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/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include "ctree.h"
  46 #include "disk-io.h"
  47 #include "transaction.h"
  48 #include "btrfs_inode.h"
  49 #include "print-tree.h"
  50 #include "ordered-data.h"
  51 #include "xattr.h"
  52 #include "tree-log.h"
  53 #include "volumes.h"
  54 #include "compression.h"
  55 #include "locking.h"
  56 #include "free-space-cache.h"
  57 #include "inode-map.h"
  58 #include "backref.h"
  59 #include "hash.h"
  60 #include "props.h"
  61 #include "qgroup.h"
  62 #include "dedupe.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 struct btrfs_dio_data {
  70         u64 outstanding_extents;
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74 };
  75
  76 static const struct inode_operations btrfs_dir_inode_operations;
  77 static const struct inode_operations btrfs_symlink_inode_operations;
  78 static const struct inode_operations btrfs_dir_ro_inode_operations;
  79 static const struct inode_operations btrfs_special_inode_operations;
  80 static const struct inode_operations btrfs_file_inode_operations;
  81 static const struct address_space_operations btrfs_aops;
  82 static const struct address_space_operations btrfs_symlink_aops;
  83 static const struct file_operations btrfs_dir_file_operations;
  84 static const struct extent_io_ops btrfs_extent_io_ops;
  85
  86 static struct kmem_cache *btrfs_inode_cachep;
  87 struct kmem_cache *btrfs_trans_handle_cachep;
  88 struct kmem_cache *btrfs_transaction_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, u64 delalloc_end,
 109                                    int *page_started, unsigned long *nr_written,
 110                                    int unlock, struct btrfs_dedupe_hash *hash);
 111 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 112                                            u64 len, u64 orig_start,
 113                                            u64 block_start, u64 block_len,
 114                                            u64 orig_block_len, u64 ram_bytes,
 115                                            int type);
 116
 117 static int btrfs_dirty_inode(struct inode *inode);
 118
 119 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 120 void btrfs_test_inode_set_ops(struct inode *inode)
 121 {
 122         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 123 }
 124 #endif
 125
 126 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 127                                      struct inode *inode,  struct inode *dir,
 128                                      const struct qstr *qstr)
 129 {
 130         int err;
 131
 132         err = btrfs_init_acl(trans, inode, dir);
 133         if (!err)
 134                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 135         return err;
 136 }
 137
 138 /*
 139  * this does all the hard work for inserting an inline extent into
 140  * the btree.  The caller should have done a btrfs_drop_extents so that
 141  * no overlapping inline items exist in the btree
 142  */
 143 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 144                                 struct btrfs_path *path, int extent_inserted,
 145                                 struct btrfs_root *root, struct inode *inode,
 146                                 u64 start, size_t size, size_t compressed_size,
 147                                 int compress_type,
 148                                 struct page **compressed_pages)
 149 {
 150         struct extent_buffer *leaf;
 151         struct page *page = NULL;
 152         char *kaddr;
 153         unsigned long ptr;
 154         struct btrfs_file_extent_item *ei;
 155         int err = 0;
 156         int ret;
 157         size_t cur_size = size;
 158         unsigned long offset;
 159
 160         if (compressed_size && compressed_pages)
 161                 cur_size = compressed_size;
 162
 163         inode_add_bytes(inode, size);
 164
 165         if (!extent_inserted) {
 166                 struct btrfs_key key;
 167                 size_t datasize;
 168
 169                 key.objectid = btrfs_ino(inode);
 170                 key.offset = start;
 171                 key.type = BTRFS_EXTENT_DATA_KEY;
 172
 173                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 174                 path->leave_spinning = 1;
 175                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 176                                               datasize);
 177                 if (ret) {
 178                         err = ret;
 179                         goto fail;
 180                 }
 181         }
 182         leaf = path->nodes[0];
 183         ei = btrfs_item_ptr(leaf, path->slots[0],
 184                             struct btrfs_file_extent_item);
 185         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 186         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 187         btrfs_set_file_extent_encryption(leaf, ei, 0);
 188         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 189         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 190         ptr = btrfs_file_extent_inline_start(ei);
 191
 192         if (compress_type != BTRFS_COMPRESS_NONE) {
 193                 struct page *cpage;
 194                 int i = 0;
 195                 while (compressed_size > 0) {
 196                         cpage = compressed_pages[i];
 197                         cur_size = min_t(unsigned long, compressed_size,
 198                                        PAGE_SIZE);
 199
 200                         kaddr = kmap_atomic(cpage);
 201                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 202                         kunmap_atomic(kaddr);
 203
 204                         i++;
 205                         ptr += cur_size;
 206                         compressed_size -= cur_size;
 207                 }
 208                 btrfs_set_file_extent_compression(leaf, ei,
 209                                                   compress_type);
 210         } else {
 211                 page = find_get_page(inode->i_mapping,
 212                                      start >> PAGE_SHIFT);
 213                 btrfs_set_file_extent_compression(leaf, ei, 0);
 214                 kaddr = kmap_atomic(page);
 215                 offset = start & (PAGE_SIZE - 1);
 216                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 217                 kunmap_atomic(kaddr);
 218                 put_page(page);
 219         }
 220         btrfs_mark_buffer_dirty(leaf);
 221         btrfs_release_path(path);
 222
 223         /*
 224          * we're an inline extent, so nobody can
 225          * extend the file past i_size without locking
 226          * a page we already have locked.
 227          *
 228          * We must do any isize and inode updates
 229          * before we unlock the pages.  Otherwise we
 230          * could end up racing with unlink.
 231          */
 232         BTRFS_I(inode)->disk_i_size = inode->i_size;
 233         ret = btrfs_update_inode(trans, root, inode);
 234
 235         return ret;
 236 fail:
 237         return err;
 238 }
 239
 240
 241 /*
 242  * conditionally insert an inline extent into the file.  This
 243  * does the checks required to make sure the data is small enough
 244  * to fit as an inline extent.
 245  */
 246 static noinline int cow_file_range_inline(struct btrfs_root *root,
 247                                           struct inode *inode, u64 start,
 248                                           u64 end, size_t compressed_size,
 249                                           int compress_type,
 250                                           struct page **compressed_pages)
 251 {
 252         struct btrfs_fs_info *fs_info = root->fs_info;
 253         struct btrfs_trans_handle *trans;
 254         u64 isize = i_size_read(inode);
 255         u64 actual_end = min(end + 1, isize);
 256         u64 inline_len = actual_end - start;
 257         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 258         u64 data_len = inline_len;
 259         int ret;
 260         struct btrfs_path *path;
 261         int extent_inserted = 0;
 262         u32 extent_item_size;
 263
 264         if (compressed_size)
 265                 data_len = compressed_size;
 266
 267         if (start > 0 ||
 268             actual_end > fs_info->sectorsize ||
 269             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 270             (!compressed_size &&
 271             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 272             end + 1 < isize ||
 273             data_len > fs_info->max_inline) {
 274                 return 1;
 275         }
 276
 277         path = btrfs_alloc_path();
 278         if (!path)
 279                 return -ENOMEM;
 280
 281         trans = btrfs_join_transaction(root);
 282         if (IS_ERR(trans)) {
 283                 btrfs_free_path(path);
 284                 return PTR_ERR(trans);
 285         }
 286         trans->block_rsv = &fs_info->delalloc_block_rsv;
 287
 288         if (compressed_size && compressed_pages)
 289                 extent_item_size = btrfs_file_extent_calc_inline_size(
 290                    compressed_size);
 291         else
 292                 extent_item_size = btrfs_file_extent_calc_inline_size(
 293                     inline_len);
 294
 295         ret = __btrfs_drop_extents(trans, root, inode, path,
 296                                    start, aligned_end, NULL,
 297                                    1, 1, extent_item_size, &extent_inserted);
 298         if (ret) {
 299                 btrfs_abort_transaction(trans, ret);
 300                 goto out;
 301         }
 302
 303         if (isize > actual_end)
 304                 inline_len = min_t(u64, isize, actual_end);
 305         ret = insert_inline_extent(trans, path, extent_inserted,
 306                                    root, inode, start,
 307                                    inline_len, compressed_size,
 308                                    compress_type, compressed_pages);
 309         if (ret && ret != -ENOSPC) {
 310                 btrfs_abort_transaction(trans, ret);
 311                 goto out;
 312         } else if (ret == -ENOSPC) {
 313                 ret = 1;
 314                 goto out;
 315         }
 316
 317         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 318         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 319         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 320 out:
 321         /*
 322          * Don't forget to free the reserved space, as for inlined extent
 323          * it won't count as data extent, free them directly here.
 324          * And at reserve time, it's always aligned to page size, so
 325          * just free one page here.
 326          */
 327         btrfs_qgroup_free_data(inode, 0, PAGE_SIZE);
 328         btrfs_free_path(path);
 329         btrfs_end_transaction(trans);
 330         return ret;
 331 }
 332
 333 struct async_extent {
 334         u64 start;
 335         u64 ram_size;
 336         u64 compressed_size;
 337         struct page **pages;
 338         unsigned long nr_pages;
 339         int compress_type;
 340         struct list_head list;
 341 };
 342
 343 struct async_cow {
 344         struct inode *inode;
 345         struct btrfs_root *root;
 346         struct page *locked_page;
 347         u64 start;
 348         u64 end;
 349         struct list_head extents;
 350         struct btrfs_work work;
 351 };
 352
 353 static noinline int add_async_extent(struct async_cow *cow,
 354                                      u64 start, u64 ram_size,
 355                                      u64 compressed_size,
 356                                      struct page **pages,
 357                                      unsigned long nr_pages,
 358                                      int compress_type)
 359 {
 360         struct async_extent *async_extent;
 361
 362         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 363         BUG_ON(!async_extent); /* -ENOMEM */
 364         async_extent->start = start;
 365         async_extent->ram_size = ram_size;
 366         async_extent->compressed_size = compressed_size;
 367         async_extent->pages = pages;
 368         async_extent->nr_pages = nr_pages;
 369         async_extent->compress_type = compress_type;
 370         list_add_tail(&async_extent->list, &cow->extents);
 371         return 0;
 372 }
 373
 374 static inline int inode_need_compress(struct inode *inode)
 375 {
 376         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 377
 378         /* force compress */
 379         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 380                 return 1;
 381         /* bad compression ratios */
 382         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 383                 return 0;
 384         if (btrfs_test_opt(fs_info, COMPRESS) ||
 385             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 386             BTRFS_I(inode)->force_compress)
 387                 return 1;
 388         return 0;
 389 }
 390
 391 /*
 392  * we create compressed extents in two phases.  The first
 393  * phase compresses a range of pages that have already been
 394  * locked (both pages and state bits are locked).
 395  *
 396  * This is done inside an ordered work queue, and the compression
 397  * is spread across many cpus.  The actual IO submission is step
 398  * two, and the ordered work queue takes care of making sure that
 399  * happens in the same order things were put onto the queue by
 400  * writepages and friends.
 401  *
 402  * If this code finds it can't get good compression, it puts an
 403  * entry onto the work queue to write the uncompressed bytes.  This
 404  * makes sure that both compressed inodes and uncompressed inodes
 405  * are written in the same order that the flusher thread sent them
 406  * down.
 407  */
 408 static noinline void compress_file_range(struct inode *inode,
 409                                         struct page *locked_page,
 410                                         u64 start, u64 end,
 411                                         struct async_cow *async_cow,
 412                                         int *num_added)
 413 {
 414         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 415         struct btrfs_root *root = BTRFS_I(inode)->root;
 416         u64 num_bytes;
 417         u64 blocksize = fs_info->sectorsize;
 418         u64 actual_end;
 419         u64 isize = i_size_read(inode);
 420         int ret = 0;
 421         struct page **pages = NULL;
 422         unsigned long nr_pages;
 423         unsigned long nr_pages_ret = 0;
 424         unsigned long total_compressed = 0;
 425         unsigned long total_in = 0;
 426         unsigned long max_compressed = SZ_128K;
 427         unsigned long max_uncompressed = SZ_128K;
 428         int i;
 429         int will_compress;
 430         int compress_type = fs_info->compress_type;
 431         int redirty = 0;
 432
 433         /* if this is a small write inside eof, kick off a defrag */
 434         if ((end - start + 1) < SZ_16K &&
 435             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 436                 btrfs_add_inode_defrag(NULL, inode);
 437
 438         actual_end = min_t(u64, isize, end + 1);
 439 again:
 440         will_compress = 0;
 441         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 442         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_SIZE);
 443
 444         /*
 445          * we don't want to send crud past the end of i_size through
 446          * compression, that's just a waste of CPU time.  So, if the
 447          * end of the file is before the start of our current
 448          * requested range of bytes, we bail out to the uncompressed
 449          * cleanup code that can deal with all of this.
 450          *
 451          * It isn't really the fastest way to fix things, but this is a
 452          * very uncommon corner.
 453          */
 454         if (actual_end <= start)
 455                 goto cleanup_and_bail_uncompressed;
 456
 457         total_compressed = actual_end - start;
 458
 459         /*
 460          * skip compression for a small file range(<=blocksize) that
 461          * isn't an inline extent, since it doesn't save disk space at all.
 462          */
 463         if (total_compressed <= blocksize &&
 464            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 465                 goto cleanup_and_bail_uncompressed;
 466
 467         /* we want to make sure that amount of ram required to uncompress
 468          * an extent is reasonable, so we limit the total size in ram
 469          * of a compressed extent to 128k.  This is a crucial number
 470          * because it also controls how easily we can spread reads across
 471          * cpus for decompression.
 472          *
 473          * We also want to make sure the amount of IO required to do
 474          * a random read is reasonably small, so we limit the size of
 475          * a compressed extent to 128k.
 476          */
 477         total_compressed = min(total_compressed, max_uncompressed);
 478         num_bytes = ALIGN(end - start + 1, blocksize);
 479         num_bytes = max(blocksize,  num_bytes);
 480         total_in = 0;
 481         ret = 0;
 482
 483         /*
 484          * we do compression for mount -o compress and when the
 485          * inode has not been flagged as nocompress.  This flag can
 486          * change at any time if we discover bad compression ratios.
 487          */
 488         if (inode_need_compress(inode)) {
 489                 WARN_ON(pages);
 490                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 491                 if (!pages) {
 492                         /* just bail out to the uncompressed code */
 493                         goto cont;
 494                 }
 495
 496                 if (BTRFS_I(inode)->force_compress)
 497                         compress_type = BTRFS_I(inode)->force_compress;
 498
 499                 /*
 500                  * we need to call clear_page_dirty_for_io on each
 501                  * page in the range.  Otherwise applications with the file
 502                  * mmap'd can wander in and change the page contents while
 503                  * we are compressing them.
 504                  *
 505                  * If the compression fails for any reason, we set the pages
 506                  * dirty again later on.
 507                  */
 508                 extent_range_clear_dirty_for_io(inode, start, end);
 509                 redirty = 1;
 510                 ret = btrfs_compress_pages(compress_type,
 511                                            inode->i_mapping, start,
 512                                            total_compressed, pages,
 513                                            nr_pages, &nr_pages_ret,
 514                                            &total_in,
 515                                            &total_compressed,
 516                                            max_compressed);
 517
 518                 if (!ret) {
 519                         unsigned long offset = total_compressed &
 520                                 (PAGE_SIZE - 1);
 521                         struct page *page = pages[nr_pages_ret - 1];
 522                         char *kaddr;
 523
 524                         /* zero the tail end of the last page, we might be
 525                          * sending it down to disk
 526                          */
 527                         if (offset) {
 528                                 kaddr = kmap_atomic(page);
 529                                 memset(kaddr + offset, 0,
 530                                        PAGE_SIZE - offset);
 531                                 kunmap_atomic(kaddr);
 532                         }
 533                         will_compress = 1;
 534                 }
 535         }
 536 cont:
 537         if (start == 0) {
 538                 /* lets try to make an inline extent */
 539                 if (ret || total_in < (actual_end - start)) {
 540                         /* we didn't compress the entire range, try
 541                          * to make an uncompressed inline extent.
 542                          */
 543                         ret = cow_file_range_inline(root, inode, start, end,
 544                                                     0, 0, NULL);
 545                 } else {
 546                         /* try making a compressed inline extent */
 547                         ret = cow_file_range_inline(root, inode, start, end,
 548                                                     total_compressed,
 549                                                     compress_type, pages);
 550                 }
 551                 if (ret <= 0) {
 552                         unsigned long clear_flags = EXTENT_DELALLOC |
 553                                 EXTENT_DEFRAG;
 554                         unsigned long page_error_op;
 555
 556                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 557                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 558
 559                         /*
 560                          * inline extent creation worked or returned error,
 561                          * we don't need to create any more async work items.
 562                          * Unlock and free up our temp pages.
 563                          */
 564                         extent_clear_unlock_delalloc(inode, start, end, end,
 565                                                      NULL, clear_flags,
 566                                                      PAGE_UNLOCK |
 567                                                      PAGE_CLEAR_DIRTY |
 568                                                      PAGE_SET_WRITEBACK |
 569                                                      page_error_op |
 570                                                      PAGE_END_WRITEBACK);
 571                         btrfs_free_reserved_data_space_noquota(inode, start,
 572                                                 end - start + 1);
 573                         goto free_pages_out;
 574                 }
 575         }
 576
 577         if (will_compress) {
 578                 /*
 579                  * we aren't doing an inline extent round the compressed size
 580                  * up to a block size boundary so the allocator does sane
 581                  * things
 582                  */
 583                 total_compressed = ALIGN(total_compressed, blocksize);
 584
 585                 /*
 586                  * one last check to make sure the compression is really a
 587                  * win, compare the page count read with the blocks on disk
 588                  */
 589                 total_in = ALIGN(total_in, PAGE_SIZE);
 590                 if (total_compressed >= total_in) {
 591                         will_compress = 0;
 592                 } else {
 593                         num_bytes = total_in;
 594                         *num_added += 1;
 595
 596                         /*
 597                          * The async work queues will take care of doing actual
 598                          * allocation on disk for these compressed pages, and
 599                          * will submit them to the elevator.
 600                          */
 601                         add_async_extent(async_cow, start, num_bytes,
 602                                         total_compressed, pages, nr_pages_ret,
 603                                         compress_type);
 604
 605                         if (start + num_bytes < end) {
 606                                 start += num_bytes;
 607                                 pages = NULL;
 608                                 cond_resched();
 609                                 goto again;
 610                         }
 611                         return;
 612                 }
 613         }
 614         if (pages) {
 615                 /*
 616                  * the compression code ran but failed to make things smaller,
 617                  * free any pages it allocated and our page pointer array
 618                  */
 619                 for (i = 0; i < nr_pages_ret; i++) {
 620                         WARN_ON(pages[i]->mapping);
 621                         put_page(pages[i]);
 622                 }
 623                 kfree(pages);
 624                 pages = NULL;
 625                 total_compressed = 0;
 626                 nr_pages_ret = 0;
 627
 628                 /* flag the file so we don't compress in the future */
 629                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 630                     !(BTRFS_I(inode)->force_compress)) {
 631                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 632                 }
 633         }
 634 cleanup_and_bail_uncompressed:
 635         /*
 636          * No compression, but we still need to write the pages in the file
 637          * we've been given so far.  redirty the locked page if it corresponds
 638          * to our extent and set things up for the async work queue to run
 639          * cow_file_range to do the normal delalloc dance.
 640          */
 641         if (page_offset(locked_page) >= start &&
 642             page_offset(locked_page) <= end)
 643                 __set_page_dirty_nobuffers(locked_page);
 644                 /* unlocked later on in the async handlers */
 645
 646         if (redirty)
 647                 extent_range_redirty_for_io(inode, start, end);
 648         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 649                          BTRFS_COMPRESS_NONE);
 650         *num_added += 1;
 651
 652         return;
 653
 654 free_pages_out:
 655         for (i = 0; i < nr_pages_ret; i++) {
 656                 WARN_ON(pages[i]->mapping);
 657                 put_page(pages[i]);
 658         }
 659         kfree(pages);
 660 }
 661
 662 static void free_async_extent_pages(struct async_extent *async_extent)
 663 {
 664         int i;
 665
 666         if (!async_extent->pages)
 667                 return;
 668
 669         for (i = 0; i < async_extent->nr_pages; i++) {
 670                 WARN_ON(async_extent->pages[i]->mapping);
 671                 put_page(async_extent->pages[i]);
 672         }
 673         kfree(async_extent->pages);
 674         async_extent->nr_pages = 0;
 675         async_extent->pages = NULL;
 676 }
 677
 678 /*
 679  * phase two of compressed writeback.  This is the ordered portion
 680  * of the code, which only gets called in the order the work was
 681  * queued.  We walk all the async extents created by compress_file_range
 682  * and send them down to the disk.
 683  */
 684 static noinline void submit_compressed_extents(struct inode *inode,
 685                                               struct async_cow *async_cow)
 686 {
 687         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 688         struct async_extent *async_extent;
 689         u64 alloc_hint = 0;
 690         struct btrfs_key ins;
 691         struct extent_map *em;
 692         struct btrfs_root *root = BTRFS_I(inode)->root;
 693         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 694         struct extent_io_tree *io_tree;
 695         int ret = 0;
 696
 697 again:
 698         while (!list_empty(&async_cow->extents)) {
 699                 async_extent = list_entry(async_cow->extents.next,
 700                                           struct async_extent, list);
 701                 list_del(&async_extent->list);
 702
 703                 io_tree = &BTRFS_I(inode)->io_tree;
 704
 705 retry:
 706                 /* did the compression code fall back to uncompressed IO? */
 707                 if (!async_extent->pages) {
 708                         int page_started = 0;
 709                         unsigned long nr_written = 0;
 710
 711                         lock_extent(io_tree, async_extent->start,
 712                                          async_extent->start +
 713                                          async_extent->ram_size - 1);
 714
 715                         /* allocate blocks */
 716                         ret = cow_file_range(inode, async_cow->locked_page,
 717                                              async_extent->start,
 718                                              async_extent->start +
 719                                              async_extent->ram_size - 1,
 720                                              async_extent->start +
 721                                              async_extent->ram_size - 1,
 722                                              &page_started, &nr_written, 0,
 723                                              NULL);
 724
 725                         /* JDM XXX */
 726
 727                         /*
 728                          * if page_started, cow_file_range inserted an
 729                          * inline extent and took care of all the unlocking
 730                          * and IO for us.  Otherwise, we need to submit
 731                          * all those pages down to the drive.
 732                          */
 733                         if (!page_started && !ret)
 734                                 extent_write_locked_range(io_tree,
 735                                                   inode, async_extent->start,
 736                                                   async_extent->start +
 737                                                   async_extent->ram_size - 1,
 738                                                   btrfs_get_extent,
 739                                                   WB_SYNC_ALL);
 740                         else if (ret)
 741                                 unlock_page(async_cow->locked_page);
 742                         kfree(async_extent);
 743                         cond_resched();
 744                         continue;
 745                 }
 746
 747                 lock_extent(io_tree, async_extent->start,
 748                             async_extent->start + async_extent->ram_size - 1);
 749
 750                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 751                                            async_extent->compressed_size,
 752                                            async_extent->compressed_size,
 753                                            0, alloc_hint, &ins, 1, 1);
 754                 if (ret) {
 755                         free_async_extent_pages(async_extent);
 756
 757                         if (ret == -ENOSPC) {
 758                                 unlock_extent(io_tree, async_extent->start,
 759                                               async_extent->start +
 760                                               async_extent->ram_size - 1);
 761
 762                                 /*
 763                                  * we need to redirty the pages if we decide to
 764                                  * fallback to uncompressed IO, otherwise we
 765                                  * will not submit these pages down to lower
 766                                  * layers.
 767                                  */
 768                                 extent_range_redirty_for_io(inode,
 769                                                 async_extent->start,
 770                                                 async_extent->start +
 771                                                 async_extent->ram_size - 1);
 772
 773                                 goto retry;
 774                         }
 775                         goto out_free;
 776                 }
 777                 /*
 778                  * here we're doing allocation and writeback of the
 779                  * compressed pages
 780                  */
 781                 btrfs_drop_extent_cache(inode, async_extent->start,
 782                                         async_extent->start +
 783                                         async_extent->ram_size - 1, 0);
 784
 785                 em = alloc_extent_map();
 786                 if (!em) {
 787                         ret = -ENOMEM;
 788                         goto out_free_reserve;
 789                 }
 790                 em->start = async_extent->start;
 791                 em->len = async_extent->ram_size;
 792                 em->orig_start = em->start;
 793                 em->mod_start = em->start;
 794                 em->mod_len = em->len;
 795
 796                 em->block_start = ins.objectid;
 797                 em->block_len = ins.offset;
 798                 em->orig_block_len = ins.offset;
 799                 em->ram_bytes = async_extent->ram_size;
 800                 em->bdev = fs_info->fs_devices->latest_bdev;
 801                 em->compress_type = async_extent->compress_type;
 802                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 803                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 804                 em->generation = -1;
 805
 806                 while (1) {
 807                         write_lock(&em_tree->lock);
 808                         ret = add_extent_mapping(em_tree, em, 1);
 809                         write_unlock(&em_tree->lock);
 810                         if (ret != -EEXIST) {
 811                                 free_extent_map(em);
 812                                 break;
 813                         }
 814                         btrfs_drop_extent_cache(inode, async_extent->start,
 815                                                 async_extent->start +
 816                                                 async_extent->ram_size - 1, 0);
 817                 }
 818
 819                 if (ret)
 820                         goto out_free_reserve;
 821
 822                 ret = btrfs_add_ordered_extent_compress(inode,
 823                                                 async_extent->start,
 824                                                 ins.objectid,
 825                                                 async_extent->ram_size,
 826                                                 ins.offset,
 827                                                 BTRFS_ORDERED_COMPRESSED,
 828                                                 async_extent->compress_type);
 829                 if (ret) {
 830                         btrfs_drop_extent_cache(inode, async_extent->start,
 831                                                 async_extent->start +
 832                                                 async_extent->ram_size - 1, 0);
 833                         goto out_free_reserve;
 834                 }
 835                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 836
 837                 /*
 838                  * clear dirty, set writeback and unlock the pages.
 839                  */
 840                 extent_clear_unlock_delalloc(inode, async_extent->start,
 841                                 async_extent->start +
 842                                 async_extent->ram_size - 1,
 843                                 async_extent->start +
 844                                 async_extent->ram_size - 1,
 845                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 846                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 847                                 PAGE_SET_WRITEBACK);
 848                 ret = btrfs_submit_compressed_write(inode,
 849                                     async_extent->start,
 850                                     async_extent->ram_size,
 851                                     ins.objectid,
 852                                     ins.offset, async_extent->pages,
 853                                     async_extent->nr_pages);
 854                 if (ret) {
 855                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 856                         struct page *p = async_extent->pages[0];
 857                         const u64 start = async_extent->start;
 858                         const u64 end = start + async_extent->ram_size - 1;
 859
 860                         p->mapping = inode->i_mapping;
 861                         tree->ops->writepage_end_io_hook(p, start, end,
 862                                                          NULL, 0);
 863                         p->mapping = NULL;
 864                         extent_clear_unlock_delalloc(inode, start, end, end,
 865                                                      NULL, 0,
 866                                                      PAGE_END_WRITEBACK |
 867                                                      PAGE_SET_ERROR);
 868                         free_async_extent_pages(async_extent);
 869                 }
 870                 alloc_hint = ins.objectid + ins.offset;
 871                 kfree(async_extent);
 872                 cond_resched();
 873         }
 874         return;
 875 out_free_reserve:
 876         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 877         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 878 out_free:
 879         extent_clear_unlock_delalloc(inode, async_extent->start,
 880                                      async_extent->start +
 881                                      async_extent->ram_size - 1,
 882                                      async_extent->start +
 883                                      async_extent->ram_size - 1,
 884                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 885                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 886                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 887                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 888                                      PAGE_SET_ERROR);
 889         free_async_extent_pages(async_extent);
 890         kfree(async_extent);
 891         goto again;
 892 }
 893
 894 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 895                                       u64 num_bytes)
 896 {
 897         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 898         struct extent_map *em;
 899         u64 alloc_hint = 0;
 900
 901         read_lock(&em_tree->lock);
 902         em = search_extent_mapping(em_tree, start, num_bytes);
 903         if (em) {
 904                 /*
 905                  * if block start isn't an actual block number then find the
 906                  * first block in this inode and use that as a hint.  If that
 907                  * block is also bogus then just don't worry about it.
 908                  */
 909                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 910                         free_extent_map(em);
 911                         em = search_extent_mapping(em_tree, 0, 0);
 912                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 913                                 alloc_hint = em->block_start;
 914                         if (em)
 915                                 free_extent_map(em);
 916                 } else {
 917                         alloc_hint = em->block_start;
 918                         free_extent_map(em);
 919                 }
 920         }
 921         read_unlock(&em_tree->lock);
 922
 923         return alloc_hint;
 924 }
 925
 926 /*
 927  * when extent_io.c finds a delayed allocation range in the file,
 928  * the call backs end up in this code.  The basic idea is to
 929  * allocate extents on disk for the range, and create ordered data structs
 930  * in ram to track those extents.
 931  *
 932  * locked_page is the page that writepage had locked already.  We use
 933  * it to make sure we don't do extra locks or unlocks.
 934  *
 935  * *page_started is set to one if we unlock locked_page and do everything
 936  * required to start IO on it.  It may be clean and already done with
 937  * IO when we return.
 938  */
 939 static noinline int cow_file_range(struct inode *inode,
 940                                    struct page *locked_page,
 941                                    u64 start, u64 end, u64 delalloc_end,
 942                                    int *page_started, unsigned long *nr_written,
 943                                    int unlock, struct btrfs_dedupe_hash *hash)
 944 {
 945         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 946         struct btrfs_root *root = BTRFS_I(inode)->root;
 947         u64 alloc_hint = 0;
 948         u64 num_bytes;
 949         unsigned long ram_size;
 950         u64 disk_num_bytes;
 951         u64 cur_alloc_size;
 952         u64 blocksize = fs_info->sectorsize;
 953         struct btrfs_key ins;
 954         struct extent_map *em;
 955         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 956         int ret = 0;
 957
 958         if (btrfs_is_free_space_inode(inode)) {
 959                 WARN_ON_ONCE(1);
 960                 ret = -EINVAL;
 961                 goto out_unlock;
 962         }
 963
 964         num_bytes = ALIGN(end - start + 1, blocksize);
 965         num_bytes = max(blocksize,  num_bytes);
 966         disk_num_bytes = num_bytes;
 967
 968         /* if this is a small write inside eof, kick off defrag */
 969         if (num_bytes < SZ_64K &&
 970             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 971                 btrfs_add_inode_defrag(NULL, inode);
 972
 973         if (start == 0) {
 974                 /* lets try to make an inline extent */
 975                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 976                                             NULL);
 977                 if (ret == 0) {
 978                         extent_clear_unlock_delalloc(inode, start, end,
 979                                      delalloc_end, NULL,
 980                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 981                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 982                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 983                                      PAGE_END_WRITEBACK);
 984                         btrfs_free_reserved_data_space_noquota(inode, start,
 985                                                 end - start + 1);
 986                         *nr_written = *nr_written +
 987                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 988                         *page_started = 1;
 989                         goto out;
 990                 } else if (ret < 0) {
 991                         goto out_unlock;
 992                 }
 993         }
 994
 995         BUG_ON(disk_num_bytes >
 996                btrfs_super_total_bytes(fs_info->super_copy));
 997
 998         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 999         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
1000
1001         while (disk_num_bytes > 0) {
1002                 unsigned long op;
1003
1004                 cur_alloc_size = disk_num_bytes;
1005                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1006                                            fs_info->sectorsize, 0, alloc_hint,
1007                                            &ins, 1, 1);
1008                 if (ret < 0)
1009                         goto out_unlock;
1010
1011                 em = alloc_extent_map();
1012                 if (!em) {
1013                         ret = -ENOMEM;
1014                         goto out_reserve;
1015                 }
1016                 em->start = start;
1017                 em->orig_start = em->start;
1018                 ram_size = ins.offset;
1019                 em->len = ins.offset;
1020                 em->mod_start = em->start;
1021                 em->mod_len = em->len;
1022
1023                 em->block_start = ins.objectid;
1024                 em->block_len = ins.offset;
1025                 em->orig_block_len = ins.offset;
1026                 em->ram_bytes = ram_size;
1027                 em->bdev = fs_info->fs_devices->latest_bdev;
1028                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1029                 em->generation = -1;
1030
1031                 while (1) {
1032                         write_lock(&em_tree->lock);
1033                         ret = add_extent_mapping(em_tree, em, 1);
1034                         write_unlock(&em_tree->lock);
1035                         if (ret != -EEXIST) {
1036                                 free_extent_map(em);
1037                                 break;
1038                         }
1039                         btrfs_drop_extent_cache(inode, start,
1040                                                 start + ram_size - 1, 0);
1041                 }
1042                 if (ret)
1043                         goto out_reserve;
1044
1045                 cur_alloc_size = ins.offset;
1046                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1047                                                ram_size, cur_alloc_size, 0);
1048                 if (ret)
1049                         goto out_drop_extent_cache;
1050
1051                 if (root->root_key.objectid ==
1052                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1053                         ret = btrfs_reloc_clone_csums(inode, start,
1054                                                       cur_alloc_size);
1055                         if (ret)
1056                                 goto out_drop_extent_cache;
1057                 }
1058
1059                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1060
1061                 if (disk_num_bytes < cur_alloc_size)
1062                         break;
1063
1064                 /* we're not doing compressed IO, don't unlock the first
1065                  * page (which the caller expects to stay locked), don't
1066                  * clear any dirty bits and don't set any writeback bits
1067                  *
1068                  * Do set the Private2 bit so we know this page was properly
1069                  * setup for writepage
1070                  */
1071                 op = unlock ? PAGE_UNLOCK : 0;
1072                 op |= PAGE_SET_PRIVATE2;
1073
1074                 extent_clear_unlock_delalloc(inode, start,
1075                                              start + ram_size - 1,
1076                                              delalloc_end, locked_page,
1077                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1078                                              op);
1079                 disk_num_bytes -= cur_alloc_size;
1080                 num_bytes -= cur_alloc_size;
1081                 alloc_hint = ins.objectid + ins.offset;
1082                 start += cur_alloc_size;
1083         }
1084 out:
1085         return ret;
1086
1087 out_drop_extent_cache:
1088         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1089 out_reserve:
1090         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1091         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1092 out_unlock:
1093         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1094                                      locked_page,
1095                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1096                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1097                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1098                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1099         goto out;
1100 }
1101
1102 /*
1103  * work queue call back to started compression on a file and pages
1104  */
1105 static noinline void async_cow_start(struct btrfs_work *work)
1106 {
1107         struct async_cow *async_cow;
1108         int num_added = 0;
1109         async_cow = container_of(work, struct async_cow, work);
1110
1111         compress_file_range(async_cow->inode, async_cow->locked_page,
1112                             async_cow->start, async_cow->end, async_cow,
1113                             &num_added);
1114         if (num_added == 0) {
1115                 btrfs_add_delayed_iput(async_cow->inode);
1116                 async_cow->inode = NULL;
1117         }
1118 }
1119
1120 /*
1121  * work queue call back to submit previously compressed pages
1122  */
1123 static noinline void async_cow_submit(struct btrfs_work *work)
1124 {
1125         struct btrfs_fs_info *fs_info;
1126         struct async_cow *async_cow;
1127         struct btrfs_root *root;
1128         unsigned long nr_pages;
1129
1130         async_cow = container_of(work, struct async_cow, work);
1131
1132         root = async_cow->root;
1133         fs_info = root->fs_info;
1134         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1135                 PAGE_SHIFT;
1136
1137         /*
1138          * atomic_sub_return implies a barrier for waitqueue_active
1139          */
1140         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1141             5 * SZ_1M &&
1142             waitqueue_active(&fs_info->async_submit_wait))
1143                 wake_up(&fs_info->async_submit_wait);
1144
1145         if (async_cow->inode)
1146                 submit_compressed_extents(async_cow->inode, async_cow);
1147 }
1148
1149 static noinline void async_cow_free(struct btrfs_work *work)
1150 {
1151         struct async_cow *async_cow;
1152         async_cow = container_of(work, struct async_cow, work);
1153         if (async_cow->inode)
1154                 btrfs_add_delayed_iput(async_cow->inode);
1155         kfree(async_cow);
1156 }
1157
1158 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1159                                 u64 start, u64 end, int *page_started,
1160                                 unsigned long *nr_written)
1161 {
1162         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1163         struct async_cow *async_cow;
1164         struct btrfs_root *root = BTRFS_I(inode)->root;
1165         unsigned long nr_pages;
1166         u64 cur_end;
1167         int limit = 10 * SZ_1M;
1168
1169         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1170                          1, 0, NULL, GFP_NOFS);
1171         while (start < end) {
1172                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1173                 BUG_ON(!async_cow); /* -ENOMEM */
1174                 async_cow->inode = igrab(inode);
1175                 async_cow->root = root;
1176                 async_cow->locked_page = locked_page;
1177                 async_cow->start = start;
1178
1179                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1180                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1181                         cur_end = end;
1182                 else
1183                         cur_end = min(end, start + SZ_512K - 1);
1184
1185                 async_cow->end = cur_end;
1186                 INIT_LIST_HEAD(&async_cow->extents);
1187
1188                 btrfs_init_work(&async_cow->work,
1189                                 btrfs_delalloc_helper,
1190                                 async_cow_start, async_cow_submit,
1191                                 async_cow_free);
1192
1193                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1194                         PAGE_SHIFT;
1195                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1196
1197                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1198
1199                 if (atomic_read(&fs_info->async_delalloc_pages) > limit) {
1200                         wait_event(fs_info->async_submit_wait,
1201                                    (atomic_read(&fs_info->async_delalloc_pages) <
1202                                     limit));
1203                 }
1204
1205                 while (atomic_read(&fs_info->async_submit_draining) &&
1206                        atomic_read(&fs_info->async_delalloc_pages)) {
1207                         wait_event(fs_info->async_submit_wait,
1208                                    (atomic_read(&fs_info->async_delalloc_pages) ==
1209                                     0));
1210                 }
1211
1212                 *nr_written += nr_pages;
1213                 start = cur_end + 1;
1214         }
1215         *page_started = 1;
1216         return 0;
1217 }
1218
1219 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1220                                         u64 bytenr, u64 num_bytes)
1221 {
1222         int ret;
1223         struct btrfs_ordered_sum *sums;
1224         LIST_HEAD(list);
1225
1226         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1227                                        bytenr + num_bytes - 1, &list, 0);
1228         if (ret == 0 && list_empty(&list))
1229                 return 0;
1230
1231         while (!list_empty(&list)) {
1232                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1233                 list_del(&sums->list);
1234                 kfree(sums);
1235         }
1236         return 1;
1237 }
1238
1239 /*
1240  * when nowcow writeback call back.  This checks for snapshots or COW copies
1241  * of the extents that exist in the file, and COWs the file as required.
1242  *
1243  * If no cow copies or snapshots exist, we write directly to the existing
1244  * blocks on disk
1245  */
1246 static noinline int run_delalloc_nocow(struct inode *inode,
1247                                        struct page *locked_page,
1248                               u64 start, u64 end, int *page_started, int force,
1249                               unsigned long *nr_written)
1250 {
1251         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1252         struct btrfs_root *root = BTRFS_I(inode)->root;
1253         struct btrfs_trans_handle *trans;
1254         struct extent_buffer *leaf;
1255         struct btrfs_path *path;
1256         struct btrfs_file_extent_item *fi;
1257         struct btrfs_key found_key;
1258         u64 cow_start;
1259         u64 cur_offset;
1260         u64 extent_end;
1261         u64 extent_offset;
1262         u64 disk_bytenr;
1263         u64 num_bytes;
1264         u64 disk_num_bytes;
1265         u64 ram_bytes;
1266         int extent_type;
1267         int ret, err;
1268         int type;
1269         int nocow;
1270         int check_prev = 1;
1271         bool nolock;
1272         u64 ino = btrfs_ino(inode);
1273
1274         path = btrfs_alloc_path();
1275         if (!path) {
1276                 extent_clear_unlock_delalloc(inode, start, end, end,
1277                                              locked_page,
1278                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1279                                              EXTENT_DO_ACCOUNTING |
1280                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1281                                              PAGE_CLEAR_DIRTY |
1282                                              PAGE_SET_WRITEBACK |
1283                                              PAGE_END_WRITEBACK);
1284                 return -ENOMEM;
1285         }
1286
1287         nolock = btrfs_is_free_space_inode(inode);
1288
1289         if (nolock)
1290                 trans = btrfs_join_transaction_nolock(root);
1291         else
1292                 trans = btrfs_join_transaction(root);
1293
1294         if (IS_ERR(trans)) {
1295                 extent_clear_unlock_delalloc(inode, start, end, end,
1296                                              locked_page,
1297                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1298                                              EXTENT_DO_ACCOUNTING |
1299                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1300                                              PAGE_CLEAR_DIRTY |
1301                                              PAGE_SET_WRITEBACK |
1302                                              PAGE_END_WRITEBACK);
1303                 btrfs_free_path(path);
1304                 return PTR_ERR(trans);
1305         }
1306
1307         trans->block_rsv = &fs_info->delalloc_block_rsv;
1308
1309         cow_start = (u64)-1;
1310         cur_offset = start;
1311         while (1) {
1312                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1313                                                cur_offset, 0);
1314                 if (ret < 0)
1315                         goto error;
1316                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1317                         leaf = path->nodes[0];
1318                         btrfs_item_key_to_cpu(leaf, &found_key,
1319                                               path->slots[0] - 1);
1320                         if (found_key.objectid == ino &&
1321                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1322                                 path->slots[0]--;
1323                 }
1324                 check_prev = 0;
1325 next_slot:
1326                 leaf = path->nodes[0];
1327                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1328                         ret = btrfs_next_leaf(root, path);
1329                         if (ret < 0)
1330                                 goto error;
1331                         if (ret > 0)
1332                                 break;
1333                         leaf = path->nodes[0];
1334                 }
1335
1336                 nocow = 0;
1337                 disk_bytenr = 0;
1338                 num_bytes = 0;
1339                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1340
1341                 if (found_key.objectid > ino)
1342                         break;
1343                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1344                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1345                         path->slots[0]++;
1346                         goto next_slot;
1347                 }
1348                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1349                     found_key.offset > end)
1350                         break;
1351
1352                 if (found_key.offset > cur_offset) {
1353                         extent_end = found_key.offset;
1354                         extent_type = 0;
1355                         goto out_check;
1356                 }
1357
1358                 fi = btrfs_item_ptr(leaf, path->slots[0],
1359                                     struct btrfs_file_extent_item);
1360                 extent_type = btrfs_file_extent_type(leaf, fi);
1361
1362                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1363                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1364                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1365                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1366                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1367                         extent_end = found_key.offset +
1368                                 btrfs_file_extent_num_bytes(leaf, fi);
1369                         disk_num_bytes =
1370                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1371                         if (extent_end <= start) {
1372                                 path->slots[0]++;
1373                                 goto next_slot;
1374                         }
1375                         if (disk_bytenr == 0)
1376                                 goto out_check;
1377                         if (btrfs_file_extent_compression(leaf, fi) ||
1378                             btrfs_file_extent_encryption(leaf, fi) ||
1379                             btrfs_file_extent_other_encoding(leaf, fi))
1380                                 goto out_check;
1381                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1382                                 goto out_check;
1383                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1384                                 goto out_check;
1385                         if (btrfs_cross_ref_exist(trans, root, ino,
1386                                                   found_key.offset -
1387                                                   extent_offset, disk_bytenr))
1388                                 goto out_check;
1389                         disk_bytenr += extent_offset;
1390                         disk_bytenr += cur_offset - found_key.offset;
1391                         num_bytes = min(end + 1, extent_end) - cur_offset;
1392                         /*
1393                          * if there are pending snapshots for this root,
1394                          * we fall into common COW way.
1395                          */
1396                         if (!nolock) {
1397                                 err = btrfs_start_write_no_snapshoting(root);
1398                                 if (!err)
1399                                         goto out_check;
1400                         }
1401                         /*
1402                          * force cow if csum exists in the range.
1403                          * this ensure that csum for a given extent are
1404                          * either valid or do not exist.
1405                          */
1406                         if (csum_exist_in_range(fs_info, disk_bytenr,
1407                                                 num_bytes))
1408                                 goto out_check;
1409                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1410                                 goto out_check;
1411                         nocow = 1;
1412                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1413                         extent_end = found_key.offset +
1414                                 btrfs_file_extent_inline_len(leaf,
1415                                                      path->slots[0], fi);
1416                         extent_end = ALIGN(extent_end,
1417                                            fs_info->sectorsize);
1418                 } else {
1419                         BUG_ON(1);
1420                 }
1421 out_check:
1422                 if (extent_end <= start) {
1423                         path->slots[0]++;
1424                         if (!nolock && nocow)
1425                                 btrfs_end_write_no_snapshoting(root);
1426                         if (nocow)
1427                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1428                         goto next_slot;
1429                 }
1430                 if (!nocow) {
1431                         if (cow_start == (u64)-1)
1432                                 cow_start = cur_offset;
1433                         cur_offset = extent_end;
1434                         if (cur_offset > end)
1435                                 break;
1436                         path->slots[0]++;
1437                         goto next_slot;
1438                 }
1439
1440                 btrfs_release_path(path);
1441                 if (cow_start != (u64)-1) {
1442                         ret = cow_file_range(inode, locked_page,
1443                                              cow_start, found_key.offset - 1,
1444                                              end, page_started, nr_written, 1,
1445                                              NULL);
1446                         if (ret) {
1447                                 if (!nolock && nocow)
1448                                         btrfs_end_write_no_snapshoting(root);
1449                                 if (nocow)
1450                                         btrfs_dec_nocow_writers(fs_info,
1451                                                                 disk_bytenr);
1452                                 goto error;
1453                         }
1454                         cow_start = (u64)-1;
1455                 }
1456
1457                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1458                         struct extent_map *em;
1459                         struct extent_map_tree *em_tree;
1460                         em_tree = &BTRFS_I(inode)->extent_tree;
1461                         em = alloc_extent_map();
1462                         BUG_ON(!em); /* -ENOMEM */
1463                         em->start = cur_offset;
1464                         em->orig_start = found_key.offset - extent_offset;
1465                         em->len = num_bytes;
1466                         em->block_len = num_bytes;
1467                         em->block_start = disk_bytenr;
1468                         em->orig_block_len = disk_num_bytes;
1469                         em->ram_bytes = ram_bytes;
1470                         em->bdev = fs_info->fs_devices->latest_bdev;
1471                         em->mod_start = em->start;
1472                         em->mod_len = em->len;
1473                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1474                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1475                         em->generation = -1;
1476                         while (1) {
1477                                 write_lock(&em_tree->lock);
1478                                 ret = add_extent_mapping(em_tree, em, 1);
1479                                 write_unlock(&em_tree->lock);
1480                                 if (ret != -EEXIST) {
1481                                         free_extent_map(em);
1482                                         break;
1483                                 }
1484                                 btrfs_drop_extent_cache(inode, em->start,
1485                                                 em->start + em->len - 1, 0);
1486                         }
1487                         type = BTRFS_ORDERED_PREALLOC;
1488                 } else {
1489                         type = BTRFS_ORDERED_NOCOW;
1490                 }
1491
1492                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1493                                                num_bytes, num_bytes, type);
1494                 if (nocow)
1495                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1496                 BUG_ON(ret); /* -ENOMEM */
1497
1498                 if (root->root_key.objectid ==
1499                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1500                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1501                                                       num_bytes);
1502                         if (ret) {
1503                                 if (!nolock && nocow)
1504                                         btrfs_end_write_no_snapshoting(root);
1505                                 goto error;
1506                         }
1507                 }
1508
1509                 extent_clear_unlock_delalloc(inode, cur_offset,
1510                                              cur_offset + num_bytes - 1, end,
1511                                              locked_page, EXTENT_LOCKED |
1512                                              EXTENT_DELALLOC |
1513                                              EXTENT_CLEAR_DATA_RESV,
1514                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1515
1516                 if (!nolock && nocow)
1517                         btrfs_end_write_no_snapshoting(root);
1518                 cur_offset = extent_end;
1519                 if (cur_offset > end)
1520                         break;
1521         }
1522         btrfs_release_path(path);
1523
1524         if (cur_offset <= end && cow_start == (u64)-1) {
1525                 cow_start = cur_offset;
1526                 cur_offset = end;
1527         }
1528
1529         if (cow_start != (u64)-1) {
1530                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1531                                      page_started, nr_written, 1, NULL);
1532                 if (ret)
1533                         goto error;
1534         }
1535
1536 error:
1537         err = btrfs_end_transaction(trans);
1538         if (!ret)
1539                 ret = err;
1540
1541         if (ret && cur_offset < end)
1542                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1543                                              locked_page, EXTENT_LOCKED |
1544                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1545                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1546                                              PAGE_CLEAR_DIRTY |
1547                                              PAGE_SET_WRITEBACK |
1548                                              PAGE_END_WRITEBACK);
1549         btrfs_free_path(path);
1550         return ret;
1551 }
1552
1553 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1554 {
1555
1556         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1557             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1558                 return 0;
1559
1560         /*
1561          * @defrag_bytes is a hint value, no spinlock held here,
1562          * if is not zero, it means the file is defragging.
1563          * Force cow if given extent needs to be defragged.
1564          */
1565         if (BTRFS_I(inode)->defrag_bytes &&
1566             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1567                            EXTENT_DEFRAG, 0, NULL))
1568                 return 1;
1569
1570         return 0;
1571 }
1572
1573 /*
1574  * extent_io.c call back to do delayed allocation processing
1575  */
1576 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1577                               u64 start, u64 end, int *page_started,
1578                               unsigned long *nr_written)
1579 {
1580         int ret;
1581         int force_cow = need_force_cow(inode, start, end);
1582
1583         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1584                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1585                                          page_started, 1, nr_written);
1586         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1587                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1588                                          page_started, 0, nr_written);
1589         } else if (!inode_need_compress(inode)) {
1590                 ret = cow_file_range(inode, locked_page, start, end, end,
1591                                       page_started, nr_written, 1, NULL);
1592         } else {
1593                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1594                         &BTRFS_I(inode)->runtime_flags);
1595                 ret = cow_file_range_async(inode, locked_page, start, end,
1596                                            page_started, nr_written);
1597         }
1598         return ret;
1599 }
1600
1601 static void btrfs_split_extent_hook(struct inode *inode,
1602                                     struct extent_state *orig, u64 split)
1603 {
1604         u64 size;
1605
1606         /* not delalloc, ignore it */
1607         if (!(orig->state & EXTENT_DELALLOC))
1608                 return;
1609
1610         size = orig->end - orig->start + 1;
1611         if (size > BTRFS_MAX_EXTENT_SIZE) {
1612                 u64 num_extents;
1613                 u64 new_size;
1614
1615                 /*
1616                  * See the explanation in btrfs_merge_extent_hook, the same
1617                  * applies here, just in reverse.
1618                  */
1619                 new_size = orig->end - split + 1;
1620                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1621                                         BTRFS_MAX_EXTENT_SIZE);
1622                 new_size = split - orig->start;
1623                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1624                                         BTRFS_MAX_EXTENT_SIZE);
1625                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1626                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1627                         return;
1628         }
1629
1630         spin_lock(&BTRFS_I(inode)->lock);
1631         BTRFS_I(inode)->outstanding_extents++;
1632         spin_unlock(&BTRFS_I(inode)->lock);
1633 }
1634
1635 /*
1636  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1637  * extents so we can keep track of new extents that are just merged onto old
1638  * extents, such as when we are doing sequential writes, so we can properly
1639  * account for the metadata space we'll need.
1640  */
1641 static void btrfs_merge_extent_hook(struct inode *inode,
1642                                     struct extent_state *new,
1643                                     struct extent_state *other)
1644 {
1645         u64 new_size, old_size;
1646         u64 num_extents;
1647
1648         /* not delalloc, ignore it */
1649         if (!(other->state & EXTENT_DELALLOC))
1650                 return;
1651
1652         if (new->start > other->start)
1653                 new_size = new->end - other->start + 1;
1654         else
1655                 new_size = other->end - new->start + 1;
1656
1657         /* we're not bigger than the max, unreserve the space and go */
1658         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1659                 spin_lock(&BTRFS_I(inode)->lock);
1660                 BTRFS_I(inode)->outstanding_extents--;
1661                 spin_unlock(&BTRFS_I(inode)->lock);
1662                 return;
1663         }
1664
1665         /*
1666          * We have to add up either side to figure out how many extents were
1667          * accounted for before we merged into one big extent.  If the number of
1668          * extents we accounted for is <= the amount we need for the new range
1669          * then we can return, otherwise drop.  Think of it like this
1670          *
1671          * [ 4k][MAX_SIZE]
1672          *
1673          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1674          * need 2 outstanding extents, on one side we have 1 and the other side
1675          * we have 1 so they are == and we can return.  But in this case
1676          *
1677          * [MAX_SIZE+4k][MAX_SIZE+4k]
1678          *
1679          * Each range on their own accounts for 2 extents, but merged together
1680          * they are only 3 extents worth of accounting, so we need to drop in
1681          * this case.
1682          */
1683         old_size = other->end - other->start + 1;
1684         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1685                                 BTRFS_MAX_EXTENT_SIZE);
1686         old_size = new->end - new->start + 1;
1687         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1688                                  BTRFS_MAX_EXTENT_SIZE);
1689
1690         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1691                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1692                 return;
1693
1694         spin_lock(&BTRFS_I(inode)->lock);
1695         BTRFS_I(inode)->outstanding_extents--;
1696         spin_unlock(&BTRFS_I(inode)->lock);
1697 }
1698
1699 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1700                                       struct inode *inode)
1701 {
1702         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1703
1704         spin_lock(&root->delalloc_lock);
1705         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1706                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1707                               &root->delalloc_inodes);
1708                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1709                         &BTRFS_I(inode)->runtime_flags);
1710                 root->nr_delalloc_inodes++;
1711                 if (root->nr_delalloc_inodes == 1) {
1712                         spin_lock(&fs_info->delalloc_root_lock);
1713                         BUG_ON(!list_empty(&root->delalloc_root));
1714                         list_add_tail(&root->delalloc_root,
1715                                       &fs_info->delalloc_roots);
1716                         spin_unlock(&fs_info->delalloc_root_lock);
1717                 }
1718         }
1719         spin_unlock(&root->delalloc_lock);
1720 }
1721
1722 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1723                                      struct inode *inode)
1724 {
1725         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1726
1727         spin_lock(&root->delalloc_lock);
1728         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1729                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1730                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1731                           &BTRFS_I(inode)->runtime_flags);
1732                 root->nr_delalloc_inodes--;
1733                 if (!root->nr_delalloc_inodes) {
1734                         spin_lock(&fs_info->delalloc_root_lock);
1735                         BUG_ON(list_empty(&root->delalloc_root));
1736                         list_del_init(&root->delalloc_root);
1737                         spin_unlock(&fs_info->delalloc_root_lock);
1738                 }
1739         }
1740         spin_unlock(&root->delalloc_lock);
1741 }
1742
1743 /*
1744  * extent_io.c set_bit_hook, used to track delayed allocation
1745  * bytes in this file, and to maintain the list of inodes that
1746  * have pending delalloc work to be done.
1747  */
1748 static void btrfs_set_bit_hook(struct inode *inode,
1749                                struct extent_state *state, unsigned *bits)
1750 {
1751
1752         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1753
1754         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1755                 WARN_ON(1);
1756         /*
1757          * set_bit and clear bit hooks normally require _irqsave/restore
1758          * but in this case, we are only testing for the DELALLOC
1759          * bit, which is only set or cleared with irqs on
1760          */
1761         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1762                 struct btrfs_root *root = BTRFS_I(inode)->root;
1763                 u64 len = state->end + 1 - state->start;
1764                 bool do_list = !btrfs_is_free_space_inode(inode);
1765
1766                 if (*bits & EXTENT_FIRST_DELALLOC) {
1767                         *bits &= ~EXTENT_FIRST_DELALLOC;
1768                 } else {
1769                         spin_lock(&BTRFS_I(inode)->lock);
1770                         BTRFS_I(inode)->outstanding_extents++;
1771                         spin_unlock(&BTRFS_I(inode)->lock);
1772                 }
1773
1774                 /* For sanity tests */
1775                 if (btrfs_is_testing(fs_info))
1776                         return;
1777
1778                 __percpu_counter_add(&fs_info->delalloc_bytes, len,
1779                                      fs_info->delalloc_batch);
1780                 spin_lock(&BTRFS_I(inode)->lock);
1781                 BTRFS_I(inode)->delalloc_bytes += len;
1782                 if (*bits & EXTENT_DEFRAG)
1783                         BTRFS_I(inode)->defrag_bytes += len;
1784                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1785                                          &BTRFS_I(inode)->runtime_flags))
1786                         btrfs_add_delalloc_inodes(root, inode);
1787                 spin_unlock(&BTRFS_I(inode)->lock);
1788         }
1789 }
1790
1791 /*
1792  * extent_io.c clear_bit_hook, see set_bit_hook for why
1793  */
1794 static void btrfs_clear_bit_hook(struct inode *inode,
1795                                  struct extent_state *state,
1796                                  unsigned *bits)
1797 {
1798         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1799         u64 len = state->end + 1 - state->start;
1800         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1801                                     BTRFS_MAX_EXTENT_SIZE);
1802
1803         spin_lock(&BTRFS_I(inode)->lock);
1804         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1805                 BTRFS_I(inode)->defrag_bytes -= len;
1806         spin_unlock(&BTRFS_I(inode)->lock);
1807
1808         /*
1809          * set_bit and clear bit hooks normally require _irqsave/restore
1810          * but in this case, we are only testing for the DELALLOC
1811          * bit, which is only set or cleared with irqs on
1812          */
1813         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1814                 struct btrfs_root *root = BTRFS_I(inode)->root;
1815                 bool do_list = !btrfs_is_free_space_inode(inode);
1816
1817                 if (*bits & EXTENT_FIRST_DELALLOC) {
1818                         *bits &= ~EXTENT_FIRST_DELALLOC;
1819                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1820                         spin_lock(&BTRFS_I(inode)->lock);
1821                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1822                         spin_unlock(&BTRFS_I(inode)->lock);
1823                 }
1824
1825                 /*
1826                  * We don't reserve metadata space for space cache inodes so we
1827                  * don't need to call dellalloc_release_metadata if there is an
1828                  * error.
1829                  */
1830                 if (*bits & EXTENT_DO_ACCOUNTING &&
1831                     root != fs_info->tree_root)
1832                         btrfs_delalloc_release_metadata(inode, len);
1833
1834                 /* For sanity tests. */
1835                 if (btrfs_is_testing(fs_info))
1836                         return;
1837
1838                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1839                     && do_list && !(state->state & EXTENT_NORESERVE)
1840                     && (*bits & (EXTENT_DO_ACCOUNTING |
1841                     EXTENT_CLEAR_DATA_RESV)))
1842                         btrfs_free_reserved_data_space_noquota(inode,
1843                                         state->start, len);
1844
1845                 __percpu_counter_add(&fs_info->delalloc_bytes, -len,
1846                                      fs_info->delalloc_batch);
1847                 spin_lock(&BTRFS_I(inode)->lock);
1848                 BTRFS_I(inode)->delalloc_bytes -= len;
1849                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1850                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1851                              &BTRFS_I(inode)->runtime_flags))
1852                         btrfs_del_delalloc_inode(root, inode);
1853                 spin_unlock(&BTRFS_I(inode)->lock);
1854         }
1855 }
1856
1857 /*
1858  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1859  * we don't create bios that span stripes or chunks
1860  *
1861  * return 1 if page cannot be merged to bio
1862  * return 0 if page can be merged to bio
1863  * return error otherwise
1864  */
1865 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1866                          size_t size, struct bio *bio,
1867                          unsigned long bio_flags)
1868 {
1869         struct inode *inode = page->mapping->host;
1870         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1871         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1872         u64 length = 0;
1873         u64 map_length;
1874         int ret;
1875
1876         if (bio_flags & EXTENT_BIO_COMPRESSED)
1877                 return 0;
1878
1879         length = bio->bi_iter.bi_size;
1880         map_length = length;
1881         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1882                               NULL, 0);
1883         if (ret < 0)
1884                 return ret;
1885         if (map_length < length + size)
1886                 return 1;
1887         return 0;
1888 }
1889
1890 /*
1891  * in order to insert checksums into the metadata in large chunks,
1892  * we wait until bio submission time.   All the pages in the bio are
1893  * checksummed and sums are attached onto the ordered extent record.
1894  *
1895  * At IO completion time the cums attached on the ordered extent record
1896  * are inserted into the btree
1897  */
1898 static int __btrfs_submit_bio_start(struct inode *inode, struct bio *bio,
1899                                     int mirror_num, unsigned long bio_flags,
1900                                     u64 bio_offset)
1901 {
1902         int ret = 0;
1903
1904         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1905         BUG_ON(ret); /* -ENOMEM */
1906         return 0;
1907 }
1908
1909 /*
1910  * in order to insert checksums into the metadata in large chunks,
1911  * we wait until bio submission time.   All the pages in the bio are
1912  * checksummed and sums are attached onto the ordered extent record.
1913  *
1914  * At IO completion time the cums attached on the ordered extent record
1915  * are inserted into the btree
1916  */
1917 static int __btrfs_submit_bio_done(struct inode *inode, struct bio *bio,
1918                           int mirror_num, unsigned long bio_flags,
1919                           u64 bio_offset)
1920 {
1921         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1922         int ret;
1923
1924         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1925         if (ret) {
1926                 bio->bi_error = ret;
1927                 bio_endio(bio);
1928         }
1929         return ret;
1930 }
1931
1932 /*
1933  * extent_io.c submission hook. This does the right thing for csum calculation
1934  * on write, or reading the csums from the tree before a read
1935  */
1936 static int btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
1937                           int mirror_num, unsigned long bio_flags,
1938                           u64 bio_offset)
1939 {
1940         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1941         struct btrfs_root *root = BTRFS_I(inode)->root;
1942         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1943         int ret = 0;
1944         int skip_sum;
1945         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1946
1947         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1948
1949         if (btrfs_is_free_space_inode(inode))
1950                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1951
1952         if (bio_op(bio) != REQ_OP_WRITE) {
1953                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1954                 if (ret)
1955                         goto out;
1956
1957                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1958                         ret = btrfs_submit_compressed_read(inode, bio,
1959                                                            mirror_num,
1960                                                            bio_flags);
1961                         goto out;
1962                 } else if (!skip_sum) {
1963                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
1964                         if (ret)
1965                                 goto out;
1966                 }
1967                 goto mapit;
1968         } else if (async && !skip_sum) {
1969                 /* csum items have already been cloned */
1970                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1971                         goto mapit;
1972                 /* we're doing a write, do the async checksumming */
1973                 ret = btrfs_wq_submit_bio(fs_info, inode, bio, mirror_num,
1974                                           bio_flags, bio_offset,
1975                                           __btrfs_submit_bio_start,
1976                                           __btrfs_submit_bio_done);
1977                 goto out;
1978         } else if (!skip_sum) {
1979                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1980                 if (ret)
1981                         goto out;
1982         }
1983
1984 mapit:
1985         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
1986
1987 out:
1988         if (ret < 0) {
1989                 bio->bi_error = ret;
1990                 bio_endio(bio);
1991         }
1992         return ret;
1993 }
1994
1995 /*
1996  * given a list of ordered sums record them in the inode.  This happens
1997  * at IO completion time based on sums calculated at bio submission time.
1998  */
1999 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2000                              struct inode *inode, u64 file_offset,
2001                              struct list_head *list)
2002 {
2003         struct btrfs_ordered_sum *sum;
2004
2005         list_for_each_entry(sum, list, list) {
2006                 trans->adding_csums = 1;
2007                 btrfs_csum_file_blocks(trans,
2008                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2009                 trans->adding_csums = 0;
2010         }
2011         return 0;
2012 }
2013
2014 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2015                               struct extent_state **cached_state, int dedupe)
2016 {
2017         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2018         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2019                                    cached_state);
2020 }
2021
2022 /* see btrfs_writepage_start_hook for details on why this is required */
2023 struct btrfs_writepage_fixup {
2024         struct page *page;
2025         struct btrfs_work work;
2026 };
2027
2028 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2029 {
2030         struct btrfs_writepage_fixup *fixup;
2031         struct btrfs_ordered_extent *ordered;
2032         struct extent_state *cached_state = NULL;
2033         struct page *page;
2034         struct inode *inode;
2035         u64 page_start;
2036         u64 page_end;
2037         int ret;
2038
2039         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2040         page = fixup->page;
2041 again:
2042         lock_page(page);
2043         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2044                 ClearPageChecked(page);
2045                 goto out_page;
2046         }
2047
2048         inode = page->mapping->host;
2049         page_start = page_offset(page);
2050         page_end = page_offset(page) + PAGE_SIZE - 1;
2051
2052         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2053                          &cached_state);
2054
2055         /* already ordered? We're done */
2056         if (PagePrivate2(page))
2057                 goto out;
2058
2059         ordered = btrfs_lookup_ordered_range(inode, page_start,
2060                                         PAGE_SIZE);
2061         if (ordered) {
2062                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2063                                      page_end, &cached_state, GFP_NOFS);
2064                 unlock_page(page);
2065                 btrfs_start_ordered_extent(inode, ordered, 1);
2066                 btrfs_put_ordered_extent(ordered);
2067                 goto again;
2068         }
2069
2070         ret = btrfs_delalloc_reserve_space(inode, page_start,
2071                                            PAGE_SIZE);
2072         if (ret) {
2073                 mapping_set_error(page->mapping, ret);
2074                 end_extent_writepage(page, ret, page_start, page_end);
2075                 ClearPageChecked(page);
2076                 goto out;
2077          }
2078
2079         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state,
2080                                   0);
2081         ClearPageChecked(page);
2082         set_page_dirty(page);
2083 out:
2084         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2085                              &cached_state, GFP_NOFS);
2086 out_page:
2087         unlock_page(page);
2088         put_page(page);
2089         kfree(fixup);
2090 }
2091
2092 /*
2093  * There are a few paths in the higher layers of the kernel that directly
2094  * set the page dirty bit without asking the filesystem if it is a
2095  * good idea.  This causes problems because we want to make sure COW
2096  * properly happens and the data=ordered rules are followed.
2097  *
2098  * In our case any range that doesn't have the ORDERED bit set
2099  * hasn't been properly setup for IO.  We kick off an async process
2100  * to fix it up.  The async helper will wait for ordered extents, set
2101  * the delalloc bit and make it safe to write the page.
2102  */
2103 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2104 {
2105         struct inode *inode = page->mapping->host;
2106         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2107         struct btrfs_writepage_fixup *fixup;
2108
2109         /* this page is properly in the ordered list */
2110         if (TestClearPagePrivate2(page))
2111                 return 0;
2112
2113         if (PageChecked(page))
2114                 return -EAGAIN;
2115
2116         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2117         if (!fixup)
2118                 return -EAGAIN;
2119
2120         SetPageChecked(page);
2121         get_page(page);
2122         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2123                         btrfs_writepage_fixup_worker, NULL, NULL);
2124         fixup->page = page;
2125         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2126         return -EBUSY;
2127 }
2128
2129 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2130                                        struct inode *inode, u64 file_pos,
2131                                        u64 disk_bytenr, u64 disk_num_bytes,
2132                                        u64 num_bytes, u64 ram_bytes,
2133                                        u8 compression, u8 encryption,
2134                                        u16 other_encoding, int extent_type)
2135 {
2136         struct btrfs_root *root = BTRFS_I(inode)->root;
2137         struct btrfs_file_extent_item *fi;
2138         struct btrfs_path *path;
2139         struct extent_buffer *leaf;
2140         struct btrfs_key ins;
2141         int extent_inserted = 0;
2142         int ret;
2143
2144         path = btrfs_alloc_path();
2145         if (!path)
2146                 return -ENOMEM;
2147
2148         /*
2149          * we may be replacing one extent in the tree with another.
2150          * The new extent is pinned in the extent map, and we don't want
2151          * to drop it from the cache until it is completely in the btree.
2152          *
2153          * So, tell btrfs_drop_extents to leave this extent in the cache.
2154          * the caller is expected to unpin it and allow it to be merged
2155          * with the others.
2156          */
2157         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2158                                    file_pos + num_bytes, NULL, 0,
2159                                    1, sizeof(*fi), &extent_inserted);
2160         if (ret)
2161                 goto out;
2162
2163         if (!extent_inserted) {
2164                 ins.objectid = btrfs_ino(inode);
2165                 ins.offset = file_pos;
2166                 ins.type = BTRFS_EXTENT_DATA_KEY;
2167
2168                 path->leave_spinning = 1;
2169                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2170                                               sizeof(*fi));
2171                 if (ret)
2172                         goto out;
2173         }
2174         leaf = path->nodes[0];
2175         fi = btrfs_item_ptr(leaf, path->slots[0],
2176                             struct btrfs_file_extent_item);
2177         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2178         btrfs_set_file_extent_type(leaf, fi, extent_type);
2179         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2180         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2181         btrfs_set_file_extent_offset(leaf, fi, 0);
2182         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2183         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2184         btrfs_set_file_extent_compression(leaf, fi, compression);
2185         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2186         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2187
2188         btrfs_mark_buffer_dirty(leaf);
2189         btrfs_release_path(path);
2190
2191         inode_add_bytes(inode, num_bytes);
2192
2193         ins.objectid = disk_bytenr;
2194         ins.offset = disk_num_bytes;
2195         ins.type = BTRFS_EXTENT_ITEM_KEY;
2196         ret = btrfs_alloc_reserved_file_extent(trans, root->root_key.objectid,
2197                                                btrfs_ino(inode), file_pos,
2198                                                ram_bytes, &ins);
2199         /*
2200          * Release the reserved range from inode dirty range map, as it is
2201          * already moved into delayed_ref_head
2202          */
2203         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2204 out:
2205         btrfs_free_path(path);
2206
2207         return ret;
2208 }
2209
2210 /* snapshot-aware defrag */
2211 struct sa_defrag_extent_backref {
2212         struct rb_node node;
2213         struct old_sa_defrag_extent *old;
2214         u64 root_id;
2215         u64 inum;
2216         u64 file_pos;
2217         u64 extent_offset;
2218         u64 num_bytes;
2219         u64 generation;
2220 };
2221
2222 struct old_sa_defrag_extent {
2223         struct list_head list;
2224         struct new_sa_defrag_extent *new;
2225
2226         u64 extent_offset;
2227         u64 bytenr;
2228         u64 offset;
2229         u64 len;
2230         int count;
2231 };
2232
2233 struct new_sa_defrag_extent {
2234         struct rb_root root;
2235         struct list_head head;
2236         struct btrfs_path *path;
2237         struct inode *inode;
2238         u64 file_pos;
2239         u64 len;
2240         u64 bytenr;
2241         u64 disk_len;
2242         u8 compress_type;
2243 };
2244
2245 static int backref_comp(struct sa_defrag_extent_backref *b1,
2246                         struct sa_defrag_extent_backref *b2)
2247 {
2248         if (b1->root_id < b2->root_id)
2249                 return -1;
2250         else if (b1->root_id > b2->root_id)
2251                 return 1;
2252
2253         if (b1->inum < b2->inum)
2254                 return -1;
2255         else if (b1->inum > b2->inum)
2256                 return 1;
2257
2258         if (b1->file_pos < b2->file_pos)
2259                 return -1;
2260         else if (b1->file_pos > b2->file_pos)
2261                 return 1;
2262
2263         /*
2264          * [------------------------------] ===> (a range of space)
2265          *     |<--->|   |<---->| =============> (fs/file tree A)
2266          * |<---------------------------->| ===> (fs/file tree B)
2267          *
2268          * A range of space can refer to two file extents in one tree while
2269          * refer to only one file extent in another tree.
2270          *
2271          * So we may process a disk offset more than one time(two extents in A)
2272          * and locate at the same extent(one extent in B), then insert two same
2273          * backrefs(both refer to the extent in B).
2274          */
2275         return 0;
2276 }
2277
2278 static void backref_insert(struct rb_root *root,
2279                            struct sa_defrag_extent_backref *backref)
2280 {
2281         struct rb_node **p = &root->rb_node;
2282         struct rb_node *parent = NULL;
2283         struct sa_defrag_extent_backref *entry;
2284         int ret;
2285
2286         while (*p) {
2287                 parent = *p;
2288                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2289
2290                 ret = backref_comp(backref, entry);
2291                 if (ret < 0)
2292                         p = &(*p)->rb_left;
2293                 else
2294                         p = &(*p)->rb_right;
2295         }
2296
2297         rb_link_node(&backref->node, parent, p);
2298         rb_insert_color(&backref->node, root);
2299 }
2300
2301 /*
2302  * Note the backref might has changed, and in this case we just return 0.
2303  */
2304 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2305                                        void *ctx)
2306 {
2307         struct btrfs_file_extent_item *extent;
2308         struct old_sa_defrag_extent *old = ctx;
2309         struct new_sa_defrag_extent *new = old->new;
2310         struct btrfs_path *path = new->path;
2311         struct btrfs_key key;
2312         struct btrfs_root *root;
2313         struct sa_defrag_extent_backref *backref;
2314         struct extent_buffer *leaf;
2315         struct inode *inode = new->inode;
2316         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2317         int slot;
2318         int ret;
2319         u64 extent_offset;
2320         u64 num_bytes;
2321
2322         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2323             inum == btrfs_ino(inode))
2324                 return 0;
2325
2326         key.objectid = root_id;
2327         key.type = BTRFS_ROOT_ITEM_KEY;
2328         key.offset = (u64)-1;
2329
2330         root = btrfs_read_fs_root_no_name(fs_info, &key);
2331         if (IS_ERR(root)) {
2332                 if (PTR_ERR(root) == -ENOENT)
2333                         return 0;
2334                 WARN_ON(1);
2335                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2336                          inum, offset, root_id);
2337                 return PTR_ERR(root);
2338         }
2339
2340         key.objectid = inum;
2341         key.type = BTRFS_EXTENT_DATA_KEY;
2342         if (offset > (u64)-1 << 32)
2343                 key.offset = 0;
2344         else
2345                 key.offset = offset;
2346
2347         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2348         if (WARN_ON(ret < 0))
2349                 return ret;
2350         ret = 0;
2351
2352         while (1) {
2353                 cond_resched();
2354
2355                 leaf = path->nodes[0];
2356                 slot = path->slots[0];
2357
2358                 if (slot >= btrfs_header_nritems(leaf)) {
2359                         ret = btrfs_next_leaf(root, path);
2360                         if (ret < 0) {
2361                                 goto out;
2362                         } else if (ret > 0) {
2363                                 ret = 0;
2364                                 goto out;
2365                         }
2366                         continue;
2367                 }
2368
2369                 path->slots[0]++;
2370
2371                 btrfs_item_key_to_cpu(leaf, &key, slot);
2372
2373                 if (key.objectid > inum)
2374                         goto out;
2375
2376                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2377                         continue;
2378
2379                 extent = btrfs_item_ptr(leaf, slot,
2380                                         struct btrfs_file_extent_item);
2381
2382                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2383                         continue;
2384
2385                 /*
2386                  * 'offset' refers to the exact key.offset,
2387                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2388                  * (key.offset - extent_offset).
2389                  */
2390                 if (key.offset != offset)
2391                         continue;
2392
2393                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2394                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2395
2396                 if (extent_offset >= old->extent_offset + old->offset +
2397                     old->len || extent_offset + num_bytes <=
2398                     old->extent_offset + old->offset)
2399                         continue;
2400                 break;
2401         }
2402
2403         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2404         if (!backref) {
2405                 ret = -ENOENT;
2406                 goto out;
2407         }
2408
2409         backref->root_id = root_id;
2410         backref->inum = inum;
2411         backref->file_pos = offset;
2412         backref->num_bytes = num_bytes;
2413         backref->extent_offset = extent_offset;
2414         backref->generation = btrfs_file_extent_generation(leaf, extent);
2415         backref->old = old;
2416         backref_insert(&new->root, backref);
2417         old->count++;
2418 out:
2419         btrfs_release_path(path);
2420         WARN_ON(ret);
2421         return ret;
2422 }
2423
2424 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2425                                    struct new_sa_defrag_extent *new)
2426 {
2427         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2428         struct old_sa_defrag_extent *old, *tmp;
2429         int ret;
2430
2431         new->path = path;
2432
2433         list_for_each_entry_safe(old, tmp, &new->head, list) {
2434                 ret = iterate_inodes_from_logical(old->bytenr +
2435                                                   old->extent_offset, fs_info,
2436                                                   path, record_one_backref,
2437                                                   old);
2438                 if (ret < 0 && ret != -ENOENT)
2439                         return false;
2440
2441                 /* no backref to be processed for this extent */
2442                 if (!old->count) {
2443                         list_del(&old->list);
2444                         kfree(old);
2445                 }
2446         }
2447
2448         if (list_empty(&new->head))
2449                 return false;
2450
2451         return true;
2452 }
2453
2454 static int relink_is_mergable(struct extent_buffer *leaf,
2455                               struct btrfs_file_extent_item *fi,
2456                               struct new_sa_defrag_extent *new)
2457 {
2458         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2459                 return 0;
2460
2461         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2462                 return 0;
2463
2464         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2465                 return 0;
2466
2467         if (btrfs_file_extent_encryption(leaf, fi) ||
2468             btrfs_file_extent_other_encoding(leaf, fi))
2469                 return 0;
2470
2471         return 1;
2472 }
2473
2474 /*
2475  * Note the backref might has changed, and in this case we just return 0.
2476  */
2477 static noinline int relink_extent_backref(struct btrfs_path *path,
2478                                  struct sa_defrag_extent_backref *prev,
2479                                  struct sa_defrag_extent_backref *backref)
2480 {
2481         struct btrfs_file_extent_item *extent;
2482         struct btrfs_file_extent_item *item;
2483         struct btrfs_ordered_extent *ordered;
2484         struct btrfs_trans_handle *trans;
2485         struct btrfs_root *root;
2486         struct btrfs_key key;
2487         struct extent_buffer *leaf;
2488         struct old_sa_defrag_extent *old = backref->old;
2489         struct new_sa_defrag_extent *new = old->new;
2490         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2491         struct inode *inode;
2492         struct extent_state *cached = NULL;
2493         int ret = 0;
2494         u64 start;
2495         u64 len;
2496         u64 lock_start;
2497         u64 lock_end;
2498         bool merge = false;
2499         int index;
2500
2501         if (prev && prev->root_id == backref->root_id &&
2502             prev->inum == backref->inum &&
2503             prev->file_pos + prev->num_bytes == backref->file_pos)
2504                 merge = true;
2505
2506         /* step 1: get root */
2507         key.objectid = backref->root_id;
2508         key.type = BTRFS_ROOT_ITEM_KEY;
2509         key.offset = (u64)-1;
2510
2511         index = srcu_read_lock(&fs_info->subvol_srcu);
2512
2513         root = btrfs_read_fs_root_no_name(fs_info, &key);
2514         if (IS_ERR(root)) {
2515                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2516                 if (PTR_ERR(root) == -ENOENT)
2517                         return 0;
2518                 return PTR_ERR(root);
2519         }
2520
2521         if (btrfs_root_readonly(root)) {
2522                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2523                 return 0;
2524         }
2525
2526         /* step 2: get inode */
2527         key.objectid = backref->inum;
2528         key.type = BTRFS_INODE_ITEM_KEY;
2529         key.offset = 0;
2530
2531         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2532         if (IS_ERR(inode)) {
2533                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2534                 return 0;
2535         }
2536
2537         srcu_read_unlock(&fs_info->subvol_srcu, index);
2538
2539         /* step 3: relink backref */
2540         lock_start = backref->file_pos;
2541         lock_end = backref->file_pos + backref->num_bytes - 1;
2542         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2543                          &cached);
2544
2545         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2546         if (ordered) {
2547                 btrfs_put_ordered_extent(ordered);
2548                 goto out_unlock;
2549         }
2550
2551         trans = btrfs_join_transaction(root);
2552         if (IS_ERR(trans)) {
2553                 ret = PTR_ERR(trans);
2554                 goto out_unlock;
2555         }
2556
2557         key.objectid = backref->inum;
2558         key.type = BTRFS_EXTENT_DATA_KEY;
2559         key.offset = backref->file_pos;
2560
2561         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2562         if (ret < 0) {
2563                 goto out_free_path;
2564         } else if (ret > 0) {
2565                 ret = 0;
2566                 goto out_free_path;
2567         }
2568
2569         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2570                                 struct btrfs_file_extent_item);
2571
2572         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2573             backref->generation)
2574                 goto out_free_path;
2575
2576         btrfs_release_path(path);
2577
2578         start = backref->file_pos;
2579         if (backref->extent_offset < old->extent_offset + old->offset)
2580                 start += old->extent_offset + old->offset -
2581                          backref->extent_offset;
2582
2583         len = min(backref->extent_offset + backref->num_bytes,
2584                   old->extent_offset + old->offset + old->len);
2585         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2586
2587         ret = btrfs_drop_extents(trans, root, inode, start,
2588                                  start + len, 1);
2589         if (ret)
2590                 goto out_free_path;
2591 again:
2592         key.objectid = btrfs_ino(inode);
2593         key.type = BTRFS_EXTENT_DATA_KEY;
2594         key.offset = start;
2595
2596         path->leave_spinning = 1;
2597         if (merge) {
2598                 struct btrfs_file_extent_item *fi;
2599                 u64 extent_len;
2600                 struct btrfs_key found_key;
2601
2602                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2603                 if (ret < 0)
2604                         goto out_free_path;
2605
2606                 path->slots[0]--;
2607                 leaf = path->nodes[0];
2608                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2609
2610                 fi = btrfs_item_ptr(leaf, path->slots[0],
2611                                     struct btrfs_file_extent_item);
2612                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2613
2614                 if (extent_len + found_key.offset == start &&
2615                     relink_is_mergable(leaf, fi, new)) {
2616                         btrfs_set_file_extent_num_bytes(leaf, fi,
2617                                                         extent_len + len);
2618                         btrfs_mark_buffer_dirty(leaf);
2619                         inode_add_bytes(inode, len);
2620
2621                         ret = 1;
2622                         goto out_free_path;
2623                 } else {
2624                         merge = false;
2625                         btrfs_release_path(path);
2626                         goto again;
2627                 }
2628         }
2629
2630         ret = btrfs_insert_empty_item(trans, root, path, &key,
2631                                         sizeof(*extent));
2632         if (ret) {
2633                 btrfs_abort_transaction(trans, ret);
2634                 goto out_free_path;
2635         }
2636
2637         leaf = path->nodes[0];
2638         item = btrfs_item_ptr(leaf, path->slots[0],
2639                                 struct btrfs_file_extent_item);
2640         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2641         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2642         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2643         btrfs_set_file_extent_num_bytes(leaf, item, len);
2644         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2645         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2646         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2647         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2648         btrfs_set_file_extent_encryption(leaf, item, 0);
2649         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2650
2651         btrfs_mark_buffer_dirty(leaf);
2652         inode_add_bytes(inode, len);
2653         btrfs_release_path(path);
2654
2655         ret = btrfs_inc_extent_ref(trans, fs_info, new->bytenr,
2656                         new->disk_len, 0,
2657                         backref->root_id, backref->inum,
2658                         new->file_pos); /* start - extent_offset */
2659         if (ret) {
2660                 btrfs_abort_transaction(trans, ret);
2661                 goto out_free_path;
2662         }
2663
2664         ret = 1;
2665 out_free_path:
2666         btrfs_release_path(path);
2667         path->leave_spinning = 0;
2668         btrfs_end_transaction(trans);
2669 out_unlock:
2670         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2671                              &cached, GFP_NOFS);
2672         iput(inode);
2673         return ret;
2674 }
2675
2676 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2677 {
2678         struct old_sa_defrag_extent *old, *tmp;
2679
2680         if (!new)
2681                 return;
2682
2683         list_for_each_entry_safe(old, tmp, &new->head, list) {
2684                 kfree(old);
2685         }
2686         kfree(new);
2687 }
2688
2689 static void relink_file_extents(struct new_sa_defrag_extent *new)
2690 {
2691         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2692         struct btrfs_path *path;
2693         struct sa_defrag_extent_backref *backref;
2694         struct sa_defrag_extent_backref *prev = NULL;
2695         struct inode *inode;
2696         struct btrfs_root *root;
2697         struct rb_node *node;
2698         int ret;
2699
2700         inode = new->inode;
2701         root = BTRFS_I(inode)->root;
2702
2703         path = btrfs_alloc_path();
2704         if (!path)
2705                 return;
2706
2707         if (!record_extent_backrefs(path, new)) {
2708                 btrfs_free_path(path);
2709                 goto out;
2710         }
2711         btrfs_release_path(path);
2712
2713         while (1) {
2714                 node = rb_first(&new->root);
2715                 if (!node)
2716                         break;
2717                 rb_erase(node, &new->root);
2718
2719                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2720
2721                 ret = relink_extent_backref(path, prev, backref);
2722                 WARN_ON(ret < 0);
2723
2724                 kfree(prev);
2725
2726                 if (ret == 1)
2727                         prev = backref;
2728                 else
2729                         prev = NULL;
2730                 cond_resched();
2731         }
2732         kfree(prev);
2733
2734         btrfs_free_path(path);
2735 out:
2736         free_sa_defrag_extent(new);
2737
2738         atomic_dec(&fs_info->defrag_running);
2739         wake_up(&fs_info->transaction_wait);
2740 }
2741
2742 static struct new_sa_defrag_extent *
2743 record_old_file_extents(struct inode *inode,
2744                         struct btrfs_ordered_extent *ordered)
2745 {
2746         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2747         struct btrfs_root *root = BTRFS_I(inode)->root;
2748         struct btrfs_path *path;
2749         struct btrfs_key key;
2750         struct old_sa_defrag_extent *old;
2751         struct new_sa_defrag_extent *new;
2752         int ret;
2753
2754         new = kmalloc(sizeof(*new), GFP_NOFS);
2755         if (!new)
2756                 return NULL;
2757
2758         new->inode = inode;
2759         new->file_pos = ordered->file_offset;
2760         new->len = ordered->len;
2761         new->bytenr = ordered->start;
2762         new->disk_len = ordered->disk_len;
2763         new->compress_type = ordered->compress_type;
2764         new->root = RB_ROOT;
2765         INIT_LIST_HEAD(&new->head);
2766
2767         path = btrfs_alloc_path();
2768         if (!path)
2769                 goto out_kfree;
2770
2771         key.objectid = btrfs_ino(inode);
2772         key.type = BTRFS_EXTENT_DATA_KEY;
2773         key.offset = new->file_pos;
2774
2775         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2776         if (ret < 0)
2777                 goto out_free_path;
2778         if (ret > 0 && path->slots[0] > 0)
2779                 path->slots[0]--;
2780
2781         /* find out all the old extents for the file range */
2782         while (1) {
2783                 struct btrfs_file_extent_item *extent;
2784                 struct extent_buffer *l;
2785                 int slot;
2786                 u64 num_bytes;
2787                 u64 offset;
2788                 u64 end;
2789                 u64 disk_bytenr;
2790                 u64 extent_offset;
2791
2792                 l = path->nodes[0];
2793                 slot = path->slots[0];
2794
2795                 if (slot >= btrfs_header_nritems(l)) {
2796                         ret = btrfs_next_leaf(root, path);
2797                         if (ret < 0)
2798                                 goto out_free_path;
2799                         else if (ret > 0)
2800                                 break;
2801                         continue;
2802                 }
2803
2804                 btrfs_item_key_to_cpu(l, &key, slot);
2805
2806                 if (key.objectid != btrfs_ino(inode))
2807                         break;
2808                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2809                         break;
2810                 if (key.offset >= new->file_pos + new->len)
2811                         break;
2812
2813                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2814
2815                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2816                 if (key.offset + num_bytes < new->file_pos)
2817                         goto next;
2818
2819                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2820                 if (!disk_bytenr)
2821                         goto next;
2822
2823                 extent_offset = btrfs_file_extent_offset(l, extent);
2824
2825                 old = kmalloc(sizeof(*old), GFP_NOFS);
2826                 if (!old)
2827                         goto out_free_path;
2828
2829                 offset = max(new->file_pos, key.offset);
2830                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2831
2832                 old->bytenr = disk_bytenr;
2833                 old->extent_offset = extent_offset;
2834                 old->offset = offset - key.offset;
2835                 old->len = end - offset;
2836                 old->new = new;
2837                 old->count = 0;
2838                 list_add_tail(&old->list, &new->head);
2839 next:
2840                 path->slots[0]++;
2841                 cond_resched();
2842         }
2843
2844         btrfs_free_path(path);
2845         atomic_inc(&fs_info->defrag_running);
2846
2847         return new;
2848
2849 out_free_path:
2850         btrfs_free_path(path);
2851 out_kfree:
2852         free_sa_defrag_extent(new);
2853         return NULL;
2854 }
2855
2856 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2857                                          u64 start, u64 len)
2858 {
2859         struct btrfs_block_group_cache *cache;
2860
2861         cache = btrfs_lookup_block_group(fs_info, start);
2862         ASSERT(cache);
2863
2864         spin_lock(&cache->lock);
2865         cache->delalloc_bytes -= len;
2866         spin_unlock(&cache->lock);
2867
2868         btrfs_put_block_group(cache);
2869 }
2870
2871 /* as ordered data IO finishes, this gets called so we can finish
2872  * an ordered extent if the range of bytes in the file it covers are
2873  * fully written.
2874  */
2875 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2876 {
2877         struct inode *inode = ordered_extent->inode;
2878         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2879         struct btrfs_root *root = BTRFS_I(inode)->root;
2880         struct btrfs_trans_handle *trans = NULL;
2881         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2882         struct extent_state *cached_state = NULL;
2883         struct new_sa_defrag_extent *new = NULL;
2884         int compress_type = 0;
2885         int ret = 0;
2886         u64 logical_len = ordered_extent->len;
2887         bool nolock;
2888         bool truncated = false;
2889
2890         nolock = btrfs_is_free_space_inode(inode);
2891
2892         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2893                 ret = -EIO;
2894                 goto out;
2895         }
2896
2897         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2898                                      ordered_extent->file_offset +
2899                                      ordered_extent->len - 1);
2900
2901         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2902                 truncated = true;
2903                 logical_len = ordered_extent->truncated_len;
2904                 /* Truncated the entire extent, don't bother adding */
2905                 if (!logical_len)
2906                         goto out;
2907         }
2908
2909         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2910                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2911
2912                 /*
2913                  * For mwrite(mmap + memset to write) case, we still reserve
2914                  * space for NOCOW range.
2915                  * As NOCOW won't cause a new delayed ref, just free the space
2916                  */
2917                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2918                                        ordered_extent->len);
2919                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2920                 if (nolock)
2921                         trans = btrfs_join_transaction_nolock(root);
2922                 else
2923                         trans = btrfs_join_transaction(root);
2924                 if (IS_ERR(trans)) {
2925                         ret = PTR_ERR(trans);
2926                         trans = NULL;
2927                         goto out;
2928                 }
2929                 trans->block_rsv = &fs_info->delalloc_block_rsv;
2930                 ret = btrfs_update_inode_fallback(trans, root, inode);
2931                 if (ret) /* -ENOMEM or corruption */
2932                         btrfs_abort_transaction(trans, ret);
2933                 goto out;
2934         }
2935
2936         lock_extent_bits(io_tree, ordered_extent->file_offset,
2937                          ordered_extent->file_offset + ordered_extent->len - 1,
2938                          &cached_state);
2939
2940         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2941                         ordered_extent->file_offset + ordered_extent->len - 1,
2942                         EXTENT_DEFRAG, 1, cached_state);
2943         if (ret) {
2944                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2945                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2946                         /* the inode is shared */
2947                         new = record_old_file_extents(inode, ordered_extent);
2948
2949                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2950                         ordered_extent->file_offset + ordered_extent->len - 1,
2951                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2952         }
2953
2954         if (nolock)
2955                 trans = btrfs_join_transaction_nolock(root);
2956         else
2957                 trans = btrfs_join_transaction(root);
2958         if (IS_ERR(trans)) {
2959                 ret = PTR_ERR(trans);
2960                 trans = NULL;
2961                 goto out_unlock;
2962         }
2963
2964         trans->block_rsv = &fs_info->delalloc_block_rsv;
2965
2966         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2967                 compress_type = ordered_extent->compress_type;
2968         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2969                 BUG_ON(compress_type);
2970                 ret = btrfs_mark_extent_written(trans, inode,
2971                                                 ordered_extent->file_offset,
2972                                                 ordered_extent->file_offset +
2973                                                 logical_len);
2974         } else {
2975                 BUG_ON(root == fs_info->tree_root);
2976                 ret = insert_reserved_file_extent(trans, inode,
2977                                                 ordered_extent->file_offset,
2978                                                 ordered_extent->start,
2979                                                 ordered_extent->disk_len,
2980                                                 logical_len, logical_len,
2981                                                 compress_type, 0, 0,
2982                                                 BTRFS_FILE_EXTENT_REG);
2983                 if (!ret)
2984                         btrfs_release_delalloc_bytes(fs_info,
2985                                                      ordered_extent->start,
2986                                                      ordered_extent->disk_len);
2987         }
2988         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2989                            ordered_extent->file_offset, ordered_extent->len,
2990                            trans->transid);
2991         if (ret < 0) {
2992                 btrfs_abort_transaction(trans, ret);
2993                 goto out_unlock;
2994         }
2995
2996         add_pending_csums(trans, inode, ordered_extent->file_offset,
2997                           &ordered_extent->list);
2998
2999         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3000         ret = btrfs_update_inode_fallback(trans, root, inode);
3001         if (ret) { /* -ENOMEM or corruption */
3002                 btrfs_abort_transaction(trans, ret);
3003                 goto out_unlock;
3004         }
3005         ret = 0;
3006 out_unlock:
3007         unlock_extent_cached(io_tree, ordered_extent->file_offset,
3008                              ordered_extent->file_offset +
3009                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
3010 out:
3011         if (root != fs_info->tree_root)
3012                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
3013         if (trans)
3014                 btrfs_end_transaction(trans);
3015
3016         if (ret || truncated) {
3017                 u64 start, end;
3018
3019                 if (truncated)
3020                         start = ordered_extent->file_offset + logical_len;
3021                 else
3022                         start = ordered_extent->file_offset;
3023                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3024                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
3025
3026                 /* Drop the cache for the part of the extent we didn't write. */
3027                 btrfs_drop_extent_cache(inode, start, end, 0);
3028
3029                 /*
3030                  * If the ordered extent had an IOERR or something else went
3031                  * wrong we need to return the space for this ordered extent
3032                  * back to the allocator.  We only free the extent in the
3033                  * truncated case if we didn't write out the extent at all.
3034                  */
3035                 if ((ret || !logical_len) &&
3036                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3037                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3038                         btrfs_free_reserved_extent(fs_info,
3039                                                    ordered_extent->start,
3040                                                    ordered_extent->disk_len, 1);
3041         }
3042
3043
3044         /*
3045          * This needs to be done to make sure anybody waiting knows we are done
3046          * updating everything for this ordered extent.
3047          */
3048         btrfs_remove_ordered_extent(inode, ordered_extent);
3049
3050         /* for snapshot-aware defrag */
3051         if (new) {
3052                 if (ret) {
3053                         free_sa_defrag_extent(new);
3054                         atomic_dec(&fs_info->defrag_running);
3055                 } else {
3056                         relink_file_extents(new);
3057                 }
3058         }
3059
3060         /* once for us */
3061         btrfs_put_ordered_extent(ordered_extent);
3062         /* once for the tree */
3063         btrfs_put_ordered_extent(ordered_extent);
3064
3065         return ret;
3066 }
3067
3068 static void finish_ordered_fn(struct btrfs_work *work)
3069 {
3070         struct btrfs_ordered_extent *ordered_extent;
3071         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3072         btrfs_finish_ordered_io(ordered_extent);
3073 }
3074
3075 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3076                                 struct extent_state *state, int uptodate)
3077 {
3078         struct inode *inode = page->mapping->host;
3079         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3080         struct btrfs_ordered_extent *ordered_extent = NULL;
3081         struct btrfs_workqueue *wq;
3082         btrfs_work_func_t func;
3083
3084         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3085
3086         ClearPagePrivate2(page);
3087         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3088                                             end - start + 1, uptodate))
3089                 return 0;
3090
3091         if (btrfs_is_free_space_inode(inode)) {
3092                 wq = fs_info->endio_freespace_worker;
3093                 func = btrfs_freespace_write_helper;
3094         } else {
3095                 wq = fs_info->endio_write_workers;
3096                 func = btrfs_endio_write_helper;
3097         }
3098
3099         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3100                         NULL);
3101         btrfs_queue_work(wq, &ordered_extent->work);
3102
3103         return 0;
3104 }
3105
3106 static int __readpage_endio_check(struct inode *inode,
3107                                   struct btrfs_io_bio *io_bio,
3108                                   int icsum, struct page *page,
3109                                   int pgoff, u64 start, size_t len)
3110 {
3111         char *kaddr;
3112         u32 csum_expected;
3113         u32 csum = ~(u32)0;
3114
3115         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3116
3117         kaddr = kmap_atomic(page);
3118         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3119         btrfs_csum_final(csum, (u8 *)&csum);
3120         if (csum != csum_expected)
3121                 goto zeroit;
3122
3123         kunmap_atomic(kaddr);
3124         return 0;
3125 zeroit:
3126         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3127                 "csum failed ino %llu off %llu csum %u expected csum %u",
3128                            btrfs_ino(inode), start, csum, csum_expected);
3129         memset(kaddr + pgoff, 1, len);
3130         flush_dcache_page(page);
3131         kunmap_atomic(kaddr);
3132         if (csum_expected == 0)
3133                 return 0;
3134         return -EIO;
3135 }
3136
3137 /*
3138  * when reads are done, we need to check csums to verify the data is correct
3139  * if there's a match, we allow the bio to finish.  If not, the code in
3140  * extent_io.c will try to find good copies for us.
3141  */
3142 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3143                                       u64 phy_offset, struct page *page,
3144                                       u64 start, u64 end, int mirror)
3145 {
3146         size_t offset = start - page_offset(page);
3147         struct inode *inode = page->mapping->host;
3148         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3149         struct btrfs_root *root = BTRFS_I(inode)->root;
3150
3151         if (PageChecked(page)) {
3152                 ClearPageChecked(page);
3153                 return 0;
3154         }
3155
3156         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3157                 return 0;
3158
3159         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3160             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3161                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3162                 return 0;
3163         }
3164
3165         phy_offset >>= inode->i_sb->s_blocksize_bits;
3166         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3167                                       start, (size_t)(end - start + 1));
3168 }
3169
3170 void btrfs_add_delayed_iput(struct inode *inode)
3171 {
3172         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3173         struct btrfs_inode *binode = BTRFS_I(inode);
3174
3175         if (atomic_add_unless(&inode->i_count, -1, 1))
3176                 return;
3177
3178         spin_lock(&fs_info->delayed_iput_lock);
3179         if (binode->delayed_iput_count == 0) {
3180                 ASSERT(list_empty(&binode->delayed_iput));
3181                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3182         } else {
3183                 binode->delayed_iput_count++;
3184         }
3185         spin_unlock(&fs_info->delayed_iput_lock);
3186 }
3187
3188 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3189 {
3190
3191         spin_lock(&fs_info->delayed_iput_lock);
3192         while (!list_empty(&fs_info->delayed_iputs)) {
3193                 struct btrfs_inode *inode;
3194
3195                 inode = list_first_entry(&fs_info->delayed_iputs,
3196                                 struct btrfs_inode, delayed_iput);
3197                 if (inode->delayed_iput_count) {
3198                         inode->delayed_iput_count--;
3199                         list_move_tail(&inode->delayed_iput,
3200                                         &fs_info->delayed_iputs);
3201                 } else {
3202                         list_del_init(&inode->delayed_iput);
3203                 }
3204                 spin_unlock(&fs_info->delayed_iput_lock);
3205                 iput(&inode->vfs_inode);
3206                 spin_lock(&fs_info->delayed_iput_lock);
3207         }
3208         spin_unlock(&fs_info->delayed_iput_lock);
3209 }
3210
3211 /*
3212  * This is called in transaction commit time. If there are no orphan
3213  * files in the subvolume, it removes orphan item and frees block_rsv
3214  * structure.
3215  */
3216 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3217                               struct btrfs_root *root)
3218 {
3219         struct btrfs_fs_info *fs_info = root->fs_info;
3220         struct btrfs_block_rsv *block_rsv;
3221         int ret;
3222
3223         if (atomic_read(&root->orphan_inodes) ||
3224             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3225                 return;
3226
3227         spin_lock(&root->orphan_lock);
3228         if (atomic_read(&root->orphan_inodes)) {
3229                 spin_unlock(&root->orphan_lock);
3230                 return;
3231         }
3232
3233         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3234                 spin_unlock(&root->orphan_lock);
3235                 return;
3236         }
3237
3238         block_rsv = root->orphan_block_rsv;
3239         root->orphan_block_rsv = NULL;
3240         spin_unlock(&root->orphan_lock);
3241
3242         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3243             btrfs_root_refs(&root->root_item) > 0) {
3244                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3245                                             root->root_key.objectid);
3246                 if (ret)
3247                         btrfs_abort_transaction(trans, ret);
3248                 else
3249                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3250                                   &root->state);
3251         }
3252
3253         if (block_rsv) {
3254                 WARN_ON(block_rsv->size > 0);
3255                 btrfs_free_block_rsv(fs_info, block_rsv);
3256         }
3257 }
3258
3259 /*
3260  * This creates an orphan entry for the given inode in case something goes
3261  * wrong in the middle of an unlink/truncate.
3262  *
3263  * NOTE: caller of this function should reserve 5 units of metadata for
3264  *       this function.
3265  */
3266 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3267 {
3268         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3269         struct btrfs_root *root = BTRFS_I(inode)->root;
3270         struct btrfs_block_rsv *block_rsv = NULL;
3271         int reserve = 0;
3272         int insert = 0;
3273         int ret;
3274
3275         if (!root->orphan_block_rsv) {
3276                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3277                                                   BTRFS_BLOCK_RSV_TEMP);
3278                 if (!block_rsv)
3279                         return -ENOMEM;
3280         }
3281
3282         spin_lock(&root->orphan_lock);
3283         if (!root->orphan_block_rsv) {
3284                 root->orphan_block_rsv = block_rsv;
3285         } else if (block_rsv) {
3286                 btrfs_free_block_rsv(fs_info, block_rsv);
3287                 block_rsv = NULL;
3288         }
3289
3290         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3291                               &BTRFS_I(inode)->runtime_flags)) {
3292 #if 0
3293                 /*
3294                  * For proper ENOSPC handling, we should do orphan
3295                  * cleanup when mounting. But this introduces backward
3296                  * compatibility issue.
3297                  */
3298                 if (!xchg(&root->orphan_item_inserted, 1))
3299                         insert = 2;
3300                 else
3301                         insert = 1;
3302 #endif
3303                 insert = 1;
3304                 atomic_inc(&root->orphan_inodes);
3305         }
3306
3307         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3308                               &BTRFS_I(inode)->runtime_flags))
3309                 reserve = 1;
3310         spin_unlock(&root->orphan_lock);
3311
3312         /* grab metadata reservation from transaction handle */
3313         if (reserve) {
3314                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3315                 ASSERT(!ret);
3316                 if (ret) {
3317                         atomic_dec(&root->orphan_inodes);
3318                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3319                                   &BTRFS_I(inode)->runtime_flags);
3320                         if (insert)
3321                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3322                                           &BTRFS_I(inode)->runtime_flags);
3323                         return ret;
3324                 }
3325         }
3326
3327         /* insert an orphan item to track this unlinked/truncated file */
3328         if (insert >= 1) {
3329                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3330                 if (ret) {
3331                         atomic_dec(&root->orphan_inodes);
3332                         if (reserve) {
3333                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3334                                           &BTRFS_I(inode)->runtime_flags);
3335                                 btrfs_orphan_release_metadata(inode);
3336                         }
3337                         if (ret != -EEXIST) {
3338                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3339                                           &BTRFS_I(inode)->runtime_flags);
3340                                 btrfs_abort_transaction(trans, ret);
3341                                 return ret;
3342                         }
3343                 }
3344                 ret = 0;
3345         }
3346
3347         /* insert an orphan item to track subvolume contains orphan files */
3348         if (insert >= 2) {
3349                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3350                                                root->root_key.objectid);
3351                 if (ret && ret != -EEXIST) {
3352                         btrfs_abort_transaction(trans, ret);
3353                         return ret;
3354                 }
3355         }
3356         return 0;
3357 }
3358
3359 /*
3360  * We have done the truncate/delete so we can go ahead and remove the orphan
3361  * item for this particular inode.
3362  */
3363 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3364                             struct inode *inode)
3365 {
3366         struct btrfs_root *root = BTRFS_I(inode)->root;
3367         int delete_item = 0;
3368         int release_rsv = 0;
3369         int ret = 0;
3370
3371         spin_lock(&root->orphan_lock);
3372         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3373                                &BTRFS_I(inode)->runtime_flags))
3374                 delete_item = 1;
3375
3376         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3377                                &BTRFS_I(inode)->runtime_flags))
3378                 release_rsv = 1;
3379         spin_unlock(&root->orphan_lock);
3380
3381         if (delete_item) {
3382                 atomic_dec(&root->orphan_inodes);
3383                 if (trans)
3384                         ret = btrfs_del_orphan_item(trans, root,
3385                                                     btrfs_ino(inode));
3386         }
3387
3388         if (release_rsv)
3389                 btrfs_orphan_release_metadata(inode);
3390
3391         return ret;
3392 }
3393
3394 /*
3395  * this cleans up any orphans that may be left on the list from the last use
3396  * of this root.
3397  */
3398 int btrfs_orphan_cleanup(struct btrfs_root *root)
3399 {
3400         struct btrfs_fs_info *fs_info = root->fs_info;
3401         struct btrfs_path *path;
3402         struct extent_buffer *leaf;
3403         struct btrfs_key key, found_key;
3404         struct btrfs_trans_handle *trans;
3405         struct inode *inode;
3406         u64 last_objectid = 0;
3407         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3408
3409         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3410                 return 0;
3411
3412         path = btrfs_alloc_path();
3413         if (!path) {
3414                 ret = -ENOMEM;
3415                 goto out;
3416         }
3417         path->reada = READA_BACK;
3418
3419         key.objectid = BTRFS_ORPHAN_OBJECTID;
3420         key.type = BTRFS_ORPHAN_ITEM_KEY;
3421         key.offset = (u64)-1;
3422
3423         while (1) {
3424                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3425                 if (ret < 0)
3426                         goto out;
3427
3428                 /*
3429                  * if ret == 0 means we found what we were searching for, which
3430                  * is weird, but possible, so only screw with path if we didn't
3431                  * find the key and see if we have stuff that matches
3432                  */
3433                 if (ret > 0) {
3434                         ret = 0;
3435                         if (path->slots[0] == 0)
3436                                 break;
3437                         path->slots[0]--;
3438                 }
3439
3440                 /* pull out the item */
3441                 leaf = path->nodes[0];
3442                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3443
3444                 /* make sure the item matches what we want */
3445                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3446                         break;
3447                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3448                         break;
3449
3450                 /* release the path since we're done with it */
3451                 btrfs_release_path(path);
3452
3453                 /*
3454                  * this is where we are basically btrfs_lookup, without the
3455                  * crossing root thing.  we store the inode number in the
3456                  * offset of the orphan item.
3457                  */
3458
3459                 if (found_key.offset == last_objectid) {
3460                         btrfs_err(fs_info,
3461                                   "Error removing orphan entry, stopping orphan cleanup");
3462                         ret = -EINVAL;
3463                         goto out;
3464                 }
3465
3466                 last_objectid = found_key.offset;
3467
3468                 found_key.objectid = found_key.offset;
3469                 found_key.type = BTRFS_INODE_ITEM_KEY;
3470                 found_key.offset = 0;
3471                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3472                 ret = PTR_ERR_OR_ZERO(inode);
3473                 if (ret && ret != -ENOENT)
3474                         goto out;
3475
3476                 if (ret == -ENOENT && root == fs_info->tree_root) {
3477                         struct btrfs_root *dead_root;
3478                         struct btrfs_fs_info *fs_info = root->fs_info;
3479                         int is_dead_root = 0;
3480
3481                         /*
3482                          * this is an orphan in the tree root. Currently these
3483                          * could come from 2 sources:
3484                          *  a) a snapshot deletion in progress
3485                          *  b) a free space cache inode
3486                          * We need to distinguish those two, as the snapshot
3487                          * orphan must not get deleted.
3488                          * find_dead_roots already ran before us, so if this
3489                          * is a snapshot deletion, we should find the root
3490                          * in the dead_roots list
3491                          */
3492                         spin_lock(&fs_info->trans_lock);
3493                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3494                                             root_list) {
3495                                 if (dead_root->root_key.objectid ==
3496                                     found_key.objectid) {
3497                                         is_dead_root = 1;
3498                                         break;
3499                                 }
3500                         }
3501                         spin_unlock(&fs_info->trans_lock);
3502                         if (is_dead_root) {
3503                                 /* prevent this orphan from being found again */
3504                                 key.offset = found_key.objectid - 1;
3505                                 continue;
3506                         }
3507                 }
3508                 /*
3509                  * Inode is already gone but the orphan item is still there,
3510                  * kill the orphan item.
3511                  */
3512                 if (ret == -ENOENT) {
3513                         trans = btrfs_start_transaction(root, 1);
3514                         if (IS_ERR(trans)) {
3515                                 ret = PTR_ERR(trans);
3516                                 goto out;
3517                         }
3518                         btrfs_debug(fs_info, "auto deleting %Lu",
3519                                     found_key.objectid);
3520                         ret = btrfs_del_orphan_item(trans, root,
3521                                                     found_key.objectid);
3522                         btrfs_end_transaction(trans);
3523                         if (ret)
3524                                 goto out;
3525                         continue;
3526                 }
3527
3528                 /*
3529                  * add this inode to the orphan list so btrfs_orphan_del does
3530                  * the proper thing when we hit it
3531                  */
3532                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3533                         &BTRFS_I(inode)->runtime_flags);
3534                 atomic_inc(&root->orphan_inodes);
3535
3536                 /* if we have links, this was a truncate, lets do that */
3537                 if (inode->i_nlink) {
3538                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3539                                 iput(inode);
3540                                 continue;
3541                         }
3542                         nr_truncate++;
3543
3544                         /* 1 for the orphan item deletion. */
3545                         trans = btrfs_start_transaction(root, 1);
3546                         if (IS_ERR(trans)) {
3547                                 iput(inode);
3548                                 ret = PTR_ERR(trans);
3549                                 goto out;
3550                         }
3551                         ret = btrfs_orphan_add(trans, inode);
3552                         btrfs_end_transaction(trans);
3553                         if (ret) {
3554                                 iput(inode);
3555                                 goto out;
3556                         }
3557
3558                         ret = btrfs_truncate(inode);
3559                         if (ret)
3560                                 btrfs_orphan_del(NULL, inode);
3561                 } else {
3562                         nr_unlink++;
3563                 }
3564
3565                 /* this will do delete_inode and everything for us */
3566                 iput(inode);
3567                 if (ret)
3568                         goto out;
3569         }
3570         /* release the path since we're done with it */
3571         btrfs_release_path(path);
3572
3573         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3574
3575         if (root->orphan_block_rsv)
3576                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3577                                         (u64)-1);
3578
3579         if (root->orphan_block_rsv ||
3580             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3581                 trans = btrfs_join_transaction(root);
3582                 if (!IS_ERR(trans))
3583                         btrfs_end_transaction(trans);
3584         }
3585
3586         if (nr_unlink)
3587                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3588         if (nr_truncate)
3589                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3590
3591 out:
3592         if (ret)
3593                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3594         btrfs_free_path(path);
3595         return ret;
3596 }
3597
3598 /*
3599  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3600  * don't find any xattrs, we know there can't be any acls.
3601  *
3602  * slot is the slot the inode is in, objectid is the objectid of the inode
3603  */
3604 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3605                                           int slot, u64 objectid,
3606                                           int *first_xattr_slot)
3607 {
3608         u32 nritems = btrfs_header_nritems(leaf);
3609         struct btrfs_key found_key;
3610         static u64 xattr_access = 0;
3611         static u64 xattr_default = 0;
3612         int scanned = 0;
3613
3614         if (!xattr_access) {
3615                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3616                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3617                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3618                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3619         }
3620
3621         slot++;
3622         *first_xattr_slot = -1;
3623         while (slot < nritems) {
3624                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3625
3626                 /* we found a different objectid, there must not be acls */
3627                 if (found_key.objectid != objectid)
3628                         return 0;
3629
3630                 /* we found an xattr, assume we've got an acl */
3631                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3632                         if (*first_xattr_slot == -1)
3633                                 *first_xattr_slot = slot;
3634                         if (found_key.offset == xattr_access ||
3635                             found_key.offset == xattr_default)
3636                                 return 1;
3637                 }
3638
3639                 /*
3640                  * we found a key greater than an xattr key, there can't
3641                  * be any acls later on
3642                  */
3643                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3644                         return 0;
3645
3646                 slot++;
3647                 scanned++;
3648
3649                 /*
3650                  * it goes inode, inode backrefs, xattrs, extents,
3651                  * so if there are a ton of hard links to an inode there can
3652                  * be a lot of backrefs.  Don't waste time searching too hard,
3653                  * this is just an optimization
3654                  */
3655                 if (scanned >= 8)
3656                         break;
3657         }
3658         /* we hit the end of the leaf before we found an xattr or
3659          * something larger than an xattr.  We have to assume the inode
3660          * has acls
3661          */
3662         if (*first_xattr_slot == -1)
3663                 *first_xattr_slot = slot;
3664         return 1;
3665 }
3666
3667 /*
3668  * read an inode from the btree into the in-memory inode
3669  */
3670 static int btrfs_read_locked_inode(struct inode *inode)
3671 {
3672         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3673         struct btrfs_path *path;
3674         struct extent_buffer *leaf;
3675         struct btrfs_inode_item *inode_item;
3676         struct btrfs_root *root = BTRFS_I(inode)->root;
3677         struct btrfs_key location;
3678         unsigned long ptr;
3679         int maybe_acls;
3680         u32 rdev;
3681         int ret;
3682         bool filled = false;
3683         int first_xattr_slot;
3684
3685         ret = btrfs_fill_inode(inode, &rdev);
3686         if (!ret)
3687                 filled = true;
3688
3689         path = btrfs_alloc_path();
3690         if (!path) {
3691                 ret = -ENOMEM;
3692                 goto make_bad;
3693         }
3694
3695         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3696
3697         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3698         if (ret) {
3699                 if (ret > 0)
3700                         ret = -ENOENT;
3701                 goto make_bad;
3702         }
3703
3704         leaf = path->nodes[0];
3705
3706         if (filled)
3707                 goto cache_index;
3708
3709         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3710                                     struct btrfs_inode_item);
3711         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3712         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3713         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3714         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3715         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3716
3717         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3718         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3719
3720         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3721         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3722
3723         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3724         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3725
3726         BTRFS_I(inode)->i_otime.tv_sec =
3727                 btrfs_timespec_sec(leaf, &inode_item->otime);
3728         BTRFS_I(inode)->i_otime.tv_nsec =
3729                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3730
3731         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3732         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3733         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3734
3735         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3736         inode->i_generation = BTRFS_I(inode)->generation;
3737         inode->i_rdev = 0;
3738         rdev = btrfs_inode_rdev(leaf, inode_item);
3739
3740         BTRFS_I(inode)->index_cnt = (u64)-1;
3741         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3742
3743 cache_index:
3744         /*
3745          * If we were modified in the current generation and evicted from memory
3746          * and then re-read we need to do a full sync since we don't have any
3747          * idea about which extents were modified before we were evicted from
3748          * cache.
3749          *
3750          * This is required for both inode re-read from disk and delayed inode
3751          * in delayed_nodes_tree.
3752          */
3753         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3754                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3755                         &BTRFS_I(inode)->runtime_flags);
3756
3757         /*
3758          * We don't persist the id of the transaction where an unlink operation
3759          * against the inode was last made. So here we assume the inode might
3760          * have been evicted, and therefore the exact value of last_unlink_trans
3761          * lost, and set it to last_trans to avoid metadata inconsistencies
3762          * between the inode and its parent if the inode is fsync'ed and the log
3763          * replayed. For example, in the scenario:
3764          *
3765          * touch mydir/foo
3766          * ln mydir/foo mydir/bar
3767          * sync
3768          * unlink mydir/bar
3769          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3770          * xfs_io -c fsync mydir/foo
3771          * <power failure>
3772          * mount fs, triggers fsync log replay
3773          *
3774          * We must make sure that when we fsync our inode foo we also log its
3775          * parent inode, otherwise after log replay the parent still has the
3776          * dentry with the "bar" name but our inode foo has a link count of 1
3777          * and doesn't have an inode ref with the name "bar" anymore.
3778          *
3779          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3780          * but it guarantees correctness at the expense of occasional full
3781          * transaction commits on fsync if our inode is a directory, or if our
3782          * inode is not a directory, logging its parent unnecessarily.
3783          */
3784         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3785
3786         path->slots[0]++;
3787         if (inode->i_nlink != 1 ||
3788             path->slots[0] >= btrfs_header_nritems(leaf))
3789                 goto cache_acl;
3790
3791         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3792         if (location.objectid != btrfs_ino(inode))
3793                 goto cache_acl;
3794
3795         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3796         if (location.type == BTRFS_INODE_REF_KEY) {
3797                 struct btrfs_inode_ref *ref;
3798
3799                 ref = (struct btrfs_inode_ref *)ptr;
3800                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3801         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3802                 struct btrfs_inode_extref *extref;
3803
3804                 extref = (struct btrfs_inode_extref *)ptr;
3805                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3806                                                                      extref);
3807         }
3808 cache_acl:
3809         /*
3810          * try to precache a NULL acl entry for files that don't have
3811          * any xattrs or acls
3812          */
3813         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3814                                            btrfs_ino(inode), &first_xattr_slot);
3815         if (first_xattr_slot != -1) {
3816                 path->slots[0] = first_xattr_slot;
3817                 ret = btrfs_load_inode_props(inode, path);
3818                 if (ret)
3819                         btrfs_err(fs_info,
3820                                   "error loading props for ino %llu (root %llu): %d",
3821                                   btrfs_ino(inode),
3822                                   root->root_key.objectid, ret);
3823         }
3824         btrfs_free_path(path);
3825
3826         if (!maybe_acls)
3827                 cache_no_acl(inode);
3828
3829         switch (inode->i_mode & S_IFMT) {
3830         case S_IFREG:
3831                 inode->i_mapping->a_ops = &btrfs_aops;
3832                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3833                 inode->i_fop = &btrfs_file_operations;
3834                 inode->i_op = &btrfs_file_inode_operations;
3835                 break;
3836         case S_IFDIR:
3837                 inode->i_fop = &btrfs_dir_file_operations;
3838                 inode->i_op = &btrfs_dir_inode_operations;
3839                 break;
3840         case S_IFLNK:
3841                 inode->i_op = &btrfs_symlink_inode_operations;
3842                 inode_nohighmem(inode);
3843                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3844                 break;
3845         default:
3846                 inode->i_op = &btrfs_special_inode_operations;
3847                 init_special_inode(inode, inode->i_mode, rdev);
3848                 break;
3849         }
3850
3851         btrfs_update_iflags(inode);
3852         return 0;
3853
3854 make_bad:
3855         btrfs_free_path(path);
3856         make_bad_inode(inode);
3857         return ret;
3858 }
3859
3860 /*
3861  * given a leaf and an inode, copy the inode fields into the leaf
3862  */
3863 static void fill_inode_item(struct btrfs_trans_handle *trans,
3864                             struct extent_buffer *leaf,
3865                             struct btrfs_inode_item *item,
3866                             struct inode *inode)
3867 {
3868         struct btrfs_map_token token;
3869
3870         btrfs_init_map_token(&token);
3871
3872         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3873         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3874         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3875                                    &token);
3876         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3877         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3878
3879         btrfs_set_token_timespec_sec(leaf, &item->atime,
3880                                      inode->i_atime.tv_sec, &token);
3881         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3882                                       inode->i_atime.tv_nsec, &token);
3883
3884         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3885                                      inode->i_mtime.tv_sec, &token);
3886         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3887                                       inode->i_mtime.tv_nsec, &token);
3888
3889         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3890                                      inode->i_ctime.tv_sec, &token);
3891         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3892                                       inode->i_ctime.tv_nsec, &token);
3893
3894         btrfs_set_token_timespec_sec(leaf, &item->otime,
3895                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3896         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3897                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3898
3899         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3900                                      &token);
3901         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3902                                          &token);
3903         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3904         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3905         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3906         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3907         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3908 }
3909
3910 /*
3911  * copy everything in the in-memory inode into the btree.
3912  */
3913 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3914                                 struct btrfs_root *root, struct inode *inode)
3915 {
3916         struct btrfs_inode_item *inode_item;
3917         struct btrfs_path *path;
3918         struct extent_buffer *leaf;
3919         int ret;
3920
3921         path = btrfs_alloc_path();
3922         if (!path)
3923                 return -ENOMEM;
3924
3925         path->leave_spinning = 1;
3926         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3927                                  1);
3928         if (ret) {
3929                 if (ret > 0)
3930                         ret = -ENOENT;
3931                 goto failed;
3932         }
3933
3934         leaf = path->nodes[0];
3935         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3936                                     struct btrfs_inode_item);
3937
3938         fill_inode_item(trans, leaf, inode_item, inode);
3939         btrfs_mark_buffer_dirty(leaf);
3940         btrfs_set_inode_last_trans(trans, inode);
3941         ret = 0;
3942 failed:
3943         btrfs_free_path(path);
3944         return ret;
3945 }
3946
3947 /*
3948  * copy everything in the in-memory inode into the btree.
3949  */
3950 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3951                                 struct btrfs_root *root, struct inode *inode)
3952 {
3953         struct btrfs_fs_info *fs_info = root->fs_info;
3954         int ret;
3955
3956         /*
3957          * If the inode is a free space inode, we can deadlock during commit
3958          * if we put it into the delayed code.
3959          *
3960          * The data relocation inode should also be directly updated
3961          * without delay
3962          */
3963         if (!btrfs_is_free_space_inode(inode)
3964             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3965             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3966                 btrfs_update_root_times(trans, root);
3967
3968                 ret = btrfs_delayed_update_inode(trans, root, inode);
3969                 if (!ret)
3970                         btrfs_set_inode_last_trans(trans, inode);
3971                 return ret;
3972         }
3973
3974         return btrfs_update_inode_item(trans, root, inode);
3975 }
3976
3977 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3978                                          struct btrfs_root *root,
3979                                          struct inode *inode)
3980 {
3981         int ret;
3982
3983         ret = btrfs_update_inode(trans, root, inode);
3984         if (ret == -ENOSPC)
3985                 return btrfs_update_inode_item(trans, root, inode);
3986         return ret;
3987 }
3988
3989 /*
3990  * unlink helper that gets used here in inode.c and in the tree logging
3991  * recovery code.  It remove a link in a directory with a given name, and
3992  * also drops the back refs in the inode to the directory
3993  */
3994 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3995                                 struct btrfs_root *root,
3996                                 struct inode *dir, struct inode *inode,
3997                                 const char *name, int name_len)
3998 {
3999         struct btrfs_fs_info *fs_info = root->fs_info;
4000         struct btrfs_path *path;
4001         int ret = 0;
4002         struct extent_buffer *leaf;
4003         struct btrfs_dir_item *di;
4004         struct btrfs_key key;
4005         u64 index;
4006         u64 ino = btrfs_ino(inode);
4007         u64 dir_ino = btrfs_ino(dir);
4008
4009         path = btrfs_alloc_path();
4010         if (!path) {
4011                 ret = -ENOMEM;
4012                 goto out;
4013         }
4014
4015         path->leave_spinning = 1;
4016         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4017                                     name, name_len, -1);
4018         if (IS_ERR(di)) {
4019                 ret = PTR_ERR(di);
4020                 goto err;
4021         }
4022         if (!di) {
4023                 ret = -ENOENT;
4024                 goto err;
4025         }
4026         leaf = path->nodes[0];
4027         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4028         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4029         if (ret)
4030                 goto err;
4031         btrfs_release_path(path);
4032
4033         /*
4034          * If we don't have dir index, we have to get it by looking up
4035          * the inode ref, since we get the inode ref, remove it directly,
4036          * it is unnecessary to do delayed deletion.
4037          *
4038          * But if we have dir index, needn't search inode ref to get it.
4039          * Since the inode ref is close to the inode item, it is better
4040          * that we delay to delete it, and just do this deletion when
4041          * we update the inode item.
4042          */
4043         if (BTRFS_I(inode)->dir_index) {
4044                 ret = btrfs_delayed_delete_inode_ref(inode);
4045                 if (!ret) {
4046                         index = BTRFS_I(inode)->dir_index;
4047                         goto skip_backref;
4048                 }
4049         }
4050
4051         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4052                                   dir_ino, &index);
4053         if (ret) {
4054                 btrfs_info(fs_info,
4055                         "failed to delete reference to %.*s, inode %llu parent %llu",
4056                         name_len, name, ino, dir_ino);
4057                 btrfs_abort_transaction(trans, ret);
4058                 goto err;
4059         }
4060 skip_backref:
4061         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4062         if (ret) {
4063                 btrfs_abort_transaction(trans, ret);
4064                 goto err;
4065         }
4066
4067         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
4068                                          inode, dir_ino);
4069         if (ret != 0 && ret != -ENOENT) {
4070                 btrfs_abort_transaction(trans, ret);
4071                 goto err;
4072         }
4073
4074         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4075                                            dir, index);
4076         if (ret == -ENOENT)
4077                 ret = 0;
4078         else if (ret)
4079                 btrfs_abort_transaction(trans, ret);
4080 err:
4081         btrfs_free_path(path);
4082         if (ret)
4083                 goto out;
4084
4085         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4086         inode_inc_iversion(inode);
4087         inode_inc_iversion(dir);
4088         inode->i_ctime = dir->i_mtime =
4089                 dir->i_ctime = current_time(inode);
4090         ret = btrfs_update_inode(trans, root, dir);
4091 out:
4092         return ret;
4093 }
4094
4095 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4096                        struct btrfs_root *root,
4097                        struct inode *dir, struct inode *inode,
4098                        const char *name, int name_len)
4099 {
4100         int ret;
4101         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4102         if (!ret) {
4103                 drop_nlink(inode);
4104                 ret = btrfs_update_inode(trans, root, inode);
4105         }
4106         return ret;
4107 }
4108
4109 /*
4110  * helper to start transaction for unlink and rmdir.
4111  *
4112  * unlink and rmdir are special in btrfs, they do not always free space, so
4113  * if we cannot make our reservations the normal way try and see if there is
4114  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4115  * allow the unlink to occur.
4116  */
4117 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4118 {
4119         struct btrfs_root *root = BTRFS_I(dir)->root;
4120
4121         /*
4122          * 1 for the possible orphan item
4123          * 1 for the dir item
4124          * 1 for the dir index
4125          * 1 for the inode ref
4126          * 1 for the inode
4127          */
4128         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4129 }
4130
4131 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4132 {
4133         struct btrfs_root *root = BTRFS_I(dir)->root;
4134         struct btrfs_trans_handle *trans;
4135         struct inode *inode = d_inode(dentry);
4136         int ret;
4137
4138         trans = __unlink_start_trans(dir);
4139         if (IS_ERR(trans))
4140                 return PTR_ERR(trans);
4141
4142         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4143
4144         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4145                                  dentry->d_name.name, dentry->d_name.len);
4146         if (ret)
4147                 goto out;
4148
4149         if (inode->i_nlink == 0) {
4150                 ret = btrfs_orphan_add(trans, inode);
4151                 if (ret)
4152                         goto out;
4153         }
4154
4155 out:
4156         btrfs_end_transaction(trans);
4157         btrfs_btree_balance_dirty(root->fs_info);
4158         return ret;
4159 }
4160
4161 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4162                         struct btrfs_root *root,
4163                         struct inode *dir, u64 objectid,
4164                         const char *name, int name_len)
4165 {
4166         struct btrfs_fs_info *fs_info = root->fs_info;
4167         struct btrfs_path *path;
4168         struct extent_buffer *leaf;
4169         struct btrfs_dir_item *di;
4170         struct btrfs_key key;
4171         u64 index;
4172         int ret;
4173         u64 dir_ino = btrfs_ino(dir);
4174
4175         path = btrfs_alloc_path();
4176         if (!path)
4177                 return -ENOMEM;
4178
4179         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4180                                    name, name_len, -1);
4181         if (IS_ERR_OR_NULL(di)) {
4182                 if (!di)
4183                         ret = -ENOENT;
4184                 else
4185                         ret = PTR_ERR(di);
4186                 goto out;
4187         }
4188
4189         leaf = path->nodes[0];
4190         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4191         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4192         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4193         if (ret) {
4194                 btrfs_abort_transaction(trans, ret);
4195                 goto out;
4196         }
4197         btrfs_release_path(path);
4198
4199         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4200                                  root->root_key.objectid, dir_ino,
4201                                  &index, name, name_len);
4202         if (ret < 0) {
4203                 if (ret != -ENOENT) {
4204                         btrfs_abort_transaction(trans, ret);
4205                         goto out;
4206                 }
4207                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4208                                                  name, name_len);
4209                 if (IS_ERR_OR_NULL(di)) {
4210                         if (!di)
4211                                 ret = -ENOENT;
4212                         else
4213                                 ret = PTR_ERR(di);
4214                         btrfs_abort_transaction(trans, ret);
4215                         goto out;
4216                 }
4217
4218                 leaf = path->nodes[0];
4219                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4220                 btrfs_release_path(path);
4221                 index = key.offset;
4222         }
4223         btrfs_release_path(path);
4224
4225         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4226         if (ret) {
4227                 btrfs_abort_transaction(trans, ret);
4228                 goto out;
4229         }
4230
4231         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4232         inode_inc_iversion(dir);
4233         dir->i_mtime = dir->i_ctime = current_time(dir);
4234         ret = btrfs_update_inode_fallback(trans, root, dir);
4235         if (ret)
4236                 btrfs_abort_transaction(trans, ret);
4237 out:
4238         btrfs_free_path(path);
4239         return ret;
4240 }
4241
4242 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4243 {
4244         struct inode *inode = d_inode(dentry);
4245         int err = 0;
4246         struct btrfs_root *root = BTRFS_I(dir)->root;
4247         struct btrfs_trans_handle *trans;
4248         u64 last_unlink_trans;
4249
4250         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4251                 return -ENOTEMPTY;
4252         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4253                 return -EPERM;
4254
4255         trans = __unlink_start_trans(dir);
4256         if (IS_ERR(trans))
4257                 return PTR_ERR(trans);
4258
4259         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4260                 err = btrfs_unlink_subvol(trans, root, dir,
4261                                           BTRFS_I(inode)->location.objectid,
4262                                           dentry->d_name.name,
4263                                           dentry->d_name.len);
4264                 goto out;
4265         }
4266
4267         err = btrfs_orphan_add(trans, inode);
4268         if (err)
4269                 goto out;
4270
4271         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4272
4273         /* now the directory is empty */
4274         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4275                                  dentry->d_name.name, dentry->d_name.len);
4276         if (!err) {
4277                 btrfs_i_size_write(inode, 0);
4278                 /*
4279                  * Propagate the last_unlink_trans value of the deleted dir to
4280                  * its parent directory. This is to prevent an unrecoverable
4281                  * log tree in the case we do something like this:
4282                  * 1) create dir foo
4283                  * 2) create snapshot under dir foo
4284                  * 3) delete the snapshot
4285                  * 4) rmdir foo
4286                  * 5) mkdir foo
4287                  * 6) fsync foo or some file inside foo
4288                  */
4289                 if (last_unlink_trans >= trans->transid)
4290                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4291         }
4292 out:
4293         btrfs_end_transaction(trans);
4294         btrfs_btree_balance_dirty(root->fs_info);
4295
4296         return err;
4297 }
4298
4299 static int truncate_space_check(struct btrfs_trans_handle *trans,
4300                                 struct btrfs_root *root,
4301                                 u64 bytes_deleted)
4302 {
4303         struct btrfs_fs_info *fs_info = root->fs_info;
4304         int ret;
4305
4306         /*
4307          * This is only used to apply pressure to the enospc system, we don't
4308          * intend to use this reservation at all.
4309          */
4310         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4311         bytes_deleted *= fs_info->nodesize;
4312         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4313                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4314         if (!ret) {
4315                 trace_btrfs_space_reservation(fs_info, "transaction",
4316                                               trans->transid,
4317                                               bytes_deleted, 1);
4318                 trans->bytes_reserved += bytes_deleted;
4319         }
4320         return ret;
4321
4322 }
4323
4324 static int truncate_inline_extent(struct inode *inode,
4325                                   struct btrfs_path *path,
4326                                   struct btrfs_key *found_key,
4327                                   const u64 item_end,
4328                                   const u64 new_size)
4329 {
4330         struct extent_buffer *leaf = path->nodes[0];
4331         int slot = path->slots[0];
4332         struct btrfs_file_extent_item *fi;
4333         u32 size = (u32)(new_size - found_key->offset);
4334         struct btrfs_root *root = BTRFS_I(inode)->root;
4335
4336         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4337
4338         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4339                 loff_t offset = new_size;
4340                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4341
4342                 /*
4343                  * Zero out the remaining of the last page of our inline extent,
4344                  * instead of directly truncating our inline extent here - that
4345                  * would be much more complex (decompressing all the data, then
4346                  * compressing the truncated data, which might be bigger than
4347                  * the size of the inline extent, resize the extent, etc).
4348                  * We release the path because to get the page we might need to
4349                  * read the extent item from disk (data not in the page cache).
4350                  */
4351                 btrfs_release_path(path);
4352                 return btrfs_truncate_block(inode, offset, page_end - offset,
4353                                         0);
4354         }
4355
4356         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4357         size = btrfs_file_extent_calc_inline_size(size);
4358         btrfs_truncate_item(root->fs_info, path, size, 1);
4359
4360         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4361                 inode_sub_bytes(inode, item_end + 1 - new_size);
4362
4363         return 0;
4364 }
4365
4366 /*
4367  * this can truncate away extent items, csum items and directory items.
4368  * It starts at a high offset and removes keys until it can't find
4369  * any higher than new_size
4370  *
4371  * csum items that cross the new i_size are truncated to the new size
4372  * as well.
4373  *
4374  * min_type is the minimum key type to truncate down to.  If set to 0, this
4375  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4376  */
4377 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4378                                struct btrfs_root *root,
4379                                struct inode *inode,
4380                                u64 new_size, u32 min_type)
4381 {
4382         struct btrfs_fs_info *fs_info = root->fs_info;
4383         struct btrfs_path *path;
4384         struct extent_buffer *leaf;
4385         struct btrfs_file_extent_item *fi;
4386         struct btrfs_key key;
4387         struct btrfs_key found_key;
4388         u64 extent_start = 0;
4389         u64 extent_num_bytes = 0;
4390         u64 extent_offset = 0;
4391         u64 item_end = 0;
4392         u64 last_size = new_size;
4393         u32 found_type = (u8)-1;
4394         int found_extent;
4395         int del_item;
4396         int pending_del_nr = 0;
4397         int pending_del_slot = 0;
4398         int extent_type = -1;
4399         int ret;
4400         int err = 0;
4401         u64 ino = btrfs_ino(inode);
4402         u64 bytes_deleted = 0;
4403         bool be_nice = 0;
4404         bool should_throttle = 0;
4405         bool should_end = 0;
4406
4407         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4408
4409         /*
4410          * for non-free space inodes and ref cows, we want to back off from
4411          * time to time
4412          */
4413         if (!btrfs_is_free_space_inode(inode) &&
4414             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4415                 be_nice = 1;
4416
4417         path = btrfs_alloc_path();
4418         if (!path)
4419                 return -ENOMEM;
4420         path->reada = READA_BACK;
4421
4422         /*
4423          * We want to drop from the next block forward in case this new size is
4424          * not block aligned since we will be keeping the last block of the
4425          * extent just the way it is.
4426          */
4427         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4428             root == fs_info->tree_root)
4429                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4430                                         fs_info->sectorsize),
4431                                         (u64)-1, 0);
4432
4433         /*
4434          * This function is also used to drop the items in the log tree before
4435          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4436          * it is used to drop the loged items. So we shouldn't kill the delayed
4437          * items.
4438          */
4439         if (min_type == 0 && root == BTRFS_I(inode)->root)
4440                 btrfs_kill_delayed_inode_items(inode);
4441
4442         key.objectid = ino;
4443         key.offset = (u64)-1;
4444         key.type = (u8)-1;
4445
4446 search_again:
4447         /*
4448          * with a 16K leaf size and 128MB extents, you can actually queue
4449          * up a huge file in a single leaf.  Most of the time that
4450          * bytes_deleted is > 0, it will be huge by the time we get here
4451          */
4452         if (be_nice && bytes_deleted > SZ_32M) {
4453                 if (btrfs_should_end_transaction(trans)) {
4454                         err = -EAGAIN;
4455                         goto error;
4456                 }
4457         }
4458
4459
4460         path->leave_spinning = 1;
4461         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4462         if (ret < 0) {
4463                 err = ret;
4464                 goto out;
4465         }
4466
4467         if (ret > 0) {
4468                 /* there are no items in the tree for us to truncate, we're
4469                  * done
4470                  */
4471                 if (path->slots[0] == 0)
4472                         goto out;
4473                 path->slots[0]--;
4474         }
4475
4476         while (1) {
4477                 fi = NULL;
4478                 leaf = path->nodes[0];
4479                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4480                 found_type = found_key.type;
4481
4482                 if (found_key.objectid != ino)
4483                         break;
4484
4485                 if (found_type < min_type)
4486                         break;
4487
4488                 item_end = found_key.offset;
4489                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4490                         fi = btrfs_item_ptr(leaf, path->slots[0],
4491                                             struct btrfs_file_extent_item);
4492                         extent_type = btrfs_file_extent_type(leaf, fi);
4493                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4494                                 item_end +=
4495                                     btrfs_file_extent_num_bytes(leaf, fi);
4496                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4497                                 item_end += btrfs_file_extent_inline_len(leaf,
4498                                                          path->slots[0], fi);
4499                         }
4500                         item_end--;
4501                 }
4502                 if (found_type > min_type) {
4503                         del_item = 1;
4504                 } else {
4505                         if (item_end < new_size) {
4506                                 /*
4507                                  * With NO_HOLES mode, for the following mapping
4508                                  *
4509                                  * [0-4k][hole][8k-12k]
4510                                  *
4511                                  * if truncating isize down to 6k, it ends up
4512                                  * isize being 8k.
4513                                  */
4514                                 if (btrfs_fs_incompat(root->fs_info, NO_HOLES))
4515                                         last_size = new_size;
4516                                 break;
4517                         }
4518                         if (found_key.offset >= new_size)
4519                                 del_item = 1;
4520                         else
4521                                 del_item = 0;
4522                 }
4523                 found_extent = 0;
4524                 /* FIXME, shrink the extent if the ref count is only 1 */
4525                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4526                         goto delete;
4527
4528                 if (del_item)
4529                         last_size = found_key.offset;
4530                 else
4531                         last_size = new_size;
4532
4533                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4534                         u64 num_dec;
4535                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4536                         if (!del_item) {
4537                                 u64 orig_num_bytes =
4538                                         btrfs_file_extent_num_bytes(leaf, fi);
4539                                 extent_num_bytes = ALIGN(new_size -
4540                                                 found_key.offset,
4541                                                 fs_info->sectorsize);
4542                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4543                                                          extent_num_bytes);
4544                                 num_dec = (orig_num_bytes -
4545                                            extent_num_bytes);
4546                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4547                                              &root->state) &&
4548                                     extent_start != 0)
4549                                         inode_sub_bytes(inode, num_dec);
4550                                 btrfs_mark_buffer_dirty(leaf);
4551                         } else {
4552                                 extent_num_bytes =
4553                                         btrfs_file_extent_disk_num_bytes(leaf,
4554                                                                          fi);
4555                                 extent_offset = found_key.offset -
4556                                         btrfs_file_extent_offset(leaf, fi);
4557
4558                                 /* FIXME blocksize != 4096 */
4559                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4560                                 if (extent_start != 0) {
4561                                         found_extent = 1;
4562                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4563                                                      &root->state))
4564                                                 inode_sub_bytes(inode, num_dec);
4565                                 }
4566                         }
4567                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4568                         /*
4569                          * we can't truncate inline items that have had
4570                          * special encodings
4571                          */
4572                         if (!del_item &&
4573                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4574                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4575
4576                                 /*
4577                                  * Need to release path in order to truncate a
4578                                  * compressed extent. So delete any accumulated
4579                                  * extent items so far.
4580                                  */
4581                                 if (btrfs_file_extent_compression(leaf, fi) !=
4582                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4583                                         err = btrfs_del_items(trans, root, path,
4584                                                               pending_del_slot,
4585                                                               pending_del_nr);
4586                                         if (err) {
4587                                                 btrfs_abort_transaction(trans,
4588                                                                         err);
4589                                                 goto error;
4590                                         }
4591                                         pending_del_nr = 0;
4592                                 }
4593
4594                                 err = truncate_inline_extent(inode, path,
4595                                                              &found_key,
4596                                                              item_end,
4597                                                              new_size);
4598                                 if (err) {
4599                                         btrfs_abort_transaction(trans, err);
4600                                         goto error;
4601                                 }
4602                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4603                                             &root->state)) {
4604                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4605                         }
4606                 }
4607 delete:
4608                 if (del_item) {
4609                         if (!pending_del_nr) {
4610                                 /* no pending yet, add ourselves */
4611                                 pending_del_slot = path->slots[0];
4612                                 pending_del_nr = 1;
4613                         } else if (pending_del_nr &&
4614                                    path->slots[0] + 1 == pending_del_slot) {
4615                                 /* hop on the pending chunk */
4616                                 pending_del_nr++;
4617                                 pending_del_slot = path->slots[0];
4618                         } else {
4619                                 BUG();
4620                         }
4621                 } else {
4622                         break;
4623                 }
4624                 should_throttle = 0;
4625
4626                 if (found_extent &&
4627                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4628                      root == fs_info->tree_root)) {
4629                         btrfs_set_path_blocking(path);
4630                         bytes_deleted += extent_num_bytes;
4631                         ret = btrfs_free_extent(trans, fs_info, extent_start,
4632                                                 extent_num_bytes, 0,
4633                                                 btrfs_header_owner(leaf),
4634                                                 ino, extent_offset);
4635                         BUG_ON(ret);
4636                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4637                                 btrfs_async_run_delayed_refs(fs_info,
4638                                         trans->delayed_ref_updates * 2,
4639                                         trans->transid, 0);
4640                         if (be_nice) {
4641                                 if (truncate_space_check(trans, root,
4642                                                          extent_num_bytes)) {
4643                                         should_end = 1;
4644                                 }
4645                                 if (btrfs_should_throttle_delayed_refs(trans,
4646                                                                        fs_info))
4647                                         should_throttle = 1;
4648                         }
4649                 }
4650
4651                 if (found_type == BTRFS_INODE_ITEM_KEY)
4652                         break;
4653
4654                 if (path->slots[0] == 0 ||
4655                     path->slots[0] != pending_del_slot ||
4656                     should_throttle || should_end) {
4657                         if (pending_del_nr) {
4658                                 ret = btrfs_del_items(trans, root, path,
4659                                                 pending_del_slot,
4660                                                 pending_del_nr);
4661                                 if (ret) {
4662                                         btrfs_abort_transaction(trans, ret);
4663                                         goto error;
4664                                 }
4665                                 pending_del_nr = 0;
4666                         }
4667                         btrfs_release_path(path);
4668                         if (should_throttle) {
4669                                 unsigned long updates = trans->delayed_ref_updates;
4670                                 if (updates) {
4671                                         trans->delayed_ref_updates = 0;
4672                                         ret = btrfs_run_delayed_refs(trans,
4673                                                                    fs_info,
4674                                                                    updates * 2);
4675                                         if (ret && !err)
4676                                                 err = ret;
4677                                 }
4678                         }
4679                         /*
4680                          * if we failed to refill our space rsv, bail out
4681                          * and let the transaction restart
4682                          */
4683                         if (should_end) {
4684                                 err = -EAGAIN;
4685                                 goto error;
4686                         }
4687                         goto search_again;
4688                 } else {
4689                         path->slots[0]--;
4690                 }
4691         }
4692 out:
4693         if (pending_del_nr) {
4694                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4695                                       pending_del_nr);
4696                 if (ret)
4697                         btrfs_abort_transaction(trans, ret);
4698         }
4699 error:
4700         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4701                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4702
4703         btrfs_free_path(path);
4704
4705         if (be_nice && bytes_deleted > SZ_32M) {
4706                 unsigned long updates = trans->delayed_ref_updates;
4707                 if (updates) {
4708                         trans->delayed_ref_updates = 0;
4709                         ret = btrfs_run_delayed_refs(trans, fs_info,
4710                                                      updates * 2);
4711                         if (ret && !err)
4712                                 err = ret;
4713                 }
4714         }
4715         return err;
4716 }
4717
4718 /*
4719  * btrfs_truncate_block - read, zero a chunk and write a block
4720  * @inode - inode that we're zeroing
4721  * @from - the offset to start zeroing
4722  * @len - the length to zero, 0 to zero the entire range respective to the
4723  *      offset
4724  * @front - zero up to the offset instead of from the offset on
4725  *
4726  * This will find the block for the "from" offset and cow the block and zero the
4727  * part we want to zero.  This is used with truncate and hole punching.
4728  */
4729 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4730                         int front)
4731 {
4732         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4733         struct address_space *mapping = inode->i_mapping;
4734         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4735         struct btrfs_ordered_extent *ordered;
4736         struct extent_state *cached_state = NULL;
4737         char *kaddr;
4738         u32 blocksize = fs_info->sectorsize;
4739         pgoff_t index = from >> PAGE_SHIFT;
4740         unsigned offset = from & (blocksize - 1);
4741         struct page *page;
4742         gfp_t mask = btrfs_alloc_write_mask(mapping);
4743         int ret = 0;
4744         u64 block_start;
4745         u64 block_end;
4746
4747         if ((offset & (blocksize - 1)) == 0 &&
4748             (!len || ((len & (blocksize - 1)) == 0)))
4749                 goto out;
4750
4751         ret = btrfs_delalloc_reserve_space(inode,
4752                         round_down(from, blocksize), blocksize);
4753         if (ret)
4754                 goto out;
4755
4756 again:
4757         page = find_or_create_page(mapping, index, mask);
4758         if (!page) {
4759                 btrfs_delalloc_release_space(inode,
4760                                 round_down(from, blocksize),
4761                                 blocksize);
4762                 ret = -ENOMEM;
4763                 goto out;
4764         }
4765
4766         block_start = round_down(from, blocksize);
4767         block_end = block_start + blocksize - 1;
4768
4769         if (!PageUptodate(page)) {
4770                 ret = btrfs_readpage(NULL, page);
4771                 lock_page(page);
4772                 if (page->mapping != mapping) {
4773                         unlock_page(page);
4774                         put_page(page);
4775                         goto again;
4776                 }
4777                 if (!PageUptodate(page)) {
4778                         ret = -EIO;
4779                         goto out_unlock;
4780                 }
4781         }
4782         wait_on_page_writeback(page);
4783
4784         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4785         set_page_extent_mapped(page);
4786
4787         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4788         if (ordered) {
4789                 unlock_extent_cached(io_tree, block_start, block_end,
4790                                      &cached_state, GFP_NOFS);
4791                 unlock_page(page);
4792                 put_page(page);
4793                 btrfs_start_ordered_extent(inode, ordered, 1);
4794                 btrfs_put_ordered_extent(ordered);
4795                 goto again;
4796         }
4797
4798         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4799                           EXTENT_DIRTY | EXTENT_DELALLOC |
4800                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4801                           0, 0, &cached_state, GFP_NOFS);
4802
4803         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4804                                         &cached_state, 0);
4805         if (ret) {
4806                 unlock_extent_cached(io_tree, block_start, block_end,
4807                                      &cached_state, GFP_NOFS);
4808                 goto out_unlock;
4809         }
4810
4811         if (offset != blocksize) {
4812                 if (!len)
4813                         len = blocksize - offset;
4814                 kaddr = kmap(page);
4815                 if (front)
4816                         memset(kaddr + (block_start - page_offset(page)),
4817                                 0, offset);
4818                 else
4819                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4820                                 0, len);
4821                 flush_dcache_page(page);
4822                 kunmap(page);
4823         }
4824         ClearPageChecked(page);
4825         set_page_dirty(page);
4826         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4827                              GFP_NOFS);
4828
4829 out_unlock:
4830         if (ret)
4831                 btrfs_delalloc_release_space(inode, block_start,
4832                                              blocksize);
4833         unlock_page(page);
4834         put_page(page);
4835 out:
4836         return ret;
4837 }
4838
4839 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4840                              u64 offset, u64 len)
4841 {
4842         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4843         struct btrfs_trans_handle *trans;
4844         int ret;
4845
4846         /*
4847          * Still need to make sure the inode looks like it's been updated so
4848          * that any holes get logged if we fsync.
4849          */
4850         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4851                 BTRFS_I(inode)->last_trans = fs_info->generation;
4852                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4853                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4854                 return 0;
4855         }
4856
4857         /*
4858          * 1 - for the one we're dropping
4859          * 1 - for the one we're adding
4860          * 1 - for updating the inode.
4861          */
4862         trans = btrfs_start_transaction(root, 3);
4863         if (IS_ERR(trans))
4864                 return PTR_ERR(trans);
4865
4866         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4867         if (ret) {
4868                 btrfs_abort_transaction(trans, ret);
4869                 btrfs_end_transaction(trans);
4870                 return ret;
4871         }
4872
4873         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4874                                        0, 0, len, 0, len, 0, 0, 0);
4875         if (ret)
4876                 btrfs_abort_transaction(trans, ret);
4877         else
4878                 btrfs_update_inode(trans, root, inode);
4879         btrfs_end_transaction(trans);
4880         return ret;
4881 }
4882
4883 /*
4884  * This function puts in dummy file extents for the area we're creating a hole
4885  * for.  So if we are truncating this file to a larger size we need to insert
4886  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4887  * the range between oldsize and size
4888  */
4889 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4890 {
4891         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4892         struct btrfs_root *root = BTRFS_I(inode)->root;
4893         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4894         struct extent_map *em = NULL;
4895         struct extent_state *cached_state = NULL;
4896         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4897         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4898         u64 block_end = ALIGN(size, fs_info->sectorsize);
4899         u64 last_byte;
4900         u64 cur_offset;
4901         u64 hole_size;
4902         int err = 0;
4903
4904         /*
4905          * If our size started in the middle of a block we need to zero out the
4906          * rest of the block before we expand the i_size, otherwise we could
4907          * expose stale data.
4908          */
4909         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4910         if (err)
4911                 return err;
4912
4913         if (size <= hole_start)
4914                 return 0;
4915
4916         while (1) {
4917                 struct btrfs_ordered_extent *ordered;
4918
4919                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4920                                  &cached_state);
4921                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4922                                                      block_end - hole_start);
4923                 if (!ordered)
4924                         break;
4925                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4926                                      &cached_state, GFP_NOFS);
4927                 btrfs_start_ordered_extent(inode, ordered, 1);
4928                 btrfs_put_ordered_extent(ordered);
4929         }
4930
4931         cur_offset = hole_start;
4932         while (1) {
4933                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4934                                 block_end - cur_offset, 0);
4935                 if (IS_ERR(em)) {
4936                         err = PTR_ERR(em);
4937                         em = NULL;
4938                         break;
4939                 }
4940                 last_byte = min(extent_map_end(em), block_end);
4941                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4942                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4943                         struct extent_map *hole_em;
4944                         hole_size = last_byte - cur_offset;
4945
4946                         err = maybe_insert_hole(root, inode, cur_offset,
4947                                                 hole_size);
4948                         if (err)
4949                                 break;
4950                         btrfs_drop_extent_cache(inode, cur_offset,
4951                                                 cur_offset + hole_size - 1, 0);
4952                         hole_em = alloc_extent_map();
4953                         if (!hole_em) {
4954                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4955                                         &BTRFS_I(inode)->runtime_flags);
4956                                 goto next;
4957                         }
4958                         hole_em->start = cur_offset;
4959                         hole_em->len = hole_size;
4960                         hole_em->orig_start = cur_offset;
4961
4962                         hole_em->block_start = EXTENT_MAP_HOLE;
4963                         hole_em->block_len = 0;
4964                         hole_em->orig_block_len = 0;
4965                         hole_em->ram_bytes = hole_size;
4966                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
4967                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4968                         hole_em->generation = fs_info->generation;
4969
4970                         while (1) {
4971                                 write_lock(&em_tree->lock);
4972                                 err = add_extent_mapping(em_tree, hole_em, 1);
4973                                 write_unlock(&em_tree->lock);
4974                                 if (err != -EEXIST)
4975                                         break;
4976                                 btrfs_drop_extent_cache(inode, cur_offset,
4977                                                         cur_offset +
4978                                                         hole_size - 1, 0);
4979                         }
4980                         free_extent_map(hole_em);
4981                 }
4982 next:
4983                 free_extent_map(em);
4984                 em = NULL;
4985                 cur_offset = last_byte;
4986                 if (cur_offset >= block_end)
4987                         break;
4988         }
4989         free_extent_map(em);
4990         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4991                              GFP_NOFS);
4992         return err;
4993 }
4994
4995 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4996 {
4997         struct btrfs_root *root = BTRFS_I(inode)->root;
4998         struct btrfs_trans_handle *trans;
4999         loff_t oldsize = i_size_read(inode);
5000         loff_t newsize = attr->ia_size;
5001         int mask = attr->ia_valid;
5002         int ret;
5003
5004         /*
5005          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5006          * special case where we need to update the times despite not having
5007          * these flags set.  For all other operations the VFS set these flags
5008          * explicitly if it wants a timestamp update.
5009          */
5010         if (newsize != oldsize) {
5011                 inode_inc_iversion(inode);
5012                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5013                         inode->i_ctime = inode->i_mtime =
5014                                 current_time(inode);
5015         }
5016
5017         if (newsize > oldsize) {
5018                 /*
5019                  * Don't do an expanding truncate while snapshoting is ongoing.
5020                  * This is to ensure the snapshot captures a fully consistent
5021                  * state of this file - if the snapshot captures this expanding
5022                  * truncation, it must capture all writes that happened before
5023                  * this truncation.
5024                  */
5025                 btrfs_wait_for_snapshot_creation(root);
5026                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5027                 if (ret) {
5028                         btrfs_end_write_no_snapshoting(root);
5029                         return ret;
5030                 }
5031
5032                 trans = btrfs_start_transaction(root, 1);
5033                 if (IS_ERR(trans)) {
5034                         btrfs_end_write_no_snapshoting(root);
5035                         return PTR_ERR(trans);
5036                 }
5037
5038                 i_size_write(inode, newsize);
5039                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5040                 pagecache_isize_extended(inode, oldsize, newsize);
5041                 ret = btrfs_update_inode(trans, root, inode);
5042                 btrfs_end_write_no_snapshoting(root);
5043                 btrfs_end_transaction(trans);
5044         } else {
5045
5046                 /*
5047                  * We're truncating a file that used to have good data down to
5048                  * zero. Make sure it gets into the ordered flush list so that
5049                  * any new writes get down to disk quickly.
5050                  */
5051                 if (newsize == 0)
5052                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5053                                 &BTRFS_I(inode)->runtime_flags);
5054
5055                 /*
5056                  * 1 for the orphan item we're going to add
5057                  * 1 for the orphan item deletion.
5058                  */
5059                 trans = btrfs_start_transaction(root, 2);
5060                 if (IS_ERR(trans))
5061                         return PTR_ERR(trans);
5062
5063                 /*
5064                  * We need to do this in case we fail at _any_ point during the
5065                  * actual truncate.  Once we do the truncate_setsize we could
5066                  * invalidate pages which forces any outstanding ordered io to
5067                  * be instantly completed which will give us extents that need
5068                  * to be truncated.  If we fail to get an orphan inode down we
5069                  * could have left over extents that were never meant to live,
5070                  * so we need to guarantee from this point on that everything
5071                  * will be consistent.
5072                  */
5073                 ret = btrfs_orphan_add(trans, inode);
5074                 btrfs_end_transaction(trans);
5075                 if (ret)
5076                         return ret;
5077
5078                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5079                 truncate_setsize(inode, newsize);
5080
5081                 /* Disable nonlocked read DIO to avoid the end less truncate */
5082                 btrfs_inode_block_unlocked_dio(inode);
5083                 inode_dio_wait(inode);
5084                 btrfs_inode_resume_unlocked_dio(inode);
5085
5086                 ret = btrfs_truncate(inode);
5087                 if (ret && inode->i_nlink) {
5088                         int err;
5089
5090                         /*
5091                          * failed to truncate, disk_i_size is only adjusted down
5092                          * as we remove extents, so it should represent the true
5093                          * size of the inode, so reset the in memory size and
5094                          * delete our orphan entry.
5095                          */
5096                         trans = btrfs_join_transaction(root);
5097                         if (IS_ERR(trans)) {
5098                                 btrfs_orphan_del(NULL, inode);
5099                                 return ret;
5100                         }
5101                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5102                         err = btrfs_orphan_del(trans, inode);
5103                         if (err)
5104                                 btrfs_abort_transaction(trans, err);
5105                         btrfs_end_transaction(trans);
5106                 }
5107         }
5108
5109         return ret;
5110 }
5111
5112 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5113 {
5114         struct inode *inode = d_inode(dentry);
5115         struct btrfs_root *root = BTRFS_I(inode)->root;
5116         int err;
5117
5118         if (btrfs_root_readonly(root))
5119                 return -EROFS;
5120
5121         err = setattr_prepare(dentry, attr);
5122         if (err)
5123                 return err;
5124
5125         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5126                 err = btrfs_setsize(inode, attr);
5127                 if (err)
5128                         return err;
5129         }
5130
5131         if (attr->ia_valid) {
5132                 setattr_copy(inode, attr);
5133                 inode_inc_iversion(inode);
5134                 err = btrfs_dirty_inode(inode);
5135
5136                 if (!err && attr->ia_valid & ATTR_MODE)
5137                         err = posix_acl_chmod(inode, inode->i_mode);
5138         }
5139
5140         return err;
5141 }
5142
5143 /*
5144  * While truncating the inode pages during eviction, we get the VFS calling
5145  * btrfs_invalidatepage() against each page of the inode. This is slow because
5146  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5147  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5148  * extent_state structures over and over, wasting lots of time.
5149  *
5150  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5151  * those expensive operations on a per page basis and do only the ordered io
5152  * finishing, while we release here the extent_map and extent_state structures,
5153  * without the excessive merging and splitting.
5154  */
5155 static void evict_inode_truncate_pages(struct inode *inode)
5156 {
5157         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5158         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5159         struct rb_node *node;
5160
5161         ASSERT(inode->i_state & I_FREEING);
5162         truncate_inode_pages_final(&inode->i_data);
5163
5164         write_lock(&map_tree->lock);
5165         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5166                 struct extent_map *em;
5167
5168                 node = rb_first(&map_tree->map);
5169                 em = rb_entry(node, struct extent_map, rb_node);
5170                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5171                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5172                 remove_extent_mapping(map_tree, em);
5173                 free_extent_map(em);
5174                 if (need_resched()) {
5175                         write_unlock(&map_tree->lock);
5176                         cond_resched();
5177                         write_lock(&map_tree->lock);
5178                 }
5179         }
5180         write_unlock(&map_tree->lock);
5181
5182         /*
5183          * Keep looping until we have no more ranges in the io tree.
5184          * We can have ongoing bios started by readpages (called from readahead)
5185          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5186          * still in progress (unlocked the pages in the bio but did not yet
5187          * unlocked the ranges in the io tree). Therefore this means some
5188          * ranges can still be locked and eviction started because before
5189          * submitting those bios, which are executed by a separate task (work
5190          * queue kthread), inode references (inode->i_count) were not taken
5191          * (which would be dropped in the end io callback of each bio).
5192          * Therefore here we effectively end up waiting for those bios and
5193          * anyone else holding locked ranges without having bumped the inode's
5194          * reference count - if we don't do it, when they access the inode's
5195          * io_tree to unlock a range it may be too late, leading to an
5196          * use-after-free issue.
5197          */
5198         spin_lock(&io_tree->lock);
5199         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5200                 struct extent_state *state;
5201                 struct extent_state *cached_state = NULL;
5202                 u64 start;
5203                 u64 end;
5204
5205                 node = rb_first(&io_tree->state);
5206                 state = rb_entry(node, struct extent_state, rb_node);
5207                 start = state->start;
5208                 end = state->end;
5209                 spin_unlock(&io_tree->lock);
5210
5211                 lock_extent_bits(io_tree, start, end, &cached_state);
5212
5213                 /*
5214                  * If still has DELALLOC flag, the extent didn't reach disk,
5215                  * and its reserved space won't be freed by delayed_ref.
5216                  * So we need to free its reserved space here.
5217                  * (Refer to comment in btrfs_invalidatepage, case 2)
5218                  *
5219                  * Note, end is the bytenr of last byte, so we need + 1 here.
5220                  */
5221                 if (state->state & EXTENT_DELALLOC)
5222                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5223
5224                 clear_extent_bit(io_tree, start, end,
5225                                  EXTENT_LOCKED | EXTENT_DIRTY |
5226                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5227                                  EXTENT_DEFRAG, 1, 1,
5228                                  &cached_state, GFP_NOFS);
5229
5230                 cond_resched();
5231                 spin_lock(&io_tree->lock);
5232         }
5233         spin_unlock(&io_tree->lock);
5234 }
5235
5236 void btrfs_evict_inode(struct inode *inode)
5237 {
5238         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5239         struct btrfs_trans_handle *trans;
5240         struct btrfs_root *root = BTRFS_I(inode)->root;
5241         struct btrfs_block_rsv *rsv, *global_rsv;
5242         int steal_from_global = 0;
5243         u64 min_size;
5244         int ret;
5245
5246         trace_btrfs_inode_evict(inode);
5247
5248         if (!root) {
5249                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5250                 return;
5251         }
5252
5253         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5254
5255         evict_inode_truncate_pages(inode);
5256
5257         if (inode->i_nlink &&
5258             ((btrfs_root_refs(&root->root_item) != 0 &&
5259               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5260              btrfs_is_free_space_inode(inode)))
5261                 goto no_delete;
5262
5263         if (is_bad_inode(inode)) {
5264                 btrfs_orphan_del(NULL, inode);
5265                 goto no_delete;
5266         }
5267         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5268         if (!special_file(inode->i_mode))
5269                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5270
5271         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5272
5273         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5274                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5275                                  &BTRFS_I(inode)->runtime_flags));
5276                 goto no_delete;
5277         }
5278
5279         if (inode->i_nlink > 0) {
5280                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5281                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5282                 goto no_delete;
5283         }
5284
5285         ret = btrfs_commit_inode_delayed_inode(inode);
5286         if (ret) {
5287                 btrfs_orphan_del(NULL, inode);
5288                 goto no_delete;
5289         }
5290
5291         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5292         if (!rsv) {
5293                 btrfs_orphan_del(NULL, inode);
5294                 goto no_delete;
5295         }
5296         rsv->size = min_size;
5297         rsv->failfast = 1;
5298         global_rsv = &fs_info->global_block_rsv;
5299
5300         btrfs_i_size_write(inode, 0);
5301
5302         /*
5303          * This is a bit simpler than btrfs_truncate since we've already
5304          * reserved our space for our orphan item in the unlink, so we just
5305          * need to reserve some slack space in case we add bytes and update
5306          * inode item when doing the truncate.
5307          */
5308         while (1) {
5309                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5310                                              BTRFS_RESERVE_FLUSH_LIMIT);
5311
5312                 /*
5313                  * Try and steal from the global reserve since we will
5314                  * likely not use this space anyway, we want to try as
5315                  * hard as possible to get this to work.
5316                  */
5317                 if (ret)
5318                         steal_from_global++;
5319                 else
5320                         steal_from_global = 0;
5321                 ret = 0;
5322
5323                 /*
5324                  * steal_from_global == 0: we reserved stuff, hooray!
5325                  * steal_from_global == 1: we didn't reserve stuff, boo!
5326                  * steal_from_global == 2: we've committed, still not a lot of
5327                  * room but maybe we'll have room in the global reserve this
5328                  * time.
5329                  * steal_from_global == 3: abandon all hope!
5330                  */
5331                 if (steal_from_global > 2) {
5332                         btrfs_warn(fs_info,
5333                                    "Could not get space for a delete, will truncate on mount %d",
5334                                    ret);
5335                         btrfs_orphan_del(NULL, inode);
5336                         btrfs_free_block_rsv(fs_info, rsv);
5337                         goto no_delete;
5338                 }
5339
5340                 trans = btrfs_join_transaction(root);
5341                 if (IS_ERR(trans)) {
5342                         btrfs_orphan_del(NULL, inode);
5343                         btrfs_free_block_rsv(fs_info, rsv);
5344                         goto no_delete;
5345                 }
5346
5347                 /*
5348                  * We can't just steal from the global reserve, we need to make
5349                  * sure there is room to do it, if not we need to commit and try
5350                  * again.
5351                  */
5352                 if (steal_from_global) {
5353                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5354                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5355                                                               min_size, 0);
5356                         else
5357                                 ret = -ENOSPC;
5358                 }
5359
5360                 /*
5361                  * Couldn't steal from the global reserve, we have too much
5362                  * pending stuff built up, commit the transaction and try it
5363                  * again.
5364                  */
5365                 if (ret) {
5366                         ret = btrfs_commit_transaction(trans);
5367                         if (ret) {
5368                                 btrfs_orphan_del(NULL, inode);
5369                                 btrfs_free_block_rsv(fs_info, rsv);
5370                                 goto no_delete;
5371                         }
5372                         continue;
5373                 } else {
5374                         steal_from_global = 0;
5375                 }
5376
5377                 trans->block_rsv = rsv;
5378
5379                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5380                 if (ret != -ENOSPC && ret != -EAGAIN)
5381                         break;
5382
5383                 trans->block_rsv = &fs_info->trans_block_rsv;
5384                 btrfs_end_transaction(trans);
5385                 trans = NULL;
5386                 btrfs_btree_balance_dirty(fs_info);
5387         }
5388
5389         btrfs_free_block_rsv(fs_info, rsv);
5390
5391         /*
5392          * Errors here aren't a big deal, it just means we leave orphan items
5393          * in the tree.  They will be cleaned up on the next mount.
5394          */
5395         if (ret == 0) {
5396                 trans->block_rsv = root->orphan_block_rsv;
5397                 btrfs_orphan_del(trans, inode);
5398         } else {
5399                 btrfs_orphan_del(NULL, inode);
5400         }
5401
5402         trans->block_rsv = &fs_info->trans_block_rsv;
5403         if (!(root == fs_info->tree_root ||
5404               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5405                 btrfs_return_ino(root, btrfs_ino(inode));
5406
5407         btrfs_end_transaction(trans);
5408         btrfs_btree_balance_dirty(fs_info);
5409 no_delete:
5410         btrfs_remove_delayed_node(inode);
5411         clear_inode(inode);
5412 }
5413
5414 /*
5415  * this returns the key found in the dir entry in the location pointer.
5416  * If no dir entries were found, location->objectid is 0.
5417  */
5418 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5419                                struct btrfs_key *location)
5420 {
5421         const char *name = dentry->d_name.name;
5422         int namelen = dentry->d_name.len;
5423         struct btrfs_dir_item *di;
5424         struct btrfs_path *path;
5425         struct btrfs_root *root = BTRFS_I(dir)->root;
5426         int ret = 0;
5427
5428         path = btrfs_alloc_path();
5429         if (!path)
5430                 return -ENOMEM;
5431
5432         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5433                                     namelen, 0);
5434         if (IS_ERR(di))
5435                 ret = PTR_ERR(di);
5436
5437         if (IS_ERR_OR_NULL(di))
5438                 goto out_err;
5439
5440         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5441 out:
5442         btrfs_free_path(path);
5443         return ret;
5444 out_err:
5445         location->objectid = 0;
5446         goto out;
5447 }
5448
5449 /*
5450  * when we hit a tree root in a directory, the btrfs part of the inode
5451  * needs to be changed to reflect the root directory of the tree root.  This
5452  * is kind of like crossing a mount point.
5453  */
5454 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5455                                     struct inode *dir,
5456                                     struct dentry *dentry,
5457                                     struct btrfs_key *location,
5458                                     struct btrfs_root **sub_root)
5459 {
5460         struct btrfs_path *path;
5461         struct btrfs_root *new_root;
5462         struct btrfs_root_ref *ref;
5463         struct extent_buffer *leaf;
5464         struct btrfs_key key;
5465         int ret;
5466         int err = 0;
5467
5468         path = btrfs_alloc_path();
5469         if (!path) {
5470                 err = -ENOMEM;
5471                 goto out;
5472         }
5473
5474         err = -ENOENT;
5475         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5476         key.type = BTRFS_ROOT_REF_KEY;
5477         key.offset = location->objectid;
5478
5479         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5480         if (ret) {
5481                 if (ret < 0)
5482                         err = ret;
5483                 goto out;
5484         }
5485
5486         leaf = path->nodes[0];
5487         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5488         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5489             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5490                 goto out;
5491
5492         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5493                                    (unsigned long)(ref + 1),
5494                                    dentry->d_name.len);
5495         if (ret)
5496                 goto out;
5497
5498         btrfs_release_path(path);
5499
5500         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5501         if (IS_ERR(new_root)) {
5502                 err = PTR_ERR(new_root);
5503                 goto out;
5504         }
5505
5506         *sub_root = new_root;
5507         location->objectid = btrfs_root_dirid(&new_root->root_item);
5508         location->type = BTRFS_INODE_ITEM_KEY;
5509         location->offset = 0;
5510         err = 0;
5511 out:
5512         btrfs_free_path(path);
5513         return err;
5514 }
5515
5516 static void inode_tree_add(struct inode *inode)
5517 {
5518         struct btrfs_root *root = BTRFS_I(inode)->root;
5519         struct btrfs_inode *entry;
5520         struct rb_node **p;
5521         struct rb_node *parent;
5522         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5523         u64 ino = btrfs_ino(inode);
5524
5525         if (inode_unhashed(inode))
5526                 return;
5527         parent = NULL;
5528         spin_lock(&root->inode_lock);
5529         p = &root->inode_tree.rb_node;
5530         while (*p) {
5531                 parent = *p;
5532                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5533
5534                 if (ino < btrfs_ino(&entry->vfs_inode))
5535                         p = &parent->rb_left;
5536                 else if (ino > btrfs_ino(&entry->vfs_inode))
5537                         p = &parent->rb_right;
5538                 else {
5539                         WARN_ON(!(entry->vfs_inode.i_state &
5540                                   (I_WILL_FREE | I_FREEING)));
5541                         rb_replace_node(parent, new, &root->inode_tree);
5542                         RB_CLEAR_NODE(parent);
5543                         spin_unlock(&root->inode_lock);
5544                         return;
5545                 }
5546         }
5547         rb_link_node(new, parent, p);
5548         rb_insert_color(new, &root->inode_tree);
5549         spin_unlock(&root->inode_lock);
5550 }
5551
5552 static void inode_tree_del(struct inode *inode)
5553 {
5554         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5555         struct btrfs_root *root = BTRFS_I(inode)->root;
5556         int empty = 0;
5557
5558         spin_lock(&root->inode_lock);
5559         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5560                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5561                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5562                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5563         }
5564         spin_unlock(&root->inode_lock);
5565
5566         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5567                 synchronize_srcu(&fs_info->subvol_srcu);
5568                 spin_lock(&root->inode_lock);
5569                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5570                 spin_unlock(&root->inode_lock);
5571                 if (empty)
5572                         btrfs_add_dead_root(root);
5573         }
5574 }
5575
5576 void btrfs_invalidate_inodes(struct btrfs_root *root)
5577 {
5578         struct btrfs_fs_info *fs_info = root->fs_info;
5579         struct rb_node *node;
5580         struct rb_node *prev;
5581         struct btrfs_inode *entry;
5582         struct inode *inode;
5583         u64 objectid = 0;
5584
5585         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5586                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5587
5588         spin_lock(&root->inode_lock);
5589 again:
5590         node = root->inode_tree.rb_node;
5591         prev = NULL;
5592         while (node) {
5593                 prev = node;
5594                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5595
5596                 if (objectid < btrfs_ino(&entry->vfs_inode))
5597                         node = node->rb_left;
5598                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5599                         node = node->rb_right;
5600                 else
5601                         break;
5602         }
5603         if (!node) {
5604                 while (prev) {
5605                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5606                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5607                                 node = prev;
5608                                 break;
5609                         }
5610                         prev = rb_next(prev);
5611                 }
5612         }
5613         while (node) {
5614                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5615                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5616                 inode = igrab(&entry->vfs_inode);
5617                 if (inode) {
5618                         spin_unlock(&root->inode_lock);
5619                         if (atomic_read(&inode->i_count) > 1)
5620                                 d_prune_aliases(inode);
5621                         /*
5622                          * btrfs_drop_inode will have it removed from
5623                          * the inode cache when its usage count
5624                          * hits zero.
5625                          */
5626                         iput(inode);
5627                         cond_resched();
5628                         spin_lock(&root->inode_lock);
5629                         goto again;
5630                 }
5631
5632                 if (cond_resched_lock(&root->inode_lock))
5633                         goto again;
5634
5635                 node = rb_next(node);
5636         }
5637         spin_unlock(&root->inode_lock);
5638 }
5639
5640 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5641 {
5642         struct btrfs_iget_args *args = p;
5643         inode->i_ino = args->location->objectid;
5644         memcpy(&BTRFS_I(inode)->location, args->location,
5645                sizeof(*args->location));
5646         BTRFS_I(inode)->root = args->root;
5647         return 0;
5648 }
5649
5650 static int btrfs_find_actor(struct inode *inode, void *opaque)
5651 {
5652         struct btrfs_iget_args *args = opaque;
5653         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5654                 args->root == BTRFS_I(inode)->root;
5655 }
5656
5657 static struct inode *btrfs_iget_locked(struct super_block *s,
5658                                        struct btrfs_key *location,
5659                                        struct btrfs_root *root)
5660 {
5661         struct inode *inode;
5662         struct btrfs_iget_args args;
5663         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5664
5665         args.location = location;
5666         args.root = root;
5667
5668         inode = iget5_locked(s, hashval, btrfs_find_actor,
5669                              btrfs_init_locked_inode,
5670                              (void *)&args);
5671         return inode;
5672 }
5673
5674 /* Get an inode object given its location and corresponding root.
5675  * Returns in *is_new if the inode was read from disk
5676  */
5677 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5678                          struct btrfs_root *root, int *new)
5679 {
5680         struct inode *inode;
5681
5682         inode = btrfs_iget_locked(s, location, root);
5683         if (!inode)
5684                 return ERR_PTR(-ENOMEM);
5685
5686         if (inode->i_state & I_NEW) {
5687                 int ret;
5688
5689                 ret = btrfs_read_locked_inode(inode);
5690                 if (!is_bad_inode(inode)) {
5691                         inode_tree_add(inode);
5692                         unlock_new_inode(inode);
5693                         if (new)
5694                                 *new = 1;
5695                 } else {
5696                         unlock_new_inode(inode);
5697                         iput(inode);
5698                         ASSERT(ret < 0);
5699                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5700                 }
5701         }
5702
5703         return inode;
5704 }
5705
5706 static struct inode *new_simple_dir(struct super_block *s,
5707                                     struct btrfs_key *key,
5708                                     struct btrfs_root *root)
5709 {
5710         struct inode *inode = new_inode(s);
5711
5712         if (!inode)
5713                 return ERR_PTR(-ENOMEM);
5714
5715         BTRFS_I(inode)->root = root;
5716         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5717         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5718
5719         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5720         inode->i_op = &btrfs_dir_ro_inode_operations;
5721         inode->i_opflags &= ~IOP_XATTR;
5722         inode->i_fop = &simple_dir_operations;
5723         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5724         inode->i_mtime = current_time(inode);
5725         inode->i_atime = inode->i_mtime;
5726         inode->i_ctime = inode->i_mtime;
5727         BTRFS_I(inode)->i_otime = inode->i_mtime;
5728
5729         return inode;
5730 }
5731
5732 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5733 {
5734         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5735         struct inode *inode;
5736         struct btrfs_root *root = BTRFS_I(dir)->root;
5737         struct btrfs_root *sub_root = root;
5738         struct btrfs_key location;
5739         int index;
5740         int ret = 0;
5741
5742         if (dentry->d_name.len > BTRFS_NAME_LEN)
5743                 return ERR_PTR(-ENAMETOOLONG);
5744
5745         ret = btrfs_inode_by_name(dir, dentry, &location);
5746         if (ret < 0)
5747                 return ERR_PTR(ret);
5748
5749         if (location.objectid == 0)
5750                 return ERR_PTR(-ENOENT);
5751
5752         if (location.type == BTRFS_INODE_ITEM_KEY) {
5753                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5754                 return inode;
5755         }
5756
5757         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5758
5759         index = srcu_read_lock(&fs_info->subvol_srcu);
5760         ret = fixup_tree_root_location(fs_info, dir, dentry,
5761                                        &location, &sub_root);
5762         if (ret < 0) {
5763                 if (ret != -ENOENT)
5764                         inode = ERR_PTR(ret);
5765                 else
5766                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5767         } else {
5768                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5769         }
5770         srcu_read_unlock(&fs_info->subvol_srcu, index);
5771
5772         if (!IS_ERR(inode) && root != sub_root) {
5773                 down_read(&fs_info->cleanup_work_sem);
5774                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5775                         ret = btrfs_orphan_cleanup(sub_root);
5776                 up_read(&fs_info->cleanup_work_sem);
5777                 if (ret) {
5778                         iput(inode);
5779                         inode = ERR_PTR(ret);
5780                 }
5781         }
5782
5783         return inode;
5784 }
5785
5786 static int btrfs_dentry_delete(const struct dentry *dentry)
5787 {
5788         struct btrfs_root *root;
5789         struct inode *inode = d_inode(dentry);
5790
5791         if (!inode && !IS_ROOT(dentry))
5792                 inode = d_inode(dentry->d_parent);
5793
5794         if (inode) {
5795                 root = BTRFS_I(inode)->root;
5796                 if (btrfs_root_refs(&root->root_item) == 0)
5797                         return 1;
5798
5799                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5800                         return 1;
5801         }
5802         return 0;
5803 }
5804
5805 static void btrfs_dentry_release(struct dentry *dentry)
5806 {
5807         kfree(dentry->d_fsdata);
5808 }
5809
5810 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5811                                    unsigned int flags)
5812 {
5813         struct inode *inode;
5814
5815         inode = btrfs_lookup_dentry(dir, dentry);
5816         if (IS_ERR(inode)) {
5817                 if (PTR_ERR(inode) == -ENOENT)
5818                         inode = NULL;
5819                 else
5820                         return ERR_CAST(inode);
5821         }
5822
5823         return d_splice_alias(inode, dentry);
5824 }
5825
5826 unsigned char btrfs_filetype_table[] = {
5827         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5828 };
5829
5830 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5831 {
5832         struct inode *inode = file_inode(file);
5833         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5834         struct btrfs_root *root = BTRFS_I(inode)->root;
5835         struct btrfs_item *item;
5836         struct btrfs_dir_item *di;
5837         struct btrfs_key key;
5838         struct btrfs_key found_key;
5839         struct btrfs_path *path;
5840         struct list_head ins_list;
5841         struct list_head del_list;
5842         int ret;
5843         struct extent_buffer *leaf;
5844         int slot;
5845         unsigned char d_type;
5846         int over = 0;
5847         char tmp_name[32];
5848         char *name_ptr;
5849         int name_len;
5850         bool put = false;
5851         struct btrfs_key location;
5852
5853         if (!dir_emit_dots(file, ctx))
5854                 return 0;
5855
5856         path = btrfs_alloc_path();
5857         if (!path)
5858                 return -ENOMEM;
5859
5860         path->reada = READA_FORWARD;
5861
5862         INIT_LIST_HEAD(&ins_list);
5863         INIT_LIST_HEAD(&del_list);
5864         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5865
5866         key.type = BTRFS_DIR_INDEX_KEY;
5867         key.offset = ctx->pos;
5868         key.objectid = btrfs_ino(inode);
5869
5870         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5871         if (ret < 0)
5872                 goto err;
5873
5874         while (1) {
5875                 leaf = path->nodes[0];
5876                 slot = path->slots[0];
5877                 if (slot >= btrfs_header_nritems(leaf)) {
5878                         ret = btrfs_next_leaf(root, path);
5879                         if (ret < 0)
5880                                 goto err;
5881                         else if (ret > 0)
5882                                 break;
5883                         continue;
5884                 }
5885
5886                 item = btrfs_item_nr(slot);
5887                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5888
5889                 if (found_key.objectid != key.objectid)
5890                         break;
5891                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5892                         break;
5893                 if (found_key.offset < ctx->pos)
5894                         goto next;
5895                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5896                         goto next;
5897
5898                 ctx->pos = found_key.offset;
5899
5900                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5901                 if (verify_dir_item(fs_info, leaf, di))
5902                         goto next;
5903
5904                 name_len = btrfs_dir_name_len(leaf, di);
5905                 if (name_len <= sizeof(tmp_name)) {
5906                         name_ptr = tmp_name;
5907                 } else {
5908                         name_ptr = kmalloc(name_len, GFP_KERNEL);
5909                         if (!name_ptr) {
5910                                 ret = -ENOMEM;
5911                                 goto err;
5912                         }
5913                 }
5914                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5915                                    name_len);
5916
5917                 d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5918                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5919
5920                 over = !dir_emit(ctx, name_ptr, name_len, location.objectid,
5921                                  d_type);
5922
5923                 if (name_ptr != tmp_name)
5924                         kfree(name_ptr);
5925
5926                 if (over)
5927                         goto nopos;
5928                 ctx->pos++;
5929 next:
5930                 path->slots[0]++;
5931         }
5932
5933         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5934         if (ret)
5935                 goto nopos;
5936
5937         /*
5938          * Stop new entries from being returned after we return the last
5939          * entry.
5940          *
5941          * New directory entries are assigned a strictly increasing
5942          * offset.  This means that new entries created during readdir
5943          * are *guaranteed* to be seen in the future by that readdir.
5944          * This has broken buggy programs which operate on names as
5945          * they're returned by readdir.  Until we re-use freed offsets
5946          * we have this hack to stop new entries from being returned
5947          * under the assumption that they'll never reach this huge
5948          * offset.
5949          *
5950          * This is being careful not to overflow 32bit loff_t unless the
5951          * last entry requires it because doing so has broken 32bit apps
5952          * in the past.
5953          */
5954         if (ctx->pos >= INT_MAX)
5955                 ctx->pos = LLONG_MAX;
5956         else
5957                 ctx->pos = INT_MAX;
5958 nopos:
5959         ret = 0;
5960 err:
5961         if (put)
5962                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5963         btrfs_free_path(path);
5964         return ret;
5965 }
5966
5967 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5968 {
5969         struct btrfs_root *root = BTRFS_I(inode)->root;
5970         struct btrfs_trans_handle *trans;
5971         int ret = 0;
5972         bool nolock = false;
5973
5974         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5975                 return 0;
5976
5977         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5978                 nolock = true;
5979
5980         if (wbc->sync_mode == WB_SYNC_ALL) {
5981                 if (nolock)
5982                         trans = btrfs_join_transaction_nolock(root);
5983                 else
5984                         trans = btrfs_join_transaction(root);
5985                 if (IS_ERR(trans))
5986                         return PTR_ERR(trans);
5987                 ret = btrfs_commit_transaction(trans);
5988         }
5989         return ret;
5990 }
5991
5992 /*
5993  * This is somewhat expensive, updating the tree every time the
5994  * inode changes.  But, it is most likely to find the inode in cache.
5995  * FIXME, needs more benchmarking...there are no reasons other than performance
5996  * to keep or drop this code.
5997  */
5998 static int btrfs_dirty_inode(struct inode *inode)
5999 {
6000         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6001         struct btrfs_root *root = BTRFS_I(inode)->root;
6002         struct btrfs_trans_handle *trans;
6003         int ret;
6004
6005         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6006                 return 0;
6007
6008         trans = btrfs_join_transaction(root);
6009         if (IS_ERR(trans))
6010                 return PTR_ERR(trans);
6011
6012         ret = btrfs_update_inode(trans, root, inode);
6013         if (ret && ret == -ENOSPC) {
6014                 /* whoops, lets try again with the full transaction */
6015                 btrfs_end_transaction(trans);
6016                 trans = btrfs_start_transaction(root, 1);
6017                 if (IS_ERR(trans))
6018                         return PTR_ERR(trans);
6019
6020                 ret = btrfs_update_inode(trans, root, inode);
6021         }
6022         btrfs_end_transaction(trans);
6023         if (BTRFS_I(inode)->delayed_node)
6024                 btrfs_balance_delayed_items(fs_info);
6025
6026         return ret;
6027 }
6028
6029 /*
6030  * This is a copy of file_update_time.  We need this so we can return error on
6031  * ENOSPC for updating the inode in the case of file write and mmap writes.
6032  */
6033 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6034                              int flags)
6035 {
6036         struct btrfs_root *root = BTRFS_I(inode)->root;
6037
6038         if (btrfs_root_readonly(root))
6039                 return -EROFS;
6040
6041         if (flags & S_VERSION)
6042                 inode_inc_iversion(inode);
6043         if (flags & S_CTIME)
6044                 inode->i_ctime = *now;
6045         if (flags & S_MTIME)
6046                 inode->i_mtime = *now;
6047         if (flags & S_ATIME)
6048                 inode->i_atime = *now;
6049         return btrfs_dirty_inode(inode);
6050 }
6051
6052 /*
6053  * find the highest existing sequence number in a directory
6054  * and then set the in-memory index_cnt variable to reflect
6055  * free sequence numbers
6056  */
6057 static int btrfs_set_inode_index_count(struct inode *inode)
6058 {
6059         struct btrfs_root *root = BTRFS_I(inode)->root;
6060         struct btrfs_key key, found_key;
6061         struct btrfs_path *path;
6062         struct extent_buffer *leaf;
6063         int ret;
6064
6065         key.objectid = btrfs_ino(inode);
6066         key.type = BTRFS_DIR_INDEX_KEY;
6067         key.offset = (u64)-1;
6068
6069         path = btrfs_alloc_path();
6070         if (!path)
6071                 return -ENOMEM;
6072
6073         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6074         if (ret < 0)
6075                 goto out;
6076         /* FIXME: we should be able to handle this */
6077         if (ret == 0)
6078                 goto out;
6079         ret = 0;
6080
6081         /*
6082          * MAGIC NUMBER EXPLANATION:
6083          * since we search a directory based on f_pos we have to start at 2
6084          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6085          * else has to start at 2
6086          */
6087         if (path->slots[0] == 0) {
6088                 BTRFS_I(inode)->index_cnt = 2;
6089                 goto out;
6090         }
6091
6092         path->slots[0]--;
6093
6094         leaf = path->nodes[0];
6095         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6096
6097         if (found_key.objectid != btrfs_ino(inode) ||
6098             found_key.type != BTRFS_DIR_INDEX_KEY) {
6099                 BTRFS_I(inode)->index_cnt = 2;
6100                 goto out;
6101         }
6102
6103         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6104 out:
6105         btrfs_free_path(path);
6106         return ret;
6107 }
6108
6109 /*
6110  * helper to find a free sequence number in a given directory.  This current
6111  * code is very simple, later versions will do smarter things in the btree
6112  */
6113 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6114 {
6115         int ret = 0;
6116
6117         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6118                 ret = btrfs_inode_delayed_dir_index_count(dir);
6119                 if (ret) {
6120                         ret = btrfs_set_inode_index_count(dir);
6121                         if (ret)
6122                                 return ret;
6123                 }
6124         }
6125
6126         *index = BTRFS_I(dir)->index_cnt;
6127         BTRFS_I(dir)->index_cnt++;
6128
6129         return ret;
6130 }
6131
6132 static int btrfs_insert_inode_locked(struct inode *inode)
6133 {
6134         struct btrfs_iget_args args;
6135         args.location = &BTRFS_I(inode)->location;
6136         args.root = BTRFS_I(inode)->root;
6137
6138         return insert_inode_locked4(inode,
6139                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6140                    btrfs_find_actor, &args);
6141 }
6142
6143 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6144                                      struct btrfs_root *root,
6145                                      struct inode *dir,
6146                                      const char *name, int name_len,
6147                                      u64 ref_objectid, u64 objectid,
6148                                      umode_t mode, u64 *index)
6149 {
6150         struct btrfs_fs_info *fs_info = root->fs_info;
6151         struct inode *inode;
6152         struct btrfs_inode_item *inode_item;
6153         struct btrfs_key *location;
6154         struct btrfs_path *path;
6155         struct btrfs_inode_ref *ref;
6156         struct btrfs_key key[2];
6157         u32 sizes[2];
6158         int nitems = name ? 2 : 1;
6159         unsigned long ptr;
6160         int ret;
6161
6162         path = btrfs_alloc_path();
6163         if (!path)
6164                 return ERR_PTR(-ENOMEM);
6165
6166         inode = new_inode(fs_info->sb);
6167         if (!inode) {
6168                 btrfs_free_path(path);
6169                 return ERR_PTR(-ENOMEM);
6170         }
6171
6172         /*
6173          * O_TMPFILE, set link count to 0, so that after this point,
6174          * we fill in an inode item with the correct link count.
6175          */
6176         if (!name)
6177                 set_nlink(inode, 0);
6178
6179         /*
6180          * we have to initialize this early, so we can reclaim the inode
6181          * number if we fail afterwards in this function.
6182          */
6183         inode->i_ino = objectid;
6184
6185         if (dir && name) {
6186                 trace_btrfs_inode_request(dir);
6187
6188                 ret = btrfs_set_inode_index(dir, index);
6189                 if (ret) {
6190                         btrfs_free_path(path);
6191                         iput(inode);
6192                         return ERR_PTR(ret);
6193                 }
6194         } else if (dir) {
6195                 *index = 0;
6196         }
6197         /*
6198          * index_cnt is ignored for everything but a dir,
6199          * btrfs_get_inode_index_count has an explanation for the magic
6200          * number
6201          */
6202         BTRFS_I(inode)->index_cnt = 2;
6203         BTRFS_I(inode)->dir_index = *index;
6204         BTRFS_I(inode)->root = root;
6205         BTRFS_I(inode)->generation = trans->transid;
6206         inode->i_generation = BTRFS_I(inode)->generation;
6207
6208         /*
6209          * We could have gotten an inode number from somebody who was fsynced
6210          * and then removed in this same transaction, so let's just set full
6211          * sync since it will be a full sync anyway and this will blow away the
6212          * old info in the log.
6213          */
6214         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6215
6216         key[0].objectid = objectid;
6217         key[0].type = BTRFS_INODE_ITEM_KEY;
6218         key[0].offset = 0;
6219
6220         sizes[0] = sizeof(struct btrfs_inode_item);
6221
6222         if (name) {
6223                 /*
6224                  * Start new inodes with an inode_ref. This is slightly more
6225                  * efficient for small numbers of hard links since they will
6226                  * be packed into one item. Extended refs will kick in if we
6227                  * add more hard links than can fit in the ref item.
6228                  */
6229                 key[1].objectid = objectid;
6230                 key[1].type = BTRFS_INODE_REF_KEY;
6231                 key[1].offset = ref_objectid;
6232
6233                 sizes[1] = name_len + sizeof(*ref);
6234         }
6235
6236         location = &BTRFS_I(inode)->location;
6237         location->objectid = objectid;
6238         location->offset = 0;
6239         location->type = BTRFS_INODE_ITEM_KEY;
6240
6241         ret = btrfs_insert_inode_locked(inode);
6242         if (ret < 0)
6243                 goto fail;
6244
6245         path->leave_spinning = 1;
6246         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6247         if (ret != 0)
6248                 goto fail_unlock;
6249
6250         inode_init_owner(inode, dir, mode);
6251         inode_set_bytes(inode, 0);
6252
6253         inode->i_mtime = current_time(inode);
6254         inode->i_atime = inode->i_mtime;
6255         inode->i_ctime = inode->i_mtime;
6256         BTRFS_I(inode)->i_otime = inode->i_mtime;
6257
6258         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6259                                   struct btrfs_inode_item);
6260         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6261                              sizeof(*inode_item));
6262         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6263
6264         if (name) {
6265                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6266                                      struct btrfs_inode_ref);
6267                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6268                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6269                 ptr = (unsigned long)(ref + 1);
6270                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6271         }
6272
6273         btrfs_mark_buffer_dirty(path->nodes[0]);
6274         btrfs_free_path(path);
6275
6276         btrfs_inherit_iflags(inode, dir);
6277
6278         if (S_ISREG(mode)) {
6279                 if (btrfs_test_opt(fs_info, NODATASUM))
6280                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6281                 if (btrfs_test_opt(fs_info, NODATACOW))
6282                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6283                                 BTRFS_INODE_NODATASUM;
6284         }
6285
6286         inode_tree_add(inode);
6287
6288         trace_btrfs_inode_new(inode);
6289         btrfs_set_inode_last_trans(trans, inode);
6290
6291         btrfs_update_root_times(trans, root);
6292
6293         ret = btrfs_inode_inherit_props(trans, inode, dir);
6294         if (ret)
6295                 btrfs_err(fs_info,
6296                           "error inheriting props for ino %llu (root %llu): %d",
6297                           btrfs_ino(inode), root->root_key.objectid, ret);
6298
6299         return inode;
6300
6301 fail_unlock:
6302         unlock_new_inode(inode);
6303 fail:
6304         if (dir && name)
6305                 BTRFS_I(dir)->index_cnt--;
6306         btrfs_free_path(path);
6307         iput(inode);
6308         return ERR_PTR(ret);
6309 }
6310
6311 static inline u8 btrfs_inode_type(struct inode *inode)
6312 {
6313         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6314 }
6315
6316 /*
6317  * utility function to add 'inode' into 'parent_inode' with
6318  * a give name and a given sequence number.
6319  * if 'add_backref' is true, also insert a backref from the
6320  * inode to the parent directory.
6321  */
6322 int btrfs_add_link(struct btrfs_trans_handle *trans,
6323                    struct inode *parent_inode, struct inode *inode,
6324                    const char *name, int name_len, int add_backref, u64 index)
6325 {
6326         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6327         int ret = 0;
6328         struct btrfs_key key;
6329         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6330         u64 ino = btrfs_ino(inode);
6331         u64 parent_ino = btrfs_ino(parent_inode);
6332
6333         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6334                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6335         } else {
6336                 key.objectid = ino;
6337                 key.type = BTRFS_INODE_ITEM_KEY;
6338                 key.offset = 0;
6339         }
6340
6341         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6342                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6343                                          root->root_key.objectid, parent_ino,
6344                                          index, name, name_len);
6345         } else if (add_backref) {
6346                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6347                                              parent_ino, index);
6348         }
6349
6350         /* Nothing to clean up yet */
6351         if (ret)
6352                 return ret;
6353
6354         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6355                                     parent_inode, &key,
6356                                     btrfs_inode_type(inode), index);
6357         if (ret == -EEXIST || ret == -EOVERFLOW)
6358                 goto fail_dir_item;
6359         else if (ret) {
6360                 btrfs_abort_transaction(trans, ret);
6361                 return ret;
6362         }
6363
6364         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6365                            name_len * 2);
6366         inode_inc_iversion(parent_inode);
6367         parent_inode->i_mtime = parent_inode->i_ctime =
6368                 current_time(parent_inode);
6369         ret = btrfs_update_inode(trans, root, parent_inode);
6370         if (ret)
6371                 btrfs_abort_transaction(trans, ret);
6372         return ret;
6373
6374 fail_dir_item:
6375         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6376                 u64 local_index;
6377                 int err;
6378                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6379                                          root->root_key.objectid, parent_ino,
6380                                          &local_index, name, name_len);
6381
6382         } else if (add_backref) {
6383                 u64 local_index;
6384                 int err;
6385
6386                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6387                                           ino, parent_ino, &local_index);
6388         }
6389         return ret;
6390 }
6391
6392 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6393                             struct inode *dir, struct dentry *dentry,
6394                             struct inode *inode, int backref, u64 index)
6395 {
6396         int err = btrfs_add_link(trans, dir, inode,
6397                                  dentry->d_name.name, dentry->d_name.len,
6398                                  backref, index);
6399         if (err > 0)
6400                 err = -EEXIST;
6401         return err;
6402 }
6403
6404 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6405                         umode_t mode, dev_t rdev)
6406 {
6407         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6408         struct btrfs_trans_handle *trans;
6409         struct btrfs_root *root = BTRFS_I(dir)->root;
6410         struct inode *inode = NULL;
6411         int err;
6412         int drop_inode = 0;
6413         u64 objectid;
6414         u64 index = 0;
6415
6416         /*
6417          * 2 for inode item and ref
6418          * 2 for dir items
6419          * 1 for xattr if selinux is on
6420          */
6421         trans = btrfs_start_transaction(root, 5);
6422         if (IS_ERR(trans))
6423                 return PTR_ERR(trans);
6424
6425         err = btrfs_find_free_ino(root, &objectid);
6426         if (err)
6427                 goto out_unlock;
6428
6429         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6430                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6431                                 mode, &index);
6432         if (IS_ERR(inode)) {
6433                 err = PTR_ERR(inode);
6434                 goto out_unlock;
6435         }
6436
6437         /*
6438         * If the active LSM wants to access the inode during
6439         * d_instantiate it needs these. Smack checks to see
6440         * if the filesystem supports xattrs by looking at the
6441         * ops vector.
6442         */
6443         inode->i_op = &btrfs_special_inode_operations;
6444         init_special_inode(inode, inode->i_mode, rdev);
6445
6446         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6447         if (err)
6448                 goto out_unlock_inode;
6449
6450         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6451         if (err) {
6452                 goto out_unlock_inode;
6453         } else {
6454                 btrfs_update_inode(trans, root, inode);
6455                 unlock_new_inode(inode);
6456                 d_instantiate(dentry, inode);
6457         }
6458
6459 out_unlock:
6460         btrfs_end_transaction(trans);
6461         btrfs_balance_delayed_items(fs_info);
6462         btrfs_btree_balance_dirty(fs_info);
6463         if (drop_inode) {
6464                 inode_dec_link_count(inode);
6465                 iput(inode);
6466         }
6467         return err;
6468
6469 out_unlock_inode:
6470         drop_inode = 1;
6471         unlock_new_inode(inode);
6472         goto out_unlock;
6473
6474 }
6475
6476 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6477                         umode_t mode, bool excl)
6478 {
6479         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6480         struct btrfs_trans_handle *trans;
6481         struct btrfs_root *root = BTRFS_I(dir)->root;
6482         struct inode *inode = NULL;
6483         int drop_inode_on_err = 0;
6484         int err;
6485         u64 objectid;
6486         u64 index = 0;
6487
6488         /*
6489          * 2 for inode item and ref
6490          * 2 for dir items
6491          * 1 for xattr if selinux is on
6492          */
6493         trans = btrfs_start_transaction(root, 5);
6494         if (IS_ERR(trans))
6495                 return PTR_ERR(trans);
6496
6497         err = btrfs_find_free_ino(root, &objectid);
6498         if (err)
6499                 goto out_unlock;
6500
6501         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6502                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6503                                 mode, &index);
6504         if (IS_ERR(inode)) {
6505                 err = PTR_ERR(inode);
6506                 goto out_unlock;
6507         }
6508         drop_inode_on_err = 1;
6509         /*
6510         * If the active LSM wants to access the inode during
6511         * d_instantiate it needs these. Smack checks to see
6512         * if the filesystem supports xattrs by looking at the
6513         * ops vector.
6514         */
6515         inode->i_fop = &btrfs_file_operations;
6516         inode->i_op = &btrfs_file_inode_operations;
6517         inode->i_mapping->a_ops = &btrfs_aops;
6518
6519         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6520         if (err)
6521                 goto out_unlock_inode;
6522
6523         err = btrfs_update_inode(trans, root, inode);
6524         if (err)
6525                 goto out_unlock_inode;
6526
6527         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6528         if (err)
6529                 goto out_unlock_inode;
6530
6531         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6532         unlock_new_inode(inode);
6533         d_instantiate(dentry, inode);
6534
6535 out_unlock:
6536         btrfs_end_transaction(trans);
6537         if (err && drop_inode_on_err) {
6538                 inode_dec_link_count(inode);
6539                 iput(inode);
6540         }
6541         btrfs_balance_delayed_items(fs_info);
6542         btrfs_btree_balance_dirty(fs_info);
6543         return err;
6544
6545 out_unlock_inode:
6546         unlock_new_inode(inode);
6547         goto out_unlock;
6548
6549 }
6550
6551 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6552                       struct dentry *dentry)
6553 {
6554         struct btrfs_trans_handle *trans = NULL;
6555         struct btrfs_root *root = BTRFS_I(dir)->root;
6556         struct inode *inode = d_inode(old_dentry);
6557         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6558         u64 index;
6559         int err;
6560         int drop_inode = 0;
6561
6562         /* do not allow sys_link's with other subvols of the same device */
6563         if (root->objectid != BTRFS_I(inode)->root->objectid)
6564                 return -EXDEV;
6565
6566         if (inode->i_nlink >= BTRFS_LINK_MAX)
6567                 return -EMLINK;
6568
6569         err = btrfs_set_inode_index(dir, &index);
6570         if (err)
6571                 goto fail;
6572
6573         /*
6574          * 2 items for inode and inode ref
6575          * 2 items for dir items
6576          * 1 item for parent inode
6577          */
6578         trans = btrfs_start_transaction(root, 5);
6579         if (IS_ERR(trans)) {
6580                 err = PTR_ERR(trans);
6581                 trans = NULL;
6582                 goto fail;
6583         }
6584
6585         /* There are several dir indexes for this inode, clear the cache. */
6586         BTRFS_I(inode)->dir_index = 0ULL;
6587         inc_nlink(inode);
6588         inode_inc_iversion(inode);
6589         inode->i_ctime = current_time(inode);
6590         ihold(inode);
6591         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6592
6593         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6594
6595         if (err) {
6596                 drop_inode = 1;
6597         } else {
6598                 struct dentry *parent = dentry->d_parent;
6599                 err = btrfs_update_inode(trans, root, inode);
6600                 if (err)
6601                         goto fail;
6602                 if (inode->i_nlink == 1) {
6603                         /*
6604                          * If new hard link count is 1, it's a file created
6605                          * with open(2) O_TMPFILE flag.
6606                          */
6607                         err = btrfs_orphan_del(trans, inode);
6608                         if (err)
6609                                 goto fail;
6610                 }
6611                 d_instantiate(dentry, inode);
6612                 btrfs_log_new_name(trans, inode, NULL, parent);
6613         }
6614
6615         btrfs_balance_delayed_items(fs_info);
6616 fail:
6617         if (trans)
6618                 btrfs_end_transaction(trans);
6619         if (drop_inode) {
6620                 inode_dec_link_count(inode);
6621                 iput(inode);
6622         }
6623         btrfs_btree_balance_dirty(fs_info);
6624         return err;
6625 }
6626
6627 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6628 {
6629         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6630         struct inode *inode = NULL;
6631         struct btrfs_trans_handle *trans;
6632         struct btrfs_root *root = BTRFS_I(dir)->root;
6633         int err = 0;
6634         int drop_on_err = 0;
6635         u64 objectid = 0;
6636         u64 index = 0;
6637
6638         /*
6639          * 2 items for inode and ref
6640          * 2 items for dir items
6641          * 1 for xattr if selinux is on
6642          */
6643         trans = btrfs_start_transaction(root, 5);
6644         if (IS_ERR(trans))
6645                 return PTR_ERR(trans);
6646
6647         err = btrfs_find_free_ino(root, &objectid);
6648         if (err)
6649                 goto out_fail;
6650
6651         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6652                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6653                                 S_IFDIR | mode, &index);
6654         if (IS_ERR(inode)) {
6655                 err = PTR_ERR(inode);
6656                 goto out_fail;
6657         }
6658
6659         drop_on_err = 1;
6660         /* these must be set before we unlock the inode */
6661         inode->i_op = &btrfs_dir_inode_operations;
6662         inode->i_fop = &btrfs_dir_file_operations;
6663
6664         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6665         if (err)
6666                 goto out_fail_inode;
6667
6668         btrfs_i_size_write(inode, 0);
6669         err = btrfs_update_inode(trans, root, inode);
6670         if (err)
6671                 goto out_fail_inode;
6672
6673         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6674                              dentry->d_name.len, 0, index);
6675         if (err)
6676                 goto out_fail_inode;
6677
6678         d_instantiate(dentry, inode);
6679         /*
6680          * mkdir is special.  We're unlocking after we call d_instantiate
6681          * to avoid a race with nfsd calling d_instantiate.
6682          */
6683         unlock_new_inode(inode);
6684         drop_on_err = 0;
6685
6686 out_fail:
6687         btrfs_end_transaction(trans);
6688         if (drop_on_err) {
6689                 inode_dec_link_count(inode);
6690                 iput(inode);
6691         }
6692         btrfs_balance_delayed_items(fs_info);
6693         btrfs_btree_balance_dirty(fs_info);
6694         return err;
6695
6696 out_fail_inode:
6697         unlock_new_inode(inode);
6698         goto out_fail;
6699 }
6700
6701 /* Find next extent map of a given extent map, caller needs to ensure locks */
6702 static struct extent_map *next_extent_map(struct extent_map *em)
6703 {
6704         struct rb_node *next;
6705
6706         next = rb_next(&em->rb_node);
6707         if (!next)
6708                 return NULL;
6709         return container_of(next, struct extent_map, rb_node);
6710 }
6711
6712 static struct extent_map *prev_extent_map(struct extent_map *em)
6713 {
6714         struct rb_node *prev;
6715
6716         prev = rb_prev(&em->rb_node);
6717         if (!prev)
6718                 return NULL;
6719         return container_of(prev, struct extent_map, rb_node);
6720 }
6721
6722 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6723  * the existing extent is the nearest extent to map_start,
6724  * and an extent that you want to insert, deal with overlap and insert
6725  * the best fitted new extent into the tree.
6726  */
6727 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6728                                 struct extent_map *existing,
6729                                 struct extent_map *em,
6730                                 u64 map_start)
6731 {
6732         struct extent_map *prev;
6733         struct extent_map *next;
6734         u64 start;
6735         u64 end;
6736         u64 start_diff;
6737
6738         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6739
6740         if (existing->start > map_start) {
6741                 next = existing;
6742                 prev = prev_extent_map(next);
6743         } else {
6744                 prev = existing;
6745                 next = next_extent_map(prev);
6746         }
6747
6748         start = prev ? extent_map_end(prev) : em->start;
6749         start = max_t(u64, start, em->start);
6750         end = next ? next->start : extent_map_end(em);
6751         end = min_t(u64, end, extent_map_end(em));
6752         start_diff = start - em->start;
6753         em->start = start;
6754         em->len = end - start;
6755         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6756             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6757                 em->block_start += start_diff;
6758                 em->block_len -= start_diff;
6759         }
6760         return add_extent_mapping(em_tree, em, 0);
6761 }
6762
6763 static noinline int uncompress_inline(struct btrfs_path *path,
6764                                       struct page *page,
6765                                       size_t pg_offset, u64 extent_offset,
6766                                       struct btrfs_file_extent_item *item)
6767 {
6768         int ret;
6769         struct extent_buffer *leaf = path->nodes[0];
6770         char *tmp;
6771         size_t max_size;
6772         unsigned long inline_size;
6773         unsigned long ptr;
6774         int compress_type;
6775
6776         WARN_ON(pg_offset != 0);
6777         compress_type = btrfs_file_extent_compression(leaf, item);
6778         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6779         inline_size = btrfs_file_extent_inline_item_len(leaf,
6780                                         btrfs_item_nr(path->slots[0]));
6781         tmp = kmalloc(inline_size, GFP_NOFS);
6782         if (!tmp)
6783                 return -ENOMEM;
6784         ptr = btrfs_file_extent_inline_start(item);
6785
6786         read_extent_buffer(leaf, tmp, ptr, inline_size);
6787
6788         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6789         ret = btrfs_decompress(compress_type, tmp, page,
6790                                extent_offset, inline_size, max_size);
6791         kfree(tmp);
6792         return ret;
6793 }
6794
6795 /*
6796  * a bit scary, this does extent mapping from logical file offset to the disk.
6797  * the ugly parts come from merging extents from the disk with the in-ram
6798  * representation.  This gets more complex because of the data=ordered code,
6799  * where the in-ram extents might be locked pending data=ordered completion.
6800  *
6801  * This also copies inline extents directly into the page.
6802  */
6803
6804 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6805                                     size_t pg_offset, u64 start, u64 len,
6806                                     int create)
6807 {
6808         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6809         int ret;
6810         int err = 0;
6811         u64 extent_start = 0;
6812         u64 extent_end = 0;
6813         u64 objectid = btrfs_ino(inode);
6814         u32 found_type;
6815         struct btrfs_path *path = NULL;
6816         struct btrfs_root *root = BTRFS_I(inode)->root;
6817         struct btrfs_file_extent_item *item;
6818         struct extent_buffer *leaf;
6819         struct btrfs_key found_key;
6820         struct extent_map *em = NULL;
6821         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6822         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6823         struct btrfs_trans_handle *trans = NULL;
6824         const bool new_inline = !page || create;
6825
6826 again:
6827         read_lock(&em_tree->lock);
6828         em = lookup_extent_mapping(em_tree, start, len);
6829         if (em)
6830                 em->bdev = fs_info->fs_devices->latest_bdev;
6831         read_unlock(&em_tree->lock);
6832
6833         if (em) {
6834                 if (em->start > start || em->start + em->len <= start)
6835                         free_extent_map(em);
6836                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6837                         free_extent_map(em);
6838                 else
6839                         goto out;
6840         }
6841         em = alloc_extent_map();
6842         if (!em) {
6843                 err = -ENOMEM;
6844                 goto out;
6845         }
6846         em->bdev = fs_info->fs_devices->latest_bdev;
6847         em->start = EXTENT_MAP_HOLE;
6848         em->orig_start = EXTENT_MAP_HOLE;
6849         em->len = (u64)-1;
6850         em->block_len = (u64)-1;
6851
6852         if (!path) {
6853                 path = btrfs_alloc_path();
6854                 if (!path) {
6855                         err = -ENOMEM;
6856                         goto out;
6857                 }
6858                 /*
6859                  * Chances are we'll be called again, so go ahead and do
6860                  * readahead
6861                  */
6862                 path->reada = READA_FORWARD;
6863         }
6864
6865         ret = btrfs_lookup_file_extent(trans, root, path,
6866                                        objectid, start, trans != NULL);
6867         if (ret < 0) {
6868                 err = ret;
6869                 goto out;
6870         }
6871
6872         if (ret != 0) {
6873                 if (path->slots[0] == 0)
6874                         goto not_found;
6875                 path->slots[0]--;
6876         }
6877
6878         leaf = path->nodes[0];
6879         item = btrfs_item_ptr(leaf, path->slots[0],
6880                               struct btrfs_file_extent_item);
6881         /* are we inside the extent that was found? */
6882         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6883         found_type = found_key.type;
6884         if (found_key.objectid != objectid ||
6885             found_type != BTRFS_EXTENT_DATA_KEY) {
6886                 /*
6887                  * If we backup past the first extent we want to move forward
6888                  * and see if there is an extent in front of us, otherwise we'll
6889                  * say there is a hole for our whole search range which can
6890                  * cause problems.
6891                  */
6892                 extent_end = start;
6893                 goto next;
6894         }
6895
6896         found_type = btrfs_file_extent_type(leaf, item);
6897         extent_start = found_key.offset;
6898         if (found_type == BTRFS_FILE_EXTENT_REG ||
6899             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6900                 extent_end = extent_start +
6901                        btrfs_file_extent_num_bytes(leaf, item);
6902         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6903                 size_t size;
6904                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6905                 extent_end = ALIGN(extent_start + size,
6906                                    fs_info->sectorsize);
6907         }
6908 next:
6909         if (start >= extent_end) {
6910                 path->slots[0]++;
6911                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6912                         ret = btrfs_next_leaf(root, path);
6913                         if (ret < 0) {
6914                                 err = ret;
6915                                 goto out;
6916                         }
6917                         if (ret > 0)
6918                                 goto not_found;
6919                         leaf = path->nodes[0];
6920                 }
6921                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6922                 if (found_key.objectid != objectid ||
6923                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6924                         goto not_found;
6925                 if (start + len <= found_key.offset)
6926                         goto not_found;
6927                 if (start > found_key.offset)
6928                         goto next;
6929                 em->start = start;
6930                 em->orig_start = start;
6931                 em->len = found_key.offset - start;
6932                 goto not_found_em;
6933         }
6934
6935         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6936
6937         if (found_type == BTRFS_FILE_EXTENT_REG ||
6938             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6939                 goto insert;
6940         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6941                 unsigned long ptr;
6942                 char *map;
6943                 size_t size;
6944                 size_t extent_offset;
6945                 size_t copy_size;
6946
6947                 if (new_inline)
6948                         goto out;
6949
6950                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6951                 extent_offset = page_offset(page) + pg_offset - extent_start;
6952                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6953                                   size - extent_offset);
6954                 em->start = extent_start + extent_offset;
6955                 em->len = ALIGN(copy_size, fs_info->sectorsize);
6956                 em->orig_block_len = em->len;
6957                 em->orig_start = em->start;
6958                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6959                 if (create == 0 && !PageUptodate(page)) {
6960                         if (btrfs_file_extent_compression(leaf, item) !=
6961                             BTRFS_COMPRESS_NONE) {
6962                                 ret = uncompress_inline(path, page, pg_offset,
6963                                                         extent_offset, item);
6964                                 if (ret) {
6965                                         err = ret;
6966                                         goto out;
6967                                 }
6968                         } else {
6969                                 map = kmap(page);
6970                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6971                                                    copy_size);
6972                                 if (pg_offset + copy_size < PAGE_SIZE) {
6973                                         memset(map + pg_offset + copy_size, 0,
6974                                                PAGE_SIZE - pg_offset -
6975                                                copy_size);
6976                                 }
6977                                 kunmap(page);
6978                         }
6979                         flush_dcache_page(page);
6980                 } else if (create && PageUptodate(page)) {
6981                         BUG();
6982                         if (!trans) {
6983                                 kunmap(page);
6984                                 free_extent_map(em);
6985                                 em = NULL;
6986
6987                                 btrfs_release_path(path);
6988                                 trans = btrfs_join_transaction(root);
6989
6990                                 if (IS_ERR(trans))
6991                                         return ERR_CAST(trans);
6992                                 goto again;
6993                         }
6994                         map = kmap(page);
6995                         write_extent_buffer(leaf, map + pg_offset, ptr,
6996                                             copy_size);
6997                         kunmap(page);
6998                         btrfs_mark_buffer_dirty(leaf);
6999                 }
7000                 set_extent_uptodate(io_tree, em->start,
7001                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7002                 goto insert;
7003         }
7004 not_found:
7005         em->start = start;
7006         em->orig_start = start;
7007         em->len = len;
7008 not_found_em:
7009         em->block_start = EXTENT_MAP_HOLE;
7010         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
7011 insert:
7012         btrfs_release_path(path);
7013         if (em->start > start || extent_map_end(em) <= start) {
7014                 btrfs_err(fs_info,
7015                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7016                           em->start, em->len, start, len);
7017                 err = -EIO;
7018                 goto out;
7019         }
7020
7021         err = 0;
7022         write_lock(&em_tree->lock);
7023         ret = add_extent_mapping(em_tree, em, 0);
7024         /* it is possible that someone inserted the extent into the tree
7025          * while we had the lock dropped.  It is also possible that
7026          * an overlapping map exists in the tree
7027          */
7028         if (ret == -EEXIST) {
7029                 struct extent_map *existing;
7030
7031                 ret = 0;
7032
7033                 existing = search_extent_mapping(em_tree, start, len);
7034                 /*
7035                  * existing will always be non-NULL, since there must be
7036                  * extent causing the -EEXIST.
7037                  */
7038                 if (existing->start == em->start &&
7039                     extent_map_end(existing) >= extent_map_end(em) &&
7040                     em->block_start == existing->block_start) {
7041                         /*
7042                          * The existing extent map already encompasses the
7043                          * entire extent map we tried to add.
7044                          */
7045                         free_extent_map(em);
7046                         em = existing;
7047                         err = 0;
7048
7049                 } else if (start >= extent_map_end(existing) ||
7050                     start <= existing->start) {
7051                         /*
7052                          * The existing extent map is the one nearest to
7053                          * the [start, start + len) range which overlaps
7054                          */
7055                         err = merge_extent_mapping(em_tree, existing,
7056                                                    em, start);
7057                         free_extent_map(existing);
7058                         if (err) {
7059                                 free_extent_map(em);
7060                                 em = NULL;
7061                         }
7062                 } else {
7063                         free_extent_map(em);
7064                         em = existing;
7065                         err = 0;
7066                 }
7067         }
7068         write_unlock(&em_tree->lock);
7069 out:
7070
7071         trace_btrfs_get_extent(root, inode, em);
7072
7073         btrfs_free_path(path);
7074         if (trans) {
7075                 ret = btrfs_end_transaction(trans);
7076                 if (!err)
7077                         err = ret;
7078         }
7079         if (err) {
7080                 free_extent_map(em);
7081                 return ERR_PTR(err);
7082         }
7083         BUG_ON(!em); /* Error is always set */
7084         return em;
7085 }
7086
7087 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7088                                            size_t pg_offset, u64 start, u64 len,
7089                                            int create)
7090 {
7091         struct extent_map *em;
7092         struct extent_map *hole_em = NULL;
7093         u64 range_start = start;
7094         u64 end;
7095         u64 found;
7096         u64 found_end;
7097         int err = 0;
7098
7099         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7100         if (IS_ERR(em))
7101                 return em;
7102         if (em) {
7103                 /*
7104                  * if our em maps to
7105                  * -  a hole or
7106                  * -  a pre-alloc extent,
7107                  * there might actually be delalloc bytes behind it.
7108                  */
7109                 if (em->block_start != EXTENT_MAP_HOLE &&
7110                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7111                         return em;
7112                 else
7113                         hole_em = em;
7114         }
7115
7116         /* check to see if we've wrapped (len == -1 or similar) */
7117         end = start + len;
7118         if (end < start)
7119                 end = (u64)-1;
7120         else
7121                 end -= 1;
7122
7123         em = NULL;
7124
7125         /* ok, we didn't find anything, lets look for delalloc */
7126         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7127                                  end, len, EXTENT_DELALLOC, 1);
7128         found_end = range_start + found;
7129         if (found_end < range_start)
7130                 found_end = (u64)-1;
7131
7132         /*
7133          * we didn't find anything useful, return
7134          * the original results from get_extent()
7135          */
7136         if (range_start > end || found_end <= start) {
7137                 em = hole_em;
7138                 hole_em = NULL;
7139                 goto out;
7140         }
7141
7142         /* adjust the range_start to make sure it doesn't
7143          * go backwards from the start they passed in
7144          */
7145         range_start = max(start, range_start);
7146         found = found_end - range_start;
7147
7148         if (found > 0) {
7149                 u64 hole_start = start;
7150                 u64 hole_len = len;
7151
7152                 em = alloc_extent_map();
7153                 if (!em) {
7154                         err = -ENOMEM;
7155                         goto out;
7156                 }
7157                 /*
7158                  * when btrfs_get_extent can't find anything it
7159                  * returns one huge hole
7160                  *
7161                  * make sure what it found really fits our range, and
7162                  * adjust to make sure it is based on the start from
7163                  * the caller
7164                  */
7165                 if (hole_em) {
7166                         u64 calc_end = extent_map_end(hole_em);
7167
7168                         if (calc_end <= start || (hole_em->start > end)) {
7169                                 free_extent_map(hole_em);
7170                                 hole_em = NULL;
7171                         } else {
7172                                 hole_start = max(hole_em->start, start);
7173                                 hole_len = calc_end - hole_start;
7174                         }
7175                 }
7176                 em->bdev = NULL;
7177                 if (hole_em && range_start > hole_start) {
7178                         /* our hole starts before our delalloc, so we
7179                          * have to return just the parts of the hole
7180                          * that go until  the delalloc starts
7181                          */
7182                         em->len = min(hole_len,
7183                                       range_start - hole_start);
7184                         em->start = hole_start;
7185                         em->orig_start = hole_start;
7186                         /*
7187                          * don't adjust block start at all,
7188                          * it is fixed at EXTENT_MAP_HOLE
7189                          */
7190                         em->block_start = hole_em->block_start;
7191                         em->block_len = hole_len;
7192                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7193                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7194                 } else {
7195                         em->start = range_start;
7196                         em->len = found;
7197                         em->orig_start = range_start;
7198                         em->block_start = EXTENT_MAP_DELALLOC;
7199                         em->block_len = found;
7200                 }
7201         } else if (hole_em) {
7202                 return hole_em;
7203         }
7204 out:
7205
7206         free_extent_map(hole_em);
7207         if (err) {
7208                 free_extent_map(em);
7209                 return ERR_PTR(err);
7210         }
7211         return em;
7212 }
7213
7214 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7215                                                   const u64 start,
7216                                                   const u64 len,
7217                                                   const u64 orig_start,
7218                                                   const u64 block_start,
7219                                                   const u64 block_len,
7220                                                   const u64 orig_block_len,
7221                                                   const u64 ram_bytes,
7222                                                   const int type)
7223 {
7224         struct extent_map *em = NULL;
7225         int ret;
7226
7227         if (type != BTRFS_ORDERED_NOCOW) {
7228                 em = create_pinned_em(inode, start, len, orig_start,
7229                                       block_start, block_len, orig_block_len,
7230                                       ram_bytes, type);
7231                 if (IS_ERR(em))
7232                         goto out;
7233         }
7234         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7235                                            len, block_len, type);
7236         if (ret) {
7237                 if (em) {
7238                         free_extent_map(em);
7239                         btrfs_drop_extent_cache(inode, start,
7240                                                 start + len - 1, 0);
7241                 }
7242                 em = ERR_PTR(ret);
7243         }
7244  out:
7245
7246         return em;
7247 }
7248
7249 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7250                                                   u64 start, u64 len)
7251 {
7252         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7253         struct btrfs_root *root = BTRFS_I(inode)->root;
7254         struct extent_map *em;
7255         struct btrfs_key ins;
7256         u64 alloc_hint;
7257         int ret;
7258
7259         alloc_hint = get_extent_allocation_hint(inode, start, len);
7260         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7261                                    0, alloc_hint, &ins, 1, 1);
7262         if (ret)
7263                 return ERR_PTR(ret);
7264
7265         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7266                                      ins.objectid, ins.offset, ins.offset,
7267                                      ins.offset, 0);
7268         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7269         if (IS_ERR(em))
7270                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7271                                            ins.offset, 1);
7272
7273         return em;
7274 }
7275
7276 /*
7277  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7278  * block must be cow'd
7279  */
7280 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7281                               u64 *orig_start, u64 *orig_block_len,
7282                               u64 *ram_bytes)
7283 {
7284         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7285         struct btrfs_trans_handle *trans;
7286         struct btrfs_path *path;
7287         int ret;
7288         struct extent_buffer *leaf;
7289         struct btrfs_root *root = BTRFS_I(inode)->root;
7290         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7291         struct btrfs_file_extent_item *fi;
7292         struct btrfs_key key;
7293         u64 disk_bytenr;
7294         u64 backref_offset;
7295         u64 extent_end;
7296         u64 num_bytes;
7297         int slot;
7298         int found_type;
7299         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7300
7301         path = btrfs_alloc_path();
7302         if (!path)
7303                 return -ENOMEM;
7304
7305         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7306                                        offset, 0);
7307         if (ret < 0)
7308                 goto out;
7309
7310         slot = path->slots[0];
7311         if (ret == 1) {
7312                 if (slot == 0) {
7313                         /* can't find the item, must cow */
7314                         ret = 0;
7315                         goto out;
7316                 }
7317                 slot--;
7318         }
7319         ret = 0;
7320         leaf = path->nodes[0];
7321         btrfs_item_key_to_cpu(leaf, &key, slot);
7322         if (key.objectid != btrfs_ino(inode) ||
7323             key.type != BTRFS_EXTENT_DATA_KEY) {
7324                 /* not our file or wrong item type, must cow */
7325                 goto out;
7326         }
7327
7328         if (key.offset > offset) {
7329                 /* Wrong offset, must cow */
7330                 goto out;
7331         }
7332
7333         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7334         found_type = btrfs_file_extent_type(leaf, fi);
7335         if (found_type != BTRFS_FILE_EXTENT_REG &&
7336             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7337                 /* not a regular extent, must cow */
7338                 goto out;
7339         }
7340
7341         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7342                 goto out;
7343
7344         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7345         if (extent_end <= offset)
7346                 goto out;
7347
7348         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7349         if (disk_bytenr == 0)
7350                 goto out;
7351
7352         if (btrfs_file_extent_compression(leaf, fi) ||
7353             btrfs_file_extent_encryption(leaf, fi) ||
7354             btrfs_file_extent_other_encoding(leaf, fi))
7355                 goto out;
7356
7357         backref_offset = btrfs_file_extent_offset(leaf, fi);
7358
7359         if (orig_start) {
7360                 *orig_start = key.offset - backref_offset;
7361                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7362                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7363         }
7364
7365         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7366                 goto out;
7367
7368         num_bytes = min(offset + *len, extent_end) - offset;
7369         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7370                 u64 range_end;
7371
7372                 range_end = round_up(offset + num_bytes,
7373                                      root->fs_info->sectorsize) - 1;
7374                 ret = test_range_bit(io_tree, offset, range_end,
7375                                      EXTENT_DELALLOC, 0, NULL);
7376                 if (ret) {
7377                         ret = -EAGAIN;
7378                         goto out;
7379                 }
7380         }
7381
7382         btrfs_release_path(path);
7383
7384         /*
7385          * look for other files referencing this extent, if we
7386          * find any we must cow
7387          */
7388         trans = btrfs_join_transaction(root);
7389         if (IS_ERR(trans)) {
7390                 ret = 0;
7391                 goto out;
7392         }
7393
7394         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7395                                     key.offset - backref_offset, disk_bytenr);
7396         btrfs_end_transaction(trans);
7397         if (ret) {
7398                 ret = 0;
7399                 goto out;
7400         }
7401
7402         /*
7403          * adjust disk_bytenr and num_bytes to cover just the bytes
7404          * in this extent we are about to write.  If there
7405          * are any csums in that range we have to cow in order
7406          * to keep the csums correct
7407          */
7408         disk_bytenr += backref_offset;
7409         disk_bytenr += offset - key.offset;
7410         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7411                 goto out;
7412         /*
7413          * all of the above have passed, it is safe to overwrite this extent
7414          * without cow
7415          */
7416         *len = num_bytes;
7417         ret = 1;
7418 out:
7419         btrfs_free_path(path);
7420         return ret;
7421 }
7422
7423 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7424 {
7425         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7426         int found = false;
7427         void **pagep = NULL;
7428         struct page *page = NULL;
7429         int start_idx;
7430         int end_idx;
7431
7432         start_idx = start >> PAGE_SHIFT;
7433
7434         /*
7435          * end is the last byte in the last page.  end == start is legal
7436          */
7437         end_idx = end >> PAGE_SHIFT;
7438
7439         rcu_read_lock();
7440
7441         /* Most of the code in this while loop is lifted from
7442          * find_get_page.  It's been modified to begin searching from a
7443          * page and return just the first page found in that range.  If the
7444          * found idx is less than or equal to the end idx then we know that
7445          * a page exists.  If no pages are found or if those pages are
7446          * outside of the range then we're fine (yay!) */
7447         while (page == NULL &&
7448                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7449                 page = radix_tree_deref_slot(pagep);
7450                 if (unlikely(!page))
7451                         break;
7452
7453                 if (radix_tree_exception(page)) {
7454                         if (radix_tree_deref_retry(page)) {
7455                                 page = NULL;
7456                                 continue;
7457                         }
7458                         /*
7459                          * Otherwise, shmem/tmpfs must be storing a swap entry
7460                          * here as an exceptional entry: so return it without
7461                          * attempting to raise page count.
7462                          */
7463                         page = NULL;
7464                         break; /* TODO: Is this relevant for this use case? */
7465                 }
7466
7467                 if (!page_cache_get_speculative(page)) {
7468                         page = NULL;
7469                         continue;
7470                 }
7471
7472                 /*
7473                  * Has the page moved?
7474                  * This is part of the lockless pagecache protocol. See
7475                  * include/linux/pagemap.h for details.
7476                  */
7477                 if (unlikely(page != *pagep)) {
7478                         put_page(page);
7479                         page = NULL;
7480                 }
7481         }
7482
7483         if (page) {
7484                 if (page->index <= end_idx)
7485                         found = true;
7486                 put_page(page);
7487         }
7488
7489         rcu_read_unlock();
7490         return found;
7491 }
7492
7493 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7494                               struct extent_state **cached_state, int writing)
7495 {
7496         struct btrfs_ordered_extent *ordered;
7497         int ret = 0;
7498
7499         while (1) {
7500                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7501                                  cached_state);
7502                 /*
7503                  * We're concerned with the entire range that we're going to be
7504                  * doing DIO to, so we need to make sure there's no ordered
7505                  * extents in this range.
7506                  */
7507                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7508                                                      lockend - lockstart + 1);
7509
7510                 /*
7511                  * We need to make sure there are no buffered pages in this
7512                  * range either, we could have raced between the invalidate in
7513                  * generic_file_direct_write and locking the extent.  The
7514                  * invalidate needs to happen so that reads after a write do not
7515                  * get stale data.
7516                  */
7517                 if (!ordered &&
7518                     (!writing ||
7519                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7520                         break;
7521
7522                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7523                                      cached_state, GFP_NOFS);
7524
7525                 if (ordered) {
7526                         /*
7527                          * If we are doing a DIO read and the ordered extent we
7528                          * found is for a buffered write, we can not wait for it
7529                          * to complete and retry, because if we do so we can
7530                          * deadlock with concurrent buffered writes on page
7531                          * locks. This happens only if our DIO read covers more
7532                          * than one extent map, if at this point has already
7533                          * created an ordered extent for a previous extent map
7534                          * and locked its range in the inode's io tree, and a
7535                          * concurrent write against that previous extent map's
7536                          * range and this range started (we unlock the ranges
7537                          * in the io tree only when the bios complete and
7538                          * buffered writes always lock pages before attempting
7539                          * to lock range in the io tree).
7540                          */
7541                         if (writing ||
7542                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7543                                 btrfs_start_ordered_extent(inode, ordered, 1);
7544                         else
7545                                 ret = -ENOTBLK;
7546                         btrfs_put_ordered_extent(ordered);
7547                 } else {
7548                         /*
7549                          * We could trigger writeback for this range (and wait
7550                          * for it to complete) and then invalidate the pages for
7551                          * this range (through invalidate_inode_pages2_range()),
7552                          * but that can lead us to a deadlock with a concurrent
7553                          * call to readpages() (a buffered read or a defrag call
7554                          * triggered a readahead) on a page lock due to an
7555                          * ordered dio extent we created before but did not have
7556                          * yet a corresponding bio submitted (whence it can not
7557                          * complete), which makes readpages() wait for that
7558                          * ordered extent to complete while holding a lock on
7559                          * that page.
7560                          */
7561                         ret = -ENOTBLK;
7562                 }
7563
7564                 if (ret)
7565                         break;
7566
7567                 cond_resched();
7568         }
7569
7570         return ret;
7571 }
7572
7573 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7574                                            u64 len, u64 orig_start,
7575                                            u64 block_start, u64 block_len,
7576                                            u64 orig_block_len, u64 ram_bytes,
7577                                            int type)
7578 {
7579         struct extent_map_tree *em_tree;
7580         struct extent_map *em;
7581         struct btrfs_root *root = BTRFS_I(inode)->root;
7582         int ret;
7583
7584         em_tree = &BTRFS_I(inode)->extent_tree;
7585         em = alloc_extent_map();
7586         if (!em)
7587                 return ERR_PTR(-ENOMEM);
7588
7589         em->start = start;
7590         em->orig_start = orig_start;
7591         em->mod_start = start;
7592         em->mod_len = len;
7593         em->len = len;
7594         em->block_len = block_len;
7595         em->block_start = block_start;
7596         em->bdev = root->fs_info->fs_devices->latest_bdev;
7597         em->orig_block_len = orig_block_len;
7598         em->ram_bytes = ram_bytes;
7599         em->generation = -1;
7600         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7601         if (type == BTRFS_ORDERED_PREALLOC)
7602                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7603
7604         do {
7605                 btrfs_drop_extent_cache(inode, em->start,
7606                                 em->start + em->len - 1, 0);
7607                 write_lock(&em_tree->lock);
7608                 ret = add_extent_mapping(em_tree, em, 1);
7609                 write_unlock(&em_tree->lock);
7610         } while (ret == -EEXIST);
7611
7612         if (ret) {
7613                 free_extent_map(em);
7614                 return ERR_PTR(ret);
7615         }
7616
7617         return em;
7618 }
7619
7620 static void adjust_dio_outstanding_extents(struct inode *inode,
7621                                            struct btrfs_dio_data *dio_data,
7622                                            const u64 len)
7623 {
7624         unsigned num_extents;
7625
7626         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7627                                            BTRFS_MAX_EXTENT_SIZE);
7628         /*
7629          * If we have an outstanding_extents count still set then we're
7630          * within our reservation, otherwise we need to adjust our inode
7631          * counter appropriately.
7632          */
7633         if (dio_data->outstanding_extents >= num_extents) {
7634                 dio_data->outstanding_extents -= num_extents;
7635         } else {
7636                 /*
7637                  * If dio write length has been split due to no large enough
7638                  * contiguous space, we need to compensate our inode counter
7639                  * appropriately.
7640                  */
7641                 u64 num_needed = num_extents - dio_data->outstanding_extents;
7642
7643                 spin_lock(&BTRFS_I(inode)->lock);
7644                 BTRFS_I(inode)->outstanding_extents += num_needed;
7645                 spin_unlock(&BTRFS_I(inode)->lock);
7646         }
7647 }
7648
7649 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7650                                    struct buffer_head *bh_result, int create)
7651 {
7652         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7653         struct extent_map *em;
7654         struct extent_state *cached_state = NULL;
7655         struct btrfs_dio_data *dio_data = NULL;
7656         u64 start = iblock << inode->i_blkbits;
7657         u64 lockstart, lockend;
7658         u64 len = bh_result->b_size;
7659         int unlock_bits = EXTENT_LOCKED;
7660         int ret = 0;
7661
7662         if (create)
7663                 unlock_bits |= EXTENT_DIRTY;
7664         else
7665                 len = min_t(u64, len, fs_info->sectorsize);
7666
7667         lockstart = start;
7668         lockend = start + len - 1;
7669
7670         if (current->journal_info) {
7671                 /*
7672                  * Need to pull our outstanding extents and set journal_info to NULL so
7673                  * that anything that needs to check if there's a transaction doesn't get
7674                  * confused.
7675                  */
7676                 dio_data = current->journal_info;
7677                 current->journal_info = NULL;
7678         }
7679
7680         /*
7681          * If this errors out it's because we couldn't invalidate pagecache for
7682          * this range and we need to fallback to buffered.
7683          */
7684         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7685                                create)) {
7686                 ret = -ENOTBLK;
7687                 goto err;
7688         }
7689
7690         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7691         if (IS_ERR(em)) {
7692                 ret = PTR_ERR(em);
7693                 goto unlock_err;
7694         }
7695
7696         /*
7697          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7698          * io.  INLINE is special, and we could probably kludge it in here, but
7699          * it's still buffered so for safety lets just fall back to the generic
7700          * buffered path.
7701          *
7702          * For COMPRESSED we _have_ to read the entire extent in so we can
7703          * decompress it, so there will be buffering required no matter what we
7704          * do, so go ahead and fallback to buffered.
7705          *
7706          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7707          * to buffered IO.  Don't blame me, this is the price we pay for using
7708          * the generic code.
7709          */
7710         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7711             em->block_start == EXTENT_MAP_INLINE) {
7712                 free_extent_map(em);
7713                 ret = -ENOTBLK;
7714                 goto unlock_err;
7715         }
7716
7717         /* Just a good old fashioned hole, return */
7718         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7719                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7720                 free_extent_map(em);
7721                 goto unlock_err;
7722         }
7723
7724         /*
7725          * We don't allocate a new extent in the following cases
7726          *
7727          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7728          * existing extent.
7729          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7730          * just use the extent.
7731          *
7732          */
7733         if (!create) {
7734                 len = min(len, em->len - (start - em->start));
7735                 lockstart = start + len;
7736                 goto unlock;
7737         }
7738
7739         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7740             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7741              em->block_start != EXTENT_MAP_HOLE)) {
7742                 int type;
7743                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7744
7745                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7746                         type = BTRFS_ORDERED_PREALLOC;
7747                 else
7748                         type = BTRFS_ORDERED_NOCOW;
7749                 len = min(len, em->len - (start - em->start));
7750                 block_start = em->block_start + (start - em->start);
7751
7752                 if (can_nocow_extent(inode, start, &len, &orig_start,
7753                                      &orig_block_len, &ram_bytes) == 1 &&
7754                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7755                         struct extent_map *em2;
7756
7757                         em2 = btrfs_create_dio_extent(inode, start, len,
7758                                                       orig_start, block_start,
7759                                                       len, orig_block_len,
7760                                                       ram_bytes, type);
7761                         btrfs_dec_nocow_writers(fs_info, block_start);
7762                         if (type == BTRFS_ORDERED_PREALLOC) {
7763                                 free_extent_map(em);
7764                                 em = em2;
7765                         }
7766                         if (em2 && IS_ERR(em2)) {
7767                                 ret = PTR_ERR(em2);
7768                                 goto unlock_err;
7769                         }
7770                         /*
7771                          * For inode marked NODATACOW or extent marked PREALLOC,
7772                          * use the existing or preallocated extent, so does not
7773                          * need to adjust btrfs_space_info's bytes_may_use.
7774                          */
7775                         btrfs_free_reserved_data_space_noquota(inode,
7776                                         start, len);
7777                         goto unlock;
7778                 }
7779         }
7780
7781         /*
7782          * this will cow the extent, reset the len in case we changed
7783          * it above
7784          */
7785         len = bh_result->b_size;
7786         free_extent_map(em);
7787         em = btrfs_new_extent_direct(inode, start, len);
7788         if (IS_ERR(em)) {
7789                 ret = PTR_ERR(em);
7790                 goto unlock_err;
7791         }
7792         len = min(len, em->len - (start - em->start));
7793 unlock:
7794         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7795                 inode->i_blkbits;
7796         bh_result->b_size = len;
7797         bh_result->b_bdev = em->bdev;
7798         set_buffer_mapped(bh_result);
7799         if (create) {
7800                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7801                         set_buffer_new(bh_result);
7802
7803                 /*
7804                  * Need to update the i_size under the extent lock so buffered
7805                  * readers will get the updated i_size when we unlock.
7806                  */
7807                 if (start + len > i_size_read(inode))
7808                         i_size_write(inode, start + len);
7809
7810                 adjust_dio_outstanding_extents(inode, dio_data, len);
7811                 WARN_ON(dio_data->reserve < len);
7812                 dio_data->reserve -= len;
7813                 dio_data->unsubmitted_oe_range_end = start + len;
7814                 current->journal_info = dio_data;
7815         }
7816
7817         /*
7818          * In the case of write we need to clear and unlock the entire range,
7819          * in the case of read we need to unlock only the end area that we
7820          * aren't using if there is any left over space.
7821          */
7822         if (lockstart < lockend) {
7823                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7824                                  lockend, unlock_bits, 1, 0,
7825                                  &cached_state, GFP_NOFS);
7826         } else {
7827                 free_extent_state(cached_state);
7828         }
7829
7830         free_extent_map(em);
7831
7832         return 0;
7833
7834 unlock_err:
7835         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7836                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7837 err:
7838         if (dio_data)
7839                 current->journal_info = dio_data;
7840         /*
7841          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7842          * write less data then expected, so that we don't underflow our inode's
7843          * outstanding extents counter.
7844          */
7845         if (create && dio_data)
7846                 adjust_dio_outstanding_extents(inode, dio_data, len);
7847
7848         return ret;
7849 }
7850
7851 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7852                                         int mirror_num)
7853 {
7854         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7855         int ret;
7856
7857         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7858
7859         bio_get(bio);
7860
7861         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7862         if (ret)
7863                 goto err;
7864
7865         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7866 err:
7867         bio_put(bio);
7868         return ret;
7869 }
7870
7871 static int btrfs_check_dio_repairable(struct inode *inode,
7872                                       struct bio *failed_bio,
7873                                       struct io_failure_record *failrec,
7874                                       int failed_mirror)
7875 {
7876         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7877         int num_copies;
7878
7879         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7880         if (num_copies == 1) {
7881                 /*
7882                  * we only have a single copy of the data, so don't bother with
7883                  * all the retry and error correction code that follows. no
7884                  * matter what the error is, it is very likely to persist.
7885                  */
7886                 btrfs_debug(fs_info,
7887                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7888                         num_copies, failrec->this_mirror, failed_mirror);
7889                 return 0;
7890         }
7891
7892         failrec->failed_mirror = failed_mirror;
7893         failrec->this_mirror++;
7894         if (failrec->this_mirror == failed_mirror)
7895                 failrec->this_mirror++;
7896
7897         if (failrec->this_mirror > num_copies) {
7898                 btrfs_debug(fs_info,
7899                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7900                         num_copies, failrec->this_mirror, failed_mirror);
7901                 return 0;
7902         }
7903
7904         return 1;
7905 }
7906
7907 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7908                         struct page *page, unsigned int pgoff,
7909                         u64 start, u64 end, int failed_mirror,
7910                         bio_end_io_t *repair_endio, void *repair_arg)
7911 {
7912         struct io_failure_record *failrec;
7913         struct bio *bio;
7914         int isector;
7915         int read_mode = 0;
7916         int ret;
7917
7918         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7919
7920         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7921         if (ret)
7922                 return ret;
7923
7924         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7925                                          failed_mirror);
7926         if (!ret) {
7927                 free_io_failure(inode, failrec);
7928                 return -EIO;
7929         }
7930
7931         if ((failed_bio->bi_vcnt > 1)
7932                 || (failed_bio->bi_io_vec->bv_len
7933                         > btrfs_inode_sectorsize(inode)))
7934                 read_mode |= REQ_FAILFAST_DEV;
7935
7936         isector = start - btrfs_io_bio(failed_bio)->logical;
7937         isector >>= inode->i_sb->s_blocksize_bits;
7938         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7939                                 pgoff, isector, repair_endio, repair_arg);
7940         if (!bio) {
7941                 free_io_failure(inode, failrec);
7942                 return -EIO;
7943         }
7944         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
7945
7946         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7947                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7948                     read_mode, failrec->this_mirror, failrec->in_validation);
7949
7950         ret = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7951         if (ret) {
7952                 free_io_failure(inode, failrec);
7953                 bio_put(bio);
7954         }
7955
7956         return ret;
7957 }
7958
7959 struct btrfs_retry_complete {
7960         struct completion done;
7961         struct inode *inode;
7962         u64 start;
7963         int uptodate;
7964 };
7965
7966 static void btrfs_retry_endio_nocsum(struct bio *bio)
7967 {
7968         struct btrfs_retry_complete *done = bio->bi_private;
7969         struct inode *inode;
7970         struct bio_vec *bvec;
7971         int i;
7972
7973         if (bio->bi_error)
7974                 goto end;
7975
7976         ASSERT(bio->bi_vcnt == 1);
7977         inode = bio->bi_io_vec->bv_page->mapping->host;
7978         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(inode));
7979
7980         done->uptodate = 1;
7981         bio_for_each_segment_all(bvec, bio, i)
7982                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7983 end:
7984         complete(&done->done);
7985         bio_put(bio);
7986 }
7987
7988 static int __btrfs_correct_data_nocsum(struct inode *inode,
7989                                        struct btrfs_io_bio *io_bio)
7990 {
7991         struct btrfs_fs_info *fs_info;
7992         struct bio_vec *bvec;
7993         struct btrfs_retry_complete done;
7994         u64 start;
7995         unsigned int pgoff;
7996         u32 sectorsize;
7997         int nr_sectors;
7998         int i;
7999         int ret;
8000
8001         fs_info = BTRFS_I(inode)->root->fs_info;
8002         sectorsize = fs_info->sectorsize;
8003
8004         start = io_bio->logical;
8005         done.inode = inode;
8006
8007         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8008                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8009                 pgoff = bvec->bv_offset;
8010
8011 next_block_or_try_again:
8012                 done.uptodate = 0;
8013                 done.start = start;
8014                 init_completion(&done.done);
8015
8016                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8017                                 pgoff, start, start + sectorsize - 1,
8018                                 io_bio->mirror_num,
8019                                 btrfs_retry_endio_nocsum, &done);
8020                 if (ret)
8021                         return ret;
8022
8023                 wait_for_completion(&done.done);
8024
8025                 if (!done.uptodate) {
8026                         /* We might have another mirror, so try again */
8027                         goto next_block_or_try_again;
8028                 }
8029
8030                 start += sectorsize;
8031
8032                 if (nr_sectors--) {
8033                         pgoff += sectorsize;
8034                         goto next_block_or_try_again;
8035                 }
8036         }
8037
8038         return 0;
8039 }
8040
8041 static void btrfs_retry_endio(struct bio *bio)
8042 {
8043         struct btrfs_retry_complete *done = bio->bi_private;
8044         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8045         struct inode *inode;
8046         struct bio_vec *bvec;
8047         u64 start;
8048         int uptodate;
8049         int ret;
8050         int i;
8051
8052         if (bio->bi_error)
8053                 goto end;
8054
8055         uptodate = 1;
8056
8057         start = done->start;
8058
8059         ASSERT(bio->bi_vcnt == 1);
8060         inode = bio->bi_io_vec->bv_page->mapping->host;
8061         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(inode));
8062
8063         bio_for_each_segment_all(bvec, bio, i) {
8064                 ret = __readpage_endio_check(done->inode, io_bio, i,
8065                                         bvec->bv_page, bvec->bv_offset,
8066                                         done->start, bvec->bv_len);
8067                 if (!ret)
8068                         clean_io_failure(done->inode, done->start,
8069                                         bvec->bv_page, bvec->bv_offset);
8070                 else
8071                         uptodate = 0;
8072         }
8073
8074         done->uptodate = uptodate;
8075 end:
8076         complete(&done->done);
8077         bio_put(bio);
8078 }
8079
8080 static int __btrfs_subio_endio_read(struct inode *inode,
8081                                     struct btrfs_io_bio *io_bio, int err)
8082 {
8083         struct btrfs_fs_info *fs_info;
8084         struct bio_vec *bvec;
8085         struct btrfs_retry_complete done;
8086         u64 start;
8087         u64 offset = 0;
8088         u32 sectorsize;
8089         int nr_sectors;
8090         unsigned int pgoff;
8091         int csum_pos;
8092         int i;
8093         int ret;
8094
8095         fs_info = BTRFS_I(inode)->root->fs_info;
8096         sectorsize = fs_info->sectorsize;
8097
8098         err = 0;
8099         start = io_bio->logical;
8100         done.inode = inode;
8101
8102         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8103                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8104
8105                 pgoff = bvec->bv_offset;
8106 next_block:
8107                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8108                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8109                                         bvec->bv_page, pgoff, start,
8110                                         sectorsize);
8111                 if (likely(!ret))
8112                         goto next;
8113 try_again:
8114                 done.uptodate = 0;
8115                 done.start = start;
8116                 init_completion(&done.done);
8117
8118                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8119                                 pgoff, start, start + sectorsize - 1,
8120                                 io_bio->mirror_num,
8121                                 btrfs_retry_endio, &done);
8122                 if (ret) {
8123                         err = ret;
8124                         goto next;
8125                 }
8126
8127                 wait_for_completion(&done.done);
8128
8129                 if (!done.uptodate) {
8130                         /* We might have another mirror, so try again */
8131                         goto try_again;
8132                 }
8133 next:
8134                 offset += sectorsize;
8135                 start += sectorsize;
8136
8137                 ASSERT(nr_sectors);
8138
8139                 if (--nr_sectors) {
8140                         pgoff += sectorsize;
8141                         goto next_block;
8142                 }
8143         }
8144
8145         return err;
8146 }
8147
8148 static int btrfs_subio_endio_read(struct inode *inode,
8149                                   struct btrfs_io_bio *io_bio, int err)
8150 {
8151         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8152
8153         if (skip_csum) {
8154                 if (unlikely(err))
8155                         return __btrfs_correct_data_nocsum(inode, io_bio);
8156                 else
8157                         return 0;
8158         } else {
8159                 return __btrfs_subio_endio_read(inode, io_bio, err);
8160         }
8161 }
8162
8163 static void btrfs_endio_direct_read(struct bio *bio)
8164 {
8165         struct btrfs_dio_private *dip = bio->bi_private;
8166         struct inode *inode = dip->inode;
8167         struct bio *dio_bio;
8168         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8169         int err = bio->bi_error;
8170
8171         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8172                 err = btrfs_subio_endio_read(inode, io_bio, err);
8173
8174         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8175                       dip->logical_offset + dip->bytes - 1);
8176         dio_bio = dip->dio_bio;
8177
8178         kfree(dip);
8179
8180         dio_bio->bi_error = bio->bi_error;
8181         dio_end_io(dio_bio, bio->bi_error);
8182
8183         if (io_bio->end_io)
8184                 io_bio->end_io(io_bio, err);
8185         bio_put(bio);
8186 }
8187
8188 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8189                                                     const u64 offset,
8190                                                     const u64 bytes,
8191                                                     const int uptodate)
8192 {
8193         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8194         struct btrfs_ordered_extent *ordered = NULL;
8195         u64 ordered_offset = offset;
8196         u64 ordered_bytes = bytes;
8197         int ret;
8198
8199 again:
8200         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8201                                                    &ordered_offset,
8202                                                    ordered_bytes,
8203                                                    uptodate);
8204         if (!ret)
8205                 goto out_test;
8206
8207         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8208                         finish_ordered_fn, NULL, NULL);
8209         btrfs_queue_work(fs_info->endio_write_workers, &ordered->work);
8210 out_test:
8211         /*
8212          * our bio might span multiple ordered extents.  If we haven't
8213          * completed the accounting for the whole dio, go back and try again
8214          */
8215         if (ordered_offset < offset + bytes) {
8216                 ordered_bytes = offset + bytes - ordered_offset;
8217                 ordered = NULL;
8218                 goto again;
8219         }
8220 }
8221
8222 static void btrfs_endio_direct_write(struct bio *bio)
8223 {
8224         struct btrfs_dio_private *dip = bio->bi_private;
8225         struct bio *dio_bio = dip->dio_bio;
8226
8227         btrfs_endio_direct_write_update_ordered(dip->inode,
8228                                                 dip->logical_offset,
8229                                                 dip->bytes,
8230                                                 !bio->bi_error);
8231
8232         kfree(dip);
8233
8234         dio_bio->bi_error = bio->bi_error;
8235         dio_end_io(dio_bio, bio->bi_error);
8236         bio_put(bio);
8237 }
8238
8239 static int __btrfs_submit_bio_start_direct_io(struct inode *inode,
8240                                     struct bio *bio, int mirror_num,
8241                                     unsigned long bio_flags, u64 offset)
8242 {
8243         int ret;
8244         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8245         BUG_ON(ret); /* -ENOMEM */
8246         return 0;
8247 }
8248
8249 static void btrfs_end_dio_bio(struct bio *bio)
8250 {
8251         struct btrfs_dio_private *dip = bio->bi_private;
8252         int err = bio->bi_error;
8253
8254         if (err)
8255                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8256                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8257                            btrfs_ino(dip->inode), bio_op(bio), bio->bi_opf,
8258                            (unsigned long long)bio->bi_iter.bi_sector,
8259                            bio->bi_iter.bi_size, err);
8260
8261         if (dip->subio_endio)
8262                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8263
8264         if (err) {
8265                 dip->errors = 1;
8266
8267                 /*
8268                  * before atomic variable goto zero, we must make sure
8269                  * dip->errors is perceived to be set.
8270                  */
8271                 smp_mb__before_atomic();
8272         }
8273
8274         /* if there are more bios still pending for this dio, just exit */
8275         if (!atomic_dec_and_test(&dip->pending_bios))
8276                 goto out;
8277
8278         if (dip->errors) {
8279                 bio_io_error(dip->orig_bio);
8280         } else {
8281                 dip->dio_bio->bi_error = 0;
8282                 bio_endio(dip->orig_bio);
8283         }
8284 out:
8285         bio_put(bio);
8286 }
8287
8288 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8289                                        u64 first_sector, gfp_t gfp_flags)
8290 {
8291         struct bio *bio;
8292         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8293         if (bio)
8294                 bio_associate_current(bio);
8295         return bio;
8296 }
8297
8298 static inline int btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8299                                                  struct btrfs_dio_private *dip,
8300                                                  struct bio *bio,
8301                                                  u64 file_offset)
8302 {
8303         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8304         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8305         int ret;
8306
8307         /*
8308          * We load all the csum data we need when we submit
8309          * the first bio to reduce the csum tree search and
8310          * contention.
8311          */
8312         if (dip->logical_offset == file_offset) {
8313                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8314                                                 file_offset);
8315                 if (ret)
8316                         return ret;
8317         }
8318
8319         if (bio == dip->orig_bio)
8320                 return 0;
8321
8322         file_offset -= dip->logical_offset;
8323         file_offset >>= inode->i_sb->s_blocksize_bits;
8324         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8325
8326         return 0;
8327 }
8328
8329 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8330                                          u64 file_offset, int skip_sum,
8331                                          int async_submit)
8332 {
8333         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8334         struct btrfs_dio_private *dip = bio->bi_private;
8335         bool write = bio_op(bio) == REQ_OP_WRITE;
8336         int ret;
8337
8338         if (async_submit)
8339                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8340
8341         bio_get(bio);
8342
8343         if (!write) {
8344                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8345                 if (ret)
8346                         goto err;
8347         }
8348
8349         if (skip_sum)
8350                 goto map;
8351
8352         if (write && async_submit) {
8353                 ret = btrfs_wq_submit_bio(fs_info, inode, bio, 0, 0,
8354                                           file_offset,
8355                                           __btrfs_submit_bio_start_direct_io,
8356                                           __btrfs_submit_bio_done);
8357                 goto err;
8358         } else if (write) {
8359                 /*
8360                  * If we aren't doing async submit, calculate the csum of the
8361                  * bio now.
8362                  */
8363                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8364                 if (ret)
8365                         goto err;
8366         } else {
8367                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8368                                                      file_offset);
8369                 if (ret)
8370                         goto err;
8371         }
8372 map:
8373         ret = btrfs_map_bio(fs_info, bio, 0, async_submit);
8374 err:
8375         bio_put(bio);
8376         return ret;
8377 }
8378
8379 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip,
8380                                     int skip_sum)
8381 {
8382         struct inode *inode = dip->inode;
8383         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8384         struct btrfs_root *root = BTRFS_I(inode)->root;
8385         struct bio *bio;
8386         struct bio *orig_bio = dip->orig_bio;
8387         struct bio_vec *bvec;
8388         u64 start_sector = orig_bio->bi_iter.bi_sector;
8389         u64 file_offset = dip->logical_offset;
8390         u64 submit_len = 0;
8391         u64 map_length;
8392         u32 blocksize = fs_info->sectorsize;
8393         int async_submit = 0;
8394         int nr_sectors;
8395         int ret;
8396         int i, j;
8397
8398         map_length = orig_bio->bi_iter.bi_size;
8399         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8400                               &map_length, NULL, 0);
8401         if (ret)
8402                 return -EIO;
8403
8404         if (map_length >= orig_bio->bi_iter.bi_size) {
8405                 bio = orig_bio;
8406                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8407                 goto submit;
8408         }
8409
8410         /* async crcs make it difficult to collect full stripe writes. */
8411         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8412                 async_submit = 0;
8413         else
8414                 async_submit = 1;
8415
8416         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8417         if (!bio)
8418                 return -ENOMEM;
8419
8420         bio->bi_opf = orig_bio->bi_opf;
8421         bio->bi_private = dip;
8422         bio->bi_end_io = btrfs_end_dio_bio;
8423         btrfs_io_bio(bio)->logical = file_offset;
8424         atomic_inc(&dip->pending_bios);
8425
8426         bio_for_each_segment_all(bvec, orig_bio, j) {
8427                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8428                 i = 0;
8429 next_block:
8430                 if (unlikely(map_length < submit_len + blocksize ||
8431                     bio_add_page(bio, bvec->bv_page, blocksize,
8432                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8433                         /*
8434                          * inc the count before we submit the bio so
8435                          * we know the end IO handler won't happen before
8436                          * we inc the count. Otherwise, the dip might get freed
8437                          * before we're done setting it up
8438                          */
8439                         atomic_inc(&dip->pending_bios);
8440                         ret = __btrfs_submit_dio_bio(bio, inode,
8441                                                      file_offset, skip_sum,
8442                                                      async_submit);
8443                         if (ret) {
8444                                 bio_put(bio);
8445                                 atomic_dec(&dip->pending_bios);
8446                                 goto out_err;
8447                         }
8448
8449                         start_sector += submit_len >> 9;
8450                         file_offset += submit_len;
8451
8452                         submit_len = 0;
8453
8454                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8455                                                   start_sector, GFP_NOFS);
8456                         if (!bio)
8457                                 goto out_err;
8458                         bio->bi_opf = orig_bio->bi_opf;
8459                         bio->bi_private = dip;
8460                         bio->bi_end_io = btrfs_end_dio_bio;
8461                         btrfs_io_bio(bio)->logical = file_offset;
8462
8463                         map_length = orig_bio->bi_iter.bi_size;
8464                         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8465                                               start_sector << 9,
8466                                               &map_length, NULL, 0);
8467                         if (ret) {
8468                                 bio_put(bio);
8469                                 goto out_err;
8470                         }
8471
8472                         goto next_block;
8473                 } else {
8474                         submit_len += blocksize;
8475                         if (--nr_sectors) {
8476                                 i++;
8477                                 goto next_block;
8478                         }
8479                 }
8480         }
8481
8482 submit:
8483         ret = __btrfs_submit_dio_bio(bio, inode, file_offset, skip_sum,
8484                                      async_submit);
8485         if (!ret)
8486                 return 0;
8487
8488         bio_put(bio);
8489 out_err:
8490         dip->errors = 1;
8491         /*
8492          * before atomic variable goto zero, we must
8493          * make sure dip->errors is perceived to be set.
8494          */
8495         smp_mb__before_atomic();
8496         if (atomic_dec_and_test(&dip->pending_bios))
8497                 bio_io_error(dip->orig_bio);
8498
8499         /* bio_end_io() will handle error, so we needn't return it */
8500         return 0;
8501 }
8502
8503 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8504                                 loff_t file_offset)
8505 {
8506         struct btrfs_dio_private *dip = NULL;
8507         struct bio *io_bio = NULL;
8508         struct btrfs_io_bio *btrfs_bio;
8509         int skip_sum;
8510         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8511         int ret = 0;
8512
8513         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8514
8515         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8516         if (!io_bio) {
8517                 ret = -ENOMEM;
8518                 goto free_ordered;
8519         }
8520
8521         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8522         if (!dip) {
8523                 ret = -ENOMEM;
8524                 goto free_ordered;
8525         }
8526
8527         dip->private = dio_bio->bi_private;
8528         dip->inode = inode;
8529         dip->logical_offset = file_offset;
8530         dip->bytes = dio_bio->bi_iter.bi_size;
8531         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8532         io_bio->bi_private = dip;
8533         dip->orig_bio = io_bio;
8534         dip->dio_bio = dio_bio;
8535         atomic_set(&dip->pending_bios, 0);
8536         btrfs_bio = btrfs_io_bio(io_bio);
8537         btrfs_bio->logical = file_offset;
8538
8539         if (write) {
8540                 io_bio->bi_end_io = btrfs_endio_direct_write;
8541         } else {
8542                 io_bio->bi_end_io = btrfs_endio_direct_read;
8543                 dip->subio_endio = btrfs_subio_endio_read;
8544         }
8545
8546         /*
8547          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8548          * even if we fail to submit a bio, because in such case we do the
8549          * corresponding error handling below and it must not be done a second
8550          * time by btrfs_direct_IO().
8551          */
8552         if (write) {
8553                 struct btrfs_dio_data *dio_data = current->journal_info;
8554
8555                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8556                         dip->bytes;
8557                 dio_data->unsubmitted_oe_range_start =
8558                         dio_data->unsubmitted_oe_range_end;
8559         }
8560
8561         ret = btrfs_submit_direct_hook(dip, skip_sum);
8562         if (!ret)
8563                 return;
8564
8565         if (btrfs_bio->end_io)
8566                 btrfs_bio->end_io(btrfs_bio, ret);
8567
8568 free_ordered:
8569         /*
8570          * If we arrived here it means either we failed to submit the dip
8571          * or we either failed to clone the dio_bio or failed to allocate the
8572          * dip. If we cloned the dio_bio and allocated the dip, we can just
8573          * call bio_endio against our io_bio so that we get proper resource
8574          * cleanup if we fail to submit the dip, otherwise, we must do the
8575          * same as btrfs_endio_direct_[write|read] because we can't call these
8576          * callbacks - they require an allocated dip and a clone of dio_bio.
8577          */
8578         if (io_bio && dip) {
8579                 io_bio->bi_error = -EIO;
8580                 bio_endio(io_bio);
8581                 /*
8582                  * The end io callbacks free our dip, do the final put on io_bio
8583                  * and all the cleanup and final put for dio_bio (through
8584                  * dio_end_io()).
8585                  */
8586                 dip = NULL;
8587                 io_bio = NULL;
8588         } else {
8589                 if (write)
8590                         btrfs_endio_direct_write_update_ordered(inode,
8591                                                 file_offset,
8592                                                 dio_bio->bi_iter.bi_size,
8593                                                 0);
8594                 else
8595                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8596                               file_offset + dio_bio->bi_iter.bi_size - 1);
8597
8598                 dio_bio->bi_error = -EIO;
8599                 /*
8600                  * Releases and cleans up our dio_bio, no need to bio_put()
8601                  * nor bio_endio()/bio_io_error() against dio_bio.
8602                  */
8603                 dio_end_io(dio_bio, ret);
8604         }
8605         if (io_bio)
8606                 bio_put(io_bio);
8607         kfree(dip);
8608 }
8609
8610 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8611                                struct kiocb *iocb,
8612                                const struct iov_iter *iter, loff_t offset)
8613 {
8614         int seg;
8615         int i;
8616         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8617         ssize_t retval = -EINVAL;
8618
8619         if (offset & blocksize_mask)
8620                 goto out;
8621
8622         if (iov_iter_alignment(iter) & blocksize_mask)
8623                 goto out;
8624
8625         /* If this is a write we don't need to check anymore */
8626         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8627                 return 0;
8628         /*
8629          * Check to make sure we don't have duplicate iov_base's in this
8630          * iovec, if so return EINVAL, otherwise we'll get csum errors
8631          * when reading back.
8632          */
8633         for (seg = 0; seg < iter->nr_segs; seg++) {
8634                 for (i = seg + 1; i < iter->nr_segs; i++) {
8635                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8636                                 goto out;
8637                 }
8638         }
8639         retval = 0;
8640 out:
8641         return retval;
8642 }
8643
8644 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8645 {
8646         struct file *file = iocb->ki_filp;
8647         struct inode *inode = file->f_mapping->host;
8648         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8649         struct btrfs_dio_data dio_data = { 0 };
8650         loff_t offset = iocb->ki_pos;
8651         size_t count = 0;
8652         int flags = 0;
8653         bool wakeup = true;
8654         bool relock = false;
8655         ssize_t ret;
8656
8657         if (check_direct_IO(fs_info, iocb, iter, offset))
8658                 return 0;
8659
8660         inode_dio_begin(inode);
8661         smp_mb__after_atomic();
8662
8663         /*
8664          * The generic stuff only does filemap_write_and_wait_range, which
8665          * isn't enough if we've written compressed pages to this area, so
8666          * we need to flush the dirty pages again to make absolutely sure
8667          * that any outstanding dirty pages are on disk.
8668          */
8669         count = iov_iter_count(iter);
8670         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8671                      &BTRFS_I(inode)->runtime_flags))
8672                 filemap_fdatawrite_range(inode->i_mapping, offset,
8673                                          offset + count - 1);
8674
8675         if (iov_iter_rw(iter) == WRITE) {
8676                 /*
8677                  * If the write DIO is beyond the EOF, we need update
8678                  * the isize, but it is protected by i_mutex. So we can
8679                  * not unlock the i_mutex at this case.
8680                  */
8681                 if (offset + count <= inode->i_size) {
8682                         inode_unlock(inode);
8683                         relock = true;
8684                 }
8685                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8686                 if (ret)
8687                         goto out;
8688                 dio_data.outstanding_extents = div64_u64(count +
8689                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8690                                                 BTRFS_MAX_EXTENT_SIZE);
8691
8692                 /*
8693                  * We need to know how many extents we reserved so that we can
8694                  * do the accounting properly if we go over the number we
8695                  * originally calculated.  Abuse current->journal_info for this.
8696                  */
8697                 dio_data.reserve = round_up(count,
8698                                             fs_info->sectorsize);
8699                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8700                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8701                 current->journal_info = &dio_data;
8702                 down_read(&BTRFS_I(inode)->dio_sem);
8703         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8704                                      &BTRFS_I(inode)->runtime_flags)) {
8705                 inode_dio_end(inode);
8706                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8707                 wakeup = false;
8708         }
8709
8710         ret = __blockdev_direct_IO(iocb, inode,
8711                                    fs_info->fs_devices->latest_bdev,
8712                                    iter, btrfs_get_blocks_direct, NULL,
8713                                    btrfs_submit_direct, flags);
8714         if (iov_iter_rw(iter) == WRITE) {
8715                 up_read(&BTRFS_I(inode)->dio_sem);
8716                 current->journal_info = NULL;
8717                 if (ret < 0 && ret != -EIOCBQUEUED) {
8718                         if (dio_data.reserve)
8719                                 btrfs_delalloc_release_space(inode, offset,
8720                                                              dio_data.reserve);
8721                         /*
8722                          * On error we might have left some ordered extents
8723                          * without submitting corresponding bios for them, so
8724                          * cleanup them up to avoid other tasks getting them
8725                          * and waiting for them to complete forever.
8726                          */
8727                         if (dio_data.unsubmitted_oe_range_start <
8728                             dio_data.unsubmitted_oe_range_end)
8729                                 btrfs_endio_direct_write_update_ordered(inode,
8730                                         dio_data.unsubmitted_oe_range_start,
8731                                         dio_data.unsubmitted_oe_range_end -
8732                                         dio_data.unsubmitted_oe_range_start,
8733                                         0);
8734                 } else if (ret >= 0 && (size_t)ret < count)
8735                         btrfs_delalloc_release_space(inode, offset,
8736                                                      count - (size_t)ret);
8737         }
8738 out:
8739         if (wakeup)
8740                 inode_dio_end(inode);
8741         if (relock)
8742                 inode_lock(inode);
8743
8744         return ret;
8745 }
8746
8747 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8748
8749 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8750                 __u64 start, __u64 len)
8751 {
8752         int     ret;
8753
8754         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8755         if (ret)
8756                 return ret;
8757
8758         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8759 }
8760
8761 int btrfs_readpage(struct file *file, struct page *page)
8762 {
8763         struct extent_io_tree *tree;
8764         tree = &BTRFS_I(page->mapping->host)->io_tree;
8765         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8766 }
8767
8768 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8769 {
8770         struct extent_io_tree *tree;
8771         struct inode *inode = page->mapping->host;
8772         int ret;
8773
8774         if (current->flags & PF_MEMALLOC) {
8775                 redirty_page_for_writepage(wbc, page);
8776                 unlock_page(page);
8777                 return 0;
8778         }
8779
8780         /*
8781          * If we are under memory pressure we will call this directly from the
8782          * VM, we need to make sure we have the inode referenced for the ordered
8783          * extent.  If not just return like we didn't do anything.
8784          */
8785         if (!igrab(inode)) {
8786                 redirty_page_for_writepage(wbc, page);
8787                 return AOP_WRITEPAGE_ACTIVATE;
8788         }
8789         tree = &BTRFS_I(page->mapping->host)->io_tree;
8790         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8791         btrfs_add_delayed_iput(inode);
8792         return ret;
8793 }
8794
8795 static int btrfs_writepages(struct address_space *mapping,
8796                             struct writeback_control *wbc)
8797 {
8798         struct extent_io_tree *tree;
8799
8800         tree = &BTRFS_I(mapping->host)->io_tree;
8801         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8802 }
8803
8804 static int
8805 btrfs_readpages(struct file *file, struct address_space *mapping,
8806                 struct list_head *pages, unsigned nr_pages)
8807 {
8808         struct extent_io_tree *tree;
8809         tree = &BTRFS_I(mapping->host)->io_tree;
8810         return extent_readpages(tree, mapping, pages, nr_pages,
8811                                 btrfs_get_extent);
8812 }
8813 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8814 {
8815         struct extent_io_tree *tree;
8816         struct extent_map_tree *map;
8817         int ret;
8818
8819         tree = &BTRFS_I(page->mapping->host)->io_tree;
8820         map = &BTRFS_I(page->mapping->host)->extent_tree;
8821         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8822         if (ret == 1) {
8823                 ClearPagePrivate(page);
8824                 set_page_private(page, 0);
8825                 put_page(page);
8826         }
8827         return ret;
8828 }
8829
8830 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8831 {
8832         if (PageWriteback(page) || PageDirty(page))
8833                 return 0;
8834         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8835 }
8836
8837 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8838                                  unsigned int length)
8839 {
8840         struct inode *inode = page->mapping->host;
8841         struct extent_io_tree *tree;
8842         struct btrfs_ordered_extent *ordered;
8843         struct extent_state *cached_state = NULL;
8844         u64 page_start = page_offset(page);
8845         u64 page_end = page_start + PAGE_SIZE - 1;
8846         u64 start;
8847         u64 end;
8848         int inode_evicting = inode->i_state & I_FREEING;
8849
8850         /*
8851          * we have the page locked, so new writeback can't start,
8852          * and the dirty bit won't be cleared while we are here.
8853          *
8854          * Wait for IO on this page so that we can safely clear
8855          * the PagePrivate2 bit and do ordered accounting
8856          */
8857         wait_on_page_writeback(page);
8858
8859         tree = &BTRFS_I(inode)->io_tree;
8860         if (offset) {
8861                 btrfs_releasepage(page, GFP_NOFS);
8862                 return;
8863         }
8864
8865         if (!inode_evicting)
8866                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8867 again:
8868         start = page_start;
8869         ordered = btrfs_lookup_ordered_range(inode, start,
8870                                         page_end - start + 1);
8871         if (ordered) {
8872                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8873                 /*
8874                  * IO on this page will never be started, so we need
8875                  * to account for any ordered extents now
8876                  */
8877                 if (!inode_evicting)
8878                         clear_extent_bit(tree, start, end,
8879                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8880                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8881                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8882                                          GFP_NOFS);
8883                 /*
8884                  * whoever cleared the private bit is responsible
8885                  * for the finish_ordered_io
8886                  */
8887                 if (TestClearPagePrivate2(page)) {
8888                         struct btrfs_ordered_inode_tree *tree;
8889                         u64 new_len;
8890
8891                         tree = &BTRFS_I(inode)->ordered_tree;
8892
8893                         spin_lock_irq(&tree->lock);
8894                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8895                         new_len = start - ordered->file_offset;
8896                         if (new_len < ordered->truncated_len)
8897                                 ordered->truncated_len = new_len;
8898                         spin_unlock_irq(&tree->lock);
8899
8900                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8901                                                            start,
8902                                                            end - start + 1, 1))
8903                                 btrfs_finish_ordered_io(ordered);
8904                 }
8905                 btrfs_put_ordered_extent(ordered);
8906                 if (!inode_evicting) {
8907                         cached_state = NULL;
8908                         lock_extent_bits(tree, start, end,
8909                                          &cached_state);
8910                 }
8911
8912                 start = end + 1;
8913                 if (start < page_end)
8914                         goto again;
8915         }
8916
8917         /*
8918          * Qgroup reserved space handler
8919          * Page here will be either
8920          * 1) Already written to disk
8921          *    In this case, its reserved space is released from data rsv map
8922          *    and will be freed by delayed_ref handler finally.
8923          *    So even we call qgroup_free_data(), it won't decrease reserved
8924          *    space.
8925          * 2) Not written to disk
8926          *    This means the reserved space should be freed here. However,
8927          *    if a truncate invalidates the page (by clearing PageDirty)
8928          *    and the page is accounted for while allocating extent
8929          *    in btrfs_check_data_free_space() we let delayed_ref to
8930          *    free the entire extent.
8931          */
8932         if (PageDirty(page))
8933                 btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
8934         if (!inode_evicting) {
8935                 clear_extent_bit(tree, page_start, page_end,
8936                                  EXTENT_LOCKED | EXTENT_DIRTY |
8937                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8938                                  EXTENT_DEFRAG, 1, 1,
8939                                  &cached_state, GFP_NOFS);
8940
8941                 __btrfs_releasepage(page, GFP_NOFS);
8942         }
8943
8944         ClearPageChecked(page);
8945         if (PagePrivate(page)) {
8946                 ClearPagePrivate(page);
8947                 set_page_private(page, 0);
8948                 put_page(page);
8949         }
8950 }
8951
8952 /*
8953  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8954  * called from a page fault handler when a page is first dirtied. Hence we must
8955  * be careful to check for EOF conditions here. We set the page up correctly
8956  * for a written page which means we get ENOSPC checking when writing into
8957  * holes and correct delalloc and unwritten extent mapping on filesystems that
8958  * support these features.
8959  *
8960  * We are not allowed to take the i_mutex here so we have to play games to
8961  * protect against truncate races as the page could now be beyond EOF.  Because
8962  * vmtruncate() writes the inode size before removing pages, once we have the
8963  * page lock we can determine safely if the page is beyond EOF. If it is not
8964  * beyond EOF, then the page is guaranteed safe against truncation until we
8965  * unlock the page.
8966  */
8967 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8968 {
8969         struct page *page = vmf->page;
8970         struct inode *inode = file_inode(vma->vm_file);
8971         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8972         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8973         struct btrfs_ordered_extent *ordered;
8974         struct extent_state *cached_state = NULL;
8975         char *kaddr;
8976         unsigned long zero_start;
8977         loff_t size;
8978         int ret;
8979         int reserved = 0;
8980         u64 reserved_space;
8981         u64 page_start;
8982         u64 page_end;
8983         u64 end;
8984
8985         reserved_space = PAGE_SIZE;
8986
8987         sb_start_pagefault(inode->i_sb);
8988         page_start = page_offset(page);
8989         page_end = page_start + PAGE_SIZE - 1;
8990         end = page_end;
8991
8992         /*
8993          * Reserving delalloc space after obtaining the page lock can lead to
8994          * deadlock. For example, if a dirty page is locked by this function
8995          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8996          * dirty page write out, then the btrfs_writepage() function could
8997          * end up waiting indefinitely to get a lock on the page currently
8998          * being processed by btrfs_page_mkwrite() function.
8999          */
9000         ret = btrfs_delalloc_reserve_space(inode, page_start,
9001                                            reserved_space);
9002         if (!ret) {
9003                 ret = file_update_time(vma->vm_file);
9004                 reserved = 1;
9005         }
9006         if (ret) {
9007                 if (ret == -ENOMEM)
9008                         ret = VM_FAULT_OOM;
9009                 else /* -ENOSPC, -EIO, etc */
9010                         ret = VM_FAULT_SIGBUS;
9011                 if (reserved)
9012                         goto out;
9013                 goto out_noreserve;
9014         }
9015
9016         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9017 again:
9018         lock_page(page);
9019         size = i_size_read(inode);
9020
9021         if ((page->mapping != inode->i_mapping) ||
9022             (page_start >= size)) {
9023                 /* page got truncated out from underneath us */
9024                 goto out_unlock;
9025         }
9026         wait_on_page_writeback(page);
9027
9028         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9029         set_page_extent_mapped(page);
9030
9031         /*
9032          * we can't set the delalloc bits if there are pending ordered
9033          * extents.  Drop our locks and wait for them to finish
9034          */
9035         ordered = btrfs_lookup_ordered_range(inode, page_start, page_end);
9036         if (ordered) {
9037                 unlock_extent_cached(io_tree, page_start, page_end,
9038                                      &cached_state, GFP_NOFS);
9039                 unlock_page(page);
9040                 btrfs_start_ordered_extent(inode, ordered, 1);
9041                 btrfs_put_ordered_extent(ordered);
9042                 goto again;
9043         }
9044
9045         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9046                 reserved_space = round_up(size - page_start,
9047                                           fs_info->sectorsize);
9048                 if (reserved_space < PAGE_SIZE) {
9049                         end = page_start + reserved_space - 1;
9050                         spin_lock(&BTRFS_I(inode)->lock);
9051                         BTRFS_I(inode)->outstanding_extents++;
9052                         spin_unlock(&BTRFS_I(inode)->lock);
9053                         btrfs_delalloc_release_space(inode, page_start,
9054                                                 PAGE_SIZE - reserved_space);
9055                 }
9056         }
9057
9058         /*
9059          * XXX - page_mkwrite gets called every time the page is dirtied, even
9060          * if it was already dirty, so for space accounting reasons we need to
9061          * clear any delalloc bits for the range we are fixing to save.  There
9062          * is probably a better way to do this, but for now keep consistent with
9063          * prepare_pages in the normal write path.
9064          */
9065         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9066                           EXTENT_DIRTY | EXTENT_DELALLOC |
9067                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9068                           0, 0, &cached_state, GFP_NOFS);
9069
9070         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9071                                         &cached_state, 0);
9072         if (ret) {
9073                 unlock_extent_cached(io_tree, page_start, page_end,
9074                                      &cached_state, GFP_NOFS);
9075                 ret = VM_FAULT_SIGBUS;
9076                 goto out_unlock;
9077         }
9078         ret = 0;
9079
9080         /* page is wholly or partially inside EOF */
9081         if (page_start + PAGE_SIZE > size)
9082                 zero_start = size & ~PAGE_MASK;
9083         else
9084                 zero_start = PAGE_SIZE;
9085
9086         if (zero_start != PAGE_SIZE) {
9087                 kaddr = kmap(page);
9088                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9089                 flush_dcache_page(page);
9090                 kunmap(page);
9091         }
9092         ClearPageChecked(page);
9093         set_page_dirty(page);
9094         SetPageUptodate(page);
9095
9096         BTRFS_I(inode)->last_trans = fs_info->generation;
9097         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9098         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9099
9100         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9101
9102 out_unlock:
9103         if (!ret) {
9104                 sb_end_pagefault(inode->i_sb);
9105                 return VM_FAULT_LOCKED;
9106         }
9107         unlock_page(page);
9108 out:
9109         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9110 out_noreserve:
9111         sb_end_pagefault(inode->i_sb);
9112         return ret;
9113 }
9114
9115 static int btrfs_truncate(struct inode *inode)
9116 {
9117         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9118         struct btrfs_root *root = BTRFS_I(inode)->root;
9119         struct btrfs_block_rsv *rsv;
9120         int ret = 0;
9121         int err = 0;
9122         struct btrfs_trans_handle *trans;
9123         u64 mask = fs_info->sectorsize - 1;
9124         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9125
9126         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9127                                        (u64)-1);
9128         if (ret)
9129                 return ret;
9130
9131         /*
9132          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9133          * 3 things going on here
9134          *
9135          * 1) We need to reserve space for our orphan item and the space to
9136          * delete our orphan item.  Lord knows we don't want to have a dangling
9137          * orphan item because we didn't reserve space to remove it.
9138          *
9139          * 2) We need to reserve space to update our inode.
9140          *
9141          * 3) We need to have something to cache all the space that is going to
9142          * be free'd up by the truncate operation, but also have some slack
9143          * space reserved in case it uses space during the truncate (thank you
9144          * very much snapshotting).
9145          *
9146          * And we need these to all be separate.  The fact is we can use a lot of
9147          * space doing the truncate, and we have no earthly idea how much space
9148          * we will use, so we need the truncate reservation to be separate so it
9149          * doesn't end up using space reserved for updating the inode or
9150          * removing the orphan item.  We also need to be able to stop the
9151          * transaction and start a new one, which means we need to be able to
9152          * update the inode several times, and we have no idea of knowing how
9153          * many times that will be, so we can't just reserve 1 item for the
9154          * entirety of the operation, so that has to be done separately as well.
9155          * Then there is the orphan item, which does indeed need to be held on
9156          * to for the whole operation, and we need nobody to touch this reserved
9157          * space except the orphan code.
9158          *
9159          * So that leaves us with
9160          *
9161          * 1) root->orphan_block_rsv - for the orphan deletion.
9162          * 2) rsv - for the truncate reservation, which we will steal from the
9163          * transaction reservation.
9164          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9165          * updating the inode.
9166          */
9167         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9168         if (!rsv)
9169                 return -ENOMEM;
9170         rsv->size = min_size;
9171         rsv->failfast = 1;
9172
9173         /*
9174          * 1 for the truncate slack space
9175          * 1 for updating the inode.
9176          */
9177         trans = btrfs_start_transaction(root, 2);
9178         if (IS_ERR(trans)) {
9179                 err = PTR_ERR(trans);
9180                 goto out;
9181         }
9182
9183         /* Migrate the slack space for the truncate to our reserve */
9184         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9185                                       min_size, 0);
9186         BUG_ON(ret);
9187
9188         /*
9189          * So if we truncate and then write and fsync we normally would just
9190          * write the extents that changed, which is a problem if we need to
9191          * first truncate that entire inode.  So set this flag so we write out
9192          * all of the extents in the inode to the sync log so we're completely
9193          * safe.
9194          */
9195         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9196         trans->block_rsv = rsv;
9197
9198         while (1) {
9199                 ret = btrfs_truncate_inode_items(trans, root, inode,
9200                                                  inode->i_size,
9201                                                  BTRFS_EXTENT_DATA_KEY);
9202                 if (ret != -ENOSPC && ret != -EAGAIN) {
9203                         err = ret;
9204                         break;
9205                 }
9206
9207                 trans->block_rsv = &fs_info->trans_block_rsv;
9208                 ret = btrfs_update_inode(trans, root, inode);
9209                 if (ret) {
9210                         err = ret;
9211                         break;
9212                 }
9213
9214                 btrfs_end_transaction(trans);
9215                 btrfs_btree_balance_dirty(fs_info);
9216
9217                 trans = btrfs_start_transaction(root, 2);
9218                 if (IS_ERR(trans)) {
9219                         ret = err = PTR_ERR(trans);
9220                         trans = NULL;
9221                         break;
9222                 }
9223
9224                 btrfs_block_rsv_release(fs_info, rsv, -1);
9225                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9226                                               rsv, min_size, 0);
9227                 BUG_ON(ret);    /* shouldn't happen */
9228                 trans->block_rsv = rsv;
9229         }
9230
9231         if (ret == 0 && inode->i_nlink > 0) {
9232                 trans->block_rsv = root->orphan_block_rsv;
9233                 ret = btrfs_orphan_del(trans, inode);
9234                 if (ret)
9235                         err = ret;
9236         }
9237
9238         if (trans) {
9239                 trans->block_rsv = &fs_info->trans_block_rsv;
9240                 ret = btrfs_update_inode(trans, root, inode);
9241                 if (ret && !err)
9242                         err = ret;
9243
9244                 ret = btrfs_end_transaction(trans);
9245                 btrfs_btree_balance_dirty(fs_info);
9246         }
9247 out:
9248         btrfs_free_block_rsv(fs_info, rsv);
9249
9250         if (ret && !err)
9251                 err = ret;
9252
9253         return err;
9254 }
9255
9256 /*
9257  * create a new subvolume directory/inode (helper for the ioctl).
9258  */
9259 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9260                              struct btrfs_root *new_root,
9261                              struct btrfs_root *parent_root,
9262                              u64 new_dirid)
9263 {
9264         struct inode *inode;
9265         int err;
9266         u64 index = 0;
9267
9268         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9269                                 new_dirid, new_dirid,
9270                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9271                                 &index);
9272         if (IS_ERR(inode))
9273                 return PTR_ERR(inode);
9274         inode->i_op = &btrfs_dir_inode_operations;
9275         inode->i_fop = &btrfs_dir_file_operations;
9276
9277         set_nlink(inode, 1);
9278         btrfs_i_size_write(inode, 0);
9279         unlock_new_inode(inode);
9280
9281         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9282         if (err)
9283                 btrfs_err(new_root->fs_info,
9284                           "error inheriting subvolume %llu properties: %d",
9285                           new_root->root_key.objectid, err);
9286
9287         err = btrfs_update_inode(trans, new_root, inode);
9288
9289         iput(inode);
9290         return err;
9291 }
9292
9293 struct inode *btrfs_alloc_inode(struct super_block *sb)
9294 {
9295         struct btrfs_inode *ei;
9296         struct inode *inode;
9297
9298         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9299         if (!ei)
9300                 return NULL;
9301
9302         ei->root = NULL;
9303         ei->generation = 0;
9304         ei->last_trans = 0;
9305         ei->last_sub_trans = 0;
9306         ei->logged_trans = 0;
9307         ei->delalloc_bytes = 0;
9308         ei->defrag_bytes = 0;
9309         ei->disk_i_size = 0;
9310         ei->flags = 0;
9311         ei->csum_bytes = 0;
9312         ei->index_cnt = (u64)-1;
9313         ei->dir_index = 0;
9314         ei->last_unlink_trans = 0;
9315         ei->last_log_commit = 0;
9316         ei->delayed_iput_count = 0;
9317
9318         spin_lock_init(&ei->lock);
9319         ei->outstanding_extents = 0;
9320         ei->reserved_extents = 0;
9321
9322         ei->runtime_flags = 0;
9323         ei->force_compress = BTRFS_COMPRESS_NONE;
9324
9325         ei->delayed_node = NULL;
9326
9327         ei->i_otime.tv_sec = 0;
9328         ei->i_otime.tv_nsec = 0;
9329
9330         inode = &ei->vfs_inode;
9331         extent_map_tree_init(&ei->extent_tree);
9332         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9333         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9334         ei->io_tree.track_uptodate = 1;
9335         ei->io_failure_tree.track_uptodate = 1;
9336         atomic_set(&ei->sync_writers, 0);
9337         mutex_init(&ei->log_mutex);
9338         mutex_init(&ei->delalloc_mutex);
9339         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9340         INIT_LIST_HEAD(&ei->delalloc_inodes);
9341         INIT_LIST_HEAD(&ei->delayed_iput);
9342         RB_CLEAR_NODE(&ei->rb_node);
9343         init_rwsem(&ei->dio_sem);
9344
9345         return inode;
9346 }
9347
9348 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9349 void btrfs_test_destroy_inode(struct inode *inode)
9350 {
9351         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9352         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9353 }
9354 #endif
9355
9356 static void btrfs_i_callback(struct rcu_head *head)
9357 {
9358         struct inode *inode = container_of(head, struct inode, i_rcu);
9359         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9360 }
9361
9362 void btrfs_destroy_inode(struct inode *inode)
9363 {
9364         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9365         struct btrfs_ordered_extent *ordered;
9366         struct btrfs_root *root = BTRFS_I(inode)->root;
9367
9368         WARN_ON(!hlist_empty(&inode->i_dentry));
9369         WARN_ON(inode->i_data.nrpages);
9370         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9371         WARN_ON(BTRFS_I(inode)->reserved_extents);
9372         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9373         WARN_ON(BTRFS_I(inode)->csum_bytes);
9374         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9375
9376         /*
9377          * This can happen where we create an inode, but somebody else also
9378          * created the same inode and we need to destroy the one we already
9379          * created.
9380          */
9381         if (!root)
9382                 goto free;
9383
9384         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9385                      &BTRFS_I(inode)->runtime_flags)) {
9386                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9387                            btrfs_ino(inode));
9388                 atomic_dec(&root->orphan_inodes);
9389         }
9390
9391         while (1) {
9392                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9393                 if (!ordered)
9394                         break;
9395                 else {
9396                         btrfs_err(fs_info,
9397                                   "found ordered extent %llu %llu on inode cleanup",
9398                                   ordered->file_offset, ordered->len);
9399                         btrfs_remove_ordered_extent(inode, ordered);
9400                         btrfs_put_ordered_extent(ordered);
9401                         btrfs_put_ordered_extent(ordered);
9402                 }
9403         }
9404         btrfs_qgroup_check_reserved_leak(inode);
9405         inode_tree_del(inode);
9406         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9407 free:
9408         call_rcu(&inode->i_rcu, btrfs_i_callback);
9409 }
9410
9411 int btrfs_drop_inode(struct inode *inode)
9412 {
9413         struct btrfs_root *root = BTRFS_I(inode)->root;
9414
9415         if (root == NULL)
9416                 return 1;
9417
9418         /* the snap/subvol tree is on deleting */
9419         if (btrfs_root_refs(&root->root_item) == 0)
9420                 return 1;
9421         else
9422                 return generic_drop_inode(inode);
9423 }
9424
9425 static void init_once(void *foo)
9426 {
9427         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9428
9429         inode_init_once(&ei->vfs_inode);
9430 }
9431
9432 void btrfs_destroy_cachep(void)
9433 {
9434         /*
9435          * Make sure all delayed rcu free inodes are flushed before we
9436          * destroy cache.
9437          */
9438         rcu_barrier();
9439         kmem_cache_destroy(btrfs_inode_cachep);
9440         kmem_cache_destroy(btrfs_trans_handle_cachep);
9441         kmem_cache_destroy(btrfs_transaction_cachep);
9442         kmem_cache_destroy(btrfs_path_cachep);
9443         kmem_cache_destroy(btrfs_free_space_cachep);
9444 }
9445
9446 int btrfs_init_cachep(void)
9447 {
9448         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9449                         sizeof(struct btrfs_inode), 0,
9450                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9451                         init_once);
9452         if (!btrfs_inode_cachep)
9453                 goto fail;
9454
9455         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9456                         sizeof(struct btrfs_trans_handle), 0,
9457                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9458         if (!btrfs_trans_handle_cachep)
9459                 goto fail;
9460
9461         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9462                         sizeof(struct btrfs_transaction), 0,
9463                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9464         if (!btrfs_transaction_cachep)
9465                 goto fail;
9466
9467         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9468                         sizeof(struct btrfs_path), 0,
9469                         SLAB_MEM_SPREAD, NULL);
9470         if (!btrfs_path_cachep)
9471                 goto fail;
9472
9473         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9474                         sizeof(struct btrfs_free_space), 0,
9475                         SLAB_MEM_SPREAD, NULL);
9476         if (!btrfs_free_space_cachep)
9477                 goto fail;
9478
9479         return 0;
9480 fail:
9481         btrfs_destroy_cachep();
9482         return -ENOMEM;
9483 }
9484
9485 static int btrfs_getattr(struct vfsmount *mnt,
9486                          struct dentry *dentry, struct kstat *stat)
9487 {
9488         u64 delalloc_bytes;
9489         struct inode *inode = d_inode(dentry);
9490         u32 blocksize = inode->i_sb->s_blocksize;
9491
9492         generic_fillattr(inode, stat);
9493         stat->dev = BTRFS_I(inode)->root->anon_dev;
9494
9495         spin_lock(&BTRFS_I(inode)->lock);
9496         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9497         spin_unlock(&BTRFS_I(inode)->lock);
9498         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9499                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9500         return 0;
9501 }
9502
9503 static int btrfs_rename_exchange(struct inode *old_dir,
9504                               struct dentry *old_dentry,
9505                               struct inode *new_dir,
9506                               struct dentry *new_dentry)
9507 {
9508         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9509         struct btrfs_trans_handle *trans;
9510         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9511         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9512         struct inode *new_inode = new_dentry->d_inode;
9513         struct inode *old_inode = old_dentry->d_inode;
9514         struct timespec ctime = current_time(old_inode);
9515         struct dentry *parent;
9516         u64 old_ino = btrfs_ino(old_inode);
9517         u64 new_ino = btrfs_ino(new_inode);
9518         u64 old_idx = 0;
9519         u64 new_idx = 0;
9520         u64 root_objectid;
9521         int ret;
9522         bool root_log_pinned = false;
9523         bool dest_log_pinned = false;
9524
9525         /* we only allow rename subvolume link between subvolumes */
9526         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9527                 return -EXDEV;
9528
9529         /* close the race window with snapshot create/destroy ioctl */
9530         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9531                 down_read(&fs_info->subvol_sem);
9532         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9533                 down_read(&fs_info->subvol_sem);
9534
9535         /*
9536          * We want to reserve the absolute worst case amount of items.  So if
9537          * both inodes are subvols and we need to unlink them then that would
9538          * require 4 item modifications, but if they are both normal inodes it
9539          * would require 5 item modifications, so we'll assume their normal
9540          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9541          * should cover the worst case number of items we'll modify.
9542          */
9543         trans = btrfs_start_transaction(root, 12);
9544         if (IS_ERR(trans)) {
9545                 ret = PTR_ERR(trans);
9546                 goto out_notrans;
9547         }
9548
9549         /*
9550          * We need to find a free sequence number both in the source and
9551          * in the destination directory for the exchange.
9552          */
9553         ret = btrfs_set_inode_index(new_dir, &old_idx);
9554         if (ret)
9555                 goto out_fail;
9556         ret = btrfs_set_inode_index(old_dir, &new_idx);
9557         if (ret)
9558                 goto out_fail;
9559
9560         BTRFS_I(old_inode)->dir_index = 0ULL;
9561         BTRFS_I(new_inode)->dir_index = 0ULL;
9562
9563         /* Reference for the source. */
9564         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9565                 /* force full log commit if subvolume involved. */
9566                 btrfs_set_log_full_commit(fs_info, trans);
9567         } else {
9568                 btrfs_pin_log_trans(root);
9569                 root_log_pinned = true;
9570                 ret = btrfs_insert_inode_ref(trans, dest,
9571                                              new_dentry->d_name.name,
9572                                              new_dentry->d_name.len,
9573                                              old_ino,
9574                                              btrfs_ino(new_dir), old_idx);
9575                 if (ret)
9576                         goto out_fail;
9577         }
9578
9579         /* And now for the dest. */
9580         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9581                 /* force full log commit if subvolume involved. */
9582                 btrfs_set_log_full_commit(fs_info, trans);
9583         } else {
9584                 btrfs_pin_log_trans(dest);
9585                 dest_log_pinned = true;
9586                 ret = btrfs_insert_inode_ref(trans, root,
9587                                              old_dentry->d_name.name,
9588                                              old_dentry->d_name.len,
9589                                              new_ino,
9590                                              btrfs_ino(old_dir), new_idx);
9591                 if (ret)
9592                         goto out_fail;
9593         }
9594
9595         /* Update inode version and ctime/mtime. */
9596         inode_inc_iversion(old_dir);
9597         inode_inc_iversion(new_dir);
9598         inode_inc_iversion(old_inode);
9599         inode_inc_iversion(new_inode);
9600         old_dir->i_ctime = old_dir->i_mtime = ctime;
9601         new_dir->i_ctime = new_dir->i_mtime = ctime;
9602         old_inode->i_ctime = ctime;
9603         new_inode->i_ctime = ctime;
9604
9605         if (old_dentry->d_parent != new_dentry->d_parent) {
9606                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9607                 btrfs_record_unlink_dir(trans, new_dir, new_inode, 1);
9608         }
9609
9610         /* src is a subvolume */
9611         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9612                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9613                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9614                                           root_objectid,
9615                                           old_dentry->d_name.name,
9616                                           old_dentry->d_name.len);
9617         } else { /* src is an inode */
9618                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9619                                            old_dentry->d_inode,
9620                                            old_dentry->d_name.name,
9621                                            old_dentry->d_name.len);
9622                 if (!ret)
9623                         ret = btrfs_update_inode(trans, root, old_inode);
9624         }
9625         if (ret) {
9626                 btrfs_abort_transaction(trans, ret);
9627                 goto out_fail;
9628         }
9629
9630         /* dest is a subvolume */
9631         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9632                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9633                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9634                                           root_objectid,
9635                                           new_dentry->d_name.name,
9636                                           new_dentry->d_name.len);
9637         } else { /* dest is an inode */
9638                 ret = __btrfs_unlink_inode(trans, dest, new_dir,
9639                                            new_dentry->d_inode,
9640                                            new_dentry->d_name.name,
9641                                            new_dentry->d_name.len);
9642                 if (!ret)
9643                         ret = btrfs_update_inode(trans, dest, new_inode);
9644         }
9645         if (ret) {
9646                 btrfs_abort_transaction(trans, ret);
9647                 goto out_fail;
9648         }
9649
9650         ret = btrfs_add_link(trans, new_dir, old_inode,
9651                              new_dentry->d_name.name,
9652                              new_dentry->d_name.len, 0, old_idx);
9653         if (ret) {
9654                 btrfs_abort_transaction(trans, ret);
9655                 goto out_fail;
9656         }
9657
9658         ret = btrfs_add_link(trans, old_dir, new_inode,
9659                              old_dentry->d_name.name,
9660                              old_dentry->d_name.len, 0, new_idx);
9661         if (ret) {
9662                 btrfs_abort_transaction(trans, ret);
9663                 goto out_fail;
9664         }
9665
9666         if (old_inode->i_nlink == 1)
9667                 BTRFS_I(old_inode)->dir_index = old_idx;
9668         if (new_inode->i_nlink == 1)
9669                 BTRFS_I(new_inode)->dir_index = new_idx;
9670
9671         if (root_log_pinned) {
9672                 parent = new_dentry->d_parent;
9673                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9674                 btrfs_end_log_trans(root);
9675                 root_log_pinned = false;
9676         }
9677         if (dest_log_pinned) {
9678                 parent = old_dentry->d_parent;
9679                 btrfs_log_new_name(trans, new_inode, new_dir, parent);
9680                 btrfs_end_log_trans(dest);
9681                 dest_log_pinned = false;
9682         }
9683 out_fail:
9684         /*
9685          * If we have pinned a log and an error happened, we unpin tasks
9686          * trying to sync the log and force them to fallback to a transaction
9687          * commit if the log currently contains any of the inodes involved in
9688          * this rename operation (to ensure we do not persist a log with an
9689          * inconsistent state for any of these inodes or leading to any
9690          * inconsistencies when replayed). If the transaction was aborted, the
9691          * abortion reason is propagated to userspace when attempting to commit
9692          * the transaction. If the log does not contain any of these inodes, we
9693          * allow the tasks to sync it.
9694          */
9695         if (ret && (root_log_pinned || dest_log_pinned)) {
9696                 if (btrfs_inode_in_log(old_dir, fs_info->generation) ||
9697                     btrfs_inode_in_log(new_dir, fs_info->generation) ||
9698                     btrfs_inode_in_log(old_inode, fs_info->generation) ||
9699                     (new_inode &&
9700                      btrfs_inode_in_log(new_inode, fs_info->generation)))
9701                         btrfs_set_log_full_commit(fs_info, trans);
9702
9703                 if (root_log_pinned) {
9704                         btrfs_end_log_trans(root);
9705                         root_log_pinned = false;
9706                 }
9707                 if (dest_log_pinned) {
9708                         btrfs_end_log_trans(dest);
9709                         dest_log_pinned = false;
9710                 }
9711         }
9712         ret = btrfs_end_transaction(trans);
9713 out_notrans:
9714         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9715                 up_read(&fs_info->subvol_sem);
9716         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9717                 up_read(&fs_info->subvol_sem);
9718
9719         return ret;
9720 }
9721
9722 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9723                                      struct btrfs_root *root,
9724                                      struct inode *dir,
9725                                      struct dentry *dentry)
9726 {
9727         int ret;
9728         struct inode *inode;
9729         u64 objectid;
9730         u64 index;
9731
9732         ret = btrfs_find_free_ino(root, &objectid);
9733         if (ret)
9734                 return ret;
9735
9736         inode = btrfs_new_inode(trans, root, dir,
9737                                 dentry->d_name.name,
9738                                 dentry->d_name.len,
9739                                 btrfs_ino(dir),
9740                                 objectid,
9741                                 S_IFCHR | WHITEOUT_MODE,
9742                                 &index);
9743
9744         if (IS_ERR(inode)) {
9745                 ret = PTR_ERR(inode);
9746                 return ret;
9747         }
9748
9749         inode->i_op = &btrfs_special_inode_operations;
9750         init_special_inode(inode, inode->i_mode,
9751                 WHITEOUT_DEV);
9752
9753         ret = btrfs_init_inode_security(trans, inode, dir,
9754                                 &dentry->d_name);
9755         if (ret)
9756                 goto out;
9757
9758         ret = btrfs_add_nondir(trans, dir, dentry,
9759                                 inode, 0, index);
9760         if (ret)
9761                 goto out;
9762
9763         ret = btrfs_update_inode(trans, root, inode);
9764 out:
9765         unlock_new_inode(inode);
9766         if (ret)
9767                 inode_dec_link_count(inode);
9768         iput(inode);
9769
9770         return ret;
9771 }
9772
9773 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9774                            struct inode *new_dir, struct dentry *new_dentry,
9775                            unsigned int flags)
9776 {
9777         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9778         struct btrfs_trans_handle *trans;
9779         unsigned int trans_num_items;
9780         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9781         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9782         struct inode *new_inode = d_inode(new_dentry);
9783         struct inode *old_inode = d_inode(old_dentry);
9784         u64 index = 0;
9785         u64 root_objectid;
9786         int ret;
9787         u64 old_ino = btrfs_ino(old_inode);
9788         bool log_pinned = false;
9789
9790         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9791                 return -EPERM;
9792
9793         /* we only allow rename subvolume link between subvolumes */
9794         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9795                 return -EXDEV;
9796
9797         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9798             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9799                 return -ENOTEMPTY;
9800
9801         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9802             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9803                 return -ENOTEMPTY;
9804
9805
9806         /* check for collisions, even if the  name isn't there */
9807         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9808                              new_dentry->d_name.name,
9809                              new_dentry->d_name.len);
9810
9811         if (ret) {
9812                 if (ret == -EEXIST) {
9813                         /* we shouldn't get
9814                          * eexist without a new_inode */
9815                         if (WARN_ON(!new_inode)) {
9816                                 return ret;
9817                         }
9818                 } else {
9819                         /* maybe -EOVERFLOW */
9820                         return ret;
9821                 }
9822         }
9823         ret = 0;
9824
9825         /*
9826          * we're using rename to replace one file with another.  Start IO on it
9827          * now so  we don't add too much work to the end of the transaction
9828          */
9829         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9830                 filemap_flush(old_inode->i_mapping);
9831
9832         /* close the racy window with snapshot create/destroy ioctl */
9833         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9834                 down_read(&fs_info->subvol_sem);
9835         /*
9836          * We want to reserve the absolute worst case amount of items.  So if
9837          * both inodes are subvols and we need to unlink them then that would
9838          * require 4 item modifications, but if they are both normal inodes it
9839          * would require 5 item modifications, so we'll assume they are normal
9840          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9841          * should cover the worst case number of items we'll modify.
9842          * If our rename has the whiteout flag, we need more 5 units for the
9843          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9844          * when selinux is enabled).
9845          */
9846         trans_num_items = 11;
9847         if (flags & RENAME_WHITEOUT)
9848                 trans_num_items += 5;
9849         trans = btrfs_start_transaction(root, trans_num_items);
9850         if (IS_ERR(trans)) {
9851                 ret = PTR_ERR(trans);
9852                 goto out_notrans;
9853         }
9854
9855         if (dest != root)
9856                 btrfs_record_root_in_trans(trans, dest);
9857
9858         ret = btrfs_set_inode_index(new_dir, &index);
9859         if (ret)
9860                 goto out_fail;
9861
9862         BTRFS_I(old_inode)->dir_index = 0ULL;
9863         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9864                 /* force full log commit if subvolume involved. */
9865                 btrfs_set_log_full_commit(fs_info, trans);
9866         } else {
9867                 btrfs_pin_log_trans(root);
9868                 log_pinned = true;
9869                 ret = btrfs_insert_inode_ref(trans, dest,
9870                                              new_dentry->d_name.name,
9871                                              new_dentry->d_name.len,
9872                                              old_ino,
9873                                              btrfs_ino(new_dir), index);
9874                 if (ret)
9875                         goto out_fail;
9876         }
9877
9878         inode_inc_iversion(old_dir);
9879         inode_inc_iversion(new_dir);
9880         inode_inc_iversion(old_inode);
9881         old_dir->i_ctime = old_dir->i_mtime =
9882         new_dir->i_ctime = new_dir->i_mtime =
9883         old_inode->i_ctime = current_time(old_dir);
9884
9885         if (old_dentry->d_parent != new_dentry->d_parent)
9886                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9887
9888         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9889                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9890                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9891                                         old_dentry->d_name.name,
9892                                         old_dentry->d_name.len);
9893         } else {
9894                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9895                                         d_inode(old_dentry),
9896                                         old_dentry->d_name.name,
9897                                         old_dentry->d_name.len);
9898                 if (!ret)
9899                         ret = btrfs_update_inode(trans, root, old_inode);
9900         }
9901         if (ret) {
9902                 btrfs_abort_transaction(trans, ret);
9903                 goto out_fail;
9904         }
9905
9906         if (new_inode) {
9907                 inode_inc_iversion(new_inode);
9908                 new_inode->i_ctime = current_time(new_inode);
9909                 if (unlikely(btrfs_ino(new_inode) ==
9910                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9911                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9912                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9913                                                 root_objectid,
9914                                                 new_dentry->d_name.name,
9915                                                 new_dentry->d_name.len);
9916                         BUG_ON(new_inode->i_nlink == 0);
9917                 } else {
9918                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9919                                                  d_inode(new_dentry),
9920                                                  new_dentry->d_name.name,
9921                                                  new_dentry->d_name.len);
9922                 }
9923                 if (!ret && new_inode->i_nlink == 0)
9924                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9925                 if (ret) {
9926                         btrfs_abort_transaction(trans, ret);
9927                         goto out_fail;
9928                 }
9929         }
9930
9931         ret = btrfs_add_link(trans, new_dir, old_inode,
9932                              new_dentry->d_name.name,
9933                              new_dentry->d_name.len, 0, index);
9934         if (ret) {
9935                 btrfs_abort_transaction(trans, ret);
9936                 goto out_fail;
9937         }
9938
9939         if (old_inode->i_nlink == 1)
9940                 BTRFS_I(old_inode)->dir_index = index;
9941
9942         if (log_pinned) {
9943                 struct dentry *parent = new_dentry->d_parent;
9944
9945                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9946                 btrfs_end_log_trans(root);
9947                 log_pinned = false;
9948         }
9949
9950         if (flags & RENAME_WHITEOUT) {
9951                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9952                                                 old_dentry);
9953
9954                 if (ret) {
9955                         btrfs_abort_transaction(trans, ret);
9956                         goto out_fail;
9957                 }
9958         }
9959 out_fail:
9960         /*
9961          * If we have pinned the log and an error happened, we unpin tasks
9962          * trying to sync the log and force them to fallback to a transaction
9963          * commit if the log currently contains any of the inodes involved in
9964          * this rename operation (to ensure we do not persist a log with an
9965          * inconsistent state for any of these inodes or leading to any
9966          * inconsistencies when replayed). If the transaction was aborted, the
9967          * abortion reason is propagated to userspace when attempting to commit
9968          * the transaction. If the log does not contain any of these inodes, we
9969          * allow the tasks to sync it.
9970          */
9971         if (ret && log_pinned) {
9972                 if (btrfs_inode_in_log(old_dir, fs_info->generation) ||
9973                     btrfs_inode_in_log(new_dir, fs_info->generation) ||
9974                     btrfs_inode_in_log(old_inode, fs_info->generation) ||
9975                     (new_inode &&
9976                      btrfs_inode_in_log(new_inode, fs_info->generation)))
9977                         btrfs_set_log_full_commit(fs_info, trans);
9978
9979                 btrfs_end_log_trans(root);
9980                 log_pinned = false;
9981         }
9982         btrfs_end_transaction(trans);
9983 out_notrans:
9984         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9985                 up_read(&fs_info->subvol_sem);
9986
9987         return ret;
9988 }
9989
9990 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9991                          struct inode *new_dir, struct dentry *new_dentry,
9992                          unsigned int flags)
9993 {
9994         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9995                 return -EINVAL;
9996
9997         if (flags & RENAME_EXCHANGE)
9998                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9999                                           new_dentry);
10000
10001         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10002 }
10003
10004 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10005 {
10006         struct btrfs_delalloc_work *delalloc_work;
10007         struct inode *inode;
10008
10009         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10010                                      work);
10011         inode = delalloc_work->inode;
10012         filemap_flush(inode->i_mapping);
10013         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10014                                 &BTRFS_I(inode)->runtime_flags))
10015                 filemap_flush(inode->i_mapping);
10016
10017         if (delalloc_work->delay_iput)
10018                 btrfs_add_delayed_iput(inode);
10019         else
10020                 iput(inode);
10021         complete(&delalloc_work->completion);
10022 }
10023
10024 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10025                                                     int delay_iput)
10026 {
10027         struct btrfs_delalloc_work *work;
10028
10029         work = kmalloc(sizeof(*work), GFP_NOFS);
10030         if (!work)
10031                 return NULL;
10032
10033         init_completion(&work->completion);
10034         INIT_LIST_HEAD(&work->list);
10035         work->inode = inode;
10036         work->delay_iput = delay_iput;
10037         WARN_ON_ONCE(!inode);
10038         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10039                         btrfs_run_delalloc_work, NULL, NULL);
10040
10041         return work;
10042 }
10043
10044 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10045 {
10046         wait_for_completion(&work->completion);
10047         kfree(work);
10048 }
10049
10050 /*
10051  * some fairly slow code that needs optimization. This walks the list
10052  * of all the inodes with pending delalloc and forces them to disk.
10053  */
10054 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10055                                    int nr)
10056 {
10057         struct btrfs_inode *binode;
10058         struct inode *inode;
10059         struct btrfs_delalloc_work *work, *next;
10060         struct list_head works;
10061         struct list_head splice;
10062         int ret = 0;
10063
10064         INIT_LIST_HEAD(&works);
10065         INIT_LIST_HEAD(&splice);
10066
10067         mutex_lock(&root->delalloc_mutex);
10068         spin_lock(&root->delalloc_lock);
10069         list_splice_init(&root->delalloc_inodes, &splice);
10070         while (!list_empty(&splice)) {
10071                 binode = list_entry(splice.next, struct btrfs_inode,
10072                                     delalloc_inodes);
10073
10074                 list_move_tail(&binode->delalloc_inodes,
10075                                &root->delalloc_inodes);
10076                 inode = igrab(&binode->vfs_inode);
10077                 if (!inode) {
10078                         cond_resched_lock(&root->delalloc_lock);
10079                         continue;
10080                 }
10081                 spin_unlock(&root->delalloc_lock);
10082
10083                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10084                 if (!work) {
10085                         if (delay_iput)
10086                                 btrfs_add_delayed_iput(inode);
10087                         else
10088                                 iput(inode);
10089                         ret = -ENOMEM;
10090                         goto out;
10091                 }
10092                 list_add_tail(&work->list, &works);
10093                 btrfs_queue_work(root->fs_info->flush_workers,
10094                                  &work->work);
10095                 ret++;
10096                 if (nr != -1 && ret >= nr)
10097                         goto out;
10098                 cond_resched();
10099                 spin_lock(&root->delalloc_lock);
10100         }
10101         spin_unlock(&root->delalloc_lock);
10102
10103 out:
10104         list_for_each_entry_safe(work, next, &works, list) {
10105                 list_del_init(&work->list);
10106                 btrfs_wait_and_free_delalloc_work(work);
10107         }
10108
10109         if (!list_empty_careful(&splice)) {
10110                 spin_lock(&root->delalloc_lock);
10111                 list_splice_tail(&splice, &root->delalloc_inodes);
10112                 spin_unlock(&root->delalloc_lock);
10113         }
10114         mutex_unlock(&root->delalloc_mutex);
10115         return ret;
10116 }
10117
10118 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10119 {
10120         struct btrfs_fs_info *fs_info = root->fs_info;
10121         int ret;
10122
10123         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10124                 return -EROFS;
10125
10126         ret = __start_delalloc_inodes(root, delay_iput, -1);
10127         if (ret > 0)
10128                 ret = 0;
10129         /*
10130          * the filemap_flush will queue IO into the worker threads, but
10131          * we have to make sure the IO is actually started and that
10132          * ordered extents get created before we return
10133          */
10134         atomic_inc(&fs_info->async_submit_draining);
10135         while (atomic_read(&fs_info->nr_async_submits) ||
10136                atomic_read(&fs_info->async_delalloc_pages)) {
10137                 wait_event(fs_info->async_submit_wait,
10138                            (atomic_read(&fs_info->nr_async_submits) == 0 &&
10139                             atomic_read(&fs_info->async_delalloc_pages) == 0));
10140         }
10141         atomic_dec(&fs_info->async_submit_draining);
10142         return ret;
10143 }
10144
10145 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10146                                int nr)
10147 {
10148         struct btrfs_root *root;
10149         struct list_head splice;
10150         int ret;
10151
10152         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10153                 return -EROFS;
10154
10155         INIT_LIST_HEAD(&splice);
10156
10157         mutex_lock(&fs_info->delalloc_root_mutex);
10158         spin_lock(&fs_info->delalloc_root_lock);
10159         list_splice_init(&fs_info->delalloc_roots, &splice);
10160         while (!list_empty(&splice) && nr) {
10161                 root = list_first_entry(&splice, struct btrfs_root,
10162                                         delalloc_root);
10163                 root = btrfs_grab_fs_root(root);
10164                 BUG_ON(!root);
10165                 list_move_tail(&root->delalloc_root,
10166                                &fs_info->delalloc_roots);
10167                 spin_unlock(&fs_info->delalloc_root_lock);
10168
10169                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10170                 btrfs_put_fs_root(root);
10171                 if (ret < 0)
10172                         goto out;
10173
10174                 if (nr != -1) {
10175                         nr -= ret;
10176                         WARN_ON(nr < 0);
10177                 }
10178                 spin_lock(&fs_info->delalloc_root_lock);
10179         }
10180         spin_unlock(&fs_info->delalloc_root_lock);
10181
10182         ret = 0;
10183         atomic_inc(&fs_info->async_submit_draining);
10184         while (atomic_read(&fs_info->nr_async_submits) ||
10185               atomic_read(&fs_info->async_delalloc_pages)) {
10186                 wait_event(fs_info->async_submit_wait,
10187                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10188                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10189         }
10190         atomic_dec(&fs_info->async_submit_draining);
10191 out:
10192         if (!list_empty_careful(&splice)) {
10193                 spin_lock(&fs_info->delalloc_root_lock);
10194                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10195                 spin_unlock(&fs_info->delalloc_root_lock);
10196         }
10197         mutex_unlock(&fs_info->delalloc_root_mutex);
10198         return ret;
10199 }
10200
10201 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10202                          const char *symname)
10203 {
10204         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10205         struct btrfs_trans_handle *trans;
10206         struct btrfs_root *root = BTRFS_I(dir)->root;
10207         struct btrfs_path *path;
10208         struct btrfs_key key;
10209         struct inode *inode = NULL;
10210         int err;
10211         int drop_inode = 0;
10212         u64 objectid;
10213         u64 index = 0;
10214         int name_len;
10215         int datasize;
10216         unsigned long ptr;
10217         struct btrfs_file_extent_item *ei;
10218         struct extent_buffer *leaf;
10219
10220         name_len = strlen(symname);
10221         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10222                 return -ENAMETOOLONG;
10223
10224         /*
10225          * 2 items for inode item and ref
10226          * 2 items for dir items
10227          * 1 item for updating parent inode item
10228          * 1 item for the inline extent item
10229          * 1 item for xattr if selinux is on
10230          */
10231         trans = btrfs_start_transaction(root, 7);
10232         if (IS_ERR(trans))
10233                 return PTR_ERR(trans);
10234
10235         err = btrfs_find_free_ino(root, &objectid);
10236         if (err)
10237                 goto out_unlock;
10238
10239         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10240                                 dentry->d_name.len, btrfs_ino(dir), objectid,
10241                                 S_IFLNK|S_IRWXUGO, &index);
10242         if (IS_ERR(inode)) {
10243                 err = PTR_ERR(inode);
10244                 goto out_unlock;
10245         }
10246
10247         /*
10248         * If the active LSM wants to access the inode during
10249         * d_instantiate it needs these. Smack checks to see
10250         * if the filesystem supports xattrs by looking at the
10251         * ops vector.
10252         */
10253         inode->i_fop = &btrfs_file_operations;
10254         inode->i_op = &btrfs_file_inode_operations;
10255         inode->i_mapping->a_ops = &btrfs_aops;
10256         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10257
10258         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10259         if (err)
10260                 goto out_unlock_inode;
10261
10262         path = btrfs_alloc_path();
10263         if (!path) {
10264                 err = -ENOMEM;
10265                 goto out_unlock_inode;
10266         }
10267         key.objectid = btrfs_ino(inode);
10268         key.offset = 0;
10269         key.type = BTRFS_EXTENT_DATA_KEY;
10270         datasize = btrfs_file_extent_calc_inline_size(name_len);
10271         err = btrfs_insert_empty_item(trans, root, path, &key,
10272                                       datasize);
10273         if (err) {
10274                 btrfs_free_path(path);
10275                 goto out_unlock_inode;
10276         }
10277         leaf = path->nodes[0];
10278         ei = btrfs_item_ptr(leaf, path->slots[0],
10279                             struct btrfs_file_extent_item);
10280         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10281         btrfs_set_file_extent_type(leaf, ei,
10282                                    BTRFS_FILE_EXTENT_INLINE);
10283         btrfs_set_file_extent_encryption(leaf, ei, 0);
10284         btrfs_set_file_extent_compression(leaf, ei, 0);
10285         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10286         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10287
10288         ptr = btrfs_file_extent_inline_start(ei);
10289         write_extent_buffer(leaf, symname, ptr, name_len);
10290         btrfs_mark_buffer_dirty(leaf);
10291         btrfs_free_path(path);
10292
10293         inode->i_op = &btrfs_symlink_inode_operations;
10294         inode_nohighmem(inode);
10295         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10296         inode_set_bytes(inode, name_len);
10297         btrfs_i_size_write(inode, name_len);
10298         err = btrfs_update_inode(trans, root, inode);
10299         /*
10300          * Last step, add directory indexes for our symlink inode. This is the
10301          * last step to avoid extra cleanup of these indexes if an error happens
10302          * elsewhere above.
10303          */
10304         if (!err)
10305                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
10306         if (err) {
10307                 drop_inode = 1;
10308                 goto out_unlock_inode;
10309         }
10310
10311         unlock_new_inode(inode);
10312         d_instantiate(dentry, inode);
10313
10314 out_unlock:
10315         btrfs_end_transaction(trans);
10316         if (drop_inode) {
10317                 inode_dec_link_count(inode);
10318                 iput(inode);
10319         }
10320         btrfs_btree_balance_dirty(fs_info);
10321         return err;
10322
10323 out_unlock_inode:
10324         drop_inode = 1;
10325         unlock_new_inode(inode);
10326         goto out_unlock;
10327 }
10328
10329 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10330                                        u64 start, u64 num_bytes, u64 min_size,
10331                                        loff_t actual_len, u64 *alloc_hint,
10332                                        struct btrfs_trans_handle *trans)
10333 {
10334         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10335         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10336         struct extent_map *em;
10337         struct btrfs_root *root = BTRFS_I(inode)->root;
10338         struct btrfs_key ins;
10339         u64 cur_offset = start;
10340         u64 i_size;
10341         u64 cur_bytes;
10342         u64 last_alloc = (u64)-1;
10343         int ret = 0;
10344         bool own_trans = true;
10345         u64 end = start + num_bytes - 1;
10346
10347         if (trans)
10348                 own_trans = false;
10349         while (num_bytes > 0) {
10350                 if (own_trans) {
10351                         trans = btrfs_start_transaction(root, 3);
10352                         if (IS_ERR(trans)) {
10353                                 ret = PTR_ERR(trans);
10354                                 break;
10355                         }
10356                 }
10357
10358                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10359                 cur_bytes = max(cur_bytes, min_size);
10360                 /*
10361                  * If we are severely fragmented we could end up with really
10362                  * small allocations, so if the allocator is returning small
10363                  * chunks lets make its job easier by only searching for those
10364                  * sized chunks.
10365                  */
10366                 cur_bytes = min(cur_bytes, last_alloc);
10367                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10368                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10369                 if (ret) {
10370                         if (own_trans)
10371                                 btrfs_end_transaction(trans);
10372                         break;
10373                 }
10374                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10375
10376                 last_alloc = ins.offset;
10377                 ret = insert_reserved_file_extent(trans, inode,
10378                                                   cur_offset, ins.objectid,
10379                                                   ins.offset, ins.offset,
10380                                                   ins.offset, 0, 0, 0,
10381                                                   BTRFS_FILE_EXTENT_PREALLOC);
10382                 if (ret) {
10383                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10384                                                    ins.offset, 0);
10385                         btrfs_abort_transaction(trans, ret);
10386                         if (own_trans)
10387                                 btrfs_end_transaction(trans);
10388                         break;
10389                 }
10390
10391                 btrfs_drop_extent_cache(inode, cur_offset,
10392                                         cur_offset + ins.offset -1, 0);
10393
10394                 em = alloc_extent_map();
10395                 if (!em) {
10396                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10397                                 &BTRFS_I(inode)->runtime_flags);
10398                         goto next;
10399                 }
10400
10401                 em->start = cur_offset;
10402                 em->orig_start = cur_offset;
10403                 em->len = ins.offset;
10404                 em->block_start = ins.objectid;
10405                 em->block_len = ins.offset;
10406                 em->orig_block_len = ins.offset;
10407                 em->ram_bytes = ins.offset;
10408                 em->bdev = fs_info->fs_devices->latest_bdev;
10409                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10410                 em->generation = trans->transid;
10411
10412                 while (1) {
10413                         write_lock(&em_tree->lock);
10414                         ret = add_extent_mapping(em_tree, em, 1);
10415                         write_unlock(&em_tree->lock);
10416                         if (ret != -EEXIST)
10417                                 break;
10418                         btrfs_drop_extent_cache(inode, cur_offset,
10419                                                 cur_offset + ins.offset - 1,
10420                                                 0);
10421                 }
10422                 free_extent_map(em);
10423 next:
10424                 num_bytes -= ins.offset;
10425                 cur_offset += ins.offset;
10426                 *alloc_hint = ins.objectid + ins.offset;
10427
10428                 inode_inc_iversion(inode);
10429                 inode->i_ctime = current_time(inode);
10430                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10431                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10432                     (actual_len > inode->i_size) &&
10433                     (cur_offset > inode->i_size)) {
10434                         if (cur_offset > actual_len)
10435                                 i_size = actual_len;
10436                         else
10437                                 i_size = cur_offset;
10438                         i_size_write(inode, i_size);
10439                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10440                 }
10441
10442                 ret = btrfs_update_inode(trans, root, inode);
10443
10444                 if (ret) {
10445                         btrfs_abort_transaction(trans, ret);
10446                         if (own_trans)
10447                                 btrfs_end_transaction(trans);
10448                         break;
10449                 }
10450
10451                 if (own_trans)
10452                         btrfs_end_transaction(trans);
10453         }
10454         if (cur_offset < end)
10455                 btrfs_free_reserved_data_space(inode, cur_offset,
10456                         end - cur_offset + 1);
10457         return ret;
10458 }
10459
10460 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10461                               u64 start, u64 num_bytes, u64 min_size,
10462                               loff_t actual_len, u64 *alloc_hint)
10463 {
10464         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10465                                            min_size, actual_len, alloc_hint,
10466                                            NULL);
10467 }
10468
10469 int btrfs_prealloc_file_range_trans(struct inode *inode,
10470                                     struct btrfs_trans_handle *trans, int mode,
10471                                     u64 start, u64 num_bytes, u64 min_size,
10472                                     loff_t actual_len, u64 *alloc_hint)
10473 {
10474         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10475                                            min_size, actual_len, alloc_hint, trans);
10476 }
10477
10478 static int btrfs_set_page_dirty(struct page *page)
10479 {
10480         return __set_page_dirty_nobuffers(page);
10481 }
10482
10483 static int btrfs_permission(struct inode *inode, int mask)
10484 {
10485         struct btrfs_root *root = BTRFS_I(inode)->root;
10486         umode_t mode = inode->i_mode;
10487
10488         if (mask & MAY_WRITE &&
10489             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10490                 if (btrfs_root_readonly(root))
10491                         return -EROFS;
10492                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10493                         return -EACCES;
10494         }
10495         return generic_permission(inode, mask);
10496 }
10497
10498 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10499 {
10500         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10501         struct btrfs_trans_handle *trans;
10502         struct btrfs_root *root = BTRFS_I(dir)->root;
10503         struct inode *inode = NULL;
10504         u64 objectid;
10505         u64 index;
10506         int ret = 0;
10507
10508         /*
10509          * 5 units required for adding orphan entry
10510          */
10511         trans = btrfs_start_transaction(root, 5);
10512         if (IS_ERR(trans))
10513                 return PTR_ERR(trans);
10514
10515         ret = btrfs_find_free_ino(root, &objectid);
10516         if (ret)
10517                 goto out;
10518
10519         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10520                                 btrfs_ino(dir), objectid, mode, &index);
10521         if (IS_ERR(inode)) {
10522                 ret = PTR_ERR(inode);
10523                 inode = NULL;
10524                 goto out;
10525         }
10526
10527         inode->i_fop = &btrfs_file_operations;
10528         inode->i_op = &btrfs_file_inode_operations;
10529
10530         inode->i_mapping->a_ops = &btrfs_aops;
10531         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10532
10533         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10534         if (ret)
10535                 goto out_inode;
10536
10537         ret = btrfs_update_inode(trans, root, inode);
10538         if (ret)
10539                 goto out_inode;
10540         ret = btrfs_orphan_add(trans, inode);
10541         if (ret)
10542                 goto out_inode;
10543
10544         /*
10545          * We set number of links to 0 in btrfs_new_inode(), and here we set
10546          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10547          * through:
10548          *
10549          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10550          */
10551         set_nlink(inode, 1);
10552         unlock_new_inode(inode);
10553         d_tmpfile(dentry, inode);
10554         mark_inode_dirty(inode);
10555
10556 out:
10557         btrfs_end_transaction(trans);
10558         if (ret)
10559                 iput(inode);
10560         btrfs_balance_delayed_items(fs_info);
10561         btrfs_btree_balance_dirty(fs_info);
10562         return ret;
10563
10564 out_inode:
10565         unlock_new_inode(inode);
10566         goto out;
10567
10568 }
10569
10570 static const struct inode_operations btrfs_dir_inode_operations = {
10571         .getattr        = btrfs_getattr,
10572         .lookup         = btrfs_lookup,
10573         .create         = btrfs_create,
10574         .unlink         = btrfs_unlink,
10575         .link           = btrfs_link,
10576         .mkdir          = btrfs_mkdir,
10577         .rmdir          = btrfs_rmdir,
10578         .rename         = btrfs_rename2,
10579         .symlink        = btrfs_symlink,
10580         .setattr        = btrfs_setattr,
10581         .mknod          = btrfs_mknod,
10582         .listxattr      = btrfs_listxattr,
10583         .permission     = btrfs_permission,
10584         .get_acl        = btrfs_get_acl,
10585         .set_acl        = btrfs_set_acl,
10586         .update_time    = btrfs_update_time,
10587         .tmpfile        = btrfs_tmpfile,
10588 };
10589 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10590         .lookup         = btrfs_lookup,
10591         .permission     = btrfs_permission,
10592         .update_time    = btrfs_update_time,
10593 };
10594
10595 static const struct file_operations btrfs_dir_file_operations = {
10596         .llseek         = generic_file_llseek,
10597         .read           = generic_read_dir,
10598         .iterate_shared = btrfs_real_readdir,
10599         .unlocked_ioctl = btrfs_ioctl,
10600 #ifdef CONFIG_COMPAT
10601         .compat_ioctl   = btrfs_compat_ioctl,
10602 #endif
10603         .release        = btrfs_release_file,
10604         .fsync          = btrfs_sync_file,
10605 };
10606
10607 static const struct extent_io_ops btrfs_extent_io_ops = {
10608         .fill_delalloc = run_delalloc_range,
10609         .submit_bio_hook = btrfs_submit_bio_hook,
10610         .merge_bio_hook = btrfs_merge_bio_hook,
10611         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10612         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10613         .writepage_start_hook = btrfs_writepage_start_hook,
10614         .set_bit_hook = btrfs_set_bit_hook,
10615         .clear_bit_hook = btrfs_clear_bit_hook,
10616         .merge_extent_hook = btrfs_merge_extent_hook,
10617         .split_extent_hook = btrfs_split_extent_hook,
10618 };
10619
10620 /*
10621  * btrfs doesn't support the bmap operation because swapfiles
10622  * use bmap to make a mapping of extents in the file.  They assume
10623  * these extents won't change over the life of the file and they
10624  * use the bmap result to do IO directly to the drive.
10625  *
10626  * the btrfs bmap call would return logical addresses that aren't
10627  * suitable for IO and they also will change frequently as COW
10628  * operations happen.  So, swapfile + btrfs == corruption.
10629  *
10630  * For now we're avoiding this by dropping bmap.
10631  */
10632 static const struct address_space_operations btrfs_aops = {
10633         .readpage       = btrfs_readpage,
10634         .writepage      = btrfs_writepage,
10635         .writepages     = btrfs_writepages,
10636         .readpages      = btrfs_readpages,
10637         .direct_IO      = btrfs_direct_IO,
10638         .invalidatepage = btrfs_invalidatepage,
10639         .releasepage    = btrfs_releasepage,
10640         .set_page_dirty = btrfs_set_page_dirty,
10641         .error_remove_page = generic_error_remove_page,
10642 };
10643
10644 static const struct address_space_operations btrfs_symlink_aops = {
10645         .readpage       = btrfs_readpage,
10646         .writepage      = btrfs_writepage,
10647         .invalidatepage = btrfs_invalidatepage,
10648         .releasepage    = btrfs_releasepage,
10649 };
10650
10651 static const struct inode_operations btrfs_file_inode_operations = {
10652         .getattr        = btrfs_getattr,
10653         .setattr        = btrfs_setattr,
10654         .listxattr      = btrfs_listxattr,
10655         .permission     = btrfs_permission,
10656         .fiemap         = btrfs_fiemap,
10657         .get_acl        = btrfs_get_acl,
10658         .set_acl        = btrfs_set_acl,
10659         .update_time    = btrfs_update_time,
10660 };
10661 static const struct inode_operations btrfs_special_inode_operations = {
10662         .getattr        = btrfs_getattr,
10663         .setattr        = btrfs_setattr,
10664         .permission     = btrfs_permission,
10665         .listxattr      = btrfs_listxattr,
10666         .get_acl        = btrfs_get_acl,
10667         .set_acl        = btrfs_set_acl,
10668         .update_time    = btrfs_update_time,
10669 };
10670 static const struct inode_operations btrfs_symlink_inode_operations = {
10671         .get_link       = page_get_link,
10672         .getattr        = btrfs_getattr,
10673         .setattr        = btrfs_setattr,
10674         .permission     = btrfs_permission,
10675         .listxattr      = btrfs_listxattr,
10676         .update_time    = btrfs_update_time,
10677 };
10678
10679 const struct dentry_operations btrfs_dentry_operations = {
10680         .d_delete       = btrfs_dentry_delete,
10681         .d_release      = btrfs_dentry_release,
10682 };