Re: net: GPF in netlink_getsockbyportid

Re: net: GPF in netlink_getsockbyportid

[Florian Westphal]

Dmitry Vyukov <dvyukov@google.com> wrote:

[ CC nf-devel, not sure if its nfnetlink fault or NETLINK_MMAP ]

> The following program causes GPF in netlink_getsockbyportid:
> 
> // autogenerated by syzkaller (http://github.com/google/syzkaller)
> #include <pthread.h>
> #include <stdint.h>
> #include <string.h>
> #include <sys/syscall.h>
> #include <unistd.h>
> 
> int main()
> {
>   syscall(SYS_mmap, 0x20000000ul, 0xe65000ul, 0x3ul, 0x32ul,
>                  0xfffffffffffffffful, 0x0ul);
>   int fd = syscall(SYS_socket, 0x10ul, 0x803ul, 0xcul, 0, 0, 0);
>   *(uint32_t*)0x20e64000 = (uint32_t)0x28;
>   *(uint32_t*)0x20e64004 = (uint32_t)0x10;
>   *(uint64_t*)0x20e64008 = (uint64_t)0x0;
>   *(uint64_t*)0x20e64010 = (uint64_t)0x3;
>   *(uint64_t*)0x20e64018 = (uint64_t)0xfff;
>   *(uint16_t*)0x20e64020 = (uint16_t)0x5;
>   syscall(SYS_write, fd, 0x20e64000ul, 0x28ul, 0, 0, 0);
>   return 0;
> }

CONFIG_NETLINK_MMAP and nfnetlink batching strike in unison :-/

root cause is in nfnetlink_rcv_batch():

296 replay:
297         status = 0;
298 
299         skb = netlink_skb_clone(oskb, GFP_KERNEL);

The clone op doesn't copy oskb->sk, so we oops in
__netlink_alloc_skb -> netlink_getsockbyportid() when nfnetlink_rcv_batch
tries to send netlink ack.

Re: net: GPF in netlink_getsockbyportid

[Daniel Borkmann]

On 01/23/2016 08:25 PM, Florian Westphal wrote:
> Dmitry Vyukov <dvyukov@google.com> wrote:
>
> [ CC nf-devel, not sure if its nfnetlink fault or NETLINK_MMAP ]
>
>> The following program causes GPF in netlink_getsockbyportid:
>>
>> // autogenerated by syzkaller (http://github.com/google/syzkaller)
>> #include <pthread.h>
>> #include <stdint.h>
>> #include <string.h>
>> #include <sys/syscall.h>
>> #include <unistd.h>
>>
>> int main()
>> {
>>    syscall(SYS_mmap, 0x20000000ul, 0xe65000ul, 0x3ul, 0x32ul,
>>                   0xfffffffffffffffful, 0x0ul);
>>    int fd = syscall(SYS_socket, 0x10ul, 0x803ul, 0xcul, 0, 0, 0);
>>    *(uint32_t*)0x20e64000 = (uint32_t)0x28;
>>    *(uint32_t*)0x20e64004 = (uint32_t)0x10;
>>    *(uint64_t*)0x20e64008 = (uint64_t)0x0;
>>    *(uint64_t*)0x20e64010 = (uint64_t)0x3;
>>    *(uint64_t*)0x20e64018 = (uint64_t)0xfff;
>>    *(uint16_t*)0x20e64020 = (uint16_t)0x5;
>>    syscall(SYS_write, fd, 0x20e64000ul, 0x28ul, 0, 0, 0);
>>    return 0;
>> }
>
> CONFIG_NETLINK_MMAP and nfnetlink batching strike in unison :-/
>
> root cause is in nfnetlink_rcv_batch():
>
> 296 replay:
> 297         status = 0;
> 298
> 299         skb = netlink_skb_clone(oskb, GFP_KERNEL);
>
> The clone op doesn't copy oskb->sk, so we oops in
> __netlink_alloc_skb -> netlink_getsockbyportid() when nfnetlink_rcv_batch
> tries to send netlink ack.

If indeed oskb is the mmap'ed netlink skb, then it's not even allowed
to call into skb_clone() as it would access skb shared info data that
can be controlled by the user space mmap buffer, iirc, we had that in
the past with nlmon where skb_clone() was accidentally used.

Re: net: GPF in netlink_getsockbyportid

[Florian Westphal]

Daniel Borkmann <daniel@iogearbox.net> wrote:
> On 01/23/2016 08:25 PM, Florian Westphal wrote:
> >Dmitry Vyukov <dvyukov@google.com> wrote:
> >
> >[ CC nf-devel, not sure if its nfnetlink fault or NETLINK_MMAP ]
> >
> >>The following program causes GPF in netlink_getsockbyportid:
[..]

> >CONFIG_NETLINK_MMAP and nfnetlink batching strike in unison :-/
> >
> >root cause is in nfnetlink_rcv_batch():
> >
> >296 replay:
> >297         status = 0;
> >298
> >299         skb = netlink_skb_clone(oskb, GFP_KERNEL);
> >
> >The clone op doesn't copy oskb->sk, so we oops in
> >__netlink_alloc_skb -> netlink_getsockbyportid() when nfnetlink_rcv_batch
> >tries to send netlink ack.
> 
> If indeed oskb is the mmap'ed netlink skb, then it's not even allowed
> to call into skb_clone()

Right, but in this case there is no mmap'd netlink sk involved -- we
crash when we try to look up dst netlink socket to see if there is an
mmap'd ring attached.

[ and that code isn't there with CONFIG_NETLINK_MMAP=n ].

Re: net: GPF in netlink_getsockbyportid

[Herbert Xu]

On Sun, Jan 24, 2016 at 01:11:03AM +0100, Florian Westphal wrote:
> Daniel Borkmann <daniel@iogearbox.net> wrote:
> > On 01/23/2016 08:25 PM, Florian Westphal wrote:
> > >Dmitry Vyukov <dvyukov@google.com> wrote:
> > >
> > >[ CC nf-devel, not sure if its nfnetlink fault or NETLINK_MMAP ]
> > >
> > >>The following program causes GPF in netlink_getsockbyportid:
> [..]
> 
> > >CONFIG_NETLINK_MMAP and nfnetlink batching strike in unison :-/
> > >
> > >root cause is in nfnetlink_rcv_batch():
> > >
> > >296 replay:
> > >297         status = 0;
> > >298
> > >299         skb = netlink_skb_clone(oskb, GFP_KERNEL);
> > >
> > >The clone op doesn't copy oskb->sk, so we oops in
> > >__netlink_alloc_skb -> netlink_getsockbyportid() when nfnetlink_rcv_batch
> > >tries to send netlink ack.
> > 
> > If indeed oskb is the mmap'ed netlink skb, then it's not even allowed
> > to call into skb_clone()
> 
> Right, but in this case there is no mmap'd netlink sk involved -- we
> crash when we try to look up dst netlink socket to see if there is an
> mmap'd ring attached.
> 
> [ and that code isn't there with CONFIG_NETLINK_MMAP=n ].

Let's CC Pablo since he wrote the code in question.

Thanks,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: net: GPF in netlink_getsockbyportid

[Pablo Neira Ayuso]

On Mon, Jan 25, 2016 at 06:03:41PM +0800, Herbert Xu wrote:
> On Sun, Jan 24, 2016 at 01:11:03AM +0100, Florian Westphal wrote:
> > Daniel Borkmann <daniel@iogearbox.net> wrote:
> > > On 01/23/2016 08:25 PM, Florian Westphal wrote:
> > > >Dmitry Vyukov <dvyukov@google.com> wrote:
> > > >
> > > >[ CC nf-devel, not sure if its nfnetlink fault or NETLINK_MMAP ]
> > > >
> > > >>The following program causes GPF in netlink_getsockbyportid:
> > [..]
> > 
> > > >CONFIG_NETLINK_MMAP and nfnetlink batching strike in unison :-/
> > > >
> > > >root cause is in nfnetlink_rcv_batch():
> > > >
> > > >296 replay:
> > > >297         status = 0;
> > > >298
> > > >299         skb = netlink_skb_clone(oskb, GFP_KERNEL);
> > > >
> > > >The clone op doesn't copy oskb->sk, so we oops in
> > > >__netlink_alloc_skb -> netlink_getsockbyportid() when nfnetlink_rcv_batch
> > > >tries to send netlink ack.
> > > 
> > > If indeed oskb is the mmap'ed netlink skb, then it's not even allowed
> > > to call into skb_clone()
> > 
> > Right, but in this case there is no mmap'd netlink sk involved -- we
> > crash when we try to look up dst netlink socket to see if there is an
> > mmap'd ring attached.
> > 
> > [ and that code isn't there with CONFIG_NETLINK_MMAP=n ].
> 
> Let's CC Pablo since he wrote the code in question.

I have just sent this patch:

http://patchwork.ozlabs.org/patch/572489/