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 = -BCH_ERR_need_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 = -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(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
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(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
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_cmp(expected_start, cur->data->min_key) > 0 &&
278 BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
279 /* cur overwrites prev: */
281 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
282 cur->data->min_key) >= 0, 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_cmp(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_cmp(expected_start,
305 cur->data->max_key) >= 0, 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_cmp(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_cmp(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_cmp(k.k->p, b->data->min_key) < 0);
378 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
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);
575 if (c->opts.reconstruct_alloc ||
576 fsck_err_on(!g->gen_valid, c,
577 "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 (fsck_err_on(gen_cmp(p.ptr.gen, g->gen) > 0, c,
593 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
595 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
596 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
598 (printbuf_reset(&buf),
599 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
604 g->dirty_sectors = 0;
605 g->cached_sectors = 0;
606 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
612 if (fsck_err_on(gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
613 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
615 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
616 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
618 (printbuf_reset(&buf),
619 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
622 if (fsck_err_on(!p.ptr.cached &&
623 gen_cmp(p.ptr.gen, g->gen) < 0, c,
624 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
626 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
627 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
629 (printbuf_reset(&buf),
630 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
633 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
636 if (fsck_err_on(g->data_type &&
637 g->data_type != data_type, c,
638 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
640 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
641 bch2_data_types[g->data_type],
642 bch2_data_types[data_type],
643 (printbuf_reset(&buf),
644 bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
645 if (data_type == BCH_DATA_btree) {
646 g->data_type = data_type;
647 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
654 struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
656 if (fsck_err_on(!m || !m->alive, c,
657 "pointer to nonexistent stripe %llu\n"
660 (printbuf_reset(&buf),
661 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
664 if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
665 "pointer does not match stripe %llu\n"
668 (printbuf_reset(&buf),
669 bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
675 struct bkey_ptrs ptrs;
676 union bch_extent_entry *entry;
677 struct bch_extent_ptr *ptr;
681 bch_err(c, "cannot update btree roots yet");
686 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
688 bch_err(c, "%s: error allocating new key", __func__);
693 bkey_reassemble(new, *k);
697 * We don't want to drop btree node pointers - if the
698 * btree node isn't there anymore, the read path will
701 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
702 bkey_for_each_ptr(ptrs, ptr) {
703 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
704 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
709 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
710 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
711 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
712 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
715 (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
717 gen_cmp(ptr->gen, g->gen) < 0) ||
718 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
720 g->data_type != data_type);
723 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
724 bkey_extent_entry_for_each(ptrs, entry) {
725 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
726 struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
727 entry->stripe_ptr.idx);
728 union bch_extent_entry *next_ptr;
730 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
731 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
736 bch_err(c, "aieee, found stripe ptr with no data ptr");
740 if (!m || !m->alive ||
741 !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
744 bch2_bkey_extent_entry_drop(new, entry);
751 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
758 bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
760 if (c->opts.verbose) {
761 printbuf_reset(&buf);
762 bch2_bkey_val_to_text(&buf, c, *k);
763 bch_info(c, "updated %s", buf.buf);
765 printbuf_reset(&buf);
766 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
767 bch_info(c, "new key %s", buf.buf);
770 *k = bkey_i_to_s_c(new);
778 /* marking of btree keys/nodes: */
780 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
781 unsigned level, bool is_root,
785 struct bch_fs *c = trans->c;
786 struct bkey deleted = KEY(0, 0, 0);
787 struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
790 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
796 BUG_ON(bch2_journal_seq_verify &&
797 k->k->version.lo > atomic64_read(&c->journal.seq));
799 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
803 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
804 "key version number higher than recorded: %llu > %llu",
806 atomic64_read(&c->key_version)))
807 atomic64_set(&c->key_version, k->k->version.lo);
810 ret = commit_do(trans, NULL, NULL, 0,
811 bch2_mark_key(trans, old, *k, flags));
815 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
819 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
821 struct bch_fs *c = trans->c;
822 struct btree_node_iter iter;
823 struct bkey unpacked;
825 struct bkey_buf prev, cur;
828 if (!btree_node_type_needs_gc(btree_node_type(b)))
831 bch2_btree_node_iter_init_from_start(&iter, b);
832 bch2_bkey_buf_init(&prev);
833 bch2_bkey_buf_init(&cur);
834 bkey_init(&prev.k->k);
836 while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
837 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
842 bch2_btree_node_iter_advance(&iter, b);
845 bch2_bkey_buf_reassemble(&cur, c, k);
847 ret = bch2_gc_check_topology(c, b, &prev, cur,
848 bch2_btree_node_iter_end(&iter));
854 bch2_bkey_buf_exit(&cur, c);
855 bch2_bkey_buf_exit(&prev, c);
859 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
860 bool initial, bool metadata_only)
862 struct bch_fs *c = trans->c;
863 struct btree_iter iter;
865 unsigned depth = metadata_only ? 1 : 0;
868 gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
870 __for_each_btree_node(trans, iter, btree_id, POS_MIN,
871 0, depth, BTREE_ITER_PREFETCH, b, ret) {
872 bch2_verify_btree_nr_keys(b);
874 gc_pos_set(c, gc_pos_btree_node(b));
876 ret = btree_gc_mark_node(trans, b, initial);
880 bch2_trans_iter_exit(trans, &iter);
885 mutex_lock(&c->btree_root_lock);
886 b = c->btree_roots[btree_id].b;
887 if (!btree_node_fake(b)) {
888 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
890 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
893 gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
894 mutex_unlock(&c->btree_root_lock);
899 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
900 unsigned target_depth)
902 struct bch_fs *c = trans->c;
903 struct btree_and_journal_iter iter;
905 struct bkey_buf cur, prev;
906 struct printbuf buf = PRINTBUF;
909 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
910 bch2_bkey_buf_init(&prev);
911 bch2_bkey_buf_init(&cur);
912 bkey_init(&prev.k->k);
914 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
915 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
916 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
918 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
921 bch_err(c, "%s: error from bch2_gc_mark_key: %s",
922 __func__, bch2_err_str(ret));
927 bch2_bkey_buf_reassemble(&cur, c, k);
928 k = bkey_i_to_s_c(cur.k);
930 bch2_btree_and_journal_iter_advance(&iter);
932 ret = bch2_gc_check_topology(c, b,
934 !bch2_btree_and_journal_iter_peek(&iter).k);
938 bch2_btree_and_journal_iter_advance(&iter);
942 if (b->c.level > target_depth) {
943 bch2_btree_and_journal_iter_exit(&iter);
944 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
946 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
949 bch2_bkey_buf_reassemble(&cur, c, k);
950 bch2_btree_and_journal_iter_advance(&iter);
952 child = bch2_btree_node_get_noiter(trans, cur.k,
953 b->c.btree_id, b->c.level - 1,
955 ret = PTR_ERR_OR_ZERO(child);
958 bch2_topology_error(c);
964 "Unreadable btree node at btree %s level %u:\n"
966 bch2_btree_ids[b->c.btree_id],
968 (printbuf_reset(&buf),
969 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
970 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
971 ret = -BCH_ERR_need_topology_repair;
972 bch_info(c, "Halting mark and sweep to start topology repair pass");
975 /* Continue marking when opted to not
978 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
982 bch_err(c, "%s: error getting btree node: %s",
983 __func__, bch2_err_str(ret));
987 ret = bch2_gc_btree_init_recurse(trans, child,
989 six_unlock_read(&child->c.lock);
996 bch2_bkey_buf_exit(&cur, c);
997 bch2_bkey_buf_exit(&prev, c);
998 bch2_btree_and_journal_iter_exit(&iter);
1003 static int bch2_gc_btree_init(struct btree_trans *trans,
1004 enum btree_id btree_id,
1007 struct bch_fs *c = trans->c;
1009 unsigned target_depth = metadata_only ? 1 : 0;
1010 struct printbuf buf = PRINTBUF;
1013 b = c->btree_roots[btree_id].b;
1015 if (btree_node_fake(b))
1018 six_lock_read(&b->c.lock, NULL, NULL);
1019 printbuf_reset(&buf);
1020 bch2_bpos_to_text(&buf, b->data->min_key);
1021 if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
1022 "btree root with incorrect min_key: %s", buf.buf)) {
1023 bch_err(c, "repair unimplemented");
1024 ret = -BCH_ERR_fsck_repair_unimplemented;
1028 printbuf_reset(&buf);
1029 bch2_bpos_to_text(&buf, b->data->max_key);
1030 if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
1031 "btree root with incorrect max_key: %s", buf.buf)) {
1032 bch_err(c, "repair unimplemented");
1033 ret = -BCH_ERR_fsck_repair_unimplemented;
1037 if (b->c.level >= target_depth)
1038 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1041 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1043 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1047 six_unlock_read(&b->c.lock);
1050 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1051 printbuf_exit(&buf);
1055 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1057 return (int) btree_id_to_gc_phase(l) -
1058 (int) btree_id_to_gc_phase(r);
1061 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1063 struct btree_trans trans;
1064 enum btree_id ids[BTREE_ID_NR];
1068 bch2_trans_init(&trans, c, 0, 0);
1071 trans.is_initial_gc = true;
1073 for (i = 0; i < BTREE_ID_NR; i++)
1075 bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1077 for (i = 0; i < BTREE_ID_NR && !ret; i++)
1079 ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1080 : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1083 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1085 bch2_trans_exit(&trans);
1089 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1091 enum bch_data_type type,
1094 u64 b = sector_to_bucket(ca, start);
1098 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1100 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1101 gc_phase(GC_PHASE_SB), flags);
1104 } while (start < end);
1107 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1110 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1114 for (i = 0; i < layout->nr_superblocks; i++) {
1115 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1117 if (offset == BCH_SB_SECTOR)
1118 mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1119 BCH_DATA_sb, flags);
1121 mark_metadata_sectors(c, ca, offset,
1122 offset + (1 << layout->sb_max_size_bits),
1123 BCH_DATA_sb, flags);
1126 for (i = 0; i < ca->journal.nr; i++) {
1127 b = ca->journal.buckets[i];
1128 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1130 gc_phase(GC_PHASE_SB), flags);
1134 static void bch2_mark_superblocks(struct bch_fs *c)
1139 mutex_lock(&c->sb_lock);
1140 gc_pos_set(c, gc_phase(GC_PHASE_SB));
1142 for_each_online_member(ca, c, i)
1143 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1144 mutex_unlock(&c->sb_lock);
1148 /* Also see bch2_pending_btree_node_free_insert_done() */
1149 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1151 struct btree_update *as;
1152 struct pending_btree_node_free *d;
1154 mutex_lock(&c->btree_interior_update_lock);
1155 gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1157 for_each_pending_btree_node_free(c, as, d)
1158 if (d->index_update_done)
1159 bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1161 mutex_unlock(&c->btree_interior_update_lock);
1165 static void bch2_gc_free(struct bch_fs *c)
1170 genradix_free(&c->reflink_gc_table);
1171 genradix_free(&c->gc_stripes);
1173 for_each_member_device(ca, c, i) {
1174 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1175 sizeof(struct bucket_array) +
1176 ca->mi.nbuckets * sizeof(struct bucket));
1177 ca->buckets_gc = NULL;
1179 free_percpu(ca->usage_gc);
1180 ca->usage_gc = NULL;
1183 free_percpu(c->usage_gc);
1187 static int bch2_gc_done(struct bch_fs *c,
1188 bool initial, bool metadata_only)
1190 struct bch_dev *ca = NULL;
1191 struct printbuf buf = PRINTBUF;
1192 bool verify = !metadata_only &&
1193 !c->opts.reconstruct_alloc &&
1194 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1198 percpu_down_write(&c->mark_lock);
1200 #define copy_field(_f, _msg, ...) \
1201 if (dst->_f != src->_f && \
1203 fsck_err(c, _msg ": got %llu, should be %llu" \
1204 , ##__VA_ARGS__, dst->_f, src->_f))) \
1206 #define copy_stripe_field(_f, _msg, ...) \
1207 if (dst->_f != src->_f && \
1209 fsck_err(c, "stripe %zu has wrong "_msg \
1210 ": got %u, should be %u", \
1211 iter.pos, ##__VA_ARGS__, \
1212 dst->_f, src->_f))) \
1214 #define copy_dev_field(_f, _msg, ...) \
1215 copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1216 #define copy_fs_field(_f, _msg, ...) \
1217 copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1219 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1220 bch2_fs_usage_acc_to_base(c, i);
1222 for_each_member_device(ca, c, dev) {
1223 struct bch_dev_usage *dst = ca->usage_base;
1224 struct bch_dev_usage *src = (void *)
1225 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1228 copy_dev_field(buckets_ec, "buckets_ec");
1230 for (i = 0; i < BCH_DATA_NR; i++) {
1231 copy_dev_field(d[i].buckets, "%s buckets", bch2_data_types[i]);
1232 copy_dev_field(d[i].sectors, "%s sectors", bch2_data_types[i]);
1233 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1238 unsigned nr = fs_usage_u64s(c);
1239 struct bch_fs_usage *dst = c->usage_base;
1240 struct bch_fs_usage *src = (void *)
1241 bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1243 copy_fs_field(hidden, "hidden");
1244 copy_fs_field(btree, "btree");
1246 if (!metadata_only) {
1247 copy_fs_field(data, "data");
1248 copy_fs_field(cached, "cached");
1249 copy_fs_field(reserved, "reserved");
1250 copy_fs_field(nr_inodes,"nr_inodes");
1252 for (i = 0; i < BCH_REPLICAS_MAX; i++)
1253 copy_fs_field(persistent_reserved[i],
1254 "persistent_reserved[%i]", i);
1257 for (i = 0; i < c->replicas.nr; i++) {
1258 struct bch_replicas_entry *e =
1259 cpu_replicas_entry(&c->replicas, i);
1261 if (metadata_only &&
1262 (e->data_type == BCH_DATA_user ||
1263 e->data_type == BCH_DATA_cached))
1266 printbuf_reset(&buf);
1267 bch2_replicas_entry_to_text(&buf, e);
1269 copy_fs_field(replicas[i], "%s", buf.buf);
1273 #undef copy_fs_field
1274 #undef copy_dev_field
1275 #undef copy_stripe_field
1279 percpu_ref_put(&ca->ref);
1281 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1283 percpu_up_write(&c->mark_lock);
1284 printbuf_exit(&buf);
1288 static int bch2_gc_start(struct bch_fs *c,
1291 struct bch_dev *ca = NULL;
1294 BUG_ON(c->usage_gc);
1296 c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1297 sizeof(u64), GFP_KERNEL);
1299 bch_err(c, "error allocating c->usage_gc");
1303 for_each_member_device(ca, c, i) {
1304 BUG_ON(ca->buckets_gc);
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);
1314 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1315 ca->mi.nbuckets - ca->mi.first_bucket);
1321 /* returns true if not equal */
1322 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1323 struct bch_alloc_v4 r)
1325 return l.gen != r.gen ||
1326 l.oldest_gen != r.oldest_gen ||
1327 l.data_type != r.data_type ||
1328 l.dirty_sectors != r.dirty_sectors ||
1329 l.cached_sectors != r.cached_sectors ||
1330 l.stripe_redundancy != r.stripe_redundancy ||
1331 l.stripe != r.stripe;
1334 static int bch2_alloc_write_key(struct btree_trans *trans,
1335 struct btree_iter *iter,
1339 struct bch_fs *c = trans->c;
1340 struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1341 struct bucket gc, *b;
1342 struct bkey_i_alloc_v4 *a;
1343 struct bch_alloc_v4 old, new;
1344 enum bch_data_type type;
1347 if (bkey_cmp(iter->pos, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
1350 bch2_alloc_to_v4(k, &old);
1353 percpu_down_read(&c->mark_lock);
1354 b = gc_bucket(ca, iter->pos.offset);
1357 * b->data_type doesn't yet include need_discard & need_gc_gen states -
1360 type = __alloc_data_type(b->dirty_sectors,
1365 if (b->data_type != type) {
1366 struct bch_dev_usage *u;
1369 u = this_cpu_ptr(ca->usage_gc);
1370 u->d[b->data_type].buckets--;
1371 b->data_type = type;
1372 u->d[b->data_type].buckets++;
1377 percpu_up_read(&c->mark_lock);
1379 if (metadata_only &&
1380 gc.data_type != BCH_DATA_sb &&
1381 gc.data_type != BCH_DATA_journal &&
1382 gc.data_type != BCH_DATA_btree)
1385 if (gen_after(old.gen, gc.gen))
1388 #define copy_bucket_field(_f) \
1389 if (c->opts.reconstruct_alloc || \
1390 fsck_err_on(new._f != gc._f, c, \
1391 "bucket %llu:%llu gen %u data type %s has wrong " #_f \
1392 ": got %u, should be %u", \
1393 iter->pos.inode, iter->pos.offset, \
1395 bch2_data_types[gc.data_type], \
1399 copy_bucket_field(gen);
1400 copy_bucket_field(data_type);
1401 copy_bucket_field(dirty_sectors);
1402 copy_bucket_field(cached_sectors);
1403 copy_bucket_field(stripe_redundancy);
1404 copy_bucket_field(stripe);
1405 #undef copy_bucket_field
1407 if (!bch2_alloc_v4_cmp(old, new))
1410 a = bch2_alloc_to_v4_mut(trans, k);
1411 ret = PTR_ERR_OR_ZERO(a);
1418 * The trigger normally makes sure this is set, but we're not running
1421 if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1422 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1424 ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1429 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1431 struct btree_trans trans;
1432 struct btree_iter iter;
1438 bch2_trans_init(&trans, c, 0, 0);
1440 for_each_member_device(ca, c, i) {
1441 ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1442 POS(ca->dev_idx, ca->mi.first_bucket),
1443 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1444 NULL, NULL, BTREE_INSERT_LAZY_RW,
1445 bch2_alloc_write_key(&trans, &iter, k, metadata_only));
1448 bch_err(c, "error writing alloc info: %s", bch2_err_str(ret));
1449 percpu_ref_put(&ca->ref);
1454 bch2_trans_exit(&trans);
1455 return ret < 0 ? ret : 0;
1458 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1461 struct btree_trans trans;
1462 struct btree_iter iter;
1465 struct bch_alloc_v4 a;
1469 for_each_member_device(ca, c, i) {
1470 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1471 ca->mi.nbuckets * sizeof(struct bucket),
1472 GFP_KERNEL|__GFP_ZERO);
1474 percpu_ref_put(&ca->ref);
1475 bch_err(c, "error allocating ca->buckets[gc]");
1479 buckets->first_bucket = ca->mi.first_bucket;
1480 buckets->nbuckets = ca->mi.nbuckets;
1481 rcu_assign_pointer(ca->buckets_gc, buckets);
1484 bch2_trans_init(&trans, c, 0, 0);
1486 for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1487 BTREE_ITER_PREFETCH, k, ret) {
1488 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1489 g = gc_bucket(ca, k.k->p.offset);
1491 bch2_alloc_to_v4(k, &a);
1496 if (metadata_only &&
1497 (a.data_type == BCH_DATA_user ||
1498 a.data_type == BCH_DATA_cached ||
1499 a.data_type == BCH_DATA_parity)) {
1500 g->data_type = a.data_type;
1501 g->dirty_sectors = a.dirty_sectors;
1502 g->cached_sectors = a.cached_sectors;
1503 g->stripe = a.stripe;
1504 g->stripe_redundancy = a.stripe_redundancy;
1507 bch2_trans_iter_exit(&trans, &iter);
1509 bch2_trans_exit(&trans);
1512 bch_err(c, "error reading alloc info at gc start: %s", bch2_err_str(ret));
1517 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1522 for_each_member_device(ca, c, i) {
1523 struct bucket_array *buckets = gc_bucket_array(ca);
1526 for_each_bucket(g, buckets) {
1527 if (metadata_only &&
1528 (g->data_type == BCH_DATA_user ||
1529 g->data_type == BCH_DATA_cached ||
1530 g->data_type == BCH_DATA_parity))
1533 g->dirty_sectors = 0;
1534 g->cached_sectors = 0;
1539 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1540 struct btree_iter *iter,
1544 struct bch_fs *c = trans->c;
1545 const __le64 *refcount = bkey_refcount_c(k);
1546 struct printbuf buf = PRINTBUF;
1547 struct reflink_gc *r;
1553 while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1554 r->offset < k.k->p.offset)
1558 r->offset != k.k->p.offset ||
1559 r->size != k.k->size) {
1560 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1564 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1565 "reflink key has wrong refcount:\n"
1568 (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1572 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1573 ret = PTR_ERR_OR_ZERO(new);
1577 bkey_reassemble(new, k);
1580 new->k.type = KEY_TYPE_deleted;
1582 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1584 ret = bch2_trans_update(trans, iter, new, 0);
1587 printbuf_exit(&buf);
1591 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1593 struct btree_trans trans;
1594 struct btree_iter iter;
1602 bch2_trans_init(&trans, c, 0, 0);
1604 ret = for_each_btree_key_commit(&trans, iter,
1605 BTREE_ID_reflink, POS_MIN,
1606 BTREE_ITER_PREFETCH, k,
1607 NULL, NULL, BTREE_INSERT_NOFAIL,
1608 bch2_gc_write_reflink_key(&trans, &iter, k, &idx));
1610 c->reflink_gc_nr = 0;
1611 bch2_trans_exit(&trans);
1615 static int bch2_gc_reflink_start(struct bch_fs *c,
1618 struct btree_trans trans;
1619 struct btree_iter iter;
1621 struct reflink_gc *r;
1627 bch2_trans_init(&trans, c, 0, 0);
1628 c->reflink_gc_nr = 0;
1630 for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1631 BTREE_ITER_PREFETCH, k, ret) {
1632 const __le64 *refcount = bkey_refcount_c(k);
1637 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1644 r->offset = k.k->p.offset;
1645 r->size = k.k->size;
1648 bch2_trans_iter_exit(&trans, &iter);
1650 bch2_trans_exit(&trans);
1654 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1656 struct genradix_iter iter;
1657 struct reflink_gc *r;
1659 genradix_for_each(&c->reflink_gc_table, iter, r)
1663 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1664 struct btree_iter *iter,
1667 struct bch_fs *c = trans->c;
1668 struct printbuf buf = PRINTBUF;
1669 const struct bch_stripe *s;
1670 struct gc_stripe *m;
1674 if (k.k->type != KEY_TYPE_stripe)
1677 s = bkey_s_c_to_stripe(k).v;
1678 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1680 for (i = 0; i < s->nr_blocks; i++)
1681 if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1685 if (fsck_err_on(true, c,
1686 "stripe has wrong block sector count %u:\n"
1689 (printbuf_reset(&buf),
1690 bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1691 m ? m->block_sectors[i] : 0)) {
1692 struct bkey_i_stripe *new;
1694 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1695 ret = PTR_ERR_OR_ZERO(new);
1699 bkey_reassemble(&new->k_i, k);
1701 for (i = 0; i < new->v.nr_blocks; i++)
1702 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1704 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1707 printbuf_exit(&buf);
1711 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1713 struct btree_trans trans;
1714 struct btree_iter iter;
1721 bch2_trans_init(&trans, c, 0, 0);
1723 ret = for_each_btree_key_commit(&trans, iter,
1724 BTREE_ID_stripes, POS_MIN,
1725 BTREE_ITER_PREFETCH, k,
1726 NULL, NULL, BTREE_INSERT_NOFAIL,
1727 bch2_gc_write_stripes_key(&trans, &iter, k));
1729 bch2_trans_exit(&trans);
1733 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1735 genradix_free(&c->gc_stripes);
1739 * bch2_gc - walk _all_ references to buckets, and recompute them:
1741 * Order matters here:
1742 * - Concurrent GC relies on the fact that we have a total ordering for
1743 * everything that GC walks - see gc_will_visit_node(),
1744 * gc_will_visit_root()
1746 * - also, references move around in the course of index updates and
1747 * various other crap: everything needs to agree on the ordering
1748 * references are allowed to move around in - e.g., we're allowed to
1749 * start with a reference owned by an open_bucket (the allocator) and
1750 * move it to the btree, but not the reverse.
1752 * This is necessary to ensure that gc doesn't miss references that
1753 * move around - if references move backwards in the ordering GC
1754 * uses, GC could skip past them
1756 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1761 lockdep_assert_held(&c->state_lock);
1763 down_write(&c->gc_lock);
1765 bch2_btree_interior_updates_flush(c);
1767 ret = bch2_gc_start(c, metadata_only) ?:
1768 bch2_gc_alloc_start(c, metadata_only) ?:
1769 bch2_gc_reflink_start(c, metadata_only);
1773 gc_pos_set(c, gc_phase(GC_PHASE_START));
1775 bch2_mark_superblocks(c);
1777 if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1778 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1779 c->opts.fix_errors != FSCK_OPT_NO) {
1780 bch_info(c, "Starting topology repair pass");
1781 ret = bch2_repair_topology(c);
1784 bch_info(c, "Topology repair pass done");
1786 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1789 ret = bch2_gc_btrees(c, initial, metadata_only);
1791 if (ret == -BCH_ERR_need_topology_repair &&
1792 !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1793 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1794 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1795 SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, true);
1799 if (ret == -BCH_ERR_need_topology_repair)
1800 ret = -BCH_ERR_fsck_errors_not_fixed;
1806 bch2_mark_pending_btree_node_frees(c);
1810 if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1811 (!iter && bch2_test_restart_gc)) {
1813 bch_info(c, "Unable to fix bucket gens, looping");
1819 * XXX: make sure gens we fixed got saved
1821 bch_info(c, "Second GC pass needed, restarting:");
1822 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1823 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1825 bch2_gc_stripes_reset(c, metadata_only);
1826 bch2_gc_alloc_reset(c, metadata_only);
1827 bch2_gc_reflink_reset(c, metadata_only);
1829 /* flush fsck errors, reset counters */
1830 bch2_flush_fsck_errs(c);
1835 bch2_journal_block(&c->journal);
1837 ret = bch2_gc_stripes_done(c, metadata_only) ?:
1838 bch2_gc_reflink_done(c, metadata_only) ?:
1839 bch2_gc_alloc_done(c, metadata_only) ?:
1840 bch2_gc_done(c, initial, metadata_only);
1842 bch2_journal_unblock(&c->journal);
1845 percpu_down_write(&c->mark_lock);
1846 /* Indicates that gc is no longer in progress: */
1847 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1850 percpu_up_write(&c->mark_lock);
1852 up_write(&c->gc_lock);
1855 * At startup, allocations can happen directly instead of via the
1856 * allocator thread - issue wakeup in case they blocked on gc_lock:
1858 closure_wake_up(&c->freelist_wait);
1862 static int gc_btree_gens_key(struct btree_trans *trans,
1863 struct btree_iter *iter,
1866 struct bch_fs *c = trans->c;
1867 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1868 const struct bch_extent_ptr *ptr;
1872 percpu_down_read(&c->mark_lock);
1873 bkey_for_each_ptr(ptrs, ptr) {
1874 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1876 if (ptr_stale(ca, ptr) > 16) {
1877 percpu_up_read(&c->mark_lock);
1882 bkey_for_each_ptr(ptrs, ptr) {
1883 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1884 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1886 if (gen_after(*gen, ptr->gen))
1889 percpu_up_read(&c->mark_lock);
1892 u = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1893 ret = PTR_ERR_OR_ZERO(u);
1897 bkey_reassemble(u, k);
1899 bch2_extent_normalize(c, bkey_i_to_s(u));
1900 return bch2_trans_update(trans, iter, u, 0);
1903 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1906 struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1907 struct bch_alloc_v4 a;
1908 struct bkey_i_alloc_v4 *a_mut;
1911 bch2_alloc_to_v4(k, &a);
1913 if (a.oldest_gen == ca->oldest_gen[iter->pos.offset])
1916 a_mut = bch2_alloc_to_v4_mut(trans, k);
1917 ret = PTR_ERR_OR_ZERO(a_mut);
1921 a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1922 a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1924 return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1927 int bch2_gc_gens(struct bch_fs *c)
1929 struct btree_trans trans;
1930 struct btree_iter iter;
1933 u64 b, start_time = local_clock();
1938 * Ideally we would be using state_lock and not gc_lock here, but that
1939 * introduces a deadlock in the RO path - we currently take the state
1940 * lock at the start of going RO, thus the gc thread may get stuck:
1942 if (!mutex_trylock(&c->gc_gens_lock))
1945 trace_and_count(c, gc_gens_start, c);
1946 down_read(&c->gc_lock);
1947 bch2_trans_init(&trans, c, 0, 0);
1949 for_each_member_device(ca, c, i) {
1950 struct bucket_gens *gens;
1952 BUG_ON(ca->oldest_gen);
1954 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1955 if (!ca->oldest_gen) {
1956 percpu_ref_put(&ca->ref);
1961 gens = bucket_gens(ca);
1963 for (b = gens->first_bucket;
1964 b < gens->nbuckets; b++)
1965 ca->oldest_gen[b] = gens->b[b];
1968 for (i = 0; i < BTREE_ID_NR; i++)
1969 if (btree_type_has_ptrs(i)) {
1970 struct btree_iter iter;
1973 c->gc_gens_btree = i;
1974 c->gc_gens_pos = POS_MIN;
1975 ret = for_each_btree_key_commit(&trans, iter, i,
1977 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
1980 BTREE_INSERT_NOFAIL,
1981 gc_btree_gens_key(&trans, &iter, k));
1983 bch_err(c, "error recalculating oldest_gen: %s", bch2_err_str(ret));
1988 ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1990 BTREE_ITER_PREFETCH,
1993 BTREE_INSERT_NOFAIL,
1994 bch2_alloc_write_oldest_gen(&trans, &iter, k));
1996 bch_err(c, "error writing oldest_gen: %s", bch2_err_str(ret));
2000 c->gc_gens_btree = 0;
2001 c->gc_gens_pos = POS_MIN;
2005 bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2006 trace_and_count(c, gc_gens_end, c);
2008 for_each_member_device(ca, c, i) {
2009 kvfree(ca->oldest_gen);
2010 ca->oldest_gen = NULL;
2013 bch2_trans_exit(&trans);
2014 up_read(&c->gc_lock);
2015 mutex_unlock(&c->gc_gens_lock);
2019 static int bch2_gc_thread(void *arg)
2021 struct bch_fs *c = arg;
2022 struct io_clock *clock = &c->io_clock[WRITE];
2023 unsigned long last = atomic64_read(&clock->now);
2024 unsigned last_kick = atomic_read(&c->kick_gc);
2031 set_current_state(TASK_INTERRUPTIBLE);
2033 if (kthread_should_stop()) {
2034 __set_current_state(TASK_RUNNING);
2038 if (atomic_read(&c->kick_gc) != last_kick)
2041 if (c->btree_gc_periodic) {
2042 unsigned long next = last + c->capacity / 16;
2044 if (atomic64_read(&clock->now) >= next)
2047 bch2_io_clock_schedule_timeout(clock, next);
2054 __set_current_state(TASK_RUNNING);
2056 last = atomic64_read(&clock->now);
2057 last_kick = atomic_read(&c->kick_gc);
2060 * Full gc is currently incompatible with btree key cache:
2063 ret = bch2_gc(c, false, false);
2065 ret = bch2_gc_gens(c);
2068 bch_err(c, "btree gc failed: %s", bch2_err_str(ret));
2070 debug_check_no_locks_held();
2076 void bch2_gc_thread_stop(struct bch_fs *c)
2078 struct task_struct *p;
2081 c->gc_thread = NULL;
2089 int bch2_gc_thread_start(struct bch_fs *c)
2091 struct task_struct *p;
2096 p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2098 bch_err(c, "error creating gc thread: %s", bch2_err_str(PTR_ERR(p)));