net/xdp: Potential issue with xp_aligned_validate_desc() in multi-buffer packets

net/xdp: Potential issue with xp_aligned_validate_desc() in multi-buffer packets

[Rao, Nikhil]

Hi XDP maintainers,

I have noticed a potential issue with multi-buffer packet support in the
xp_aligned_validate_desc() function in net/xdp/xsk_queue.h.

In xp_aligned_validate_desc(), the validation always adjusts the address
and length by pool->tx_metadata_len:

   u64 addr = desc->addr - pool->tx_metadata_len;
   u64 len = desc->len + pool->tx_metadata_len;

However, in multi-buffer XDP packets, only the first descriptor contains
TX metadata. Subsequent descriptors do not have
metadata and should not be validated with this adjustment.

This causes validation failures for the 2nd buffer onward in
multi-buffer packets when TX metadata is enabled.

The same issue exists in xp_unaligned_validate_desc() as well.

I'm happy to submit a patch if this is indeed a bug. Please let me know
your thoughts.

Thanks,
Nikhil

Re: net/xdp: Potential issue with xp_aligned_validate_desc() in multi-buffer packets

[Stanislav Fomichev]

On 02/03, Rao, Nikhil wrote:
> Hi XDP maintainers,
> 
> I have noticed a potential issue with multi-buffer packet support in the
> xp_aligned_validate_desc() function in net/xdp/xsk_queue.h.
> 
> In xp_aligned_validate_desc(), the validation always adjusts the address
> and length by pool->tx_metadata_len:
> 
>   u64 addr = desc->addr - pool->tx_metadata_len;
>   u64 len = desc->len + pool->tx_metadata_len;
> 
> However, in multi-buffer XDP packets, only the first descriptor contains
> TX metadata. Subsequent descriptors do not have
> metadata and should not be validated with this adjustment.
> 
> This causes validation failures for the 2nd buffer onward in
> multi-buffer packets when TX metadata is enabled.
> 
> The same issue exists in xp_unaligned_validate_desc() as well.
> 
> I'm happy to submit a patch if this is indeed a bug. Please let me know
> your thoughts.

I think your assessment is correct. Can you try to extend
prog_tests/xdp_metadata.c to send a mbuf packet? I don't think we have
any real test coverage for the metadata+mbuf..