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
do_trace_key_cache_fill(struct btree_trans * trans,struct btree_path * ck_path,struct bkey_s_c k)290 static noinline_for_stack void do_trace_key_cache_fill(struct btree_trans *trans,
291 struct btree_path *ck_path,
292 struct bkey_s_c k)
293 {
294 struct printbuf buf = PRINTBUF;
295
296 bch2_bpos_to_text(&buf, ck_path->pos);
297 prt_char(&buf, ' ');
298 bch2_bkey_val_to_text(&buf, trans->c, k);
299 trace_key_cache_fill(trans, buf.buf);
300 printbuf_exit(&buf);
301 }
302
btree_key_cache_fill(struct btree_trans * trans,struct btree_path * ck_path,unsigned flags)303 static noinline int btree_key_cache_fill(struct btree_trans *trans,
304 struct btree_path *ck_path,
305 unsigned flags)
306 {
307 if (flags & BTREE_ITER_cached_nofill) {
308 ck_path->l[0].b = NULL;
309 return 0;
310 }
311
312 struct bch_fs *c = trans->c;
313 struct btree_iter iter;
314 struct bkey_s_c k;
315 int ret;
316
317 bch2_trans_iter_init(trans, &iter, ck_path->btree_id, ck_path->pos,
318 BTREE_ITER_intent|
319 BTREE_ITER_key_cache_fill|
320 BTREE_ITER_cached_nofill);
321 iter.flags &= ~BTREE_ITER_with_journal;
322 k = bch2_btree_iter_peek_slot(trans, &iter);
323 ret = bkey_err(k);
324 if (ret)
325 goto err;
326
327 /* Recheck after btree lookup, before allocating: */
328 ret = bch2_btree_key_cache_find(c, ck_path->btree_id, ck_path->pos) ? -EEXIST : 0;
329 if (unlikely(ret))
330 goto out;
331
332 ret = btree_key_cache_create(trans, btree_iter_path(trans, &iter), ck_path, k);
333 if (ret)
334 goto err;
335
336 if (trace_key_cache_fill_enabled())
337 do_trace_key_cache_fill(trans, ck_path, k);
338 out:
339 /* We're not likely to need this iterator again: */
340 bch2_set_btree_iter_dontneed(trans, &iter);
341 err:
342 bch2_trans_iter_exit(trans, &iter);
343 return ret;
344 }
345
btree_path_traverse_cached_fast(struct btree_trans * trans,struct btree_path * path)346 static inline int btree_path_traverse_cached_fast(struct btree_trans *trans,
347 struct btree_path *path)
348 {
349 struct bch_fs *c = trans->c;
350 struct bkey_cached *ck;
351 retry:
352 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
353 if (!ck)
354 return -ENOENT;
355
356 enum six_lock_type lock_want = __btree_lock_want(path, 0);
357
358 int ret = btree_node_lock(trans, path, (void *) ck, 0, lock_want, _THIS_IP_);
359 if (ret)
360 return ret;
361
362 if (ck->key.btree_id != path->btree_id ||
363 !bpos_eq(ck->key.pos, path->pos)) {
364 six_unlock_type(&ck->c.lock, lock_want);
365 goto retry;
366 }
367
368 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
369 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
370
371 btree_path_cached_set(trans, path, ck, (enum btree_node_locked_type) lock_want);
372 path->uptodate = BTREE_ITER_UPTODATE;
373 return 0;
374 }
375
bch2_btree_path_traverse_cached(struct btree_trans * trans,struct btree_path * path,unsigned flags)376 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
377 unsigned flags)
378 {
379 EBUG_ON(path->level);
380
381 path->l[1].b = NULL;
382
383 int ret;
384 do {
385 ret = btree_path_traverse_cached_fast(trans, path);
386 if (unlikely(ret == -ENOENT))
387 ret = btree_key_cache_fill(trans, path, flags);
388 } while (ret == -EEXIST);
389
390 if (unlikely(ret)) {
391 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
392 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
393 btree_node_unlock(trans, path, 0);
394 path->l[0].b = ERR_PTR(ret);
395 }
396 }
397 return ret;
398 }
399
btree_key_cache_flush_pos(struct btree_trans * trans,struct bkey_cached_key key,u64 journal_seq,unsigned commit_flags,bool evict)400 static int btree_key_cache_flush_pos(struct btree_trans *trans,
401 struct bkey_cached_key key,
402 u64 journal_seq,
403 unsigned commit_flags,
404 bool evict)
405 {
406 struct bch_fs *c = trans->c;
407 struct journal *j = &c->journal;
408 struct btree_iter c_iter, b_iter;
409 struct bkey_cached *ck = NULL;
410 int ret;
411
412 bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
413 BTREE_ITER_slots|
414 BTREE_ITER_intent|
415 BTREE_ITER_all_snapshots);
416 bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
417 BTREE_ITER_cached|
418 BTREE_ITER_intent);
419 b_iter.flags &= ~BTREE_ITER_with_key_cache;
420
421 ret = bch2_btree_iter_traverse(trans, &c_iter);
422 if (ret)
423 goto out;
424
425 ck = (void *) btree_iter_path(trans, &c_iter)->l[0].b;
426 if (!ck)
427 goto out;
428
429 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
430 if (evict)
431 goto evict;
432 goto out;
433 }
434
435 if (journal_seq && ck->journal.seq != journal_seq)
436 goto out;
437
438 trans->journal_res.seq = ck->journal.seq;
439
440 /*
441 * If we're at the end of the journal, we really want to free up space
442 * in the journal right away - we don't want to pin that old journal
443 * sequence number with a new btree node write, we want to re-journal
444 * the update
445 */
446 if (ck->journal.seq == journal_last_seq(j))
447 commit_flags |= BCH_WATERMARK_reclaim;
448
449 if (ck->journal.seq != journal_last_seq(j) ||
450 !test_bit(JOURNAL_space_low, &c->journal.flags))
451 commit_flags |= BCH_TRANS_COMMIT_no_journal_res;
452
453 struct bkey_s_c btree_k = bch2_btree_iter_peek_slot(trans, &b_iter);
454 ret = bkey_err(btree_k);
455 if (ret)
456 goto err;
457
458 /* * Check that we're not violating cache coherency rules: */
459 BUG_ON(bkey_deleted(btree_k.k));
460
461 ret = bch2_trans_update(trans, &b_iter, ck->k,
462 BTREE_UPDATE_key_cache_reclaim|
463 BTREE_UPDATE_internal_snapshot_node|
464 BTREE_TRIGGER_norun) ?:
465 bch2_trans_commit(trans, NULL, NULL,
466 BCH_TRANS_COMMIT_no_check_rw|
467 BCH_TRANS_COMMIT_no_enospc|
468 commit_flags);
469 err:
470 bch2_fs_fatal_err_on(ret &&
471 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
472 !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
473 !bch2_journal_error(j), c,
474 "flushing key cache: %s", bch2_err_str(ret));
475 if (ret)
476 goto out;
477
478 bch2_journal_pin_drop(j, &ck->journal);
479
480 struct btree_path *path = btree_iter_path(trans, &c_iter);
481 BUG_ON(!btree_node_locked(path, 0));
482
483 if (!evict) {
484 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
485 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
486 atomic_long_dec(&c->btree_key_cache.nr_dirty);
487 }
488 } else {
489 struct btree_path *path2;
490 unsigned i;
491 evict:
492 trans_for_each_path(trans, path2, i)
493 if (path2 != path)
494 __bch2_btree_path_unlock(trans, path2);
495
496 bch2_btree_node_lock_write_nofail(trans, path, &ck->c);
497
498 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
499 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
500 atomic_long_dec(&c->btree_key_cache.nr_dirty);
501 }
502
503 mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
504 if (bkey_cached_evict(&c->btree_key_cache, ck)) {
505 bkey_cached_free(&c->btree_key_cache, ck);
506 } else {
507 six_unlock_write(&ck->c.lock);
508 six_unlock_intent(&ck->c.lock);
509 }
510 }
511 out:
512 bch2_trans_iter_exit(trans, &b_iter);
513 bch2_trans_iter_exit(trans, &c_iter);
514 return ret;
515 }
516
bch2_btree_key_cache_journal_flush(struct journal * j,struct journal_entry_pin * pin,u64 seq)517 int bch2_btree_key_cache_journal_flush(struct journal *j,
518 struct journal_entry_pin *pin, u64 seq)
519 {
520 struct bch_fs *c = container_of(j, struct bch_fs, journal);
521 struct bkey_cached *ck =
522 container_of(pin, struct bkey_cached, journal);
523 struct bkey_cached_key key;
524 struct btree_trans *trans = bch2_trans_get(c);
525 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
526 int ret = 0;
527
528 btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
529 key = ck->key;
530
531 if (ck->journal.seq != seq ||
532 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
533 six_unlock_read(&ck->c.lock);
534 goto unlock;
535 }
536
537 if (ck->seq != seq) {
538 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
539 bch2_btree_key_cache_journal_flush);
540 six_unlock_read(&ck->c.lock);
541 goto unlock;
542 }
543 six_unlock_read(&ck->c.lock);
544
545 ret = lockrestart_do(trans,
546 btree_key_cache_flush_pos(trans, key, seq,
547 BCH_TRANS_COMMIT_journal_reclaim, false));
548 unlock:
549 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
550
551 bch2_trans_put(trans);
552 return ret;
553 }
554
bch2_btree_insert_key_cached(struct btree_trans * trans,unsigned flags,struct btree_insert_entry * insert_entry)555 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
556 unsigned flags,
557 struct btree_insert_entry *insert_entry)
558 {
559 struct bch_fs *c = trans->c;
560 struct bkey_cached *ck = (void *) (trans->paths + insert_entry->path)->l[0].b;
561 struct bkey_i *insert = insert_entry->k;
562 bool kick_reclaim = false;
563
564 BUG_ON(insert->k.u64s > ck->u64s);
565
566 bkey_copy(ck->k, insert);
567
568 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
569 EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
570 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
571 atomic_long_inc(&c->btree_key_cache.nr_dirty);
572
573 if (bch2_nr_btree_keys_need_flush(c))
574 kick_reclaim = true;
575 }
576
577 /*
578 * To minimize lock contention, we only add the journal pin here and
579 * defer pin updates to the flush callback via ->seq. Be careful not to
580 * update ->seq on nojournal commits because we don't want to update the
581 * pin to a seq that doesn't include journal updates on disk. Otherwise
582 * we risk losing the update after a crash.
583 *
584 * The only exception is if the pin is not active in the first place. We
585 * have to add the pin because journal reclaim drives key cache
586 * flushing. The flush callback will not proceed unless ->seq matches
587 * the latest pin, so make sure it starts with a consistent value.
588 */
589 if (!(insert_entry->flags & BTREE_UPDATE_nojournal) ||
590 !journal_pin_active(&ck->journal)) {
591 ck->seq = trans->journal_res.seq;
592 }
593 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
594 &ck->journal, bch2_btree_key_cache_journal_flush);
595
596 if (kick_reclaim)
597 journal_reclaim_kick(&c->journal);
598 return true;
599 }
600
bch2_btree_key_cache_drop(struct btree_trans * trans,struct btree_path * path)601 void bch2_btree_key_cache_drop(struct btree_trans *trans,
602 struct btree_path *path)
603 {
604 struct bch_fs *c = trans->c;
605 struct btree_key_cache *bc = &c->btree_key_cache;
606 struct bkey_cached *ck = (void *) path->l[0].b;
607
608 /*
609 * We just did an update to the btree, bypassing the key cache: the key
610 * cache key is now stale and must be dropped, even if dirty:
611 */
612 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
613 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
614 atomic_long_dec(&c->btree_key_cache.nr_dirty);
615 bch2_journal_pin_drop(&c->journal, &ck->journal);
616 }
617
618 bkey_cached_evict(bc, ck);
619 bkey_cached_free(bc, ck);
620
621 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
622
623 struct btree_path *path2;
624 unsigned i;
625 trans_for_each_path(trans, path2, i)
626 if (path2->l[0].b == (void *) ck) {
627 __bch2_btree_path_unlock(trans, path2);
628 path2->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_drop);
629 path2->should_be_locked = false;
630 btree_path_set_dirty(path2, BTREE_ITER_NEED_TRAVERSE);
631 }
632
633 bch2_trans_verify_locks(trans);
634 }
635
bch2_btree_key_cache_scan(struct shrinker * shrink,struct shrink_control * sc)636 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
637 struct shrink_control *sc)
638 {
639 struct bch_fs *c = shrink->private_data;
640 struct btree_key_cache *bc = &c->btree_key_cache;
641 struct bucket_table *tbl;
642 struct bkey_cached *ck;
643 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
644 unsigned iter, start;
645 int srcu_idx;
646
647 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
648 rcu_read_lock();
649
650 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
651
652 /*
653 * Scanning is expensive while a rehash is in progress - most elements
654 * will be on the new hashtable, if it's in progress
655 *
656 * A rehash could still start while we're scanning - that's ok, we'll
657 * still see most elements.
658 */
659 if (unlikely(tbl->nest)) {
660 rcu_read_unlock();
661 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
662 return SHRINK_STOP;
663 }
664
665 iter = bc->shrink_iter;
666 if (iter >= tbl->size)
667 iter = 0;
668 start = iter;
669
670 do {
671 struct rhash_head *pos, *next;
672
673 pos = rht_ptr_rcu(&tbl->buckets[iter]);
674
675 while (!rht_is_a_nulls(pos)) {
676 next = rht_dereference_bucket_rcu(pos->next, tbl, iter);
677 ck = container_of(pos, struct bkey_cached, hash);
678
679 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
680 bc->skipped_dirty++;
681 } else if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags)) {
682 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
683 bc->skipped_accessed++;
684 } else if (!bkey_cached_lock_for_evict(ck)) {
685 bc->skipped_lock_fail++;
686 } else if (bkey_cached_evict(bc, ck)) {
687 bkey_cached_free(bc, ck);
688 bc->freed++;
689 freed++;
690 } else {
691 six_unlock_write(&ck->c.lock);
692 six_unlock_intent(&ck->c.lock);
693 }
694
695 scanned++;
696 if (scanned >= nr)
697 goto out;
698
699 pos = next;
700 }
701
702 iter++;
703 if (iter >= tbl->size)
704 iter = 0;
705 } while (scanned < nr && iter != start);
706 out:
707 bc->shrink_iter = iter;
708
709 rcu_read_unlock();
710 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
711
712 return freed;
713 }
714
bch2_btree_key_cache_count(struct shrinker * shrink,struct shrink_control * sc)715 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
716 struct shrink_control *sc)
717 {
718 struct bch_fs *c = shrink->private_data;
719 struct btree_key_cache *bc = &c->btree_key_cache;
720 long nr = atomic_long_read(&bc->nr_keys) -
721 atomic_long_read(&bc->nr_dirty);
722
723 /*
724 * Avoid hammering our shrinker too much if it's nearly empty - the
725 * shrinker code doesn't take into account how big our cache is, if it's
726 * mostly empty but the system is under memory pressure it causes nasty
727 * lock contention:
728 */
729 nr -= 128;
730
731 return max(0L, nr);
732 }
733
bch2_fs_btree_key_cache_exit(struct btree_key_cache * bc)734 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
735 {
736 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
737 struct bucket_table *tbl;
738 struct bkey_cached *ck;
739 struct rhash_head *pos;
740 LIST_HEAD(items);
741 unsigned i;
742
743 shrinker_free(bc->shrink);
744
745 /*
746 * The loop is needed to guard against racing with rehash:
747 */
748 while (atomic_long_read(&bc->nr_keys)) {
749 rcu_read_lock();
750 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
751 if (tbl) {
752 if (tbl->nest) {
753 /* wait for in progress rehash */
754 rcu_read_unlock();
755 mutex_lock(&bc->table.mutex);
756 mutex_unlock(&bc->table.mutex);
757 continue;
758 }
759 for (i = 0; i < tbl->size; i++)
760 while (pos = rht_ptr_rcu(&tbl->buckets[i]), !rht_is_a_nulls(pos)) {
761 ck = container_of(pos, struct bkey_cached, hash);
762 BUG_ON(!bkey_cached_evict(bc, ck));
763 kfree(ck->k);
764 kmem_cache_free(bch2_key_cache, ck);
765 }
766 }
767 rcu_read_unlock();
768 }
769
770 if (atomic_long_read(&bc->nr_dirty) &&
771 !bch2_journal_error(&c->journal) &&
772 test_bit(BCH_FS_was_rw, &c->flags))
773 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
774 atomic_long_read(&bc->nr_dirty));
775
776 if (atomic_long_read(&bc->nr_keys))
777 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
778 atomic_long_read(&bc->nr_keys));
779
780 if (bc->table_init_done)
781 rhashtable_destroy(&bc->table);
782
783 rcu_pending_exit(&bc->pending[0]);
784 rcu_pending_exit(&bc->pending[1]);
785
786 free_percpu(bc->nr_pending);
787 }
788
bch2_fs_btree_key_cache_init_early(struct btree_key_cache * c)789 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
790 {
791 }
792
bch2_fs_btree_key_cache_init(struct btree_key_cache * bc)793 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
794 {
795 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
796 struct shrinker *shrink;
797
798 bc->nr_pending = alloc_percpu(size_t);
799 if (!bc->nr_pending)
800 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
801
802 if (rcu_pending_init(&bc->pending[0], &c->btree_trans_barrier, __bkey_cached_free) ||
803 rcu_pending_init(&bc->pending[1], &c->btree_trans_barrier, __bkey_cached_free))
804 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
805
806 if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
807 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
808
809 bc->table_init_done = true;
810
811 shrink = shrinker_alloc(0, "%s-btree_key_cache", c->name);
812 if (!shrink)
813 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
814 bc->shrink = shrink;
815 shrink->count_objects = bch2_btree_key_cache_count;
816 shrink->scan_objects = bch2_btree_key_cache_scan;
817 shrink->batch = 1 << 14;
818 shrink->seeks = 0;
819 shrink->private_data = c;
820 shrinker_register(shrink);
821 return 0;
822 }
823
bch2_btree_key_cache_to_text(struct printbuf * out,struct btree_key_cache * bc)824 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *bc)
825 {
826 printbuf_tabstop_push(out, 24);
827 printbuf_tabstop_push(out, 12);
828
829 prt_printf(out, "keys:\t%lu\r\n", atomic_long_read(&bc->nr_keys));
830 prt_printf(out, "dirty:\t%lu\r\n", atomic_long_read(&bc->nr_dirty));
831 prt_printf(out, "table size:\t%u\r\n", bc->table.tbl->size);
832 prt_newline(out);
833 prt_printf(out, "shrinker:\n");
834 prt_printf(out, "requested_to_free:\t%lu\r\n", bc->requested_to_free);
835 prt_printf(out, "freed:\t%lu\r\n", bc->freed);
836 prt_printf(out, "skipped_dirty:\t%lu\r\n", bc->skipped_dirty);
837 prt_printf(out, "skipped_accessed:\t%lu\r\n", bc->skipped_accessed);
838 prt_printf(out, "skipped_lock_fail:\t%lu\r\n", bc->skipped_lock_fail);
839 prt_newline(out);
840 prt_printf(out, "pending:\t%zu\r\n", per_cpu_sum(bc->nr_pending));
841 }
842
bch2_btree_key_cache_exit(void)843 void bch2_btree_key_cache_exit(void)
844 {
845 kmem_cache_destroy(bch2_key_cache);
846 }
847
bch2_btree_key_cache_init(void)848 int __init bch2_btree_key_cache_init(void)
849 {
850 bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
851 if (!bch2_key_cache)
852 return -ENOMEM;
853
854 return 0;
855 }
856