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
 278         if (bch2_btree_shrinker_disabled)
 279                 return SHRINK_STOP;
 280
 281         /* Return -1 if we can't do anything right now */
 282         if (sc->gfp_mask & __GFP_FS)
 283                 mutex_lock(&bc->lock);
 284         else if (!mutex_trylock(&bc->lock))
 285                 return -1;
 286
 287         flags = memalloc_nofs_save();
 288
 289         /*
 290          * It's _really_ critical that we don't free too many btree nodes - we
 291          * have to always leave ourselves a reserve. The reserve is how we
 292          * guarantee that allocating memory for a new btree node can always
 293          * succeed, so that inserting keys into the btree can always succeed and
 294          * IO can always make forward progress:
 295          */
 296         nr /= btree_pages(c);
 297         can_free = btree_cache_can_free(bc);
 298         nr = min_t(unsigned long, nr, can_free);
 299
 300         i = 0;
 301         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 302                 touched++;
 303
 304                 if (freed >= nr)
 305                         break;
 306
 307                 if (++i > 3 &&
 308                     !btree_node_reclaim(c, b)) {
 309                         btree_node_data_free(c, b);
 310                         six_unlock_write(&b->c.lock);
 311                         six_unlock_intent(&b->c.lock);
 312                         freed++;
 313                 }
 314         }
 315 restart:
 316         list_for_each_entry_safe(b, t, &bc->live, list) {
 317                 touched++;
 318
 319                 if (freed >= nr) {
 320                         /* Save position */
 321                         if (&t->list != &bc->live)
 322                                 list_move_tail(&bc->live, &t->list);
 323                         break;
 324                 }
 325
 326                 if (!btree_node_accessed(b) &&
 327                     !btree_node_reclaim(c, b)) {
 328                         /* can't call bch2_btree_node_hash_remove under lock  */
 329                         freed++;
 330                         if (&t->list != &bc->live)
 331                                 list_move_tail(&bc->live, &t->list);
 332
 333                         btree_node_data_free(c, b);
 334                         mutex_unlock(&bc->lock);
 335
 336                         bch2_btree_node_hash_remove(bc, b);
 337                         six_unlock_write(&b->c.lock);
 338                         six_unlock_intent(&b->c.lock);
 339
 340                         if (freed >= nr)
 341                                 goto out;
 342
 343                         if (sc->gfp_mask & __GFP_FS)
 344                                 mutex_lock(&bc->lock);
 345                         else if (!mutex_trylock(&bc->lock))
 346                                 goto out;
 347                         goto restart;
 348                 } else
 349                         clear_btree_node_accessed(b);
 350         }
 351
 352         mutex_unlock(&bc->lock);
 353 out:
 354         memalloc_nofs_restore(flags);
 355         return (unsigned long) freed * btree_pages(c);
 356 }
 357
 358 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 359                                             struct shrink_control *sc)
 360 {
 361         struct bch_fs *c = container_of(shrink, struct bch_fs,
 362                                         btree_cache.shrink);
 363         struct btree_cache *bc = &c->btree_cache;
 364
 365         if (bch2_btree_shrinker_disabled)
 366                 return 0;
 367
 368         return btree_cache_can_free(bc) * btree_pages(c);
 369 }
 370
 371 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 372 {
 373         struct btree_cache *bc = &c->btree_cache;
 374         struct btree *b;
 375         unsigned i, flags;
 376
 377         if (bc->shrink.list.next)
 378                 unregister_shrinker(&bc->shrink);
 379
 380         /* vfree() can allocate memory: */
 381         flags = memalloc_nofs_save();
 382         mutex_lock(&bc->lock);
 383
 384         if (c->verify_data)
 385                 list_move(&c->verify_data->list, &bc->live);
 386
 387         kvpfree(c->verify_ondisk, btree_bytes(c));
 388
 389         for (i = 0; i < BTREE_ID_NR; i++)
 390                 if (c->btree_roots[i].b)
 391                         list_add(&c->btree_roots[i].b->list, &bc->live);
 392
 393         list_splice(&bc->freeable, &bc->live);
 394
 395         while (!list_empty(&bc->live)) {
 396                 b = list_first_entry(&bc->live, struct btree, list);
 397
 398                 BUG_ON(btree_node_read_in_flight(b) ||
 399                        btree_node_write_in_flight(b));
 400
 401                 if (btree_node_dirty(b))
 402                         bch2_btree_complete_write(c, b, btree_current_write(b));
 403                 clear_btree_node_dirty(c, b);
 404
 405                 btree_node_data_free(c, b);
 406         }
 407
 408         BUG_ON(atomic_read(&c->btree_cache.dirty));
 409
 410         while (!list_empty(&bc->freed)) {
 411                 b = list_first_entry(&bc->freed, struct btree, list);
 412                 list_del(&b->list);
 413                 six_lock_pcpu_free(&b->c.lock);
 414                 kfree(b);
 415         }
 416
 417         mutex_unlock(&bc->lock);
 418         memalloc_nofs_restore(flags);
 419
 420         if (bc->table_init_done)
 421                 rhashtable_destroy(&bc->table);
 422 }
 423
 424 int bch2_fs_btree_cache_init(struct bch_fs *c)
 425 {
 426         struct btree_cache *bc = &c->btree_cache;
 427         unsigned i;
 428         int ret = 0;
 429
 430         pr_verbose_init(c->opts, "");
 431
 432         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 433         if (ret)
 434                 goto out;
 435
 436         bc->table_init_done = true;
 437
 438         bch2_recalc_btree_reserve(c);
 439
 440         for (i = 0; i < bc->reserve; i++)
 441                 if (!__bch2_btree_node_mem_alloc(c)) {
 442                         ret = -ENOMEM;
 443                         goto out;
 444                 }
 445
 446         list_splice_init(&bc->live, &bc->freeable);
 447
 448         mutex_init(&c->verify_lock);
 449
 450         bc->shrink.count_objects        = bch2_btree_cache_count;
 451         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 452         bc->shrink.seeks                = 4;
 453         bc->shrink.batch                = btree_pages(c) * 2;
 454         ret = register_shrinker(&bc->shrink);
 455 out:
 456         pr_verbose_init(c->opts, "ret %i", ret);
 457         return ret;
 458 }
 459
 460 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 461 {
 462         mutex_init(&bc->lock);
 463         INIT_LIST_HEAD(&bc->live);
 464         INIT_LIST_HEAD(&bc->freeable);
 465         INIT_LIST_HEAD(&bc->freed);
 466 }
 467
 468 /*
 469  * We can only have one thread cannibalizing other cached btree nodes at a time,
 470  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 471  * cannibalize_bucket() will take. This means every time we unlock the root of
 472  * the btree, we need to release this lock if we have it held.
 473  */
 474 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 475 {
 476         struct btree_cache *bc = &c->btree_cache;
 477
 478         if (bc->alloc_lock == current) {
 479                 trace_btree_node_cannibalize_unlock(c);
 480                 bc->alloc_lock = NULL;
 481                 closure_wake_up(&bc->alloc_wait);
 482         }
 483 }
 484
 485 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 486 {
 487         struct btree_cache *bc = &c->btree_cache;
 488         struct task_struct *old;
 489
 490         old = cmpxchg(&bc->alloc_lock, NULL, current);
 491         if (old == NULL || old == current)
 492                 goto success;
 493
 494         if (!cl) {
 495                 trace_btree_node_cannibalize_lock_fail(c);
 496                 return -ENOMEM;
 497         }
 498
 499         closure_wait(&bc->alloc_wait, cl);
 500
 501         /* Try again, after adding ourselves to waitlist */
 502         old = cmpxchg(&bc->alloc_lock, NULL, current);
 503         if (old == NULL || old == current) {
 504                 /* We raced */
 505                 closure_wake_up(&bc->alloc_wait);
 506                 goto success;
 507         }
 508
 509         trace_btree_node_cannibalize_lock_fail(c);
 510         return -EAGAIN;
 511
 512 success:
 513         trace_btree_node_cannibalize_lock(c);
 514         return 0;
 515 }
 516
 517 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 518 {
 519         struct btree_cache *bc = &c->btree_cache;
 520         struct btree *b;
 521
 522         list_for_each_entry_reverse(b, &bc->live, list)
 523                 if (!btree_node_reclaim(c, b))
 524                         return b;
 525
 526         while (1) {
 527                 list_for_each_entry_reverse(b, &bc->live, list)
 528                         if (!btree_node_write_and_reclaim(c, b))
 529                                 return b;
 530
 531                 /*
 532                  * Rare case: all nodes were intent-locked.
 533                  * Just busy-wait.
 534                  */
 535                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 536                 cond_resched();
 537         }
 538 }
 539
 540 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 541 {
 542         struct btree_cache *bc = &c->btree_cache;
 543         struct btree *b;
 544         u64 start_time = local_clock();
 545         unsigned flags;
 546
 547         flags = memalloc_nofs_save();
 548         mutex_lock(&bc->lock);
 549
 550         /*
 551          * btree_free() doesn't free memory; it sticks the node on the end of
 552          * the list. Check if there's any freed nodes there:
 553          */
 554         list_for_each_entry(b, &bc->freeable, list)
 555                 if (!btree_node_reclaim(c, b))
 556                         goto got_node;
 557
 558         /*
 559          * We never free struct btree itself, just the memory that holds the on
 560          * disk node. Check the freed list before allocating a new one:
 561          */
 562         list_for_each_entry(b, &bc->freed, list)
 563                 if (!btree_node_reclaim(c, b))
 564                         goto got_node;
 565
 566         b = NULL;
 567 got_node:
 568         if (b)
 569                 list_del_init(&b->list);
 570         mutex_unlock(&bc->lock);
 571
 572         if (!b) {
 573                 b = __btree_node_mem_alloc(c);
 574                 if (!b)
 575                         goto err;
 576
 577                 BUG_ON(!six_trylock_intent(&b->c.lock));
 578                 BUG_ON(!six_trylock_write(&b->c.lock));
 579         }
 580
 581         if (!b->data) {
 582                 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 583                         goto err;
 584
 585                 mutex_lock(&bc->lock);
 586                 bc->used++;
 587                 mutex_unlock(&bc->lock);
 588         }
 589
 590         BUG_ON(btree_node_hashed(b));
 591         BUG_ON(btree_node_dirty(b));
 592         BUG_ON(btree_node_write_in_flight(b));
 593 out:
 594         b->flags                = 0;
 595         b->written              = 0;
 596         b->nsets                = 0;
 597         b->sib_u64s[0]          = 0;
 598         b->sib_u64s[1]          = 0;
 599         b->whiteout_u64s        = 0;
 600         bch2_btree_keys_init(b);
 601         set_btree_node_accessed(b);
 602
 603         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 604                                start_time);
 605
 606         memalloc_nofs_restore(flags);
 607         return b;
 608 err:
 609         mutex_lock(&bc->lock);
 610
 611         if (b) {
 612                 list_add(&b->list, &bc->freed);
 613                 six_unlock_write(&b->c.lock);
 614                 six_unlock_intent(&b->c.lock);
 615         }
 616
 617         /* Try to cannibalize another cached btree node: */
 618         if (bc->alloc_lock == current) {
 619                 b = btree_node_cannibalize(c);
 620                 list_del_init(&b->list);
 621                 mutex_unlock(&bc->lock);
 622
 623                 bch2_btree_node_hash_remove(bc, b);
 624
 625                 trace_btree_node_cannibalize(c);
 626                 goto out;
 627         }
 628
 629         mutex_unlock(&bc->lock);
 630         memalloc_nofs_restore(flags);
 631         return ERR_PTR(-ENOMEM);
 632 }
 633
 634 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 635 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 636                                 struct btree_trans *trans,
 637                                 struct btree_path *path,
 638                                 const struct bkey_i *k,
 639                                 enum btree_id btree_id,
 640                                 unsigned level,
 641                                 enum six_lock_type lock_type,
 642                                 bool sync)
 643 {
 644         struct btree_cache *bc = &c->btree_cache;
 645         struct btree *b;
 646         u32 seq;
 647
 648         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 649         /*
 650          * Parent node must be locked, else we could read in a btree node that's
 651          * been freed:
 652          */
 653         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 654                 btree_trans_restart(trans);
 655                 return ERR_PTR(-EINTR);
 656         }
 657
 658         b = bch2_btree_node_mem_alloc(c);
 659         if (IS_ERR(b))
 660                 return b;
 661
 662         bkey_copy(&b->key, k);
 663         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 664                 /* raced with another fill: */
 665
 666                 /* mark as unhashed... */
 667                 b->hash_val = 0;
 668
 669                 mutex_lock(&bc->lock);
 670                 list_add(&b->list, &bc->freeable);
 671                 mutex_unlock(&bc->lock);
 672
 673                 six_unlock_write(&b->c.lock);
 674                 six_unlock_intent(&b->c.lock);
 675                 return NULL;
 676         }
 677
 678         set_btree_node_read_in_flight(b);
 679
 680         six_unlock_write(&b->c.lock);
 681         seq = b->c.lock.state.seq;
 682         six_unlock_intent(&b->c.lock);
 683
 684         /* Unlock before doing IO: */
 685         if (trans && sync)
 686                 bch2_trans_unlock(trans);
 687
 688         bch2_btree_node_read(c, b, sync);
 689
 690         if (!sync)
 691                 return NULL;
 692
 693         if (trans &&
 694             (!bch2_trans_relock(trans) ||
 695              !bch2_btree_path_relock_intent(trans, path))) {
 696                 BUG_ON(!trans->restarted);
 697                 return ERR_PTR(-EINTR);
 698         }
 699
 700         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 701                 btree_trans_restart(trans);
 702                 return ERR_PTR(-EINTR);
 703         }
 704
 705         return b;
 706 }
 707
 708 static int lock_node_check_fn(struct six_lock *lock, void *p)
 709 {
 710         struct btree *b = container_of(lock, struct btree, c.lock);
 711         const struct bkey_i *k = p;
 712
 713         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 714 }
 715
 716 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 717 {
 718         char buf1[200], buf2[100], buf3[100];
 719
 720         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 721                 return;
 722
 723         bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&b->key));
 724         bch2_bpos_to_text(&PBUF(buf2), b->data->min_key);
 725         bch2_bpos_to_text(&PBUF(buf3), b->data->max_key);
 726
 727         bch2_fs_inconsistent(c, "btree node header doesn't match ptr\n"
 728                              "btree %s level %u\n"
 729                              "ptr: %s\n"
 730                              "header: btree %s level %llu\n"
 731                              "min %s max %s\n",
 732                              bch2_btree_ids[b->c.btree_id], b->c.level,
 733                              buf1,
 734                              bch2_btree_ids[BTREE_NODE_ID(b->data)],
 735                              BTREE_NODE_LEVEL(b->data),
 736                              buf2, buf3);
 737 }
 738
 739 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 740 {
 741         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 742             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 743             bpos_cmp(b->data->max_key, b->key.k.p) ||
 744             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 745              bpos_cmp(b->data->min_key,
 746                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 747                 btree_bad_header(c, b);
 748 }
 749
 750 /**
 751  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 752  * in from disk if necessary.
 753  *
 754  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 755  *
 756  * The btree node will have either a read or a write lock held, depending on
 757  * the @write parameter.
 758  */
 759 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 760                                   const struct bkey_i *k, unsigned level,
 761                                   enum six_lock_type lock_type,
 762                                   unsigned long trace_ip)
 763 {
 764         struct bch_fs *c = trans->c;
 765         struct btree_cache *bc = &c->btree_cache;
 766         struct btree *b;
 767         struct bset_tree *t;
 768
 769         EBUG_ON(level >= BTREE_MAX_DEPTH);
 770
 771         if (c->opts.btree_node_mem_ptr_optimization) {
 772                 b = btree_node_mem_ptr(k);
 773                 if (b)
 774                         goto lock_node;
 775         }
 776 retry:
 777         b = btree_cache_find(bc, k);
 778         if (unlikely(!b)) {
 779                 /*
 780                  * We must have the parent locked to call bch2_btree_node_fill(),
 781                  * else we could read in a btree node from disk that's been
 782                  * freed:
 783                  */
 784                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 785                                          level, lock_type, true);
 786
 787                 /* We raced and found the btree node in the cache */
 788                 if (!b)
 789                         goto retry;
 790
 791                 if (IS_ERR(b))
 792                         return b;
 793         } else {
 794 lock_node:
 795                 /*
 796                  * There's a potential deadlock with splits and insertions into
 797                  * interior nodes we have to avoid:
 798                  *
 799                  * The other thread might be holding an intent lock on the node
 800                  * we want, and they want to update its parent node so they're
 801                  * going to upgrade their intent lock on the parent node to a
 802                  * write lock.
 803                  *
 804                  * But if we're holding a read lock on the parent, and we're
 805                  * trying to get the intent lock they're holding, we deadlock.
 806                  *
 807                  * So to avoid this we drop the read locks on parent nodes when
 808                  * we're starting to take intent locks - and handle the race.
 809                  *
 810                  * The race is that they might be about to free the node we
 811                  * want, and dropping our read lock on the parent node lets them
 812                  * update the parent marking the node we want as freed, and then
 813                  * free it:
 814                  *
 815                  * To guard against this, btree nodes are evicted from the cache
 816                  * when they're freed - and b->hash_val is zeroed out, which we
 817                  * check for after we lock the node.
 818                  *
 819                  * Then, bch2_btree_node_relock() on the parent will fail - because
 820                  * the parent was modified, when the pointer to the node we want
 821                  * was removed - and we'll bail out:
 822                  */
 823                 if (btree_node_read_locked(path, level + 1))
 824                         btree_node_unlock(path, level + 1);
 825
 826                 if (!btree_node_lock(trans, path, b, k->k.p, level, lock_type,
 827                                      lock_node_check_fn, (void *) k, trace_ip)) {
 828                         if (!trans->restarted)
 829                                 goto retry;
 830                         return ERR_PTR(-EINTR);
 831                 }
 832
 833                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 834                              b->c.level != level ||
 835                              race_fault())) {
 836                         six_unlock_type(&b->c.lock, lock_type);
 837                         if (bch2_btree_node_relock(trans, path, level + 1))
 838                                 goto retry;
 839
 840                         trace_trans_restart_btree_node_reused(trans->ip,
 841                                                               trace_ip,
 842                                                               path->btree_id,
 843                                                               &path->pos);
 844                         btree_trans_restart(trans);
 845                         return ERR_PTR(-EINTR);
 846                 }
 847         }
 848
 849         if (unlikely(btree_node_read_in_flight(b))) {
 850                 u32 seq = b->c.lock.state.seq;
 851
 852                 six_unlock_type(&b->c.lock, lock_type);
 853                 bch2_trans_unlock(trans);
 854
 855                 bch2_btree_node_wait_on_read(b);
 856
 857                 /*
 858                  * should_be_locked is not set on this path yet, so we need to
 859                  * relock it specifically:
 860                  */
 861                 if (trans &&
 862                     (!bch2_trans_relock(trans) ||
 863                      !bch2_btree_path_relock_intent(trans, path))) {
 864                         BUG_ON(!trans->restarted);
 865                         return ERR_PTR(-EINTR);
 866                 }
 867
 868                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 869                         goto retry;
 870         }
 871
 872         prefetch(b->aux_data);
 873
 874         for_each_bset(b, t) {
 875                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 876
 877                 prefetch(p + L1_CACHE_BYTES * 0);
 878                 prefetch(p + L1_CACHE_BYTES * 1);
 879                 prefetch(p + L1_CACHE_BYTES * 2);
 880         }
 881
 882         /* avoid atomic set bit if it's not needed: */
 883         if (!btree_node_accessed(b))
 884                 set_btree_node_accessed(b);
 885
 886         if (unlikely(btree_node_read_error(b))) {
 887                 six_unlock_type(&b->c.lock, lock_type);
 888                 return ERR_PTR(-EIO);
 889         }
 890
 891         EBUG_ON(b->c.btree_id != path->btree_id);
 892         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 893         btree_check_header(c, b);
 894
 895         return b;
 896 }
 897
 898 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 899                                          const struct bkey_i *k,
 900                                          enum btree_id btree_id,
 901                                          unsigned level,
 902                                          bool nofill)
 903 {
 904         struct btree_cache *bc = &c->btree_cache;
 905         struct btree *b;
 906         struct bset_tree *t;
 907         int ret;
 908
 909         EBUG_ON(level >= BTREE_MAX_DEPTH);
 910
 911         if (c->opts.btree_node_mem_ptr_optimization) {
 912                 b = btree_node_mem_ptr(k);
 913                 if (b)
 914                         goto lock_node;
 915         }
 916 retry:
 917         b = btree_cache_find(bc, k);
 918         if (unlikely(!b)) {
 919                 if (nofill)
 920                         goto out;
 921
 922                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
 923                                          level, SIX_LOCK_read, true);
 924
 925                 /* We raced and found the btree node in the cache */
 926                 if (!b)
 927                         goto retry;
 928
 929                 if (IS_ERR(b) &&
 930                     !bch2_btree_cache_cannibalize_lock(c, NULL))
 931                         goto retry;
 932
 933                 if (IS_ERR(b))
 934                         goto out;
 935         } else {
 936 lock_node:
 937                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
 938                 if (ret)
 939                         goto retry;
 940
 941                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 942                              b->c.btree_id != btree_id ||
 943                              b->c.level != level)) {
 944                         six_unlock_read(&b->c.lock);
 945                         goto retry;
 946                 }
 947         }
 948
 949         /* XXX: waiting on IO with btree locks held: */
 950         __bch2_btree_node_wait_on_read(b);
 951
 952         prefetch(b->aux_data);
 953
 954         for_each_bset(b, t) {
 955                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 956
 957                 prefetch(p + L1_CACHE_BYTES * 0);
 958                 prefetch(p + L1_CACHE_BYTES * 1);
 959                 prefetch(p + L1_CACHE_BYTES * 2);
 960         }
 961
 962         /* avoid atomic set bit if it's not needed: */
 963         if (!btree_node_accessed(b))
 964                 set_btree_node_accessed(b);
 965
 966         if (unlikely(btree_node_read_error(b))) {
 967                 six_unlock_read(&b->c.lock);
 968                 b = ERR_PTR(-EIO);
 969                 goto out;
 970         }
 971
 972         EBUG_ON(b->c.btree_id != btree_id);
 973         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 974         btree_check_header(c, b);
 975 out:
 976         bch2_btree_cache_cannibalize_unlock(c);
 977         return b;
 978 }
 979
 980 int bch2_btree_node_prefetch(struct bch_fs *c,
 981                              struct btree_trans *trans,
 982                              struct btree_path *path,
 983                              const struct bkey_i *k,
 984                              enum btree_id btree_id, unsigned level)
 985 {
 986         struct btree_cache *bc = &c->btree_cache;
 987         struct btree *b;
 988
 989         BUG_ON(trans && !btree_node_locked(path, level + 1));
 990         BUG_ON(level >= BTREE_MAX_DEPTH);
 991
 992         b = btree_cache_find(bc, k);
 993         if (b)
 994                 return 0;
 995
 996         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
 997                                  level, SIX_LOCK_read, false);
 998         return PTR_ERR_OR_ZERO(b);
 999 }
1000
1001 void bch2_btree_node_evict(struct bch_fs *c, const struct bkey_i *k)
1002 {
1003         struct btree_cache *bc = &c->btree_cache;
1004         struct btree *b;
1005
1006         b = btree_cache_find(bc, k);
1007         if (!b)
1008                 return;
1009 wait_on_io:
1010         /* not allowed to wait on io with btree locks held: */
1011
1012         /* XXX we're called from btree_gc which will be holding other btree
1013          * nodes locked
1014          * */
1015         __bch2_btree_node_wait_on_read(b);
1016         __bch2_btree_node_wait_on_write(b);
1017
1018         six_lock_intent(&b->c.lock, NULL, NULL);
1019         six_lock_write(&b->c.lock, NULL, NULL);
1020
1021         if (btree_node_dirty(b)) {
1022                 __bch2_btree_node_write(c, b, false);
1023                 six_unlock_write(&b->c.lock);
1024                 six_unlock_intent(&b->c.lock);
1025                 goto wait_on_io;
1026         }
1027
1028         BUG_ON(btree_node_dirty(b));
1029
1030         mutex_lock(&bc->lock);
1031         btree_node_data_free(c, b);
1032         bch2_btree_node_hash_remove(bc, b);
1033         mutex_unlock(&bc->lock);
1034
1035         six_unlock_write(&b->c.lock);
1036         six_unlock_intent(&b->c.lock);
1037 }
1038
1039 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1040                              struct btree *b)
1041 {
1042         const struct bkey_format *f = &b->format;
1043         struct bset_stats stats;
1044
1045         memset(&stats, 0, sizeof(stats));
1046
1047         bch2_btree_keys_stats(b, &stats);
1048
1049         pr_buf(out, "l %u ", b->c.level);
1050         bch2_bpos_to_text(out, b->data->min_key);
1051         pr_buf(out, " - ");
1052         bch2_bpos_to_text(out, b->data->max_key);
1053         pr_buf(out, ":\n"
1054                "    ptrs: ");
1055         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1056
1057         pr_buf(out, "\n"
1058                "    format: u64s %u fields %u %u %u %u %u\n"
1059                "    unpack fn len: %u\n"
1060                "    bytes used %zu/%zu (%zu%% full)\n"
1061                "    sib u64s: %u, %u (merge threshold %u)\n"
1062                "    nr packed keys %u\n"
1063                "    nr unpacked keys %u\n"
1064                "    floats %zu\n"
1065                "    failed unpacked %zu\n",
1066                f->key_u64s,
1067                f->bits_per_field[0],
1068                f->bits_per_field[1],
1069                f->bits_per_field[2],
1070                f->bits_per_field[3],
1071                f->bits_per_field[4],
1072                b->unpack_fn_len,
1073                b->nr.live_u64s * sizeof(u64),
1074                btree_bytes(c) - sizeof(struct btree_node),
1075                b->nr.live_u64s * 100 / btree_max_u64s(c),
1076                b->sib_u64s[0],
1077                b->sib_u64s[1],
1078                c->btree_foreground_merge_threshold,
1079                b->nr.packed_keys,
1080                b->nr.unpacked_keys,
1081                stats.floats,
1082                stats.failed);
1083 }
1084
1085 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1086 {
1087         pr_buf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1088         pr_buf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1089         pr_buf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1090 }