Bug in skb_gro_receive - possible bad page state problems?

Bug in skb_gro_receive - possible bad page state problems?

[Anand H. Krishnan]

Hello,

Is skb_gro_receive doing the right thing for cloned packets?

When we are merging fragments, we do not seem to be taking a reference
to the underlying page. To me, it looks like it should work fine for non-cloned
packets. However, for cloned packets, when the gro-ed packet is eventually
freed (because the original skb was not cloned and hence reference was 1),
the merged skb's frags also get freed (put_page-ed) without taking into account
the other references that were held for the fragments (dataref).

We saw crashes because of this behavior. Our setup had a third party kernel
forwarding module which uses GRO (napi_gro_receive). Doing iperf3 with small
packets and doing tcpdump on the receiving tap interface results in the problem.
With DEBUG_VM enabled, put page crashes. Without DEBUG_VM, bad page
state results.

Your thoughts (please CC me, since I am not part of this list).

Thanks,
Anand

Re: Bug in skb_gro_receive - possible bad page state problems?

[Eric Dumazet]

On Fri, 2017-05-05 at 08:57 +0530, Anand H. Krishnan wrote:
> Hello,
> 
> Is skb_gro_receive doing the right thing for cloned packets?
> 
> When we are merging fragments, we do not seem to be taking a reference
> to the underlying page. To me, it looks like it should work fine for non-cloned
> packets. However, for cloned packets, when the gro-ed packet is eventually
> freed (because the original skb was not cloned and hence reference was 1),
> the merged skb's frags also get freed (put_page-ed) without taking into account
> the other references that were held for the fragments (dataref).
> 
> We saw crashes because of this behavior. Our setup had a third party kernel
> forwarding module which uses GRO (napi_gro_receive). Doing iperf3 with small
> packets and doing tcpdump on the receiving tap interface results in the problem.
> With DEBUG_VM enabled, put page crashes. Without DEBUG_VM, bad page
> state results.

Yep, GRO must not be used with cloned skb.

This is why gro_cells_receive() has this check :

if (!gcells->cells || skb_cloned(skb) || netif_elide_gro(dev))
        return netif_rx(skb);

(But not the main napi_gro_receive() that is supposed to be used by
driver before any tap)