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 #include "zoned.h"
  13 #include "fs.h"
  14 #include "accessors.h"
  15 #include "extent-tree.h"
  16
  17 /*
  18  * HOW DOES SPACE RESERVATION WORK
  19  *
  20  * If you want to know about delalloc specifically, there is a separate comment
  21  * for that with the delalloc code.  This comment is about how the whole system
  22  * works generally.
  23  *
  24  * BASIC CONCEPTS
  25  *
  26  *   1) space_info.  This is the ultimate arbiter of how much space we can use.
  27  *   There's a description of the bytes_ fields with the struct declaration,
  28  *   refer to that for specifics on each field.  Suffice it to say that for
  29  *   reservations we care about total_bytes - SUM(space_info->bytes_) when
  30  *   determining if there is space to make an allocation.  There is a space_info
  31  *   for METADATA, SYSTEM, and DATA areas.
  32  *
  33  *   2) block_rsv's.  These are basically buckets for every different type of
  34  *   metadata reservation we have.  You can see the comment in the block_rsv
  35  *   code on the rules for each type, but generally block_rsv->reserved is how
  36  *   much space is accounted for in space_info->bytes_may_use.
  37  *
  38  *   3) btrfs_calc*_size.  These are the worst case calculations we used based
  39  *   on the number of items we will want to modify.  We have one for changing
  40  *   items, and one for inserting new items.  Generally we use these helpers to
  41  *   determine the size of the block reserves, and then use the actual bytes
  42  *   values to adjust the space_info counters.
  43  *
  44  * MAKING RESERVATIONS, THE NORMAL CASE
  45  *
  46  *   We call into either btrfs_reserve_data_bytes() or
  47  *   btrfs_reserve_metadata_bytes(), depending on which we're looking for, with
  48  *   num_bytes we want to reserve.
  49  *
  50  *   ->reserve
  51  *     space_info->bytes_may_reserve += num_bytes
  52  *
  53  *   ->extent allocation
  54  *     Call btrfs_add_reserved_bytes() which does
  55  *     space_info->bytes_may_reserve -= num_bytes
  56  *     space_info->bytes_reserved += extent_bytes
  57  *
  58  *   ->insert reference
  59  *     Call btrfs_update_block_group() which does
  60  *     space_info->bytes_reserved -= extent_bytes
  61  *     space_info->bytes_used += extent_bytes
  62  *
  63  * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
  64  *
  65  *   Assume we are unable to simply make the reservation because we do not have
  66  *   enough space
  67  *
  68  *   -> __reserve_bytes
  69  *     create a reserve_ticket with ->bytes set to our reservation, add it to
  70  *     the tail of space_info->tickets, kick async flush thread
  71  *
  72  *   ->handle_reserve_ticket
  73  *     wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
  74  *     on the ticket.
  75  *
  76  *   -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
  77  *     Flushes various things attempting to free up space.
  78  *
  79  *   -> btrfs_try_granting_tickets()
  80  *     This is called by anything that either subtracts space from
  81  *     space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the
  82  *     space_info->total_bytes.  This loops through the ->priority_tickets and
  83  *     then the ->tickets list checking to see if the reservation can be
  84  *     completed.  If it can the space is added to space_info->bytes_may_use and
  85  *     the ticket is woken up.
  86  *
  87  *   -> ticket wakeup
  88  *     Check if ->bytes == 0, if it does we got our reservation and we can carry
  89  *     on, if not return the appropriate error (ENOSPC, but can be EINTR if we
  90  *     were interrupted.)
  91  *
  92  * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
  93  *
  94  *   Same as the above, except we add ourselves to the
  95  *   space_info->priority_tickets, and we do not use ticket->wait, we simply
  96  *   call flush_space() ourselves for the states that are safe for us to call
  97  *   without deadlocking and hope for the best.
  98  *
  99  * THE FLUSHING STATES
 100  *
 101  *   Generally speaking we will have two cases for each state, a "nice" state
 102  *   and a "ALL THE THINGS" state.  In btrfs we delay a lot of work in order to
 103  *   reduce the locking over head on the various trees, and even to keep from
 104  *   doing any work at all in the case of delayed refs.  Each of these delayed
 105  *   things however hold reservations, and so letting them run allows us to
 106  *   reclaim space so we can make new reservations.
 107  *
 108  *   FLUSH_DELAYED_ITEMS
 109  *     Every inode has a delayed item to update the inode.  Take a simple write
 110  *     for example, we would update the inode item at write time to update the
 111  *     mtime, and then again at finish_ordered_io() time in order to update the
 112  *     isize or bytes.  We keep these delayed items to coalesce these operations
 113  *     into a single operation done on demand.  These are an easy way to reclaim
 114  *     metadata space.
 115  *
 116  *   FLUSH_DELALLOC
 117  *     Look at the delalloc comment to get an idea of how much space is reserved
 118  *     for delayed allocation.  We can reclaim some of this space simply by
 119  *     running delalloc, but usually we need to wait for ordered extents to
 120  *     reclaim the bulk of this space.
 121  *
 122  *   FLUSH_DELAYED_REFS
 123  *     We have a block reserve for the outstanding delayed refs space, and every
 124  *     delayed ref operation holds a reservation.  Running these is a quick way
 125  *     to reclaim space, but we want to hold this until the end because COW can
 126  *     churn a lot and we can avoid making some extent tree modifications if we
 127  *     are able to delay for as long as possible.
 128  *
 129  *   ALLOC_CHUNK
 130  *     We will skip this the first time through space reservation, because of
 131  *     overcommit and we don't want to have a lot of useless metadata space when
 132  *     our worst case reservations will likely never come true.
 133  *
 134  *   RUN_DELAYED_IPUTS
 135  *     If we're freeing inodes we're likely freeing checksums, file extent
 136  *     items, and extent tree items.  Loads of space could be freed up by these
 137  *     operations, however they won't be usable until the transaction commits.
 138  *
 139  *   COMMIT_TRANS
 140  *     This will commit the transaction.  Historically we had a lot of logic
 141  *     surrounding whether or not we'd commit the transaction, but this waits born
 142  *     out of a pre-tickets era where we could end up committing the transaction
 143  *     thousands of times in a row without making progress.  Now thanks to our
 144  *     ticketing system we know if we're not making progress and can error
 145  *     everybody out after a few commits rather than burning the disk hoping for
 146  *     a different answer.
 147  *
 148  * OVERCOMMIT
 149  *
 150  *   Because we hold so many reservations for metadata we will allow you to
 151  *   reserve more space than is currently free in the currently allocate
 152  *   metadata space.  This only happens with metadata, data does not allow
 153  *   overcommitting.
 154  *
 155  *   You can see the current logic for when we allow overcommit in
 156  *   btrfs_can_overcommit(), but it only applies to unallocated space.  If there
 157  *   is no unallocated space to be had, all reservations are kept within the
 158  *   free space in the allocated metadata chunks.
 159  *
 160  *   Because of overcommitting, you generally want to use the
 161  *   btrfs_can_overcommit() logic for metadata allocations, as it does the right
 162  *   thing with or without extra unallocated space.
 163  */
 164
 165 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
 166                           bool may_use_included)
 167 {
 168         ASSERT(s_info);
 169         return s_info->bytes_used + s_info->bytes_reserved +
 170                 s_info->bytes_pinned + s_info->bytes_readonly +
 171                 s_info->bytes_zone_unusable +
 172                 (may_use_included ? s_info->bytes_may_use : 0);
 173 }
 174
 175 /*
 176  * after adding space to the filesystem, we need to clear the full flags
 177  * on all the space infos.
 178  */
 179 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 180 {
 181         struct list_head *head = &info->space_info;
 182         struct btrfs_space_info *found;
 183
 184         list_for_each_entry(found, head, list)
 185                 found->full = 0;
 186 }
 187
 188 /*
 189  * Block groups with more than this value (percents) of unusable space will be
 190  * scheduled for background reclaim.
 191  */
 192 #define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH                      (75)
 193
 194 /*
 195  * Calculate chunk size depending on volume type (regular or zoned).
 196  */
 197 static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags)
 198 {
 199         if (btrfs_is_zoned(fs_info))
 200                 return fs_info->zone_size;
 201
 202         ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
 203
 204         if (flags & BTRFS_BLOCK_GROUP_DATA)
 205                 return BTRFS_MAX_DATA_CHUNK_SIZE;
 206         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 207                 return SZ_32M;
 208
 209         /* Handle BTRFS_BLOCK_GROUP_METADATA */
 210         if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G)
 211                 return SZ_1G;
 212
 213         return SZ_256M;
 214 }
 215
 216 /*
 217  * Update default chunk size.
 218  */
 219 void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info,
 220                                         u64 chunk_size)
 221 {
 222         WRITE_ONCE(space_info->chunk_size, chunk_size);
 223 }
 224
 225 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
 226 {
 227
 228         struct btrfs_space_info *space_info;
 229         int i;
 230         int ret;
 231
 232         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
 233         if (!space_info)
 234                 return -ENOMEM;
 235
 236         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
 237                 INIT_LIST_HEAD(&space_info->block_groups[i]);
 238         init_rwsem(&space_info->groups_sem);
 239         spin_lock_init(&space_info->lock);
 240         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
 241         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
 242         INIT_LIST_HEAD(&space_info->ro_bgs);
 243         INIT_LIST_HEAD(&space_info->tickets);
 244         INIT_LIST_HEAD(&space_info->priority_tickets);
 245         space_info->clamp = 1;
 246         btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags));
 247
 248         if (btrfs_is_zoned(info))
 249                 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH;
 250
 251         ret = btrfs_sysfs_add_space_info_type(info, space_info);
 252         if (ret)
 253                 return ret;
 254
 255         list_add(&space_info->list, &info->space_info);
 256         if (flags & BTRFS_BLOCK_GROUP_DATA)
 257                 info->data_sinfo = space_info;
 258
 259         return ret;
 260 }
 261
 262 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
 263 {
 264         struct btrfs_super_block *disk_super;
 265         u64 features;
 266         u64 flags;
 267         int mixed = 0;
 268         int ret;
 269
 270         disk_super = fs_info->super_copy;
 271         if (!btrfs_super_root(disk_super))
 272                 return -EINVAL;
 273
 274         features = btrfs_super_incompat_flags(disk_super);
 275         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
 276                 mixed = 1;
 277
 278         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 279         ret = create_space_info(fs_info, flags);
 280         if (ret)
 281                 goto out;
 282
 283         if (mixed) {
 284                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
 285                 ret = create_space_info(fs_info, flags);
 286         } else {
 287                 flags = BTRFS_BLOCK_GROUP_METADATA;
 288                 ret = create_space_info(fs_info, flags);
 289                 if (ret)
 290                         goto out;
 291
 292                 flags = BTRFS_BLOCK_GROUP_DATA;
 293                 ret = create_space_info(fs_info, flags);
 294         }
 295 out:
 296         return ret;
 297 }
 298
 299 void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info,
 300                                 struct btrfs_block_group *block_group)
 301 {
 302         struct btrfs_space_info *found;
 303         int factor, index;
 304
 305         factor = btrfs_bg_type_to_factor(block_group->flags);
 306
 307         found = btrfs_find_space_info(info, block_group->flags);
 308         ASSERT(found);
 309         spin_lock(&found->lock);
 310         found->total_bytes += block_group->length;
 311         if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags))
 312                 found->active_total_bytes += block_group->length;
 313         found->disk_total += block_group->length * factor;
 314         found->bytes_used += block_group->used;
 315         found->disk_used += block_group->used * factor;
 316         found->bytes_readonly += block_group->bytes_super;
 317         found->bytes_zone_unusable += block_group->zone_unusable;
 318         if (block_group->length > 0)
 319                 found->full = 0;
 320         btrfs_try_granting_tickets(info, found);
 321         spin_unlock(&found->lock);
 322
 323         block_group->space_info = found;
 324
 325         index = btrfs_bg_flags_to_raid_index(block_group->flags);
 326         down_write(&found->groups_sem);
 327         list_add_tail(&block_group->list, &found->block_groups[index]);
 328         up_write(&found->groups_sem);
 329 }
 330
 331 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
 332                                                u64 flags)
 333 {
 334         struct list_head *head = &info->space_info;
 335         struct btrfs_space_info *found;
 336
 337         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 338
 339         list_for_each_entry(found, head, list) {
 340                 if (found->flags & flags)
 341                         return found;
 342         }
 343         return NULL;
 344 }
 345
 346 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
 347                           struct btrfs_space_info *space_info,
 348                           enum btrfs_reserve_flush_enum flush)
 349 {
 350         u64 profile;
 351         u64 avail;
 352         int factor;
 353
 354         if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
 355                 profile = btrfs_system_alloc_profile(fs_info);
 356         else
 357                 profile = btrfs_metadata_alloc_profile(fs_info);
 358
 359         avail = atomic64_read(&fs_info->free_chunk_space);
 360
 361         /*
 362          * If we have dup, raid1 or raid10 then only half of the free
 363          * space is actually usable.  For raid56, the space info used
 364          * doesn't include the parity drive, so we don't have to
 365          * change the math
 366          */
 367         factor = btrfs_bg_type_to_factor(profile);
 368         avail = div_u64(avail, factor);
 369
 370         /*
 371          * If we aren't flushing all things, let us overcommit up to
 372          * 1/2th of the space. If we can flush, don't let us overcommit
 373          * too much, let it overcommit up to 1/8 of the space.
 374          */
 375         if (flush == BTRFS_RESERVE_FLUSH_ALL)
 376                 avail >>= 3;
 377         else
 378                 avail >>= 1;
 379         return avail;
 380 }
 381
 382 static inline u64 writable_total_bytes(struct btrfs_fs_info *fs_info,
 383                                        struct btrfs_space_info *space_info)
 384 {
 385         /*
 386          * On regular filesystem, all total_bytes are always writable. On zoned
 387          * filesystem, there may be a limitation imposed by max_active_zones.
 388          * For metadata allocation, we cannot finish an existing active block
 389          * group to avoid a deadlock. Thus, we need to consider only the active
 390          * groups to be writable for metadata space.
 391          */
 392         if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA))
 393                 return space_info->total_bytes;
 394
 395         return space_info->active_total_bytes;
 396 }
 397
 398 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
 399                          struct btrfs_space_info *space_info, u64 bytes,
 400                          enum btrfs_reserve_flush_enum flush)
 401 {
 402         u64 avail;
 403         u64 used;
 404
 405         /* Don't overcommit when in mixed mode */
 406         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
 407                 return 0;
 408
 409         used = btrfs_space_info_used(space_info, true);
 410         if (test_bit(BTRFS_FS_NO_OVERCOMMIT, &fs_info->flags) &&
 411             (space_info->flags & BTRFS_BLOCK_GROUP_METADATA))
 412                 avail = 0;
 413         else
 414                 avail = calc_available_free_space(fs_info, space_info, flush);
 415
 416         if (used + bytes < writable_total_bytes(fs_info, space_info) + avail)
 417                 return 1;
 418         return 0;
 419 }
 420
 421 static void remove_ticket(struct btrfs_space_info *space_info,
 422                           struct reserve_ticket *ticket)
 423 {
 424         if (!list_empty(&ticket->list)) {
 425                 list_del_init(&ticket->list);
 426                 ASSERT(space_info->reclaim_size >= ticket->bytes);
 427                 space_info->reclaim_size -= ticket->bytes;
 428         }
 429 }
 430
 431 /*
 432  * This is for space we already have accounted in space_info->bytes_may_use, so
 433  * basically when we're returning space from block_rsv's.
 434  */
 435 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
 436                                 struct btrfs_space_info *space_info)
 437 {
 438         struct list_head *head;
 439         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
 440
 441         lockdep_assert_held(&space_info->lock);
 442
 443         head = &space_info->priority_tickets;
 444 again:
 445         while (!list_empty(head)) {
 446                 struct reserve_ticket *ticket;
 447                 u64 used = btrfs_space_info_used(space_info, true);
 448
 449                 ticket = list_first_entry(head, struct reserve_ticket, list);
 450
 451                 /* Check and see if our ticket can be satisfied now. */
 452                 if ((used + ticket->bytes <= writable_total_bytes(fs_info, space_info)) ||
 453                     btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
 454                                          flush)) {
 455                         btrfs_space_info_update_bytes_may_use(fs_info,
 456                                                               space_info,
 457                                                               ticket->bytes);
 458                         remove_ticket(space_info, ticket);
 459                         ticket->bytes = 0;
 460                         space_info->tickets_id++;
 461                         wake_up(&ticket->wait);
 462                 } else {
 463                         break;
 464                 }
 465         }
 466
 467         if (head == &space_info->priority_tickets) {
 468                 head = &space_info->tickets;
 469                 flush = BTRFS_RESERVE_FLUSH_ALL;
 470                 goto again;
 471         }
 472 }
 473
 474 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
 475 do {                                                                    \
 476         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
 477         spin_lock(&__rsv->lock);                                        \
 478         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
 479                    __rsv->size, __rsv->reserved);                       \
 480         spin_unlock(&__rsv->lock);                                      \
 481 } while (0)
 482
 483 static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info)
 484 {
 485         switch (space_info->flags) {
 486         case BTRFS_BLOCK_GROUP_SYSTEM:
 487                 return "SYSTEM";
 488         case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA:
 489                 return "DATA+METADATA";
 490         case BTRFS_BLOCK_GROUP_DATA:
 491                 return "DATA";
 492         case BTRFS_BLOCK_GROUP_METADATA:
 493                 return "METADATA";
 494         default:
 495                 return "UNKNOWN";
 496         }
 497 }
 498
 499 static void dump_global_block_rsv(struct btrfs_fs_info *fs_info)
 500 {
 501         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
 502         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
 503         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
 504         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
 505         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
 506 }
 507
 508 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 509                                     struct btrfs_space_info *info)
 510 {
 511         const char *flag_str = space_info_flag_to_str(info);
 512         lockdep_assert_held(&info->lock);
 513
 514         /* The free space could be negative in case of overcommit */
 515         btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
 516                    flag_str,
 517                    (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
 518                    info->full ? "" : "not ");
 519         btrfs_info(fs_info,
 520 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
 521                 info->total_bytes, info->bytes_used, info->bytes_pinned,
 522                 info->bytes_reserved, info->bytes_may_use,
 523                 info->bytes_readonly, info->bytes_zone_unusable);
 524 }
 525
 526 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 527                            struct btrfs_space_info *info, u64 bytes,
 528                            int dump_block_groups)
 529 {
 530         struct btrfs_block_group *cache;
 531         int index = 0;
 532
 533         spin_lock(&info->lock);
 534         __btrfs_dump_space_info(fs_info, info);
 535         dump_global_block_rsv(fs_info);
 536         spin_unlock(&info->lock);
 537
 538         if (!dump_block_groups)
 539                 return;
 540
 541         down_read(&info->groups_sem);
 542 again:
 543         list_for_each_entry(cache, &info->block_groups[index], list) {
 544                 spin_lock(&cache->lock);
 545                 btrfs_info(fs_info,
 546                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s",
 547                         cache->start, cache->length, cache->used, cache->pinned,
 548                         cache->reserved, cache->zone_unusable,
 549                         cache->ro ? "[readonly]" : "");
 550                 spin_unlock(&cache->lock);
 551                 btrfs_dump_free_space(cache, bytes);
 552         }
 553         if (++index < BTRFS_NR_RAID_TYPES)
 554                 goto again;
 555         up_read(&info->groups_sem);
 556 }
 557
 558 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
 559                                         u64 to_reclaim)
 560 {
 561         u64 bytes;
 562         u64 nr;
 563
 564         bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 565         nr = div64_u64(to_reclaim, bytes);
 566         if (!nr)
 567                 nr = 1;
 568         return nr;
 569 }
 570
 571 #define EXTENT_SIZE_PER_ITEM    SZ_256K
 572
 573 /*
 574  * shrink metadata reservation for delalloc
 575  */
 576 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
 577                             struct btrfs_space_info *space_info,
 578                             u64 to_reclaim, bool wait_ordered,
 579                             bool for_preempt)
 580 {
 581         struct btrfs_trans_handle *trans;
 582         u64 delalloc_bytes;
 583         u64 ordered_bytes;
 584         u64 items;
 585         long time_left;
 586         int loops;
 587
 588         delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
 589         ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
 590         if (delalloc_bytes == 0 && ordered_bytes == 0)
 591                 return;
 592
 593         /* Calc the number of the pages we need flush for space reservation */
 594         if (to_reclaim == U64_MAX) {
 595                 items = U64_MAX;
 596         } else {
 597                 /*
 598                  * to_reclaim is set to however much metadata we need to
 599                  * reclaim, but reclaiming that much data doesn't really track
 600                  * exactly.  What we really want to do is reclaim full inode's
 601                  * worth of reservations, however that's not available to us
 602                  * here.  We will take a fraction of the delalloc bytes for our
 603                  * flushing loops and hope for the best.  Delalloc will expand
 604                  * the amount we write to cover an entire dirty extent, which
 605                  * will reclaim the metadata reservation for that range.  If
 606                  * it's not enough subsequent flush stages will be more
 607                  * aggressive.
 608                  */
 609                 to_reclaim = max(to_reclaim, delalloc_bytes >> 3);
 610                 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
 611         }
 612
 613         trans = current->journal_info;
 614
 615         /*
 616          * If we are doing more ordered than delalloc we need to just wait on
 617          * ordered extents, otherwise we'll waste time trying to flush delalloc
 618          * that likely won't give us the space back we need.
 619          */
 620         if (ordered_bytes > delalloc_bytes && !for_preempt)
 621                 wait_ordered = true;
 622
 623         loops = 0;
 624         while ((delalloc_bytes || ordered_bytes) && loops < 3) {
 625                 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
 626                 long nr_pages = min_t(u64, temp, LONG_MAX);
 627                 int async_pages;
 628
 629                 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
 630
 631                 /*
 632                  * We need to make sure any outstanding async pages are now
 633                  * processed before we continue.  This is because things like
 634                  * sync_inode() try to be smart and skip writing if the inode is
 635                  * marked clean.  We don't use filemap_fwrite for flushing
 636                  * because we want to control how many pages we write out at a
 637                  * time, thus this is the only safe way to make sure we've
 638                  * waited for outstanding compressed workers to have started
 639                  * their jobs and thus have ordered extents set up properly.
 640                  *
 641                  * This exists because we do not want to wait for each
 642                  * individual inode to finish its async work, we simply want to
 643                  * start the IO on everybody, and then come back here and wait
 644                  * for all of the async work to catch up.  Once we're done with
 645                  * that we know we'll have ordered extents for everything and we
 646                  * can decide if we wait for that or not.
 647                  *
 648                  * If we choose to replace this in the future, make absolutely
 649                  * sure that the proper waiting is being done in the async case,
 650                  * as there have been bugs in that area before.
 651                  */
 652                 async_pages = atomic_read(&fs_info->async_delalloc_pages);
 653                 if (!async_pages)
 654                         goto skip_async;
 655
 656                 /*
 657                  * We don't want to wait forever, if we wrote less pages in this
 658                  * loop than we have outstanding, only wait for that number of
 659                  * pages, otherwise we can wait for all async pages to finish
 660                  * before continuing.
 661                  */
 662                 if (async_pages > nr_pages)
 663                         async_pages -= nr_pages;
 664                 else
 665                         async_pages = 0;
 666                 wait_event(fs_info->async_submit_wait,
 667                            atomic_read(&fs_info->async_delalloc_pages) <=
 668                            async_pages);
 669 skip_async:
 670                 loops++;
 671                 if (wait_ordered && !trans) {
 672                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
 673                 } else {
 674                         time_left = schedule_timeout_killable(1);
 675                         if (time_left)
 676                                 break;
 677                 }
 678
 679                 /*
 680                  * If we are for preemption we just want a one-shot of delalloc
 681                  * flushing so we can stop flushing if we decide we don't need
 682                  * to anymore.
 683                  */
 684                 if (for_preempt)
 685                         break;
 686
 687                 spin_lock(&space_info->lock);
 688                 if (list_empty(&space_info->tickets) &&
 689                     list_empty(&space_info->priority_tickets)) {
 690                         spin_unlock(&space_info->lock);
 691                         break;
 692                 }
 693                 spin_unlock(&space_info->lock);
 694
 695                 delalloc_bytes = percpu_counter_sum_positive(
 696                                                 &fs_info->delalloc_bytes);
 697                 ordered_bytes = percpu_counter_sum_positive(
 698                                                 &fs_info->ordered_bytes);
 699         }
 700 }
 701
 702 /*
 703  * Try to flush some data based on policy set by @state. This is only advisory
 704  * and may fail for various reasons. The caller is supposed to examine the
 705  * state of @space_info to detect the outcome.
 706  */
 707 static void flush_space(struct btrfs_fs_info *fs_info,
 708                        struct btrfs_space_info *space_info, u64 num_bytes,
 709                        enum btrfs_flush_state state, bool for_preempt)
 710 {
 711         struct btrfs_root *root = fs_info->tree_root;
 712         struct btrfs_trans_handle *trans;
 713         int nr;
 714         int ret = 0;
 715
 716         switch (state) {
 717         case FLUSH_DELAYED_ITEMS_NR:
 718         case FLUSH_DELAYED_ITEMS:
 719                 if (state == FLUSH_DELAYED_ITEMS_NR)
 720                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
 721                 else
 722                         nr = -1;
 723
 724                 trans = btrfs_join_transaction(root);
 725                 if (IS_ERR(trans)) {
 726                         ret = PTR_ERR(trans);
 727                         break;
 728                 }
 729                 ret = btrfs_run_delayed_items_nr(trans, nr);
 730                 btrfs_end_transaction(trans);
 731                 break;
 732         case FLUSH_DELALLOC:
 733         case FLUSH_DELALLOC_WAIT:
 734         case FLUSH_DELALLOC_FULL:
 735                 if (state == FLUSH_DELALLOC_FULL)
 736                         num_bytes = U64_MAX;
 737                 shrink_delalloc(fs_info, space_info, num_bytes,
 738                                 state != FLUSH_DELALLOC, for_preempt);
 739                 break;
 740         case FLUSH_DELAYED_REFS_NR:
 741         case FLUSH_DELAYED_REFS:
 742                 trans = btrfs_join_transaction(root);
 743                 if (IS_ERR(trans)) {
 744                         ret = PTR_ERR(trans);
 745                         break;
 746                 }
 747                 if (state == FLUSH_DELAYED_REFS_NR)
 748                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
 749                 else
 750                         nr = 0;
 751                 btrfs_run_delayed_refs(trans, nr);
 752                 btrfs_end_transaction(trans);
 753                 break;
 754         case ALLOC_CHUNK:
 755         case ALLOC_CHUNK_FORCE:
 756                 /*
 757                  * For metadata space on zoned filesystem, reaching here means we
 758                  * don't have enough space left in active_total_bytes. Try to
 759                  * activate a block group first, because we may have inactive
 760                  * block group already allocated.
 761                  */
 762                 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, false);
 763                 if (ret < 0)
 764                         break;
 765                 else if (ret == 1)
 766                         break;
 767
 768                 trans = btrfs_join_transaction(root);
 769                 if (IS_ERR(trans)) {
 770                         ret = PTR_ERR(trans);
 771                         break;
 772                 }
 773                 ret = btrfs_chunk_alloc(trans,
 774                                 btrfs_get_alloc_profile(fs_info, space_info->flags),
 775                                 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
 776                                         CHUNK_ALLOC_FORCE);
 777                 btrfs_end_transaction(trans);
 778
 779                 /*
 780                  * For metadata space on zoned filesystem, allocating a new chunk
 781                  * is not enough. We still need to activate the block * group.
 782                  * Active the newly allocated block group by (maybe) finishing
 783                  * a block group.
 784                  */
 785                 if (ret == 1) {
 786                         ret = btrfs_zoned_activate_one_bg(fs_info, space_info, true);
 787                         /*
 788                          * Revert to the original ret regardless we could finish
 789                          * one block group or not.
 790                          */
 791                         if (ret >= 0)
 792                                 ret = 1;
 793                 }
 794
 795                 if (ret > 0 || ret == -ENOSPC)
 796                         ret = 0;
 797                 break;
 798         case RUN_DELAYED_IPUTS:
 799                 /*
 800                  * If we have pending delayed iputs then we could free up a
 801                  * bunch of pinned space, so make sure we run the iputs before
 802                  * we do our pinned bytes check below.
 803                  */
 804                 btrfs_run_delayed_iputs(fs_info);
 805                 btrfs_wait_on_delayed_iputs(fs_info);
 806                 break;
 807         case COMMIT_TRANS:
 808                 ASSERT(current->journal_info == NULL);
 809                 trans = btrfs_join_transaction(root);
 810                 if (IS_ERR(trans)) {
 811                         ret = PTR_ERR(trans);
 812                         break;
 813                 }
 814                 ret = btrfs_commit_transaction(trans);
 815                 break;
 816         default:
 817                 ret = -ENOSPC;
 818                 break;
 819         }
 820
 821         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
 822                                 ret, for_preempt);
 823         return;
 824 }
 825
 826 static inline u64
 827 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
 828                                  struct btrfs_space_info *space_info)
 829 {
 830         u64 used;
 831         u64 avail;
 832         u64 total;
 833         u64 to_reclaim = space_info->reclaim_size;
 834
 835         lockdep_assert_held(&space_info->lock);
 836
 837         avail = calc_available_free_space(fs_info, space_info,
 838                                           BTRFS_RESERVE_FLUSH_ALL);
 839         used = btrfs_space_info_used(space_info, true);
 840
 841         /*
 842          * We may be flushing because suddenly we have less space than we had
 843          * before, and now we're well over-committed based on our current free
 844          * space.  If that's the case add in our overage so we make sure to put
 845          * appropriate pressure on the flushing state machine.
 846          */
 847         total = writable_total_bytes(fs_info, space_info);
 848         if (total + avail < used)
 849                 to_reclaim += used - (total + avail);
 850
 851         return to_reclaim;
 852 }
 853
 854 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
 855                                     struct btrfs_space_info *space_info)
 856 {
 857         u64 global_rsv_size = fs_info->global_block_rsv.reserved;
 858         u64 ordered, delalloc;
 859         u64 total = writable_total_bytes(fs_info, space_info);
 860         u64 thresh;
 861         u64 used;
 862
 863         thresh = mult_perc(total, 90);
 864
 865         lockdep_assert_held(&space_info->lock);
 866
 867         /* If we're just plain full then async reclaim just slows us down. */
 868         if ((space_info->bytes_used + space_info->bytes_reserved +
 869              global_rsv_size) >= thresh)
 870                 return false;
 871
 872         used = space_info->bytes_may_use + space_info->bytes_pinned;
 873
 874         /* The total flushable belongs to the global rsv, don't flush. */
 875         if (global_rsv_size >= used)
 876                 return false;
 877
 878         /*
 879          * 128MiB is 1/4 of the maximum global rsv size.  If we have less than
 880          * that devoted to other reservations then there's no sense in flushing,
 881          * we don't have a lot of things that need flushing.
 882          */
 883         if (used - global_rsv_size <= SZ_128M)
 884                 return false;
 885
 886         /*
 887          * We have tickets queued, bail so we don't compete with the async
 888          * flushers.
 889          */
 890         if (space_info->reclaim_size)
 891                 return false;
 892
 893         /*
 894          * If we have over half of the free space occupied by reservations or
 895          * pinned then we want to start flushing.
 896          *
 897          * We do not do the traditional thing here, which is to say
 898          *
 899          *   if (used >= ((total_bytes + avail) / 2))
 900          *     return 1;
 901          *
 902          * because this doesn't quite work how we want.  If we had more than 50%
 903          * of the space_info used by bytes_used and we had 0 available we'd just
 904          * constantly run the background flusher.  Instead we want it to kick in
 905          * if our reclaimable space exceeds our clamped free space.
 906          *
 907          * Our clamping range is 2^1 -> 2^8.  Practically speaking that means
 908          * the following:
 909          *
 910          * Amount of RAM        Minimum threshold       Maximum threshold
 911          *
 912          *        256GiB                     1GiB                  128GiB
 913          *        128GiB                   512MiB                   64GiB
 914          *         64GiB                   256MiB                   32GiB
 915          *         32GiB                   128MiB                   16GiB
 916          *         16GiB                    64MiB                    8GiB
 917          *
 918          * These are the range our thresholds will fall in, corresponding to how
 919          * much delalloc we need for the background flusher to kick in.
 920          */
 921
 922         thresh = calc_available_free_space(fs_info, space_info,
 923                                            BTRFS_RESERVE_FLUSH_ALL);
 924         used = space_info->bytes_used + space_info->bytes_reserved +
 925                space_info->bytes_readonly + global_rsv_size;
 926         if (used < total)
 927                 thresh += total - used;
 928         thresh >>= space_info->clamp;
 929
 930         used = space_info->bytes_pinned;
 931
 932         /*
 933          * If we have more ordered bytes than delalloc bytes then we're either
 934          * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
 935          * around.  Preemptive flushing is only useful in that it can free up
 936          * space before tickets need to wait for things to finish.  In the case
 937          * of ordered extents, preemptively waiting on ordered extents gets us
 938          * nothing, if our reservations are tied up in ordered extents we'll
 939          * simply have to slow down writers by forcing them to wait on ordered
 940          * extents.
 941          *
 942          * In the case that ordered is larger than delalloc, only include the
 943          * block reserves that we would actually be able to directly reclaim
 944          * from.  In this case if we're heavy on metadata operations this will
 945          * clearly be heavy enough to warrant preemptive flushing.  In the case
 946          * of heavy DIO or ordered reservations, preemptive flushing will just
 947          * waste time and cause us to slow down.
 948          *
 949          * We want to make sure we truly are maxed out on ordered however, so
 950          * cut ordered in half, and if it's still higher than delalloc then we
 951          * can keep flushing.  This is to avoid the case where we start
 952          * flushing, and now delalloc == ordered and we stop preemptively
 953          * flushing when we could still have several gigs of delalloc to flush.
 954          */
 955         ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
 956         delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
 957         if (ordered >= delalloc)
 958                 used += fs_info->delayed_refs_rsv.reserved +
 959                         fs_info->delayed_block_rsv.reserved;
 960         else
 961                 used += space_info->bytes_may_use - global_rsv_size;
 962
 963         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
 964                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
 965 }
 966
 967 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
 968                                   struct btrfs_space_info *space_info,
 969                                   struct reserve_ticket *ticket)
 970 {
 971         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
 972         u64 min_bytes;
 973
 974         if (!ticket->steal)
 975                 return false;
 976
 977         if (global_rsv->space_info != space_info)
 978                 return false;
 979
 980         spin_lock(&global_rsv->lock);
 981         min_bytes = mult_perc(global_rsv->size, 10);
 982         if (global_rsv->reserved < min_bytes + ticket->bytes) {
 983                 spin_unlock(&global_rsv->lock);
 984                 return false;
 985         }
 986         global_rsv->reserved -= ticket->bytes;
 987         remove_ticket(space_info, ticket);
 988         ticket->bytes = 0;
 989         wake_up(&ticket->wait);
 990         space_info->tickets_id++;
 991         if (global_rsv->reserved < global_rsv->size)
 992                 global_rsv->full = 0;
 993         spin_unlock(&global_rsv->lock);
 994
 995         return true;
 996 }
 997
 998 /*
 999  * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
1000  * @fs_info - fs_info for this fs
1001  * @space_info - the space info we were flushing
1002  *
1003  * We call this when we've exhausted our flushing ability and haven't made
1004  * progress in satisfying tickets.  The reservation code handles tickets in
1005  * order, so if there is a large ticket first and then smaller ones we could
1006  * very well satisfy the smaller tickets.  This will attempt to wake up any
1007  * tickets in the list to catch this case.
1008  *
1009  * This function returns true if it was able to make progress by clearing out
1010  * other tickets, or if it stumbles across a ticket that was smaller than the
1011  * first ticket.
1012  */
1013 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
1014                                    struct btrfs_space_info *space_info)
1015 {
1016         struct reserve_ticket *ticket;
1017         u64 tickets_id = space_info->tickets_id;
1018         const bool aborted = BTRFS_FS_ERROR(fs_info);
1019
1020         trace_btrfs_fail_all_tickets(fs_info, space_info);
1021
1022         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1023                 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
1024                 __btrfs_dump_space_info(fs_info, space_info);
1025         }
1026
1027         while (!list_empty(&space_info->tickets) &&
1028                tickets_id == space_info->tickets_id) {
1029                 ticket = list_first_entry(&space_info->tickets,
1030                                           struct reserve_ticket, list);
1031
1032                 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket))
1033                         return true;
1034
1035                 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1036                         btrfs_info(fs_info, "failing ticket with %llu bytes",
1037                                    ticket->bytes);
1038
1039                 remove_ticket(space_info, ticket);
1040                 if (aborted)
1041                         ticket->error = -EIO;
1042                 else
1043                         ticket->error = -ENOSPC;
1044                 wake_up(&ticket->wait);
1045
1046                 /*
1047                  * We're just throwing tickets away, so more flushing may not
1048                  * trip over btrfs_try_granting_tickets, so we need to call it
1049                  * here to see if we can make progress with the next ticket in
1050                  * the list.
1051                  */
1052                 if (!aborted)
1053                         btrfs_try_granting_tickets(fs_info, space_info);
1054         }
1055         return (tickets_id != space_info->tickets_id);
1056 }
1057
1058 /*
1059  * This is for normal flushers, we can wait all goddamned day if we want to.  We
1060  * will loop and continuously try to flush as long as we are making progress.
1061  * We count progress as clearing off tickets each time we have to loop.
1062  */
1063 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
1064 {
1065         struct btrfs_fs_info *fs_info;
1066         struct btrfs_space_info *space_info;
1067         u64 to_reclaim;
1068         enum btrfs_flush_state flush_state;
1069         int commit_cycles = 0;
1070         u64 last_tickets_id;
1071
1072         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
1073         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1074
1075         spin_lock(&space_info->lock);
1076         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1077         if (!to_reclaim) {
1078                 space_info->flush = 0;
1079                 spin_unlock(&space_info->lock);
1080                 return;
1081         }
1082         last_tickets_id = space_info->tickets_id;
1083         spin_unlock(&space_info->lock);
1084
1085         flush_state = FLUSH_DELAYED_ITEMS_NR;
1086         do {
1087                 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
1088                 spin_lock(&space_info->lock);
1089                 if (list_empty(&space_info->tickets)) {
1090                         space_info->flush = 0;
1091                         spin_unlock(&space_info->lock);
1092                         return;
1093                 }
1094                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1095                                                               space_info);
1096                 if (last_tickets_id == space_info->tickets_id) {
1097                         flush_state++;
1098                 } else {
1099                         last_tickets_id = space_info->tickets_id;
1100                         flush_state = FLUSH_DELAYED_ITEMS_NR;
1101                         if (commit_cycles)
1102                                 commit_cycles--;
1103                 }
1104
1105                 /*
1106                  * We do not want to empty the system of delalloc unless we're
1107                  * under heavy pressure, so allow one trip through the flushing
1108                  * logic before we start doing a FLUSH_DELALLOC_FULL.
1109                  */
1110                 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles)
1111                         flush_state++;
1112
1113                 /*
1114                  * We don't want to force a chunk allocation until we've tried
1115                  * pretty hard to reclaim space.  Think of the case where we
1116                  * freed up a bunch of space and so have a lot of pinned space
1117                  * to reclaim.  We would rather use that than possibly create a
1118                  * underutilized metadata chunk.  So if this is our first run
1119                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
1120                  * commit the transaction.  If nothing has changed the next go
1121                  * around then we can force a chunk allocation.
1122                  */
1123                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1124                         flush_state++;
1125
1126                 if (flush_state > COMMIT_TRANS) {
1127                         commit_cycles++;
1128                         if (commit_cycles > 2) {
1129                                 if (maybe_fail_all_tickets(fs_info, space_info)) {
1130                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
1131                                         commit_cycles--;
1132                                 } else {
1133                                         space_info->flush = 0;
1134                                 }
1135                         } else {
1136                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
1137                         }
1138                 }
1139                 spin_unlock(&space_info->lock);
1140         } while (flush_state <= COMMIT_TRANS);
1141 }
1142
1143 /*
1144  * This handles pre-flushing of metadata space before we get to the point that
1145  * we need to start blocking threads on tickets.  The logic here is different
1146  * from the other flush paths because it doesn't rely on tickets to tell us how
1147  * much we need to flush, instead it attempts to keep us below the 80% full
1148  * watermark of space by flushing whichever reservation pool is currently the
1149  * largest.
1150  */
1151 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1152 {
1153         struct btrfs_fs_info *fs_info;
1154         struct btrfs_space_info *space_info;
1155         struct btrfs_block_rsv *delayed_block_rsv;
1156         struct btrfs_block_rsv *delayed_refs_rsv;
1157         struct btrfs_block_rsv *global_rsv;
1158         struct btrfs_block_rsv *trans_rsv;
1159         int loops = 0;
1160
1161         fs_info = container_of(work, struct btrfs_fs_info,
1162                                preempt_reclaim_work);
1163         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1164         delayed_block_rsv = &fs_info->delayed_block_rsv;
1165         delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1166         global_rsv = &fs_info->global_block_rsv;
1167         trans_rsv = &fs_info->trans_block_rsv;
1168
1169         spin_lock(&space_info->lock);
1170         while (need_preemptive_reclaim(fs_info, space_info)) {
1171                 enum btrfs_flush_state flush;
1172                 u64 delalloc_size = 0;
1173                 u64 to_reclaim, block_rsv_size;
1174                 u64 global_rsv_size = global_rsv->reserved;
1175
1176                 loops++;
1177
1178                 /*
1179                  * We don't have a precise counter for the metadata being
1180                  * reserved for delalloc, so we'll approximate it by subtracting
1181                  * out the block rsv's space from the bytes_may_use.  If that
1182                  * amount is higher than the individual reserves, then we can
1183                  * assume it's tied up in delalloc reservations.
1184                  */
1185                 block_rsv_size = global_rsv_size +
1186                         delayed_block_rsv->reserved +
1187                         delayed_refs_rsv->reserved +
1188                         trans_rsv->reserved;
1189                 if (block_rsv_size < space_info->bytes_may_use)
1190                         delalloc_size = space_info->bytes_may_use - block_rsv_size;
1191
1192                 /*
1193                  * We don't want to include the global_rsv in our calculation,
1194                  * because that's space we can't touch.  Subtract it from the
1195                  * block_rsv_size for the next checks.
1196                  */
1197                 block_rsv_size -= global_rsv_size;
1198
1199                 /*
1200                  * We really want to avoid flushing delalloc too much, as it
1201                  * could result in poor allocation patterns, so only flush it if
1202                  * it's larger than the rest of the pools combined.
1203                  */
1204                 if (delalloc_size > block_rsv_size) {
1205                         to_reclaim = delalloc_size;
1206                         flush = FLUSH_DELALLOC;
1207                 } else if (space_info->bytes_pinned >
1208                            (delayed_block_rsv->reserved +
1209                             delayed_refs_rsv->reserved)) {
1210                         to_reclaim = space_info->bytes_pinned;
1211                         flush = COMMIT_TRANS;
1212                 } else if (delayed_block_rsv->reserved >
1213                            delayed_refs_rsv->reserved) {
1214                         to_reclaim = delayed_block_rsv->reserved;
1215                         flush = FLUSH_DELAYED_ITEMS_NR;
1216                 } else {
1217                         to_reclaim = delayed_refs_rsv->reserved;
1218                         flush = FLUSH_DELAYED_REFS_NR;
1219                 }
1220
1221                 spin_unlock(&space_info->lock);
1222
1223                 /*
1224                  * We don't want to reclaim everything, just a portion, so scale
1225                  * down the to_reclaim by 1/4.  If it takes us down to 0,
1226                  * reclaim 1 items worth.
1227                  */
1228                 to_reclaim >>= 2;
1229                 if (!to_reclaim)
1230                         to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1231                 flush_space(fs_info, space_info, to_reclaim, flush, true);
1232                 cond_resched();
1233                 spin_lock(&space_info->lock);
1234         }
1235
1236         /* We only went through once, back off our clamping. */
1237         if (loops == 1 && !space_info->reclaim_size)
1238                 space_info->clamp = max(1, space_info->clamp - 1);
1239         trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1240         spin_unlock(&space_info->lock);
1241 }
1242
1243 /*
1244  * FLUSH_DELALLOC_WAIT:
1245  *   Space is freed from flushing delalloc in one of two ways.
1246  *
1247  *   1) compression is on and we allocate less space than we reserved
1248  *   2) we are overwriting existing space
1249  *
1250  *   For #1 that extra space is reclaimed as soon as the delalloc pages are
1251  *   COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1252  *   length to ->bytes_reserved, and subtracts the reserved space from
1253  *   ->bytes_may_use.
1254  *
1255  *   For #2 this is trickier.  Once the ordered extent runs we will drop the
1256  *   extent in the range we are overwriting, which creates a delayed ref for
1257  *   that freed extent.  This however is not reclaimed until the transaction
1258  *   commits, thus the next stages.
1259  *
1260  * RUN_DELAYED_IPUTS
1261  *   If we are freeing inodes, we want to make sure all delayed iputs have
1262  *   completed, because they could have been on an inode with i_nlink == 0, and
1263  *   thus have been truncated and freed up space.  But again this space is not
1264  *   immediately re-usable, it comes in the form of a delayed ref, which must be
1265  *   run and then the transaction must be committed.
1266  *
1267  * COMMIT_TRANS
1268  *   This is where we reclaim all of the pinned space generated by running the
1269  *   iputs
1270  *
1271  * ALLOC_CHUNK_FORCE
1272  *   For data we start with alloc chunk force, however we could have been full
1273  *   before, and then the transaction commit could have freed new block groups,
1274  *   so if we now have space to allocate do the force chunk allocation.
1275  */
1276 static const enum btrfs_flush_state data_flush_states[] = {
1277         FLUSH_DELALLOC_FULL,
1278         RUN_DELAYED_IPUTS,
1279         COMMIT_TRANS,
1280         ALLOC_CHUNK_FORCE,
1281 };
1282
1283 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1284 {
1285         struct btrfs_fs_info *fs_info;
1286         struct btrfs_space_info *space_info;
1287         u64 last_tickets_id;
1288         enum btrfs_flush_state flush_state = 0;
1289
1290         fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1291         space_info = fs_info->data_sinfo;
1292
1293         spin_lock(&space_info->lock);
1294         if (list_empty(&space_info->tickets)) {
1295                 space_info->flush = 0;
1296                 spin_unlock(&space_info->lock);
1297                 return;
1298         }
1299         last_tickets_id = space_info->tickets_id;
1300         spin_unlock(&space_info->lock);
1301
1302         while (!space_info->full) {
1303                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1304                 spin_lock(&space_info->lock);
1305                 if (list_empty(&space_info->tickets)) {
1306                         space_info->flush = 0;
1307                         spin_unlock(&space_info->lock);
1308                         return;
1309                 }
1310
1311                 /* Something happened, fail everything and bail. */
1312                 if (BTRFS_FS_ERROR(fs_info))
1313                         goto aborted_fs;
1314                 last_tickets_id = space_info->tickets_id;
1315                 spin_unlock(&space_info->lock);
1316         }
1317
1318         while (flush_state < ARRAY_SIZE(data_flush_states)) {
1319                 flush_space(fs_info, space_info, U64_MAX,
1320                             data_flush_states[flush_state], false);
1321                 spin_lock(&space_info->lock);
1322                 if (list_empty(&space_info->tickets)) {
1323                         space_info->flush = 0;
1324                         spin_unlock(&space_info->lock);
1325                         return;
1326                 }
1327
1328                 if (last_tickets_id == space_info->tickets_id) {
1329                         flush_state++;
1330                 } else {
1331                         last_tickets_id = space_info->tickets_id;
1332                         flush_state = 0;
1333                 }
1334
1335                 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1336                         if (space_info->full) {
1337                                 if (maybe_fail_all_tickets(fs_info, space_info))
1338                                         flush_state = 0;
1339                                 else
1340                                         space_info->flush = 0;
1341                         } else {
1342                                 flush_state = 0;
1343                         }
1344
1345                         /* Something happened, fail everything and bail. */
1346                         if (BTRFS_FS_ERROR(fs_info))
1347                                 goto aborted_fs;
1348
1349                 }
1350                 spin_unlock(&space_info->lock);
1351         }
1352         return;
1353
1354 aborted_fs:
1355         maybe_fail_all_tickets(fs_info, space_info);
1356         space_info->flush = 0;
1357         spin_unlock(&space_info->lock);
1358 }
1359
1360 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1361 {
1362         INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1363         INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1364         INIT_WORK(&fs_info->preempt_reclaim_work,
1365                   btrfs_preempt_reclaim_metadata_space);
1366 }
1367
1368 static const enum btrfs_flush_state priority_flush_states[] = {
1369         FLUSH_DELAYED_ITEMS_NR,
1370         FLUSH_DELAYED_ITEMS,
1371         ALLOC_CHUNK,
1372 };
1373
1374 static const enum btrfs_flush_state evict_flush_states[] = {
1375         FLUSH_DELAYED_ITEMS_NR,
1376         FLUSH_DELAYED_ITEMS,
1377         FLUSH_DELAYED_REFS_NR,
1378         FLUSH_DELAYED_REFS,
1379         FLUSH_DELALLOC,
1380         FLUSH_DELALLOC_WAIT,
1381         FLUSH_DELALLOC_FULL,
1382         ALLOC_CHUNK,
1383         COMMIT_TRANS,
1384 };
1385
1386 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1387                                 struct btrfs_space_info *space_info,
1388                                 struct reserve_ticket *ticket,
1389                                 const enum btrfs_flush_state *states,
1390                                 int states_nr)
1391 {
1392         u64 to_reclaim;
1393         int flush_state = 0;
1394
1395         spin_lock(&space_info->lock);
1396         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1397         /*
1398          * This is the priority reclaim path, so to_reclaim could be >0 still
1399          * because we may have only satisfied the priority tickets and still
1400          * left non priority tickets on the list.  We would then have
1401          * to_reclaim but ->bytes == 0.
1402          */
1403         if (ticket->bytes == 0) {
1404                 spin_unlock(&space_info->lock);
1405                 return;
1406         }
1407
1408         while (flush_state < states_nr) {
1409                 spin_unlock(&space_info->lock);
1410                 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1411                             false);
1412                 flush_state++;
1413                 spin_lock(&space_info->lock);
1414                 if (ticket->bytes == 0) {
1415                         spin_unlock(&space_info->lock);
1416                         return;
1417                 }
1418         }
1419
1420         /* Attempt to steal from the global rsv if we can. */
1421         if (!steal_from_global_rsv(fs_info, space_info, ticket)) {
1422                 ticket->error = -ENOSPC;
1423                 remove_ticket(space_info, ticket);
1424         }
1425
1426         /*
1427          * We must run try_granting_tickets here because we could be a large
1428          * ticket in front of a smaller ticket that can now be satisfied with
1429          * the available space.
1430          */
1431         btrfs_try_granting_tickets(fs_info, space_info);
1432         spin_unlock(&space_info->lock);
1433 }
1434
1435 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1436                                         struct btrfs_space_info *space_info,
1437                                         struct reserve_ticket *ticket)
1438 {
1439         spin_lock(&space_info->lock);
1440
1441         /* We could have been granted before we got here. */
1442         if (ticket->bytes == 0) {
1443                 spin_unlock(&space_info->lock);
1444                 return;
1445         }
1446
1447         while (!space_info->full) {
1448                 spin_unlock(&space_info->lock);
1449                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1450                 spin_lock(&space_info->lock);
1451                 if (ticket->bytes == 0) {
1452                         spin_unlock(&space_info->lock);
1453                         return;
1454                 }
1455         }
1456
1457         ticket->error = -ENOSPC;
1458         remove_ticket(space_info, ticket);
1459         btrfs_try_granting_tickets(fs_info, space_info);
1460         spin_unlock(&space_info->lock);
1461 }
1462
1463 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1464                                 struct btrfs_space_info *space_info,
1465                                 struct reserve_ticket *ticket)
1466
1467 {
1468         DEFINE_WAIT(wait);
1469         int ret = 0;
1470
1471         spin_lock(&space_info->lock);
1472         while (ticket->bytes > 0 && ticket->error == 0) {
1473                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1474                 if (ret) {
1475                         /*
1476                          * Delete us from the list. After we unlock the space
1477                          * info, we don't want the async reclaim job to reserve
1478                          * space for this ticket. If that would happen, then the
1479                          * ticket's task would not known that space was reserved
1480                          * despite getting an error, resulting in a space leak
1481                          * (bytes_may_use counter of our space_info).
1482                          */
1483                         remove_ticket(space_info, ticket);
1484                         ticket->error = -EINTR;
1485                         break;
1486                 }
1487                 spin_unlock(&space_info->lock);
1488
1489                 schedule();
1490
1491                 finish_wait(&ticket->wait, &wait);
1492                 spin_lock(&space_info->lock);
1493         }
1494         spin_unlock(&space_info->lock);
1495 }
1496
1497 /*
1498  * Do the appropriate flushing and waiting for a ticket.
1499  *
1500  * @fs_info:    the filesystem
1501  * @space_info: space info for the reservation
1502  * @ticket:     ticket for the reservation
1503  * @start_ns:   timestamp when the reservation started
1504  * @orig_bytes: amount of bytes originally reserved
1505  * @flush:      how much we can flush
1506  *
1507  * This does the work of figuring out how to flush for the ticket, waiting for
1508  * the reservation, and returning the appropriate error if there is one.
1509  */
1510 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1511                                  struct btrfs_space_info *space_info,
1512                                  struct reserve_ticket *ticket,
1513                                  u64 start_ns, u64 orig_bytes,
1514                                  enum btrfs_reserve_flush_enum flush)
1515 {
1516         int ret;
1517
1518         switch (flush) {
1519         case BTRFS_RESERVE_FLUSH_DATA:
1520         case BTRFS_RESERVE_FLUSH_ALL:
1521         case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1522                 wait_reserve_ticket(fs_info, space_info, ticket);
1523                 break;
1524         case BTRFS_RESERVE_FLUSH_LIMIT:
1525                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1526                                                 priority_flush_states,
1527                                                 ARRAY_SIZE(priority_flush_states));
1528                 break;
1529         case BTRFS_RESERVE_FLUSH_EVICT:
1530                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1531                                                 evict_flush_states,
1532                                                 ARRAY_SIZE(evict_flush_states));
1533                 break;
1534         case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1535                 priority_reclaim_data_space(fs_info, space_info, ticket);
1536                 break;
1537         default:
1538                 ASSERT(0);
1539                 break;
1540         }
1541
1542         ret = ticket->error;
1543         ASSERT(list_empty(&ticket->list));
1544         /*
1545          * Check that we can't have an error set if the reservation succeeded,
1546          * as that would confuse tasks and lead them to error out without
1547          * releasing reserved space (if an error happens the expectation is that
1548          * space wasn't reserved at all).
1549          */
1550         ASSERT(!(ticket->bytes == 0 && ticket->error));
1551         trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1552                                    start_ns, flush, ticket->error);
1553         return ret;
1554 }
1555
1556 /*
1557  * This returns true if this flush state will go through the ordinary flushing
1558  * code.
1559  */
1560 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1561 {
1562         return  (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1563                 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1564 }
1565
1566 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1567                                        struct btrfs_space_info *space_info)
1568 {
1569         u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1570         u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1571
1572         /*
1573          * If we're heavy on ordered operations then clamping won't help us.  We
1574          * need to clamp specifically to keep up with dirty'ing buffered
1575          * writers, because there's not a 1:1 correlation of writing delalloc
1576          * and freeing space, like there is with flushing delayed refs or
1577          * delayed nodes.  If we're already more ordered than delalloc then
1578          * we're keeping up, otherwise we aren't and should probably clamp.
1579          */
1580         if (ordered < delalloc)
1581                 space_info->clamp = min(space_info->clamp + 1, 8);
1582 }
1583
1584 static inline bool can_steal(enum btrfs_reserve_flush_enum flush)
1585 {
1586         return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1587                 flush == BTRFS_RESERVE_FLUSH_EVICT);
1588 }
1589
1590 /*
1591  * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to
1592  * fail as quickly as possible.
1593  */
1594 static inline bool can_ticket(enum btrfs_reserve_flush_enum flush)
1595 {
1596         return (flush != BTRFS_RESERVE_NO_FLUSH &&
1597                 flush != BTRFS_RESERVE_FLUSH_EMERGENCY);
1598 }
1599
1600 /*
1601  * Try to reserve bytes from the block_rsv's space.
1602  *
1603  * @fs_info:    the filesystem
1604  * @space_info: space info we want to allocate from
1605  * @orig_bytes: number of bytes we want
1606  * @flush:      whether or not we can flush to make our reservation
1607  *
1608  * This will reserve orig_bytes number of bytes from the space info associated
1609  * with the block_rsv.  If there is not enough space it will make an attempt to
1610  * flush out space to make room.  It will do this by flushing delalloc if
1611  * possible or committing the transaction.  If flush is 0 then no attempts to
1612  * regain reservations will be made and this will fail if there is not enough
1613  * space already.
1614  */
1615 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1616                            struct btrfs_space_info *space_info, u64 orig_bytes,
1617                            enum btrfs_reserve_flush_enum flush)
1618 {
1619         struct work_struct *async_work;
1620         struct reserve_ticket ticket;
1621         u64 start_ns = 0;
1622         u64 used;
1623         int ret = 0;
1624         bool pending_tickets;
1625
1626         ASSERT(orig_bytes);
1627         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1628
1629         if (flush == BTRFS_RESERVE_FLUSH_DATA)
1630                 async_work = &fs_info->async_data_reclaim_work;
1631         else
1632                 async_work = &fs_info->async_reclaim_work;
1633
1634         spin_lock(&space_info->lock);
1635         ret = -ENOSPC;
1636         used = btrfs_space_info_used(space_info, true);
1637
1638         /*
1639          * We don't want NO_FLUSH allocations to jump everybody, they can
1640          * generally handle ENOSPC in a different way, so treat them the same as
1641          * normal flushers when it comes to skipping pending tickets.
1642          */
1643         if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1644                 pending_tickets = !list_empty(&space_info->tickets) ||
1645                         !list_empty(&space_info->priority_tickets);
1646         else
1647                 pending_tickets = !list_empty(&space_info->priority_tickets);
1648
1649         /*
1650          * Carry on if we have enough space (short-circuit) OR call
1651          * can_overcommit() to ensure we can overcommit to continue.
1652          */
1653         if (!pending_tickets &&
1654             ((used + orig_bytes <= writable_total_bytes(fs_info, space_info)) ||
1655              btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1656                 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1657                                                       orig_bytes);
1658                 ret = 0;
1659         }
1660
1661         /*
1662          * Things are dire, we need to make a reservation so we don't abort.  We
1663          * will let this reservation go through as long as we have actual space
1664          * left to allocate for the block.
1665          */
1666         if (ret && unlikely(flush == BTRFS_RESERVE_FLUSH_EMERGENCY)) {
1667                 used = btrfs_space_info_used(space_info, false);
1668                 if (used + orig_bytes <=
1669                     writable_total_bytes(fs_info, space_info)) {
1670                         btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1671                                                               orig_bytes);
1672                         ret = 0;
1673                 }
1674         }
1675
1676         /*
1677          * If we couldn't make a reservation then setup our reservation ticket
1678          * and kick the async worker if it's not already running.
1679          *
1680          * If we are a priority flusher then we just need to add our ticket to
1681          * the list and we will do our own flushing further down.
1682          */
1683         if (ret && can_ticket(flush)) {
1684                 ticket.bytes = orig_bytes;
1685                 ticket.error = 0;
1686                 space_info->reclaim_size += ticket.bytes;
1687                 init_waitqueue_head(&ticket.wait);
1688                 ticket.steal = can_steal(flush);
1689                 if (trace_btrfs_reserve_ticket_enabled())
1690                         start_ns = ktime_get_ns();
1691
1692                 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1693                     flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1694                     flush == BTRFS_RESERVE_FLUSH_DATA) {
1695                         list_add_tail(&ticket.list, &space_info->tickets);
1696                         if (!space_info->flush) {
1697                                 /*
1698                                  * We were forced to add a reserve ticket, so
1699                                  * our preemptive flushing is unable to keep
1700                                  * up.  Clamp down on the threshold for the
1701                                  * preemptive flushing in order to keep up with
1702                                  * the workload.
1703                                  */
1704                                 maybe_clamp_preempt(fs_info, space_info);
1705
1706                                 space_info->flush = 1;
1707                                 trace_btrfs_trigger_flush(fs_info,
1708                                                           space_info->flags,
1709                                                           orig_bytes, flush,
1710                                                           "enospc");
1711                                 queue_work(system_unbound_wq, async_work);
1712                         }
1713                 } else {
1714                         list_add_tail(&ticket.list,
1715                                       &space_info->priority_tickets);
1716                 }
1717         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1718                 /*
1719                  * We will do the space reservation dance during log replay,
1720                  * which means we won't have fs_info->fs_root set, so don't do
1721                  * the async reclaim as we will panic.
1722                  */
1723                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1724                     !work_busy(&fs_info->preempt_reclaim_work) &&
1725                     need_preemptive_reclaim(fs_info, space_info)) {
1726                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
1727                                                   orig_bytes, flush, "preempt");
1728                         queue_work(system_unbound_wq,
1729                                    &fs_info->preempt_reclaim_work);
1730                 }
1731         }
1732         spin_unlock(&space_info->lock);
1733         if (!ret || !can_ticket(flush))
1734                 return ret;
1735
1736         return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1737                                      orig_bytes, flush);
1738 }
1739
1740 /*
1741  * Try to reserve metadata bytes from the block_rsv's space.
1742  *
1743  * @fs_info:    the filesystem
1744  * @block_rsv:  block_rsv we're allocating for
1745  * @orig_bytes: number of bytes we want
1746  * @flush:      whether or not we can flush to make our reservation
1747  *
1748  * This will reserve orig_bytes number of bytes from the space info associated
1749  * with the block_rsv.  If there is not enough space it will make an attempt to
1750  * flush out space to make room.  It will do this by flushing delalloc if
1751  * possible or committing the transaction.  If flush is 0 then no attempts to
1752  * regain reservations will be made and this will fail if there is not enough
1753  * space already.
1754  */
1755 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1756                                  struct btrfs_block_rsv *block_rsv,
1757                                  u64 orig_bytes,
1758                                  enum btrfs_reserve_flush_enum flush)
1759 {
1760         int ret;
1761
1762         ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush);
1763         if (ret == -ENOSPC) {
1764                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1765                                               block_rsv->space_info->flags,
1766                                               orig_bytes, 1);
1767
1768                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1769                         btrfs_dump_space_info(fs_info, block_rsv->space_info,
1770                                               orig_bytes, 0);
1771         }
1772         return ret;
1773 }
1774
1775 /*
1776  * Try to reserve data bytes for an allocation.
1777  *
1778  * @fs_info: the filesystem
1779  * @bytes:   number of bytes we need
1780  * @flush:   how we are allowed to flush
1781  *
1782  * This will reserve bytes from the data space info.  If there is not enough
1783  * space then we will attempt to flush space as specified by flush.
1784  */
1785 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1786                              enum btrfs_reserve_flush_enum flush)
1787 {
1788         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1789         int ret;
1790
1791         ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1792                flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE ||
1793                flush == BTRFS_RESERVE_NO_FLUSH);
1794         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1795
1796         ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1797         if (ret == -ENOSPC) {
1798                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1799                                               data_sinfo->flags, bytes, 1);
1800                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1801                         btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1802         }
1803         return ret;
1804 }
1805
1806 /* Dump all the space infos when we abort a transaction due to ENOSPC. */
1807 __cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info)
1808 {
1809         struct btrfs_space_info *space_info;
1810
1811         btrfs_info(fs_info, "dumping space info:");
1812         list_for_each_entry(space_info, &fs_info->space_info, list) {
1813                 spin_lock(&space_info->lock);
1814                 __btrfs_dump_space_info(fs_info, space_info);
1815                 spin_unlock(&space_info->lock);
1816         }
1817         dump_global_block_rsv(fs_info);
1818 }
1819
1820 /*
1821  * Account the unused space of all the readonly block group in the space_info.
1822  * takes mirrors into account.
1823  */
1824 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
1825 {
1826         struct btrfs_block_group *block_group;
1827         u64 free_bytes = 0;
1828         int factor;
1829
1830         /* It's df, we don't care if it's racy */
1831         if (list_empty(&sinfo->ro_bgs))
1832                 return 0;
1833
1834         spin_lock(&sinfo->lock);
1835         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
1836                 spin_lock(&block_group->lock);
1837
1838                 if (!block_group->ro) {
1839                         spin_unlock(&block_group->lock);
1840                         continue;
1841                 }
1842
1843                 factor = btrfs_bg_type_to_factor(block_group->flags);
1844                 free_bytes += (block_group->length -
1845                                block_group->used) * factor;
1846
1847                 spin_unlock(&block_group->lock);
1848         }
1849         spin_unlock(&sinfo->lock);
1850
1851         return free_bytes;
1852 }