This helps ensure key cache reclaim isn't contending with threads
waiting for the key cache to be helped, and fixes a severe performance
bug.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
return nr_dirty > max_dirty;
}
+static inline bool bch2_btree_key_cache_wait_done(struct bch_fs *c)
+{
+ size_t nr_dirty = atomic_long_read(&c->btree_key_cache.nr_dirty);
+ size_t nr_keys = atomic_long_read(&c->btree_key_cache.nr_keys);
+ size_t max_dirty = 2048 + (nr_keys * 5) / 8;
+
+ return nr_dirty <= max_dirty;
+}
+
int bch2_btree_key_cache_journal_flush(struct journal *,
struct journal_entry_pin *, u64);
static int journal_reclaim_wait_done(struct bch_fs *c)
{
int ret = bch2_journal_error(&c->journal) ?:
- !bch2_btree_key_cache_must_wait(c);
+ bch2_btree_key_cache_wait_done(c);
if (!ret)
journal_reclaim_kick(&c->journal);