fs/btrfs/inode.c

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