Signed-off-by: Kent Overstreet <kmo@daterainc.com>
else if (!mutex_trylock(&c->bucket_lock))
return -1;
+ /*
+ * It's _really_ critical that we don't free too many btree nodes - we
+ * have to always leave ourselves a reserve. The reserve is how we
+ * guarantee that allocating memory for a new btree node can always
+ * succeed, so that inserting keys into the btree can always succeed and
+ * IO can always make forward progress:
+ */
nr /= c->btree_pages;
nr = min_t(unsigned long, nr, mca_can_free(c));