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 <trace/events/bcachefs.h>
  11
  12 #define DEF_BTREE_ID(kwd, val, name) name,
  13
  14 const char * const bch2_btree_ids[] = {
  15         DEFINE_BCH_BTREE_IDS()
  16         NULL
  17 };
  18
  19 #undef DEF_BTREE_ID
  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 #define mca_can_free(c)                                         \
  37         max_t(int, 0, c->btree_cache_used - c->btree_cache_reserve)
  38
  39 static void __mca_data_free(struct bch_fs *c, struct btree *b)
  40 {
  41         EBUG_ON(btree_node_write_in_flight(b));
  42
  43         free_pages((unsigned long) b->data, btree_page_order(c));
  44         b->data = NULL;
  45         bch2_btree_keys_free(b);
  46 }
  47
  48 static void mca_data_free(struct bch_fs *c, struct btree *b)
  49 {
  50         __mca_data_free(c, b);
  51         c->btree_cache_used--;
  52         list_move(&b->list, &c->btree_cache_freed);
  53 }
  54
  55 #define PTR_HASH(_k)    (bkey_i_to_extent_c(_k)->v._data[0])
  56
  57 static const struct rhashtable_params bch_btree_cache_params = {
  58         .head_offset    = offsetof(struct btree, hash),
  59         .key_offset     = offsetof(struct btree, key.v),
  60         .key_len        = sizeof(struct bch_extent_ptr),
  61 };
  62
  63 static void mca_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  64 {
  65         unsigned order = ilog2(btree_pages(c));
  66
  67         b->data = (void *) __get_free_pages(gfp, order);
  68         if (!b->data)
  69                 goto err;
  70
  71         if (bch2_btree_keys_alloc(b, order, gfp))
  72                 goto err;
  73
  74         c->btree_cache_used++;
  75         list_move(&b->list, &c->btree_cache_freeable);
  76         return;
  77 err:
  78         free_pages((unsigned long) b->data, order);
  79         b->data = NULL;
  80         list_move(&b->list, &c->btree_cache_freed);
  81 }
  82
  83 static struct btree *mca_bucket_alloc(struct bch_fs *c, gfp_t gfp)
  84 {
  85         struct btree *b = kzalloc(sizeof(struct btree), gfp);
  86         if (!b)
  87                 return NULL;
  88
  89         six_lock_init(&b->lock);
  90         INIT_LIST_HEAD(&b->list);
  91         INIT_LIST_HEAD(&b->write_blocked);
  92
  93         mca_data_alloc(c, b, gfp);
  94         return b->data ? b : NULL;
  95 }
  96
  97 /* Btree in memory cache - hash table */
  98
  99 void bch2_btree_node_hash_remove(struct bch_fs *c, struct btree *b)
 100 {
 101         BUG_ON(btree_node_dirty(b));
 102
 103         b->nsets = 0;
 104
 105         rhashtable_remove_fast(&c->btree_cache_table, &b->hash,
 106                                bch_btree_cache_params);
 107
 108         /* Cause future lookups for this node to fail: */
 109         bkey_i_to_extent(&b->key)->v._data[0] = 0;
 110 }
 111
 112 int bch2_btree_node_hash_insert(struct bch_fs *c, struct btree *b,
 113                     unsigned level, enum btree_id id)
 114 {
 115         int ret;
 116         b->level        = level;
 117         b->btree_id     = id;
 118
 119         ret = rhashtable_lookup_insert_fast(&c->btree_cache_table, &b->hash,
 120                                             bch_btree_cache_params);
 121         if (ret)
 122                 return ret;
 123
 124         mutex_lock(&c->btree_cache_lock);
 125         list_add(&b->list, &c->btree_cache);
 126         mutex_unlock(&c->btree_cache_lock);
 127
 128         return 0;
 129 }
 130
 131 __flatten
 132 static inline struct btree *mca_find(struct bch_fs *c,
 133                                      const struct bkey_i *k)
 134 {
 135         return rhashtable_lookup_fast(&c->btree_cache_table, &PTR_HASH(k),
 136                                       bch_btree_cache_params);
 137 }
 138
 139 /*
 140  * this version is for btree nodes that have already been freed (we're not
 141  * reaping a real btree node)
 142  */
 143 static int mca_reap_notrace(struct bch_fs *c, struct btree *b, bool flush)
 144 {
 145         lockdep_assert_held(&c->btree_cache_lock);
 146
 147         if (!six_trylock_intent(&b->lock))
 148                 return -ENOMEM;
 149
 150         if (!six_trylock_write(&b->lock))
 151                 goto out_unlock_intent;
 152
 153         if (btree_node_write_error(b) ||
 154             btree_node_noevict(b))
 155                 goto out_unlock;
 156
 157         if (!list_empty(&b->write_blocked))
 158                 goto out_unlock;
 159
 160         if (!flush &&
 161             (btree_node_dirty(b) ||
 162              btree_node_write_in_flight(b)))
 163                 goto out_unlock;
 164
 165         /*
 166          * Using the underscore version because we don't want to compact bsets
 167          * after the write, since this node is about to be evicted - unless
 168          * btree verify mode is enabled, since it runs out of the post write
 169          * cleanup:
 170          */
 171         if (btree_node_dirty(b)) {
 172                 if (verify_btree_ondisk(c))
 173                         bch2_btree_node_write(c, b, NULL, SIX_LOCK_intent, -1);
 174                 else
 175                         __bch2_btree_node_write(c, b, NULL, SIX_LOCK_read, -1);
 176         }
 177
 178         /* wait for any in flight btree write */
 179         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
 180                        TASK_UNINTERRUPTIBLE);
 181
 182         return 0;
 183 out_unlock:
 184         six_unlock_write(&b->lock);
 185 out_unlock_intent:
 186         six_unlock_intent(&b->lock);
 187         return -ENOMEM;
 188 }
 189
 190 static int mca_reap(struct bch_fs *c, struct btree *b, bool flush)
 191 {
 192         int ret = mca_reap_notrace(c, b, flush);
 193
 194         trace_btree_node_reap(c, b, ret);
 195         return ret;
 196 }
 197
 198 static unsigned long bch2_mca_scan(struct shrinker *shrink,
 199                                    struct shrink_control *sc)
 200 {
 201         struct bch_fs *c = container_of(shrink, struct bch_fs,
 202                                            btree_cache_shrink);
 203         struct btree *b, *t;
 204         unsigned long nr = sc->nr_to_scan;
 205         unsigned long can_free;
 206         unsigned long touched = 0;
 207         unsigned long freed = 0;
 208         unsigned i;
 209
 210         if (btree_shrinker_disabled(c))
 211                 return SHRINK_STOP;
 212
 213         if (c->btree_cache_alloc_lock)
 214                 return SHRINK_STOP;
 215
 216         /* Return -1 if we can't do anything right now */
 217         if (sc->gfp_mask & __GFP_IO)
 218                 mutex_lock(&c->btree_cache_lock);
 219         else if (!mutex_trylock(&c->btree_cache_lock))
 220                 return -1;
 221
 222         /*
 223          * It's _really_ critical that we don't free too many btree nodes - we
 224          * have to always leave ourselves a reserve. The reserve is how we
 225          * guarantee that allocating memory for a new btree node can always
 226          * succeed, so that inserting keys into the btree can always succeed and
 227          * IO can always make forward progress:
 228          */
 229         nr /= btree_pages(c);
 230         can_free = mca_can_free(c);
 231         nr = min_t(unsigned long, nr, can_free);
 232
 233         i = 0;
 234         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 235                 touched++;
 236
 237                 if (freed >= nr)
 238                         break;
 239
 240                 if (++i > 3 &&
 241                     !mca_reap_notrace(c, b, false)) {
 242                         mca_data_free(c, b);
 243                         six_unlock_write(&b->lock);
 244                         six_unlock_intent(&b->lock);
 245                         freed++;
 246                 }
 247         }
 248 restart:
 249         list_for_each_entry_safe(b, t, &c->btree_cache, list) {
 250                 touched++;
 251
 252                 if (freed >= nr) {
 253                         /* Save position */
 254                         if (&t->list != &c->btree_cache)
 255                                 list_move_tail(&c->btree_cache, &t->list);
 256                         break;
 257                 }
 258
 259                 if (!btree_node_accessed(b) &&
 260                     !mca_reap(c, b, false)) {
 261                         /* can't call bch2_btree_node_hash_remove under btree_cache_lock  */
 262                         freed++;
 263                         if (&t->list != &c->btree_cache)
 264                                 list_move_tail(&c->btree_cache, &t->list);
 265
 266                         mca_data_free(c, b);
 267                         mutex_unlock(&c->btree_cache_lock);
 268
 269                         bch2_btree_node_hash_remove(c, b);
 270                         six_unlock_write(&b->lock);
 271                         six_unlock_intent(&b->lock);
 272
 273                         if (freed >= nr)
 274                                 goto out;
 275
 276                         if (sc->gfp_mask & __GFP_IO)
 277                                 mutex_lock(&c->btree_cache_lock);
 278                         else if (!mutex_trylock(&c->btree_cache_lock))
 279                                 goto out;
 280                         goto restart;
 281                 } else
 282                         clear_btree_node_accessed(b);
 283         }
 284
 285         mutex_unlock(&c->btree_cache_lock);
 286 out:
 287         return (unsigned long) freed * btree_pages(c);
 288 }
 289
 290 static unsigned long bch2_mca_count(struct shrinker *shrink,
 291                                     struct shrink_control *sc)
 292 {
 293         struct bch_fs *c = container_of(shrink, struct bch_fs,
 294                                            btree_cache_shrink);
 295
 296         if (btree_shrinker_disabled(c))
 297                 return 0;
 298
 299         if (c->btree_cache_alloc_lock)
 300                 return 0;
 301
 302         return mca_can_free(c) * btree_pages(c);
 303 }
 304
 305 void bch2_fs_btree_exit(struct bch_fs *c)
 306 {
 307         struct btree *b;
 308         unsigned i;
 309
 310         if (c->btree_cache_shrink.list.next)
 311                 unregister_shrinker(&c->btree_cache_shrink);
 312
 313         mutex_lock(&c->btree_cache_lock);
 314
 315 #ifdef CONFIG_BCACHEFS_DEBUG
 316         if (c->verify_data)
 317                 list_move(&c->verify_data->list, &c->btree_cache);
 318
 319         free_pages((unsigned long) c->verify_ondisk, ilog2(btree_pages(c)));
 320 #endif
 321
 322         for (i = 0; i < BTREE_ID_NR; i++)
 323                 if (c->btree_roots[i].b)
 324                         list_add(&c->btree_roots[i].b->list, &c->btree_cache);
 325
 326         list_splice(&c->btree_cache_freeable,
 327                     &c->btree_cache);
 328
 329         while (!list_empty(&c->btree_cache)) {
 330                 b = list_first_entry(&c->btree_cache, struct btree, list);
 331
 332                 if (btree_node_dirty(b))
 333                         bch2_btree_complete_write(c, b, btree_current_write(b));
 334                 clear_btree_node_dirty(b);
 335
 336                 mca_data_free(c, b);
 337         }
 338
 339         while (!list_empty(&c->btree_cache_freed)) {
 340                 b = list_first_entry(&c->btree_cache_freed,
 341                                      struct btree, list);
 342                 list_del(&b->list);
 343                 kfree(b);
 344         }
 345
 346         mutex_unlock(&c->btree_cache_lock);
 347
 348         if (c->btree_cache_table_init_done)
 349                 rhashtable_destroy(&c->btree_cache_table);
 350 }
 351
 352 int bch2_fs_btree_init(struct bch_fs *c)
 353 {
 354         unsigned i;
 355         int ret;
 356
 357         ret = rhashtable_init(&c->btree_cache_table, &bch_btree_cache_params);
 358         if (ret)
 359                 return ret;
 360
 361         c->btree_cache_table_init_done = true;
 362
 363         bch2_recalc_btree_reserve(c);
 364
 365         for (i = 0; i < c->btree_cache_reserve; i++)
 366                 if (!mca_bucket_alloc(c, GFP_KERNEL))
 367                         return -ENOMEM;
 368
 369         list_splice_init(&c->btree_cache,
 370                          &c->btree_cache_freeable);
 371
 372 #ifdef CONFIG_BCACHEFS_DEBUG
 373         mutex_init(&c->verify_lock);
 374
 375         c->verify_ondisk = (void *)
 376                 __get_free_pages(GFP_KERNEL, ilog2(btree_pages(c)));
 377         if (!c->verify_ondisk)
 378                 return -ENOMEM;
 379
 380         c->verify_data = mca_bucket_alloc(c, GFP_KERNEL);
 381         if (!c->verify_data)
 382                 return -ENOMEM;
 383
 384         list_del_init(&c->verify_data->list);
 385 #endif
 386
 387         c->btree_cache_shrink.count_objects = bch2_mca_count;
 388         c->btree_cache_shrink.scan_objects = bch2_mca_scan;
 389         c->btree_cache_shrink.seeks = 4;
 390         c->btree_cache_shrink.batch = btree_pages(c) * 2;
 391         register_shrinker(&c->btree_cache_shrink);
 392
 393         return 0;
 394 }
 395
 396 /*
 397  * We can only have one thread cannibalizing other cached btree nodes at a time,
 398  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 399  * cannibalize_bucket() will take. This means every time we unlock the root of
 400  * the btree, we need to release this lock if we have it held.
 401  */
 402 void bch2_btree_node_cannibalize_unlock(struct bch_fs *c)
 403 {
 404         if (c->btree_cache_alloc_lock == current) {
 405                 trace_btree_node_cannibalize_unlock(c);
 406                 c->btree_cache_alloc_lock = NULL;
 407                 closure_wake_up(&c->mca_wait);
 408         }
 409 }
 410
 411 int bch2_btree_node_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 412 {
 413         struct task_struct *old;
 414
 415         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 416         if (old == NULL || old == current)
 417                 goto success;
 418
 419         if (!cl) {
 420                 trace_btree_node_cannibalize_lock_fail(c);
 421                 return -ENOMEM;
 422         }
 423
 424         closure_wait(&c->mca_wait, cl);
 425
 426         /* Try again, after adding ourselves to waitlist */
 427         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 428         if (old == NULL || old == current) {
 429                 /* We raced */
 430                 closure_wake_up(&c->mca_wait);
 431                 goto success;
 432         }
 433
 434         trace_btree_node_cannibalize_lock_fail(c);
 435         return -EAGAIN;
 436
 437 success:
 438         trace_btree_node_cannibalize_lock(c);
 439         return 0;
 440 }
 441
 442 static struct btree *mca_cannibalize(struct bch_fs *c)
 443 {
 444         struct btree *b;
 445
 446         list_for_each_entry_reverse(b, &c->btree_cache, list)
 447                 if (!mca_reap(c, b, false))
 448                         return b;
 449
 450         while (1) {
 451                 list_for_each_entry_reverse(b, &c->btree_cache, list)
 452                         if (!mca_reap(c, b, true))
 453                                 return b;
 454
 455                 /*
 456                  * Rare case: all nodes were intent-locked.
 457                  * Just busy-wait.
 458                  */
 459                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 460                 cond_resched();
 461         }
 462 }
 463
 464 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 465 {
 466         struct btree *b;
 467         u64 start_time = local_clock();
 468
 469         mutex_lock(&c->btree_cache_lock);
 470
 471         /*
 472          * btree_free() doesn't free memory; it sticks the node on the end of
 473          * the list. Check if there's any freed nodes there:
 474          */
 475         list_for_each_entry(b, &c->btree_cache_freeable, list)
 476                 if (!mca_reap_notrace(c, b, false))
 477                         goto out_unlock;
 478
 479         /*
 480          * We never free struct btree itself, just the memory that holds the on
 481          * disk node. Check the freed list before allocating a new one:
 482          */
 483         list_for_each_entry(b, &c->btree_cache_freed, list)
 484                 if (!mca_reap_notrace(c, b, false)) {
 485                         mca_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
 486                         if (b->data)
 487                                 goto out_unlock;
 488
 489                         six_unlock_write(&b->lock);
 490                         six_unlock_intent(&b->lock);
 491                         goto err;
 492                 }
 493
 494         b = mca_bucket_alloc(c, __GFP_NOWARN|GFP_NOIO);
 495         if (!b)
 496                 goto err;
 497
 498         BUG_ON(!six_trylock_intent(&b->lock));
 499         BUG_ON(!six_trylock_write(&b->lock));
 500 out_unlock:
 501         BUG_ON(bkey_extent_is_data(&b->key.k) && PTR_HASH(&b->key));
 502         BUG_ON(btree_node_write_in_flight(b));
 503
 504         list_del_init(&b->list);
 505         mutex_unlock(&c->btree_cache_lock);
 506 out:
 507         b->flags                = 0;
 508         b->written              = 0;
 509         b->nsets                = 0;
 510         b->sib_u64s[0]          = 0;
 511         b->sib_u64s[1]          = 0;
 512         b->whiteout_u64s        = 0;
 513         b->uncompacted_whiteout_u64s = 0;
 514         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 515
 516         bch2_time_stats_update(&c->btree_node_mem_alloc_time, start_time);
 517
 518         return b;
 519 err:
 520         /* Try to cannibalize another cached btree node: */
 521         if (c->btree_cache_alloc_lock == current) {
 522                 b = mca_cannibalize(c);
 523                 list_del_init(&b->list);
 524                 mutex_unlock(&c->btree_cache_lock);
 525
 526                 bch2_btree_node_hash_remove(c, b);
 527
 528                 trace_btree_node_cannibalize(c);
 529                 goto out;
 530         }
 531
 532         mutex_unlock(&c->btree_cache_lock);
 533         return ERR_PTR(-ENOMEM);
 534 }
 535
 536 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 537 static noinline struct btree *bch2_btree_node_fill(struct btree_iter *iter,
 538                                                    const struct bkey_i *k,
 539                                                    unsigned level,
 540                                                    enum six_lock_type lock_type)
 541 {
 542         struct bch_fs *c = iter->c;
 543         struct btree *b;
 544
 545         b = bch2_btree_node_mem_alloc(c);
 546         if (IS_ERR(b))
 547                 return b;
 548
 549         bkey_copy(&b->key, k);
 550         if (bch2_btree_node_hash_insert(c, b, level, iter->btree_id)) {
 551                 /* raced with another fill: */
 552
 553                 /* mark as unhashed... */
 554                 bkey_i_to_extent(&b->key)->v._data[0] = 0;
 555
 556                 mutex_lock(&c->btree_cache_lock);
 557                 list_add(&b->list, &c->btree_cache_freeable);
 558                 mutex_unlock(&c->btree_cache_lock);
 559
 560                 six_unlock_write(&b->lock);
 561                 six_unlock_intent(&b->lock);
 562                 return NULL;
 563         }
 564
 565         /*
 566          * If the btree node wasn't cached, we can't drop our lock on
 567          * the parent until after it's added to the cache - because
 568          * otherwise we could race with a btree_split() freeing the node
 569          * we're trying to lock.
 570          *
 571          * But the deadlock described below doesn't exist in this case,
 572          * so it's safe to not drop the parent lock until here:
 573          */
 574         if (btree_node_read_locked(iter, level + 1))
 575                 btree_node_unlock(iter, level + 1);
 576
 577         bch2_btree_node_read(c, b);
 578         six_unlock_write(&b->lock);
 579
 580         if (lock_type == SIX_LOCK_read)
 581                 six_lock_downgrade(&b->lock);
 582
 583         return b;
 584 }
 585
 586 /**
 587  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 588  * in from disk if necessary.
 589  *
 590  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 591  *
 592  * The btree node will have either a read or a write lock held, depending on
 593  * the @write parameter.
 594  */
 595 struct btree *bch2_btree_node_get(struct btree_iter *iter,
 596                                   const struct bkey_i *k, unsigned level,
 597                                   enum six_lock_type lock_type)
 598 {
 599         struct btree *b;
 600         struct bset_tree *t;
 601
 602         BUG_ON(level >= BTREE_MAX_DEPTH);
 603 retry:
 604         rcu_read_lock();
 605         b = mca_find(iter->c, k);
 606         rcu_read_unlock();
 607
 608         if (unlikely(!b)) {
 609                 /*
 610                  * We must have the parent locked to call bch2_btree_node_fill(),
 611                  * else we could read in a btree node from disk that's been
 612                  * freed:
 613                  */
 614                 b = bch2_btree_node_fill(iter, k, level, lock_type);
 615
 616                 /* We raced and found the btree node in the cache */
 617                 if (!b)
 618                         goto retry;
 619
 620                 if (IS_ERR(b))
 621                         return b;
 622         } else {
 623                 /*
 624                  * There's a potential deadlock with splits and insertions into
 625                  * interior nodes we have to avoid:
 626                  *
 627                  * The other thread might be holding an intent lock on the node
 628                  * we want, and they want to update its parent node so they're
 629                  * going to upgrade their intent lock on the parent node to a
 630                  * write lock.
 631                  *
 632                  * But if we're holding a read lock on the parent, and we're
 633                  * trying to get the intent lock they're holding, we deadlock.
 634                  *
 635                  * So to avoid this we drop the read locks on parent nodes when
 636                  * we're starting to take intent locks - and handle the race.
 637                  *
 638                  * The race is that they might be about to free the node we
 639                  * want, and dropping our read lock on the parent node lets them
 640                  * update the parent marking the node we want as freed, and then
 641                  * free it:
 642                  *
 643                  * To guard against this, btree nodes are evicted from the cache
 644                  * when they're freed - and PTR_HASH() is zeroed out, which we
 645                  * check for after we lock the node.
 646                  *
 647                  * Then, bch2_btree_node_relock() on the parent will fail - because
 648                  * the parent was modified, when the pointer to the node we want
 649                  * was removed - and we'll bail out:
 650                  */
 651                 if (btree_node_read_locked(iter, level + 1))
 652                         btree_node_unlock(iter, level + 1);
 653
 654                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
 655                         return ERR_PTR(-EINTR);
 656
 657                 if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
 658                              b->level != level ||
 659                              race_fault())) {
 660                         six_unlock_type(&b->lock, lock_type);
 661                         if (bch2_btree_node_relock(iter, level + 1))
 662                                 goto retry;
 663
 664                         return ERR_PTR(-EINTR);
 665                 }
 666         }
 667
 668         prefetch(b->aux_data);
 669
 670         for_each_bset(b, t) {
 671                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 672
 673                 prefetch(p + L1_CACHE_BYTES * 0);
 674                 prefetch(p + L1_CACHE_BYTES * 1);
 675                 prefetch(p + L1_CACHE_BYTES * 2);
 676         }
 677
 678         /* avoid atomic set bit if it's not needed: */
 679         if (btree_node_accessed(b))
 680                 set_btree_node_accessed(b);
 681
 682         if (unlikely(btree_node_read_error(b))) {
 683                 six_unlock_type(&b->lock, lock_type);
 684                 return ERR_PTR(-EIO);
 685         }
 686
 687         EBUG_ON(!b->written);
 688         EBUG_ON(b->btree_id != iter->btree_id ||
 689                 BTREE_NODE_LEVEL(b->data) != level ||
 690                 bkey_cmp(b->data->max_key, k->k.p));
 691
 692         return b;
 693 }
 694
 695 int bch2_print_btree_node(struct bch_fs *c, struct btree *b,
 696                           char *buf, size_t len)
 697 {
 698         const struct bkey_format *f = &b->format;
 699         struct bset_stats stats;
 700         char ptrs[100];
 701
 702         memset(&stats, 0, sizeof(stats));
 703
 704         bch2_val_to_text(c, BKEY_TYPE_BTREE, ptrs, sizeof(ptrs),
 705                         bkey_i_to_s_c(&b->key));
 706         bch2_btree_keys_stats(b, &stats);
 707
 708         return scnprintf(buf, len,
 709                          "l %u %llu:%llu - %llu:%llu:\n"
 710                          "    ptrs: %s\n"
 711                          "    format: u64s %u fields %u %u %u %u %u\n"
 712                          "    unpack fn len: %u\n"
 713                          "    bytes used %zu/%zu (%zu%% full)\n"
 714                          "    sib u64s: %u, %u (merge threshold %zu)\n"
 715                          "    nr packed keys %u\n"
 716                          "    nr unpacked keys %u\n"
 717                          "    floats %zu\n"
 718                          "    failed unpacked %zu\n"
 719                          "    failed prev %zu\n"
 720                          "    failed overflow %zu\n",
 721                          b->level,
 722                          b->data->min_key.inode,
 723                          b->data->min_key.offset,
 724                          b->data->max_key.inode,
 725                          b->data->max_key.offset,
 726                          ptrs,
 727                          f->key_u64s,
 728                          f->bits_per_field[0],
 729                          f->bits_per_field[1],
 730                          f->bits_per_field[2],
 731                          f->bits_per_field[3],
 732                          f->bits_per_field[4],
 733                          b->unpack_fn_len,
 734                          b->nr.live_u64s * sizeof(u64),
 735                          btree_bytes(c) - sizeof(struct btree_node),
 736                          b->nr.live_u64s * 100 / btree_max_u64s(c),
 737                          b->sib_u64s[0],
 738                          b->sib_u64s[1],
 739                          BTREE_FOREGROUND_MERGE_THRESHOLD(c),
 740                          b->nr.packed_keys,
 741                          b->nr.unpacked_keys,
 742                          stats.floats,
 743                          stats.failed_unpacked,
 744                          stats.failed_prev,
 745                          stats.failed_overflow);
 746 }