Skip to contentSkip to conversation

Subject: [RFC PATCH] kvm/x86: Fix 'lpages' kvm stat for TDM MMU

PATCH Applied Shahin, Md Shahadat Hossain 4 participants 5 messages Apr 29, 2021

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

Large pages not being created properly may result in increased memory access time. The 'lpages' kvm stat used to keep track of the current number of large pages in the system, but with TDP MMU enabled the stat is not showing the correct number.

This patch extends the lpages counter to cover the TDP case.

---
 arch/x86/kvm/mmu/tdp_mmu.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/mmu/tdp_mmu.c +6 -0
index 34207b874886..1e2a3cb33568 100644
@@ -425,6 +425,12 @@ static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
 
 	if (old_spte == new_spte)
 		return;
+	
+	if (is_large_pte(old_spte))
+		--kvm->stat.lpages;
+	
+	if (is_large_pte(new_spte))
+		++kvm->stat.lpages;
 
 	trace_kvm_tdp_mmu_spte_changed(as_id, gfn, level, old_spte, new_spte);
 
Signature · 11 lines
2.17.1



Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss
Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
Sitz: Berlin
Ust-ID: DE 289 237 879
maintainer Applied
#1 lore raw
Quoted 21 lines · Md Shahadat Hossain

Large pages not being created properly may result in increased memory access time. The 'lpages' kvm stat used to keep track of the current number of large pages in the system, but with TDP MMU enabled the stat is not showing the correct number.

This patch extends the lpages counter to cover the TDP case.

---
  arch/x86/kvm/mmu/tdp_mmu.c | 6 ++++++
  1 file changed, 6 insertions(+)
arch/x86/kvm/mmu/tdp_mmu.c +0 -0
index 34207b874886..1e2a3cb33568 100644
@@ -425,6 +425,12 @@ static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
 
  	if (old_spte == new_spte)

arch/x86/kvm/mmu/tdp_mmu.c static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn, · line 436

Quoted 10 lines
arch/x86/kvm/mmu/tdp_mmu.c +6 -0
  		return;
+	
+	if (is_large_pte(old_spte))
+		--kvm->stat.lpages;
+	
+	if (is_large_pte(new_spte))
+		++kvm->stat.lpages;
 
  	trace_kvm_tdp_mmu_spte_changed(as_id, gfn, level, old_spte, new_spte);
 

Queued, thanks.

Paolo

#2 lore raw
In reply to Shahin, Md Shahadat Hossain
Quoted 18 lines · Md Shahadat Hossain

Large pages not being created properly may result in increased memory access time. The 'lpages' kvm stat used to keep track of the current number of large pages in the system, but with TDP MMU enabled the stat is not showing the correct number.

This patch extends the lpages counter to cover the TDP case.

---
 arch/x86/kvm/mmu/tdp_mmu.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kvm/mmu/tdp_mmu.c +0 -0
index 34207b874886..1e2a3cb33568 100644

arch/x86/kvm/mmu/tdp_mmu.c static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn, · line 433

Quoted 9 lines
arch/x86/kvm/mmu/tdp_mmu.c +6 -0
@@ -425,6 +425,12 @@ static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
 
 	if (old_spte == new_spte)
 		return;
+	
+	if (is_large_pte(old_spte))
+		--kvm->stat.lpages;
+	
+	if (is_large_pte(new_spte))
+		++kvm->stat.lpages;

Hrm, kvm->stat.lpages could get corrupted when __handle_changed_spte() is called under read lock, e.g. if multiple vCPUs are faulting in memory.

maintainer
#3 lore raw
Quoted 20 lines · Sean Christopherson
Quoted 19 lines · Md Shahadat Hossain

Large pages not being created properly may result in increased memory access time. The 'lpages' kvm stat used to keep track of the current number of large pages in the system, but with TDP MMU enabled the stat is not showing the correct number.

This patch extends the lpages counter to cover the TDP case.

---
  arch/x86/kvm/mmu/tdp_mmu.c | 6 ++++++
  1 file changed, 6 insertions(+)
arch/x86/kvm/mmu/tdp_mmu.c +0 -0
index 34207b874886..1e2a3cb33568 100644
@@ -425,6 +425,12 @@ static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
 
Quoted 11 lines

arch/x86/kvm/mmu/tdp_mmu.c static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn, · line 433

Quoted 8 lines
arch/x86/kvm/mmu/tdp_mmu.c +6 -0
  	if (old_spte == new_spte)
  		return;
+	
+	if (is_large_pte(old_spte))
+		--kvm->stat.lpages;
+	
+	if (is_large_pte(new_spte))
+		++kvm->stat.lpages;

Hrm, kvm->stat.lpages could get corrupted when __handle_changed_spte() is called under read lock, e.g. if multiple vCPUs are faulting in memory.

Ouch, indeed!

One way to fix it without needing an atomic operation is to make it a per-vcpu stat. It would be a bit weird for the binary stats because we would have to hide this one from the vCPU statistics file descriptor (and only aggregate it in the VM statistics).

Alternatively, you can do the atomic_add only if is_large_pte(old_spte) != is_large_pte(new_spte), casting &kvm->stat.lpages to an atomic64_t*.

Paolo

#4 lore raw
Quoted 36 lines · Paolo Bonzini
Quoted 20 lines · Sean Christopherson
Quoted 19 lines · Md Shahadat Hossain

Large pages not being created properly may result in increased memory access time. The 'lpages' kvm stat used to keep track of the current number of large pages in the system, but with TDP MMU enabled the stat is not showing the correct number.

This patch extends the lpages counter to cover the TDP case.

---
  arch/x86/kvm/mmu/tdp_mmu.c | 6 ++++++
  1 file changed, 6 insertions(+)
arch/x86/kvm/mmu/tdp_mmu.c +0 -0
index 34207b874886..1e2a3cb33568 100644
@@ -425,6 +425,12 @@ static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
 
Quoted 11 lines

arch/x86/kvm/mmu/tdp_mmu.c static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn, · line 433

Quoted 8 lines
arch/x86/kvm/mmu/tdp_mmu.c +6 -0
     if (old_spte == new_spte)
             return;
+
+    if (is_large_pte(old_spte))
+            --kvm->stat.lpages;
+
+    if (is_large_pte(new_spte))
+            ++kvm->stat.lpages;

Hrm, kvm->stat.lpages could get corrupted when __handle_changed_spte() is called under read lock, e.g. if multiple vCPUs are faulting in memory.

Ouch, indeed!

Quoted 7 lines

One way to fix it without needing an atomic operation is to make it a per-vcpu stat. It would be a bit weird for the binary stats because we would have to hide this one from the vCPU statistics file descriptor (and only aggregate it in the VM statistics).

Alternatively, you can do the atomic_add only if is_large_pte(old_spte) != is_large_pte(new_spte), casting &kvm->stat.lpages to an atomic64_t*.

I forgot that the lpages stat existed upstream. Internally at Google we also maintain a count for each mapping level and just update it atomically. The per-vCPU stat approach would work too, but I doubt it's worth the complexity unless we have other use cases for a similarly aggregated stat.

Quoted 3 lines

Paolo