Re: [PATCH] locking/atomic/x86: Introduce try_cmpxchg64

From: Uros Bizjak
Date: Wed May 11 2022 - 04:24:41 EST


On Wed, May 11, 2022 at 9:54 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Tue, May 10, 2022 at 07:07:25PM +0200, Uros Bizjak wrote:
> > On Tue, May 10, 2022 at 6:55 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > On Tue, May 10, 2022 at 05:42:17PM +0200, Uros Bizjak wrote:
> > > > This patch adds try_cmpxchg64 to improve code around cmpxchg8b. While
> > > > the resulting code improvements on x86_64 are minor (a compare and a move saved),
> > > > the improvements on x86_32 are quite noticeable. The code improves from:
> > >
> > > What user of cmpxchg64 is this?
> >
> > This is cmpxchg64 in pi_try_set_control from
> > arch/x86/kvm/vmx/posted_intr.c, as shown in a RFC patch [1].
>
> I can't read that code, my brain is hard wired to read pi as priority
> inheritance/inversion.
>
> Still, does 32bit actually support that stuff?

Unfortunately, it does:

kvm-intel-y += vmx/vmx.o vmx/vmenter.o vmx/pmu_intel.o vmx/vmcs12.o \
vmx/evmcs.o vmx/nested.o vmx/posted_intr.o

And when existing cmpxchg64 is substituted with cmpxchg, the
compilation dies for 32bits with:

error: call to ‘__cmpxchg_wrong_size’ declared with attribute error:
Bad argument size for cmpxchg

So, the majority of the patch deals with 32bits and tries to implement
the inlined insn correctly for all cases. The 64bit part is simply a
call to arch_try_cmpxchg, and the rest is auto-generated from scripts.

>
> > There are some more opportunities for try_cmpxchg64 in KVM, namely
> > fast_pf_fix_direct_spte in arch/x86/kvm/mmu/mmu.c and
> > tdp_mmu_set_spte_atomic in arch/x86/kvm/mmu/tdp_mmu.c
>
> tdp_mmu is definitely 64bit only and as such shouldn't need to use
> cmpxchg64.

Indeed.

>
> Anyway, your patch looks about right, but I find it *really* hard to
> care about 32bit code these days.

Thanks, this is also my sentiment, but I hope the patch will enable
better code and perhaps ease similar situation I have had elsewhere.

Uros.

On Wed, May 11, 2022 at 9:54 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Tue, May 10, 2022 at 07:07:25PM +0200, Uros Bizjak wrote:
> > On Tue, May 10, 2022 at 6:55 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > On Tue, May 10, 2022 at 05:42:17PM +0200, Uros Bizjak wrote:
> > > > This patch adds try_cmpxchg64 to improve code around cmpxchg8b. While
> > > > the resulting code improvements on x86_64 are minor (a compare and a move saved),
> > > > the improvements on x86_32 are quite noticeable. The code improves from:
> > >
> > > What user of cmpxchg64 is this?
> >
> > This is cmpxchg64 in pi_try_set_control from
> > arch/x86/kvm/vmx/posted_intr.c, as shown in a RFC patch [1].
>
> I can't read that code, my brain is hard wired to read pi as priority
> inheritance/inversion.
>
> Still, does 32bit actually support that stuff?

Unfortunately, it does:

kvm-intel-y += vmx/vmx.o vmx/vmenter.o vmx/pmu_intel.o vmx/vmcs12.o \
vmx/evmcs.o vmx/nested.o vmx/posted_intr.o

And when cmpxchg64 is substituted with cmpxchg, the compilation dies
for 32bits with:

error: call to ‘__cmpxchg_wrong_size’ declared with attribute error:
Bad argument size for cmpxchg

So, the majority of the patch deals with 32bits and tries to implement
the inlined insn correctly for all cases. The 64bit part is simply a
call to arch_try_cmpxchg, and the rest is auto-generated from scripts.

>
> > There are some more opportunities for try_cmpxchg64 in KVM, namely
> > fast_pf_fix_direct_spte in arch/x86/kvm/mmu/mmu.c and
> > tdp_mmu_set_spte_atomic in arch/x86/kvm/mmu/tdp_mmu.c
>
> tdp_mmu is definitely 64bit only and as such shouldn't need to use
> cmpxchg64.

Indeed.

>
> Anyway, your patch looks about right, but I find it *really* hard to
> care about 32bit code these days.

Thanks, this is also my sentiment, but I hope the patch will enable
better code and perhaps ease a similar situation elsewhere in the
sources.

Uros.