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;
 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:
 364         mutex_unlock(&bc->lock);
 365 out_nounlock:
 366         ret = freed;
 367         memalloc_nofs_restore(flags);
 368         trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
 369         return ret;
 370 }
 371
 372 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 373                                             struct shrink_control *sc)
 374 {
 375         struct bch_fs *c = container_of(shrink, struct bch_fs,
 376                                         btree_cache.shrink);
 377         struct btree_cache *bc = &c->btree_cache;
 378
 379         if (bch2_btree_shrinker_disabled)
 380                 return 0;
 381
 382         return btree_cache_can_free(bc);
 383 }
 384
 385 static void bch2_btree_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
 386 {
 387         struct bch_fs *c = container_of(shrink, struct bch_fs,
 388                                         btree_cache.shrink);
 389
 390         bch2_btree_cache_to_text(out, c);
 391 }
 392
 393 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 394 {
 395         struct btree_cache *bc = &c->btree_cache;
 396         struct btree *b;
 397         unsigned i, flags;
 398
 399         if (bc->shrink.list.next)
 400                 unregister_shrinker(&bc->shrink);
 401
 402         /* vfree() can allocate memory: */
 403         flags = memalloc_nofs_save();
 404         mutex_lock(&bc->lock);
 405
 406         if (c->verify_data)
 407                 list_move(&c->verify_data->list, &bc->live);
 408
 409         kvpfree(c->verify_ondisk, btree_bytes(c));
 410
 411         for (i = 0; i < BTREE_ID_NR; i++)
 412                 if (c->btree_roots[i].b)
 413                         list_add(&c->btree_roots[i].b->list, &bc->live);
 414
 415         list_splice(&bc->freeable, &bc->live);
 416
 417         while (!list_empty(&bc->live)) {
 418                 b = list_first_entry(&bc->live, struct btree, list);
 419
 420                 BUG_ON(btree_node_read_in_flight(b) ||
 421                        btree_node_write_in_flight(b));
 422
 423                 if (btree_node_dirty(b))
 424                         bch2_btree_complete_write(c, b, btree_current_write(b));
 425                 clear_btree_node_dirty_acct(c, b);
 426
 427                 btree_node_data_free(c, b);
 428         }
 429
 430         BUG_ON(atomic_read(&c->btree_cache.dirty));
 431
 432         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 433
 434         while (!list_empty(&bc->freed_nonpcpu)) {
 435                 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
 436                 list_del(&b->list);
 437                 six_lock_pcpu_free(&b->c.lock);
 438                 kfree(b);
 439         }
 440
 441         mutex_unlock(&bc->lock);
 442         memalloc_nofs_restore(flags);
 443
 444         if (bc->table_init_done)
 445                 rhashtable_destroy(&bc->table);
 446 }
 447
 448 int bch2_fs_btree_cache_init(struct bch_fs *c)
 449 {
 450         struct btree_cache *bc = &c->btree_cache;
 451         unsigned i;
 452         int ret = 0;
 453
 454         pr_verbose_init(c->opts, "");
 455
 456         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 457         if (ret)
 458                 goto out;
 459
 460         bc->table_init_done = true;
 461
 462         bch2_recalc_btree_reserve(c);
 463
 464         for (i = 0; i < bc->reserve; i++)
 465                 if (!__bch2_btree_node_mem_alloc(c)) {
 466                         ret = -ENOMEM;
 467                         goto out;
 468                 }
 469
 470         list_splice_init(&bc->live, &bc->freeable);
 471
 472         mutex_init(&c->verify_lock);
 473
 474         bc->shrink.count_objects        = bch2_btree_cache_count;
 475         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 476         bc->shrink.to_text              = bch2_btree_cache_shrinker_to_text;
 477         bc->shrink.seeks                = 4;
 478         ret = register_shrinker(&bc->shrink);
 479 out:
 480         pr_verbose_init(c->opts, "ret %i", ret);
 481         return ret;
 482 }
 483
 484 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 485 {
 486         mutex_init(&bc->lock);
 487         INIT_LIST_HEAD(&bc->live);
 488         INIT_LIST_HEAD(&bc->freeable);
 489         INIT_LIST_HEAD(&bc->freed_pcpu);
 490         INIT_LIST_HEAD(&bc->freed_nonpcpu);
 491 }
 492
 493 /*
 494  * We can only have one thread cannibalizing other cached btree nodes at a time,
 495  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 496  * cannibalize_bucket() will take. This means every time we unlock the root of
 497  * the btree, we need to release this lock if we have it held.
 498  */
 499 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 500 {
 501         struct btree_cache *bc = &c->btree_cache;
 502
 503         if (bc->alloc_lock == current) {
 504                 trace_and_count(c, btree_cache_cannibalize_unlock, c);
 505                 bc->alloc_lock = NULL;
 506                 closure_wake_up(&bc->alloc_wait);
 507         }
 508 }
 509
 510 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 511 {
 512         struct btree_cache *bc = &c->btree_cache;
 513         struct task_struct *old;
 514
 515         old = cmpxchg(&bc->alloc_lock, NULL, current);
 516         if (old == NULL || old == current)
 517                 goto success;
 518
 519         if (!cl) {
 520                 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 521                 return -ENOMEM;
 522         }
 523
 524         closure_wait(&bc->alloc_wait, cl);
 525
 526         /* Try again, after adding ourselves to waitlist */
 527         old = cmpxchg(&bc->alloc_lock, NULL, current);
 528         if (old == NULL || old == current) {
 529                 /* We raced */
 530                 closure_wake_up(&bc->alloc_wait);
 531                 goto success;
 532         }
 533
 534         trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
 535         return -EAGAIN;
 536
 537 success:
 538         trace_and_count(c, btree_cache_cannibalize_lock, c);
 539         return 0;
 540 }
 541
 542 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 543 {
 544         struct btree_cache *bc = &c->btree_cache;
 545         struct btree *b;
 546
 547         list_for_each_entry_reverse(b, &bc->live, list)
 548                 if (!btree_node_reclaim(c, b))
 549                         return b;
 550
 551         while (1) {
 552                 list_for_each_entry_reverse(b, &bc->live, list)
 553                         if (!btree_node_write_and_reclaim(c, b))
 554                                 return b;
 555
 556                 /*
 557                  * Rare case: all nodes were intent-locked.
 558                  * Just busy-wait.
 559                  */
 560                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 561                 cond_resched();
 562         }
 563 }
 564
 565 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
 566 {
 567         struct btree_cache *bc = &c->btree_cache;
 568         struct list_head *freed = pcpu_read_locks
 569                 ? &bc->freed_pcpu
 570                 : &bc->freed_nonpcpu;
 571         struct btree *b, *b2;
 572         u64 start_time = local_clock();
 573         unsigned flags;
 574
 575         flags = memalloc_nofs_save();
 576         mutex_lock(&bc->lock);
 577
 578         /*
 579          * We never free struct btree itself, just the memory that holds the on
 580          * disk node. Check the freed list before allocating a new one:
 581          */
 582         list_for_each_entry(b, freed, list)
 583                 if (!btree_node_reclaim(c, b)) {
 584                         list_del_init(&b->list);
 585                         goto got_node;
 586                 }
 587
 588         b = __btree_node_mem_alloc(c, __GFP_NOWARN);
 589         if (!b) {
 590                 mutex_unlock(&bc->lock);
 591                 b = __btree_node_mem_alloc(c, GFP_KERNEL);
 592                 if (!b)
 593                         goto err;
 594                 mutex_lock(&bc->lock);
 595         }
 596
 597         if (pcpu_read_locks)
 598                 six_lock_pcpu_alloc(&b->c.lock);
 599
 600         BUG_ON(!six_trylock_intent(&b->c.lock));
 601         BUG_ON(!six_trylock_write(&b->c.lock));
 602 got_node:
 603
 604         /*
 605          * btree_free() doesn't free memory; it sticks the node on the end of
 606          * the list. Check if there's any freed nodes there:
 607          */
 608         list_for_each_entry(b2, &bc->freeable, list)
 609                 if (!btree_node_reclaim(c, b2)) {
 610                         swap(b->data, b2->data);
 611                         swap(b->aux_data, b2->aux_data);
 612                         btree_node_to_freedlist(bc, b2);
 613                         six_unlock_write(&b2->c.lock);
 614                         six_unlock_intent(&b2->c.lock);
 615                         goto got_mem;
 616                 }
 617
 618         mutex_unlock(&bc->lock);
 619
 620         if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
 621                 goto err;
 622
 623         mutex_lock(&bc->lock);
 624         bc->used++;
 625 got_mem:
 626         mutex_unlock(&bc->lock);
 627
 628         BUG_ON(btree_node_hashed(b));
 629         BUG_ON(btree_node_dirty(b));
 630         BUG_ON(btree_node_write_in_flight(b));
 631 out:
 632         b->flags                = 0;
 633         b->written              = 0;
 634         b->nsets                = 0;
 635         b->sib_u64s[0]          = 0;
 636         b->sib_u64s[1]          = 0;
 637         b->whiteout_u64s        = 0;
 638         bch2_btree_keys_init(b);
 639         set_btree_node_accessed(b);
 640
 641         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 642                                start_time);
 643
 644         memalloc_nofs_restore(flags);
 645         return b;
 646 err:
 647         mutex_lock(&bc->lock);
 648
 649         /* Try to cannibalize another cached btree node: */
 650         if (bc->alloc_lock == current) {
 651                 b2 = btree_node_cannibalize(c);
 652                 bch2_btree_node_hash_remove(bc, b2);
 653
 654                 if (b) {
 655                         swap(b->data, b2->data);
 656                         swap(b->aux_data, b2->aux_data);
 657                         btree_node_to_freedlist(bc, b2);
 658                         six_unlock_write(&b2->c.lock);
 659                         six_unlock_intent(&b2->c.lock);
 660                 } else {
 661                         b = b2;
 662                         list_del_init(&b->list);
 663                 }
 664
 665                 mutex_unlock(&bc->lock);
 666
 667                 trace_and_count(c, btree_cache_cannibalize, c);
 668                 goto out;
 669         }
 670
 671         mutex_unlock(&bc->lock);
 672         memalloc_nofs_restore(flags);
 673         return ERR_PTR(-ENOMEM);
 674 }
 675
 676 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 677 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 678                                 struct btree_trans *trans,
 679                                 struct btree_path *path,
 680                                 const struct bkey_i *k,
 681                                 enum btree_id btree_id,
 682                                 unsigned level,
 683                                 enum six_lock_type lock_type,
 684                                 bool sync)
 685 {
 686         struct btree_cache *bc = &c->btree_cache;
 687         struct btree *b;
 688         u32 seq;
 689
 690         BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 691         /*
 692          * Parent node must be locked, else we could read in a btree node that's
 693          * been freed:
 694          */
 695         if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
 696                 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
 697                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
 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_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
 705                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
 706         }
 707
 708         if (IS_ERR(b))
 709                 return b;
 710
 711         bkey_copy(&b->key, k);
 712         if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 713                 /* raced with another fill: */
 714
 715                 /* mark as unhashed... */
 716                 b->hash_val = 0;
 717
 718                 mutex_lock(&bc->lock);
 719                 list_add(&b->list, &bc->freeable);
 720                 mutex_unlock(&bc->lock);
 721
 722                 six_unlock_write(&b->c.lock);
 723                 six_unlock_intent(&b->c.lock);
 724                 return NULL;
 725         }
 726
 727         set_btree_node_read_in_flight(b);
 728
 729         six_unlock_write(&b->c.lock);
 730         seq = b->c.lock.state.seq;
 731         six_unlock_intent(&b->c.lock);
 732
 733         /* Unlock before doing IO: */
 734         if (trans && sync)
 735                 bch2_trans_unlock(trans);
 736
 737         bch2_btree_node_read(c, b, sync);
 738
 739         if (!sync)
 740                 return NULL;
 741
 742         if (trans) {
 743                 int ret = bch2_trans_relock(trans) ?:
 744                         bch2_btree_path_relock_intent(trans, path);
 745                 if (ret) {
 746                         BUG_ON(!trans->restarted);
 747                         return ERR_PTR(ret);
 748                 }
 749         }
 750
 751         if (!six_relock_type(&b->c.lock, lock_type, seq)) {
 752                 if (trans)
 753                         trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
 754                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
 755         }
 756
 757         return b;
 758 }
 759
 760 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 761 {
 762         struct printbuf buf = PRINTBUF;
 763
 764         if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
 765                 return;
 766
 767         prt_printf(&buf,
 768                "btree node header doesn't match ptr\n"
 769                "btree %s level %u\n"
 770                "ptr: ",
 771                bch2_btree_ids[b->c.btree_id], b->c.level);
 772         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 773
 774         prt_printf(&buf, "\nheader: btree %s level %llu\n"
 775                "min ",
 776                bch2_btree_ids[BTREE_NODE_ID(b->data)],
 777                BTREE_NODE_LEVEL(b->data));
 778         bch2_bpos_to_text(&buf, b->data->min_key);
 779
 780         prt_printf(&buf, "\nmax ");
 781         bch2_bpos_to_text(&buf, b->data->max_key);
 782
 783         bch2_fs_inconsistent(c, "%s", buf.buf);
 784         printbuf_exit(&buf);
 785 }
 786
 787 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
 788 {
 789         if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
 790             b->c.level != BTREE_NODE_LEVEL(b->data) ||
 791             bpos_cmp(b->data->max_key, b->key.k.p) ||
 792             (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 793              bpos_cmp(b->data->min_key,
 794                       bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
 795                 btree_bad_header(c, b);
 796 }
 797
 798 /**
 799  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 800  * in from disk if necessary.
 801  *
 802  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 803  *
 804  * The btree node will have either a read or a write lock held, depending on
 805  * the @write parameter.
 806  */
 807 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
 808                                   const struct bkey_i *k, unsigned level,
 809                                   enum six_lock_type lock_type,
 810                                   unsigned long trace_ip)
 811 {
 812         struct bch_fs *c = trans->c;
 813         struct btree_cache *bc = &c->btree_cache;
 814         struct btree *b;
 815         struct bset_tree *t;
 816         int ret;
 817
 818         EBUG_ON(level >= BTREE_MAX_DEPTH);
 819
 820         b = btree_node_mem_ptr(k);
 821
 822         /*
 823          * Check b->hash_val _before_ calling btree_node_lock() - this might not
 824          * be the node we want anymore, and trying to lock the wrong node could
 825          * cause an unneccessary transaction restart:
 826          */
 827         if (likely(c->opts.btree_node_mem_ptr_optimization &&
 828                    b &&
 829                    b->hash_val == btree_ptr_hash_val(k)))
 830                         goto lock_node;
 831 retry:
 832         b = btree_cache_find(bc, k);
 833         if (unlikely(!b)) {
 834                 /*
 835                  * We must have the parent locked to call bch2_btree_node_fill(),
 836                  * else we could read in a btree node from disk that's been
 837                  * freed:
 838                  */
 839                 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
 840                                          level, lock_type, true);
 841
 842                 /* We raced and found the btree node in the cache */
 843                 if (!b)
 844                         goto retry;
 845
 846                 if (IS_ERR(b))
 847                         return b;
 848         } else {
 849 lock_node:
 850                 /*
 851                  * There's a potential deadlock with splits and insertions into
 852                  * interior nodes we have to avoid:
 853                  *
 854                  * The other thread might be holding an intent lock on the node
 855                  * we want, and they want to update its parent node so they're
 856                  * going to upgrade their intent lock on the parent node to a
 857                  * write lock.
 858                  *
 859                  * But if we're holding a read lock on the parent, and we're
 860                  * trying to get the intent lock they're holding, we deadlock.
 861                  *
 862                  * So to avoid this we drop the read locks on parent nodes when
 863                  * we're starting to take intent locks - and handle the race.
 864                  *
 865                  * The race is that they might be about to free the node we
 866                  * want, and dropping our read lock on the parent node lets them
 867                  * update the parent marking the node we want as freed, and then
 868                  * free it:
 869                  *
 870                  * To guard against this, btree nodes are evicted from the cache
 871                  * when they're freed - and b->hash_val is zeroed out, which we
 872                  * check for after we lock the node.
 873                  *
 874                  * Then, bch2_btree_node_relock() on the parent will fail - because
 875                  * the parent was modified, when the pointer to the node we want
 876                  * was removed - and we'll bail out:
 877                  */
 878                 if (btree_node_read_locked(path, level + 1))
 879                         btree_node_unlock(trans, path, level + 1);
 880
 881                 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
 882                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 883                         return ERR_PTR(ret);
 884
 885                 BUG_ON(ret);
 886
 887                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 888                              b->c.level != level ||
 889                              race_fault())) {
 890                         six_unlock_type(&b->c.lock, lock_type);
 891                         if (bch2_btree_node_relock(trans, path, level + 1))
 892                                 goto retry;
 893
 894                         trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
 895                         return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
 896                 }
 897         }
 898
 899         if (unlikely(btree_node_read_in_flight(b))) {
 900                 u32 seq = b->c.lock.state.seq;
 901
 902                 six_unlock_type(&b->c.lock, lock_type);
 903                 bch2_trans_unlock(trans);
 904
 905                 bch2_btree_node_wait_on_read(b);
 906
 907                 /*
 908                  * should_be_locked is not set on this path yet, so we need to
 909                  * relock it specifically:
 910                  */
 911                 if (trans) {
 912                         int ret = bch2_trans_relock(trans) ?:
 913                                 bch2_btree_path_relock_intent(trans, path);
 914                         if (ret) {
 915                                 BUG_ON(!trans->restarted);
 916                                 return ERR_PTR(ret);
 917                         }
 918                 }
 919
 920                 if (!six_relock_type(&b->c.lock, lock_type, seq))
 921                         goto retry;
 922         }
 923
 924         prefetch(b->aux_data);
 925
 926         for_each_bset(b, t) {
 927                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 928
 929                 prefetch(p + L1_CACHE_BYTES * 0);
 930                 prefetch(p + L1_CACHE_BYTES * 1);
 931                 prefetch(p + L1_CACHE_BYTES * 2);
 932         }
 933
 934         /* avoid atomic set bit if it's not needed: */
 935         if (!btree_node_accessed(b))
 936                 set_btree_node_accessed(b);
 937
 938         if (unlikely(btree_node_read_error(b))) {
 939                 six_unlock_type(&b->c.lock, lock_type);
 940                 return ERR_PTR(-EIO);
 941         }
 942
 943         EBUG_ON(b->c.btree_id != path->btree_id);
 944         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
 945         btree_check_header(c, b);
 946
 947         return b;
 948 }
 949
 950 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
 951                                          const struct bkey_i *k,
 952                                          enum btree_id btree_id,
 953                                          unsigned level,
 954                                          bool nofill)
 955 {
 956         struct bch_fs *c = trans->c;
 957         struct btree_cache *bc = &c->btree_cache;
 958         struct btree *b;
 959         struct bset_tree *t;
 960         int ret;
 961
 962         EBUG_ON(level >= BTREE_MAX_DEPTH);
 963
 964         if (c->opts.btree_node_mem_ptr_optimization) {
 965                 b = btree_node_mem_ptr(k);
 966                 if (b)
 967                         goto lock_node;
 968         }
 969 retry:
 970         b = btree_cache_find(bc, k);
 971         if (unlikely(!b)) {
 972                 if (nofill)
 973                         goto out;
 974
 975                 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
 976                                          level, SIX_LOCK_read, true);
 977
 978                 /* We raced and found the btree node in the cache */
 979                 if (!b)
 980                         goto retry;
 981
 982                 if (IS_ERR(b) &&
 983                     !bch2_btree_cache_cannibalize_lock(c, NULL))
 984                         goto retry;
 985
 986                 if (IS_ERR(b))
 987                         goto out;
 988         } else {
 989 lock_node:
 990                 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read);
 991                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 992                         return ERR_PTR(ret);
 993
 994                 BUG_ON(ret);
 995
 996                 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 997                              b->c.btree_id != btree_id ||
 998                              b->c.level != level)) {
 999                         six_unlock_read(&b->c.lock);
1000                         goto retry;
1001                 }
1002         }
1003
1004         /* XXX: waiting on IO with btree locks held: */
1005         __bch2_btree_node_wait_on_read(b);
1006
1007         prefetch(b->aux_data);
1008
1009         for_each_bset(b, t) {
1010                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1011
1012                 prefetch(p + L1_CACHE_BYTES * 0);
1013                 prefetch(p + L1_CACHE_BYTES * 1);
1014                 prefetch(p + L1_CACHE_BYTES * 2);
1015         }
1016
1017         /* avoid atomic set bit if it's not needed: */
1018         if (!btree_node_accessed(b))
1019                 set_btree_node_accessed(b);
1020
1021         if (unlikely(btree_node_read_error(b))) {
1022                 six_unlock_read(&b->c.lock);
1023                 b = ERR_PTR(-EIO);
1024                 goto out;
1025         }
1026
1027         EBUG_ON(b->c.btree_id != btree_id);
1028         EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1029         btree_check_header(c, b);
1030 out:
1031         bch2_btree_cache_cannibalize_unlock(c);
1032         return b;
1033 }
1034
1035 int bch2_btree_node_prefetch(struct bch_fs *c,
1036                              struct btree_trans *trans,
1037                              struct btree_path *path,
1038                              const struct bkey_i *k,
1039                              enum btree_id btree_id, unsigned level)
1040 {
1041         struct btree_cache *bc = &c->btree_cache;
1042         struct btree *b;
1043
1044         BUG_ON(trans && !btree_node_locked(path, level + 1));
1045         BUG_ON(level >= BTREE_MAX_DEPTH);
1046
1047         b = btree_cache_find(bc, k);
1048         if (b)
1049                 return 0;
1050
1051         b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1052                                  level, SIX_LOCK_read, false);
1053         return PTR_ERR_OR_ZERO(b);
1054 }
1055
1056 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1057 {
1058         struct bch_fs *c = trans->c;
1059         struct btree_cache *bc = &c->btree_cache;
1060         struct btree *b;
1061
1062         b = btree_cache_find(bc, k);
1063         if (!b)
1064                 return;
1065 wait_on_io:
1066         /* not allowed to wait on io with btree locks held: */
1067
1068         /* XXX we're called from btree_gc which will be holding other btree
1069          * nodes locked
1070          * */
1071         __bch2_btree_node_wait_on_read(b);
1072         __bch2_btree_node_wait_on_write(b);
1073
1074         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1075         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1076
1077         if (btree_node_dirty(b)) {
1078                 __bch2_btree_node_write(c, b, 0);
1079                 six_unlock_write(&b->c.lock);
1080                 six_unlock_intent(&b->c.lock);
1081                 goto wait_on_io;
1082         }
1083
1084         BUG_ON(btree_node_dirty(b));
1085
1086         mutex_lock(&bc->lock);
1087         btree_node_data_free(c, b);
1088         bch2_btree_node_hash_remove(bc, b);
1089         mutex_unlock(&bc->lock);
1090
1091         six_unlock_write(&b->c.lock);
1092         six_unlock_intent(&b->c.lock);
1093 }
1094
1095 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1096                              struct btree *b)
1097 {
1098         const struct bkey_format *f = &b->format;
1099         struct bset_stats stats;
1100
1101         memset(&stats, 0, sizeof(stats));
1102
1103         bch2_btree_keys_stats(b, &stats);
1104
1105         prt_printf(out, "l %u ", b->c.level);
1106         bch2_bpos_to_text(out, b->data->min_key);
1107         prt_printf(out, " - ");
1108         bch2_bpos_to_text(out, b->data->max_key);
1109         prt_printf(out, ":\n"
1110                "    ptrs: ");
1111         bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1112
1113         prt_printf(out, "\n"
1114                "    format: u64s %u fields %u %u %u %u %u\n"
1115                "    unpack fn len: %u\n"
1116                "    bytes used %zu/%zu (%zu%% full)\n"
1117                "    sib u64s: %u, %u (merge threshold %u)\n"
1118                "    nr packed keys %u\n"
1119                "    nr unpacked keys %u\n"
1120                "    floats %zu\n"
1121                "    failed unpacked %zu\n",
1122                f->key_u64s,
1123                f->bits_per_field[0],
1124                f->bits_per_field[1],
1125                f->bits_per_field[2],
1126                f->bits_per_field[3],
1127                f->bits_per_field[4],
1128                b->unpack_fn_len,
1129                b->nr.live_u64s * sizeof(u64),
1130                btree_bytes(c) - sizeof(struct btree_node),
1131                b->nr.live_u64s * 100 / btree_max_u64s(c),
1132                b->sib_u64s[0],
1133                b->sib_u64s[1],
1134                c->btree_foreground_merge_threshold,
1135                b->nr.packed_keys,
1136                b->nr.unpacked_keys,
1137                stats.floats,
1138                stats.failed);
1139 }
1140
1141 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1142 {
1143         prt_printf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1144         prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1145         prt_printf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1146 }