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 "backpointers.h"
11 #include "bkey_methods.h"
13 #include "btree_journal_iter.h"
14 #include "btree_key_cache.h"
15 #include "btree_locking.h"
16 #include "btree_node_scan.h"
17 #include "btree_update_interior.h"
29 #include "recovery_passes.h"
35 #include <linux/slab.h>
36 #include <linux/bitops.h>
37 #include <linux/freezer.h>
38 #include <linux/kthread.h>
39 #include <linux/preempt.h>
40 #include <linux/rcupdate.h>
41 #include <linux/sched/task.h>
43 #define DROP_THIS_NODE 10
44 #define DROP_PREV_NODE 11
45 #define DID_FILL_FROM_SCAN 12
47 static struct bkey_s unsafe_bkey_s_c_to_s(struct bkey_s_c k)
49 return (struct bkey_s) {{{
51 (struct bch_val *) k.v
55 static bool should_restart_for_topology_repair(struct bch_fs *c)
57 return c->opts.fix_errors != FSCK_FIX_no &&
58 !(c->recovery_passes_complete & BIT_ULL(BCH_RECOVERY_PASS_check_topology));
61 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
64 write_seqcount_begin(&c->gc_pos_lock);
66 write_seqcount_end(&c->gc_pos_lock);
70 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
72 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
73 __gc_pos_set(c, new_pos);
76 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
78 switch (b->key.k.type) {
79 case KEY_TYPE_btree_ptr: {
80 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
84 dst->v.seq = b->data->keys.seq;
85 dst->v.sectors_written = 0;
87 dst->v.min_key = b->data->min_key;
88 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
89 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
92 case KEY_TYPE_btree_ptr_v2:
93 bkey_copy(&dst->k_i, &b->key);
100 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
101 enum btree_id btree, unsigned level,
102 struct bkey_s_c old, struct bkey_i *new)
104 struct bch_fs *c = trans->c;
109 bch2_bkey_buf_init(&tmp);
110 bch2_bkey_buf_reassemble(&tmp, c, old);
112 b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
113 if (!IS_ERR_OR_NULL(b)) {
114 mutex_lock(&c->btree_cache.lock);
116 bch2_btree_node_hash_remove(&c->btree_cache, b);
118 bkey_copy(&b->key, new);
119 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
122 mutex_unlock(&c->btree_cache.lock);
123 six_unlock_read(&b->c.lock);
126 bch2_bkey_buf_exit(&tmp, c);
129 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
131 struct bkey_i_btree_ptr_v2 *new;
134 if (c->opts.verbose) {
135 struct printbuf buf = PRINTBUF;
137 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
138 prt_str(&buf, " -> ");
139 bch2_bpos_to_text(&buf, new_min);
141 bch_info(c, "%s(): %s", __func__, buf.buf);
145 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
147 return -BCH_ERR_ENOMEM_gc_repair_key;
149 btree_ptr_to_v2(b, new);
150 b->data->min_key = new_min;
151 new->v.min_key = new_min;
152 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
154 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
160 bch2_btree_node_drop_keys_outside_node(b);
161 bkey_copy(&b->key, &new->k_i);
165 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
167 struct bkey_i_btree_ptr_v2 *new;
170 if (c->opts.verbose) {
171 struct printbuf buf = PRINTBUF;
173 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
174 prt_str(&buf, " -> ");
175 bch2_bpos_to_text(&buf, new_max);
177 bch_info(c, "%s(): %s", __func__, buf.buf);
181 ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
185 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
187 return -BCH_ERR_ENOMEM_gc_repair_key;
189 btree_ptr_to_v2(b, new);
190 b->data->max_key = new_max;
192 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
194 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
200 bch2_btree_node_drop_keys_outside_node(b);
202 mutex_lock(&c->btree_cache.lock);
203 bch2_btree_node_hash_remove(&c->btree_cache, b);
205 bkey_copy(&b->key, &new->k_i);
206 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
208 mutex_unlock(&c->btree_cache.lock);
212 static int btree_check_node_boundaries(struct bch_fs *c, struct btree *b,
213 struct btree *prev, struct btree *cur,
214 struct bpos *pulled_from_scan)
216 struct bpos expected_start = !prev
218 : bpos_successor(prev->key.k.p);
219 struct printbuf buf = PRINTBUF;
222 BUG_ON(b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
223 !bpos_eq(bkey_i_to_btree_ptr_v2(&b->key)->v.min_key,
226 if (bpos_eq(expected_start, cur->data->min_key))
229 prt_printf(&buf, " at btree %s level %u:\n parent: ",
230 bch2_btree_id_str(b->c.btree_id), b->c.level);
231 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
234 prt_printf(&buf, "\n prev: ");
235 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&prev->key));
238 prt_str(&buf, "\n next: ");
239 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&cur->key));
241 if (bpos_lt(expected_start, cur->data->min_key)) { /* gap */
242 if (b->c.level == 1 &&
243 bpos_lt(*pulled_from_scan, cur->data->min_key)) {
244 ret = bch2_get_scanned_nodes(c, b->c.btree_id, 0,
246 bpos_predecessor(cur->data->min_key));
250 *pulled_from_scan = cur->data->min_key;
251 ret = DID_FILL_FROM_SCAN;
253 if (mustfix_fsck_err(c, btree_node_topology_bad_min_key,
254 "btree node with incorrect min_key%s", buf.buf))
255 ret = set_node_min(c, cur, expected_start);
257 } else { /* overlap */
258 if (prev && BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) { /* cur overwrites prev */
259 if (bpos_ge(prev->data->min_key, cur->data->min_key)) { /* fully? */
260 if (mustfix_fsck_err(c, btree_node_topology_overwritten_by_next_node,
261 "btree node overwritten by next node%s", buf.buf))
262 ret = DROP_PREV_NODE;
264 if (mustfix_fsck_err(c, btree_node_topology_bad_max_key,
265 "btree node with incorrect max_key%s", buf.buf))
266 ret = set_node_max(c, prev,
267 bpos_predecessor(cur->data->min_key));
270 if (bpos_ge(expected_start, cur->data->max_key)) { /* fully? */
271 if (mustfix_fsck_err(c, btree_node_topology_overwritten_by_prev_node,
272 "btree node overwritten by prev node%s", buf.buf))
273 ret = DROP_THIS_NODE;
275 if (mustfix_fsck_err(c, btree_node_topology_bad_min_key,
276 "btree node with incorrect min_key%s", buf.buf))
277 ret = set_node_min(c, cur, expected_start);
287 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
288 struct btree *child, struct bpos *pulled_from_scan)
290 struct printbuf buf = PRINTBUF;
293 if (bpos_eq(child->key.k.p, b->key.k.p))
296 prt_printf(&buf, "at btree %s level %u:\n parent: ",
297 bch2_btree_id_str(b->c.btree_id), b->c.level);
298 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
300 prt_str(&buf, "\n child: ");
301 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&child->key));
303 if (mustfix_fsck_err(c, btree_node_topology_bad_max_key,
304 "btree node with incorrect max_key%s", buf.buf)) {
305 if (b->c.level == 1 &&
306 bpos_lt(*pulled_from_scan, b->key.k.p)) {
307 ret = bch2_get_scanned_nodes(c, b->c.btree_id, 0,
308 bpos_successor(child->key.k.p), b->key.k.p);
312 *pulled_from_scan = b->key.k.p;
313 ret = DID_FILL_FROM_SCAN;
315 ret = set_node_max(c, child, b->key.k.p);
324 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b,
325 struct bpos *pulled_from_scan)
327 struct bch_fs *c = trans->c;
328 struct btree_and_journal_iter iter;
330 struct bkey_buf prev_k, cur_k;
331 struct btree *prev = NULL, *cur = NULL;
332 bool have_child, new_pass = false;
333 struct printbuf buf = PRINTBUF;
339 bch2_bkey_buf_init(&prev_k);
340 bch2_bkey_buf_init(&cur_k);
343 have_child = new_pass = false;
344 bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
345 iter.prefetch = true;
347 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
348 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
349 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
351 bch2_btree_and_journal_iter_advance(&iter);
352 bch2_bkey_buf_reassemble(&cur_k, c, k);
354 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
355 b->c.btree_id, b->c.level - 1,
357 ret = PTR_ERR_OR_ZERO(cur);
359 printbuf_reset(&buf);
360 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
362 if (mustfix_fsck_err_on(bch2_err_matches(ret, EIO), c,
363 btree_node_unreadable,
364 "Topology repair: unreadable btree node at btree %s level %u:\n"
366 bch2_btree_id_str(b->c.btree_id),
369 bch2_btree_node_evict(trans, cur_k.k);
371 ret = bch2_journal_key_delete(c, b->c.btree_id,
372 b->c.level, cur_k.k->k.p);
376 if (!btree_id_is_alloc(b->c.btree_id)) {
377 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_scan_for_btree_nodes);
384 bch_err_msg(c, ret, "getting btree node");
388 if (bch2_btree_node_is_stale(c, cur)) {
389 bch_info(c, "btree node %s older than nodes found by scanning", buf.buf);
390 six_unlock_read(&cur->c.lock);
391 bch2_btree_node_evict(trans, cur_k.k);
392 ret = bch2_journal_key_delete(c, b->c.btree_id,
393 b->c.level, cur_k.k->k.p);
400 ret = btree_check_node_boundaries(c, b, prev, cur, pulled_from_scan);
401 if (ret == DID_FILL_FROM_SCAN) {
406 if (ret == DROP_THIS_NODE) {
407 six_unlock_read(&cur->c.lock);
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 six_unlock_read(&prev->c.lock);
421 if (ret == DROP_PREV_NODE) {
422 bch_info(c, "dropped prev node");
423 bch2_btree_node_evict(trans, prev_k.k);
424 ret = bch2_journal_key_delete(c, b->c.btree_id,
425 b->c.level, prev_k.k->k.p);
429 bch2_btree_and_journal_iter_exit(&iter);
436 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
439 if (!ret && !IS_ERR_OR_NULL(prev)) {
441 ret = btree_repair_node_end(c, b, prev, pulled_from_scan);
442 if (ret == DID_FILL_FROM_SCAN) {
448 if (!IS_ERR_OR_NULL(prev))
449 six_unlock_read(&prev->c.lock);
451 if (!IS_ERR_OR_NULL(cur))
452 six_unlock_read(&cur->c.lock);
458 bch2_btree_and_journal_iter_exit(&iter);
463 bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
464 iter.prefetch = true;
466 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
467 bch2_bkey_buf_reassemble(&cur_k, c, k);
468 bch2_btree_and_journal_iter_advance(&iter);
470 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
471 b->c.btree_id, b->c.level - 1,
473 ret = PTR_ERR_OR_ZERO(cur);
475 bch_err_msg(c, ret, "getting btree node");
479 ret = bch2_btree_repair_topology_recurse(trans, cur, pulled_from_scan);
480 six_unlock_read(&cur->c.lock);
483 if (ret == DROP_THIS_NODE) {
484 bch2_btree_node_evict(trans, cur_k.k);
485 ret = bch2_journal_key_delete(c, b->c.btree_id,
486 b->c.level, cur_k.k->k.p);
496 printbuf_reset(&buf);
497 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
499 if (mustfix_fsck_err_on(!have_child, c,
500 btree_node_topology_interior_node_empty,
501 "empty interior btree node at btree %s level %u\n"
503 bch2_btree_id_str(b->c.btree_id),
504 b->c.level, buf.buf))
505 ret = DROP_THIS_NODE;
508 if (!IS_ERR_OR_NULL(prev))
509 six_unlock_read(&prev->c.lock);
510 if (!IS_ERR_OR_NULL(cur))
511 six_unlock_read(&cur->c.lock);
513 bch2_btree_and_journal_iter_exit(&iter);
515 if (!ret && new_pass)
518 BUG_ON(!ret && bch2_btree_node_check_topology(trans, b));
520 bch2_bkey_buf_exit(&prev_k, c);
521 bch2_bkey_buf_exit(&cur_k, c);
526 int bch2_check_topology(struct bch_fs *c)
528 struct btree_trans *trans = bch2_trans_get(c);
529 struct bpos pulled_from_scan = POS_MIN;
532 for (unsigned i = 0; i < btree_id_nr_alive(c) && !ret; i++) {
533 struct btree_root *r = bch2_btree_id_root(c, i);
534 bool reconstructed_root = false;
537 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_scan_for_btree_nodes);
541 bch_info(c, "btree root %s unreadable, must recover from scan", bch2_btree_id_str(i));
546 if (!bch2_btree_has_scanned_nodes(c, i)) {
547 mustfix_fsck_err(c, btree_root_unreadable_and_scan_found_nothing,
548 "no nodes found for btree %s, continue?", bch2_btree_id_str(i));
549 bch2_btree_root_alloc_fake(c, i, 0);
551 bch2_btree_root_alloc_fake(c, i, 1);
552 bch2_shoot_down_journal_keys(c, i, 1, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
553 ret = bch2_get_scanned_nodes(c, i, 0, POS_MIN, SPOS_MAX);
558 reconstructed_root = true;
561 struct btree *b = r->b;
563 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
564 ret = bch2_btree_repair_topology_recurse(trans, b, &pulled_from_scan);
565 six_unlock_read(&b->c.lock);
567 if (ret == DROP_THIS_NODE) {
568 bch2_btree_node_hash_remove(&c->btree_cache, b);
569 mutex_lock(&c->btree_cache.lock);
570 list_move(&b->list, &c->btree_cache.freeable);
571 mutex_unlock(&c->btree_cache.lock);
575 if (!reconstructed_root)
576 goto reconstruct_root;
578 bch_err(c, "empty btree root %s", bch2_btree_id_str(i));
579 bch2_btree_root_alloc_fake(c, i, 0);
585 bch2_trans_put(trans);
589 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
590 unsigned level, bool is_root,
593 struct bch_fs *c = trans->c;
594 struct bkey_ptrs_c ptrs_c = bch2_bkey_ptrs_c(*k);
595 const union bch_extent_entry *entry_c;
596 struct extent_ptr_decoded p = { 0 };
597 bool do_update = false;
598 struct printbuf buf = PRINTBUF;
603 * use check_bucket_ref here
605 bkey_for_each_ptr_decode(k->k, ptrs_c, p, entry_c) {
606 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
607 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
608 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, p, entry_c);
610 if (fsck_err_on(!g->gen_valid,
611 c, ptr_to_missing_alloc_key,
612 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
614 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
615 bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
617 (printbuf_reset(&buf),
618 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
627 if (fsck_err_on(gen_cmp(p.ptr.gen, g->gen) > 0,
628 c, ptr_gen_newer_than_bucket_gen,
629 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
631 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
632 bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
634 (printbuf_reset(&buf),
635 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
640 g->dirty_sectors = 0;
641 g->cached_sectors = 0;
642 set_bit(BCH_FS_need_another_gc, &c->flags);
648 if (fsck_err_on(gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX,
649 c, ptr_gen_newer_than_bucket_gen,
650 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
652 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
653 bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
655 (printbuf_reset(&buf),
656 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
659 if (fsck_err_on(!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0,
661 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
663 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
664 bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
666 (printbuf_reset(&buf),
667 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
670 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
673 if (fsck_err_on(bucket_data_type(g->data_type) &&
674 bucket_data_type(g->data_type) !=
675 bucket_data_type(data_type), c,
676 ptr_bucket_data_type_mismatch,
677 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
679 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
680 bch2_data_type_str(g->data_type),
681 bch2_data_type_str(data_type),
682 (printbuf_reset(&buf),
683 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
684 if (data_type == BCH_DATA_btree) {
685 g->data_type = data_type;
686 set_bit(BCH_FS_need_another_gc, &c->flags);
693 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
695 if (fsck_err_on(!m || !m->alive, c,
696 ptr_to_missing_stripe,
697 "pointer to nonexistent stripe %llu\n"
700 (printbuf_reset(&buf),
701 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
704 if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
705 ptr_to_incorrect_stripe,
706 "pointer does not match stripe %llu\n"
709 (printbuf_reset(&buf),
710 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
717 bch_err(c, "cannot update btree roots yet");
722 struct bkey_i *new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
724 ret = -BCH_ERR_ENOMEM_gc_repair_key;
725 bch_err_msg(c, ret, "allocating new key");
729 bkey_reassemble(new, *k);
733 * We don't want to drop btree node pointers - if the
734 * btree node isn't there anymore, the read path will
737 struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
738 bkey_for_each_ptr(ptrs, ptr) {
739 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
740 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
745 struct bkey_ptrs ptrs;
746 union bch_extent_entry *entry;
748 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
749 bkey_for_each_ptr_decode(bkey_i_to_s(new).k, ptrs, p, entry) {
750 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
751 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
752 enum bch_data_type data_type = bch2_bkey_ptr_data_type(bkey_i_to_s_c(new), p, entry);
755 (!g->gen_valid || gen_cmp(p.ptr.gen, g->gen) > 0)) ||
757 gen_cmp(p.ptr.gen, g->gen) < 0) ||
758 gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX ||
760 g->data_type != data_type)) {
761 bch2_bkey_drop_ptr(bkey_i_to_s(new), &entry->ptr);
762 goto restart_drop_ptrs;
766 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
767 bkey_extent_entry_for_each(ptrs, entry) {
768 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
769 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
770 entry->stripe_ptr.idx);
771 union bch_extent_entry *next_ptr;
773 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
774 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
779 bch_err(c, "aieee, found stripe ptr with no data ptr");
783 if (!m || !m->alive ||
784 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
787 bch2_bkey_extent_entry_drop(new, entry);
795 bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
798 printbuf_reset(&buf);
799 bch2_bkey_val_to_text(&buf, c, *k);
800 bch_info(c, "updated %s", buf.buf);
802 printbuf_reset(&buf);
803 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
804 bch_info(c, "new key %s", buf.buf);
807 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
813 *k = bkey_i_to_s_c(new);
821 /* marking of btree keys/nodes: */
823 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
824 unsigned level, bool is_root,
828 struct bch_fs *c = trans->c;
829 struct bkey deleted = KEY(0, 0, 0);
830 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
831 struct printbuf buf = PRINTBUF;
837 BUG_ON(bch2_journal_seq_verify &&
838 k->k->version.lo > atomic64_read(&c->journal.seq));
840 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
841 bkey_version_in_future,
842 "key version number higher than recorded: %llu > %llu",
844 atomic64_read(&c->key_version)))
845 atomic64_set(&c->key_version, k->k->version.lo);
848 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
852 if (mustfix_fsck_err_on(level && !bch2_dev_btree_bitmap_marked(c, *k),
853 c, btree_bitmap_not_marked,
854 "btree ptr not marked in member info btree allocated bitmap\n %s",
855 (bch2_bkey_val_to_text(&buf, c, *k),
857 mutex_lock(&c->sb_lock);
858 bch2_dev_btree_bitmap_mark(c, *k);
860 mutex_unlock(&c->sb_lock);
863 ret = commit_do(trans, NULL, NULL, 0,
864 bch2_key_trigger(trans, btree_id, level, old,
865 unsafe_bkey_s_c_to_s(*k), BTREE_TRIGGER_GC));
873 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
875 struct btree_node_iter iter;
876 struct bkey unpacked;
880 ret = bch2_btree_node_check_topology(trans, b);
884 if (!btree_node_type_needs_gc(btree_node_type(b)))
887 bch2_btree_node_iter_init_from_start(&iter, b);
889 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
890 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
895 bch2_btree_node_iter_advance(&iter, b);
901 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
902 bool initial, bool metadata_only)
904 struct bch_fs *c = trans->c;
905 struct btree_iter iter;
907 unsigned depth = metadata_only ? 1 : 0;
910 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
912 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
913 0, depth, BTREE_ITER_PREFETCH, b, ret) {
914 bch2_verify_btree_nr_keys(b);
916 gc_pos_set(c, gc_pos_btree_node(b));
918 ret = btree_gc_mark_node(trans, b, initial);
922 bch2_trans_iter_exit(trans, &iter);
927 mutex_lock(&c->btree_root_lock);
928 b = bch2_btree_id_root(c, btree_id)->b;
929 if (!btree_node_fake(b)) {
930 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
932 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
935 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
936 mutex_unlock(&c->btree_root_lock);
941 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
942 unsigned target_depth)
944 struct bch_fs *c = trans->c;
945 struct btree_and_journal_iter iter;
948 struct printbuf buf = PRINTBUF;
951 ret = bch2_btree_node_check_topology(trans, b);
955 bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
956 bch2_bkey_buf_init(&cur);
958 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
959 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
960 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
962 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
967 bch2_btree_and_journal_iter_advance(&iter);
970 if (b->c.level > target_depth) {
971 bch2_btree_and_journal_iter_exit(&iter);
972 bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
973 iter.prefetch = true;
975 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
978 bch2_bkey_buf_reassemble(&cur, c, k);
979 bch2_btree_and_journal_iter_advance(&iter);
981 child = bch2_btree_node_get_noiter(trans, cur.k,
982 b->c.btree_id, b->c.level - 1,
984 ret = PTR_ERR_OR_ZERO(child);
986 if (bch2_err_matches(ret, EIO)) {
987 bch2_topology_error(c);
993 btree_node_read_error,
994 "Unreadable btree node at btree %s level %u:\n"
996 bch2_btree_id_str(b->c.btree_id),
998 (printbuf_reset(&buf),
999 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
1000 should_restart_for_topology_repair(c)) {
1001 bch_info(c, "Halting mark and sweep to start topology repair pass");
1002 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
1005 /* Continue marking when opted to not
1008 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
1012 bch_err_msg(c, ret, "getting btree node");
1016 ret = bch2_gc_btree_init_recurse(trans, child,
1018 six_unlock_read(&child->c.lock);
1025 bch2_bkey_buf_exit(&cur, c);
1026 bch2_btree_and_journal_iter_exit(&iter);
1027 printbuf_exit(&buf);
1031 static int bch2_gc_btree_init(struct btree_trans *trans,
1032 enum btree_id btree_id,
1035 struct bch_fs *c = trans->c;
1037 unsigned target_depth = metadata_only ? 1 : 0;
1038 struct printbuf buf = PRINTBUF;
1041 b = bch2_btree_id_root(c, btree_id)->b;
1043 six_lock_read(&b->c.lock, NULL, NULL);
1044 printbuf_reset(&buf);
1045 bch2_bpos_to_text(&buf, b->data->min_key);
1046 if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1047 btree_root_bad_min_key,
1048 "btree root with incorrect min_key: %s", buf.buf)) {
1049 bch_err(c, "repair unimplemented");
1050 ret = -BCH_ERR_fsck_repair_unimplemented;
1054 printbuf_reset(&buf);
1055 bch2_bpos_to_text(&buf, b->data->max_key);
1056 if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1057 btree_root_bad_max_key,
1058 "btree root with incorrect max_key: %s", buf.buf)) {
1059 bch_err(c, "repair unimplemented");
1060 ret = -BCH_ERR_fsck_repair_unimplemented;
1064 if (b->c.level >= target_depth)
1065 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1068 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1070 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1074 six_unlock_read(&b->c.lock);
1077 printbuf_exit(&buf);
1081 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1083 return (int) btree_id_to_gc_phase(l) -
1084 (int) btree_id_to_gc_phase(r);
1087 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1089 struct btree_trans *trans = bch2_trans_get(c);
1090 enum btree_id ids[BTREE_ID_NR];
1094 for (i = 0; i < BTREE_ID_NR; i++)
1096 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1098 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1100 ? bch2_gc_btree_init(trans, ids[i], metadata_only)
1101 : bch2_gc_btree(trans, ids[i], initial, metadata_only);
1103 for (i = BTREE_ID_NR; i < btree_id_nr_alive(c) && !ret; i++) {
1104 if (!bch2_btree_id_root(c, i)->alive)
1108 ? bch2_gc_btree_init(trans, i, metadata_only)
1109 : bch2_gc_btree(trans, i, initial, metadata_only);
1112 bch2_trans_put(trans);
1117 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1119 enum bch_data_type type,
1122 u64 b = sector_to_bucket(ca, start);
1126 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1128 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1129 gc_phase(GC_PHASE_SB), flags);
1132 } while (start < end);
1135 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1138 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1142 for (i = 0; i < layout->nr_superblocks; i++) {
1143 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1145 if (offset == BCH_SB_SECTOR)
1146 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1147 BCH_DATA_sb, flags);
1149 mark_metadata_sectors(c, ca, offset,
1150 offset + (1 << layout->sb_max_size_bits),
1151 BCH_DATA_sb, flags);
1154 for (i = 0; i < ca->journal.nr; i++) {
1155 b = ca->journal.buckets[i];
1156 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1158 gc_phase(GC_PHASE_SB), flags);
1162 static void bch2_mark_superblocks(struct bch_fs *c)
1164 mutex_lock(&c->sb_lock);
1165 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1167 for_each_online_member(c, ca)
1168 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1169 mutex_unlock(&c->sb_lock);
1173 /* Also see bch2_pending_btree_node_free_insert_done() */
1174 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1176 struct btree_update *as;
1177 struct pending_btree_node_free *d;
1179 mutex_lock(&c->btree_interior_update_lock);
1180 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1182 for_each_pending_btree_node_free(c, as, d)
1183 if (d->index_update_done)
1184 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1186 mutex_unlock(&c->btree_interior_update_lock);
1190 static void bch2_gc_free(struct bch_fs *c)
1192 genradix_free(&c->reflink_gc_table);
1193 genradix_free(&c->gc_stripes);
1195 for_each_member_device(c, ca) {
1196 kvfree(rcu_dereference_protected(ca->buckets_gc, 1));
1197 ca->buckets_gc = NULL;
1199 free_percpu(ca->usage_gc);
1200 ca->usage_gc = NULL;
1203 free_percpu(c->usage_gc);
1207 static int bch2_gc_done(struct bch_fs *c,
1208 bool initial, bool metadata_only)
1210 struct bch_dev *ca = NULL;
1211 struct printbuf buf = PRINTBUF;
1212 bool verify = !metadata_only &&
1213 !c->opts.reconstruct_alloc &&
1214 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1218 percpu_down_write(&c->mark_lock);
1220 #define copy_field(_err, _f, _msg, ...) \
1221 if (dst->_f != src->_f && \
1223 fsck_err(c, _err, _msg ": got %llu, should be %llu" \
1224 , ##__VA_ARGS__, dst->_f, src->_f))) \
1226 #define copy_dev_field(_err, _f, _msg, ...) \
1227 copy_field(_err, _f, "dev %u has wrong " _msg, ca->dev_idx, ##__VA_ARGS__)
1228 #define copy_fs_field(_err, _f, _msg, ...) \
1229 copy_field(_err, _f, "fs has wrong " _msg, ##__VA_ARGS__)
1231 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1232 bch2_fs_usage_acc_to_base(c, i);
1234 __for_each_member_device(c, ca) {
1235 struct bch_dev_usage *dst = ca->usage_base;
1236 struct bch_dev_usage *src = (void *)
1237 bch2_acc_percpu_u64s((u64 __percpu *) ca->usage_gc,
1240 for (i = 0; i < BCH_DATA_NR; i++) {
1241 copy_dev_field(dev_usage_buckets_wrong,
1242 d[i].buckets, "%s buckets", bch2_data_type_str(i));
1243 copy_dev_field(dev_usage_sectors_wrong,
1244 d[i].sectors, "%s sectors", bch2_data_type_str(i));
1245 copy_dev_field(dev_usage_fragmented_wrong,
1246 d[i].fragmented, "%s fragmented", bch2_data_type_str(i));
1251 unsigned nr = fs_usage_u64s(c);
1252 struct bch_fs_usage *dst = c->usage_base;
1253 struct bch_fs_usage *src = (void *)
1254 bch2_acc_percpu_u64s((u64 __percpu *) c->usage_gc, nr);
1256 copy_fs_field(fs_usage_hidden_wrong,
1257 b.hidden, "hidden");
1258 copy_fs_field(fs_usage_btree_wrong,
1261 if (!metadata_only) {
1262 copy_fs_field(fs_usage_data_wrong,
1264 copy_fs_field(fs_usage_cached_wrong,
1265 b.cached, "cached");
1266 copy_fs_field(fs_usage_reserved_wrong,
1267 b.reserved, "reserved");
1268 copy_fs_field(fs_usage_nr_inodes_wrong,
1269 b.nr_inodes,"nr_inodes");
1271 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1272 copy_fs_field(fs_usage_persistent_reserved_wrong,
1273 persistent_reserved[i],
1274 "persistent_reserved[%i]", i);
1277 for (i = 0; i < c->replicas.nr; i++) {
1278 struct bch_replicas_entry_v1 *e =
1279 cpu_replicas_entry(&c->replicas, i);
1281 if (metadata_only &&
1282 (e->data_type == BCH_DATA_user ||
1283 e->data_type == BCH_DATA_cached))
1286 printbuf_reset(&buf);
1287 bch2_replicas_entry_to_text(&buf, e);
1289 copy_fs_field(fs_usage_replicas_wrong,
1290 replicas[i], "%s", buf.buf);
1294 #undef copy_fs_field
1295 #undef copy_dev_field
1296 #undef copy_stripe_field
1300 percpu_ref_put(&ca->ref);
1303 percpu_up_write(&c->mark_lock);
1304 printbuf_exit(&buf);
1308 static int bch2_gc_start(struct bch_fs *c)
1310 BUG_ON(c->usage_gc);
1312 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1313 sizeof(u64), GFP_KERNEL);
1315 bch_err(c, "error allocating c->usage_gc");
1316 return -BCH_ERR_ENOMEM_gc_start;
1319 for_each_member_device(c, ca) {
1320 BUG_ON(ca->usage_gc);
1322 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1323 if (!ca->usage_gc) {
1324 bch_err(c, "error allocating ca->usage_gc");
1325 percpu_ref_put(&ca->ref);
1326 return -BCH_ERR_ENOMEM_gc_start;
1329 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1330 ca->mi.nbuckets - ca->mi.first_bucket);
1336 static int bch2_gc_reset(struct bch_fs *c)
1338 for_each_member_device(c, ca) {
1339 free_percpu(ca->usage_gc);
1340 ca->usage_gc = NULL;
1343 free_percpu(c->usage_gc);
1346 return bch2_gc_start(c);
1349 /* returns true if not equal */
1350 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1351 struct bch_alloc_v4 r)
1353 return l.gen != r.gen ||
1354 l.oldest_gen != r.oldest_gen ||
1355 l.data_type != r.data_type ||
1356 l.dirty_sectors != r.dirty_sectors ||
1357 l.cached_sectors != r.cached_sectors ||
1358 l.stripe_redundancy != r.stripe_redundancy ||
1359 l.stripe != r.stripe;
1362 static int bch2_alloc_write_key(struct btree_trans *trans,
1363 struct btree_iter *iter,
1367 struct bch_fs *c = trans->c;
1368 struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1369 struct bucket old_gc, gc, *b;
1370 struct bkey_i_alloc_v4 *a;
1371 struct bch_alloc_v4 old_convert, new;
1372 const struct bch_alloc_v4 *old;
1375 old = bch2_alloc_to_v4(k, &old_convert);
1378 percpu_down_read(&c->mark_lock);
1379 b = gc_bucket(ca, iter->pos.offset);
1382 if ((old->data_type == BCH_DATA_sb ||
1383 old->data_type == BCH_DATA_journal) &&
1384 !bch2_dev_is_online(ca)) {
1385 b->data_type = old->data_type;
1386 b->dirty_sectors = old->dirty_sectors;
1390 * b->data_type doesn't yet include need_discard & need_gc_gen states -
1393 b->data_type = __alloc_data_type(b->dirty_sectors,
1400 if (gc.data_type != old_gc.data_type ||
1401 gc.dirty_sectors != old_gc.dirty_sectors)
1402 bch2_dev_usage_update_m(c, ca, &old_gc, &gc);
1403 percpu_up_read(&c->mark_lock);
1405 if (metadata_only &&
1406 gc.data_type != BCH_DATA_sb &&
1407 gc.data_type != BCH_DATA_journal &&
1408 gc.data_type != BCH_DATA_btree)
1411 if (gen_after(old->gen, gc.gen))
1414 if (fsck_err_on(new.data_type != gc.data_type, c,
1415 alloc_key_data_type_wrong,
1416 "bucket %llu:%llu gen %u has wrong data_type"
1417 ": got %s, should be %s",
1418 iter->pos.inode, iter->pos.offset,
1420 bch2_data_type_str(new.data_type),
1421 bch2_data_type_str(gc.data_type)))
1422 new.data_type = gc.data_type;
1424 #define copy_bucket_field(_errtype, _f) \
1425 if (fsck_err_on(new._f != gc._f, c, _errtype, \
1426 "bucket %llu:%llu gen %u data type %s has wrong " #_f \
1427 ": got %u, should be %u", \
1428 iter->pos.inode, iter->pos.offset, \
1430 bch2_data_type_str(gc.data_type), \
1434 copy_bucket_field(alloc_key_gen_wrong,
1436 copy_bucket_field(alloc_key_dirty_sectors_wrong,
1438 copy_bucket_field(alloc_key_cached_sectors_wrong,
1440 copy_bucket_field(alloc_key_stripe_wrong,
1442 copy_bucket_field(alloc_key_stripe_redundancy_wrong,
1444 #undef copy_bucket_field
1446 if (!bch2_alloc_v4_cmp(*old, new))
1449 a = bch2_alloc_to_v4_mut(trans, k);
1450 ret = PTR_ERR_OR_ZERO(a);
1457 * The trigger normally makes sure this is set, but we're not running
1460 if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1461 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1463 ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1468 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1472 for_each_member_device(c, ca) {
1473 ret = bch2_trans_run(c,
1474 for_each_btree_key_upto_commit(trans, iter, BTREE_ID_alloc,
1475 POS(ca->dev_idx, ca->mi.first_bucket),
1476 POS(ca->dev_idx, ca->mi.nbuckets - 1),
1477 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1478 NULL, NULL, BCH_TRANS_COMMIT_lazy_rw,
1479 bch2_alloc_write_key(trans, &iter, k, metadata_only)));
1481 percpu_ref_put(&ca->ref);
1490 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1492 for_each_member_device(c, ca) {
1493 struct bucket_array *buckets = kvmalloc(sizeof(struct bucket_array) +
1494 ca->mi.nbuckets * sizeof(struct bucket),
1495 GFP_KERNEL|__GFP_ZERO);
1497 percpu_ref_put(&ca->ref);
1498 bch_err(c, "error allocating ca->buckets[gc]");
1499 return -BCH_ERR_ENOMEM_gc_alloc_start;
1502 buckets->first_bucket = ca->mi.first_bucket;
1503 buckets->nbuckets = ca->mi.nbuckets;
1504 rcu_assign_pointer(ca->buckets_gc, buckets);
1507 int ret = bch2_trans_run(c,
1508 for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
1509 BTREE_ITER_PREFETCH, k, ({
1510 struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
1511 struct bucket *g = gc_bucket(ca, k.k->p.offset);
1513 struct bch_alloc_v4 a_convert;
1514 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1519 if (metadata_only &&
1520 (a->data_type == BCH_DATA_user ||
1521 a->data_type == BCH_DATA_cached ||
1522 a->data_type == BCH_DATA_parity)) {
1523 g->data_type = a->data_type;
1524 g->dirty_sectors = a->dirty_sectors;
1525 g->cached_sectors = a->cached_sectors;
1526 g->stripe = a->stripe;
1527 g->stripe_redundancy = a->stripe_redundancy;
1536 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1538 for_each_member_device(c, ca) {
1539 struct bucket_array *buckets = gc_bucket_array(ca);
1542 for_each_bucket(g, buckets) {
1543 if (metadata_only &&
1544 (g->data_type == BCH_DATA_user ||
1545 g->data_type == BCH_DATA_cached ||
1546 g->data_type == BCH_DATA_parity))
1549 g->dirty_sectors = 0;
1550 g->cached_sectors = 0;
1555 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1556 struct btree_iter *iter,
1560 struct bch_fs *c = trans->c;
1561 const __le64 *refcount = bkey_refcount_c(k);
1562 struct printbuf buf = PRINTBUF;
1563 struct reflink_gc *r;
1569 while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1570 r->offset < k.k->p.offset)
1574 r->offset != k.k->p.offset ||
1575 r->size != k.k->size) {
1576 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1580 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1581 reflink_v_refcount_wrong,
1582 "reflink key has wrong refcount:\n"
1585 (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1587 struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
1588 ret = PTR_ERR_OR_ZERO(new);
1593 new->k.type = KEY_TYPE_deleted;
1595 *bkey_refcount(bkey_i_to_s(new)) = cpu_to_le64(r->refcount);
1596 ret = bch2_trans_update(trans, iter, new, 0);
1600 printbuf_exit(&buf);
1604 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1611 int ret = bch2_trans_run(c,
1612 for_each_btree_key_commit(trans, iter,
1613 BTREE_ID_reflink, POS_MIN,
1614 BTREE_ITER_PREFETCH, k,
1615 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1616 bch2_gc_write_reflink_key(trans, &iter, k, &idx)));
1617 c->reflink_gc_nr = 0;
1621 static int bch2_gc_reflink_start(struct bch_fs *c,
1628 c->reflink_gc_nr = 0;
1630 int ret = bch2_trans_run(c,
1631 for_each_btree_key(trans, iter, BTREE_ID_reflink, POS_MIN,
1632 BTREE_ITER_PREFETCH, k, ({
1633 const __le64 *refcount = bkey_refcount_c(k);
1638 struct reflink_gc *r = genradix_ptr_alloc(&c->reflink_gc_table,
1639 c->reflink_gc_nr++, GFP_KERNEL);
1641 ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1645 r->offset = k.k->p.offset;
1646 r->size = k.k->size;
1655 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1657 struct genradix_iter iter;
1658 struct reflink_gc *r;
1660 genradix_for_each(&c->reflink_gc_table, iter, r)
1664 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1665 struct btree_iter *iter,
1668 struct bch_fs *c = trans->c;
1669 struct printbuf buf = PRINTBUF;
1670 const struct bch_stripe *s;
1671 struct gc_stripe *m;
1676 if (k.k->type != KEY_TYPE_stripe)
1679 s = bkey_s_c_to_stripe(k).v;
1680 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1682 for (i = 0; i < s->nr_blocks; i++) {
1683 u32 old = stripe_blockcount_get(s, i);
1684 u32 new = (m ? m->block_sectors[i] : 0);
1687 prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1694 bch2_bkey_val_to_text(&buf, c, k);
1696 if (fsck_err_on(bad, c, stripe_sector_count_wrong,
1698 struct bkey_i_stripe *new;
1700 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1701 ret = PTR_ERR_OR_ZERO(new);
1705 bkey_reassemble(&new->k_i, k);
1707 for (i = 0; i < new->v.nr_blocks; i++)
1708 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1710 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1713 printbuf_exit(&buf);
1717 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1722 return bch2_trans_run(c,
1723 for_each_btree_key_commit(trans, iter,
1724 BTREE_ID_stripes, POS_MIN,
1725 BTREE_ITER_PREFETCH, k,
1726 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1727 bch2_gc_write_stripes_key(trans, &iter, k)));
1730 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1732 genradix_free(&c->gc_stripes);
1736 * bch2_gc - walk _all_ references to buckets, and recompute them:
1738 * @c: filesystem object
1739 * @initial: are we in recovery?
1740 * @metadata_only: are we just checking metadata references, or everything?
1742 * Returns: 0 on success, or standard errcode on failure
1744 * Order matters here:
1745 * - Concurrent GC relies on the fact that we have a total ordering for
1746 * everything that GC walks - see gc_will_visit_node(),
1747 * gc_will_visit_root()
1749 * - also, references move around in the course of index updates and
1750 * various other crap: everything needs to agree on the ordering
1751 * references are allowed to move around in - e.g., we're allowed to
1752 * start with a reference owned by an open_bucket (the allocator) and
1753 * move it to the btree, but not the reverse.
1755 * This is necessary to ensure that gc doesn't miss references that
1756 * move around - if references move backwards in the ordering GC
1757 * uses, GC could skip past them
1759 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1764 lockdep_assert_held(&c->state_lock);
1766 down_write(&c->gc_lock);
1768 bch2_btree_interior_updates_flush(c);
1770 ret = bch2_gc_start(c) ?:
1771 bch2_gc_alloc_start(c, metadata_only) ?:
1772 bch2_gc_reflink_start(c, metadata_only);
1776 gc_pos_set(c, gc_phase(GC_PHASE_START));
1778 bch2_mark_superblocks(c);
1780 ret = bch2_gc_btrees(c, initial, metadata_only);
1786 bch2_mark_pending_btree_node_frees(c);
1790 if (test_bit(BCH_FS_need_another_gc, &c->flags) ||
1791 (!iter && bch2_test_restart_gc)) {
1793 bch_info(c, "Unable to fix bucket gens, looping");
1799 * XXX: make sure gens we fixed got saved
1801 bch_info(c, "Second GC pass needed, restarting:");
1802 clear_bit(BCH_FS_need_another_gc, &c->flags);
1803 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1805 bch2_gc_stripes_reset(c, metadata_only);
1806 bch2_gc_alloc_reset(c, metadata_only);
1807 bch2_gc_reflink_reset(c, metadata_only);
1808 ret = bch2_gc_reset(c);
1812 /* flush fsck errors, reset counters */
1813 bch2_flush_fsck_errs(c);
1818 bch2_journal_block(&c->journal);
1820 ret = bch2_gc_alloc_done(c, metadata_only) ?:
1821 bch2_gc_done(c, initial, metadata_only) ?:
1822 bch2_gc_stripes_done(c, metadata_only) ?:
1823 bch2_gc_reflink_done(c, metadata_only);
1825 bch2_journal_unblock(&c->journal);
1828 percpu_down_write(&c->mark_lock);
1829 /* Indicates that gc is no longer in progress: */
1830 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1833 percpu_up_write(&c->mark_lock);
1835 up_write(&c->gc_lock);
1838 * At startup, allocations can happen directly instead of via the
1839 * allocator thread - issue wakeup in case they blocked on gc_lock:
1841 closure_wake_up(&c->freelist_wait);
1846 static int gc_btree_gens_key(struct btree_trans *trans,
1847 struct btree_iter *iter,
1850 struct bch_fs *c = trans->c;
1851 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1855 percpu_down_read(&c->mark_lock);
1856 bkey_for_each_ptr(ptrs, ptr) {
1857 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1859 if (ptr_stale(ca, ptr) > 16) {
1860 percpu_up_read(&c->mark_lock);
1865 bkey_for_each_ptr(ptrs, ptr) {
1866 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1867 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1869 if (gen_after(*gen, ptr->gen))
1872 percpu_up_read(&c->mark_lock);
1875 u = bch2_bkey_make_mut(trans, iter, &k, 0);
1876 ret = PTR_ERR_OR_ZERO(u);
1880 bch2_extent_normalize(c, bkey_i_to_s(u));
1884 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1887 struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1888 struct bch_alloc_v4 a_convert;
1889 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1890 struct bkey_i_alloc_v4 *a_mut;
1893 if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1896 a_mut = bch2_alloc_to_v4_mut(trans, k);
1897 ret = PTR_ERR_OR_ZERO(a_mut);
1901 a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1902 a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1904 return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1907 int bch2_gc_gens(struct bch_fs *c)
1909 u64 b, start_time = local_clock();
1913 * Ideally we would be using state_lock and not gc_lock here, but that
1914 * introduces a deadlock in the RO path - we currently take the state
1915 * lock at the start of going RO, thus the gc thread may get stuck:
1917 if (!mutex_trylock(&c->gc_gens_lock))
1920 trace_and_count(c, gc_gens_start, c);
1921 down_read(&c->gc_lock);
1923 for_each_member_device(c, ca) {
1924 struct bucket_gens *gens = bucket_gens(ca);
1926 BUG_ON(ca->oldest_gen);
1928 ca->oldest_gen = kvmalloc(gens->nbuckets, GFP_KERNEL);
1929 if (!ca->oldest_gen) {
1930 percpu_ref_put(&ca->ref);
1931 ret = -BCH_ERR_ENOMEM_gc_gens;
1935 for (b = gens->first_bucket;
1936 b < gens->nbuckets; b++)
1937 ca->oldest_gen[b] = gens->b[b];
1940 for (unsigned i = 0; i < BTREE_ID_NR; i++)
1941 if (btree_type_has_ptrs(i)) {
1942 c->gc_gens_btree = i;
1943 c->gc_gens_pos = POS_MIN;
1945 ret = bch2_trans_run(c,
1946 for_each_btree_key_commit(trans, iter, i,
1948 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
1951 BCH_TRANS_COMMIT_no_enospc,
1952 gc_btree_gens_key(trans, &iter, k)));
1957 ret = bch2_trans_run(c,
1958 for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1960 BTREE_ITER_PREFETCH,
1963 BCH_TRANS_COMMIT_no_enospc,
1964 bch2_alloc_write_oldest_gen(trans, &iter, k)));
1968 c->gc_gens_btree = 0;
1969 c->gc_gens_pos = POS_MIN;
1973 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1974 trace_and_count(c, gc_gens_end, c);
1976 for_each_member_device(c, ca) {
1977 kvfree(ca->oldest_gen);
1978 ca->oldest_gen = NULL;
1981 up_read(&c->gc_lock);
1982 mutex_unlock(&c->gc_gens_lock);
1983 if (!bch2_err_matches(ret, EROFS))
1988 static int bch2_gc_thread(void *arg)
1990 struct bch_fs *c = arg;
1991 struct io_clock *clock = &c->io_clock[WRITE];
1992 unsigned long last = atomic64_read(&clock->now);
1993 unsigned last_kick = atomic_read(&c->kick_gc);
1999 set_current_state(TASK_INTERRUPTIBLE);
2001 if (kthread_should_stop()) {
2002 __set_current_state(TASK_RUNNING);
2006 if (atomic_read(&c->kick_gc) != last_kick)
2009 if (c->btree_gc_periodic) {
2010 unsigned long next = last + c->capacity / 16;
2012 if (atomic64_read(&clock->now) >= next)
2015 bch2_io_clock_schedule_timeout(clock, next);
2022 __set_current_state(TASK_RUNNING);
2024 last = atomic64_read(&clock->now);
2025 last_kick = atomic_read(&c->kick_gc);
2028 * Full gc is currently incompatible with btree key cache:
2031 ret = bch2_gc(c, false, false);
2035 debug_check_no_locks_held();
2041 void bch2_gc_thread_stop(struct bch_fs *c)
2043 struct task_struct *p;
2046 c->gc_thread = NULL;
2054 int bch2_gc_thread_start(struct bch_fs *c)
2056 struct task_struct *p;
2061 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2063 bch_err_fn(c, PTR_ERR(p));