6904                         if (ret < 0) {
6905                                 err = ret;
6906                                 goto out;
6907                         }
6908                         if (ret > 0)
6909                                 goto not_found;
6910                         leaf = path->nodes[0];
6911                 }
6912                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6913                 if (found_key.objectid != objectid ||
6914                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6915                         goto not_found;
6916                 if (start + len <= found_key.offset)
6917                         goto not_found;
6918                 if (start > found_key.offset)
6919                         goto next;
6920                 em->start = start;
6921                 em->orig_start = start;
6922                 em->len = found_key.offset - start;
6923                 goto not_found_em;
6924         }
6925
6926         btrfs_extent_item_to_extent_map(inode, path, item,
6927                         new_inline, em);
6928
6929         if (found_type == BTRFS_FILE_EXTENT_REG ||
6930             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6931                 goto insert;
6932         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6933                 unsigned long ptr;
6934                 char *map;
6935                 size_t size;
6936                 size_t extent_offset;
6937                 size_t copy_size;
6938
6939                 if (new_inline)
6940                         goto out;
6941
6942                 size = btrfs_file_extent_ram_bytes(leaf, item);
6943                 extent_offset = page_offset(page) + pg_offset - extent_start;
6944                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6945                                   size - extent_offset);
6946                 em->start = extent_start + extent_offset;
6947                 em->len = ALIGN(copy_size, fs_info->sectorsize);
6948                 em->orig_block_len = em->len;
6949                 em->orig_start = em->start;
6950                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6951                 if (!PageUptodate(page)) {
6952                         if (btrfs_file_extent_compression(leaf, item) !=
6953                             BTRFS_COMPRESS_NONE) {
6954                                 ret = uncompress_inline(path, page, pg_offset,
6955                                                         extent_offset, item);
6956                                 if (ret) {
6957                                         err = ret;
6958                                         goto out;
6959                                 }
6960                         } else {
6961                                 map = kmap(page);
6962                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6963                                                    copy_size);
6964                                 if (pg_offset + copy_size < PAGE_SIZE) {
6965                                         memset(map + pg_offset + copy_size, 0,
6966                                                PAGE_SIZE - pg_offset -
6967                                                copy_size);
6968                                 }
6969                                 kunmap(page);
6970                         }
6971                         flush_dcache_page(page);
6972                 }
6973                 set_extent_uptodate(io_tree, em->start,
6974                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6975                 goto insert;
6976         }
6977 not_found:
6978         em->start = start;
6979         em->orig_start = start;
6980         em->len = len;
6981 not_found_em:
6982         em->block_start = EXTENT_MAP_HOLE;
6983 insert:
6984         btrfs_release_path(path);
6985         if (em->start > start || extent_map_end(em) <= start) {
6986                 btrfs_err(fs_info,
6987                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
6988                           em->start, em->len, start, len);
6989                 err = -EIO;
6990                 goto out;
6991         }
6992
6993         err = 0;
6994         write_lock(&em_tree->lock);
6995         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
6996         write_unlock(&em_tree->lock);
6997 out:
6998
6999         trace_btrfs_get_extent(root, inode, em);
7000
7001         btrfs_free_path(path);
7002         if (err) {
7003                 free_extent_map(em);
7004                 return ERR_PTR(err);
7005         }
7006         BUG_ON(!em); /* Error is always set */
7007         return em;
7008 }
7009
7010 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7011                 struct page *page,
7012                 size_t pg_offset, u64 start, u64 len,
7013                 int create)
7014 {
7015         struct extent_map *em;
7016         struct extent_map *hole_em = NULL;
7017         u64 range_start = start;
7018         u64 end;
7019         u64 found;
7020         u64 found_end;
7021         int err = 0;
7022
7023         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7024         if (IS_ERR(em))
7025                 return em;
7026         /*
7027          * If our em maps to:
7028          * - a hole or
7029          * - a pre-alloc extent,
7030          * there might actually be delalloc bytes behind it.
7031          */
7032         if (em->block_start != EXTENT_MAP_HOLE &&
7033             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7034                 return em;
7035         else
7036                 hole_em = em;
7037
7038         /* check to see if we've wrapped (len == -1 or similar) */
7039         end = start + len;
7040         if (end < start)
7041                 end = (u64)-1;
7042         else
7043                 end -= 1;
7044
7045         em = NULL;
7046
7047         /* ok, we didn't find anything, lets look for delalloc */
7048         found = count_range_bits(&inode->io_tree, &range_start,
7049                                  end, len, EXTENT_DELALLOC, 1);
7050         found_end = range_start + found;
7051         if (found_end < range_start)
7052                 found_end = (u64)-1;
7053
7054         /*
7055          * we didn't find anything useful, return
7056          * the original results from get_extent()
7057          */
7058         if (range_start > end || found_end <= start) {
7059                 em = hole_em;
7060                 hole_em = NULL;
7061                 goto out;
7062         }
7063
7064         /* adjust the range_start to make sure it doesn't
7065          * go backwards from the start they passed in
7066          */
7067         range_start = max(start, range_start);
7068         found = found_end - range_start;
7069
7070         if (found > 0) {
7071                 u64 hole_start = start;
7072                 u64 hole_len = len;
7073
7074                 em = alloc_extent_map();
7075                 if (!em) {
7076                         err = -ENOMEM;
7077                         goto out;
7078                 }
7079                 /*
7080                  * when btrfs_get_extent can't find anything it
7081                  * returns one huge hole
7082                  *
7083                  * make sure what it found really fits our range, and
7084                  * adjust to make sure it is based on the start from
7085                  * the caller
7086                  */
7087                 if (hole_em) {
7088                         u64 calc_end = extent_map_end(hole_em);
7089
7090                         if (calc_end <= start || (hole_em->start > end)) {
7091                                 free_extent_map(hole_em);
7092                                 hole_em = NULL;
7093                         } else {
7094                                 hole_start = max(hole_em->start, start);
7095                                 hole_len = calc_end - hole_start;
7096                         }
7097                 }
7098                 em->bdev = NULL;
7099                 if (hole_em && range_start > hole_start) {
7100                         /* our hole starts before our delalloc, so we
7101                          * have to return just the parts of the hole
7102                          * that go until  the delalloc starts
7103                          */
7104                         em->len = min(hole_len,
7105                                       range_start - hole_start);
7106                         em->start = hole_start;
7107                         em->orig_start = hole_start;
7108                         /*
7109                          * don't adjust block start at all,
7110                          * it is fixed at EXTENT_MAP_HOLE
7111                          */
7112                         em->block_start = hole_em->block_start;
7113                         em->block_len = hole_len;
7114                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7115                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7116                 } else {
7117                         em->start = range_start;
7118                         em->len = found;
7119                         em->orig_start = range_start;
7120                         em->block_start = EXTENT_MAP_DELALLOC;
7121                         em->block_len = found;
7122                 }
7123         } else {
7124                 return hole_em;
7125         }
7126 out:
7127
7128         free_extent_map(hole_em);
7129         if (err) {
7130                 free_extent_map(em);
7131                 return ERR_PTR(err);
7132         }
7133         return em;
7134 }
7135
7136 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7137                                                   const u64 start,
7138                                                   const u64 len,
7139                                                   const u64 orig_start,
7140                                                   const u64 block_start,
7141                                                   const u64 block_len,
7142                                                   const u64 orig_block_len,
7143                                                   const u64 ram_bytes,
7144                                                   const int type)
7145 {
7146         struct extent_map *em = NULL;
7147         int ret;
7148
7149         if (type != BTRFS_ORDERED_NOCOW) {
7150                 em = create_io_em(inode, start, len, orig_start,
7151                                   block_start, block_len, orig_block_len,
7152                                   ram_bytes,
7153                                   BTRFS_COMPRESS_NONE, /* compress_type */
7154                                   type);
7155                 if (IS_ERR(em))
7156                         goto out;
7157         }
7158         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7159                                            len, block_len, type);
7160         if (ret) {
7161                 if (em) {
7162                         free_extent_map(em);
7163                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7164                                                 start + len - 1, 0);
7165                 }
7166                 em = ERR_PTR(ret);
7167         }
7168  out:
7169
7170         return em;
7171 }
7172
7173 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7174                                                   u64 start, u64 len)
7175 {
7176         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7177         struct btrfs_root *root = BTRFS_I(inode)->root;
7178         struct extent_map *em;
7179         struct btrfs_key ins;
7180         u64 alloc_hint;
7181         int ret;
7182
7183         alloc_hint = get_extent_allocation_hint(inode, start, len);
7184         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7185                                    0, alloc_hint, &ins, 1, 1);
7186         if (ret)
7187                 return ERR_PTR(ret);
7188
7189         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7190                                      ins.objectid, ins.offset, ins.offset,
7191                                      ins.offset, BTRFS_ORDERED_REGULAR);
7192         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7193         if (IS_ERR(em))
7194                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7195                                            ins.offset, 1);
7196
7197         return em;
7198 }
7199
7200 /*
7201  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7202  * block must be cow'd
7203  */
7204 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7205                               u64 *orig_start, u64 *orig_block_len,
7206                               u64 *ram_bytes)
7207 {
7208         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7209         struct btrfs_path *path;
7210         int ret;
7211         struct extent_buffer *leaf;
7212         struct btrfs_root *root = BTRFS_I(inode)->root;
7213         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7214         struct btrfs_file_extent_item *fi;
7215         struct btrfs_key key;
7216         u64 disk_bytenr;
7217         u64 backref_offset;
7218         u64 extent_end;
7219         u64 num_bytes;
7220         int slot;
7221         int found_type;
7222         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7223
7224         path = btrfs_alloc_path();
7225         if (!path)
7226                 return -ENOMEM;
7227
7228         ret = btrfs_lookup_file_extent(NULL, root, path,
7229                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7230         if (ret < 0)
7231                 goto out;
7232
7233         slot = path->slots[0];
7234         if (ret == 1) {
7235                 if (slot == 0) {
7236                         /* can't find the item, must cow */
7237                         ret = 0;
7238                         goto out;
7239                 }
7240                 slot--;
7241         }
7242         ret = 0;
7243         leaf = path->nodes[0];
7244         btrfs_item_key_to_cpu(leaf, &key, slot);
7245         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7246             key.type != BTRFS_EXTENT_DATA_KEY) {
7247                 /* not our file or wrong item type, must cow */
7248                 goto out;
7249         }
7250
7251         if (key.offset > offset) {
7252                 /* Wrong offset, must cow */
7253                 goto out;
7254         }
7255
7256         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7257         found_type = btrfs_file_extent_type(leaf, fi);
7258         if (found_type != BTRFS_FILE_EXTENT_REG &&
7259             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7260                 /* not a regular extent, must cow */
7261                 goto out;
7262         }
7263
7264         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7265                 goto out;
7266
7267         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7268         if (extent_end <= offset)
7269                 goto out;
7270
7271         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7272         if (disk_bytenr == 0)
7273                 goto out;
7274
7275         if (btrfs_file_extent_compression(leaf, fi) ||
7276             btrfs_file_extent_encryption(leaf, fi) ||
7277             btrfs_file_extent_other_encoding(leaf, fi))
7278                 goto out;
7279
7280         /*
7281          * Do the same check as in btrfs_cross_ref_exist but without the
7282          * unnecessary search.
7283          */
7284         if (btrfs_file_extent_generation(leaf, fi) <=
7285             btrfs_root_last_snapshot(&root->root_item))
7286                 goto out;
7287
7288         backref_offset = btrfs_file_extent_offset(leaf, fi);
7289
7290         if (orig_start) {
7291                 *orig_start = key.offset - backref_offset;
7292                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7293                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7294         }
7295
7296         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7297                 goto out;
7298
7299         num_bytes = min(offset + *len, extent_end) - offset;
7300         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7301                 u64 range_end;
7302
7303                 range_end = round_up(offset + num_bytes,
7304                                      root->fs_info->sectorsize) - 1;
7305                 ret = test_range_bit(io_tree, offset, range_end,
7306                                      EXTENT_DELALLOC, 0, NULL);
7307                 if (ret) {
7308                         ret = -EAGAIN;
7309                         goto out;
7310                 }
7311         }
7312
7313         btrfs_release_path(path);
7314
7315         /*
7316          * look for other files referencing this extent, if we
7317          * find any we must cow
7318          */
7319
7320         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7321                                     key.offset - backref_offset, disk_bytenr);
7322         if (ret) {
7323                 ret = 0;
7324                 goto out;
7325         }
7326
7327         /*
7328          * adjust disk_bytenr and num_bytes to cover just the bytes
7329          * in this extent we are about to write.  If there
7330          * are any csums in that range we have to cow in order
7331          * to keep the csums correct
7332          */
7333         disk_bytenr += backref_offset;
7334         disk_bytenr += offset - key.offset;
7335         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7336                 goto out;
7337         /*
7338          * all of the above have passed, it is safe to overwrite this extent
7339          * without cow
7340          */
7341         *len = num_bytes;
7342         ret = 1;
7343 out:
7344         btrfs_free_path(path);
7345         return ret;
7346 }
7347
7348 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7349                               struct extent_state **cached_state, int writing)
7350 {
7351         struct btrfs_ordered_extent *ordered;
7352         int ret = 0;
7353
7354         while (1) {
7355                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7356                                  cached_state);
7357                 /*
7358                  * We're concerned with the entire range that we're going to be
7359                  * doing DIO to, so we need to make sure there's no ordered
7360                  * extents in this range.
7361                  */
7362                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7363                                                      lockend - lockstart + 1);
7364
7365                 /*
7366                  * We need to make sure there are no buffered pages in this
7367                  * range either, we could have raced between the invalidate in
7368                  * generic_file_direct_write and locking the extent.  The
7369                  * invalidate needs to happen so that reads after a write do not
7370                  * get stale data.
7371                  */
7372                 if (!ordered &&
7373                     (!writing || !filemap_range_has_page(inode->i_mapping,
7374                                                          lockstart, lockend)))
7375                         break;
7376
7377                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7378                                      cached_state);
7379
7380                 if (ordered) {
7381                         /*
7382                          * If we are doing a DIO read and the ordered extent we
7383                          * found is for a buffered write, we can not wait for it
7384                          * to complete and retry, because if we do so we can
7385                          * deadlock with concurrent buffered writes on page
7386                          * locks. This happens only if our DIO read covers more
7387                          * than one extent map, if at this point has already
7388                          * created an ordered extent for a previous extent map
7389                          * and locked its range in the inode's io tree, and a
7390                          * concurrent write against that previous extent map's
7391                          * range and this range started (we unlock the ranges
7392                          * in the io tree only when the bios complete and
7393                          * buffered writes always lock pages before attempting
7394                          * to lock range in the io tree).
7395                          */
7396                         if (writing ||
7397                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7398                                 btrfs_start_ordered_extent(inode, ordered, 1);
7399                         else
7400                                 ret = -ENOTBLK;
7401                         btrfs_put_ordered_extent(ordered);
7402                 } else {
7403                         /*
7404                          * We could trigger writeback for this range (and wait
7405                          * for it to complete) and then invalidate the pages for
7406                          * this range (through invalidate_inode_pages2_range()),
7407                          * but that can lead us to a deadlock with a concurrent
7408                          * call to readpages() (a buffered read or a defrag call
7409                          * triggered a readahead) on a page lock due to an
7410                          * ordered dio extent we created before but did not have
7411                          * yet a corresponding bio submitted (whence it can not
7412                          * complete), which makes readpages() wait for that
7413                          * ordered extent to complete while holding a lock on
7414                          * that page.
7415                          */
7416                         ret = -ENOTBLK;
7417                 }
7418
7419                 if (ret)
7420                         break;
7421
7422                 cond_resched();
7423         }
7424
7425         return ret;
7426 }
7427
7428 /* The callers of this must take lock_extent() */
7429 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7430                                        u64 orig_start, u64 block_start,
7431                                        u64 block_len, u64 orig_block_len,
7432                                        u64 ram_bytes, int compress_type,
7433                                        int type)
7434 {
7435         struct extent_map_tree *em_tree;
7436         struct extent_map *em;
7437         struct btrfs_root *root = BTRFS_I(inode)->root;
7438         int ret;
7439
7440         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7441                type == BTRFS_ORDERED_COMPRESSED ||
7442                type == BTRFS_ORDERED_NOCOW ||
7443                type == BTRFS_ORDERED_REGULAR);
7444
7445         em_tree = &BTRFS_I(inode)->extent_tree;
7446         em = alloc_extent_map();
7447         if (!em)
7448                 return ERR_PTR(-ENOMEM);
7449
7450         em->start = start;
7451         em->orig_start = orig_start;
7452         em->len = len;
7453         em->block_len = block_len;
7454         em->block_start = block_start;
7455         em->bdev = root->fs_info->fs_devices->latest_bdev;
7456         em->orig_block_len = orig_block_len;
7457         em->ram_bytes = ram_bytes;
7458         em->generation = -1;
7459         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7460         if (type == BTRFS_ORDERED_PREALLOC) {
7461                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7462         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7463                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7464                 em->compress_type = compress_type;
7465         }
7466
7467         do {
7468                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7469                                 em->start + em->len - 1, 0);
7470                 write_lock(&em_tree->lock);
7471                 ret = add_extent_mapping(em_tree, em, 1);
7472                 write_unlock(&em_tree->lock);
7473                 /*
7474                  * The caller has taken lock_extent(), who could race with us
7475                  * to add em?
7476                  */
7477         } while (ret == -EEXIST);
7478
7479         if (ret) {
7480                 free_extent_map(em);
7481                 return ERR_PTR(ret);
7482         }
7483
7484         /* em got 2 refs now, callers needs to do free_extent_map once. */
7485         return em;
7486 }
7487
7488
7489 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7490                                         struct buffer_head *bh_result,
7491                                         struct inode *inode,
7492                                         u64 start, u64 len)
7493 {
7494         if (em->block_start == EXTENT_MAP_HOLE ||
7495                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7496                 return -ENOENT;
7497
7498         len = min(len, em->len - (start - em->start));
7499
7500         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7501                 inode->i_blkbits;
7502         bh_result->b_size = len;
7503         bh_result->b_bdev = em->bdev;
7504         set_buffer_mapped(bh_result);
7505
7506         return 0;
7507 }
7508
7509 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7510                                          struct buffer_head *bh_result,
7511                                          struct inode *inode,
7512                                          struct btrfs_dio_data *dio_data,
7513                                          u64 start, u64 len)
7514 {
7515         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7516         struct extent_map *em = *map;
7517         int ret = 0;
7518
7519         /*
7520          * We don't allocate a new extent in the following cases
7521          *
7522          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7523          * existing extent.
7524          * 2) The extent is marked as PREALLOC. We're good to go here and can
7525          * just use the extent.
7526          *
7527          */
7528         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7529             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7530              em->block_start != EXTENT_MAP_HOLE)) {
7531                 int type;
7532                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7533
7534                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7535                         type = BTRFS_ORDERED_PREALLOC;
7536                 else
7537                         type = BTRFS_ORDERED_NOCOW;
7538                 len = min(len, em->len - (start - em->start));
7539                 block_start = em->block_start + (start - em->start);
7540
7541                 if (can_nocow_extent(inode, start, &len, &orig_start,
7542                                      &orig_block_len, &ram_bytes) == 1 &&
7543                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7544                         struct extent_map *em2;
7545
7546                         em2 = btrfs_create_dio_extent(inode, start, len,
7547                                                       orig_start, block_start,
7548                                                       len, orig_block_len,
7549                                                       ram_bytes, type);
7550                         btrfs_dec_nocow_writers(fs_info, block_start);
7551                         if (type == BTRFS_ORDERED_PREALLOC) {
7552                                 free_extent_map(em);
7553                                 *map = em = em2;
7554                         }
7555
7556                         if (em2 && IS_ERR(em2)) {
7557                                 ret = PTR_ERR(em2);
7558                                 goto out;
7559                         }
7560                         /*
7561                          * For inode marked NODATACOW or extent marked PREALLOC,
7562                          * use the existing or preallocated extent, so does not
7563                          * need to adjust btrfs_space_info's bytes_may_use.
7564                          */
7565                         btrfs_free_reserved_data_space_noquota(inode, start,
7566                                                                len);
7567                         goto skip_cow;
7568                 }
7569         }
7570
7571         /* this will cow the extent */
7572         len = bh_result->b_size;
7573         free_extent_map(em);
7574         *map = em = btrfs_new_extent_direct(inode, start, len);
7575         if (IS_ERR(em)) {
7576                 ret = PTR_ERR(em);
7577                 goto out;
7578         }
7579
7580         len = min(len, em->len - (start - em->start));
7581
7582 skip_cow:
7583         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7584                 inode->i_blkbits;
7585         bh_result->b_size = len;
7586         bh_result->b_bdev = em->bdev;
7587         set_buffer_mapped(bh_result);
7588
7589         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7590                 set_buffer_new(bh_result);
7591
7592         /*
7593          * Need to update the i_size under the extent lock so buffered
7594          * readers will get the updated i_size when we unlock.
7595          */
7596         if (!dio_data->overwrite && start + len > i_size_read(inode))
7597                 i_size_write(inode, start + len);
7598
7599         WARN_ON(dio_data->reserve < len);
7600         dio_data->reserve -= len;
7601         dio_data->unsubmitted_oe_range_end = start + len;
7602         current->journal_info = dio_data;
7603 out:
7604         return ret;
7605 }
7606
7607 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7608                                    struct buffer_head *bh_result, int create)
7609 {
7610         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7611         struct extent_map *em;
7612         struct extent_state *cached_state = NULL;
7613         struct btrfs_dio_data *dio_data = NULL;
7614         u64 start = iblock << inode->i_blkbits;
7615         u64 lockstart, lockend;
7616         u64 len = bh_result->b_size;
7617         int unlock_bits = EXTENT_LOCKED;
7618         int ret = 0;
7619
7620         if (create)
7621                 unlock_bits |= EXTENT_DIRTY;
7622         else
7623                 len = min_t(u64, len, fs_info->sectorsize);
7624
7625         lockstart = start;
7626         lockend = start + len - 1;
7627
7628         if (current->journal_info) {
7629                 /*
7630                  * Need to pull our outstanding extents and set journal_info to NULL so
7631                  * that anything that needs to check if there's a transaction doesn't get
7632                  * confused.
7633                  */
7634                 dio_data = current->journal_info;
7635                 current->journal_info = NULL;
7636         }
7637
7638         /*
7639          * If this errors out it's because we couldn't invalidate pagecache for
7640          * this range and we need to fallback to buffered.
7641          */
7642         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7643                                create)) {
7644                 ret = -ENOTBLK;
7645                 goto err;
7646         }
7647
7648         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7649         if (IS_ERR(em)) {
7650                 ret = PTR_ERR(em);
7651                 goto unlock_err;
7652         }
7653
7654         /*
7655          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7656          * io.  INLINE is special, and we could probably kludge it in here, but
7657          * it's still buffered so for safety lets just fall back to the generic
7658          * buffered path.
7659          *
7660          * For COMPRESSED we _have_ to read the entire extent in so we can
7661          * decompress it, so there will be buffering required no matter what we
7662          * do, so go ahead and fallback to buffered.
7663          *
7664          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7665          * to buffered IO.  Don't blame me, this is the price we pay for using
7666          * the generic code.
7667          */
7668         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7669             em->block_start == EXTENT_MAP_INLINE) {
7670                 free_extent_map(em);
7671                 ret = -ENOTBLK;
7672                 goto unlock_err;
7673         }
7674
7675         if (create) {
7676                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7677                                                     dio_data, start, len);
7678                 if (ret < 0)
7679                         goto unlock_err;
7680
7681                 /* clear and unlock the entire range */
7682                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7683                                  unlock_bits, 1, 0, &cached_state);
7684         } else {
7685                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7686                                                    start, len);
7687                 /* Can be negative only if we read from a hole */
7688                 if (ret < 0) {
7689                         ret = 0;
7690                         free_extent_map(em);
7691                         goto unlock_err;
7692                 }
7693                 /*
7694                  * We need to unlock only the end area that we aren't using.
7695                  * The rest is going to be unlocked by the endio routine.
7696                  */
7697                 lockstart = start + bh_result->b_size;
7698                 if (lockstart < lockend) {
7699                         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7700                                          lockend, unlock_bits, 1, 0,
7701                                          &cached_state);
7702                 } else {
7703                         free_extent_state(cached_state);
7704                 }
7705         }
7706
7707         free_extent_map(em);
7708
7709         return 0;
7710
7711 unlock_err:
7712         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7713                          unlock_bits, 1, 0, &cached_state);
7714 err:
7715         if (dio_data)
7716                 current->journal_info = dio_data;
7717         return ret;
7718 }
7719
7720 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7721                                                  struct bio *bio,
7722                                                  int mirror_num)
7723 {
7724         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7725         blk_status_t ret;
7726
7727         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7728
7729         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7730         if (ret)
7731                 return ret;
7732
7733         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7734
7735         return ret;
7736 }
7737
7738 static int btrfs_check_dio_repairable(struct inode *inode,
7739                                       struct bio *failed_bio,
7740                                       struct io_failure_record *failrec,
7741                                       int failed_mirror)
7742 {
7743         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7744         int num_copies;
7745
7746         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7747         if (num_copies == 1) {
7748                 /*
7749                  * we only have a single copy of the data, so don't bother with
7750                  * all the retry and error correction code that follows. no
7751                  * matter what the error is, it is very likely to persist.
7752                  */
7753                 btrfs_debug(fs_info,
7754                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7755                         num_copies, failrec->this_mirror, failed_mirror);
7756                 return 0;
7757         }
7758
7759         failrec->failed_mirror = failed_mirror;
7760         failrec->this_mirror++;
7761         if (failrec->this_mirror == failed_mirror)
7762                 failrec->this_mirror++;
7763
7764         if (failrec->this_mirror > num_copies) {
7765                 btrfs_debug(fs_info,
7766                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7767                         num_copies, failrec->this_mirror, failed_mirror);
7768                 return 0;
7769         }
7770
7771         return 1;
7772 }
7773
7774 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7775                                    struct page *page, unsigned int pgoff,
7776                                    u64 start, u64 end, int failed_mirror,
7777                                    bio_end_io_t *repair_endio, void *repair_arg)
7778 {
7779         struct io_failure_record *failrec;
7780         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7781         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7782         struct bio *bio;
7783         int isector;
7784         unsigned int read_mode = 0;
7785         int segs;
7786         int ret;
7787         blk_status_t status;
7788         struct bio_vec bvec;
7789
7790         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7791
7792         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7793         if (ret)
7794                 return errno_to_blk_status(ret);
7795
7796         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7797                                          failed_mirror);
7798         if (!ret) {
7799                 free_io_failure(failure_tree, io_tree, failrec);
7800                 return BLK_STS_IOERR;
7801         }
7802
7803         segs = bio_segments(failed_bio);
7804         bio_get_first_bvec(failed_bio, &bvec);
7805         if (segs > 1 ||
7806             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7807                 read_mode |= REQ_FAILFAST_DEV;
7808
7809         isector = start - btrfs_io_bio(failed_bio)->logical;
7810         isector >>= inode->i_sb->s_blocksize_bits;
7811         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7812                                 pgoff, isector, repair_endio, repair_arg);
7813         bio->bi_opf = REQ_OP_READ | read_mode;
7814
7815         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7816                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7817                     read_mode, failrec->this_mirror, failrec->in_validation);
7818
7819         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7820         if (status) {
7821                 free_io_failure(failure_tree, io_tree, failrec);
7822                 bio_put(bio);
7823         }
7824
7825         return status;
7826 }
7827
7828 struct btrfs_retry_complete {
7829         struct completion done;
7830         struct inode *inode;
7831         u64 start;
7832         int uptodate;
7833 };
7834
7835 static void btrfs_retry_endio_nocsum(struct bio *bio)
7836 {
7837         struct btrfs_retry_complete *done = bio->bi_private;
7838         struct inode *inode = done->inode;
7839         struct bio_vec *bvec;
7840         struct extent_io_tree *io_tree, *failure_tree;
7841         int i;
7842
7843         if (bio->bi_status)
7844                 goto end;
7845
7846         ASSERT(bio->bi_vcnt == 1);
7847         io_tree = &BTRFS_I(inode)->io_tree;
7848         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7849         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
7850
7851         done->uptodate = 1;
7852         ASSERT(!bio_flagged(bio, BIO_CLONED));
7853         bio_for_each_segment_all(bvec, bio, i)
7854                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7855                                  io_tree, done->start, bvec->bv_page,
7856                                  btrfs_ino(BTRFS_I(inode)), 0);
7857 end:
7858         complete(&done->done);
7859         bio_put(bio);
7860 }
7861
7862 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7863                                                 struct btrfs_io_bio *io_bio)
7864 {
7865         struct btrfs_fs_info *fs_info;
7866         struct bio_vec bvec;
7867         struct bvec_iter iter;
7868         struct btrfs_retry_complete done;
7869         u64 start;
7870         unsigned int pgoff;
7871         u32 sectorsize;
7872         int nr_sectors;
7873         blk_status_t ret;
7874         blk_status_t err = BLK_STS_OK;
7875
7876         fs_info = BTRFS_I(inode)->root->fs_info;
7877         sectorsize = fs_info->sectorsize;
7878
7879         start = io_bio->logical;
7880         done.inode = inode;
7881         io_bio->bio.bi_iter = io_bio->iter;
7882
7883         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7884                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7885                 pgoff = bvec.bv_offset;
7886
7887 next_block_or_try_again:
7888                 done.uptodate = 0;
7889                 done.start = start;
7890                 init_completion(&done.done);
7891
7892                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7893                                 pgoff, start, start + sectorsize - 1,
7894                                 io_bio->mirror_num,
7895                                 btrfs_retry_endio_nocsum, &done);
7896                 if (ret) {
7897                         err = ret;
7898                         goto next;
7899                 }
7900
7901                 wait_for_completion_io(&done.done);
7902
7903                 if (!done.uptodate) {
7904                         /* We might have another mirror, so try again */
7905                         goto next_block_or_try_again;
7906                 }
7907
7908 next:
7909                 start += sectorsize;
7910
7911                 nr_sectors--;
7912                 if (nr_sectors) {
7913                         pgoff += sectorsize;
7914                         ASSERT(pgoff < PAGE_SIZE);
7915                         goto next_block_or_try_again;
7916                 }
7917         }
7918
7919         return err;
7920 }
7921
7922 static void btrfs_retry_endio(struct bio *bio)
7923 {
7924         struct btrfs_retry_complete *done = bio->bi_private;
7925         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7926         struct extent_io_tree *io_tree, *failure_tree;
7927         struct inode *inode = done->inode;
7928         struct bio_vec *bvec;
7929         int uptodate;
7930         int ret;
7931         int i;
7932
7933         if (bio->bi_status)
7934                 goto end;
7935
7936         uptodate = 1;
7937
7938         ASSERT(bio->bi_vcnt == 1);
7939         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
7940
7941         io_tree = &BTRFS_I(inode)->io_tree;
7942         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7943
7944         ASSERT(!bio_flagged(bio, BIO_CLONED));
7945         bio_for_each_segment_all(bvec, bio, i) {
7946                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7947                                              bvec->bv_offset, done->start,
7948                                              bvec->bv_len);
7949                 if (!ret)
7950                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
7951                                          failure_tree, io_tree, done->start,
7952                                          bvec->bv_page,
7953                                          btrfs_ino(BTRFS_I(inode)),
7954                                          bvec->bv_offset);
7955                 else
7956                         uptodate = 0;
7957         }
7958
7959         done->uptodate = uptodate;
7960 end:
7961         complete(&done->done);
7962         bio_put(bio);
7963 }
7964
7965 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
7966                 struct btrfs_io_bio *io_bio, blk_status_t err)
7967 {
7968         struct btrfs_fs_info *fs_info;
7969         struct bio_vec bvec;
7970         struct bvec_iter iter;
7971         struct btrfs_retry_complete done;
7972         u64 start;
7973         u64 offset = 0;
7974         u32 sectorsize;
7975         int nr_sectors;
7976         unsigned int pgoff;
7977         int csum_pos;
7978         bool uptodate = (err == 0);
7979         int ret;
7980         blk_status_t status;
7981
7982         fs_info = BTRFS_I(inode)->root->fs_info;
7983         sectorsize = fs_info->sectorsize;
7984
7985         err = BLK_STS_OK;
7986         start = io_bio->logical;
7987         done.inode = inode;
7988         io_bio->bio.bi_iter = io_bio->iter;
7989
7990         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7991                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7992
7993                 pgoff = bvec.bv_offset;
7994 next_block:
7995                 if (uptodate) {
7996                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
7997                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
7998                                         bvec.bv_page, pgoff, start, sectorsize);
7999                         if (likely(!ret))
8000                                 goto next;
8001                 }
8002 try_again:
8003                 done.uptodate = 0;
8004                 done.start = start;
8005                 init_completion(&done.done);
8006
8007                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8008                                         pgoff, start, start + sectorsize - 1,
8009                                         io_bio->mirror_num, btrfs_retry_endio,
8010                                         &done);
8011                 if (status) {
8012                         err = status;
8013                         goto next;
8014                 }
8015
8016                 wait_for_completion_io(&done.done);
8017
8018                 if (!done.uptodate) {
8019                         /* We might have another mirror, so try again */
8020                         goto try_again;
8021                 }
8022 next:
8023                 offset += sectorsize;
8024                 start += sectorsize;
8025
8026                 ASSERT(nr_sectors);
8027
8028                 nr_sectors--;
8029                 if (nr_sectors) {
8030                         pgoff += sectorsize;
8031                         ASSERT(pgoff < PAGE_SIZE);
8032                         goto next_block;
8033                 }
8034         }
8035
8036         return err;
8037 }
8038
8039 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8040                 struct btrfs_io_bio *io_bio, blk_status_t err)
8041 {
8042         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8043
8044         if (skip_csum) {
8045                 if (unlikely(err))
8046                         return __btrfs_correct_data_nocsum(inode, io_bio);
8047                 else
8048                         return BLK_STS_OK;
8049         } else {
8050                 return __btrfs_subio_endio_read(inode, io_bio, err);
8051         }
8052 }
8053
8054 static void btrfs_endio_direct_read(struct bio *bio)
8055 {
8056         struct btrfs_dio_private *dip = bio->bi_private;
8057         struct inode *inode = dip->inode;
8058         struct bio *dio_bio;
8059         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8060         blk_status_t err = bio->bi_status;
8061
8062         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8063                 err = btrfs_subio_endio_read(inode, io_bio, err);
8064
8065         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8066                       dip->logical_offset + dip->bytes - 1);
8067         dio_bio = dip->dio_bio;
8068
8069         kfree(dip);
8070
8071         dio_bio->bi_status = err;
8072         dio_end_io(dio_bio);
8073
8074         if (io_bio->end_io)
8075                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8076         bio_put(bio);
8077 }
8078
8079 static void __endio_write_update_ordered(struct inode *inode,
8080                                          const u64 offset, const u64 bytes,
8081                                          const bool uptodate)
8082 {
8083         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8084         struct btrfs_ordered_extent *ordered = NULL;
8085         struct btrfs_workqueue *wq;
8086         btrfs_work_func_t func;
8087         u64 ordered_offset = offset;
8088         u64 ordered_bytes = bytes;
8089         u64 last_offset;
8090
8091         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8092                 wq = fs_info->endio_freespace_worker;
8093                 func = btrfs_freespace_write_helper;
8094         } else {
8095                 wq = fs_info->endio_write_workers;
8096                 func = btrfs_endio_write_helper;
8097         }
8098
8099         while (ordered_offset < offset + bytes) {
8100                 last_offset = ordered_offset;
8101                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
8102                                                            &ordered_offset,
8103                                                            ordered_bytes,
8104                                                            uptodate)) {
8105                         btrfs_init_work(&ordered->work, func,
8106                                         finish_ordered_fn,
8107                                         NULL, NULL);
8108                         btrfs_queue_work(wq, &ordered->work);
8109                 }
8110                 /*
8111                  * If btrfs_dec_test_ordered_pending does not find any ordered
8112                  * extent in the range, we can exit.
8113                  */
8114                 if (ordered_offset == last_offset)
8115                         return;
8116                 /*
8117                  * Our bio might span multiple ordered extents. In this case
8118                  * we keep goin until we have accounted the whole dio.
8119                  */
8120                 if (ordered_offset < offset + bytes) {
8121                         ordered_bytes = offset + bytes - ordered_offset;
8122                         ordered = NULL;
8123                 }
8124         }
8125 }
8126
8127 static void btrfs_endio_direct_write(struct bio *bio)
8128 {
8129         struct btrfs_dio_private *dip = bio->bi_private;
8130         struct bio *dio_bio = dip->dio_bio;
8131
8132         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8133                                      dip->bytes, !bio->bi_status);
8134
8135         kfree(dip);
8136
8137         dio_bio->bi_status = bio->bi_status;
8138         dio_end_io(dio_bio);
8139         bio_put(bio);
8140 }
8141
8142 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8143                                     struct bio *bio, u64 offset)
8144 {
8145         struct inode *inode = private_data;
8146         blk_status_t ret;
8147         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8148         BUG_ON(ret); /* -ENOMEM */
8149         return 0;
8150 }
8151
8152 static void btrfs_end_dio_bio(struct bio *bio)
8153 {
8154         struct btrfs_dio_private *dip = bio->bi_private;
8155         blk_status_t err = bio->bi_status;
8156
8157         if (err)
8158                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8159                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8160                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8161                            bio->bi_opf,
8162                            (unsigned long long)bio->bi_iter.bi_sector,
8163                            bio->bi_iter.bi_size, err);
8164
8165         if (dip->subio_endio)
8166                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8167
8168         if (err) {
8169                 /*
8170                  * We want to perceive the errors flag being set before
8171                  * decrementing the reference count. We don't need a barrier
8172                  * since atomic operations with a return value are fully
8173                  * ordered as per atomic_t.txt
8174                  */
8175                 dip->errors = 1;
8176         }
8177
8178         /* if there are more bios still pending for this dio, just exit */
8179         if (!atomic_dec_and_test(&dip->pending_bios))
8180                 goto out;
8181
8182         if (dip->errors) {
8183                 bio_io_error(dip->orig_bio);
8184         } else {
8185                 dip->dio_bio->bi_status = BLK_STS_OK;
8186                 bio_endio(dip->orig_bio);
8187         }
8188 out:
8189         bio_put(bio);
8190 }
8191
8192 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8193                                                  struct btrfs_dio_private *dip,
8194                                                  struct bio *bio,
8195                                                  u64 file_offset)
8196 {
8197         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8198         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8199         blk_status_t ret;
8200
8201         /*
8202          * We load all the csum data we need when we submit
8203          * the first bio to reduce the csum tree search and
8204          * contention.
8205          */
8206         if (dip->logical_offset == file_offset) {
8207                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8208                                                 file_offset);
8209                 if (ret)
8210                         return ret;
8211         }
8212
8213         if (bio == dip->orig_bio)
8214                 return 0;
8215
8216         file_offset -= dip->logical_offset;
8217         file_offset >>= inode->i_sb->s_blocksize_bits;
8218         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8219
8220         return 0;
8221 }
8222
8223 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8224                 struct inode *inode, u64 file_offset, int async_submit)
8225 {
8226         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8227         struct btrfs_dio_private *dip = bio->bi_private;
8228         bool write = bio_op(bio) == REQ_OP_WRITE;
8229         blk_status_t ret;
8230
8231         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8232         if (async_submit)
8233                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8234
8235         if (!write) {
8236                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8237                 if (ret)
8238                         goto err;
8239         }
8240
8241         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8242                 goto map;
8243
8244         if (write && async_submit) {
8245                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8246                                           file_offset, inode,
8247                                           btrfs_submit_bio_start_direct_io);
8248                 goto err;
8249         } else if (write) {
8250                 /*
8251                  * If we aren't doing async submit, calculate the csum of the
8252                  * bio now.
8253                  */
8254                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8255                 if (ret)
8256                         goto err;
8257         } else {
8258                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8259                                                      file_offset);
8260                 if (ret)
8261                         goto err;
8262         }
8263 map:
8264         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8265 err:
8266         return ret;
8267 }
8268
8269 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8270 {
8271         struct inode *inode = dip->inode;
8272         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8273         struct bio *bio;
8274         struct bio *orig_bio = dip->orig_bio;
8275         u64 start_sector = orig_bio->bi_iter.bi_sector;
8276         u64 file_offset = dip->logical_offset;
8277         u64 map_length;
8278         int async_submit = 0;
8279         u64 submit_len;
8280         int clone_offset = 0;
8281         int clone_len;
8282         int ret;
8283         blk_status_t status;
8284
8285         map_length = orig_bio->bi_iter.bi_size;
8286         submit_len = map_length;
8287         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8288                               &map_length, NULL, 0);
8289         if (ret)
8290                 return -EIO;
8291
8292         if (map_length >= submit_len) {
8293                 bio = orig_bio;
8294                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8295                 goto submit;
8296         }
8297
8298         /* async crcs make it difficult to collect full stripe writes. */
8299         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8300                 async_submit = 0;
8301         else
8302                 async_submit = 1;
8303
8304         /* bio split */
8305         ASSERT(map_length <= INT_MAX);
8306         atomic_inc(&dip->pending_bios);
8307         do {
8308                 clone_len = min_t(int, submit_len, map_length);
8309
8310                 /*
8311                  * This will never fail as it's passing GPF_NOFS and
8312                  * the allocation is backed by btrfs_bioset.
8313                  */
8314                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8315                                               clone_len);
8316                 bio->bi_private = dip;
8317                 bio->bi_end_io = btrfs_end_dio_bio;
8318                 btrfs_io_bio(bio)->logical = file_offset;
8319
8320                 ASSERT(submit_len >= clone_len);
8321                 submit_len -= clone_len;
8322                 if (submit_len == 0)
8323                         break;
8324
8325                 /*
8326                  * Increase the count before we submit the bio so we know
8327                  * the end IO handler won't happen before we increase the
8328                  * count. Otherwise, the dip might get freed before we're
8329                  * done setting it up.
8330                  */
8331                 atomic_inc(&dip->pending_bios);
8332
8333                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8334                                                 async_submit);
8335                 if (status) {
8336                         bio_put(bio);
8337                         atomic_dec(&dip->pending_bios);
8338                         goto out_err;
8339                 }
8340
8341                 clone_offset += clone_len;
8342                 start_sector += clone_len >> 9;
8343                 file_offset += clone_len;
8344
8345                 map_length = submit_len;
8346                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8347                                       start_sector << 9, &map_length, NULL, 0);
8348                 if (ret)
8349                         goto out_err;
8350         } while (submit_len > 0);
8351
8352 submit:
8353         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8354         if (!status)
8355                 return 0;
8356
8357         bio_put(bio);
8358 out_err:
8359         dip->errors = 1;
8360         /*
8361          * Before atomic variable goto zero, we must  make sure dip->errors is
8362          * perceived to be set. This ordering is ensured by the fact that an
8363          * atomic operations with a return value are fully ordered as per
8364          * atomic_t.txt
8365          */
8366         if (atomic_dec_and_test(&dip->pending_bios))
8367                 bio_io_error(dip->orig_bio);
8368
8369         /* bio_end_io() will handle error, so we needn't return it */
8370         return 0;
8371 }
8372
8373 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8374                                 loff_t file_offset)
8375 {
8376         struct btrfs_dio_private *dip = NULL;
8377         struct bio *bio = NULL;
8378         struct btrfs_io_bio *io_bio;
8379         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8380         int ret = 0;
8381
8382         bio = btrfs_bio_clone(dio_bio);
8383
8384         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8385         if (!dip) {
8386                 ret = -ENOMEM;
8387                 goto free_ordered;
8388         }
8389
8390         dip->private = dio_bio->bi_private;
8391         dip->inode = inode;
8392         dip->logical_offset = file_offset;
8393         dip->bytes = dio_bio->bi_iter.bi_size;
8394         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8395         bio->bi_private = dip;
8396         dip->orig_bio = bio;
8397         dip->dio_bio = dio_bio;
8398         atomic_set(&dip->pending_bios, 0);
8399         io_bio = btrfs_io_bio(bio);
8400         io_bio->logical = file_offset;
8401
8402         if (write) {
8403                 bio->bi_end_io = btrfs_endio_direct_write;
8404         } else {
8405                 bio->bi_end_io = btrfs_endio_direct_read;
8406                 dip->subio_endio = btrfs_subio_endio_read;
8407         }
8408
8409         /*
8410          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8411          * even if we fail to submit a bio, because in such case we do the
8412          * corresponding error handling below and it must not be done a second
8413          * time by btrfs_direct_IO().
8414          */
8415         if (write) {
8416                 struct btrfs_dio_data *dio_data = current->journal_info;
8417
8418                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8419                         dip->bytes;
8420                 dio_data->unsubmitted_oe_range_start =
8421                         dio_data->unsubmitted_oe_range_end;
8422         }
8423
8424         ret = btrfs_submit_direct_hook(dip);
8425         if (!ret)
8426                 return;
8427
8428         if (io_bio->end_io)
8429                 io_bio->end_io(io_bio, ret);
8430
8431 free_ordered:
8432         /*
8433          * If we arrived here it means either we failed to submit the dip
8434          * or we either failed to clone the dio_bio or failed to allocate the
8435          * dip. If we cloned the dio_bio and allocated the dip, we can just
8436          * call bio_endio against our io_bio so that we get proper resource
8437          * cleanup if we fail to submit the dip, otherwise, we must do the
8438          * same as btrfs_endio_direct_[write|read] because we can't call these
8439          * callbacks - they require an allocated dip and a clone of dio_bio.
8440          */
8441         if (bio && dip) {
8442                 bio_io_error(bio);
8443                 /*
8444                  * The end io callbacks free our dip, do the final put on bio
8445                  * and all the cleanup and final put for dio_bio (through
8446                  * dio_end_io()).
8447                  */
8448                 dip = NULL;
8449                 bio = NULL;
8450         } else {
8451                 if (write)
8452                         __endio_write_update_ordered(inode,
8453                                                 file_offset,
8454                                                 dio_bio->bi_iter.bi_size,
8455                                                 false);
8456                 else
8457                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8458                               file_offset + dio_bio->bi_iter.bi_size - 1);
8459
8460                 dio_bio->bi_status = BLK_STS_IOERR;
8461                 /*
8462                  * Releases and cleans up our dio_bio, no need to bio_put()
8463                  * nor bio_endio()/bio_io_error() against dio_bio.
8464                  */
8465                 dio_end_io(dio_bio);
8466         }
8467         if (bio)
8468                 bio_put(bio);
8469         kfree(dip);
8470 }
8471
8472 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8473                                const struct iov_iter *iter, loff_t offset)
8474 {
8475         int seg;
8476         int i;
8477         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8478         ssize_t retval = -EINVAL;
8479
8480         if (offset & blocksize_mask)
8481                 goto out;
8482
8483         if (iov_iter_alignment(iter) & blocksize_mask)
8484                 goto out;
8485
8486         /* If this is a write we don't need to check anymore */
8487         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8488                 return 0;
8489         /*
8490          * Check to make sure we don't have duplicate iov_base's in this
8491          * iovec, if so return EINVAL, otherwise we'll get csum errors
8492          * when reading back.
8493          */
8494         for (seg = 0; seg < iter->nr_segs; seg++) {
8495                 for (i = seg + 1; i < iter->nr_segs; i++) {
8496                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8497                                 goto out;
8498                 }
8499         }
8500         retval = 0;
8501 out:
8502         return retval;
8503 }
8504
8505 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8506 {
8507         struct file *file = iocb->ki_filp;
8508         struct inode *inode = file->f_mapping->host;
8509         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8510         struct btrfs_dio_data dio_data = { 0 };
8511         struct extent_changeset *data_reserved = NULL;
8512         loff_t offset = iocb->ki_pos;
8513         size_t count = 0;
8514         int flags = 0;
8515         bool wakeup = true;
8516         bool relock = false;
8517         ssize_t ret;
8518
8519         if (check_direct_IO(fs_info, iter, offset))
8520                 return 0;
8521
8522         inode_dio_begin(inode);
8523
8524         /*
8525          * The generic stuff only does filemap_write_and_wait_range, which
8526          * isn't enough if we've written compressed pages to this area, so
8527          * we need to flush the dirty pages again to make absolutely sure
8528          * that any outstanding dirty pages are on disk.
8529          */
8530         count = iov_iter_count(iter);
8531         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8532                      &BTRFS_I(inode)->runtime_flags))
8533                 filemap_fdatawrite_range(inode->i_mapping, offset,
8534                                          offset + count - 1);
8535
8536         if (iov_iter_rw(iter) == WRITE) {
8537                 /*
8538                  * If the write DIO is beyond the EOF, we need update
8539                  * the isize, but it is protected by i_mutex. So we can
8540                  * not unlock the i_mutex at this case.
8541                  */
8542                 if (offset + count <= inode->i_size) {
8543                         dio_data.overwrite = 1;
8544                         inode_unlock(inode);
8545                         relock = true;
8546                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8547                         ret = -EAGAIN;
8548                         goto out;
8549                 }
8550                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8551                                                    offset, count);
8552                 if (ret)
8553                         goto out;
8554
8555                 /*
8556                  * We need to know how many extents we reserved so that we can
8557                  * do the accounting properly if we go over the number we
8558                  * originally calculated.  Abuse current->journal_info for this.
8559                  */
8560                 dio_data.reserve = round_up(count,
8561                                             fs_info->sectorsize);
8562                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8563                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8564                 current->journal_info = &dio_data;
8565                 down_read(&BTRFS_I(inode)->dio_sem);
8566         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8567                                      &BTRFS_I(inode)->runtime_flags)) {
8568                 inode_dio_end(inode);
8569                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8570                 wakeup = false;
8571         }
8572
8573         ret = __blockdev_direct_IO(iocb, inode,
8574                                    fs_info->fs_devices->latest_bdev,
8575                                    iter, btrfs_get_blocks_direct, NULL,
8576                                    btrfs_submit_direct, flags);
8577         if (iov_iter_rw(iter) == WRITE) {
8578                 up_read(&BTRFS_I(inode)->dio_sem);
8579                 current->journal_info = NULL;
8580                 if (ret < 0 && ret != -EIOCBQUEUED) {
8581                         if (dio_data.reserve)
8582                                 btrfs_delalloc_release_space(inode, data_reserved,
8583                                         offset, dio_data.reserve, true);
8584                         /*
8585                          * On error we might have left some ordered extents
8586                          * without submitting corresponding bios for them, so
8587                          * cleanup them up to avoid other tasks getting them
8588                          * and waiting for them to complete forever.
8589                          */
8590                         if (dio_data.unsubmitted_oe_range_start <
8591                             dio_data.unsubmitted_oe_range_end)
8592                                 __endio_write_update_ordered(inode,
8593                                         dio_data.unsubmitted_oe_range_start,
8594                                         dio_data.unsubmitted_oe_range_end -
8595                                         dio_data.unsubmitted_oe_range_start,
8596                                         false);
8597                 } else if (ret >= 0 && (size_t)ret < count)
8598                         btrfs_delalloc_release_space(inode, data_reserved,
8599                                         offset, count - (size_t)ret, true);
8600                 btrfs_delalloc_release_extents(BTRFS_I(inode), count, false);
8601         }
8602 out:
8603         if (wakeup)
8604                 inode_dio_end(inode);
8605         if (relock)
8606                 inode_lock(inode);
8607
8608         extent_changeset_free(data_reserved);
8609         return ret;
8610 }
8611
8612 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8613
8614 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8615                 __u64 start, __u64 len)
8616 {
8617         int     ret;
8618
8619         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8620         if (ret)
8621                 return ret;
8622
8623         return extent_fiemap(inode, fieinfo, start, len);
8624 }
8625
8626 int btrfs_readpage(struct file *file, struct page *page)
8627 {
8628         struct extent_io_tree *tree;
8629         tree = &BTRFS_I(page->mapping->host)->io_tree;
8630         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8631 }
8632
8633 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8634 {
8635         struct inode *inode = page->mapping->host;
8636         int ret;
8637
8638         if (current->flags & PF_MEMALLOC) {
8639                 redirty_page_for_writepage(wbc, page);
8640                 unlock_page(page);
8641                 return 0;
8642         }
8643
8644         /*
8645          * If we are under memory pressure we will call this directly from the
8646          * VM, we need to make sure we have the inode referenced for the ordered
8647          * extent.  If not just return like we didn't do anything.
8648          */
8649         if (!igrab(inode)) {
8650                 redirty_page_for_writepage(wbc, page);
8651                 return AOP_WRITEPAGE_ACTIVATE;
8652         }
8653         ret = extent_write_full_page(page, wbc);
8654         btrfs_add_delayed_iput(inode);
8655         return ret;
8656 }
8657
8658 static int btrfs_writepages(struct address_space *mapping,
8659                             struct writeback_control *wbc)
8660 {
8661         return extent_writepages(mapping, wbc);
8662 }
8663
8664 static int
8665 btrfs_readpages(struct file *file, struct address_space *mapping,
8666                 struct list_head *pages, unsigned nr_pages)
8667 {
8668         return extent_readpages(mapping, pages, nr_pages);
8669 }
8670
8671 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8672 {
8673         int ret = try_release_extent_mapping(page, gfp_flags);
8674         if (ret == 1) {
8675                 ClearPagePrivate(page);
8676                 set_page_private(page, 0);
8677                 put_page(page);
8678         }
8679         return ret;
8680 }
8681
8682 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8683 {
8684         if (PageWriteback(page) || PageDirty(page))
8685                 return 0;
8686         return __btrfs_releasepage(page, gfp_flags);
8687 }
8688
8689 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8690                                  unsigned int length)
8691 {
8692         struct inode *inode = page->mapping->host;
8693         struct extent_io_tree *tree;
8694         struct btrfs_ordered_extent *ordered;
8695         struct extent_state *cached_state = NULL;
8696         u64 page_start = page_offset(page);
8697         u64 page_end = page_start + PAGE_SIZE - 1;
8698         u64 start;
8699         u64 end;
8700         int inode_evicting = inode->i_state & I_FREEING;
8701
8702         /*
8703          * we have the page locked, so new writeback can't start,
8704          * and the dirty bit won't be cleared while we are here.
8705          *
8706          * Wait for IO on this page so that we can safely clear
8707          * the PagePrivate2 bit and do ordered accounting
8708          */
8709         wait_on_page_writeback(page);
8710
8711         tree = &BTRFS_I(inode)->io_tree;
8712         if (offset) {
8713                 btrfs_releasepage(page, GFP_NOFS);
8714                 return;
8715         }
8716
8717         if (!inode_evicting)
8718                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8719 again:
8720         start = page_start;
8721         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8722                                         page_end - start + 1);
8723         if (ordered) {
8724                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8725                 /*
8726                  * IO on this page will never be started, so we need
8727                  * to account for any ordered extents now
8728                  */
8729                 if (!inode_evicting)
8730                         clear_extent_bit(tree, start, end,
8731                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8732                                          EXTENT_DELALLOC_NEW |
8733                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8734                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8735                 /*
8736                  * whoever cleared the private bit is responsible
8737                  * for the finish_ordered_io
8738                  */
8739                 if (TestClearPagePrivate2(page)) {
8740                         struct btrfs_ordered_inode_tree *tree;
8741                         u64 new_len;
8742
8743                         tree = &BTRFS_I(inode)->ordered_tree;
8744
8745                         spin_lock_irq(&tree->lock);
8746                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8747                         new_len = start - ordered->file_offset;
8748                         if (new_len < ordered->truncated_len)
8749                                 ordered->truncated_len = new_len;
8750                         spin_unlock_irq(&tree->lock);
8751
8752                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8753                                                            start,
8754                                                            end - start + 1, 1))
8755                                 btrfs_finish_ordered_io(ordered);
8756                 }
8757                 btrfs_put_ordered_extent(ordered);
8758                 if (!inode_evicting) {
8759                         cached_state = NULL;
8760                         lock_extent_bits(tree, start, end,
8761                                          &cached_state);
8762                 }
8763
8764                 start = end + 1;
8765                 if (start < page_end)
8766                         goto again;
8767         }
8768
8769         /*
8770          * Qgroup reserved space handler
8771          * Page here will be either
8772          * 1) Already written to disk
8773          *    In this case, its reserved space is released from data rsv map
8774          *    and will be freed by delayed_ref handler finally.
8775          *    So even we call qgroup_free_data(), it won't decrease reserved
8776          *    space.
8777          * 2) Not written to disk
8778          *    This means the reserved space should be freed here. However,
8779          *    if a truncate invalidates the page (by clearing PageDirty)
8780          *    and the page is accounted for while allocating extent
8781          *    in btrfs_check_data_free_space() we let delayed_ref to
8782          *    free the entire extent.
8783          */
8784         if (PageDirty(page))
8785                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8786         if (!inode_evicting) {
8787                 clear_extent_bit(tree, page_start, page_end,
8788                                  EXTENT_LOCKED | EXTENT_DIRTY |
8789                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8790                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8791                                  &cached_state);
8792
8793                 __btrfs_releasepage(page, GFP_NOFS);
8794         }
8795
8796         ClearPageChecked(page);
8797         if (PagePrivate(page)) {
8798                 ClearPagePrivate(page);
8799                 set_page_private(page, 0);
8800                 put_page(page);
8801         }
8802 }
8803
8804 /*
8805  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8806  * called from a page fault handler when a page is first dirtied. Hence we must
8807  * be careful to check for EOF conditions here. We set the page up correctly
8808  * for a written page which means we get ENOSPC checking when writing into
8809  * holes and correct delalloc and unwritten extent mapping on filesystems that
8810  * support these features.
8811  *
8812  * We are not allowed to take the i_mutex here so we have to play games to
8813  * protect against truncate races as the page could now be beyond EOF.  Because
8814  * truncate_setsize() writes the inode size before removing pages, once we have
8815  * the page lock we can determine safely if the page is beyond EOF. If it is not
8816  * beyond EOF, then the page is guaranteed safe against truncation until we
8817  * unlock the page.
8818  */
8819 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8820 {
8821         struct page *page = vmf->page;
8822         struct inode *inode = file_inode(vmf->vma->vm_file);
8823         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8824         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8825         struct btrfs_ordered_extent *ordered;
8826         struct extent_state *cached_state = NULL;
8827         struct extent_changeset *data_reserved = NULL;
8828         char *kaddr;
8829         unsigned long zero_start;
8830         loff_t size;
8831         vm_fault_t ret;
8832         int ret2;
8833         int reserved = 0;
8834         u64 reserved_space;
8835         u64 page_start;
8836         u64 page_end;
8837         u64 end;
8838
8839         reserved_space = PAGE_SIZE;
8840
8841         sb_start_pagefault(inode->i_sb);
8842         page_start = page_offset(page);
8843         page_end = page_start + PAGE_SIZE - 1;
8844         end = page_end;
8845
8846         /*
8847          * Reserving delalloc space after obtaining the page lock can lead to
8848          * deadlock. For example, if a dirty page is locked by this function
8849          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8850          * dirty page write out, then the btrfs_writepage() function could
8851          * end up waiting indefinitely to get a lock on the page currently
8852          * being processed by btrfs_page_mkwrite() function.
8853          */
8854         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8855                                            reserved_space);
8856         if (!ret2) {
8857                 ret2 = file_update_time(vmf->vma->vm_file);
8858                 reserved = 1;
8859         }
8860         if (ret2) {
8861                 ret = vmf_error(ret2);
8862                 if (reserved)
8863                         goto out;
8864                 goto out_noreserve;
8865         }
8866
8867         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8868 again:
8869         lock_page(page);
8870         size = i_size_read(inode);
8871
8872         if ((page->mapping != inode->i_mapping) ||
8873             (page_start >= size)) {
8874                 /* page got truncated out from underneath us */
8875                 goto out_unlock;
8876         }
8877         wait_on_page_writeback(page);
8878
8879         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8880         set_page_extent_mapped(page);
8881
8882         /*
8883          * we can't set the delalloc bits if there are pending ordered
8884          * extents.  Drop our locks and wait for them to finish
8885          */
8886         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8887                         PAGE_SIZE);
8888         if (ordered) {
8889                 unlock_extent_cached(io_tree, page_start, page_end,
8890                                      &cached_state);
8891                 unlock_page(page);
8892                 btrfs_start_ordered_extent(inode, ordered, 1);
8893                 btrfs_put_ordered_extent(ordered);
8894                 goto again;
8895         }
8896
8897         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8898                 reserved_space = round_up(size - page_start,
8899                                           fs_info->sectorsize);
8900                 if (reserved_space < PAGE_SIZE) {
8901                         end = page_start + reserved_space - 1;
8902                         btrfs_delalloc_release_space(inode, data_reserved,
8903                                         page_start, PAGE_SIZE - reserved_space,
8904                                         true);
8905                 }
8906         }
8907
8908         /*
8909          * page_mkwrite gets called when the page is firstly dirtied after it's
8910          * faulted in, but write(2) could also dirty a page and set delalloc
8911          * bits, thus in this case for space account reason, we still need to
8912          * clear any delalloc bits within this page range since we have to
8913          * reserve data&meta space before lock_page() (see above comments).
8914          */
8915         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8916                           EXTENT_DIRTY | EXTENT_DELALLOC |
8917                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8918                           0, 0, &cached_state);
8919
8920         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
8921                                         &cached_state, 0);
8922         if (ret2) {
8923                 unlock_extent_cached(io_tree, page_start, page_end,
8924                                      &cached_state);
8925                 ret = VM_FAULT_SIGBUS;
8926                 goto out_unlock;
8927         }
8928         ret2 = 0;
8929
8930         /* page is wholly or partially inside EOF */
8931         if (page_start + PAGE_SIZE > size)
8932                 zero_start = size & ~PAGE_MASK;
8933         else
8934                 zero_start = PAGE_SIZE;
8935
8936         if (zero_start != PAGE_SIZE) {
8937                 kaddr = kmap(page);
8938                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
8939                 flush_dcache_page(page);
8940                 kunmap(page);
8941         }
8942         ClearPageChecked(page);
8943         set_page_dirty(page);
8944         SetPageUptodate(page);
8945
8946         BTRFS_I(inode)->last_trans = fs_info->generation;
8947         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8948         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8949
8950         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
8951
8952         if (!ret2) {
8953                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, true);
8954                 sb_end_pagefault(inode->i_sb);
8955                 extent_changeset_free(data_reserved);
8956                 return VM_FAULT_LOCKED;
8957         }
8958
8959 out_unlock:
8960         unlock_page(page);
8961 out:
8962         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, (ret != 0));
8963         btrfs_delalloc_release_space(inode, data_reserved, page_start,
8964                                      reserved_space, (ret != 0));
8965 out_noreserve:
8966         sb_end_pagefault(inode->i_sb);
8967         extent_changeset_free(data_reserved);
8968         return ret;
8969 }
8970
8971 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
8972 {
8973         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8974         struct btrfs_root *root = BTRFS_I(inode)->root;
8975         struct btrfs_block_rsv *rsv;
8976         int ret;
8977         struct btrfs_trans_handle *trans;
8978         u64 mask = fs_info->sectorsize - 1;
8979         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
8980
8981         if (!skip_writeback) {
8982                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8983                                                (u64)-1);
8984                 if (ret)
8985                         return ret;
8986         }
8987
8988         /*
8989          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
8990          * things going on here:
8991          *
8992          * 1) We need to reserve space to update our inode.
8993          *
8994          * 2) We need to have something to cache all the space that is going to
8995          * be free'd up by the truncate operation, but also have some slack
8996          * space reserved in case it uses space during the truncate (thank you
8997          * very much snapshotting).
8998          *
8999          * And we need these to be separate.  The fact is we can use a lot of
9000          * space doing the truncate, and we have no earthly idea how much space
9001          * we will use, so we need the truncate reservation to be separate so it
9002          * doesn't end up using space reserved for updating the inode.  We also
9003          * need to be able to stop the transaction and start a new one, which
9004          * means we need to be able to update the inode several times, and we
9005          * have no idea of knowing how many times that will be, so we can't just
9006          * reserve 1 item for the entirety of the operation, so that has to be
9007          * done separately as well.
9008          *
9009          * So that leaves us with
9010          *
9011          * 1) rsv - for the truncate reservation, which we will steal from the
9012          * transaction reservation.
9013          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
9014          * updating the inode.
9015          */
9016         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9017         if (!rsv)
9018                 return -ENOMEM;
9019         rsv->size = min_size;
9020         rsv->failfast = 1;
9021
9022         /*
9023          * 1 for the truncate slack space
9024          * 1 for updating the inode.
9025          */
9026         trans = btrfs_start_transaction(root, 2);
9027         if (IS_ERR(trans)) {
9028                 ret = PTR_ERR(trans);
9029                 goto out;
9030         }
9031
9032         /* Migrate the slack space for the truncate to our reserve */
9033         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9034                                       min_size, 0);
9035         BUG_ON(ret);
9036
9037         /*
9038          * So if we truncate and then write and fsync we normally would just
9039          * write the extents that changed, which is a problem if we need to
9040          * first truncate that entire inode.  So set this flag so we write out
9041          * all of the extents in the inode to the sync log so we're completely
9042          * safe.
9043          */
9044         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9045         trans->block_rsv = rsv;
9046
9047         while (1) {
9048                 ret = btrfs_truncate_inode_items(trans, root, inode,
9049                                                  inode->i_size,
9050                                                  BTRFS_EXTENT_DATA_KEY);
9051                 trans->block_rsv = &fs_info->trans_block_rsv;
9052                 if (ret != -ENOSPC && ret != -EAGAIN)
9053                         break;
9054
9055                 ret = btrfs_update_inode(trans, root, inode);
9056                 if (ret)
9057                         break;
9058
9059                 btrfs_end_transaction(trans);
9060                 btrfs_btree_balance_dirty(fs_info);
9061
9062                 trans = btrfs_start_transaction(root, 2);
9063                 if (IS_ERR(trans)) {
9064                         ret = PTR_ERR(trans);
9065                         trans = NULL;
9066                         break;
9067                 }
9068
9069                 btrfs_block_rsv_release(fs_info, rsv, -1);
9070                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9071                                               rsv, min_size, 0);
9072                 BUG_ON(ret);    /* shouldn't happen */
9073                 trans->block_rsv = rsv;
9074         }
9075
9076         /*
9077          * We can't call btrfs_truncate_block inside a trans handle as we could
9078          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9079          * we've truncated everything except the last little bit, and can do
9080          * btrfs_truncate_block and then update the disk_i_size.
9081          */
9082         if (ret == NEED_TRUNCATE_BLOCK) {
9083                 btrfs_end_transaction(trans);
9084                 btrfs_btree_balance_dirty(fs_info);
9085
9086                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9087                 if (ret)
9088                         goto out;
9089                 trans = btrfs_start_transaction(root, 1);
9090                 if (IS_ERR(trans)) {
9091                         ret = PTR_ERR(trans);
9092                         goto out;
9093                 }
9094                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9095         }
9096
9097         if (trans) {
9098                 int ret2;
9099
9100                 trans->block_rsv = &fs_info->trans_block_rsv;
9101                 ret2 = btrfs_update_inode(trans, root, inode);
9102                 if (ret2 && !ret)
9103                         ret = ret2;
9104
9105                 ret2 = btrfs_end_transaction(trans);
9106                 if (ret2 && !ret)
9107                         ret = ret2;
9108                 btrfs_btree_balance_dirty(fs_info);
9109         }
9110 out:
9111         btrfs_free_block_rsv(fs_info, rsv);
9112
9113         return ret;
9114 }
9115
9116 /*
9117  * create a new subvolume directory/inode (helper for the ioctl).
9118  */
9119 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9120                              struct btrfs_root *new_root,
9121                              struct btrfs_root *parent_root,
9122                              u64 new_dirid)
9123 {
9124         struct inode *inode;
9125         int err;
9126         u64 index = 0;
9127
9128         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9129                                 new_dirid, new_dirid,
9130                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9131                                 &index);
9132         if (IS_ERR(inode))
9133                 return PTR_ERR(inode);
9134         inode->i_op = &btrfs_dir_inode_operations;
9135         inode->i_fop = &btrfs_dir_file_operations;
9136
9137         set_nlink(inode, 1);
9138         btrfs_i_size_write(BTRFS_I(inode), 0);
9139         unlock_new_inode(inode);
9140
9141         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9142         if (err)
9143                 btrfs_err(new_root->fs_info,
9144                           "error inheriting subvolume %llu properties: %d",
9145                           new_root->root_key.objectid, err);
9146
9147         err = btrfs_update_inode(trans, new_root, inode);
9148
9149         iput(inode);
9150         return err;
9151 }
9152
9153 struct inode *btrfs_alloc_inode(struct super_block *sb)
9154 {
9155         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9156         struct btrfs_inode *ei;
9157         struct inode *inode;
9158
9159         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9160         if (!ei)
9161                 return NULL;
9162
9163         ei->root = NULL;
9164         ei->generation = 0;
9165         ei->last_trans = 0;
9166         ei->last_sub_trans = 0;
9167         ei->logged_trans = 0;
9168         ei->delalloc_bytes = 0;
9169         ei->new_delalloc_bytes = 0;
9170         ei->defrag_bytes = 0;
9171         ei->disk_i_size = 0;
9172         ei->flags = 0;
9173         ei->csum_bytes = 0;
9174         ei->index_cnt = (u64)-1;
9175         ei->dir_index = 0;
9176         ei->last_unlink_trans = 0;
9177         ei->last_log_commit = 0;
9178
9179         spin_lock_init(&ei->lock);
9180         ei->outstanding_extents = 0;
9181         if (sb->s_magic != BTRFS_TEST_MAGIC)
9182                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9183                                               BTRFS_BLOCK_RSV_DELALLOC);
9184         ei->runtime_flags = 0;
9185         ei->prop_compress = BTRFS_COMPRESS_NONE;
9186         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9187
9188         ei->delayed_node = NULL;
9189
9190         ei->i_otime.tv_sec = 0;
9191         ei->i_otime.tv_nsec = 0;
9192
9193         inode = &ei->vfs_inode;
9194         extent_map_tree_init(&ei->extent_tree);
9195         extent_io_tree_init(&ei->io_tree, inode);
9196         extent_io_tree_init(&ei->io_failure_tree, inode);
9197         ei->io_tree.track_uptodate = 1;
9198         ei->io_failure_tree.track_uptodate = 1;
9199         atomic_set(&ei->sync_writers, 0);
9200         mutex_init(&ei->log_mutex);
9201         mutex_init(&ei->delalloc_mutex);
9202         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9203         INIT_LIST_HEAD(&ei->delalloc_inodes);
9204         INIT_LIST_HEAD(&ei->delayed_iput);
9205         RB_CLEAR_NODE(&ei->rb_node);
9206         init_rwsem(&ei->dio_sem);
9207
9208         return inode;
9209 }
9210
9211 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9212 void btrfs_test_destroy_inode(struct inode *inode)
9213 {
9214         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9215         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9216 }
9217 #endif
9218
9219 static void btrfs_i_callback(struct rcu_head *head)
9220 {
9221         struct inode *inode = container_of(head, struct inode, i_rcu);
9222         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9223 }
9224
9225 void btrfs_destroy_inode(struct inode *inode)
9226 {
9227         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9228         struct btrfs_ordered_extent *ordered;
9229         struct btrfs_root *root = BTRFS_I(inode)->root;
9230
9231         WARN_ON(!hlist_empty(&inode->i_dentry));
9232         WARN_ON(inode->i_data.nrpages);
9233         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9234         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9235         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9236         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9237         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9238         WARN_ON(BTRFS_I(inode)->csum_bytes);
9239         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9240
9241         /*
9242          * This can happen where we create an inode, but somebody else also
9243          * created the same inode and we need to destroy the one we already
9244          * created.
9245          */
9246         if (!root)
9247                 goto free;
9248
9249         while (1) {
9250                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9251                 if (!ordered)
9252                         break;
9253                 else {
9254                         btrfs_err(fs_info,
9255                                   "found ordered extent %llu %llu on inode cleanup",
9256                                   ordered->file_offset, ordered->len);
9257                         btrfs_remove_ordered_extent(inode, ordered);
9258                         btrfs_put_ordered_extent(ordered);
9259                         btrfs_put_ordered_extent(ordered);
9260                 }
9261         }
9262         btrfs_qgroup_check_reserved_leak(inode);
9263         inode_tree_del(inode);
9264         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9265 free:
9266         call_rcu(&inode->i_rcu, btrfs_i_callback);
9267 }
9268
9269 int btrfs_drop_inode(struct inode *inode)
9270 {
9271         struct btrfs_root *root = BTRFS_I(inode)->root;
9272
9273         if (root == NULL)
9274                 return 1;
9275
9276         /* the snap/subvol tree is on deleting */
9277         if (btrfs_root_refs(&root->root_item) == 0)
9278                 return 1;
9279         else
9280                 return generic_drop_inode(inode);
9281 }
9282
9283 static void init_once(void *foo)
9284 {
9285         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9286
9287         inode_init_once(&ei->vfs_inode);
9288 }
9289
9290 void __cold btrfs_destroy_cachep(void)
9291 {
9292         /*
9293          * Make sure all delayed rcu free inodes are flushed before we
9294          * destroy cache.
9295          */
9296         rcu_barrier();
9297         kmem_cache_destroy(btrfs_inode_cachep);
9298         kmem_cache_destroy(btrfs_trans_handle_cachep);
9299         kmem_cache_destroy(btrfs_path_cachep);
9300         kmem_cache_destroy(btrfs_free_space_cachep);
9301 }
9302
9303 int __init btrfs_init_cachep(void)
9304 {
9305         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9306                         sizeof(struct btrfs_inode), 0,
9307                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9308                         init_once);
9309         if (!btrfs_inode_cachep)
9310                 goto fail;
9311
9312         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9313                         sizeof(struct btrfs_trans_handle), 0,
9314                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9315         if (!btrfs_trans_handle_cachep)
9316                 goto fail;
9317
9318         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9319                         sizeof(struct btrfs_path), 0,
9320                         SLAB_MEM_SPREAD, NULL);
9321         if (!btrfs_path_cachep)
9322                 goto fail;
9323
9324         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9325                         sizeof(struct btrfs_free_space), 0,
9326                         SLAB_MEM_SPREAD, NULL);
9327         if (!btrfs_free_space_cachep)
9328                 goto fail;
9329
9330         return 0;
9331 fail:
9332         btrfs_destroy_cachep();
9333         return -ENOMEM;
9334 }
9335
9336 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9337                          u32 request_mask, unsigned int flags)
9338 {
9339         u64 delalloc_bytes;
9340         struct inode *inode = d_inode(path->dentry);
9341         u32 blocksize = inode->i_sb->s_blocksize;
9342         u32 bi_flags = BTRFS_I(inode)->flags;
9343
9344         stat->result_mask |= STATX_BTIME;
9345         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9346         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9347         if (bi_flags & BTRFS_INODE_APPEND)
9348                 stat->attributes |= STATX_ATTR_APPEND;
9349         if (bi_flags & BTRFS_INODE_COMPRESS)
9350                 stat->attributes |= STATX_ATTR_COMPRESSED;
9351         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9352                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9353         if (bi_flags & BTRFS_INODE_NODUMP)
9354                 stat->attributes |= STATX_ATTR_NODUMP;
9355
9356         stat->attributes_mask |= (STATX_ATTR_APPEND |
9357                                   STATX_ATTR_COMPRESSED |
9358                                   STATX_ATTR_IMMUTABLE |
9359                                   STATX_ATTR_NODUMP);
9360
9361         generic_fillattr(inode, stat);
9362         stat->dev = BTRFS_I(inode)->root->anon_dev;
9363
9364         spin_lock(&BTRFS_I(inode)->lock);
9365         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9366         spin_unlock(&BTRFS_I(inode)->lock);
9367         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9368                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9369         return 0;
9370 }
9371
9372 static int btrfs_rename_exchange(struct inode *old_dir,
9373                               struct dentry *old_dentry,
9374                               struct inode *new_dir,
9375                               struct dentry *new_dentry)
9376 {
9377         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9378         struct btrfs_trans_handle *trans;
9379         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9380         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9381         struct inode *new_inode = new_dentry->d_inode;
9382         struct inode *old_inode = old_dentry->d_inode;
9383         struct timespec64 ctime = current_time(old_inode);
9384         struct dentry *parent;
9385         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9386         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9387         u64 old_idx = 0;
9388         u64 new_idx = 0;
9389         u64 root_objectid;
9390         int ret;
9391         int ret2;
9392         bool root_log_pinned = false;
9393         bool dest_log_pinned = false;
9394
9395         /* we only allow rename subvolume link between subvolumes */
9396         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9397                 return -EXDEV;
9398
9399         /* close the race window with snapshot create/destroy ioctl */
9400         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9401                 down_read(&fs_info->subvol_sem);
9402         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9403                 down_read(&fs_info->subvol_sem);
9404
9405         /*
9406          * We want to reserve the absolute worst case amount of items.  So if
9407          * both inodes are subvols and we need to unlink them then that would
9408          * require 4 item modifications, but if they are both normal inodes it
9409          * would require 5 item modifications, so we'll assume their normal
9410          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9411          * should cover the worst case number of items we'll modify.
9412          */
9413         trans = btrfs_start_transaction(root, 12);
9414         if (IS_ERR(trans)) {
9415                 ret = PTR_ERR(trans);
9416                 goto out_notrans;
9417         }
9418
9419         /*
9420          * We need to find a free sequence number both in the source and
9421          * in the destination directory for the exchange.
9422          */
9423         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9424         if (ret)
9425                 goto out_fail;
9426         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9427         if (ret)
9428                 goto out_fail;
9429
9430         BTRFS_I(old_inode)->dir_index = 0ULL;
9431         BTRFS_I(new_inode)->dir_index = 0ULL;
9432
9433         /* Reference for the source. */
9434         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9435                 /* force full log commit if subvolume involved. */
9436                 btrfs_set_log_full_commit(fs_info, trans);
9437         } else {
9438                 btrfs_pin_log_trans(root);
9439                 root_log_pinned = true;
9440                 ret = btrfs_insert_inode_ref(trans, dest,
9441                                              new_dentry->d_name.name,
9442                                              new_dentry->d_name.len,
9443                                              old_ino,
9444                                              btrfs_ino(BTRFS_I(new_dir)),
9445                                              old_idx);
9446                 if (ret)
9447                         goto out_fail;
9448         }
9449
9450         /* And now for the dest. */
9451         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9452                 /* force full log commit if subvolume involved. */
9453                 btrfs_set_log_full_commit(fs_info, trans);
9454         } else {
9455                 btrfs_pin_log_trans(dest);
9456                 dest_log_pinned = true;
9457                 ret = btrfs_insert_inode_ref(trans, root,
9458                                              old_dentry->d_name.name,
9459                                              old_dentry->d_name.len,
9460                                              new_ino,
9461                                              btrfs_ino(BTRFS_I(old_dir)),
9462                                              new_idx);
9463                 if (ret)
9464                         goto out_fail;
9465         }
9466
9467         /* Update inode version and ctime/mtime. */
9468         inode_inc_iversion(old_dir);
9469         inode_inc_iversion(new_dir);
9470         inode_inc_iversion(old_inode);
9471         inode_inc_iversion(new_inode);
9472         old_dir->i_ctime = old_dir->i_mtime = ctime;
9473         new_dir->i_ctime = new_dir->i_mtime = ctime;
9474         old_inode->i_ctime = ctime;
9475         new_inode->i_ctime = ctime;
9476
9477         if (old_dentry->d_parent != new_dentry->d_parent) {
9478                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9479                                 BTRFS_I(old_inode), 1);
9480                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9481                                 BTRFS_I(new_inode), 1);
9482         }
9483
9484         /* src is a subvolume */
9485         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9486                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9487                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9488                                           old_dentry->d_name.name,
9489                                           old_dentry->d_name.len);
9490         } else { /* src is an inode */
9491                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9492                                            BTRFS_I(old_dentry->d_inode),
9493                                            old_dentry->d_name.name,
9494                                            old_dentry->d_name.len);
9495                 if (!ret)
9496                         ret = btrfs_update_inode(trans, root, old_inode);
9497         }
9498         if (ret) {
9499                 btrfs_abort_transaction(trans, ret);
9500                 goto out_fail;
9501         }
9502
9503         /* dest is a subvolume */
9504         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9505                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9506                 ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9507                                           new_dentry->d_name.name,
9508                                           new_dentry->d_name.len);
9509         } else { /* dest is an inode */
9510                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9511                                            BTRFS_I(new_dentry->d_inode),
9512                                            new_dentry->d_name.name,
9513                                            new_dentry->d_name.len);
9514                 if (!ret)
9515                         ret = btrfs_update_inode(trans, dest, new_inode);
9516         }
9517         if (ret) {
9518                 btrfs_abort_transaction(trans, ret);
9519                 goto out_fail;
9520         }
9521
9522         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9523                              new_dentry->d_name.name,
9524                              new_dentry->d_name.len, 0, old_idx);
9525         if (ret) {
9526                 btrfs_abort_transaction(trans, ret);
9527                 goto out_fail;
9528         }
9529
9530         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9531                              old_dentry->d_name.name,
9532                              old_dentry->d_name.len, 0, new_idx);
9533         if (ret) {
9534                 btrfs_abort_transaction(trans, ret);
9535                 goto out_fail;
9536         }
9537
9538         if (old_inode->i_nlink == 1)
9539                 BTRFS_I(old_inode)->dir_index = old_idx;
9540         if (new_inode->i_nlink == 1)
9541                 BTRFS_I(new_inode)->dir_index = new_idx;
9542
9543         if (root_log_pinned) {
9544                 parent = new_dentry->d_parent;
9545                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9546                                 parent);
9547                 btrfs_end_log_trans(root);
9548                 root_log_pinned = false;
9549         }
9550         if (dest_log_pinned) {
9551                 parent = old_dentry->d_parent;
9552                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9553                                 parent);
9554                 btrfs_end_log_trans(dest);
9555                 dest_log_pinned = false;
9556         }
9557 out_fail:
9558         /*
9559          * If we have pinned a log and an error happened, we unpin tasks
9560          * trying to sync the log and force them to fallback to a transaction
9561          * commit if the log currently contains any of the inodes involved in
9562          * this rename operation (to ensure we do not persist a log with an
9563          * inconsistent state for any of these inodes or leading to any
9564          * inconsistencies when replayed). If the transaction was aborted, the
9565          * abortion reason is propagated to userspace when attempting to commit
9566          * the transaction. If the log does not contain any of these inodes, we
9567          * allow the tasks to sync it.
9568          */
9569         if (ret && (root_log_pinned || dest_log_pinned)) {
9570                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9571                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9572                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9573                     (new_inode &&
9574                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9575                         btrfs_set_log_full_commit(fs_info, trans);
9576
9577                 if (root_log_pinned) {
9578                         btrfs_end_log_trans(root);
9579                         root_log_pinned = false;
9580                 }
9581                 if (dest_log_pinned) {
9582                         btrfs_end_log_trans(dest);
9583                         dest_log_pinned = false;
9584                 }
9585         }
9586         ret2 = btrfs_end_transaction(trans);
9587         ret = ret ? ret : ret2;
9588 out_notrans:
9589         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9590                 up_read(&fs_info->subvol_sem);
9591         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9592                 up_read(&fs_info->subvol_sem);
9593
9594         return ret;
9595 }
9596
9597 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9598                                      struct btrfs_root *root,
9599                                      struct inode *dir,
9600                                      struct dentry *dentry)
9601 {
9602         int ret;
9603         struct inode *inode;
9604         u64 objectid;
9605         u64 index;
9606
9607         ret = btrfs_find_free_ino(root, &objectid);
9608         if (ret)
9609                 return ret;
9610
9611         inode = btrfs_new_inode(trans, root, dir,
9612                                 dentry->d_name.name,
9613                                 dentry->d_name.len,
9614                                 btrfs_ino(BTRFS_I(dir)),
9615                                 objectid,
9616                                 S_IFCHR | WHITEOUT_MODE,
9617                                 &index);
9618
9619         if (IS_ERR(inode)) {
9620                 ret = PTR_ERR(inode);
9621                 return ret;
9622         }
9623
9624         inode->i_op = &btrfs_special_inode_operations;
9625         init_special_inode(inode, inode->i_mode,
9626                 WHITEOUT_DEV);
9627
9628         ret = btrfs_init_inode_security(trans, inode, dir,
9629                                 &dentry->d_name);
9630         if (ret)
9631                 goto out;
9632
9633         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9634                                 BTRFS_I(inode), 0, index);
9635         if (ret)
9636                 goto out;
9637
9638         ret = btrfs_update_inode(trans, root, inode);
9639 out:
9640         unlock_new_inode(inode);
9641         if (ret)
9642                 inode_dec_link_count(inode);
9643         iput(inode);
9644
9645         return ret;
9646 }
9647
9648 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9649                            struct inode *new_dir, struct dentry *new_dentry,
9650                            unsigned int flags)
9651 {
9652         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9653         struct btrfs_trans_handle *trans;
9654         unsigned int trans_num_items;
9655         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9656         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9657         struct inode *new_inode = d_inode(new_dentry);
9658         struct inode *old_inode = d_inode(old_dentry);
9659         u64 index = 0;
9660         u64 root_objectid;
9661         int ret;
9662         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9663         bool log_pinned = false;
9664
9665         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9666                 return -EPERM;
9667
9668         /* we only allow rename subvolume link between subvolumes */
9669         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9670                 return -EXDEV;
9671
9672         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9673             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9674                 return -ENOTEMPTY;
9675
9676         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9677             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9678                 return -ENOTEMPTY;
9679
9680
9681         /* check for collisions, even if the  name isn't there */
9682         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9683                              new_dentry->d_name.name,
9684                              new_dentry->d_name.len);
9685
9686         if (ret) {
9687                 if (ret == -EEXIST) {
9688                         /* we shouldn't get
9689                          * eexist without a new_inode */
9690                         if (WARN_ON(!new_inode)) {
9691                                 return ret;
9692                         }
9693                 } else {
9694                         /* maybe -EOVERFLOW */
9695                         return ret;
9696                 }
9697         }
9698         ret = 0;
9699
9700         /*
9701          * we're using rename to replace one file with another.  Start IO on it
9702          * now so  we don't add too much work to the end of the transaction
9703          */
9704         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9705                 filemap_flush(old_inode->i_mapping);
9706
9707         /* close the racy window with snapshot create/destroy ioctl */
9708         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9709                 down_read(&fs_info->subvol_sem);
9710         /*
9711          * We want to reserve the absolute worst case amount of items.  So if
9712          * both inodes are subvols and we need to unlink them then that would
9713          * require 4 item modifications, but if they are both normal inodes it
9714          * would require 5 item modifications, so we'll assume they are normal
9715          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9716          * should cover the worst case number of items we'll modify.
9717          * If our rename has the whiteout flag, we need more 5 units for the
9718          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9719          * when selinux is enabled).
9720          */
9721         trans_num_items = 11;
9722         if (flags & RENAME_WHITEOUT)
9723                 trans_num_items += 5;
9724         trans = btrfs_start_transaction(root, trans_num_items);
9725         if (IS_ERR(trans)) {
9726                 ret = PTR_ERR(trans);
9727                 goto out_notrans;
9728         }
9729
9730         if (dest != root)
9731                 btrfs_record_root_in_trans(trans, dest);
9732
9733         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9734         if (ret)
9735                 goto out_fail;
9736
9737         BTRFS_I(old_inode)->dir_index = 0ULL;
9738         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9739                 /* force full log commit if subvolume involved. */
9740                 btrfs_set_log_full_commit(fs_info, trans);
9741         } else {
9742                 btrfs_pin_log_trans(root);
9743                 log_pinned = true;
9744                 ret = btrfs_insert_inode_ref(trans, dest,
9745                                              new_dentry->d_name.name,
9746                                              new_dentry->d_name.len,
9747                                              old_ino,
9748                                              btrfs_ino(BTRFS_I(new_dir)), index);
9749                 if (ret)
9750                         goto out_fail;
9751         }
9752
9753         inode_inc_iversion(old_dir);
9754         inode_inc_iversion(new_dir);
9755         inode_inc_iversion(old_inode);
9756         old_dir->i_ctime = old_dir->i_mtime =
9757         new_dir->i_ctime = new_dir->i_mtime =
9758         old_inode->i_ctime = current_time(old_dir);
9759
9760         if (old_dentry->d_parent != new_dentry->d_parent)
9761                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9762                                 BTRFS_I(old_inode), 1);
9763
9764         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9765                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9766                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9767                                         old_dentry->d_name.name,
9768                                         old_dentry->d_name.len);
9769         } else {
9770                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9771                                         BTRFS_I(d_inode(old_dentry)),
9772                                         old_dentry->d_name.name,
9773                                         old_dentry->d_name.len);
9774                 if (!ret)
9775                         ret = btrfs_update_inode(trans, root, old_inode);
9776         }
9777         if (ret) {
9778                 btrfs_abort_transaction(trans, ret);
9779                 goto out_fail;
9780         }
9781
9782         if (new_inode) {
9783                 inode_inc_iversion(new_inode);
9784                 new_inode->i_ctime = current_time(new_inode);
9785                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9786                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9787                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9788                         ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9789                                                 new_dentry->d_name.name,
9790                                                 new_dentry->d_name.len);
9791                         BUG_ON(new_inode->i_nlink == 0);
9792                 } else {
9793                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9794                                                  BTRFS_I(d_inode(new_dentry)),
9795                                                  new_dentry->d_name.name,
9796                                                  new_dentry->d_name.len);
9797                 }
9798                 if (!ret && new_inode->i_nlink == 0)
9799                         ret = btrfs_orphan_add(trans,
9800                                         BTRFS_I(d_inode(new_dentry)));
9801                 if (ret) {
9802                         btrfs_abort_transaction(trans, ret);
9803                         goto out_fail;
9804                 }
9805         }
9806
9807         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9808                              new_dentry->d_name.name,
9809                              new_dentry->d_name.len, 0, index);
9810         if (ret) {
9811                 btrfs_abort_transaction(trans, ret);
9812                 goto out_fail;
9813         }
9814
9815         if (old_inode->i_nlink == 1)
9816                 BTRFS_I(old_inode)->dir_index = index;
9817
9818         if (log_pinned) {
9819                 struct dentry *parent = new_dentry->d_parent;
9820
9821                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9822                                 parent);
9823                 btrfs_end_log_trans(root);
9824                 log_pinned = false;
9825         }
9826
9827         if (flags & RENAME_WHITEOUT) {
9828                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9829                                                 old_dentry);
9830
9831                 if (ret) {
9832                         btrfs_abort_transaction(trans, ret);
9833                         goto out_fail;
9834                 }
9835         }
9836 out_fail:
9837         /*
9838          * If we have pinned the log and an error happened, we unpin tasks
9839          * trying to sync the log and force them to fallback to a transaction
9840          * commit if the log currently contains any of the inodes involved in
9841          * this rename operation (to ensure we do not persist a log with an
9842          * inconsistent state for any of these inodes or leading to any
9843          * inconsistencies when replayed). If the transaction was aborted, the
9844          * abortion reason is propagated to userspace when attempting to commit
9845          * the transaction. If the log does not contain any of these inodes, we
9846          * allow the tasks to sync it.
9847          */
9848         if (ret && log_pinned) {
9849                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9850                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9851                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9852                     (new_inode &&
9853                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9854                         btrfs_set_log_full_commit(fs_info, trans);
9855
9856                 btrfs_end_log_trans(root);
9857                 log_pinned = false;
9858         }
9859         btrfs_end_transaction(trans);
9860 out_notrans:
9861         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9862                 up_read(&fs_info->subvol_sem);
9863
9864         return ret;
9865 }
9866
9867 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9868                          struct inode *new_dir, struct dentry *new_dentry,
9869                          unsigned int flags)
9870 {
9871         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9872                 return -EINVAL;
9873
9874         if (flags & RENAME_EXCHANGE)
9875                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9876                                           new_dentry);
9877
9878         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9879 }
9880
9881 struct btrfs_delalloc_work {
9882         struct inode *inode;
9883         struct completion completion;
9884         struct list_head list;
9885         struct btrfs_work work;
9886 };
9887
9888 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9889 {
9890         struct btrfs_delalloc_work *delalloc_work;
9891         struct inode *inode;
9892
9893         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9894                                      work);
9895         inode = delalloc_work->inode;
9896         filemap_flush(inode->i_mapping);
9897         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9898                                 &BTRFS_I(inode)->runtime_flags))
9899                 filemap_flush(inode->i_mapping);
9900
9901         iput(inode);
9902         complete(&delalloc_work->completion);
9903 }
9904
9905 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
9906 {
9907         struct btrfs_delalloc_work *work;
9908
9909         work = kmalloc(sizeof(*work), GFP_NOFS);
9910         if (!work)
9911                 return NULL;
9912
9913         init_completion(&work->completion);
9914         INIT_LIST_HEAD(&work->list);
9915         work->inode = inode;
9916         WARN_ON_ONCE(!inode);
9917         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9918                         btrfs_run_delalloc_work, NULL, NULL);
9919
9920         return work;
9921 }
9922
9923 /*
9924  * some fairly slow code that needs optimization. This walks the list
9925  * of all the inodes with pending delalloc and forces them to disk.
9926  */
9927 static int start_delalloc_inodes(struct btrfs_root *root, int nr)
9928 {
9929         struct btrfs_inode *binode;
9930         struct inode *inode;
9931         struct btrfs_delalloc_work *work, *next;
9932         struct list_head works;
9933         struct list_head splice;
9934         int ret = 0;
9935
9936         INIT_LIST_HEAD(&works);
9937         INIT_LIST_HEAD(&splice);
9938
9939         mutex_lock(&root->delalloc_mutex);
9940         spin_lock(&root->delalloc_lock);
9941         list_splice_init(&root->delalloc_inodes, &splice);
9942         while (!list_empty(&splice)) {
9943                 binode = list_entry(splice.next, struct btrfs_inode,
9944                                     delalloc_inodes);
9945
9946                 list_move_tail(&binode->delalloc_inodes,
9947                                &root->delalloc_inodes);
9948                 inode = igrab(&binode->vfs_inode);
9949                 if (!inode) {
9950                         cond_resched_lock(&root->delalloc_lock);
9951                         continue;
9952                 }
9953                 spin_unlock(&root->delalloc_lock);
9954
9955                 work = btrfs_alloc_delalloc_work(inode);
9956                 if (!work) {
9957                         iput(inode);
9958                         ret = -ENOMEM;
9959                         goto out;
9960                 }
9961                 list_add_tail(&work->list, &works);
9962                 btrfs_queue_work(root->fs_info->flush_workers,
9963                                  &work->work);
9964                 ret++;
9965                 if (nr != -1 && ret >= nr)
9966                         goto out;
9967                 cond_resched();
9968                 spin_lock(&root->delalloc_lock);
9969         }
9970         spin_unlock(&root->delalloc_lock);
9971
9972 out:
9973         list_for_each_entry_safe(work, next, &works, list) {
9974                 list_del_init(&work->list);
9975                 wait_for_completion(&work->completion);
9976                 kfree(work);
9977         }
9978
9979         if (!list_empty(&splice)) {
9980                 spin_lock(&root->delalloc_lock);
9981                 list_splice_tail(&splice, &root->delalloc_inodes);
9982                 spin_unlock(&root->delalloc_lock);
9983         }
9984         mutex_unlock(&root->delalloc_mutex);
9985         return ret;
9986 }
9987
9988 int btrfs_start_delalloc_inodes(struct btrfs_root *root)
9989 {
9990         struct btrfs_fs_info *fs_info = root->fs_info;
9991         int ret;
9992
9993         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9994                 return -EROFS;
9995
9996         ret = start_delalloc_inodes(root, -1);
9997         if (ret > 0)
9998                 ret = 0;
9999         return ret;
10000 }
10001
10002 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
10003 {
10004         struct btrfs_root *root;
10005         struct list_head splice;
10006         int ret;
10007
10008         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10009                 return -EROFS;
10010
10011         INIT_LIST_HEAD(&splice);
10012
10013         mutex_lock(&fs_info->delalloc_root_mutex);
10014         spin_lock(&fs_info->delalloc_root_lock);
10015         list_splice_init(&fs_info->delalloc_roots, &splice);
10016         while (!list_empty(&splice) && nr) {
10017                 root = list_first_entry(&splice, struct btrfs_root,
10018                                         delalloc_root);
10019                 root = btrfs_grab_fs_root(root);
10020                 BUG_ON(!root);
10021                 list_move_tail(&root->delalloc_root,
10022                                &fs_info->delalloc_roots);
10023                 spin_unlock(&fs_info->delalloc_root_lock);
10024
10025                 ret = start_delalloc_inodes(root, nr);
10026                 btrfs_put_fs_root(root);
10027                 if (ret < 0)
10028                         goto out;
10029
10030                 if (nr != -1) {
10031                         nr -= ret;
10032                         WARN_ON(nr < 0);
10033                 }
10034                 spin_lock(&fs_info->delalloc_root_lock);
10035         }
10036         spin_unlock(&fs_info->delalloc_root_lock);
10037
10038         ret = 0;
10039 out:
10040         if (!list_empty(&splice)) {
10041                 spin_lock(&fs_info->delalloc_root_lock);
10042                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10043                 spin_unlock(&fs_info->delalloc_root_lock);
10044         }
10045         mutex_unlock(&fs_info->delalloc_root_mutex);
10046         return ret;
10047 }
10048
10049 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10050                          const char *symname)
10051 {
10052         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10053         struct btrfs_trans_handle *trans;
10054         struct btrfs_root *root = BTRFS_I(dir)->root;
10055         struct btrfs_path *path;
10056         struct btrfs_key key;
10057         struct inode *inode = NULL;
10058         int err;
10059         u64 objectid;
10060         u64 index = 0;
10061         int name_len;
10062         int datasize;
10063         unsigned long ptr;
10064         struct btrfs_file_extent_item *ei;
10065         struct extent_buffer *leaf;
10066
10067         name_len = strlen(symname);
10068         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10069                 return -ENAMETOOLONG;
10070
10071         /*
10072          * 2 items for inode item and ref
10073          * 2 items for dir items
10074          * 1 item for updating parent inode item
10075          * 1 item for the inline extent item
10076          * 1 item for xattr if selinux is on
10077          */
10078         trans = btrfs_start_transaction(root, 7);
10079         if (IS_ERR(trans))
10080                 return PTR_ERR(trans);
10081
10082         err = btrfs_find_free_ino(root, &objectid);
10083         if (err)
10084                 goto out_unlock;
10085
10086         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10087                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10088                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10089         if (IS_ERR(inode)) {
10090                 err = PTR_ERR(inode);
10091                 inode = NULL;
10092                 goto out_unlock;
10093         }
10094
10095         /*
10096         * If the active LSM wants to access the inode during
10097         * d_instantiate it needs these. Smack checks to see
10098         * if the filesystem supports xattrs by looking at the
10099         * ops vector.
10100         */
10101         inode->i_fop = &btrfs_file_operations;
10102         inode->i_op = &btrfs_file_inode_operations;
10103         inode->i_mapping->a_ops = &btrfs_aops;
10104         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10105
10106         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10107         if (err)
10108                 goto out_unlock;
10109
10110         path = btrfs_alloc_path();
10111         if (!path) {
10112                 err = -ENOMEM;
10113                 goto out_unlock;
10114         }
10115         key.objectid = btrfs_ino(BTRFS_I(inode));
10116         key.offset = 0;
10117         key.type = BTRFS_EXTENT_DATA_KEY;
10118         datasize = btrfs_file_extent_calc_inline_size(name_len);
10119         err = btrfs_insert_empty_item(trans, root, path, &key,
10120                                       datasize);
10121         if (err) {
10122                 btrfs_free_path(path);
10123                 goto out_unlock;
10124         }
10125         leaf = path->nodes[0];
10126         ei = btrfs_item_ptr(leaf, path->slots[0],
10127                             struct btrfs_file_extent_item);
10128         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10129         btrfs_set_file_extent_type(leaf, ei,
10130                                    BTRFS_FILE_EXTENT_INLINE);
10131         btrfs_set_file_extent_encryption(leaf, ei, 0);
10132         btrfs_set_file_extent_compression(leaf, ei, 0);
10133         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10134         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10135
10136         ptr = btrfs_file_extent_inline_start(ei);
10137         write_extent_buffer(leaf, symname, ptr, name_len);
10138         btrfs_mark_buffer_dirty(leaf);
10139         btrfs_free_path(path);
10140
10141         inode->i_op = &btrfs_symlink_inode_operations;
10142         inode_nohighmem(inode);
10143         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10144         inode_set_bytes(inode, name_len);
10145         btrfs_i_size_write(BTRFS_I(inode), name_len);
10146         err = btrfs_update_inode(trans, root, inode);
10147         /*
10148          * Last step, add directory indexes for our symlink inode. This is the
10149          * last step to avoid extra cleanup of these indexes if an error happens
10150          * elsewhere above.
10151          */
10152         if (!err)
10153                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10154                                 BTRFS_I(inode), 0, index);
10155         if (err)
10156                 goto out_unlock;
10157
10158         d_instantiate_new(dentry, inode);
10159
10160 out_unlock:
10161         btrfs_end_transaction(trans);
10162         if (err && inode) {
10163                 inode_dec_link_count(inode);
10164                 discard_new_inode(inode);
10165         }
10166         btrfs_btree_balance_dirty(fs_info);
10167         return err;
10168 }
10169
10170 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10171                                        u64 start, u64 num_bytes, u64 min_size,
10172                                        loff_t actual_len, u64 *alloc_hint,
10173                                        struct btrfs_trans_handle *trans)
10174 {
10175         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10176         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10177         struct extent_map *em;
10178         struct btrfs_root *root = BTRFS_I(inode)->root;
10179         struct btrfs_key ins;
10180         u64 cur_offset = start;
10181         u64 i_size;
10182         u64 cur_bytes;
10183         u64 last_alloc = (u64)-1;
10184         int ret = 0;
10185         bool own_trans = true;
10186         u64 end = start + num_bytes - 1;
10187
10188         if (trans)
10189                 own_trans = false;
10190         while (num_bytes > 0) {
10191                 if (own_trans) {
10192                         trans = btrfs_start_transaction(root, 3);
10193                         if (IS_ERR(trans)) {
10194                                 ret = PTR_ERR(trans);
10195                                 break;
10196                         }
10197                 }
10198
10199                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10200                 cur_bytes = max(cur_bytes, min_size);
10201                 /*
10202                  * If we are severely fragmented we could end up with really
10203                  * small allocations, so if the allocator is returning small
10204                  * chunks lets make its job easier by only searching for those
10205                  * sized chunks.
10206                  */
10207                 cur_bytes = min(cur_bytes, last_alloc);
10208                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10209                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10210                 if (ret) {
10211                         if (own_trans)
10212                                 btrfs_end_transaction(trans);
10213                         break;
10214                 }
10215                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10216
10217                 last_alloc = ins.offset;
10218                 ret = insert_reserved_file_extent(trans, inode,
10219                                                   cur_offset, ins.objectid,
10220                                                   ins.offset, ins.offset,
10221                                                   ins.offset, 0, 0, 0,
10222                                                   BTRFS_FILE_EXTENT_PREALLOC);
10223                 if (ret) {
10224                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10225                                                    ins.offset, 0);
10226                         btrfs_abort_transaction(trans, ret);
10227                         if (own_trans)
10228                                 btrfs_end_transaction(trans);
10229                         break;
10230                 }
10231
10232                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10233                                         cur_offset + ins.offset -1, 0);
10234
10235                 em = alloc_extent_map();
10236                 if (!em) {
10237                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10238                                 &BTRFS_I(inode)->runtime_flags);
10239                         goto next;
10240                 }
10241
10242                 em->start = cur_offset;
10243                 em->orig_start = cur_offset;
10244                 em->len = ins.offset;
10245                 em->block_start = ins.objectid;
10246                 em->block_len = ins.offset;
10247                 em->orig_block_len = ins.offset;
10248                 em->ram_bytes = ins.offset;
10249                 em->bdev = fs_info->fs_devices->latest_bdev;
10250                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10251                 em->generation = trans->transid;
10252
10253                 while (1) {
10254                         write_lock(&em_tree->lock);
10255                         ret = add_extent_mapping(em_tree, em, 1);
10256                         write_unlock(&em_tree->lock);
10257                         if (ret != -EEXIST)
10258                                 break;
10259                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10260                                                 cur_offset + ins.offset - 1,
10261                                                 0);
10262                 }
10263                 free_extent_map(em);
10264 next:
10265                 num_bytes -= ins.offset;
10266                 cur_offset += ins.offset;
10267                 *alloc_hint = ins.objectid + ins.offset;
10268
10269                 inode_inc_iversion(inode);
10270                 inode->i_ctime = current_time(inode);
10271                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10272                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10273                     (actual_len > inode->i_size) &&
10274                     (cur_offset > inode->i_size)) {
10275                         if (cur_offset > actual_len)
10276                                 i_size = actual_len;
10277                         else
10278                                 i_size = cur_offset;
10279                         i_size_write(inode, i_size);
10280                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10281                 }
10282
10283                 ret = btrfs_update_inode(trans, root, inode);
10284
10285                 if (ret) {
10286                         btrfs_abort_transaction(trans, ret);
10287                         if (own_trans)
10288                                 btrfs_end_transaction(trans);
10289                         break;
10290                 }
10291
10292                 if (own_trans)
10293                         btrfs_end_transaction(trans);
10294         }
10295         if (cur_offset < end)
10296                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10297                         end - cur_offset + 1);
10298         return ret;
10299 }
10300
10301 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10302                               u64 start, u64 num_bytes, u64 min_size,
10303                               loff_t actual_len, u64 *alloc_hint)
10304 {
10305         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10306                                            min_size, actual_len, alloc_hint,
10307                                            NULL);
10308 }
10309
10310 int btrfs_prealloc_file_range_trans(struct inode *inode,
10311                                     struct btrfs_trans_handle *trans, int mode,
10312                                     u64 start, u64 num_bytes, u64 min_size,
10313                                     loff_t actual_len, u64 *alloc_hint)
10314 {
10315         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10316                                            min_size, actual_len, alloc_hint, trans);
10317 }
10318
10319 static int btrfs_set_page_dirty(struct page *page)
10320 {
10321         return __set_page_dirty_nobuffers(page);
10322 }
10323
10324 static int btrfs_permission(struct inode *inode, int mask)
10325 {
10326         struct btrfs_root *root = BTRFS_I(inode)->root;
10327         umode_t mode = inode->i_mode;
10328
10329         if (mask & MAY_WRITE &&
10330             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10331                 if (btrfs_root_readonly(root))
10332                         return -EROFS;
10333                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10334                         return -EACCES;
10335         }
10336         return generic_permission(inode, mask);
10337 }
10338
10339 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10340 {
10341         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10342         struct btrfs_trans_handle *trans;
10343         struct btrfs_root *root = BTRFS_I(dir)->root;
10344         struct inode *inode = NULL;
10345         u64 objectid;
10346         u64 index;
10347         int ret = 0;
10348
10349         /*
10350          * 5 units required for adding orphan entry
10351          */
10352         trans = btrfs_start_transaction(root, 5);
10353         if (IS_ERR(trans))
10354                 return PTR_ERR(trans);
10355
10356         ret = btrfs_find_free_ino(root, &objectid);
10357         if (ret)
10358                 goto out;
10359
10360         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10361                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10362         if (IS_ERR(inode)) {
10363                 ret = PTR_ERR(inode);
10364                 inode = NULL;
10365                 goto out;
10366         }
10367
10368         inode->i_fop = &btrfs_file_operations;
10369         inode->i_op = &btrfs_file_inode_operations;
10370
10371         inode->i_mapping->a_ops = &btrfs_aops;
10372         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10373
10374         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10375         if (ret)
10376                 goto out;
10377
10378         ret = btrfs_update_inode(trans, root, inode);
10379         if (ret)
10380                 goto out;
10381         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10382         if (ret)
10383                 goto out;
10384
10385         /*
10386          * We set number of links to 0 in btrfs_new_inode(), and here we set
10387          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10388          * through:
10389          *
10390          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10391          */
10392         set_nlink(inode, 1);
10393         d_tmpfile(dentry, inode);
10394         unlock_new_inode(inode);
10395         mark_inode_dirty(inode);
10396 out:
10397         btrfs_end_transaction(trans);
10398         if (ret && inode)
10399                 discard_new_inode(inode);
10400         btrfs_btree_balance_dirty(fs_info);
10401         return ret;
10402 }
10403
10404 __attribute__((const))
10405 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10406 {
10407         return -EAGAIN;
10408 }
10409
10410 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10411                                         u64 start, u64 end)
10412 {
10413         struct inode *inode = private_data;
10414         u64 isize;
10415
10416         isize = i_size_read(inode);
10417         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10418                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10419                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10420                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10421         }
10422 }
10423
10424 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10425 {
10426         struct inode *inode = tree->private_data;
10427         unsigned long index = start >> PAGE_SHIFT;
10428         unsigned long end_index = end >> PAGE_SHIFT;
10429         struct page *page;
10430
10431         while (index <= end_index) {
10432                 page = find_get_page(inode->i_mapping, index);
10433                 ASSERT(page); /* Pages should be in the extent_io_tree */
10434                 set_page_writeback(page);
10435                 put_page(page);
10436                 index++;
10437         }
10438 }
10439
10440 static const struct inode_operations btrfs_dir_inode_operations = {
10441         .getattr        = btrfs_getattr,
10442         .lookup         = btrfs_lookup,
10443         .create         = btrfs_create,
10444         .unlink         = btrfs_unlink,
10445         .link           = btrfs_link,
10446         .mkdir          = btrfs_mkdir,
10447         .rmdir          = btrfs_rmdir,
10448         .rename         = btrfs_rename2,
10449         .symlink        = btrfs_symlink,
10450         .setattr        = btrfs_setattr,
10451         .mknod          = btrfs_mknod,
10452         .listxattr      = btrfs_listxattr,
10453         .permission     = btrfs_permission,
10454         .get_acl        = btrfs_get_acl,
10455         .set_acl        = btrfs_set_acl,
10456         .update_time    = btrfs_update_time,
10457         .tmpfile        = btrfs_tmpfile,
10458 };
10459 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10460         .lookup         = btrfs_lookup,
10461         .permission     = btrfs_permission,
10462         .update_time    = btrfs_update_time,
10463 };
10464
10465 static const struct file_operations btrfs_dir_file_operations = {
10466         .llseek         = generic_file_llseek,
10467         .read           = generic_read_dir,
10468         .iterate_shared = btrfs_real_readdir,
10469         .open           = btrfs_opendir,
10470         .unlocked_ioctl = btrfs_ioctl,
10471 #ifdef CONFIG_COMPAT
10472         .compat_ioctl   = btrfs_compat_ioctl,
10473 #endif
10474         .release        = btrfs_release_file,
10475         .fsync          = btrfs_sync_file,
10476 };
10477
10478 static const struct extent_io_ops btrfs_extent_io_ops = {
10479         /* mandatory callbacks */
10480         .submit_bio_hook = btrfs_submit_bio_hook,
10481         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10482         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10483
10484         /* optional callbacks */
10485         .fill_delalloc = run_delalloc_range,
10486         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10487         .writepage_start_hook = btrfs_writepage_start_hook,
10488         .set_bit_hook = btrfs_set_bit_hook,
10489         .clear_bit_hook = btrfs_clear_bit_hook,
10490         .merge_extent_hook = btrfs_merge_extent_hook,
10491         .split_extent_hook = btrfs_split_extent_hook,
10492         .check_extent_io_range = btrfs_check_extent_io_range,
10493 };
10494
10495 /*
10496  * btrfs doesn't support the bmap operation because swapfiles
10497  * use bmap to make a mapping of extents in the file.  They assume
10498  * these extents won't change over the life of the file and they
10499  * use the bmap result to do IO directly to the drive.
10500  *
10501  * the btrfs bmap call would return logical addresses that aren't
10502  * suitable for IO and they also will change frequently as COW
10503  * operations happen.  So, swapfile + btrfs == corruption.
10504  *
10505  * For now we're avoiding this by dropping bmap.
10506  */
10507 static const struct address_space_operations btrfs_aops = {
10508         .readpage       = btrfs_readpage,
10509         .writepage      = btrfs_writepage,
10510         .writepages     = btrfs_writepages,
10511         .readpages      = btrfs_readpages,
10512         .direct_IO      = btrfs_direct_IO,
10513         .invalidatepage = btrfs_invalidatepage,
10514         .releasepage    = btrfs_releasepage,
10515         .set_page_dirty = btrfs_set_page_dirty,
10516         .error_remove_page = generic_error_remove_page,
10517 };
10518
10519 static const struct address_space_operations btrfs_symlink_aops = {
10520         .readpage       = btrfs_readpage,
10521         .writepage      = btrfs_writepage,
10522         .invalidatepage = btrfs_invalidatepage,
10523         .releasepage    = btrfs_releasepage,
10524 };
10525
10526 static const struct inode_operations btrfs_file_inode_operations = {
10527         .getattr        = btrfs_getattr,
10528         .setattr        = btrfs_setattr,
10529         .listxattr      = btrfs_listxattr,
10530         .permission     = btrfs_permission,
10531         .fiemap         = btrfs_fiemap,
10532         .get_acl        = btrfs_get_acl,
10533         .set_acl        = btrfs_set_acl,
10534         .update_time    = btrfs_update_time,
10535 };
10536 static const struct inode_operations btrfs_special_inode_operations = {
10537         .getattr        = btrfs_getattr,
10538         .setattr        = btrfs_setattr,
10539         .permission     = btrfs_permission,
10540         .listxattr      = btrfs_listxattr,
10541         .get_acl        = btrfs_get_acl,
10542         .set_acl        = btrfs_set_acl,
10543         .update_time    = btrfs_update_time,
10544 };
10545 static const struct inode_operations btrfs_symlink_inode_operations = {
10546         .get_link       = page_get_link,
10547         .getattr        = btrfs_getattr,
10548         .setattr        = btrfs_setattr,
10549         .permission     = btrfs_permission,
10550         .listxattr      = btrfs_listxattr,
10551         .update_time    = btrfs_update_time,
10552 };
10553
10554 const struct dentry_operations btrfs_dentry_operations = {
10555         .d_delete       = btrfs_dentry_delete,
10556 };