#include <linux/sched/task.h>
#include <linux/sort.h>
+static void bch2_discard_one_bucket_fast(struct bch_fs *c, struct bpos bucket);
+
/* Persistent alloc info: */
static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
bkey_fsck_err_on(!bch2_bucket_sectors_dirty(*a.v),
c, err, alloc_key_dirty_sectors_0,
"data_type %s but dirty_sectors==0",
- bch2_data_types[a.v->data_type]);
+ bch2_data_type_str(a.v->data_type));
break;
case BCH_DATA_cached:
bkey_fsck_err_on(!a.v->cached_sectors ||
{
struct bch_alloc_v4 _a;
const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
- unsigned i;
prt_newline(out);
printbuf_indent_add(out, 2);
- prt_printf(out, "gen %u oldest_gen %u data_type %s",
- a->gen, a->oldest_gen,
- a->data_type < BCH_DATA_NR
- ? bch2_data_types[a->data_type]
- : "(invalid data type)");
+ prt_printf(out, "gen %u oldest_gen %u data_type ", a->gen, a->oldest_gen);
+ bch2_prt_data_type(out, a->data_type);
prt_newline(out);
prt_printf(out, "journal_seq %llu", a->journal_seq);
prt_newline(out);
prt_printf(out, "fragmentation %llu", a->fragmentation_lru);
prt_newline(out);
prt_printf(out, "bp_start %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
- prt_newline(out);
-
- if (BCH_ALLOC_V4_NR_BACKPOINTERS(a)) {
- struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
- const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
-
- prt_printf(out, "backpointers: %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
- printbuf_indent_add(out, 2);
-
- for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) {
- prt_newline(out);
- bch2_backpointer_to_text(out, &bps[i]);
- }
-
- printbuf_indent_sub(out, 2);
- }
-
printbuf_indent_sub(out, 2);
}
return ret;
}
-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)
+int bch2_trigger_alloc(struct btree_trans *trans,
+ enum btree_id btree, unsigned level,
+ struct bkey_s_c old, struct bkey_s new,
+ unsigned flags)
{
struct bch_fs *c = trans->c;
- struct bch_alloc_v4 old_a_convert, *new_a;
- const struct bch_alloc_v4 *old_a;
- u64 old_lru, new_lru;
int ret = 0;
- /*
- * Deletion only happens in the device removal path, with
- * BTREE_TRIGGER_NORUN:
- */
- BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
+ if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
+ "alloc key for invalid device or bucket"))
+ return -EIO;
- old_a = bch2_alloc_to_v4(old, &old_a_convert);
- new_a = &bkey_i_to_alloc_v4(new)->v;
+ struct bch_dev *ca = bch_dev_bkey_exists(c, new.k->p.inode);
- new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
+ struct bch_alloc_v4 old_a_convert;
+ const struct bch_alloc_v4 *old_a = bch2_alloc_to_v4(old, &old_a_convert);
- if (bch2_bucket_sectors(*new_a) > bch2_bucket_sectors(*old_a)) {
- 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);
- }
+ if (flags & BTREE_TRIGGER_TRANSACTIONAL) {
+ struct bch_alloc_v4 *new_a = bkey_s_to_alloc_v4(new).v;
- 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);
- }
+ new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
- 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;
- }
+ if (bch2_bucket_sectors(*new_a) > bch2_bucket_sectors(*old_a)) {
+ 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);
+ }
+
+ 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);
+ }
- 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));
+ 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, new.s_c, new_a, true);
+ if (ret)
+ return ret;
+ }
- old_lru = alloc_lru_idx_read(*old_a);
- new_lru = alloc_lru_idx_read(*new_a);
+ 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));
- if (old_lru != new_lru) {
- ret = bch2_lru_change(trans, new->k.p.inode,
- bucket_to_u64(new->k.p),
- old_lru, new_lru);
- if (ret)
- return ret;
- }
+ u64 old_lru = alloc_lru_idx_read(*old_a);
+ u64 new_lru = alloc_lru_idx_read(*new_a);
+ if (old_lru != new_lru) {
+ ret = bch2_lru_change(trans, new.k->p.inode,
+ bucket_to_u64(new.k->p),
+ old_lru, new_lru);
+ if (ret)
+ return ret;
+ }
- new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
- bch_dev_bkey_exists(c, new->k.p.inode));
+ new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
+ bch_dev_bkey_exists(c, new.k->p.inode));
+ if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
+ ret = bch2_lru_change(trans,
+ BCH_LRU_FRAGMENTATION_START,
+ bucket_to_u64(new.k->p),
+ old_a->fragmentation_lru, new_a->fragmentation_lru);
+ if (ret)
+ return ret;
+ }
- if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
- ret = bch2_lru_change(trans,
- BCH_LRU_FRAGMENTATION_START,
- bucket_to_u64(new->k.p),
- old_a->fragmentation_lru, new_a->fragmentation_lru);
- if (ret)
- return ret;
+ if (old_a->gen != new_a->gen) {
+ ret = bch2_bucket_gen_update(trans, new.k->p, new_a->gen);
+ if (ret)
+ return ret;
+ }
+
+ /*
+ * need to know if we're getting called from the invalidate path or
+ * not:
+ */
+
+ if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
+ old_a->cached_sectors) {
+ ret = bch2_update_cached_sectors_list(trans, new.k->p.inode,
+ -((s64) old_a->cached_sectors));
+ if (ret)
+ return ret;
+ }
}
- if (old_a->gen != new_a->gen) {
- ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
- if (ret)
- return ret;
+ if ((flags & BTREE_TRIGGER_ATOMIC) && (flags & BTREE_TRIGGER_INSERT)) {
+ struct bch_alloc_v4 *new_a = bkey_s_to_alloc_v4(new).v;
+ u64 journal_seq = trans->journal_res.seq;
+ u64 bucket_journal_seq = new_a->journal_seq;
+
+ if ((flags & BTREE_TRIGGER_INSERT) &&
+ data_type_is_empty(old_a->data_type) !=
+ data_type_is_empty(new_a->data_type) &&
+ new.k->type == KEY_TYPE_alloc_v4) {
+ struct bch_alloc_v4 *v = bkey_s_to_alloc_v4(new).v;
+
+ /*
+ * If the btree updates referring to a bucket weren't flushed
+ * before the bucket became empty again, then the we don't have
+ * to wait on a journal flush before we can reuse the bucket:
+ */
+ v->journal_seq = bucket_journal_seq =
+ data_type_is_empty(new_a->data_type) &&
+ (journal_seq == v->journal_seq ||
+ bch2_journal_noflush_seq(&c->journal, v->journal_seq))
+ ? 0 : journal_seq;
+ }
+
+ if (!data_type_is_empty(old_a->data_type) &&
+ data_type_is_empty(new_a->data_type) &&
+ bucket_journal_seq) {
+ ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
+ c->journal.flushed_seq_ondisk,
+ new.k->p.inode, new.k->p.offset,
+ bucket_journal_seq);
+ if (ret) {
+ bch2_fs_fatal_error(c,
+ "error setting bucket_needs_journal_commit: %i", ret);
+ return ret;
+ }
+ }
+
+ percpu_down_read(&c->mark_lock);
+ if (new_a->gen != old_a->gen)
+ *bucket_gen(ca, new.k->p.offset) = new_a->gen;
+
+ bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, false);
+ percpu_up_read(&c->mark_lock);
+
+#define eval_state(_a, expr) ({ const struct bch_alloc_v4 *a = _a; expr; })
+#define statechange(expr) !eval_state(old_a, expr) && eval_state(new_a, expr)
+#define bucket_flushed(a) (!a->journal_seq || a->journal_seq <= c->journal.flushed_seq_ondisk)
+
+ if (statechange(a->data_type == BCH_DATA_free) &&
+ bucket_flushed(new_a))
+ closure_wake_up(&c->freelist_wait);
+
+ if (statechange(a->data_type == BCH_DATA_need_discard) &&
+ !bch2_bucket_is_open(c, new.k->p.inode, new.k->p.offset) &&
+ bucket_flushed(new_a))
+ bch2_discard_one_bucket_fast(c, new.k->p);
+
+ if (statechange(a->data_type == BCH_DATA_cached) &&
+ !bch2_bucket_is_open(c, new.k->p.inode, new.k->p.offset) &&
+ should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
+ bch2_do_invalidates(c);
+
+ if (statechange(a->data_type == BCH_DATA_need_gc_gens))
+ bch2_do_gc_gens(c);
}
- /*
- * need to know if we're getting called from the invalidate path or
- * not:
- */
+ if ((flags & BTREE_TRIGGER_GC) &&
+ (flags & BTREE_TRIGGER_BUCKET_INVALIDATE)) {
+ struct bch_alloc_v4 new_a_convert;
+ const struct bch_alloc_v4 *new_a = bch2_alloc_to_v4(new.s_c, &new_a_convert);
- if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
- old_a->cached_sectors) {
- ret = bch2_update_cached_sectors_list(trans, new->k.p.inode,
- -((s64) old_a->cached_sectors));
- if (ret)
- return ret;
+ percpu_down_read(&c->mark_lock);
+ struct bucket *g = gc_bucket(ca, new.k->p.offset);
+
+ bucket_lock(g);
+
+ g->gen_valid = 1;
+ g->gen = new_a->gen;
+ g->data_type = new_a->data_type;
+ g->stripe = new_a->stripe;
+ g->stripe_redundancy = new_a->stripe_redundancy;
+ g->dirty_sectors = new_a->dirty_sectors;
+ g->cached_sectors = new_a->cached_sectors;
+
+ bucket_unlock(g);
+ percpu_up_read(&c->mark_lock);
}
return 0;
unsigned i, gens_offset, gens_end_offset;
int ret;
- if (c->sb.version < bcachefs_metadata_version_bucket_gens)
- return 0;
-
bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
k = bch2_btree_iter_peek_slot(bucket_gens_iter);
return ret;
}
+static int discard_in_flight_add(struct bch_fs *c, struct bpos bucket)
+{
+ int ret;
+
+ mutex_lock(&c->discard_buckets_in_flight_lock);
+ darray_for_each(c->discard_buckets_in_flight, i)
+ if (bkey_eq(*i, bucket)) {
+ ret = -EEXIST;
+ goto out;
+ }
+
+ ret = darray_push(&c->discard_buckets_in_flight, bucket);
+out:
+ mutex_unlock(&c->discard_buckets_in_flight_lock);
+ return ret;
+}
+
+static void discard_in_flight_remove(struct bch_fs *c, struct bpos bucket)
+{
+ mutex_lock(&c->discard_buckets_in_flight_lock);
+ darray_for_each(c->discard_buckets_in_flight, i)
+ if (bkey_eq(*i, bucket)) {
+ darray_remove_item(&c->discard_buckets_in_flight, i);
+ goto found;
+ }
+ BUG();
+found:
+ mutex_unlock(&c->discard_buckets_in_flight_lock);
+}
+
+struct discard_buckets_state {
+ u64 seen;
+ u64 open;
+ u64 need_journal_commit;
+ u64 discarded;
+ struct bch_dev *ca;
+ u64 need_journal_commit_this_dev;
+};
+
+static void discard_buckets_next_dev(struct bch_fs *c, struct discard_buckets_state *s, struct bch_dev *ca)
+{
+ if (s->ca == ca)
+ return;
+
+ if (s->ca && s->need_journal_commit_this_dev >
+ bch2_dev_usage_read(s->ca).d[BCH_DATA_free].buckets)
+ bch2_journal_flush_async(&c->journal, NULL);
+
+ if (s->ca)
+ percpu_ref_put(&s->ca->ref);
+ if (ca)
+ percpu_ref_get(&ca->ref);
+ s->ca = ca;
+ s->need_journal_commit_this_dev = 0;
+}
+
static int bch2_discard_one_bucket(struct btree_trans *trans,
struct btree_iter *need_discard_iter,
struct bpos *discard_pos_done,
- u64 *seen,
- u64 *open,
- u64 *need_journal_commit,
- u64 *discarded)
+ struct discard_buckets_state *s)
{
struct bch_fs *c = trans->c;
struct bpos pos = need_discard_iter->pos;
struct bch_dev *ca;
struct bkey_i_alloc_v4 *a;
struct printbuf buf = PRINTBUF;
+ bool discard_locked = false;
int ret = 0;
ca = bch_dev_bkey_exists(c, pos.inode);
+
if (!percpu_ref_tryget(&ca->io_ref)) {
bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
return 0;
}
+ discard_buckets_next_dev(c, s, ca);
+
if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
- (*open)++;
+ s->open++;
goto out;
}
if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
c->journal.flushed_seq_ondisk,
pos.inode, pos.offset)) {
- (*need_journal_commit)++;
+ s->need_journal_commit++;
+ s->need_journal_commit_this_dev++;
goto out;
}
goto out;
}
+ if (discard_in_flight_add(c, SPOS(iter.pos.inode, iter.pos.offset, true)))
+ goto out;
+
+ discard_locked = true;
+
if (!bkey_eq(*discard_pos_done, iter.pos) &&
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);
+ bch2_trans_unlock_long(trans);
blkdev_issue_discard(ca->disk_sb.bdev,
k.k->p.offset * ca->mi.bucket_size,
ca->mi.bucket_size,
goto out;
count_event(c, bucket_discard);
- (*discarded)++;
+ s->discarded++;
out:
- (*seen)++;
+ if (discard_locked)
+ discard_in_flight_remove(c, iter.pos);
+ s->seen++;
bch2_trans_iter_exit(trans, &iter);
percpu_ref_put(&ca->io_ref);
printbuf_exit(&buf);
static void bch2_do_discards_work(struct work_struct *work)
{
struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
- u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
+ struct discard_buckets_state s = {};
struct bpos discard_pos_done = POS_MAX;
int ret;
ret = bch2_trans_run(c,
for_each_btree_key(trans, iter,
BTREE_ID_need_discard, POS_MIN, 0, k,
- bch2_discard_one_bucket(trans, &iter, &discard_pos_done,
- &seen,
- &open,
- &need_journal_commit,
- &discarded)));
-
- if (need_journal_commit * 2 > seen)
- bch2_journal_flush_async(&c->journal, NULL);
+ bch2_discard_one_bucket(trans, &iter, &discard_pos_done, &s)));
- bch2_write_ref_put(c, BCH_WRITE_REF_discard);
+ discard_buckets_next_dev(c, &s, NULL);
- trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
+ trace_discard_buckets(c, s.seen, s.open, s.need_journal_commit, s.discarded,
bch2_err_str(ret));
+
+ bch2_write_ref_put(c, BCH_WRITE_REF_discard);
}
void bch2_do_discards(struct bch_fs *c)
bch2_write_ref_put(c, BCH_WRITE_REF_discard);
}
+static int bch2_clear_bucket_needs_discard(struct btree_trans *trans, struct bpos bucket)
+{
+ struct btree_iter iter;
+ bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, bucket, BTREE_ITER_INTENT);
+ struct bkey_s_c k = bch2_btree_iter_peek_slot(&iter);
+ int ret = bkey_err(k);
+ if (ret)
+ goto err;
+
+ struct bkey_i_alloc_v4 *a = bch2_alloc_to_v4_mut(trans, k);
+ ret = PTR_ERR_OR_ZERO(a);
+ if (ret)
+ goto err;
+
+ SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
+ a->v.data_type = alloc_data_type(a->v, a->v.data_type);
+
+ ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
+err:
+ bch2_trans_iter_exit(trans, &iter);
+ return ret;
+}
+
+static void bch2_do_discards_fast_work(struct work_struct *work)
+{
+ struct bch_fs *c = container_of(work, struct bch_fs, discard_fast_work);
+
+ while (1) {
+ bool got_bucket = false;
+ struct bpos bucket;
+ struct bch_dev *ca;
+
+ mutex_lock(&c->discard_buckets_in_flight_lock);
+ darray_for_each(c->discard_buckets_in_flight, i) {
+ if (i->snapshot)
+ continue;
+
+ ca = bch_dev_bkey_exists(c, i->inode);
+
+ if (!percpu_ref_tryget(&ca->io_ref)) {
+ darray_remove_item(&c->discard_buckets_in_flight, i);
+ continue;
+ }
+
+ got_bucket = true;
+ bucket = *i;
+ i->snapshot = true;
+ break;
+ }
+ mutex_unlock(&c->discard_buckets_in_flight_lock);
+
+ if (!got_bucket)
+ break;
+
+ blkdev_issue_discard(ca->disk_sb.bdev,
+ bucket.offset * ca->mi.bucket_size,
+ ca->mi.bucket_size,
+ GFP_KERNEL);
+
+ int ret = bch2_trans_do(c, NULL, NULL,
+ BCH_WATERMARK_btree|
+ BCH_TRANS_COMMIT_no_enospc,
+ bch2_clear_bucket_needs_discard(trans, bucket));
+ bch_err_fn(c, ret);
+
+ percpu_ref_put(&ca->io_ref);
+ discard_in_flight_remove(c, bucket);
+
+ if (ret)
+ break;
+ }
+
+ bch2_write_ref_put(c, BCH_WRITE_REF_discard_fast);
+}
+
+static void bch2_discard_one_bucket_fast(struct bch_fs *c, struct bpos bucket)
+{
+ struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
+
+ if (!percpu_ref_is_dying(&ca->io_ref) &&
+ !discard_in_flight_add(c, bucket) &&
+ bch2_write_ref_tryget(c, BCH_WRITE_REF_discard_fast) &&
+ !queue_work(c->write_ref_wq, &c->discard_fast_work))
+ bch2_write_ref_put(c, BCH_WRITE_REF_discard_fast);
+}
+
static int invalidate_one_bucket(struct btree_trans *trans,
struct btree_iter *lru_iter,
struct bkey_s_c lru_k,
set_bit(ca->dev_idx, c->rw_devs[i].d);
}
+void bch2_fs_allocator_background_exit(struct bch_fs *c)
+{
+ darray_exit(&c->discard_buckets_in_flight);
+}
+
void bch2_fs_allocator_background_init(struct bch_fs *c)
{
spin_lock_init(&c->freelist_lock);
+ mutex_init(&c->discard_buckets_in_flight_lock);
INIT_WORK(&c->discard_work, bch2_do_discards_work);
+ INIT_WORK(&c->discard_fast_work, bch2_do_discards_fast_work);
INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
}
projects / bcachefs-tools-debian / blobdiff