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  *   FORCE_COMMIT_TRANS
 144  *     For use by the preemptive flusher.  We use this to bypass the ticketing
 145  *     checks in may_commit_transaction, as we have more information about the
 146  *     overall state of the system and may want to commit the transaction ahead
 147  *     of actual ENOSPC conditions.
 148  *
 149  * OVERCOMMIT
 150  *
 151  *   Because we hold so many reservations for metadata we will allow you to
 152  *   reserve more space than is currently free in the currently allocate
 153  *   metadata space.  This only happens with metadata, data does not allow
 154  *   overcommitting.
 155  *
 156  *   You can see the current logic for when we allow overcommit in
 157  *   btrfs_can_overcommit(), but it only applies to unallocated space.  If there
 158  *   is no unallocated space to be had, all reservations are kept within the
 159  *   free space in the allocated metadata chunks.
 160  *
 161  *   Because of overcommitting, you generally want to use the
 162  *   btrfs_can_overcommit() logic for metadata allocations, as it does the right
 163  *   thing with or without extra unallocated space.
 164  */
 165
 166 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
 167                           bool may_use_included)
 168 {
 169         ASSERT(s_info);
 170         return s_info->bytes_used + s_info->bytes_reserved +
 171                 s_info->bytes_pinned + s_info->bytes_readonly +
 172                 s_info->bytes_zone_unusable +
 173                 (may_use_included ? s_info->bytes_may_use : 0);
 174 }
 175
 176 /*
 177  * after adding space to the filesystem, we need to clear the full flags
 178  * on all the space infos.
 179  */
 180 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 181 {
 182         struct list_head *head = &info->space_info;
 183         struct btrfs_space_info *found;
 184
 185         list_for_each_entry(found, head, list)
 186                 found->full = 0;
 187 }
 188
 189 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
 190 {
 191
 192         struct btrfs_space_info *space_info;
 193         int i;
 194         int ret;
 195
 196         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
 197         if (!space_info)
 198                 return -ENOMEM;
 199
 200         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
 201                                  GFP_KERNEL);
 202         if (ret) {
 203                 kfree(space_info);
 204                 return ret;
 205         }
 206
 207         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
 208                 INIT_LIST_HEAD(&space_info->block_groups[i]);
 209         init_rwsem(&space_info->groups_sem);
 210         spin_lock_init(&space_info->lock);
 211         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
 212         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
 213         INIT_LIST_HEAD(&space_info->ro_bgs);
 214         INIT_LIST_HEAD(&space_info->tickets);
 215         INIT_LIST_HEAD(&space_info->priority_tickets);
 216         space_info->clamp = 1;
 217
 218         ret = btrfs_sysfs_add_space_info_type(info, space_info);
 219         if (ret)
 220                 return ret;
 221
 222         list_add(&space_info->list, &info->space_info);
 223         if (flags & BTRFS_BLOCK_GROUP_DATA)
 224                 info->data_sinfo = space_info;
 225
 226         return ret;
 227 }
 228
 229 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
 230 {
 231         struct btrfs_super_block *disk_super;
 232         u64 features;
 233         u64 flags;
 234         int mixed = 0;
 235         int ret;
 236
 237         disk_super = fs_info->super_copy;
 238         if (!btrfs_super_root(disk_super))
 239                 return -EINVAL;
 240
 241         features = btrfs_super_incompat_flags(disk_super);
 242         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
 243                 mixed = 1;
 244
 245         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 246         ret = create_space_info(fs_info, flags);
 247         if (ret)
 248                 goto out;
 249
 250         if (mixed) {
 251                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
 252                 ret = create_space_info(fs_info, flags);
 253         } else {
 254                 flags = BTRFS_BLOCK_GROUP_METADATA;
 255                 ret = create_space_info(fs_info, flags);
 256                 if (ret)
 257                         goto out;
 258
 259                 flags = BTRFS_BLOCK_GROUP_DATA;
 260                 ret = create_space_info(fs_info, flags);
 261         }
 262 out:
 263         return ret;
 264 }
 265
 266 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags,
 267                              u64 total_bytes, u64 bytes_used,
 268                              u64 bytes_readonly, u64 bytes_zone_unusable,
 269                              struct btrfs_space_info **space_info)
 270 {
 271         struct btrfs_space_info *found;
 272         int factor;
 273
 274         factor = btrfs_bg_type_to_factor(flags);
 275
 276         found = btrfs_find_space_info(info, flags);
 277         ASSERT(found);
 278         spin_lock(&found->lock);
 279         found->total_bytes += total_bytes;
 280         found->disk_total += total_bytes * factor;
 281         found->bytes_used += bytes_used;
 282         found->disk_used += bytes_used * factor;
 283         found->bytes_readonly += bytes_readonly;
 284         found->bytes_zone_unusable += bytes_zone_unusable;
 285         if (total_bytes > 0)
 286                 found->full = 0;
 287         btrfs_try_granting_tickets(info, found);
 288         spin_unlock(&found->lock);
 289         *space_info = found;
 290 }
 291
 292 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
 293                                                u64 flags)
 294 {
 295         struct list_head *head = &info->space_info;
 296         struct btrfs_space_info *found;
 297
 298         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 299
 300         list_for_each_entry(found, head, list) {
 301                 if (found->flags & flags)
 302                         return found;
 303         }
 304         return NULL;
 305 }
 306
 307 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
 308                           struct btrfs_space_info *space_info,
 309                           enum btrfs_reserve_flush_enum flush)
 310 {
 311         u64 profile;
 312         u64 avail;
 313         int factor;
 314
 315         if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
 316                 profile = btrfs_system_alloc_profile(fs_info);
 317         else
 318                 profile = btrfs_metadata_alloc_profile(fs_info);
 319
 320         avail = atomic64_read(&fs_info->free_chunk_space);
 321
 322         /*
 323          * If we have dup, raid1 or raid10 then only half of the free
 324          * space is actually usable.  For raid56, the space info used
 325          * doesn't include the parity drive, so we don't have to
 326          * change the math
 327          */
 328         factor = btrfs_bg_type_to_factor(profile);
 329         avail = div_u64(avail, factor);
 330
 331         /*
 332          * If we aren't flushing all things, let us overcommit up to
 333          * 1/2th of the space. If we can flush, don't let us overcommit
 334          * too much, let it overcommit up to 1/8 of the space.
 335          */
 336         if (flush == BTRFS_RESERVE_FLUSH_ALL)
 337                 avail >>= 3;
 338         else
 339                 avail >>= 1;
 340         return avail;
 341 }
 342
 343 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
 344                          struct btrfs_space_info *space_info, u64 bytes,
 345                          enum btrfs_reserve_flush_enum flush)
 346 {
 347         u64 avail;
 348         u64 used;
 349
 350         /* Don't overcommit when in mixed mode */
 351         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
 352                 return 0;
 353
 354         used = btrfs_space_info_used(space_info, true);
 355         avail = calc_available_free_space(fs_info, space_info, flush);
 356
 357         if (used + bytes < space_info->total_bytes + avail)
 358                 return 1;
 359         return 0;
 360 }
 361
 362 static void remove_ticket(struct btrfs_space_info *space_info,
 363                           struct reserve_ticket *ticket)
 364 {
 365         if (!list_empty(&ticket->list)) {
 366                 list_del_init(&ticket->list);
 367                 ASSERT(space_info->reclaim_size >= ticket->bytes);
 368                 space_info->reclaim_size -= ticket->bytes;
 369         }
 370 }
 371
 372 /*
 373  * This is for space we already have accounted in space_info->bytes_may_use, so
 374  * basically when we're returning space from block_rsv's.
 375  */
 376 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
 377                                 struct btrfs_space_info *space_info)
 378 {
 379         struct list_head *head;
 380         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
 381
 382         lockdep_assert_held(&space_info->lock);
 383
 384         head = &space_info->priority_tickets;
 385 again:
 386         while (!list_empty(head)) {
 387                 struct reserve_ticket *ticket;
 388                 u64 used = btrfs_space_info_used(space_info, true);
 389
 390                 ticket = list_first_entry(head, struct reserve_ticket, list);
 391
 392                 /* Check and see if our ticket can be satisified now. */
 393                 if ((used + ticket->bytes <= space_info->total_bytes) ||
 394                     btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
 395                                          flush)) {
 396                         btrfs_space_info_update_bytes_may_use(fs_info,
 397                                                               space_info,
 398                                                               ticket->bytes);
 399                         remove_ticket(space_info, ticket);
 400                         ticket->bytes = 0;
 401                         space_info->tickets_id++;
 402                         wake_up(&ticket->wait);
 403                 } else {
 404                         break;
 405                 }
 406         }
 407
 408         if (head == &space_info->priority_tickets) {
 409                 head = &space_info->tickets;
 410                 flush = BTRFS_RESERVE_FLUSH_ALL;
 411                 goto again;
 412         }
 413 }
 414
 415 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
 416 do {                                                                    \
 417         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
 418         spin_lock(&__rsv->lock);                                        \
 419         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
 420                    __rsv->size, __rsv->reserved);                       \
 421         spin_unlock(&__rsv->lock);                                      \
 422 } while (0)
 423
 424 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 425                                     struct btrfs_space_info *info)
 426 {
 427         lockdep_assert_held(&info->lock);
 428
 429         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
 430                    info->flags,
 431                    info->total_bytes - btrfs_space_info_used(info, true),
 432                    info->full ? "" : "not ");
 433         btrfs_info(fs_info,
 434                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
 435                 info->total_bytes, info->bytes_used, info->bytes_pinned,
 436                 info->bytes_reserved, info->bytes_may_use,
 437                 info->bytes_readonly, info->bytes_zone_unusable);
 438
 439         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
 440         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
 441         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
 442         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
 443         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
 444
 445 }
 446
 447 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 448                            struct btrfs_space_info *info, u64 bytes,
 449                            int dump_block_groups)
 450 {
 451         struct btrfs_block_group *cache;
 452         int index = 0;
 453
 454         spin_lock(&info->lock);
 455         __btrfs_dump_space_info(fs_info, info);
 456         spin_unlock(&info->lock);
 457
 458         if (!dump_block_groups)
 459                 return;
 460
 461         down_read(&info->groups_sem);
 462 again:
 463         list_for_each_entry(cache, &info->block_groups[index], list) {
 464                 spin_lock(&cache->lock);
 465                 btrfs_info(fs_info,
 466                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s",
 467                         cache->start, cache->length, cache->used, cache->pinned,
 468                         cache->reserved, cache->zone_unusable,
 469                         cache->ro ? "[readonly]" : "");
 470                 spin_unlock(&cache->lock);
 471                 btrfs_dump_free_space(cache, bytes);
 472         }
 473         if (++index < BTRFS_NR_RAID_TYPES)
 474                 goto again;
 475         up_read(&info->groups_sem);
 476 }
 477
 478 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
 479                                         u64 to_reclaim)
 480 {
 481         u64 bytes;
 482         u64 nr;
 483
 484         bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 485         nr = div64_u64(to_reclaim, bytes);
 486         if (!nr)
 487                 nr = 1;
 488         return nr;
 489 }
 490
 491 #define EXTENT_SIZE_PER_ITEM    SZ_256K
 492
 493 /*
 494  * shrink metadata reservation for delalloc
 495  */
 496 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
 497                             struct btrfs_space_info *space_info,
 498                             u64 to_reclaim, bool wait_ordered,
 499                             bool for_preempt)
 500 {
 501         struct btrfs_trans_handle *trans;
 502         u64 delalloc_bytes;
 503         u64 ordered_bytes;
 504         u64 items;
 505         long time_left;
 506         int loops;
 507
 508         /* Calc the number of the pages we need flush for space reservation */
 509         if (to_reclaim == U64_MAX) {
 510                 items = U64_MAX;
 511         } else {
 512                 /*
 513                  * to_reclaim is set to however much metadata we need to
 514                  * reclaim, but reclaiming that much data doesn't really track
 515                  * exactly, so increase the amount to reclaim by 2x in order to
 516                  * make sure we're flushing enough delalloc to hopefully reclaim
 517                  * some metadata reservations.
 518                  */
 519                 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
 520                 to_reclaim = items * EXTENT_SIZE_PER_ITEM;
 521         }
 522
 523         trans = (struct btrfs_trans_handle *)current->journal_info;
 524
 525         delalloc_bytes = percpu_counter_sum_positive(
 526                                                 &fs_info->delalloc_bytes);
 527         ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
 528         if (delalloc_bytes == 0 && ordered_bytes == 0)
 529                 return;
 530
 531         /*
 532          * If we are doing more ordered than delalloc we need to just wait on
 533          * ordered extents, otherwise we'll waste time trying to flush delalloc
 534          * that likely won't give us the space back we need.
 535          */
 536         if (ordered_bytes > delalloc_bytes && !for_preempt)
 537                 wait_ordered = true;
 538
 539         loops = 0;
 540         while ((delalloc_bytes || ordered_bytes) && loops < 3) {
 541                 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
 542                 long nr_pages = min_t(u64, temp, LONG_MAX);
 543
 544                 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
 545
 546                 loops++;
 547                 if (wait_ordered && !trans) {
 548                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
 549                 } else {
 550                         time_left = schedule_timeout_killable(1);
 551                         if (time_left)
 552                                 break;
 553                 }
 554
 555                 /*
 556                  * If we are for preemption we just want a one-shot of delalloc
 557                  * flushing so we can stop flushing if we decide we don't need
 558                  * to anymore.
 559                  */
 560                 if (for_preempt)
 561                         break;
 562
 563                 spin_lock(&space_info->lock);
 564                 if (list_empty(&space_info->tickets) &&
 565                     list_empty(&space_info->priority_tickets)) {
 566                         spin_unlock(&space_info->lock);
 567                         break;
 568                 }
 569                 spin_unlock(&space_info->lock);
 570
 571                 delalloc_bytes = percpu_counter_sum_positive(
 572                                                 &fs_info->delalloc_bytes);
 573                 ordered_bytes = percpu_counter_sum_positive(
 574                                                 &fs_info->ordered_bytes);
 575         }
 576 }
 577
 578 /**
 579  * Possibly commit the transaction if its ok to
 580  *
 581  * @fs_info:    the filesystem
 582  * @space_info: space_info we are checking for commit, either data or metadata
 583  *
 584  * This will check to make sure that committing the transaction will actually
 585  * get us somewhere and then commit the transaction if it does.  Otherwise it
 586  * will return -ENOSPC.
 587  */
 588 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
 589                                   struct btrfs_space_info *space_info)
 590 {
 591         struct reserve_ticket *ticket = NULL;
 592         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
 593         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 594         struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv;
 595         struct btrfs_trans_handle *trans;
 596         u64 reclaim_bytes = 0;
 597         u64 bytes_needed = 0;
 598         u64 cur_free_bytes = 0;
 599
 600         trans = (struct btrfs_trans_handle *)current->journal_info;
 601         if (trans)
 602                 return -EAGAIN;
 603
 604         spin_lock(&space_info->lock);
 605         cur_free_bytes = btrfs_space_info_used(space_info, true);
 606         if (cur_free_bytes < space_info->total_bytes)
 607                 cur_free_bytes = space_info->total_bytes - cur_free_bytes;
 608         else
 609                 cur_free_bytes = 0;
 610
 611         if (!list_empty(&space_info->priority_tickets))
 612                 ticket = list_first_entry(&space_info->priority_tickets,
 613                                           struct reserve_ticket, list);
 614         else if (!list_empty(&space_info->tickets))
 615                 ticket = list_first_entry(&space_info->tickets,
 616                                           struct reserve_ticket, list);
 617         if (ticket)
 618                 bytes_needed = ticket->bytes;
 619
 620         if (bytes_needed > cur_free_bytes)
 621                 bytes_needed -= cur_free_bytes;
 622         else
 623                 bytes_needed = 0;
 624         spin_unlock(&space_info->lock);
 625
 626         if (!bytes_needed)
 627                 return 0;
 628
 629         trans = btrfs_join_transaction(fs_info->extent_root);
 630         if (IS_ERR(trans))
 631                 return PTR_ERR(trans);
 632
 633         /*
 634          * See if there is enough pinned space to make this reservation, or if
 635          * we have block groups that are going to be freed, allowing us to
 636          * possibly do a chunk allocation the next loop through.
 637          */
 638         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
 639             __percpu_counter_compare(&space_info->total_bytes_pinned,
 640                                      bytes_needed,
 641                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
 642                 goto commit;
 643
 644         /*
 645          * See if there is some space in the delayed insertion reserve for this
 646          * reservation.  If the space_info's don't match (like for DATA or
 647          * SYSTEM) then just go enospc, reclaiming this space won't recover any
 648          * space to satisfy those reservations.
 649          */
 650         if (space_info != delayed_rsv->space_info)
 651                 goto enospc;
 652
 653         spin_lock(&delayed_rsv->lock);
 654         reclaim_bytes += delayed_rsv->reserved;
 655         spin_unlock(&delayed_rsv->lock);
 656
 657         spin_lock(&delayed_refs_rsv->lock);
 658         reclaim_bytes += delayed_refs_rsv->reserved;
 659         spin_unlock(&delayed_refs_rsv->lock);
 660
 661         spin_lock(&trans_rsv->lock);
 662         reclaim_bytes += trans_rsv->reserved;
 663         spin_unlock(&trans_rsv->lock);
 664
 665         if (reclaim_bytes >= bytes_needed)
 666                 goto commit;
 667         bytes_needed -= reclaim_bytes;
 668
 669         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
 670                                    bytes_needed,
 671                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
 672                 goto enospc;
 673
 674 commit:
 675         return btrfs_commit_transaction(trans);
 676 enospc:
 677         btrfs_end_transaction(trans);
 678         return -ENOSPC;
 679 }
 680
 681 /*
 682  * Try to flush some data based on policy set by @state. This is only advisory
 683  * and may fail for various reasons. The caller is supposed to examine the
 684  * state of @space_info to detect the outcome.
 685  */
 686 static void flush_space(struct btrfs_fs_info *fs_info,
 687                        struct btrfs_space_info *space_info, u64 num_bytes,
 688                        enum btrfs_flush_state state, bool for_preempt)
 689 {
 690         struct btrfs_root *root = fs_info->extent_root;
 691         struct btrfs_trans_handle *trans;
 692         int nr;
 693         int ret = 0;
 694
 695         switch (state) {
 696         case FLUSH_DELAYED_ITEMS_NR:
 697         case FLUSH_DELAYED_ITEMS:
 698                 if (state == FLUSH_DELAYED_ITEMS_NR)
 699                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
 700                 else
 701                         nr = -1;
 702
 703                 trans = btrfs_join_transaction(root);
 704                 if (IS_ERR(trans)) {
 705                         ret = PTR_ERR(trans);
 706                         break;
 707                 }
 708                 ret = btrfs_run_delayed_items_nr(trans, nr);
 709                 btrfs_end_transaction(trans);
 710                 break;
 711         case FLUSH_DELALLOC:
 712         case FLUSH_DELALLOC_WAIT:
 713                 shrink_delalloc(fs_info, space_info, num_bytes,
 714                                 state == FLUSH_DELALLOC_WAIT, for_preempt);
 715                 break;
 716         case FLUSH_DELAYED_REFS_NR:
 717         case FLUSH_DELAYED_REFS:
 718                 trans = btrfs_join_transaction(root);
 719                 if (IS_ERR(trans)) {
 720                         ret = PTR_ERR(trans);
 721                         break;
 722                 }
 723                 if (state == FLUSH_DELAYED_REFS_NR)
 724                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
 725                 else
 726                         nr = 0;
 727                 btrfs_run_delayed_refs(trans, nr);
 728                 btrfs_end_transaction(trans);
 729                 break;
 730         case ALLOC_CHUNK:
 731         case ALLOC_CHUNK_FORCE:
 732                 trans = btrfs_join_transaction(root);
 733                 if (IS_ERR(trans)) {
 734                         ret = PTR_ERR(trans);
 735                         break;
 736                 }
 737                 ret = btrfs_chunk_alloc(trans,
 738                                 btrfs_get_alloc_profile(fs_info, space_info->flags),
 739                                 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
 740                                         CHUNK_ALLOC_FORCE);
 741                 btrfs_end_transaction(trans);
 742                 if (ret > 0 || ret == -ENOSPC)
 743                         ret = 0;
 744                 break;
 745         case RUN_DELAYED_IPUTS:
 746                 /*
 747                  * If we have pending delayed iputs then we could free up a
 748                  * bunch of pinned space, so make sure we run the iputs before
 749                  * we do our pinned bytes check below.
 750                  */
 751                 btrfs_run_delayed_iputs(fs_info);
 752                 btrfs_wait_on_delayed_iputs(fs_info);
 753                 break;
 754         case COMMIT_TRANS:
 755                 ret = may_commit_transaction(fs_info, space_info);
 756                 break;
 757         case FORCE_COMMIT_TRANS:
 758                 trans = btrfs_join_transaction(root);
 759                 if (IS_ERR(trans)) {
 760                         ret = PTR_ERR(trans);
 761                         break;
 762                 }
 763                 ret = btrfs_commit_transaction(trans);
 764                 break;
 765         default:
 766                 ret = -ENOSPC;
 767                 break;
 768         }
 769
 770         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
 771                                 ret, for_preempt);
 772         return;
 773 }
 774
 775 static inline u64
 776 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
 777                                  struct btrfs_space_info *space_info)
 778 {
 779         u64 used;
 780         u64 avail;
 781         u64 to_reclaim = space_info->reclaim_size;
 782
 783         lockdep_assert_held(&space_info->lock);
 784
 785         avail = calc_available_free_space(fs_info, space_info,
 786                                           BTRFS_RESERVE_FLUSH_ALL);
 787         used = btrfs_space_info_used(space_info, true);
 788
 789         /*
 790          * We may be flushing because suddenly we have less space than we had
 791          * before, and now we're well over-committed based on our current free
 792          * space.  If that's the case add in our overage so we make sure to put
 793          * appropriate pressure on the flushing state machine.
 794          */
 795         if (space_info->total_bytes + avail < used)
 796                 to_reclaim += used - (space_info->total_bytes + avail);
 797
 798         return to_reclaim;
 799 }
 800
 801 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
 802                                     struct btrfs_space_info *space_info)
 803 {
 804         u64 global_rsv_size = fs_info->global_block_rsv.reserved;
 805         u64 ordered, delalloc;
 806         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
 807         u64 used;
 808
 809         /* If we're just plain full then async reclaim just slows us down. */
 810         if ((space_info->bytes_used + space_info->bytes_reserved +
 811              global_rsv_size) >= thresh)
 812                 return false;
 813
 814         /*
 815          * We have tickets queued, bail so we don't compete with the async
 816          * flushers.
 817          */
 818         if (space_info->reclaim_size)
 819                 return false;
 820
 821         /*
 822          * If we have over half of the free space occupied by reservations or
 823          * pinned then we want to start flushing.
 824          *
 825          * We do not do the traditional thing here, which is to say
 826          *
 827          *   if (used >= ((total_bytes + avail) / 2))
 828          *     return 1;
 829          *
 830          * because this doesn't quite work how we want.  If we had more than 50%
 831          * of the space_info used by bytes_used and we had 0 available we'd just
 832          * constantly run the background flusher.  Instead we want it to kick in
 833          * if our reclaimable space exceeds our clamped free space.
 834          *
 835          * Our clamping range is 2^1 -> 2^8.  Practically speaking that means
 836          * the following:
 837          *
 838          * Amount of RAM        Minimum threshold       Maximum threshold
 839          *
 840          *        256GiB                     1GiB                  128GiB
 841          *        128GiB                   512MiB                   64GiB
 842          *         64GiB                   256MiB                   32GiB
 843          *         32GiB                   128MiB                   16GiB
 844          *         16GiB                    64MiB                    8GiB
 845          *
 846          * These are the range our thresholds will fall in, corresponding to how
 847          * much delalloc we need for the background flusher to kick in.
 848          */
 849
 850         thresh = calc_available_free_space(fs_info, space_info,
 851                                            BTRFS_RESERVE_FLUSH_ALL);
 852         used = space_info->bytes_used + space_info->bytes_reserved +
 853                space_info->bytes_readonly + global_rsv_size;
 854         if (used < space_info->total_bytes)
 855                 thresh += space_info->total_bytes - used;
 856         thresh >>= space_info->clamp;
 857
 858         used = space_info->bytes_pinned;
 859
 860         /*
 861          * If we have more ordered bytes than delalloc bytes then we're either
 862          * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
 863          * around.  Preemptive flushing is only useful in that it can free up
 864          * space before tickets need to wait for things to finish.  In the case
 865          * of ordered extents, preemptively waiting on ordered extents gets us
 866          * nothing, if our reservations are tied up in ordered extents we'll
 867          * simply have to slow down writers by forcing them to wait on ordered
 868          * extents.
 869          *
 870          * In the case that ordered is larger than delalloc, only include the
 871          * block reserves that we would actually be able to directly reclaim
 872          * from.  In this case if we're heavy on metadata operations this will
 873          * clearly be heavy enough to warrant preemptive flushing.  In the case
 874          * of heavy DIO or ordered reservations, preemptive flushing will just
 875          * waste time and cause us to slow down.
 876          *
 877          * We want to make sure we truly are maxed out on ordered however, so
 878          * cut ordered in half, and if it's still higher than delalloc then we
 879          * can keep flushing.  This is to avoid the case where we start
 880          * flushing, and now delalloc == ordered and we stop preemptively
 881          * flushing when we could still have several gigs of delalloc to flush.
 882          */
 883         ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
 884         delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
 885         if (ordered >= delalloc)
 886                 used += fs_info->delayed_refs_rsv.reserved +
 887                         fs_info->delayed_block_rsv.reserved;
 888         else
 889                 used += space_info->bytes_may_use - global_rsv_size;
 890
 891         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
 892                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
 893 }
 894
 895 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
 896                                   struct btrfs_space_info *space_info,
 897                                   struct reserve_ticket *ticket)
 898 {
 899         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
 900         u64 min_bytes;
 901
 902         if (global_rsv->space_info != space_info)
 903                 return false;
 904
 905         spin_lock(&global_rsv->lock);
 906         min_bytes = div_factor(global_rsv->size, 1);
 907         if (global_rsv->reserved < min_bytes + ticket->bytes) {
 908                 spin_unlock(&global_rsv->lock);
 909                 return false;
 910         }
 911         global_rsv->reserved -= ticket->bytes;
 912         remove_ticket(space_info, ticket);
 913         ticket->bytes = 0;
 914         wake_up(&ticket->wait);
 915         space_info->tickets_id++;
 916         if (global_rsv->reserved < global_rsv->size)
 917                 global_rsv->full = 0;
 918         spin_unlock(&global_rsv->lock);
 919
 920         return true;
 921 }
 922
 923 /*
 924  * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
 925  * @fs_info - fs_info for this fs
 926  * @space_info - the space info we were flushing
 927  *
 928  * We call this when we've exhausted our flushing ability and haven't made
 929  * progress in satisfying tickets.  The reservation code handles tickets in
 930  * order, so if there is a large ticket first and then smaller ones we could
 931  * very well satisfy the smaller tickets.  This will attempt to wake up any
 932  * tickets in the list to catch this case.
 933  *
 934  * This function returns true if it was able to make progress by clearing out
 935  * other tickets, or if it stumbles across a ticket that was smaller than the
 936  * first ticket.
 937  */
 938 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
 939                                    struct btrfs_space_info *space_info)
 940 {
 941         struct reserve_ticket *ticket;
 942         u64 tickets_id = space_info->tickets_id;
 943         u64 first_ticket_bytes = 0;
 944
 945         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
 946                 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
 947                 __btrfs_dump_space_info(fs_info, space_info);
 948         }
 949
 950         while (!list_empty(&space_info->tickets) &&
 951                tickets_id == space_info->tickets_id) {
 952                 ticket = list_first_entry(&space_info->tickets,
 953                                           struct reserve_ticket, list);
 954
 955                 if (ticket->steal &&
 956                     steal_from_global_rsv(fs_info, space_info, ticket))
 957                         return true;
 958
 959                 /*
 960                  * may_commit_transaction will avoid committing the transaction
 961                  * if it doesn't feel like the space reclaimed by the commit
 962                  * would result in the ticket succeeding.  However if we have a
 963                  * smaller ticket in the queue it may be small enough to be
 964                  * satisified by committing the transaction, so if any
 965                  * subsequent ticket is smaller than the first ticket go ahead
 966                  * and send us back for another loop through the enospc flushing
 967                  * code.
 968                  */
 969                 if (first_ticket_bytes == 0)
 970                         first_ticket_bytes = ticket->bytes;
 971                 else if (first_ticket_bytes > ticket->bytes)
 972                         return true;
 973
 974                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
 975                         btrfs_info(fs_info, "failing ticket with %llu bytes",
 976                                    ticket->bytes);
 977
 978                 remove_ticket(space_info, ticket);
 979                 ticket->error = -ENOSPC;
 980                 wake_up(&ticket->wait);
 981
 982                 /*
 983                  * We're just throwing tickets away, so more flushing may not
 984                  * trip over btrfs_try_granting_tickets, so we need to call it
 985                  * here to see if we can make progress with the next ticket in
 986                  * the list.
 987                  */
 988                 btrfs_try_granting_tickets(fs_info, space_info);
 989         }
 990         return (tickets_id != space_info->tickets_id);
 991 }
 992
 993 /*
 994  * This is for normal flushers, we can wait all goddamned day if we want to.  We
 995  * will loop and continuously try to flush as long as we are making progress.
 996  * We count progress as clearing off tickets each time we have to loop.
 997  */
 998 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
 999 {
1000         struct btrfs_fs_info *fs_info;
1001         struct btrfs_space_info *space_info;
1002         u64 to_reclaim;
1003         enum btrfs_flush_state flush_state;
1004         int commit_cycles = 0;
1005         u64 last_tickets_id;
1006
1007         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
1008         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1009
1010         spin_lock(&space_info->lock);
1011         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1012         if (!to_reclaim) {
1013                 space_info->flush = 0;
1014                 spin_unlock(&space_info->lock);
1015                 return;
1016         }
1017         last_tickets_id = space_info->tickets_id;
1018         spin_unlock(&space_info->lock);
1019
1020         flush_state = FLUSH_DELAYED_ITEMS_NR;
1021         do {
1022                 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
1023                 spin_lock(&space_info->lock);
1024                 if (list_empty(&space_info->tickets)) {
1025                         space_info->flush = 0;
1026                         spin_unlock(&space_info->lock);
1027                         return;
1028                 }
1029                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1030                                                               space_info);
1031                 if (last_tickets_id == space_info->tickets_id) {
1032                         flush_state++;
1033                 } else {
1034                         last_tickets_id = space_info->tickets_id;
1035                         flush_state = FLUSH_DELAYED_ITEMS_NR;
1036                         if (commit_cycles)
1037                                 commit_cycles--;
1038                 }
1039
1040                 /*
1041                  * We don't want to force a chunk allocation until we've tried
1042                  * pretty hard to reclaim space.  Think of the case where we
1043                  * freed up a bunch of space and so have a lot of pinned space
1044                  * to reclaim.  We would rather use that than possibly create a
1045                  * underutilized metadata chunk.  So if this is our first run
1046                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
1047                  * commit the transaction.  If nothing has changed the next go
1048                  * around then we can force a chunk allocation.
1049                  */
1050                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1051                         flush_state++;
1052
1053                 if (flush_state > COMMIT_TRANS) {
1054                         commit_cycles++;
1055                         if (commit_cycles > 2) {
1056                                 if (maybe_fail_all_tickets(fs_info, space_info)) {
1057                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
1058                                         commit_cycles--;
1059                                 } else {
1060                                         space_info->flush = 0;
1061                                 }
1062                         } else {
1063                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
1064                         }
1065                 }
1066                 spin_unlock(&space_info->lock);
1067         } while (flush_state <= COMMIT_TRANS);
1068 }
1069
1070 /*
1071  * This handles pre-flushing of metadata space before we get to the point that
1072  * we need to start blocking threads on tickets.  The logic here is different
1073  * from the other flush paths because it doesn't rely on tickets to tell us how
1074  * much we need to flush, instead it attempts to keep us below the 80% full
1075  * watermark of space by flushing whichever reservation pool is currently the
1076  * largest.
1077  */
1078 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1079 {
1080         struct btrfs_fs_info *fs_info;
1081         struct btrfs_space_info *space_info;
1082         struct btrfs_block_rsv *delayed_block_rsv;
1083         struct btrfs_block_rsv *delayed_refs_rsv;
1084         struct btrfs_block_rsv *global_rsv;
1085         struct btrfs_block_rsv *trans_rsv;
1086         int loops = 0;
1087
1088         fs_info = container_of(work, struct btrfs_fs_info,
1089                                preempt_reclaim_work);
1090         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1091         delayed_block_rsv = &fs_info->delayed_block_rsv;
1092         delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1093         global_rsv = &fs_info->global_block_rsv;
1094         trans_rsv = &fs_info->trans_block_rsv;
1095
1096         spin_lock(&space_info->lock);
1097         while (need_preemptive_reclaim(fs_info, space_info)) {
1098                 enum btrfs_flush_state flush;
1099                 u64 delalloc_size = 0;
1100                 u64 to_reclaim, block_rsv_size;
1101                 u64 global_rsv_size = global_rsv->reserved;
1102
1103                 loops++;
1104
1105                 /*
1106                  * We don't have a precise counter for the metadata being
1107                  * reserved for delalloc, so we'll approximate it by subtracting
1108                  * out the block rsv's space from the bytes_may_use.  If that
1109                  * amount is higher than the individual reserves, then we can
1110                  * assume it's tied up in delalloc reservations.
1111                  */
1112                 block_rsv_size = global_rsv_size +
1113                         delayed_block_rsv->reserved +
1114                         delayed_refs_rsv->reserved +
1115                         trans_rsv->reserved;
1116                 if (block_rsv_size < space_info->bytes_may_use)
1117                         delalloc_size = space_info->bytes_may_use - block_rsv_size;
1118                 spin_unlock(&space_info->lock);
1119
1120                 /*
1121                  * We don't want to include the global_rsv in our calculation,
1122                  * because that's space we can't touch.  Subtract it from the
1123                  * block_rsv_size for the next checks.
1124                  */
1125                 block_rsv_size -= global_rsv_size;
1126
1127                 /*
1128                  * We really want to avoid flushing delalloc too much, as it
1129                  * could result in poor allocation patterns, so only flush it if
1130                  * it's larger than the rest of the pools combined.
1131                  */
1132                 if (delalloc_size > block_rsv_size) {
1133                         to_reclaim = delalloc_size;
1134                         flush = FLUSH_DELALLOC;
1135                 } else if (space_info->bytes_pinned >
1136                            (delayed_block_rsv->reserved +
1137                             delayed_refs_rsv->reserved)) {
1138                         to_reclaim = space_info->bytes_pinned;
1139                         flush = FORCE_COMMIT_TRANS;
1140                 } else if (delayed_block_rsv->reserved >
1141                            delayed_refs_rsv->reserved) {
1142                         to_reclaim = delayed_block_rsv->reserved;
1143                         flush = FLUSH_DELAYED_ITEMS_NR;
1144                 } else {
1145                         to_reclaim = delayed_refs_rsv->reserved;
1146                         flush = FLUSH_DELAYED_REFS_NR;
1147                 }
1148
1149                 /*
1150                  * We don't want to reclaim everything, just a portion, so scale
1151                  * down the to_reclaim by 1/4.  If it takes us down to 0,
1152                  * reclaim 1 items worth.
1153                  */
1154                 to_reclaim >>= 2;
1155                 if (!to_reclaim)
1156                         to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1157                 flush_space(fs_info, space_info, to_reclaim, flush, true);
1158                 cond_resched();
1159                 spin_lock(&space_info->lock);
1160         }
1161
1162         /* We only went through once, back off our clamping. */
1163         if (loops == 1 && !space_info->reclaim_size)
1164                 space_info->clamp = max(1, space_info->clamp - 1);
1165         trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1166         spin_unlock(&space_info->lock);
1167 }
1168
1169 /*
1170  * FLUSH_DELALLOC_WAIT:
1171  *   Space is freed from flushing delalloc in one of two ways.
1172  *
1173  *   1) compression is on and we allocate less space than we reserved
1174  *   2) we are overwriting existing space
1175  *
1176  *   For #1 that extra space is reclaimed as soon as the delalloc pages are
1177  *   COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1178  *   length to ->bytes_reserved, and subtracts the reserved space from
1179  *   ->bytes_may_use.
1180  *
1181  *   For #2 this is trickier.  Once the ordered extent runs we will drop the
1182  *   extent in the range we are overwriting, which creates a delayed ref for
1183  *   that freed extent.  This however is not reclaimed until the transaction
1184  *   commits, thus the next stages.
1185  *
1186  * RUN_DELAYED_IPUTS
1187  *   If we are freeing inodes, we want to make sure all delayed iputs have
1188  *   completed, because they could have been on an inode with i_nlink == 0, and
1189  *   thus have been truncated and freed up space.  But again this space is not
1190  *   immediately re-usable, it comes in the form of a delayed ref, which must be
1191  *   run and then the transaction must be committed.
1192  *
1193  * FLUSH_DELAYED_REFS
1194  *   The above two cases generate delayed refs that will affect
1195  *   ->total_bytes_pinned.  However this counter can be inconsistent with
1196  *   reality if there are outstanding delayed refs.  This is because we adjust
1197  *   the counter based solely on the current set of delayed refs and disregard
1198  *   any on-disk state which might include more refs.  So for example, if we
1199  *   have an extent with 2 references, but we only drop 1, we'll see that there
1200  *   is a negative delayed ref count for the extent and assume that the space
1201  *   will be freed, and thus increase ->total_bytes_pinned.
1202  *
1203  *   Running the delayed refs gives us the actual real view of what will be
1204  *   freed at the transaction commit time.  This stage will not actually free
1205  *   space for us, it just makes sure that may_commit_transaction() has all of
1206  *   the information it needs to make the right decision.
1207  *
1208  * COMMIT_TRANS
1209  *   This is where we reclaim all of the pinned space generated by the previous
1210  *   two stages.  We will not commit the transaction if we don't think we're
1211  *   likely to satisfy our request, which means if our current free space +
1212  *   total_bytes_pinned < reservation we will not commit.  This is why the
1213  *   previous states are actually important, to make sure we know for sure
1214  *   whether committing the transaction will allow us to make progress.
1215  *
1216  * ALLOC_CHUNK_FORCE
1217  *   For data we start with alloc chunk force, however we could have been full
1218  *   before, and then the transaction commit could have freed new block groups,
1219  *   so if we now have space to allocate do the force chunk allocation.
1220  */
1221 static const enum btrfs_flush_state data_flush_states[] = {
1222         FLUSH_DELALLOC_WAIT,
1223         RUN_DELAYED_IPUTS,
1224         FLUSH_DELAYED_REFS,
1225         COMMIT_TRANS,
1226         ALLOC_CHUNK_FORCE,
1227 };
1228
1229 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1230 {
1231         struct btrfs_fs_info *fs_info;
1232         struct btrfs_space_info *space_info;
1233         u64 last_tickets_id;
1234         enum btrfs_flush_state flush_state = 0;
1235
1236         fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1237         space_info = fs_info->data_sinfo;
1238
1239         spin_lock(&space_info->lock);
1240         if (list_empty(&space_info->tickets)) {
1241                 space_info->flush = 0;
1242                 spin_unlock(&space_info->lock);
1243                 return;
1244         }
1245         last_tickets_id = space_info->tickets_id;
1246         spin_unlock(&space_info->lock);
1247
1248         while (!space_info->full) {
1249                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1250                 spin_lock(&space_info->lock);
1251                 if (list_empty(&space_info->tickets)) {
1252                         space_info->flush = 0;
1253                         spin_unlock(&space_info->lock);
1254                         return;
1255                 }
1256                 last_tickets_id = space_info->tickets_id;
1257                 spin_unlock(&space_info->lock);
1258         }
1259
1260         while (flush_state < ARRAY_SIZE(data_flush_states)) {
1261                 flush_space(fs_info, space_info, U64_MAX,
1262                             data_flush_states[flush_state], false);
1263                 spin_lock(&space_info->lock);
1264                 if (list_empty(&space_info->tickets)) {
1265                         space_info->flush = 0;
1266                         spin_unlock(&space_info->lock);
1267                         return;
1268                 }
1269
1270                 if (last_tickets_id == space_info->tickets_id) {
1271                         flush_state++;
1272                 } else {
1273                         last_tickets_id = space_info->tickets_id;
1274                         flush_state = 0;
1275                 }
1276
1277                 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1278                         if (space_info->full) {
1279                                 if (maybe_fail_all_tickets(fs_info, space_info))
1280                                         flush_state = 0;
1281                                 else
1282                                         space_info->flush = 0;
1283                         } else {
1284                                 flush_state = 0;
1285                         }
1286                 }
1287                 spin_unlock(&space_info->lock);
1288         }
1289 }
1290
1291 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1292 {
1293         INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1294         INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1295         INIT_WORK(&fs_info->preempt_reclaim_work,
1296                   btrfs_preempt_reclaim_metadata_space);
1297 }
1298
1299 static const enum btrfs_flush_state priority_flush_states[] = {
1300         FLUSH_DELAYED_ITEMS_NR,
1301         FLUSH_DELAYED_ITEMS,
1302         ALLOC_CHUNK,
1303 };
1304
1305 static const enum btrfs_flush_state evict_flush_states[] = {
1306         FLUSH_DELAYED_ITEMS_NR,
1307         FLUSH_DELAYED_ITEMS,
1308         FLUSH_DELAYED_REFS_NR,
1309         FLUSH_DELAYED_REFS,
1310         FLUSH_DELALLOC,
1311         FLUSH_DELALLOC_WAIT,
1312         ALLOC_CHUNK,
1313         COMMIT_TRANS,
1314 };
1315
1316 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1317                                 struct btrfs_space_info *space_info,
1318                                 struct reserve_ticket *ticket,
1319                                 const enum btrfs_flush_state *states,
1320                                 int states_nr)
1321 {
1322         u64 to_reclaim;
1323         int flush_state;
1324
1325         spin_lock(&space_info->lock);
1326         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1327         if (!to_reclaim) {
1328                 spin_unlock(&space_info->lock);
1329                 return;
1330         }
1331         spin_unlock(&space_info->lock);
1332
1333         flush_state = 0;
1334         do {
1335                 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1336                             false);
1337                 flush_state++;
1338                 spin_lock(&space_info->lock);
1339                 if (ticket->bytes == 0) {
1340                         spin_unlock(&space_info->lock);
1341                         return;
1342                 }
1343                 spin_unlock(&space_info->lock);
1344         } while (flush_state < states_nr);
1345 }
1346
1347 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1348                                         struct btrfs_space_info *space_info,
1349                                         struct reserve_ticket *ticket)
1350 {
1351         while (!space_info->full) {
1352                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1353                 spin_lock(&space_info->lock);
1354                 if (ticket->bytes == 0) {
1355                         spin_unlock(&space_info->lock);
1356                         return;
1357                 }
1358                 spin_unlock(&space_info->lock);
1359         }
1360 }
1361
1362 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1363                                 struct btrfs_space_info *space_info,
1364                                 struct reserve_ticket *ticket)
1365
1366 {
1367         DEFINE_WAIT(wait);
1368         int ret = 0;
1369
1370         spin_lock(&space_info->lock);
1371         while (ticket->bytes > 0 && ticket->error == 0) {
1372                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1373                 if (ret) {
1374                         /*
1375                          * Delete us from the list. After we unlock the space
1376                          * info, we don't want the async reclaim job to reserve
1377                          * space for this ticket. If that would happen, then the
1378                          * ticket's task would not known that space was reserved
1379                          * despite getting an error, resulting in a space leak
1380                          * (bytes_may_use counter of our space_info).
1381                          */
1382                         remove_ticket(space_info, ticket);
1383                         ticket->error = -EINTR;
1384                         break;
1385                 }
1386                 spin_unlock(&space_info->lock);
1387
1388                 schedule();
1389
1390                 finish_wait(&ticket->wait, &wait);
1391                 spin_lock(&space_info->lock);
1392         }
1393         spin_unlock(&space_info->lock);
1394 }
1395
1396 /**
1397  * Do the appropriate flushing and waiting for a ticket
1398  *
1399  * @fs_info:    the filesystem
1400  * @space_info: space info for the reservation
1401  * @ticket:     ticket for the reservation
1402  * @start_ns:   timestamp when the reservation started
1403  * @orig_bytes: amount of bytes originally reserved
1404  * @flush:      how much we can flush
1405  *
1406  * This does the work of figuring out how to flush for the ticket, waiting for
1407  * the reservation, and returning the appropriate error if there is one.
1408  */
1409 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1410                                  struct btrfs_space_info *space_info,
1411                                  struct reserve_ticket *ticket,
1412                                  u64 start_ns, u64 orig_bytes,
1413                                  enum btrfs_reserve_flush_enum flush)
1414 {
1415         int ret;
1416
1417         switch (flush) {
1418         case BTRFS_RESERVE_FLUSH_DATA:
1419         case BTRFS_RESERVE_FLUSH_ALL:
1420         case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1421                 wait_reserve_ticket(fs_info, space_info, ticket);
1422                 break;
1423         case BTRFS_RESERVE_FLUSH_LIMIT:
1424                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1425                                                 priority_flush_states,
1426                                                 ARRAY_SIZE(priority_flush_states));
1427                 break;
1428         case BTRFS_RESERVE_FLUSH_EVICT:
1429                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1430                                                 evict_flush_states,
1431                                                 ARRAY_SIZE(evict_flush_states));
1432                 break;
1433         case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1434                 priority_reclaim_data_space(fs_info, space_info, ticket);
1435                 break;
1436         default:
1437                 ASSERT(0);
1438                 break;
1439         }
1440
1441         spin_lock(&space_info->lock);
1442         ret = ticket->error;
1443         if (ticket->bytes || ticket->error) {
1444                 /*
1445                  * We were a priority ticket, so we need to delete ourselves
1446                  * from the list.  Because we could have other priority tickets
1447                  * behind us that require less space, run
1448                  * btrfs_try_granting_tickets() to see if their reservations can
1449                  * now be made.
1450                  */
1451                 if (!list_empty(&ticket->list)) {
1452                         remove_ticket(space_info, ticket);
1453                         btrfs_try_granting_tickets(fs_info, space_info);
1454                 }
1455
1456                 if (!ret)
1457                         ret = -ENOSPC;
1458         }
1459         spin_unlock(&space_info->lock);
1460         ASSERT(list_empty(&ticket->list));
1461         /*
1462          * Check that we can't have an error set if the reservation succeeded,
1463          * as that would confuse tasks and lead them to error out without
1464          * releasing reserved space (if an error happens the expectation is that
1465          * space wasn't reserved at all).
1466          */
1467         ASSERT(!(ticket->bytes == 0 && ticket->error));
1468         trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1469                                    start_ns, flush, ticket->error);
1470         return ret;
1471 }
1472
1473 /*
1474  * This returns true if this flush state will go through the ordinary flushing
1475  * code.
1476  */
1477 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1478 {
1479         return  (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1480                 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1481 }
1482
1483 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1484                                        struct btrfs_space_info *space_info)
1485 {
1486         u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1487         u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1488
1489         /*
1490          * If we're heavy on ordered operations then clamping won't help us.  We
1491          * need to clamp specifically to keep up with dirty'ing buffered
1492          * writers, because there's not a 1:1 correlation of writing delalloc
1493          * and freeing space, like there is with flushing delayed refs or
1494          * delayed nodes.  If we're already more ordered than delalloc then
1495          * we're keeping up, otherwise we aren't and should probably clamp.
1496          */
1497         if (ordered < delalloc)
1498                 space_info->clamp = min(space_info->clamp + 1, 8);
1499 }
1500
1501 /**
1502  * Try to reserve bytes from the block_rsv's space
1503  *
1504  * @fs_info:    the filesystem
1505  * @space_info: space info we want to allocate from
1506  * @orig_bytes: number of bytes we want
1507  * @flush:      whether or not we can flush to make our reservation
1508  *
1509  * This will reserve orig_bytes number of bytes from the space info associated
1510  * with the block_rsv.  If there is not enough space it will make an attempt to
1511  * flush out space to make room.  It will do this by flushing delalloc if
1512  * possible or committing the transaction.  If flush is 0 then no attempts to
1513  * regain reservations will be made and this will fail if there is not enough
1514  * space already.
1515  */
1516 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1517                            struct btrfs_space_info *space_info, u64 orig_bytes,
1518                            enum btrfs_reserve_flush_enum flush)
1519 {
1520         struct work_struct *async_work;
1521         struct reserve_ticket ticket;
1522         u64 start_ns = 0;
1523         u64 used;
1524         int ret = 0;
1525         bool pending_tickets;
1526
1527         ASSERT(orig_bytes);
1528         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1529
1530         if (flush == BTRFS_RESERVE_FLUSH_DATA)
1531                 async_work = &fs_info->async_data_reclaim_work;
1532         else
1533                 async_work = &fs_info->async_reclaim_work;
1534
1535         spin_lock(&space_info->lock);
1536         ret = -ENOSPC;
1537         used = btrfs_space_info_used(space_info, true);
1538
1539         /*
1540          * We don't want NO_FLUSH allocations to jump everybody, they can
1541          * generally handle ENOSPC in a different way, so treat them the same as
1542          * normal flushers when it comes to skipping pending tickets.
1543          */
1544         if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1545                 pending_tickets = !list_empty(&space_info->tickets) ||
1546                         !list_empty(&space_info->priority_tickets);
1547         else
1548                 pending_tickets = !list_empty(&space_info->priority_tickets);
1549
1550         /*
1551          * Carry on if we have enough space (short-circuit) OR call
1552          * can_overcommit() to ensure we can overcommit to continue.
1553          */
1554         if (!pending_tickets &&
1555             ((used + orig_bytes <= space_info->total_bytes) ||
1556              btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1557                 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1558                                                       orig_bytes);
1559                 ret = 0;
1560         }
1561
1562         /*
1563          * If we couldn't make a reservation then setup our reservation ticket
1564          * and kick the async worker if it's not already running.
1565          *
1566          * If we are a priority flusher then we just need to add our ticket to
1567          * the list and we will do our own flushing further down.
1568          */
1569         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1570                 ticket.bytes = orig_bytes;
1571                 ticket.error = 0;
1572                 space_info->reclaim_size += ticket.bytes;
1573                 init_waitqueue_head(&ticket.wait);
1574                 ticket.steal = (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1575                 if (trace_btrfs_reserve_ticket_enabled())
1576                         start_ns = ktime_get_ns();
1577
1578                 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1579                     flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1580                     flush == BTRFS_RESERVE_FLUSH_DATA) {
1581                         list_add_tail(&ticket.list, &space_info->tickets);
1582                         if (!space_info->flush) {
1583                                 /*
1584                                  * We were forced to add a reserve ticket, so
1585                                  * our preemptive flushing is unable to keep
1586                                  * up.  Clamp down on the threshold for the
1587                                  * preemptive flushing in order to keep up with
1588                                  * the workload.
1589                                  */
1590                                 maybe_clamp_preempt(fs_info, space_info);
1591
1592                                 space_info->flush = 1;
1593                                 trace_btrfs_trigger_flush(fs_info,
1594                                                           space_info->flags,
1595                                                           orig_bytes, flush,
1596                                                           "enospc");
1597                                 queue_work(system_unbound_wq, async_work);
1598                         }
1599                 } else {
1600                         list_add_tail(&ticket.list,
1601                                       &space_info->priority_tickets);
1602                 }
1603         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1604                 used += orig_bytes;
1605                 /*
1606                  * We will do the space reservation dance during log replay,
1607                  * which means we won't have fs_info->fs_root set, so don't do
1608                  * the async reclaim as we will panic.
1609                  */
1610                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1611                     !work_busy(&fs_info->preempt_reclaim_work) &&
1612                     need_preemptive_reclaim(fs_info, space_info)) {
1613                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
1614                                                   orig_bytes, flush, "preempt");
1615                         queue_work(system_unbound_wq,
1616                                    &fs_info->preempt_reclaim_work);
1617                 }
1618         }
1619         spin_unlock(&space_info->lock);
1620         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1621                 return ret;
1622
1623         return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1624                                      orig_bytes, flush);
1625 }
1626
1627 /**
1628  * Trye to reserve metadata bytes from the block_rsv's space
1629  *
1630  * @root:       the root we're allocating for
1631  * @block_rsv:  block_rsv we're allocating for
1632  * @orig_bytes: number of bytes we want
1633  * @flush:      whether or not we can flush to make our reservation
1634  *
1635  * This will reserve orig_bytes number of bytes from the space info associated
1636  * with the block_rsv.  If there is not enough space it will make an attempt to
1637  * flush out space to make room.  It will do this by flushing delalloc if
1638  * possible or committing the transaction.  If flush is 0 then no attempts to
1639  * regain reservations will be made and this will fail if there is not enough
1640  * space already.
1641  */
1642 int btrfs_reserve_metadata_bytes(struct btrfs_root *root,
1643                                  struct btrfs_block_rsv *block_rsv,
1644                                  u64 orig_bytes,
1645                                  enum btrfs_reserve_flush_enum flush)
1646 {
1647         struct btrfs_fs_info *fs_info = root->fs_info;
1648         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
1649         int ret;
1650
1651         ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush);
1652         if (ret == -ENOSPC &&
1653             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
1654                 if (block_rsv != global_rsv &&
1655                     !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes))
1656                         ret = 0;
1657         }
1658         if (ret == -ENOSPC) {
1659                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1660                                               block_rsv->space_info->flags,
1661                                               orig_bytes, 1);
1662
1663                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1664                         btrfs_dump_space_info(fs_info, block_rsv->space_info,
1665                                               orig_bytes, 0);
1666         }
1667         return ret;
1668 }
1669
1670 /**
1671  * Try to reserve data bytes for an allocation
1672  *
1673  * @fs_info: the filesystem
1674  * @bytes:   number of bytes we need
1675  * @flush:   how we are allowed to flush
1676  *
1677  * This will reserve bytes from the data space info.  If there is not enough
1678  * space then we will attempt to flush space as specified by flush.
1679  */
1680 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1681                              enum btrfs_reserve_flush_enum flush)
1682 {
1683         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1684         int ret;
1685
1686         ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1687                flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE);
1688         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1689
1690         ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1691         if (ret == -ENOSPC) {
1692                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1693                                               data_sinfo->flags, bytes, 1);
1694                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1695                         btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1696         }
1697         return ret;
1698 }