From: Alexander Lobakin <aleksander.lobakin@xxxxxxxxx>
Date: Fri, 1 Aug 2025 15:12:43 +0200
> From: Kees Cook <kees@xxxxxxxxxx>
> Date: Thu, 31 Jul 2025 10:05:47 -0700
>
>> On Thu, Jul 31, 2025 at 01:37:34PM +0100, Simon Horman wrote:
>>> While I appreciate the desire for improved performance and nicer code
>>> generation. I think the idea of writing 64 bits of data to the
>>> address of a 32 bit member of a structure goes against the direction
>>> of hardening work by Kees and others.
>>
>> Agreed: it's better to avoid obscuring these details from the compiler
>> so it can have an "actual" view of the object sizes involved.
>>
>>> Indeed, it seems to me this is the kind of thing that struct_group()
>>> aims to avoid.
>>>
>>> In this case struct group() doesn't seem like the best option,
>>> because it would provide a 64-bit buffer that we can memcpy into.
>>> But it seems altogether better to simply assign u64 value to a u64 member.
>>
>> Agreed: with struct_group you get a sized pointer, and while you can
>> provide a struct tag to make it an assignable object, it doesn't make
>> too much sense here.
>>
>>> So I'm wondering if an approach along the following lines is appropriate
>>> (Very lightly compile tested only!).
>>>
>>> And yes, there is room for improvement of the wording of the comment
>>> I included below.
>>>
>>> diff --git a/include/net/libeth/xdp.h b/include/net/libeth/xdp.h
>>> index f4880b50e804..a7d3d8e44aa6 100644
>>> --- a/include/net/libeth/xdp.h
>>> +++ b/include/net/libeth/xdp.h
>>> @@ -1283,11 +1283,7 @@ static inline void libeth_xdp_prepare_buff(struct libeth_xdp_buff *xdp,
>>> const struct page *page = __netmem_to_page(fqe->netmem);
>>>
>>> #ifdef __LIBETH_WORD_ACCESS
>>> - static_assert(offsetofend(typeof(xdp->base), flags) -
>>> - offsetof(typeof(xdp->base), frame_sz) ==
>>> - sizeof(u64));
>>> -
>>> - *(u64 *)&xdp->base.frame_sz = fqe->truesize;
>>> + xdp->base.frame_sz_le_qword = fqe->truesize;
>>> #else
>>> xdp_init_buff(&xdp->base, fqe->truesize, xdp->base.rxq);
>>> #endif
>>> diff --git a/include/net/xdp.h b/include/net/xdp.h
>>> index b40f1f96cb11..b5eedeb82c9b 100644
>>> --- a/include/net/xdp.h
>>> +++ b/include/net/xdp.h
>>> @@ -85,8 +85,19 @@ struct xdp_buff {
>>> void *data_hard_start;
>>> struct xdp_rxq_info *rxq;
>>> struct xdp_txq_info *txq;
>>> - u32 frame_sz; /* frame size to deduce data_hard_end/reserved tailroom*/
>>> - u32 flags; /* supported values defined in xdp_buff_flags */
>>> + union {
>>> + /* Allow setting frame_sz and flags as a single u64 on
>>> + * little endian systems. This may may give optimal
>>> + * performance. */
>>> + u64 frame_sz_le_qword;
>>> + struct {
>>> + /* Frame size to deduce data_hard_end/reserved
>>> + * tailroom. */
>>> + u32 frame_sz;
>>> + /* Supported values defined in xdp_buff_flags. */
>>> + u32 flags;
>>> + };
>>> + };
>>> };
>>
>> Yeah, this looks like a nice way to express this, and is way more
>> descriptive than "(u64 *)&xdp->base.frame_sz" :)
>
> Sounds good to me!
>
> Let me send v4 where I'll fix this.
Note: would it be okay if I send v4 with this fix when the window opens,
while our validation will retest v3 from Tony's tree in meantine? It's a
cosmetic change anyway and does not involve any functional changes.
Thanks,
Olek
