On Fri, Oct 11, 2024, Yan Zhao wrote:
> On Thu, Oct 10, 2024 at 09:14:41AM -0700, Sean Christopherson wrote:
> > On Thu, Oct 10, 2024, Yan Zhao wrote:
> > > On Wed, Oct 09, 2024 at 12:23:44PM -0700, Sean Christopherson wrote:
> > > > Add a lockdep assertion in kvm_unmap_gfn_range() to ensure that either
> > > > mmu_invalidate_in_progress is elevated, or that the range is being zapped
> > > > due to memslot removal (loosely detected by slots_lock being held).
> > > > Zapping SPTEs without mmu_invalidate_{in_progress,seq} protection is unsafe
> > > > as KVM's page fault path snapshots state before acquiring mmu_lock, and
> > > > thus can create SPTEs with stale information if vCPUs aren't forced to
> > > > retry faults (due to seeing an in-progress or past MMU invalidation).
> > > >
> > > > Memslot removal is a special case, as the memslot is retrieved outside of
> > > > mmu_invalidate_seq, i.e. doesn't use the "standard" protections, and
> > > > instead relies on SRCU synchronization to ensure any in-flight page faults
> > > > are fully resolved before zapping SPTEs.
> > > >
> > > > Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> > > > ---
> > > > arch/x86/kvm/mmu/mmu.c | 10 ++++++++++
> > > > 1 file changed, 10 insertions(+)
> > > >
> > > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> > > > index 09494d01c38e..c6716fd3666f 100644
> > > > --- a/arch/x86/kvm/mmu/mmu.c
> > > > +++ b/arch/x86/kvm/mmu/mmu.c
> > > > @@ -1556,6 +1556,16 @@ bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
> > > > {
> > > > bool flush = false;
> > > >
> > > > + /*
> > > > + * To prevent races with vCPUs faulting in a gfn using stale data,
> > > > + * zapping a gfn range must be protected by mmu_invalidate_in_progress
> > > > + * (and mmu_invalidate_seq). The only exception is memslot deletion,
> > > > + * in which case SRCU synchronization ensures SPTEs a zapped after all
> > > > + * vCPUs have unlocked SRCU and are guaranteed to see the invalid slot.
> > > > + */
> > > > + lockdep_assert_once(kvm->mmu_invalidate_in_progress ||
> > > > + lockdep_is_held(&kvm->slots_lock));
> > > > +
> > > Is the detection of slots_lock too loose?
> >
> > Yes, but I can't think of an easy way to tighten it. My original thought was to
> > require range->slot to be invalid, but KVM (correctly) passes in the old, valid
> > memslot to kvm_arch_flush_shadow_memslot().
> >
> > The goal with the assert is to detect as many bugs as possible, without adding
> > too much complexity, and also to document the rules for using kvm_unmap_gfn_range().
> >
> > Actually, we can tighten the check, by verifying that the slot being unmapped is
> > valid, but that the slot that KVM sees is invalid. I'm not sure I love it though,
> > as it's absurdly specific.
> Right. It doesn't reflect the wait in kvm_swap_active_memslots() for the old
> slot.
>
> CPU 0 CPU 1
> 1. fault on old begins
> 2. swap to new
> 3. zap old
> 4. fault on old ends
>
> Without CPU 1 waiting for 1&4 complete between 2&3, stale data is still
> possible.
>
> So, the detection in kvm_memslot_is_being_invalidated() only indicates the
> caller is from kvm_arch_flush_shadow_memslot() with current code.
Yep, which is why I don't love it.
> Given that, how do you feel about passing in a "bool is_flush_slot" to indicate
> the caller and asserting?
I like it even less than the ugliness I proposed :-) It'd basically be a "I pinky
swear I know what I'm doing" flag, and I think the downsides of having true/false
literals in the code would outweigh the upside of the precise assertion.
