Re: linux-next: build failure after merge of the bpf-next tree

[Alexei Starovoitov]

On Wed, Mar 19, 2025 at 12:44 PM Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
>
> On Wed, Mar 19, 2025 at 7:56 PM Alexei Starovoitov
> <alexei.starovoitov@xxxxxxxxx> wrote:
> >
> > On Wed, Mar 19, 2025 at 9:06 AM Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
> > >
> > > On Wed, Mar 19, 2025 at 3:55 PM Alexei Starovoitov
> > > <alexei.starovoitov@xxxxxxxxx> wrote:
> > > >
> > > > On Wed, Mar 19, 2025 at 7:36 AM Kumar Kartikeya Dwivedi
> > > > <memxor@xxxxxxxxx> wrote:
> > > > >
> > > > > > >
> > > > > > > I've sent a fix [0], but unfortunately I was unable to reproduce the
> > > > > > > problem with an LLVM >= 19 build, idk why. I will try with GCC >= 14
> > > > > > > as the patches require to confirm, but based on the error I am 99%
> > > > > > > sure it will fix the problem.
> > > > > >
> > > > > > Probably because __seg_gs has CC_HAS_NAMED_AS depends on CC_IS_GCC.
> > > > > > Let me give it a go with GCC.
> > > > > >
> > > > >
> > > > > Can confirm now that this fixes it, I just did a build with GCC 14
> > > > > where Uros's __percpu checks kick in.
> > > >
> > > > Great. Thanks for checking and quick fix.
> > > >
> > > > btw clang supports it with __attribute__((address_space(256))),
> > > > so CC_IS_GCC probably should be relaxed.
> > >
> > > https://github.com/llvm/llvm-project/issues/93449
> > >
> > > needs to be fixed first. Also, the feature has to be thoroughly tested
> > > (preferably by someone having a deep knowledge of clang) before it is
> > > enabled by default.
> >
> > clang error makes sense to me.
>
> It is not an error, but an internal compiler error. This should never happen.

Not quite. llvm backends don't have a good way to explain the error,
but this is invalid condition.
Arguably llvm should do a better job in such cases instead of
printing stack trace.

>
> > What does it even mean to do addr space cast from percpu to normal address:
> >
> > __typeof__(int __seg_gs) const_pcpu_hot;
> > void *__attribute____UNIQUE_ID___addressable_const_pcpu_hot612 =
> >     (void *)(long)&const_pcpu_hot;
>
> Please see [1] for an explanation.
>
> [1] https://gcc.gnu.org/onlinedocs/gcc/Named-Address-Spaces.html#x86-Named-Address-Spaces

You didn't answer my question.

As suspected, gcc is producing garbage code.

See:
https://godbolt.org/z/ozozYY3nv

For
void *ptr = (void *)(long)&pcpu_hot;

gcc emits
.quad pcpu_hot
which is nonsensical, while clang refuses to produce garbage
and dumps stack.

Sadly, both compilers produce garbage for ret_addr()

and both compilers produce correct code for ret_value().
At least something.

Uros,
your percpu code is broken.
you shouldn't rely on gcc producing garbage.
Sooner or later gcc will start erroring on it just as clang.