[PATCH] arm64: Kconfig: Don't select CPUMASK_OFFSTACK with 64K pages

[Will Deacon]

On Fri, Oct 30, 2026 at 03:46:03AM -0700, Breno Leitao wrote:
> Hello Will,
> 
> On Fri, Oct 30, 2026 at 10:28:04AM +0000, Will Deacon wrote:
> > On Tue, Oct 27, 2026 at 04:59:52AM -0700, Breno Leitao wrote:
> > > On Wed, Oct 07, 2026 at 12:58:16AM -0700, Breno Leitao wrote:
> > > > arm64 selects CPUMASK_OFFSTACK when NR_CPUS is above 256, which keeps
> > > > cpumask_var_t off the kernel stack. With 64K pages the kernel stack is
> > > > 64K as well, we have enough room to keep cpumask on the stack, avoiding
> > > > extra alocation and possible performance penalties.
> > > > 
> > > > Skip the select when ARM64_64K_PAGES is set. A 64K kernel can then raise
> > > > NR_CPUS above 256 and keep cpumask_var_t as an inline array. 4K and 16K
> > > > page kernels are unchanged.
> > > > 
> > > > Signed-off-by: Breno Leitao <leitao at debian.org>
> > > > ---
> > > > Based on Will presentation at LPC, I've decided to have a look at Meta's
> > > > stack usage, and I found that this might be an small optimization!?
> > > > ---
> > > >  arch/arm64/Kconfig | 2 +-
> > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > > 
> > > > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > > > index 5f2db2cecf9772..b73371327d2a47 100644
> > > > --- a/arch/arm64/Kconfig
> > > > +++ b/arch/arm64/Kconfig
> > > > @@ -111,7 +111,7 @@ config ARM64
> > > >  	select CLONE_BACKWARDS
> > > >  	select COMMON_CLK
> > > >  	select CPU_PM if (SUSPEND || CPU_IDLE)
> > > > -	select CPUMASK_OFFSTACK if NR_CPUS > 256
> > > > +	select CPUMASK_OFFSTACK if NR_CPUS > 256 && !ARM64_64K_PAGES
> > > 
> > > Does this one sound sane?
> > 
> > It doesn't make an awful lot of sense to me, to be honest. The page size
> > and the number of CPUs are two independent variables, so tying them
> > together like this looks a little odd.
> 
> Sure, but with 64K pages the kernel stack is _huge_. Why do we still
> need to keep the cpumask off the stack frame in that case?

Because you might have thousands of CPUs? Because copying these things by
value could be expensive?

Will