Re: [PATCH 2/2] KVM: SVM: drop useless cpumask_test_cpu() in pre_sev_run()

[Yury Norov <yury.norov@gmail.com>]

On Mon, Aug 11, 2025 at 03:04:50PM -0700, Sean Christopherson wrote:
> On Mon, Aug 11, 2025, Yury Norov wrote:
> > On Mon, Aug 11, 2025 at 01:45:46PM -0700, Sean Christopherson wrote:
> > > On Mon, Aug 11, 2025, Yury Norov wrote:
> > > > Testing cpumask for a CPU to be cleared just before setting the exact
> > > > same CPU is useless because the end result is always the same: CPU is
> > > > set.
> > > 
> > > No, it is not useless.  Blindly writing to the variable will unnecessarily bounce
> > > the cacheline, and this is a hot path.
> > 
> > How hot is that path?
> 
> Very, it gets hit on every VM-Exit => VM-Entry.  For context, putting a single
> printk anywhere in KVM's exit=>entry path can completely prevent forward progress
> in the guest (for some workloads/patterns).
> 
> > How bad the cache contention is?
> 
> I would expect it to be "fatally" bad for some workloads and setups.  Not literally
> fatal, but bad enough that it would require an urgent fix.
> 
> > Is there any evidence that conditional cpumask_set_cpu() worth the effort?
> 
> I don't have evidence for this specific code flow, but there is plenty of evidence
> that shows that generating atomic accesses, especially across sockets, can have a
> significant negative impact on performance.
> 
> I didn't ask for performance numbers for optimizing setting the mask because (a)
> I know the VM-Entry path can be extremely hot, (b) I know that dueling atomics
> can be hugely problematic, and (c) I don't see the separate test + set logic as
> being at all notable in terms of effort.
> 
> > The original patch doesn't discuss that at all, and without any comment the
> > code looks just buggy.
> 
> FWIW, there was discussion in a previous version of the series, but no hard
> numbers on the perf impact.
> 
> https://lore.kernel.org/all/Z75se_OZQvaeQE-4@google.com

OK, I see. So, as I said I don't insist on moving this patch. Let's
drop it if you think that cache contention is critical.

I probably need to think in the opposite direction - if some code
pieces trash caches by concurrent accessing the whole cache line just
to set a single bit, and people try to minimize that by using
conditional set/clear_bit(), we need to make it as effective as we
can, and a part of the API.

Thanks for the discussion, Sean!