Feature #13362: [PATCH] socket: avoid fcntl for read/write_nonblock on Linux

• Status: Closed
• Assignee: -
• Target version: -

[normalperson (Eric Wong)]

On platforms where MSG_DONTWAIT works reliably on all sockets
(so far, I know of Linux), we can avoid fcntl syscalls and
implement IO#write_nonblock and IO#read_nonblock in terms of the
socket-specific send and recv family of syscalls.

This avoids side effects on the socket, and also encourages
generic code to be written in cases where IO wrappers like
OpenSSL::SSL::SSLSocket are used.

I could've sworn I sent something like this years ago, but
maybe I'm confusing it with kgio or socket_dontwait.

[normalperson (Eric Wong)]

[e-mail] wrote:

Feature #13362: [PATCH] socket: avoid fcntl for read/write_nonblock on Linux
https://bugs.ruby-lang.org/issues/13362

Any comment?

Also, this patch avoids the race condition where another thread
or process can clear the nonblocking flag via fcntl immediately
after {read,write}_nonblock sets it and causes the
{read,write}_nonblock caller to get stuck.

Thanks.

[akr (Akira Tanaka)]

• Status changed from Open to Feedback

I think it's possible on such platforms.

If some non-Ruby application depend on nonblocking flag set by Ruby,
such application will be affected, though.
It is better style that such application set nonblocking flag explicitly.

[normalperson (Eric Wong)]

[e-mail] wrote:

Issue #13362 has been updated by akr (Akira Tanaka).

Status changed from Open to Feedback

I think it's possible on such platforms.

If some non-Ruby application depend on nonblocking flag set by Ruby,
such application will be affected, though.

Yes, there may be some code which depends on side-effects.
However, I think it is a minor concern and anything which
inherits sockets ought to know to set flags correctly.

Normal Ruby code does not have this problem, since blocking
methods know to call rb_io_wait_*able on EAGAIN/EWOULDBLOCK.

It is better style that such application set nonblocking flag explicitly.

Right.

Should I commit?

[akr (Akira Tanaka)]

I think we can try it to see the incompatibility is really minor or not.
So, it should be described in NEWS.

• Status changed from Feedback to Closed

Applied in changeset trunk|r58400.

---

socket: avoid fcntl for read/write_nonblock on Linux

On platforms where MSG_DONTWAIT works reliably on all sockets
(so far, I know of Linux), we can avoid fcntl syscalls and
implement IO#write_nonblock and IO#read_nonblock in terms of the
socket-specific send and recv family of syscalls.

This avoids side effects on the socket, and also encourages
generic code to be written in cases where IO wrappers like
OpenSSL::SSL::SSLSocket are used.

Perhaps in the future, side-effect-free non-blocking I/O can
be standard on all files and OSes: https://cr.yp.to/unix/nonblock.html

• ext/socket/lib/socket.rb (read_nonblock, write_nonblock):
  Linux-specific wrapper without side effects
  [ruby-core:80780] [Feature #13362]
• test/socket/test_basicsocket.rb (test_read_write_nonblock):
  new test

[normalperson (Eric Wong)]

[e-mail] wrote:

I think we can try it to see the incompatibility is really minor or not.
So, it should be described in NEWS.

OK, r58400.

I think Ruby 1.8 => 1.9 introduced similar minor incompatibility
for all pipe and sockets, too, since 1.8 always used
non-blocking I/O to implement blocking IO with green threads.