Message ID | 1458032624-139688-1-git-send-email-glider@google.com |
---|---|
State | Rejected, archived |
Delegated to: | David Miller |
Headers | show |
On Tue, 2016-03-15 at 10:03 +0100, Alexander Potapenko wrote: > According to IEEE Std 1003.1, 2013, sending data to a SOCK_SEQPACKET > socketpair with MSG_NOSIGNAL flag set must result in a SIGPIPE if the > socket is no longer connected. I find this sentence slightly confusing ? If MSG_NOSIGNAL is set, then SIGPIPE should not be generated. s/with/without/ maybe ?
From: Alexander Potapenko > Sent: 15 March 2016 09:04 > According to IEEE Std 1003.1, 2013, sending data to a SOCK_SEQPACKET > socketpair with MSG_NOSIGNAL flag set must result in a SIGPIPE if the > socket is no longer connected. ... > Without the below patch the behavior is as follows: > > $ ./sock seqpacket > sendmsg: Broken pipe ... > The behavior of the patched kernel complies with POSIX: > > $ ./sock seqpacket > Killed by SIGPIPE ... While POSIX might specify this behaviour, changing the behaviour could easily break applications. Basically this change (more or less) require every application that uses SOCK_SEQPACKED to be audited to ensure that MSG_NOSIGNAL is set on ever send/write to the socket. Personally I think the whole SIGPIPE on sockets should never have been allowed to get into the standard. I don't remember MSG_NOSIGNAL being present in SYSV. The only time you want a write into a pipe to generate SIGPIPE is for pipes generates by the shell that feed stdout to stdin of the next process in the pipeline. If pipes are implemented as unix-domain socketpairs (no one does that any more) then it would require the SIGPIPE for write(). David
On Tue, Mar 15, 2016 at 2:20 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote: > On Tue, 2016-03-15 at 10:03 +0100, Alexander Potapenko wrote: >> According to IEEE Std 1003.1, 2013, sending data to a SOCK_SEQPACKET >> socketpair with MSG_NOSIGNAL flag set must result in a SIGPIPE if the >> socket is no longer connected. > > I find this sentence slightly confusing ? > > If MSG_NOSIGNAL is set, then SIGPIPE should not be generated. > > s/with/without/ maybe ? > Agreed. I've rewritten this a couple of times and failed to proof-read it for the last time.
On Tue, Mar 15, 2016 at 2:46 PM, David Laight <David.Laight@aculab.com> wrote: > From: Alexander Potapenko >> Sent: 15 March 2016 09:04 >> According to IEEE Std 1003.1, 2013, sending data to a SOCK_SEQPACKET >> socketpair with MSG_NOSIGNAL flag set must result in a SIGPIPE if the >> socket is no longer connected. > ... >> Without the below patch the behavior is as follows: >> >> $ ./sock seqpacket >> sendmsg: Broken pipe > ... >> The behavior of the patched kernel complies with POSIX: >> >> $ ./sock seqpacket >> Killed by SIGPIPE > ... > > While POSIX might specify this behaviour, changing the behaviour > could easily break applications. > Basically this change (more or less) require every application that > uses SOCK_SEQPACKED to be audited to ensure that MSG_NOSIGNAL is set > on ever send/write to the socket. This is true, but the drawback of maintaining a non-standard behavior is that people can't write portable code. I couldn't find the exact place where the bug has been introduced, but according to http://lxr.free-electrons.com/ SOCK_SEQPACKET sockets did not respect MSG_NOSIGNAL from the very beginning (http://lxr.free-electrons.com/source/net/unix/af_unix.c?v=3.8#L1723, there was no such thing as SOCK_SEQPACKET in AF_UNIX in 2.4.37) Unfortunately I don't know how to estimate the number of existing users depending on this oddity, as well as the number of people that have to work around it, so leaving it up to maintainers to decide whether the fix is needed. > Personally I think the whole SIGPIPE on sockets should never have been > allowed to get into the standard. > I don't remember MSG_NOSIGNAL being present in SYSV. I think it was not. > The only time you want a write into a pipe to generate SIGPIPE is > for pipes generates by the shell that feed stdout to stdin of the > next process in the pipeline. > If pipes are implemented as unix-domain socketpairs (no one does that > any more) then it would require the SIGPIPE for write(). > > David > >
From: Alexander Potapenko <glider@google.com> Date: Tue, 15 Mar 2016 10:03:44 +0100 > According to IEEE Std 1003.1, 2013, sending data to a SOCK_SEQPACKET > socketpair with MSG_NOSIGNAL flag set must result in a SIGPIPE if the > socket is no longer connected. > > I used the following program to check the kernel behavior: ... > Signed-off-by: Alexander Potapenko <glider@google.com> In the final analysis I don't think we can do this. Properly working programs will start to mysteriously receive signals in this situation and fail. I'm sorry but I think we're stuck with this behavior.
diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 8fbe6d7..ba34c73 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -1645,6 +1645,7 @@ static int unix_dgram_sendmsg(struct socket *sock, struct msghdr *msg, struct sock *other = NULL; int namelen = 0; /* fake GCC */ int err; + bool send_sigpipe = false; unsigned int hash; struct sk_buff *skb; long timeo; @@ -1675,6 +1676,12 @@ static int unix_dgram_sendmsg(struct socket *sock, struct msghdr *msg, goto out; } + if (sk->sk_shutdown & SEND_SHUTDOWN && sock->type == SOCK_SEQPACKET) { + send_sigpipe = true; + err = -EPIPE; + goto out; + } + if (test_bit(SOCK_PASSCRED, &sock->flags) && !u->addr && (err = unix_autobind(sock)) != 0) goto out; @@ -1769,8 +1776,11 @@ restart_locked: } err = -EPIPE; - if (other->sk_shutdown & RCV_SHUTDOWN) + if (other->sk_shutdown & RCV_SHUTDOWN) { + if (sock->type == SOCK_SEQPACKET) + send_sigpipe = true; goto out_unlock; + } if (sk->sk_type != SOCK_SEQPACKET) { err = security_unix_may_send(sk->sk_socket, other->sk_socket); @@ -1837,6 +1847,8 @@ out: if (other) sock_put(other); scm_destroy(&scm); + if (send_sigpipe && !(msg->msg_flags & MSG_NOSIGNAL)) + send_sig(SIGPIPE, current, 0); return err; }
According to IEEE Std 1003.1, 2013, sending data to a SOCK_SEQPACKET socketpair with MSG_NOSIGNAL flag set must result in a SIGPIPE if the socket is no longer connected. I used the following program to check the kernel behavior: /*****************/ #include <signal.h> #include <stdio.h> #include <string.h> #include <sys/socket.h> #include <unistd.h> void PipeHandler(int sig) { fprintf(stderr, "Killed by SIGPIPE\n"); _exit(1); } int main(int argc, char *argv[]) { if (argc < 2) return 0; struct sigaction act; act.sa_handler = PipeHandler; sigaction(SIGPIPE, &act, NULL); int fds[2]; if (strcmp(argv[1], "seqpacket") == 0) { if (socketpair(AF_UNIX, SOCK_SEQPACKET, 0, fds) == -1) return -1; } else { if (strcmp(argv[1], "stream") == 0) { if (socketpair(AF_UNIX, SOCK_STREAM, 0, fds) == -1) return -1; } else { return -1; } } int flags = 0; if ((argc > 2) && (strcmp(argv[2], "nosignal") == 0)) flags = MSG_NOSIGNAL; struct msghdr msg; memset(&msg, 0, sizeof(msg)); close(fds[0]); const ssize_t r = sendmsg(fds[1], &msg, flags); if (r == -1) perror("sendmsg"); return 0; } /*****************/ Without the below patch the behavior is as follows: $ ./sock seqpacket sendmsg: Broken pipe $ ./sock stream Killed by SIGPIPE $ ./sock seqpacket nosignal sendmsg: Broken pipe $ ./sock stream nosignal sendmsg: Broken pipe The behavior of the patched kernel complies with POSIX: $ ./sock seqpacket Killed by SIGPIPE $ ./sock stream Killed by SIGPIPE $ ./sock seqpacket nosignal sendmsg: Broken pipe $ ./sock stream nosignal sendmsg: Broken pipe Signed-off-by: Alexander Potapenko <glider@google.com> --- net/unix/af_unix.c | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-)