git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_cache.c

   1
   2 #include "bcachefs.h"
   3 #include "btree_cache.h"
   4 #include "btree_io.h"
   5 #include "btree_iter.h"
   6 #include "btree_locking.h"
   7 #include "debug.h"
   8 #include "extents.h"
   9
  10 #include <linux/prefetch.h>
  11 #include <trace/events/bcachefs.h>
  12
  13 #define DEF_BTREE_ID(kwd, val, name) name,
  14
  15 const char * const bch2_btree_ids[] = {
  16         DEFINE_BCH_BTREE_IDS()
  17         NULL
  18 };
  19
  20 #undef DEF_BTREE_ID
  21
  22 void bch2_recalc_btree_reserve(struct bch_fs *c)
  23 {
  24         unsigned i, reserve = 16;
  25
  26         if (!c->btree_roots[0].b)
  27                 reserve += 8;
  28
  29         for (i = 0; i < BTREE_ID_NR; i++)
  30                 if (c->btree_roots[i].b)
  31                         reserve += min_t(unsigned, 1,
  32                                          c->btree_roots[i].b->level) * 8;
  33
  34         c->btree_cache.reserve = reserve;
  35 }
  36
  37 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
  38 {
  39         return max_t(int, 0, bc->used - bc->reserve);
  40 }
  41
  42 static void __btree_node_data_free(struct bch_fs *c, struct btree *b)
  43 {
  44         EBUG_ON(btree_node_write_in_flight(b));
  45
  46         kvpfree(b->data, btree_bytes(c));
  47         b->data = NULL;
  48         bch2_btree_keys_free(b);
  49 }
  50
  51 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
  52 {
  53         struct btree_cache *bc = &c->btree_cache;
  54
  55         __btree_node_data_free(c, b);
  56         bc->used--;
  57         list_move(&b->list, &bc->freed);
  58 }
  59
  60 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  61                                    const void *obj)
  62 {
  63         const struct btree *b = obj;
  64         const u64 *v = arg->key;
  65
  66         return PTR_HASH(&b->key) == *v ? 0 : 1;
  67 }
  68
  69 static const struct rhashtable_params bch_btree_cache_params = {
  70         .head_offset    = offsetof(struct btree, hash),
  71         .key_offset     = offsetof(struct btree, key.v),
  72         .key_len        = sizeof(struct bch_extent_ptr),
  73         .obj_cmpfn      = bch2_btree_cache_cmp_fn,
  74 };
  75
  76 static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
  77 {
  78         struct btree_cache *bc = &c->btree_cache;
  79
  80         b->data = kvpmalloc(btree_bytes(c), gfp);
  81         if (!b->data)
  82                 goto err;
  83
  84         if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp))
  85                 goto err;
  86
  87         bc->used++;
  88         list_move(&b->list, &bc->freeable);
  89         return;
  90 err:
  91         kvpfree(b->data, btree_bytes(c));
  92         b->data = NULL;
  93         list_move(&b->list, &bc->freed);
  94 }
  95
  96 static struct btree *btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
  97 {
  98         struct btree *b = kzalloc(sizeof(struct btree), gfp);
  99         if (!b)
 100                 return NULL;
 101
 102         bkey_extent_init(&b->key);
 103         six_lock_init(&b->lock);
 104         INIT_LIST_HEAD(&b->list);
 105         INIT_LIST_HEAD(&b->write_blocked);
 106
 107         btree_node_data_alloc(c, b, gfp);
 108         return b->data ? b : NULL;
 109 }
 110
 111 /* Btree in memory cache - hash table */
 112
 113 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 114 {
 115         rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 116
 117         /* Cause future lookups for this node to fail: */
 118         bkey_i_to_extent(&b->key)->v._data[0] = 0;
 119 }
 120
 121 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 122 {
 123         return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 124                                              bch_btree_cache_params);
 125 }
 126
 127 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 128                                 unsigned level, enum btree_id id)
 129 {
 130         int ret;
 131
 132         b->level        = level;
 133         b->btree_id     = id;
 134
 135         mutex_lock(&bc->lock);
 136         ret = __bch2_btree_node_hash_insert(bc, b);
 137         if (!ret)
 138                 list_add(&b->list, &bc->live);
 139         mutex_unlock(&bc->lock);
 140
 141         return ret;
 142 }
 143
 144 __flatten
 145 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 146                                      const struct bkey_i *k)
 147 {
 148         return rhashtable_lookup_fast(&bc->table, &PTR_HASH(k),
 149                                       bch_btree_cache_params);
 150 }
 151
 152 /*
 153  * this version is for btree nodes that have already been freed (we're not
 154  * reaping a real btree node)
 155  */
 156 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 157 {
 158         struct btree_cache *bc = &c->btree_cache;
 159         int ret = 0;
 160
 161         lockdep_assert_held(&bc->lock);
 162
 163         if (!six_trylock_intent(&b->lock))
 164                 return -ENOMEM;
 165
 166         if (!six_trylock_write(&b->lock))
 167                 goto out_unlock_intent;
 168
 169         if (btree_node_noevict(b))
 170                 goto out_unlock;
 171
 172         if (!btree_node_may_write(b))
 173                 goto out_unlock;
 174
 175         if (btree_node_dirty(b) ||
 176             btree_node_write_in_flight(b) ||
 177             btree_node_read_in_flight(b)) {
 178                 if (!flush)
 179                         goto out_unlock;
 180
 181                 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 182                                TASK_UNINTERRUPTIBLE);
 183
 184                 /*
 185                  * Using the underscore version because we don't want to compact
 186                  * bsets after the write, since this node is about to be evicted
 187                  * - unless btree verify mode is enabled, since it runs out of
 188                  * the post write cleanup:
 189                  */
 190                 if (verify_btree_ondisk(c))
 191                         bch2_btree_node_write(c, b, SIX_LOCK_intent);
 192                 else
 193                         __bch2_btree_node_write(c, b, SIX_LOCK_read);
 194
 195                 /* wait for any in flight btree write */
 196                 btree_node_wait_on_io(b);
 197         }
 198 out:
 199         if (PTR_HASH(&b->key) && !ret)
 200                 trace_btree_node_reap(c, b);
 201         return ret;
 202 out_unlock:
 203         six_unlock_write(&b->lock);
 204 out_unlock_intent:
 205         six_unlock_intent(&b->lock);
 206         ret = -ENOMEM;
 207         goto out;
 208 }
 209
 210 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
 211 {
 212         return __btree_node_reclaim(c, b, false);
 213 }
 214
 215 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 216 {
 217         return __btree_node_reclaim(c, b, true);
 218 }
 219
 220 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 221                                            struct shrink_control *sc)
 222 {
 223         struct bch_fs *c = container_of(shrink, struct bch_fs,
 224                                         btree_cache.shrink);
 225         struct btree_cache *bc = &c->btree_cache;
 226         struct btree *b, *t;
 227         unsigned long nr = sc->nr_to_scan;
 228         unsigned long can_free;
 229         unsigned long touched = 0;
 230         unsigned long freed = 0;
 231         unsigned i;
 232
 233         if (btree_shrinker_disabled(c))
 234                 return SHRINK_STOP;
 235
 236         /* Return -1 if we can't do anything right now */
 237         if (sc->gfp_mask & __GFP_IO)
 238                 mutex_lock(&bc->lock);
 239         else if (!mutex_trylock(&bc->lock))
 240                 return -1;
 241
 242         /*
 243          * It's _really_ critical that we don't free too many btree nodes - we
 244          * have to always leave ourselves a reserve. The reserve is how we
 245          * guarantee that allocating memory for a new btree node can always
 246          * succeed, so that inserting keys into the btree can always succeed and
 247          * IO can always make forward progress:
 248          */
 249         nr /= btree_pages(c);
 250         can_free = btree_cache_can_free(bc);
 251         nr = min_t(unsigned long, nr, can_free);
 252
 253         i = 0;
 254         list_for_each_entry_safe(b, t, &bc->freeable, list) {
 255                 touched++;
 256
 257                 if (freed >= nr)
 258                         break;
 259
 260                 if (++i > 3 &&
 261                     !btree_node_reclaim(c, b)) {
 262                         btree_node_data_free(c, b);
 263                         six_unlock_write(&b->lock);
 264                         six_unlock_intent(&b->lock);
 265                         freed++;
 266                 }
 267         }
 268 restart:
 269         list_for_each_entry_safe(b, t, &bc->live, list) {
 270                 touched++;
 271
 272                 if (freed >= nr) {
 273                         /* Save position */
 274                         if (&t->list != &bc->live)
 275                                 list_move_tail(&bc->live, &t->list);
 276                         break;
 277                 }
 278
 279                 if (!btree_node_accessed(b) &&
 280                     !btree_node_reclaim(c, b)) {
 281                         /* can't call bch2_btree_node_hash_remove under lock  */
 282                         freed++;
 283                         if (&t->list != &bc->live)
 284                                 list_move_tail(&bc->live, &t->list);
 285
 286                         btree_node_data_free(c, b);
 287                         mutex_unlock(&bc->lock);
 288
 289                         bch2_btree_node_hash_remove(bc, b);
 290                         six_unlock_write(&b->lock);
 291                         six_unlock_intent(&b->lock);
 292
 293                         if (freed >= nr)
 294                                 goto out;
 295
 296                         if (sc->gfp_mask & __GFP_IO)
 297                                 mutex_lock(&bc->lock);
 298                         else if (!mutex_trylock(&bc->lock))
 299                                 goto out;
 300                         goto restart;
 301                 } else
 302                         clear_btree_node_accessed(b);
 303         }
 304
 305         mutex_unlock(&bc->lock);
 306 out:
 307         return (unsigned long) freed * btree_pages(c);
 308 }
 309
 310 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
 311                                             struct shrink_control *sc)
 312 {
 313         struct bch_fs *c = container_of(shrink, struct bch_fs,
 314                                         btree_cache.shrink);
 315         struct btree_cache *bc = &c->btree_cache;
 316
 317         if (btree_shrinker_disabled(c))
 318                 return 0;
 319
 320         return btree_cache_can_free(bc) * btree_pages(c);
 321 }
 322
 323 void bch2_fs_btree_cache_exit(struct bch_fs *c)
 324 {
 325         struct btree_cache *bc = &c->btree_cache;
 326         struct btree *b;
 327         unsigned i;
 328
 329         if (bc->shrink.list.next)
 330                 unregister_shrinker(&bc->shrink);
 331
 332         mutex_lock(&bc->lock);
 333
 334 #ifdef CONFIG_BCACHEFS_DEBUG
 335         if (c->verify_data)
 336                 list_move(&c->verify_data->list, &bc->live);
 337
 338         kvpfree(c->verify_ondisk, btree_bytes(c));
 339 #endif
 340
 341         for (i = 0; i < BTREE_ID_NR; i++)
 342                 if (c->btree_roots[i].b)
 343                         list_add(&c->btree_roots[i].b->list, &bc->live);
 344
 345         list_splice(&bc->freeable, &bc->live);
 346
 347         while (!list_empty(&bc->live)) {
 348                 b = list_first_entry(&bc->live, struct btree, list);
 349
 350                 BUG_ON(btree_node_read_in_flight(b) ||
 351                        btree_node_write_in_flight(b));
 352
 353                 if (btree_node_dirty(b))
 354                         bch2_btree_complete_write(c, b, btree_current_write(b));
 355                 clear_btree_node_dirty(b);
 356
 357                 btree_node_data_free(c, b);
 358         }
 359
 360         while (!list_empty(&bc->freed)) {
 361                 b = list_first_entry(&bc->freed, struct btree, list);
 362                 list_del(&b->list);
 363                 kfree(b);
 364         }
 365
 366         mutex_unlock(&bc->lock);
 367
 368         if (bc->table_init_done)
 369                 rhashtable_destroy(&bc->table);
 370 }
 371
 372 int bch2_fs_btree_cache_init(struct bch_fs *c)
 373 {
 374         struct btree_cache *bc = &c->btree_cache;
 375         unsigned i;
 376         int ret = 0;
 377
 378         pr_verbose_init(c->opts, "");
 379
 380         ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
 381         if (ret)
 382                 goto out;
 383
 384         bc->table_init_done = true;
 385
 386         bch2_recalc_btree_reserve(c);
 387
 388         for (i = 0; i < bc->reserve; i++)
 389                 if (!btree_node_mem_alloc(c, GFP_KERNEL)) {
 390                         ret = -ENOMEM;
 391                         goto out;
 392                 }
 393
 394         list_splice_init(&bc->live, &bc->freeable);
 395
 396 #ifdef CONFIG_BCACHEFS_DEBUG
 397         mutex_init(&c->verify_lock);
 398
 399         c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
 400         if (!c->verify_ondisk) {
 401                 ret = -ENOMEM;
 402                 goto out;
 403         }
 404
 405         c->verify_data = btree_node_mem_alloc(c, GFP_KERNEL);
 406         if (!c->verify_data) {
 407                 ret = -ENOMEM;
 408                 goto out;
 409         }
 410
 411         list_del_init(&c->verify_data->list);
 412 #endif
 413
 414         bc->shrink.count_objects        = bch2_btree_cache_count;
 415         bc->shrink.scan_objects         = bch2_btree_cache_scan;
 416         bc->shrink.seeks                = 4;
 417         bc->shrink.batch                = btree_pages(c) * 2;
 418         register_shrinker(&bc->shrink);
 419 out:
 420         pr_verbose_init(c->opts, "ret %i", ret);
 421         return ret;
 422 }
 423
 424 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
 425 {
 426         mutex_init(&bc->lock);
 427         INIT_LIST_HEAD(&bc->live);
 428         INIT_LIST_HEAD(&bc->freeable);
 429         INIT_LIST_HEAD(&bc->freed);
 430 }
 431
 432 /*
 433  * We can only have one thread cannibalizing other cached btree nodes at a time,
 434  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 435  * cannibalize_bucket() will take. This means every time we unlock the root of
 436  * the btree, we need to release this lock if we have it held.
 437  */
 438 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
 439 {
 440         struct btree_cache *bc = &c->btree_cache;
 441
 442         if (bc->alloc_lock == current) {
 443                 trace_btree_node_cannibalize_unlock(c);
 444                 bc->alloc_lock = NULL;
 445                 closure_wake_up(&bc->alloc_wait);
 446         }
 447 }
 448
 449 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
 450 {
 451         struct btree_cache *bc = &c->btree_cache;
 452         struct task_struct *old;
 453
 454         old = cmpxchg(&bc->alloc_lock, NULL, current);
 455         if (old == NULL || old == current)
 456                 goto success;
 457
 458         if (!cl) {
 459                 trace_btree_node_cannibalize_lock_fail(c);
 460                 return -ENOMEM;
 461         }
 462
 463         closure_wait(&bc->alloc_wait, cl);
 464
 465         /* Try again, after adding ourselves to waitlist */
 466         old = cmpxchg(&bc->alloc_lock, NULL, current);
 467         if (old == NULL || old == current) {
 468                 /* We raced */
 469                 closure_wake_up(&bc->alloc_wait);
 470                 goto success;
 471         }
 472
 473         trace_btree_node_cannibalize_lock_fail(c);
 474         return -EAGAIN;
 475
 476 success:
 477         trace_btree_node_cannibalize_lock(c);
 478         return 0;
 479 }
 480
 481 static struct btree *btree_node_cannibalize(struct bch_fs *c)
 482 {
 483         struct btree_cache *bc = &c->btree_cache;
 484         struct btree *b;
 485
 486         list_for_each_entry_reverse(b, &bc->live, list)
 487                 if (!btree_node_reclaim(c, b))
 488                         return b;
 489
 490         while (1) {
 491                 list_for_each_entry_reverse(b, &bc->live, list)
 492                         if (!btree_node_write_and_reclaim(c, b))
 493                                 return b;
 494
 495                 /*
 496                  * Rare case: all nodes were intent-locked.
 497                  * Just busy-wait.
 498                  */
 499                 WARN_ONCE(1, "btree cache cannibalize failed\n");
 500                 cond_resched();
 501         }
 502 }
 503
 504 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
 505 {
 506         struct btree_cache *bc = &c->btree_cache;
 507         struct btree *b;
 508         u64 start_time = local_clock();
 509
 510         mutex_lock(&bc->lock);
 511
 512         /*
 513          * btree_free() doesn't free memory; it sticks the node on the end of
 514          * the list. Check if there's any freed nodes there:
 515          */
 516         list_for_each_entry(b, &bc->freeable, list)
 517                 if (!btree_node_reclaim(c, b))
 518                         goto out_unlock;
 519
 520         /*
 521          * We never free struct btree itself, just the memory that holds the on
 522          * disk node. Check the freed list before allocating a new one:
 523          */
 524         list_for_each_entry(b, &bc->freed, list)
 525                 if (!btree_node_reclaim(c, b)) {
 526                         btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
 527                         if (b->data)
 528                                 goto out_unlock;
 529
 530                         six_unlock_write(&b->lock);
 531                         six_unlock_intent(&b->lock);
 532                         goto err;
 533                 }
 534
 535         b = btree_node_mem_alloc(c, __GFP_NOWARN|GFP_NOIO);
 536         if (!b)
 537                 goto err;
 538
 539         BUG_ON(!six_trylock_intent(&b->lock));
 540         BUG_ON(!six_trylock_write(&b->lock));
 541 out_unlock:
 542         BUG_ON(btree_node_hashed(b));
 543         BUG_ON(btree_node_write_in_flight(b));
 544
 545         list_del_init(&b->list);
 546         mutex_unlock(&bc->lock);
 547 out:
 548         b->flags                = 0;
 549         b->written              = 0;
 550         b->nsets                = 0;
 551         b->sib_u64s[0]          = 0;
 552         b->sib_u64s[1]          = 0;
 553         b->whiteout_u64s        = 0;
 554         b->uncompacted_whiteout_u64s = 0;
 555         bch2_btree_keys_init(b, &c->expensive_debug_checks);
 556
 557         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
 558                                start_time);
 559
 560         return b;
 561 err:
 562         /* Try to cannibalize another cached btree node: */
 563         if (bc->alloc_lock == current) {
 564                 b = btree_node_cannibalize(c);
 565                 list_del_init(&b->list);
 566                 mutex_unlock(&bc->lock);
 567
 568                 bch2_btree_node_hash_remove(bc, b);
 569
 570                 trace_btree_node_cannibalize(c);
 571                 goto out;
 572         }
 573
 574         mutex_unlock(&bc->lock);
 575         return ERR_PTR(-ENOMEM);
 576 }
 577
 578 /* Slowpath, don't want it inlined into btree_iter_traverse() */
 579 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 580                                 struct btree_iter *iter,
 581                                 const struct bkey_i *k,
 582                                 unsigned level,
 583                                 enum six_lock_type lock_type,
 584                                 bool sync)
 585 {
 586         struct btree_cache *bc = &c->btree_cache;
 587         struct btree *b;
 588
 589         /*
 590          * Parent node must be locked, else we could read in a btree node that's
 591          * been freed:
 592          */
 593         BUG_ON(!btree_node_locked(iter, level + 1));
 594         BUG_ON(level >= BTREE_MAX_DEPTH);
 595
 596         b = bch2_btree_node_mem_alloc(c);
 597         if (IS_ERR(b))
 598                 return b;
 599
 600         bkey_copy(&b->key, k);
 601         if (bch2_btree_node_hash_insert(bc, b, level, iter->btree_id)) {
 602                 /* raced with another fill: */
 603
 604                 /* mark as unhashed... */
 605                 bkey_i_to_extent(&b->key)->v._data[0] = 0;
 606
 607                 mutex_lock(&bc->lock);
 608                 list_add(&b->list, &bc->freeable);
 609                 mutex_unlock(&bc->lock);
 610
 611                 six_unlock_write(&b->lock);
 612                 six_unlock_intent(&b->lock);
 613                 return NULL;
 614         }
 615
 616         /*
 617          * If the btree node wasn't cached, we can't drop our lock on
 618          * the parent until after it's added to the cache - because
 619          * otherwise we could race with a btree_split() freeing the node
 620          * we're trying to lock.
 621          *
 622          * But the deadlock described below doesn't exist in this case,
 623          * so it's safe to not drop the parent lock until here:
 624          */
 625         if (btree_node_read_locked(iter, level + 1))
 626                 btree_node_unlock(iter, level + 1);
 627
 628         bch2_btree_node_read(c, b, sync);
 629
 630         six_unlock_write(&b->lock);
 631
 632         if (!sync) {
 633                 six_unlock_intent(&b->lock);
 634                 return NULL;
 635         }
 636
 637         if (lock_type == SIX_LOCK_read)
 638                 six_lock_downgrade(&b->lock);
 639
 640         return b;
 641 }
 642
 643 /**
 644  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 645  * in from disk if necessary.
 646  *
 647  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 648  *
 649  * The btree node will have either a read or a write lock held, depending on
 650  * the @write parameter.
 651  */
 652 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 653                                   const struct bkey_i *k, unsigned level,
 654                                   enum six_lock_type lock_type,
 655                                   bool may_drop_locks)
 656 {
 657         struct btree_cache *bc = &c->btree_cache;
 658         struct btree *b;
 659         struct bset_tree *t;
 660
 661         /*
 662          * XXX: locking optimization
 663          *
 664          * we can make the locking looser here - caller can drop lock on parent
 665          * node before locking child node (and potentially blocking): we just
 666          * have to have bch2_btree_node_fill() call relock on the parent and
 667          * return -EINTR if that fails
 668          */
 669         EBUG_ON(!btree_node_locked(iter, level + 1));
 670         EBUG_ON(level >= BTREE_MAX_DEPTH);
 671 retry:
 672         rcu_read_lock();
 673         b = btree_cache_find(bc, k);
 674         rcu_read_unlock();
 675
 676         if (unlikely(!b)) {
 677                 /*
 678                  * We must have the parent locked to call bch2_btree_node_fill(),
 679                  * else we could read in a btree node from disk that's been
 680                  * freed:
 681                  */
 682                 b = bch2_btree_node_fill(c, iter, k, level, lock_type, true);
 683
 684                 /* We raced and found the btree node in the cache */
 685                 if (!b)
 686                         goto retry;
 687
 688                 if (IS_ERR(b))
 689                         return b;
 690         } else {
 691                 /*
 692                  * There's a potential deadlock with splits and insertions into
 693                  * interior nodes we have to avoid:
 694                  *
 695                  * The other thread might be holding an intent lock on the node
 696                  * we want, and they want to update its parent node so they're
 697                  * going to upgrade their intent lock on the parent node to a
 698                  * write lock.
 699                  *
 700                  * But if we're holding a read lock on the parent, and we're
 701                  * trying to get the intent lock they're holding, we deadlock.
 702                  *
 703                  * So to avoid this we drop the read locks on parent nodes when
 704                  * we're starting to take intent locks - and handle the race.
 705                  *
 706                  * The race is that they might be about to free the node we
 707                  * want, and dropping our read lock on the parent node lets them
 708                  * update the parent marking the node we want as freed, and then
 709                  * free it:
 710                  *
 711                  * To guard against this, btree nodes are evicted from the cache
 712                  * when they're freed - and PTR_HASH() is zeroed out, which we
 713                  * check for after we lock the node.
 714                  *
 715                  * Then, bch2_btree_node_relock() on the parent will fail - because
 716                  * the parent was modified, when the pointer to the node we want
 717                  * was removed - and we'll bail out:
 718                  */
 719                 if (btree_node_read_locked(iter, level + 1))
 720                         btree_node_unlock(iter, level + 1);
 721
 722                 if (!btree_node_lock(b, k->k.p, level, iter,
 723                                      lock_type, may_drop_locks))
 724                         return ERR_PTR(-EINTR);
 725
 726                 if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
 727                              b->level != level ||
 728                              race_fault())) {
 729                         six_unlock_type(&b->lock, lock_type);
 730                         if (bch2_btree_node_relock(iter, level + 1))
 731                                 goto retry;
 732
 733                         trans_restart();
 734                         return ERR_PTR(-EINTR);
 735                 }
 736         }
 737
 738         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 739                        TASK_UNINTERRUPTIBLE);
 740
 741         prefetch(b->aux_data);
 742
 743         for_each_bset(b, t) {
 744                 void *p = (u64 *) b->aux_data + t->aux_data_offset;
 745
 746                 prefetch(p + L1_CACHE_BYTES * 0);
 747                 prefetch(p + L1_CACHE_BYTES * 1);
 748                 prefetch(p + L1_CACHE_BYTES * 2);
 749         }
 750
 751         /* avoid atomic set bit if it's not needed: */
 752         if (btree_node_accessed(b))
 753                 set_btree_node_accessed(b);
 754
 755         if (unlikely(btree_node_read_error(b))) {
 756                 six_unlock_type(&b->lock, lock_type);
 757                 return ERR_PTR(-EIO);
 758         }
 759
 760         EBUG_ON(b->btree_id != iter->btree_id ||
 761                 BTREE_NODE_LEVEL(b->data) != level ||
 762                 bkey_cmp(b->data->max_key, k->k.p));
 763
 764         return b;
 765 }
 766
 767 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 768                                           struct btree_iter *iter,
 769                                           struct btree *b,
 770                                           bool may_drop_locks,
 771                                           enum btree_node_sibling sib)
 772 {
 773         struct btree *parent;
 774         struct btree_node_iter node_iter;
 775         struct bkey_packed *k;
 776         BKEY_PADDED(k) tmp;
 777         struct btree *ret = NULL;
 778         unsigned level = b->level;
 779
 780         parent = btree_iter_node(iter, level + 1);
 781         if (!parent)
 782                 return NULL;
 783
 784         if (!bch2_btree_node_relock(iter, level + 1))
 785                 goto out_upgrade;
 786
 787         node_iter = iter->l[parent->level].iter;
 788
 789         k = bch2_btree_node_iter_peek_all(&node_iter, parent);
 790         BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
 791
 792         k = sib == btree_prev_sib
 793                 ? bch2_btree_node_iter_prev(&node_iter, parent)
 794                 : (bch2_btree_node_iter_advance(&node_iter, parent),
 795                    bch2_btree_node_iter_peek(&node_iter, parent));
 796         if (!k)
 797                 goto out;
 798
 799         bch2_bkey_unpack(parent, &tmp.k, k);
 800
 801         ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 802                                   SIX_LOCK_intent, may_drop_locks);
 803
 804         if (PTR_ERR_OR_ZERO(ret) == -EINTR && may_drop_locks) {
 805                 struct btree_iter *linked;
 806
 807                 if (!bch2_btree_node_relock(iter, level + 1))
 808                         goto out_upgrade;
 809
 810                 /*
 811                  * We might have got -EINTR because trylock failed, and we're
 812                  * holding other locks that would cause us to deadlock:
 813                  */
 814                 for_each_linked_btree_iter(iter, linked)
 815                         if (btree_iter_cmp(iter, linked) < 0)
 816                                 __bch2_btree_iter_unlock(linked);
 817
 818                 if (sib == btree_prev_sib)
 819                         btree_node_unlock(iter, level);
 820
 821                 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
 822                                           SIX_LOCK_intent, may_drop_locks);
 823
 824                 /*
 825                  * before btree_iter_relock() calls btree_iter_verify_locks():
 826                  */
 827                 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 828                         btree_node_unlock(iter, level + 1);
 829
 830                 if (!bch2_btree_node_relock(iter, level)) {
 831                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
 832
 833                         if (!IS_ERR(ret)) {
 834                                 six_unlock_intent(&ret->lock);
 835                                 ret = ERR_PTR(-EINTR);
 836                         }
 837                 }
 838
 839                 bch2_btree_iter_relock(iter);
 840         }
 841 out:
 842         if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
 843                 btree_node_unlock(iter, level + 1);
 844
 845         bch2_btree_iter_verify_locks(iter);
 846
 847         BUG_ON((!may_drop_locks || !IS_ERR(ret)) &&
 848                (iter->uptodate >= BTREE_ITER_NEED_RELOCK ||
 849                 !btree_node_locked(iter, level)));
 850
 851         if (!IS_ERR_OR_NULL(ret)) {
 852                 struct btree *n1 = ret, *n2 = b;
 853
 854                 if (sib != btree_prev_sib)
 855                         swap(n1, n2);
 856
 857                 BUG_ON(bkey_cmp(btree_type_successor(n1->btree_id,
 858                                                      n1->key.k.p),
 859                                 n2->data->min_key));
 860         }
 861
 862         return ret;
 863 out_upgrade:
 864         if (may_drop_locks)
 865                 bch2_btree_iter_upgrade(iter, level + 2, true);
 866         ret = ERR_PTR(-EINTR);
 867         goto out;
 868 }
 869
 870 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 871                               const struct bkey_i *k, unsigned level)
 872 {
 873         struct btree_cache *bc = &c->btree_cache;
 874         struct btree *b;
 875
 876         BUG_ON(!btree_node_locked(iter, level + 1));
 877         BUG_ON(level >= BTREE_MAX_DEPTH);
 878
 879         rcu_read_lock();
 880         b = btree_cache_find(bc, k);
 881         rcu_read_unlock();
 882
 883         if (b)
 884                 return;
 885
 886         bch2_btree_node_fill(c, iter, k, level, SIX_LOCK_read, false);
 887 }
 888
 889 int bch2_print_btree_node(struct bch_fs *c, struct btree *b,
 890                           char *buf, size_t len)
 891 {
 892         const struct bkey_format *f = &b->format;
 893         struct bset_stats stats;
 894         char ptrs[100];
 895
 896         memset(&stats, 0, sizeof(stats));
 897
 898         bch2_val_to_text(c, BKEY_TYPE_BTREE, ptrs, sizeof(ptrs),
 899                         bkey_i_to_s_c(&b->key));
 900         bch2_btree_keys_stats(b, &stats);
 901
 902         return scnprintf(buf, len,
 903                          "l %u %llu:%llu - %llu:%llu:\n"
 904                          "    ptrs: %s\n"
 905                          "    format: u64s %u fields %u %u %u %u %u\n"
 906                          "    unpack fn len: %u\n"
 907                          "    bytes used %zu/%zu (%zu%% full)\n"
 908                          "    sib u64s: %u, %u (merge threshold %zu)\n"
 909                          "    nr packed keys %u\n"
 910                          "    nr unpacked keys %u\n"
 911                          "    floats %zu\n"
 912                          "    failed unpacked %zu\n"
 913                          "    failed prev %zu\n"
 914                          "    failed overflow %zu\n",
 915                          b->level,
 916                          b->data->min_key.inode,
 917                          b->data->min_key.offset,
 918                          b->data->max_key.inode,
 919                          b->data->max_key.offset,
 920                          ptrs,
 921                          f->key_u64s,
 922                          f->bits_per_field[0],
 923                          f->bits_per_field[1],
 924                          f->bits_per_field[2],
 925                          f->bits_per_field[3],
 926                          f->bits_per_field[4],
 927                          b->unpack_fn_len,
 928                          b->nr.live_u64s * sizeof(u64),
 929                          btree_bytes(c) - sizeof(struct btree_node),
 930                          b->nr.live_u64s * 100 / btree_max_u64s(c),
 931                          b->sib_u64s[0],
 932                          b->sib_u64s[1],
 933                          BTREE_FOREGROUND_MERGE_THRESHOLD(c),
 934                          b->nr.packed_keys,
 935                          b->nr.unpacked_keys,
 936                          stats.floats,
 937                          stats.failed_unpacked,
 938                          stats.failed_prev,
 939                          stats.failed_overflow);
 940 }