Re: [PATCH] riscv: Add native this_cpu_cmpxchg() support

[Paul Walmsley]

Hi, 

On Thu, 30 Jul 2026, Xie Bo wrote:

> RISC-V falls back to the generic this_cpu_cmpxchg() implementation,
> which serializes the operation by disabling local interrupts.
> Consequently, HAVE_CMPXCHG_LOCAL is unset and users such as
> percpu_counter cannot use their cmpxchg-based fast paths.
> 
> Implement the 4-byte this_cpu_cmpxchg() operation with cmpxchg_local(),
> and provide the 8-byte operation on RV64. Pin execution while resolving
> the current CPU pointer so the LR/SC loop operates on one per-CPU
> instance, while allowing interrupt-context updates to race through the
> atomic operation.
> 
> Copy the old and new values to private temporaries before invoking
> cmpxchg_local(). This avoids collisions between local variable names in
> nested statement-expression macros.
> 
> The 1- and 2-byte operations continue to use the generic fallback.
> 
> Select HAVE_CMPXCHG_LOCAL and update the architecture feature matrix.
> 
> A percpu_counter_add() benchmark using the default batch value reduced
> the median time per operation by 77.3% (10 runs of 5,000,000
> operations).
> 
> In three 60-second stress-ng fork runs, median throughput increased by
> 4.6%. A 120-second combined fork and VM stress test completed without
> rss-counter errors or validation failures.
> 
> Signed-off-by: Xie Bo <xb@xxxxxxxxxxxxx>

This patch seems to be missing 128-bit cmpxchg support.  Care to add it?


thanks,

- Paul