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_key_cache.h"
13 #include "btree_locking.h"
14 #include "btree_update_interior.h"
32 #include <linux/slab.h>
33 #include <linux/bitops.h>
34 #include <linux/freezer.h>
35 #include <linux/kthread.h>
36 #include <linux/preempt.h>
37 #include <linux/rcupdate.h>
38 #include <linux/sched/task.h>
40 #define DROP_THIS_NODE 10
41 #define DROP_PREV_NODE 11
43 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
46 write_seqcount_begin(&c->gc_pos_lock);
48 write_seqcount_end(&c->gc_pos_lock);
52 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
54 BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
55 __gc_pos_set(c, new_pos);
59 * Missing: if an interior btree node is empty, we need to do something -
60 * perhaps just kill it
62 static int bch2_gc_check_topology(struct bch_fs *c,
64 struct bkey_buf *prev,
68 struct bpos node_start = b->data->min_key;
69 struct bpos node_end = b->data->max_key;
70 struct bpos expected_start = bkey_deleted(&prev->k->k)
72 : bpos_successor(prev->k->k.p);
73 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
76 if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
77 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
79 if (!bpos_eq(expected_start, bp->v.min_key)) {
80 bch2_topology_error(c);
82 if (bkey_deleted(&prev->k->k)) {
83 prt_printf(&buf1, "start of node: ");
84 bch2_bpos_to_text(&buf1, node_start);
86 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
88 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
94 "btree node with incorrect min_key at btree %s level %u:\n"
97 bch2_btree_ids[b->c.btree_id], b->c.level,
98 buf1.buf, buf2.buf) &&
99 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
100 bch_info(c, "Halting mark and sweep to start topology repair pass");
101 ret = -BCH_ERR_need_topology_repair;
104 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
109 if (is_last && !bpos_eq(cur.k->k.p, node_end)) {
110 bch2_topology_error(c);
112 printbuf_reset(&buf1);
113 printbuf_reset(&buf2);
115 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
116 bch2_bpos_to_text(&buf2, node_end);
122 "btree node with incorrect max_key at btree %s level %u:\n"
125 bch2_btree_ids[b->c.btree_id], b->c.level,
126 buf1.buf, buf2.buf) &&
127 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
128 bch_info(c, "Halting mark and sweep to start topology repair pass");
129 ret = -BCH_ERR_need_topology_repair;
132 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
136 bch2_bkey_buf_copy(prev, c, cur.k);
139 printbuf_exit(&buf2);
140 printbuf_exit(&buf1);
144 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
146 switch (b->key.k.type) {
147 case KEY_TYPE_btree_ptr: {
148 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
152 dst->v.seq = b->data->keys.seq;
153 dst->v.sectors_written = 0;
155 dst->v.min_key = b->data->min_key;
156 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
157 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
160 case KEY_TYPE_btree_ptr_v2:
161 bkey_copy(&dst->k_i, &b->key);
168 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
169 enum btree_id btree, unsigned level,
170 struct bkey_s_c old, struct bkey_i *new)
172 struct bch_fs *c = trans->c;
177 bch2_bkey_buf_init(&tmp);
178 bch2_bkey_buf_reassemble(&tmp, c, old);
180 b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
181 if (!IS_ERR_OR_NULL(b)) {
182 mutex_lock(&c->btree_cache.lock);
184 bch2_btree_node_hash_remove(&c->btree_cache, b);
186 bkey_copy(&b->key, new);
187 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
190 mutex_unlock(&c->btree_cache.lock);
191 six_unlock_read(&b->c.lock);
194 bch2_bkey_buf_exit(&tmp, c);
197 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
199 struct bkey_i_btree_ptr_v2 *new;
202 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
204 return -BCH_ERR_ENOMEM_gc_repair_key;
206 btree_ptr_to_v2(b, new);
207 b->data->min_key = new_min;
208 new->v.min_key = new_min;
209 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
211 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
217 bch2_btree_node_drop_keys_outside_node(b);
218 bkey_copy(&b->key, &new->k_i);
222 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
224 struct bkey_i_btree_ptr_v2 *new;
227 ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
231 new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
233 return -BCH_ERR_ENOMEM_gc_repair_key;
235 btree_ptr_to_v2(b, new);
236 b->data->max_key = new_max;
238 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
240 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
246 bch2_btree_node_drop_keys_outside_node(b);
248 mutex_lock(&c->btree_cache.lock);
249 bch2_btree_node_hash_remove(&c->btree_cache, b);
251 bkey_copy(&b->key, &new->k_i);
252 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
254 mutex_unlock(&c->btree_cache.lock);
258 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
259 struct btree *prev, struct btree *cur)
261 struct bpos expected_start = !prev
263 : bpos_successor(prev->key.k.p);
264 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
268 prt_printf(&buf1, "start of node: ");
269 bch2_bpos_to_text(&buf1, b->data->min_key);
271 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
274 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
277 bpos_gt(expected_start, cur->data->min_key) &&
278 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
279 /* cur overwrites prev: */
281 if (mustfix_fsck_err_on(bpos_ge(prev->data->min_key,
282 cur->data->min_key), c,
283 "btree node overwritten by next node at btree %s level %u:\n"
286 bch2_btree_ids[b->c.btree_id], b->c.level,
287 buf1.buf, buf2.buf)) {
288 ret = DROP_PREV_NODE;
292 if (mustfix_fsck_err_on(!bpos_eq(prev->key.k.p,
293 bpos_predecessor(cur->data->min_key)), c,
294 "btree node with incorrect max_key at btree %s level %u:\n"
297 bch2_btree_ids[b->c.btree_id], b->c.level,
299 ret = set_node_max(c, prev,
300 bpos_predecessor(cur->data->min_key));
302 /* prev overwrites cur: */
304 if (mustfix_fsck_err_on(bpos_ge(expected_start,
305 cur->data->max_key), c,
306 "btree node overwritten by prev node at btree %s level %u:\n"
309 bch2_btree_ids[b->c.btree_id], b->c.level,
310 buf1.buf, buf2.buf)) {
311 ret = DROP_THIS_NODE;
315 if (mustfix_fsck_err_on(!bpos_eq(expected_start, cur->data->min_key), c,
316 "btree node with incorrect min_key at btree %s level %u:\n"
319 bch2_btree_ids[b->c.btree_id], b->c.level,
321 ret = set_node_min(c, cur, expected_start);
325 printbuf_exit(&buf2);
326 printbuf_exit(&buf1);
330 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
333 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
336 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
337 bch2_bpos_to_text(&buf2, b->key.k.p);
339 if (mustfix_fsck_err_on(!bpos_eq(child->key.k.p, b->key.k.p), c,
340 "btree node with incorrect max_key at btree %s level %u:\n"
343 bch2_btree_ids[b->c.btree_id], b->c.level,
344 buf1.buf, buf2.buf)) {
345 ret = set_node_max(c, child, b->key.k.p);
351 printbuf_exit(&buf2);
352 printbuf_exit(&buf1);
356 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b)
358 struct bch_fs *c = trans->c;
359 struct btree_and_journal_iter iter;
361 struct bkey_buf prev_k, cur_k;
362 struct btree *prev = NULL, *cur = NULL;
363 bool have_child, dropped_children = false;
364 struct printbuf buf = PRINTBUF;
371 have_child = dropped_children = false;
372 bch2_bkey_buf_init(&prev_k);
373 bch2_bkey_buf_init(&cur_k);
374 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
376 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
377 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
378 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
380 bch2_btree_and_journal_iter_advance(&iter);
381 bch2_bkey_buf_reassemble(&cur_k, c, k);
383 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
384 b->c.btree_id, b->c.level - 1,
386 ret = PTR_ERR_OR_ZERO(cur);
388 printbuf_reset(&buf);
389 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
391 if (mustfix_fsck_err_on(ret == -EIO, c,
392 "Topology repair: unreadable btree node at btree %s level %u:\n"
394 bch2_btree_ids[b->c.btree_id],
397 bch2_btree_node_evict(trans, cur_k.k);
398 ret = bch2_journal_key_delete(c, b->c.btree_id,
399 b->c.level, cur_k.k->k.p);
407 bch_err(c, "%s: error getting btree node: %s",
408 __func__, bch2_err_str(ret));
412 ret = btree_repair_node_boundaries(c, b, prev, cur);
414 if (ret == DROP_THIS_NODE) {
415 six_unlock_read(&cur->c.lock);
416 bch2_btree_node_evict(trans, cur_k.k);
417 ret = bch2_journal_key_delete(c, b->c.btree_id,
418 b->c.level, cur_k.k->k.p);
426 six_unlock_read(&prev->c.lock);
429 if (ret == DROP_PREV_NODE) {
430 bch2_btree_node_evict(trans, prev_k.k);
431 ret = bch2_journal_key_delete(c, b->c.btree_id,
432 b->c.level, prev_k.k->k.p);
436 bch2_btree_and_journal_iter_exit(&iter);
437 bch2_bkey_buf_exit(&prev_k, c);
438 bch2_bkey_buf_exit(&cur_k, c);
445 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
448 if (!ret && !IS_ERR_OR_NULL(prev)) {
450 ret = btree_repair_node_end(c, b, prev);
453 if (!IS_ERR_OR_NULL(prev))
454 six_unlock_read(&prev->c.lock);
456 if (!IS_ERR_OR_NULL(cur))
457 six_unlock_read(&cur->c.lock);
463 bch2_btree_and_journal_iter_exit(&iter);
464 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
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);
476 bch_err(c, "%s: error getting btree node: %s",
477 __func__, bch2_err_str(ret));
481 ret = bch2_btree_repair_topology_recurse(trans, cur);
482 six_unlock_read(&cur->c.lock);
485 if (ret == DROP_THIS_NODE) {
486 bch2_btree_node_evict(trans, cur_k.k);
487 ret = bch2_journal_key_delete(c, b->c.btree_id,
488 b->c.level, cur_k.k->k.p);
489 dropped_children = true;
498 printbuf_reset(&buf);
499 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
501 if (mustfix_fsck_err_on(!have_child, c,
502 "empty interior btree node at btree %s level %u\n"
504 bch2_btree_ids[b->c.btree_id],
505 b->c.level, buf.buf))
506 ret = DROP_THIS_NODE;
509 if (!IS_ERR_OR_NULL(prev))
510 six_unlock_read(&prev->c.lock);
511 if (!IS_ERR_OR_NULL(cur))
512 six_unlock_read(&cur->c.lock);
514 bch2_btree_and_journal_iter_exit(&iter);
515 bch2_bkey_buf_exit(&prev_k, c);
516 bch2_bkey_buf_exit(&cur_k, c);
518 if (!ret && dropped_children)
525 static int bch2_repair_topology(struct bch_fs *c)
527 struct btree_trans trans;
532 bch2_trans_init(&trans, c, 0, 0);
534 for (i = 0; i < BTREE_ID_NR && !ret; i++) {
535 b = c->btree_roots[i].b;
536 if (btree_node_fake(b))
539 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
540 ret = bch2_btree_repair_topology_recurse(&trans, b);
541 six_unlock_read(&b->c.lock);
543 if (ret == DROP_THIS_NODE) {
544 bch_err(c, "empty btree root - repair unimplemented");
545 ret = -BCH_ERR_fsck_repair_unimplemented;
549 bch2_trans_exit(&trans);
554 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
555 unsigned level, bool is_root,
558 struct bch_fs *c = trans->c;
559 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
560 const union bch_extent_entry *entry;
561 struct extent_ptr_decoded p = { 0 };
562 bool do_update = false;
563 struct printbuf buf = PRINTBUF;
568 * use check_bucket_ref here
570 bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
571 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
572 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
573 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
576 (c->opts.reconstruct_alloc ||
577 fsck_err(c, "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
579 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
580 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
582 (printbuf_reset(&buf),
583 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
592 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
593 (c->opts.reconstruct_alloc ||
594 fsck_err(c, "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
596 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
597 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
599 (printbuf_reset(&buf),
600 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
605 g->dirty_sectors = 0;
606 g->cached_sectors = 0;
607 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
613 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
614 (c->opts.reconstruct_alloc ||
615 fsck_err(c, "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
617 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
618 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
620 (printbuf_reset(&buf),
621 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
624 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
625 (c->opts.reconstruct_alloc ||
626 fsck_err(c, "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
628 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
629 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
631 (printbuf_reset(&buf),
632 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
635 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
638 if (fsck_err_on(bucket_data_type(g->data_type) &&
639 bucket_data_type(g->data_type) != data_type, c,
640 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
642 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
643 bch2_data_types[g->data_type],
644 bch2_data_types[data_type],
645 (printbuf_reset(&buf),
646 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
647 if (data_type == BCH_DATA_btree) {
648 g->data_type = data_type;
649 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
656 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
658 if (fsck_err_on(!m || !m->alive, c,
659 "pointer to nonexistent stripe %llu\n"
662 (printbuf_reset(&buf),
663 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
666 if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
667 "pointer does not match stripe %llu\n"
670 (printbuf_reset(&buf),
671 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
677 struct bkey_ptrs ptrs;
678 union bch_extent_entry *entry;
679 struct bch_extent_ptr *ptr;
683 bch_err(c, "cannot update btree roots yet");
688 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
690 bch_err(c, "%s: error allocating new key", __func__);
691 ret = -BCH_ERR_ENOMEM_gc_repair_key;
695 bkey_reassemble(new, *k);
699 * We don't want to drop btree node pointers - if the
700 * btree node isn't there anymore, the read path will
703 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
704 bkey_for_each_ptr(ptrs, ptr) {
705 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
706 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
711 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
712 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
713 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
714 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
717 (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
719 gen_cmp(ptr->gen, g->gen) < 0) ||
720 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
722 g->data_type != data_type);
725 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
726 bkey_extent_entry_for_each(ptrs, entry) {
727 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
728 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
729 entry->stripe_ptr.idx);
730 union bch_extent_entry *next_ptr;
732 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
733 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
738 bch_err(c, "aieee, found stripe ptr with no data ptr");
742 if (!m || !m->alive ||
743 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
746 bch2_bkey_extent_entry_drop(new, entry);
753 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
760 bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
763 printbuf_reset(&buf);
764 bch2_bkey_val_to_text(&buf, c, *k);
765 bch_info(c, "updated %s", buf.buf);
767 printbuf_reset(&buf);
768 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
769 bch_info(c, "new key %s", buf.buf);
772 *k = bkey_i_to_s_c(new);
780 /* marking of btree keys/nodes: */
782 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
783 unsigned level, bool is_root,
787 struct bch_fs *c = trans->c;
788 struct bkey deleted = KEY(0, 0, 0);
789 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
792 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
798 BUG_ON(bch2_journal_seq_verify &&
799 k->k->version.lo > atomic64_read(&c->journal.seq));
801 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
805 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
806 "key version number higher than recorded: %llu > %llu",
808 atomic64_read(&c->key_version)))
809 atomic64_set(&c->key_version, k->k->version.lo);
812 ret = commit_do(trans, NULL, NULL, 0,
813 bch2_mark_key(trans, btree_id, level, old, *k, flags));
817 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
821 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
823 struct bch_fs *c = trans->c;
824 struct btree_node_iter iter;
825 struct bkey unpacked;
827 struct bkey_buf prev, cur;
830 if (!btree_node_type_needs_gc(btree_node_type(b)))
833 bch2_btree_node_iter_init_from_start(&iter, b);
834 bch2_bkey_buf_init(&prev);
835 bch2_bkey_buf_init(&cur);
836 bkey_init(&prev.k->k);
838 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
839 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
844 bch2_btree_node_iter_advance(&iter, b);
847 bch2_bkey_buf_reassemble(&cur, c, k);
849 ret = bch2_gc_check_topology(c, b, &prev, cur,
850 bch2_btree_node_iter_end(&iter));
856 bch2_bkey_buf_exit(&cur, c);
857 bch2_bkey_buf_exit(&prev, c);
861 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
862 bool initial, bool metadata_only)
864 struct bch_fs *c = trans->c;
865 struct btree_iter iter;
867 unsigned depth = metadata_only ? 1 : 0;
870 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
872 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
873 0, depth, BTREE_ITER_PREFETCH, b, ret) {
874 bch2_verify_btree_nr_keys(b);
876 gc_pos_set(c, gc_pos_btree_node(b));
878 ret = btree_gc_mark_node(trans, b, initial);
882 bch2_trans_iter_exit(trans, &iter);
887 mutex_lock(&c->btree_root_lock);
888 b = c->btree_roots[btree_id].b;
889 if (!btree_node_fake(b)) {
890 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
892 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
895 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
896 mutex_unlock(&c->btree_root_lock);
901 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
902 unsigned target_depth)
904 struct bch_fs *c = trans->c;
905 struct btree_and_journal_iter iter;
907 struct bkey_buf cur, prev;
908 struct printbuf buf = PRINTBUF;
911 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
912 bch2_bkey_buf_init(&prev);
913 bch2_bkey_buf_init(&cur);
914 bkey_init(&prev.k->k);
916 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
917 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
918 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
920 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
923 bch_err(c, "%s: error from bch2_gc_mark_key: %s",
924 __func__, bch2_err_str(ret));
929 bch2_bkey_buf_reassemble(&cur, c, k);
930 k = bkey_i_to_s_c(cur.k);
932 bch2_btree_and_journal_iter_advance(&iter);
934 ret = bch2_gc_check_topology(c, b,
936 !bch2_btree_and_journal_iter_peek(&iter).k);
940 bch2_btree_and_journal_iter_advance(&iter);
944 if (b->c.level > target_depth) {
945 bch2_btree_and_journal_iter_exit(&iter);
946 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
948 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
951 bch2_bkey_buf_reassemble(&cur, c, k);
952 bch2_btree_and_journal_iter_advance(&iter);
954 child = bch2_btree_node_get_noiter(trans, cur.k,
955 b->c.btree_id, b->c.level - 1,
957 ret = PTR_ERR_OR_ZERO(child);
960 bch2_topology_error(c);
966 "Unreadable btree node at btree %s level %u:\n"
968 bch2_btree_ids[b->c.btree_id],
970 (printbuf_reset(&buf),
971 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
972 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
973 ret = -BCH_ERR_need_topology_repair;
974 bch_info(c, "Halting mark and sweep to start topology repair pass");
977 /* Continue marking when opted to not
980 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
984 bch_err(c, "%s: error getting btree node: %s",
985 __func__, bch2_err_str(ret));
989 ret = bch2_gc_btree_init_recurse(trans, child,
991 six_unlock_read(&child->c.lock);
998 bch2_bkey_buf_exit(&cur, c);
999 bch2_bkey_buf_exit(&prev, c);
1000 bch2_btree_and_journal_iter_exit(&iter);
1001 printbuf_exit(&buf);
1005 static int bch2_gc_btree_init(struct btree_trans *trans,
1006 enum btree_id btree_id,
1009 struct bch_fs *c = trans->c;
1011 unsigned target_depth = metadata_only ? 1 : 0;
1012 struct printbuf buf = PRINTBUF;
1015 b = c->btree_roots[btree_id].b;
1017 if (btree_node_fake(b))
1020 six_lock_read(&b->c.lock, NULL, NULL);
1021 printbuf_reset(&buf);
1022 bch2_bpos_to_text(&buf, b->data->min_key);
1023 if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1024 "btree root with incorrect min_key: %s", buf.buf)) {
1025 bch_err(c, "repair unimplemented");
1026 ret = -BCH_ERR_fsck_repair_unimplemented;
1030 printbuf_reset(&buf);
1031 bch2_bpos_to_text(&buf, b->data->max_key);
1032 if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1033 "btree root with incorrect max_key: %s", buf.buf)) {
1034 bch_err(c, "repair unimplemented");
1035 ret = -BCH_ERR_fsck_repair_unimplemented;
1039 if (b->c.level >= target_depth)
1040 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1043 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1045 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1049 six_unlock_read(&b->c.lock);
1052 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1053 printbuf_exit(&buf);
1057 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1059 return (int) btree_id_to_gc_phase(l) -
1060 (int) btree_id_to_gc_phase(r);
1063 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1065 struct btree_trans trans;
1066 enum btree_id ids[BTREE_ID_NR];
1070 bch2_trans_init(&trans, c, 0, 0);
1073 trans.is_initial_gc = true;
1075 for (i = 0; i < BTREE_ID_NR; i++)
1077 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1079 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1081 ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1082 : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1085 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1087 bch2_trans_exit(&trans);
1091 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1093 enum bch_data_type type,
1096 u64 b = sector_to_bucket(ca, start);
1100 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1102 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1103 gc_phase(GC_PHASE_SB), flags);
1106 } while (start < end);
1109 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1112 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1116 for (i = 0; i < layout->nr_superblocks; i++) {
1117 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1119 if (offset == BCH_SB_SECTOR)
1120 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1121 BCH_DATA_sb, flags);
1123 mark_metadata_sectors(c, ca, offset,
1124 offset + (1 << layout->sb_max_size_bits),
1125 BCH_DATA_sb, flags);
1128 for (i = 0; i < ca->journal.nr; i++) {
1129 b = ca->journal.buckets[i];
1130 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1132 gc_phase(GC_PHASE_SB), flags);
1136 static void bch2_mark_superblocks(struct bch_fs *c)
1141 mutex_lock(&c->sb_lock);
1142 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1144 for_each_online_member(ca, c, i)
1145 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1146 mutex_unlock(&c->sb_lock);
1150 /* Also see bch2_pending_btree_node_free_insert_done() */
1151 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1153 struct btree_update *as;
1154 struct pending_btree_node_free *d;
1156 mutex_lock(&c->btree_interior_update_lock);
1157 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1159 for_each_pending_btree_node_free(c, as, d)
1160 if (d->index_update_done)
1161 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1163 mutex_unlock(&c->btree_interior_update_lock);
1167 static void bch2_gc_free(struct bch_fs *c)
1172 genradix_free(&c->reflink_gc_table);
1173 genradix_free(&c->gc_stripes);
1175 for_each_member_device(ca, c, i) {
1176 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1177 sizeof(struct bucket_array) +
1178 ca->mi.nbuckets * sizeof(struct bucket));
1179 ca->buckets_gc = NULL;
1181 free_percpu(ca->usage_gc);
1182 ca->usage_gc = NULL;
1185 free_percpu(c->usage_gc);
1189 static int bch2_gc_done(struct bch_fs *c,
1190 bool initial, bool metadata_only)
1192 struct bch_dev *ca = NULL;
1193 struct printbuf buf = PRINTBUF;
1194 bool verify = !metadata_only &&
1195 !c->opts.reconstruct_alloc &&
1196 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1200 percpu_down_write(&c->mark_lock);
1202 #define copy_field(_f, _msg, ...) \
1203 if (dst->_f != src->_f && \
1205 fsck_err(c, _msg ": got %llu, should be %llu" \
1206 , ##__VA_ARGS__, dst->_f, src->_f))) \
1208 #define copy_stripe_field(_f, _msg, ...) \
1209 if (dst->_f != src->_f && \
1211 fsck_err(c, "stripe %zu has wrong "_msg \
1212 ": got %u, should be %u", \
1213 iter.pos, ##__VA_ARGS__, \
1214 dst->_f, src->_f))) \
1216 #define copy_dev_field(_f, _msg, ...) \
1217 copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1218 #define copy_fs_field(_f, _msg, ...) \
1219 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1221 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1222 bch2_fs_usage_acc_to_base(c, i);
1224 for_each_member_device(ca, c, dev) {
1225 struct bch_dev_usage *dst = ca->usage_base;
1226 struct bch_dev_usage *src = (void *)
1227 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1230 copy_dev_field(buckets_ec, "buckets_ec");
1232 for (i = 0; i < BCH_DATA_NR; i++) {
1233 copy_dev_field(d[i].buckets, "%s buckets", bch2_data_types[i]);
1234 copy_dev_field(d[i].sectors, "%s sectors", bch2_data_types[i]);
1235 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1240 unsigned nr = fs_usage_u64s(c);
1241 struct bch_fs_usage *dst = c->usage_base;
1242 struct bch_fs_usage *src = (void *)
1243 bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1245 copy_fs_field(hidden, "hidden");
1246 copy_fs_field(btree, "btree");
1248 if (!metadata_only) {
1249 copy_fs_field(data, "data");
1250 copy_fs_field(cached, "cached");
1251 copy_fs_field(reserved, "reserved");
1252 copy_fs_field(nr_inodes,"nr_inodes");
1254 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1255 copy_fs_field(persistent_reserved[i],
1256 "persistent_reserved[%i]", i);
1259 for (i = 0; i < c->replicas.nr; i++) {
1260 struct bch_replicas_entry *e =
1261 cpu_replicas_entry(&c->replicas, i);
1263 if (metadata_only &&
1264 (e->data_type == BCH_DATA_user ||
1265 e->data_type == BCH_DATA_cached))
1268 printbuf_reset(&buf);
1269 bch2_replicas_entry_to_text(&buf, e);
1271 copy_fs_field(replicas[i], "%s", buf.buf);
1275 #undef copy_fs_field
1276 #undef copy_dev_field
1277 #undef copy_stripe_field
1281 percpu_ref_put(&ca->ref);
1283 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1285 percpu_up_write(&c->mark_lock);
1286 printbuf_exit(&buf);
1290 static int bch2_gc_start(struct bch_fs *c)
1292 struct bch_dev *ca = NULL;
1295 BUG_ON(c->usage_gc);
1297 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1298 sizeof(u64), GFP_KERNEL);
1300 bch_err(c, "error allocating c->usage_gc");
1301 return -BCH_ERR_ENOMEM_gc_start;
1304 for_each_member_device(ca, c, i) {
1305 BUG_ON(ca->usage_gc);
1307 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1308 if (!ca->usage_gc) {
1309 bch_err(c, "error allocating ca->usage_gc");
1310 percpu_ref_put(&ca->ref);
1311 return -BCH_ERR_ENOMEM_gc_start;
1314 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1315 ca->mi.nbuckets - ca->mi.first_bucket);
1321 static int bch2_gc_reset(struct bch_fs *c)
1326 for_each_member_device(ca, c, i) {
1327 free_percpu(ca->usage_gc);
1328 ca->usage_gc = NULL;
1331 free_percpu(c->usage_gc);
1334 return bch2_gc_start(c);
1337 /* returns true if not equal */
1338 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1339 struct bch_alloc_v4 r)
1341 return l.gen != r.gen ||
1342 l.oldest_gen != r.oldest_gen ||
1343 l.data_type != r.data_type ||
1344 l.dirty_sectors != r.dirty_sectors ||
1345 l.cached_sectors != r.cached_sectors ||
1346 l.stripe_redundancy != r.stripe_redundancy ||
1347 l.stripe != r.stripe;
1350 static int bch2_alloc_write_key(struct btree_trans *trans,
1351 struct btree_iter *iter,
1355 struct bch_fs *c = trans->c;
1356 struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1357 struct bucket gc, *b;
1358 struct bkey_i_alloc_v4 *a;
1359 struct bch_alloc_v4 old_convert, new;
1360 const struct bch_alloc_v4 *old;
1361 enum bch_data_type type;
1364 if (bkey_ge(iter->pos, POS(ca->dev_idx, ca->mi.nbuckets)))
1367 old = bch2_alloc_to_v4(k, &old_convert);
1370 percpu_down_read(&c->mark_lock);
1371 b = gc_bucket(ca, iter->pos.offset);
1374 * b->data_type doesn't yet include need_discard & need_gc_gen states -
1377 type = __alloc_data_type(b->dirty_sectors,
1382 if (b->data_type != type) {
1383 struct bch_dev_usage *u;
1386 u = this_cpu_ptr(ca->usage_gc);
1387 u->d[b->data_type].buckets--;
1388 b->data_type = type;
1389 u->d[b->data_type].buckets++;
1394 percpu_up_read(&c->mark_lock);
1396 if (metadata_only &&
1397 gc.data_type != BCH_DATA_sb &&
1398 gc.data_type != BCH_DATA_journal &&
1399 gc.data_type != BCH_DATA_btree)
1402 if (gen_after(old->gen, gc.gen))
1405 if (c->opts.reconstruct_alloc ||
1406 fsck_err_on(new.data_type != gc.data_type, c,
1407 "bucket %llu:%llu gen %u has wrong data_type"
1408 ": got %s, should be %s",
1409 iter->pos.inode, iter->pos.offset,
1411 bch2_data_types[new.data_type],
1412 bch2_data_types[gc.data_type]))
1413 new.data_type = gc.data_type;
1415 #define copy_bucket_field(_f) \
1416 if (c->opts.reconstruct_alloc || \
1417 fsck_err_on(new._f != gc._f, c, \
1418 "bucket %llu:%llu gen %u data type %s has wrong " #_f \
1419 ": got %u, should be %u", \
1420 iter->pos.inode, iter->pos.offset, \
1422 bch2_data_types[gc.data_type], \
1426 copy_bucket_field(gen);
1427 copy_bucket_field(dirty_sectors);
1428 copy_bucket_field(cached_sectors);
1429 copy_bucket_field(stripe_redundancy);
1430 copy_bucket_field(stripe);
1431 #undef copy_bucket_field
1433 if (!bch2_alloc_v4_cmp(*old, new))
1436 a = bch2_alloc_to_v4_mut(trans, k);
1437 ret = PTR_ERR_OR_ZERO(a);
1444 * The trigger normally makes sure this is set, but we're not running
1447 if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1448 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1450 ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1455 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1457 struct btree_trans trans;
1458 struct btree_iter iter;
1464 bch2_trans_init(&trans, c, 0, 0);
1466 for_each_member_device(ca, c, i) {
1467 ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1468 POS(ca->dev_idx, ca->mi.first_bucket),
1469 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1470 NULL, NULL, BTREE_INSERT_LAZY_RW,
1471 bch2_alloc_write_key(&trans, &iter, k, metadata_only));
1474 bch_err(c, "error writing alloc info: %s", bch2_err_str(ret));
1475 percpu_ref_put(&ca->ref);
1480 bch2_trans_exit(&trans);
1481 return ret < 0 ? ret : 0;
1484 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1487 struct btree_trans trans;
1488 struct btree_iter iter;
1491 struct bch_alloc_v4 a_convert;
1492 const struct bch_alloc_v4 *a;
1496 for_each_member_device(ca, c, i) {
1497 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1498 ca->mi.nbuckets * sizeof(struct bucket),
1499 GFP_KERNEL|__GFP_ZERO);
1501 percpu_ref_put(&ca->ref);
1502 bch_err(c, "error allocating ca->buckets[gc]");
1503 return -BCH_ERR_ENOMEM_gc_alloc_start;
1506 buckets->first_bucket = ca->mi.first_bucket;
1507 buckets->nbuckets = ca->mi.nbuckets;
1508 rcu_assign_pointer(ca->buckets_gc, buckets);
1511 bch2_trans_init(&trans, c, 0, 0);
1513 for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1514 BTREE_ITER_PREFETCH, k, ret) {
1515 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1516 g = gc_bucket(ca, k.k->p.offset);
1518 a = bch2_alloc_to_v4(k, &a_convert);
1523 if (metadata_only &&
1524 (a->data_type == BCH_DATA_user ||
1525 a->data_type == BCH_DATA_cached ||
1526 a->data_type == BCH_DATA_parity)) {
1527 g->data_type = a->data_type;
1528 g->dirty_sectors = a->dirty_sectors;
1529 g->cached_sectors = a->cached_sectors;
1530 g->stripe = a->stripe;
1531 g->stripe_redundancy = a->stripe_redundancy;
1534 bch2_trans_iter_exit(&trans, &iter);
1536 bch2_trans_exit(&trans);
1539 bch_err(c, "error reading alloc info at gc start: %s", bch2_err_str(ret));
1544 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1549 for_each_member_device(ca, c, i) {
1550 struct bucket_array *buckets = gc_bucket_array(ca);
1553 for_each_bucket(g, buckets) {
1554 if (metadata_only &&
1555 (g->data_type == BCH_DATA_user ||
1556 g->data_type == BCH_DATA_cached ||
1557 g->data_type == BCH_DATA_parity))
1560 g->dirty_sectors = 0;
1561 g->cached_sectors = 0;
1566 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1567 struct btree_iter *iter,
1571 struct bch_fs *c = trans->c;
1572 const __le64 *refcount = bkey_refcount_c(k);
1573 struct printbuf buf = PRINTBUF;
1574 struct reflink_gc *r;
1580 while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1581 r->offset < k.k->p.offset)
1585 r->offset != k.k->p.offset ||
1586 r->size != k.k->size) {
1587 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1591 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1592 "reflink key has wrong refcount:\n"
1595 (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1597 struct bkey_i *new = bch2_bkey_make_mut(trans, iter, k, 0);
1599 ret = PTR_ERR_OR_ZERO(new);
1604 new->k.type = KEY_TYPE_deleted;
1606 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1609 printbuf_exit(&buf);
1613 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1615 struct btree_trans trans;
1616 struct btree_iter iter;
1624 bch2_trans_init(&trans, c, 0, 0);
1626 ret = for_each_btree_key_commit(&trans, iter,
1627 BTREE_ID_reflink, POS_MIN,
1628 BTREE_ITER_PREFETCH, k,
1629 NULL, NULL, BTREE_INSERT_NOFAIL,
1630 bch2_gc_write_reflink_key(&trans, &iter, k, &idx));
1632 c->reflink_gc_nr = 0;
1633 bch2_trans_exit(&trans);
1637 static int bch2_gc_reflink_start(struct bch_fs *c,
1640 struct btree_trans trans;
1641 struct btree_iter iter;
1643 struct reflink_gc *r;
1649 bch2_trans_init(&trans, c, 0, 0);
1650 c->reflink_gc_nr = 0;
1652 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1653 BTREE_ITER_PREFETCH, k, ret) {
1654 const __le64 *refcount = bkey_refcount_c(k);
1659 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1662 ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1666 r->offset = k.k->p.offset;
1667 r->size = k.k->size;
1670 bch2_trans_iter_exit(&trans, &iter);
1672 bch2_trans_exit(&trans);
1676 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1678 struct genradix_iter iter;
1679 struct reflink_gc *r;
1681 genradix_for_each(&c->reflink_gc_table, iter, r)
1685 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1686 struct btree_iter *iter,
1689 struct bch_fs *c = trans->c;
1690 struct printbuf buf = PRINTBUF;
1691 const struct bch_stripe *s;
1692 struct gc_stripe *m;
1697 if (k.k->type != KEY_TYPE_stripe)
1700 s = bkey_s_c_to_stripe(k).v;
1701 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1703 for (i = 0; i < s->nr_blocks; i++) {
1704 u32 old = stripe_blockcount_get(s, i);
1705 u32 new = (m ? m->block_sectors[i] : 0);
1708 prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1715 bch2_bkey_val_to_text(&buf, c, k);
1717 if (fsck_err_on(bad, c, "%s", buf.buf)) {
1718 struct bkey_i_stripe *new;
1720 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1721 ret = PTR_ERR_OR_ZERO(new);
1725 bkey_reassemble(&new->k_i, k);
1727 for (i = 0; i < new->v.nr_blocks; i++)
1728 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1730 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1733 printbuf_exit(&buf);
1737 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1739 struct btree_trans trans;
1740 struct btree_iter iter;
1747 bch2_trans_init(&trans, c, 0, 0);
1749 ret = for_each_btree_key_commit(&trans, iter,
1750 BTREE_ID_stripes, POS_MIN,
1751 BTREE_ITER_PREFETCH, k,
1752 NULL, NULL, BTREE_INSERT_NOFAIL,
1753 bch2_gc_write_stripes_key(&trans, &iter, k));
1755 bch2_trans_exit(&trans);
1759 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1761 genradix_free(&c->gc_stripes);
1765 * bch2_gc - walk _all_ references to buckets, and recompute them:
1767 * Order matters here:
1768 * - Concurrent GC relies on the fact that we have a total ordering for
1769 * everything that GC walks - see gc_will_visit_node(),
1770 * gc_will_visit_root()
1772 * - also, references move around in the course of index updates and
1773 * various other crap: everything needs to agree on the ordering
1774 * references are allowed to move around in - e.g., we're allowed to
1775 * start with a reference owned by an open_bucket (the allocator) and
1776 * move it to the btree, but not the reverse.
1778 * This is necessary to ensure that gc doesn't miss references that
1779 * move around - if references move backwards in the ordering GC
1780 * uses, GC could skip past them
1782 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1787 lockdep_assert_held(&c->state_lock);
1789 down_write(&c->gc_lock);
1791 bch2_btree_interior_updates_flush(c);
1793 ret = bch2_gc_start(c) ?:
1794 bch2_gc_alloc_start(c, metadata_only) ?:
1795 bch2_gc_reflink_start(c, metadata_only);
1799 gc_pos_set(c, gc_phase(GC_PHASE_START));
1801 bch2_mark_superblocks(c);
1803 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) ||
1804 (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1805 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1806 c->opts.fix_errors != FSCK_OPT_NO)) {
1807 bch_info(c, "Starting topology repair pass");
1808 ret = bch2_repair_topology(c);
1811 bch_info(c, "Topology repair pass done");
1813 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1816 ret = bch2_gc_btrees(c, initial, metadata_only);
1818 if (ret == -BCH_ERR_need_topology_repair &&
1819 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1820 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1821 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1822 SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, true);
1826 if (ret == -BCH_ERR_need_topology_repair)
1827 ret = -BCH_ERR_fsck_errors_not_fixed;
1833 bch2_mark_pending_btree_node_frees(c);
1837 if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1838 (!iter && bch2_test_restart_gc)) {
1840 bch_info(c, "Unable to fix bucket gens, looping");
1846 * XXX: make sure gens we fixed got saved
1848 bch_info(c, "Second GC pass needed, restarting:");
1849 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1850 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1852 bch2_gc_stripes_reset(c, metadata_only);
1853 bch2_gc_alloc_reset(c, metadata_only);
1854 bch2_gc_reflink_reset(c, metadata_only);
1855 ret = bch2_gc_reset(c);
1859 /* flush fsck errors, reset counters */
1860 bch2_flush_fsck_errs(c);
1865 bch2_journal_block(&c->journal);
1867 ret = bch2_gc_stripes_done(c, metadata_only) ?:
1868 bch2_gc_reflink_done(c, metadata_only) ?:
1869 bch2_gc_alloc_done(c, metadata_only) ?:
1870 bch2_gc_done(c, initial, metadata_only);
1872 bch2_journal_unblock(&c->journal);
1875 percpu_down_write(&c->mark_lock);
1876 /* Indicates that gc is no longer in progress: */
1877 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1880 percpu_up_write(&c->mark_lock);
1882 up_write(&c->gc_lock);
1885 * At startup, allocations can happen directly instead of via the
1886 * allocator thread - issue wakeup in case they blocked on gc_lock:
1888 closure_wake_up(&c->freelist_wait);
1892 static int gc_btree_gens_key(struct btree_trans *trans,
1893 struct btree_iter *iter,
1896 struct bch_fs *c = trans->c;
1897 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1898 const struct bch_extent_ptr *ptr;
1902 percpu_down_read(&c->mark_lock);
1903 bkey_for_each_ptr(ptrs, ptr) {
1904 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1906 if (ptr_stale(ca, ptr) > 16) {
1907 percpu_up_read(&c->mark_lock);
1912 bkey_for_each_ptr(ptrs, ptr) {
1913 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1914 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1916 if (gen_after(*gen, ptr->gen))
1919 percpu_up_read(&c->mark_lock);
1922 u = bch2_bkey_make_mut(trans, iter, k, 0);
1923 ret = PTR_ERR_OR_ZERO(u);
1927 bch2_extent_normalize(c, bkey_i_to_s(u));
1931 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1934 struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1935 struct bch_alloc_v4 a_convert;
1936 const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1937 struct bkey_i_alloc_v4 *a_mut;
1940 if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1943 a_mut = bch2_alloc_to_v4_mut(trans, k);
1944 ret = PTR_ERR_OR_ZERO(a_mut);
1948 a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1949 a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1951 return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1954 int bch2_gc_gens(struct bch_fs *c)
1956 struct btree_trans trans;
1957 struct btree_iter iter;
1960 u64 b, start_time = local_clock();
1965 * Ideally we would be using state_lock and not gc_lock here, but that
1966 * introduces a deadlock in the RO path - we currently take the state
1967 * lock at the start of going RO, thus the gc thread may get stuck:
1969 if (!mutex_trylock(&c->gc_gens_lock))
1972 trace_and_count(c, gc_gens_start, c);
1973 down_read(&c->gc_lock);
1974 bch2_trans_init(&trans, c, 0, 0);
1976 for_each_member_device(ca, c, i) {
1977 struct bucket_gens *gens;
1979 BUG_ON(ca->oldest_gen);
1981 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1982 if (!ca->oldest_gen) {
1983 percpu_ref_put(&ca->ref);
1984 ret = -BCH_ERR_ENOMEM_gc_gens;
1988 gens = bucket_gens(ca);
1990 for (b = gens->first_bucket;
1991 b < gens->nbuckets; b++)
1992 ca->oldest_gen[b] = gens->b[b];
1995 for (i = 0; i < BTREE_ID_NR; i++)
1996 if (btree_type_has_ptrs(i)) {
1997 struct btree_iter iter;
2000 c->gc_gens_btree = i;
2001 c->gc_gens_pos = POS_MIN;
2002 ret = for_each_btree_key_commit(&trans, iter, i,
2004 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
2007 BTREE_INSERT_NOFAIL,
2008 gc_btree_gens_key(&trans, &iter, k));
2009 if (ret && !bch2_err_matches(ret, EROFS))
2010 bch_err(c, "error recalculating oldest_gen: %s", bch2_err_str(ret));
2015 ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
2017 BTREE_ITER_PREFETCH,
2020 BTREE_INSERT_NOFAIL,
2021 bch2_alloc_write_oldest_gen(&trans, &iter, k));
2022 if (ret && !bch2_err_matches(ret, EROFS))
2023 bch_err(c, "error writing oldest_gen: %s", bch2_err_str(ret));
2027 c->gc_gens_btree = 0;
2028 c->gc_gens_pos = POS_MIN;
2032 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2033 trace_and_count(c, gc_gens_end, c);
2035 for_each_member_device(ca, c, i) {
2036 kvfree(ca->oldest_gen);
2037 ca->oldest_gen = NULL;
2040 bch2_trans_exit(&trans);
2041 up_read(&c->gc_lock);
2042 mutex_unlock(&c->gc_gens_lock);
2046 static int bch2_gc_thread(void *arg)
2048 struct bch_fs *c = arg;
2049 struct io_clock *clock = &c->io_clock[WRITE];
2050 unsigned long last = atomic64_read(&clock->now);
2051 unsigned last_kick = atomic_read(&c->kick_gc);
2058 set_current_state(TASK_INTERRUPTIBLE);
2060 if (kthread_should_stop()) {
2061 __set_current_state(TASK_RUNNING);
2065 if (atomic_read(&c->kick_gc) != last_kick)
2068 if (c->btree_gc_periodic) {
2069 unsigned long next = last + c->capacity / 16;
2071 if (atomic64_read(&clock->now) >= next)
2074 bch2_io_clock_schedule_timeout(clock, next);
2081 __set_current_state(TASK_RUNNING);
2083 last = atomic64_read(&clock->now);
2084 last_kick = atomic_read(&c->kick_gc);
2087 * Full gc is currently incompatible with btree key cache:
2090 ret = bch2_gc(c, false, false);
2092 ret = bch2_gc_gens(c);
2095 bch_err(c, "btree gc failed: %s", bch2_err_str(ret));
2097 debug_check_no_locks_held();
2103 void bch2_gc_thread_stop(struct bch_fs *c)
2105 struct task_struct *p;
2108 c->gc_thread = NULL;
2116 int bch2_gc_thread_start(struct bch_fs *c)
2118 struct task_struct *p;
2123 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2125 bch_err(c, "error creating gc thread: %s", bch2_err_str(PTR_ERR(p)));