Re: [PATCH 09/22] KVM: x86/mmu: Try "unprotect for retry" iff there are indirect SPs

[Sean Christopherson <seanjc@google.com>]

On Thu, Aug 15, 2024, Paolo Bonzini wrote:
> On Thu, Aug 15, 2024 at 4:09 PM Sean Christopherson <seanjc@google.com> wrote:
> > > (This is preexisting in reexecute_instruction() and goes away in patch 18, if
> > > I'm pre-reading that part of the series correctly).
> > >
> > > Bonus points for opportunistically adding a READ_ONCE() here and in
> > > kvm_mmu_track_write().
> >
> > Hmm, right, this one should have a READ_ONCE(), but I don't see any reason to
> > add one in kvm_mmu_track_write().  If the compiler was crazy and generate multiple
> > loads between the smp_mb() and write_lock(), _and_ the value transitioned from
> > 1->0, reading '0' on the second go is totally fine because it means the last
> > shadow page was zapped.  Amusingly, it'd actually be "better" in that it would
> > avoid unnecessary taking mmu_lock.
> 
> Your call, but I have started leaning towards always using
> READ_ONCE(), similar to all atomic_t accesses are done with
> atomic_read(); that is, just as much as a marker for cross-thread
> lock-free accesses, in addition to limiting the compiler's
> optimizations.
> 
> tools/memory-model/Documentation/access-marking.txt also suggests
> using READ_ONCE() and WRITE_ONCE() always except in special cases.
> They are also more friendly to KCSAN (though I have never used it).
> 
> This of course has the issue of being yet another unfinished transition.

I opted to fix the kvm_vcpu_exit_request() case[*], and add the READ_ONCE() to
this patch, but left kvm_mmu_track_write() as-is.

My reasoning, and what I think makes for a decent policy, is that while I 100%
agree lockless accesses need _some_ form of protection/documentation, I think
adding READ_ONCE() (and WRITE_ONCE()) on top adds confusion and makes the actual
requirement unclear.

In other words, if there's already an smp_rmb() or smp_wmb() (or similar), then
don't add READ/WRITE_ONCE() (unless that's also necesary for some reason) because
doing so detracts from the barriers that are actually necessary.

[*] https://lore.kernel.org/all/20240828232013.768446-1-seanjc@google.com

> > Obviously the READ_ONCE() would be harmless, but IMO it would be more confusing
> > than helpful, e.g. would beg the question of why kvm_vcpu_exit_request() doesn't
> > wrap vcpu->mode with READ_ONCE().  Heh, though arguably vcpu->mode should be
> > wrapped with READ_ONCE() since it's a helper and could be called multiple times
> > without any code in between that would guarantee a reload.
> 
> Indeed, who said I wouldn't change that one as well? :)
> 
> Paolo
>