xref: /linux/fs/bcachefs/btree_cache.c (revision b4db9f840283caca0d904436f187ef56a9126eaa)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "bcachefs.h"
4 #include "bbpos.h"
5 #include "bkey_buf.h"
6 #include "btree_cache.h"
7 #include "btree_io.h"
8 #include "btree_iter.h"
9 #include "btree_locking.h"
10 #include "debug.h"
11 #include "errcode.h"
12 #include "error.h"
13 #include "journal.h"
14 #include "trace.h"
15 
16 #include <linux/prefetch.h>
17 #include <linux/sched/mm.h>
18 
19 const char * const bch2_btree_node_flags[] = {
20 #define x(f)	#f,
21 	BTREE_FLAGS()
22 #undef x
23 	NULL
24 };
25 
26 void bch2_recalc_btree_reserve(struct bch_fs *c)
27 {
28 	unsigned i, reserve = 16;
29 
30 	if (!c->btree_roots_known[0].b)
31 		reserve += 8;
32 
33 	for (i = 0; i < btree_id_nr_alive(c); i++) {
34 		struct btree_root *r = bch2_btree_id_root(c, i);
35 
36 		if (r->b)
37 			reserve += min_t(unsigned, 1, r->b->c.level) * 8;
38 	}
39 
40 	c->btree_cache.reserve = reserve;
41 }
42 
43 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
44 {
45 	return max_t(int, 0, bc->used - bc->reserve);
46 }
47 
48 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
49 {
50 	if (b->c.lock.readers)
51 		list_move(&b->list, &bc->freed_pcpu);
52 	else
53 		list_move(&b->list, &bc->freed_nonpcpu);
54 }
55 
56 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
57 {
58 	struct btree_cache *bc = &c->btree_cache;
59 
60 	EBUG_ON(btree_node_write_in_flight(b));
61 
62 	clear_btree_node_just_written(b);
63 
64 	kvfree(b->data);
65 	b->data = NULL;
66 #ifdef __KERNEL__
67 	kvfree(b->aux_data);
68 #else
69 	munmap(b->aux_data, btree_aux_data_bytes(b));
70 #endif
71 	b->aux_data = NULL;
72 
73 	bc->used--;
74 
75 	btree_node_to_freedlist(bc, b);
76 }
77 
78 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
79 				   const void *obj)
80 {
81 	const struct btree *b = obj;
82 	const u64 *v = arg->key;
83 
84 	return b->hash_val == *v ? 0 : 1;
85 }
86 
87 static const struct rhashtable_params bch_btree_cache_params = {
88 	.head_offset	= offsetof(struct btree, hash),
89 	.key_offset	= offsetof(struct btree, hash_val),
90 	.key_len	= sizeof(u64),
91 	.obj_cmpfn	= bch2_btree_cache_cmp_fn,
92 };
93 
94 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
95 {
96 	BUG_ON(b->data || b->aux_data);
97 
98 	b->data = kvmalloc(btree_buf_bytes(b), gfp);
99 	if (!b->data)
100 		return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
101 #ifdef __KERNEL__
102 	b->aux_data = kvmalloc(btree_aux_data_bytes(b), gfp);
103 #else
104 	b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
105 			   PROT_READ|PROT_WRITE|PROT_EXEC,
106 			   MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
107 	if (b->aux_data == MAP_FAILED)
108 		b->aux_data = NULL;
109 #endif
110 	if (!b->aux_data) {
111 		kvfree(b->data);
112 		b->data = NULL;
113 		return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
114 	}
115 
116 	return 0;
117 }
118 
119 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
120 {
121 	struct btree *b;
122 
123 	b = kzalloc(sizeof(struct btree), gfp);
124 	if (!b)
125 		return NULL;
126 
127 	bkey_btree_ptr_init(&b->key);
128 	INIT_LIST_HEAD(&b->list);
129 	INIT_LIST_HEAD(&b->write_blocked);
130 	b->byte_order = ilog2(c->opts.btree_node_size);
131 	return b;
132 }
133 
134 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
135 {
136 	struct btree_cache *bc = &c->btree_cache;
137 	struct btree *b;
138 
139 	b = __btree_node_mem_alloc(c, GFP_KERNEL);
140 	if (!b)
141 		return NULL;
142 
143 	if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
144 		kfree(b);
145 		return NULL;
146 	}
147 
148 	bch2_btree_lock_init(&b->c, 0);
149 
150 	bc->used++;
151 	list_add(&b->list, &bc->freeable);
152 	return b;
153 }
154 
155 /* Btree in memory cache - hash table */
156 
157 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
158 {
159 	int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
160 
161 	BUG_ON(ret);
162 
163 	/* Cause future lookups for this node to fail: */
164 	b->hash_val = 0;
165 }
166 
167 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
168 {
169 	BUG_ON(b->hash_val);
170 	b->hash_val = btree_ptr_hash_val(&b->key);
171 
172 	return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
173 					     bch_btree_cache_params);
174 }
175 
176 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
177 				unsigned level, enum btree_id id)
178 {
179 	int ret;
180 
181 	b->c.level	= level;
182 	b->c.btree_id	= id;
183 
184 	mutex_lock(&bc->lock);
185 	ret = __bch2_btree_node_hash_insert(bc, b);
186 	if (!ret)
187 		list_add_tail(&b->list, &bc->live);
188 	mutex_unlock(&bc->lock);
189 
190 	return ret;
191 }
192 
193 __flatten
194 static inline struct btree *btree_cache_find(struct btree_cache *bc,
195 				     const struct bkey_i *k)
196 {
197 	u64 v = btree_ptr_hash_val(k);
198 
199 	return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
200 }
201 
202 /*
203  * this version is for btree nodes that have already been freed (we're not
204  * reaping a real btree node)
205  */
206 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
207 {
208 	struct btree_cache *bc = &c->btree_cache;
209 	int ret = 0;
210 
211 	lockdep_assert_held(&bc->lock);
212 
213 	struct bbpos pos = BBPOS(b->c.btree_id, b->key.k.p);
214 
215 	u64 mask = b->c.level
216 		? bc->pinned_nodes_interior_mask
217 		: bc->pinned_nodes_leaf_mask;
218 
219 	if ((mask & BIT_ULL(b->c.btree_id)) &&
220 	    bbpos_cmp(bc->pinned_nodes_start, pos) < 0 &&
221 	    bbpos_cmp(bc->pinned_nodes_end, pos) >= 0)
222 		return -BCH_ERR_ENOMEM_btree_node_reclaim;
223 
224 wait_on_io:
225 	if (b->flags & ((1U << BTREE_NODE_dirty)|
226 			(1U << BTREE_NODE_read_in_flight)|
227 			(1U << BTREE_NODE_write_in_flight))) {
228 		if (!flush)
229 			return -BCH_ERR_ENOMEM_btree_node_reclaim;
230 
231 		/* XXX: waiting on IO with btree cache lock held */
232 		bch2_btree_node_wait_on_read(b);
233 		bch2_btree_node_wait_on_write(b);
234 	}
235 
236 	if (!six_trylock_intent(&b->c.lock))
237 		return -BCH_ERR_ENOMEM_btree_node_reclaim;
238 
239 	if (!six_trylock_write(&b->c.lock))
240 		goto out_unlock_intent;
241 
242 	/* recheck under lock */
243 	if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
244 			(1U << BTREE_NODE_write_in_flight))) {
245 		if (!flush)
246 			goto out_unlock;
247 		six_unlock_write(&b->c.lock);
248 		six_unlock_intent(&b->c.lock);
249 		goto wait_on_io;
250 	}
251 
252 	if (btree_node_noevict(b) ||
253 	    btree_node_write_blocked(b) ||
254 	    btree_node_will_make_reachable(b))
255 		goto out_unlock;
256 
257 	if (btree_node_dirty(b)) {
258 		if (!flush)
259 			goto out_unlock;
260 		/*
261 		 * Using the underscore version because we don't want to compact
262 		 * bsets after the write, since this node is about to be evicted
263 		 * - unless btree verify mode is enabled, since it runs out of
264 		 * the post write cleanup:
265 		 */
266 		if (bch2_verify_btree_ondisk)
267 			bch2_btree_node_write(c, b, SIX_LOCK_intent,
268 					      BTREE_WRITE_cache_reclaim);
269 		else
270 			__bch2_btree_node_write(c, b,
271 						BTREE_WRITE_cache_reclaim);
272 
273 		six_unlock_write(&b->c.lock);
274 		six_unlock_intent(&b->c.lock);
275 		goto wait_on_io;
276 	}
277 out:
278 	if (b->hash_val && !ret)
279 		trace_and_count(c, btree_cache_reap, c, b);
280 	return ret;
281 out_unlock:
282 	six_unlock_write(&b->c.lock);
283 out_unlock_intent:
284 	six_unlock_intent(&b->c.lock);
285 	ret = -BCH_ERR_ENOMEM_btree_node_reclaim;
286 	goto out;
287 }
288 
289 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
290 {
291 	return __btree_node_reclaim(c, b, false);
292 }
293 
294 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
295 {
296 	return __btree_node_reclaim(c, b, true);
297 }
298 
299 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
300 					   struct shrink_control *sc)
301 {
302 	struct bch_fs *c = shrink->private_data;
303 	struct btree_cache *bc = &c->btree_cache;
304 	struct btree *b, *t;
305 	unsigned long nr = sc->nr_to_scan;
306 	unsigned long can_free = 0;
307 	unsigned long freed = 0;
308 	unsigned long touched = 0;
309 	unsigned i, flags;
310 	unsigned long ret = SHRINK_STOP;
311 	bool trigger_writes = atomic_read(&bc->dirty) + nr >=
312 		bc->used * 3 / 4;
313 
314 	if (bch2_btree_shrinker_disabled)
315 		return SHRINK_STOP;
316 
317 	mutex_lock(&bc->lock);
318 	flags = memalloc_nofs_save();
319 
320 	/*
321 	 * It's _really_ critical that we don't free too many btree nodes - we
322 	 * have to always leave ourselves a reserve. The reserve is how we
323 	 * guarantee that allocating memory for a new btree node can always
324 	 * succeed, so that inserting keys into the btree can always succeed and
325 	 * IO can always make forward progress:
326 	 */
327 	can_free = btree_cache_can_free(bc);
328 	nr = min_t(unsigned long, nr, can_free);
329 
330 	i = 0;
331 	list_for_each_entry_safe(b, t, &bc->freeable, list) {
332 		/*
333 		 * Leave a few nodes on the freeable list, so that a btree split
334 		 * won't have to hit the system allocator:
335 		 */
336 		if (++i <= 3)
337 			continue;
338 
339 		touched++;
340 
341 		if (touched >= nr)
342 			goto out;
343 
344 		if (!btree_node_reclaim(c, b)) {
345 			btree_node_data_free(c, b);
346 			six_unlock_write(&b->c.lock);
347 			six_unlock_intent(&b->c.lock);
348 			freed++;
349 		}
350 	}
351 restart:
352 	list_for_each_entry_safe(b, t, &bc->live, list) {
353 		touched++;
354 
355 		if (btree_node_accessed(b)) {
356 			clear_btree_node_accessed(b);
357 		} else if (!btree_node_reclaim(c, b)) {
358 			freed++;
359 			btree_node_data_free(c, b);
360 
361 			bch2_btree_node_hash_remove(bc, b);
362 			six_unlock_write(&b->c.lock);
363 			six_unlock_intent(&b->c.lock);
364 
365 			if (freed == nr)
366 				goto out_rotate;
367 		} else if (trigger_writes &&
368 			   btree_node_dirty(b) &&
369 			   !btree_node_will_make_reachable(b) &&
370 			   !btree_node_write_blocked(b) &&
371 			   six_trylock_read(&b->c.lock)) {
372 			list_move(&bc->live, &b->list);
373 			mutex_unlock(&bc->lock);
374 			__bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
375 			six_unlock_read(&b->c.lock);
376 			if (touched >= nr)
377 				goto out_nounlock;
378 			mutex_lock(&bc->lock);
379 			goto restart;
380 		}
381 
382 		if (touched >= nr)
383 			break;
384 	}
385 out_rotate:
386 	if (&t->list != &bc->live)
387 		list_move_tail(&bc->live, &t->list);
388 out:
389 	mutex_unlock(&bc->lock);
390 out_nounlock:
391 	ret = freed;
392 	memalloc_nofs_restore(flags);
393 	trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
394 	return ret;
395 }
396 
397 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
398 					    struct shrink_control *sc)
399 {
400 	struct bch_fs *c = shrink->private_data;
401 	struct btree_cache *bc = &c->btree_cache;
402 
403 	if (bch2_btree_shrinker_disabled)
404 		return 0;
405 
406 	return btree_cache_can_free(bc);
407 }
408 
409 void bch2_fs_btree_cache_exit(struct bch_fs *c)
410 {
411 	struct btree_cache *bc = &c->btree_cache;
412 	struct btree *b;
413 	unsigned i, flags;
414 
415 	shrinker_free(bc->shrink);
416 
417 	/* vfree() can allocate memory: */
418 	flags = memalloc_nofs_save();
419 	mutex_lock(&bc->lock);
420 
421 	if (c->verify_data)
422 		list_move(&c->verify_data->list, &bc->live);
423 
424 	kvfree(c->verify_ondisk);
425 
426 	for (i = 0; i < btree_id_nr_alive(c); i++) {
427 		struct btree_root *r = bch2_btree_id_root(c, i);
428 
429 		if (r->b)
430 			list_add(&r->b->list, &bc->live);
431 	}
432 
433 	list_splice(&bc->freeable, &bc->live);
434 
435 	while (!list_empty(&bc->live)) {
436 		b = list_first_entry(&bc->live, struct btree, list);
437 
438 		BUG_ON(btree_node_read_in_flight(b) ||
439 		       btree_node_write_in_flight(b));
440 
441 		btree_node_data_free(c, b);
442 	}
443 
444 	BUG_ON(!bch2_journal_error(&c->journal) &&
445 	       atomic_read(&c->btree_cache.dirty));
446 
447 	list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
448 
449 	while (!list_empty(&bc->freed_nonpcpu)) {
450 		b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
451 		list_del(&b->list);
452 		six_lock_exit(&b->c.lock);
453 		kfree(b);
454 	}
455 
456 	mutex_unlock(&bc->lock);
457 	memalloc_nofs_restore(flags);
458 
459 	if (bc->table_init_done)
460 		rhashtable_destroy(&bc->table);
461 }
462 
463 int bch2_fs_btree_cache_init(struct bch_fs *c)
464 {
465 	struct btree_cache *bc = &c->btree_cache;
466 	struct shrinker *shrink;
467 	unsigned i;
468 	int ret = 0;
469 
470 	ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
471 	if (ret)
472 		goto err;
473 
474 	bc->table_init_done = true;
475 
476 	bch2_recalc_btree_reserve(c);
477 
478 	for (i = 0; i < bc->reserve; i++)
479 		if (!__bch2_btree_node_mem_alloc(c))
480 			goto err;
481 
482 	list_splice_init(&bc->live, &bc->freeable);
483 
484 	mutex_init(&c->verify_lock);
485 
486 	shrink = shrinker_alloc(0, "%s-btree_cache", c->name);
487 	if (!shrink)
488 		goto err;
489 	bc->shrink = shrink;
490 	shrink->count_objects	= bch2_btree_cache_count;
491 	shrink->scan_objects	= bch2_btree_cache_scan;
492 	shrink->seeks		= 4;
493 	shrink->private_data	= c;
494 	shrinker_register(shrink);
495 
496 	return 0;
497 err:
498 	return -BCH_ERR_ENOMEM_fs_btree_cache_init;
499 }
500 
501 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
502 {
503 	mutex_init(&bc->lock);
504 	INIT_LIST_HEAD(&bc->live);
505 	INIT_LIST_HEAD(&bc->freeable);
506 	INIT_LIST_HEAD(&bc->freed_pcpu);
507 	INIT_LIST_HEAD(&bc->freed_nonpcpu);
508 }
509 
510 /*
511  * We can only have one thread cannibalizing other cached btree nodes at a time,
512  * or we'll deadlock. We use an open coded mutex to ensure that, which a
513  * cannibalize_bucket() will take. This means every time we unlock the root of
514  * the btree, we need to release this lock if we have it held.
515  */
516 void bch2_btree_cache_cannibalize_unlock(struct btree_trans *trans)
517 {
518 	struct bch_fs *c = trans->c;
519 	struct btree_cache *bc = &c->btree_cache;
520 
521 	if (bc->alloc_lock == current) {
522 		trace_and_count(c, btree_cache_cannibalize_unlock, trans);
523 		bc->alloc_lock = NULL;
524 		closure_wake_up(&bc->alloc_wait);
525 	}
526 }
527 
528 int bch2_btree_cache_cannibalize_lock(struct btree_trans *trans, struct closure *cl)
529 {
530 	struct bch_fs *c = trans->c;
531 	struct btree_cache *bc = &c->btree_cache;
532 	struct task_struct *old;
533 
534 	old = cmpxchg(&bc->alloc_lock, NULL, current);
535 	if (old == NULL || old == current)
536 		goto success;
537 
538 	if (!cl) {
539 		trace_and_count(c, btree_cache_cannibalize_lock_fail, trans);
540 		return -BCH_ERR_ENOMEM_btree_cache_cannibalize_lock;
541 	}
542 
543 	closure_wait(&bc->alloc_wait, cl);
544 
545 	/* Try again, after adding ourselves to waitlist */
546 	old = cmpxchg(&bc->alloc_lock, NULL, current);
547 	if (old == NULL || old == current) {
548 		/* We raced */
549 		closure_wake_up(&bc->alloc_wait);
550 		goto success;
551 	}
552 
553 	trace_and_count(c, btree_cache_cannibalize_lock_fail, trans);
554 	return -BCH_ERR_btree_cache_cannibalize_lock_blocked;
555 
556 success:
557 	trace_and_count(c, btree_cache_cannibalize_lock, trans);
558 	return 0;
559 }
560 
561 static struct btree *btree_node_cannibalize(struct bch_fs *c)
562 {
563 	struct btree_cache *bc = &c->btree_cache;
564 	struct btree *b;
565 
566 	list_for_each_entry_reverse(b, &bc->live, list)
567 		if (!btree_node_reclaim(c, b))
568 			return b;
569 
570 	while (1) {
571 		list_for_each_entry_reverse(b, &bc->live, list)
572 			if (!btree_node_write_and_reclaim(c, b))
573 				return b;
574 
575 		/*
576 		 * Rare case: all nodes were intent-locked.
577 		 * Just busy-wait.
578 		 */
579 		WARN_ONCE(1, "btree cache cannibalize failed\n");
580 		cond_resched();
581 	}
582 }
583 
584 struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_read_locks)
585 {
586 	struct bch_fs *c = trans->c;
587 	struct btree_cache *bc = &c->btree_cache;
588 	struct list_head *freed = pcpu_read_locks
589 		? &bc->freed_pcpu
590 		: &bc->freed_nonpcpu;
591 	struct btree *b, *b2;
592 	u64 start_time = local_clock();
593 	unsigned flags;
594 
595 	flags = memalloc_nofs_save();
596 	mutex_lock(&bc->lock);
597 
598 	/*
599 	 * We never free struct btree itself, just the memory that holds the on
600 	 * disk node. Check the freed list before allocating a new one:
601 	 */
602 	list_for_each_entry(b, freed, list)
603 		if (!btree_node_reclaim(c, b)) {
604 			list_del_init(&b->list);
605 			goto got_node;
606 		}
607 
608 	b = __btree_node_mem_alloc(c, GFP_NOWAIT|__GFP_NOWARN);
609 	if (!b) {
610 		mutex_unlock(&bc->lock);
611 		bch2_trans_unlock(trans);
612 		b = __btree_node_mem_alloc(c, GFP_KERNEL);
613 		if (!b)
614 			goto err;
615 		mutex_lock(&bc->lock);
616 	}
617 
618 	bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0);
619 
620 	BUG_ON(!six_trylock_intent(&b->c.lock));
621 	BUG_ON(!six_trylock_write(&b->c.lock));
622 got_node:
623 
624 	/*
625 	 * btree_free() doesn't free memory; it sticks the node on the end of
626 	 * the list. Check if there's any freed nodes there:
627 	 */
628 	list_for_each_entry(b2, &bc->freeable, list)
629 		if (!btree_node_reclaim(c, b2)) {
630 			swap(b->data, b2->data);
631 			swap(b->aux_data, b2->aux_data);
632 			btree_node_to_freedlist(bc, b2);
633 			six_unlock_write(&b2->c.lock);
634 			six_unlock_intent(&b2->c.lock);
635 			goto got_mem;
636 		}
637 
638 	mutex_unlock(&bc->lock);
639 
640 	if (btree_node_data_alloc(c, b, GFP_NOWAIT|__GFP_NOWARN)) {
641 		bch2_trans_unlock(trans);
642 		if (btree_node_data_alloc(c, b, GFP_KERNEL|__GFP_NOWARN))
643 			goto err;
644 	}
645 
646 	mutex_lock(&bc->lock);
647 	bc->used++;
648 got_mem:
649 	mutex_unlock(&bc->lock);
650 
651 	BUG_ON(btree_node_hashed(b));
652 	BUG_ON(btree_node_dirty(b));
653 	BUG_ON(btree_node_write_in_flight(b));
654 out:
655 	b->flags		= 0;
656 	b->written		= 0;
657 	b->nsets		= 0;
658 	b->sib_u64s[0]		= 0;
659 	b->sib_u64s[1]		= 0;
660 	b->whiteout_u64s	= 0;
661 	bch2_btree_keys_init(b);
662 	set_btree_node_accessed(b);
663 
664 	bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
665 			       start_time);
666 
667 	memalloc_nofs_restore(flags);
668 	return b;
669 err:
670 	mutex_lock(&bc->lock);
671 
672 	/* Try to cannibalize another cached btree node: */
673 	if (bc->alloc_lock == current) {
674 		b2 = btree_node_cannibalize(c);
675 		clear_btree_node_just_written(b2);
676 		bch2_btree_node_hash_remove(bc, b2);
677 
678 		if (b) {
679 			swap(b->data, b2->data);
680 			swap(b->aux_data, b2->aux_data);
681 			btree_node_to_freedlist(bc, b2);
682 			six_unlock_write(&b2->c.lock);
683 			six_unlock_intent(&b2->c.lock);
684 		} else {
685 			b = b2;
686 			list_del_init(&b->list);
687 		}
688 
689 		mutex_unlock(&bc->lock);
690 
691 		trace_and_count(c, btree_cache_cannibalize, trans);
692 		goto out;
693 	}
694 
695 	mutex_unlock(&bc->lock);
696 	memalloc_nofs_restore(flags);
697 	return ERR_PTR(-BCH_ERR_ENOMEM_btree_node_mem_alloc);
698 }
699 
700 /* Slowpath, don't want it inlined into btree_iter_traverse() */
701 static noinline struct btree *bch2_btree_node_fill(struct btree_trans *trans,
702 				struct btree_path *path,
703 				const struct bkey_i *k,
704 				enum btree_id btree_id,
705 				unsigned level,
706 				enum six_lock_type lock_type,
707 				bool sync)
708 {
709 	struct bch_fs *c = trans->c;
710 	struct btree_cache *bc = &c->btree_cache;
711 	struct btree *b;
712 	u32 seq;
713 
714 	BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
715 	/*
716 	 * Parent node must be locked, else we could read in a btree node that's
717 	 * been freed:
718 	 */
719 	if (path && !bch2_btree_node_relock(trans, path, level + 1)) {
720 		trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
721 		return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
722 	}
723 
724 	b = bch2_btree_node_mem_alloc(trans, level != 0);
725 
726 	if (bch2_err_matches(PTR_ERR_OR_ZERO(b), ENOMEM)) {
727 		if (!path)
728 			return b;
729 
730 		trans->memory_allocation_failure = true;
731 		trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
732 		return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
733 	}
734 
735 	if (IS_ERR(b))
736 		return b;
737 
738 	bkey_copy(&b->key, k);
739 	if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
740 		/* raced with another fill: */
741 
742 		/* mark as unhashed... */
743 		b->hash_val = 0;
744 
745 		mutex_lock(&bc->lock);
746 		list_add(&b->list, &bc->freeable);
747 		mutex_unlock(&bc->lock);
748 
749 		six_unlock_write(&b->c.lock);
750 		six_unlock_intent(&b->c.lock);
751 		return NULL;
752 	}
753 
754 	set_btree_node_read_in_flight(b);
755 
756 	six_unlock_write(&b->c.lock);
757 	seq = six_lock_seq(&b->c.lock);
758 	six_unlock_intent(&b->c.lock);
759 
760 	/* Unlock before doing IO: */
761 	if (path && sync)
762 		bch2_trans_unlock_noassert(trans);
763 
764 	bch2_btree_node_read(trans, b, sync);
765 
766 	if (!sync)
767 		return NULL;
768 
769 	if (path) {
770 		int ret = bch2_trans_relock(trans) ?:
771 			bch2_btree_path_relock_intent(trans, path);
772 		if (ret) {
773 			BUG_ON(!trans->restarted);
774 			return ERR_PTR(ret);
775 		}
776 	}
777 
778 	if (!six_relock_type(&b->c.lock, lock_type, seq)) {
779 		BUG_ON(!path);
780 
781 		trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
782 		return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
783 	}
784 
785 	return b;
786 }
787 
788 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
789 {
790 	struct printbuf buf = PRINTBUF;
791 
792 	if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_allocations)
793 		return;
794 
795 	prt_printf(&buf,
796 	       "btree node header doesn't match ptr\n"
797 	       "btree %s level %u\n"
798 	       "ptr: ",
799 	       bch2_btree_id_str(b->c.btree_id), b->c.level);
800 	bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
801 
802 	prt_printf(&buf, "\nheader: btree %s level %llu\n"
803 	       "min ",
804 	       bch2_btree_id_str(BTREE_NODE_ID(b->data)),
805 	       BTREE_NODE_LEVEL(b->data));
806 	bch2_bpos_to_text(&buf, b->data->min_key);
807 
808 	prt_printf(&buf, "\nmax ");
809 	bch2_bpos_to_text(&buf, b->data->max_key);
810 
811 	bch2_fs_topology_error(c, "%s", buf.buf);
812 
813 	printbuf_exit(&buf);
814 }
815 
816 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
817 {
818 	if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
819 	    b->c.level != BTREE_NODE_LEVEL(b->data) ||
820 	    !bpos_eq(b->data->max_key, b->key.k.p) ||
821 	    (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
822 	     !bpos_eq(b->data->min_key,
823 		      bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
824 		btree_bad_header(c, b);
825 }
826 
827 static struct btree *__bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
828 					   const struct bkey_i *k, unsigned level,
829 					   enum six_lock_type lock_type,
830 					   unsigned long trace_ip)
831 {
832 	struct bch_fs *c = trans->c;
833 	struct btree_cache *bc = &c->btree_cache;
834 	struct btree *b;
835 	struct bset_tree *t;
836 	bool need_relock = false;
837 	int ret;
838 
839 	EBUG_ON(level >= BTREE_MAX_DEPTH);
840 retry:
841 	b = btree_cache_find(bc, k);
842 	if (unlikely(!b)) {
843 		/*
844 		 * We must have the parent locked to call bch2_btree_node_fill(),
845 		 * else we could read in a btree node from disk that's been
846 		 * freed:
847 		 */
848 		b = bch2_btree_node_fill(trans, path, k, path->btree_id,
849 					 level, lock_type, true);
850 		need_relock = true;
851 
852 		/* We raced and found the btree node in the cache */
853 		if (!b)
854 			goto retry;
855 
856 		if (IS_ERR(b))
857 			return b;
858 	} else {
859 		if (btree_node_read_locked(path, level + 1))
860 			btree_node_unlock(trans, path, level + 1);
861 
862 		ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
863 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
864 			return ERR_PTR(ret);
865 
866 		BUG_ON(ret);
867 
868 		if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
869 			     b->c.level != level ||
870 			     race_fault())) {
871 			six_unlock_type(&b->c.lock, lock_type);
872 			if (bch2_btree_node_relock(trans, path, level + 1))
873 				goto retry;
874 
875 			trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
876 			return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
877 		}
878 
879 		/* avoid atomic set bit if it's not needed: */
880 		if (!btree_node_accessed(b))
881 			set_btree_node_accessed(b);
882 	}
883 
884 	if (unlikely(btree_node_read_in_flight(b))) {
885 		u32 seq = six_lock_seq(&b->c.lock);
886 
887 		six_unlock_type(&b->c.lock, lock_type);
888 		bch2_trans_unlock(trans);
889 		need_relock = true;
890 
891 		bch2_btree_node_wait_on_read(b);
892 
893 		/*
894 		 * should_be_locked is not set on this path yet, so we need to
895 		 * relock it specifically:
896 		 */
897 		if (!six_relock_type(&b->c.lock, lock_type, seq))
898 			goto retry;
899 	}
900 
901 	if (unlikely(need_relock)) {
902 		ret = bch2_trans_relock(trans) ?:
903 			bch2_btree_path_relock_intent(trans, path);
904 		if (ret) {
905 			six_unlock_type(&b->c.lock, lock_type);
906 			return ERR_PTR(ret);
907 		}
908 	}
909 
910 	prefetch(b->aux_data);
911 
912 	for_each_bset(b, t) {
913 		void *p = (u64 *) b->aux_data + t->aux_data_offset;
914 
915 		prefetch(p + L1_CACHE_BYTES * 0);
916 		prefetch(p + L1_CACHE_BYTES * 1);
917 		prefetch(p + L1_CACHE_BYTES * 2);
918 	}
919 
920 	if (unlikely(btree_node_read_error(b))) {
921 		six_unlock_type(&b->c.lock, lock_type);
922 		return ERR_PTR(-BCH_ERR_btree_node_read_error);
923 	}
924 
925 	EBUG_ON(b->c.btree_id != path->btree_id);
926 	EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
927 	btree_check_header(c, b);
928 
929 	return b;
930 }
931 
932 /**
933  * bch2_btree_node_get - find a btree node in the cache and lock it, reading it
934  * in from disk if necessary.
935  *
936  * @trans:	btree transaction object
937  * @path:	btree_path being traversed
938  * @k:		pointer to btree node (generally KEY_TYPE_btree_ptr_v2)
939  * @level:	level of btree node being looked up (0 == leaf node)
940  * @lock_type:	SIX_LOCK_read or SIX_LOCK_intent
941  * @trace_ip:	ip of caller of btree iterator code (i.e. caller of bch2_btree_iter_peek())
942  *
943  * The btree node will have either a read or a write lock held, depending on
944  * the @write parameter.
945  *
946  * Returns: btree node or ERR_PTR()
947  */
948 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
949 				  const struct bkey_i *k, unsigned level,
950 				  enum six_lock_type lock_type,
951 				  unsigned long trace_ip)
952 {
953 	struct bch_fs *c = trans->c;
954 	struct btree *b;
955 	struct bset_tree *t;
956 	int ret;
957 
958 	EBUG_ON(level >= BTREE_MAX_DEPTH);
959 
960 	b = btree_node_mem_ptr(k);
961 
962 	/*
963 	 * Check b->hash_val _before_ calling btree_node_lock() - this might not
964 	 * be the node we want anymore, and trying to lock the wrong node could
965 	 * cause an unneccessary transaction restart:
966 	 */
967 	if (unlikely(!c->opts.btree_node_mem_ptr_optimization ||
968 		     !b ||
969 		     b->hash_val != btree_ptr_hash_val(k)))
970 		return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
971 
972 	if (btree_node_read_locked(path, level + 1))
973 		btree_node_unlock(trans, path, level + 1);
974 
975 	ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
976 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
977 		return ERR_PTR(ret);
978 
979 	BUG_ON(ret);
980 
981 	if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
982 		     b->c.level != level ||
983 		     race_fault())) {
984 		six_unlock_type(&b->c.lock, lock_type);
985 		if (bch2_btree_node_relock(trans, path, level + 1))
986 			return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
987 
988 		trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
989 		return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
990 	}
991 
992 	if (unlikely(btree_node_read_in_flight(b))) {
993 		six_unlock_type(&b->c.lock, lock_type);
994 		return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
995 	}
996 
997 	prefetch(b->aux_data);
998 
999 	for_each_bset(b, t) {
1000 		void *p = (u64 *) b->aux_data + t->aux_data_offset;
1001 
1002 		prefetch(p + L1_CACHE_BYTES * 0);
1003 		prefetch(p + L1_CACHE_BYTES * 1);
1004 		prefetch(p + L1_CACHE_BYTES * 2);
1005 	}
1006 
1007 	/* avoid atomic set bit if it's not needed: */
1008 	if (!btree_node_accessed(b))
1009 		set_btree_node_accessed(b);
1010 
1011 	if (unlikely(btree_node_read_error(b))) {
1012 		six_unlock_type(&b->c.lock, lock_type);
1013 		return ERR_PTR(-BCH_ERR_btree_node_read_error);
1014 	}
1015 
1016 	EBUG_ON(b->c.btree_id != path->btree_id);
1017 	EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1018 	btree_check_header(c, b);
1019 
1020 	return b;
1021 }
1022 
1023 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
1024 					 const struct bkey_i *k,
1025 					 enum btree_id btree_id,
1026 					 unsigned level,
1027 					 bool nofill)
1028 {
1029 	struct bch_fs *c = trans->c;
1030 	struct btree_cache *bc = &c->btree_cache;
1031 	struct btree *b;
1032 	struct bset_tree *t;
1033 	int ret;
1034 
1035 	EBUG_ON(level >= BTREE_MAX_DEPTH);
1036 
1037 	if (c->opts.btree_node_mem_ptr_optimization) {
1038 		b = btree_node_mem_ptr(k);
1039 		if (b)
1040 			goto lock_node;
1041 	}
1042 retry:
1043 	b = btree_cache_find(bc, k);
1044 	if (unlikely(!b)) {
1045 		if (nofill)
1046 			goto out;
1047 
1048 		b = bch2_btree_node_fill(trans, NULL, k, btree_id,
1049 					 level, SIX_LOCK_read, true);
1050 
1051 		/* We raced and found the btree node in the cache */
1052 		if (!b)
1053 			goto retry;
1054 
1055 		if (IS_ERR(b) &&
1056 		    !bch2_btree_cache_cannibalize_lock(trans, NULL))
1057 			goto retry;
1058 
1059 		if (IS_ERR(b))
1060 			goto out;
1061 	} else {
1062 lock_node:
1063 		ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read, _THIS_IP_);
1064 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1065 			return ERR_PTR(ret);
1066 
1067 		BUG_ON(ret);
1068 
1069 		if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1070 			     b->c.btree_id != btree_id ||
1071 			     b->c.level != level)) {
1072 			six_unlock_read(&b->c.lock);
1073 			goto retry;
1074 		}
1075 	}
1076 
1077 	/* XXX: waiting on IO with btree locks held: */
1078 	__bch2_btree_node_wait_on_read(b);
1079 
1080 	prefetch(b->aux_data);
1081 
1082 	for_each_bset(b, t) {
1083 		void *p = (u64 *) b->aux_data + t->aux_data_offset;
1084 
1085 		prefetch(p + L1_CACHE_BYTES * 0);
1086 		prefetch(p + L1_CACHE_BYTES * 1);
1087 		prefetch(p + L1_CACHE_BYTES * 2);
1088 	}
1089 
1090 	/* avoid atomic set bit if it's not needed: */
1091 	if (!btree_node_accessed(b))
1092 		set_btree_node_accessed(b);
1093 
1094 	if (unlikely(btree_node_read_error(b))) {
1095 		six_unlock_read(&b->c.lock);
1096 		b = ERR_PTR(-BCH_ERR_btree_node_read_error);
1097 		goto out;
1098 	}
1099 
1100 	EBUG_ON(b->c.btree_id != btree_id);
1101 	EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1102 	btree_check_header(c, b);
1103 out:
1104 	bch2_btree_cache_cannibalize_unlock(trans);
1105 	return b;
1106 }
1107 
1108 int bch2_btree_node_prefetch(struct btree_trans *trans,
1109 			     struct btree_path *path,
1110 			     const struct bkey_i *k,
1111 			     enum btree_id btree_id, unsigned level)
1112 {
1113 	struct bch_fs *c = trans->c;
1114 	struct btree_cache *bc = &c->btree_cache;
1115 	struct btree *b;
1116 
1117 	BUG_ON(path && !btree_node_locked(path, level + 1));
1118 	BUG_ON(level >= BTREE_MAX_DEPTH);
1119 
1120 	b = btree_cache_find(bc, k);
1121 	if (b)
1122 		return 0;
1123 
1124 	b = bch2_btree_node_fill(trans, path, k, btree_id,
1125 				 level, SIX_LOCK_read, false);
1126 	return PTR_ERR_OR_ZERO(b);
1127 }
1128 
1129 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1130 {
1131 	struct bch_fs *c = trans->c;
1132 	struct btree_cache *bc = &c->btree_cache;
1133 	struct btree *b;
1134 
1135 	b = btree_cache_find(bc, k);
1136 	if (!b)
1137 		return;
1138 
1139 	BUG_ON(b == btree_node_root(trans->c, b));
1140 wait_on_io:
1141 	/* not allowed to wait on io with btree locks held: */
1142 
1143 	/* XXX we're called from btree_gc which will be holding other btree
1144 	 * nodes locked
1145 	 */
1146 	__bch2_btree_node_wait_on_read(b);
1147 	__bch2_btree_node_wait_on_write(b);
1148 
1149 	btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1150 	btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1151 
1152 	if (btree_node_dirty(b)) {
1153 		__bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
1154 		six_unlock_write(&b->c.lock);
1155 		six_unlock_intent(&b->c.lock);
1156 		goto wait_on_io;
1157 	}
1158 
1159 	BUG_ON(btree_node_dirty(b));
1160 
1161 	mutex_lock(&bc->lock);
1162 	btree_node_data_free(c, b);
1163 	bch2_btree_node_hash_remove(bc, b);
1164 	mutex_unlock(&bc->lock);
1165 
1166 	six_unlock_write(&b->c.lock);
1167 	six_unlock_intent(&b->c.lock);
1168 }
1169 
1170 const char *bch2_btree_id_str(enum btree_id btree)
1171 {
1172 	return btree < BTREE_ID_NR ? __bch2_btree_ids[btree] : "(unknown)";
1173 }
1174 
1175 void bch2_btree_pos_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
1176 {
1177 	prt_printf(out, "%s level %u/%u\n  ",
1178 	       bch2_btree_id_str(b->c.btree_id),
1179 	       b->c.level,
1180 	       bch2_btree_id_root(c, b->c.btree_id)->level);
1181 	bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1182 }
1183 
1184 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
1185 {
1186 	struct bset_stats stats;
1187 
1188 	memset(&stats, 0, sizeof(stats));
1189 
1190 	bch2_btree_keys_stats(b, &stats);
1191 
1192 	prt_printf(out, "l %u ", b->c.level);
1193 	bch2_bpos_to_text(out, b->data->min_key);
1194 	prt_printf(out, " - ");
1195 	bch2_bpos_to_text(out, b->data->max_key);
1196 	prt_printf(out, ":\n"
1197 	       "    ptrs: ");
1198 	bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1199 	prt_newline(out);
1200 
1201 	prt_printf(out,
1202 	       "    format: ");
1203 	bch2_bkey_format_to_text(out, &b->format);
1204 
1205 	prt_printf(out,
1206 	       "    unpack fn len: %u\n"
1207 	       "    bytes used %zu/%zu (%zu%% full)\n"
1208 	       "    sib u64s: %u, %u (merge threshold %u)\n"
1209 	       "    nr packed keys %u\n"
1210 	       "    nr unpacked keys %u\n"
1211 	       "    floats %zu\n"
1212 	       "    failed unpacked %zu\n",
1213 	       b->unpack_fn_len,
1214 	       b->nr.live_u64s * sizeof(u64),
1215 	       btree_buf_bytes(b) - sizeof(struct btree_node),
1216 	       b->nr.live_u64s * 100 / btree_max_u64s(c),
1217 	       b->sib_u64s[0],
1218 	       b->sib_u64s[1],
1219 	       c->btree_foreground_merge_threshold,
1220 	       b->nr.packed_keys,
1221 	       b->nr.unpacked_keys,
1222 	       stats.floats,
1223 	       stats.failed);
1224 }
1225 
1226 void bch2_btree_cache_to_text(struct printbuf *out, const struct bch_fs *c)
1227 {
1228 	prt_printf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1229 	prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1230 	prt_printf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);
1231 }
1232