fs/bcachefs/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 #include "trace.h"
  13
  14 #include <linux/prefetch.h>
  15 #include <linux/sched/mm.h>
  16
  17 const char * const bch2_btree_node_flags[] = {
  18 #define x(f)    #f,
  19         BTREE_FLAGS()
  20 #undef x
  21         NULL
  22 };
  23
  24 void bch2_recalc_btree_reserve(struct bch_fs *c)
  25 {
  26         unsigned i, reserve = 16;
  27
  28         if (!c->btree_roots_known[0].b)
  29                 reserve += 8;
  30
  31         for (i = 0; i < btree_id_nr_alive(c); i++) {
  32                 struct btree_root *r = bch2_btree_id_root(c, i);
  33
  34                 if (r->b)
  35                         reserve += min_t(unsigned, 1, r->b->c.level) * 8;
  36         }
  37
  38         c->btree_cache.reserve = reserve;
  39 }
  40
  41 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  42 {
  43         return max_t(int, 0, bc->used - bc->reserve);
  44 }
  45
  46 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
  47 {
  48         if (b->c.lock.readers)
  49                 list_move(&b->list, &bc->freed_pcpu);
  50         else
  51                 list_move(&b->list, &bc->freed_nonpcpu);
  52 }
  53
  54 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  55 {
  56         struct btree_cache *bc = &c->btree_cache;
  57
  58         EBUG_ON(btree_node_write_in_flight(b));
  59
  60         clear_btree_node_just_written(b);
  61
  62         kvpfree(b->data, btree_bytes(c));
  63         b->data = NULL;
  64 #ifdef __KERNEL__
  65         kvfree(b->aux_data);
  66 #else
  67         munmap(b->aux_data, btree_aux_data_bytes(b));
  68 #endif
  69         b->aux_data = NULL;
  70
  71         bc->used--;
  72
  73         btree_node_to_freedlist(bc, b);
  74 }
  75
  76 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  77                                    const void *obj)
  78 {
  79         const struct btree *b = obj;
  80         const u64 *v = arg->key;
  81
  82         return b->hash_val == *v ? 0 : 1;
  83 }
  84
  85 static const struct rhashtable_params bch_btree_cache_params = {
  86         .head_offset    = offsetof(struct btree, hash),
  87         .key_offset     = offsetof(struct btree, hash_val),
  88         .key_len        = sizeof(u64),
  89         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  90 };
  91
  92 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  93 {
  94         BUG_ON(b->data || b->aux_data);
  95
  96         b->data = kvpmalloc(btree_bytes(c), gfp);
  97         if (!b->data)
  98                 return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
  99 #ifdef __KERNEL__
 100         b->aux_data = kvmalloc(btree_aux_data_bytes(b), gfp);
 101 #else
 102         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
 103                            PROT_READ|PROT_WRITE|PROT_EXEC,
 104                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 105         if (b->aux_data == MAP_FAILED)
 106                 b->aux_data = NULL;
 107 #endif
 108         if (!b->aux_data) {
 109                 kvpfree(b->data, btree_bytes(c));
 110                 b->data = NULL;
 111                 return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
 112         }
 113
 114         return 0;
 115 }
 116
 117 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
 118 {
 119         struct btree *b;
 120
 121         b = kzalloc(sizeof(struct btree), gfp);
 122         if (!b)
 123                 return NULL;
 124
 125         bkey_btree_ptr_init(&b->key);
 126         INIT_LIST_HEAD(&b->list);
 127         INIT_LIST_HEAD(&b->write_blocked);
 128         b->byte_order = ilog2(btree_bytes(c));
 129         return b;
 130 }
 131
 132 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 133 {
 134         struct btree_cache *bc = &c->btree_cache;
 135         struct btree *b;
 136
 137         b = __btree_node_mem_alloc(c, GFP_KERNEL);
 138         if (!b)
 139                 return NULL;
 140
 141         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 142                 kfree(b);
 143                 return NULL;
 144         }
 145
 146         bch2_btree_lock_init(&b->c, 0);
 147
 148         bc->used++;
 149         list_add(&b->list, &bc->freeable);
 150         return b;
 151 }
 152
 153 /* Btree in memory cache - hash table */
 154
 155 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 156 {
 157         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 158
 159         BUG_ON(ret);
 160
 161         /* Cause future lookups for this node to fail: */
 162         b->hash_val = 0;
 163 }
 164
 165 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 166 {
 167         BUG_ON(b->hash_val);
 168         b->hash_val = btree_ptr_hash_val(&b->key);
 169
 170         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 171                                              bch_btree_cache_params);
 172 }
 173
 174 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 175                                 unsigned level, enum btree_id id)
 176 {
 177         int ret;
 178
 179         b->c.level      = level;
 180         b->c.btree_id   = id;
 181
 182         mutex_lock(&bc->lock);
 183         ret = __bch2_btree_node_hash_insert(bc, b);
 184         if (!ret)
 185                 list_add_tail(&b->list, &bc->live);
 186         mutex_unlock(&bc->lock);
 187
 188         return ret;
 189 }
 190
 191 __flatten
 192 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 193                                      const struct bkey_i *k)
 194 {
 195         u64 v = btree_ptr_hash_val(k);
 196
 197         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 198 }
 199
 200 /*
 201  * this version is for btree nodes that have already been freed (we're not
 202  * reaping a real btree node)
 203  */
 204 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 205 {
 206         struct btree_cache *bc = &c->btree_cache;
 207         int ret = 0;
 208
 209         lockdep_assert_held(&bc->lock);
 210 wait_on_io:
 211         if (b->flags & ((1U << BTREE_NODE_dirty)|
 212                         (1U << BTREE_NODE_read_in_flight)|
 213                         (1U << BTREE_NODE_write_in_flight))) {
 214                 if (!flush)
 215                         return -BCH_ERR_ENOMEM_btree_node_reclaim;
 216
 217                 /* XXX: waiting on IO with btree cache lock held */
 218                 bch2_btree_node_wait_on_read(b);
 219                 bch2_btree_node_wait_on_write(b);
 220         }
 221
 222         if (!six_trylock_intent(&b->c.lock))
 223                 return -BCH_ERR_ENOMEM_btree_node_reclaim;
 224
 225         if (!six_trylock_write(&b->c.lock))
 226                 goto out_unlock_intent;
 227
 228         /* recheck under lock */
 229         if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 230                         (1U << BTREE_NODE_write_in_flight))) {
 231                 if (!flush)
 232                         goto out_unlock;
 233                 six_unlock_write(&b->c.lock);
 234                 six_unlock_intent(&b->c.lock);
 235                 goto wait_on_io;
 236         }
 237
 238         if (btree_node_noevict(b) ||
 239             btree_node_write_blocked(b) ||
 240             btree_node_will_make_reachable(b))
 241                 goto out_unlock;
 242
 243         if (btree_node_dirty(b)) {
 244                 if (!flush)
 245                         goto out_unlock;
 246                 /*
 247                  * Using the underscore version because we don't want to compact
 248                  * bsets after the write, since this node is about to be evicted
 249                  * - unless btree verify mode is enabled, since it runs out of
 250                  * the post write cleanup:
 251                  */
 252                 if (bch2_verify_btree_ondisk)
 253                         bch2_btree_node_write(c, b, SIX_LOCK_intent,
 254                                               BTREE_WRITE_cache_reclaim);
 255                 else
 256                         __bch2_btree_node_write(c, b,
 257                                                 BTREE_WRITE_cache_reclaim);
 258
 259                 six_unlock_write(&b->c.lock);
 260                 six_unlock_intent(&b->c.lock);
 261                 goto wait_on_io;
 262         }
 263 out:
 264         if (b->hash_val && !ret)
 265                 trace_and_count(c, btree_cache_reap, c, b);
 266         return ret;
 267 out_unlock:
 268         six_unlock_write(&b->c.lock);
 269 out_unlock_intent:
 270         six_unlock_intent(&b->c.lock);
 271         ret = -BCH_ERR_ENOMEM_btree_node_reclaim;
 272         goto out;
 273 }
 274
 275 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 276 {
 277         return __btree_node_reclaim(c, b, false);
 278 }
 279
 280 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 281 {
 282         return __btree_node_reclaim(c, b, true);
 283 }
 284
 285 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 286                                            struct shrink_control *sc)
 287 {
 288         struct bch_fs *c = container_of(shrink, struct bch_fs,
 289                                         btree_cache.shrink);
 290         struct btree_cache *bc = &c->btree_cache;
 291         struct btree *b, *t;
 292         unsigned long nr = sc->nr_to_scan;
 293         unsigned long can_free = 0;
 294         unsigned long freed = 0;
 295         unsigned long touched = 0;
 296         unsigned i, flags;
 297         unsigned long ret = SHRINK_STOP;
 298         bool trigger_writes = atomic_read(&bc->dirty) + nr >=
 299                 bc->used * 3 / 4;
 300
 301         if (bch2_btree_shrinker_disabled)
 302                 return SHRINK_STOP;
 303
 304         mutex_lock(&bc->lock);
 305         flags = memalloc_nofs_save();
 306
 307         /*
 308          * It's _really_ critical that we don't free too many btree nodes - we
 309          * have to always leave ourselves a reserve. The reserve is how we
 310          * guarantee that allocating memory for a new btree node can always
 311          * succeed, so that inserting keys into the btree can always succeed and
 312          * IO can always make forward progress:
 313          */
 314         can_free = btree_cache_can_free(bc);
 315         nr = min_t(unsigned long, nr, can_free);
 316
 317         i = 0;
 318         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 319                 /*
 320                  * Leave a few nodes on the freeable list, so that a btree split
 321                  * won't have to hit the system allocator:
 322                  */
 323                 if (++i <= 3)
 324                         continue;
 325
 326                 touched++;
 327
 328                 if (touched >= nr)
 329                         goto out;
 330
 331                 if (!btree_node_reclaim(c, b)) {
 332                         btree_node_data_free(c, b);
 333                         six_unlock_write(&b->c.lock);
 334                         six_unlock_intent(&b->c.lock);
 335                         freed++;
 336                 }
 337         }
 338 restart:
 339         list_for_each_entry_safe(b, t, &bc->live, list) {
 340                 touched++;
 341
 342                 if (btree_node_accessed(b)) {
 343                         clear_btree_node_accessed(b);
 344                 } else if (!btree_node_reclaim(c, b)) {
 345                         freed++;
 346                         btree_node_data_free(c, b);
 347
 348                         bch2_btree_node_hash_remove(bc, b);
 349                         six_unlock_write(&b->c.lock);
 350                         six_unlock_intent(&b->c.lock);
 351
 352                         if (freed == nr)
 353                                 goto out_rotate;
 354                 } else if (trigger_writes &&
 355                            btree_node_dirty(b) &&
 356                            !btree_node_will_make_reachable(b) &&
 357                            !btree_node_write_blocked(b) &&
 358                            six_trylock_read(&b->c.lock)) {
 359                         list_move(&bc->live, &b->list);
 360                         mutex_unlock(&bc->lock);
 361                         __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
 362                         six_unlock_read(&b->c.lock);
 363                         if (touched >= nr)
 364                                 goto out_nounlock;
 365                         mutex_lock(&bc->lock);
 366                         goto restart;
 367                 }
 368
 369                 if (touched >= nr)
 370                         break;
 371         }
 372 out_rotate:
 373         if (&t->list != &bc->live)
 374                 list_move_tail(&bc->live, &t->list);
 375 out:
 376         mutex_unlock(&bc->lock);
 377 out_nounlock:
 378         ret = freed;
 379         memalloc_nofs_restore(flags);
 380         trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
 381         return ret;
 382 }
 383
 384 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 385                                             struct shrink_control *sc)
 386 {
 387         struct bch_fs *c = container_of(shrink, struct bch_fs,
 388                                         btree_cache.shrink);
 389         struct btree_cache *bc = &c->btree_cache;
 390
 391         if (bch2_btree_shrinker_disabled)
 392                 return 0;
 393
 394         return btree_cache_can_free(bc);
 395 }
 396
 397 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 398 {
 399         struct btree_cache *bc = &c->btree_cache;
 400         struct btree *b;
 401         unsigned i, flags;
 402
 403         unregister_shrinker(&bc->shrink);
 404
 405         /* vfree() can allocate memory: */
 406         flags = memalloc_nofs_save();
 407         mutex_lock(&bc->lock);
 408
 409         if (c->verify_data)
 410                 list_move(&c->verify_data->list, &bc->live);
 411
 412         kvpfree(c->verify_ondisk, btree_bytes(c));
 413
 414         for (i = 0; i < btree_id_nr_alive(c); i++) {
 415                 struct btree_root *r = bch2_btree_id_root(c, i);
 416
 417                 if (r->b)
 418                         list_add(&r->b->list, &bc->live);
 419         }
 420
 421         list_splice(&bc->freeable, &bc->live);
 422
 423         while (!list_empty(&bc->live)) {
 424                 b = list_first_entry(&bc->live, struct btree, list);
 425
 426                 BUG_ON(btree_node_read_in_flight(b) ||
 427                        btree_node_write_in_flight(b));
 428
 429                 if (btree_node_dirty(b))
 430                         bch2_btree_complete_write(c, b, btree_current_write(b));
 431                 clear_btree_node_dirty_acct(c, b);
 432
 433                 btree_node_data_free(c, b);
 434         }
 435
 436         BUG_ON(atomic_read(&c->btree_cache.dirty));
 437
 438         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 439
 440         while (!list_empty(&bc->freed_nonpcpu)) {
 441                 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
 442                 list_del(&b->list);
 443                 six_lock_exit(&b->c.lock);
 444                 kfree(b);
 445         }
 446
 447         mutex_unlock(&bc->lock);
 448         memalloc_nofs_restore(flags);
 449
 450         if (bc->table_init_done)
 451                 rhashtable_destroy(&bc->table);
 452 }
 453
 454 int bch2_fs_btree_cache_init(struct bch_fs *c)
 455 {
 456         struct btree_cache *bc = &c->btree_cache;
 457         unsigned i;
 458         int ret = 0;
 459
 460         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 461         if (ret)
 462                 goto err;
 463
 464         bc->table_init_done = true;
 465
 466         bch2_recalc_btree_reserve(c);
 467
 468         for (i = 0; i < bc->reserve; i++)
 469                 if (!__bch2_btree_node_mem_alloc(c))
 470                         goto err;
 471
 472         list_splice_init(&bc->live, &bc->freeable);
 473
 474         mutex_init(&c->verify_lock);
 475
 476         bc->shrink.count_objects        = bch2_btree_cache_count;
 477         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 478         bc->shrink.seeks                = 4;
 479         ret = register_shrinker(&bc->shrink, "%s-btree_cache", c->name);
 480         if (ret)
 481                 goto err;
 482
 483         return 0;
 484 err:
 485         return -BCH_ERR_ENOMEM_fs_btree_cache_init;
 486 }
 487
 488 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 489 {
 490         mutex_init(&bc->lock);
 491         INIT_LIST_HEAD(&bc->live);
 492         INIT_LIST_HEAD(&bc->freeable);
 493         INIT_LIST_HEAD(&bc->freed_pcpu);
 494         INIT_LIST_HEAD(&bc->freed_nonpcpu);
 495 }
 496
 497 /*
 498  * We can only have one thread cannibalizing other cached btree nodes at a time,
 499  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 500  * cannibalize_bucket() will take. This means every time we unlock the root of
 501  * the btree, we need to release this lock if we have it held.
 502  */
 503 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 504 {
 505         struct btree_cache *bc = &c->btree_cache;
 506
 507         if (bc->alloc_lock == current) {
 508                 trace_and_count(c, btree_cache_cannibalize_unlock, c);
 509                 bc->alloc_lock = NULL;
 510                 closure_wake_up(&bc->alloc_wait);
 511         }
 512 }
 513
 514 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 515 {
 516         struct btree_cache *bc = &c->btree_cache;
 517         struct task_struct *old;
 518
 519         old = cmpxchg(&bc->alloc_lock, NULL, current);
 520         if (old == NULL || old == current)
 521                 goto success;
 522
 523         if (!cl) {
 524                 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 525                 return -BCH_ERR_ENOMEM_btree_cache_cannibalize_lock;
 526         }
 527
 528         closure_wait(&bc->alloc_wait, cl);
 529
 530         /* Try again, after adding ourselves to waitlist */
 531         old = cmpxchg(&bc->alloc_lock, NULL, current);
 532         if (old == NULL || old == current) {
 533                 /* We raced */
 534                 closure_wake_up(&bc->alloc_wait);
 535                 goto success;
 536         }
 537
 538         trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 539         return -BCH_ERR_btree_cache_cannibalize_lock_blocked;
 540
 541 success:
 542         trace_and_count(c, btree_cache_cannibalize_lock, c);
 543         return 0;
 544 }
 545
 546 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 547 {
 548         struct btree_cache *bc = &c->btree_cache;
 549         struct btree *b;
 550
 551         list_for_each_entry_reverse(b, &bc->live, list)
 552                 if (!btree_node_reclaim(c, b))
 553                         return b;
 554
 555         while (1) {
 556                 list_for_each_entry_reverse(b, &bc->live, list)
 557                         if (!btree_node_write_and_reclaim(c, b))
 558                                 return b;
 559
 560                 /*
 561                  * Rare case: all nodes were intent-locked.
 562                  * Just busy-wait.
 563                  */
 564                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 565                 cond_resched();
 566         }
 567 }
 568
 569 struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_read_locks)
 570 {
 571         struct bch_fs *c = trans->c;
 572         struct btree_cache *bc = &c->btree_cache;
 573         struct list_head *freed = pcpu_read_locks
 574                 ? &bc->freed_pcpu
 575                 : &bc->freed_nonpcpu;
 576         struct btree *b, *b2;
 577         u64 start_time = local_clock();
 578         unsigned flags;
 579
 580         flags = memalloc_nofs_save();
 581         mutex_lock(&bc->lock);
 582
 583         /*
 584          * We never free struct btree itself, just the memory that holds the on
 585          * disk node. Check the freed list before allocating a new one:
 586          */
 587         list_for_each_entry(b, freed, list)
 588                 if (!btree_node_reclaim(c, b)) {
 589                         list_del_init(&b->list);
 590                         goto got_node;
 591                 }
 592
 593         b = __btree_node_mem_alloc(c, GFP_NOWAIT|__GFP_NOWARN);
 594         if (!b) {
 595                 mutex_unlock(&bc->lock);
 596                 bch2_trans_unlock(trans);
 597                 b = __btree_node_mem_alloc(c, GFP_KERNEL);
 598                 if (!b)
 599                         goto err;
 600                 mutex_lock(&bc->lock);
 601         }
 602
 603         bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0);
 604
 605         BUG_ON(!six_trylock_intent(&b->c.lock));
 606         BUG_ON(!six_trylock_write(&b->c.lock));
 607 got_node:
 608
 609         /*
 610          * btree_free() doesn't free memory; it sticks the node on the end of
 611          * the list. Check if there's any freed nodes there:
 612          */
 613         list_for_each_entry(b2, &bc->freeable, list)
 614                 if (!btree_node_reclaim(c, b2)) {
 615                         swap(b->data, b2->data);
 616                         swap(b->aux_data, b2->aux_data);
 617                         btree_node_to_freedlist(bc, b2);
 618                         six_unlock_write(&b2->c.lock);
 619                         six_unlock_intent(&b2->c.lock);
 620                         goto got_mem;
 621                 }
 622
 623         mutex_unlock(&bc->lock);
 624
 625         if (btree_node_data_alloc(c, b, GFP_NOWAIT|__GFP_NOWARN)) {
 626                 bch2_trans_unlock(trans);
 627                 if (btree_node_data_alloc(c, b, GFP_KERNEL|__GFP_NOWARN))
 628                         goto err;
 629         }
 630
 631         mutex_lock(&bc->lock);
 632         bc->used++;
 633 got_mem:
 634         mutex_unlock(&bc->lock);
 635
 636         BUG_ON(btree_node_hashed(b));
 637         BUG_ON(btree_node_dirty(b));
 638         BUG_ON(btree_node_write_in_flight(b));
 639 out:
 640         b->flags                = 0;
 641         b->written              = 0;
 642         b->nsets                = 0;
 643         b->sib_u64s[0]          = 0;
 644         b->sib_u64s[1]          = 0;
 645         b->whiteout_u64s        = 0;
 646         bch2_btree_keys_init(b);
 647         set_btree_node_accessed(b);
 648
 649         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 650                                start_time);
 651
 652         memalloc_nofs_restore(flags);
 653         return b;
 654 err:
 655         mutex_lock(&bc->lock);
 656
 657         /* Try to cannibalize another cached btree node: */
 658         if (bc->alloc_lock == current) {
 659                 b2 = btree_node_cannibalize(c);
 660                 clear_btree_node_just_written(b2);
 661                 bch2_btree_node_hash_remove(bc, b2);
 662
 663                 if (b) {
 664                         swap(b->data, b2->data);
 665                         swap(b->aux_data, b2->aux_data);
 666                         btree_node_to_freedlist(bc, b2);
 667                         six_unlock_write(&b2->c.lock);
 668                         six_unlock_intent(&b2->c.lock);
 669                 } else {
 670                         b = b2;
 671                         list_del_init(&b->list);
 672                 }
 673
 674                 mutex_unlock(&bc->lock);
 675
 676                 trace_and_count(c, btree_cache_cannibalize, c);
 677                 goto out;
 678         }
 679
 680         mutex_unlock(&bc->lock);
 681         memalloc_nofs_restore(flags);
 682         return ERR_PTR(-BCH_ERR_ENOMEM_btree_node_mem_alloc);
 683 }
 684
 685 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 686 static noinline struct btree *bch2_btree_node_fill(struct btree_trans *trans,
 687                                 struct btree_path *path,
 688                                 const struct bkey_i *k,
 689                                 enum btree_id btree_id,
 690                                 unsigned level,
 691                                 enum six_lock_type lock_type,
 692                                 bool sync)
 693 {
 694         struct bch_fs *c = trans->c;
 695         struct btree_cache *bc = &c->btree_cache;
 696         struct btree *b;
 697         u32 seq;
 698
 699         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 700         /*
 701          * Parent node must be locked, else we could read in a btree node that's
 702          * been freed:
 703          */
 704         if (path && !bch2_btree_node_relock(trans, path, level + 1)) {
 705                 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
 706                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
 707         }
 708
 709         b = bch2_btree_node_mem_alloc(trans, level != 0);
 710
 711         if (bch2_err_matches(PTR_ERR_OR_ZERO(b), ENOMEM)) {
 712                 trans->memory_allocation_failure = true;
 713                 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
 714                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
 715         }
 716
 717         if (IS_ERR(b))
 718                 return b;
 719
 720         /*
 721          * Btree nodes read in from disk should not have the accessed bit set
 722          * initially, so that linear scans don't thrash the cache:
 723          */
 724         clear_btree_node_accessed(b);
 725
 726         bkey_copy(&b->key, k);
 727         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 728                 /* raced with another fill: */
 729
 730                 /* mark as unhashed... */
 731                 b->hash_val = 0;
 732
 733                 mutex_lock(&bc->lock);
 734                 list_add(&b->list, &bc->freeable);
 735                 mutex_unlock(&bc->lock);
 736
 737                 six_unlock_write(&b->c.lock);
 738                 six_unlock_intent(&b->c.lock);
 739                 return NULL;
 740         }
 741
 742         set_btree_node_read_in_flight(b);
 743
 744         six_unlock_write(&b->c.lock);
 745         seq = six_lock_seq(&b->c.lock);
 746         six_unlock_intent(&b->c.lock);
 747
 748         /* Unlock before doing IO: */
 749         if (path && sync)
 750                 bch2_trans_unlock_noassert(trans);
 751
 752         bch2_btree_node_read(c, b, sync);
 753
 754         if (!sync)
 755                 return NULL;
 756
 757         if (path) {
 758                 int ret = bch2_trans_relock(trans) ?:
 759                         bch2_btree_path_relock_intent(trans, path);
 760                 if (ret) {
 761                         BUG_ON(!trans->restarted);
 762                         return ERR_PTR(ret);
 763                 }
 764         }
 765
 766         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 767                 if (path)
 768                         trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
 769                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
 770         }
 771
 772         return b;
 773 }
 774
 775 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 776 {
 777         struct printbuf buf = PRINTBUF;
 778
 779         if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_allocations)
 780                 return;
 781
 782         prt_printf(&buf,
 783                "btree node header doesn't match ptr\n"
 784                "btree %s level %u\n"
 785                "ptr: ",
 786                bch2_btree_id_str(b->c.btree_id), b->c.level);
 787         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 788
 789         prt_printf(&buf, "\nheader: btree %s level %llu\n"
 790                "min ",
 791                bch2_btree_id_str(BTREE_NODE_ID(b->data)),
 792                BTREE_NODE_LEVEL(b->data));
 793         bch2_bpos_to_text(&buf, b->data->min_key);
 794
 795         prt_printf(&buf, "\nmax ");
 796         bch2_bpos_to_text(&buf, b->data->max_key);
 797
 798         bch2_fs_inconsistent(c, "%s", buf.buf);
 799         printbuf_exit(&buf);
 800 }
 801
 802 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 803 {
 804         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 805             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 806             !bpos_eq(b->data->max_key, b->key.k.p) ||
 807             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 808              !bpos_eq(b->data->min_key,
 809                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 810                 btree_bad_header(c, b);
 811 }
 812
 813 static struct btree *__bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 814                                            const struct bkey_i *k, unsigned level,
 815                                            enum six_lock_type lock_type,
 816                                            unsigned long trace_ip)
 817 {
 818         struct bch_fs *c = trans->c;
 819         struct btree_cache *bc = &c->btree_cache;
 820         struct btree *b;
 821         struct bset_tree *t;
 822         bool need_relock = false;
 823         int ret;
 824
 825         EBUG_ON(level >= BTREE_MAX_DEPTH);
 826 retry:
 827         b = btree_cache_find(bc, k);
 828         if (unlikely(!b)) {
 829                 /*
 830                  * We must have the parent locked to call bch2_btree_node_fill(),
 831                  * else we could read in a btree node from disk that's been
 832                  * freed:
 833                  */
 834                 b = bch2_btree_node_fill(trans, path, k, path->btree_id,
 835                                          level, lock_type, true);
 836                 need_relock = true;
 837
 838                 /* We raced and found the btree node in the cache */
 839                 if (!b)
 840                         goto retry;
 841
 842                 if (IS_ERR(b))
 843                         return b;
 844         } else {
 845                 if (btree_node_read_locked(path, level + 1))
 846                         btree_node_unlock(trans, path, level + 1);
 847
 848                 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
 849                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 850                         return ERR_PTR(ret);
 851
 852                 BUG_ON(ret);
 853
 854                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 855                              b->c.level != level ||
 856                              race_fault())) {
 857                         six_unlock_type(&b->c.lock, lock_type);
 858                         if (bch2_btree_node_relock(trans, path, level + 1))
 859                                 goto retry;
 860
 861                         trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
 862                         return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
 863                 }
 864
 865                 /* avoid atomic set bit if it's not needed: */
 866                 if (!btree_node_accessed(b))
 867                         set_btree_node_accessed(b);
 868         }
 869
 870         if (unlikely(btree_node_read_in_flight(b))) {
 871                 u32 seq = six_lock_seq(&b->c.lock);
 872
 873                 six_unlock_type(&b->c.lock, lock_type);
 874                 bch2_trans_unlock(trans);
 875                 need_relock = true;
 876
 877                 bch2_btree_node_wait_on_read(b);
 878
 879                 /*
 880                  * should_be_locked is not set on this path yet, so we need to
 881                  * relock it specifically:
 882                  */
 883                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 884                         goto retry;
 885         }
 886
 887         if (unlikely(need_relock)) {
 888                 ret = bch2_trans_relock(trans) ?:
 889                         bch2_btree_path_relock_intent(trans, path);
 890                 if (ret) {
 891                         six_unlock_type(&b->c.lock, lock_type);
 892                         return ERR_PTR(ret);
 893                 }
 894         }
 895
 896         prefetch(b->aux_data);
 897
 898         for_each_bset(b, t) {
 899                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 900
 901                 prefetch(p + L1_CACHE_BYTES * 0);
 902                 prefetch(p + L1_CACHE_BYTES * 1);
 903                 prefetch(p + L1_CACHE_BYTES * 2);
 904         }
 905
 906         if (unlikely(btree_node_read_error(b))) {
 907                 six_unlock_type(&b->c.lock, lock_type);
 908                 return ERR_PTR(-EIO);
 909         }
 910
 911         EBUG_ON(b->c.btree_id != path->btree_id);
 912         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 913         btree_check_header(c, b);
 914
 915         return b;
 916 }
 917
 918 /**
 919  * bch2_btree_node_get - find a btree node in the cache and lock it, reading it
 920  * in from disk if necessary.
 921  *
 922  * @trans:      btree transaction object
 923  * @path:       btree_path being traversed
 924  * @k:          pointer to btree node (generally KEY_TYPE_btree_ptr_v2)
 925  * @level:      level of btree node being looked up (0 == leaf node)
 926  * @lock_type:  SIX_LOCK_read or SIX_LOCK_intent
 927  * @trace_ip:   ip of caller of btree iterator code (i.e. caller of bch2_btree_iter_peek())
 928  *
 929  * The btree node will have either a read or a write lock held, depending on
 930  * the @write parameter.
 931  *
 932  * Returns: btree node or ERR_PTR()
 933  */
 934 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 935                                   const struct bkey_i *k, unsigned level,
 936                                   enum six_lock_type lock_type,
 937                                   unsigned long trace_ip)
 938 {
 939         struct bch_fs *c = trans->c;
 940         struct btree *b;
 941         struct bset_tree *t;
 942         int ret;
 943
 944         EBUG_ON(level >= BTREE_MAX_DEPTH);
 945
 946         b = btree_node_mem_ptr(k);
 947
 948         /*
 949          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 950          * be the node we want anymore, and trying to lock the wrong node could
 951          * cause an unneccessary transaction restart:
 952          */
 953         if (unlikely(!c->opts.btree_node_mem_ptr_optimization ||
 954                      !b ||
 955                      b->hash_val != btree_ptr_hash_val(k)))
 956                 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
 957
 958         if (btree_node_read_locked(path, level + 1))
 959                 btree_node_unlock(trans, path, level + 1);
 960
 961         ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
 962         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 963                 return ERR_PTR(ret);
 964
 965         BUG_ON(ret);
 966
 967         if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 968                      b->c.level != level ||
 969                      race_fault())) {
 970                 six_unlock_type(&b->c.lock, lock_type);
 971                 if (bch2_btree_node_relock(trans, path, level + 1))
 972                         return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
 973
 974                 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
 975                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
 976         }
 977
 978         if (unlikely(btree_node_read_in_flight(b))) {
 979                 six_unlock_type(&b->c.lock, lock_type);
 980                 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
 981         }
 982
 983         prefetch(b->aux_data);
 984
 985         for_each_bset(b, t) {
 986                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 987
 988                 prefetch(p + L1_CACHE_BYTES * 0);
 989                 prefetch(p + L1_CACHE_BYTES * 1);
 990                 prefetch(p + L1_CACHE_BYTES * 2);
 991         }
 992
 993         /* avoid atomic set bit if it's not needed: */
 994         if (!btree_node_accessed(b))
 995                 set_btree_node_accessed(b);
 996
 997         if (unlikely(btree_node_read_error(b))) {
 998                 six_unlock_type(&b->c.lock, lock_type);
 999                 return ERR_PTR(-EIO);
1000         }
1001
1002         EBUG_ON(b->c.btree_id != path->btree_id);
1003         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1004         btree_check_header(c, b);
1005
1006         return b;
1007 }
1008
1009 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
1010                                          const struct bkey_i *k,
1011                                          enum btree_id btree_id,
1012                                          unsigned level,
1013                                          bool nofill)
1014 {
1015         struct bch_fs *c = trans->c;
1016         struct btree_cache *bc = &c->btree_cache;
1017         struct btree *b;
1018         struct bset_tree *t;
1019         int ret;
1020
1021         EBUG_ON(level >= BTREE_MAX_DEPTH);
1022
1023         if (c->opts.btree_node_mem_ptr_optimization) {
1024                 b = btree_node_mem_ptr(k);
1025                 if (b)
1026                         goto lock_node;
1027         }
1028 retry:
1029         b = btree_cache_find(bc, k);
1030         if (unlikely(!b)) {
1031                 if (nofill)
1032                         goto out;
1033
1034                 b = bch2_btree_node_fill(trans, NULL, k, btree_id,
1035                                          level, SIX_LOCK_read, true);
1036
1037                 /* We raced and found the btree node in the cache */
1038                 if (!b)
1039                         goto retry;
1040
1041                 if (IS_ERR(b) &&
1042                     !bch2_btree_cache_cannibalize_lock(c, NULL))
1043                         goto retry;
1044
1045                 if (IS_ERR(b))
1046                         goto out;
1047         } else {
1048 lock_node:
1049                 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read, _THIS_IP_);
1050                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1051                         return ERR_PTR(ret);
1052
1053                 BUG_ON(ret);
1054
1055                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1056                              b->c.btree_id != btree_id ||
1057                              b->c.level != level)) {
1058                         six_unlock_read(&b->c.lock);
1059                         goto retry;
1060                 }
1061         }
1062
1063         /* XXX: waiting on IO with btree locks held: */
1064         __bch2_btree_node_wait_on_read(b);
1065
1066         prefetch(b->aux_data);
1067
1068         for_each_bset(b, t) {
1069                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1070
1071                 prefetch(p + L1_CACHE_BYTES * 0);
1072                 prefetch(p + L1_CACHE_BYTES * 1);
1073                 prefetch(p + L1_CACHE_BYTES * 2);
1074         }
1075
1076         /* avoid atomic set bit if it's not needed: */
1077         if (!btree_node_accessed(b))
1078                 set_btree_node_accessed(b);
1079
1080         if (unlikely(btree_node_read_error(b))) {
1081                 six_unlock_read(&b->c.lock);
1082                 b = ERR_PTR(-EIO);
1083                 goto out;
1084         }
1085
1086         EBUG_ON(b->c.btree_id != btree_id);
1087         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1088         btree_check_header(c, b);
1089 out:
1090         bch2_btree_cache_cannibalize_unlock(c);
1091         return b;
1092 }
1093
1094 int bch2_btree_node_prefetch(struct btree_trans *trans,
1095                              struct btree_path *path,
1096                              const struct bkey_i *k,
1097                              enum btree_id btree_id, unsigned level)
1098 {
1099         struct bch_fs *c = trans->c;
1100         struct btree_cache *bc = &c->btree_cache;
1101         struct btree *b;
1102
1103         BUG_ON(trans && !btree_node_locked(path, level + 1));
1104         BUG_ON(level >= BTREE_MAX_DEPTH);
1105
1106         b = btree_cache_find(bc, k);
1107         if (b)
1108                 return 0;
1109
1110         b = bch2_btree_node_fill(trans, path, k, btree_id,
1111                                  level, SIX_LOCK_read, false);
1112         return PTR_ERR_OR_ZERO(b);
1113 }
1114
1115 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1116 {
1117         struct bch_fs *c = trans->c;
1118         struct btree_cache *bc = &c->btree_cache;
1119         struct btree *b;
1120
1121         b = btree_cache_find(bc, k);
1122         if (!b)
1123                 return;
1124 wait_on_io:
1125         /* not allowed to wait on io with btree locks held: */
1126
1127         /* XXX we're called from btree_gc which will be holding other btree
1128          * nodes locked
1129          */
1130         __bch2_btree_node_wait_on_read(b);
1131         __bch2_btree_node_wait_on_write(b);
1132
1133         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1134         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1135
1136         if (btree_node_dirty(b)) {
1137                 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
1138                 six_unlock_write(&b->c.lock);
1139                 six_unlock_intent(&b->c.lock);
1140                 goto wait_on_io;
1141         }
1142
1143         BUG_ON(btree_node_dirty(b));
1144
1145         mutex_lock(&bc->lock);
1146         btree_node_data_free(c, b);
1147         bch2_btree_node_hash_remove(bc, b);
1148         mutex_unlock(&bc->lock);
1149
1150         six_unlock_write(&b->c.lock);
1151         six_unlock_intent(&b->c.lock);
1152 }
1153
1154 const char *bch2_btree_id_str(enum btree_id btree)
1155 {
1156         return btree < BTREE_ID_NR ? __bch2_btree_ids[btree] : "(unknown)";
1157 }
1158
1159 void bch2_btree_pos_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
1160 {
1161         prt_printf(out, "%s level %u/%u\n  ",
1162                bch2_btree_id_str(b->c.btree_id),
1163                b->c.level,
1164                bch2_btree_id_root(c, b->c.btree_id)->level);
1165         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1166 }
1167
1168 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
1169 {
1170         struct bset_stats stats;
1171
1172         memset(&stats, 0, sizeof(stats));
1173
1174         bch2_btree_keys_stats(b, &stats);
1175
1176         prt_printf(out, "l %u ", b->c.level);
1177         bch2_bpos_to_text(out, b->data->min_key);
1178         prt_printf(out, " - ");
1179         bch2_bpos_to_text(out, b->data->max_key);
1180         prt_printf(out, ":\n"
1181                "    ptrs: ");
1182         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1183         prt_newline(out);
1184
1185         prt_printf(out,
1186                "    format: ");
1187         bch2_bkey_format_to_text(out, &b->format);
1188
1189         prt_printf(out,
1190                "    unpack fn len: %u\n"
1191                "    bytes used %zu/%zu (%zu%% full)\n"
1192                "    sib u64s: %u, %u (merge threshold %u)\n"
1193                "    nr packed keys %u\n"
1194                "    nr unpacked keys %u\n"
1195                "    floats %zu\n"
1196                "    failed unpacked %zu\n",
1197                b->unpack_fn_len,
1198                b->nr.live_u64s * sizeof(u64),
1199                btree_bytes(c) - sizeof(struct btree_node),
1200                b->nr.live_u64s * 100 / btree_max_u64s(c),
1201                b->sib_u64s[0],
1202                b->sib_u64s[1],
1203                c->btree_foreground_merge_threshold,
1204                b->nr.packed_keys,
1205                b->nr.unpacked_keys,
1206                stats.floats,
1207                stats.failed);
1208 }
1209
1210 void bch2_btree_cache_to_text(struct printbuf *out, const struct bch_fs *c)
1211 {
1212         prt_printf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1213         prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1214         prt_printf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1215 }