Re: [PATCH RESEND 1/1] lib/Kconfig: remove DEBUG_PER_CPU_MAPS dependency for CPUMASK_OFFSTACK

[Libo Chen]

On 4/14/22 04:41, Arnd Bergmann wrote:
> On Wed, Apr 13, 2022 at 11:50 PM Libo Chen <libo.chen@xxxxxxxxxx> wrote:
> > On 4/13/22 13:52, Arnd Bergmann wrote:
> > > > > Yes, it is. I don't know that the problem is...
> > > > Masahiro explained that CPUMASK_OFFSTACK can only be configured by
> > > > options not users if DEBUG_PER_CPU_MASK is not enabled. This doesn't
> > > > seem to be what we want.
> > > I think the correct way to do it is to follow x86 and powerpc, and tying
> > > CPUMASK_OFFSTACK to "large" values of CONFIG_NR_CPUS.
> > > For smaller values of NR_CPUS, the onstack masks are obviously
> > > cheaper, we just need to decide what the cut-off point is.
> > I agree. It appears enabling CPUMASK_OFFSTACK breaks kernel builds on
> > some architectures such as parisc and nios2 as reported by kernel test
> > robot. Maybe it makes sense to use DEBUG_PER_CPU_MAPS as some kind of
> > guard on CPUMASK_OFFSTACK.
> NIOS2 does not support SMP builds at all, so it should never be possible to
> select CPUMASK_OFFSTACK there. We may want to guard
> DEBUG_PER_CPU_MAPS by adding a 'depends on SMP' in order to
> prevent it from getting selected.
>
> For PARISC, the largest configuration is 32-way SMP, so CPUMASK_OFFSTACK
> is clearly pointless there as well, even though it should technically
> be possible
> to support. What is the build error on parisc?
Similar to NIOS2, A bunch of undefined references to *_cpumask_var() 
calls.  It seems that PARISC doesn't support the cpumask offstack API at all

> > > In x86, the onstack masks can be selected for normal SMP builds with
> > > up to 512 CPUs, while CONFIG_MAXSMP=y raises the limit to 8192
> > > CPUs while selecting CPUMASK_OFFSTACK.
> > > PowerPC does it the other way round, selecting CPUMASK_OFFSTACK
> > > implicitly whenever NR_CPUS is set to 8192 or more.
> > >
> > > I think we can easily do the same as powerpc on arm64. With the
> > I am leaning more towards x86's way because even NR_CPUS=160 is too
> > expensive for 4-core arm64 VMs according to apachebench. I highly doubt
> > that there is a good cut-off point to make everybody happy (or not unhappy).
> It seems surprising that you would see any improvement for offstack masks
> when using NR_CPUS=160, that is just three 64-bit words worth of data, but
> it requires allocating the mask dynamically, which takes way more memory
> to initialize.
>
> > > ApacheBench test you cite in the patch description, what is the
> > > value of NR_CPUS at which you start seeing a noticeable
> > > benefit for offstack masks? Can you do the same test for
> > > NR_CPUS=1024 or 2048?
> > As mentioned above, a good cut-off point moves depends on the actual
> > number of CPUs. But yeah I can do the same test for 1024 or even smaller
> > NR_CPUs values on the same 64-core arm64 VM setup.
> If you see an improvement for small NR_CPUS values using offstack masks,
> it's possible that the actual difference is something completely
> different and we
> can just make the on-stack case faster, possibly the cause is something about
> cacheline alignment or inlining decisions using your specific kernel config.
>
> Are you able to compare the 'perf report' output between runs with either
> size to see where the extra time gets spent?
Okay yeah I will take some time to gather more data. It does appear that 
something else may also contribute to the performance difference.

Thanks
Libo
>          Arnd