[ruby-core:103580] [Ruby master Bug#17527] rb_io_wait_readable/writable with scheduler don't check errno

Issue #17527 has been updated by nagachika (Tomoyuki Chikanaga).


Added https://bugs.ruby-lang.org/projects/ruby-master/repository/git/revisions/a9c5c2d614f30a616970245fef3e7ffc151e2ecf as a related commit.
I will backport the changesets.

----------------------------------------
Bug #17527: rb_io_wait_readable/writable with scheduler don't check errno 
https://bugs.ruby-lang.org/issues/17527#change-91679

* Author: ysbaddaden (Julien Portalier)
* Status: Closed
* Priority: Normal
* Assignee: ioquatix (Samuel Williams)
* ruby -v: ruby 3.0.0p0 (2020-12-25 revision 95aff21468) [x86_64-linux]
* Backport: 2.5: DONTNEED, 2.6: DONTNEED, 2.7: DONTNEED, 3.0: REQUIRED
----------------------------------------
## Problem

Playing with the new Fiber Scheduler, I noticed that `TCPServer#accept` would hung forever after closing the server from another Fiber. I expected it to be resumed and fail with IOError, as it happens with threads.

## Analysis

What happens is that the `accept4` call in `rsock_s_accept` fails and sets errno to `Errno::EBADF`, it then checks a few memory/limit related errnos, then calls `rb_io_wait_readable` expecting it to handle the current errno for IO errors. But when a scheduler is set, it immediately delegates to `Scheduler#io_wait` and doesn't check the current errno! In my case (nio4r), the `io_wait` hook returns a ready state, which causes `rsock_s_accept` to loop forever.

I tried to manually check in the `io_wait` hook whether the IO is closed, but the fd is never updated (AFAIK never set to -1) so `io.closed?` is always false. I'm not sure schedulers should check whether the fd is closed, thought.

## Proposed solution

A solution is to follow what happens for threads, and only check the scheduler when errno is EAGAIN or EWOULDBLOCK. I believe it's the only errors where we're expected to wait. This change also means that EINTR will be handled, too, and other errnos to raise an exception.

Instead of raising `IOError.new("closed stream")` as it happens for threads, it raises `Errno::EBADF` when a Scheduler is set. I suppose in the thread branches, it updates the IO at some point and calls `rb_io_check_closed` with the updated fd —maybe with `GetOpenFile` (`RB_IO_POINTER`) — and we ought to do the same at some point?

Another solution it to not delegate to the scheduler inside `rb_io_wait_readable` because it will eventually call `rb_wait_for_single_fd` that will check for the scheduler, but we can avoid some function calls, as well as thread-related debug information that could be confusing. It also won't raise help to raise IOError.

I'm attaching a patch that implements the first solution. It fixes both `rb_io_wait_readable` and `rb_io_wait_writable` since the latter may exhibit the same kind of issue in another scenario. This is speculative, I didn't hit one, yet.

---Files--------------------------------
rb_io_wait_methods_with_scheduler_skip_errno_checks.patch (1.52 KB)


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

Unsubscribe: <mailto:[e-mail]?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>