[ovs-dev] [PATCH] netlink-socket: Check for null sock in nl_sock_recv__()

[Ilya Maximets]

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 at 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.

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...

Best regards, Ilya Maximets.