fs/btrfs/space-info.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "misc.h"
   4 #include "ctree.h"
   5 #include "space-info.h"
   6 #include "sysfs.h"
   7 #include "volumes.h"
   8 #include "free-space-cache.h"
   9 #include "ordered-data.h"
  10 #include "transaction.h"
  11 #include "block-group.h"
  12
  13 /*
  14  * HOW DOES SPACE RESERVATION WORK
  15  *
  16  * If you want to know about delalloc specifically, there is a separate comment
  17  * for that with the delalloc code.  This comment is about how the whole system
  18  * works generally.
  19  *
  20  * BASIC CONCEPTS
  21  *
  22  *   1) space_info.  This is the ultimate arbiter of how much space we can use.
  23  *   There's a description of the bytes_ fields with the struct declaration,
  24  *   refer to that for specifics on each field.  Suffice it to say that for
  25  *   reservations we care about total_bytes - SUM(space_info->bytes_) when
  26  *   determining if there is space to make an allocation.  There is a space_info
  27  *   for METADATA, SYSTEM, and DATA areas.
  28  *
  29  *   2) block_rsv's.  These are basically buckets for every different type of
  30  *   metadata reservation we have.  You can see the comment in the block_rsv
  31  *   code on the rules for each type, but generally block_rsv->reserved is how
  32  *   much space is accounted for in space_info->bytes_may_use.
  33  *
  34  *   3) btrfs_calc*_size.  These are the worst case calculations we used based
  35  *   on the number of items we will want to modify.  We have one for changing
  36  *   items, and one for inserting new items.  Generally we use these helpers to
  37  *   determine the size of the block reserves, and then use the actual bytes
  38  *   values to adjust the space_info counters.
  39  *
  40  * MAKING RESERVATIONS, THE NORMAL CASE
  41  *
  42  *   We call into either btrfs_reserve_data_bytes() or
  43  *   btrfs_reserve_metadata_bytes(), depending on which we're looking for, with
  44  *   num_bytes we want to reserve.
  45  *
  46  *   ->reserve
  47  *     space_info->bytes_may_reserve += num_bytes
  48  *
  49  *   ->extent allocation
  50  *     Call btrfs_add_reserved_bytes() which does
  51  *     space_info->bytes_may_reserve -= num_bytes
  52  *     space_info->bytes_reserved += extent_bytes
  53  *
  54  *   ->insert reference
  55  *     Call btrfs_update_block_group() which does
  56  *     space_info->bytes_reserved -= extent_bytes
  57  *     space_info->bytes_used += extent_bytes
  58  *
  59  * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
  60  *
  61  *   Assume we are unable to simply make the reservation because we do not have
  62  *   enough space
  63  *
  64  *   -> __reserve_bytes
  65  *     create a reserve_ticket with ->bytes set to our reservation, add it to
  66  *     the tail of space_info->tickets, kick async flush thread
  67  *
  68  *   ->handle_reserve_ticket
  69  *     wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
  70  *     on the ticket.
  71  *
  72  *   -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
  73  *     Flushes various things attempting to free up space.
  74  *
  75  *   -> btrfs_try_granting_tickets()
  76  *     This is called by anything that either subtracts space from
  77  *     space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the
  78  *     space_info->total_bytes.  This loops through the ->priority_tickets and
  79  *     then the ->tickets list checking to see if the reservation can be
  80  *     completed.  If it can the space is added to space_info->bytes_may_use and
  81  *     the ticket is woken up.
  82  *
  83  *   -> ticket wakeup
  84  *     Check if ->bytes == 0, if it does we got our reservation and we can carry
  85  *     on, if not return the appropriate error (ENOSPC, but can be EINTR if we
  86  *     were interrupted.)
  87  *
  88  * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
  89  *
  90  *   Same as the above, except we add ourselves to the
  91  *   space_info->priority_tickets, and we do not use ticket->wait, we simply
  92  *   call flush_space() ourselves for the states that are safe for us to call
  93  *   without deadlocking and hope for the best.
  94  *
  95  * THE FLUSHING STATES
  96  *
  97  *   Generally speaking we will have two cases for each state, a "nice" state
  98  *   and a "ALL THE THINGS" state.  In btrfs we delay a lot of work in order to
  99  *   reduce the locking over head on the various trees, and even to keep from
 100  *   doing any work at all in the case of delayed refs.  Each of these delayed
 101  *   things however hold reservations, and so letting them run allows us to
 102  *   reclaim space so we can make new reservations.
 103  *
 104  *   FLUSH_DELAYED_ITEMS
 105  *     Every inode has a delayed item to update the inode.  Take a simple write
 106  *     for example, we would update the inode item at write time to update the
 107  *     mtime, and then again at finish_ordered_io() time in order to update the
 108  *     isize or bytes.  We keep these delayed items to coalesce these operations
 109  *     into a single operation done on demand.  These are an easy way to reclaim
 110  *     metadata space.
 111  *
 112  *   FLUSH_DELALLOC
 113  *     Look at the delalloc comment to get an idea of how much space is reserved
 114  *     for delayed allocation.  We can reclaim some of this space simply by
 115  *     running delalloc, but usually we need to wait for ordered extents to
 116  *     reclaim the bulk of this space.
 117  *
 118  *   FLUSH_DELAYED_REFS
 119  *     We have a block reserve for the outstanding delayed refs space, and every
 120  *     delayed ref operation holds a reservation.  Running these is a quick way
 121  *     to reclaim space, but we want to hold this until the end because COW can
 122  *     churn a lot and we can avoid making some extent tree modifications if we
 123  *     are able to delay for as long as possible.
 124  *
 125  *   ALLOC_CHUNK
 126  *     We will skip this the first time through space reservation, because of
 127  *     overcommit and we don't want to have a lot of useless metadata space when
 128  *     our worst case reservations will likely never come true.
 129  *
 130  *   RUN_DELAYED_IPUTS
 131  *     If we're freeing inodes we're likely freeing checksums, file extent
 132  *     items, and extent tree items.  Loads of space could be freed up by these
 133  *     operations, however they won't be usable until the transaction commits.
 134  *
 135  *   COMMIT_TRANS
 136  *     may_commit_transaction() is the ultimate arbiter on whether we commit the
 137  *     transaction or not.  In order to avoid constantly churning we do all the
 138  *     above flushing first and then commit the transaction as the last resort.
 139  *     However we need to take into account things like pinned space that would
 140  *     be freed, plus any delayed work we may not have gotten rid of in the case
 141  *     of metadata.
 142  *
 143  * OVERCOMMIT
 144  *
 145  *   Because we hold so many reservations for metadata we will allow you to
 146  *   reserve more space than is currently free in the currently allocate
 147  *   metadata space.  This only happens with metadata, data does not allow
 148  *   overcommitting.
 149  *
 150  *   You can see the current logic for when we allow overcommit in
 151  *   btrfs_can_overcommit(), but it only applies to unallocated space.  If there
 152  *   is no unallocated space to be had, all reservations are kept within the
 153  *   free space in the allocated metadata chunks.
 154  *
 155  *   Because of overcommitting, you generally want to use the
 156  *   btrfs_can_overcommit() logic for metadata allocations, as it does the right
 157  *   thing with or without extra unallocated space.
 158  */
 159
 160 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
 161                           bool may_use_included)
 162 {
 163         ASSERT(s_info);
 164         return s_info->bytes_used + s_info->bytes_reserved +
 165                 s_info->bytes_pinned + s_info->bytes_readonly +
 166                 (may_use_included ? s_info->bytes_may_use : 0);
 167 }
 168
 169 /*
 170  * after adding space to the filesystem, we need to clear the full flags
 171  * on all the space infos.
 172  */
 173 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 174 {
 175         struct list_head *head = &info->space_info;
 176         struct btrfs_space_info *found;
 177
 178         rcu_read_lock();
 179         list_for_each_entry_rcu(found, head, list)
 180                 found->full = 0;
 181         rcu_read_unlock();
 182 }
 183
 184 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
 185 {
 186
 187         struct btrfs_space_info *space_info;
 188         int i;
 189         int ret;
 190
 191         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
 192         if (!space_info)
 193                 return -ENOMEM;
 194
 195         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
 196                                  GFP_KERNEL);
 197         if (ret) {
 198                 kfree(space_info);
 199                 return ret;
 200         }
 201
 202         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
 203                 INIT_LIST_HEAD(&space_info->block_groups[i]);
 204         init_rwsem(&space_info->groups_sem);
 205         spin_lock_init(&space_info->lock);
 206         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
 207         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
 208         INIT_LIST_HEAD(&space_info->ro_bgs);
 209         INIT_LIST_HEAD(&space_info->tickets);
 210         INIT_LIST_HEAD(&space_info->priority_tickets);
 211
 212         ret = btrfs_sysfs_add_space_info_type(info, space_info);
 213         if (ret)
 214                 return ret;
 215
 216         list_add_rcu(&space_info->list, &info->space_info);
 217         if (flags & BTRFS_BLOCK_GROUP_DATA)
 218                 info->data_sinfo = space_info;
 219
 220         return ret;
 221 }
 222
 223 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
 224 {
 225         struct btrfs_super_block *disk_super;
 226         u64 features;
 227         u64 flags;
 228         int mixed = 0;
 229         int ret;
 230
 231         disk_super = fs_info->super_copy;
 232         if (!btrfs_super_root(disk_super))
 233                 return -EINVAL;
 234
 235         features = btrfs_super_incompat_flags(disk_super);
 236         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
 237                 mixed = 1;
 238
 239         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 240         ret = create_space_info(fs_info, flags);
 241         if (ret)
 242                 goto out;
 243
 244         if (mixed) {
 245                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
 246                 ret = create_space_info(fs_info, flags);
 247         } else {
 248                 flags = BTRFS_BLOCK_GROUP_METADATA;
 249                 ret = create_space_info(fs_info, flags);
 250                 if (ret)
 251                         goto out;
 252
 253                 flags = BTRFS_BLOCK_GROUP_DATA;
 254                 ret = create_space_info(fs_info, flags);
 255         }
 256 out:
 257         return ret;
 258 }
 259
 260 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags,
 261                              u64 total_bytes, u64 bytes_used,
 262                              u64 bytes_readonly,
 263                              struct btrfs_space_info **space_info)
 264 {
 265         struct btrfs_space_info *found;
 266         int factor;
 267
 268         factor = btrfs_bg_type_to_factor(flags);
 269
 270         found = btrfs_find_space_info(info, flags);
 271         ASSERT(found);
 272         spin_lock(&found->lock);
 273         found->total_bytes += total_bytes;
 274         found->disk_total += total_bytes * factor;
 275         found->bytes_used += bytes_used;
 276         found->disk_used += bytes_used * factor;
 277         found->bytes_readonly += bytes_readonly;
 278         if (total_bytes > 0)
 279                 found->full = 0;
 280         btrfs_try_granting_tickets(info, found);
 281         spin_unlock(&found->lock);
 282         *space_info = found;
 283 }
 284
 285 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
 286                                                u64 flags)
 287 {
 288         struct list_head *head = &info->space_info;
 289         struct btrfs_space_info *found;
 290
 291         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 292
 293         rcu_read_lock();
 294         list_for_each_entry_rcu(found, head, list) {
 295                 if (found->flags & flags) {
 296                         rcu_read_unlock();
 297                         return found;
 298                 }
 299         }
 300         rcu_read_unlock();
 301         return NULL;
 302 }
 303
 304 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
 305 {
 306         return (global->size << 1);
 307 }
 308
 309 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
 310                           struct btrfs_space_info *space_info,
 311                           enum btrfs_reserve_flush_enum flush)
 312 {
 313         u64 profile;
 314         u64 avail;
 315         int factor;
 316
 317         if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
 318                 profile = btrfs_system_alloc_profile(fs_info);
 319         else
 320                 profile = btrfs_metadata_alloc_profile(fs_info);
 321
 322         avail = atomic64_read(&fs_info->free_chunk_space);
 323
 324         /*
 325          * If we have dup, raid1 or raid10 then only half of the free
 326          * space is actually usable.  For raid56, the space info used
 327          * doesn't include the parity drive, so we don't have to
 328          * change the math
 329          */
 330         factor = btrfs_bg_type_to_factor(profile);
 331         avail = div_u64(avail, factor);
 332
 333         /*
 334          * If we aren't flushing all things, let us overcommit up to
 335          * 1/2th of the space. If we can flush, don't let us overcommit
 336          * too much, let it overcommit up to 1/8 of the space.
 337          */
 338         if (flush == BTRFS_RESERVE_FLUSH_ALL)
 339                 avail >>= 3;
 340         else
 341                 avail >>= 1;
 342         return avail;
 343 }
 344
 345 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
 346                          struct btrfs_space_info *space_info, u64 bytes,
 347                          enum btrfs_reserve_flush_enum flush)
 348 {
 349         u64 avail;
 350         u64 used;
 351
 352         /* Don't overcommit when in mixed mode */
 353         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
 354                 return 0;
 355
 356         used = btrfs_space_info_used(space_info, true);
 357         avail = calc_available_free_space(fs_info, space_info, flush);
 358
 359         if (used + bytes < space_info->total_bytes + avail)
 360                 return 1;
 361         return 0;
 362 }
 363
 364 /*
 365  * This is for space we already have accounted in space_info->bytes_may_use, so
 366  * basically when we're returning space from block_rsv's.
 367  */
 368 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
 369                                 struct btrfs_space_info *space_info)
 370 {
 371         struct list_head *head;
 372         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
 373
 374         lockdep_assert_held(&space_info->lock);
 375
 376         head = &space_info->priority_tickets;
 377 again:
 378         while (!list_empty(head)) {
 379                 struct reserve_ticket *ticket;
 380                 u64 used = btrfs_space_info_used(space_info, true);
 381
 382                 ticket = list_first_entry(head, struct reserve_ticket, list);
 383
 384                 /* Check and see if our ticket can be satisified now. */
 385                 if ((used + ticket->bytes <= space_info->total_bytes) ||
 386                     btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
 387                                          flush)) {
 388                         btrfs_space_info_update_bytes_may_use(fs_info,
 389                                                               space_info,
 390                                                               ticket->bytes);
 391                         list_del_init(&ticket->list);
 392                         ASSERT(space_info->reclaim_size >= ticket->bytes);
 393                         space_info->reclaim_size -= ticket->bytes;
 394                         ticket->bytes = 0;
 395                         space_info->tickets_id++;
 396                         wake_up(&ticket->wait);
 397                 } else {
 398                         break;
 399                 }
 400         }
 401
 402         if (head == &space_info->priority_tickets) {
 403                 head = &space_info->tickets;
 404                 flush = BTRFS_RESERVE_FLUSH_ALL;
 405                 goto again;
 406         }
 407 }
 408
 409 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
 410 do {                                                                    \
 411         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
 412         spin_lock(&__rsv->lock);                                        \
 413         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
 414                    __rsv->size, __rsv->reserved);                       \
 415         spin_unlock(&__rsv->lock);                                      \
 416 } while (0)
 417
 418 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 419                                     struct btrfs_space_info *info)
 420 {
 421         lockdep_assert_held(&info->lock);
 422
 423         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
 424                    info->flags,
 425                    info->total_bytes - btrfs_space_info_used(info, true),
 426                    info->full ? "" : "not ");
 427         btrfs_info(fs_info,
 428                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
 429                 info->total_bytes, info->bytes_used, info->bytes_pinned,
 430                 info->bytes_reserved, info->bytes_may_use,
 431                 info->bytes_readonly);
 432
 433         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
 434         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
 435         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
 436         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
 437         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
 438
 439 }
 440
 441 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 442                            struct btrfs_space_info *info, u64 bytes,
 443                            int dump_block_groups)
 444 {
 445         struct btrfs_block_group *cache;
 446         int index = 0;
 447
 448         spin_lock(&info->lock);
 449         __btrfs_dump_space_info(fs_info, info);
 450         spin_unlock(&info->lock);
 451
 452         if (!dump_block_groups)
 453                 return;
 454
 455         down_read(&info->groups_sem);
 456 again:
 457         list_for_each_entry(cache, &info->block_groups[index], list) {
 458                 spin_lock(&cache->lock);
 459                 btrfs_info(fs_info,
 460                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
 461                         cache->start, cache->length, cache->used, cache->pinned,
 462                         cache->reserved, cache->ro ? "[readonly]" : "");
 463                 btrfs_dump_free_space(cache, bytes);
 464                 spin_unlock(&cache->lock);
 465         }
 466         if (++index < BTRFS_NR_RAID_TYPES)
 467                 goto again;
 468         up_read(&info->groups_sem);
 469 }
 470
 471 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
 472                                          unsigned long nr_pages, int nr_items)
 473 {
 474         struct super_block *sb = fs_info->sb;
 475
 476         if (down_read_trylock(&sb->s_umount)) {
 477                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
 478                 up_read(&sb->s_umount);
 479         } else {
 480                 /*
 481                  * We needn't worry the filesystem going from r/w to r/o though
 482                  * we don't acquire ->s_umount mutex, because the filesystem
 483                  * should guarantee the delalloc inodes list be empty after
 484                  * the filesystem is readonly(all dirty pages are written to
 485                  * the disk).
 486                  */
 487                 btrfs_start_delalloc_roots(fs_info, nr_items);
 488                 if (!current->journal_info)
 489                         btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
 490         }
 491 }
 492
 493 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
 494                                         u64 to_reclaim)
 495 {
 496         u64 bytes;
 497         u64 nr;
 498
 499         bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 500         nr = div64_u64(to_reclaim, bytes);
 501         if (!nr)
 502                 nr = 1;
 503         return nr;
 504 }
 505
 506 #define EXTENT_SIZE_PER_ITEM    SZ_256K
 507
 508 /*
 509  * shrink metadata reservation for delalloc
 510  */
 511 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
 512                             u64 orig, bool wait_ordered)
 513 {
 514         struct btrfs_space_info *space_info;
 515         struct btrfs_trans_handle *trans;
 516         u64 delalloc_bytes;
 517         u64 dio_bytes;
 518         u64 async_pages;
 519         u64 items;
 520         long time_left;
 521         unsigned long nr_pages;
 522         int loops;
 523
 524         /* Calc the number of the pages we need flush for space reservation */
 525         items = calc_reclaim_items_nr(fs_info, to_reclaim);
 526         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
 527
 528         trans = (struct btrfs_trans_handle *)current->journal_info;
 529         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
 530
 531         delalloc_bytes = percpu_counter_sum_positive(
 532                                                 &fs_info->delalloc_bytes);
 533         dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
 534         if (delalloc_bytes == 0 && dio_bytes == 0) {
 535                 if (trans)
 536                         return;
 537                 if (wait_ordered)
 538                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
 539                 return;
 540         }
 541
 542         /*
 543          * If we are doing more ordered than delalloc we need to just wait on
 544          * ordered extents, otherwise we'll waste time trying to flush delalloc
 545          * that likely won't give us the space back we need.
 546          */
 547         if (dio_bytes > delalloc_bytes)
 548                 wait_ordered = true;
 549
 550         loops = 0;
 551         while ((delalloc_bytes || dio_bytes) && loops < 3) {
 552                 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
 553
 554                 /*
 555                  * Triggers inode writeback for up to nr_pages. This will invoke
 556                  * ->writepages callback and trigger delalloc filling
 557                  *  (btrfs_run_delalloc_range()).
 558                  */
 559                 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
 560
 561                 /*
 562                  * We need to wait for the compressed pages to start before
 563                  * we continue.
 564                  */
 565                 async_pages = atomic_read(&fs_info->async_delalloc_pages);
 566                 if (!async_pages)
 567                         goto skip_async;
 568
 569                 /*
 570                  * Calculate how many compressed pages we want to be written
 571                  * before we continue. I.e if there are more async pages than we
 572                  * require wait_event will wait until nr_pages are written.
 573                  */
 574                 if (async_pages <= nr_pages)
 575                         async_pages = 0;
 576                 else
 577                         async_pages -= nr_pages;
 578
 579                 wait_event(fs_info->async_submit_wait,
 580                            atomic_read(&fs_info->async_delalloc_pages) <=
 581                            (int)async_pages);
 582 skip_async:
 583                 spin_lock(&space_info->lock);
 584                 if (list_empty(&space_info->tickets) &&
 585                     list_empty(&space_info->priority_tickets)) {
 586                         spin_unlock(&space_info->lock);
 587                         break;
 588                 }
 589                 spin_unlock(&space_info->lock);
 590
 591                 loops++;
 592                 if (wait_ordered && !trans) {
 593                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
 594                 } else {
 595                         time_left = schedule_timeout_killable(1);
 596                         if (time_left)
 597                                 break;
 598                 }
 599                 delalloc_bytes = percpu_counter_sum_positive(
 600                                                 &fs_info->delalloc_bytes);
 601                 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
 602         }
 603 }
 604
 605 /**
 606  * maybe_commit_transaction - possibly commit the transaction if its ok to
 607  * @root - the root we're allocating for
 608  * @bytes - the number of bytes we want to reserve
 609  * @force - force the commit
 610  *
 611  * This will check to make sure that committing the transaction will actually
 612  * get us somewhere and then commit the transaction if it does.  Otherwise it
 613  * will return -ENOSPC.
 614  */
 615 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
 616                                   struct btrfs_space_info *space_info)
 617 {
 618         struct reserve_ticket *ticket = NULL;
 619         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
 620         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 621         struct btrfs_trans_handle *trans;
 622         u64 bytes_needed;
 623         u64 reclaim_bytes = 0;
 624         u64 cur_free_bytes = 0;
 625
 626         trans = (struct btrfs_trans_handle *)current->journal_info;
 627         if (trans)
 628                 return -EAGAIN;
 629
 630         spin_lock(&space_info->lock);
 631         cur_free_bytes = btrfs_space_info_used(space_info, true);
 632         if (cur_free_bytes < space_info->total_bytes)
 633                 cur_free_bytes = space_info->total_bytes - cur_free_bytes;
 634         else
 635                 cur_free_bytes = 0;
 636
 637         if (!list_empty(&space_info->priority_tickets))
 638                 ticket = list_first_entry(&space_info->priority_tickets,
 639                                           struct reserve_ticket, list);
 640         else if (!list_empty(&space_info->tickets))
 641                 ticket = list_first_entry(&space_info->tickets,
 642                                           struct reserve_ticket, list);
 643         bytes_needed = (ticket) ? ticket->bytes : 0;
 644
 645         if (bytes_needed > cur_free_bytes)
 646                 bytes_needed -= cur_free_bytes;
 647         else
 648                 bytes_needed = 0;
 649         spin_unlock(&space_info->lock);
 650
 651         if (!bytes_needed)
 652                 return 0;
 653
 654         trans = btrfs_join_transaction(fs_info->extent_root);
 655         if (IS_ERR(trans))
 656                 return PTR_ERR(trans);
 657
 658         /*
 659          * See if there is enough pinned space to make this reservation, or if
 660          * we have block groups that are going to be freed, allowing us to
 661          * possibly do a chunk allocation the next loop through.
 662          */
 663         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
 664             __percpu_counter_compare(&space_info->total_bytes_pinned,
 665                                      bytes_needed,
 666                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
 667                 goto commit;
 668
 669         /*
 670          * See if there is some space in the delayed insertion reservation for
 671          * this reservation.
 672          */
 673         if (space_info != delayed_rsv->space_info)
 674                 goto enospc;
 675
 676         spin_lock(&delayed_rsv->lock);
 677         reclaim_bytes += delayed_rsv->reserved;
 678         spin_unlock(&delayed_rsv->lock);
 679
 680         spin_lock(&delayed_refs_rsv->lock);
 681         reclaim_bytes += delayed_refs_rsv->reserved;
 682         spin_unlock(&delayed_refs_rsv->lock);
 683         if (reclaim_bytes >= bytes_needed)
 684                 goto commit;
 685         bytes_needed -= reclaim_bytes;
 686
 687         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
 688                                    bytes_needed,
 689                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
 690                 goto enospc;
 691
 692 commit:
 693         return btrfs_commit_transaction(trans);
 694 enospc:
 695         btrfs_end_transaction(trans);
 696         return -ENOSPC;
 697 }
 698
 699 /*
 700  * Try to flush some data based on policy set by @state. This is only advisory
 701  * and may fail for various reasons. The caller is supposed to examine the
 702  * state of @space_info to detect the outcome.
 703  */
 704 static void flush_space(struct btrfs_fs_info *fs_info,
 705                        struct btrfs_space_info *space_info, u64 num_bytes,
 706                        int state)
 707 {
 708         struct btrfs_root *root = fs_info->extent_root;
 709         struct btrfs_trans_handle *trans;
 710         int nr;
 711         int ret = 0;
 712
 713         switch (state) {
 714         case FLUSH_DELAYED_ITEMS_NR:
 715         case FLUSH_DELAYED_ITEMS:
 716                 if (state == FLUSH_DELAYED_ITEMS_NR)
 717                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
 718                 else
 719                         nr = -1;
 720
 721                 trans = btrfs_join_transaction(root);
 722                 if (IS_ERR(trans)) {
 723                         ret = PTR_ERR(trans);
 724                         break;
 725                 }
 726                 ret = btrfs_run_delayed_items_nr(trans, nr);
 727                 btrfs_end_transaction(trans);
 728                 break;
 729         case FLUSH_DELALLOC:
 730         case FLUSH_DELALLOC_WAIT:
 731                 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
 732                                 state == FLUSH_DELALLOC_WAIT);
 733                 break;
 734         case FLUSH_DELAYED_REFS_NR:
 735         case FLUSH_DELAYED_REFS:
 736                 trans = btrfs_join_transaction(root);
 737                 if (IS_ERR(trans)) {
 738                         ret = PTR_ERR(trans);
 739                         break;
 740                 }
 741                 if (state == FLUSH_DELAYED_REFS_NR)
 742                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
 743                 else
 744                         nr = 0;
 745                 btrfs_run_delayed_refs(trans, nr);
 746                 btrfs_end_transaction(trans);
 747                 break;
 748         case ALLOC_CHUNK:
 749         case ALLOC_CHUNK_FORCE:
 750                 trans = btrfs_join_transaction(root);
 751                 if (IS_ERR(trans)) {
 752                         ret = PTR_ERR(trans);
 753                         break;
 754                 }
 755                 ret = btrfs_chunk_alloc(trans,
 756                                 btrfs_metadata_alloc_profile(fs_info),
 757                                 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
 758                                         CHUNK_ALLOC_FORCE);
 759                 btrfs_end_transaction(trans);
 760                 if (ret > 0 || ret == -ENOSPC)
 761                         ret = 0;
 762                 break;
 763         case RUN_DELAYED_IPUTS:
 764                 /*
 765                  * If we have pending delayed iputs then we could free up a
 766                  * bunch of pinned space, so make sure we run the iputs before
 767                  * we do our pinned bytes check below.
 768                  */
 769                 btrfs_run_delayed_iputs(fs_info);
 770                 btrfs_wait_on_delayed_iputs(fs_info);
 771                 break;
 772         case COMMIT_TRANS:
 773                 ret = may_commit_transaction(fs_info, space_info);
 774                 break;
 775         default:
 776                 ret = -ENOSPC;
 777                 break;
 778         }
 779
 780         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
 781                                 ret);
 782         return;
 783 }
 784
 785 static inline u64
 786 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
 787                                  struct btrfs_space_info *space_info)
 788 {
 789         u64 used;
 790         u64 avail;
 791         u64 expected;
 792         u64 to_reclaim = space_info->reclaim_size;
 793
 794         lockdep_assert_held(&space_info->lock);
 795
 796         avail = calc_available_free_space(fs_info, space_info,
 797                                           BTRFS_RESERVE_FLUSH_ALL);
 798         used = btrfs_space_info_used(space_info, true);
 799
 800         /*
 801          * We may be flushing because suddenly we have less space than we had
 802          * before, and now we're well over-committed based on our current free
 803          * space.  If that's the case add in our overage so we make sure to put
 804          * appropriate pressure on the flushing state machine.
 805          */
 806         if (space_info->total_bytes + avail < used)
 807                 to_reclaim += used - (space_info->total_bytes + avail);
 808
 809         if (to_reclaim)
 810                 return to_reclaim;
 811
 812         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
 813         if (btrfs_can_overcommit(fs_info, space_info, to_reclaim,
 814                                  BTRFS_RESERVE_FLUSH_ALL))
 815                 return 0;
 816
 817         used = btrfs_space_info_used(space_info, true);
 818
 819         if (btrfs_can_overcommit(fs_info, space_info, SZ_1M,
 820                                  BTRFS_RESERVE_FLUSH_ALL))
 821                 expected = div_factor_fine(space_info->total_bytes, 95);
 822         else
 823                 expected = div_factor_fine(space_info->total_bytes, 90);
 824
 825         if (used > expected)
 826                 to_reclaim = used - expected;
 827         else
 828                 to_reclaim = 0;
 829         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
 830                                      space_info->bytes_reserved);
 831         return to_reclaim;
 832 }
 833
 834 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
 835                                         struct btrfs_space_info *space_info,
 836                                         u64 used)
 837 {
 838         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
 839
 840         /* If we're just plain full then async reclaim just slows us down. */
 841         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
 842                 return 0;
 843
 844         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info))
 845                 return 0;
 846
 847         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
 848                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
 849 }
 850
 851 /*
 852  * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
 853  * @fs_info - fs_info for this fs
 854  * @space_info - the space info we were flushing
 855  *
 856  * We call this when we've exhausted our flushing ability and haven't made
 857  * progress in satisfying tickets.  The reservation code handles tickets in
 858  * order, so if there is a large ticket first and then smaller ones we could
 859  * very well satisfy the smaller tickets.  This will attempt to wake up any
 860  * tickets in the list to catch this case.
 861  *
 862  * This function returns true if it was able to make progress by clearing out
 863  * other tickets, or if it stumbles across a ticket that was smaller than the
 864  * first ticket.
 865  */
 866 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
 867                                    struct btrfs_space_info *space_info)
 868 {
 869         struct reserve_ticket *ticket;
 870         u64 tickets_id = space_info->tickets_id;
 871         u64 first_ticket_bytes = 0;
 872
 873         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
 874                 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
 875                 __btrfs_dump_space_info(fs_info, space_info);
 876         }
 877
 878         while (!list_empty(&space_info->tickets) &&
 879                tickets_id == space_info->tickets_id) {
 880                 ticket = list_first_entry(&space_info->tickets,
 881                                           struct reserve_ticket, list);
 882
 883                 /*
 884                  * may_commit_transaction will avoid committing the transaction
 885                  * if it doesn't feel like the space reclaimed by the commit
 886                  * would result in the ticket succeeding.  However if we have a
 887                  * smaller ticket in the queue it may be small enough to be
 888                  * satisified by committing the transaction, so if any
 889                  * subsequent ticket is smaller than the first ticket go ahead
 890                  * and send us back for another loop through the enospc flushing
 891                  * code.
 892                  */
 893                 if (first_ticket_bytes == 0)
 894                         first_ticket_bytes = ticket->bytes;
 895                 else if (first_ticket_bytes > ticket->bytes)
 896                         return true;
 897
 898                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
 899                         btrfs_info(fs_info, "failing ticket with %llu bytes",
 900                                    ticket->bytes);
 901
 902                 list_del_init(&ticket->list);
 903                 ticket->error = -ENOSPC;
 904                 wake_up(&ticket->wait);
 905
 906                 /*
 907                  * We're just throwing tickets away, so more flushing may not
 908                  * trip over btrfs_try_granting_tickets, so we need to call it
 909                  * here to see if we can make progress with the next ticket in
 910                  * the list.
 911                  */
 912                 btrfs_try_granting_tickets(fs_info, space_info);
 913         }
 914         return (tickets_id != space_info->tickets_id);
 915 }
 916
 917 /*
 918  * This is for normal flushers, we can wait all goddamned day if we want to.  We
 919  * will loop and continuously try to flush as long as we are making progress.
 920  * We count progress as clearing off tickets each time we have to loop.
 921  */
 922 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
 923 {
 924         struct btrfs_fs_info *fs_info;
 925         struct btrfs_space_info *space_info;
 926         u64 to_reclaim;
 927         int flush_state;
 928         int commit_cycles = 0;
 929         u64 last_tickets_id;
 930
 931         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
 932         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
 933
 934         spin_lock(&space_info->lock);
 935         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
 936         if (!to_reclaim) {
 937                 space_info->flush = 0;
 938                 spin_unlock(&space_info->lock);
 939                 return;
 940         }
 941         last_tickets_id = space_info->tickets_id;
 942         spin_unlock(&space_info->lock);
 943
 944         flush_state = FLUSH_DELAYED_ITEMS_NR;
 945         do {
 946                 flush_space(fs_info, space_info, to_reclaim, flush_state);
 947                 spin_lock(&space_info->lock);
 948                 if (list_empty(&space_info->tickets)) {
 949                         space_info->flush = 0;
 950                         spin_unlock(&space_info->lock);
 951                         return;
 952                 }
 953                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
 954                                                               space_info);
 955                 if (last_tickets_id == space_info->tickets_id) {
 956                         flush_state++;
 957                 } else {
 958                         last_tickets_id = space_info->tickets_id;
 959                         flush_state = FLUSH_DELAYED_ITEMS_NR;
 960                         if (commit_cycles)
 961                                 commit_cycles--;
 962                 }
 963
 964                 /*
 965                  * We don't want to force a chunk allocation until we've tried
 966                  * pretty hard to reclaim space.  Think of the case where we
 967                  * freed up a bunch of space and so have a lot of pinned space
 968                  * to reclaim.  We would rather use that than possibly create a
 969                  * underutilized metadata chunk.  So if this is our first run
 970                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
 971                  * commit the transaction.  If nothing has changed the next go
 972                  * around then we can force a chunk allocation.
 973                  */
 974                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
 975                         flush_state++;
 976
 977                 if (flush_state > COMMIT_TRANS) {
 978                         commit_cycles++;
 979                         if (commit_cycles > 2) {
 980                                 if (maybe_fail_all_tickets(fs_info, space_info)) {
 981                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
 982                                         commit_cycles--;
 983                                 } else {
 984                                         space_info->flush = 0;
 985                                 }
 986                         } else {
 987                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
 988                         }
 989                 }
 990                 spin_unlock(&space_info->lock);
 991         } while (flush_state <= COMMIT_TRANS);
 992 }
 993
 994 void btrfs_init_async_reclaim_work(struct work_struct *work)
 995 {
 996         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
 997 }
 998
 999 static const enum btrfs_flush_state priority_flush_states[] = {
1000         FLUSH_DELAYED_ITEMS_NR,
1001         FLUSH_DELAYED_ITEMS,
1002         ALLOC_CHUNK,
1003 };
1004
1005 static const enum btrfs_flush_state evict_flush_states[] = {
1006         FLUSH_DELAYED_ITEMS_NR,
1007         FLUSH_DELAYED_ITEMS,
1008         FLUSH_DELAYED_REFS_NR,
1009         FLUSH_DELAYED_REFS,
1010         FLUSH_DELALLOC,
1011         FLUSH_DELALLOC_WAIT,
1012         ALLOC_CHUNK,
1013         COMMIT_TRANS,
1014 };
1015
1016 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1017                                 struct btrfs_space_info *space_info,
1018                                 struct reserve_ticket *ticket,
1019                                 const enum btrfs_flush_state *states,
1020                                 int states_nr)
1021 {
1022         u64 to_reclaim;
1023         int flush_state;
1024
1025         spin_lock(&space_info->lock);
1026         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1027         if (!to_reclaim) {
1028                 spin_unlock(&space_info->lock);
1029                 return;
1030         }
1031         spin_unlock(&space_info->lock);
1032
1033         flush_state = 0;
1034         do {
1035                 flush_space(fs_info, space_info, to_reclaim, states[flush_state]);
1036                 flush_state++;
1037                 spin_lock(&space_info->lock);
1038                 if (ticket->bytes == 0) {
1039                         spin_unlock(&space_info->lock);
1040                         return;
1041                 }
1042                 spin_unlock(&space_info->lock);
1043         } while (flush_state < states_nr);
1044 }
1045
1046 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1047                                 struct btrfs_space_info *space_info,
1048                                 struct reserve_ticket *ticket)
1049
1050 {
1051         DEFINE_WAIT(wait);
1052         int ret = 0;
1053
1054         spin_lock(&space_info->lock);
1055         while (ticket->bytes > 0 && ticket->error == 0) {
1056                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1057                 if (ret) {
1058                         /*
1059                          * Delete us from the list. After we unlock the space
1060                          * info, we don't want the async reclaim job to reserve
1061                          * space for this ticket. If that would happen, then the
1062                          * ticket's task would not known that space was reserved
1063                          * despite getting an error, resulting in a space leak
1064                          * (bytes_may_use counter of our space_info).
1065                          */
1066                         list_del_init(&ticket->list);
1067                         ticket->error = -EINTR;
1068                         break;
1069                 }
1070                 spin_unlock(&space_info->lock);
1071
1072                 schedule();
1073
1074                 finish_wait(&ticket->wait, &wait);
1075                 spin_lock(&space_info->lock);
1076         }
1077         spin_unlock(&space_info->lock);
1078 }
1079
1080 /**
1081  * handle_reserve_ticket - do the appropriate flushing and waiting for a ticket
1082  * @fs_info - the fs
1083  * @space_info - the space_info for the reservation
1084  * @ticket - the ticket for the reservation
1085  * @flush - how much we can flush
1086  *
1087  * This does the work of figuring out how to flush for the ticket, waiting for
1088  * the reservation, and returning the appropriate error if there is one.
1089  */
1090 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1091                                  struct btrfs_space_info *space_info,
1092                                  struct reserve_ticket *ticket,
1093                                  enum btrfs_reserve_flush_enum flush)
1094 {
1095         int ret;
1096
1097         switch (flush) {
1098         case BTRFS_RESERVE_FLUSH_ALL:
1099                 wait_reserve_ticket(fs_info, space_info, ticket);
1100                 break;
1101         case BTRFS_RESERVE_FLUSH_LIMIT:
1102                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1103                                                 priority_flush_states,
1104                                                 ARRAY_SIZE(priority_flush_states));
1105                 break;
1106         case BTRFS_RESERVE_FLUSH_EVICT:
1107                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1108                                                 evict_flush_states,
1109                                                 ARRAY_SIZE(evict_flush_states));
1110                 break;
1111         default:
1112                 ASSERT(0);
1113                 break;
1114         }
1115
1116         spin_lock(&space_info->lock);
1117         ret = ticket->error;
1118         if (ticket->bytes || ticket->error) {
1119                 /*
1120                  * Need to delete here for priority tickets. For regular tickets
1121                  * either the async reclaim job deletes the ticket from the list
1122                  * or we delete it ourselves at wait_reserve_ticket().
1123                  */
1124                 list_del_init(&ticket->list);
1125                 if (!ret)
1126                         ret = -ENOSPC;
1127         }
1128         spin_unlock(&space_info->lock);
1129         ASSERT(list_empty(&ticket->list));
1130         /*
1131          * Check that we can't have an error set if the reservation succeeded,
1132          * as that would confuse tasks and lead them to error out without
1133          * releasing reserved space (if an error happens the expectation is that
1134          * space wasn't reserved at all).
1135          */
1136         ASSERT(!(ticket->bytes == 0 && ticket->error));
1137         return ret;
1138 }
1139
1140 /**
1141  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1142  * @root - the root we're allocating for
1143  * @space_info - the space info we want to allocate from
1144  * @orig_bytes - the number of bytes we want
1145  * @flush - whether or not we can flush to make our reservation
1146  *
1147  * This will reserve orig_bytes number of bytes from the space info associated
1148  * with the block_rsv.  If there is not enough space it will make an attempt to
1149  * flush out space to make room.  It will do this by flushing delalloc if
1150  * possible or committing the transaction.  If flush is 0 then no attempts to
1151  * regain reservations will be made and this will fail if there is not enough
1152  * space already.
1153  */
1154 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1155                                     struct btrfs_space_info *space_info,
1156                                     u64 orig_bytes,
1157                                     enum btrfs_reserve_flush_enum flush)
1158 {
1159         struct reserve_ticket ticket;
1160         u64 used;
1161         int ret = 0;
1162         bool pending_tickets;
1163
1164         ASSERT(orig_bytes);
1165         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1166
1167         spin_lock(&space_info->lock);
1168         ret = -ENOSPC;
1169         used = btrfs_space_info_used(space_info, true);
1170         pending_tickets = !list_empty(&space_info->tickets) ||
1171                 !list_empty(&space_info->priority_tickets);
1172
1173         /*
1174          * Carry on if we have enough space (short-circuit) OR call
1175          * can_overcommit() to ensure we can overcommit to continue.
1176          */
1177         if (!pending_tickets &&
1178             ((used + orig_bytes <= space_info->total_bytes) ||
1179              btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1180                 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1181                                                       orig_bytes);
1182                 ret = 0;
1183         }
1184
1185         /*
1186          * If we couldn't make a reservation then setup our reservation ticket
1187          * and kick the async worker if it's not already running.
1188          *
1189          * If we are a priority flusher then we just need to add our ticket to
1190          * the list and we will do our own flushing further down.
1191          */
1192         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1193                 ASSERT(space_info->reclaim_size >= 0);
1194                 ticket.bytes = orig_bytes;
1195                 ticket.error = 0;
1196                 space_info->reclaim_size += ticket.bytes;
1197                 init_waitqueue_head(&ticket.wait);
1198                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
1199                         list_add_tail(&ticket.list, &space_info->tickets);
1200                         if (!space_info->flush) {
1201                                 space_info->flush = 1;
1202                                 trace_btrfs_trigger_flush(fs_info,
1203                                                           space_info->flags,
1204                                                           orig_bytes, flush,
1205                                                           "enospc");
1206                                 queue_work(system_unbound_wq,
1207                                            &fs_info->async_reclaim_work);
1208                         }
1209                 } else {
1210                         list_add_tail(&ticket.list,
1211                                       &space_info->priority_tickets);
1212                 }
1213         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1214                 used += orig_bytes;
1215                 /*
1216                  * We will do the space reservation dance during log replay,
1217                  * which means we won't have fs_info->fs_root set, so don't do
1218                  * the async reclaim as we will panic.
1219                  */
1220                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1221                     need_do_async_reclaim(fs_info, space_info, used) &&
1222                     !work_busy(&fs_info->async_reclaim_work)) {
1223                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
1224                                                   orig_bytes, flush, "preempt");
1225                         queue_work(system_unbound_wq,
1226                                    &fs_info->async_reclaim_work);
1227                 }
1228         }
1229         spin_unlock(&space_info->lock);
1230         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1231                 return ret;
1232
1233         return handle_reserve_ticket(fs_info, space_info, &ticket, flush);
1234 }
1235
1236 /**
1237  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1238  * @root - the root we're allocating for
1239  * @block_rsv - the block_rsv we're allocating for
1240  * @orig_bytes - the number of bytes we want
1241  * @flush - whether or not we can flush to make our reservation
1242  *
1243  * This will reserve orig_bytes number of bytes from the space info associated
1244  * with the block_rsv.  If there is not enough space it will make an attempt to
1245  * flush out space to make room.  It will do this by flushing delalloc if
1246  * possible or committing the transaction.  If flush is 0 then no attempts to
1247  * regain reservations will be made and this will fail if there is not enough
1248  * space already.
1249  */
1250 int btrfs_reserve_metadata_bytes(struct btrfs_root *root,
1251                                  struct btrfs_block_rsv *block_rsv,
1252                                  u64 orig_bytes,
1253                                  enum btrfs_reserve_flush_enum flush)
1254 {
1255         struct btrfs_fs_info *fs_info = root->fs_info;
1256         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
1257         int ret;
1258
1259         ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
1260                                        orig_bytes, flush);
1261         if (ret == -ENOSPC &&
1262             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
1263                 if (block_rsv != global_rsv &&
1264                     !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes))
1265                         ret = 0;
1266         }
1267         if (ret == -ENOSPC) {
1268                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1269                                               block_rsv->space_info->flags,
1270                                               orig_bytes, 1);
1271
1272                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1273                         btrfs_dump_space_info(fs_info, block_rsv->space_info,
1274                                               orig_bytes, 0);
1275         }
1276         return ret;
1277 }