net/dst_cache.c: preemption bug in net/dst_cache.c

net/dst_cache.c: preemption bug in net/dst_cache.c

[Bharath Vedartham]

Hi all,

I just want to bring attention to the syzbot bug [1]

Even though syzbot claims the bug to be in net/tipc, I feel it is in
net/dst_cache.c. Please correct me if I am wrong.

This bug is being triggered a lot of times by syzbot since the day it
was reported. Also given that this is core networking code, I felt it
was important to bring this to attention.

It looks like preemption needs to be disabled before using this_cpu_ptr
or maybe we would be better of using a get_cpu_var and put_cpu_var combo
here.

[1] https://syzkaller.appspot.com/bug?id=dc6352b92862eb79373fe03fdf9af5928753e057

Thank you
Bharath

Re: net/dst_cache.c: preemption bug in net/dst_cache.c

[Xin Long]

On Fri, Aug 23, 2019 at 3:58 PM Bharath Vedartham <linux.bhar@gmail.com> wrote:
>
> Hi all,
>
> I just want to bring attention to the syzbot bug [1]
>
> Even though syzbot claims the bug to be in net/tipc, I feel it is in
> net/dst_cache.c. Please correct me if I am wrong.
>
> This bug is being triggered a lot of times by syzbot since the day it
> was reported. Also given that this is core networking code, I felt it
> was important to bring this to attention.
>
> It looks like preemption needs to be disabled before using this_cpu_ptr
> or maybe we would be better of using a get_cpu_var and put_cpu_var combo
> here.
b->media->send_msg (tipc_udp_send_msg)
-> tipc_udp_xmit() -> dst_cache_get()

send_msg() is always called under the protection of rcu_read_lock(), which
already disabled preemption. If not, there must be some unbalanced calls of
disable/enable preemption elsewhere.

Agree that this could be a serious issue, do you have any reproducer for this?

Thanks.

>
> [1] https://syzkaller.appspot.com/bug?id=dc6352b92862eb79373fe03fdf9af5928753e057
>
> Thank you
> Bharath

Re: net/dst_cache.c: preemption bug in net/dst_cache.c

[Bharath Vedartham]

On Mon, Sep 09, 2019 at 05:48:25PM +0800, Xin Long wrote:
> On Fri, Aug 23, 2019 at 3:58 PM Bharath Vedartham <linux.bhar@gmail.com> wrote:
> >
> > Hi all,
> >
> > I just want to bring attention to the syzbot bug [1]
> >
> > Even though syzbot claims the bug to be in net/tipc, I feel it is in
> > net/dst_cache.c. Please correct me if I am wrong.
> >
> > This bug is being triggered a lot of times by syzbot since the day it
> > was reported. Also given that this is core networking code, I felt it
> > was important to bring this to attention.
> >
> > It looks like preemption needs to be disabled before using this_cpu_ptr
> > or maybe we would be better of using a get_cpu_var and put_cpu_var combo
> > here.
> b->media->send_msg (tipc_udp_send_msg)
> -> tipc_udp_xmit() -> dst_cache_get()
> 
> send_msg() is always called under the protection of rcu_read_lock(), which
> already disabled preemption. If not, there must be some unbalanced calls of
> disable/enable preemption elsewhere.
> 
> Agree that this could be a serious issue, do you have any reproducer for this?
> 
> Thanks.
Hi Xin,

Sorry for the delayed response. I do not have a reproducer for this. You
can submit a patch to syzbot which can run the patch on the same system
on which it found the bug.

Thank you
Bharath
> >
> > [1] https://syzkaller.appspot.com/bug?id=dc6352b92862eb79373fe03fdf9af5928753e057
> >
> > Thank you
> > Bharath