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  *     This will commit the transaction.  Historically we had a lot of logic
 137  *     surrounding whether or not we'd commit the transaction, but this waits born
 138  *     out of a pre-tickets era where we could end up committing the transaction
 139  *     thousands of times in a row without making progress.  Now thanks to our
 140  *     ticketing system we know if we're not making progress and can error
 141  *     everybody out after a few commits rather than burning the disk hoping for
 142  *     a different answer.
 143  *
 144  * OVERCOMMIT
 145  *
 146  *   Because we hold so many reservations for metadata we will allow you to
 147  *   reserve more space than is currently free in the currently allocate
 148  *   metadata space.  This only happens with metadata, data does not allow
 149  *   overcommitting.
 150  *
 151  *   You can see the current logic for when we allow overcommit in
 152  *   btrfs_can_overcommit(), but it only applies to unallocated space.  If there
 153  *   is no unallocated space to be had, all reservations are kept within the
 154  *   free space in the allocated metadata chunks.
 155  *
 156  *   Because of overcommitting, you generally want to use the
 157  *   btrfs_can_overcommit() logic for metadata allocations, as it does the right
 158  *   thing with or without extra unallocated space.
 159  */
 160
 161 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
 162                           bool may_use_included)
 163 {
 164         ASSERT(s_info);
 165         return s_info->bytes_used + s_info->bytes_reserved +
 166                 s_info->bytes_pinned + s_info->bytes_readonly +
 167                 s_info->bytes_zone_unusable +
 168                 (may_use_included ? s_info->bytes_may_use : 0);
 169 }
 170
 171 /*
 172  * after adding space to the filesystem, we need to clear the full flags
 173  * on all the space infos.
 174  */
 175 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 176 {
 177         struct list_head *head = &info->space_info;
 178         struct btrfs_space_info *found;
 179
 180         list_for_each_entry(found, head, list)
 181                 found->full = 0;
 182 }
 183
 184 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
 185 {
 186
 187         struct btrfs_space_info *space_info;
 188         int i;
 189         int ret;
 190
 191         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
 192         if (!space_info)
 193                 return -ENOMEM;
 194
 195         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
 196                 INIT_LIST_HEAD(&space_info->block_groups[i]);
 197         init_rwsem(&space_info->groups_sem);
 198         spin_lock_init(&space_info->lock);
 199         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
 200         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
 201         INIT_LIST_HEAD(&space_info->ro_bgs);
 202         INIT_LIST_HEAD(&space_info->tickets);
 203         INIT_LIST_HEAD(&space_info->priority_tickets);
 204         space_info->clamp = 1;
 205
 206         ret = btrfs_sysfs_add_space_info_type(info, space_info);
 207         if (ret)
 208                 return ret;
 209
 210         list_add(&space_info->list, &info->space_info);
 211         if (flags & BTRFS_BLOCK_GROUP_DATA)
 212                 info->data_sinfo = space_info;
 213
 214         return ret;
 215 }
 216
 217 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
 218 {
 219         struct btrfs_super_block *disk_super;
 220         u64 features;
 221         u64 flags;
 222         int mixed = 0;
 223         int ret;
 224
 225         disk_super = fs_info->super_copy;
 226         if (!btrfs_super_root(disk_super))
 227                 return -EINVAL;
 228
 229         features = btrfs_super_incompat_flags(disk_super);
 230         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
 231                 mixed = 1;
 232
 233         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 234         ret = create_space_info(fs_info, flags);
 235         if (ret)
 236                 goto out;
 237
 238         if (mixed) {
 239                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
 240                 ret = create_space_info(fs_info, flags);
 241         } else {
 242                 flags = BTRFS_BLOCK_GROUP_METADATA;
 243                 ret = create_space_info(fs_info, flags);
 244                 if (ret)
 245                         goto out;
 246
 247                 flags = BTRFS_BLOCK_GROUP_DATA;
 248                 ret = create_space_info(fs_info, flags);
 249         }
 250 out:
 251         return ret;
 252 }
 253
 254 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags,
 255                              u64 total_bytes, u64 bytes_used,
 256                              u64 bytes_readonly, u64 bytes_zone_unusable,
 257                              struct btrfs_space_info **space_info)
 258 {
 259         struct btrfs_space_info *found;
 260         int factor;
 261
 262         factor = btrfs_bg_type_to_factor(flags);
 263
 264         found = btrfs_find_space_info(info, flags);
 265         ASSERT(found);
 266         spin_lock(&found->lock);
 267         found->total_bytes += total_bytes;
 268         found->disk_total += total_bytes * factor;
 269         found->bytes_used += bytes_used;
 270         found->disk_used += bytes_used * factor;
 271         found->bytes_readonly += bytes_readonly;
 272         found->bytes_zone_unusable += bytes_zone_unusable;
 273         if (total_bytes > 0)
 274                 found->full = 0;
 275         btrfs_try_granting_tickets(info, found);
 276         spin_unlock(&found->lock);
 277         *space_info = found;
 278 }
 279
 280 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
 281                                                u64 flags)
 282 {
 283         struct list_head *head = &info->space_info;
 284         struct btrfs_space_info *found;
 285
 286         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 287
 288         list_for_each_entry(found, head, list) {
 289                 if (found->flags & flags)
 290                         return found;
 291         }
 292         return NULL;
 293 }
 294
 295 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
 296                           struct btrfs_space_info *space_info,
 297                           enum btrfs_reserve_flush_enum flush)
 298 {
 299         u64 profile;
 300         u64 avail;
 301         int factor;
 302
 303         if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
 304                 profile = btrfs_system_alloc_profile(fs_info);
 305         else
 306                 profile = btrfs_metadata_alloc_profile(fs_info);
 307
 308         avail = atomic64_read(&fs_info->free_chunk_space);
 309
 310         /*
 311          * If we have dup, raid1 or raid10 then only half of the free
 312          * space is actually usable.  For raid56, the space info used
 313          * doesn't include the parity drive, so we don't have to
 314          * change the math
 315          */
 316         factor = btrfs_bg_type_to_factor(profile);
 317         avail = div_u64(avail, factor);
 318
 319         /*
 320          * If we aren't flushing all things, let us overcommit up to
 321          * 1/2th of the space. If we can flush, don't let us overcommit
 322          * too much, let it overcommit up to 1/8 of the space.
 323          */
 324         if (flush == BTRFS_RESERVE_FLUSH_ALL)
 325                 avail >>= 3;
 326         else
 327                 avail >>= 1;
 328         return avail;
 329 }
 330
 331 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
 332                          struct btrfs_space_info *space_info, u64 bytes,
 333                          enum btrfs_reserve_flush_enum flush)
 334 {
 335         u64 avail;
 336         u64 used;
 337
 338         /* Don't overcommit when in mixed mode */
 339         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
 340                 return 0;
 341
 342         used = btrfs_space_info_used(space_info, true);
 343         avail = calc_available_free_space(fs_info, space_info, flush);
 344
 345         if (used + bytes < space_info->total_bytes + avail)
 346                 return 1;
 347         return 0;
 348 }
 349
 350 static void remove_ticket(struct btrfs_space_info *space_info,
 351                           struct reserve_ticket *ticket)
 352 {
 353         if (!list_empty(&ticket->list)) {
 354                 list_del_init(&ticket->list);
 355                 ASSERT(space_info->reclaim_size >= ticket->bytes);
 356                 space_info->reclaim_size -= ticket->bytes;
 357         }
 358 }
 359
 360 /*
 361  * This is for space we already have accounted in space_info->bytes_may_use, so
 362  * basically when we're returning space from block_rsv's.
 363  */
 364 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
 365                                 struct btrfs_space_info *space_info)
 366 {
 367         struct list_head *head;
 368         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
 369
 370         lockdep_assert_held(&space_info->lock);
 371
 372         head = &space_info->priority_tickets;
 373 again:
 374         while (!list_empty(head)) {
 375                 struct reserve_ticket *ticket;
 376                 u64 used = btrfs_space_info_used(space_info, true);
 377
 378                 ticket = list_first_entry(head, struct reserve_ticket, list);
 379
 380                 /* Check and see if our ticket can be satisfied now. */
 381                 if ((used + ticket->bytes <= space_info->total_bytes) ||
 382                     btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
 383                                          flush)) {
 384                         btrfs_space_info_update_bytes_may_use(fs_info,
 385                                                               space_info,
 386                                                               ticket->bytes);
 387                         remove_ticket(space_info, ticket);
 388                         ticket->bytes = 0;
 389                         space_info->tickets_id++;
 390                         wake_up(&ticket->wait);
 391                 } else {
 392                         break;
 393                 }
 394         }
 395
 396         if (head == &space_info->priority_tickets) {
 397                 head = &space_info->tickets;
 398                 flush = BTRFS_RESERVE_FLUSH_ALL;
 399                 goto again;
 400         }
 401 }
 402
 403 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
 404 do {                                                                    \
 405         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
 406         spin_lock(&__rsv->lock);                                        \
 407         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
 408                    __rsv->size, __rsv->reserved);                       \
 409         spin_unlock(&__rsv->lock);                                      \
 410 } while (0)
 411
 412 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 413                                     struct btrfs_space_info *info)
 414 {
 415         lockdep_assert_held(&info->lock);
 416
 417         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
 418                    info->flags,
 419                    info->total_bytes - btrfs_space_info_used(info, true),
 420                    info->full ? "" : "not ");
 421         btrfs_info(fs_info,
 422                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
 423                 info->total_bytes, info->bytes_used, info->bytes_pinned,
 424                 info->bytes_reserved, info->bytes_may_use,
 425                 info->bytes_readonly, info->bytes_zone_unusable);
 426
 427         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
 428         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
 429         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
 430         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
 431         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
 432
 433 }
 434
 435 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 436                            struct btrfs_space_info *info, u64 bytes,
 437                            int dump_block_groups)
 438 {
 439         struct btrfs_block_group *cache;
 440         int index = 0;
 441
 442         spin_lock(&info->lock);
 443         __btrfs_dump_space_info(fs_info, info);
 444         spin_unlock(&info->lock);
 445
 446         if (!dump_block_groups)
 447                 return;
 448
 449         down_read(&info->groups_sem);
 450 again:
 451         list_for_each_entry(cache, &info->block_groups[index], list) {
 452                 spin_lock(&cache->lock);
 453                 btrfs_info(fs_info,
 454                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s",
 455                         cache->start, cache->length, cache->used, cache->pinned,
 456                         cache->reserved, cache->zone_unusable,
 457                         cache->ro ? "[readonly]" : "");
 458                 spin_unlock(&cache->lock);
 459                 btrfs_dump_free_space(cache, bytes);
 460         }
 461         if (++index < BTRFS_NR_RAID_TYPES)
 462                 goto again;
 463         up_read(&info->groups_sem);
 464 }
 465
 466 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
 467                                         u64 to_reclaim)
 468 {
 469         u64 bytes;
 470         u64 nr;
 471
 472         bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 473         nr = div64_u64(to_reclaim, bytes);
 474         if (!nr)
 475                 nr = 1;
 476         return nr;
 477 }
 478
 479 #define EXTENT_SIZE_PER_ITEM    SZ_256K
 480
 481 /*
 482  * shrink metadata reservation for delalloc
 483  */
 484 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
 485                             struct btrfs_space_info *space_info,
 486                             u64 to_reclaim, bool wait_ordered,
 487                             bool for_preempt)
 488 {
 489         struct btrfs_trans_handle *trans;
 490         u64 delalloc_bytes;
 491         u64 ordered_bytes;
 492         u64 items;
 493         long time_left;
 494         int loops;
 495
 496         /* Calc the number of the pages we need flush for space reservation */
 497         if (to_reclaim == U64_MAX) {
 498                 items = U64_MAX;
 499         } else {
 500                 /*
 501                  * to_reclaim is set to however much metadata we need to
 502                  * reclaim, but reclaiming that much data doesn't really track
 503                  * exactly, so increase the amount to reclaim by 2x in order to
 504                  * make sure we're flushing enough delalloc to hopefully reclaim
 505                  * some metadata reservations.
 506                  */
 507                 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
 508                 to_reclaim = items * EXTENT_SIZE_PER_ITEM;
 509         }
 510
 511         trans = (struct btrfs_trans_handle *)current->journal_info;
 512
 513         delalloc_bytes = percpu_counter_sum_positive(
 514                                                 &fs_info->delalloc_bytes);
 515         ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
 516         if (delalloc_bytes == 0 && ordered_bytes == 0)
 517                 return;
 518
 519         /*
 520          * If we are doing more ordered than delalloc we need to just wait on
 521          * ordered extents, otherwise we'll waste time trying to flush delalloc
 522          * that likely won't give us the space back we need.
 523          */
 524         if (ordered_bytes > delalloc_bytes && !for_preempt)
 525                 wait_ordered = true;
 526
 527         loops = 0;
 528         while ((delalloc_bytes || ordered_bytes) && loops < 3) {
 529                 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
 530                 long nr_pages = min_t(u64, temp, LONG_MAX);
 531
 532                 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
 533
 534                 loops++;
 535                 if (wait_ordered && !trans) {
 536                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
 537                 } else {
 538                         time_left = schedule_timeout_killable(1);
 539                         if (time_left)
 540                                 break;
 541                 }
 542
 543                 /*
 544                  * If we are for preemption we just want a one-shot of delalloc
 545                  * flushing so we can stop flushing if we decide we don't need
 546                  * to anymore.
 547                  */
 548                 if (for_preempt)
 549                         break;
 550
 551                 spin_lock(&space_info->lock);
 552                 if (list_empty(&space_info->tickets) &&
 553                     list_empty(&space_info->priority_tickets)) {
 554                         spin_unlock(&space_info->lock);
 555                         break;
 556                 }
 557                 spin_unlock(&space_info->lock);
 558
 559                 delalloc_bytes = percpu_counter_sum_positive(
 560                                                 &fs_info->delalloc_bytes);
 561                 ordered_bytes = percpu_counter_sum_positive(
 562                                                 &fs_info->ordered_bytes);
 563         }
 564 }
 565
 566 /*
 567  * Try to flush some data based on policy set by @state. This is only advisory
 568  * and may fail for various reasons. The caller is supposed to examine the
 569  * state of @space_info to detect the outcome.
 570  */
 571 static void flush_space(struct btrfs_fs_info *fs_info,
 572                        struct btrfs_space_info *space_info, u64 num_bytes,
 573                        enum btrfs_flush_state state, bool for_preempt)
 574 {
 575         struct btrfs_root *root = fs_info->extent_root;
 576         struct btrfs_trans_handle *trans;
 577         int nr;
 578         int ret = 0;
 579
 580         switch (state) {
 581         case FLUSH_DELAYED_ITEMS_NR:
 582         case FLUSH_DELAYED_ITEMS:
 583                 if (state == FLUSH_DELAYED_ITEMS_NR)
 584                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
 585                 else
 586                         nr = -1;
 587
 588                 trans = btrfs_join_transaction(root);
 589                 if (IS_ERR(trans)) {
 590                         ret = PTR_ERR(trans);
 591                         break;
 592                 }
 593                 ret = btrfs_run_delayed_items_nr(trans, nr);
 594                 btrfs_end_transaction(trans);
 595                 break;
 596         case FLUSH_DELALLOC:
 597         case FLUSH_DELALLOC_WAIT:
 598                 shrink_delalloc(fs_info, space_info, num_bytes,
 599                                 state == FLUSH_DELALLOC_WAIT, for_preempt);
 600                 break;
 601         case FLUSH_DELAYED_REFS_NR:
 602         case FLUSH_DELAYED_REFS:
 603                 trans = btrfs_join_transaction(root);
 604                 if (IS_ERR(trans)) {
 605                         ret = PTR_ERR(trans);
 606                         break;
 607                 }
 608                 if (state == FLUSH_DELAYED_REFS_NR)
 609                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
 610                 else
 611                         nr = 0;
 612                 btrfs_run_delayed_refs(trans, nr);
 613                 btrfs_end_transaction(trans);
 614                 break;
 615         case ALLOC_CHUNK:
 616         case ALLOC_CHUNK_FORCE:
 617                 trans = btrfs_join_transaction(root);
 618                 if (IS_ERR(trans)) {
 619                         ret = PTR_ERR(trans);
 620                         break;
 621                 }
 622                 ret = btrfs_chunk_alloc(trans,
 623                                 btrfs_get_alloc_profile(fs_info, space_info->flags),
 624                                 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
 625                                         CHUNK_ALLOC_FORCE);
 626                 btrfs_end_transaction(trans);
 627                 if (ret > 0 || ret == -ENOSPC)
 628                         ret = 0;
 629                 break;
 630         case RUN_DELAYED_IPUTS:
 631                 /*
 632                  * If we have pending delayed iputs then we could free up a
 633                  * bunch of pinned space, so make sure we run the iputs before
 634                  * we do our pinned bytes check below.
 635                  */
 636                 btrfs_run_delayed_iputs(fs_info);
 637                 btrfs_wait_on_delayed_iputs(fs_info);
 638                 break;
 639         case COMMIT_TRANS:
 640                 ASSERT(current->journal_info == NULL);
 641                 trans = btrfs_join_transaction(root);
 642                 if (IS_ERR(trans)) {
 643                         ret = PTR_ERR(trans);
 644                         break;
 645                 }
 646                 ret = btrfs_commit_transaction(trans);
 647                 break;
 648         default:
 649                 ret = -ENOSPC;
 650                 break;
 651         }
 652
 653         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
 654                                 ret, for_preempt);
 655         return;
 656 }
 657
 658 static inline u64
 659 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
 660                                  struct btrfs_space_info *space_info)
 661 {
 662         u64 used;
 663         u64 avail;
 664         u64 to_reclaim = space_info->reclaim_size;
 665
 666         lockdep_assert_held(&space_info->lock);
 667
 668         avail = calc_available_free_space(fs_info, space_info,
 669                                           BTRFS_RESERVE_FLUSH_ALL);
 670         used = btrfs_space_info_used(space_info, true);
 671
 672         /*
 673          * We may be flushing because suddenly we have less space than we had
 674          * before, and now we're well over-committed based on our current free
 675          * space.  If that's the case add in our overage so we make sure to put
 676          * appropriate pressure on the flushing state machine.
 677          */
 678         if (space_info->total_bytes + avail < used)
 679                 to_reclaim += used - (space_info->total_bytes + avail);
 680
 681         return to_reclaim;
 682 }
 683
 684 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
 685                                     struct btrfs_space_info *space_info)
 686 {
 687         u64 global_rsv_size = fs_info->global_block_rsv.reserved;
 688         u64 ordered, delalloc;
 689         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
 690         u64 used;
 691
 692         /* If we're just plain full then async reclaim just slows us down. */
 693         if ((space_info->bytes_used + space_info->bytes_reserved +
 694              global_rsv_size) >= thresh)
 695                 return false;
 696
 697         /*
 698          * We have tickets queued, bail so we don't compete with the async
 699          * flushers.
 700          */
 701         if (space_info->reclaim_size)
 702                 return false;
 703
 704         /*
 705          * If we have over half of the free space occupied by reservations or
 706          * pinned then we want to start flushing.
 707          *
 708          * We do not do the traditional thing here, which is to say
 709          *
 710          *   if (used >= ((total_bytes + avail) / 2))
 711          *     return 1;
 712          *
 713          * because this doesn't quite work how we want.  If we had more than 50%
 714          * of the space_info used by bytes_used and we had 0 available we'd just
 715          * constantly run the background flusher.  Instead we want it to kick in
 716          * if our reclaimable space exceeds our clamped free space.
 717          *
 718          * Our clamping range is 2^1 -> 2^8.  Practically speaking that means
 719          * the following:
 720          *
 721          * Amount of RAM        Minimum threshold       Maximum threshold
 722          *
 723          *        256GiB                     1GiB                  128GiB
 724          *        128GiB                   512MiB                   64GiB
 725          *         64GiB                   256MiB                   32GiB
 726          *         32GiB                   128MiB                   16GiB
 727          *         16GiB                    64MiB                    8GiB
 728          *
 729          * These are the range our thresholds will fall in, corresponding to how
 730          * much delalloc we need for the background flusher to kick in.
 731          */
 732
 733         thresh = calc_available_free_space(fs_info, space_info,
 734                                            BTRFS_RESERVE_FLUSH_ALL);
 735         used = space_info->bytes_used + space_info->bytes_reserved +
 736                space_info->bytes_readonly + global_rsv_size;
 737         if (used < space_info->total_bytes)
 738                 thresh += space_info->total_bytes - used;
 739         thresh >>= space_info->clamp;
 740
 741         used = space_info->bytes_pinned;
 742
 743         /*
 744          * If we have more ordered bytes than delalloc bytes then we're either
 745          * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
 746          * around.  Preemptive flushing is only useful in that it can free up
 747          * space before tickets need to wait for things to finish.  In the case
 748          * of ordered extents, preemptively waiting on ordered extents gets us
 749          * nothing, if our reservations are tied up in ordered extents we'll
 750          * simply have to slow down writers by forcing them to wait on ordered
 751          * extents.
 752          *
 753          * In the case that ordered is larger than delalloc, only include the
 754          * block reserves that we would actually be able to directly reclaim
 755          * from.  In this case if we're heavy on metadata operations this will
 756          * clearly be heavy enough to warrant preemptive flushing.  In the case
 757          * of heavy DIO or ordered reservations, preemptive flushing will just
 758          * waste time and cause us to slow down.
 759          *
 760          * We want to make sure we truly are maxed out on ordered however, so
 761          * cut ordered in half, and if it's still higher than delalloc then we
 762          * can keep flushing.  This is to avoid the case where we start
 763          * flushing, and now delalloc == ordered and we stop preemptively
 764          * flushing when we could still have several gigs of delalloc to flush.
 765          */
 766         ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
 767         delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
 768         if (ordered >= delalloc)
 769                 used += fs_info->delayed_refs_rsv.reserved +
 770                         fs_info->delayed_block_rsv.reserved;
 771         else
 772                 used += space_info->bytes_may_use - global_rsv_size;
 773
 774         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
 775                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
 776 }
 777
 778 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
 779                                   struct btrfs_space_info *space_info,
 780                                   struct reserve_ticket *ticket)
 781 {
 782         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
 783         u64 min_bytes;
 784
 785         if (global_rsv->space_info != space_info)
 786                 return false;
 787
 788         spin_lock(&global_rsv->lock);
 789         min_bytes = div_factor(global_rsv->size, 1);
 790         if (global_rsv->reserved < min_bytes + ticket->bytes) {
 791                 spin_unlock(&global_rsv->lock);
 792                 return false;
 793         }
 794         global_rsv->reserved -= ticket->bytes;
 795         remove_ticket(space_info, ticket);
 796         ticket->bytes = 0;
 797         wake_up(&ticket->wait);
 798         space_info->tickets_id++;
 799         if (global_rsv->reserved < global_rsv->size)
 800                 global_rsv->full = 0;
 801         spin_unlock(&global_rsv->lock);
 802
 803         return true;
 804 }
 805
 806 /*
 807  * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
 808  * @fs_info - fs_info for this fs
 809  * @space_info - the space info we were flushing
 810  *
 811  * We call this when we've exhausted our flushing ability and haven't made
 812  * progress in satisfying tickets.  The reservation code handles tickets in
 813  * order, so if there is a large ticket first and then smaller ones we could
 814  * very well satisfy the smaller tickets.  This will attempt to wake up any
 815  * tickets in the list to catch this case.
 816  *
 817  * This function returns true if it was able to make progress by clearing out
 818  * other tickets, or if it stumbles across a ticket that was smaller than the
 819  * first ticket.
 820  */
 821 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
 822                                    struct btrfs_space_info *space_info)
 823 {
 824         struct reserve_ticket *ticket;
 825         u64 tickets_id = space_info->tickets_id;
 826
 827         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
 828                 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
 829                 __btrfs_dump_space_info(fs_info, space_info);
 830         }
 831
 832         while (!list_empty(&space_info->tickets) &&
 833                tickets_id == space_info->tickets_id) {
 834                 ticket = list_first_entry(&space_info->tickets,
 835                                           struct reserve_ticket, list);
 836
 837                 if (ticket->steal &&
 838                     steal_from_global_rsv(fs_info, space_info, ticket))
 839                         return true;
 840
 841                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
 842                         btrfs_info(fs_info, "failing ticket with %llu bytes",
 843                                    ticket->bytes);
 844
 845                 remove_ticket(space_info, ticket);
 846                 ticket->error = -ENOSPC;
 847                 wake_up(&ticket->wait);
 848
 849                 /*
 850                  * We're just throwing tickets away, so more flushing may not
 851                  * trip over btrfs_try_granting_tickets, so we need to call it
 852                  * here to see if we can make progress with the next ticket in
 853                  * the list.
 854                  */
 855                 btrfs_try_granting_tickets(fs_info, space_info);
 856         }
 857         return (tickets_id != space_info->tickets_id);
 858 }
 859
 860 /*
 861  * This is for normal flushers, we can wait all goddamned day if we want to.  We
 862  * will loop and continuously try to flush as long as we are making progress.
 863  * We count progress as clearing off tickets each time we have to loop.
 864  */
 865 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
 866 {
 867         struct btrfs_fs_info *fs_info;
 868         struct btrfs_space_info *space_info;
 869         u64 to_reclaim;
 870         enum btrfs_flush_state flush_state;
 871         int commit_cycles = 0;
 872         u64 last_tickets_id;
 873
 874         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
 875         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
 876
 877         spin_lock(&space_info->lock);
 878         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
 879         if (!to_reclaim) {
 880                 space_info->flush = 0;
 881                 spin_unlock(&space_info->lock);
 882                 return;
 883         }
 884         last_tickets_id = space_info->tickets_id;
 885         spin_unlock(&space_info->lock);
 886
 887         flush_state = FLUSH_DELAYED_ITEMS_NR;
 888         do {
 889                 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
 890                 spin_lock(&space_info->lock);
 891                 if (list_empty(&space_info->tickets)) {
 892                         space_info->flush = 0;
 893                         spin_unlock(&space_info->lock);
 894                         return;
 895                 }
 896                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
 897                                                               space_info);
 898                 if (last_tickets_id == space_info->tickets_id) {
 899                         flush_state++;
 900                 } else {
 901                         last_tickets_id = space_info->tickets_id;
 902                         flush_state = FLUSH_DELAYED_ITEMS_NR;
 903                         if (commit_cycles)
 904                                 commit_cycles--;
 905                 }
 906
 907                 /*
 908                  * We don't want to force a chunk allocation until we've tried
 909                  * pretty hard to reclaim space.  Think of the case where we
 910                  * freed up a bunch of space and so have a lot of pinned space
 911                  * to reclaim.  We would rather use that than possibly create a
 912                  * underutilized metadata chunk.  So if this is our first run
 913                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
 914                  * commit the transaction.  If nothing has changed the next go
 915                  * around then we can force a chunk allocation.
 916                  */
 917                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
 918                         flush_state++;
 919
 920                 if (flush_state > COMMIT_TRANS) {
 921                         commit_cycles++;
 922                         if (commit_cycles > 2) {
 923                                 if (maybe_fail_all_tickets(fs_info, space_info)) {
 924                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
 925                                         commit_cycles--;
 926                                 } else {
 927                                         space_info->flush = 0;
 928                                 }
 929                         } else {
 930                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
 931                         }
 932                 }
 933                 spin_unlock(&space_info->lock);
 934         } while (flush_state <= COMMIT_TRANS);
 935 }
 936
 937 /*
 938  * This handles pre-flushing of metadata space before we get to the point that
 939  * we need to start blocking threads on tickets.  The logic here is different
 940  * from the other flush paths because it doesn't rely on tickets to tell us how
 941  * much we need to flush, instead it attempts to keep us below the 80% full
 942  * watermark of space by flushing whichever reservation pool is currently the
 943  * largest.
 944  */
 945 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
 946 {
 947         struct btrfs_fs_info *fs_info;
 948         struct btrfs_space_info *space_info;
 949         struct btrfs_block_rsv *delayed_block_rsv;
 950         struct btrfs_block_rsv *delayed_refs_rsv;
 951         struct btrfs_block_rsv *global_rsv;
 952         struct btrfs_block_rsv *trans_rsv;
 953         int loops = 0;
 954
 955         fs_info = container_of(work, struct btrfs_fs_info,
 956                                preempt_reclaim_work);
 957         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
 958         delayed_block_rsv = &fs_info->delayed_block_rsv;
 959         delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 960         global_rsv = &fs_info->global_block_rsv;
 961         trans_rsv = &fs_info->trans_block_rsv;
 962
 963         spin_lock(&space_info->lock);
 964         while (need_preemptive_reclaim(fs_info, space_info)) {
 965                 enum btrfs_flush_state flush;
 966                 u64 delalloc_size = 0;
 967                 u64 to_reclaim, block_rsv_size;
 968                 u64 global_rsv_size = global_rsv->reserved;
 969
 970                 loops++;
 971
 972                 /*
 973                  * We don't have a precise counter for the metadata being
 974                  * reserved for delalloc, so we'll approximate it by subtracting
 975                  * out the block rsv's space from the bytes_may_use.  If that
 976                  * amount is higher than the individual reserves, then we can
 977                  * assume it's tied up in delalloc reservations.
 978                  */
 979                 block_rsv_size = global_rsv_size +
 980                         delayed_block_rsv->reserved +
 981                         delayed_refs_rsv->reserved +
 982                         trans_rsv->reserved;
 983                 if (block_rsv_size < space_info->bytes_may_use)
 984                         delalloc_size = space_info->bytes_may_use - block_rsv_size;
 985                 spin_unlock(&space_info->lock);
 986
 987                 /*
 988                  * We don't want to include the global_rsv in our calculation,
 989                  * because that's space we can't touch.  Subtract it from the
 990                  * block_rsv_size for the next checks.
 991                  */
 992                 block_rsv_size -= global_rsv_size;
 993
 994                 /*
 995                  * We really want to avoid flushing delalloc too much, as it
 996                  * could result in poor allocation patterns, so only flush it if
 997                  * it's larger than the rest of the pools combined.
 998                  */
 999                 if (delalloc_size > block_rsv_size) {
1000                         to_reclaim = delalloc_size;
1001                         flush = FLUSH_DELALLOC;
1002                 } else if (space_info->bytes_pinned >
1003                            (delayed_block_rsv->reserved +
1004                             delayed_refs_rsv->reserved)) {
1005                         to_reclaim = space_info->bytes_pinned;
1006                         flush = COMMIT_TRANS;
1007                 } else if (delayed_block_rsv->reserved >
1008                            delayed_refs_rsv->reserved) {
1009                         to_reclaim = delayed_block_rsv->reserved;
1010                         flush = FLUSH_DELAYED_ITEMS_NR;
1011                 } else {
1012                         to_reclaim = delayed_refs_rsv->reserved;
1013                         flush = FLUSH_DELAYED_REFS_NR;
1014                 }
1015
1016                 /*
1017                  * We don't want to reclaim everything, just a portion, so scale
1018                  * down the to_reclaim by 1/4.  If it takes us down to 0,
1019                  * reclaim 1 items worth.
1020                  */
1021                 to_reclaim >>= 2;
1022                 if (!to_reclaim)
1023                         to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1024                 flush_space(fs_info, space_info, to_reclaim, flush, true);
1025                 cond_resched();
1026                 spin_lock(&space_info->lock);
1027         }
1028
1029         /* We only went through once, back off our clamping. */
1030         if (loops == 1 && !space_info->reclaim_size)
1031                 space_info->clamp = max(1, space_info->clamp - 1);
1032         trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1033         spin_unlock(&space_info->lock);
1034 }
1035
1036 /*
1037  * FLUSH_DELALLOC_WAIT:
1038  *   Space is freed from flushing delalloc in one of two ways.
1039  *
1040  *   1) compression is on and we allocate less space than we reserved
1041  *   2) we are overwriting existing space
1042  *
1043  *   For #1 that extra space is reclaimed as soon as the delalloc pages are
1044  *   COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1045  *   length to ->bytes_reserved, and subtracts the reserved space from
1046  *   ->bytes_may_use.
1047  *
1048  *   For #2 this is trickier.  Once the ordered extent runs we will drop the
1049  *   extent in the range we are overwriting, which creates a delayed ref for
1050  *   that freed extent.  This however is not reclaimed until the transaction
1051  *   commits, thus the next stages.
1052  *
1053  * RUN_DELAYED_IPUTS
1054  *   If we are freeing inodes, we want to make sure all delayed iputs have
1055  *   completed, because they could have been on an inode with i_nlink == 0, and
1056  *   thus have been truncated and freed up space.  But again this space is not
1057  *   immediately re-usable, it comes in the form of a delayed ref, which must be
1058  *   run and then the transaction must be committed.
1059  *
1060  * COMMIT_TRANS
1061  *   This is where we reclaim all of the pinned space generated by running the
1062  *   iputs
1063  *
1064  * ALLOC_CHUNK_FORCE
1065  *   For data we start with alloc chunk force, however we could have been full
1066  *   before, and then the transaction commit could have freed new block groups,
1067  *   so if we now have space to allocate do the force chunk allocation.
1068  */
1069 static const enum btrfs_flush_state data_flush_states[] = {
1070         FLUSH_DELALLOC_WAIT,
1071         RUN_DELAYED_IPUTS,
1072         COMMIT_TRANS,
1073         ALLOC_CHUNK_FORCE,
1074 };
1075
1076 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1077 {
1078         struct btrfs_fs_info *fs_info;
1079         struct btrfs_space_info *space_info;
1080         u64 last_tickets_id;
1081         enum btrfs_flush_state flush_state = 0;
1082
1083         fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1084         space_info = fs_info->data_sinfo;
1085
1086         spin_lock(&space_info->lock);
1087         if (list_empty(&space_info->tickets)) {
1088                 space_info->flush = 0;
1089                 spin_unlock(&space_info->lock);
1090                 return;
1091         }
1092         last_tickets_id = space_info->tickets_id;
1093         spin_unlock(&space_info->lock);
1094
1095         while (!space_info->full) {
1096                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1097                 spin_lock(&space_info->lock);
1098                 if (list_empty(&space_info->tickets)) {
1099                         space_info->flush = 0;
1100                         spin_unlock(&space_info->lock);
1101                         return;
1102                 }
1103                 last_tickets_id = space_info->tickets_id;
1104                 spin_unlock(&space_info->lock);
1105         }
1106
1107         while (flush_state < ARRAY_SIZE(data_flush_states)) {
1108                 flush_space(fs_info, space_info, U64_MAX,
1109                             data_flush_states[flush_state], false);
1110                 spin_lock(&space_info->lock);
1111                 if (list_empty(&space_info->tickets)) {
1112                         space_info->flush = 0;
1113                         spin_unlock(&space_info->lock);
1114                         return;
1115                 }
1116
1117                 if (last_tickets_id == space_info->tickets_id) {
1118                         flush_state++;
1119                 } else {
1120                         last_tickets_id = space_info->tickets_id;
1121                         flush_state = 0;
1122                 }
1123
1124                 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1125                         if (space_info->full) {
1126                                 if (maybe_fail_all_tickets(fs_info, space_info))
1127                                         flush_state = 0;
1128                                 else
1129                                         space_info->flush = 0;
1130                         } else {
1131                                 flush_state = 0;
1132                         }
1133                 }
1134                 spin_unlock(&space_info->lock);
1135         }
1136 }
1137
1138 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1139 {
1140         INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1141         INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1142         INIT_WORK(&fs_info->preempt_reclaim_work,
1143                   btrfs_preempt_reclaim_metadata_space);
1144 }
1145
1146 static const enum btrfs_flush_state priority_flush_states[] = {
1147         FLUSH_DELAYED_ITEMS_NR,
1148         FLUSH_DELAYED_ITEMS,
1149         ALLOC_CHUNK,
1150 };
1151
1152 static const enum btrfs_flush_state evict_flush_states[] = {
1153         FLUSH_DELAYED_ITEMS_NR,
1154         FLUSH_DELAYED_ITEMS,
1155         FLUSH_DELAYED_REFS_NR,
1156         FLUSH_DELAYED_REFS,
1157         FLUSH_DELALLOC,
1158         FLUSH_DELALLOC_WAIT,
1159         ALLOC_CHUNK,
1160         COMMIT_TRANS,
1161 };
1162
1163 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1164                                 struct btrfs_space_info *space_info,
1165                                 struct reserve_ticket *ticket,
1166                                 const enum btrfs_flush_state *states,
1167                                 int states_nr)
1168 {
1169         u64 to_reclaim;
1170         int flush_state;
1171
1172         spin_lock(&space_info->lock);
1173         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1174         if (!to_reclaim) {
1175                 spin_unlock(&space_info->lock);
1176                 return;
1177         }
1178         spin_unlock(&space_info->lock);
1179
1180         flush_state = 0;
1181         do {
1182                 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1183                             false);
1184                 flush_state++;
1185                 spin_lock(&space_info->lock);
1186                 if (ticket->bytes == 0) {
1187                         spin_unlock(&space_info->lock);
1188                         return;
1189                 }
1190                 spin_unlock(&space_info->lock);
1191         } while (flush_state < states_nr);
1192 }
1193
1194 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1195                                         struct btrfs_space_info *space_info,
1196                                         struct reserve_ticket *ticket)
1197 {
1198         while (!space_info->full) {
1199                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1200                 spin_lock(&space_info->lock);
1201                 if (ticket->bytes == 0) {
1202                         spin_unlock(&space_info->lock);
1203                         return;
1204                 }
1205                 spin_unlock(&space_info->lock);
1206         }
1207 }
1208
1209 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1210                                 struct btrfs_space_info *space_info,
1211                                 struct reserve_ticket *ticket)
1212
1213 {
1214         DEFINE_WAIT(wait);
1215         int ret = 0;
1216
1217         spin_lock(&space_info->lock);
1218         while (ticket->bytes > 0 && ticket->error == 0) {
1219                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1220                 if (ret) {
1221                         /*
1222                          * Delete us from the list. After we unlock the space
1223                          * info, we don't want the async reclaim job to reserve
1224                          * space for this ticket. If that would happen, then the
1225                          * ticket's task would not known that space was reserved
1226                          * despite getting an error, resulting in a space leak
1227                          * (bytes_may_use counter of our space_info).
1228                          */
1229                         remove_ticket(space_info, ticket);
1230                         ticket->error = -EINTR;
1231                         break;
1232                 }
1233                 spin_unlock(&space_info->lock);
1234
1235                 schedule();
1236
1237                 finish_wait(&ticket->wait, &wait);
1238                 spin_lock(&space_info->lock);
1239         }
1240         spin_unlock(&space_info->lock);
1241 }
1242
1243 /**
1244  * Do the appropriate flushing and waiting for a ticket
1245  *
1246  * @fs_info:    the filesystem
1247  * @space_info: space info for the reservation
1248  * @ticket:     ticket for the reservation
1249  * @start_ns:   timestamp when the reservation started
1250  * @orig_bytes: amount of bytes originally reserved
1251  * @flush:      how much we can flush
1252  *
1253  * This does the work of figuring out how to flush for the ticket, waiting for
1254  * the reservation, and returning the appropriate error if there is one.
1255  */
1256 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1257                                  struct btrfs_space_info *space_info,
1258                                  struct reserve_ticket *ticket,
1259                                  u64 start_ns, u64 orig_bytes,
1260                                  enum btrfs_reserve_flush_enum flush)
1261 {
1262         int ret;
1263
1264         switch (flush) {
1265         case BTRFS_RESERVE_FLUSH_DATA:
1266         case BTRFS_RESERVE_FLUSH_ALL:
1267         case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1268                 wait_reserve_ticket(fs_info, space_info, ticket);
1269                 break;
1270         case BTRFS_RESERVE_FLUSH_LIMIT:
1271                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1272                                                 priority_flush_states,
1273                                                 ARRAY_SIZE(priority_flush_states));
1274                 break;
1275         case BTRFS_RESERVE_FLUSH_EVICT:
1276                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1277                                                 evict_flush_states,
1278                                                 ARRAY_SIZE(evict_flush_states));
1279                 break;
1280         case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1281                 priority_reclaim_data_space(fs_info, space_info, ticket);
1282                 break;
1283         default:
1284                 ASSERT(0);
1285                 break;
1286         }
1287
1288         spin_lock(&space_info->lock);
1289         ret = ticket->error;
1290         if (ticket->bytes || ticket->error) {
1291                 /*
1292                  * We were a priority ticket, so we need to delete ourselves
1293                  * from the list.  Because we could have other priority tickets
1294                  * behind us that require less space, run
1295                  * btrfs_try_granting_tickets() to see if their reservations can
1296                  * now be made.
1297                  */
1298                 if (!list_empty(&ticket->list)) {
1299                         remove_ticket(space_info, ticket);
1300                         btrfs_try_granting_tickets(fs_info, space_info);
1301                 }
1302
1303                 if (!ret)
1304                         ret = -ENOSPC;
1305         }
1306         spin_unlock(&space_info->lock);
1307         ASSERT(list_empty(&ticket->list));
1308         /*
1309          * Check that we can't have an error set if the reservation succeeded,
1310          * as that would confuse tasks and lead them to error out without
1311          * releasing reserved space (if an error happens the expectation is that
1312          * space wasn't reserved at all).
1313          */
1314         ASSERT(!(ticket->bytes == 0 && ticket->error));
1315         trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1316                                    start_ns, flush, ticket->error);
1317         return ret;
1318 }
1319
1320 /*
1321  * This returns true if this flush state will go through the ordinary flushing
1322  * code.
1323  */
1324 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1325 {
1326         return  (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1327                 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1328 }
1329
1330 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1331                                        struct btrfs_space_info *space_info)
1332 {
1333         u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1334         u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1335
1336         /*
1337          * If we're heavy on ordered operations then clamping won't help us.  We
1338          * need to clamp specifically to keep up with dirty'ing buffered
1339          * writers, because there's not a 1:1 correlation of writing delalloc
1340          * and freeing space, like there is with flushing delayed refs or
1341          * delayed nodes.  If we're already more ordered than delalloc then
1342          * we're keeping up, otherwise we aren't and should probably clamp.
1343          */
1344         if (ordered < delalloc)
1345                 space_info->clamp = min(space_info->clamp + 1, 8);
1346 }
1347
1348 /**
1349  * Try to reserve bytes from the block_rsv's space
1350  *
1351  * @fs_info:    the filesystem
1352  * @space_info: space info we want to allocate from
1353  * @orig_bytes: number of bytes we want
1354  * @flush:      whether or not we can flush to make our reservation
1355  *
1356  * This will reserve orig_bytes number of bytes from the space info associated
1357  * with the block_rsv.  If there is not enough space it will make an attempt to
1358  * flush out space to make room.  It will do this by flushing delalloc if
1359  * possible or committing the transaction.  If flush is 0 then no attempts to
1360  * regain reservations will be made and this will fail if there is not enough
1361  * space already.
1362  */
1363 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1364                            struct btrfs_space_info *space_info, u64 orig_bytes,
1365                            enum btrfs_reserve_flush_enum flush)
1366 {
1367         struct work_struct *async_work;
1368         struct reserve_ticket ticket;
1369         u64 start_ns = 0;
1370         u64 used;
1371         int ret = 0;
1372         bool pending_tickets;
1373
1374         ASSERT(orig_bytes);
1375         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1376
1377         if (flush == BTRFS_RESERVE_FLUSH_DATA)
1378                 async_work = &fs_info->async_data_reclaim_work;
1379         else
1380                 async_work = &fs_info->async_reclaim_work;
1381
1382         spin_lock(&space_info->lock);
1383         ret = -ENOSPC;
1384         used = btrfs_space_info_used(space_info, true);
1385
1386         /*
1387          * We don't want NO_FLUSH allocations to jump everybody, they can
1388          * generally handle ENOSPC in a different way, so treat them the same as
1389          * normal flushers when it comes to skipping pending tickets.
1390          */
1391         if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1392                 pending_tickets = !list_empty(&space_info->tickets) ||
1393                         !list_empty(&space_info->priority_tickets);
1394         else
1395                 pending_tickets = !list_empty(&space_info->priority_tickets);
1396
1397         /*
1398          * Carry on if we have enough space (short-circuit) OR call
1399          * can_overcommit() to ensure we can overcommit to continue.
1400          */
1401         if (!pending_tickets &&
1402             ((used + orig_bytes <= space_info->total_bytes) ||
1403              btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1404                 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1405                                                       orig_bytes);
1406                 ret = 0;
1407         }
1408
1409         /*
1410          * If we couldn't make a reservation then setup our reservation ticket
1411          * and kick the async worker if it's not already running.
1412          *
1413          * If we are a priority flusher then we just need to add our ticket to
1414          * the list and we will do our own flushing further down.
1415          */
1416         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1417                 ticket.bytes = orig_bytes;
1418                 ticket.error = 0;
1419                 space_info->reclaim_size += ticket.bytes;
1420                 init_waitqueue_head(&ticket.wait);
1421                 ticket.steal = (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1422                 if (trace_btrfs_reserve_ticket_enabled())
1423                         start_ns = ktime_get_ns();
1424
1425                 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1426                     flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1427                     flush == BTRFS_RESERVE_FLUSH_DATA) {
1428                         list_add_tail(&ticket.list, &space_info->tickets);
1429                         if (!space_info->flush) {
1430                                 /*
1431                                  * We were forced to add a reserve ticket, so
1432                                  * our preemptive flushing is unable to keep
1433                                  * up.  Clamp down on the threshold for the
1434                                  * preemptive flushing in order to keep up with
1435                                  * the workload.
1436                                  */
1437                                 maybe_clamp_preempt(fs_info, space_info);
1438
1439                                 space_info->flush = 1;
1440                                 trace_btrfs_trigger_flush(fs_info,
1441                                                           space_info->flags,
1442                                                           orig_bytes, flush,
1443                                                           "enospc");
1444                                 queue_work(system_unbound_wq, async_work);
1445                         }
1446                 } else {
1447                         list_add_tail(&ticket.list,
1448                                       &space_info->priority_tickets);
1449                 }
1450         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1451                 used += orig_bytes;
1452                 /*
1453                  * We will do the space reservation dance during log replay,
1454                  * which means we won't have fs_info->fs_root set, so don't do
1455                  * the async reclaim as we will panic.
1456                  */
1457                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1458                     !work_busy(&fs_info->preempt_reclaim_work) &&
1459                     need_preemptive_reclaim(fs_info, space_info)) {
1460                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
1461                                                   orig_bytes, flush, "preempt");
1462                         queue_work(system_unbound_wq,
1463                                    &fs_info->preempt_reclaim_work);
1464                 }
1465         }
1466         spin_unlock(&space_info->lock);
1467         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1468                 return ret;
1469
1470         return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1471                                      orig_bytes, flush);
1472 }
1473
1474 /**
1475  * Trye to reserve metadata bytes from the block_rsv's space
1476  *
1477  * @root:       the root we're allocating for
1478  * @block_rsv:  block_rsv we're allocating for
1479  * @orig_bytes: number of bytes we want
1480  * @flush:      whether or not we can flush to make our reservation
1481  *
1482  * This will reserve orig_bytes number of bytes from the space info associated
1483  * with the block_rsv.  If there is not enough space it will make an attempt to
1484  * flush out space to make room.  It will do this by flushing delalloc if
1485  * possible or committing the transaction.  If flush is 0 then no attempts to
1486  * regain reservations will be made and this will fail if there is not enough
1487  * space already.
1488  */
1489 int btrfs_reserve_metadata_bytes(struct btrfs_root *root,
1490                                  struct btrfs_block_rsv *block_rsv,
1491                                  u64 orig_bytes,
1492                                  enum btrfs_reserve_flush_enum flush)
1493 {
1494         struct btrfs_fs_info *fs_info = root->fs_info;
1495         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
1496         int ret;
1497
1498         ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush);
1499         if (ret == -ENOSPC &&
1500             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
1501                 if (block_rsv != global_rsv &&
1502                     !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes))
1503                         ret = 0;
1504         }
1505         if (ret == -ENOSPC) {
1506                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1507                                               block_rsv->space_info->flags,
1508                                               orig_bytes, 1);
1509
1510                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1511                         btrfs_dump_space_info(fs_info, block_rsv->space_info,
1512                                               orig_bytes, 0);
1513         }
1514         return ret;
1515 }
1516
1517 /**
1518  * Try to reserve data bytes for an allocation
1519  *
1520  * @fs_info: the filesystem
1521  * @bytes:   number of bytes we need
1522  * @flush:   how we are allowed to flush
1523  *
1524  * This will reserve bytes from the data space info.  If there is not enough
1525  * space then we will attempt to flush space as specified by flush.
1526  */
1527 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1528                              enum btrfs_reserve_flush_enum flush)
1529 {
1530         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1531         int ret;
1532
1533         ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1534                flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE);
1535         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1536
1537         ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1538         if (ret == -ENOSPC) {
1539                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1540                                               data_sinfo->flags, bytes, 1);
1541                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1542                         btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1543         }
1544         return ret;
1545 }