Re: non-blocking connect for kernel SCTP sockets

[Michal Kubecek <mkubecek@suse.cz>]

On Wed, May 02, 2018 at 05:46:23PM +0800, Xin Long wrote:
> On Wed, May 2, 2018 at 5:06 PM, Michal Kubecek <mkubecek@suse.cz> wrote:
> > Hello,
> >
> > while investigating a bug, we noticed that DLM tries to connect an SCTP
> > socket in non-blocking mode using
> >
> >         result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len,
> >                                     O_NONBLOCK);
> >
> > which does not work. The reason is that inet_dgram_connect() cannot pass
> > its flags argument to sctp_connect() so that __sctp_connect() which does
> > the actual waiting resorts to checking sk->sk_socket->file->f_flags
> > instead. As the socket used by DLM is a kernel socket with no associated
> > file, it ends up blocking.
> >
> > TCP doesn't suffer from this problem as for TCP, the waiting is done in
> > inet_stream_connect() which has the flags argument. I also checked other
> > proto::connect handlers and sctp_connect() seems to be the only one with
> > this kind of problem.
> >
> > This could be worked around in DLM and further experiments indicate
> > current DLM code wouldn't actually handle the non-blocking connect
> > properly. But I still feel ignoring the flags argument is rather a trap
> > that should be fixed.
> It is a bug, https://bugzilla.redhat.com/show_bug.cgi?id=1251530

Not authorized. :-)

> We have the fix which also includes some cleanup, and needs to do
> more testing.

OK, I'll wait for your submission.

> > I have prepared a series adding flags argument to proto::connect and
> > using it in sctp_connect() and __sctp_connect(). But I'm not sure if
> > it's not too big hammer to address issue only affecting one handler.
> > So my question is: would such generic approach be preferred or should we,
> > rather make SCTP work the way TCP does, i.e. move the waiting from,
> > proto::connect() to proto_ops::connect()? This would require introducing
> > inet_seqpacket_connect() as inet_dgram_connect() is primarily intended
> > for use with UDP.)
> We don't fix it in the generic proto::connect, which will afftect
> many other places.

That was my concern, too. On the other hand, the TCP specific waiting
code in inet_stream_connect() makes me wonder if it wouldn't be cleaner
to move it into the TCP specific handler as well (which is something
this approach would allow).

> We're replacing only sctp's proto_ops::connect with sctp_connect and
> leave its proto::connect as NULL, so that it can get this flags param
> without touching the generic struct and code.

Yes, that should do the trick (and makes backporting to distribution
kernels with frozen kABI much easier). I guess I was too fixed on the
split between proto_ops::connect and proto::connect to see this
solution.

Michal Kubecek