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>
15 static int bch2_btree_write_buffer_journal_flush(struct journal *,
16 struct journal_entry_pin *, u64);
18 static int bch2_journal_keys_to_write_buffer(struct bch_fs *, struct journal_buf *);
20 static inline bool __wb_key_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
22 return (cmp_int(l->hi, r->hi) ?:
23 cmp_int(l->mi, r->mi) ?:
24 cmp_int(l->lo, r->lo)) >= 0;
27 static inline bool wb_key_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
32 asm("mov (%[l]), %%rax;"
36 "mov 16(%[l]), %%rax;"
37 "sbb 16(%[r]), %%rax;"
39 : [l] "r" (l), [r] "r" (r)
42 EBUG_ON(cmp != __wb_key_cmp(l, r));
45 return __wb_key_cmp(l, r);
49 /* Compare excluding idx, the low 24 bits: */
50 static inline bool wb_key_eq(const void *_l, const void *_r)
52 const struct wb_key_ref *l = _l;
53 const struct wb_key_ref *r = _r;
55 return !((l->hi ^ r->hi)|
57 ((l->lo >> 24) ^ (r->lo >> 24)));
60 static noinline void wb_sort(struct wb_key_ref *base, size_t num)
62 size_t n = num, a = num / 2;
64 if (!a) /* num < 2 || size == 0 */
70 if (a) /* Building heap: sift down --a */
72 else if (--n) /* Sorting: Extract root to --n */
73 swap(base[0], base[n]);
74 else /* Sort complete */
78 * Sift element at "a" down into heap. This is the
79 * "bottom-up" variant, which significantly reduces
80 * calls to cmp_func(): we find the sift-down path all
81 * the way to the leaves (one compare per level), then
82 * backtrack to find where to insert the target element.
84 * Because elements tend to sift down close to the leaves,
85 * this uses fewer compares than doing two per level
86 * on the way down. (A bit more than half as many on
87 * average, 3/4 worst-case.)
89 for (b = a; c = 2*b + 1, (d = c + 1) < n;)
90 b = wb_key_cmp(base + c, base + d) ? c : d;
91 if (d == n) /* Special case last leaf with no sibling */
94 /* Now backtrack from "b" to the correct location for "a" */
95 while (b != a && wb_key_cmp(base + a, base + b))
97 c = b; /* Where "a" belongs */
98 while (b != a) { /* Shift it into place */
100 swap(base[b], base[c]);
105 static noinline int wb_flush_one_slowpath(struct btree_trans *trans,
106 struct btree_iter *iter,
107 struct btree_write_buffered_key *wb)
109 struct btree_path *path = btree_iter_path(trans, iter);
111 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
113 trans->journal_res.seq = wb->journal_seq;
115 return bch2_trans_update(trans, iter, &wb->k,
116 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
117 bch2_trans_commit(trans, NULL, NULL,
118 BCH_TRANS_COMMIT_no_enospc|
119 BCH_TRANS_COMMIT_no_check_rw|
120 BCH_TRANS_COMMIT_no_journal_res|
121 BCH_TRANS_COMMIT_journal_reclaim);
124 static inline int wb_flush_one(struct btree_trans *trans, struct btree_iter *iter,
125 struct btree_write_buffered_key *wb,
126 bool *write_locked, size_t *fast)
128 struct bch_fs *c = trans->c;
129 struct btree_path *path;
132 EBUG_ON(!wb->journal_seq);
133 EBUG_ON(!c->btree_write_buffer.flushing.pin.seq);
134 EBUG_ON(c->btree_write_buffer.flushing.pin.seq > wb->journal_seq);
136 ret = bch2_btree_iter_traverse(iter);
141 * We can't clone a path that has write locks: unshare it now, before
142 * set_pos and traverse():
144 if (btree_iter_path(trans, iter)->ref > 1)
145 iter->path = __bch2_btree_path_make_mut(trans, iter->path, true, _THIS_IP_);
147 path = btree_iter_path(trans, iter);
149 if (!*write_locked) {
150 ret = bch2_btree_node_lock_write(trans, path, &path->l[0].b->c);
154 bch2_btree_node_prep_for_write(trans, path, path->l[0].b);
155 *write_locked = true;
158 if (unlikely(!bch2_btree_node_insert_fits(c, path->l[0].b, wb->k.k.u64s))) {
159 *write_locked = false;
160 return wb_flush_one_slowpath(trans, iter, wb);
163 bch2_btree_insert_key_leaf(trans, path, &wb->k, wb->journal_seq);
169 * Update a btree with a write buffered key using the journal seq of the
170 * original write buffer insert.
172 * It is not safe to rejournal the key once it has been inserted into the write
173 * buffer because that may break recovery ordering. For example, the key may
174 * have already been modified in the active write buffer in a seq that comes
175 * before the current transaction. If we were to journal this key again and
176 * crash, recovery would process updates in the wrong order.
179 btree_write_buffered_insert(struct btree_trans *trans,
180 struct btree_write_buffered_key *wb)
182 struct btree_iter iter;
185 bch2_trans_iter_init(trans, &iter, wb->btree, bkey_start_pos(&wb->k.k),
186 BTREE_ITER_CACHED|BTREE_ITER_INTENT);
188 trans->journal_res.seq = wb->journal_seq;
190 ret = bch2_btree_iter_traverse(&iter) ?:
191 bch2_trans_update(trans, &iter, &wb->k,
192 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
193 bch2_trans_iter_exit(trans, &iter);
197 static void move_keys_from_inc_to_flushing(struct btree_write_buffer *wb)
199 struct bch_fs *c = container_of(wb, struct bch_fs, btree_write_buffer);
200 struct journal *j = &c->journal;
202 if (!wb->inc.keys.nr)
205 bch2_journal_pin_add(j, wb->inc.keys.data[0].journal_seq, &wb->flushing.pin,
206 bch2_btree_write_buffer_journal_flush);
208 darray_resize(&wb->flushing.keys, min_t(size_t, 1U << 20, wb->flushing.keys.nr + wb->inc.keys.nr));
209 darray_resize(&wb->sorted, wb->flushing.keys.size);
211 if (!wb->flushing.keys.nr && wb->sorted.size >= wb->inc.keys.nr) {
212 swap(wb->flushing.keys, wb->inc.keys);
216 size_t nr = min(darray_room(wb->flushing.keys),
217 wb->sorted.size - wb->flushing.keys.nr);
218 nr = min(nr, wb->inc.keys.nr);
220 memcpy(&darray_top(wb->flushing.keys),
222 sizeof(wb->inc.keys.data[0]) * nr);
224 memmove(wb->inc.keys.data,
225 wb->inc.keys.data + nr,
226 sizeof(wb->inc.keys.data[0]) * (wb->inc.keys.nr - nr));
228 wb->flushing.keys.nr += nr;
229 wb->inc.keys.nr -= nr;
231 if (!wb->inc.keys.nr)
232 bch2_journal_pin_drop(j, &wb->inc.pin);
234 bch2_journal_pin_update(j, wb->inc.keys.data[0].journal_seq, &wb->inc.pin,
235 bch2_btree_write_buffer_journal_flush);
239 bch2_journal_set_watermark(j);
240 spin_unlock(&j->lock);
243 BUG_ON(wb->sorted.size < wb->flushing.keys.nr);
246 static int bch2_btree_write_buffer_flush_locked(struct btree_trans *trans)
248 struct bch_fs *c = trans->c;
249 struct journal *j = &c->journal;
250 struct btree_write_buffer *wb = &c->btree_write_buffer;
251 struct wb_key_ref *i;
252 struct btree_iter iter = { NULL };
253 size_t skipped = 0, fast = 0, slowpath = 0;
254 bool write_locked = false;
257 bch2_trans_unlock(trans);
258 bch2_trans_begin(trans);
260 mutex_lock(&wb->inc.lock);
261 move_keys_from_inc_to_flushing(wb);
262 mutex_unlock(&wb->inc.lock);
264 for (size_t i = 0; i < wb->flushing.keys.nr; i++) {
265 wb->sorted.data[i].idx = i;
266 wb->sorted.data[i].btree = wb->flushing.keys.data[i].btree;
267 memcpy(&wb->sorted.data[i].pos, &wb->flushing.keys.data[i].k.k.p, sizeof(struct bpos));
269 wb->sorted.nr = wb->flushing.keys.nr;
272 * We first sort so that we can detect and skip redundant updates, and
273 * then we attempt to flush in sorted btree order, as this is most
276 * However, since we're not flushing in the order they appear in the
277 * journal we won't be able to drop our journal pin until everything is
278 * flushed - which means this could deadlock the journal if we weren't
279 * passing BCH_TRANS_COMMIT_journal_reclaim. This causes the update to fail
280 * if it would block taking a journal reservation.
282 * If that happens, simply skip the key so we can optimistically insert
283 * as many keys as possible in the fast path.
285 wb_sort(wb->sorted.data, wb->sorted.nr);
287 darray_for_each(wb->sorted, i) {
288 struct btree_write_buffered_key *k = &wb->flushing.keys.data[i->idx];
290 for (struct wb_key_ref *n = i + 1; n < min(i + 4, &darray_top(wb->sorted)); n++)
291 prefetch(&wb->flushing.keys.data[n->idx]);
293 BUG_ON(!k->journal_seq);
295 if (i + 1 < &darray_top(wb->sorted) &&
296 wb_key_eq(i, i + 1)) {
297 struct btree_write_buffered_key *n = &wb->flushing.keys.data[i[1].idx];
300 n->journal_seq = min_t(u64, n->journal_seq, k->journal_seq);
306 struct btree_path *path = btree_iter_path(trans, &iter);
308 if (path->btree_id != i->btree ||
309 bpos_gt(k->k.k.p, path->l[0].b->key.k.p)) {
310 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
311 write_locked = false;
315 if (!iter.path || iter.btree_id != k->btree) {
316 bch2_trans_iter_exit(trans, &iter);
317 bch2_trans_iter_init(trans, &iter, k->btree, k->k.k.p,
318 BTREE_ITER_INTENT|BTREE_ITER_ALL_SNAPSHOTS);
321 bch2_btree_iter_set_pos(&iter, k->k.k.p);
322 btree_iter_path(trans, &iter)->preserve = false;
326 ret = -BCH_ERR_journal_reclaim_would_deadlock;
330 ret = wb_flush_one(trans, &iter, k, &write_locked, &fast);
332 bch2_trans_begin(trans);
333 } while (bch2_err_matches(ret, BCH_ERR_transaction_restart));
337 } else if (ret == -BCH_ERR_journal_reclaim_would_deadlock) {
345 struct btree_path *path = btree_iter_path(trans, &iter);
346 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
348 bch2_trans_iter_exit(trans, &iter);
355 * Flush in the order they were present in the journal, so that
356 * we can release journal pins:
357 * The fastpath zapped the seq of keys that were successfully flushed so
358 * we can skip those here.
360 trace_and_count(c, write_buffer_flush_slowpath, trans, slowpath, wb->flushing.keys.nr);
362 struct btree_write_buffered_key *i;
363 darray_for_each(wb->flushing.keys, i) {
367 bch2_journal_pin_update(j, i->journal_seq, &wb->flushing.pin,
368 bch2_btree_write_buffer_journal_flush);
370 bch2_trans_begin(trans);
372 ret = commit_do(trans, NULL, NULL,
373 BCH_WATERMARK_reclaim|
374 BCH_TRANS_COMMIT_no_check_rw|
375 BCH_TRANS_COMMIT_no_enospc|
376 BCH_TRANS_COMMIT_no_journal_res|
377 BCH_TRANS_COMMIT_journal_reclaim,
378 btree_write_buffered_insert(trans, i));
384 bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret));
385 trace_write_buffer_flush(trans, wb->flushing.keys.nr, skipped, fast, 0);
386 bch2_journal_pin_drop(j, &wb->flushing.pin);
387 wb->flushing.keys.nr = 0;
391 static int fetch_wb_keys_from_journal(struct bch_fs *c, u64 seq)
393 struct journal *j = &c->journal;
394 struct journal_buf *buf;
397 while (!ret && (buf = bch2_next_write_buffer_flush_journal_buf(j, seq))) {
398 ret = bch2_journal_keys_to_write_buffer(c, buf);
399 mutex_unlock(&j->buf_lock);
405 static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq)
407 struct bch_fs *c = trans->c;
408 struct btree_write_buffer *wb = &c->btree_write_buffer;
409 int ret = 0, fetch_from_journal_err;
412 bch2_trans_unlock(trans);
414 fetch_from_journal_err = fetch_wb_keys_from_journal(c, seq);
417 * On memory allocation failure, bch2_btree_write_buffer_flush_locked()
418 * is not guaranteed to empty wb->inc:
420 mutex_lock(&wb->flushing.lock);
421 ret = bch2_btree_write_buffer_flush_locked(trans);
422 mutex_unlock(&wb->flushing.lock);
424 (fetch_from_journal_err ||
425 (wb->inc.pin.seq && wb->inc.pin.seq <= seq) ||
426 (wb->flushing.pin.seq && wb->flushing.pin.seq <= seq)));
431 static int bch2_btree_write_buffer_journal_flush(struct journal *j,
432 struct journal_entry_pin *_pin, u64 seq)
434 struct bch_fs *c = container_of(j, struct bch_fs, journal);
436 return bch2_trans_run(c, btree_write_buffer_flush_seq(trans, seq));
439 int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
441 struct bch_fs *c = trans->c;
443 trace_and_count(c, write_buffer_flush_sync, trans, _RET_IP_);
445 return btree_write_buffer_flush_seq(trans, journal_cur_seq(&c->journal));
448 int bch2_btree_write_buffer_flush_nocheck_rw(struct btree_trans *trans)
450 struct bch_fs *c = trans->c;
451 struct btree_write_buffer *wb = &c->btree_write_buffer;
454 if (mutex_trylock(&wb->flushing.lock)) {
455 ret = bch2_btree_write_buffer_flush_locked(trans);
456 mutex_unlock(&wb->flushing.lock);
462 int bch2_btree_write_buffer_tryflush(struct btree_trans *trans)
464 struct bch_fs *c = trans->c;
466 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer))
467 return -BCH_ERR_erofs_no_writes;
469 int ret = bch2_btree_write_buffer_flush_nocheck_rw(trans);
470 bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
474 static void bch2_btree_write_buffer_flush_work(struct work_struct *work)
476 struct bch_fs *c = container_of(work, struct bch_fs, btree_write_buffer.flush_work);
477 struct btree_write_buffer *wb = &c->btree_write_buffer;
480 mutex_lock(&wb->flushing.lock);
482 ret = bch2_trans_run(c, bch2_btree_write_buffer_flush_locked(trans));
483 } while (!ret && bch2_btree_write_buffer_should_flush(c));
484 mutex_unlock(&wb->flushing.lock);
486 bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
489 int __bch2_journal_key_to_wb(struct bch_fs *c,
490 struct journal_keys_to_wb *dst,
491 enum btree_id btree, struct bkey_i *k)
493 struct btree_write_buffer *wb = &c->btree_write_buffer;
496 ret = darray_make_room_gfp(&dst->wb->keys, 1, GFP_KERNEL);
497 if (!ret && dst->wb == &wb->flushing)
498 ret = darray_resize(&wb->sorted, wb->flushing.keys.size);
501 if (dst->wb == &c->btree_write_buffer.flushing) {
502 mutex_unlock(&dst->wb->lock);
503 dst->wb = &c->btree_write_buffer.inc;
504 bch2_journal_pin_add(&c->journal, dst->seq, &dst->wb->pin,
505 bch2_btree_write_buffer_journal_flush);
512 dst->room = darray_room(dst->wb->keys);
513 if (dst->wb == &wb->flushing)
514 dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
518 struct btree_write_buffered_key *wb_k = &darray_top(dst->wb->keys);
519 wb_k->journal_seq = dst->seq;
521 bkey_copy(&wb_k->k, k);
527 void bch2_journal_keys_to_write_buffer_start(struct bch_fs *c, struct journal_keys_to_wb *dst, u64 seq)
529 struct btree_write_buffer *wb = &c->btree_write_buffer;
531 if (mutex_trylock(&wb->flushing.lock)) {
532 mutex_lock(&wb->inc.lock);
533 move_keys_from_inc_to_flushing(wb);
536 * Attempt to skip wb->inc, and add keys directly to
537 * wb->flushing, saving us a copy later:
540 if (!wb->inc.keys.nr) {
541 dst->wb = &wb->flushing;
543 mutex_unlock(&wb->flushing.lock);
547 mutex_lock(&wb->inc.lock);
551 dst->room = darray_room(dst->wb->keys);
552 if (dst->wb == &wb->flushing)
553 dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
556 bch2_journal_pin_add(&c->journal, seq, &dst->wb->pin,
557 bch2_btree_write_buffer_journal_flush);
560 void bch2_journal_keys_to_write_buffer_end(struct bch_fs *c, struct journal_keys_to_wb *dst)
562 struct btree_write_buffer *wb = &c->btree_write_buffer;
564 if (!dst->wb->keys.nr)
565 bch2_journal_pin_drop(&c->journal, &dst->wb->pin);
567 if (bch2_btree_write_buffer_should_flush(c) &&
568 __bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer) &&
569 !queue_work(system_unbound_wq, &c->btree_write_buffer.flush_work))
570 bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
572 if (dst->wb == &wb->flushing)
573 mutex_unlock(&wb->flushing.lock);
574 mutex_unlock(&wb->inc.lock);
577 static int bch2_journal_keys_to_write_buffer(struct bch_fs *c, struct journal_buf *buf)
579 struct journal_keys_to_wb dst;
580 struct jset_entry *entry;
584 bch2_journal_keys_to_write_buffer_start(c, &dst, le64_to_cpu(buf->data->seq));
586 for_each_jset_entry_type(entry, buf->data, BCH_JSET_ENTRY_write_buffer_keys) {
587 jset_entry_for_each_key(entry, k) {
588 ret = bch2_journal_key_to_wb(c, &dst, entry->btree_id, k);
593 entry->type = BCH_JSET_ENTRY_btree_keys;
596 buf->need_flush_to_write_buffer = false;
598 bch2_journal_keys_to_write_buffer_end(c, &dst);
602 static int wb_keys_resize(struct btree_write_buffer_keys *wb, size_t new_size)
604 if (wb->keys.size >= new_size)
607 if (!mutex_trylock(&wb->lock))
610 int ret = darray_resize(&wb->keys, new_size);
611 mutex_unlock(&wb->lock);
615 int bch2_btree_write_buffer_resize(struct bch_fs *c, size_t new_size)
617 struct btree_write_buffer *wb = &c->btree_write_buffer;
619 return wb_keys_resize(&wb->flushing, new_size) ?:
620 wb_keys_resize(&wb->inc, new_size);
623 void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
625 struct btree_write_buffer *wb = &c->btree_write_buffer;
627 BUG_ON((wb->inc.keys.nr || wb->flushing.keys.nr) &&
628 !bch2_journal_error(&c->journal));
630 darray_exit(&wb->sorted);
631 darray_exit(&wb->flushing.keys);
632 darray_exit(&wb->inc.keys);
635 int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
637 struct btree_write_buffer *wb = &c->btree_write_buffer;
639 mutex_init(&wb->inc.lock);
640 mutex_init(&wb->flushing.lock);
641 INIT_WORK(&wb->flush_work, bch2_btree_write_buffer_flush_work);
643 /* Will be resized by journal as needed: */
644 unsigned initial_size = 1 << 16;
646 return darray_make_room(&wb->inc.keys, initial_size) ?:
647 darray_make_room(&wb->flushing.keys, initial_size) ?:
648 darray_make_room(&wb->sorted, initial_size);