[ruby-core:98304] [Ruby master Feature#16792] Make Mutex held per Fiber instead of per Thread
From:
ko1@...
Date:
2020-05-12 22:21:09 UTC
List:
ruby-core #98304
Issue #16792 has been updated by ko1 (Koichi Sasada). I think it seems difficult to implement it. If an interpreter manages everything, it is easy (at least I can image how to implement it). (1) API I'm not sure we can implement Mutex scheduling with the hooks introduced at #10. If there is only one thread, maybe it is easy to implement. Maybe `notify_mutex(mutex)` will be called with locked mutex by the interpreter. It means we need to introduce lock_nonblock or similar API for Mutex (maybe `_nonblock` is not good name because it is different from `IO#read_nonblock`). Also we need to wait this notification with wait for the ready of IO operations. how to write it? Use pipe trick or interrupt mechanism? (2) Queue/SizedQueue/CV there are several blocking operations because of the thread synchronization, how to treat them? Introduce unified hook method like `wait_synchronization(sync_object)` and `notify_synchronization(sync_object)`? Introduce `pop_nonblock` similar to `Mutex#sync_nonblock`? (3) implemented by all schedulers? maybe most of code are same between scheduler implementations. can we provide a framework to implement it? ---------------------------------------- Feature #16792: Make Mutex held per Fiber instead of per Thread https://bugs.ruby-lang.org/issues/16792#change-85542 * Author: Eregon (Benoit Daloze) * Status: Open * Priority: Normal ---------------------------------------- Currently, Mutex in CRuby is held per Thread. In JRuby and TruffleRuby, Mutex is held per Fiber (because it's simply easier implementation-wise). While a user could theoretically notice the difference, it seems extremely uncommon in practice (probably incorrect synchronization). The usage pattern for a Mutex is using #synchronize or lock+unlock. Such a pattern protects/surrounds a region of code using some resource, and such a region of code is always on the same Fiber since it's on a given Ruby "stack". With #16786 it becomes more relevant to have Mutex held per Fiber, otherwise Mutex#lock will hurt scalability of that proposal significantly. This means, if a Fiber does Mutex#lock and it's already held by another Fiber of the same Thread, and the Thread#scheduler is enabled, instead of just raising an error (which made sense before, because it would be a deadlock, but no longer the case with scheduler), or disabling fiber scheduling entirely until #unlock (current state in #16786, makes Mutex#lock special and hurts scalability), we would just go to the scheduler and schedule another Fiber (for instance, the one holding that Mutex, or any other ready to be run Fiber). This is not a new idea and in fact Crystal already does this with its non-blocking Fibers, which is very similar with #16786: https://github.com/crystal-lang/crystal/blob/612825a53c831ce7d17368c8211342b199ca02ff/src/mutex.cr#L72 Mutex#lock is just like other blocking operations, so let's make it so building on #16786. I believe it's the natural and intuitive thing to do for Fiber concurrency with a scheduler. Queue#pop and SizedQueue#push could be other candidates to handle in a similar way. Here is an early commit to make Mutex held per Fiber, it's quite trivial as you can see: https://github.com/ruby/ruby/compare/master...eregon:mutex-per-fiber It passes test-all and test-spec. -- https://bugs.ruby-lang.org/ Unsubscribe: <mailto:[email protected]?subject=unsubscribe> <http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>
