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_noassert(struct btree_key_cache * bc,struct bkey_cached * ck)104 static inline void bkey_cached_free_noassert(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_free(struct btree_trans * trans,struct btree_key_cache * bc,struct bkey_cached * ck)119 static void bkey_cached_free(struct btree_trans *trans,
120 struct btree_key_cache *bc,
121 struct bkey_cached *ck)
122 {
123 /*
124 * we'll hit strange issues in the SRCU code if we aren't holding an
125 * SRCU read lock...
126 */
127 EBUG_ON(!trans->srcu_held);
128
129 bkey_cached_free_noassert(bc, ck);
130 }
131
__bkey_cached_alloc(unsigned key_u64s,gfp_t gfp)132 static struct bkey_cached *__bkey_cached_alloc(unsigned key_u64s, gfp_t gfp)
133 {
134 gfp |= __GFP_ACCOUNT|__GFP_RECLAIMABLE;
135
136 struct bkey_cached *ck = kmem_cache_zalloc(bch2_key_cache, gfp);
137 if (unlikely(!ck))
138 return NULL;
139 ck->k = kmalloc(key_u64s * sizeof(u64), gfp);
140 if (unlikely(!ck->k)) {
141 kmem_cache_free(bch2_key_cache, ck);
142 return NULL;
143 }
144 ck->u64s = key_u64s;
145 return ck;
146 }
147
148 static struct bkey_cached *
bkey_cached_alloc(struct btree_trans * trans,struct btree_path * path,unsigned key_u64s)149 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path, unsigned key_u64s)
150 {
151 struct bch_fs *c = trans->c;
152 struct btree_key_cache *bc = &c->btree_key_cache;
153 bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
154 int ret;
155
156 struct bkey_cached *ck = container_of_or_null(
157 rcu_pending_dequeue(&bc->pending[pcpu_readers]),
158 struct bkey_cached, rcu);
159 if (ck)
160 goto lock;
161
162 ck = allocate_dropping_locks(trans, ret,
163 __bkey_cached_alloc(key_u64s, _gfp));
164 if (ret) {
165 if (ck)
166 kfree(ck->k);
167 kmem_cache_free(bch2_key_cache, ck);
168 return ERR_PTR(ret);
169 }
170
171 if (ck) {
172 bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0, GFP_KERNEL);
173 ck->c.cached = true;
174 goto lock;
175 }
176
177 ck = container_of_or_null(rcu_pending_dequeue_from_all(&bc->pending[pcpu_readers]),
178 struct bkey_cached, rcu);
179 if (ck)
180 goto lock;
181 lock:
182 six_lock_intent(&ck->c.lock, NULL, NULL);
183 six_lock_write(&ck->c.lock, NULL, NULL);
184 return ck;
185 }
186
187 static struct bkey_cached *
bkey_cached_reuse(struct btree_key_cache * c)188 bkey_cached_reuse(struct btree_key_cache *c)
189 {
190
191 guard(rcu)();
192 struct bucket_table *tbl = rht_dereference_rcu(c->table.tbl, &c->table);
193 struct rhash_head *pos;
194 struct bkey_cached *ck;
195
196 for (unsigned i = 0; i < tbl->size; i++)
197 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
198 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
199 bkey_cached_lock_for_evict(ck)) {
200 if (bkey_cached_evict(c, ck))
201 return ck;
202 six_unlock_write(&ck->c.lock);
203 six_unlock_intent(&ck->c.lock);
204 }
205 }
206 return NULL;
207 }
208
btree_key_cache_create(struct btree_trans * trans,struct btree_path * path,struct btree_path * ck_path,struct bkey_s_c k)209 static int btree_key_cache_create(struct btree_trans *trans,
210 struct btree_path *path,
211 struct btree_path *ck_path,
212 struct bkey_s_c k)
213 {
214 struct bch_fs *c = trans->c;
215 struct btree_key_cache *bc = &c->btree_key_cache;
216
217 /*
218 * bch2_varint_decode can read past the end of the buffer by at
219 * most 7 bytes (it won't be used):
220 */
221 unsigned key_u64s = k.k->u64s + 1;
222
223 /*
224 * Allocate some extra space so that the transaction commit path is less
225 * likely to have to reallocate, since that requires a transaction
226 * restart:
227 */
228 key_u64s = min(256U, (key_u64s * 3) / 2);
229 key_u64s = roundup_pow_of_two(key_u64s);
230
231 struct bkey_cached *ck = bkey_cached_alloc(trans, ck_path, key_u64s);
232 int ret = PTR_ERR_OR_ZERO(ck);
233 if (ret)
234 return ret;
235
236 if (unlikely(!ck)) {
237 ck = bkey_cached_reuse(bc);
238 if (unlikely(!ck)) {
239 bch_err(c, "error allocating memory for key cache item, btree %s",
240 bch2_btree_id_str(ck_path->btree_id));
241 return bch_err_throw(c, ENOMEM_btree_key_cache_create);
242 }
243 }
244
245 ck->c.level = 0;
246 ck->c.btree_id = ck_path->btree_id;
247 ck->key.btree_id = ck_path->btree_id;
248 ck->key.pos = ck_path->pos;
249 ck->flags = 1U << BKEY_CACHED_ACCESSED;
250
251 if (unlikely(key_u64s > ck->u64s)) {
252 mark_btree_node_locked_noreset(ck_path, 0, BTREE_NODE_UNLOCKED);
253
254 struct bkey_i *new_k = allocate_dropping_locks(trans, ret,
255 kmalloc(key_u64s * sizeof(u64), _gfp));
256 if (unlikely(!new_k)) {
257 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
258 bch2_btree_id_str(ck->key.btree_id), key_u64s);
259 ret = bch_err_throw(c, ENOMEM_btree_key_cache_fill);
260 } else if (ret) {
261 kfree(new_k);
262 goto err;
263 }
264
265 kfree(ck->k);
266 ck->k = new_k;
267 ck->u64s = key_u64s;
268 }
269
270 bkey_reassemble(ck->k, k);
271
272 ret = bch2_btree_node_lock_write(trans, path, &path_l(path)->b->c);
273 if (unlikely(ret))
274 goto err;
275
276 ret = rhashtable_lookup_insert_fast(&bc->table, &ck->hash, bch2_btree_key_cache_params);
277
278 bch2_btree_node_unlock_write(trans, path, path_l(path)->b);
279
280 if (unlikely(ret)) /* raced with another fill? */
281 goto err;
282
283 atomic_long_inc(&bc->nr_keys);
284 six_unlock_write(&ck->c.lock);
285
286 enum six_lock_type lock_want = __btree_lock_want(ck_path, 0);
287 if (lock_want == SIX_LOCK_read)
288 six_lock_downgrade(&ck->c.lock);
289 btree_path_cached_set(trans, ck_path, ck, (enum btree_node_locked_type) lock_want);
290 ck_path->uptodate = BTREE_ITER_UPTODATE;
291 return 0;
292 err:
293 bkey_cached_free(trans, bc, ck);
294 mark_btree_node_locked_noreset(ck_path, 0, BTREE_NODE_UNLOCKED);
295
296 return ret;
297 }
298
do_trace_key_cache_fill(struct btree_trans * trans,struct btree_path * ck_path,struct bkey_s_c k)299 static noinline_for_stack void do_trace_key_cache_fill(struct btree_trans *trans,
300 struct btree_path *ck_path,
301 struct bkey_s_c k)
302 {
303 struct printbuf buf = PRINTBUF;
304
305 bch2_bpos_to_text(&buf, ck_path->pos);
306 prt_char(&buf, ' ');
307 bch2_bkey_val_to_text(&buf, trans->c, k);
308 trace_key_cache_fill(trans, buf.buf);
309 printbuf_exit(&buf);
310 }
311
btree_key_cache_fill(struct btree_trans * trans,btree_path_idx_t ck_path_idx,unsigned flags)312 static noinline int btree_key_cache_fill(struct btree_trans *trans,
313 btree_path_idx_t ck_path_idx,
314 unsigned flags)
315 {
316 struct btree_path *ck_path = trans->paths + ck_path_idx;
317
318 if (flags & BTREE_ITER_cached_nofill) {
319 ck_path->l[0].b = NULL;
320 return 0;
321 }
322
323 struct bch_fs *c = trans->c;
324 struct btree_iter iter;
325 struct bkey_s_c k;
326 int ret;
327
328 bch2_trans_iter_init(trans, &iter, ck_path->btree_id, ck_path->pos,
329 BTREE_ITER_intent|
330 BTREE_ITER_key_cache_fill|
331 BTREE_ITER_cached_nofill);
332 iter.flags &= ~BTREE_ITER_with_journal;
333 k = bch2_btree_iter_peek_slot(trans, &iter);
334 ret = bkey_err(k);
335 if (ret)
336 goto err;
337
338 /* Recheck after btree lookup, before allocating: */
339 ck_path = trans->paths + ck_path_idx;
340 ret = bch2_btree_key_cache_find(c, ck_path->btree_id, ck_path->pos) ? -EEXIST : 0;
341 if (unlikely(ret))
342 goto out;
343
344 ret = btree_key_cache_create(trans, btree_iter_path(trans, &iter), ck_path, k);
345 if (ret)
346 goto err;
347
348 if (trace_key_cache_fill_enabled())
349 do_trace_key_cache_fill(trans, ck_path, k);
350 out:
351 /* We're not likely to need this iterator again: */
352 bch2_set_btree_iter_dontneed(trans, &iter);
353 err:
354 bch2_trans_iter_exit(trans, &iter);
355 return ret;
356 }
357
btree_path_traverse_cached_fast(struct btree_trans * trans,btree_path_idx_t path_idx)358 static inline int btree_path_traverse_cached_fast(struct btree_trans *trans,
359 btree_path_idx_t path_idx)
360 {
361 struct bch_fs *c = trans->c;
362 struct bkey_cached *ck;
363 struct btree_path *path = trans->paths + path_idx;
364 retry:
365 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
366 if (!ck)
367 return -ENOENT;
368
369 enum six_lock_type lock_want = __btree_lock_want(path, 0);
370
371 int ret = btree_node_lock(trans, path, (void *) ck, 0, lock_want, _THIS_IP_);
372 if (ret)
373 return ret;
374
375 if (ck->key.btree_id != path->btree_id ||
376 !bpos_eq(ck->key.pos, path->pos)) {
377 six_unlock_type(&ck->c.lock, lock_want);
378 goto retry;
379 }
380
381 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
382 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
383
384 btree_path_cached_set(trans, path, ck, (enum btree_node_locked_type) lock_want);
385 path->uptodate = BTREE_ITER_UPTODATE;
386 return 0;
387 }
388
bch2_btree_path_traverse_cached(struct btree_trans * trans,btree_path_idx_t path_idx,unsigned flags)389 int bch2_btree_path_traverse_cached(struct btree_trans *trans,
390 btree_path_idx_t path_idx,
391 unsigned flags)
392 {
393 EBUG_ON(trans->paths[path_idx].level);
394
395 int ret;
396 do {
397 ret = btree_path_traverse_cached_fast(trans, path_idx);
398 if (unlikely(ret == -ENOENT))
399 ret = btree_key_cache_fill(trans, path_idx, flags);
400 } while (ret == -EEXIST);
401
402 struct btree_path *path = trans->paths + path_idx;
403
404 if (unlikely(ret)) {
405 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
406 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
407 btree_node_unlock(trans, path, 0);
408 path->l[0].b = ERR_PTR(ret);
409 }
410 } else {
411 BUG_ON(path->uptodate);
412 BUG_ON(!path->nodes_locked);
413 }
414
415 return ret;
416 }
417
btree_key_cache_flush_pos(struct btree_trans * trans,struct bkey_cached_key key,u64 journal_seq,unsigned commit_flags,bool evict)418 static int btree_key_cache_flush_pos(struct btree_trans *trans,
419 struct bkey_cached_key key,
420 u64 journal_seq,
421 unsigned commit_flags,
422 bool evict)
423 {
424 struct bch_fs *c = trans->c;
425 struct journal *j = &c->journal;
426 struct btree_iter c_iter, b_iter;
427 struct bkey_cached *ck = NULL;
428 int ret;
429
430 bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
431 BTREE_ITER_slots|
432 BTREE_ITER_intent|
433 BTREE_ITER_all_snapshots);
434 bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
435 BTREE_ITER_cached|
436 BTREE_ITER_intent);
437 b_iter.flags &= ~BTREE_ITER_with_key_cache;
438
439 ret = bch2_btree_iter_traverse(trans, &c_iter);
440 if (ret)
441 goto out;
442
443 ck = (void *) btree_iter_path(trans, &c_iter)->l[0].b;
444 if (!ck)
445 goto out;
446
447 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
448 if (evict)
449 goto evict;
450 goto out;
451 }
452
453 if (journal_seq && ck->journal.seq != journal_seq)
454 goto out;
455
456 trans->journal_res.seq = ck->journal.seq;
457
458 /*
459 * If we're at the end of the journal, we really want to free up space
460 * in the journal right away - we don't want to pin that old journal
461 * sequence number with a new btree node write, we want to re-journal
462 * the update
463 */
464 if (ck->journal.seq == journal_last_seq(j))
465 commit_flags |= BCH_WATERMARK_reclaim;
466
467 if (ck->journal.seq != journal_last_seq(j) ||
468 !test_bit(JOURNAL_space_low, &c->journal.flags))
469 commit_flags |= BCH_TRANS_COMMIT_no_journal_res;
470
471 struct bkey_s_c btree_k = bch2_btree_iter_peek_slot(trans, &b_iter);
472 ret = bkey_err(btree_k);
473 if (ret)
474 goto err;
475
476 /* * Check that we're not violating cache coherency rules: */
477 BUG_ON(bkey_deleted(btree_k.k));
478
479 ret = bch2_trans_update(trans, &b_iter, ck->k,
480 BTREE_UPDATE_key_cache_reclaim|
481 BTREE_UPDATE_internal_snapshot_node|
482 BTREE_TRIGGER_norun) ?:
483 bch2_trans_commit(trans, NULL, NULL,
484 BCH_TRANS_COMMIT_no_check_rw|
485 BCH_TRANS_COMMIT_no_enospc|
486 commit_flags);
487 err:
488 bch2_fs_fatal_err_on(ret &&
489 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
490 !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
491 !bch2_journal_error(j), c,
492 "flushing key cache: %s", bch2_err_str(ret));
493 if (ret)
494 goto out;
495
496 bch2_journal_pin_drop(j, &ck->journal);
497
498 struct btree_path *path = btree_iter_path(trans, &c_iter);
499 BUG_ON(!btree_node_locked(path, 0));
500
501 if (!evict) {
502 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
503 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
504 atomic_long_dec(&c->btree_key_cache.nr_dirty);
505 }
506 } else {
507 struct btree_path *path2;
508 unsigned i;
509 evict:
510 trans_for_each_path(trans, path2, i)
511 if (path2 != path)
512 __bch2_btree_path_unlock(trans, path2);
513
514 bch2_btree_node_lock_write_nofail(trans, path, &ck->c);
515
516 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
517 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
518 atomic_long_dec(&c->btree_key_cache.nr_dirty);
519 }
520
521 mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
522 if (bkey_cached_evict(&c->btree_key_cache, ck)) {
523 bkey_cached_free(trans, &c->btree_key_cache, ck);
524 } else {
525 six_unlock_write(&ck->c.lock);
526 six_unlock_intent(&ck->c.lock);
527 }
528 }
529 out:
530 bch2_trans_iter_exit(trans, &b_iter);
531 bch2_trans_iter_exit(trans, &c_iter);
532 return ret;
533 }
534
bch2_btree_key_cache_journal_flush(struct journal * j,struct journal_entry_pin * pin,u64 seq)535 int bch2_btree_key_cache_journal_flush(struct journal *j,
536 struct journal_entry_pin *pin, u64 seq)
537 {
538 struct bch_fs *c = container_of(j, struct bch_fs, journal);
539 struct bkey_cached *ck =
540 container_of(pin, struct bkey_cached, journal);
541 struct bkey_cached_key key;
542 struct btree_trans *trans = bch2_trans_get(c);
543 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
544 int ret = 0;
545
546 btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
547 key = ck->key;
548
549 if (ck->journal.seq != seq ||
550 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
551 six_unlock_read(&ck->c.lock);
552 goto unlock;
553 }
554
555 if (ck->seq != seq) {
556 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
557 bch2_btree_key_cache_journal_flush);
558 six_unlock_read(&ck->c.lock);
559 goto unlock;
560 }
561 six_unlock_read(&ck->c.lock);
562
563 ret = lockrestart_do(trans,
564 btree_key_cache_flush_pos(trans, key, seq,
565 BCH_TRANS_COMMIT_journal_reclaim, false));
566 unlock:
567 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
568
569 bch2_trans_put(trans);
570 return ret;
571 }
572
bch2_btree_insert_key_cached(struct btree_trans * trans,unsigned flags,struct btree_insert_entry * insert_entry)573 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
574 unsigned flags,
575 struct btree_insert_entry *insert_entry)
576 {
577 struct bch_fs *c = trans->c;
578 struct bkey_cached *ck = (void *) (trans->paths + insert_entry->path)->l[0].b;
579 struct bkey_i *insert = insert_entry->k;
580 bool kick_reclaim = false;
581
582 BUG_ON(insert->k.u64s > ck->u64s);
583
584 bkey_copy(ck->k, insert);
585
586 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
587 EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
588 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
589 atomic_long_inc(&c->btree_key_cache.nr_dirty);
590
591 if (bch2_nr_btree_keys_need_flush(c))
592 kick_reclaim = true;
593 }
594
595 /*
596 * To minimize lock contention, we only add the journal pin here and
597 * defer pin updates to the flush callback via ->seq. Be careful not to
598 * update ->seq on nojournal commits because we don't want to update the
599 * pin to a seq that doesn't include journal updates on disk. Otherwise
600 * we risk losing the update after a crash.
601 *
602 * The only exception is if the pin is not active in the first place. We
603 * have to add the pin because journal reclaim drives key cache
604 * flushing. The flush callback will not proceed unless ->seq matches
605 * the latest pin, so make sure it starts with a consistent value.
606 */
607 if (!(insert_entry->flags & BTREE_UPDATE_nojournal) ||
608 !journal_pin_active(&ck->journal)) {
609 ck->seq = trans->journal_res.seq;
610 }
611 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
612 &ck->journal, bch2_btree_key_cache_journal_flush);
613
614 if (kick_reclaim)
615 journal_reclaim_kick(&c->journal);
616 return true;
617 }
618
bch2_btree_key_cache_drop(struct btree_trans * trans,struct btree_path * path)619 void bch2_btree_key_cache_drop(struct btree_trans *trans,
620 struct btree_path *path)
621 {
622 struct bch_fs *c = trans->c;
623 struct btree_key_cache *bc = &c->btree_key_cache;
624 struct bkey_cached *ck = (void *) path->l[0].b;
625
626 /*
627 * We just did an update to the btree, bypassing the key cache: the key
628 * cache key is now stale and must be dropped, even if dirty:
629 */
630 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
631 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
632 atomic_long_dec(&c->btree_key_cache.nr_dirty);
633 bch2_journal_pin_drop(&c->journal, &ck->journal);
634 }
635
636 bkey_cached_evict(bc, ck);
637 bkey_cached_free(trans, bc, ck);
638
639 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
640
641 struct btree_path *path2;
642 unsigned i;
643 trans_for_each_path(trans, path2, i)
644 if (path2->l[0].b == (void *) ck) {
645 /*
646 * It's safe to clear should_be_locked here because
647 * we're evicting from the key cache, and we still have
648 * the underlying btree locked: filling into the key
649 * cache would require taking a write lock on the btree
650 * node
651 */
652 path2->should_be_locked = false;
653 __bch2_btree_path_unlock(trans, path2);
654 path2->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_drop);
655 btree_path_set_dirty(trans, path2, BTREE_ITER_NEED_TRAVERSE);
656 }
657
658 bch2_trans_verify_locks(trans);
659 }
660
bch2_btree_key_cache_scan(struct shrinker * shrink,struct shrink_control * sc)661 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
662 struct shrink_control *sc)
663 {
664 struct bch_fs *c = shrink->private_data;
665 struct btree_key_cache *bc = &c->btree_key_cache;
666 struct bucket_table *tbl;
667 struct bkey_cached *ck;
668 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
669 unsigned iter, start;
670 int srcu_idx;
671
672 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
673 rcu_read_lock();
674
675 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
676
677 /*
678 * Scanning is expensive while a rehash is in progress - most elements
679 * will be on the new hashtable, if it's in progress
680 *
681 * A rehash could still start while we're scanning - that's ok, we'll
682 * still see most elements.
683 */
684 if (unlikely(tbl->nest)) {
685 rcu_read_unlock();
686 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
687 return SHRINK_STOP;
688 }
689
690 iter = bc->shrink_iter;
691 if (iter >= tbl->size)
692 iter = 0;
693 start = iter;
694
695 do {
696 struct rhash_head *pos, *next;
697
698 pos = rht_ptr_rcu(&tbl->buckets[iter]);
699
700 while (!rht_is_a_nulls(pos)) {
701 next = rht_dereference_bucket_rcu(pos->next, tbl, iter);
702 ck = container_of(pos, struct bkey_cached, hash);
703
704 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
705 bc->skipped_dirty++;
706 } else if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags)) {
707 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
708 bc->skipped_accessed++;
709 } else if (!bkey_cached_lock_for_evict(ck)) {
710 bc->skipped_lock_fail++;
711 } else if (bkey_cached_evict(bc, ck)) {
712 bkey_cached_free_noassert(bc, ck);
713 bc->freed++;
714 freed++;
715 } else {
716 six_unlock_write(&ck->c.lock);
717 six_unlock_intent(&ck->c.lock);
718 }
719
720 scanned++;
721 if (scanned >= nr)
722 goto out;
723
724 pos = next;
725 }
726
727 iter++;
728 if (iter >= tbl->size)
729 iter = 0;
730 } while (scanned < nr && iter != start);
731 out:
732 bc->shrink_iter = iter;
733
734 rcu_read_unlock();
735 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
736
737 return freed;
738 }
739
bch2_btree_key_cache_count(struct shrinker * shrink,struct shrink_control * sc)740 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
741 struct shrink_control *sc)
742 {
743 struct bch_fs *c = shrink->private_data;
744 struct btree_key_cache *bc = &c->btree_key_cache;
745 long nr = atomic_long_read(&bc->nr_keys) -
746 atomic_long_read(&bc->nr_dirty);
747
748 /*
749 * Avoid hammering our shrinker too much if it's nearly empty - the
750 * shrinker code doesn't take into account how big our cache is, if it's
751 * mostly empty but the system is under memory pressure it causes nasty
752 * lock contention:
753 */
754 nr -= 128;
755
756 return max(0L, nr);
757 }
758
bch2_fs_btree_key_cache_exit(struct btree_key_cache * bc)759 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
760 {
761 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
762 struct bucket_table *tbl;
763 struct bkey_cached *ck;
764 struct rhash_head *pos;
765 LIST_HEAD(items);
766 unsigned i;
767
768 shrinker_free(bc->shrink);
769
770 /*
771 * The loop is needed to guard against racing with rehash:
772 */
773 while (atomic_long_read(&bc->nr_keys)) {
774 rcu_read_lock();
775 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
776 if (tbl) {
777 if (tbl->nest) {
778 /* wait for in progress rehash */
779 rcu_read_unlock();
780 mutex_lock(&bc->table.mutex);
781 mutex_unlock(&bc->table.mutex);
782 continue;
783 }
784 for (i = 0; i < tbl->size; i++)
785 while (pos = rht_ptr_rcu(&tbl->buckets[i]), !rht_is_a_nulls(pos)) {
786 ck = container_of(pos, struct bkey_cached, hash);
787 BUG_ON(!bkey_cached_evict(bc, ck));
788 kfree(ck->k);
789 kmem_cache_free(bch2_key_cache, ck);
790 }
791 }
792 rcu_read_unlock();
793 }
794
795 if (atomic_long_read(&bc->nr_dirty) &&
796 !bch2_journal_error(&c->journal) &&
797 test_bit(BCH_FS_was_rw, &c->flags))
798 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
799 atomic_long_read(&bc->nr_dirty));
800
801 if (atomic_long_read(&bc->nr_keys))
802 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
803 atomic_long_read(&bc->nr_keys));
804
805 if (bc->table_init_done)
806 rhashtable_destroy(&bc->table);
807
808 rcu_pending_exit(&bc->pending[0]);
809 rcu_pending_exit(&bc->pending[1]);
810
811 free_percpu(bc->nr_pending);
812 }
813
bch2_fs_btree_key_cache_init_early(struct btree_key_cache * c)814 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
815 {
816 }
817
bch2_fs_btree_key_cache_init(struct btree_key_cache * bc)818 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
819 {
820 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
821 struct shrinker *shrink;
822
823 bc->nr_pending = alloc_percpu(size_t);
824 if (!bc->nr_pending)
825 return bch_err_throw(c, ENOMEM_fs_btree_cache_init);
826
827 if (rcu_pending_init(&bc->pending[0], &c->btree_trans_barrier, __bkey_cached_free) ||
828 rcu_pending_init(&bc->pending[1], &c->btree_trans_barrier, __bkey_cached_free))
829 return bch_err_throw(c, ENOMEM_fs_btree_cache_init);
830
831 if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
832 return bch_err_throw(c, ENOMEM_fs_btree_cache_init);
833
834 bc->table_init_done = true;
835
836 shrink = shrinker_alloc(0, "%s-btree_key_cache", c->name);
837 if (!shrink)
838 return bch_err_throw(c, ENOMEM_fs_btree_cache_init);
839 bc->shrink = shrink;
840 shrink->count_objects = bch2_btree_key_cache_count;
841 shrink->scan_objects = bch2_btree_key_cache_scan;
842 shrink->batch = 1 << 14;
843 shrink->seeks = 0;
844 shrink->private_data = c;
845 shrinker_register(shrink);
846 return 0;
847 }
848
bch2_btree_key_cache_to_text(struct printbuf * out,struct btree_key_cache * bc)849 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *bc)
850 {
851 printbuf_tabstop_push(out, 24);
852 printbuf_tabstop_push(out, 12);
853
854 prt_printf(out, "keys:\t%lu\r\n", atomic_long_read(&bc->nr_keys));
855 prt_printf(out, "dirty:\t%lu\r\n", atomic_long_read(&bc->nr_dirty));
856 prt_printf(out, "table size:\t%u\r\n", bc->table.tbl->size);
857 prt_newline(out);
858 prt_printf(out, "shrinker:\n");
859 prt_printf(out, "requested_to_free:\t%lu\r\n", bc->requested_to_free);
860 prt_printf(out, "freed:\t%lu\r\n", bc->freed);
861 prt_printf(out, "skipped_dirty:\t%lu\r\n", bc->skipped_dirty);
862 prt_printf(out, "skipped_accessed:\t%lu\r\n", bc->skipped_accessed);
863 prt_printf(out, "skipped_lock_fail:\t%lu\r\n", bc->skipped_lock_fail);
864 prt_newline(out);
865 prt_printf(out, "pending:\t%zu\r\n", per_cpu_sum(bc->nr_pending));
866 }
867
bch2_btree_key_cache_exit(void)868 void bch2_btree_key_cache_exit(void)
869 {
870 kmem_cache_destroy(bch2_key_cache);
871 }
872
bch2_btree_key_cache_init(void)873 int __init bch2_btree_key_cache_init(void)
874 {
875 bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
876 if (!bch2_key_cache)
877 return -ENOMEM;
878
879 return 0;
880 }
881