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 struct lock_class_key bch2_btree_node_lock_key;
  17
  18 const char * const bch2_btree_node_flags[] = {
  19 #define x(f)    #f,
  20         BTREE_FLAGS()
  21 #undef x
  22         NULL
  23 };
  24
  25 void bch2_recalc_btree_reserve(struct bch_fs *c)
  26 {
  27         unsigned i, reserve = 16;
  28
  29         if (!c->btree_roots[0].b)
  30                 reserve += 8;
  31
  32         for (i = 0; i < BTREE_ID_NR; i++)
  33                 if (c->btree_roots[i].b)
  34                         reserve += min_t(unsigned, 1,
  35                                          c->btree_roots[i].b->c.level) * 8;
  36
  37         c->btree_cache.reserve = reserve;
  38 }
  39
  40 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  41 {
  42         return max_t(int, 0, bc->used - bc->reserve);
  43 }
  44
  45 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  46 {
  47         struct btree_cache *bc = &c->btree_cache;
  48
  49         EBUG_ON(btree_node_write_in_flight(b));
  50
  51         kvpfree(b->data, btree_bytes(c));
  52         b->data = NULL;
  53 #ifdef __KERNEL__
  54         vfree(b->aux_data);
  55 #else
  56         munmap(b->aux_data, btree_aux_data_bytes(b));
  57 #endif
  58         b->aux_data = NULL;
  59
  60         bc->used--;
  61         list_move(&b->list, &bc->freed);
  62 }
  63
  64 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  65                                    const void *obj)
  66 {
  67         const struct btree *b = obj;
  68         const u64 *v = arg->key;
  69
  70         return b->hash_val == *v ? 0 : 1;
  71 }
  72
  73 static const struct rhashtable_params bch_btree_cache_params = {
  74         .head_offset    = offsetof(struct btree, hash),
  75         .key_offset     = offsetof(struct btree, hash_val),
  76         .key_len        = sizeof(u64),
  77         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  78 };
  79
  80 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  81 {
  82         BUG_ON(b->data || b->aux_data);
  83
  84         b->data = kvpmalloc(btree_bytes(c), gfp);
  85         if (!b->data)
  86                 return -ENOMEM;
  87 #ifdef __KERNEL__
  88         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  89 #else
  90         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
  91                            PROT_READ|PROT_WRITE|PROT_EXEC,
  92                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
  93         if (b->aux_data == MAP_FAILED)
  94                 b->aux_data = NULL;
  95 #endif
  96         if (!b->aux_data) {
  97                 kvpfree(b->data, btree_bytes(c));
  98                 b->data = NULL;
  99                 return -ENOMEM;
 100         }
 101
 102         return 0;
 103 }
 104
 105 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
 106 {
 107         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
 108         if (!b)
 109                 return NULL;
 110
 111         bkey_btree_ptr_init(&b->key);
 112         __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 113         INIT_LIST_HEAD(&b->list);
 114         INIT_LIST_HEAD(&b->write_blocked);
 115         b->byte_order = ilog2(btree_bytes(c));
 116         return b;
 117 }
 118
 119 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 120 {
 121         struct btree_cache *bc = &c->btree_cache;
 122         struct btree *b = __btree_node_mem_alloc(c);
 123         if (!b)
 124                 return NULL;
 125
 126         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 127                 kfree(b);
 128                 return NULL;
 129         }
 130
 131         bc->used++;
 132         list_add(&b->list, &bc->freeable);
 133         return b;
 134 }
 135
 136 /* Btree in memory cache - hash table */
 137
 138 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 139 {
 140         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 141         BUG_ON(ret);
 142
 143         /* Cause future lookups for this node to fail: */
 144         b->hash_val = 0;
 145
 146         six_lock_wakeup_all(&b->c.lock);
 147 }
 148
 149 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 150 {
 151         BUG_ON(b->hash_val);
 152         b->hash_val = btree_ptr_hash_val(&b->key);
 153
 154         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 155                                              bch_btree_cache_params);
 156 }
 157
 158 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 159                                 unsigned level, enum btree_id id)
 160 {
 161         int ret;
 162
 163         b->c.level      = level;
 164         b->c.btree_id   = id;
 165
 166         if (level)
 167                 six_lock_pcpu_alloc(&b->c.lock);
 168         else
 169                 six_lock_pcpu_free_rcu(&b->c.lock);
 170
 171         mutex_lock(&bc->lock);
 172         ret = __bch2_btree_node_hash_insert(bc, b);
 173         if (!ret)
 174                 list_add(&b->list, &bc->live);
 175         mutex_unlock(&bc->lock);
 176
 177         return ret;
 178 }
 179
 180 __flatten
 181 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 182                                      const struct bkey_i *k)
 183 {
 184         u64 v = btree_ptr_hash_val(k);
 185
 186         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 187 }
 188
 189 /*
 190  * this version is for btree nodes that have already been freed (we're not
 191  * reaping a real btree node)
 192  */
 193 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 194 {
 195         struct btree_cache *bc = &c->btree_cache;
 196         int ret = 0;
 197
 198         lockdep_assert_held(&bc->lock);
 199 wait_on_io:
 200         if (b->flags & ((1U << BTREE_NODE_dirty)|
 201                         (1U << BTREE_NODE_read_in_flight)|
 202                         (1U << BTREE_NODE_write_in_flight))) {
 203                 if (!flush)
 204                         return -ENOMEM;
 205
 206                 /* XXX: waiting on IO with btree cache lock held */
 207                 bch2_btree_node_wait_on_read(b);
 208                 bch2_btree_node_wait_on_write(b);
 209         }
 210
 211         if (!six_trylock_intent(&b->c.lock))
 212                 return -ENOMEM;
 213
 214         if (!six_trylock_write(&b->c.lock))
 215                 goto out_unlock_intent;
 216
 217         /* recheck under lock */
 218         if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 219                         (1U << BTREE_NODE_write_in_flight))) {
 220                 if (!flush)
 221                         goto out_unlock;
 222                 six_unlock_write(&b->c.lock);
 223                 six_unlock_intent(&b->c.lock);
 224                 goto wait_on_io;
 225         }
 226
 227         if (btree_node_noevict(b) ||
 228             btree_node_write_blocked(b) ||
 229             btree_node_will_make_reachable(b))
 230                 goto out_unlock;
 231
 232         if (btree_node_dirty(b)) {
 233                 if (!flush)
 234                         goto out_unlock;
 235                 /*
 236                  * Using the underscore version because we don't want to compact
 237                  * bsets after the write, since this node is about to be evicted
 238                  * - unless btree verify mode is enabled, since it runs out of
 239                  * the post write cleanup:
 240                  */
 241                 if (bch2_verify_btree_ondisk)
 242                         bch2_btree_node_write(c, b, SIX_LOCK_intent, 0);
 243                 else
 244                         __bch2_btree_node_write(c, b, 0);
 245
 246                 six_unlock_write(&b->c.lock);
 247                 six_unlock_intent(&b->c.lock);
 248                 goto wait_on_io;
 249         }
 250 out:
 251         if (b->hash_val && !ret)
 252                 trace_btree_node_reap(c, b);
 253         return ret;
 254 out_unlock:
 255         six_unlock_write(&b->c.lock);
 256 out_unlock_intent:
 257         six_unlock_intent(&b->c.lock);
 258         ret = -ENOMEM;
 259         goto out;
 260 }
 261
 262 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 263 {
 264         return __btree_node_reclaim(c, b, false);
 265 }
 266
 267 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 268 {
 269         return __btree_node_reclaim(c, b, true);
 270 }
 271
 272 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 273                                            struct shrink_control *sc)
 274 {
 275         struct bch_fs *c = container_of(shrink, struct bch_fs,
 276                                         btree_cache.shrink);
 277         struct btree_cache *bc = &c->btree_cache;
 278         struct btree *b, *t;
 279         unsigned long nr = sc->nr_to_scan;
 280         unsigned long can_free;
 281         unsigned long touched = 0;
 282         unsigned long freed = 0;
 283         unsigned i, flags;
 284         unsigned long ret = SHRINK_STOP;
 285
 286         if (bch2_btree_shrinker_disabled)
 287                 return SHRINK_STOP;
 288
 289         /* Return -1 if we can't do anything right now */
 290         if (sc->gfp_mask & __GFP_FS)
 291                 mutex_lock(&bc->lock);
 292         else if (!mutex_trylock(&bc->lock))
 293                 goto out_norestore;
 294
 295         flags = memalloc_nofs_save();
 296
 297         /*
 298          * It's _really_ critical that we don't free too many btree nodes - we
 299          * have to always leave ourselves a reserve. The reserve is how we
 300          * guarantee that allocating memory for a new btree node can always
 301          * succeed, so that inserting keys into the btree can always succeed and
 302          * IO can always make forward progress:
 303          */
 304         nr /= btree_pages(c);
 305         can_free = btree_cache_can_free(bc);
 306         nr = min_t(unsigned long, nr, can_free);
 307
 308         i = 0;
 309         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 310                 /*
 311                  * Leave a few nodes on the freeable list, so that a btree split
 312                  * won't have to hit the system allocator:
 313                  */
 314                 if (++i <= 3)
 315                         continue;
 316
 317                 touched++;
 318
 319                 if (touched >= nr)
 320                         break;
 321
 322                 if (!btree_node_reclaim(c, b)) {
 323                         btree_node_data_free(c, b);
 324                         six_unlock_write(&b->c.lock);
 325                         six_unlock_intent(&b->c.lock);
 326                         freed++;
 327                 }
 328         }
 329 restart:
 330         list_for_each_entry_safe(b, t, &bc->live, list) {
 331                 touched++;
 332
 333                 if (touched >= nr) {
 334                         /* Save position */
 335                         if (&t->list != &bc->live)
 336                                 list_move_tail(&bc->live, &t->list);
 337                         break;
 338                 }
 339
 340                 if (!btree_node_accessed(b) &&
 341                     !btree_node_reclaim(c, b)) {
 342                         /* can't call bch2_btree_node_hash_remove under lock  */
 343                         freed++;
 344                         if (&t->list != &bc->live)
 345                                 list_move_tail(&bc->live, &t->list);
 346
 347                         btree_node_data_free(c, b);
 348                         mutex_unlock(&bc->lock);
 349
 350                         bch2_btree_node_hash_remove(bc, b);
 351                         six_unlock_write(&b->c.lock);
 352                         six_unlock_intent(&b->c.lock);
 353
 354                         if (freed >= nr)
 355                                 goto out;
 356
 357                         if (sc->gfp_mask & __GFP_FS)
 358                                 mutex_lock(&bc->lock);
 359                         else if (!mutex_trylock(&bc->lock))
 360                                 goto out;
 361                         goto restart;
 362                 } else
 363                         clear_btree_node_accessed(b);
 364         }
 365
 366         mutex_unlock(&bc->lock);
 367 out:
 368         ret = (unsigned long) freed * btree_pages(c);
 369         memalloc_nofs_restore(flags);
 370 out_norestore:
 371         trace_btree_cache_scan(sc->nr_to_scan,
 372                                sc->nr_to_scan / btree_pages(c),
 373                                btree_cache_can_free(bc),
 374                                ret);
 375         return ret;
 376 }
 377
 378 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 379                                             struct shrink_control *sc)
 380 {
 381         struct bch_fs *c = container_of(shrink, struct bch_fs,
 382                                         btree_cache.shrink);
 383         struct btree_cache *bc = &c->btree_cache;
 384
 385         if (bch2_btree_shrinker_disabled)
 386                 return 0;
 387
 388         return btree_cache_can_free(bc) * btree_pages(c);
 389 }
 390
 391 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 392 {
 393         struct btree_cache *bc = &c->btree_cache;
 394         struct btree *b;
 395         unsigned i, flags;
 396
 397         if (bc->shrink.list.next)
 398                 unregister_shrinker(&bc->shrink);
 399
 400         /* vfree() can allocate memory: */
 401         flags = memalloc_nofs_save();
 402         mutex_lock(&bc->lock);
 403
 404         if (c->verify_data)
 405                 list_move(&c->verify_data->list, &bc->live);
 406
 407         kvpfree(c->verify_ondisk, btree_bytes(c));
 408
 409         for (i = 0; i < BTREE_ID_NR; i++)
 410                 if (c->btree_roots[i].b)
 411                         list_add(&c->btree_roots[i].b->list, &bc->live);
 412
 413         list_splice(&bc->freeable, &bc->live);
 414
 415         while (!list_empty(&bc->live)) {
 416                 b = list_first_entry(&bc->live, struct btree, list);
 417
 418                 BUG_ON(btree_node_read_in_flight(b) ||
 419                        btree_node_write_in_flight(b));
 420
 421                 if (btree_node_dirty(b))
 422                         bch2_btree_complete_write(c, b, btree_current_write(b));
 423                 clear_btree_node_dirty_acct(c, b);
 424
 425                 btree_node_data_free(c, b);
 426         }
 427
 428         BUG_ON(atomic_read(&c->btree_cache.dirty));
 429
 430         while (!list_empty(&bc->freed)) {
 431                 b = list_first_entry(&bc->freed, struct btree, list);
 432                 list_del(&b->list);
 433                 six_lock_pcpu_free(&b->c.lock);
 434                 kfree(b);
 435         }
 436
 437         mutex_unlock(&bc->lock);
 438         memalloc_nofs_restore(flags);
 439
 440         if (bc->table_init_done)
 441                 rhashtable_destroy(&bc->table);
 442 }
 443
 444 int bch2_fs_btree_cache_init(struct bch_fs *c)
 445 {
 446         struct btree_cache *bc = &c->btree_cache;
 447         unsigned i;
 448         int ret = 0;
 449
 450         pr_verbose_init(c->opts, "");
 451
 452         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 453         if (ret)
 454                 goto out;
 455
 456         bc->table_init_done = true;
 457
 458         bch2_recalc_btree_reserve(c);
 459
 460         for (i = 0; i < bc->reserve; i++)
 461                 if (!__bch2_btree_node_mem_alloc(c)) {
 462                         ret = -ENOMEM;
 463                         goto out;
 464                 }
 465
 466         list_splice_init(&bc->live, &bc->freeable);
 467
 468         mutex_init(&c->verify_lock);
 469
 470         bc->shrink.count_objects        = bch2_btree_cache_count;
 471         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 472         bc->shrink.seeks                = 4;
 473         bc->shrink.batch                = btree_pages(c) * 2;
 474         ret = register_shrinker(&bc->shrink);
 475 out:
 476         pr_verbose_init(c->opts, "ret %i", ret);
 477         return ret;
 478 }
 479
 480 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 481 {
 482         mutex_init(&bc->lock);
 483         INIT_LIST_HEAD(&bc->live);
 484         INIT_LIST_HEAD(&bc->freeable);
 485         INIT_LIST_HEAD(&bc->freed);
 486 }
 487
 488 /*
 489  * We can only have one thread cannibalizing other cached btree nodes at a time,
 490  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 491  * cannibalize_bucket() will take. This means every time we unlock the root of
 492  * the btree, we need to release this lock if we have it held.
 493  */
 494 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 495 {
 496         struct btree_cache *bc = &c->btree_cache;
 497
 498         if (bc->alloc_lock == current) {
 499                 trace_btree_node_cannibalize_unlock(c);
 500                 bc->alloc_lock = NULL;
 501                 closure_wake_up(&bc->alloc_wait);
 502         }
 503 }
 504
 505 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 506 {
 507         struct btree_cache *bc = &c->btree_cache;
 508         struct task_struct *old;
 509
 510         old = cmpxchg(&bc->alloc_lock, NULL, current);
 511         if (old == NULL || old == current)
 512                 goto success;
 513
 514         if (!cl) {
 515                 trace_btree_node_cannibalize_lock_fail(c);
 516                 return -ENOMEM;
 517         }
 518
 519         closure_wait(&bc->alloc_wait, cl);
 520
 521         /* Try again, after adding ourselves to waitlist */
 522         old = cmpxchg(&bc->alloc_lock, NULL, current);
 523         if (old == NULL || old == current) {
 524                 /* We raced */
 525                 closure_wake_up(&bc->alloc_wait);
 526                 goto success;
 527         }
 528
 529         trace_btree_node_cannibalize_lock_fail(c);
 530         return -EAGAIN;
 531
 532 success:
 533         trace_btree_node_cannibalize_lock(c);
 534         return 0;
 535 }
 536
 537 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 538 {
 539         struct btree_cache *bc = &c->btree_cache;
 540         struct btree *b;
 541
 542         list_for_each_entry_reverse(b, &bc->live, list)
 543                 if (!btree_node_reclaim(c, b))
 544                         return b;
 545
 546         while (1) {
 547                 list_for_each_entry_reverse(b, &bc->live, list)
 548                         if (!btree_node_write_and_reclaim(c, b))
 549                                 return b;
 550
 551                 /*
 552                  * Rare case: all nodes were intent-locked.
 553                  * Just busy-wait.
 554                  */
 555                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 556                 cond_resched();
 557         }
 558 }
 559
 560 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 561 {
 562         struct btree_cache *bc = &c->btree_cache;
 563         struct btree *b;
 564         u64 start_time = local_clock();
 565         unsigned flags;
 566
 567         flags = memalloc_nofs_save();
 568         mutex_lock(&bc->lock);
 569
 570         /*
 571          * btree_free() doesn't free memory; it sticks the node on the end of
 572          * the list. Check if there's any freed nodes there:
 573          */
 574         list_for_each_entry(b, &bc->freeable, list)
 575                 if (!btree_node_reclaim(c, b))
 576                         goto got_node;
 577
 578         /*
 579          * We never free struct btree itself, just the memory that holds the on
 580          * disk node. Check the freed list before allocating a new one:
 581          */
 582         list_for_each_entry(b, &bc->freed, list)
 583                 if (!btree_node_reclaim(c, b))
 584                         goto got_node;
 585
 586         b = NULL;
 587 got_node:
 588         if (b)
 589                 list_del_init(&b->list);
 590         mutex_unlock(&bc->lock);
 591
 592         if (!b) {
 593                 b = __btree_node_mem_alloc(c);
 594                 if (!b)
 595                         goto err;
 596
 597                 BUG_ON(!six_trylock_intent(&b->c.lock));
 598                 BUG_ON(!six_trylock_write(&b->c.lock));
 599         }
 600
 601         if (!b->data) {
 602                 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 603                         goto err;
 604
 605                 mutex_lock(&bc->lock);
 606                 bc->used++;
 607                 mutex_unlock(&bc->lock);
 608         }
 609
 610         BUG_ON(btree_node_hashed(b));
 611         BUG_ON(btree_node_dirty(b));
 612         BUG_ON(btree_node_write_in_flight(b));
 613 out:
 614         b->flags                = 0;
 615         b->written              = 0;
 616         b->nsets                = 0;
 617         b->sib_u64s[0]          = 0;
 618         b->sib_u64s[1]          = 0;
 619         b->whiteout_u64s        = 0;
 620         bch2_btree_keys_init(b);
 621         set_btree_node_accessed(b);
 622
 623         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 624                                start_time);
 625
 626         memalloc_nofs_restore(flags);
 627         return b;
 628 err:
 629         mutex_lock(&bc->lock);
 630
 631         if (b) {
 632                 list_add(&b->list, &bc->freed);
 633                 six_unlock_write(&b->c.lock);
 634                 six_unlock_intent(&b->c.lock);
 635         }
 636
 637         /* Try to cannibalize another cached btree node: */
 638         if (bc->alloc_lock == current) {
 639                 b = btree_node_cannibalize(c);
 640                 list_del_init(&b->list);
 641                 mutex_unlock(&bc->lock);
 642
 643                 bch2_btree_node_hash_remove(bc, b);
 644
 645                 trace_btree_node_cannibalize(c);
 646                 goto out;
 647         }
 648
 649         mutex_unlock(&bc->lock);
 650         memalloc_nofs_restore(flags);
 651         return ERR_PTR(-ENOMEM);
 652 }
 653
 654 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 655 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 656                                 struct btree_trans *trans,
 657                                 struct btree_path *path,
 658                                 const struct bkey_i *k,
 659                                 enum btree_id btree_id,
 660                                 unsigned level,
 661                                 enum six_lock_type lock_type,
 662                                 bool sync)
 663 {
 664         struct btree_cache *bc = &c->btree_cache;
 665         struct btree *b;
 666         u32 seq;
 667
 668         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 669         /*
 670          * Parent node must be locked, else we could read in a btree node that's
 671          * been freed:
 672          */
 673         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 674                 trace_trans_restart_relock_parent_for_fill(trans->fn,
 675                                         _THIS_IP_, btree_id, &path->pos);
 676                 btree_trans_restart(trans);
 677                 return ERR_PTR(-EINTR);
 678         }
 679
 680         b = bch2_btree_node_mem_alloc(c);
 681
 682         if (trans && b == ERR_PTR(-ENOMEM)) {
 683                 trans->memory_allocation_failure = true;
 684                 trace_trans_restart_memory_allocation_failure(trans->fn,
 685                                 _THIS_IP_, btree_id, &path->pos);
 686                 btree_trans_restart(trans);
 687                 return ERR_PTR(-EINTR);
 688         }
 689
 690         if (IS_ERR(b))
 691                 return b;
 692
 693         bkey_copy(&b->key, k);
 694         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 695                 /* raced with another fill: */
 696
 697                 /* mark as unhashed... */
 698                 b->hash_val = 0;
 699
 700                 mutex_lock(&bc->lock);
 701                 list_add(&b->list, &bc->freeable);
 702                 mutex_unlock(&bc->lock);
 703
 704                 six_unlock_write(&b->c.lock);
 705                 six_unlock_intent(&b->c.lock);
 706                 return NULL;
 707         }
 708
 709         set_btree_node_read_in_flight(b);
 710
 711         six_unlock_write(&b->c.lock);
 712         seq = b->c.lock.state.seq;
 713         six_unlock_intent(&b->c.lock);
 714
 715         /* Unlock before doing IO: */
 716         if (trans && sync)
 717                 bch2_trans_unlock(trans);
 718
 719         bch2_btree_node_read(c, b, sync);
 720
 721         if (!sync)
 722                 return NULL;
 723
 724         if (trans &&
 725             (!bch2_trans_relock(trans) ||
 726              !bch2_btree_path_relock_intent(trans, path))) {
 727                 BUG_ON(!trans->restarted);
 728                 return ERR_PTR(-EINTR);
 729         }
 730
 731         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 732                 trace_trans_restart_relock_after_fill(trans->fn, _THIS_IP_,
 733                                            btree_id, &path->pos);
 734                 btree_trans_restart(trans);
 735                 return ERR_PTR(-EINTR);
 736         }
 737
 738         return b;
 739 }
 740
 741 static int lock_node_check_fn(struct six_lock *lock, void *p)
 742 {
 743         struct btree *b = container_of(lock, struct btree, c.lock);
 744         const struct bkey_i *k = p;
 745
 746         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 747 }
 748
 749 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 750 {
 751         struct printbuf buf1 = PRINTBUF;
 752         struct printbuf buf2 = PRINTBUF;
 753         struct printbuf buf3 = PRINTBUF;
 754
 755         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 756                 return;
 757
 758         bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&b->key));
 759         bch2_bpos_to_text(&buf2, b->data->min_key);
 760         bch2_bpos_to_text(&buf3, b->data->max_key);
 761
 762         bch2_fs_inconsistent(c, "btree node header doesn't match ptr\n"
 763                              "btree %s level %u\n"
 764                              "ptr: %s\n"
 765                              "header: btree %s level %llu\n"
 766                              "min %s max %s\n",
 767                              bch2_btree_ids[b->c.btree_id], b->c.level,
 768                              buf1.buf,
 769                              bch2_btree_ids[BTREE_NODE_ID(b->data)],
 770                              BTREE_NODE_LEVEL(b->data),
 771                              buf2.buf, buf3.buf);
 772
 773         printbuf_exit(&buf3);
 774         printbuf_exit(&buf2);
 775         printbuf_exit(&buf1);
 776 }
 777
 778 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 779 {
 780         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 781             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 782             bpos_cmp(b->data->max_key, b->key.k.p) ||
 783             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 784              bpos_cmp(b->data->min_key,
 785                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 786                 btree_bad_header(c, b);
 787 }
 788
 789 /**
 790  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 791  * in from disk if necessary.
 792  *
 793  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 794  *
 795  * The btree node will have either a read or a write lock held, depending on
 796  * the @write parameter.
 797  */
 798 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 799                                   const struct bkey_i *k, unsigned level,
 800                                   enum six_lock_type lock_type,
 801                                   unsigned long trace_ip)
 802 {
 803         struct bch_fs *c = trans->c;
 804         struct btree_cache *bc = &c->btree_cache;
 805         struct btree *b;
 806         struct bset_tree *t;
 807
 808         EBUG_ON(level >= BTREE_MAX_DEPTH);
 809
 810         b = btree_node_mem_ptr(k);
 811
 812         /*
 813          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 814          * be the node we want anymore, and trying to lock the wrong node could
 815          * cause an unneccessary transaction restart:
 816          */
 817         if (likely(c->opts.btree_node_mem_ptr_optimization &&
 818                    b &&
 819                    b->hash_val == btree_ptr_hash_val(k)))
 820                         goto lock_node;
 821 retry:
 822         b = btree_cache_find(bc, k);
 823         if (unlikely(!b)) {
 824                 /*
 825                  * We must have the parent locked to call bch2_btree_node_fill(),
 826                  * else we could read in a btree node from disk that's been
 827                  * freed:
 828                  */
 829                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 830                                          level, lock_type, true);
 831
 832                 /* We raced and found the btree node in the cache */
 833                 if (!b)
 834                         goto retry;
 835
 836                 if (IS_ERR(b))
 837                         return b;
 838         } else {
 839 lock_node:
 840                 /*
 841                  * There's a potential deadlock with splits and insertions into
 842                  * interior nodes we have to avoid:
 843                  *
 844                  * The other thread might be holding an intent lock on the node
 845                  * we want, and they want to update its parent node so they're
 846                  * going to upgrade their intent lock on the parent node to a
 847                  * write lock.
 848                  *
 849                  * But if we're holding a read lock on the parent, and we're
 850                  * trying to get the intent lock they're holding, we deadlock.
 851                  *
 852                  * So to avoid this we drop the read locks on parent nodes when
 853                  * we're starting to take intent locks - and handle the race.
 854                  *
 855                  * The race is that they might be about to free the node we
 856                  * want, and dropping our read lock on the parent node lets them
 857                  * update the parent marking the node we want as freed, and then
 858                  * free it:
 859                  *
 860                  * To guard against this, btree nodes are evicted from the cache
 861                  * when they're freed - and b->hash_val is zeroed out, which we
 862                  * check for after we lock the node.
 863                  *
 864                  * Then, bch2_btree_node_relock() on the parent will fail - because
 865                  * the parent was modified, when the pointer to the node we want
 866                  * was removed - and we'll bail out:
 867                  */
 868                 if (btree_node_read_locked(path, level + 1))
 869                         btree_node_unlock(path, level + 1);
 870
 871                 if (!btree_node_lock(trans, path, b, k->k.p, level, lock_type,
 872                                      lock_node_check_fn, (void *) k, trace_ip)) {
 873                         if (!trans->restarted)
 874                                 goto retry;
 875                         return ERR_PTR(-EINTR);
 876                 }
 877
 878                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 879                              b->c.level != level ||
 880                              race_fault())) {
 881                         six_unlock_type(&b->c.lock, lock_type);
 882                         if (bch2_btree_node_relock(trans, path, level + 1))
 883                                 goto retry;
 884
 885                         trace_trans_restart_btree_node_reused(trans->fn,
 886                                                               trace_ip,
 887                                                               path->btree_id,
 888                                                               &path->pos);
 889                         btree_trans_restart(trans);
 890                         return ERR_PTR(-EINTR);
 891                 }
 892         }
 893
 894         if (unlikely(btree_node_read_in_flight(b))) {
 895                 u32 seq = b->c.lock.state.seq;
 896
 897                 six_unlock_type(&b->c.lock, lock_type);
 898                 bch2_trans_unlock(trans);
 899
 900                 bch2_btree_node_wait_on_read(b);
 901
 902                 /*
 903                  * should_be_locked is not set on this path yet, so we need to
 904                  * relock it specifically:
 905                  */
 906                 if (trans &&
 907                     (!bch2_trans_relock(trans) ||
 908                      !bch2_btree_path_relock_intent(trans, path))) {
 909                         BUG_ON(!trans->restarted);
 910                         return ERR_PTR(-EINTR);
 911                 }
 912
 913                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 914                         goto retry;
 915         }
 916
 917         prefetch(b->aux_data);
 918
 919         for_each_bset(b, t) {
 920                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 921
 922                 prefetch(p + L1_CACHE_BYTES * 0);
 923                 prefetch(p + L1_CACHE_BYTES * 1);
 924                 prefetch(p + L1_CACHE_BYTES * 2);
 925         }
 926
 927         /* avoid atomic set bit if it's not needed: */
 928         if (!btree_node_accessed(b))
 929                 set_btree_node_accessed(b);
 930
 931         if (unlikely(btree_node_read_error(b))) {
 932                 six_unlock_type(&b->c.lock, lock_type);
 933                 return ERR_PTR(-EIO);
 934         }
 935
 936         EBUG_ON(b->c.btree_id != path->btree_id);
 937         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 938         btree_check_header(c, b);
 939
 940         return b;
 941 }
 942
 943 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 944                                          const struct bkey_i *k,
 945                                          enum btree_id btree_id,
 946                                          unsigned level,
 947                                          bool nofill)
 948 {
 949         struct btree_cache *bc = &c->btree_cache;
 950         struct btree *b;
 951         struct bset_tree *t;
 952         int ret;
 953
 954         EBUG_ON(level >= BTREE_MAX_DEPTH);
 955
 956         if (c->opts.btree_node_mem_ptr_optimization) {
 957                 b = btree_node_mem_ptr(k);
 958                 if (b)
 959                         goto lock_node;
 960         }
 961 retry:
 962         b = btree_cache_find(bc, k);
 963         if (unlikely(!b)) {
 964                 if (nofill)
 965                         goto out;
 966
 967                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
 968                                          level, SIX_LOCK_read, true);
 969
 970                 /* We raced and found the btree node in the cache */
 971                 if (!b)
 972                         goto retry;
 973
 974                 if (IS_ERR(b) &&
 975                     !bch2_btree_cache_cannibalize_lock(c, NULL))
 976                         goto retry;
 977
 978                 if (IS_ERR(b))
 979                         goto out;
 980         } else {
 981 lock_node:
 982                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 983                 if (ret)
 984                         goto retry;
 985
 986                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 987                              b->c.btree_id != btree_id ||
 988                              b->c.level != level)) {
 989                         six_unlock_read(&b->c.lock);
 990                         goto retry;
 991                 }
 992         }
 993
 994         /* XXX: waiting on IO with btree locks held: */
 995         __bch2_btree_node_wait_on_read(b);
 996
 997         prefetch(b->aux_data);
 998
 999         for_each_bset(b, t) {
1000                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1001
1002                 prefetch(p + L1_CACHE_BYTES * 0);
1003                 prefetch(p + L1_CACHE_BYTES * 1);
1004                 prefetch(p + L1_CACHE_BYTES * 2);
1005         }
1006
1007         /* avoid atomic set bit if it's not needed: */
1008         if (!btree_node_accessed(b))
1009                 set_btree_node_accessed(b);
1010
1011         if (unlikely(btree_node_read_error(b))) {
1012                 six_unlock_read(&b->c.lock);
1013                 b = ERR_PTR(-EIO);
1014                 goto out;
1015         }
1016
1017         EBUG_ON(b->c.btree_id != btree_id);
1018         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1019         btree_check_header(c, b);
1020 out:
1021         bch2_btree_cache_cannibalize_unlock(c);
1022         return b;
1023 }
1024
1025 int bch2_btree_node_prefetch(struct bch_fs *c,
1026                              struct btree_trans *trans,
1027                              struct btree_path *path,
1028                              const struct bkey_i *k,
1029                              enum btree_id btree_id, unsigned level)
1030 {
1031         struct btree_cache *bc = &c->btree_cache;
1032         struct btree *b;
1033
1034         BUG_ON(trans && !btree_node_locked(path, level + 1));
1035         BUG_ON(level >= BTREE_MAX_DEPTH);
1036
1037         b = btree_cache_find(bc, k);
1038         if (b)
1039                 return 0;
1040
1041         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1042                                  level, SIX_LOCK_read, false);
1043         return PTR_ERR_OR_ZERO(b);
1044 }
1045
1046 void bch2_btree_node_evict(struct bch_fs *c, const struct bkey_i *k)
1047 {
1048         struct btree_cache *bc = &c->btree_cache;
1049         struct btree *b;
1050
1051         b = btree_cache_find(bc, k);
1052         if (!b)
1053                 return;
1054 wait_on_io:
1055         /* not allowed to wait on io with btree locks held: */
1056
1057         /* XXX we're called from btree_gc which will be holding other btree
1058          * nodes locked
1059          * */
1060         __bch2_btree_node_wait_on_read(b);
1061         __bch2_btree_node_wait_on_write(b);
1062
1063         six_lock_intent(&b->c.lock, NULL, NULL);
1064         six_lock_write(&b->c.lock, NULL, NULL);
1065
1066         if (btree_node_dirty(b)) {
1067                 __bch2_btree_node_write(c, b, 0);
1068                 six_unlock_write(&b->c.lock);
1069                 six_unlock_intent(&b->c.lock);
1070                 goto wait_on_io;
1071         }
1072
1073         BUG_ON(btree_node_dirty(b));
1074
1075         mutex_lock(&bc->lock);
1076         btree_node_data_free(c, b);
1077         bch2_btree_node_hash_remove(bc, b);
1078         mutex_unlock(&bc->lock);
1079
1080         six_unlock_write(&b->c.lock);
1081         six_unlock_intent(&b->c.lock);
1082 }
1083
1084 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1085                              struct btree *b)
1086 {
1087         const struct bkey_format *f = &b->format;
1088         struct bset_stats stats;
1089
1090         memset(&stats, 0, sizeof(stats));
1091
1092         bch2_btree_keys_stats(b, &stats);
1093
1094         pr_buf(out, "l %u ", b->c.level);
1095         bch2_bpos_to_text(out, b->data->min_key);
1096         pr_buf(out, " - ");
1097         bch2_bpos_to_text(out, b->data->max_key);
1098         pr_buf(out, ":\n"
1099                "    ptrs: ");
1100         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1101
1102         pr_buf(out, "\n"
1103                "    format: u64s %u fields %u %u %u %u %u\n"
1104                "    unpack fn len: %u\n"
1105                "    bytes used %zu/%zu (%zu%% full)\n"
1106                "    sib u64s: %u, %u (merge threshold %u)\n"
1107                "    nr packed keys %u\n"
1108                "    nr unpacked keys %u\n"
1109                "    floats %zu\n"
1110                "    failed unpacked %zu\n",
1111                f->key_u64s,
1112                f->bits_per_field[0],
1113                f->bits_per_field[1],
1114                f->bits_per_field[2],
1115                f->bits_per_field[3],
1116                f->bits_per_field[4],
1117                b->unpack_fn_len,
1118                b->nr.live_u64s * sizeof(u64),
1119                btree_bytes(c) - sizeof(struct btree_node),
1120                b->nr.live_u64s * 100 / btree_max_u64s(c),
1121                b->sib_u64s[0],
1122                b->sib_u64s[1],
1123                c->btree_foreground_merge_threshold,
1124                b->nr.packed_keys,
1125                b->nr.unpacked_keys,
1126                stats.floats,
1127                stats.failed);
1128 }
1129
1130 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1131 {
1132         pr_buf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1133         pr_buf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1134         pr_buf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1135 }