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 #include "delalloc-space.h"
  30 #include "reflink.h"
  31
  32 static struct kmem_cache *btrfs_inode_defrag_cachep;
  33 /*
  34  * when auto defrag is enabled we
  35  * queue up these defrag structs to remember which
  36  * inodes need defragging passes
  37  */
  38 struct inode_defrag {
  39         struct rb_node rb_node;
  40         /* objectid */
  41         u64 ino;
  42         /*
  43          * transid where the defrag was added, we search for
  44          * extents newer than this
  45          */
  46         u64 transid;
  47
  48         /* root objectid */
  49         u64 root;
  50
  51         /* last offset we were able to defrag */
  52         u64 last_offset;
  53
  54         /* if we've wrapped around back to zero once already */
  55         int cycled;
  56 };
  57
  58 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  59                                   struct inode_defrag *defrag2)
  60 {
  61         if (defrag1->root > defrag2->root)
  62                 return 1;
  63         else if (defrag1->root < defrag2->root)
  64                 return -1;
  65         else if (defrag1->ino > defrag2->ino)
  66                 return 1;
  67         else if (defrag1->ino < defrag2->ino)
  68                 return -1;
  69         else
  70                 return 0;
  71 }
  72
  73 /* pop a record for an inode into the defrag tree.  The lock
  74  * must be held already
  75  *
  76  * If you're inserting a record for an older transid than an
  77  * existing record, the transid already in the tree is lowered
  78  *
  79  * If an existing record is found the defrag item you
  80  * pass in is freed
  81  */
  82 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  83                                     struct inode_defrag *defrag)
  84 {
  85         struct btrfs_fs_info *fs_info = inode->root->fs_info;
  86         struct inode_defrag *entry;
  87         struct rb_node **p;
  88         struct rb_node *parent = NULL;
  89         int ret;
  90
  91         p = &fs_info->defrag_inodes.rb_node;
  92         while (*p) {
  93                 parent = *p;
  94                 entry = rb_entry(parent, struct inode_defrag, rb_node);
  95
  96                 ret = __compare_inode_defrag(defrag, entry);
  97                 if (ret < 0)
  98                         p = &parent->rb_left;
  99                 else if (ret > 0)
 100                         p = &parent->rb_right;
 101                 else {
 102                         /* if we're reinserting an entry for
 103                          * an old defrag run, make sure to
 104                          * lower the transid of our existing record
 105                          */
 106                         if (defrag->transid < entry->transid)
 107                                 entry->transid = defrag->transid;
 108                         if (defrag->last_offset > entry->last_offset)
 109                                 entry->last_offset = defrag->last_offset;
 110                         return -EEXIST;
 111                 }
 112         }
 113         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 114         rb_link_node(&defrag->rb_node, parent, p);
 115         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 116         return 0;
 117 }
 118
 119 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 120 {
 121         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 122                 return 0;
 123
 124         if (btrfs_fs_closing(fs_info))
 125                 return 0;
 126
 127         return 1;
 128 }
 129
 130 /*
 131  * insert a defrag record for this inode if auto defrag is
 132  * enabled
 133  */
 134 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 135                            struct btrfs_inode *inode)
 136 {
 137         struct btrfs_root *root = inode->root;
 138         struct btrfs_fs_info *fs_info = root->fs_info;
 139         struct inode_defrag *defrag;
 140         u64 transid;
 141         int ret;
 142
 143         if (!__need_auto_defrag(fs_info))
 144                 return 0;
 145
 146         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 147                 return 0;
 148
 149         if (trans)
 150                 transid = trans->transid;
 151         else
 152                 transid = inode->root->last_trans;
 153
 154         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 155         if (!defrag)
 156                 return -ENOMEM;
 157
 158         defrag->ino = btrfs_ino(inode);
 159         defrag->transid = transid;
 160         defrag->root = root->root_key.objectid;
 161
 162         spin_lock(&fs_info->defrag_inodes_lock);
 163         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 164                 /*
 165                  * If we set IN_DEFRAG flag and evict the inode from memory,
 166                  * and then re-read this inode, this new inode doesn't have
 167                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 168                  */
 169                 ret = __btrfs_add_inode_defrag(inode, defrag);
 170                 if (ret)
 171                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 172         } else {
 173                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 174         }
 175         spin_unlock(&fs_info->defrag_inodes_lock);
 176         return 0;
 177 }
 178
 179 /*
 180  * Requeue the defrag object. If there is a defrag object that points to
 181  * the same inode in the tree, we will merge them together (by
 182  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 183  */
 184 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 185                                        struct inode_defrag *defrag)
 186 {
 187         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 188         int ret;
 189
 190         if (!__need_auto_defrag(fs_info))
 191                 goto out;
 192
 193         /*
 194          * Here we don't check the IN_DEFRAG flag, because we need merge
 195          * them together.
 196          */
 197         spin_lock(&fs_info->defrag_inodes_lock);
 198         ret = __btrfs_add_inode_defrag(inode, defrag);
 199         spin_unlock(&fs_info->defrag_inodes_lock);
 200         if (ret)
 201                 goto out;
 202         return;
 203 out:
 204         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 205 }
 206
 207 /*
 208  * pick the defragable inode that we want, if it doesn't exist, we will get
 209  * the next one.
 210  */
 211 static struct inode_defrag *
 212 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 213 {
 214         struct inode_defrag *entry = NULL;
 215         struct inode_defrag tmp;
 216         struct rb_node *p;
 217         struct rb_node *parent = NULL;
 218         int ret;
 219
 220         tmp.ino = ino;
 221         tmp.root = root;
 222
 223         spin_lock(&fs_info->defrag_inodes_lock);
 224         p = fs_info->defrag_inodes.rb_node;
 225         while (p) {
 226                 parent = p;
 227                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 228
 229                 ret = __compare_inode_defrag(&tmp, entry);
 230                 if (ret < 0)
 231                         p = parent->rb_left;
 232                 else if (ret > 0)
 233                         p = parent->rb_right;
 234                 else
 235                         goto out;
 236         }
 237
 238         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 239                 parent = rb_next(parent);
 240                 if (parent)
 241                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 242                 else
 243                         entry = NULL;
 244         }
 245 out:
 246         if (entry)
 247                 rb_erase(parent, &fs_info->defrag_inodes);
 248         spin_unlock(&fs_info->defrag_inodes_lock);
 249         return entry;
 250 }
 251
 252 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 253 {
 254         struct inode_defrag *defrag;
 255         struct rb_node *node;
 256
 257         spin_lock(&fs_info->defrag_inodes_lock);
 258         node = rb_first(&fs_info->defrag_inodes);
 259         while (node) {
 260                 rb_erase(node, &fs_info->defrag_inodes);
 261                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 262                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 263
 264                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 265
 266                 node = rb_first(&fs_info->defrag_inodes);
 267         }
 268         spin_unlock(&fs_info->defrag_inodes_lock);
 269 }
 270
 271 #define BTRFS_DEFRAG_BATCH      1024
 272
 273 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 274                                     struct inode_defrag *defrag)
 275 {
 276         struct btrfs_root *inode_root;
 277         struct inode *inode;
 278         struct btrfs_key key;
 279         struct btrfs_ioctl_defrag_range_args range;
 280         int num_defrag;
 281         int ret;
 282
 283         /* get the inode */
 284         key.objectid = defrag->root;
 285         key.type = BTRFS_ROOT_ITEM_KEY;
 286         key.offset = (u64)-1;
 287
 288         inode_root = btrfs_get_fs_root(fs_info, &key, true);
 289         if (IS_ERR(inode_root)) {
 290                 ret = PTR_ERR(inode_root);
 291                 goto cleanup;
 292         }
 293
 294         key.objectid = defrag->ino;
 295         key.type = BTRFS_INODE_ITEM_KEY;
 296         key.offset = 0;
 297         inode = btrfs_iget(fs_info->sb, &key, inode_root);
 298         btrfs_put_root(inode_root);
 299         if (IS_ERR(inode)) {
 300                 ret = PTR_ERR(inode);
 301                 goto cleanup;
 302         }
 303
 304         /* do a chunk of defrag */
 305         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 306         memset(&range, 0, sizeof(range));
 307         range.len = (u64)-1;
 308         range.start = defrag->last_offset;
 309
 310         sb_start_write(fs_info->sb);
 311         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 312                                        BTRFS_DEFRAG_BATCH);
 313         sb_end_write(fs_info->sb);
 314         /*
 315          * if we filled the whole defrag batch, there
 316          * must be more work to do.  Queue this defrag
 317          * again
 318          */
 319         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 320                 defrag->last_offset = range.start;
 321                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 322         } else if (defrag->last_offset && !defrag->cycled) {
 323                 /*
 324                  * we didn't fill our defrag batch, but
 325                  * we didn't start at zero.  Make sure we loop
 326                  * around to the start of the file.
 327                  */
 328                 defrag->last_offset = 0;
 329                 defrag->cycled = 1;
 330                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 331         } else {
 332                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 333         }
 334
 335         iput(inode);
 336         return 0;
 337 cleanup:
 338         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 339         return ret;
 340 }
 341
 342 /*
 343  * run through the list of inodes in the FS that need
 344  * defragging
 345  */
 346 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 347 {
 348         struct inode_defrag *defrag;
 349         u64 first_ino = 0;
 350         u64 root_objectid = 0;
 351
 352         atomic_inc(&fs_info->defrag_running);
 353         while (1) {
 354                 /* Pause the auto defragger. */
 355                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 356                              &fs_info->fs_state))
 357                         break;
 358
 359                 if (!__need_auto_defrag(fs_info))
 360                         break;
 361
 362                 /* find an inode to defrag */
 363                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 364                                                  first_ino);
 365                 if (!defrag) {
 366                         if (root_objectid || first_ino) {
 367                                 root_objectid = 0;
 368                                 first_ino = 0;
 369                                 continue;
 370                         } else {
 371                                 break;
 372                         }
 373                 }
 374
 375                 first_ino = defrag->ino + 1;
 376                 root_objectid = defrag->root;
 377
 378                 __btrfs_run_defrag_inode(fs_info, defrag);
 379         }
 380         atomic_dec(&fs_info->defrag_running);
 381
 382         /*
 383          * during unmount, we use the transaction_wait queue to
 384          * wait for the defragger to stop
 385          */
 386         wake_up(&fs_info->transaction_wait);
 387         return 0;
 388 }
 389
 390 /* simple helper to fault in pages and copy.  This should go away
 391  * and be replaced with calls into generic code.
 392  */
 393 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 394                                          struct page **prepared_pages,
 395                                          struct iov_iter *i)
 396 {
 397         size_t copied = 0;
 398         size_t total_copied = 0;
 399         int pg = 0;
 400         int offset = offset_in_page(pos);
 401
 402         while (write_bytes > 0) {
 403                 size_t count = min_t(size_t,
 404                                      PAGE_SIZE - offset, write_bytes);
 405                 struct page *page = prepared_pages[pg];
 406                 /*
 407                  * Copy data from userspace to the current page
 408                  */
 409                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 410
 411                 /* Flush processor's dcache for this page */
 412                 flush_dcache_page(page);
 413
 414                 /*
 415                  * if we get a partial write, we can end up with
 416                  * partially up to date pages.  These add
 417                  * a lot of complexity, so make sure they don't
 418                  * happen by forcing this copy to be retried.
 419                  *
 420                  * The rest of the btrfs_file_write code will fall
 421                  * back to page at a time copies after we return 0.
 422                  */
 423                 if (!PageUptodate(page) && copied < count)
 424                         copied = 0;
 425
 426                 iov_iter_advance(i, copied);
 427                 write_bytes -= copied;
 428                 total_copied += copied;
 429
 430                 /* Return to btrfs_file_write_iter to fault page */
 431                 if (unlikely(copied == 0))
 432                         break;
 433
 434                 if (copied < PAGE_SIZE - offset) {
 435                         offset += copied;
 436                 } else {
 437                         pg++;
 438                         offset = 0;
 439                 }
 440         }
 441         return total_copied;
 442 }
 443
 444 /*
 445  * unlocks pages after btrfs_file_write is done with them
 446  */
 447 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 448 {
 449         size_t i;
 450         for (i = 0; i < num_pages; i++) {
 451                 /* page checked is some magic around finding pages that
 452                  * have been modified without going through btrfs_set_page_dirty
 453                  * clear it here. There should be no need to mark the pages
 454                  * accessed as prepare_pages should have marked them accessed
 455                  * in prepare_pages via find_or_create_page()
 456                  */
 457                 ClearPageChecked(pages[i]);
 458                 unlock_page(pages[i]);
 459                 put_page(pages[i]);
 460         }
 461 }
 462
 463 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
 464                                          const u64 start,
 465                                          const u64 len,
 466                                          struct extent_state **cached_state)
 467 {
 468         u64 search_start = start;
 469         const u64 end = start + len - 1;
 470
 471         while (search_start < end) {
 472                 const u64 search_len = end - search_start + 1;
 473                 struct extent_map *em;
 474                 u64 em_len;
 475                 int ret = 0;
 476
 477                 em = btrfs_get_extent(inode, NULL, 0, search_start, search_len);
 478                 if (IS_ERR(em))
 479                         return PTR_ERR(em);
 480
 481                 if (em->block_start != EXTENT_MAP_HOLE)
 482                         goto next;
 483
 484                 em_len = em->len;
 485                 if (em->start < search_start)
 486                         em_len -= search_start - em->start;
 487                 if (em_len > search_len)
 488                         em_len = search_len;
 489
 490                 ret = set_extent_bit(&inode->io_tree, search_start,
 491                                      search_start + em_len - 1,
 492                                      EXTENT_DELALLOC_NEW,
 493                                      NULL, cached_state, GFP_NOFS);
 494 next:
 495                 search_start = extent_map_end(em);
 496                 free_extent_map(em);
 497                 if (ret)
 498                         return ret;
 499         }
 500         return 0;
 501 }
 502
 503 /*
 504  * after copy_from_user, pages need to be dirtied and we need to make
 505  * sure holes are created between the current EOF and the start of
 506  * any next extents (if required).
 507  *
 508  * this also makes the decision about creating an inline extent vs
 509  * doing real data extents, marking pages dirty and delalloc as required.
 510  */
 511 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
 512                       size_t num_pages, loff_t pos, size_t write_bytes,
 513                       struct extent_state **cached)
 514 {
 515         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 516         int err = 0;
 517         int i;
 518         u64 num_bytes;
 519         u64 start_pos;
 520         u64 end_of_last_block;
 521         u64 end_pos = pos + write_bytes;
 522         loff_t isize = i_size_read(inode);
 523         unsigned int extra_bits = 0;
 524
 525         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 526         num_bytes = round_up(write_bytes + pos - start_pos,
 527                              fs_info->sectorsize);
 528
 529         end_of_last_block = start_pos + num_bytes - 1;
 530
 531         /*
 532          * The pages may have already been dirty, clear out old accounting so
 533          * we can set things up properly
 534          */
 535         clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos, end_of_last_block,
 536                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
 537                          0, 0, cached);
 538
 539         if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
 540                 if (start_pos >= isize &&
 541                     !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
 542                         /*
 543                          * There can't be any extents following eof in this case
 544                          * so just set the delalloc new bit for the range
 545                          * directly.
 546                          */
 547                         extra_bits |= EXTENT_DELALLOC_NEW;
 548                 } else {
 549                         err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
 550                                                             start_pos,
 551                                                             num_bytes, cached);
 552                         if (err)
 553                                 return err;
 554                 }
 555         }
 556
 557         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 558                                         extra_bits, cached);
 559         if (err)
 560                 return err;
 561
 562         for (i = 0; i < num_pages; i++) {
 563                 struct page *p = pages[i];
 564                 SetPageUptodate(p);
 565                 ClearPageChecked(p);
 566                 set_page_dirty(p);
 567         }
 568
 569         /*
 570          * we've only changed i_size in ram, and we haven't updated
 571          * the disk i_size.  There is no need to log the inode
 572          * at this time.
 573          */
 574         if (end_pos > isize)
 575                 i_size_write(inode, end_pos);
 576         return 0;
 577 }
 578
 579 /*
 580  * this drops all the extents in the cache that intersect the range
 581  * [start, end].  Existing extents are split as required.
 582  */
 583 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 584                              int skip_pinned)
 585 {
 586         struct extent_map *em;
 587         struct extent_map *split = NULL;
 588         struct extent_map *split2 = NULL;
 589         struct extent_map_tree *em_tree = &inode->extent_tree;
 590         u64 len = end - start + 1;
 591         u64 gen;
 592         int ret;
 593         int testend = 1;
 594         unsigned long flags;
 595         int compressed = 0;
 596         bool modified;
 597
 598         WARN_ON(end < start);
 599         if (end == (u64)-1) {
 600                 len = (u64)-1;
 601                 testend = 0;
 602         }
 603         while (1) {
 604                 int no_splits = 0;
 605
 606                 modified = false;
 607                 if (!split)
 608                         split = alloc_extent_map();
 609                 if (!split2)
 610                         split2 = alloc_extent_map();
 611                 if (!split || !split2)
 612                         no_splits = 1;
 613
 614                 write_lock(&em_tree->lock);
 615                 em = lookup_extent_mapping(em_tree, start, len);
 616                 if (!em) {
 617                         write_unlock(&em_tree->lock);
 618                         break;
 619                 }
 620                 flags = em->flags;
 621                 gen = em->generation;
 622                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 623                         if (testend && em->start + em->len >= start + len) {
 624                                 free_extent_map(em);
 625                                 write_unlock(&em_tree->lock);
 626                                 break;
 627                         }
 628                         start = em->start + em->len;
 629                         if (testend)
 630                                 len = start + len - (em->start + em->len);
 631                         free_extent_map(em);
 632                         write_unlock(&em_tree->lock);
 633                         continue;
 634                 }
 635                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 636                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 637                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 638                 modified = !list_empty(&em->list);
 639                 if (no_splits)
 640                         goto next;
 641
 642                 if (em->start < start) {
 643                         split->start = em->start;
 644                         split->len = start - em->start;
 645
 646                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 647                                 split->orig_start = em->orig_start;
 648                                 split->block_start = em->block_start;
 649
 650                                 if (compressed)
 651                                         split->block_len = em->block_len;
 652                                 else
 653                                         split->block_len = split->len;
 654                                 split->orig_block_len = max(split->block_len,
 655                                                 em->orig_block_len);
 656                                 split->ram_bytes = em->ram_bytes;
 657                         } else {
 658                                 split->orig_start = split->start;
 659                                 split->block_len = 0;
 660                                 split->block_start = em->block_start;
 661                                 split->orig_block_len = 0;
 662                                 split->ram_bytes = split->len;
 663                         }
 664
 665                         split->generation = gen;
 666                         split->flags = flags;
 667                         split->compress_type = em->compress_type;
 668                         replace_extent_mapping(em_tree, em, split, modified);
 669                         free_extent_map(split);
 670                         split = split2;
 671                         split2 = NULL;
 672                 }
 673                 if (testend && em->start + em->len > start + len) {
 674                         u64 diff = start + len - em->start;
 675
 676                         split->start = start + len;
 677                         split->len = em->start + em->len - (start + len);
 678                         split->flags = flags;
 679                         split->compress_type = em->compress_type;
 680                         split->generation = gen;
 681
 682                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 683                                 split->orig_block_len = max(em->block_len,
 684                                                     em->orig_block_len);
 685
 686                                 split->ram_bytes = em->ram_bytes;
 687                                 if (compressed) {
 688                                         split->block_len = em->block_len;
 689                                         split->block_start = em->block_start;
 690                                         split->orig_start = em->orig_start;
 691                                 } else {
 692                                         split->block_len = split->len;
 693                                         split->block_start = em->block_start
 694                                                 + diff;
 695                                         split->orig_start = em->orig_start;
 696                                 }
 697                         } else {
 698                                 split->ram_bytes = split->len;
 699                                 split->orig_start = split->start;
 700                                 split->block_len = 0;
 701                                 split->block_start = em->block_start;
 702                                 split->orig_block_len = 0;
 703                         }
 704
 705                         if (extent_map_in_tree(em)) {
 706                                 replace_extent_mapping(em_tree, em, split,
 707                                                        modified);
 708                         } else {
 709                                 ret = add_extent_mapping(em_tree, split,
 710                                                          modified);
 711                                 ASSERT(ret == 0); /* Logic error */
 712                         }
 713                         free_extent_map(split);
 714                         split = NULL;
 715                 }
 716 next:
 717                 if (extent_map_in_tree(em))
 718                         remove_extent_mapping(em_tree, em);
 719                 write_unlock(&em_tree->lock);
 720
 721                 /* once for us */
 722                 free_extent_map(em);
 723                 /* once for the tree*/
 724                 free_extent_map(em);
 725         }
 726         if (split)
 727                 free_extent_map(split);
 728         if (split2)
 729                 free_extent_map(split2);
 730 }
 731
 732 /*
 733  * this is very complex, but the basic idea is to drop all extents
 734  * in the range start - end.  hint_block is filled in with a block number
 735  * that would be a good hint to the block allocator for this file.
 736  *
 737  * If an extent intersects the range but is not entirely inside the range
 738  * it is either truncated or split.  Anything entirely inside the range
 739  * is deleted from the tree.
 740  */
 741 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 742                          struct btrfs_root *root, struct inode *inode,
 743                          struct btrfs_path *path, u64 start, u64 end,
 744                          u64 *drop_end, int drop_cache,
 745                          int replace_extent,
 746                          u32 extent_item_size,
 747                          int *key_inserted)
 748 {
 749         struct btrfs_fs_info *fs_info = root->fs_info;
 750         struct extent_buffer *leaf;
 751         struct btrfs_file_extent_item *fi;
 752         struct btrfs_ref ref = { 0 };
 753         struct btrfs_key key;
 754         struct btrfs_key new_key;
 755         u64 ino = btrfs_ino(BTRFS_I(inode));
 756         u64 search_start = start;
 757         u64 disk_bytenr = 0;
 758         u64 num_bytes = 0;
 759         u64 extent_offset = 0;
 760         u64 extent_end = 0;
 761         u64 last_end = start;
 762         int del_nr = 0;
 763         int del_slot = 0;
 764         int extent_type;
 765         int recow;
 766         int ret;
 767         int modify_tree = -1;
 768         int update_refs;
 769         int found = 0;
 770         int leafs_visited = 0;
 771
 772         if (drop_cache)
 773                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
 774
 775         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 776                 modify_tree = 0;
 777
 778         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 779                        root == fs_info->tree_root);
 780         while (1) {
 781                 recow = 0;
 782                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 783                                                search_start, modify_tree);
 784                 if (ret < 0)
 785                         break;
 786                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 787                         leaf = path->nodes[0];
 788                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 789                         if (key.objectid == ino &&
 790                             key.type == BTRFS_EXTENT_DATA_KEY)
 791                                 path->slots[0]--;
 792                 }
 793                 ret = 0;
 794                 leafs_visited++;
 795 next_slot:
 796                 leaf = path->nodes[0];
 797                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 798                         BUG_ON(del_nr > 0);
 799                         ret = btrfs_next_leaf(root, path);
 800                         if (ret < 0)
 801                                 break;
 802                         if (ret > 0) {
 803                                 ret = 0;
 804                                 break;
 805                         }
 806                         leafs_visited++;
 807                         leaf = path->nodes[0];
 808                         recow = 1;
 809                 }
 810
 811                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 812
 813                 if (key.objectid > ino)
 814                         break;
 815                 if (WARN_ON_ONCE(key.objectid < ino) ||
 816                     key.type < BTRFS_EXTENT_DATA_KEY) {
 817                         ASSERT(del_nr == 0);
 818                         path->slots[0]++;
 819                         goto next_slot;
 820                 }
 821                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 822                         break;
 823
 824                 fi = btrfs_item_ptr(leaf, path->slots[0],
 825                                     struct btrfs_file_extent_item);
 826                 extent_type = btrfs_file_extent_type(leaf, fi);
 827
 828                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 829                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 830                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 831                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 832                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 833                         extent_end = key.offset +
 834                                 btrfs_file_extent_num_bytes(leaf, fi);
 835                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 836                         extent_end = key.offset +
 837                                 btrfs_file_extent_ram_bytes(leaf, fi);
 838                 } else {
 839                         /* can't happen */
 840                         BUG();
 841                 }
 842
 843                 /*
 844                  * Don't skip extent items representing 0 byte lengths. They
 845                  * used to be created (bug) if while punching holes we hit
 846                  * -ENOSPC condition. So if we find one here, just ensure we
 847                  * delete it, otherwise we would insert a new file extent item
 848                  * with the same key (offset) as that 0 bytes length file
 849                  * extent item in the call to setup_items_for_insert() later
 850                  * in this function.
 851                  */
 852                 if (extent_end == key.offset && extent_end >= search_start) {
 853                         last_end = extent_end;
 854                         goto delete_extent_item;
 855                 }
 856
 857                 if (extent_end <= search_start) {
 858                         path->slots[0]++;
 859                         goto next_slot;
 860                 }
 861
 862                 found = 1;
 863                 search_start = max(key.offset, start);
 864                 if (recow || !modify_tree) {
 865                         modify_tree = -1;
 866                         btrfs_release_path(path);
 867                         continue;
 868                 }
 869
 870                 /*
 871                  *     | - range to drop - |
 872                  *  | -------- extent -------- |
 873                  */
 874                 if (start > key.offset && end < extent_end) {
 875                         BUG_ON(del_nr > 0);
 876                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 877                                 ret = -EOPNOTSUPP;
 878                                 break;
 879                         }
 880
 881                         memcpy(&new_key, &key, sizeof(new_key));
 882                         new_key.offset = start;
 883                         ret = btrfs_duplicate_item(trans, root, path,
 884                                                    &new_key);
 885                         if (ret == -EAGAIN) {
 886                                 btrfs_release_path(path);
 887                                 continue;
 888                         }
 889                         if (ret < 0)
 890                                 break;
 891
 892                         leaf = path->nodes[0];
 893                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 894                                             struct btrfs_file_extent_item);
 895                         btrfs_set_file_extent_num_bytes(leaf, fi,
 896                                                         start - key.offset);
 897
 898                         fi = btrfs_item_ptr(leaf, path->slots[0],
 899                                             struct btrfs_file_extent_item);
 900
 901                         extent_offset += start - key.offset;
 902                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 903                         btrfs_set_file_extent_num_bytes(leaf, fi,
 904                                                         extent_end - start);
 905                         btrfs_mark_buffer_dirty(leaf);
 906
 907                         if (update_refs && disk_bytenr > 0) {
 908                                 btrfs_init_generic_ref(&ref,
 909                                                 BTRFS_ADD_DELAYED_REF,
 910                                                 disk_bytenr, num_bytes, 0);
 911                                 btrfs_init_data_ref(&ref,
 912                                                 root->root_key.objectid,
 913                                                 new_key.objectid,
 914                                                 start - extent_offset);
 915                                 ret = btrfs_inc_extent_ref(trans, &ref);
 916                                 BUG_ON(ret); /* -ENOMEM */
 917                         }
 918                         key.offset = start;
 919                 }
 920                 /*
 921                  * From here on out we will have actually dropped something, so
 922                  * last_end can be updated.
 923                  */
 924                 last_end = extent_end;
 925
 926                 /*
 927                  *  | ---- range to drop ----- |
 928                  *      | -------- extent -------- |
 929                  */
 930                 if (start <= key.offset && end < extent_end) {
 931                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 932                                 ret = -EOPNOTSUPP;
 933                                 break;
 934                         }
 935
 936                         memcpy(&new_key, &key, sizeof(new_key));
 937                         new_key.offset = end;
 938                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 939
 940                         extent_offset += end - key.offset;
 941                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 942                         btrfs_set_file_extent_num_bytes(leaf, fi,
 943                                                         extent_end - end);
 944                         btrfs_mark_buffer_dirty(leaf);
 945                         if (update_refs && disk_bytenr > 0)
 946                                 inode_sub_bytes(inode, end - key.offset);
 947                         break;
 948                 }
 949
 950                 search_start = extent_end;
 951                 /*
 952                  *       | ---- range to drop ----- |
 953                  *  | -------- extent -------- |
 954                  */
 955                 if (start > key.offset && end >= extent_end) {
 956                         BUG_ON(del_nr > 0);
 957                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 958                                 ret = -EOPNOTSUPP;
 959                                 break;
 960                         }
 961
 962                         btrfs_set_file_extent_num_bytes(leaf, fi,
 963                                                         start - key.offset);
 964                         btrfs_mark_buffer_dirty(leaf);
 965                         if (update_refs && disk_bytenr > 0)
 966                                 inode_sub_bytes(inode, extent_end - start);
 967                         if (end == extent_end)
 968                                 break;
 969
 970                         path->slots[0]++;
 971                         goto next_slot;
 972                 }
 973
 974                 /*
 975                  *  | ---- range to drop ----- |
 976                  *    | ------ extent ------ |
 977                  */
 978                 if (start <= key.offset && end >= extent_end) {
 979 delete_extent_item:
 980                         if (del_nr == 0) {
 981                                 del_slot = path->slots[0];
 982                                 del_nr = 1;
 983                         } else {
 984                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 985                                 del_nr++;
 986                         }
 987
 988                         if (update_refs &&
 989                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 990                                 inode_sub_bytes(inode,
 991                                                 extent_end - key.offset);
 992                                 extent_end = ALIGN(extent_end,
 993                                                    fs_info->sectorsize);
 994                         } else if (update_refs && disk_bytenr > 0) {
 995                                 btrfs_init_generic_ref(&ref,
 996                                                 BTRFS_DROP_DELAYED_REF,
 997                                                 disk_bytenr, num_bytes, 0);
 998                                 btrfs_init_data_ref(&ref,
 999                                                 root->root_key.objectid,
1000                                                 key.objectid,
1001                                                 key.offset - extent_offset);
1002                                 ret = btrfs_free_extent(trans, &ref);
1003                                 BUG_ON(ret); /* -ENOMEM */
1004                                 inode_sub_bytes(inode,
1005                                                 extent_end - key.offset);
1006                         }
1007
1008                         if (end == extent_end)
1009                                 break;
1010
1011                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1012                                 path->slots[0]++;
1013                                 goto next_slot;
1014                         }
1015
1016                         ret = btrfs_del_items(trans, root, path, del_slot,
1017                                               del_nr);
1018                         if (ret) {
1019                                 btrfs_abort_transaction(trans, ret);
1020                                 break;
1021                         }
1022
1023                         del_nr = 0;
1024                         del_slot = 0;
1025
1026                         btrfs_release_path(path);
1027                         continue;
1028                 }
1029
1030                 BUG();
1031         }
1032
1033         if (!ret && del_nr > 0) {
1034                 /*
1035                  * Set path->slots[0] to first slot, so that after the delete
1036                  * if items are move off from our leaf to its immediate left or
1037                  * right neighbor leafs, we end up with a correct and adjusted
1038                  * path->slots[0] for our insertion (if replace_extent != 0).
1039                  */
1040                 path->slots[0] = del_slot;
1041                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1042                 if (ret)
1043                         btrfs_abort_transaction(trans, ret);
1044         }
1045
1046         leaf = path->nodes[0];
1047         /*
1048          * If btrfs_del_items() was called, it might have deleted a leaf, in
1049          * which case it unlocked our path, so check path->locks[0] matches a
1050          * write lock.
1051          */
1052         if (!ret && replace_extent && leafs_visited == 1 &&
1053             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1054              path->locks[0] == BTRFS_WRITE_LOCK) &&
1055             btrfs_leaf_free_space(leaf) >=
1056             sizeof(struct btrfs_item) + extent_item_size) {
1057
1058                 key.objectid = ino;
1059                 key.type = BTRFS_EXTENT_DATA_KEY;
1060                 key.offset = start;
1061                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1062                         struct btrfs_key slot_key;
1063
1064                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1065                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1066                                 path->slots[0]++;
1067                 }
1068                 setup_items_for_insert(root, path, &key,
1069                                        &extent_item_size,
1070                                        extent_item_size,
1071                                        sizeof(struct btrfs_item) +
1072                                        extent_item_size, 1);
1073                 *key_inserted = 1;
1074         }
1075
1076         if (!replace_extent || !(*key_inserted))
1077                 btrfs_release_path(path);
1078         if (drop_end)
1079                 *drop_end = found ? min(end, last_end) : end;
1080         return ret;
1081 }
1082
1083 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1084                        struct btrfs_root *root, struct inode *inode, u64 start,
1085                        u64 end, int drop_cache)
1086 {
1087         struct btrfs_path *path;
1088         int ret;
1089
1090         path = btrfs_alloc_path();
1091         if (!path)
1092                 return -ENOMEM;
1093         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1094                                    drop_cache, 0, 0, NULL);
1095         btrfs_free_path(path);
1096         return ret;
1097 }
1098
1099 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1100                             u64 objectid, u64 bytenr, u64 orig_offset,
1101                             u64 *start, u64 *end)
1102 {
1103         struct btrfs_file_extent_item *fi;
1104         struct btrfs_key key;
1105         u64 extent_end;
1106
1107         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1108                 return 0;
1109
1110         btrfs_item_key_to_cpu(leaf, &key, slot);
1111         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1112                 return 0;
1113
1114         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1115         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1116             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1117             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1118             btrfs_file_extent_compression(leaf, fi) ||
1119             btrfs_file_extent_encryption(leaf, fi) ||
1120             btrfs_file_extent_other_encoding(leaf, fi))
1121                 return 0;
1122
1123         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1124         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1125                 return 0;
1126
1127         *start = key.offset;
1128         *end = extent_end;
1129         return 1;
1130 }
1131
1132 /*
1133  * Mark extent in the range start - end as written.
1134  *
1135  * This changes extent type from 'pre-allocated' to 'regular'. If only
1136  * part of extent is marked as written, the extent will be split into
1137  * two or three.
1138  */
1139 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1140                               struct btrfs_inode *inode, u64 start, u64 end)
1141 {
1142         struct btrfs_fs_info *fs_info = trans->fs_info;
1143         struct btrfs_root *root = inode->root;
1144         struct extent_buffer *leaf;
1145         struct btrfs_path *path;
1146         struct btrfs_file_extent_item *fi;
1147         struct btrfs_ref ref = { 0 };
1148         struct btrfs_key key;
1149         struct btrfs_key new_key;
1150         u64 bytenr;
1151         u64 num_bytes;
1152         u64 extent_end;
1153         u64 orig_offset;
1154         u64 other_start;
1155         u64 other_end;
1156         u64 split;
1157         int del_nr = 0;
1158         int del_slot = 0;
1159         int recow;
1160         int ret;
1161         u64 ino = btrfs_ino(inode);
1162
1163         path = btrfs_alloc_path();
1164         if (!path)
1165                 return -ENOMEM;
1166 again:
1167         recow = 0;
1168         split = start;
1169         key.objectid = ino;
1170         key.type = BTRFS_EXTENT_DATA_KEY;
1171         key.offset = split;
1172
1173         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1174         if (ret < 0)
1175                 goto out;
1176         if (ret > 0 && path->slots[0] > 0)
1177                 path->slots[0]--;
1178
1179         leaf = path->nodes[0];
1180         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1181         if (key.objectid != ino ||
1182             key.type != BTRFS_EXTENT_DATA_KEY) {
1183                 ret = -EINVAL;
1184                 btrfs_abort_transaction(trans, ret);
1185                 goto out;
1186         }
1187         fi = btrfs_item_ptr(leaf, path->slots[0],
1188                             struct btrfs_file_extent_item);
1189         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1190                 ret = -EINVAL;
1191                 btrfs_abort_transaction(trans, ret);
1192                 goto out;
1193         }
1194         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1195         if (key.offset > start || extent_end < end) {
1196                 ret = -EINVAL;
1197                 btrfs_abort_transaction(trans, ret);
1198                 goto out;
1199         }
1200
1201         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1202         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1203         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1204         memcpy(&new_key, &key, sizeof(new_key));
1205
1206         if (start == key.offset && end < extent_end) {
1207                 other_start = 0;
1208                 other_end = start;
1209                 if (extent_mergeable(leaf, path->slots[0] - 1,
1210                                      ino, bytenr, orig_offset,
1211                                      &other_start, &other_end)) {
1212                         new_key.offset = end;
1213                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1214                         fi = btrfs_item_ptr(leaf, path->slots[0],
1215                                             struct btrfs_file_extent_item);
1216                         btrfs_set_file_extent_generation(leaf, fi,
1217                                                          trans->transid);
1218                         btrfs_set_file_extent_num_bytes(leaf, fi,
1219                                                         extent_end - end);
1220                         btrfs_set_file_extent_offset(leaf, fi,
1221                                                      end - orig_offset);
1222                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1223                                             struct btrfs_file_extent_item);
1224                         btrfs_set_file_extent_generation(leaf, fi,
1225                                                          trans->transid);
1226                         btrfs_set_file_extent_num_bytes(leaf, fi,
1227                                                         end - other_start);
1228                         btrfs_mark_buffer_dirty(leaf);
1229                         goto out;
1230                 }
1231         }
1232
1233         if (start > key.offset && end == extent_end) {
1234                 other_start = end;
1235                 other_end = 0;
1236                 if (extent_mergeable(leaf, path->slots[0] + 1,
1237                                      ino, bytenr, orig_offset,
1238                                      &other_start, &other_end)) {
1239                         fi = btrfs_item_ptr(leaf, path->slots[0],
1240                                             struct btrfs_file_extent_item);
1241                         btrfs_set_file_extent_num_bytes(leaf, fi,
1242                                                         start - key.offset);
1243                         btrfs_set_file_extent_generation(leaf, fi,
1244                                                          trans->transid);
1245                         path->slots[0]++;
1246                         new_key.offset = start;
1247                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1248
1249                         fi = btrfs_item_ptr(leaf, path->slots[0],
1250                                             struct btrfs_file_extent_item);
1251                         btrfs_set_file_extent_generation(leaf, fi,
1252                                                          trans->transid);
1253                         btrfs_set_file_extent_num_bytes(leaf, fi,
1254                                                         other_end - start);
1255                         btrfs_set_file_extent_offset(leaf, fi,
1256                                                      start - orig_offset);
1257                         btrfs_mark_buffer_dirty(leaf);
1258                         goto out;
1259                 }
1260         }
1261
1262         while (start > key.offset || end < extent_end) {
1263                 if (key.offset == start)
1264                         split = end;
1265
1266                 new_key.offset = split;
1267                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1268                 if (ret == -EAGAIN) {
1269                         btrfs_release_path(path);
1270                         goto again;
1271                 }
1272                 if (ret < 0) {
1273                         btrfs_abort_transaction(trans, ret);
1274                         goto out;
1275                 }
1276
1277                 leaf = path->nodes[0];
1278                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1279                                     struct btrfs_file_extent_item);
1280                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1281                 btrfs_set_file_extent_num_bytes(leaf, fi,
1282                                                 split - key.offset);
1283
1284                 fi = btrfs_item_ptr(leaf, path->slots[0],
1285                                     struct btrfs_file_extent_item);
1286
1287                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1288                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1289                 btrfs_set_file_extent_num_bytes(leaf, fi,
1290                                                 extent_end - split);
1291                 btrfs_mark_buffer_dirty(leaf);
1292
1293                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1294                                        num_bytes, 0);
1295                 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1296                                     orig_offset);
1297                 ret = btrfs_inc_extent_ref(trans, &ref);
1298                 if (ret) {
1299                         btrfs_abort_transaction(trans, ret);
1300                         goto out;
1301                 }
1302
1303                 if (split == start) {
1304                         key.offset = start;
1305                 } else {
1306                         if (start != key.offset) {
1307                                 ret = -EINVAL;
1308                                 btrfs_abort_transaction(trans, ret);
1309                                 goto out;
1310                         }
1311                         path->slots[0]--;
1312                         extent_end = end;
1313                 }
1314                 recow = 1;
1315         }
1316
1317         other_start = end;
1318         other_end = 0;
1319         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1320                                num_bytes, 0);
1321         btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1322         if (extent_mergeable(leaf, path->slots[0] + 1,
1323                              ino, bytenr, orig_offset,
1324                              &other_start, &other_end)) {
1325                 if (recow) {
1326                         btrfs_release_path(path);
1327                         goto again;
1328                 }
1329                 extent_end = other_end;
1330                 del_slot = path->slots[0] + 1;
1331                 del_nr++;
1332                 ret = btrfs_free_extent(trans, &ref);
1333                 if (ret) {
1334                         btrfs_abort_transaction(trans, ret);
1335                         goto out;
1336                 }
1337         }
1338         other_start = 0;
1339         other_end = start;
1340         if (extent_mergeable(leaf, path->slots[0] - 1,
1341                              ino, bytenr, orig_offset,
1342                              &other_start, &other_end)) {
1343                 if (recow) {
1344                         btrfs_release_path(path);
1345                         goto again;
1346                 }
1347                 key.offset = other_start;
1348                 del_slot = path->slots[0];
1349                 del_nr++;
1350                 ret = btrfs_free_extent(trans, &ref);
1351                 if (ret) {
1352                         btrfs_abort_transaction(trans, ret);
1353                         goto out;
1354                 }
1355         }
1356         if (del_nr == 0) {
1357                 fi = btrfs_item_ptr(leaf, path->slots[0],
1358                            struct btrfs_file_extent_item);
1359                 btrfs_set_file_extent_type(leaf, fi,
1360                                            BTRFS_FILE_EXTENT_REG);
1361                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1362                 btrfs_mark_buffer_dirty(leaf);
1363         } else {
1364                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1365                            struct btrfs_file_extent_item);
1366                 btrfs_set_file_extent_type(leaf, fi,
1367                                            BTRFS_FILE_EXTENT_REG);
1368                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1369                 btrfs_set_file_extent_num_bytes(leaf, fi,
1370                                                 extent_end - key.offset);
1371                 btrfs_mark_buffer_dirty(leaf);
1372
1373                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1374                 if (ret < 0) {
1375                         btrfs_abort_transaction(trans, ret);
1376                         goto out;
1377                 }
1378         }
1379 out:
1380         btrfs_free_path(path);
1381         return 0;
1382 }
1383
1384 /*
1385  * on error we return an unlocked page and the error value
1386  * on success we return a locked page and 0
1387  */
1388 static int prepare_uptodate_page(struct inode *inode,
1389                                  struct page *page, u64 pos,
1390                                  bool force_uptodate)
1391 {
1392         int ret = 0;
1393
1394         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1395             !PageUptodate(page)) {
1396                 ret = btrfs_readpage(NULL, page);
1397                 if (ret)
1398                         return ret;
1399                 lock_page(page);
1400                 if (!PageUptodate(page)) {
1401                         unlock_page(page);
1402                         return -EIO;
1403                 }
1404                 if (page->mapping != inode->i_mapping) {
1405                         unlock_page(page);
1406                         return -EAGAIN;
1407                 }
1408         }
1409         return 0;
1410 }
1411
1412 /*
1413  * this just gets pages into the page cache and locks them down.
1414  */
1415 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1416                                   size_t num_pages, loff_t pos,
1417                                   size_t write_bytes, bool force_uptodate)
1418 {
1419         int i;
1420         unsigned long index = pos >> PAGE_SHIFT;
1421         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1422         int err = 0;
1423         int faili;
1424
1425         for (i = 0; i < num_pages; i++) {
1426 again:
1427                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1428                                                mask | __GFP_WRITE);
1429                 if (!pages[i]) {
1430                         faili = i - 1;
1431                         err = -ENOMEM;
1432                         goto fail;
1433                 }
1434
1435                 if (i == 0)
1436                         err = prepare_uptodate_page(inode, pages[i], pos,
1437                                                     force_uptodate);
1438                 if (!err && i == num_pages - 1)
1439                         err = prepare_uptodate_page(inode, pages[i],
1440                                                     pos + write_bytes, false);
1441                 if (err) {
1442                         put_page(pages[i]);
1443                         if (err == -EAGAIN) {
1444                                 err = 0;
1445                                 goto again;
1446                         }
1447                         faili = i - 1;
1448                         goto fail;
1449                 }
1450                 wait_on_page_writeback(pages[i]);
1451         }
1452
1453         return 0;
1454 fail:
1455         while (faili >= 0) {
1456                 unlock_page(pages[faili]);
1457                 put_page(pages[faili]);
1458                 faili--;
1459         }
1460         return err;
1461
1462 }
1463
1464 /*
1465  * This function locks the extent and properly waits for data=ordered extents
1466  * to finish before allowing the pages to be modified if need.
1467  *
1468  * The return value:
1469  * 1 - the extent is locked
1470  * 0 - the extent is not locked, and everything is OK
1471  * -EAGAIN - need re-prepare the pages
1472  * the other < 0 number - Something wrong happens
1473  */
1474 static noinline int
1475 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1476                                 size_t num_pages, loff_t pos,
1477                                 size_t write_bytes,
1478                                 u64 *lockstart, u64 *lockend,
1479                                 struct extent_state **cached_state)
1480 {
1481         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1482         u64 start_pos;
1483         u64 last_pos;
1484         int i;
1485         int ret = 0;
1486
1487         start_pos = round_down(pos, fs_info->sectorsize);
1488         last_pos = start_pos
1489                 + round_up(pos + write_bytes - start_pos,
1490                            fs_info->sectorsize) - 1;
1491
1492         if (start_pos < inode->vfs_inode.i_size) {
1493                 struct btrfs_ordered_extent *ordered;
1494
1495                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1496                                 cached_state);
1497                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1498                                                      last_pos - start_pos + 1);
1499                 if (ordered &&
1500                     ordered->file_offset + ordered->num_bytes > start_pos &&
1501                     ordered->file_offset <= last_pos) {
1502                         unlock_extent_cached(&inode->io_tree, start_pos,
1503                                         last_pos, cached_state);
1504                         for (i = 0; i < num_pages; i++) {
1505                                 unlock_page(pages[i]);
1506                                 put_page(pages[i]);
1507                         }
1508                         btrfs_start_ordered_extent(&inode->vfs_inode,
1509                                         ordered, 1);
1510                         btrfs_put_ordered_extent(ordered);
1511                         return -EAGAIN;
1512                 }
1513                 if (ordered)
1514                         btrfs_put_ordered_extent(ordered);
1515
1516                 *lockstart = start_pos;
1517                 *lockend = last_pos;
1518                 ret = 1;
1519         }
1520
1521         /*
1522          * It's possible the pages are dirty right now, but we don't want
1523          * to clean them yet because copy_from_user may catch a page fault
1524          * and we might have to fall back to one page at a time.  If that
1525          * happens, we'll unlock these pages and we'd have a window where
1526          * reclaim could sneak in and drop the once-dirty page on the floor
1527          * without writing it.
1528          *
1529          * We have the pages locked and the extent range locked, so there's
1530          * no way someone can start IO on any dirty pages in this range.
1531          *
1532          * We'll call btrfs_dirty_pages() later on, and that will flip around
1533          * delalloc bits and dirty the pages as required.
1534          */
1535         for (i = 0; i < num_pages; i++) {
1536                 set_page_extent_mapped(pages[i]);
1537                 WARN_ON(!PageLocked(pages[i]));
1538         }
1539
1540         return ret;
1541 }
1542
1543 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1544                                     size_t *write_bytes)
1545 {
1546         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1547         struct btrfs_root *root = inode->root;
1548         u64 lockstart, lockend;
1549         u64 num_bytes;
1550         int ret;
1551
1552         if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1553                 return -EAGAIN;
1554
1555         lockstart = round_down(pos, fs_info->sectorsize);
1556         lockend = round_up(pos + *write_bytes,
1557                            fs_info->sectorsize) - 1;
1558
1559         btrfs_lock_and_flush_ordered_range(inode, lockstart,
1560                                            lockend, NULL);
1561
1562         num_bytes = lockend - lockstart + 1;
1563         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1564                         NULL, NULL, NULL);
1565         if (ret <= 0) {
1566                 ret = 0;
1567                 btrfs_drew_write_unlock(&root->snapshot_lock);
1568         } else {
1569                 *write_bytes = min_t(size_t, *write_bytes ,
1570                                      num_bytes - pos + lockstart);
1571         }
1572
1573         unlock_extent(&inode->io_tree, lockstart, lockend);
1574
1575         return ret;
1576 }
1577
1578 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1579                                                struct iov_iter *i)
1580 {
1581         struct file *file = iocb->ki_filp;
1582         loff_t pos = iocb->ki_pos;
1583         struct inode *inode = file_inode(file);
1584         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1585         struct btrfs_root *root = BTRFS_I(inode)->root;
1586         struct page **pages = NULL;
1587         struct extent_changeset *data_reserved = NULL;
1588         u64 release_bytes = 0;
1589         u64 lockstart;
1590         u64 lockend;
1591         size_t num_written = 0;
1592         int nrptrs;
1593         int ret = 0;
1594         bool only_release_metadata = false;
1595         bool force_page_uptodate = false;
1596
1597         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1598                         PAGE_SIZE / (sizeof(struct page *)));
1599         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1600         nrptrs = max(nrptrs, 8);
1601         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1602         if (!pages)
1603                 return -ENOMEM;
1604
1605         while (iov_iter_count(i) > 0) {
1606                 struct extent_state *cached_state = NULL;
1607                 size_t offset = offset_in_page(pos);
1608                 size_t sector_offset;
1609                 size_t write_bytes = min(iov_iter_count(i),
1610                                          nrptrs * (size_t)PAGE_SIZE -
1611                                          offset);
1612                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1613                                                 PAGE_SIZE);
1614                 size_t reserve_bytes;
1615                 size_t dirty_pages;
1616                 size_t copied;
1617                 size_t dirty_sectors;
1618                 size_t num_sectors;
1619                 int extents_locked;
1620
1621                 WARN_ON(num_pages > nrptrs);
1622
1623                 /*
1624                  * Fault pages before locking them in prepare_pages
1625                  * to avoid recursive lock
1626                  */
1627                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1628                         ret = -EFAULT;
1629                         break;
1630                 }
1631
1632                 only_release_metadata = false;
1633                 sector_offset = pos & (fs_info->sectorsize - 1);
1634                 reserve_bytes = round_up(write_bytes + sector_offset,
1635                                 fs_info->sectorsize);
1636
1637                 extent_changeset_release(data_reserved);
1638                 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1639                                                   write_bytes);
1640                 if (ret < 0) {
1641                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1642                                                       BTRFS_INODE_PREALLOC)) &&
1643                             check_can_nocow(BTRFS_I(inode), pos,
1644                                         &write_bytes) > 0) {
1645                                 /*
1646                                  * For nodata cow case, no need to reserve
1647                                  * data space.
1648                                  */
1649                                 only_release_metadata = true;
1650                                 /*
1651                                  * our prealloc extent may be smaller than
1652                                  * write_bytes, so scale down.
1653                                  */
1654                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1655                                                          PAGE_SIZE);
1656                                 reserve_bytes = round_up(write_bytes +
1657                                                          sector_offset,
1658                                                          fs_info->sectorsize);
1659                         } else {
1660                                 break;
1661                         }
1662                 }
1663
1664                 WARN_ON(reserve_bytes == 0);
1665                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1666                                 reserve_bytes);
1667                 if (ret) {
1668                         if (!only_release_metadata)
1669                                 btrfs_free_reserved_data_space(inode,
1670                                                 data_reserved, pos,
1671                                                 write_bytes);
1672                         else
1673                                 btrfs_drew_write_unlock(&root->snapshot_lock);
1674                         break;
1675                 }
1676
1677                 release_bytes = reserve_bytes;
1678 again:
1679                 /*
1680                  * This is going to setup the pages array with the number of
1681                  * pages we want, so we don't really need to worry about the
1682                  * contents of pages from loop to loop
1683                  */
1684                 ret = prepare_pages(inode, pages, num_pages,
1685                                     pos, write_bytes,
1686                                     force_page_uptodate);
1687                 if (ret) {
1688                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1689                                                        reserve_bytes);
1690                         break;
1691                 }
1692
1693                 extents_locked = lock_and_cleanup_extent_if_need(
1694                                 BTRFS_I(inode), pages,
1695                                 num_pages, pos, write_bytes, &lockstart,
1696                                 &lockend, &cached_state);
1697                 if (extents_locked < 0) {
1698                         if (extents_locked == -EAGAIN)
1699                                 goto again;
1700                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1701                                                        reserve_bytes);
1702                         ret = extents_locked;
1703                         break;
1704                 }
1705
1706                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1707
1708                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1709                 dirty_sectors = round_up(copied + sector_offset,
1710                                         fs_info->sectorsize);
1711                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1712
1713                 /*
1714                  * if we have trouble faulting in the pages, fall
1715                  * back to one page at a time
1716                  */
1717                 if (copied < write_bytes)
1718                         nrptrs = 1;
1719
1720                 if (copied == 0) {
1721                         force_page_uptodate = true;
1722                         dirty_sectors = 0;
1723                         dirty_pages = 0;
1724                 } else {
1725                         force_page_uptodate = false;
1726                         dirty_pages = DIV_ROUND_UP(copied + offset,
1727                                                    PAGE_SIZE);
1728                 }
1729
1730                 if (num_sectors > dirty_sectors) {
1731                         /* release everything except the sectors we dirtied */
1732                         release_bytes -= dirty_sectors <<
1733                                                 fs_info->sb->s_blocksize_bits;
1734                         if (only_release_metadata) {
1735                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1736                                                         release_bytes, true);
1737                         } else {
1738                                 u64 __pos;
1739
1740                                 __pos = round_down(pos,
1741                                                    fs_info->sectorsize) +
1742                                         (dirty_pages << PAGE_SHIFT);
1743                                 btrfs_delalloc_release_space(inode,
1744                                                 data_reserved, __pos,
1745                                                 release_bytes, true);
1746                         }
1747                 }
1748
1749                 release_bytes = round_up(copied + sector_offset,
1750                                         fs_info->sectorsize);
1751
1752                 if (copied > 0)
1753                         ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1754                                                 pos, copied, &cached_state);
1755
1756                 /*
1757                  * If we have not locked the extent range, because the range's
1758                  * start offset is >= i_size, we might still have a non-NULL
1759                  * cached extent state, acquired while marking the extent range
1760                  * as delalloc through btrfs_dirty_pages(). Therefore free any
1761                  * possible cached extent state to avoid a memory leak.
1762                  */
1763                 if (extents_locked)
1764                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1765                                              lockstart, lockend, &cached_state);
1766                 else
1767                         free_extent_state(cached_state);
1768
1769                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1770                 if (ret) {
1771                         btrfs_drop_pages(pages, num_pages);
1772                         break;
1773                 }
1774
1775                 release_bytes = 0;
1776                 if (only_release_metadata)
1777                         btrfs_drew_write_unlock(&root->snapshot_lock);
1778
1779                 if (only_release_metadata && copied > 0) {
1780                         lockstart = round_down(pos,
1781                                                fs_info->sectorsize);
1782                         lockend = round_up(pos + copied,
1783                                            fs_info->sectorsize) - 1;
1784
1785                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1786                                        lockend, EXTENT_NORESERVE, NULL,
1787                                        NULL, GFP_NOFS);
1788                 }
1789
1790                 btrfs_drop_pages(pages, num_pages);
1791
1792                 cond_resched();
1793
1794                 balance_dirty_pages_ratelimited(inode->i_mapping);
1795                 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1796                         btrfs_btree_balance_dirty(fs_info);
1797
1798                 pos += copied;
1799                 num_written += copied;
1800         }
1801
1802         kfree(pages);
1803
1804         if (release_bytes) {
1805                 if (only_release_metadata) {
1806                         btrfs_drew_write_unlock(&root->snapshot_lock);
1807                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1808                                         release_bytes, true);
1809                 } else {
1810                         btrfs_delalloc_release_space(inode, data_reserved,
1811                                         round_down(pos, fs_info->sectorsize),
1812                                         release_bytes, true);
1813                 }
1814         }
1815
1816         extent_changeset_free(data_reserved);
1817         return num_written ? num_written : ret;
1818 }
1819
1820 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1821 {
1822         struct file *file = iocb->ki_filp;
1823         struct inode *inode = file_inode(file);
1824         loff_t pos;
1825         ssize_t written;
1826         ssize_t written_buffered;
1827         loff_t endbyte;
1828         int err;
1829
1830         written = generic_file_direct_write(iocb, from);
1831
1832         if (written < 0 || !iov_iter_count(from))
1833                 return written;
1834
1835         pos = iocb->ki_pos;
1836         written_buffered = btrfs_buffered_write(iocb, from);
1837         if (written_buffered < 0) {
1838                 err = written_buffered;
1839                 goto out;
1840         }
1841         /*
1842          * Ensure all data is persisted. We want the next direct IO read to be
1843          * able to read what was just written.
1844          */
1845         endbyte = pos + written_buffered - 1;
1846         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1847         if (err)
1848                 goto out;
1849         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1850         if (err)
1851                 goto out;
1852         written += written_buffered;
1853         iocb->ki_pos = pos + written_buffered;
1854         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1855                                  endbyte >> PAGE_SHIFT);
1856 out:
1857         return written ? written : err;
1858 }
1859
1860 static void update_time_for_write(struct inode *inode)
1861 {
1862         struct timespec64 now;
1863
1864         if (IS_NOCMTIME(inode))
1865                 return;
1866
1867         now = current_time(inode);
1868         if (!timespec64_equal(&inode->i_mtime, &now))
1869                 inode->i_mtime = now;
1870
1871         if (!timespec64_equal(&inode->i_ctime, &now))
1872                 inode->i_ctime = now;
1873
1874         if (IS_I_VERSION(inode))
1875                 inode_inc_iversion(inode);
1876 }
1877
1878 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1879                                     struct iov_iter *from)
1880 {
1881         struct file *file = iocb->ki_filp;
1882         struct inode *inode = file_inode(file);
1883         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1884         struct btrfs_root *root = BTRFS_I(inode)->root;
1885         u64 start_pos;
1886         u64 end_pos;
1887         ssize_t num_written = 0;
1888         const bool sync = iocb->ki_flags & IOCB_DSYNC;
1889         ssize_t err;
1890         loff_t pos;
1891         size_t count;
1892         loff_t oldsize;
1893         int clean_page = 0;
1894
1895         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1896             (iocb->ki_flags & IOCB_NOWAIT))
1897                 return -EOPNOTSUPP;
1898
1899         if (iocb->ki_flags & IOCB_NOWAIT) {
1900                 if (!inode_trylock(inode))
1901                         return -EAGAIN;
1902         } else {
1903                 inode_lock(inode);
1904         }
1905
1906         err = generic_write_checks(iocb, from);
1907         if (err <= 0) {
1908                 inode_unlock(inode);
1909                 return err;
1910         }
1911
1912         pos = iocb->ki_pos;
1913         count = iov_iter_count(from);
1914         if (iocb->ki_flags & IOCB_NOWAIT) {
1915                 /*
1916                  * We will allocate space in case nodatacow is not set,
1917                  * so bail
1918                  */
1919                 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1920                                               BTRFS_INODE_PREALLOC)) ||
1921                     check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1922                         inode_unlock(inode);
1923                         return -EAGAIN;
1924                 }
1925         }
1926
1927         current->backing_dev_info = inode_to_bdi(inode);
1928         err = file_remove_privs(file);
1929         if (err) {
1930                 inode_unlock(inode);
1931                 goto out;
1932         }
1933
1934         /*
1935          * If BTRFS flips readonly due to some impossible error
1936          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1937          * although we have opened a file as writable, we have
1938          * to stop this write operation to ensure FS consistency.
1939          */
1940         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1941                 inode_unlock(inode);
1942                 err = -EROFS;
1943                 goto out;
1944         }
1945
1946         /*
1947          * We reserve space for updating the inode when we reserve space for the
1948          * extent we are going to write, so we will enospc out there.  We don't
1949          * need to start yet another transaction to update the inode as we will
1950          * update the inode when we finish writing whatever data we write.
1951          */
1952         update_time_for_write(inode);
1953
1954         start_pos = round_down(pos, fs_info->sectorsize);
1955         oldsize = i_size_read(inode);
1956         if (start_pos > oldsize) {
1957                 /* Expand hole size to cover write data, preventing empty gap */
1958                 end_pos = round_up(pos + count,
1959                                    fs_info->sectorsize);
1960                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1961                 if (err) {
1962                         inode_unlock(inode);
1963                         goto out;
1964                 }
1965                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1966                         clean_page = 1;
1967         }
1968
1969         if (sync)
1970                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1971
1972         if (iocb->ki_flags & IOCB_DIRECT) {
1973                 num_written = __btrfs_direct_write(iocb, from);
1974         } else {
1975                 num_written = btrfs_buffered_write(iocb, from);
1976                 if (num_written > 0)
1977                         iocb->ki_pos = pos + num_written;
1978                 if (clean_page)
1979                         pagecache_isize_extended(inode, oldsize,
1980                                                 i_size_read(inode));
1981         }
1982
1983         inode_unlock(inode);
1984
1985         /*
1986          * We also have to set last_sub_trans to the current log transid,
1987          * otherwise subsequent syncs to a file that's been synced in this
1988          * transaction will appear to have already occurred.
1989          */
1990         spin_lock(&BTRFS_I(inode)->lock);
1991         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1992         spin_unlock(&BTRFS_I(inode)->lock);
1993         if (num_written > 0)
1994                 num_written = generic_write_sync(iocb, num_written);
1995
1996         if (sync)
1997                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1998 out:
1999         current->backing_dev_info = NULL;
2000         return num_written ? num_written : err;
2001 }
2002
2003 int btrfs_release_file(struct inode *inode, struct file *filp)
2004 {
2005         struct btrfs_file_private *private = filp->private_data;
2006
2007         if (private && private->filldir_buf)
2008                 kfree(private->filldir_buf);
2009         kfree(private);
2010         filp->private_data = NULL;
2011
2012         /*
2013          * ordered_data_close is set by setattr when we are about to truncate
2014          * a file from a non-zero size to a zero size.  This tries to
2015          * flush down new bytes that may have been written if the
2016          * application were using truncate to replace a file in place.
2017          */
2018         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2019                                &BTRFS_I(inode)->runtime_flags))
2020                         filemap_flush(inode->i_mapping);
2021         return 0;
2022 }
2023
2024 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2025 {
2026         int ret;
2027         struct blk_plug plug;
2028
2029         /*
2030          * This is only called in fsync, which would do synchronous writes, so
2031          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2032          * multiple disks using raid profile, a large IO can be split to
2033          * several segments of stripe length (currently 64K).
2034          */
2035         blk_start_plug(&plug);
2036         atomic_inc(&BTRFS_I(inode)->sync_writers);
2037         ret = btrfs_fdatawrite_range(inode, start, end);
2038         atomic_dec(&BTRFS_I(inode)->sync_writers);
2039         blk_finish_plug(&plug);
2040
2041         return ret;
2042 }
2043
2044 /*
2045  * fsync call for both files and directories.  This logs the inode into
2046  * the tree log instead of forcing full commits whenever possible.
2047  *
2048  * It needs to call filemap_fdatawait so that all ordered extent updates are
2049  * in the metadata btree are up to date for copying to the log.
2050  *
2051  * It drops the inode mutex before doing the tree log commit.  This is an
2052  * important optimization for directories because holding the mutex prevents
2053  * new operations on the dir while we write to disk.
2054  */
2055 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2056 {
2057         struct dentry *dentry = file_dentry(file);
2058         struct inode *inode = d_inode(dentry);
2059         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2060         struct btrfs_root *root = BTRFS_I(inode)->root;
2061         struct btrfs_trans_handle *trans;
2062         struct btrfs_log_ctx ctx;
2063         int ret = 0, err;
2064
2065         trace_btrfs_sync_file(file, datasync);
2066
2067         btrfs_init_log_ctx(&ctx, inode);
2068
2069         /*
2070          * Set the range to full if the NO_HOLES feature is not enabled.
2071          * This is to avoid missing file extent items representing holes after
2072          * replaying the log.
2073          */
2074         if (!btrfs_fs_incompat(fs_info, NO_HOLES)) {
2075                 start = 0;
2076                 end = LLONG_MAX;
2077         }
2078
2079         /*
2080          * We write the dirty pages in the range and wait until they complete
2081          * out of the ->i_mutex. If so, we can flush the dirty pages by
2082          * multi-task, and make the performance up.  See
2083          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2084          */
2085         ret = start_ordered_ops(inode, start, end);
2086         if (ret)
2087                 goto out;
2088
2089         inode_lock(inode);
2090
2091         /*
2092          * We take the dio_sem here because the tree log stuff can race with
2093          * lockless dio writes and get an extent map logged for an extent we
2094          * never waited on.  We need it this high up for lockdep reasons.
2095          */
2096         down_write(&BTRFS_I(inode)->dio_sem);
2097
2098         atomic_inc(&root->log_batch);
2099
2100         /*
2101          * If the inode needs a full sync, make sure we use a full range to
2102          * avoid log tree corruption, due to hole detection racing with ordered
2103          * extent completion for adjacent ranges and races between logging and
2104          * completion of ordered extents for adjancent ranges - both races
2105          * could lead to file extent items in the log with overlapping ranges.
2106          * Do this while holding the inode lock, to avoid races with other
2107          * tasks.
2108          */
2109         if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2110                      &BTRFS_I(inode)->runtime_flags)) {
2111                 start = 0;
2112                 end = LLONG_MAX;
2113         }
2114
2115         /*
2116          * Before we acquired the inode's lock, someone may have dirtied more
2117          * pages in the target range. We need to make sure that writeback for
2118          * any such pages does not start while we are logging the inode, because
2119          * if it does, any of the following might happen when we are not doing a
2120          * full inode sync:
2121          *
2122          * 1) We log an extent after its writeback finishes but before its
2123          *    checksums are added to the csum tree, leading to -EIO errors
2124          *    when attempting to read the extent after a log replay.
2125          *
2126          * 2) We can end up logging an extent before its writeback finishes.
2127          *    Therefore after the log replay we will have a file extent item
2128          *    pointing to an unwritten extent (and no data checksums as well).
2129          *
2130          * So trigger writeback for any eventual new dirty pages and then we
2131          * wait for all ordered extents to complete below.
2132          */
2133         ret = start_ordered_ops(inode, start, end);
2134         if (ret) {
2135                 up_write(&BTRFS_I(inode)->dio_sem);
2136                 inode_unlock(inode);
2137                 goto out;
2138         }
2139
2140         /*
2141          * We have to do this here to avoid the priority inversion of waiting on
2142          * IO of a lower priority task while holding a transaction open.
2143          *
2144          * Also, the range length can be represented by u64, we have to do the
2145          * typecasts to avoid signed overflow if it's [0, LLONG_MAX].
2146          */
2147         ret = btrfs_wait_ordered_range(inode, start, (u64)end - (u64)start + 1);
2148         if (ret) {
2149                 up_write(&BTRFS_I(inode)->dio_sem);
2150                 inode_unlock(inode);
2151                 goto out;
2152         }
2153         atomic_inc(&root->log_batch);
2154
2155         smp_mb();
2156         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2157             BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2158                 /*
2159                  * We've had everything committed since the last time we were
2160                  * modified so clear this flag in case it was set for whatever
2161                  * reason, it's no longer relevant.
2162                  */
2163                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2164                           &BTRFS_I(inode)->runtime_flags);
2165                 /*
2166                  * An ordered extent might have started before and completed
2167                  * already with io errors, in which case the inode was not
2168                  * updated and we end up here. So check the inode's mapping
2169                  * for any errors that might have happened since we last
2170                  * checked called fsync.
2171                  */
2172                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2173                 up_write(&BTRFS_I(inode)->dio_sem);
2174                 inode_unlock(inode);
2175                 goto out;
2176         }
2177
2178         /*
2179          * We use start here because we will need to wait on the IO to complete
2180          * in btrfs_sync_log, which could require joining a transaction (for
2181          * example checking cross references in the nocow path).  If we use join
2182          * here we could get into a situation where we're waiting on IO to
2183          * happen that is blocked on a transaction trying to commit.  With start
2184          * we inc the extwriter counter, so we wait for all extwriters to exit
2185          * before we start blocking joiners.  This comment is to keep somebody
2186          * from thinking they are super smart and changing this to
2187          * btrfs_join_transaction *cough*Josef*cough*.
2188          */
2189         trans = btrfs_start_transaction(root, 0);
2190         if (IS_ERR(trans)) {
2191                 ret = PTR_ERR(trans);
2192                 up_write(&BTRFS_I(inode)->dio_sem);
2193                 inode_unlock(inode);
2194                 goto out;
2195         }
2196
2197         ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2198         if (ret < 0) {
2199                 /* Fallthrough and commit/free transaction. */
2200                 ret = 1;
2201         }
2202
2203         /* we've logged all the items and now have a consistent
2204          * version of the file in the log.  It is possible that
2205          * someone will come in and modify the file, but that's
2206          * fine because the log is consistent on disk, and we
2207          * have references to all of the file's extents
2208          *
2209          * It is possible that someone will come in and log the
2210          * file again, but that will end up using the synchronization
2211          * inside btrfs_sync_log to keep things safe.
2212          */
2213         up_write(&BTRFS_I(inode)->dio_sem);
2214         inode_unlock(inode);
2215
2216         if (ret != BTRFS_NO_LOG_SYNC) {
2217                 if (!ret) {
2218                         ret = btrfs_sync_log(trans, root, &ctx);
2219                         if (!ret) {
2220                                 ret = btrfs_end_transaction(trans);
2221                                 goto out;
2222                         }
2223                 }
2224                 ret = btrfs_commit_transaction(trans);
2225         } else {
2226                 ret = btrfs_end_transaction(trans);
2227         }
2228 out:
2229         ASSERT(list_empty(&ctx.list));
2230         err = file_check_and_advance_wb_err(file);
2231         if (!ret)
2232                 ret = err;
2233         return ret > 0 ? -EIO : ret;
2234 }
2235
2236 static const struct vm_operations_struct btrfs_file_vm_ops = {
2237         .fault          = filemap_fault,
2238         .map_pages      = filemap_map_pages,
2239         .page_mkwrite   = btrfs_page_mkwrite,
2240 };
2241
2242 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2243 {
2244         struct address_space *mapping = filp->f_mapping;
2245
2246         if (!mapping->a_ops->readpage)
2247                 return -ENOEXEC;
2248
2249         file_accessed(filp);
2250         vma->vm_ops = &btrfs_file_vm_ops;
2251
2252         return 0;
2253 }
2254
2255 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2256                           int slot, u64 start, u64 end)
2257 {
2258         struct btrfs_file_extent_item *fi;
2259         struct btrfs_key key;
2260
2261         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2262                 return 0;
2263
2264         btrfs_item_key_to_cpu(leaf, &key, slot);
2265         if (key.objectid != btrfs_ino(inode) ||
2266             key.type != BTRFS_EXTENT_DATA_KEY)
2267                 return 0;
2268
2269         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2270
2271         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2272                 return 0;
2273
2274         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2275                 return 0;
2276
2277         if (key.offset == end)
2278                 return 1;
2279         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2280                 return 1;
2281         return 0;
2282 }
2283
2284 static int fill_holes(struct btrfs_trans_handle *trans,
2285                 struct btrfs_inode *inode,
2286                 struct btrfs_path *path, u64 offset, u64 end)
2287 {
2288         struct btrfs_fs_info *fs_info = trans->fs_info;
2289         struct btrfs_root *root = inode->root;
2290         struct extent_buffer *leaf;
2291         struct btrfs_file_extent_item *fi;
2292         struct extent_map *hole_em;
2293         struct extent_map_tree *em_tree = &inode->extent_tree;
2294         struct btrfs_key key;
2295         int ret;
2296
2297         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2298                 goto out;
2299
2300         key.objectid = btrfs_ino(inode);
2301         key.type = BTRFS_EXTENT_DATA_KEY;
2302         key.offset = offset;
2303
2304         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2305         if (ret <= 0) {
2306                 /*
2307                  * We should have dropped this offset, so if we find it then
2308                  * something has gone horribly wrong.
2309                  */
2310                 if (ret == 0)
2311                         ret = -EINVAL;
2312                 return ret;
2313         }
2314
2315         leaf = path->nodes[0];
2316         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2317                 u64 num_bytes;
2318
2319                 path->slots[0]--;
2320                 fi = btrfs_item_ptr(leaf, path->slots[0],
2321                                     struct btrfs_file_extent_item);
2322                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2323                         end - offset;
2324                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2325                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2326                 btrfs_set_file_extent_offset(leaf, fi, 0);
2327                 btrfs_mark_buffer_dirty(leaf);
2328                 goto out;
2329         }
2330
2331         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2332                 u64 num_bytes;
2333
2334                 key.offset = offset;
2335                 btrfs_set_item_key_safe(fs_info, path, &key);
2336                 fi = btrfs_item_ptr(leaf, path->slots[0],
2337                                     struct btrfs_file_extent_item);
2338                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2339                         offset;
2340                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2341                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2342                 btrfs_set_file_extent_offset(leaf, fi, 0);
2343                 btrfs_mark_buffer_dirty(leaf);
2344                 goto out;
2345         }
2346         btrfs_release_path(path);
2347
2348         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2349                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2350         if (ret)
2351                 return ret;
2352
2353 out:
2354         btrfs_release_path(path);
2355
2356         hole_em = alloc_extent_map();
2357         if (!hole_em) {
2358                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2359                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2360         } else {
2361                 hole_em->start = offset;
2362                 hole_em->len = end - offset;
2363                 hole_em->ram_bytes = hole_em->len;
2364                 hole_em->orig_start = offset;
2365
2366                 hole_em->block_start = EXTENT_MAP_HOLE;
2367                 hole_em->block_len = 0;
2368                 hole_em->orig_block_len = 0;
2369                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2370                 hole_em->generation = trans->transid;
2371
2372                 do {
2373                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2374                         write_lock(&em_tree->lock);
2375                         ret = add_extent_mapping(em_tree, hole_em, 1);
2376                         write_unlock(&em_tree->lock);
2377                 } while (ret == -EEXIST);
2378                 free_extent_map(hole_em);
2379                 if (ret)
2380                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2381                                         &inode->runtime_flags);
2382         }
2383
2384         return 0;
2385 }
2386
2387 /*
2388  * Find a hole extent on given inode and change start/len to the end of hole
2389  * extent.(hole/vacuum extent whose em->start <= start &&
2390  *         em->start + em->len > start)
2391  * When a hole extent is found, return 1 and modify start/len.
2392  */
2393 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2394 {
2395         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2396         struct extent_map *em;
2397         int ret = 0;
2398
2399         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2400                               round_down(*start, fs_info->sectorsize),
2401                               round_up(*len, fs_info->sectorsize));
2402         if (IS_ERR(em))
2403                 return PTR_ERR(em);
2404
2405         /* Hole or vacuum extent(only exists in no-hole mode) */
2406         if (em->block_start == EXTENT_MAP_HOLE) {
2407                 ret = 1;
2408                 *len = em->start + em->len > *start + *len ?
2409                        0 : *start + *len - em->start - em->len;
2410                 *start = em->start + em->len;
2411         }
2412         free_extent_map(em);
2413         return ret;
2414 }
2415
2416 static int btrfs_punch_hole_lock_range(struct inode *inode,
2417                                        const u64 lockstart,
2418                                        const u64 lockend,
2419                                        struct extent_state **cached_state)
2420 {
2421         while (1) {
2422                 struct btrfs_ordered_extent *ordered;
2423                 int ret;
2424
2425                 truncate_pagecache_range(inode, lockstart, lockend);
2426
2427                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2428                                  cached_state);
2429                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2430
2431                 /*
2432                  * We need to make sure we have no ordered extents in this range
2433                  * and nobody raced in and read a page in this range, if we did
2434                  * we need to try again.
2435                  */
2436                 if ((!ordered ||
2437                     (ordered->file_offset + ordered->num_bytes <= lockstart ||
2438                      ordered->file_offset > lockend)) &&
2439                      !filemap_range_has_page(inode->i_mapping,
2440                                              lockstart, lockend)) {
2441                         if (ordered)
2442                                 btrfs_put_ordered_extent(ordered);
2443                         break;
2444                 }
2445                 if (ordered)
2446                         btrfs_put_ordered_extent(ordered);
2447                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2448                                      lockend, cached_state);
2449                 ret = btrfs_wait_ordered_range(inode, lockstart,
2450                                                lockend - lockstart + 1);
2451                 if (ret)
2452                         return ret;
2453         }
2454         return 0;
2455 }
2456
2457 static int btrfs_insert_clone_extent(struct btrfs_trans_handle *trans,
2458                                      struct inode *inode,
2459                                      struct btrfs_path *path,
2460                                      struct btrfs_clone_extent_info *clone_info,
2461                                      const u64 clone_len)
2462 {
2463         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2464         struct btrfs_root *root = BTRFS_I(inode)->root;
2465         struct btrfs_file_extent_item *extent;
2466         struct extent_buffer *leaf;
2467         struct btrfs_key key;
2468         int slot;
2469         struct btrfs_ref ref = { 0 };
2470         u64 ref_offset;
2471         int ret;
2472
2473         if (clone_len == 0)
2474                 return 0;
2475
2476         if (clone_info->disk_offset == 0 &&
2477             btrfs_fs_incompat(fs_info, NO_HOLES))
2478                 return 0;
2479
2480         key.objectid = btrfs_ino(BTRFS_I(inode));
2481         key.type = BTRFS_EXTENT_DATA_KEY;
2482         key.offset = clone_info->file_offset;
2483         ret = btrfs_insert_empty_item(trans, root, path, &key,
2484                                       clone_info->item_size);
2485         if (ret)
2486                 return ret;
2487         leaf = path->nodes[0];
2488         slot = path->slots[0];
2489         write_extent_buffer(leaf, clone_info->extent_buf,
2490                             btrfs_item_ptr_offset(leaf, slot),
2491                             clone_info->item_size);
2492         extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2493         btrfs_set_file_extent_offset(leaf, extent, clone_info->data_offset);
2494         btrfs_set_file_extent_num_bytes(leaf, extent, clone_len);
2495         btrfs_mark_buffer_dirty(leaf);
2496         btrfs_release_path(path);
2497
2498         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
2499                         clone_info->file_offset, clone_len);
2500         if (ret)
2501                 return ret;
2502
2503         /* If it's a hole, nothing more needs to be done. */
2504         if (clone_info->disk_offset == 0)
2505                 return 0;
2506
2507         inode_add_bytes(inode, clone_len);
2508         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2509                                clone_info->disk_offset,
2510                                clone_info->disk_len, 0);
2511         ref_offset = clone_info->file_offset - clone_info->data_offset;
2512         btrfs_init_data_ref(&ref, root->root_key.objectid,
2513                             btrfs_ino(BTRFS_I(inode)), ref_offset);
2514         ret = btrfs_inc_extent_ref(trans, &ref);
2515
2516         return ret;
2517 }
2518
2519 /*
2520  * The respective range must have been previously locked, as well as the inode.
2521  * The end offset is inclusive (last byte of the range).
2522  * @clone_info is NULL for fallocate's hole punching and non-NULL for extent
2523  * cloning.
2524  * When cloning, we don't want to end up in a state where we dropped extents
2525  * without inserting a new one, so we must abort the transaction to avoid a
2526  * corruption.
2527  */
2528 int btrfs_punch_hole_range(struct inode *inode, struct btrfs_path *path,
2529                            const u64 start, const u64 end,
2530                            struct btrfs_clone_extent_info *clone_info,
2531                            struct btrfs_trans_handle **trans_out)
2532 {
2533         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2534         u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2535         u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2536         struct btrfs_root *root = BTRFS_I(inode)->root;
2537         struct btrfs_trans_handle *trans = NULL;
2538         struct btrfs_block_rsv *rsv;
2539         unsigned int rsv_count;
2540         u64 cur_offset;
2541         u64 drop_end;
2542         u64 len = end - start;
2543         int ret = 0;
2544
2545         if (end <= start)
2546                 return -EINVAL;
2547
2548         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2549         if (!rsv) {
2550                 ret = -ENOMEM;
2551                 goto out;
2552         }
2553         rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2554         rsv->failfast = 1;
2555
2556         /*
2557          * 1 - update the inode
2558          * 1 - removing the extents in the range
2559          * 1 - adding the hole extent if no_holes isn't set or if we are cloning
2560          *     an extent
2561          */
2562         if (!btrfs_fs_incompat(fs_info, NO_HOLES) || clone_info)
2563                 rsv_count = 3;
2564         else
2565                 rsv_count = 2;
2566
2567         trans = btrfs_start_transaction(root, rsv_count);
2568         if (IS_ERR(trans)) {
2569                 ret = PTR_ERR(trans);
2570                 trans = NULL;
2571                 goto out_free;
2572         }
2573
2574         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2575                                       min_size, false);
2576         BUG_ON(ret);
2577         trans->block_rsv = rsv;
2578
2579         cur_offset = start;
2580         while (cur_offset < end) {
2581                 ret = __btrfs_drop_extents(trans, root, inode, path,
2582                                            cur_offset, end + 1, &drop_end,
2583                                            1, 0, 0, NULL);
2584                 if (ret != -ENOSPC) {
2585                         /*
2586                          * When cloning we want to avoid transaction aborts when
2587                          * nothing was done and we are attempting to clone parts
2588                          * of inline extents, in such cases -EOPNOTSUPP is
2589                          * returned by __btrfs_drop_extents() without having
2590                          * changed anything in the file.
2591                          */
2592                         if (clone_info && ret && ret != -EOPNOTSUPP)
2593                                 btrfs_abort_transaction(trans, ret);
2594                         break;
2595                 }
2596
2597                 trans->block_rsv = &fs_info->trans_block_rsv;
2598
2599                 if (!clone_info && cur_offset < drop_end &&
2600                     cur_offset < ino_size) {
2601                         ret = fill_holes(trans, BTRFS_I(inode), path,
2602                                         cur_offset, drop_end);
2603                         if (ret) {
2604                                 /*
2605                                  * If we failed then we didn't insert our hole
2606                                  * entries for the area we dropped, so now the
2607                                  * fs is corrupted, so we must abort the
2608                                  * transaction.
2609                                  */
2610                                 btrfs_abort_transaction(trans, ret);
2611                                 break;
2612                         }
2613                 } else if (!clone_info && cur_offset < drop_end) {
2614                         /*
2615                          * We are past the i_size here, but since we didn't
2616                          * insert holes we need to clear the mapped area so we
2617                          * know to not set disk_i_size in this area until a new
2618                          * file extent is inserted here.
2619                          */
2620                         ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2621                                         cur_offset, drop_end - cur_offset);
2622                         if (ret) {
2623                                 /*
2624                                  * We couldn't clear our area, so we could
2625                                  * presumably adjust up and corrupt the fs, so
2626                                  * we need to abort.
2627                                  */
2628                                 btrfs_abort_transaction(trans, ret);
2629                                 break;
2630                         }
2631                 }
2632
2633                 if (clone_info && drop_end > clone_info->file_offset) {
2634                         u64 clone_len = drop_end - clone_info->file_offset;
2635
2636                         ret = btrfs_insert_clone_extent(trans, inode, path,
2637                                                         clone_info, clone_len);
2638                         if (ret) {
2639                                 btrfs_abort_transaction(trans, ret);
2640                                 break;
2641                         }
2642                         clone_info->data_len -= clone_len;
2643                         clone_info->data_offset += clone_len;
2644                         clone_info->file_offset += clone_len;
2645                 }
2646
2647                 cur_offset = drop_end;
2648
2649                 ret = btrfs_update_inode(trans, root, inode);
2650                 if (ret)
2651                         break;
2652
2653                 btrfs_end_transaction(trans);
2654                 btrfs_btree_balance_dirty(fs_info);
2655
2656                 trans = btrfs_start_transaction(root, rsv_count);
2657                 if (IS_ERR(trans)) {
2658                         ret = PTR_ERR(trans);
2659                         trans = NULL;
2660                         break;
2661                 }
2662
2663                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2664                                               rsv, min_size, false);
2665                 BUG_ON(ret);    /* shouldn't happen */
2666                 trans->block_rsv = rsv;
2667
2668                 if (!clone_info) {
2669                         ret = find_first_non_hole(inode, &cur_offset, &len);
2670                         if (unlikely(ret < 0))
2671                                 break;
2672                         if (ret && !len) {
2673                                 ret = 0;
2674                                 break;
2675                         }
2676                 }
2677         }
2678
2679         /*
2680          * If we were cloning, force the next fsync to be a full one since we
2681          * we replaced (or just dropped in the case of cloning holes when
2682          * NO_HOLES is enabled) extents and extent maps.
2683          * This is for the sake of simplicity, and cloning into files larger
2684          * than 16Mb would force the full fsync any way (when
2685          * try_release_extent_mapping() is invoked during page cache truncation.
2686          */
2687         if (clone_info)
2688                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2689                         &BTRFS_I(inode)->runtime_flags);
2690
2691         if (ret)
2692                 goto out_trans;
2693
2694         trans->block_rsv = &fs_info->trans_block_rsv;
2695         /*
2696          * If we are using the NO_HOLES feature we might have had already an
2697          * hole that overlaps a part of the region [lockstart, lockend] and
2698          * ends at (or beyond) lockend. Since we have no file extent items to
2699          * represent holes, drop_end can be less than lockend and so we must
2700          * make sure we have an extent map representing the existing hole (the
2701          * call to __btrfs_drop_extents() might have dropped the existing extent
2702          * map representing the existing hole), otherwise the fast fsync path
2703          * will not record the existence of the hole region
2704          * [existing_hole_start, lockend].
2705          */
2706         if (drop_end <= end)
2707                 drop_end = end + 1;
2708         /*
2709          * Don't insert file hole extent item if it's for a range beyond eof
2710          * (because it's useless) or if it represents a 0 bytes range (when
2711          * cur_offset == drop_end).
2712          */
2713         if (!clone_info && cur_offset < ino_size && cur_offset < drop_end) {
2714                 ret = fill_holes(trans, BTRFS_I(inode), path,
2715                                 cur_offset, drop_end);
2716                 if (ret) {
2717                         /* Same comment as above. */
2718                         btrfs_abort_transaction(trans, ret);
2719                         goto out_trans;
2720                 }
2721         } else if (!clone_info && cur_offset < drop_end) {
2722                 /* See the comment in the loop above for the reasoning here. */
2723                 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2724                                         cur_offset, drop_end - cur_offset);
2725                 if (ret) {
2726                         btrfs_abort_transaction(trans, ret);
2727                         goto out_trans;
2728                 }
2729
2730         }
2731         if (clone_info) {
2732                 ret = btrfs_insert_clone_extent(trans, inode, path, clone_info,
2733                                                 clone_info->data_len);
2734                 if (ret) {
2735                         btrfs_abort_transaction(trans, ret);
2736                         goto out_trans;
2737                 }
2738         }
2739
2740 out_trans:
2741         if (!trans)
2742                 goto out_free;
2743
2744         trans->block_rsv = &fs_info->trans_block_rsv;
2745         if (ret)
2746                 btrfs_end_transaction(trans);
2747         else
2748                 *trans_out = trans;
2749 out_free:
2750         btrfs_free_block_rsv(fs_info, rsv);
2751 out:
2752         return ret;
2753 }
2754
2755 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2756 {
2757         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2758         struct btrfs_root *root = BTRFS_I(inode)->root;
2759         struct extent_state *cached_state = NULL;
2760         struct btrfs_path *path;
2761         struct btrfs_trans_handle *trans = NULL;
2762         u64 lockstart;
2763         u64 lockend;
2764         u64 tail_start;
2765         u64 tail_len;
2766         u64 orig_start = offset;
2767         int ret = 0;
2768         bool same_block;
2769         u64 ino_size;
2770         bool truncated_block = false;
2771         bool updated_inode = false;
2772
2773         ret = btrfs_wait_ordered_range(inode, offset, len);
2774         if (ret)
2775                 return ret;
2776
2777         inode_lock(inode);
2778         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2779         ret = find_first_non_hole(inode, &offset, &len);
2780         if (ret < 0)
2781                 goto out_only_mutex;
2782         if (ret && !len) {
2783                 /* Already in a large hole */
2784                 ret = 0;
2785                 goto out_only_mutex;
2786         }
2787
2788         lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2789         lockend = round_down(offset + len,
2790                              btrfs_inode_sectorsize(inode)) - 1;
2791         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2792                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2793         /*
2794          * We needn't truncate any block which is beyond the end of the file
2795          * because we are sure there is no data there.
2796          */
2797         /*
2798          * Only do this if we are in the same block and we aren't doing the
2799          * entire block.
2800          */
2801         if (same_block && len < fs_info->sectorsize) {
2802                 if (offset < ino_size) {
2803                         truncated_block = true;
2804                         ret = btrfs_truncate_block(inode, offset, len, 0);
2805                 } else {
2806                         ret = 0;
2807                 }
2808                 goto out_only_mutex;
2809         }
2810
2811         /* zero back part of the first block */
2812         if (offset < ino_size) {
2813                 truncated_block = true;
2814                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2815                 if (ret) {
2816                         inode_unlock(inode);
2817                         return ret;
2818                 }
2819         }
2820
2821         /* Check the aligned pages after the first unaligned page,
2822          * if offset != orig_start, which means the first unaligned page
2823          * including several following pages are already in holes,
2824          * the extra check can be skipped */
2825         if (offset == orig_start) {
2826                 /* after truncate page, check hole again */
2827                 len = offset + len - lockstart;
2828                 offset = lockstart;
2829                 ret = find_first_non_hole(inode, &offset, &len);
2830                 if (ret < 0)
2831                         goto out_only_mutex;
2832                 if (ret && !len) {
2833                         ret = 0;
2834                         goto out_only_mutex;
2835                 }
2836                 lockstart = offset;
2837         }
2838
2839         /* Check the tail unaligned part is in a hole */
2840         tail_start = lockend + 1;
2841         tail_len = offset + len - tail_start;
2842         if (tail_len) {
2843                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2844                 if (unlikely(ret < 0))
2845                         goto out_only_mutex;
2846                 if (!ret) {
2847                         /* zero the front end of the last page */
2848                         if (tail_start + tail_len < ino_size) {
2849                                 truncated_block = true;
2850                                 ret = btrfs_truncate_block(inode,
2851                                                         tail_start + tail_len,
2852                                                         0, 1);
2853                                 if (ret)
2854                                         goto out_only_mutex;
2855                         }
2856                 }
2857         }
2858
2859         if (lockend < lockstart) {
2860                 ret = 0;
2861                 goto out_only_mutex;
2862         }
2863
2864         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2865                                           &cached_state);
2866         if (ret)
2867                 goto out_only_mutex;
2868
2869         path = btrfs_alloc_path();
2870         if (!path) {
2871                 ret = -ENOMEM;
2872                 goto out;
2873         }
2874
2875         ret = btrfs_punch_hole_range(inode, path, lockstart, lockend, NULL,
2876                                      &trans);
2877         btrfs_free_path(path);
2878         if (ret)
2879                 goto out;
2880
2881         ASSERT(trans != NULL);
2882         inode_inc_iversion(inode);
2883         inode->i_mtime = inode->i_ctime = current_time(inode);
2884         ret = btrfs_update_inode(trans, root, inode);
2885         updated_inode = true;
2886         btrfs_end_transaction(trans);
2887         btrfs_btree_balance_dirty(fs_info);
2888 out:
2889         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2890                              &cached_state);
2891 out_only_mutex:
2892         if (!updated_inode && truncated_block && !ret) {
2893                 /*
2894                  * If we only end up zeroing part of a page, we still need to
2895                  * update the inode item, so that all the time fields are
2896                  * updated as well as the necessary btrfs inode in memory fields
2897                  * for detecting, at fsync time, if the inode isn't yet in the
2898                  * log tree or it's there but not up to date.
2899                  */
2900                 struct timespec64 now = current_time(inode);
2901
2902                 inode_inc_iversion(inode);
2903                 inode->i_mtime = now;
2904                 inode->i_ctime = now;
2905                 trans = btrfs_start_transaction(root, 1);
2906                 if (IS_ERR(trans)) {
2907                         ret = PTR_ERR(trans);
2908                 } else {
2909                         int ret2;
2910
2911                         ret = btrfs_update_inode(trans, root, inode);
2912                         ret2 = btrfs_end_transaction(trans);
2913                         if (!ret)
2914                                 ret = ret2;
2915                 }
2916         }
2917         inode_unlock(inode);
2918         return ret;
2919 }
2920
2921 /* Helper structure to record which range is already reserved */
2922 struct falloc_range {
2923         struct list_head list;
2924         u64 start;
2925         u64 len;
2926 };
2927
2928 /*
2929  * Helper function to add falloc range
2930  *
2931  * Caller should have locked the larger range of extent containing
2932  * [start, len)
2933  */
2934 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2935 {
2936         struct falloc_range *prev = NULL;
2937         struct falloc_range *range = NULL;
2938
2939         if (list_empty(head))
2940                 goto insert;
2941
2942         /*
2943          * As fallocate iterate by bytenr order, we only need to check
2944          * the last range.
2945          */
2946         prev = list_entry(head->prev, struct falloc_range, list);
2947         if (prev->start + prev->len == start) {
2948                 prev->len += len;
2949                 return 0;
2950         }
2951 insert:
2952         range = kmalloc(sizeof(*range), GFP_KERNEL);
2953         if (!range)
2954                 return -ENOMEM;
2955         range->start = start;
2956         range->len = len;
2957         list_add_tail(&range->list, head);
2958         return 0;
2959 }
2960
2961 static int btrfs_fallocate_update_isize(struct inode *inode,
2962                                         const u64 end,
2963                                         const int mode)
2964 {
2965         struct btrfs_trans_handle *trans;
2966         struct btrfs_root *root = BTRFS_I(inode)->root;
2967         int ret;
2968         int ret2;
2969
2970         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2971                 return 0;
2972
2973         trans = btrfs_start_transaction(root, 1);
2974         if (IS_ERR(trans))
2975                 return PTR_ERR(trans);
2976
2977         inode->i_ctime = current_time(inode);
2978         i_size_write(inode, end);
2979         btrfs_inode_safe_disk_i_size_write(inode, 0);
2980         ret = btrfs_update_inode(trans, root, inode);
2981         ret2 = btrfs_end_transaction(trans);
2982
2983         return ret ? ret : ret2;
2984 }
2985
2986 enum {
2987         RANGE_BOUNDARY_WRITTEN_EXTENT,
2988         RANGE_BOUNDARY_PREALLOC_EXTENT,
2989         RANGE_BOUNDARY_HOLE,
2990 };
2991
2992 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2993                                                  u64 offset)
2994 {
2995         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2996         struct extent_map *em;
2997         int ret;
2998
2999         offset = round_down(offset, sectorsize);
3000         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize);
3001         if (IS_ERR(em))
3002                 return PTR_ERR(em);
3003
3004         if (em->block_start == EXTENT_MAP_HOLE)
3005                 ret = RANGE_BOUNDARY_HOLE;
3006         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3007                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3008         else
3009                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3010
3011         free_extent_map(em);
3012         return ret;
3013 }
3014
3015 static int btrfs_zero_range(struct inode *inode,
3016                             loff_t offset,
3017                             loff_t len,
3018                             const int mode)
3019 {
3020         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3021         struct extent_map *em;
3022         struct extent_changeset *data_reserved = NULL;
3023         int ret;
3024         u64 alloc_hint = 0;
3025         const u64 sectorsize = btrfs_inode_sectorsize(inode);
3026         u64 alloc_start = round_down(offset, sectorsize);
3027         u64 alloc_end = round_up(offset + len, sectorsize);
3028         u64 bytes_to_reserve = 0;
3029         bool space_reserved = false;
3030
3031         inode_dio_wait(inode);
3032
3033         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3034                               alloc_end - alloc_start);
3035         if (IS_ERR(em)) {
3036                 ret = PTR_ERR(em);
3037                 goto out;
3038         }
3039
3040         /*
3041          * Avoid hole punching and extent allocation for some cases. More cases
3042          * could be considered, but these are unlikely common and we keep things
3043          * as simple as possible for now. Also, intentionally, if the target
3044          * range contains one or more prealloc extents together with regular
3045          * extents and holes, we drop all the existing extents and allocate a
3046          * new prealloc extent, so that we get a larger contiguous disk extent.
3047          */
3048         if (em->start <= alloc_start &&
3049             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3050                 const u64 em_end = em->start + em->len;
3051
3052                 if (em_end >= offset + len) {
3053                         /*
3054                          * The whole range is already a prealloc extent,
3055                          * do nothing except updating the inode's i_size if
3056                          * needed.
3057                          */
3058                         free_extent_map(em);
3059                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3060                                                            mode);
3061                         goto out;
3062                 }
3063                 /*
3064                  * Part of the range is already a prealloc extent, so operate
3065                  * only on the remaining part of the range.
3066                  */
3067                 alloc_start = em_end;
3068                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3069                 len = offset + len - alloc_start;
3070                 offset = alloc_start;
3071                 alloc_hint = em->block_start + em->len;
3072         }
3073         free_extent_map(em);
3074
3075         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3076             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3077                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3078                                       sectorsize);
3079                 if (IS_ERR(em)) {
3080                         ret = PTR_ERR(em);
3081                         goto out;
3082                 }
3083
3084                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3085                         free_extent_map(em);
3086                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3087                                                            mode);
3088                         goto out;
3089                 }
3090                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3091                         free_extent_map(em);
3092                         ret = btrfs_truncate_block(inode, offset, len, 0);
3093                         if (!ret)
3094                                 ret = btrfs_fallocate_update_isize(inode,
3095                                                                    offset + len,
3096                                                                    mode);
3097                         return ret;
3098                 }
3099                 free_extent_map(em);
3100                 alloc_start = round_down(offset, sectorsize);
3101                 alloc_end = alloc_start + sectorsize;
3102                 goto reserve_space;
3103         }
3104
3105         alloc_start = round_up(offset, sectorsize);
3106         alloc_end = round_down(offset + len, sectorsize);
3107
3108         /*
3109          * For unaligned ranges, check the pages at the boundaries, they might
3110          * map to an extent, in which case we need to partially zero them, or
3111          * they might map to a hole, in which case we need our allocation range
3112          * to cover them.
3113          */
3114         if (!IS_ALIGNED(offset, sectorsize)) {
3115                 ret = btrfs_zero_range_check_range_boundary(inode, offset);
3116                 if (ret < 0)
3117                         goto out;
3118                 if (ret == RANGE_BOUNDARY_HOLE) {
3119                         alloc_start = round_down(offset, sectorsize);
3120                         ret = 0;
3121                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3122                         ret = btrfs_truncate_block(inode, offset, 0, 0);
3123                         if (ret)
3124                                 goto out;
3125                 } else {
3126                         ret = 0;
3127                 }
3128         }
3129
3130         if (!IS_ALIGNED(offset + len, sectorsize)) {
3131                 ret = btrfs_zero_range_check_range_boundary(inode,
3132                                                             offset + len);
3133                 if (ret < 0)
3134                         goto out;
3135                 if (ret == RANGE_BOUNDARY_HOLE) {
3136                         alloc_end = round_up(offset + len, sectorsize);
3137                         ret = 0;
3138                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3139                         ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3140                         if (ret)
3141                                 goto out;
3142                 } else {
3143                         ret = 0;
3144                 }
3145         }
3146
3147 reserve_space:
3148         if (alloc_start < alloc_end) {
3149                 struct extent_state *cached_state = NULL;
3150                 const u64 lockstart = alloc_start;
3151                 const u64 lockend = alloc_end - 1;
3152
3153                 bytes_to_reserve = alloc_end - alloc_start;
3154                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3155                                                       bytes_to_reserve);
3156                 if (ret < 0)
3157                         goto out;
3158                 space_reserved = true;
3159                 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3160                                                 alloc_start, bytes_to_reserve);
3161                 if (ret)
3162                         goto out;
3163                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3164                                                   &cached_state);
3165                 if (ret)
3166                         goto out;
3167                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3168                                                 alloc_end - alloc_start,
3169                                                 i_blocksize(inode),
3170                                                 offset + len, &alloc_hint);
3171                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3172                                      lockend, &cached_state);
3173                 /* btrfs_prealloc_file_range releases reserved space on error */
3174                 if (ret) {
3175                         space_reserved = false;
3176                         goto out;
3177                 }
3178         }
3179         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3180  out:
3181         if (ret && space_reserved)
3182                 btrfs_free_reserved_data_space(inode, data_reserved,
3183                                                alloc_start, bytes_to_reserve);
3184         extent_changeset_free(data_reserved);
3185
3186         return ret;
3187 }
3188
3189 static long btrfs_fallocate(struct file *file, int mode,
3190                             loff_t offset, loff_t len)
3191 {
3192         struct inode *inode = file_inode(file);
3193         struct extent_state *cached_state = NULL;
3194         struct extent_changeset *data_reserved = NULL;
3195         struct falloc_range *range;
3196         struct falloc_range *tmp;
3197         struct list_head reserve_list;
3198         u64 cur_offset;
3199         u64 last_byte;
3200         u64 alloc_start;
3201         u64 alloc_end;
3202         u64 alloc_hint = 0;
3203         u64 locked_end;
3204         u64 actual_end = 0;
3205         struct extent_map *em;
3206         int blocksize = btrfs_inode_sectorsize(inode);
3207         int ret;
3208
3209         alloc_start = round_down(offset, blocksize);
3210         alloc_end = round_up(offset + len, blocksize);
3211         cur_offset = alloc_start;
3212
3213         /* Make sure we aren't being give some crap mode */
3214         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3215                      FALLOC_FL_ZERO_RANGE))
3216                 return -EOPNOTSUPP;
3217
3218         if (mode & FALLOC_FL_PUNCH_HOLE)
3219                 return btrfs_punch_hole(inode, offset, len);
3220
3221         /*
3222          * Only trigger disk allocation, don't trigger qgroup reserve
3223          *
3224          * For qgroup space, it will be checked later.
3225          */
3226         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3227                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3228                                                       alloc_end - alloc_start);
3229                 if (ret < 0)
3230                         return ret;
3231         }
3232
3233         inode_lock(inode);
3234
3235         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3236                 ret = inode_newsize_ok(inode, offset + len);
3237                 if (ret)
3238                         goto out;
3239         }
3240
3241         /*
3242          * TODO: Move these two operations after we have checked
3243          * accurate reserved space, or fallocate can still fail but
3244          * with page truncated or size expanded.
3245          *
3246          * But that's a minor problem and won't do much harm BTW.
3247          */
3248         if (alloc_start > inode->i_size) {
3249                 ret = btrfs_cont_expand(inode, i_size_read(inode),
3250                                         alloc_start);
3251                 if (ret)
3252                         goto out;
3253         } else if (offset + len > inode->i_size) {
3254                 /*
3255                  * If we are fallocating from the end of the file onward we
3256                  * need to zero out the end of the block if i_size lands in the
3257                  * middle of a block.
3258                  */
3259                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3260                 if (ret)
3261                         goto out;
3262         }
3263
3264         /*
3265          * wait for ordered IO before we have any locks.  We'll loop again
3266          * below with the locks held.
3267          */
3268         ret = btrfs_wait_ordered_range(inode, alloc_start,
3269                                        alloc_end - alloc_start);
3270         if (ret)
3271                 goto out;
3272
3273         if (mode & FALLOC_FL_ZERO_RANGE) {
3274                 ret = btrfs_zero_range(inode, offset, len, mode);
3275                 inode_unlock(inode);
3276                 return ret;
3277         }
3278
3279         locked_end = alloc_end - 1;
3280         while (1) {
3281                 struct btrfs_ordered_extent *ordered;
3282
3283                 /* the extent lock is ordered inside the running
3284                  * transaction
3285                  */
3286                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3287                                  locked_end, &cached_state);
3288                 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3289
3290                 if (ordered &&
3291                     ordered->file_offset + ordered->num_bytes > alloc_start &&
3292                     ordered->file_offset < alloc_end) {
3293                         btrfs_put_ordered_extent(ordered);
3294                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3295                                              alloc_start, locked_end,
3296                                              &cached_state);
3297                         /*
3298                          * we can't wait on the range with the transaction
3299                          * running or with the extent lock held
3300                          */
3301                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3302                                                        alloc_end - alloc_start);
3303                         if (ret)
3304                                 goto out;
3305                 } else {
3306                         if (ordered)
3307                                 btrfs_put_ordered_extent(ordered);
3308                         break;
3309                 }
3310         }
3311
3312         /* First, check if we exceed the qgroup limit */
3313         INIT_LIST_HEAD(&reserve_list);
3314         while (cur_offset < alloc_end) {
3315                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3316                                       alloc_end - cur_offset);
3317                 if (IS_ERR(em)) {
3318                         ret = PTR_ERR(em);
3319                         break;
3320                 }
3321                 last_byte = min(extent_map_end(em), alloc_end);
3322                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3323                 last_byte = ALIGN(last_byte, blocksize);
3324                 if (em->block_start == EXTENT_MAP_HOLE ||
3325                     (cur_offset >= inode->i_size &&
3326                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3327                         ret = add_falloc_range(&reserve_list, cur_offset,
3328                                                last_byte - cur_offset);
3329                         if (ret < 0) {
3330                                 free_extent_map(em);
3331                                 break;
3332                         }
3333                         ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3334                                         cur_offset, last_byte - cur_offset);
3335                         if (ret < 0) {
3336                                 cur_offset = last_byte;
3337                                 free_extent_map(em);
3338                                 break;
3339                         }
3340                 } else {
3341                         /*
3342                          * Do not need to reserve unwritten extent for this
3343                          * range, free reserved data space first, otherwise
3344                          * it'll result in false ENOSPC error.
3345                          */
3346                         btrfs_free_reserved_data_space(inode, data_reserved,
3347                                         cur_offset, last_byte - cur_offset);
3348                 }
3349                 free_extent_map(em);
3350                 cur_offset = last_byte;
3351         }
3352
3353         /*
3354          * If ret is still 0, means we're OK to fallocate.
3355          * Or just cleanup the list and exit.
3356          */
3357         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3358                 if (!ret)
3359                         ret = btrfs_prealloc_file_range(inode, mode,
3360                                         range->start,
3361                                         range->len, i_blocksize(inode),
3362                                         offset + len, &alloc_hint);
3363                 else
3364                         btrfs_free_reserved_data_space(inode,
3365                                         data_reserved, range->start,
3366                                         range->len);
3367                 list_del(&range->list);
3368                 kfree(range);
3369         }
3370         if (ret < 0)
3371                 goto out_unlock;
3372
3373         /*
3374          * We didn't need to allocate any more space, but we still extended the
3375          * size of the file so we need to update i_size and the inode item.
3376          */
3377         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3378 out_unlock:
3379         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3380                              &cached_state);
3381 out:
3382         inode_unlock(inode);
3383         /* Let go of our reservation. */
3384         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3385                 btrfs_free_reserved_data_space(inode, data_reserved,
3386                                 cur_offset, alloc_end - cur_offset);
3387         extent_changeset_free(data_reserved);
3388         return ret;
3389 }
3390
3391 static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3392                                   int whence)
3393 {
3394         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3395         struct extent_map *em = NULL;
3396         struct extent_state *cached_state = NULL;
3397         loff_t i_size = inode->i_size;
3398         u64 lockstart;
3399         u64 lockend;
3400         u64 start;
3401         u64 len;
3402         int ret = 0;
3403
3404         if (i_size == 0 || offset >= i_size)
3405                 return -ENXIO;
3406
3407         /*
3408          * offset can be negative, in this case we start finding DATA/HOLE from
3409          * the very start of the file.
3410          */
3411         start = max_t(loff_t, 0, offset);
3412
3413         lockstart = round_down(start, fs_info->sectorsize);
3414         lockend = round_up(i_size, fs_info->sectorsize);
3415         if (lockend <= lockstart)
3416                 lockend = lockstart + fs_info->sectorsize;
3417         lockend--;
3418         len = lockend - lockstart + 1;
3419
3420         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3421                          &cached_state);
3422
3423         while (start < i_size) {
3424                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3425                 if (IS_ERR(em)) {
3426                         ret = PTR_ERR(em);
3427                         em = NULL;
3428                         break;
3429                 }
3430
3431                 if (whence == SEEK_HOLE &&
3432                     (em->block_start == EXTENT_MAP_HOLE ||
3433                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3434                         break;
3435                 else if (whence == SEEK_DATA &&
3436                            (em->block_start != EXTENT_MAP_HOLE &&
3437                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3438                         break;
3439
3440                 start = em->start + em->len;
3441                 free_extent_map(em);
3442                 em = NULL;
3443                 cond_resched();
3444         }
3445         free_extent_map(em);
3446         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3447                              &cached_state);
3448         if (ret) {
3449                 offset = ret;
3450         } else {
3451                 if (whence == SEEK_DATA && start >= i_size)
3452                         offset = -ENXIO;
3453                 else
3454                         offset = min_t(loff_t, start, i_size);
3455         }
3456
3457         return offset;
3458 }
3459
3460 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3461 {
3462         struct inode *inode = file->f_mapping->host;
3463
3464         switch (whence) {
3465         default:
3466                 return generic_file_llseek(file, offset, whence);
3467         case SEEK_DATA:
3468         case SEEK_HOLE:
3469                 inode_lock_shared(inode);
3470                 offset = find_desired_extent(inode, offset, whence);
3471                 inode_unlock_shared(inode);
3472                 break;
3473         }
3474
3475         if (offset < 0)
3476                 return offset;
3477
3478         return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3479 }
3480
3481 static int btrfs_file_open(struct inode *inode, struct file *filp)
3482 {
3483         filp->f_mode |= FMODE_NOWAIT;
3484         return generic_file_open(inode, filp);
3485 }
3486
3487 const struct file_operations btrfs_file_operations = {
3488         .llseek         = btrfs_file_llseek,
3489         .read_iter      = generic_file_read_iter,
3490         .splice_read    = generic_file_splice_read,
3491         .write_iter     = btrfs_file_write_iter,
3492         .mmap           = btrfs_file_mmap,
3493         .open           = btrfs_file_open,
3494         .release        = btrfs_release_file,
3495         .fsync          = btrfs_sync_file,
3496         .fallocate      = btrfs_fallocate,
3497         .unlocked_ioctl = btrfs_ioctl,
3498 #ifdef CONFIG_COMPAT
3499         .compat_ioctl   = btrfs_compat_ioctl,
3500 #endif
3501         .remap_file_range = btrfs_remap_file_range,
3502 };
3503
3504 void __cold btrfs_auto_defrag_exit(void)
3505 {
3506         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3507 }
3508
3509 int __init btrfs_auto_defrag_init(void)
3510 {
3511         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3512                                         sizeof(struct inode_defrag), 0,
3513                                         SLAB_MEM_SPREAD,
3514                                         NULL);
3515         if (!btrfs_inode_defrag_cachep)
3516                 return -ENOMEM;
3517
3518         return 0;
3519 }
3520
3521 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3522 {
3523         int ret;
3524
3525         /*
3526          * So with compression we will find and lock a dirty page and clear the
3527          * first one as dirty, setup an async extent, and immediately return
3528          * with the entire range locked but with nobody actually marked with
3529          * writeback.  So we can't just filemap_write_and_wait_range() and
3530          * expect it to work since it will just kick off a thread to do the
3531          * actual work.  So we need to call filemap_fdatawrite_range _again_
3532          * since it will wait on the page lock, which won't be unlocked until
3533          * after the pages have been marked as writeback and so we're good to go
3534          * from there.  We have to do this otherwise we'll miss the ordered
3535          * extents and that results in badness.  Please Josef, do not think you
3536          * know better and pull this out at some point in the future, it is
3537          * right and you are wrong.
3538          */
3539         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3540         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3541                              &BTRFS_I(inode)->runtime_flags))
3542                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3543
3544         return ret;
3545 }