On Mon, Nov 22, 2021 at 03:46:28PM -0300, Murilo Opsfelder Araújo wrote:
> Hi, Ilya Maximets.
>
> On 11/19/21 13:23, Ilya Maximets wrote:
> > On 11/18/21 22:19, David Christensen wrote:
> > >
> > >
> > > On 11/18/21 11:56 AM, Murilo Opsfelder Araújo wrote:
> > > > On 11/16/21 19:31, Ilya Maximets wrote:
> > > > > On 10/25/21 19:45, David Christensen wrote:
> > > > > > In certain high load situations, such as when creating a large
> > > > > > number of
> > > > > > ports on a switch, the parameter 'sock' may be passed to
> > > > > > nl_sock_recv__()
> > > > > > as null, resulting in a segmentation fault when 'sock' is later
> > > > > > dereferenced, such as when calling recvmsg().
> > > > >
> > > > > Hi, David. Thanks for the patch.
> > > > >
> > > > > It's OK to check for a NULL pointer there, I guess. However,
> > > > > do you know from where it was actually called? This function,
> > > > > in general, should not be called without the actual socket,
> > > > > so we, probably, should fix the caller instead.
> > > > >
> > > > > Best regards, Ilya Maximets.
> > > >
> > > > Hi, Ilya Maximets.
> > > >
> > > > When I looked at the coredump file, ch->sock was nil and was passed to
> > > > nl_sock_recv():
> > > >
> > > > (gdb) l
> > > > 2701
> > > > 2702 while (handler->event_offset < handler->n_events) {
> > > > 2703 int idx =
> > > > handler->epoll_events[handler->event_offset].data.u32;
> > > > 2704 struct dpif_channel *ch = &dpif->channels[idx];
> > > >
> > > > (gdb) p idx
> > > > $26 = 4
> > > > (gdb) p *dpif->channels@5
> > > > $27 = {{sock = 0x1001ae88240, last_poll = -9223372036854775808}, {sock
> > > > = 0x1001aa9a8a0, last_poll = -9223372036854775808}, {sock =
> > > > 0x1001ae09510, last_poll = 60634070}, {sock = 0x1001a9dbb60, last_poll
> > > > = 60756950}, {sock = 0x0,
> > > > last_poll = 61340749}}
> > > >
> > > >
> > > > The above snippet is from lib/dpif-netlink.c and the caller is
> > > > dpif_netlink_recv_vport_dispatch().
> > > >
> > > > The channel at idx=4 had sock=0x0, which was passed to nl_sock_recv()
> > > > via ch->sock parameter.
> > > > In that function, it tried to access sock->fd when calling recvmsg(),
> > > > causing the segfault.
> > > >
> > > > I'm not enough experienced in Open vSwitch to explain why sock was nil
> > > > at that given index.
> > > > The fix seems worth, though.
> > >
> > > A few other points of note:
> > >
> > > - Test system was a very large configuration (2K CPUs, > 1TB RAM)
> > > - OVS Switch was configured with 6K ports as follows:
> > >
> > > # b=br0; cmds=; for i in {1..6000}; do cmds+=" -- add-port $b p$i -- set
> > > interface p$i type=internal"; done
> > > # time sudo ovs-vsctl $cmds
> > >
> > > - OVS was installed from RHEL RPM. Build from source did not exhibit the
> > > same problem.
> > > - Unable to reproduce on a different system (128 CPUs, 256GB RAM), even
> > > with 10K ports.
> > >
> > > Dave
> >
> > Thanks for all the information. Having a NULL socket in the 'channels'
> > array doesn't look a good sign. This may mean that even if OVS will
> > not crash, it may not receive upcalls for certain ports, which is not
> > good eihter.
> >
> > Could you provide more info on which RHEL package did you use and which
> > version you built from source? Maybe we can find the difference.
>
> When the bug was reported internally, the distro package was actually from
> Fedora 33 ppc64le.
>
> One difference I've found at that time was regarding the configure options.
>
> From the logs at [0], the distro package was configured with these options:
>
> + ./configure --build=ppc64le-redhat-linux-gnu
> --host=ppc64le-redhat-linux-gnu --program-prefix=
> --disable-dependency-tracking --prefix=/usr --exec-prefix=/usr
> --bindir=/usr/bin --sbindir=/usr/sbin --sysconfdir=/etc --datadir=/usr/share
> --includedir=/usr/include --libdir=/usr/lib64 --libexecdir=/usr/libexec
> --localstatedir=/var --sharedstatedir=/var/lib --mandir=/usr/share/man
> --infodir=/usr/share/info --enable-libcapng --disable-static --enable-shared
> --enable-ssl --with-dpdk=shared --with-pkidir=/var/lib/openvswitch/pki
>
> And the built-from-source binaries from Dave were built like this:
>
> git clone https://github.com/openvswitch/ovs.git
> cd ovs/
> git checkout v2.15.0
Do you mean 2.16.0? Because the dpif_netlink_recv_vport_dispatch()
from the snippet above doesn't exist in 2.15.0.
And the FC33 uses 2.15.0:
https://koji.fedoraproject.org/koji/buildinfo?buildID=1713250
Can you reproduce the issue?
The reason I am asking is because in 2.16.0 we added another way to
handle upcalls, so I am wondering if the bug was introduced with
that change 2.16.0 or if it was already there in 2.15.0.
Also, could you provide the full backtrace of all threads?
(gdb) thread apply all bt
It would be nice to have a dump of the struct handler too.
Thanks,
fbl
> ./boot.sh
> ./configure
> make
> sudo make install
>
> At that time, I suspected that one (or a combination) of the configure
> options was contributing to the issue to manifest.
>
> [0]
> https://kojipkgs.fedoraproject.org//packages/openvswitch/2.15.0/1.fc33/data/logs/ppc64le/build.log
>
> >
> > Was there any suspicious warnings/errors in the OVS log file prior to
> > the failure? Maybe we run out of file descriptors somewhere due to
> > the machine size...
>
> None that I could tell. Only the following in the ovs-vswitchd.log:
>
> fatal_signal(revalidator138)|WARN|terminating with signal 11 (signal 11)
>
> >
> > Best regards, Ilya Maximets.
> >
>
> Cheers!
>
> --
> Murilo
> _______________________________________________
> dev mailing list
> d...@openvswitch.org
> https://mail.openvswitch.org/mailman/listinfo/ovs-dev
--
fbl
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
