X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Falloc_background.c;h=ca1f45cc80b7f24de2be1cf86886e7cb796ecbfb;hb=79d39bd7cec19ea03356652d7add5e99b4da3a97;hp=744addb0019d3c12539cd173cfe960b455d07730;hpb=7e3538959966b80be3d4cfcb7076c8ab40a4be11;p=bcachefs-tools-debian diff --git a/libbcachefs/alloc_background.c b/libbcachefs/alloc_background.c index 744addb..ca1f45c 100644 --- a/libbcachefs/alloc_background.c +++ b/libbcachefs/alloc_background.c @@ -1,17 +1,23 @@ +// SPDX-License-Identifier: GPL-2.0 #include "bcachefs.h" #include "alloc_background.h" #include "alloc_foreground.h" +#include "backpointers.h" #include "btree_cache.h" #include "btree_io.h" +#include "btree_key_cache.h" #include "btree_update.h" #include "btree_update_interior.h" #include "btree_gc.h" #include "buckets.h" +#include "buckets_waiting_for_journal.h" #include "clock.h" #include "debug.h" #include "ec.h" #include "error.h" +#include "lru.h" #include "recovery.h" +#include "varint.h" #include #include @@ -22,56 +28,30 @@ #include #include -static const char * const bch2_alloc_field_names[] = { -#define x(name, bytes) #name, - BCH_ALLOC_FIELDS() +/* Persistent alloc info: */ + +static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = { +#define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8, + BCH_ALLOC_FIELDS_V1() #undef x - NULL }; -static void bch2_recalc_oldest_io(struct bch_fs *, struct bch_dev *, int); - -/* Ratelimiting/PD controllers */ - -static void pd_controllers_update(struct work_struct *work) -{ - struct bch_fs *c = container_of(to_delayed_work(work), - struct bch_fs, - pd_controllers_update); - struct bch_dev *ca; - unsigned i; - - for_each_member_device(ca, c, i) { - struct bch_dev_usage stats = bch2_dev_usage_read(c, ca); - - u64 free = bucket_to_sector(ca, - __dev_buckets_free(ca, stats)) << 9; - /* - * Bytes of internal fragmentation, which can be - * reclaimed by copy GC - */ - s64 fragmented = (bucket_to_sector(ca, - stats.buckets[BCH_DATA_USER] + - stats.buckets[BCH_DATA_CACHED]) - - (stats.sectors[BCH_DATA_USER] + - stats.sectors[BCH_DATA_CACHED])) << 9; - - fragmented = max(0LL, fragmented); - - bch2_pd_controller_update(&ca->copygc_pd, - free, fragmented, -1); - } - - schedule_delayed_work(&c->pd_controllers_update, - c->pd_controllers_update_seconds * HZ); -} - -/* Persistent alloc info: */ +struct bkey_alloc_unpacked { + u64 journal_seq; + u8 gen; + u8 oldest_gen; + u8 data_type; + bool need_discard:1; + bool need_inc_gen:1; +#define x(_name, _bits) u##_bits _name; + BCH_ALLOC_FIELDS_V2() +#undef x +}; -static inline u64 get_alloc_field(const struct bch_alloc *a, - const void **p, unsigned field) +static inline u64 alloc_field_v1_get(const struct bch_alloc *a, + const void **p, unsigned field) { - unsigned bytes = BCH_ALLOC_FIELD_BYTES[field]; + unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field]; u64 v; if (!(a->fields & (1 << field))) @@ -98,10 +78,10 @@ static inline u64 get_alloc_field(const struct bch_alloc *a, return v; } -static inline void put_alloc_field(struct bkey_i_alloc *a, void **p, - unsigned field, u64 v) +static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p, + unsigned field, u64 v) { - unsigned bytes = BCH_ALLOC_FIELD_BYTES[field]; + unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field]; if (!v) return; @@ -128,1105 +108,1314 @@ static inline void put_alloc_field(struct bkey_i_alloc *a, void **p, *p += bytes; } -struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k) +static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out, + struct bkey_s_c k) { - struct bkey_alloc_unpacked ret = { .gen = 0 }; + const struct bch_alloc *in = bkey_s_c_to_alloc(k).v; + const void *d = in->data; + unsigned idx = 0; - if (k.k->type == KEY_TYPE_alloc) { - const struct bch_alloc *a = bkey_s_c_to_alloc(k).v; - const void *d = a->data; - unsigned idx = 0; + out->gen = in->gen; - ret.gen = a->gen; +#define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++); + BCH_ALLOC_FIELDS_V1() +#undef x +} -#define x(_name, _bits) ret._name = get_alloc_field(a, &d, idx++); - BCH_ALLOC_FIELDS() +static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out, + struct bkey_s_c k) +{ + struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k); + const u8 *in = a.v->data; + const u8 *end = bkey_val_end(a); + unsigned fieldnr = 0; + int ret; + u64 v; + + out->gen = a.v->gen; + out->oldest_gen = a.v->oldest_gen; + out->data_type = a.v->data_type; + +#define x(_name, _bits) \ + if (fieldnr < a.v->nr_fields) { \ + ret = bch2_varint_decode_fast(in, end, &v); \ + if (ret < 0) \ + return ret; \ + in += ret; \ + } else { \ + v = 0; \ + } \ + out->_name = v; \ + if (v != out->_name) \ + return -1; \ + fieldnr++; + + BCH_ALLOC_FIELDS_V2() #undef x + return 0; +} + +static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out, + struct bkey_s_c k) +{ + struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k); + const u8 *in = a.v->data; + const u8 *end = bkey_val_end(a); + unsigned fieldnr = 0; + int ret; + u64 v; + + out->gen = a.v->gen; + out->oldest_gen = a.v->oldest_gen; + out->data_type = a.v->data_type; + out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v); + out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v); + out->journal_seq = le64_to_cpu(a.v->journal_seq); + +#define x(_name, _bits) \ + if (fieldnr < a.v->nr_fields) { \ + ret = bch2_varint_decode_fast(in, end, &v); \ + if (ret < 0) \ + return ret; \ + in += ret; \ + } else { \ + v = 0; \ + } \ + out->_name = v; \ + if (v != out->_name) \ + return -1; \ + fieldnr++; + + BCH_ALLOC_FIELDS_V2() +#undef x + return 0; +} + +static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k) +{ + struct bkey_alloc_unpacked ret = { .gen = 0 }; + + switch (k.k->type) { + case KEY_TYPE_alloc: + bch2_alloc_unpack_v1(&ret, k); + break; + case KEY_TYPE_alloc_v2: + bch2_alloc_unpack_v2(&ret, k); + break; + case KEY_TYPE_alloc_v3: + bch2_alloc_unpack_v3(&ret, k); + break; } + return ret; } -void bch2_alloc_pack(struct bkey_i_alloc *dst, - const struct bkey_alloc_unpacked src) +struct bkey_i_alloc_v4 * +bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter, + struct bpos pos) { - unsigned idx = 0; - void *d = dst->v.data; - - dst->v.fields = 0; - dst->v.gen = src.gen; + struct bkey_s_c k; + struct bkey_i_alloc_v4 *a; + int ret; -#define x(_name, _bits) put_alloc_field(dst, &d, idx++, src._name); - BCH_ALLOC_FIELDS() -#undef x + bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos, + BTREE_ITER_WITH_UPDATES| + BTREE_ITER_CACHED| + BTREE_ITER_INTENT); + k = bch2_btree_iter_peek_slot(iter); + ret = bkey_err(k); + if (ret) { + bch2_trans_iter_exit(trans, iter); + return ERR_PTR(ret); + } - set_bkey_val_bytes(&dst->k, (void *) d - (void *) &dst->v); + a = bch2_alloc_to_v4_mut(trans, k); + if (IS_ERR(a)) + bch2_trans_iter_exit(trans, iter); + return a; } -static unsigned bch_alloc_val_u64s(const struct bch_alloc *a) +static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a) { unsigned i, bytes = offsetof(struct bch_alloc, data); - for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_FIELD_BYTES); i++) + for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++) if (a->fields & (1 << i)) - bytes += BCH_ALLOC_FIELD_BYTES[i]; + bytes += BCH_ALLOC_V1_FIELD_BYTES[i]; return DIV_ROUND_UP(bytes, sizeof(u64)); } -const char *bch2_alloc_invalid(const struct bch_fs *c, struct bkey_s_c k) +int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k, + int rw, struct printbuf *err) { struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k); - if (k.k->p.inode >= c->sb.nr_devices || - !c->devs[k.k->p.inode]) - return "invalid device"; - /* allow for unknown fields */ - if (bkey_val_u64s(a.k) < bch_alloc_val_u64s(a.v)) - return "incorrect value size"; + if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) { + prt_printf(err, "incorrect value size (%zu < %u)", + bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v)); + return -EINVAL; + } - return NULL; + return 0; } -void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, - struct bkey_s_c k) +int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k, + int rw, struct printbuf *err) { - struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k); - const void *d = a.v->data; - unsigned i; + struct bkey_alloc_unpacked u; - pr_buf(out, "gen %u", a.v->gen); + if (bch2_alloc_unpack_v2(&u, k)) { + prt_printf(err, "unpack error"); + return -EINVAL; + } - for (i = 0; i < BCH_ALLOC_FIELD_NR; i++) - if (a.v->fields & (1 << i)) - pr_buf(out, " %s %llu", - bch2_alloc_field_names[i], - get_alloc_field(a.v, &d, i)); + return 0; } -static inline struct bkey_alloc_unpacked -alloc_mem_to_key(struct bucket *g, struct bucket_mark m) +int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k, + int rw, struct printbuf *err) { - return (struct bkey_alloc_unpacked) { - .gen = m.gen, - .oldest_gen = g->oldest_gen, - .data_type = m.data_type, - .dirty_sectors = m.dirty_sectors, - .cached_sectors = m.cached_sectors, - .read_time = g->io_time[READ], - .write_time = g->io_time[WRITE], - }; -} + struct bkey_alloc_unpacked u; -int bch2_alloc_read(struct bch_fs *c, struct journal_keys *journal_keys) -{ - struct btree_trans trans; - struct btree_iter *iter; - struct bkey_s_c k; - struct bch_dev *ca; - struct journal_key *j; - unsigned i; - int ret; + if (bch2_alloc_unpack_v3(&u, k)) { + prt_printf(err, "unpack error"); + return -EINVAL; + } - bch2_trans_init(&trans, c, 0, 0); + return 0; +} - for_each_btree_key(&trans, iter, BTREE_ID_ALLOC, POS_MIN, 0, k, ret) - bch2_mark_key(c, k, 0, NULL, 0, - BCH_BUCKET_MARK_ALLOC_READ| - BCH_BUCKET_MARK_NOATOMIC); +int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k, + int rw, struct printbuf *err) +{ + struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k); - ret = bch2_trans_exit(&trans) ?: ret; - if (ret) { - bch_err(c, "error reading alloc info: %i", ret); - return ret; + if (alloc_v4_u64s(a.v) != bkey_val_u64s(k.k)) { + prt_printf(err, "bad val size (%lu != %u)", + bkey_val_u64s(k.k), alloc_v4_u64s(a.v)); + return -EINVAL; } - for_each_journal_key(*journal_keys, j) - if (j->btree_id == BTREE_ID_ALLOC) - bch2_mark_key(c, bkey_i_to_s_c(j->k), 0, NULL, 0, - BCH_BUCKET_MARK_ALLOC_READ| - BCH_BUCKET_MARK_NOATOMIC); + if (!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) && + BCH_ALLOC_V4_NR_BACKPOINTERS(a.v)) { + prt_printf(err, "invalid backpointers_start"); + return -EINVAL; + } - percpu_down_write(&c->mark_lock); - bch2_dev_usage_from_buckets(c); - percpu_up_write(&c->mark_lock); + if (rw == WRITE) { + if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) { + prt_printf(err, "invalid data type (got %u should be %u)", + a.v->data_type, alloc_data_type(*a.v, a.v->data_type)); + return -EINVAL; + } - mutex_lock(&c->bucket_clock[READ].lock); - for_each_member_device(ca, c, i) { - down_read(&ca->bucket_lock); - bch2_recalc_oldest_io(c, ca, READ); - up_read(&ca->bucket_lock); - } - mutex_unlock(&c->bucket_clock[READ].lock); + switch (a.v->data_type) { + case BCH_DATA_free: + case BCH_DATA_need_gc_gens: + case BCH_DATA_need_discard: + if (a.v->dirty_sectors || + a.v->cached_sectors || + a.v->stripe) { + prt_printf(err, "empty data type free but have data"); + return -EINVAL; + } + break; + case BCH_DATA_sb: + case BCH_DATA_journal: + case BCH_DATA_btree: + case BCH_DATA_user: + case BCH_DATA_parity: + if (!a.v->dirty_sectors) { + prt_printf(err, "data_type %s but dirty_sectors==0", + bch2_data_types[a.v->data_type]); + return -EINVAL; + } + break; + case BCH_DATA_cached: + if (!a.v->cached_sectors || + a.v->dirty_sectors || + a.v->stripe) { + prt_printf(err, "data type inconsistency"); + return -EINVAL; + } - mutex_lock(&c->bucket_clock[WRITE].lock); - for_each_member_device(ca, c, i) { - down_read(&ca->bucket_lock); - bch2_recalc_oldest_io(c, ca, WRITE); - up_read(&ca->bucket_lock); + if (!a.v->io_time[READ] && + test_bit(BCH_FS_CHECK_ALLOC_TO_LRU_REFS_DONE, &c->flags)) { + prt_printf(err, "cached bucket with read_time == 0"); + return -EINVAL; + } + break; + case BCH_DATA_stripe: + if (!a.v->stripe) { + prt_printf(err, "data_type %s but stripe==0", + bch2_data_types[a.v->data_type]); + return -EINVAL; + } + break; + } } - mutex_unlock(&c->bucket_clock[WRITE].lock); return 0; } -int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k) +static inline u64 swab40(u64 x) { - struct btree_trans trans; - struct btree_iter *iter; - struct bch_dev *ca; - int ret; + return (((x & 0x00000000ffULL) << 32)| + ((x & 0x000000ff00ULL) << 16)| + ((x & 0x0000ff0000ULL) >> 0)| + ((x & 0x00ff000000ULL) >> 16)| + ((x & 0xff00000000ULL) >> 32)); +} - if (k->k.p.inode >= c->sb.nr_devices || - !c->devs[k->k.p.inode]) - return 0; +void bch2_alloc_v4_swab(struct bkey_s k) +{ + struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v; + struct bch_backpointer *bp, *bps; + + a->journal_seq = swab64(a->journal_seq); + a->flags = swab32(a->flags); + a->dirty_sectors = swab32(a->dirty_sectors); + a->cached_sectors = swab32(a->cached_sectors); + a->io_time[0] = swab64(a->io_time[0]); + a->io_time[1] = swab64(a->io_time[1]); + a->stripe = swab32(a->stripe); + a->nr_external_backpointers = swab32(a->nr_external_backpointers); + + bps = alloc_v4_backpointers(a); + for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) { + bp->bucket_offset = swab40(bp->bucket_offset); + bp->bucket_len = swab32(bp->bucket_len); + bch2_bpos_swab(&bp->pos); + } +} - ca = bch_dev_bkey_exists(c, k->k.p.inode); +void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k) +{ + struct bch_alloc_v4 _a; + const struct bch_alloc_v4 *a = &_a; + const struct bch_backpointer *bps; + unsigned i; - if (k->k.p.offset >= ca->mi.nbuckets) - return 0; + if (k.k->type == KEY_TYPE_alloc_v4) + a = bkey_s_c_to_alloc_v4(k).v; + else + bch2_alloc_to_v4(k, &_a); + + prt_newline(out); + printbuf_indent_add(out, 2); + + prt_printf(out, "gen %u oldest_gen %u data_type %s", + a->gen, a->oldest_gen, bch2_data_types[a->data_type]); + prt_newline(out); + prt_printf(out, "journal_seq %llu", a->journal_seq); + prt_newline(out); + prt_printf(out, "need_discard %llu", BCH_ALLOC_V4_NEED_DISCARD(a)); + prt_newline(out); + prt_printf(out, "need_inc_gen %llu", BCH_ALLOC_V4_NEED_INC_GEN(a)); + prt_newline(out); + prt_printf(out, "dirty_sectors %u", a->dirty_sectors); + prt_newline(out); + prt_printf(out, "cached_sectors %u", a->cached_sectors); + prt_newline(out); + prt_printf(out, "stripe %u", a->stripe); + prt_newline(out); + prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy); + prt_newline(out); + prt_printf(out, "io_time[READ] %llu", a->io_time[READ]); + prt_newline(out); + prt_printf(out, "io_time[WRITE] %llu", a->io_time[WRITE]); + prt_newline(out); + prt_printf(out, "backpointers: %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a)); + printbuf_indent_add(out, 2); + + bps = alloc_v4_backpointers_c(a); + for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a); i++) { + prt_newline(out); + bch2_backpointer_to_text(out, &bps[i]); + } - bch2_trans_init(&trans, c, 0, 0); + printbuf_indent_sub(out, 4); +} - iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, k->k.p, - BTREE_ITER_INTENT); +void bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out) +{ + if (k.k->type == KEY_TYPE_alloc_v4) { + int d; + + *out = *bkey_s_c_to_alloc_v4(k).v; + + d = (int) BCH_ALLOC_V4_U64s - + (int) (BCH_ALLOC_V4_BACKPOINTERS_START(out) ?: BCH_ALLOC_V4_U64s_V0); + if (unlikely(d > 0)) { + memset((u64 *) out + BCH_ALLOC_V4_BACKPOINTERS_START(out), + 0, + d * sizeof(u64)); + SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s); + } + } else { + struct bkey_alloc_unpacked u = bch2_alloc_unpack(k); + + *out = (struct bch_alloc_v4) { + .journal_seq = u.journal_seq, + .flags = u.need_discard, + .gen = u.gen, + .oldest_gen = u.oldest_gen, + .data_type = u.data_type, + .stripe_redundancy = u.stripe_redundancy, + .dirty_sectors = u.dirty_sectors, + .cached_sectors = u.cached_sectors, + .io_time[READ] = u.read_time, + .io_time[WRITE] = u.write_time, + .stripe = u.stripe, + }; + + SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s); + } +} - ret = bch2_btree_iter_traverse(iter); - if (ret) - goto err; +struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k) +{ + unsigned bytes = k.k->type == KEY_TYPE_alloc_v4 + ? bkey_bytes(k.k) + : sizeof(struct bkey_i_alloc_v4); + struct bkey_i_alloc_v4 *ret; - /* check buckets_written with btree node locked: */ - if (test_bit(k->k.p.offset, ca->buckets_written)) { - ret = 0; - goto err; - } + /* + * Reserve space for one more backpointer here: + * Not sketchy at doing it this way, nope... + */ + ret = bch2_trans_kmalloc(trans, bytes + sizeof(struct bch_backpointer)); + if (IS_ERR(ret)) + return ret; - bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, k)); + if (k.k->type == KEY_TYPE_alloc_v4) { + bkey_reassemble(&ret->k_i, k); - ret = bch2_trans_commit(&trans, NULL, NULL, - BTREE_INSERT_NOFAIL| - BTREE_INSERT_LAZY_RW| - BTREE_INSERT_JOURNAL_REPLAY| - BTREE_INSERT_NOMARK); -err: - bch2_trans_exit(&trans); + if (BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v) < BCH_ALLOC_V4_U64s) { + struct bch_backpointer *src, *dst; + + src = alloc_v4_backpointers(&ret->v); + SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s); + dst = alloc_v4_backpointers(&ret->v); + + memmove(dst, src, BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v) * + sizeof(struct bch_backpointer)); + memset(src, 0, dst - src); + set_alloc_v4_u64s(ret); + } + } else { + bkey_alloc_v4_init(&ret->k_i); + ret->k.p = k.k->p; + bch2_alloc_to_v4(k, &ret->v); + } return ret; } -int bch2_alloc_write(struct bch_fs *c, unsigned flags, bool *wrote) +int bch2_alloc_read(struct bch_fs *c) { struct btree_trans trans; - struct btree_iter *iter; - struct bucket_array *buckets; - struct bch_dev *ca; - struct bucket *g; - struct bucket_mark m, new; - struct bkey_alloc_unpacked old_u, new_u; - __BKEY_PADDED(k, 8) alloc_key; /* hack: */ - struct bkey_i_alloc *a; + struct btree_iter iter; struct bkey_s_c k; - unsigned i; - size_t b; - int ret = 0; - - BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8); + struct bch_alloc_v4 a; + struct bch_dev *ca; + int ret; bch2_trans_init(&trans, c, 0, 0); - iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, POS_MIN, - BTREE_ITER_SLOTS|BTREE_ITER_INTENT); - - for_each_rw_member(ca, c, i) { - down_read(&ca->bucket_lock); -restart: - buckets = bucket_array(ca); - - for (b = buckets->first_bucket; - b < buckets->nbuckets; - b++) { - if (!buckets->b[b].mark.dirty) - continue; + for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN, + BTREE_ITER_PREFETCH, k, ret) { + /* + * Not a fsck error because this is checked/repaired by + * bch2_check_alloc_key() which runs later: + */ + if (!bch2_dev_bucket_exists(c, k.k->p)) + continue; - bch2_btree_iter_set_pos(iter, POS(i, b)); - k = bch2_btree_iter_peek_slot(iter); - ret = bkey_err(k); - if (ret) - goto err; + ca = bch_dev_bkey_exists(c, k.k->p.inode); + bch2_alloc_to_v4(k, &a); - old_u = bch2_alloc_unpack(k); + *bucket_gen(ca, k.k->p.offset) = a.gen; + } + bch2_trans_iter_exit(&trans, &iter); - percpu_down_read_preempt_disable(&c->mark_lock); - g = bucket(ca, b); - m = READ_ONCE(g->mark); - new_u = alloc_mem_to_key(g, m); - percpu_up_read_preempt_enable(&c->mark_lock); + bch2_trans_exit(&trans); - if (!m.dirty) - continue; + if (ret) + bch_err(c, "error reading alloc info: %i", ret); - if ((flags & BTREE_INSERT_LAZY_RW) && - percpu_ref_is_zero(&c->writes)) { - up_read(&ca->bucket_lock); - bch2_trans_unlock(&trans); + return ret; +} - ret = bch2_fs_read_write_early(c); - down_read(&ca->bucket_lock); +/* Free space/discard btree: */ - if (ret) - goto err; - goto restart; - } +static int bch2_bucket_do_index(struct btree_trans *trans, + struct bkey_s_c alloc_k, + const struct bch_alloc_v4 *a, + bool set) +{ + struct bch_fs *c = trans->c; + struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode); + struct btree_iter iter; + struct bkey_s_c old; + struct bkey_i *k; + enum btree_id btree; + enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted; + enum bch_bkey_type new_type = set ? KEY_TYPE_set : KEY_TYPE_deleted; + struct printbuf buf = PRINTBUF; + int ret; - a = bkey_alloc_init(&alloc_key.k); - a->k.p = iter->pos; - bch2_alloc_pack(a, new_u); + if (a->data_type != BCH_DATA_free && + a->data_type != BCH_DATA_need_discard) + return 0; - bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &a->k_i)); - ret = bch2_trans_commit(&trans, NULL, NULL, - BTREE_INSERT_NOFAIL| - BTREE_INSERT_NOMARK| - flags); -err: - if (ret && !test_bit(BCH_FS_EMERGENCY_RO, &c->flags)) { - bch_err(c, "error %i writing alloc info", ret); - printk(KERN_CONT "dev %llu bucket %llu\n", - iter->pos.inode, iter->pos.offset); - printk(KERN_CONT "gen %u -> %u\n", old_u.gen, new_u.gen); -#define x(_name, _bits) printk(KERN_CONT #_name " %u -> %u\n", old_u._name, new_u._name); - BCH_ALLOC_FIELDS() -#undef x - } - if (ret) - break; + k = bch2_trans_kmalloc(trans, sizeof(*k)); + if (IS_ERR(k)) + return PTR_ERR(k); - new = m; - new.dirty = false; - atomic64_cmpxchg(&g->_mark.v, m.v.counter, new.v.counter); + bkey_init(&k->k); + k->k.type = new_type; - if (ca->buckets_written) - set_bit(b, ca->buckets_written); + switch (a->data_type) { + case BCH_DATA_free: + btree = BTREE_ID_freespace; + k->k.p = alloc_freespace_pos(alloc_k.k->p, *a); + bch2_key_resize(&k->k, 1); + break; + case BCH_DATA_need_discard: + btree = BTREE_ID_need_discard; + k->k.p = alloc_k.k->p; + break; + default: + return 0; + } - bch2_trans_cond_resched(&trans); - *wrote = true; - } - up_read(&ca->bucket_lock); + bch2_trans_iter_init(trans, &iter, btree, + bkey_start_pos(&k->k), + BTREE_ITER_INTENT); + old = bch2_btree_iter_peek_slot(&iter); + ret = bkey_err(old); + if (ret) + goto err; - if (ret) { - percpu_ref_put(&ca->io_ref); - break; - } + if (ca->mi.freespace_initialized && + bch2_trans_inconsistent_on(old.k->type != old_type, trans, + "incorrect key when %s %s btree (got %s should be %s)\n" + " for %s", + set ? "setting" : "clearing", + bch2_btree_ids[btree], + bch2_bkey_types[old.k->type], + bch2_bkey_types[old_type], + (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) { + ret = -EIO; + goto err; } - bch2_trans_exit(&trans); - + ret = bch2_trans_update(trans, &iter, k, 0); +err: + bch2_trans_iter_exit(trans, &iter); + printbuf_exit(&buf); return ret; } -/* Bucket IO clocks: */ - -static void bch2_recalc_oldest_io(struct bch_fs *c, struct bch_dev *ca, int rw) +int bch2_trans_mark_alloc(struct btree_trans *trans, + enum btree_id btree_id, unsigned level, + struct bkey_s_c old, struct bkey_i *new, + unsigned flags) { - struct bucket_clock *clock = &c->bucket_clock[rw]; - struct bucket_array *buckets = bucket_array(ca); - struct bucket *g; - u16 max_last_io = 0; - unsigned i; - - lockdep_assert_held(&c->bucket_clock[rw].lock); - - /* Recalculate max_last_io for this device: */ - for_each_bucket(g, buckets) - max_last_io = max(max_last_io, bucket_last_io(c, g, rw)); + struct bch_fs *c = trans->c; + struct bch_alloc_v4 old_a, *new_a; + u64 old_lru, new_lru; + int ret = 0; - ca->max_last_bucket_io[rw] = max_last_io; + /* + * Deletion only happens in the device removal path, with + * BTREE_TRIGGER_NORUN: + */ + BUG_ON(new->k.type != KEY_TYPE_alloc_v4); - /* Recalculate global max_last_io: */ - max_last_io = 0; + bch2_alloc_to_v4(old, &old_a); + new_a = &bkey_i_to_alloc_v4(new)->v; - for_each_member_device(ca, c, i) - max_last_io = max(max_last_io, ca->max_last_bucket_io[rw]); + new_a->data_type = alloc_data_type(*new_a, new_a->data_type); - clock->max_last_io = max_last_io; -} + if (new_a->dirty_sectors > old_a.dirty_sectors || + new_a->cached_sectors > old_a.cached_sectors) { + new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now)); + new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now)); + SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true); + SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true); + } -static void bch2_rescale_bucket_io_times(struct bch_fs *c, int rw) -{ - struct bucket_clock *clock = &c->bucket_clock[rw]; - struct bucket_array *buckets; - struct bch_dev *ca; - struct bucket *g; - unsigned i; + if (data_type_is_empty(new_a->data_type) && + BCH_ALLOC_V4_NEED_INC_GEN(new_a) && + !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) { + new_a->gen++; + SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false); + } - trace_rescale_prios(c); + if (old_a.data_type != new_a->data_type || + (new_a->data_type == BCH_DATA_free && + alloc_freespace_genbits(old_a) != alloc_freespace_genbits(*new_a))) { + ret = bch2_bucket_do_index(trans, old, &old_a, false) ?: + bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true); + if (ret) + return ret; + } - for_each_member_device(ca, c, i) { - down_read(&ca->bucket_lock); - buckets = bucket_array(ca); + if (new_a->data_type == BCH_DATA_cached && + !new_a->io_time[READ]) + new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now)); - for_each_bucket(g, buckets) - g->io_time[rw] = clock->hand - - bucket_last_io(c, g, rw) / 2; + old_lru = alloc_lru_idx(old_a); + new_lru = alloc_lru_idx(*new_a); - bch2_recalc_oldest_io(c, ca, rw); + if (old_lru != new_lru) { + ret = bch2_lru_change(trans, new->k.p.inode, new->k.p.offset, + old_lru, &new_lru, old); + if (ret) + return ret; - up_read(&ca->bucket_lock); + if (new_a->data_type == BCH_DATA_cached) + new_a->io_time[READ] = new_lru; } -} -static inline u64 bucket_clock_freq(u64 capacity) -{ - return max(capacity >> 10, 2028ULL); + return 0; } -static void bch2_inc_clock_hand(struct io_timer *timer) +static int bch2_check_alloc_key(struct btree_trans *trans, + struct btree_iter *alloc_iter, + struct btree_iter *discard_iter, + struct btree_iter *freespace_iter) { - struct bucket_clock *clock = container_of(timer, - struct bucket_clock, rescale); - struct bch_fs *c = container_of(clock, - struct bch_fs, bucket_clock[clock->rw]); + struct bch_fs *c = trans->c; struct bch_dev *ca; - u64 capacity; - unsigned i; + struct bch_alloc_v4 a; + unsigned discard_key_type, freespace_key_type; + struct bkey_s_c alloc_k, k; + struct printbuf buf = PRINTBUF; + int ret; - mutex_lock(&clock->lock); + alloc_k = bch2_dev_bucket_exists(c, alloc_iter->pos) + ? bch2_btree_iter_peek_slot(alloc_iter) + : bch2_btree_iter_peek(alloc_iter); + if (!alloc_k.k) + return 1; - /* if clock cannot be advanced more, rescale prio */ - if (clock->max_last_io >= U16_MAX - 2) - bch2_rescale_bucket_io_times(c, clock->rw); + ret = bkey_err(alloc_k); + if (ret) + return ret; - BUG_ON(clock->max_last_io >= U16_MAX - 2); + if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c, + "alloc key for invalid device:bucket %llu:%llu", + alloc_k.k->p.inode, alloc_k.k->p.offset)) + return bch2_btree_delete_at(trans, alloc_iter, 0); - for_each_member_device(ca, c, i) - ca->max_last_bucket_io[clock->rw]++; - clock->max_last_io++; - clock->hand++; + ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode); + if (!ca->mi.freespace_initialized) + return 0; - mutex_unlock(&clock->lock); + bch2_alloc_to_v4(alloc_k, &a); - capacity = READ_ONCE(c->capacity); + discard_key_type = a.data_type == BCH_DATA_need_discard + ? KEY_TYPE_set : 0; + freespace_key_type = a.data_type == BCH_DATA_free + ? KEY_TYPE_set : 0; - if (!capacity) - return; + bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p); + bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, a)); - /* - * we only increment when 0.1% of the filesystem capacity has been read - * or written too, this determines if it's time - * - * XXX: we shouldn't really be going off of the capacity of devices in - * RW mode (that will be 0 when we're RO, yet we can still service - * reads) - */ - timer->expire += bucket_clock_freq(capacity); + k = bch2_btree_iter_peek_slot(discard_iter); + ret = bkey_err(k); + if (ret) + goto err; - bch2_io_timer_add(&c->io_clock[clock->rw], timer); -} + if (k.k->type != discard_key_type && + (c->opts.reconstruct_alloc || + fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n" + " %s", + bch2_bkey_types[k.k->type], + bch2_bkey_types[discard_key_type], + (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) { + struct bkey_i *update = + bch2_trans_kmalloc(trans, sizeof(*update)); + + ret = PTR_ERR_OR_ZERO(update); + if (ret) + goto err; -static void bch2_bucket_clock_init(struct bch_fs *c, int rw) -{ - struct bucket_clock *clock = &c->bucket_clock[rw]; + bkey_init(&update->k); + update->k.type = discard_key_type; + update->k.p = discard_iter->pos; - clock->hand = 1; - clock->rw = rw; - clock->rescale.fn = bch2_inc_clock_hand; - clock->rescale.expire = bucket_clock_freq(c->capacity); - mutex_init(&clock->lock); -} + ret = bch2_trans_update(trans, discard_iter, update, 0); + if (ret) + goto err; + } -/* Background allocator thread: */ + k = bch2_btree_iter_peek_slot(freespace_iter); + ret = bkey_err(k); + if (ret) + goto err; -/* - * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens - * (marking them as invalidated on disk), then optionally issues discard - * commands to the newly free buckets, then puts them on the various freelists. - */ + if (k.k->type != freespace_key_type && + (c->opts.reconstruct_alloc || + fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n" + " %s", + bch2_bkey_types[k.k->type], + bch2_bkey_types[freespace_key_type], + (printbuf_reset(&buf), + bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) { + struct bkey_i *update = + bch2_trans_kmalloc(trans, sizeof(*update)); + + ret = PTR_ERR_OR_ZERO(update); + if (ret) + goto err; -#define BUCKET_GC_GEN_MAX 96U + bkey_init(&update->k); + update->k.type = freespace_key_type; + update->k.p = freespace_iter->pos; + bch2_key_resize(&update->k, 1); -/** - * wait_buckets_available - wait on reclaimable buckets - * - * If there aren't enough available buckets to fill up free_inc, wait until - * there are. - */ -static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca) + ret = bch2_trans_update(trans, freespace_iter, update, 0); + if (ret) + goto err; + } +err: +fsck_err: + printbuf_exit(&buf); + return ret; +} + +static int bch2_check_discard_freespace_key(struct btree_trans *trans, + struct btree_iter *iter) { - unsigned long gc_count = c->gc_count; - int ret = 0; + struct bch_fs *c = trans->c; + struct btree_iter alloc_iter; + struct bkey_s_c k, freespace_k; + struct bch_alloc_v4 a; + u64 genbits; + struct bpos pos; + enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard + ? BCH_DATA_need_discard + : BCH_DATA_free; + struct printbuf buf = PRINTBUF; + int ret; - ca->allocator_state = ALLOCATOR_BLOCKED; - closure_wake_up(&c->freelist_wait); + freespace_k = bch2_btree_iter_peek(iter); + if (!freespace_k.k) + return 1; - while (1) { - set_current_state(TASK_INTERRUPTIBLE); - if (kthread_should_stop()) { - ret = 1; - break; - } + ret = bkey_err(freespace_k); + if (ret) + return ret; - if (gc_count != c->gc_count) - ca->inc_gen_really_needs_gc = 0; + pos = iter->pos; + pos.offset &= ~(~0ULL << 56); + genbits = iter->pos.offset & (~0ULL << 56); - if ((ssize_t) (dev_buckets_available(c, ca) - - ca->inc_gen_really_needs_gc) >= - (ssize_t) fifo_free(&ca->free_inc)) - break; + bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0); - up_read(&c->gc_lock); - schedule(); - try_to_freeze(); - down_read(&c->gc_lock); - } + if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c, + "entry in %s btree for nonexistant dev:bucket %llu:%llu", + bch2_btree_ids[iter->btree_id], pos.inode, pos.offset)) + goto delete; - __set_current_state(TASK_RUNNING); - ca->allocator_state = ALLOCATOR_RUNNING; - closure_wake_up(&c->freelist_wait); + k = bch2_btree_iter_peek_slot(&alloc_iter); + ret = bkey_err(k); + if (ret) + goto err; + bch2_alloc_to_v4(k, &a); + + if (fsck_err_on(a.data_type != state || + (state == BCH_DATA_free && + genbits != alloc_freespace_genbits(a)), c, + "%s\n incorrectly set in %s index (free %u, genbits %llu should be %llu)", + (bch2_bkey_val_to_text(&buf, c, k), buf.buf), + bch2_btree_ids[iter->btree_id], + a.data_type == state, + genbits >> 56, alloc_freespace_genbits(a) >> 56)) + goto delete; +out: +err: +fsck_err: + bch2_trans_iter_exit(trans, &alloc_iter); + printbuf_exit(&buf); return ret; +delete: + ret = bch2_btree_delete_extent_at(trans, iter, + iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0); + goto out; } -static bool bch2_can_invalidate_bucket(struct bch_dev *ca, - size_t bucket, - struct bucket_mark mark) +int bch2_check_alloc_info(struct bch_fs *c) { - u8 gc_gen; - - if (!is_available_bucket(mark)) - return false; - - if (ca->buckets_nouse && - test_bit(bucket, ca->buckets_nouse)) - return false; - - gc_gen = bucket_gc_gen(ca, bucket); + struct btree_trans trans; + struct btree_iter iter, discard_iter, freespace_iter; + int ret = 0; - if (gc_gen >= BUCKET_GC_GEN_MAX / 2) - ca->inc_gen_needs_gc++; + bch2_trans_init(&trans, c, 0, 0); - if (gc_gen >= BUCKET_GC_GEN_MAX) - ca->inc_gen_really_needs_gc++; + bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN, + BTREE_ITER_PREFETCH); + bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN, + BTREE_ITER_PREFETCH); + bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN, + BTREE_ITER_PREFETCH); + while (1) { + ret = commit_do(&trans, NULL, NULL, + BTREE_INSERT_NOFAIL| + BTREE_INSERT_LAZY_RW, + bch2_check_alloc_key(&trans, &iter, + &discard_iter, + &freespace_iter)); + if (ret) + break; - return gc_gen < BUCKET_GC_GEN_MAX; -} + bch2_btree_iter_advance(&iter); + } + bch2_trans_iter_exit(&trans, &freespace_iter); + bch2_trans_iter_exit(&trans, &discard_iter); + bch2_trans_iter_exit(&trans, &iter); -/* - * Determines what order we're going to reuse buckets, smallest bucket_key() - * first. - * - * - * - We take into account the read prio of the bucket, which gives us an - * indication of how hot the data is -- we scale the prio so that the prio - * farthest from the clock is worth 1/8th of the closest. - * - * - The number of sectors of cached data in the bucket, which gives us an - * indication of the cost in cache misses this eviction will cause. - * - * - If hotness * sectors used compares equal, we pick the bucket with the - * smallest bucket_gc_gen() - since incrementing the same bucket's generation - * number repeatedly forces us to run mark and sweep gc to avoid generation - * number wraparound. - */ - -static unsigned long bucket_sort_key(struct bch_fs *c, struct bch_dev *ca, - size_t b, struct bucket_mark m) -{ - unsigned last_io = bucket_last_io(c, bucket(ca, b), READ); - unsigned max_last_io = ca->max_last_bucket_io[READ]; + if (ret < 0) + goto err; - /* - * Time since last read, scaled to [0, 8) where larger value indicates - * more recently read data: - */ - unsigned long hotness = (max_last_io - last_io) * 7 / max_last_io; + bch2_trans_iter_init(&trans, &iter, BTREE_ID_need_discard, POS_MIN, + BTREE_ITER_PREFETCH); + while (1) { + ret = commit_do(&trans, NULL, NULL, + BTREE_INSERT_NOFAIL| + BTREE_INSERT_LAZY_RW, + bch2_check_discard_freespace_key(&trans, &iter)); + if (ret) + break; - /* How much we want to keep the data in this bucket: */ - unsigned long data_wantness = - (hotness + 1) * bucket_sectors_used(m); + bch2_btree_iter_advance(&iter); + } + bch2_trans_iter_exit(&trans, &iter); - unsigned long needs_journal_commit = - bucket_needs_journal_commit(m, c->journal.last_seq_ondisk); + if (ret < 0) + goto err; - return (data_wantness << 9) | - (needs_journal_commit << 8) | - (bucket_gc_gen(ca, b) / 16); -} + bch2_trans_iter_init(&trans, &iter, BTREE_ID_freespace, POS_MIN, + BTREE_ITER_PREFETCH); + while (1) { + ret = commit_do(&trans, NULL, NULL, + BTREE_INSERT_NOFAIL| + BTREE_INSERT_LAZY_RW, + bch2_check_discard_freespace_key(&trans, &iter)); + if (ret) + break; -static inline int bucket_alloc_cmp(alloc_heap *h, - struct alloc_heap_entry l, - struct alloc_heap_entry r) -{ - return cmp_int(l.key, r.key) ?: - cmp_int(r.nr, l.nr) ?: - cmp_int(l.bucket, r.bucket); + bch2_btree_iter_advance(&iter); + } + bch2_trans_iter_exit(&trans, &iter); +err: + bch2_trans_exit(&trans); + return ret < 0 ? ret : 0; } -static inline int bucket_idx_cmp(const void *_l, const void *_r) +static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans, + struct btree_iter *alloc_iter) { - const struct alloc_heap_entry *l = _l, *r = _r; + struct bch_fs *c = trans->c; + struct btree_iter lru_iter; + struct bch_alloc_v4 a; + struct bkey_s_c alloc_k, k; + struct printbuf buf = PRINTBUF; + struct printbuf buf2 = PRINTBUF; + int ret; - return cmp_int(l->bucket, r->bucket); -} + alloc_k = bch2_btree_iter_peek(alloc_iter); + if (!alloc_k.k) + return 0; -static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca) -{ - struct bucket_array *buckets; - struct alloc_heap_entry e = { 0 }; - size_t b, i, nr = 0; + ret = bkey_err(alloc_k); + if (ret) + return ret; - ca->alloc_heap.used = 0; + bch2_alloc_to_v4(alloc_k, &a); - mutex_lock(&c->bucket_clock[READ].lock); - down_read(&ca->bucket_lock); + if (a.data_type != BCH_DATA_cached) + return 0; - buckets = bucket_array(ca); + bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru, + POS(alloc_k.k->p.inode, a.io_time[READ]), 0); - bch2_recalc_oldest_io(c, ca, READ); + k = bch2_btree_iter_peek_slot(&lru_iter); + ret = bkey_err(k); + if (ret) + goto err; - /* - * Find buckets with lowest read priority, by building a maxheap sorted - * by read priority and repeatedly replacing the maximum element until - * all buckets have been visited. - */ - for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) { - struct bucket_mark m = READ_ONCE(buckets->b[b].mark); - unsigned long key = bucket_sort_key(c, ca, b, m); + if (fsck_err_on(!a.io_time[READ], c, + "cached bucket with read_time 0\n" + " %s", + (printbuf_reset(&buf), + bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) || + fsck_err_on(k.k->type != KEY_TYPE_lru || + le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != alloc_k.k->p.offset, c, + "incorrect/missing lru entry\n" + " %s\n" + " %s", + (printbuf_reset(&buf), + bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf), + (bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) { + u64 read_time = a.io_time[READ]; + + if (!a.io_time[READ]) + a.io_time[READ] = atomic64_read(&c->io_clock[READ].now); + + ret = bch2_lru_set(trans, + alloc_k.k->p.inode, + alloc_k.k->p.offset, + &a.io_time[READ]); + if (ret) + goto err; - if (!bch2_can_invalidate_bucket(ca, b, m)) - continue; + if (a.io_time[READ] != read_time) { + struct bkey_i_alloc_v4 *a_mut = + bch2_alloc_to_v4_mut(trans, alloc_k); + ret = PTR_ERR_OR_ZERO(a_mut); + if (ret) + goto err; - if (e.nr && e.bucket + e.nr == b && e.key == key) { - e.nr++; - } else { - if (e.nr) - heap_add_or_replace(&ca->alloc_heap, e, - -bucket_alloc_cmp, NULL); - - e = (struct alloc_heap_entry) { - .bucket = b, - .nr = 1, - .key = key, - }; + a_mut->v.io_time[READ] = a.io_time[READ]; + ret = bch2_trans_update(trans, alloc_iter, + &a_mut->k_i, BTREE_TRIGGER_NORUN); + if (ret) + goto err; } - - cond_resched(); } - - if (e.nr) - heap_add_or_replace(&ca->alloc_heap, e, - -bucket_alloc_cmp, NULL); - - for (i = 0; i < ca->alloc_heap.used; i++) - nr += ca->alloc_heap.data[i].nr; - - while (nr - ca->alloc_heap.data[0].nr >= ALLOC_SCAN_BATCH(ca)) { - nr -= ca->alloc_heap.data[0].nr; - heap_pop(&ca->alloc_heap, e, -bucket_alloc_cmp, NULL); - } - - up_read(&ca->bucket_lock); - mutex_unlock(&c->bucket_clock[READ].lock); +err: +fsck_err: + bch2_trans_iter_exit(trans, &lru_iter); + printbuf_exit(&buf2); + printbuf_exit(&buf); + return ret; } -static void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca) +int bch2_check_alloc_to_lru_refs(struct bch_fs *c) { - struct bucket_array *buckets = bucket_array(ca); - struct bucket_mark m; - size_t b, start; - - if (ca->fifo_last_bucket < ca->mi.first_bucket || - ca->fifo_last_bucket >= ca->mi.nbuckets) - ca->fifo_last_bucket = ca->mi.first_bucket; - - start = ca->fifo_last_bucket; - - do { - ca->fifo_last_bucket++; - if (ca->fifo_last_bucket == ca->mi.nbuckets) - ca->fifo_last_bucket = ca->mi.first_bucket; - - b = ca->fifo_last_bucket; - m = READ_ONCE(buckets->b[b].mark); + struct btree_trans trans; + struct btree_iter iter; + struct bkey_s_c k; + int ret = 0; - if (bch2_can_invalidate_bucket(ca, b, m)) { - struct alloc_heap_entry e = { .bucket = b, .nr = 1, }; + bch2_trans_init(&trans, c, 0, 0); - heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); - if (heap_full(&ca->alloc_heap)) - break; - } + for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc, + POS_MIN, BTREE_ITER_PREFETCH, k, + NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW, + bch2_check_alloc_to_lru_ref(&trans, &iter)); - cond_resched(); - } while (ca->fifo_last_bucket != start); + bch2_trans_exit(&trans); + return ret < 0 ? ret : 0; } -static void find_reclaimable_buckets_random(struct bch_fs *c, struct bch_dev *ca) +static int bch2_clear_need_discard(struct btree_trans *trans, struct bpos pos, + struct bch_dev *ca, bool *discard_done) { - struct bucket_array *buckets = bucket_array(ca); - struct bucket_mark m; - size_t checked, i; + struct bch_fs *c = trans->c; + struct btree_iter iter; + struct bkey_s_c k; + struct bkey_i_alloc_v4 *a; + struct printbuf buf = PRINTBUF; + int ret; - for (checked = 0; - checked < ca->mi.nbuckets / 2; - checked++) { - size_t b = bch2_rand_range(ca->mi.nbuckets - - ca->mi.first_bucket) + - ca->mi.first_bucket; + bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, pos, + BTREE_ITER_CACHED); + k = bch2_btree_iter_peek_slot(&iter); + ret = bkey_err(k); + if (ret) + goto out; - m = READ_ONCE(buckets->b[b].mark); + a = bch2_alloc_to_v4_mut(trans, k); + ret = PTR_ERR_OR_ZERO(a); + if (ret) + goto out; - if (bch2_can_invalidate_bucket(ca, b, m)) { - struct alloc_heap_entry e = { .bucket = b, .nr = 1, }; + if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) { + a->v.gen++; + SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false); + goto write; + } - heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); - if (heap_full(&ca->alloc_heap)) - break; - } + if (bch2_trans_inconsistent_on(a->v.journal_seq > c->journal.flushed_seq_ondisk, trans, + "clearing need_discard but journal_seq %llu > flushed_seq %llu\n" + "%s", + a->v.journal_seq, + c->journal.flushed_seq_ondisk, + (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) { + ret = -EIO; + goto out; + } - cond_resched(); + if (bch2_trans_inconsistent_on(a->v.data_type != BCH_DATA_need_discard, trans, + "bucket incorrectly set in need_discard btree\n" + "%s", + (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) { + ret = -EIO; + goto out; } - sort(ca->alloc_heap.data, - ca->alloc_heap.used, - sizeof(ca->alloc_heap.data[0]), - bucket_idx_cmp, NULL); + if (!*discard_done && ca->mi.discard && !c->opts.nochanges) { + /* + * This works without any other locks because this is the only + * thread that removes items from the need_discard tree + */ + bch2_trans_unlock(trans); + blkdev_issue_discard(ca->disk_sb.bdev, + k.k->p.offset * ca->mi.bucket_size, + ca->mi.bucket_size, + GFP_KERNEL, 0); + *discard_done = true; + + ret = bch2_trans_relock(trans) ? 0 : -EINTR; + if (ret) + goto out; + } - /* remove duplicates: */ - for (i = 0; i + 1 < ca->alloc_heap.used; i++) - if (ca->alloc_heap.data[i].bucket == - ca->alloc_heap.data[i + 1].bucket) - ca->alloc_heap.data[i].nr = 0; + SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false); + a->v.data_type = alloc_data_type(a->v, a->v.data_type); +write: + ret = bch2_trans_update(trans, &iter, &a->k_i, 0); +out: + bch2_trans_iter_exit(trans, &iter); + printbuf_exit(&buf); + return ret; } -static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca) +static void bch2_do_discards_work(struct work_struct *work) { - size_t i, nr = 0; - - ca->inc_gen_needs_gc = 0; - - switch (ca->mi.replacement) { - case CACHE_REPLACEMENT_LRU: - find_reclaimable_buckets_lru(c, ca); - break; - case CACHE_REPLACEMENT_FIFO: - find_reclaimable_buckets_fifo(c, ca); - break; - case CACHE_REPLACEMENT_RANDOM: - find_reclaimable_buckets_random(c, ca); - break; - } + struct bch_fs *c = container_of(work, struct bch_fs, discard_work); + struct bch_dev *ca = NULL; + struct btree_trans trans; + struct btree_iter iter; + struct bkey_s_c k; + u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0; + int ret; - heap_resort(&ca->alloc_heap, bucket_alloc_cmp, NULL); + bch2_trans_init(&trans, c, 0, 0); - for (i = 0; i < ca->alloc_heap.used; i++) - nr += ca->alloc_heap.data[i].nr; + for_each_btree_key(&trans, iter, BTREE_ID_need_discard, + POS_MIN, 0, k, ret) { + bool discard_done = false; - return nr; -} + if (ca && k.k->p.inode != ca->dev_idx) { + percpu_ref_put(&ca->io_ref); + ca = NULL; + } -static inline long next_alloc_bucket(struct bch_dev *ca) -{ - struct alloc_heap_entry e, *top = ca->alloc_heap.data; + if (!ca) { + ca = bch_dev_bkey_exists(c, k.k->p.inode); + if (!percpu_ref_tryget(&ca->io_ref)) { + ca = NULL; + bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0)); + continue; + } + } + + seen++; - while (ca->alloc_heap.used) { - if (top->nr) { - size_t b = top->bucket; + if (bch2_bucket_is_open_safe(c, k.k->p.inode, k.k->p.offset)) { + open++; + continue; + } - top->bucket++; - top->nr--; - return b; + if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal, + c->journal.flushed_seq_ondisk, + k.k->p.inode, k.k->p.offset)) { + need_journal_commit++; + continue; } - heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); + ret = commit_do(&trans, NULL, NULL, + BTREE_INSERT_USE_RESERVE| + BTREE_INSERT_NOFAIL, + bch2_clear_need_discard(&trans, k.k->p, ca, &discard_done)); + if (ret) + break; + + this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]); + discarded++; } + bch2_trans_iter_exit(&trans, &iter); - return -1; -} + if (ca) + percpu_ref_put(&ca->io_ref); -/* - * returns sequence number of most recent journal entry that updated this - * bucket: - */ -static u64 bucket_journal_seq(struct bch_fs *c, struct bucket_mark m) -{ - if (m.journal_seq_valid) { - u64 journal_seq = atomic64_read(&c->journal.seq); - u64 bucket_seq = journal_seq; + bch2_trans_exit(&trans); - bucket_seq &= ~((u64) U16_MAX); - bucket_seq |= m.journal_seq; + if (need_journal_commit * 2 > seen) + bch2_journal_flush_async(&c->journal, NULL); - if (bucket_seq > journal_seq) - bucket_seq -= 1 << 16; + percpu_ref_put(&c->writes); - return bucket_seq; - } else { - return 0; - } + trace_discard_buckets(c, seen, open, need_journal_commit, discarded, ret); +} + +void bch2_do_discards(struct bch_fs *c) +{ + if (percpu_ref_tryget_live(&c->writes) && + !queue_work(system_long_wq, &c->discard_work)) + percpu_ref_put(&c->writes); } -static int bch2_invalidate_one_bucket2(struct btree_trans *trans, - struct bch_dev *ca, - struct btree_iter *iter, - u64 *journal_seq, unsigned flags) +static int invalidate_one_bucket(struct btree_trans *trans, struct bch_dev *ca, + struct bpos *bucket_pos, unsigned *cached_sectors) { -#if 0 - __BKEY_PADDED(k, BKEY_ALLOC_VAL_U64s_MAX) alloc_key; -#else - /* hack: */ - __BKEY_PADDED(k, 8) alloc_key; -#endif struct bch_fs *c = trans->c; - struct bkey_i_alloc *a; - struct bkey_alloc_unpacked u; - struct bucket *g; - struct bucket_mark m; + struct btree_iter lru_iter, alloc_iter = { NULL }; struct bkey_s_c k; - bool invalidating_cached_data; - size_t b; + struct bkey_i_alloc_v4 *a; + u64 bucket, idx; + struct printbuf buf = PRINTBUF; int ret; - BUG_ON(!ca->alloc_heap.used || - !ca->alloc_heap.data[0].nr); - b = ca->alloc_heap.data[0].bucket; - - /* first, put on free_inc and mark as owned by allocator: */ - percpu_down_read_preempt_disable(&c->mark_lock); - spin_lock(&c->freelist_lock); - - verify_not_on_freelist(c, ca, b); - - BUG_ON(!fifo_push(&ca->free_inc, b)); - - bch2_mark_alloc_bucket(c, ca, b, true, gc_pos_alloc(c, NULL), 0); - - spin_unlock(&c->freelist_lock); - percpu_up_read_preempt_enable(&c->mark_lock); - - BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8); - - bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b)); -retry: - k = bch2_btree_iter_peek_slot(iter); + bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru, + POS(ca->dev_idx, 0), 0); +next_lru: + k = bch2_btree_iter_peek(&lru_iter); ret = bkey_err(k); if (ret) - return ret; - - /* - * The allocator has to start before journal replay is finished - thus, - * we have to trust the in memory bucket @m, not the version in the - * btree: - */ - percpu_down_read_preempt_disable(&c->mark_lock); - g = bucket(ca, b); - m = READ_ONCE(g->mark); - u = alloc_mem_to_key(g, m); - percpu_up_read_preempt_enable(&c->mark_lock); + goto out; - invalidating_cached_data = m.cached_sectors != 0; - - u.gen++; - u.data_type = 0; - u.dirty_sectors = 0; - u.cached_sectors = 0; - u.read_time = c->bucket_clock[READ].hand; - u.write_time = c->bucket_clock[WRITE].hand; + if (!k.k || k.k->p.inode != ca->dev_idx) { + ret = 1; + goto out; + } - a = bkey_alloc_init(&alloc_key.k); - a->k.p = iter->pos; - bch2_alloc_pack(a, u); + if (k.k->type != KEY_TYPE_lru) { + prt_printf(&buf, "non lru key in lru btree:\n "); + bch2_bkey_val_to_text(&buf, c, k); - bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &a->k_i)); + if (!test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) { + bch_err(c, "%s", buf.buf); + bch2_btree_iter_advance(&lru_iter); + goto next_lru; + } else { + bch2_trans_inconsistent(trans, "%s", buf.buf); + ret = -EINVAL; + goto out; + } + } - /* - * XXX: - * when using deferred btree updates, we have journal reclaim doing - * btree updates and thus requiring the allocator to make forward - * progress, and here the allocator is requiring space in the journal - - * so we need a journal pre-reservation: - */ - ret = bch2_trans_commit(trans, NULL, - invalidating_cached_data ? journal_seq : NULL, - BTREE_INSERT_ATOMIC| - BTREE_INSERT_NOUNLOCK| - BTREE_INSERT_NOCHECK_RW| - BTREE_INSERT_NOFAIL| - BTREE_INSERT_USE_RESERVE| - BTREE_INSERT_USE_ALLOC_RESERVE| - BTREE_INSERT_BUCKET_INVALIDATE| - flags); - if (ret == -EINTR) - goto retry; - - if (!ret) { - /* remove from alloc_heap: */ - struct alloc_heap_entry e, *top = ca->alloc_heap.data; - - top->bucket++; - top->nr--; - - if (!top->nr) - heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL); - - /* with btree still locked: */ - if (ca->buckets_written) - set_bit(b, ca->buckets_written); + idx = k.k->p.offset; + bucket = le64_to_cpu(bkey_s_c_to_lru(k).v->idx); - /* - * Make sure we flush the last journal entry that updated this - * bucket (i.e. deleting the last reference) before writing to - * this bucket again: - */ - *journal_seq = max(*journal_seq, bucket_journal_seq(c, m)); - } else { - size_t b2; + *bucket_pos = POS(ca->dev_idx, bucket); - /* remove from free_inc: */ - percpu_down_read_preempt_disable(&c->mark_lock); - spin_lock(&c->freelist_lock); + a = bch2_trans_start_alloc_update(trans, &alloc_iter, *bucket_pos); + ret = PTR_ERR_OR_ZERO(a); + if (ret) + goto out; + + if (idx != alloc_lru_idx(a->v)) { + prt_printf(&buf, "alloc key does not point back to lru entry when invalidating bucket:\n "); + bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i)); + prt_printf(&buf, "\n "); + bch2_bkey_val_to_text(&buf, c, k); + + if (!test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) { + bch_err(c, "%s", buf.buf); + bch2_btree_iter_advance(&lru_iter); + goto next_lru; + } else { + bch2_trans_inconsistent(trans, "%s", buf.buf); + ret = -EINVAL; + goto out; + } + } - bch2_mark_alloc_bucket(c, ca, b, false, - gc_pos_alloc(c, NULL), 0); + if (!a->v.cached_sectors) + bch_err(c, "invalidating empty bucket, confused"); - BUG_ON(!fifo_pop_back(&ca->free_inc, b2)); - BUG_ON(b != b2); + *cached_sectors = a->v.cached_sectors; - spin_unlock(&c->freelist_lock); - percpu_up_read_preempt_enable(&c->mark_lock); - } + SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false); + a->v.gen++; + a->v.data_type = 0; + a->v.dirty_sectors = 0; + a->v.cached_sectors = 0; + a->v.io_time[READ] = atomic64_read(&c->io_clock[READ].now); + a->v.io_time[WRITE] = atomic64_read(&c->io_clock[WRITE].now); + ret = bch2_trans_update(trans, &alloc_iter, &a->k_i, + BTREE_TRIGGER_BUCKET_INVALIDATE); + if (ret) + goto out; +out: + bch2_trans_iter_exit(trans, &alloc_iter); + bch2_trans_iter_exit(trans, &lru_iter); + printbuf_exit(&buf); return ret; } -static bool bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca, - size_t bucket, u64 *flush_seq) +static void bch2_do_invalidates_work(struct work_struct *work) { - struct bucket_mark m; + struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work); + struct bch_dev *ca; + struct btree_trans trans; + struct bpos bucket; + unsigned i, sectors; + int ret = 0; - percpu_down_read_preempt_disable(&c->mark_lock); - spin_lock(&c->freelist_lock); + bch2_trans_init(&trans, c, 0, 0); - bch2_invalidate_bucket(c, ca, bucket, &m); + for_each_member_device(ca, c, i) { + s64 nr_to_invalidate = + should_invalidate_buckets(ca, bch2_dev_usage_read(ca)); + + while (nr_to_invalidate-- >= 0) { + ret = commit_do(&trans, NULL, NULL, + BTREE_INSERT_USE_RESERVE| + BTREE_INSERT_NOFAIL, + invalidate_one_bucket(&trans, ca, &bucket, + §ors)); + if (ret) + break; - verify_not_on_freelist(c, ca, bucket); - BUG_ON(!fifo_push(&ca->free_inc, bucket)); + trace_invalidate_bucket(c, bucket.inode, bucket.offset, sectors); + this_cpu_inc(c->counters[BCH_COUNTER_bucket_invalidate]); + } + } - spin_unlock(&c->freelist_lock); + bch2_trans_exit(&trans); + percpu_ref_put(&c->writes); +} - bucket_io_clock_reset(c, ca, bucket, READ); - bucket_io_clock_reset(c, ca, bucket, WRITE); +void bch2_do_invalidates(struct bch_fs *c) +{ + if (percpu_ref_tryget_live(&c->writes) && + !queue_work(system_long_wq, &c->invalidate_work)) + percpu_ref_put(&c->writes); +} - percpu_up_read_preempt_enable(&c->mark_lock); +static int bucket_freespace_init(struct btree_trans *trans, struct btree_iter *iter) +{ + struct bch_alloc_v4 a; + struct bkey_s_c k; + int ret; - *flush_seq = max(*flush_seq, bucket_journal_seq(c, m)); + k = bch2_btree_iter_peek_slot(iter); + ret = bkey_err(k); + if (ret) + return ret; - return m.cached_sectors != 0; + bch2_alloc_to_v4(k, &a); + return bch2_bucket_do_index(trans, k, &a, true); } -/* - * Pull buckets off ca->alloc_heap, invalidate them, move them to ca->free_inc: - */ -static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca) +static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca) { struct btree_trans trans; - struct btree_iter *iter; - u64 journal_seq = 0; - int ret = 0; + struct btree_iter iter; + struct bkey_s_c k; + struct bch_member *m; + int ret; bch2_trans_init(&trans, c, 0, 0); - iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, - POS(ca->dev_idx, 0), - BTREE_ITER_SLOTS|BTREE_ITER_INTENT); + for_each_btree_key(&trans, iter, BTREE_ID_alloc, + POS(ca->dev_idx, ca->mi.first_bucket), + BTREE_ITER_SLOTS| + BTREE_ITER_PREFETCH, k, ret) { + if (iter.pos.offset >= ca->mi.nbuckets) + break; - /* Only use nowait if we've already invalidated at least one bucket: */ - while (!ret && - !fifo_full(&ca->free_inc) && - ca->alloc_heap.used) - ret = bch2_invalidate_one_bucket2(&trans, ca, iter, &journal_seq, - BTREE_INSERT_GC_LOCK_HELD| - (!fifo_empty(&ca->free_inc) - ? BTREE_INSERT_NOWAIT : 0)); + ret = commit_do(&trans, NULL, NULL, + BTREE_INSERT_LAZY_RW, + bucket_freespace_init(&trans, &iter)); + if (ret) + break; + } + bch2_trans_iter_exit(&trans, &iter); bch2_trans_exit(&trans); - /* If we used NOWAIT, don't return the error: */ - if (!fifo_empty(&ca->free_inc)) - ret = 0; if (ret) { - bch_err(ca, "error invalidating buckets: %i", ret); + bch_err(ca, "error initializing free space: %i", ret); return ret; } - if (journal_seq) - ret = bch2_journal_flush_seq(&c->journal, journal_seq); - if (ret) { - bch_err(ca, "journal error: %i", ret); - return ret; - } + mutex_lock(&c->sb_lock); + m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx; + SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true); + mutex_unlock(&c->sb_lock); - return 0; + return ret; } -static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, size_t bucket) +int bch2_fs_freespace_init(struct bch_fs *c) { + struct bch_dev *ca; unsigned i; int ret = 0; + bool doing_init = false; - while (1) { - set_current_state(TASK_INTERRUPTIBLE); - - spin_lock(&c->freelist_lock); - for (i = 0; i < RESERVE_NR; i++) - if (fifo_push(&ca->free[i], bucket)) { - fifo_pop(&ca->free_inc, bucket); - - closure_wake_up(&c->freelist_wait); - ca->allocator_state = ALLOCATOR_RUNNING; + /* + * We can crash during the device add path, so we need to check this on + * every mount: + */ - spin_unlock(&c->freelist_lock); - goto out; - } + for_each_member_device(ca, c, i) { + if (ca->mi.freespace_initialized) + continue; - if (ca->allocator_state != ALLOCATOR_BLOCKED_FULL) { - ca->allocator_state = ALLOCATOR_BLOCKED_FULL; - closure_wake_up(&c->freelist_wait); + if (!doing_init) { + bch_info(c, "initializing freespace"); + doing_init = true; } - spin_unlock(&c->freelist_lock); - - if ((current->flags & PF_KTHREAD) && - kthread_should_stop()) { - ret = 1; - break; + ret = bch2_dev_freespace_init(c, ca); + if (ret) { + percpu_ref_put(&ca->ref); + return ret; } - - schedule(); - try_to_freeze(); } -out: - __set_current_state(TASK_RUNNING); - return ret; -} - -/* - * Pulls buckets off free_inc, discards them (if enabled), then adds them to - * freelists, waiting until there's room if necessary: - */ -static int discard_invalidated_buckets(struct bch_fs *c, struct bch_dev *ca) -{ - while (!fifo_empty(&ca->free_inc)) { - size_t bucket = fifo_peek(&ca->free_inc); - if (ca->mi.discard && - blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev))) - blkdev_issue_discard(ca->disk_sb.bdev, - bucket_to_sector(ca, bucket), - ca->mi.bucket_size, GFP_NOIO, 0); + if (doing_init) { + mutex_lock(&c->sb_lock); + bch2_write_super(c); + mutex_unlock(&c->sb_lock); - if (push_invalidated_bucket(c, ca, bucket)) - return 1; + bch_verbose(c, "done initializing freespace"); } - return 0; + return ret; } -/** - * bch_allocator_thread - move buckets from free_inc to reserves - * - * The free_inc FIFO is populated by find_reclaimable_buckets(), and - * the reserves are depleted by bucket allocation. When we run out - * of free_inc, try to invalidate some buckets and write out - * prios and gens. - */ -static int bch2_allocator_thread(void *arg) -{ - struct bch_dev *ca = arg; - struct bch_fs *c = ca->fs; - size_t nr; - int ret; - - set_freezable(); - ca->allocator_state = ALLOCATOR_RUNNING; - - while (1) { - cond_resched(); - - pr_debug("discarding %zu invalidated buckets", - fifo_used(&ca->free_inc)); - - ret = discard_invalidated_buckets(c, ca); - if (ret) - goto stop; - - down_read(&c->gc_lock); - - ret = bch2_invalidate_buckets(c, ca); - if (ret) { - up_read(&c->gc_lock); - goto stop; - } - - if (!fifo_empty(&ca->free_inc)) { - up_read(&c->gc_lock); - continue; - } - - pr_debug("free_inc now empty"); - - do { - /* - * Find some buckets that we can invalidate, either - * they're completely unused, or only contain clean data - * that's been written back to the backing device or - * another cache tier - */ - - pr_debug("scanning for reclaimable buckets"); - - nr = find_reclaimable_buckets(c, ca); - - pr_debug("found %zu buckets", nr); - - trace_alloc_batch(ca, nr, ca->alloc_heap.size); - - if ((ca->inc_gen_needs_gc >= ALLOC_SCAN_BATCH(ca) || - ca->inc_gen_really_needs_gc) && - c->gc_thread) { - atomic_inc(&c->kick_gc); - wake_up_process(c->gc_thread); - } +/* Bucket IO clocks: */ - /* - * If we found any buckets, we have to invalidate them - * before we scan for more - but if we didn't find very - * many we may want to wait on more buckets being - * available so we don't spin: - */ - if (!nr || - (nr < ALLOC_SCAN_BATCH(ca) && - !fifo_full(&ca->free[RESERVE_MOVINGGC]))) { - ret = wait_buckets_available(c, ca); - if (ret) { - up_read(&c->gc_lock); - goto stop; - } - } - } while (!nr); +int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev, + size_t bucket_nr, int rw) +{ + struct bch_fs *c = trans->c; + struct btree_iter iter; + struct bkey_i_alloc_v4 *a; + u64 now; + int ret = 0; - up_read(&c->gc_lock); + a = bch2_trans_start_alloc_update(trans, &iter, POS(dev, bucket_nr)); + ret = PTR_ERR_OR_ZERO(a); + if (ret) + return ret; - pr_debug("%zu buckets to invalidate", nr); + now = atomic64_read(&c->io_clock[rw].now); + if (a->v.io_time[rw] == now) + goto out; - /* - * alloc_heap is now full of newly-invalidated buckets: next, - * write out the new bucket gens: - */ - } + a->v.io_time[rw] = now; -stop: - pr_debug("alloc thread stopping (ret %i)", ret); - ca->allocator_state = ALLOCATOR_STOPPED; - closure_wake_up(&c->freelist_wait); - return 0; + ret = bch2_trans_update(trans, &iter, &a->k_i, 0) ?: + bch2_trans_commit(trans, NULL, NULL, 0); +out: + bch2_trans_iter_exit(trans, &iter); + return ret; } /* Startup/shutdown (ro/rw): */ @@ -1237,12 +1426,12 @@ void bch2_recalc_capacity(struct bch_fs *c) u64 capacity = 0, reserved_sectors = 0, gc_reserve; unsigned bucket_size_max = 0; unsigned long ra_pages = 0; - unsigned i, j; + unsigned i; lockdep_assert_held(&c->state_lock); for_each_online_member(ca, c, i) { - struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_bdi; + struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi; ra_pages += bdi->ra_pages; } @@ -1268,8 +1457,9 @@ void bch2_recalc_capacity(struct bch_fs *c) * allocations for foreground writes must wait - * not -ENOSPC calculations. */ - for (j = 0; j < RESERVE_NONE; j++) - dev_reserve += ca->free[j].size; + + dev_reserve += ca->nr_btree_reserve * 2; + dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */ dev_reserve += 1; /* btree write point */ dev_reserve += 1; /* copygc write point */ @@ -1277,8 +1467,6 @@ void bch2_recalc_capacity(struct bch_fs *c) dev_reserve *= ca->mi.bucket_size; - ca->copygc_threshold = dev_reserve; - capacity += bucket_to_sector(ca, ca->mi.nbuckets - ca->mi.first_bucket); @@ -1300,18 +1488,6 @@ void bch2_recalc_capacity(struct bch_fs *c) c->bucket_size_max = bucket_size_max; - if (c->capacity) { - bch2_io_timer_add(&c->io_clock[READ], - &c->bucket_clock[READ].rescale); - bch2_io_timer_add(&c->io_clock[WRITE], - &c->bucket_clock[WRITE].rescale); - } else { - bch2_io_timer_del(&c->io_clock[READ], - &c->bucket_clock[READ].rescale); - bch2_io_timer_del(&c->io_clock[WRITE], - &c->bucket_clock[WRITE].rescale); - } - /* Wake up case someone was waiting for buckets */ closure_wake_up(&c->freelist_wait); } @@ -1326,7 +1502,7 @@ static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca) ob++) { spin_lock(&ob->lock); if (ob->valid && !ob->on_partial_list && - ob->ptr.dev == ca->dev_idx) + ob->dev == ca->dev_idx) ret = true; spin_unlock(&ob->lock); } @@ -1339,8 +1515,6 @@ void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca) { unsigned i; - BUG_ON(ca->alloc_thread); - /* First, remove device from allocation groups: */ for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++) @@ -1355,7 +1529,7 @@ void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca) for (i = 0; i < ARRAY_SIZE(c->write_points); i++) bch2_writepoint_stop(c, ca, &c->write_points[i]); - bch2_writepoint_stop(c, ca, &ca->copygc_write_point); + bch2_writepoint_stop(c, ca, &c->copygc_write_point); bch2_writepoint_stop(c, ca, &c->rebalance_write_point); bch2_writepoint_stop(c, ca, &c->btree_write_point); @@ -1414,276 +1588,9 @@ void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca) set_bit(ca->dev_idx, c->rw_devs[i].d); } -void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca) -{ - if (ca->alloc_thread) - closure_wait_event(&c->freelist_wait, - ca->allocator_state != ALLOCATOR_RUNNING); -} - -/* stop allocator thread: */ -void bch2_dev_allocator_stop(struct bch_dev *ca) -{ - struct task_struct *p; - - p = rcu_dereference_protected(ca->alloc_thread, 1); - ca->alloc_thread = NULL; - - /* - * We need an rcu barrier between setting ca->alloc_thread = NULL and - * the thread shutting down to avoid bch2_wake_allocator() racing: - * - * XXX: it would be better to have the rcu barrier be asynchronous - * instead of blocking us here - */ - synchronize_rcu(); - - if (p) { - kthread_stop(p); - put_task_struct(p); - } -} - -/* start allocator thread: */ -int bch2_dev_allocator_start(struct bch_dev *ca) -{ - struct task_struct *p; - - /* - * allocator thread already started? - */ - if (ca->alloc_thread) - return 0; - - p = kthread_create(bch2_allocator_thread, ca, - "bch_alloc[%s]", ca->name); - if (IS_ERR(p)) - return PTR_ERR(p); - - get_task_struct(p); - rcu_assign_pointer(ca->alloc_thread, p); - wake_up_process(p); - return 0; -} - -static bool flush_held_btree_writes(struct bch_fs *c) -{ - struct bucket_table *tbl; - struct rhash_head *pos; - struct btree *b; - bool nodes_unwritten; - size_t i; -again: - cond_resched(); - nodes_unwritten = false; - - rcu_read_lock(); - for_each_cached_btree(b, c, tbl, i, pos) - if (btree_node_need_write(b)) { - if (btree_node_may_write(b)) { - rcu_read_unlock(); - btree_node_lock_type(c, b, SIX_LOCK_read); - bch2_btree_node_write(c, b, SIX_LOCK_read); - six_unlock_read(&b->lock); - goto again; - } else { - nodes_unwritten = true; - } - } - rcu_read_unlock(); - - if (c->btree_roots_dirty) { - bch2_journal_meta(&c->journal); - goto again; - } - - return !nodes_unwritten && - !bch2_btree_interior_updates_nr_pending(c); -} - -static void allocator_start_issue_discards(struct bch_fs *c) -{ - struct bch_dev *ca; - unsigned dev_iter; - size_t bu; - - for_each_rw_member(ca, c, dev_iter) - while (fifo_pop(&ca->free_inc, bu)) - blkdev_issue_discard(ca->disk_sb.bdev, - bucket_to_sector(ca, bu), - ca->mi.bucket_size, GFP_NOIO, 0); -} - -static int resize_free_inc(struct bch_dev *ca) -{ - alloc_fifo free_inc; - - if (!fifo_full(&ca->free_inc)) - return 0; - - if (!init_fifo(&free_inc, - ca->free_inc.size * 2, - GFP_KERNEL)) - return -ENOMEM; - - fifo_move(&free_inc, &ca->free_inc); - swap(free_inc, ca->free_inc); - free_fifo(&free_inc); - return 0; -} - -static bool bch2_fs_allocator_start_fast(struct bch_fs *c) -{ - struct bch_dev *ca; - unsigned dev_iter; - bool ret = true; - - if (test_alloc_startup(c)) - return false; - - down_read(&c->gc_lock); - - /* Scan for buckets that are already invalidated: */ - for_each_rw_member(ca, c, dev_iter) { - struct bucket_array *buckets; - struct bucket_mark m; - long bu; - - down_read(&ca->bucket_lock); - buckets = bucket_array(ca); - - for (bu = buckets->first_bucket; - bu < buckets->nbuckets; bu++) { - m = READ_ONCE(buckets->b[bu].mark); - - if (!buckets->b[bu].gen_valid || - !is_available_bucket(m) || - m.cached_sectors || - (ca->buckets_nouse && - test_bit(bu, ca->buckets_nouse))) - continue; - - percpu_down_read_preempt_disable(&c->mark_lock); - bch2_mark_alloc_bucket(c, ca, bu, true, - gc_pos_alloc(c, NULL), 0); - percpu_up_read_preempt_enable(&c->mark_lock); - - fifo_push(&ca->free_inc, bu); - - discard_invalidated_buckets(c, ca); - - if (fifo_full(&ca->free[RESERVE_BTREE])) - break; - } - up_read(&ca->bucket_lock); - } - - up_read(&c->gc_lock); - - /* did we find enough buckets? */ - for_each_rw_member(ca, c, dev_iter) - if (!fifo_full(&ca->free[RESERVE_BTREE])) - ret = false; - - return ret; -} - -int bch2_fs_allocator_start(struct bch_fs *c) -{ - struct bch_dev *ca; - unsigned dev_iter; - u64 journal_seq = 0; - bool wrote; - long bu; - int ret = 0; - - if (!test_alloc_startup(c) && - bch2_fs_allocator_start_fast(c)) - return 0; - - pr_debug("not enough empty buckets; scanning for reclaimable buckets"); - - /* - * We're moving buckets to freelists _before_ they've been marked as - * invalidated on disk - we have to so that we can allocate new btree - * nodes to mark them as invalidated on disk. - * - * However, we can't _write_ to any of these buckets yet - they might - * have cached data in them, which is live until they're marked as - * invalidated on disk: - */ - set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags); - - down_read(&c->gc_lock); - do { - wrote = false; - - for_each_rw_member(ca, c, dev_iter) { - find_reclaimable_buckets(c, ca); - - while (!fifo_full(&ca->free[RESERVE_BTREE]) && - (bu = next_alloc_bucket(ca)) >= 0) { - ret = resize_free_inc(ca); - if (ret) { - percpu_ref_put(&ca->io_ref); - up_read(&c->gc_lock); - goto err; - } - - bch2_invalidate_one_bucket(c, ca, bu, - &journal_seq); - - fifo_push(&ca->free[RESERVE_BTREE], bu); - } - } - - pr_debug("done scanning for reclaimable buckets"); - - /* - * XXX: it's possible for this to deadlock waiting on journal reclaim, - * since we're holding btree writes. What then? - */ - ret = bch2_alloc_write(c, - BTREE_INSERT_NOCHECK_RW| - BTREE_INSERT_USE_ALLOC_RESERVE| - BTREE_INSERT_NOWAIT, &wrote); - - /* - * If bch2_alloc_write() did anything, it may have used some - * buckets, and we need the RESERVE_BTREE freelist full - so we - * need to loop and scan again. - * And if it errored, it may have been because there weren't - * enough buckets, so just scan and loop again as long as it - * made some progress: - */ - } while (wrote); - up_read(&c->gc_lock); - - if (ret) - goto err; - - pr_debug("flushing journal"); - - ret = bch2_journal_flush(&c->journal); - if (ret) - goto err; - - pr_debug("issuing discards"); - allocator_start_issue_discards(c); -err: - clear_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags); - closure_wait_event(&c->btree_interior_update_wait, - flush_held_btree_writes(c)); - - return ret; -} - void bch2_fs_allocator_background_init(struct bch_fs *c) { spin_lock_init(&c->freelist_lock); - bch2_bucket_clock_init(c, READ); - bch2_bucket_clock_init(c, WRITE); - - c->pd_controllers_update_seconds = 5; - INIT_DELAYED_WORK(&c->pd_controllers_update, pd_controllers_update); + INIT_WORK(&c->discard_work, bch2_do_discards_work); + INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work); }