Re: [PATCH v3] KVM: x86: Fix recording of guest steal time / preempted status

[David Woodhouse <dwmw2@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 2137 bytes --]

On Thu, 2021-11-11 at 14:23 +0100, Paolo Bonzini wrote:
> Queued with these changes.

Thanks. I note...

    [I didn't entirely agree with David's assessment of the
     usefulness of the gfn_to_pfn cache, and integrated the outcome
     of the discussion in the above commit message. - Paolo]

 ... the key change being the 'Until' in:

    Until the gfn_to_pfn cache handles the remapping automatically by
    integrating with the MMU notifiers, we might as well not get a
    kernel mapping of it...

I do not recall that we'd actually reached a conclusion that we *will*
make the gfn_to_pfn cache generally usable in that fashion. The latest
I knew of that discussion was my message at 
https://lore.kernel.org/kvm/55a5d4e3fbd29dd55e276b97eeaefd0411b3290b.camel@infradead.org/
in which I said I'd be a whole lot happier with that if we could do it
with RCU instead of an rwlock — but I don't think we can because we'd
need to call synchronize_srcu() in the MMU notifier callback that might
not be permitted to sleep?

I'm also slightly less comfortable with having the MMU notifier work
through an arbitrary *list* of gfn_to_pfn caches that it potentially
needs to invalidate, but that is very much a minor concern compared
with the first.

I started looking through the nested code which is the big user of this
facility. The important part of the gfn_to_pfn mapping as I've used it
for Xen event channel delivery is the fast path in the common case,
falling back to a slow path that needs to sleep, to revalidate the
mapping. That fast vs. slow path (with a workqueue) already existed for
irqfd delivery and I just needed to hook into it in the right places.

I didn't see anything in nested code that would benefit from that same
setup, and AFAICT it should all be running with current->mm == kvm->mm
so surely it ought to be able to just access things using the userspace
HVA and sleep if necessary?

(There's an *entirely* gratuitous one in nested_cache_shadow_vmcs12()
which does a map/memcpy/unmap that really ought to be kvm_read_guest().
I'll send a patch for that shortly)



[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 5174 bytes --]