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"
11 #include "btree_write_buffer.h"
16 #include "extent_update.h"
18 #include "journal_reclaim.h"
21 #include "subvolume.h"
25 #include <linux/prefetch.h>
26 #include <linux/sort.h>
29 * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
32 static struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
34 struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
36 if (k.k && bpos_eq(path->pos, k.k->p))
41 return (struct bkey_s_c) { u, NULL };
44 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
46 #ifdef CONFIG_BCACHEFS_DEBUG
47 struct bch_fs *c = trans->c;
49 struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u);
51 if (unlikely(trans->journal_replay_not_finished)) {
53 bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
56 k = bkey_i_to_s_c(j_k);
60 u.needs_whiteout = i->old_k.needs_whiteout;
62 BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
63 BUG_ON(i->old_v != k.v);
67 static int __must_check
68 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
69 struct bkey_i *, enum btree_update_flags,
72 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
73 const struct btree_insert_entry *r)
75 return cmp_int(l->btree_id, r->btree_id) ?:
76 cmp_int(l->cached, r->cached) ?:
77 -cmp_int(l->level, r->level) ?:
78 bpos_cmp(l->k->k.p, r->k->k.p);
81 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
83 return i->path->l + i->level;
86 static inline bool same_leaf_as_prev(struct btree_trans *trans,
87 struct btree_insert_entry *i)
89 return i != trans->updates &&
90 insert_l(&i[0])->b == insert_l(&i[-1])->b;
93 static inline bool same_leaf_as_next(struct btree_trans *trans,
94 struct btree_insert_entry *i)
96 return i + 1 < trans->updates + trans->nr_updates &&
97 insert_l(&i[0])->b == insert_l(&i[1])->b;
100 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
101 struct btree_path *path,
104 struct bch_fs *c = trans->c;
106 if (unlikely(btree_node_just_written(b)) &&
107 bch2_btree_post_write_cleanup(c, b))
108 bch2_trans_node_reinit_iter(trans, b);
111 * If the last bset has been written, or if it's gotten too big - start
112 * a new bset to insert into:
114 if (want_new_bset(c, b))
115 bch2_btree_init_next(trans, b);
118 /* Inserting into a given leaf node (last stage of insert): */
120 /* Handle overwrites and do insert, for non extents: */
121 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
122 struct btree_path *path,
124 struct btree_node_iter *node_iter,
125 struct bkey_i *insert)
127 struct bkey_packed *k;
128 unsigned clobber_u64s = 0, new_u64s = 0;
130 EBUG_ON(btree_node_just_written(b));
131 EBUG_ON(bset_written(b, btree_bset_last(b)));
132 EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
133 EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
134 EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
135 EBUG_ON(insert->k.u64s >
136 bch_btree_keys_u64s_remaining(trans->c, b));
138 k = bch2_btree_node_iter_peek_all(node_iter, b);
139 if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
142 /* @k is the key being overwritten/deleted, if any: */
143 EBUG_ON(k && bkey_deleted(k));
145 /* Deleting, but not found? nothing to do: */
146 if (bkey_deleted(&insert->k) && !k)
149 if (bkey_deleted(&insert->k)) {
151 btree_account_key_drop(b, k);
152 k->type = KEY_TYPE_deleted;
154 if (k->needs_whiteout)
155 push_whiteout(trans->c, b, insert->k.p);
156 k->needs_whiteout = false;
158 if (k >= btree_bset_last(b)->start) {
159 clobber_u64s = k->u64s;
160 bch2_bset_delete(b, k, clobber_u64s);
163 bch2_btree_path_fix_key_modified(trans, b, k);
171 btree_account_key_drop(b, k);
172 k->type = KEY_TYPE_deleted;
174 insert->k.needs_whiteout = k->needs_whiteout;
175 k->needs_whiteout = false;
177 if (k >= btree_bset_last(b)->start) {
178 clobber_u64s = k->u64s;
181 bch2_btree_path_fix_key_modified(trans, b, k);
185 k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
187 bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
190 if (clobber_u64s != new_u64s)
191 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
192 clobber_u64s, new_u64s);
196 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
199 struct bch_fs *c = container_of(j, struct bch_fs, journal);
200 struct btree_write *w = container_of(pin, struct btree_write, journal);
201 struct btree *b = container_of(w, struct btree, writes[i]);
202 struct btree_trans trans;
203 unsigned long old, new, v;
204 unsigned idx = w - b->writes;
206 bch2_trans_init(&trans, c, 0, 0);
208 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
209 v = READ_ONCE(b->flags);
214 if (!(old & (1 << BTREE_NODE_dirty)) ||
215 !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
216 w->journal.seq != seq)
219 new &= ~BTREE_WRITE_TYPE_MASK;
220 new |= BTREE_WRITE_journal_reclaim;
221 new |= 1 << BTREE_NODE_need_write;
222 } while ((v = cmpxchg(&b->flags, old, new)) != old);
224 btree_node_write_if_need(c, b, SIX_LOCK_read);
225 six_unlock_read(&b->c.lock);
227 bch2_trans_exit(&trans);
231 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
233 return __btree_node_flush(j, pin, 0, seq);
236 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
238 return __btree_node_flush(j, pin, 1, seq);
241 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
242 struct btree *b, u64 seq)
244 struct btree_write *w = btree_current_write(b);
246 bch2_journal_pin_add(&c->journal, seq, &w->journal,
247 btree_node_write_idx(b) == 0
248 ? bch2_btree_node_flush0
249 : bch2_btree_node_flush1);
253 * btree_insert_key - insert a key one key into a leaf node
255 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
256 struct btree_path *path,
257 struct bkey_i *insert,
260 struct bch_fs *c = trans->c;
261 struct btree *b = path_l(path)->b;
262 struct bset_tree *t = bset_tree_last(b);
263 struct bset *i = bset(b, t);
264 int old_u64s = bset_u64s(t);
265 int old_live_u64s = b->nr.live_u64s;
266 int live_u64s_added, u64s_added;
268 if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
269 &path_l(path)->iter, insert)))
272 i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
274 bch2_btree_add_journal_pin(c, b, journal_seq);
276 if (unlikely(!btree_node_dirty(b))) {
277 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
278 set_btree_node_dirty_acct(c, b);
281 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
282 u64s_added = (int) bset_u64s(t) - old_u64s;
284 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
285 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
286 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
287 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
289 if (u64s_added > live_u64s_added &&
290 bch2_maybe_compact_whiteouts(c, b))
291 bch2_trans_node_reinit_iter(trans, b);
294 /* Cached btree updates: */
296 /* Normal update interface: */
298 static inline void btree_insert_entry_checks(struct btree_trans *trans,
299 struct btree_insert_entry *i)
301 BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
302 BUG_ON(i->cached != i->path->cached);
303 BUG_ON(i->level != i->path->level);
304 BUG_ON(i->btree_id != i->path->btree_id);
306 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
307 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
308 i->k->k.p.snapshot &&
309 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
313 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned flags,
314 unsigned long trace_ip)
316 return drop_locks_do(trans,
317 bch2_journal_preres_get(&trans->c->journal,
318 &trans->journal_preres,
319 trans->journal_preres_u64s,
320 (flags & BCH_WATERMARK_MASK)));
323 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
326 return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
327 trans->journal_u64s, flags);
330 #define JSET_ENTRY_LOG_U64s 4
332 static noinline void journal_transaction_name(struct btree_trans *trans)
334 struct bch_fs *c = trans->c;
335 struct journal *j = &c->journal;
336 struct jset_entry *entry =
337 bch2_journal_add_entry(j, &trans->journal_res,
338 BCH_JSET_ENTRY_log, 0, 0,
339 JSET_ENTRY_LOG_U64s);
340 struct jset_entry_log *l =
341 container_of(entry, struct jset_entry_log, entry);
343 strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
346 static inline int btree_key_can_insert(struct btree_trans *trans,
347 struct btree *b, unsigned u64s)
349 struct bch_fs *c = trans->c;
351 if (!bch2_btree_node_insert_fits(c, b, u64s))
352 return -BCH_ERR_btree_insert_btree_node_full;
357 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
358 struct btree_path *path, unsigned u64s)
360 struct bch_fs *c = trans->c;
361 struct bkey_cached *ck = (void *) path->l[0].b;
362 struct btree_insert_entry *i;
364 struct bkey_i *new_k;
366 EBUG_ON(path->level);
368 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
369 bch2_btree_key_cache_must_wait(c) &&
370 !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
371 return -BCH_ERR_btree_insert_need_journal_reclaim;
374 * bch2_varint_decode can read past the end of the buffer by at most 7
375 * bytes (it won't be used):
379 if (u64s <= ck->u64s)
382 new_u64s = roundup_pow_of_two(u64s);
383 new_k = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
385 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
386 bch2_btree_ids[path->btree_id], new_u64s);
387 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
390 trans_for_each_update(trans, i)
391 if (i->old_v == &ck->k->v)
392 i->old_v = &new_k->v;
401 static int run_one_mem_trigger(struct btree_trans *trans,
402 struct btree_insert_entry *i,
405 struct bkey_s_c old = { &i->old_k, i->old_v };
406 struct bkey_i *new = i->k;
407 const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
408 const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
411 verify_update_old_key(trans, i);
413 if (unlikely(flags & BTREE_TRIGGER_NORUN))
416 if (!btree_node_type_needs_gc((enum btree_node_type) i->btree_id))
419 if (old_ops->atomic_trigger == new_ops->atomic_trigger &&
420 ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
421 ret = bch2_mark_key(trans, i->btree_id, i->level,
422 old, bkey_i_to_s_c(new),
423 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
425 struct bkey _deleted = KEY(0, 0, 0);
426 struct bkey_s_c deleted = (struct bkey_s_c) { &_deleted, NULL };
428 _deleted.p = i->path->pos;
430 ret = bch2_mark_key(trans, i->btree_id, i->level,
431 deleted, bkey_i_to_s_c(new),
432 BTREE_TRIGGER_INSERT|flags) ?:
433 bch2_mark_key(trans, i->btree_id, i->level,
435 BTREE_TRIGGER_OVERWRITE|flags);
441 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
445 * Transactional triggers create new btree_insert_entries, so we can't
446 * pass them a pointer to a btree_insert_entry, that memory is going to
449 struct bkey old_k = i->old_k;
450 struct bkey_s_c old = { &old_k, i->old_v };
451 const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
452 const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
454 verify_update_old_key(trans, i);
456 if ((i->flags & BTREE_TRIGGER_NORUN) ||
457 !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
460 if (!i->insert_trigger_run &&
461 !i->overwrite_trigger_run &&
462 old_ops->trans_trigger == new_ops->trans_trigger &&
463 ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
464 i->overwrite_trigger_run = true;
465 i->insert_trigger_run = true;
466 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
467 BTREE_TRIGGER_INSERT|
468 BTREE_TRIGGER_OVERWRITE|
470 } else if (overwrite && !i->overwrite_trigger_run) {
471 i->overwrite_trigger_run = true;
472 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
473 } else if (!overwrite && !i->insert_trigger_run) {
474 i->insert_trigger_run = true;
475 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
481 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
482 struct btree_insert_entry *btree_id_start)
484 struct btree_insert_entry *i;
485 bool trans_trigger_run;
488 for (overwrite = 1; overwrite >= 0; --overwrite) {
491 * Running triggers will append more updates to the list of updates as
495 trans_trigger_run = false;
497 for (i = btree_id_start;
498 i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
500 if (i->btree_id != btree_id)
503 ret = run_one_trans_trigger(trans, i, overwrite);
507 trans_trigger_run = true;
509 } while (trans_trigger_run);
515 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
517 struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
518 unsigned btree_id = 0;
523 * For a given btree, this algorithm runs insert triggers before
524 * overwrite triggers: this is so that when extents are being moved
525 * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
528 for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
529 if (btree_id == BTREE_ID_alloc)
532 while (btree_id_start < trans->updates + trans->nr_updates &&
533 btree_id_start->btree_id < btree_id)
536 ret = run_btree_triggers(trans, btree_id, btree_id_start);
541 trans_for_each_update(trans, i) {
542 if (i->btree_id > BTREE_ID_alloc)
544 if (i->btree_id == BTREE_ID_alloc) {
545 ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
552 #ifdef CONFIG_BCACHEFS_DEBUG
553 trans_for_each_update(trans, i)
554 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
555 (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
556 (!i->insert_trigger_run || !i->overwrite_trigger_run));
561 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
563 struct bch_fs *c = trans->c;
564 struct btree_insert_entry *i;
567 trans_for_each_update(trans, i) {
569 * XXX: synchronization of cached update triggers with gc
570 * XXX: synchronization of interior node updates with gc
572 BUG_ON(i->cached || i->level);
574 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
575 ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
585 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
586 struct btree_insert_entry **stopped_at,
587 unsigned long trace_ip)
589 struct bch_fs *c = trans->c;
590 struct btree_insert_entry *i;
591 struct btree_write_buffered_key *wb;
592 struct btree_trans_commit_hook *h;
594 bool marking = false;
598 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
599 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
603 * Check if the insert will fit in the leaf node with the write lock
604 * held, otherwise another thread could write the node changing the
605 * amount of space available:
608 prefetch(&trans->c->journal.flags);
610 trans_for_each_update(trans, i) {
611 /* Multiple inserts might go to same leaf: */
612 if (!same_leaf_as_prev(trans, i))
615 u64s += i->k->k.u64s;
617 ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
618 : btree_key_can_insert_cached(trans, flags, i->path, u64s);
624 if (btree_node_type_needs_gc(i->bkey_type))
628 if (trans->nr_wb_updates &&
629 trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
630 return -BCH_ERR_btree_insert_need_flush_buffer;
633 * Don't get journal reservation until after we know insert will
636 if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
637 ret = bch2_trans_journal_res_get(trans,
638 (flags & BCH_WATERMARK_MASK)|
639 JOURNAL_RES_GET_NONBLOCK);
643 if (unlikely(trans->journal_transaction_names))
644 journal_transaction_name(trans);
646 trans->journal_res.seq = c->journal.replay_journal_seq;
650 * Not allowed to fail after we've gotten our journal reservation - we
654 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
655 !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
656 if (bch2_journal_seq_verify)
657 trans_for_each_update(trans, i)
658 i->k->k.version.lo = trans->journal_res.seq;
659 else if (bch2_inject_invalid_keys)
660 trans_for_each_update(trans, i)
661 i->k->k.version = MAX_VERSION;
664 if (trans->fs_usage_deltas &&
665 bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
666 return -BCH_ERR_btree_insert_need_mark_replicas;
668 if (trans->nr_wb_updates) {
669 EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
671 ret = bch2_btree_insert_keys_write_buffer(trans);
673 goto revert_fs_usage;
678 ret = h->fn(trans, h);
680 goto revert_fs_usage;
684 trans_for_each_update(trans, i)
685 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
686 ret = run_one_mem_trigger(trans, i, i->flags);
691 if (unlikely(c->gc_pos.phase)) {
692 ret = bch2_trans_commit_run_gc_triggers(trans);
697 if (unlikely(trans->extra_journal_entries.nr)) {
698 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
699 trans->extra_journal_entries.data,
700 trans->extra_journal_entries.nr);
702 trans->journal_res.offset += trans->extra_journal_entries.nr;
703 trans->journal_res.u64s -= trans->extra_journal_entries.nr;
706 if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
707 struct journal *j = &c->journal;
708 struct jset_entry *entry;
710 trans_for_each_update(trans, i) {
711 if (i->key_cache_already_flushed)
714 if (i->flags & BTREE_UPDATE_NOJOURNAL)
717 verify_update_old_key(trans, i);
719 if (trans->journal_transaction_names) {
720 entry = bch2_journal_add_entry(j, &trans->journal_res,
721 BCH_JSET_ENTRY_overwrite,
722 i->btree_id, i->level,
724 bkey_reassemble(&entry->start[0],
725 (struct bkey_s_c) { &i->old_k, i->old_v });
728 entry = bch2_journal_add_entry(j, &trans->journal_res,
729 BCH_JSET_ENTRY_btree_keys,
730 i->btree_id, i->level,
732 bkey_copy(&entry->start[0], i->k);
735 trans_for_each_wb_update(trans, wb) {
736 entry = bch2_journal_add_entry(j, &trans->journal_res,
737 BCH_JSET_ENTRY_btree_keys,
740 bkey_copy(&entry->start[0], &wb->k);
743 if (trans->journal_seq)
744 *trans->journal_seq = trans->journal_res.seq;
747 trans_for_each_update(trans, i) {
748 i->k->k.needs_whiteout = false;
751 u64 seq = trans->journal_res.seq;
753 if (i->flags & BTREE_UPDATE_PREJOURNAL)
756 bch2_btree_insert_key_leaf(trans, i->path, i->k, seq);
757 } else if (!i->key_cache_already_flushed)
758 bch2_btree_insert_key_cached(trans, flags, i);
760 bch2_btree_key_cache_drop(trans, i->path);
761 btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
769 if (trans->fs_usage_deltas)
770 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
774 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
776 while (--i >= trans->updates) {
777 if (same_leaf_as_prev(trans, i))
780 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
783 trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
784 return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
787 static inline int trans_lock_write(struct btree_trans *trans)
789 struct btree_insert_entry *i;
791 trans_for_each_update(trans, i) {
792 if (same_leaf_as_prev(trans, i))
795 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
796 return trans_lock_write_fail(trans, i);
799 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
805 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
807 struct btree_insert_entry *i;
808 struct btree_write_buffered_key *wb;
810 trans_for_each_update(trans, i)
811 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
813 trans_for_each_wb_update(trans, wb)
814 bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
817 #ifdef CONFIG_BCACHEFS_DEBUG
818 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
819 struct btree_insert_entry *i,
820 struct printbuf *err)
822 struct bch_fs *c = trans->c;
823 int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
826 prt_printf(err, "invalid bkey on insert from %s -> %ps",
827 trans->fn, (void *) i->ip_allocated);
829 printbuf_indent_add(err, 2);
831 bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
834 bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
835 i->bkey_type, rw, err);
836 bch2_print_string_as_lines(KERN_ERR, err->buf);
838 bch2_inconsistent_error(c);
839 bch2_dump_trans_updates(trans);
847 * Get journal reservation, take write locks, and attempt to do btree update(s):
849 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
850 struct btree_insert_entry **stopped_at,
851 unsigned long trace_ip)
853 struct bch_fs *c = trans->c;
854 struct btree_insert_entry *i;
855 int ret = 0, u64s_delta = 0;
857 #ifdef CONFIG_BCACHEFS_DEBUG
858 trans_for_each_update(trans, i) {
859 struct printbuf buf = PRINTBUF;
860 enum bkey_invalid_flags invalid_flags = 0;
862 if (!(flags & BTREE_INSERT_JOURNAL_REPLAY))
863 invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
865 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
866 i->bkey_type, invalid_flags, &buf)))
867 ret = bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
868 btree_insert_entry_checks(trans, i);
876 trans_for_each_update(trans, i) {
880 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
881 u64s_delta -= i->old_btree_u64s;
883 if (!same_leaf_as_next(trans, i)) {
884 if (u64s_delta <= 0) {
885 ret = bch2_foreground_maybe_merge(trans, i->path,
895 ret = bch2_journal_preres_get(&c->journal,
896 &trans->journal_preres, trans->journal_preres_u64s,
897 (flags & BCH_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
898 if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
899 ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
903 ret = trans_lock_write(trans);
907 ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
909 if (!ret && unlikely(trans->journal_replay_not_finished))
910 bch2_drop_overwrites_from_journal(trans);
912 trans_for_each_update(trans, i)
913 if (!same_leaf_as_prev(trans, i))
914 bch2_btree_node_unlock_write_inlined(trans, i->path,
917 if (!ret && trans->journal_pin)
918 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
919 trans->journal_pin, NULL);
922 * Drop journal reservation after dropping write locks, since dropping
923 * the journal reservation may kick off a journal write:
925 bch2_journal_res_put(&c->journal, &trans->journal_res);
930 bch2_trans_downgrade(trans);
935 static int journal_reclaim_wait_done(struct bch_fs *c)
937 int ret = bch2_journal_error(&c->journal) ?:
938 !bch2_btree_key_cache_must_wait(c);
941 journal_reclaim_kick(&c->journal);
946 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
947 struct btree_insert_entry *i,
948 int ret, unsigned long trace_ip)
950 struct bch_fs *c = trans->c;
953 case -BCH_ERR_btree_insert_btree_node_full:
954 ret = bch2_btree_split_leaf(trans, i->path, flags);
955 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
956 trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
958 case -BCH_ERR_btree_insert_need_mark_replicas:
959 ret = drop_locks_do(trans,
960 bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
962 case -BCH_ERR_journal_res_get_blocked:
964 * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
967 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
968 (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
969 ret = -BCH_ERR_journal_reclaim_would_deadlock;
973 ret = drop_locks_do(trans,
974 bch2_trans_journal_res_get(trans,
975 (flags & BCH_WATERMARK_MASK)|
976 JOURNAL_RES_GET_CHECK));
978 case -BCH_ERR_btree_insert_need_journal_reclaim:
979 bch2_trans_unlock(trans);
981 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
983 wait_event_freezable(c->journal.reclaim_wait,
984 (ret = journal_reclaim_wait_done(c)));
988 ret = bch2_trans_relock(trans);
990 case -BCH_ERR_btree_insert_need_flush_buffer: {
991 struct btree_write_buffer *wb = &c->btree_write_buffer;
995 if (wb->state.nr > wb->size * 3 / 4) {
996 bch2_trans_unlock(trans);
997 mutex_lock(&wb->flush_lock);
999 if (wb->state.nr > wb->size * 3 / 4) {
1000 bch2_trans_begin(trans);
1001 ret = __bch2_btree_write_buffer_flush(trans,
1002 flags|BTREE_INSERT_NOCHECK_RW, true);
1004 trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1005 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1008 mutex_unlock(&wb->flush_lock);
1009 ret = bch2_trans_relock(trans);
1019 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
1021 bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
1022 !(flags & BTREE_INSERT_NOWAIT) &&
1023 (flags & BTREE_INSERT_NOFAIL), c,
1024 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
1030 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
1032 struct bch_fs *c = trans->c;
1035 if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
1036 test_bit(BCH_FS_STARTED, &c->flags))
1037 return -BCH_ERR_erofs_trans_commit;
1039 ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
1043 bch2_write_ref_get(c, BCH_WRITE_REF_trans);
1048 * This is for updates done in the early part of fsck - btree_gc - before we've
1049 * gone RW. we only add the new key to the list of keys for journal replay to
1053 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1055 struct bch_fs *c = trans->c;
1056 struct btree_insert_entry *i;
1059 trans_for_each_update(trans, i) {
1060 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1068 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1070 struct bch_fs *c = trans->c;
1071 struct btree_insert_entry *i = NULL;
1072 struct btree_write_buffered_key *wb;
1076 if (!trans->nr_updates &&
1077 !trans->nr_wb_updates &&
1078 !trans->extra_journal_entries.nr)
1081 if (flags & BTREE_INSERT_GC_LOCK_HELD)
1082 lockdep_assert_held(&c->gc_lock);
1084 ret = bch2_trans_commit_run_triggers(trans);
1088 if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1089 ret = do_bch2_trans_commit_to_journal_replay(trans);
1093 if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1094 unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1095 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1100 if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1101 mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1102 bch2_trans_begin(trans);
1103 bch2_trans_unlock(trans);
1105 ret = __bch2_btree_write_buffer_flush(trans,
1106 flags|BTREE_INSERT_NOCHECK_RW, true);
1108 trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1109 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1114 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1116 memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1118 trans->journal_u64s = trans->extra_journal_entries.nr;
1119 trans->journal_preres_u64s = 0;
1121 trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1123 if (trans->journal_transaction_names)
1124 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1126 trans_for_each_update(trans, i) {
1127 EBUG_ON(!i->path->should_be_locked);
1129 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1133 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1135 if (i->key_cache_already_flushed)
1138 /* we're going to journal the key being updated: */
1139 u64s = jset_u64s(i->k->k.u64s);
1141 likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1142 trans->journal_preres_u64s += u64s;
1144 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1147 trans->journal_u64s += u64s;
1149 /* and we're also going to log the overwrite: */
1150 if (trans->journal_transaction_names)
1151 trans->journal_u64s += jset_u64s(i->old_k.u64s);
1154 trans_for_each_wb_update(trans, wb)
1155 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1157 if (trans->extra_journal_res) {
1158 ret = bch2_disk_reservation_add(c, trans->disk_res,
1159 trans->extra_journal_res,
1160 (flags & BTREE_INSERT_NOFAIL)
1161 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1166 bch2_trans_verify_not_in_restart(trans);
1167 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1169 ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1171 /* make sure we didn't drop or screw up locks: */
1172 bch2_trans_verify_locks(trans);
1177 trace_and_count(c, transaction_commit, trans, _RET_IP_);
1179 bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1181 if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1182 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1184 bch2_trans_reset_updates(trans);
1188 ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1195 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1199 struct bch_fs *c = trans->c;
1200 struct btree_iter iter;
1204 bch2_trans_iter_init(trans, &iter, id, pos,
1205 BTREE_ITER_NOT_EXTENTS|
1206 BTREE_ITER_ALL_SNAPSHOTS);
1208 k = bch2_btree_iter_prev(&iter);
1216 if (!bkey_eq(pos, k.k->p))
1219 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1224 bch2_trans_iter_exit(trans, &iter);
1229 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1233 if (!btree_type_has_snapshots(id) ||
1234 bch2_snapshot_is_leaf(trans->c, pos.snapshot))
1237 return __check_pos_snapshot_overwritten(trans, id, pos);
1240 static noinline int extent_front_merge(struct btree_trans *trans,
1241 struct btree_iter *iter,
1243 struct bkey_i **insert,
1244 enum btree_update_flags flags)
1246 struct bch_fs *c = trans->c;
1247 struct bkey_i *update;
1250 update = bch2_bkey_make_mut_noupdate(trans, k);
1251 ret = PTR_ERR_OR_ZERO(update);
1255 if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
1258 ret = check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
1259 check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
1265 ret = bch2_btree_delete_at(trans, iter, flags);
1273 static noinline int extent_back_merge(struct btree_trans *trans,
1274 struct btree_iter *iter,
1275 struct bkey_i *insert,
1278 struct bch_fs *c = trans->c;
1281 ret = check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
1282 check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
1288 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1293 * When deleting, check if we need to emit a whiteout (because we're overwriting
1294 * something in an ancestor snapshot)
1296 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1297 enum btree_id btree_id, struct bpos pos)
1299 struct btree_iter iter;
1301 u32 snapshot = pos.snapshot;
1304 if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1309 for_each_btree_key_norestart(trans, iter, btree_id, pos,
1310 BTREE_ITER_ALL_SNAPSHOTS|
1311 BTREE_ITER_NOPRESERVE, k, ret) {
1312 if (!bkey_eq(k.k->p, pos))
1315 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1317 ret = !bkey_whiteout(k.k);
1321 bch2_trans_iter_exit(trans, &iter);
1326 int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
1328 struct bpos old_pos,
1329 struct bpos new_pos)
1331 struct bch_fs *c = trans->c;
1332 struct btree_iter old_iter, new_iter = { NULL };
1333 struct bkey_s_c old_k, new_k;
1335 struct bkey_i *update;
1338 if (!bch2_snapshot_has_children(c, old_pos.snapshot))
1343 bch2_trans_iter_init(trans, &old_iter, id, old_pos,
1344 BTREE_ITER_NOT_EXTENTS|
1345 BTREE_ITER_ALL_SNAPSHOTS);
1346 while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
1347 !(ret = bkey_err(old_k)) &&
1348 bkey_eq(old_pos, old_k.k->p)) {
1349 struct bpos whiteout_pos =
1350 SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
1352 if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
1353 snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
1356 new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
1357 BTREE_ITER_NOT_EXTENTS|
1359 ret = bkey_err(new_k);
1363 if (new_k.k->type == KEY_TYPE_deleted) {
1364 update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
1365 ret = PTR_ERR_OR_ZERO(update);
1369 bkey_init(&update->k);
1370 update->k.p = whiteout_pos;
1371 update->k.type = KEY_TYPE_whiteout;
1373 ret = bch2_trans_update(trans, &new_iter, update,
1374 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1376 bch2_trans_iter_exit(trans, &new_iter);
1378 ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
1382 bch2_trans_iter_exit(trans, &new_iter);
1383 bch2_trans_iter_exit(trans, &old_iter);
1389 int bch2_trans_update_extent_overwrite(struct btree_trans *trans,
1390 struct btree_iter *iter,
1391 enum btree_update_flags flags,
1392 struct bkey_s_c old,
1393 struct bkey_s_c new)
1395 enum btree_id btree_id = iter->btree_id;
1396 struct bkey_i *update;
1397 struct bpos new_start = bkey_start_pos(new.k);
1398 bool front_split = bkey_lt(bkey_start_pos(old.k), new_start);
1399 bool back_split = bkey_gt(old.k->p, new.k->p);
1400 int ret = 0, compressed_sectors;
1403 * If we're going to be splitting a compressed extent, note it
1404 * so that __bch2_trans_commit() can increase our disk
1407 if (((front_split && back_split) ||
1408 ((front_split || back_split) && old.k->p.snapshot != new.k->p.snapshot)) &&
1409 (compressed_sectors = bch2_bkey_sectors_compressed(old)))
1410 trans->extra_journal_res += compressed_sectors;
1413 update = bch2_bkey_make_mut_noupdate(trans, old);
1414 if ((ret = PTR_ERR_OR_ZERO(update)))
1417 bch2_cut_back(new_start, update);
1419 ret = bch2_insert_snapshot_whiteouts(trans, btree_id,
1420 old.k->p, update->k.p) ?:
1421 bch2_btree_insert_nonextent(trans, btree_id, update,
1422 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1427 /* If we're overwriting in a different snapshot - middle split: */
1428 if (old.k->p.snapshot != new.k->p.snapshot &&
1429 (front_split || back_split)) {
1430 update = bch2_bkey_make_mut_noupdate(trans, old);
1431 if ((ret = PTR_ERR_OR_ZERO(update)))
1434 bch2_cut_front(new_start, update);
1435 bch2_cut_back(new.k->p, update);
1437 ret = bch2_insert_snapshot_whiteouts(trans, btree_id,
1438 old.k->p, update->k.p) ?:
1439 bch2_btree_insert_nonextent(trans, btree_id, update,
1440 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1445 if (bkey_le(old.k->p, new.k->p)) {
1446 update = bch2_trans_kmalloc(trans, sizeof(*update));
1447 if ((ret = PTR_ERR_OR_ZERO(update)))
1450 bkey_init(&update->k);
1451 update->k.p = old.k->p;
1452 update->k.p.snapshot = new.k->p.snapshot;
1454 if (new.k->p.snapshot != old.k->p.snapshot) {
1455 update->k.type = KEY_TYPE_whiteout;
1456 } else if (btree_type_has_snapshots(btree_id)) {
1457 ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
1461 update->k.type = KEY_TYPE_whiteout;
1464 ret = bch2_btree_insert_nonextent(trans, btree_id, update,
1465 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1471 update = bch2_bkey_make_mut_noupdate(trans, old);
1472 if ((ret = PTR_ERR_OR_ZERO(update)))
1475 bch2_cut_front(new.k->p, update);
1477 ret = bch2_trans_update_by_path(trans, iter->path, update,
1478 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1487 static int bch2_trans_update_extent(struct btree_trans *trans,
1488 struct btree_iter *orig_iter,
1489 struct bkey_i *insert,
1490 enum btree_update_flags flags)
1492 struct btree_iter iter;
1494 enum btree_id btree_id = orig_iter->btree_id;
1497 bch2_trans_iter_init(trans, &iter, btree_id, bkey_start_pos(&insert->k),
1499 BTREE_ITER_WITH_UPDATES|
1500 BTREE_ITER_NOT_EXTENTS);
1501 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1502 if ((ret = bkey_err(k)))
1507 if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
1508 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1509 ret = extent_front_merge(trans, &iter, k, &insert, flags);
1517 while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1518 bool done = bkey_lt(insert->k.p, k.k->p);
1520 ret = bch2_trans_update_extent_overwrite(trans, &iter, flags, k, bkey_i_to_s_c(insert));
1527 bch2_btree_iter_advance(&iter);
1528 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1529 if ((ret = bkey_err(k)))
1535 if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1536 ret = extent_back_merge(trans, &iter, insert, k);
1541 if (!bkey_deleted(&insert->k))
1542 ret = bch2_btree_insert_nonextent(trans, btree_id, insert, flags);
1544 bch2_trans_iter_exit(trans, &iter);
1549 static noinline int flush_new_cached_update(struct btree_trans *trans,
1550 struct btree_path *path,
1551 struct btree_insert_entry *i,
1552 enum btree_update_flags flags,
1555 struct btree_path *btree_path;
1559 btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1560 BTREE_ITER_INTENT, _THIS_IP_);
1561 ret = bch2_btree_path_traverse(trans, btree_path, 0);
1566 * The old key in the insert entry might actually refer to an existing
1567 * key in the btree that has been deleted from cache and not yet
1568 * flushed. Check for this and skip the flush so we don't run triggers
1569 * against a stale key.
1571 bch2_btree_path_peek_slot_exact(btree_path, &k);
1572 if (!bkey_deleted(&k))
1575 i->key_cache_already_flushed = true;
1576 i->flags |= BTREE_TRIGGER_NORUN;
1578 btree_path_set_should_be_locked(btree_path);
1579 ret = bch2_trans_update_by_path(trans, btree_path, i->k, flags, ip);
1581 bch2_path_put(trans, btree_path, true);
1585 static int __must_check
1586 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1587 struct bkey_i *k, enum btree_update_flags flags,
1590 struct bch_fs *c = trans->c;
1591 struct btree_insert_entry *i, n;
1595 EBUG_ON(!path->should_be_locked);
1596 EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1597 EBUG_ON(!bpos_eq(k->k.p, path->pos));
1600 * The transaction journal res hasn't been allocated at this point.
1601 * That occurs at commit time. Reuse the seq field to pass in the seq
1602 * of a prejournaled key.
1604 if (flags & BTREE_UPDATE_PREJOURNAL)
1605 seq = trans->journal_res.seq;
1607 n = (struct btree_insert_entry) {
1609 .bkey_type = __btree_node_type(path->level, path->btree_id),
1610 .btree_id = path->btree_id,
1611 .level = path->level,
1612 .cached = path->cached,
1619 #ifdef CONFIG_BCACHEFS_DEBUG
1620 trans_for_each_update(trans, i)
1621 BUG_ON(i != trans->updates &&
1622 btree_insert_entry_cmp(i - 1, i) >= 0);
1626 * Pending updates are kept sorted: first, find position of new update,
1627 * then delete/trim any updates the new update overwrites:
1629 trans_for_each_update(trans, i) {
1630 cmp = btree_insert_entry_cmp(&n, i);
1635 if (!cmp && i < trans->updates + trans->nr_updates) {
1636 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1638 bch2_path_put(trans, i->path, true);
1640 i->cached = n.cached;
1644 i->ip_allocated = n.ip_allocated;
1646 array_insert_item(trans->updates, trans->nr_updates,
1647 i - trans->updates, n);
1649 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
1650 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1652 if (unlikely(trans->journal_replay_not_finished)) {
1653 struct bkey_i *j_k =
1654 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1663 __btree_path_get(i->path, true);
1666 * If a key is present in the key cache, it must also exist in the
1667 * btree - this is necessary for cache coherency. When iterating over
1668 * a btree that's cached in the key cache, the btree iter code checks
1669 * the key cache - but the key has to exist in the btree for that to
1672 if (path->cached && bkey_deleted(&i->old_k))
1673 return flush_new_cached_update(trans, path, i, flags, ip);
1678 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1679 struct bkey_i *k, enum btree_update_flags flags)
1681 struct btree_path *path = iter->update_path ?: iter->path;
1682 struct bkey_cached *ck;
1685 if (iter->flags & BTREE_ITER_IS_EXTENTS)
1686 return bch2_trans_update_extent(trans, iter, k, flags);
1688 if (bkey_deleted(&k->k) &&
1689 !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1690 (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1691 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1692 if (unlikely(ret < 0))
1696 k->k.type = KEY_TYPE_whiteout;
1700 * Ensure that updates to cached btrees go to the key cache:
1702 if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1705 btree_id_cached(trans->c, path->btree_id)) {
1706 if (!iter->key_cache_path ||
1707 !iter->key_cache_path->should_be_locked ||
1708 !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1709 if (!iter->key_cache_path)
1710 iter->key_cache_path =
1711 bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1713 BTREE_ITER_CACHED, _THIS_IP_);
1715 iter->key_cache_path =
1716 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1717 iter->flags & BTREE_ITER_INTENT,
1720 ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1725 ck = (void *) iter->key_cache_path->l[0].b;
1727 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1728 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1729 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1732 btree_path_set_should_be_locked(iter->key_cache_path);
1735 path = iter->key_cache_path;
1738 return bch2_trans_update_by_path(trans, path, k, flags, _RET_IP_);
1742 * Add a transaction update for a key that has already been journaled.
1744 int __must_check bch2_trans_update_seq(struct btree_trans *trans, u64 seq,
1745 struct btree_iter *iter, struct bkey_i *k,
1746 enum btree_update_flags flags)
1748 trans->journal_res.seq = seq;
1749 return bch2_trans_update(trans, iter, k, flags|BTREE_UPDATE_NOJOURNAL|
1750 BTREE_UPDATE_PREJOURNAL);
1753 int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
1754 enum btree_id btree,
1757 struct btree_write_buffered_key *i;
1760 EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
1761 EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
1763 trans_for_each_wb_update(trans, i) {
1764 if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
1765 bkey_copy(&i->k, k);
1770 if (!trans->wb_updates ||
1771 trans->nr_wb_updates == trans->wb_updates_size) {
1772 struct btree_write_buffered_key *u;
1774 if (trans->nr_wb_updates == trans->wb_updates_size) {
1775 struct btree_transaction_stats *s = btree_trans_stats(trans);
1777 BUG_ON(trans->wb_updates_size > U8_MAX / 2);
1778 trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
1780 s->wb_updates_size = trans->wb_updates_size;
1783 u = bch2_trans_kmalloc_nomemzero(trans,
1784 trans->wb_updates_size *
1785 sizeof(struct btree_write_buffered_key));
1786 ret = PTR_ERR_OR_ZERO(u);
1790 if (trans->nr_wb_updates)
1791 memcpy(u, trans->wb_updates, trans->nr_wb_updates *
1792 sizeof(struct btree_write_buffered_key));
1793 trans->wb_updates = u;
1796 trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
1800 bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
1801 trans->nr_wb_updates++;
1806 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
1807 enum btree_id btree, struct bpos end)
1812 bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
1813 k = bch2_btree_iter_prev(iter);
1818 bch2_btree_iter_advance(iter);
1819 k = bch2_btree_iter_peek_slot(iter);
1824 BUG_ON(k.k->type != KEY_TYPE_deleted);
1826 if (bkey_gt(k.k->p, end)) {
1827 ret = -BCH_ERR_ENOSPC_btree_slot;
1833 bch2_trans_iter_exit(trans, iter);
1837 void bch2_trans_commit_hook(struct btree_trans *trans,
1838 struct btree_trans_commit_hook *h)
1840 h->next = trans->hooks;
1844 int bch2_btree_insert_nonextent(struct btree_trans *trans,
1845 enum btree_id btree, struct bkey_i *k,
1846 enum btree_update_flags flags)
1848 struct btree_iter iter;
1851 bch2_trans_iter_init(trans, &iter, btree, k->k.p,
1852 BTREE_ITER_NOT_EXTENTS|
1854 ret = bch2_btree_iter_traverse(&iter) ?:
1855 bch2_trans_update(trans, &iter, k, flags);
1856 bch2_trans_iter_exit(trans, &iter);
1860 int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
1861 struct bkey_i *k, enum btree_update_flags flags)
1863 struct btree_iter iter;
1866 bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1869 ret = bch2_btree_iter_traverse(&iter) ?:
1870 bch2_trans_update(trans, &iter, k, flags);
1871 bch2_trans_iter_exit(trans, &iter);
1876 * bch2_btree_insert - insert keys into the extent btree
1877 * @c: pointer to struct bch_fs
1878 * @id: btree to insert into
1879 * @insert_keys: list of keys to insert
1880 * @hook: insert callback
1882 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1884 struct disk_reservation *disk_res,
1885 u64 *journal_seq, int flags)
1887 return bch2_trans_do(c, disk_res, journal_seq, flags,
1888 __bch2_btree_insert(&trans, id, k, 0));
1891 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1892 unsigned len, unsigned update_flags)
1896 k = bch2_trans_kmalloc(trans, sizeof(*k));
1902 bch2_key_resize(&k->k, len);
1903 return bch2_trans_update(trans, iter, k, update_flags);
1906 int bch2_btree_delete_at(struct btree_trans *trans,
1907 struct btree_iter *iter, unsigned update_flags)
1909 return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1912 int bch2_btree_delete_at_buffered(struct btree_trans *trans,
1913 enum btree_id btree, struct bpos pos)
1917 k = bch2_trans_kmalloc(trans, sizeof(*k));
1923 return bch2_trans_update_buffered(trans, btree, k);
1926 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1927 struct bpos start, struct bpos end,
1928 unsigned update_flags,
1931 u32 restart_count = trans->restart_count;
1932 struct btree_iter iter;
1936 bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1937 while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1938 struct disk_reservation disk_res =
1939 bch2_disk_reservation_init(trans->c, 0);
1940 struct bkey_i delete;
1946 bkey_init(&delete.k);
1949 * This could probably be more efficient for extents:
1953 * For extents, iter.pos won't necessarily be the same as
1954 * bkey_start_pos(k.k) (for non extents they always will be the
1955 * same). It's important that we delete starting from iter.pos
1956 * because the range we want to delete could start in the middle
1959 * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1960 * bkey_start_pos(k.k)).
1962 delete.k.p = iter.pos;
1964 if (iter.flags & BTREE_ITER_IS_EXTENTS)
1965 bch2_key_resize(&delete.k,
1966 bpos_min(end, k.k->p).offset -
1969 ret = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1970 bch2_trans_commit(trans, &disk_res, journal_seq,
1971 BTREE_INSERT_NOFAIL);
1972 bch2_disk_reservation_put(trans->c, &disk_res);
1975 * the bch2_trans_begin() call is in a weird place because we
1976 * need to call it after every transaction commit, to avoid path
1977 * overflow, but don't want to call it if the delete operation
1978 * is a no-op and we have no work to do:
1980 bch2_trans_begin(trans);
1982 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1987 bch2_trans_iter_exit(trans, &iter);
1989 if (!ret && trans_was_restarted(trans, restart_count))
1990 ret = -BCH_ERR_transaction_restart_nested;
1995 * bch_btree_delete_range - delete everything within a given range
1997 * Range is a half open interval - [start, end)
1999 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
2000 struct bpos start, struct bpos end,
2001 unsigned update_flags,
2004 int ret = bch2_trans_run(c,
2005 bch2_btree_delete_range_trans(&trans, id, start, end,
2006 update_flags, journal_seq));
2007 if (ret == -BCH_ERR_transaction_restart_nested)
2012 int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
2013 struct bpos pos, bool set)
2018 k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
2019 ret = PTR_ERR_OR_ZERO(k);
2024 k->k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
2027 return bch2_trans_update_buffered(trans, btree, k);
2030 static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
2032 struct printbuf buf = PRINTBUF;
2033 struct jset_entry_log *l;
2037 prt_vprintf(&buf, fmt, args);
2038 ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
2042 u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
2044 ret = darray_make_room(entries, jset_u64s(u64s));
2048 l = (void *) &darray_top(*entries);
2049 l->entry.u64s = cpu_to_le16(u64s);
2050 l->entry.btree_id = 0;
2052 l->entry.type = BCH_JSET_ENTRY_log;
2053 l->entry.pad[0] = 0;
2054 l->entry.pad[1] = 0;
2055 l->entry.pad[2] = 0;
2056 memcpy(l->d, buf.buf, buf.pos);
2058 l->d[buf.pos++] = '\0';
2060 entries->nr += jset_u64s(u64s);
2062 printbuf_exit(&buf);
2067 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
2072 if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
2073 ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
2075 ret = bch2_trans_do(c, NULL, NULL,
2076 BTREE_INSERT_LAZY_RW|commit_flags,
2077 __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
2083 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
2088 va_start(args, fmt);
2089 ret = __bch2_fs_log_msg(c, 0, fmt, args);
2095 * Use for logging messages during recovery to enable reserved space and avoid
2098 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
2103 va_start(args, fmt);
2104 ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);