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 (verify_btree_ondisk(c))
 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 (btree_shrinker_disabled(c))
 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         memalloc_nofs_restore(flags);
 332         mutex_unlock(&bc->lock);
 333 out:
 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 (btree_shrinker_disabled(c))
 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(b);
 385
 386                 btree_node_data_free(c, b);
 387         }
 388
 389         while (!list_empty(&bc->freed)) {
 390                 b = list_first_entry(&bc->freed, struct btree, list);
 391                 list_del(&b->list);
 392                 kfree(b);
 393         }
 394
 395         mutex_unlock(&bc->lock);
 396         memalloc_nofs_restore(flags);
 397
 398         if (bc->table_init_done)
 399                 rhashtable_destroy(&bc->table);
 400 }
 401
 402 int bch2_fs_btree_cache_init(struct bch_fs *c)
 403 {
 404         struct btree_cache *bc = &c->btree_cache;
 405         unsigned i;
 406         int ret = 0;
 407
 408         pr_verbose_init(c->opts, "");
 409
 410         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 411         if (ret)
 412                 goto out;
 413
 414         bc->table_init_done = true;
 415
 416         bch2_recalc_btree_reserve(c);
 417
 418         for (i = 0; i < bc->reserve; i++)
 419                 if (!btree_node_mem_alloc(c)) {
 420                         ret = -ENOMEM;
 421                         goto out;
 422                 }
 423
 424         list_splice_init(&bc->live, &bc->freeable);
 425
 426 #ifdef CONFIG_BCACHEFS_DEBUG
 427         mutex_init(&c->verify_lock);
 428
 429         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 430         if (!c->verify_ondisk) {
 431                 ret = -ENOMEM;
 432                 goto out;
 433         }
 434
 435         c->verify_data = btree_node_mem_alloc(c);
 436         if (!c->verify_data) {
 437                 ret = -ENOMEM;
 438                 goto out;
 439         }
 440
 441         list_del_init(&c->verify_data->list);
 442 #endif
 443
 444         bc->shrink.count_objects        = bch2_btree_cache_count;
 445         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 446         bc->shrink.seeks                = 4;
 447         bc->shrink.batch                = btree_pages(c) * 2;
 448         register_shrinker(&bc->shrink);
 449 out:
 450         pr_verbose_init(c->opts, "ret %i", ret);
 451         return ret;
 452 }
 453
 454 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 455 {
 456         mutex_init(&bc->lock);
 457         INIT_LIST_HEAD(&bc->live);
 458         INIT_LIST_HEAD(&bc->freeable);
 459         INIT_LIST_HEAD(&bc->freed);
 460 }
 461
 462 /*
 463  * We can only have one thread cannibalizing other cached btree nodes at a time,
 464  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 465  * cannibalize_bucket() will take. This means every time we unlock the root of
 466  * the btree, we need to release this lock if we have it held.
 467  */
 468 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 469 {
 470         struct btree_cache *bc = &c->btree_cache;
 471
 472         if (bc->alloc_lock == current) {
 473                 trace_btree_node_cannibalize_unlock(c);
 474                 bc->alloc_lock = NULL;
 475                 closure_wake_up(&bc->alloc_wait);
 476         }
 477 }
 478
 479 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 480 {
 481         struct btree_cache *bc = &c->btree_cache;
 482         struct task_struct *old;
 483
 484         old = cmpxchg(&bc->alloc_lock, NULL, current);
 485         if (old == NULL || old == current)
 486                 goto success;
 487
 488         if (!cl) {
 489                 trace_btree_node_cannibalize_lock_fail(c);
 490                 return -ENOMEM;
 491         }
 492
 493         closure_wait(&bc->alloc_wait, cl);
 494
 495         /* Try again, after adding ourselves to waitlist */
 496         old = cmpxchg(&bc->alloc_lock, NULL, current);
 497         if (old == NULL || old == current) {
 498                 /* We raced */
 499                 closure_wake_up(&bc->alloc_wait);
 500                 goto success;
 501         }
 502
 503         trace_btree_node_cannibalize_lock_fail(c);
 504         return -EAGAIN;
 505
 506 success:
 507         trace_btree_node_cannibalize_lock(c);
 508         return 0;
 509 }
 510
 511 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 512 {
 513         struct btree_cache *bc = &c->btree_cache;
 514         struct btree *b;
 515
 516         list_for_each_entry_reverse(b, &bc->live, list)
 517                 if (!btree_node_reclaim(c, b))
 518                         return b;
 519
 520         while (1) {
 521                 list_for_each_entry_reverse(b, &bc->live, list)
 522                         if (!btree_node_write_and_reclaim(c, b))
 523                                 return b;
 524
 525                 /*
 526                  * Rare case: all nodes were intent-locked.
 527                  * Just busy-wait.
 528                  */
 529                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 530                 cond_resched();
 531         }
 532 }
 533
 534 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 535 {
 536         struct btree_cache *bc = &c->btree_cache;
 537         struct btree *b;
 538         u64 start_time = local_clock();
 539         unsigned flags;
 540
 541         flags = memalloc_nofs_save();
 542         mutex_lock(&bc->lock);
 543
 544         /*
 545          * btree_free() doesn't free memory; it sticks the node on the end of
 546          * the list. Check if there's any freed nodes there:
 547          */
 548         list_for_each_entry(b, &bc->freeable, list)
 549                 if (!btree_node_reclaim(c, b))
 550                         goto got_node;
 551
 552         /*
 553          * We never free struct btree itself, just the memory that holds the on
 554          * disk node. Check the freed list before allocating a new one:
 555          */
 556         list_for_each_entry(b, &bc->freed, list)
 557                 if (!btree_node_reclaim(c, b))
 558                         goto got_node;
 559
 560         b = NULL;
 561 got_node:
 562         if (b)
 563                 list_del_init(&b->list);
 564         mutex_unlock(&bc->lock);
 565
 566         if (!b) {
 567                 b = __btree_node_mem_alloc(c);
 568                 if (!b)
 569                         goto err;
 570
 571                 BUG_ON(!six_trylock_intent(&b->c.lock));
 572                 BUG_ON(!six_trylock_write(&b->c.lock));
 573         }
 574
 575         if (!b->data) {
 576                 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 577                         goto err;
 578
 579                 mutex_lock(&bc->lock);
 580                 bc->used++;
 581                 mutex_unlock(&bc->lock);
 582         }
 583
 584         BUG_ON(btree_node_hashed(b));
 585         BUG_ON(btree_node_write_in_flight(b));
 586 out:
 587         b->flags                = 0;
 588         b->written              = 0;
 589         b->nsets                = 0;
 590         b->sib_u64s[0]          = 0;
 591         b->sib_u64s[1]          = 0;
 592         b->whiteout_u64s        = 0;
 593         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 594
 595         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 596                                start_time);
 597
 598         memalloc_nofs_restore(flags);
 599         return b;
 600 err:
 601         mutex_lock(&bc->lock);
 602
 603         if (b) {
 604                 list_add(&b->list, &bc->freed);
 605                 six_unlock_write(&b->c.lock);
 606                 six_unlock_intent(&b->c.lock);
 607         }
 608
 609         /* Try to cannibalize another cached btree node: */
 610         if (bc->alloc_lock == current) {
 611                 b = btree_node_cannibalize(c);
 612                 list_del_init(&b->list);
 613                 mutex_unlock(&bc->lock);
 614
 615                 bch2_btree_node_hash_remove(bc, b);
 616
 617                 trace_btree_node_cannibalize(c);
 618                 goto out;
 619         }
 620
 621         mutex_unlock(&bc->lock);
 622         memalloc_nofs_restore(flags);
 623         return ERR_PTR(-ENOMEM);
 624 }
 625
 626 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 627 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 628                                 struct btree_iter *iter,
 629                                 const struct bkey_i *k,
 630                                 enum btree_id btree_id,
 631                                 unsigned level,
 632                                 enum six_lock_type lock_type,
 633                                 bool sync)
 634 {
 635         struct btree_cache *bc = &c->btree_cache;
 636         struct btree *b;
 637
 638         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 639         /*
 640          * Parent node must be locked, else we could read in a btree node that's
 641          * been freed:
 642          */
 643         if (iter && !bch2_btree_node_relock(iter, level + 1))
 644                 return ERR_PTR(-EINTR);
 645
 646         b = bch2_btree_node_mem_alloc(c);
 647         if (IS_ERR(b))
 648                 return b;
 649
 650         bkey_copy(&b->key, k);
 651         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 652                 /* raced with another fill: */
 653
 654                 /* mark as unhashed... */
 655                 b->hash_val = 0;
 656
 657                 mutex_lock(&bc->lock);
 658                 list_add(&b->list, &bc->freeable);
 659                 mutex_unlock(&bc->lock);
 660
 661                 six_unlock_write(&b->c.lock);
 662                 six_unlock_intent(&b->c.lock);
 663                 return NULL;
 664         }
 665
 666         /*
 667          * Unlock before doing IO:
 668          *
 669          * XXX: ideally should be dropping all btree node locks here
 670          */
 671         if (iter && btree_node_read_locked(iter, level + 1))
 672                 btree_node_unlock(iter, level + 1);
 673
 674         bch2_btree_node_read(c, b, sync);
 675
 676         six_unlock_write(&b->c.lock);
 677
 678         if (!sync) {
 679                 six_unlock_intent(&b->c.lock);
 680                 return NULL;
 681         }
 682
 683         if (lock_type == SIX_LOCK_read)
 684                 six_lock_downgrade(&b->c.lock);
 685
 686         return b;
 687 }
 688
 689 static int lock_node_check_fn(struct six_lock *lock, void *p)
 690 {
 691         struct btree *b = container_of(lock, struct btree, c.lock);
 692         const struct bkey_i *k = p;
 693
 694         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 695 }
 696
 697 /**
 698  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 699  * in from disk if necessary.
 700  *
 701  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 702  *
 703  * The btree node will have either a read or a write lock held, depending on
 704  * the @write parameter.
 705  */
 706 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 707                                   const struct bkey_i *k, unsigned level,
 708                                   enum six_lock_type lock_type)
 709 {
 710         struct btree_cache *bc = &c->btree_cache;
 711         struct btree *b;
 712         struct bset_tree *t;
 713
 714         EBUG_ON(level >= BTREE_MAX_DEPTH);
 715
 716         b = btree_node_mem_ptr(k);
 717         if (b)
 718                 goto lock_node;
 719 retry:
 720         b = btree_cache_find(bc, k);
 721         if (unlikely(!b)) {
 722                 /*
 723                  * We must have the parent locked to call bch2_btree_node_fill(),
 724                  * else we could read in a btree node from disk that's been
 725                  * freed:
 726                  */
 727                 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
 728                                          level, lock_type, true);
 729
 730                 /* We raced and found the btree node in the cache */
 731                 if (!b)
 732                         goto retry;
 733
 734                 if (IS_ERR(b))
 735                         return b;
 736         } else {
 737 lock_node:
 738                 /*
 739                  * There's a potential deadlock with splits and insertions into
 740                  * interior nodes we have to avoid:
 741                  *
 742                  * The other thread might be holding an intent lock on the node
 743                  * we want, and they want to update its parent node so they're
 744                  * going to upgrade their intent lock on the parent node to a
 745                  * write lock.
 746                  *
 747                  * But if we're holding a read lock on the parent, and we're
 748                  * trying to get the intent lock they're holding, we deadlock.
 749                  *
 750                  * So to avoid this we drop the read locks on parent nodes when
 751                  * we're starting to take intent locks - and handle the race.
 752                  *
 753                  * The race is that they might be about to free the node we
 754                  * want, and dropping our read lock on the parent node lets them
 755                  * update the parent marking the node we want as freed, and then
 756                  * free it:
 757                  *
 758                  * To guard against this, btree nodes are evicted from the cache
 759                  * when they're freed - and b->hash_val is zeroed out, which we
 760                  * check for after we lock the node.
 761                  *
 762                  * Then, bch2_btree_node_relock() on the parent will fail - because
 763                  * the parent was modified, when the pointer to the node we want
 764                  * was removed - and we'll bail out:
 765                  */
 766                 if (btree_node_read_locked(iter, level + 1))
 767                         btree_node_unlock(iter, level + 1);
 768
 769                 if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
 770                                      lock_node_check_fn, (void *) k)) {
 771                         if (b->hash_val != btree_ptr_hash_val(k))
 772                                 goto retry;
 773                         return ERR_PTR(-EINTR);
 774                 }
 775
 776                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 777                              b->c.level != level ||
 778                              race_fault())) {
 779                         six_unlock_type(&b->c.lock, lock_type);
 780                         if (bch2_btree_node_relock(iter, level + 1))
 781                                 goto retry;
 782
 783                         trace_trans_restart_btree_node_reused(iter->trans->ip);
 784                         return ERR_PTR(-EINTR);
 785                 }
 786         }
 787
 788         /* XXX: waiting on IO with btree locks held: */
 789         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 790                        TASK_UNINTERRUPTIBLE);
 791
 792         prefetch(b->aux_data);
 793
 794         for_each_bset(b, t) {
 795                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 796
 797                 prefetch(p + L1_CACHE_BYTES * 0);
 798                 prefetch(p + L1_CACHE_BYTES * 1);
 799                 prefetch(p + L1_CACHE_BYTES * 2);
 800         }
 801
 802         /* avoid atomic set bit if it's not needed: */
 803         if (!btree_node_accessed(b))
 804                 set_btree_node_accessed(b);
 805
 806         if (unlikely(btree_node_read_error(b))) {
 807                 six_unlock_type(&b->c.lock, lock_type);
 808                 return ERR_PTR(-EIO);
 809         }
 810
 811         EBUG_ON(b->c.btree_id != iter->btree_id ||
 812                 BTREE_NODE_LEVEL(b->data) != level ||
 813                 bkey_cmp(b->data->max_key, k->k.p));
 814
 815         return b;
 816 }
 817
 818 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 819                                          const struct bkey_i *k,
 820                                          enum btree_id btree_id,
 821                                          unsigned level)
 822 {
 823         struct btree_cache *bc = &c->btree_cache;
 824         struct btree *b;
 825         struct bset_tree *t;
 826         int ret;
 827
 828         EBUG_ON(level >= BTREE_MAX_DEPTH);
 829
 830         b = btree_node_mem_ptr(k);
 831         if (b)
 832                 goto lock_node;
 833 retry:
 834         b = btree_cache_find(bc, k);
 835         if (unlikely(!b)) {
 836                 b = bch2_btree_node_fill(c, NULL, k, btree_id,
 837                                          level, SIX_LOCK_read, true);
 838
 839                 /* We raced and found the btree node in the cache */
 840                 if (!b)
 841                         goto retry;
 842
 843                 if (IS_ERR(b))
 844                         return b;
 845         } else {
 846 lock_node:
 847                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 848                 if (ret)
 849                         goto retry;
 850
 851                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 852                              b->c.btree_id != btree_id ||
 853                              b->c.level != level)) {
 854                         six_unlock_read(&b->c.lock);
 855                         goto retry;
 856                 }
 857         }
 858
 859         /* XXX: waiting on IO with btree locks held: */
 860         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 861                        TASK_UNINTERRUPTIBLE);
 862
 863         prefetch(b->aux_data);
 864
 865         for_each_bset(b, t) {
 866                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 867
 868                 prefetch(p + L1_CACHE_BYTES * 0);
 869                 prefetch(p + L1_CACHE_BYTES * 1);
 870                 prefetch(p + L1_CACHE_BYTES * 2);
 871         }
 872
 873         /* avoid atomic set bit if it's not needed: */
 874         if (!btree_node_accessed(b))
 875                 set_btree_node_accessed(b);
 876
 877         if (unlikely(btree_node_read_error(b))) {
 878                 six_unlock_read(&b->c.lock);
 879                 return ERR_PTR(-EIO);
 880         }
 881
 882         EBUG_ON(b->c.btree_id != btree_id ||
 883                 BTREE_NODE_LEVEL(b->data) != level ||
 884                 bkey_cmp(b->data->max_key, k->k.p));
 885
 886         return b;
 887 }
 888
 889 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 890                                           struct btree_iter *iter,
 891                                           struct btree *b,
 892                                           enum btree_node_sibling sib)
 893 {
 894         struct btree_trans *trans = iter->trans;
 895         struct btree *parent;
 896         struct btree_node_iter node_iter;
 897         struct bkey_packed *k;
 898         BKEY_PADDED(k) tmp;
 899         struct btree *ret = NULL;
 900         unsigned level = b->c.level;
 901
 902         parent = btree_iter_node(iter, level + 1);
 903         if (!parent)
 904                 return NULL;
 905
 906         /*
 907          * There's a corner case where a btree_iter might have a node locked
 908          * that is just outside its current pos - when
 909          * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
 910          *
 911          * But the lock ordering checks in __bch2_btree_node_lock() go off of
 912          * iter->pos, not the node's key: so if the iterator is marked as
 913          * needing to be traversed, we risk deadlock if we don't bail out here:
 914          */
 915         if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
 916                 return ERR_PTR(-EINTR);
 917
 918         if (!bch2_btree_node_relock(iter, level + 1)) {
 919                 ret = ERR_PTR(-EINTR);
 920                 goto out;
 921         }
 922
 923         node_iter = iter->l[parent->c.level].iter;
 924
 925         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 926         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 927
 928         k = sib == btree_prev_sib
 929                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 930                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 931                    bch2_btree_node_iter_peek(&node_iter, parent));
 932         if (!k)
 933                 goto out;
 934
 935         bch2_bkey_unpack(parent, &tmp.k, k);
 936
 937         ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 938                                   SIX_LOCK_intent);
 939
 940         if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
 941                 struct btree_iter *linked;
 942
 943                 if (!bch2_btree_node_relock(iter, level + 1))
 944                         goto out;
 945
 946                 /*
 947                  * We might have got -EINTR because trylock failed, and we're
 948                  * holding other locks that would cause us to deadlock:
 949                  */
 950                 trans_for_each_iter(trans, linked)
 951                         if (btree_iter_cmp(iter, linked) < 0)
 952                                 __bch2_btree_iter_unlock(linked);
 953
 954                 if (sib == btree_prev_sib)
 955                         btree_node_unlock(iter, level);
 956
 957                 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 958                                           SIX_LOCK_intent);
 959
 960                 /*
 961                  * before btree_iter_relock() calls btree_iter_verify_locks():
 962                  */
 963                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 964                         btree_node_unlock(iter, level + 1);
 965
 966                 if (!bch2_btree_node_relock(iter, level)) {
 967                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 968
 969                         if (!IS_ERR(ret)) {
 970                                 six_unlock_intent(&ret->c.lock);
 971                                 ret = ERR_PTR(-EINTR);
 972                         }
 973                 }
 974
 975                 bch2_trans_relock(trans);
 976         }
 977 out:
 978         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 979                 btree_node_unlock(iter, level + 1);
 980
 981         if (PTR_ERR_OR_ZERO(ret) == -EINTR)
 982                 bch2_btree_iter_upgrade(iter, level + 2);
 983
 984         BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
 985
 986         if (!IS_ERR_OR_NULL(ret)) {
 987                 struct btree *n1 = ret, *n2 = b;
 988
 989                 if (sib != btree_prev_sib)
 990                         swap(n1, n2);
 991
 992                 BUG_ON(bkey_cmp(bkey_successor(n1->key.k.p),
 993                                 n2->data->min_key));
 994         }
 995
 996         bch2_btree_trans_verify_locks(trans);
 997
 998         return ret;
 999 }
1000
1001 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
1002                               const struct bkey_i *k, unsigned level)
1003 {
1004         struct btree_cache *bc = &c->btree_cache;
1005         struct btree *b;
1006
1007         BUG_ON(!btree_node_locked(iter, level + 1));
1008         BUG_ON(level >= BTREE_MAX_DEPTH);
1009
1010         b = btree_cache_find(bc, k);
1011         if (b)
1012                 return;
1013
1014         bch2_btree_node_fill(c, iter, k, iter->btree_id,
1015                              level, SIX_LOCK_read, false);
1016 }
1017
1018 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1019                              struct btree *b)
1020 {
1021         const struct bkey_format *f = &b->format;
1022         struct bset_stats stats;
1023
1024         memset(&stats, 0, sizeof(stats));
1025
1026         bch2_btree_keys_stats(b, &stats);
1027
1028         pr_buf(out,
1029                "l %u %llu:%llu - %llu:%llu:\n"
1030                "    ptrs: ",
1031                b->c.level,
1032                b->data->min_key.inode,
1033                b->data->min_key.offset,
1034                b->data->max_key.inode,
1035                b->data->max_key.offset);
1036         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1037         pr_buf(out, "\n"
1038                "    format: u64s %u fields %u %u %u %u %u\n"
1039                "    unpack fn len: %u\n"
1040                "    bytes used %zu/%zu (%zu%% full)\n"
1041                "    sib u64s: %u, %u (merge threshold %zu)\n"
1042                "    nr packed keys %u\n"
1043                "    nr unpacked keys %u\n"
1044                "    floats %zu\n"
1045                "    failed unpacked %zu\n",
1046                f->key_u64s,
1047                f->bits_per_field[0],
1048                f->bits_per_field[1],
1049                f->bits_per_field[2],
1050                f->bits_per_field[3],
1051                f->bits_per_field[4],
1052                b->unpack_fn_len,
1053                b->nr.live_u64s * sizeof(u64),
1054                btree_bytes(c) - sizeof(struct btree_node),
1055                b->nr.live_u64s * 100 / btree_max_u64s(c),
1056                b->sib_u64s[0],
1057                b->sib_u64s[1],
1058                BTREE_FOREGROUND_MERGE_THRESHOLD(c),
1059                b->nr.packed_keys,
1060                b->nr.unpacked_keys,
1061                stats.floats,
1062                stats.failed);
1063 }