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 "errcode.h"
  11 #include "error.h"
  12
  13 #include <linux/prefetch.h>
  14 #include <linux/sched/mm.h>
  15 #include <trace/events/bcachefs.h>
  16
  17 const char * const bch2_btree_node_flags[] = {
  18 #define x(f)    #f,
  19         BTREE_FLAGS()
  20 #undef x
  21         NULL
  22 };
  23
  24 void bch2_recalc_btree_reserve(struct bch_fs *c)
  25 {
  26         unsigned i, reserve = 16;
  27
  28         if (!c->btree_roots[0].b)
  29                 reserve += 8;
  30
  31         for (i = 0; i < BTREE_ID_NR; i++)
  32                 if (c->btree_roots[i].b)
  33                         reserve += min_t(unsigned, 1,
  34                                          c->btree_roots[i].b->c.level) * 8;
  35
  36         c->btree_cache.reserve = reserve;
  37 }
  38
  39 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  40 {
  41         return max_t(int, 0, bc->used - bc->reserve);
  42 }
  43
  44 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
  45 {
  46         if (b->c.lock.readers)
  47                 list_move(&b->list, &bc->freed_pcpu);
  48         else
  49                 list_move(&b->list, &bc->freed_nonpcpu);
  50 }
  51
  52 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  53 {
  54         struct btree_cache *bc = &c->btree_cache;
  55
  56         EBUG_ON(btree_node_write_in_flight(b));
  57
  58         kvpfree(b->data, btree_bytes(c));
  59         b->data = NULL;
  60 #ifdef __KERNEL__
  61         vfree(b->aux_data);
  62 #else
  63         munmap(b->aux_data, btree_aux_data_bytes(b));
  64 #endif
  65         b->aux_data = NULL;
  66
  67         bc->used--;
  68
  69         btree_node_to_freedlist(bc, b);
  70 }
  71
  72 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  73                                    const void *obj)
  74 {
  75         const struct btree *b = obj;
  76         const u64 *v = arg->key;
  77
  78         return b->hash_val == *v ? 0 : 1;
  79 }
  80
  81 static const struct rhashtable_params bch_btree_cache_params = {
  82         .head_offset    = offsetof(struct btree, hash),
  83         .key_offset     = offsetof(struct btree, hash_val),
  84         .key_len        = sizeof(u64),
  85         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  86 };
  87
  88 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  89 {
  90         BUG_ON(b->data || b->aux_data);
  91
  92         b->data = kvpmalloc(btree_bytes(c), gfp);
  93         if (!b->data)
  94                 return -ENOMEM;
  95 #ifdef __KERNEL__
  96         b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
  97 #else
  98         b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
  99                            PROT_READ|PROT_WRITE|PROT_EXEC,
 100                            MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
 101         if (b->aux_data == MAP_FAILED)
 102                 b->aux_data = NULL;
 103 #endif
 104         if (!b->aux_data) {
 105                 kvpfree(b->data, btree_bytes(c));
 106                 b->data = NULL;
 107                 return -ENOMEM;
 108         }
 109
 110         return 0;
 111 }
 112
 113 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
 114 {
 115         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 116         if (!b)
 117                 return NULL;
 118
 119         bkey_btree_ptr_init(&b->key);
 120         __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 121 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 122         lockdep_set_no_check_recursion(&b->c.lock.dep_map);
 123 #endif
 124         INIT_LIST_HEAD(&b->list);
 125         INIT_LIST_HEAD(&b->write_blocked);
 126         b->byte_order = ilog2(btree_bytes(c));
 127         return b;
 128 }
 129
 130 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
 131 {
 132         struct btree_cache *bc = &c->btree_cache;
 133         struct btree *b = __btree_node_mem_alloc(c, GFP_KERNEL);
 134         if (!b)
 135                 return NULL;
 136
 137         if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
 138                 kfree(b);
 139                 return NULL;
 140         }
 141
 142         bc->used++;
 143         list_add(&b->list, &bc->freeable);
 144         return b;
 145 }
 146
 147 /* Btree in memory cache - hash table */
 148
 149 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 150 {
 151         int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 152         BUG_ON(ret);
 153
 154         /* Cause future lookups for this node to fail: */
 155         b->hash_val = 0;
 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_and_count(c, btree_cache_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 freed = 0;
 286         unsigned long touched = 0;
 287         unsigned i, flags;
 288         unsigned long ret = SHRINK_STOP;
 289         bool trigger_writes = atomic_read(&bc->dirty) + nr >=
 290                 bc->used * 3 / 4;
 291
 292         if (bch2_btree_shrinker_disabled)
 293                 return SHRINK_STOP;
 294
 295         mutex_lock(&bc->lock);
 296         flags = memalloc_nofs_save();
 297
 298         /*
 299          * It's _really_ critical that we don't free too many btree nodes - we
 300          * have to always leave ourselves a reserve. The reserve is how we
 301          * guarantee that allocating memory for a new btree node can always
 302          * succeed, so that inserting keys into the btree can always succeed and
 303          * IO can always make forward progress:
 304          */
 305         can_free = btree_cache_can_free(bc);
 306         nr = min_t(unsigned long, nr, can_free);
 307
 308         i = 0;
 309         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 310                 /*
 311                  * Leave a few nodes on the freeable list, so that a btree split
 312                  * won't have to hit the system allocator:
 313                  */
 314                 if (++i <= 3)
 315                         continue;
 316
 317                 touched++;
 318
 319                 if (touched >= nr)
 320                         goto out;
 321
 322                 if (!btree_node_reclaim(c, b)) {
 323                         btree_node_data_free(c, b);
 324                         six_unlock_write(&b->c.lock);
 325                         six_unlock_intent(&b->c.lock);
 326                         freed++;
 327                 }
 328         }
 329 restart:
 330         list_for_each_entry_safe(b, t, &bc->live, list) {
 331                 touched++;
 332
 333                 if (btree_node_accessed(b)) {
 334                         clear_btree_node_accessed(b);
 335                 } else if (!btree_node_reclaim(c, b)) {
 336                         freed++;
 337                         btree_node_data_free(c, b);
 338
 339                         bch2_btree_node_hash_remove(bc, b);
 340                         six_unlock_write(&b->c.lock);
 341                         six_unlock_intent(&b->c.lock);
 342
 343                         if (freed == nr)
 344                                 goto out_rotate;
 345                 } else if (trigger_writes &&
 346                            btree_node_dirty(b) &&
 347                            !btree_node_will_make_reachable(b) &&
 348                            !btree_node_write_blocked(b) &&
 349                            six_trylock_read(&b->c.lock)) {
 350                         list_move(&bc->live, &b->list);
 351                         mutex_unlock(&bc->lock);
 352                         __bch2_btree_node_write(c, b, 0);
 353                         six_unlock_read(&b->c.lock);
 354                         if (touched >= nr)
 355                                 goto out_nounlock;
 356                         mutex_lock(&bc->lock);
 357                         goto restart;
 358                 }
 359
 360                 if (touched >= nr)
 361                         break;
 362         }
 363 out_rotate:
 364         if (&t->list != &bc->live)
 365                 list_move_tail(&bc->live, &t->list);
 366 out:
 367         mutex_unlock(&bc->lock);
 368 out_nounlock:
 369         ret = freed;
 370         memalloc_nofs_restore(flags);
 371         trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
 372         return ret;
 373 }
 374
 375 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 376                                             struct shrink_control *sc)
 377 {
 378         struct bch_fs *c = container_of(shrink, struct bch_fs,
 379                                         btree_cache.shrink);
 380         struct btree_cache *bc = &c->btree_cache;
 381
 382         if (bch2_btree_shrinker_disabled)
 383                 return 0;
 384
 385         return btree_cache_can_free(bc);
 386 }
 387
 388 static void bch2_btree_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
 389 {
 390         struct bch_fs *c = container_of(shrink, struct bch_fs,
 391                                         btree_cache.shrink);
 392
 393         bch2_btree_cache_to_text(out, c);
 394 }
 395
 396 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 397 {
 398         struct btree_cache *bc = &c->btree_cache;
 399         struct btree *b;
 400         unsigned i, flags;
 401
 402         if (bc->shrink.list.next)
 403                 unregister_shrinker(&bc->shrink);
 404
 405         /* vfree() can allocate memory: */
 406         flags = memalloc_nofs_save();
 407         mutex_lock(&bc->lock);
 408
 409         if (c->verify_data)
 410                 list_move(&c->verify_data->list, &bc->live);
 411
 412         kvpfree(c->verify_ondisk, btree_bytes(c));
 413
 414         for (i = 0; i < BTREE_ID_NR; i++)
 415                 if (c->btree_roots[i].b)
 416                         list_add(&c->btree_roots[i].b->list, &bc->live);
 417
 418         list_splice(&bc->freeable, &bc->live);
 419
 420         while (!list_empty(&bc->live)) {
 421                 b = list_first_entry(&bc->live, struct btree, list);
 422
 423                 BUG_ON(btree_node_read_in_flight(b) ||
 424                        btree_node_write_in_flight(b));
 425
 426                 if (btree_node_dirty(b))
 427                         bch2_btree_complete_write(c, b, btree_current_write(b));
 428                 clear_btree_node_dirty_acct(c, b);
 429
 430                 btree_node_data_free(c, b);
 431         }
 432
 433         BUG_ON(atomic_read(&c->btree_cache.dirty));
 434
 435         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 436
 437         while (!list_empty(&bc->freed_nonpcpu)) {
 438                 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
 439                 list_del(&b->list);
 440                 six_lock_pcpu_free(&b->c.lock);
 441                 kfree(b);
 442         }
 443
 444         mutex_unlock(&bc->lock);
 445         memalloc_nofs_restore(flags);
 446
 447         if (bc->table_init_done)
 448                 rhashtable_destroy(&bc->table);
 449 }
 450
 451 int bch2_fs_btree_cache_init(struct bch_fs *c)
 452 {
 453         struct btree_cache *bc = &c->btree_cache;
 454         unsigned i;
 455         int ret = 0;
 456
 457         pr_verbose_init(c->opts, "");
 458
 459         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 460         if (ret)
 461                 goto out;
 462
 463         bc->table_init_done = true;
 464
 465         bch2_recalc_btree_reserve(c);
 466
 467         for (i = 0; i < bc->reserve; i++)
 468                 if (!__bch2_btree_node_mem_alloc(c)) {
 469                         ret = -ENOMEM;
 470                         goto out;
 471                 }
 472
 473         list_splice_init(&bc->live, &bc->freeable);
 474
 475         mutex_init(&c->verify_lock);
 476
 477         bc->shrink.count_objects        = bch2_btree_cache_count;
 478         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 479         bc->shrink.to_text              = bch2_btree_cache_shrinker_to_text;
 480         bc->shrink.seeks                = 4;
 481         ret = register_shrinker(&bc->shrink, "%s/btree_cache", c->name);
 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_and_count(c, btree_cache_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_and_count(c, btree_cache_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_and_count(c, btree_cache_cannibalize_lock_fail, c);
 538         return -EAGAIN;
 539
 540 success:
 541         trace_and_count(c, btree_cache_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, __GFP_NOWARN);
 592         if (!b) {
 593                 mutex_unlock(&bc->lock);
 594                 b = __btree_node_mem_alloc(c, GFP_KERNEL);
 595                 if (!b)
 596                         goto err;
 597                 mutex_lock(&bc->lock);
 598         }
 599
 600         if (pcpu_read_locks)
 601                 six_lock_pcpu_alloc(&b->c.lock);
 602
 603         BUG_ON(!six_trylock_intent(&b->c.lock));
 604         BUG_ON(!six_trylock_write(&b->c.lock));
 605 got_node:
 606
 607         /*
 608          * btree_free() doesn't free memory; it sticks the node on the end of
 609          * the list. Check if there's any freed nodes there:
 610          */
 611         list_for_each_entry(b2, &bc->freeable, list)
 612                 if (!btree_node_reclaim(c, b2)) {
 613                         swap(b->data, b2->data);
 614                         swap(b->aux_data, b2->aux_data);
 615                         btree_node_to_freedlist(bc, b2);
 616                         six_unlock_write(&b2->c.lock);
 617                         six_unlock_intent(&b2->c.lock);
 618                         goto got_mem;
 619                 }
 620
 621         mutex_unlock(&bc->lock);
 622
 623         if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 624                 goto err;
 625
 626         mutex_lock(&bc->lock);
 627         bc->used++;
 628 got_mem:
 629         mutex_unlock(&bc->lock);
 630
 631         BUG_ON(btree_node_hashed(b));
 632         BUG_ON(btree_node_dirty(b));
 633         BUG_ON(btree_node_write_in_flight(b));
 634 out:
 635         b->flags                = 0;
 636         b->written              = 0;
 637         b->nsets                = 0;
 638         b->sib_u64s[0]          = 0;
 639         b->sib_u64s[1]          = 0;
 640         b->whiteout_u64s        = 0;
 641         bch2_btree_keys_init(b);
 642         set_btree_node_accessed(b);
 643
 644         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 645                                start_time);
 646
 647         memalloc_nofs_restore(flags);
 648         return b;
 649 err:
 650         mutex_lock(&bc->lock);
 651
 652         /* Try to cannibalize another cached btree node: */
 653         if (bc->alloc_lock == current) {
 654                 b2 = btree_node_cannibalize(c);
 655                 bch2_btree_node_hash_remove(bc, b2);
 656
 657                 if (b) {
 658                         swap(b->data, b2->data);
 659                         swap(b->aux_data, b2->aux_data);
 660                         btree_node_to_freedlist(bc, b2);
 661                         six_unlock_write(&b2->c.lock);
 662                         six_unlock_intent(&b2->c.lock);
 663                 } else {
 664                         b = b2;
 665                         list_del_init(&b->list);
 666                 }
 667
 668                 mutex_unlock(&bc->lock);
 669
 670                 trace_and_count(c, btree_cache_cannibalize, c);
 671                 goto out;
 672         }
 673
 674         mutex_unlock(&bc->lock);
 675         memalloc_nofs_restore(flags);
 676         return ERR_PTR(-ENOMEM);
 677 }
 678
 679 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 680 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 681                                 struct btree_trans *trans,
 682                                 struct btree_path *path,
 683                                 const struct bkey_i *k,
 684                                 enum btree_id btree_id,
 685                                 unsigned level,
 686                                 enum six_lock_type lock_type,
 687                                 bool sync)
 688 {
 689         struct btree_cache *bc = &c->btree_cache;
 690         struct btree *b;
 691         u32 seq;
 692
 693         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 694         /*
 695          * Parent node must be locked, else we could read in a btree node that's
 696          * been freed:
 697          */
 698         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 699                 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
 700                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
 701         }
 702
 703         b = bch2_btree_node_mem_alloc(c, level != 0);
 704
 705         if (trans && b == ERR_PTR(-ENOMEM)) {
 706                 trans->memory_allocation_failure = true;
 707                 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
 708                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
 709         }
 710
 711         if (IS_ERR(b))
 712                 return b;
 713
 714         bkey_copy(&b->key, k);
 715         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 716                 /* raced with another fill: */
 717
 718                 /* mark as unhashed... */
 719                 b->hash_val = 0;
 720
 721                 mutex_lock(&bc->lock);
 722                 list_add(&b->list, &bc->freeable);
 723                 mutex_unlock(&bc->lock);
 724
 725                 six_unlock_write(&b->c.lock);
 726                 six_unlock_intent(&b->c.lock);
 727                 return NULL;
 728         }
 729
 730         set_btree_node_read_in_flight(b);
 731
 732         six_unlock_write(&b->c.lock);
 733         seq = b->c.lock.state.seq;
 734         six_unlock_intent(&b->c.lock);
 735
 736         /* Unlock before doing IO: */
 737         if (trans && sync)
 738                 bch2_trans_unlock(trans);
 739
 740         bch2_btree_node_read(c, b, sync);
 741
 742         if (!sync)
 743                 return NULL;
 744
 745         if (trans) {
 746                 int ret = bch2_trans_relock(trans) ?:
 747                         bch2_btree_path_relock_intent(trans, path);
 748                 if (ret) {
 749                         BUG_ON(!trans->restarted);
 750                         return ERR_PTR(ret);
 751                 }
 752         }
 753
 754         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 755                 if (trans)
 756                         trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
 757                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
 758         }
 759
 760         return b;
 761 }
 762
 763 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 764 {
 765         struct printbuf buf = PRINTBUF;
 766
 767         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 768                 return;
 769
 770         prt_printf(&buf,
 771                "btree node header doesn't match ptr\n"
 772                "btree %s level %u\n"
 773                "ptr: ",
 774                bch2_btree_ids[b->c.btree_id], b->c.level);
 775         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 776
 777         prt_printf(&buf, "\nheader: btree %s level %llu\n"
 778                "min ",
 779                bch2_btree_ids[BTREE_NODE_ID(b->data)],
 780                BTREE_NODE_LEVEL(b->data));
 781         bch2_bpos_to_text(&buf, b->data->min_key);
 782
 783         prt_printf(&buf, "\nmax ");
 784         bch2_bpos_to_text(&buf, b->data->max_key);
 785
 786         bch2_fs_inconsistent(c, "%s", buf.buf);
 787         printbuf_exit(&buf);
 788 }
 789
 790 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 791 {
 792         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 793             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 794             bpos_cmp(b->data->max_key, b->key.k.p) ||
 795             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 796              bpos_cmp(b->data->min_key,
 797                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 798                 btree_bad_header(c, b);
 799 }
 800
 801 /**
 802  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 803  * in from disk if necessary.
 804  *
 805  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 806  *
 807  * The btree node will have either a read or a write lock held, depending on
 808  * the @write parameter.
 809  */
 810 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 811                                   const struct bkey_i *k, unsigned level,
 812                                   enum six_lock_type lock_type,
 813                                   unsigned long trace_ip)
 814 {
 815         struct bch_fs *c = trans->c;
 816         struct btree_cache *bc = &c->btree_cache;
 817         struct btree *b;
 818         struct bset_tree *t;
 819         int ret;
 820
 821         EBUG_ON(level >= BTREE_MAX_DEPTH);
 822
 823         b = btree_node_mem_ptr(k);
 824
 825         /*
 826          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 827          * be the node we want anymore, and trying to lock the wrong node could
 828          * cause an unneccessary transaction restart:
 829          */
 830         if (likely(c->opts.btree_node_mem_ptr_optimization &&
 831                    b &&
 832                    b->hash_val == btree_ptr_hash_val(k)))
 833                         goto lock_node;
 834 retry:
 835         b = btree_cache_find(bc, k);
 836         if (unlikely(!b)) {
 837                 /*
 838                  * We must have the parent locked to call bch2_btree_node_fill(),
 839                  * else we could read in a btree node from disk that's been
 840                  * freed:
 841                  */
 842                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 843                                          level, lock_type, true);
 844
 845                 /* We raced and found the btree node in the cache */
 846                 if (!b)
 847                         goto retry;
 848
 849                 if (IS_ERR(b))
 850                         return b;
 851         } else {
 852 lock_node:
 853                 /*
 854                  * There's a potential deadlock with splits and insertions into
 855                  * interior nodes we have to avoid:
 856                  *
 857                  * The other thread might be holding an intent lock on the node
 858                  * we want, and they want to update its parent node so they're
 859                  * going to upgrade their intent lock on the parent node to a
 860                  * write lock.
 861                  *
 862                  * But if we're holding a read lock on the parent, and we're
 863                  * trying to get the intent lock they're holding, we deadlock.
 864                  *
 865                  * So to avoid this we drop the read locks on parent nodes when
 866                  * we're starting to take intent locks - and handle the race.
 867                  *
 868                  * The race is that they might be about to free the node we
 869                  * want, and dropping our read lock on the parent node lets them
 870                  * update the parent marking the node we want as freed, and then
 871                  * free it:
 872                  *
 873                  * To guard against this, btree nodes are evicted from the cache
 874                  * when they're freed - and b->hash_val is zeroed out, which we
 875                  * check for after we lock the node.
 876                  *
 877                  * Then, bch2_btree_node_relock() on the parent will fail - because
 878                  * the parent was modified, when the pointer to the node we want
 879                  * was removed - and we'll bail out:
 880                  */
 881                 if (btree_node_read_locked(path, level + 1))
 882                         btree_node_unlock(trans, path, level + 1);
 883
 884                 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
 885                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 886                         return ERR_PTR(ret);
 887
 888                 BUG_ON(ret);
 889
 890                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 891                              b->c.level != level ||
 892                              race_fault())) {
 893                         six_unlock_type(&b->c.lock, lock_type);
 894                         if (bch2_btree_node_relock(trans, path, level + 1))
 895                                 goto retry;
 896
 897                         trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
 898                         return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
 899                 }
 900         }
 901
 902         if (unlikely(btree_node_read_in_flight(b))) {
 903                 u32 seq = b->c.lock.state.seq;
 904
 905                 six_unlock_type(&b->c.lock, lock_type);
 906                 bch2_trans_unlock(trans);
 907
 908                 bch2_btree_node_wait_on_read(b);
 909
 910                 /*
 911                  * should_be_locked is not set on this path yet, so we need to
 912                  * relock it specifically:
 913                  */
 914                 if (trans) {
 915                         int ret = bch2_trans_relock(trans) ?:
 916                                 bch2_btree_path_relock_intent(trans, path);
 917                         if (ret) {
 918                                 BUG_ON(!trans->restarted);
 919                                 return ERR_PTR(ret);
 920                         }
 921                 }
 922
 923                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 924                         goto retry;
 925         }
 926
 927         prefetch(b->aux_data);
 928
 929         for_each_bset(b, t) {
 930                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 931
 932                 prefetch(p + L1_CACHE_BYTES * 0);
 933                 prefetch(p + L1_CACHE_BYTES * 1);
 934                 prefetch(p + L1_CACHE_BYTES * 2);
 935         }
 936
 937         /* avoid atomic set bit if it's not needed: */
 938         if (!btree_node_accessed(b))
 939                 set_btree_node_accessed(b);
 940
 941         if (unlikely(btree_node_read_error(b))) {
 942                 six_unlock_type(&b->c.lock, lock_type);
 943                 return ERR_PTR(-EIO);
 944         }
 945
 946         EBUG_ON(b->c.btree_id != path->btree_id);
 947         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 948         btree_check_header(c, b);
 949
 950         return b;
 951 }
 952
 953 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
 954                                          const struct bkey_i *k,
 955                                          enum btree_id btree_id,
 956                                          unsigned level,
 957                                          bool nofill)
 958 {
 959         struct bch_fs *c = trans->c;
 960         struct btree_cache *bc = &c->btree_cache;
 961         struct btree *b;
 962         struct bset_tree *t;
 963         int ret;
 964
 965         EBUG_ON(level >= BTREE_MAX_DEPTH);
 966
 967         if (c->opts.btree_node_mem_ptr_optimization) {
 968                 b = btree_node_mem_ptr(k);
 969                 if (b)
 970                         goto lock_node;
 971         }
 972 retry:
 973         b = btree_cache_find(bc, k);
 974         if (unlikely(!b)) {
 975                 if (nofill)
 976                         goto out;
 977
 978                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
 979                                          level, SIX_LOCK_read, true);
 980
 981                 /* We raced and found the btree node in the cache */
 982                 if (!b)
 983                         goto retry;
 984
 985                 if (IS_ERR(b) &&
 986                     !bch2_btree_cache_cannibalize_lock(c, NULL))
 987                         goto retry;
 988
 989                 if (IS_ERR(b))
 990                         goto out;
 991         } else {
 992 lock_node:
 993                 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read);
 994                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 995                         return ERR_PTR(ret);
 996
 997                 BUG_ON(ret);
 998
 999                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1000                              b->c.btree_id != btree_id ||
1001                              b->c.level != level)) {
1002                         six_unlock_read(&b->c.lock);
1003                         goto retry;
1004                 }
1005         }
1006
1007         /* XXX: waiting on IO with btree locks held: */
1008         __bch2_btree_node_wait_on_read(b);
1009
1010         prefetch(b->aux_data);
1011
1012         for_each_bset(b, t) {
1013                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1014
1015                 prefetch(p + L1_CACHE_BYTES * 0);
1016                 prefetch(p + L1_CACHE_BYTES * 1);
1017                 prefetch(p + L1_CACHE_BYTES * 2);
1018         }
1019
1020         /* avoid atomic set bit if it's not needed: */
1021         if (!btree_node_accessed(b))
1022                 set_btree_node_accessed(b);
1023
1024         if (unlikely(btree_node_read_error(b))) {
1025                 six_unlock_read(&b->c.lock);
1026                 b = ERR_PTR(-EIO);
1027                 goto out;
1028         }
1029
1030         EBUG_ON(b->c.btree_id != btree_id);
1031         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1032         btree_check_header(c, b);
1033 out:
1034         bch2_btree_cache_cannibalize_unlock(c);
1035         return b;
1036 }
1037
1038 int bch2_btree_node_prefetch(struct bch_fs *c,
1039                              struct btree_trans *trans,
1040                              struct btree_path *path,
1041                              const struct bkey_i *k,
1042                              enum btree_id btree_id, unsigned level)
1043 {
1044         struct btree_cache *bc = &c->btree_cache;
1045         struct btree *b;
1046
1047         BUG_ON(trans && !btree_node_locked(path, level + 1));
1048         BUG_ON(level >= BTREE_MAX_DEPTH);
1049
1050         b = btree_cache_find(bc, k);
1051         if (b)
1052                 return 0;
1053
1054         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1055                                  level, SIX_LOCK_read, false);
1056         return PTR_ERR_OR_ZERO(b);
1057 }
1058
1059 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1060 {
1061         struct bch_fs *c = trans->c;
1062         struct btree_cache *bc = &c->btree_cache;
1063         struct btree *b;
1064
1065         b = btree_cache_find(bc, k);
1066         if (!b)
1067                 return;
1068 wait_on_io:
1069         /* not allowed to wait on io with btree locks held: */
1070
1071         /* XXX we're called from btree_gc which will be holding other btree
1072          * nodes locked
1073          * */
1074         __bch2_btree_node_wait_on_read(b);
1075         __bch2_btree_node_wait_on_write(b);
1076
1077         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1078         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1079
1080         if (btree_node_dirty(b)) {
1081                 __bch2_btree_node_write(c, b, 0);
1082                 six_unlock_write(&b->c.lock);
1083                 six_unlock_intent(&b->c.lock);
1084                 goto wait_on_io;
1085         }
1086
1087         BUG_ON(btree_node_dirty(b));
1088
1089         mutex_lock(&bc->lock);
1090         btree_node_data_free(c, b);
1091         bch2_btree_node_hash_remove(bc, b);
1092         mutex_unlock(&bc->lock);
1093
1094         six_unlock_write(&b->c.lock);
1095         six_unlock_intent(&b->c.lock);
1096 }
1097
1098 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1099                              struct btree *b)
1100 {
1101         const struct bkey_format *f = &b->format;
1102         struct bset_stats stats;
1103
1104         memset(&stats, 0, sizeof(stats));
1105
1106         bch2_btree_keys_stats(b, &stats);
1107
1108         prt_printf(out, "l %u ", b->c.level);
1109         bch2_bpos_to_text(out, b->data->min_key);
1110         prt_printf(out, " - ");
1111         bch2_bpos_to_text(out, b->data->max_key);
1112         prt_printf(out, ":\n"
1113                "    ptrs: ");
1114         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1115
1116         prt_printf(out, "\n"
1117                "    format: u64s %u fields %u %u %u %u %u\n"
1118                "    unpack fn len: %u\n"
1119                "    bytes used %zu/%zu (%zu%% full)\n"
1120                "    sib u64s: %u, %u (merge threshold %u)\n"
1121                "    nr packed keys %u\n"
1122                "    nr unpacked keys %u\n"
1123                "    floats %zu\n"
1124                "    failed unpacked %zu\n",
1125                f->key_u64s,
1126                f->bits_per_field[0],
1127                f->bits_per_field[1],
1128                f->bits_per_field[2],
1129                f->bits_per_field[3],
1130                f->bits_per_field[4],
1131                b->unpack_fn_len,
1132                b->nr.live_u64s * sizeof(u64),
1133                btree_bytes(c) - sizeof(struct btree_node),
1134                b->nr.live_u64s * 100 / btree_max_u64s(c),
1135                b->sib_u64s[0],
1136                b->sib_u64s[1],
1137                c->btree_foreground_merge_threshold,
1138                b->nr.packed_keys,
1139                b->nr.unpacked_keys,
1140                stats.floats,
1141                stats.failed);
1142 }
1143
1144 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1145 {
1146         prt_printf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1147         prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1148         prt_printf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1149 }