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 "bkey_buf.h"
   5 #include "btree_cache.h"
   6 #include "btree_io.h"
   7 #include "btree_iter.h"
   8 #include "btree_locking.h"
   9 #include "debug.h"
  10 #include "error.h"
  11
  12 #include <linux/prefetch.h>
  13 #include <linux/sched/mm.h>
  14 #include <trace/events/bcachefs.h>
  15
  16 void bch2_recalc_btree_reserve(struct bch_fs *c)
  17 {
  18         unsigned i, reserve = 16;
  19
  20         if (!c->btree_roots[0].b)
  21                 reserve += 8;
  22
  23         for (i = 0; i < BTREE_ID_NR; i++)
  24                 if (c->btree_roots[i].b)
  25                         reserve += min_t(unsigned, 1,
  26                                          c->btree_roots[i].b->c.level) * 8;
  27
  28         c->btree_cache.reserve = reserve;
  29 }
  30
  31 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  32 {
  33         return max_t(int, 0, bc->used - bc->reserve);
  34 }
  35
  36 static void __btree_node_data_free(struct bch_fs *c, struct btree *b)
  37 {
  38         EBUG_ON(btree_node_write_in_flight(b));
  39
  40         kvpfree(b->data, btree_bytes(c));
  41         b->data = NULL;
  42         vfree(b->aux_data);
  43         b->aux_data = NULL;
  44 }
  45
  46 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  47 {
  48         struct btree_cache *bc = &c->btree_cache;
  49
  50         __btree_node_data_free(c, b);
  51         bc->used--;
  52         list_move(&b->list, &bc->freed);
  53 }
  54
  55 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  56                                    const void *obj)
  57 {
  58         const struct btree *b = obj;
  59         const u64 *v = arg->key;
  60
  61         return b->hash_val == *v ? 0 : 1;
  62 }
  63
  64 static const struct rhashtable_params bch_btree_cache_params = {
  65         .head_offset    = offsetof(struct btree, hash),
  66         .key_offset     = offsetof(struct btree, hash_val),
  67         .key_len        = sizeof(u64),
  68         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  69 };
  70
  71 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  72 {
  73         BUG_ON(b->data || b->aux_data);
  74
  75         b->data = kvpmalloc(btree_bytes(c), gfp);
  76         if (!b->data)
  77                 return -ENOMEM;
  78
  79         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  80         if (!b->aux_data) {
  81                 kvpfree(b->data, btree_bytes(c));
  82                 b->data = NULL;
  83                 return -ENOMEM;
  84         }
  85
  86         return 0;
  87 }
  88
  89 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
  90 {
  91         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
  92         if (!b)
  93                 return NULL;
  94
  95         bkey_btree_ptr_init(&b->key);
  96         six_lock_init(&b->c.lock);
  97         INIT_LIST_HEAD(&b->list);
  98         INIT_LIST_HEAD(&b->write_blocked);
  99         b->byte_order = ilog2(btree_bytes(c));
 100         return b;
 101 }
 102
 103 static struct btree *btree_node_mem_alloc(struct bch_fs *c)
 104 {
 105         struct btree_cache *bc = &c->btree_cache;
 106         struct btree *b = __btree_node_mem_alloc(c);
 107         if (!b)
 108                 return NULL;
 109
 110         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 111                 kfree(b);
 112                 return NULL;
 113         }
 114
 115         bc->used++;
 116         list_add(&b->list, &bc->freeable);
 117         return b;
 118 }
 119
 120 /* Btree in memory cache - hash table */
 121
 122 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 123 {
 124         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 125
 126         /* Cause future lookups for this node to fail: */
 127         b->hash_val = 0;
 128
 129         six_lock_wakeup_all(&b->c.lock);
 130 }
 131
 132 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 133 {
 134         BUG_ON(b->hash_val);
 135         b->hash_val = btree_ptr_hash_val(&b->key);
 136
 137         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 138                                              bch_btree_cache_params);
 139 }
 140
 141 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 142                                 unsigned level, enum btree_id id)
 143 {
 144         int ret;
 145
 146         b->c.level      = level;
 147         b->c.btree_id   = id;
 148
 149         if (level)
 150                 six_lock_pcpu_alloc(&b->c.lock);
 151         else
 152                 six_lock_pcpu_free_rcu(&b->c.lock);
 153
 154         mutex_lock(&bc->lock);
 155         ret = __bch2_btree_node_hash_insert(bc, b);
 156         if (!ret)
 157                 list_add(&b->list, &bc->live);
 158         mutex_unlock(&bc->lock);
 159
 160         return ret;
 161 }
 162
 163 __flatten
 164 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 165                                      const struct bkey_i *k)
 166 {
 167         u64 v = btree_ptr_hash_val(k);
 168
 169         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 170 }
 171
 172 /*
 173  * this version is for btree nodes that have already been freed (we're not
 174  * reaping a real btree node)
 175  */
 176 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 177 {
 178         struct btree_cache *bc = &c->btree_cache;
 179         int ret = 0;
 180
 181         lockdep_assert_held(&bc->lock);
 182
 183         if (!six_trylock_intent(&b->c.lock))
 184                 return -ENOMEM;
 185
 186         if (!six_trylock_write(&b->c.lock))
 187                 goto out_unlock_intent;
 188
 189         if (btree_node_noevict(b))
 190                 goto out_unlock;
 191
 192         if (!btree_node_may_write(b))
 193                 goto out_unlock;
 194
 195         if (btree_node_dirty(b) &&
 196             test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
 197                 goto out_unlock;
 198
 199         if (btree_node_dirty(b) ||
 200             btree_node_write_in_flight(b) ||
 201             btree_node_read_in_flight(b)) {
 202                 if (!flush)
 203                         goto out_unlock;
 204
 205                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 206                                TASK_UNINTERRUPTIBLE);
 207
 208                 /*
 209                  * Using the underscore version because we don't want to compact
 210                  * bsets after the write, since this node is about to be evicted
 211                  * - unless btree verify mode is enabled, since it runs out of
 212                  * the post write cleanup:
 213                  */
 214                 if (bch2_verify_btree_ondisk)
 215                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 216                 else
 217                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 218
 219                 /* wait for any in flight btree write */
 220                 btree_node_wait_on_io(b);
 221         }
 222 out:
 223         if (b->hash_val && !ret)
 224                 trace_btree_node_reap(c, b);
 225         return ret;
 226 out_unlock:
 227         six_unlock_write(&b->c.lock);
 228 out_unlock_intent:
 229         six_unlock_intent(&b->c.lock);
 230         ret = -ENOMEM;
 231         goto out;
 232 }
 233
 234 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 235 {
 236         return __btree_node_reclaim(c, b, false);
 237 }
 238
 239 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 240 {
 241         return __btree_node_reclaim(c, b, true);
 242 }
 243
 244 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 245                                            struct shrink_control *sc)
 246 {
 247         struct bch_fs *c = container_of(shrink, struct bch_fs,
 248                                         btree_cache.shrink);
 249         struct btree_cache *bc = &c->btree_cache;
 250         struct btree *b, *t;
 251         unsigned long nr = sc->nr_to_scan;
 252         unsigned long can_free;
 253         unsigned long touched = 0;
 254         unsigned long freed = 0;
 255         unsigned i, flags;
 256
 257         if (bch2_btree_shrinker_disabled)
 258                 return SHRINK_STOP;
 259
 260         /* Return -1 if we can't do anything right now */
 261         if (sc->gfp_mask & __GFP_FS)
 262                 mutex_lock(&bc->lock);
 263         else if (!mutex_trylock(&bc->lock))
 264                 return -1;
 265
 266         flags = memalloc_nofs_save();
 267
 268         /*
 269          * It's _really_ critical that we don't free too many btree nodes - we
 270          * have to always leave ourselves a reserve. The reserve is how we
 271          * guarantee that allocating memory for a new btree node can always
 272          * succeed, so that inserting keys into the btree can always succeed and
 273          * IO can always make forward progress:
 274          */
 275         nr /= btree_pages(c);
 276         can_free = btree_cache_can_free(bc);
 277         nr = min_t(unsigned long, nr, can_free);
 278
 279         i = 0;
 280         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 281                 touched++;
 282
 283                 if (freed >= nr)
 284                         break;
 285
 286                 if (++i > 3 &&
 287                     !btree_node_reclaim(c, b)) {
 288                         btree_node_data_free(c, b);
 289                         six_unlock_write(&b->c.lock);
 290                         six_unlock_intent(&b->c.lock);
 291                         freed++;
 292                 }
 293         }
 294 restart:
 295         list_for_each_entry_safe(b, t, &bc->live, list) {
 296                 touched++;
 297
 298                 if (freed >= nr) {
 299                         /* Save position */
 300                         if (&t->list != &bc->live)
 301                                 list_move_tail(&bc->live, &t->list);
 302                         break;
 303                 }
 304
 305                 if (!btree_node_accessed(b) &&
 306                     !btree_node_reclaim(c, b)) {
 307                         /* can't call bch2_btree_node_hash_remove under lock  */
 308                         freed++;
 309                         if (&t->list != &bc->live)
 310                                 list_move_tail(&bc->live, &t->list);
 311
 312                         btree_node_data_free(c, b);
 313                         mutex_unlock(&bc->lock);
 314
 315                         bch2_btree_node_hash_remove(bc, b);
 316                         six_unlock_write(&b->c.lock);
 317                         six_unlock_intent(&b->c.lock);
 318
 319                         if (freed >= nr)
 320                                 goto out;
 321
 322                         if (sc->gfp_mask & __GFP_FS)
 323                                 mutex_lock(&bc->lock);
 324                         else if (!mutex_trylock(&bc->lock))
 325                                 goto out;
 326                         goto restart;
 327                 } else
 328                         clear_btree_node_accessed(b);
 329         }
 330
 331         mutex_unlock(&bc->lock);
 332 out:
 333         memalloc_nofs_restore(flags);
 334         return (unsigned long) freed * btree_pages(c);
 335 }
 336
 337 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 338                                             struct shrink_control *sc)
 339 {
 340         struct bch_fs *c = container_of(shrink, struct bch_fs,
 341                                         btree_cache.shrink);
 342         struct btree_cache *bc = &c->btree_cache;
 343
 344         if (bch2_btree_shrinker_disabled)
 345                 return 0;
 346
 347         return btree_cache_can_free(bc) * btree_pages(c);
 348 }
 349
 350 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 351 {
 352         struct btree_cache *bc = &c->btree_cache;
 353         struct btree *b;
 354         unsigned i, flags;
 355
 356         if (bc->shrink.list.next)
 357                 unregister_shrinker(&bc->shrink);
 358
 359         /* vfree() can allocate memory: */
 360         flags = memalloc_nofs_save();
 361         mutex_lock(&bc->lock);
 362
 363 #ifdef CONFIG_BCACHEFS_DEBUG
 364         if (c->verify_data)
 365                 list_move(&c->verify_data->list, &bc->live);
 366
 367         kvpfree(c->verify_ondisk, btree_bytes(c));
 368 #endif
 369
 370         for (i = 0; i < BTREE_ID_NR; i++)
 371                 if (c->btree_roots[i].b)
 372                         list_add(&c->btree_roots[i].b->list, &bc->live);
 373
 374         list_splice(&bc->freeable, &bc->live);
 375
 376         while (!list_empty(&bc->live)) {
 377                 b = list_first_entry(&bc->live, struct btree, list);
 378
 379                 BUG_ON(btree_node_read_in_flight(b) ||
 380                        btree_node_write_in_flight(b));
 381
 382                 if (btree_node_dirty(b))
 383                         bch2_btree_complete_write(c, b, btree_current_write(b));
 384                 clear_btree_node_dirty(c, b);
 385
 386                 btree_node_data_free(c, b);
 387         }
 388
 389         BUG_ON(atomic_read(&c->btree_cache.dirty));
 390
 391         while (!list_empty(&bc->freed)) {
 392                 b = list_first_entry(&bc->freed, struct btree, list);
 393                 list_del(&b->list);
 394                 six_lock_pcpu_free(&b->c.lock);
 395                 kfree(b);
 396         }
 397
 398         mutex_unlock(&bc->lock);
 399         memalloc_nofs_restore(flags);
 400
 401         if (bc->table_init_done)
 402                 rhashtable_destroy(&bc->table);
 403 }
 404
 405 int bch2_fs_btree_cache_init(struct bch_fs *c)
 406 {
 407         struct btree_cache *bc = &c->btree_cache;
 408         unsigned i;
 409         int ret = 0;
 410
 411         pr_verbose_init(c->opts, "");
 412
 413         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 414         if (ret)
 415                 goto out;
 416
 417         bc->table_init_done = true;
 418
 419         bch2_recalc_btree_reserve(c);
 420
 421         for (i = 0; i < bc->reserve; i++)
 422                 if (!btree_node_mem_alloc(c)) {
 423                         ret = -ENOMEM;
 424                         goto out;
 425                 }
 426
 427         list_splice_init(&bc->live, &bc->freeable);
 428
 429 #ifdef CONFIG_BCACHEFS_DEBUG
 430         mutex_init(&c->verify_lock);
 431
 432         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 433         if (!c->verify_ondisk) {
 434                 ret = -ENOMEM;
 435                 goto out;
 436         }
 437
 438         c->verify_data = btree_node_mem_alloc(c);
 439         if (!c->verify_data) {
 440                 ret = -ENOMEM;
 441                 goto out;
 442         }
 443
 444         list_del_init(&c->verify_data->list);
 445 #endif
 446
 447         bc->shrink.count_objects        = bch2_btree_cache_count;
 448         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 449         bc->shrink.seeks                = 4;
 450         bc->shrink.batch                = btree_pages(c) * 2;
 451         ret = register_shrinker(&bc->shrink);
 452 out:
 453         pr_verbose_init(c->opts, "ret %i", ret);
 454         return ret;
 455 }
 456
 457 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 458 {
 459         mutex_init(&bc->lock);
 460         INIT_LIST_HEAD(&bc->live);
 461         INIT_LIST_HEAD(&bc->freeable);
 462         INIT_LIST_HEAD(&bc->freed);
 463 }
 464
 465 /*
 466  * We can only have one thread cannibalizing other cached btree nodes at a time,
 467  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 468  * cannibalize_bucket() will take. This means every time we unlock the root of
 469  * the btree, we need to release this lock if we have it held.
 470  */
 471 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 472 {
 473         struct btree_cache *bc = &c->btree_cache;
 474
 475         if (bc->alloc_lock == current) {
 476                 trace_btree_node_cannibalize_unlock(c);
 477                 bc->alloc_lock = NULL;
 478                 closure_wake_up(&bc->alloc_wait);
 479         }
 480 }
 481
 482 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 483 {
 484         struct btree_cache *bc = &c->btree_cache;
 485         struct task_struct *old;
 486
 487         old = cmpxchg(&bc->alloc_lock, NULL, current);
 488         if (old == NULL || old == current)
 489                 goto success;
 490
 491         if (!cl) {
 492                 trace_btree_node_cannibalize_lock_fail(c);
 493                 return -ENOMEM;
 494         }
 495
 496         closure_wait(&bc->alloc_wait, cl);
 497
 498         /* Try again, after adding ourselves to waitlist */
 499         old = cmpxchg(&bc->alloc_lock, NULL, current);
 500         if (old == NULL || old == current) {
 501                 /* We raced */
 502                 closure_wake_up(&bc->alloc_wait);
 503                 goto success;
 504         }
 505
 506         trace_btree_node_cannibalize_lock_fail(c);
 507         return -EAGAIN;
 508
 509 success:
 510         trace_btree_node_cannibalize_lock(c);
 511         return 0;
 512 }
 513
 514 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 515 {
 516         struct btree_cache *bc = &c->btree_cache;
 517         struct btree *b;
 518
 519         list_for_each_entry_reverse(b, &bc->live, list)
 520                 if (!btree_node_reclaim(c, b))
 521                         return b;
 522
 523         while (1) {
 524                 list_for_each_entry_reverse(b, &bc->live, list)
 525                         if (!btree_node_write_and_reclaim(c, b))
 526                                 return b;
 527
 528                 /*
 529                  * Rare case: all nodes were intent-locked.
 530                  * Just busy-wait.
 531                  */
 532                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 533                 cond_resched();
 534         }
 535 }
 536
 537 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 538 {
 539         struct btree_cache *bc = &c->btree_cache;
 540         struct btree *b;
 541         u64 start_time = local_clock();
 542         unsigned flags;
 543
 544         flags = memalloc_nofs_save();
 545         mutex_lock(&bc->lock);
 546
 547         /*
 548          * btree_free() doesn't free memory; it sticks the node on the end of
 549          * the list. Check if there's any freed nodes there:
 550          */
 551         list_for_each_entry(b, &bc->freeable, list)
 552                 if (!btree_node_reclaim(c, b))
 553                         goto got_node;
 554
 555         /*
 556          * We never free struct btree itself, just the memory that holds the on
 557          * disk node. Check the freed list before allocating a new one:
 558          */
 559         list_for_each_entry(b, &bc->freed, list)
 560                 if (!btree_node_reclaim(c, b))
 561                         goto got_node;
 562
 563         b = NULL;
 564 got_node:
 565         if (b)
 566                 list_del_init(&b->list);
 567         mutex_unlock(&bc->lock);
 568
 569         if (!b) {
 570                 b = __btree_node_mem_alloc(c);
 571                 if (!b)
 572                         goto err;
 573
 574                 BUG_ON(!six_trylock_intent(&b->c.lock));
 575                 BUG_ON(!six_trylock_write(&b->c.lock));
 576         }
 577
 578         if (!b->data) {
 579                 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 580                         goto err;
 581
 582                 mutex_lock(&bc->lock);
 583                 bc->used++;
 584                 mutex_unlock(&bc->lock);
 585         }
 586
 587         BUG_ON(btree_node_hashed(b));
 588         BUG_ON(btree_node_write_in_flight(b));
 589 out:
 590         b->flags                = 0;
 591         b->written              = 0;
 592         b->nsets                = 0;
 593         b->sib_u64s[0]          = 0;
 594         b->sib_u64s[1]          = 0;
 595         b->whiteout_u64s        = 0;
 596         bch2_btree_keys_init(b);
 597
 598         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 599                                start_time);
 600
 601         memalloc_nofs_restore(flags);
 602         return b;
 603 err:
 604         mutex_lock(&bc->lock);
 605
 606         if (b) {
 607                 list_add(&b->list, &bc->freed);
 608                 six_unlock_write(&b->c.lock);
 609                 six_unlock_intent(&b->c.lock);
 610         }
 611
 612         /* Try to cannibalize another cached btree node: */
 613         if (bc->alloc_lock == current) {
 614                 b = btree_node_cannibalize(c);
 615                 list_del_init(&b->list);
 616                 mutex_unlock(&bc->lock);
 617
 618                 bch2_btree_node_hash_remove(bc, b);
 619
 620                 trace_btree_node_cannibalize(c);
 621                 goto out;
 622         }
 623
 624         mutex_unlock(&bc->lock);
 625         memalloc_nofs_restore(flags);
 626         return ERR_PTR(-ENOMEM);
 627 }
 628
 629 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 630 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 631                                 struct btree_iter *iter,
 632                                 const struct bkey_i *k,
 633                                 enum btree_id btree_id,
 634                                 unsigned level,
 635                                 enum six_lock_type lock_type,
 636                                 bool sync)
 637 {
 638         struct btree_cache *bc = &c->btree_cache;
 639         struct btree *b;
 640
 641         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 642         /*
 643          * Parent node must be locked, else we could read in a btree node that's
 644          * been freed:
 645          */
 646         if (iter && !bch2_btree_node_relock(iter, level + 1))
 647                 return ERR_PTR(-EINTR);
 648
 649         b = bch2_btree_node_mem_alloc(c);
 650         if (IS_ERR(b))
 651                 return b;
 652
 653         bkey_copy(&b->key, k);
 654         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 655                 /* raced with another fill: */
 656
 657                 /* mark as unhashed... */
 658                 b->hash_val = 0;
 659
 660                 mutex_lock(&bc->lock);
 661                 list_add(&b->list, &bc->freeable);
 662                 mutex_unlock(&bc->lock);
 663
 664                 six_unlock_write(&b->c.lock);
 665                 six_unlock_intent(&b->c.lock);
 666                 return NULL;
 667         }
 668
 669         /*
 670          * Unlock before doing IO:
 671          *
 672          * XXX: ideally should be dropping all btree node locks here
 673          */
 674         if (iter && btree_node_read_locked(iter, level + 1))
 675                 btree_node_unlock(iter, level + 1);
 676
 677         bch2_btree_node_read(c, b, sync);
 678
 679         six_unlock_write(&b->c.lock);
 680
 681         if (!sync) {
 682                 six_unlock_intent(&b->c.lock);
 683                 return NULL;
 684         }
 685
 686         if (lock_type == SIX_LOCK_read)
 687                 six_lock_downgrade(&b->c.lock);
 688
 689         return b;
 690 }
 691
 692 static int lock_node_check_fn(struct six_lock *lock, void *p)
 693 {
 694         struct btree *b = container_of(lock, struct btree, c.lock);
 695         const struct bkey_i *k = p;
 696
 697         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 698 }
 699
 700 /**
 701  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 702  * in from disk if necessary.
 703  *
 704  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 705  *
 706  * The btree node will have either a read or a write lock held, depending on
 707  * the @write parameter.
 708  */
 709 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 710                                   const struct bkey_i *k, unsigned level,
 711                                   enum six_lock_type lock_type,
 712                                   unsigned long trace_ip)
 713 {
 714         struct btree_cache *bc = &c->btree_cache;
 715         struct btree *b;
 716         struct bset_tree *t;
 717
 718         EBUG_ON(level >= BTREE_MAX_DEPTH);
 719
 720         b = btree_node_mem_ptr(k);
 721         if (b)
 722                 goto lock_node;
 723 retry:
 724         b = btree_cache_find(bc, k);
 725         if (unlikely(!b)) {
 726                 /*
 727                  * We must have the parent locked to call bch2_btree_node_fill(),
 728                  * else we could read in a btree node from disk that's been
 729                  * freed:
 730                  */
 731                 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
 732                                          level, lock_type, true);
 733
 734                 /* We raced and found the btree node in the cache */
 735                 if (!b)
 736                         goto retry;
 737
 738                 if (IS_ERR(b))
 739                         return b;
 740         } else {
 741 lock_node:
 742                 /*
 743                  * There's a potential deadlock with splits and insertions into
 744                  * interior nodes we have to avoid:
 745                  *
 746                  * The other thread might be holding an intent lock on the node
 747                  * we want, and they want to update its parent node so they're
 748                  * going to upgrade their intent lock on the parent node to a
 749                  * write lock.
 750                  *
 751                  * But if we're holding a read lock on the parent, and we're
 752                  * trying to get the intent lock they're holding, we deadlock.
 753                  *
 754                  * So to avoid this we drop the read locks on parent nodes when
 755                  * we're starting to take intent locks - and handle the race.
 756                  *
 757                  * The race is that they might be about to free the node we
 758                  * want, and dropping our read lock on the parent node lets them
 759                  * update the parent marking the node we want as freed, and then
 760                  * free it:
 761                  *
 762                  * To guard against this, btree nodes are evicted from the cache
 763                  * when they're freed - and b->hash_val is zeroed out, which we
 764                  * check for after we lock the node.
 765                  *
 766                  * Then, bch2_btree_node_relock() on the parent will fail - because
 767                  * the parent was modified, when the pointer to the node we want
 768                  * was removed - and we'll bail out:
 769                  */
 770                 if (btree_node_read_locked(iter, level + 1))
 771                         btree_node_unlock(iter, level + 1);
 772
 773                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
 774                                      lock_node_check_fn, (void *) k, trace_ip)) {
 775                         if (b->hash_val != btree_ptr_hash_val(k))
 776                                 goto retry;
 777                         return ERR_PTR(-EINTR);
 778                 }
 779
 780                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 781                              b->c.level != level ||
 782                              race_fault())) {
 783                         six_unlock_type(&b->c.lock, lock_type);
 784                         if (bch2_btree_node_relock(iter, level + 1))
 785                                 goto retry;
 786
 787                         trace_trans_restart_btree_node_reused(iter->trans->ip);
 788                         return ERR_PTR(-EINTR);
 789                 }
 790         }
 791
 792         /* XXX: waiting on IO with btree locks held: */
 793         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 794                        TASK_UNINTERRUPTIBLE);
 795
 796         prefetch(b->aux_data);
 797
 798         for_each_bset(b, t) {
 799                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 800
 801                 prefetch(p + L1_CACHE_BYTES * 0);
 802                 prefetch(p + L1_CACHE_BYTES * 1);
 803                 prefetch(p + L1_CACHE_BYTES * 2);
 804         }
 805
 806         /* avoid atomic set bit if it's not needed: */
 807         if (!btree_node_accessed(b))
 808                 set_btree_node_accessed(b);
 809
 810         if (unlikely(btree_node_read_error(b))) {
 811                 six_unlock_type(&b->c.lock, lock_type);
 812                 return ERR_PTR(-EIO);
 813         }
 814
 815         EBUG_ON(b->c.btree_id != iter->btree_id);
 816         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 817         EBUG_ON(bpos_cmp(b->data->max_key, k->k.p));
 818         EBUG_ON(b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 819                 bpos_cmp(b->data->min_key,
 820                          bkey_i_to_btree_ptr_v2(&b->key)->v.min_key));
 821
 822         return b;
 823 }
 824
 825 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 826                                          const struct bkey_i *k,
 827                                          enum btree_id btree_id,
 828                                          unsigned level,
 829                                          bool nofill)
 830 {
 831         struct btree_cache *bc = &c->btree_cache;
 832         struct btree *b;
 833         struct bset_tree *t;
 834         int ret;
 835
 836         EBUG_ON(level >= BTREE_MAX_DEPTH);
 837
 838         b = btree_node_mem_ptr(k);
 839         if (b)
 840                 goto lock_node;
 841 retry:
 842         b = btree_cache_find(bc, k);
 843         if (unlikely(!b)) {
 844                 if (nofill)
 845                         goto out;
 846
 847                 b = bch2_btree_node_fill(c, NULL, k, btree_id,
 848                                          level, SIX_LOCK_read, true);
 849
 850                 /* We raced and found the btree node in the cache */
 851                 if (!b)
 852                         goto retry;
 853
 854                 if (IS_ERR(b) &&
 855                     !bch2_btree_cache_cannibalize_lock(c, NULL))
 856                         goto retry;
 857
 858                 if (IS_ERR(b))
 859                         goto out;
 860         } else {
 861 lock_node:
 862                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 863                 if (ret)
 864                         goto retry;
 865
 866                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 867                              b->c.btree_id != btree_id ||
 868                              b->c.level != level)) {
 869                         six_unlock_read(&b->c.lock);
 870                         goto retry;
 871                 }
 872         }
 873
 874         /* XXX: waiting on IO with btree locks held: */
 875         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 876                        TASK_UNINTERRUPTIBLE);
 877
 878         prefetch(b->aux_data);
 879
 880         for_each_bset(b, t) {
 881                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 882
 883                 prefetch(p + L1_CACHE_BYTES * 0);
 884                 prefetch(p + L1_CACHE_BYTES * 1);
 885                 prefetch(p + L1_CACHE_BYTES * 2);
 886         }
 887
 888         /* avoid atomic set bit if it's not needed: */
 889         if (!btree_node_accessed(b))
 890                 set_btree_node_accessed(b);
 891
 892         if (unlikely(btree_node_read_error(b))) {
 893                 six_unlock_read(&b->c.lock);
 894                 b = ERR_PTR(-EIO);
 895                 goto out;
 896         }
 897
 898         EBUG_ON(b->c.btree_id != btree_id);
 899         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 900         EBUG_ON(bpos_cmp(b->data->max_key, k->k.p));
 901         EBUG_ON(b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 902                 bpos_cmp(b->data->min_key,
 903                          bkey_i_to_btree_ptr_v2(&b->key)->v.min_key));
 904 out:
 905         bch2_btree_cache_cannibalize_unlock(c);
 906         return b;
 907 }
 908
 909 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 910                               const struct bkey_i *k,
 911                               enum btree_id btree_id, unsigned level)
 912 {
 913         struct btree_cache *bc = &c->btree_cache;
 914         struct btree *b;
 915
 916         BUG_ON(iter && !btree_node_locked(iter, level + 1));
 917         BUG_ON(level >= BTREE_MAX_DEPTH);
 918
 919         b = btree_cache_find(bc, k);
 920         if (b)
 921                 return;
 922
 923         bch2_btree_node_fill(c, iter, k, btree_id, level, SIX_LOCK_read, false);
 924 }
 925
 926 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
 927                              struct btree *b)
 928 {
 929         const struct bkey_format *f = &b->format;
 930         struct bset_stats stats;
 931
 932         memset(&stats, 0, sizeof(stats));
 933
 934         bch2_btree_keys_stats(b, &stats);
 935
 936         pr_buf(out, "l %u ", b->c.level);
 937         bch2_bpos_to_text(out, b->data->min_key);
 938         pr_buf(out, " - ");
 939         bch2_bpos_to_text(out, b->data->max_key);
 940         pr_buf(out, ":\n"
 941                "    ptrs: ");
 942         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 943
 944         pr_buf(out, "\n"
 945                "    format: u64s %u fields %u %u %u %u %u\n"
 946                "    unpack fn len: %u\n"
 947                "    bytes used %zu/%zu (%zu%% full)\n"
 948                "    sib u64s: %u, %u (merge threshold %u)\n"
 949                "    nr packed keys %u\n"
 950                "    nr unpacked keys %u\n"
 951                "    floats %zu\n"
 952                "    failed unpacked %zu\n",
 953                f->key_u64s,
 954                f->bits_per_field[0],
 955                f->bits_per_field[1],
 956                f->bits_per_field[2],
 957                f->bits_per_field[3],
 958                f->bits_per_field[4],
 959                b->unpack_fn_len,
 960                b->nr.live_u64s * sizeof(u64),
 961                btree_bytes(c) - sizeof(struct btree_node),
 962                b->nr.live_u64s * 100 / btree_max_u64s(c),
 963                b->sib_u64s[0],
 964                b->sib_u64s[1],
 965                c->btree_foreground_merge_threshold,
 966                b->nr.packed_keys,
 967                b->nr.unpacked_keys,
 968                stats.floats,
 969                stats.failed);
 970 }
 971
 972 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
 973 {
 974         pr_buf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
 975         pr_buf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
 976         pr_buf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
 977 }