From: "Eregon (Benoit Daloze)" <noreply@...>
Date: 2022-11-23T11:09:28+00:00
Subject: [ruby-core:110866] [Ruby master Feature#19141] Add thread-owned Monitor to protect thread-local resources

Issue #19141 has been updated by Eregon (Benoit Daloze).


chrisseaton (Chris Seaton) wrote in #note-11:
> Maybe `Enumerator` should not use a full fiber - but something less that behaves like the thread it's run on?

It still needs its own stack so it can suspend anywhere.
For Enumerators created by e.g. `Array#each` and other Enumerable methods without block for a core enumerable object it may be fine to not have its own Fiber, but for `Enumerator.new` or with a custom (not core) `each` that's much less clear, because there can be arbitrary logic in such an Enumerator, notably those database accesses in that Rails issue. And in that case it might not be correct to share the fiber locals, the owned locks, use the same Fiber for `Fiber.current`, etc.
It would also mean `Fiber.current` inside that Enumerator could be changing (if Enumerator#next is called by different Fibers).

> An underlying issue is that the user doesn't know they're using a fiber, and really how can we expect them to?

We should document that in `Enumerator#{next*,peek*}` clearly and the pitfalls of it.
There is already a mention of it but it's no obvious (`See class-level notes about external iterators.`).

----------------------------------------
Feature #19141: Add thread-owned Monitor to protect thread-local resources
https://bugs.ruby-lang.org/issues/19141#change-100226

* Author: wildmaples (Maple Ong)
* Status: Open
* Priority: Normal
----------------------------------------
### Background 

In Ruby v3.0.2, Monitor was modified to be owned by fibers instead of threads [for reasons as described in this issue](https://bugs.ruby-lang.org/issues/17827) and so it is also consistent with Mutex. Before the change to Monitor, Mutex was modified to per-fiber instead of thread ([issue](https://bugs.ruby-lang.org/issues/16792), [PR](https://github.com/ruby/ruby/commit/178c1b0922dc727897d81d7cfe9c97d5ffa97fd9)) which caused some problems (See: [comment](https://bugs.ruby-lang.org/issues/17827#note-1)).  

### Problem

We are now encountering a problem where using Enumerator (implemented transparently using Fiber, so the user is not aware) within a Fiber-owned process, which causes a deadlock. That means any framework using Monitor is incompatible to be used with Enumerator. 

In general, there are many types of thread-local resources (connections for example), so it would make sense to have a thread-owned monitor to protect them. I think few resources are really fiber-owned.

#### Specifics 
* Concurrent Ruby is still designed with per-thread locking, which causes similar incompatibilities. (Read: [issue](https://github.com/ruby-concurrency/concurrent-ruby/issues/962))
* Systems test in Rails implements locking using Monitor, resulting in deadlock in these known cases:
  * when cache clearing (Read: [issue](https://github.com/rails/rails/issues/45994))
  * database transactions when used with Enumerator (Read: [comment](https://github.com/rails/rails/issues/45994#issuecomment-1304306575)) 

### Demo

```ruby
# ruby 2.7.6p219 (2022-04-12 revision c9c2245c0a) [arm64-darwin21]
# Thread #<Thread:0x000000014a8eb228 demo.rb:8 run>, fiber #<Fiber:0x000000014a8eaf80 (resumed)>, locked true, owned true
# Thread #<Thread:0x000000014a8eb228 demo.rb:8 run>, fiber #<Fiber:0x000000014a8eacb0 demo.rb:13 (resumed)>, locked true, owned true

# ruby 3.1.2p20 (2022-04-12 revision 4491bb740a) [arm64-darwin21]
# Thread #<Thread:0x0000000102329a08 demo.rb:8 run>, fiber #<Fiber:0x0000000102329828 (resumed)>, locked true, owned true
# Thread #<Thread:0x0000000102329a08 demo.rb:8 run>, fiber #<Fiber:0x00000001023294e0 demo.rb:13 (resumed)>, locked true, owned false

require 'fiber'
require 'monitor'

puts RUBY_DESCRIPTION

# We have a single monitor - we're pretending it protects some thread-local resources
m = Monitor.new

# We'll create an explicit thread
t = Thread.new do
  # Lock to protect our thread-local resource
  m.enter

  puts "Thread #{Thread.current}, fiber #{Fiber.current}, locked #{m.mon_locked?}, owned #{m.mon_owned?}"

  # The Enumerator here creates a fiber, which runs on the same thread, so would want to use the same thread-local resource
  e = Enumerator.new do |y|
    # In 2.7 this is fine, in 3.0 it's not, as the fiber thinks it doesn't have the lock
    puts "Thread #{Thread.current}, fiber #{Fiber.current}, locked #{m.mon_locked?}, owned #{m.mon_owned?}"
    
    # This would deadlock
    # m.enter

    y.yield 1
  end
  e.next
end

t.join
```

### Possible Solutions

* Allow `Monitor` to be per thread or fiber through a flag
* Having `Thread::Monitor` and `Fiber::Monitor` as two separate classes. Leave `Monitor` as it is right now. However, this may not be possible due to the `Thread::Mutex` alias 

These options would give us more flexibility in which type of Monitor to use. 



-- 
https://bugs.ruby-lang.org/

Unsubscribe: <mailto:ruby-core-request@ruby-lang.org?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>