Bug #20723
closed`IO#close` is broken on Ruby 3.3+ when using the Fiber scheduler.
Description
The following program seems to work okay on Ruby 3.2 but hangs on Ruby 3.3:
#!/usr/bin/env ruby
require 'bundler/inline'
gemfile do
source 'https://rubygems.org'
gem 'async'
end
require 'socket'
def close_while_reading(io)
thread = Thread.new do
Thread.current.report_on_exception = false
io.wait_readable
end
# Wait until the thread is blocked on read:
Thread.pass until thread.status == "sleep"
Async do
io.close
end
thread.join
end
begin
client, server = Socket.pair(:UNIX, :STREAM)
close_while_reading(client)
rescue => error
$stderr.puts error.full_message
end
Updated by ioquatix (Samuel Williams) about 2 months ago
It may be related to https://github.com/ruby/ruby/commit/66871c5a06d723f8350935ced1e88d8cc929d809
Updated by ioquatix (Samuel Williams) about 2 months ago
I've added the following work-around: https://github.com/socketry/io-stream/commit/7d1546fa829d3fe046f66f559d9a774497390f3e
Updated by kjtsanaktsidis (KJ Tsanaktsidis) about 2 months ago
Sorry about this. I think https://github.com/ruby/ruby/pull/11614 is the smallest diff that will fix the issue (and this should probably be backported to 3.3).
Separately to that, I wonder if we need to wrap up some common function for "wake up this fiber, either with the fiber scheduler or with the thread directly". We've implemented this logic in a number of places... but let's keep that refactor out of the fix PR (since it shouldn't be backported).
Updated by kjtsanaktsidis (KJ Tsanaktsidis) about 2 months ago
- Status changed from Open to Closed
Applied in changeset git|e08d5239b68ad61a731f4938cf963e37a5e88c25.
Ensure fiber scheduler is woken up when close interrupts read
If one thread is reading and another closes that socket, the close
blocks waiting for the read to abort cleanly. This ensures that Ruby is
totally done with the file descriptor BEFORE we tell the OS to close
and potentially re-use it.
When the read is correctly terminated, the close should be unblocked.
That currently works if closing is happening on a thread, but if it's
happening on a fiber with a fiber scheduler, it does NOT work.
This patch ensures that if the close happened in a fiber scheduled
thread, that the scheduler is notified that the fiber is unblocked.
[Bug #20723]
Updated by kjtsanaktsidis (KJ Tsanaktsidis) about 1 month ago
Backport PR for 3.3 - https://github.com/ruby/ruby/pull/11664
No backport for 3.2 is required because this locking around close didn't exist there.
Updated by k0kubun (Takashi Kokubun) about 2 hours ago
- Backport changed from 3.3: REQUIRED to 3.3: DONE
ruby_3_3 5b6009870dff883a8e71a05e60f175cea1d00d55.