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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_cache.h"
   5 #include "btree_io.h"
   6 #include "btree_iter.h"
   7 #include "btree_locking.h"
   8 #include "debug.h"
   9
  10 #include <linux/prefetch.h>
  11 #include <linux/sched/mm.h>
  12 #include <trace/events/bcachefs.h>
  13
  14 const char * const bch2_btree_ids[] = {
  15 #define x(kwd, val, name) name,
  16         BCH_BTREE_IDS()
  17 #undef x
  18         NULL
  19 };
  20
  21 void bch2_recalc_btree_reserve(struct bch_fs *c)
  22 {
  23         unsigned i, reserve = 16;
  24
  25         if (!c->btree_roots[0].b)
  26                 reserve += 8;
  27
  28         for (i = 0; i < BTREE_ID_NR; i++)
  29                 if (c->btree_roots[i].b)
  30                         reserve += min_t(unsigned, 1,
  31                                          c->btree_roots[i].b->c.level) * 8;
  32
  33         c->btree_cache.reserve = reserve;
  34 }
  35
  36 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  37 {
  38         return max_t(int, 0, bc->used - bc->reserve);
  39 }
  40
  41 static void __btree_node_data_free(struct bch_fs *c, struct btree *b)
  42 {
  43         EBUG_ON(btree_node_write_in_flight(b));
  44
  45         kvpfree(b->data, btree_bytes(c));
  46         b->data = NULL;
  47         vfree(b->aux_data);
  48         b->aux_data = NULL;
  49 }
  50
  51 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  52 {
  53         struct btree_cache *bc = &c->btree_cache;
  54
  55         __btree_node_data_free(c, b);
  56         bc->used--;
  57         list_move(&b->list, &bc->freed);
  58 }
  59
  60 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  61                                    const void *obj)
  62 {
  63         const struct btree *b = obj;
  64         const u64 *v = arg->key;
  65
  66         return b->hash_val == *v ? 0 : 1;
  67 }
  68
  69 static const struct rhashtable_params bch_btree_cache_params = {
  70         .head_offset    = offsetof(struct btree, hash),
  71         .key_offset     = offsetof(struct btree, hash_val),
  72         .key_len        = sizeof(u64),
  73         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  74 };
  75
  76 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  77 {
  78         BUG_ON(b->data || b->aux_data);
  79
  80         b->data = kvpmalloc(btree_bytes(c), gfp);
  81         if (!b->data)
  82                 return -ENOMEM;
  83
  84         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  85         if (!b->aux_data) {
  86                 kvpfree(b->data, btree_bytes(c));
  87                 b->data = NULL;
  88                 return -ENOMEM;
  89         }
  90
  91         return 0;
  92 }
  93
  94 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
  95 {
  96         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
  97         if (!b)
  98                 return NULL;
  99
 100         bkey_btree_ptr_init(&b->key);
 101         six_lock_init(&b->c.lock);
 102         INIT_LIST_HEAD(&b->list);
 103         INIT_LIST_HEAD(&b->write_blocked);
 104         b->byte_order = ilog2(btree_bytes(c));
 105         return b;
 106 }
 107
 108 static struct btree *btree_node_mem_alloc(struct bch_fs *c)
 109 {
 110         struct btree_cache *bc = &c->btree_cache;
 111         struct btree *b = __btree_node_mem_alloc(c);
 112         if (!b)
 113                 return NULL;
 114
 115         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 116                 kfree(b);
 117                 return NULL;
 118         }
 119
 120         bc->used++;
 121         list_add(&b->list, &bc->freeable);
 122         return b;
 123 }
 124
 125 /* Btree in memory cache - hash table */
 126
 127 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 128 {
 129         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 130
 131         /* Cause future lookups for this node to fail: */
 132         b->hash_val = 0;
 133
 134         six_lock_wakeup_all(&b->c.lock);
 135 }
 136
 137 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 138 {
 139         BUG_ON(b->hash_val);
 140         b->hash_val = btree_ptr_hash_val(&b->key);
 141
 142         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 143                                              bch_btree_cache_params);
 144 }
 145
 146 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 147                                 unsigned level, enum btree_id id)
 148 {
 149         int ret;
 150
 151         b->c.level      = level;
 152         b->c.btree_id   = id;
 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                 kfree(b);
 395         }
 396
 397         mutex_unlock(&bc->lock);
 398         memalloc_nofs_restore(flags);
 399
 400         if (bc->table_init_done)
 401                 rhashtable_destroy(&bc->table);
 402 }
 403
 404 int bch2_fs_btree_cache_init(struct bch_fs *c)
 405 {
 406         struct btree_cache *bc = &c->btree_cache;
 407         unsigned i;
 408         int ret = 0;
 409
 410         pr_verbose_init(c->opts, "");
 411
 412         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 413         if (ret)
 414                 goto out;
 415
 416         bc->table_init_done = true;
 417
 418         bch2_recalc_btree_reserve(c);
 419
 420         for (i = 0; i < bc->reserve; i++)
 421                 if (!btree_node_mem_alloc(c)) {
 422                         ret = -ENOMEM;
 423                         goto out;
 424                 }
 425
 426         list_splice_init(&bc->live, &bc->freeable);
 427
 428 #ifdef CONFIG_BCACHEFS_DEBUG
 429         mutex_init(&c->verify_lock);
 430
 431         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 432         if (!c->verify_ondisk) {
 433                 ret = -ENOMEM;
 434                 goto out;
 435         }
 436
 437         c->verify_data = btree_node_mem_alloc(c);
 438         if (!c->verify_data) {
 439                 ret = -ENOMEM;
 440                 goto out;
 441         }
 442
 443         list_del_init(&c->verify_data->list);
 444 #endif
 445
 446         bc->shrink.count_objects        = bch2_btree_cache_count;
 447         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 448         bc->shrink.seeks                = 4;
 449         bc->shrink.batch                = btree_pages(c) * 2;
 450         ret = register_shrinker(&bc->shrink);
 451 out:
 452         pr_verbose_init(c->opts, "ret %i", ret);
 453         return ret;
 454 }
 455
 456 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 457 {
 458         mutex_init(&bc->lock);
 459         INIT_LIST_HEAD(&bc->live);
 460         INIT_LIST_HEAD(&bc->freeable);
 461         INIT_LIST_HEAD(&bc->freed);
 462 }
 463
 464 /*
 465  * We can only have one thread cannibalizing other cached btree nodes at a time,
 466  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 467  * cannibalize_bucket() will take. This means every time we unlock the root of
 468  * the btree, we need to release this lock if we have it held.
 469  */
 470 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 471 {
 472         struct btree_cache *bc = &c->btree_cache;
 473
 474         if (bc->alloc_lock == current) {
 475                 trace_btree_node_cannibalize_unlock(c);
 476                 bc->alloc_lock = NULL;
 477                 closure_wake_up(&bc->alloc_wait);
 478         }
 479 }
 480
 481 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 482 {
 483         struct btree_cache *bc = &c->btree_cache;
 484         struct task_struct *old;
 485
 486         old = cmpxchg(&bc->alloc_lock, NULL, current);
 487         if (old == NULL || old == current)
 488                 goto success;
 489
 490         if (!cl) {
 491                 trace_btree_node_cannibalize_lock_fail(c);
 492                 return -ENOMEM;
 493         }
 494
 495         closure_wait(&bc->alloc_wait, cl);
 496
 497         /* Try again, after adding ourselves to waitlist */
 498         old = cmpxchg(&bc->alloc_lock, NULL, current);
 499         if (old == NULL || old == current) {
 500                 /* We raced */
 501                 closure_wake_up(&bc->alloc_wait);
 502                 goto success;
 503         }
 504
 505         trace_btree_node_cannibalize_lock_fail(c);
 506         return -EAGAIN;
 507
 508 success:
 509         trace_btree_node_cannibalize_lock(c);
 510         return 0;
 511 }
 512
 513 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 514 {
 515         struct btree_cache *bc = &c->btree_cache;
 516         struct btree *b;
 517
 518         list_for_each_entry_reverse(b, &bc->live, list)
 519                 if (!btree_node_reclaim(c, b))
 520                         return b;
 521
 522         while (1) {
 523                 list_for_each_entry_reverse(b, &bc->live, list)
 524                         if (!btree_node_write_and_reclaim(c, b))
 525                                 return b;
 526
 527                 /*
 528                  * Rare case: all nodes were intent-locked.
 529                  * Just busy-wait.
 530                  */
 531                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 532                 cond_resched();
 533         }
 534 }
 535
 536 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 537 {
 538         struct btree_cache *bc = &c->btree_cache;
 539         struct btree *b;
 540         u64 start_time = local_clock();
 541         unsigned flags;
 542
 543         flags = memalloc_nofs_save();
 544         mutex_lock(&bc->lock);
 545
 546         /*
 547          * btree_free() doesn't free memory; it sticks the node on the end of
 548          * the list. Check if there's any freed nodes there:
 549          */
 550         list_for_each_entry(b, &bc->freeable, list)
 551                 if (!btree_node_reclaim(c, b))
 552                         goto got_node;
 553
 554         /*
 555          * We never free struct btree itself, just the memory that holds the on
 556          * disk node. Check the freed list before allocating a new one:
 557          */
 558         list_for_each_entry(b, &bc->freed, list)
 559                 if (!btree_node_reclaim(c, b))
 560                         goto got_node;
 561
 562         b = NULL;
 563 got_node:
 564         if (b)
 565                 list_del_init(&b->list);
 566         mutex_unlock(&bc->lock);
 567
 568         if (!b) {
 569                 b = __btree_node_mem_alloc(c);
 570                 if (!b)
 571                         goto err;
 572
 573                 BUG_ON(!six_trylock_intent(&b->c.lock));
 574                 BUG_ON(!six_trylock_write(&b->c.lock));
 575         }
 576
 577         if (!b->data) {
 578                 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 579                         goto err;
 580
 581                 mutex_lock(&bc->lock);
 582                 bc->used++;
 583                 mutex_unlock(&bc->lock);
 584         }
 585
 586         BUG_ON(btree_node_hashed(b));
 587         BUG_ON(btree_node_write_in_flight(b));
 588 out:
 589         b->flags                = 0;
 590         b->written              = 0;
 591         b->nsets                = 0;
 592         b->sib_u64s[0]          = 0;
 593         b->sib_u64s[1]          = 0;
 594         b->whiteout_u64s        = 0;
 595         bch2_btree_keys_init(b);
 596
 597         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 598                                start_time);
 599
 600         memalloc_nofs_restore(flags);
 601         return b;
 602 err:
 603         mutex_lock(&bc->lock);
 604
 605         if (b) {
 606                 list_add(&b->list, &bc->freed);
 607                 six_unlock_write(&b->c.lock);
 608                 six_unlock_intent(&b->c.lock);
 609         }
 610
 611         /* Try to cannibalize another cached btree node: */
 612         if (bc->alloc_lock == current) {
 613                 b = btree_node_cannibalize(c);
 614                 list_del_init(&b->list);
 615                 mutex_unlock(&bc->lock);
 616
 617                 bch2_btree_node_hash_remove(bc, b);
 618
 619                 trace_btree_node_cannibalize(c);
 620                 goto out;
 621         }
 622
 623         mutex_unlock(&bc->lock);
 624         memalloc_nofs_restore(flags);
 625         return ERR_PTR(-ENOMEM);
 626 }
 627
 628 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 629 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 630                                 struct btree_iter *iter,
 631                                 const struct bkey_i *k,
 632                                 enum btree_id btree_id,
 633                                 unsigned level,
 634                                 enum six_lock_type lock_type,
 635                                 bool sync)
 636 {
 637         struct btree_cache *bc = &c->btree_cache;
 638         struct btree *b;
 639
 640         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 641         /*
 642          * Parent node must be locked, else we could read in a btree node that's
 643          * been freed:
 644          */
 645         if (iter && !bch2_btree_node_relock(iter, level + 1))
 646                 return ERR_PTR(-EINTR);
 647
 648         b = bch2_btree_node_mem_alloc(c);
 649         if (IS_ERR(b))
 650                 return b;
 651
 652         bkey_copy(&b->key, k);
 653         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 654                 /* raced with another fill: */
 655
 656                 /* mark as unhashed... */
 657                 b->hash_val = 0;
 658
 659                 mutex_lock(&bc->lock);
 660                 list_add(&b->list, &bc->freeable);
 661                 mutex_unlock(&bc->lock);
 662
 663                 six_unlock_write(&b->c.lock);
 664                 six_unlock_intent(&b->c.lock);
 665                 return NULL;
 666         }
 667
 668         /*
 669          * Unlock before doing IO:
 670          *
 671          * XXX: ideally should be dropping all btree node locks here
 672          */
 673         if (iter && btree_node_read_locked(iter, level + 1))
 674                 btree_node_unlock(iter, level + 1);
 675
 676         bch2_btree_node_read(c, b, sync);
 677
 678         six_unlock_write(&b->c.lock);
 679
 680         if (!sync) {
 681                 six_unlock_intent(&b->c.lock);
 682                 return NULL;
 683         }
 684
 685         if (lock_type == SIX_LOCK_read)
 686                 six_lock_downgrade(&b->c.lock);
 687
 688         return b;
 689 }
 690
 691 static int lock_node_check_fn(struct six_lock *lock, void *p)
 692 {
 693         struct btree *b = container_of(lock, struct btree, c.lock);
 694         const struct bkey_i *k = p;
 695
 696         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 697 }
 698
 699 /**
 700  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 701  * in from disk if necessary.
 702  *
 703  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 704  *
 705  * The btree node will have either a read or a write lock held, depending on
 706  * the @write parameter.
 707  */
 708 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 709                                   const struct bkey_i *k, unsigned level,
 710                                   enum six_lock_type lock_type,
 711                                   unsigned long trace_ip)
 712 {
 713         struct btree_cache *bc = &c->btree_cache;
 714         struct btree *b;
 715         struct bset_tree *t;
 716
 717         EBUG_ON(level >= BTREE_MAX_DEPTH);
 718
 719         b = btree_node_mem_ptr(k);
 720         if (b)
 721                 goto lock_node;
 722 retry:
 723         b = btree_cache_find(bc, k);
 724         if (unlikely(!b)) {
 725                 /*
 726                  * We must have the parent locked to call bch2_btree_node_fill(),
 727                  * else we could read in a btree node from disk that's been
 728                  * freed:
 729                  */
 730                 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
 731                                          level, lock_type, true);
 732
 733                 /* We raced and found the btree node in the cache */
 734                 if (!b)
 735                         goto retry;
 736
 737                 if (IS_ERR(b))
 738                         return b;
 739         } else {
 740 lock_node:
 741                 /*
 742                  * There's a potential deadlock with splits and insertions into
 743                  * interior nodes we have to avoid:
 744                  *
 745                  * The other thread might be holding an intent lock on the node
 746                  * we want, and they want to update its parent node so they're
 747                  * going to upgrade their intent lock on the parent node to a
 748                  * write lock.
 749                  *
 750                  * But if we're holding a read lock on the parent, and we're
 751                  * trying to get the intent lock they're holding, we deadlock.
 752                  *
 753                  * So to avoid this we drop the read locks on parent nodes when
 754                  * we're starting to take intent locks - and handle the race.
 755                  *
 756                  * The race is that they might be about to free the node we
 757                  * want, and dropping our read lock on the parent node lets them
 758                  * update the parent marking the node we want as freed, and then
 759                  * free it:
 760                  *
 761                  * To guard against this, btree nodes are evicted from the cache
 762                  * when they're freed - and b->hash_val is zeroed out, which we
 763                  * check for after we lock the node.
 764                  *
 765                  * Then, bch2_btree_node_relock() on the parent will fail - because
 766                  * the parent was modified, when the pointer to the node we want
 767                  * was removed - and we'll bail out:
 768                  */
 769                 if (btree_node_read_locked(iter, level + 1))
 770                         btree_node_unlock(iter, level + 1);
 771
 772                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
 773                                      lock_node_check_fn, (void *) k, trace_ip)) {
 774                         if (b->hash_val != btree_ptr_hash_val(k))
 775                                 goto retry;
 776                         return ERR_PTR(-EINTR);
 777                 }
 778
 779                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 780                              b->c.level != level ||
 781                              race_fault())) {
 782                         six_unlock_type(&b->c.lock, lock_type);
 783                         if (bch2_btree_node_relock(iter, level + 1))
 784                                 goto retry;
 785
 786                         trace_trans_restart_btree_node_reused(iter->trans->ip);
 787                         return ERR_PTR(-EINTR);
 788                 }
 789         }
 790
 791         /* XXX: waiting on IO with btree locks held: */
 792         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 793                        TASK_UNINTERRUPTIBLE);
 794
 795         prefetch(b->aux_data);
 796
 797         for_each_bset(b, t) {
 798                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 799
 800                 prefetch(p + L1_CACHE_BYTES * 0);
 801                 prefetch(p + L1_CACHE_BYTES * 1);
 802                 prefetch(p + L1_CACHE_BYTES * 2);
 803         }
 804
 805         /* avoid atomic set bit if it's not needed: */
 806         if (!btree_node_accessed(b))
 807                 set_btree_node_accessed(b);
 808
 809         if (unlikely(btree_node_read_error(b))) {
 810                 six_unlock_type(&b->c.lock, lock_type);
 811                 return ERR_PTR(-EIO);
 812         }
 813
 814         EBUG_ON(b->c.btree_id != iter->btree_id ||
 815                 BTREE_NODE_LEVEL(b->data) != level ||
 816                 bkey_cmp(b->data->max_key, k->k.p));
 817
 818         return b;
 819 }
 820
 821 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 822                                          const struct bkey_i *k,
 823                                          enum btree_id btree_id,
 824                                          unsigned level)
 825 {
 826         struct btree_cache *bc = &c->btree_cache;
 827         struct btree *b;
 828         struct bset_tree *t;
 829         int ret;
 830
 831         EBUG_ON(level >= BTREE_MAX_DEPTH);
 832
 833         b = btree_node_mem_ptr(k);
 834         if (b)
 835                 goto lock_node;
 836 retry:
 837         b = btree_cache_find(bc, k);
 838         if (unlikely(!b)) {
 839                 b = bch2_btree_node_fill(c, NULL, k, btree_id,
 840                                          level, SIX_LOCK_read, true);
 841
 842                 /* We raced and found the btree node in the cache */
 843                 if (!b)
 844                         goto retry;
 845
 846                 if (IS_ERR(b))
 847                         return b;
 848         } else {
 849 lock_node:
 850                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 851                 if (ret)
 852                         goto retry;
 853
 854                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 855                              b->c.btree_id != btree_id ||
 856                              b->c.level != level)) {
 857                         six_unlock_read(&b->c.lock);
 858                         goto retry;
 859                 }
 860         }
 861
 862         /* XXX: waiting on IO with btree locks held: */
 863         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 864                        TASK_UNINTERRUPTIBLE);
 865
 866         prefetch(b->aux_data);
 867
 868         for_each_bset(b, t) {
 869                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 870
 871                 prefetch(p + L1_CACHE_BYTES * 0);
 872                 prefetch(p + L1_CACHE_BYTES * 1);
 873                 prefetch(p + L1_CACHE_BYTES * 2);
 874         }
 875
 876         /* avoid atomic set bit if it's not needed: */
 877         if (!btree_node_accessed(b))
 878                 set_btree_node_accessed(b);
 879
 880         if (unlikely(btree_node_read_error(b))) {
 881                 six_unlock_read(&b->c.lock);
 882                 return ERR_PTR(-EIO);
 883         }
 884
 885         EBUG_ON(b->c.btree_id != btree_id ||
 886                 BTREE_NODE_LEVEL(b->data) != level ||
 887                 bkey_cmp(b->data->max_key, k->k.p));
 888
 889         return b;
 890 }
 891
 892 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 893                                           struct btree_iter *iter,
 894                                           struct btree *b,
 895                                           enum btree_node_sibling sib)
 896 {
 897         struct btree_trans *trans = iter->trans;
 898         struct btree *parent;
 899         struct btree_node_iter node_iter;
 900         struct bkey_packed *k;
 901         BKEY_PADDED(k) tmp;
 902         struct btree *ret = NULL;
 903         unsigned level = b->c.level;
 904
 905         parent = btree_iter_node(iter, level + 1);
 906         if (!parent)
 907                 return NULL;
 908
 909         /*
 910          * There's a corner case where a btree_iter might have a node locked
 911          * that is just outside its current pos - when
 912          * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
 913          *
 914          * But the lock ordering checks in __bch2_btree_node_lock() go off of
 915          * iter->pos, not the node's key: so if the iterator is marked as
 916          * needing to be traversed, we risk deadlock if we don't bail out here:
 917          */
 918         if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
 919                 return ERR_PTR(-EINTR);
 920
 921         if (!bch2_btree_node_relock(iter, level + 1)) {
 922                 ret = ERR_PTR(-EINTR);
 923                 goto out;
 924         }
 925
 926         node_iter = iter->l[parent->c.level].iter;
 927
 928         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 929         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 930
 931         k = sib == btree_prev_sib
 932                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 933                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 934                    bch2_btree_node_iter_peek(&node_iter, parent));
 935         if (!k)
 936                 goto out;
 937
 938         bch2_bkey_unpack(parent, &tmp.k, k);
 939
 940         ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 941                                   SIX_LOCK_intent, _THIS_IP_);
 942
 943         if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
 944                 struct btree_iter *linked;
 945
 946                 if (!bch2_btree_node_relock(iter, level + 1))
 947                         goto out;
 948
 949                 /*
 950                  * We might have got -EINTR because trylock failed, and we're
 951                  * holding other locks that would cause us to deadlock:
 952                  */
 953                 trans_for_each_iter(trans, linked)
 954                         if (btree_iter_lock_cmp(iter, linked) < 0)
 955                                 __bch2_btree_iter_unlock(linked);
 956
 957                 if (sib == btree_prev_sib)
 958                         btree_node_unlock(iter, level);
 959
 960                 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 961                                           SIX_LOCK_intent, _THIS_IP_);
 962
 963                 /*
 964                  * before btree_iter_relock() calls btree_iter_verify_locks():
 965                  */
 966                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 967                         btree_node_unlock(iter, level + 1);
 968
 969                 if (!bch2_btree_node_relock(iter, level)) {
 970                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 971
 972                         if (!IS_ERR(ret)) {
 973                                 six_unlock_intent(&ret->c.lock);
 974                                 ret = ERR_PTR(-EINTR);
 975                         }
 976                 }
 977
 978                 bch2_trans_relock(trans);
 979         }
 980 out:
 981         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 982                 btree_node_unlock(iter, level + 1);
 983
 984         if (PTR_ERR_OR_ZERO(ret) == -EINTR)
 985                 bch2_btree_iter_upgrade(iter, level + 2);
 986
 987         BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
 988
 989         if (!IS_ERR_OR_NULL(ret)) {
 990                 struct btree *n1 = ret, *n2 = b;
 991
 992                 if (sib != btree_prev_sib)
 993                         swap(n1, n2);
 994
 995                 BUG_ON(bkey_cmp(bkey_successor(n1->key.k.p),
 996                                 n2->data->min_key));
 997         }
 998
 999         bch2_btree_trans_verify_locks(trans);
1000
1001         return ret;
1002 }
1003
1004 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
1005                               const struct bkey_i *k, unsigned level)
1006 {
1007         struct btree_cache *bc = &c->btree_cache;
1008         struct btree *b;
1009
1010         BUG_ON(!btree_node_locked(iter, level + 1));
1011         BUG_ON(level >= BTREE_MAX_DEPTH);
1012
1013         b = btree_cache_find(bc, k);
1014         if (b)
1015                 return;
1016
1017         bch2_btree_node_fill(c, iter, k, iter->btree_id,
1018                              level, SIX_LOCK_read, false);
1019 }
1020
1021 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1022                              struct btree *b)
1023 {
1024         const struct bkey_format *f = &b->format;
1025         struct bset_stats stats;
1026
1027         memset(&stats, 0, sizeof(stats));
1028
1029         bch2_btree_keys_stats(b, &stats);
1030
1031         pr_buf(out,
1032                "l %u %llu:%llu - %llu:%llu:\n"
1033                "    ptrs: ",
1034                b->c.level,
1035                b->data->min_key.inode,
1036                b->data->min_key.offset,
1037                b->data->max_key.inode,
1038                b->data->max_key.offset);
1039         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1040         pr_buf(out, "\n"
1041                "    format: u64s %u fields %u %u %u %u %u\n"
1042                "    unpack fn len: %u\n"
1043                "    bytes used %zu/%zu (%zu%% full)\n"
1044                "    sib u64s: %u, %u (merge threshold %zu)\n"
1045                "    nr packed keys %u\n"
1046                "    nr unpacked keys %u\n"
1047                "    floats %zu\n"
1048                "    failed unpacked %zu\n",
1049                f->key_u64s,
1050                f->bits_per_field[0],
1051                f->bits_per_field[1],
1052                f->bits_per_field[2],
1053                f->bits_per_field[3],
1054                f->bits_per_field[4],
1055                b->unpack_fn_len,
1056                b->nr.live_u64s * sizeof(u64),
1057                btree_bytes(c) - sizeof(struct btree_node),
1058                b->nr.live_u64s * 100 / btree_max_u64s(c),
1059                b->sib_u64s[0],
1060                b->sib_u64s[1],
1061                BTREE_FOREGROUND_MERGE_THRESHOLD(c),
1062                b->nr.packed_keys,
1063                b->nr.unpacked_keys,
1064                stats.floats,
1065                stats.failed);
1066 }