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 #define mca_can_free(c)                                         \
  38         max_t(int, 0, c->btree_cache_used - c->btree_cache_reserve)
  39
  40 static void __mca_data_free(struct bch_fs *c, struct btree *b)
  41 {
  42         EBUG_ON(btree_node_write_in_flight(b));
  43
  44         free_pages((unsigned long) b->data, btree_page_order(c));
  45         b->data = NULL;
  46         bch2_btree_keys_free(b);
  47 }
  48
  49 static void mca_data_free(struct bch_fs *c, struct btree *b)
  50 {
  51         __mca_data_free(c, b);
  52         c->btree_cache_used--;
  53         list_move(&b->list, &c->btree_cache_freed);
  54 }
  55
  56 #define PTR_HASH(_k)    (bkey_i_to_extent_c(_k)->v._data[0])
  57
  58 static const struct rhashtable_params bch_btree_cache_params = {
  59         .head_offset    = offsetof(struct btree, hash),
  60         .key_offset     = offsetof(struct btree, key.v),
  61         .key_len        = sizeof(struct bch_extent_ptr),
  62 };
  63
  64 static void mca_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  65 {
  66         unsigned order = ilog2(btree_pages(c));
  67
  68         b->data = (void *) __get_free_pages(gfp, order);
  69         if (!b->data)
  70                 goto err;
  71
  72         if (bch2_btree_keys_alloc(b, order, gfp))
  73                 goto err;
  74
  75         c->btree_cache_used++;
  76         list_move(&b->list, &c->btree_cache_freeable);
  77         return;
  78 err:
  79         free_pages((unsigned long) b->data, order);
  80         b->data = NULL;
  81         list_move(&b->list, &c->btree_cache_freed);
  82 }
  83
  84 static struct btree *mca_bucket_alloc(struct bch_fs *c, gfp_t gfp)
  85 {
  86         struct btree *b = kzalloc(sizeof(struct btree), gfp);
  87         if (!b)
  88                 return NULL;
  89
  90         six_lock_init(&b->lock);
  91         INIT_LIST_HEAD(&b->list);
  92         INIT_LIST_HEAD(&b->write_blocked);
  93
  94         mca_data_alloc(c, b, gfp);
  95         return b->data ? b : NULL;
  96 }
  97
  98 /* Btree in memory cache - hash table */
  99
 100 void bch2_btree_node_hash_remove(struct bch_fs *c, struct btree *b)
 101 {
 102         BUG_ON(btree_node_dirty(b));
 103
 104         b->nsets = 0;
 105
 106         rhashtable_remove_fast(&c->btree_cache_table, &b->hash,
 107                                bch_btree_cache_params);
 108
 109         /* Cause future lookups for this node to fail: */
 110         bkey_i_to_extent(&b->key)->v._data[0] = 0;
 111 }
 112
 113 int bch2_btree_node_hash_insert(struct bch_fs *c, struct btree *b,
 114                     unsigned level, enum btree_id id)
 115 {
 116         int ret;
 117         b->level        = level;
 118         b->btree_id     = id;
 119
 120         ret = rhashtable_lookup_insert_fast(&c->btree_cache_table, &b->hash,
 121                                             bch_btree_cache_params);
 122         if (ret)
 123                 return ret;
 124
 125         mutex_lock(&c->btree_cache_lock);
 126         list_add(&b->list, &c->btree_cache);
 127         mutex_unlock(&c->btree_cache_lock);
 128
 129         return 0;
 130 }
 131
 132 __flatten
 133 static inline struct btree *mca_find(struct bch_fs *c,
 134                                      const struct bkey_i *k)
 135 {
 136         return rhashtable_lookup_fast(&c->btree_cache_table, &PTR_HASH(k),
 137                                       bch_btree_cache_params);
 138 }
 139
 140 /*
 141  * this version is for btree nodes that have already been freed (we're not
 142  * reaping a real btree node)
 143  */
 144 static int mca_reap_notrace(struct bch_fs *c, struct btree *b, bool flush)
 145 {
 146         lockdep_assert_held(&c->btree_cache_lock);
 147
 148         if (!six_trylock_intent(&b->lock))
 149                 return -ENOMEM;
 150
 151         if (!six_trylock_write(&b->lock))
 152                 goto out_unlock_intent;
 153
 154         if (btree_node_write_error(b) ||
 155             btree_node_noevict(b))
 156                 goto out_unlock;
 157
 158         if (!list_empty(&b->write_blocked))
 159                 goto out_unlock;
 160
 161         if (!flush &&
 162             (btree_node_dirty(b) ||
 163              btree_node_write_in_flight(b)))
 164                 goto out_unlock;
 165
 166         /*
 167          * Using the underscore version because we don't want to compact bsets
 168          * after the write, since this node is about to be evicted - unless
 169          * btree verify mode is enabled, since it runs out of the post write
 170          * cleanup:
 171          */
 172         if (btree_node_dirty(b)) {
 173                 if (verify_btree_ondisk(c))
 174                         bch2_btree_node_write(c, b, NULL, SIX_LOCK_intent, -1);
 175                 else
 176                         __bch2_btree_node_write(c, b, NULL, SIX_LOCK_read, -1);
 177         }
 178
 179         /* wait for any in flight btree write */
 180         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
 181                        TASK_UNINTERRUPTIBLE);
 182
 183         return 0;
 184 out_unlock:
 185         six_unlock_write(&b->lock);
 186 out_unlock_intent:
 187         six_unlock_intent(&b->lock);
 188         return -ENOMEM;
 189 }
 190
 191 static int mca_reap(struct bch_fs *c, struct btree *b, bool flush)
 192 {
 193         int ret = mca_reap_notrace(c, b, flush);
 194
 195         trace_btree_node_reap(c, b, ret);
 196         return ret;
 197 }
 198
 199 static unsigned long bch2_mca_scan(struct shrinker *shrink,
 200                                    struct shrink_control *sc)
 201 {
 202         struct bch_fs *c = container_of(shrink, struct bch_fs,
 203                                            btree_cache_shrink);
 204         struct btree *b, *t;
 205         unsigned long nr = sc->nr_to_scan;
 206         unsigned long can_free;
 207         unsigned long touched = 0;
 208         unsigned long freed = 0;
 209         unsigned i;
 210
 211         if (btree_shrinker_disabled(c))
 212                 return SHRINK_STOP;
 213
 214         if (c->btree_cache_alloc_lock)
 215                 return SHRINK_STOP;
 216
 217         /* Return -1 if we can't do anything right now */
 218         if (sc->gfp_mask & __GFP_IO)
 219                 mutex_lock(&c->btree_cache_lock);
 220         else if (!mutex_trylock(&c->btree_cache_lock))
 221                 return -1;
 222
 223         /*
 224          * It's _really_ critical that we don't free too many btree nodes - we
 225          * have to always leave ourselves a reserve. The reserve is how we
 226          * guarantee that allocating memory for a new btree node can always
 227          * succeed, so that inserting keys into the btree can always succeed and
 228          * IO can always make forward progress:
 229          */
 230         nr /= btree_pages(c);
 231         can_free = mca_can_free(c);
 232         nr = min_t(unsigned long, nr, can_free);
 233
 234         i = 0;
 235         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 236                 touched++;
 237
 238                 if (freed >= nr)
 239                         break;
 240
 241                 if (++i > 3 &&
 242                     !mca_reap_notrace(c, b, false)) {
 243                         mca_data_free(c, b);
 244                         six_unlock_write(&b->lock);
 245                         six_unlock_intent(&b->lock);
 246                         freed++;
 247                 }
 248         }
 249 restart:
 250         list_for_each_entry_safe(b, t, &c->btree_cache, list) {
 251                 touched++;
 252
 253                 if (freed >= nr) {
 254                         /* Save position */
 255                         if (&t->list != &c->btree_cache)
 256                                 list_move_tail(&c->btree_cache, &t->list);
 257                         break;
 258                 }
 259
 260                 if (!btree_node_accessed(b) &&
 261                     !mca_reap(c, b, false)) {
 262                         /* can't call bch2_btree_node_hash_remove under btree_cache_lock  */
 263                         freed++;
 264                         if (&t->list != &c->btree_cache)
 265                                 list_move_tail(&c->btree_cache, &t->list);
 266
 267                         mca_data_free(c, b);
 268                         mutex_unlock(&c->btree_cache_lock);
 269
 270                         bch2_btree_node_hash_remove(c, b);
 271                         six_unlock_write(&b->lock);
 272                         six_unlock_intent(&b->lock);
 273
 274                         if (freed >= nr)
 275                                 goto out;
 276
 277                         if (sc->gfp_mask & __GFP_IO)
 278                                 mutex_lock(&c->btree_cache_lock);
 279                         else if (!mutex_trylock(&c->btree_cache_lock))
 280                                 goto out;
 281                         goto restart;
 282                 } else
 283                         clear_btree_node_accessed(b);
 284         }
 285
 286         mutex_unlock(&c->btree_cache_lock);
 287 out:
 288         return (unsigned long) freed * btree_pages(c);
 289 }
 290
 291 static unsigned long bch2_mca_count(struct shrinker *shrink,
 292                                     struct shrink_control *sc)
 293 {
 294         struct bch_fs *c = container_of(shrink, struct bch_fs,
 295                                            btree_cache_shrink);
 296
 297         if (btree_shrinker_disabled(c))
 298                 return 0;
 299
 300         if (c->btree_cache_alloc_lock)
 301                 return 0;
 302
 303         return mca_can_free(c) * btree_pages(c);
 304 }
 305
 306 void bch2_fs_btree_exit(struct bch_fs *c)
 307 {
 308         struct btree *b;
 309         unsigned i;
 310
 311         if (c->btree_cache_shrink.list.next)
 312                 unregister_shrinker(&c->btree_cache_shrink);
 313
 314         mutex_lock(&c->btree_cache_lock);
 315
 316 #ifdef CONFIG_BCACHEFS_DEBUG
 317         if (c->verify_data)
 318                 list_move(&c->verify_data->list, &c->btree_cache);
 319
 320         free_pages((unsigned long) c->verify_ondisk, ilog2(btree_pages(c)));
 321 #endif
 322
 323         for (i = 0; i < BTREE_ID_NR; i++)
 324                 if (c->btree_roots[i].b)
 325                         list_add(&c->btree_roots[i].b->list, &c->btree_cache);
 326
 327         list_splice(&c->btree_cache_freeable,
 328                     &c->btree_cache);
 329
 330         while (!list_empty(&c->btree_cache)) {
 331                 b = list_first_entry(&c->btree_cache, struct btree, list);
 332
 333                 if (btree_node_dirty(b))
 334                         bch2_btree_complete_write(c, b, btree_current_write(b));
 335                 clear_btree_node_dirty(b);
 336
 337                 mca_data_free(c, b);
 338         }
 339
 340         while (!list_empty(&c->btree_cache_freed)) {
 341                 b = list_first_entry(&c->btree_cache_freed,
 342                                      struct btree, list);
 343                 list_del(&b->list);
 344                 kfree(b);
 345         }
 346
 347         mutex_unlock(&c->btree_cache_lock);
 348
 349         if (c->btree_cache_table_init_done)
 350                 rhashtable_destroy(&c->btree_cache_table);
 351 }
 352
 353 int bch2_fs_btree_init(struct bch_fs *c)
 354 {
 355         unsigned i;
 356         int ret;
 357
 358         ret = rhashtable_init(&c->btree_cache_table, &bch_btree_cache_params);
 359         if (ret)
 360                 return ret;
 361
 362         c->btree_cache_table_init_done = true;
 363
 364         bch2_recalc_btree_reserve(c);
 365
 366         for (i = 0; i < c->btree_cache_reserve; i++)
 367                 if (!mca_bucket_alloc(c, GFP_KERNEL))
 368                         return -ENOMEM;
 369
 370         list_splice_init(&c->btree_cache,
 371                          &c->btree_cache_freeable);
 372
 373 #ifdef CONFIG_BCACHEFS_DEBUG
 374         mutex_init(&c->verify_lock);
 375
 376         c->verify_ondisk = (void *)
 377                 __get_free_pages(GFP_KERNEL, ilog2(btree_pages(c)));
 378         if (!c->verify_ondisk)
 379                 return -ENOMEM;
 380
 381         c->verify_data = mca_bucket_alloc(c, GFP_KERNEL);
 382         if (!c->verify_data)
 383                 return -ENOMEM;
 384
 385         list_del_init(&c->verify_data->list);
 386 #endif
 387
 388         c->btree_cache_shrink.count_objects = bch2_mca_count;
 389         c->btree_cache_shrink.scan_objects = bch2_mca_scan;
 390         c->btree_cache_shrink.seeks = 4;
 391         c->btree_cache_shrink.batch = btree_pages(c) * 2;
 392         register_shrinker(&c->btree_cache_shrink);
 393
 394         return 0;
 395 }
 396
 397 /*
 398  * We can only have one thread cannibalizing other cached btree nodes at a time,
 399  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 400  * cannibalize_bucket() will take. This means every time we unlock the root of
 401  * the btree, we need to release this lock if we have it held.
 402  */
 403 void bch2_btree_node_cannibalize_unlock(struct bch_fs *c)
 404 {
 405         if (c->btree_cache_alloc_lock == current) {
 406                 trace_btree_node_cannibalize_unlock(c);
 407                 c->btree_cache_alloc_lock = NULL;
 408                 closure_wake_up(&c->mca_wait);
 409         }
 410 }
 411
 412 int bch2_btree_node_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 413 {
 414         struct task_struct *old;
 415
 416         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 417         if (old == NULL || old == current)
 418                 goto success;
 419
 420         if (!cl) {
 421                 trace_btree_node_cannibalize_lock_fail(c);
 422                 return -ENOMEM;
 423         }
 424
 425         closure_wait(&c->mca_wait, cl);
 426
 427         /* Try again, after adding ourselves to waitlist */
 428         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 429         if (old == NULL || old == current) {
 430                 /* We raced */
 431                 closure_wake_up(&c->mca_wait);
 432                 goto success;
 433         }
 434
 435         trace_btree_node_cannibalize_lock_fail(c);
 436         return -EAGAIN;
 437
 438 success:
 439         trace_btree_node_cannibalize_lock(c);
 440         return 0;
 441 }
 442
 443 static struct btree *mca_cannibalize(struct bch_fs *c)
 444 {
 445         struct btree *b;
 446
 447         list_for_each_entry_reverse(b, &c->btree_cache, list)
 448                 if (!mca_reap(c, b, false))
 449                         return b;
 450
 451         while (1) {
 452                 list_for_each_entry_reverse(b, &c->btree_cache, list)
 453                         if (!mca_reap(c, b, true))
 454                                 return b;
 455
 456                 /*
 457                  * Rare case: all nodes were intent-locked.
 458                  * Just busy-wait.
 459                  */
 460                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 461                 cond_resched();
 462         }
 463 }
 464
 465 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 466 {
 467         struct btree *b;
 468         u64 start_time = local_clock();
 469
 470         mutex_lock(&c->btree_cache_lock);
 471
 472         /*
 473          * btree_free() doesn't free memory; it sticks the node on the end of
 474          * the list. Check if there's any freed nodes there:
 475          */
 476         list_for_each_entry(b, &c->btree_cache_freeable, list)
 477                 if (!mca_reap_notrace(c, b, false))
 478                         goto out_unlock;
 479
 480         /*
 481          * We never free struct btree itself, just the memory that holds the on
 482          * disk node. Check the freed list before allocating a new one:
 483          */
 484         list_for_each_entry(b, &c->btree_cache_freed, list)
 485                 if (!mca_reap_notrace(c, b, false)) {
 486                         mca_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
 487                         if (b->data)
 488                                 goto out_unlock;
 489
 490                         six_unlock_write(&b->lock);
 491                         six_unlock_intent(&b->lock);
 492                         goto err;
 493                 }
 494
 495         b = mca_bucket_alloc(c, __GFP_NOWARN|GFP_NOIO);
 496         if (!b)
 497                 goto err;
 498
 499         BUG_ON(!six_trylock_intent(&b->lock));
 500         BUG_ON(!six_trylock_write(&b->lock));
 501 out_unlock:
 502         BUG_ON(bkey_extent_is_data(&b->key.k) && PTR_HASH(&b->key));
 503         BUG_ON(btree_node_write_in_flight(b));
 504
 505         list_del_init(&b->list);
 506         mutex_unlock(&c->btree_cache_lock);
 507 out:
 508         b->flags                = 0;
 509         b->written              = 0;
 510         b->nsets                = 0;
 511         b->sib_u64s[0]          = 0;
 512         b->sib_u64s[1]          = 0;
 513         b->whiteout_u64s        = 0;
 514         b->uncompacted_whiteout_u64s = 0;
 515         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 516
 517         bch2_time_stats_update(&c->btree_node_mem_alloc_time, start_time);
 518
 519         return b;
 520 err:
 521         /* Try to cannibalize another cached btree node: */
 522         if (c->btree_cache_alloc_lock == current) {
 523                 b = mca_cannibalize(c);
 524                 list_del_init(&b->list);
 525                 mutex_unlock(&c->btree_cache_lock);
 526
 527                 bch2_btree_node_hash_remove(c, b);
 528
 529                 trace_btree_node_cannibalize(c);
 530                 goto out;
 531         }
 532
 533         mutex_unlock(&c->btree_cache_lock);
 534         return ERR_PTR(-ENOMEM);
 535 }
 536
 537 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 538 static noinline struct btree *bch2_btree_node_fill(struct btree_iter *iter,
 539                                                    const struct bkey_i *k,
 540                                                    unsigned level,
 541                                                    enum six_lock_type lock_type)
 542 {
 543         struct bch_fs *c = iter->c;
 544         struct btree *b;
 545
 546         b = bch2_btree_node_mem_alloc(c);
 547         if (IS_ERR(b))
 548                 return b;
 549
 550         bkey_copy(&b->key, k);
 551         if (bch2_btree_node_hash_insert(c, b, level, iter->btree_id)) {
 552                 /* raced with another fill: */
 553
 554                 /* mark as unhashed... */
 555                 bkey_i_to_extent(&b->key)->v._data[0] = 0;
 556
 557                 mutex_lock(&c->btree_cache_lock);
 558                 list_add(&b->list, &c->btree_cache_freeable);
 559                 mutex_unlock(&c->btree_cache_lock);
 560
 561                 six_unlock_write(&b->lock);
 562                 six_unlock_intent(&b->lock);
 563                 return NULL;
 564         }
 565
 566         /*
 567          * If the btree node wasn't cached, we can't drop our lock on
 568          * the parent until after it's added to the cache - because
 569          * otherwise we could race with a btree_split() freeing the node
 570          * we're trying to lock.
 571          *
 572          * But the deadlock described below doesn't exist in this case,
 573          * so it's safe to not drop the parent lock until here:
 574          */
 575         if (btree_node_read_locked(iter, level + 1))
 576                 btree_node_unlock(iter, level + 1);
 577
 578         bch2_btree_node_read(c, b);
 579         six_unlock_write(&b->lock);
 580
 581         if (lock_type == SIX_LOCK_read)
 582                 six_lock_downgrade(&b->lock);
 583
 584         return b;
 585 }
 586
 587 /**
 588  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 589  * in from disk if necessary.
 590  *
 591  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 592  *
 593  * The btree node will have either a read or a write lock held, depending on
 594  * the @write parameter.
 595  */
 596 struct btree *bch2_btree_node_get(struct btree_iter *iter,
 597                                   const struct bkey_i *k, unsigned level,
 598                                   enum six_lock_type lock_type)
 599 {
 600         struct btree *b;
 601         struct bset_tree *t;
 602
 603         BUG_ON(level >= BTREE_MAX_DEPTH);
 604 retry:
 605         rcu_read_lock();
 606         b = mca_find(iter->c, k);
 607         rcu_read_unlock();
 608
 609         if (unlikely(!b)) {
 610                 /*
 611                  * We must have the parent locked to call bch2_btree_node_fill(),
 612                  * else we could read in a btree node from disk that's been
 613                  * freed:
 614                  */
 615                 b = bch2_btree_node_fill(iter, k, level, lock_type);
 616
 617                 /* We raced and found the btree node in the cache */
 618                 if (!b)
 619                         goto retry;
 620
 621                 if (IS_ERR(b))
 622                         return b;
 623         } else {
 624                 /*
 625                  * There's a potential deadlock with splits and insertions into
 626                  * interior nodes we have to avoid:
 627                  *
 628                  * The other thread might be holding an intent lock on the node
 629                  * we want, and they want to update its parent node so they're
 630                  * going to upgrade their intent lock on the parent node to a
 631                  * write lock.
 632                  *
 633                  * But if we're holding a read lock on the parent, and we're
 634                  * trying to get the intent lock they're holding, we deadlock.
 635                  *
 636                  * So to avoid this we drop the read locks on parent nodes when
 637                  * we're starting to take intent locks - and handle the race.
 638                  *
 639                  * The race is that they might be about to free the node we
 640                  * want, and dropping our read lock on the parent node lets them
 641                  * update the parent marking the node we want as freed, and then
 642                  * free it:
 643                  *
 644                  * To guard against this, btree nodes are evicted from the cache
 645                  * when they're freed - and PTR_HASH() is zeroed out, which we
 646                  * check for after we lock the node.
 647                  *
 648                  * Then, bch2_btree_node_relock() on the parent will fail - because
 649                  * the parent was modified, when the pointer to the node we want
 650                  * was removed - and we'll bail out:
 651                  */
 652                 if (btree_node_read_locked(iter, level + 1))
 653                         btree_node_unlock(iter, level + 1);
 654
 655                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
 656                         return ERR_PTR(-EINTR);
 657
 658                 if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
 659                              b->level != level ||
 660                              race_fault())) {
 661                         six_unlock_type(&b->lock, lock_type);
 662                         if (bch2_btree_node_relock(iter, level + 1))
 663                                 goto retry;
 664
 665                         return ERR_PTR(-EINTR);
 666                 }
 667         }
 668
 669         prefetch(b->aux_data);
 670
 671         for_each_bset(b, t) {
 672                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 673
 674                 prefetch(p + L1_CACHE_BYTES * 0);
 675                 prefetch(p + L1_CACHE_BYTES * 1);
 676                 prefetch(p + L1_CACHE_BYTES * 2);
 677         }
 678
 679         /* avoid atomic set bit if it's not needed: */
 680         if (btree_node_accessed(b))
 681                 set_btree_node_accessed(b);
 682
 683         if (unlikely(btree_node_read_error(b))) {
 684                 six_unlock_type(&b->lock, lock_type);
 685                 return ERR_PTR(-EIO);
 686         }
 687
 688         EBUG_ON(!b->written);
 689         EBUG_ON(b->btree_id != iter->btree_id ||
 690                 BTREE_NODE_LEVEL(b->data) != level ||
 691                 bkey_cmp(b->data->max_key, k->k.p));
 692
 693         return b;
 694 }
 695
 696 int bch2_print_btree_node(struct bch_fs *c, struct btree *b,
 697                           char *buf, size_t len)
 698 {
 699         const struct bkey_format *f = &b->format;
 700         struct bset_stats stats;
 701         char ptrs[100];
 702
 703         memset(&stats, 0, sizeof(stats));
 704
 705         bch2_val_to_text(c, BKEY_TYPE_BTREE, ptrs, sizeof(ptrs),
 706                         bkey_i_to_s_c(&b->key));
 707         bch2_btree_keys_stats(b, &stats);
 708
 709         return scnprintf(buf, len,
 710                          "l %u %llu:%llu - %llu:%llu:\n"
 711                          "    ptrs: %s\n"
 712                          "    format: u64s %u fields %u %u %u %u %u\n"
 713                          "    unpack fn len: %u\n"
 714                          "    bytes used %zu/%zu (%zu%% full)\n"
 715                          "    sib u64s: %u, %u (merge threshold %zu)\n"
 716                          "    nr packed keys %u\n"
 717                          "    nr unpacked keys %u\n"
 718                          "    floats %zu\n"
 719                          "    failed unpacked %zu\n"
 720                          "    failed prev %zu\n"
 721                          "    failed overflow %zu\n",
 722                          b->level,
 723                          b->data->min_key.inode,
 724                          b->data->min_key.offset,
 725                          b->data->max_key.inode,
 726                          b->data->max_key.offset,
 727                          ptrs,
 728                          f->key_u64s,
 729                          f->bits_per_field[0],
 730                          f->bits_per_field[1],
 731                          f->bits_per_field[2],
 732                          f->bits_per_field[3],
 733                          f->bits_per_field[4],
 734                          b->unpack_fn_len,
 735                          b->nr.live_u64s * sizeof(u64),
 736                          btree_bytes(c) - sizeof(struct btree_node),
 737                          b->nr.live_u64s * 100 / btree_max_u64s(c),
 738                          b->sib_u64s[0],
 739                          b->sib_u64s[1],
 740                          BTREE_FOREGROUND_MERGE_THRESHOLD(c),
 741                          b->nr.packed_keys,
 742                          b->nr.unpacked_keys,
 743                          stats.floats,
 744                          stats.failed_unpacked,
 745                          stats.failed_prev,
 746                          stats.failed_overflow);
 747 }