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_ioctl_defrag_range_args range;
 279         int num_defrag;
 280         int ret;
 281
 282         /* get the inode */
 283         inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
 284         if (IS_ERR(inode_root)) {
 285                 ret = PTR_ERR(inode_root);
 286                 goto cleanup;
 287         }
 288
 289         inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
 290         btrfs_put_root(inode_root);
 291         if (IS_ERR(inode)) {
 292                 ret = PTR_ERR(inode);
 293                 goto cleanup;
 294         }
 295
 296         /* do a chunk of defrag */
 297         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 298         memset(&range, 0, sizeof(range));
 299         range.len = (u64)-1;
 300         range.start = defrag->last_offset;
 301
 302         sb_start_write(fs_info->sb);
 303         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 304                                        BTRFS_DEFRAG_BATCH);
 305         sb_end_write(fs_info->sb);
 306         /*
 307          * if we filled the whole defrag batch, there
 308          * must be more work to do.  Queue this defrag
 309          * again
 310          */
 311         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 312                 defrag->last_offset = range.start;
 313                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 314         } else if (defrag->last_offset && !defrag->cycled) {
 315                 /*
 316                  * we didn't fill our defrag batch, but
 317                  * we didn't start at zero.  Make sure we loop
 318                  * around to the start of the file.
 319                  */
 320                 defrag->last_offset = 0;
 321                 defrag->cycled = 1;
 322                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 323         } else {
 324                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 325         }
 326
 327         iput(inode);
 328         return 0;
 329 cleanup:
 330         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 331         return ret;
 332 }
 333
 334 /*
 335  * run through the list of inodes in the FS that need
 336  * defragging
 337  */
 338 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 339 {
 340         struct inode_defrag *defrag;
 341         u64 first_ino = 0;
 342         u64 root_objectid = 0;
 343
 344         atomic_inc(&fs_info->defrag_running);
 345         while (1) {
 346                 /* Pause the auto defragger. */
 347                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 348                              &fs_info->fs_state))
 349                         break;
 350
 351                 if (!__need_auto_defrag(fs_info))
 352                         break;
 353
 354                 /* find an inode to defrag */
 355                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 356                                                  first_ino);
 357                 if (!defrag) {
 358                         if (root_objectid || first_ino) {
 359                                 root_objectid = 0;
 360                                 first_ino = 0;
 361                                 continue;
 362                         } else {
 363                                 break;
 364                         }
 365                 }
 366
 367                 first_ino = defrag->ino + 1;
 368                 root_objectid = defrag->root;
 369
 370                 __btrfs_run_defrag_inode(fs_info, defrag);
 371         }
 372         atomic_dec(&fs_info->defrag_running);
 373
 374         /*
 375          * during unmount, we use the transaction_wait queue to
 376          * wait for the defragger to stop
 377          */
 378         wake_up(&fs_info->transaction_wait);
 379         return 0;
 380 }
 381
 382 /* simple helper to fault in pages and copy.  This should go away
 383  * and be replaced with calls into generic code.
 384  */
 385 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 386                                          struct page **prepared_pages,
 387                                          struct iov_iter *i)
 388 {
 389         size_t copied = 0;
 390         size_t total_copied = 0;
 391         int pg = 0;
 392         int offset = offset_in_page(pos);
 393
 394         while (write_bytes > 0) {
 395                 size_t count = min_t(size_t,
 396                                      PAGE_SIZE - offset, write_bytes);
 397                 struct page *page = prepared_pages[pg];
 398                 /*
 399                  * Copy data from userspace to the current page
 400                  */
 401                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 402
 403                 /* Flush processor's dcache for this page */
 404                 flush_dcache_page(page);
 405
 406                 /*
 407                  * if we get a partial write, we can end up with
 408                  * partially up to date pages.  These add
 409                  * a lot of complexity, so make sure they don't
 410                  * happen by forcing this copy to be retried.
 411                  *
 412                  * The rest of the btrfs_file_write code will fall
 413                  * back to page at a time copies after we return 0.
 414                  */
 415                 if (!PageUptodate(page) && copied < count)
 416                         copied = 0;
 417
 418                 iov_iter_advance(i, copied);
 419                 write_bytes -= copied;
 420                 total_copied += copied;
 421
 422                 /* Return to btrfs_file_write_iter to fault page */
 423                 if (unlikely(copied == 0))
 424                         break;
 425
 426                 if (copied < PAGE_SIZE - offset) {
 427                         offset += copied;
 428                 } else {
 429                         pg++;
 430                         offset = 0;
 431                 }
 432         }
 433         return total_copied;
 434 }
 435
 436 /*
 437  * unlocks pages after btrfs_file_write is done with them
 438  */
 439 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 440 {
 441         size_t i;
 442         for (i = 0; i < num_pages; i++) {
 443                 /* page checked is some magic around finding pages that
 444                  * have been modified without going through btrfs_set_page_dirty
 445                  * clear it here. There should be no need to mark the pages
 446                  * accessed as prepare_pages should have marked them accessed
 447                  * in prepare_pages via find_or_create_page()
 448                  */
 449                 ClearPageChecked(pages[i]);
 450                 unlock_page(pages[i]);
 451                 put_page(pages[i]);
 452         }
 453 }
 454
 455 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
 456                                          const u64 start,
 457                                          const u64 len,
 458                                          struct extent_state **cached_state)
 459 {
 460         u64 search_start = start;
 461         const u64 end = start + len - 1;
 462
 463         while (search_start < end) {
 464                 const u64 search_len = end - search_start + 1;
 465                 struct extent_map *em;
 466                 u64 em_len;
 467                 int ret = 0;
 468
 469                 em = btrfs_get_extent(inode, NULL, 0, search_start, search_len);
 470                 if (IS_ERR(em))
 471                         return PTR_ERR(em);
 472
 473                 if (em->block_start != EXTENT_MAP_HOLE)
 474                         goto next;
 475
 476                 em_len = em->len;
 477                 if (em->start < search_start)
 478                         em_len -= search_start - em->start;
 479                 if (em_len > search_len)
 480                         em_len = search_len;
 481
 482                 ret = set_extent_bit(&inode->io_tree, search_start,
 483                                      search_start + em_len - 1,
 484                                      EXTENT_DELALLOC_NEW,
 485                                      NULL, cached_state, GFP_NOFS);
 486 next:
 487                 search_start = extent_map_end(em);
 488                 free_extent_map(em);
 489                 if (ret)
 490                         return ret;
 491         }
 492         return 0;
 493 }
 494
 495 /*
 496  * after copy_from_user, pages need to be dirtied and we need to make
 497  * sure holes are created between the current EOF and the start of
 498  * any next extents (if required).
 499  *
 500  * this also makes the decision about creating an inline extent vs
 501  * doing real data extents, marking pages dirty and delalloc as required.
 502  */
 503 int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
 504                       size_t num_pages, loff_t pos, size_t write_bytes,
 505                       struct extent_state **cached)
 506 {
 507         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 508         int err = 0;
 509         int i;
 510         u64 num_bytes;
 511         u64 start_pos;
 512         u64 end_of_last_block;
 513         u64 end_pos = pos + write_bytes;
 514         loff_t isize = i_size_read(&inode->vfs_inode);
 515         unsigned int extra_bits = 0;
 516
 517         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 518         num_bytes = round_up(write_bytes + pos - start_pos,
 519                              fs_info->sectorsize);
 520
 521         end_of_last_block = start_pos + num_bytes - 1;
 522
 523         /*
 524          * The pages may have already been dirty, clear out old accounting so
 525          * we can set things up properly
 526          */
 527         clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
 528                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
 529                          0, 0, cached);
 530
 531         if (!btrfs_is_free_space_inode(inode)) {
 532                 if (start_pos >= isize &&
 533                     !(inode->flags & BTRFS_INODE_PREALLOC)) {
 534                         /*
 535                          * There can't be any extents following eof in this case
 536                          * so just set the delalloc new bit for the range
 537                          * directly.
 538                          */
 539                         extra_bits |= EXTENT_DELALLOC_NEW;
 540                 } else {
 541                         err = btrfs_find_new_delalloc_bytes(inode, start_pos,
 542                                                             num_bytes, cached);
 543                         if (err)
 544                                 return err;
 545                 }
 546         }
 547
 548         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 549                                         extra_bits, cached);
 550         if (err)
 551                 return err;
 552
 553         for (i = 0; i < num_pages; i++) {
 554                 struct page *p = pages[i];
 555                 SetPageUptodate(p);
 556                 ClearPageChecked(p);
 557                 set_page_dirty(p);
 558         }
 559
 560         /*
 561          * we've only changed i_size in ram, and we haven't updated
 562          * the disk i_size.  There is no need to log the inode
 563          * at this time.
 564          */
 565         if (end_pos > isize)
 566                 i_size_write(&inode->vfs_inode, end_pos);
 567         return 0;
 568 }
 569
 570 /*
 571  * this drops all the extents in the cache that intersect the range
 572  * [start, end].  Existing extents are split as required.
 573  */
 574 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 575                              int skip_pinned)
 576 {
 577         struct extent_map *em;
 578         struct extent_map *split = NULL;
 579         struct extent_map *split2 = NULL;
 580         struct extent_map_tree *em_tree = &inode->extent_tree;
 581         u64 len = end - start + 1;
 582         u64 gen;
 583         int ret;
 584         int testend = 1;
 585         unsigned long flags;
 586         int compressed = 0;
 587         bool modified;
 588
 589         WARN_ON(end < start);
 590         if (end == (u64)-1) {
 591                 len = (u64)-1;
 592                 testend = 0;
 593         }
 594         while (1) {
 595                 int no_splits = 0;
 596
 597                 modified = false;
 598                 if (!split)
 599                         split = alloc_extent_map();
 600                 if (!split2)
 601                         split2 = alloc_extent_map();
 602                 if (!split || !split2)
 603                         no_splits = 1;
 604
 605                 write_lock(&em_tree->lock);
 606                 em = lookup_extent_mapping(em_tree, start, len);
 607                 if (!em) {
 608                         write_unlock(&em_tree->lock);
 609                         break;
 610                 }
 611                 flags = em->flags;
 612                 gen = em->generation;
 613                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 614                         if (testend && em->start + em->len >= start + len) {
 615                                 free_extent_map(em);
 616                                 write_unlock(&em_tree->lock);
 617                                 break;
 618                         }
 619                         start = em->start + em->len;
 620                         if (testend)
 621                                 len = start + len - (em->start + em->len);
 622                         free_extent_map(em);
 623                         write_unlock(&em_tree->lock);
 624                         continue;
 625                 }
 626                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 627                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 628                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 629                 modified = !list_empty(&em->list);
 630                 if (no_splits)
 631                         goto next;
 632
 633                 if (em->start < start) {
 634                         split->start = em->start;
 635                         split->len = start - em->start;
 636
 637                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 638                                 split->orig_start = em->orig_start;
 639                                 split->block_start = em->block_start;
 640
 641                                 if (compressed)
 642                                         split->block_len = em->block_len;
 643                                 else
 644                                         split->block_len = split->len;
 645                                 split->orig_block_len = max(split->block_len,
 646                                                 em->orig_block_len);
 647                                 split->ram_bytes = em->ram_bytes;
 648                         } else {
 649                                 split->orig_start = split->start;
 650                                 split->block_len = 0;
 651                                 split->block_start = em->block_start;
 652                                 split->orig_block_len = 0;
 653                                 split->ram_bytes = split->len;
 654                         }
 655
 656                         split->generation = gen;
 657                         split->flags = flags;
 658                         split->compress_type = em->compress_type;
 659                         replace_extent_mapping(em_tree, em, split, modified);
 660                         free_extent_map(split);
 661                         split = split2;
 662                         split2 = NULL;
 663                 }
 664                 if (testend && em->start + em->len > start + len) {
 665                         u64 diff = start + len - em->start;
 666
 667                         split->start = start + len;
 668                         split->len = em->start + em->len - (start + len);
 669                         split->flags = flags;
 670                         split->compress_type = em->compress_type;
 671                         split->generation = gen;
 672
 673                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 674                                 split->orig_block_len = max(em->block_len,
 675                                                     em->orig_block_len);
 676
 677                                 split->ram_bytes = em->ram_bytes;
 678                                 if (compressed) {
 679                                         split->block_len = em->block_len;
 680                                         split->block_start = em->block_start;
 681                                         split->orig_start = em->orig_start;
 682                                 } else {
 683                                         split->block_len = split->len;
 684                                         split->block_start = em->block_start
 685                                                 + diff;
 686                                         split->orig_start = em->orig_start;
 687                                 }
 688                         } else {
 689                                 split->ram_bytes = split->len;
 690                                 split->orig_start = split->start;
 691                                 split->block_len = 0;
 692                                 split->block_start = em->block_start;
 693                                 split->orig_block_len = 0;
 694                         }
 695
 696                         if (extent_map_in_tree(em)) {
 697                                 replace_extent_mapping(em_tree, em, split,
 698                                                        modified);
 699                         } else {
 700                                 ret = add_extent_mapping(em_tree, split,
 701                                                          modified);
 702                                 ASSERT(ret == 0); /* Logic error */
 703                         }
 704                         free_extent_map(split);
 705                         split = NULL;
 706                 }
 707 next:
 708                 if (extent_map_in_tree(em))
 709                         remove_extent_mapping(em_tree, em);
 710                 write_unlock(&em_tree->lock);
 711
 712                 /* once for us */
 713                 free_extent_map(em);
 714                 /* once for the tree*/
 715                 free_extent_map(em);
 716         }
 717         if (split)
 718                 free_extent_map(split);
 719         if (split2)
 720                 free_extent_map(split2);
 721 }
 722
 723 /*
 724  * this is very complex, but the basic idea is to drop all extents
 725  * in the range start - end.  hint_block is filled in with a block number
 726  * that would be a good hint to the block allocator for this file.
 727  *
 728  * If an extent intersects the range but is not entirely inside the range
 729  * it is either truncated or split.  Anything entirely inside the range
 730  * is deleted from the tree.
 731  */
 732 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 733                          struct btrfs_root *root, struct btrfs_inode *inode,
 734                          struct btrfs_path *path, u64 start, u64 end,
 735                          u64 *drop_end, int drop_cache,
 736                          int replace_extent,
 737                          u32 extent_item_size,
 738                          int *key_inserted)
 739 {
 740         struct btrfs_fs_info *fs_info = root->fs_info;
 741         struct extent_buffer *leaf;
 742         struct btrfs_file_extent_item *fi;
 743         struct btrfs_ref ref = { 0 };
 744         struct btrfs_key key;
 745         struct btrfs_key new_key;
 746         struct inode *vfs_inode = &inode->vfs_inode;
 747         u64 ino = btrfs_ino(inode);
 748         u64 search_start = start;
 749         u64 disk_bytenr = 0;
 750         u64 num_bytes = 0;
 751         u64 extent_offset = 0;
 752         u64 extent_end = 0;
 753         u64 last_end = start;
 754         int del_nr = 0;
 755         int del_slot = 0;
 756         int extent_type;
 757         int recow;
 758         int ret;
 759         int modify_tree = -1;
 760         int update_refs;
 761         int found = 0;
 762         int leafs_visited = 0;
 763
 764         if (drop_cache)
 765                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 766
 767         if (start >= inode->disk_i_size && !replace_extent)
 768                 modify_tree = 0;
 769
 770         update_refs = (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
 771                        root == fs_info->tree_root);
 772         while (1) {
 773                 recow = 0;
 774                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 775                                                search_start, modify_tree);
 776                 if (ret < 0)
 777                         break;
 778                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 779                         leaf = path->nodes[0];
 780                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 781                         if (key.objectid == ino &&
 782                             key.type == BTRFS_EXTENT_DATA_KEY)
 783                                 path->slots[0]--;
 784                 }
 785                 ret = 0;
 786                 leafs_visited++;
 787 next_slot:
 788                 leaf = path->nodes[0];
 789                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 790                         BUG_ON(del_nr > 0);
 791                         ret = btrfs_next_leaf(root, path);
 792                         if (ret < 0)
 793                                 break;
 794                         if (ret > 0) {
 795                                 ret = 0;
 796                                 break;
 797                         }
 798                         leafs_visited++;
 799                         leaf = path->nodes[0];
 800                         recow = 1;
 801                 }
 802
 803                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 804
 805                 if (key.objectid > ino)
 806                         break;
 807                 if (WARN_ON_ONCE(key.objectid < ino) ||
 808                     key.type < BTRFS_EXTENT_DATA_KEY) {
 809                         ASSERT(del_nr == 0);
 810                         path->slots[0]++;
 811                         goto next_slot;
 812                 }
 813                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 814                         break;
 815
 816                 fi = btrfs_item_ptr(leaf, path->slots[0],
 817                                     struct btrfs_file_extent_item);
 818                 extent_type = btrfs_file_extent_type(leaf, fi);
 819
 820                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 821                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 822                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 823                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 824                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 825                         extent_end = key.offset +
 826                                 btrfs_file_extent_num_bytes(leaf, fi);
 827                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 828                         extent_end = key.offset +
 829                                 btrfs_file_extent_ram_bytes(leaf, fi);
 830                 } else {
 831                         /* can't happen */
 832                         BUG();
 833                 }
 834
 835                 /*
 836                  * Don't skip extent items representing 0 byte lengths. They
 837                  * used to be created (bug) if while punching holes we hit
 838                  * -ENOSPC condition. So if we find one here, just ensure we
 839                  * delete it, otherwise we would insert a new file extent item
 840                  * with the same key (offset) as that 0 bytes length file
 841                  * extent item in the call to setup_items_for_insert() later
 842                  * in this function.
 843                  */
 844                 if (extent_end == key.offset && extent_end >= search_start) {
 845                         last_end = extent_end;
 846                         goto delete_extent_item;
 847                 }
 848
 849                 if (extent_end <= search_start) {
 850                         path->slots[0]++;
 851                         goto next_slot;
 852                 }
 853
 854                 found = 1;
 855                 search_start = max(key.offset, start);
 856                 if (recow || !modify_tree) {
 857                         modify_tree = -1;
 858                         btrfs_release_path(path);
 859                         continue;
 860                 }
 861
 862                 /*
 863                  *     | - range to drop - |
 864                  *  | -------- extent -------- |
 865                  */
 866                 if (start > key.offset && end < extent_end) {
 867                         BUG_ON(del_nr > 0);
 868                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 869                                 ret = -EOPNOTSUPP;
 870                                 break;
 871                         }
 872
 873                         memcpy(&new_key, &key, sizeof(new_key));
 874                         new_key.offset = start;
 875                         ret = btrfs_duplicate_item(trans, root, path,
 876                                                    &new_key);
 877                         if (ret == -EAGAIN) {
 878                                 btrfs_release_path(path);
 879                                 continue;
 880                         }
 881                         if (ret < 0)
 882                                 break;
 883
 884                         leaf = path->nodes[0];
 885                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 886                                             struct btrfs_file_extent_item);
 887                         btrfs_set_file_extent_num_bytes(leaf, fi,
 888                                                         start - key.offset);
 889
 890                         fi = btrfs_item_ptr(leaf, path->slots[0],
 891                                             struct btrfs_file_extent_item);
 892
 893                         extent_offset += start - key.offset;
 894                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 895                         btrfs_set_file_extent_num_bytes(leaf, fi,
 896                                                         extent_end - start);
 897                         btrfs_mark_buffer_dirty(leaf);
 898
 899                         if (update_refs && disk_bytenr > 0) {
 900                                 btrfs_init_generic_ref(&ref,
 901                                                 BTRFS_ADD_DELAYED_REF,
 902                                                 disk_bytenr, num_bytes, 0);
 903                                 btrfs_init_data_ref(&ref,
 904                                                 root->root_key.objectid,
 905                                                 new_key.objectid,
 906                                                 start - extent_offset);
 907                                 ret = btrfs_inc_extent_ref(trans, &ref);
 908                                 BUG_ON(ret); /* -ENOMEM */
 909                         }
 910                         key.offset = start;
 911                 }
 912                 /*
 913                  * From here on out we will have actually dropped something, so
 914                  * last_end can be updated.
 915                  */
 916                 last_end = extent_end;
 917
 918                 /*
 919                  *  | ---- range to drop ----- |
 920                  *      | -------- extent -------- |
 921                  */
 922                 if (start <= key.offset && end < extent_end) {
 923                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 924                                 ret = -EOPNOTSUPP;
 925                                 break;
 926                         }
 927
 928                         memcpy(&new_key, &key, sizeof(new_key));
 929                         new_key.offset = end;
 930                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 931
 932                         extent_offset += end - key.offset;
 933                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 934                         btrfs_set_file_extent_num_bytes(leaf, fi,
 935                                                         extent_end - end);
 936                         btrfs_mark_buffer_dirty(leaf);
 937                         if (update_refs && disk_bytenr > 0)
 938                                 inode_sub_bytes(vfs_inode, end - key.offset);
 939                         break;
 940                 }
 941
 942                 search_start = extent_end;
 943                 /*
 944                  *       | ---- range to drop ----- |
 945                  *  | -------- extent -------- |
 946                  */
 947                 if (start > key.offset && end >= extent_end) {
 948                         BUG_ON(del_nr > 0);
 949                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 950                                 ret = -EOPNOTSUPP;
 951                                 break;
 952                         }
 953
 954                         btrfs_set_file_extent_num_bytes(leaf, fi,
 955                                                         start - key.offset);
 956                         btrfs_mark_buffer_dirty(leaf);
 957                         if (update_refs && disk_bytenr > 0)
 958                                 inode_sub_bytes(vfs_inode, extent_end - start);
 959                         if (end == extent_end)
 960                                 break;
 961
 962                         path->slots[0]++;
 963                         goto next_slot;
 964                 }
 965
 966                 /*
 967                  *  | ---- range to drop ----- |
 968                  *    | ------ extent ------ |
 969                  */
 970                 if (start <= key.offset && end >= extent_end) {
 971 delete_extent_item:
 972                         if (del_nr == 0) {
 973                                 del_slot = path->slots[0];
 974                                 del_nr = 1;
 975                         } else {
 976                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 977                                 del_nr++;
 978                         }
 979
 980                         if (update_refs &&
 981                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 982                                 inode_sub_bytes(vfs_inode,
 983                                                 extent_end - key.offset);
 984                                 extent_end = ALIGN(extent_end,
 985                                                    fs_info->sectorsize);
 986                         } else if (update_refs && disk_bytenr > 0) {
 987                                 btrfs_init_generic_ref(&ref,
 988                                                 BTRFS_DROP_DELAYED_REF,
 989                                                 disk_bytenr, num_bytes, 0);
 990                                 btrfs_init_data_ref(&ref,
 991                                                 root->root_key.objectid,
 992                                                 key.objectid,
 993                                                 key.offset - extent_offset);
 994                                 ret = btrfs_free_extent(trans, &ref);
 995                                 BUG_ON(ret); /* -ENOMEM */
 996                                 inode_sub_bytes(vfs_inode,
 997                                                 extent_end - key.offset);
 998                         }
 999
1000                         if (end == extent_end)
1001                                 break;
1002
1003                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1004                                 path->slots[0]++;
1005                                 goto next_slot;
1006                         }
1007
1008                         ret = btrfs_del_items(trans, root, path, del_slot,
1009                                               del_nr);
1010                         if (ret) {
1011                                 btrfs_abort_transaction(trans, ret);
1012                                 break;
1013                         }
1014
1015                         del_nr = 0;
1016                         del_slot = 0;
1017
1018                         btrfs_release_path(path);
1019                         continue;
1020                 }
1021
1022                 BUG();
1023         }
1024
1025         if (!ret && del_nr > 0) {
1026                 /*
1027                  * Set path->slots[0] to first slot, so that after the delete
1028                  * if items are move off from our leaf to its immediate left or
1029                  * right neighbor leafs, we end up with a correct and adjusted
1030                  * path->slots[0] for our insertion (if replace_extent != 0).
1031                  */
1032                 path->slots[0] = del_slot;
1033                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1034                 if (ret)
1035                         btrfs_abort_transaction(trans, ret);
1036         }
1037
1038         leaf = path->nodes[0];
1039         /*
1040          * If btrfs_del_items() was called, it might have deleted a leaf, in
1041          * which case it unlocked our path, so check path->locks[0] matches a
1042          * write lock.
1043          */
1044         if (!ret && replace_extent && leafs_visited == 1 &&
1045             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1046              path->locks[0] == BTRFS_WRITE_LOCK) &&
1047             btrfs_leaf_free_space(leaf) >=
1048             sizeof(struct btrfs_item) + extent_item_size) {
1049
1050                 key.objectid = ino;
1051                 key.type = BTRFS_EXTENT_DATA_KEY;
1052                 key.offset = start;
1053                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1054                         struct btrfs_key slot_key;
1055
1056                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1057                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1058                                 path->slots[0]++;
1059                 }
1060                 setup_items_for_insert(root, path, &key, &extent_item_size, 1);
1061                 *key_inserted = 1;
1062         }
1063
1064         if (!replace_extent || !(*key_inserted))
1065                 btrfs_release_path(path);
1066         if (drop_end)
1067                 *drop_end = found ? min(end, last_end) : end;
1068         return ret;
1069 }
1070
1071 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1072                        struct btrfs_root *root, struct inode *inode, u64 start,
1073                        u64 end, int drop_cache)
1074 {
1075         struct btrfs_path *path;
1076         int ret;
1077
1078         path = btrfs_alloc_path();
1079         if (!path)
1080                 return -ENOMEM;
1081         ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path, start,
1082                                    end, NULL, drop_cache, 0, 0, NULL);
1083         btrfs_free_path(path);
1084         return ret;
1085 }
1086
1087 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1088                             u64 objectid, u64 bytenr, u64 orig_offset,
1089                             u64 *start, u64 *end)
1090 {
1091         struct btrfs_file_extent_item *fi;
1092         struct btrfs_key key;
1093         u64 extent_end;
1094
1095         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1096                 return 0;
1097
1098         btrfs_item_key_to_cpu(leaf, &key, slot);
1099         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1100                 return 0;
1101
1102         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1103         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1104             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1105             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1106             btrfs_file_extent_compression(leaf, fi) ||
1107             btrfs_file_extent_encryption(leaf, fi) ||
1108             btrfs_file_extent_other_encoding(leaf, fi))
1109                 return 0;
1110
1111         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1112         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1113                 return 0;
1114
1115         *start = key.offset;
1116         *end = extent_end;
1117         return 1;
1118 }
1119
1120 /*
1121  * Mark extent in the range start - end as written.
1122  *
1123  * This changes extent type from 'pre-allocated' to 'regular'. If only
1124  * part of extent is marked as written, the extent will be split into
1125  * two or three.
1126  */
1127 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1128                               struct btrfs_inode *inode, u64 start, u64 end)
1129 {
1130         struct btrfs_fs_info *fs_info = trans->fs_info;
1131         struct btrfs_root *root = inode->root;
1132         struct extent_buffer *leaf;
1133         struct btrfs_path *path;
1134         struct btrfs_file_extent_item *fi;
1135         struct btrfs_ref ref = { 0 };
1136         struct btrfs_key key;
1137         struct btrfs_key new_key;
1138         u64 bytenr;
1139         u64 num_bytes;
1140         u64 extent_end;
1141         u64 orig_offset;
1142         u64 other_start;
1143         u64 other_end;
1144         u64 split;
1145         int del_nr = 0;
1146         int del_slot = 0;
1147         int recow;
1148         int ret;
1149         u64 ino = btrfs_ino(inode);
1150
1151         path = btrfs_alloc_path();
1152         if (!path)
1153                 return -ENOMEM;
1154 again:
1155         recow = 0;
1156         split = start;
1157         key.objectid = ino;
1158         key.type = BTRFS_EXTENT_DATA_KEY;
1159         key.offset = split;
1160
1161         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1162         if (ret < 0)
1163                 goto out;
1164         if (ret > 0 && path->slots[0] > 0)
1165                 path->slots[0]--;
1166
1167         leaf = path->nodes[0];
1168         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1169         if (key.objectid != ino ||
1170             key.type != BTRFS_EXTENT_DATA_KEY) {
1171                 ret = -EINVAL;
1172                 btrfs_abort_transaction(trans, ret);
1173                 goto out;
1174         }
1175         fi = btrfs_item_ptr(leaf, path->slots[0],
1176                             struct btrfs_file_extent_item);
1177         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1178                 ret = -EINVAL;
1179                 btrfs_abort_transaction(trans, ret);
1180                 goto out;
1181         }
1182         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1183         if (key.offset > start || extent_end < end) {
1184                 ret = -EINVAL;
1185                 btrfs_abort_transaction(trans, ret);
1186                 goto out;
1187         }
1188
1189         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1190         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1191         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1192         memcpy(&new_key, &key, sizeof(new_key));
1193
1194         if (start == key.offset && end < extent_end) {
1195                 other_start = 0;
1196                 other_end = start;
1197                 if (extent_mergeable(leaf, path->slots[0] - 1,
1198                                      ino, bytenr, orig_offset,
1199                                      &other_start, &other_end)) {
1200                         new_key.offset = end;
1201                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1202                         fi = btrfs_item_ptr(leaf, path->slots[0],
1203                                             struct btrfs_file_extent_item);
1204                         btrfs_set_file_extent_generation(leaf, fi,
1205                                                          trans->transid);
1206                         btrfs_set_file_extent_num_bytes(leaf, fi,
1207                                                         extent_end - end);
1208                         btrfs_set_file_extent_offset(leaf, fi,
1209                                                      end - orig_offset);
1210                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1211                                             struct btrfs_file_extent_item);
1212                         btrfs_set_file_extent_generation(leaf, fi,
1213                                                          trans->transid);
1214                         btrfs_set_file_extent_num_bytes(leaf, fi,
1215                                                         end - other_start);
1216                         btrfs_mark_buffer_dirty(leaf);
1217                         goto out;
1218                 }
1219         }
1220
1221         if (start > key.offset && end == extent_end) {
1222                 other_start = end;
1223                 other_end = 0;
1224                 if (extent_mergeable(leaf, path->slots[0] + 1,
1225                                      ino, bytenr, orig_offset,
1226                                      &other_start, &other_end)) {
1227                         fi = btrfs_item_ptr(leaf, path->slots[0],
1228                                             struct btrfs_file_extent_item);
1229                         btrfs_set_file_extent_num_bytes(leaf, fi,
1230                                                         start - key.offset);
1231                         btrfs_set_file_extent_generation(leaf, fi,
1232                                                          trans->transid);
1233                         path->slots[0]++;
1234                         new_key.offset = start;
1235                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1236
1237                         fi = btrfs_item_ptr(leaf, path->slots[0],
1238                                             struct btrfs_file_extent_item);
1239                         btrfs_set_file_extent_generation(leaf, fi,
1240                                                          trans->transid);
1241                         btrfs_set_file_extent_num_bytes(leaf, fi,
1242                                                         other_end - start);
1243                         btrfs_set_file_extent_offset(leaf, fi,
1244                                                      start - orig_offset);
1245                         btrfs_mark_buffer_dirty(leaf);
1246                         goto out;
1247                 }
1248         }
1249
1250         while (start > key.offset || end < extent_end) {
1251                 if (key.offset == start)
1252                         split = end;
1253
1254                 new_key.offset = split;
1255                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1256                 if (ret == -EAGAIN) {
1257                         btrfs_release_path(path);
1258                         goto again;
1259                 }
1260                 if (ret < 0) {
1261                         btrfs_abort_transaction(trans, ret);
1262                         goto out;
1263                 }
1264
1265                 leaf = path->nodes[0];
1266                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1267                                     struct btrfs_file_extent_item);
1268                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1269                 btrfs_set_file_extent_num_bytes(leaf, fi,
1270                                                 split - key.offset);
1271
1272                 fi = btrfs_item_ptr(leaf, path->slots[0],
1273                                     struct btrfs_file_extent_item);
1274
1275                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1276                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1277                 btrfs_set_file_extent_num_bytes(leaf, fi,
1278                                                 extent_end - split);
1279                 btrfs_mark_buffer_dirty(leaf);
1280
1281                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1282                                        num_bytes, 0);
1283                 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1284                                     orig_offset);
1285                 ret = btrfs_inc_extent_ref(trans, &ref);
1286                 if (ret) {
1287                         btrfs_abort_transaction(trans, ret);
1288                         goto out;
1289                 }
1290
1291                 if (split == start) {
1292                         key.offset = start;
1293                 } else {
1294                         if (start != key.offset) {
1295                                 ret = -EINVAL;
1296                                 btrfs_abort_transaction(trans, ret);
1297                                 goto out;
1298                         }
1299                         path->slots[0]--;
1300                         extent_end = end;
1301                 }
1302                 recow = 1;
1303         }
1304
1305         other_start = end;
1306         other_end = 0;
1307         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1308                                num_bytes, 0);
1309         btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1310         if (extent_mergeable(leaf, path->slots[0] + 1,
1311                              ino, bytenr, orig_offset,
1312                              &other_start, &other_end)) {
1313                 if (recow) {
1314                         btrfs_release_path(path);
1315                         goto again;
1316                 }
1317                 extent_end = other_end;
1318                 del_slot = path->slots[0] + 1;
1319                 del_nr++;
1320                 ret = btrfs_free_extent(trans, &ref);
1321                 if (ret) {
1322                         btrfs_abort_transaction(trans, ret);
1323                         goto out;
1324                 }
1325         }
1326         other_start = 0;
1327         other_end = start;
1328         if (extent_mergeable(leaf, path->slots[0] - 1,
1329                              ino, bytenr, orig_offset,
1330                              &other_start, &other_end)) {
1331                 if (recow) {
1332                         btrfs_release_path(path);
1333                         goto again;
1334                 }
1335                 key.offset = other_start;
1336                 del_slot = path->slots[0];
1337                 del_nr++;
1338                 ret = btrfs_free_extent(trans, &ref);
1339                 if (ret) {
1340                         btrfs_abort_transaction(trans, ret);
1341                         goto out;
1342                 }
1343         }
1344         if (del_nr == 0) {
1345                 fi = btrfs_item_ptr(leaf, path->slots[0],
1346                            struct btrfs_file_extent_item);
1347                 btrfs_set_file_extent_type(leaf, fi,
1348                                            BTRFS_FILE_EXTENT_REG);
1349                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1350                 btrfs_mark_buffer_dirty(leaf);
1351         } else {
1352                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1353                            struct btrfs_file_extent_item);
1354                 btrfs_set_file_extent_type(leaf, fi,
1355                                            BTRFS_FILE_EXTENT_REG);
1356                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1357                 btrfs_set_file_extent_num_bytes(leaf, fi,
1358                                                 extent_end - key.offset);
1359                 btrfs_mark_buffer_dirty(leaf);
1360
1361                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1362                 if (ret < 0) {
1363                         btrfs_abort_transaction(trans, ret);
1364                         goto out;
1365                 }
1366         }
1367 out:
1368         btrfs_free_path(path);
1369         return 0;
1370 }
1371
1372 /*
1373  * on error we return an unlocked page and the error value
1374  * on success we return a locked page and 0
1375  */
1376 static int prepare_uptodate_page(struct inode *inode,
1377                                  struct page *page, u64 pos,
1378                                  bool force_uptodate)
1379 {
1380         int ret = 0;
1381
1382         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1383             !PageUptodate(page)) {
1384                 ret = btrfs_readpage(NULL, page);
1385                 if (ret)
1386                         return ret;
1387                 lock_page(page);
1388                 if (!PageUptodate(page)) {
1389                         unlock_page(page);
1390                         return -EIO;
1391                 }
1392                 if (page->mapping != inode->i_mapping) {
1393                         unlock_page(page);
1394                         return -EAGAIN;
1395                 }
1396         }
1397         return 0;
1398 }
1399
1400 /*
1401  * this just gets pages into the page cache and locks them down.
1402  */
1403 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1404                                   size_t num_pages, loff_t pos,
1405                                   size_t write_bytes, bool force_uptodate)
1406 {
1407         int i;
1408         unsigned long index = pos >> PAGE_SHIFT;
1409         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1410         int err = 0;
1411         int faili;
1412
1413         for (i = 0; i < num_pages; i++) {
1414 again:
1415                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1416                                                mask | __GFP_WRITE);
1417                 if (!pages[i]) {
1418                         faili = i - 1;
1419                         err = -ENOMEM;
1420                         goto fail;
1421                 }
1422
1423                 if (i == 0)
1424                         err = prepare_uptodate_page(inode, pages[i], pos,
1425                                                     force_uptodate);
1426                 if (!err && i == num_pages - 1)
1427                         err = prepare_uptodate_page(inode, pages[i],
1428                                                     pos + write_bytes, false);
1429                 if (err) {
1430                         put_page(pages[i]);
1431                         if (err == -EAGAIN) {
1432                                 err = 0;
1433                                 goto again;
1434                         }
1435                         faili = i - 1;
1436                         goto fail;
1437                 }
1438                 wait_on_page_writeback(pages[i]);
1439         }
1440
1441         return 0;
1442 fail:
1443         while (faili >= 0) {
1444                 unlock_page(pages[faili]);
1445                 put_page(pages[faili]);
1446                 faili--;
1447         }
1448         return err;
1449
1450 }
1451
1452 /*
1453  * This function locks the extent and properly waits for data=ordered extents
1454  * to finish before allowing the pages to be modified if need.
1455  *
1456  * The return value:
1457  * 1 - the extent is locked
1458  * 0 - the extent is not locked, and everything is OK
1459  * -EAGAIN - need re-prepare the pages
1460  * the other < 0 number - Something wrong happens
1461  */
1462 static noinline int
1463 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1464                                 size_t num_pages, loff_t pos,
1465                                 size_t write_bytes,
1466                                 u64 *lockstart, u64 *lockend,
1467                                 struct extent_state **cached_state)
1468 {
1469         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1470         u64 start_pos;
1471         u64 last_pos;
1472         int i;
1473         int ret = 0;
1474
1475         start_pos = round_down(pos, fs_info->sectorsize);
1476         last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1477
1478         if (start_pos < inode->vfs_inode.i_size) {
1479                 struct btrfs_ordered_extent *ordered;
1480
1481                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1482                                 cached_state);
1483                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1484                                                      last_pos - start_pos + 1);
1485                 if (ordered &&
1486                     ordered->file_offset + ordered->num_bytes > start_pos &&
1487                     ordered->file_offset <= last_pos) {
1488                         unlock_extent_cached(&inode->io_tree, start_pos,
1489                                         last_pos, cached_state);
1490                         for (i = 0; i < num_pages; i++) {
1491                                 unlock_page(pages[i]);
1492                                 put_page(pages[i]);
1493                         }
1494                         btrfs_start_ordered_extent(ordered, 1);
1495                         btrfs_put_ordered_extent(ordered);
1496                         return -EAGAIN;
1497                 }
1498                 if (ordered)
1499                         btrfs_put_ordered_extent(ordered);
1500
1501                 *lockstart = start_pos;
1502                 *lockend = last_pos;
1503                 ret = 1;
1504         }
1505
1506         /*
1507          * It's possible the pages are dirty right now, but we don't want
1508          * to clean them yet because copy_from_user may catch a page fault
1509          * and we might have to fall back to one page at a time.  If that
1510          * happens, we'll unlock these pages and we'd have a window where
1511          * reclaim could sneak in and drop the once-dirty page on the floor
1512          * without writing it.
1513          *
1514          * We have the pages locked and the extent range locked, so there's
1515          * no way someone can start IO on any dirty pages in this range.
1516          *
1517          * We'll call btrfs_dirty_pages() later on, and that will flip around
1518          * delalloc bits and dirty the pages as required.
1519          */
1520         for (i = 0; i < num_pages; i++) {
1521                 set_page_extent_mapped(pages[i]);
1522                 WARN_ON(!PageLocked(pages[i]));
1523         }
1524
1525         return ret;
1526 }
1527
1528 static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1529                            size_t *write_bytes, bool nowait)
1530 {
1531         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1532         struct btrfs_root *root = inode->root;
1533         u64 lockstart, lockend;
1534         u64 num_bytes;
1535         int ret;
1536
1537         if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1538                 return 0;
1539
1540         if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1541                 return -EAGAIN;
1542
1543         lockstart = round_down(pos, fs_info->sectorsize);
1544         lockend = round_up(pos + *write_bytes,
1545                            fs_info->sectorsize) - 1;
1546         num_bytes = lockend - lockstart + 1;
1547
1548         if (nowait) {
1549                 struct btrfs_ordered_extent *ordered;
1550
1551                 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1552                         return -EAGAIN;
1553
1554                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1555                                                      num_bytes);
1556                 if (ordered) {
1557                         btrfs_put_ordered_extent(ordered);
1558                         ret = -EAGAIN;
1559                         goto out_unlock;
1560                 }
1561         } else {
1562                 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1563                                                    lockend, NULL);
1564         }
1565
1566         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1567                         NULL, NULL, NULL, false);
1568         if (ret <= 0) {
1569                 ret = 0;
1570                 if (!nowait)
1571                         btrfs_drew_write_unlock(&root->snapshot_lock);
1572         } else {
1573                 *write_bytes = min_t(size_t, *write_bytes ,
1574                                      num_bytes - pos + lockstart);
1575         }
1576 out_unlock:
1577         unlock_extent(&inode->io_tree, lockstart, lockend);
1578
1579         return ret;
1580 }
1581
1582 static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1583                               size_t *write_bytes)
1584 {
1585         return check_can_nocow(inode, pos, write_bytes, true);
1586 }
1587
1588 /*
1589  * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1590  *
1591  * @pos:         File offset
1592  * @write_bytes: The length to write, will be updated to the nocow writeable
1593  *               range
1594  *
1595  * This function will flush ordered extents in the range to ensure proper
1596  * nocow checks.
1597  *
1598  * Return:
1599  * >0           and update @write_bytes if we can do nocow write
1600  *  0           if we can't do nocow write
1601  * -EAGAIN      if we can't get the needed lock or there are ordered extents
1602  *              for * (nowait == true) case
1603  * <0           if other error happened
1604  *
1605  * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1606  */
1607 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1608                            size_t *write_bytes)
1609 {
1610         return check_can_nocow(inode, pos, write_bytes, false);
1611 }
1612
1613 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1614 {
1615         btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1616 }
1617
1618 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1619                                                struct iov_iter *i)
1620 {
1621         struct file *file = iocb->ki_filp;
1622         loff_t pos = iocb->ki_pos;
1623         struct inode *inode = file_inode(file);
1624         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1625         struct page **pages = NULL;
1626         struct extent_changeset *data_reserved = NULL;
1627         u64 release_bytes = 0;
1628         u64 lockstart;
1629         u64 lockend;
1630         size_t num_written = 0;
1631         int nrptrs;
1632         int ret = 0;
1633         bool only_release_metadata = false;
1634         bool force_page_uptodate = false;
1635
1636         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1637                         PAGE_SIZE / (sizeof(struct page *)));
1638         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1639         nrptrs = max(nrptrs, 8);
1640         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1641         if (!pages)
1642                 return -ENOMEM;
1643
1644         while (iov_iter_count(i) > 0) {
1645                 struct extent_state *cached_state = NULL;
1646                 size_t offset = offset_in_page(pos);
1647                 size_t sector_offset;
1648                 size_t write_bytes = min(iov_iter_count(i),
1649                                          nrptrs * (size_t)PAGE_SIZE -
1650                                          offset);
1651                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1652                                                 PAGE_SIZE);
1653                 size_t reserve_bytes;
1654                 size_t dirty_pages;
1655                 size_t copied;
1656                 size_t dirty_sectors;
1657                 size_t num_sectors;
1658                 int extents_locked;
1659
1660                 WARN_ON(num_pages > nrptrs);
1661
1662                 /*
1663                  * Fault pages before locking them in prepare_pages
1664                  * to avoid recursive lock
1665                  */
1666                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1667                         ret = -EFAULT;
1668                         break;
1669                 }
1670
1671                 only_release_metadata = false;
1672                 sector_offset = pos & (fs_info->sectorsize - 1);
1673                 reserve_bytes = round_up(write_bytes + sector_offset,
1674                                 fs_info->sectorsize);
1675
1676                 extent_changeset_release(data_reserved);
1677                 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1678                                                   &data_reserved, pos,
1679                                                   write_bytes);
1680                 if (ret < 0) {
1681                         if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1682                                                    &write_bytes) > 0) {
1683                                 /*
1684                                  * For nodata cow case, no need to reserve
1685                                  * data space.
1686                                  */
1687                                 only_release_metadata = true;
1688                                 /*
1689                                  * our prealloc extent may be smaller than
1690                                  * write_bytes, so scale down.
1691                                  */
1692                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1693                                                          PAGE_SIZE);
1694                                 reserve_bytes = round_up(write_bytes +
1695                                                          sector_offset,
1696                                                          fs_info->sectorsize);
1697                         } else {
1698                                 break;
1699                         }
1700                 }
1701
1702                 WARN_ON(reserve_bytes == 0);
1703                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1704                                 reserve_bytes);
1705                 if (ret) {
1706                         if (!only_release_metadata)
1707                                 btrfs_free_reserved_data_space(BTRFS_I(inode),
1708                                                 data_reserved, pos,
1709                                                 write_bytes);
1710                         else
1711                                 btrfs_check_nocow_unlock(BTRFS_I(inode));
1712                         break;
1713                 }
1714
1715                 release_bytes = reserve_bytes;
1716 again:
1717                 /*
1718                  * This is going to setup the pages array with the number of
1719                  * pages we want, so we don't really need to worry about the
1720                  * contents of pages from loop to loop
1721                  */
1722                 ret = prepare_pages(inode, pages, num_pages,
1723                                     pos, write_bytes,
1724                                     force_page_uptodate);
1725                 if (ret) {
1726                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1727                                                        reserve_bytes);
1728                         break;
1729                 }
1730
1731                 extents_locked = lock_and_cleanup_extent_if_need(
1732                                 BTRFS_I(inode), pages,
1733                                 num_pages, pos, write_bytes, &lockstart,
1734                                 &lockend, &cached_state);
1735                 if (extents_locked < 0) {
1736                         if (extents_locked == -EAGAIN)
1737                                 goto again;
1738                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1739                                                        reserve_bytes);
1740                         ret = extents_locked;
1741                         break;
1742                 }
1743
1744                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1745
1746                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1747                 dirty_sectors = round_up(copied + sector_offset,
1748                                         fs_info->sectorsize);
1749                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1750
1751                 /*
1752                  * if we have trouble faulting in the pages, fall
1753                  * back to one page at a time
1754                  */
1755                 if (copied < write_bytes)
1756                         nrptrs = 1;
1757
1758                 if (copied == 0) {
1759                         force_page_uptodate = true;
1760                         dirty_sectors = 0;
1761                         dirty_pages = 0;
1762                 } else {
1763                         force_page_uptodate = false;
1764                         dirty_pages = DIV_ROUND_UP(copied + offset,
1765                                                    PAGE_SIZE);
1766                 }
1767
1768                 if (num_sectors > dirty_sectors) {
1769                         /* release everything except the sectors we dirtied */
1770                         release_bytes -= dirty_sectors <<
1771                                                 fs_info->sb->s_blocksize_bits;
1772                         if (only_release_metadata) {
1773                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1774                                                         release_bytes, true);
1775                         } else {
1776                                 u64 __pos;
1777
1778                                 __pos = round_down(pos,
1779                                                    fs_info->sectorsize) +
1780                                         (dirty_pages << PAGE_SHIFT);
1781                                 btrfs_delalloc_release_space(BTRFS_I(inode),
1782                                                 data_reserved, __pos,
1783                                                 release_bytes, true);
1784                         }
1785                 }
1786
1787                 release_bytes = round_up(copied + sector_offset,
1788                                         fs_info->sectorsize);
1789
1790                 if (copied > 0)
1791                         ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1792                                                 dirty_pages, pos, copied,
1793                                                 &cached_state);
1794
1795                 /*
1796                  * If we have not locked the extent range, because the range's
1797                  * start offset is >= i_size, we might still have a non-NULL
1798                  * cached extent state, acquired while marking the extent range
1799                  * as delalloc through btrfs_dirty_pages(). Therefore free any
1800                  * possible cached extent state to avoid a memory leak.
1801                  */
1802                 if (extents_locked)
1803                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1804                                              lockstart, lockend, &cached_state);
1805                 else
1806                         free_extent_state(cached_state);
1807
1808                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1809                 if (ret) {
1810                         btrfs_drop_pages(pages, num_pages);
1811                         break;
1812                 }
1813
1814                 release_bytes = 0;
1815                 if (only_release_metadata)
1816                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1817
1818                 if (only_release_metadata && copied > 0) {
1819                         lockstart = round_down(pos,
1820                                                fs_info->sectorsize);
1821                         lockend = round_up(pos + copied,
1822                                            fs_info->sectorsize) - 1;
1823
1824                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1825                                        lockend, EXTENT_NORESERVE, NULL,
1826                                        NULL, GFP_NOFS);
1827                 }
1828
1829                 btrfs_drop_pages(pages, num_pages);
1830
1831                 cond_resched();
1832
1833                 balance_dirty_pages_ratelimited(inode->i_mapping);
1834
1835                 pos += copied;
1836                 num_written += copied;
1837         }
1838
1839         kfree(pages);
1840
1841         if (release_bytes) {
1842                 if (only_release_metadata) {
1843                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1844                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1845                                         release_bytes, true);
1846                 } else {
1847                         btrfs_delalloc_release_space(BTRFS_I(inode),
1848                                         data_reserved,
1849                                         round_down(pos, fs_info->sectorsize),
1850                                         release_bytes, true);
1851                 }
1852         }
1853
1854         extent_changeset_free(data_reserved);
1855         return num_written ? num_written : ret;
1856 }
1857
1858 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1859 {
1860         struct file *file = iocb->ki_filp;
1861         struct inode *inode = file_inode(file);
1862         loff_t pos;
1863         ssize_t written;
1864         ssize_t written_buffered;
1865         loff_t endbyte;
1866         int err;
1867
1868         written = btrfs_direct_IO(iocb, from);
1869
1870         if (written < 0 || !iov_iter_count(from))
1871                 return written;
1872
1873         pos = iocb->ki_pos;
1874         written_buffered = btrfs_buffered_write(iocb, from);
1875         if (written_buffered < 0) {
1876                 err = written_buffered;
1877                 goto out;
1878         }
1879         /*
1880          * Ensure all data is persisted. We want the next direct IO read to be
1881          * able to read what was just written.
1882          */
1883         endbyte = pos + written_buffered - 1;
1884         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1885         if (err)
1886                 goto out;
1887         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1888         if (err)
1889                 goto out;
1890         written += written_buffered;
1891         iocb->ki_pos = pos + written_buffered;
1892         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1893                                  endbyte >> PAGE_SHIFT);
1894 out:
1895         return written ? written : err;
1896 }
1897
1898 static void update_time_for_write(struct inode *inode)
1899 {
1900         struct timespec64 now;
1901
1902         if (IS_NOCMTIME(inode))
1903                 return;
1904
1905         now = current_time(inode);
1906         if (!timespec64_equal(&inode->i_mtime, &now))
1907                 inode->i_mtime = now;
1908
1909         if (!timespec64_equal(&inode->i_ctime, &now))
1910                 inode->i_ctime = now;
1911
1912         if (IS_I_VERSION(inode))
1913                 inode_inc_iversion(inode);
1914 }
1915
1916 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1917                                     struct iov_iter *from)
1918 {
1919         struct file *file = iocb->ki_filp;
1920         struct inode *inode = file_inode(file);
1921         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1922         struct btrfs_root *root = BTRFS_I(inode)->root;
1923         u64 start_pos;
1924         u64 end_pos;
1925         ssize_t num_written = 0;
1926         const bool sync = iocb->ki_flags & IOCB_DSYNC;
1927         ssize_t err;
1928         loff_t pos;
1929         size_t count;
1930         loff_t oldsize;
1931         int clean_page = 0;
1932
1933         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1934             (iocb->ki_flags & IOCB_NOWAIT))
1935                 return -EOPNOTSUPP;
1936
1937         if (iocb->ki_flags & IOCB_NOWAIT) {
1938                 if (!inode_trylock(inode))
1939                         return -EAGAIN;
1940         } else {
1941                 inode_lock(inode);
1942         }
1943
1944         err = generic_write_checks(iocb, from);
1945         if (err <= 0) {
1946                 inode_unlock(inode);
1947                 return err;
1948         }
1949
1950         pos = iocb->ki_pos;
1951         count = iov_iter_count(from);
1952         if (iocb->ki_flags & IOCB_NOWAIT) {
1953                 size_t nocow_bytes = count;
1954
1955                 /*
1956                  * We will allocate space in case nodatacow is not set,
1957                  * so bail
1958                  */
1959                 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes)
1960                     <= 0) {
1961                         inode_unlock(inode);
1962                         return -EAGAIN;
1963                 }
1964                 /*
1965                  * There are holes in the range or parts of the range that must
1966                  * be COWed (shared extents, RO block groups, etc), so just bail
1967                  * out.
1968                  */
1969                 if (nocow_bytes < count) {
1970                         inode_unlock(inode);
1971                         return -EAGAIN;
1972                 }
1973         }
1974
1975         current->backing_dev_info = inode_to_bdi(inode);
1976         err = file_remove_privs(file);
1977         if (err) {
1978                 inode_unlock(inode);
1979                 goto out;
1980         }
1981
1982         /*
1983          * If BTRFS flips readonly due to some impossible error
1984          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1985          * although we have opened a file as writable, we have
1986          * to stop this write operation to ensure FS consistency.
1987          */
1988         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1989                 inode_unlock(inode);
1990                 err = -EROFS;
1991                 goto out;
1992         }
1993
1994         /*
1995          * We reserve space for updating the inode when we reserve space for the
1996          * extent we are going to write, so we will enospc out there.  We don't
1997          * need to start yet another transaction to update the inode as we will
1998          * update the inode when we finish writing whatever data we write.
1999          */
2000         update_time_for_write(inode);
2001
2002         start_pos = round_down(pos, fs_info->sectorsize);
2003         oldsize = i_size_read(inode);
2004         if (start_pos > oldsize) {
2005                 /* Expand hole size to cover write data, preventing empty gap */
2006                 end_pos = round_up(pos + count,
2007                                    fs_info->sectorsize);
2008                 err = btrfs_cont_expand(inode, oldsize, end_pos);
2009                 if (err) {
2010                         inode_unlock(inode);
2011                         goto out;
2012                 }
2013                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
2014                         clean_page = 1;
2015         }
2016
2017         if (sync)
2018                 atomic_inc(&BTRFS_I(inode)->sync_writers);
2019
2020         if (iocb->ki_flags & IOCB_DIRECT) {
2021                 /*
2022                  * 1. We must always clear IOCB_DSYNC in order to not deadlock
2023                  *    in iomap, as it calls generic_write_sync() in this case.
2024                  * 2. If we are async, we can call iomap_dio_complete() either
2025                  *    in
2026                  *
2027                  *    2.1. A worker thread from the last bio completed.  In this
2028                  *         case we need to mark the btrfs_dio_data that it is
2029                  *         async in order to call generic_write_sync() properly.
2030                  *         This is handled by setting BTRFS_DIO_SYNC_STUB in the
2031                  *         current->journal_info.
2032                  *    2.2  The submitter context, because all IO completed
2033                  *         before we exited iomap_dio_rw().  In this case we can
2034                  *         just re-set the IOCB_DSYNC on the iocb and we'll do
2035                  *         the sync below.  If our ->end_io() gets called and
2036                  *         current->journal_info is set, then we know we're in
2037                  *         our current context and we will clear
2038                  *         current->journal_info to indicate that we need to
2039                  *         sync below.
2040                  */
2041                 if (sync) {
2042                         ASSERT(current->journal_info == NULL);
2043                         iocb->ki_flags &= ~IOCB_DSYNC;
2044                         current->journal_info = BTRFS_DIO_SYNC_STUB;
2045                 }
2046                 num_written = __btrfs_direct_write(iocb, from);
2047
2048                 /*
2049                  * As stated above, we cleared journal_info, so we need to do
2050                  * the sync ourselves.
2051                  */
2052                 if (sync && current->journal_info == NULL)
2053                         iocb->ki_flags |= IOCB_DSYNC;
2054                 current->journal_info = NULL;
2055         } else {
2056                 num_written = btrfs_buffered_write(iocb, from);
2057                 if (num_written > 0)
2058                         iocb->ki_pos = pos + num_written;
2059                 if (clean_page)
2060                         pagecache_isize_extended(inode, oldsize,
2061                                                 i_size_read(inode));
2062         }
2063
2064         inode_unlock(inode);
2065
2066         /*
2067          * We also have to set last_sub_trans to the current log transid,
2068          * otherwise subsequent syncs to a file that's been synced in this
2069          * transaction will appear to have already occurred.
2070          */
2071         spin_lock(&BTRFS_I(inode)->lock);
2072         BTRFS_I(inode)->last_sub_trans = root->log_transid;
2073         spin_unlock(&BTRFS_I(inode)->lock);
2074         if (num_written > 0)
2075                 num_written = generic_write_sync(iocb, num_written);
2076
2077         if (sync)
2078                 atomic_dec(&BTRFS_I(inode)->sync_writers);
2079 out:
2080         current->backing_dev_info = NULL;
2081         return num_written ? num_written : err;
2082 }
2083
2084 int btrfs_release_file(struct inode *inode, struct file *filp)
2085 {
2086         struct btrfs_file_private *private = filp->private_data;
2087
2088         if (private && private->filldir_buf)
2089                 kfree(private->filldir_buf);
2090         kfree(private);
2091         filp->private_data = NULL;
2092
2093         /*
2094          * Set by setattr when we are about to truncate a file from a non-zero
2095          * size to a zero size.  This tries to flush down new bytes that may
2096          * have been written if the application were using truncate to replace
2097          * a file in place.
2098          */
2099         if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2100                                &BTRFS_I(inode)->runtime_flags))
2101                         filemap_flush(inode->i_mapping);
2102         return 0;
2103 }
2104
2105 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2106 {
2107         int ret;
2108         struct blk_plug plug;
2109
2110         /*
2111          * This is only called in fsync, which would do synchronous writes, so
2112          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2113          * multiple disks using raid profile, a large IO can be split to
2114          * several segments of stripe length (currently 64K).
2115          */
2116         blk_start_plug(&plug);
2117         atomic_inc(&BTRFS_I(inode)->sync_writers);
2118         ret = btrfs_fdatawrite_range(inode, start, end);
2119         atomic_dec(&BTRFS_I(inode)->sync_writers);
2120         blk_finish_plug(&plug);
2121
2122         return ret;
2123 }
2124
2125 /*
2126  * fsync call for both files and directories.  This logs the inode into
2127  * the tree log instead of forcing full commits whenever possible.
2128  *
2129  * It needs to call filemap_fdatawait so that all ordered extent updates are
2130  * in the metadata btree are up to date for copying to the log.
2131  *
2132  * It drops the inode mutex before doing the tree log commit.  This is an
2133  * important optimization for directories because holding the mutex prevents
2134  * new operations on the dir while we write to disk.
2135  */
2136 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2137 {
2138         struct dentry *dentry = file_dentry(file);
2139         struct inode *inode = d_inode(dentry);
2140         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2141         struct btrfs_root *root = BTRFS_I(inode)->root;
2142         struct btrfs_trans_handle *trans;
2143         struct btrfs_log_ctx ctx;
2144         int ret = 0, err;
2145         u64 len;
2146         bool full_sync;
2147
2148         trace_btrfs_sync_file(file, datasync);
2149
2150         btrfs_init_log_ctx(&ctx, inode);
2151
2152         /*
2153          * Always set the range to a full range, otherwise we can get into
2154          * several problems, from missing file extent items to represent holes
2155          * when not using the NO_HOLES feature, to log tree corruption due to
2156          * races between hole detection during logging and completion of ordered
2157          * extents outside the range, to missing checksums due to ordered extents
2158          * for which we flushed only a subset of their pages.
2159          */
2160         start = 0;
2161         end = LLONG_MAX;
2162         len = (u64)LLONG_MAX + 1;
2163
2164         /*
2165          * We write the dirty pages in the range and wait until they complete
2166          * out of the ->i_mutex. If so, we can flush the dirty pages by
2167          * multi-task, and make the performance up.  See
2168          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2169          */
2170         ret = start_ordered_ops(inode, start, end);
2171         if (ret)
2172                 goto out;
2173
2174         inode_lock(inode);
2175
2176         /*
2177          * We take the dio_sem here because the tree log stuff can race with
2178          * lockless dio writes and get an extent map logged for an extent we
2179          * never waited on.  We need it this high up for lockdep reasons.
2180          */
2181         down_write(&BTRFS_I(inode)->dio_sem);
2182
2183         atomic_inc(&root->log_batch);
2184
2185         /*
2186          * Always check for the full sync flag while holding the inode's lock,
2187          * to avoid races with other tasks. The flag must be either set all the
2188          * time during logging or always off all the time while logging.
2189          */
2190         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2191                              &BTRFS_I(inode)->runtime_flags);
2192
2193         /*
2194          * Before we acquired the inode's lock, someone may have dirtied more
2195          * pages in the target range. We need to make sure that writeback for
2196          * any such pages does not start while we are logging the inode, because
2197          * if it does, any of the following might happen when we are not doing a
2198          * full inode sync:
2199          *
2200          * 1) We log an extent after its writeback finishes but before its
2201          *    checksums are added to the csum tree, leading to -EIO errors
2202          *    when attempting to read the extent after a log replay.
2203          *
2204          * 2) We can end up logging an extent before its writeback finishes.
2205          *    Therefore after the log replay we will have a file extent item
2206          *    pointing to an unwritten extent (and no data checksums as well).
2207          *
2208          * So trigger writeback for any eventual new dirty pages and then we
2209          * wait for all ordered extents to complete below.
2210          */
2211         ret = start_ordered_ops(inode, start, end);
2212         if (ret) {
2213                 up_write(&BTRFS_I(inode)->dio_sem);
2214                 inode_unlock(inode);
2215                 goto out;
2216         }
2217
2218         /*
2219          * We have to do this here to avoid the priority inversion of waiting on
2220          * IO of a lower priority task while holding a transaction open.
2221          *
2222          * For a full fsync we wait for the ordered extents to complete while
2223          * for a fast fsync we wait just for writeback to complete, and then
2224          * attach the ordered extents to the transaction so that a transaction
2225          * commit waits for their completion, to avoid data loss if we fsync,
2226          * the current transaction commits before the ordered extents complete
2227          * and a power failure happens right after that.
2228          */
2229         if (full_sync) {
2230                 ret = btrfs_wait_ordered_range(inode, start, len);
2231         } else {
2232                 /*
2233                  * Get our ordered extents as soon as possible to avoid doing
2234                  * checksum lookups in the csum tree, and use instead the
2235                  * checksums attached to the ordered extents.
2236                  */
2237                 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2238                                                       &ctx.ordered_extents);
2239                 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2240         }
2241
2242         if (ret)
2243                 goto out_release_extents;
2244
2245         atomic_inc(&root->log_batch);
2246
2247         /*
2248          * If we are doing a fast fsync we can not bail out if the inode's
2249          * last_trans is <= then the last committed transaction, because we only
2250          * update the last_trans of the inode during ordered extent completion,
2251          * and for a fast fsync we don't wait for that, we only wait for the
2252          * writeback to complete.
2253          */
2254         smp_mb();
2255         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2256             (BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed &&
2257              (full_sync || list_empty(&ctx.ordered_extents)))) {
2258                 /*
2259                  * We've had everything committed since the last time we were
2260                  * modified so clear this flag in case it was set for whatever
2261                  * reason, it's no longer relevant.
2262                  */
2263                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2264                           &BTRFS_I(inode)->runtime_flags);
2265                 /*
2266                  * An ordered extent might have started before and completed
2267                  * already with io errors, in which case the inode was not
2268                  * updated and we end up here. So check the inode's mapping
2269                  * for any errors that might have happened since we last
2270                  * checked called fsync.
2271                  */
2272                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2273                 goto out_release_extents;
2274         }
2275
2276         /*
2277          * We use start here because we will need to wait on the IO to complete
2278          * in btrfs_sync_log, which could require joining a transaction (for
2279          * example checking cross references in the nocow path).  If we use join
2280          * here we could get into a situation where we're waiting on IO to
2281          * happen that is blocked on a transaction trying to commit.  With start
2282          * we inc the extwriter counter, so we wait for all extwriters to exit
2283          * before we start blocking joiners.  This comment is to keep somebody
2284          * from thinking they are super smart and changing this to
2285          * btrfs_join_transaction *cough*Josef*cough*.
2286          */
2287         trans = btrfs_start_transaction(root, 0);
2288         if (IS_ERR(trans)) {
2289                 ret = PTR_ERR(trans);
2290                 goto out_release_extents;
2291         }
2292
2293         ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2294         btrfs_release_log_ctx_extents(&ctx);
2295         if (ret < 0) {
2296                 /* Fallthrough and commit/free transaction. */
2297                 ret = 1;
2298         }
2299
2300         /* we've logged all the items and now have a consistent
2301          * version of the file in the log.  It is possible that
2302          * someone will come in and modify the file, but that's
2303          * fine because the log is consistent on disk, and we
2304          * have references to all of the file's extents
2305          *
2306          * It is possible that someone will come in and log the
2307          * file again, but that will end up using the synchronization
2308          * inside btrfs_sync_log to keep things safe.
2309          */
2310         up_write(&BTRFS_I(inode)->dio_sem);
2311         inode_unlock(inode);
2312
2313         if (ret != BTRFS_NO_LOG_SYNC) {
2314                 if (!ret) {
2315                         ret = btrfs_sync_log(trans, root, &ctx);
2316                         if (!ret) {
2317                                 ret = btrfs_end_transaction(trans);
2318                                 goto out;
2319                         }
2320                 }
2321                 if (!full_sync) {
2322                         ret = btrfs_wait_ordered_range(inode, start, len);
2323                         if (ret) {
2324                                 btrfs_end_transaction(trans);
2325                                 goto out;
2326                         }
2327                 }
2328                 ret = btrfs_commit_transaction(trans);
2329         } else {
2330                 ret = btrfs_end_transaction(trans);
2331         }
2332 out:
2333         ASSERT(list_empty(&ctx.list));
2334         err = file_check_and_advance_wb_err(file);
2335         if (!ret)
2336                 ret = err;
2337         return ret > 0 ? -EIO : ret;
2338
2339 out_release_extents:
2340         btrfs_release_log_ctx_extents(&ctx);
2341         up_write(&BTRFS_I(inode)->dio_sem);
2342         inode_unlock(inode);
2343         goto out;
2344 }
2345
2346 static const struct vm_operations_struct btrfs_file_vm_ops = {
2347         .fault          = filemap_fault,
2348         .map_pages      = filemap_map_pages,
2349         .page_mkwrite   = btrfs_page_mkwrite,
2350 };
2351
2352 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2353 {
2354         struct address_space *mapping = filp->f_mapping;
2355
2356         if (!mapping->a_ops->readpage)
2357                 return -ENOEXEC;
2358
2359         file_accessed(filp);
2360         vma->vm_ops = &btrfs_file_vm_ops;
2361
2362         return 0;
2363 }
2364
2365 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2366                           int slot, u64 start, u64 end)
2367 {
2368         struct btrfs_file_extent_item *fi;
2369         struct btrfs_key key;
2370
2371         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2372                 return 0;
2373
2374         btrfs_item_key_to_cpu(leaf, &key, slot);
2375         if (key.objectid != btrfs_ino(inode) ||
2376             key.type != BTRFS_EXTENT_DATA_KEY)
2377                 return 0;
2378
2379         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2380
2381         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2382                 return 0;
2383
2384         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2385                 return 0;
2386
2387         if (key.offset == end)
2388                 return 1;
2389         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2390                 return 1;
2391         return 0;
2392 }
2393
2394 static int fill_holes(struct btrfs_trans_handle *trans,
2395                 struct btrfs_inode *inode,
2396                 struct btrfs_path *path, u64 offset, u64 end)
2397 {
2398         struct btrfs_fs_info *fs_info = trans->fs_info;
2399         struct btrfs_root *root = inode->root;
2400         struct extent_buffer *leaf;
2401         struct btrfs_file_extent_item *fi;
2402         struct extent_map *hole_em;
2403         struct extent_map_tree *em_tree = &inode->extent_tree;
2404         struct btrfs_key key;
2405         int ret;
2406
2407         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2408                 goto out;
2409
2410         key.objectid = btrfs_ino(inode);
2411         key.type = BTRFS_EXTENT_DATA_KEY;
2412         key.offset = offset;
2413
2414         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2415         if (ret <= 0) {
2416                 /*
2417                  * We should have dropped this offset, so if we find it then
2418                  * something has gone horribly wrong.
2419                  */
2420                 if (ret == 0)
2421                         ret = -EINVAL;
2422                 return ret;
2423         }
2424
2425         leaf = path->nodes[0];
2426         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2427                 u64 num_bytes;
2428
2429                 path->slots[0]--;
2430                 fi = btrfs_item_ptr(leaf, path->slots[0],
2431                                     struct btrfs_file_extent_item);
2432                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2433                         end - offset;
2434                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2435                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2436                 btrfs_set_file_extent_offset(leaf, fi, 0);
2437                 btrfs_mark_buffer_dirty(leaf);
2438                 goto out;
2439         }
2440
2441         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2442                 u64 num_bytes;
2443
2444                 key.offset = offset;
2445                 btrfs_set_item_key_safe(fs_info, path, &key);
2446                 fi = btrfs_item_ptr(leaf, path->slots[0],
2447                                     struct btrfs_file_extent_item);
2448                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2449                         offset;
2450                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2451                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2452                 btrfs_set_file_extent_offset(leaf, fi, 0);
2453                 btrfs_mark_buffer_dirty(leaf);
2454                 goto out;
2455         }
2456         btrfs_release_path(path);
2457
2458         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2459                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2460         if (ret)
2461                 return ret;
2462
2463 out:
2464         btrfs_release_path(path);
2465
2466         hole_em = alloc_extent_map();
2467         if (!hole_em) {
2468                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2469                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2470         } else {
2471                 hole_em->start = offset;
2472                 hole_em->len = end - offset;
2473                 hole_em->ram_bytes = hole_em->len;
2474                 hole_em->orig_start = offset;
2475
2476                 hole_em->block_start = EXTENT_MAP_HOLE;
2477                 hole_em->block_len = 0;
2478                 hole_em->orig_block_len = 0;
2479                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2480                 hole_em->generation = trans->transid;
2481
2482                 do {
2483                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2484                         write_lock(&em_tree->lock);
2485                         ret = add_extent_mapping(em_tree, hole_em, 1);
2486                         write_unlock(&em_tree->lock);
2487                 } while (ret == -EEXIST);
2488                 free_extent_map(hole_em);
2489                 if (ret)
2490                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2491                                         &inode->runtime_flags);
2492         }
2493
2494         return 0;
2495 }
2496
2497 /*
2498  * Find a hole extent on given inode and change start/len to the end of hole
2499  * extent.(hole/vacuum extent whose em->start <= start &&
2500  *         em->start + em->len > start)
2501  * When a hole extent is found, return 1 and modify start/len.
2502  */
2503 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2504 {
2505         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2506         struct extent_map *em;
2507         int ret = 0;
2508
2509         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2510                               round_down(*start, fs_info->sectorsize),
2511                               round_up(*len, fs_info->sectorsize));
2512         if (IS_ERR(em))
2513                 return PTR_ERR(em);
2514
2515         /* Hole or vacuum extent(only exists in no-hole mode) */
2516         if (em->block_start == EXTENT_MAP_HOLE) {
2517                 ret = 1;
2518                 *len = em->start + em->len > *start + *len ?
2519                        0 : *start + *len - em->start - em->len;
2520                 *start = em->start + em->len;
2521         }
2522         free_extent_map(em);
2523         return ret;
2524 }
2525
2526 static int btrfs_punch_hole_lock_range(struct inode *inode,
2527                                        const u64 lockstart,
2528                                        const u64 lockend,
2529                                        struct extent_state **cached_state)
2530 {
2531         while (1) {
2532                 struct btrfs_ordered_extent *ordered;
2533                 int ret;
2534
2535                 truncate_pagecache_range(inode, lockstart, lockend);
2536
2537                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2538                                  cached_state);
2539                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2540                                                             lockend);
2541
2542                 /*
2543                  * We need to make sure we have no ordered extents in this range
2544                  * and nobody raced in and read a page in this range, if we did
2545                  * we need to try again.
2546                  */
2547                 if ((!ordered ||
2548                     (ordered->file_offset + ordered->num_bytes <= lockstart ||
2549                      ordered->file_offset > lockend)) &&
2550                      !filemap_range_has_page(inode->i_mapping,
2551                                              lockstart, lockend)) {
2552                         if (ordered)
2553                                 btrfs_put_ordered_extent(ordered);
2554                         break;
2555                 }
2556                 if (ordered)
2557                         btrfs_put_ordered_extent(ordered);
2558                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2559                                      lockend, cached_state);
2560                 ret = btrfs_wait_ordered_range(inode, lockstart,
2561                                                lockend - lockstart + 1);
2562                 if (ret)
2563                         return ret;
2564         }
2565         return 0;
2566 }
2567
2568 static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2569                                      struct inode *inode,
2570                                      struct btrfs_path *path,
2571                                      struct btrfs_replace_extent_info *extent_info,
2572                                      const u64 replace_len)
2573 {
2574         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2575         struct btrfs_root *root = BTRFS_I(inode)->root;
2576         struct btrfs_file_extent_item *extent;
2577         struct extent_buffer *leaf;
2578         struct btrfs_key key;
2579         int slot;
2580         struct btrfs_ref ref = { 0 };
2581         int ret;
2582
2583         if (replace_len == 0)
2584                 return 0;
2585
2586         if (extent_info->disk_offset == 0 &&
2587             btrfs_fs_incompat(fs_info, NO_HOLES))
2588                 return 0;
2589
2590         key.objectid = btrfs_ino(BTRFS_I(inode));
2591         key.type = BTRFS_EXTENT_DATA_KEY;
2592         key.offset = extent_info->file_offset;
2593         ret = btrfs_insert_empty_item(trans, root, path, &key,
2594                                       sizeof(struct btrfs_file_extent_item));
2595         if (ret)
2596                 return ret;
2597         leaf = path->nodes[0];
2598         slot = path->slots[0];
2599         write_extent_buffer(leaf, extent_info->extent_buf,
2600                             btrfs_item_ptr_offset(leaf, slot),
2601                             sizeof(struct btrfs_file_extent_item));
2602         extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2603         ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2604         btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2605         btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2606         if (extent_info->is_new_extent)
2607                 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2608         btrfs_mark_buffer_dirty(leaf);
2609         btrfs_release_path(path);
2610
2611         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
2612                         extent_info->file_offset, replace_len);
2613         if (ret)
2614                 return ret;
2615
2616         /* If it's a hole, nothing more needs to be done. */
2617         if (extent_info->disk_offset == 0)
2618                 return 0;
2619
2620         inode_add_bytes(inode, replace_len);
2621
2622         if (extent_info->is_new_extent && extent_info->insertions == 0) {
2623                 key.objectid = extent_info->disk_offset;
2624                 key.type = BTRFS_EXTENT_ITEM_KEY;
2625                 key.offset = extent_info->disk_len;
2626                 ret = btrfs_alloc_reserved_file_extent(trans, root,
2627                                                        btrfs_ino(BTRFS_I(inode)),
2628                                                        extent_info->file_offset,
2629                                                        extent_info->qgroup_reserved,
2630                                                        &key);
2631         } else {
2632                 u64 ref_offset;
2633
2634                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2635                                        extent_info->disk_offset,
2636                                        extent_info->disk_len, 0);
2637                 ref_offset = extent_info->file_offset - extent_info->data_offset;
2638                 btrfs_init_data_ref(&ref, root->root_key.objectid,
2639                                     btrfs_ino(BTRFS_I(inode)), ref_offset);
2640                 ret = btrfs_inc_extent_ref(trans, &ref);
2641         }
2642
2643         extent_info->insertions++;
2644
2645         return ret;
2646 }
2647
2648 /*
2649  * The respective range must have been previously locked, as well as the inode.
2650  * The end offset is inclusive (last byte of the range).
2651  * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2652  * the file range with an extent.
2653  * When not punching a hole, we don't want to end up in a state where we dropped
2654  * extents without inserting a new one, so we must abort the transaction to avoid
2655  * a corruption.
2656  */
2657 int btrfs_replace_file_extents(struct inode *inode, struct btrfs_path *path,
2658                            const u64 start, const u64 end,
2659                            struct btrfs_replace_extent_info *extent_info,
2660                            struct btrfs_trans_handle **trans_out)
2661 {
2662         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2663         u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2664         u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2665         struct btrfs_root *root = BTRFS_I(inode)->root;
2666         struct btrfs_trans_handle *trans = NULL;
2667         struct btrfs_block_rsv *rsv;
2668         unsigned int rsv_count;
2669         u64 cur_offset;
2670         u64 drop_end;
2671         u64 len = end - start;
2672         int ret = 0;
2673
2674         if (end <= start)
2675                 return -EINVAL;
2676
2677         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2678         if (!rsv) {
2679                 ret = -ENOMEM;
2680                 goto out;
2681         }
2682         rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2683         rsv->failfast = 1;
2684
2685         /*
2686          * 1 - update the inode
2687          * 1 - removing the extents in the range
2688          * 1 - adding the hole extent if no_holes isn't set or if we are
2689          *     replacing the range with a new extent
2690          */
2691         if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2692                 rsv_count = 3;
2693         else
2694                 rsv_count = 2;
2695
2696         trans = btrfs_start_transaction(root, rsv_count);
2697         if (IS_ERR(trans)) {
2698                 ret = PTR_ERR(trans);
2699                 trans = NULL;
2700                 goto out_free;
2701         }
2702
2703         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2704                                       min_size, false);
2705         BUG_ON(ret);
2706         trans->block_rsv = rsv;
2707
2708         cur_offset = start;
2709         while (cur_offset < end) {
2710                 ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path,
2711                                            cur_offset, end + 1, &drop_end,
2712                                            1, 0, 0, NULL);
2713                 if (ret != -ENOSPC) {
2714                         /*
2715                          * When cloning we want to avoid transaction aborts when
2716                          * nothing was done and we are attempting to clone parts
2717                          * of inline extents, in such cases -EOPNOTSUPP is
2718                          * returned by __btrfs_drop_extents() without having
2719                          * changed anything in the file.
2720                          */
2721                         if (extent_info && !extent_info->is_new_extent &&
2722                             ret && ret != -EOPNOTSUPP)
2723                                 btrfs_abort_transaction(trans, ret);
2724                         break;
2725                 }
2726
2727                 trans->block_rsv = &fs_info->trans_block_rsv;
2728
2729                 if (!extent_info && cur_offset < drop_end &&
2730                     cur_offset < ino_size) {
2731                         ret = fill_holes(trans, BTRFS_I(inode), path,
2732                                         cur_offset, drop_end);
2733                         if (ret) {
2734                                 /*
2735                                  * If we failed then we didn't insert our hole
2736                                  * entries for the area we dropped, so now the
2737                                  * fs is corrupted, so we must abort the
2738                                  * transaction.
2739                                  */
2740                                 btrfs_abort_transaction(trans, ret);
2741                                 break;
2742                         }
2743                 } else if (!extent_info && cur_offset < drop_end) {
2744                         /*
2745                          * We are past the i_size here, but since we didn't
2746                          * insert holes we need to clear the mapped area so we
2747                          * know to not set disk_i_size in this area until a new
2748                          * file extent is inserted here.
2749                          */
2750                         ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2751                                         cur_offset, drop_end - cur_offset);
2752                         if (ret) {
2753                                 /*
2754                                  * We couldn't clear our area, so we could
2755                                  * presumably adjust up and corrupt the fs, so
2756                                  * we need to abort.
2757                                  */
2758                                 btrfs_abort_transaction(trans, ret);
2759                                 break;
2760                         }
2761                 }
2762
2763                 if (extent_info && drop_end > extent_info->file_offset) {
2764                         u64 replace_len = drop_end - extent_info->file_offset;
2765
2766                         ret = btrfs_insert_replace_extent(trans, inode, path,
2767                                                         extent_info, replace_len);
2768                         if (ret) {
2769                                 btrfs_abort_transaction(trans, ret);
2770                                 break;
2771                         }
2772                         extent_info->data_len -= replace_len;
2773                         extent_info->data_offset += replace_len;
2774                         extent_info->file_offset += replace_len;
2775                 }
2776
2777                 cur_offset = drop_end;
2778
2779                 ret = btrfs_update_inode(trans, root, inode);
2780                 if (ret)
2781                         break;
2782
2783                 btrfs_end_transaction(trans);
2784                 btrfs_btree_balance_dirty(fs_info);
2785
2786                 trans = btrfs_start_transaction(root, rsv_count);
2787                 if (IS_ERR(trans)) {
2788                         ret = PTR_ERR(trans);
2789                         trans = NULL;
2790                         break;
2791                 }
2792
2793                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2794                                               rsv, min_size, false);
2795                 BUG_ON(ret);    /* shouldn't happen */
2796                 trans->block_rsv = rsv;
2797
2798                 if (!extent_info) {
2799                         ret = find_first_non_hole(inode, &cur_offset, &len);
2800                         if (unlikely(ret < 0))
2801                                 break;
2802                         if (ret && !len) {
2803                                 ret = 0;
2804                                 break;
2805                         }
2806                 }
2807         }
2808
2809         /*
2810          * If we were cloning, force the next fsync to be a full one since we
2811          * we replaced (or just dropped in the case of cloning holes when
2812          * NO_HOLES is enabled) extents and extent maps.
2813          * This is for the sake of simplicity, and cloning into files larger
2814          * than 16Mb would force the full fsync any way (when
2815          * try_release_extent_mapping() is invoked during page cache truncation.
2816          */
2817         if (extent_info && !extent_info->is_new_extent)
2818                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2819                         &BTRFS_I(inode)->runtime_flags);
2820
2821         if (ret)
2822                 goto out_trans;
2823
2824         trans->block_rsv = &fs_info->trans_block_rsv;
2825         /*
2826          * If we are using the NO_HOLES feature we might have had already an
2827          * hole that overlaps a part of the region [lockstart, lockend] and
2828          * ends at (or beyond) lockend. Since we have no file extent items to
2829          * represent holes, drop_end can be less than lockend and so we must
2830          * make sure we have an extent map representing the existing hole (the
2831          * call to __btrfs_drop_extents() might have dropped the existing extent
2832          * map representing the existing hole), otherwise the fast fsync path
2833          * will not record the existence of the hole region
2834          * [existing_hole_start, lockend].
2835          */
2836         if (drop_end <= end)
2837                 drop_end = end + 1;
2838         /*
2839          * Don't insert file hole extent item if it's for a range beyond eof
2840          * (because it's useless) or if it represents a 0 bytes range (when
2841          * cur_offset == drop_end).
2842          */
2843         if (!extent_info && cur_offset < ino_size && cur_offset < drop_end) {
2844                 ret = fill_holes(trans, BTRFS_I(inode), path,
2845                                 cur_offset, drop_end);
2846                 if (ret) {
2847                         /* Same comment as above. */
2848                         btrfs_abort_transaction(trans, ret);
2849                         goto out_trans;
2850                 }
2851         } else if (!extent_info && cur_offset < drop_end) {
2852                 /* See the comment in the loop above for the reasoning here. */
2853                 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2854                                         cur_offset, drop_end - cur_offset);
2855                 if (ret) {
2856                         btrfs_abort_transaction(trans, ret);
2857                         goto out_trans;
2858                 }
2859
2860         }
2861         if (extent_info) {
2862                 ret = btrfs_insert_replace_extent(trans, inode, path, extent_info,
2863                                                 extent_info->data_len);
2864                 if (ret) {
2865                         btrfs_abort_transaction(trans, ret);
2866                         goto out_trans;
2867                 }
2868         }
2869
2870 out_trans:
2871         if (!trans)
2872                 goto out_free;
2873
2874         trans->block_rsv = &fs_info->trans_block_rsv;
2875         if (ret)
2876                 btrfs_end_transaction(trans);
2877         else
2878                 *trans_out = trans;
2879 out_free:
2880         btrfs_free_block_rsv(fs_info, rsv);
2881 out:
2882         return ret;
2883 }
2884
2885 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2886 {
2887         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2888         struct btrfs_root *root = BTRFS_I(inode)->root;
2889         struct extent_state *cached_state = NULL;
2890         struct btrfs_path *path;
2891         struct btrfs_trans_handle *trans = NULL;
2892         u64 lockstart;
2893         u64 lockend;
2894         u64 tail_start;
2895         u64 tail_len;
2896         u64 orig_start = offset;
2897         int ret = 0;
2898         bool same_block;
2899         u64 ino_size;
2900         bool truncated_block = false;
2901         bool updated_inode = false;
2902
2903         ret = btrfs_wait_ordered_range(inode, offset, len);
2904         if (ret)
2905                 return ret;
2906
2907         inode_lock(inode);
2908         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2909         ret = find_first_non_hole(inode, &offset, &len);
2910         if (ret < 0)
2911                 goto out_only_mutex;
2912         if (ret && !len) {
2913                 /* Already in a large hole */
2914                 ret = 0;
2915                 goto out_only_mutex;
2916         }
2917
2918         lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
2919         lockend = round_down(offset + len,
2920                              btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
2921         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2922                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2923         /*
2924          * We needn't truncate any block which is beyond the end of the file
2925          * because we are sure there is no data there.
2926          */
2927         /*
2928          * Only do this if we are in the same block and we aren't doing the
2929          * entire block.
2930          */
2931         if (same_block && len < fs_info->sectorsize) {
2932                 if (offset < ino_size) {
2933                         truncated_block = true;
2934                         ret = btrfs_truncate_block(inode, offset, len, 0);
2935                 } else {
2936                         ret = 0;
2937                 }
2938                 goto out_only_mutex;
2939         }
2940
2941         /* zero back part of the first block */
2942         if (offset < ino_size) {
2943                 truncated_block = true;
2944                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2945                 if (ret) {
2946                         inode_unlock(inode);
2947                         return ret;
2948                 }
2949         }
2950
2951         /* Check the aligned pages after the first unaligned page,
2952          * if offset != orig_start, which means the first unaligned page
2953          * including several following pages are already in holes,
2954          * the extra check can be skipped */
2955         if (offset == orig_start) {
2956                 /* after truncate page, check hole again */
2957                 len = offset + len - lockstart;
2958                 offset = lockstart;
2959                 ret = find_first_non_hole(inode, &offset, &len);
2960                 if (ret < 0)
2961                         goto out_only_mutex;
2962                 if (ret && !len) {
2963                         ret = 0;
2964                         goto out_only_mutex;
2965                 }
2966                 lockstart = offset;
2967         }
2968
2969         /* Check the tail unaligned part is in a hole */
2970         tail_start = lockend + 1;
2971         tail_len = offset + len - tail_start;
2972         if (tail_len) {
2973                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2974                 if (unlikely(ret < 0))
2975                         goto out_only_mutex;
2976                 if (!ret) {
2977                         /* zero the front end of the last page */
2978                         if (tail_start + tail_len < ino_size) {
2979                                 truncated_block = true;
2980                                 ret = btrfs_truncate_block(inode,
2981                                                         tail_start + tail_len,
2982                                                         0, 1);
2983                                 if (ret)
2984                                         goto out_only_mutex;
2985                         }
2986                 }
2987         }
2988
2989         if (lockend < lockstart) {
2990                 ret = 0;
2991                 goto out_only_mutex;
2992         }
2993
2994         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2995                                           &cached_state);
2996         if (ret)
2997                 goto out_only_mutex;
2998
2999         path = btrfs_alloc_path();
3000         if (!path) {
3001                 ret = -ENOMEM;
3002                 goto out;
3003         }
3004
3005         ret = btrfs_replace_file_extents(inode, path, lockstart, lockend, NULL,
3006                                      &trans);
3007         btrfs_free_path(path);
3008         if (ret)
3009                 goto out;
3010
3011         ASSERT(trans != NULL);
3012         inode_inc_iversion(inode);
3013         inode->i_mtime = inode->i_ctime = current_time(inode);
3014         ret = btrfs_update_inode(trans, root, inode);
3015         updated_inode = true;
3016         btrfs_end_transaction(trans);
3017         btrfs_btree_balance_dirty(fs_info);
3018 out:
3019         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3020                              &cached_state);
3021 out_only_mutex:
3022         if (!updated_inode && truncated_block && !ret) {
3023                 /*
3024                  * If we only end up zeroing part of a page, we still need to
3025                  * update the inode item, so that all the time fields are
3026                  * updated as well as the necessary btrfs inode in memory fields
3027                  * for detecting, at fsync time, if the inode isn't yet in the
3028                  * log tree or it's there but not up to date.
3029                  */
3030                 struct timespec64 now = current_time(inode);
3031
3032                 inode_inc_iversion(inode);
3033                 inode->i_mtime = now;
3034                 inode->i_ctime = now;
3035                 trans = btrfs_start_transaction(root, 1);
3036                 if (IS_ERR(trans)) {
3037                         ret = PTR_ERR(trans);
3038                 } else {
3039                         int ret2;
3040
3041                         ret = btrfs_update_inode(trans, root, inode);
3042                         ret2 = btrfs_end_transaction(trans);
3043                         if (!ret)
3044                                 ret = ret2;
3045                 }
3046         }
3047         inode_unlock(inode);
3048         return ret;
3049 }
3050
3051 /* Helper structure to record which range is already reserved */
3052 struct falloc_range {
3053         struct list_head list;
3054         u64 start;
3055         u64 len;
3056 };
3057
3058 /*
3059  * Helper function to add falloc range
3060  *
3061  * Caller should have locked the larger range of extent containing
3062  * [start, len)
3063  */
3064 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3065 {
3066         struct falloc_range *prev = NULL;
3067         struct falloc_range *range = NULL;
3068
3069         if (list_empty(head))
3070                 goto insert;
3071
3072         /*
3073          * As fallocate iterate by bytenr order, we only need to check
3074          * the last range.
3075          */
3076         prev = list_entry(head->prev, struct falloc_range, list);
3077         if (prev->start + prev->len == start) {
3078                 prev->len += len;
3079                 return 0;
3080         }
3081 insert:
3082         range = kmalloc(sizeof(*range), GFP_KERNEL);
3083         if (!range)
3084                 return -ENOMEM;
3085         range->start = start;
3086         range->len = len;
3087         list_add_tail(&range->list, head);
3088         return 0;
3089 }
3090
3091 static int btrfs_fallocate_update_isize(struct inode *inode,
3092                                         const u64 end,
3093                                         const int mode)
3094 {
3095         struct btrfs_trans_handle *trans;
3096         struct btrfs_root *root = BTRFS_I(inode)->root;
3097         int ret;
3098         int ret2;
3099
3100         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3101                 return 0;
3102
3103         trans = btrfs_start_transaction(root, 1);
3104         if (IS_ERR(trans))
3105                 return PTR_ERR(trans);
3106
3107         inode->i_ctime = current_time(inode);
3108         i_size_write(inode, end);
3109         btrfs_inode_safe_disk_i_size_write(inode, 0);
3110         ret = btrfs_update_inode(trans, root, inode);
3111         ret2 = btrfs_end_transaction(trans);
3112
3113         return ret ? ret : ret2;
3114 }
3115
3116 enum {
3117         RANGE_BOUNDARY_WRITTEN_EXTENT,
3118         RANGE_BOUNDARY_PREALLOC_EXTENT,
3119         RANGE_BOUNDARY_HOLE,
3120 };
3121
3122 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3123                                                  u64 offset)
3124 {
3125         const u64 sectorsize = btrfs_inode_sectorsize(inode);
3126         struct extent_map *em;
3127         int ret;
3128
3129         offset = round_down(offset, sectorsize);
3130         em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3131         if (IS_ERR(em))
3132                 return PTR_ERR(em);
3133
3134         if (em->block_start == EXTENT_MAP_HOLE)
3135                 ret = RANGE_BOUNDARY_HOLE;
3136         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3137                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3138         else
3139                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3140
3141         free_extent_map(em);
3142         return ret;
3143 }
3144
3145 static int btrfs_zero_range(struct inode *inode,
3146                             loff_t offset,
3147                             loff_t len,
3148                             const int mode)
3149 {
3150         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3151         struct extent_map *em;
3152         struct extent_changeset *data_reserved = NULL;
3153         int ret;
3154         u64 alloc_hint = 0;
3155         const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3156         u64 alloc_start = round_down(offset, sectorsize);
3157         u64 alloc_end = round_up(offset + len, sectorsize);
3158         u64 bytes_to_reserve = 0;
3159         bool space_reserved = false;
3160
3161         inode_dio_wait(inode);
3162
3163         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3164                               alloc_end - alloc_start);
3165         if (IS_ERR(em)) {
3166                 ret = PTR_ERR(em);
3167                 goto out;
3168         }
3169
3170         /*
3171          * Avoid hole punching and extent allocation for some cases. More cases
3172          * could be considered, but these are unlikely common and we keep things
3173          * as simple as possible for now. Also, intentionally, if the target
3174          * range contains one or more prealloc extents together with regular
3175          * extents and holes, we drop all the existing extents and allocate a
3176          * new prealloc extent, so that we get a larger contiguous disk extent.
3177          */
3178         if (em->start <= alloc_start &&
3179             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3180                 const u64 em_end = em->start + em->len;
3181
3182                 if (em_end >= offset + len) {
3183                         /*
3184                          * The whole range is already a prealloc extent,
3185                          * do nothing except updating the inode's i_size if
3186                          * needed.
3187                          */
3188                         free_extent_map(em);
3189                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3190                                                            mode);
3191                         goto out;
3192                 }
3193                 /*
3194                  * Part of the range is already a prealloc extent, so operate
3195                  * only on the remaining part of the range.
3196                  */
3197                 alloc_start = em_end;
3198                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3199                 len = offset + len - alloc_start;
3200                 offset = alloc_start;
3201                 alloc_hint = em->block_start + em->len;
3202         }
3203         free_extent_map(em);
3204
3205         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3206             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3207                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3208                                       sectorsize);
3209                 if (IS_ERR(em)) {
3210                         ret = PTR_ERR(em);
3211                         goto out;
3212                 }
3213
3214                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3215                         free_extent_map(em);
3216                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3217                                                            mode);
3218                         goto out;
3219                 }
3220                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3221                         free_extent_map(em);
3222                         ret = btrfs_truncate_block(inode, offset, len, 0);
3223                         if (!ret)
3224                                 ret = btrfs_fallocate_update_isize(inode,
3225                                                                    offset + len,
3226                                                                    mode);
3227                         return ret;
3228                 }
3229                 free_extent_map(em);
3230                 alloc_start = round_down(offset, sectorsize);
3231                 alloc_end = alloc_start + sectorsize;
3232                 goto reserve_space;
3233         }
3234
3235         alloc_start = round_up(offset, sectorsize);
3236         alloc_end = round_down(offset + len, sectorsize);
3237
3238         /*
3239          * For unaligned ranges, check the pages at the boundaries, they might
3240          * map to an extent, in which case we need to partially zero them, or
3241          * they might map to a hole, in which case we need our allocation range
3242          * to cover them.
3243          */
3244         if (!IS_ALIGNED(offset, sectorsize)) {
3245                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3246                                                             offset);
3247                 if (ret < 0)
3248                         goto out;
3249                 if (ret == RANGE_BOUNDARY_HOLE) {
3250                         alloc_start = round_down(offset, sectorsize);
3251                         ret = 0;
3252                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3253                         ret = btrfs_truncate_block(inode, offset, 0, 0);
3254                         if (ret)
3255                                 goto out;
3256                 } else {
3257                         ret = 0;
3258                 }
3259         }
3260
3261         if (!IS_ALIGNED(offset + len, sectorsize)) {
3262                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3263                                                             offset + len);
3264                 if (ret < 0)
3265                         goto out;
3266                 if (ret == RANGE_BOUNDARY_HOLE) {
3267                         alloc_end = round_up(offset + len, sectorsize);
3268                         ret = 0;
3269                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3270                         ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3271                         if (ret)
3272                                 goto out;
3273                 } else {
3274                         ret = 0;
3275                 }
3276         }
3277
3278 reserve_space:
3279         if (alloc_start < alloc_end) {
3280                 struct extent_state *cached_state = NULL;
3281                 const u64 lockstart = alloc_start;
3282                 const u64 lockend = alloc_end - 1;
3283
3284                 bytes_to_reserve = alloc_end - alloc_start;
3285                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3286                                                       bytes_to_reserve);
3287                 if (ret < 0)
3288                         goto out;
3289                 space_reserved = true;
3290                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3291                                                   &cached_state);
3292                 if (ret)
3293                         goto out;
3294                 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3295                                                 alloc_start, bytes_to_reserve);
3296                 if (ret)
3297                         goto out;
3298                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3299                                                 alloc_end - alloc_start,
3300                                                 i_blocksize(inode),
3301                                                 offset + len, &alloc_hint);
3302                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3303                                      lockend, &cached_state);
3304                 /* btrfs_prealloc_file_range releases reserved space on error */
3305                 if (ret) {
3306                         space_reserved = false;
3307                         goto out;
3308                 }
3309         }
3310         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3311  out:
3312         if (ret && space_reserved)
3313                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3314                                                alloc_start, bytes_to_reserve);
3315         extent_changeset_free(data_reserved);
3316
3317         return ret;
3318 }
3319
3320 static long btrfs_fallocate(struct file *file, int mode,
3321                             loff_t offset, loff_t len)
3322 {
3323         struct inode *inode = file_inode(file);
3324         struct extent_state *cached_state = NULL;
3325         struct extent_changeset *data_reserved = NULL;
3326         struct falloc_range *range;
3327         struct falloc_range *tmp;
3328         struct list_head reserve_list;
3329         u64 cur_offset;
3330         u64 last_byte;
3331         u64 alloc_start;
3332         u64 alloc_end;
3333         u64 alloc_hint = 0;
3334         u64 locked_end;
3335         u64 actual_end = 0;
3336         struct extent_map *em;
3337         int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3338         int ret;
3339
3340         alloc_start = round_down(offset, blocksize);
3341         alloc_end = round_up(offset + len, blocksize);
3342         cur_offset = alloc_start;
3343
3344         /* Make sure we aren't being give some crap mode */
3345         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3346                      FALLOC_FL_ZERO_RANGE))
3347                 return -EOPNOTSUPP;
3348
3349         if (mode & FALLOC_FL_PUNCH_HOLE)
3350                 return btrfs_punch_hole(inode, offset, len);
3351
3352         /*
3353          * Only trigger disk allocation, don't trigger qgroup reserve
3354          *
3355          * For qgroup space, it will be checked later.
3356          */
3357         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3358                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3359                                                       alloc_end - alloc_start);
3360                 if (ret < 0)
3361                         return ret;
3362         }
3363
3364         inode_lock(inode);
3365
3366         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3367                 ret = inode_newsize_ok(inode, offset + len);
3368                 if (ret)
3369                         goto out;
3370         }
3371
3372         /*
3373          * TODO: Move these two operations after we have checked
3374          * accurate reserved space, or fallocate can still fail but
3375          * with page truncated or size expanded.
3376          *
3377          * But that's a minor problem and won't do much harm BTW.
3378          */
3379         if (alloc_start > inode->i_size) {
3380                 ret = btrfs_cont_expand(inode, i_size_read(inode),
3381                                         alloc_start);
3382                 if (ret)
3383                         goto out;
3384         } else if (offset + len > inode->i_size) {
3385                 /*
3386                  * If we are fallocating from the end of the file onward we
3387                  * need to zero out the end of the block if i_size lands in the
3388                  * middle of a block.
3389                  */
3390                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3391                 if (ret)
3392                         goto out;
3393         }
3394
3395         /*
3396          * wait for ordered IO before we have any locks.  We'll loop again
3397          * below with the locks held.
3398          */
3399         ret = btrfs_wait_ordered_range(inode, alloc_start,
3400                                        alloc_end - alloc_start);
3401         if (ret)
3402                 goto out;
3403
3404         if (mode & FALLOC_FL_ZERO_RANGE) {
3405                 ret = btrfs_zero_range(inode, offset, len, mode);
3406                 inode_unlock(inode);
3407                 return ret;
3408         }
3409
3410         locked_end = alloc_end - 1;
3411         while (1) {
3412                 struct btrfs_ordered_extent *ordered;
3413
3414                 /* the extent lock is ordered inside the running
3415                  * transaction
3416                  */
3417                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3418                                  locked_end, &cached_state);
3419                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3420                                                             locked_end);
3421
3422                 if (ordered &&
3423                     ordered->file_offset + ordered->num_bytes > alloc_start &&
3424                     ordered->file_offset < alloc_end) {
3425                         btrfs_put_ordered_extent(ordered);
3426                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3427                                              alloc_start, locked_end,
3428                                              &cached_state);
3429                         /*
3430                          * we can't wait on the range with the transaction
3431                          * running or with the extent lock held
3432                          */
3433                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3434                                                        alloc_end - alloc_start);
3435                         if (ret)
3436                                 goto out;
3437                 } else {
3438                         if (ordered)
3439                                 btrfs_put_ordered_extent(ordered);
3440                         break;
3441                 }
3442         }
3443
3444         /* First, check if we exceed the qgroup limit */
3445         INIT_LIST_HEAD(&reserve_list);
3446         while (cur_offset < alloc_end) {
3447                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3448                                       alloc_end - cur_offset);
3449                 if (IS_ERR(em)) {
3450                         ret = PTR_ERR(em);
3451                         break;
3452                 }
3453                 last_byte = min(extent_map_end(em), alloc_end);
3454                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3455                 last_byte = ALIGN(last_byte, blocksize);
3456                 if (em->block_start == EXTENT_MAP_HOLE ||
3457                     (cur_offset >= inode->i_size &&
3458                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3459                         ret = add_falloc_range(&reserve_list, cur_offset,
3460                                                last_byte - cur_offset);
3461                         if (ret < 0) {
3462                                 free_extent_map(em);
3463                                 break;
3464                         }
3465                         ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3466                                         &data_reserved, cur_offset,
3467                                         last_byte - cur_offset);
3468                         if (ret < 0) {
3469                                 cur_offset = last_byte;
3470                                 free_extent_map(em);
3471                                 break;
3472                         }
3473                 } else {
3474                         /*
3475                          * Do not need to reserve unwritten extent for this
3476                          * range, free reserved data space first, otherwise
3477                          * it'll result in false ENOSPC error.
3478                          */
3479                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3480                                 data_reserved, cur_offset,
3481                                 last_byte - cur_offset);
3482                 }
3483                 free_extent_map(em);
3484                 cur_offset = last_byte;
3485         }
3486
3487         /*
3488          * If ret is still 0, means we're OK to fallocate.
3489          * Or just cleanup the list and exit.
3490          */
3491         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3492                 if (!ret)
3493                         ret = btrfs_prealloc_file_range(inode, mode,
3494                                         range->start,
3495                                         range->len, i_blocksize(inode),
3496                                         offset + len, &alloc_hint);
3497                 else
3498                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3499                                         data_reserved, range->start,
3500                                         range->len);
3501                 list_del(&range->list);
3502                 kfree(range);
3503         }
3504         if (ret < 0)
3505                 goto out_unlock;
3506
3507         /*
3508          * We didn't need to allocate any more space, but we still extended the
3509          * size of the file so we need to update i_size and the inode item.
3510          */
3511         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3512 out_unlock:
3513         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3514                              &cached_state);
3515 out:
3516         inode_unlock(inode);
3517         /* Let go of our reservation. */
3518         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3519                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3520                                 cur_offset, alloc_end - cur_offset);
3521         extent_changeset_free(data_reserved);
3522         return ret;
3523 }
3524
3525 static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3526                                   int whence)
3527 {
3528         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3529         struct extent_map *em = NULL;
3530         struct extent_state *cached_state = NULL;
3531         loff_t i_size = inode->i_size;
3532         u64 lockstart;
3533         u64 lockend;
3534         u64 start;
3535         u64 len;
3536         int ret = 0;
3537
3538         if (i_size == 0 || offset >= i_size)
3539                 return -ENXIO;
3540
3541         /*
3542          * offset can be negative, in this case we start finding DATA/HOLE from
3543          * the very start of the file.
3544          */
3545         start = max_t(loff_t, 0, offset);
3546
3547         lockstart = round_down(start, fs_info->sectorsize);
3548         lockend = round_up(i_size, fs_info->sectorsize);
3549         if (lockend <= lockstart)
3550                 lockend = lockstart + fs_info->sectorsize;
3551         lockend--;
3552         len = lockend - lockstart + 1;
3553
3554         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3555                          &cached_state);
3556
3557         while (start < i_size) {
3558                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3559                 if (IS_ERR(em)) {
3560                         ret = PTR_ERR(em);
3561                         em = NULL;
3562                         break;
3563                 }
3564
3565                 if (whence == SEEK_HOLE &&
3566                     (em->block_start == EXTENT_MAP_HOLE ||
3567                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3568                         break;
3569                 else if (whence == SEEK_DATA &&
3570                            (em->block_start != EXTENT_MAP_HOLE &&
3571                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3572                         break;
3573
3574                 start = em->start + em->len;
3575                 free_extent_map(em);
3576                 em = NULL;
3577                 cond_resched();
3578         }
3579         free_extent_map(em);
3580         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3581                              &cached_state);
3582         if (ret) {
3583                 offset = ret;
3584         } else {
3585                 if (whence == SEEK_DATA && start >= i_size)
3586                         offset = -ENXIO;
3587                 else
3588                         offset = min_t(loff_t, start, i_size);
3589         }
3590
3591         return offset;
3592 }
3593
3594 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3595 {
3596         struct inode *inode = file->f_mapping->host;
3597
3598         switch (whence) {
3599         default:
3600                 return generic_file_llseek(file, offset, whence);
3601         case SEEK_DATA:
3602         case SEEK_HOLE:
3603                 inode_lock_shared(inode);
3604                 offset = find_desired_extent(inode, offset, whence);
3605                 inode_unlock_shared(inode);
3606                 break;
3607         }
3608
3609         if (offset < 0)
3610                 return offset;
3611
3612         return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3613 }
3614
3615 static int btrfs_file_open(struct inode *inode, struct file *filp)
3616 {
3617         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3618         return generic_file_open(inode, filp);
3619 }
3620
3621 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3622 {
3623         ssize_t ret = 0;
3624
3625         if (iocb->ki_flags & IOCB_DIRECT) {
3626                 struct inode *inode = file_inode(iocb->ki_filp);
3627
3628                 inode_lock_shared(inode);
3629                 ret = btrfs_direct_IO(iocb, to);
3630                 inode_unlock_shared(inode);
3631                 if (ret < 0 || !iov_iter_count(to) ||
3632                     iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3633                         return ret;
3634         }
3635
3636         return generic_file_buffered_read(iocb, to, ret);
3637 }
3638
3639 const struct file_operations btrfs_file_operations = {
3640         .llseek         = btrfs_file_llseek,
3641         .read_iter      = btrfs_file_read_iter,
3642         .splice_read    = generic_file_splice_read,
3643         .write_iter     = btrfs_file_write_iter,
3644         .splice_write   = iter_file_splice_write,
3645         .mmap           = btrfs_file_mmap,
3646         .open           = btrfs_file_open,
3647         .release        = btrfs_release_file,
3648         .fsync          = btrfs_sync_file,
3649         .fallocate      = btrfs_fallocate,
3650         .unlocked_ioctl = btrfs_ioctl,
3651 #ifdef CONFIG_COMPAT
3652         .compat_ioctl   = btrfs_compat_ioctl,
3653 #endif
3654         .remap_file_range = btrfs_remap_file_range,
3655 };
3656
3657 void __cold btrfs_auto_defrag_exit(void)
3658 {
3659         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3660 }
3661
3662 int __init btrfs_auto_defrag_init(void)
3663 {
3664         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3665                                         sizeof(struct inode_defrag), 0,
3666                                         SLAB_MEM_SPREAD,
3667                                         NULL);
3668         if (!btrfs_inode_defrag_cachep)
3669                 return -ENOMEM;
3670
3671         return 0;
3672 }
3673
3674 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3675 {
3676         int ret;
3677
3678         /*
3679          * So with compression we will find and lock a dirty page and clear the
3680          * first one as dirty, setup an async extent, and immediately return
3681          * with the entire range locked but with nobody actually marked with
3682          * writeback.  So we can't just filemap_write_and_wait_range() and
3683          * expect it to work since it will just kick off a thread to do the
3684          * actual work.  So we need to call filemap_fdatawrite_range _again_
3685          * since it will wait on the page lock, which won't be unlocked until
3686          * after the pages have been marked as writeback and so we're good to go
3687          * from there.  We have to do this otherwise we'll miss the ordered
3688          * extents and that results in badness.  Please Josef, do not think you
3689          * know better and pull this out at some point in the future, it is
3690          * right and you are wrong.
3691          */
3692         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3693         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3694                              &BTRFS_I(inode)->runtime_flags))
3695                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3696
3697         return ret;
3698 }