Feature #11145: [PATCH] socket: avoid redundant fcntl with MSG_DONTWAIT

• Status: Closed
• Assignee: akr (Akira Tanaka)
• Target version: -

[normalperson (Eric Wong)]

MSG_DONTWAIT is enough to force non-blocking I/O under Linux,
so avoid changing the state of a socket. This will allow certain
threads to do a non-destructive non-blocking "peek" while others
block (without relying on an extra ppoll syscall).

I suspect any OS with the MSG_DONTWAIT flag implements it correctly. AFAIK,
MSG_DONTWAIT is only on FreeBSD and Linux, and I've only tested on Linux.

• ext/socket/init.c (rsock_s_recvfrom_nonblock): avoid redundant fcntl

[normalperson (Eric Wong)]

• File 0001-socket-avoid-redundant-fcntl-on-Linux.patch 0001-socket-avoid-redundant-fcntl-on-Linux.patch added

Better patch to support sending, too. I'm limiting this to Linux-only since
it seems newer FreeBSD has support for MSG_DONTWAIT, but I'm not sure it is
complete.

I'll commit in a few days

• Status changed from Open to Closed

Applied in changeset r50666.

---

socket: avoid redundant fcntl on Linux

• ext/socket/ancdata.c (bsock_sendmsg_internal,
  bsock_recvmsg_internal):
  avoid redundant fcntl on Linux
  [ruby-core:69154] [Feature #11145]
• ext/socket/init.c (rsock_s_recvfrom_nonblock): ditto
• ext/socket/rubysocket.h (MSG_DONTWAIT_RELIABLE): new macro

MSG_DONTWAIT is enough to force non-blocking I/O under Linux,
so avoid changing the state of a socket. This will allow certain
threads to do a non-destructive non-blocking "peek" while others
block (without relying on an extra ppoll syscall).

We shall be conservative about enabling this feature since some
OSes may have incomplete support for MSG_DONTWAIT. I shall
defer to a FreeBSD expert to enable that for FreeBSD.