On Wed, Dec 15, 2021, Sean Christopherson wrote:
> On Wed, Dec 15, 2021, Paolo Bonzini wrote:
> > On 12/15/21 02:15, Sean Christopherson wrote:
> > > Patches 01-03 implement a bug fix by ensuring KVM zaps both valid and
> > > invalid roots when unmapping a gfn range (including the magic "all" range).
> > > Failure to zap invalid roots means KVM doesn't honor the mmu_notifier's
> > > requirement that all references are dropped.
> > >
> > > set_nx_huge_pages() is the most blatant offender, as it doesn't elevate
> > > mm_users and so a VM's entire mm can be released, but the same underlying
> > > bug exists for any "unmap" command from the mmu_notifier in combination
> > > with a memslot update. E.g. if KVM is deleting a memslot, and a
> > > mmu_notifier hook acquires mmu_lock while it's dropped by
> > > kvm_mmu_zap_all_fast(), the mmu_notifier hook will see the to-be-deleted
> > > memslot but won't zap entries from the invalid roots.
> > >
> > > Patch 04 is cleanup to reuse the common iterator for walking _only_
> > > invalid roots.
> > >
> > > Sean Christopherson (4):
> > > KVM: x86/mmu: Use common TDP MMU zap helper for MMU notifier unmap
> > > hook
> > > KVM: x86/mmu: Move "invalid" check out of kvm_tdp_mmu_get_root()
> > > KVM: x86/mmu: Zap _all_ roots when unmapping gfn range in TDP MMU
> > > KVM: x86/mmu: Use common iterator for walking invalid TDP MMU roots
> > >
> > > arch/x86/kvm/mmu/tdp_mmu.c | 116 +++++++++++++++++--------------------
> > > arch/x86/kvm/mmu/tdp_mmu.h | 3 -
> > > 2 files changed, 53 insertions(+), 66 deletions(-)
> > >
> >
> > Queued 1-3 for 5.16 and 4 for 5.17.
>
> Actually, can you please unqueue patch 4? I think we can actually kill off
> kvm_tdp_mmu_zap_invalidated_roots() entirely. I don't know if that code will be
> ready for 5.17, but if it is then this patch is unnecesary. And if not, it
> shouldn't be difficult to re-queue this a bit later.
Cancel that request, the comment above kvm_tdp_mmu_zap_invalidated_roots() lies,
as do the changelogs for commits b7cccd397f31 ("KVM: x86/mmu: Fast invalidation
for TDP MMU") and 4c6654bd160d ("KVM: x86/mmu: Tear down roots before
kvm_mmu_zap_all_fast returns"), and the fact that they are even separate commits.
KVM _must_ zap invalid roots before returning from kvm_mmu_zap_all_fast(), because
when it's called from kvm_mmu_invalidate_zap_pages_in_memslot(), KVM is relying on
it to fully remove all references to the memslot. Once the memslot is gone, KVM's
mmu_notifier hooks will be unable to find the stale references as the hva=>gfn
translation is done via the memslots. If userspace unmaps a range after deleting
a memslot, KVM will fail to zap in response to the mmu_notifier due to not finding
a memslot corresponding to the notifier's range, which leads to another variation
of the splat I've come to know and love.
WARNING: CPU: 33 PID: 44927 at arch/x86/kvm/../../../virt/kvm/kvm_main.c:173
RIP: 0010:kvm_is_zone_device_pfn+0x96/0xa0 [kvm]
kvm_set_pfn_dirty+0xa8/0xe0 [kvm]
__handle_changed_spte+0x2f7/0x5b0 [kvm]
__handle_changed_spte+0x2f7/0x5b0 [kvm]
__tdp_mmu_set_spte+0x64/0x170 [kvm]
tdp_mmu_zap_root+0x1f5/0x220 [kvm]
kvm_tdp_mmu_zap_all+0x47/0x60 [kvm]
kvm_mmu_zap_all+0xf0/0x100 [kvm]
kvm_mmu_notifier_release+0x2b/0x60 [kvm]
mmu_notifier_unregister+0x48/0xe0
kvm_destroy_vm+0x129/0x2a0 [kvm]
kvm_vm_release+0x1d/0x30 [kvm]
__fput+0x82/0x240
task_work_run+0x5c/0x90
exit_to_user_mode_prepare+0x114/0x120
syscall_exit_to_user_mode+0x1d/0x40
do_syscall_64+0x48/0xc0
entry_SYSCALL_64_after_hwframe+0x44/0xae
I'll include a patch in my flush+zap rework series to reword that comment, because
it is very, very misleading.
