1 // SPDX-License-Identifier: GPL-2.0
4 #include "btree_update.h"
5 #include "btree_update_interior.h"
8 #include "btree_iter.h"
9 #include "btree_key_cache.h"
10 #include "btree_locking.h"
14 #include "extent_update.h"
16 #include "journal_reclaim.h"
19 #include "subvolume.h"
22 #include <linux/prefetch.h>
23 #include <linux/sort.h>
24 #include <trace/events/bcachefs.h>
26 static int __must_check
27 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
28 struct bkey_i *, enum btree_update_flags);
30 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
31 const struct btree_insert_entry *r)
33 return cmp_int(l->btree_id, r->btree_id) ?:
34 -cmp_int(l->level, r->level) ?:
35 bpos_cmp(l->k->k.p, r->k->k.p);
38 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
40 return i->path->l + i->level;
43 static inline bool same_leaf_as_prev(struct btree_trans *trans,
44 struct btree_insert_entry *i)
46 return i != trans->updates &&
47 insert_l(&i[0])->b == insert_l(&i[-1])->b;
50 static inline bool same_leaf_as_next(struct btree_trans *trans,
51 struct btree_insert_entry *i)
53 return i + 1 < trans->updates + trans->nr_updates &&
54 insert_l(&i[0])->b == insert_l(&i[1])->b;
57 static inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
58 struct btree_path *path,
61 struct bch_fs *c = trans->c;
66 if (unlikely(btree_node_just_written(b)) &&
67 bch2_btree_post_write_cleanup(c, b))
68 bch2_trans_node_reinit_iter(trans, b);
71 * If the last bset has been written, or if it's gotten too big - start
72 * a new bset to insert into:
74 if (want_new_bset(c, b))
75 bch2_btree_init_next(trans, b);
78 void bch2_btree_node_lock_for_insert(struct btree_trans *trans,
79 struct btree_path *path,
82 bch2_btree_node_lock_write(trans, path, b);
83 bch2_btree_node_prep_for_write(trans, path, b);
86 /* Inserting into a given leaf node (last stage of insert): */
88 /* Handle overwrites and do insert, for non extents: */
89 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
90 struct btree_path *path,
92 struct btree_node_iter *node_iter,
93 struct bkey_i *insert)
95 struct bkey_packed *k;
96 unsigned clobber_u64s = 0, new_u64s = 0;
98 EBUG_ON(btree_node_just_written(b));
99 EBUG_ON(bset_written(b, btree_bset_last(b)));
100 EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
101 EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
102 EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
103 EBUG_ON(insert->k.u64s >
104 bch_btree_keys_u64s_remaining(trans->c, b));
106 k = bch2_btree_node_iter_peek_all(node_iter, b);
107 if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
110 /* @k is the key being overwritten/deleted, if any: */
111 EBUG_ON(k && bkey_deleted(k));
113 /* Deleting, but not found? nothing to do: */
114 if (bkey_deleted(&insert->k) && !k)
117 if (bkey_deleted(&insert->k)) {
119 btree_account_key_drop(b, k);
120 k->type = KEY_TYPE_deleted;
122 if (k->needs_whiteout)
123 push_whiteout(trans->c, b, insert->k.p);
124 k->needs_whiteout = false;
126 if (k >= btree_bset_last(b)->start) {
127 clobber_u64s = k->u64s;
128 bch2_bset_delete(b, k, clobber_u64s);
131 bch2_btree_path_fix_key_modified(trans, b, k);
139 btree_account_key_drop(b, k);
140 k->type = KEY_TYPE_deleted;
142 insert->k.needs_whiteout = k->needs_whiteout;
143 k->needs_whiteout = false;
145 if (k >= btree_bset_last(b)->start) {
146 clobber_u64s = k->u64s;
149 bch2_btree_path_fix_key_modified(trans, b, k);
153 k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
155 bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
158 if (clobber_u64s != new_u64s)
159 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
160 clobber_u64s, new_u64s);
164 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
167 struct bch_fs *c = container_of(j, struct bch_fs, journal);
168 struct btree_write *w = container_of(pin, struct btree_write, journal);
169 struct btree *b = container_of(w, struct btree, writes[i]);
170 unsigned long old, new, v;
171 unsigned idx = w - b->writes;
173 six_lock_read(&b->c.lock, NULL, NULL);
174 v = READ_ONCE(b->flags);
179 if (!(old & (1 << BTREE_NODE_dirty)) ||
180 !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
181 w->journal.seq != seq)
184 new |= 1 << BTREE_NODE_need_write;
185 } while ((v = cmpxchg(&b->flags, old, new)) != old);
187 btree_node_write_if_need(c, b, SIX_LOCK_read);
188 six_unlock_read(&b->c.lock);
192 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
194 return __btree_node_flush(j, pin, 0, seq);
197 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
199 return __btree_node_flush(j, pin, 1, seq);
202 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
203 struct btree *b, u64 seq)
205 struct btree_write *w = btree_current_write(b);
207 bch2_journal_pin_add(&c->journal, seq, &w->journal,
208 btree_node_write_idx(b) == 0
210 : btree_node_flush1);
214 * btree_insert_key - insert a key one key into a leaf node
216 static bool btree_insert_key_leaf(struct btree_trans *trans,
217 struct btree_insert_entry *insert)
219 struct bch_fs *c = trans->c;
220 struct btree *b = insert_l(insert)->b;
221 struct bset_tree *t = bset_tree_last(b);
222 struct bset *i = bset(b, t);
223 int old_u64s = bset_u64s(t);
224 int old_live_u64s = b->nr.live_u64s;
225 int live_u64s_added, u64s_added;
227 if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
228 &insert_l(insert)->iter, insert->k)))
231 i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
232 le64_to_cpu(i->journal_seq)));
234 bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
236 if (unlikely(!btree_node_dirty(b)))
237 set_btree_node_dirty_acct(c, b);
239 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
240 u64s_added = (int) bset_u64s(t) - old_u64s;
242 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
243 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
244 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
245 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
247 if (u64s_added > live_u64s_added &&
248 bch2_maybe_compact_whiteouts(c, b))
249 bch2_trans_node_reinit_iter(trans, b);
254 /* Cached btree updates: */
256 /* Normal update interface: */
258 static inline void btree_insert_entry_checks(struct btree_trans *trans,
259 struct btree_insert_entry *i)
261 BUG_ON(bpos_cmp(i->k->k.p, i->path->pos));
262 BUG_ON(i->cached != i->path->cached);
263 BUG_ON(i->level != i->path->level);
264 BUG_ON(i->btree_id != i->path->btree_id);
266 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
267 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
268 i->k->k.p.snapshot &&
269 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
273 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
274 unsigned long trace_ip)
276 struct bch_fs *c = trans->c;
279 bch2_trans_unlock(trans);
281 ret = bch2_journal_preres_get(&c->journal,
282 &trans->journal_preres, u64s, 0);
286 if (!bch2_trans_relock(trans)) {
287 trace_trans_restart_journal_preres_get(trans->fn, trace_ip);
294 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
297 struct bch_fs *c = trans->c;
300 if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
301 flags |= JOURNAL_RES_GET_RESERVED;
303 ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
304 trans->journal_u64s, flags);
306 return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
309 #define JSET_ENTRY_LOG_U64s 4
311 static noinline void journal_transaction_name(struct btree_trans *trans)
313 struct bch_fs *c = trans->c;
314 struct jset_entry *entry = journal_res_entry(&c->journal, &trans->journal_res);
315 struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
316 unsigned u64s = JSET_ENTRY_LOG_U64s - 1;
317 unsigned b, buflen = u64s * sizeof(u64);
319 l->entry.u64s = cpu_to_le16(u64s);
320 l->entry.btree_id = 0;
322 l->entry.type = BCH_JSET_ENTRY_log;
326 b = min_t(unsigned, strlen(trans->fn), buflen);
327 memcpy(l->d, trans->fn, b);
331 trans->journal_res.offset += JSET_ENTRY_LOG_U64s;
332 trans->journal_res.u64s -= JSET_ENTRY_LOG_U64s;
335 static inline enum btree_insert_ret
336 btree_key_can_insert(struct btree_trans *trans,
340 struct bch_fs *c = trans->c;
342 if (!bch2_btree_node_insert_fits(c, b, u64s))
343 return BTREE_INSERT_BTREE_NODE_FULL;
345 return BTREE_INSERT_OK;
348 static enum btree_insert_ret
349 btree_key_can_insert_cached(struct btree_trans *trans,
350 struct btree_path *path,
353 struct bch_fs *c = trans->c;
354 struct bkey_cached *ck = (void *) path->l[0].b;
356 struct bkey_i *new_k;
358 EBUG_ON(path->level);
360 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
361 bch2_btree_key_cache_must_wait(c) &&
362 !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
363 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
366 * bch2_varint_decode can read past the end of the buffer by at most 7
367 * bytes (it won't be used):
371 if (u64s <= ck->u64s)
372 return BTREE_INSERT_OK;
374 new_u64s = roundup_pow_of_two(u64s);
375 new_k = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
377 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
378 bch2_btree_ids[path->btree_id], new_u64s);
385 * Keys returned by peek() are no longer valid pointers, so we need a
386 * transaction restart:
388 trace_trans_restart_key_cache_key_realloced(trans->fn, _RET_IP_,
389 path->btree_id, &path->pos);
391 * Not using btree_trans_restart() because we can't unlock here, we have
394 trans->restarted = true;
398 static inline void do_btree_insert_one(struct btree_trans *trans,
399 struct btree_insert_entry *i)
401 struct bch_fs *c = trans->c;
402 struct journal *j = &c->journal;
405 EBUG_ON(trans->journal_res.ref !=
406 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
408 i->k->k.needs_whiteout = false;
410 did_work = !i->cached
411 ? btree_insert_key_leaf(trans, i)
412 : bch2_btree_insert_key_cached(trans, i->path, i->k);
416 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
417 bch2_journal_add_keys(j, &trans->journal_res,
422 if (trans->journal_seq)
423 *trans->journal_seq = trans->journal_res.seq;
429 static int run_one_mem_trigger(struct btree_trans *trans,
430 struct btree_insert_entry *i,
433 struct bkey_s_c old = { &i->old_k, i->old_v };
434 struct bkey_i *new = i->k;
437 if (unlikely(flags & BTREE_TRIGGER_NORUN))
440 if (!btree_node_type_needs_gc(i->btree_id))
443 if (old.k->type == new->k.type &&
444 ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
445 ret = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
446 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
448 struct bkey _deleted = KEY(0, 0, 0);
449 struct bkey_s_c deleted = (struct bkey_s_c) { &_deleted, NULL };
451 _deleted.p = i->path->pos;
453 ret = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
454 BTREE_TRIGGER_INSERT|flags) ?:
455 bch2_mark_key(trans, old, deleted,
456 BTREE_TRIGGER_OVERWRITE|flags);
462 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
465 struct bkey_s_c old = { &i->old_k, i->old_v };
468 if ((i->flags & BTREE_TRIGGER_NORUN) ||
469 !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
473 if (i->insert_trigger_run)
476 BUG_ON(i->overwrite_trigger_run);
477 i->insert_trigger_run = true;
479 if (i->overwrite_trigger_run)
482 BUG_ON(!i->insert_trigger_run);
483 i->overwrite_trigger_run = true;
487 ret = bch2_trans_mark_old(trans, old, i->flags);
488 } else if (old.k->type == i->k->k.type &&
489 ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
490 i->overwrite_trigger_run = true;
491 ret = bch2_trans_mark_key(trans, old, i->k,
492 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|i->flags);
494 ret = bch2_trans_mark_new(trans, i->k, i->flags);
498 trace_trans_restart_mark(trans->fn, _RET_IP_,
499 i->btree_id, &i->path->pos);
503 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
504 struct btree_insert_entry *btree_id_start)
506 struct btree_insert_entry *i;
507 bool trans_trigger_run;
510 for (overwrite = 0; overwrite < 2; overwrite++) {
513 * Running triggers will append more updates to the list of updates as
517 trans_trigger_run = false;
519 for (i = btree_id_start;
520 i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
522 if (i->btree_id != btree_id)
525 ret = run_one_trans_trigger(trans, i, overwrite);
529 trans_trigger_run = true;
531 } while (trans_trigger_run);
537 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
539 struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
540 unsigned btree_id = 0;
545 * For a given btree, this algorithm runs insert triggers before
546 * overwrite triggers: this is so that when extents are being moved
547 * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
550 for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
551 if (btree_id == BTREE_ID_alloc)
554 while (btree_id_start < trans->updates + trans->nr_updates &&
555 btree_id_start->btree_id < btree_id)
558 ret = run_btree_triggers(trans, btree_id, btree_id_start);
563 trans_for_each_update(trans, i) {
564 if (i->btree_id > BTREE_ID_alloc)
566 if (i->btree_id == BTREE_ID_alloc) {
567 ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
574 trans_for_each_update(trans, i)
575 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
576 (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
577 (!i->insert_trigger_run || !i->overwrite_trigger_run));
582 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
584 struct bch_fs *c = trans->c;
585 struct btree_insert_entry *i;
588 trans_for_each_update(trans, i) {
590 * XXX: synchronization of cached update triggers with gc
591 * XXX: synchronization of interior node updates with gc
593 BUG_ON(i->cached || i->level);
595 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
596 ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
606 bch2_trans_commit_write_locked(struct btree_trans *trans,
607 struct btree_insert_entry **stopped_at,
608 unsigned long trace_ip)
610 struct bch_fs *c = trans->c;
611 struct btree_insert_entry *i;
612 struct btree_trans_commit_hook *h;
614 bool marking = false;
618 trace_trans_restart_fault_inject(trans->fn, trace_ip);
619 trans->restarted = true;
624 * Check if the insert will fit in the leaf node with the write lock
625 * held, otherwise another thread could write the node changing the
626 * amount of space available:
629 prefetch(&trans->c->journal.flags);
633 ret = h->fn(trans, h);
639 trans_for_each_update(trans, i) {
640 /* Multiple inserts might go to same leaf: */
641 if (!same_leaf_as_prev(trans, i))
644 u64s += i->k->k.u64s;
646 ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
647 : btree_key_can_insert_cached(trans, i->path, u64s);
653 if (btree_node_type_needs_gc(i->bkey_type))
658 * Don't get journal reservation until after we know insert will
661 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
662 ret = bch2_trans_journal_res_get(trans,
663 JOURNAL_RES_GET_NONBLOCK);
667 if (unlikely(trans->journal_transaction_names))
668 journal_transaction_name(trans);
670 trans->journal_res.seq = c->journal.replay_journal_seq;
673 if (unlikely(trans->extra_journal_entry_u64s)) {
674 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
675 trans->extra_journal_entries,
676 trans->extra_journal_entry_u64s);
678 trans->journal_res.offset += trans->extra_journal_entry_u64s;
679 trans->journal_res.u64s -= trans->extra_journal_entry_u64s;
683 * Not allowed to fail after we've gotten our journal reservation - we
687 if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
688 if (bch2_journal_seq_verify)
689 trans_for_each_update(trans, i)
690 i->k->k.version.lo = trans->journal_res.seq;
691 else if (bch2_inject_invalid_keys)
692 trans_for_each_update(trans, i)
693 i->k->k.version = MAX_VERSION;
696 if (trans->fs_usage_deltas &&
697 bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
698 return BTREE_INSERT_NEED_MARK_REPLICAS;
700 trans_for_each_update(trans, i)
701 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
702 ret = run_one_mem_trigger(trans, i, i->flags);
707 if (unlikely(c->gc_pos.phase)) {
708 ret = bch2_trans_commit_run_gc_triggers(trans);
713 trans_for_each_update(trans, i)
714 do_btree_insert_one(trans, i);
719 static inline void path_upgrade_readers(struct btree_trans *trans, struct btree_path *path)
723 for (l = 0; l < BTREE_MAX_DEPTH; l++)
724 if (btree_node_read_locked(path, l))
725 BUG_ON(!bch2_btree_node_upgrade(trans, path, l));
728 static inline void upgrade_readers(struct btree_trans *trans, struct btree_path *path)
730 struct btree *b = path_l(path)->b;
733 if (path->nodes_locked &&
734 path->nodes_locked != path->nodes_intent_locked)
735 path_upgrade_readers(trans, path);
736 } while ((path = prev_btree_path(trans, path)) &&
737 path_l(path)->b == b);
741 * Check for nodes that we have both read and intent locks on, and upgrade the
744 static inline void normalize_read_intent_locks(struct btree_trans *trans)
746 struct btree_path *path;
747 unsigned i, nr_read = 0, nr_intent = 0;
749 trans_for_each_path_inorder(trans, path, i) {
750 struct btree_path *next = i + 1 < trans->nr_sorted
751 ? trans->paths + trans->sorted[i + 1]
754 if (path->nodes_locked) {
755 if (path->nodes_intent_locked)
761 if (!next || path_l(path)->b != path_l(next)->b) {
762 if (nr_read && nr_intent)
763 upgrade_readers(trans, path);
765 nr_read = nr_intent = 0;
769 bch2_trans_verify_locks(trans);
772 static inline bool have_conflicting_read_lock(struct btree_trans *trans, struct btree_path *pos)
774 struct btree_path *path;
777 trans_for_each_path_inorder(trans, path, i) {
781 if (path->nodes_locked != path->nodes_intent_locked &&
782 !bch2_btree_path_upgrade(trans, path, path->level + 1))
789 static inline int trans_lock_write(struct btree_trans *trans)
791 struct btree_insert_entry *i;
793 trans_for_each_update(trans, i) {
794 if (same_leaf_as_prev(trans, i))
797 if (!six_trylock_write(&insert_l(i)->b->c.lock)) {
798 if (have_conflicting_read_lock(trans, i->path))
801 btree_node_lock_type(trans, i->path,
803 i->path->pos, i->level,
804 SIX_LOCK_write, NULL, NULL);
807 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
812 while (--i >= trans->updates) {
813 if (same_leaf_as_prev(trans, i))
816 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
819 trace_trans_restart_would_deadlock_write(trans->fn);
820 return btree_trans_restart(trans);
823 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
825 struct btree_insert_entry *i;
827 trans_for_each_update(trans, i)
828 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
832 * Get journal reservation, take write locks, and attempt to do btree update(s):
834 static inline int do_bch2_trans_commit(struct btree_trans *trans,
835 struct btree_insert_entry **stopped_at,
836 unsigned long trace_ip)
838 struct bch_fs *c = trans->c;
839 struct btree_insert_entry *i;
840 int ret, u64s_delta = 0;
842 trans_for_each_update(trans, i) {
843 const char *invalid = bch2_bkey_invalid(c,
844 bkey_i_to_s_c(i->k), i->bkey_type);
846 struct printbuf buf = PRINTBUF;
848 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
849 bch2_fs_fatal_error(c, "invalid bkey %s on insert from %s -> %ps: %s\n",
850 buf.buf, trans->fn, (void *) i->ip_allocated, invalid);
854 btree_insert_entry_checks(trans, i);
857 trans_for_each_update(trans, i) {
861 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
862 u64s_delta -= i->old_btree_u64s;
864 if (!same_leaf_as_next(trans, i)) {
865 if (u64s_delta <= 0) {
866 ret = bch2_foreground_maybe_merge(trans, i->path,
867 i->level, trans->flags);
876 ret = bch2_journal_preres_get(&c->journal,
877 &trans->journal_preres, trans->journal_preres_u64s,
878 JOURNAL_RES_GET_NONBLOCK|
879 ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
880 ? JOURNAL_RES_GET_RESERVED : 0));
881 if (unlikely(ret == -EAGAIN))
882 ret = bch2_trans_journal_preres_get_cold(trans,
883 trans->journal_preres_u64s, trace_ip);
887 normalize_read_intent_locks(trans);
889 ret = trans_lock_write(trans);
893 ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
895 if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
896 bch2_drop_overwrites_from_journal(trans);
898 trans_for_each_update(trans, i)
899 if (!same_leaf_as_prev(trans, i))
900 bch2_btree_node_unlock_write_inlined(trans, i->path,
903 if (!ret && trans->journal_pin)
904 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
905 trans->journal_pin, NULL);
908 * Drop journal reservation after dropping write locks, since dropping
909 * the journal reservation may kick off a journal write:
911 bch2_journal_res_put(&c->journal, &trans->journal_res);
916 bch2_trans_downgrade(trans);
921 static int journal_reclaim_wait_done(struct bch_fs *c)
923 int ret = bch2_journal_error(&c->journal) ?:
924 !bch2_btree_key_cache_must_wait(c);
927 journal_reclaim_kick(&c->journal);
932 int bch2_trans_commit_error(struct btree_trans *trans,
933 struct btree_insert_entry *i,
934 int ret, unsigned long trace_ip)
936 struct bch_fs *c = trans->c;
939 case BTREE_INSERT_BTREE_NODE_FULL:
940 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
945 trace_trans_restart_btree_node_split(trans->fn, trace_ip,
946 i->btree_id, &i->path->pos);
948 case BTREE_INSERT_NEED_MARK_REPLICAS:
949 bch2_trans_unlock(trans);
951 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
955 if (bch2_trans_relock(trans))
958 trace_trans_restart_mark_replicas(trans->fn, trace_ip);
961 case BTREE_INSERT_NEED_JOURNAL_RES:
962 bch2_trans_unlock(trans);
964 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
965 !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED)) {
966 trans->restarted = true;
971 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
975 if (bch2_trans_relock(trans))
978 trace_trans_restart_journal_res_get(trans->fn, trace_ip);
981 case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
982 bch2_trans_unlock(trans);
984 trace_trans_blocked_journal_reclaim(trans->fn, trace_ip);
986 wait_event_freezable(c->journal.reclaim_wait,
987 (ret = journal_reclaim_wait_done(c)));
991 if (bch2_trans_relock(trans))
994 trace_trans_restart_journal_reclaim(trans->fn, trace_ip);
1002 BUG_ON((ret == EINTR || ret == -EAGAIN) && !trans->restarted);
1003 BUG_ON(ret == -ENOSPC &&
1004 !(trans->flags & BTREE_INSERT_NOWAIT) &&
1005 (trans->flags & BTREE_INSERT_NOFAIL));
1011 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
1013 struct bch_fs *c = trans->c;
1016 if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
1017 test_bit(BCH_FS_STARTED, &c->flags))
1020 bch2_trans_unlock(trans);
1022 ret = bch2_fs_read_write_early(c);
1026 if (!bch2_trans_relock(trans))
1029 percpu_ref_get(&c->writes);
1034 * This is for updates done in the early part of fsck - btree_gc - before we've
1035 * gone RW. we only add the new key to the list of keys for journal replay to
1039 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1041 struct bch_fs *c = trans->c;
1042 struct btree_insert_entry *i;
1045 trans_for_each_update(trans, i) {
1046 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1054 int __bch2_trans_commit(struct btree_trans *trans)
1056 struct bch_fs *c = trans->c;
1057 struct btree_insert_entry *i = NULL;
1061 if (!trans->nr_updates &&
1062 !trans->extra_journal_entry_u64s)
1065 if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
1066 lockdep_assert_held(&c->gc_lock);
1068 ret = bch2_trans_commit_run_triggers(trans);
1072 if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1073 ret = do_bch2_trans_commit_to_journal_replay(trans);
1077 if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1078 unlikely(!percpu_ref_tryget(&c->writes))) {
1079 ret = bch2_trans_commit_get_rw_cold(trans);
1084 memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1086 trans->journal_u64s = trans->extra_journal_entry_u64s;
1087 trans->journal_preres_u64s = 0;
1089 trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1091 if (trans->journal_transaction_names)
1092 trans->journal_u64s += JSET_ENTRY_LOG_U64s;
1094 trans_for_each_update(trans, i) {
1095 BUG_ON(!i->path->should_be_locked);
1097 if (unlikely(!bch2_btree_path_upgrade(trans, i->path, i->level + 1))) {
1098 trace_trans_restart_upgrade(trans->fn, _RET_IP_,
1099 i->btree_id, &i->path->pos);
1100 ret = btree_trans_restart(trans);
1104 BUG_ON(!btree_node_intent_locked(i->path, i->level));
1106 u64s = jset_u64s(i->k->k.u64s);
1108 likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1109 trans->journal_preres_u64s += u64s;
1110 trans->journal_u64s += u64s;
1113 if (trans->extra_journal_res) {
1114 ret = bch2_disk_reservation_add(c, trans->disk_res,
1115 trans->extra_journal_res,
1116 (trans->flags & BTREE_INSERT_NOFAIL)
1117 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1122 BUG_ON(trans->restarted);
1123 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1125 ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1127 /* make sure we didn't drop or screw up locks: */
1128 bch2_trans_verify_locks(trans);
1133 bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1135 if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1136 percpu_ref_put(&c->writes);
1138 trans_for_each_update(trans, i)
1139 bch2_path_put(trans, i->path, true);
1141 trans->extra_journal_res = 0;
1142 trans->nr_updates = 0;
1143 trans->hooks = NULL;
1144 trans->extra_journal_entries = NULL;
1145 trans->extra_journal_entry_u64s = 0;
1147 if (trans->fs_usage_deltas) {
1148 trans->fs_usage_deltas->used = 0;
1149 memset(&trans->fs_usage_deltas->memset_start, 0,
1150 (void *) &trans->fs_usage_deltas->memset_end -
1151 (void *) &trans->fs_usage_deltas->memset_start);
1156 ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1163 static int check_pos_snapshot_overwritten(struct btree_trans *trans,
1167 struct bch_fs *c = trans->c;
1168 struct btree_iter iter;
1172 if (!btree_type_has_snapshots(id))
1175 if (!snapshot_t(c, pos.snapshot)->children[0])
1178 bch2_trans_iter_init(trans, &iter, id, pos,
1179 BTREE_ITER_NOT_EXTENTS|
1180 BTREE_ITER_ALL_SNAPSHOTS);
1182 k = bch2_btree_iter_prev(&iter);
1190 if (bkey_cmp(pos, k.k->p))
1193 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1198 bch2_trans_iter_exit(trans, &iter);
1203 int bch2_trans_update_extent(struct btree_trans *trans,
1204 struct btree_iter *orig_iter,
1205 struct bkey_i *insert,
1206 enum btree_update_flags flags)
1208 struct bch_fs *c = trans->c;
1209 struct btree_iter iter, update_iter;
1210 struct bpos start = bkey_start_pos(&insert->k);
1211 struct bkey_i *update;
1213 enum btree_id btree_id = orig_iter->btree_id;
1214 int ret = 0, compressed_sectors;
1216 bch2_trans_iter_init(trans, &iter, btree_id, start,
1218 BTREE_ITER_WITH_UPDATES|
1219 BTREE_ITER_NOT_EXTENTS);
1220 k = bch2_btree_iter_peek(&iter);
1221 if ((ret = bkey_err(k)))
1226 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1228 * We can't merge extents if they belong to interior snapshot
1229 * tree nodes, and there's a snapshot in which one extent is
1230 * visible and the other is not - i.e. if visibility is
1233 * Instead of checking if visibilitiy of the two extents is
1234 * different, for now we just check if either has been
1237 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1243 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1249 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1250 if ((ret = PTR_ERR_OR_ZERO(update)))
1253 bkey_reassemble(update, k);
1255 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1256 ret = bch2_btree_delete_at(trans, &iter, flags);
1266 if (!bkey_cmp(k.k->p, start))
1269 while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1270 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1271 bool back_split = bkey_cmp(k.k->p, insert->k.p) > 0;
1274 * If we're going to be splitting a compressed extent, note it
1275 * so that __bch2_trans_commit() can increase our disk
1278 if (((front_split && back_split) ||
1279 ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1280 (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1281 trans->extra_journal_res += compressed_sectors;
1284 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1285 if ((ret = PTR_ERR_OR_ZERO(update)))
1288 bkey_reassemble(update, k);
1290 bch2_cut_back(start, update);
1292 bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1293 BTREE_ITER_NOT_EXTENTS|
1294 BTREE_ITER_ALL_SNAPSHOTS|
1296 ret = bch2_btree_iter_traverse(&update_iter) ?:
1297 bch2_trans_update(trans, &update_iter, update,
1298 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1300 bch2_trans_iter_exit(trans, &update_iter);
1306 if (k.k->p.snapshot != insert->k.p.snapshot &&
1307 (front_split || back_split)) {
1308 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1309 if ((ret = PTR_ERR_OR_ZERO(update)))
1312 bkey_reassemble(update, k);
1314 bch2_cut_front(start, update);
1315 bch2_cut_back(insert->k.p, update);
1317 bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1318 BTREE_ITER_NOT_EXTENTS|
1319 BTREE_ITER_ALL_SNAPSHOTS|
1321 ret = bch2_btree_iter_traverse(&update_iter) ?:
1322 bch2_trans_update(trans, &update_iter, update,
1323 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1325 bch2_trans_iter_exit(trans, &update_iter);
1330 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1331 update = bch2_trans_kmalloc(trans, sizeof(*update));
1332 if ((ret = PTR_ERR_OR_ZERO(update)))
1335 bkey_init(&update->k);
1336 update->k.p = k.k->p;
1338 if (insert->k.p.snapshot != k.k->p.snapshot) {
1339 update->k.p.snapshot = insert->k.p.snapshot;
1340 update->k.type = KEY_TYPE_whiteout;
1343 bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1344 BTREE_ITER_NOT_EXTENTS|
1346 ret = bch2_btree_iter_traverse(&update_iter) ?:
1347 bch2_trans_update(trans, &update_iter, update,
1348 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1350 bch2_trans_iter_exit(trans, &update_iter);
1357 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1358 if ((ret = PTR_ERR_OR_ZERO(update)))
1361 bkey_reassemble(update, k);
1362 bch2_cut_front(insert->k.p, update);
1364 ret = bch2_trans_update_by_path(trans, iter.path, update,
1365 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1372 k = bch2_btree_iter_next(&iter);
1373 if ((ret = bkey_err(k)))
1379 if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1380 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1386 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1392 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1397 if (!bkey_deleted(&insert->k)) {
1399 * Rewinding iterators is expensive: get a new one and the one
1400 * that points to the start of insert will be cloned from:
1402 bch2_trans_iter_exit(trans, &iter);
1403 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1404 BTREE_ITER_NOT_EXTENTS|
1406 ret = bch2_btree_iter_traverse(&iter) ?:
1407 bch2_trans_update(trans, &iter, insert, flags);
1410 bch2_trans_iter_exit(trans, &iter);
1416 * When deleting, check if we need to emit a whiteout (because we're overwriting
1417 * something in an ancestor snapshot)
1419 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1420 enum btree_id btree_id, struct bpos pos)
1422 struct btree_iter iter;
1424 u32 snapshot = pos.snapshot;
1427 if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1432 for_each_btree_key_norestart(trans, iter, btree_id, pos,
1433 BTREE_ITER_ALL_SNAPSHOTS|
1434 BTREE_ITER_NOPRESERVE, k, ret) {
1435 if (bkey_cmp(k.k->p, pos))
1438 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1440 ret = !bkey_whiteout(k.k);
1444 bch2_trans_iter_exit(trans, &iter);
1449 static int __must_check
1450 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1451 struct bkey_i *k, enum btree_update_flags flags)
1453 struct bch_fs *c = trans->c;
1454 struct btree_insert_entry *i, n;
1456 BUG_ON(!path->should_be_locked);
1458 BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1459 BUG_ON(bpos_cmp(k->k.p, path->pos));
1461 n = (struct btree_insert_entry) {
1463 .bkey_type = __btree_node_type(path->level, path->btree_id),
1464 .btree_id = path->btree_id,
1465 .level = path->level,
1466 .cached = path->cached,
1469 .ip_allocated = _RET_IP_,
1472 #ifdef CONFIG_BCACHEFS_DEBUG
1473 trans_for_each_update(trans, i)
1474 BUG_ON(i != trans->updates &&
1475 btree_insert_entry_cmp(i - 1, i) >= 0);
1479 * Pending updates are kept sorted: first, find position of new update,
1480 * then delete/trim any updates the new update overwrites:
1482 trans_for_each_update(trans, i)
1483 if (btree_insert_entry_cmp(&n, i) <= 0)
1486 if (i < trans->updates + trans->nr_updates &&
1487 !btree_insert_entry_cmp(&n, i)) {
1488 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1490 bch2_path_put(trans, i->path, true);
1492 i->cached = n.cached;
1495 i->ip_allocated = n.ip_allocated;
1497 array_insert_item(trans->updates, trans->nr_updates,
1498 i - trans->updates, n);
1500 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1501 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1503 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
1504 struct bkey_i *j_k =
1505 bch2_journal_keys_peek(c, n.btree_id, n.level, k->k.p);
1507 if (j_k && !bpos_cmp(j_k->k.p, i->k->k.p)) {
1514 __btree_path_get(n.path, true);
1518 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1519 struct bkey_i *k, enum btree_update_flags flags)
1521 struct btree_path *path = iter->update_path ?: iter->path;
1522 struct bkey_cached *ck;
1525 if (iter->flags & BTREE_ITER_IS_EXTENTS)
1526 return bch2_trans_update_extent(trans, iter, k, flags);
1528 if (bkey_deleted(&k->k) &&
1529 !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1530 (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1531 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1532 if (unlikely(ret < 0))
1536 k->k.type = KEY_TYPE_whiteout;
1539 if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1542 btree_id_cached(trans->c, path->btree_id)) {
1543 if (!iter->key_cache_path ||
1544 !iter->key_cache_path->should_be_locked ||
1545 bpos_cmp(iter->key_cache_path->pos, k->k.p)) {
1546 if (!iter->key_cache_path)
1547 iter->key_cache_path =
1548 bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1550 BTREE_ITER_CACHED, _THIS_IP_);
1552 iter->key_cache_path =
1553 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1554 iter->flags & BTREE_ITER_INTENT,
1557 ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1562 ck = (void *) iter->key_cache_path->l[0].b;
1564 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1565 trace_trans_restart_key_cache_raced(trans->fn, _RET_IP_);
1566 btree_trans_restart(trans);
1570 iter->key_cache_path->should_be_locked = true;
1573 path = iter->key_cache_path;
1576 return bch2_trans_update_by_path(trans, path, k, flags);
1579 void bch2_trans_commit_hook(struct btree_trans *trans,
1580 struct btree_trans_commit_hook *h)
1582 h->next = trans->hooks;
1586 int __bch2_btree_insert(struct btree_trans *trans,
1587 enum btree_id id, struct bkey_i *k)
1589 struct btree_iter iter;
1592 bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1594 ret = bch2_btree_iter_traverse(&iter) ?:
1595 bch2_trans_update(trans, &iter, k, 0);
1596 bch2_trans_iter_exit(trans, &iter);
1601 * bch2_btree_insert - insert keys into the extent btree
1602 * @c: pointer to struct bch_fs
1603 * @id: btree to insert into
1604 * @insert_keys: list of keys to insert
1605 * @hook: insert callback
1607 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1609 struct disk_reservation *disk_res,
1610 u64 *journal_seq, int flags)
1612 return bch2_trans_do(c, disk_res, journal_seq, flags,
1613 __bch2_btree_insert(&trans, id, k));
1616 int bch2_btree_delete_at(struct btree_trans *trans,
1617 struct btree_iter *iter, unsigned update_flags)
1621 k = bch2_trans_kmalloc(trans, sizeof(*k));
1627 return bch2_trans_update(trans, iter, k, update_flags);
1630 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1631 struct bpos start, struct bpos end,
1632 unsigned iter_flags,
1635 struct btree_iter iter;
1639 bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT|iter_flags);
1641 while ((bch2_trans_begin(trans),
1642 (k = bch2_btree_iter_peek(&iter)).k) &&
1643 !(ret = bkey_err(k)) &&
1644 bkey_cmp(iter.pos, end) < 0) {
1645 struct disk_reservation disk_res =
1646 bch2_disk_reservation_init(trans->c, 0);
1647 struct bkey_i delete;
1649 bkey_init(&delete.k);
1652 * This could probably be more efficient for extents:
1656 * For extents, iter.pos won't necessarily be the same as
1657 * bkey_start_pos(k.k) (for non extents they always will be the
1658 * same). It's important that we delete starting from iter.pos
1659 * because the range we want to delete could start in the middle
1662 * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1663 * bkey_start_pos(k.k)).
1665 delete.k.p = iter.pos;
1667 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1668 unsigned max_sectors =
1669 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1671 /* create the biggest key we can */
1672 bch2_key_resize(&delete.k, max_sectors);
1673 bch2_cut_back(end, &delete);
1675 ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1680 ret = bch2_trans_update(trans, &iter, &delete, 0) ?:
1681 bch2_trans_commit(trans, &disk_res, journal_seq,
1682 BTREE_INSERT_NOFAIL);
1683 bch2_disk_reservation_put(trans->c, &disk_res);
1688 if (ret == -EINTR) {
1693 bch2_trans_iter_exit(trans, &iter);
1698 * bch_btree_delete_range - delete everything within a given range
1700 * Range is a half open interval - [start, end)
1702 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1703 struct bpos start, struct bpos end,
1704 unsigned iter_flags,
1707 return bch2_trans_do(c, NULL, journal_seq, 0,
1708 bch2_btree_delete_range_trans(&trans, id, start, end,
1709 iter_flags, journal_seq));