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 "btree_cache.h"
   5 #include "btree_io.h"
   6 #include "btree_iter.h"
   7 #include "btree_locking.h"
   8 #include "debug.h"
   9
  10 #include <linux/prefetch.h>
  11 #include <linux/sched/mm.h>
  12 #include <trace/events/bcachefs.h>
  13
  14 const char * const bch2_btree_ids[] = {
  15 #define x(kwd, val, name) name,
  16         BCH_BTREE_IDS()
  17 #undef x
  18         NULL
  19 };
  20
  21 void bch2_recalc_btree_reserve(struct bch_fs *c)
  22 {
  23         unsigned i, reserve = 16;
  24
  25         if (!c->btree_roots[0].b)
  26                 reserve += 8;
  27
  28         for (i = 0; i < BTREE_ID_NR; i++)
  29                 if (c->btree_roots[i].b)
  30                         reserve += min_t(unsigned, 1,
  31                                          c->btree_roots[i].b->level) * 8;
  32
  33         c->btree_cache.reserve = reserve;
  34 }
  35
  36 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  37 {
  38         return max_t(int, 0, bc->used - bc->reserve);
  39 }
  40
  41 static void __btree_node_data_free(struct bch_fs *c, struct btree *b)
  42 {
  43         EBUG_ON(btree_node_write_in_flight(b));
  44
  45         kvpfree(b->data, btree_bytes(c));
  46         b->data = NULL;
  47         bch2_btree_keys_free(b);
  48 }
  49
  50 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  51 {
  52         struct btree_cache *bc = &c->btree_cache;
  53
  54         __btree_node_data_free(c, b);
  55         bc->used--;
  56         list_move(&b->list, &bc->freed);
  57 }
  58
  59 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  60                                    const void *obj)
  61 {
  62         const struct btree *b = obj;
  63         const u64 *v = arg->key;
  64
  65         return b->hash_val == *v ? 0 : 1;
  66 }
  67
  68 static const struct rhashtable_params bch_btree_cache_params = {
  69         .head_offset    = offsetof(struct btree, hash),
  70         .key_offset     = offsetof(struct btree, hash_val),
  71         .key_len        = sizeof(u64),
  72         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  73 };
  74
  75 static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  76 {
  77         struct btree_cache *bc = &c->btree_cache;
  78
  79         b->data = kvpmalloc(btree_bytes(c), gfp);
  80         if (!b->data)
  81                 goto err;
  82
  83         if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp))
  84                 goto err;
  85
  86         bc->used++;
  87         list_move(&b->list, &bc->freeable);
  88         return;
  89 err:
  90         kvpfree(b->data, btree_bytes(c));
  91         b->data = NULL;
  92         list_move(&b->list, &bc->freed);
  93 }
  94
  95 static struct btree *btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
  96 {
  97         struct btree *b = kzalloc(sizeof(struct btree), gfp);
  98         if (!b)
  99                 return NULL;
 100
 101         bkey_btree_ptr_init(&b->key);
 102         six_lock_init(&b->lock);
 103         INIT_LIST_HEAD(&b->list);
 104         INIT_LIST_HEAD(&b->write_blocked);
 105
 106         btree_node_data_alloc(c, b, gfp);
 107         return b->data ? b : NULL;
 108 }
 109
 110 /* Btree in memory cache - hash table */
 111
 112 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 113 {
 114         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 115
 116         /* Cause future lookups for this node to fail: */
 117         b->hash_val = 0;
 118 }
 119
 120 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 121 {
 122         BUG_ON(b->hash_val);
 123         b->hash_val = btree_ptr_hash_val(&b->key);
 124
 125         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 126                                              bch_btree_cache_params);
 127 }
 128
 129 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 130                                 unsigned level, enum btree_id id)
 131 {
 132         int ret;
 133
 134         b->level        = level;
 135         b->btree_id     = id;
 136
 137         mutex_lock(&bc->lock);
 138         ret = __bch2_btree_node_hash_insert(bc, b);
 139         if (!ret)
 140                 list_add(&b->list, &bc->live);
 141         mutex_unlock(&bc->lock);
 142
 143         return ret;
 144 }
 145
 146 __flatten
 147 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 148                                      const struct bkey_i *k)
 149 {
 150         u64 v = btree_ptr_hash_val(k);
 151
 152         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 153 }
 154
 155 /*
 156  * this version is for btree nodes that have already been freed (we're not
 157  * reaping a real btree node)
 158  */
 159 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 160 {
 161         struct btree_cache *bc = &c->btree_cache;
 162         int ret = 0;
 163
 164         lockdep_assert_held(&bc->lock);
 165
 166         if (!six_trylock_intent(&b->lock))
 167                 return -ENOMEM;
 168
 169         if (!six_trylock_write(&b->lock))
 170                 goto out_unlock_intent;
 171
 172         if (btree_node_noevict(b))
 173                 goto out_unlock;
 174
 175         if (!btree_node_may_write(b))
 176                 goto out_unlock;
 177
 178         if (btree_node_dirty(b) &&
 179             test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 180                 goto out_unlock;
 181
 182         if (btree_node_dirty(b) ||
 183             btree_node_write_in_flight(b) ||
 184             btree_node_read_in_flight(b)) {
 185                 if (!flush)
 186                         goto out_unlock;
 187
 188                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 189                                TASK_UNINTERRUPTIBLE);
 190
 191                 /*
 192                  * Using the underscore version because we don't want to compact
 193                  * bsets after the write, since this node is about to be evicted
 194                  * - unless btree verify mode is enabled, since it runs out of
 195                  * the post write cleanup:
 196                  */
 197                 if (verify_btree_ondisk(c))
 198                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 199                 else
 200                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 201
 202                 /* wait for any in flight btree write */
 203                 btree_node_wait_on_io(b);
 204         }
 205 out:
 206         if (b->hash_val && !ret)
 207                 trace_btree_node_reap(c, b);
 208         return ret;
 209 out_unlock:
 210         six_unlock_write(&b->lock);
 211 out_unlock_intent:
 212         six_unlock_intent(&b->lock);
 213         ret = -ENOMEM;
 214         goto out;
 215 }
 216
 217 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 218 {
 219         return __btree_node_reclaim(c, b, false);
 220 }
 221
 222 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 223 {
 224         return __btree_node_reclaim(c, b, true);
 225 }
 226
 227 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 228                                            struct shrink_control *sc)
 229 {
 230         struct bch_fs *c = container_of(shrink, struct bch_fs,
 231                                         btree_cache.shrink);
 232         struct btree_cache *bc = &c->btree_cache;
 233         struct btree *b, *t;
 234         unsigned long nr = sc->nr_to_scan;
 235         unsigned long can_free;
 236         unsigned long touched = 0;
 237         unsigned long freed = 0;
 238         unsigned i;
 239
 240         if (btree_shrinker_disabled(c))
 241                 return SHRINK_STOP;
 242
 243         /* Return -1 if we can't do anything right now */
 244         if (sc->gfp_mask & __GFP_IO)
 245                 mutex_lock(&bc->lock);
 246         else if (!mutex_trylock(&bc->lock))
 247                 return -1;
 248
 249         /*
 250          * It's _really_ critical that we don't free too many btree nodes - we
 251          * have to always leave ourselves a reserve. The reserve is how we
 252          * guarantee that allocating memory for a new btree node can always
 253          * succeed, so that inserting keys into the btree can always succeed and
 254          * IO can always make forward progress:
 255          */
 256         nr /= btree_pages(c);
 257         can_free = btree_cache_can_free(bc);
 258         nr = min_t(unsigned long, nr, can_free);
 259
 260         i = 0;
 261         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 262                 touched++;
 263
 264                 if (freed >= nr)
 265                         break;
 266
 267                 if (++i > 3 &&
 268                     !btree_node_reclaim(c, b)) {
 269                         btree_node_data_free(c, b);
 270                         six_unlock_write(&b->lock);
 271                         six_unlock_intent(&b->lock);
 272                         freed++;
 273                 }
 274         }
 275 restart:
 276         list_for_each_entry_safe(b, t, &bc->live, list) {
 277                 touched++;
 278
 279                 if (freed >= nr) {
 280                         /* Save position */
 281                         if (&t->list != &bc->live)
 282                                 list_move_tail(&bc->live, &t->list);
 283                         break;
 284                 }
 285
 286                 if (!btree_node_accessed(b) &&
 287                     !btree_node_reclaim(c, b)) {
 288                         /* can't call bch2_btree_node_hash_remove under lock  */
 289                         freed++;
 290                         if (&t->list != &bc->live)
 291                                 list_move_tail(&bc->live, &t->list);
 292
 293                         btree_node_data_free(c, b);
 294                         mutex_unlock(&bc->lock);
 295
 296                         bch2_btree_node_hash_remove(bc, b);
 297                         six_unlock_write(&b->lock);
 298                         six_unlock_intent(&b->lock);
 299
 300                         if (freed >= nr)
 301                                 goto out;
 302
 303                         if (sc->gfp_mask & __GFP_IO)
 304                                 mutex_lock(&bc->lock);
 305                         else if (!mutex_trylock(&bc->lock))
 306                                 goto out;
 307                         goto restart;
 308                 } else
 309                         clear_btree_node_accessed(b);
 310         }
 311
 312         mutex_unlock(&bc->lock);
 313 out:
 314         return (unsigned long) freed * btree_pages(c);
 315 }
 316
 317 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 318                                             struct shrink_control *sc)
 319 {
 320         struct bch_fs *c = container_of(shrink, struct bch_fs,
 321                                         btree_cache.shrink);
 322         struct btree_cache *bc = &c->btree_cache;
 323
 324         if (btree_shrinker_disabled(c))
 325                 return 0;
 326
 327         return btree_cache_can_free(bc) * btree_pages(c);
 328 }
 329
 330 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 331 {
 332         struct btree_cache *bc = &c->btree_cache;
 333         struct btree *b;
 334         unsigned i;
 335
 336         if (bc->shrink.list.next)
 337                 unregister_shrinker(&bc->shrink);
 338
 339         mutex_lock(&bc->lock);
 340
 341 #ifdef CONFIG_BCACHEFS_DEBUG
 342         if (c->verify_data)
 343                 list_move(&c->verify_data->list, &bc->live);
 344
 345         kvpfree(c->verify_ondisk, btree_bytes(c));
 346 #endif
 347
 348         for (i = 0; i < BTREE_ID_NR; i++)
 349                 if (c->btree_roots[i].b)
 350                         list_add(&c->btree_roots[i].b->list, &bc->live);
 351
 352         list_splice(&bc->freeable, &bc->live);
 353
 354         while (!list_empty(&bc->live)) {
 355                 b = list_first_entry(&bc->live, struct btree, list);
 356
 357                 BUG_ON(btree_node_read_in_flight(b) ||
 358                        btree_node_write_in_flight(b));
 359
 360                 if (btree_node_dirty(b))
 361                         bch2_btree_complete_write(c, b, btree_current_write(b));
 362                 clear_btree_node_dirty(b);
 363
 364                 btree_node_data_free(c, b);
 365         }
 366
 367         while (!list_empty(&bc->freed)) {
 368                 b = list_first_entry(&bc->freed, struct btree, list);
 369                 list_del(&b->list);
 370                 kfree(b);
 371         }
 372
 373         mutex_unlock(&bc->lock);
 374
 375         if (bc->table_init_done)
 376                 rhashtable_destroy(&bc->table);
 377 }
 378
 379 int bch2_fs_btree_cache_init(struct bch_fs *c)
 380 {
 381         struct btree_cache *bc = &c->btree_cache;
 382         unsigned i;
 383         int ret = 0;
 384
 385         pr_verbose_init(c->opts, "");
 386
 387         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 388         if (ret)
 389                 goto out;
 390
 391         bc->table_init_done = true;
 392
 393         bch2_recalc_btree_reserve(c);
 394
 395         for (i = 0; i < bc->reserve; i++)
 396                 if (!btree_node_mem_alloc(c, GFP_KERNEL)) {
 397                         ret = -ENOMEM;
 398                         goto out;
 399                 }
 400
 401         list_splice_init(&bc->live, &bc->freeable);
 402
 403 #ifdef CONFIG_BCACHEFS_DEBUG
 404         mutex_init(&c->verify_lock);
 405
 406         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 407         if (!c->verify_ondisk) {
 408                 ret = -ENOMEM;
 409                 goto out;
 410         }
 411
 412         c->verify_data = btree_node_mem_alloc(c, GFP_KERNEL);
 413         if (!c->verify_data) {
 414                 ret = -ENOMEM;
 415                 goto out;
 416         }
 417
 418         list_del_init(&c->verify_data->list);
 419 #endif
 420
 421         bc->shrink.count_objects        = bch2_btree_cache_count;
 422         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 423         bc->shrink.seeks                = 4;
 424         bc->shrink.batch                = btree_pages(c) * 2;
 425         register_shrinker(&bc->shrink);
 426 out:
 427         pr_verbose_init(c->opts, "ret %i", ret);
 428         return ret;
 429 }
 430
 431 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 432 {
 433         mutex_init(&bc->lock);
 434         INIT_LIST_HEAD(&bc->live);
 435         INIT_LIST_HEAD(&bc->freeable);
 436         INIT_LIST_HEAD(&bc->freed);
 437 }
 438
 439 /*
 440  * We can only have one thread cannibalizing other cached btree nodes at a time,
 441  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 442  * cannibalize_bucket() will take. This means every time we unlock the root of
 443  * the btree, we need to release this lock if we have it held.
 444  */
 445 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 446 {
 447         struct btree_cache *bc = &c->btree_cache;
 448
 449         if (bc->alloc_lock == current) {
 450                 trace_btree_node_cannibalize_unlock(c);
 451                 bc->alloc_lock = NULL;
 452                 closure_wake_up(&bc->alloc_wait);
 453         }
 454 }
 455
 456 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 457 {
 458         struct btree_cache *bc = &c->btree_cache;
 459         struct task_struct *old;
 460
 461         old = cmpxchg(&bc->alloc_lock, NULL, current);
 462         if (old == NULL || old == current)
 463                 goto success;
 464
 465         if (!cl) {
 466                 trace_btree_node_cannibalize_lock_fail(c);
 467                 return -ENOMEM;
 468         }
 469
 470         closure_wait(&bc->alloc_wait, cl);
 471
 472         /* Try again, after adding ourselves to waitlist */
 473         old = cmpxchg(&bc->alloc_lock, NULL, current);
 474         if (old == NULL || old == current) {
 475                 /* We raced */
 476                 closure_wake_up(&bc->alloc_wait);
 477                 goto success;
 478         }
 479
 480         trace_btree_node_cannibalize_lock_fail(c);
 481         return -EAGAIN;
 482
 483 success:
 484         trace_btree_node_cannibalize_lock(c);
 485         return 0;
 486 }
 487
 488 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 489 {
 490         struct btree_cache *bc = &c->btree_cache;
 491         struct btree *b;
 492
 493         list_for_each_entry_reverse(b, &bc->live, list)
 494                 if (!btree_node_reclaim(c, b))
 495                         return b;
 496
 497         while (1) {
 498                 list_for_each_entry_reverse(b, &bc->live, list)
 499                         if (!btree_node_write_and_reclaim(c, b))
 500                                 return b;
 501
 502                 /*
 503                  * Rare case: all nodes were intent-locked.
 504                  * Just busy-wait.
 505                  */
 506                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 507                 cond_resched();
 508         }
 509 }
 510
 511 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 512 {
 513         struct btree_cache *bc = &c->btree_cache;
 514         struct btree *b;
 515         u64 start_time = local_clock();
 516         unsigned flags;
 517
 518         flags = memalloc_nofs_save();
 519         mutex_lock(&bc->lock);
 520
 521         /*
 522          * btree_free() doesn't free memory; it sticks the node on the end of
 523          * the list. Check if there's any freed nodes there:
 524          */
 525         list_for_each_entry(b, &bc->freeable, list)
 526                 if (!btree_node_reclaim(c, b))
 527                         goto out_unlock;
 528
 529         /*
 530          * We never free struct btree itself, just the memory that holds the on
 531          * disk node. Check the freed list before allocating a new one:
 532          */
 533         list_for_each_entry(b, &bc->freed, list)
 534                 if (!btree_node_reclaim(c, b)) {
 535                         btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
 536                         if (b->data)
 537                                 goto out_unlock;
 538
 539                         six_unlock_write(&b->lock);
 540                         six_unlock_intent(&b->lock);
 541                         goto err;
 542                 }
 543
 544         b = btree_node_mem_alloc(c, __GFP_NOWARN|GFP_NOIO);
 545         if (!b)
 546                 goto err;
 547
 548         BUG_ON(!six_trylock_intent(&b->lock));
 549         BUG_ON(!six_trylock_write(&b->lock));
 550 out_unlock:
 551         BUG_ON(btree_node_hashed(b));
 552         BUG_ON(btree_node_write_in_flight(b));
 553
 554         list_del_init(&b->list);
 555         mutex_unlock(&bc->lock);
 556         memalloc_nofs_restore(flags);
 557 out:
 558         b->flags                = 0;
 559         b->written              = 0;
 560         b->nsets                = 0;
 561         b->sib_u64s[0]          = 0;
 562         b->sib_u64s[1]          = 0;
 563         b->whiteout_u64s        = 0;
 564         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 565
 566         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 567                                start_time);
 568
 569         return b;
 570 err:
 571         /* Try to cannibalize another cached btree node: */
 572         if (bc->alloc_lock == current) {
 573                 b = btree_node_cannibalize(c);
 574                 list_del_init(&b->list);
 575                 mutex_unlock(&bc->lock);
 576
 577                 bch2_btree_node_hash_remove(bc, b);
 578
 579                 trace_btree_node_cannibalize(c);
 580                 goto out;
 581         }
 582
 583         mutex_unlock(&bc->lock);
 584         return ERR_PTR(-ENOMEM);
 585 }
 586
 587 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 588 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 589                                 struct btree_iter *iter,
 590                                 const struct bkey_i *k,
 591                                 enum btree_id btree_id,
 592                                 unsigned level,
 593                                 enum six_lock_type lock_type,
 594                                 bool sync)
 595 {
 596         struct btree_cache *bc = &c->btree_cache;
 597         struct btree *b;
 598
 599         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 600         /*
 601          * Parent node must be locked, else we could read in a btree node that's
 602          * been freed:
 603          */
 604         if (iter && !bch2_btree_node_relock(iter, level + 1))
 605                 return ERR_PTR(-EINTR);
 606
 607         b = bch2_btree_node_mem_alloc(c);
 608         if (IS_ERR(b))
 609                 return b;
 610
 611         bkey_copy(&b->key, k);
 612         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 613                 /* raced with another fill: */
 614
 615                 /* mark as unhashed... */
 616                 b->hash_val = 0;
 617
 618                 mutex_lock(&bc->lock);
 619                 list_add(&b->list, &bc->freeable);
 620                 mutex_unlock(&bc->lock);
 621
 622                 six_unlock_write(&b->lock);
 623                 six_unlock_intent(&b->lock);
 624                 return NULL;
 625         }
 626
 627         /*
 628          * Unlock before doing IO:
 629          *
 630          * XXX: ideally should be dropping all btree node locks here
 631          */
 632         if (iter && btree_node_read_locked(iter, level + 1))
 633                 btree_node_unlock(iter, level + 1);
 634
 635         bch2_btree_node_read(c, b, sync);
 636
 637         six_unlock_write(&b->lock);
 638
 639         if (!sync) {
 640                 six_unlock_intent(&b->lock);
 641                 return NULL;
 642         }
 643
 644         if (lock_type == SIX_LOCK_read)
 645                 six_lock_downgrade(&b->lock);
 646
 647         return b;
 648 }
 649
 650 /**
 651  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 652  * in from disk if necessary.
 653  *
 654  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 655  *
 656  * The btree node will have either a read or a write lock held, depending on
 657  * the @write parameter.
 658  */
 659 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 660                                   const struct bkey_i *k, unsigned level,
 661                                   enum six_lock_type lock_type)
 662 {
 663         struct btree_cache *bc = &c->btree_cache;
 664         struct btree *b;
 665         struct bset_tree *t;
 666
 667         EBUG_ON(level >= BTREE_MAX_DEPTH);
 668
 669         b = btree_node_mem_ptr(k);
 670         if (b)
 671                 goto lock_node;
 672 retry:
 673         b = btree_cache_find(bc, k);
 674         if (unlikely(!b)) {
 675                 /*
 676                  * We must have the parent locked to call bch2_btree_node_fill(),
 677                  * else we could read in a btree node from disk that's been
 678                  * freed:
 679                  */
 680                 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
 681                                          level, lock_type, true);
 682
 683                 /* We raced and found the btree node in the cache */
 684                 if (!b)
 685                         goto retry;
 686
 687                 if (IS_ERR(b))
 688                         return b;
 689         } else {
 690 lock_node:
 691                 /*
 692                  * There's a potential deadlock with splits and insertions into
 693                  * interior nodes we have to avoid:
 694                  *
 695                  * The other thread might be holding an intent lock on the node
 696                  * we want, and they want to update its parent node so they're
 697                  * going to upgrade their intent lock on the parent node to a
 698                  * write lock.
 699                  *
 700                  * But if we're holding a read lock on the parent, and we're
 701                  * trying to get the intent lock they're holding, we deadlock.
 702                  *
 703                  * So to avoid this we drop the read locks on parent nodes when
 704                  * we're starting to take intent locks - and handle the race.
 705                  *
 706                  * The race is that they might be about to free the node we
 707                  * want, and dropping our read lock on the parent node lets them
 708                  * update the parent marking the node we want as freed, and then
 709                  * free it:
 710                  *
 711                  * To guard against this, btree nodes are evicted from the cache
 712                  * when they're freed - and b->hash_val is zeroed out, which we
 713                  * check for after we lock the node.
 714                  *
 715                  * Then, bch2_btree_node_relock() on the parent will fail - because
 716                  * the parent was modified, when the pointer to the node we want
 717                  * was removed - and we'll bail out:
 718                  */
 719                 if (btree_node_read_locked(iter, level + 1))
 720                         btree_node_unlock(iter, level + 1);
 721
 722                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
 723                         return ERR_PTR(-EINTR);
 724
 725                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 726                              b->level != level ||
 727                              race_fault())) {
 728                         six_unlock_type(&b->lock, lock_type);
 729                         if (bch2_btree_node_relock(iter, level + 1))
 730                                 goto retry;
 731
 732                         trace_trans_restart_btree_node_reused(iter->trans->ip);
 733                         return ERR_PTR(-EINTR);
 734                 }
 735         }
 736
 737         /* XXX: waiting on IO with btree locks held: */
 738         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 739                        TASK_UNINTERRUPTIBLE);
 740
 741         prefetch(b->aux_data);
 742
 743         for_each_bset(b, t) {
 744                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 745
 746                 prefetch(p + L1_CACHE_BYTES * 0);
 747                 prefetch(p + L1_CACHE_BYTES * 1);
 748                 prefetch(p + L1_CACHE_BYTES * 2);
 749         }
 750
 751         /* avoid atomic set bit if it's not needed: */
 752         if (!btree_node_accessed(b))
 753                 set_btree_node_accessed(b);
 754
 755         if (unlikely(btree_node_read_error(b))) {
 756                 six_unlock_type(&b->lock, lock_type);
 757                 return ERR_PTR(-EIO);
 758         }
 759
 760         EBUG_ON(b->btree_id != iter->btree_id ||
 761                 BTREE_NODE_LEVEL(b->data) != level ||
 762                 bkey_cmp(b->data->max_key, k->k.p));
 763
 764         return b;
 765 }
 766
 767 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 768                                          const struct bkey_i *k,
 769                                          enum btree_id btree_id,
 770                                          unsigned level)
 771 {
 772         struct btree_cache *bc = &c->btree_cache;
 773         struct btree *b;
 774         struct bset_tree *t;
 775
 776         EBUG_ON(level >= BTREE_MAX_DEPTH);
 777
 778         b = btree_node_mem_ptr(k);
 779         if (b)
 780                 goto lock_node;
 781 retry:
 782         b = btree_cache_find(bc, k);
 783         if (unlikely(!b)) {
 784                 b = bch2_btree_node_fill(c, NULL, k, btree_id,
 785                                          level, SIX_LOCK_read, true);
 786
 787                 /* We raced and found the btree node in the cache */
 788                 if (!b)
 789                         goto retry;
 790
 791                 if (IS_ERR(b))
 792                         return b;
 793         } else {
 794 lock_node:
 795                 six_lock_read(&b->lock);
 796
 797                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 798                              b->btree_id != btree_id ||
 799                              b->level != level)) {
 800                         six_unlock_read(&b->lock);
 801                         goto retry;
 802                 }
 803         }
 804
 805         /* XXX: waiting on IO with btree locks held: */
 806         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 807                        TASK_UNINTERRUPTIBLE);
 808
 809         prefetch(b->aux_data);
 810
 811         for_each_bset(b, t) {
 812                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 813
 814                 prefetch(p + L1_CACHE_BYTES * 0);
 815                 prefetch(p + L1_CACHE_BYTES * 1);
 816                 prefetch(p + L1_CACHE_BYTES * 2);
 817         }
 818
 819         /* avoid atomic set bit if it's not needed: */
 820         if (!btree_node_accessed(b))
 821                 set_btree_node_accessed(b);
 822
 823         if (unlikely(btree_node_read_error(b))) {
 824                 six_unlock_read(&b->lock);
 825                 return ERR_PTR(-EIO);
 826         }
 827
 828         EBUG_ON(b->btree_id != btree_id ||
 829                 BTREE_NODE_LEVEL(b->data) != level ||
 830                 bkey_cmp(b->data->max_key, k->k.p));
 831
 832         return b;
 833 }
 834
 835 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 836                                           struct btree_iter *iter,
 837                                           struct btree *b,
 838                                           enum btree_node_sibling sib)
 839 {
 840         struct btree_trans *trans = iter->trans;
 841         struct btree *parent;
 842         struct btree_node_iter node_iter;
 843         struct bkey_packed *k;
 844         BKEY_PADDED(k) tmp;
 845         struct btree *ret = NULL;
 846         unsigned level = b->level;
 847
 848         parent = btree_iter_node(iter, level + 1);
 849         if (!parent)
 850                 return NULL;
 851
 852         if (!bch2_btree_node_relock(iter, level + 1)) {
 853                 ret = ERR_PTR(-EINTR);
 854                 goto out;
 855         }
 856
 857         node_iter = iter->l[parent->level].iter;
 858
 859         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 860         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 861
 862         k = sib == btree_prev_sib
 863                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 864                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 865                    bch2_btree_node_iter_peek(&node_iter, parent));
 866         if (!k)
 867                 goto out;
 868
 869         bch2_bkey_unpack(parent, &tmp.k, k);
 870
 871         ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 872                                   SIX_LOCK_intent);
 873
 874         if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
 875                 struct btree_iter *linked;
 876
 877                 if (!bch2_btree_node_relock(iter, level + 1))
 878                         goto out;
 879
 880                 /*
 881                  * We might have got -EINTR because trylock failed, and we're
 882                  * holding other locks that would cause us to deadlock:
 883                  */
 884                 trans_for_each_iter(trans, linked)
 885                         if (btree_iter_cmp(iter, linked) < 0)
 886                                 __bch2_btree_iter_unlock(linked);
 887
 888                 if (sib == btree_prev_sib)
 889                         btree_node_unlock(iter, level);
 890
 891                 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 892                                           SIX_LOCK_intent);
 893
 894                 /*
 895                  * before btree_iter_relock() calls btree_iter_verify_locks():
 896                  */
 897                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 898                         btree_node_unlock(iter, level + 1);
 899
 900                 if (!bch2_btree_node_relock(iter, level)) {
 901                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 902
 903                         if (!IS_ERR(ret)) {
 904                                 six_unlock_intent(&ret->lock);
 905                                 ret = ERR_PTR(-EINTR);
 906                         }
 907                 }
 908
 909                 bch2_trans_relock(trans);
 910         }
 911 out:
 912         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 913                 btree_node_unlock(iter, level + 1);
 914
 915         if (PTR_ERR_OR_ZERO(ret) == -EINTR)
 916                 bch2_btree_iter_upgrade(iter, level + 2);
 917
 918         BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
 919
 920         if (!IS_ERR_OR_NULL(ret)) {
 921                 struct btree *n1 = ret, *n2 = b;
 922
 923                 if (sib != btree_prev_sib)
 924                         swap(n1, n2);
 925
 926                 BUG_ON(bkey_cmp(btree_type_successor(n1->btree_id,
 927                                                      n1->key.k.p),
 928                                 n2->data->min_key));
 929         }
 930
 931         bch2_btree_trans_verify_locks(trans);
 932
 933         return ret;
 934 }
 935
 936 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 937                               const struct bkey_i *k, unsigned level)
 938 {
 939         struct btree_cache *bc = &c->btree_cache;
 940         struct btree *b;
 941
 942         BUG_ON(!btree_node_locked(iter, level + 1));
 943         BUG_ON(level >= BTREE_MAX_DEPTH);
 944
 945         b = btree_cache_find(bc, k);
 946         if (b)
 947                 return;
 948
 949         bch2_btree_node_fill(c, iter, k, iter->btree_id,
 950                              level, SIX_LOCK_read, false);
 951 }
 952
 953 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
 954                              struct btree *b)
 955 {
 956         const struct bkey_format *f = &b->format;
 957         struct bset_stats stats;
 958
 959         memset(&stats, 0, sizeof(stats));
 960
 961         bch2_btree_keys_stats(b, &stats);
 962
 963         pr_buf(out,
 964                "l %u %llu:%llu - %llu:%llu:\n"
 965                "    ptrs: ",
 966                b->level,
 967                b->data->min_key.inode,
 968                b->data->min_key.offset,
 969                b->data->max_key.inode,
 970                b->data->max_key.offset);
 971         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 972         pr_buf(out, "\n"
 973                "    format: u64s %u fields %u %u %u %u %u\n"
 974                "    unpack fn len: %u\n"
 975                "    bytes used %zu/%zu (%zu%% full)\n"
 976                "    sib u64s: %u, %u (merge threshold %zu)\n"
 977                "    nr packed keys %u\n"
 978                "    nr unpacked keys %u\n"
 979                "    floats %zu\n"
 980                "    failed unpacked %zu\n",
 981                f->key_u64s,
 982                f->bits_per_field[0],
 983                f->bits_per_field[1],
 984                f->bits_per_field[2],
 985                f->bits_per_field[3],
 986                f->bits_per_field[4],
 987                b->unpack_fn_len,
 988                b->nr.live_u64s * sizeof(u64),
 989                btree_bytes(c) - sizeof(struct btree_node),
 990                b->nr.live_u64s * 100 / btree_max_u64s(c),
 991                b->sib_u64s[0],
 992                b->sib_u64s[1],
 993                BTREE_FOREGROUND_MERGE_THRESHOLD(c),
 994                b->nr.packed_keys,
 995                b->nr.unpacked_keys,
 996                stats.floats,
 997                stats.failed);
 998 }