Re: [PATCH net v2] xsk: reject undersized tx_metadata at runtime validation

[Stanislav Fomichev <sdf.kernel@gmail.com>]

On 07/23, Alexander Lobakin wrote:
> From: Stanislav Fomichev <sdf.kernel@gmail.com>
> Date: Wed, 22 Jul 2026 10:55:17 -0700
> 
> > On 07/22, Cen Zhang (Microsoft) wrote:
> >> There is no metadata length check before accessing struct xsk_tx_metadata
> >> fields in both xdp_umem_reg() (registration) and
> >> xsk_buff_valid_tx_metadata() (runtime).  Thus, an invalid small
> >> tx_metadata_len (e.g. 8 bytes) passes registration and later causes an
> >> out-of-bounds read when xsk_skb_metadata() accesses fields beyond the
> >> declared metadata region.
> 
> [...]
> 
> > @@ -773,7 +773,7 @@ struct xdp_desc_ctx xp_raw_get_ctx(const struct xsk_buff_pool *pool, u64 addr)
> >  	addr = __xp_raw_get_addr(pool, addr);
> >  
> >  	ret.dma = __xp_raw_get_dma(pool, addr);
> > -	ret.meta = __xsk_buff_get_metadata(pool, __xp_raw_get_data(pool, addr));
> > +	xsk_buff_read_metadata(pool, addr, &ret.meta);
> 
> That effectively disables some micro-optimizations that were the reason
> I created this helper -- __xp_raw_get_data() is inline, while
> xsk_buff_read_metadata() uses the external version, probably some other
> stuff (note the underscored __xsk_buff_get_metadata()) =\ This function
> expanded to have no external calls on clang -O2, with this patch I'm not
> sure it would.

Any better ideas on how to solve it? I can convert all non libeth stuff
to these new xsk_buff_read_metadata helpers and keep xp_raw_get_ctx
to you? Will that work?