Re: tc: RTM_GETQDISC causes kernel OOPS

[Patrick McHardy <kaber@trash.net>]

Eric Dumazet wrote:
> Le vendredi 21 mai 2010 à 15:42 -0700, Ben Pfaff a écrit :
>> Hi.  While working on some library code for working with qdiscs and
>> classes I came upon a kernel OOPS.  Originally I came across it with a
>> 2.6.26 kernel, but I can also reproduce it with unmodified v2.6.34 from
>> kernel.org.
>>
>> At the end of this mail I'm appending both an example of the OOPS and a
>> simple test program that reliably reproduces the problem for me when I
>> invoke it with "lo" as argument.  The program does not need to be run as
>> root.
>>
>> After the OOPS, a lot of networking and other system functions stop
>> working, so it seems to me a serious issue.
>>
>> The null pointer dereference that causes the OOPS is the dereference of
>> the return value of qdisc_dev() in tc_fill_qdisc() in
>> net/sched/sch_api.c line 1163:
>>
>>     1161	tcm->tcm__pad1 = 0;
>>     1162	tcm->tcm__pad2 = 0;
>>     1163	tcm->tcm_ifindex = qdisc_dev(q)->ifindex;
>>     1164	tcm->tcm_parent = clid;
>>     1165	tcm->tcm_handle = q->handle;
>>
>> I am pretty sure about that, because if I add "WARN_ON(!qdisc_dev(q));"
>> just before line 1163 then that warning triggers.
>>
>> Thanks,
> 
> Indeed, thanks for this very useful report !
> 
> We could add a check for TCQ_F_BUILTIN flag, or just make 
> qdisc_notify() checks consistent for both old and new qdisc
> 
> What other people thinks ?

We already use TCQ_F_BUILTIN in tc_qdisc_dump_ignore(), so I
think it would be more consistent than checking for a handle
to use it here as well.