1 // SPDX-License-Identifier: GPL-2.0
4 #include "btree_locking.h"
5 #include "btree_update.h"
6 #include "btree_update_interior.h"
7 #include "btree_write_buffer.h"
10 #include "journal_io.h"
11 #include "journal_reclaim.h"
13 #include <linux/prefetch.h>
14 #include <linux/sort.h>
16 static int bch2_btree_write_buffer_journal_flush(struct journal *,
17 struct journal_entry_pin *, u64);
19 static int bch2_journal_keys_to_write_buffer(struct bch_fs *, struct journal_buf *);
21 static inline bool __wb_key_ref_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
23 return (cmp_int(l->hi, r->hi) ?:
24 cmp_int(l->mi, r->mi) ?:
25 cmp_int(l->lo, r->lo)) >= 0;
28 static inline bool wb_key_ref_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
33 asm("mov (%[l]), %%rax;"
37 "mov 16(%[l]), %%rax;"
38 "sbb 16(%[r]), %%rax;"
40 : [l] "r" (l), [r] "r" (r)
43 EBUG_ON(cmp != __wb_key_ref_cmp(l, r));
46 return __wb_key_ref_cmp(l, r);
50 static int wb_key_seq_cmp(const void *_l, const void *_r)
52 const struct btree_write_buffered_key *l = _l;
53 const struct btree_write_buffered_key *r = _r;
55 return cmp_int(l->journal_seq, r->journal_seq);
58 /* Compare excluding idx, the low 24 bits: */
59 static inline bool wb_key_eq(const void *_l, const void *_r)
61 const struct wb_key_ref *l = _l;
62 const struct wb_key_ref *r = _r;
64 return !((l->hi ^ r->hi)|
66 ((l->lo >> 24) ^ (r->lo >> 24)));
69 static noinline void wb_sort(struct wb_key_ref *base, size_t num)
71 size_t n = num, a = num / 2;
73 if (!a) /* num < 2 || size == 0 */
79 if (a) /* Building heap: sift down --a */
81 else if (--n) /* Sorting: Extract root to --n */
82 swap(base[0], base[n]);
83 else /* Sort complete */
87 * Sift element at "a" down into heap. This is the
88 * "bottom-up" variant, which significantly reduces
89 * calls to cmp_func(): we find the sift-down path all
90 * the way to the leaves (one compare per level), then
91 * backtrack to find where to insert the target element.
93 * Because elements tend to sift down close to the leaves,
94 * this uses fewer compares than doing two per level
95 * on the way down. (A bit more than half as many on
96 * average, 3/4 worst-case.)
98 for (b = a; c = 2*b + 1, (d = c + 1) < n;)
99 b = wb_key_ref_cmp(base + c, base + d) ? c : d;
100 if (d == n) /* Special case last leaf with no sibling */
103 /* Now backtrack from "b" to the correct location for "a" */
104 while (b != a && wb_key_ref_cmp(base + a, base + b))
106 c = b; /* Where "a" belongs */
107 while (b != a) { /* Shift it into place */
109 swap(base[b], base[c]);
114 static noinline int wb_flush_one_slowpath(struct btree_trans *trans,
115 struct btree_iter *iter,
116 struct btree_write_buffered_key *wb)
118 struct btree_path *path = btree_iter_path(trans, iter);
120 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
122 trans->journal_res.seq = wb->journal_seq;
124 return bch2_trans_update(trans, iter, &wb->k,
125 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
126 bch2_trans_commit(trans, NULL, NULL,
127 BCH_TRANS_COMMIT_no_enospc|
128 BCH_TRANS_COMMIT_no_check_rw|
129 BCH_TRANS_COMMIT_no_journal_res|
130 BCH_TRANS_COMMIT_journal_reclaim);
133 static inline int wb_flush_one(struct btree_trans *trans, struct btree_iter *iter,
134 struct btree_write_buffered_key *wb,
135 bool *write_locked, size_t *fast)
137 struct btree_path *path;
140 EBUG_ON(!wb->journal_seq);
141 EBUG_ON(!trans->c->btree_write_buffer.flushing.pin.seq);
142 EBUG_ON(trans->c->btree_write_buffer.flushing.pin.seq > wb->journal_seq);
144 ret = bch2_btree_iter_traverse(iter);
149 * We can't clone a path that has write locks: unshare it now, before
150 * set_pos and traverse():
152 if (btree_iter_path(trans, iter)->ref > 1)
153 iter->path = __bch2_btree_path_make_mut(trans, iter->path, true, _THIS_IP_);
155 path = btree_iter_path(trans, iter);
157 if (!*write_locked) {
158 ret = bch2_btree_node_lock_write(trans, path, &path->l[0].b->c);
162 bch2_btree_node_prep_for_write(trans, path, path->l[0].b);
163 *write_locked = true;
166 if (unlikely(!bch2_btree_node_insert_fits(path->l[0].b, wb->k.k.u64s))) {
167 *write_locked = false;
168 return wb_flush_one_slowpath(trans, iter, wb);
171 bch2_btree_insert_key_leaf(trans, path, &wb->k, wb->journal_seq);
177 * Update a btree with a write buffered key using the journal seq of the
178 * original write buffer insert.
180 * It is not safe to rejournal the key once it has been inserted into the write
181 * buffer because that may break recovery ordering. For example, the key may
182 * have already been modified in the active write buffer in a seq that comes
183 * before the current transaction. If we were to journal this key again and
184 * crash, recovery would process updates in the wrong order.
187 btree_write_buffered_insert(struct btree_trans *trans,
188 struct btree_write_buffered_key *wb)
190 struct btree_iter iter;
193 bch2_trans_iter_init(trans, &iter, wb->btree, bkey_start_pos(&wb->k.k),
194 BTREE_ITER_CACHED|BTREE_ITER_INTENT);
196 trans->journal_res.seq = wb->journal_seq;
198 ret = bch2_btree_iter_traverse(&iter) ?:
199 bch2_trans_update(trans, &iter, &wb->k,
200 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
201 bch2_trans_iter_exit(trans, &iter);
205 static void move_keys_from_inc_to_flushing(struct btree_write_buffer *wb)
207 struct bch_fs *c = container_of(wb, struct bch_fs, btree_write_buffer);
208 struct journal *j = &c->journal;
210 if (!wb->inc.keys.nr)
213 bch2_journal_pin_add(j, wb->inc.keys.data[0].journal_seq, &wb->flushing.pin,
214 bch2_btree_write_buffer_journal_flush);
216 darray_resize(&wb->flushing.keys, min_t(size_t, 1U << 20, wb->flushing.keys.nr + wb->inc.keys.nr));
217 darray_resize(&wb->sorted, wb->flushing.keys.size);
219 if (!wb->flushing.keys.nr && wb->sorted.size >= wb->inc.keys.nr) {
220 swap(wb->flushing.keys, wb->inc.keys);
224 size_t nr = min(darray_room(wb->flushing.keys),
225 wb->sorted.size - wb->flushing.keys.nr);
226 nr = min(nr, wb->inc.keys.nr);
228 memcpy(&darray_top(wb->flushing.keys),
230 sizeof(wb->inc.keys.data[0]) * nr);
232 memmove(wb->inc.keys.data,
233 wb->inc.keys.data + nr,
234 sizeof(wb->inc.keys.data[0]) * (wb->inc.keys.nr - nr));
236 wb->flushing.keys.nr += nr;
237 wb->inc.keys.nr -= nr;
239 if (!wb->inc.keys.nr)
240 bch2_journal_pin_drop(j, &wb->inc.pin);
242 bch2_journal_pin_update(j, wb->inc.keys.data[0].journal_seq, &wb->inc.pin,
243 bch2_btree_write_buffer_journal_flush);
247 bch2_journal_set_watermark(j);
248 spin_unlock(&j->lock);
251 BUG_ON(wb->sorted.size < wb->flushing.keys.nr);
254 static int bch2_btree_write_buffer_flush_locked(struct btree_trans *trans)
256 struct bch_fs *c = trans->c;
257 struct journal *j = &c->journal;
258 struct btree_write_buffer *wb = &c->btree_write_buffer;
259 struct btree_iter iter = { NULL };
260 size_t skipped = 0, fast = 0, slowpath = 0;
261 bool write_locked = false;
264 bch2_trans_unlock(trans);
265 bch2_trans_begin(trans);
267 mutex_lock(&wb->inc.lock);
268 move_keys_from_inc_to_flushing(wb);
269 mutex_unlock(&wb->inc.lock);
271 for (size_t i = 0; i < wb->flushing.keys.nr; i++) {
272 wb->sorted.data[i].idx = i;
273 wb->sorted.data[i].btree = wb->flushing.keys.data[i].btree;
274 memcpy(&wb->sorted.data[i].pos, &wb->flushing.keys.data[i].k.k.p, sizeof(struct bpos));
276 wb->sorted.nr = wb->flushing.keys.nr;
279 * We first sort so that we can detect and skip redundant updates, and
280 * then we attempt to flush in sorted btree order, as this is most
283 * However, since we're not flushing in the order they appear in the
284 * journal we won't be able to drop our journal pin until everything is
285 * flushed - which means this could deadlock the journal if we weren't
286 * passing BCH_TRANS_COMMIT_journal_reclaim. This causes the update to fail
287 * if it would block taking a journal reservation.
289 * If that happens, simply skip the key so we can optimistically insert
290 * as many keys as possible in the fast path.
292 wb_sort(wb->sorted.data, wb->sorted.nr);
294 darray_for_each(wb->sorted, i) {
295 struct btree_write_buffered_key *k = &wb->flushing.keys.data[i->idx];
297 for (struct wb_key_ref *n = i + 1; n < min(i + 4, &darray_top(wb->sorted)); n++)
298 prefetch(&wb->flushing.keys.data[n->idx]);
300 BUG_ON(!k->journal_seq);
302 if (i + 1 < &darray_top(wb->sorted) &&
303 wb_key_eq(i, i + 1)) {
304 struct btree_write_buffered_key *n = &wb->flushing.keys.data[i[1].idx];
307 n->journal_seq = min_t(u64, n->journal_seq, k->journal_seq);
313 struct btree_path *path = btree_iter_path(trans, &iter);
315 if (path->btree_id != i->btree ||
316 bpos_gt(k->k.k.p, path->l[0].b->key.k.p)) {
317 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
318 write_locked = false;
320 ret = lockrestart_do(trans,
321 bch2_btree_iter_traverse(&iter) ?:
322 bch2_foreground_maybe_merge(trans, iter.path, 0,
323 BCH_WATERMARK_reclaim|
324 BCH_TRANS_COMMIT_journal_reclaim|
325 BCH_TRANS_COMMIT_no_check_rw|
326 BCH_TRANS_COMMIT_no_enospc));
332 if (!iter.path || iter.btree_id != k->btree) {
333 bch2_trans_iter_exit(trans, &iter);
334 bch2_trans_iter_init(trans, &iter, k->btree, k->k.k.p,
335 BTREE_ITER_INTENT|BTREE_ITER_ALL_SNAPSHOTS);
338 bch2_btree_iter_set_pos(&iter, k->k.k.p);
339 btree_iter_path(trans, &iter)->preserve = false;
343 ret = -BCH_ERR_journal_reclaim_would_deadlock;
347 ret = wb_flush_one(trans, &iter, k, &write_locked, &fast);
349 bch2_trans_begin(trans);
350 } while (bch2_err_matches(ret, BCH_ERR_transaction_restart));
354 } else if (ret == -BCH_ERR_journal_reclaim_would_deadlock) {
362 struct btree_path *path = btree_iter_path(trans, &iter);
363 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
365 bch2_trans_iter_exit(trans, &iter);
372 * Flush in the order they were present in the journal, so that
373 * we can release journal pins:
374 * The fastpath zapped the seq of keys that were successfully flushed so
375 * we can skip those here.
377 trace_and_count(c, write_buffer_flush_slowpath, trans, slowpath, wb->flushing.keys.nr);
379 sort(wb->flushing.keys.data,
380 wb->flushing.keys.nr,
381 sizeof(wb->flushing.keys.data[0]),
382 wb_key_seq_cmp, NULL);
384 darray_for_each(wb->flushing.keys, i) {
388 bch2_journal_pin_update(j, i->journal_seq, &wb->flushing.pin,
389 bch2_btree_write_buffer_journal_flush);
391 bch2_trans_begin(trans);
393 ret = commit_do(trans, NULL, NULL,
394 BCH_WATERMARK_reclaim|
395 BCH_TRANS_COMMIT_journal_reclaim|
396 BCH_TRANS_COMMIT_no_check_rw|
397 BCH_TRANS_COMMIT_no_enospc|
398 BCH_TRANS_COMMIT_no_journal_res ,
399 btree_write_buffered_insert(trans, i));
405 bch2_fs_fatal_err_on(ret, c, "%s", bch2_err_str(ret));
406 trace_write_buffer_flush(trans, wb->flushing.keys.nr, skipped, fast, 0);
407 bch2_journal_pin_drop(j, &wb->flushing.pin);
408 wb->flushing.keys.nr = 0;
412 static int fetch_wb_keys_from_journal(struct bch_fs *c, u64 seq)
414 struct journal *j = &c->journal;
415 struct journal_buf *buf;
418 while (!ret && (buf = bch2_next_write_buffer_flush_journal_buf(j, seq))) {
419 ret = bch2_journal_keys_to_write_buffer(c, buf);
420 mutex_unlock(&j->buf_lock);
426 static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq)
428 struct bch_fs *c = trans->c;
429 struct btree_write_buffer *wb = &c->btree_write_buffer;
430 int ret = 0, fetch_from_journal_err;
433 bch2_trans_unlock(trans);
435 fetch_from_journal_err = fetch_wb_keys_from_journal(c, seq);
438 * On memory allocation failure, bch2_btree_write_buffer_flush_locked()
439 * is not guaranteed to empty wb->inc:
441 mutex_lock(&wb->flushing.lock);
442 ret = bch2_btree_write_buffer_flush_locked(trans);
443 mutex_unlock(&wb->flushing.lock);
445 (fetch_from_journal_err ||
446 (wb->inc.pin.seq && wb->inc.pin.seq <= seq) ||
447 (wb->flushing.pin.seq && wb->flushing.pin.seq <= seq)));
452 static int bch2_btree_write_buffer_journal_flush(struct journal *j,
453 struct journal_entry_pin *_pin, u64 seq)
455 struct bch_fs *c = container_of(j, struct bch_fs, journal);
457 return bch2_trans_run(c, btree_write_buffer_flush_seq(trans, seq));
460 int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
462 struct bch_fs *c = trans->c;
464 trace_and_count(c, write_buffer_flush_sync, trans, _RET_IP_);
466 return btree_write_buffer_flush_seq(trans, journal_cur_seq(&c->journal));
469 int bch2_btree_write_buffer_flush_nocheck_rw(struct btree_trans *trans)
471 struct bch_fs *c = trans->c;
472 struct btree_write_buffer *wb = &c->btree_write_buffer;
475 if (mutex_trylock(&wb->flushing.lock)) {
476 ret = bch2_btree_write_buffer_flush_locked(trans);
477 mutex_unlock(&wb->flushing.lock);
483 int bch2_btree_write_buffer_tryflush(struct btree_trans *trans)
485 struct bch_fs *c = trans->c;
487 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer))
488 return -BCH_ERR_erofs_no_writes;
490 int ret = bch2_btree_write_buffer_flush_nocheck_rw(trans);
491 bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
495 static void bch2_btree_write_buffer_flush_work(struct work_struct *work)
497 struct bch_fs *c = container_of(work, struct bch_fs, btree_write_buffer.flush_work);
498 struct btree_write_buffer *wb = &c->btree_write_buffer;
501 mutex_lock(&wb->flushing.lock);
503 ret = bch2_trans_run(c, bch2_btree_write_buffer_flush_locked(trans));
504 } while (!ret && bch2_btree_write_buffer_should_flush(c));
505 mutex_unlock(&wb->flushing.lock);
507 bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
510 int bch2_journal_key_to_wb_slowpath(struct bch_fs *c,
511 struct journal_keys_to_wb *dst,
512 enum btree_id btree, struct bkey_i *k)
514 struct btree_write_buffer *wb = &c->btree_write_buffer;
517 ret = darray_make_room_gfp(&dst->wb->keys, 1, GFP_KERNEL);
518 if (!ret && dst->wb == &wb->flushing)
519 ret = darray_resize(&wb->sorted, wb->flushing.keys.size);
522 if (dst->wb == &c->btree_write_buffer.flushing) {
523 mutex_unlock(&dst->wb->lock);
524 dst->wb = &c->btree_write_buffer.inc;
525 bch2_journal_pin_add(&c->journal, dst->seq, &dst->wb->pin,
526 bch2_btree_write_buffer_journal_flush);
533 dst->room = darray_room(dst->wb->keys);
534 if (dst->wb == &wb->flushing)
535 dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
539 struct btree_write_buffered_key *wb_k = &darray_top(dst->wb->keys);
540 wb_k->journal_seq = dst->seq;
542 bkey_copy(&wb_k->k, k);
548 void bch2_journal_keys_to_write_buffer_start(struct bch_fs *c, struct journal_keys_to_wb *dst, u64 seq)
550 struct btree_write_buffer *wb = &c->btree_write_buffer;
552 if (mutex_trylock(&wb->flushing.lock)) {
553 mutex_lock(&wb->inc.lock);
554 move_keys_from_inc_to_flushing(wb);
557 * Attempt to skip wb->inc, and add keys directly to
558 * wb->flushing, saving us a copy later:
561 if (!wb->inc.keys.nr) {
562 dst->wb = &wb->flushing;
564 mutex_unlock(&wb->flushing.lock);
568 mutex_lock(&wb->inc.lock);
572 dst->room = darray_room(dst->wb->keys);
573 if (dst->wb == &wb->flushing)
574 dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
577 bch2_journal_pin_add(&c->journal, seq, &dst->wb->pin,
578 bch2_btree_write_buffer_journal_flush);
581 void bch2_journal_keys_to_write_buffer_end(struct bch_fs *c, struct journal_keys_to_wb *dst)
583 struct btree_write_buffer *wb = &c->btree_write_buffer;
585 if (!dst->wb->keys.nr)
586 bch2_journal_pin_drop(&c->journal, &dst->wb->pin);
588 if (bch2_btree_write_buffer_should_flush(c) &&
589 __bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer) &&
590 !queue_work(system_unbound_wq, &c->btree_write_buffer.flush_work))
591 bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
593 if (dst->wb == &wb->flushing)
594 mutex_unlock(&wb->flushing.lock);
595 mutex_unlock(&wb->inc.lock);
598 static int bch2_journal_keys_to_write_buffer(struct bch_fs *c, struct journal_buf *buf)
600 struct journal_keys_to_wb dst;
603 bch2_journal_keys_to_write_buffer_start(c, &dst, le64_to_cpu(buf->data->seq));
605 for_each_jset_entry_type(entry, buf->data, BCH_JSET_ENTRY_write_buffer_keys) {
606 jset_entry_for_each_key(entry, k) {
607 ret = bch2_journal_key_to_wb(c, &dst, entry->btree_id, k);
612 entry->type = BCH_JSET_ENTRY_btree_keys;
615 spin_lock(&c->journal.lock);
616 buf->need_flush_to_write_buffer = false;
617 spin_unlock(&c->journal.lock);
619 bch2_journal_keys_to_write_buffer_end(c, &dst);
623 static int wb_keys_resize(struct btree_write_buffer_keys *wb, size_t new_size)
625 if (wb->keys.size >= new_size)
628 if (!mutex_trylock(&wb->lock))
631 int ret = darray_resize(&wb->keys, new_size);
632 mutex_unlock(&wb->lock);
636 int bch2_btree_write_buffer_resize(struct bch_fs *c, size_t new_size)
638 struct btree_write_buffer *wb = &c->btree_write_buffer;
640 return wb_keys_resize(&wb->flushing, new_size) ?:
641 wb_keys_resize(&wb->inc, new_size);
644 void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
646 struct btree_write_buffer *wb = &c->btree_write_buffer;
648 BUG_ON((wb->inc.keys.nr || wb->flushing.keys.nr) &&
649 !bch2_journal_error(&c->journal));
651 darray_exit(&wb->sorted);
652 darray_exit(&wb->flushing.keys);
653 darray_exit(&wb->inc.keys);
656 int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
658 struct btree_write_buffer *wb = &c->btree_write_buffer;
660 mutex_init(&wb->inc.lock);
661 mutex_init(&wb->flushing.lock);
662 INIT_WORK(&wb->flush_work, bch2_btree_write_buffer_flush_work);
664 /* Will be resized by journal as needed: */
665 unsigned initial_size = 1 << 16;
667 return darray_make_room(&wb->inc.keys, initial_size) ?:
668 darray_make_room(&wb->flushing.keys, initial_size) ?:
669 darray_make_room(&wb->sorted, initial_size);