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

   1
   2 #include "bcachefs.h"
   3 #include "btree_cache.h"
   4 #include "btree_io.h"
   5 #include "btree_iter.h"
   6 #include "btree_locking.h"
   7 #include "debug.h"
   8 #include "extents.h"
   9
  10 #include <linux/prefetch.h>
  11 #include <trace/events/bcachefs.h>
  12
  13 #define DEF_BTREE_ID(kwd, val, name) name,
  14
  15 const char * const bch2_btree_ids[] = {
  16         DEFINE_BCH_BTREE_IDS()
  17         NULL
  18 };
  19
  20 #undef DEF_BTREE_ID
  21
  22 void bch2_recalc_btree_reserve(struct bch_fs *c)
  23 {
  24         unsigned i, reserve = 16;
  25
  26         if (!c->btree_roots[0].b)
  27                 reserve += 8;
  28
  29         for (i = 0; i < BTREE_ID_NR; i++)
  30                 if (c->btree_roots[i].b)
  31                         reserve += min_t(unsigned, 1,
  32                                          c->btree_roots[i].b->level) * 8;
  33
  34         c->btree_cache.reserve = reserve;
  35 }
  36
  37 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  38 {
  39         return max_t(int, 0, bc->used - bc->reserve);
  40 }
  41
  42 static void __btree_node_data_free(struct bch_fs *c, struct btree *b)
  43 {
  44         EBUG_ON(btree_node_write_in_flight(b));
  45
  46         kvpfree(b->data, btree_bytes(c));
  47         b->data = NULL;
  48         bch2_btree_keys_free(b);
  49 }
  50
  51 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  52 {
  53         struct btree_cache *bc = &c->btree_cache;
  54
  55         __btree_node_data_free(c, b);
  56         bc->used--;
  57         list_move(&b->list, &bc->freed);
  58 }
  59
  60 static const struct rhashtable_params bch_btree_cache_params = {
  61         .head_offset    = offsetof(struct btree, hash),
  62         .key_offset     = offsetof(struct btree, key.v),
  63         .key_len        = sizeof(struct bch_extent_ptr),
  64 };
  65
  66 static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  67 {
  68         struct btree_cache *bc = &c->btree_cache;
  69
  70         b->data = kvpmalloc(btree_bytes(c), gfp);
  71         if (!b->data)
  72                 goto err;
  73
  74         if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp))
  75                 goto err;
  76
  77         bc->used++;
  78         list_move(&b->list, &bc->freeable);
  79         return;
  80 err:
  81         kvpfree(b->data, btree_bytes(c));
  82         b->data = NULL;
  83         list_move(&b->list, &bc->freed);
  84 }
  85
  86 static struct btree *btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
  87 {
  88         struct btree *b = kzalloc(sizeof(struct btree), gfp);
  89         if (!b)
  90                 return NULL;
  91
  92         bkey_extent_init(&b->key);
  93         six_lock_init(&b->lock);
  94         INIT_LIST_HEAD(&b->list);
  95         INIT_LIST_HEAD(&b->write_blocked);
  96
  97         btree_node_data_alloc(c, b, gfp);
  98         return b->data ? b : NULL;
  99 }
 100
 101 /* Btree in memory cache - hash table */
 102
 103 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 104 {
 105         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 106
 107         /* Cause future lookups for this node to fail: */
 108         bkey_i_to_extent(&b->key)->v._data[0] = 0;
 109 }
 110
 111 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 112 {
 113         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 114                                              bch_btree_cache_params);
 115 }
 116
 117 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 118                                 unsigned level, enum btree_id id)
 119 {
 120         int ret;
 121
 122         b->level        = level;
 123         b->btree_id     = id;
 124
 125         mutex_lock(&bc->lock);
 126         ret = __bch2_btree_node_hash_insert(bc, b);
 127         if (!ret)
 128                 list_add(&b->list, &bc->live);
 129         mutex_unlock(&bc->lock);
 130
 131         return ret;
 132 }
 133
 134 __flatten
 135 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 136                                      const struct bkey_i *k)
 137 {
 138         return rhashtable_lookup_fast(&bc->table, &PTR_HASH(k),
 139                                       bch_btree_cache_params);
 140 }
 141
 142 /*
 143  * this version is for btree nodes that have already been freed (we're not
 144  * reaping a real btree node)
 145  */
 146 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 147 {
 148         struct btree_cache *bc = &c->btree_cache;
 149         int ret = 0;
 150
 151         lockdep_assert_held(&bc->lock);
 152
 153         if (!six_trylock_intent(&b->lock))
 154                 return -ENOMEM;
 155
 156         if (!six_trylock_write(&b->lock))
 157                 goto out_unlock_intent;
 158
 159         if (btree_node_noevict(b))
 160                 goto out_unlock;
 161
 162         if (!btree_node_may_write(b))
 163                 goto out_unlock;
 164
 165         if (btree_node_dirty(b) ||
 166             btree_node_write_in_flight(b) ||
 167             btree_node_read_in_flight(b)) {
 168                 if (!flush)
 169                         goto out_unlock;
 170
 171                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 172                                TASK_UNINTERRUPTIBLE);
 173
 174                 /*
 175                  * Using the underscore version because we don't want to compact
 176                  * bsets after the write, since this node is about to be evicted
 177                  * - unless btree verify mode is enabled, since it runs out of
 178                  * the post write cleanup:
 179                  */
 180                 if (verify_btree_ondisk(c))
 181                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 182                 else
 183                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 184
 185                 /* wait for any in flight btree write */
 186                 btree_node_wait_on_io(b);
 187         }
 188 out:
 189         if (PTR_HASH(&b->key) && !ret)
 190                 trace_btree_node_reap(c, b);
 191         return ret;
 192 out_unlock:
 193         six_unlock_write(&b->lock);
 194 out_unlock_intent:
 195         six_unlock_intent(&b->lock);
 196         ret = -ENOMEM;
 197         goto out;
 198 }
 199
 200 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 201 {
 202         return __btree_node_reclaim(c, b, false);
 203 }
 204
 205 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 206 {
 207         return __btree_node_reclaim(c, b, true);
 208 }
 209
 210 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 211                                            struct shrink_control *sc)
 212 {
 213         struct bch_fs *c = container_of(shrink, struct bch_fs,
 214                                         btree_cache.shrink);
 215         struct btree_cache *bc = &c->btree_cache;
 216         struct btree *b, *t;
 217         unsigned long nr = sc->nr_to_scan;
 218         unsigned long can_free;
 219         unsigned long touched = 0;
 220         unsigned long freed = 0;
 221         unsigned i;
 222
 223         if (btree_shrinker_disabled(c))
 224                 return SHRINK_STOP;
 225
 226         /* Return -1 if we can't do anything right now */
 227         if (sc->gfp_mask & __GFP_IO)
 228                 mutex_lock(&bc->lock);
 229         else if (!mutex_trylock(&bc->lock))
 230                 return -1;
 231
 232         /*
 233          * It's _really_ critical that we don't free too many btree nodes - we
 234          * have to always leave ourselves a reserve. The reserve is how we
 235          * guarantee that allocating memory for a new btree node can always
 236          * succeed, so that inserting keys into the btree can always succeed and
 237          * IO can always make forward progress:
 238          */
 239         nr /= btree_pages(c);
 240         can_free = btree_cache_can_free(bc);
 241         nr = min_t(unsigned long, nr, can_free);
 242
 243         i = 0;
 244         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 245                 touched++;
 246
 247                 if (freed >= nr)
 248                         break;
 249
 250                 if (++i > 3 &&
 251                     !btree_node_reclaim(c, b)) {
 252                         btree_node_data_free(c, b);
 253                         six_unlock_write(&b->lock);
 254                         six_unlock_intent(&b->lock);
 255                         freed++;
 256                 }
 257         }
 258 restart:
 259         list_for_each_entry_safe(b, t, &bc->live, list) {
 260                 touched++;
 261
 262                 if (freed >= nr) {
 263                         /* Save position */
 264                         if (&t->list != &bc->live)
 265                                 list_move_tail(&bc->live, &t->list);
 266                         break;
 267                 }
 268
 269                 if (!btree_node_accessed(b) &&
 270                     !btree_node_reclaim(c, b)) {
 271                         /* can't call bch2_btree_node_hash_remove under lock  */
 272                         freed++;
 273                         if (&t->list != &bc->live)
 274                                 list_move_tail(&bc->live, &t->list);
 275
 276                         btree_node_data_free(c, b);
 277                         mutex_unlock(&bc->lock);
 278
 279                         bch2_btree_node_hash_remove(bc, b);
 280                         six_unlock_write(&b->lock);
 281                         six_unlock_intent(&b->lock);
 282
 283                         if (freed >= nr)
 284                                 goto out;
 285
 286                         if (sc->gfp_mask & __GFP_IO)
 287                                 mutex_lock(&bc->lock);
 288                         else if (!mutex_trylock(&bc->lock))
 289                                 goto out;
 290                         goto restart;
 291                 } else
 292                         clear_btree_node_accessed(b);
 293         }
 294
 295         mutex_unlock(&bc->lock);
 296 out:
 297         return (unsigned long) freed * btree_pages(c);
 298 }
 299
 300 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 301                                             struct shrink_control *sc)
 302 {
 303         struct bch_fs *c = container_of(shrink, struct bch_fs,
 304                                         btree_cache.shrink);
 305         struct btree_cache *bc = &c->btree_cache;
 306
 307         if (btree_shrinker_disabled(c))
 308                 return 0;
 309
 310         return btree_cache_can_free(bc) * btree_pages(c);
 311 }
 312
 313 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 314 {
 315         struct btree_cache *bc = &c->btree_cache;
 316         struct btree *b;
 317         unsigned i;
 318
 319         if (bc->shrink.list.next)
 320                 unregister_shrinker(&bc->shrink);
 321
 322         mutex_lock(&bc->lock);
 323
 324 #ifdef CONFIG_BCACHEFS_DEBUG
 325         if (c->verify_data)
 326                 list_move(&c->verify_data->list, &bc->live);
 327
 328         kvpfree(c->verify_ondisk, btree_bytes(c));
 329 #endif
 330
 331         for (i = 0; i < BTREE_ID_NR; i++)
 332                 if (c->btree_roots[i].b)
 333                         list_add(&c->btree_roots[i].b->list, &bc->live);
 334
 335         list_splice(&bc->freeable, &bc->live);
 336
 337         while (!list_empty(&bc->live)) {
 338                 b = list_first_entry(&bc->live, struct btree, list);
 339
 340                 if (btree_node_dirty(b))
 341                         bch2_btree_complete_write(c, b, btree_current_write(b));
 342                 clear_btree_node_dirty(b);
 343
 344                 btree_node_data_free(c, b);
 345         }
 346
 347         while (!list_empty(&bc->freed)) {
 348                 b = list_first_entry(&bc->freed, struct btree, list);
 349                 list_del(&b->list);
 350                 kfree(b);
 351         }
 352
 353         mutex_unlock(&bc->lock);
 354
 355         if (bc->table_init_done)
 356                 rhashtable_destroy(&bc->table);
 357 }
 358
 359 int bch2_fs_btree_cache_init(struct bch_fs *c)
 360 {
 361         struct btree_cache *bc = &c->btree_cache;
 362         unsigned i;
 363         int ret;
 364
 365         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 366         if (ret)
 367                 return ret;
 368
 369         bc->table_init_done = true;
 370
 371         bch2_recalc_btree_reserve(c);
 372
 373         for (i = 0; i < bc->reserve; i++)
 374                 if (!btree_node_mem_alloc(c, GFP_KERNEL))
 375                         return -ENOMEM;
 376
 377         list_splice_init(&bc->live, &bc->freeable);
 378
 379 #ifdef CONFIG_BCACHEFS_DEBUG
 380         mutex_init(&c->verify_lock);
 381
 382         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 383         if (!c->verify_ondisk)
 384                 return -ENOMEM;
 385
 386         c->verify_data = btree_node_mem_alloc(c, GFP_KERNEL);
 387         if (!c->verify_data)
 388                 return -ENOMEM;
 389
 390         list_del_init(&c->verify_data->list);
 391 #endif
 392
 393         bc->shrink.count_objects        = bch2_btree_cache_count;
 394         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 395         bc->shrink.seeks                = 4;
 396         bc->shrink.batch                = btree_pages(c) * 2;
 397         register_shrinker(&bc->shrink);
 398
 399         return 0;
 400 }
 401
 402 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 403 {
 404         mutex_init(&bc->lock);
 405         INIT_LIST_HEAD(&bc->live);
 406         INIT_LIST_HEAD(&bc->freeable);
 407         INIT_LIST_HEAD(&bc->freed);
 408 }
 409
 410 /*
 411  * We can only have one thread cannibalizing other cached btree nodes at a time,
 412  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 413  * cannibalize_bucket() will take. This means every time we unlock the root of
 414  * the btree, we need to release this lock if we have it held.
 415  */
 416 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 417 {
 418         struct btree_cache *bc = &c->btree_cache;
 419
 420         if (bc->alloc_lock == current) {
 421                 trace_btree_node_cannibalize_unlock(c);
 422                 bc->alloc_lock = NULL;
 423                 closure_wake_up(&bc->alloc_wait);
 424         }
 425 }
 426
 427 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 428 {
 429         struct btree_cache *bc = &c->btree_cache;
 430         struct task_struct *old;
 431
 432         old = cmpxchg(&bc->alloc_lock, NULL, current);
 433         if (old == NULL || old == current)
 434                 goto success;
 435
 436         if (!cl) {
 437                 trace_btree_node_cannibalize_lock_fail(c);
 438                 return -ENOMEM;
 439         }
 440
 441         closure_wait(&bc->alloc_wait, cl);
 442
 443         /* Try again, after adding ourselves to waitlist */
 444         old = cmpxchg(&bc->alloc_lock, NULL, current);
 445         if (old == NULL || old == current) {
 446                 /* We raced */
 447                 closure_wake_up(&bc->alloc_wait);
 448                 goto success;
 449         }
 450
 451         trace_btree_node_cannibalize_lock_fail(c);
 452         return -EAGAIN;
 453
 454 success:
 455         trace_btree_node_cannibalize_lock(c);
 456         return 0;
 457 }
 458
 459 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 460 {
 461         struct btree_cache *bc = &c->btree_cache;
 462         struct btree *b;
 463
 464         list_for_each_entry_reverse(b, &bc->live, list)
 465                 if (!btree_node_reclaim(c, b))
 466                         return b;
 467
 468         while (1) {
 469                 list_for_each_entry_reverse(b, &bc->live, list)
 470                         if (!btree_node_write_and_reclaim(c, b))
 471                                 return b;
 472
 473                 /*
 474                  * Rare case: all nodes were intent-locked.
 475                  * Just busy-wait.
 476                  */
 477                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 478                 cond_resched();
 479         }
 480 }
 481
 482 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 483 {
 484         struct btree_cache *bc = &c->btree_cache;
 485         struct btree *b;
 486         u64 start_time = local_clock();
 487
 488         mutex_lock(&bc->lock);
 489
 490         /*
 491          * btree_free() doesn't free memory; it sticks the node on the end of
 492          * the list. Check if there's any freed nodes there:
 493          */
 494         list_for_each_entry(b, &bc->freeable, list)
 495                 if (!btree_node_reclaim(c, b))
 496                         goto out_unlock;
 497
 498         /*
 499          * We never free struct btree itself, just the memory that holds the on
 500          * disk node. Check the freed list before allocating a new one:
 501          */
 502         list_for_each_entry(b, &bc->freed, list)
 503                 if (!btree_node_reclaim(c, b)) {
 504                         btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
 505                         if (b->data)
 506                                 goto out_unlock;
 507
 508                         six_unlock_write(&b->lock);
 509                         six_unlock_intent(&b->lock);
 510                         goto err;
 511                 }
 512
 513         b = btree_node_mem_alloc(c, __GFP_NOWARN|GFP_NOIO);
 514         if (!b)
 515                 goto err;
 516
 517         BUG_ON(!six_trylock_intent(&b->lock));
 518         BUG_ON(!six_trylock_write(&b->lock));
 519 out_unlock:
 520         BUG_ON(btree_node_hashed(b));
 521         BUG_ON(btree_node_write_in_flight(b));
 522
 523         list_del_init(&b->list);
 524         mutex_unlock(&bc->lock);
 525 out:
 526         b->flags                = 0;
 527         b->written              = 0;
 528         b->nsets                = 0;
 529         b->sib_u64s[0]          = 0;
 530         b->sib_u64s[1]          = 0;
 531         b->whiteout_u64s        = 0;
 532         b->uncompacted_whiteout_u64s = 0;
 533         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 534
 535         bch2_time_stats_update(&c->btree_node_mem_alloc_time, start_time);
 536
 537         return b;
 538 err:
 539         /* Try to cannibalize another cached btree node: */
 540         if (bc->alloc_lock == current) {
 541                 b = btree_node_cannibalize(c);
 542                 list_del_init(&b->list);
 543                 mutex_unlock(&bc->lock);
 544
 545                 bch2_btree_node_hash_remove(bc, b);
 546
 547                 trace_btree_node_cannibalize(c);
 548                 goto out;
 549         }
 550
 551         mutex_unlock(&bc->lock);
 552         return ERR_PTR(-ENOMEM);
 553 }
 554
 555 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 556 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 557                                                    struct btree_iter *iter,
 558                                                    const struct bkey_i *k,
 559                                                    unsigned level,
 560                                                    enum six_lock_type lock_type)
 561 {
 562         struct btree_cache *bc = &c->btree_cache;
 563         struct btree *b;
 564
 565         /*
 566          * Parent node must be locked, else we could read in a btree node that's
 567          * been freed:
 568          */
 569         BUG_ON(!btree_node_locked(iter, level + 1));
 570
 571         b = bch2_btree_node_mem_alloc(c);
 572         if (IS_ERR(b))
 573                 return b;
 574
 575         bkey_copy(&b->key, k);
 576         if (bch2_btree_node_hash_insert(bc, b, level, iter->btree_id)) {
 577                 /* raced with another fill: */
 578
 579                 /* mark as unhashed... */
 580                 bkey_i_to_extent(&b->key)->v._data[0] = 0;
 581
 582                 mutex_lock(&bc->lock);
 583                 list_add(&b->list, &bc->freeable);
 584                 mutex_unlock(&bc->lock);
 585
 586                 six_unlock_write(&b->lock);
 587                 six_unlock_intent(&b->lock);
 588                 return NULL;
 589         }
 590
 591         /*
 592          * If the btree node wasn't cached, we can't drop our lock on
 593          * the parent until after it's added to the cache - because
 594          * otherwise we could race with a btree_split() freeing the node
 595          * we're trying to lock.
 596          *
 597          * But the deadlock described below doesn't exist in this case,
 598          * so it's safe to not drop the parent lock until here:
 599          */
 600         if (btree_node_read_locked(iter, level + 1))
 601                 btree_node_unlock(iter, level + 1);
 602
 603         bch2_btree_node_read(c, b, true);
 604         six_unlock_write(&b->lock);
 605
 606         if (lock_type == SIX_LOCK_read)
 607                 six_lock_downgrade(&b->lock);
 608
 609         return b;
 610 }
 611
 612 /**
 613  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 614  * in from disk if necessary.
 615  *
 616  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 617  *
 618  * The btree node will have either a read or a write lock held, depending on
 619  * the @write parameter.
 620  */
 621 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 622                                   const struct bkey_i *k, unsigned level,
 623                                   enum six_lock_type lock_type)
 624 {
 625         struct btree_cache *bc = &c->btree_cache;
 626         struct btree *b;
 627         struct bset_tree *t;
 628
 629         /* btree_node_fill() requires parent to be locked: */
 630         EBUG_ON(!btree_node_locked(iter, level + 1));
 631         EBUG_ON(level >= BTREE_MAX_DEPTH);
 632 retry:
 633         rcu_read_lock();
 634         b = btree_cache_find(bc, k);
 635         rcu_read_unlock();
 636
 637         if (unlikely(!b)) {
 638                 /*
 639                  * We must have the parent locked to call bch2_btree_node_fill(),
 640                  * else we could read in a btree node from disk that's been
 641                  * freed:
 642                  */
 643                 b = bch2_btree_node_fill(c, iter, k, level, lock_type);
 644
 645                 /* We raced and found the btree node in the cache */
 646                 if (!b)
 647                         goto retry;
 648
 649                 if (IS_ERR(b))
 650                         return b;
 651         } else {
 652                 /*
 653                  * There's a potential deadlock with splits and insertions into
 654                  * interior nodes we have to avoid:
 655                  *
 656                  * The other thread might be holding an intent lock on the node
 657                  * we want, and they want to update its parent node so they're
 658                  * going to upgrade their intent lock on the parent node to a
 659                  * write lock.
 660                  *
 661                  * But if we're holding a read lock on the parent, and we're
 662                  * trying to get the intent lock they're holding, we deadlock.
 663                  *
 664                  * So to avoid this we drop the read locks on parent nodes when
 665                  * we're starting to take intent locks - and handle the race.
 666                  *
 667                  * The race is that they might be about to free the node we
 668                  * want, and dropping our read lock on the parent node lets them
 669                  * update the parent marking the node we want as freed, and then
 670                  * free it:
 671                  *
 672                  * To guard against this, btree nodes are evicted from the cache
 673                  * when they're freed - and PTR_HASH() is zeroed out, which we
 674                  * check for after we lock the node.
 675                  *
 676                  * Then, bch2_btree_node_relock() on the parent will fail - because
 677                  * the parent was modified, when the pointer to the node we want
 678                  * was removed - and we'll bail out:
 679                  */
 680                 if (btree_node_read_locked(iter, level + 1))
 681                         btree_node_unlock(iter, level + 1);
 682
 683                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
 684                         return ERR_PTR(-EINTR);
 685
 686                 if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
 687                              b->level != level ||
 688                              race_fault())) {
 689                         six_unlock_type(&b->lock, lock_type);
 690                         if (bch2_btree_node_relock(iter, level + 1))
 691                                 goto retry;
 692
 693                         return ERR_PTR(-EINTR);
 694                 }
 695         }
 696
 697         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 698                        TASK_UNINTERRUPTIBLE);
 699
 700         prefetch(b->aux_data);
 701
 702         for_each_bset(b, t) {
 703                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 704
 705                 prefetch(p + L1_CACHE_BYTES * 0);
 706                 prefetch(p + L1_CACHE_BYTES * 1);
 707                 prefetch(p + L1_CACHE_BYTES * 2);
 708         }
 709
 710         /* avoid atomic set bit if it's not needed: */
 711         if (btree_node_accessed(b))
 712                 set_btree_node_accessed(b);
 713
 714         if (unlikely(btree_node_read_error(b))) {
 715                 six_unlock_type(&b->lock, lock_type);
 716                 return ERR_PTR(-EIO);
 717         }
 718
 719         EBUG_ON(!b->written);
 720         EBUG_ON(b->btree_id != iter->btree_id ||
 721                 BTREE_NODE_LEVEL(b->data) != level ||
 722                 bkey_cmp(b->data->max_key, k->k.p));
 723
 724         return b;
 725 }
 726
 727 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 728                                           struct btree_iter *iter,
 729                                           struct btree *b,
 730                                           enum btree_node_sibling sib)
 731 {
 732         struct btree *parent;
 733         struct btree_node_iter node_iter;
 734         struct bkey_packed *k;
 735         BKEY_PADDED(k) tmp;
 736         struct btree *ret;
 737         unsigned level = b->level;
 738
 739         parent = iter->nodes[level + 1];
 740         if (!parent)
 741                 return NULL;
 742
 743         if (!bch2_btree_node_relock(iter, level + 1)) {
 744                 bch2_btree_iter_set_locks_want(iter, level + 2);
 745                 return ERR_PTR(-EINTR);
 746         }
 747
 748         node_iter = iter->node_iters[parent->level];
 749
 750         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 751         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 752
 753         do {
 754                 k = sib == btree_prev_sib
 755                         ? bch2_btree_node_iter_prev_all(&node_iter, parent)
 756                         : (bch2_btree_node_iter_advance(&node_iter, parent),
 757                            bch2_btree_node_iter_peek_all(&node_iter, parent));
 758                 if (!k)
 759                         return NULL;
 760         } while (bkey_deleted(k));
 761
 762         bch2_bkey_unpack(parent, &tmp.k, k);
 763
 764         ret = bch2_btree_node_get(c, iter, &tmp.k, level, SIX_LOCK_intent);
 765
 766         if (IS_ERR(ret) && PTR_ERR(ret) == -EINTR) {
 767                 btree_node_unlock(iter, level);
 768
 769                 if (!bch2_btree_node_relock(iter, level + 1)) {
 770                         bch2_btree_iter_set_locks_want(iter, level + 2);
 771                         return ERR_PTR(-EINTR);
 772                 }
 773
 774                 ret = bch2_btree_node_get(c, iter, &tmp.k, level, SIX_LOCK_intent);
 775         }
 776
 777         if (!IS_ERR(ret) && !bch2_btree_node_relock(iter, level)) {
 778                 six_unlock_intent(&ret->lock);
 779                 ret = ERR_PTR(-EINTR);
 780         }
 781
 782         return ret;
 783 }
 784
 785 void bch2_btree_node_prefetch(struct bch_fs *c, const struct bkey_i *k,
 786                               unsigned level, enum btree_id btree_id)
 787 {
 788         struct btree_cache *bc = &c->btree_cache;
 789         struct btree *b;
 790
 791         BUG_ON(level >= BTREE_MAX_DEPTH);
 792
 793         rcu_read_lock();
 794         b = btree_cache_find(bc, k);
 795         rcu_read_unlock();
 796
 797         if (b)
 798                 return;
 799
 800         b = bch2_btree_node_mem_alloc(c);
 801         if (IS_ERR(b))
 802                 return;
 803
 804         bkey_copy(&b->key, k);
 805         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 806                 /* raced with another fill: */
 807
 808                 /* mark as unhashed... */
 809                 bkey_i_to_extent(&b->key)->v._data[0] = 0;
 810
 811                 mutex_lock(&bc->lock);
 812                 list_add(&b->list, &bc->freeable);
 813                 mutex_unlock(&bc->lock);
 814                 goto out;
 815         }
 816
 817         bch2_btree_node_read(c, b, false);
 818 out:
 819         six_unlock_write(&b->lock);
 820         six_unlock_intent(&b->lock);
 821 }
 822
 823 int bch2_print_btree_node(struct bch_fs *c, struct btree *b,
 824                           char *buf, size_t len)
 825 {
 826         const struct bkey_format *f = &b->format;
 827         struct bset_stats stats;
 828         char ptrs[100];
 829
 830         memset(&stats, 0, sizeof(stats));
 831
 832         bch2_val_to_text(c, BKEY_TYPE_BTREE, ptrs, sizeof(ptrs),
 833                         bkey_i_to_s_c(&b->key));
 834         bch2_btree_keys_stats(b, &stats);
 835
 836         return scnprintf(buf, len,
 837                          "l %u %llu:%llu - %llu:%llu:\n"
 838                          "    ptrs: %s\n"
 839                          "    format: u64s %u fields %u %u %u %u %u\n"
 840                          "    unpack fn len: %u\n"
 841                          "    bytes used %zu/%zu (%zu%% full)\n"
 842                          "    sib u64s: %u, %u (merge threshold %zu)\n"
 843                          "    nr packed keys %u\n"
 844                          "    nr unpacked keys %u\n"
 845                          "    floats %zu\n"
 846                          "    failed unpacked %zu\n"
 847                          "    failed prev %zu\n"
 848                          "    failed overflow %zu\n",
 849                          b->level,
 850                          b->data->min_key.inode,
 851                          b->data->min_key.offset,
 852                          b->data->max_key.inode,
 853                          b->data->max_key.offset,
 854                          ptrs,
 855                          f->key_u64s,
 856                          f->bits_per_field[0],
 857                          f->bits_per_field[1],
 858                          f->bits_per_field[2],
 859                          f->bits_per_field[3],
 860                          f->bits_per_field[4],
 861                          b->unpack_fn_len,
 862                          b->nr.live_u64s * sizeof(u64),
 863                          btree_bytes(c) - sizeof(struct btree_node),
 864                          b->nr.live_u64s * 100 / btree_max_u64s(c),
 865                          b->sib_u64s[0],
 866                          b->sib_u64s[1],
 867                          BTREE_FOREGROUND_MERGE_THRESHOLD(c),
 868                          b->nr.packed_keys,
 869                          b->nr.unpacked_keys,
 870                          stats.floats,
 871                          stats.failed_unpacked,
 872                          stats.failed_prev,
 873                          stats.failed_overflow);
 874 }