git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_cache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_buf.h"
   5 #include "btree_cache.h"
   6 #include "btree_io.h"
   7 #include "btree_iter.h"
   8 #include "btree_locking.h"
   9 #include "debug.h"
  10 #include "errcode.h"
  11 #include "error.h"
  12
  13 #include <linux/prefetch.h>
  14 #include <linux/sched/mm.h>
  15 #include <trace/events/bcachefs.h>
  16
  17 #define BTREE_CACHE_NOT_FREED_INCREMENT(counter) \
  18 do {                                             \
  19         if (shrinker_counter)                    \
  20                 bc->not_freed_##counter++;       \
  21 } while (0)
  22
  23 const char * const bch2_btree_node_flags[] = {
  24 #define x(f)    #f,
  25         BTREE_FLAGS()
  26 #undef x
  27         NULL
  28 };
  29
  30 void bch2_recalc_btree_reserve(struct bch_fs *c)
  31 {
  32         unsigned i, reserve = 16;
  33
  34         if (!c->btree_roots[0].b)
  35                 reserve += 8;
  36
  37         for (i = 0; i < BTREE_ID_NR; i++)
  38                 if (c->btree_roots[i].b)
  39                         reserve += min_t(unsigned, 1,
  40                                          c->btree_roots[i].b->c.level) * 8;
  41
  42         c->btree_cache.reserve = reserve;
  43 }
  44
  45 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  46 {
  47         return max_t(int, 0, bc->used - bc->reserve);
  48 }
  49
  50 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
  51 {
  52         if (b->c.lock.readers)
  53                 list_move(&b->list, &bc->freed_pcpu);
  54         else
  55                 list_move(&b->list, &bc->freed_nonpcpu);
  56 }
  57
  58 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  59 {
  60         struct btree_cache *bc = &c->btree_cache;
  61
  62         EBUG_ON(btree_node_write_in_flight(b));
  63
  64         kvpfree(b->data, btree_bytes(c));
  65         b->data = NULL;
  66 #ifdef __KERNEL__
  67         vfree(b->aux_data);
  68 #else
  69         munmap(b->aux_data, btree_aux_data_bytes(b));
  70 #endif
  71         b->aux_data = NULL;
  72
  73         bc->used--;
  74
  75         btree_node_to_freedlist(bc, b);
  76 }
  77
  78 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  79                                    const void *obj)
  80 {
  81         const struct btree *b = obj;
  82         const u64 *v = arg->key;
  83
  84         return b->hash_val == *v ? 0 : 1;
  85 }
  86
  87 static const struct rhashtable_params bch_btree_cache_params = {
  88         .head_offset    = offsetof(struct btree, hash),
  89         .key_offset     = offsetof(struct btree, hash_val),
  90         .key_len        = sizeof(u64),
  91         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  92 };
  93
  94 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  95 {
  96         BUG_ON(b->data || b->aux_data);
  97
  98         b->data = kvpmalloc(btree_bytes(c), gfp);
  99         if (!b->data)
 100                 return -ENOMEM;
 101 #ifdef __KERNEL__
 102         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
 103 #else
 104         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
 105                            PROT_READ|PROT_WRITE|PROT_EXEC,
 106                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 107         if (b->aux_data == MAP_FAILED)
 108                 b->aux_data = NULL;
 109 #endif
 110         if (!b->aux_data) {
 111                 kvpfree(b->data, btree_bytes(c));
 112                 b->data = NULL;
 113                 return -ENOMEM;
 114         }
 115
 116         return 0;
 117 }
 118
 119 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
 120 {
 121         struct btree *b;
 122
 123         b = kzalloc(sizeof(struct btree), gfp);
 124         if (!b)
 125                 return NULL;
 126
 127         bkey_btree_ptr_init(&b->key);
 128         __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 129 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 130         lockdep_set_no_check_recursion(&b->c.lock.dep_map);
 131 #endif
 132         INIT_LIST_HEAD(&b->list);
 133         INIT_LIST_HEAD(&b->write_blocked);
 134         b->byte_order = ilog2(btree_bytes(c));
 135         return b;
 136 }
 137
 138 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 139 {
 140         struct btree_cache *bc = &c->btree_cache;
 141         struct btree *b;
 142
 143         b = __btree_node_mem_alloc(c, GFP_KERNEL);
 144         if (!b)
 145                 return NULL;
 146
 147         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 148                 kfree(b);
 149                 return NULL;
 150         }
 151
 152         bc->used++;
 153         list_add(&b->list, &bc->freeable);
 154         return b;
 155 }
 156
 157 /* Btree in memory cache - hash table */
 158
 159 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 160 {
 161         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 162
 163         BUG_ON(ret);
 164
 165         /* Cause future lookups for this node to fail: */
 166         b->hash_val = 0;
 167 }
 168
 169 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 170 {
 171         BUG_ON(b->hash_val);
 172         b->hash_val = btree_ptr_hash_val(&b->key);
 173
 174         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 175                                              bch_btree_cache_params);
 176 }
 177
 178 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 179                                 unsigned level, enum btree_id id)
 180 {
 181         int ret;
 182
 183         b->c.level      = level;
 184         b->c.btree_id   = id;
 185
 186         mutex_lock(&bc->lock);
 187         ret = __bch2_btree_node_hash_insert(bc, b);
 188         if (!ret)
 189                 list_add_tail(&b->list, &bc->live);
 190         mutex_unlock(&bc->lock);
 191
 192         return ret;
 193 }
 194
 195 __flatten
 196 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 197                                      const struct bkey_i *k)
 198 {
 199         u64 v = btree_ptr_hash_val(k);
 200
 201         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 202 }
 203
 204 /*
 205  * this version is for btree nodes that have already been freed (we're not
 206  * reaping a real btree node)
 207  */
 208 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush, bool shrinker_counter)
 209 {
 210         struct btree_cache *bc = &c->btree_cache;
 211         int ret = 0;
 212
 213         lockdep_assert_held(&bc->lock);
 214 wait_on_io:
 215         if (b->flags & ((1U << BTREE_NODE_dirty)|
 216                         (1U << BTREE_NODE_read_in_flight)|
 217                         (1U << BTREE_NODE_write_in_flight))) {
 218                 if (!flush) {
 219                         if (btree_node_dirty(b))
 220                                 BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
 221                         else if (btree_node_read_in_flight(b))
 222                                 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
 223                         else if (btree_node_write_in_flight(b))
 224                                 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
 225                         return -ENOMEM;
 226                 }
 227
 228                 /* XXX: waiting on IO with btree cache lock held */
 229                 bch2_btree_node_wait_on_read(b);
 230                 bch2_btree_node_wait_on_write(b);
 231         }
 232
 233         if (!six_trylock_intent(&b->c.lock)) {
 234                 BTREE_CACHE_NOT_FREED_INCREMENT(lock_intent);
 235                 return -ENOMEM;
 236         }
 237
 238         if (!six_trylock_write(&b->c.lock)) {
 239                 BTREE_CACHE_NOT_FREED_INCREMENT(lock_write);
 240                 goto out_unlock_intent;
 241         }
 242
 243         /* recheck under lock */
 244         if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 245                         (1U << BTREE_NODE_write_in_flight))) {
 246                 if (!flush) {
 247                         if (btree_node_read_in_flight(b))
 248                                 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
 249                         else if (btree_node_write_in_flight(b))
 250                                 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
 251                         goto out_unlock;
 252                 }
 253                 six_unlock_write(&b->c.lock);
 254                 six_unlock_intent(&b->c.lock);
 255                 goto wait_on_io;
 256         }
 257
 258         if (btree_node_noevict(b)) {
 259                 BTREE_CACHE_NOT_FREED_INCREMENT(noevict);
 260                 goto out_unlock;
 261         }
 262         if (btree_node_write_blocked(b)) {
 263                 BTREE_CACHE_NOT_FREED_INCREMENT(write_blocked);
 264                 goto out_unlock;
 265         }
 266         if (btree_node_will_make_reachable(b)) {
 267                 BTREE_CACHE_NOT_FREED_INCREMENT(will_make_reachable);
 268                 goto out_unlock;
 269         }
 270
 271         if (btree_node_dirty(b)) {
 272                 if (!flush) {
 273                         BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
 274                         goto out_unlock;
 275                 }
 276                 /*
 277                  * Using the underscore version because we don't want to compact
 278                  * bsets after the write, since this node is about to be evicted
 279                  * - unless btree verify mode is enabled, since it runs out of
 280                  * the post write cleanup:
 281                  */
 282                 if (bch2_verify_btree_ondisk)
 283                         bch2_btree_node_write(c, b, SIX_LOCK_intent,
 284                                               BTREE_WRITE_cache_reclaim);
 285                 else
 286                         __bch2_btree_node_write(c, b,
 287                                                 BTREE_WRITE_cache_reclaim);
 288
 289                 six_unlock_write(&b->c.lock);
 290                 six_unlock_intent(&b->c.lock);
 291                 goto wait_on_io;
 292         }
 293 out:
 294         if (b->hash_val && !ret)
 295                 trace_and_count(c, btree_cache_reap, c, b);
 296         return ret;
 297 out_unlock:
 298         six_unlock_write(&b->c.lock);
 299 out_unlock_intent:
 300         six_unlock_intent(&b->c.lock);
 301         ret = -ENOMEM;
 302         goto out;
 303 }
 304
 305 static int btree_node_reclaim(struct bch_fs *c, struct btree *b, bool shrinker_counter)
 306 {
 307         return __btree_node_reclaim(c, b, false, shrinker_counter);
 308 }
 309
 310 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 311 {
 312         return __btree_node_reclaim(c, b, true, false);
 313 }
 314
 315 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 316                                            struct shrink_control *sc)
 317 {
 318         struct bch_fs *c = container_of(shrink, struct bch_fs,
 319                                         btree_cache.shrink);
 320         struct btree_cache *bc = &c->btree_cache;
 321         struct btree *b, *t;
 322         unsigned long nr = sc->nr_to_scan;
 323         unsigned long can_free = 0;
 324         unsigned long freed = 0;
 325         unsigned long touched = 0;
 326         unsigned i, flags;
 327         unsigned long ret = SHRINK_STOP;
 328         bool trigger_writes = atomic_read(&bc->dirty) + nr >=
 329                 bc->used * 3 / 4;
 330
 331         if (bch2_btree_shrinker_disabled)
 332                 return SHRINK_STOP;
 333
 334         mutex_lock(&bc->lock);
 335         flags = memalloc_nofs_save();
 336
 337         /*
 338          * It's _really_ critical that we don't free too many btree nodes - we
 339          * have to always leave ourselves a reserve. The reserve is how we
 340          * guarantee that allocating memory for a new btree node can always
 341          * succeed, so that inserting keys into the btree can always succeed and
 342          * IO can always make forward progress:
 343          */
 344         can_free = btree_cache_can_free(bc);
 345         nr = min_t(unsigned long, nr, can_free);
 346
 347         i = 0;
 348         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 349                 /*
 350                  * Leave a few nodes on the freeable list, so that a btree split
 351                  * won't have to hit the system allocator:
 352                  */
 353                 if (++i <= 3)
 354                         continue;
 355
 356                 touched++;
 357
 358                 if (touched >= nr)
 359                         goto out;
 360
 361                 if (!btree_node_reclaim(c, b, true)) {
 362                         btree_node_data_free(c, b);
 363                         six_unlock_write(&b->c.lock);
 364                         six_unlock_intent(&b->c.lock);
 365                         freed++;
 366                         bc->freed++;
 367                 }
 368         }
 369 restart:
 370         list_for_each_entry_safe(b, t, &bc->live, list) {
 371                 touched++;
 372
 373                 if (btree_node_accessed(b)) {
 374                         clear_btree_node_accessed(b);
 375                         bc->not_freed_access_bit++;
 376                 } else if (!btree_node_reclaim(c, b, true)) {
 377                         freed++;
 378                         btree_node_data_free(c, b);
 379                         bc->freed++;
 380
 381                         bch2_btree_node_hash_remove(bc, b);
 382                         six_unlock_write(&b->c.lock);
 383                         six_unlock_intent(&b->c.lock);
 384
 385                         if (freed == nr)
 386                                 goto out_rotate;
 387                 } else if (trigger_writes &&
 388                            btree_node_dirty(b) &&
 389                            !btree_node_will_make_reachable(b) &&
 390                            !btree_node_write_blocked(b) &&
 391                            six_trylock_read(&b->c.lock)) {
 392                         list_move(&bc->live, &b->list);
 393                         mutex_unlock(&bc->lock);
 394                         __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
 395                         six_unlock_read(&b->c.lock);
 396                         if (touched >= nr)
 397                                 goto out_nounlock;
 398                         mutex_lock(&bc->lock);
 399                         goto restart;
 400                 }
 401
 402                 if (touched >= nr)
 403                         break;
 404         }
 405 out_rotate:
 406         if (&t->list != &bc->live)
 407                 list_move_tail(&bc->live, &t->list);
 408 out:
 409         mutex_unlock(&bc->lock);
 410 out_nounlock:
 411         ret = freed;
 412         memalloc_nofs_restore(flags);
 413         trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
 414         return ret;
 415 }
 416
 417 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 418                                             struct shrink_control *sc)
 419 {
 420         struct bch_fs *c = container_of(shrink, struct bch_fs,
 421                                         btree_cache.shrink);
 422         struct btree_cache *bc = &c->btree_cache;
 423
 424         if (bch2_btree_shrinker_disabled)
 425                 return 0;
 426
 427         return btree_cache_can_free(bc);
 428 }
 429
 430 static void bch2_btree_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
 431 {
 432         struct bch_fs *c = container_of(shrink, struct bch_fs,
 433                                         btree_cache.shrink);
 434
 435         bch2_btree_cache_to_text(out, &c->btree_cache);
 436 }
 437
 438 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 439 {
 440         struct btree_cache *bc = &c->btree_cache;
 441         struct btree *b;
 442         unsigned i, flags;
 443
 444         if (bc->shrink.list.next)
 445                 unregister_shrinker(&bc->shrink);
 446
 447         /* vfree() can allocate memory: */
 448         flags = memalloc_nofs_save();
 449         mutex_lock(&bc->lock);
 450
 451         if (c->verify_data)
 452                 list_move(&c->verify_data->list, &bc->live);
 453
 454         kvpfree(c->verify_ondisk, btree_bytes(c));
 455
 456         for (i = 0; i < BTREE_ID_NR; i++)
 457                 if (c->btree_roots[i].b)
 458                         list_add(&c->btree_roots[i].b->list, &bc->live);
 459
 460         list_splice(&bc->freeable, &bc->live);
 461
 462         while (!list_empty(&bc->live)) {
 463                 b = list_first_entry(&bc->live, struct btree, list);
 464
 465                 BUG_ON(btree_node_read_in_flight(b) ||
 466                        btree_node_write_in_flight(b));
 467
 468                 if (btree_node_dirty(b))
 469                         bch2_btree_complete_write(c, b, btree_current_write(b));
 470                 clear_btree_node_dirty_acct(c, b);
 471
 472                 btree_node_data_free(c, b);
 473         }
 474
 475         BUG_ON(atomic_read(&c->btree_cache.dirty));
 476
 477         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 478
 479         while (!list_empty(&bc->freed_nonpcpu)) {
 480                 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
 481                 list_del(&b->list);
 482                 six_lock_pcpu_free(&b->c.lock);
 483                 kfree(b);
 484         }
 485
 486         mutex_unlock(&bc->lock);
 487         memalloc_nofs_restore(flags);
 488
 489         if (bc->table_init_done)
 490                 rhashtable_destroy(&bc->table);
 491 }
 492
 493 int bch2_fs_btree_cache_init(struct bch_fs *c)
 494 {
 495         struct btree_cache *bc = &c->btree_cache;
 496         unsigned i;
 497         int ret = 0;
 498
 499         pr_verbose_init(c->opts, "");
 500
 501         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 502         if (ret)
 503                 goto out;
 504
 505         bc->table_init_done = true;
 506
 507         bch2_recalc_btree_reserve(c);
 508
 509         for (i = 0; i < bc->reserve; i++)
 510                 if (!__bch2_btree_node_mem_alloc(c)) {
 511                         ret = -ENOMEM;
 512                         goto out;
 513                 }
 514
 515         list_splice_init(&bc->live, &bc->freeable);
 516
 517         mutex_init(&c->verify_lock);
 518
 519         bc->shrink.count_objects        = bch2_btree_cache_count;
 520         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 521         bc->shrink.to_text              = bch2_btree_cache_shrinker_to_text;
 522         bc->shrink.seeks                = 4;
 523         ret = register_shrinker(&bc->shrink, "%s/btree_cache", c->name);
 524 out:
 525         pr_verbose_init(c->opts, "ret %i", ret);
 526         return ret;
 527 }
 528
 529 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 530 {
 531         mutex_init(&bc->lock);
 532         INIT_LIST_HEAD(&bc->live);
 533         INIT_LIST_HEAD(&bc->freeable);
 534         INIT_LIST_HEAD(&bc->freed_pcpu);
 535         INIT_LIST_HEAD(&bc->freed_nonpcpu);
 536 }
 537
 538 /*
 539  * We can only have one thread cannibalizing other cached btree nodes at a time,
 540  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 541  * cannibalize_bucket() will take. This means every time we unlock the root of
 542  * the btree, we need to release this lock if we have it held.
 543  */
 544 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 545 {
 546         struct btree_cache *bc = &c->btree_cache;
 547
 548         if (bc->alloc_lock == current) {
 549                 trace_and_count(c, btree_cache_cannibalize_unlock, c);
 550                 bc->alloc_lock = NULL;
 551                 closure_wake_up(&bc->alloc_wait);
 552         }
 553 }
 554
 555 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 556 {
 557         struct btree_cache *bc = &c->btree_cache;
 558         struct task_struct *old;
 559
 560         old = cmpxchg(&bc->alloc_lock, NULL, current);
 561         if (old == NULL || old == current)
 562                 goto success;
 563
 564         if (!cl) {
 565                 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 566                 return -ENOMEM;
 567         }
 568
 569         closure_wait(&bc->alloc_wait, cl);
 570
 571         /* Try again, after adding ourselves to waitlist */
 572         old = cmpxchg(&bc->alloc_lock, NULL, current);
 573         if (old == NULL || old == current) {
 574                 /* We raced */
 575                 closure_wake_up(&bc->alloc_wait);
 576                 goto success;
 577         }
 578
 579         trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 580         return -EAGAIN;
 581
 582 success:
 583         trace_and_count(c, btree_cache_cannibalize_lock, c);
 584         return 0;
 585 }
 586
 587 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 588 {
 589         struct btree_cache *bc = &c->btree_cache;
 590         struct btree *b;
 591
 592         list_for_each_entry_reverse(b, &bc->live, list)
 593                 if (!btree_node_reclaim(c, b, false))
 594                         return b;
 595
 596         while (1) {
 597                 list_for_each_entry_reverse(b, &bc->live, list)
 598                         if (!btree_node_write_and_reclaim(c, b))
 599                                 return b;
 600
 601                 /*
 602                  * Rare case: all nodes were intent-locked.
 603                  * Just busy-wait.
 604                  */
 605                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 606                 cond_resched();
 607         }
 608 }
 609
 610 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
 611 {
 612         struct btree_cache *bc = &c->btree_cache;
 613         struct list_head *freed = pcpu_read_locks
 614                 ? &bc->freed_pcpu
 615                 : &bc->freed_nonpcpu;
 616         struct btree *b, *b2;
 617         u64 start_time = local_clock();
 618         unsigned flags;
 619
 620         flags = memalloc_nofs_save();
 621         mutex_lock(&bc->lock);
 622
 623         /*
 624          * We never free struct btree itself, just the memory that holds the on
 625          * disk node. Check the freed list before allocating a new one:
 626          */
 627         list_for_each_entry(b, freed, list)
 628                 if (!btree_node_reclaim(c, b, false)) {
 629                         list_del_init(&b->list);
 630                         goto got_node;
 631                 }
 632
 633         b = __btree_node_mem_alloc(c, __GFP_NOWARN);
 634         if (!b) {
 635                 mutex_unlock(&bc->lock);
 636                 b = __btree_node_mem_alloc(c, GFP_KERNEL);
 637                 if (!b)
 638                         goto err;
 639                 mutex_lock(&bc->lock);
 640         }
 641
 642         if (pcpu_read_locks)
 643                 six_lock_pcpu_alloc(&b->c.lock);
 644
 645         BUG_ON(!six_trylock_intent(&b->c.lock));
 646         BUG_ON(!six_trylock_write(&b->c.lock));
 647 got_node:
 648
 649         /*
 650          * btree_free() doesn't free memory; it sticks the node on the end of
 651          * the list. Check if there's any freed nodes there:
 652          */
 653         list_for_each_entry(b2, &bc->freeable, list)
 654                 if (!btree_node_reclaim(c, b2, false)) {
 655                         swap(b->data, b2->data);
 656                         swap(b->aux_data, b2->aux_data);
 657                         btree_node_to_freedlist(bc, b2);
 658                         six_unlock_write(&b2->c.lock);
 659                         six_unlock_intent(&b2->c.lock);
 660                         goto got_mem;
 661                 }
 662
 663         mutex_unlock(&bc->lock);
 664
 665         if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 666                 goto err;
 667
 668         mutex_lock(&bc->lock);
 669         bc->used++;
 670 got_mem:
 671         mutex_unlock(&bc->lock);
 672
 673         BUG_ON(btree_node_hashed(b));
 674         BUG_ON(btree_node_dirty(b));
 675         BUG_ON(btree_node_write_in_flight(b));
 676 out:
 677         b->flags                = 0;
 678         b->written              = 0;
 679         b->nsets                = 0;
 680         b->sib_u64s[0]          = 0;
 681         b->sib_u64s[1]          = 0;
 682         b->whiteout_u64s        = 0;
 683         bch2_btree_keys_init(b);
 684         set_btree_node_accessed(b);
 685
 686         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 687                                start_time);
 688
 689         memalloc_nofs_restore(flags);
 690         return b;
 691 err:
 692         mutex_lock(&bc->lock);
 693
 694         /* Try to cannibalize another cached btree node: */
 695         if (bc->alloc_lock == current) {
 696                 b2 = btree_node_cannibalize(c);
 697                 bch2_btree_node_hash_remove(bc, b2);
 698
 699                 if (b) {
 700                         swap(b->data, b2->data);
 701                         swap(b->aux_data, b2->aux_data);
 702                         btree_node_to_freedlist(bc, b2);
 703                         six_unlock_write(&b2->c.lock);
 704                         six_unlock_intent(&b2->c.lock);
 705                 } else {
 706                         b = b2;
 707                         list_del_init(&b->list);
 708                 }
 709
 710                 mutex_unlock(&bc->lock);
 711
 712                 trace_and_count(c, btree_cache_cannibalize, c);
 713                 goto out;
 714         }
 715
 716         mutex_unlock(&bc->lock);
 717         memalloc_nofs_restore(flags);
 718         return ERR_PTR(-ENOMEM);
 719 }
 720
 721 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 722 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 723                                 struct btree_trans *trans,
 724                                 struct btree_path *path,
 725                                 const struct bkey_i *k,
 726                                 enum btree_id btree_id,
 727                                 unsigned level,
 728                                 enum six_lock_type lock_type,
 729                                 bool sync)
 730 {
 731         struct btree_cache *bc = &c->btree_cache;
 732         struct btree *b;
 733         u32 seq;
 734
 735         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 736         /*
 737          * Parent node must be locked, else we could read in a btree node that's
 738          * been freed:
 739          */
 740         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 741                 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
 742                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
 743         }
 744
 745         b = bch2_btree_node_mem_alloc(c, level != 0);
 746
 747         if (trans && b == ERR_PTR(-ENOMEM)) {
 748                 trans->memory_allocation_failure = true;
 749                 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
 750                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
 751         }
 752
 753         if (IS_ERR(b))
 754                 return b;
 755
 756         bkey_copy(&b->key, k);
 757         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 758                 /* raced with another fill: */
 759
 760                 /* mark as unhashed... */
 761                 b->hash_val = 0;
 762
 763                 mutex_lock(&bc->lock);
 764                 list_add(&b->list, &bc->freeable);
 765                 mutex_unlock(&bc->lock);
 766
 767                 six_unlock_write(&b->c.lock);
 768                 six_unlock_intent(&b->c.lock);
 769                 return NULL;
 770         }
 771
 772         set_btree_node_read_in_flight(b);
 773
 774         six_unlock_write(&b->c.lock);
 775         seq = b->c.lock.state.seq;
 776         six_unlock_intent(&b->c.lock);
 777
 778         /* Unlock before doing IO: */
 779         if (trans && sync)
 780                 bch2_trans_unlock(trans);
 781
 782         bch2_btree_node_read(c, b, sync);
 783
 784         if (!sync)
 785                 return NULL;
 786
 787         if (trans) {
 788                 int ret = bch2_trans_relock(trans) ?:
 789                         bch2_btree_path_relock_intent(trans, path);
 790                 if (ret) {
 791                         BUG_ON(!trans->restarted);
 792                         return ERR_PTR(ret);
 793                 }
 794         }
 795
 796         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 797                 if (trans)
 798                         trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
 799                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
 800         }
 801
 802         return b;
 803 }
 804
 805 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 806 {
 807         struct printbuf buf = PRINTBUF;
 808
 809         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 810                 return;
 811
 812         prt_printf(&buf,
 813                "btree node header doesn't match ptr\n"
 814                "btree %s level %u\n"
 815                "ptr: ",
 816                bch2_btree_ids[b->c.btree_id], b->c.level);
 817         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 818
 819         prt_printf(&buf, "\nheader: btree %s level %llu\n"
 820                "min ",
 821                bch2_btree_ids[BTREE_NODE_ID(b->data)],
 822                BTREE_NODE_LEVEL(b->data));
 823         bch2_bpos_to_text(&buf, b->data->min_key);
 824
 825         prt_printf(&buf, "\nmax ");
 826         bch2_bpos_to_text(&buf, b->data->max_key);
 827
 828         bch2_fs_inconsistent(c, "%s", buf.buf);
 829         printbuf_exit(&buf);
 830 }
 831
 832 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 833 {
 834         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 835             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 836             bpos_cmp(b->data->max_key, b->key.k.p) ||
 837             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 838              bpos_cmp(b->data->min_key,
 839                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 840                 btree_bad_header(c, b);
 841 }
 842
 843 /**
 844  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 845  * in from disk if necessary.
 846  *
 847  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 848  *
 849  * The btree node will have either a read or a write lock held, depending on
 850  * the @write parameter.
 851  */
 852 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 853                                   const struct bkey_i *k, unsigned level,
 854                                   enum six_lock_type lock_type,
 855                                   unsigned long trace_ip)
 856 {
 857         struct bch_fs *c = trans->c;
 858         struct btree_cache *bc = &c->btree_cache;
 859         struct btree *b;
 860         struct bset_tree *t;
 861         int ret;
 862
 863         EBUG_ON(level >= BTREE_MAX_DEPTH);
 864
 865         b = btree_node_mem_ptr(k);
 866
 867         /*
 868          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 869          * be the node we want anymore, and trying to lock the wrong node could
 870          * cause an unneccessary transaction restart:
 871          */
 872         if (likely(c->opts.btree_node_mem_ptr_optimization &&
 873                    b &&
 874                    b->hash_val == btree_ptr_hash_val(k)))
 875                 goto lock_node;
 876 retry:
 877         b = btree_cache_find(bc, k);
 878         if (unlikely(!b)) {
 879                 /*
 880                  * We must have the parent locked to call bch2_btree_node_fill(),
 881                  * else we could read in a btree node from disk that's been
 882                  * freed:
 883                  */
 884                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 885                                          level, lock_type, true);
 886
 887                 /* We raced and found the btree node in the cache */
 888                 if (!b)
 889                         goto retry;
 890
 891                 if (IS_ERR(b))
 892                         return b;
 893         } else {
 894 lock_node:
 895                 /*
 896                  * There's a potential deadlock with splits and insertions into
 897                  * interior nodes we have to avoid:
 898                  *
 899                  * The other thread might be holding an intent lock on the node
 900                  * we want, and they want to update its parent node so they're
 901                  * going to upgrade their intent lock on the parent node to a
 902                  * write lock.
 903                  *
 904                  * But if we're holding a read lock on the parent, and we're
 905                  * trying to get the intent lock they're holding, we deadlock.
 906                  *
 907                  * So to avoid this we drop the read locks on parent nodes when
 908                  * we're starting to take intent locks - and handle the race.
 909                  *
 910                  * The race is that they might be about to free the node we
 911                  * want, and dropping our read lock on the parent node lets them
 912                  * update the parent marking the node we want as freed, and then
 913                  * free it:
 914                  *
 915                  * To guard against this, btree nodes are evicted from the cache
 916                  * when they're freed - and b->hash_val is zeroed out, which we
 917                  * check for after we lock the node.
 918                  *
 919                  * Then, bch2_btree_node_relock() on the parent will fail - because
 920                  * the parent was modified, when the pointer to the node we want
 921                  * was removed - and we'll bail out:
 922                  */
 923                 if (btree_node_read_locked(path, level + 1))
 924                         btree_node_unlock(trans, path, level + 1);
 925
 926                 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
 927                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 928                         return ERR_PTR(ret);
 929
 930                 BUG_ON(ret);
 931
 932                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 933                              b->c.level != level ||
 934                              race_fault())) {
 935                         six_unlock_type(&b->c.lock, lock_type);
 936                         if (bch2_btree_node_relock(trans, path, level + 1))
 937                                 goto retry;
 938
 939                         trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
 940                         return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
 941                 }
 942         }
 943
 944         if (unlikely(btree_node_read_in_flight(b))) {
 945                 u32 seq = b->c.lock.state.seq;
 946
 947                 six_unlock_type(&b->c.lock, lock_type);
 948                 bch2_trans_unlock(trans);
 949
 950                 bch2_btree_node_wait_on_read(b);
 951
 952                 /*
 953                  * should_be_locked is not set on this path yet, so we need to
 954                  * relock it specifically:
 955                  */
 956                 if (trans) {
 957                         int ret = bch2_trans_relock(trans) ?:
 958                                 bch2_btree_path_relock_intent(trans, path);
 959                         if (ret) {
 960                                 BUG_ON(!trans->restarted);
 961                                 return ERR_PTR(ret);
 962                         }
 963                 }
 964
 965                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 966                         goto retry;
 967         }
 968
 969         prefetch(b->aux_data);
 970
 971         for_each_bset(b, t) {
 972                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 973
 974                 prefetch(p + L1_CACHE_BYTES * 0);
 975                 prefetch(p + L1_CACHE_BYTES * 1);
 976                 prefetch(p + L1_CACHE_BYTES * 2);
 977         }
 978
 979         /* avoid atomic set bit if it's not needed: */
 980         if (!btree_node_accessed(b))
 981                 set_btree_node_accessed(b);
 982
 983         if (unlikely(btree_node_read_error(b))) {
 984                 six_unlock_type(&b->c.lock, lock_type);
 985                 return ERR_PTR(-EIO);
 986         }
 987
 988         EBUG_ON(b->c.btree_id != path->btree_id);
 989         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 990         btree_check_header(c, b);
 991
 992         return b;
 993 }
 994
 995 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
 996                                          const struct bkey_i *k,
 997                                          enum btree_id btree_id,
 998                                          unsigned level,
 999                                          bool nofill)
1000 {
1001         struct bch_fs *c = trans->c;
1002         struct btree_cache *bc = &c->btree_cache;
1003         struct btree *b;
1004         struct bset_tree *t;
1005         int ret;
1006
1007         EBUG_ON(level >= BTREE_MAX_DEPTH);
1008
1009         if (c->opts.btree_node_mem_ptr_optimization) {
1010                 b = btree_node_mem_ptr(k);
1011                 if (b)
1012                         goto lock_node;
1013         }
1014 retry:
1015         b = btree_cache_find(bc, k);
1016         if (unlikely(!b)) {
1017                 if (nofill)
1018                         goto out;
1019
1020                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
1021                                          level, SIX_LOCK_read, true);
1022
1023                 /* We raced and found the btree node in the cache */
1024                 if (!b)
1025                         goto retry;
1026
1027                 if (IS_ERR(b) &&
1028                     !bch2_btree_cache_cannibalize_lock(c, NULL))
1029                         goto retry;
1030
1031                 if (IS_ERR(b))
1032                         goto out;
1033         } else {
1034 lock_node:
1035                 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read);
1036                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1037                         return ERR_PTR(ret);
1038
1039                 BUG_ON(ret);
1040
1041                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1042                              b->c.btree_id != btree_id ||
1043                              b->c.level != level)) {
1044                         six_unlock_read(&b->c.lock);
1045                         goto retry;
1046                 }
1047         }
1048
1049         /* XXX: waiting on IO with btree locks held: */
1050         __bch2_btree_node_wait_on_read(b);
1051
1052         prefetch(b->aux_data);
1053
1054         for_each_bset(b, t) {
1055                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1056
1057                 prefetch(p + L1_CACHE_BYTES * 0);
1058                 prefetch(p + L1_CACHE_BYTES * 1);
1059                 prefetch(p + L1_CACHE_BYTES * 2);
1060         }
1061
1062         /* avoid atomic set bit if it's not needed: */
1063         if (!btree_node_accessed(b))
1064                 set_btree_node_accessed(b);
1065
1066         if (unlikely(btree_node_read_error(b))) {
1067                 six_unlock_read(&b->c.lock);
1068                 b = ERR_PTR(-EIO);
1069                 goto out;
1070         }
1071
1072         EBUG_ON(b->c.btree_id != btree_id);
1073         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1074         btree_check_header(c, b);
1075 out:
1076         bch2_btree_cache_cannibalize_unlock(c);
1077         return b;
1078 }
1079
1080 int bch2_btree_node_prefetch(struct bch_fs *c,
1081                              struct btree_trans *trans,
1082                              struct btree_path *path,
1083                              const struct bkey_i *k,
1084                              enum btree_id btree_id, unsigned level)
1085 {
1086         struct btree_cache *bc = &c->btree_cache;
1087         struct btree *b;
1088
1089         BUG_ON(trans && !btree_node_locked(path, level + 1));
1090         BUG_ON(level >= BTREE_MAX_DEPTH);
1091
1092         b = btree_cache_find(bc, k);
1093         if (b)
1094                 return 0;
1095
1096         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1097                                  level, SIX_LOCK_read, false);
1098         return PTR_ERR_OR_ZERO(b);
1099 }
1100
1101 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1102 {
1103         struct bch_fs *c = trans->c;
1104         struct btree_cache *bc = &c->btree_cache;
1105         struct btree *b;
1106
1107         b = btree_cache_find(bc, k);
1108         if (!b)
1109                 return;
1110 wait_on_io:
1111         /* not allowed to wait on io with btree locks held: */
1112
1113         /* XXX we're called from btree_gc which will be holding other btree
1114          * nodes locked
1115          */
1116         __bch2_btree_node_wait_on_read(b);
1117         __bch2_btree_node_wait_on_write(b);
1118
1119         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1120         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1121
1122         if (btree_node_dirty(b)) {
1123                 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
1124                 six_unlock_write(&b->c.lock);
1125                 six_unlock_intent(&b->c.lock);
1126                 goto wait_on_io;
1127         }
1128
1129         BUG_ON(btree_node_dirty(b));
1130
1131         mutex_lock(&bc->lock);
1132         btree_node_data_free(c, b);
1133         bch2_btree_node_hash_remove(bc, b);
1134         mutex_unlock(&bc->lock);
1135
1136         six_unlock_write(&b->c.lock);
1137         six_unlock_intent(&b->c.lock);
1138 }
1139
1140 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1141                              struct btree *b)
1142 {
1143         const struct bkey_format *f = &b->format;
1144         struct bset_stats stats;
1145
1146         memset(&stats, 0, sizeof(stats));
1147
1148         bch2_btree_keys_stats(b, &stats);
1149
1150         prt_printf(out, "l %u ", b->c.level);
1151         bch2_bpos_to_text(out, b->data->min_key);
1152         prt_printf(out, " - ");
1153         bch2_bpos_to_text(out, b->data->max_key);
1154         prt_printf(out, ":\n"
1155                "    ptrs: ");
1156         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1157
1158         prt_printf(out, "\n"
1159                "    format: u64s %u fields %u %u %u %u %u\n"
1160                "    unpack fn len: %u\n"
1161                "    bytes used %zu/%zu (%zu%% full)\n"
1162                "    sib u64s: %u, %u (merge threshold %u)\n"
1163                "    nr packed keys %u\n"
1164                "    nr unpacked keys %u\n"
1165                "    floats %zu\n"
1166                "    failed unpacked %zu\n",
1167                f->key_u64s,
1168                f->bits_per_field[0],
1169                f->bits_per_field[1],
1170                f->bits_per_field[2],
1171                f->bits_per_field[3],
1172                f->bits_per_field[4],
1173                b->unpack_fn_len,
1174                b->nr.live_u64s * sizeof(u64),
1175                btree_bytes(c) - sizeof(struct btree_node),
1176                b->nr.live_u64s * 100 / btree_max_u64s(c),
1177                b->sib_u64s[0],
1178                b->sib_u64s[1],
1179                c->btree_foreground_merge_threshold,
1180                b->nr.packed_keys,
1181                b->nr.unpacked_keys,
1182                stats.floats,
1183                stats.failed);
1184 }
1185
1186 void bch2_btree_cache_to_text(struct printbuf *out, struct btree_cache *bc)
1187 {
1188         prt_printf(out, "nr nodes:\t\t%u\n", bc->used);
1189         prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&bc->dirty));
1190         prt_printf(out, "cannibalize lock:\t%p\n", bc->alloc_lock);
1191
1192         prt_printf(out, "freed:\t\t\t\t%u\n", bc->freed);
1193         prt_printf(out, "not freed, dirty:\t\t%u\n", bc->not_freed_dirty);
1194         prt_printf(out, "not freed, write in flight:\t%u\n", bc->not_freed_write_in_flight);
1195         prt_printf(out, "not freed, read in flight:\t%u\n", bc->not_freed_read_in_flight);
1196         prt_printf(out, "not freed, lock intent failed:\t%u\n", bc->not_freed_lock_intent);
1197         prt_printf(out, "not freed, lock write failed:\t%u\n", bc->not_freed_lock_write);
1198         prt_printf(out, "not freed, access bit:\t\t%u\n", bc->not_freed_access_bit);
1199         prt_printf(out, "not freed, no evict failed:\t%u\n", bc->not_freed_noevict);
1200         prt_printf(out, "not freed, write blocked:\t%u\n", bc->not_freed_write_blocked);
1201         prt_printf(out, "not freed, will make reachable:\t%u\n", bc->not_freed_will_make_reachable);
1202
1203 }