Support for atomic Compare-And-Swap instructions has been in the RISC-V
port of the Linux kernel for a long time. That being said, we apparently
never bothered to set HAVE_CMPXCHG_DOUBLE and HAVE_CMPXCHG_LOCAL in the
Kconfig, despite having all the framework to support them.

Signed-off-by: Miquel Sabaté Solà <mssola@mssola.com>
---
This is a resend of [1], rebased on top of the latest commit from the
for-next branch.

I have built this patch with multiple configurations and ran it with KVM
(the VisionFive2 board that I have lacks the needed extensions). All seems
to work, but I do wonder if we did not enable these for a reason or this
just slipped through. So far in the code I believe everything is in place,
and I haven't seen any commit in the git log stating otherwise.

[1] https://lore.kernel.org/all/20260220074449.8526-1-mssola@mssola.com/

 Documentation/features/locking/cmpxchg-local/arch-support.txt | 2 +-
 arch/riscv/Kconfig                                            | 2 ++
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/Documentation/features/locking/cmpxchg-local/arch-support.txt b/Documentation/features/locking/cmpxchg-local/arch-support.txt
index 2c3a4b91f16d..28d5fa8c3b4f 100644
--- a/Documentation/features/locking/cmpxchg-local/arch-support.txt
+++ b/Documentation/features/locking/cmpxchg-local/arch-support.txt
@@ -20,7 +20,7 @@
     |    openrisc: | TODO |
     |      parisc: | TODO |
     |     powerpc: | TODO |
-    |       riscv: | TODO |
+    |       riscv: |  ok  |
     |        s390: |  ok  |
     |          sh: | TODO |
     |       sparc: | TODO |
diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
index 1955fcc5effd..b8ca5792a392 100644
--- a/arch/riscv/Kconfig
+++ b/arch/riscv/Kconfig
@@ -152,6 +152,8 @@ config RISCV
 	select HAVE_ARCH_USERFAULTFD_WP if 64BIT && MMU && USERFAULTFD && RISCV_ISA_SVRSW60T59B
 	select HAVE_ARCH_VMAP_STACK if MMU && 64BIT
 	select HAVE_ASM_MODVERSIONS
+	select HAVE_CMPXCHG_DOUBLE if RISCV_ISA_ZACAS && RISCV_ISA_ZABHA
+	select HAVE_CMPXCHG_LOCAL if RISCV_ISA_ZACAS && RISCV_ISA_ZABHA
 	select HAVE_BUILDTIME_MCOUNT_SORT
 	select HAVE_CONTEXT_TRACKING_USER
 	select HAVE_DEBUG_KMEMLEAK
--
2.54.0

Hi,

On Fri, 5 Jun 2026, Miquel Sabaté Solà wrote:

> Support for atomic Compare-And-Swap instructions has been in the RISC-V
> port of the Linux kernel for a long time. That being said, we apparently
> never bothered to set HAVE_CMPXCHG_DOUBLE and HAVE_CMPXCHG_LOCAL in the
> Kconfig, despite having all the framework to support them.
> 
> Signed-off-by: Miquel Sabaté Solà <mssola@mssola.com>
> ---
> This is a resend of [1], rebased on top of the latest commit from the
> for-next branch.
> 
> I have built this patch with multiple configurations and ran it with KVM
> (the VisionFive2 board that I have lacks the needed extensions). All seems
> to work, but I do wonder if we did not enable these for a reason or this
> just slipped through. So far in the code I believe everything is in place,
> and I haven't seen any commit in the git log stating otherwise.
> 
> [1] https://lore.kernel.org/all/20260220074449.8526-1-mssola@mssola.com/

Thanks for the patch.  Your comments above are why I've been hesitant to 
merge it.  I'm not aware of any publicly available hardware that supports 
Zacas/Zabha.  No one has stepped forward to provide any Tested-by:s on 
hardware that hasn't been released yet.  You mention that you tested on 
your VisionFive2 board, but it would not have exercised those code paths.

Of course, we already have Zacas/Zabha support, merged back in 2024, in 
cmpxchg.h.  I assume (?) that it was tested in QEMU, but I don't see any 
comments about that in the patch series.  No one sent any Tested-by:s 
then, either.

It would be good if you (and ideally others) could put this patch through 
some testing on QEMU with Zacas and Zabha enabled, before we merge it.  
The affected code paths for HAVE_CMPXCHG_LOCAL seem to primarily involve 
per-CPU counters and MM zone counters, so those would be the areas to 
focus.  HAVE_CMPXCHG_DOUBLE seems to do nothing useful other than 
preventing the AMD IOMMU driver from being selected if it's not present, 
so that part of the patch seems fairly useless.  In fact I'd suggest 
dropping that from the patch and just sending a separate patch to remove 
HAVE_CMPXCHG_DOUBLE from the kernel completely.


- Paul

Hi,

Paul Walmsley @ 2026-06-06 18:50 -06:

> Hi,
>
> On Fri, 5 Jun 2026, Miquel Sabaté Solà wrote:
>
>> Support for atomic Compare-And-Swap instructions has been in the RISC-V
>> port of the Linux kernel for a long time. That being said, we apparently
>> never bothered to set HAVE_CMPXCHG_DOUBLE and HAVE_CMPXCHG_LOCAL in the
>> Kconfig, despite having all the framework to support them.
>>
>> Signed-off-by: Miquel Sabaté Solà <mssola@mssola.com>
>> ---
>> This is a resend of [1], rebased on top of the latest commit from the
>> for-next branch.
>>
>> I have built this patch with multiple configurations and ran it with KVM
>> (the VisionFive2 board that I have lacks the needed extensions). All seems
>> to work, but I do wonder if we did not enable these for a reason or this
>> just slipped through. So far in the code I believe everything is in place,
>> and I haven't seen any commit in the git log stating otherwise.
>>
>> [1] https://lore.kernel.org/all/20260220074449.8526-1-mssola@mssola.com/
>
> Thanks for the patch.  Your comments above are why I've been hesitant to
> merge it.  I'm not aware of any publicly available hardware that supports
> Zacas/Zabha.  No one has stepped forward to provide any Tested-by:s on
> hardware that hasn't been released yet.  You mention that you tested on
> your VisionFive2 board, but it would not have exercised those code paths.

No, I mention that I ran it _only_ on KVM, as my VisionFive2 board lacks
these extensions and hence I couldn't possible have tested this there :)

>
> Of course, we already have Zacas/Zabha support, merged back in 2024, in
> cmpxchg.h.  I assume (?) that it was tested in QEMU, but I don't see any
> comments about that in the patch series.  No one sent any Tested-by:s
> then, either.
>
> It would be good if you (and ideally others) could put this patch through
> some testing on QEMU with Zacas and Zabha enabled, before we merge it.
> The affected code paths for HAVE_CMPXCHG_LOCAL seem to primarily involve
> per-CPU counters and MM zone counters, so those would be the areas to
> focus.  HAVE_CMPXCHG_DOUBLE seems to do nothing useful other than
> preventing the AMD IOMMU driver from being selected if it's not present,
> so that part of the patch seems fairly useless.  In fact I'd suggest
> dropping that from the patch and just sending a separate patch to remove
> HAVE_CMPXCHG_DOUBLE from the kernel completely.

To be fair, on QEMU I only "tested" it by booting it, running a few
things for some time and ensuring that nothing got totally broken in the
process while taking a look at the kernel logs.

In any case, let me double check with QEMU with these extensions enabled
and I'll try to be more thorough about it. I'll do just that whenever I
have some spare time during the following week :)

As for HAVE_CMPXCHG_DOUBLE, removing it makes sense. Let me just take
another look and I will send a separate patch whenever I'm ready for it.

>
>
> - Paul

Thanks for your input!
Miquel