On Wed, 2017-02-08 at 19:55 +0100, Dmitry Vyukov wrote: > On Wed, Feb 8, 2017 at 6:58 PM, Eric Dumazet <eric.duma...@gmail.com> wrote: > > On Wed, 2017-02-08 at 18:36 +0100, Dmitry Vyukov wrote: > >> On Tue, Jan 24, 2017 at 4:52 PM, Eric Dumazet <eduma...@google.com> wrote: > >> > On Tue, Jan 24, 2017 at 7:06 AM, Dmitry Vyukov <dvyu...@google.com> > >> > wrote: > >> >>> > >> >>> This code was changed a long time ago : > >> >>> > >> >>> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=ed2e923945892a8372ab70d2f61d364b0b6d9054 > >> >>> > >> >>> So I suspect a recent patch broke the logic. > >> >>> > >> >>> You might start a bisection : > >> >>> > >> >>> I would check if 4.7 and 4.8 trigger the issue you noticed. > >> >> > >> >> > >> >> It happens with too low rate for bisecting (few times per day). I > >> >> could add some additional checks into code, but I don't know what > >> >> checks could be useful. > >> > > >> > If you can not tell if 4.7 and/or 4.8 have the problem, I am not sure > >> > we are able to help. > >> > >> > >> There are also chances that the problem is older. > >> > >> Looking at the code, this part of inet_twsk_purge looks fishy: > >> > >> 285 if (unlikely((tw->tw_family != family) || > >> 286 > >> atomic_read(&twsk_net(tw)->count))) { > >> > >> It uses net->count == 0 check to find the right sockets. But what if > >> there are several nets with count == 0 in flight, can't there be > >> several inet_twsk_purge calls running concurrently freeing each other > >> sockets? If so it looks like inet_twsk_purge can call > >> inet_twsk_deschedule_put twice for a socket. Namely, two calls for > >> different nets discover the socket, check that net->count==0 and both > >> call inet_twsk_deschedule_put. Shouldn't we just give inet_twsk_purge > >> net that it needs to purge? > > > > Yes, atomic_read() is not a proper sync point. > > Do you mean that it does not include read barrier? > I more mean that we can call inet_twsk_deschedule_put twice for the same > socket.
I meant that this code assumed RTNL being held. This might not be the case now, after some old change.