[PATCH] socket: avoid redundant fcntl on Linux

[PATCH] socket: avoid redundant fcntl on Linux

[Eric Wong]

* 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.
---
 ext/socket/ancdata.c    |  4 ++--
 ext/socket/init.c       |  5 ++++-
 ext/socket/rubysocket.h | 11 +++++++++++
 3 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/ext/socket/ancdata.c b/ext/socket/ancdata.c
index 3eaed1c..0a94149 100644
--- a/ext/socket/ancdata.c
+++ b/ext/socket/ancdata.c
@@ -1277,7 +1277,7 @@ bsock_sendmsg_internal(int argc, VALUE *argv, VALUE sock, int nonblock)
 #endif
 
     rb_io_check_closed(fptr);
-    if (nonblock)
+    if (nonblock && !MSG_DONTWAIT_RELIABLE)
         rb_io_set_nonblock(fptr);
 
     ss = rb_sendmsg(fptr->fd, &mh, flags);
@@ -1595,7 +1595,7 @@ bsock_recvmsg_internal(int argc, VALUE *argv, VALUE sock, int nonblock)
         flags |= MSG_PEEK;
 
     rb_io_check_closed(fptr);
-    if (nonblock)
+    if (nonblock && !MSG_DONTWAIT_RELIABLE)
         rb_io_set_nonblock(fptr);
 
     ss = rb_recvmsg(fptr->fd, &mh, flags);
diff --git a/ext/socket/init.c b/ext/socket/init.c
index 317dd67..455652d 100644
--- a/ext/socket/init.c
+++ b/ext/socket/init.c
@@ -211,7 +211,10 @@ rsock_s_recvfrom_nonblock(VALUE sock, int argc, VALUE *argv, enum sock_recv_type
     str = rb_tainted_str_new(0, buflen);
 
     rb_io_check_closed(fptr);
-    rb_io_set_nonblock(fptr);
+
+    if (!MSG_DONTWAIT_RELIABLE)
+	rb_io_set_nonblock(fptr);
+
     len0 = alen;
     slen = recvfrom(fd, RSTRING_PTR(str), buflen, flags, &buf.addr, &alen);
     if (slen != -1 && len0 < alen)
diff --git a/ext/socket/rubysocket.h b/ext/socket/rubysocket.h
index 689a7a8..359d28e 100644
--- a/ext/socket/rubysocket.h
+++ b/ext/socket/rubysocket.h
@@ -407,9 +407,20 @@ NORETURN(void rsock_sys_fail_raddrinfo_or_sockaddr(const char *, VALUE addr, VAL
 #if defined(__linux__)
 static inline int rsock_maybe_fd_writable(int fd) { return 1; }
 static inline void rsock_maybe_wait_fd(int fd) { }
+#  ifdef MSG_DONTWAIT
+#    define MSG_DONTWAIT_RELIABLE 1
+#  endif
 #else /* some systems (mswin/mingw) need these.  ref: r36946 */
 #  define rsock_maybe_fd_writable(fd) rb_thread_fd_writable((fd))
 #  define rsock_maybe_wait_fd(fd) rb_thread_wait_fd((fd))
 #endif
 
+/*
+ * some OSes may support MSG_DONTWAIT inconsistently depending on socket
+ * type, we only expect Linux to support it consistently for all socket types.
+ */
+#ifndef MSG_DONTWAIT_RELIABLE
+#  define MSG_DONTWAIT_RELIABLE 0
+#endif
+
 #endif
-- 
EW