-
Kent Overstreet authored
We've outgrown our own deadlock avoidance strategy. The btree iterator API provides an interface where the user doesn't need to concern themselves with lock ordering - different btree iterators can be traversed in any order. Without special care, this will lead to deadlocks. Our previous strategy was to define a lock ordering internally, and whenever we attempt to take a lock and trylock() fails, we'd check if the current btree transaction is holding any locks that cause a lock ordering violation. If so, we'd issue a transaction restart, and then bch2_trans_begin() would re-traverse all previously used iterators, but in the correct order. That approach had some issues, though. - Sometimes we'd issue transaction restarts unnecessarily, when no deadlock would have actually occured. Lock ordering restarts have become our primary cause of transaction restarts, on some workloads totally 20% of actual transaction commits. - To avoid deadlock or livelock, we'd often have to take intent locks when we only wanted a read lock: with the lock ordering approach, it is actually illegal to hold _any_ read lock while blocking on an intent lock, and this has been causing us unnecessary lock contention. - It was getting fragile - the various lock ordering rules are not trivial, and we'd been seeing occasional livelock issues related to this machinery. So, since bcachefs is already a relational database masquerading as a filesystem, we're stealing the next traditional database technique and switching to a cycle detector for avoiding deadlocks. When we block taking a btree lock, after adding ourself to the waitlist but before sleeping, we do a DFS of btree transactions waiting on other btree transactions, starting with the current transaction and walking our held locks, and transactions blocking on our held locks. If we find a cycle, we emit a transaction restart. Occasionally (e.g. the btree split path) we can not allow the lock() operation to fail, so if necessary we'll tell another transaction that it has to fail. Result: trans_restart_would_deadlock events are reduced by a factor of 10 to 100, and we'll be able to delete a whole bunch of grotty, fragile code. Signed-off-by:
Kent Overstreet <kent.overstreet@gmail.com>
33bd5d06