xref: /linux/fs/bcachefs/btree_key_cache.c (revision dd83757f6e686a2188997cb58b5975f744bb7786)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "bcachefs.h"
4 #include "btree_cache.h"
5 #include "btree_iter.h"
6 #include "btree_key_cache.h"
7 #include "btree_locking.h"
8 #include "btree_update.h"
9 #include "errcode.h"
10 #include "error.h"
11 #include "journal.h"
12 #include "journal_reclaim.h"
13 #include "trace.h"
14 
15 #include <linux/sched/mm.h>
16 
btree_uses_pcpu_readers(enum btree_id id)17 static inline bool btree_uses_pcpu_readers(enum btree_id id)
18 {
19 	return id == BTREE_ID_subvolumes;
20 }
21 
22 static struct kmem_cache *bch2_key_cache;
23 
bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg * arg,const void * obj)24 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
25 				       const void *obj)
26 {
27 	const struct bkey_cached *ck = obj;
28 	const struct bkey_cached_key *key = arg->key;
29 
30 	return ck->key.btree_id != key->btree_id ||
31 		!bpos_eq(ck->key.pos, key->pos);
32 }
33 
34 static const struct rhashtable_params bch2_btree_key_cache_params = {
35 	.head_offset		= offsetof(struct bkey_cached, hash),
36 	.key_offset		= offsetof(struct bkey_cached, key),
37 	.key_len		= sizeof(struct bkey_cached_key),
38 	.obj_cmpfn		= bch2_btree_key_cache_cmp_fn,
39 	.automatic_shrinking	= true,
40 };
41 
btree_path_cached_set(struct btree_trans * trans,struct btree_path * path,struct bkey_cached * ck,enum btree_node_locked_type lock_held)42 static inline void btree_path_cached_set(struct btree_trans *trans, struct btree_path *path,
43 					 struct bkey_cached *ck,
44 					 enum btree_node_locked_type lock_held)
45 {
46 	path->l[0].lock_seq	= six_lock_seq(&ck->c.lock);
47 	path->l[0].b		= (void *) ck;
48 	mark_btree_node_locked(trans, path, 0, lock_held);
49 }
50 
51 __flatten
52 inline struct bkey_cached *
bch2_btree_key_cache_find(struct bch_fs * c,enum btree_id btree_id,struct bpos pos)53 bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
54 {
55 	struct bkey_cached_key key = {
56 		.btree_id	= btree_id,
57 		.pos		= pos,
58 	};
59 
60 	return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
61 				      bch2_btree_key_cache_params);
62 }
63 
bkey_cached_lock_for_evict(struct bkey_cached * ck)64 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
65 {
66 	if (!six_trylock_intent(&ck->c.lock))
67 		return false;
68 
69 	if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
70 		six_unlock_intent(&ck->c.lock);
71 		return false;
72 	}
73 
74 	if (!six_trylock_write(&ck->c.lock)) {
75 		six_unlock_intent(&ck->c.lock);
76 		return false;
77 	}
78 
79 	return true;
80 }
81 
bkey_cached_evict(struct btree_key_cache * c,struct bkey_cached * ck)82 static bool bkey_cached_evict(struct btree_key_cache *c,
83 			      struct bkey_cached *ck)
84 {
85 	bool ret = !rhashtable_remove_fast(&c->table, &ck->hash,
86 				      bch2_btree_key_cache_params);
87 	if (ret) {
88 		memset(&ck->key, ~0, sizeof(ck->key));
89 		atomic_long_dec(&c->nr_keys);
90 	}
91 
92 	return ret;
93 }
94 
__bkey_cached_free(struct rcu_pending * pending,struct rcu_head * rcu)95 static void __bkey_cached_free(struct rcu_pending *pending, struct rcu_head *rcu)
96 {
97 	struct bch_fs *c = container_of(pending->srcu, struct bch_fs, btree_trans_barrier);
98 	struct bkey_cached *ck = container_of(rcu, struct bkey_cached, rcu);
99 
100 	this_cpu_dec(*c->btree_key_cache.nr_pending);
101 	kmem_cache_free(bch2_key_cache, ck);
102 }
103 
bkey_cached_free(struct btree_key_cache * bc,struct bkey_cached * ck)104 static void bkey_cached_free(struct btree_key_cache *bc,
105 			     struct bkey_cached *ck)
106 {
107 	kfree(ck->k);
108 	ck->k		= NULL;
109 	ck->u64s	= 0;
110 
111 	six_unlock_write(&ck->c.lock);
112 	six_unlock_intent(&ck->c.lock);
113 
114 	bool pcpu_readers = ck->c.lock.readers != NULL;
115 	rcu_pending_enqueue(&bc->pending[pcpu_readers], &ck->rcu);
116 	this_cpu_inc(*bc->nr_pending);
117 }
118 
__bkey_cached_alloc(unsigned key_u64s,gfp_t gfp)119 static struct bkey_cached *__bkey_cached_alloc(unsigned key_u64s, gfp_t gfp)
120 {
121 	gfp |= __GFP_ACCOUNT|__GFP_RECLAIMABLE;
122 
123 	struct bkey_cached *ck = kmem_cache_zalloc(bch2_key_cache, gfp);
124 	if (unlikely(!ck))
125 		return NULL;
126 	ck->k = kmalloc(key_u64s * sizeof(u64), gfp);
127 	if (unlikely(!ck->k)) {
128 		kmem_cache_free(bch2_key_cache, ck);
129 		return NULL;
130 	}
131 	ck->u64s = key_u64s;
132 	return ck;
133 }
134 
135 static struct bkey_cached *
bkey_cached_alloc(struct btree_trans * trans,struct btree_path * path,unsigned key_u64s)136 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path, unsigned key_u64s)
137 {
138 	struct bch_fs *c = trans->c;
139 	struct btree_key_cache *bc = &c->btree_key_cache;
140 	bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
141 	int ret;
142 
143 	struct bkey_cached *ck = container_of_or_null(
144 				rcu_pending_dequeue(&bc->pending[pcpu_readers]),
145 				struct bkey_cached, rcu);
146 	if (ck)
147 		goto lock;
148 
149 	ck = allocate_dropping_locks(trans, ret,
150 				     __bkey_cached_alloc(key_u64s, _gfp));
151 	if (ret) {
152 		if (ck)
153 			kfree(ck->k);
154 		kmem_cache_free(bch2_key_cache, ck);
155 		return ERR_PTR(ret);
156 	}
157 
158 	if (ck) {
159 		bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0, GFP_KERNEL);
160 		ck->c.cached = true;
161 		goto lock;
162 	}
163 
164 	ck = container_of_or_null(rcu_pending_dequeue_from_all(&bc->pending[pcpu_readers]),
165 				  struct bkey_cached, rcu);
166 	if (ck)
167 		goto lock;
168 lock:
169 	six_lock_intent(&ck->c.lock, NULL, NULL);
170 	six_lock_write(&ck->c.lock, NULL, NULL);
171 	return ck;
172 }
173 
174 static struct bkey_cached *
bkey_cached_reuse(struct btree_key_cache * c)175 bkey_cached_reuse(struct btree_key_cache *c)
176 {
177 	struct bucket_table *tbl;
178 	struct rhash_head *pos;
179 	struct bkey_cached *ck;
180 	unsigned i;
181 
182 	rcu_read_lock();
183 	tbl = rht_dereference_rcu(c->table.tbl, &c->table);
184 	for (i = 0; i < tbl->size; i++)
185 		rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
186 			if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
187 			    bkey_cached_lock_for_evict(ck)) {
188 				if (bkey_cached_evict(c, ck))
189 					goto out;
190 				six_unlock_write(&ck->c.lock);
191 				six_unlock_intent(&ck->c.lock);
192 			}
193 		}
194 	ck = NULL;
195 out:
196 	rcu_read_unlock();
197 	return ck;
198 }
199 
btree_key_cache_create(struct btree_trans * trans,struct btree_path * path,struct btree_path * ck_path,struct bkey_s_c k)200 static int btree_key_cache_create(struct btree_trans *trans,
201 				  struct btree_path *path,
202 				  struct btree_path *ck_path,
203 				  struct bkey_s_c k)
204 {
205 	struct bch_fs *c = trans->c;
206 	struct btree_key_cache *bc = &c->btree_key_cache;
207 
208 	/*
209 	 * bch2_varint_decode can read past the end of the buffer by at
210 	 * most 7 bytes (it won't be used):
211 	 */
212 	unsigned key_u64s = k.k->u64s + 1;
213 
214 	/*
215 	 * Allocate some extra space so that the transaction commit path is less
216 	 * likely to have to reallocate, since that requires a transaction
217 	 * restart:
218 	 */
219 	key_u64s = min(256U, (key_u64s * 3) / 2);
220 	key_u64s = roundup_pow_of_two(key_u64s);
221 
222 	struct bkey_cached *ck = bkey_cached_alloc(trans, ck_path, key_u64s);
223 	int ret = PTR_ERR_OR_ZERO(ck);
224 	if (ret)
225 		return ret;
226 
227 	if (unlikely(!ck)) {
228 		ck = bkey_cached_reuse(bc);
229 		if (unlikely(!ck)) {
230 			bch_err(c, "error allocating memory for key cache item, btree %s",
231 				bch2_btree_id_str(ck_path->btree_id));
232 			return -BCH_ERR_ENOMEM_btree_key_cache_create;
233 		}
234 	}
235 
236 	ck->c.level		= 0;
237 	ck->c.btree_id		= ck_path->btree_id;
238 	ck->key.btree_id	= ck_path->btree_id;
239 	ck->key.pos		= ck_path->pos;
240 	ck->flags		= 1U << BKEY_CACHED_ACCESSED;
241 
242 	if (unlikely(key_u64s > ck->u64s)) {
243 		mark_btree_node_locked_noreset(ck_path, 0, BTREE_NODE_UNLOCKED);
244 
245 		struct bkey_i *new_k = allocate_dropping_locks(trans, ret,
246 				kmalloc(key_u64s * sizeof(u64), _gfp));
247 		if (unlikely(!new_k)) {
248 			bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
249 				bch2_btree_id_str(ck->key.btree_id), key_u64s);
250 			ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
251 		} else if (ret) {
252 			kfree(new_k);
253 			goto err;
254 		}
255 
256 		kfree(ck->k);
257 		ck->k = new_k;
258 		ck->u64s = key_u64s;
259 	}
260 
261 	bkey_reassemble(ck->k, k);
262 
263 	ret = bch2_btree_node_lock_write(trans, path, &path_l(path)->b->c);
264 	if (unlikely(ret))
265 		goto err;
266 
267 	ret = rhashtable_lookup_insert_fast(&bc->table, &ck->hash, bch2_btree_key_cache_params);
268 
269 	bch2_btree_node_unlock_write(trans, path, path_l(path)->b);
270 
271 	if (unlikely(ret)) /* raced with another fill? */
272 		goto err;
273 
274 	atomic_long_inc(&bc->nr_keys);
275 	six_unlock_write(&ck->c.lock);
276 
277 	enum six_lock_type lock_want = __btree_lock_want(ck_path, 0);
278 	if (lock_want == SIX_LOCK_read)
279 		six_lock_downgrade(&ck->c.lock);
280 	btree_path_cached_set(trans, ck_path, ck, (enum btree_node_locked_type) lock_want);
281 	ck_path->uptodate = BTREE_ITER_UPTODATE;
282 	return 0;
283 err:
284 	bkey_cached_free(bc, ck);
285 	mark_btree_node_locked_noreset(ck_path, 0, BTREE_NODE_UNLOCKED);
286 
287 	return ret;
288 }
289 
btree_key_cache_fill(struct btree_trans * trans,struct btree_path * ck_path,unsigned flags)290 static noinline int btree_key_cache_fill(struct btree_trans *trans,
291 					 struct btree_path *ck_path,
292 					 unsigned flags)
293 {
294 	if (flags & BTREE_ITER_cached_nofill) {
295 		ck_path->l[0].b = NULL;
296 		return 0;
297 	}
298 
299 	struct bch_fs *c = trans->c;
300 	struct btree_iter iter;
301 	struct bkey_s_c k;
302 	int ret;
303 
304 	bch2_trans_iter_init(trans, &iter, ck_path->btree_id, ck_path->pos,
305 			     BTREE_ITER_intent|
306 			     BTREE_ITER_key_cache_fill|
307 			     BTREE_ITER_cached_nofill);
308 	iter.flags &= ~BTREE_ITER_with_journal;
309 	k = bch2_btree_iter_peek_slot(&iter);
310 	ret = bkey_err(k);
311 	if (ret)
312 		goto err;
313 
314 	/* Recheck after btree lookup, before allocating: */
315 	ret = bch2_btree_key_cache_find(c, ck_path->btree_id, ck_path->pos) ? -EEXIST : 0;
316 	if (unlikely(ret))
317 		goto out;
318 
319 	ret = btree_key_cache_create(trans, btree_iter_path(trans, &iter), ck_path, k);
320 	if (ret)
321 		goto err;
322 
323 	if (trace_key_cache_fill_enabled()) {
324 		struct printbuf buf = PRINTBUF;
325 
326 		bch2_bpos_to_text(&buf, ck_path->pos);
327 		prt_char(&buf, ' ');
328 		bch2_bkey_val_to_text(&buf, trans->c, k);
329 		trace_key_cache_fill(trans, buf.buf);
330 		printbuf_exit(&buf);
331 	}
332 out:
333 	/* We're not likely to need this iterator again: */
334 	bch2_set_btree_iter_dontneed(&iter);
335 err:
336 	bch2_trans_iter_exit(trans, &iter);
337 	return ret;
338 }
339 
btree_path_traverse_cached_fast(struct btree_trans * trans,struct btree_path * path)340 static inline int btree_path_traverse_cached_fast(struct btree_trans *trans,
341 						  struct btree_path *path)
342 {
343 	struct bch_fs *c = trans->c;
344 	struct bkey_cached *ck;
345 retry:
346 	ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
347 	if (!ck)
348 		return -ENOENT;
349 
350 	enum six_lock_type lock_want = __btree_lock_want(path, 0);
351 
352 	int ret = btree_node_lock(trans, path, (void *) ck, 0, lock_want, _THIS_IP_);
353 	if (ret)
354 		return ret;
355 
356 	if (ck->key.btree_id != path->btree_id ||
357 	    !bpos_eq(ck->key.pos, path->pos)) {
358 		six_unlock_type(&ck->c.lock, lock_want);
359 		goto retry;
360 	}
361 
362 	if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
363 		set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
364 
365 	btree_path_cached_set(trans, path, ck, (enum btree_node_locked_type) lock_want);
366 	path->uptodate = BTREE_ITER_UPTODATE;
367 	return 0;
368 }
369 
bch2_btree_path_traverse_cached(struct btree_trans * trans,struct btree_path * path,unsigned flags)370 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
371 				    unsigned flags)
372 {
373 	EBUG_ON(path->level);
374 
375 	path->l[1].b = NULL;
376 
377 	int ret;
378 	do {
379 		ret = btree_path_traverse_cached_fast(trans, path);
380 		if (unlikely(ret == -ENOENT))
381 			ret = btree_key_cache_fill(trans, path, flags);
382 	} while (ret == -EEXIST);
383 
384 	if (unlikely(ret)) {
385 		path->uptodate = BTREE_ITER_NEED_TRAVERSE;
386 		if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
387 			btree_node_unlock(trans, path, 0);
388 			path->l[0].b = ERR_PTR(ret);
389 		}
390 	}
391 	return ret;
392 }
393 
btree_key_cache_flush_pos(struct btree_trans * trans,struct bkey_cached_key key,u64 journal_seq,unsigned commit_flags,bool evict)394 static int btree_key_cache_flush_pos(struct btree_trans *trans,
395 				     struct bkey_cached_key key,
396 				     u64 journal_seq,
397 				     unsigned commit_flags,
398 				     bool evict)
399 {
400 	struct bch_fs *c = trans->c;
401 	struct journal *j = &c->journal;
402 	struct btree_iter c_iter, b_iter;
403 	struct bkey_cached *ck = NULL;
404 	int ret;
405 
406 	bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
407 			     BTREE_ITER_slots|
408 			     BTREE_ITER_intent|
409 			     BTREE_ITER_all_snapshots);
410 	bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
411 			     BTREE_ITER_cached|
412 			     BTREE_ITER_intent);
413 	b_iter.flags &= ~BTREE_ITER_with_key_cache;
414 
415 	ret = bch2_btree_iter_traverse(&c_iter);
416 	if (ret)
417 		goto out;
418 
419 	ck = (void *) btree_iter_path(trans, &c_iter)->l[0].b;
420 	if (!ck)
421 		goto out;
422 
423 	if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
424 		if (evict)
425 			goto evict;
426 		goto out;
427 	}
428 
429 	if (journal_seq && ck->journal.seq != journal_seq)
430 		goto out;
431 
432 	trans->journal_res.seq = ck->journal.seq;
433 
434 	/*
435 	 * If we're at the end of the journal, we really want to free up space
436 	 * in the journal right away - we don't want to pin that old journal
437 	 * sequence number with a new btree node write, we want to re-journal
438 	 * the update
439 	 */
440 	if (ck->journal.seq == journal_last_seq(j))
441 		commit_flags |= BCH_WATERMARK_reclaim;
442 
443 	if (ck->journal.seq != journal_last_seq(j) ||
444 	    !test_bit(JOURNAL_space_low, &c->journal.flags))
445 		commit_flags |= BCH_TRANS_COMMIT_no_journal_res;
446 
447 	struct bkey_s_c btree_k = bch2_btree_iter_peek_slot(&b_iter);
448 	ret = bkey_err(btree_k);
449 	if (ret)
450 		goto err;
451 
452 	/* * Check that we're not violating cache coherency rules: */
453 	BUG_ON(bkey_deleted(btree_k.k));
454 
455 	ret   = bch2_trans_update(trans, &b_iter, ck->k,
456 				  BTREE_UPDATE_key_cache_reclaim|
457 				  BTREE_UPDATE_internal_snapshot_node|
458 				  BTREE_TRIGGER_norun) ?:
459 		bch2_trans_commit(trans, NULL, NULL,
460 				  BCH_TRANS_COMMIT_no_check_rw|
461 				  BCH_TRANS_COMMIT_no_enospc|
462 				  commit_flags);
463 err:
464 	bch2_fs_fatal_err_on(ret &&
465 			     !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
466 			     !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
467 			     !bch2_journal_error(j), c,
468 			     "flushing key cache: %s", bch2_err_str(ret));
469 	if (ret)
470 		goto out;
471 
472 	bch2_journal_pin_drop(j, &ck->journal);
473 
474 	struct btree_path *path = btree_iter_path(trans, &c_iter);
475 	BUG_ON(!btree_node_locked(path, 0));
476 
477 	if (!evict) {
478 		if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
479 			clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
480 			atomic_long_dec(&c->btree_key_cache.nr_dirty);
481 		}
482 	} else {
483 		struct btree_path *path2;
484 		unsigned i;
485 evict:
486 		trans_for_each_path(trans, path2, i)
487 			if (path2 != path)
488 				__bch2_btree_path_unlock(trans, path2);
489 
490 		bch2_btree_node_lock_write_nofail(trans, path, &ck->c);
491 
492 		if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
493 			clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
494 			atomic_long_dec(&c->btree_key_cache.nr_dirty);
495 		}
496 
497 		mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
498 		if (bkey_cached_evict(&c->btree_key_cache, ck)) {
499 			bkey_cached_free(&c->btree_key_cache, ck);
500 		} else {
501 			six_unlock_write(&ck->c.lock);
502 			six_unlock_intent(&ck->c.lock);
503 		}
504 	}
505 out:
506 	bch2_trans_iter_exit(trans, &b_iter);
507 	bch2_trans_iter_exit(trans, &c_iter);
508 	return ret;
509 }
510 
bch2_btree_key_cache_journal_flush(struct journal * j,struct journal_entry_pin * pin,u64 seq)511 int bch2_btree_key_cache_journal_flush(struct journal *j,
512 				struct journal_entry_pin *pin, u64 seq)
513 {
514 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
515 	struct bkey_cached *ck =
516 		container_of(pin, struct bkey_cached, journal);
517 	struct bkey_cached_key key;
518 	struct btree_trans *trans = bch2_trans_get(c);
519 	int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
520 	int ret = 0;
521 
522 	btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
523 	key = ck->key;
524 
525 	if (ck->journal.seq != seq ||
526 	    !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
527 		six_unlock_read(&ck->c.lock);
528 		goto unlock;
529 	}
530 
531 	if (ck->seq != seq) {
532 		bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
533 					bch2_btree_key_cache_journal_flush);
534 		six_unlock_read(&ck->c.lock);
535 		goto unlock;
536 	}
537 	six_unlock_read(&ck->c.lock);
538 
539 	ret = lockrestart_do(trans,
540 		btree_key_cache_flush_pos(trans, key, seq,
541 				BCH_TRANS_COMMIT_journal_reclaim, false));
542 unlock:
543 	srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
544 
545 	bch2_trans_put(trans);
546 	return ret;
547 }
548 
bch2_btree_insert_key_cached(struct btree_trans * trans,unsigned flags,struct btree_insert_entry * insert_entry)549 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
550 				  unsigned flags,
551 				  struct btree_insert_entry *insert_entry)
552 {
553 	struct bch_fs *c = trans->c;
554 	struct bkey_cached *ck = (void *) (trans->paths + insert_entry->path)->l[0].b;
555 	struct bkey_i *insert = insert_entry->k;
556 	bool kick_reclaim = false;
557 
558 	BUG_ON(insert->k.u64s > ck->u64s);
559 
560 	bkey_copy(ck->k, insert);
561 
562 	if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
563 		EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
564 		set_bit(BKEY_CACHED_DIRTY, &ck->flags);
565 		atomic_long_inc(&c->btree_key_cache.nr_dirty);
566 
567 		if (bch2_nr_btree_keys_need_flush(c))
568 			kick_reclaim = true;
569 	}
570 
571 	/*
572 	 * To minimize lock contention, we only add the journal pin here and
573 	 * defer pin updates to the flush callback via ->seq. Be careful not to
574 	 * update ->seq on nojournal commits because we don't want to update the
575 	 * pin to a seq that doesn't include journal updates on disk. Otherwise
576 	 * we risk losing the update after a crash.
577 	 *
578 	 * The only exception is if the pin is not active in the first place. We
579 	 * have to add the pin because journal reclaim drives key cache
580 	 * flushing. The flush callback will not proceed unless ->seq matches
581 	 * the latest pin, so make sure it starts with a consistent value.
582 	 */
583 	if (!(insert_entry->flags & BTREE_UPDATE_nojournal) ||
584 	    !journal_pin_active(&ck->journal)) {
585 		ck->seq = trans->journal_res.seq;
586 	}
587 	bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
588 			     &ck->journal, bch2_btree_key_cache_journal_flush);
589 
590 	if (kick_reclaim)
591 		journal_reclaim_kick(&c->journal);
592 	return true;
593 }
594 
bch2_btree_key_cache_drop(struct btree_trans * trans,struct btree_path * path)595 void bch2_btree_key_cache_drop(struct btree_trans *trans,
596 			       struct btree_path *path)
597 {
598 	struct bch_fs *c = trans->c;
599 	struct btree_key_cache *bc = &c->btree_key_cache;
600 	struct bkey_cached *ck = (void *) path->l[0].b;
601 
602 	/*
603 	 * We just did an update to the btree, bypassing the key cache: the key
604 	 * cache key is now stale and must be dropped, even if dirty:
605 	 */
606 	if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
607 		clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
608 		atomic_long_dec(&c->btree_key_cache.nr_dirty);
609 		bch2_journal_pin_drop(&c->journal, &ck->journal);
610 	}
611 
612 	bkey_cached_evict(bc, ck);
613 	bkey_cached_free(bc, ck);
614 
615 	mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
616 
617 	struct btree_path *path2;
618 	unsigned i;
619 	trans_for_each_path(trans, path2, i)
620 		if (path2->l[0].b == (void *) ck) {
621 			__bch2_btree_path_unlock(trans, path2);
622 			path2->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_drop);
623 			path2->should_be_locked = false;
624 			btree_path_set_dirty(path2, BTREE_ITER_NEED_TRAVERSE);
625 		}
626 
627 	bch2_trans_verify_locks(trans);
628 }
629 
bch2_btree_key_cache_scan(struct shrinker * shrink,struct shrink_control * sc)630 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
631 					   struct shrink_control *sc)
632 {
633 	struct bch_fs *c = shrink->private_data;
634 	struct btree_key_cache *bc = &c->btree_key_cache;
635 	struct bucket_table *tbl;
636 	struct bkey_cached *ck;
637 	size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
638 	unsigned iter, start;
639 	int srcu_idx;
640 
641 	srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
642 	rcu_read_lock();
643 
644 	tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
645 
646 	/*
647 	 * Scanning is expensive while a rehash is in progress - most elements
648 	 * will be on the new hashtable, if it's in progress
649 	 *
650 	 * A rehash could still start while we're scanning - that's ok, we'll
651 	 * still see most elements.
652 	 */
653 	if (unlikely(tbl->nest)) {
654 		rcu_read_unlock();
655 		srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
656 		return SHRINK_STOP;
657 	}
658 
659 	iter = bc->shrink_iter;
660 	if (iter >= tbl->size)
661 		iter = 0;
662 	start = iter;
663 
664 	do {
665 		struct rhash_head *pos, *next;
666 
667 		pos = rht_ptr_rcu(&tbl->buckets[iter]);
668 
669 		while (!rht_is_a_nulls(pos)) {
670 			next = rht_dereference_bucket_rcu(pos->next, tbl, iter);
671 			ck = container_of(pos, struct bkey_cached, hash);
672 
673 			if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
674 				bc->skipped_dirty++;
675 			} else if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags)) {
676 				clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
677 				bc->skipped_accessed++;
678 			} else if (!bkey_cached_lock_for_evict(ck)) {
679 				bc->skipped_lock_fail++;
680 			} else if (bkey_cached_evict(bc, ck)) {
681 				bkey_cached_free(bc, ck);
682 				bc->freed++;
683 				freed++;
684 			} else {
685 				six_unlock_write(&ck->c.lock);
686 				six_unlock_intent(&ck->c.lock);
687 			}
688 
689 			scanned++;
690 			if (scanned >= nr)
691 				goto out;
692 
693 			pos = next;
694 		}
695 
696 		iter++;
697 		if (iter >= tbl->size)
698 			iter = 0;
699 	} while (scanned < nr && iter != start);
700 out:
701 	bc->shrink_iter = iter;
702 
703 	rcu_read_unlock();
704 	srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
705 
706 	return freed;
707 }
708 
bch2_btree_key_cache_count(struct shrinker * shrink,struct shrink_control * sc)709 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
710 					    struct shrink_control *sc)
711 {
712 	struct bch_fs *c = shrink->private_data;
713 	struct btree_key_cache *bc = &c->btree_key_cache;
714 	long nr = atomic_long_read(&bc->nr_keys) -
715 		atomic_long_read(&bc->nr_dirty);
716 
717 	/*
718 	 * Avoid hammering our shrinker too much if it's nearly empty - the
719 	 * shrinker code doesn't take into account how big our cache is, if it's
720 	 * mostly empty but the system is under memory pressure it causes nasty
721 	 * lock contention:
722 	 */
723 	nr -= 128;
724 
725 	return max(0L, nr);
726 }
727 
bch2_fs_btree_key_cache_exit(struct btree_key_cache * bc)728 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
729 {
730 	struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
731 	struct bucket_table *tbl;
732 	struct bkey_cached *ck;
733 	struct rhash_head *pos;
734 	LIST_HEAD(items);
735 	unsigned i;
736 
737 	shrinker_free(bc->shrink);
738 
739 	/*
740 	 * The loop is needed to guard against racing with rehash:
741 	 */
742 	while (atomic_long_read(&bc->nr_keys)) {
743 		rcu_read_lock();
744 		tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
745 		if (tbl) {
746 			if (tbl->nest) {
747 				/* wait for in progress rehash */
748 				rcu_read_unlock();
749 				mutex_lock(&bc->table.mutex);
750 				mutex_unlock(&bc->table.mutex);
751 				continue;
752 			}
753 			for (i = 0; i < tbl->size; i++)
754 				while (pos = rht_ptr_rcu(&tbl->buckets[i]), !rht_is_a_nulls(pos)) {
755 					ck = container_of(pos, struct bkey_cached, hash);
756 					BUG_ON(!bkey_cached_evict(bc, ck));
757 					kfree(ck->k);
758 					kmem_cache_free(bch2_key_cache, ck);
759 				}
760 		}
761 		rcu_read_unlock();
762 	}
763 
764 	if (atomic_long_read(&bc->nr_dirty) &&
765 	    !bch2_journal_error(&c->journal) &&
766 	    test_bit(BCH_FS_was_rw, &c->flags))
767 		panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
768 		      atomic_long_read(&bc->nr_dirty));
769 
770 	if (atomic_long_read(&bc->nr_keys))
771 		panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
772 		      atomic_long_read(&bc->nr_keys));
773 
774 	if (bc->table_init_done)
775 		rhashtable_destroy(&bc->table);
776 
777 	rcu_pending_exit(&bc->pending[0]);
778 	rcu_pending_exit(&bc->pending[1]);
779 
780 	free_percpu(bc->nr_pending);
781 }
782 
bch2_fs_btree_key_cache_init_early(struct btree_key_cache * c)783 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
784 {
785 }
786 
bch2_fs_btree_key_cache_init(struct btree_key_cache * bc)787 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
788 {
789 	struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
790 	struct shrinker *shrink;
791 
792 	bc->nr_pending = alloc_percpu(size_t);
793 	if (!bc->nr_pending)
794 		return -BCH_ERR_ENOMEM_fs_btree_cache_init;
795 
796 	if (rcu_pending_init(&bc->pending[0], &c->btree_trans_barrier, __bkey_cached_free) ||
797 	    rcu_pending_init(&bc->pending[1], &c->btree_trans_barrier, __bkey_cached_free))
798 		return -BCH_ERR_ENOMEM_fs_btree_cache_init;
799 
800 	if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
801 		return -BCH_ERR_ENOMEM_fs_btree_cache_init;
802 
803 	bc->table_init_done = true;
804 
805 	shrink = shrinker_alloc(0, "%s-btree_key_cache", c->name);
806 	if (!shrink)
807 		return -BCH_ERR_ENOMEM_fs_btree_cache_init;
808 	bc->shrink = shrink;
809 	shrink->count_objects	= bch2_btree_key_cache_count;
810 	shrink->scan_objects	= bch2_btree_key_cache_scan;
811 	shrink->batch		= 1 << 14;
812 	shrink->seeks		= 0;
813 	shrink->private_data	= c;
814 	shrinker_register(shrink);
815 	return 0;
816 }
817 
bch2_btree_key_cache_to_text(struct printbuf * out,struct btree_key_cache * bc)818 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *bc)
819 {
820 	printbuf_tabstop_push(out, 24);
821 	printbuf_tabstop_push(out, 12);
822 
823 	prt_printf(out, "keys:\t%lu\r\n",		atomic_long_read(&bc->nr_keys));
824 	prt_printf(out, "dirty:\t%lu\r\n",		atomic_long_read(&bc->nr_dirty));
825 	prt_printf(out, "table size:\t%u\r\n",		bc->table.tbl->size);
826 	prt_newline(out);
827 	prt_printf(out, "shrinker:\n");
828 	prt_printf(out, "requested_to_free:\t%lu\r\n",	bc->requested_to_free);
829 	prt_printf(out, "freed:\t%lu\r\n",		bc->freed);
830 	prt_printf(out, "skipped_dirty:\t%lu\r\n",	bc->skipped_dirty);
831 	prt_printf(out, "skipped_accessed:\t%lu\r\n",	bc->skipped_accessed);
832 	prt_printf(out, "skipped_lock_fail:\t%lu\r\n",	bc->skipped_lock_fail);
833 	prt_newline(out);
834 	prt_printf(out, "pending:\t%zu\r\n",		per_cpu_sum(bc->nr_pending));
835 }
836 
bch2_btree_key_cache_exit(void)837 void bch2_btree_key_cache_exit(void)
838 {
839 	kmem_cache_destroy(bch2_key_cache);
840 }
841 
bch2_btree_key_cache_init(void)842 int __init bch2_btree_key_cache_init(void)
843 {
844 	bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
845 	if (!bch2_key_cache)
846 		return -ENOMEM;
847 
848 	return 0;
849 }
850