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"
31 #include <linux/slab.h>
32 #include <linux/bitops.h>
33 #include <linux/freezer.h>
34 #include <linux/kthread.h>
35 #include <linux/preempt.h>
36 #include <linux/rcupdate.h>
37 #include <linux/sched/task.h>
38 #include <trace/events/bcachefs.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_cmp(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 = FSCK_ERR_START_TOPOLOGY_REPAIR;
104 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
109 if (is_last && bpos_cmp(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 = FSCK_ERR_START_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 bch_fs *c,
169 enum btree_id btree, unsigned level,
170 struct bkey_s_c old, struct bkey_i *new)
176 bch2_bkey_buf_init(&tmp);
177 bch2_bkey_buf_reassemble(&tmp, c, old);
179 b = bch2_btree_node_get_noiter(c, tmp.k, btree, level, true);
180 if (!IS_ERR_OR_NULL(b)) {
181 mutex_lock(&c->btree_cache.lock);
183 bch2_btree_node_hash_remove(&c->btree_cache, b);
185 bkey_copy(&b->key, new);
186 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
189 mutex_unlock(&c->btree_cache.lock);
190 six_unlock_read(&b->c.lock);
193 bch2_bkey_buf_exit(&tmp, c);
196 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
198 struct bkey_i_btree_ptr_v2 *new;
201 new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
205 btree_ptr_to_v2(b, new);
206 b->data->min_key = new_min;
207 new->v.min_key = new_min;
208 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
210 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
216 bch2_btree_node_drop_keys_outside_node(b);
217 bkey_copy(&b->key, &new->k_i);
221 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
223 struct bkey_i_btree_ptr_v2 *new;
226 ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
230 new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
234 btree_ptr_to_v2(b, new);
235 b->data->max_key = new_max;
237 SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
239 ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
245 bch2_btree_node_drop_keys_outside_node(b);
247 mutex_lock(&c->btree_cache.lock);
248 bch2_btree_node_hash_remove(&c->btree_cache, b);
250 bkey_copy(&b->key, &new->k_i);
251 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
253 mutex_unlock(&c->btree_cache.lock);
257 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
258 struct btree *prev, struct btree *cur)
260 struct bpos expected_start = !prev
262 : bpos_successor(prev->key.k.p);
263 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
267 prt_printf(&buf1, "start of node: ");
268 bch2_bpos_to_text(&buf1, b->data->min_key);
270 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
273 bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
276 bpos_cmp(expected_start, cur->data->min_key) > 0 &&
277 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
278 /* cur overwrites prev: */
280 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
281 cur->data->min_key) >= 0, c,
282 "btree node overwritten by next node at btree %s level %u:\n"
285 bch2_btree_ids[b->c.btree_id], b->c.level,
286 buf1.buf, buf2.buf)) {
287 ret = DROP_PREV_NODE;
291 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
292 bpos_predecessor(cur->data->min_key)), c,
293 "btree node with incorrect max_key at btree %s level %u:\n"
296 bch2_btree_ids[b->c.btree_id], b->c.level,
298 ret = set_node_max(c, prev,
299 bpos_predecessor(cur->data->min_key));
301 /* prev overwrites cur: */
303 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
304 cur->data->max_key) >= 0, c,
305 "btree node overwritten by prev node at btree %s level %u:\n"
308 bch2_btree_ids[b->c.btree_id], b->c.level,
309 buf1.buf, buf2.buf)) {
310 ret = DROP_THIS_NODE;
314 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
315 "btree node with incorrect min_key at btree %s level %u:\n"
318 bch2_btree_ids[b->c.btree_id], b->c.level,
320 ret = set_node_min(c, cur, expected_start);
324 printbuf_exit(&buf2);
325 printbuf_exit(&buf1);
329 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
332 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
335 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
336 bch2_bpos_to_text(&buf2, b->key.k.p);
338 if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
339 "btree node with incorrect max_key at btree %s level %u:\n"
342 bch2_btree_ids[b->c.btree_id], b->c.level,
343 buf1.buf, buf2.buf)) {
344 ret = set_node_max(c, child, b->key.k.p);
350 printbuf_exit(&buf2);
351 printbuf_exit(&buf1);
355 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
357 struct btree_and_journal_iter iter;
359 struct bkey_buf prev_k, cur_k;
360 struct btree *prev = NULL, *cur = NULL;
361 bool have_child, dropped_children = false;
362 struct printbuf buf = PRINTBUF;
369 have_child = dropped_children = false;
370 bch2_bkey_buf_init(&prev_k);
371 bch2_bkey_buf_init(&cur_k);
372 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
374 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
375 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
376 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
378 bch2_btree_and_journal_iter_advance(&iter);
379 bch2_bkey_buf_reassemble(&cur_k, c, k);
381 cur = bch2_btree_node_get_noiter(c, cur_k.k,
382 b->c.btree_id, b->c.level - 1,
384 ret = PTR_ERR_OR_ZERO(cur);
386 printbuf_reset(&buf);
387 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
389 if (mustfix_fsck_err_on(ret == -EIO, c,
390 "Topology repair: unreadable btree node at btree %s level %u:\n"
392 bch2_btree_ids[b->c.btree_id],
395 bch2_btree_node_evict(c, cur_k.k);
396 ret = bch2_journal_key_delete(c, b->c.btree_id,
397 b->c.level, cur_k.k->k.p);
405 bch_err(c, "%s: error %i getting btree node",
410 ret = btree_repair_node_boundaries(c, b, prev, cur);
412 if (ret == DROP_THIS_NODE) {
413 six_unlock_read(&cur->c.lock);
414 bch2_btree_node_evict(c, cur_k.k);
415 ret = bch2_journal_key_delete(c, b->c.btree_id,
416 b->c.level, cur_k.k->k.p);
424 six_unlock_read(&prev->c.lock);
427 if (ret == DROP_PREV_NODE) {
428 bch2_btree_node_evict(c, prev_k.k);
429 ret = bch2_journal_key_delete(c, b->c.btree_id,
430 b->c.level, prev_k.k->k.p);
434 bch2_btree_and_journal_iter_exit(&iter);
435 bch2_bkey_buf_exit(&prev_k, c);
436 bch2_bkey_buf_exit(&cur_k, c);
443 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
446 if (!ret && !IS_ERR_OR_NULL(prev)) {
448 ret = btree_repair_node_end(c, b, prev);
451 if (!IS_ERR_OR_NULL(prev))
452 six_unlock_read(&prev->c.lock);
454 if (!IS_ERR_OR_NULL(cur))
455 six_unlock_read(&cur->c.lock);
461 bch2_btree_and_journal_iter_exit(&iter);
462 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
464 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
465 bch2_bkey_buf_reassemble(&cur_k, c, k);
466 bch2_btree_and_journal_iter_advance(&iter);
468 cur = bch2_btree_node_get_noiter(c, cur_k.k,
469 b->c.btree_id, b->c.level - 1,
471 ret = PTR_ERR_OR_ZERO(cur);
474 bch_err(c, "%s: error %i getting btree node",
479 ret = bch2_btree_repair_topology_recurse(c, cur);
480 six_unlock_read(&cur->c.lock);
483 if (ret == DROP_THIS_NODE) {
484 bch2_btree_node_evict(c, cur_k.k);
485 ret = bch2_journal_key_delete(c, b->c.btree_id,
486 b->c.level, cur_k.k->k.p);
487 dropped_children = true;
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 "empty interior btree node at btree %s level %u\n"
502 bch2_btree_ids[b->c.btree_id],
503 b->c.level, buf.buf))
504 ret = DROP_THIS_NODE;
507 if (!IS_ERR_OR_NULL(prev))
508 six_unlock_read(&prev->c.lock);
509 if (!IS_ERR_OR_NULL(cur))
510 six_unlock_read(&cur->c.lock);
512 bch2_btree_and_journal_iter_exit(&iter);
513 bch2_bkey_buf_exit(&prev_k, c);
514 bch2_bkey_buf_exit(&cur_k, c);
516 if (!ret && dropped_children)
523 static int bch2_repair_topology(struct bch_fs *c)
529 for (i = 0; i < BTREE_ID_NR && !ret; i++) {
530 b = c->btree_roots[i].b;
531 if (btree_node_fake(b))
534 six_lock_read(&b->c.lock, NULL, NULL);
535 ret = bch2_btree_repair_topology_recurse(c, b);
536 six_unlock_read(&b->c.lock);
538 if (ret == DROP_THIS_NODE) {
539 bch_err(c, "empty btree root - repair unimplemented");
547 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
548 unsigned level, bool is_root,
551 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
552 const union bch_extent_entry *entry;
553 struct extent_ptr_decoded p = { 0 };
554 bool do_update = false;
555 struct printbuf buf = PRINTBUF;
560 * use check_bucket_ref here
562 bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
563 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
564 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
565 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
567 if (c->opts.reconstruct_alloc ||
568 fsck_err_on(!g->gen_valid, c,
569 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
571 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
572 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
574 (printbuf_reset(&buf),
575 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
584 if (fsck_err_on(gen_cmp(p.ptr.gen, g->gen) > 0, c,
585 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
587 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
588 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
590 (printbuf_reset(&buf),
591 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
596 g->dirty_sectors = 0;
597 g->cached_sectors = 0;
598 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
604 if (fsck_err_on(gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
605 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
607 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
608 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
610 (printbuf_reset(&buf),
611 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
614 if (fsck_err_on(!p.ptr.cached &&
615 gen_cmp(p.ptr.gen, g->gen) < 0, c,
616 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
618 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
619 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
621 (printbuf_reset(&buf),
622 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
625 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
628 if (fsck_err_on(g->data_type &&
629 g->data_type != data_type, c,
630 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
632 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
633 bch2_data_types[g->data_type],
634 bch2_data_types[data_type],
635 (printbuf_reset(&buf),
636 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
637 if (data_type == BCH_DATA_btree) {
638 g->data_type = data_type;
639 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
646 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
648 if (fsck_err_on(!m || !m->alive, c,
649 "pointer to nonexistent stripe %llu\n"
652 (printbuf_reset(&buf),
653 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
656 if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
657 "pointer does not match stripe %llu\n"
660 (printbuf_reset(&buf),
661 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
667 struct bkey_ptrs ptrs;
668 union bch_extent_entry *entry;
669 struct bch_extent_ptr *ptr;
673 bch_err(c, "cannot update btree roots yet");
678 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
680 bch_err(c, "%s: error allocating new key", __func__);
685 bkey_reassemble(new, *k);
689 * We don't want to drop btree node pointers - if the
690 * btree node isn't there anymore, the read path will
693 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
694 bkey_for_each_ptr(ptrs, ptr) {
695 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
696 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
701 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
702 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
703 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
704 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
707 (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
709 gen_cmp(ptr->gen, g->gen) < 0) ||
710 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
712 g->data_type != data_type);
715 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
716 bkey_extent_entry_for_each(ptrs, entry) {
717 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
718 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
719 entry->stripe_ptr.idx);
720 union bch_extent_entry *next_ptr;
722 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
723 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
728 bch_err(c, "aieee, found stripe ptr with no data ptr");
732 if (!m || !m->alive ||
733 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
736 bch2_bkey_extent_entry_drop(new, entry);
743 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
750 bch2_btree_node_update_key_early(c, btree_id, level - 1, *k, new);
752 if (c->opts.verbose) {
753 printbuf_reset(&buf);
754 bch2_bkey_val_to_text(&buf, c, *k);
755 bch_info(c, "updated %s", buf.buf);
757 printbuf_reset(&buf);
758 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
759 bch_info(c, "new key %s", buf.buf);
762 *k = bkey_i_to_s_c(new);
770 /* marking of btree keys/nodes: */
772 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
773 unsigned level, bool is_root,
777 struct bch_fs *c = trans->c;
778 struct bkey deleted = KEY(0, 0, 0);
779 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
782 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
788 BUG_ON(bch2_journal_seq_verify &&
789 k->k->version.lo > atomic64_read(&c->journal.seq));
791 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
795 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
796 "key version number higher than recorded: %llu > %llu",
798 atomic64_read(&c->key_version)))
799 atomic64_set(&c->key_version, k->k->version.lo);
802 ret = commit_do(trans, NULL, NULL, 0,
803 bch2_mark_key(trans, old, *k, flags));
807 bch_err(c, "%s: ret %i", __func__, ret);
811 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
813 struct bch_fs *c = trans->c;
814 struct btree_node_iter iter;
815 struct bkey unpacked;
817 struct bkey_buf prev, cur;
820 if (!btree_node_type_needs_gc(btree_node_type(b)))
823 bch2_btree_node_iter_init_from_start(&iter, b);
824 bch2_bkey_buf_init(&prev);
825 bch2_bkey_buf_init(&cur);
826 bkey_init(&prev.k->k);
828 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
829 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
834 bch2_btree_node_iter_advance(&iter, b);
837 bch2_bkey_buf_reassemble(&cur, c, k);
839 ret = bch2_gc_check_topology(c, b, &prev, cur,
840 bch2_btree_node_iter_end(&iter));
846 bch2_bkey_buf_exit(&cur, c);
847 bch2_bkey_buf_exit(&prev, c);
851 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
852 bool initial, bool metadata_only)
854 struct bch_fs *c = trans->c;
855 struct btree_iter iter;
857 unsigned depth = metadata_only ? 1 : 0;
860 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
862 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
863 0, depth, BTREE_ITER_PREFETCH, b, ret) {
864 bch2_verify_btree_nr_keys(b);
866 gc_pos_set(c, gc_pos_btree_node(b));
868 ret = btree_gc_mark_node(trans, b, initial);
872 bch2_trans_iter_exit(trans, &iter);
877 mutex_lock(&c->btree_root_lock);
878 b = c->btree_roots[btree_id].b;
879 if (!btree_node_fake(b)) {
880 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
882 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
885 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
886 mutex_unlock(&c->btree_root_lock);
891 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
892 unsigned target_depth)
894 struct bch_fs *c = trans->c;
895 struct btree_and_journal_iter iter;
897 struct bkey_buf cur, prev;
898 struct printbuf buf = PRINTBUF;
901 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
902 bch2_bkey_buf_init(&prev);
903 bch2_bkey_buf_init(&cur);
904 bkey_init(&prev.k->k);
906 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
907 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
908 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
910 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
913 bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
918 bch2_bkey_buf_reassemble(&cur, c, k);
919 k = bkey_i_to_s_c(cur.k);
921 bch2_btree_and_journal_iter_advance(&iter);
923 ret = bch2_gc_check_topology(c, b,
925 !bch2_btree_and_journal_iter_peek(&iter).k);
929 bch2_btree_and_journal_iter_advance(&iter);
933 if (b->c.level > target_depth) {
934 bch2_btree_and_journal_iter_exit(&iter);
935 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
937 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
940 bch2_bkey_buf_reassemble(&cur, c, k);
941 bch2_btree_and_journal_iter_advance(&iter);
943 child = bch2_btree_node_get_noiter(c, cur.k,
944 b->c.btree_id, b->c.level - 1,
946 ret = PTR_ERR_OR_ZERO(child);
949 bch2_topology_error(c);
955 "Unreadable btree node at btree %s level %u:\n"
957 bch2_btree_ids[b->c.btree_id],
959 (printbuf_reset(&buf),
960 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
961 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
962 ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
963 bch_info(c, "Halting mark and sweep to start topology repair pass");
966 /* Continue marking when opted to not
969 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
973 bch_err(c, "%s: error %i getting btree node",
978 ret = bch2_gc_btree_init_recurse(trans, child,
980 six_unlock_read(&child->c.lock);
987 bch2_bkey_buf_exit(&cur, c);
988 bch2_bkey_buf_exit(&prev, c);
989 bch2_btree_and_journal_iter_exit(&iter);
994 static int bch2_gc_btree_init(struct btree_trans *trans,
995 enum btree_id btree_id,
998 struct bch_fs *c = trans->c;
1000 unsigned target_depth = metadata_only ? 1 : 0;
1001 struct printbuf buf = PRINTBUF;
1004 b = c->btree_roots[btree_id].b;
1006 if (btree_node_fake(b))
1009 six_lock_read(&b->c.lock, NULL, NULL);
1010 printbuf_reset(&buf);
1011 bch2_bpos_to_text(&buf, b->data->min_key);
1012 if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
1013 "btree root with incorrect min_key: %s", buf.buf)) {
1014 bch_err(c, "repair unimplemented");
1015 ret = FSCK_ERR_EXIT;
1019 printbuf_reset(&buf);
1020 bch2_bpos_to_text(&buf, b->data->max_key);
1021 if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
1022 "btree root with incorrect max_key: %s", buf.buf)) {
1023 bch_err(c, "repair unimplemented");
1024 ret = FSCK_ERR_EXIT;
1028 if (b->c.level >= target_depth)
1029 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1032 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1034 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1038 six_unlock_read(&b->c.lock);
1041 bch_err(c, "%s: ret %i", __func__, ret);
1042 printbuf_exit(&buf);
1046 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1048 return (int) btree_id_to_gc_phase(l) -
1049 (int) btree_id_to_gc_phase(r);
1052 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1054 struct btree_trans trans;
1055 enum btree_id ids[BTREE_ID_NR];
1059 bch2_trans_init(&trans, c, 0, 0);
1062 trans.is_initial_gc = true;
1064 for (i = 0; i < BTREE_ID_NR; i++)
1066 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1068 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1070 ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1071 : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1074 bch_err(c, "%s: ret %i", __func__, ret);
1076 bch2_trans_exit(&trans);
1080 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1082 enum bch_data_type type,
1085 u64 b = sector_to_bucket(ca, start);
1089 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1091 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1092 gc_phase(GC_PHASE_SB), flags);
1095 } while (start < end);
1098 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1101 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1105 for (i = 0; i < layout->nr_superblocks; i++) {
1106 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1108 if (offset == BCH_SB_SECTOR)
1109 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1110 BCH_DATA_sb, flags);
1112 mark_metadata_sectors(c, ca, offset,
1113 offset + (1 << layout->sb_max_size_bits),
1114 BCH_DATA_sb, flags);
1117 for (i = 0; i < ca->journal.nr; i++) {
1118 b = ca->journal.buckets[i];
1119 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1121 gc_phase(GC_PHASE_SB), flags);
1125 static void bch2_mark_superblocks(struct bch_fs *c)
1130 mutex_lock(&c->sb_lock);
1131 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1133 for_each_online_member(ca, c, i)
1134 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1135 mutex_unlock(&c->sb_lock);
1139 /* Also see bch2_pending_btree_node_free_insert_done() */
1140 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1142 struct btree_update *as;
1143 struct pending_btree_node_free *d;
1145 mutex_lock(&c->btree_interior_update_lock);
1146 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1148 for_each_pending_btree_node_free(c, as, d)
1149 if (d->index_update_done)
1150 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1152 mutex_unlock(&c->btree_interior_update_lock);
1156 static void bch2_gc_free(struct bch_fs *c)
1161 genradix_free(&c->reflink_gc_table);
1162 genradix_free(&c->gc_stripes);
1164 for_each_member_device(ca, c, i) {
1165 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1166 sizeof(struct bucket_array) +
1167 ca->mi.nbuckets * sizeof(struct bucket));
1168 ca->buckets_gc = NULL;
1170 free_percpu(ca->usage_gc);
1171 ca->usage_gc = NULL;
1174 free_percpu(c->usage_gc);
1178 static int bch2_gc_done(struct bch_fs *c,
1179 bool initial, bool metadata_only)
1181 struct bch_dev *ca = NULL;
1182 struct printbuf buf = PRINTBUF;
1183 bool verify = !metadata_only &&
1184 !c->opts.reconstruct_alloc &&
1185 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1189 percpu_down_write(&c->mark_lock);
1191 #define copy_field(_f, _msg, ...) \
1192 if (dst->_f != src->_f && \
1194 fsck_err(c, _msg ": got %llu, should be %llu" \
1195 , ##__VA_ARGS__, dst->_f, src->_f))) \
1197 #define copy_stripe_field(_f, _msg, ...) \
1198 if (dst->_f != src->_f && \
1200 fsck_err(c, "stripe %zu has wrong "_msg \
1201 ": got %u, should be %u", \
1202 iter.pos, ##__VA_ARGS__, \
1203 dst->_f, src->_f))) \
1205 #define copy_dev_field(_f, _msg, ...) \
1206 copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1207 #define copy_fs_field(_f, _msg, ...) \
1208 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1210 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1211 bch2_fs_usage_acc_to_base(c, i);
1213 for_each_member_device(ca, c, dev) {
1214 struct bch_dev_usage *dst = ca->usage_base;
1215 struct bch_dev_usage *src = (void *)
1216 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1219 copy_dev_field(buckets_ec, "buckets_ec");
1221 for (i = 0; i < BCH_DATA_NR; i++) {
1222 copy_dev_field(d[i].buckets, "%s buckets", bch2_data_types[i]);
1223 copy_dev_field(d[i].sectors, "%s sectors", bch2_data_types[i]);
1224 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1229 unsigned nr = fs_usage_u64s(c);
1230 struct bch_fs_usage *dst = c->usage_base;
1231 struct bch_fs_usage *src = (void *)
1232 bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1234 copy_fs_field(hidden, "hidden");
1235 copy_fs_field(btree, "btree");
1237 if (!metadata_only) {
1238 copy_fs_field(data, "data");
1239 copy_fs_field(cached, "cached");
1240 copy_fs_field(reserved, "reserved");
1241 copy_fs_field(nr_inodes,"nr_inodes");
1243 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1244 copy_fs_field(persistent_reserved[i],
1245 "persistent_reserved[%i]", i);
1248 for (i = 0; i < c->replicas.nr; i++) {
1249 struct bch_replicas_entry *e =
1250 cpu_replicas_entry(&c->replicas, i);
1252 if (metadata_only &&
1253 (e->data_type == BCH_DATA_user ||
1254 e->data_type == BCH_DATA_cached))
1257 printbuf_reset(&buf);
1258 bch2_replicas_entry_to_text(&buf, e);
1260 copy_fs_field(replicas[i], "%s", buf.buf);
1264 #undef copy_fs_field
1265 #undef copy_dev_field
1266 #undef copy_stripe_field
1270 percpu_ref_put(&ca->ref);
1272 bch_err(c, "%s: ret %i", __func__, ret);
1274 percpu_up_write(&c->mark_lock);
1275 printbuf_exit(&buf);
1279 static int bch2_gc_start(struct bch_fs *c,
1282 struct bch_dev *ca = NULL;
1285 BUG_ON(c->usage_gc);
1287 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1288 sizeof(u64), GFP_KERNEL);
1290 bch_err(c, "error allocating c->usage_gc");
1294 for_each_member_device(ca, c, i) {
1295 BUG_ON(ca->buckets_gc);
1296 BUG_ON(ca->usage_gc);
1298 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1299 if (!ca->usage_gc) {
1300 bch_err(c, "error allocating ca->usage_gc");
1301 percpu_ref_put(&ca->ref);
1305 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1306 ca->mi.nbuckets - ca->mi.first_bucket);
1312 /* returns true if not equal */
1313 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1314 struct bch_alloc_v4 r)
1316 return l.gen != r.gen ||
1317 l.oldest_gen != r.oldest_gen ||
1318 l.data_type != r.data_type ||
1319 l.dirty_sectors != r.dirty_sectors ||
1320 l.cached_sectors != r.cached_sectors ||
1321 l.stripe_redundancy != r.stripe_redundancy ||
1322 l.stripe != r.stripe;
1325 static int bch2_alloc_write_key(struct btree_trans *trans,
1326 struct btree_iter *iter,
1329 struct bch_fs *c = trans->c;
1330 struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1331 struct bucket gc, *b;
1333 struct bkey_i_alloc_v4 *a;
1334 struct bch_alloc_v4 old, new;
1335 enum bch_data_type type;
1338 k = bch2_btree_iter_peek_slot(iter);
1343 bch2_alloc_to_v4(k, &old);
1346 percpu_down_read(&c->mark_lock);
1347 b = gc_bucket(ca, iter->pos.offset);
1350 * b->data_type doesn't yet include need_discard & need_gc_gen states -
1353 type = __alloc_data_type(b->dirty_sectors,
1358 if (b->data_type != type) {
1359 struct bch_dev_usage *u;
1362 u = this_cpu_ptr(ca->usage_gc);
1363 u->d[b->data_type].buckets--;
1364 b->data_type = type;
1365 u->d[b->data_type].buckets++;
1370 percpu_up_read(&c->mark_lock);
1372 if (metadata_only &&
1373 gc.data_type != BCH_DATA_sb &&
1374 gc.data_type != BCH_DATA_journal &&
1375 gc.data_type != BCH_DATA_btree)
1378 if (gen_after(old.gen, gc.gen))
1381 #define copy_bucket_field(_f) \
1382 if (c->opts.reconstruct_alloc || \
1383 fsck_err_on(new._f != gc._f, c, \
1384 "bucket %llu:%llu gen %u data type %s has wrong " #_f \
1385 ": got %u, should be %u", \
1386 iter->pos.inode, iter->pos.offset, \
1388 bch2_data_types[gc.data_type], \
1392 copy_bucket_field(gen);
1393 copy_bucket_field(data_type);
1394 copy_bucket_field(dirty_sectors);
1395 copy_bucket_field(cached_sectors);
1396 copy_bucket_field(stripe_redundancy);
1397 copy_bucket_field(stripe);
1398 #undef copy_bucket_field
1400 if (!bch2_alloc_v4_cmp(old, new))
1403 a = bch2_alloc_to_v4_mut(trans, k);
1404 ret = PTR_ERR_OR_ZERO(a);
1411 * The trigger normally makes sure this is set, but we're not running
1414 if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1415 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1417 ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1422 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1424 struct btree_trans trans;
1425 struct btree_iter iter;
1431 bch2_trans_init(&trans, c, 0, 0);
1433 for_each_member_device(ca, c, i) {
1434 for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1435 POS(ca->dev_idx, ca->mi.first_bucket),
1437 BTREE_ITER_PREFETCH, k, ret) {
1438 if (bkey_cmp(iter.pos, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
1441 ret = commit_do(&trans, NULL, NULL,
1442 BTREE_INSERT_LAZY_RW,
1443 bch2_alloc_write_key(&trans, &iter,
1448 bch2_trans_iter_exit(&trans, &iter);
1451 bch_err(c, "error writing alloc info: %i", ret);
1452 percpu_ref_put(&ca->ref);
1457 bch2_trans_exit(&trans);
1461 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1464 struct btree_trans trans;
1465 struct btree_iter iter;
1468 struct bch_alloc_v4 a;
1472 for_each_member_device(ca, c, i) {
1473 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1474 ca->mi.nbuckets * sizeof(struct bucket),
1475 GFP_KERNEL|__GFP_ZERO);
1477 percpu_ref_put(&ca->ref);
1478 bch_err(c, "error allocating ca->buckets[gc]");
1482 buckets->first_bucket = ca->mi.first_bucket;
1483 buckets->nbuckets = ca->mi.nbuckets;
1484 rcu_assign_pointer(ca->buckets_gc, buckets);
1487 bch2_trans_init(&trans, c, 0, 0);
1489 for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1490 BTREE_ITER_PREFETCH, k, ret) {
1491 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1492 g = gc_bucket(ca, k.k->p.offset);
1494 bch2_alloc_to_v4(k, &a);
1499 if (metadata_only &&
1500 (a.data_type == BCH_DATA_user ||
1501 a.data_type == BCH_DATA_cached ||
1502 a.data_type == BCH_DATA_parity)) {
1503 g->data_type = a.data_type;
1504 g->dirty_sectors = a.dirty_sectors;
1505 g->cached_sectors = a.cached_sectors;
1506 g->stripe = a.stripe;
1507 g->stripe_redundancy = a.stripe_redundancy;
1510 bch2_trans_iter_exit(&trans, &iter);
1512 bch2_trans_exit(&trans);
1515 bch_err(c, "error reading alloc info at gc start: %i", ret);
1520 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1525 for_each_member_device(ca, c, i) {
1526 struct bucket_array *buckets = gc_bucket_array(ca);
1529 for_each_bucket(g, buckets) {
1530 if (metadata_only &&
1531 (g->data_type == BCH_DATA_user ||
1532 g->data_type == BCH_DATA_cached ||
1533 g->data_type == BCH_DATA_parity))
1536 g->dirty_sectors = 0;
1537 g->cached_sectors = 0;
1542 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1544 struct btree_trans trans;
1545 struct btree_iter iter;
1547 struct reflink_gc *r;
1549 struct printbuf buf = PRINTBUF;
1555 bch2_trans_init(&trans, c, 0, 0);
1557 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1558 BTREE_ITER_PREFETCH, k, ret) {
1559 const __le64 *refcount = bkey_refcount_c(k);
1564 r = genradix_ptr(&c->reflink_gc_table, idx++);
1566 r->offset != k.k->p.offset ||
1567 r->size != k.k->size) {
1568 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1573 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1574 "reflink key has wrong refcount:\n"
1577 (printbuf_reset(&buf),
1578 bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1582 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1588 bkey_reassemble(new, k);
1591 new->k.type = KEY_TYPE_deleted;
1593 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1595 ret = commit_do(&trans, NULL, NULL, 0,
1596 __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1604 bch2_trans_iter_exit(&trans, &iter);
1605 c->reflink_gc_nr = 0;
1606 bch2_trans_exit(&trans);
1607 printbuf_exit(&buf);
1611 static int bch2_gc_reflink_start(struct bch_fs *c,
1614 struct btree_trans trans;
1615 struct btree_iter iter;
1617 struct reflink_gc *r;
1623 bch2_trans_init(&trans, c, 0, 0);
1624 c->reflink_gc_nr = 0;
1626 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1627 BTREE_ITER_PREFETCH, k, ret) {
1628 const __le64 *refcount = bkey_refcount_c(k);
1633 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1640 r->offset = k.k->p.offset;
1641 r->size = k.k->size;
1644 bch2_trans_iter_exit(&trans, &iter);
1646 bch2_trans_exit(&trans);
1650 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1652 struct genradix_iter iter;
1653 struct reflink_gc *r;
1655 genradix_for_each(&c->reflink_gc_table, iter, r)
1659 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1661 struct btree_trans trans;
1662 struct btree_iter iter;
1664 struct gc_stripe *m;
1665 const struct bch_stripe *s;
1666 struct printbuf buf = PRINTBUF;
1673 bch2_trans_init(&trans, c, 0, 0);
1675 for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1676 BTREE_ITER_PREFETCH, k, ret) {
1677 if (k.k->type != KEY_TYPE_stripe)
1680 s = bkey_s_c_to_stripe(k).v;
1681 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1683 for (i = 0; i < s->nr_blocks; i++)
1684 if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1688 if (fsck_err_on(true, c,
1689 "stripe has wrong block sector count %u:\n"
1692 (printbuf_reset(&buf),
1693 bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1694 m ? m->block_sectors[i] : 0)) {
1695 struct bkey_i_stripe *new;
1697 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1703 bkey_reassemble(&new->k_i, k);
1705 for (i = 0; i < new->v.nr_blocks; i++)
1706 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1708 ret = commit_do(&trans, NULL, NULL, 0,
1709 __bch2_btree_insert(&trans, BTREE_ID_reflink, &new->k_i));
1714 bch2_trans_iter_exit(&trans, &iter);
1716 bch2_trans_exit(&trans);
1718 printbuf_exit(&buf);
1722 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1724 genradix_free(&c->gc_stripes);
1728 * bch2_gc - walk _all_ references to buckets, and recompute them:
1730 * Order matters here:
1731 * - Concurrent GC relies on the fact that we have a total ordering for
1732 * everything that GC walks - see gc_will_visit_node(),
1733 * gc_will_visit_root()
1735 * - also, references move around in the course of index updates and
1736 * various other crap: everything needs to agree on the ordering
1737 * references are allowed to move around in - e.g., we're allowed to
1738 * start with a reference owned by an open_bucket (the allocator) and
1739 * move it to the btree, but not the reverse.
1741 * This is necessary to ensure that gc doesn't miss references that
1742 * move around - if references move backwards in the ordering GC
1743 * uses, GC could skip past them
1745 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1750 lockdep_assert_held(&c->state_lock);
1752 down_write(&c->gc_lock);
1754 bch2_btree_interior_updates_flush(c);
1756 ret = bch2_gc_start(c, metadata_only) ?:
1757 bch2_gc_alloc_start(c, metadata_only) ?:
1758 bch2_gc_reflink_start(c, metadata_only);
1762 gc_pos_set(c, gc_phase(GC_PHASE_START));
1764 bch2_mark_superblocks(c);
1766 if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1767 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1768 c->opts.fix_errors != FSCK_OPT_NO) {
1769 bch_info(c, "Starting topology repair pass");
1770 ret = bch2_repair_topology(c);
1773 bch_info(c, "Topology repair pass done");
1775 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1778 ret = bch2_gc_btrees(c, initial, metadata_only);
1780 if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1781 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1782 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1783 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1784 SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, true);
1788 if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1789 ret = FSCK_ERR_EXIT;
1795 bch2_mark_pending_btree_node_frees(c);
1799 if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1800 (!iter && bch2_test_restart_gc)) {
1802 bch_info(c, "Unable to fix bucket gens, looping");
1808 * XXX: make sure gens we fixed got saved
1810 bch_info(c, "Second GC pass needed, restarting:");
1811 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1812 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1814 bch2_gc_stripes_reset(c, metadata_only);
1815 bch2_gc_alloc_reset(c, metadata_only);
1816 bch2_gc_reflink_reset(c, metadata_only);
1818 /* flush fsck errors, reset counters */
1819 bch2_flush_fsck_errs(c);
1824 bch2_journal_block(&c->journal);
1826 ret = bch2_gc_stripes_done(c, metadata_only) ?:
1827 bch2_gc_reflink_done(c, metadata_only) ?:
1828 bch2_gc_alloc_done(c, metadata_only) ?:
1829 bch2_gc_done(c, initial, metadata_only);
1831 bch2_journal_unblock(&c->journal);
1834 percpu_down_write(&c->mark_lock);
1835 /* Indicates that gc is no longer in progress: */
1836 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1839 percpu_up_write(&c->mark_lock);
1841 up_write(&c->gc_lock);
1844 * At startup, allocations can happen directly instead of via the
1845 * allocator thread - issue wakeup in case they blocked on gc_lock:
1847 closure_wake_up(&c->freelist_wait);
1851 static int gc_btree_gens_key(struct btree_trans *trans,
1852 struct btree_iter *iter,
1855 struct bch_fs *c = trans->c;
1856 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1857 const struct bch_extent_ptr *ptr;
1861 percpu_down_read(&c->mark_lock);
1862 bkey_for_each_ptr(ptrs, ptr) {
1863 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1865 if (ptr_stale(ca, ptr) > 16) {
1866 percpu_up_read(&c->mark_lock);
1871 bkey_for_each_ptr(ptrs, ptr) {
1872 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1873 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1875 if (gen_after(*gen, ptr->gen))
1878 percpu_up_read(&c->mark_lock);
1881 u = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1882 ret = PTR_ERR_OR_ZERO(u);
1886 bkey_reassemble(u, k);
1888 bch2_extent_normalize(c, bkey_i_to_s(u));
1889 return bch2_trans_update(trans, iter, u, 0);
1892 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1895 struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1896 struct bch_alloc_v4 a;
1897 struct bkey_i_alloc_v4 *a_mut;
1900 bch2_alloc_to_v4(k, &a);
1902 if (a.oldest_gen == ca->oldest_gen[iter->pos.offset])
1905 a_mut = bch2_alloc_to_v4_mut(trans, k);
1906 ret = PTR_ERR_OR_ZERO(a_mut);
1910 a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1911 a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1913 return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1916 int bch2_gc_gens(struct bch_fs *c)
1918 struct btree_trans trans;
1919 struct btree_iter iter;
1922 u64 b, start_time = local_clock();
1927 * Ideally we would be using state_lock and not gc_lock here, but that
1928 * introduces a deadlock in the RO path - we currently take the state
1929 * lock at the start of going RO, thus the gc thread may get stuck:
1931 if (!mutex_trylock(&c->gc_gens_lock))
1934 trace_gc_gens_start(c);
1935 down_read(&c->gc_lock);
1936 bch2_trans_init(&trans, c, 0, 0);
1938 for_each_member_device(ca, c, i) {
1939 struct bucket_gens *gens;
1941 BUG_ON(ca->oldest_gen);
1943 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1944 if (!ca->oldest_gen) {
1945 percpu_ref_put(&ca->ref);
1950 gens = bucket_gens(ca);
1952 for (b = gens->first_bucket;
1953 b < gens->nbuckets; b++)
1954 ca->oldest_gen[b] = gens->b[b];
1957 for (i = 0; i < BTREE_ID_NR; i++)
1958 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1959 struct btree_iter iter;
1962 c->gc_gens_btree = i;
1963 c->gc_gens_pos = POS_MIN;
1964 ret = for_each_btree_key_commit(&trans, iter, i,
1966 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
1969 BTREE_INSERT_NOFAIL,
1970 gc_btree_gens_key(&trans, &iter, k));
1972 bch_err(c, "error recalculating oldest_gen: %i", ret);
1977 ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1979 BTREE_ITER_PREFETCH,
1982 BTREE_INSERT_NOFAIL,
1983 bch2_alloc_write_oldest_gen(&trans, &iter, k));
1985 bch_err(c, "error writing oldest_gen: %i", ret);
1989 c->gc_gens_btree = 0;
1990 c->gc_gens_pos = POS_MIN;
1994 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1995 trace_gc_gens_end(c);
1997 for_each_member_device(ca, c, i) {
1998 kvfree(ca->oldest_gen);
1999 ca->oldest_gen = NULL;
2002 bch2_trans_exit(&trans);
2003 up_read(&c->gc_lock);
2004 mutex_unlock(&c->gc_gens_lock);
2008 static int bch2_gc_thread(void *arg)
2010 struct bch_fs *c = arg;
2011 struct io_clock *clock = &c->io_clock[WRITE];
2012 unsigned long last = atomic64_read(&clock->now);
2013 unsigned last_kick = atomic_read(&c->kick_gc);
2020 set_current_state(TASK_INTERRUPTIBLE);
2022 if (kthread_should_stop()) {
2023 __set_current_state(TASK_RUNNING);
2027 if (atomic_read(&c->kick_gc) != last_kick)
2030 if (c->btree_gc_periodic) {
2031 unsigned long next = last + c->capacity / 16;
2033 if (atomic64_read(&clock->now) >= next)
2036 bch2_io_clock_schedule_timeout(clock, next);
2043 __set_current_state(TASK_RUNNING);
2045 last = atomic64_read(&clock->now);
2046 last_kick = atomic_read(&c->kick_gc);
2049 * Full gc is currently incompatible with btree key cache:
2052 ret = bch2_gc(c, false, false);
2054 ret = bch2_gc_gens(c);
2057 bch_err(c, "btree gc failed: %i", ret);
2059 debug_check_no_locks_held();
2065 void bch2_gc_thread_stop(struct bch_fs *c)
2067 struct task_struct *p;
2070 c->gc_thread = NULL;
2078 int bch2_gc_thread_start(struct bch_fs *c)
2080 struct task_struct *p;
2085 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2087 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));