From: akr@...
Date: 2014-01-18T22:53:01+00:00
Subject: [ruby-core:59857] [ruby-trunk - Feature #9330] [PATCH 0/3] avoid redundant fcntl/fstat syscalls for cloexec sockets
Issue #9330 has been updated by Akira Tanaka.
The patch 1/3 and 3/3 are also have difficulty with rb_update_max_fd.
For example, rsock_socket0 returns fd without calling rb_update_max_fd.
This style makes us difficult to confirm that rb_update_max_fd is called for all fd allocation.
----------------------------------------
Feature #9330: [PATCH 0/3] avoid redundant fcntl/fstat syscalls for cloexec sockets
https://bugs.ruby-lang.org/issues/9330#change-44423
* Author: Eric Wong
* Status: Open
* Priority: Low
* Assignee:
* Category: ext
* Target version: next minor
----------------------------------------
commit 52525b673669019dc3c03474e613937fd3587187
Author: Eric Wong <e@80x24.org>
Date: Tue Dec 31 02:09:27 2013 +0000
socket: avoid redundant fcntl calls for FD passing
If MSG_CMSG_CLOEXEC works on the first recvmsg syscall, we should
assume it always works and avoid checking it with every single FD
received.
While we're at it, it is pointless to set FD_CLOEXEC immediately
before we close it in discard_cmsg; so only update the max FD (the
FD is likely to be recycled, so we should update even if we close
it).
Note: I think it's safe to assume any system with MSG_CMSG_CLOEXEC
will have implemented SOCK_CLOEXEC; too, so we reuse use
rsock_detect_cloexec instead of defining a new function wrapped
by #ifdef MSG_CMSG_CLOEXEC.
commit aa9ed17a5f10cf53f72dc5ab1f4ff8d8b5e4dd84
Author: Eric Wong <e@80x24.org>
Date: Tue Dec 31 01:36:27 2013 +0000
socket: avoid redundant calls to rb_update_max_fd
rsock_init_sock already calls rb_update_max_fd, so do not
waste time making redundant fstat() and atomic CAS instructions.
commit b1694d7bac55bd9f537426624fe599c46b10dee6
Author: Eric Wong <e@80x24.org>
Date: Mon Dec 30 23:58:32 2013 +0000
socket: avoid redundant fcntl(fd, F_GETFD) with SOCK_CLOEXEC
If socket or socketpair succeeds with SOCK_CLOEXEC once, we should
expect it to succeed for all subsequent invocations of each of those
syscalls. This reduces the number of syscalls made and makes strace
output cleaner.
While we're at it, swap some remaining rb_fd_fix_cloexec calls for
rb_maygvl_fd_fix_cloexec, as the former calls rb_update_max_fd
redundantly. A future commit will reduce rb_update_max_fd calls
further.
The following changes since commit d6c7ab9ffeff31fb5a49776ab74a61b478344107:
* 2013-12-31 (2013-12-30 19:48:05 +0000)
are available in the git repository at:
git://bogomips.org/ruby.git socket_cloexec_once
for you to fetch changes up to 52525b673669019dc3c03474e613937fd3587187:
socket: avoid redundant fcntl calls for FD passing (2013-12-31 02:17:53 +0000)
----------------------------------------------------------------
Eric Wong (3):
socket: avoid redundant fcntl(fd, F_GETFD) with SOCK_CLOEXEC
socket: avoid redundant calls to rb_update_max_fd
socket: avoid redundant fcntl calls for FD passing
ext/socket/ancdata.c | 9 +++++++--
ext/socket/init.c | 49 ++++++++++++++++++++++++++++++++++++++-----------
ext/socket/rubysocket.h | 2 ++
ext/socket/socket.c | 31 ++++++++++++++++++++++---------
ext/socket/unixsocket.c | 10 ++++++++--
5 files changed, 77 insertions(+), 24 deletions(-)
---Files--------------------------------
0001-socket-avoid-redundant-fcntl-fd-F_GETFD-with-SOCK_CL.patch (5.87 KB)
0002-socket-avoid-redundant-calls-to-rb_update_max_fd.patch (1.35 KB)
0003-socket-avoid-redundant-fcntl-calls-for-FD-passing.patch (3.27 KB)
--
http://bugs.ruby-lang.org/