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