1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
4 * Copyright (C) 2014 Datera Inc.
8 #include "alloc_background.h"
9 #include "alloc_foreground.h"
10 #include "bkey_methods.h"
12 #include "btree_locking.h"
13 #include "btree_update_interior.h"
30 #include <linux/slab.h>
31 #include <linux/bitops.h>
32 #include <linux/freezer.h>
33 #include <linux/kthread.h>
34 #include <linux/preempt.h>
35 #include <linux/rcupdate.h>
36 #include <linux/sched/task.h>
37 #include <trace/events/bcachefs.h>
39 #define DROP_THIS_NODE 10
40 #define DROP_PREV_NODE 11
42 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
45 write_seqcount_begin(&c->gc_pos_lock);
47 write_seqcount_end(&c->gc_pos_lock);
51 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
53 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
54 __gc_pos_set(c, new_pos);
58 * Missing: if an interior btree node is empty, we need to do something -
59 * perhaps just kill it
61 static int bch2_gc_check_topology(struct bch_fs *c,
63 struct bkey_buf *prev,
67 struct bpos node_start = b->data->min_key;
68 struct bpos node_end = b->data->max_key;
69 struct bpos expected_start = bkey_deleted(&prev->k->k)
71 : bpos_successor(prev->k->k.p);
72 char buf1[200], buf2[200];
75 if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
76 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
78 if (bkey_deleted(&prev->k->k)) {
79 struct printbuf out = PBUF(buf1);
80 pr_buf(&out, "start of node: ");
81 bch2_bpos_to_text(&out, node_start);
83 bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(prev->k));
86 if (bpos_cmp(expected_start, bp->v.min_key)) {
87 bch2_topology_error(c);
93 "btree node with incorrect min_key at btree %s level %u:\n"
96 bch2_btree_ids[b->c.btree_id], b->c.level,
98 (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(cur.k)), buf2)) &&
99 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
100 bch_info(c, "Halting mark and sweep to start topology repair pass");
101 return FSCK_ERR_START_TOPOLOGY_REPAIR;
103 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
108 if (is_last && bpos_cmp(cur.k->k.p, node_end)) {
109 bch2_topology_error(c);
115 "btree node with incorrect max_key at btree %s level %u:\n"
118 bch2_btree_ids[b->c.btree_id], b->c.level,
119 (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(cur.k)), buf1),
120 (bch2_bpos_to_text(&PBUF(buf2), node_end), buf2)) &&
121 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
122 bch_info(c, "Halting mark and sweep to start topology repair pass");
123 return FSCK_ERR_START_TOPOLOGY_REPAIR;
125 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
129 bch2_bkey_buf_copy(prev, c, cur.k);
134 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
136 switch (b->key.k.type) {
137 case KEY_TYPE_btree_ptr: {
138 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
142 dst->v.seq = b->data->keys.seq;
143 dst->v.sectors_written = 0;
145 dst->v.min_key = b->data->min_key;
146 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
147 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
150 case KEY_TYPE_btree_ptr_v2:
151 bkey_copy(&dst->k_i, &b->key);
158 static void bch2_btree_node_update_key_early(struct bch_fs *c,
159 enum btree_id btree, unsigned level,
160 struct bkey_s_c old, struct bkey_i *new)
166 bch2_bkey_buf_init(&tmp);
167 bch2_bkey_buf_reassemble(&tmp, c, old);
169 b = bch2_btree_node_get_noiter(c, tmp.k, btree, level, true);
170 if (!IS_ERR_OR_NULL(b)) {
171 mutex_lock(&c->btree_cache.lock);
173 bch2_btree_node_hash_remove(&c->btree_cache, b);
175 bkey_copy(&b->key, new);
176 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
179 mutex_unlock(&c->btree_cache.lock);
180 six_unlock_read(&b->c.lock);
183 bch2_bkey_buf_exit(&tmp, c);
186 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
188 struct bkey_i_btree_ptr_v2 *new;
191 new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
195 btree_ptr_to_v2(b, new);
196 b->data->min_key = new_min;
197 new->v.min_key = new_min;
198 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
200 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
206 bch2_btree_node_drop_keys_outside_node(b);
211 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
213 struct bkey_i_btree_ptr_v2 *new;
216 ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
220 new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
224 btree_ptr_to_v2(b, new);
225 b->data->max_key = new_max;
227 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
229 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
235 bch2_btree_node_drop_keys_outside_node(b);
237 mutex_lock(&c->btree_cache.lock);
238 bch2_btree_node_hash_remove(&c->btree_cache, b);
240 bkey_copy(&b->key, &new->k_i);
241 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
243 mutex_unlock(&c->btree_cache.lock);
247 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
248 struct btree *prev, struct btree *cur)
250 struct bpos expected_start = !prev
252 : bpos_successor(prev->key.k.p);
253 char buf1[200], buf2[200];
257 struct printbuf out = PBUF(buf1);
258 pr_buf(&out, "start of node: ");
259 bch2_bpos_to_text(&out, b->data->min_key);
261 bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&prev->key));
264 bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&cur->key));
267 bpos_cmp(expected_start, cur->data->min_key) > 0 &&
268 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
269 /* cur overwrites prev: */
271 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
272 cur->data->min_key) >= 0, c,
273 "btree node overwritten by next node at btree %s level %u:\n"
276 bch2_btree_ids[b->c.btree_id], b->c.level,
278 return DROP_PREV_NODE;
280 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
281 bpos_predecessor(cur->data->min_key)), c,
282 "btree node with incorrect max_key at btree %s level %u:\n"
285 bch2_btree_ids[b->c.btree_id], b->c.level,
287 ret = set_node_max(c, prev,
288 bpos_predecessor(cur->data->min_key));
290 /* prev overwrites cur: */
292 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
293 cur->data->max_key) >= 0, c,
294 "btree node overwritten by prev node at btree %s level %u:\n"
297 bch2_btree_ids[b->c.btree_id], b->c.level,
299 return DROP_THIS_NODE;
301 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
302 "btree node with incorrect min_key at btree %s level %u:\n"
305 bch2_btree_ids[b->c.btree_id], b->c.level,
307 ret = set_node_min(c, cur, expected_start);
313 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
316 char buf1[200], buf2[200];
319 if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
320 "btree node with incorrect max_key at btree %s level %u:\n"
323 bch2_btree_ids[b->c.btree_id], b->c.level,
324 (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&child->key)), buf1),
325 (bch2_bpos_to_text(&PBUF(buf2), b->key.k.p), buf2))) {
326 ret = set_node_max(c, child, b->key.k.p);
334 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
336 struct btree_and_journal_iter iter;
338 struct bkey_buf prev_k, cur_k;
339 struct btree *prev = NULL, *cur = NULL;
340 bool have_child, dropped_children = false;
348 have_child = dropped_children = false;
349 bch2_bkey_buf_init(&prev_k);
350 bch2_bkey_buf_init(&cur_k);
351 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
353 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
354 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
355 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
357 bch2_btree_and_journal_iter_advance(&iter);
358 bch2_bkey_buf_reassemble(&cur_k, c, k);
360 cur = bch2_btree_node_get_noiter(c, cur_k.k,
361 b->c.btree_id, b->c.level - 1,
363 ret = PTR_ERR_OR_ZERO(cur);
365 if (mustfix_fsck_err_on(ret == -EIO, c,
366 "Unreadable btree node at btree %s level %u:\n"
368 bch2_btree_ids[b->c.btree_id],
370 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur_k.k)), buf))) {
371 bch2_btree_node_evict(c, cur_k.k);
372 ret = bch2_journal_key_delete(c, b->c.btree_id,
373 b->c.level, cur_k.k->k.p);
380 bch_err(c, "%s: error %i getting btree node",
385 ret = btree_repair_node_boundaries(c, b, prev, cur);
387 if (ret == DROP_THIS_NODE) {
388 six_unlock_read(&cur->c.lock);
389 bch2_btree_node_evict(c, cur_k.k);
390 ret = bch2_journal_key_delete(c, b->c.btree_id,
391 b->c.level, cur_k.k->k.p);
398 six_unlock_read(&prev->c.lock);
401 if (ret == DROP_PREV_NODE) {
402 bch2_btree_node_evict(c, prev_k.k);
403 ret = bch2_journal_key_delete(c, b->c.btree_id,
404 b->c.level, prev_k.k->k.p);
408 bch2_btree_and_journal_iter_exit(&iter);
409 bch2_bkey_buf_exit(&prev_k, c);
410 bch2_bkey_buf_exit(&cur_k, c);
417 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
420 if (!ret && !IS_ERR_OR_NULL(prev)) {
422 ret = btree_repair_node_end(c, b, prev);
425 if (!IS_ERR_OR_NULL(prev))
426 six_unlock_read(&prev->c.lock);
428 if (!IS_ERR_OR_NULL(cur))
429 six_unlock_read(&cur->c.lock);
435 bch2_btree_and_journal_iter_exit(&iter);
436 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
438 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
439 bch2_bkey_buf_reassemble(&cur_k, c, k);
440 bch2_btree_and_journal_iter_advance(&iter);
442 cur = bch2_btree_node_get_noiter(c, cur_k.k,
443 b->c.btree_id, b->c.level - 1,
445 ret = PTR_ERR_OR_ZERO(cur);
448 bch_err(c, "%s: error %i getting btree node",
453 ret = bch2_btree_repair_topology_recurse(c, cur);
454 six_unlock_read(&cur->c.lock);
457 if (ret == DROP_THIS_NODE) {
458 bch2_btree_node_evict(c, cur_k.k);
459 ret = bch2_journal_key_delete(c, b->c.btree_id,
460 b->c.level, cur_k.k->k.p);
461 dropped_children = true;
470 if (mustfix_fsck_err_on(!have_child, c,
471 "empty interior btree node at btree %s level %u\n"
473 bch2_btree_ids[b->c.btree_id],
475 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(&b->key)), buf)))
476 ret = DROP_THIS_NODE;
479 if (!IS_ERR_OR_NULL(prev))
480 six_unlock_read(&prev->c.lock);
481 if (!IS_ERR_OR_NULL(cur))
482 six_unlock_read(&cur->c.lock);
484 bch2_btree_and_journal_iter_exit(&iter);
485 bch2_bkey_buf_exit(&prev_k, c);
486 bch2_bkey_buf_exit(&cur_k, c);
488 if (!ret && dropped_children)
494 static int bch2_repair_topology(struct bch_fs *c)
500 for (i = 0; i < BTREE_ID_NR && !ret; i++) {
501 b = c->btree_roots[i].b;
502 if (btree_node_fake(b))
505 six_lock_read(&b->c.lock, NULL, NULL);
506 ret = bch2_btree_repair_topology_recurse(c, b);
507 six_unlock_read(&b->c.lock);
509 if (ret == DROP_THIS_NODE) {
510 bch_err(c, "empty btree root - repair unimplemented");
518 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
519 unsigned level, bool is_root,
522 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
523 const union bch_extent_entry *entry;
524 struct extent_ptr_decoded p = { 0 };
525 bool do_update = false;
531 * use check_bucket_ref here
533 bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
534 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
535 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
536 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr);
537 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
539 if (fsck_err_on(!g->gen_valid, c,
540 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
542 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
543 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
545 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
547 g2->_mark.gen = g->_mark.gen = p.ptr.gen;
548 g2->gen_valid = g->gen_valid = true;
549 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
555 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
556 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
558 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
559 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
560 p.ptr.gen, g->mark.gen,
561 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
563 g2->_mark.gen = g->_mark.gen = p.ptr.gen;
564 g2->gen_valid = g->gen_valid = true;
565 g2->_mark.data_type = 0;
566 g2->_mark.dirty_sectors = 0;
567 g2->_mark.cached_sectors = 0;
568 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
569 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
575 if (fsck_err_on(gen_cmp(g->mark.gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
576 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
578 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->mark.gen,
579 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
581 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
584 if (fsck_err_on(!p.ptr.cached &&
585 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
586 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
588 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
589 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
590 p.ptr.gen, g->mark.gen,
591 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
594 if (data_type != BCH_DATA_btree && p.ptr.gen != g->mark.gen)
597 if (fsck_err_on(g->mark.data_type &&
598 g->mark.data_type != data_type, c,
599 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
601 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
602 bch2_data_types[g->mark.data_type],
603 bch2_data_types[data_type],
604 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
605 if (data_type == BCH_DATA_btree) {
606 g2->_mark.data_type = g->_mark.data_type = data_type;
607 g2->gen_valid = g->gen_valid = true;
608 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
615 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
617 if (fsck_err_on(!m || !m->alive, c,
618 "pointer to nonexistent stripe %llu\n"
621 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
624 if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
625 "pointer does not match stripe %llu\n"
628 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
634 struct bkey_ptrs ptrs;
635 union bch_extent_entry *entry;
636 struct bch_extent_ptr *ptr;
640 bch_err(c, "cannot update btree roots yet");
644 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
646 bch_err(c, "%s: error allocating new key", __func__);
650 bkey_reassemble(new, *k);
654 * We don't want to drop btree node pointers - if the
655 * btree node isn't there anymore, the read path will
658 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
659 bkey_for_each_ptr(ptrs, ptr) {
660 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
661 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
663 ptr->gen = g->mark.gen;
666 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
667 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
668 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
669 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
672 (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
674 gen_cmp(ptr->gen, g->mark.gen) < 0) ||
675 gen_cmp(g->mark.gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
676 (g->mark.data_type &&
677 g->mark.data_type != data_type);
680 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
681 bkey_extent_entry_for_each(ptrs, entry) {
682 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
683 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
684 entry->stripe_ptr.idx);
685 union bch_extent_entry *next_ptr;
687 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
688 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
693 bch_err(c, "aieee, found stripe ptr with no data ptr");
697 if (!m || !m->alive ||
698 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
701 bch2_bkey_extent_entry_drop(new, entry);
708 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
715 bch2_btree_node_update_key_early(c, btree_id, level - 1, *k, new);
717 bch2_bkey_val_to_text(&PBUF(buf), c, *k);
718 bch_info(c, "updated %s", buf);
719 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(new));
720 bch_info(c, "new key %s", buf);
721 *k = bkey_i_to_s_c(new);
727 /* marking of btree keys/nodes: */
729 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
730 unsigned level, bool is_root,
732 u8 *max_stale, bool initial)
734 struct bch_fs *c = trans->c;
735 struct bkey_ptrs_c ptrs;
736 const struct bch_extent_ptr *ptr;
737 struct bkey deleted = KEY(0, 0, 0);
738 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
741 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
747 BUG_ON(bch2_journal_seq_verify &&
748 k->k->version.lo > journal_cur_seq(&c->journal));
750 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
754 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
755 "key version number higher than recorded: %llu > %llu",
757 atomic64_read(&c->key_version)))
758 atomic64_set(&c->key_version, k->k->version.lo);
761 ptrs = bch2_bkey_ptrs_c(*k);
762 bkey_for_each_ptr(ptrs, ptr) {
763 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
764 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
766 if (gen_after(g->oldest_gen, ptr->gen))
767 g->oldest_gen = ptr->gen;
769 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
772 ret = bch2_mark_key(trans, old, *k, flags);
776 bch_err(c, "%s: ret %i", __func__, ret);
780 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, u8 *max_stale,
783 struct bch_fs *c = trans->c;
784 struct btree_node_iter iter;
785 struct bkey unpacked;
787 struct bkey_buf prev, cur;
792 if (!btree_node_type_needs_gc(btree_node_type(b)))
795 bch2_btree_node_iter_init_from_start(&iter, b);
796 bch2_bkey_buf_init(&prev);
797 bch2_bkey_buf_init(&cur);
798 bkey_init(&prev.k->k);
800 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
801 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
802 &k, max_stale, initial);
806 bch2_btree_node_iter_advance(&iter, b);
809 bch2_bkey_buf_reassemble(&cur, c, k);
811 ret = bch2_gc_check_topology(c, b, &prev, cur,
812 bch2_btree_node_iter_end(&iter));
818 bch2_bkey_buf_exit(&cur, c);
819 bch2_bkey_buf_exit(&prev, c);
823 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
824 bool initial, bool metadata_only)
826 struct bch_fs *c = trans->c;
827 struct btree_iter iter;
829 unsigned depth = metadata_only ? 1
830 : bch2_expensive_debug_checks ? 0
831 : !btree_node_type_needs_gc(btree_id) ? 1
836 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
838 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
839 0, depth, BTREE_ITER_PREFETCH, b, ret) {
840 bch2_verify_btree_nr_keys(b);
842 gc_pos_set(c, gc_pos_btree_node(b));
844 ret = btree_gc_mark_node(trans, b, &max_stale, initial);
850 bch2_btree_node_rewrite(trans, &iter, b,
852 BTREE_INSERT_GC_LOCK_HELD);
853 else if (!bch2_btree_gc_rewrite_disabled &&
854 (bch2_btree_gc_always_rewrite || max_stale > 16))
855 bch2_btree_node_rewrite(trans, &iter,
856 b, BTREE_INSERT_NOWAIT|
857 BTREE_INSERT_GC_LOCK_HELD);
860 bch2_trans_iter_exit(trans, &iter);
865 mutex_lock(&c->btree_root_lock);
866 b = c->btree_roots[btree_id].b;
867 if (!btree_node_fake(b)) {
868 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
870 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
871 &k, &max_stale, initial);
873 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
874 mutex_unlock(&c->btree_root_lock);
879 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
880 unsigned target_depth)
882 struct bch_fs *c = trans->c;
883 struct btree_and_journal_iter iter;
885 struct bkey_buf cur, prev;
890 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
891 bch2_bkey_buf_init(&prev);
892 bch2_bkey_buf_init(&cur);
893 bkey_init(&prev.k->k);
895 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
896 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
897 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
899 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
900 &k, &max_stale, true);
902 bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
907 bch2_bkey_buf_reassemble(&cur, c, k);
908 k = bkey_i_to_s_c(cur.k);
910 bch2_btree_and_journal_iter_advance(&iter);
912 ret = bch2_gc_check_topology(c, b,
914 !bch2_btree_and_journal_iter_peek(&iter).k);
918 bch2_btree_and_journal_iter_advance(&iter);
922 if (b->c.level > target_depth) {
923 bch2_btree_and_journal_iter_exit(&iter);
924 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
926 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
929 bch2_bkey_buf_reassemble(&cur, c, k);
930 bch2_btree_and_journal_iter_advance(&iter);
932 child = bch2_btree_node_get_noiter(c, cur.k,
933 b->c.btree_id, b->c.level - 1,
935 ret = PTR_ERR_OR_ZERO(child);
938 bch2_topology_error(c);
944 "Unreadable btree node at btree %s level %u:\n"
946 bch2_btree_ids[b->c.btree_id],
948 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf)) &&
949 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
950 ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
951 bch_info(c, "Halting mark and sweep to start topology repair pass");
954 /* Continue marking when opted to not
957 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
961 bch_err(c, "%s: error %i getting btree node",
966 ret = bch2_gc_btree_init_recurse(trans, child,
968 six_unlock_read(&child->c.lock);
975 bch2_bkey_buf_exit(&cur, c);
976 bch2_bkey_buf_exit(&prev, c);
977 bch2_btree_and_journal_iter_exit(&iter);
981 static int bch2_gc_btree_init(struct btree_trans *trans,
982 enum btree_id btree_id,
985 struct bch_fs *c = trans->c;
987 unsigned target_depth = metadata_only ? 1
988 : bch2_expensive_debug_checks ? 0
989 : !btree_node_type_needs_gc(btree_id) ? 1
995 b = c->btree_roots[btree_id].b;
997 if (btree_node_fake(b))
1000 six_lock_read(&b->c.lock, NULL, NULL);
1001 if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
1002 "btree root with incorrect min_key: %s",
1003 (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
1004 bch_err(c, "repair unimplemented");
1005 ret = FSCK_ERR_EXIT;
1009 if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
1010 "btree root with incorrect max_key: %s",
1011 (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
1012 bch_err(c, "repair unimplemented");
1013 ret = FSCK_ERR_EXIT;
1017 if (b->c.level >= target_depth)
1018 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1021 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1023 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1024 &k, &max_stale, true);
1027 six_unlock_read(&b->c.lock);
1030 bch_err(c, "%s: ret %i", __func__, ret);
1034 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1036 return (int) btree_id_to_gc_phase(l) -
1037 (int) btree_id_to_gc_phase(r);
1040 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1042 struct btree_trans trans;
1043 enum btree_id ids[BTREE_ID_NR];
1047 bch2_trans_init(&trans, c, 0, 0);
1049 for (i = 0; i < BTREE_ID_NR; i++)
1051 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1053 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1055 ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1056 : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1059 bch_err(c, "%s: ret %i", __func__, ret);
1061 bch2_trans_exit(&trans);
1065 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1067 enum bch_data_type type,
1070 u64 b = sector_to_bucket(ca, start);
1074 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1076 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1077 gc_phase(GC_PHASE_SB), flags);
1080 } while (start < end);
1083 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1086 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1090 for (i = 0; i < layout->nr_superblocks; i++) {
1091 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1093 if (offset == BCH_SB_SECTOR)
1094 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1095 BCH_DATA_sb, flags);
1097 mark_metadata_sectors(c, ca, offset,
1098 offset + (1 << layout->sb_max_size_bits),
1099 BCH_DATA_sb, flags);
1102 for (i = 0; i < ca->journal.nr; i++) {
1103 b = ca->journal.buckets[i];
1104 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1106 gc_phase(GC_PHASE_SB), flags);
1110 static void bch2_mark_superblocks(struct bch_fs *c)
1115 mutex_lock(&c->sb_lock);
1116 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1118 for_each_online_member(ca, c, i)
1119 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1120 mutex_unlock(&c->sb_lock);
1124 /* Also see bch2_pending_btree_node_free_insert_done() */
1125 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1127 struct btree_update *as;
1128 struct pending_btree_node_free *d;
1130 mutex_lock(&c->btree_interior_update_lock);
1131 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1133 for_each_pending_btree_node_free(c, as, d)
1134 if (d->index_update_done)
1135 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1137 mutex_unlock(&c->btree_interior_update_lock);
1141 static void bch2_gc_free(struct bch_fs *c)
1146 genradix_free(&c->reflink_gc_table);
1147 genradix_free(&c->gc_stripes);
1149 for_each_member_device(ca, c, i) {
1150 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1151 sizeof(struct bucket_array) +
1152 ca->mi.nbuckets * sizeof(struct bucket));
1153 ca->buckets[1] = NULL;
1155 free_percpu(ca->usage_gc);
1156 ca->usage_gc = NULL;
1159 free_percpu(c->usage_gc);
1163 static int bch2_gc_done(struct bch_fs *c,
1164 bool initial, bool metadata_only)
1166 struct bch_dev *ca = NULL;
1167 bool verify = !metadata_only && (!initial ||
1168 (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1172 #define copy_field(_f, _msg, ...) \
1173 if (dst->_f != src->_f) { \
1175 fsck_err(c, _msg ": got %llu, should be %llu" \
1176 , ##__VA_ARGS__, dst->_f, src->_f); \
1177 dst->_f = src->_f; \
1178 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
1180 #define copy_stripe_field(_f, _msg, ...) \
1181 if (dst->_f != src->_f) { \
1183 fsck_err(c, "stripe %zu has wrong "_msg \
1184 ": got %u, should be %u", \
1185 iter.pos, ##__VA_ARGS__, \
1186 dst->_f, src->_f); \
1187 dst->_f = src->_f; \
1188 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
1190 #define copy_bucket_field(_f) \
1191 if (dst->b[b]._f != src->b[b]._f) { \
1193 fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f \
1194 ": got %u, should be %u", dev, b, \
1195 dst->b[b].mark.gen, \
1196 bch2_data_types[dst->b[b].mark.data_type],\
1197 dst->b[b]._f, src->b[b]._f); \
1198 dst->b[b]._f = src->b[b]._f; \
1199 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
1201 #define copy_dev_field(_f, _msg, ...) \
1202 copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1203 #define copy_fs_field(_f, _msg, ...) \
1204 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1206 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1207 bch2_fs_usage_acc_to_base(c, i);
1209 for_each_member_device(ca, c, dev) {
1210 struct bucket_array *dst = __bucket_array(ca, 0);
1211 struct bucket_array *src = __bucket_array(ca, 1);
1214 for (b = 0; b < src->nbuckets; b++) {
1215 copy_bucket_field(_mark.gen);
1216 copy_bucket_field(_mark.data_type);
1217 copy_bucket_field(_mark.stripe);
1218 copy_bucket_field(_mark.dirty_sectors);
1219 copy_bucket_field(_mark.cached_sectors);
1220 copy_bucket_field(stripe_redundancy);
1221 copy_bucket_field(stripe);
1223 dst->b[b].oldest_gen = src->b[b].oldest_gen;
1227 struct bch_dev_usage *dst = ca->usage_base;
1228 struct bch_dev_usage *src = (void *)
1229 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1232 copy_dev_field(buckets_ec, "buckets_ec");
1233 copy_dev_field(buckets_unavailable, "buckets_unavailable");
1235 for (i = 0; i < BCH_DATA_NR; i++) {
1236 copy_dev_field(d[i].buckets, "%s buckets", bch2_data_types[i]);
1237 copy_dev_field(d[i].sectors, "%s sectors", bch2_data_types[i]);
1238 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1244 unsigned nr = fs_usage_u64s(c);
1245 struct bch_fs_usage *dst = c->usage_base;
1246 struct bch_fs_usage *src = (void *)
1247 bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1249 copy_fs_field(hidden, "hidden");
1250 copy_fs_field(btree, "btree");
1252 if (!metadata_only) {
1253 copy_fs_field(data, "data");
1254 copy_fs_field(cached, "cached");
1255 copy_fs_field(reserved, "reserved");
1256 copy_fs_field(nr_inodes,"nr_inodes");
1258 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1259 copy_fs_field(persistent_reserved[i],
1260 "persistent_reserved[%i]", i);
1263 for (i = 0; i < c->replicas.nr; i++) {
1264 struct bch_replicas_entry *e =
1265 cpu_replicas_entry(&c->replicas, i);
1268 if (metadata_only &&
1269 (e->data_type == BCH_DATA_user ||
1270 e->data_type == BCH_DATA_cached))
1273 bch2_replicas_entry_to_text(&PBUF(buf), e);
1275 copy_fs_field(replicas[i], "%s", buf);
1279 #undef copy_fs_field
1280 #undef copy_dev_field
1281 #undef copy_bucket_field
1282 #undef copy_stripe_field
1286 percpu_ref_put(&ca->ref);
1288 bch_err(c, "%s: ret %i", __func__, ret);
1292 static int bch2_gc_start(struct bch_fs *c,
1295 struct bch_dev *ca = NULL;
1298 BUG_ON(c->usage_gc);
1300 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1301 sizeof(u64), GFP_KERNEL);
1303 bch_err(c, "error allocating c->usage_gc");
1307 for_each_member_device(ca, c, i) {
1308 BUG_ON(ca->buckets[1]);
1309 BUG_ON(ca->usage_gc);
1311 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
1312 ca->mi.nbuckets * sizeof(struct bucket),
1313 GFP_KERNEL|__GFP_ZERO);
1314 if (!ca->buckets[1]) {
1315 percpu_ref_put(&ca->ref);
1316 bch_err(c, "error allocating ca->buckets[gc]");
1320 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1321 if (!ca->usage_gc) {
1322 bch_err(c, "error allocating ca->usage_gc");
1323 percpu_ref_put(&ca->ref);
1328 percpu_down_write(&c->mark_lock);
1331 * indicate to stripe code that we need to allocate for the gc stripes
1334 gc_pos_set(c, gc_phase(GC_PHASE_START));
1336 for_each_member_device(ca, c, i) {
1337 struct bucket_array *dst = __bucket_array(ca, 1);
1338 struct bucket_array *src = __bucket_array(ca, 0);
1341 dst->first_bucket = src->first_bucket;
1342 dst->nbuckets = src->nbuckets;
1344 for (b = 0; b < src->nbuckets; b++) {
1345 struct bucket *d = &dst->b[b];
1346 struct bucket *s = &src->b[b];
1348 d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
1349 d->gen_valid = s->gen_valid;
1351 if (metadata_only &&
1352 (s->mark.data_type == BCH_DATA_user ||
1353 s->mark.data_type == BCH_DATA_cached))
1358 percpu_up_write(&c->mark_lock);
1363 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1366 struct btree_trans trans;
1367 struct btree_iter iter;
1369 struct reflink_gc *r;
1377 bch2_trans_init(&trans, c, 0, 0);
1379 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1380 BTREE_ITER_PREFETCH, k, ret) {
1381 const __le64 *refcount = bkey_refcount_c(k);
1386 r = genradix_ptr(&c->reflink_gc_table, idx++);
1388 r->offset != k.k->p.offset ||
1389 r->size != k.k->size) {
1390 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1395 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1396 "reflink key has wrong refcount:\n"
1399 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1403 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1409 bkey_reassemble(new, k);
1412 new->k.type = KEY_TYPE_deleted;
1414 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1417 ? bch2_journal_key_insert(c, BTREE_ID_stripes, 0, new)
1418 : __bch2_trans_do(&trans, NULL, NULL, 0,
1419 __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1427 bch2_trans_iter_exit(&trans, &iter);
1428 c->reflink_gc_nr = 0;
1429 bch2_trans_exit(&trans);
1433 static int bch2_gc_stripes_done(struct bch_fs *c, bool initial,
1436 struct btree_trans trans;
1437 struct btree_iter iter;
1439 struct gc_stripe *m;
1440 const struct bch_stripe *s;
1448 bch2_trans_init(&trans, c, 0, 0);
1450 for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1451 BTREE_ITER_PREFETCH, k, ret) {
1452 if (k.k->type != KEY_TYPE_stripe)
1455 s = bkey_s_c_to_stripe(k).v;
1456 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1458 for (i = 0; i < s->nr_blocks; i++)
1459 if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1463 if (fsck_err_on(true, c,
1464 "stripe has wrong block sector count %u:\n"
1467 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1468 m ? m->block_sectors[i] : 0)) {
1469 struct bkey_i_stripe *new;
1471 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1477 bkey_reassemble(&new->k_i, k);
1479 for (i = 0; i < new->v.nr_blocks; i++)
1480 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1483 ? bch2_journal_key_insert(c, BTREE_ID_stripes, 0, &new->k_i)
1484 : __bch2_trans_do(&trans, NULL, NULL, 0,
1485 __bch2_btree_insert(&trans, BTREE_ID_reflink, &new->k_i));
1490 bch2_trans_iter_exit(&trans, &iter);
1492 bch2_trans_exit(&trans);
1496 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1499 struct btree_trans trans;
1500 struct btree_iter iter;
1502 struct reflink_gc *r;
1508 bch2_trans_init(&trans, c, 0, 0);
1509 c->reflink_gc_nr = 0;
1511 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1512 BTREE_ITER_PREFETCH, k, ret) {
1513 const __le64 *refcount = bkey_refcount_c(k);
1518 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1525 r->offset = k.k->p.offset;
1526 r->size = k.k->size;
1529 bch2_trans_iter_exit(&trans, &iter);
1531 bch2_trans_exit(&trans);
1536 * bch2_gc - walk _all_ references to buckets, and recompute them:
1538 * Order matters here:
1539 * - Concurrent GC relies on the fact that we have a total ordering for
1540 * everything that GC walks - see gc_will_visit_node(),
1541 * gc_will_visit_root()
1543 * - also, references move around in the course of index updates and
1544 * various other crap: everything needs to agree on the ordering
1545 * references are allowed to move around in - e.g., we're allowed to
1546 * start with a reference owned by an open_bucket (the allocator) and
1547 * move it to the btree, but not the reverse.
1549 * This is necessary to ensure that gc doesn't miss references that
1550 * move around - if references move backwards in the ordering GC
1551 * uses, GC could skip past them
1553 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1556 u64 start_time = local_clock();
1557 unsigned i, iter = 0;
1560 lockdep_assert_held(&c->state_lock);
1563 down_write(&c->gc_lock);
1565 /* flush interior btree updates: */
1566 closure_wait_event(&c->btree_interior_update_wait,
1567 !bch2_btree_interior_updates_nr_pending(c));
1569 ret = bch2_gc_start(c, metadata_only) ?:
1570 bch2_gc_reflink_start(c, initial, metadata_only);
1574 bch2_mark_superblocks(c);
1576 if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1577 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1578 c->opts.fix_errors != FSCK_OPT_NO) {
1579 bch_info(c, "starting topology repair pass");
1580 ret = bch2_repair_topology(c);
1583 bch_info(c, "topology repair pass done");
1585 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1588 ret = bch2_gc_btrees(c, initial, metadata_only);
1590 if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1591 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1592 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1593 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1597 if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1598 ret = FSCK_ERR_EXIT;
1604 bch2_mark_pending_btree_node_frees(c);
1608 if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1609 (!iter && bch2_test_restart_gc)) {
1611 * XXX: make sure gens we fixed got saved
1614 bch_info(c, "Second GC pass needed, restarting:");
1615 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1616 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1618 percpu_down_write(&c->mark_lock);
1620 percpu_up_write(&c->mark_lock);
1621 /* flush fsck errors, reset counters */
1622 bch2_flush_fsck_errs(c);
1627 bch_info(c, "Unable to fix bucket gens, looping");
1632 bch2_journal_block(&c->journal);
1634 percpu_down_write(&c->mark_lock);
1635 ret = bch2_gc_reflink_done(c, initial, metadata_only) ?:
1636 bch2_gc_stripes_done(c, initial, metadata_only) ?:
1637 bch2_gc_done(c, initial, metadata_only);
1639 bch2_journal_unblock(&c->journal);
1641 percpu_down_write(&c->mark_lock);
1644 /* Indicates that gc is no longer in progress: */
1645 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1648 percpu_up_write(&c->mark_lock);
1650 up_write(&c->gc_lock);
1653 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1656 * Wake up allocator in case it was waiting for buckets
1657 * because of not being able to inc gens
1659 for_each_member_device(ca, c, i)
1660 bch2_wake_allocator(ca);
1663 * At startup, allocations can happen directly instead of via the
1664 * allocator thread - issue wakeup in case they blocked on gc_lock:
1666 closure_wake_up(&c->freelist_wait);
1670 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1672 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1673 const struct bch_extent_ptr *ptr;
1675 percpu_down_read(&c->mark_lock);
1676 bkey_for_each_ptr(ptrs, ptr) {
1677 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1678 struct bucket *g = PTR_BUCKET(ca, ptr);
1680 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1681 percpu_up_read(&c->mark_lock);
1686 bkey_for_each_ptr(ptrs, ptr) {
1687 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1688 struct bucket *g = PTR_BUCKET(ca, ptr);
1690 if (gen_after(g->gc_gen, ptr->gen))
1691 g->gc_gen = ptr->gen;
1693 percpu_up_read(&c->mark_lock);
1699 * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1700 * node pointers currently never have cached pointers that can become stale:
1702 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1704 struct btree_trans trans;
1705 struct btree_iter iter;
1708 int ret = 0, commit_err = 0;
1710 bch2_bkey_buf_init(&sk);
1711 bch2_trans_init(&trans, c, 0, 0);
1713 bch2_trans_iter_init(&trans, &iter, btree_id, POS_MIN,
1714 BTREE_ITER_PREFETCH|
1715 BTREE_ITER_NOT_EXTENTS|
1716 BTREE_ITER_ALL_SNAPSHOTS);
1718 while ((bch2_trans_begin(&trans),
1719 k = bch2_btree_iter_peek(&iter)).k) {
1727 c->gc_gens_pos = iter.pos;
1729 if (gc_btree_gens_key(c, k) && !commit_err) {
1730 bch2_bkey_buf_reassemble(&sk, c, k);
1731 bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1734 bch2_trans_update(&trans, &iter, sk.k, 0) ?:
1735 bch2_trans_commit(&trans, NULL, NULL,
1736 BTREE_INSERT_NOWAIT|
1737 BTREE_INSERT_NOFAIL);
1738 if (commit_err == -EINTR) {
1744 bch2_btree_iter_advance(&iter);
1746 bch2_trans_iter_exit(&trans, &iter);
1748 bch2_trans_exit(&trans);
1749 bch2_bkey_buf_exit(&sk, c);
1754 int bch2_gc_gens(struct bch_fs *c)
1757 struct bucket_array *buckets;
1759 u64 start_time = local_clock();
1764 * Ideally we would be using state_lock and not gc_lock here, but that
1765 * introduces a deadlock in the RO path - we currently take the state
1766 * lock at the start of going RO, thus the gc thread may get stuck:
1768 down_read(&c->gc_lock);
1770 for_each_member_device(ca, c, i) {
1771 down_read(&ca->bucket_lock);
1772 buckets = bucket_array(ca);
1774 for_each_bucket(g, buckets)
1775 g->gc_gen = g->mark.gen;
1776 up_read(&ca->bucket_lock);
1779 for (i = 0; i < BTREE_ID_NR; i++)
1780 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1781 c->gc_gens_btree = i;
1782 c->gc_gens_pos = POS_MIN;
1783 ret = bch2_gc_btree_gens(c, i);
1785 bch_err(c, "error recalculating oldest_gen: %i", ret);
1790 for_each_member_device(ca, c, i) {
1791 down_read(&ca->bucket_lock);
1792 buckets = bucket_array(ca);
1794 for_each_bucket(g, buckets)
1795 g->oldest_gen = g->gc_gen;
1796 up_read(&ca->bucket_lock);
1799 c->gc_gens_btree = 0;
1800 c->gc_gens_pos = POS_MIN;
1804 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1806 up_read(&c->gc_lock);
1810 static int bch2_gc_thread(void *arg)
1812 struct bch_fs *c = arg;
1813 struct io_clock *clock = &c->io_clock[WRITE];
1814 unsigned long last = atomic64_read(&clock->now);
1815 unsigned last_kick = atomic_read(&c->kick_gc);
1822 set_current_state(TASK_INTERRUPTIBLE);
1824 if (kthread_should_stop()) {
1825 __set_current_state(TASK_RUNNING);
1829 if (atomic_read(&c->kick_gc) != last_kick)
1832 if (c->btree_gc_periodic) {
1833 unsigned long next = last + c->capacity / 16;
1835 if (atomic64_read(&clock->now) >= next)
1838 bch2_io_clock_schedule_timeout(clock, next);
1845 __set_current_state(TASK_RUNNING);
1847 last = atomic64_read(&clock->now);
1848 last_kick = atomic_read(&c->kick_gc);
1851 * Full gc is currently incompatible with btree key cache:
1854 ret = bch2_gc(c, false, false);
1856 ret = bch2_gc_gens(c);
1859 bch_err(c, "btree gc failed: %i", ret);
1861 debug_check_no_locks_held();
1867 void bch2_gc_thread_stop(struct bch_fs *c)
1869 struct task_struct *p;
1872 c->gc_thread = NULL;
1880 int bch2_gc_thread_start(struct bch_fs *c)
1882 struct task_struct *p;
1887 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1889 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));