Re: One question about API cpumask_setall()

[Yury Norov <ynorov@nvidia.com>]

On Thu, Jul 30, 2026 at 02:41:42PM +0800, Bibo Mao wrote:
> 
> 
> On 2026/7/30 上午1:57, Yury Norov wrote:
> > On Wed, Jul 29, 2026 at 10:13:56AM +0800, Bibo Mao wrote:
> > > 
> > > 
> > > On 2026/7/29 上午9:55, Yury Norov wrote:
> > > > On Tue, Jul 28, 2026 at 04:16:27PM +0800, Bibo Mao wrote:
> > > > > Hi,
> > > > > 
> > > > > I am preparing to use API cpumask_setall() for TLB flush on LoongArch KVM
> > > > > side. There are cpumask_test_and_clear_cpu/cpumask_setall() APIs, it seems
> > > > > that API cpumask_setall() does not support atomic operation.
> > > > > 
> > > > > Do you think it is reasonable to add new API similar with cpumask_setall()
> > > > > which supports atomic operation?
> > > > 
> > > > Maybe reasonable, but it's impossible. cpumask is a bitmap, which in
> > > > turn is an array of unsigned longs. There's no atomic operations on
> > > > arrays.
> > > It should be workable with atomic operation on unsigned long, rather than
> > > the array of unsigned longs, the API cpumask_test_and_clear_cpu is atomic
> > > operation on unsigned long also.
> > 
> > No, cpumask_test_and_clear_cpu is atomic() operates on cpumask_t*,
> > which is a pointer to unsigned long.
> > > Is it ok to use smp_store_mb() to fill the bitmap with the similar API
> > > cpumask_setall()? where xchg() is used in smp_store_mb() wrapper instead.
> > 
> > If your bitmap is longer than 1 word, you'll have to issue 2 or more
> > smp_store_mb()'s. Then, a series of atomic operations is not atomic
> > itself.
> I understand what you say, the whole series of atomic operations is not
> atomic.
> 
> My concern is whether cpumask_test_and_clear_cpu() operation on cpumask_t*
> will invalidate non-atomic write operation on another CPU. From MESI/MOSI
> protocol, this will not happen, only that order of write and atomic
> test_and_clear operation may be different.

You're changing the topic now. You asked about the atomic version of
cpumask_setall(), and I've pointed that it's impossible.

Now you're asking about invalidation of caches. I can't say for each
and every architecture out there, and all their caching modes.

If you need to run cpumask_setall() concurrently with
cpumask_test_and_clear_cpu(), please use external locking mechanism.

Thanks,
Yury