[01/10] kvm: mmu: spte_write_protect optimization

Commit Message

Junaid Shahid Oct. 20, 2018, 3:15 a.m. UTC

This change uses a lighter-weight function instead of mmu_spte_update()
in the common case in spte_write_protect(). This helps speed up the
get_dirty_log IOCTL.

Signed-off-by: Junaid Shahid <[e-mail]>
---
 arch/x86/kvm/mmu.c | 25 +++++++++++++++++++++----
 1 file changed, 21 insertions(+), 4 deletions(-)

Comments

Junaid Shahid Oct. 30, 2018, 9:37 p.m. UTC | #1

On 10/19/2018 10:49 PM, Peter Feiner wrote:
> 
> What about handling an spte with AD bits? Don't you have to mark the pfn dirty if the D bit is set? And that would require a TLB shootdown *after* clearing the D bit, right? Which would require 2 EPT walks to be implemented using a single shoot down. ... What am I missing here?

This function is only used for write-protection, not clearing D bits. The D-bit test/clear happens in a different code path, through spte_test_and_clear_dirty(). We do need a TLB shootdown afterwards, but it can be done just once after going over the whole memory slot.

Thanks,
Junaid

Patch

diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
index 4cf43ce42959..189e21c77525 100644
--- a/arch/x86/kvm/mmu.c
+++ b/arch/x86/kvm/mmu.c
@@ -1460,6 +1460,22 @@  static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
 		kvm_flush_remote_tlbs(vcpu->kvm);
 }
 
+static bool spte_test_and_clear_writable(u64 *sptep)
+{
+	u64 spte = *sptep;
+
+	if (spte & PT_WRITABLE_MASK) {
+		clear_bit(PT_WRITABLE_SHIFT, (ulong *)sptep);
+
+		if (!spte_ad_enabled(spte))
+			kvm_set_pfn_dirty(spte_to_pfn(spte));
+
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Write-protect on the specified @sptep, @pt_protect indicates whether
  * spte write-protection is caused by protecting shadow page table.
@@ -1483,11 +1499,12 @@  static bool spte_write_protect(u64 *sptep, bool pt_protect)
 
 	rmap_printk("rmap_write_protect: spte %p %llx\n", sptep, *sptep);
 
-	if (pt_protect)
-		spte &= ~SPTE_MMU_WRITEABLE;
-	spte = spte & ~PT_WRITABLE_MASK;
+	if (pt_protect) {
+		spte &= ~(PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE);
+		return mmu_spte_update(sptep, spte);
+	}
 
-	return mmu_spte_update(sptep, spte);
+	return spte_test_and_clear_writable(sptep);
 }
 
 static bool __rmap_write_protect(struct kvm *kvm,