Bug #21166
Updated by ioquatix (Samuel Williams) 8 months ago
## Background
Ruby's `IO#close` can cause `IO#read`, `IO#write`, `IO#wait`, `IO#wait_readable` and `IO#wait_writable` to be interrupted with an IOError (stream closed in another thread). For reference, `IO#select` cannot be interrupted in this way.
```ruby
r, w = IO.pipe
thread = Thread.new do
r.read(1)
end
Thread.pass until thread.status == "sleep"
r.close
thread.join
# ./test.rb:6:in 'IO#read': stream closed in another thread (IOError)
```
## Problem
The fiber scheduler provides hooks for `io_read`, `io_write` and `io_wait` which are used by `IO#read`, `IO#write`, `IO#wait`, `IO#wait_readable` and `IO#wait_writable`, but those hooks are not interrupted when `IO#close` is invoked. That is because `rb_notify_fd_close` is not scheduler aware, and the fiber scheduler is unable to register itself into the "waiting file descriptor" list.
```ruby
#!/usr/bin/env ruby
require 'async'
r, w = IO.pipe
thread = Thread.new do
Async do
r.wait_readable
end
end
Thread.pass until thread.status == "sleep"
r.close
thread.join
```
In this test program, `rb_notify_fd_close` will incorrectly terminate the entire fiber scheduler thread:
```
#<Thread:0x00007faa5b161bf8 /home/samuel/Developer/socketry/io-event/test.rb:7 run> terminated with exception (report_on_exception is true):
/home/samuel/Developer/socketry/io-event/lib/io/event/selector/select.rb:470:in 'IO.select': closed stream (IOError)
from /home/samuel/Developer/socketry/io-event/lib/io/event/selector/select.rb:470:in 'block in IO::Event::Selector::Select#select'
from /home/samuel/Developer/socketry/io-event/lib/io/event/selector/select.rb:468:in 'Thread.handle_interrupt'
from /home/samuel/Developer/socketry/io-event/lib/io/event/selector/select.rb:468:in 'IO::Event::Selector::Select#select'
from /home/samuel/.gem/ruby/3.4.1/gems/async-2.23.0/lib/async/scheduler.rb:396:in 'Async::Scheduler#run_once!'
...
```
## Solution
This PR introduces some new functions:
- `VALUE rb_io_interruptible_operation(VALUE self, VALUE(*function)(VALUE), VALUE argument)` for wrapping user IO operations so they can be interrupted.
- `IO#interruptable_operation(&block)` the same as above.
- `VALUE rb_fiber_scheduler_fiber_interrupt(VALUE scheduler, VALUE fiber, VALUE exception)` for interrupting a specific fiber on the fiber scheduler.
`rb_notify_fd_close` is modified so that it is fiber scheduler aware and uses `rb_fiber_scheduler_fiber_interrupt` to interrupt a fiber. In addition, we also change the internal `struct waiting_fd` to track the `rb_execution_context_t` rather than just the `rb_thread_t` instance, so that we can correctly wake up either the waiting thread or fiber.
The public interface `rb_io_interruptible_operation` and `IO#interruptible_operation` are introduced so that the scheduler implementation can wrap IO operations that should be interruptible, e.g.
```ruby
Fiber.schedule do
io.interruptible_operation do
io.wait_readable
end
end
# Will interrupt above fiber:
io.close
```
See <https://github.com/ruby/ruby/pull/12585> for the proposed implementation and <https://github.com/socketry/io-event/pull/130> for example of how `io-event` gem uses both the C and Ruby interfaces. implementation.