1 // SPDX-License-Identifier: GPL-2.0 2 3 #include "misc.h" 4 #include "ctree.h" 5 #include "space-info.h" 6 #include "sysfs.h" 7 #include "volumes.h" 8 #include "free-space-cache.h" 9 #include "ordered-data.h" 10 #include "transaction.h" 11 #include "block-group.h" 12 #include "zoned.h" 13 #include "fs.h" 14 #include "accessors.h" 15 #include "extent-tree.h" 16 17 /* 18 * HOW DOES SPACE RESERVATION WORK 19 * 20 * If you want to know about delalloc specifically, there is a separate comment 21 * for that with the delalloc code. This comment is about how the whole system 22 * works generally. 23 * 24 * BASIC CONCEPTS 25 * 26 * 1) space_info. This is the ultimate arbiter of how much space we can use. 27 * There's a description of the bytes_ fields with the struct declaration, 28 * refer to that for specifics on each field. Suffice it to say that for 29 * reservations we care about total_bytes - SUM(space_info->bytes_) when 30 * determining if there is space to make an allocation. There is a space_info 31 * for METADATA, SYSTEM, and DATA areas. 32 * 33 * 2) block_rsv's. These are basically buckets for every different type of 34 * metadata reservation we have. You can see the comment in the block_rsv 35 * code on the rules for each type, but generally block_rsv->reserved is how 36 * much space is accounted for in space_info->bytes_may_use. 37 * 38 * 3) btrfs_calc*_size. These are the worst case calculations we used based 39 * on the number of items we will want to modify. We have one for changing 40 * items, and one for inserting new items. Generally we use these helpers to 41 * determine the size of the block reserves, and then use the actual bytes 42 * values to adjust the space_info counters. 43 * 44 * MAKING RESERVATIONS, THE NORMAL CASE 45 * 46 * We call into either btrfs_reserve_data_bytes() or 47 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with 48 * num_bytes we want to reserve. 49 * 50 * ->reserve 51 * space_info->bytes_may_reserve += num_bytes 52 * 53 * ->extent allocation 54 * Call btrfs_add_reserved_bytes() which does 55 * space_info->bytes_may_reserve -= num_bytes 56 * space_info->bytes_reserved += extent_bytes 57 * 58 * ->insert reference 59 * Call btrfs_update_block_group() which does 60 * space_info->bytes_reserved -= extent_bytes 61 * space_info->bytes_used += extent_bytes 62 * 63 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority) 64 * 65 * Assume we are unable to simply make the reservation because we do not have 66 * enough space 67 * 68 * -> __reserve_bytes 69 * create a reserve_ticket with ->bytes set to our reservation, add it to 70 * the tail of space_info->tickets, kick async flush thread 71 * 72 * ->handle_reserve_ticket 73 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set 74 * on the ticket. 75 * 76 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space 77 * Flushes various things attempting to free up space. 78 * 79 * -> btrfs_try_granting_tickets() 80 * This is called by anything that either subtracts space from 81 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the 82 * space_info->total_bytes. This loops through the ->priority_tickets and 83 * then the ->tickets list checking to see if the reservation can be 84 * completed. If it can the space is added to space_info->bytes_may_use and 85 * the ticket is woken up. 86 * 87 * -> ticket wakeup 88 * Check if ->bytes == 0, if it does we got our reservation and we can carry 89 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we 90 * were interrupted.) 91 * 92 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY 93 * 94 * Same as the above, except we add ourselves to the 95 * space_info->priority_tickets, and we do not use ticket->wait, we simply 96 * call flush_space() ourselves for the states that are safe for us to call 97 * without deadlocking and hope for the best. 98 * 99 * THE FLUSHING STATES 100 * 101 * Generally speaking we will have two cases for each state, a "nice" state 102 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to 103 * reduce the locking over head on the various trees, and even to keep from 104 * doing any work at all in the case of delayed refs. Each of these delayed 105 * things however hold reservations, and so letting them run allows us to 106 * reclaim space so we can make new reservations. 107 * 108 * FLUSH_DELAYED_ITEMS 109 * Every inode has a delayed item to update the inode. Take a simple write 110 * for example, we would update the inode item at write time to update the 111 * mtime, and then again at finish_ordered_io() time in order to update the 112 * isize or bytes. We keep these delayed items to coalesce these operations 113 * into a single operation done on demand. These are an easy way to reclaim 114 * metadata space. 115 * 116 * FLUSH_DELALLOC 117 * Look at the delalloc comment to get an idea of how much space is reserved 118 * for delayed allocation. We can reclaim some of this space simply by 119 * running delalloc, but usually we need to wait for ordered extents to 120 * reclaim the bulk of this space. 121 * 122 * FLUSH_DELAYED_REFS 123 * We have a block reserve for the outstanding delayed refs space, and every 124 * delayed ref operation holds a reservation. Running these is a quick way 125 * to reclaim space, but we want to hold this until the end because COW can 126 * churn a lot and we can avoid making some extent tree modifications if we 127 * are able to delay for as long as possible. 128 * 129 * ALLOC_CHUNK 130 * We will skip this the first time through space reservation, because of 131 * overcommit and we don't want to have a lot of useless metadata space when 132 * our worst case reservations will likely never come true. 133 * 134 * RUN_DELAYED_IPUTS 135 * If we're freeing inodes we're likely freeing checksums, file extent 136 * items, and extent tree items. Loads of space could be freed up by these 137 * operations, however they won't be usable until the transaction commits. 138 * 139 * COMMIT_TRANS 140 * This will commit the transaction. Historically we had a lot of logic 141 * surrounding whether or not we'd commit the transaction, but this waits born 142 * out of a pre-tickets era where we could end up committing the transaction 143 * thousands of times in a row without making progress. Now thanks to our 144 * ticketing system we know if we're not making progress and can error 145 * everybody out after a few commits rather than burning the disk hoping for 146 * a different answer. 147 * 148 * OVERCOMMIT 149 * 150 * Because we hold so many reservations for metadata we will allow you to 151 * reserve more space than is currently free in the currently allocate 152 * metadata space. This only happens with metadata, data does not allow 153 * overcommitting. 154 * 155 * You can see the current logic for when we allow overcommit in 156 * btrfs_can_overcommit(), but it only applies to unallocated space. If there 157 * is no unallocated space to be had, all reservations are kept within the 158 * free space in the allocated metadata chunks. 159 * 160 * Because of overcommitting, you generally want to use the 161 * btrfs_can_overcommit() logic for metadata allocations, as it does the right 162 * thing with or without extra unallocated space. 163 */ 164 165 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info, 166 bool may_use_included) 167 { 168 ASSERT(s_info); 169 return s_info->bytes_used + s_info->bytes_reserved + 170 s_info->bytes_pinned + s_info->bytes_readonly + 171 s_info->bytes_zone_unusable + 172 (may_use_included ? s_info->bytes_may_use : 0); 173 } 174 175 /* 176 * after adding space to the filesystem, we need to clear the full flags 177 * on all the space infos. 178 */ 179 void btrfs_clear_space_info_full(struct btrfs_fs_info *info) 180 { 181 struct list_head *head = &info->space_info; 182 struct btrfs_space_info *found; 183 184 list_for_each_entry(found, head, list) 185 found->full = 0; 186 } 187 188 /* 189 * Block groups with more than this value (percents) of unusable space will be 190 * scheduled for background reclaim. 191 */ 192 #define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75) 193 194 /* 195 * Calculate chunk size depending on volume type (regular or zoned). 196 */ 197 static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags) 198 { 199 if (btrfs_is_zoned(fs_info)) 200 return fs_info->zone_size; 201 202 ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK); 203 204 if (flags & BTRFS_BLOCK_GROUP_DATA) 205 return BTRFS_MAX_DATA_CHUNK_SIZE; 206 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM) 207 return SZ_32M; 208 209 /* Handle BTRFS_BLOCK_GROUP_METADATA */ 210 if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G) 211 return SZ_1G; 212 213 return SZ_256M; 214 } 215 216 /* 217 * Update default chunk size. 218 */ 219 void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info, 220 u64 chunk_size) 221 { 222 WRITE_ONCE(space_info->chunk_size, chunk_size); 223 } 224 225 static int create_space_info(struct btrfs_fs_info *info, u64 flags) 226 { 227 228 struct btrfs_space_info *space_info; 229 int i; 230 int ret; 231 232 space_info = kzalloc(sizeof(*space_info), GFP_NOFS); 233 if (!space_info) 234 return -ENOMEM; 235 236 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) 237 INIT_LIST_HEAD(&space_info->block_groups[i]); 238 init_rwsem(&space_info->groups_sem); 239 spin_lock_init(&space_info->lock); 240 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK; 241 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; 242 INIT_LIST_HEAD(&space_info->ro_bgs); 243 INIT_LIST_HEAD(&space_info->tickets); 244 INIT_LIST_HEAD(&space_info->priority_tickets); 245 space_info->clamp = 1; 246 btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags)); 247 248 if (btrfs_is_zoned(info)) 249 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH; 250 251 ret = btrfs_sysfs_add_space_info_type(info, space_info); 252 if (ret) 253 return ret; 254 255 list_add(&space_info->list, &info->space_info); 256 if (flags & BTRFS_BLOCK_GROUP_DATA) 257 info->data_sinfo = space_info; 258 259 return ret; 260 } 261 262 int btrfs_init_space_info(struct btrfs_fs_info *fs_info) 263 { 264 struct btrfs_super_block *disk_super; 265 u64 features; 266 u64 flags; 267 int mixed = 0; 268 int ret; 269 270 disk_super = fs_info->super_copy; 271 if (!btrfs_super_root(disk_super)) 272 return -EINVAL; 273 274 features = btrfs_super_incompat_flags(disk_super); 275 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) 276 mixed = 1; 277 278 flags = BTRFS_BLOCK_GROUP_SYSTEM; 279 ret = create_space_info(fs_info, flags); 280 if (ret) 281 goto out; 282 283 if (mixed) { 284 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA; 285 ret = create_space_info(fs_info, flags); 286 } else { 287 flags = BTRFS_BLOCK_GROUP_METADATA; 288 ret = create_space_info(fs_info, flags); 289 if (ret) 290 goto out; 291 292 flags = BTRFS_BLOCK_GROUP_DATA; 293 ret = create_space_info(fs_info, flags); 294 } 295 out: 296 return ret; 297 } 298 299 void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info, 300 struct btrfs_block_group *block_group) 301 { 302 struct btrfs_space_info *found; 303 int factor, index; 304 305 factor = btrfs_bg_type_to_factor(block_group->flags); 306 307 found = btrfs_find_space_info(info, block_group->flags); 308 ASSERT(found); 309 spin_lock(&found->lock); 310 found->total_bytes += block_group->length; 311 found->disk_total += block_group->length * factor; 312 found->bytes_used += block_group->used; 313 found->disk_used += block_group->used * factor; 314 found->bytes_readonly += block_group->bytes_super; 315 found->bytes_zone_unusable += block_group->zone_unusable; 316 if (block_group->length > 0) 317 found->full = 0; 318 btrfs_try_granting_tickets(info, found); 319 spin_unlock(&found->lock); 320 321 block_group->space_info = found; 322 323 index = btrfs_bg_flags_to_raid_index(block_group->flags); 324 down_write(&found->groups_sem); 325 list_add_tail(&block_group->list, &found->block_groups[index]); 326 up_write(&found->groups_sem); 327 } 328 329 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info, 330 u64 flags) 331 { 332 struct list_head *head = &info->space_info; 333 struct btrfs_space_info *found; 334 335 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK; 336 337 list_for_each_entry(found, head, list) { 338 if (found->flags & flags) 339 return found; 340 } 341 return NULL; 342 } 343 344 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info, 345 struct btrfs_space_info *space_info, 346 enum btrfs_reserve_flush_enum flush) 347 { 348 u64 profile; 349 u64 avail; 350 int factor; 351 352 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM) 353 profile = btrfs_system_alloc_profile(fs_info); 354 else 355 profile = btrfs_metadata_alloc_profile(fs_info); 356 357 avail = atomic64_read(&fs_info->free_chunk_space); 358 359 /* 360 * If we have dup, raid1 or raid10 then only half of the free 361 * space is actually usable. For raid56, the space info used 362 * doesn't include the parity drive, so we don't have to 363 * change the math 364 */ 365 factor = btrfs_bg_type_to_factor(profile); 366 avail = div_u64(avail, factor); 367 368 /* 369 * If we aren't flushing all things, let us overcommit up to 370 * 1/2th of the space. If we can flush, don't let us overcommit 371 * too much, let it overcommit up to 1/8 of the space. 372 */ 373 if (flush == BTRFS_RESERVE_FLUSH_ALL) 374 avail >>= 3; 375 else 376 avail >>= 1; 377 return avail; 378 } 379 380 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info, 381 struct btrfs_space_info *space_info, u64 bytes, 382 enum btrfs_reserve_flush_enum flush) 383 { 384 u64 avail; 385 u64 used; 386 387 /* Don't overcommit when in mixed mode */ 388 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA) 389 return 0; 390 391 used = btrfs_space_info_used(space_info, true); 392 if (test_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags) && 393 (space_info->flags & BTRFS_BLOCK_GROUP_METADATA)) 394 avail = 0; 395 else 396 avail = calc_available_free_space(fs_info, space_info, flush); 397 398 if (used + bytes < space_info->total_bytes + avail) 399 return 1; 400 return 0; 401 } 402 403 static void remove_ticket(struct btrfs_space_info *space_info, 404 struct reserve_ticket *ticket) 405 { 406 if (!list_empty(&ticket->list)) { 407 list_del_init(&ticket->list); 408 ASSERT(space_info->reclaim_size >= ticket->bytes); 409 space_info->reclaim_size -= ticket->bytes; 410 } 411 } 412 413 /* 414 * This is for space we already have accounted in space_info->bytes_may_use, so 415 * basically when we're returning space from block_rsv's. 416 */ 417 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info, 418 struct btrfs_space_info *space_info) 419 { 420 struct list_head *head; 421 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH; 422 423 lockdep_assert_held(&space_info->lock); 424 425 head = &space_info->priority_tickets; 426 again: 427 while (!list_empty(head)) { 428 struct reserve_ticket *ticket; 429 u64 used = btrfs_space_info_used(space_info, true); 430 431 ticket = list_first_entry(head, struct reserve_ticket, list); 432 433 /* Check and see if our ticket can be satisfied now. */ 434 if ((used + ticket->bytes <= space_info->total_bytes) || 435 btrfs_can_overcommit(fs_info, space_info, ticket->bytes, 436 flush)) { 437 btrfs_space_info_update_bytes_may_use(fs_info, 438 space_info, 439 ticket->bytes); 440 remove_ticket(space_info, ticket); 441 ticket->bytes = 0; 442 space_info->tickets_id++; 443 wake_up(&ticket->wait); 444 } else { 445 break; 446 } 447 } 448 449 if (head == &space_info->priority_tickets) { 450 head = &space_info->tickets; 451 flush = BTRFS_RESERVE_FLUSH_ALL; 452 goto again; 453 } 454 } 455 456 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \ 457 do { \ 458 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \ 459 spin_lock(&__rsv->lock); \ 460 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \ 461 __rsv->size, __rsv->reserved); \ 462 spin_unlock(&__rsv->lock); \ 463 } while (0) 464 465 static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info) 466 { 467 switch (space_info->flags) { 468 case BTRFS_BLOCK_GROUP_SYSTEM: 469 return "SYSTEM"; 470 case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA: 471 return "DATA+METADATA"; 472 case BTRFS_BLOCK_GROUP_DATA: 473 return "DATA"; 474 case BTRFS_BLOCK_GROUP_METADATA: 475 return "METADATA"; 476 default: 477 return "UNKNOWN"; 478 } 479 } 480 481 static void dump_global_block_rsv(struct btrfs_fs_info *fs_info) 482 { 483 DUMP_BLOCK_RSV(fs_info, global_block_rsv); 484 DUMP_BLOCK_RSV(fs_info, trans_block_rsv); 485 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv); 486 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv); 487 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv); 488 } 489 490 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 491 struct btrfs_space_info *info) 492 { 493 const char *flag_str = space_info_flag_to_str(info); 494 lockdep_assert_held(&info->lock); 495 496 /* The free space could be negative in case of overcommit */ 497 btrfs_info(fs_info, "space_info %s has %lld free, is %sfull", 498 flag_str, 499 (s64)(info->total_bytes - btrfs_space_info_used(info, true)), 500 info->full ? "" : "not "); 501 btrfs_info(fs_info, 502 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu", 503 info->total_bytes, info->bytes_used, info->bytes_pinned, 504 info->bytes_reserved, info->bytes_may_use, 505 info->bytes_readonly, info->bytes_zone_unusable); 506 } 507 508 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 509 struct btrfs_space_info *info, u64 bytes, 510 int dump_block_groups) 511 { 512 struct btrfs_block_group *cache; 513 int index = 0; 514 515 spin_lock(&info->lock); 516 __btrfs_dump_space_info(fs_info, info); 517 dump_global_block_rsv(fs_info); 518 spin_unlock(&info->lock); 519 520 if (!dump_block_groups) 521 return; 522 523 down_read(&info->groups_sem); 524 again: 525 list_for_each_entry(cache, &info->block_groups[index], list) { 526 u64 avail; 527 528 spin_lock(&cache->lock); 529 avail = cache->length - cache->used - cache->pinned - 530 cache->reserved - cache->delalloc_bytes - 531 cache->bytes_super - cache->zone_unusable; 532 btrfs_info(fs_info, 533 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu delalloc %llu super %llu zone_unusable (%llu bytes available) %s", 534 cache->start, cache->length, cache->used, cache->pinned, 535 cache->reserved, cache->delalloc_bytes, 536 cache->bytes_super, cache->zone_unusable, 537 avail, cache->ro ? "[readonly]" : ""); 538 spin_unlock(&cache->lock); 539 btrfs_dump_free_space(cache, bytes); 540 } 541 if (++index < BTRFS_NR_RAID_TYPES) 542 goto again; 543 up_read(&info->groups_sem); 544 } 545 546 static inline u64 calc_reclaim_items_nr(const struct btrfs_fs_info *fs_info, 547 u64 to_reclaim) 548 { 549 u64 bytes; 550 u64 nr; 551 552 bytes = btrfs_calc_insert_metadata_size(fs_info, 1); 553 nr = div64_u64(to_reclaim, bytes); 554 if (!nr) 555 nr = 1; 556 return nr; 557 } 558 559 static inline u64 calc_delayed_refs_nr(const struct btrfs_fs_info *fs_info, 560 u64 to_reclaim) 561 { 562 const u64 bytes = btrfs_calc_delayed_ref_bytes(fs_info, 1); 563 u64 nr; 564 565 nr = div64_u64(to_reclaim, bytes); 566 if (!nr) 567 nr = 1; 568 return nr; 569 } 570 571 #define EXTENT_SIZE_PER_ITEM SZ_256K 572 573 /* 574 * shrink metadata reservation for delalloc 575 */ 576 static void shrink_delalloc(struct btrfs_fs_info *fs_info, 577 struct btrfs_space_info *space_info, 578 u64 to_reclaim, bool wait_ordered, 579 bool for_preempt) 580 { 581 struct btrfs_trans_handle *trans; 582 u64 delalloc_bytes; 583 u64 ordered_bytes; 584 u64 items; 585 long time_left; 586 int loops; 587 588 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 589 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes); 590 if (delalloc_bytes == 0 && ordered_bytes == 0) 591 return; 592 593 /* Calc the number of the pages we need flush for space reservation */ 594 if (to_reclaim == U64_MAX) { 595 items = U64_MAX; 596 } else { 597 /* 598 * to_reclaim is set to however much metadata we need to 599 * reclaim, but reclaiming that much data doesn't really track 600 * exactly. What we really want to do is reclaim full inode's 601 * worth of reservations, however that's not available to us 602 * here. We will take a fraction of the delalloc bytes for our 603 * flushing loops and hope for the best. Delalloc will expand 604 * the amount we write to cover an entire dirty extent, which 605 * will reclaim the metadata reservation for that range. If 606 * it's not enough subsequent flush stages will be more 607 * aggressive. 608 */ 609 to_reclaim = max(to_reclaim, delalloc_bytes >> 3); 610 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2; 611 } 612 613 trans = current->journal_info; 614 615 /* 616 * If we are doing more ordered than delalloc we need to just wait on 617 * ordered extents, otherwise we'll waste time trying to flush delalloc 618 * that likely won't give us the space back we need. 619 */ 620 if (ordered_bytes > delalloc_bytes && !for_preempt) 621 wait_ordered = true; 622 623 loops = 0; 624 while ((delalloc_bytes || ordered_bytes) && loops < 3) { 625 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT; 626 long nr_pages = min_t(u64, temp, LONG_MAX); 627 int async_pages; 628 629 btrfs_start_delalloc_roots(fs_info, nr_pages, true); 630 631 /* 632 * We need to make sure any outstanding async pages are now 633 * processed before we continue. This is because things like 634 * sync_inode() try to be smart and skip writing if the inode is 635 * marked clean. We don't use filemap_fwrite for flushing 636 * because we want to control how many pages we write out at a 637 * time, thus this is the only safe way to make sure we've 638 * waited for outstanding compressed workers to have started 639 * their jobs and thus have ordered extents set up properly. 640 * 641 * This exists because we do not want to wait for each 642 * individual inode to finish its async work, we simply want to 643 * start the IO on everybody, and then come back here and wait 644 * for all of the async work to catch up. Once we're done with 645 * that we know we'll have ordered extents for everything and we 646 * can decide if we wait for that or not. 647 * 648 * If we choose to replace this in the future, make absolutely 649 * sure that the proper waiting is being done in the async case, 650 * as there have been bugs in that area before. 651 */ 652 async_pages = atomic_read(&fs_info->async_delalloc_pages); 653 if (!async_pages) 654 goto skip_async; 655 656 /* 657 * We don't want to wait forever, if we wrote less pages in this 658 * loop than we have outstanding, only wait for that number of 659 * pages, otherwise we can wait for all async pages to finish 660 * before continuing. 661 */ 662 if (async_pages > nr_pages) 663 async_pages -= nr_pages; 664 else 665 async_pages = 0; 666 wait_event(fs_info->async_submit_wait, 667 atomic_read(&fs_info->async_delalloc_pages) <= 668 async_pages); 669 skip_async: 670 loops++; 671 if (wait_ordered && !trans) { 672 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); 673 } else { 674 time_left = schedule_timeout_killable(1); 675 if (time_left) 676 break; 677 } 678 679 /* 680 * If we are for preemption we just want a one-shot of delalloc 681 * flushing so we can stop flushing if we decide we don't need 682 * to anymore. 683 */ 684 if (for_preempt) 685 break; 686 687 spin_lock(&space_info->lock); 688 if (list_empty(&space_info->tickets) && 689 list_empty(&space_info->priority_tickets)) { 690 spin_unlock(&space_info->lock); 691 break; 692 } 693 spin_unlock(&space_info->lock); 694 695 delalloc_bytes = percpu_counter_sum_positive( 696 &fs_info->delalloc_bytes); 697 ordered_bytes = percpu_counter_sum_positive( 698 &fs_info->ordered_bytes); 699 } 700 } 701 702 /* 703 * Try to flush some data based on policy set by @state. This is only advisory 704 * and may fail for various reasons. The caller is supposed to examine the 705 * state of @space_info to detect the outcome. 706 */ 707 static void flush_space(struct btrfs_fs_info *fs_info, 708 struct btrfs_space_info *space_info, u64 num_bytes, 709 enum btrfs_flush_state state, bool for_preempt) 710 { 711 struct btrfs_root *root = fs_info->tree_root; 712 struct btrfs_trans_handle *trans; 713 int nr; 714 int ret = 0; 715 716 switch (state) { 717 case FLUSH_DELAYED_ITEMS_NR: 718 case FLUSH_DELAYED_ITEMS: 719 if (state == FLUSH_DELAYED_ITEMS_NR) 720 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2; 721 else 722 nr = -1; 723 724 trans = btrfs_join_transaction(root); 725 if (IS_ERR(trans)) { 726 ret = PTR_ERR(trans); 727 break; 728 } 729 ret = btrfs_run_delayed_items_nr(trans, nr); 730 btrfs_end_transaction(trans); 731 break; 732 case FLUSH_DELALLOC: 733 case FLUSH_DELALLOC_WAIT: 734 case FLUSH_DELALLOC_FULL: 735 if (state == FLUSH_DELALLOC_FULL) 736 num_bytes = U64_MAX; 737 shrink_delalloc(fs_info, space_info, num_bytes, 738 state != FLUSH_DELALLOC, for_preempt); 739 break; 740 case FLUSH_DELAYED_REFS_NR: 741 case FLUSH_DELAYED_REFS: 742 trans = btrfs_join_transaction(root); 743 if (IS_ERR(trans)) { 744 ret = PTR_ERR(trans); 745 break; 746 } 747 if (state == FLUSH_DELAYED_REFS_NR) 748 nr = calc_delayed_refs_nr(fs_info, num_bytes); 749 else 750 nr = 0; 751 btrfs_run_delayed_refs(trans, nr); 752 btrfs_end_transaction(trans); 753 break; 754 case ALLOC_CHUNK: 755 case ALLOC_CHUNK_FORCE: 756 /* 757 * For metadata space on zoned filesystem, reaching here means we 758 * don't have enough space left in active_total_bytes. Try to 759 * activate a block group first, because we may have inactive 760 * block group already allocated. 761 */ 762 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, false); 763 if (ret < 0) 764 break; 765 else if (ret == 1) 766 break; 767 768 trans = btrfs_join_transaction(root); 769 if (IS_ERR(trans)) { 770 ret = PTR_ERR(trans); 771 break; 772 } 773 ret = btrfs_chunk_alloc(trans, 774 btrfs_get_alloc_profile(fs_info, space_info->flags), 775 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE : 776 CHUNK_ALLOC_FORCE); 777 btrfs_end_transaction(trans); 778 779 /* 780 * For metadata space on zoned filesystem, allocating a new chunk 781 * is not enough. We still need to activate the block * group. 782 * Active the newly allocated block group by (maybe) finishing 783 * a block group. 784 */ 785 if (ret == 1) { 786 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, true); 787 /* 788 * Revert to the original ret regardless we could finish 789 * one block group or not. 790 */ 791 if (ret >= 0) 792 ret = 1; 793 } 794 795 if (ret > 0 || ret == -ENOSPC) 796 ret = 0; 797 break; 798 case RUN_DELAYED_IPUTS: 799 /* 800 * If we have pending delayed iputs then we could free up a 801 * bunch of pinned space, so make sure we run the iputs before 802 * we do our pinned bytes check below. 803 */ 804 btrfs_run_delayed_iputs(fs_info); 805 btrfs_wait_on_delayed_iputs(fs_info); 806 break; 807 case COMMIT_TRANS: 808 ASSERT(current->journal_info == NULL); 809 trans = btrfs_join_transaction(root); 810 if (IS_ERR(trans)) { 811 ret = PTR_ERR(trans); 812 break; 813 } 814 ret = btrfs_commit_transaction(trans); 815 break; 816 default: 817 ret = -ENOSPC; 818 break; 819 } 820 821 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state, 822 ret, for_preempt); 823 return; 824 } 825 826 static inline u64 827 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info, 828 struct btrfs_space_info *space_info) 829 { 830 u64 used; 831 u64 avail; 832 u64 to_reclaim = space_info->reclaim_size; 833 834 lockdep_assert_held(&space_info->lock); 835 836 avail = calc_available_free_space(fs_info, space_info, 837 BTRFS_RESERVE_FLUSH_ALL); 838 used = btrfs_space_info_used(space_info, true); 839 840 /* 841 * We may be flushing because suddenly we have less space than we had 842 * before, and now we're well over-committed based on our current free 843 * space. If that's the case add in our overage so we make sure to put 844 * appropriate pressure on the flushing state machine. 845 */ 846 if (space_info->total_bytes + avail < used) 847 to_reclaim += used - (space_info->total_bytes + avail); 848 849 return to_reclaim; 850 } 851 852 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info, 853 struct btrfs_space_info *space_info) 854 { 855 u64 global_rsv_size = fs_info->global_block_rsv.reserved; 856 u64 ordered, delalloc; 857 u64 thresh; 858 u64 used; 859 860 thresh = mult_perc(space_info->total_bytes, 90); 861 862 lockdep_assert_held(&space_info->lock); 863 864 /* If we're just plain full then async reclaim just slows us down. */ 865 if ((space_info->bytes_used + space_info->bytes_reserved + 866 global_rsv_size) >= thresh) 867 return false; 868 869 used = space_info->bytes_may_use + space_info->bytes_pinned; 870 871 /* The total flushable belongs to the global rsv, don't flush. */ 872 if (global_rsv_size >= used) 873 return false; 874 875 /* 876 * 128MiB is 1/4 of the maximum global rsv size. If we have less than 877 * that devoted to other reservations then there's no sense in flushing, 878 * we don't have a lot of things that need flushing. 879 */ 880 if (used - global_rsv_size <= SZ_128M) 881 return false; 882 883 /* 884 * We have tickets queued, bail so we don't compete with the async 885 * flushers. 886 */ 887 if (space_info->reclaim_size) 888 return false; 889 890 /* 891 * If we have over half of the free space occupied by reservations or 892 * pinned then we want to start flushing. 893 * 894 * We do not do the traditional thing here, which is to say 895 * 896 * if (used >= ((total_bytes + avail) / 2)) 897 * return 1; 898 * 899 * because this doesn't quite work how we want. If we had more than 50% 900 * of the space_info used by bytes_used and we had 0 available we'd just 901 * constantly run the background flusher. Instead we want it to kick in 902 * if our reclaimable space exceeds our clamped free space. 903 * 904 * Our clamping range is 2^1 -> 2^8. Practically speaking that means 905 * the following: 906 * 907 * Amount of RAM Minimum threshold Maximum threshold 908 * 909 * 256GiB 1GiB 128GiB 910 * 128GiB 512MiB 64GiB 911 * 64GiB 256MiB 32GiB 912 * 32GiB 128MiB 16GiB 913 * 16GiB 64MiB 8GiB 914 * 915 * These are the range our thresholds will fall in, corresponding to how 916 * much delalloc we need for the background flusher to kick in. 917 */ 918 919 thresh = calc_available_free_space(fs_info, space_info, 920 BTRFS_RESERVE_FLUSH_ALL); 921 used = space_info->bytes_used + space_info->bytes_reserved + 922 space_info->bytes_readonly + global_rsv_size; 923 if (used < space_info->total_bytes) 924 thresh += space_info->total_bytes - used; 925 thresh >>= space_info->clamp; 926 927 used = space_info->bytes_pinned; 928 929 /* 930 * If we have more ordered bytes than delalloc bytes then we're either 931 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting 932 * around. Preemptive flushing is only useful in that it can free up 933 * space before tickets need to wait for things to finish. In the case 934 * of ordered extents, preemptively waiting on ordered extents gets us 935 * nothing, if our reservations are tied up in ordered extents we'll 936 * simply have to slow down writers by forcing them to wait on ordered 937 * extents. 938 * 939 * In the case that ordered is larger than delalloc, only include the 940 * block reserves that we would actually be able to directly reclaim 941 * from. In this case if we're heavy on metadata operations this will 942 * clearly be heavy enough to warrant preemptive flushing. In the case 943 * of heavy DIO or ordered reservations, preemptive flushing will just 944 * waste time and cause us to slow down. 945 * 946 * We want to make sure we truly are maxed out on ordered however, so 947 * cut ordered in half, and if it's still higher than delalloc then we 948 * can keep flushing. This is to avoid the case where we start 949 * flushing, and now delalloc == ordered and we stop preemptively 950 * flushing when we could still have several gigs of delalloc to flush. 951 */ 952 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1; 953 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes); 954 if (ordered >= delalloc) 955 used += fs_info->delayed_refs_rsv.reserved + 956 fs_info->delayed_block_rsv.reserved; 957 else 958 used += space_info->bytes_may_use - global_rsv_size; 959 960 return (used >= thresh && !btrfs_fs_closing(fs_info) && 961 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state)); 962 } 963 964 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info, 965 struct btrfs_space_info *space_info, 966 struct reserve_ticket *ticket) 967 { 968 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 969 u64 min_bytes; 970 971 if (!ticket->steal) 972 return false; 973 974 if (global_rsv->space_info != space_info) 975 return false; 976 977 spin_lock(&global_rsv->lock); 978 min_bytes = mult_perc(global_rsv->size, 10); 979 if (global_rsv->reserved < min_bytes + ticket->bytes) { 980 spin_unlock(&global_rsv->lock); 981 return false; 982 } 983 global_rsv->reserved -= ticket->bytes; 984 remove_ticket(space_info, ticket); 985 ticket->bytes = 0; 986 wake_up(&ticket->wait); 987 space_info->tickets_id++; 988 if (global_rsv->reserved < global_rsv->size) 989 global_rsv->full = 0; 990 spin_unlock(&global_rsv->lock); 991 992 return true; 993 } 994 995 /* 996 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets 997 * @fs_info - fs_info for this fs 998 * @space_info - the space info we were flushing 999 * 1000 * We call this when we've exhausted our flushing ability and haven't made 1001 * progress in satisfying tickets. The reservation code handles tickets in 1002 * order, so if there is a large ticket first and then smaller ones we could 1003 * very well satisfy the smaller tickets. This will attempt to wake up any 1004 * tickets in the list to catch this case. 1005 * 1006 * This function returns true if it was able to make progress by clearing out 1007 * other tickets, or if it stumbles across a ticket that was smaller than the 1008 * first ticket. 1009 */ 1010 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info, 1011 struct btrfs_space_info *space_info) 1012 { 1013 struct reserve_ticket *ticket; 1014 u64 tickets_id = space_info->tickets_id; 1015 const bool aborted = BTRFS_FS_ERROR(fs_info); 1016 1017 trace_btrfs_fail_all_tickets(fs_info, space_info); 1018 1019 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 1020 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info"); 1021 __btrfs_dump_space_info(fs_info, space_info); 1022 } 1023 1024 while (!list_empty(&space_info->tickets) && 1025 tickets_id == space_info->tickets_id) { 1026 ticket = list_first_entry(&space_info->tickets, 1027 struct reserve_ticket, list); 1028 1029 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket)) 1030 return true; 1031 1032 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1033 btrfs_info(fs_info, "failing ticket with %llu bytes", 1034 ticket->bytes); 1035 1036 remove_ticket(space_info, ticket); 1037 if (aborted) 1038 ticket->error = -EIO; 1039 else 1040 ticket->error = -ENOSPC; 1041 wake_up(&ticket->wait); 1042 1043 /* 1044 * We're just throwing tickets away, so more flushing may not 1045 * trip over btrfs_try_granting_tickets, so we need to call it 1046 * here to see if we can make progress with the next ticket in 1047 * the list. 1048 */ 1049 if (!aborted) 1050 btrfs_try_granting_tickets(fs_info, space_info); 1051 } 1052 return (tickets_id != space_info->tickets_id); 1053 } 1054 1055 /* 1056 * This is for normal flushers, we can wait all goddamned day if we want to. We 1057 * will loop and continuously try to flush as long as we are making progress. 1058 * We count progress as clearing off tickets each time we have to loop. 1059 */ 1060 static void btrfs_async_reclaim_metadata_space(struct work_struct *work) 1061 { 1062 struct btrfs_fs_info *fs_info; 1063 struct btrfs_space_info *space_info; 1064 u64 to_reclaim; 1065 enum btrfs_flush_state flush_state; 1066 int commit_cycles = 0; 1067 u64 last_tickets_id; 1068 1069 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work); 1070 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 1071 1072 spin_lock(&space_info->lock); 1073 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1074 if (!to_reclaim) { 1075 space_info->flush = 0; 1076 spin_unlock(&space_info->lock); 1077 return; 1078 } 1079 last_tickets_id = space_info->tickets_id; 1080 spin_unlock(&space_info->lock); 1081 1082 flush_state = FLUSH_DELAYED_ITEMS_NR; 1083 do { 1084 flush_space(fs_info, space_info, to_reclaim, flush_state, false); 1085 spin_lock(&space_info->lock); 1086 if (list_empty(&space_info->tickets)) { 1087 space_info->flush = 0; 1088 spin_unlock(&space_info->lock); 1089 return; 1090 } 1091 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, 1092 space_info); 1093 if (last_tickets_id == space_info->tickets_id) { 1094 flush_state++; 1095 } else { 1096 last_tickets_id = space_info->tickets_id; 1097 flush_state = FLUSH_DELAYED_ITEMS_NR; 1098 if (commit_cycles) 1099 commit_cycles--; 1100 } 1101 1102 /* 1103 * We do not want to empty the system of delalloc unless we're 1104 * under heavy pressure, so allow one trip through the flushing 1105 * logic before we start doing a FLUSH_DELALLOC_FULL. 1106 */ 1107 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles) 1108 flush_state++; 1109 1110 /* 1111 * We don't want to force a chunk allocation until we've tried 1112 * pretty hard to reclaim space. Think of the case where we 1113 * freed up a bunch of space and so have a lot of pinned space 1114 * to reclaim. We would rather use that than possibly create a 1115 * underutilized metadata chunk. So if this is our first run 1116 * through the flushing state machine skip ALLOC_CHUNK_FORCE and 1117 * commit the transaction. If nothing has changed the next go 1118 * around then we can force a chunk allocation. 1119 */ 1120 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles) 1121 flush_state++; 1122 1123 if (flush_state > COMMIT_TRANS) { 1124 commit_cycles++; 1125 if (commit_cycles > 2) { 1126 if (maybe_fail_all_tickets(fs_info, space_info)) { 1127 flush_state = FLUSH_DELAYED_ITEMS_NR; 1128 commit_cycles--; 1129 } else { 1130 space_info->flush = 0; 1131 } 1132 } else { 1133 flush_state = FLUSH_DELAYED_ITEMS_NR; 1134 } 1135 } 1136 spin_unlock(&space_info->lock); 1137 } while (flush_state <= COMMIT_TRANS); 1138 } 1139 1140 /* 1141 * This handles pre-flushing of metadata space before we get to the point that 1142 * we need to start blocking threads on tickets. The logic here is different 1143 * from the other flush paths because it doesn't rely on tickets to tell us how 1144 * much we need to flush, instead it attempts to keep us below the 80% full 1145 * watermark of space by flushing whichever reservation pool is currently the 1146 * largest. 1147 */ 1148 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work) 1149 { 1150 struct btrfs_fs_info *fs_info; 1151 struct btrfs_space_info *space_info; 1152 struct btrfs_block_rsv *delayed_block_rsv; 1153 struct btrfs_block_rsv *delayed_refs_rsv; 1154 struct btrfs_block_rsv *global_rsv; 1155 struct btrfs_block_rsv *trans_rsv; 1156 int loops = 0; 1157 1158 fs_info = container_of(work, struct btrfs_fs_info, 1159 preempt_reclaim_work); 1160 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 1161 delayed_block_rsv = &fs_info->delayed_block_rsv; 1162 delayed_refs_rsv = &fs_info->delayed_refs_rsv; 1163 global_rsv = &fs_info->global_block_rsv; 1164 trans_rsv = &fs_info->trans_block_rsv; 1165 1166 spin_lock(&space_info->lock); 1167 while (need_preemptive_reclaim(fs_info, space_info)) { 1168 enum btrfs_flush_state flush; 1169 u64 delalloc_size = 0; 1170 u64 to_reclaim, block_rsv_size; 1171 u64 global_rsv_size = global_rsv->reserved; 1172 1173 loops++; 1174 1175 /* 1176 * We don't have a precise counter for the metadata being 1177 * reserved for delalloc, so we'll approximate it by subtracting 1178 * out the block rsv's space from the bytes_may_use. If that 1179 * amount is higher than the individual reserves, then we can 1180 * assume it's tied up in delalloc reservations. 1181 */ 1182 block_rsv_size = global_rsv_size + 1183 delayed_block_rsv->reserved + 1184 delayed_refs_rsv->reserved + 1185 trans_rsv->reserved; 1186 if (block_rsv_size < space_info->bytes_may_use) 1187 delalloc_size = space_info->bytes_may_use - block_rsv_size; 1188 1189 /* 1190 * We don't want to include the global_rsv in our calculation, 1191 * because that's space we can't touch. Subtract it from the 1192 * block_rsv_size for the next checks. 1193 */ 1194 block_rsv_size -= global_rsv_size; 1195 1196 /* 1197 * We really want to avoid flushing delalloc too much, as it 1198 * could result in poor allocation patterns, so only flush it if 1199 * it's larger than the rest of the pools combined. 1200 */ 1201 if (delalloc_size > block_rsv_size) { 1202 to_reclaim = delalloc_size; 1203 flush = FLUSH_DELALLOC; 1204 } else if (space_info->bytes_pinned > 1205 (delayed_block_rsv->reserved + 1206 delayed_refs_rsv->reserved)) { 1207 to_reclaim = space_info->bytes_pinned; 1208 flush = COMMIT_TRANS; 1209 } else if (delayed_block_rsv->reserved > 1210 delayed_refs_rsv->reserved) { 1211 to_reclaim = delayed_block_rsv->reserved; 1212 flush = FLUSH_DELAYED_ITEMS_NR; 1213 } else { 1214 to_reclaim = delayed_refs_rsv->reserved; 1215 flush = FLUSH_DELAYED_REFS_NR; 1216 } 1217 1218 spin_unlock(&space_info->lock); 1219 1220 /* 1221 * We don't want to reclaim everything, just a portion, so scale 1222 * down the to_reclaim by 1/4. If it takes us down to 0, 1223 * reclaim 1 items worth. 1224 */ 1225 to_reclaim >>= 2; 1226 if (!to_reclaim) 1227 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1); 1228 flush_space(fs_info, space_info, to_reclaim, flush, true); 1229 cond_resched(); 1230 spin_lock(&space_info->lock); 1231 } 1232 1233 /* We only went through once, back off our clamping. */ 1234 if (loops == 1 && !space_info->reclaim_size) 1235 space_info->clamp = max(1, space_info->clamp - 1); 1236 trace_btrfs_done_preemptive_reclaim(fs_info, space_info); 1237 spin_unlock(&space_info->lock); 1238 } 1239 1240 /* 1241 * FLUSH_DELALLOC_WAIT: 1242 * Space is freed from flushing delalloc in one of two ways. 1243 * 1244 * 1) compression is on and we allocate less space than we reserved 1245 * 2) we are overwriting existing space 1246 * 1247 * For #1 that extra space is reclaimed as soon as the delalloc pages are 1248 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent 1249 * length to ->bytes_reserved, and subtracts the reserved space from 1250 * ->bytes_may_use. 1251 * 1252 * For #2 this is trickier. Once the ordered extent runs we will drop the 1253 * extent in the range we are overwriting, which creates a delayed ref for 1254 * that freed extent. This however is not reclaimed until the transaction 1255 * commits, thus the next stages. 1256 * 1257 * RUN_DELAYED_IPUTS 1258 * If we are freeing inodes, we want to make sure all delayed iputs have 1259 * completed, because they could have been on an inode with i_nlink == 0, and 1260 * thus have been truncated and freed up space. But again this space is not 1261 * immediately re-usable, it comes in the form of a delayed ref, which must be 1262 * run and then the transaction must be committed. 1263 * 1264 * COMMIT_TRANS 1265 * This is where we reclaim all of the pinned space generated by running the 1266 * iputs 1267 * 1268 * ALLOC_CHUNK_FORCE 1269 * For data we start with alloc chunk force, however we could have been full 1270 * before, and then the transaction commit could have freed new block groups, 1271 * so if we now have space to allocate do the force chunk allocation. 1272 */ 1273 static const enum btrfs_flush_state data_flush_states[] = { 1274 FLUSH_DELALLOC_FULL, 1275 RUN_DELAYED_IPUTS, 1276 COMMIT_TRANS, 1277 ALLOC_CHUNK_FORCE, 1278 }; 1279 1280 static void btrfs_async_reclaim_data_space(struct work_struct *work) 1281 { 1282 struct btrfs_fs_info *fs_info; 1283 struct btrfs_space_info *space_info; 1284 u64 last_tickets_id; 1285 enum btrfs_flush_state flush_state = 0; 1286 1287 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work); 1288 space_info = fs_info->data_sinfo; 1289 1290 spin_lock(&space_info->lock); 1291 if (list_empty(&space_info->tickets)) { 1292 space_info->flush = 0; 1293 spin_unlock(&space_info->lock); 1294 return; 1295 } 1296 last_tickets_id = space_info->tickets_id; 1297 spin_unlock(&space_info->lock); 1298 1299 while (!space_info->full) { 1300 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1301 spin_lock(&space_info->lock); 1302 if (list_empty(&space_info->tickets)) { 1303 space_info->flush = 0; 1304 spin_unlock(&space_info->lock); 1305 return; 1306 } 1307 1308 /* Something happened, fail everything and bail. */ 1309 if (BTRFS_FS_ERROR(fs_info)) 1310 goto aborted_fs; 1311 last_tickets_id = space_info->tickets_id; 1312 spin_unlock(&space_info->lock); 1313 } 1314 1315 while (flush_state < ARRAY_SIZE(data_flush_states)) { 1316 flush_space(fs_info, space_info, U64_MAX, 1317 data_flush_states[flush_state], false); 1318 spin_lock(&space_info->lock); 1319 if (list_empty(&space_info->tickets)) { 1320 space_info->flush = 0; 1321 spin_unlock(&space_info->lock); 1322 return; 1323 } 1324 1325 if (last_tickets_id == space_info->tickets_id) { 1326 flush_state++; 1327 } else { 1328 last_tickets_id = space_info->tickets_id; 1329 flush_state = 0; 1330 } 1331 1332 if (flush_state >= ARRAY_SIZE(data_flush_states)) { 1333 if (space_info->full) { 1334 if (maybe_fail_all_tickets(fs_info, space_info)) 1335 flush_state = 0; 1336 else 1337 space_info->flush = 0; 1338 } else { 1339 flush_state = 0; 1340 } 1341 1342 /* Something happened, fail everything and bail. */ 1343 if (BTRFS_FS_ERROR(fs_info)) 1344 goto aborted_fs; 1345 1346 } 1347 spin_unlock(&space_info->lock); 1348 } 1349 return; 1350 1351 aborted_fs: 1352 maybe_fail_all_tickets(fs_info, space_info); 1353 space_info->flush = 0; 1354 spin_unlock(&space_info->lock); 1355 } 1356 1357 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info) 1358 { 1359 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space); 1360 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space); 1361 INIT_WORK(&fs_info->preempt_reclaim_work, 1362 btrfs_preempt_reclaim_metadata_space); 1363 } 1364 1365 static const enum btrfs_flush_state priority_flush_states[] = { 1366 FLUSH_DELAYED_ITEMS_NR, 1367 FLUSH_DELAYED_ITEMS, 1368 ALLOC_CHUNK, 1369 }; 1370 1371 static const enum btrfs_flush_state evict_flush_states[] = { 1372 FLUSH_DELAYED_ITEMS_NR, 1373 FLUSH_DELAYED_ITEMS, 1374 FLUSH_DELAYED_REFS_NR, 1375 FLUSH_DELAYED_REFS, 1376 FLUSH_DELALLOC, 1377 FLUSH_DELALLOC_WAIT, 1378 FLUSH_DELALLOC_FULL, 1379 ALLOC_CHUNK, 1380 COMMIT_TRANS, 1381 }; 1382 1383 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info, 1384 struct btrfs_space_info *space_info, 1385 struct reserve_ticket *ticket, 1386 const enum btrfs_flush_state *states, 1387 int states_nr) 1388 { 1389 u64 to_reclaim; 1390 int flush_state = 0; 1391 1392 spin_lock(&space_info->lock); 1393 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1394 /* 1395 * This is the priority reclaim path, so to_reclaim could be >0 still 1396 * because we may have only satisfied the priority tickets and still 1397 * left non priority tickets on the list. We would then have 1398 * to_reclaim but ->bytes == 0. 1399 */ 1400 if (ticket->bytes == 0) { 1401 spin_unlock(&space_info->lock); 1402 return; 1403 } 1404 1405 while (flush_state < states_nr) { 1406 spin_unlock(&space_info->lock); 1407 flush_space(fs_info, space_info, to_reclaim, states[flush_state], 1408 false); 1409 flush_state++; 1410 spin_lock(&space_info->lock); 1411 if (ticket->bytes == 0) { 1412 spin_unlock(&space_info->lock); 1413 return; 1414 } 1415 } 1416 1417 /* Attempt to steal from the global rsv if we can. */ 1418 if (!steal_from_global_rsv(fs_info, space_info, ticket)) { 1419 ticket->error = -ENOSPC; 1420 remove_ticket(space_info, ticket); 1421 } 1422 1423 /* 1424 * We must run try_granting_tickets here because we could be a large 1425 * ticket in front of a smaller ticket that can now be satisfied with 1426 * the available space. 1427 */ 1428 btrfs_try_granting_tickets(fs_info, space_info); 1429 spin_unlock(&space_info->lock); 1430 } 1431 1432 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info, 1433 struct btrfs_space_info *space_info, 1434 struct reserve_ticket *ticket) 1435 { 1436 spin_lock(&space_info->lock); 1437 1438 /* We could have been granted before we got here. */ 1439 if (ticket->bytes == 0) { 1440 spin_unlock(&space_info->lock); 1441 return; 1442 } 1443 1444 while (!space_info->full) { 1445 spin_unlock(&space_info->lock); 1446 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1447 spin_lock(&space_info->lock); 1448 if (ticket->bytes == 0) { 1449 spin_unlock(&space_info->lock); 1450 return; 1451 } 1452 } 1453 1454 ticket->error = -ENOSPC; 1455 remove_ticket(space_info, ticket); 1456 btrfs_try_granting_tickets(fs_info, space_info); 1457 spin_unlock(&space_info->lock); 1458 } 1459 1460 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info, 1461 struct btrfs_space_info *space_info, 1462 struct reserve_ticket *ticket) 1463 1464 { 1465 DEFINE_WAIT(wait); 1466 int ret = 0; 1467 1468 spin_lock(&space_info->lock); 1469 while (ticket->bytes > 0 && ticket->error == 0) { 1470 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE); 1471 if (ret) { 1472 /* 1473 * Delete us from the list. After we unlock the space 1474 * info, we don't want the async reclaim job to reserve 1475 * space for this ticket. If that would happen, then the 1476 * ticket's task would not known that space was reserved 1477 * despite getting an error, resulting in a space leak 1478 * (bytes_may_use counter of our space_info). 1479 */ 1480 remove_ticket(space_info, ticket); 1481 ticket->error = -EINTR; 1482 break; 1483 } 1484 spin_unlock(&space_info->lock); 1485 1486 schedule(); 1487 1488 finish_wait(&ticket->wait, &wait); 1489 spin_lock(&space_info->lock); 1490 } 1491 spin_unlock(&space_info->lock); 1492 } 1493 1494 /* 1495 * Do the appropriate flushing and waiting for a ticket. 1496 * 1497 * @fs_info: the filesystem 1498 * @space_info: space info for the reservation 1499 * @ticket: ticket for the reservation 1500 * @start_ns: timestamp when the reservation started 1501 * @orig_bytes: amount of bytes originally reserved 1502 * @flush: how much we can flush 1503 * 1504 * This does the work of figuring out how to flush for the ticket, waiting for 1505 * the reservation, and returning the appropriate error if there is one. 1506 */ 1507 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info, 1508 struct btrfs_space_info *space_info, 1509 struct reserve_ticket *ticket, 1510 u64 start_ns, u64 orig_bytes, 1511 enum btrfs_reserve_flush_enum flush) 1512 { 1513 int ret; 1514 1515 switch (flush) { 1516 case BTRFS_RESERVE_FLUSH_DATA: 1517 case BTRFS_RESERVE_FLUSH_ALL: 1518 case BTRFS_RESERVE_FLUSH_ALL_STEAL: 1519 wait_reserve_ticket(fs_info, space_info, ticket); 1520 break; 1521 case BTRFS_RESERVE_FLUSH_LIMIT: 1522 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1523 priority_flush_states, 1524 ARRAY_SIZE(priority_flush_states)); 1525 break; 1526 case BTRFS_RESERVE_FLUSH_EVICT: 1527 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1528 evict_flush_states, 1529 ARRAY_SIZE(evict_flush_states)); 1530 break; 1531 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE: 1532 priority_reclaim_data_space(fs_info, space_info, ticket); 1533 break; 1534 default: 1535 ASSERT(0); 1536 break; 1537 } 1538 1539 ret = ticket->error; 1540 ASSERT(list_empty(&ticket->list)); 1541 /* 1542 * Check that we can't have an error set if the reservation succeeded, 1543 * as that would confuse tasks and lead them to error out without 1544 * releasing reserved space (if an error happens the expectation is that 1545 * space wasn't reserved at all). 1546 */ 1547 ASSERT(!(ticket->bytes == 0 && ticket->error)); 1548 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes, 1549 start_ns, flush, ticket->error); 1550 return ret; 1551 } 1552 1553 /* 1554 * This returns true if this flush state will go through the ordinary flushing 1555 * code. 1556 */ 1557 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush) 1558 { 1559 return (flush == BTRFS_RESERVE_FLUSH_ALL) || 1560 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1561 } 1562 1563 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info, 1564 struct btrfs_space_info *space_info) 1565 { 1566 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes); 1567 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 1568 1569 /* 1570 * If we're heavy on ordered operations then clamping won't help us. We 1571 * need to clamp specifically to keep up with dirty'ing buffered 1572 * writers, because there's not a 1:1 correlation of writing delalloc 1573 * and freeing space, like there is with flushing delayed refs or 1574 * delayed nodes. If we're already more ordered than delalloc then 1575 * we're keeping up, otherwise we aren't and should probably clamp. 1576 */ 1577 if (ordered < delalloc) 1578 space_info->clamp = min(space_info->clamp + 1, 8); 1579 } 1580 1581 static inline bool can_steal(enum btrfs_reserve_flush_enum flush) 1582 { 1583 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1584 flush == BTRFS_RESERVE_FLUSH_EVICT); 1585 } 1586 1587 /* 1588 * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to 1589 * fail as quickly as possible. 1590 */ 1591 static inline bool can_ticket(enum btrfs_reserve_flush_enum flush) 1592 { 1593 return (flush != BTRFS_RESERVE_NO_FLUSH && 1594 flush != BTRFS_RESERVE_FLUSH_EMERGENCY); 1595 } 1596 1597 /* 1598 * Try to reserve bytes from the block_rsv's space. 1599 * 1600 * @fs_info: the filesystem 1601 * @space_info: space info we want to allocate from 1602 * @orig_bytes: number of bytes we want 1603 * @flush: whether or not we can flush to make our reservation 1604 * 1605 * This will reserve orig_bytes number of bytes from the space info associated 1606 * with the block_rsv. If there is not enough space it will make an attempt to 1607 * flush out space to make room. It will do this by flushing delalloc if 1608 * possible or committing the transaction. If flush is 0 then no attempts to 1609 * regain reservations will be made and this will fail if there is not enough 1610 * space already. 1611 */ 1612 static int __reserve_bytes(struct btrfs_fs_info *fs_info, 1613 struct btrfs_space_info *space_info, u64 orig_bytes, 1614 enum btrfs_reserve_flush_enum flush) 1615 { 1616 struct work_struct *async_work; 1617 struct reserve_ticket ticket; 1618 u64 start_ns = 0; 1619 u64 used; 1620 int ret = -ENOSPC; 1621 bool pending_tickets; 1622 1623 ASSERT(orig_bytes); 1624 /* 1625 * If have a transaction handle (current->journal_info != NULL), then 1626 * the flush method can not be neither BTRFS_RESERVE_FLUSH_ALL* nor 1627 * BTRFS_RESERVE_FLUSH_EVICT, as we could deadlock because those 1628 * flushing methods can trigger transaction commits. 1629 */ 1630 if (current->journal_info) { 1631 /* One assert per line for easier debugging. */ 1632 ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL); 1633 ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL_STEAL); 1634 ASSERT(flush != BTRFS_RESERVE_FLUSH_EVICT); 1635 } 1636 1637 if (flush == BTRFS_RESERVE_FLUSH_DATA) 1638 async_work = &fs_info->async_data_reclaim_work; 1639 else 1640 async_work = &fs_info->async_reclaim_work; 1641 1642 spin_lock(&space_info->lock); 1643 used = btrfs_space_info_used(space_info, true); 1644 1645 /* 1646 * We don't want NO_FLUSH allocations to jump everybody, they can 1647 * generally handle ENOSPC in a different way, so treat them the same as 1648 * normal flushers when it comes to skipping pending tickets. 1649 */ 1650 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH)) 1651 pending_tickets = !list_empty(&space_info->tickets) || 1652 !list_empty(&space_info->priority_tickets); 1653 else 1654 pending_tickets = !list_empty(&space_info->priority_tickets); 1655 1656 /* 1657 * Carry on if we have enough space (short-circuit) OR call 1658 * can_overcommit() to ensure we can overcommit to continue. 1659 */ 1660 if (!pending_tickets && 1661 ((used + orig_bytes <= space_info->total_bytes) || 1662 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) { 1663 btrfs_space_info_update_bytes_may_use(fs_info, space_info, 1664 orig_bytes); 1665 ret = 0; 1666 } 1667 1668 /* 1669 * Things are dire, we need to make a reservation so we don't abort. We 1670 * will let this reservation go through as long as we have actual space 1671 * left to allocate for the block. 1672 */ 1673 if (ret && unlikely(flush == BTRFS_RESERVE_FLUSH_EMERGENCY)) { 1674 used = btrfs_space_info_used(space_info, false); 1675 if (used + orig_bytes <= space_info->total_bytes) { 1676 btrfs_space_info_update_bytes_may_use(fs_info, space_info, 1677 orig_bytes); 1678 ret = 0; 1679 } 1680 } 1681 1682 /* 1683 * If we couldn't make a reservation then setup our reservation ticket 1684 * and kick the async worker if it's not already running. 1685 * 1686 * If we are a priority flusher then we just need to add our ticket to 1687 * the list and we will do our own flushing further down. 1688 */ 1689 if (ret && can_ticket(flush)) { 1690 ticket.bytes = orig_bytes; 1691 ticket.error = 0; 1692 space_info->reclaim_size += ticket.bytes; 1693 init_waitqueue_head(&ticket.wait); 1694 ticket.steal = can_steal(flush); 1695 if (trace_btrfs_reserve_ticket_enabled()) 1696 start_ns = ktime_get_ns(); 1697 1698 if (flush == BTRFS_RESERVE_FLUSH_ALL || 1699 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1700 flush == BTRFS_RESERVE_FLUSH_DATA) { 1701 list_add_tail(&ticket.list, &space_info->tickets); 1702 if (!space_info->flush) { 1703 /* 1704 * We were forced to add a reserve ticket, so 1705 * our preemptive flushing is unable to keep 1706 * up. Clamp down on the threshold for the 1707 * preemptive flushing in order to keep up with 1708 * the workload. 1709 */ 1710 maybe_clamp_preempt(fs_info, space_info); 1711 1712 space_info->flush = 1; 1713 trace_btrfs_trigger_flush(fs_info, 1714 space_info->flags, 1715 orig_bytes, flush, 1716 "enospc"); 1717 queue_work(system_unbound_wq, async_work); 1718 } 1719 } else { 1720 list_add_tail(&ticket.list, 1721 &space_info->priority_tickets); 1722 } 1723 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { 1724 /* 1725 * We will do the space reservation dance during log replay, 1726 * which means we won't have fs_info->fs_root set, so don't do 1727 * the async reclaim as we will panic. 1728 */ 1729 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) && 1730 !work_busy(&fs_info->preempt_reclaim_work) && 1731 need_preemptive_reclaim(fs_info, space_info)) { 1732 trace_btrfs_trigger_flush(fs_info, space_info->flags, 1733 orig_bytes, flush, "preempt"); 1734 queue_work(system_unbound_wq, 1735 &fs_info->preempt_reclaim_work); 1736 } 1737 } 1738 spin_unlock(&space_info->lock); 1739 if (!ret || !can_ticket(flush)) 1740 return ret; 1741 1742 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns, 1743 orig_bytes, flush); 1744 } 1745 1746 /* 1747 * Try to reserve metadata bytes from the block_rsv's space. 1748 * 1749 * @fs_info: the filesystem 1750 * @block_rsv: block_rsv we're allocating for 1751 * @orig_bytes: number of bytes we want 1752 * @flush: whether or not we can flush to make our reservation 1753 * 1754 * This will reserve orig_bytes number of bytes from the space info associated 1755 * with the block_rsv. If there is not enough space it will make an attempt to 1756 * flush out space to make room. It will do this by flushing delalloc if 1757 * possible or committing the transaction. If flush is 0 then no attempts to 1758 * regain reservations will be made and this will fail if there is not enough 1759 * space already. 1760 */ 1761 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info, 1762 struct btrfs_block_rsv *block_rsv, 1763 u64 orig_bytes, 1764 enum btrfs_reserve_flush_enum flush) 1765 { 1766 int ret; 1767 1768 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush); 1769 if (ret == -ENOSPC) { 1770 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1771 block_rsv->space_info->flags, 1772 orig_bytes, 1); 1773 1774 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1775 btrfs_dump_space_info(fs_info, block_rsv->space_info, 1776 orig_bytes, 0); 1777 } 1778 return ret; 1779 } 1780 1781 /* 1782 * Try to reserve data bytes for an allocation. 1783 * 1784 * @fs_info: the filesystem 1785 * @bytes: number of bytes we need 1786 * @flush: how we are allowed to flush 1787 * 1788 * This will reserve bytes from the data space info. If there is not enough 1789 * space then we will attempt to flush space as specified by flush. 1790 */ 1791 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, 1792 enum btrfs_reserve_flush_enum flush) 1793 { 1794 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo; 1795 int ret; 1796 1797 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA || 1798 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE || 1799 flush == BTRFS_RESERVE_NO_FLUSH); 1800 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA); 1801 1802 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush); 1803 if (ret == -ENOSPC) { 1804 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1805 data_sinfo->flags, bytes, 1); 1806 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1807 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0); 1808 } 1809 return ret; 1810 } 1811 1812 /* Dump all the space infos when we abort a transaction due to ENOSPC. */ 1813 __cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info) 1814 { 1815 struct btrfs_space_info *space_info; 1816 1817 btrfs_info(fs_info, "dumping space info:"); 1818 list_for_each_entry(space_info, &fs_info->space_info, list) { 1819 spin_lock(&space_info->lock); 1820 __btrfs_dump_space_info(fs_info, space_info); 1821 spin_unlock(&space_info->lock); 1822 } 1823 dump_global_block_rsv(fs_info); 1824 } 1825 1826 /* 1827 * Account the unused space of all the readonly block group in the space_info. 1828 * takes mirrors into account. 1829 */ 1830 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo) 1831 { 1832 struct btrfs_block_group *block_group; 1833 u64 free_bytes = 0; 1834 int factor; 1835 1836 /* It's df, we don't care if it's racy */ 1837 if (list_empty(&sinfo->ro_bgs)) 1838 return 0; 1839 1840 spin_lock(&sinfo->lock); 1841 list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) { 1842 spin_lock(&block_group->lock); 1843 1844 if (!block_group->ro) { 1845 spin_unlock(&block_group->lock); 1846 continue; 1847 } 1848 1849 factor = btrfs_bg_type_to_factor(block_group->flags); 1850 free_bytes += (block_group->length - 1851 block_group->used) * factor; 1852 1853 spin_unlock(&block_group->lock); 1854 } 1855 spin_unlock(&sinfo->lock); 1856 1857 return free_bytes; 1858 } 1859