fs/btrfs/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/fs.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/time.h>
   9 #include <linux/init.h>
  10 #include <linux/string.h>
  11 #include <linux/backing-dev.h>
  12 #include <linux/falloc.h>
  13 #include <linux/writeback.h>
  14 #include <linux/compat.h>
  15 #include <linux/slab.h>
  16 #include <linux/btrfs.h>
  17 #include <linux/uio.h>
  18 #include <linux/iversion.h>
  19 #include "ctree.h"
  20 #include "disk-io.h"
  21 #include "transaction.h"
  22 #include "btrfs_inode.h"
  23 #include "print-tree.h"
  24 #include "tree-log.h"
  25 #include "locking.h"
  26 #include "volumes.h"
  27 #include "qgroup.h"
  28 #include "compression.h"
  29
  30 static struct kmem_cache *btrfs_inode_defrag_cachep;
  31 /*
  32  * when auto defrag is enabled we
  33  * queue up these defrag structs to remember which
  34  * inodes need defragging passes
  35  */
  36 struct inode_defrag {
  37         struct rb_node rb_node;
  38         /* objectid */
  39         u64 ino;
  40         /*
  41          * transid where the defrag was added, we search for
  42          * extents newer than this
  43          */
  44         u64 transid;
  45
  46         /* root objectid */
  47         u64 root;
  48
  49         /* last offset we were able to defrag */
  50         u64 last_offset;
  51
  52         /* if we've wrapped around back to zero once already */
  53         int cycled;
  54 };
  55
  56 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  57                                   struct inode_defrag *defrag2)
  58 {
  59         if (defrag1->root > defrag2->root)
  60                 return 1;
  61         else if (defrag1->root < defrag2->root)
  62                 return -1;
  63         else if (defrag1->ino > defrag2->ino)
  64                 return 1;
  65         else if (defrag1->ino < defrag2->ino)
  66                 return -1;
  67         else
  68                 return 0;
  69 }
  70
  71 /* pop a record for an inode into the defrag tree.  The lock
  72  * must be held already
  73  *
  74  * If you're inserting a record for an older transid than an
  75  * existing record, the transid already in the tree is lowered
  76  *
  77  * If an existing record is found the defrag item you
  78  * pass in is freed
  79  */
  80 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  81                                     struct inode_defrag *defrag)
  82 {
  83         struct btrfs_fs_info *fs_info = inode->root->fs_info;
  84         struct inode_defrag *entry;
  85         struct rb_node **p;
  86         struct rb_node *parent = NULL;
  87         int ret;
  88
  89         p = &fs_info->defrag_inodes.rb_node;
  90         while (*p) {
  91                 parent = *p;
  92                 entry = rb_entry(parent, struct inode_defrag, rb_node);
  93
  94                 ret = __compare_inode_defrag(defrag, entry);
  95                 if (ret < 0)
  96                         p = &parent->rb_left;
  97                 else if (ret > 0)
  98                         p = &parent->rb_right;
  99                 else {
 100                         /* if we're reinserting an entry for
 101                          * an old defrag run, make sure to
 102                          * lower the transid of our existing record
 103                          */
 104                         if (defrag->transid < entry->transid)
 105                                 entry->transid = defrag->transid;
 106                         if (defrag->last_offset > entry->last_offset)
 107                                 entry->last_offset = defrag->last_offset;
 108                         return -EEXIST;
 109                 }
 110         }
 111         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 112         rb_link_node(&defrag->rb_node, parent, p);
 113         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 114         return 0;
 115 }
 116
 117 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 118 {
 119         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 120                 return 0;
 121
 122         if (btrfs_fs_closing(fs_info))
 123                 return 0;
 124
 125         return 1;
 126 }
 127
 128 /*
 129  * insert a defrag record for this inode if auto defrag is
 130  * enabled
 131  */
 132 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 133                            struct btrfs_inode *inode)
 134 {
 135         struct btrfs_root *root = inode->root;
 136         struct btrfs_fs_info *fs_info = root->fs_info;
 137         struct inode_defrag *defrag;
 138         u64 transid;
 139         int ret;
 140
 141         if (!__need_auto_defrag(fs_info))
 142                 return 0;
 143
 144         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 145                 return 0;
 146
 147         if (trans)
 148                 transid = trans->transid;
 149         else
 150                 transid = inode->root->last_trans;
 151
 152         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 153         if (!defrag)
 154                 return -ENOMEM;
 155
 156         defrag->ino = btrfs_ino(inode);
 157         defrag->transid = transid;
 158         defrag->root = root->root_key.objectid;
 159
 160         spin_lock(&fs_info->defrag_inodes_lock);
 161         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 162                 /*
 163                  * If we set IN_DEFRAG flag and evict the inode from memory,
 164                  * and then re-read this inode, this new inode doesn't have
 165                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 166                  */
 167                 ret = __btrfs_add_inode_defrag(inode, defrag);
 168                 if (ret)
 169                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 170         } else {
 171                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 172         }
 173         spin_unlock(&fs_info->defrag_inodes_lock);
 174         return 0;
 175 }
 176
 177 /*
 178  * Requeue the defrag object. If there is a defrag object that points to
 179  * the same inode in the tree, we will merge them together (by
 180  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 181  */
 182 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 183                                        struct inode_defrag *defrag)
 184 {
 185         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 186         int ret;
 187
 188         if (!__need_auto_defrag(fs_info))
 189                 goto out;
 190
 191         /*
 192          * Here we don't check the IN_DEFRAG flag, because we need merge
 193          * them together.
 194          */
 195         spin_lock(&fs_info->defrag_inodes_lock);
 196         ret = __btrfs_add_inode_defrag(inode, defrag);
 197         spin_unlock(&fs_info->defrag_inodes_lock);
 198         if (ret)
 199                 goto out;
 200         return;
 201 out:
 202         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 203 }
 204
 205 /*
 206  * pick the defragable inode that we want, if it doesn't exist, we will get
 207  * the next one.
 208  */
 209 static struct inode_defrag *
 210 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 211 {
 212         struct inode_defrag *entry = NULL;
 213         struct inode_defrag tmp;
 214         struct rb_node *p;
 215         struct rb_node *parent = NULL;
 216         int ret;
 217
 218         tmp.ino = ino;
 219         tmp.root = root;
 220
 221         spin_lock(&fs_info->defrag_inodes_lock);
 222         p = fs_info->defrag_inodes.rb_node;
 223         while (p) {
 224                 parent = p;
 225                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 226
 227                 ret = __compare_inode_defrag(&tmp, entry);
 228                 if (ret < 0)
 229                         p = parent->rb_left;
 230                 else if (ret > 0)
 231                         p = parent->rb_right;
 232                 else
 233                         goto out;
 234         }
 235
 236         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 237                 parent = rb_next(parent);
 238                 if (parent)
 239                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 240                 else
 241                         entry = NULL;
 242         }
 243 out:
 244         if (entry)
 245                 rb_erase(parent, &fs_info->defrag_inodes);
 246         spin_unlock(&fs_info->defrag_inodes_lock);
 247         return entry;
 248 }
 249
 250 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 251 {
 252         struct inode_defrag *defrag;
 253         struct rb_node *node;
 254
 255         spin_lock(&fs_info->defrag_inodes_lock);
 256         node = rb_first(&fs_info->defrag_inodes);
 257         while (node) {
 258                 rb_erase(node, &fs_info->defrag_inodes);
 259                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 260                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 261
 262                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 263
 264                 node = rb_first(&fs_info->defrag_inodes);
 265         }
 266         spin_unlock(&fs_info->defrag_inodes_lock);
 267 }
 268
 269 #define BTRFS_DEFRAG_BATCH      1024
 270
 271 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 272                                     struct inode_defrag *defrag)
 273 {
 274         struct btrfs_root *inode_root;
 275         struct inode *inode;
 276         struct btrfs_key key;
 277         struct btrfs_ioctl_defrag_range_args range;
 278         int num_defrag;
 279         int index;
 280         int ret;
 281
 282         /* get the inode */
 283         key.objectid = defrag->root;
 284         key.type = BTRFS_ROOT_ITEM_KEY;
 285         key.offset = (u64)-1;
 286
 287         index = srcu_read_lock(&fs_info->subvol_srcu);
 288
 289         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 290         if (IS_ERR(inode_root)) {
 291                 ret = PTR_ERR(inode_root);
 292                 goto cleanup;
 293         }
 294
 295         key.objectid = defrag->ino;
 296         key.type = BTRFS_INODE_ITEM_KEY;
 297         key.offset = 0;
 298         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 299         if (IS_ERR(inode)) {
 300                 ret = PTR_ERR(inode);
 301                 goto cleanup;
 302         }
 303         srcu_read_unlock(&fs_info->subvol_srcu, index);
 304
 305         /* do a chunk of defrag */
 306         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 307         memset(&range, 0, sizeof(range));
 308         range.len = (u64)-1;
 309         range.start = defrag->last_offset;
 310
 311         sb_start_write(fs_info->sb);
 312         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 313                                        BTRFS_DEFRAG_BATCH);
 314         sb_end_write(fs_info->sb);
 315         /*
 316          * if we filled the whole defrag batch, there
 317          * must be more work to do.  Queue this defrag
 318          * again
 319          */
 320         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 321                 defrag->last_offset = range.start;
 322                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 323         } else if (defrag->last_offset && !defrag->cycled) {
 324                 /*
 325                  * we didn't fill our defrag batch, but
 326                  * we didn't start at zero.  Make sure we loop
 327                  * around to the start of the file.
 328                  */
 329                 defrag->last_offset = 0;
 330                 defrag->cycled = 1;
 331                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 332         } else {
 333                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 334         }
 335
 336         iput(inode);
 337         return 0;
 338 cleanup:
 339         srcu_read_unlock(&fs_info->subvol_srcu, index);
 340         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 341         return ret;
 342 }
 343
 344 /*
 345  * run through the list of inodes in the FS that need
 346  * defragging
 347  */
 348 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 349 {
 350         struct inode_defrag *defrag;
 351         u64 first_ino = 0;
 352         u64 root_objectid = 0;
 353
 354         atomic_inc(&fs_info->defrag_running);
 355         while (1) {
 356                 /* Pause the auto defragger. */
 357                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 358                              &fs_info->fs_state))
 359                         break;
 360
 361                 if (!__need_auto_defrag(fs_info))
 362                         break;
 363
 364                 /* find an inode to defrag */
 365                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 366                                                  first_ino);
 367                 if (!defrag) {
 368                         if (root_objectid || first_ino) {
 369                                 root_objectid = 0;
 370                                 first_ino = 0;
 371                                 continue;
 372                         } else {
 373                                 break;
 374                         }
 375                 }
 376
 377                 first_ino = defrag->ino + 1;
 378                 root_objectid = defrag->root;
 379
 380                 __btrfs_run_defrag_inode(fs_info, defrag);
 381         }
 382         atomic_dec(&fs_info->defrag_running);
 383
 384         /*
 385          * during unmount, we use the transaction_wait queue to
 386          * wait for the defragger to stop
 387          */
 388         wake_up(&fs_info->transaction_wait);
 389         return 0;
 390 }
 391
 392 /* simple helper to fault in pages and copy.  This should go away
 393  * and be replaced with calls into generic code.
 394  */
 395 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 396                                          struct page **prepared_pages,
 397                                          struct iov_iter *i)
 398 {
 399         size_t copied = 0;
 400         size_t total_copied = 0;
 401         int pg = 0;
 402         int offset = offset_in_page(pos);
 403
 404         while (write_bytes > 0) {
 405                 size_t count = min_t(size_t,
 406                                      PAGE_SIZE - offset, write_bytes);
 407                 struct page *page = prepared_pages[pg];
 408                 /*
 409                  * Copy data from userspace to the current page
 410                  */
 411                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 412
 413                 /* Flush processor's dcache for this page */
 414                 flush_dcache_page(page);
 415
 416                 /*
 417                  * if we get a partial write, we can end up with
 418                  * partially up to date pages.  These add
 419                  * a lot of complexity, so make sure they don't
 420                  * happen by forcing this copy to be retried.
 421                  *
 422                  * The rest of the btrfs_file_write code will fall
 423                  * back to page at a time copies after we return 0.
 424                  */
 425                 if (!PageUptodate(page) && copied < count)
 426                         copied = 0;
 427
 428                 iov_iter_advance(i, copied);
 429                 write_bytes -= copied;
 430                 total_copied += copied;
 431
 432                 /* Return to btrfs_file_write_iter to fault page */
 433                 if (unlikely(copied == 0))
 434                         break;
 435
 436                 if (copied < PAGE_SIZE - offset) {
 437                         offset += copied;
 438                 } else {
 439                         pg++;
 440                         offset = 0;
 441                 }
 442         }
 443         return total_copied;
 444 }
 445
 446 /*
 447  * unlocks pages after btrfs_file_write is done with them
 448  */
 449 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 450 {
 451         size_t i;
 452         for (i = 0; i < num_pages; i++) {
 453                 /* page checked is some magic around finding pages that
 454                  * have been modified without going through btrfs_set_page_dirty
 455                  * clear it here. There should be no need to mark the pages
 456                  * accessed as prepare_pages should have marked them accessed
 457                  * in prepare_pages via find_or_create_page()
 458                  */
 459                 ClearPageChecked(pages[i]);
 460                 unlock_page(pages[i]);
 461                 put_page(pages[i]);
 462         }
 463 }
 464
 465 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
 466                                          const u64 start,
 467                                          const u64 len,
 468                                          struct extent_state **cached_state)
 469 {
 470         u64 search_start = start;
 471         const u64 end = start + len - 1;
 472
 473         while (search_start < end) {
 474                 const u64 search_len = end - search_start + 1;
 475                 struct extent_map *em;
 476                 u64 em_len;
 477                 int ret = 0;
 478
 479                 em = btrfs_get_extent(inode, NULL, 0, search_start,
 480                                       search_len, 0);
 481                 if (IS_ERR(em))
 482                         return PTR_ERR(em);
 483
 484                 if (em->block_start != EXTENT_MAP_HOLE)
 485                         goto next;
 486
 487                 em_len = em->len;
 488                 if (em->start < search_start)
 489                         em_len -= search_start - em->start;
 490                 if (em_len > search_len)
 491                         em_len = search_len;
 492
 493                 ret = set_extent_bit(&inode->io_tree, search_start,
 494                                      search_start + em_len - 1,
 495                                      EXTENT_DELALLOC_NEW,
 496                                      NULL, cached_state, GFP_NOFS);
 497 next:
 498                 search_start = extent_map_end(em);
 499                 free_extent_map(em);
 500                 if (ret)
 501                         return ret;
 502         }
 503         return 0;
 504 }
 505
 506 /*
 507  * after copy_from_user, pages need to be dirtied and we need to make
 508  * sure holes are created between the current EOF and the start of
 509  * any next extents (if required).
 510  *
 511  * this also makes the decision about creating an inline extent vs
 512  * doing real data extents, marking pages dirty and delalloc as required.
 513  */
 514 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
 515                       size_t num_pages, loff_t pos, size_t write_bytes,
 516                       struct extent_state **cached)
 517 {
 518         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 519         int err = 0;
 520         int i;
 521         u64 num_bytes;
 522         u64 start_pos;
 523         u64 end_of_last_block;
 524         u64 end_pos = pos + write_bytes;
 525         loff_t isize = i_size_read(inode);
 526         unsigned int extra_bits = 0;
 527
 528         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 529         num_bytes = round_up(write_bytes + pos - start_pos,
 530                              fs_info->sectorsize);
 531
 532         end_of_last_block = start_pos + num_bytes - 1;
 533
 534         /*
 535          * The pages may have already been dirty, clear out old accounting so
 536          * we can set things up properly
 537          */
 538         clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos, end_of_last_block,
 539                          EXTENT_DIRTY | EXTENT_DELALLOC |
 540                          EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 0, 0, cached);
 541
 542         if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
 543                 if (start_pos >= isize &&
 544                     !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
 545                         /*
 546                          * There can't be any extents following eof in this case
 547                          * so just set the delalloc new bit for the range
 548                          * directly.
 549                          */
 550                         extra_bits |= EXTENT_DELALLOC_NEW;
 551                 } else {
 552                         err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
 553                                                             start_pos,
 554                                                             num_bytes, cached);
 555                         if (err)
 556                                 return err;
 557                 }
 558         }
 559
 560         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 561                                         extra_bits, cached, 0);
 562         if (err)
 563                 return err;
 564
 565         for (i = 0; i < num_pages; i++) {
 566                 struct page *p = pages[i];
 567                 SetPageUptodate(p);
 568                 ClearPageChecked(p);
 569                 set_page_dirty(p);
 570         }
 571
 572         /*
 573          * we've only changed i_size in ram, and we haven't updated
 574          * the disk i_size.  There is no need to log the inode
 575          * at this time.
 576          */
 577         if (end_pos > isize)
 578                 i_size_write(inode, end_pos);
 579         return 0;
 580 }
 581
 582 /*
 583  * this drops all the extents in the cache that intersect the range
 584  * [start, end].  Existing extents are split as required.
 585  */
 586 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 587                              int skip_pinned)
 588 {
 589         struct extent_map *em;
 590         struct extent_map *split = NULL;
 591         struct extent_map *split2 = NULL;
 592         struct extent_map_tree *em_tree = &inode->extent_tree;
 593         u64 len = end - start + 1;
 594         u64 gen;
 595         int ret;
 596         int testend = 1;
 597         unsigned long flags;
 598         int compressed = 0;
 599         bool modified;
 600
 601         WARN_ON(end < start);
 602         if (end == (u64)-1) {
 603                 len = (u64)-1;
 604                 testend = 0;
 605         }
 606         while (1) {
 607                 int no_splits = 0;
 608
 609                 modified = false;
 610                 if (!split)
 611                         split = alloc_extent_map();
 612                 if (!split2)
 613                         split2 = alloc_extent_map();
 614                 if (!split || !split2)
 615                         no_splits = 1;
 616
 617                 write_lock(&em_tree->lock);
 618                 em = lookup_extent_mapping(em_tree, start, len);
 619                 if (!em) {
 620                         write_unlock(&em_tree->lock);
 621                         break;
 622                 }
 623                 flags = em->flags;
 624                 gen = em->generation;
 625                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 626                         if (testend && em->start + em->len >= start + len) {
 627                                 free_extent_map(em);
 628                                 write_unlock(&em_tree->lock);
 629                                 break;
 630                         }
 631                         start = em->start + em->len;
 632                         if (testend)
 633                                 len = start + len - (em->start + em->len);
 634                         free_extent_map(em);
 635                         write_unlock(&em_tree->lock);
 636                         continue;
 637                 }
 638                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 639                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 640                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 641                 modified = !list_empty(&em->list);
 642                 if (no_splits)
 643                         goto next;
 644
 645                 if (em->start < start) {
 646                         split->start = em->start;
 647                         split->len = start - em->start;
 648
 649                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 650                                 split->orig_start = em->orig_start;
 651                                 split->block_start = em->block_start;
 652
 653                                 if (compressed)
 654                                         split->block_len = em->block_len;
 655                                 else
 656                                         split->block_len = split->len;
 657                                 split->orig_block_len = max(split->block_len,
 658                                                 em->orig_block_len);
 659                                 split->ram_bytes = em->ram_bytes;
 660                         } else {
 661                                 split->orig_start = split->start;
 662                                 split->block_len = 0;
 663                                 split->block_start = em->block_start;
 664                                 split->orig_block_len = 0;
 665                                 split->ram_bytes = split->len;
 666                         }
 667
 668                         split->generation = gen;
 669                         split->bdev = em->bdev;
 670                         split->flags = flags;
 671                         split->compress_type = em->compress_type;
 672                         replace_extent_mapping(em_tree, em, split, modified);
 673                         free_extent_map(split);
 674                         split = split2;
 675                         split2 = NULL;
 676                 }
 677                 if (testend && em->start + em->len > start + len) {
 678                         u64 diff = start + len - em->start;
 679
 680                         split->start = start + len;
 681                         split->len = em->start + em->len - (start + len);
 682                         split->bdev = em->bdev;
 683                         split->flags = flags;
 684                         split->compress_type = em->compress_type;
 685                         split->generation = gen;
 686
 687                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 688                                 split->orig_block_len = max(em->block_len,
 689                                                     em->orig_block_len);
 690
 691                                 split->ram_bytes = em->ram_bytes;
 692                                 if (compressed) {
 693                                         split->block_len = em->block_len;
 694                                         split->block_start = em->block_start;
 695                                         split->orig_start = em->orig_start;
 696                                 } else {
 697                                         split->block_len = split->len;
 698                                         split->block_start = em->block_start
 699                                                 + diff;
 700                                         split->orig_start = em->orig_start;
 701                                 }
 702                         } else {
 703                                 split->ram_bytes = split->len;
 704                                 split->orig_start = split->start;
 705                                 split->block_len = 0;
 706                                 split->block_start = em->block_start;
 707                                 split->orig_block_len = 0;
 708                         }
 709
 710                         if (extent_map_in_tree(em)) {
 711                                 replace_extent_mapping(em_tree, em, split,
 712                                                        modified);
 713                         } else {
 714                                 ret = add_extent_mapping(em_tree, split,
 715                                                          modified);
 716                                 ASSERT(ret == 0); /* Logic error */
 717                         }
 718                         free_extent_map(split);
 719                         split = NULL;
 720                 }
 721 next:
 722                 if (extent_map_in_tree(em))
 723                         remove_extent_mapping(em_tree, em);
 724                 write_unlock(&em_tree->lock);
 725
 726                 /* once for us */
 727                 free_extent_map(em);
 728                 /* once for the tree*/
 729                 free_extent_map(em);
 730         }
 731         if (split)
 732                 free_extent_map(split);
 733         if (split2)
 734                 free_extent_map(split2);
 735 }
 736
 737 /*
 738  * this is very complex, but the basic idea is to drop all extents
 739  * in the range start - end.  hint_block is filled in with a block number
 740  * that would be a good hint to the block allocator for this file.
 741  *
 742  * If an extent intersects the range but is not entirely inside the range
 743  * it is either truncated or split.  Anything entirely inside the range
 744  * is deleted from the tree.
 745  */
 746 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 747                          struct btrfs_root *root, struct inode *inode,
 748                          struct btrfs_path *path, u64 start, u64 end,
 749                          u64 *drop_end, int drop_cache,
 750                          int replace_extent,
 751                          u32 extent_item_size,
 752                          int *key_inserted)
 753 {
 754         struct btrfs_fs_info *fs_info = root->fs_info;
 755         struct extent_buffer *leaf;
 756         struct btrfs_file_extent_item *fi;
 757         struct btrfs_ref ref = { 0 };
 758         struct btrfs_key key;
 759         struct btrfs_key new_key;
 760         u64 ino = btrfs_ino(BTRFS_I(inode));
 761         u64 search_start = start;
 762         u64 disk_bytenr = 0;
 763         u64 num_bytes = 0;
 764         u64 extent_offset = 0;
 765         u64 extent_end = 0;
 766         u64 last_end = start;
 767         int del_nr = 0;
 768         int del_slot = 0;
 769         int extent_type;
 770         int recow;
 771         int ret;
 772         int modify_tree = -1;
 773         int update_refs;
 774         int found = 0;
 775         int leafs_visited = 0;
 776
 777         if (drop_cache)
 778                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
 779
 780         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 781                 modify_tree = 0;
 782
 783         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 784                        root == fs_info->tree_root);
 785         while (1) {
 786                 recow = 0;
 787                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 788                                                search_start, modify_tree);
 789                 if (ret < 0)
 790                         break;
 791                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 792                         leaf = path->nodes[0];
 793                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 794                         if (key.objectid == ino &&
 795                             key.type == BTRFS_EXTENT_DATA_KEY)
 796                                 path->slots[0]--;
 797                 }
 798                 ret = 0;
 799                 leafs_visited++;
 800 next_slot:
 801                 leaf = path->nodes[0];
 802                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 803                         BUG_ON(del_nr > 0);
 804                         ret = btrfs_next_leaf(root, path);
 805                         if (ret < 0)
 806                                 break;
 807                         if (ret > 0) {
 808                                 ret = 0;
 809                                 break;
 810                         }
 811                         leafs_visited++;
 812                         leaf = path->nodes[0];
 813                         recow = 1;
 814                 }
 815
 816                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 817
 818                 if (key.objectid > ino)
 819                         break;
 820                 if (WARN_ON_ONCE(key.objectid < ino) ||
 821                     key.type < BTRFS_EXTENT_DATA_KEY) {
 822                         ASSERT(del_nr == 0);
 823                         path->slots[0]++;
 824                         goto next_slot;
 825                 }
 826                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 827                         break;
 828
 829                 fi = btrfs_item_ptr(leaf, path->slots[0],
 830                                     struct btrfs_file_extent_item);
 831                 extent_type = btrfs_file_extent_type(leaf, fi);
 832
 833                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 834                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 835                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 836                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 837                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 838                         extent_end = key.offset +
 839                                 btrfs_file_extent_num_bytes(leaf, fi);
 840                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 841                         extent_end = key.offset +
 842                                 btrfs_file_extent_ram_bytes(leaf, fi);
 843                 } else {
 844                         /* can't happen */
 845                         BUG();
 846                 }
 847
 848                 /*
 849                  * Don't skip extent items representing 0 byte lengths. They
 850                  * used to be created (bug) if while punching holes we hit
 851                  * -ENOSPC condition. So if we find one here, just ensure we
 852                  * delete it, otherwise we would insert a new file extent item
 853                  * with the same key (offset) as that 0 bytes length file
 854                  * extent item in the call to setup_items_for_insert() later
 855                  * in this function.
 856                  */
 857                 if (extent_end == key.offset && extent_end >= search_start) {
 858                         last_end = extent_end;
 859                         goto delete_extent_item;
 860                 }
 861
 862                 if (extent_end <= search_start) {
 863                         path->slots[0]++;
 864                         goto next_slot;
 865                 }
 866
 867                 found = 1;
 868                 search_start = max(key.offset, start);
 869                 if (recow || !modify_tree) {
 870                         modify_tree = -1;
 871                         btrfs_release_path(path);
 872                         continue;
 873                 }
 874
 875                 /*
 876                  *     | - range to drop - |
 877                  *  | -------- extent -------- |
 878                  */
 879                 if (start > key.offset && end < extent_end) {
 880                         BUG_ON(del_nr > 0);
 881                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 882                                 ret = -EOPNOTSUPP;
 883                                 break;
 884                         }
 885
 886                         memcpy(&new_key, &key, sizeof(new_key));
 887                         new_key.offset = start;
 888                         ret = btrfs_duplicate_item(trans, root, path,
 889                                                    &new_key);
 890                         if (ret == -EAGAIN) {
 891                                 btrfs_release_path(path);
 892                                 continue;
 893                         }
 894                         if (ret < 0)
 895                                 break;
 896
 897                         leaf = path->nodes[0];
 898                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 899                                             struct btrfs_file_extent_item);
 900                         btrfs_set_file_extent_num_bytes(leaf, fi,
 901                                                         start - key.offset);
 902
 903                         fi = btrfs_item_ptr(leaf, path->slots[0],
 904                                             struct btrfs_file_extent_item);
 905
 906                         extent_offset += start - key.offset;
 907                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 908                         btrfs_set_file_extent_num_bytes(leaf, fi,
 909                                                         extent_end - start);
 910                         btrfs_mark_buffer_dirty(leaf);
 911
 912                         if (update_refs && disk_bytenr > 0) {
 913                                 btrfs_init_generic_ref(&ref,
 914                                                 BTRFS_ADD_DELAYED_REF,
 915                                                 disk_bytenr, num_bytes, 0);
 916                                 btrfs_init_data_ref(&ref,
 917                                                 root->root_key.objectid,
 918                                                 new_key.objectid,
 919                                                 start - extent_offset);
 920                                 ret = btrfs_inc_extent_ref(trans, &ref);
 921                                 BUG_ON(ret); /* -ENOMEM */
 922                         }
 923                         key.offset = start;
 924                 }
 925                 /*
 926                  * From here on out we will have actually dropped something, so
 927                  * last_end can be updated.
 928                  */
 929                 last_end = extent_end;
 930
 931                 /*
 932                  *  | ---- range to drop ----- |
 933                  *      | -------- extent -------- |
 934                  */
 935                 if (start <= key.offset && end < extent_end) {
 936                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 937                                 ret = -EOPNOTSUPP;
 938                                 break;
 939                         }
 940
 941                         memcpy(&new_key, &key, sizeof(new_key));
 942                         new_key.offset = end;
 943                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 944
 945                         extent_offset += end - key.offset;
 946                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 947                         btrfs_set_file_extent_num_bytes(leaf, fi,
 948                                                         extent_end - end);
 949                         btrfs_mark_buffer_dirty(leaf);
 950                         if (update_refs && disk_bytenr > 0)
 951                                 inode_sub_bytes(inode, end - key.offset);
 952                         break;
 953                 }
 954
 955                 search_start = extent_end;
 956                 /*
 957                  *       | ---- range to drop ----- |
 958                  *  | -------- extent -------- |
 959                  */
 960                 if (start > key.offset && end >= extent_end) {
 961                         BUG_ON(del_nr > 0);
 962                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 963                                 ret = -EOPNOTSUPP;
 964                                 break;
 965                         }
 966
 967                         btrfs_set_file_extent_num_bytes(leaf, fi,
 968                                                         start - key.offset);
 969                         btrfs_mark_buffer_dirty(leaf);
 970                         if (update_refs && disk_bytenr > 0)
 971                                 inode_sub_bytes(inode, extent_end - start);
 972                         if (end == extent_end)
 973                                 break;
 974
 975                         path->slots[0]++;
 976                         goto next_slot;
 977                 }
 978
 979                 /*
 980                  *  | ---- range to drop ----- |
 981                  *    | ------ extent ------ |
 982                  */
 983                 if (start <= key.offset && end >= extent_end) {
 984 delete_extent_item:
 985                         if (del_nr == 0) {
 986                                 del_slot = path->slots[0];
 987                                 del_nr = 1;
 988                         } else {
 989                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 990                                 del_nr++;
 991                         }
 992
 993                         if (update_refs &&
 994                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 995                                 inode_sub_bytes(inode,
 996                                                 extent_end - key.offset);
 997                                 extent_end = ALIGN(extent_end,
 998                                                    fs_info->sectorsize);
 999                         } else if (update_refs && disk_bytenr > 0) {
1000                                 btrfs_init_generic_ref(&ref,
1001                                                 BTRFS_DROP_DELAYED_REF,
1002                                                 disk_bytenr, num_bytes, 0);
1003                                 btrfs_init_data_ref(&ref,
1004                                                 root->root_key.objectid,
1005                                                 key.objectid,
1006                                                 key.offset - extent_offset);
1007                                 ret = btrfs_free_extent(trans, &ref);
1008                                 BUG_ON(ret); /* -ENOMEM */
1009                                 inode_sub_bytes(inode,
1010                                                 extent_end - key.offset);
1011                         }
1012
1013                         if (end == extent_end)
1014                                 break;
1015
1016                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1017                                 path->slots[0]++;
1018                                 goto next_slot;
1019                         }
1020
1021                         ret = btrfs_del_items(trans, root, path, del_slot,
1022                                               del_nr);
1023                         if (ret) {
1024                                 btrfs_abort_transaction(trans, ret);
1025                                 break;
1026                         }
1027
1028                         del_nr = 0;
1029                         del_slot = 0;
1030
1031                         btrfs_release_path(path);
1032                         continue;
1033                 }
1034
1035                 BUG();
1036         }
1037
1038         if (!ret && del_nr > 0) {
1039                 /*
1040                  * Set path->slots[0] to first slot, so that after the delete
1041                  * if items are move off from our leaf to its immediate left or
1042                  * right neighbor leafs, we end up with a correct and adjusted
1043                  * path->slots[0] for our insertion (if replace_extent != 0).
1044                  */
1045                 path->slots[0] = del_slot;
1046                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1047                 if (ret)
1048                         btrfs_abort_transaction(trans, ret);
1049         }
1050
1051         leaf = path->nodes[0];
1052         /*
1053          * If btrfs_del_items() was called, it might have deleted a leaf, in
1054          * which case it unlocked our path, so check path->locks[0] matches a
1055          * write lock.
1056          */
1057         if (!ret && replace_extent && leafs_visited == 1 &&
1058             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1059              path->locks[0] == BTRFS_WRITE_LOCK) &&
1060             btrfs_leaf_free_space(leaf) >=
1061             sizeof(struct btrfs_item) + extent_item_size) {
1062
1063                 key.objectid = ino;
1064                 key.type = BTRFS_EXTENT_DATA_KEY;
1065                 key.offset = start;
1066                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1067                         struct btrfs_key slot_key;
1068
1069                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1070                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1071                                 path->slots[0]++;
1072                 }
1073                 setup_items_for_insert(root, path, &key,
1074                                        &extent_item_size,
1075                                        extent_item_size,
1076                                        sizeof(struct btrfs_item) +
1077                                        extent_item_size, 1);
1078                 *key_inserted = 1;
1079         }
1080
1081         if (!replace_extent || !(*key_inserted))
1082                 btrfs_release_path(path);
1083         if (drop_end)
1084                 *drop_end = found ? min(end, last_end) : end;
1085         return ret;
1086 }
1087
1088 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1089                        struct btrfs_root *root, struct inode *inode, u64 start,
1090                        u64 end, int drop_cache)
1091 {
1092         struct btrfs_path *path;
1093         int ret;
1094
1095         path = btrfs_alloc_path();
1096         if (!path)
1097                 return -ENOMEM;
1098         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1099                                    drop_cache, 0, 0, NULL);
1100         btrfs_free_path(path);
1101         return ret;
1102 }
1103
1104 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1105                             u64 objectid, u64 bytenr, u64 orig_offset,
1106                             u64 *start, u64 *end)
1107 {
1108         struct btrfs_file_extent_item *fi;
1109         struct btrfs_key key;
1110         u64 extent_end;
1111
1112         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1113                 return 0;
1114
1115         btrfs_item_key_to_cpu(leaf, &key, slot);
1116         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1117                 return 0;
1118
1119         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1120         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1121             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1122             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1123             btrfs_file_extent_compression(leaf, fi) ||
1124             btrfs_file_extent_encryption(leaf, fi) ||
1125             btrfs_file_extent_other_encoding(leaf, fi))
1126                 return 0;
1127
1128         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1129         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1130                 return 0;
1131
1132         *start = key.offset;
1133         *end = extent_end;
1134         return 1;
1135 }
1136
1137 /*
1138  * Mark extent in the range start - end as written.
1139  *
1140  * This changes extent type from 'pre-allocated' to 'regular'. If only
1141  * part of extent is marked as written, the extent will be split into
1142  * two or three.
1143  */
1144 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1145                               struct btrfs_inode *inode, u64 start, u64 end)
1146 {
1147         struct btrfs_fs_info *fs_info = trans->fs_info;
1148         struct btrfs_root *root = inode->root;
1149         struct extent_buffer *leaf;
1150         struct btrfs_path *path;
1151         struct btrfs_file_extent_item *fi;
1152         struct btrfs_ref ref = { 0 };
1153         struct btrfs_key key;
1154         struct btrfs_key new_key;
1155         u64 bytenr;
1156         u64 num_bytes;
1157         u64 extent_end;
1158         u64 orig_offset;
1159         u64 other_start;
1160         u64 other_end;
1161         u64 split;
1162         int del_nr = 0;
1163         int del_slot = 0;
1164         int recow;
1165         int ret;
1166         u64 ino = btrfs_ino(inode);
1167
1168         path = btrfs_alloc_path();
1169         if (!path)
1170                 return -ENOMEM;
1171 again:
1172         recow = 0;
1173         split = start;
1174         key.objectid = ino;
1175         key.type = BTRFS_EXTENT_DATA_KEY;
1176         key.offset = split;
1177
1178         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1179         if (ret < 0)
1180                 goto out;
1181         if (ret > 0 && path->slots[0] > 0)
1182                 path->slots[0]--;
1183
1184         leaf = path->nodes[0];
1185         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1186         if (key.objectid != ino ||
1187             key.type != BTRFS_EXTENT_DATA_KEY) {
1188                 ret = -EINVAL;
1189                 btrfs_abort_transaction(trans, ret);
1190                 goto out;
1191         }
1192         fi = btrfs_item_ptr(leaf, path->slots[0],
1193                             struct btrfs_file_extent_item);
1194         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1195                 ret = -EINVAL;
1196                 btrfs_abort_transaction(trans, ret);
1197                 goto out;
1198         }
1199         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1200         if (key.offset > start || extent_end < end) {
1201                 ret = -EINVAL;
1202                 btrfs_abort_transaction(trans, ret);
1203                 goto out;
1204         }
1205
1206         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1207         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1208         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1209         memcpy(&new_key, &key, sizeof(new_key));
1210
1211         if (start == key.offset && end < extent_end) {
1212                 other_start = 0;
1213                 other_end = start;
1214                 if (extent_mergeable(leaf, path->slots[0] - 1,
1215                                      ino, bytenr, orig_offset,
1216                                      &other_start, &other_end)) {
1217                         new_key.offset = end;
1218                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1219                         fi = btrfs_item_ptr(leaf, path->slots[0],
1220                                             struct btrfs_file_extent_item);
1221                         btrfs_set_file_extent_generation(leaf, fi,
1222                                                          trans->transid);
1223                         btrfs_set_file_extent_num_bytes(leaf, fi,
1224                                                         extent_end - end);
1225                         btrfs_set_file_extent_offset(leaf, fi,
1226                                                      end - orig_offset);
1227                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1228                                             struct btrfs_file_extent_item);
1229                         btrfs_set_file_extent_generation(leaf, fi,
1230                                                          trans->transid);
1231                         btrfs_set_file_extent_num_bytes(leaf, fi,
1232                                                         end - other_start);
1233                         btrfs_mark_buffer_dirty(leaf);
1234                         goto out;
1235                 }
1236         }
1237
1238         if (start > key.offset && end == extent_end) {
1239                 other_start = end;
1240                 other_end = 0;
1241                 if (extent_mergeable(leaf, path->slots[0] + 1,
1242                                      ino, bytenr, orig_offset,
1243                                      &other_start, &other_end)) {
1244                         fi = btrfs_item_ptr(leaf, path->slots[0],
1245                                             struct btrfs_file_extent_item);
1246                         btrfs_set_file_extent_num_bytes(leaf, fi,
1247                                                         start - key.offset);
1248                         btrfs_set_file_extent_generation(leaf, fi,
1249                                                          trans->transid);
1250                         path->slots[0]++;
1251                         new_key.offset = start;
1252                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1253
1254                         fi = btrfs_item_ptr(leaf, path->slots[0],
1255                                             struct btrfs_file_extent_item);
1256                         btrfs_set_file_extent_generation(leaf, fi,
1257                                                          trans->transid);
1258                         btrfs_set_file_extent_num_bytes(leaf, fi,
1259                                                         other_end - start);
1260                         btrfs_set_file_extent_offset(leaf, fi,
1261                                                      start - orig_offset);
1262                         btrfs_mark_buffer_dirty(leaf);
1263                         goto out;
1264                 }
1265         }
1266
1267         while (start > key.offset || end < extent_end) {
1268                 if (key.offset == start)
1269                         split = end;
1270
1271                 new_key.offset = split;
1272                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1273                 if (ret == -EAGAIN) {
1274                         btrfs_release_path(path);
1275                         goto again;
1276                 }
1277                 if (ret < 0) {
1278                         btrfs_abort_transaction(trans, ret);
1279                         goto out;
1280                 }
1281
1282                 leaf = path->nodes[0];
1283                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1284                                     struct btrfs_file_extent_item);
1285                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1286                 btrfs_set_file_extent_num_bytes(leaf, fi,
1287                                                 split - key.offset);
1288
1289                 fi = btrfs_item_ptr(leaf, path->slots[0],
1290                                     struct btrfs_file_extent_item);
1291
1292                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1293                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1294                 btrfs_set_file_extent_num_bytes(leaf, fi,
1295                                                 extent_end - split);
1296                 btrfs_mark_buffer_dirty(leaf);
1297
1298                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1299                                        num_bytes, 0);
1300                 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1301                                     orig_offset);
1302                 ret = btrfs_inc_extent_ref(trans, &ref);
1303                 if (ret) {
1304                         btrfs_abort_transaction(trans, ret);
1305                         goto out;
1306                 }
1307
1308                 if (split == start) {
1309                         key.offset = start;
1310                 } else {
1311                         if (start != key.offset) {
1312                                 ret = -EINVAL;
1313                                 btrfs_abort_transaction(trans, ret);
1314                                 goto out;
1315                         }
1316                         path->slots[0]--;
1317                         extent_end = end;
1318                 }
1319                 recow = 1;
1320         }
1321
1322         other_start = end;
1323         other_end = 0;
1324         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1325                                num_bytes, 0);
1326         btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1327         if (extent_mergeable(leaf, path->slots[0] + 1,
1328                              ino, bytenr, orig_offset,
1329                              &other_start, &other_end)) {
1330                 if (recow) {
1331                         btrfs_release_path(path);
1332                         goto again;
1333                 }
1334                 extent_end = other_end;
1335                 del_slot = path->slots[0] + 1;
1336                 del_nr++;
1337                 ret = btrfs_free_extent(trans, &ref);
1338                 if (ret) {
1339                         btrfs_abort_transaction(trans, ret);
1340                         goto out;
1341                 }
1342         }
1343         other_start = 0;
1344         other_end = start;
1345         if (extent_mergeable(leaf, path->slots[0] - 1,
1346                              ino, bytenr, orig_offset,
1347                              &other_start, &other_end)) {
1348                 if (recow) {
1349                         btrfs_release_path(path);
1350                         goto again;
1351                 }
1352                 key.offset = other_start;
1353                 del_slot = path->slots[0];
1354                 del_nr++;
1355                 ret = btrfs_free_extent(trans, &ref);
1356                 if (ret) {
1357                         btrfs_abort_transaction(trans, ret);
1358                         goto out;
1359                 }
1360         }
1361         if (del_nr == 0) {
1362                 fi = btrfs_item_ptr(leaf, path->slots[0],
1363                            struct btrfs_file_extent_item);
1364                 btrfs_set_file_extent_type(leaf, fi,
1365                                            BTRFS_FILE_EXTENT_REG);
1366                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1367                 btrfs_mark_buffer_dirty(leaf);
1368         } else {
1369                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1370                            struct btrfs_file_extent_item);
1371                 btrfs_set_file_extent_type(leaf, fi,
1372                                            BTRFS_FILE_EXTENT_REG);
1373                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1374                 btrfs_set_file_extent_num_bytes(leaf, fi,
1375                                                 extent_end - key.offset);
1376                 btrfs_mark_buffer_dirty(leaf);
1377
1378                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1379                 if (ret < 0) {
1380                         btrfs_abort_transaction(trans, ret);
1381                         goto out;
1382                 }
1383         }
1384 out:
1385         btrfs_free_path(path);
1386         return 0;
1387 }
1388
1389 /*
1390  * on error we return an unlocked page and the error value
1391  * on success we return a locked page and 0
1392  */
1393 static int prepare_uptodate_page(struct inode *inode,
1394                                  struct page *page, u64 pos,
1395                                  bool force_uptodate)
1396 {
1397         int ret = 0;
1398
1399         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1400             !PageUptodate(page)) {
1401                 ret = btrfs_readpage(NULL, page);
1402                 if (ret)
1403                         return ret;
1404                 lock_page(page);
1405                 if (!PageUptodate(page)) {
1406                         unlock_page(page);
1407                         return -EIO;
1408                 }
1409                 if (page->mapping != inode->i_mapping) {
1410                         unlock_page(page);
1411                         return -EAGAIN;
1412                 }
1413         }
1414         return 0;
1415 }
1416
1417 /*
1418  * this just gets pages into the page cache and locks them down.
1419  */
1420 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1421                                   size_t num_pages, loff_t pos,
1422                                   size_t write_bytes, bool force_uptodate)
1423 {
1424         int i;
1425         unsigned long index = pos >> PAGE_SHIFT;
1426         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1427         int err = 0;
1428         int faili;
1429
1430         for (i = 0; i < num_pages; i++) {
1431 again:
1432                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1433                                                mask | __GFP_WRITE);
1434                 if (!pages[i]) {
1435                         faili = i - 1;
1436                         err = -ENOMEM;
1437                         goto fail;
1438                 }
1439
1440                 if (i == 0)
1441                         err = prepare_uptodate_page(inode, pages[i], pos,
1442                                                     force_uptodate);
1443                 if (!err && i == num_pages - 1)
1444                         err = prepare_uptodate_page(inode, pages[i],
1445                                                     pos + write_bytes, false);
1446                 if (err) {
1447                         put_page(pages[i]);
1448                         if (err == -EAGAIN) {
1449                                 err = 0;
1450                                 goto again;
1451                         }
1452                         faili = i - 1;
1453                         goto fail;
1454                 }
1455                 wait_on_page_writeback(pages[i]);
1456         }
1457
1458         return 0;
1459 fail:
1460         while (faili >= 0) {
1461                 unlock_page(pages[faili]);
1462                 put_page(pages[faili]);
1463                 faili--;
1464         }
1465         return err;
1466
1467 }
1468
1469 /*
1470  * This function locks the extent and properly waits for data=ordered extents
1471  * to finish before allowing the pages to be modified if need.
1472  *
1473  * The return value:
1474  * 1 - the extent is locked
1475  * 0 - the extent is not locked, and everything is OK
1476  * -EAGAIN - need re-prepare the pages
1477  * the other < 0 number - Something wrong happens
1478  */
1479 static noinline int
1480 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1481                                 size_t num_pages, loff_t pos,
1482                                 size_t write_bytes,
1483                                 u64 *lockstart, u64 *lockend,
1484                                 struct extent_state **cached_state)
1485 {
1486         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1487         u64 start_pos;
1488         u64 last_pos;
1489         int i;
1490         int ret = 0;
1491
1492         start_pos = round_down(pos, fs_info->sectorsize);
1493         last_pos = start_pos
1494                 + round_up(pos + write_bytes - start_pos,
1495                            fs_info->sectorsize) - 1;
1496
1497         if (start_pos < inode->vfs_inode.i_size) {
1498                 struct btrfs_ordered_extent *ordered;
1499
1500                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1501                                 cached_state);
1502                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1503                                                      last_pos - start_pos + 1);
1504                 if (ordered &&
1505                     ordered->file_offset + ordered->len > start_pos &&
1506                     ordered->file_offset <= last_pos) {
1507                         unlock_extent_cached(&inode->io_tree, start_pos,
1508                                         last_pos, cached_state);
1509                         for (i = 0; i < num_pages; i++) {
1510                                 unlock_page(pages[i]);
1511                                 put_page(pages[i]);
1512                         }
1513                         btrfs_start_ordered_extent(&inode->vfs_inode,
1514                                         ordered, 1);
1515                         btrfs_put_ordered_extent(ordered);
1516                         return -EAGAIN;
1517                 }
1518                 if (ordered)
1519                         btrfs_put_ordered_extent(ordered);
1520
1521                 *lockstart = start_pos;
1522                 *lockend = last_pos;
1523                 ret = 1;
1524         }
1525
1526         /*
1527          * It's possible the pages are dirty right now, but we don't want
1528          * to clean them yet because copy_from_user may catch a page fault
1529          * and we might have to fall back to one page at a time.  If that
1530          * happens, we'll unlock these pages and we'd have a window where
1531          * reclaim could sneak in and drop the once-dirty page on the floor
1532          * without writing it.
1533          *
1534          * We have the pages locked and the extent range locked, so there's
1535          * no way someone can start IO on any dirty pages in this range.
1536          *
1537          * We'll call btrfs_dirty_pages() later on, and that will flip around
1538          * delalloc bits and dirty the pages as required.
1539          */
1540         for (i = 0; i < num_pages; i++) {
1541                 set_page_extent_mapped(pages[i]);
1542                 WARN_ON(!PageLocked(pages[i]));
1543         }
1544
1545         return ret;
1546 }
1547
1548 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1549                                     size_t *write_bytes)
1550 {
1551         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1552         struct btrfs_root *root = inode->root;
1553         struct btrfs_ordered_extent *ordered;
1554         u64 lockstart, lockend;
1555         u64 num_bytes;
1556         int ret;
1557
1558         ret = btrfs_start_write_no_snapshotting(root);
1559         if (!ret)
1560                 return -ENOSPC;
1561
1562         lockstart = round_down(pos, fs_info->sectorsize);
1563         lockend = round_up(pos + *write_bytes,
1564                            fs_info->sectorsize) - 1;
1565
1566         while (1) {
1567                 lock_extent(&inode->io_tree, lockstart, lockend);
1568                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1569                                                      lockend - lockstart + 1);
1570                 if (!ordered) {
1571                         break;
1572                 }
1573                 unlock_extent(&inode->io_tree, lockstart, lockend);
1574                 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1575                 btrfs_put_ordered_extent(ordered);
1576         }
1577
1578         num_bytes = lockend - lockstart + 1;
1579         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1580                         NULL, NULL, NULL);
1581         if (ret <= 0) {
1582                 ret = 0;
1583                 btrfs_end_write_no_snapshotting(root);
1584         } else {
1585                 *write_bytes = min_t(size_t, *write_bytes ,
1586                                      num_bytes - pos + lockstart);
1587         }
1588
1589         unlock_extent(&inode->io_tree, lockstart, lockend);
1590
1591         return ret;
1592 }
1593
1594 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1595                                                struct iov_iter *i)
1596 {
1597         struct file *file = iocb->ki_filp;
1598         loff_t pos = iocb->ki_pos;
1599         struct inode *inode = file_inode(file);
1600         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1601         struct btrfs_root *root = BTRFS_I(inode)->root;
1602         struct page **pages = NULL;
1603         struct extent_state *cached_state = NULL;
1604         struct extent_changeset *data_reserved = NULL;
1605         u64 release_bytes = 0;
1606         u64 lockstart;
1607         u64 lockend;
1608         size_t num_written = 0;
1609         int nrptrs;
1610         int ret = 0;
1611         bool only_release_metadata = false;
1612         bool force_page_uptodate = false;
1613
1614         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1615                         PAGE_SIZE / (sizeof(struct page *)));
1616         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1617         nrptrs = max(nrptrs, 8);
1618         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1619         if (!pages)
1620                 return -ENOMEM;
1621
1622         while (iov_iter_count(i) > 0) {
1623                 size_t offset = offset_in_page(pos);
1624                 size_t sector_offset;
1625                 size_t write_bytes = min(iov_iter_count(i),
1626                                          nrptrs * (size_t)PAGE_SIZE -
1627                                          offset);
1628                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1629                                                 PAGE_SIZE);
1630                 size_t reserve_bytes;
1631                 size_t dirty_pages;
1632                 size_t copied;
1633                 size_t dirty_sectors;
1634                 size_t num_sectors;
1635                 int extents_locked;
1636
1637                 WARN_ON(num_pages > nrptrs);
1638
1639                 /*
1640                  * Fault pages before locking them in prepare_pages
1641                  * to avoid recursive lock
1642                  */
1643                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1644                         ret = -EFAULT;
1645                         break;
1646                 }
1647
1648                 sector_offset = pos & (fs_info->sectorsize - 1);
1649                 reserve_bytes = round_up(write_bytes + sector_offset,
1650                                 fs_info->sectorsize);
1651
1652                 extent_changeset_release(data_reserved);
1653                 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1654                                                   write_bytes);
1655                 if (ret < 0) {
1656                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1657                                                       BTRFS_INODE_PREALLOC)) &&
1658                             check_can_nocow(BTRFS_I(inode), pos,
1659                                         &write_bytes) > 0) {
1660                                 /*
1661                                  * For nodata cow case, no need to reserve
1662                                  * data space.
1663                                  */
1664                                 only_release_metadata = true;
1665                                 /*
1666                                  * our prealloc extent may be smaller than
1667                                  * write_bytes, so scale down.
1668                                  */
1669                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1670                                                          PAGE_SIZE);
1671                                 reserve_bytes = round_up(write_bytes +
1672                                                          sector_offset,
1673                                                          fs_info->sectorsize);
1674                         } else {
1675                                 break;
1676                         }
1677                 }
1678
1679                 WARN_ON(reserve_bytes == 0);
1680                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1681                                 reserve_bytes);
1682                 if (ret) {
1683                         if (!only_release_metadata)
1684                                 btrfs_free_reserved_data_space(inode,
1685                                                 data_reserved, pos,
1686                                                 write_bytes);
1687                         else
1688                                 btrfs_end_write_no_snapshotting(root);
1689                         break;
1690                 }
1691
1692                 release_bytes = reserve_bytes;
1693 again:
1694                 /*
1695                  * This is going to setup the pages array with the number of
1696                  * pages we want, so we don't really need to worry about the
1697                  * contents of pages from loop to loop
1698                  */
1699                 ret = prepare_pages(inode, pages, num_pages,
1700                                     pos, write_bytes,
1701                                     force_page_uptodate);
1702                 if (ret) {
1703                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1704                                                        reserve_bytes, true);
1705                         break;
1706                 }
1707
1708                 extents_locked = lock_and_cleanup_extent_if_need(
1709                                 BTRFS_I(inode), pages,
1710                                 num_pages, pos, write_bytes, &lockstart,
1711                                 &lockend, &cached_state);
1712                 if (extents_locked < 0) {
1713                         if (extents_locked == -EAGAIN)
1714                                 goto again;
1715                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1716                                                        reserve_bytes, true);
1717                         ret = extents_locked;
1718                         break;
1719                 }
1720
1721                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1722
1723                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1724                 dirty_sectors = round_up(copied + sector_offset,
1725                                         fs_info->sectorsize);
1726                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1727
1728                 /*
1729                  * if we have trouble faulting in the pages, fall
1730                  * back to one page at a time
1731                  */
1732                 if (copied < write_bytes)
1733                         nrptrs = 1;
1734
1735                 if (copied == 0) {
1736                         force_page_uptodate = true;
1737                         dirty_sectors = 0;
1738                         dirty_pages = 0;
1739                 } else {
1740                         force_page_uptodate = false;
1741                         dirty_pages = DIV_ROUND_UP(copied + offset,
1742                                                    PAGE_SIZE);
1743                 }
1744
1745                 if (num_sectors > dirty_sectors) {
1746                         /* release everything except the sectors we dirtied */
1747                         release_bytes -= dirty_sectors <<
1748                                                 fs_info->sb->s_blocksize_bits;
1749                         if (only_release_metadata) {
1750                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1751                                                         release_bytes, true);
1752                         } else {
1753                                 u64 __pos;
1754
1755                                 __pos = round_down(pos,
1756                                                    fs_info->sectorsize) +
1757                                         (dirty_pages << PAGE_SHIFT);
1758                                 btrfs_delalloc_release_space(inode,
1759                                                 data_reserved, __pos,
1760                                                 release_bytes, true);
1761                         }
1762                 }
1763
1764                 release_bytes = round_up(copied + sector_offset,
1765                                         fs_info->sectorsize);
1766
1767                 if (copied > 0)
1768                         ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1769                                                 pos, copied, &cached_state);
1770                 if (extents_locked)
1771                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1772                                              lockstart, lockend, &cached_state);
1773                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes,
1774                                                true);
1775                 if (ret) {
1776                         btrfs_drop_pages(pages, num_pages);
1777                         break;
1778                 }
1779
1780                 release_bytes = 0;
1781                 if (only_release_metadata)
1782                         btrfs_end_write_no_snapshotting(root);
1783
1784                 if (only_release_metadata && copied > 0) {
1785                         lockstart = round_down(pos,
1786                                                fs_info->sectorsize);
1787                         lockend = round_up(pos + copied,
1788                                            fs_info->sectorsize) - 1;
1789
1790                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1791                                        lockend, EXTENT_NORESERVE, NULL,
1792                                        NULL, GFP_NOFS);
1793                         only_release_metadata = false;
1794                 }
1795
1796                 btrfs_drop_pages(pages, num_pages);
1797
1798                 cond_resched();
1799
1800                 balance_dirty_pages_ratelimited(inode->i_mapping);
1801                 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1802                         btrfs_btree_balance_dirty(fs_info);
1803
1804                 pos += copied;
1805                 num_written += copied;
1806         }
1807
1808         kfree(pages);
1809
1810         if (release_bytes) {
1811                 if (only_release_metadata) {
1812                         btrfs_end_write_no_snapshotting(root);
1813                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1814                                         release_bytes, true);
1815                 } else {
1816                         btrfs_delalloc_release_space(inode, data_reserved,
1817                                         round_down(pos, fs_info->sectorsize),
1818                                         release_bytes, true);
1819                 }
1820         }
1821
1822         extent_changeset_free(data_reserved);
1823         return num_written ? num_written : ret;
1824 }
1825
1826 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1827 {
1828         struct file *file = iocb->ki_filp;
1829         struct inode *inode = file_inode(file);
1830         loff_t pos;
1831         ssize_t written;
1832         ssize_t written_buffered;
1833         loff_t endbyte;
1834         int err;
1835
1836         written = generic_file_direct_write(iocb, from);
1837
1838         if (written < 0 || !iov_iter_count(from))
1839                 return written;
1840
1841         pos = iocb->ki_pos;
1842         written_buffered = btrfs_buffered_write(iocb, from);
1843         if (written_buffered < 0) {
1844                 err = written_buffered;
1845                 goto out;
1846         }
1847         /*
1848          * Ensure all data is persisted. We want the next direct IO read to be
1849          * able to read what was just written.
1850          */
1851         endbyte = pos + written_buffered - 1;
1852         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1853         if (err)
1854                 goto out;
1855         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1856         if (err)
1857                 goto out;
1858         written += written_buffered;
1859         iocb->ki_pos = pos + written_buffered;
1860         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1861                                  endbyte >> PAGE_SHIFT);
1862 out:
1863         return written ? written : err;
1864 }
1865
1866 static void update_time_for_write(struct inode *inode)
1867 {
1868         struct timespec64 now;
1869
1870         if (IS_NOCMTIME(inode))
1871                 return;
1872
1873         now = current_time(inode);
1874         if (!timespec64_equal(&inode->i_mtime, &now))
1875                 inode->i_mtime = now;
1876
1877         if (!timespec64_equal(&inode->i_ctime, &now))
1878                 inode->i_ctime = now;
1879
1880         if (IS_I_VERSION(inode))
1881                 inode_inc_iversion(inode);
1882 }
1883
1884 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1885                                     struct iov_iter *from)
1886 {
1887         struct file *file = iocb->ki_filp;
1888         struct inode *inode = file_inode(file);
1889         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1890         struct btrfs_root *root = BTRFS_I(inode)->root;
1891         u64 start_pos;
1892         u64 end_pos;
1893         ssize_t num_written = 0;
1894         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1895         ssize_t err;
1896         loff_t pos;
1897         size_t count = iov_iter_count(from);
1898         loff_t oldsize;
1899         int clean_page = 0;
1900
1901         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1902             (iocb->ki_flags & IOCB_NOWAIT))
1903                 return -EOPNOTSUPP;
1904
1905         if (!inode_trylock(inode)) {
1906                 if (iocb->ki_flags & IOCB_NOWAIT)
1907                         return -EAGAIN;
1908                 inode_lock(inode);
1909         }
1910
1911         err = generic_write_checks(iocb, from);
1912         if (err <= 0) {
1913                 inode_unlock(inode);
1914                 return err;
1915         }
1916
1917         pos = iocb->ki_pos;
1918         if (iocb->ki_flags & IOCB_NOWAIT) {
1919                 /*
1920                  * We will allocate space in case nodatacow is not set,
1921                  * so bail
1922                  */
1923                 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1924                                               BTRFS_INODE_PREALLOC)) ||
1925                     check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1926                         inode_unlock(inode);
1927                         return -EAGAIN;
1928                 }
1929         }
1930
1931         current->backing_dev_info = inode_to_bdi(inode);
1932         err = file_remove_privs(file);
1933         if (err) {
1934                 inode_unlock(inode);
1935                 goto out;
1936         }
1937
1938         /*
1939          * If BTRFS flips readonly due to some impossible error
1940          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1941          * although we have opened a file as writable, we have
1942          * to stop this write operation to ensure FS consistency.
1943          */
1944         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1945                 inode_unlock(inode);
1946                 err = -EROFS;
1947                 goto out;
1948         }
1949
1950         /*
1951          * We reserve space for updating the inode when we reserve space for the
1952          * extent we are going to write, so we will enospc out there.  We don't
1953          * need to start yet another transaction to update the inode as we will
1954          * update the inode when we finish writing whatever data we write.
1955          */
1956         update_time_for_write(inode);
1957
1958         start_pos = round_down(pos, fs_info->sectorsize);
1959         oldsize = i_size_read(inode);
1960         if (start_pos > oldsize) {
1961                 /* Expand hole size to cover write data, preventing empty gap */
1962                 end_pos = round_up(pos + count,
1963                                    fs_info->sectorsize);
1964                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1965                 if (err) {
1966                         inode_unlock(inode);
1967                         goto out;
1968                 }
1969                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1970                         clean_page = 1;
1971         }
1972
1973         if (sync)
1974                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1975
1976         if (iocb->ki_flags & IOCB_DIRECT) {
1977                 num_written = __btrfs_direct_write(iocb, from);
1978         } else {
1979                 num_written = btrfs_buffered_write(iocb, from);
1980                 if (num_written > 0)
1981                         iocb->ki_pos = pos + num_written;
1982                 if (clean_page)
1983                         pagecache_isize_extended(inode, oldsize,
1984                                                 i_size_read(inode));
1985         }
1986
1987         inode_unlock(inode);
1988
1989         /*
1990          * We also have to set last_sub_trans to the current log transid,
1991          * otherwise subsequent syncs to a file that's been synced in this
1992          * transaction will appear to have already occurred.
1993          */
1994         spin_lock(&BTRFS_I(inode)->lock);
1995         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1996         spin_unlock(&BTRFS_I(inode)->lock);
1997         if (num_written > 0)
1998                 num_written = generic_write_sync(iocb, num_written);
1999
2000         if (sync)
2001                 atomic_dec(&BTRFS_I(inode)->sync_writers);
2002 out:
2003         current->backing_dev_info = NULL;
2004         return num_written ? num_written : err;
2005 }
2006
2007 int btrfs_release_file(struct inode *inode, struct file *filp)
2008 {
2009         struct btrfs_file_private *private = filp->private_data;
2010
2011         if (private && private->filldir_buf)
2012                 kfree(private->filldir_buf);
2013         kfree(private);
2014         filp->private_data = NULL;
2015
2016         /*
2017          * ordered_data_close is set by setattr when we are about to truncate
2018          * a file from a non-zero size to a zero size.  This tries to
2019          * flush down new bytes that may have been written if the
2020          * application were using truncate to replace a file in place.
2021          */
2022         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2023                                &BTRFS_I(inode)->runtime_flags))
2024                         filemap_flush(inode->i_mapping);
2025         return 0;
2026 }
2027
2028 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2029 {
2030         int ret;
2031         struct blk_plug plug;
2032
2033         /*
2034          * This is only called in fsync, which would do synchronous writes, so
2035          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2036          * multiple disks using raid profile, a large IO can be split to
2037          * several segments of stripe length (currently 64K).
2038          */
2039         blk_start_plug(&plug);
2040         atomic_inc(&BTRFS_I(inode)->sync_writers);
2041         ret = btrfs_fdatawrite_range(inode, start, end);
2042         atomic_dec(&BTRFS_I(inode)->sync_writers);
2043         blk_finish_plug(&plug);
2044
2045         return ret;
2046 }
2047
2048 /*
2049  * fsync call for both files and directories.  This logs the inode into
2050  * the tree log instead of forcing full commits whenever possible.
2051  *
2052  * It needs to call filemap_fdatawait so that all ordered extent updates are
2053  * in the metadata btree are up to date for copying to the log.
2054  *
2055  * It drops the inode mutex before doing the tree log commit.  This is an
2056  * important optimization for directories because holding the mutex prevents
2057  * new operations on the dir while we write to disk.
2058  */
2059 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2060 {
2061         struct dentry *dentry = file_dentry(file);
2062         struct inode *inode = d_inode(dentry);
2063         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2064         struct btrfs_root *root = BTRFS_I(inode)->root;
2065         struct btrfs_trans_handle *trans;
2066         struct btrfs_log_ctx ctx;
2067         int ret = 0, err;
2068         u64 len;
2069
2070         /*
2071          * If the inode needs a full sync, make sure we use a full range to
2072          * avoid log tree corruption, due to hole detection racing with ordered
2073          * extent completion for adjacent ranges, and assertion failures during
2074          * hole detection.
2075          */
2076         if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2077                      &BTRFS_I(inode)->runtime_flags)) {
2078                 start = 0;
2079                 end = LLONG_MAX;
2080         }
2081
2082         /*
2083          * The range length can be represented by u64, we have to do the typecasts
2084          * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2085          */
2086         len = (u64)end - (u64)start + 1;
2087         trace_btrfs_sync_file(file, datasync);
2088
2089         btrfs_init_log_ctx(&ctx, inode);
2090
2091         /*
2092          * We write the dirty pages in the range and wait until they complete
2093          * out of the ->i_mutex. If so, we can flush the dirty pages by
2094          * multi-task, and make the performance up.  See
2095          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2096          */
2097         ret = start_ordered_ops(inode, start, end);
2098         if (ret)
2099                 goto out;
2100
2101         inode_lock(inode);
2102
2103         /*
2104          * We take the dio_sem here because the tree log stuff can race with
2105          * lockless dio writes and get an extent map logged for an extent we
2106          * never waited on.  We need it this high up for lockdep reasons.
2107          */
2108         down_write(&BTRFS_I(inode)->dio_sem);
2109
2110         atomic_inc(&root->log_batch);
2111
2112         /*
2113          * Before we acquired the inode's lock, someone may have dirtied more
2114          * pages in the target range. We need to make sure that writeback for
2115          * any such pages does not start while we are logging the inode, because
2116          * if it does, any of the following might happen when we are not doing a
2117          * full inode sync:
2118          *
2119          * 1) We log an extent after its writeback finishes but before its
2120          *    checksums are added to the csum tree, leading to -EIO errors
2121          *    when attempting to read the extent after a log replay.
2122          *
2123          * 2) We can end up logging an extent before its writeback finishes.
2124          *    Therefore after the log replay we will have a file extent item
2125          *    pointing to an unwritten extent (and no data checksums as well).
2126          *
2127          * So trigger writeback for any eventual new dirty pages and then we
2128          * wait for all ordered extents to complete below.
2129          */
2130         ret = start_ordered_ops(inode, start, end);
2131         if (ret) {
2132                 inode_unlock(inode);
2133                 goto out;
2134         }
2135
2136         /*
2137          * We have to do this here to avoid the priority inversion of waiting on
2138          * IO of a lower priority task while holding a transaction open.
2139          */
2140         ret = btrfs_wait_ordered_range(inode, start, len);
2141         if (ret) {
2142                 up_write(&BTRFS_I(inode)->dio_sem);
2143                 inode_unlock(inode);
2144                 goto out;
2145         }
2146         atomic_inc(&root->log_batch);
2147
2148         smp_mb();
2149         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2150             BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2151                 /*
2152                  * We've had everything committed since the last time we were
2153                  * modified so clear this flag in case it was set for whatever
2154                  * reason, it's no longer relevant.
2155                  */
2156                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2157                           &BTRFS_I(inode)->runtime_flags);
2158                 /*
2159                  * An ordered extent might have started before and completed
2160                  * already with io errors, in which case the inode was not
2161                  * updated and we end up here. So check the inode's mapping
2162                  * for any errors that might have happened since we last
2163                  * checked called fsync.
2164                  */
2165                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2166                 up_write(&BTRFS_I(inode)->dio_sem);
2167                 inode_unlock(inode);
2168                 goto out;
2169         }
2170
2171         /*
2172          * We use start here because we will need to wait on the IO to complete
2173          * in btrfs_sync_log, which could require joining a transaction (for
2174          * example checking cross references in the nocow path).  If we use join
2175          * here we could get into a situation where we're waiting on IO to
2176          * happen that is blocked on a transaction trying to commit.  With start
2177          * we inc the extwriter counter, so we wait for all extwriters to exit
2178          * before we start blocking joiners.  This comment is to keep somebody
2179          * from thinking they are super smart and changing this to
2180          * btrfs_join_transaction *cough*Josef*cough*.
2181          */
2182         trans = btrfs_start_transaction(root, 0);
2183         if (IS_ERR(trans)) {
2184                 ret = PTR_ERR(trans);
2185                 up_write(&BTRFS_I(inode)->dio_sem);
2186                 inode_unlock(inode);
2187                 goto out;
2188         }
2189
2190         ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2191         if (ret < 0) {
2192                 /* Fallthrough and commit/free transaction. */
2193                 ret = 1;
2194         }
2195
2196         /* we've logged all the items and now have a consistent
2197          * version of the file in the log.  It is possible that
2198          * someone will come in and modify the file, but that's
2199          * fine because the log is consistent on disk, and we
2200          * have references to all of the file's extents
2201          *
2202          * It is possible that someone will come in and log the
2203          * file again, but that will end up using the synchronization
2204          * inside btrfs_sync_log to keep things safe.
2205          */
2206         up_write(&BTRFS_I(inode)->dio_sem);
2207         inode_unlock(inode);
2208
2209         if (ret != BTRFS_NO_LOG_SYNC) {
2210                 if (!ret) {
2211                         ret = btrfs_sync_log(trans, root, &ctx);
2212                         if (!ret) {
2213                                 ret = btrfs_end_transaction(trans);
2214                                 goto out;
2215                         }
2216                 }
2217                 ret = btrfs_commit_transaction(trans);
2218         } else {
2219                 ret = btrfs_end_transaction(trans);
2220         }
2221 out:
2222         ASSERT(list_empty(&ctx.list));
2223         err = file_check_and_advance_wb_err(file);
2224         if (!ret)
2225                 ret = err;
2226         return ret > 0 ? -EIO : ret;
2227 }
2228
2229 static const struct vm_operations_struct btrfs_file_vm_ops = {
2230         .fault          = filemap_fault,
2231         .map_pages      = filemap_map_pages,
2232         .page_mkwrite   = btrfs_page_mkwrite,
2233 };
2234
2235 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2236 {
2237         struct address_space *mapping = filp->f_mapping;
2238
2239         if (!mapping->a_ops->readpage)
2240                 return -ENOEXEC;
2241
2242         file_accessed(filp);
2243         vma->vm_ops = &btrfs_file_vm_ops;
2244
2245         return 0;
2246 }
2247
2248 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2249                           int slot, u64 start, u64 end)
2250 {
2251         struct btrfs_file_extent_item *fi;
2252         struct btrfs_key key;
2253
2254         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2255                 return 0;
2256
2257         btrfs_item_key_to_cpu(leaf, &key, slot);
2258         if (key.objectid != btrfs_ino(inode) ||
2259             key.type != BTRFS_EXTENT_DATA_KEY)
2260                 return 0;
2261
2262         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2263
2264         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2265                 return 0;
2266
2267         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2268                 return 0;
2269
2270         if (key.offset == end)
2271                 return 1;
2272         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2273                 return 1;
2274         return 0;
2275 }
2276
2277 static int fill_holes(struct btrfs_trans_handle *trans,
2278                 struct btrfs_inode *inode,
2279                 struct btrfs_path *path, u64 offset, u64 end)
2280 {
2281         struct btrfs_fs_info *fs_info = trans->fs_info;
2282         struct btrfs_root *root = inode->root;
2283         struct extent_buffer *leaf;
2284         struct btrfs_file_extent_item *fi;
2285         struct extent_map *hole_em;
2286         struct extent_map_tree *em_tree = &inode->extent_tree;
2287         struct btrfs_key key;
2288         int ret;
2289
2290         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2291                 goto out;
2292
2293         key.objectid = btrfs_ino(inode);
2294         key.type = BTRFS_EXTENT_DATA_KEY;
2295         key.offset = offset;
2296
2297         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2298         if (ret <= 0) {
2299                 /*
2300                  * We should have dropped this offset, so if we find it then
2301                  * something has gone horribly wrong.
2302                  */
2303                 if (ret == 0)
2304                         ret = -EINVAL;
2305                 return ret;
2306         }
2307
2308         leaf = path->nodes[0];
2309         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2310                 u64 num_bytes;
2311
2312                 path->slots[0]--;
2313                 fi = btrfs_item_ptr(leaf, path->slots[0],
2314                                     struct btrfs_file_extent_item);
2315                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2316                         end - offset;
2317                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2318                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2319                 btrfs_set_file_extent_offset(leaf, fi, 0);
2320                 btrfs_mark_buffer_dirty(leaf);
2321                 goto out;
2322         }
2323
2324         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2325                 u64 num_bytes;
2326
2327                 key.offset = offset;
2328                 btrfs_set_item_key_safe(fs_info, path, &key);
2329                 fi = btrfs_item_ptr(leaf, path->slots[0],
2330                                     struct btrfs_file_extent_item);
2331                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2332                         offset;
2333                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2334                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2335                 btrfs_set_file_extent_offset(leaf, fi, 0);
2336                 btrfs_mark_buffer_dirty(leaf);
2337                 goto out;
2338         }
2339         btrfs_release_path(path);
2340
2341         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2342                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2343         if (ret)
2344                 return ret;
2345
2346 out:
2347         btrfs_release_path(path);
2348
2349         hole_em = alloc_extent_map();
2350         if (!hole_em) {
2351                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2352                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2353         } else {
2354                 hole_em->start = offset;
2355                 hole_em->len = end - offset;
2356                 hole_em->ram_bytes = hole_em->len;
2357                 hole_em->orig_start = offset;
2358
2359                 hole_em->block_start = EXTENT_MAP_HOLE;
2360                 hole_em->block_len = 0;
2361                 hole_em->orig_block_len = 0;
2362                 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2363                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2364                 hole_em->generation = trans->transid;
2365
2366                 do {
2367                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2368                         write_lock(&em_tree->lock);
2369                         ret = add_extent_mapping(em_tree, hole_em, 1);
2370                         write_unlock(&em_tree->lock);
2371                 } while (ret == -EEXIST);
2372                 free_extent_map(hole_em);
2373                 if (ret)
2374                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2375                                         &inode->runtime_flags);
2376         }
2377
2378         return 0;
2379 }
2380
2381 /*
2382  * Find a hole extent on given inode and change start/len to the end of hole
2383  * extent.(hole/vacuum extent whose em->start <= start &&
2384  *         em->start + em->len > start)
2385  * When a hole extent is found, return 1 and modify start/len.
2386  */
2387 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2388 {
2389         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2390         struct extent_map *em;
2391         int ret = 0;
2392
2393         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2394                               round_down(*start, fs_info->sectorsize),
2395                               round_up(*len, fs_info->sectorsize), 0);
2396         if (IS_ERR(em))
2397                 return PTR_ERR(em);
2398
2399         /* Hole or vacuum extent(only exists in no-hole mode) */
2400         if (em->block_start == EXTENT_MAP_HOLE) {
2401                 ret = 1;
2402                 *len = em->start + em->len > *start + *len ?
2403                        0 : *start + *len - em->start - em->len;
2404                 *start = em->start + em->len;
2405         }
2406         free_extent_map(em);
2407         return ret;
2408 }
2409
2410 static int btrfs_punch_hole_lock_range(struct inode *inode,
2411                                        const u64 lockstart,
2412                                        const u64 lockend,
2413                                        struct extent_state **cached_state)
2414 {
2415         while (1) {
2416                 struct btrfs_ordered_extent *ordered;
2417                 int ret;
2418
2419                 truncate_pagecache_range(inode, lockstart, lockend);
2420
2421                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2422                                  cached_state);
2423                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2424
2425                 /*
2426                  * We need to make sure we have no ordered extents in this range
2427                  * and nobody raced in and read a page in this range, if we did
2428                  * we need to try again.
2429                  */
2430                 if ((!ordered ||
2431                     (ordered->file_offset + ordered->len <= lockstart ||
2432                      ordered->file_offset > lockend)) &&
2433                      !filemap_range_has_page(inode->i_mapping,
2434                                              lockstart, lockend)) {
2435                         if (ordered)
2436                                 btrfs_put_ordered_extent(ordered);
2437                         break;
2438                 }
2439                 if (ordered)
2440                         btrfs_put_ordered_extent(ordered);
2441                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2442                                      lockend, cached_state);
2443                 ret = btrfs_wait_ordered_range(inode, lockstart,
2444                                                lockend - lockstart + 1);
2445                 if (ret)
2446                         return ret;
2447         }
2448         return 0;
2449 }
2450
2451 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2452 {
2453         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2454         struct btrfs_root *root = BTRFS_I(inode)->root;
2455         struct extent_state *cached_state = NULL;
2456         struct btrfs_path *path;
2457         struct btrfs_block_rsv *rsv;
2458         struct btrfs_trans_handle *trans;
2459         u64 lockstart;
2460         u64 lockend;
2461         u64 tail_start;
2462         u64 tail_len;
2463         u64 orig_start = offset;
2464         u64 cur_offset;
2465         u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2466         u64 drop_end;
2467         int ret = 0;
2468         int err = 0;
2469         unsigned int rsv_count;
2470         bool same_block;
2471         bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2472         u64 ino_size;
2473         bool truncated_block = false;
2474         bool updated_inode = false;
2475
2476         ret = btrfs_wait_ordered_range(inode, offset, len);
2477         if (ret)
2478                 return ret;
2479
2480         inode_lock(inode);
2481         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2482         ret = find_first_non_hole(inode, &offset, &len);
2483         if (ret < 0)
2484                 goto out_only_mutex;
2485         if (ret && !len) {
2486                 /* Already in a large hole */
2487                 ret = 0;
2488                 goto out_only_mutex;
2489         }
2490
2491         lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2492         lockend = round_down(offset + len,
2493                              btrfs_inode_sectorsize(inode)) - 1;
2494         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2495                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2496         /*
2497          * We needn't truncate any block which is beyond the end of the file
2498          * because we are sure there is no data there.
2499          */
2500         /*
2501          * Only do this if we are in the same block and we aren't doing the
2502          * entire block.
2503          */
2504         if (same_block && len < fs_info->sectorsize) {
2505                 if (offset < ino_size) {
2506                         truncated_block = true;
2507                         ret = btrfs_truncate_block(inode, offset, len, 0);
2508                 } else {
2509                         ret = 0;
2510                 }
2511                 goto out_only_mutex;
2512         }
2513
2514         /* zero back part of the first block */
2515         if (offset < ino_size) {
2516                 truncated_block = true;
2517                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2518                 if (ret) {
2519                         inode_unlock(inode);
2520                         return ret;
2521                 }
2522         }
2523
2524         /* Check the aligned pages after the first unaligned page,
2525          * if offset != orig_start, which means the first unaligned page
2526          * including several following pages are already in holes,
2527          * the extra check can be skipped */
2528         if (offset == orig_start) {
2529                 /* after truncate page, check hole again */
2530                 len = offset + len - lockstart;
2531                 offset = lockstart;
2532                 ret = find_first_non_hole(inode, &offset, &len);
2533                 if (ret < 0)
2534                         goto out_only_mutex;
2535                 if (ret && !len) {
2536                         ret = 0;
2537                         goto out_only_mutex;
2538                 }
2539                 lockstart = offset;
2540         }
2541
2542         /* Check the tail unaligned part is in a hole */
2543         tail_start = lockend + 1;
2544         tail_len = offset + len - tail_start;
2545         if (tail_len) {
2546                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2547                 if (unlikely(ret < 0))
2548                         goto out_only_mutex;
2549                 if (!ret) {
2550                         /* zero the front end of the last page */
2551                         if (tail_start + tail_len < ino_size) {
2552                                 truncated_block = true;
2553                                 ret = btrfs_truncate_block(inode,
2554                                                         tail_start + tail_len,
2555                                                         0, 1);
2556                                 if (ret)
2557                                         goto out_only_mutex;
2558                         }
2559                 }
2560         }
2561
2562         if (lockend < lockstart) {
2563                 ret = 0;
2564                 goto out_only_mutex;
2565         }
2566
2567         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2568                                           &cached_state);
2569         if (ret)
2570                 goto out_only_mutex;
2571
2572         path = btrfs_alloc_path();
2573         if (!path) {
2574                 ret = -ENOMEM;
2575                 goto out;
2576         }
2577
2578         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2579         if (!rsv) {
2580                 ret = -ENOMEM;
2581                 goto out_free;
2582         }
2583         rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2584         rsv->failfast = 1;
2585
2586         /*
2587          * 1 - update the inode
2588          * 1 - removing the extents in the range
2589          * 1 - adding the hole extent if no_holes isn't set
2590          */
2591         rsv_count = no_holes ? 2 : 3;
2592         trans = btrfs_start_transaction(root, rsv_count);
2593         if (IS_ERR(trans)) {
2594                 err = PTR_ERR(trans);
2595                 goto out_free;
2596         }
2597
2598         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2599                                       min_size, false);
2600         BUG_ON(ret);
2601         trans->block_rsv = rsv;
2602
2603         cur_offset = lockstart;
2604         len = lockend - cur_offset;
2605         while (cur_offset < lockend) {
2606                 ret = __btrfs_drop_extents(trans, root, inode, path,
2607                                            cur_offset, lockend + 1,
2608                                            &drop_end, 1, 0, 0, NULL);
2609                 if (ret != -ENOSPC)
2610                         break;
2611
2612                 trans->block_rsv = &fs_info->trans_block_rsv;
2613
2614                 if (cur_offset < drop_end && cur_offset < ino_size) {
2615                         ret = fill_holes(trans, BTRFS_I(inode), path,
2616                                         cur_offset, drop_end);
2617                         if (ret) {
2618                                 /*
2619                                  * If we failed then we didn't insert our hole
2620                                  * entries for the area we dropped, so now the
2621                                  * fs is corrupted, so we must abort the
2622                                  * transaction.
2623                                  */
2624                                 btrfs_abort_transaction(trans, ret);
2625                                 err = ret;
2626                                 break;
2627                         }
2628                 }
2629
2630                 cur_offset = drop_end;
2631
2632                 ret = btrfs_update_inode(trans, root, inode);
2633                 if (ret) {
2634                         err = ret;
2635                         break;
2636                 }
2637
2638                 btrfs_end_transaction(trans);
2639                 btrfs_btree_balance_dirty(fs_info);
2640
2641                 trans = btrfs_start_transaction(root, rsv_count);
2642                 if (IS_ERR(trans)) {
2643                         ret = PTR_ERR(trans);
2644                         trans = NULL;
2645                         break;
2646                 }
2647
2648                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2649                                               rsv, min_size, false);
2650                 BUG_ON(ret);    /* shouldn't happen */
2651                 trans->block_rsv = rsv;
2652
2653                 ret = find_first_non_hole(inode, &cur_offset, &len);
2654                 if (unlikely(ret < 0))
2655                         break;
2656                 if (ret && !len) {
2657                         ret = 0;
2658                         break;
2659                 }
2660         }
2661
2662         if (ret) {
2663                 err = ret;
2664                 goto out_trans;
2665         }
2666
2667         trans->block_rsv = &fs_info->trans_block_rsv;
2668         /*
2669          * If we are using the NO_HOLES feature we might have had already an
2670          * hole that overlaps a part of the region [lockstart, lockend] and
2671          * ends at (or beyond) lockend. Since we have no file extent items to
2672          * represent holes, drop_end can be less than lockend and so we must
2673          * make sure we have an extent map representing the existing hole (the
2674          * call to __btrfs_drop_extents() might have dropped the existing extent
2675          * map representing the existing hole), otherwise the fast fsync path
2676          * will not record the existence of the hole region
2677          * [existing_hole_start, lockend].
2678          */
2679         if (drop_end <= lockend)
2680                 drop_end = lockend + 1;
2681         /*
2682          * Don't insert file hole extent item if it's for a range beyond eof
2683          * (because it's useless) or if it represents a 0 bytes range (when
2684          * cur_offset == drop_end).
2685          */
2686         if (cur_offset < ino_size && cur_offset < drop_end) {
2687                 ret = fill_holes(trans, BTRFS_I(inode), path,
2688                                 cur_offset, drop_end);
2689                 if (ret) {
2690                         /* Same comment as above. */
2691                         btrfs_abort_transaction(trans, ret);
2692                         err = ret;
2693                         goto out_trans;
2694                 }
2695         }
2696
2697 out_trans:
2698         if (!trans)
2699                 goto out_free;
2700
2701         inode_inc_iversion(inode);
2702         inode->i_mtime = inode->i_ctime = current_time(inode);
2703
2704         trans->block_rsv = &fs_info->trans_block_rsv;
2705         ret = btrfs_update_inode(trans, root, inode);
2706         updated_inode = true;
2707         btrfs_end_transaction(trans);
2708         btrfs_btree_balance_dirty(fs_info);
2709 out_free:
2710         btrfs_free_path(path);
2711         btrfs_free_block_rsv(fs_info, rsv);
2712 out:
2713         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2714                              &cached_state);
2715 out_only_mutex:
2716         if (!updated_inode && truncated_block && !ret && !err) {
2717                 /*
2718                  * If we only end up zeroing part of a page, we still need to
2719                  * update the inode item, so that all the time fields are
2720                  * updated as well as the necessary btrfs inode in memory fields
2721                  * for detecting, at fsync time, if the inode isn't yet in the
2722                  * log tree or it's there but not up to date.
2723                  */
2724                 trans = btrfs_start_transaction(root, 1);
2725                 if (IS_ERR(trans)) {
2726                         err = PTR_ERR(trans);
2727                 } else {
2728                         err = btrfs_update_inode(trans, root, inode);
2729                         ret = btrfs_end_transaction(trans);
2730                 }
2731         }
2732         inode_unlock(inode);
2733         if (ret && !err)
2734                 err = ret;
2735         return err;
2736 }
2737
2738 /* Helper structure to record which range is already reserved */
2739 struct falloc_range {
2740         struct list_head list;
2741         u64 start;
2742         u64 len;
2743 };
2744
2745 /*
2746  * Helper function to add falloc range
2747  *
2748  * Caller should have locked the larger range of extent containing
2749  * [start, len)
2750  */
2751 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2752 {
2753         struct falloc_range *prev = NULL;
2754         struct falloc_range *range = NULL;
2755
2756         if (list_empty(head))
2757                 goto insert;
2758
2759         /*
2760          * As fallocate iterate by bytenr order, we only need to check
2761          * the last range.
2762          */
2763         prev = list_entry(head->prev, struct falloc_range, list);
2764         if (prev->start + prev->len == start) {
2765                 prev->len += len;
2766                 return 0;
2767         }
2768 insert:
2769         range = kmalloc(sizeof(*range), GFP_KERNEL);
2770         if (!range)
2771                 return -ENOMEM;
2772         range->start = start;
2773         range->len = len;
2774         list_add_tail(&range->list, head);
2775         return 0;
2776 }
2777
2778 static int btrfs_fallocate_update_isize(struct inode *inode,
2779                                         const u64 end,
2780                                         const int mode)
2781 {
2782         struct btrfs_trans_handle *trans;
2783         struct btrfs_root *root = BTRFS_I(inode)->root;
2784         int ret;
2785         int ret2;
2786
2787         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2788                 return 0;
2789
2790         trans = btrfs_start_transaction(root, 1);
2791         if (IS_ERR(trans))
2792                 return PTR_ERR(trans);
2793
2794         inode->i_ctime = current_time(inode);
2795         i_size_write(inode, end);
2796         btrfs_ordered_update_i_size(inode, end, NULL);
2797         ret = btrfs_update_inode(trans, root, inode);
2798         ret2 = btrfs_end_transaction(trans);
2799
2800         return ret ? ret : ret2;
2801 }
2802
2803 enum {
2804         RANGE_BOUNDARY_WRITTEN_EXTENT = 0,
2805         RANGE_BOUNDARY_PREALLOC_EXTENT = 1,
2806         RANGE_BOUNDARY_HOLE = 2,
2807 };
2808
2809 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2810                                                  u64 offset)
2811 {
2812         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2813         struct extent_map *em;
2814         int ret;
2815
2816         offset = round_down(offset, sectorsize);
2817         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2818         if (IS_ERR(em))
2819                 return PTR_ERR(em);
2820
2821         if (em->block_start == EXTENT_MAP_HOLE)
2822                 ret = RANGE_BOUNDARY_HOLE;
2823         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2824                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2825         else
2826                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2827
2828         free_extent_map(em);
2829         return ret;
2830 }
2831
2832 static int btrfs_zero_range(struct inode *inode,
2833                             loff_t offset,
2834                             loff_t len,
2835                             const int mode)
2836 {
2837         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2838         struct extent_map *em;
2839         struct extent_changeset *data_reserved = NULL;
2840         int ret;
2841         u64 alloc_hint = 0;
2842         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2843         u64 alloc_start = round_down(offset, sectorsize);
2844         u64 alloc_end = round_up(offset + len, sectorsize);
2845         u64 bytes_to_reserve = 0;
2846         bool space_reserved = false;
2847
2848         inode_dio_wait(inode);
2849
2850         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2851                               alloc_start, alloc_end - alloc_start, 0);
2852         if (IS_ERR(em)) {
2853                 ret = PTR_ERR(em);
2854                 goto out;
2855         }
2856
2857         /*
2858          * Avoid hole punching and extent allocation for some cases. More cases
2859          * could be considered, but these are unlikely common and we keep things
2860          * as simple as possible for now. Also, intentionally, if the target
2861          * range contains one or more prealloc extents together with regular
2862          * extents and holes, we drop all the existing extents and allocate a
2863          * new prealloc extent, so that we get a larger contiguous disk extent.
2864          */
2865         if (em->start <= alloc_start &&
2866             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2867                 const u64 em_end = em->start + em->len;
2868
2869                 if (em_end >= offset + len) {
2870                         /*
2871                          * The whole range is already a prealloc extent,
2872                          * do nothing except updating the inode's i_size if
2873                          * needed.
2874                          */
2875                         free_extent_map(em);
2876                         ret = btrfs_fallocate_update_isize(inode, offset + len,
2877                                                            mode);
2878                         goto out;
2879                 }
2880                 /*
2881                  * Part of the range is already a prealloc extent, so operate
2882                  * only on the remaining part of the range.
2883                  */
2884                 alloc_start = em_end;
2885                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2886                 len = offset + len - alloc_start;
2887                 offset = alloc_start;
2888                 alloc_hint = em->block_start + em->len;
2889         }
2890         free_extent_map(em);
2891
2892         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2893             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2894                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2895                                       alloc_start, sectorsize, 0);
2896                 if (IS_ERR(em)) {
2897                         ret = PTR_ERR(em);
2898                         goto out;
2899                 }
2900
2901                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2902                         free_extent_map(em);
2903                         ret = btrfs_fallocate_update_isize(inode, offset + len,
2904                                                            mode);
2905                         goto out;
2906                 }
2907                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2908                         free_extent_map(em);
2909                         ret = btrfs_truncate_block(inode, offset, len, 0);
2910                         if (!ret)
2911                                 ret = btrfs_fallocate_update_isize(inode,
2912                                                                    offset + len,
2913                                                                    mode);
2914                         return ret;
2915                 }
2916                 free_extent_map(em);
2917                 alloc_start = round_down(offset, sectorsize);
2918                 alloc_end = alloc_start + sectorsize;
2919                 goto reserve_space;
2920         }
2921
2922         alloc_start = round_up(offset, sectorsize);
2923         alloc_end = round_down(offset + len, sectorsize);
2924
2925         /*
2926          * For unaligned ranges, check the pages at the boundaries, they might
2927          * map to an extent, in which case we need to partially zero them, or
2928          * they might map to a hole, in which case we need our allocation range
2929          * to cover them.
2930          */
2931         if (!IS_ALIGNED(offset, sectorsize)) {
2932                 ret = btrfs_zero_range_check_range_boundary(inode, offset);
2933                 if (ret < 0)
2934                         goto out;
2935                 if (ret == RANGE_BOUNDARY_HOLE) {
2936                         alloc_start = round_down(offset, sectorsize);
2937                         ret = 0;
2938                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2939                         ret = btrfs_truncate_block(inode, offset, 0, 0);
2940                         if (ret)
2941                                 goto out;
2942                 } else {
2943                         ret = 0;
2944                 }
2945         }
2946
2947         if (!IS_ALIGNED(offset + len, sectorsize)) {
2948                 ret = btrfs_zero_range_check_range_boundary(inode,
2949                                                             offset + len);
2950                 if (ret < 0)
2951                         goto out;
2952                 if (ret == RANGE_BOUNDARY_HOLE) {
2953                         alloc_end = round_up(offset + len, sectorsize);
2954                         ret = 0;
2955                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2956                         ret = btrfs_truncate_block(inode, offset + len, 0, 1);
2957                         if (ret)
2958                                 goto out;
2959                 } else {
2960                         ret = 0;
2961                 }
2962         }
2963
2964 reserve_space:
2965         if (alloc_start < alloc_end) {
2966                 struct extent_state *cached_state = NULL;
2967                 const u64 lockstart = alloc_start;
2968                 const u64 lockend = alloc_end - 1;
2969
2970                 bytes_to_reserve = alloc_end - alloc_start;
2971                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2972                                                       bytes_to_reserve);
2973                 if (ret < 0)
2974                         goto out;
2975                 space_reserved = true;
2976                 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
2977                                                 alloc_start, bytes_to_reserve);
2978                 if (ret)
2979                         goto out;
2980                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2981                                                   &cached_state);
2982                 if (ret)
2983                         goto out;
2984                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
2985                                                 alloc_end - alloc_start,
2986                                                 i_blocksize(inode),
2987                                                 offset + len, &alloc_hint);
2988                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2989                                      lockend, &cached_state);
2990                 /* btrfs_prealloc_file_range releases reserved space on error */
2991                 if (ret) {
2992                         space_reserved = false;
2993                         goto out;
2994                 }
2995         }
2996         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
2997  out:
2998         if (ret && space_reserved)
2999                 btrfs_free_reserved_data_space(inode, data_reserved,
3000                                                alloc_start, bytes_to_reserve);
3001         extent_changeset_free(data_reserved);
3002
3003         return ret;
3004 }
3005
3006 static long btrfs_fallocate(struct file *file, int mode,
3007                             loff_t offset, loff_t len)
3008 {
3009         struct inode *inode = file_inode(file);
3010         struct extent_state *cached_state = NULL;
3011         struct extent_changeset *data_reserved = NULL;
3012         struct falloc_range *range;
3013         struct falloc_range *tmp;
3014         struct list_head reserve_list;
3015         u64 cur_offset;
3016         u64 last_byte;
3017         u64 alloc_start;
3018         u64 alloc_end;
3019         u64 alloc_hint = 0;
3020         u64 locked_end;
3021         u64 actual_end = 0;
3022         struct extent_map *em;
3023         int blocksize = btrfs_inode_sectorsize(inode);
3024         int ret;
3025
3026         alloc_start = round_down(offset, blocksize);
3027         alloc_end = round_up(offset + len, blocksize);
3028         cur_offset = alloc_start;
3029
3030         /* Make sure we aren't being give some crap mode */
3031         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3032                      FALLOC_FL_ZERO_RANGE))
3033                 return -EOPNOTSUPP;
3034
3035         if (mode & FALLOC_FL_PUNCH_HOLE)
3036                 return btrfs_punch_hole(inode, offset, len);
3037
3038         /*
3039          * Only trigger disk allocation, don't trigger qgroup reserve
3040          *
3041          * For qgroup space, it will be checked later.
3042          */
3043         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3044                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3045                                                       alloc_end - alloc_start);
3046                 if (ret < 0)
3047                         return ret;
3048         }
3049
3050         inode_lock(inode);
3051
3052         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3053                 ret = inode_newsize_ok(inode, offset + len);
3054                 if (ret)
3055                         goto out;
3056         }
3057
3058         /*
3059          * TODO: Move these two operations after we have checked
3060          * accurate reserved space, or fallocate can still fail but
3061          * with page truncated or size expanded.
3062          *
3063          * But that's a minor problem and won't do much harm BTW.
3064          */
3065         if (alloc_start > inode->i_size) {
3066                 ret = btrfs_cont_expand(inode, i_size_read(inode),
3067                                         alloc_start);
3068                 if (ret)
3069                         goto out;
3070         } else if (offset + len > inode->i_size) {
3071                 /*
3072                  * If we are fallocating from the end of the file onward we
3073                  * need to zero out the end of the block if i_size lands in the
3074                  * middle of a block.
3075                  */
3076                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3077                 if (ret)
3078                         goto out;
3079         }
3080
3081         /*
3082          * wait for ordered IO before we have any locks.  We'll loop again
3083          * below with the locks held.
3084          */
3085         ret = btrfs_wait_ordered_range(inode, alloc_start,
3086                                        alloc_end - alloc_start);
3087         if (ret)
3088                 goto out;
3089
3090         if (mode & FALLOC_FL_ZERO_RANGE) {
3091                 ret = btrfs_zero_range(inode, offset, len, mode);
3092                 inode_unlock(inode);
3093                 return ret;
3094         }
3095
3096         locked_end = alloc_end - 1;
3097         while (1) {
3098                 struct btrfs_ordered_extent *ordered;
3099
3100                 /* the extent lock is ordered inside the running
3101                  * transaction
3102                  */
3103                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3104                                  locked_end, &cached_state);
3105                 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3106
3107                 if (ordered &&
3108                     ordered->file_offset + ordered->len > alloc_start &&
3109                     ordered->file_offset < alloc_end) {
3110                         btrfs_put_ordered_extent(ordered);
3111                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3112                                              alloc_start, locked_end,
3113                                              &cached_state);
3114                         /*
3115                          * we can't wait on the range with the transaction
3116                          * running or with the extent lock held
3117                          */
3118                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3119                                                        alloc_end - alloc_start);
3120                         if (ret)
3121                                 goto out;
3122                 } else {
3123                         if (ordered)
3124                                 btrfs_put_ordered_extent(ordered);
3125                         break;
3126                 }
3127         }
3128
3129         /* First, check if we exceed the qgroup limit */
3130         INIT_LIST_HEAD(&reserve_list);
3131         while (cur_offset < alloc_end) {
3132                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3133                                       alloc_end - cur_offset, 0);
3134                 if (IS_ERR(em)) {
3135                         ret = PTR_ERR(em);
3136                         break;
3137                 }
3138                 last_byte = min(extent_map_end(em), alloc_end);
3139                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3140                 last_byte = ALIGN(last_byte, blocksize);
3141                 if (em->block_start == EXTENT_MAP_HOLE ||
3142                     (cur_offset >= inode->i_size &&
3143                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3144                         ret = add_falloc_range(&reserve_list, cur_offset,
3145                                                last_byte - cur_offset);
3146                         if (ret < 0) {
3147                                 free_extent_map(em);
3148                                 break;
3149                         }
3150                         ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3151                                         cur_offset, last_byte - cur_offset);
3152                         if (ret < 0) {
3153                                 cur_offset = last_byte;
3154                                 free_extent_map(em);
3155                                 break;
3156                         }
3157                 } else {
3158                         /*
3159                          * Do not need to reserve unwritten extent for this
3160                          * range, free reserved data space first, otherwise
3161                          * it'll result in false ENOSPC error.
3162                          */
3163                         btrfs_free_reserved_data_space(inode, data_reserved,
3164                                         cur_offset, last_byte - cur_offset);
3165                 }
3166                 free_extent_map(em);
3167                 cur_offset = last_byte;
3168         }
3169
3170         /*
3171          * If ret is still 0, means we're OK to fallocate.
3172          * Or just cleanup the list and exit.
3173          */
3174         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3175                 if (!ret)
3176                         ret = btrfs_prealloc_file_range(inode, mode,
3177                                         range->start,
3178                                         range->len, i_blocksize(inode),
3179                                         offset + len, &alloc_hint);
3180                 else
3181                         btrfs_free_reserved_data_space(inode,
3182                                         data_reserved, range->start,
3183                                         range->len);
3184                 list_del(&range->list);
3185                 kfree(range);
3186         }
3187         if (ret < 0)
3188                 goto out_unlock;
3189
3190         /*
3191          * We didn't need to allocate any more space, but we still extended the
3192          * size of the file so we need to update i_size and the inode item.
3193          */
3194         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3195 out_unlock:
3196         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3197                              &cached_state);
3198 out:
3199         inode_unlock(inode);
3200         /* Let go of our reservation. */
3201         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3202                 btrfs_free_reserved_data_space(inode, data_reserved,
3203                                 cur_offset, alloc_end - cur_offset);
3204         extent_changeset_free(data_reserved);
3205         return ret;
3206 }
3207
3208 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3209 {
3210         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3211         struct extent_map *em = NULL;
3212         struct extent_state *cached_state = NULL;
3213         u64 lockstart;
3214         u64 lockend;
3215         u64 start;
3216         u64 len;
3217         int ret = 0;
3218
3219         if (inode->i_size == 0)
3220                 return -ENXIO;
3221
3222         /*
3223          * *offset can be negative, in this case we start finding DATA/HOLE from
3224          * the very start of the file.
3225          */
3226         start = max_t(loff_t, 0, *offset);
3227
3228         lockstart = round_down(start, fs_info->sectorsize);
3229         lockend = round_up(i_size_read(inode),
3230                            fs_info->sectorsize);
3231         if (lockend <= lockstart)
3232                 lockend = lockstart + fs_info->sectorsize;
3233         lockend--;
3234         len = lockend - lockstart + 1;
3235
3236         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3237                          &cached_state);
3238
3239         while (start < inode->i_size) {
3240                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3241                 if (IS_ERR(em)) {
3242                         ret = PTR_ERR(em);
3243                         em = NULL;
3244                         break;
3245                 }
3246
3247                 if (whence == SEEK_HOLE &&
3248                     (em->block_start == EXTENT_MAP_HOLE ||
3249                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3250                         break;
3251                 else if (whence == SEEK_DATA &&
3252                            (em->block_start != EXTENT_MAP_HOLE &&
3253                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3254                         break;
3255
3256                 start = em->start + em->len;
3257                 free_extent_map(em);
3258                 em = NULL;
3259                 cond_resched();
3260         }
3261         free_extent_map(em);
3262         if (!ret) {
3263                 if (whence == SEEK_DATA && start >= inode->i_size)
3264                         ret = -ENXIO;
3265                 else
3266                         *offset = min_t(loff_t, start, inode->i_size);
3267         }
3268         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3269                              &cached_state);
3270         return ret;
3271 }
3272
3273 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3274 {
3275         struct inode *inode = file->f_mapping->host;
3276         int ret;
3277
3278         inode_lock(inode);
3279         switch (whence) {
3280         case SEEK_END:
3281         case SEEK_CUR:
3282                 offset = generic_file_llseek(file, offset, whence);
3283                 goto out;
3284         case SEEK_DATA:
3285         case SEEK_HOLE:
3286                 if (offset >= i_size_read(inode)) {
3287                         inode_unlock(inode);
3288                         return -ENXIO;
3289                 }
3290
3291                 ret = find_desired_extent(inode, &offset, whence);
3292                 if (ret) {
3293                         inode_unlock(inode);
3294                         return ret;
3295                 }
3296         }
3297
3298         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3299 out:
3300         inode_unlock(inode);
3301         return offset;
3302 }
3303
3304 static int btrfs_file_open(struct inode *inode, struct file *filp)
3305 {
3306         filp->f_mode |= FMODE_NOWAIT;
3307         return generic_file_open(inode, filp);
3308 }
3309
3310 const struct file_operations btrfs_file_operations = {
3311         .llseek         = btrfs_file_llseek,
3312         .read_iter      = generic_file_read_iter,
3313         .splice_read    = generic_file_splice_read,
3314         .write_iter     = btrfs_file_write_iter,
3315         .mmap           = btrfs_file_mmap,
3316         .open           = btrfs_file_open,
3317         .release        = btrfs_release_file,
3318         .fsync          = btrfs_sync_file,
3319         .fallocate      = btrfs_fallocate,
3320         .unlocked_ioctl = btrfs_ioctl,
3321 #ifdef CONFIG_COMPAT
3322         .compat_ioctl   = btrfs_compat_ioctl,
3323 #endif
3324         .remap_file_range = btrfs_remap_file_range,
3325 };
3326
3327 void __cold btrfs_auto_defrag_exit(void)
3328 {
3329         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3330 }
3331
3332 int __init btrfs_auto_defrag_init(void)
3333 {
3334         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3335                                         sizeof(struct inode_defrag), 0,
3336                                         SLAB_MEM_SPREAD,
3337                                         NULL);
3338         if (!btrfs_inode_defrag_cachep)
3339                 return -ENOMEM;
3340
3341         return 0;
3342 }
3343
3344 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3345 {
3346         int ret;
3347
3348         /*
3349          * So with compression we will find and lock a dirty page and clear the
3350          * first one as dirty, setup an async extent, and immediately return
3351          * with the entire range locked but with nobody actually marked with
3352          * writeback.  So we can't just filemap_write_and_wait_range() and
3353          * expect it to work since it will just kick off a thread to do the
3354          * actual work.  So we need to call filemap_fdatawrite_range _again_
3355          * since it will wait on the page lock, which won't be unlocked until
3356          * after the pages have been marked as writeback and so we're good to go
3357          * from there.  We have to do this otherwise we'll miss the ordered
3358          * extents and that results in badness.  Please Josef, do not think you
3359          * know better and pull this out at some point in the future, it is
3360          * right and you are wrong.
3361          */
3362         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3363         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3364                              &BTRFS_I(inode)->runtime_flags))
3365                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3366
3367         return ret;
3368 }