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_to_freedlist(struct btree_cache *bc, struct btree *b)
  46 {
  47         if (b->c.lock.readers)
  48                 list_move(&b->list, &bc->freed_pcpu);
  49         else
  50                 list_move(&b->list, &bc->freed_nonpcpu);
  51 }
  52
  53 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  54 {
  55         struct btree_cache *bc = &c->btree_cache;
  56
  57         EBUG_ON(btree_node_write_in_flight(b));
  58
  59         kvpfree(b->data, btree_bytes(c));
  60         b->data = NULL;
  61 #ifdef __KERNEL__
  62         vfree(b->aux_data);
  63 #else
  64         munmap(b->aux_data, btree_aux_data_bytes(b));
  65 #endif
  66         b->aux_data = NULL;
  67
  68         bc->used--;
  69
  70         btree_node_to_freedlist(bc, b);
  71 }
  72
  73 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  74                                    const void *obj)
  75 {
  76         const struct btree *b = obj;
  77         const u64 *v = arg->key;
  78
  79         return b->hash_val == *v ? 0 : 1;
  80 }
  81
  82 static const struct rhashtable_params bch_btree_cache_params = {
  83         .head_offset    = offsetof(struct btree, hash),
  84         .key_offset     = offsetof(struct btree, hash_val),
  85         .key_len        = sizeof(u64),
  86         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  87 };
  88
  89 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  90 {
  91         BUG_ON(b->data || b->aux_data);
  92
  93         b->data = kvpmalloc(btree_bytes(c), gfp);
  94         if (!b->data)
  95                 return -ENOMEM;
  96 #ifdef __KERNEL__
  97         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  98 #else
  99         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
 100                            PROT_READ|PROT_WRITE|PROT_EXEC,
 101                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 102         if (b->aux_data == MAP_FAILED)
 103                 b->aux_data = NULL;
 104 #endif
 105         if (!b->aux_data) {
 106                 kvpfree(b->data, btree_bytes(c));
 107                 b->data = NULL;
 108                 return -ENOMEM;
 109         }
 110
 111         return 0;
 112 }
 113
 114 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
 115 {
 116         struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
 117         if (!b)
 118                 return NULL;
 119
 120         bkey_btree_ptr_init(&b->key);
 121         __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 122         INIT_LIST_HEAD(&b->list);
 123         INIT_LIST_HEAD(&b->write_blocked);
 124         b->byte_order = ilog2(btree_bytes(c));
 125         return b;
 126 }
 127
 128 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 129 {
 130         struct btree_cache *bc = &c->btree_cache;
 131         struct btree *b = __btree_node_mem_alloc(c);
 132         if (!b)
 133                 return NULL;
 134
 135         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 136                 kfree(b);
 137                 return NULL;
 138         }
 139
 140         bc->used++;
 141         list_add(&b->list, &bc->freeable);
 142         return b;
 143 }
 144
 145 /* Btree in memory cache - hash table */
 146
 147 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 148 {
 149         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 150         BUG_ON(ret);
 151
 152         /* Cause future lookups for this node to fail: */
 153         b->hash_val = 0;
 154
 155         six_lock_wakeup_all(&b->c.lock);
 156 }
 157
 158 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 159 {
 160         BUG_ON(b->hash_val);
 161         b->hash_val = btree_ptr_hash_val(&b->key);
 162
 163         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 164                                              bch_btree_cache_params);
 165 }
 166
 167 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 168                                 unsigned level, enum btree_id id)
 169 {
 170         int ret;
 171
 172         b->c.level      = level;
 173         b->c.btree_id   = id;
 174
 175         mutex_lock(&bc->lock);
 176         ret = __bch2_btree_node_hash_insert(bc, b);
 177         if (!ret)
 178                 list_add(&b->list, &bc->live);
 179         mutex_unlock(&bc->lock);
 180
 181         return ret;
 182 }
 183
 184 __flatten
 185 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 186                                      const struct bkey_i *k)
 187 {
 188         u64 v = btree_ptr_hash_val(k);
 189
 190         return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 191 }
 192
 193 /*
 194  * this version is for btree nodes that have already been freed (we're not
 195  * reaping a real btree node)
 196  */
 197 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 198 {
 199         struct btree_cache *bc = &c->btree_cache;
 200         int ret = 0;
 201
 202         lockdep_assert_held(&bc->lock);
 203 wait_on_io:
 204         if (b->flags & ((1U << BTREE_NODE_dirty)|
 205                         (1U << BTREE_NODE_read_in_flight)|
 206                         (1U << BTREE_NODE_write_in_flight))) {
 207                 if (!flush)
 208                         return -ENOMEM;
 209
 210                 /* XXX: waiting on IO with btree cache lock held */
 211                 bch2_btree_node_wait_on_read(b);
 212                 bch2_btree_node_wait_on_write(b);
 213         }
 214
 215         if (!six_trylock_intent(&b->c.lock))
 216                 return -ENOMEM;
 217
 218         if (!six_trylock_write(&b->c.lock))
 219                 goto out_unlock_intent;
 220
 221         /* recheck under lock */
 222         if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 223                         (1U << BTREE_NODE_write_in_flight))) {
 224                 if (!flush)
 225                         goto out_unlock;
 226                 six_unlock_write(&b->c.lock);
 227                 six_unlock_intent(&b->c.lock);
 228                 goto wait_on_io;
 229         }
 230
 231         if (btree_node_noevict(b) ||
 232             btree_node_write_blocked(b) ||
 233             btree_node_will_make_reachable(b))
 234                 goto out_unlock;
 235
 236         if (btree_node_dirty(b)) {
 237                 if (!flush)
 238                         goto out_unlock;
 239                 /*
 240                  * Using the underscore version because we don't want to compact
 241                  * bsets after the write, since this node is about to be evicted
 242                  * - unless btree verify mode is enabled, since it runs out of
 243                  * the post write cleanup:
 244                  */
 245                 if (bch2_verify_btree_ondisk)
 246                         bch2_btree_node_write(c, b, SIX_LOCK_intent, 0);
 247                 else
 248                         __bch2_btree_node_write(c, b, 0);
 249
 250                 six_unlock_write(&b->c.lock);
 251                 six_unlock_intent(&b->c.lock);
 252                 goto wait_on_io;
 253         }
 254 out:
 255         if (b->hash_val && !ret)
 256                 trace_btree_node_reap(c, b);
 257         return ret;
 258 out_unlock:
 259         six_unlock_write(&b->c.lock);
 260 out_unlock_intent:
 261         six_unlock_intent(&b->c.lock);
 262         ret = -ENOMEM;
 263         goto out;
 264 }
 265
 266 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 267 {
 268         return __btree_node_reclaim(c, b, false);
 269 }
 270
 271 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 272 {
 273         return __btree_node_reclaim(c, b, true);
 274 }
 275
 276 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 277                                            struct shrink_control *sc)
 278 {
 279         struct bch_fs *c = container_of(shrink, struct bch_fs,
 280                                         btree_cache.shrink);
 281         struct btree_cache *bc = &c->btree_cache;
 282         struct btree *b, *t;
 283         unsigned long nr = sc->nr_to_scan;
 284         unsigned long can_free = 0;
 285         unsigned long touched = 0;
 286         unsigned long freed = 0;
 287         unsigned i, flags;
 288         unsigned long ret = SHRINK_STOP;
 289
 290         if (bch2_btree_shrinker_disabled)
 291                 return SHRINK_STOP;
 292
 293         /* Return -1 if we can't do anything right now */
 294         if (sc->gfp_mask & __GFP_FS)
 295                 mutex_lock(&bc->lock);
 296         else if (!mutex_trylock(&bc->lock))
 297                 goto out_norestore;
 298
 299         flags = memalloc_nofs_save();
 300
 301         /*
 302          * It's _really_ critical that we don't free too many btree nodes - we
 303          * have to always leave ourselves a reserve. The reserve is how we
 304          * guarantee that allocating memory for a new btree node can always
 305          * succeed, so that inserting keys into the btree can always succeed and
 306          * IO can always make forward progress:
 307          */
 308         can_free = btree_cache_can_free(bc);
 309         nr = min_t(unsigned long, nr, can_free);
 310
 311         i = 0;
 312         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 313                 /*
 314                  * Leave a few nodes on the freeable list, so that a btree split
 315                  * won't have to hit the system allocator:
 316                  */
 317                 if (++i <= 3)
 318                         continue;
 319
 320                 touched++;
 321
 322                 if (touched >= nr)
 323                         break;
 324
 325                 if (!btree_node_reclaim(c, b)) {
 326                         btree_node_data_free(c, b);
 327                         six_unlock_write(&b->c.lock);
 328                         six_unlock_intent(&b->c.lock);
 329                         freed++;
 330                 }
 331         }
 332 restart:
 333         list_for_each_entry_safe(b, t, &bc->live, list) {
 334                 /* tweak this */
 335                 if (btree_node_accessed(b)) {
 336                         clear_btree_node_accessed(b);
 337                         goto touched;
 338                 }
 339
 340                 if (!btree_node_reclaim(c, b)) {
 341                         /* can't call bch2_btree_node_hash_remove under lock  */
 342                         freed++;
 343                         if (&t->list != &bc->live)
 344                                 list_move_tail(&bc->live, &t->list);
 345
 346                         btree_node_data_free(c, b);
 347                         mutex_unlock(&bc->lock);
 348
 349                         bch2_btree_node_hash_remove(bc, b);
 350                         six_unlock_write(&b->c.lock);
 351                         six_unlock_intent(&b->c.lock);
 352
 353                         if (freed >= nr)
 354                                 goto out;
 355
 356                         if (sc->gfp_mask & __GFP_FS)
 357                                 mutex_lock(&bc->lock);
 358                         else if (!mutex_trylock(&bc->lock))
 359                                 goto out;
 360                         goto restart;
 361                 } else {
 362                         continue;
 363                 }
 364 touched:
 365                 touched++;
 366
 367                 if (touched >= nr) {
 368                         /* Save position */
 369                         if (&t->list != &bc->live)
 370                                 list_move_tail(&bc->live, &t->list);
 371                         break;
 372                 }
 373         }
 374
 375         mutex_unlock(&bc->lock);
 376 out:
 377         ret = freed;
 378         memalloc_nofs_restore(flags);
 379 out_norestore:
 380         trace_btree_cache_scan(sc->nr_to_scan, can_free, ret);
 381         return ret;
 382 }
 383
 384 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 385                                             struct shrink_control *sc)
 386 {
 387         struct bch_fs *c = container_of(shrink, struct bch_fs,
 388                                         btree_cache.shrink);
 389         struct btree_cache *bc = &c->btree_cache;
 390
 391         if (bch2_btree_shrinker_disabled)
 392                 return 0;
 393
 394         return btree_cache_can_free(bc);
 395 }
 396
 397 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 398 {
 399         struct btree_cache *bc = &c->btree_cache;
 400         struct btree *b;
 401         unsigned i, flags;
 402
 403         if (bc->shrink.list.next)
 404                 unregister_shrinker(&bc->shrink);
 405
 406         /* vfree() can allocate memory: */
 407         flags = memalloc_nofs_save();
 408         mutex_lock(&bc->lock);
 409
 410         if (c->verify_data)
 411                 list_move(&c->verify_data->list, &bc->live);
 412
 413         kvpfree(c->verify_ondisk, btree_bytes(c));
 414
 415         for (i = 0; i < BTREE_ID_NR; i++)
 416                 if (c->btree_roots[i].b)
 417                         list_add(&c->btree_roots[i].b->list, &bc->live);
 418
 419         list_splice(&bc->freeable, &bc->live);
 420
 421         while (!list_empty(&bc->live)) {
 422                 b = list_first_entry(&bc->live, struct btree, list);
 423
 424                 BUG_ON(btree_node_read_in_flight(b) ||
 425                        btree_node_write_in_flight(b));
 426
 427                 if (btree_node_dirty(b))
 428                         bch2_btree_complete_write(c, b, btree_current_write(b));
 429                 clear_btree_node_dirty_acct(c, b);
 430
 431                 btree_node_data_free(c, b);
 432         }
 433
 434         BUG_ON(atomic_read(&c->btree_cache.dirty));
 435
 436         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 437
 438         while (!list_empty(&bc->freed_nonpcpu)) {
 439                 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
 440                 list_del(&b->list);
 441                 six_lock_pcpu_free(&b->c.lock);
 442                 kfree(b);
 443         }
 444
 445         mutex_unlock(&bc->lock);
 446         memalloc_nofs_restore(flags);
 447
 448         if (bc->table_init_done)
 449                 rhashtable_destroy(&bc->table);
 450 }
 451
 452 int bch2_fs_btree_cache_init(struct bch_fs *c)
 453 {
 454         struct btree_cache *bc = &c->btree_cache;
 455         unsigned i;
 456         int ret = 0;
 457
 458         pr_verbose_init(c->opts, "");
 459
 460         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 461         if (ret)
 462                 goto out;
 463
 464         bc->table_init_done = true;
 465
 466         bch2_recalc_btree_reserve(c);
 467
 468         for (i = 0; i < bc->reserve; i++)
 469                 if (!__bch2_btree_node_mem_alloc(c)) {
 470                         ret = -ENOMEM;
 471                         goto out;
 472                 }
 473
 474         list_splice_init(&bc->live, &bc->freeable);
 475
 476         mutex_init(&c->verify_lock);
 477
 478         bc->shrink.count_objects        = bch2_btree_cache_count;
 479         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 480         bc->shrink.seeks                = 4;
 481         ret = register_shrinker(&bc->shrink);
 482 out:
 483         pr_verbose_init(c->opts, "ret %i", ret);
 484         return ret;
 485 }
 486
 487 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 488 {
 489         mutex_init(&bc->lock);
 490         INIT_LIST_HEAD(&bc->live);
 491         INIT_LIST_HEAD(&bc->freeable);
 492         INIT_LIST_HEAD(&bc->freed_pcpu);
 493         INIT_LIST_HEAD(&bc->freed_nonpcpu);
 494 }
 495
 496 /*
 497  * We can only have one thread cannibalizing other cached btree nodes at a time,
 498  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 499  * cannibalize_bucket() will take. This means every time we unlock the root of
 500  * the btree, we need to release this lock if we have it held.
 501  */
 502 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 503 {
 504         struct btree_cache *bc = &c->btree_cache;
 505
 506         if (bc->alloc_lock == current) {
 507                 trace_btree_node_cannibalize_unlock(c);
 508                 bc->alloc_lock = NULL;
 509                 closure_wake_up(&bc->alloc_wait);
 510         }
 511 }
 512
 513 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 514 {
 515         struct btree_cache *bc = &c->btree_cache;
 516         struct task_struct *old;
 517
 518         old = cmpxchg(&bc->alloc_lock, NULL, current);
 519         if (old == NULL || old == current)
 520                 goto success;
 521
 522         if (!cl) {
 523                 trace_btree_node_cannibalize_lock_fail(c);
 524                 return -ENOMEM;
 525         }
 526
 527         closure_wait(&bc->alloc_wait, cl);
 528
 529         /* Try again, after adding ourselves to waitlist */
 530         old = cmpxchg(&bc->alloc_lock, NULL, current);
 531         if (old == NULL || old == current) {
 532                 /* We raced */
 533                 closure_wake_up(&bc->alloc_wait);
 534                 goto success;
 535         }
 536
 537         trace_btree_node_cannibalize_lock_fail(c);
 538         return -EAGAIN;
 539
 540 success:
 541         trace_btree_node_cannibalize_lock(c);
 542         return 0;
 543 }
 544
 545 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 546 {
 547         struct btree_cache *bc = &c->btree_cache;
 548         struct btree *b;
 549
 550         list_for_each_entry_reverse(b, &bc->live, list)
 551                 if (!btree_node_reclaim(c, b))
 552                         return b;
 553
 554         while (1) {
 555                 list_for_each_entry_reverse(b, &bc->live, list)
 556                         if (!btree_node_write_and_reclaim(c, b))
 557                                 return b;
 558
 559                 /*
 560                  * Rare case: all nodes were intent-locked.
 561                  * Just busy-wait.
 562                  */
 563                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 564                 cond_resched();
 565         }
 566 }
 567
 568 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
 569 {
 570         struct btree_cache *bc = &c->btree_cache;
 571         struct list_head *freed = pcpu_read_locks
 572                 ? &bc->freed_pcpu
 573                 : &bc->freed_nonpcpu;
 574         struct btree *b, *b2;
 575         u64 start_time = local_clock();
 576         unsigned flags;
 577
 578         flags = memalloc_nofs_save();
 579         mutex_lock(&bc->lock);
 580
 581         /*
 582          * We never free struct btree itself, just the memory that holds the on
 583          * disk node. Check the freed list before allocating a new one:
 584          */
 585         list_for_each_entry(b, freed, list)
 586                 if (!btree_node_reclaim(c, b)) {
 587                         list_del_init(&b->list);
 588                         goto got_node;
 589                 }
 590
 591         b = __btree_node_mem_alloc(c);
 592         if (!b)
 593                 goto err_locked;
 594
 595         if (pcpu_read_locks)
 596                 six_lock_pcpu_alloc(&b->c.lock);
 597
 598         BUG_ON(!six_trylock_intent(&b->c.lock));
 599         BUG_ON(!six_trylock_write(&b->c.lock));
 600 got_node:
 601
 602         /*
 603          * btree_free() doesn't free memory; it sticks the node on the end of
 604          * the list. Check if there's any freed nodes there:
 605          */
 606         list_for_each_entry(b2, &bc->freeable, list)
 607                 if (!btree_node_reclaim(c, b2)) {
 608                         swap(b->data, b2->data);
 609                         swap(b->aux_data, b2->aux_data);
 610                         btree_node_to_freedlist(bc, b2);
 611                         six_unlock_write(&b2->c.lock);
 612                         six_unlock_intent(&b2->c.lock);
 613                         goto got_mem;
 614                 }
 615
 616         mutex_unlock(&bc->lock);
 617
 618         if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 619                 goto err;
 620
 621         mutex_lock(&bc->lock);
 622         bc->used++;
 623 got_mem:
 624         mutex_unlock(&bc->lock);
 625
 626         BUG_ON(btree_node_hashed(b));
 627         BUG_ON(btree_node_dirty(b));
 628         BUG_ON(btree_node_write_in_flight(b));
 629 out:
 630         b->flags                = 0;
 631         b->written              = 0;
 632         b->nsets                = 0;
 633         b->sib_u64s[0]          = 0;
 634         b->sib_u64s[1]          = 0;
 635         b->whiteout_u64s        = 0;
 636         bch2_btree_keys_init(b);
 637         set_btree_node_accessed(b);
 638
 639         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 640                                start_time);
 641
 642         memalloc_nofs_restore(flags);
 643         return b;
 644 err:
 645         mutex_lock(&bc->lock);
 646 err_locked:
 647         /* Try to cannibalize another cached btree node: */
 648         if (bc->alloc_lock == current) {
 649                 b2 = btree_node_cannibalize(c);
 650                 bch2_btree_node_hash_remove(bc, b2);
 651
 652                 if (b) {
 653                         swap(b->data, b2->data);
 654                         swap(b->aux_data, b2->aux_data);
 655                         btree_node_to_freedlist(bc, b2);
 656                         six_unlock_write(&b2->c.lock);
 657                         six_unlock_intent(&b2->c.lock);
 658                 } else {
 659                         b = b2;
 660                         list_del_init(&b->list);
 661                 }
 662
 663                 mutex_unlock(&bc->lock);
 664
 665                 trace_btree_node_cannibalize(c);
 666                 goto out;
 667         }
 668
 669         mutex_unlock(&bc->lock);
 670         memalloc_nofs_restore(flags);
 671         return ERR_PTR(-ENOMEM);
 672 }
 673
 674 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 675 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 676                                 struct btree_trans *trans,
 677                                 struct btree_path *path,
 678                                 const struct bkey_i *k,
 679                                 enum btree_id btree_id,
 680                                 unsigned level,
 681                                 enum six_lock_type lock_type,
 682                                 bool sync)
 683 {
 684         struct btree_cache *bc = &c->btree_cache;
 685         struct btree *b;
 686         u32 seq;
 687
 688         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 689         /*
 690          * Parent node must be locked, else we could read in a btree node that's
 691          * been freed:
 692          */
 693         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 694                 trace_trans_restart_relock_parent_for_fill(trans->fn,
 695                                         _THIS_IP_, btree_id, &path->pos);
 696                 btree_trans_restart(trans);
 697                 return ERR_PTR(-EINTR);
 698         }
 699
 700         b = bch2_btree_node_mem_alloc(c, level != 0);
 701
 702         if (trans && b == ERR_PTR(-ENOMEM)) {
 703                 trans->memory_allocation_failure = true;
 704                 trace_trans_restart_memory_allocation_failure(trans->fn,
 705                                 _THIS_IP_, btree_id, &path->pos);
 706                 btree_trans_restart(trans);
 707                 return ERR_PTR(-EINTR);
 708         }
 709
 710         if (IS_ERR(b))
 711                 return b;
 712
 713         bkey_copy(&b->key, k);
 714         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 715                 /* raced with another fill: */
 716
 717                 /* mark as unhashed... */
 718                 b->hash_val = 0;
 719
 720                 mutex_lock(&bc->lock);
 721                 list_add(&b->list, &bc->freeable);
 722                 mutex_unlock(&bc->lock);
 723
 724                 six_unlock_write(&b->c.lock);
 725                 six_unlock_intent(&b->c.lock);
 726                 return NULL;
 727         }
 728
 729         set_btree_node_read_in_flight(b);
 730
 731         six_unlock_write(&b->c.lock);
 732         seq = b->c.lock.state.seq;
 733         six_unlock_intent(&b->c.lock);
 734
 735         /* Unlock before doing IO: */
 736         if (trans && sync)
 737                 bch2_trans_unlock(trans);
 738
 739         bch2_btree_node_read(c, b, sync);
 740
 741         if (!sync)
 742                 return NULL;
 743
 744         if (trans &&
 745             (!bch2_trans_relock(trans) ||
 746              !bch2_btree_path_relock_intent(trans, path))) {
 747                 BUG_ON(!trans->restarted);
 748                 return ERR_PTR(-EINTR);
 749         }
 750
 751         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 752                 trace_trans_restart_relock_after_fill(trans->fn, _THIS_IP_,
 753                                            btree_id, &path->pos);
 754                 btree_trans_restart(trans);
 755                 return ERR_PTR(-EINTR);
 756         }
 757
 758         return b;
 759 }
 760
 761 static int lock_node_check_fn(struct six_lock *lock, void *p)
 762 {
 763         struct btree *b = container_of(lock, struct btree, c.lock);
 764         const struct bkey_i *k = p;
 765
 766         return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
 767 }
 768
 769 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 770 {
 771         struct printbuf buf = PRINTBUF;
 772
 773         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 774                 return;
 775
 776         pr_buf(&buf,
 777                "btree node header doesn't match ptr\n"
 778                "btree %s level %u\n"
 779                "ptr: ",
 780                bch2_btree_ids[b->c.btree_id], b->c.level);
 781         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 782
 783         pr_buf(&buf, "\nheader: btree %s level %llu\n"
 784                "min ",
 785                bch2_btree_ids[BTREE_NODE_ID(b->data)],
 786                BTREE_NODE_LEVEL(b->data));
 787         bch2_bpos_to_text(&buf, b->data->min_key);
 788
 789         pr_buf(&buf, "\nmax ");
 790         bch2_bpos_to_text(&buf, b->data->max_key);
 791
 792         bch2_fs_inconsistent(c, "%s", buf.buf);
 793         printbuf_exit(&buf);
 794 }
 795
 796 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 797 {
 798         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 799             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 800             bpos_cmp(b->data->max_key, b->key.k.p) ||
 801             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 802              bpos_cmp(b->data->min_key,
 803                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 804                 btree_bad_header(c, b);
 805 }
 806
 807 /**
 808  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 809  * in from disk if necessary.
 810  *
 811  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 812  *
 813  * The btree node will have either a read or a write lock held, depending on
 814  * the @write parameter.
 815  */
 816 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 817                                   const struct bkey_i *k, unsigned level,
 818                                   enum six_lock_type lock_type,
 819                                   unsigned long trace_ip)
 820 {
 821         struct bch_fs *c = trans->c;
 822         struct btree_cache *bc = &c->btree_cache;
 823         struct btree *b;
 824         struct bset_tree *t;
 825
 826         EBUG_ON(level >= BTREE_MAX_DEPTH);
 827
 828         b = btree_node_mem_ptr(k);
 829
 830         /*
 831          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 832          * be the node we want anymore, and trying to lock the wrong node could
 833          * cause an unneccessary transaction restart:
 834          */
 835         if (likely(c->opts.btree_node_mem_ptr_optimization &&
 836                    b &&
 837                    b->hash_val == btree_ptr_hash_val(k)))
 838                         goto lock_node;
 839 retry:
 840         b = btree_cache_find(bc, k);
 841         if (unlikely(!b)) {
 842                 /*
 843                  * We must have the parent locked to call bch2_btree_node_fill(),
 844                  * else we could read in a btree node from disk that's been
 845                  * freed:
 846                  */
 847                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 848                                          level, lock_type, true);
 849
 850                 /* We raced and found the btree node in the cache */
 851                 if (!b)
 852                         goto retry;
 853
 854                 if (IS_ERR(b))
 855                         return b;
 856         } else {
 857 lock_node:
 858                 /*
 859                  * There's a potential deadlock with splits and insertions into
 860                  * interior nodes we have to avoid:
 861                  *
 862                  * The other thread might be holding an intent lock on the node
 863                  * we want, and they want to update its parent node so they're
 864                  * going to upgrade their intent lock on the parent node to a
 865                  * write lock.
 866                  *
 867                  * But if we're holding a read lock on the parent, and we're
 868                  * trying to get the intent lock they're holding, we deadlock.
 869                  *
 870                  * So to avoid this we drop the read locks on parent nodes when
 871                  * we're starting to take intent locks - and handle the race.
 872                  *
 873                  * The race is that they might be about to free the node we
 874                  * want, and dropping our read lock on the parent node lets them
 875                  * update the parent marking the node we want as freed, and then
 876                  * free it:
 877                  *
 878                  * To guard against this, btree nodes are evicted from the cache
 879                  * when they're freed - and b->hash_val is zeroed out, which we
 880                  * check for after we lock the node.
 881                  *
 882                  * Then, bch2_btree_node_relock() on the parent will fail - because
 883                  * the parent was modified, when the pointer to the node we want
 884                  * was removed - and we'll bail out:
 885                  */
 886                 if (btree_node_read_locked(path, level + 1))
 887                         btree_node_unlock(path, level + 1);
 888
 889                 if (!btree_node_lock(trans, path, b, k->k.p, level, lock_type,
 890                                      lock_node_check_fn, (void *) k, trace_ip)) {
 891                         if (!trans->restarted)
 892                                 goto retry;
 893                         return ERR_PTR(-EINTR);
 894                 }
 895
 896                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 897                              b->c.level != level ||
 898                              race_fault())) {
 899                         six_unlock_type(&b->c.lock, lock_type);
 900                         if (bch2_btree_node_relock(trans, path, level + 1))
 901                                 goto retry;
 902
 903                         trace_trans_restart_btree_node_reused(trans->fn,
 904                                                               trace_ip,
 905                                                               path->btree_id,
 906                                                               &path->pos);
 907                         btree_trans_restart(trans);
 908                         return ERR_PTR(-EINTR);
 909                 }
 910         }
 911
 912         if (unlikely(btree_node_read_in_flight(b))) {
 913                 u32 seq = b->c.lock.state.seq;
 914
 915                 six_unlock_type(&b->c.lock, lock_type);
 916                 bch2_trans_unlock(trans);
 917
 918                 bch2_btree_node_wait_on_read(b);
 919
 920                 /*
 921                  * should_be_locked is not set on this path yet, so we need to
 922                  * relock it specifically:
 923                  */
 924                 if (trans &&
 925                     (!bch2_trans_relock(trans) ||
 926                      !bch2_btree_path_relock_intent(trans, path))) {
 927                         BUG_ON(!trans->restarted);
 928                         return ERR_PTR(-EINTR);
 929                 }
 930
 931                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 932                         goto retry;
 933         }
 934
 935         prefetch(b->aux_data);
 936
 937         for_each_bset(b, t) {
 938                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 939
 940                 prefetch(p + L1_CACHE_BYTES * 0);
 941                 prefetch(p + L1_CACHE_BYTES * 1);
 942                 prefetch(p + L1_CACHE_BYTES * 2);
 943         }
 944
 945         /* avoid atomic set bit if it's not needed: */
 946         if (!btree_node_accessed(b))
 947                 set_btree_node_accessed(b);
 948
 949         if (unlikely(btree_node_read_error(b))) {
 950                 six_unlock_type(&b->c.lock, lock_type);
 951                 return ERR_PTR(-EIO);
 952         }
 953
 954         EBUG_ON(b->c.btree_id != path->btree_id);
 955         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 956         btree_check_header(c, b);
 957
 958         return b;
 959 }
 960
 961 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
 962                                          const struct bkey_i *k,
 963                                          enum btree_id btree_id,
 964                                          unsigned level,
 965                                          bool nofill)
 966 {
 967         struct btree_cache *bc = &c->btree_cache;
 968         struct btree *b;
 969         struct bset_tree *t;
 970         int ret;
 971
 972         EBUG_ON(level >= BTREE_MAX_DEPTH);
 973
 974         if (c->opts.btree_node_mem_ptr_optimization) {
 975                 b = btree_node_mem_ptr(k);
 976                 if (b)
 977                         goto lock_node;
 978         }
 979 retry:
 980         b = btree_cache_find(bc, k);
 981         if (unlikely(!b)) {
 982                 if (nofill)
 983                         goto out;
 984
 985                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
 986                                          level, SIX_LOCK_read, true);
 987
 988                 /* We raced and found the btree node in the cache */
 989                 if (!b)
 990                         goto retry;
 991
 992                 if (IS_ERR(b) &&
 993                     !bch2_btree_cache_cannibalize_lock(c, NULL))
 994                         goto retry;
 995
 996                 if (IS_ERR(b))
 997                         goto out;
 998         } else {
 999 lock_node:
1000                 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
1001                 if (ret)
1002                         goto retry;
1003
1004                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1005                              b->c.btree_id != btree_id ||
1006                              b->c.level != level)) {
1007                         six_unlock_read(&b->c.lock);
1008                         goto retry;
1009                 }
1010         }
1011
1012         /* XXX: waiting on IO with btree locks held: */
1013         __bch2_btree_node_wait_on_read(b);
1014
1015         prefetch(b->aux_data);
1016
1017         for_each_bset(b, t) {
1018                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1019
1020                 prefetch(p + L1_CACHE_BYTES * 0);
1021                 prefetch(p + L1_CACHE_BYTES * 1);
1022                 prefetch(p + L1_CACHE_BYTES * 2);
1023         }
1024
1025         /* avoid atomic set bit if it's not needed: */
1026         if (!btree_node_accessed(b))
1027                 set_btree_node_accessed(b);
1028
1029         if (unlikely(btree_node_read_error(b))) {
1030                 six_unlock_read(&b->c.lock);
1031                 b = ERR_PTR(-EIO);
1032                 goto out;
1033         }
1034
1035         EBUG_ON(b->c.btree_id != btree_id);
1036         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1037         btree_check_header(c, b);
1038 out:
1039         bch2_btree_cache_cannibalize_unlock(c);
1040         return b;
1041 }
1042
1043 int bch2_btree_node_prefetch(struct bch_fs *c,
1044                              struct btree_trans *trans,
1045                              struct btree_path *path,
1046                              const struct bkey_i *k,
1047                              enum btree_id btree_id, unsigned level)
1048 {
1049         struct btree_cache *bc = &c->btree_cache;
1050         struct btree *b;
1051
1052         BUG_ON(trans && !btree_node_locked(path, level + 1));
1053         BUG_ON(level >= BTREE_MAX_DEPTH);
1054
1055         b = btree_cache_find(bc, k);
1056         if (b)
1057                 return 0;
1058
1059         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1060                                  level, SIX_LOCK_read, false);
1061         return PTR_ERR_OR_ZERO(b);
1062 }
1063
1064 void bch2_btree_node_evict(struct bch_fs *c, const struct bkey_i *k)
1065 {
1066         struct btree_cache *bc = &c->btree_cache;
1067         struct btree *b;
1068
1069         b = btree_cache_find(bc, k);
1070         if (!b)
1071                 return;
1072 wait_on_io:
1073         /* not allowed to wait on io with btree locks held: */
1074
1075         /* XXX we're called from btree_gc which will be holding other btree
1076          * nodes locked
1077          * */
1078         __bch2_btree_node_wait_on_read(b);
1079         __bch2_btree_node_wait_on_write(b);
1080
1081         six_lock_intent(&b->c.lock, NULL, NULL);
1082         six_lock_write(&b->c.lock, NULL, NULL);
1083
1084         if (btree_node_dirty(b)) {
1085                 __bch2_btree_node_write(c, b, 0);
1086                 six_unlock_write(&b->c.lock);
1087                 six_unlock_intent(&b->c.lock);
1088                 goto wait_on_io;
1089         }
1090
1091         BUG_ON(btree_node_dirty(b));
1092
1093         mutex_lock(&bc->lock);
1094         btree_node_data_free(c, b);
1095         bch2_btree_node_hash_remove(bc, b);
1096         mutex_unlock(&bc->lock);
1097
1098         six_unlock_write(&b->c.lock);
1099         six_unlock_intent(&b->c.lock);
1100 }
1101
1102 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1103                              struct btree *b)
1104 {
1105         const struct bkey_format *f = &b->format;
1106         struct bset_stats stats;
1107
1108         memset(&stats, 0, sizeof(stats));
1109
1110         bch2_btree_keys_stats(b, &stats);
1111
1112         pr_buf(out, "l %u ", b->c.level);
1113         bch2_bpos_to_text(out, b->data->min_key);
1114         pr_buf(out, " - ");
1115         bch2_bpos_to_text(out, b->data->max_key);
1116         pr_buf(out, ":\n"
1117                "    ptrs: ");
1118         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1119
1120         pr_buf(out, "\n"
1121                "    format: u64s %u fields %u %u %u %u %u\n"
1122                "    unpack fn len: %u\n"
1123                "    bytes used %zu/%zu (%zu%% full)\n"
1124                "    sib u64s: %u, %u (merge threshold %u)\n"
1125                "    nr packed keys %u\n"
1126                "    nr unpacked keys %u\n"
1127                "    floats %zu\n"
1128                "    failed unpacked %zu\n",
1129                f->key_u64s,
1130                f->bits_per_field[0],
1131                f->bits_per_field[1],
1132                f->bits_per_field[2],
1133                f->bits_per_field[3],
1134                f->bits_per_field[4],
1135                b->unpack_fn_len,
1136                b->nr.live_u64s * sizeof(u64),
1137                btree_bytes(c) - sizeof(struct btree_node),
1138                b->nr.live_u64s * 100 / btree_max_u64s(c),
1139                b->sib_u64s[0],
1140                b->sib_u64s[1],
1141                c->btree_foreground_merge_threshold,
1142                b->nr.packed_keys,
1143                b->nr.unpacked_keys,
1144                stats.floats,
1145                stats.failed);
1146 }
1147
1148 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1149 {
1150         pr_buf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1151         pr_buf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1152         pr_buf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1153 }