Re: [PATCH] locking/local_lock: Reduce local_[un]lock_nested_bh() overhead

[Uros Bizjak]

On Wed, Mar 11, 2026 at 5:32 PM Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
>
> On Wed, Mar 11, 2026 at 4:55 PM Sebastian Andrzej Siewior
> <bigeasy@xxxxxxxxxxxxx> wrote:
> >
> > On 2026-03-09 15:52:34 [+0100], Eric Dumazet wrote:
> > > > +void ericeric(void);
> > > > +void ericeric(void)
> > > > +{
> > > > +       raw_cpu_read_long(this_cpu_off);
> > > > +       raw_cpu_read_long(this_cpu_off);
> > > > +}
> > > >
> > > > I am guessing __raw_cpu_read() is forcing the asm ?
> > >
> > > Might be a clang issue. Oh well.
> >
> > So the difference is that with gcc we have USE_X86_SEG_SUPPORT and with
> > llvm we don't. This leads to two asm statements with LLVM of which only
> > one is eliminated. This optimisation origins in commit ca4256348660c
> > ("x86/percpu: Use C for percpu read/write accessors").
> >
> > __seg_fs and __seg_gs is supported by LLVM but enabling it leads to tons
> > warnings and aborts later.
>
> Tons of warnings is just due to clang being picky and warns for
> duplicated qualifiers, such as "__seg_gs __seg_gs var". This can be
> fixed with:
>
> https://lore.kernel.org/lkml/20240526175655.227798-1-ubizjak@xxxxxxxxx/
>
> > Is there something missing in LLVM? The generated code for
> > raw_cpu_read_long(this_cpu_off) looks fine.
>
> Yes:
>
> 1. The %fs: and %gs: prefix does not get emitted in inline assembly.
>
> 2. An internal compiler error when addressing symbols directly:
> https://github.com/llvm/llvm-project/issues/93449
>
> 3. Wrong named address space for anonymous struct:
> https://github.com/llvm/llvm-project/issues/119705

BTW: A related issue is that ASAN fails to handle gs: prefixed
addresses. For GCC, we have had to disable ASAN instrumentation for
all locations in non-default address spaces. With asm accessors, the
access is hidden from ASAN, and the memory access is not instrumented
anyways.

Uros.