fs/btrfs/file.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/time.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/mpage.h>
  27 #include <linux/falloc.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/statfs.h>
  31 #include <linux/compat.h>
  32 #include <linux/slab.h>
  33 #include <linux/btrfs.h>
  34 #include <linux/uio.h>
  35 #include "ctree.h"
  36 #include "disk-io.h"
  37 #include "transaction.h"
  38 #include "btrfs_inode.h"
  39 #include "print-tree.h"
  40 #include "tree-log.h"
  41 #include "locking.h"
  42 #include "volumes.h"
  43 #include "qgroup.h"
  44 #include "compression.h"
  45
  46 static struct kmem_cache *btrfs_inode_defrag_cachep;
  47 /*
  48  * when auto defrag is enabled we
  49  * queue up these defrag structs to remember which
  50  * inodes need defragging passes
  51  */
  52 struct inode_defrag {
  53         struct rb_node rb_node;
  54         /* objectid */
  55         u64 ino;
  56         /*
  57          * transid where the defrag was added, we search for
  58          * extents newer than this
  59          */
  60         u64 transid;
  61
  62         /* root objectid */
  63         u64 root;
  64
  65         /* last offset we were able to defrag */
  66         u64 last_offset;
  67
  68         /* if we've wrapped around back to zero once already */
  69         int cycled;
  70 };
  71
  72 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  73                                   struct inode_defrag *defrag2)
  74 {
  75         if (defrag1->root > defrag2->root)
  76                 return 1;
  77         else if (defrag1->root < defrag2->root)
  78                 return -1;
  79         else if (defrag1->ino > defrag2->ino)
  80                 return 1;
  81         else if (defrag1->ino < defrag2->ino)
  82                 return -1;
  83         else
  84                 return 0;
  85 }
  86
  87 /* pop a record for an inode into the defrag tree.  The lock
  88  * must be held already
  89  *
  90  * If you're inserting a record for an older transid than an
  91  * existing record, the transid already in the tree is lowered
  92  *
  93  * If an existing record is found the defrag item you
  94  * pass in is freed
  95  */
  96 static int __btrfs_add_inode_defrag(struct inode *inode,
  97                                     struct inode_defrag *defrag)
  98 {
  99         struct btrfs_root *root = BTRFS_I(inode)->root;
 100         struct inode_defrag *entry;
 101         struct rb_node **p;
 102         struct rb_node *parent = NULL;
 103         int ret;
 104
 105         p = &root->fs_info->defrag_inodes.rb_node;
 106         while (*p) {
 107                 parent = *p;
 108                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 109
 110                 ret = __compare_inode_defrag(defrag, entry);
 111                 if (ret < 0)
 112                         p = &parent->rb_left;
 113                 else if (ret > 0)
 114                         p = &parent->rb_right;
 115                 else {
 116                         /* if we're reinserting an entry for
 117                          * an old defrag run, make sure to
 118                          * lower the transid of our existing record
 119                          */
 120                         if (defrag->transid < entry->transid)
 121                                 entry->transid = defrag->transid;
 122                         if (defrag->last_offset > entry->last_offset)
 123                                 entry->last_offset = defrag->last_offset;
 124                         return -EEXIST;
 125                 }
 126         }
 127         set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 128         rb_link_node(&defrag->rb_node, parent, p);
 129         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 130         return 0;
 131 }
 132
 133 static inline int __need_auto_defrag(struct btrfs_root *root)
 134 {
 135         if (!btrfs_test_opt(root->fs_info, AUTO_DEFRAG))
 136                 return 0;
 137
 138         if (btrfs_fs_closing(root->fs_info))
 139                 return 0;
 140
 141         return 1;
 142 }
 143
 144 /*
 145  * insert a defrag record for this inode if auto defrag is
 146  * enabled
 147  */
 148 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 149                            struct inode *inode)
 150 {
 151         struct btrfs_root *root = BTRFS_I(inode)->root;
 152         struct inode_defrag *defrag;
 153         u64 transid;
 154         int ret;
 155
 156         if (!__need_auto_defrag(root))
 157                 return 0;
 158
 159         if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 160                 return 0;
 161
 162         if (trans)
 163                 transid = trans->transid;
 164         else
 165                 transid = BTRFS_I(inode)->root->last_trans;
 166
 167         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 168         if (!defrag)
 169                 return -ENOMEM;
 170
 171         defrag->ino = btrfs_ino(inode);
 172         defrag->transid = transid;
 173         defrag->root = root->root_key.objectid;
 174
 175         spin_lock(&root->fs_info->defrag_inodes_lock);
 176         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
 177                 /*
 178                  * If we set IN_DEFRAG flag and evict the inode from memory,
 179                  * and then re-read this inode, this new inode doesn't have
 180                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 181                  */
 182                 ret = __btrfs_add_inode_defrag(inode, defrag);
 183                 if (ret)
 184                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 185         } else {
 186                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 187         }
 188         spin_unlock(&root->fs_info->defrag_inodes_lock);
 189         return 0;
 190 }
 191
 192 /*
 193  * Requeue the defrag object. If there is a defrag object that points to
 194  * the same inode in the tree, we will merge them together (by
 195  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 196  */
 197 static void btrfs_requeue_inode_defrag(struct inode *inode,
 198                                        struct inode_defrag *defrag)
 199 {
 200         struct btrfs_root *root = BTRFS_I(inode)->root;
 201         int ret;
 202
 203         if (!__need_auto_defrag(root))
 204                 goto out;
 205
 206         /*
 207          * Here we don't check the IN_DEFRAG flag, because we need merge
 208          * them together.
 209          */
 210         spin_lock(&root->fs_info->defrag_inodes_lock);
 211         ret = __btrfs_add_inode_defrag(inode, defrag);
 212         spin_unlock(&root->fs_info->defrag_inodes_lock);
 213         if (ret)
 214                 goto out;
 215         return;
 216 out:
 217         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 218 }
 219
 220 /*
 221  * pick the defragable inode that we want, if it doesn't exist, we will get
 222  * the next one.
 223  */
 224 static struct inode_defrag *
 225 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 226 {
 227         struct inode_defrag *entry = NULL;
 228         struct inode_defrag tmp;
 229         struct rb_node *p;
 230         struct rb_node *parent = NULL;
 231         int ret;
 232
 233         tmp.ino = ino;
 234         tmp.root = root;
 235
 236         spin_lock(&fs_info->defrag_inodes_lock);
 237         p = fs_info->defrag_inodes.rb_node;
 238         while (p) {
 239                 parent = p;
 240                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 241
 242                 ret = __compare_inode_defrag(&tmp, entry);
 243                 if (ret < 0)
 244                         p = parent->rb_left;
 245                 else if (ret > 0)
 246                         p = parent->rb_right;
 247                 else
 248                         goto out;
 249         }
 250
 251         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 252                 parent = rb_next(parent);
 253                 if (parent)
 254                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 255                 else
 256                         entry = NULL;
 257         }
 258 out:
 259         if (entry)
 260                 rb_erase(parent, &fs_info->defrag_inodes);
 261         spin_unlock(&fs_info->defrag_inodes_lock);
 262         return entry;
 263 }
 264
 265 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 266 {
 267         struct inode_defrag *defrag;
 268         struct rb_node *node;
 269
 270         spin_lock(&fs_info->defrag_inodes_lock);
 271         node = rb_first(&fs_info->defrag_inodes);
 272         while (node) {
 273                 rb_erase(node, &fs_info->defrag_inodes);
 274                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 275                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 276
 277                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 278
 279                 node = rb_first(&fs_info->defrag_inodes);
 280         }
 281         spin_unlock(&fs_info->defrag_inodes_lock);
 282 }
 283
 284 #define BTRFS_DEFRAG_BATCH      1024
 285
 286 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 287                                     struct inode_defrag *defrag)
 288 {
 289         struct btrfs_root *inode_root;
 290         struct inode *inode;
 291         struct btrfs_key key;
 292         struct btrfs_ioctl_defrag_range_args range;
 293         int num_defrag;
 294         int index;
 295         int ret;
 296
 297         /* get the inode */
 298         key.objectid = defrag->root;
 299         key.type = BTRFS_ROOT_ITEM_KEY;
 300         key.offset = (u64)-1;
 301
 302         index = srcu_read_lock(&fs_info->subvol_srcu);
 303
 304         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 305         if (IS_ERR(inode_root)) {
 306                 ret = PTR_ERR(inode_root);
 307                 goto cleanup;
 308         }
 309
 310         key.objectid = defrag->ino;
 311         key.type = BTRFS_INODE_ITEM_KEY;
 312         key.offset = 0;
 313         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 314         if (IS_ERR(inode)) {
 315                 ret = PTR_ERR(inode);
 316                 goto cleanup;
 317         }
 318         srcu_read_unlock(&fs_info->subvol_srcu, index);
 319
 320         /* do a chunk of defrag */
 321         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 322         memset(&range, 0, sizeof(range));
 323         range.len = (u64)-1;
 324         range.start = defrag->last_offset;
 325
 326         sb_start_write(fs_info->sb);
 327         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 328                                        BTRFS_DEFRAG_BATCH);
 329         sb_end_write(fs_info->sb);
 330         /*
 331          * if we filled the whole defrag batch, there
 332          * must be more work to do.  Queue this defrag
 333          * again
 334          */
 335         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 336                 defrag->last_offset = range.start;
 337                 btrfs_requeue_inode_defrag(inode, defrag);
 338         } else if (defrag->last_offset && !defrag->cycled) {
 339                 /*
 340                  * we didn't fill our defrag batch, but
 341                  * we didn't start at zero.  Make sure we loop
 342                  * around to the start of the file.
 343                  */
 344                 defrag->last_offset = 0;
 345                 defrag->cycled = 1;
 346                 btrfs_requeue_inode_defrag(inode, defrag);
 347         } else {
 348                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 349         }
 350
 351         iput(inode);
 352         return 0;
 353 cleanup:
 354         srcu_read_unlock(&fs_info->subvol_srcu, index);
 355         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 356         return ret;
 357 }
 358
 359 /*
 360  * run through the list of inodes in the FS that need
 361  * defragging
 362  */
 363 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 364 {
 365         struct inode_defrag *defrag;
 366         u64 first_ino = 0;
 367         u64 root_objectid = 0;
 368
 369         atomic_inc(&fs_info->defrag_running);
 370         while (1) {
 371                 /* Pause the auto defragger. */
 372                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 373                              &fs_info->fs_state))
 374                         break;
 375
 376                 if (!__need_auto_defrag(fs_info->tree_root))
 377                         break;
 378
 379                 /* find an inode to defrag */
 380                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 381                                                  first_ino);
 382                 if (!defrag) {
 383                         if (root_objectid || first_ino) {
 384                                 root_objectid = 0;
 385                                 first_ino = 0;
 386                                 continue;
 387                         } else {
 388                                 break;
 389                         }
 390                 }
 391
 392                 first_ino = defrag->ino + 1;
 393                 root_objectid = defrag->root;
 394
 395                 __btrfs_run_defrag_inode(fs_info, defrag);
 396         }
 397         atomic_dec(&fs_info->defrag_running);
 398
 399         /*
 400          * during unmount, we use the transaction_wait queue to
 401          * wait for the defragger to stop
 402          */
 403         wake_up(&fs_info->transaction_wait);
 404         return 0;
 405 }
 406
 407 /* simple helper to fault in pages and copy.  This should go away
 408  * and be replaced with calls into generic code.
 409  */
 410 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 411                                          struct page **prepared_pages,
 412                                          struct iov_iter *i)
 413 {
 414         size_t copied = 0;
 415         size_t total_copied = 0;
 416         int pg = 0;
 417         int offset = pos & (PAGE_SIZE - 1);
 418
 419         while (write_bytes > 0) {
 420                 size_t count = min_t(size_t,
 421                                      PAGE_SIZE - offset, write_bytes);
 422                 struct page *page = prepared_pages[pg];
 423                 /*
 424                  * Copy data from userspace to the current page
 425                  */
 426                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 427
 428                 /* Flush processor's dcache for this page */
 429                 flush_dcache_page(page);
 430
 431                 /*
 432                  * if we get a partial write, we can end up with
 433                  * partially up to date pages.  These add
 434                  * a lot of complexity, so make sure they don't
 435                  * happen by forcing this copy to be retried.
 436                  *
 437                  * The rest of the btrfs_file_write code will fall
 438                  * back to page at a time copies after we return 0.
 439                  */
 440                 if (!PageUptodate(page) && copied < count)
 441                         copied = 0;
 442
 443                 iov_iter_advance(i, copied);
 444                 write_bytes -= copied;
 445                 total_copied += copied;
 446
 447                 /* Return to btrfs_file_write_iter to fault page */
 448                 if (unlikely(copied == 0))
 449                         break;
 450
 451                 if (copied < PAGE_SIZE - offset) {
 452                         offset += copied;
 453                 } else {
 454                         pg++;
 455                         offset = 0;
 456                 }
 457         }
 458         return total_copied;
 459 }
 460
 461 /*
 462  * unlocks pages after btrfs_file_write is done with them
 463  */
 464 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 465 {
 466         size_t i;
 467         for (i = 0; i < num_pages; i++) {
 468                 /* page checked is some magic around finding pages that
 469                  * have been modified without going through btrfs_set_page_dirty
 470                  * clear it here. There should be no need to mark the pages
 471                  * accessed as prepare_pages should have marked them accessed
 472                  * in prepare_pages via find_or_create_page()
 473                  */
 474                 ClearPageChecked(pages[i]);
 475                 unlock_page(pages[i]);
 476                 put_page(pages[i]);
 477         }
 478 }
 479
 480 /*
 481  * after copy_from_user, pages need to be dirtied and we need to make
 482  * sure holes are created between the current EOF and the start of
 483  * any next extents (if required).
 484  *
 485  * this also makes the decision about creating an inline extent vs
 486  * doing real data extents, marking pages dirty and delalloc as required.
 487  */
 488 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 489                              struct page **pages, size_t num_pages,
 490                              loff_t pos, size_t write_bytes,
 491                              struct extent_state **cached)
 492 {
 493         int err = 0;
 494         int i;
 495         u64 num_bytes;
 496         u64 start_pos;
 497         u64 end_of_last_block;
 498         u64 end_pos = pos + write_bytes;
 499         loff_t isize = i_size_read(inode);
 500
 501         start_pos = pos & ~((u64)root->sectorsize - 1);
 502         num_bytes = round_up(write_bytes + pos - start_pos, root->sectorsize);
 503
 504         end_of_last_block = start_pos + num_bytes - 1;
 505         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 506                                         cached, 0);
 507         if (err)
 508                 return err;
 509
 510         for (i = 0; i < num_pages; i++) {
 511                 struct page *p = pages[i];
 512                 SetPageUptodate(p);
 513                 ClearPageChecked(p);
 514                 set_page_dirty(p);
 515         }
 516
 517         /*
 518          * we've only changed i_size in ram, and we haven't updated
 519          * the disk i_size.  There is no need to log the inode
 520          * at this time.
 521          */
 522         if (end_pos > isize)
 523                 i_size_write(inode, end_pos);
 524         return 0;
 525 }
 526
 527 /*
 528  * this drops all the extents in the cache that intersect the range
 529  * [start, end].  Existing extents are split as required.
 530  */
 531 void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 532                              int skip_pinned)
 533 {
 534         struct extent_map *em;
 535         struct extent_map *split = NULL;
 536         struct extent_map *split2 = NULL;
 537         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 538         u64 len = end - start + 1;
 539         u64 gen;
 540         int ret;
 541         int testend = 1;
 542         unsigned long flags;
 543         int compressed = 0;
 544         bool modified;
 545
 546         WARN_ON(end < start);
 547         if (end == (u64)-1) {
 548                 len = (u64)-1;
 549                 testend = 0;
 550         }
 551         while (1) {
 552                 int no_splits = 0;
 553
 554                 modified = false;
 555                 if (!split)
 556                         split = alloc_extent_map();
 557                 if (!split2)
 558                         split2 = alloc_extent_map();
 559                 if (!split || !split2)
 560                         no_splits = 1;
 561
 562                 write_lock(&em_tree->lock);
 563                 em = lookup_extent_mapping(em_tree, start, len);
 564                 if (!em) {
 565                         write_unlock(&em_tree->lock);
 566                         break;
 567                 }
 568                 flags = em->flags;
 569                 gen = em->generation;
 570                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 571                         if (testend && em->start + em->len >= start + len) {
 572                                 free_extent_map(em);
 573                                 write_unlock(&em_tree->lock);
 574                                 break;
 575                         }
 576                         start = em->start + em->len;
 577                         if (testend)
 578                                 len = start + len - (em->start + em->len);
 579                         free_extent_map(em);
 580                         write_unlock(&em_tree->lock);
 581                         continue;
 582                 }
 583                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 584                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 585                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 586                 modified = !list_empty(&em->list);
 587                 if (no_splits)
 588                         goto next;
 589
 590                 if (em->start < start) {
 591                         split->start = em->start;
 592                         split->len = start - em->start;
 593
 594                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 595                                 split->orig_start = em->orig_start;
 596                                 split->block_start = em->block_start;
 597
 598                                 if (compressed)
 599                                         split->block_len = em->block_len;
 600                                 else
 601                                         split->block_len = split->len;
 602                                 split->orig_block_len = max(split->block_len,
 603                                                 em->orig_block_len);
 604                                 split->ram_bytes = em->ram_bytes;
 605                         } else {
 606                                 split->orig_start = split->start;
 607                                 split->block_len = 0;
 608                                 split->block_start = em->block_start;
 609                                 split->orig_block_len = 0;
 610                                 split->ram_bytes = split->len;
 611                         }
 612
 613                         split->generation = gen;
 614                         split->bdev = em->bdev;
 615                         split->flags = flags;
 616                         split->compress_type = em->compress_type;
 617                         replace_extent_mapping(em_tree, em, split, modified);
 618                         free_extent_map(split);
 619                         split = split2;
 620                         split2 = NULL;
 621                 }
 622                 if (testend && em->start + em->len > start + len) {
 623                         u64 diff = start + len - em->start;
 624
 625                         split->start = start + len;
 626                         split->len = em->start + em->len - (start + len);
 627                         split->bdev = em->bdev;
 628                         split->flags = flags;
 629                         split->compress_type = em->compress_type;
 630                         split->generation = gen;
 631
 632                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 633                                 split->orig_block_len = max(em->block_len,
 634                                                     em->orig_block_len);
 635
 636                                 split->ram_bytes = em->ram_bytes;
 637                                 if (compressed) {
 638                                         split->block_len = em->block_len;
 639                                         split->block_start = em->block_start;
 640                                         split->orig_start = em->orig_start;
 641                                 } else {
 642                                         split->block_len = split->len;
 643                                         split->block_start = em->block_start
 644                                                 + diff;
 645                                         split->orig_start = em->orig_start;
 646                                 }
 647                         } else {
 648                                 split->ram_bytes = split->len;
 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                         }
 654
 655                         if (extent_map_in_tree(em)) {
 656                                 replace_extent_mapping(em_tree, em, split,
 657                                                        modified);
 658                         } else {
 659                                 ret = add_extent_mapping(em_tree, split,
 660                                                          modified);
 661                                 ASSERT(ret == 0); /* Logic error */
 662                         }
 663                         free_extent_map(split);
 664                         split = NULL;
 665                 }
 666 next:
 667                 if (extent_map_in_tree(em))
 668                         remove_extent_mapping(em_tree, em);
 669                 write_unlock(&em_tree->lock);
 670
 671                 /* once for us */
 672                 free_extent_map(em);
 673                 /* once for the tree*/
 674                 free_extent_map(em);
 675         }
 676         if (split)
 677                 free_extent_map(split);
 678         if (split2)
 679                 free_extent_map(split2);
 680 }
 681
 682 /*
 683  * this is very complex, but the basic idea is to drop all extents
 684  * in the range start - end.  hint_block is filled in with a block number
 685  * that would be a good hint to the block allocator for this file.
 686  *
 687  * If an extent intersects the range but is not entirely inside the range
 688  * it is either truncated or split.  Anything entirely inside the range
 689  * is deleted from the tree.
 690  */
 691 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 692                          struct btrfs_root *root, struct inode *inode,
 693                          struct btrfs_path *path, u64 start, u64 end,
 694                          u64 *drop_end, int drop_cache,
 695                          int replace_extent,
 696                          u32 extent_item_size,
 697                          int *key_inserted)
 698 {
 699         struct extent_buffer *leaf;
 700         struct btrfs_file_extent_item *fi;
 701         struct btrfs_key key;
 702         struct btrfs_key new_key;
 703         u64 ino = btrfs_ino(inode);
 704         u64 search_start = start;
 705         u64 disk_bytenr = 0;
 706         u64 num_bytes = 0;
 707         u64 extent_offset = 0;
 708         u64 extent_end = 0;
 709         int del_nr = 0;
 710         int del_slot = 0;
 711         int extent_type;
 712         int recow;
 713         int ret;
 714         int modify_tree = -1;
 715         int update_refs;
 716         int found = 0;
 717         int leafs_visited = 0;
 718
 719         if (drop_cache)
 720                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 721
 722         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 723                 modify_tree = 0;
 724
 725         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 726                        root == root->fs_info->tree_root);
 727         while (1) {
 728                 recow = 0;
 729                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 730                                                search_start, modify_tree);
 731                 if (ret < 0)
 732                         break;
 733                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 734                         leaf = path->nodes[0];
 735                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 736                         if (key.objectid == ino &&
 737                             key.type == BTRFS_EXTENT_DATA_KEY)
 738                                 path->slots[0]--;
 739                 }
 740                 ret = 0;
 741                 leafs_visited++;
 742 next_slot:
 743                 leaf = path->nodes[0];
 744                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 745                         BUG_ON(del_nr > 0);
 746                         ret = btrfs_next_leaf(root, path);
 747                         if (ret < 0)
 748                                 break;
 749                         if (ret > 0) {
 750                                 ret = 0;
 751                                 break;
 752                         }
 753                         leafs_visited++;
 754                         leaf = path->nodes[0];
 755                         recow = 1;
 756                 }
 757
 758                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 759
 760                 if (key.objectid > ino)
 761                         break;
 762                 if (WARN_ON_ONCE(key.objectid < ino) ||
 763                     key.type < BTRFS_EXTENT_DATA_KEY) {
 764                         ASSERT(del_nr == 0);
 765                         path->slots[0]++;
 766                         goto next_slot;
 767                 }
 768                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 769                         break;
 770
 771                 fi = btrfs_item_ptr(leaf, path->slots[0],
 772                                     struct btrfs_file_extent_item);
 773                 extent_type = btrfs_file_extent_type(leaf, fi);
 774
 775                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 776                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 777                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 778                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 779                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 780                         extent_end = key.offset +
 781                                 btrfs_file_extent_num_bytes(leaf, fi);
 782                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 783                         extent_end = key.offset +
 784                                 btrfs_file_extent_inline_len(leaf,
 785                                                      path->slots[0], fi);
 786                 } else {
 787                         /* can't happen */
 788                         BUG();
 789                 }
 790
 791                 /*
 792                  * Don't skip extent items representing 0 byte lengths. They
 793                  * used to be created (bug) if while punching holes we hit
 794                  * -ENOSPC condition. So if we find one here, just ensure we
 795                  * delete it, otherwise we would insert a new file extent item
 796                  * with the same key (offset) as that 0 bytes length file
 797                  * extent item in the call to setup_items_for_insert() later
 798                  * in this function.
 799                  */
 800                 if (extent_end == key.offset && extent_end >= search_start)
 801                         goto delete_extent_item;
 802
 803                 if (extent_end <= search_start) {
 804                         path->slots[0]++;
 805                         goto next_slot;
 806                 }
 807
 808                 found = 1;
 809                 search_start = max(key.offset, start);
 810                 if (recow || !modify_tree) {
 811                         modify_tree = -1;
 812                         btrfs_release_path(path);
 813                         continue;
 814                 }
 815
 816                 /*
 817                  *     | - range to drop - |
 818                  *  | -------- extent -------- |
 819                  */
 820                 if (start > key.offset && end < extent_end) {
 821                         BUG_ON(del_nr > 0);
 822                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 823                                 ret = -EOPNOTSUPP;
 824                                 break;
 825                         }
 826
 827                         memcpy(&new_key, &key, sizeof(new_key));
 828                         new_key.offset = start;
 829                         ret = btrfs_duplicate_item(trans, root, path,
 830                                                    &new_key);
 831                         if (ret == -EAGAIN) {
 832                                 btrfs_release_path(path);
 833                                 continue;
 834                         }
 835                         if (ret < 0)
 836                                 break;
 837
 838                         leaf = path->nodes[0];
 839                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 840                                             struct btrfs_file_extent_item);
 841                         btrfs_set_file_extent_num_bytes(leaf, fi,
 842                                                         start - key.offset);
 843
 844                         fi = btrfs_item_ptr(leaf, path->slots[0],
 845                                             struct btrfs_file_extent_item);
 846
 847                         extent_offset += start - key.offset;
 848                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 849                         btrfs_set_file_extent_num_bytes(leaf, fi,
 850                                                         extent_end - start);
 851                         btrfs_mark_buffer_dirty(leaf);
 852
 853                         if (update_refs && disk_bytenr > 0) {
 854                                 ret = btrfs_inc_extent_ref(trans, root,
 855                                                 disk_bytenr, num_bytes, 0,
 856                                                 root->root_key.objectid,
 857                                                 new_key.objectid,
 858                                                 start - extent_offset);
 859                                 BUG_ON(ret); /* -ENOMEM */
 860                         }
 861                         key.offset = start;
 862                 }
 863                 /*
 864                  *  | ---- range to drop ----- |
 865                  *      | -------- extent -------- |
 866                  */
 867                 if (start <= key.offset && end < extent_end) {
 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 = end;
 875                         btrfs_set_item_key_safe(root->fs_info, path, &new_key);
 876
 877                         extent_offset += end - key.offset;
 878                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 879                         btrfs_set_file_extent_num_bytes(leaf, fi,
 880                                                         extent_end - end);
 881                         btrfs_mark_buffer_dirty(leaf);
 882                         if (update_refs && disk_bytenr > 0)
 883                                 inode_sub_bytes(inode, end - key.offset);
 884                         break;
 885                 }
 886
 887                 search_start = extent_end;
 888                 /*
 889                  *       | ---- range to drop ----- |
 890                  *  | -------- extent -------- |
 891                  */
 892                 if (start > key.offset && end >= extent_end) {
 893                         BUG_ON(del_nr > 0);
 894                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 895                                 ret = -EOPNOTSUPP;
 896                                 break;
 897                         }
 898
 899                         btrfs_set_file_extent_num_bytes(leaf, fi,
 900                                                         start - key.offset);
 901                         btrfs_mark_buffer_dirty(leaf);
 902                         if (update_refs && disk_bytenr > 0)
 903                                 inode_sub_bytes(inode, extent_end - start);
 904                         if (end == extent_end)
 905                                 break;
 906
 907                         path->slots[0]++;
 908                         goto next_slot;
 909                 }
 910
 911                 /*
 912                  *  | ---- range to drop ----- |
 913                  *    | ------ extent ------ |
 914                  */
 915                 if (start <= key.offset && end >= extent_end) {
 916 delete_extent_item:
 917                         if (del_nr == 0) {
 918                                 del_slot = path->slots[0];
 919                                 del_nr = 1;
 920                         } else {
 921                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 922                                 del_nr++;
 923                         }
 924
 925                         if (update_refs &&
 926                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 927                                 inode_sub_bytes(inode,
 928                                                 extent_end - key.offset);
 929                                 extent_end = ALIGN(extent_end,
 930                                                    root->sectorsize);
 931                         } else if (update_refs && disk_bytenr > 0) {
 932                                 ret = btrfs_free_extent(trans, root,
 933                                                 disk_bytenr, num_bytes, 0,
 934                                                 root->root_key.objectid,
 935                                                 key.objectid, key.offset -
 936                                                 extent_offset);
 937                                 BUG_ON(ret); /* -ENOMEM */
 938                                 inode_sub_bytes(inode,
 939                                                 extent_end - key.offset);
 940                         }
 941
 942                         if (end == extent_end)
 943                                 break;
 944
 945                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 946                                 path->slots[0]++;
 947                                 goto next_slot;
 948                         }
 949
 950                         ret = btrfs_del_items(trans, root, path, del_slot,
 951                                               del_nr);
 952                         if (ret) {
 953                                 btrfs_abort_transaction(trans, ret);
 954                                 break;
 955                         }
 956
 957                         del_nr = 0;
 958                         del_slot = 0;
 959
 960                         btrfs_release_path(path);
 961                         continue;
 962                 }
 963
 964                 BUG_ON(1);
 965         }
 966
 967         if (!ret && del_nr > 0) {
 968                 /*
 969                  * Set path->slots[0] to first slot, so that after the delete
 970                  * if items are move off from our leaf to its immediate left or
 971                  * right neighbor leafs, we end up with a correct and adjusted
 972                  * path->slots[0] for our insertion (if replace_extent != 0).
 973                  */
 974                 path->slots[0] = del_slot;
 975                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 976                 if (ret)
 977                         btrfs_abort_transaction(trans, ret);
 978         }
 979
 980         leaf = path->nodes[0];
 981         /*
 982          * If btrfs_del_items() was called, it might have deleted a leaf, in
 983          * which case it unlocked our path, so check path->locks[0] matches a
 984          * write lock.
 985          */
 986         if (!ret && replace_extent && leafs_visited == 1 &&
 987             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 988              path->locks[0] == BTRFS_WRITE_LOCK) &&
 989             btrfs_leaf_free_space(root, leaf) >=
 990             sizeof(struct btrfs_item) + extent_item_size) {
 991
 992                 key.objectid = ino;
 993                 key.type = BTRFS_EXTENT_DATA_KEY;
 994                 key.offset = start;
 995                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
 996                         struct btrfs_key slot_key;
 997
 998                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
 999                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1000                                 path->slots[0]++;
1001                 }
1002                 setup_items_for_insert(root, path, &key,
1003                                        &extent_item_size,
1004                                        extent_item_size,
1005                                        sizeof(struct btrfs_item) +
1006                                        extent_item_size, 1);
1007                 *key_inserted = 1;
1008         }
1009
1010         if (!replace_extent || !(*key_inserted))
1011                 btrfs_release_path(path);
1012         if (drop_end)
1013                 *drop_end = found ? min(end, extent_end) : end;
1014         return ret;
1015 }
1016
1017 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1018                        struct btrfs_root *root, struct inode *inode, u64 start,
1019                        u64 end, int drop_cache)
1020 {
1021         struct btrfs_path *path;
1022         int ret;
1023
1024         path = btrfs_alloc_path();
1025         if (!path)
1026                 return -ENOMEM;
1027         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1028                                    drop_cache, 0, 0, NULL);
1029         btrfs_free_path(path);
1030         return ret;
1031 }
1032
1033 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1034                             u64 objectid, u64 bytenr, u64 orig_offset,
1035                             u64 *start, u64 *end)
1036 {
1037         struct btrfs_file_extent_item *fi;
1038         struct btrfs_key key;
1039         u64 extent_end;
1040
1041         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1042                 return 0;
1043
1044         btrfs_item_key_to_cpu(leaf, &key, slot);
1045         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1046                 return 0;
1047
1048         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1049         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1050             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1051             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1052             btrfs_file_extent_compression(leaf, fi) ||
1053             btrfs_file_extent_encryption(leaf, fi) ||
1054             btrfs_file_extent_other_encoding(leaf, fi))
1055                 return 0;
1056
1057         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1058         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1059                 return 0;
1060
1061         *start = key.offset;
1062         *end = extent_end;
1063         return 1;
1064 }
1065
1066 /*
1067  * Mark extent in the range start - end as written.
1068  *
1069  * This changes extent type from 'pre-allocated' to 'regular'. If only
1070  * part of extent is marked as written, the extent will be split into
1071  * two or three.
1072  */
1073 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1074                               struct inode *inode, u64 start, u64 end)
1075 {
1076         struct btrfs_root *root = BTRFS_I(inode)->root;
1077         struct extent_buffer *leaf;
1078         struct btrfs_path *path;
1079         struct btrfs_file_extent_item *fi;
1080         struct btrfs_key key;
1081         struct btrfs_key new_key;
1082         u64 bytenr;
1083         u64 num_bytes;
1084         u64 extent_end;
1085         u64 orig_offset;
1086         u64 other_start;
1087         u64 other_end;
1088         u64 split;
1089         int del_nr = 0;
1090         int del_slot = 0;
1091         int recow;
1092         int ret = 0;
1093         u64 ino = btrfs_ino(inode);
1094
1095         path = btrfs_alloc_path();
1096         if (!path)
1097                 return -ENOMEM;
1098 again:
1099         recow = 0;
1100         split = start;
1101         key.objectid = ino;
1102         key.type = BTRFS_EXTENT_DATA_KEY;
1103         key.offset = split;
1104
1105         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1106         if (ret < 0)
1107                 goto out;
1108         if (ret > 0 && path->slots[0] > 0)
1109                 path->slots[0]--;
1110
1111         leaf = path->nodes[0];
1112         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1113         if (key.objectid != ino ||
1114             key.type != BTRFS_EXTENT_DATA_KEY) {
1115                 ret = -EINVAL;
1116                 btrfs_abort_transaction(trans, ret);
1117                 goto out;
1118         }
1119         fi = btrfs_item_ptr(leaf, path->slots[0],
1120                             struct btrfs_file_extent_item);
1121         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1122                 ret = -EINVAL;
1123                 btrfs_abort_transaction(trans, ret);
1124                 goto out;
1125         }
1126         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1127         if (key.offset > start || extent_end < end) {
1128                 ret = -EINVAL;
1129                 btrfs_abort_transaction(trans, ret);
1130                 goto out;
1131         }
1132
1133         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1134         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1135         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1136         memcpy(&new_key, &key, sizeof(new_key));
1137
1138         if (start == key.offset && end < extent_end) {
1139                 other_start = 0;
1140                 other_end = start;
1141                 if (extent_mergeable(leaf, path->slots[0] - 1,
1142                                      ino, bytenr, orig_offset,
1143                                      &other_start, &other_end)) {
1144                         new_key.offset = end;
1145                         btrfs_set_item_key_safe(root->fs_info, path, &new_key);
1146                         fi = btrfs_item_ptr(leaf, path->slots[0],
1147                                             struct btrfs_file_extent_item);
1148                         btrfs_set_file_extent_generation(leaf, fi,
1149                                                          trans->transid);
1150                         btrfs_set_file_extent_num_bytes(leaf, fi,
1151                                                         extent_end - end);
1152                         btrfs_set_file_extent_offset(leaf, fi,
1153                                                      end - orig_offset);
1154                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1155                                             struct btrfs_file_extent_item);
1156                         btrfs_set_file_extent_generation(leaf, fi,
1157                                                          trans->transid);
1158                         btrfs_set_file_extent_num_bytes(leaf, fi,
1159                                                         end - other_start);
1160                         btrfs_mark_buffer_dirty(leaf);
1161                         goto out;
1162                 }
1163         }
1164
1165         if (start > key.offset && end == extent_end) {
1166                 other_start = end;
1167                 other_end = 0;
1168                 if (extent_mergeable(leaf, path->slots[0] + 1,
1169                                      ino, bytenr, orig_offset,
1170                                      &other_start, &other_end)) {
1171                         fi = btrfs_item_ptr(leaf, path->slots[0],
1172                                             struct btrfs_file_extent_item);
1173                         btrfs_set_file_extent_num_bytes(leaf, fi,
1174                                                         start - key.offset);
1175                         btrfs_set_file_extent_generation(leaf, fi,
1176                                                          trans->transid);
1177                         path->slots[0]++;
1178                         new_key.offset = start;
1179                         btrfs_set_item_key_safe(root->fs_info, path, &new_key);
1180
1181                         fi = btrfs_item_ptr(leaf, path->slots[0],
1182                                             struct btrfs_file_extent_item);
1183                         btrfs_set_file_extent_generation(leaf, fi,
1184                                                          trans->transid);
1185                         btrfs_set_file_extent_num_bytes(leaf, fi,
1186                                                         other_end - start);
1187                         btrfs_set_file_extent_offset(leaf, fi,
1188                                                      start - orig_offset);
1189                         btrfs_mark_buffer_dirty(leaf);
1190                         goto out;
1191                 }
1192         }
1193
1194         while (start > key.offset || end < extent_end) {
1195                 if (key.offset == start)
1196                         split = end;
1197
1198                 new_key.offset = split;
1199                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1200                 if (ret == -EAGAIN) {
1201                         btrfs_release_path(path);
1202                         goto again;
1203                 }
1204                 if (ret < 0) {
1205                         btrfs_abort_transaction(trans, ret);
1206                         goto out;
1207                 }
1208
1209                 leaf = path->nodes[0];
1210                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1211                                     struct btrfs_file_extent_item);
1212                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1213                 btrfs_set_file_extent_num_bytes(leaf, fi,
1214                                                 split - key.offset);
1215
1216                 fi = btrfs_item_ptr(leaf, path->slots[0],
1217                                     struct btrfs_file_extent_item);
1218
1219                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1220                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1221                 btrfs_set_file_extent_num_bytes(leaf, fi,
1222                                                 extent_end - split);
1223                 btrfs_mark_buffer_dirty(leaf);
1224
1225                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1226                                            root->root_key.objectid,
1227                                            ino, orig_offset);
1228                 if (ret) {
1229                         btrfs_abort_transaction(trans, ret);
1230                         goto out;
1231                 }
1232
1233                 if (split == start) {
1234                         key.offset = start;
1235                 } else {
1236                         if (start != key.offset) {
1237                                 ret = -EINVAL;
1238                                 btrfs_abort_transaction(trans, ret);
1239                                 goto out;
1240                         }
1241                         path->slots[0]--;
1242                         extent_end = end;
1243                 }
1244                 recow = 1;
1245         }
1246
1247         other_start = end;
1248         other_end = 0;
1249         if (extent_mergeable(leaf, path->slots[0] + 1,
1250                              ino, bytenr, orig_offset,
1251                              &other_start, &other_end)) {
1252                 if (recow) {
1253                         btrfs_release_path(path);
1254                         goto again;
1255                 }
1256                 extent_end = other_end;
1257                 del_slot = path->slots[0] + 1;
1258                 del_nr++;
1259                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1260                                         0, root->root_key.objectid,
1261                                         ino, orig_offset);
1262                 if (ret) {
1263                         btrfs_abort_transaction(trans, ret);
1264                         goto out;
1265                 }
1266         }
1267         other_start = 0;
1268         other_end = start;
1269         if (extent_mergeable(leaf, path->slots[0] - 1,
1270                              ino, bytenr, orig_offset,
1271                              &other_start, &other_end)) {
1272                 if (recow) {
1273                         btrfs_release_path(path);
1274                         goto again;
1275                 }
1276                 key.offset = other_start;
1277                 del_slot = path->slots[0];
1278                 del_nr++;
1279                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1280                                         0, root->root_key.objectid,
1281                                         ino, orig_offset);
1282                 if (ret) {
1283                         btrfs_abort_transaction(trans, ret);
1284                         goto out;
1285                 }
1286         }
1287         if (del_nr == 0) {
1288                 fi = btrfs_item_ptr(leaf, path->slots[0],
1289                            struct btrfs_file_extent_item);
1290                 btrfs_set_file_extent_type(leaf, fi,
1291                                            BTRFS_FILE_EXTENT_REG);
1292                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1293                 btrfs_mark_buffer_dirty(leaf);
1294         } else {
1295                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1296                            struct btrfs_file_extent_item);
1297                 btrfs_set_file_extent_type(leaf, fi,
1298                                            BTRFS_FILE_EXTENT_REG);
1299                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1300                 btrfs_set_file_extent_num_bytes(leaf, fi,
1301                                                 extent_end - key.offset);
1302                 btrfs_mark_buffer_dirty(leaf);
1303
1304                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1305                 if (ret < 0) {
1306                         btrfs_abort_transaction(trans, ret);
1307                         goto out;
1308                 }
1309         }
1310 out:
1311         btrfs_free_path(path);
1312         return ret;
1313 }
1314
1315 /*
1316  * on error we return an unlocked page and the error value
1317  * on success we return a locked page and 0
1318  */
1319 static int prepare_uptodate_page(struct inode *inode,
1320                                  struct page *page, u64 pos,
1321                                  bool force_uptodate)
1322 {
1323         int ret = 0;
1324
1325         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1326             !PageUptodate(page)) {
1327                 ret = btrfs_readpage(NULL, page);
1328                 if (ret)
1329                         return ret;
1330                 lock_page(page);
1331                 if (!PageUptodate(page)) {
1332                         unlock_page(page);
1333                         return -EIO;
1334                 }
1335                 if (page->mapping != inode->i_mapping) {
1336                         unlock_page(page);
1337                         return -EAGAIN;
1338                 }
1339         }
1340         return 0;
1341 }
1342
1343 /*
1344  * this just gets pages into the page cache and locks them down.
1345  */
1346 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1347                                   size_t num_pages, loff_t pos,
1348                                   size_t write_bytes, bool force_uptodate)
1349 {
1350         int i;
1351         unsigned long index = pos >> PAGE_SHIFT;
1352         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1353         int err = 0;
1354         int faili;
1355
1356         for (i = 0; i < num_pages; i++) {
1357 again:
1358                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1359                                                mask | __GFP_WRITE);
1360                 if (!pages[i]) {
1361                         faili = i - 1;
1362                         err = -ENOMEM;
1363                         goto fail;
1364                 }
1365
1366                 if (i == 0)
1367                         err = prepare_uptodate_page(inode, pages[i], pos,
1368                                                     force_uptodate);
1369                 if (!err && i == num_pages - 1)
1370                         err = prepare_uptodate_page(inode, pages[i],
1371                                                     pos + write_bytes, false);
1372                 if (err) {
1373                         put_page(pages[i]);
1374                         if (err == -EAGAIN) {
1375                                 err = 0;
1376                                 goto again;
1377                         }
1378                         faili = i - 1;
1379                         goto fail;
1380                 }
1381                 wait_on_page_writeback(pages[i]);
1382         }
1383
1384         return 0;
1385 fail:
1386         while (faili >= 0) {
1387                 unlock_page(pages[faili]);
1388                 put_page(pages[faili]);
1389                 faili--;
1390         }
1391         return err;
1392
1393 }
1394
1395 /*
1396  * This function locks the extent and properly waits for data=ordered extents
1397  * to finish before allowing the pages to be modified if need.
1398  *
1399  * The return value:
1400  * 1 - the extent is locked
1401  * 0 - the extent is not locked, and everything is OK
1402  * -EAGAIN - need re-prepare the pages
1403  * the other < 0 number - Something wrong happens
1404  */
1405 static noinline int
1406 lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1407                                 size_t num_pages, loff_t pos,
1408                                 size_t write_bytes,
1409                                 u64 *lockstart, u64 *lockend,
1410                                 struct extent_state **cached_state)
1411 {
1412         struct btrfs_root *root = BTRFS_I(inode)->root;
1413         u64 start_pos;
1414         u64 last_pos;
1415         int i;
1416         int ret = 0;
1417
1418         start_pos = round_down(pos, root->sectorsize);
1419         last_pos = start_pos
1420                 + round_up(pos + write_bytes - start_pos, root->sectorsize) - 1;
1421
1422         if (start_pos < inode->i_size) {
1423                 struct btrfs_ordered_extent *ordered;
1424                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1425                                  start_pos, last_pos, cached_state);
1426                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1427                                                      last_pos - start_pos + 1);
1428                 if (ordered &&
1429                     ordered->file_offset + ordered->len > start_pos &&
1430                     ordered->file_offset <= last_pos) {
1431                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1432                                              start_pos, last_pos,
1433                                              cached_state, GFP_NOFS);
1434                         for (i = 0; i < num_pages; i++) {
1435                                 unlock_page(pages[i]);
1436                                 put_page(pages[i]);
1437                         }
1438                         btrfs_start_ordered_extent(inode, ordered, 1);
1439                         btrfs_put_ordered_extent(ordered);
1440                         return -EAGAIN;
1441                 }
1442                 if (ordered)
1443                         btrfs_put_ordered_extent(ordered);
1444
1445                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1446                                   last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1447                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1448                                   0, 0, cached_state, GFP_NOFS);
1449                 *lockstart = start_pos;
1450                 *lockend = last_pos;
1451                 ret = 1;
1452         }
1453
1454         for (i = 0; i < num_pages; i++) {
1455                 if (clear_page_dirty_for_io(pages[i]))
1456                         account_page_redirty(pages[i]);
1457                 set_page_extent_mapped(pages[i]);
1458                 WARN_ON(!PageLocked(pages[i]));
1459         }
1460
1461         return ret;
1462 }
1463
1464 static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1465                                     size_t *write_bytes)
1466 {
1467         struct btrfs_root *root = BTRFS_I(inode)->root;
1468         struct btrfs_ordered_extent *ordered;
1469         u64 lockstart, lockend;
1470         u64 num_bytes;
1471         int ret;
1472
1473         ret = btrfs_start_write_no_snapshoting(root);
1474         if (!ret)
1475                 return -ENOSPC;
1476
1477         lockstart = round_down(pos, root->sectorsize);
1478         lockend = round_up(pos + *write_bytes, root->sectorsize) - 1;
1479
1480         while (1) {
1481                 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1482                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1483                                                      lockend - lockstart + 1);
1484                 if (!ordered) {
1485                         break;
1486                 }
1487                 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1488                 btrfs_start_ordered_extent(inode, ordered, 1);
1489                 btrfs_put_ordered_extent(ordered);
1490         }
1491
1492         num_bytes = lockend - lockstart + 1;
1493         ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1494         if (ret <= 0) {
1495                 ret = 0;
1496                 btrfs_end_write_no_snapshoting(root);
1497         } else {
1498                 *write_bytes = min_t(size_t, *write_bytes ,
1499                                      num_bytes - pos + lockstart);
1500         }
1501
1502         unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1503
1504         return ret;
1505 }
1506
1507 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1508                                                struct iov_iter *i,
1509                                                loff_t pos)
1510 {
1511         struct inode *inode = file_inode(file);
1512         struct btrfs_root *root = BTRFS_I(inode)->root;
1513         struct page **pages = NULL;
1514         struct extent_state *cached_state = NULL;
1515         u64 release_bytes = 0;
1516         u64 lockstart;
1517         u64 lockend;
1518         size_t num_written = 0;
1519         int nrptrs;
1520         int ret = 0;
1521         bool only_release_metadata = false;
1522         bool force_page_uptodate = false;
1523         bool need_unlock;
1524
1525         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1526                         PAGE_SIZE / (sizeof(struct page *)));
1527         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1528         nrptrs = max(nrptrs, 8);
1529         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1530         if (!pages)
1531                 return -ENOMEM;
1532
1533         while (iov_iter_count(i) > 0) {
1534                 size_t offset = pos & (PAGE_SIZE - 1);
1535                 size_t sector_offset;
1536                 size_t write_bytes = min(iov_iter_count(i),
1537                                          nrptrs * (size_t)PAGE_SIZE -
1538                                          offset);
1539                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1540                                                 PAGE_SIZE);
1541                 size_t reserve_bytes;
1542                 size_t dirty_pages;
1543                 size_t copied;
1544                 size_t dirty_sectors;
1545                 size_t num_sectors;
1546
1547                 WARN_ON(num_pages > nrptrs);
1548
1549                 /*
1550                  * Fault pages before locking them in prepare_pages
1551                  * to avoid recursive lock
1552                  */
1553                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1554                         ret = -EFAULT;
1555                         break;
1556                 }
1557
1558                 only_release_metadata = false;
1559                 sector_offset = pos & (root->sectorsize - 1);
1560                 reserve_bytes = round_up(write_bytes + sector_offset,
1561                                 root->sectorsize);
1562
1563                 ret = btrfs_check_data_free_space(inode, pos, write_bytes);
1564                 if (ret < 0) {
1565                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1566                                                       BTRFS_INODE_PREALLOC)) &&
1567                             check_can_nocow(inode, pos, &write_bytes) > 0) {
1568                                 /*
1569                                  * For nodata cow case, no need to reserve
1570                                  * data space.
1571                                  */
1572                                 only_release_metadata = true;
1573                                 /*
1574                                  * our prealloc extent may be smaller than
1575                                  * write_bytes, so scale down.
1576                                  */
1577                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1578                                                          PAGE_SIZE);
1579                                 reserve_bytes = round_up(write_bytes +
1580                                                          sector_offset,
1581                                                          root->sectorsize);
1582                         } else {
1583                                 break;
1584                         }
1585                 }
1586
1587                 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1588                 if (ret) {
1589                         if (!only_release_metadata)
1590                                 btrfs_free_reserved_data_space(inode, pos,
1591                                                                write_bytes);
1592                         else
1593                                 btrfs_end_write_no_snapshoting(root);
1594                         break;
1595                 }
1596
1597                 release_bytes = reserve_bytes;
1598                 need_unlock = false;
1599 again:
1600                 /*
1601                  * This is going to setup the pages array with the number of
1602                  * pages we want, so we don't really need to worry about the
1603                  * contents of pages from loop to loop
1604                  */
1605                 ret = prepare_pages(inode, pages, num_pages,
1606                                     pos, write_bytes,
1607                                     force_page_uptodate);
1608                 if (ret)
1609                         break;
1610
1611                 ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1612                                                 pos, write_bytes, &lockstart,
1613                                                 &lockend, &cached_state);
1614                 if (ret < 0) {
1615                         if (ret == -EAGAIN)
1616                                 goto again;
1617                         break;
1618                 } else if (ret > 0) {
1619                         need_unlock = true;
1620                         ret = 0;
1621                 }
1622
1623                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1624
1625                 num_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
1626                                                 reserve_bytes);
1627                 dirty_sectors = round_up(copied + sector_offset,
1628                                         root->sectorsize);
1629                 dirty_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
1630                                                 dirty_sectors);
1631
1632                 /*
1633                  * if we have trouble faulting in the pages, fall
1634                  * back to one page at a time
1635                  */
1636                 if (copied < write_bytes)
1637                         nrptrs = 1;
1638
1639                 if (copied == 0) {
1640                         force_page_uptodate = true;
1641                         dirty_sectors = 0;
1642                         dirty_pages = 0;
1643                 } else {
1644                         force_page_uptodate = false;
1645                         dirty_pages = DIV_ROUND_UP(copied + offset,
1646                                                    PAGE_SIZE);
1647                 }
1648
1649                 /*
1650                  * If we had a short copy we need to release the excess delaloc
1651                  * bytes we reserved.  We need to increment outstanding_extents
1652                  * because btrfs_delalloc_release_space and
1653                  * btrfs_delalloc_release_metadata will decrement it, but
1654                  * we still have an outstanding extent for the chunk we actually
1655                  * managed to copy.
1656                  */
1657                 if (num_sectors > dirty_sectors) {
1658
1659                         /* release everything except the sectors we dirtied */
1660                         release_bytes -= dirty_sectors <<
1661                                 root->fs_info->sb->s_blocksize_bits;
1662
1663                         if (copied > 0) {
1664                                 spin_lock(&BTRFS_I(inode)->lock);
1665                                 BTRFS_I(inode)->outstanding_extents++;
1666                                 spin_unlock(&BTRFS_I(inode)->lock);
1667                         }
1668                         if (only_release_metadata) {
1669                                 btrfs_delalloc_release_metadata(inode,
1670                                                                 release_bytes);
1671                         } else {
1672                                 u64 __pos;
1673
1674                                 __pos = round_down(pos, root->sectorsize) +
1675                                         (dirty_pages << PAGE_SHIFT);
1676                                 btrfs_delalloc_release_space(inode, __pos,
1677                                                              release_bytes);
1678                         }
1679                 }
1680
1681                 release_bytes = round_up(copied + sector_offset,
1682                                         root->sectorsize);
1683
1684                 if (copied > 0)
1685                         ret = btrfs_dirty_pages(root, inode, pages,
1686                                                 dirty_pages, pos, copied,
1687                                                 NULL);
1688                 if (need_unlock)
1689                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1690                                              lockstart, lockend, &cached_state,
1691                                              GFP_NOFS);
1692                 if (ret) {
1693                         btrfs_drop_pages(pages, num_pages);
1694                         break;
1695                 }
1696
1697                 release_bytes = 0;
1698                 if (only_release_metadata)
1699                         btrfs_end_write_no_snapshoting(root);
1700
1701                 if (only_release_metadata && copied > 0) {
1702                         lockstart = round_down(pos, root->sectorsize);
1703                         lockend = round_up(pos + copied, root->sectorsize) - 1;
1704
1705                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1706                                        lockend, EXTENT_NORESERVE, NULL,
1707                                        NULL, GFP_NOFS);
1708                 }
1709
1710                 btrfs_drop_pages(pages, num_pages);
1711
1712                 cond_resched();
1713
1714                 balance_dirty_pages_ratelimited(inode->i_mapping);
1715                 if (dirty_pages < (root->nodesize >> PAGE_SHIFT) + 1)
1716                         btrfs_btree_balance_dirty(root);
1717
1718                 pos += copied;
1719                 num_written += copied;
1720         }
1721
1722         kfree(pages);
1723
1724         if (release_bytes) {
1725                 if (only_release_metadata) {
1726                         btrfs_end_write_no_snapshoting(root);
1727                         btrfs_delalloc_release_metadata(inode, release_bytes);
1728                 } else {
1729                         btrfs_delalloc_release_space(inode,
1730                                                 round_down(pos, root->sectorsize),
1731                                                 release_bytes);
1732                 }
1733         }
1734
1735         return num_written ? num_written : ret;
1736 }
1737
1738 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1739 {
1740         struct file *file = iocb->ki_filp;
1741         struct inode *inode = file_inode(file);
1742         loff_t pos = iocb->ki_pos;
1743         ssize_t written;
1744         ssize_t written_buffered;
1745         loff_t endbyte;
1746         int err;
1747
1748         written = generic_file_direct_write(iocb, from);
1749
1750         if (written < 0 || !iov_iter_count(from))
1751                 return written;
1752
1753         pos += written;
1754         written_buffered = __btrfs_buffered_write(file, from, pos);
1755         if (written_buffered < 0) {
1756                 err = written_buffered;
1757                 goto out;
1758         }
1759         /*
1760          * Ensure all data is persisted. We want the next direct IO read to be
1761          * able to read what was just written.
1762          */
1763         endbyte = pos + written_buffered - 1;
1764         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1765         if (err)
1766                 goto out;
1767         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1768         if (err)
1769                 goto out;
1770         written += written_buffered;
1771         iocb->ki_pos = pos + written_buffered;
1772         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1773                                  endbyte >> PAGE_SHIFT);
1774 out:
1775         return written ? written : err;
1776 }
1777
1778 static void update_time_for_write(struct inode *inode)
1779 {
1780         struct timespec now;
1781
1782         if (IS_NOCMTIME(inode))
1783                 return;
1784
1785         now = current_time(inode);
1786         if (!timespec_equal(&inode->i_mtime, &now))
1787                 inode->i_mtime = now;
1788
1789         if (!timespec_equal(&inode->i_ctime, &now))
1790                 inode->i_ctime = now;
1791
1792         if (IS_I_VERSION(inode))
1793                 inode_inc_iversion(inode);
1794 }
1795
1796 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1797                                     struct iov_iter *from)
1798 {
1799         struct file *file = iocb->ki_filp;
1800         struct inode *inode = file_inode(file);
1801         struct btrfs_root *root = BTRFS_I(inode)->root;
1802         u64 start_pos;
1803         u64 end_pos;
1804         ssize_t num_written = 0;
1805         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1806         ssize_t err;
1807         loff_t pos;
1808         size_t count;
1809         loff_t oldsize;
1810         int clean_page = 0;
1811
1812         inode_lock(inode);
1813         err = generic_write_checks(iocb, from);
1814         if (err <= 0) {
1815                 inode_unlock(inode);
1816                 return err;
1817         }
1818
1819         current->backing_dev_info = inode_to_bdi(inode);
1820         err = file_remove_privs(file);
1821         if (err) {
1822                 inode_unlock(inode);
1823                 goto out;
1824         }
1825
1826         /*
1827          * If BTRFS flips readonly due to some impossible error
1828          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1829          * although we have opened a file as writable, we have
1830          * to stop this write operation to ensure FS consistency.
1831          */
1832         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1833                 inode_unlock(inode);
1834                 err = -EROFS;
1835                 goto out;
1836         }
1837
1838         /*
1839          * We reserve space for updating the inode when we reserve space for the
1840          * extent we are going to write, so we will enospc out there.  We don't
1841          * need to start yet another transaction to update the inode as we will
1842          * update the inode when we finish writing whatever data we write.
1843          */
1844         update_time_for_write(inode);
1845
1846         pos = iocb->ki_pos;
1847         count = iov_iter_count(from);
1848         start_pos = round_down(pos, root->sectorsize);
1849         oldsize = i_size_read(inode);
1850         if (start_pos > oldsize) {
1851                 /* Expand hole size to cover write data, preventing empty gap */
1852                 end_pos = round_up(pos + count, root->sectorsize);
1853                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1854                 if (err) {
1855                         inode_unlock(inode);
1856                         goto out;
1857                 }
1858                 if (start_pos > round_up(oldsize, root->sectorsize))
1859                         clean_page = 1;
1860         }
1861
1862         if (sync)
1863                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1864
1865         if (iocb->ki_flags & IOCB_DIRECT) {
1866                 num_written = __btrfs_direct_write(iocb, from);
1867         } else {
1868                 num_written = __btrfs_buffered_write(file, from, pos);
1869                 if (num_written > 0)
1870                         iocb->ki_pos = pos + num_written;
1871                 if (clean_page)
1872                         pagecache_isize_extended(inode, oldsize,
1873                                                 i_size_read(inode));
1874         }
1875
1876         inode_unlock(inode);
1877
1878         /*
1879          * We also have to set last_sub_trans to the current log transid,
1880          * otherwise subsequent syncs to a file that's been synced in this
1881          * transaction will appear to have already occurred.
1882          */
1883         spin_lock(&BTRFS_I(inode)->lock);
1884         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1885         spin_unlock(&BTRFS_I(inode)->lock);
1886         if (num_written > 0)
1887                 num_written = generic_write_sync(iocb, num_written);
1888
1889         if (sync)
1890                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1891 out:
1892         current->backing_dev_info = NULL;
1893         return num_written ? num_written : err;
1894 }
1895
1896 int btrfs_release_file(struct inode *inode, struct file *filp)
1897 {
1898         if (filp->private_data)
1899                 btrfs_ioctl_trans_end(filp);
1900         /*
1901          * ordered_data_close is set by settattr when we are about to truncate
1902          * a file from a non-zero size to a zero size.  This tries to
1903          * flush down new bytes that may have been written if the
1904          * application were using truncate to replace a file in place.
1905          */
1906         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1907                                &BTRFS_I(inode)->runtime_flags))
1908                         filemap_flush(inode->i_mapping);
1909         return 0;
1910 }
1911
1912 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1913 {
1914         int ret;
1915         struct blk_plug plug;
1916
1917         /*
1918          * This is only called in fsync, which would do synchronous writes, so
1919          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
1920          * multiple disks using raid profile, a large IO can be split to
1921          * several segments of stripe length (currently 64K).
1922          */
1923         blk_start_plug(&plug);
1924         atomic_inc(&BTRFS_I(inode)->sync_writers);
1925         ret = btrfs_fdatawrite_range(inode, start, end);
1926         atomic_dec(&BTRFS_I(inode)->sync_writers);
1927         blk_finish_plug(&plug);
1928
1929         return ret;
1930 }
1931
1932 /*
1933  * fsync call for both files and directories.  This logs the inode into
1934  * the tree log instead of forcing full commits whenever possible.
1935  *
1936  * It needs to call filemap_fdatawait so that all ordered extent updates are
1937  * in the metadata btree are up to date for copying to the log.
1938  *
1939  * It drops the inode mutex before doing the tree log commit.  This is an
1940  * important optimization for directories because holding the mutex prevents
1941  * new operations on the dir while we write to disk.
1942  */
1943 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1944 {
1945         struct dentry *dentry = file_dentry(file);
1946         struct inode *inode = d_inode(dentry);
1947         struct btrfs_root *root = BTRFS_I(inode)->root;
1948         struct btrfs_trans_handle *trans;
1949         struct btrfs_log_ctx ctx;
1950         int ret = 0;
1951         bool full_sync = 0;
1952         u64 len;
1953
1954         /*
1955          * If the inode needs a full sync, make sure we use a full range to
1956          * avoid log tree corruption, due to hole detection racing with ordered
1957          * extent completion for adjacent ranges, and assertion failures during
1958          * hole detection.
1959          */
1960         if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1961                      &BTRFS_I(inode)->runtime_flags)) {
1962                 start = 0;
1963                 end = LLONG_MAX;
1964         }
1965
1966         /*
1967          * The range length can be represented by u64, we have to do the typecasts
1968          * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
1969          */
1970         len = (u64)end - (u64)start + 1;
1971         trace_btrfs_sync_file(file, datasync);
1972
1973         /*
1974          * We write the dirty pages in the range and wait until they complete
1975          * out of the ->i_mutex. If so, we can flush the dirty pages by
1976          * multi-task, and make the performance up.  See
1977          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1978          */
1979         ret = start_ordered_ops(inode, start, end);
1980         if (ret)
1981                 return ret;
1982
1983         inode_lock(inode);
1984         atomic_inc(&root->log_batch);
1985         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1986                              &BTRFS_I(inode)->runtime_flags);
1987         /*
1988          * We might have have had more pages made dirty after calling
1989          * start_ordered_ops and before acquiring the inode's i_mutex.
1990          */
1991         if (full_sync) {
1992                 /*
1993                  * For a full sync, we need to make sure any ordered operations
1994                  * start and finish before we start logging the inode, so that
1995                  * all extents are persisted and the respective file extent
1996                  * items are in the fs/subvol btree.
1997                  */
1998                 ret = btrfs_wait_ordered_range(inode, start, len);
1999         } else {
2000                 /*
2001                  * Start any new ordered operations before starting to log the
2002                  * inode. We will wait for them to finish in btrfs_sync_log().
2003                  *
2004                  * Right before acquiring the inode's mutex, we might have new
2005                  * writes dirtying pages, which won't immediately start the
2006                  * respective ordered operations - that is done through the
2007                  * fill_delalloc callbacks invoked from the writepage and
2008                  * writepages address space operations. So make sure we start
2009                  * all ordered operations before starting to log our inode. Not
2010                  * doing this means that while logging the inode, writeback
2011                  * could start and invoke writepage/writepages, which would call
2012                  * the fill_delalloc callbacks (cow_file_range,
2013                  * submit_compressed_extents). These callbacks add first an
2014                  * extent map to the modified list of extents and then create
2015                  * the respective ordered operation, which means in
2016                  * tree-log.c:btrfs_log_inode() we might capture all existing
2017                  * ordered operations (with btrfs_get_logged_extents()) before
2018                  * the fill_delalloc callback adds its ordered operation, and by
2019                  * the time we visit the modified list of extent maps (with
2020                  * btrfs_log_changed_extents()), we see and process the extent
2021                  * map they created. We then use the extent map to construct a
2022                  * file extent item for logging without waiting for the
2023                  * respective ordered operation to finish - this file extent
2024                  * item points to a disk location that might not have yet been
2025                  * written to, containing random data - so after a crash a log
2026                  * replay will make our inode have file extent items that point
2027                  * to disk locations containing invalid data, as we returned
2028                  * success to userspace without waiting for the respective
2029                  * ordered operation to finish, because it wasn't captured by
2030                  * btrfs_get_logged_extents().
2031                  */
2032                 ret = start_ordered_ops(inode, start, end);
2033         }
2034         if (ret) {
2035                 inode_unlock(inode);
2036                 goto out;
2037         }
2038         atomic_inc(&root->log_batch);
2039
2040         /*
2041          * If the last transaction that changed this file was before the current
2042          * transaction and we have the full sync flag set in our inode, we can
2043          * bail out now without any syncing.
2044          *
2045          * Note that we can't bail out if the full sync flag isn't set. This is
2046          * because when the full sync flag is set we start all ordered extents
2047          * and wait for them to fully complete - when they complete they update
2048          * the inode's last_trans field through:
2049          *
2050          *     btrfs_finish_ordered_io() ->
2051          *         btrfs_update_inode_fallback() ->
2052          *             btrfs_update_inode() ->
2053          *                 btrfs_set_inode_last_trans()
2054          *
2055          * So we are sure that last_trans is up to date and can do this check to
2056          * bail out safely. For the fast path, when the full sync flag is not
2057          * set in our inode, we can not do it because we start only our ordered
2058          * extents and don't wait for them to complete (that is when
2059          * btrfs_finish_ordered_io runs), so here at this point their last_trans
2060          * value might be less than or equals to fs_info->last_trans_committed,
2061          * and setting a speculative last_trans for an inode when a buffered
2062          * write is made (such as fs_info->generation + 1 for example) would not
2063          * be reliable since after setting the value and before fsync is called
2064          * any number of transactions can start and commit (transaction kthread
2065          * commits the current transaction periodically), and a transaction
2066          * commit does not start nor waits for ordered extents to complete.
2067          */
2068         smp_mb();
2069         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
2070             (full_sync && BTRFS_I(inode)->last_trans <=
2071              root->fs_info->last_trans_committed) ||
2072             (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2073              BTRFS_I(inode)->last_trans
2074              <= root->fs_info->last_trans_committed)) {
2075                 /*
2076                  * We've had everything committed since the last time we were
2077                  * modified so clear this flag in case it was set for whatever
2078                  * reason, it's no longer relevant.
2079                  */
2080                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2081                           &BTRFS_I(inode)->runtime_flags);
2082                 /*
2083                  * An ordered extent might have started before and completed
2084                  * already with io errors, in which case the inode was not
2085                  * updated and we end up here. So check the inode's mapping
2086                  * flags for any errors that might have happened while doing
2087                  * writeback of file data.
2088                  */
2089                 ret = filemap_check_errors(inode->i_mapping);
2090                 inode_unlock(inode);
2091                 goto out;
2092         }
2093
2094         /*
2095          * ok we haven't committed the transaction yet, lets do a commit
2096          */
2097         if (file->private_data)
2098                 btrfs_ioctl_trans_end(file);
2099
2100         /*
2101          * We use start here because we will need to wait on the IO to complete
2102          * in btrfs_sync_log, which could require joining a transaction (for
2103          * example checking cross references in the nocow path).  If we use join
2104          * here we could get into a situation where we're waiting on IO to
2105          * happen that is blocked on a transaction trying to commit.  With start
2106          * we inc the extwriter counter, so we wait for all extwriters to exit
2107          * before we start blocking join'ers.  This comment is to keep somebody
2108          * from thinking they are super smart and changing this to
2109          * btrfs_join_transaction *cough*Josef*cough*.
2110          */
2111         trans = btrfs_start_transaction(root, 0);
2112         if (IS_ERR(trans)) {
2113                 ret = PTR_ERR(trans);
2114                 inode_unlock(inode);
2115                 goto out;
2116         }
2117         trans->sync = true;
2118
2119         btrfs_init_log_ctx(&ctx, inode);
2120
2121         ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2122         if (ret < 0) {
2123                 /* Fallthrough and commit/free transaction. */
2124                 ret = 1;
2125         }
2126
2127         /* we've logged all the items and now have a consistent
2128          * version of the file in the log.  It is possible that
2129          * someone will come in and modify the file, but that's
2130          * fine because the log is consistent on disk, and we
2131          * have references to all of the file's extents
2132          *
2133          * It is possible that someone will come in and log the
2134          * file again, but that will end up using the synchronization
2135          * inside btrfs_sync_log to keep things safe.
2136          */
2137         inode_unlock(inode);
2138
2139         /*
2140          * If any of the ordered extents had an error, just return it to user
2141          * space, so that the application knows some writes didn't succeed and
2142          * can take proper action (retry for e.g.). Blindly committing the
2143          * transaction in this case, would fool userspace that everything was
2144          * successful. And we also want to make sure our log doesn't contain
2145          * file extent items pointing to extents that weren't fully written to -
2146          * just like in the non fast fsync path, where we check for the ordered
2147          * operation's error flag before writing to the log tree and return -EIO
2148          * if any of them had this flag set (btrfs_wait_ordered_range) -
2149          * therefore we need to check for errors in the ordered operations,
2150          * which are indicated by ctx.io_err.
2151          */
2152         if (ctx.io_err) {
2153                 btrfs_end_transaction(trans, root);
2154                 ret = ctx.io_err;
2155                 goto out;
2156         }
2157
2158         if (ret != BTRFS_NO_LOG_SYNC) {
2159                 if (!ret) {
2160                         ret = btrfs_sync_log(trans, root, &ctx);
2161                         if (!ret) {
2162                                 ret = btrfs_end_transaction(trans, root);
2163                                 goto out;
2164                         }
2165                 }
2166                 if (!full_sync) {
2167                         ret = btrfs_wait_ordered_range(inode, start, len);
2168                         if (ret) {
2169                                 btrfs_end_transaction(trans, root);
2170                                 goto out;
2171                         }
2172                 }
2173                 ret = btrfs_commit_transaction(trans, root);
2174         } else {
2175                 ret = btrfs_end_transaction(trans, root);
2176         }
2177 out:
2178         return ret > 0 ? -EIO : ret;
2179 }
2180
2181 static const struct vm_operations_struct btrfs_file_vm_ops = {
2182         .fault          = filemap_fault,
2183         .map_pages      = filemap_map_pages,
2184         .page_mkwrite   = btrfs_page_mkwrite,
2185 };
2186
2187 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2188 {
2189         struct address_space *mapping = filp->f_mapping;
2190
2191         if (!mapping->a_ops->readpage)
2192                 return -ENOEXEC;
2193
2194         file_accessed(filp);
2195         vma->vm_ops = &btrfs_file_vm_ops;
2196
2197         return 0;
2198 }
2199
2200 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2201                           int slot, u64 start, u64 end)
2202 {
2203         struct btrfs_file_extent_item *fi;
2204         struct btrfs_key key;
2205
2206         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2207                 return 0;
2208
2209         btrfs_item_key_to_cpu(leaf, &key, slot);
2210         if (key.objectid != btrfs_ino(inode) ||
2211             key.type != BTRFS_EXTENT_DATA_KEY)
2212                 return 0;
2213
2214         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2215
2216         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2217                 return 0;
2218
2219         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2220                 return 0;
2221
2222         if (key.offset == end)
2223                 return 1;
2224         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2225                 return 1;
2226         return 0;
2227 }
2228
2229 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2230                       struct btrfs_path *path, u64 offset, u64 end)
2231 {
2232         struct btrfs_root *root = BTRFS_I(inode)->root;
2233         struct extent_buffer *leaf;
2234         struct btrfs_file_extent_item *fi;
2235         struct extent_map *hole_em;
2236         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2237         struct btrfs_key key;
2238         int ret;
2239
2240         if (btrfs_fs_incompat(root->fs_info, NO_HOLES))
2241                 goto out;
2242
2243         key.objectid = btrfs_ino(inode);
2244         key.type = BTRFS_EXTENT_DATA_KEY;
2245         key.offset = offset;
2246
2247         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2248         if (ret < 0)
2249                 return ret;
2250         BUG_ON(!ret);
2251
2252         leaf = path->nodes[0];
2253         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2254                 u64 num_bytes;
2255
2256                 path->slots[0]--;
2257                 fi = btrfs_item_ptr(leaf, path->slots[0],
2258                                     struct btrfs_file_extent_item);
2259                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2260                         end - offset;
2261                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2262                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2263                 btrfs_set_file_extent_offset(leaf, fi, 0);
2264                 btrfs_mark_buffer_dirty(leaf);
2265                 goto out;
2266         }
2267
2268         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2269                 u64 num_bytes;
2270
2271                 key.offset = offset;
2272                 btrfs_set_item_key_safe(root->fs_info, path, &key);
2273                 fi = btrfs_item_ptr(leaf, path->slots[0],
2274                                     struct btrfs_file_extent_item);
2275                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2276                         offset;
2277                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2278                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2279                 btrfs_set_file_extent_offset(leaf, fi, 0);
2280                 btrfs_mark_buffer_dirty(leaf);
2281                 goto out;
2282         }
2283         btrfs_release_path(path);
2284
2285         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2286                                        0, 0, end - offset, 0, end - offset,
2287                                        0, 0, 0);
2288         if (ret)
2289                 return ret;
2290
2291 out:
2292         btrfs_release_path(path);
2293
2294         hole_em = alloc_extent_map();
2295         if (!hole_em) {
2296                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2297                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2298                         &BTRFS_I(inode)->runtime_flags);
2299         } else {
2300                 hole_em->start = offset;
2301                 hole_em->len = end - offset;
2302                 hole_em->ram_bytes = hole_em->len;
2303                 hole_em->orig_start = offset;
2304
2305                 hole_em->block_start = EXTENT_MAP_HOLE;
2306                 hole_em->block_len = 0;
2307                 hole_em->orig_block_len = 0;
2308                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2309                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2310                 hole_em->generation = trans->transid;
2311
2312                 do {
2313                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2314                         write_lock(&em_tree->lock);
2315                         ret = add_extent_mapping(em_tree, hole_em, 1);
2316                         write_unlock(&em_tree->lock);
2317                 } while (ret == -EEXIST);
2318                 free_extent_map(hole_em);
2319                 if (ret)
2320                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2321                                 &BTRFS_I(inode)->runtime_flags);
2322         }
2323
2324         return 0;
2325 }
2326
2327 /*
2328  * Find a hole extent on given inode and change start/len to the end of hole
2329  * extent.(hole/vacuum extent whose em->start <= start &&
2330  *         em->start + em->len > start)
2331  * When a hole extent is found, return 1 and modify start/len.
2332  */
2333 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2334 {
2335         struct extent_map *em;
2336         int ret = 0;
2337
2338         em = btrfs_get_extent(inode, NULL, 0, *start, *len, 0);
2339         if (IS_ERR_OR_NULL(em)) {
2340                 if (!em)
2341                         ret = -ENOMEM;
2342                 else
2343                         ret = PTR_ERR(em);
2344                 return ret;
2345         }
2346
2347         /* Hole or vacuum extent(only exists in no-hole mode) */
2348         if (em->block_start == EXTENT_MAP_HOLE) {
2349                 ret = 1;
2350                 *len = em->start + em->len > *start + *len ?
2351                        0 : *start + *len - em->start - em->len;
2352                 *start = em->start + em->len;
2353         }
2354         free_extent_map(em);
2355         return ret;
2356 }
2357
2358 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2359 {
2360         struct btrfs_root *root = BTRFS_I(inode)->root;
2361         struct extent_state *cached_state = NULL;
2362         struct btrfs_path *path;
2363         struct btrfs_block_rsv *rsv;
2364         struct btrfs_trans_handle *trans;
2365         u64 lockstart;
2366         u64 lockend;
2367         u64 tail_start;
2368         u64 tail_len;
2369         u64 orig_start = offset;
2370         u64 cur_offset;
2371         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2372         u64 drop_end;
2373         int ret = 0;
2374         int err = 0;
2375         unsigned int rsv_count;
2376         bool same_block;
2377         bool no_holes = btrfs_fs_incompat(root->fs_info, NO_HOLES);
2378         u64 ino_size;
2379         bool truncated_block = false;
2380         bool updated_inode = false;
2381
2382         ret = btrfs_wait_ordered_range(inode, offset, len);
2383         if (ret)
2384                 return ret;
2385
2386         inode_lock(inode);
2387         ino_size = round_up(inode->i_size, root->sectorsize);
2388         ret = find_first_non_hole(inode, &offset, &len);
2389         if (ret < 0)
2390                 goto out_only_mutex;
2391         if (ret && !len) {
2392                 /* Already in a large hole */
2393                 ret = 0;
2394                 goto out_only_mutex;
2395         }
2396
2397         lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
2398         lockend = round_down(offset + len,
2399                              BTRFS_I(inode)->root->sectorsize) - 1;
2400         same_block = (BTRFS_BYTES_TO_BLKS(root->fs_info, offset))
2401                 == (BTRFS_BYTES_TO_BLKS(root->fs_info, offset + len - 1));
2402         /*
2403          * We needn't truncate any block which is beyond the end of the file
2404          * because we are sure there is no data there.
2405          */
2406         /*
2407          * Only do this if we are in the same block and we aren't doing the
2408          * entire block.
2409          */
2410         if (same_block && len < root->sectorsize) {
2411                 if (offset < ino_size) {
2412                         truncated_block = true;
2413                         ret = btrfs_truncate_block(inode, offset, len, 0);
2414                 } else {
2415                         ret = 0;
2416                 }
2417                 goto out_only_mutex;
2418         }
2419
2420         /* zero back part of the first block */
2421         if (offset < ino_size) {
2422                 truncated_block = true;
2423                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2424                 if (ret) {
2425                         inode_unlock(inode);
2426                         return ret;
2427                 }
2428         }
2429
2430         /* Check the aligned pages after the first unaligned page,
2431          * if offset != orig_start, which means the first unaligned page
2432          * including several following pages are already in holes,
2433          * the extra check can be skipped */
2434         if (offset == orig_start) {
2435                 /* after truncate page, check hole again */
2436                 len = offset + len - lockstart;
2437                 offset = lockstart;
2438                 ret = find_first_non_hole(inode, &offset, &len);
2439                 if (ret < 0)
2440                         goto out_only_mutex;
2441                 if (ret && !len) {
2442                         ret = 0;
2443                         goto out_only_mutex;
2444                 }
2445                 lockstart = offset;
2446         }
2447
2448         /* Check the tail unaligned part is in a hole */
2449         tail_start = lockend + 1;
2450         tail_len = offset + len - tail_start;
2451         if (tail_len) {
2452                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2453                 if (unlikely(ret < 0))
2454                         goto out_only_mutex;
2455                 if (!ret) {
2456                         /* zero the front end of the last page */
2457                         if (tail_start + tail_len < ino_size) {
2458                                 truncated_block = true;
2459                                 ret = btrfs_truncate_block(inode,
2460                                                         tail_start + tail_len,
2461                                                         0, 1);
2462                                 if (ret)
2463                                         goto out_only_mutex;
2464                         }
2465                 }
2466         }
2467
2468         if (lockend < lockstart) {
2469                 ret = 0;
2470                 goto out_only_mutex;
2471         }
2472
2473         while (1) {
2474                 struct btrfs_ordered_extent *ordered;
2475
2476                 truncate_pagecache_range(inode, lockstart, lockend);
2477
2478                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2479                                  &cached_state);
2480                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2481
2482                 /*
2483                  * We need to make sure we have no ordered extents in this range
2484                  * and nobody raced in and read a page in this range, if we did
2485                  * we need to try again.
2486                  */
2487                 if ((!ordered ||
2488                     (ordered->file_offset + ordered->len <= lockstart ||
2489                      ordered->file_offset > lockend)) &&
2490                      !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2491                         if (ordered)
2492                                 btrfs_put_ordered_extent(ordered);
2493                         break;
2494                 }
2495                 if (ordered)
2496                         btrfs_put_ordered_extent(ordered);
2497                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2498                                      lockend, &cached_state, GFP_NOFS);
2499                 ret = btrfs_wait_ordered_range(inode, lockstart,
2500                                                lockend - lockstart + 1);
2501                 if (ret) {
2502                         inode_unlock(inode);
2503                         return ret;
2504                 }
2505         }
2506
2507         path = btrfs_alloc_path();
2508         if (!path) {
2509                 ret = -ENOMEM;
2510                 goto out;
2511         }
2512
2513         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2514         if (!rsv) {
2515                 ret = -ENOMEM;
2516                 goto out_free;
2517         }
2518         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2519         rsv->failfast = 1;
2520
2521         /*
2522          * 1 - update the inode
2523          * 1 - removing the extents in the range
2524          * 1 - adding the hole extent if no_holes isn't set
2525          */
2526         rsv_count = no_holes ? 2 : 3;
2527         trans = btrfs_start_transaction(root, rsv_count);
2528         if (IS_ERR(trans)) {
2529                 err = PTR_ERR(trans);
2530                 goto out_free;
2531         }
2532
2533         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2534                                       min_size, 0);
2535         BUG_ON(ret);
2536         trans->block_rsv = rsv;
2537
2538         cur_offset = lockstart;
2539         len = lockend - cur_offset;
2540         while (cur_offset < lockend) {
2541                 ret = __btrfs_drop_extents(trans, root, inode, path,
2542                                            cur_offset, lockend + 1,
2543                                            &drop_end, 1, 0, 0, NULL);
2544                 if (ret != -ENOSPC)
2545                         break;
2546
2547                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2548
2549                 if (cur_offset < ino_size) {
2550                         ret = fill_holes(trans, inode, path, cur_offset,
2551                                          drop_end);
2552                         if (ret) {
2553                                 err = ret;
2554                                 break;
2555                         }
2556                 }
2557
2558                 cur_offset = drop_end;
2559
2560                 ret = btrfs_update_inode(trans, root, inode);
2561                 if (ret) {
2562                         err = ret;
2563                         break;
2564                 }
2565
2566                 btrfs_end_transaction(trans, root);
2567                 btrfs_btree_balance_dirty(root);
2568
2569                 trans = btrfs_start_transaction(root, rsv_count);
2570                 if (IS_ERR(trans)) {
2571                         ret = PTR_ERR(trans);
2572                         trans = NULL;
2573                         break;
2574                 }
2575
2576                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2577                                               rsv, min_size, 0);
2578                 BUG_ON(ret);    /* shouldn't happen */
2579                 trans->block_rsv = rsv;
2580
2581                 ret = find_first_non_hole(inode, &cur_offset, &len);
2582                 if (unlikely(ret < 0))
2583                         break;
2584                 if (ret && !len) {
2585                         ret = 0;
2586                         break;
2587                 }
2588         }
2589
2590         if (ret) {
2591                 err = ret;
2592                 goto out_trans;
2593         }
2594
2595         trans->block_rsv = &root->fs_info->trans_block_rsv;
2596         /*
2597          * If we are using the NO_HOLES feature we might have had already an
2598          * hole that overlaps a part of the region [lockstart, lockend] and
2599          * ends at (or beyond) lockend. Since we have no file extent items to
2600          * represent holes, drop_end can be less than lockend and so we must
2601          * make sure we have an extent map representing the existing hole (the
2602          * call to __btrfs_drop_extents() might have dropped the existing extent
2603          * map representing the existing hole), otherwise the fast fsync path
2604          * will not record the existence of the hole region
2605          * [existing_hole_start, lockend].
2606          */
2607         if (drop_end <= lockend)
2608                 drop_end = lockend + 1;
2609         /*
2610          * Don't insert file hole extent item if it's for a range beyond eof
2611          * (because it's useless) or if it represents a 0 bytes range (when
2612          * cur_offset == drop_end).
2613          */
2614         if (cur_offset < ino_size && cur_offset < drop_end) {
2615                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2616                 if (ret) {
2617                         err = ret;
2618                         goto out_trans;
2619                 }
2620         }
2621
2622 out_trans:
2623         if (!trans)
2624                 goto out_free;
2625
2626         inode_inc_iversion(inode);
2627         inode->i_mtime = inode->i_ctime = current_time(inode);
2628
2629         trans->block_rsv = &root->fs_info->trans_block_rsv;
2630         ret = btrfs_update_inode(trans, root, inode);
2631         updated_inode = true;
2632         btrfs_end_transaction(trans, root);
2633         btrfs_btree_balance_dirty(root);
2634 out_free:
2635         btrfs_free_path(path);
2636         btrfs_free_block_rsv(root, rsv);
2637 out:
2638         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2639                              &cached_state, GFP_NOFS);
2640 out_only_mutex:
2641         if (!updated_inode && truncated_block && !ret && !err) {
2642                 /*
2643                  * If we only end up zeroing part of a page, we still need to
2644                  * update the inode item, so that all the time fields are
2645                  * updated as well as the necessary btrfs inode in memory fields
2646                  * for detecting, at fsync time, if the inode isn't yet in the
2647                  * log tree or it's there but not up to date.
2648                  */
2649                 struct timespec now = current_time(inode);
2650
2651                 inode_inc_iversion(inode);
2652                 inode->i_mtime = now;
2653                 inode->i_ctime = now;
2654                 trans = btrfs_start_transaction(root, 1);
2655                 if (IS_ERR(trans)) {
2656                         err = PTR_ERR(trans);
2657                 } else {
2658                         err = btrfs_update_inode(trans, root, inode);
2659                         ret = btrfs_end_transaction(trans, root);
2660                 }
2661         }
2662         inode_unlock(inode);
2663         if (ret && !err)
2664                 err = ret;
2665         return err;
2666 }
2667
2668 /* Helper structure to record which range is already reserved */
2669 struct falloc_range {
2670         struct list_head list;
2671         u64 start;
2672         u64 len;
2673 };
2674
2675 /*
2676  * Helper function to add falloc range
2677  *
2678  * Caller should have locked the larger range of extent containing
2679  * [start, len)
2680  */
2681 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2682 {
2683         struct falloc_range *prev = NULL;
2684         struct falloc_range *range = NULL;
2685
2686         if (list_empty(head))
2687                 goto insert;
2688
2689         /*
2690          * As fallocate iterate by bytenr order, we only need to check
2691          * the last range.
2692          */
2693         prev = list_entry(head->prev, struct falloc_range, list);
2694         if (prev->start + prev->len == start) {
2695                 prev->len += len;
2696                 return 0;
2697         }
2698 insert:
2699         range = kmalloc(sizeof(*range), GFP_KERNEL);
2700         if (!range)
2701                 return -ENOMEM;
2702         range->start = start;
2703         range->len = len;
2704         list_add_tail(&range->list, head);
2705         return 0;
2706 }
2707
2708 static long btrfs_fallocate(struct file *file, int mode,
2709                             loff_t offset, loff_t len)
2710 {
2711         struct inode *inode = file_inode(file);
2712         struct extent_state *cached_state = NULL;
2713         struct falloc_range *range;
2714         struct falloc_range *tmp;
2715         struct list_head reserve_list;
2716         u64 cur_offset;
2717         u64 last_byte;
2718         u64 alloc_start;
2719         u64 alloc_end;
2720         u64 alloc_hint = 0;
2721         u64 locked_end;
2722         u64 actual_end = 0;
2723         struct extent_map *em;
2724         int blocksize = BTRFS_I(inode)->root->sectorsize;
2725         int ret;
2726
2727         alloc_start = round_down(offset, blocksize);
2728         alloc_end = round_up(offset + len, blocksize);
2729         cur_offset = alloc_start;
2730
2731         /* Make sure we aren't being give some crap mode */
2732         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2733                 return -EOPNOTSUPP;
2734
2735         if (mode & FALLOC_FL_PUNCH_HOLE)
2736                 return btrfs_punch_hole(inode, offset, len);
2737
2738         /*
2739          * Only trigger disk allocation, don't trigger qgroup reserve
2740          *
2741          * For qgroup space, it will be checked later.
2742          */
2743         ret = btrfs_alloc_data_chunk_ondemand(inode, alloc_end - alloc_start);
2744         if (ret < 0)
2745                 return ret;
2746
2747         inode_lock(inode);
2748
2749         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2750                 ret = inode_newsize_ok(inode, offset + len);
2751                 if (ret)
2752                         goto out;
2753         }
2754
2755         /*
2756          * TODO: Move these two operations after we have checked
2757          * accurate reserved space, or fallocate can still fail but
2758          * with page truncated or size expanded.
2759          *
2760          * But that's a minor problem and won't do much harm BTW.
2761          */
2762         if (alloc_start > inode->i_size) {
2763                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2764                                         alloc_start);
2765                 if (ret)
2766                         goto out;
2767         } else if (offset + len > inode->i_size) {
2768                 /*
2769                  * If we are fallocating from the end of the file onward we
2770                  * need to zero out the end of the block if i_size lands in the
2771                  * middle of a block.
2772                  */
2773                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2774                 if (ret)
2775                         goto out;
2776         }
2777
2778         /*
2779          * wait for ordered IO before we have any locks.  We'll loop again
2780          * below with the locks held.
2781          */
2782         ret = btrfs_wait_ordered_range(inode, alloc_start,
2783                                        alloc_end - alloc_start);
2784         if (ret)
2785                 goto out;
2786
2787         locked_end = alloc_end - 1;
2788         while (1) {
2789                 struct btrfs_ordered_extent *ordered;
2790
2791                 /* the extent lock is ordered inside the running
2792                  * transaction
2793                  */
2794                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2795                                  locked_end, &cached_state);
2796                 ordered = btrfs_lookup_first_ordered_extent(inode,
2797                                                             alloc_end - 1);
2798                 if (ordered &&
2799                     ordered->file_offset + ordered->len > alloc_start &&
2800                     ordered->file_offset < alloc_end) {
2801                         btrfs_put_ordered_extent(ordered);
2802                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2803                                              alloc_start, locked_end,
2804                                              &cached_state, GFP_KERNEL);
2805                         /*
2806                          * we can't wait on the range with the transaction
2807                          * running or with the extent lock held
2808                          */
2809                         ret = btrfs_wait_ordered_range(inode, alloc_start,
2810                                                        alloc_end - alloc_start);
2811                         if (ret)
2812                                 goto out;
2813                 } else {
2814                         if (ordered)
2815                                 btrfs_put_ordered_extent(ordered);
2816                         break;
2817                 }
2818         }
2819
2820         /* First, check if we exceed the qgroup limit */
2821         INIT_LIST_HEAD(&reserve_list);
2822         while (1) {
2823                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2824                                       alloc_end - cur_offset, 0);
2825                 if (IS_ERR_OR_NULL(em)) {
2826                         if (!em)
2827                                 ret = -ENOMEM;
2828                         else
2829                                 ret = PTR_ERR(em);
2830                         break;
2831                 }
2832                 last_byte = min(extent_map_end(em), alloc_end);
2833                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2834                 last_byte = ALIGN(last_byte, blocksize);
2835                 if (em->block_start == EXTENT_MAP_HOLE ||
2836                     (cur_offset >= inode->i_size &&
2837                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2838                         ret = add_falloc_range(&reserve_list, cur_offset,
2839                                                last_byte - cur_offset);
2840                         if (ret < 0) {
2841                                 free_extent_map(em);
2842                                 break;
2843                         }
2844                         ret = btrfs_qgroup_reserve_data(inode, cur_offset,
2845                                         last_byte - cur_offset);
2846                         if (ret < 0) {
2847                                 free_extent_map(em);
2848                                 break;
2849                         }
2850                 } else {
2851                         /*
2852                          * Do not need to reserve unwritten extent for this
2853                          * range, free reserved data space first, otherwise
2854                          * it'll result in false ENOSPC error.
2855                          */
2856                         btrfs_free_reserved_data_space(inode, cur_offset,
2857                                 last_byte - cur_offset);
2858                 }
2859                 free_extent_map(em);
2860                 cur_offset = last_byte;
2861                 if (cur_offset >= alloc_end)
2862                         break;
2863         }
2864
2865         /*
2866          * If ret is still 0, means we're OK to fallocate.
2867          * Or just cleanup the list and exit.
2868          */
2869         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2870                 if (!ret)
2871                         ret = btrfs_prealloc_file_range(inode, mode,
2872                                         range->start,
2873                                         range->len, i_blocksize(inode),
2874                                         offset + len, &alloc_hint);
2875                 else
2876                         btrfs_free_reserved_data_space(inode, range->start,
2877                                                        range->len);
2878                 list_del(&range->list);
2879                 kfree(range);
2880         }
2881         if (ret < 0)
2882                 goto out_unlock;
2883
2884         if (actual_end > inode->i_size &&
2885             !(mode & FALLOC_FL_KEEP_SIZE)) {
2886                 struct btrfs_trans_handle *trans;
2887                 struct btrfs_root *root = BTRFS_I(inode)->root;
2888
2889                 /*
2890                  * We didn't need to allocate any more space, but we
2891                  * still extended the size of the file so we need to
2892                  * update i_size and the inode item.
2893                  */
2894                 trans = btrfs_start_transaction(root, 1);
2895                 if (IS_ERR(trans)) {
2896                         ret = PTR_ERR(trans);
2897                 } else {
2898                         inode->i_ctime = current_time(inode);
2899                         i_size_write(inode, actual_end);
2900                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2901                         ret = btrfs_update_inode(trans, root, inode);
2902                         if (ret)
2903                                 btrfs_end_transaction(trans, root);
2904                         else
2905                                 ret = btrfs_end_transaction(trans, root);
2906                 }
2907         }
2908 out_unlock:
2909         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2910                              &cached_state, GFP_KERNEL);
2911 out:
2912         inode_unlock(inode);
2913         /* Let go of our reservation. */
2914         if (ret != 0)
2915                 btrfs_free_reserved_data_space(inode, alloc_start,
2916                                        alloc_end - cur_offset);
2917         return ret;
2918 }
2919
2920 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2921 {
2922         struct btrfs_root *root = BTRFS_I(inode)->root;
2923         struct extent_map *em = NULL;
2924         struct extent_state *cached_state = NULL;
2925         u64 lockstart;
2926         u64 lockend;
2927         u64 start;
2928         u64 len;
2929         int ret = 0;
2930
2931         if (inode->i_size == 0)
2932                 return -ENXIO;
2933
2934         /*
2935          * *offset can be negative, in this case we start finding DATA/HOLE from
2936          * the very start of the file.
2937          */
2938         start = max_t(loff_t, 0, *offset);
2939
2940         lockstart = round_down(start, root->sectorsize);
2941         lockend = round_up(i_size_read(inode), root->sectorsize);
2942         if (lockend <= lockstart)
2943                 lockend = lockstart + root->sectorsize;
2944         lockend--;
2945         len = lockend - lockstart + 1;
2946
2947         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2948                          &cached_state);
2949
2950         while (start < inode->i_size) {
2951                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2952                 if (IS_ERR(em)) {
2953                         ret = PTR_ERR(em);
2954                         em = NULL;
2955                         break;
2956                 }
2957
2958                 if (whence == SEEK_HOLE &&
2959                     (em->block_start == EXTENT_MAP_HOLE ||
2960                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2961                         break;
2962                 else if (whence == SEEK_DATA &&
2963                            (em->block_start != EXTENT_MAP_HOLE &&
2964                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2965                         break;
2966
2967                 start = em->start + em->len;
2968                 free_extent_map(em);
2969                 em = NULL;
2970                 cond_resched();
2971         }
2972         free_extent_map(em);
2973         if (!ret) {
2974                 if (whence == SEEK_DATA && start >= inode->i_size)
2975                         ret = -ENXIO;
2976                 else
2977                         *offset = min_t(loff_t, start, inode->i_size);
2978         }
2979         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2980                              &cached_state, GFP_NOFS);
2981         return ret;
2982 }
2983
2984 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2985 {
2986         struct inode *inode = file->f_mapping->host;
2987         int ret;
2988
2989         inode_lock(inode);
2990         switch (whence) {
2991         case SEEK_END:
2992         case SEEK_CUR:
2993                 offset = generic_file_llseek(file, offset, whence);
2994                 goto out;
2995         case SEEK_DATA:
2996         case SEEK_HOLE:
2997                 if (offset >= i_size_read(inode)) {
2998                         inode_unlock(inode);
2999                         return -ENXIO;
3000                 }
3001
3002                 ret = find_desired_extent(inode, &offset, whence);
3003                 if (ret) {
3004                         inode_unlock(inode);
3005                         return ret;
3006                 }
3007         }
3008
3009         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3010 out:
3011         inode_unlock(inode);
3012         return offset;
3013 }
3014
3015 const struct file_operations btrfs_file_operations = {
3016         .llseek         = btrfs_file_llseek,
3017         .read_iter      = generic_file_read_iter,
3018         .splice_read    = generic_file_splice_read,
3019         .write_iter     = btrfs_file_write_iter,
3020         .mmap           = btrfs_file_mmap,
3021         .open           = generic_file_open,
3022         .release        = btrfs_release_file,
3023         .fsync          = btrfs_sync_file,
3024         .fallocate      = btrfs_fallocate,
3025         .unlocked_ioctl = btrfs_ioctl,
3026 #ifdef CONFIG_COMPAT
3027         .compat_ioctl   = btrfs_compat_ioctl,
3028 #endif
3029         .copy_file_range = btrfs_copy_file_range,
3030         .clone_file_range = btrfs_clone_file_range,
3031         .dedupe_file_range = btrfs_dedupe_file_range,
3032 };
3033
3034 void btrfs_auto_defrag_exit(void)
3035 {
3036         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3037 }
3038
3039 int btrfs_auto_defrag_init(void)
3040 {
3041         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3042                                         sizeof(struct inode_defrag), 0,
3043                                         SLAB_MEM_SPREAD,
3044                                         NULL);
3045         if (!btrfs_inode_defrag_cachep)
3046                 return -ENOMEM;
3047
3048         return 0;
3049 }
3050
3051 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3052 {
3053         int ret;
3054
3055         /*
3056          * So with compression we will find and lock a dirty page and clear the
3057          * first one as dirty, setup an async extent, and immediately return
3058          * with the entire range locked but with nobody actually marked with
3059          * writeback.  So we can't just filemap_write_and_wait_range() and
3060          * expect it to work since it will just kick off a thread to do the
3061          * actual work.  So we need to call filemap_fdatawrite_range _again_
3062          * since it will wait on the page lock, which won't be unlocked until
3063          * after the pages have been marked as writeback and so we're good to go
3064          * from there.  We have to do this otherwise we'll miss the ordered
3065          * extents and that results in badness.  Please Josef, do not think you
3066          * know better and pull this out at some point in the future, it is
3067          * right and you are wrong.
3068          */
3069         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3070         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3071                              &BTRFS_I(inode)->runtime_flags))
3072                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3073
3074         return ret;
3075 }