Re: [PATCH v2 6/8] checksum: use csum_ip4_header() in udp.c and tcp.c

[Stefano Brivio <sbrivio@redhat.com>]

On Sat, 17 Feb 2024 15:22:12 +0100
Laurent Vivier <lvivier@redhat.com> wrote:

> On 2/16/24 19:24, Stefano Brivio wrote:
> > On Fri, 16 Feb 2024 19:05:39 +0100
> > Laurent Vivier <lvivier@redhat.com> wrote:
> >
> > ...  
> >> I think I will send the v3 of my series without fixing that because I don't have enough
> >> time this week. I will address the problem later.  
> > No problem! I will also try to spend a moment and see if there's some
> > reasonable solution I can suggest. Thanks,
> >  
> I can imagine 4 solutions:
> 
>   * to use inline functions (could it helps the compiler to manage the alignment problem?)

I guess in practice yes, but it could be formally complicated for a
compiler to make sure no instructions dereferencing those pointers will
be emitted, plus this is on the packet path and if the compiler decides
to *not* inline, we shouldn't force that.

>   * to use C macros

I'm not sure exactly how, I have some vague idea of what you might
mean, it could be quite awkward though.

>   * to use these new functions only with vhost-user as we know pointers will be aligned.

This is quite unlikely to help: the problem is that 802.3 (Ethernet)
frame headers are (without VLANs) 14 bytes.

If you align the start of the frame, and we need those frames (and
pointers to them) whenever we talk Layer-2, the rest can't be aligned
to 4-bytes boundary.

>   * to include structure we want to address in a generic wrapperstructure that will
>     unalign it as it is done with the current structure.

This sounds like the easiest and safest way to me. Note that pointers
to 'taph' can be happily dereferenced, too. You can pass around
pointers to that, instead of using 'iph'.

I used (almost everywhere?) the start of the buffer, but 'taph' is fine
as well.

-- 
Stefano