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_start(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 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
237 "btree node with incorrect min_key at btree %s level %u:\n"
240 bch2_btree_ids[b->c.btree_id], b->c.level,
242 (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&cur->key)), buf2))) {
244 bpos_cmp(expected_start, cur->data->min_key) > 0 &&
245 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
246 if (bkey_cmp(prev->data->min_key,
247 cur->data->min_key) <= 0)
248 return DROP_PREV_NODE;
250 ret = set_node_max(c, prev,
251 bpos_predecessor(cur->data->min_key));
253 if (bkey_cmp(expected_start, b->data->max_key) >= 0)
254 return DROP_THIS_NODE;
256 ret = set_node_min(c, cur, expected_start);
265 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
268 char buf1[200], buf2[200];
271 if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
272 "btree node with incorrect max_key at btree %s level %u:\n"
275 bch2_btree_ids[b->c.btree_id], b->c.level,
276 (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&child->key)), buf1),
277 (bch2_bpos_to_text(&PBUF(buf2), b->key.k.p), buf2))) {
278 ret = set_node_max(c, child, b->key.k.p);
286 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
288 struct btree_and_journal_iter iter;
290 struct bkey_buf prev_k, cur_k;
291 struct btree *prev = NULL, *cur = NULL;
292 bool have_child, dropped_children = false;
300 have_child = dropped_children = false;
301 bch2_bkey_buf_init(&prev_k);
302 bch2_bkey_buf_init(&cur_k);
303 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
305 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
306 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
307 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
309 bch2_btree_and_journal_iter_advance(&iter);
310 bch2_bkey_buf_reassemble(&cur_k, c, k);
312 cur = bch2_btree_node_get_noiter(c, cur_k.k,
313 b->c.btree_id, b->c.level - 1,
315 ret = PTR_ERR_OR_ZERO(cur);
317 if (mustfix_fsck_err_on(ret == -EIO, c,
318 "Unreadable btree node at btree %s level %u:\n"
320 bch2_btree_ids[b->c.btree_id],
322 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur_k.k)), buf))) {
323 bch2_btree_node_evict(c, cur_k.k);
324 ret = bch2_journal_key_delete(c, b->c.btree_id,
325 b->c.level, cur_k.k->k.p);
332 bch_err(c, "%s: error %i getting btree node",
337 ret = btree_repair_node_start(c, b, prev, cur);
339 if (ret == DROP_THIS_NODE) {
340 six_unlock_read(&cur->c.lock);
341 bch2_btree_node_evict(c, cur_k.k);
342 ret = bch2_journal_key_delete(c, b->c.btree_id,
343 b->c.level, cur_k.k->k.p);
350 six_unlock_read(&prev->c.lock);
353 if (ret == DROP_PREV_NODE) {
354 bch2_btree_node_evict(c, prev_k.k);
355 ret = bch2_journal_key_delete(c, b->c.btree_id,
356 b->c.level, prev_k.k->k.p);
360 bch2_btree_and_journal_iter_exit(&iter);
361 bch2_bkey_buf_exit(&prev_k, c);
362 bch2_bkey_buf_exit(&cur_k, c);
369 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
372 if (!ret && !IS_ERR_OR_NULL(prev)) {
374 ret = btree_repair_node_end(c, b, prev);
377 if (!IS_ERR_OR_NULL(prev))
378 six_unlock_read(&prev->c.lock);
380 if (!IS_ERR_OR_NULL(cur))
381 six_unlock_read(&cur->c.lock);
387 bch2_btree_and_journal_iter_exit(&iter);
388 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
390 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
391 bch2_bkey_buf_reassemble(&cur_k, c, k);
392 bch2_btree_and_journal_iter_advance(&iter);
394 cur = bch2_btree_node_get_noiter(c, cur_k.k,
395 b->c.btree_id, b->c.level - 1,
397 ret = PTR_ERR_OR_ZERO(cur);
400 bch_err(c, "%s: error %i getting btree node",
405 ret = bch2_btree_repair_topology_recurse(c, cur);
406 six_unlock_read(&cur->c.lock);
409 if (ret == DROP_THIS_NODE) {
410 bch2_btree_node_evict(c, cur_k.k);
411 ret = bch2_journal_key_delete(c, b->c.btree_id,
412 b->c.level, cur_k.k->k.p);
413 dropped_children = true;
422 if (mustfix_fsck_err_on(!have_child, c,
423 "empty interior btree node at btree %s level %u\n"
425 bch2_btree_ids[b->c.btree_id],
427 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(&b->key)), buf)))
428 ret = DROP_THIS_NODE;
431 if (!IS_ERR_OR_NULL(prev))
432 six_unlock_read(&prev->c.lock);
433 if (!IS_ERR_OR_NULL(cur))
434 six_unlock_read(&cur->c.lock);
436 bch2_btree_and_journal_iter_exit(&iter);
437 bch2_bkey_buf_exit(&prev_k, c);
438 bch2_bkey_buf_exit(&cur_k, c);
440 if (!ret && dropped_children)
446 static int bch2_repair_topology(struct bch_fs *c)
452 for (i = 0; i < BTREE_ID_NR && !ret; i++) {
453 b = c->btree_roots[i].b;
454 if (btree_node_fake(b))
457 six_lock_read(&b->c.lock, NULL, NULL);
458 ret = bch2_btree_repair_topology_recurse(c, b);
459 six_unlock_read(&b->c.lock);
461 if (ret == DROP_THIS_NODE) {
462 bch_err(c, "empty btree root - repair unimplemented");
470 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
471 unsigned level, bool is_root,
474 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
475 const union bch_extent_entry *entry;
476 struct extent_ptr_decoded p = { 0 };
477 bool do_update = false;
481 bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
482 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
483 struct bucket *g = PTR_BUCKET(ca, &p.ptr, true);
484 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr, false);
485 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
487 if (fsck_err_on(g->mark.data_type &&
488 g->mark.data_type != data_type, c,
489 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
491 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
492 bch2_data_types[g->mark.data_type],
493 bch2_data_types[data_type],
494 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
495 if (data_type == BCH_DATA_btree) {
496 g2->_mark.data_type = g->_mark.data_type = data_type;
497 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
503 if (fsck_err_on(!g->gen_valid, c,
504 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
506 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
507 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
509 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
511 g2->_mark.gen = g->_mark.gen = p.ptr.gen;
512 g2->gen_valid = g->gen_valid = true;
513 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
519 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
520 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
522 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
523 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
524 p.ptr.gen, g->mark.gen,
525 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
527 g2->_mark.gen = g->_mark.gen = p.ptr.gen;
528 g2->gen_valid = g->gen_valid = true;
529 g2->_mark.data_type = 0;
530 g2->_mark.dirty_sectors = 0;
531 g2->_mark.cached_sectors = 0;
532 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
533 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
539 if (fsck_err_on(!p.ptr.cached &&
540 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
541 "bucket %u:%zu data type %s stale dirty ptr: %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)))
550 struct stripe *m = genradix_ptr(&c->stripes[true], p.ec.idx);
552 if (fsck_err_on(!m || !m->alive, c,
553 "pointer to nonexistent stripe %llu\n"
556 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
559 if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
560 "pointer does not match stripe %llu\n"
563 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
569 struct bkey_ptrs ptrs;
570 union bch_extent_entry *entry;
571 struct bch_extent_ptr *ptr;
575 bch_err(c, "cannot update btree roots yet");
579 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
581 bch_err(c, "%s: error allocating new key", __func__);
585 bkey_reassemble(new, *k);
589 * We don't want to drop btree node pointers - if the
590 * btree node isn't there anymore, the read path will
593 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
594 bkey_for_each_ptr(ptrs, ptr) {
595 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
596 struct bucket *g = PTR_BUCKET(ca, ptr, true);
598 ptr->gen = g->mark.gen;
601 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
602 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
603 struct bucket *g = PTR_BUCKET(ca, ptr, true);
604 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
607 (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
609 gen_cmp(ptr->gen, g->mark.gen) < 0) ||
610 (g->mark.data_type &&
611 g->mark.data_type != data_type);
614 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
615 bkey_extent_entry_for_each(ptrs, entry) {
616 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
617 struct stripe *m = genradix_ptr(&c->stripes[true],
618 entry->stripe_ptr.idx);
619 union bch_extent_entry *next_ptr;
621 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
622 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
627 bch_err(c, "aieee, found stripe ptr with no data ptr");
631 if (!m || !m->alive ||
632 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
635 bch2_bkey_extent_entry_drop(new, entry);
642 ret = bch2_journal_key_insert(c, btree_id, level, new);
646 *k = bkey_i_to_s_c(new);
652 /* marking of btree keys/nodes: */
654 static int bch2_gc_mark_key(struct bch_fs *c, enum btree_id btree_id,
655 unsigned level, bool is_root,
657 u8 *max_stale, bool initial)
659 struct bkey_ptrs_c ptrs;
660 const struct bch_extent_ptr *ptr;
662 BTREE_TRIGGER_INSERT|
664 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
668 BUG_ON(bch2_journal_seq_verify &&
669 k->k->version.lo > journal_cur_seq(&c->journal));
671 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
675 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
676 "key version number higher than recorded: %llu > %llu",
678 atomic64_read(&c->key_version)))
679 atomic64_set(&c->key_version, k->k->version.lo);
681 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
682 fsck_err_on(!bch2_bkey_replicas_marked(c, *k), c,
683 "superblock not marked as containing replicas (type %u)",
685 ret = bch2_mark_bkey_replicas(c, *k);
687 bch_err(c, "error marking bkey replicas: %i", ret);
693 ptrs = bch2_bkey_ptrs_c(*k);
694 bkey_for_each_ptr(ptrs, ptr) {
695 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
696 struct bucket *g = PTR_BUCKET(ca, ptr, true);
698 if (gen_after(g->oldest_gen, ptr->gen))
699 g->oldest_gen = ptr->gen;
701 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
704 bch2_mark_key(c, *k, flags);
708 bch_err(c, "%s: ret %i", __func__, ret);
712 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
715 struct btree_node_iter iter;
716 struct bkey unpacked;
718 struct bkey_buf prev, cur;
723 if (!btree_node_type_needs_gc(btree_node_type(b)))
726 bch2_btree_node_iter_init_from_start(&iter, b);
727 bch2_bkey_buf_init(&prev);
728 bch2_bkey_buf_init(&cur);
729 bkey_init(&prev.k->k);
731 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
732 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
733 &k, max_stale, initial);
737 bch2_btree_node_iter_advance(&iter, b);
740 bch2_bkey_buf_reassemble(&cur, c, k);
742 ret = bch2_gc_check_topology(c, b, &prev, cur,
743 bch2_btree_node_iter_end(&iter));
749 bch2_bkey_buf_exit(&cur, c);
750 bch2_bkey_buf_exit(&prev, c);
754 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
755 bool initial, bool metadata_only)
757 struct btree_trans trans;
758 struct btree_iter *iter;
760 unsigned depth = metadata_only ? 1
761 : bch2_expensive_debug_checks ? 0
762 : !btree_node_type_needs_gc(btree_id) ? 1
767 bch2_trans_init(&trans, c, 0, 0);
769 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
771 __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
772 0, depth, BTREE_ITER_PREFETCH, b) {
773 bch2_verify_btree_nr_keys(b);
775 gc_pos_set(c, gc_pos_btree_node(b));
777 ret = btree_gc_mark_node(c, b, &max_stale, initial);
783 bch2_btree_node_rewrite(c, iter,
786 BTREE_INSERT_GC_LOCK_HELD);
787 else if (!bch2_btree_gc_rewrite_disabled &&
788 (bch2_btree_gc_always_rewrite || max_stale > 16))
789 bch2_btree_node_rewrite(c, iter,
792 BTREE_INSERT_GC_LOCK_HELD);
795 bch2_trans_cond_resched(&trans);
797 bch2_trans_iter_put(&trans, iter);
799 ret = bch2_trans_exit(&trans) ?: ret;
803 mutex_lock(&c->btree_root_lock);
804 b = c->btree_roots[btree_id].b;
805 if (!btree_node_fake(b)) {
806 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
808 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
809 &k, &max_stale, initial);
811 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
812 mutex_unlock(&c->btree_root_lock);
817 static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
818 unsigned target_depth)
820 struct btree_and_journal_iter iter;
822 struct bkey_buf cur, prev;
827 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
828 bch2_bkey_buf_init(&prev);
829 bch2_bkey_buf_init(&cur);
830 bkey_init(&prev.k->k);
832 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
833 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
834 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
836 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
837 &k, &max_stale, true);
839 bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
844 bch2_bkey_buf_reassemble(&cur, c, k);
845 k = bkey_i_to_s_c(cur.k);
847 bch2_btree_and_journal_iter_advance(&iter);
849 ret = bch2_gc_check_topology(c, b,
851 !bch2_btree_and_journal_iter_peek(&iter).k);
855 bch2_btree_and_journal_iter_advance(&iter);
859 if (b->c.level > target_depth) {
860 bch2_btree_and_journal_iter_exit(&iter);
861 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
863 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
866 bch2_bkey_buf_reassemble(&cur, c, k);
867 bch2_btree_and_journal_iter_advance(&iter);
869 child = bch2_btree_node_get_noiter(c, cur.k,
870 b->c.btree_id, b->c.level - 1,
872 ret = PTR_ERR_OR_ZERO(child);
875 bch2_topology_error(c);
881 "Unreadable btree node at btree %s level %u:\n"
883 bch2_btree_ids[b->c.btree_id],
885 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf)) &&
886 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
887 ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
888 bch_info(c, "Halting mark and sweep to start topology repair pass");
891 /* Continue marking when opted to not
894 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
898 bch_err(c, "%s: error %i getting btree node",
903 ret = bch2_gc_btree_init_recurse(c, child,
905 six_unlock_read(&child->c.lock);
912 bch2_bkey_buf_exit(&cur, c);
913 bch2_bkey_buf_exit(&prev, c);
914 bch2_btree_and_journal_iter_exit(&iter);
918 static int bch2_gc_btree_init(struct bch_fs *c,
919 enum btree_id btree_id,
923 unsigned target_depth = metadata_only ? 1
924 : bch2_expensive_debug_checks ? 0
925 : !btree_node_type_needs_gc(btree_id) ? 1
931 b = c->btree_roots[btree_id].b;
933 if (btree_node_fake(b))
936 six_lock_read(&b->c.lock, NULL, NULL);
937 if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
938 "btree root with incorrect min_key: %s",
939 (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
940 bch_err(c, "repair unimplemented");
945 if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, POS_MAX), c,
946 "btree root with incorrect max_key: %s",
947 (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
948 bch_err(c, "repair unimplemented");
953 if (b->c.level >= target_depth)
954 ret = bch2_gc_btree_init_recurse(c, b, target_depth);
957 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
959 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
960 &k, &max_stale, true);
963 six_unlock_read(&b->c.lock);
966 bch_err(c, "%s: ret %i", __func__, ret);
970 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
972 return (int) btree_id_to_gc_phase(l) -
973 (int) btree_id_to_gc_phase(r);
976 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
978 enum btree_id ids[BTREE_ID_NR];
982 for (i = 0; i < BTREE_ID_NR; i++)
984 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
986 for (i = 0; i < BTREE_ID_NR && !ret; i++)
988 ? bch2_gc_btree_init(c, ids[i], metadata_only)
989 : bch2_gc_btree(c, ids[i], initial, metadata_only);
992 bch_err(c, "%s: ret %i", __func__, ret);
996 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
998 enum bch_data_type type,
1001 u64 b = sector_to_bucket(ca, start);
1005 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1007 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1008 gc_phase(GC_PHASE_SB), flags);
1011 } while (start < end);
1014 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1017 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1022 * This conditional is kind of gross, but we may be called from the
1023 * device add path, before the new device has actually been added to the
1024 * running filesystem:
1027 lockdep_assert_held(&c->sb_lock);
1028 percpu_down_read(&c->mark_lock);
1031 for (i = 0; i < layout->nr_superblocks; i++) {
1032 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1034 if (offset == BCH_SB_SECTOR)
1035 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1036 BCH_DATA_sb, flags);
1038 mark_metadata_sectors(c, ca, offset,
1039 offset + (1 << layout->sb_max_size_bits),
1040 BCH_DATA_sb, flags);
1043 for (i = 0; i < ca->journal.nr; i++) {
1044 b = ca->journal.buckets[i];
1045 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1047 gc_phase(GC_PHASE_SB), flags);
1051 percpu_up_read(&c->mark_lock);
1054 static void bch2_mark_superblocks(struct bch_fs *c)
1059 mutex_lock(&c->sb_lock);
1060 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1062 for_each_online_member(ca, c, i)
1063 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1064 mutex_unlock(&c->sb_lock);
1068 /* Also see bch2_pending_btree_node_free_insert_done() */
1069 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1071 struct btree_update *as;
1072 struct pending_btree_node_free *d;
1074 mutex_lock(&c->btree_interior_update_lock);
1075 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1077 for_each_pending_btree_node_free(c, as, d)
1078 if (d->index_update_done)
1079 bch2_mark_key(c, bkey_i_to_s_c(&d->key),
1080 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_GC);
1082 mutex_unlock(&c->btree_interior_update_lock);
1086 static void bch2_gc_free(struct bch_fs *c)
1091 genradix_free(&c->stripes[1]);
1093 for_each_member_device(ca, c, i) {
1094 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1095 sizeof(struct bucket_array) +
1096 ca->mi.nbuckets * sizeof(struct bucket));
1097 ca->buckets[1] = NULL;
1099 free_percpu(ca->usage_gc);
1100 ca->usage_gc = NULL;
1103 free_percpu(c->usage_gc);
1107 static int bch2_gc_done(struct bch_fs *c,
1108 bool initial, bool metadata_only)
1110 struct bch_dev *ca = NULL;
1111 bool verify = !metadata_only && (!initial ||
1112 (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1116 #define copy_field(_f, _msg, ...) \
1117 if (dst->_f != src->_f) { \
1119 fsck_err(c, _msg ": got %llu, should be %llu" \
1120 , ##__VA_ARGS__, dst->_f, src->_f); \
1121 dst->_f = src->_f; \
1122 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
1124 #define copy_stripe_field(_f, _msg, ...) \
1125 if (dst->_f != src->_f) { \
1127 fsck_err(c, "stripe %zu has wrong "_msg \
1128 ": got %u, should be %u", \
1129 iter.pos, ##__VA_ARGS__, \
1130 dst->_f, src->_f); \
1131 dst->_f = src->_f; \
1132 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
1134 #define copy_bucket_field(_f) \
1135 if (dst->b[b].mark._f != src->b[b].mark._f) { \
1137 fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f \
1138 ": got %u, should be %u", dev, b, \
1139 dst->b[b].mark.gen, \
1140 bch2_data_types[dst->b[b].mark.data_type],\
1141 dst->b[b].mark._f, src->b[b].mark._f); \
1142 dst->b[b]._mark._f = src->b[b].mark._f; \
1143 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags); \
1145 #define copy_dev_field(_f, _msg, ...) \
1146 copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1147 #define copy_fs_field(_f, _msg, ...) \
1148 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1150 if (!metadata_only) {
1151 struct genradix_iter iter = genradix_iter_init(&c->stripes[1], 0);
1152 struct stripe *dst, *src;
1154 while ((src = genradix_iter_peek(&iter, &c->stripes[1]))) {
1155 dst = genradix_ptr_alloc(&c->stripes[0], iter.pos, GFP_KERNEL);
1157 if (dst->alive != src->alive ||
1158 dst->sectors != src->sectors ||
1159 dst->algorithm != src->algorithm ||
1160 dst->nr_blocks != src->nr_blocks ||
1161 dst->nr_redundant != src->nr_redundant) {
1162 bch_err(c, "unexpected stripe inconsistency at bch2_gc_done, confused");
1167 for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
1168 copy_stripe_field(block_sectors[i],
1169 "block_sectors[%u]", i);
1171 dst->blocks_nonempty = 0;
1172 for (i = 0; i < dst->nr_blocks; i++)
1173 dst->blocks_nonempty += dst->block_sectors[i] != 0;
1175 genradix_iter_advance(&iter, &c->stripes[1]);
1179 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1180 bch2_fs_usage_acc_to_base(c, i);
1182 for_each_member_device(ca, c, dev) {
1183 struct bucket_array *dst = __bucket_array(ca, 0);
1184 struct bucket_array *src = __bucket_array(ca, 1);
1187 for (b = 0; b < src->nbuckets; b++) {
1188 copy_bucket_field(gen);
1189 copy_bucket_field(data_type);
1190 copy_bucket_field(stripe);
1191 copy_bucket_field(dirty_sectors);
1192 copy_bucket_field(cached_sectors);
1194 dst->b[b].oldest_gen = src->b[b].oldest_gen;
1198 struct bch_dev_usage *dst = ca->usage_base;
1199 struct bch_dev_usage *src = (void *)
1200 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1203 copy_dev_field(buckets_ec, "buckets_ec");
1204 copy_dev_field(buckets_unavailable, "buckets_unavailable");
1206 for (i = 0; i < BCH_DATA_NR; i++) {
1207 copy_dev_field(d[i].buckets, "%s buckets", bch2_data_types[i]);
1208 copy_dev_field(d[i].sectors, "%s sectors", bch2_data_types[i]);
1209 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1215 unsigned nr = fs_usage_u64s(c);
1216 struct bch_fs_usage *dst = c->usage_base;
1217 struct bch_fs_usage *src = (void *)
1218 bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1220 copy_fs_field(hidden, "hidden");
1221 copy_fs_field(btree, "btree");
1223 if (!metadata_only) {
1224 copy_fs_field(data, "data");
1225 copy_fs_field(cached, "cached");
1226 copy_fs_field(reserved, "reserved");
1227 copy_fs_field(nr_inodes,"nr_inodes");
1229 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1230 copy_fs_field(persistent_reserved[i],
1231 "persistent_reserved[%i]", i);
1234 for (i = 0; i < c->replicas.nr; i++) {
1235 struct bch_replicas_entry *e =
1236 cpu_replicas_entry(&c->replicas, i);
1239 if (metadata_only &&
1240 (e->data_type == BCH_DATA_user ||
1241 e->data_type == BCH_DATA_cached))
1244 bch2_replicas_entry_to_text(&PBUF(buf), e);
1246 copy_fs_field(replicas[i], "%s", buf);
1250 #undef copy_fs_field
1251 #undef copy_dev_field
1252 #undef copy_bucket_field
1253 #undef copy_stripe_field
1257 percpu_ref_put(&ca->ref);
1259 bch_err(c, "%s: ret %i", __func__, ret);
1263 static int bch2_gc_start(struct bch_fs *c,
1266 struct bch_dev *ca = NULL;
1270 BUG_ON(c->usage_gc);
1272 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1273 sizeof(u64), GFP_KERNEL);
1275 bch_err(c, "error allocating c->usage_gc");
1279 for_each_member_device(ca, c, i) {
1280 BUG_ON(ca->buckets[1]);
1281 BUG_ON(ca->usage_gc);
1283 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
1284 ca->mi.nbuckets * sizeof(struct bucket),
1285 GFP_KERNEL|__GFP_ZERO);
1286 if (!ca->buckets[1]) {
1287 percpu_ref_put(&ca->ref);
1288 bch_err(c, "error allocating ca->buckets[gc]");
1292 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1293 if (!ca->usage_gc) {
1294 bch_err(c, "error allocating ca->usage_gc");
1295 percpu_ref_put(&ca->ref);
1300 ret = bch2_ec_mem_alloc(c, true);
1302 bch_err(c, "error allocating ec gc mem");
1306 percpu_down_write(&c->mark_lock);
1309 * indicate to stripe code that we need to allocate for the gc stripes
1312 gc_pos_set(c, gc_phase(GC_PHASE_START));
1314 for_each_member_device(ca, c, i) {
1315 struct bucket_array *dst = __bucket_array(ca, 1);
1316 struct bucket_array *src = __bucket_array(ca, 0);
1319 dst->first_bucket = src->first_bucket;
1320 dst->nbuckets = src->nbuckets;
1322 for (b = 0; b < src->nbuckets; b++) {
1323 struct bucket *d = &dst->b[b];
1324 struct bucket *s = &src->b[b];
1326 d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
1327 d->gen_valid = s->gen_valid;
1329 if (metadata_only &&
1330 (s->mark.data_type == BCH_DATA_user ||
1331 s->mark.data_type == BCH_DATA_cached))
1336 percpu_up_write(&c->mark_lock);
1341 static int bch2_gc_reflink_done_initial_fn(struct bch_fs *c, struct bkey_s_c k)
1343 struct reflink_gc *r;
1344 const __le64 *refcount = bkey_refcount_c(k);
1351 r = genradix_ptr(&c->reflink_gc_table, c->reflink_gc_idx++);
1356 r->offset != k.k->p.offset ||
1357 r->size != k.k->size) {
1358 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1362 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1363 "reflink key has wrong refcount:\n"
1366 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1370 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1376 bkey_reassemble(new, k);
1379 new->k.type = KEY_TYPE_deleted;
1382 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1385 ret = bch2_journal_key_insert(c, BTREE_ID_reflink, 0, new);
1393 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1396 struct btree_trans trans;
1397 struct btree_iter *iter;
1399 struct reflink_gc *r;
1408 c->reflink_gc_idx = 0;
1410 ret = bch2_btree_and_journal_walk(c, BTREE_ID_reflink,
1411 bch2_gc_reflink_done_initial_fn);
1415 bch2_trans_init(&trans, c, 0, 0);
1417 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1418 BTREE_ITER_PREFETCH, k, ret) {
1419 const __le64 *refcount = bkey_refcount_c(k);
1424 r = genradix_ptr(&c->reflink_gc_table, idx);
1426 r->offset != k.k->p.offset ||
1427 r->size != k.k->size) {
1428 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1433 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1434 "reflink key has wrong refcount:\n"
1437 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1441 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1447 bkey_reassemble(new, k);
1450 new->k.type = KEY_TYPE_deleted;
1452 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1454 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1455 __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1463 bch2_trans_iter_put(&trans, iter);
1464 bch2_trans_exit(&trans);
1466 genradix_free(&c->reflink_gc_table);
1467 c->reflink_gc_nr = 0;
1471 static int bch2_gc_reflink_start_initial_fn(struct bch_fs *c, struct bkey_s_c k)
1474 struct reflink_gc *r;
1475 const __le64 *refcount = bkey_refcount_c(k);
1480 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1485 r->offset = k.k->p.offset;
1486 r->size = k.k->size;
1491 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1494 struct btree_trans trans;
1495 struct btree_iter *iter;
1497 struct reflink_gc *r;
1503 genradix_free(&c->reflink_gc_table);
1504 c->reflink_gc_nr = 0;
1507 return bch2_btree_and_journal_walk(c, BTREE_ID_reflink,
1508 bch2_gc_reflink_start_initial_fn);
1510 bch2_trans_init(&trans, c, 0, 0);
1512 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1513 BTREE_ITER_PREFETCH, k, ret) {
1514 const __le64 *refcount = bkey_refcount_c(k);
1519 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1526 r->offset = k.k->p.offset;
1527 r->size = k.k->size;
1530 bch2_trans_iter_put(&trans, iter);
1532 bch2_trans_exit(&trans);
1537 * bch2_gc - walk _all_ references to buckets, and recompute them:
1539 * Order matters here:
1540 * - Concurrent GC relies on the fact that we have a total ordering for
1541 * everything that GC walks - see gc_will_visit_node(),
1542 * gc_will_visit_root()
1544 * - also, references move around in the course of index updates and
1545 * various other crap: everything needs to agree on the ordering
1546 * references are allowed to move around in - e.g., we're allowed to
1547 * start with a reference owned by an open_bucket (the allocator) and
1548 * move it to the btree, but not the reverse.
1550 * This is necessary to ensure that gc doesn't miss references that
1551 * move around - if references move backwards in the ordering GC
1552 * uses, GC could skip past them
1554 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1557 u64 start_time = local_clock();
1558 unsigned i, iter = 0;
1561 lockdep_assert_held(&c->state_lock);
1564 down_write(&c->gc_lock);
1566 /* flush interior btree updates: */
1567 closure_wait_event(&c->btree_interior_update_wait,
1568 !bch2_btree_interior_updates_nr_pending(c));
1570 ret = bch2_gc_start(c, metadata_only) ?:
1571 bch2_gc_reflink_start(c, initial, metadata_only);
1575 bch2_mark_superblocks(c);
1577 if (test_bit(BCH_FS_TOPOLOGY_ERROR, &c->flags) &&
1578 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1579 c->opts.fix_errors != FSCK_OPT_NO) {
1580 bch_info(c, "starting topology repair pass");
1581 ret = bch2_repair_topology(c);
1584 bch_info(c, "topology repair pass done");
1586 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1589 ret = bch2_gc_btrees(c, initial, metadata_only);
1591 if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1592 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1593 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1594 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1598 if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1599 ret = FSCK_ERR_EXIT;
1605 bch2_mark_pending_btree_node_frees(c);
1609 if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1610 (!iter && bch2_test_restart_gc)) {
1612 * XXX: make sure gens we fixed got saved
1615 bch_info(c, "Second GC pass needed, restarting:");
1616 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1617 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1619 percpu_down_write(&c->mark_lock);
1621 percpu_up_write(&c->mark_lock);
1622 /* flush fsck errors, reset counters */
1623 bch2_flush_fsck_errs(c);
1628 bch_info(c, "Unable to fix bucket gens, looping");
1633 bch2_journal_block(&c->journal);
1635 percpu_down_write(&c->mark_lock);
1636 ret = bch2_gc_reflink_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, false);
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, false);
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 iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
1714 BTREE_ITER_PREFETCH|
1715 BTREE_ITER_NOT_EXTENTS|
1716 BTREE_ITER_ALL_SNAPSHOTS);
1718 while ((k = bch2_btree_iter_peek(iter)).k &&
1719 !(ret = bkey_err(k))) {
1720 c->gc_gens_pos = iter->pos;
1722 if (gc_btree_gens_key(c, k) && !commit_err) {
1723 bch2_bkey_buf_reassemble(&sk, c, k);
1724 bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1728 bch2_trans_update(&trans, iter, sk.k, 0) ?:
1729 bch2_trans_commit(&trans, NULL, NULL,
1730 BTREE_INSERT_NOWAIT|
1731 BTREE_INSERT_NOFAIL);
1732 if (commit_err == -EINTR) {
1738 bch2_btree_iter_advance(iter);
1740 bch2_trans_iter_put(&trans, iter);
1742 bch2_trans_exit(&trans);
1743 bch2_bkey_buf_exit(&sk, c);
1748 int bch2_gc_gens(struct bch_fs *c)
1751 struct bucket_array *buckets;
1757 * Ideally we would be using state_lock and not gc_lock here, but that
1758 * introduces a deadlock in the RO path - we currently take the state
1759 * lock at the start of going RO, thus the gc thread may get stuck:
1761 down_read(&c->gc_lock);
1763 for_each_member_device(ca, c, i) {
1764 down_read(&ca->bucket_lock);
1765 buckets = bucket_array(ca);
1767 for_each_bucket(g, buckets)
1768 g->gc_gen = g->mark.gen;
1769 up_read(&ca->bucket_lock);
1772 for (i = 0; i < BTREE_ID_NR; i++)
1773 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1774 c->gc_gens_btree = i;
1775 c->gc_gens_pos = POS_MIN;
1776 ret = bch2_gc_btree_gens(c, i);
1778 bch_err(c, "error recalculating oldest_gen: %i", ret);
1783 for_each_member_device(ca, c, i) {
1784 down_read(&ca->bucket_lock);
1785 buckets = bucket_array(ca);
1787 for_each_bucket(g, buckets)
1788 g->oldest_gen = g->gc_gen;
1789 up_read(&ca->bucket_lock);
1792 c->gc_gens_btree = 0;
1793 c->gc_gens_pos = POS_MIN;
1797 up_read(&c->gc_lock);
1801 static int bch2_gc_thread(void *arg)
1803 struct bch_fs *c = arg;
1804 struct io_clock *clock = &c->io_clock[WRITE];
1805 unsigned long last = atomic64_read(&clock->now);
1806 unsigned last_kick = atomic_read(&c->kick_gc);
1813 set_current_state(TASK_INTERRUPTIBLE);
1815 if (kthread_should_stop()) {
1816 __set_current_state(TASK_RUNNING);
1820 if (atomic_read(&c->kick_gc) != last_kick)
1823 if (c->btree_gc_periodic) {
1824 unsigned long next = last + c->capacity / 16;
1826 if (atomic64_read(&clock->now) >= next)
1829 bch2_io_clock_schedule_timeout(clock, next);
1836 __set_current_state(TASK_RUNNING);
1838 last = atomic64_read(&clock->now);
1839 last_kick = atomic_read(&c->kick_gc);
1842 * Full gc is currently incompatible with btree key cache:
1845 ret = bch2_gc(c, false, false);
1847 ret = bch2_gc_gens(c);
1850 bch_err(c, "btree gc failed: %i", ret);
1852 debug_check_no_locks_held();
1858 void bch2_gc_thread_stop(struct bch_fs *c)
1860 struct task_struct *p;
1863 c->gc_thread = NULL;
1871 int bch2_gc_thread_start(struct bch_fs *c)
1873 struct task_struct *p;
1878 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1880 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));