fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 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, int *page_started,
 109                                    unsigned long *nr_written, int unlock);
 110 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 111                                            u64 len, u64 orig_start,
 112                                            u64 block_start, u64 block_len,
 113                                            u64 orig_block_len, u64 ram_bytes,
 114                                            int type);
 115
 116 static int btrfs_dirty_inode(struct inode *inode);
 117
 118 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 119 void btrfs_test_inode_set_ops(struct inode *inode)
 120 {
 121         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 122 }
 123 #endif
 124
 125 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 126                                      struct inode *inode,  struct inode *dir,
 127                                      const struct qstr *qstr)
 128 {
 129         int err;
 130
 131         err = btrfs_init_acl(trans, inode, dir);
 132         if (!err)
 133                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 134         return err;
 135 }
 136
 137 /*
 138  * this does all the hard work for inserting an inline extent into
 139  * the btree.  The caller should have done a btrfs_drop_extents so that
 140  * no overlapping inline items exist in the btree
 141  */
 142 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 143                                 struct btrfs_path *path, int extent_inserted,
 144                                 struct btrfs_root *root, struct inode *inode,
 145                                 u64 start, size_t size, size_t compressed_size,
 146                                 int compress_type,
 147                                 struct page **compressed_pages)
 148 {
 149         struct extent_buffer *leaf;
 150         struct page *page = NULL;
 151         char *kaddr;
 152         unsigned long ptr;
 153         struct btrfs_file_extent_item *ei;
 154         int err = 0;
 155         int ret;
 156         size_t cur_size = size;
 157         unsigned long offset;
 158
 159         if (compressed_size && compressed_pages)
 160                 cur_size = compressed_size;
 161
 162         inode_add_bytes(inode, size);
 163
 164         if (!extent_inserted) {
 165                 struct btrfs_key key;
 166                 size_t datasize;
 167
 168                 key.objectid = btrfs_ino(inode);
 169                 key.offset = start;
 170                 key.type = BTRFS_EXTENT_DATA_KEY;
 171
 172                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 173                 path->leave_spinning = 1;
 174                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                               datasize);
 176                 if (ret) {
 177                         err = ret;
 178                         goto fail;
 179                 }
 180         }
 181         leaf = path->nodes[0];
 182         ei = btrfs_item_ptr(leaf, path->slots[0],
 183                             struct btrfs_file_extent_item);
 184         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 185         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 186         btrfs_set_file_extent_encryption(leaf, ei, 0);
 187         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 188         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 189         ptr = btrfs_file_extent_inline_start(ei);
 190
 191         if (compress_type != BTRFS_COMPRESS_NONE) {
 192                 struct page *cpage;
 193                 int i = 0;
 194                 while (compressed_size > 0) {
 195                         cpage = compressed_pages[i];
 196                         cur_size = min_t(unsigned long, compressed_size,
 197                                        PAGE_SIZE);
 198
 199                         kaddr = kmap_atomic(cpage);
 200                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 201                         kunmap_atomic(kaddr);
 202
 203                         i++;
 204                         ptr += cur_size;
 205                         compressed_size -= cur_size;
 206                 }
 207                 btrfs_set_file_extent_compression(leaf, ei,
 208                                                   compress_type);
 209         } else {
 210                 page = find_get_page(inode->i_mapping,
 211                                      start >> PAGE_SHIFT);
 212                 btrfs_set_file_extent_compression(leaf, ei, 0);
 213                 kaddr = kmap_atomic(page);
 214                 offset = start & (PAGE_SIZE - 1);
 215                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 216                 kunmap_atomic(kaddr);
 217                 put_page(page);
 218         }
 219         btrfs_mark_buffer_dirty(leaf);
 220         btrfs_release_path(path);
 221
 222         /*
 223          * we're an inline extent, so nobody can
 224          * extend the file past i_size without locking
 225          * a page we already have locked.
 226          *
 227          * We must do any isize and inode updates
 228          * before we unlock the pages.  Otherwise we
 229          * could end up racing with unlink.
 230          */
 231         BTRFS_I(inode)->disk_i_size = inode->i_size;
 232         ret = btrfs_update_inode(trans, root, inode);
 233
 234         return ret;
 235 fail:
 236         return err;
 237 }
 238
 239
 240 /*
 241  * conditionally insert an inline extent into the file.  This
 242  * does the checks required to make sure the data is small enough
 243  * to fit as an inline extent.
 244  */
 245 static noinline int cow_file_range_inline(struct btrfs_root *root,
 246                                           struct inode *inode, u64 start,
 247                                           u64 end, size_t compressed_size,
 248                                           int compress_type,
 249                                           struct page **compressed_pages)
 250 {
 251         struct btrfs_trans_handle *trans;
 252         u64 isize = i_size_read(inode);
 253         u64 actual_end = min(end + 1, isize);
 254         u64 inline_len = actual_end - start;
 255         u64 aligned_end = ALIGN(end, root->sectorsize);
 256         u64 data_len = inline_len;
 257         int ret;
 258         struct btrfs_path *path;
 259         int extent_inserted = 0;
 260         u32 extent_item_size;
 261
 262         if (compressed_size)
 263                 data_len = compressed_size;
 264
 265         if (start > 0 ||
 266             actual_end > root->sectorsize ||
 267             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 268             (!compressed_size &&
 269             (actual_end & (root->sectorsize - 1)) == 0) ||
 270             end + 1 < isize ||
 271             data_len > root->fs_info->max_inline) {
 272                 return 1;
 273         }
 274
 275         path = btrfs_alloc_path();
 276         if (!path)
 277                 return -ENOMEM;
 278
 279         trans = btrfs_join_transaction(root);
 280         if (IS_ERR(trans)) {
 281                 btrfs_free_path(path);
 282                 return PTR_ERR(trans);
 283         }
 284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 285
 286         if (compressed_size && compressed_pages)
 287                 extent_item_size = btrfs_file_extent_calc_inline_size(
 288                    compressed_size);
 289         else
 290                 extent_item_size = btrfs_file_extent_calc_inline_size(
 291                     inline_len);
 292
 293         ret = __btrfs_drop_extents(trans, root, inode, path,
 294                                    start, aligned_end, NULL,
 295                                    1, 1, extent_item_size, &extent_inserted);
 296         if (ret) {
 297                 btrfs_abort_transaction(trans, root, ret);
 298                 goto out;
 299         }
 300
 301         if (isize > actual_end)
 302                 inline_len = min_t(u64, isize, actual_end);
 303         ret = insert_inline_extent(trans, path, extent_inserted,
 304                                    root, inode, start,
 305                                    inline_len, compressed_size,
 306                                    compress_type, compressed_pages);
 307         if (ret && ret != -ENOSPC) {
 308                 btrfs_abort_transaction(trans, root, ret);
 309                 goto out;
 310         } else if (ret == -ENOSPC) {
 311                 ret = 1;
 312                 goto out;
 313         }
 314
 315         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 316         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 317         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 318 out:
 319         /*
 320          * Don't forget to free the reserved space, as for inlined extent
 321          * it won't count as data extent, free them directly here.
 322          * And at reserve time, it's always aligned to page size, so
 323          * just free one page here.
 324          */
 325         btrfs_qgroup_free_data(inode, 0, PAGE_SIZE);
 326         btrfs_free_path(path);
 327         btrfs_end_transaction(trans, root);
 328         return ret;
 329 }
 330
 331 struct async_extent {
 332         u64 start;
 333         u64 ram_size;
 334         u64 compressed_size;
 335         struct page **pages;
 336         unsigned long nr_pages;
 337         int compress_type;
 338         struct list_head list;
 339 };
 340
 341 struct async_cow {
 342         struct inode *inode;
 343         struct btrfs_root *root;
 344         struct page *locked_page;
 345         u64 start;
 346         u64 end;
 347         struct list_head extents;
 348         struct btrfs_work work;
 349 };
 350
 351 static noinline int add_async_extent(struct async_cow *cow,
 352                                      u64 start, u64 ram_size,
 353                                      u64 compressed_size,
 354                                      struct page **pages,
 355                                      unsigned long nr_pages,
 356                                      int compress_type)
 357 {
 358         struct async_extent *async_extent;
 359
 360         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 361         BUG_ON(!async_extent); /* -ENOMEM */
 362         async_extent->start = start;
 363         async_extent->ram_size = ram_size;
 364         async_extent->compressed_size = compressed_size;
 365         async_extent->pages = pages;
 366         async_extent->nr_pages = nr_pages;
 367         async_extent->compress_type = compress_type;
 368         list_add_tail(&async_extent->list, &cow->extents);
 369         return 0;
 370 }
 371
 372 static inline int inode_need_compress(struct inode *inode)
 373 {
 374         struct btrfs_root *root = BTRFS_I(inode)->root;
 375
 376         /* force compress */
 377         if (btrfs_test_opt(root, FORCE_COMPRESS))
 378                 return 1;
 379         /* bad compression ratios */
 380         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 381                 return 0;
 382         if (btrfs_test_opt(root, COMPRESS) ||
 383             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 384             BTRFS_I(inode)->force_compress)
 385                 return 1;
 386         return 0;
 387 }
 388
 389 /*
 390  * we create compressed extents in two phases.  The first
 391  * phase compresses a range of pages that have already been
 392  * locked (both pages and state bits are locked).
 393  *
 394  * This is done inside an ordered work queue, and the compression
 395  * is spread across many cpus.  The actual IO submission is step
 396  * two, and the ordered work queue takes care of making sure that
 397  * happens in the same order things were put onto the queue by
 398  * writepages and friends.
 399  *
 400  * If this code finds it can't get good compression, it puts an
 401  * entry onto the work queue to write the uncompressed bytes.  This
 402  * makes sure that both compressed inodes and uncompressed inodes
 403  * are written in the same order that the flusher thread sent them
 404  * down.
 405  */
 406 static noinline void compress_file_range(struct inode *inode,
 407                                         struct page *locked_page,
 408                                         u64 start, u64 end,
 409                                         struct async_cow *async_cow,
 410                                         int *num_added)
 411 {
 412         struct btrfs_root *root = BTRFS_I(inode)->root;
 413         u64 num_bytes;
 414         u64 blocksize = root->sectorsize;
 415         u64 actual_end;
 416         u64 isize = i_size_read(inode);
 417         int ret = 0;
 418         struct page **pages = NULL;
 419         unsigned long nr_pages;
 420         unsigned long nr_pages_ret = 0;
 421         unsigned long total_compressed = 0;
 422         unsigned long total_in = 0;
 423         unsigned long max_compressed = SZ_128K;
 424         unsigned long max_uncompressed = SZ_128K;
 425         int i;
 426         int will_compress;
 427         int compress_type = root->fs_info->compress_type;
 428         int redirty = 0;
 429
 430         /* if this is a small write inside eof, kick off a defrag */
 431         if ((end - start + 1) < SZ_16K &&
 432             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434
 435         actual_end = min_t(u64, isize, end + 1);
 436 again:
 437         will_compress = 0;
 438         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 439         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_SIZE);
 440
 441         /*
 442          * we don't want to send crud past the end of i_size through
 443          * compression, that's just a waste of CPU time.  So, if the
 444          * end of the file is before the start of our current
 445          * requested range of bytes, we bail out to the uncompressed
 446          * cleanup code that can deal with all of this.
 447          *
 448          * It isn't really the fastest way to fix things, but this is a
 449          * very uncommon corner.
 450          */
 451         if (actual_end <= start)
 452                 goto cleanup_and_bail_uncompressed;
 453
 454         total_compressed = actual_end - start;
 455
 456         /*
 457          * skip compression for a small file range(<=blocksize) that
 458          * isn't an inline extent, since it doesn't save disk space at all.
 459          */
 460         if (total_compressed <= blocksize &&
 461            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 462                 goto cleanup_and_bail_uncompressed;
 463
 464         /* we want to make sure that amount of ram required to uncompress
 465          * an extent is reasonable, so we limit the total size in ram
 466          * of a compressed extent to 128k.  This is a crucial number
 467          * because it also controls how easily we can spread reads across
 468          * cpus for decompression.
 469          *
 470          * We also want to make sure the amount of IO required to do
 471          * a random read is reasonably small, so we limit the size of
 472          * a compressed extent to 128k.
 473          */
 474         total_compressed = min(total_compressed, max_uncompressed);
 475         num_bytes = ALIGN(end - start + 1, blocksize);
 476         num_bytes = max(blocksize,  num_bytes);
 477         total_in = 0;
 478         ret = 0;
 479
 480         /*
 481          * we do compression for mount -o compress and when the
 482          * inode has not been flagged as nocompress.  This flag can
 483          * change at any time if we discover bad compression ratios.
 484          */
 485         if (inode_need_compress(inode)) {
 486                 WARN_ON(pages);
 487                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 488                 if (!pages) {
 489                         /* just bail out to the uncompressed code */
 490                         goto cont;
 491                 }
 492
 493                 if (BTRFS_I(inode)->force_compress)
 494                         compress_type = BTRFS_I(inode)->force_compress;
 495
 496                 /*
 497                  * we need to call clear_page_dirty_for_io on each
 498                  * page in the range.  Otherwise applications with the file
 499                  * mmap'd can wander in and change the page contents while
 500                  * we are compressing them.
 501                  *
 502                  * If the compression fails for any reason, we set the pages
 503                  * dirty again later on.
 504                  */
 505                 extent_range_clear_dirty_for_io(inode, start, end);
 506                 redirty = 1;
 507                 ret = btrfs_compress_pages(compress_type,
 508                                            inode->i_mapping, start,
 509                                            total_compressed, pages,
 510                                            nr_pages, &nr_pages_ret,
 511                                            &total_in,
 512                                            &total_compressed,
 513                                            max_compressed);
 514
 515                 if (!ret) {
 516                         unsigned long offset = total_compressed &
 517                                 (PAGE_SIZE - 1);
 518                         struct page *page = pages[nr_pages_ret - 1];
 519                         char *kaddr;
 520
 521                         /* zero the tail end of the last page, we might be
 522                          * sending it down to disk
 523                          */
 524                         if (offset) {
 525                                 kaddr = kmap_atomic(page);
 526                                 memset(kaddr + offset, 0,
 527                                        PAGE_SIZE - offset);
 528                                 kunmap_atomic(kaddr);
 529                         }
 530                         will_compress = 1;
 531                 }
 532         }
 533 cont:
 534         if (start == 0) {
 535                 /* lets try to make an inline extent */
 536                 if (ret || total_in < (actual_end - start)) {
 537                         /* we didn't compress the entire range, try
 538                          * to make an uncompressed inline extent.
 539                          */
 540                         ret = cow_file_range_inline(root, inode, start, end,
 541                                                     0, 0, NULL);
 542                 } else {
 543                         /* try making a compressed inline extent */
 544                         ret = cow_file_range_inline(root, inode, start, end,
 545                                                     total_compressed,
 546                                                     compress_type, pages);
 547                 }
 548                 if (ret <= 0) {
 549                         unsigned long clear_flags = EXTENT_DELALLOC |
 550                                 EXTENT_DEFRAG;
 551                         unsigned long page_error_op;
 552
 553                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 554                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 555
 556                         /*
 557                          * inline extent creation worked or returned error,
 558                          * we don't need to create any more async work items.
 559                          * Unlock and free up our temp pages.
 560                          */
 561                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 562                                                      clear_flags, PAGE_UNLOCK |
 563                                                      PAGE_CLEAR_DIRTY |
 564                                                      PAGE_SET_WRITEBACK |
 565                                                      page_error_op |
 566                                                      PAGE_END_WRITEBACK);
 567                         goto free_pages_out;
 568                 }
 569         }
 570
 571         if (will_compress) {
 572                 /*
 573                  * we aren't doing an inline extent round the compressed size
 574                  * up to a block size boundary so the allocator does sane
 575                  * things
 576                  */
 577                 total_compressed = ALIGN(total_compressed, blocksize);
 578
 579                 /*
 580                  * one last check to make sure the compression is really a
 581                  * win, compare the page count read with the blocks on disk
 582                  */
 583                 total_in = ALIGN(total_in, PAGE_SIZE);
 584                 if (total_compressed >= total_in) {
 585                         will_compress = 0;
 586                 } else {
 587                         num_bytes = total_in;
 588                 }
 589         }
 590         if (!will_compress && pages) {
 591                 /*
 592                  * the compression code ran but failed to make things smaller,
 593                  * free any pages it allocated and our page pointer array
 594                  */
 595                 for (i = 0; i < nr_pages_ret; i++) {
 596                         WARN_ON(pages[i]->mapping);
 597                         put_page(pages[i]);
 598                 }
 599                 kfree(pages);
 600                 pages = NULL;
 601                 total_compressed = 0;
 602                 nr_pages_ret = 0;
 603
 604                 /* flag the file so we don't compress in the future */
 605                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 606                     !(BTRFS_I(inode)->force_compress)) {
 607                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 608                 }
 609         }
 610         if (will_compress) {
 611                 *num_added += 1;
 612
 613                 /* the async work queues will take care of doing actual
 614                  * allocation on disk for these compressed pages,
 615                  * and will submit them to the elevator.
 616                  */
 617                 add_async_extent(async_cow, start, num_bytes,
 618                                  total_compressed, pages, nr_pages_ret,
 619                                  compress_type);
 620
 621                 if (start + num_bytes < end) {
 622                         start += num_bytes;
 623                         pages = NULL;
 624                         cond_resched();
 625                         goto again;
 626                 }
 627         } else {
 628 cleanup_and_bail_uncompressed:
 629                 /*
 630                  * No compression, but we still need to write the pages in
 631                  * the file we've been given so far.  redirty the locked
 632                  * page if it corresponds to our extent and set things up
 633                  * for the async work queue to run cow_file_range to do
 634                  * the normal delalloc dance
 635                  */
 636                 if (page_offset(locked_page) >= start &&
 637                     page_offset(locked_page) <= end) {
 638                         __set_page_dirty_nobuffers(locked_page);
 639                         /* unlocked later on in the async handlers */
 640                 }
 641                 if (redirty)
 642                         extent_range_redirty_for_io(inode, start, end);
 643                 add_async_extent(async_cow, start, end - start + 1,
 644                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 645                 *num_added += 1;
 646         }
 647
 648         return;
 649
 650 free_pages_out:
 651         for (i = 0; i < nr_pages_ret; i++) {
 652                 WARN_ON(pages[i]->mapping);
 653                 put_page(pages[i]);
 654         }
 655         kfree(pages);
 656 }
 657
 658 static void free_async_extent_pages(struct async_extent *async_extent)
 659 {
 660         int i;
 661
 662         if (!async_extent->pages)
 663                 return;
 664
 665         for (i = 0; i < async_extent->nr_pages; i++) {
 666                 WARN_ON(async_extent->pages[i]->mapping);
 667                 put_page(async_extent->pages[i]);
 668         }
 669         kfree(async_extent->pages);
 670         async_extent->nr_pages = 0;
 671         async_extent->pages = NULL;
 672 }
 673
 674 /*
 675  * phase two of compressed writeback.  This is the ordered portion
 676  * of the code, which only gets called in the order the work was
 677  * queued.  We walk all the async extents created by compress_file_range
 678  * and send them down to the disk.
 679  */
 680 static noinline void submit_compressed_extents(struct inode *inode,
 681                                               struct async_cow *async_cow)
 682 {
 683         struct async_extent *async_extent;
 684         u64 alloc_hint = 0;
 685         struct btrfs_key ins;
 686         struct extent_map *em;
 687         struct btrfs_root *root = BTRFS_I(inode)->root;
 688         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 689         struct extent_io_tree *io_tree;
 690         int ret = 0;
 691
 692 again:
 693         while (!list_empty(&async_cow->extents)) {
 694                 async_extent = list_entry(async_cow->extents.next,
 695                                           struct async_extent, list);
 696                 list_del(&async_extent->list);
 697
 698                 io_tree = &BTRFS_I(inode)->io_tree;
 699
 700 retry:
 701                 /* did the compression code fall back to uncompressed IO? */
 702                 if (!async_extent->pages) {
 703                         int page_started = 0;
 704                         unsigned long nr_written = 0;
 705
 706                         lock_extent(io_tree, async_extent->start,
 707                                          async_extent->start +
 708                                          async_extent->ram_size - 1);
 709
 710                         /* allocate blocks */
 711                         ret = cow_file_range(inode, async_cow->locked_page,
 712                                              async_extent->start,
 713                                              async_extent->start +
 714                                              async_extent->ram_size - 1,
 715                                              &page_started, &nr_written, 0);
 716
 717                         /* JDM XXX */
 718
 719                         /*
 720                          * if page_started, cow_file_range inserted an
 721                          * inline extent and took care of all the unlocking
 722                          * and IO for us.  Otherwise, we need to submit
 723                          * all those pages down to the drive.
 724                          */
 725                         if (!page_started && !ret)
 726                                 extent_write_locked_range(io_tree,
 727                                                   inode, async_extent->start,
 728                                                   async_extent->start +
 729                                                   async_extent->ram_size - 1,
 730                                                   btrfs_get_extent,
 731                                                   WB_SYNC_ALL);
 732                         else if (ret)
 733                                 unlock_page(async_cow->locked_page);
 734                         kfree(async_extent);
 735                         cond_resched();
 736                         continue;
 737                 }
 738
 739                 lock_extent(io_tree, async_extent->start,
 740                             async_extent->start + async_extent->ram_size - 1);
 741
 742                 ret = btrfs_reserve_extent(root,
 743                                            async_extent->compressed_size,
 744                                            async_extent->compressed_size,
 745                                            0, alloc_hint, &ins, 1, 1);
 746                 if (ret) {
 747                         free_async_extent_pages(async_extent);
 748
 749                         if (ret == -ENOSPC) {
 750                                 unlock_extent(io_tree, async_extent->start,
 751                                               async_extent->start +
 752                                               async_extent->ram_size - 1);
 753
 754                                 /*
 755                                  * we need to redirty the pages if we decide to
 756                                  * fallback to uncompressed IO, otherwise we
 757                                  * will not submit these pages down to lower
 758                                  * layers.
 759                                  */
 760                                 extent_range_redirty_for_io(inode,
 761                                                 async_extent->start,
 762                                                 async_extent->start +
 763                                                 async_extent->ram_size - 1);
 764
 765                                 goto retry;
 766                         }
 767                         goto out_free;
 768                 }
 769                 /*
 770                  * here we're doing allocation and writeback of the
 771                  * compressed pages
 772                  */
 773                 btrfs_drop_extent_cache(inode, async_extent->start,
 774                                         async_extent->start +
 775                                         async_extent->ram_size - 1, 0);
 776
 777                 em = alloc_extent_map();
 778                 if (!em) {
 779                         ret = -ENOMEM;
 780                         goto out_free_reserve;
 781                 }
 782                 em->start = async_extent->start;
 783                 em->len = async_extent->ram_size;
 784                 em->orig_start = em->start;
 785                 em->mod_start = em->start;
 786                 em->mod_len = em->len;
 787
 788                 em->block_start = ins.objectid;
 789                 em->block_len = ins.offset;
 790                 em->orig_block_len = ins.offset;
 791                 em->ram_bytes = async_extent->ram_size;
 792                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 793                 em->compress_type = async_extent->compress_type;
 794                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 795                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 796                 em->generation = -1;
 797
 798                 while (1) {
 799                         write_lock(&em_tree->lock);
 800                         ret = add_extent_mapping(em_tree, em, 1);
 801                         write_unlock(&em_tree->lock);
 802                         if (ret != -EEXIST) {
 803                                 free_extent_map(em);
 804                                 break;
 805                         }
 806                         btrfs_drop_extent_cache(inode, async_extent->start,
 807                                                 async_extent->start +
 808                                                 async_extent->ram_size - 1, 0);
 809                 }
 810
 811                 if (ret)
 812                         goto out_free_reserve;
 813
 814                 ret = btrfs_add_ordered_extent_compress(inode,
 815                                                 async_extent->start,
 816                                                 ins.objectid,
 817                                                 async_extent->ram_size,
 818                                                 ins.offset,
 819                                                 BTRFS_ORDERED_COMPRESSED,
 820                                                 async_extent->compress_type);
 821                 if (ret) {
 822                         btrfs_drop_extent_cache(inode, async_extent->start,
 823                                                 async_extent->start +
 824                                                 async_extent->ram_size - 1, 0);
 825                         goto out_free_reserve;
 826                 }
 827                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 828
 829                 /*
 830                  * clear dirty, set writeback and unlock the pages.
 831                  */
 832                 extent_clear_unlock_delalloc(inode, async_extent->start,
 833                                 async_extent->start +
 834                                 async_extent->ram_size - 1,
 835                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 836                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 837                                 PAGE_SET_WRITEBACK);
 838                 ret = btrfs_submit_compressed_write(inode,
 839                                     async_extent->start,
 840                                     async_extent->ram_size,
 841                                     ins.objectid,
 842                                     ins.offset, async_extent->pages,
 843                                     async_extent->nr_pages);
 844                 if (ret) {
 845                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 846                         struct page *p = async_extent->pages[0];
 847                         const u64 start = async_extent->start;
 848                         const u64 end = start + async_extent->ram_size - 1;
 849
 850                         p->mapping = inode->i_mapping;
 851                         tree->ops->writepage_end_io_hook(p, start, end,
 852                                                          NULL, 0);
 853                         p->mapping = NULL;
 854                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 855                                                      PAGE_END_WRITEBACK |
 856                                                      PAGE_SET_ERROR);
 857                         free_async_extent_pages(async_extent);
 858                 }
 859                 alloc_hint = ins.objectid + ins.offset;
 860                 kfree(async_extent);
 861                 cond_resched();
 862         }
 863         return;
 864 out_free_reserve:
 865         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 866         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 867 out_free:
 868         extent_clear_unlock_delalloc(inode, async_extent->start,
 869                                      async_extent->start +
 870                                      async_extent->ram_size - 1,
 871                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 872                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 873                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 874                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 875                                      PAGE_SET_ERROR);
 876         free_async_extent_pages(async_extent);
 877         kfree(async_extent);
 878         goto again;
 879 }
 880
 881 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 882                                       u64 num_bytes)
 883 {
 884         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 885         struct extent_map *em;
 886         u64 alloc_hint = 0;
 887
 888         read_lock(&em_tree->lock);
 889         em = search_extent_mapping(em_tree, start, num_bytes);
 890         if (em) {
 891                 /*
 892                  * if block start isn't an actual block number then find the
 893                  * first block in this inode and use that as a hint.  If that
 894                  * block is also bogus then just don't worry about it.
 895                  */
 896                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 897                         free_extent_map(em);
 898                         em = search_extent_mapping(em_tree, 0, 0);
 899                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 900                                 alloc_hint = em->block_start;
 901                         if (em)
 902                                 free_extent_map(em);
 903                 } else {
 904                         alloc_hint = em->block_start;
 905                         free_extent_map(em);
 906                 }
 907         }
 908         read_unlock(&em_tree->lock);
 909
 910         return alloc_hint;
 911 }
 912
 913 /*
 914  * when extent_io.c finds a delayed allocation range in the file,
 915  * the call backs end up in this code.  The basic idea is to
 916  * allocate extents on disk for the range, and create ordered data structs
 917  * in ram to track those extents.
 918  *
 919  * locked_page is the page that writepage had locked already.  We use
 920  * it to make sure we don't do extra locks or unlocks.
 921  *
 922  * *page_started is set to one if we unlock locked_page and do everything
 923  * required to start IO on it.  It may be clean and already done with
 924  * IO when we return.
 925  */
 926 static noinline int cow_file_range(struct inode *inode,
 927                                    struct page *locked_page,
 928                                    u64 start, u64 end, int *page_started,
 929                                    unsigned long *nr_written,
 930                                    int unlock)
 931 {
 932         struct btrfs_root *root = BTRFS_I(inode)->root;
 933         u64 alloc_hint = 0;
 934         u64 num_bytes;
 935         unsigned long ram_size;
 936         u64 disk_num_bytes;
 937         u64 cur_alloc_size;
 938         u64 blocksize = root->sectorsize;
 939         struct btrfs_key ins;
 940         struct extent_map *em;
 941         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 942         int ret = 0;
 943
 944         if (btrfs_is_free_space_inode(inode)) {
 945                 WARN_ON_ONCE(1);
 946                 ret = -EINVAL;
 947                 goto out_unlock;
 948         }
 949
 950         num_bytes = ALIGN(end - start + 1, blocksize);
 951         num_bytes = max(blocksize,  num_bytes);
 952         disk_num_bytes = num_bytes;
 953
 954         /* if this is a small write inside eof, kick off defrag */
 955         if (num_bytes < SZ_64K &&
 956             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 957                 btrfs_add_inode_defrag(NULL, inode);
 958
 959         if (start == 0) {
 960                 /* lets try to make an inline extent */
 961                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 962                                             NULL);
 963                 if (ret == 0) {
 964                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 965                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 966                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 967                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 968                                      PAGE_END_WRITEBACK);
 969
 970                         *nr_written = *nr_written +
 971                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 972                         *page_started = 1;
 973                         goto out;
 974                 } else if (ret < 0) {
 975                         goto out_unlock;
 976                 }
 977         }
 978
 979         BUG_ON(disk_num_bytes >
 980                btrfs_super_total_bytes(root->fs_info->super_copy));
 981
 982         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 983         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 984
 985         while (disk_num_bytes > 0) {
 986                 unsigned long op;
 987
 988                 cur_alloc_size = disk_num_bytes;
 989                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 990                                            root->sectorsize, 0, alloc_hint,
 991                                            &ins, 1, 1);
 992                 if (ret < 0)
 993                         goto out_unlock;
 994
 995                 em = alloc_extent_map();
 996                 if (!em) {
 997                         ret = -ENOMEM;
 998                         goto out_reserve;
 999                 }
1000                 em->start = start;
1001                 em->orig_start = em->start;
1002                 ram_size = ins.offset;
1003                 em->len = ins.offset;
1004                 em->mod_start = em->start;
1005                 em->mod_len = em->len;
1006
1007                 em->block_start = ins.objectid;
1008                 em->block_len = ins.offset;
1009                 em->orig_block_len = ins.offset;
1010                 em->ram_bytes = ram_size;
1011                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1012                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1013                 em->generation = -1;
1014
1015                 while (1) {
1016                         write_lock(&em_tree->lock);
1017                         ret = add_extent_mapping(em_tree, em, 1);
1018                         write_unlock(&em_tree->lock);
1019                         if (ret != -EEXIST) {
1020                                 free_extent_map(em);
1021                                 break;
1022                         }
1023                         btrfs_drop_extent_cache(inode, start,
1024                                                 start + ram_size - 1, 0);
1025                 }
1026                 if (ret)
1027                         goto out_reserve;
1028
1029                 cur_alloc_size = ins.offset;
1030                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1031                                                ram_size, cur_alloc_size, 0);
1032                 if (ret)
1033                         goto out_drop_extent_cache;
1034
1035                 if (root->root_key.objectid ==
1036                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1037                         ret = btrfs_reloc_clone_csums(inode, start,
1038                                                       cur_alloc_size);
1039                         if (ret)
1040                                 goto out_drop_extent_cache;
1041                 }
1042
1043                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1044
1045                 if (disk_num_bytes < cur_alloc_size)
1046                         break;
1047
1048                 /* we're not doing compressed IO, don't unlock the first
1049                  * page (which the caller expects to stay locked), don't
1050                  * clear any dirty bits and don't set any writeback bits
1051                  *
1052                  * Do set the Private2 bit so we know this page was properly
1053                  * setup for writepage
1054                  */
1055                 op = unlock ? PAGE_UNLOCK : 0;
1056                 op |= PAGE_SET_PRIVATE2;
1057
1058                 extent_clear_unlock_delalloc(inode, start,
1059                                              start + ram_size - 1, locked_page,
1060                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1061                                              op);
1062                 disk_num_bytes -= cur_alloc_size;
1063                 num_bytes -= cur_alloc_size;
1064                 alloc_hint = ins.objectid + ins.offset;
1065                 start += cur_alloc_size;
1066         }
1067 out:
1068         return ret;
1069
1070 out_drop_extent_cache:
1071         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1072 out_reserve:
1073         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1074         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1075 out_unlock:
1076         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1077                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1078                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1079                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1080                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1081         goto out;
1082 }
1083
1084 /*
1085  * work queue call back to started compression on a file and pages
1086  */
1087 static noinline void async_cow_start(struct btrfs_work *work)
1088 {
1089         struct async_cow *async_cow;
1090         int num_added = 0;
1091         async_cow = container_of(work, struct async_cow, work);
1092
1093         compress_file_range(async_cow->inode, async_cow->locked_page,
1094                             async_cow->start, async_cow->end, async_cow,
1095                             &num_added);
1096         if (num_added == 0) {
1097                 btrfs_add_delayed_iput(async_cow->inode);
1098                 async_cow->inode = NULL;
1099         }
1100 }
1101
1102 /*
1103  * work queue call back to submit previously compressed pages
1104  */
1105 static noinline void async_cow_submit(struct btrfs_work *work)
1106 {
1107         struct async_cow *async_cow;
1108         struct btrfs_root *root;
1109         unsigned long nr_pages;
1110
1111         async_cow = container_of(work, struct async_cow, work);
1112
1113         root = async_cow->root;
1114         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1115                 PAGE_SHIFT;
1116
1117         /*
1118          * atomic_sub_return implies a barrier for waitqueue_active
1119          */
1120         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1121             5 * SZ_1M &&
1122             waitqueue_active(&root->fs_info->async_submit_wait))
1123                 wake_up(&root->fs_info->async_submit_wait);
1124
1125         if (async_cow->inode)
1126                 submit_compressed_extents(async_cow->inode, async_cow);
1127 }
1128
1129 static noinline void async_cow_free(struct btrfs_work *work)
1130 {
1131         struct async_cow *async_cow;
1132         async_cow = container_of(work, struct async_cow, work);
1133         if (async_cow->inode)
1134                 btrfs_add_delayed_iput(async_cow->inode);
1135         kfree(async_cow);
1136 }
1137
1138 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1139                                 u64 start, u64 end, int *page_started,
1140                                 unsigned long *nr_written)
1141 {
1142         struct async_cow *async_cow;
1143         struct btrfs_root *root = BTRFS_I(inode)->root;
1144         unsigned long nr_pages;
1145         u64 cur_end;
1146         int limit = 10 * SZ_1M;
1147
1148         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1149                          1, 0, NULL, GFP_NOFS);
1150         while (start < end) {
1151                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1152                 BUG_ON(!async_cow); /* -ENOMEM */
1153                 async_cow->inode = igrab(inode);
1154                 async_cow->root = root;
1155                 async_cow->locked_page = locked_page;
1156                 async_cow->start = start;
1157
1158                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1159                     !btrfs_test_opt(root, FORCE_COMPRESS))
1160                         cur_end = end;
1161                 else
1162                         cur_end = min(end, start + SZ_512K - 1);
1163
1164                 async_cow->end = cur_end;
1165                 INIT_LIST_HEAD(&async_cow->extents);
1166
1167                 btrfs_init_work(&async_cow->work,
1168                                 btrfs_delalloc_helper,
1169                                 async_cow_start, async_cow_submit,
1170                                 async_cow_free);
1171
1172                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1173                         PAGE_SHIFT;
1174                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1175
1176                 btrfs_queue_work(root->fs_info->delalloc_workers,
1177                                  &async_cow->work);
1178
1179                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1180                         wait_event(root->fs_info->async_submit_wait,
1181                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1182                             limit));
1183                 }
1184
1185                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1186                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1187                         wait_event(root->fs_info->async_submit_wait,
1188                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1189                            0));
1190                 }
1191
1192                 *nr_written += nr_pages;
1193                 start = cur_end + 1;
1194         }
1195         *page_started = 1;
1196         return 0;
1197 }
1198
1199 static noinline int csum_exist_in_range(struct btrfs_root *root,
1200                                         u64 bytenr, u64 num_bytes)
1201 {
1202         int ret;
1203         struct btrfs_ordered_sum *sums;
1204         LIST_HEAD(list);
1205
1206         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1207                                        bytenr + num_bytes - 1, &list, 0);
1208         if (ret == 0 && list_empty(&list))
1209                 return 0;
1210
1211         while (!list_empty(&list)) {
1212                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1213                 list_del(&sums->list);
1214                 kfree(sums);
1215         }
1216         return 1;
1217 }
1218
1219 /*
1220  * when nowcow writeback call back.  This checks for snapshots or COW copies
1221  * of the extents that exist in the file, and COWs the file as required.
1222  *
1223  * If no cow copies or snapshots exist, we write directly to the existing
1224  * blocks on disk
1225  */
1226 static noinline int run_delalloc_nocow(struct inode *inode,
1227                                        struct page *locked_page,
1228                               u64 start, u64 end, int *page_started, int force,
1229                               unsigned long *nr_written)
1230 {
1231         struct btrfs_root *root = BTRFS_I(inode)->root;
1232         struct btrfs_trans_handle *trans;
1233         struct extent_buffer *leaf;
1234         struct btrfs_path *path;
1235         struct btrfs_file_extent_item *fi;
1236         struct btrfs_key found_key;
1237         u64 cow_start;
1238         u64 cur_offset;
1239         u64 extent_end;
1240         u64 extent_offset;
1241         u64 disk_bytenr;
1242         u64 num_bytes;
1243         u64 disk_num_bytes;
1244         u64 ram_bytes;
1245         int extent_type;
1246         int ret, err;
1247         int type;
1248         int nocow;
1249         int check_prev = 1;
1250         bool nolock;
1251         u64 ino = btrfs_ino(inode);
1252
1253         path = btrfs_alloc_path();
1254         if (!path) {
1255                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1256                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1257                                              EXTENT_DO_ACCOUNTING |
1258                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1259                                              PAGE_CLEAR_DIRTY |
1260                                              PAGE_SET_WRITEBACK |
1261                                              PAGE_END_WRITEBACK);
1262                 return -ENOMEM;
1263         }
1264
1265         nolock = btrfs_is_free_space_inode(inode);
1266
1267         if (nolock)
1268                 trans = btrfs_join_transaction_nolock(root);
1269         else
1270                 trans = btrfs_join_transaction(root);
1271
1272         if (IS_ERR(trans)) {
1273                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1274                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1275                                              EXTENT_DO_ACCOUNTING |
1276                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1277                                              PAGE_CLEAR_DIRTY |
1278                                              PAGE_SET_WRITEBACK |
1279                                              PAGE_END_WRITEBACK);
1280                 btrfs_free_path(path);
1281                 return PTR_ERR(trans);
1282         }
1283
1284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1285
1286         cow_start = (u64)-1;
1287         cur_offset = start;
1288         while (1) {
1289                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1290                                                cur_offset, 0);
1291                 if (ret < 0)
1292                         goto error;
1293                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1294                         leaf = path->nodes[0];
1295                         btrfs_item_key_to_cpu(leaf, &found_key,
1296                                               path->slots[0] - 1);
1297                         if (found_key.objectid == ino &&
1298                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1299                                 path->slots[0]--;
1300                 }
1301                 check_prev = 0;
1302 next_slot:
1303                 leaf = path->nodes[0];
1304                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1305                         ret = btrfs_next_leaf(root, path);
1306                         if (ret < 0)
1307                                 goto error;
1308                         if (ret > 0)
1309                                 break;
1310                         leaf = path->nodes[0];
1311                 }
1312
1313                 nocow = 0;
1314                 disk_bytenr = 0;
1315                 num_bytes = 0;
1316                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1317
1318                 if (found_key.objectid > ino)
1319                         break;
1320                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1321                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1322                         path->slots[0]++;
1323                         goto next_slot;
1324                 }
1325                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1326                     found_key.offset > end)
1327                         break;
1328
1329                 if (found_key.offset > cur_offset) {
1330                         extent_end = found_key.offset;
1331                         extent_type = 0;
1332                         goto out_check;
1333                 }
1334
1335                 fi = btrfs_item_ptr(leaf, path->slots[0],
1336                                     struct btrfs_file_extent_item);
1337                 extent_type = btrfs_file_extent_type(leaf, fi);
1338
1339                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1340                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1341                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1342                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1343                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1344                         extent_end = found_key.offset +
1345                                 btrfs_file_extent_num_bytes(leaf, fi);
1346                         disk_num_bytes =
1347                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1348                         if (extent_end <= start) {
1349                                 path->slots[0]++;
1350                                 goto next_slot;
1351                         }
1352                         if (disk_bytenr == 0)
1353                                 goto out_check;
1354                         if (btrfs_file_extent_compression(leaf, fi) ||
1355                             btrfs_file_extent_encryption(leaf, fi) ||
1356                             btrfs_file_extent_other_encoding(leaf, fi))
1357                                 goto out_check;
1358                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1359                                 goto out_check;
1360                         if (btrfs_extent_readonly(root, disk_bytenr))
1361                                 goto out_check;
1362                         if (btrfs_cross_ref_exist(trans, root, ino,
1363                                                   found_key.offset -
1364                                                   extent_offset, disk_bytenr))
1365                                 goto out_check;
1366                         disk_bytenr += extent_offset;
1367                         disk_bytenr += cur_offset - found_key.offset;
1368                         num_bytes = min(end + 1, extent_end) - cur_offset;
1369                         /*
1370                          * if there are pending snapshots for this root,
1371                          * we fall into common COW way.
1372                          */
1373                         if (!nolock) {
1374                                 err = btrfs_start_write_no_snapshoting(root);
1375                                 if (!err)
1376                                         goto out_check;
1377                         }
1378                         /*
1379                          * force cow if csum exists in the range.
1380                          * this ensure that csum for a given extent are
1381                          * either valid or do not exist.
1382                          */
1383                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1384                                 goto out_check;
1385                         if (!btrfs_inc_nocow_writers(root->fs_info,
1386                                                      disk_bytenr))
1387                                 goto out_check;
1388                         nocow = 1;
1389                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1390                         extent_end = found_key.offset +
1391                                 btrfs_file_extent_inline_len(leaf,
1392                                                      path->slots[0], fi);
1393                         extent_end = ALIGN(extent_end, root->sectorsize);
1394                 } else {
1395                         BUG_ON(1);
1396                 }
1397 out_check:
1398                 if (extent_end <= start) {
1399                         path->slots[0]++;
1400                         if (!nolock && nocow)
1401                                 btrfs_end_write_no_snapshoting(root);
1402                         if (nocow)
1403                                 btrfs_dec_nocow_writers(root->fs_info,
1404                                                         disk_bytenr);
1405                         goto next_slot;
1406                 }
1407                 if (!nocow) {
1408                         if (cow_start == (u64)-1)
1409                                 cow_start = cur_offset;
1410                         cur_offset = extent_end;
1411                         if (cur_offset > end)
1412                                 break;
1413                         path->slots[0]++;
1414                         goto next_slot;
1415                 }
1416
1417                 btrfs_release_path(path);
1418                 if (cow_start != (u64)-1) {
1419                         ret = cow_file_range(inode, locked_page,
1420                                              cow_start, found_key.offset - 1,
1421                                              page_started, nr_written, 1);
1422                         if (ret) {
1423                                 if (!nolock && nocow)
1424                                         btrfs_end_write_no_snapshoting(root);
1425                                 if (nocow)
1426                                         btrfs_dec_nocow_writers(root->fs_info,
1427                                                                 disk_bytenr);
1428                                 goto error;
1429                         }
1430                         cow_start = (u64)-1;
1431                 }
1432
1433                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1434                         struct extent_map *em;
1435                         struct extent_map_tree *em_tree;
1436                         em_tree = &BTRFS_I(inode)->extent_tree;
1437                         em = alloc_extent_map();
1438                         BUG_ON(!em); /* -ENOMEM */
1439                         em->start = cur_offset;
1440                         em->orig_start = found_key.offset - extent_offset;
1441                         em->len = num_bytes;
1442                         em->block_len = num_bytes;
1443                         em->block_start = disk_bytenr;
1444                         em->orig_block_len = disk_num_bytes;
1445                         em->ram_bytes = ram_bytes;
1446                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1447                         em->mod_start = em->start;
1448                         em->mod_len = em->len;
1449                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1450                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1451                         em->generation = -1;
1452                         while (1) {
1453                                 write_lock(&em_tree->lock);
1454                                 ret = add_extent_mapping(em_tree, em, 1);
1455                                 write_unlock(&em_tree->lock);
1456                                 if (ret != -EEXIST) {
1457                                         free_extent_map(em);
1458                                         break;
1459                                 }
1460                                 btrfs_drop_extent_cache(inode, em->start,
1461                                                 em->start + em->len - 1, 0);
1462                         }
1463                         type = BTRFS_ORDERED_PREALLOC;
1464                 } else {
1465                         type = BTRFS_ORDERED_NOCOW;
1466                 }
1467
1468                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1469                                                num_bytes, num_bytes, type);
1470                 if (nocow)
1471                         btrfs_dec_nocow_writers(root->fs_info, disk_bytenr);
1472                 BUG_ON(ret); /* -ENOMEM */
1473
1474                 if (root->root_key.objectid ==
1475                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1476                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1477                                                       num_bytes);
1478                         if (ret) {
1479                                 if (!nolock && nocow)
1480                                         btrfs_end_write_no_snapshoting(root);
1481                                 goto error;
1482                         }
1483                 }
1484
1485                 extent_clear_unlock_delalloc(inode, cur_offset,
1486                                              cur_offset + num_bytes - 1,
1487                                              locked_page, EXTENT_LOCKED |
1488                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1489                                              PAGE_SET_PRIVATE2);
1490                 if (!nolock && nocow)
1491                         btrfs_end_write_no_snapshoting(root);
1492                 cur_offset = extent_end;
1493                 if (cur_offset > end)
1494                         break;
1495         }
1496         btrfs_release_path(path);
1497
1498         if (cur_offset <= end && cow_start == (u64)-1) {
1499                 cow_start = cur_offset;
1500                 cur_offset = end;
1501         }
1502
1503         if (cow_start != (u64)-1) {
1504                 ret = cow_file_range(inode, locked_page, cow_start, end,
1505                                      page_started, nr_written, 1);
1506                 if (ret)
1507                         goto error;
1508         }
1509
1510 error:
1511         err = btrfs_end_transaction(trans, root);
1512         if (!ret)
1513                 ret = err;
1514
1515         if (ret && cur_offset < end)
1516                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1517                                              locked_page, EXTENT_LOCKED |
1518                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1519                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1520                                              PAGE_CLEAR_DIRTY |
1521                                              PAGE_SET_WRITEBACK |
1522                                              PAGE_END_WRITEBACK);
1523         btrfs_free_path(path);
1524         return ret;
1525 }
1526
1527 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1528 {
1529
1530         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1531             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1532                 return 0;
1533
1534         /*
1535          * @defrag_bytes is a hint value, no spinlock held here,
1536          * if is not zero, it means the file is defragging.
1537          * Force cow if given extent needs to be defragged.
1538          */
1539         if (BTRFS_I(inode)->defrag_bytes &&
1540             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1541                            EXTENT_DEFRAG, 0, NULL))
1542                 return 1;
1543
1544         return 0;
1545 }
1546
1547 /*
1548  * extent_io.c call back to do delayed allocation processing
1549  */
1550 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1551                               u64 start, u64 end, int *page_started,
1552                               unsigned long *nr_written)
1553 {
1554         int ret;
1555         int force_cow = need_force_cow(inode, start, end);
1556
1557         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1558                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1559                                          page_started, 1, nr_written);
1560         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1561                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1562                                          page_started, 0, nr_written);
1563         } else if (!inode_need_compress(inode)) {
1564                 ret = cow_file_range(inode, locked_page, start, end,
1565                                       page_started, nr_written, 1);
1566         } else {
1567                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1568                         &BTRFS_I(inode)->runtime_flags);
1569                 ret = cow_file_range_async(inode, locked_page, start, end,
1570                                            page_started, nr_written);
1571         }
1572         return ret;
1573 }
1574
1575 static void btrfs_split_extent_hook(struct inode *inode,
1576                                     struct extent_state *orig, u64 split)
1577 {
1578         u64 size;
1579
1580         /* not delalloc, ignore it */
1581         if (!(orig->state & EXTENT_DELALLOC))
1582                 return;
1583
1584         size = orig->end - orig->start + 1;
1585         if (size > BTRFS_MAX_EXTENT_SIZE) {
1586                 u64 num_extents;
1587                 u64 new_size;
1588
1589                 /*
1590                  * See the explanation in btrfs_merge_extent_hook, the same
1591                  * applies here, just in reverse.
1592                  */
1593                 new_size = orig->end - split + 1;
1594                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1595                                         BTRFS_MAX_EXTENT_SIZE);
1596                 new_size = split - orig->start;
1597                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1598                                         BTRFS_MAX_EXTENT_SIZE);
1599                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1600                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1601                         return;
1602         }
1603
1604         spin_lock(&BTRFS_I(inode)->lock);
1605         BTRFS_I(inode)->outstanding_extents++;
1606         spin_unlock(&BTRFS_I(inode)->lock);
1607 }
1608
1609 /*
1610  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1611  * extents so we can keep track of new extents that are just merged onto old
1612  * extents, such as when we are doing sequential writes, so we can properly
1613  * account for the metadata space we'll need.
1614  */
1615 static void btrfs_merge_extent_hook(struct inode *inode,
1616                                     struct extent_state *new,
1617                                     struct extent_state *other)
1618 {
1619         u64 new_size, old_size;
1620         u64 num_extents;
1621
1622         /* not delalloc, ignore it */
1623         if (!(other->state & EXTENT_DELALLOC))
1624                 return;
1625
1626         if (new->start > other->start)
1627                 new_size = new->end - other->start + 1;
1628         else
1629                 new_size = other->end - new->start + 1;
1630
1631         /* we're not bigger than the max, unreserve the space and go */
1632         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1633                 spin_lock(&BTRFS_I(inode)->lock);
1634                 BTRFS_I(inode)->outstanding_extents--;
1635                 spin_unlock(&BTRFS_I(inode)->lock);
1636                 return;
1637         }
1638
1639         /*
1640          * We have to add up either side to figure out how many extents were
1641          * accounted for before we merged into one big extent.  If the number of
1642          * extents we accounted for is <= the amount we need for the new range
1643          * then we can return, otherwise drop.  Think of it like this
1644          *
1645          * [ 4k][MAX_SIZE]
1646          *
1647          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1648          * need 2 outstanding extents, on one side we have 1 and the other side
1649          * we have 1 so they are == and we can return.  But in this case
1650          *
1651          * [MAX_SIZE+4k][MAX_SIZE+4k]
1652          *
1653          * Each range on their own accounts for 2 extents, but merged together
1654          * they are only 3 extents worth of accounting, so we need to drop in
1655          * this case.
1656          */
1657         old_size = other->end - other->start + 1;
1658         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1659                                 BTRFS_MAX_EXTENT_SIZE);
1660         old_size = new->end - new->start + 1;
1661         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1662                                  BTRFS_MAX_EXTENT_SIZE);
1663
1664         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1665                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1666                 return;
1667
1668         spin_lock(&BTRFS_I(inode)->lock);
1669         BTRFS_I(inode)->outstanding_extents--;
1670         spin_unlock(&BTRFS_I(inode)->lock);
1671 }
1672
1673 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1674                                       struct inode *inode)
1675 {
1676         spin_lock(&root->delalloc_lock);
1677         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1678                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1679                               &root->delalloc_inodes);
1680                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1681                         &BTRFS_I(inode)->runtime_flags);
1682                 root->nr_delalloc_inodes++;
1683                 if (root->nr_delalloc_inodes == 1) {
1684                         spin_lock(&root->fs_info->delalloc_root_lock);
1685                         BUG_ON(!list_empty(&root->delalloc_root));
1686                         list_add_tail(&root->delalloc_root,
1687                                       &root->fs_info->delalloc_roots);
1688                         spin_unlock(&root->fs_info->delalloc_root_lock);
1689                 }
1690         }
1691         spin_unlock(&root->delalloc_lock);
1692 }
1693
1694 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1695                                      struct inode *inode)
1696 {
1697         spin_lock(&root->delalloc_lock);
1698         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1699                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1700                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1701                           &BTRFS_I(inode)->runtime_flags);
1702                 root->nr_delalloc_inodes--;
1703                 if (!root->nr_delalloc_inodes) {
1704                         spin_lock(&root->fs_info->delalloc_root_lock);
1705                         BUG_ON(list_empty(&root->delalloc_root));
1706                         list_del_init(&root->delalloc_root);
1707                         spin_unlock(&root->fs_info->delalloc_root_lock);
1708                 }
1709         }
1710         spin_unlock(&root->delalloc_lock);
1711 }
1712
1713 /*
1714  * extent_io.c set_bit_hook, used to track delayed allocation
1715  * bytes in this file, and to maintain the list of inodes that
1716  * have pending delalloc work to be done.
1717  */
1718 static void btrfs_set_bit_hook(struct inode *inode,
1719                                struct extent_state *state, unsigned *bits)
1720 {
1721
1722         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1723                 WARN_ON(1);
1724         /*
1725          * set_bit and clear bit hooks normally require _irqsave/restore
1726          * but in this case, we are only testing for the DELALLOC
1727          * bit, which is only set or cleared with irqs on
1728          */
1729         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1730                 struct btrfs_root *root = BTRFS_I(inode)->root;
1731                 u64 len = state->end + 1 - state->start;
1732                 bool do_list = !btrfs_is_free_space_inode(inode);
1733
1734                 if (*bits & EXTENT_FIRST_DELALLOC) {
1735                         *bits &= ~EXTENT_FIRST_DELALLOC;
1736                 } else {
1737                         spin_lock(&BTRFS_I(inode)->lock);
1738                         BTRFS_I(inode)->outstanding_extents++;
1739                         spin_unlock(&BTRFS_I(inode)->lock);
1740                 }
1741
1742                 /* For sanity tests */
1743                 if (btrfs_test_is_dummy_root(root))
1744                         return;
1745
1746                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1747                                      root->fs_info->delalloc_batch);
1748                 spin_lock(&BTRFS_I(inode)->lock);
1749                 BTRFS_I(inode)->delalloc_bytes += len;
1750                 if (*bits & EXTENT_DEFRAG)
1751                         BTRFS_I(inode)->defrag_bytes += len;
1752                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1753                                          &BTRFS_I(inode)->runtime_flags))
1754                         btrfs_add_delalloc_inodes(root, inode);
1755                 spin_unlock(&BTRFS_I(inode)->lock);
1756         }
1757 }
1758
1759 /*
1760  * extent_io.c clear_bit_hook, see set_bit_hook for why
1761  */
1762 static void btrfs_clear_bit_hook(struct inode *inode,
1763                                  struct extent_state *state,
1764                                  unsigned *bits)
1765 {
1766         u64 len = state->end + 1 - state->start;
1767         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1768                                     BTRFS_MAX_EXTENT_SIZE);
1769
1770         spin_lock(&BTRFS_I(inode)->lock);
1771         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1772                 BTRFS_I(inode)->defrag_bytes -= len;
1773         spin_unlock(&BTRFS_I(inode)->lock);
1774
1775         /*
1776          * set_bit and clear bit hooks normally require _irqsave/restore
1777          * but in this case, we are only testing for the DELALLOC
1778          * bit, which is only set or cleared with irqs on
1779          */
1780         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1781                 struct btrfs_root *root = BTRFS_I(inode)->root;
1782                 bool do_list = !btrfs_is_free_space_inode(inode);
1783
1784                 if (*bits & EXTENT_FIRST_DELALLOC) {
1785                         *bits &= ~EXTENT_FIRST_DELALLOC;
1786                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1787                         spin_lock(&BTRFS_I(inode)->lock);
1788                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1789                         spin_unlock(&BTRFS_I(inode)->lock);
1790                 }
1791
1792                 /*
1793                  * We don't reserve metadata space for space cache inodes so we
1794                  * don't need to call dellalloc_release_metadata if there is an
1795                  * error.
1796                  */
1797                 if (*bits & EXTENT_DO_ACCOUNTING &&
1798                     root != root->fs_info->tree_root)
1799                         btrfs_delalloc_release_metadata(inode, len);
1800
1801                 /* For sanity tests. */
1802                 if (btrfs_test_is_dummy_root(root))
1803                         return;
1804
1805                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1806                     && do_list && !(state->state & EXTENT_NORESERVE))
1807                         btrfs_free_reserved_data_space_noquota(inode,
1808                                         state->start, len);
1809
1810                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1811                                      root->fs_info->delalloc_batch);
1812                 spin_lock(&BTRFS_I(inode)->lock);
1813                 BTRFS_I(inode)->delalloc_bytes -= len;
1814                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1815                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1816                              &BTRFS_I(inode)->runtime_flags))
1817                         btrfs_del_delalloc_inode(root, inode);
1818                 spin_unlock(&BTRFS_I(inode)->lock);
1819         }
1820 }
1821
1822 /*
1823  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1824  * we don't create bios that span stripes or chunks
1825  */
1826 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1827                          size_t size, struct bio *bio,
1828                          unsigned long bio_flags)
1829 {
1830         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1831         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1832         u64 length = 0;
1833         u64 map_length;
1834         int ret;
1835
1836         if (bio_flags & EXTENT_BIO_COMPRESSED)
1837                 return 0;
1838
1839         length = bio->bi_iter.bi_size;
1840         map_length = length;
1841         ret = btrfs_map_block(root->fs_info, rw, logical,
1842                               &map_length, NULL, 0);
1843         /* Will always return 0 with map_multi == NULL */
1844         BUG_ON(ret < 0);
1845         if (map_length < length + size)
1846                 return 1;
1847         return 0;
1848 }
1849
1850 /*
1851  * in order to insert checksums into the metadata in large chunks,
1852  * we wait until bio submission time.   All the pages in the bio are
1853  * checksummed and sums are attached onto the ordered extent record.
1854  *
1855  * At IO completion time the cums attached on the ordered extent record
1856  * are inserted into the btree
1857  */
1858 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1859                                     struct bio *bio, int mirror_num,
1860                                     unsigned long bio_flags,
1861                                     u64 bio_offset)
1862 {
1863         struct btrfs_root *root = BTRFS_I(inode)->root;
1864         int ret = 0;
1865
1866         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1867         BUG_ON(ret); /* -ENOMEM */
1868         return 0;
1869 }
1870
1871 /*
1872  * in order to insert checksums into the metadata in large chunks,
1873  * we wait until bio submission time.   All the pages in the bio are
1874  * checksummed and sums are attached onto the ordered extent record.
1875  *
1876  * At IO completion time the cums attached on the ordered extent record
1877  * are inserted into the btree
1878  */
1879 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1880                           int mirror_num, unsigned long bio_flags,
1881                           u64 bio_offset)
1882 {
1883         struct btrfs_root *root = BTRFS_I(inode)->root;
1884         int ret;
1885
1886         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1887         if (ret) {
1888                 bio->bi_error = ret;
1889                 bio_endio(bio);
1890         }
1891         return ret;
1892 }
1893
1894 /*
1895  * extent_io.c submission hook. This does the right thing for csum calculation
1896  * on write, or reading the csums from the tree before a read
1897  */
1898 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1899                           int mirror_num, unsigned long bio_flags,
1900                           u64 bio_offset)
1901 {
1902         struct btrfs_root *root = BTRFS_I(inode)->root;
1903         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1904         int ret = 0;
1905         int skip_sum;
1906         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1907
1908         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1909
1910         if (btrfs_is_free_space_inode(inode))
1911                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1912
1913         if (!(rw & REQ_WRITE)) {
1914                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1915                 if (ret)
1916                         goto out;
1917
1918                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1919                         ret = btrfs_submit_compressed_read(inode, bio,
1920                                                            mirror_num,
1921                                                            bio_flags);
1922                         goto out;
1923                 } else if (!skip_sum) {
1924                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1925                         if (ret)
1926                                 goto out;
1927                 }
1928                 goto mapit;
1929         } else if (async && !skip_sum) {
1930                 /* csum items have already been cloned */
1931                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1932                         goto mapit;
1933                 /* we're doing a write, do the async checksumming */
1934                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1935                                    inode, rw, bio, mirror_num,
1936                                    bio_flags, bio_offset,
1937                                    __btrfs_submit_bio_start,
1938                                    __btrfs_submit_bio_done);
1939                 goto out;
1940         } else if (!skip_sum) {
1941                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1942                 if (ret)
1943                         goto out;
1944         }
1945
1946 mapit:
1947         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1948
1949 out:
1950         if (ret < 0) {
1951                 bio->bi_error = ret;
1952                 bio_endio(bio);
1953         }
1954         return ret;
1955 }
1956
1957 /*
1958  * given a list of ordered sums record them in the inode.  This happens
1959  * at IO completion time based on sums calculated at bio submission time.
1960  */
1961 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1962                              struct inode *inode, u64 file_offset,
1963                              struct list_head *list)
1964 {
1965         struct btrfs_ordered_sum *sum;
1966
1967         list_for_each_entry(sum, list, list) {
1968                 trans->adding_csums = 1;
1969                 btrfs_csum_file_blocks(trans,
1970                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1971                 trans->adding_csums = 0;
1972         }
1973         return 0;
1974 }
1975
1976 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1977                               struct extent_state **cached_state)
1978 {
1979         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
1980         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1981                                    cached_state);
1982 }
1983
1984 /* see btrfs_writepage_start_hook for details on why this is required */
1985 struct btrfs_writepage_fixup {
1986         struct page *page;
1987         struct btrfs_work work;
1988 };
1989
1990 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1991 {
1992         struct btrfs_writepage_fixup *fixup;
1993         struct btrfs_ordered_extent *ordered;
1994         struct extent_state *cached_state = NULL;
1995         struct page *page;
1996         struct inode *inode;
1997         u64 page_start;
1998         u64 page_end;
1999         int ret;
2000
2001         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2002         page = fixup->page;
2003 again:
2004         lock_page(page);
2005         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2006                 ClearPageChecked(page);
2007                 goto out_page;
2008         }
2009
2010         inode = page->mapping->host;
2011         page_start = page_offset(page);
2012         page_end = page_offset(page) + PAGE_SIZE - 1;
2013
2014         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2015                          &cached_state);
2016
2017         /* already ordered? We're done */
2018         if (PagePrivate2(page))
2019                 goto out;
2020
2021         ordered = btrfs_lookup_ordered_range(inode, page_start,
2022                                         PAGE_SIZE);
2023         if (ordered) {
2024                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2025                                      page_end, &cached_state, GFP_NOFS);
2026                 unlock_page(page);
2027                 btrfs_start_ordered_extent(inode, ordered, 1);
2028                 btrfs_put_ordered_extent(ordered);
2029                 goto again;
2030         }
2031
2032         ret = btrfs_delalloc_reserve_space(inode, page_start,
2033                                            PAGE_SIZE);
2034         if (ret) {
2035                 mapping_set_error(page->mapping, ret);
2036                 end_extent_writepage(page, ret, page_start, page_end);
2037                 ClearPageChecked(page);
2038                 goto out;
2039          }
2040
2041         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2042         ClearPageChecked(page);
2043         set_page_dirty(page);
2044 out:
2045         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2046                              &cached_state, GFP_NOFS);
2047 out_page:
2048         unlock_page(page);
2049         put_page(page);
2050         kfree(fixup);
2051 }
2052
2053 /*
2054  * There are a few paths in the higher layers of the kernel that directly
2055  * set the page dirty bit without asking the filesystem if it is a
2056  * good idea.  This causes problems because we want to make sure COW
2057  * properly happens and the data=ordered rules are followed.
2058  *
2059  * In our case any range that doesn't have the ORDERED bit set
2060  * hasn't been properly setup for IO.  We kick off an async process
2061  * to fix it up.  The async helper will wait for ordered extents, set
2062  * the delalloc bit and make it safe to write the page.
2063  */
2064 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2065 {
2066         struct inode *inode = page->mapping->host;
2067         struct btrfs_writepage_fixup *fixup;
2068         struct btrfs_root *root = BTRFS_I(inode)->root;
2069
2070         /* this page is properly in the ordered list */
2071         if (TestClearPagePrivate2(page))
2072                 return 0;
2073
2074         if (PageChecked(page))
2075                 return -EAGAIN;
2076
2077         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2078         if (!fixup)
2079                 return -EAGAIN;
2080
2081         SetPageChecked(page);
2082         get_page(page);
2083         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2084                         btrfs_writepage_fixup_worker, NULL, NULL);
2085         fixup->page = page;
2086         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2087         return -EBUSY;
2088 }
2089
2090 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2091                                        struct inode *inode, u64 file_pos,
2092                                        u64 disk_bytenr, u64 disk_num_bytes,
2093                                        u64 num_bytes, u64 ram_bytes,
2094                                        u8 compression, u8 encryption,
2095                                        u16 other_encoding, int extent_type)
2096 {
2097         struct btrfs_root *root = BTRFS_I(inode)->root;
2098         struct btrfs_file_extent_item *fi;
2099         struct btrfs_path *path;
2100         struct extent_buffer *leaf;
2101         struct btrfs_key ins;
2102         int extent_inserted = 0;
2103         int ret;
2104
2105         path = btrfs_alloc_path();
2106         if (!path)
2107                 return -ENOMEM;
2108
2109         /*
2110          * we may be replacing one extent in the tree with another.
2111          * The new extent is pinned in the extent map, and we don't want
2112          * to drop it from the cache until it is completely in the btree.
2113          *
2114          * So, tell btrfs_drop_extents to leave this extent in the cache.
2115          * the caller is expected to unpin it and allow it to be merged
2116          * with the others.
2117          */
2118         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2119                                    file_pos + num_bytes, NULL, 0,
2120                                    1, sizeof(*fi), &extent_inserted);
2121         if (ret)
2122                 goto out;
2123
2124         if (!extent_inserted) {
2125                 ins.objectid = btrfs_ino(inode);
2126                 ins.offset = file_pos;
2127                 ins.type = BTRFS_EXTENT_DATA_KEY;
2128
2129                 path->leave_spinning = 1;
2130                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2131                                               sizeof(*fi));
2132                 if (ret)
2133                         goto out;
2134         }
2135         leaf = path->nodes[0];
2136         fi = btrfs_item_ptr(leaf, path->slots[0],
2137                             struct btrfs_file_extent_item);
2138         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2139         btrfs_set_file_extent_type(leaf, fi, extent_type);
2140         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2141         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2142         btrfs_set_file_extent_offset(leaf, fi, 0);
2143         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2144         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2145         btrfs_set_file_extent_compression(leaf, fi, compression);
2146         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2147         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2148
2149         btrfs_mark_buffer_dirty(leaf);
2150         btrfs_release_path(path);
2151
2152         inode_add_bytes(inode, num_bytes);
2153
2154         ins.objectid = disk_bytenr;
2155         ins.offset = disk_num_bytes;
2156         ins.type = BTRFS_EXTENT_ITEM_KEY;
2157         ret = btrfs_alloc_reserved_file_extent(trans, root,
2158                                         root->root_key.objectid,
2159                                         btrfs_ino(inode), file_pos,
2160                                         ram_bytes, &ins);
2161         /*
2162          * Release the reserved range from inode dirty range map, as it is
2163          * already moved into delayed_ref_head
2164          */
2165         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2166 out:
2167         btrfs_free_path(path);
2168
2169         return ret;
2170 }
2171
2172 /* snapshot-aware defrag */
2173 struct sa_defrag_extent_backref {
2174         struct rb_node node;
2175         struct old_sa_defrag_extent *old;
2176         u64 root_id;
2177         u64 inum;
2178         u64 file_pos;
2179         u64 extent_offset;
2180         u64 num_bytes;
2181         u64 generation;
2182 };
2183
2184 struct old_sa_defrag_extent {
2185         struct list_head list;
2186         struct new_sa_defrag_extent *new;
2187
2188         u64 extent_offset;
2189         u64 bytenr;
2190         u64 offset;
2191         u64 len;
2192         int count;
2193 };
2194
2195 struct new_sa_defrag_extent {
2196         struct rb_root root;
2197         struct list_head head;
2198         struct btrfs_path *path;
2199         struct inode *inode;
2200         u64 file_pos;
2201         u64 len;
2202         u64 bytenr;
2203         u64 disk_len;
2204         u8 compress_type;
2205 };
2206
2207 static int backref_comp(struct sa_defrag_extent_backref *b1,
2208                         struct sa_defrag_extent_backref *b2)
2209 {
2210         if (b1->root_id < b2->root_id)
2211                 return -1;
2212         else if (b1->root_id > b2->root_id)
2213                 return 1;
2214
2215         if (b1->inum < b2->inum)
2216                 return -1;
2217         else if (b1->inum > b2->inum)
2218                 return 1;
2219
2220         if (b1->file_pos < b2->file_pos)
2221                 return -1;
2222         else if (b1->file_pos > b2->file_pos)
2223                 return 1;
2224
2225         /*
2226          * [------------------------------] ===> (a range of space)
2227          *     |<--->|   |<---->| =============> (fs/file tree A)
2228          * |<---------------------------->| ===> (fs/file tree B)
2229          *
2230          * A range of space can refer to two file extents in one tree while
2231          * refer to only one file extent in another tree.
2232          *
2233          * So we may process a disk offset more than one time(two extents in A)
2234          * and locate at the same extent(one extent in B), then insert two same
2235          * backrefs(both refer to the extent in B).
2236          */
2237         return 0;
2238 }
2239
2240 static void backref_insert(struct rb_root *root,
2241                            struct sa_defrag_extent_backref *backref)
2242 {
2243         struct rb_node **p = &root->rb_node;
2244         struct rb_node *parent = NULL;
2245         struct sa_defrag_extent_backref *entry;
2246         int ret;
2247
2248         while (*p) {
2249                 parent = *p;
2250                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2251
2252                 ret = backref_comp(backref, entry);
2253                 if (ret < 0)
2254                         p = &(*p)->rb_left;
2255                 else
2256                         p = &(*p)->rb_right;
2257         }
2258
2259         rb_link_node(&backref->node, parent, p);
2260         rb_insert_color(&backref->node, root);
2261 }
2262
2263 /*
2264  * Note the backref might has changed, and in this case we just return 0.
2265  */
2266 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2267                                        void *ctx)
2268 {
2269         struct btrfs_file_extent_item *extent;
2270         struct btrfs_fs_info *fs_info;
2271         struct old_sa_defrag_extent *old = ctx;
2272         struct new_sa_defrag_extent *new = old->new;
2273         struct btrfs_path *path = new->path;
2274         struct btrfs_key key;
2275         struct btrfs_root *root;
2276         struct sa_defrag_extent_backref *backref;
2277         struct extent_buffer *leaf;
2278         struct inode *inode = new->inode;
2279         int slot;
2280         int ret;
2281         u64 extent_offset;
2282         u64 num_bytes;
2283
2284         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2285             inum == btrfs_ino(inode))
2286                 return 0;
2287
2288         key.objectid = root_id;
2289         key.type = BTRFS_ROOT_ITEM_KEY;
2290         key.offset = (u64)-1;
2291
2292         fs_info = BTRFS_I(inode)->root->fs_info;
2293         root = btrfs_read_fs_root_no_name(fs_info, &key);
2294         if (IS_ERR(root)) {
2295                 if (PTR_ERR(root) == -ENOENT)
2296                         return 0;
2297                 WARN_ON(1);
2298                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2299                          inum, offset, root_id);
2300                 return PTR_ERR(root);
2301         }
2302
2303         key.objectid = inum;
2304         key.type = BTRFS_EXTENT_DATA_KEY;
2305         if (offset > (u64)-1 << 32)
2306                 key.offset = 0;
2307         else
2308                 key.offset = offset;
2309
2310         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2311         if (WARN_ON(ret < 0))
2312                 return ret;
2313         ret = 0;
2314
2315         while (1) {
2316                 cond_resched();
2317
2318                 leaf = path->nodes[0];
2319                 slot = path->slots[0];
2320
2321                 if (slot >= btrfs_header_nritems(leaf)) {
2322                         ret = btrfs_next_leaf(root, path);
2323                         if (ret < 0) {
2324                                 goto out;
2325                         } else if (ret > 0) {
2326                                 ret = 0;
2327                                 goto out;
2328                         }
2329                         continue;
2330                 }
2331
2332                 path->slots[0]++;
2333
2334                 btrfs_item_key_to_cpu(leaf, &key, slot);
2335
2336                 if (key.objectid > inum)
2337                         goto out;
2338
2339                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2340                         continue;
2341
2342                 extent = btrfs_item_ptr(leaf, slot,
2343                                         struct btrfs_file_extent_item);
2344
2345                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2346                         continue;
2347
2348                 /*
2349                  * 'offset' refers to the exact key.offset,
2350                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2351                  * (key.offset - extent_offset).
2352                  */
2353                 if (key.offset != offset)
2354                         continue;
2355
2356                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2357                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2358
2359                 if (extent_offset >= old->extent_offset + old->offset +
2360                     old->len || extent_offset + num_bytes <=
2361                     old->extent_offset + old->offset)
2362                         continue;
2363                 break;
2364         }
2365
2366         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2367         if (!backref) {
2368                 ret = -ENOENT;
2369                 goto out;
2370         }
2371
2372         backref->root_id = root_id;
2373         backref->inum = inum;
2374         backref->file_pos = offset;
2375         backref->num_bytes = num_bytes;
2376         backref->extent_offset = extent_offset;
2377         backref->generation = btrfs_file_extent_generation(leaf, extent);
2378         backref->old = old;
2379         backref_insert(&new->root, backref);
2380         old->count++;
2381 out:
2382         btrfs_release_path(path);
2383         WARN_ON(ret);
2384         return ret;
2385 }
2386
2387 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2388                                    struct new_sa_defrag_extent *new)
2389 {
2390         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2391         struct old_sa_defrag_extent *old, *tmp;
2392         int ret;
2393
2394         new->path = path;
2395
2396         list_for_each_entry_safe(old, tmp, &new->head, list) {
2397                 ret = iterate_inodes_from_logical(old->bytenr +
2398                                                   old->extent_offset, fs_info,
2399                                                   path, record_one_backref,
2400                                                   old);
2401                 if (ret < 0 && ret != -ENOENT)
2402                         return false;
2403
2404                 /* no backref to be processed for this extent */
2405                 if (!old->count) {
2406                         list_del(&old->list);
2407                         kfree(old);
2408                 }
2409         }
2410
2411         if (list_empty(&new->head))
2412                 return false;
2413
2414         return true;
2415 }
2416
2417 static int relink_is_mergable(struct extent_buffer *leaf,
2418                               struct btrfs_file_extent_item *fi,
2419                               struct new_sa_defrag_extent *new)
2420 {
2421         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2422                 return 0;
2423
2424         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2425                 return 0;
2426
2427         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2428                 return 0;
2429
2430         if (btrfs_file_extent_encryption(leaf, fi) ||
2431             btrfs_file_extent_other_encoding(leaf, fi))
2432                 return 0;
2433
2434         return 1;
2435 }
2436
2437 /*
2438  * Note the backref might has changed, and in this case we just return 0.
2439  */
2440 static noinline int relink_extent_backref(struct btrfs_path *path,
2441                                  struct sa_defrag_extent_backref *prev,
2442                                  struct sa_defrag_extent_backref *backref)
2443 {
2444         struct btrfs_file_extent_item *extent;
2445         struct btrfs_file_extent_item *item;
2446         struct btrfs_ordered_extent *ordered;
2447         struct btrfs_trans_handle *trans;
2448         struct btrfs_fs_info *fs_info;
2449         struct btrfs_root *root;
2450         struct btrfs_key key;
2451         struct extent_buffer *leaf;
2452         struct old_sa_defrag_extent *old = backref->old;
2453         struct new_sa_defrag_extent *new = old->new;
2454         struct inode *src_inode = new->inode;
2455         struct inode *inode;
2456         struct extent_state *cached = NULL;
2457         int ret = 0;
2458         u64 start;
2459         u64 len;
2460         u64 lock_start;
2461         u64 lock_end;
2462         bool merge = false;
2463         int index;
2464
2465         if (prev && prev->root_id == backref->root_id &&
2466             prev->inum == backref->inum &&
2467             prev->file_pos + prev->num_bytes == backref->file_pos)
2468                 merge = true;
2469
2470         /* step 1: get root */
2471         key.objectid = backref->root_id;
2472         key.type = BTRFS_ROOT_ITEM_KEY;
2473         key.offset = (u64)-1;
2474
2475         fs_info = BTRFS_I(src_inode)->root->fs_info;
2476         index = srcu_read_lock(&fs_info->subvol_srcu);
2477
2478         root = btrfs_read_fs_root_no_name(fs_info, &key);
2479         if (IS_ERR(root)) {
2480                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2481                 if (PTR_ERR(root) == -ENOENT)
2482                         return 0;
2483                 return PTR_ERR(root);
2484         }
2485
2486         if (btrfs_root_readonly(root)) {
2487                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2488                 return 0;
2489         }
2490
2491         /* step 2: get inode */
2492         key.objectid = backref->inum;
2493         key.type = BTRFS_INODE_ITEM_KEY;
2494         key.offset = 0;
2495
2496         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2497         if (IS_ERR(inode)) {
2498                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2499                 return 0;
2500         }
2501
2502         srcu_read_unlock(&fs_info->subvol_srcu, index);
2503
2504         /* step 3: relink backref */
2505         lock_start = backref->file_pos;
2506         lock_end = backref->file_pos + backref->num_bytes - 1;
2507         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2508                          &cached);
2509
2510         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2511         if (ordered) {
2512                 btrfs_put_ordered_extent(ordered);
2513                 goto out_unlock;
2514         }
2515
2516         trans = btrfs_join_transaction(root);
2517         if (IS_ERR(trans)) {
2518                 ret = PTR_ERR(trans);
2519                 goto out_unlock;
2520         }
2521
2522         key.objectid = backref->inum;
2523         key.type = BTRFS_EXTENT_DATA_KEY;
2524         key.offset = backref->file_pos;
2525
2526         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2527         if (ret < 0) {
2528                 goto out_free_path;
2529         } else if (ret > 0) {
2530                 ret = 0;
2531                 goto out_free_path;
2532         }
2533
2534         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2535                                 struct btrfs_file_extent_item);
2536
2537         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2538             backref->generation)
2539                 goto out_free_path;
2540
2541         btrfs_release_path(path);
2542
2543         start = backref->file_pos;
2544         if (backref->extent_offset < old->extent_offset + old->offset)
2545                 start += old->extent_offset + old->offset -
2546                          backref->extent_offset;
2547
2548         len = min(backref->extent_offset + backref->num_bytes,
2549                   old->extent_offset + old->offset + old->len);
2550         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2551
2552         ret = btrfs_drop_extents(trans, root, inode, start,
2553                                  start + len, 1);
2554         if (ret)
2555                 goto out_free_path;
2556 again:
2557         key.objectid = btrfs_ino(inode);
2558         key.type = BTRFS_EXTENT_DATA_KEY;
2559         key.offset = start;
2560
2561         path->leave_spinning = 1;
2562         if (merge) {
2563                 struct btrfs_file_extent_item *fi;
2564                 u64 extent_len;
2565                 struct btrfs_key found_key;
2566
2567                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2568                 if (ret < 0)
2569                         goto out_free_path;
2570
2571                 path->slots[0]--;
2572                 leaf = path->nodes[0];
2573                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2574
2575                 fi = btrfs_item_ptr(leaf, path->slots[0],
2576                                     struct btrfs_file_extent_item);
2577                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2578
2579                 if (extent_len + found_key.offset == start &&
2580                     relink_is_mergable(leaf, fi, new)) {
2581                         btrfs_set_file_extent_num_bytes(leaf, fi,
2582                                                         extent_len + len);
2583                         btrfs_mark_buffer_dirty(leaf);
2584                         inode_add_bytes(inode, len);
2585
2586                         ret = 1;
2587                         goto out_free_path;
2588                 } else {
2589                         merge = false;
2590                         btrfs_release_path(path);
2591                         goto again;
2592                 }
2593         }
2594
2595         ret = btrfs_insert_empty_item(trans, root, path, &key,
2596                                         sizeof(*extent));
2597         if (ret) {
2598                 btrfs_abort_transaction(trans, root, ret);
2599                 goto out_free_path;
2600         }
2601
2602         leaf = path->nodes[0];
2603         item = btrfs_item_ptr(leaf, path->slots[0],
2604                                 struct btrfs_file_extent_item);
2605         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2606         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2607         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2608         btrfs_set_file_extent_num_bytes(leaf, item, len);
2609         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2610         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2611         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2612         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2613         btrfs_set_file_extent_encryption(leaf, item, 0);
2614         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2615
2616         btrfs_mark_buffer_dirty(leaf);
2617         inode_add_bytes(inode, len);
2618         btrfs_release_path(path);
2619
2620         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2621                         new->disk_len, 0,
2622                         backref->root_id, backref->inum,
2623                         new->file_pos); /* start - extent_offset */
2624         if (ret) {
2625                 btrfs_abort_transaction(trans, root, ret);
2626                 goto out_free_path;
2627         }
2628
2629         ret = 1;
2630 out_free_path:
2631         btrfs_release_path(path);
2632         path->leave_spinning = 0;
2633         btrfs_end_transaction(trans, root);
2634 out_unlock:
2635         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2636                              &cached, GFP_NOFS);
2637         iput(inode);
2638         return ret;
2639 }
2640
2641 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2642 {
2643         struct old_sa_defrag_extent *old, *tmp;
2644
2645         if (!new)
2646                 return;
2647
2648         list_for_each_entry_safe(old, tmp, &new->head, list) {
2649                 kfree(old);
2650         }
2651         kfree(new);
2652 }
2653
2654 static void relink_file_extents(struct new_sa_defrag_extent *new)
2655 {
2656         struct btrfs_path *path;
2657         struct sa_defrag_extent_backref *backref;
2658         struct sa_defrag_extent_backref *prev = NULL;
2659         struct inode *inode;
2660         struct btrfs_root *root;
2661         struct rb_node *node;
2662         int ret;
2663
2664         inode = new->inode;
2665         root = BTRFS_I(inode)->root;
2666
2667         path = btrfs_alloc_path();
2668         if (!path)
2669                 return;
2670
2671         if (!record_extent_backrefs(path, new)) {
2672                 btrfs_free_path(path);
2673                 goto out;
2674         }
2675         btrfs_release_path(path);
2676
2677         while (1) {
2678                 node = rb_first(&new->root);
2679                 if (!node)
2680                         break;
2681                 rb_erase(node, &new->root);
2682
2683                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2684
2685                 ret = relink_extent_backref(path, prev, backref);
2686                 WARN_ON(ret < 0);
2687
2688                 kfree(prev);
2689
2690                 if (ret == 1)
2691                         prev = backref;
2692                 else
2693                         prev = NULL;
2694                 cond_resched();
2695         }
2696         kfree(prev);
2697
2698         btrfs_free_path(path);
2699 out:
2700         free_sa_defrag_extent(new);
2701
2702         atomic_dec(&root->fs_info->defrag_running);
2703         wake_up(&root->fs_info->transaction_wait);
2704 }
2705
2706 static struct new_sa_defrag_extent *
2707 record_old_file_extents(struct inode *inode,
2708                         struct btrfs_ordered_extent *ordered)
2709 {
2710         struct btrfs_root *root = BTRFS_I(inode)->root;
2711         struct btrfs_path *path;
2712         struct btrfs_key key;
2713         struct old_sa_defrag_extent *old;
2714         struct new_sa_defrag_extent *new;
2715         int ret;
2716
2717         new = kmalloc(sizeof(*new), GFP_NOFS);
2718         if (!new)
2719                 return NULL;
2720
2721         new->inode = inode;
2722         new->file_pos = ordered->file_offset;
2723         new->len = ordered->len;
2724         new->bytenr = ordered->start;
2725         new->disk_len = ordered->disk_len;
2726         new->compress_type = ordered->compress_type;
2727         new->root = RB_ROOT;
2728         INIT_LIST_HEAD(&new->head);
2729
2730         path = btrfs_alloc_path();
2731         if (!path)
2732                 goto out_kfree;
2733
2734         key.objectid = btrfs_ino(inode);
2735         key.type = BTRFS_EXTENT_DATA_KEY;
2736         key.offset = new->file_pos;
2737
2738         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2739         if (ret < 0)
2740                 goto out_free_path;
2741         if (ret > 0 && path->slots[0] > 0)
2742                 path->slots[0]--;
2743
2744         /* find out all the old extents for the file range */
2745         while (1) {
2746                 struct btrfs_file_extent_item *extent;
2747                 struct extent_buffer *l;
2748                 int slot;
2749                 u64 num_bytes;
2750                 u64 offset;
2751                 u64 end;
2752                 u64 disk_bytenr;
2753                 u64 extent_offset;
2754
2755                 l = path->nodes[0];
2756                 slot = path->slots[0];
2757
2758                 if (slot >= btrfs_header_nritems(l)) {
2759                         ret = btrfs_next_leaf(root, path);
2760                         if (ret < 0)
2761                                 goto out_free_path;
2762                         else if (ret > 0)
2763                                 break;
2764                         continue;
2765                 }
2766
2767                 btrfs_item_key_to_cpu(l, &key, slot);
2768
2769                 if (key.objectid != btrfs_ino(inode))
2770                         break;
2771                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2772                         break;
2773                 if (key.offset >= new->file_pos + new->len)
2774                         break;
2775
2776                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2777
2778                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2779                 if (key.offset + num_bytes < new->file_pos)
2780                         goto next;
2781
2782                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2783                 if (!disk_bytenr)
2784                         goto next;
2785
2786                 extent_offset = btrfs_file_extent_offset(l, extent);
2787
2788                 old = kmalloc(sizeof(*old), GFP_NOFS);
2789                 if (!old)
2790                         goto out_free_path;
2791
2792                 offset = max(new->file_pos, key.offset);
2793                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2794
2795                 old->bytenr = disk_bytenr;
2796                 old->extent_offset = extent_offset;
2797                 old->offset = offset - key.offset;
2798                 old->len = end - offset;
2799                 old->new = new;
2800                 old->count = 0;
2801                 list_add_tail(&old->list, &new->head);
2802 next:
2803                 path->slots[0]++;
2804                 cond_resched();
2805         }
2806
2807         btrfs_free_path(path);
2808         atomic_inc(&root->fs_info->defrag_running);
2809
2810         return new;
2811
2812 out_free_path:
2813         btrfs_free_path(path);
2814 out_kfree:
2815         free_sa_defrag_extent(new);
2816         return NULL;
2817 }
2818
2819 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2820                                          u64 start, u64 len)
2821 {
2822         struct btrfs_block_group_cache *cache;
2823
2824         cache = btrfs_lookup_block_group(root->fs_info, start);
2825         ASSERT(cache);
2826
2827         spin_lock(&cache->lock);
2828         cache->delalloc_bytes -= len;
2829         spin_unlock(&cache->lock);
2830
2831         btrfs_put_block_group(cache);
2832 }
2833
2834 /* as ordered data IO finishes, this gets called so we can finish
2835  * an ordered extent if the range of bytes in the file it covers are
2836  * fully written.
2837  */
2838 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2839 {
2840         struct inode *inode = ordered_extent->inode;
2841         struct btrfs_root *root = BTRFS_I(inode)->root;
2842         struct btrfs_trans_handle *trans = NULL;
2843         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2844         struct extent_state *cached_state = NULL;
2845         struct new_sa_defrag_extent *new = NULL;
2846         int compress_type = 0;
2847         int ret = 0;
2848         u64 logical_len = ordered_extent->len;
2849         bool nolock;
2850         bool truncated = false;
2851
2852         nolock = btrfs_is_free_space_inode(inode);
2853
2854         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2855                 ret = -EIO;
2856                 goto out;
2857         }
2858
2859         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2860                                      ordered_extent->file_offset +
2861                                      ordered_extent->len - 1);
2862
2863         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2864                 truncated = true;
2865                 logical_len = ordered_extent->truncated_len;
2866                 /* Truncated the entire extent, don't bother adding */
2867                 if (!logical_len)
2868                         goto out;
2869         }
2870
2871         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2872                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2873
2874                 /*
2875                  * For mwrite(mmap + memset to write) case, we still reserve
2876                  * space for NOCOW range.
2877                  * As NOCOW won't cause a new delayed ref, just free the space
2878                  */
2879                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2880                                        ordered_extent->len);
2881                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2882                 if (nolock)
2883                         trans = btrfs_join_transaction_nolock(root);
2884                 else
2885                         trans = btrfs_join_transaction(root);
2886                 if (IS_ERR(trans)) {
2887                         ret = PTR_ERR(trans);
2888                         trans = NULL;
2889                         goto out;
2890                 }
2891                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2892                 ret = btrfs_update_inode_fallback(trans, root, inode);
2893                 if (ret) /* -ENOMEM or corruption */
2894                         btrfs_abort_transaction(trans, root, ret);
2895                 goto out;
2896         }
2897
2898         lock_extent_bits(io_tree, ordered_extent->file_offset,
2899                          ordered_extent->file_offset + ordered_extent->len - 1,
2900                          &cached_state);
2901
2902         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2903                         ordered_extent->file_offset + ordered_extent->len - 1,
2904                         EXTENT_DEFRAG, 1, cached_state);
2905         if (ret) {
2906                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2907                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2908                         /* the inode is shared */
2909                         new = record_old_file_extents(inode, ordered_extent);
2910
2911                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2912                         ordered_extent->file_offset + ordered_extent->len - 1,
2913                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2914         }
2915
2916         if (nolock)
2917                 trans = btrfs_join_transaction_nolock(root);
2918         else
2919                 trans = btrfs_join_transaction(root);
2920         if (IS_ERR(trans)) {
2921                 ret = PTR_ERR(trans);
2922                 trans = NULL;
2923                 goto out_unlock;
2924         }
2925
2926         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2927
2928         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2929                 compress_type = ordered_extent->compress_type;
2930         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2931                 BUG_ON(compress_type);
2932                 ret = btrfs_mark_extent_written(trans, inode,
2933                                                 ordered_extent->file_offset,
2934                                                 ordered_extent->file_offset +
2935                                                 logical_len);
2936         } else {
2937                 BUG_ON(root == root->fs_info->tree_root);
2938                 ret = insert_reserved_file_extent(trans, inode,
2939                                                 ordered_extent->file_offset,
2940                                                 ordered_extent->start,
2941                                                 ordered_extent->disk_len,
2942                                                 logical_len, logical_len,
2943                                                 compress_type, 0, 0,
2944                                                 BTRFS_FILE_EXTENT_REG);
2945                 if (!ret)
2946                         btrfs_release_delalloc_bytes(root,
2947                                                      ordered_extent->start,
2948                                                      ordered_extent->disk_len);
2949         }
2950         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2951                            ordered_extent->file_offset, ordered_extent->len,
2952                            trans->transid);
2953         if (ret < 0) {
2954                 btrfs_abort_transaction(trans, root, ret);
2955                 goto out_unlock;
2956         }
2957
2958         add_pending_csums(trans, inode, ordered_extent->file_offset,
2959                           &ordered_extent->list);
2960
2961         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2962         ret = btrfs_update_inode_fallback(trans, root, inode);
2963         if (ret) { /* -ENOMEM or corruption */
2964                 btrfs_abort_transaction(trans, root, ret);
2965                 goto out_unlock;
2966         }
2967         ret = 0;
2968 out_unlock:
2969         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2970                              ordered_extent->file_offset +
2971                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2972 out:
2973         if (root != root->fs_info->tree_root)
2974                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2975         if (trans)
2976                 btrfs_end_transaction(trans, root);
2977
2978         if (ret || truncated) {
2979                 u64 start, end;
2980
2981                 if (truncated)
2982                         start = ordered_extent->file_offset + logical_len;
2983                 else
2984                         start = ordered_extent->file_offset;
2985                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2986                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2987
2988                 /* Drop the cache for the part of the extent we didn't write. */
2989                 btrfs_drop_extent_cache(inode, start, end, 0);
2990
2991                 /*
2992                  * If the ordered extent had an IOERR or something else went
2993                  * wrong we need to return the space for this ordered extent
2994                  * back to the allocator.  We only free the extent in the
2995                  * truncated case if we didn't write out the extent at all.
2996                  */
2997                 if ((ret || !logical_len) &&
2998                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2999                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3000                         btrfs_free_reserved_extent(root, ordered_extent->start,
3001                                                    ordered_extent->disk_len, 1);
3002         }
3003
3004
3005         /*
3006          * This needs to be done to make sure anybody waiting knows we are done
3007          * updating everything for this ordered extent.
3008          */
3009         btrfs_remove_ordered_extent(inode, ordered_extent);
3010
3011         /* for snapshot-aware defrag */
3012         if (new) {
3013                 if (ret) {
3014                         free_sa_defrag_extent(new);
3015                         atomic_dec(&root->fs_info->defrag_running);
3016                 } else {
3017                         relink_file_extents(new);
3018                 }
3019         }
3020
3021         /* once for us */
3022         btrfs_put_ordered_extent(ordered_extent);
3023         /* once for the tree */
3024         btrfs_put_ordered_extent(ordered_extent);
3025
3026         return ret;
3027 }
3028
3029 static void finish_ordered_fn(struct btrfs_work *work)
3030 {
3031         struct btrfs_ordered_extent *ordered_extent;
3032         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3033         btrfs_finish_ordered_io(ordered_extent);
3034 }
3035
3036 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3037                                 struct extent_state *state, int uptodate)
3038 {
3039         struct inode *inode = page->mapping->host;
3040         struct btrfs_root *root = BTRFS_I(inode)->root;
3041         struct btrfs_ordered_extent *ordered_extent = NULL;
3042         struct btrfs_workqueue *wq;
3043         btrfs_work_func_t func;
3044
3045         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3046
3047         ClearPagePrivate2(page);
3048         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3049                                             end - start + 1, uptodate))
3050                 return 0;
3051
3052         if (btrfs_is_free_space_inode(inode)) {
3053                 wq = root->fs_info->endio_freespace_worker;
3054                 func = btrfs_freespace_write_helper;
3055         } else {
3056                 wq = root->fs_info->endio_write_workers;
3057                 func = btrfs_endio_write_helper;
3058         }
3059
3060         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3061                         NULL);
3062         btrfs_queue_work(wq, &ordered_extent->work);
3063
3064         return 0;
3065 }
3066
3067 static int __readpage_endio_check(struct inode *inode,
3068                                   struct btrfs_io_bio *io_bio,
3069                                   int icsum, struct page *page,
3070                                   int pgoff, u64 start, size_t len)
3071 {
3072         char *kaddr;
3073         u32 csum_expected;
3074         u32 csum = ~(u32)0;
3075
3076         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3077
3078         kaddr = kmap_atomic(page);
3079         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3080         btrfs_csum_final(csum, (char *)&csum);
3081         if (csum != csum_expected)
3082                 goto zeroit;
3083
3084         kunmap_atomic(kaddr);
3085         return 0;
3086 zeroit:
3087         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3088                 "csum failed ino %llu off %llu csum %u expected csum %u",
3089                            btrfs_ino(inode), start, csum, csum_expected);
3090         memset(kaddr + pgoff, 1, len);
3091         flush_dcache_page(page);
3092         kunmap_atomic(kaddr);
3093         if (csum_expected == 0)
3094                 return 0;
3095         return -EIO;
3096 }
3097
3098 /*
3099  * when reads are done, we need to check csums to verify the data is correct
3100  * if there's a match, we allow the bio to finish.  If not, the code in
3101  * extent_io.c will try to find good copies for us.
3102  */
3103 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3104                                       u64 phy_offset, struct page *page,
3105                                       u64 start, u64 end, int mirror)
3106 {
3107         size_t offset = start - page_offset(page);
3108         struct inode *inode = page->mapping->host;
3109         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3110         struct btrfs_root *root = BTRFS_I(inode)->root;
3111
3112         if (PageChecked(page)) {
3113                 ClearPageChecked(page);
3114                 return 0;
3115         }
3116
3117         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3118                 return 0;
3119
3120         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3121             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3122                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3123                 return 0;
3124         }
3125
3126         phy_offset >>= inode->i_sb->s_blocksize_bits;
3127         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3128                                       start, (size_t)(end - start + 1));
3129 }
3130
3131 void btrfs_add_delayed_iput(struct inode *inode)
3132 {
3133         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3134         struct btrfs_inode *binode = BTRFS_I(inode);
3135
3136         if (atomic_add_unless(&inode->i_count, -1, 1))
3137                 return;
3138
3139         spin_lock(&fs_info->delayed_iput_lock);
3140         if (binode->delayed_iput_count == 0) {
3141                 ASSERT(list_empty(&binode->delayed_iput));
3142                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3143         } else {
3144                 binode->delayed_iput_count++;
3145         }
3146         spin_unlock(&fs_info->delayed_iput_lock);
3147 }
3148
3149 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3150 {
3151         struct btrfs_fs_info *fs_info = root->fs_info;
3152
3153         spin_lock(&fs_info->delayed_iput_lock);
3154         while (!list_empty(&fs_info->delayed_iputs)) {
3155                 struct btrfs_inode *inode;
3156
3157                 inode = list_first_entry(&fs_info->delayed_iputs,
3158                                 struct btrfs_inode, delayed_iput);
3159                 if (inode->delayed_iput_count) {
3160                         inode->delayed_iput_count--;
3161                         list_move_tail(&inode->delayed_iput,
3162                                         &fs_info->delayed_iputs);
3163                 } else {
3164                         list_del_init(&inode->delayed_iput);
3165                 }
3166                 spin_unlock(&fs_info->delayed_iput_lock);
3167                 iput(&inode->vfs_inode);
3168                 spin_lock(&fs_info->delayed_iput_lock);
3169         }
3170         spin_unlock(&fs_info->delayed_iput_lock);
3171 }
3172
3173 /*
3174  * This is called in transaction commit time. If there are no orphan
3175  * files in the subvolume, it removes orphan item and frees block_rsv
3176  * structure.
3177  */
3178 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3179                               struct btrfs_root *root)
3180 {
3181         struct btrfs_block_rsv *block_rsv;
3182         int ret;
3183
3184         if (atomic_read(&root->orphan_inodes) ||
3185             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3186                 return;
3187
3188         spin_lock(&root->orphan_lock);
3189         if (atomic_read(&root->orphan_inodes)) {
3190                 spin_unlock(&root->orphan_lock);
3191                 return;
3192         }
3193
3194         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3195                 spin_unlock(&root->orphan_lock);
3196                 return;
3197         }
3198
3199         block_rsv = root->orphan_block_rsv;
3200         root->orphan_block_rsv = NULL;
3201         spin_unlock(&root->orphan_lock);
3202
3203         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3204             btrfs_root_refs(&root->root_item) > 0) {
3205                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3206                                             root->root_key.objectid);
3207                 if (ret)
3208                         btrfs_abort_transaction(trans, root, ret);
3209                 else
3210                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3211                                   &root->state);
3212         }
3213
3214         if (block_rsv) {
3215                 WARN_ON(block_rsv->size > 0);
3216                 btrfs_free_block_rsv(root, block_rsv);
3217         }
3218 }
3219
3220 /*
3221  * This creates an orphan entry for the given inode in case something goes
3222  * wrong in the middle of an unlink/truncate.
3223  *
3224  * NOTE: caller of this function should reserve 5 units of metadata for
3225  *       this function.
3226  */
3227 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3228 {
3229         struct btrfs_root *root = BTRFS_I(inode)->root;
3230         struct btrfs_block_rsv *block_rsv = NULL;
3231         int reserve = 0;
3232         int insert = 0;
3233         int ret;
3234
3235         if (!root->orphan_block_rsv) {
3236                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3237                 if (!block_rsv)
3238                         return -ENOMEM;
3239         }
3240
3241         spin_lock(&root->orphan_lock);
3242         if (!root->orphan_block_rsv) {
3243                 root->orphan_block_rsv = block_rsv;
3244         } else if (block_rsv) {
3245                 btrfs_free_block_rsv(root, block_rsv);
3246                 block_rsv = NULL;
3247         }
3248
3249         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3250                               &BTRFS_I(inode)->runtime_flags)) {
3251 #if 0
3252                 /*
3253                  * For proper ENOSPC handling, we should do orphan
3254                  * cleanup when mounting. But this introduces backward
3255                  * compatibility issue.
3256                  */
3257                 if (!xchg(&root->orphan_item_inserted, 1))
3258                         insert = 2;
3259                 else
3260                         insert = 1;
3261 #endif
3262                 insert = 1;
3263                 atomic_inc(&root->orphan_inodes);
3264         }
3265
3266         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3267                               &BTRFS_I(inode)->runtime_flags))
3268                 reserve = 1;
3269         spin_unlock(&root->orphan_lock);
3270
3271         /* grab metadata reservation from transaction handle */
3272         if (reserve) {
3273                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3274                 ASSERT(!ret);
3275                 if (ret) {
3276                         atomic_dec(&root->orphan_inodes);
3277                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3278                                   &BTRFS_I(inode)->runtime_flags);
3279                         if (insert)
3280                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3281                                           &BTRFS_I(inode)->runtime_flags);
3282                         return ret;
3283                 }
3284         }
3285
3286         /* insert an orphan item to track this unlinked/truncated file */
3287         if (insert >= 1) {
3288                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3289                 if (ret) {
3290                         atomic_dec(&root->orphan_inodes);
3291                         if (reserve) {
3292                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3293                                           &BTRFS_I(inode)->runtime_flags);
3294                                 btrfs_orphan_release_metadata(inode);
3295                         }
3296                         if (ret != -EEXIST) {
3297                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3298                                           &BTRFS_I(inode)->runtime_flags);
3299                                 btrfs_abort_transaction(trans, root, ret);
3300                                 return ret;
3301                         }
3302                 }
3303                 ret = 0;
3304         }
3305
3306         /* insert an orphan item to track subvolume contains orphan files */
3307         if (insert >= 2) {
3308                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3309                                                root->root_key.objectid);
3310                 if (ret && ret != -EEXIST) {
3311                         btrfs_abort_transaction(trans, root, ret);
3312                         return ret;
3313                 }
3314         }
3315         return 0;
3316 }
3317
3318 /*
3319  * We have done the truncate/delete so we can go ahead and remove the orphan
3320  * item for this particular inode.
3321  */
3322 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3323                             struct inode *inode)
3324 {
3325         struct btrfs_root *root = BTRFS_I(inode)->root;
3326         int delete_item = 0;
3327         int release_rsv = 0;
3328         int ret = 0;
3329
3330         spin_lock(&root->orphan_lock);
3331         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3332                                &BTRFS_I(inode)->runtime_flags))
3333                 delete_item = 1;
3334
3335         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3336                                &BTRFS_I(inode)->runtime_flags))
3337                 release_rsv = 1;
3338         spin_unlock(&root->orphan_lock);
3339
3340         if (delete_item) {
3341                 atomic_dec(&root->orphan_inodes);
3342                 if (trans)
3343                         ret = btrfs_del_orphan_item(trans, root,
3344                                                     btrfs_ino(inode));
3345         }
3346
3347         if (release_rsv)
3348                 btrfs_orphan_release_metadata(inode);
3349
3350         return ret;
3351 }
3352
3353 /*
3354  * this cleans up any orphans that may be left on the list from the last use
3355  * of this root.
3356  */
3357 int btrfs_orphan_cleanup(struct btrfs_root *root)
3358 {
3359         struct btrfs_path *path;
3360         struct extent_buffer *leaf;
3361         struct btrfs_key key, found_key;
3362         struct btrfs_trans_handle *trans;
3363         struct inode *inode;
3364         u64 last_objectid = 0;
3365         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3366
3367         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3368                 return 0;
3369
3370         path = btrfs_alloc_path();
3371         if (!path) {
3372                 ret = -ENOMEM;
3373                 goto out;
3374         }
3375         path->reada = READA_BACK;
3376
3377         key.objectid = BTRFS_ORPHAN_OBJECTID;
3378         key.type = BTRFS_ORPHAN_ITEM_KEY;
3379         key.offset = (u64)-1;
3380
3381         while (1) {
3382                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3383                 if (ret < 0)
3384                         goto out;
3385
3386                 /*
3387                  * if ret == 0 means we found what we were searching for, which
3388                  * is weird, but possible, so only screw with path if we didn't
3389                  * find the key and see if we have stuff that matches
3390                  */
3391                 if (ret > 0) {
3392                         ret = 0;
3393                         if (path->slots[0] == 0)
3394                                 break;
3395                         path->slots[0]--;
3396                 }
3397
3398                 /* pull out the item */
3399                 leaf = path->nodes[0];
3400                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3401
3402                 /* make sure the item matches what we want */
3403                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3404                         break;
3405                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3406                         break;
3407
3408                 /* release the path since we're done with it */
3409                 btrfs_release_path(path);
3410
3411                 /*
3412                  * this is where we are basically btrfs_lookup, without the
3413                  * crossing root thing.  we store the inode number in the
3414                  * offset of the orphan item.
3415                  */
3416
3417                 if (found_key.offset == last_objectid) {
3418                         btrfs_err(root->fs_info,
3419                                 "Error removing orphan entry, stopping orphan cleanup");
3420                         ret = -EINVAL;
3421                         goto out;
3422                 }
3423
3424                 last_objectid = found_key.offset;
3425
3426                 found_key.objectid = found_key.offset;
3427                 found_key.type = BTRFS_INODE_ITEM_KEY;
3428                 found_key.offset = 0;
3429                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3430                 ret = PTR_ERR_OR_ZERO(inode);
3431                 if (ret && ret != -ESTALE)
3432                         goto out;
3433
3434                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3435                         struct btrfs_root *dead_root;
3436                         struct btrfs_fs_info *fs_info = root->fs_info;
3437                         int is_dead_root = 0;
3438
3439                         /*
3440                          * this is an orphan in the tree root. Currently these
3441                          * could come from 2 sources:
3442                          *  a) a snapshot deletion in progress
3443                          *  b) a free space cache inode
3444                          * We need to distinguish those two, as the snapshot
3445                          * orphan must not get deleted.
3446                          * find_dead_roots already ran before us, so if this
3447                          * is a snapshot deletion, we should find the root
3448                          * in the dead_roots list
3449                          */
3450                         spin_lock(&fs_info->trans_lock);
3451                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3452                                             root_list) {
3453                                 if (dead_root->root_key.objectid ==
3454                                     found_key.objectid) {
3455                                         is_dead_root = 1;
3456                                         break;
3457                                 }
3458                         }
3459                         spin_unlock(&fs_info->trans_lock);
3460                         if (is_dead_root) {
3461                                 /* prevent this orphan from being found again */
3462                                 key.offset = found_key.objectid - 1;
3463                                 continue;
3464                         }
3465                 }
3466                 /*
3467                  * Inode is already gone but the orphan item is still there,
3468                  * kill the orphan item.
3469                  */
3470                 if (ret == -ESTALE) {
3471                         trans = btrfs_start_transaction(root, 1);
3472                         if (IS_ERR(trans)) {
3473                                 ret = PTR_ERR(trans);
3474                                 goto out;
3475                         }
3476                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3477                                 found_key.objectid);
3478                         ret = btrfs_del_orphan_item(trans, root,
3479                                                     found_key.objectid);
3480                         btrfs_end_transaction(trans, root);
3481                         if (ret)
3482                                 goto out;
3483                         continue;
3484                 }
3485
3486                 /*
3487                  * add this inode to the orphan list so btrfs_orphan_del does
3488                  * the proper thing when we hit it
3489                  */
3490                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3491                         &BTRFS_I(inode)->runtime_flags);
3492                 atomic_inc(&root->orphan_inodes);
3493
3494                 /* if we have links, this was a truncate, lets do that */
3495                 if (inode->i_nlink) {
3496                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3497                                 iput(inode);
3498                                 continue;
3499                         }
3500                         nr_truncate++;
3501
3502                         /* 1 for the orphan item deletion. */
3503                         trans = btrfs_start_transaction(root, 1);
3504                         if (IS_ERR(trans)) {
3505                                 iput(inode);
3506                                 ret = PTR_ERR(trans);
3507                                 goto out;
3508                         }
3509                         ret = btrfs_orphan_add(trans, inode);
3510                         btrfs_end_transaction(trans, root);
3511                         if (ret) {
3512                                 iput(inode);
3513                                 goto out;
3514                         }
3515
3516                         ret = btrfs_truncate(inode);
3517                         if (ret)
3518                                 btrfs_orphan_del(NULL, inode);
3519                 } else {
3520                         nr_unlink++;
3521                 }
3522
3523                 /* this will do delete_inode and everything for us */
3524                 iput(inode);
3525                 if (ret)
3526                         goto out;
3527         }
3528         /* release the path since we're done with it */
3529         btrfs_release_path(path);
3530
3531         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3532
3533         if (root->orphan_block_rsv)
3534                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3535                                         (u64)-1);
3536
3537         if (root->orphan_block_rsv ||
3538             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3539                 trans = btrfs_join_transaction(root);
3540                 if (!IS_ERR(trans))
3541                         btrfs_end_transaction(trans, root);
3542         }
3543
3544         if (nr_unlink)
3545                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3546         if (nr_truncate)
3547                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3548
3549 out:
3550         if (ret)
3551                 btrfs_err(root->fs_info,
3552                         "could not do orphan cleanup %d", ret);
3553         btrfs_free_path(path);
3554         return ret;
3555 }
3556
3557 /*
3558  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3559  * don't find any xattrs, we know there can't be any acls.
3560  *
3561  * slot is the slot the inode is in, objectid is the objectid of the inode
3562  */
3563 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3564                                           int slot, u64 objectid,
3565                                           int *first_xattr_slot)
3566 {
3567         u32 nritems = btrfs_header_nritems(leaf);
3568         struct btrfs_key found_key;
3569         static u64 xattr_access = 0;
3570         static u64 xattr_default = 0;
3571         int scanned = 0;
3572
3573         if (!xattr_access) {
3574                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3575                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3576                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3577                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3578         }
3579
3580         slot++;
3581         *first_xattr_slot = -1;
3582         while (slot < nritems) {
3583                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3584
3585                 /* we found a different objectid, there must not be acls */
3586                 if (found_key.objectid != objectid)
3587                         return 0;
3588
3589                 /* we found an xattr, assume we've got an acl */
3590                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3591                         if (*first_xattr_slot == -1)
3592                                 *first_xattr_slot = slot;
3593                         if (found_key.offset == xattr_access ||
3594                             found_key.offset == xattr_default)
3595                                 return 1;
3596                 }
3597
3598                 /*
3599                  * we found a key greater than an xattr key, there can't
3600                  * be any acls later on
3601                  */
3602                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3603                         return 0;
3604
3605                 slot++;
3606                 scanned++;
3607
3608                 /*
3609                  * it goes inode, inode backrefs, xattrs, extents,
3610                  * so if there are a ton of hard links to an inode there can
3611                  * be a lot of backrefs.  Don't waste time searching too hard,
3612                  * this is just an optimization
3613                  */
3614                 if (scanned >= 8)
3615                         break;
3616         }
3617         /* we hit the end of the leaf before we found an xattr or
3618          * something larger than an xattr.  We have to assume the inode
3619          * has acls
3620          */
3621         if (*first_xattr_slot == -1)
3622                 *first_xattr_slot = slot;
3623         return 1;
3624 }
3625
3626 /*
3627  * read an inode from the btree into the in-memory inode
3628  */
3629 static void btrfs_read_locked_inode(struct inode *inode)
3630 {
3631         struct btrfs_path *path;
3632         struct extent_buffer *leaf;
3633         struct btrfs_inode_item *inode_item;
3634         struct btrfs_root *root = BTRFS_I(inode)->root;
3635         struct btrfs_key location;
3636         unsigned long ptr;
3637         int maybe_acls;
3638         u32 rdev;
3639         int ret;
3640         bool filled = false;
3641         int first_xattr_slot;
3642
3643         ret = btrfs_fill_inode(inode, &rdev);
3644         if (!ret)
3645                 filled = true;
3646
3647         path = btrfs_alloc_path();
3648         if (!path)
3649                 goto make_bad;
3650
3651         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3652
3653         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3654         if (ret)
3655                 goto make_bad;
3656
3657         leaf = path->nodes[0];
3658
3659         if (filled)
3660                 goto cache_index;
3661
3662         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3663                                     struct btrfs_inode_item);
3664         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3665         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3666         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3667         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3668         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3669
3670         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3671         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3672
3673         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3674         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3675
3676         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3677         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3678
3679         BTRFS_I(inode)->i_otime.tv_sec =
3680                 btrfs_timespec_sec(leaf, &inode_item->otime);
3681         BTRFS_I(inode)->i_otime.tv_nsec =
3682                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3683
3684         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3685         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3686         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3687
3688         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3689         inode->i_generation = BTRFS_I(inode)->generation;
3690         inode->i_rdev = 0;
3691         rdev = btrfs_inode_rdev(leaf, inode_item);
3692
3693         BTRFS_I(inode)->index_cnt = (u64)-1;
3694         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3695
3696 cache_index:
3697         /*
3698          * If we were modified in the current generation and evicted from memory
3699          * and then re-read we need to do a full sync since we don't have any
3700          * idea about which extents were modified before we were evicted from
3701          * cache.
3702          *
3703          * This is required for both inode re-read from disk and delayed inode
3704          * in delayed_nodes_tree.
3705          */
3706         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3707                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3708                         &BTRFS_I(inode)->runtime_flags);
3709
3710         /*
3711          * We don't persist the id of the transaction where an unlink operation
3712          * against the inode was last made. So here we assume the inode might
3713          * have been evicted, and therefore the exact value of last_unlink_trans
3714          * lost, and set it to last_trans to avoid metadata inconsistencies
3715          * between the inode and its parent if the inode is fsync'ed and the log
3716          * replayed. For example, in the scenario:
3717          *
3718          * touch mydir/foo
3719          * ln mydir/foo mydir/bar
3720          * sync
3721          * unlink mydir/bar
3722          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3723          * xfs_io -c fsync mydir/foo
3724          * <power failure>
3725          * mount fs, triggers fsync log replay
3726          *
3727          * We must make sure that when we fsync our inode foo we also log its
3728          * parent inode, otherwise after log replay the parent still has the
3729          * dentry with the "bar" name but our inode foo has a link count of 1
3730          * and doesn't have an inode ref with the name "bar" anymore.
3731          *
3732          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3733          * but it guarantees correctness at the expense of occasional full
3734          * transaction commits on fsync if our inode is a directory, or if our
3735          * inode is not a directory, logging its parent unnecessarily.
3736          */
3737         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3738
3739         path->slots[0]++;
3740         if (inode->i_nlink != 1 ||
3741             path->slots[0] >= btrfs_header_nritems(leaf))
3742                 goto cache_acl;
3743
3744         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3745         if (location.objectid != btrfs_ino(inode))
3746                 goto cache_acl;
3747
3748         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3749         if (location.type == BTRFS_INODE_REF_KEY) {
3750                 struct btrfs_inode_ref *ref;
3751
3752                 ref = (struct btrfs_inode_ref *)ptr;
3753                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3754         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3755                 struct btrfs_inode_extref *extref;
3756
3757                 extref = (struct btrfs_inode_extref *)ptr;
3758                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3759                                                                      extref);
3760         }
3761 cache_acl:
3762         /*
3763          * try to precache a NULL acl entry for files that don't have
3764          * any xattrs or acls
3765          */
3766         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3767                                            btrfs_ino(inode), &first_xattr_slot);
3768         if (first_xattr_slot != -1) {
3769                 path->slots[0] = first_xattr_slot;
3770                 ret = btrfs_load_inode_props(inode, path);
3771                 if (ret)
3772                         btrfs_err(root->fs_info,
3773                                   "error loading props for ino %llu (root %llu): %d",
3774                                   btrfs_ino(inode),
3775                                   root->root_key.objectid, ret);
3776         }
3777         btrfs_free_path(path);
3778
3779         if (!maybe_acls)
3780                 cache_no_acl(inode);
3781
3782         switch (inode->i_mode & S_IFMT) {
3783         case S_IFREG:
3784                 inode->i_mapping->a_ops = &btrfs_aops;
3785                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3786                 inode->i_fop = &btrfs_file_operations;
3787                 inode->i_op = &btrfs_file_inode_operations;
3788                 break;
3789         case S_IFDIR:
3790                 inode->i_fop = &btrfs_dir_file_operations;
3791                 if (root == root->fs_info->tree_root)
3792                         inode->i_op = &btrfs_dir_ro_inode_operations;
3793                 else
3794                         inode->i_op = &btrfs_dir_inode_operations;
3795                 break;
3796         case S_IFLNK:
3797                 inode->i_op = &btrfs_symlink_inode_operations;
3798                 inode_nohighmem(inode);
3799                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3800                 break;
3801         default:
3802                 inode->i_op = &btrfs_special_inode_operations;
3803                 init_special_inode(inode, inode->i_mode, rdev);
3804                 break;
3805         }
3806
3807         btrfs_update_iflags(inode);
3808         return;
3809
3810 make_bad:
3811         btrfs_free_path(path);
3812         make_bad_inode(inode);
3813 }
3814
3815 /*
3816  * given a leaf and an inode, copy the inode fields into the leaf
3817  */
3818 static void fill_inode_item(struct btrfs_trans_handle *trans,
3819                             struct extent_buffer *leaf,
3820                             struct btrfs_inode_item *item,
3821                             struct inode *inode)
3822 {
3823         struct btrfs_map_token token;
3824
3825         btrfs_init_map_token(&token);
3826
3827         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3828         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3829         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3830                                    &token);
3831         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3832         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3833
3834         btrfs_set_token_timespec_sec(leaf, &item->atime,
3835                                      inode->i_atime.tv_sec, &token);
3836         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3837                                       inode->i_atime.tv_nsec, &token);
3838
3839         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3840                                      inode->i_mtime.tv_sec, &token);
3841         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3842                                       inode->i_mtime.tv_nsec, &token);
3843
3844         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3845                                      inode->i_ctime.tv_sec, &token);
3846         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3847                                       inode->i_ctime.tv_nsec, &token);
3848
3849         btrfs_set_token_timespec_sec(leaf, &item->otime,
3850                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3851         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3852                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3853
3854         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3855                                      &token);
3856         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3857                                          &token);
3858         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3859         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3860         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3861         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3862         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3863 }
3864
3865 /*
3866  * copy everything in the in-memory inode into the btree.
3867  */
3868 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3869                                 struct btrfs_root *root, struct inode *inode)
3870 {
3871         struct btrfs_inode_item *inode_item;
3872         struct btrfs_path *path;
3873         struct extent_buffer *leaf;
3874         int ret;
3875
3876         path = btrfs_alloc_path();
3877         if (!path)
3878                 return -ENOMEM;
3879
3880         path->leave_spinning = 1;
3881         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3882                                  1);
3883         if (ret) {
3884                 if (ret > 0)
3885                         ret = -ENOENT;
3886                 goto failed;
3887         }
3888
3889         leaf = path->nodes[0];
3890         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3891                                     struct btrfs_inode_item);
3892
3893         fill_inode_item(trans, leaf, inode_item, inode);
3894         btrfs_mark_buffer_dirty(leaf);
3895         btrfs_set_inode_last_trans(trans, inode);
3896         ret = 0;
3897 failed:
3898         btrfs_free_path(path);
3899         return ret;
3900 }
3901
3902 /*
3903  * copy everything in the in-memory inode into the btree.
3904  */
3905 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3906                                 struct btrfs_root *root, struct inode *inode)
3907 {
3908         int ret;
3909
3910         /*
3911          * If the inode is a free space inode, we can deadlock during commit
3912          * if we put it into the delayed code.
3913          *
3914          * The data relocation inode should also be directly updated
3915          * without delay
3916          */
3917         if (!btrfs_is_free_space_inode(inode)
3918             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3919             && !root->fs_info->log_root_recovering) {
3920                 btrfs_update_root_times(trans, root);
3921
3922                 ret = btrfs_delayed_update_inode(trans, root, inode);
3923                 if (!ret)
3924                         btrfs_set_inode_last_trans(trans, inode);
3925                 return ret;
3926         }
3927
3928         return btrfs_update_inode_item(trans, root, inode);
3929 }
3930
3931 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3932                                          struct btrfs_root *root,
3933                                          struct inode *inode)
3934 {
3935         int ret;
3936
3937         ret = btrfs_update_inode(trans, root, inode);
3938         if (ret == -ENOSPC)
3939                 return btrfs_update_inode_item(trans, root, inode);
3940         return ret;
3941 }
3942
3943 /*
3944  * unlink helper that gets used here in inode.c and in the tree logging
3945  * recovery code.  It remove a link in a directory with a given name, and
3946  * also drops the back refs in the inode to the directory
3947  */
3948 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3949                                 struct btrfs_root *root,
3950                                 struct inode *dir, struct inode *inode,
3951                                 const char *name, int name_len)
3952 {
3953         struct btrfs_path *path;
3954         int ret = 0;
3955         struct extent_buffer *leaf;
3956         struct btrfs_dir_item *di;
3957         struct btrfs_key key;
3958         u64 index;
3959         u64 ino = btrfs_ino(inode);
3960         u64 dir_ino = btrfs_ino(dir);
3961
3962         path = btrfs_alloc_path();
3963         if (!path) {
3964                 ret = -ENOMEM;
3965                 goto out;
3966         }
3967
3968         path->leave_spinning = 1;
3969         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3970                                     name, name_len, -1);
3971         if (IS_ERR(di)) {
3972                 ret = PTR_ERR(di);
3973                 goto err;
3974         }
3975         if (!di) {
3976                 ret = -ENOENT;
3977                 goto err;
3978         }
3979         leaf = path->nodes[0];
3980         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3981         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3982         if (ret)
3983                 goto err;
3984         btrfs_release_path(path);
3985
3986         /*
3987          * If we don't have dir index, we have to get it by looking up
3988          * the inode ref, since we get the inode ref, remove it directly,
3989          * it is unnecessary to do delayed deletion.
3990          *
3991          * But if we have dir index, needn't search inode ref to get it.
3992          * Since the inode ref is close to the inode item, it is better
3993          * that we delay to delete it, and just do this deletion when
3994          * we update the inode item.
3995          */
3996         if (BTRFS_I(inode)->dir_index) {
3997                 ret = btrfs_delayed_delete_inode_ref(inode);
3998                 if (!ret) {
3999                         index = BTRFS_I(inode)->dir_index;
4000                         goto skip_backref;
4001                 }
4002         }
4003
4004         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4005                                   dir_ino, &index);
4006         if (ret) {
4007                 btrfs_info(root->fs_info,
4008                         "failed to delete reference to %.*s, inode %llu parent %llu",
4009                         name_len, name, ino, dir_ino);
4010                 btrfs_abort_transaction(trans, root, ret);
4011                 goto err;
4012         }
4013 skip_backref:
4014         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4015         if (ret) {
4016                 btrfs_abort_transaction(trans, root, ret);
4017                 goto err;
4018         }
4019
4020         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
4021                                          inode, dir_ino);
4022         if (ret != 0 && ret != -ENOENT) {
4023                 btrfs_abort_transaction(trans, root, ret);
4024                 goto err;
4025         }
4026
4027         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4028                                            dir, index);
4029         if (ret == -ENOENT)
4030                 ret = 0;
4031         else if (ret)
4032                 btrfs_abort_transaction(trans, root, ret);
4033 err:
4034         btrfs_free_path(path);
4035         if (ret)
4036                 goto out;
4037
4038         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4039         inode_inc_iversion(inode);
4040         inode_inc_iversion(dir);
4041         inode->i_ctime = dir->i_mtime =
4042                 dir->i_ctime = current_fs_time(inode->i_sb);
4043         ret = btrfs_update_inode(trans, root, dir);
4044 out:
4045         return ret;
4046 }
4047
4048 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4049                        struct btrfs_root *root,
4050                        struct inode *dir, struct inode *inode,
4051                        const char *name, int name_len)
4052 {
4053         int ret;
4054         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4055         if (!ret) {
4056                 drop_nlink(inode);
4057                 ret = btrfs_update_inode(trans, root, inode);
4058         }
4059         return ret;
4060 }
4061
4062 /*
4063  * helper to start transaction for unlink and rmdir.
4064  *
4065  * unlink and rmdir are special in btrfs, they do not always free space, so
4066  * if we cannot make our reservations the normal way try and see if there is
4067  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4068  * allow the unlink to occur.
4069  */
4070 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4071 {
4072         struct btrfs_root *root = BTRFS_I(dir)->root;
4073
4074         /*
4075          * 1 for the possible orphan item
4076          * 1 for the dir item
4077          * 1 for the dir index
4078          * 1 for the inode ref
4079          * 1 for the inode
4080          */
4081         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4082 }
4083
4084 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4085 {
4086         struct btrfs_root *root = BTRFS_I(dir)->root;
4087         struct btrfs_trans_handle *trans;
4088         struct inode *inode = d_inode(dentry);
4089         int ret;
4090
4091         trans = __unlink_start_trans(dir);
4092         if (IS_ERR(trans))
4093                 return PTR_ERR(trans);
4094
4095         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4096
4097         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4098                                  dentry->d_name.name, dentry->d_name.len);
4099         if (ret)
4100                 goto out;
4101
4102         if (inode->i_nlink == 0) {
4103                 ret = btrfs_orphan_add(trans, inode);
4104                 if (ret)
4105                         goto out;
4106         }
4107
4108 out:
4109         btrfs_end_transaction(trans, root);
4110         btrfs_btree_balance_dirty(root);
4111         return ret;
4112 }
4113
4114 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4115                         struct btrfs_root *root,
4116                         struct inode *dir, u64 objectid,
4117                         const char *name, int name_len)
4118 {
4119         struct btrfs_path *path;
4120         struct extent_buffer *leaf;
4121         struct btrfs_dir_item *di;
4122         struct btrfs_key key;
4123         u64 index;
4124         int ret;
4125         u64 dir_ino = btrfs_ino(dir);
4126
4127         path = btrfs_alloc_path();
4128         if (!path)
4129                 return -ENOMEM;
4130
4131         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4132                                    name, name_len, -1);
4133         if (IS_ERR_OR_NULL(di)) {
4134                 if (!di)
4135                         ret = -ENOENT;
4136                 else
4137                         ret = PTR_ERR(di);
4138                 goto out;
4139         }
4140
4141         leaf = path->nodes[0];
4142         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4143         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4144         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4145         if (ret) {
4146                 btrfs_abort_transaction(trans, root, ret);
4147                 goto out;
4148         }
4149         btrfs_release_path(path);
4150
4151         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4152                                  objectid, root->root_key.objectid,
4153                                  dir_ino, &index, name, name_len);
4154         if (ret < 0) {
4155                 if (ret != -ENOENT) {
4156                         btrfs_abort_transaction(trans, root, ret);
4157                         goto out;
4158                 }
4159                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4160                                                  name, name_len);
4161                 if (IS_ERR_OR_NULL(di)) {
4162                         if (!di)
4163                                 ret = -ENOENT;
4164                         else
4165                                 ret = PTR_ERR(di);
4166                         btrfs_abort_transaction(trans, root, ret);
4167                         goto out;
4168                 }
4169
4170                 leaf = path->nodes[0];
4171                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4172                 btrfs_release_path(path);
4173                 index = key.offset;
4174         }
4175         btrfs_release_path(path);
4176
4177         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4178         if (ret) {
4179                 btrfs_abort_transaction(trans, root, ret);
4180                 goto out;
4181         }
4182
4183         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4184         inode_inc_iversion(dir);
4185         dir->i_mtime = dir->i_ctime = current_fs_time(dir->i_sb);
4186         ret = btrfs_update_inode_fallback(trans, root, dir);
4187         if (ret)
4188                 btrfs_abort_transaction(trans, root, ret);
4189 out:
4190         btrfs_free_path(path);
4191         return ret;
4192 }
4193
4194 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4195 {
4196         struct inode *inode = d_inode(dentry);
4197         int err = 0;
4198         struct btrfs_root *root = BTRFS_I(dir)->root;
4199         struct btrfs_trans_handle *trans;
4200
4201         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4202                 return -ENOTEMPTY;
4203         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4204                 return -EPERM;
4205
4206         trans = __unlink_start_trans(dir);
4207         if (IS_ERR(trans))
4208                 return PTR_ERR(trans);
4209
4210         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4211                 err = btrfs_unlink_subvol(trans, root, dir,
4212                                           BTRFS_I(inode)->location.objectid,
4213                                           dentry->d_name.name,
4214                                           dentry->d_name.len);
4215                 goto out;
4216         }
4217
4218         err = btrfs_orphan_add(trans, inode);
4219         if (err)
4220                 goto out;
4221
4222         /* now the directory is empty */
4223         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4224                                  dentry->d_name.name, dentry->d_name.len);
4225         if (!err)
4226                 btrfs_i_size_write(inode, 0);
4227 out:
4228         btrfs_end_transaction(trans, root);
4229         btrfs_btree_balance_dirty(root);
4230
4231         return err;
4232 }
4233
4234 static int truncate_space_check(struct btrfs_trans_handle *trans,
4235                                 struct btrfs_root *root,
4236                                 u64 bytes_deleted)
4237 {
4238         int ret;
4239
4240         /*
4241          * This is only used to apply pressure to the enospc system, we don't
4242          * intend to use this reservation at all.
4243          */
4244         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4245         bytes_deleted *= root->nodesize;
4246         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4247                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4248         if (!ret) {
4249                 trace_btrfs_space_reservation(root->fs_info, "transaction",
4250                                               trans->transid,
4251                                               bytes_deleted, 1);
4252                 trans->bytes_reserved += bytes_deleted;
4253         }
4254         return ret;
4255
4256 }
4257
4258 static int truncate_inline_extent(struct inode *inode,
4259                                   struct btrfs_path *path,
4260                                   struct btrfs_key *found_key,
4261                                   const u64 item_end,
4262                                   const u64 new_size)
4263 {
4264         struct extent_buffer *leaf = path->nodes[0];
4265         int slot = path->slots[0];
4266         struct btrfs_file_extent_item *fi;
4267         u32 size = (u32)(new_size - found_key->offset);
4268         struct btrfs_root *root = BTRFS_I(inode)->root;
4269
4270         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4271
4272         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4273                 loff_t offset = new_size;
4274                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4275
4276                 /*
4277                  * Zero out the remaining of the last page of our inline extent,
4278                  * instead of directly truncating our inline extent here - that
4279                  * would be much more complex (decompressing all the data, then
4280                  * compressing the truncated data, which might be bigger than
4281                  * the size of the inline extent, resize the extent, etc).
4282                  * We release the path because to get the page we might need to
4283                  * read the extent item from disk (data not in the page cache).
4284                  */
4285                 btrfs_release_path(path);
4286                 return btrfs_truncate_block(inode, offset, page_end - offset,
4287                                         0);
4288         }
4289
4290         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4291         size = btrfs_file_extent_calc_inline_size(size);
4292         btrfs_truncate_item(root, path, size, 1);
4293
4294         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4295                 inode_sub_bytes(inode, item_end + 1 - new_size);
4296
4297         return 0;
4298 }
4299
4300 /*
4301  * this can truncate away extent items, csum items and directory items.
4302  * It starts at a high offset and removes keys until it can't find
4303  * any higher than new_size
4304  *
4305  * csum items that cross the new i_size are truncated to the new size
4306  * as well.
4307  *
4308  * min_type is the minimum key type to truncate down to.  If set to 0, this
4309  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4310  */
4311 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4312                                struct btrfs_root *root,
4313                                struct inode *inode,
4314                                u64 new_size, u32 min_type)
4315 {
4316         struct btrfs_path *path;
4317         struct extent_buffer *leaf;
4318         struct btrfs_file_extent_item *fi;
4319         struct btrfs_key key;
4320         struct btrfs_key found_key;
4321         u64 extent_start = 0;
4322         u64 extent_num_bytes = 0;
4323         u64 extent_offset = 0;
4324         u64 item_end = 0;
4325         u64 last_size = new_size;
4326         u32 found_type = (u8)-1;
4327         int found_extent;
4328         int del_item;
4329         int pending_del_nr = 0;
4330         int pending_del_slot = 0;
4331         int extent_type = -1;
4332         int ret;
4333         int err = 0;
4334         u64 ino = btrfs_ino(inode);
4335         u64 bytes_deleted = 0;
4336         bool be_nice = 0;
4337         bool should_throttle = 0;
4338         bool should_end = 0;
4339
4340         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4341
4342         /*
4343          * for non-free space inodes and ref cows, we want to back off from
4344          * time to time
4345          */
4346         if (!btrfs_is_free_space_inode(inode) &&
4347             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4348                 be_nice = 1;
4349
4350         path = btrfs_alloc_path();
4351         if (!path)
4352                 return -ENOMEM;
4353         path->reada = READA_BACK;
4354
4355         /*
4356          * We want to drop from the next block forward in case this new size is
4357          * not block aligned since we will be keeping the last block of the
4358          * extent just the way it is.
4359          */
4360         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4361             root == root->fs_info->tree_root)
4362                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4363                                         root->sectorsize), (u64)-1, 0);
4364
4365         /*
4366          * This function is also used to drop the items in the log tree before
4367          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4368          * it is used to drop the loged items. So we shouldn't kill the delayed
4369          * items.
4370          */
4371         if (min_type == 0 && root == BTRFS_I(inode)->root)
4372                 btrfs_kill_delayed_inode_items(inode);
4373
4374         key.objectid = ino;
4375         key.offset = (u64)-1;
4376         key.type = (u8)-1;
4377
4378 search_again:
4379         /*
4380          * with a 16K leaf size and 128MB extents, you can actually queue
4381          * up a huge file in a single leaf.  Most of the time that
4382          * bytes_deleted is > 0, it will be huge by the time we get here
4383          */
4384         if (be_nice && bytes_deleted > SZ_32M) {
4385                 if (btrfs_should_end_transaction(trans, root)) {
4386                         err = -EAGAIN;
4387                         goto error;
4388                 }
4389         }
4390
4391
4392         path->leave_spinning = 1;
4393         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4394         if (ret < 0) {
4395                 err = ret;
4396                 goto out;
4397         }
4398
4399         if (ret > 0) {
4400                 /* there are no items in the tree for us to truncate, we're
4401                  * done
4402                  */
4403                 if (path->slots[0] == 0)
4404                         goto out;
4405                 path->slots[0]--;
4406         }
4407
4408         while (1) {
4409                 fi = NULL;
4410                 leaf = path->nodes[0];
4411                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4412                 found_type = found_key.type;
4413
4414                 if (found_key.objectid != ino)
4415                         break;
4416
4417                 if (found_type < min_type)
4418                         break;
4419
4420                 item_end = found_key.offset;
4421                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4422                         fi = btrfs_item_ptr(leaf, path->slots[0],
4423                                             struct btrfs_file_extent_item);
4424                         extent_type = btrfs_file_extent_type(leaf, fi);
4425                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4426                                 item_end +=
4427                                     btrfs_file_extent_num_bytes(leaf, fi);
4428                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4429                                 item_end += btrfs_file_extent_inline_len(leaf,
4430                                                          path->slots[0], fi);
4431                         }
4432                         item_end--;
4433                 }
4434                 if (found_type > min_type) {
4435                         del_item = 1;
4436                 } else {
4437                         if (item_end < new_size)
4438                                 break;
4439                         if (found_key.offset >= new_size)
4440                                 del_item = 1;
4441                         else
4442                                 del_item = 0;
4443                 }
4444                 found_extent = 0;
4445                 /* FIXME, shrink the extent if the ref count is only 1 */
4446                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4447                         goto delete;
4448
4449                 if (del_item)
4450                         last_size = found_key.offset;
4451                 else
4452                         last_size = new_size;
4453
4454                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4455                         u64 num_dec;
4456                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4457                         if (!del_item) {
4458                                 u64 orig_num_bytes =
4459                                         btrfs_file_extent_num_bytes(leaf, fi);
4460                                 extent_num_bytes = ALIGN(new_size -
4461                                                 found_key.offset,
4462                                                 root->sectorsize);
4463                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4464                                                          extent_num_bytes);
4465                                 num_dec = (orig_num_bytes -
4466                                            extent_num_bytes);
4467                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4468                                              &root->state) &&
4469                                     extent_start != 0)
4470                                         inode_sub_bytes(inode, num_dec);
4471                                 btrfs_mark_buffer_dirty(leaf);
4472                         } else {
4473                                 extent_num_bytes =
4474                                         btrfs_file_extent_disk_num_bytes(leaf,
4475                                                                          fi);
4476                                 extent_offset = found_key.offset -
4477                                         btrfs_file_extent_offset(leaf, fi);
4478
4479                                 /* FIXME blocksize != 4096 */
4480                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4481                                 if (extent_start != 0) {
4482                                         found_extent = 1;
4483                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4484                                                      &root->state))
4485                                                 inode_sub_bytes(inode, num_dec);
4486                                 }
4487                         }
4488                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4489                         /*
4490                          * we can't truncate inline items that have had
4491                          * special encodings
4492                          */
4493                         if (!del_item &&
4494                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4495                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4496
4497                                 /*
4498                                  * Need to release path in order to truncate a
4499                                  * compressed extent. So delete any accumulated
4500                                  * extent items so far.
4501                                  */
4502                                 if (btrfs_file_extent_compression(leaf, fi) !=
4503                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4504                                         err = btrfs_del_items(trans, root, path,
4505                                                               pending_del_slot,
4506                                                               pending_del_nr);
4507                                         if (err) {
4508                                                 btrfs_abort_transaction(trans,
4509                                                                         root,
4510                                                                         err);
4511                                                 goto error;
4512                                         }
4513                                         pending_del_nr = 0;
4514                                 }
4515
4516                                 err = truncate_inline_extent(inode, path,
4517                                                              &found_key,
4518                                                              item_end,
4519                                                              new_size);
4520                                 if (err) {
4521                                         btrfs_abort_transaction(trans,
4522                                                                 root, err);
4523                                         goto error;
4524                                 }
4525                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4526                                             &root->state)) {
4527                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4528                         }
4529                 }
4530 delete:
4531                 if (del_item) {
4532                         if (!pending_del_nr) {
4533                                 /* no pending yet, add ourselves */
4534                                 pending_del_slot = path->slots[0];
4535                                 pending_del_nr = 1;
4536                         } else if (pending_del_nr &&
4537                                    path->slots[0] + 1 == pending_del_slot) {
4538                                 /* hop on the pending chunk */
4539                                 pending_del_nr++;
4540                                 pending_del_slot = path->slots[0];
4541                         } else {
4542                                 BUG();
4543                         }
4544                 } else {
4545                         break;
4546                 }
4547                 should_throttle = 0;
4548
4549                 if (found_extent &&
4550                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4551                      root == root->fs_info->tree_root)) {
4552                         btrfs_set_path_blocking(path);
4553                         bytes_deleted += extent_num_bytes;
4554                         ret = btrfs_free_extent(trans, root, extent_start,
4555                                                 extent_num_bytes, 0,
4556                                                 btrfs_header_owner(leaf),
4557                                                 ino, extent_offset);
4558                         BUG_ON(ret);
4559                         if (btrfs_should_throttle_delayed_refs(trans, root))
4560                                 btrfs_async_run_delayed_refs(root,
4561                                         trans->delayed_ref_updates * 2, 0);
4562                         if (be_nice) {
4563                                 if (truncate_space_check(trans, root,
4564                                                          extent_num_bytes)) {
4565                                         should_end = 1;
4566                                 }
4567                                 if (btrfs_should_throttle_delayed_refs(trans,
4568                                                                        root)) {
4569                                         should_throttle = 1;
4570                                 }
4571                         }
4572                 }
4573
4574                 if (found_type == BTRFS_INODE_ITEM_KEY)
4575                         break;
4576
4577                 if (path->slots[0] == 0 ||
4578                     path->slots[0] != pending_del_slot ||
4579                     should_throttle || should_end) {
4580                         if (pending_del_nr) {
4581                                 ret = btrfs_del_items(trans, root, path,
4582                                                 pending_del_slot,
4583                                                 pending_del_nr);
4584                                 if (ret) {
4585                                         btrfs_abort_transaction(trans,
4586                                                                 root, ret);
4587                                         goto error;
4588                                 }
4589                                 pending_del_nr = 0;
4590                         }
4591                         btrfs_release_path(path);
4592                         if (should_throttle) {
4593                                 unsigned long updates = trans->delayed_ref_updates;
4594                                 if (updates) {
4595                                         trans->delayed_ref_updates = 0;
4596                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4597                                         if (ret && !err)
4598                                                 err = ret;
4599                                 }
4600                         }
4601                         /*
4602                          * if we failed to refill our space rsv, bail out
4603                          * and let the transaction restart
4604                          */
4605                         if (should_end) {
4606                                 err = -EAGAIN;
4607                                 goto error;
4608                         }
4609                         goto search_again;
4610                 } else {
4611                         path->slots[0]--;
4612                 }
4613         }
4614 out:
4615         if (pending_del_nr) {
4616                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4617                                       pending_del_nr);
4618                 if (ret)
4619                         btrfs_abort_transaction(trans, root, ret);
4620         }
4621 error:
4622         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4623                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4624
4625         btrfs_free_path(path);
4626
4627         if (be_nice && bytes_deleted > SZ_32M) {
4628                 unsigned long updates = trans->delayed_ref_updates;
4629                 if (updates) {
4630                         trans->delayed_ref_updates = 0;
4631                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4632                         if (ret && !err)
4633                                 err = ret;
4634                 }
4635         }
4636         return err;
4637 }
4638
4639 /*
4640  * btrfs_truncate_block - read, zero a chunk and write a block
4641  * @inode - inode that we're zeroing
4642  * @from - the offset to start zeroing
4643  * @len - the length to zero, 0 to zero the entire range respective to the
4644  *      offset
4645  * @front - zero up to the offset instead of from the offset on
4646  *
4647  * This will find the block for the "from" offset and cow the block and zero the
4648  * part we want to zero.  This is used with truncate and hole punching.
4649  */
4650 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4651                         int front)
4652 {
4653         struct address_space *mapping = inode->i_mapping;
4654         struct btrfs_root *root = BTRFS_I(inode)->root;
4655         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4656         struct btrfs_ordered_extent *ordered;
4657         struct extent_state *cached_state = NULL;
4658         char *kaddr;
4659         u32 blocksize = root->sectorsize;
4660         pgoff_t index = from >> PAGE_SHIFT;
4661         unsigned offset = from & (blocksize - 1);
4662         struct page *page;
4663         gfp_t mask = btrfs_alloc_write_mask(mapping);
4664         int ret = 0;
4665         u64 block_start;
4666         u64 block_end;
4667
4668         if ((offset & (blocksize - 1)) == 0 &&
4669             (!len || ((len & (blocksize - 1)) == 0)))
4670                 goto out;
4671
4672         ret = btrfs_delalloc_reserve_space(inode,
4673                         round_down(from, blocksize), blocksize);
4674         if (ret)
4675                 goto out;
4676
4677 again:
4678         page = find_or_create_page(mapping, index, mask);
4679         if (!page) {
4680                 btrfs_delalloc_release_space(inode,
4681                                 round_down(from, blocksize),
4682                                 blocksize);
4683                 ret = -ENOMEM;
4684                 goto out;
4685         }
4686
4687         block_start = round_down(from, blocksize);
4688         block_end = block_start + blocksize - 1;
4689
4690         if (!PageUptodate(page)) {
4691                 ret = btrfs_readpage(NULL, page);
4692                 lock_page(page);
4693                 if (page->mapping != mapping) {
4694                         unlock_page(page);
4695                         put_page(page);
4696                         goto again;
4697                 }
4698                 if (!PageUptodate(page)) {
4699                         ret = -EIO;
4700                         goto out_unlock;
4701                 }
4702         }
4703         wait_on_page_writeback(page);
4704
4705         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4706         set_page_extent_mapped(page);
4707
4708         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4709         if (ordered) {
4710                 unlock_extent_cached(io_tree, block_start, block_end,
4711                                      &cached_state, GFP_NOFS);
4712                 unlock_page(page);
4713                 put_page(page);
4714                 btrfs_start_ordered_extent(inode, ordered, 1);
4715                 btrfs_put_ordered_extent(ordered);
4716                 goto again;
4717         }
4718
4719         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4720                           EXTENT_DIRTY | EXTENT_DELALLOC |
4721                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4722                           0, 0, &cached_state, GFP_NOFS);
4723
4724         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4725                                         &cached_state);
4726         if (ret) {
4727                 unlock_extent_cached(io_tree, block_start, block_end,
4728                                      &cached_state, GFP_NOFS);
4729                 goto out_unlock;
4730         }
4731
4732         if (offset != blocksize) {
4733                 if (!len)
4734                         len = blocksize - offset;
4735                 kaddr = kmap(page);
4736                 if (front)
4737                         memset(kaddr + (block_start - page_offset(page)),
4738                                 0, offset);
4739                 else
4740                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4741                                 0, len);
4742                 flush_dcache_page(page);
4743                 kunmap(page);
4744         }
4745         ClearPageChecked(page);
4746         set_page_dirty(page);
4747         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4748                              GFP_NOFS);
4749
4750 out_unlock:
4751         if (ret)
4752                 btrfs_delalloc_release_space(inode, block_start,
4753                                              blocksize);
4754         unlock_page(page);
4755         put_page(page);
4756 out:
4757         return ret;
4758 }
4759
4760 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4761                              u64 offset, u64 len)
4762 {
4763         struct btrfs_trans_handle *trans;
4764         int ret;
4765
4766         /*
4767          * Still need to make sure the inode looks like it's been updated so
4768          * that any holes get logged if we fsync.
4769          */
4770         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4771                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4772                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4773                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4774                 return 0;
4775         }
4776
4777         /*
4778          * 1 - for the one we're dropping
4779          * 1 - for the one we're adding
4780          * 1 - for updating the inode.
4781          */
4782         trans = btrfs_start_transaction(root, 3);
4783         if (IS_ERR(trans))
4784                 return PTR_ERR(trans);
4785
4786         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4787         if (ret) {
4788                 btrfs_abort_transaction(trans, root, ret);
4789                 btrfs_end_transaction(trans, root);
4790                 return ret;
4791         }
4792
4793         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4794                                        0, 0, len, 0, len, 0, 0, 0);
4795         if (ret)
4796                 btrfs_abort_transaction(trans, root, ret);
4797         else
4798                 btrfs_update_inode(trans, root, inode);
4799         btrfs_end_transaction(trans, root);
4800         return ret;
4801 }
4802
4803 /*
4804  * This function puts in dummy file extents for the area we're creating a hole
4805  * for.  So if we are truncating this file to a larger size we need to insert
4806  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4807  * the range between oldsize and size
4808  */
4809 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4810 {
4811         struct btrfs_root *root = BTRFS_I(inode)->root;
4812         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4813         struct extent_map *em = NULL;
4814         struct extent_state *cached_state = NULL;
4815         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4816         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4817         u64 block_end = ALIGN(size, root->sectorsize);
4818         u64 last_byte;
4819         u64 cur_offset;
4820         u64 hole_size;
4821         int err = 0;
4822
4823         /*
4824          * If our size started in the middle of a block we need to zero out the
4825          * rest of the block before we expand the i_size, otherwise we could
4826          * expose stale data.
4827          */
4828         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4829         if (err)
4830                 return err;
4831
4832         if (size <= hole_start)
4833                 return 0;
4834
4835         while (1) {
4836                 struct btrfs_ordered_extent *ordered;
4837
4838                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4839                                  &cached_state);
4840                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4841                                                      block_end - hole_start);
4842                 if (!ordered)
4843                         break;
4844                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4845                                      &cached_state, GFP_NOFS);
4846                 btrfs_start_ordered_extent(inode, ordered, 1);
4847                 btrfs_put_ordered_extent(ordered);
4848         }
4849
4850         cur_offset = hole_start;
4851         while (1) {
4852                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4853                                 block_end - cur_offset, 0);
4854                 if (IS_ERR(em)) {
4855                         err = PTR_ERR(em);
4856                         em = NULL;
4857                         break;
4858                 }
4859                 last_byte = min(extent_map_end(em), block_end);
4860                 last_byte = ALIGN(last_byte , root->sectorsize);
4861                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4862                         struct extent_map *hole_em;
4863                         hole_size = last_byte - cur_offset;
4864
4865                         err = maybe_insert_hole(root, inode, cur_offset,
4866                                                 hole_size);
4867                         if (err)
4868                                 break;
4869                         btrfs_drop_extent_cache(inode, cur_offset,
4870                                                 cur_offset + hole_size - 1, 0);
4871                         hole_em = alloc_extent_map();
4872                         if (!hole_em) {
4873                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4874                                         &BTRFS_I(inode)->runtime_flags);
4875                                 goto next;
4876                         }
4877                         hole_em->start = cur_offset;
4878                         hole_em->len = hole_size;
4879                         hole_em->orig_start = cur_offset;
4880
4881                         hole_em->block_start = EXTENT_MAP_HOLE;
4882                         hole_em->block_len = 0;
4883                         hole_em->orig_block_len = 0;
4884                         hole_em->ram_bytes = hole_size;
4885                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4886                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4887                         hole_em->generation = root->fs_info->generation;
4888
4889                         while (1) {
4890                                 write_lock(&em_tree->lock);
4891                                 err = add_extent_mapping(em_tree, hole_em, 1);
4892                                 write_unlock(&em_tree->lock);
4893                                 if (err != -EEXIST)
4894                                         break;
4895                                 btrfs_drop_extent_cache(inode, cur_offset,
4896                                                         cur_offset +
4897                                                         hole_size - 1, 0);
4898                         }
4899                         free_extent_map(hole_em);
4900                 }
4901 next:
4902                 free_extent_map(em);
4903                 em = NULL;
4904                 cur_offset = last_byte;
4905                 if (cur_offset >= block_end)
4906                         break;
4907         }
4908         free_extent_map(em);
4909         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4910                              GFP_NOFS);
4911         return err;
4912 }
4913
4914 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4915 {
4916         struct btrfs_root *root = BTRFS_I(inode)->root;
4917         struct btrfs_trans_handle *trans;
4918         loff_t oldsize = i_size_read(inode);
4919         loff_t newsize = attr->ia_size;
4920         int mask = attr->ia_valid;
4921         int ret;
4922
4923         /*
4924          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4925          * special case where we need to update the times despite not having
4926          * these flags set.  For all other operations the VFS set these flags
4927          * explicitly if it wants a timestamp update.
4928          */
4929         if (newsize != oldsize) {
4930                 inode_inc_iversion(inode);
4931                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4932                         inode->i_ctime = inode->i_mtime =
4933                                 current_fs_time(inode->i_sb);
4934         }
4935
4936         if (newsize > oldsize) {
4937                 /*
4938                  * Don't do an expanding truncate while snapshoting is ongoing.
4939                  * This is to ensure the snapshot captures a fully consistent
4940                  * state of this file - if the snapshot captures this expanding
4941                  * truncation, it must capture all writes that happened before
4942                  * this truncation.
4943                  */
4944                 btrfs_wait_for_snapshot_creation(root);
4945                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4946                 if (ret) {
4947                         btrfs_end_write_no_snapshoting(root);
4948                         return ret;
4949                 }
4950
4951                 trans = btrfs_start_transaction(root, 1);
4952                 if (IS_ERR(trans)) {
4953                         btrfs_end_write_no_snapshoting(root);
4954                         return PTR_ERR(trans);
4955                 }
4956
4957                 i_size_write(inode, newsize);
4958                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4959                 pagecache_isize_extended(inode, oldsize, newsize);
4960                 ret = btrfs_update_inode(trans, root, inode);
4961                 btrfs_end_write_no_snapshoting(root);
4962                 btrfs_end_transaction(trans, root);
4963         } else {
4964
4965                 /*
4966                  * We're truncating a file that used to have good data down to
4967                  * zero. Make sure it gets into the ordered flush list so that
4968                  * any new writes get down to disk quickly.
4969                  */
4970                 if (newsize == 0)
4971                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4972                                 &BTRFS_I(inode)->runtime_flags);
4973
4974                 /*
4975                  * 1 for the orphan item we're going to add
4976                  * 1 for the orphan item deletion.
4977                  */
4978                 trans = btrfs_start_transaction(root, 2);
4979                 if (IS_ERR(trans))
4980                         return PTR_ERR(trans);
4981
4982                 /*
4983                  * We need to do this in case we fail at _any_ point during the
4984                  * actual truncate.  Once we do the truncate_setsize we could
4985                  * invalidate pages which forces any outstanding ordered io to
4986                  * be instantly completed which will give us extents that need
4987                  * to be truncated.  If we fail to get an orphan inode down we
4988                  * could have left over extents that were never meant to live,
4989                  * so we need to guarantee from this point on that everything
4990                  * will be consistent.
4991                  */
4992                 ret = btrfs_orphan_add(trans, inode);
4993                 btrfs_end_transaction(trans, root);
4994                 if (ret)
4995                         return ret;
4996
4997                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4998                 truncate_setsize(inode, newsize);
4999
5000                 /* Disable nonlocked read DIO to avoid the end less truncate */
5001                 btrfs_inode_block_unlocked_dio(inode);
5002                 inode_dio_wait(inode);
5003                 btrfs_inode_resume_unlocked_dio(inode);
5004
5005                 ret = btrfs_truncate(inode);
5006                 if (ret && inode->i_nlink) {
5007                         int err;
5008
5009                         /*
5010                          * failed to truncate, disk_i_size is only adjusted down
5011                          * as we remove extents, so it should represent the true
5012                          * size of the inode, so reset the in memory size and
5013                          * delete our orphan entry.
5014                          */
5015                         trans = btrfs_join_transaction(root);
5016                         if (IS_ERR(trans)) {
5017                                 btrfs_orphan_del(NULL, inode);
5018                                 return ret;
5019                         }
5020                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5021                         err = btrfs_orphan_del(trans, inode);
5022                         if (err)
5023                                 btrfs_abort_transaction(trans, root, err);
5024                         btrfs_end_transaction(trans, root);
5025                 }
5026         }
5027
5028         return ret;
5029 }
5030
5031 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5032 {
5033         struct inode *inode = d_inode(dentry);
5034         struct btrfs_root *root = BTRFS_I(inode)->root;
5035         int err;
5036
5037         if (btrfs_root_readonly(root))
5038                 return -EROFS;
5039
5040         err = inode_change_ok(inode, attr);
5041         if (err)
5042                 return err;
5043
5044         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5045                 err = btrfs_setsize(inode, attr);
5046                 if (err)
5047                         return err;
5048         }
5049
5050         if (attr->ia_valid) {
5051                 setattr_copy(inode, attr);
5052                 inode_inc_iversion(inode);
5053                 err = btrfs_dirty_inode(inode);
5054
5055                 if (!err && attr->ia_valid & ATTR_MODE)
5056                         err = posix_acl_chmod(inode, inode->i_mode);
5057         }
5058
5059         return err;
5060 }
5061
5062 /*
5063  * While truncating the inode pages during eviction, we get the VFS calling
5064  * btrfs_invalidatepage() against each page of the inode. This is slow because
5065  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5066  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5067  * extent_state structures over and over, wasting lots of time.
5068  *
5069  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5070  * those expensive operations on a per page basis and do only the ordered io
5071  * finishing, while we release here the extent_map and extent_state structures,
5072  * without the excessive merging and splitting.
5073  */
5074 static void evict_inode_truncate_pages(struct inode *inode)
5075 {
5076         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5077         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5078         struct rb_node *node;
5079
5080         ASSERT(inode->i_state & I_FREEING);
5081         truncate_inode_pages_final(&inode->i_data);
5082
5083         write_lock(&map_tree->lock);
5084         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5085                 struct extent_map *em;
5086
5087                 node = rb_first(&map_tree->map);
5088                 em = rb_entry(node, struct extent_map, rb_node);
5089                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5090                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5091                 remove_extent_mapping(map_tree, em);
5092                 free_extent_map(em);
5093                 if (need_resched()) {
5094                         write_unlock(&map_tree->lock);
5095                         cond_resched();
5096                         write_lock(&map_tree->lock);
5097                 }
5098         }
5099         write_unlock(&map_tree->lock);
5100
5101         /*
5102          * Keep looping until we have no more ranges in the io tree.
5103          * We can have ongoing bios started by readpages (called from readahead)
5104          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5105          * still in progress (unlocked the pages in the bio but did not yet
5106          * unlocked the ranges in the io tree). Therefore this means some
5107          * ranges can still be locked and eviction started because before
5108          * submitting those bios, which are executed by a separate task (work
5109          * queue kthread), inode references (inode->i_count) were not taken
5110          * (which would be dropped in the end io callback of each bio).
5111          * Therefore here we effectively end up waiting for those bios and
5112          * anyone else holding locked ranges without having bumped the inode's
5113          * reference count - if we don't do it, when they access the inode's
5114          * io_tree to unlock a range it may be too late, leading to an
5115          * use-after-free issue.
5116          */
5117         spin_lock(&io_tree->lock);
5118         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5119                 struct extent_state *state;
5120                 struct extent_state *cached_state = NULL;
5121                 u64 start;
5122                 u64 end;
5123
5124                 node = rb_first(&io_tree->state);
5125                 state = rb_entry(node, struct extent_state, rb_node);
5126                 start = state->start;
5127                 end = state->end;
5128                 spin_unlock(&io_tree->lock);
5129
5130                 lock_extent_bits(io_tree, start, end, &cached_state);
5131
5132                 /*
5133                  * If still has DELALLOC flag, the extent didn't reach disk,
5134                  * and its reserved space won't be freed by delayed_ref.
5135                  * So we need to free its reserved space here.
5136                  * (Refer to comment in btrfs_invalidatepage, case 2)
5137                  *
5138                  * Note, end is the bytenr of last byte, so we need + 1 here.
5139                  */
5140                 if (state->state & EXTENT_DELALLOC)
5141                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5142
5143                 clear_extent_bit(io_tree, start, end,
5144                                  EXTENT_LOCKED | EXTENT_DIRTY |
5145                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5146                                  EXTENT_DEFRAG, 1, 1,
5147                                  &cached_state, GFP_NOFS);
5148
5149                 cond_resched();
5150                 spin_lock(&io_tree->lock);
5151         }
5152         spin_unlock(&io_tree->lock);
5153 }
5154
5155 void btrfs_evict_inode(struct inode *inode)
5156 {
5157         struct btrfs_trans_handle *trans;
5158         struct btrfs_root *root = BTRFS_I(inode)->root;
5159         struct btrfs_block_rsv *rsv, *global_rsv;
5160         int steal_from_global = 0;
5161         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5162         int ret;
5163
5164         trace_btrfs_inode_evict(inode);
5165
5166         evict_inode_truncate_pages(inode);
5167
5168         if (inode->i_nlink &&
5169             ((btrfs_root_refs(&root->root_item) != 0 &&
5170               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5171              btrfs_is_free_space_inode(inode)))
5172                 goto no_delete;
5173
5174         if (is_bad_inode(inode)) {
5175                 btrfs_orphan_del(NULL, inode);
5176                 goto no_delete;
5177         }
5178         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5179         if (!special_file(inode->i_mode))
5180                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5181
5182         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5183
5184         if (root->fs_info->log_root_recovering) {
5185                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5186                                  &BTRFS_I(inode)->runtime_flags));
5187                 goto no_delete;
5188         }
5189
5190         if (inode->i_nlink > 0) {
5191                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5192                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5193                 goto no_delete;
5194         }
5195
5196         ret = btrfs_commit_inode_delayed_inode(inode);
5197         if (ret) {
5198                 btrfs_orphan_del(NULL, inode);
5199                 goto no_delete;
5200         }
5201
5202         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5203         if (!rsv) {
5204                 btrfs_orphan_del(NULL, inode);
5205                 goto no_delete;
5206         }
5207         rsv->size = min_size;
5208         rsv->failfast = 1;
5209         global_rsv = &root->fs_info->global_block_rsv;
5210
5211         btrfs_i_size_write(inode, 0);
5212
5213         /*
5214          * This is a bit simpler than btrfs_truncate since we've already
5215          * reserved our space for our orphan item in the unlink, so we just
5216          * need to reserve some slack space in case we add bytes and update
5217          * inode item when doing the truncate.
5218          */
5219         while (1) {
5220                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5221                                              BTRFS_RESERVE_FLUSH_LIMIT);
5222
5223                 /*
5224                  * Try and steal from the global reserve since we will
5225                  * likely not use this space anyway, we want to try as
5226                  * hard as possible to get this to work.
5227                  */
5228                 if (ret)
5229                         steal_from_global++;
5230                 else
5231                         steal_from_global = 0;
5232                 ret = 0;
5233
5234                 /*
5235                  * steal_from_global == 0: we reserved stuff, hooray!
5236                  * steal_from_global == 1: we didn't reserve stuff, boo!
5237                  * steal_from_global == 2: we've committed, still not a lot of
5238                  * room but maybe we'll have room in the global reserve this
5239                  * time.
5240                  * steal_from_global == 3: abandon all hope!
5241                  */
5242                 if (steal_from_global > 2) {
5243                         btrfs_warn(root->fs_info,
5244                                 "Could not get space for a delete, will truncate on mount %d",
5245                                 ret);
5246                         btrfs_orphan_del(NULL, inode);
5247                         btrfs_free_block_rsv(root, rsv);
5248                         goto no_delete;
5249                 }
5250
5251                 trans = btrfs_join_transaction(root);
5252                 if (IS_ERR(trans)) {
5253                         btrfs_orphan_del(NULL, inode);
5254                         btrfs_free_block_rsv(root, rsv);
5255                         goto no_delete;
5256                 }
5257
5258                 /*
5259                  * We can't just steal from the global reserve, we need to make
5260                  * sure there is room to do it, if not we need to commit and try
5261                  * again.
5262                  */
5263                 if (steal_from_global) {
5264                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5265                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5266                                                               min_size);
5267                         else
5268                                 ret = -ENOSPC;
5269                 }
5270
5271                 /*
5272                  * Couldn't steal from the global reserve, we have too much
5273                  * pending stuff built up, commit the transaction and try it
5274                  * again.
5275                  */
5276                 if (ret) {
5277                         ret = btrfs_commit_transaction(trans, root);
5278                         if (ret) {
5279                                 btrfs_orphan_del(NULL, inode);
5280                                 btrfs_free_block_rsv(root, rsv);
5281                                 goto no_delete;
5282                         }
5283                         continue;
5284                 } else {
5285                         steal_from_global = 0;
5286                 }
5287
5288                 trans->block_rsv = rsv;
5289
5290                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5291                 if (ret != -ENOSPC && ret != -EAGAIN)
5292                         break;
5293
5294                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5295                 btrfs_end_transaction(trans, root);
5296                 trans = NULL;
5297                 btrfs_btree_balance_dirty(root);
5298         }
5299
5300         btrfs_free_block_rsv(root, rsv);
5301
5302         /*
5303          * Errors here aren't a big deal, it just means we leave orphan items
5304          * in the tree.  They will be cleaned up on the next mount.
5305          */
5306         if (ret == 0) {
5307                 trans->block_rsv = root->orphan_block_rsv;
5308                 btrfs_orphan_del(trans, inode);
5309         } else {
5310                 btrfs_orphan_del(NULL, inode);
5311         }
5312
5313         trans->block_rsv = &root->fs_info->trans_block_rsv;
5314         if (!(root == root->fs_info->tree_root ||
5315               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5316                 btrfs_return_ino(root, btrfs_ino(inode));
5317
5318         btrfs_end_transaction(trans, root);
5319         btrfs_btree_balance_dirty(root);
5320 no_delete:
5321         btrfs_remove_delayed_node(inode);
5322         clear_inode(inode);
5323 }
5324
5325 /*
5326  * this returns the key found in the dir entry in the location pointer.
5327  * If no dir entries were found, location->objectid is 0.
5328  */
5329 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5330                                struct btrfs_key *location)
5331 {
5332         const char *name = dentry->d_name.name;
5333         int namelen = dentry->d_name.len;
5334         struct btrfs_dir_item *di;
5335         struct btrfs_path *path;
5336         struct btrfs_root *root = BTRFS_I(dir)->root;
5337         int ret = 0;
5338
5339         path = btrfs_alloc_path();
5340         if (!path)
5341                 return -ENOMEM;
5342
5343         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5344                                     namelen, 0);
5345         if (IS_ERR(di))
5346                 ret = PTR_ERR(di);
5347
5348         if (IS_ERR_OR_NULL(di))
5349                 goto out_err;
5350
5351         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5352 out:
5353         btrfs_free_path(path);
5354         return ret;
5355 out_err:
5356         location->objectid = 0;
5357         goto out;
5358 }
5359
5360 /*
5361  * when we hit a tree root in a directory, the btrfs part of the inode
5362  * needs to be changed to reflect the root directory of the tree root.  This
5363  * is kind of like crossing a mount point.
5364  */
5365 static int fixup_tree_root_location(struct btrfs_root *root,
5366                                     struct inode *dir,
5367                                     struct dentry *dentry,
5368                                     struct btrfs_key *location,
5369                                     struct btrfs_root **sub_root)
5370 {
5371         struct btrfs_path *path;
5372         struct btrfs_root *new_root;
5373         struct btrfs_root_ref *ref;
5374         struct extent_buffer *leaf;
5375         struct btrfs_key key;
5376         int ret;
5377         int err = 0;
5378
5379         path = btrfs_alloc_path();
5380         if (!path) {
5381                 err = -ENOMEM;
5382                 goto out;
5383         }
5384
5385         err = -ENOENT;
5386         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5387         key.type = BTRFS_ROOT_REF_KEY;
5388         key.offset = location->objectid;
5389
5390         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5391                                 0, 0);
5392         if (ret) {
5393                 if (ret < 0)
5394                         err = ret;
5395                 goto out;
5396         }
5397
5398         leaf = path->nodes[0];
5399         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5400         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5401             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5402                 goto out;
5403
5404         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5405                                    (unsigned long)(ref + 1),
5406                                    dentry->d_name.len);
5407         if (ret)
5408                 goto out;
5409
5410         btrfs_release_path(path);
5411
5412         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5413         if (IS_ERR(new_root)) {
5414                 err = PTR_ERR(new_root);
5415                 goto out;
5416         }
5417
5418         *sub_root = new_root;
5419         location->objectid = btrfs_root_dirid(&new_root->root_item);
5420         location->type = BTRFS_INODE_ITEM_KEY;
5421         location->offset = 0;
5422         err = 0;
5423 out:
5424         btrfs_free_path(path);
5425         return err;
5426 }
5427
5428 static void inode_tree_add(struct inode *inode)
5429 {
5430         struct btrfs_root *root = BTRFS_I(inode)->root;
5431         struct btrfs_inode *entry;
5432         struct rb_node **p;
5433         struct rb_node *parent;
5434         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5435         u64 ino = btrfs_ino(inode);
5436
5437         if (inode_unhashed(inode))
5438                 return;
5439         parent = NULL;
5440         spin_lock(&root->inode_lock);
5441         p = &root->inode_tree.rb_node;
5442         while (*p) {
5443                 parent = *p;
5444                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5445
5446                 if (ino < btrfs_ino(&entry->vfs_inode))
5447                         p = &parent->rb_left;
5448                 else if (ino > btrfs_ino(&entry->vfs_inode))
5449                         p = &parent->rb_right;
5450                 else {
5451                         WARN_ON(!(entry->vfs_inode.i_state &
5452                                   (I_WILL_FREE | I_FREEING)));
5453                         rb_replace_node(parent, new, &root->inode_tree);
5454                         RB_CLEAR_NODE(parent);
5455                         spin_unlock(&root->inode_lock);
5456                         return;
5457                 }
5458         }
5459         rb_link_node(new, parent, p);
5460         rb_insert_color(new, &root->inode_tree);
5461         spin_unlock(&root->inode_lock);
5462 }
5463
5464 static void inode_tree_del(struct inode *inode)
5465 {
5466         struct btrfs_root *root = BTRFS_I(inode)->root;
5467         int empty = 0;
5468
5469         spin_lock(&root->inode_lock);
5470         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5471                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5472                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5473                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5474         }
5475         spin_unlock(&root->inode_lock);
5476
5477         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5478                 synchronize_srcu(&root->fs_info->subvol_srcu);
5479                 spin_lock(&root->inode_lock);
5480                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5481                 spin_unlock(&root->inode_lock);
5482                 if (empty)
5483                         btrfs_add_dead_root(root);
5484         }
5485 }
5486
5487 void btrfs_invalidate_inodes(struct btrfs_root *root)
5488 {
5489         struct rb_node *node;
5490         struct rb_node *prev;
5491         struct btrfs_inode *entry;
5492         struct inode *inode;
5493         u64 objectid = 0;
5494
5495         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5496                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5497
5498         spin_lock(&root->inode_lock);
5499 again:
5500         node = root->inode_tree.rb_node;
5501         prev = NULL;
5502         while (node) {
5503                 prev = node;
5504                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5505
5506                 if (objectid < btrfs_ino(&entry->vfs_inode))
5507                         node = node->rb_left;
5508                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5509                         node = node->rb_right;
5510                 else
5511                         break;
5512         }
5513         if (!node) {
5514                 while (prev) {
5515                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5516                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5517                                 node = prev;
5518                                 break;
5519                         }
5520                         prev = rb_next(prev);
5521                 }
5522         }
5523         while (node) {
5524                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5525                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5526                 inode = igrab(&entry->vfs_inode);
5527                 if (inode) {
5528                         spin_unlock(&root->inode_lock);
5529                         if (atomic_read(&inode->i_count) > 1)
5530                                 d_prune_aliases(inode);
5531                         /*
5532                          * btrfs_drop_inode will have it removed from
5533                          * the inode cache when its usage count
5534                          * hits zero.
5535                          */
5536                         iput(inode);
5537                         cond_resched();
5538                         spin_lock(&root->inode_lock);
5539                         goto again;
5540                 }
5541
5542                 if (cond_resched_lock(&root->inode_lock))
5543                         goto again;
5544
5545                 node = rb_next(node);
5546         }
5547         spin_unlock(&root->inode_lock);
5548 }
5549
5550 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5551 {
5552         struct btrfs_iget_args *args = p;
5553         inode->i_ino = args->location->objectid;
5554         memcpy(&BTRFS_I(inode)->location, args->location,
5555                sizeof(*args->location));
5556         BTRFS_I(inode)->root = args->root;
5557         return 0;
5558 }
5559
5560 static int btrfs_find_actor(struct inode *inode, void *opaque)
5561 {
5562         struct btrfs_iget_args *args = opaque;
5563         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5564                 args->root == BTRFS_I(inode)->root;
5565 }
5566
5567 static struct inode *btrfs_iget_locked(struct super_block *s,
5568                                        struct btrfs_key *location,
5569                                        struct btrfs_root *root)
5570 {
5571         struct inode *inode;
5572         struct btrfs_iget_args args;
5573         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5574
5575         args.location = location;
5576         args.root = root;
5577
5578         inode = iget5_locked(s, hashval, btrfs_find_actor,
5579                              btrfs_init_locked_inode,
5580                              (void *)&args);
5581         return inode;
5582 }
5583
5584 /* Get an inode object given its location and corresponding root.
5585  * Returns in *is_new if the inode was read from disk
5586  */
5587 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5588                          struct btrfs_root *root, int *new)
5589 {
5590         struct inode *inode;
5591
5592         inode = btrfs_iget_locked(s, location, root);
5593         if (!inode)
5594                 return ERR_PTR(-ENOMEM);
5595
5596         if (inode->i_state & I_NEW) {
5597                 btrfs_read_locked_inode(inode);
5598                 if (!is_bad_inode(inode)) {
5599                         inode_tree_add(inode);
5600                         unlock_new_inode(inode);
5601                         if (new)
5602                                 *new = 1;
5603                 } else {
5604                         unlock_new_inode(inode);
5605                         iput(inode);
5606                         inode = ERR_PTR(-ESTALE);
5607                 }
5608         }
5609
5610         return inode;
5611 }
5612
5613 static struct inode *new_simple_dir(struct super_block *s,
5614                                     struct btrfs_key *key,
5615                                     struct btrfs_root *root)
5616 {
5617         struct inode *inode = new_inode(s);
5618
5619         if (!inode)
5620                 return ERR_PTR(-ENOMEM);
5621
5622         BTRFS_I(inode)->root = root;
5623         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5624         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5625
5626         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5627         inode->i_op = &btrfs_dir_ro_inode_operations;
5628         inode->i_fop = &simple_dir_operations;
5629         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5630         inode->i_mtime = current_fs_time(inode->i_sb);
5631         inode->i_atime = inode->i_mtime;
5632         inode->i_ctime = inode->i_mtime;
5633         BTRFS_I(inode)->i_otime = inode->i_mtime;
5634
5635         return inode;
5636 }
5637
5638 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5639 {
5640         struct inode *inode;
5641         struct btrfs_root *root = BTRFS_I(dir)->root;
5642         struct btrfs_root *sub_root = root;
5643         struct btrfs_key location;
5644         int index;
5645         int ret = 0;
5646
5647         if (dentry->d_name.len > BTRFS_NAME_LEN)
5648                 return ERR_PTR(-ENAMETOOLONG);
5649
5650         ret = btrfs_inode_by_name(dir, dentry, &location);
5651         if (ret < 0)
5652                 return ERR_PTR(ret);
5653
5654         if (location.objectid == 0)
5655                 return ERR_PTR(-ENOENT);
5656
5657         if (location.type == BTRFS_INODE_ITEM_KEY) {
5658                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5659                 return inode;
5660         }
5661
5662         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5663
5664         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5665         ret = fixup_tree_root_location(root, dir, dentry,
5666                                        &location, &sub_root);
5667         if (ret < 0) {
5668                 if (ret != -ENOENT)
5669                         inode = ERR_PTR(ret);
5670                 else
5671                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5672         } else {
5673                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5674         }
5675         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5676
5677         if (!IS_ERR(inode) && root != sub_root) {
5678                 down_read(&root->fs_info->cleanup_work_sem);
5679                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5680                         ret = btrfs_orphan_cleanup(sub_root);
5681                 up_read(&root->fs_info->cleanup_work_sem);
5682                 if (ret) {
5683                         iput(inode);
5684                         inode = ERR_PTR(ret);
5685                 }
5686         }
5687
5688         return inode;
5689 }
5690
5691 static int btrfs_dentry_delete(const struct dentry *dentry)
5692 {
5693         struct btrfs_root *root;
5694         struct inode *inode = d_inode(dentry);
5695
5696         if (!inode && !IS_ROOT(dentry))
5697                 inode = d_inode(dentry->d_parent);
5698
5699         if (inode) {
5700                 root = BTRFS_I(inode)->root;
5701                 if (btrfs_root_refs(&root->root_item) == 0)
5702                         return 1;
5703
5704                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5705                         return 1;
5706         }
5707         return 0;
5708 }
5709
5710 static void btrfs_dentry_release(struct dentry *dentry)
5711 {
5712         kfree(dentry->d_fsdata);
5713 }
5714
5715 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5716                                    unsigned int flags)
5717 {
5718         struct inode *inode;
5719
5720         inode = btrfs_lookup_dentry(dir, dentry);
5721         if (IS_ERR(inode)) {
5722                 if (PTR_ERR(inode) == -ENOENT)
5723                         inode = NULL;
5724                 else
5725                         return ERR_CAST(inode);
5726         }
5727
5728         return d_splice_alias(inode, dentry);
5729 }
5730
5731 unsigned char btrfs_filetype_table[] = {
5732         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5733 };
5734
5735 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5736 {
5737         struct inode *inode = file_inode(file);
5738         struct btrfs_root *root = BTRFS_I(inode)->root;
5739         struct btrfs_item *item;
5740         struct btrfs_dir_item *di;
5741         struct btrfs_key key;
5742         struct btrfs_key found_key;
5743         struct btrfs_path *path;
5744         struct list_head ins_list;
5745         struct list_head del_list;
5746         int ret;
5747         struct extent_buffer *leaf;
5748         int slot;
5749         unsigned char d_type;
5750         int over = 0;
5751         u32 di_cur;
5752         u32 di_total;
5753         u32 di_len;
5754         int key_type = BTRFS_DIR_INDEX_KEY;
5755         char tmp_name[32];
5756         char *name_ptr;
5757         int name_len;
5758         int is_curr = 0;        /* ctx->pos points to the current index? */
5759         bool emitted;
5760
5761         /* FIXME, use a real flag for deciding about the key type */
5762         if (root->fs_info->tree_root == root)
5763                 key_type = BTRFS_DIR_ITEM_KEY;
5764
5765         if (!dir_emit_dots(file, ctx))
5766                 return 0;
5767
5768         path = btrfs_alloc_path();
5769         if (!path)
5770                 return -ENOMEM;
5771
5772         path->reada = READA_FORWARD;
5773
5774         if (key_type == BTRFS_DIR_INDEX_KEY) {
5775                 INIT_LIST_HEAD(&ins_list);
5776                 INIT_LIST_HEAD(&del_list);
5777                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5778         }
5779
5780         key.type = key_type;
5781         key.offset = ctx->pos;
5782         key.objectid = btrfs_ino(inode);
5783
5784         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5785         if (ret < 0)
5786                 goto err;
5787
5788         emitted = false;
5789         while (1) {
5790                 leaf = path->nodes[0];
5791                 slot = path->slots[0];
5792                 if (slot >= btrfs_header_nritems(leaf)) {
5793                         ret = btrfs_next_leaf(root, path);
5794                         if (ret < 0)
5795                                 goto err;
5796                         else if (ret > 0)
5797                                 break;
5798                         continue;
5799                 }
5800
5801                 item = btrfs_item_nr(slot);
5802                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5803
5804                 if (found_key.objectid != key.objectid)
5805                         break;
5806                 if (found_key.type != key_type)
5807                         break;
5808                 if (found_key.offset < ctx->pos)
5809                         goto next;
5810                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5811                     btrfs_should_delete_dir_index(&del_list,
5812                                                   found_key.offset))
5813                         goto next;
5814
5815                 ctx->pos = found_key.offset;
5816                 is_curr = 1;
5817
5818                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5819                 di_cur = 0;
5820                 di_total = btrfs_item_size(leaf, item);
5821
5822                 while (di_cur < di_total) {
5823                         struct btrfs_key location;
5824
5825                         if (verify_dir_item(root, leaf, di))
5826                                 break;
5827
5828                         name_len = btrfs_dir_name_len(leaf, di);
5829                         if (name_len <= sizeof(tmp_name)) {
5830                                 name_ptr = tmp_name;
5831                         } else {
5832                                 name_ptr = kmalloc(name_len, GFP_KERNEL);
5833                                 if (!name_ptr) {
5834                                         ret = -ENOMEM;
5835                                         goto err;
5836                                 }
5837                         }
5838                         read_extent_buffer(leaf, name_ptr,
5839                                            (unsigned long)(di + 1), name_len);
5840
5841                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5842                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5843
5844
5845                         /* is this a reference to our own snapshot? If so
5846                          * skip it.
5847                          *
5848                          * In contrast to old kernels, we insert the snapshot's
5849                          * dir item and dir index after it has been created, so
5850                          * we won't find a reference to our own snapshot. We
5851                          * still keep the following code for backward
5852                          * compatibility.
5853                          */
5854                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5855                             location.objectid == root->root_key.objectid) {
5856                                 over = 0;
5857                                 goto skip;
5858                         }
5859                         over = !dir_emit(ctx, name_ptr, name_len,
5860                                        location.objectid, d_type);
5861
5862 skip:
5863                         if (name_ptr != tmp_name)
5864                                 kfree(name_ptr);
5865
5866                         if (over)
5867                                 goto nopos;
5868                         emitted = true;
5869                         di_len = btrfs_dir_name_len(leaf, di) +
5870                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5871                         di_cur += di_len;
5872                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5873                 }
5874 next:
5875                 path->slots[0]++;
5876         }
5877
5878         if (key_type == BTRFS_DIR_INDEX_KEY) {
5879                 if (is_curr)
5880                         ctx->pos++;
5881                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list, &emitted);
5882                 if (ret)
5883                         goto nopos;
5884         }
5885
5886         /*
5887          * If we haven't emitted any dir entry, we must not touch ctx->pos as
5888          * it was was set to the termination value in previous call. We assume
5889          * that "." and ".." were emitted if we reach this point and set the
5890          * termination value as well for an empty directory.
5891          */
5892         if (ctx->pos > 2 && !emitted)
5893                 goto nopos;
5894
5895         /* Reached end of directory/root. Bump pos past the last item. */
5896         ctx->pos++;
5897
5898         /*
5899          * Stop new entries from being returned after we return the last
5900          * entry.
5901          *
5902          * New directory entries are assigned a strictly increasing
5903          * offset.  This means that new entries created during readdir
5904          * are *guaranteed* to be seen in the future by that readdir.
5905          * This has broken buggy programs which operate on names as
5906          * they're returned by readdir.  Until we re-use freed offsets
5907          * we have this hack to stop new entries from being returned
5908          * under the assumption that they'll never reach this huge
5909          * offset.
5910          *
5911          * This is being careful not to overflow 32bit loff_t unless the
5912          * last entry requires it because doing so has broken 32bit apps
5913          * in the past.
5914          */
5915         if (key_type == BTRFS_DIR_INDEX_KEY) {
5916                 if (ctx->pos >= INT_MAX)
5917                         ctx->pos = LLONG_MAX;
5918                 else
5919                         ctx->pos = INT_MAX;
5920         }
5921 nopos:
5922         ret = 0;
5923 err:
5924         if (key_type == BTRFS_DIR_INDEX_KEY)
5925                 btrfs_put_delayed_items(&ins_list, &del_list);
5926         btrfs_free_path(path);
5927         return ret;
5928 }
5929
5930 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5931 {
5932         struct btrfs_root *root = BTRFS_I(inode)->root;
5933         struct btrfs_trans_handle *trans;
5934         int ret = 0;
5935         bool nolock = false;
5936
5937         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5938                 return 0;
5939
5940         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5941                 nolock = true;
5942
5943         if (wbc->sync_mode == WB_SYNC_ALL) {
5944                 if (nolock)
5945                         trans = btrfs_join_transaction_nolock(root);
5946                 else
5947                         trans = btrfs_join_transaction(root);
5948                 if (IS_ERR(trans))
5949                         return PTR_ERR(trans);
5950                 ret = btrfs_commit_transaction(trans, root);
5951         }
5952         return ret;
5953 }
5954
5955 /*
5956  * This is somewhat expensive, updating the tree every time the
5957  * inode changes.  But, it is most likely to find the inode in cache.
5958  * FIXME, needs more benchmarking...there are no reasons other than performance
5959  * to keep or drop this code.
5960  */
5961 static int btrfs_dirty_inode(struct inode *inode)
5962 {
5963         struct btrfs_root *root = BTRFS_I(inode)->root;
5964         struct btrfs_trans_handle *trans;
5965         int ret;
5966
5967         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5968                 return 0;
5969
5970         trans = btrfs_join_transaction(root);
5971         if (IS_ERR(trans))
5972                 return PTR_ERR(trans);
5973
5974         ret = btrfs_update_inode(trans, root, inode);
5975         if (ret && ret == -ENOSPC) {
5976                 /* whoops, lets try again with the full transaction */
5977                 btrfs_end_transaction(trans, root);
5978                 trans = btrfs_start_transaction(root, 1);
5979                 if (IS_ERR(trans))
5980                         return PTR_ERR(trans);
5981
5982                 ret = btrfs_update_inode(trans, root, inode);
5983         }
5984         btrfs_end_transaction(trans, root);
5985         if (BTRFS_I(inode)->delayed_node)
5986                 btrfs_balance_delayed_items(root);
5987
5988         return ret;
5989 }
5990
5991 /*
5992  * This is a copy of file_update_time.  We need this so we can return error on
5993  * ENOSPC for updating the inode in the case of file write and mmap writes.
5994  */
5995 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5996                              int flags)
5997 {
5998         struct btrfs_root *root = BTRFS_I(inode)->root;
5999
6000         if (btrfs_root_readonly(root))
6001                 return -EROFS;
6002
6003         if (flags & S_VERSION)
6004                 inode_inc_iversion(inode);
6005         if (flags & S_CTIME)
6006                 inode->i_ctime = *now;
6007         if (flags & S_MTIME)
6008                 inode->i_mtime = *now;
6009         if (flags & S_ATIME)
6010                 inode->i_atime = *now;
6011         return btrfs_dirty_inode(inode);
6012 }
6013
6014 /*
6015  * find the highest existing sequence number in a directory
6016  * and then set the in-memory index_cnt variable to reflect
6017  * free sequence numbers
6018  */
6019 static int btrfs_set_inode_index_count(struct inode *inode)
6020 {
6021         struct btrfs_root *root = BTRFS_I(inode)->root;
6022         struct btrfs_key key, found_key;
6023         struct btrfs_path *path;
6024         struct extent_buffer *leaf;
6025         int ret;
6026
6027         key.objectid = btrfs_ino(inode);
6028         key.type = BTRFS_DIR_INDEX_KEY;
6029         key.offset = (u64)-1;
6030
6031         path = btrfs_alloc_path();
6032         if (!path)
6033                 return -ENOMEM;
6034
6035         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6036         if (ret < 0)
6037                 goto out;
6038         /* FIXME: we should be able to handle this */
6039         if (ret == 0)
6040                 goto out;
6041         ret = 0;
6042
6043         /*
6044          * MAGIC NUMBER EXPLANATION:
6045          * since we search a directory based on f_pos we have to start at 2
6046          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6047          * else has to start at 2
6048          */
6049         if (path->slots[0] == 0) {
6050                 BTRFS_I(inode)->index_cnt = 2;
6051                 goto out;
6052         }
6053
6054         path->slots[0]--;
6055
6056         leaf = path->nodes[0];
6057         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6058
6059         if (found_key.objectid != btrfs_ino(inode) ||
6060             found_key.type != BTRFS_DIR_INDEX_KEY) {
6061                 BTRFS_I(inode)->index_cnt = 2;
6062                 goto out;
6063         }
6064
6065         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6066 out:
6067         btrfs_free_path(path);
6068         return ret;
6069 }
6070
6071 /*
6072  * helper to find a free sequence number in a given directory.  This current
6073  * code is very simple, later versions will do smarter things in the btree
6074  */
6075 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6076 {
6077         int ret = 0;
6078
6079         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6080                 ret = btrfs_inode_delayed_dir_index_count(dir);
6081                 if (ret) {
6082                         ret = btrfs_set_inode_index_count(dir);
6083                         if (ret)
6084                                 return ret;
6085                 }
6086         }
6087
6088         *index = BTRFS_I(dir)->index_cnt;
6089         BTRFS_I(dir)->index_cnt++;
6090
6091         return ret;
6092 }
6093
6094 static int btrfs_insert_inode_locked(struct inode *inode)
6095 {
6096         struct btrfs_iget_args args;
6097         args.location = &BTRFS_I(inode)->location;
6098         args.root = BTRFS_I(inode)->root;
6099
6100         return insert_inode_locked4(inode,
6101                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6102                    btrfs_find_actor, &args);
6103 }
6104
6105 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6106                                      struct btrfs_root *root,
6107                                      struct inode *dir,
6108                                      const char *name, int name_len,
6109                                      u64 ref_objectid, u64 objectid,
6110                                      umode_t mode, u64 *index)
6111 {
6112         struct inode *inode;
6113         struct btrfs_inode_item *inode_item;
6114         struct btrfs_key *location;
6115         struct btrfs_path *path;
6116         struct btrfs_inode_ref *ref;
6117         struct btrfs_key key[2];
6118         u32 sizes[2];
6119         int nitems = name ? 2 : 1;
6120         unsigned long ptr;
6121         int ret;
6122
6123         path = btrfs_alloc_path();
6124         if (!path)
6125                 return ERR_PTR(-ENOMEM);
6126
6127         inode = new_inode(root->fs_info->sb);
6128         if (!inode) {
6129                 btrfs_free_path(path);
6130                 return ERR_PTR(-ENOMEM);
6131         }
6132
6133         /*
6134          * O_TMPFILE, set link count to 0, so that after this point,
6135          * we fill in an inode item with the correct link count.
6136          */
6137         if (!name)
6138                 set_nlink(inode, 0);
6139
6140         /*
6141          * we have to initialize this early, so we can reclaim the inode
6142          * number if we fail afterwards in this function.
6143          */
6144         inode->i_ino = objectid;
6145
6146         if (dir && name) {
6147                 trace_btrfs_inode_request(dir);
6148
6149                 ret = btrfs_set_inode_index(dir, index);
6150                 if (ret) {
6151                         btrfs_free_path(path);
6152                         iput(inode);
6153                         return ERR_PTR(ret);
6154                 }
6155         } else if (dir) {
6156                 *index = 0;
6157         }
6158         /*
6159          * index_cnt is ignored for everything but a dir,
6160          * btrfs_get_inode_index_count has an explanation for the magic
6161          * number
6162          */
6163         BTRFS_I(inode)->index_cnt = 2;
6164         BTRFS_I(inode)->dir_index = *index;
6165         BTRFS_I(inode)->root = root;
6166         BTRFS_I(inode)->generation = trans->transid;
6167         inode->i_generation = BTRFS_I(inode)->generation;
6168
6169         /*
6170          * We could have gotten an inode number from somebody who was fsynced
6171          * and then removed in this same transaction, so let's just set full
6172          * sync since it will be a full sync anyway and this will blow away the
6173          * old info in the log.
6174          */
6175         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6176
6177         key[0].objectid = objectid;
6178         key[0].type = BTRFS_INODE_ITEM_KEY;
6179         key[0].offset = 0;
6180
6181         sizes[0] = sizeof(struct btrfs_inode_item);
6182
6183         if (name) {
6184                 /*
6185                  * Start new inodes with an inode_ref. This is slightly more
6186                  * efficient for small numbers of hard links since they will
6187                  * be packed into one item. Extended refs will kick in if we
6188                  * add more hard links than can fit in the ref item.
6189                  */
6190                 key[1].objectid = objectid;
6191                 key[1].type = BTRFS_INODE_REF_KEY;
6192                 key[1].offset = ref_objectid;
6193
6194                 sizes[1] = name_len + sizeof(*ref);
6195         }
6196
6197         location = &BTRFS_I(inode)->location;
6198         location->objectid = objectid;
6199         location->offset = 0;
6200         location->type = BTRFS_INODE_ITEM_KEY;
6201
6202         ret = btrfs_insert_inode_locked(inode);
6203         if (ret < 0)
6204                 goto fail;
6205
6206         path->leave_spinning = 1;
6207         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6208         if (ret != 0)
6209                 goto fail_unlock;
6210
6211         inode_init_owner(inode, dir, mode);
6212         inode_set_bytes(inode, 0);
6213
6214         inode->i_mtime = current_fs_time(inode->i_sb);
6215         inode->i_atime = inode->i_mtime;
6216         inode->i_ctime = inode->i_mtime;
6217         BTRFS_I(inode)->i_otime = inode->i_mtime;
6218
6219         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6220                                   struct btrfs_inode_item);
6221         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6222                              sizeof(*inode_item));
6223         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6224
6225         if (name) {
6226                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6227                                      struct btrfs_inode_ref);
6228                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6229                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6230                 ptr = (unsigned long)(ref + 1);
6231                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6232         }
6233
6234         btrfs_mark_buffer_dirty(path->nodes[0]);
6235         btrfs_free_path(path);
6236
6237         btrfs_inherit_iflags(inode, dir);
6238
6239         if (S_ISREG(mode)) {
6240                 if (btrfs_test_opt(root, NODATASUM))
6241                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6242                 if (btrfs_test_opt(root, NODATACOW))
6243                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6244                                 BTRFS_INODE_NODATASUM;
6245         }
6246
6247         inode_tree_add(inode);
6248
6249         trace_btrfs_inode_new(inode);
6250         btrfs_set_inode_last_trans(trans, inode);
6251
6252         btrfs_update_root_times(trans, root);
6253
6254         ret = btrfs_inode_inherit_props(trans, inode, dir);
6255         if (ret)
6256                 btrfs_err(root->fs_info,
6257                           "error inheriting props for ino %llu (root %llu): %d",
6258                           btrfs_ino(inode), root->root_key.objectid, ret);
6259
6260         return inode;
6261
6262 fail_unlock:
6263         unlock_new_inode(inode);
6264 fail:
6265         if (dir && name)
6266                 BTRFS_I(dir)->index_cnt--;
6267         btrfs_free_path(path);
6268         iput(inode);
6269         return ERR_PTR(ret);
6270 }
6271
6272 static inline u8 btrfs_inode_type(struct inode *inode)
6273 {
6274         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6275 }
6276
6277 /*
6278  * utility function to add 'inode' into 'parent_inode' with
6279  * a give name and a given sequence number.
6280  * if 'add_backref' is true, also insert a backref from the
6281  * inode to the parent directory.
6282  */
6283 int btrfs_add_link(struct btrfs_trans_handle *trans,
6284                    struct inode *parent_inode, struct inode *inode,
6285                    const char *name, int name_len, int add_backref, u64 index)
6286 {
6287         int ret = 0;
6288         struct btrfs_key key;
6289         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6290         u64 ino = btrfs_ino(inode);
6291         u64 parent_ino = btrfs_ino(parent_inode);
6292
6293         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6294                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6295         } else {
6296                 key.objectid = ino;
6297                 key.type = BTRFS_INODE_ITEM_KEY;
6298                 key.offset = 0;
6299         }
6300
6301         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6302                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6303                                          key.objectid, root->root_key.objectid,
6304                                          parent_ino, index, name, name_len);
6305         } else if (add_backref) {
6306                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6307                                              parent_ino, index);
6308         }
6309
6310         /* Nothing to clean up yet */
6311         if (ret)
6312                 return ret;
6313
6314         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6315                                     parent_inode, &key,
6316                                     btrfs_inode_type(inode), index);
6317         if (ret == -EEXIST || ret == -EOVERFLOW)
6318                 goto fail_dir_item;
6319         else if (ret) {
6320                 btrfs_abort_transaction(trans, root, ret);
6321                 return ret;
6322         }
6323
6324         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6325                            name_len * 2);
6326         inode_inc_iversion(parent_inode);
6327         parent_inode->i_mtime = parent_inode->i_ctime =
6328                 current_fs_time(parent_inode->i_sb);
6329         ret = btrfs_update_inode(trans, root, parent_inode);
6330         if (ret)
6331                 btrfs_abort_transaction(trans, root, ret);
6332         return ret;
6333
6334 fail_dir_item:
6335         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6336                 u64 local_index;
6337                 int err;
6338                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6339                                  key.objectid, root->root_key.objectid,
6340                                  parent_ino, &local_index, name, name_len);
6341
6342         } else if (add_backref) {
6343                 u64 local_index;
6344                 int err;
6345
6346                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6347                                           ino, parent_ino, &local_index);
6348         }
6349         return ret;
6350 }
6351
6352 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6353                             struct inode *dir, struct dentry *dentry,
6354                             struct inode *inode, int backref, u64 index)
6355 {
6356         int err = btrfs_add_link(trans, dir, inode,
6357                                  dentry->d_name.name, dentry->d_name.len,
6358                                  backref, index);
6359         if (err > 0)
6360                 err = -EEXIST;
6361         return err;
6362 }
6363
6364 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6365                         umode_t mode, dev_t rdev)
6366 {
6367         struct btrfs_trans_handle *trans;
6368         struct btrfs_root *root = BTRFS_I(dir)->root;
6369         struct inode *inode = NULL;
6370         int err;
6371         int drop_inode = 0;
6372         u64 objectid;
6373         u64 index = 0;
6374
6375         /*
6376          * 2 for inode item and ref
6377          * 2 for dir items
6378          * 1 for xattr if selinux is on
6379          */
6380         trans = btrfs_start_transaction(root, 5);
6381         if (IS_ERR(trans))
6382                 return PTR_ERR(trans);
6383
6384         err = btrfs_find_free_ino(root, &objectid);
6385         if (err)
6386                 goto out_unlock;
6387
6388         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6389                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6390                                 mode, &index);
6391         if (IS_ERR(inode)) {
6392                 err = PTR_ERR(inode);
6393                 goto out_unlock;
6394         }
6395
6396         /*
6397         * If the active LSM wants to access the inode during
6398         * d_instantiate it needs these. Smack checks to see
6399         * if the filesystem supports xattrs by looking at the
6400         * ops vector.
6401         */
6402         inode->i_op = &btrfs_special_inode_operations;
6403         init_special_inode(inode, inode->i_mode, rdev);
6404
6405         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6406         if (err)
6407                 goto out_unlock_inode;
6408
6409         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6410         if (err) {
6411                 goto out_unlock_inode;
6412         } else {
6413                 btrfs_update_inode(trans, root, inode);
6414                 unlock_new_inode(inode);
6415                 d_instantiate(dentry, inode);
6416         }
6417
6418 out_unlock:
6419         btrfs_end_transaction(trans, root);
6420         btrfs_balance_delayed_items(root);
6421         btrfs_btree_balance_dirty(root);
6422         if (drop_inode) {
6423                 inode_dec_link_count(inode);
6424                 iput(inode);
6425         }
6426         return err;
6427
6428 out_unlock_inode:
6429         drop_inode = 1;
6430         unlock_new_inode(inode);
6431         goto out_unlock;
6432
6433 }
6434
6435 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6436                         umode_t mode, bool excl)
6437 {
6438         struct btrfs_trans_handle *trans;
6439         struct btrfs_root *root = BTRFS_I(dir)->root;
6440         struct inode *inode = NULL;
6441         int drop_inode_on_err = 0;
6442         int err;
6443         u64 objectid;
6444         u64 index = 0;
6445
6446         /*
6447          * 2 for inode item and ref
6448          * 2 for dir items
6449          * 1 for xattr if selinux is on
6450          */
6451         trans = btrfs_start_transaction(root, 5);
6452         if (IS_ERR(trans))
6453                 return PTR_ERR(trans);
6454
6455         err = btrfs_find_free_ino(root, &objectid);
6456         if (err)
6457                 goto out_unlock;
6458
6459         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6460                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6461                                 mode, &index);
6462         if (IS_ERR(inode)) {
6463                 err = PTR_ERR(inode);
6464                 goto out_unlock;
6465         }
6466         drop_inode_on_err = 1;
6467         /*
6468         * If the active LSM wants to access the inode during
6469         * d_instantiate it needs these. Smack checks to see
6470         * if the filesystem supports xattrs by looking at the
6471         * ops vector.
6472         */
6473         inode->i_fop = &btrfs_file_operations;
6474         inode->i_op = &btrfs_file_inode_operations;
6475         inode->i_mapping->a_ops = &btrfs_aops;
6476
6477         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6478         if (err)
6479                 goto out_unlock_inode;
6480
6481         err = btrfs_update_inode(trans, root, inode);
6482         if (err)
6483                 goto out_unlock_inode;
6484
6485         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6486         if (err)
6487                 goto out_unlock_inode;
6488
6489         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6490         unlock_new_inode(inode);
6491         d_instantiate(dentry, inode);
6492
6493 out_unlock:
6494         btrfs_end_transaction(trans, root);
6495         if (err && drop_inode_on_err) {
6496                 inode_dec_link_count(inode);
6497                 iput(inode);
6498         }
6499         btrfs_balance_delayed_items(root);
6500         btrfs_btree_balance_dirty(root);
6501         return err;
6502
6503 out_unlock_inode:
6504         unlock_new_inode(inode);
6505         goto out_unlock;
6506
6507 }
6508
6509 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6510                       struct dentry *dentry)
6511 {
6512         struct btrfs_trans_handle *trans = NULL;
6513         struct btrfs_root *root = BTRFS_I(dir)->root;
6514         struct inode *inode = d_inode(old_dentry);
6515         u64 index;
6516         int err;
6517         int drop_inode = 0;
6518
6519         /* do not allow sys_link's with other subvols of the same device */
6520         if (root->objectid != BTRFS_I(inode)->root->objectid)
6521                 return -EXDEV;
6522
6523         if (inode->i_nlink >= BTRFS_LINK_MAX)
6524                 return -EMLINK;
6525
6526         err = btrfs_set_inode_index(dir, &index);
6527         if (err)
6528                 goto fail;
6529
6530         /*
6531          * 2 items for inode and inode ref
6532          * 2 items for dir items
6533          * 1 item for parent inode
6534          */
6535         trans = btrfs_start_transaction(root, 5);
6536         if (IS_ERR(trans)) {
6537                 err = PTR_ERR(trans);
6538                 trans = NULL;
6539                 goto fail;
6540         }
6541
6542         /* There are several dir indexes for this inode, clear the cache. */
6543         BTRFS_I(inode)->dir_index = 0ULL;
6544         inc_nlink(inode);
6545         inode_inc_iversion(inode);
6546         inode->i_ctime = current_fs_time(inode->i_sb);
6547         ihold(inode);
6548         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6549
6550         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6551
6552         if (err) {
6553                 drop_inode = 1;
6554         } else {
6555                 struct dentry *parent = dentry->d_parent;
6556                 err = btrfs_update_inode(trans, root, inode);
6557                 if (err)
6558                         goto fail;
6559                 if (inode->i_nlink == 1) {
6560                         /*
6561                          * If new hard link count is 1, it's a file created
6562                          * with open(2) O_TMPFILE flag.
6563                          */
6564                         err = btrfs_orphan_del(trans, inode);
6565                         if (err)
6566                                 goto fail;
6567                 }
6568                 d_instantiate(dentry, inode);
6569                 btrfs_log_new_name(trans, inode, NULL, parent);
6570         }
6571
6572         btrfs_balance_delayed_items(root);
6573 fail:
6574         if (trans)
6575                 btrfs_end_transaction(trans, root);
6576         if (drop_inode) {
6577                 inode_dec_link_count(inode);
6578                 iput(inode);
6579         }
6580         btrfs_btree_balance_dirty(root);
6581         return err;
6582 }
6583
6584 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6585 {
6586         struct inode *inode = NULL;
6587         struct btrfs_trans_handle *trans;
6588         struct btrfs_root *root = BTRFS_I(dir)->root;
6589         int err = 0;
6590         int drop_on_err = 0;
6591         u64 objectid = 0;
6592         u64 index = 0;
6593
6594         /*
6595          * 2 items for inode and ref
6596          * 2 items for dir items
6597          * 1 for xattr if selinux is on
6598          */
6599         trans = btrfs_start_transaction(root, 5);
6600         if (IS_ERR(trans))
6601                 return PTR_ERR(trans);
6602
6603         err = btrfs_find_free_ino(root, &objectid);
6604         if (err)
6605                 goto out_fail;
6606
6607         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6608                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6609                                 S_IFDIR | mode, &index);
6610         if (IS_ERR(inode)) {
6611                 err = PTR_ERR(inode);
6612                 goto out_fail;
6613         }
6614
6615         drop_on_err = 1;
6616         /* these must be set before we unlock the inode */
6617         inode->i_op = &btrfs_dir_inode_operations;
6618         inode->i_fop = &btrfs_dir_file_operations;
6619
6620         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6621         if (err)
6622                 goto out_fail_inode;
6623
6624         btrfs_i_size_write(inode, 0);
6625         err = btrfs_update_inode(trans, root, inode);
6626         if (err)
6627                 goto out_fail_inode;
6628
6629         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6630                              dentry->d_name.len, 0, index);
6631         if (err)
6632                 goto out_fail_inode;
6633
6634         d_instantiate(dentry, inode);
6635         /*
6636          * mkdir is special.  We're unlocking after we call d_instantiate
6637          * to avoid a race with nfsd calling d_instantiate.
6638          */
6639         unlock_new_inode(inode);
6640         drop_on_err = 0;
6641
6642 out_fail:
6643         btrfs_end_transaction(trans, root);
6644         if (drop_on_err) {
6645                 inode_dec_link_count(inode);
6646                 iput(inode);
6647         }
6648         btrfs_balance_delayed_items(root);
6649         btrfs_btree_balance_dirty(root);
6650         return err;
6651
6652 out_fail_inode:
6653         unlock_new_inode(inode);
6654         goto out_fail;
6655 }
6656
6657 /* Find next extent map of a given extent map, caller needs to ensure locks */
6658 static struct extent_map *next_extent_map(struct extent_map *em)
6659 {
6660         struct rb_node *next;
6661
6662         next = rb_next(&em->rb_node);
6663         if (!next)
6664                 return NULL;
6665         return container_of(next, struct extent_map, rb_node);
6666 }
6667
6668 static struct extent_map *prev_extent_map(struct extent_map *em)
6669 {
6670         struct rb_node *prev;
6671
6672         prev = rb_prev(&em->rb_node);
6673         if (!prev)
6674                 return NULL;
6675         return container_of(prev, struct extent_map, rb_node);
6676 }
6677
6678 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6679  * the existing extent is the nearest extent to map_start,
6680  * and an extent that you want to insert, deal with overlap and insert
6681  * the best fitted new extent into the tree.
6682  */
6683 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6684                                 struct extent_map *existing,
6685                                 struct extent_map *em,
6686                                 u64 map_start)
6687 {
6688         struct extent_map *prev;
6689         struct extent_map *next;
6690         u64 start;
6691         u64 end;
6692         u64 start_diff;
6693
6694         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6695
6696         if (existing->start > map_start) {
6697                 next = existing;
6698                 prev = prev_extent_map(next);
6699         } else {
6700                 prev = existing;
6701                 next = next_extent_map(prev);
6702         }
6703
6704         start = prev ? extent_map_end(prev) : em->start;
6705         start = max_t(u64, start, em->start);
6706         end = next ? next->start : extent_map_end(em);
6707         end = min_t(u64, end, extent_map_end(em));
6708         start_diff = start - em->start;
6709         em->start = start;
6710         em->len = end - start;
6711         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6712             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6713                 em->block_start += start_diff;
6714                 em->block_len -= start_diff;
6715         }
6716         return add_extent_mapping(em_tree, em, 0);
6717 }
6718
6719 static noinline int uncompress_inline(struct btrfs_path *path,
6720                                       struct page *page,
6721                                       size_t pg_offset, u64 extent_offset,
6722                                       struct btrfs_file_extent_item *item)
6723 {
6724         int ret;
6725         struct extent_buffer *leaf = path->nodes[0];
6726         char *tmp;
6727         size_t max_size;
6728         unsigned long inline_size;
6729         unsigned long ptr;
6730         int compress_type;
6731
6732         WARN_ON(pg_offset != 0);
6733         compress_type = btrfs_file_extent_compression(leaf, item);
6734         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6735         inline_size = btrfs_file_extent_inline_item_len(leaf,
6736                                         btrfs_item_nr(path->slots[0]));
6737         tmp = kmalloc(inline_size, GFP_NOFS);
6738         if (!tmp)
6739                 return -ENOMEM;
6740         ptr = btrfs_file_extent_inline_start(item);
6741
6742         read_extent_buffer(leaf, tmp, ptr, inline_size);
6743
6744         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6745         ret = btrfs_decompress(compress_type, tmp, page,
6746                                extent_offset, inline_size, max_size);
6747         kfree(tmp);
6748         return ret;
6749 }
6750
6751 /*
6752  * a bit scary, this does extent mapping from logical file offset to the disk.
6753  * the ugly parts come from merging extents from the disk with the in-ram
6754  * representation.  This gets more complex because of the data=ordered code,
6755  * where the in-ram extents might be locked pending data=ordered completion.
6756  *
6757  * This also copies inline extents directly into the page.
6758  */
6759
6760 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6761                                     size_t pg_offset, u64 start, u64 len,
6762                                     int create)
6763 {
6764         int ret;
6765         int err = 0;
6766         u64 extent_start = 0;
6767         u64 extent_end = 0;
6768         u64 objectid = btrfs_ino(inode);
6769         u32 found_type;
6770         struct btrfs_path *path = NULL;
6771         struct btrfs_root *root = BTRFS_I(inode)->root;
6772         struct btrfs_file_extent_item *item;
6773         struct extent_buffer *leaf;
6774         struct btrfs_key found_key;
6775         struct extent_map *em = NULL;
6776         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6777         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6778         struct btrfs_trans_handle *trans = NULL;
6779         const bool new_inline = !page || create;
6780
6781 again:
6782         read_lock(&em_tree->lock);
6783         em = lookup_extent_mapping(em_tree, start, len);
6784         if (em)
6785                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6786         read_unlock(&em_tree->lock);
6787
6788         if (em) {
6789                 if (em->start > start || em->start + em->len <= start)
6790                         free_extent_map(em);
6791                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6792                         free_extent_map(em);
6793                 else
6794                         goto out;
6795         }
6796         em = alloc_extent_map();
6797         if (!em) {
6798                 err = -ENOMEM;
6799                 goto out;
6800         }
6801         em->bdev = root->fs_info->fs_devices->latest_bdev;
6802         em->start = EXTENT_MAP_HOLE;
6803         em->orig_start = EXTENT_MAP_HOLE;
6804         em->len = (u64)-1;
6805         em->block_len = (u64)-1;
6806
6807         if (!path) {
6808                 path = btrfs_alloc_path();
6809                 if (!path) {
6810                         err = -ENOMEM;
6811                         goto out;
6812                 }
6813                 /*
6814                  * Chances are we'll be called again, so go ahead and do
6815                  * readahead
6816                  */
6817                 path->reada = READA_FORWARD;
6818         }
6819
6820         ret = btrfs_lookup_file_extent(trans, root, path,
6821                                        objectid, start, trans != NULL);
6822         if (ret < 0) {
6823                 err = ret;
6824                 goto out;
6825         }
6826
6827         if (ret != 0) {
6828                 if (path->slots[0] == 0)
6829                         goto not_found;
6830                 path->slots[0]--;
6831         }
6832
6833         leaf = path->nodes[0];
6834         item = btrfs_item_ptr(leaf, path->slots[0],
6835                               struct btrfs_file_extent_item);
6836         /* are we inside the extent that was found? */
6837         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6838         found_type = found_key.type;
6839         if (found_key.objectid != objectid ||
6840             found_type != BTRFS_EXTENT_DATA_KEY) {
6841                 /*
6842                  * If we backup past the first extent we want to move forward
6843                  * and see if there is an extent in front of us, otherwise we'll
6844                  * say there is a hole for our whole search range which can
6845                  * cause problems.
6846                  */
6847                 extent_end = start;
6848                 goto next;
6849         }
6850
6851         found_type = btrfs_file_extent_type(leaf, item);
6852         extent_start = found_key.offset;
6853         if (found_type == BTRFS_FILE_EXTENT_REG ||
6854             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6855                 extent_end = extent_start +
6856                        btrfs_file_extent_num_bytes(leaf, item);
6857         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6858                 size_t size;
6859                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6860                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6861         }
6862 next:
6863         if (start >= extent_end) {
6864                 path->slots[0]++;
6865                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6866                         ret = btrfs_next_leaf(root, path);
6867                         if (ret < 0) {
6868                                 err = ret;
6869                                 goto out;
6870                         }
6871                         if (ret > 0)
6872                                 goto not_found;
6873                         leaf = path->nodes[0];
6874                 }
6875                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6876                 if (found_key.objectid != objectid ||
6877                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6878                         goto not_found;
6879                 if (start + len <= found_key.offset)
6880                         goto not_found;
6881                 if (start > found_key.offset)
6882                         goto next;
6883                 em->start = start;
6884                 em->orig_start = start;
6885                 em->len = found_key.offset - start;
6886                 goto not_found_em;
6887         }
6888
6889         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6890
6891         if (found_type == BTRFS_FILE_EXTENT_REG ||
6892             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6893                 goto insert;
6894         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6895                 unsigned long ptr;
6896                 char *map;
6897                 size_t size;
6898                 size_t extent_offset;
6899                 size_t copy_size;
6900
6901                 if (new_inline)
6902                         goto out;
6903
6904                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6905                 extent_offset = page_offset(page) + pg_offset - extent_start;
6906                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6907                                   size - extent_offset);
6908                 em->start = extent_start + extent_offset;
6909                 em->len = ALIGN(copy_size, root->sectorsize);
6910                 em->orig_block_len = em->len;
6911                 em->orig_start = em->start;
6912                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6913                 if (create == 0 && !PageUptodate(page)) {
6914                         if (btrfs_file_extent_compression(leaf, item) !=
6915                             BTRFS_COMPRESS_NONE) {
6916                                 ret = uncompress_inline(path, page, pg_offset,
6917                                                         extent_offset, item);
6918                                 if (ret) {
6919                                         err = ret;
6920                                         goto out;
6921                                 }
6922                         } else {
6923                                 map = kmap(page);
6924                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6925                                                    copy_size);
6926                                 if (pg_offset + copy_size < PAGE_SIZE) {
6927                                         memset(map + pg_offset + copy_size, 0,
6928                                                PAGE_SIZE - pg_offset -
6929                                                copy_size);
6930                                 }
6931                                 kunmap(page);
6932                         }
6933                         flush_dcache_page(page);
6934                 } else if (create && PageUptodate(page)) {
6935                         BUG();
6936                         if (!trans) {
6937                                 kunmap(page);
6938                                 free_extent_map(em);
6939                                 em = NULL;
6940
6941                                 btrfs_release_path(path);
6942                                 trans = btrfs_join_transaction(root);
6943
6944                                 if (IS_ERR(trans))
6945                                         return ERR_CAST(trans);
6946                                 goto again;
6947                         }
6948                         map = kmap(page);
6949                         write_extent_buffer(leaf, map + pg_offset, ptr,
6950                                             copy_size);
6951                         kunmap(page);
6952                         btrfs_mark_buffer_dirty(leaf);
6953                 }
6954                 set_extent_uptodate(io_tree, em->start,
6955                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6956                 goto insert;
6957         }
6958 not_found:
6959         em->start = start;
6960         em->orig_start = start;
6961         em->len = len;
6962 not_found_em:
6963         em->block_start = EXTENT_MAP_HOLE;
6964         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6965 insert:
6966         btrfs_release_path(path);
6967         if (em->start > start || extent_map_end(em) <= start) {
6968                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6969                         em->start, em->len, start, len);
6970                 err = -EIO;
6971                 goto out;
6972         }
6973
6974         err = 0;
6975         write_lock(&em_tree->lock);
6976         ret = add_extent_mapping(em_tree, em, 0);
6977         /* it is possible that someone inserted the extent into the tree
6978          * while we had the lock dropped.  It is also possible that
6979          * an overlapping map exists in the tree
6980          */
6981         if (ret == -EEXIST) {
6982                 struct extent_map *existing;
6983
6984                 ret = 0;
6985
6986                 existing = search_extent_mapping(em_tree, start, len);
6987                 /*
6988                  * existing will always be non-NULL, since there must be
6989                  * extent causing the -EEXIST.
6990                  */
6991                 if (existing->start == em->start &&
6992                     extent_map_end(existing) == extent_map_end(em) &&
6993                     em->block_start == existing->block_start) {
6994                         /*
6995                          * these two extents are the same, it happens
6996                          * with inlines especially
6997                          */
6998                         free_extent_map(em);
6999                         em = existing;
7000                         err = 0;
7001
7002                 } else if (start >= extent_map_end(existing) ||
7003                     start <= existing->start) {
7004                         /*
7005                          * The existing extent map is the one nearest to
7006                          * the [start, start + len) range which overlaps
7007                          */
7008                         err = merge_extent_mapping(em_tree, existing,
7009                                                    em, start);
7010                         free_extent_map(existing);
7011                         if (err) {
7012                                 free_extent_map(em);
7013                                 em = NULL;
7014                         }
7015                 } else {
7016                         free_extent_map(em);
7017                         em = existing;
7018                         err = 0;
7019                 }
7020         }
7021         write_unlock(&em_tree->lock);
7022 out:
7023
7024         trace_btrfs_get_extent(root, em);
7025
7026         btrfs_free_path(path);
7027         if (trans) {
7028                 ret = btrfs_end_transaction(trans, root);
7029                 if (!err)
7030                         err = ret;
7031         }
7032         if (err) {
7033                 free_extent_map(em);
7034                 return ERR_PTR(err);
7035         }
7036         BUG_ON(!em); /* Error is always set */
7037         return em;
7038 }
7039
7040 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7041                                            size_t pg_offset, u64 start, u64 len,
7042                                            int create)
7043 {
7044         struct extent_map *em;
7045         struct extent_map *hole_em = NULL;
7046         u64 range_start = start;
7047         u64 end;
7048         u64 found;
7049         u64 found_end;
7050         int err = 0;
7051
7052         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7053         if (IS_ERR(em))
7054                 return em;
7055         if (em) {
7056                 /*
7057                  * if our em maps to
7058                  * -  a hole or
7059                  * -  a pre-alloc extent,
7060                  * there might actually be delalloc bytes behind it.
7061                  */
7062                 if (em->block_start != EXTENT_MAP_HOLE &&
7063                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7064                         return em;
7065                 else
7066                         hole_em = em;
7067         }
7068
7069         /* check to see if we've wrapped (len == -1 or similar) */
7070         end = start + len;
7071         if (end < start)
7072                 end = (u64)-1;
7073         else
7074                 end -= 1;
7075
7076         em = NULL;
7077
7078         /* ok, we didn't find anything, lets look for delalloc */
7079         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7080                                  end, len, EXTENT_DELALLOC, 1);
7081         found_end = range_start + found;
7082         if (found_end < range_start)
7083                 found_end = (u64)-1;
7084
7085         /*
7086          * we didn't find anything useful, return
7087          * the original results from get_extent()
7088          */
7089         if (range_start > end || found_end <= start) {
7090                 em = hole_em;
7091                 hole_em = NULL;
7092                 goto out;
7093         }
7094
7095         /* adjust the range_start to make sure it doesn't
7096          * go backwards from the start they passed in
7097          */
7098         range_start = max(start, range_start);
7099         found = found_end - range_start;
7100
7101         if (found > 0) {
7102                 u64 hole_start = start;
7103                 u64 hole_len = len;
7104
7105                 em = alloc_extent_map();
7106                 if (!em) {
7107                         err = -ENOMEM;
7108                         goto out;
7109                 }
7110                 /*
7111                  * when btrfs_get_extent can't find anything it
7112                  * returns one huge hole
7113                  *
7114                  * make sure what it found really fits our range, and
7115                  * adjust to make sure it is based on the start from
7116                  * the caller
7117                  */
7118                 if (hole_em) {
7119                         u64 calc_end = extent_map_end(hole_em);
7120
7121                         if (calc_end <= start || (hole_em->start > end)) {
7122                                 free_extent_map(hole_em);
7123                                 hole_em = NULL;
7124                         } else {
7125                                 hole_start = max(hole_em->start, start);
7126                                 hole_len = calc_end - hole_start;
7127                         }
7128                 }
7129                 em->bdev = NULL;
7130                 if (hole_em && range_start > hole_start) {
7131                         /* our hole starts before our delalloc, so we
7132                          * have to return just the parts of the hole
7133                          * that go until  the delalloc starts
7134                          */
7135                         em->len = min(hole_len,
7136                                       range_start - hole_start);
7137                         em->start = hole_start;
7138                         em->orig_start = hole_start;
7139                         /*
7140                          * don't adjust block start at all,
7141                          * it is fixed at EXTENT_MAP_HOLE
7142                          */
7143                         em->block_start = hole_em->block_start;
7144                         em->block_len = hole_len;
7145                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7146                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7147                 } else {
7148                         em->start = range_start;
7149                         em->len = found;
7150                         em->orig_start = range_start;
7151                         em->block_start = EXTENT_MAP_DELALLOC;
7152                         em->block_len = found;
7153                 }
7154         } else if (hole_em) {
7155                 return hole_em;
7156         }
7157 out:
7158
7159         free_extent_map(hole_em);
7160         if (err) {
7161                 free_extent_map(em);
7162                 return ERR_PTR(err);
7163         }
7164         return em;
7165 }
7166
7167 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7168                                                   const u64 start,
7169                                                   const u64 len,
7170                                                   const u64 orig_start,
7171                                                   const u64 block_start,
7172                                                   const u64 block_len,
7173                                                   const u64 orig_block_len,
7174                                                   const u64 ram_bytes,
7175                                                   const int type)
7176 {
7177         struct extent_map *em = NULL;
7178         int ret;
7179
7180         down_read(&BTRFS_I(inode)->dio_sem);
7181         if (type != BTRFS_ORDERED_NOCOW) {
7182                 em = create_pinned_em(inode, start, len, orig_start,
7183                                       block_start, block_len, orig_block_len,
7184                                       ram_bytes, type);
7185                 if (IS_ERR(em))
7186                         goto out;
7187         }
7188         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7189                                            len, block_len, type);
7190         if (ret) {
7191                 if (em) {
7192                         free_extent_map(em);
7193                         btrfs_drop_extent_cache(inode, start,
7194                                                 start + len - 1, 0);
7195                 }
7196                 em = ERR_PTR(ret);
7197         }
7198  out:
7199         up_read(&BTRFS_I(inode)->dio_sem);
7200
7201         return em;
7202 }
7203
7204 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7205                                                   u64 start, u64 len)
7206 {
7207         struct btrfs_root *root = BTRFS_I(inode)->root;
7208         struct extent_map *em;
7209         struct btrfs_key ins;
7210         u64 alloc_hint;
7211         int ret;
7212
7213         alloc_hint = get_extent_allocation_hint(inode, start, len);
7214         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7215                                    alloc_hint, &ins, 1, 1);
7216         if (ret)
7217                 return ERR_PTR(ret);
7218
7219         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7220                                      ins.objectid, ins.offset, ins.offset,
7221                                      ins.offset, 0);
7222         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
7223         if (IS_ERR(em))
7224                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7225
7226         return em;
7227 }
7228
7229 /*
7230  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7231  * block must be cow'd
7232  */
7233 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7234                               u64 *orig_start, u64 *orig_block_len,
7235                               u64 *ram_bytes)
7236 {
7237         struct btrfs_trans_handle *trans;
7238         struct btrfs_path *path;
7239         int ret;
7240         struct extent_buffer *leaf;
7241         struct btrfs_root *root = BTRFS_I(inode)->root;
7242         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7243         struct btrfs_file_extent_item *fi;
7244         struct btrfs_key key;
7245         u64 disk_bytenr;
7246         u64 backref_offset;
7247         u64 extent_end;
7248         u64 num_bytes;
7249         int slot;
7250         int found_type;
7251         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7252
7253         path = btrfs_alloc_path();
7254         if (!path)
7255                 return -ENOMEM;
7256
7257         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7258                                        offset, 0);
7259         if (ret < 0)
7260                 goto out;
7261
7262         slot = path->slots[0];
7263         if (ret == 1) {
7264                 if (slot == 0) {
7265                         /* can't find the item, must cow */
7266                         ret = 0;
7267                         goto out;
7268                 }
7269                 slot--;
7270         }
7271         ret = 0;
7272         leaf = path->nodes[0];
7273         btrfs_item_key_to_cpu(leaf, &key, slot);
7274         if (key.objectid != btrfs_ino(inode) ||
7275             key.type != BTRFS_EXTENT_DATA_KEY) {
7276                 /* not our file or wrong item type, must cow */
7277                 goto out;
7278         }
7279
7280         if (key.offset > offset) {
7281                 /* Wrong offset, must cow */
7282                 goto out;
7283         }
7284
7285         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7286         found_type = btrfs_file_extent_type(leaf, fi);
7287         if (found_type != BTRFS_FILE_EXTENT_REG &&
7288             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7289                 /* not a regular extent, must cow */
7290                 goto out;
7291         }
7292
7293         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7294                 goto out;
7295
7296         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7297         if (extent_end <= offset)
7298                 goto out;
7299
7300         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7301         if (disk_bytenr == 0)
7302                 goto out;
7303
7304         if (btrfs_file_extent_compression(leaf, fi) ||
7305             btrfs_file_extent_encryption(leaf, fi) ||
7306             btrfs_file_extent_other_encoding(leaf, fi))
7307                 goto out;
7308
7309         backref_offset = btrfs_file_extent_offset(leaf, fi);
7310
7311         if (orig_start) {
7312                 *orig_start = key.offset - backref_offset;
7313                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7314                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7315         }
7316
7317         if (btrfs_extent_readonly(root, disk_bytenr))
7318                 goto out;
7319
7320         num_bytes = min(offset + *len, extent_end) - offset;
7321         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7322                 u64 range_end;
7323
7324                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7325                 ret = test_range_bit(io_tree, offset, range_end,
7326                                      EXTENT_DELALLOC, 0, NULL);
7327                 if (ret) {
7328                         ret = -EAGAIN;
7329                         goto out;
7330                 }
7331         }
7332
7333         btrfs_release_path(path);
7334
7335         /*
7336          * look for other files referencing this extent, if we
7337          * find any we must cow
7338          */
7339         trans = btrfs_join_transaction(root);
7340         if (IS_ERR(trans)) {
7341                 ret = 0;
7342                 goto out;
7343         }
7344
7345         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7346                                     key.offset - backref_offset, disk_bytenr);
7347         btrfs_end_transaction(trans, root);
7348         if (ret) {
7349                 ret = 0;
7350                 goto out;
7351         }
7352
7353         /*
7354          * adjust disk_bytenr and num_bytes to cover just the bytes
7355          * in this extent we are about to write.  If there
7356          * are any csums in that range we have to cow in order
7357          * to keep the csums correct
7358          */
7359         disk_bytenr += backref_offset;
7360         disk_bytenr += offset - key.offset;
7361         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7362                                 goto out;
7363         /*
7364          * all of the above have passed, it is safe to overwrite this extent
7365          * without cow
7366          */
7367         *len = num_bytes;
7368         ret = 1;
7369 out:
7370         btrfs_free_path(path);
7371         return ret;
7372 }
7373
7374 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7375 {
7376         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7377         int found = false;
7378         void **pagep = NULL;
7379         struct page *page = NULL;
7380         int start_idx;
7381         int end_idx;
7382
7383         start_idx = start >> PAGE_SHIFT;
7384
7385         /*
7386          * end is the last byte in the last page.  end == start is legal
7387          */
7388         end_idx = end >> PAGE_SHIFT;
7389
7390         rcu_read_lock();
7391
7392         /* Most of the code in this while loop is lifted from
7393          * find_get_page.  It's been modified to begin searching from a
7394          * page and return just the first page found in that range.  If the
7395          * found idx is less than or equal to the end idx then we know that
7396          * a page exists.  If no pages are found or if those pages are
7397          * outside of the range then we're fine (yay!) */
7398         while (page == NULL &&
7399                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7400                 page = radix_tree_deref_slot(pagep);
7401                 if (unlikely(!page))
7402                         break;
7403
7404                 if (radix_tree_exception(page)) {
7405                         if (radix_tree_deref_retry(page)) {
7406                                 page = NULL;
7407                                 continue;
7408                         }
7409                         /*
7410                          * Otherwise, shmem/tmpfs must be storing a swap entry
7411                          * here as an exceptional entry: so return it without
7412                          * attempting to raise page count.
7413                          */
7414                         page = NULL;
7415                         break; /* TODO: Is this relevant for this use case? */
7416                 }
7417
7418                 if (!page_cache_get_speculative(page)) {
7419                         page = NULL;
7420                         continue;
7421                 }
7422
7423                 /*
7424                  * Has the page moved?
7425                  * This is part of the lockless pagecache protocol. See
7426                  * include/linux/pagemap.h for details.
7427                  */
7428                 if (unlikely(page != *pagep)) {
7429                         put_page(page);
7430                         page = NULL;
7431                 }
7432         }
7433
7434         if (page) {
7435                 if (page->index <= end_idx)
7436                         found = true;
7437                 put_page(page);
7438         }
7439
7440         rcu_read_unlock();
7441         return found;
7442 }
7443
7444 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7445                               struct extent_state **cached_state, int writing)
7446 {
7447         struct btrfs_ordered_extent *ordered;
7448         int ret = 0;
7449
7450         while (1) {
7451                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7452                                  cached_state);
7453                 /*
7454                  * We're concerned with the entire range that we're going to be
7455                  * doing DIO to, so we need to make sure there's no ordered
7456                  * extents in this range.
7457                  */
7458                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7459                                                      lockend - lockstart + 1);
7460
7461                 /*
7462                  * We need to make sure there are no buffered pages in this
7463                  * range either, we could have raced between the invalidate in
7464                  * generic_file_direct_write and locking the extent.  The
7465                  * invalidate needs to happen so that reads after a write do not
7466                  * get stale data.
7467                  */
7468                 if (!ordered &&
7469                     (!writing ||
7470                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7471                         break;
7472
7473                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7474                                      cached_state, GFP_NOFS);
7475
7476                 if (ordered) {
7477                         /*
7478                          * If we are doing a DIO read and the ordered extent we
7479                          * found is for a buffered write, we can not wait for it
7480                          * to complete and retry, because if we do so we can
7481                          * deadlock with concurrent buffered writes on page
7482                          * locks. This happens only if our DIO read covers more
7483                          * than one extent map, if at this point has already
7484                          * created an ordered extent for a previous extent map
7485                          * and locked its range in the inode's io tree, and a
7486                          * concurrent write against that previous extent map's
7487                          * range and this range started (we unlock the ranges
7488                          * in the io tree only when the bios complete and
7489                          * buffered writes always lock pages before attempting
7490                          * to lock range in the io tree).
7491                          */
7492                         if (writing ||
7493                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7494                                 btrfs_start_ordered_extent(inode, ordered, 1);
7495                         else
7496                                 ret = -ENOTBLK;
7497                         btrfs_put_ordered_extent(ordered);
7498                 } else {
7499                         /*
7500                          * We could trigger writeback for this range (and wait
7501                          * for it to complete) and then invalidate the pages for
7502                          * this range (through invalidate_inode_pages2_range()),
7503                          * but that can lead us to a deadlock with a concurrent
7504                          * call to readpages() (a buffered read or a defrag call
7505                          * triggered a readahead) on a page lock due to an
7506                          * ordered dio extent we created before but did not have
7507                          * yet a corresponding bio submitted (whence it can not
7508                          * complete), which makes readpages() wait for that
7509                          * ordered extent to complete while holding a lock on
7510                          * that page.
7511                          */
7512                         ret = -ENOTBLK;
7513                 }
7514
7515                 if (ret)
7516                         break;
7517
7518                 cond_resched();
7519         }
7520
7521         return ret;
7522 }
7523
7524 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7525                                            u64 len, u64 orig_start,
7526                                            u64 block_start, u64 block_len,
7527                                            u64 orig_block_len, u64 ram_bytes,
7528                                            int type)
7529 {
7530         struct extent_map_tree *em_tree;
7531         struct extent_map *em;
7532         struct btrfs_root *root = BTRFS_I(inode)->root;
7533         int ret;
7534
7535         em_tree = &BTRFS_I(inode)->extent_tree;
7536         em = alloc_extent_map();
7537         if (!em)
7538                 return ERR_PTR(-ENOMEM);
7539
7540         em->start = start;
7541         em->orig_start = orig_start;
7542         em->mod_start = start;
7543         em->mod_len = len;
7544         em->len = len;
7545         em->block_len = block_len;
7546         em->block_start = block_start;
7547         em->bdev = root->fs_info->fs_devices->latest_bdev;
7548         em->orig_block_len = orig_block_len;
7549         em->ram_bytes = ram_bytes;
7550         em->generation = -1;
7551         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7552         if (type == BTRFS_ORDERED_PREALLOC)
7553                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7554
7555         do {
7556                 btrfs_drop_extent_cache(inode, em->start,
7557                                 em->start + em->len - 1, 0);
7558                 write_lock(&em_tree->lock);
7559                 ret = add_extent_mapping(em_tree, em, 1);
7560                 write_unlock(&em_tree->lock);
7561         } while (ret == -EEXIST);
7562
7563         if (ret) {
7564                 free_extent_map(em);
7565                 return ERR_PTR(ret);
7566         }
7567
7568         return em;
7569 }
7570
7571 static void adjust_dio_outstanding_extents(struct inode *inode,
7572                                            struct btrfs_dio_data *dio_data,
7573                                            const u64 len)
7574 {
7575         unsigned num_extents;
7576
7577         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7578                                            BTRFS_MAX_EXTENT_SIZE);
7579         /*
7580          * If we have an outstanding_extents count still set then we're
7581          * within our reservation, otherwise we need to adjust our inode
7582          * counter appropriately.
7583          */
7584         if (dio_data->outstanding_extents) {
7585                 dio_data->outstanding_extents -= num_extents;
7586         } else {
7587                 spin_lock(&BTRFS_I(inode)->lock);
7588                 BTRFS_I(inode)->outstanding_extents += num_extents;
7589                 spin_unlock(&BTRFS_I(inode)->lock);
7590         }
7591 }
7592
7593 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7594                                    struct buffer_head *bh_result, int create)
7595 {
7596         struct extent_map *em;
7597         struct btrfs_root *root = BTRFS_I(inode)->root;
7598         struct extent_state *cached_state = NULL;
7599         struct btrfs_dio_data *dio_data = NULL;
7600         u64 start = iblock << inode->i_blkbits;
7601         u64 lockstart, lockend;
7602         u64 len = bh_result->b_size;
7603         int unlock_bits = EXTENT_LOCKED;
7604         int ret = 0;
7605
7606         if (create)
7607                 unlock_bits |= EXTENT_DIRTY;
7608         else
7609                 len = min_t(u64, len, root->sectorsize);
7610
7611         lockstart = start;
7612         lockend = start + len - 1;
7613
7614         if (current->journal_info) {
7615                 /*
7616                  * Need to pull our outstanding extents and set journal_info to NULL so
7617                  * that anything that needs to check if there's a transaction doesn't get
7618                  * confused.
7619                  */
7620                 dio_data = current->journal_info;
7621                 current->journal_info = NULL;
7622         }
7623
7624         /*
7625          * If this errors out it's because we couldn't invalidate pagecache for
7626          * this range and we need to fallback to buffered.
7627          */
7628         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7629                                create)) {
7630                 ret = -ENOTBLK;
7631                 goto err;
7632         }
7633
7634         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7635         if (IS_ERR(em)) {
7636                 ret = PTR_ERR(em);
7637                 goto unlock_err;
7638         }
7639
7640         /*
7641          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7642          * io.  INLINE is special, and we could probably kludge it in here, but
7643          * it's still buffered so for safety lets just fall back to the generic
7644          * buffered path.
7645          *
7646          * For COMPRESSED we _have_ to read the entire extent in so we can
7647          * decompress it, so there will be buffering required no matter what we
7648          * do, so go ahead and fallback to buffered.
7649          *
7650          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7651          * to buffered IO.  Don't blame me, this is the price we pay for using
7652          * the generic code.
7653          */
7654         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7655             em->block_start == EXTENT_MAP_INLINE) {
7656                 free_extent_map(em);
7657                 ret = -ENOTBLK;
7658                 goto unlock_err;
7659         }
7660
7661         /* Just a good old fashioned hole, return */
7662         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7663                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7664                 free_extent_map(em);
7665                 goto unlock_err;
7666         }
7667
7668         /*
7669          * We don't allocate a new extent in the following cases
7670          *
7671          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7672          * existing extent.
7673          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7674          * just use the extent.
7675          *
7676          */
7677         if (!create) {
7678                 len = min(len, em->len - (start - em->start));
7679                 lockstart = start + len;
7680                 goto unlock;
7681         }
7682
7683         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7684             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7685              em->block_start != EXTENT_MAP_HOLE)) {
7686                 int type;
7687                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7688
7689                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7690                         type = BTRFS_ORDERED_PREALLOC;
7691                 else
7692                         type = BTRFS_ORDERED_NOCOW;
7693                 len = min(len, em->len - (start - em->start));
7694                 block_start = em->block_start + (start - em->start);
7695
7696                 if (can_nocow_extent(inode, start, &len, &orig_start,
7697                                      &orig_block_len, &ram_bytes) == 1 &&
7698                     btrfs_inc_nocow_writers(root->fs_info, block_start)) {
7699                         struct extent_map *em2;
7700
7701                         em2 = btrfs_create_dio_extent(inode, start, len,
7702                                                       orig_start, block_start,
7703                                                       len, orig_block_len,
7704                                                       ram_bytes, type);
7705                         btrfs_dec_nocow_writers(root->fs_info, block_start);
7706                         if (type == BTRFS_ORDERED_PREALLOC) {
7707                                 free_extent_map(em);
7708                                 em = em2;
7709                         }
7710                         if (em2 && IS_ERR(em2)) {
7711                                 ret = PTR_ERR(em2);
7712                                 goto unlock_err;
7713                         }
7714                         goto unlock;
7715                 }
7716         }
7717
7718         /*
7719          * this will cow the extent, reset the len in case we changed
7720          * it above
7721          */
7722         len = bh_result->b_size;
7723         free_extent_map(em);
7724         em = btrfs_new_extent_direct(inode, start, len);
7725         if (IS_ERR(em)) {
7726                 ret = PTR_ERR(em);
7727                 goto unlock_err;
7728         }
7729         len = min(len, em->len - (start - em->start));
7730 unlock:
7731         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7732                 inode->i_blkbits;
7733         bh_result->b_size = len;
7734         bh_result->b_bdev = em->bdev;
7735         set_buffer_mapped(bh_result);
7736         if (create) {
7737                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7738                         set_buffer_new(bh_result);
7739
7740                 /*
7741                  * Need to update the i_size under the extent lock so buffered
7742                  * readers will get the updated i_size when we unlock.
7743                  */
7744                 if (start + len > i_size_read(inode))
7745                         i_size_write(inode, start + len);
7746
7747                 adjust_dio_outstanding_extents(inode, dio_data, len);
7748                 btrfs_free_reserved_data_space(inode, start, len);
7749                 WARN_ON(dio_data->reserve < len);
7750                 dio_data->reserve -= len;
7751                 dio_data->unsubmitted_oe_range_end = start + len;
7752                 current->journal_info = dio_data;
7753         }
7754
7755         /*
7756          * In the case of write we need to clear and unlock the entire range,
7757          * in the case of read we need to unlock only the end area that we
7758          * aren't using if there is any left over space.
7759          */
7760         if (lockstart < lockend) {
7761                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7762                                  lockend, unlock_bits, 1, 0,
7763                                  &cached_state, GFP_NOFS);
7764         } else {
7765                 free_extent_state(cached_state);
7766         }
7767
7768         free_extent_map(em);
7769
7770         return 0;
7771
7772 unlock_err:
7773         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7774                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7775 err:
7776         if (dio_data)
7777                 current->journal_info = dio_data;
7778         /*
7779          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7780          * write less data then expected, so that we don't underflow our inode's
7781          * outstanding extents counter.
7782          */
7783         if (create && dio_data)
7784                 adjust_dio_outstanding_extents(inode, dio_data, len);
7785
7786         return ret;
7787 }
7788
7789 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7790                                         int rw, int mirror_num)
7791 {
7792         struct btrfs_root *root = BTRFS_I(inode)->root;
7793         int ret;
7794
7795         BUG_ON(rw & REQ_WRITE);
7796
7797         bio_get(bio);
7798
7799         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7800                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7801         if (ret)
7802                 goto err;
7803
7804         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7805 err:
7806         bio_put(bio);
7807         return ret;
7808 }
7809
7810 static int btrfs_check_dio_repairable(struct inode *inode,
7811                                       struct bio *failed_bio,
7812                                       struct io_failure_record *failrec,
7813                                       int failed_mirror)
7814 {
7815         int num_copies;
7816
7817         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7818                                       failrec->logical, failrec->len);
7819         if (num_copies == 1) {
7820                 /*
7821                  * we only have a single copy of the data, so don't bother with
7822                  * all the retry and error correction code that follows. no
7823                  * matter what the error is, it is very likely to persist.
7824                  */
7825                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7826                          num_copies, failrec->this_mirror, failed_mirror);
7827                 return 0;
7828         }
7829
7830         failrec->failed_mirror = failed_mirror;
7831         failrec->this_mirror++;
7832         if (failrec->this_mirror == failed_mirror)
7833                 failrec->this_mirror++;
7834
7835         if (failrec->this_mirror > num_copies) {
7836                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7837                          num_copies, failrec->this_mirror, failed_mirror);
7838                 return 0;
7839         }
7840
7841         return 1;
7842 }
7843
7844 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7845                         struct page *page, unsigned int pgoff,
7846                         u64 start, u64 end, int failed_mirror,
7847                         bio_end_io_t *repair_endio, void *repair_arg)
7848 {
7849         struct io_failure_record *failrec;
7850         struct bio *bio;
7851         int isector;
7852         int read_mode;
7853         int ret;
7854
7855         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7856
7857         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7858         if (ret)
7859                 return ret;
7860
7861         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7862                                          failed_mirror);
7863         if (!ret) {
7864                 free_io_failure(inode, failrec);
7865                 return -EIO;
7866         }
7867
7868         if ((failed_bio->bi_vcnt > 1)
7869                 || (failed_bio->bi_io_vec->bv_len
7870                         > BTRFS_I(inode)->root->sectorsize))
7871                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7872         else
7873                 read_mode = READ_SYNC;
7874
7875         isector = start - btrfs_io_bio(failed_bio)->logical;
7876         isector >>= inode->i_sb->s_blocksize_bits;
7877         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7878                                 pgoff, isector, repair_endio, repair_arg);
7879         if (!bio) {
7880                 free_io_failure(inode, failrec);
7881                 return -EIO;
7882         }
7883
7884         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7885                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7886                     read_mode, failrec->this_mirror, failrec->in_validation);
7887
7888         ret = submit_dio_repair_bio(inode, bio, read_mode,
7889                                     failrec->this_mirror);
7890         if (ret) {
7891                 free_io_failure(inode, failrec);
7892                 bio_put(bio);
7893         }
7894
7895         return ret;
7896 }
7897
7898 struct btrfs_retry_complete {
7899         struct completion done;
7900         struct inode *inode;
7901         u64 start;
7902         int uptodate;
7903 };
7904
7905 static void btrfs_retry_endio_nocsum(struct bio *bio)
7906 {
7907         struct btrfs_retry_complete *done = bio->bi_private;
7908         struct inode *inode;
7909         struct bio_vec *bvec;
7910         int i;
7911
7912         if (bio->bi_error)
7913                 goto end;
7914
7915         ASSERT(bio->bi_vcnt == 1);
7916         inode = bio->bi_io_vec->bv_page->mapping->host;
7917         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7918
7919         done->uptodate = 1;
7920         bio_for_each_segment_all(bvec, bio, i)
7921                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7922 end:
7923         complete(&done->done);
7924         bio_put(bio);
7925 }
7926
7927 static int __btrfs_correct_data_nocsum(struct inode *inode,
7928                                        struct btrfs_io_bio *io_bio)
7929 {
7930         struct btrfs_fs_info *fs_info;
7931         struct bio_vec *bvec;
7932         struct btrfs_retry_complete done;
7933         u64 start;
7934         unsigned int pgoff;
7935         u32 sectorsize;
7936         int nr_sectors;
7937         int i;
7938         int ret;
7939
7940         fs_info = BTRFS_I(inode)->root->fs_info;
7941         sectorsize = BTRFS_I(inode)->root->sectorsize;
7942
7943         start = io_bio->logical;
7944         done.inode = inode;
7945
7946         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7947                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
7948                 pgoff = bvec->bv_offset;
7949
7950 next_block_or_try_again:
7951                 done.uptodate = 0;
7952                 done.start = start;
7953                 init_completion(&done.done);
7954
7955                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
7956                                 pgoff, start, start + sectorsize - 1,
7957                                 io_bio->mirror_num,
7958                                 btrfs_retry_endio_nocsum, &done);
7959                 if (ret)
7960                         return ret;
7961
7962                 wait_for_completion(&done.done);
7963
7964                 if (!done.uptodate) {
7965                         /* We might have another mirror, so try again */
7966                         goto next_block_or_try_again;
7967                 }
7968
7969                 start += sectorsize;
7970
7971                 if (nr_sectors--) {
7972                         pgoff += sectorsize;
7973                         goto next_block_or_try_again;
7974                 }
7975         }
7976
7977         return 0;
7978 }
7979
7980 static void btrfs_retry_endio(struct bio *bio)
7981 {
7982         struct btrfs_retry_complete *done = bio->bi_private;
7983         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7984         struct inode *inode;
7985         struct bio_vec *bvec;
7986         u64 start;
7987         int uptodate;
7988         int ret;
7989         int i;
7990
7991         if (bio->bi_error)
7992                 goto end;
7993
7994         uptodate = 1;
7995
7996         start = done->start;
7997
7998         ASSERT(bio->bi_vcnt == 1);
7999         inode = bio->bi_io_vec->bv_page->mapping->host;
8000         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
8001
8002         bio_for_each_segment_all(bvec, bio, i) {
8003                 ret = __readpage_endio_check(done->inode, io_bio, i,
8004                                         bvec->bv_page, bvec->bv_offset,
8005                                         done->start, bvec->bv_len);
8006                 if (!ret)
8007                         clean_io_failure(done->inode, done->start,
8008                                         bvec->bv_page, bvec->bv_offset);
8009                 else
8010                         uptodate = 0;
8011         }
8012
8013         done->uptodate = uptodate;
8014 end:
8015         complete(&done->done);
8016         bio_put(bio);
8017 }
8018
8019 static int __btrfs_subio_endio_read(struct inode *inode,
8020                                     struct btrfs_io_bio *io_bio, int err)
8021 {
8022         struct btrfs_fs_info *fs_info;
8023         struct bio_vec *bvec;
8024         struct btrfs_retry_complete done;
8025         u64 start;
8026         u64 offset = 0;
8027         u32 sectorsize;
8028         int nr_sectors;
8029         unsigned int pgoff;
8030         int csum_pos;
8031         int i;
8032         int ret;
8033
8034         fs_info = BTRFS_I(inode)->root->fs_info;
8035         sectorsize = BTRFS_I(inode)->root->sectorsize;
8036
8037         err = 0;
8038         start = io_bio->logical;
8039         done.inode = inode;
8040
8041         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8042                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8043
8044                 pgoff = bvec->bv_offset;
8045 next_block:
8046                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8047                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8048                                         bvec->bv_page, pgoff, start,
8049                                         sectorsize);
8050                 if (likely(!ret))
8051                         goto next;
8052 try_again:
8053                 done.uptodate = 0;
8054                 done.start = start;
8055                 init_completion(&done.done);
8056
8057                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8058                                 pgoff, start, start + sectorsize - 1,
8059                                 io_bio->mirror_num,
8060                                 btrfs_retry_endio, &done);
8061                 if (ret) {
8062                         err = ret;
8063                         goto next;
8064                 }
8065
8066                 wait_for_completion(&done.done);
8067
8068                 if (!done.uptodate) {
8069                         /* We might have another mirror, so try again */
8070                         goto try_again;
8071                 }
8072 next:
8073                 offset += sectorsize;
8074                 start += sectorsize;
8075
8076                 ASSERT(nr_sectors);
8077
8078                 if (--nr_sectors) {
8079                         pgoff += sectorsize;
8080                         goto next_block;
8081                 }
8082         }
8083
8084         return err;
8085 }
8086
8087 static int btrfs_subio_endio_read(struct inode *inode,
8088                                   struct btrfs_io_bio *io_bio, int err)
8089 {
8090         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8091
8092         if (skip_csum) {
8093                 if (unlikely(err))
8094                         return __btrfs_correct_data_nocsum(inode, io_bio);
8095                 else
8096                         return 0;
8097         } else {
8098                 return __btrfs_subio_endio_read(inode, io_bio, err);
8099         }
8100 }
8101
8102 static void btrfs_endio_direct_read(struct bio *bio)
8103 {
8104         struct btrfs_dio_private *dip = bio->bi_private;
8105         struct inode *inode = dip->inode;
8106         struct bio *dio_bio;
8107         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8108         int err = bio->bi_error;
8109
8110         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8111                 err = btrfs_subio_endio_read(inode, io_bio, err);
8112
8113         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8114                       dip->logical_offset + dip->bytes - 1);
8115         dio_bio = dip->dio_bio;
8116
8117         kfree(dip);
8118
8119         dio_bio->bi_error = bio->bi_error;
8120         dio_end_io(dio_bio, bio->bi_error);
8121
8122         if (io_bio->end_io)
8123                 io_bio->end_io(io_bio, err);
8124         bio_put(bio);
8125 }
8126
8127 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8128                                                     const u64 offset,
8129                                                     const u64 bytes,
8130                                                     const int uptodate)
8131 {
8132         struct btrfs_root *root = BTRFS_I(inode)->root;
8133         struct btrfs_ordered_extent *ordered = NULL;
8134         u64 ordered_offset = offset;
8135         u64 ordered_bytes = bytes;
8136         int ret;
8137
8138 again:
8139         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8140                                                    &ordered_offset,
8141                                                    ordered_bytes,
8142                                                    uptodate);
8143         if (!ret)
8144                 goto out_test;
8145
8146         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8147                         finish_ordered_fn, NULL, NULL);
8148         btrfs_queue_work(root->fs_info->endio_write_workers,
8149                          &ordered->work);
8150 out_test:
8151         /*
8152          * our bio might span multiple ordered extents.  If we haven't
8153          * completed the accounting for the whole dio, go back and try again
8154          */
8155         if (ordered_offset < offset + bytes) {
8156                 ordered_bytes = offset + bytes - ordered_offset;
8157                 ordered = NULL;
8158                 goto again;
8159         }
8160 }
8161
8162 static void btrfs_endio_direct_write(struct bio *bio)
8163 {
8164         struct btrfs_dio_private *dip = bio->bi_private;
8165         struct bio *dio_bio = dip->dio_bio;
8166
8167         btrfs_endio_direct_write_update_ordered(dip->inode,
8168                                                 dip->logical_offset,
8169                                                 dip->bytes,
8170                                                 !bio->bi_error);
8171
8172         kfree(dip);
8173
8174         dio_bio->bi_error = bio->bi_error;
8175         dio_end_io(dio_bio, bio->bi_error);
8176         bio_put(bio);
8177 }
8178
8179 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8180                                     struct bio *bio, int mirror_num,
8181                                     unsigned long bio_flags, u64 offset)
8182 {
8183         int ret;
8184         struct btrfs_root *root = BTRFS_I(inode)->root;
8185         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8186         BUG_ON(ret); /* -ENOMEM */
8187         return 0;
8188 }
8189
8190 static void btrfs_end_dio_bio(struct bio *bio)
8191 {
8192         struct btrfs_dio_private *dip = bio->bi_private;
8193         int err = bio->bi_error;
8194
8195         if (err)
8196                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8197                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8198                            btrfs_ino(dip->inode), bio->bi_rw,
8199                            (unsigned long long)bio->bi_iter.bi_sector,
8200                            bio->bi_iter.bi_size, err);
8201
8202         if (dip->subio_endio)
8203                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8204
8205         if (err) {
8206                 dip->errors = 1;
8207
8208                 /*
8209                  * before atomic variable goto zero, we must make sure
8210                  * dip->errors is perceived to be set.
8211                  */
8212                 smp_mb__before_atomic();
8213         }
8214
8215         /* if there are more bios still pending for this dio, just exit */
8216         if (!atomic_dec_and_test(&dip->pending_bios))
8217                 goto out;
8218
8219         if (dip->errors) {
8220                 bio_io_error(dip->orig_bio);
8221         } else {
8222                 dip->dio_bio->bi_error = 0;
8223                 bio_endio(dip->orig_bio);
8224         }
8225 out:
8226         bio_put(bio);
8227 }
8228
8229 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8230                                        u64 first_sector, gfp_t gfp_flags)
8231 {
8232         struct bio *bio;
8233         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8234         if (bio)
8235                 bio_associate_current(bio);
8236         return bio;
8237 }
8238
8239 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8240                                                  struct inode *inode,
8241                                                  struct btrfs_dio_private *dip,
8242                                                  struct bio *bio,
8243                                                  u64 file_offset)
8244 {
8245         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8246         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8247         int ret;
8248
8249         /*
8250          * We load all the csum data we need when we submit
8251          * the first bio to reduce the csum tree search and
8252          * contention.
8253          */
8254         if (dip->logical_offset == file_offset) {
8255                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8256                                                 file_offset);
8257                 if (ret)
8258                         return ret;
8259         }
8260
8261         if (bio == dip->orig_bio)
8262                 return 0;
8263
8264         file_offset -= dip->logical_offset;
8265         file_offset >>= inode->i_sb->s_blocksize_bits;
8266         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8267
8268         return 0;
8269 }
8270
8271 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8272                                          int rw, u64 file_offset, int skip_sum,
8273                                          int async_submit)
8274 {
8275         struct btrfs_dio_private *dip = bio->bi_private;
8276         int write = rw & REQ_WRITE;
8277         struct btrfs_root *root = BTRFS_I(inode)->root;
8278         int ret;
8279
8280         if (async_submit)
8281                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8282
8283         bio_get(bio);
8284
8285         if (!write) {
8286                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8287                                 BTRFS_WQ_ENDIO_DATA);
8288                 if (ret)
8289                         goto err;
8290         }
8291
8292         if (skip_sum)
8293                 goto map;
8294
8295         if (write && async_submit) {
8296                 ret = btrfs_wq_submit_bio(root->fs_info,
8297                                    inode, rw, bio, 0, 0,
8298                                    file_offset,
8299                                    __btrfs_submit_bio_start_direct_io,
8300                                    __btrfs_submit_bio_done);
8301                 goto err;
8302         } else if (write) {
8303                 /*
8304                  * If we aren't doing async submit, calculate the csum of the
8305                  * bio now.
8306                  */
8307                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8308                 if (ret)
8309                         goto err;
8310         } else {
8311                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8312                                                      file_offset);
8313                 if (ret)
8314                         goto err;
8315         }
8316 map:
8317         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8318 err:
8319         bio_put(bio);
8320         return ret;
8321 }
8322
8323 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8324                                     int skip_sum)
8325 {
8326         struct inode *inode = dip->inode;
8327         struct btrfs_root *root = BTRFS_I(inode)->root;
8328         struct bio *bio;
8329         struct bio *orig_bio = dip->orig_bio;
8330         struct bio_vec *bvec = orig_bio->bi_io_vec;
8331         u64 start_sector = orig_bio->bi_iter.bi_sector;
8332         u64 file_offset = dip->logical_offset;
8333         u64 submit_len = 0;
8334         u64 map_length;
8335         u32 blocksize = root->sectorsize;
8336         int async_submit = 0;
8337         int nr_sectors;
8338         int ret;
8339         int i;
8340
8341         map_length = orig_bio->bi_iter.bi_size;
8342         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8343                               &map_length, NULL, 0);
8344         if (ret)
8345                 return -EIO;
8346
8347         if (map_length >= orig_bio->bi_iter.bi_size) {
8348                 bio = orig_bio;
8349                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8350                 goto submit;
8351         }
8352
8353         /* async crcs make it difficult to collect full stripe writes. */
8354         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8355                 async_submit = 0;
8356         else
8357                 async_submit = 1;
8358
8359         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8360         if (!bio)
8361                 return -ENOMEM;
8362
8363         bio->bi_private = dip;
8364         bio->bi_end_io = btrfs_end_dio_bio;
8365         btrfs_io_bio(bio)->logical = file_offset;
8366         atomic_inc(&dip->pending_bios);
8367
8368         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8369                 nr_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info, bvec->bv_len);
8370                 i = 0;
8371 next_block:
8372                 if (unlikely(map_length < submit_len + blocksize ||
8373                     bio_add_page(bio, bvec->bv_page, blocksize,
8374                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8375                         /*
8376                          * inc the count before we submit the bio so
8377                          * we know the end IO handler won't happen before
8378                          * we inc the count. Otherwise, the dip might get freed
8379                          * before we're done setting it up
8380                          */
8381                         atomic_inc(&dip->pending_bios);
8382                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8383                                                      file_offset, skip_sum,
8384                                                      async_submit);
8385                         if (ret) {
8386                                 bio_put(bio);
8387                                 atomic_dec(&dip->pending_bios);
8388                                 goto out_err;
8389                         }
8390
8391                         start_sector += submit_len >> 9;
8392                         file_offset += submit_len;
8393
8394                         submit_len = 0;
8395
8396                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8397                                                   start_sector, GFP_NOFS);
8398                         if (!bio)
8399                                 goto out_err;
8400                         bio->bi_private = dip;
8401                         bio->bi_end_io = btrfs_end_dio_bio;
8402                         btrfs_io_bio(bio)->logical = file_offset;
8403
8404                         map_length = orig_bio->bi_iter.bi_size;
8405                         ret = btrfs_map_block(root->fs_info, rw,
8406                                               start_sector << 9,
8407                                               &map_length, NULL, 0);
8408                         if (ret) {
8409                                 bio_put(bio);
8410                                 goto out_err;
8411                         }
8412
8413                         goto next_block;
8414                 } else {
8415                         submit_len += blocksize;
8416                         if (--nr_sectors) {
8417                                 i++;
8418                                 goto next_block;
8419                         }
8420                         bvec++;
8421                 }
8422         }
8423
8424 submit:
8425         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8426                                      async_submit);
8427         if (!ret)
8428                 return 0;
8429
8430         bio_put(bio);
8431 out_err:
8432         dip->errors = 1;
8433         /*
8434          * before atomic variable goto zero, we must
8435          * make sure dip->errors is perceived to be set.
8436          */
8437         smp_mb__before_atomic();
8438         if (atomic_dec_and_test(&dip->pending_bios))
8439                 bio_io_error(dip->orig_bio);
8440
8441         /* bio_end_io() will handle error, so we needn't return it */
8442         return 0;
8443 }
8444
8445 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8446                                 struct inode *inode, loff_t file_offset)
8447 {
8448         struct btrfs_dio_private *dip = NULL;
8449         struct bio *io_bio = NULL;
8450         struct btrfs_io_bio *btrfs_bio;
8451         int skip_sum;
8452         int write = rw & REQ_WRITE;
8453         int ret = 0;
8454
8455         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8456
8457         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8458         if (!io_bio) {
8459                 ret = -ENOMEM;
8460                 goto free_ordered;
8461         }
8462
8463         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8464         if (!dip) {
8465                 ret = -ENOMEM;
8466                 goto free_ordered;
8467         }
8468
8469         dip->private = dio_bio->bi_private;
8470         dip->inode = inode;
8471         dip->logical_offset = file_offset;
8472         dip->bytes = dio_bio->bi_iter.bi_size;
8473         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8474         io_bio->bi_private = dip;
8475         dip->orig_bio = io_bio;
8476         dip->dio_bio = dio_bio;
8477         atomic_set(&dip->pending_bios, 0);
8478         btrfs_bio = btrfs_io_bio(io_bio);
8479         btrfs_bio->logical = file_offset;
8480
8481         if (write) {
8482                 io_bio->bi_end_io = btrfs_endio_direct_write;
8483         } else {
8484                 io_bio->bi_end_io = btrfs_endio_direct_read;
8485                 dip->subio_endio = btrfs_subio_endio_read;
8486         }
8487
8488         /*
8489          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8490          * even if we fail to submit a bio, because in such case we do the
8491          * corresponding error handling below and it must not be done a second
8492          * time by btrfs_direct_IO().
8493          */
8494         if (write) {
8495                 struct btrfs_dio_data *dio_data = current->journal_info;
8496
8497                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8498                         dip->bytes;
8499                 dio_data->unsubmitted_oe_range_start =
8500                         dio_data->unsubmitted_oe_range_end;
8501         }
8502
8503         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8504         if (!ret)
8505                 return;
8506
8507         if (btrfs_bio->end_io)
8508                 btrfs_bio->end_io(btrfs_bio, ret);
8509
8510 free_ordered:
8511         /*
8512          * If we arrived here it means either we failed to submit the dip
8513          * or we either failed to clone the dio_bio or failed to allocate the
8514          * dip. If we cloned the dio_bio and allocated the dip, we can just
8515          * call bio_endio against our io_bio so that we get proper resource
8516          * cleanup if we fail to submit the dip, otherwise, we must do the
8517          * same as btrfs_endio_direct_[write|read] because we can't call these
8518          * callbacks - they require an allocated dip and a clone of dio_bio.
8519          */
8520         if (io_bio && dip) {
8521                 io_bio->bi_error = -EIO;
8522                 bio_endio(io_bio);
8523                 /*
8524                  * The end io callbacks free our dip, do the final put on io_bio
8525                  * and all the cleanup and final put for dio_bio (through
8526                  * dio_end_io()).
8527                  */
8528                 dip = NULL;
8529                 io_bio = NULL;
8530         } else {
8531                 if (write)
8532                         btrfs_endio_direct_write_update_ordered(inode,
8533                                                 file_offset,
8534                                                 dio_bio->bi_iter.bi_size,
8535                                                 0);
8536                 else
8537                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8538                               file_offset + dio_bio->bi_iter.bi_size - 1);
8539
8540                 dio_bio->bi_error = -EIO;
8541                 /*
8542                  * Releases and cleans up our dio_bio, no need to bio_put()
8543                  * nor bio_endio()/bio_io_error() against dio_bio.
8544                  */
8545                 dio_end_io(dio_bio, ret);
8546         }
8547         if (io_bio)
8548                 bio_put(io_bio);
8549         kfree(dip);
8550 }
8551
8552 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8553                         const struct iov_iter *iter, loff_t offset)
8554 {
8555         int seg;
8556         int i;
8557         unsigned blocksize_mask = root->sectorsize - 1;
8558         ssize_t retval = -EINVAL;
8559
8560         if (offset & blocksize_mask)
8561                 goto out;
8562
8563         if (iov_iter_alignment(iter) & blocksize_mask)
8564                 goto out;
8565
8566         /* If this is a write we don't need to check anymore */
8567         if (iov_iter_rw(iter) == WRITE)
8568                 return 0;
8569         /*
8570          * Check to make sure we don't have duplicate iov_base's in this
8571          * iovec, if so return EINVAL, otherwise we'll get csum errors
8572          * when reading back.
8573          */
8574         for (seg = 0; seg < iter->nr_segs; seg++) {
8575                 for (i = seg + 1; i < iter->nr_segs; i++) {
8576                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8577                                 goto out;
8578                 }
8579         }
8580         retval = 0;
8581 out:
8582         return retval;
8583 }
8584
8585 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8586 {
8587         struct file *file = iocb->ki_filp;
8588         struct inode *inode = file->f_mapping->host;
8589         struct btrfs_root *root = BTRFS_I(inode)->root;
8590         struct btrfs_dio_data dio_data = { 0 };
8591         loff_t offset = iocb->ki_pos;
8592         size_t count = 0;
8593         int flags = 0;
8594         bool wakeup = true;
8595         bool relock = false;
8596         ssize_t ret;
8597
8598         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8599                 return 0;
8600
8601         inode_dio_begin(inode);
8602         smp_mb__after_atomic();
8603
8604         /*
8605          * The generic stuff only does filemap_write_and_wait_range, which
8606          * isn't enough if we've written compressed pages to this area, so
8607          * we need to flush the dirty pages again to make absolutely sure
8608          * that any outstanding dirty pages are on disk.
8609          */
8610         count = iov_iter_count(iter);
8611         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8612                      &BTRFS_I(inode)->runtime_flags))
8613                 filemap_fdatawrite_range(inode->i_mapping, offset,
8614                                          offset + count - 1);
8615
8616         if (iov_iter_rw(iter) == WRITE) {
8617                 /*
8618                  * If the write DIO is beyond the EOF, we need update
8619                  * the isize, but it is protected by i_mutex. So we can
8620                  * not unlock the i_mutex at this case.
8621                  */
8622                 if (offset + count <= inode->i_size) {
8623                         inode_unlock(inode);
8624                         relock = true;
8625                 }
8626                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8627                 if (ret)
8628                         goto out;
8629                 dio_data.outstanding_extents = div64_u64(count +
8630                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8631                                                 BTRFS_MAX_EXTENT_SIZE);
8632
8633                 /*
8634                  * We need to know how many extents we reserved so that we can
8635                  * do the accounting properly if we go over the number we
8636                  * originally calculated.  Abuse current->journal_info for this.
8637                  */
8638                 dio_data.reserve = round_up(count, root->sectorsize);
8639                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8640                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8641                 current->journal_info = &dio_data;
8642         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8643                                      &BTRFS_I(inode)->runtime_flags)) {
8644                 inode_dio_end(inode);
8645                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8646                 wakeup = false;
8647         }
8648
8649         ret = __blockdev_direct_IO(iocb, inode,
8650                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8651                                    iter, btrfs_get_blocks_direct, NULL,
8652                                    btrfs_submit_direct, flags);
8653         if (iov_iter_rw(iter) == WRITE) {
8654                 current->journal_info = NULL;
8655                 if (ret < 0 && ret != -EIOCBQUEUED) {
8656                         if (dio_data.reserve)
8657                                 btrfs_delalloc_release_space(inode, offset,
8658                                                              dio_data.reserve);
8659                         /*
8660                          * On error we might have left some ordered extents
8661                          * without submitting corresponding bios for them, so
8662                          * cleanup them up to avoid other tasks getting them
8663                          * and waiting for them to complete forever.
8664                          */
8665                         if (dio_data.unsubmitted_oe_range_start <
8666                             dio_data.unsubmitted_oe_range_end)
8667                                 btrfs_endio_direct_write_update_ordered(inode,
8668                                         dio_data.unsubmitted_oe_range_start,
8669                                         dio_data.unsubmitted_oe_range_end -
8670                                         dio_data.unsubmitted_oe_range_start,
8671                                         0);
8672                 } else if (ret >= 0 && (size_t)ret < count)
8673                         btrfs_delalloc_release_space(inode, offset,
8674                                                      count - (size_t)ret);
8675         }
8676 out:
8677         if (wakeup)
8678                 inode_dio_end(inode);
8679         if (relock)
8680                 inode_lock(inode);
8681
8682         return ret;
8683 }
8684
8685 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8686
8687 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8688                 __u64 start, __u64 len)
8689 {
8690         int     ret;
8691
8692         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8693         if (ret)
8694                 return ret;
8695
8696         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8697 }
8698
8699 int btrfs_readpage(struct file *file, struct page *page)
8700 {
8701         struct extent_io_tree *tree;
8702         tree = &BTRFS_I(page->mapping->host)->io_tree;
8703         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8704 }
8705
8706 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8707 {
8708         struct extent_io_tree *tree;
8709         struct inode *inode = page->mapping->host;
8710         int ret;
8711
8712         if (current->flags & PF_MEMALLOC) {
8713                 redirty_page_for_writepage(wbc, page);
8714                 unlock_page(page);
8715                 return 0;
8716         }
8717
8718         /*
8719          * If we are under memory pressure we will call this directly from the
8720          * VM, we need to make sure we have the inode referenced for the ordered
8721          * extent.  If not just return like we didn't do anything.
8722          */
8723         if (!igrab(inode)) {
8724                 redirty_page_for_writepage(wbc, page);
8725                 return AOP_WRITEPAGE_ACTIVATE;
8726         }
8727         tree = &BTRFS_I(page->mapping->host)->io_tree;
8728         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8729         btrfs_add_delayed_iput(inode);
8730         return ret;
8731 }
8732
8733 static int btrfs_writepages(struct address_space *mapping,
8734                             struct writeback_control *wbc)
8735 {
8736         struct extent_io_tree *tree;
8737
8738         tree = &BTRFS_I(mapping->host)->io_tree;
8739         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8740 }
8741
8742 static int
8743 btrfs_readpages(struct file *file, struct address_space *mapping,
8744                 struct list_head *pages, unsigned nr_pages)
8745 {
8746         struct extent_io_tree *tree;
8747         tree = &BTRFS_I(mapping->host)->io_tree;
8748         return extent_readpages(tree, mapping, pages, nr_pages,
8749                                 btrfs_get_extent);
8750 }
8751 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8752 {
8753         struct extent_io_tree *tree;
8754         struct extent_map_tree *map;
8755         int ret;
8756
8757         tree = &BTRFS_I(page->mapping->host)->io_tree;
8758         map = &BTRFS_I(page->mapping->host)->extent_tree;
8759         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8760         if (ret == 1) {
8761                 ClearPagePrivate(page);
8762                 set_page_private(page, 0);
8763                 put_page(page);
8764         }
8765         return ret;
8766 }
8767
8768 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8769 {
8770         if (PageWriteback(page) || PageDirty(page))
8771                 return 0;
8772         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8773 }
8774
8775 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8776                                  unsigned int length)
8777 {
8778         struct inode *inode = page->mapping->host;
8779         struct extent_io_tree *tree;
8780         struct btrfs_ordered_extent *ordered;
8781         struct extent_state *cached_state = NULL;
8782         u64 page_start = page_offset(page);
8783         u64 page_end = page_start + PAGE_SIZE - 1;
8784         u64 start;
8785         u64 end;
8786         int inode_evicting = inode->i_state & I_FREEING;
8787
8788         /*
8789          * we have the page locked, so new writeback can't start,
8790          * and the dirty bit won't be cleared while we are here.
8791          *
8792          * Wait for IO on this page so that we can safely clear
8793          * the PagePrivate2 bit and do ordered accounting
8794          */
8795         wait_on_page_writeback(page);
8796
8797         tree = &BTRFS_I(inode)->io_tree;
8798         if (offset) {
8799                 btrfs_releasepage(page, GFP_NOFS);
8800                 return;
8801         }
8802
8803         if (!inode_evicting)
8804                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8805 again:
8806         start = page_start;
8807         ordered = btrfs_lookup_ordered_range(inode, start,
8808                                         page_end - start + 1);
8809         if (ordered) {
8810                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8811                 /*
8812                  * IO on this page will never be started, so we need
8813                  * to account for any ordered extents now
8814                  */
8815                 if (!inode_evicting)
8816                         clear_extent_bit(tree, start, end,
8817                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8818                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8819                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8820                                          GFP_NOFS);
8821                 /*
8822                  * whoever cleared the private bit is responsible
8823                  * for the finish_ordered_io
8824                  */
8825                 if (TestClearPagePrivate2(page)) {
8826                         struct btrfs_ordered_inode_tree *tree;
8827                         u64 new_len;
8828
8829                         tree = &BTRFS_I(inode)->ordered_tree;
8830
8831                         spin_lock_irq(&tree->lock);
8832                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8833                         new_len = start - ordered->file_offset;
8834                         if (new_len < ordered->truncated_len)
8835                                 ordered->truncated_len = new_len;
8836                         spin_unlock_irq(&tree->lock);
8837
8838                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8839                                                            start,
8840                                                            end - start + 1, 1))
8841                                 btrfs_finish_ordered_io(ordered);
8842                 }
8843                 btrfs_put_ordered_extent(ordered);
8844                 if (!inode_evicting) {
8845                         cached_state = NULL;
8846                         lock_extent_bits(tree, start, end,
8847                                          &cached_state);
8848                 }
8849
8850                 start = end + 1;
8851                 if (start < page_end)
8852                         goto again;
8853         }
8854
8855         /*
8856          * Qgroup reserved space handler
8857          * Page here will be either
8858          * 1) Already written to disk
8859          *    In this case, its reserved space is released from data rsv map
8860          *    and will be freed by delayed_ref handler finally.
8861          *    So even we call qgroup_free_data(), it won't decrease reserved
8862          *    space.
8863          * 2) Not written to disk
8864          *    This means the reserved space should be freed here.
8865          */
8866         btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
8867         if (!inode_evicting) {
8868                 clear_extent_bit(tree, page_start, page_end,
8869                                  EXTENT_LOCKED | EXTENT_DIRTY |
8870                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8871                                  EXTENT_DEFRAG, 1, 1,
8872                                  &cached_state, GFP_NOFS);
8873
8874                 __btrfs_releasepage(page, GFP_NOFS);
8875         }
8876
8877         ClearPageChecked(page);
8878         if (PagePrivate(page)) {
8879                 ClearPagePrivate(page);
8880                 set_page_private(page, 0);
8881                 put_page(page);
8882         }
8883 }
8884
8885 /*
8886  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8887  * called from a page fault handler when a page is first dirtied. Hence we must
8888  * be careful to check for EOF conditions here. We set the page up correctly
8889  * for a written page which means we get ENOSPC checking when writing into
8890  * holes and correct delalloc and unwritten extent mapping on filesystems that
8891  * support these features.
8892  *
8893  * We are not allowed to take the i_mutex here so we have to play games to
8894  * protect against truncate races as the page could now be beyond EOF.  Because
8895  * vmtruncate() writes the inode size before removing pages, once we have the
8896  * page lock we can determine safely if the page is beyond EOF. If it is not
8897  * beyond EOF, then the page is guaranteed safe against truncation until we
8898  * unlock the page.
8899  */
8900 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8901 {
8902         struct page *page = vmf->page;
8903         struct inode *inode = file_inode(vma->vm_file);
8904         struct btrfs_root *root = BTRFS_I(inode)->root;
8905         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8906         struct btrfs_ordered_extent *ordered;
8907         struct extent_state *cached_state = NULL;
8908         char *kaddr;
8909         unsigned long zero_start;
8910         loff_t size;
8911         int ret;
8912         int reserved = 0;
8913         u64 reserved_space;
8914         u64 page_start;
8915         u64 page_end;
8916         u64 end;
8917
8918         reserved_space = PAGE_SIZE;
8919
8920         sb_start_pagefault(inode->i_sb);
8921         page_start = page_offset(page);
8922         page_end = page_start + PAGE_SIZE - 1;
8923         end = page_end;
8924
8925         /*
8926          * Reserving delalloc space after obtaining the page lock can lead to
8927          * deadlock. For example, if a dirty page is locked by this function
8928          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8929          * dirty page write out, then the btrfs_writepage() function could
8930          * end up waiting indefinitely to get a lock on the page currently
8931          * being processed by btrfs_page_mkwrite() function.
8932          */
8933         ret = btrfs_delalloc_reserve_space(inode, page_start,
8934                                            reserved_space);
8935         if (!ret) {
8936                 ret = file_update_time(vma->vm_file);
8937                 reserved = 1;
8938         }
8939         if (ret) {
8940                 if (ret == -ENOMEM)
8941                         ret = VM_FAULT_OOM;
8942                 else /* -ENOSPC, -EIO, etc */
8943                         ret = VM_FAULT_SIGBUS;
8944                 if (reserved)
8945                         goto out;
8946                 goto out_noreserve;
8947         }
8948
8949         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8950 again:
8951         lock_page(page);
8952         size = i_size_read(inode);
8953
8954         if ((page->mapping != inode->i_mapping) ||
8955             (page_start >= size)) {
8956                 /* page got truncated out from underneath us */
8957                 goto out_unlock;
8958         }
8959         wait_on_page_writeback(page);
8960
8961         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8962         set_page_extent_mapped(page);
8963
8964         /*
8965          * we can't set the delalloc bits if there are pending ordered
8966          * extents.  Drop our locks and wait for them to finish
8967          */
8968         ordered = btrfs_lookup_ordered_range(inode, page_start, page_end);
8969         if (ordered) {
8970                 unlock_extent_cached(io_tree, page_start, page_end,
8971                                      &cached_state, GFP_NOFS);
8972                 unlock_page(page);
8973                 btrfs_start_ordered_extent(inode, ordered, 1);
8974                 btrfs_put_ordered_extent(ordered);
8975                 goto again;
8976         }
8977
8978         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8979                 reserved_space = round_up(size - page_start, root->sectorsize);
8980                 if (reserved_space < PAGE_SIZE) {
8981                         end = page_start + reserved_space - 1;
8982                         spin_lock(&BTRFS_I(inode)->lock);
8983                         BTRFS_I(inode)->outstanding_extents++;
8984                         spin_unlock(&BTRFS_I(inode)->lock);
8985                         btrfs_delalloc_release_space(inode, page_start,
8986                                                 PAGE_SIZE - reserved_space);
8987                 }
8988         }
8989
8990         /*
8991          * XXX - page_mkwrite gets called every time the page is dirtied, even
8992          * if it was already dirty, so for space accounting reasons we need to
8993          * clear any delalloc bits for the range we are fixing to save.  There
8994          * is probably a better way to do this, but for now keep consistent with
8995          * prepare_pages in the normal write path.
8996          */
8997         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8998                           EXTENT_DIRTY | EXTENT_DELALLOC |
8999                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9000                           0, 0, &cached_state, GFP_NOFS);
9001
9002         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9003                                         &cached_state);
9004         if (ret) {
9005                 unlock_extent_cached(io_tree, page_start, page_end,
9006                                      &cached_state, GFP_NOFS);
9007                 ret = VM_FAULT_SIGBUS;
9008                 goto out_unlock;
9009         }
9010         ret = 0;
9011
9012         /* page is wholly or partially inside EOF */
9013         if (page_start + PAGE_SIZE > size)
9014                 zero_start = size & ~PAGE_MASK;
9015         else
9016                 zero_start = PAGE_SIZE;
9017
9018         if (zero_start != PAGE_SIZE) {
9019                 kaddr = kmap(page);
9020                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9021                 flush_dcache_page(page);
9022                 kunmap(page);
9023         }
9024         ClearPageChecked(page);
9025         set_page_dirty(page);
9026         SetPageUptodate(page);
9027
9028         BTRFS_I(inode)->last_trans = root->fs_info->generation;
9029         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9030         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9031
9032         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9033
9034 out_unlock:
9035         if (!ret) {
9036                 sb_end_pagefault(inode->i_sb);
9037                 return VM_FAULT_LOCKED;
9038         }
9039         unlock_page(page);
9040 out:
9041         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9042 out_noreserve:
9043         sb_end_pagefault(inode->i_sb);
9044         return ret;
9045 }
9046
9047 static int btrfs_truncate(struct inode *inode)
9048 {
9049         struct btrfs_root *root = BTRFS_I(inode)->root;
9050         struct btrfs_block_rsv *rsv;
9051         int ret = 0;
9052         int err = 0;
9053         struct btrfs_trans_handle *trans;
9054         u64 mask = root->sectorsize - 1;
9055         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
9056
9057         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9058                                        (u64)-1);
9059         if (ret)
9060                 return ret;
9061
9062         /*
9063          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9064          * 3 things going on here
9065          *
9066          * 1) We need to reserve space for our orphan item and the space to
9067          * delete our orphan item.  Lord knows we don't want to have a dangling
9068          * orphan item because we didn't reserve space to remove it.
9069          *
9070          * 2) We need to reserve space to update our inode.
9071          *
9072          * 3) We need to have something to cache all the space that is going to
9073          * be free'd up by the truncate operation, but also have some slack
9074          * space reserved in case it uses space during the truncate (thank you
9075          * very much snapshotting).
9076          *
9077          * And we need these to all be separate.  The fact is we can use a lot of
9078          * space doing the truncate, and we have no earthly idea how much space
9079          * we will use, so we need the truncate reservation to be separate so it
9080          * doesn't end up using space reserved for updating the inode or
9081          * removing the orphan item.  We also need to be able to stop the
9082          * transaction and start a new one, which means we need to be able to
9083          * update the inode several times, and we have no idea of knowing how
9084          * many times that will be, so we can't just reserve 1 item for the
9085          * entirety of the operation, so that has to be done separately as well.
9086          * Then there is the orphan item, which does indeed need to be held on
9087          * to for the whole operation, and we need nobody to touch this reserved
9088          * space except the orphan code.
9089          *
9090          * So that leaves us with
9091          *
9092          * 1) root->orphan_block_rsv - for the orphan deletion.
9093          * 2) rsv - for the truncate reservation, which we will steal from the
9094          * transaction reservation.
9095          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9096          * updating the inode.
9097          */
9098         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
9099         if (!rsv)
9100                 return -ENOMEM;
9101         rsv->size = min_size;
9102         rsv->failfast = 1;
9103
9104         /*
9105          * 1 for the truncate slack space
9106          * 1 for updating the inode.
9107          */
9108         trans = btrfs_start_transaction(root, 2);
9109         if (IS_ERR(trans)) {
9110                 err = PTR_ERR(trans);
9111                 goto out;
9112         }
9113
9114         /* Migrate the slack space for the truncate to our reserve */
9115         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
9116                                       min_size);
9117         BUG_ON(ret);
9118
9119         /*
9120          * So if we truncate and then write and fsync we normally would just
9121          * write the extents that changed, which is a problem if we need to
9122          * first truncate that entire inode.  So set this flag so we write out
9123          * all of the extents in the inode to the sync log so we're completely
9124          * safe.
9125          */
9126         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9127         trans->block_rsv = rsv;
9128
9129         while (1) {
9130                 ret = btrfs_truncate_inode_items(trans, root, inode,
9131                                                  inode->i_size,
9132                                                  BTRFS_EXTENT_DATA_KEY);
9133                 if (ret != -ENOSPC && ret != -EAGAIN) {
9134                         err = ret;
9135                         break;
9136                 }
9137
9138                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9139                 ret = btrfs_update_inode(trans, root, inode);
9140                 if (ret) {
9141                         err = ret;
9142                         break;
9143                 }
9144
9145                 btrfs_end_transaction(trans, root);
9146                 btrfs_btree_balance_dirty(root);
9147
9148                 trans = btrfs_start_transaction(root, 2);
9149                 if (IS_ERR(trans)) {
9150                         ret = err = PTR_ERR(trans);
9151                         trans = NULL;
9152                         break;
9153                 }
9154
9155                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
9156                                               rsv, min_size);
9157                 BUG_ON(ret);    /* shouldn't happen */
9158                 trans->block_rsv = rsv;
9159         }
9160
9161         if (ret == 0 && inode->i_nlink > 0) {
9162                 trans->block_rsv = root->orphan_block_rsv;
9163                 ret = btrfs_orphan_del(trans, inode);
9164                 if (ret)
9165                         err = ret;
9166         }
9167
9168         if (trans) {
9169                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9170                 ret = btrfs_update_inode(trans, root, inode);
9171                 if (ret && !err)
9172                         err = ret;
9173
9174                 ret = btrfs_end_transaction(trans, root);
9175                 btrfs_btree_balance_dirty(root);
9176         }
9177
9178 out:
9179         btrfs_free_block_rsv(root, rsv);
9180
9181         if (ret && !err)
9182                 err = ret;
9183
9184         return err;
9185 }
9186
9187 /*
9188  * create a new subvolume directory/inode (helper for the ioctl).
9189  */
9190 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9191                              struct btrfs_root *new_root,
9192                              struct btrfs_root *parent_root,
9193                              u64 new_dirid)
9194 {
9195         struct inode *inode;
9196         int err;
9197         u64 index = 0;
9198
9199         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9200                                 new_dirid, new_dirid,
9201                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9202                                 &index);
9203         if (IS_ERR(inode))
9204                 return PTR_ERR(inode);
9205         inode->i_op = &btrfs_dir_inode_operations;
9206         inode->i_fop = &btrfs_dir_file_operations;
9207
9208         set_nlink(inode, 1);
9209         btrfs_i_size_write(inode, 0);
9210         unlock_new_inode(inode);
9211
9212         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9213         if (err)
9214                 btrfs_err(new_root->fs_info,
9215                           "error inheriting subvolume %llu properties: %d",
9216                           new_root->root_key.objectid, err);
9217
9218         err = btrfs_update_inode(trans, new_root, inode);
9219
9220         iput(inode);
9221         return err;
9222 }
9223
9224 struct inode *btrfs_alloc_inode(struct super_block *sb)
9225 {
9226         struct btrfs_inode *ei;
9227         struct inode *inode;
9228
9229         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9230         if (!ei)
9231                 return NULL;
9232
9233         ei->root = NULL;
9234         ei->generation = 0;
9235         ei->last_trans = 0;
9236         ei->last_sub_trans = 0;
9237         ei->logged_trans = 0;
9238         ei->delalloc_bytes = 0;
9239         ei->defrag_bytes = 0;
9240         ei->disk_i_size = 0;
9241         ei->flags = 0;
9242         ei->csum_bytes = 0;
9243         ei->index_cnt = (u64)-1;
9244         ei->dir_index = 0;
9245         ei->last_unlink_trans = 0;
9246         ei->last_log_commit = 0;
9247         ei->delayed_iput_count = 0;
9248
9249         spin_lock_init(&ei->lock);
9250         ei->outstanding_extents = 0;
9251         ei->reserved_extents = 0;
9252
9253         ei->runtime_flags = 0;
9254         ei->force_compress = BTRFS_COMPRESS_NONE;
9255
9256         ei->delayed_node = NULL;
9257
9258         ei->i_otime.tv_sec = 0;
9259         ei->i_otime.tv_nsec = 0;
9260
9261         inode = &ei->vfs_inode;
9262         extent_map_tree_init(&ei->extent_tree);
9263         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9264         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9265         ei->io_tree.track_uptodate = 1;
9266         ei->io_failure_tree.track_uptodate = 1;
9267         atomic_set(&ei->sync_writers, 0);
9268         mutex_init(&ei->log_mutex);
9269         mutex_init(&ei->delalloc_mutex);
9270         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9271         INIT_LIST_HEAD(&ei->delalloc_inodes);
9272         INIT_LIST_HEAD(&ei->delayed_iput);
9273         RB_CLEAR_NODE(&ei->rb_node);
9274         init_rwsem(&ei->dio_sem);
9275
9276         return inode;
9277 }
9278
9279 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9280 void btrfs_test_destroy_inode(struct inode *inode)
9281 {
9282         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9283         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9284 }
9285 #endif
9286
9287 static void btrfs_i_callback(struct rcu_head *head)
9288 {
9289         struct inode *inode = container_of(head, struct inode, i_rcu);
9290         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9291 }
9292
9293 void btrfs_destroy_inode(struct inode *inode)
9294 {
9295         struct btrfs_ordered_extent *ordered;
9296         struct btrfs_root *root = BTRFS_I(inode)->root;
9297
9298         WARN_ON(!hlist_empty(&inode->i_dentry));
9299         WARN_ON(inode->i_data.nrpages);
9300         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9301         WARN_ON(BTRFS_I(inode)->reserved_extents);
9302         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9303         WARN_ON(BTRFS_I(inode)->csum_bytes);
9304         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9305
9306         /*
9307          * This can happen where we create an inode, but somebody else also
9308          * created the same inode and we need to destroy the one we already
9309          * created.
9310          */
9311         if (!root)
9312                 goto free;
9313
9314         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9315                      &BTRFS_I(inode)->runtime_flags)) {
9316                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9317                         btrfs_ino(inode));
9318                 atomic_dec(&root->orphan_inodes);
9319         }
9320
9321         while (1) {
9322                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9323                 if (!ordered)
9324                         break;
9325                 else {
9326                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9327                                 ordered->file_offset, ordered->len);
9328                         btrfs_remove_ordered_extent(inode, ordered);
9329                         btrfs_put_ordered_extent(ordered);
9330                         btrfs_put_ordered_extent(ordered);
9331                 }
9332         }
9333         btrfs_qgroup_check_reserved_leak(inode);
9334         inode_tree_del(inode);
9335         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9336 free:
9337         call_rcu(&inode->i_rcu, btrfs_i_callback);
9338 }
9339
9340 int btrfs_drop_inode(struct inode *inode)
9341 {
9342         struct btrfs_root *root = BTRFS_I(inode)->root;
9343
9344         if (root == NULL)
9345                 return 1;
9346
9347         /* the snap/subvol tree is on deleting */
9348         if (btrfs_root_refs(&root->root_item) == 0)
9349                 return 1;
9350         else
9351                 return generic_drop_inode(inode);
9352 }
9353
9354 static void init_once(void *foo)
9355 {
9356         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9357
9358         inode_init_once(&ei->vfs_inode);
9359 }
9360
9361 void btrfs_destroy_cachep(void)
9362 {
9363         /*
9364          * Make sure all delayed rcu free inodes are flushed before we
9365          * destroy cache.
9366          */
9367         rcu_barrier();
9368         kmem_cache_destroy(btrfs_inode_cachep);
9369         kmem_cache_destroy(btrfs_trans_handle_cachep);
9370         kmem_cache_destroy(btrfs_transaction_cachep);
9371         kmem_cache_destroy(btrfs_path_cachep);
9372         kmem_cache_destroy(btrfs_free_space_cachep);
9373 }
9374
9375 int btrfs_init_cachep(void)
9376 {
9377         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9378                         sizeof(struct btrfs_inode), 0,
9379                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9380                         init_once);
9381         if (!btrfs_inode_cachep)
9382                 goto fail;
9383
9384         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9385                         sizeof(struct btrfs_trans_handle), 0,
9386                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9387         if (!btrfs_trans_handle_cachep)
9388                 goto fail;
9389
9390         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9391                         sizeof(struct btrfs_transaction), 0,
9392                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9393         if (!btrfs_transaction_cachep)
9394                 goto fail;
9395
9396         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9397                         sizeof(struct btrfs_path), 0,
9398                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9399         if (!btrfs_path_cachep)
9400                 goto fail;
9401
9402         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9403                         sizeof(struct btrfs_free_space), 0,
9404                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9405         if (!btrfs_free_space_cachep)
9406                 goto fail;
9407
9408         return 0;
9409 fail:
9410         btrfs_destroy_cachep();
9411         return -ENOMEM;
9412 }
9413
9414 static int btrfs_getattr(struct vfsmount *mnt,
9415                          struct dentry *dentry, struct kstat *stat)
9416 {
9417         u64 delalloc_bytes;
9418         struct inode *inode = d_inode(dentry);
9419         u32 blocksize = inode->i_sb->s_blocksize;
9420
9421         generic_fillattr(inode, stat);
9422         stat->dev = BTRFS_I(inode)->root->anon_dev;
9423
9424         spin_lock(&BTRFS_I(inode)->lock);
9425         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9426         spin_unlock(&BTRFS_I(inode)->lock);
9427         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9428                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9429         return 0;
9430 }
9431
9432 static int btrfs_rename_exchange(struct inode *old_dir,
9433                               struct dentry *old_dentry,
9434                               struct inode *new_dir,
9435                               struct dentry *new_dentry)
9436 {
9437         struct btrfs_trans_handle *trans;
9438         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9439         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9440         struct inode *new_inode = new_dentry->d_inode;
9441         struct inode *old_inode = old_dentry->d_inode;
9442         struct timespec ctime = CURRENT_TIME;
9443         struct dentry *parent;
9444         u64 old_ino = btrfs_ino(old_inode);
9445         u64 new_ino = btrfs_ino(new_inode);
9446         u64 old_idx = 0;
9447         u64 new_idx = 0;
9448         u64 root_objectid;
9449         int ret;
9450         bool root_log_pinned = false;
9451         bool dest_log_pinned = false;
9452
9453         /* we only allow rename subvolume link between subvolumes */
9454         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9455                 return -EXDEV;
9456
9457         /* close the race window with snapshot create/destroy ioctl */
9458         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9459                 down_read(&root->fs_info->subvol_sem);
9460         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9461                 down_read(&dest->fs_info->subvol_sem);
9462
9463         /*
9464          * We want to reserve the absolute worst case amount of items.  So if
9465          * both inodes are subvols and we need to unlink them then that would
9466          * require 4 item modifications, but if they are both normal inodes it
9467          * would require 5 item modifications, so we'll assume their normal
9468          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9469          * should cover the worst case number of items we'll modify.
9470          */
9471         trans = btrfs_start_transaction(root, 12);
9472         if (IS_ERR(trans)) {
9473                 ret = PTR_ERR(trans);
9474                 goto out_notrans;
9475         }
9476
9477         /*
9478          * We need to find a free sequence number both in the source and
9479          * in the destination directory for the exchange.
9480          */
9481         ret = btrfs_set_inode_index(new_dir, &old_idx);
9482         if (ret)
9483                 goto out_fail;
9484         ret = btrfs_set_inode_index(old_dir, &new_idx);
9485         if (ret)
9486                 goto out_fail;
9487
9488         BTRFS_I(old_inode)->dir_index = 0ULL;
9489         BTRFS_I(new_inode)->dir_index = 0ULL;
9490
9491         /* Reference for the source. */
9492         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9493                 /* force full log commit if subvolume involved. */
9494                 btrfs_set_log_full_commit(root->fs_info, trans);
9495         } else {
9496                 btrfs_pin_log_trans(root);
9497                 root_log_pinned = true;
9498                 ret = btrfs_insert_inode_ref(trans, dest,
9499                                              new_dentry->d_name.name,
9500                                              new_dentry->d_name.len,
9501                                              old_ino,
9502                                              btrfs_ino(new_dir), old_idx);
9503                 if (ret)
9504                         goto out_fail;
9505         }
9506
9507         /* And now for the dest. */
9508         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9509                 /* force full log commit if subvolume involved. */
9510                 btrfs_set_log_full_commit(dest->fs_info, trans);
9511         } else {
9512                 btrfs_pin_log_trans(dest);
9513                 dest_log_pinned = true;
9514                 ret = btrfs_insert_inode_ref(trans, root,
9515                                              old_dentry->d_name.name,
9516                                              old_dentry->d_name.len,
9517                                              new_ino,
9518                                              btrfs_ino(old_dir), new_idx);
9519                 if (ret)
9520                         goto out_fail;
9521         }
9522
9523         /* Update inode version and ctime/mtime. */
9524         inode_inc_iversion(old_dir);
9525         inode_inc_iversion(new_dir);
9526         inode_inc_iversion(old_inode);
9527         inode_inc_iversion(new_inode);
9528         old_dir->i_ctime = old_dir->i_mtime = ctime;
9529         new_dir->i_ctime = new_dir->i_mtime = ctime;
9530         old_inode->i_ctime = ctime;
9531         new_inode->i_ctime = ctime;
9532
9533         if (old_dentry->d_parent != new_dentry->d_parent) {
9534                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9535                 btrfs_record_unlink_dir(trans, new_dir, new_inode, 1);
9536         }
9537
9538         /* src is a subvolume */
9539         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9540                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9541                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9542                                           root_objectid,
9543                                           old_dentry->d_name.name,
9544                                           old_dentry->d_name.len);
9545         } else { /* src is an inode */
9546                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9547                                            old_dentry->d_inode,
9548                                            old_dentry->d_name.name,
9549                                            old_dentry->d_name.len);
9550                 if (!ret)
9551                         ret = btrfs_update_inode(trans, root, old_inode);
9552         }
9553         if (ret) {
9554                 btrfs_abort_transaction(trans, root, ret);
9555                 goto out_fail;
9556         }
9557
9558         /* dest is a subvolume */
9559         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9560                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9561                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9562                                           root_objectid,
9563                                           new_dentry->d_name.name,
9564                                           new_dentry->d_name.len);
9565         } else { /* dest is an inode */
9566                 ret = __btrfs_unlink_inode(trans, dest, new_dir,
9567                                            new_dentry->d_inode,
9568                                            new_dentry->d_name.name,
9569                                            new_dentry->d_name.len);
9570                 if (!ret)
9571                         ret = btrfs_update_inode(trans, dest, new_inode);
9572         }
9573         if (ret) {
9574                 btrfs_abort_transaction(trans, root, ret);
9575                 goto out_fail;
9576         }
9577
9578         ret = btrfs_add_link(trans, new_dir, old_inode,
9579                              new_dentry->d_name.name,
9580                              new_dentry->d_name.len, 0, old_idx);
9581         if (ret) {
9582                 btrfs_abort_transaction(trans, root, ret);
9583                 goto out_fail;
9584         }
9585
9586         ret = btrfs_add_link(trans, old_dir, new_inode,
9587                              old_dentry->d_name.name,
9588                              old_dentry->d_name.len, 0, new_idx);
9589         if (ret) {
9590                 btrfs_abort_transaction(trans, root, ret);
9591                 goto out_fail;
9592         }
9593
9594         if (old_inode->i_nlink == 1)
9595                 BTRFS_I(old_inode)->dir_index = old_idx;
9596         if (new_inode->i_nlink == 1)
9597                 BTRFS_I(new_inode)->dir_index = new_idx;
9598
9599         if (root_log_pinned) {
9600                 parent = new_dentry->d_parent;
9601                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9602                 btrfs_end_log_trans(root);
9603                 root_log_pinned = false;
9604         }
9605         if (dest_log_pinned) {
9606                 parent = old_dentry->d_parent;
9607                 btrfs_log_new_name(trans, new_inode, new_dir, parent);
9608                 btrfs_end_log_trans(dest);
9609                 dest_log_pinned = false;
9610         }
9611 out_fail:
9612         /*
9613          * If we have pinned a log and an error happened, we unpin tasks
9614          * trying to sync the log and force them to fallback to a transaction
9615          * commit if the log currently contains any of the inodes involved in
9616          * this rename operation (to ensure we do not persist a log with an
9617          * inconsistent state for any of these inodes or leading to any
9618          * inconsistencies when replayed). If the transaction was aborted, the
9619          * abortion reason is propagated to userspace when attempting to commit
9620          * the transaction. If the log does not contain any of these inodes, we
9621          * allow the tasks to sync it.
9622          */
9623         if (ret && (root_log_pinned || dest_log_pinned)) {
9624                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9625                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9626                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9627                     (new_inode &&
9628                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9629                     btrfs_set_log_full_commit(root->fs_info, trans);
9630
9631                 if (root_log_pinned) {
9632                         btrfs_end_log_trans(root);
9633                         root_log_pinned = false;
9634                 }
9635                 if (dest_log_pinned) {
9636                         btrfs_end_log_trans(dest);
9637                         dest_log_pinned = false;
9638                 }
9639         }
9640         ret = btrfs_end_transaction(trans, root);
9641 out_notrans:
9642         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9643                 up_read(&dest->fs_info->subvol_sem);
9644         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9645                 up_read(&root->fs_info->subvol_sem);
9646
9647         return ret;
9648 }
9649
9650 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9651                                      struct btrfs_root *root,
9652                                      struct inode *dir,
9653                                      struct dentry *dentry)
9654 {
9655         int ret;
9656         struct inode *inode;
9657         u64 objectid;
9658         u64 index;
9659
9660         ret = btrfs_find_free_ino(root, &objectid);
9661         if (ret)
9662                 return ret;
9663
9664         inode = btrfs_new_inode(trans, root, dir,
9665                                 dentry->d_name.name,
9666                                 dentry->d_name.len,
9667                                 btrfs_ino(dir),
9668                                 objectid,
9669                                 S_IFCHR | WHITEOUT_MODE,
9670                                 &index);
9671
9672         if (IS_ERR(inode)) {
9673                 ret = PTR_ERR(inode);
9674                 return ret;
9675         }
9676
9677         inode->i_op = &btrfs_special_inode_operations;
9678         init_special_inode(inode, inode->i_mode,
9679                 WHITEOUT_DEV);
9680
9681         ret = btrfs_init_inode_security(trans, inode, dir,
9682                                 &dentry->d_name);
9683         if (ret)
9684                 goto out;
9685
9686         ret = btrfs_add_nondir(trans, dir, dentry,
9687                                 inode, 0, index);
9688         if (ret)
9689                 goto out;
9690
9691         ret = btrfs_update_inode(trans, root, inode);
9692 out:
9693         unlock_new_inode(inode);
9694         if (ret)
9695                 inode_dec_link_count(inode);
9696         iput(inode);
9697
9698         return ret;
9699 }
9700
9701 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9702                            struct inode *new_dir, struct dentry *new_dentry,
9703                            unsigned int flags)
9704 {
9705         struct btrfs_trans_handle *trans;
9706         unsigned int trans_num_items;
9707         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9708         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9709         struct inode *new_inode = d_inode(new_dentry);
9710         struct inode *old_inode = d_inode(old_dentry);
9711         u64 index = 0;
9712         u64 root_objectid;
9713         int ret;
9714         u64 old_ino = btrfs_ino(old_inode);
9715         bool log_pinned = false;
9716
9717         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9718                 return -EPERM;
9719
9720         /* we only allow rename subvolume link between subvolumes */
9721         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9722                 return -EXDEV;
9723
9724         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9725             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9726                 return -ENOTEMPTY;
9727
9728         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9729             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9730                 return -ENOTEMPTY;
9731
9732
9733         /* check for collisions, even if the  name isn't there */
9734         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9735                              new_dentry->d_name.name,
9736                              new_dentry->d_name.len);
9737
9738         if (ret) {
9739                 if (ret == -EEXIST) {
9740                         /* we shouldn't get
9741                          * eexist without a new_inode */
9742                         if (WARN_ON(!new_inode)) {
9743                                 return ret;
9744                         }
9745                 } else {
9746                         /* maybe -EOVERFLOW */
9747                         return ret;
9748                 }
9749         }
9750         ret = 0;
9751
9752         /*
9753          * we're using rename to replace one file with another.  Start IO on it
9754          * now so  we don't add too much work to the end of the transaction
9755          */
9756         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9757                 filemap_flush(old_inode->i_mapping);
9758
9759         /* close the racy window with snapshot create/destroy ioctl */
9760         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9761                 down_read(&root->fs_info->subvol_sem);
9762         /*
9763          * We want to reserve the absolute worst case amount of items.  So if
9764          * both inodes are subvols and we need to unlink them then that would
9765          * require 4 item modifications, but if they are both normal inodes it
9766          * would require 5 item modifications, so we'll assume they are normal
9767          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9768          * should cover the worst case number of items we'll modify.
9769          * If our rename has the whiteout flag, we need more 5 units for the
9770          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9771          * when selinux is enabled).
9772          */
9773         trans_num_items = 11;
9774         if (flags & RENAME_WHITEOUT)
9775                 trans_num_items += 5;
9776         trans = btrfs_start_transaction(root, trans_num_items);
9777         if (IS_ERR(trans)) {
9778                 ret = PTR_ERR(trans);
9779                 goto out_notrans;
9780         }
9781
9782         if (dest != root)
9783                 btrfs_record_root_in_trans(trans, dest);
9784
9785         ret = btrfs_set_inode_index(new_dir, &index);
9786         if (ret)
9787                 goto out_fail;
9788
9789         BTRFS_I(old_inode)->dir_index = 0ULL;
9790         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9791                 /* force full log commit if subvolume involved. */
9792                 btrfs_set_log_full_commit(root->fs_info, trans);
9793         } else {
9794                 btrfs_pin_log_trans(root);
9795                 log_pinned = true;
9796                 ret = btrfs_insert_inode_ref(trans, dest,
9797                                              new_dentry->d_name.name,
9798                                              new_dentry->d_name.len,
9799                                              old_ino,
9800                                              btrfs_ino(new_dir), index);
9801                 if (ret)
9802                         goto out_fail;
9803         }
9804
9805         inode_inc_iversion(old_dir);
9806         inode_inc_iversion(new_dir);
9807         inode_inc_iversion(old_inode);
9808         old_dir->i_ctime = old_dir->i_mtime =
9809         new_dir->i_ctime = new_dir->i_mtime =
9810         old_inode->i_ctime = current_fs_time(old_dir->i_sb);
9811
9812         if (old_dentry->d_parent != new_dentry->d_parent)
9813                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9814
9815         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9816                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9817                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9818                                         old_dentry->d_name.name,
9819                                         old_dentry->d_name.len);
9820         } else {
9821                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9822                                         d_inode(old_dentry),
9823                                         old_dentry->d_name.name,
9824                                         old_dentry->d_name.len);
9825                 if (!ret)
9826                         ret = btrfs_update_inode(trans, root, old_inode);
9827         }
9828         if (ret) {
9829                 btrfs_abort_transaction(trans, root, ret);
9830                 goto out_fail;
9831         }
9832
9833         if (new_inode) {
9834                 inode_inc_iversion(new_inode);
9835                 new_inode->i_ctime = current_fs_time(new_inode->i_sb);
9836                 if (unlikely(btrfs_ino(new_inode) ==
9837                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9838                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9839                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9840                                                 root_objectid,
9841                                                 new_dentry->d_name.name,
9842                                                 new_dentry->d_name.len);
9843                         BUG_ON(new_inode->i_nlink == 0);
9844                 } else {
9845                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9846                                                  d_inode(new_dentry),
9847                                                  new_dentry->d_name.name,
9848                                                  new_dentry->d_name.len);
9849                 }
9850                 if (!ret && new_inode->i_nlink == 0)
9851                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9852                 if (ret) {
9853                         btrfs_abort_transaction(trans, root, ret);
9854                         goto out_fail;
9855                 }
9856         }
9857
9858         ret = btrfs_add_link(trans, new_dir, old_inode,
9859                              new_dentry->d_name.name,
9860                              new_dentry->d_name.len, 0, index);
9861         if (ret) {
9862                 btrfs_abort_transaction(trans, root, ret);
9863                 goto out_fail;
9864         }
9865
9866         if (old_inode->i_nlink == 1)
9867                 BTRFS_I(old_inode)->dir_index = index;
9868
9869         if (log_pinned) {
9870                 struct dentry *parent = new_dentry->d_parent;
9871
9872                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9873                 btrfs_end_log_trans(root);
9874                 log_pinned = false;
9875         }
9876
9877         if (flags & RENAME_WHITEOUT) {
9878                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9879                                                 old_dentry);
9880
9881                 if (ret) {
9882                         btrfs_abort_transaction(trans, root, ret);
9883                         goto out_fail;
9884                 }
9885         }
9886 out_fail:
9887         /*
9888          * If we have pinned the log and an error happened, we unpin tasks
9889          * trying to sync the log and force them to fallback to a transaction
9890          * commit if the log currently contains any of the inodes involved in
9891          * this rename operation (to ensure we do not persist a log with an
9892          * inconsistent state for any of these inodes or leading to any
9893          * inconsistencies when replayed). If the transaction was aborted, the
9894          * abortion reason is propagated to userspace when attempting to commit
9895          * the transaction. If the log does not contain any of these inodes, we
9896          * allow the tasks to sync it.
9897          */
9898         if (ret && log_pinned) {
9899                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9900                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9901                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9902                     (new_inode &&
9903                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9904                     btrfs_set_log_full_commit(root->fs_info, trans);
9905
9906                 btrfs_end_log_trans(root);
9907                 log_pinned = false;
9908         }
9909         btrfs_end_transaction(trans, root);
9910 out_notrans:
9911         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9912                 up_read(&root->fs_info->subvol_sem);
9913
9914         return ret;
9915 }
9916
9917 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9918                          struct inode *new_dir, struct dentry *new_dentry,
9919                          unsigned int flags)
9920 {
9921         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9922                 return -EINVAL;
9923
9924         if (flags & RENAME_EXCHANGE)
9925                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9926                                           new_dentry);
9927
9928         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9929 }
9930
9931 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9932 {
9933         struct btrfs_delalloc_work *delalloc_work;
9934         struct inode *inode;
9935
9936         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9937                                      work);
9938         inode = delalloc_work->inode;
9939         filemap_flush(inode->i_mapping);
9940         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9941                                 &BTRFS_I(inode)->runtime_flags))
9942                 filemap_flush(inode->i_mapping);
9943
9944         if (delalloc_work->delay_iput)
9945                 btrfs_add_delayed_iput(inode);
9946         else
9947                 iput(inode);
9948         complete(&delalloc_work->completion);
9949 }
9950
9951 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9952                                                     int delay_iput)
9953 {
9954         struct btrfs_delalloc_work *work;
9955
9956         work = kmalloc(sizeof(*work), GFP_NOFS);
9957         if (!work)
9958                 return NULL;
9959
9960         init_completion(&work->completion);
9961         INIT_LIST_HEAD(&work->list);
9962         work->inode = inode;
9963         work->delay_iput = delay_iput;
9964         WARN_ON_ONCE(!inode);
9965         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9966                         btrfs_run_delalloc_work, NULL, NULL);
9967
9968         return work;
9969 }
9970
9971 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9972 {
9973         wait_for_completion(&work->completion);
9974         kfree(work);
9975 }
9976
9977 /*
9978  * some fairly slow code that needs optimization. This walks the list
9979  * of all the inodes with pending delalloc and forces them to disk.
9980  */
9981 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9982                                    int nr)
9983 {
9984         struct btrfs_inode *binode;
9985         struct inode *inode;
9986         struct btrfs_delalloc_work *work, *next;
9987         struct list_head works;
9988         struct list_head splice;
9989         int ret = 0;
9990
9991         INIT_LIST_HEAD(&works);
9992         INIT_LIST_HEAD(&splice);
9993
9994         mutex_lock(&root->delalloc_mutex);
9995         spin_lock(&root->delalloc_lock);
9996         list_splice_init(&root->delalloc_inodes, &splice);
9997         while (!list_empty(&splice)) {
9998                 binode = list_entry(splice.next, struct btrfs_inode,
9999                                     delalloc_inodes);
10000
10001                 list_move_tail(&binode->delalloc_inodes,
10002                                &root->delalloc_inodes);
10003                 inode = igrab(&binode->vfs_inode);
10004                 if (!inode) {
10005                         cond_resched_lock(&root->delalloc_lock);
10006                         continue;
10007                 }
10008                 spin_unlock(&root->delalloc_lock);
10009
10010                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10011                 if (!work) {
10012                         if (delay_iput)
10013                                 btrfs_add_delayed_iput(inode);
10014                         else
10015                                 iput(inode);
10016                         ret = -ENOMEM;
10017                         goto out;
10018                 }
10019                 list_add_tail(&work->list, &works);
10020                 btrfs_queue_work(root->fs_info->flush_workers,
10021                                  &work->work);
10022                 ret++;
10023                 if (nr != -1 && ret >= nr)
10024                         goto out;
10025                 cond_resched();
10026                 spin_lock(&root->delalloc_lock);
10027         }
10028         spin_unlock(&root->delalloc_lock);
10029
10030 out:
10031         list_for_each_entry_safe(work, next, &works, list) {
10032                 list_del_init(&work->list);
10033                 btrfs_wait_and_free_delalloc_work(work);
10034         }
10035
10036         if (!list_empty_careful(&splice)) {
10037                 spin_lock(&root->delalloc_lock);
10038                 list_splice_tail(&splice, &root->delalloc_inodes);
10039                 spin_unlock(&root->delalloc_lock);
10040         }
10041         mutex_unlock(&root->delalloc_mutex);
10042         return ret;
10043 }
10044
10045 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10046 {
10047         int ret;
10048
10049         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
10050                 return -EROFS;
10051
10052         ret = __start_delalloc_inodes(root, delay_iput, -1);
10053         if (ret > 0)
10054                 ret = 0;
10055         /*
10056          * the filemap_flush will queue IO into the worker threads, but
10057          * we have to make sure the IO is actually started and that
10058          * ordered extents get created before we return
10059          */
10060         atomic_inc(&root->fs_info->async_submit_draining);
10061         while (atomic_read(&root->fs_info->nr_async_submits) ||
10062               atomic_read(&root->fs_info->async_delalloc_pages)) {
10063                 wait_event(root->fs_info->async_submit_wait,
10064                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
10065                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
10066         }
10067         atomic_dec(&root->fs_info->async_submit_draining);
10068         return ret;
10069 }
10070
10071 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10072                                int nr)
10073 {
10074         struct btrfs_root *root;
10075         struct list_head splice;
10076         int ret;
10077
10078         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10079                 return -EROFS;
10080
10081         INIT_LIST_HEAD(&splice);
10082
10083         mutex_lock(&fs_info->delalloc_root_mutex);
10084         spin_lock(&fs_info->delalloc_root_lock);
10085         list_splice_init(&fs_info->delalloc_roots, &splice);
10086         while (!list_empty(&splice) && nr) {
10087                 root = list_first_entry(&splice, struct btrfs_root,
10088                                         delalloc_root);
10089                 root = btrfs_grab_fs_root(root);
10090                 BUG_ON(!root);
10091                 list_move_tail(&root->delalloc_root,
10092                                &fs_info->delalloc_roots);
10093                 spin_unlock(&fs_info->delalloc_root_lock);
10094
10095                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10096                 btrfs_put_fs_root(root);
10097                 if (ret < 0)
10098                         goto out;
10099
10100                 if (nr != -1) {
10101                         nr -= ret;
10102                         WARN_ON(nr < 0);
10103                 }
10104                 spin_lock(&fs_info->delalloc_root_lock);
10105         }
10106         spin_unlock(&fs_info->delalloc_root_lock);
10107
10108         ret = 0;
10109         atomic_inc(&fs_info->async_submit_draining);
10110         while (atomic_read(&fs_info->nr_async_submits) ||
10111               atomic_read(&fs_info->async_delalloc_pages)) {
10112                 wait_event(fs_info->async_submit_wait,
10113                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10114                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10115         }
10116         atomic_dec(&fs_info->async_submit_draining);
10117 out:
10118         if (!list_empty_careful(&splice)) {
10119                 spin_lock(&fs_info->delalloc_root_lock);
10120                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10121                 spin_unlock(&fs_info->delalloc_root_lock);
10122         }
10123         mutex_unlock(&fs_info->delalloc_root_mutex);
10124         return ret;
10125 }
10126
10127 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10128                          const char *symname)
10129 {
10130         struct btrfs_trans_handle *trans;
10131         struct btrfs_root *root = BTRFS_I(dir)->root;
10132         struct btrfs_path *path;
10133         struct btrfs_key key;
10134         struct inode *inode = NULL;
10135         int err;
10136         int drop_inode = 0;
10137         u64 objectid;
10138         u64 index = 0;
10139         int name_len;
10140         int datasize;
10141         unsigned long ptr;
10142         struct btrfs_file_extent_item *ei;
10143         struct extent_buffer *leaf;
10144
10145         name_len = strlen(symname);
10146         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
10147                 return -ENAMETOOLONG;
10148
10149         /*
10150          * 2 items for inode item and ref
10151          * 2 items for dir items
10152          * 1 item for updating parent inode item
10153          * 1 item for the inline extent item
10154          * 1 item for xattr if selinux is on
10155          */
10156         trans = btrfs_start_transaction(root, 7);
10157         if (IS_ERR(trans))
10158                 return PTR_ERR(trans);
10159
10160         err = btrfs_find_free_ino(root, &objectid);
10161         if (err)
10162                 goto out_unlock;
10163
10164         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10165                                 dentry->d_name.len, btrfs_ino(dir), objectid,
10166                                 S_IFLNK|S_IRWXUGO, &index);
10167         if (IS_ERR(inode)) {
10168                 err = PTR_ERR(inode);
10169                 goto out_unlock;
10170         }
10171
10172         /*
10173         * If the active LSM wants to access the inode during
10174         * d_instantiate it needs these. Smack checks to see
10175         * if the filesystem supports xattrs by looking at the
10176         * ops vector.
10177         */
10178         inode->i_fop = &btrfs_file_operations;
10179         inode->i_op = &btrfs_file_inode_operations;
10180         inode->i_mapping->a_ops = &btrfs_aops;
10181         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10182
10183         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10184         if (err)
10185                 goto out_unlock_inode;
10186
10187         path = btrfs_alloc_path();
10188         if (!path) {
10189                 err = -ENOMEM;
10190                 goto out_unlock_inode;
10191         }
10192         key.objectid = btrfs_ino(inode);
10193         key.offset = 0;
10194         key.type = BTRFS_EXTENT_DATA_KEY;
10195         datasize = btrfs_file_extent_calc_inline_size(name_len);
10196         err = btrfs_insert_empty_item(trans, root, path, &key,
10197                                       datasize);
10198         if (err) {
10199                 btrfs_free_path(path);
10200                 goto out_unlock_inode;
10201         }
10202         leaf = path->nodes[0];
10203         ei = btrfs_item_ptr(leaf, path->slots[0],
10204                             struct btrfs_file_extent_item);
10205         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10206         btrfs_set_file_extent_type(leaf, ei,
10207                                    BTRFS_FILE_EXTENT_INLINE);
10208         btrfs_set_file_extent_encryption(leaf, ei, 0);
10209         btrfs_set_file_extent_compression(leaf, ei, 0);
10210         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10211         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10212
10213         ptr = btrfs_file_extent_inline_start(ei);
10214         write_extent_buffer(leaf, symname, ptr, name_len);
10215         btrfs_mark_buffer_dirty(leaf);
10216         btrfs_free_path(path);
10217
10218         inode->i_op = &btrfs_symlink_inode_operations;
10219         inode_nohighmem(inode);
10220         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10221         inode_set_bytes(inode, name_len);
10222         btrfs_i_size_write(inode, name_len);
10223         err = btrfs_update_inode(trans, root, inode);
10224         /*
10225          * Last step, add directory indexes for our symlink inode. This is the
10226          * last step to avoid extra cleanup of these indexes if an error happens
10227          * elsewhere above.
10228          */
10229         if (!err)
10230                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
10231         if (err) {
10232                 drop_inode = 1;
10233                 goto out_unlock_inode;
10234         }
10235
10236         unlock_new_inode(inode);
10237         d_instantiate(dentry, inode);
10238
10239 out_unlock:
10240         btrfs_end_transaction(trans, root);
10241         if (drop_inode) {
10242                 inode_dec_link_count(inode);
10243                 iput(inode);
10244         }
10245         btrfs_btree_balance_dirty(root);
10246         return err;
10247
10248 out_unlock_inode:
10249         drop_inode = 1;
10250         unlock_new_inode(inode);
10251         goto out_unlock;
10252 }
10253
10254 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10255                                        u64 start, u64 num_bytes, u64 min_size,
10256                                        loff_t actual_len, u64 *alloc_hint,
10257                                        struct btrfs_trans_handle *trans)
10258 {
10259         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10260         struct extent_map *em;
10261         struct btrfs_root *root = BTRFS_I(inode)->root;
10262         struct btrfs_key ins;
10263         u64 cur_offset = start;
10264         u64 i_size;
10265         u64 cur_bytes;
10266         u64 last_alloc = (u64)-1;
10267         int ret = 0;
10268         bool own_trans = true;
10269
10270         if (trans)
10271                 own_trans = false;
10272         while (num_bytes > 0) {
10273                 if (own_trans) {
10274                         trans = btrfs_start_transaction(root, 3);
10275                         if (IS_ERR(trans)) {
10276                                 ret = PTR_ERR(trans);
10277                                 break;
10278                         }
10279                 }
10280
10281                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10282                 cur_bytes = max(cur_bytes, min_size);
10283                 /*
10284                  * If we are severely fragmented we could end up with really
10285                  * small allocations, so if the allocator is returning small
10286                  * chunks lets make its job easier by only searching for those
10287                  * sized chunks.
10288                  */
10289                 cur_bytes = min(cur_bytes, last_alloc);
10290                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
10291                                            *alloc_hint, &ins, 1, 0);
10292                 if (ret) {
10293                         if (own_trans)
10294                                 btrfs_end_transaction(trans, root);
10295                         break;
10296                 }
10297                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
10298
10299                 last_alloc = ins.offset;
10300                 ret = insert_reserved_file_extent(trans, inode,
10301                                                   cur_offset, ins.objectid,
10302                                                   ins.offset, ins.offset,
10303                                                   ins.offset, 0, 0, 0,
10304                                                   BTRFS_FILE_EXTENT_PREALLOC);
10305                 if (ret) {
10306                         btrfs_free_reserved_extent(root, ins.objectid,
10307                                                    ins.offset, 0);
10308                         btrfs_abort_transaction(trans, root, ret);
10309                         if (own_trans)
10310                                 btrfs_end_transaction(trans, root);
10311                         break;
10312                 }
10313
10314                 btrfs_drop_extent_cache(inode, cur_offset,
10315                                         cur_offset + ins.offset -1, 0);
10316
10317                 em = alloc_extent_map();
10318                 if (!em) {
10319                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10320                                 &BTRFS_I(inode)->runtime_flags);
10321                         goto next;
10322                 }
10323
10324                 em->start = cur_offset;
10325                 em->orig_start = cur_offset;
10326                 em->len = ins.offset;
10327                 em->block_start = ins.objectid;
10328                 em->block_len = ins.offset;
10329                 em->orig_block_len = ins.offset;
10330                 em->ram_bytes = ins.offset;
10331                 em->bdev = root->fs_info->fs_devices->latest_bdev;
10332                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10333                 em->generation = trans->transid;
10334
10335                 while (1) {
10336                         write_lock(&em_tree->lock);
10337                         ret = add_extent_mapping(em_tree, em, 1);
10338                         write_unlock(&em_tree->lock);
10339                         if (ret != -EEXIST)
10340                                 break;
10341                         btrfs_drop_extent_cache(inode, cur_offset,
10342                                                 cur_offset + ins.offset - 1,
10343                                                 0);
10344                 }
10345                 free_extent_map(em);
10346 next:
10347                 num_bytes -= ins.offset;
10348                 cur_offset += ins.offset;
10349                 *alloc_hint = ins.objectid + ins.offset;
10350
10351                 inode_inc_iversion(inode);
10352                 inode->i_ctime = current_fs_time(inode->i_sb);
10353                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10354                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10355                     (actual_len > inode->i_size) &&
10356                     (cur_offset > inode->i_size)) {
10357                         if (cur_offset > actual_len)
10358                                 i_size = actual_len;
10359                         else
10360                                 i_size = cur_offset;
10361                         i_size_write(inode, i_size);
10362                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10363                 }
10364
10365                 ret = btrfs_update_inode(trans, root, inode);
10366
10367                 if (ret) {
10368                         btrfs_abort_transaction(trans, root, ret);
10369                         if (own_trans)
10370                                 btrfs_end_transaction(trans, root);
10371                         break;
10372                 }
10373
10374                 if (own_trans)
10375                         btrfs_end_transaction(trans, root);
10376         }
10377         return ret;
10378 }
10379
10380 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10381                               u64 start, u64 num_bytes, u64 min_size,
10382                               loff_t actual_len, u64 *alloc_hint)
10383 {
10384         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10385                                            min_size, actual_len, alloc_hint,
10386                                            NULL);
10387 }
10388
10389 int btrfs_prealloc_file_range_trans(struct inode *inode,
10390                                     struct btrfs_trans_handle *trans, int mode,
10391                                     u64 start, u64 num_bytes, u64 min_size,
10392                                     loff_t actual_len, u64 *alloc_hint)
10393 {
10394         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10395                                            min_size, actual_len, alloc_hint, trans);
10396 }
10397
10398 static int btrfs_set_page_dirty(struct page *page)
10399 {
10400         return __set_page_dirty_nobuffers(page);
10401 }
10402
10403 static int btrfs_permission(struct inode *inode, int mask)
10404 {
10405         struct btrfs_root *root = BTRFS_I(inode)->root;
10406         umode_t mode = inode->i_mode;
10407
10408         if (mask & MAY_WRITE &&
10409             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10410                 if (btrfs_root_readonly(root))
10411                         return -EROFS;
10412                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10413                         return -EACCES;
10414         }
10415         return generic_permission(inode, mask);
10416 }
10417
10418 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10419 {
10420         struct btrfs_trans_handle *trans;
10421         struct btrfs_root *root = BTRFS_I(dir)->root;
10422         struct inode *inode = NULL;
10423         u64 objectid;
10424         u64 index;
10425         int ret = 0;
10426
10427         /*
10428          * 5 units required for adding orphan entry
10429          */
10430         trans = btrfs_start_transaction(root, 5);
10431         if (IS_ERR(trans))
10432                 return PTR_ERR(trans);
10433
10434         ret = btrfs_find_free_ino(root, &objectid);
10435         if (ret)
10436                 goto out;
10437
10438         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10439                                 btrfs_ino(dir), objectid, mode, &index);
10440         if (IS_ERR(inode)) {
10441                 ret = PTR_ERR(inode);
10442                 inode = NULL;
10443                 goto out;
10444         }
10445
10446         inode->i_fop = &btrfs_file_operations;
10447         inode->i_op = &btrfs_file_inode_operations;
10448
10449         inode->i_mapping->a_ops = &btrfs_aops;
10450         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10451
10452         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10453         if (ret)
10454                 goto out_inode;
10455
10456         ret = btrfs_update_inode(trans, root, inode);
10457         if (ret)
10458                 goto out_inode;
10459         ret = btrfs_orphan_add(trans, inode);
10460         if (ret)
10461                 goto out_inode;
10462
10463         /*
10464          * We set number of links to 0 in btrfs_new_inode(), and here we set
10465          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10466          * through:
10467          *
10468          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10469          */
10470         set_nlink(inode, 1);
10471         unlock_new_inode(inode);
10472         d_tmpfile(dentry, inode);
10473         mark_inode_dirty(inode);
10474
10475 out:
10476         btrfs_end_transaction(trans, root);
10477         if (ret)
10478                 iput(inode);
10479         btrfs_balance_delayed_items(root);
10480         btrfs_btree_balance_dirty(root);
10481         return ret;
10482
10483 out_inode:
10484         unlock_new_inode(inode);
10485         goto out;
10486
10487 }
10488
10489 /* Inspired by filemap_check_errors() */
10490 int btrfs_inode_check_errors(struct inode *inode)
10491 {
10492         int ret = 0;
10493
10494         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10495             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10496                 ret = -ENOSPC;
10497         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10498             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10499                 ret = -EIO;
10500
10501         return ret;
10502 }
10503
10504 static const struct inode_operations btrfs_dir_inode_operations = {
10505         .getattr        = btrfs_getattr,
10506         .lookup         = btrfs_lookup,
10507         .create         = btrfs_create,
10508         .unlink         = btrfs_unlink,
10509         .link           = btrfs_link,
10510         .mkdir          = btrfs_mkdir,
10511         .rmdir          = btrfs_rmdir,
10512         .rename2        = btrfs_rename2,
10513         .symlink        = btrfs_symlink,
10514         .setattr        = btrfs_setattr,
10515         .mknod          = btrfs_mknod,
10516         .setxattr       = generic_setxattr,
10517         .getxattr       = generic_getxattr,
10518         .listxattr      = btrfs_listxattr,
10519         .removexattr    = generic_removexattr,
10520         .permission     = btrfs_permission,
10521         .get_acl        = btrfs_get_acl,
10522         .set_acl        = btrfs_set_acl,
10523         .update_time    = btrfs_update_time,
10524         .tmpfile        = btrfs_tmpfile,
10525 };
10526 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10527         .lookup         = btrfs_lookup,
10528         .permission     = btrfs_permission,
10529         .get_acl        = btrfs_get_acl,
10530         .set_acl        = btrfs_set_acl,
10531         .update_time    = btrfs_update_time,
10532 };
10533
10534 static const struct file_operations btrfs_dir_file_operations = {
10535         .llseek         = generic_file_llseek,
10536         .read           = generic_read_dir,
10537         .iterate        = btrfs_real_readdir,
10538         .unlocked_ioctl = btrfs_ioctl,
10539 #ifdef CONFIG_COMPAT
10540         .compat_ioctl   = btrfs_compat_ioctl,
10541 #endif
10542         .release        = btrfs_release_file,
10543         .fsync          = btrfs_sync_file,
10544 };
10545
10546 static const struct extent_io_ops btrfs_extent_io_ops = {
10547         .fill_delalloc = run_delalloc_range,
10548         .submit_bio_hook = btrfs_submit_bio_hook,
10549         .merge_bio_hook = btrfs_merge_bio_hook,
10550         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10551         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10552         .writepage_start_hook = btrfs_writepage_start_hook,
10553         .set_bit_hook = btrfs_set_bit_hook,
10554         .clear_bit_hook = btrfs_clear_bit_hook,
10555         .merge_extent_hook = btrfs_merge_extent_hook,
10556         .split_extent_hook = btrfs_split_extent_hook,
10557 };
10558
10559 /*
10560  * btrfs doesn't support the bmap operation because swapfiles
10561  * use bmap to make a mapping of extents in the file.  They assume
10562  * these extents won't change over the life of the file and they
10563  * use the bmap result to do IO directly to the drive.
10564  *
10565  * the btrfs bmap call would return logical addresses that aren't
10566  * suitable for IO and they also will change frequently as COW
10567  * operations happen.  So, swapfile + btrfs == corruption.
10568  *
10569  * For now we're avoiding this by dropping bmap.
10570  */
10571 static const struct address_space_operations btrfs_aops = {
10572         .readpage       = btrfs_readpage,
10573         .writepage      = btrfs_writepage,
10574         .writepages     = btrfs_writepages,
10575         .readpages      = btrfs_readpages,
10576         .direct_IO      = btrfs_direct_IO,
10577         .invalidatepage = btrfs_invalidatepage,
10578         .releasepage    = btrfs_releasepage,
10579         .set_page_dirty = btrfs_set_page_dirty,
10580         .error_remove_page = generic_error_remove_page,
10581 };
10582
10583 static const struct address_space_operations btrfs_symlink_aops = {
10584         .readpage       = btrfs_readpage,
10585         .writepage      = btrfs_writepage,
10586         .invalidatepage = btrfs_invalidatepage,
10587         .releasepage    = btrfs_releasepage,
10588 };
10589
10590 static const struct inode_operations btrfs_file_inode_operations = {
10591         .getattr        = btrfs_getattr,
10592         .setattr        = btrfs_setattr,
10593         .setxattr       = generic_setxattr,
10594         .getxattr       = generic_getxattr,
10595         .listxattr      = btrfs_listxattr,
10596         .removexattr    = generic_removexattr,
10597         .permission     = btrfs_permission,
10598         .fiemap         = btrfs_fiemap,
10599         .get_acl        = btrfs_get_acl,
10600         .set_acl        = btrfs_set_acl,
10601         .update_time    = btrfs_update_time,
10602 };
10603 static const struct inode_operations btrfs_special_inode_operations = {
10604         .getattr        = btrfs_getattr,
10605         .setattr        = btrfs_setattr,
10606         .permission     = btrfs_permission,
10607         .setxattr       = generic_setxattr,
10608         .getxattr       = generic_getxattr,
10609         .listxattr      = btrfs_listxattr,
10610         .removexattr    = generic_removexattr,
10611         .get_acl        = btrfs_get_acl,
10612         .set_acl        = btrfs_set_acl,
10613         .update_time    = btrfs_update_time,
10614 };
10615 static const struct inode_operations btrfs_symlink_inode_operations = {
10616         .readlink       = generic_readlink,
10617         .get_link       = page_get_link,
10618         .getattr        = btrfs_getattr,
10619         .setattr        = btrfs_setattr,
10620         .permission     = btrfs_permission,
10621         .setxattr       = generic_setxattr,
10622         .getxattr       = generic_getxattr,
10623         .listxattr      = btrfs_listxattr,
10624         .removexattr    = generic_removexattr,
10625         .update_time    = btrfs_update_time,
10626 };
10627
10628 const struct dentry_operations btrfs_dentry_operations = {
10629         .d_delete       = btrfs_dentry_delete,
10630         .d_release      = btrfs_dentry_release,
10631 };