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 int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
160 struct bkey_i_btree_ptr_v2 *new;
163 new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
167 btree_ptr_to_v2(b, new);
168 b->data->min_key = new_min;
169 new->v.min_key = new_min;
170 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
172 ret = bch2_journal_key_insert(c, b->c.btree_id, b->c.level + 1, &new->k_i);
178 bch2_btree_node_drop_keys_outside_node(b);
183 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
185 struct bkey_i_btree_ptr_v2 *new;
188 ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
192 new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
196 btree_ptr_to_v2(b, new);
197 b->data->max_key = new_max;
199 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
201 ret = bch2_journal_key_insert(c, b->c.btree_id, b->c.level + 1, &new->k_i);
207 bch2_btree_node_drop_keys_outside_node(b);
209 mutex_lock(&c->btree_cache.lock);
210 bch2_btree_node_hash_remove(&c->btree_cache, b);
212 bkey_copy(&b->key, &new->k_i);
213 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
215 mutex_unlock(&c->btree_cache.lock);
219 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
220 struct btree *prev, struct btree *cur)
222 struct bpos expected_start = !prev
224 : bpos_successor(prev->key.k.p);
225 char buf1[200], buf2[200];
229 struct printbuf out = PBUF(buf1);
230 pr_buf(&out, "start of node: ");
231 bch2_bpos_to_text(&out, b->data->min_key);
233 bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&prev->key));
236 bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&cur->key));
239 bpos_cmp(expected_start, cur->data->min_key) > 0 &&
240 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
241 /* cur overwrites prev: */
243 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
244 cur->data->min_key) >= 0, c,
245 "btree node overwritten by next node at btree %s level %u:\n"
248 bch2_btree_ids[b->c.btree_id], b->c.level,
250 return DROP_PREV_NODE;
252 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
253 bpos_predecessor(cur->data->min_key)), c,
254 "btree node with incorrect max_key at btree %s level %u:\n"
257 bch2_btree_ids[b->c.btree_id], b->c.level,
259 ret = set_node_max(c, prev,
260 bpos_predecessor(cur->data->min_key));
262 /* prev overwrites cur: */
264 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
265 cur->data->max_key) >= 0, c,
266 "btree node overwritten by prev node at btree %s level %u:\n"
269 bch2_btree_ids[b->c.btree_id], b->c.level,
271 return DROP_THIS_NODE;
273 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
274 "btree node with incorrect min_key at btree %s level %u:\n"
277 bch2_btree_ids[b->c.btree_id], b->c.level,
279 ret = set_node_min(c, cur, expected_start);
285 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
288 char buf1[200], buf2[200];
291 if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
292 "btree node with incorrect max_key at btree %s level %u:\n"
295 bch2_btree_ids[b->c.btree_id], b->c.level,
296 (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&child->key)), buf1),
297 (bch2_bpos_to_text(&PBUF(buf2), b->key.k.p), buf2))) {
298 ret = set_node_max(c, child, b->key.k.p);
306 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
308 struct btree_and_journal_iter iter;
310 struct bkey_buf prev_k, cur_k;
311 struct btree *prev = NULL, *cur = NULL;
312 bool have_child, dropped_children = false;
320 have_child = dropped_children = false;
321 bch2_bkey_buf_init(&prev_k);
322 bch2_bkey_buf_init(&cur_k);
323 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
325 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
326 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
327 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
329 bch2_btree_and_journal_iter_advance(&iter);
330 bch2_bkey_buf_reassemble(&cur_k, c, k);
332 cur = bch2_btree_node_get_noiter(c, cur_k.k,
333 b->c.btree_id, b->c.level - 1,
335 ret = PTR_ERR_OR_ZERO(cur);
337 if (mustfix_fsck_err_on(ret == -EIO, c,
338 "Unreadable btree node at btree %s level %u:\n"
340 bch2_btree_ids[b->c.btree_id],
342 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur_k.k)), buf))) {
343 bch2_btree_node_evict(c, cur_k.k);
344 ret = bch2_journal_key_delete(c, b->c.btree_id,
345 b->c.level, cur_k.k->k.p);
352 bch_err(c, "%s: error %i getting btree node",
357 ret = btree_repair_node_boundaries(c, b, prev, cur);
359 if (ret == DROP_THIS_NODE) {
360 six_unlock_read(&cur->c.lock);
361 bch2_btree_node_evict(c, cur_k.k);
362 ret = bch2_journal_key_delete(c, b->c.btree_id,
363 b->c.level, cur_k.k->k.p);
370 six_unlock_read(&prev->c.lock);
373 if (ret == DROP_PREV_NODE) {
374 bch2_btree_node_evict(c, prev_k.k);
375 ret = bch2_journal_key_delete(c, b->c.btree_id,
376 b->c.level, prev_k.k->k.p);
380 bch2_btree_and_journal_iter_exit(&iter);
381 bch2_bkey_buf_exit(&prev_k, c);
382 bch2_bkey_buf_exit(&cur_k, c);
389 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
392 if (!ret && !IS_ERR_OR_NULL(prev)) {
394 ret = btree_repair_node_end(c, b, prev);
397 if (!IS_ERR_OR_NULL(prev))
398 six_unlock_read(&prev->c.lock);
400 if (!IS_ERR_OR_NULL(cur))
401 six_unlock_read(&cur->c.lock);
407 bch2_btree_and_journal_iter_exit(&iter);
408 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
410 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
411 bch2_bkey_buf_reassemble(&cur_k, c, k);
412 bch2_btree_and_journal_iter_advance(&iter);
414 cur = bch2_btree_node_get_noiter(c, cur_k.k,
415 b->c.btree_id, b->c.level - 1,
417 ret = PTR_ERR_OR_ZERO(cur);
420 bch_err(c, "%s: error %i getting btree node",
425 ret = bch2_btree_repair_topology_recurse(c, cur);
426 six_unlock_read(&cur->c.lock);
429 if (ret == DROP_THIS_NODE) {
430 bch2_btree_node_evict(c, cur_k.k);
431 ret = bch2_journal_key_delete(c, b->c.btree_id,
432 b->c.level, cur_k.k->k.p);
433 dropped_children = true;
442 if (mustfix_fsck_err_on(!have_child, c,
443 "empty interior btree node at btree %s level %u\n"
445 bch2_btree_ids[b->c.btree_id],
447 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(&b->key)), buf)))
448 ret = DROP_THIS_NODE;
451 if (!IS_ERR_OR_NULL(prev))
452 six_unlock_read(&prev->c.lock);
453 if (!IS_ERR_OR_NULL(cur))
454 six_unlock_read(&cur->c.lock);
456 bch2_btree_and_journal_iter_exit(&iter);
457 bch2_bkey_buf_exit(&prev_k, c);
458 bch2_bkey_buf_exit(&cur_k, c);
460 if (!ret && dropped_children)
466 static int bch2_repair_topology(struct bch_fs *c)
472 for (i = 0; i < BTREE_ID_NR && !ret; i++) {
473 b = c->btree_roots[i].b;
474 if (btree_node_fake(b))
477 six_lock_read(&b->c.lock, NULL, NULL);
478 ret = bch2_btree_repair_topology_recurse(c, b);
479 six_unlock_read(&b->c.lock);
481 if (ret == DROP_THIS_NODE) {
482 bch_err(c, "empty btree root - repair unimplemented");
490 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
491 unsigned level, bool is_root,
494 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
495 const union bch_extent_entry *entry;
496 struct extent_ptr_decoded p = { 0 };
497 bool do_update = false;
503 * use check_bucket_ref here
505 bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
506 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
507 struct bucket *g = PTR_BUCKET(ca, &p.ptr, true);
508 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr, false);
509 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
511 if (fsck_err_on(!g->gen_valid, c,
512 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
514 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
515 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
517 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
519 g2->_mark.gen = g->_mark.gen = p.ptr.gen;
520 g2->gen_valid = g->gen_valid = true;
521 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
527 if (fsck_err_on(data_type == BCH_DATA_btree &&
528 g->mark.gen != p.ptr.gen, c,
529 "bucket %u:%zu data type %s has metadata but wrong gen: %u != %u\n"
531 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
532 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
533 p.ptr.gen, g->mark.gen,
534 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
535 g2->_mark.data_type = g->_mark.data_type = data_type;
536 g2->gen_valid = g->gen_valid = true;
537 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
540 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
541 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
543 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
544 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
545 p.ptr.gen, g->mark.gen,
546 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
548 g2->_mark.gen = g->_mark.gen = p.ptr.gen;
549 g2->gen_valid = g->gen_valid = true;
550 g2->_mark.data_type = 0;
551 g2->_mark.dirty_sectors = 0;
552 g2->_mark.cached_sectors = 0;
553 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
554 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
560 if (fsck_err_on(gen_cmp(g->mark.gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
561 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
563 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->mark.gen,
564 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
566 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
569 if (fsck_err_on(!p.ptr.cached &&
570 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
571 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
573 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
574 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
575 p.ptr.gen, g->mark.gen,
576 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
579 if (p.ptr.gen != g->mark.gen)
582 if (fsck_err_on(g->mark.data_type &&
583 g->mark.data_type != data_type, c,
584 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
586 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
587 bch2_data_types[g->mark.data_type],
588 bch2_data_types[data_type],
589 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
590 if (data_type == BCH_DATA_btree) {
591 g2->_mark.data_type = g->_mark.data_type = data_type;
592 g2->gen_valid = g->gen_valid = true;
593 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
600 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
602 if (fsck_err_on(!m || !m->alive, c,
603 "pointer to nonexistent stripe %llu\n"
606 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
609 if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
610 "pointer does not match stripe %llu\n"
613 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
619 struct bkey_ptrs ptrs;
620 union bch_extent_entry *entry;
621 struct bch_extent_ptr *ptr;
625 bch_err(c, "cannot update btree roots yet");
629 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
631 bch_err(c, "%s: error allocating new key", __func__);
635 bkey_reassemble(new, *k);
639 * We don't want to drop btree node pointers - if the
640 * btree node isn't there anymore, the read path will
643 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
644 bkey_for_each_ptr(ptrs, ptr) {
645 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
646 struct bucket *g = PTR_BUCKET(ca, ptr, true);
648 ptr->gen = g->mark.gen;
651 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
652 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
653 struct bucket *g = PTR_BUCKET(ca, ptr, true);
654 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
657 (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
659 gen_cmp(ptr->gen, g->mark.gen) < 0) ||
660 gen_cmp(g->mark.gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
661 (g->mark.data_type &&
662 g->mark.data_type != data_type);
665 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
666 bkey_extent_entry_for_each(ptrs, entry) {
667 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
668 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
669 entry->stripe_ptr.idx);
670 union bch_extent_entry *next_ptr;
672 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
673 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
678 bch_err(c, "aieee, found stripe ptr with no data ptr");
682 if (!m || !m->alive ||
683 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
686 bch2_bkey_extent_entry_drop(new, entry);
693 ret = bch2_journal_key_insert(c, btree_id, level, new);
697 *k = bkey_i_to_s_c(new);
703 /* marking of btree keys/nodes: */
705 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
706 unsigned level, bool is_root,
708 u8 *max_stale, bool initial)
710 struct bch_fs *c = trans->c;
711 struct bkey_ptrs_c ptrs;
712 const struct bch_extent_ptr *ptr;
713 struct bkey deleted = KEY(0, 0, 0);
714 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
717 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
723 BUG_ON(bch2_journal_seq_verify &&
724 k->k->version.lo > journal_cur_seq(&c->journal));
726 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
730 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
731 "key version number higher than recorded: %llu > %llu",
733 atomic64_read(&c->key_version)))
734 atomic64_set(&c->key_version, k->k->version.lo);
737 ptrs = bch2_bkey_ptrs_c(*k);
738 bkey_for_each_ptr(ptrs, ptr) {
739 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
740 struct bucket *g = PTR_BUCKET(ca, ptr, true);
742 if (gen_after(g->oldest_gen, ptr->gen))
743 g->oldest_gen = ptr->gen;
745 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
748 ret = bch2_mark_key(trans, old, *k, flags);
752 bch_err(c, "%s: ret %i", __func__, ret);
756 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, u8 *max_stale,
759 struct bch_fs *c = trans->c;
760 struct btree_node_iter iter;
761 struct bkey unpacked;
763 struct bkey_buf prev, cur;
768 if (!btree_node_type_needs_gc(btree_node_type(b)))
771 bch2_btree_node_iter_init_from_start(&iter, b);
772 bch2_bkey_buf_init(&prev);
773 bch2_bkey_buf_init(&cur);
774 bkey_init(&prev.k->k);
776 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
777 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
778 &k, max_stale, initial);
782 bch2_btree_node_iter_advance(&iter, b);
785 bch2_bkey_buf_reassemble(&cur, c, k);
787 ret = bch2_gc_check_topology(c, b, &prev, cur,
788 bch2_btree_node_iter_end(&iter));
794 bch2_bkey_buf_exit(&cur, c);
795 bch2_bkey_buf_exit(&prev, c);
799 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
800 bool initial, bool metadata_only)
802 struct bch_fs *c = trans->c;
803 struct btree_iter iter;
805 unsigned depth = metadata_only ? 1
806 : bch2_expensive_debug_checks ? 0
807 : !btree_node_type_needs_gc(btree_id) ? 1
812 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
814 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
815 0, depth, BTREE_ITER_PREFETCH, b, ret) {
816 bch2_verify_btree_nr_keys(b);
818 gc_pos_set(c, gc_pos_btree_node(b));
820 ret = btree_gc_mark_node(trans, b, &max_stale, initial);
826 bch2_btree_node_rewrite(trans, &iter, b,
828 BTREE_INSERT_GC_LOCK_HELD);
829 else if (!bch2_btree_gc_rewrite_disabled &&
830 (bch2_btree_gc_always_rewrite || max_stale > 16))
831 bch2_btree_node_rewrite(trans, &iter,
832 b, BTREE_INSERT_NOWAIT|
833 BTREE_INSERT_GC_LOCK_HELD);
836 bch2_trans_iter_exit(trans, &iter);
841 mutex_lock(&c->btree_root_lock);
842 b = c->btree_roots[btree_id].b;
843 if (!btree_node_fake(b)) {
844 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
846 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
847 &k, &max_stale, initial);
849 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
850 mutex_unlock(&c->btree_root_lock);
855 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
856 unsigned target_depth)
858 struct bch_fs *c = trans->c;
859 struct btree_and_journal_iter iter;
861 struct bkey_buf cur, prev;
866 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
867 bch2_bkey_buf_init(&prev);
868 bch2_bkey_buf_init(&cur);
869 bkey_init(&prev.k->k);
871 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
872 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
873 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
875 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
876 &k, &max_stale, true);
878 bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
883 bch2_bkey_buf_reassemble(&cur, c, k);
884 k = bkey_i_to_s_c(cur.k);
886 bch2_btree_and_journal_iter_advance(&iter);
888 ret = bch2_gc_check_topology(c, b,
890 !bch2_btree_and_journal_iter_peek(&iter).k);
894 bch2_btree_and_journal_iter_advance(&iter);
898 if (b->c.level > target_depth) {
899 bch2_btree_and_journal_iter_exit(&iter);
900 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
902 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
905 bch2_bkey_buf_reassemble(&cur, c, k);
906 bch2_btree_and_journal_iter_advance(&iter);
908 child = bch2_btree_node_get_noiter(c, cur.k,
909 b->c.btree_id, b->c.level - 1,
911 ret = PTR_ERR_OR_ZERO(child);
914 bch2_topology_error(c);
920 "Unreadable btree node at btree %s level %u:\n"
922 bch2_btree_ids[b->c.btree_id],
924 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf)) &&
925 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
926 ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
927 bch_info(c, "Halting mark and sweep to start topology repair pass");
930 /* Continue marking when opted to not
933 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
937 bch_err(c, "%s: error %i getting btree node",
942 ret = bch2_gc_btree_init_recurse(trans, child,
944 six_unlock_read(&child->c.lock);
951 bch2_bkey_buf_exit(&cur, c);
952 bch2_bkey_buf_exit(&prev, c);
953 bch2_btree_and_journal_iter_exit(&iter);
957 static int bch2_gc_btree_init(struct btree_trans *trans,
958 enum btree_id btree_id,
961 struct bch_fs *c = trans->c;
963 unsigned target_depth = metadata_only ? 1
964 : bch2_expensive_debug_checks ? 0
965 : !btree_node_type_needs_gc(btree_id) ? 1
971 b = c->btree_roots[btree_id].b;
973 if (btree_node_fake(b))
976 six_lock_read(&b->c.lock, NULL, NULL);
977 if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
978 "btree root with incorrect min_key: %s",
979 (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
980 bch_err(c, "repair unimplemented");
985 if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
986 "btree root with incorrect max_key: %s",
987 (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
988 bch_err(c, "repair unimplemented");
993 if (b->c.level >= target_depth)
994 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
997 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
999 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1000 &k, &max_stale, true);
1003 six_unlock_read(&b->c.lock);
1006 bch_err(c, "%s: ret %i", __func__, ret);
1010 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1012 return (int) btree_id_to_gc_phase(l) -
1013 (int) btree_id_to_gc_phase(r);
1016 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1018 struct btree_trans trans;
1019 enum btree_id ids[BTREE_ID_NR];
1023 bch2_trans_init(&trans, c, 0, 0);
1025 for (i = 0; i < BTREE_ID_NR; i++)
1027 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1029 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1031 ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1032 : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1035 bch_err(c, "%s: ret %i", __func__, ret);
1037 bch2_trans_exit(&trans);
1041 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1043 enum bch_data_type type,
1046 u64 b = sector_to_bucket(ca, start);
1050 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1052 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1053 gc_phase(GC_PHASE_SB), flags);
1056 } while (start < end);
1059 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1062 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1067 * This conditional is kind of gross, but we may be called from the
1068 * device add path, before the new device has actually been added to the
1069 * running filesystem:
1072 lockdep_assert_held(&c->sb_lock);
1073 percpu_down_read(&c->mark_lock);
1076 for (i = 0; i < layout->nr_superblocks; i++) {
1077 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1079 if (offset == BCH_SB_SECTOR)
1080 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1081 BCH_DATA_sb, flags);
1083 mark_metadata_sectors(c, ca, offset,
1084 offset + (1 << layout->sb_max_size_bits),
1085 BCH_DATA_sb, flags);
1088 for (i = 0; i < ca->journal.nr; i++) {
1089 b = ca->journal.buckets[i];
1090 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1092 gc_phase(GC_PHASE_SB), flags);
1096 percpu_up_read(&c->mark_lock);
1099 static void bch2_mark_superblocks(struct bch_fs *c)
1104 mutex_lock(&c->sb_lock);
1105 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1107 for_each_online_member(ca, c, i)
1108 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1109 mutex_unlock(&c->sb_lock);
1113 /* Also see bch2_pending_btree_node_free_insert_done() */
1114 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1116 struct btree_update *as;
1117 struct pending_btree_node_free *d;
1119 mutex_lock(&c->btree_interior_update_lock);
1120 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1122 for_each_pending_btree_node_free(c, as, d)
1123 if (d->index_update_done)
1124 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1126 mutex_unlock(&c->btree_interior_update_lock);
1130 static void bch2_gc_free(struct bch_fs *c)
1135 genradix_free(&c->reflink_gc_table);
1136 genradix_free(&c->gc_stripes);
1138 for_each_member_device(ca, c, i) {
1139 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1140 sizeof(struct bucket_array) +
1141 ca->mi.nbuckets * sizeof(struct bucket));
1142 ca->buckets[1] = NULL;
1144 free_percpu(ca->usage_gc);
1145 ca->usage_gc = NULL;
1148 free_percpu(c->usage_gc);
1152 static int bch2_gc_done(struct bch_fs *c,
1153 bool initial, bool metadata_only)
1155 struct bch_dev *ca = NULL;
1156 bool verify = !metadata_only && (!initial ||
1157 (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1161 #define copy_field(_f, _msg, ...) \
1162 if (dst->_f != src->_f) { \
1164 fsck_err(c, _msg ": got %llu, should be %llu" \
1165 , ##__VA_ARGS__, dst->_f, src->_f); \
1166 dst->_f = src->_f; \
1167 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
1169 #define copy_stripe_field(_f, _msg, ...) \
1170 if (dst->_f != src->_f) { \
1172 fsck_err(c, "stripe %zu has wrong "_msg \
1173 ": got %u, should be %u", \
1174 iter.pos, ##__VA_ARGS__, \
1175 dst->_f, src->_f); \
1176 dst->_f = src->_f; \
1177 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
1179 #define copy_bucket_field(_f) \
1180 if (dst->b[b]._f != src->b[b]._f) { \
1182 fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f \
1183 ": got %u, should be %u", dev, b, \
1184 dst->b[b].mark.gen, \
1185 bch2_data_types[dst->b[b].mark.data_type],\
1186 dst->b[b]._f, src->b[b]._f); \
1187 dst->b[b]._f = src->b[b]._f; \
1188 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
1190 #define copy_dev_field(_f, _msg, ...) \
1191 copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1192 #define copy_fs_field(_f, _msg, ...) \
1193 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1195 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1196 bch2_fs_usage_acc_to_base(c, i);
1198 for_each_member_device(ca, c, dev) {
1199 struct bucket_array *dst = __bucket_array(ca, 0);
1200 struct bucket_array *src = __bucket_array(ca, 1);
1203 for (b = 0; b < src->nbuckets; b++) {
1204 copy_bucket_field(_mark.gen);
1205 copy_bucket_field(_mark.data_type);
1206 copy_bucket_field(_mark.stripe);
1207 copy_bucket_field(_mark.dirty_sectors);
1208 copy_bucket_field(_mark.cached_sectors);
1209 copy_bucket_field(stripe_redundancy);
1210 copy_bucket_field(stripe);
1212 dst->b[b].oldest_gen = src->b[b].oldest_gen;
1216 struct bch_dev_usage *dst = ca->usage_base;
1217 struct bch_dev_usage *src = (void *)
1218 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1221 copy_dev_field(buckets_ec, "buckets_ec");
1222 copy_dev_field(buckets_unavailable, "buckets_unavailable");
1224 for (i = 0; i < BCH_DATA_NR; i++) {
1225 copy_dev_field(d[i].buckets, "%s buckets", bch2_data_types[i]);
1226 copy_dev_field(d[i].sectors, "%s sectors", bch2_data_types[i]);
1227 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1233 unsigned nr = fs_usage_u64s(c);
1234 struct bch_fs_usage *dst = c->usage_base;
1235 struct bch_fs_usage *src = (void *)
1236 bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1238 copy_fs_field(hidden, "hidden");
1239 copy_fs_field(btree, "btree");
1241 if (!metadata_only) {
1242 copy_fs_field(data, "data");
1243 copy_fs_field(cached, "cached");
1244 copy_fs_field(reserved, "reserved");
1245 copy_fs_field(nr_inodes,"nr_inodes");
1247 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1248 copy_fs_field(persistent_reserved[i],
1249 "persistent_reserved[%i]", i);
1252 for (i = 0; i < c->replicas.nr; i++) {
1253 struct bch_replicas_entry *e =
1254 cpu_replicas_entry(&c->replicas, i);
1257 if (metadata_only &&
1258 (e->data_type == BCH_DATA_user ||
1259 e->data_type == BCH_DATA_cached))
1262 bch2_replicas_entry_to_text(&PBUF(buf), e);
1264 copy_fs_field(replicas[i], "%s", buf);
1268 #undef copy_fs_field
1269 #undef copy_dev_field
1270 #undef copy_bucket_field
1271 #undef copy_stripe_field
1275 percpu_ref_put(&ca->ref);
1277 bch_err(c, "%s: ret %i", __func__, ret);
1281 static int bch2_gc_start(struct bch_fs *c,
1284 struct bch_dev *ca = NULL;
1288 BUG_ON(c->usage_gc);
1290 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1291 sizeof(u64), GFP_KERNEL);
1293 bch_err(c, "error allocating c->usage_gc");
1297 for_each_member_device(ca, c, i) {
1298 BUG_ON(ca->buckets[1]);
1299 BUG_ON(ca->usage_gc);
1301 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
1302 ca->mi.nbuckets * sizeof(struct bucket),
1303 GFP_KERNEL|__GFP_ZERO);
1304 if (!ca->buckets[1]) {
1305 percpu_ref_put(&ca->ref);
1306 bch_err(c, "error allocating ca->buckets[gc]");
1310 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1311 if (!ca->usage_gc) {
1312 bch_err(c, "error allocating ca->usage_gc");
1313 percpu_ref_put(&ca->ref);
1318 ret = bch2_ec_mem_alloc(c, true);
1320 bch_err(c, "error allocating ec gc mem");
1324 percpu_down_write(&c->mark_lock);
1327 * indicate to stripe code that we need to allocate for the gc stripes
1330 gc_pos_set(c, gc_phase(GC_PHASE_START));
1332 for_each_member_device(ca, c, i) {
1333 struct bucket_array *dst = __bucket_array(ca, 1);
1334 struct bucket_array *src = __bucket_array(ca, 0);
1337 dst->first_bucket = src->first_bucket;
1338 dst->nbuckets = src->nbuckets;
1340 for (b = 0; b < src->nbuckets; b++) {
1341 struct bucket *d = &dst->b[b];
1342 struct bucket *s = &src->b[b];
1344 d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
1345 d->gen_valid = s->gen_valid;
1347 if (metadata_only &&
1348 (s->mark.data_type == BCH_DATA_user ||
1349 s->mark.data_type == BCH_DATA_cached))
1354 percpu_up_write(&c->mark_lock);
1359 static int bch2_gc_reflink_done_initial_fn(struct btree_trans *trans,
1362 struct bch_fs *c = trans->c;
1363 struct reflink_gc *r;
1364 const __le64 *refcount = bkey_refcount_c(k);
1371 r = genradix_ptr(&c->reflink_gc_table, c->reflink_gc_idx++);
1376 r->offset != k.k->p.offset ||
1377 r->size != k.k->size) {
1378 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1382 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1383 "reflink key has wrong refcount:\n"
1386 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1390 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1396 bkey_reassemble(new, k);
1399 new->k.type = KEY_TYPE_deleted;
1402 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1405 ret = bch2_journal_key_insert(c, BTREE_ID_reflink, 0, new);
1413 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1416 struct btree_trans trans;
1417 struct btree_iter iter;
1419 struct reflink_gc *r;
1427 bch2_trans_init(&trans, c, 0, 0);
1430 c->reflink_gc_idx = 0;
1432 ret = bch2_btree_and_journal_walk(&trans, BTREE_ID_reflink,
1433 bch2_gc_reflink_done_initial_fn);
1437 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1438 BTREE_ITER_PREFETCH, k, ret) {
1439 const __le64 *refcount = bkey_refcount_c(k);
1444 r = genradix_ptr(&c->reflink_gc_table, idx);
1446 r->offset != k.k->p.offset ||
1447 r->size != k.k->size) {
1448 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1453 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1454 "reflink key has wrong refcount:\n"
1457 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1461 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1467 bkey_reassemble(new, k);
1470 new->k.type = KEY_TYPE_deleted;
1472 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1474 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1475 __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1483 bch2_trans_iter_exit(&trans, &iter);
1485 c->reflink_gc_nr = 0;
1486 bch2_trans_exit(&trans);
1490 static int bch2_gc_stripes_done_initial_fn(struct btree_trans *trans,
1493 struct bch_fs *c = trans->c;
1494 struct gc_stripe *m;
1495 const struct bch_stripe *s;
1500 if (k.k->type != KEY_TYPE_stripe)
1503 s = bkey_s_c_to_stripe(k).v;
1505 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1507 for (i = 0; i < s->nr_blocks; i++)
1508 if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1512 if (fsck_err_on(true, c,
1513 "stripe has wrong block sector count %u:\n"
1516 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1517 m ? m->block_sectors[i] : 0)) {
1518 struct bkey_i_stripe *new;
1520 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1526 bkey_reassemble(&new->k_i, k);
1528 for (i = 0; i < new->v.nr_blocks; i++)
1529 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1531 ret = bch2_journal_key_insert(c, BTREE_ID_stripes, 0, &new->k_i);
1539 static int bch2_gc_stripes_done(struct bch_fs *c, bool initial,
1542 struct btree_trans trans;
1548 bch2_trans_init(&trans, c, 0, 0);
1551 ret = bch2_btree_and_journal_walk(&trans, BTREE_ID_stripes,
1552 bch2_gc_stripes_done_initial_fn);
1557 bch2_trans_exit(&trans);
1561 static int bch2_gc_reflink_start_initial_fn(struct btree_trans *trans,
1565 struct bch_fs *c = trans->c;
1566 struct reflink_gc *r;
1567 const __le64 *refcount = bkey_refcount_c(k);
1572 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1577 r->offset = k.k->p.offset;
1578 r->size = k.k->size;
1583 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1586 struct btree_trans trans;
1587 struct btree_iter iter;
1589 struct reflink_gc *r;
1595 bch2_trans_init(&trans, c, 0, 0);
1596 c->reflink_gc_nr = 0;
1599 ret = bch2_btree_and_journal_walk(&trans, BTREE_ID_reflink,
1600 bch2_gc_reflink_start_initial_fn);
1604 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1605 BTREE_ITER_PREFETCH, k, ret) {
1606 const __le64 *refcount = bkey_refcount_c(k);
1611 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1618 r->offset = k.k->p.offset;
1619 r->size = k.k->size;
1622 bch2_trans_iter_exit(&trans, &iter);
1624 bch2_trans_exit(&trans);
1629 * bch2_gc - walk _all_ references to buckets, and recompute them:
1631 * Order matters here:
1632 * - Concurrent GC relies on the fact that we have a total ordering for
1633 * everything that GC walks - see gc_will_visit_node(),
1634 * gc_will_visit_root()
1636 * - also, references move around in the course of index updates and
1637 * various other crap: everything needs to agree on the ordering
1638 * references are allowed to move around in - e.g., we're allowed to
1639 * start with a reference owned by an open_bucket (the allocator) and
1640 * move it to the btree, but not the reverse.
1642 * This is necessary to ensure that gc doesn't miss references that
1643 * move around - if references move backwards in the ordering GC
1644 * uses, GC could skip past them
1646 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1649 u64 start_time = local_clock();
1650 unsigned i, iter = 0;
1653 lockdep_assert_held(&c->state_lock);
1656 down_write(&c->gc_lock);
1658 /* flush interior btree updates: */
1659 closure_wait_event(&c->btree_interior_update_wait,
1660 !bch2_btree_interior_updates_nr_pending(c));
1662 ret = bch2_gc_start(c, metadata_only) ?:
1663 bch2_gc_reflink_start(c, initial, metadata_only);
1667 bch2_mark_superblocks(c);
1669 if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1670 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1671 c->opts.fix_errors != FSCK_OPT_NO) {
1672 bch_info(c, "starting topology repair pass");
1673 ret = bch2_repair_topology(c);
1676 bch_info(c, "topology repair pass done");
1678 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1681 ret = bch2_gc_btrees(c, initial, metadata_only);
1683 if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1684 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1685 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1686 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1690 if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1691 ret = FSCK_ERR_EXIT;
1697 bch2_mark_pending_btree_node_frees(c);
1701 if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1702 (!iter && bch2_test_restart_gc)) {
1704 * XXX: make sure gens we fixed got saved
1707 bch_info(c, "Second GC pass needed, restarting:");
1708 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1709 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1711 percpu_down_write(&c->mark_lock);
1713 percpu_up_write(&c->mark_lock);
1714 /* flush fsck errors, reset counters */
1715 bch2_flush_fsck_errs(c);
1720 bch_info(c, "Unable to fix bucket gens, looping");
1725 bch2_journal_block(&c->journal);
1727 percpu_down_write(&c->mark_lock);
1728 ret = bch2_gc_reflink_done(c, initial, metadata_only) ?:
1729 bch2_gc_stripes_done(c, initial, metadata_only) ?:
1730 bch2_gc_done(c, initial, metadata_only);
1732 bch2_journal_unblock(&c->journal);
1734 percpu_down_write(&c->mark_lock);
1737 /* Indicates that gc is no longer in progress: */
1738 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1741 percpu_up_write(&c->mark_lock);
1743 up_write(&c->gc_lock);
1746 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1749 * Wake up allocator in case it was waiting for buckets
1750 * because of not being able to inc gens
1752 for_each_member_device(ca, c, i)
1753 bch2_wake_allocator(ca);
1756 * At startup, allocations can happen directly instead of via the
1757 * allocator thread - issue wakeup in case they blocked on gc_lock:
1759 closure_wake_up(&c->freelist_wait);
1763 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1765 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1766 const struct bch_extent_ptr *ptr;
1768 percpu_down_read(&c->mark_lock);
1769 bkey_for_each_ptr(ptrs, ptr) {
1770 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1771 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1773 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1774 percpu_up_read(&c->mark_lock);
1779 bkey_for_each_ptr(ptrs, ptr) {
1780 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1781 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1783 if (gen_after(g->gc_gen, ptr->gen))
1784 g->gc_gen = ptr->gen;
1786 percpu_up_read(&c->mark_lock);
1792 * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1793 * node pointers currently never have cached pointers that can become stale:
1795 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1797 struct btree_trans trans;
1798 struct btree_iter iter;
1801 int ret = 0, commit_err = 0;
1803 bch2_bkey_buf_init(&sk);
1804 bch2_trans_init(&trans, c, 0, 0);
1806 bch2_trans_iter_init(&trans, &iter, btree_id, POS_MIN,
1807 BTREE_ITER_PREFETCH|
1808 BTREE_ITER_NOT_EXTENTS|
1809 BTREE_ITER_ALL_SNAPSHOTS);
1811 while ((bch2_trans_begin(&trans),
1812 k = bch2_btree_iter_peek(&iter)).k) {
1820 c->gc_gens_pos = iter.pos;
1822 if (gc_btree_gens_key(c, k) && !commit_err) {
1823 bch2_bkey_buf_reassemble(&sk, c, k);
1824 bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1827 bch2_trans_update(&trans, &iter, sk.k, 0) ?:
1828 bch2_trans_commit(&trans, NULL, NULL,
1829 BTREE_INSERT_NOWAIT|
1830 BTREE_INSERT_NOFAIL);
1831 if (commit_err == -EINTR) {
1837 bch2_btree_iter_advance(&iter);
1839 bch2_trans_iter_exit(&trans, &iter);
1841 bch2_trans_exit(&trans);
1842 bch2_bkey_buf_exit(&sk, c);
1847 int bch2_gc_gens(struct bch_fs *c)
1850 struct bucket_array *buckets;
1856 * Ideally we would be using state_lock and not gc_lock here, but that
1857 * introduces a deadlock in the RO path - we currently take the state
1858 * lock at the start of going RO, thus the gc thread may get stuck:
1860 down_read(&c->gc_lock);
1862 for_each_member_device(ca, c, i) {
1863 down_read(&ca->bucket_lock);
1864 buckets = bucket_array(ca);
1866 for_each_bucket(g, buckets)
1867 g->gc_gen = g->mark.gen;
1868 up_read(&ca->bucket_lock);
1871 for (i = 0; i < BTREE_ID_NR; i++)
1872 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1873 c->gc_gens_btree = i;
1874 c->gc_gens_pos = POS_MIN;
1875 ret = bch2_gc_btree_gens(c, i);
1877 bch_err(c, "error recalculating oldest_gen: %i", ret);
1882 for_each_member_device(ca, c, i) {
1883 down_read(&ca->bucket_lock);
1884 buckets = bucket_array(ca);
1886 for_each_bucket(g, buckets)
1887 g->oldest_gen = g->gc_gen;
1888 up_read(&ca->bucket_lock);
1891 c->gc_gens_btree = 0;
1892 c->gc_gens_pos = POS_MIN;
1896 up_read(&c->gc_lock);
1900 static int bch2_gc_thread(void *arg)
1902 struct bch_fs *c = arg;
1903 struct io_clock *clock = &c->io_clock[WRITE];
1904 unsigned long last = atomic64_read(&clock->now);
1905 unsigned last_kick = atomic_read(&c->kick_gc);
1912 set_current_state(TASK_INTERRUPTIBLE);
1914 if (kthread_should_stop()) {
1915 __set_current_state(TASK_RUNNING);
1919 if (atomic_read(&c->kick_gc) != last_kick)
1922 if (c->btree_gc_periodic) {
1923 unsigned long next = last + c->capacity / 16;
1925 if (atomic64_read(&clock->now) >= next)
1928 bch2_io_clock_schedule_timeout(clock, next);
1935 __set_current_state(TASK_RUNNING);
1937 last = atomic64_read(&clock->now);
1938 last_kick = atomic_read(&c->kick_gc);
1941 * Full gc is currently incompatible with btree key cache:
1944 ret = bch2_gc(c, false, false);
1946 ret = bch2_gc_gens(c);
1949 bch_err(c, "btree gc failed: %i", ret);
1951 debug_check_no_locks_held();
1957 void bch2_gc_thread_stop(struct bch_fs *c)
1959 struct task_struct *p;
1962 c->gc_thread = NULL;
1970 int bch2_gc_thread_start(struct bch_fs *c)
1972 struct task_struct *p;
1977 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1979 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));