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_journal_iter.h"
13 #include "btree_key_cache.h"
14 #include "btree_locking.h"
15 #include "btree_update_interior.h"
33 #include <linux/slab.h>
34 #include <linux/bitops.h>
35 #include <linux/freezer.h>
36 #include <linux/kthread.h>
37 #include <linux/preempt.h>
38 #include <linux/rcupdate.h>
39 #include <linux/sched/task.h>
41 #define DROP_THIS_NODE 10
42 #define DROP_PREV_NODE 11
44 static bool should_restart_for_topology_repair(struct bch_fs *c)
46 return c->opts.fix_errors != FSCK_FIX_no &&
47 !(c->recovery_passes_complete & BIT_ULL(BCH_RECOVERY_PASS_check_topology));
50 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
53 write_seqcount_begin(&c->gc_pos_lock);
55 write_seqcount_end(&c->gc_pos_lock);
59 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
61 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
62 __gc_pos_set(c, new_pos);
66 * Missing: if an interior btree node is empty, we need to do something -
67 * perhaps just kill it
69 static int bch2_gc_check_topology(struct bch_fs *c,
71 struct bkey_buf *prev,
75 struct bpos node_start = b->data->min_key;
76 struct bpos node_end = b->data->max_key;
77 struct bpos expected_start = bkey_deleted(&prev->k->k)
79 : bpos_successor(prev->k->k.p);
80 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
83 if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
84 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
86 if (!bpos_eq(expected_start, bp->v.min_key)) {
87 bch2_topology_error(c);
89 if (bkey_deleted(&prev->k->k)) {
90 prt_printf(&buf1, "start of node: ");
91 bch2_bpos_to_text(&buf1, node_start);
93 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
95 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
101 btree_node_topology_bad_min_key,
102 "btree node with incorrect min_key at btree %s level %u:\n"
105 bch2_btree_id_str(b->c.btree_id), b->c.level,
106 buf1.buf, buf2.buf) && should_restart_for_topology_repair(c)) {
107 bch_info(c, "Halting mark and sweep to start topology repair pass");
108 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
111 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
116 if (is_last && !bpos_eq(cur.k->k.p, node_end)) {
117 bch2_topology_error(c);
119 printbuf_reset(&buf1);
120 printbuf_reset(&buf2);
122 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
123 bch2_bpos_to_text(&buf2, node_end);
125 if (__fsck_err(c, FSCK_CAN_FIX|FSCK_CAN_IGNORE|FSCK_NO_RATELIMIT,
126 btree_node_topology_bad_max_key,
127 "btree node with incorrect max_key at btree %s level %u:\n"
130 bch2_btree_id_str(b->c.btree_id), b->c.level,
131 buf1.buf, buf2.buf) &&
132 should_restart_for_topology_repair(c)) {
133 bch_info(c, "Halting mark and sweep to start topology repair pass");
134 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
137 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
141 bch2_bkey_buf_copy(prev, c, cur.k);
144 printbuf_exit(&buf2);
145 printbuf_exit(&buf1);
149 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
151 switch (b->key.k.type) {
152 case KEY_TYPE_btree_ptr: {
153 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
157 dst->v.seq = b->data->keys.seq;
158 dst->v.sectors_written = 0;
160 dst->v.min_key = b->data->min_key;
161 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
162 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
165 case KEY_TYPE_btree_ptr_v2:
166 bkey_copy(&dst->k_i, &b->key);
173 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
174 enum btree_id btree, unsigned level,
175 struct bkey_s_c old, struct bkey_i *new)
177 struct bch_fs *c = trans->c;
182 bch2_bkey_buf_init(&tmp);
183 bch2_bkey_buf_reassemble(&tmp, c, old);
185 b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
186 if (!IS_ERR_OR_NULL(b)) {
187 mutex_lock(&c->btree_cache.lock);
189 bch2_btree_node_hash_remove(&c->btree_cache, b);
191 bkey_copy(&b->key, new);
192 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
195 mutex_unlock(&c->btree_cache.lock);
196 six_unlock_read(&b->c.lock);
199 bch2_bkey_buf_exit(&tmp, c);
202 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
204 struct bkey_i_btree_ptr_v2 *new;
207 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
209 return -BCH_ERR_ENOMEM_gc_repair_key;
211 btree_ptr_to_v2(b, new);
212 b->data->min_key = new_min;
213 new->v.min_key = new_min;
214 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
216 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
222 bch2_btree_node_drop_keys_outside_node(b);
223 bkey_copy(&b->key, &new->k_i);
227 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
229 struct bkey_i_btree_ptr_v2 *new;
232 ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
236 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
238 return -BCH_ERR_ENOMEM_gc_repair_key;
240 btree_ptr_to_v2(b, new);
241 b->data->max_key = new_max;
243 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
245 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
251 bch2_btree_node_drop_keys_outside_node(b);
253 mutex_lock(&c->btree_cache.lock);
254 bch2_btree_node_hash_remove(&c->btree_cache, b);
256 bkey_copy(&b->key, &new->k_i);
257 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
259 mutex_unlock(&c->btree_cache.lock);
263 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
264 struct btree *prev, struct btree *cur)
266 struct bpos expected_start = !prev
268 : bpos_successor(prev->key.k.p);
269 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
273 prt_printf(&buf1, "start of node: ");
274 bch2_bpos_to_text(&buf1, b->data->min_key);
276 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
279 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
282 bpos_gt(expected_start, cur->data->min_key) &&
283 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
284 /* cur overwrites prev: */
286 if (mustfix_fsck_err_on(bpos_ge(prev->data->min_key,
287 cur->data->min_key), c,
288 btree_node_topology_overwritten_by_next_node,
289 "btree node overwritten by next node at btree %s level %u:\n"
292 bch2_btree_id_str(b->c.btree_id), b->c.level,
293 buf1.buf, buf2.buf)) {
294 ret = DROP_PREV_NODE;
298 if (mustfix_fsck_err_on(!bpos_eq(prev->key.k.p,
299 bpos_predecessor(cur->data->min_key)), c,
300 btree_node_topology_bad_max_key,
301 "btree node with incorrect max_key at btree %s level %u:\n"
304 bch2_btree_id_str(b->c.btree_id), b->c.level,
306 ret = set_node_max(c, prev,
307 bpos_predecessor(cur->data->min_key));
309 /* prev overwrites cur: */
311 if (mustfix_fsck_err_on(bpos_ge(expected_start,
312 cur->data->max_key), c,
313 btree_node_topology_overwritten_by_prev_node,
314 "btree node overwritten by prev node at btree %s level %u:\n"
317 bch2_btree_id_str(b->c.btree_id), b->c.level,
318 buf1.buf, buf2.buf)) {
319 ret = DROP_THIS_NODE;
323 if (mustfix_fsck_err_on(!bpos_eq(expected_start, cur->data->min_key), c,
324 btree_node_topology_bad_min_key,
325 "btree node with incorrect min_key at btree %s level %u:\n"
328 bch2_btree_id_str(b->c.btree_id), b->c.level,
330 ret = set_node_min(c, cur, expected_start);
334 printbuf_exit(&buf2);
335 printbuf_exit(&buf1);
339 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
342 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
345 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
346 bch2_bpos_to_text(&buf2, b->key.k.p);
348 if (mustfix_fsck_err_on(!bpos_eq(child->key.k.p, b->key.k.p), c,
349 btree_node_topology_bad_max_key,
350 "btree node with incorrect max_key at btree %s level %u:\n"
353 bch2_btree_id_str(b->c.btree_id), b->c.level,
354 buf1.buf, buf2.buf)) {
355 ret = set_node_max(c, child, b->key.k.p);
361 printbuf_exit(&buf2);
362 printbuf_exit(&buf1);
366 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b)
368 struct bch_fs *c = trans->c;
369 struct btree_and_journal_iter iter;
371 struct bkey_buf prev_k, cur_k;
372 struct btree *prev = NULL, *cur = NULL;
373 bool have_child, dropped_children = false;
374 struct printbuf buf = PRINTBUF;
381 have_child = dropped_children = false;
382 bch2_bkey_buf_init(&prev_k);
383 bch2_bkey_buf_init(&cur_k);
384 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
386 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
387 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
388 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
390 bch2_btree_and_journal_iter_advance(&iter);
391 bch2_bkey_buf_reassemble(&cur_k, c, k);
393 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
394 b->c.btree_id, b->c.level - 1,
396 ret = PTR_ERR_OR_ZERO(cur);
398 printbuf_reset(&buf);
399 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
401 if (mustfix_fsck_err_on(ret == -EIO, c,
402 btree_node_unreadable,
403 "Topology repair: unreadable btree node at btree %s level %u:\n"
405 bch2_btree_id_str(b->c.btree_id),
408 bch2_btree_node_evict(trans, cur_k.k);
409 ret = bch2_journal_key_delete(c, b->c.btree_id,
410 b->c.level, cur_k.k->k.p);
418 bch_err_msg(c, ret, "getting btree node");
422 ret = btree_repair_node_boundaries(c, b, prev, cur);
424 if (ret == DROP_THIS_NODE) {
425 six_unlock_read(&cur->c.lock);
426 bch2_btree_node_evict(trans, cur_k.k);
427 ret = bch2_journal_key_delete(c, b->c.btree_id,
428 b->c.level, cur_k.k->k.p);
436 six_unlock_read(&prev->c.lock);
439 if (ret == DROP_PREV_NODE) {
440 bch2_btree_node_evict(trans, prev_k.k);
441 ret = bch2_journal_key_delete(c, b->c.btree_id,
442 b->c.level, prev_k.k->k.p);
446 bch2_btree_and_journal_iter_exit(&iter);
447 bch2_bkey_buf_exit(&prev_k, c);
448 bch2_bkey_buf_exit(&cur_k, c);
455 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
458 if (!ret && !IS_ERR_OR_NULL(prev)) {
460 ret = btree_repair_node_end(c, b, prev);
463 if (!IS_ERR_OR_NULL(prev))
464 six_unlock_read(&prev->c.lock);
466 if (!IS_ERR_OR_NULL(cur))
467 six_unlock_read(&cur->c.lock);
473 bch2_btree_and_journal_iter_exit(&iter);
474 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
476 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
477 bch2_bkey_buf_reassemble(&cur_k, c, k);
478 bch2_btree_and_journal_iter_advance(&iter);
480 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
481 b->c.btree_id, b->c.level - 1,
483 ret = PTR_ERR_OR_ZERO(cur);
486 bch_err_msg(c, ret, "getting btree node");
490 ret = bch2_btree_repair_topology_recurse(trans, cur);
491 six_unlock_read(&cur->c.lock);
494 if (ret == DROP_THIS_NODE) {
495 bch2_btree_node_evict(trans, cur_k.k);
496 ret = bch2_journal_key_delete(c, b->c.btree_id,
497 b->c.level, cur_k.k->k.p);
498 dropped_children = true;
507 printbuf_reset(&buf);
508 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
510 if (mustfix_fsck_err_on(!have_child, c,
511 btree_node_topology_interior_node_empty,
512 "empty interior btree node at btree %s level %u\n"
514 bch2_btree_id_str(b->c.btree_id),
515 b->c.level, buf.buf))
516 ret = DROP_THIS_NODE;
519 if (!IS_ERR_OR_NULL(prev))
520 six_unlock_read(&prev->c.lock);
521 if (!IS_ERR_OR_NULL(cur))
522 six_unlock_read(&cur->c.lock);
524 bch2_btree_and_journal_iter_exit(&iter);
525 bch2_bkey_buf_exit(&prev_k, c);
526 bch2_bkey_buf_exit(&cur_k, c);
528 if (!ret && dropped_children)
535 int bch2_check_topology(struct bch_fs *c)
537 struct btree_trans *trans = bch2_trans_get(c);
542 for (i = 0; i < btree_id_nr_alive(c) && !ret; i++) {
543 struct btree_root *r = bch2_btree_id_root(c, i);
549 if (btree_node_fake(b))
552 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
553 ret = bch2_btree_repair_topology_recurse(trans, b);
554 six_unlock_read(&b->c.lock);
556 if (ret == DROP_THIS_NODE) {
557 bch_err(c, "empty btree root - repair unimplemented");
558 ret = -BCH_ERR_fsck_repair_unimplemented;
562 bch2_trans_put(trans);
567 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
568 unsigned level, bool is_root,
571 struct bch_fs *c = trans->c;
572 struct bkey_ptrs_c ptrs_c = bch2_bkey_ptrs_c(*k);
573 const union bch_extent_entry *entry_c;
574 struct extent_ptr_decoded p = { 0 };
575 bool do_update = false;
576 struct printbuf buf = PRINTBUF;
581 * use check_bucket_ref here
583 bkey_for_each_ptr_decode(k->k, ptrs_c, p, entry_c) {
584 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
585 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
586 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry_c->ptr);
589 (c->opts.reconstruct_alloc ||
590 fsck_err(c, ptr_to_missing_alloc_key,
591 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
593 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
594 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
596 (printbuf_reset(&buf),
597 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
606 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
607 (c->opts.reconstruct_alloc ||
608 fsck_err(c, ptr_gen_newer_than_bucket_gen,
609 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
611 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
612 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
614 (printbuf_reset(&buf),
615 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
620 g->dirty_sectors = 0;
621 g->cached_sectors = 0;
622 set_bit(BCH_FS_need_another_gc, &c->flags);
628 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
629 (c->opts.reconstruct_alloc ||
630 fsck_err(c, ptr_gen_newer_than_bucket_gen,
631 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
633 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
634 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
636 (printbuf_reset(&buf),
637 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
640 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
641 (c->opts.reconstruct_alloc ||
642 fsck_err(c, stale_dirty_ptr,
643 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
645 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
646 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
648 (printbuf_reset(&buf),
649 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
652 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
655 if (fsck_err_on(bucket_data_type(g->data_type) &&
656 bucket_data_type(g->data_type) != data_type, c,
657 ptr_bucket_data_type_mismatch,
658 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
660 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
661 bch2_data_types[g->data_type],
662 bch2_data_types[data_type],
663 (printbuf_reset(&buf),
664 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
665 if (data_type == BCH_DATA_btree) {
666 g->data_type = data_type;
667 set_bit(BCH_FS_need_another_gc, &c->flags);
674 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
676 if (fsck_err_on(!m || !m->alive, c,
677 ptr_to_missing_stripe,
678 "pointer to nonexistent stripe %llu\n"
681 (printbuf_reset(&buf),
682 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
685 if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
686 ptr_to_incorrect_stripe,
687 "pointer does not match stripe %llu\n"
690 (printbuf_reset(&buf),
691 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
697 struct bkey_ptrs ptrs;
698 union bch_extent_entry *entry;
699 struct bch_extent_ptr *ptr;
703 bch_err(c, "cannot update btree roots yet");
708 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
710 bch_err_msg(c, ret, "allocating new key");
711 ret = -BCH_ERR_ENOMEM_gc_repair_key;
715 bkey_reassemble(new, *k);
719 * We don't want to drop btree node pointers - if the
720 * btree node isn't there anymore, the read path will
723 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
724 bkey_for_each_ptr(ptrs, ptr) {
725 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
726 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
731 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
732 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
733 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
734 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
737 (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
739 gen_cmp(ptr->gen, g->gen) < 0) ||
740 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
742 g->data_type != data_type);
745 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
746 bkey_extent_entry_for_each(ptrs, entry) {
747 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
748 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
749 entry->stripe_ptr.idx);
750 union bch_extent_entry *next_ptr;
752 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
753 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
758 bch_err(c, "aieee, found stripe ptr with no data ptr");
762 if (!m || !m->alive ||
763 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
766 bch2_bkey_extent_entry_drop(new, entry);
773 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
780 bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
783 printbuf_reset(&buf);
784 bch2_bkey_val_to_text(&buf, c, *k);
785 bch_info(c, "updated %s", buf.buf);
787 printbuf_reset(&buf);
788 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
789 bch_info(c, "new key %s", buf.buf);
792 *k = bkey_i_to_s_c(new);
800 /* marking of btree keys/nodes: */
802 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
803 unsigned level, bool is_root,
807 struct bch_fs *c = trans->c;
808 struct bkey deleted = KEY(0, 0, 0);
809 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
812 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
818 BUG_ON(bch2_journal_seq_verify &&
819 k->k->version.lo > atomic64_read(&c->journal.seq));
821 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
825 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
826 bkey_version_in_future,
827 "key version number higher than recorded: %llu > %llu",
829 atomic64_read(&c->key_version)))
830 atomic64_set(&c->key_version, k->k->version.lo);
833 ret = commit_do(trans, NULL, NULL, 0,
834 bch2_mark_key(trans, btree_id, level, old, *k, flags));
842 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
844 struct bch_fs *c = trans->c;
845 struct btree_node_iter iter;
846 struct bkey unpacked;
848 struct bkey_buf prev, cur;
851 if (!btree_node_type_needs_gc(btree_node_type(b)))
854 bch2_btree_node_iter_init_from_start(&iter, b);
855 bch2_bkey_buf_init(&prev);
856 bch2_bkey_buf_init(&cur);
857 bkey_init(&prev.k->k);
859 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
860 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
865 bch2_btree_node_iter_advance(&iter, b);
868 bch2_bkey_buf_reassemble(&cur, c, k);
870 ret = bch2_gc_check_topology(c, b, &prev, cur,
871 bch2_btree_node_iter_end(&iter));
877 bch2_bkey_buf_exit(&cur, c);
878 bch2_bkey_buf_exit(&prev, c);
882 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
883 bool initial, bool metadata_only)
885 struct bch_fs *c = trans->c;
886 struct btree_iter iter;
888 unsigned depth = metadata_only ? 1 : 0;
891 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
893 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
894 0, depth, BTREE_ITER_PREFETCH, b, ret) {
895 bch2_verify_btree_nr_keys(b);
897 gc_pos_set(c, gc_pos_btree_node(b));
899 ret = btree_gc_mark_node(trans, b, initial);
903 bch2_trans_iter_exit(trans, &iter);
908 mutex_lock(&c->btree_root_lock);
909 b = bch2_btree_id_root(c, btree_id)->b;
910 if (!btree_node_fake(b)) {
911 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
913 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
916 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
917 mutex_unlock(&c->btree_root_lock);
922 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
923 unsigned target_depth)
925 struct bch_fs *c = trans->c;
926 struct btree_and_journal_iter iter;
928 struct bkey_buf cur, prev;
929 struct printbuf buf = PRINTBUF;
932 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
933 bch2_bkey_buf_init(&prev);
934 bch2_bkey_buf_init(&cur);
935 bkey_init(&prev.k->k);
937 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
938 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
939 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
941 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
947 bch2_bkey_buf_reassemble(&cur, c, k);
948 k = bkey_i_to_s_c(cur.k);
950 bch2_btree_and_journal_iter_advance(&iter);
952 ret = bch2_gc_check_topology(c, b,
954 !bch2_btree_and_journal_iter_peek(&iter).k);
958 bch2_btree_and_journal_iter_advance(&iter);
962 if (b->c.level > target_depth) {
963 bch2_btree_and_journal_iter_exit(&iter);
964 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
966 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
969 bch2_bkey_buf_reassemble(&cur, c, k);
970 bch2_btree_and_journal_iter_advance(&iter);
972 child = bch2_btree_node_get_noiter(trans, cur.k,
973 b->c.btree_id, b->c.level - 1,
975 ret = PTR_ERR_OR_ZERO(child);
978 bch2_topology_error(c);
984 btree_node_read_error,
985 "Unreadable btree node at btree %s level %u:\n"
987 bch2_btree_id_str(b->c.btree_id),
989 (printbuf_reset(&buf),
990 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
991 should_restart_for_topology_repair(c)) {
992 bch_info(c, "Halting mark and sweep to start topology repair pass");
993 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
996 /* Continue marking when opted to not
999 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
1003 bch_err_msg(c, ret, "getting btree node");
1007 ret = bch2_gc_btree_init_recurse(trans, child,
1009 six_unlock_read(&child->c.lock);
1016 bch2_bkey_buf_exit(&cur, c);
1017 bch2_bkey_buf_exit(&prev, c);
1018 bch2_btree_and_journal_iter_exit(&iter);
1019 printbuf_exit(&buf);
1023 static int bch2_gc_btree_init(struct btree_trans *trans,
1024 enum btree_id btree_id,
1027 struct bch_fs *c = trans->c;
1029 unsigned target_depth = metadata_only ? 1 : 0;
1030 struct printbuf buf = PRINTBUF;
1033 b = bch2_btree_id_root(c, btree_id)->b;
1035 if (btree_node_fake(b))
1038 six_lock_read(&b->c.lock, NULL, NULL);
1039 printbuf_reset(&buf);
1040 bch2_bpos_to_text(&buf, b->data->min_key);
1041 if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1042 btree_root_bad_min_key,
1043 "btree root with incorrect min_key: %s", buf.buf)) {
1044 bch_err(c, "repair unimplemented");
1045 ret = -BCH_ERR_fsck_repair_unimplemented;
1049 printbuf_reset(&buf);
1050 bch2_bpos_to_text(&buf, b->data->max_key);
1051 if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1052 btree_root_bad_max_key,
1053 "btree root with incorrect max_key: %s", buf.buf)) {
1054 bch_err(c, "repair unimplemented");
1055 ret = -BCH_ERR_fsck_repair_unimplemented;
1059 if (b->c.level >= target_depth)
1060 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1063 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1065 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1069 six_unlock_read(&b->c.lock);
1073 printbuf_exit(&buf);
1077 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1079 return (int) btree_id_to_gc_phase(l) -
1080 (int) btree_id_to_gc_phase(r);
1083 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1085 struct btree_trans *trans = bch2_trans_get(c);
1086 enum btree_id ids[BTREE_ID_NR];
1090 for (i = 0; i < BTREE_ID_NR; i++)
1092 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1094 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1096 ? bch2_gc_btree_init(trans, ids[i], metadata_only)
1097 : bch2_gc_btree(trans, ids[i], initial, metadata_only);
1099 for (i = BTREE_ID_NR; i < btree_id_nr_alive(c) && !ret; i++) {
1100 if (!bch2_btree_id_root(c, i)->alive)
1104 ? bch2_gc_btree_init(trans, i, metadata_only)
1105 : bch2_gc_btree(trans, i, initial, metadata_only);
1111 bch2_trans_put(trans);
1115 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1117 enum bch_data_type type,
1120 u64 b = sector_to_bucket(ca, start);
1124 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1126 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1127 gc_phase(GC_PHASE_SB), flags);
1130 } while (start < end);
1133 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1136 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1140 for (i = 0; i < layout->nr_superblocks; i++) {
1141 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1143 if (offset == BCH_SB_SECTOR)
1144 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1145 BCH_DATA_sb, flags);
1147 mark_metadata_sectors(c, ca, offset,
1148 offset + (1 << layout->sb_max_size_bits),
1149 BCH_DATA_sb, flags);
1152 for (i = 0; i < ca->journal.nr; i++) {
1153 b = ca->journal.buckets[i];
1154 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1156 gc_phase(GC_PHASE_SB), flags);
1160 static void bch2_mark_superblocks(struct bch_fs *c)
1165 mutex_lock(&c->sb_lock);
1166 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1168 for_each_online_member(ca, c, i)
1169 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1170 mutex_unlock(&c->sb_lock);
1174 /* Also see bch2_pending_btree_node_free_insert_done() */
1175 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1177 struct btree_update *as;
1178 struct pending_btree_node_free *d;
1180 mutex_lock(&c->btree_interior_update_lock);
1181 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1183 for_each_pending_btree_node_free(c, as, d)
1184 if (d->index_update_done)
1185 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1187 mutex_unlock(&c->btree_interior_update_lock);
1191 static void bch2_gc_free(struct bch_fs *c)
1196 genradix_free(&c->reflink_gc_table);
1197 genradix_free(&c->gc_stripes);
1199 for_each_member_device(ca, c, i) {
1200 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1201 sizeof(struct bucket_array) +
1202 ca->mi.nbuckets * sizeof(struct bucket));
1203 ca->buckets_gc = NULL;
1205 free_percpu(ca->usage_gc);
1206 ca->usage_gc = NULL;
1209 free_percpu(c->usage_gc);
1213 static int bch2_gc_done(struct bch_fs *c,
1214 bool initial, bool metadata_only)
1216 struct bch_dev *ca = NULL;
1217 struct printbuf buf = PRINTBUF;
1218 bool verify = !metadata_only &&
1219 !c->opts.reconstruct_alloc &&
1220 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1224 percpu_down_write(&c->mark_lock);
1226 #define copy_field(_err, _f, _msg, ...) \
1227 if (dst->_f != src->_f && \
1229 fsck_err(c, _err, _msg ": got %llu, should be %llu" \
1230 , ##__VA_ARGS__, dst->_f, src->_f))) \
1232 #define copy_dev_field(_err, _f, _msg, ...) \
1233 copy_field(_err, _f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1234 #define copy_fs_field(_err, _f, _msg, ...) \
1235 copy_field(_err, _f, "fs has wrong " _msg, ##__VA_ARGS__)
1237 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1238 bch2_fs_usage_acc_to_base(c, i);
1240 for_each_member_device(ca, c, dev) {
1241 struct bch_dev_usage *dst = ca->usage_base;
1242 struct bch_dev_usage *src = (void *)
1243 bch2_acc_percpu_u64s((u64 __percpu *) ca->usage_gc,
1246 for (i = 0; i < BCH_DATA_NR; i++) {
1247 copy_dev_field(dev_usage_buckets_wrong,
1248 d[i].buckets, "%s buckets", bch2_data_types[i]);
1249 copy_dev_field(dev_usage_sectors_wrong,
1250 d[i].sectors, "%s sectors", bch2_data_types[i]);
1251 copy_dev_field(dev_usage_fragmented_wrong,
1252 d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1257 unsigned nr = fs_usage_u64s(c);
1258 struct bch_fs_usage *dst = c->usage_base;
1259 struct bch_fs_usage *src = (void *)
1260 bch2_acc_percpu_u64s((u64 __percpu *) c->usage_gc, nr);
1262 copy_fs_field(fs_usage_hidden_wrong,
1264 copy_fs_field(fs_usage_btree_wrong,
1267 if (!metadata_only) {
1268 copy_fs_field(fs_usage_data_wrong,
1270 copy_fs_field(fs_usage_cached_wrong,
1272 copy_fs_field(fs_usage_reserved_wrong,
1273 reserved, "reserved");
1274 copy_fs_field(fs_usage_nr_inodes_wrong,
1275 nr_inodes,"nr_inodes");
1277 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1278 copy_fs_field(fs_usage_persistent_reserved_wrong,
1279 persistent_reserved[i],
1280 "persistent_reserved[%i]", i);
1283 for (i = 0; i < c->replicas.nr; i++) {
1284 struct bch_replicas_entry_v1 *e =
1285 cpu_replicas_entry(&c->replicas, i);
1287 if (metadata_only &&
1288 (e->data_type == BCH_DATA_user ||
1289 e->data_type == BCH_DATA_cached))
1292 printbuf_reset(&buf);
1293 bch2_replicas_entry_to_text(&buf, e);
1295 copy_fs_field(fs_usage_replicas_wrong,
1296 replicas[i], "%s", buf.buf);
1300 #undef copy_fs_field
1301 #undef copy_dev_field
1302 #undef copy_stripe_field
1306 percpu_ref_put(&ca->ref);
1310 percpu_up_write(&c->mark_lock);
1311 printbuf_exit(&buf);
1315 static int bch2_gc_start(struct bch_fs *c)
1317 struct bch_dev *ca = NULL;
1320 BUG_ON(c->usage_gc);
1322 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1323 sizeof(u64), GFP_KERNEL);
1325 bch_err(c, "error allocating c->usage_gc");
1326 return -BCH_ERR_ENOMEM_gc_start;
1329 for_each_member_device(ca, c, i) {
1330 BUG_ON(ca->usage_gc);
1332 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1333 if (!ca->usage_gc) {
1334 bch_err(c, "error allocating ca->usage_gc");
1335 percpu_ref_put(&ca->ref);
1336 return -BCH_ERR_ENOMEM_gc_start;
1339 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1340 ca->mi.nbuckets - ca->mi.first_bucket);
1346 static int bch2_gc_reset(struct bch_fs *c)
1351 for_each_member_device(ca, c, i) {
1352 free_percpu(ca->usage_gc);
1353 ca->usage_gc = NULL;
1356 free_percpu(c->usage_gc);
1359 return bch2_gc_start(c);
1362 /* returns true if not equal */
1363 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1364 struct bch_alloc_v4 r)
1366 return l.gen != r.gen ||
1367 l.oldest_gen != r.oldest_gen ||
1368 l.data_type != r.data_type ||
1369 l.dirty_sectors != r.dirty_sectors ||
1370 l.cached_sectors != r.cached_sectors ||
1371 l.stripe_redundancy != r.stripe_redundancy ||
1372 l.stripe != r.stripe;
1375 static int bch2_alloc_write_key(struct btree_trans *trans,
1376 struct btree_iter *iter,
1380 struct bch_fs *c = trans->c;
1381 struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1382 struct bucket gc, *b;
1383 struct bkey_i_alloc_v4 *a;
1384 struct bch_alloc_v4 old_convert, new;
1385 const struct bch_alloc_v4 *old;
1386 enum bch_data_type type;
1389 if (bkey_ge(iter->pos, POS(ca->dev_idx, ca->mi.nbuckets)))
1392 old = bch2_alloc_to_v4(k, &old_convert);
1395 percpu_down_read(&c->mark_lock);
1396 b = gc_bucket(ca, iter->pos.offset);
1399 * b->data_type doesn't yet include need_discard & need_gc_gen states -
1402 type = __alloc_data_type(b->dirty_sectors,
1407 if (b->data_type != type) {
1408 struct bch_dev_usage *u;
1411 u = this_cpu_ptr(ca->usage_gc);
1412 u->d[b->data_type].buckets--;
1413 b->data_type = type;
1414 u->d[b->data_type].buckets++;
1419 percpu_up_read(&c->mark_lock);
1421 if (metadata_only &&
1422 gc.data_type != BCH_DATA_sb &&
1423 gc.data_type != BCH_DATA_journal &&
1424 gc.data_type != BCH_DATA_btree)
1427 if (gen_after(old->gen, gc.gen))
1430 if (c->opts.reconstruct_alloc ||
1431 fsck_err_on(new.data_type != gc.data_type, c,
1432 alloc_key_data_type_wrong,
1433 "bucket %llu:%llu gen %u has wrong data_type"
1434 ": got %s, should be %s",
1435 iter->pos.inode, iter->pos.offset,
1437 bch2_data_types[new.data_type],
1438 bch2_data_types[gc.data_type]))
1439 new.data_type = gc.data_type;
1441 #define copy_bucket_field(_errtype, _f) \
1442 if (c->opts.reconstruct_alloc || \
1443 fsck_err_on(new._f != gc._f, c, _errtype, \
1444 "bucket %llu:%llu gen %u data type %s has wrong " #_f \
1445 ": got %u, should be %u", \
1446 iter->pos.inode, iter->pos.offset, \
1448 bch2_data_types[gc.data_type], \
1452 copy_bucket_field(alloc_key_gen_wrong,
1454 copy_bucket_field(alloc_key_dirty_sectors_wrong,
1456 copy_bucket_field(alloc_key_cached_sectors_wrong,
1458 copy_bucket_field(alloc_key_stripe_wrong,
1460 copy_bucket_field(alloc_key_stripe_redundancy_wrong,
1462 #undef copy_bucket_field
1464 if (!bch2_alloc_v4_cmp(*old, new))
1467 a = bch2_alloc_to_v4_mut(trans, k);
1468 ret = PTR_ERR_OR_ZERO(a);
1475 * The trigger normally makes sure this is set, but we're not running
1478 if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1479 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1481 ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1486 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1488 struct btree_trans *trans = bch2_trans_get(c);
1489 struct btree_iter iter;
1495 for_each_member_device(ca, c, i) {
1496 ret = for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1497 POS(ca->dev_idx, ca->mi.first_bucket),
1498 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1499 NULL, NULL, BCH_TRANS_COMMIT_lazy_rw,
1500 bch2_alloc_write_key(trans, &iter, k, metadata_only));
1504 percpu_ref_put(&ca->ref);
1509 bch2_trans_put(trans);
1510 return ret < 0 ? ret : 0;
1513 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1516 struct btree_trans *trans = bch2_trans_get(c);
1517 struct btree_iter iter;
1520 struct bch_alloc_v4 a_convert;
1521 const struct bch_alloc_v4 *a;
1525 for_each_member_device(ca, c, i) {
1526 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1527 ca->mi.nbuckets * sizeof(struct bucket),
1528 GFP_KERNEL|__GFP_ZERO);
1530 percpu_ref_put(&ca->ref);
1531 bch_err(c, "error allocating ca->buckets[gc]");
1532 ret = -BCH_ERR_ENOMEM_gc_alloc_start;
1536 buckets->first_bucket = ca->mi.first_bucket;
1537 buckets->nbuckets = ca->mi.nbuckets;
1538 rcu_assign_pointer(ca->buckets_gc, buckets);
1541 ret = for_each_btree_key2(trans, iter, BTREE_ID_alloc, POS_MIN,
1542 BTREE_ITER_PREFETCH, k, ({
1543 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1544 g = gc_bucket(ca, k.k->p.offset);
1546 a = bch2_alloc_to_v4(k, &a_convert);
1551 if (metadata_only &&
1552 (a->data_type == BCH_DATA_user ||
1553 a->data_type == BCH_DATA_cached ||
1554 a->data_type == BCH_DATA_parity)) {
1555 g->data_type = a->data_type;
1556 g->dirty_sectors = a->dirty_sectors;
1557 g->cached_sectors = a->cached_sectors;
1558 g->stripe = a->stripe;
1559 g->stripe_redundancy = a->stripe_redundancy;
1565 bch2_trans_put(trans);
1571 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1576 for_each_member_device(ca, c, i) {
1577 struct bucket_array *buckets = gc_bucket_array(ca);
1580 for_each_bucket(g, buckets) {
1581 if (metadata_only &&
1582 (g->data_type == BCH_DATA_user ||
1583 g->data_type == BCH_DATA_cached ||
1584 g->data_type == BCH_DATA_parity))
1587 g->dirty_sectors = 0;
1588 g->cached_sectors = 0;
1593 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1594 struct btree_iter *iter,
1598 struct bch_fs *c = trans->c;
1599 const __le64 *refcount = bkey_refcount_c(k);
1600 struct printbuf buf = PRINTBUF;
1601 struct reflink_gc *r;
1607 while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1608 r->offset < k.k->p.offset)
1612 r->offset != k.k->p.offset ||
1613 r->size != k.k->size) {
1614 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1618 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1619 reflink_v_refcount_wrong,
1620 "reflink key has wrong refcount:\n"
1623 (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1625 struct bkey_i *new = bch2_bkey_make_mut(trans, iter, &k, 0);
1627 ret = PTR_ERR_OR_ZERO(new);
1632 new->k.type = KEY_TYPE_deleted;
1634 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1637 printbuf_exit(&buf);
1641 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1643 struct btree_trans *trans;
1644 struct btree_iter iter;
1652 trans = bch2_trans_get(c);
1654 ret = for_each_btree_key_commit(trans, iter,
1655 BTREE_ID_reflink, POS_MIN,
1656 BTREE_ITER_PREFETCH, k,
1657 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1658 bch2_gc_write_reflink_key(trans, &iter, k, &idx));
1660 c->reflink_gc_nr = 0;
1661 bch2_trans_put(trans);
1665 static int bch2_gc_reflink_start(struct bch_fs *c,
1668 struct btree_trans *trans;
1669 struct btree_iter iter;
1671 struct reflink_gc *r;
1677 trans = bch2_trans_get(c);
1678 c->reflink_gc_nr = 0;
1680 for_each_btree_key(trans, iter, BTREE_ID_reflink, POS_MIN,
1681 BTREE_ITER_PREFETCH, k, ret) {
1682 const __le64 *refcount = bkey_refcount_c(k);
1687 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1690 ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1694 r->offset = k.k->p.offset;
1695 r->size = k.k->size;
1698 bch2_trans_iter_exit(trans, &iter);
1700 bch2_trans_put(trans);
1704 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1706 struct genradix_iter iter;
1707 struct reflink_gc *r;
1709 genradix_for_each(&c->reflink_gc_table, iter, r)
1713 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1714 struct btree_iter *iter,
1717 struct bch_fs *c = trans->c;
1718 struct printbuf buf = PRINTBUF;
1719 const struct bch_stripe *s;
1720 struct gc_stripe *m;
1725 if (k.k->type != KEY_TYPE_stripe)
1728 s = bkey_s_c_to_stripe(k).v;
1729 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1731 for (i = 0; i < s->nr_blocks; i++) {
1732 u32 old = stripe_blockcount_get(s, i);
1733 u32 new = (m ? m->block_sectors[i] : 0);
1736 prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1743 bch2_bkey_val_to_text(&buf, c, k);
1745 if (fsck_err_on(bad, c, stripe_sector_count_wrong,
1747 struct bkey_i_stripe *new;
1749 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1750 ret = PTR_ERR_OR_ZERO(new);
1754 bkey_reassemble(&new->k_i, k);
1756 for (i = 0; i < new->v.nr_blocks; i++)
1757 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1759 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1762 printbuf_exit(&buf);
1766 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1768 struct btree_trans *trans;
1769 struct btree_iter iter;
1776 trans = bch2_trans_get(c);
1778 ret = for_each_btree_key_commit(trans, iter,
1779 BTREE_ID_stripes, POS_MIN,
1780 BTREE_ITER_PREFETCH, k,
1781 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1782 bch2_gc_write_stripes_key(trans, &iter, k));
1784 bch2_trans_put(trans);
1788 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1790 genradix_free(&c->gc_stripes);
1794 * bch2_gc - walk _all_ references to buckets, and recompute them:
1796 * @c: filesystem object
1797 * @initial: are we in recovery?
1798 * @metadata_only: are we just checking metadata references, or everything?
1800 * Returns: 0 on success, or standard errcode on failure
1802 * Order matters here:
1803 * - Concurrent GC relies on the fact that we have a total ordering for
1804 * everything that GC walks - see gc_will_visit_node(),
1805 * gc_will_visit_root()
1807 * - also, references move around in the course of index updates and
1808 * various other crap: everything needs to agree on the ordering
1809 * references are allowed to move around in - e.g., we're allowed to
1810 * start with a reference owned by an open_bucket (the allocator) and
1811 * move it to the btree, but not the reverse.
1813 * This is necessary to ensure that gc doesn't miss references that
1814 * move around - if references move backwards in the ordering GC
1815 * uses, GC could skip past them
1817 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1822 lockdep_assert_held(&c->state_lock);
1824 down_write(&c->gc_lock);
1826 bch2_btree_interior_updates_flush(c);
1828 ret = bch2_gc_start(c) ?:
1829 bch2_gc_alloc_start(c, metadata_only) ?:
1830 bch2_gc_reflink_start(c, metadata_only);
1834 gc_pos_set(c, gc_phase(GC_PHASE_START));
1836 bch2_mark_superblocks(c);
1838 ret = bch2_gc_btrees(c, initial, metadata_only);
1844 bch2_mark_pending_btree_node_frees(c);
1848 if (test_bit(BCH_FS_need_another_gc, &c->flags) ||
1849 (!iter && bch2_test_restart_gc)) {
1851 bch_info(c, "Unable to fix bucket gens, looping");
1857 * XXX: make sure gens we fixed got saved
1859 bch_info(c, "Second GC pass needed, restarting:");
1860 clear_bit(BCH_FS_need_another_gc, &c->flags);
1861 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1863 bch2_gc_stripes_reset(c, metadata_only);
1864 bch2_gc_alloc_reset(c, metadata_only);
1865 bch2_gc_reflink_reset(c, metadata_only);
1866 ret = bch2_gc_reset(c);
1870 /* flush fsck errors, reset counters */
1871 bch2_flush_fsck_errs(c);
1876 bch2_journal_block(&c->journal);
1878 ret = bch2_gc_stripes_done(c, metadata_only) ?:
1879 bch2_gc_reflink_done(c, metadata_only) ?:
1880 bch2_gc_alloc_done(c, metadata_only) ?:
1881 bch2_gc_done(c, initial, metadata_only);
1883 bch2_journal_unblock(&c->journal);
1886 percpu_down_write(&c->mark_lock);
1887 /* Indicates that gc is no longer in progress: */
1888 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1891 percpu_up_write(&c->mark_lock);
1893 up_write(&c->gc_lock);
1896 * At startup, allocations can happen directly instead of via the
1897 * allocator thread - issue wakeup in case they blocked on gc_lock:
1899 closure_wake_up(&c->freelist_wait);
1906 static int gc_btree_gens_key(struct btree_trans *trans,
1907 struct btree_iter *iter,
1910 struct bch_fs *c = trans->c;
1911 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1912 const struct bch_extent_ptr *ptr;
1916 percpu_down_read(&c->mark_lock);
1917 bkey_for_each_ptr(ptrs, ptr) {
1918 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1920 if (ptr_stale(ca, ptr) > 16) {
1921 percpu_up_read(&c->mark_lock);
1926 bkey_for_each_ptr(ptrs, ptr) {
1927 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1928 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1930 if (gen_after(*gen, ptr->gen))
1933 percpu_up_read(&c->mark_lock);
1936 u = bch2_bkey_make_mut(trans, iter, &k, 0);
1937 ret = PTR_ERR_OR_ZERO(u);
1941 bch2_extent_normalize(c, bkey_i_to_s(u));
1945 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1948 struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1949 struct bch_alloc_v4 a_convert;
1950 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1951 struct bkey_i_alloc_v4 *a_mut;
1954 if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1957 a_mut = bch2_alloc_to_v4_mut(trans, k);
1958 ret = PTR_ERR_OR_ZERO(a_mut);
1962 a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1963 a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1965 return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1968 int bch2_gc_gens(struct bch_fs *c)
1970 struct btree_trans *trans;
1971 struct btree_iter iter;
1974 u64 b, start_time = local_clock();
1979 * Ideally we would be using state_lock and not gc_lock here, but that
1980 * introduces a deadlock in the RO path - we currently take the state
1981 * lock at the start of going RO, thus the gc thread may get stuck:
1983 if (!mutex_trylock(&c->gc_gens_lock))
1986 trace_and_count(c, gc_gens_start, c);
1987 down_read(&c->gc_lock);
1988 trans = bch2_trans_get(c);
1990 for_each_member_device(ca, c, i) {
1991 struct bucket_gens *gens = bucket_gens(ca);
1993 BUG_ON(ca->oldest_gen);
1995 ca->oldest_gen = kvmalloc(gens->nbuckets, GFP_KERNEL);
1996 if (!ca->oldest_gen) {
1997 percpu_ref_put(&ca->ref);
1998 ret = -BCH_ERR_ENOMEM_gc_gens;
2002 for (b = gens->first_bucket;
2003 b < gens->nbuckets; b++)
2004 ca->oldest_gen[b] = gens->b[b];
2007 for (i = 0; i < BTREE_ID_NR; i++)
2008 if (btree_type_has_ptrs(i)) {
2009 c->gc_gens_btree = i;
2010 c->gc_gens_pos = POS_MIN;
2012 ret = for_each_btree_key_commit(trans, iter, i,
2014 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
2017 BCH_TRANS_COMMIT_no_enospc,
2018 gc_btree_gens_key(trans, &iter, k));
2019 if (ret && !bch2_err_matches(ret, EROFS))
2025 ret = for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
2027 BTREE_ITER_PREFETCH,
2030 BCH_TRANS_COMMIT_no_enospc,
2031 bch2_alloc_write_oldest_gen(trans, &iter, k));
2032 if (ret && !bch2_err_matches(ret, EROFS))
2037 c->gc_gens_btree = 0;
2038 c->gc_gens_pos = POS_MIN;
2042 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2043 trace_and_count(c, gc_gens_end, c);
2045 for_each_member_device(ca, c, i) {
2046 kvfree(ca->oldest_gen);
2047 ca->oldest_gen = NULL;
2050 bch2_trans_put(trans);
2051 up_read(&c->gc_lock);
2052 mutex_unlock(&c->gc_gens_lock);
2056 static int bch2_gc_thread(void *arg)
2058 struct bch_fs *c = arg;
2059 struct io_clock *clock = &c->io_clock[WRITE];
2060 unsigned long last = atomic64_read(&clock->now);
2061 unsigned last_kick = atomic_read(&c->kick_gc);
2068 set_current_state(TASK_INTERRUPTIBLE);
2070 if (kthread_should_stop()) {
2071 __set_current_state(TASK_RUNNING);
2075 if (atomic_read(&c->kick_gc) != last_kick)
2078 if (c->btree_gc_periodic) {
2079 unsigned long next = last + c->capacity / 16;
2081 if (atomic64_read(&clock->now) >= next)
2084 bch2_io_clock_schedule_timeout(clock, next);
2091 __set_current_state(TASK_RUNNING);
2093 last = atomic64_read(&clock->now);
2094 last_kick = atomic_read(&c->kick_gc);
2097 * Full gc is currently incompatible with btree key cache:
2100 ret = bch2_gc(c, false, false);
2102 ret = bch2_gc_gens(c);
2107 debug_check_no_locks_held();
2113 void bch2_gc_thread_stop(struct bch_fs *c)
2115 struct task_struct *p;
2118 c->gc_thread = NULL;
2126 int bch2_gc_thread_start(struct bch_fs *c)
2128 struct task_struct *p;
2133 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2135 bch_err_fn(c, PTR_ERR(p));