KVM: x86/mmu: Reintroduce fast invalidate/zap for flushing memslot

James Harvey reported a livelock that was introduced by commit
d012a06ab1d23 ("Revert "KVM: x86/mmu: Zap only the relevant pages when
removing a memslot"").

The livelock occurs because kvm_mmu_zap_all() as it exists today will
voluntarily reschedule and drop KVM's mmu_lock, which allows other vCPUs
to add shadow pages.  With enough vCPUs, kvm_mmu_zap_all() can get stuck
in an infinite loop as it can never zap all pages before observing lock
contention or the need to reschedule.  The equivalent of kvm_mmu_zap_all()
that was in use at the time of the reverted commit (4e103134b8623, "KVM:
x86/mmu: Zap only the relevant pages when removing a memslot") employed
a fast invalidate mechanism and was not susceptible to the above livelock.

There are three ways to fix the livelock:

- Reverting the revert (commit d012a06ab1d23) is not a viable option as
  the revert is needed to fix a regression that occurs when the guest has
  one or more assigned devices.  It's unlikely we'll root cause the device
  assignment regression soon enough to fix the regression timely.

- Remove the conditional reschedule from kvm_mmu_zap_all().  However, although
  removing the reschedule would be a smaller code change, it's less safe
  in the sense that the resulting kvm_mmu_zap_all() hasn't been used in
  the wild for flushing memslots since the fast invalidate mechanism was
  introduced by commit 6ca18b6950f8d ("KVM: x86: use the fast way to
  invalidate all pages"), back in 2013.

- Reintroduce the fast invalidate mechanism and use it when zapping shadow
  pages in response to a memslot being deleted/moved, which is what this
  patch does.

For all intents and purposes, this is a revert of commit ea145aacf4ae8
("Revert "KVM: MMU: fast invalidate all pages"") and a partial revert of
commit 7390de1e99a70 ("Revert "KVM: x86: use the fast way to invalidate
all pages""), i.e. restores the behavior of commit 5304b8d37c2a5 ("KVM:
MMU: fast invalidate all pages") and commit 6ca18b6950f8d ("KVM: x86:
use the fast way to invalidate all pages") respectively.

Fixes: d012a06ab1d23 ("Revert "KVM: x86/mmu: Zap only the relevant pages when removing a memslot"")
Reported-by: James Harvey <jamespharvey20@gmail.com>
Cc: Alex Willamson <alex.williamson@redhat.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

1 files changed, 2 insertions, 0 deletions

diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 74e88e5edd9c..bdc16b0aa7c6 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -335,6 +335,7 @@ struct kvm_mmu_page {
 	int root_count;          /* Currently serving as active root */
 	unsigned int unsync_children;
 	struct kvm_rmap_head parent_ptes; /* rmap pointers to parent sptes */
+	unsigned long mmu_valid_gen;
 	DECLARE_BITMAP(unsync_child_bitmap, 512);
 
 #ifdef CONFIG_X86_32
@@ -856,6 +857,7 @@ struct kvm_arch {
 	unsigned long n_requested_mmu_pages;
 	unsigned long n_max_mmu_pages;
 	unsigned int indirect_shadow_pages;
+	unsigned long mmu_valid_gen;
 	struct hlist_head mmu_page_hash[KVM_NUM_MMU_PAGES];
 	/*
 	 * Hash table of struct kvm_mmu_page.