1 // SPDX-License-Identifier: GPL-2.0
3 * Code for manipulating bucket marks for garbage collection.
5 * Copyright 2014 Datera, Inc.
9 #include "alloc_background.h"
12 #include "btree_update.h"
14 #include "buckets_waiting_for_journal.h"
22 #include "subvolume.h"
24 #include <linux/preempt.h>
25 #include <trace/events/bcachefs.h>
27 static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
28 enum bch_data_type data_type,
33 fs_usage->btree += sectors;
37 fs_usage->data += sectors;
40 fs_usage->cached += sectors;
47 void bch2_fs_usage_initialize(struct bch_fs *c)
49 struct bch_fs_usage *usage;
53 percpu_down_write(&c->mark_lock);
54 usage = c->usage_base;
56 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
57 bch2_fs_usage_acc_to_base(c, i);
59 for (i = 0; i < BCH_REPLICAS_MAX; i++)
60 usage->reserved += usage->persistent_reserved[i];
62 for (i = 0; i < c->replicas.nr; i++) {
63 struct bch_replicas_entry *e =
64 cpu_replicas_entry(&c->replicas, i);
66 fs_usage_data_type_to_base(usage, e->data_type, usage->replicas[i]);
69 for_each_member_device(ca, c, i) {
70 struct bch_dev_usage dev = bch2_dev_usage_read(ca);
72 usage->hidden += (dev.d[BCH_DATA_sb].buckets +
73 dev.d[BCH_DATA_journal].buckets) *
77 percpu_up_write(&c->mark_lock);
80 static inline struct bch_dev_usage *dev_usage_ptr(struct bch_dev *ca,
84 BUG_ON(!gc && !journal_seq);
86 return this_cpu_ptr(gc
88 : ca->usage[journal_seq & JOURNAL_BUF_MASK]);
91 struct bch_dev_usage bch2_dev_usage_read(struct bch_dev *ca)
93 struct bch_fs *c = ca->fs;
94 struct bch_dev_usage ret;
95 unsigned seq, i, u64s = dev_usage_u64s();
98 seq = read_seqcount_begin(&c->usage_lock);
99 memcpy(&ret, ca->usage_base, u64s * sizeof(u64));
100 for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
101 acc_u64s_percpu((u64 *) &ret, (u64 __percpu *) ca->usage[i], u64s);
102 } while (read_seqcount_retry(&c->usage_lock, seq));
107 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
108 unsigned journal_seq,
111 percpu_rwsem_assert_held(&c->mark_lock);
112 BUG_ON(!gc && !journal_seq);
114 return this_cpu_ptr(gc
116 : c->usage[journal_seq & JOURNAL_BUF_MASK]);
119 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
121 ssize_t offset = v - (u64 *) c->usage_base;
125 BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
126 percpu_rwsem_assert_held(&c->mark_lock);
129 seq = read_seqcount_begin(&c->usage_lock);
132 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
133 ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
134 } while (read_seqcount_retry(&c->usage_lock, seq));
139 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
141 struct bch_fs_usage_online *ret;
142 unsigned seq, i, u64s;
144 percpu_down_read(&c->mark_lock);
146 ret = kmalloc(sizeof(struct bch_fs_usage_online) +
147 sizeof(u64) * c->replicas.nr, GFP_NOFS);
148 if (unlikely(!ret)) {
149 percpu_up_read(&c->mark_lock);
153 ret->online_reserved = percpu_u64_get(c->online_reserved);
155 u64s = fs_usage_u64s(c);
157 seq = read_seqcount_begin(&c->usage_lock);
158 memcpy(&ret->u, c->usage_base, u64s * sizeof(u64));
159 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
160 acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i], u64s);
161 } while (read_seqcount_retry(&c->usage_lock, seq));
166 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
169 unsigned i, u64s = fs_usage_u64s(c);
171 BUG_ON(idx >= ARRAY_SIZE(c->usage));
174 write_seqcount_begin(&c->usage_lock);
176 acc_u64s_percpu((u64 *) c->usage_base,
177 (u64 __percpu *) c->usage[idx], u64s);
178 percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
181 for_each_member_device_rcu(ca, c, i, NULL) {
182 u64s = dev_usage_u64s();
184 acc_u64s_percpu((u64 *) ca->usage_base,
185 (u64 __percpu *) ca->usage[idx], u64s);
186 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
190 write_seqcount_end(&c->usage_lock);
194 void bch2_fs_usage_to_text(struct printbuf *out,
196 struct bch_fs_usage_online *fs_usage)
200 pr_buf(out, "capacity:\t\t\t%llu\n", c->capacity);
202 pr_buf(out, "hidden:\t\t\t\t%llu\n",
204 pr_buf(out, "data:\t\t\t\t%llu\n",
206 pr_buf(out, "cached:\t\t\t\t%llu\n",
208 pr_buf(out, "reserved:\t\t\t%llu\n",
209 fs_usage->u.reserved);
210 pr_buf(out, "nr_inodes:\t\t\t%llu\n",
211 fs_usage->u.nr_inodes);
212 pr_buf(out, "online reserved:\t\t%llu\n",
213 fs_usage->online_reserved);
216 i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
218 pr_buf(out, "%u replicas:\n", i + 1);
219 pr_buf(out, "\treserved:\t\t%llu\n",
220 fs_usage->u.persistent_reserved[i]);
223 for (i = 0; i < c->replicas.nr; i++) {
224 struct bch_replicas_entry *e =
225 cpu_replicas_entry(&c->replicas, i);
228 bch2_replicas_entry_to_text(out, e);
229 pr_buf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
233 static u64 reserve_factor(u64 r)
235 return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
238 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
240 return min(fs_usage->u.hidden +
243 reserve_factor(fs_usage->u.reserved +
244 fs_usage->online_reserved),
248 static struct bch_fs_usage_short
249 __bch2_fs_usage_read_short(struct bch_fs *c)
251 struct bch_fs_usage_short ret;
254 ret.capacity = c->capacity -
255 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
257 data = bch2_fs_usage_read_one(c, &c->usage_base->data) +
258 bch2_fs_usage_read_one(c, &c->usage_base->btree);
259 reserved = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
260 percpu_u64_get(c->online_reserved);
262 ret.used = min(ret.capacity, data + reserve_factor(reserved));
263 ret.free = ret.capacity - ret.used;
265 ret.nr_inodes = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
270 struct bch_fs_usage_short
271 bch2_fs_usage_read_short(struct bch_fs *c)
273 struct bch_fs_usage_short ret;
275 percpu_down_read(&c->mark_lock);
276 ret = __bch2_fs_usage_read_short(c);
277 percpu_up_read(&c->mark_lock);
282 static inline int is_unavailable_bucket(struct bucket_mark m)
284 return !is_available_bucket(m);
287 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
288 struct bucket_mark m)
290 return m.dirty_sectors
291 ? max(0, (int) ca->mi.bucket_size - (int) m.dirty_sectors)
295 static inline int is_stripe_data_bucket(struct bucket_mark m)
297 return m.stripe && m.data_type != BCH_DATA_parity;
300 static inline enum bch_data_type bucket_type(struct bucket_mark m)
302 return m.cached_sectors && !m.dirty_sectors
307 static inline void account_bucket(struct bch_fs_usage *fs_usage,
308 struct bch_dev_usage *dev_usage,
309 enum bch_data_type type,
312 if (type == BCH_DATA_sb || type == BCH_DATA_journal)
313 fs_usage->hidden += size;
315 dev_usage->d[type].buckets += nr;
318 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
319 struct bucket_mark old, struct bucket_mark new,
320 u64 journal_seq, bool gc)
322 struct bch_fs_usage *fs_usage;
323 struct bch_dev_usage *u;
326 fs_usage = fs_usage_ptr(c, journal_seq, gc);
327 u = dev_usage_ptr(ca, journal_seq, gc);
329 if (bucket_type(old))
330 account_bucket(fs_usage, u, bucket_type(old),
331 -1, -ca->mi.bucket_size);
333 if (bucket_type(new))
334 account_bucket(fs_usage, u, bucket_type(new),
335 1, ca->mi.bucket_size);
337 u->buckets_ec += (int) new.stripe - (int) old.stripe;
338 u->buckets_unavailable +=
339 is_unavailable_bucket(new) - is_unavailable_bucket(old);
341 u->d[old.data_type].sectors -= old.dirty_sectors;
342 u->d[new.data_type].sectors += new.dirty_sectors;
343 u->d[BCH_DATA_cached].sectors +=
344 (int) new.cached_sectors - (int) old.cached_sectors;
346 u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
347 u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
351 if (!is_available_bucket(old) && is_available_bucket(new))
352 bch2_wake_allocator(ca);
355 static inline int __update_replicas(struct bch_fs *c,
356 struct bch_fs_usage *fs_usage,
357 struct bch_replicas_entry *r,
360 int idx = bch2_replicas_entry_idx(c, r);
365 fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
366 fs_usage->replicas[idx] += sectors;
370 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
371 struct bch_replicas_entry *r, s64 sectors,
372 unsigned journal_seq, bool gc)
374 struct bch_fs_usage __percpu *fs_usage;
378 percpu_down_read(&c->mark_lock);
380 idx = bch2_replicas_entry_idx(c, r);
382 (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
383 fsck_err(c, "no replicas entry\n"
385 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf)))) {
386 percpu_up_read(&c->mark_lock);
387 ret = bch2_mark_replicas(c, r);
391 percpu_down_read(&c->mark_lock);
392 idx = bch2_replicas_entry_idx(c, r);
400 fs_usage = fs_usage_ptr(c, journal_seq, gc);
401 fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
402 fs_usage->replicas[idx] += sectors;
406 percpu_up_read(&c->mark_lock);
410 static inline int update_cached_sectors(struct bch_fs *c,
412 unsigned dev, s64 sectors,
413 unsigned journal_seq, bool gc)
415 struct bch_replicas_padded r;
417 bch2_replicas_entry_cached(&r.e, dev);
419 return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
422 static struct replicas_delta_list *
423 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
425 struct replicas_delta_list *d = trans->fs_usage_deltas;
426 unsigned new_size = d ? (d->size + more) * 2 : 128;
427 unsigned alloc_size = sizeof(*d) + new_size;
429 WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
431 if (!d || d->used + more > d->size) {
432 d = krealloc(d, alloc_size, GFP_NOIO|__GFP_ZERO);
434 BUG_ON(!d && alloc_size > REPLICAS_DELTA_LIST_MAX);
437 d = mempool_alloc(&trans->c->replicas_delta_pool, GFP_NOIO);
438 memset(d, 0, REPLICAS_DELTA_LIST_MAX);
440 if (trans->fs_usage_deltas)
441 memcpy(d, trans->fs_usage_deltas,
442 trans->fs_usage_deltas->size + sizeof(*d));
444 new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
445 kfree(trans->fs_usage_deltas);
449 trans->fs_usage_deltas = d;
454 static inline void update_replicas_list(struct btree_trans *trans,
455 struct bch_replicas_entry *r,
458 struct replicas_delta_list *d;
459 struct replicas_delta *n;
465 b = replicas_entry_bytes(r) + 8;
466 d = replicas_deltas_realloc(trans, b);
468 n = (void *) d->d + d->used;
470 memcpy(&n->r, r, replicas_entry_bytes(r));
471 bch2_replicas_entry_sort(&n->r);
475 static inline void update_cached_sectors_list(struct btree_trans *trans,
476 unsigned dev, s64 sectors)
478 struct bch_replicas_padded r;
480 bch2_replicas_entry_cached(&r.e, dev);
482 update_replicas_list(trans, &r.e, sectors);
485 void bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
486 size_t b, bool owned_by_allocator)
488 struct bucket *g = bucket(ca, b);
489 struct bucket_mark old, new;
491 old = bucket_cmpxchg(g, new, ({
492 new.owned_by_allocator = owned_by_allocator;
495 BUG_ON(owned_by_allocator == old.owned_by_allocator);
498 static int bch2_mark_alloc(struct btree_trans *trans,
499 struct bkey_s_c old, struct bkey_s_c new,
502 bool gc = flags & BTREE_TRIGGER_GC;
503 u64 journal_seq = trans->journal_res.seq;
504 struct bch_fs *c = trans->c;
505 struct bkey_alloc_unpacked old_u = bch2_alloc_unpack(old);
506 struct bkey_alloc_unpacked new_u = bch2_alloc_unpack(new);
509 struct bucket_mark old_m, m;
513 * alloc btree is read in by bch2_alloc_read, not gc:
515 if ((flags & BTREE_TRIGGER_GC) &&
516 !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
519 if ((flags & BTREE_TRIGGER_INSERT) &&
520 !old_u.data_type != !new_u.data_type &&
521 new.k->type == KEY_TYPE_alloc_v3) {
522 struct bch_alloc_v3 *v = (struct bch_alloc_v3 *) new.v;
523 u64 old_journal_seq = le64_to_cpu(v->journal_seq);
525 BUG_ON(!journal_seq);
528 * If the btree updates referring to a bucket weren't flushed
529 * before the bucket became empty again, then the we don't have
530 * to wait on a journal flush before we can reuse the bucket:
532 new_u.journal_seq = !new_u.data_type &&
533 (journal_seq == old_journal_seq ||
534 bch2_journal_noflush_seq(&c->journal, old_journal_seq))
536 v->journal_seq = cpu_to_le64(new_u.journal_seq);
539 if (old_u.data_type && !new_u.data_type && new_u.journal_seq) {
540 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
541 c->journal.flushed_seq_ondisk,
542 new_u.dev, new_u.bucket,
545 bch2_fs_fatal_error(c,
546 "error setting bucket_needs_journal_commit: %i", ret);
551 ca = bch_dev_bkey_exists(c, new_u.dev);
553 if (new_u.bucket >= ca->mi.nbuckets)
556 percpu_down_read(&c->mark_lock);
557 if (!gc && new_u.gen != old_u.gen)
558 *bucket_gen(ca, new_u.bucket) = new_u.gen;
560 g = __bucket(ca, new_u.bucket, gc);
562 old_m = bucket_cmpxchg(g, m, ({
564 m.data_type = new_u.data_type;
565 m.dirty_sectors = new_u.dirty_sectors;
566 m.cached_sectors = new_u.cached_sectors;
567 m.stripe = new_u.stripe != 0;
570 bch2_dev_usage_update(c, ca, old_m, m, journal_seq, gc);
572 g->io_time[READ] = new_u.read_time;
573 g->io_time[WRITE] = new_u.write_time;
574 g->oldest_gen = new_u.oldest_gen;
576 g->stripe = new_u.stripe;
577 g->stripe_redundancy = new_u.stripe_redundancy;
578 percpu_up_read(&c->mark_lock);
581 * need to know if we're getting called from the invalidate path or
585 if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
586 old_m.cached_sectors) {
587 ret = update_cached_sectors(c, new, ca->dev_idx,
588 -old_m.cached_sectors,
591 bch2_fs_fatal_error(c, "bch2_mark_alloc(): no replicas entry while updating cached sectors");
595 trace_invalidate(ca, bucket_to_sector(ca, new_u.bucket),
596 old_m.cached_sectors);
602 #define checked_add(a, b) \
604 unsigned _res = (unsigned) (a) + (b); \
605 bool overflow = _res > U16_MAX; \
612 void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
613 size_t b, enum bch_data_type data_type,
614 unsigned sectors, struct gc_pos pos,
618 struct bucket_mark old, new;
621 BUG_ON(!(flags & BTREE_TRIGGER_GC));
622 BUG_ON(data_type != BCH_DATA_sb &&
623 data_type != BCH_DATA_journal);
626 * Backup superblock might be past the end of our normal usable space:
628 if (b >= ca->mi.nbuckets)
631 percpu_down_read(&c->mark_lock);
632 g = gc_bucket(ca, b);
633 old = bucket_cmpxchg(g, new, ({
634 new.data_type = data_type;
635 overflow = checked_add(new.dirty_sectors, sectors);
638 bch2_fs_inconsistent_on(old.data_type &&
639 old.data_type != data_type, c,
640 "different types of data in same bucket: %s, %s",
641 bch2_data_types[old.data_type],
642 bch2_data_types[data_type]);
644 bch2_fs_inconsistent_on(overflow, c,
645 "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > U16_MAX",
646 ca->dev_idx, b, new.gen,
647 bch2_data_types[old.data_type ?: data_type],
648 old.dirty_sectors, sectors);
650 bch2_dev_usage_update(c, ca, old, new, 0, true);
651 percpu_up_read(&c->mark_lock);
654 static s64 ptr_disk_sectors(s64 sectors, struct extent_ptr_decoded p)
656 EBUG_ON(sectors < 0);
658 return p.crc.compression_type &&
659 p.crc.compression_type != BCH_COMPRESSION_TYPE_incompressible
660 ? DIV_ROUND_UP_ULL(sectors * p.crc.compressed_size,
661 p.crc.uncompressed_size)
665 static int check_bucket_ref(struct bch_fs *c,
667 const struct bch_extent_ptr *ptr,
668 s64 sectors, enum bch_data_type ptr_data_type,
669 u8 b_gen, u8 bucket_data_type,
670 u16 dirty_sectors, u16 cached_sectors)
672 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
673 size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
674 u16 bucket_sectors = !ptr->cached
679 if (gen_after(ptr->gen, b_gen)) {
680 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
681 "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
683 ptr->dev, bucket_nr, b_gen,
684 bch2_data_types[bucket_data_type ?: ptr_data_type],
686 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
690 if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
691 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
692 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
694 ptr->dev, bucket_nr, b_gen,
695 bch2_data_types[bucket_data_type ?: ptr_data_type],
697 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
701 if (b_gen != ptr->gen && !ptr->cached) {
702 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
703 "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
705 ptr->dev, bucket_nr, b_gen,
706 *bucket_gen(ca, bucket_nr),
707 bch2_data_types[bucket_data_type ?: ptr_data_type],
709 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
713 if (b_gen != ptr->gen)
716 if (bucket_data_type && ptr_data_type &&
717 bucket_data_type != ptr_data_type) {
718 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
719 "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
721 ptr->dev, bucket_nr, b_gen,
722 bch2_data_types[bucket_data_type],
723 bch2_data_types[ptr_data_type],
724 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
728 if ((unsigned) (bucket_sectors + sectors) > U16_MAX) {
729 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
730 "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
732 ptr->dev, bucket_nr, b_gen,
733 bch2_data_types[bucket_data_type ?: ptr_data_type],
734 bucket_sectors, sectors,
735 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
742 static int mark_stripe_bucket(struct btree_trans *trans,
747 struct bch_fs *c = trans->c;
748 u64 journal_seq = trans->journal_res.seq;
749 const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
750 unsigned nr_data = s->nr_blocks - s->nr_redundant;
751 bool parity = ptr_idx >= nr_data;
752 enum bch_data_type data_type = parity ? BCH_DATA_parity : 0;
753 s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
754 const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
755 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
757 struct bucket_mark new, old;
761 BUG_ON(!(flags & BTREE_TRIGGER_GC));
763 /* * XXX doesn't handle deletion */
765 percpu_down_read(&c->mark_lock);
766 g = PTR_GC_BUCKET(ca, ptr);
768 if (g->mark.dirty_sectors ||
769 (g->stripe && g->stripe != k.k->p.offset)) {
770 bch2_fs_inconsistent(c,
771 "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
772 ptr->dev, PTR_BUCKET_NR(ca, ptr), g->mark.gen,
773 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
778 old = bucket_cmpxchg(g, new, ({
779 ret = check_bucket_ref(c, k, ptr, sectors, data_type,
780 new.gen, new.data_type,
781 new.dirty_sectors, new.cached_sectors);
785 new.dirty_sectors += sectors;
787 new.data_type = data_type;
792 g->stripe = k.k->p.offset;
793 g->stripe_redundancy = s->nr_redundant;
795 bch2_dev_usage_update(c, ca, old, new, journal_seq, true);
797 percpu_up_read(&c->mark_lock);
802 static int __mark_pointer(struct btree_trans *trans,
804 const struct bch_extent_ptr *ptr,
805 s64 sectors, enum bch_data_type ptr_data_type,
806 u8 bucket_gen, u8 *bucket_data_type,
807 u16 *dirty_sectors, u16 *cached_sectors)
809 u16 *dst_sectors = !ptr->cached
812 int ret = check_bucket_ref(trans->c, k, ptr, sectors, ptr_data_type,
813 bucket_gen, *bucket_data_type,
814 *dirty_sectors, *cached_sectors);
819 *dst_sectors += sectors;
820 *bucket_data_type = *dirty_sectors || *cached_sectors
825 static int bch2_mark_pointer(struct btree_trans *trans,
827 struct extent_ptr_decoded p,
828 s64 sectors, enum bch_data_type data_type,
831 u64 journal_seq = trans->journal_res.seq;
832 struct bch_fs *c = trans->c;
833 struct bucket_mark old, new;
834 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
840 BUG_ON(!(flags & BTREE_TRIGGER_GC));
842 percpu_down_read(&c->mark_lock);
843 g = PTR_GC_BUCKET(ca, &p.ptr);
845 v = atomic64_read(&g->_mark.v);
847 new.v.counter = old.v.counter = v;
848 bucket_data_type = new.data_type;
850 ret = __mark_pointer(trans, k, &p.ptr, sectors,
854 &new.cached_sectors);
858 new.data_type = bucket_data_type;
860 if (flags & BTREE_TRIGGER_NOATOMIC) {
864 } while ((v = atomic64_cmpxchg(&g->_mark.v,
866 new.v.counter)) != old.v.counter);
868 bch2_dev_usage_update(c, ca, old, new, journal_seq, true);
870 percpu_up_read(&c->mark_lock);
875 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
877 struct bch_extent_stripe_ptr p,
878 enum bch_data_type data_type,
882 struct bch_fs *c = trans->c;
883 struct bch_replicas_padded r;
886 BUG_ON(!(flags & BTREE_TRIGGER_GC));
888 m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
890 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
895 spin_lock(&c->ec_stripes_heap_lock);
897 if (!m || !m->alive) {
898 spin_unlock(&c->ec_stripes_heap_lock);
899 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
901 bch2_inconsistent_error(c);
905 m->block_sectors[p.block] += sectors;
908 spin_unlock(&c->ec_stripes_heap_lock);
910 r.e.data_type = data_type;
911 update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
916 static int bch2_mark_extent(struct btree_trans *trans,
917 struct bkey_s_c old, struct bkey_s_c new,
920 u64 journal_seq = trans->journal_res.seq;
921 struct bch_fs *c = trans->c;
922 struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
923 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
924 const union bch_extent_entry *entry;
925 struct extent_ptr_decoded p;
926 struct bch_replicas_padded r;
927 enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
930 s64 sectors = bkey_is_btree_ptr(k.k)
933 s64 dirty_sectors = 0;
937 BUG_ON(!(flags & BTREE_TRIGGER_GC));
939 r.e.data_type = data_type;
943 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
944 s64 disk_sectors = ptr_disk_sectors(sectors, p);
946 if (flags & BTREE_TRIGGER_OVERWRITE)
947 disk_sectors = -disk_sectors;
949 ret = bch2_mark_pointer(trans, k, p, disk_sectors,
958 ret = update_cached_sectors(c, k, p.ptr.dev,
959 disk_sectors, journal_seq, true);
961 bch2_fs_fatal_error(c, "bch2_mark_extent(): no replicas entry while updating cached sectors");
965 } else if (!p.has_ec) {
966 dirty_sectors += disk_sectors;
967 r.e.devs[r.e.nr_devs++] = p.ptr.dev;
969 ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
970 disk_sectors, flags);
975 * There may be other dirty pointers in this extent, but
976 * if so they're not required for mounting if we have an
977 * erasure coded pointer in this extent:
984 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
988 bch2_bkey_val_to_text(&PBUF(buf), c, k);
989 bch2_fs_fatal_error(c, "no replicas entry for %s", buf);
997 static int bch2_mark_stripe(struct btree_trans *trans,
998 struct bkey_s_c old, struct bkey_s_c new,
1001 bool gc = flags & BTREE_TRIGGER_GC;
1002 u64 journal_seq = trans->journal_res.seq;
1003 struct bch_fs *c = trans->c;
1004 u64 idx = new.k->p.offset;
1005 const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1006 ? bkey_s_c_to_stripe(old).v : NULL;
1007 const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1008 ? bkey_s_c_to_stripe(new).v : NULL;
1012 BUG_ON(gc && old_s);
1015 struct stripe *m = genradix_ptr(&c->stripes, idx);
1017 if (!m || (old_s && !m->alive)) {
1018 char buf1[200], buf2[200];
1020 bch2_bkey_val_to_text(&PBUF(buf1), c, old);
1021 bch2_bkey_val_to_text(&PBUF(buf2), c, new);
1022 bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1024 "new %s", idx, buf1, buf2);
1025 bch2_inconsistent_error(c);
1030 spin_lock(&c->ec_stripes_heap_lock);
1031 bch2_stripes_heap_del(c, m, idx);
1032 spin_unlock(&c->ec_stripes_heap_lock);
1034 memset(m, 0, sizeof(*m));
1037 m->sectors = le16_to_cpu(new_s->sectors);
1038 m->algorithm = new_s->algorithm;
1039 m->nr_blocks = new_s->nr_blocks;
1040 m->nr_redundant = new_s->nr_redundant;
1041 m->blocks_nonempty = 0;
1043 for (i = 0; i < new_s->nr_blocks; i++)
1044 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1046 spin_lock(&c->ec_stripes_heap_lock);
1047 bch2_stripes_heap_update(c, m, idx);
1048 spin_unlock(&c->ec_stripes_heap_lock);
1051 struct gc_stripe *m =
1052 genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1055 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1060 * This will be wrong when we bring back runtime gc: we should
1061 * be unmarking the old key and then marking the new key
1064 m->sectors = le16_to_cpu(new_s->sectors);
1065 m->nr_blocks = new_s->nr_blocks;
1066 m->nr_redundant = new_s->nr_redundant;
1068 for (i = 0; i < new_s->nr_blocks; i++)
1069 m->ptrs[i] = new_s->ptrs[i];
1071 bch2_bkey_to_replicas(&m->r.e, new);
1074 * gc recalculates this field from stripe ptr
1077 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1079 for (i = 0; i < new_s->nr_blocks; i++) {
1080 ret = mark_stripe_bucket(trans, new, i, flags);
1085 ret = update_replicas(c, new, &m->r.e,
1086 ((s64) m->sectors * m->nr_redundant),
1091 bch2_bkey_val_to_text(&PBUF(buf), c, new);
1092 bch2_fs_fatal_error(c, "no replicas entry for %s", buf);
1100 static int bch2_mark_inode(struct btree_trans *trans,
1101 struct bkey_s_c old, struct bkey_s_c new,
1104 struct bch_fs *c = trans->c;
1105 struct bch_fs_usage __percpu *fs_usage;
1106 u64 journal_seq = trans->journal_res.seq;
1108 if (flags & BTREE_TRIGGER_INSERT) {
1109 struct bch_inode_v2 *v = (struct bch_inode_v2 *) new.v;
1111 BUG_ON(!journal_seq);
1112 BUG_ON(new.k->type != KEY_TYPE_inode_v2);
1114 v->bi_journal_seq = cpu_to_le64(journal_seq);
1117 if (flags & BTREE_TRIGGER_GC) {
1118 percpu_down_read(&c->mark_lock);
1121 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1122 fs_usage->nr_inodes += bkey_is_inode(new.k);
1123 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1126 percpu_up_read(&c->mark_lock);
1131 static int bch2_mark_reservation(struct btree_trans *trans,
1132 struct bkey_s_c old, struct bkey_s_c new,
1135 struct bch_fs *c = trans->c;
1136 struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1137 struct bch_fs_usage __percpu *fs_usage;
1138 unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1139 s64 sectors = (s64) k.k->size;
1141 BUG_ON(!(flags & BTREE_TRIGGER_GC));
1143 if (flags & BTREE_TRIGGER_OVERWRITE)
1145 sectors *= replicas;
1147 percpu_down_read(&c->mark_lock);
1150 fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1151 replicas = clamp_t(unsigned, replicas, 1,
1152 ARRAY_SIZE(fs_usage->persistent_reserved));
1154 fs_usage->reserved += sectors;
1155 fs_usage->persistent_reserved[replicas - 1] += sectors;
1158 percpu_up_read(&c->mark_lock);
1163 static s64 __bch2_mark_reflink_p(struct bch_fs *c, struct bkey_s_c_reflink_p p,
1164 u64 *idx, unsigned flags, size_t r_idx)
1166 struct reflink_gc *r;
1167 int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1170 if (r_idx >= c->reflink_gc_nr)
1173 r = genradix_ptr(&c->reflink_gc_table, r_idx);
1174 if (*idx < r->offset - r->size)
1177 BUG_ON((s64) r->refcount + add < 0);
1187 * XXX: we're replacing the entire reflink pointer with an error
1188 * key, we should just be replacing the part that was missing:
1190 if (fsck_err(c, "%llu:%llu len %u points to nonexistent indirect extent %llu",
1191 p.k->p.inode, p.k->p.offset, p.k->size, *idx)) {
1192 struct bkey_i_error new;
1195 new.k.type = KEY_TYPE_error;
1197 new.k.size = p.k->size;
1198 ret = bch2_journal_key_insert(c, BTREE_ID_extents, 0, &new.k_i);
1204 static int bch2_mark_reflink_p(struct btree_trans *trans,
1205 struct bkey_s_c old, struct bkey_s_c new,
1208 struct bch_fs *c = trans->c;
1209 struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1210 struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1211 struct reflink_gc *ref;
1213 u64 idx = le64_to_cpu(p.v->idx);
1214 u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1217 BUG_ON(!(flags & BTREE_TRIGGER_GC));
1219 if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1220 idx -= le32_to_cpu(p.v->front_pad);
1221 end += le32_to_cpu(p.v->back_pad);
1225 r = c->reflink_gc_nr;
1227 m = l + (r - l) / 2;
1229 ref = genradix_ptr(&c->reflink_gc_table, m);
1230 if (ref->offset <= idx)
1236 while (idx < end && !ret)
1237 ret = __bch2_mark_reflink_p(c, p, &idx, flags, l++);
1242 int bch2_mark_key(struct btree_trans *trans,
1243 struct bkey_s_c old,
1244 struct bkey_s_c new,
1247 struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1249 switch (k.k->type) {
1250 case KEY_TYPE_alloc:
1251 case KEY_TYPE_alloc_v2:
1252 case KEY_TYPE_alloc_v3:
1253 return bch2_mark_alloc(trans, old, new, flags);
1254 case KEY_TYPE_btree_ptr:
1255 case KEY_TYPE_btree_ptr_v2:
1256 case KEY_TYPE_extent:
1257 case KEY_TYPE_reflink_v:
1258 return bch2_mark_extent(trans, old, new, flags);
1259 case KEY_TYPE_stripe:
1260 return bch2_mark_stripe(trans, old, new, flags);
1261 case KEY_TYPE_inode:
1262 case KEY_TYPE_inode_v2:
1263 return bch2_mark_inode(trans, old, new, flags);
1264 case KEY_TYPE_reservation:
1265 return bch2_mark_reservation(trans, old, new, flags);
1266 case KEY_TYPE_reflink_p:
1267 return bch2_mark_reflink_p(trans, old, new, flags);
1268 case KEY_TYPE_snapshot:
1269 return bch2_mark_snapshot(trans, old, new, flags);
1275 int bch2_mark_update(struct btree_trans *trans, struct btree_path *path,
1276 struct bkey_i *new, unsigned flags)
1278 struct bkey _deleted = KEY(0, 0, 0);
1279 struct bkey_s_c deleted = (struct bkey_s_c) { &_deleted, NULL };
1280 struct bkey_s_c old;
1281 struct bkey unpacked;
1284 _deleted.p = path->pos;
1286 if (unlikely(flags & BTREE_TRIGGER_NORUN))
1289 if (!btree_node_type_needs_gc(path->btree_id))
1292 old = bch2_btree_path_peek_slot(path, &unpacked);
1294 if (old.k->type == new->k.type &&
1295 ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
1296 ret = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
1297 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
1299 ret = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
1300 BTREE_TRIGGER_INSERT|flags) ?:
1301 bch2_mark_key(trans, old, deleted,
1302 BTREE_TRIGGER_OVERWRITE|flags);
1308 static noinline __cold
1309 void fs_usage_apply_warn(struct btree_trans *trans,
1310 unsigned disk_res_sectors,
1311 s64 should_not_have_added)
1313 struct bch_fs *c = trans->c;
1314 struct btree_insert_entry *i;
1317 bch_err(c, "disk usage increased %lli more than %u sectors reserved",
1318 should_not_have_added, disk_res_sectors);
1320 trans_for_each_update(trans, i) {
1321 pr_err("while inserting");
1322 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
1324 pr_err("overlapping with");
1328 struct bkey_s_c k = bch2_btree_path_peek_slot(i->path, &u);
1330 bch2_bkey_val_to_text(&PBUF(buf), c, k);
1333 struct bkey_cached *ck = (void *) i->path->l[0].b;
1336 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(ck->k));
1344 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1345 struct replicas_delta_list *deltas)
1347 struct bch_fs *c = trans->c;
1348 static int warned_disk_usage = 0;
1350 unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1351 struct replicas_delta *d = deltas->d, *d2;
1352 struct replicas_delta *top = (void *) deltas->d + deltas->used;
1353 struct bch_fs_usage *dst;
1354 s64 added = 0, should_not_have_added;
1357 percpu_down_read(&c->mark_lock);
1359 dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1361 for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1362 switch (d->r.data_type) {
1363 case BCH_DATA_btree:
1365 case BCH_DATA_parity:
1369 if (__update_replicas(c, dst, &d->r, d->delta))
1373 dst->nr_inodes += deltas->nr_inodes;
1375 for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1376 added += deltas->persistent_reserved[i];
1377 dst->reserved += deltas->persistent_reserved[i];
1378 dst->persistent_reserved[i] += deltas->persistent_reserved[i];
1382 * Not allowed to reduce sectors_available except by getting a
1385 should_not_have_added = added - (s64) disk_res_sectors;
1386 if (unlikely(should_not_have_added > 0)) {
1387 u64 old, new, v = atomic64_read(&c->sectors_available);
1391 new = max_t(s64, 0, old - should_not_have_added);
1392 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1395 added -= should_not_have_added;
1400 trans->disk_res->sectors -= added;
1401 this_cpu_sub(*c->online_reserved, added);
1405 percpu_up_read(&c->mark_lock);
1407 if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1408 fs_usage_apply_warn(trans, disk_res_sectors, should_not_have_added);
1411 /* revert changes: */
1412 for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1413 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1416 percpu_up_read(&c->mark_lock);
1422 static int bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
1423 const struct bch_extent_ptr *ptr,
1424 struct bkey_alloc_unpacked *u)
1426 struct bch_fs *c = trans->c;
1427 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1431 bch2_trans_iter_init(trans, iter, BTREE_ID_alloc,
1432 POS(ptr->dev, PTR_BUCKET_NR(ca, ptr)),
1433 BTREE_ITER_WITH_UPDATES|
1436 k = bch2_btree_iter_peek_slot(iter);
1439 bch2_trans_iter_exit(trans, iter);
1443 *u = bch2_alloc_unpack(k);
1447 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1448 struct bkey_s_c k, struct extent_ptr_decoded p,
1449 s64 sectors, enum bch_data_type data_type)
1451 struct btree_iter iter;
1452 struct bkey_alloc_unpacked u;
1455 ret = bch2_trans_start_alloc_update(trans, &iter, &p.ptr, &u);
1459 ret = __mark_pointer(trans, k, &p.ptr, sectors, data_type,
1460 u.gen, &u.data_type,
1461 &u.dirty_sectors, &u.cached_sectors);
1465 ret = bch2_alloc_write(trans, &iter, &u, 0);
1469 bch2_trans_iter_exit(trans, &iter);
1473 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1474 struct extent_ptr_decoded p,
1475 s64 sectors, enum bch_data_type data_type)
1477 struct bch_fs *c = trans->c;
1478 struct btree_iter iter;
1480 struct bkey_i_stripe *s;
1481 struct bch_replicas_padded r;
1484 bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1486 BTREE_ITER_WITH_UPDATES);
1487 k = bch2_btree_iter_peek_slot(&iter);
1492 if (k.k->type != KEY_TYPE_stripe) {
1493 bch2_fs_inconsistent(c,
1494 "pointer to nonexistent stripe %llu",
1496 bch2_inconsistent_error(c);
1501 if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
1502 bch2_fs_inconsistent(c,
1503 "stripe pointer doesn't match stripe %llu",
1509 s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1510 ret = PTR_ERR_OR_ZERO(s);
1514 bkey_reassemble(&s->k_i, k);
1515 stripe_blockcount_set(&s->v, p.ec.block,
1516 stripe_blockcount_get(&s->v, p.ec.block) +
1519 ret = bch2_trans_update(trans, &iter, &s->k_i, 0);
1523 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1524 r.e.data_type = data_type;
1525 update_replicas_list(trans, &r.e, sectors);
1527 bch2_trans_iter_exit(trans, &iter);
1531 static int bch2_trans_mark_extent(struct btree_trans *trans,
1532 struct bkey_s_c k, unsigned flags)
1534 struct bch_fs *c = trans->c;
1535 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1536 const union bch_extent_entry *entry;
1537 struct extent_ptr_decoded p;
1538 struct bch_replicas_padded r;
1539 enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1542 s64 sectors = bkey_is_btree_ptr(k.k)
1545 s64 dirty_sectors = 0;
1549 r.e.data_type = data_type;
1551 r.e.nr_required = 1;
1553 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1554 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1556 if (flags & BTREE_TRIGGER_OVERWRITE)
1557 disk_sectors = -disk_sectors;
1559 ret = bch2_trans_mark_pointer(trans, k, p,
1560 disk_sectors, data_type);
1568 update_cached_sectors_list(trans, p.ptr.dev,
1570 } else if (!p.has_ec) {
1571 dirty_sectors += disk_sectors;
1572 r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1574 ret = bch2_trans_mark_stripe_ptr(trans, p,
1575 disk_sectors, data_type);
1579 r.e.nr_required = 0;
1584 update_replicas_list(trans, &r.e, dirty_sectors);
1589 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1590 struct bkey_s_c_stripe s,
1591 unsigned idx, bool deleting)
1593 struct bch_fs *c = trans->c;
1594 const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1595 struct btree_iter iter;
1596 struct bkey_alloc_unpacked u;
1597 enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1598 ? BCH_DATA_parity : 0;
1599 s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1605 ret = bch2_trans_start_alloc_update(trans, &iter, ptr, &u);
1609 ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
1611 u.dirty_sectors, u.cached_sectors);
1616 if (bch2_fs_inconsistent_on(u.stripe ||
1617 u.stripe_redundancy, c,
1618 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1619 iter.pos.inode, iter.pos.offset, u.gen,
1620 bch2_data_types[u.data_type],
1622 u.stripe, s.k->p.offset)) {
1627 if (bch2_fs_inconsistent_on(data_type && u.dirty_sectors, c,
1628 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1629 iter.pos.inode, iter.pos.offset, u.gen,
1630 bch2_data_types[u.data_type],
1637 u.stripe = s.k->p.offset;
1638 u.stripe_redundancy = s.v->nr_redundant;
1640 if (bch2_fs_inconsistent_on(u.stripe != s.k->p.offset ||
1641 u.stripe_redundancy != s.v->nr_redundant, c,
1642 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1643 iter.pos.inode, iter.pos.offset, u.gen,
1644 s.k->p.offset, u.stripe)) {
1650 u.stripe_redundancy = 0;
1653 u.dirty_sectors += sectors;
1655 u.data_type = !deleting ? data_type : 0;
1657 ret = bch2_alloc_write(trans, &iter, &u, 0);
1661 bch2_trans_iter_exit(trans, &iter);
1665 static int bch2_trans_mark_stripe(struct btree_trans *trans,
1666 struct bkey_s_c old, struct bkey_s_c new,
1669 struct bkey_s_c_stripe old_s = { .k = NULL };
1670 struct bkey_s_c_stripe new_s = { .k = NULL };
1671 struct bch_replicas_padded r;
1672 unsigned i, nr_blocks;
1675 if (old.k->type == KEY_TYPE_stripe)
1676 old_s = bkey_s_c_to_stripe(old);
1677 if (new.k->type == KEY_TYPE_stripe)
1678 new_s = bkey_s_c_to_stripe(new);
1681 * If the pointers aren't changing, we don't need to do anything:
1683 if (new_s.k && old_s.k &&
1684 new_s.v->nr_blocks == old_s.v->nr_blocks &&
1685 new_s.v->nr_redundant == old_s.v->nr_redundant &&
1686 !memcmp(old_s.v->ptrs, new_s.v->ptrs,
1687 new_s.v->nr_blocks * sizeof(struct bch_extent_ptr)))
1690 BUG_ON(new_s.k && old_s.k &&
1691 (new_s.v->nr_blocks != old_s.v->nr_blocks ||
1692 new_s.v->nr_redundant != old_s.v->nr_redundant));
1694 nr_blocks = new_s.k ? new_s.v->nr_blocks : old_s.v->nr_blocks;
1697 s64 sectors = le16_to_cpu(new_s.v->sectors);
1699 bch2_bkey_to_replicas(&r.e, new);
1700 update_replicas_list(trans, &r.e, sectors * new_s.v->nr_redundant);
1704 s64 sectors = -((s64) le16_to_cpu(old_s.v->sectors));
1706 bch2_bkey_to_replicas(&r.e, old);
1707 update_replicas_list(trans, &r.e, sectors * old_s.v->nr_redundant);
1710 for (i = 0; i < nr_blocks; i++) {
1711 if (new_s.k && old_s.k &&
1712 !memcmp(&new_s.v->ptrs[i],
1714 sizeof(new_s.v->ptrs[i])))
1718 ret = bch2_trans_mark_stripe_bucket(trans, new_s, i, false);
1724 ret = bch2_trans_mark_stripe_bucket(trans, old_s, i, true);
1733 static int bch2_trans_mark_inode(struct btree_trans *trans,
1734 struct bkey_s_c old,
1735 struct bkey_s_c new,
1738 int nr = bkey_is_inode(new.k) - bkey_is_inode(old.k);
1741 struct replicas_delta_list *d =
1742 replicas_deltas_realloc(trans, 0);
1749 static int bch2_trans_mark_reservation(struct btree_trans *trans,
1750 struct bkey_s_c k, unsigned flags)
1752 unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1753 s64 sectors = (s64) k.k->size;
1754 struct replicas_delta_list *d;
1756 if (flags & BTREE_TRIGGER_OVERWRITE)
1758 sectors *= replicas;
1760 d = replicas_deltas_realloc(trans, 0);
1762 replicas = clamp_t(unsigned, replicas, 1,
1763 ARRAY_SIZE(d->persistent_reserved));
1765 d->persistent_reserved[replicas - 1] += sectors;
1769 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1770 struct bkey_s_c_reflink_p p,
1771 u64 *idx, unsigned flags)
1773 struct bch_fs *c = trans->c;
1774 struct btree_iter iter;
1778 int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1782 bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, POS(0, *idx),
1784 BTREE_ITER_WITH_UPDATES);
1785 k = bch2_btree_iter_peek_slot(&iter);
1790 n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1791 ret = PTR_ERR_OR_ZERO(n);
1795 bkey_reassemble(n, k);
1797 refcount = bkey_refcount(n);
1799 bch2_bkey_val_to_text(&PBUF(buf), c, p.s_c);
1800 bch2_fs_inconsistent(c,
1801 "nonexistent indirect extent at %llu while marking\n %s",
1807 if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1808 bch2_bkey_val_to_text(&PBUF(buf), c, p.s_c);
1809 bch2_fs_inconsistent(c,
1810 "indirect extent refcount underflow at %llu while marking\n %s",
1816 if (flags & BTREE_TRIGGER_INSERT) {
1817 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1820 pad = max_t(s64, le32_to_cpu(v->front_pad),
1821 le64_to_cpu(v->idx) - bkey_start_offset(k.k));
1822 BUG_ON(pad > U32_MAX);
1823 v->front_pad = cpu_to_le32(pad);
1825 pad = max_t(s64, le32_to_cpu(v->back_pad),
1826 k.k->p.offset - p.k->size - le64_to_cpu(v->idx));
1827 BUG_ON(pad > U32_MAX);
1828 v->back_pad = cpu_to_le32(pad);
1831 le64_add_cpu(refcount, add);
1834 n->k.type = KEY_TYPE_deleted;
1835 set_bkey_val_u64s(&n->k, 0);
1838 bch2_btree_iter_set_pos_to_extent_start(&iter);
1839 ret = bch2_trans_update(trans, &iter, n, 0);
1843 *idx = k.k->p.offset;
1845 bch2_trans_iter_exit(trans, &iter);
1849 static int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1850 struct bkey_s_c k, unsigned flags)
1852 struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1856 if (flags & BTREE_TRIGGER_INSERT) {
1857 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1859 v->front_pad = v->back_pad = 0;
1862 idx = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1863 end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1864 le32_to_cpu(p.v->back_pad);
1866 while (idx < end_idx && !ret)
1867 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1872 int bch2_trans_mark_key(struct btree_trans *trans, struct bkey_s_c old,
1873 struct bkey_s_c new, unsigned flags)
1875 struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1877 switch (k.k->type) {
1878 case KEY_TYPE_btree_ptr:
1879 case KEY_TYPE_btree_ptr_v2:
1880 case KEY_TYPE_extent:
1881 case KEY_TYPE_reflink_v:
1882 return bch2_trans_mark_extent(trans, k, flags);
1883 case KEY_TYPE_stripe:
1884 return bch2_trans_mark_stripe(trans, old, new, flags);
1885 case KEY_TYPE_inode:
1886 case KEY_TYPE_inode_v2:
1887 return bch2_trans_mark_inode(trans, old, new, flags);
1888 case KEY_TYPE_reservation:
1889 return bch2_trans_mark_reservation(trans, k, flags);
1890 case KEY_TYPE_reflink_p:
1891 return bch2_trans_mark_reflink_p(trans, k, flags);
1897 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1898 struct bch_dev *ca, size_t b,
1899 enum bch_data_type type,
1902 struct bch_fs *c = trans->c;
1903 struct btree_iter iter;
1904 struct bkey_alloc_unpacked u;
1905 struct bch_extent_ptr ptr = {
1907 .offset = bucket_to_sector(ca, b),
1912 * Backup superblock might be past the end of our normal usable space:
1914 if (b >= ca->mi.nbuckets)
1917 ret = bch2_trans_start_alloc_update(trans, &iter, &ptr, &u);
1921 if (u.data_type && u.data_type != type) {
1922 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1923 "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1925 iter.pos.inode, iter.pos.offset, u.gen,
1926 bch2_data_types[u.data_type],
1927 bch2_data_types[type],
1928 bch2_data_types[type]);
1934 u.dirty_sectors = sectors;
1936 ret = bch2_alloc_write(trans, &iter, &u, 0);
1940 bch2_trans_iter_exit(trans, &iter);
1944 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1945 struct bch_dev *ca, size_t b,
1946 enum bch_data_type type,
1949 return __bch2_trans_do(trans, NULL, NULL, 0,
1950 __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1953 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1956 enum bch_data_type type,
1957 u64 *bucket, unsigned *bucket_sectors)
1960 u64 b = sector_to_bucket(ca, start);
1962 min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1964 if (b != *bucket && *bucket_sectors) {
1965 int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1966 type, *bucket_sectors);
1970 *bucket_sectors = 0;
1974 *bucket_sectors += sectors;
1976 } while (start < end);
1981 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1984 struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1986 unsigned i, bucket_sectors = 0;
1989 for (i = 0; i < layout->nr_superblocks; i++) {
1990 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1992 if (offset == BCH_SB_SECTOR) {
1993 ret = bch2_trans_mark_metadata_sectors(trans, ca,
1995 BCH_DATA_sb, &bucket, &bucket_sectors);
2000 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
2001 offset + (1 << layout->sb_max_size_bits),
2002 BCH_DATA_sb, &bucket, &bucket_sectors);
2007 if (bucket_sectors) {
2008 ret = bch2_trans_mark_metadata_bucket(trans, ca,
2009 bucket, BCH_DATA_sb, bucket_sectors);
2014 for (i = 0; i < ca->journal.nr; i++) {
2015 ret = bch2_trans_mark_metadata_bucket(trans, ca,
2016 ca->journal.buckets[i],
2017 BCH_DATA_journal, ca->mi.bucket_size);
2025 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
2027 return bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
2028 __bch2_trans_mark_dev_sb(&trans, ca));
2031 /* Disk reservations: */
2033 #define SECTORS_CACHE 1024
2035 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
2036 u64 sectors, int flags)
2038 struct bch_fs_pcpu *pcpu;
2040 s64 sectors_available;
2043 percpu_down_read(&c->mark_lock);
2045 pcpu = this_cpu_ptr(c->pcpu);
2047 if (sectors <= pcpu->sectors_available)
2050 v = atomic64_read(&c->sectors_available);
2053 get = min((u64) sectors + SECTORS_CACHE, old);
2055 if (get < sectors) {
2059 } while ((v = atomic64_cmpxchg(&c->sectors_available,
2060 old, old - get)) != old);
2062 pcpu->sectors_available += get;
2065 pcpu->sectors_available -= sectors;
2066 this_cpu_add(*c->online_reserved, sectors);
2067 res->sectors += sectors;
2070 percpu_up_read(&c->mark_lock);
2074 mutex_lock(&c->sectors_available_lock);
2076 percpu_u64_set(&c->pcpu->sectors_available, 0);
2077 sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2079 if (sectors <= sectors_available ||
2080 (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2081 atomic64_set(&c->sectors_available,
2082 max_t(s64, 0, sectors_available - sectors));
2083 this_cpu_add(*c->online_reserved, sectors);
2084 res->sectors += sectors;
2087 atomic64_set(&c->sectors_available, sectors_available);
2091 mutex_unlock(&c->sectors_available_lock);
2092 percpu_up_read(&c->mark_lock);
2097 /* Startup/shutdown: */
2099 static void buckets_free_rcu(struct rcu_head *rcu)
2101 struct bucket_array *buckets =
2102 container_of(rcu, struct bucket_array, rcu);
2106 buckets->nbuckets * sizeof(struct bucket));
2109 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2111 struct bucket_gens *buckets =
2112 container_of(rcu, struct bucket_gens, rcu);
2114 kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2117 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2119 struct bucket_array *buckets = NULL, *old_buckets = NULL;
2120 struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2121 unsigned long *buckets_nouse = NULL;
2122 alloc_fifo free[RESERVE_NR];
2123 alloc_fifo free_inc;
2124 alloc_heap alloc_heap;
2126 size_t btree_reserve = DIV_ROUND_UP(BTREE_NODE_RESERVE,
2127 ca->mi.bucket_size / btree_sectors(c));
2128 /* XXX: these should be tunable */
2129 size_t reserve_none = max_t(size_t, 1, nbuckets >> 9);
2130 size_t copygc_reserve = max_t(size_t, 2, nbuckets >> 6);
2131 size_t free_inc_nr = max(max_t(size_t, 1, nbuckets >> 12),
2133 bool resize = ca->buckets[0] != NULL;
2137 memset(&free, 0, sizeof(free));
2138 memset(&free_inc, 0, sizeof(free_inc));
2139 memset(&alloc_heap, 0, sizeof(alloc_heap));
2141 if (!(buckets = kvpmalloc(sizeof(struct bucket_array) +
2142 nbuckets * sizeof(struct bucket),
2143 GFP_KERNEL|__GFP_ZERO)) ||
2144 !(bucket_gens = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2145 GFP_KERNEL|__GFP_ZERO)) ||
2146 (c->opts.buckets_nouse &&
2147 !(buckets_nouse = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2148 sizeof(unsigned long),
2149 GFP_KERNEL|__GFP_ZERO))) ||
2150 !init_fifo(&free[RESERVE_MOVINGGC],
2151 copygc_reserve, GFP_KERNEL) ||
2152 !init_fifo(&free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
2153 !init_fifo(&free_inc, free_inc_nr, GFP_KERNEL) ||
2154 !init_heap(&alloc_heap, ALLOC_SCAN_BATCH(ca) << 1, GFP_KERNEL))
2157 buckets->first_bucket = ca->mi.first_bucket;
2158 buckets->nbuckets = nbuckets;
2159 bucket_gens->first_bucket = ca->mi.first_bucket;
2160 bucket_gens->nbuckets = nbuckets;
2162 bch2_copygc_stop(c);
2165 down_write(&c->gc_lock);
2166 down_write(&ca->bucket_lock);
2167 percpu_down_write(&c->mark_lock);
2170 old_buckets = bucket_array(ca);
2171 old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2174 size_t n = min(buckets->nbuckets, old_buckets->nbuckets);
2178 n * sizeof(struct bucket));
2179 memcpy(bucket_gens->b,
2183 memcpy(buckets_nouse,
2185 BITS_TO_LONGS(n) * sizeof(unsigned long));
2188 rcu_assign_pointer(ca->buckets[0], buckets);
2189 rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2190 buckets = old_buckets;
2191 bucket_gens = old_bucket_gens;
2193 swap(ca->buckets_nouse, buckets_nouse);
2196 percpu_up_write(&c->mark_lock);
2197 up_write(&c->gc_lock);
2200 spin_lock(&c->freelist_lock);
2201 for (i = 0; i < RESERVE_NR; i++) {
2202 fifo_move(&free[i], &ca->free[i]);
2203 swap(ca->free[i], free[i]);
2205 fifo_move(&free_inc, &ca->free_inc);
2206 swap(ca->free_inc, free_inc);
2207 spin_unlock(&c->freelist_lock);
2209 /* with gc lock held, alloc_heap can't be in use: */
2210 swap(ca->alloc_heap, alloc_heap);
2212 nbuckets = ca->mi.nbuckets;
2215 up_write(&ca->bucket_lock);
2219 free_heap(&alloc_heap);
2220 free_fifo(&free_inc);
2221 for (i = 0; i < RESERVE_NR; i++)
2222 free_fifo(&free[i]);
2223 kvpfree(buckets_nouse,
2224 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2226 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2228 call_rcu(&buckets->rcu, buckets_free_rcu);
2233 void bch2_dev_buckets_free(struct bch_dev *ca)
2237 free_heap(&ca->alloc_heap);
2238 free_fifo(&ca->free_inc);
2239 for (i = 0; i < RESERVE_NR; i++)
2240 free_fifo(&ca->free[i]);
2241 kvpfree(ca->buckets_nouse,
2242 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2243 kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2244 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2245 kvpfree(rcu_dereference_protected(ca->buckets[0], 1),
2246 sizeof(struct bucket_array) +
2247 ca->mi.nbuckets * sizeof(struct bucket));
2249 for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2250 free_percpu(ca->usage[i]);
2251 kfree(ca->usage_base);
2254 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2258 ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2259 if (!ca->usage_base)
2262 for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2263 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2268 return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;