Skip to contentSkip to conversation

KVM: x86/mmu: consider the hva in mmu_notifier retry

PATCH In review David Stevens 2 participants 2 messages Jan 27, 2021

Subsystems (7)
x86 KVM PowerPC MIPS +3 more subsystems
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 KVM: x86/mmu: consider the hva in mmu_notifier retry

From: David Stevens <[email protected]>

Track the range being invalidated by mmu_notifier and skip page fault retries if the fault address is not affected by the in-progress invalidation. Handle concurrent invalidations by finding the minimal range which includes all ranges being invalidated. Although the combined range may include unrelated addresses and cannot be shrunk as individual invalidation operations complete, it is unlikely the marginal gains of proper range tracking are worth the additional complexity.

The primary benefit of this change is the reduction in the likelihood of extreme latency when handing a page fault due to another thread having been preempted while modifying host virtual addresses.

---
v1 -> v2:
- improve handling of concurrent invalidation requests by unioning ranges, instead of just giving up and using [0, ULONG_MAX).
- add lockdep check
- code comments and formatting

 arch/powerpc/kvm/book3s_64_mmu_hv.c    |  2 +-
 arch/powerpc/kvm/book3s_64_mmu_radix.c |  2 +-
 arch/x86/kvm/mmu/mmu.c                 | 16 ++++++++------
 arch/x86/kvm/mmu/paging_tmpl.h         |  7 ++++---
 include/linux/kvm_host.h               | 27 +++++++++++++++++++++++-
 virt/kvm/kvm_main.c                    | 29 ++++++++++++++++++++++----
 6 files changed, 67 insertions(+), 16 deletions(-)
arch/powerpc/kvm/book3s_64_mmu_hv.c +1 -1
index 38ea396a23d6..8e06cd3f759c 100644
@@ -590,7 +590,7 @@ int kvmppc_book3s_hv_page_fault(struct kvm_vcpu *vcpu,
 	} else {
 		/* Call KVM generic code to do the slow-path check */
 		pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
-					   writing, &write_ok);
+					   writing, &write_ok, NULL);
 		if (is_error_noslot_pfn(pfn))
 			return -EFAULT;
 		page = NULL;
arch/powerpc/kvm/book3s_64_mmu_radix.c +1 -1
index bb35490400e9..e603de7ade52 100644
@@ -822,7 +822,7 @@ int kvmppc_book3s_instantiate_page(struct kvm_vcpu *vcpu,
 
 		/* Call KVM generic code to do the slow-path check */
 		pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
-					   writing, upgrade_p);
+					   writing, upgrade_p, NULL);
 		if (is_error_noslot_pfn(pfn))
 			return -EFAULT;
 		page = NULL;
arch/x86/kvm/mmu/mmu.c +10 -6
index 6d16481aa29d..79166288ed03 100644
@@ -3658,8 +3658,8 @@ static bool kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
 }
 
 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
-			 gpa_t cr2_or_gpa, kvm_pfn_t *pfn, bool write,
-			 bool *writable)
+			 gpa_t cr2_or_gpa, kvm_pfn_t *pfn, hva_t *hva,
+			 bool write, bool *writable)
 {
 	struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
 	bool async;
@@ -3672,7 +3672,8 @@ static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
 	}
 
 	async = false;
-	*pfn = __gfn_to_pfn_memslot(slot, gfn, false, &async, write, writable);
+	*pfn = __gfn_to_pfn_memslot(slot, gfn, false, &async,
+				    write, writable, hva);
 	if (!async)
 		return false; /* *pfn has correct page already */
 
@@ -3686,7 +3687,8 @@ static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
 			return true;
 	}
 
-	*pfn = __gfn_to_pfn_memslot(slot, gfn, false, NULL, write, writable);
+	*pfn = __gfn_to_pfn_memslot(slot, gfn, false, NULL,
+				    write, writable, hva);
 	return false;
 }
 
@@ -3699,6 +3701,7 @@ static int direct_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
 	gfn_t gfn = gpa >> PAGE_SHIFT;
 	unsigned long mmu_seq;
 	kvm_pfn_t pfn;
+	hva_t hva;
 	int r;
 
 	if (page_fault_handle_page_track(vcpu, error_code, gfn))
@@ -3717,7 +3720,8 @@ static int direct_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
 	mmu_seq = vcpu->kvm->mmu_notifier_seq;
 	smp_rmb();
 
-	if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, write, &map_writable))
+	if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, &hva,
+			 write, &map_writable))
 		return RET_PF_RETRY;
 
 	if (handle_abnormal_pfn(vcpu, is_tdp ? 0 : gpa, gfn, pfn, ACC_ALL, &r))
@@ -3725,7 +3729,7 @@ static int direct_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
 
 	r = RET_PF_RETRY;
 	spin_lock(&vcpu->kvm->mmu_lock);
-	if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
+	if (mmu_notifier_retry_hva(vcpu->kvm, mmu_seq, hva))
 		goto out_unlock;
 	r = make_mmu_pages_available(vcpu);
 	if (r)
arch/x86/kvm/mmu/paging_tmpl.h +4 -3
index 50e268eb8e1a..3171784139a4 100644
@@ -790,6 +790,7 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gpa_t addr, u32 error_code,
 	struct guest_walker walker;
 	int r;
 	kvm_pfn_t pfn;
+	hva_t hva;
 	unsigned long mmu_seq;
 	bool map_writable, is_self_change_mapping;
 	int max_level;
@@ -840,8 +841,8 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gpa_t addr, u32 error_code,
 	mmu_seq = vcpu->kvm->mmu_notifier_seq;
 	smp_rmb();
 
-	if (try_async_pf(vcpu, prefault, walker.gfn, addr, &pfn, write_fault,
-			 &map_writable))
+	if (try_async_pf(vcpu, prefault, walker.gfn, addr, &pfn, &hva,
+			 write_fault, &map_writable))
 		return RET_PF_RETRY;
 
 	if (handle_abnormal_pfn(vcpu, addr, walker.gfn, pfn, walker.pte_access, &r))
@@ -869,7 +870,7 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gpa_t addr, u32 error_code,
 
 	r = RET_PF_RETRY;
 	spin_lock(&vcpu->kvm->mmu_lock);
-	if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
+	if (mmu_notifier_retry_hva(vcpu->kvm, mmu_seq, hva))
 		goto out_unlock;
 
 	kvm_mmu_audit(vcpu, AUDIT_PRE_PAGE_FAULT);
include/linux/kvm_host.h +26 -1
index f3b1013fb22c..08f22c958d64 100644
@@ -11,6 +11,7 @@
 #include <linux/signal.h>
 #include <linux/sched.h>
 #include <linux/bug.h>
+#include <linux/minmax.h>
 #include <linux/mm.h>
 #include <linux/mmu_notifier.h>
 #include <linux/preempt.h>
@@ -502,6 +503,8 @@ struct kvm {
 	struct mmu_notifier mmu_notifier;
 	unsigned long mmu_notifier_seq;
 	long mmu_notifier_count;
+	unsigned long mmu_notifier_range_start;
+	unsigned long mmu_notifier_range_end;
 #endif
 	long tlbs_dirty;
 	struct list_head devices;
@@ -729,7 +732,7 @@ kvm_pfn_t gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn);
 kvm_pfn_t gfn_to_pfn_memslot_atomic(struct kvm_memory_slot *slot, gfn_t gfn);
 kvm_pfn_t __gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn,
 			       bool atomic, bool *async, bool write_fault,
-			       bool *writable);
+			       bool *writable, hva_t *hva);
 
 void kvm_release_pfn_clean(kvm_pfn_t pfn);
 void kvm_release_pfn_dirty(kvm_pfn_t pfn);
@@ -1203,6 +1206,28 @@ static inline int mmu_notifier_retry(struct kvm *kvm, unsigned long mmu_seq)
 		return 1;
 	return 0;
 }
+
+static inline int mmu_notifier_retry_hva(struct kvm *kvm,
+					 unsigned long mmu_seq,
+					 unsigned long hva)
+{
+#ifdef CONFIG_LOCKDEP
+	lockdep_is_held(&kvm->mmu_lock);
+#endif
+	/*
+	 * If mmu_notifier_count is non-zero, then the range maintained by
+	 * kvm_mmu_notifier_invalidate_range_start contains all addresses that
+	 * might be being invalidated. Note that it may include some false
+	 * positives, due to shortcuts when handing concurrent invalidations.
+	 */
+	if (unlikely(kvm->mmu_notifier_count) &&
+	    kvm->mmu_notifier_range_start <= hva &&
+	    hva < kvm->mmu_notifier_range_end)
+		return 1;
+	if (kvm->mmu_notifier_seq != mmu_seq)
+		return 1;
+	return 0;
+}
 #endif
 
 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
virt/kvm/kvm_main.c +23 -4
index fa9e3614d30e..b4b0a1eeb4c5 100644
@@ -483,6 +483,24 @@ static int kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
 	 * count is also read inside the mmu_lock critical section.
 	 */
 	kvm->mmu_notifier_count++;
+	if (likely(kvm->mmu_notifier_count = 1)) {
+		kvm->mmu_notifier_range_start = range->start;
+		kvm->mmu_notifier_range_end = range->end;
+	} else {
+		/*
+		 * Fully tracking multiple concurrent ranges has dimishing
+		 * returns. Keep things simple and just find the minimal range
+		 * which includes the current and new ranges. As there won't be
+		 * enough information to subtract a range after its invalidate
+		 * completes, any ranges invalidated concurrently will
+		 * accumulate and persist until all outstanding invalidates
+		 * complete.
+		 */
+		kvm->mmu_notifier_range_start +			min(kvm->mmu_notifier_range_start, range->start);
+		kvm->mmu_notifier_range_end +			max(kvm->mmu_notifier_range_end, range->end);
+	}
 	need_tlb_flush = kvm_unmap_hva_range(kvm, range->start, range->end,
 					     range->flags);
 	/* we've to flush the tlb before the pages can be freed */
@@ -2010,10 +2028,13 @@ static kvm_pfn_t hva_to_pfn(unsigned long addr, bool atomic, bool *async,
 
 kvm_pfn_t __gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn,
 			       bool atomic, bool *async, bool write_fault,
-			       bool *writable)
+			       bool *writable, hva_t *hva)
 {
 	unsigned long addr = __gfn_to_hva_many(slot, gfn, NULL, write_fault);
 
+	if (hva)
+		*hva = addr;
+
 	if (addr = KVM_HVA_ERR_RO_BAD) {
 		if (writable)
 			*writable = false;
@@ -2041,19 +2062,19 @@ kvm_pfn_t gfn_to_pfn_prot(struct kvm *kvm, gfn_t gfn, bool write_fault,
 		      bool *writable)
 {
 	return __gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn, false, NULL,
-				    write_fault, writable);
+				    write_fault, writable, NULL);
 }
 EXPORT_SYMBOL_GPL(gfn_to_pfn_prot);
 
 kvm_pfn_t gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn)
 {
-	return __gfn_to_pfn_memslot(slot, gfn, false, NULL, true, NULL);
+	return __gfn_to_pfn_memslot(slot, gfn, false, NULL, true, NULL, NULL);
 }
 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot);
 
 kvm_pfn_t gfn_to_pfn_memslot_atomic(struct kvm_memory_slot *slot, gfn_t gfn)
 {
-	return __gfn_to_pfn_memslot(slot, gfn, true, NULL, true, NULL);
+	return __gfn_to_pfn_memslot(slot, gfn, true, NULL, true, NULL, NULL);
 }
 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot_atomic);
 
Signature · 1 line
2.30.0.280.ga3ce27912f-goog
#1 lore raw
Quoted 14 lines · David Stevens

From: David Stevens <[email protected]>

Track the range being invalidated by mmu_notifier and skip page fault retries if the fault address is not affected by the in-progress invalidation. Handle concurrent invalidations by finding the minimal range which includes all ranges being invalidated. Although the combined range may include unrelated addresses and cannot be shrunk as individual invalidation operations complete, it is unlikely the marginal gains of proper range tracking are worth the additional complexity.

The primary benefit of this change is the reduction in the likelihood of extreme latency when handing a page fault due to another thread having been preempted while modifying host virtual addresses.

Quoted 5 lines

---
v1 -> v2:
- improve handling of concurrent invalidation requests by unioning ranges, instead of just giving up and using [0, ULONG_MAX).

Ooh, even better.

Quoted 11 lines

- add lockdep check
- code comments and formatting

 arch/powerpc/kvm/book3s_64_mmu_hv.c    |  2 +-
 arch/powerpc/kvm/book3s_64_mmu_radix.c |  2 +-
 arch/x86/kvm/mmu/mmu.c                 | 16 ++++++++------
 arch/x86/kvm/mmu/paging_tmpl.h         |  7 ++++---
 include/linux/kvm_host.h               | 27 +++++++++++++++++++++++-
 virt/kvm/kvm_main.c                    | 29 ++++++++++++++++++++++----
 6 files changed, 67 insertions(+), 16 deletions(-)

...

Quoted 11 lines
(quoted diff) +2 -1
@@ -3717,7 +3720,8 @@ static int direct_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
 	mmu_seq = vcpu->kvm->mmu_notifier_seq;
 	smp_rmb();
 
-	if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, write, &map_writable))
+	if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, &hva,
+			 write, &map_writable))
 		return RET_PF_RETRY;
 
 	if (handle_abnormal_pfn(vcpu, is_tdp ? 0 : gpa, gfn, pfn, ACC_ALL, &r))

Review static int direct_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code, · line 3732

Quoted 5 lines
(quoted diff) +1 -1
@@ -3725,7 +3729,7 @@ static int direct_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
 
 	r = RET_PF_RETRY;
 	spin_lock(&vcpu->kvm->mmu_lock);
-	if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
+	if (mmu_notifier_retry_hva(vcpu->kvm, mmu_seq, hva))

'hva' will be uninitialized at this point if the gfn did not resolve to a memslot, i.e. when handling an MMIO page fault. On the plus side, that's an opportunity for another optimization as there is no need to retry MMIO page faults on mmu_notifier invalidations. Including the attached patch as a preqreq to this will avoid consuming an uninitialized 'hva'.

Quoted 3 lines
(quoted diff) +0 -0
 		goto out_unlock;
 	r = make_mmu_pages_available(vcpu);
 	if (r)

...

Review static inline int mmu_notifier_retry(struct kvm *kvm, unsigned long mmu_seq) · line 1215

Quoted 13 lines

void kvm_release_pfn_clean(kvm_pfn_t pfn);
void kvm_release_pfn_dirty(kvm_pfn_t pfn);

(quoted diff) +7 -0
@@ -1203,6 +1206,28 @@ static inline int mmu_notifier_retry(struct kvm *kvm, unsigned long mmu_seq)
 		return 1;
 	return 0;
 }
+
+static inline int mmu_notifier_retry_hva(struct kvm *kvm,
+					 unsigned long mmu_seq,
+					 unsigned long hva)
+{
+#ifdef CONFIG_LOCKDEP
+	lockdep_is_held(&kvm->mmu_lock);

No need to manually do the #ifdef, just use lockdep_assert_held instead of lockdep_is_held.

Review static inline int mmu_notifier_retry(struct kvm *kvm, unsigned long mmu_seq) · line 1225

Quoted 10 lines
(quoted diff) +10 -0
+#endif
+	/*
+	 * If mmu_notifier_count is non-zero, then the range maintained by
+	 * kvm_mmu_notifier_invalidate_range_start contains all addresses that
+	 * might be being invalidated. Note that it may include some false
+	 * positives, due to shortcuts when handing concurrent invalidations.
+	 */
+	if (unlikely(kvm->mmu_notifier_count) &&
+	    kvm->mmu_notifier_range_start <= hva &&
+	    hva < kvm->mmu_notifier_range_end)

Uber nit: I find this easier to read if 'hva' is on the left-hand side for both checks, i.e.

	if (unlikely(kvm->mmu_notifier_count) &&
	    hva >= kvm->mmu_notifier_range_start &&
	    hva < kvm->mmu_notifier_range_end)

Review static inline int mmu_notifier_retry(struct kvm *kvm, unsigned long mmu_seq) · line 1233

Quoted 8 lines
(quoted diff) +5 -0
+		return 1;
+	if (kvm->mmu_notifier_seq != mmu_seq)
+		return 1;
+	return 0;
+}
 #endif
 
 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING

From a1bfdc6fe16582440815cfecc656313dff993003 Mon Sep 17 00:00:00 2001
From: Sean Christopherson <[email protected]>
Date: Wed, 27 Jan 2021 10:04:45 -0800
Subject: [PATCH] KVM: x86/mmu: Skip mmu_notifier check when handling MMIO page
fault

Don't retry a page fault due to an mmu_notifier invalidation when handling a page fault for a GPA that did not resolve to a memslot, i.e. an MMIO page fault. Invalidations from the mmu_notifier signal a change in a host virtual address (HVA) mapping; without a memslot, there is no HVA and thus no possibility that the invalidation is relevant to the page fault being handled.

Note, the MMIO vs. memslot generation checks handle the case where a pending memslot will create a memslot overlapping the faulting GPA. The mmu_notifier checks are orthogonal to memslot updates.

---
 arch/x86/kvm/mmu/mmu.c         | 2 +-
 arch/x86/kvm/mmu/paging_tmpl.h | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
arch/x86/kvm/mmu/mmu.c +1 -1
index 6d16481aa29d..9ac0a727015d 100644
@@ -3725,7 +3725,7 @@ static int direct_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
 
 	r = RET_PF_RETRY;
 	spin_lock(&vcpu->kvm->mmu_lock);
-	if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
+	if (!is_noslot_pfn(pfn) && mmu_notifier_retry(vcpu->kvm, mmu_seq))
 		goto out_unlock;
 	r = make_mmu_pages_available(vcpu);
 	if (r)
arch/x86/kvm/mmu/paging_tmpl.h +1 -1
index 50e268eb8e1a..ab54263d857c 100644
@@ -869,7 +869,7 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gpa_t addr, u32 error_code,
 
 	r = RET_PF_RETRY;
 	spin_lock(&vcpu->kvm->mmu_lock);
-	if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
+	if (!is_noslot_pfn(pfn) && mmu_notifier_retry(vcpu->kvm, mmu_seq))
 		goto out_unlock;
 
 	kvm_mmu_audit(vcpu, AUDIT_PRE_PAGE_FAULT);
Signature · 1 line
2.30.0.280.ga3ce27912f-goog