Re: [PATCH v3] drivers: net: prevent tun_get_user() to exceed xdp size limits

[David Ahern <dsahern@gmail.com>]

On 7/26/23 1:37 PM, David Ahern wrote:
> On 7/26/23 3:02 AM, Jesper Dangaard Brouer wrote:
>> Cc. John and Ahern
>>
>> On 26/07/2023 04.09, Jason Wang wrote:
>>> On Tue, Jul 25, 2023 at 11:54 PM Andrew Kanner
>>> <andrew.kanner@gmail.com> wrote:
>>>>
>>>> Syzkaller reported the following issue:
>>>> =======================================
>>>> Too BIG xdp->frame_sz = 131072
>>
>> Is this a contiguous physical memory allocation?
>>
>> 131072 bytes equal order 5 page.
>>
>> Looking at tun.c code I cannot find a code path that could create
>> order-5 skb->data, but only SKB with order-0 fragments.  But I guess it
>> is the netif_receive_generic_xdp() what will realloc to make this linear
>> (via skb_linearize())
> 
> 
> get_tun_user is passed an iov_iter with a single segment of 65007
> total_len. The alloc_skb path is hit with an align size of only 64. That
> is insufficient for XDP so the netif_receive_generic_xdp hits the
> pskb_expand_head path. Something is off in the math in
> netif_receive_generic_xdp resulting in the skb markers being off. That
> causes bpf_prog_run_generic_xdp to compute the wrong frame_sz.


BTW, it is pskb_expand_head that turns it from a 64kB to a 128 kB
allocation. But the 128kB part is not relevant to the "bug" here really.

The warn on getting tripped in bpf_xdp_adjust_tail is because xdp
generic path is skb based and can have a frame_sz > 4kB. That's what the
splat is about.

Perhaps the solution is to remove the WARN_ON.