Re: [PATCH] hash: make gfni stubs inline

[David Marchand <david.marchand@redhat.com>]

On Tue, Mar 5, 2024 at 6:53 PM Tyler Retzlaff
<roretzla@linux.microsoft.com> wrote:
>
> On Tue, Mar 05, 2024 at 11:14:45AM +0100, David Marchand wrote:
> > On Mon, Mar 4, 2024 at 7:45 PM Stephen Hemminger
> > <stephen@networkplumber.org> wrote:
> > >
> > > This reverts commit 07d836e5929d18ad6640ebae90dd2f81a2cafb71.
> > >
> > > Tyler found build issues with MSVC and the thash gfni stubs.
> > > The problem would be link errors from missing symbols.
> >
> > Trying to understand this link error.
> > Does it come from the fact that rte_thash_gfni/rte_thash_gfni_bulk
> > declarations are hidden under RTE_THASH_GFNI_DEFINED in
> > rte_thash_gfni.h?
> >
> > If so, why not always expose those two symbols unconditionnally and
> > link with the stub only when ! RTE_THASH_GFNI_DEFINED.
>
> So I don't have a lot of background of this lib.
>
> I think we understand that we can't conditionally expose symbols. That's
> what windows was picking up because it seems none of our CI's ever end
> up with RTE_THASH_GFNI_DEFINED but my local test system did and failed.
> (my experiments showed that Linux would complain too if it was defined)

I can't reproduce a problem when I build (gcc/clang) for a target that
has GFNI/AVX512F.
binutils ld seems to just ignore unknown symbols in the map.

With current main:
[dmarchan@dmarchan main]$ nm build/lib/librte_hash.so.24.1 | grep rte_thash_gfni
00000000000088b0 T rte_thash_gfni_supported
[dmarchan@dmarchan main]$ nm build-nogfni/lib/librte_hash.so.24.1 |
grep rte_thash_gfni
00000000000102c0 T rte_thash_gfni
00000000000102d0 T rte_thash_gfni_bulk
000000000000294e t rte_thash_gfni_bulk.cold
0000000000002918 t rte_thash_gfni.cold
000000000000d3c0 T rte_thash_gfni_supported


>
> If we always expose the symbols then as you point out we have to
> conditionally link with the stub otherwise the inline (non-stub) will be
> duplicate and build / link will fail.
>
> I guess the part I don't understand with your suggestion is how we would
> conditionally link with just the stub? We have to link with rte_hash to
> get the rest of hash and the stub. I've probably missed something here.

No we can't, Stephen suggestion is a full solution.

>
> Since we never had a release exposing the new symbols introduced by
> Stephen in question my suggestion was that we just revert for 24.03 so
> we don't end up with an ABI break later if we choose to solve the
> problem without exports.
>
> I don't know what else to do, but I think we need to decide for 24.03.

I am fully aware that we must fix this for 24.03.

I would like to be sure Stephen fix (see v3) works for you, so have a
look because I am not able to reproduce an issue and validate the fix
myself.


-- 
David Marchand