Hello,
On Mon, Oct 12, 2015, at 21:41, Jason Baron wrote:
> On 10/09/2015 10:38 AM, Hannes Frederic Sowa wrote:
> > Hi,
> >
> > Jason Baron <jba...@akamai.com> writes:
> >
> >> The unix_dgram_poll() routine calls sock_poll_wait() not only for the wait
> >> queue associated with the socket s that we are poll'ing against, but also
> >> calls
> >> sock_poll_wait() for a remote peer socket p, if it is connected. Thus,
> >> if we call poll()/select()/epoll() for the socket s, there are then
> >> a couple of code paths in which the remote peer socket p and its associated
> >> peer_wait queue can be freed before poll()/select()/epoll() have a chance
> >> to remove themselves from the remote peer socket.
> >>
> >> The way that remote peer socket can be freed are:
> >>
> >> 1. If s calls connect() to a connect to a new socket other than p, it will
> >> drop its reference on p, and thus a close() on p will free it.
> >>
> >> 2. If we call close on p(), then a subsequent sendmsg() from s, will drop
> >> the final reference to p, allowing it to be freed.
> >>
> >> Address this issue, by reverting unix_dgram_poll() to only register with
> >> the wait queue associated with s and register a callback with the remote
> >> peer
> >> socket on connect() that will wake up the wait queue associated with s. If
> >> scenarios 1 or 2 occur above we then simply remove the callback from the
> >> remote peer. This then presents the expected semantics to poll()/select()/
> >> epoll().
> >>
> >> I've implemented this for sock-type, SOCK_RAW, SOCK_DGRAM, and
> >> SOCK_SEQPACKET
> >> but not for SOCK_STREAM, since SOCK_STREAM does not use unix_dgram_poll().
> >>
> >> Introduced in commit ec0d215f9420 ("af_unix: fix 'poll for write'/connected
> >> DGRAM sockets").
> >>
> >> Tested-by: Mathias Krause <mini...@googlemail.com>
> >> Signed-off-by: Jason Baron <jba...@akamai.com>
> >
> > While I think this approach works, I haven't seen where the current code
> > leaks a reference. Assignment to unix_peer(sk) in general take spin_lock
> > and increment refcount. Are there bugs at the two places you referred
> > to?
> >
> > Is an easier fix just to use atomic_inc_not_zero(&sk->sk_refcnt) in
> > unix_peer_get() which could also help other places?
> >
>
> Hi,
>
> So we could potentially inc the refcnt on the remote peer such that the
> remote peer does not free before the socket that has connected to it.
> However, then the socket that has taken the reference against the peer
> socket has to potentially record a number of remote sockets (all the ones
> that it has connected to over its lifetime), and then drop all of their
> refcnt's when it finally closes.
>
> The reason for this is that with the current code when we do
> poll()/select()/epoll() on a socket with a peer socket, those calls
> take reference on the peer socket. Specifically, they record the remote
> peer whead, such that they can remove their callbacks when they return.
> So its not safe to just drop a reference on the remote peer when it
> closes because their might be outstanding poll()/select()/epoll()
> references pending.
Thanks for the explanation, it was very helpful. The eventpoll
infrastructure seems not to be easily able to handle these kind of
socket cross references easily, I understand.
> Normally, poll()/select()/epoll() are waiting on a whead associated
> directly with the fd/file that they are waiting for.
Exactly. The reference count is implicit by the current process to
handle the filedescriptor and deregister the wait heads during program
tear-down or close. So a sock_poll_wait call to a foreign socket's wait
queue will confuse this subsystem.
> The other point here is that the way this patch structures things is
> that when the socket connects to a new remote and hence disconnects from
> an existing remote, POLLOUT events will continue to be correctly
> delivered. That was not possible with the current structure of things
> b/c there was no way to inform poll to re-register with the remote peer
> whead. So, that means that the first test case here now works:
>
> https://lkml.org/lkml/2015/10/4/154
>
> Whereas with the old code test case would just hang for ever.
>
> So yes there is a bit of code churn here, but I think it moves the
> code-base in a direction that not only solves this issue, but corrects
> additional poll() behaviors as well.
Agreed, the new semantics make sense to me and are an improvement.
Thanks,
Hannes
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
