Thread (6 messages) flat view 6 messages, 4 authors, 2016-03-18

Re: [PATCH] af_unix: closed SOCK_SEQPACKET socketpair must get SIGPIPE

From: Alexander Potapenko <glider@google.com>
Date: 2016-03-15 16:13:12
Also in: lkml

On Tue, Mar 15, 2016 at 2:46 PM, David Laight [off-list ref] wrote:
From: Alexander Potapenko
quoted
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.
...
quoted
Without the below patch the behavior is as follows:

$ ./sock seqpacket
sendmsg: Broken pipe
...
quoted
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


-- 
Alexander Potapenko
Software Engineer

Google Germany GmbH
Erika-Mann-Straße, 33
80636 München

Geschäftsführer: Matthew Scott Sucherman, Paul Terence Manicle
Registergericht und -nummer: Hamburg, HRB 86891
Sitz der Gesellschaft: Hamburg
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help