Commit d5425a3b authored by Kent Overstreet's avatar Kent Overstreet Committed by Kent Overstreet

bcachefs: Throttle updates when btree key cache is too dirty

This is needed to ensure we don't deadlock because journal reclaim and
thus memory reclaim isn't making forward progress.
Signed-off-by: default avatarKent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: default avatarKent Overstreet <kent.overstreet@linux.dev>
parent 9d4582ff
...@@ -5,11 +5,20 @@ static inline size_t bch2_nr_btree_keys_need_flush(struct bch_fs *c) ...@@ -5,11 +5,20 @@ static inline size_t bch2_nr_btree_keys_need_flush(struct bch_fs *c)
{ {
size_t nr_dirty = READ_ONCE(c->btree_key_cache.nr_dirty); size_t nr_dirty = READ_ONCE(c->btree_key_cache.nr_dirty);
size_t nr_keys = READ_ONCE(c->btree_key_cache.nr_dirty); size_t nr_keys = READ_ONCE(c->btree_key_cache.nr_dirty);
size_t max_dirty = 1024 + (nr_keys * 3) / 4; size_t max_dirty = 4096 + nr_keys / 2;
return max_t(ssize_t, 0, nr_dirty - max_dirty); return max_t(ssize_t, 0, nr_dirty - max_dirty);
} }
static inline bool bch2_btree_key_cache_must_wait(struct bch_fs *c)
{
size_t nr_dirty = READ_ONCE(c->btree_key_cache.nr_dirty);
size_t nr_keys = READ_ONCE(c->btree_key_cache.nr_dirty);
size_t max_dirty = 4096 + (nr_keys * 3) / 4;
return nr_dirty > max_dirty;
}
struct bkey_cached * struct bkey_cached *
bch2_btree_key_cache_find(struct bch_fs *, enum btree_id, struct bpos); bch2_btree_key_cache_find(struct bch_fs *, enum btree_id, struct bpos);
......
...@@ -649,6 +649,7 @@ enum btree_insert_ret { ...@@ -649,6 +649,7 @@ enum btree_insert_ret {
BTREE_INSERT_ENOSPC, BTREE_INSERT_ENOSPC,
BTREE_INSERT_NEED_MARK_REPLICAS, BTREE_INSERT_NEED_MARK_REPLICAS,
BTREE_INSERT_NEED_JOURNAL_RES, BTREE_INSERT_NEED_JOURNAL_RES,
BTREE_INSERT_NEED_JOURNAL_RECLAIM,
}; };
enum btree_gc_coalesce_fail_reason { enum btree_gc_coalesce_fail_reason {
......
...@@ -286,6 +286,10 @@ btree_key_can_insert_cached(struct btree_trans *trans, ...@@ -286,6 +286,10 @@ btree_key_can_insert_cached(struct btree_trans *trans,
BUG_ON(iter->level); BUG_ON(iter->level);
if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
bch2_btree_key_cache_must_wait(trans->c))
return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
if (u64s <= ck->u64s) if (u64s <= ck->u64s)
return BTREE_INSERT_OK; return BTREE_INSERT_OK;
...@@ -652,6 +656,21 @@ int bch2_trans_commit_error(struct btree_trans *trans, ...@@ -652,6 +656,21 @@ int bch2_trans_commit_error(struct btree_trans *trans,
trace_trans_restart_journal_res_get(trans->ip); trace_trans_restart_journal_res_get(trans->ip);
ret = -EINTR; ret = -EINTR;
break; break;
case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
bch2_trans_unlock(trans);
while (bch2_btree_key_cache_must_wait(c)) {
mutex_lock(&c->journal.reclaim_lock);
bch2_journal_reclaim(&c->journal);
mutex_unlock(&c->journal.reclaim_lock);
}
if (bch2_trans_relock(trans))
return 0;
trace_trans_restart_journal_reclaim(trans->ip);
ret = -EINTR;
break;
default: default:
BUG_ON(ret >= 0); BUG_ON(ret >= 0);
break; break;
......
...@@ -681,6 +681,11 @@ DEFINE_EVENT(transaction_restart, trans_restart_journal_preres_get, ...@@ -681,6 +681,11 @@ DEFINE_EVENT(transaction_restart, trans_restart_journal_preres_get,
TP_ARGS(ip) TP_ARGS(ip)
); );
DEFINE_EVENT(transaction_restart, trans_restart_journal_reclaim,
TP_PROTO(unsigned long ip),
TP_ARGS(ip)
);
DEFINE_EVENT(transaction_restart, trans_restart_mark_replicas, DEFINE_EVENT(transaction_restart, trans_restart_mark_replicas,
TP_PROTO(unsigned long ip), TP_PROTO(unsigned long ip),
TP_ARGS(ip) TP_ARGS(ip)
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment