On Mon, May 15, 2017 at 02:36:58PM +0100, Suzuki K Poulose wrote:
> On 15/05/17 11:00, Christoffer Dall wrote:
> >Hi Suzuki,
> >
> >On Wed, May 03, 2017 at 03:17:52PM +0100, Suzuki K Poulose wrote:
> >>We yield the kvm->mmu_lock occassionaly while performing an operation
> >>(e.g, unmap or permission changes) on a large area of stage2 mappings.
> >>However this could possibly cause another thread to clear and free up
> >>the stage2 page tables while we were waiting for regaining the lock and
> >>thus the original thread could end up in accessing memory that was
> >>freed. This patch fixes the problem by making sure that the stage2
> >>pagetable is still valid after we regain the lock. The fact that
> >>mmu_notifer->release() could be called twice (via __mmu_notifier_release
> >>and mmu_notifier_unregsister) enhances the possibility of hitting
> >>this race where there are two threads trying to unmap the entire guest
> >>shadow pages.
> >>
> >>While at it, cleanup the redudant checks around cond_resched_lock in
> >>stage2_wp_range(), as cond_resched_lock already does the same checks.
> >>
> >>Cc: Mark Rutland <mark.rutland@xxxxxxx>
> >>Cc: Radim Krčmář <rkrcmar@xxxxxxxxxx>
> >>Cc: andreyknvl@xxxxxxxxxx
> >>Cc: Christoffer Dall <christoffer.dall@xxxxxxxxxx>
> >>Cc: Marc Zyngier <marc.zyngier@xxxxxxx>
> >>Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> >>Signed-off-by: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> >>---
> >> arch/arm/kvm/mmu.c | 17 ++++++++++++-----
> >> 1 file changed, 12 insertions(+), 5 deletions(-)
> >>
> >>diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c
> >>index 909a1a7..5b3e0db 100644
> >>--- a/arch/arm/kvm/mmu.c
> >>+++ b/arch/arm/kvm/mmu.c
> >>@@ -301,9 +301,14 @@ static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
> >> /*
> >> * If the range is too large, release the kvm->mmu_lock
> >> * to prevent starvation and lockup detector warnings.
> >>+ * Make sure the page table is still active when we regain
> >>+ * the lock.
> >> */
> >>- if (next != end)
> >>+ if (next != end) {
> >> cond_resched_lock(&kvm->mmu_lock);
> >>+ if (!READ_ONCE(kvm->arch.pgd))
> >>+ break;
> >>+ }
> >
> >So I don't think this change is wrong, but I wonder if it's sufficient.
> >For example, I can see that this function is called from
> >
> >stage2_unmsp_vm
> > -> stage2_unmap_memslot
> > -> unmap_stage2_range
> >
> >and
> >
> >kvm_arch_flush_shadow_memslot
> > -> unmap_stage2_range
> >
> >which never check if the pgd pointer is valid,
>
> You are right. Those two callers do not check it. We could fix all of this by simply
> moving the check to the beginning of the loop.
> i.e, something like this :
>
> @@ -295,6 +295,12 @@ static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
> assert_spin_locked(&kvm->mmu_lock);
> pgd = kvm->arch.pgd + stage2_pgd_index(addr);
> do {
> + /*
> + * Make sure the page table is still active, as we could
> + * another thread could have possibly freed the page table.
> + */
> + if (!READ_ONCE(kvm->arch.pgd))
> + break;
> next = stage2_pgd_addr_end(addr, end);
> if (!stage2_pgd_none(*pgd))
> unmap_stage2_puds(kvm, pgd, addr, next);
>
>
>
>
> >and finally, kvm_free_stage2_pgd also checks the pgd pointer outside of holding the
> >kvm->mmu_lock so why is this not racy?
>
> This has been fixed by patch 1 in the series. So should be fine.
>
>
> I can respin the patch with the changes if you are OK with it.
>
Yes, absolutely. I've already applied patch 1 so no need to include
that in your respin.
Thanks!
-Christoffer
