Skip to contentSkip to conversation

ARM64: Guest performance improvement during dirty

RFC PATCH In review Jing Zhang 4 participants 20 messages Jan 10, 2022 – Jan 13, 2022

Subsystems (4)
lore
Share & tools
lore

Jing Zhang

author
#0 lore raw

cover ARM64: Guest performance improvement during dirty

This patch is to reduce the performance degradation of guest workload during dirty logging on ARM64. A fast path is added to handle permission relaxation during dirty logging. The MMU lock is replaced with rwlock, by which all permision relaxations on leaf pte can be performed under the read lock. This greatly reduces the MMU lock contention during dirty logging. With this solution, the source guest workload performance degradation can be improved by more than 60%.

Problem:
* A Google internal live migration test shows that the source guest workload performance has >99% degradation for about 105 seconds, >50% degradation for about 112 seconds, >10% degradation for about 112 seconds on ARM64. This shows that most of the time, the guest workload degradtion is above 99%, which obviously needs some improvement compared to the test result on x86 (>99% for 6s, >50% for 9s, >10% for 27s).
* Tested H/W: Ampere Altra 3GHz, #CPU: 64, #Mem: 256GB
* VM spec: #vCPU: 48, #Mem/vCPU: 4GB

Analysis:
* We enabled CONFIG_LOCK_STAT in kernel and used dirty_log_perf_test to get

    the number of contentions of MMU lock and the "dirty memory time" on
    various VM spec.
    By using test command
    ./dirty_log_perf_test -b 2G -m 2 -i 2 -s anonymous_hugetlb_2mb -v [#vCPU]
    Below are the results:
    +-------+------------------------+-----------------------+
    | #vCPU | dirty memory time (ms) | number of contentions |
    +-------+------------------------+-----------------------+
    | 1     | 926                    | 0                     |
    +-------+------------------------+-----------------------+
    | 2     | 1189                   | 4732558               |
    +-------+------------------------+-----------------------+
    | 4     | 2503                   | 11527185              |
    +-------+------------------------+-----------------------+
    | 8     | 5069                   | 24881677              |
    +-------+------------------------+-----------------------+
    | 16    | 10340                  | 50347956              |
    +-------+------------------------+-----------------------+
    | 32    | 20351                  | 100605720             |
    +-------+------------------------+-----------------------+
    | 64    | 40994                  | 201442478             |
    +-------+------------------------+-----------------------+

* From the test results above, the "dirty memory time" and the number of MMU lock contention scale with the number of vCPUs. That means all the dirty memory operations from all vCPU threads have been serialized by the MMU lock. Further analysis also shows that the permission relaxation during dirty logging is where vCPU threads get serialized.

Solution:
* On ARM64, there is no mechanism as PML (Page Modification Logging) and the dirty-bit solution for dirty logging is much complicated compared to the write-protection solution. The straight way to reduce the guest performance degradation is to enhance the concurrency for the permission fault path during dirty logging.
* In this patch, we only put leaf PTE permission relaxation for dirty

    logging under read lock, all others would go under write lock.
    Below are the results based on the solution:
    +-------+------------------------+
    | #vCPU | dirty memory time (ms) |
    +-------+------------------------+
    | 1     | 803                    |
    +-------+------------------------+
    | 2     | 843                    |
    +-------+------------------------+
    | 4     | 942                    |
    +-------+------------------------+
    | 8     | 1458                   |
    +-------+------------------------+
    | 16    | 2853                   |
    +-------+------------------------+
    | 32    | 5886                   |
    +-------+------------------------+
    | 64    | 12190                  |
    +-------+------------------------+
    All "dirty memory time" have been reduced by more than 60% when the
    number of vCPU grows.

---

Jing Zhang (3):
KVM: arm64: Use read/write spin lock for MMU protection
KVM: arm64: Add fast path to handle permission relaxation during dirty
logging
KVM: selftests: Add vgic initialization for dirty log perf test for
ARM

 arch/arm64/include/asm/kvm_host.h             |  2 +
 arch/arm64/kvm/mmu.c                          | 86 +++++++++++++++----
 .../selftests/kvm/dirty_log_perf_test.c       | 10 +++
 3 files changed, 80 insertions(+), 18 deletions(-)

base-commit: fea31d1690945e6dd6c3e89ec5591490857bc3d4

Signature · 1 line
2.34.1.575.g55b058a8bb-goog
author
#1 lore raw

RFC PATCH 1/3 KVM: arm64: Use read/write spin lock for MMU protection

To reduce the contentions caused by MMU lock, some MMU operations can be performed under read lock.
One improvement is to add a fast path for permission relaxation during dirty logging under the read lock.

---
 arch/arm64/include/asm/kvm_host.h |  2 ++
 arch/arm64/kvm/mmu.c              | 36 +++++++++++++++----------------
 2 files changed, 20 insertions(+), 18 deletions(-)
arch/arm64/include/asm/kvm_host.h +2 -0
index 3b44ea17af88..6c99c0335bae 100644
@@ -50,6 +50,8 @@
 #define KVM_DIRTY_LOG_MANUAL_CAPS   (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
 				     KVM_DIRTY_LOG_INITIALLY_SET)
 
+#define KVM_HAVE_MMU_RWLOCK
+
 /*
  * Mode of operation configurable with kvm-arm.mode early param.
  * See Documentation/admin-guide/kernel-parameters.txt for more information.
arch/arm64/kvm/mmu.c +18 -18
index bc2aba953299..cafd5813c949 100644
@@ -58,7 +58,7 @@ static int stage2_apply_range(struct kvm *kvm, phys_addr_t addr,
 			break;
 
 		if (resched && next != end)
-			cond_resched_lock(&kvm->mmu_lock);
+			cond_resched_rwlock_write(&kvm->mmu_lock);
 	} while (addr = next, addr != end);
 
 	return ret;
@@ -179,7 +179,7 @@ static void __unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64
 	struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
 	phys_addr_t end = start + size;
 
-	assert_spin_locked(&kvm->mmu_lock);
+	lockdep_assert_held_write(&kvm->mmu_lock);
 	WARN_ON(size & ~PAGE_MASK);
 	WARN_ON(stage2_apply_range(kvm, start, end, kvm_pgtable_stage2_unmap,
 				   may_block));
@@ -213,13 +213,13 @@ static void stage2_flush_vm(struct kvm *kvm)
 	int idx, bkt;
 
 	idx = srcu_read_lock(&kvm->srcu);
-	spin_lock(&kvm->mmu_lock);
+	write_lock(&kvm->mmu_lock);
 
 	slots = kvm_memslots(kvm);
 	kvm_for_each_memslot(memslot, bkt, slots)
 		stage2_flush_memslot(kvm, memslot);
 
-	spin_unlock(&kvm->mmu_lock);
+	write_unlock(&kvm->mmu_lock);
 	srcu_read_unlock(&kvm->srcu, idx);
 }
 
@@ -720,13 +720,13 @@ void stage2_unmap_vm(struct kvm *kvm)
 
 	idx = srcu_read_lock(&kvm->srcu);
 	mmap_read_lock(current->mm);
-	spin_lock(&kvm->mmu_lock);
+	write_lock(&kvm->mmu_lock);
 
 	slots = kvm_memslots(kvm);
 	kvm_for_each_memslot(memslot, bkt, slots)
 		stage2_unmap_memslot(kvm, memslot);
 
-	spin_unlock(&kvm->mmu_lock);
+	write_unlock(&kvm->mmu_lock);
 	mmap_read_unlock(current->mm);
 	srcu_read_unlock(&kvm->srcu, idx);
 }
@@ -736,14 +736,14 @@ void kvm_free_stage2_pgd(struct kvm_s2_mmu *mmu)
 	struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
 	struct kvm_pgtable *pgt = NULL;
 
-	spin_lock(&kvm->mmu_lock);
+	write_lock(&kvm->mmu_lock);
 	pgt = mmu->pgt;
 	if (pgt) {
 		mmu->pgd_phys = 0;
 		mmu->pgt = NULL;
 		free_percpu(mmu->last_vcpu_ran);
 	}
-	spin_unlock(&kvm->mmu_lock);
+	write_unlock(&kvm->mmu_lock);
 
 	if (pgt) {
 		kvm_pgtable_stage2_destroy(pgt);
@@ -783,10 +783,10 @@ int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
 		if (ret)
 			break;
 
-		spin_lock(&kvm->mmu_lock);
+		write_lock(&kvm->mmu_lock);
 		ret = kvm_pgtable_stage2_map(pgt, addr, PAGE_SIZE, pa, prot,
 					     &cache);
-		spin_unlock(&kvm->mmu_lock);
+		write_unlock(&kvm->mmu_lock);
 		if (ret)
 			break;
 
@@ -834,9 +834,9 @@ static void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
 	start = memslot->base_gfn << PAGE_SHIFT;
 	end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
 
-	spin_lock(&kvm->mmu_lock);
+	write_lock(&kvm->mmu_lock);
 	stage2_wp_range(&kvm->arch.mmu, start, end);
-	spin_unlock(&kvm->mmu_lock);
+	write_unlock(&kvm->mmu_lock);
 	kvm_flush_remote_tlbs(kvm);
 }
 
@@ -1212,7 +1212,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 	if (exec_fault && device)
 		return -ENOEXEC;
 
-	spin_lock(&kvm->mmu_lock);
+	write_lock(&kvm->mmu_lock);
 	pgt = vcpu->arch.hw_mmu->pgt;
 	if (mmu_notifier_retry(kvm, mmu_seq))
 		goto out_unlock;
@@ -1271,7 +1271,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 	}
 
 out_unlock:
-	spin_unlock(&kvm->mmu_lock);
+	write_unlock(&kvm->mmu_lock);
 	kvm_set_pfn_accessed(pfn);
 	kvm_release_pfn_clean(pfn);
 	return ret != -EAGAIN ? ret : 0;
@@ -1286,10 +1286,10 @@ static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
 
 	trace_kvm_access_fault(fault_ipa);
 
-	spin_lock(&vcpu->kvm->mmu_lock);
+	write_lock(&vcpu->kvm->mmu_lock);
 	mmu = vcpu->arch.hw_mmu;
 	kpte = kvm_pgtable_stage2_mkyoung(mmu->pgt, fault_ipa);
-	spin_unlock(&vcpu->kvm->mmu_lock);
+	write_unlock(&vcpu->kvm->mmu_lock);
 
 	pte = __pte(kpte);
 	if (pte_valid(pte))
@@ -1692,9 +1692,9 @@ void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
 	gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
 	phys_addr_t size = slot->npages << PAGE_SHIFT;
 
-	spin_lock(&kvm->mmu_lock);
+	write_lock(&kvm->mmu_lock);
 	unmap_stage2_range(&kvm->arch.mmu, gpa, size);
-	spin_unlock(&kvm->mmu_lock);
+	write_unlock(&kvm->mmu_lock);
 }
 
 /*
Signature · 1 line
2.34.1.575.g55b058a8bb-goog
maintainer
#2 lore raw

Re: [RFC PATCH 1/3] KVM: arm64: Use read/write spin lock for MMU protection

Quoted 5 lines · Jing Zhang

To reduce the contentions caused by MMU lock, some MMU operations can be performed under read lock.
One improvement is to add a fast path for permission relaxation during dirty logging under the read lock.

This commit message really doesn't say what this patch does (converting our MMU spinlock to a rwlock, and replacing all instances of the lock being acquired with a write lock acquisition). Crucially, it only mention the read lock which appears *nowhere* in this patch.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author
#3 lore raw

Re: [RFC PATCH 1/3] KVM: arm64: Use read/write spin lock for MMU protection

Quoted 9 lines · Marc Zyngier
Quoted 5 lines · Jing Zhang

To reduce the contentions caused by MMU lock, some MMU operations can be performed under read lock.
One improvement is to add a fast path for permission relaxation during dirty logging under the read lock.

Quoted 5 lines

This commit message really doesn't say what this patch does (converting our MMU spinlock to a rwlock, and replacing all instances of the lock being acquired with a write lock acquisition). Crucially, it only mention the read lock which appears *nowhere* in this patch.

Good point. Will use the below message instead for future posts. "Replace MMU spinlock with rwlock and update all instances of the lock being acquired with a write lock acquisition.
Future commit will add a fast path for permission relaxation during dirty logging under a read lock."

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Thanks,
Jing

author
#4 lore raw

RFC PATCH 2/3 KVM: arm64: Add fast path to handle permission relaxation during dirty logging

In reply to Jing Zhang

To reduce MMU lock contention during dirty logging, all permission relaxation operations would be performed under read lock.

---
 arch/arm64/kvm/mmu.c | 50 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 50 insertions(+)
arch/arm64/kvm/mmu.c +50 -0
index cafd5813c949..dd1f43fba4b0 100644
@@ -1063,6 +1063,54 @@ static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn,
 	return 0;
 }
 
+static bool fast_mark_writable(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
+		struct kvm_memory_slot *memslot, unsigned long fault_status)
+{
+	int ret;
+	bool writable;
+	bool write_fault = kvm_is_write_fault(vcpu);
+	gfn_t gfn = fault_ipa >> PAGE_SHIFT;
+	kvm_pfn_t pfn;
+	struct kvm *kvm = vcpu->kvm;
+	bool logging_active = memslot_is_logging(memslot);
+	unsigned long fault_level = kvm_vcpu_trap_get_fault_level(vcpu);
+	unsigned long fault_granule;
+
+	fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
+
+	/* Make sure the fault can be handled in the fast path.
+	 * Only handle write permission fault on non-hugepage during dirty
+	 * logging period.
+	 */
+	if (fault_status != FSC_PERM || fault_granule != PAGE_SIZE
+			|| !logging_active || !write_fault)
+		return false;
+
+
+	/* Pin the pfn to make sure it couldn't be freed and be resued for
+	 * another gfn.
+	 */
+	pfn = __gfn_to_pfn_memslot(memslot, gfn, true, NULL,
+				   write_fault, &writable, NULL);
+	if (is_error_pfn(pfn) || !writable)
+		return false;
+
+	read_lock(&kvm->mmu_lock);
+	ret = kvm_pgtable_stage2_relax_perms(
+			vcpu->arch.hw_mmu->pgt, fault_ipa, PAGE_HYP);
+
+	if (!ret) {
+		kvm_set_pfn_dirty(pfn);
+		mark_page_dirty_in_slot(kvm, memslot, gfn);
+	}
+	read_unlock(&kvm->mmu_lock);
+
+	kvm_set_pfn_accessed(pfn);
+	kvm_release_pfn_clean(pfn);
+
+	return true;
+}
+
 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 			  struct kvm_memory_slot *memslot, unsigned long hva,
 			  unsigned long fault_status)
@@ -1085,6 +1133,8 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 	enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
 	struct kvm_pgtable *pgt;
 
+	if (fast_mark_writable(vcpu, fault_ipa, memslot, fault_status))
+		return 0;
 	fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
 	write_fault = kvm_is_write_fault(vcpu);
 	exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
Signature · 1 line
2.34.1.575.g55b058a8bb-goog
maintainer
#5 lore raw

Re: [RFC PATCH 2/3] KVM: arm64: Add fast path to handle permission relaxation during dirty logging

Quoted 26 lines · Jing Zhang

To reduce MMU lock contention during dirty logging, all permission relaxation operations would be performed under read lock.

---
 arch/arm64/kvm/mmu.c | 50 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 50 insertions(+)
arch/arm64/kvm/mmu.c +9 -0
index cafd5813c949..dd1f43fba4b0 100644
@@ -1063,6 +1063,54 @@ static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn,
 	return 0;
 }
 
+static bool fast_mark_writable(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
+		struct kvm_memory_slot *memslot, unsigned long fault_status)
+{
+	int ret;
+	bool writable;
+	bool write_fault = kvm_is_write_fault(vcpu);
+	gfn_t gfn = fault_ipa >> PAGE_SHIFT;
+	kvm_pfn_t pfn;
+	struct kvm *kvm = vcpu->kvm;

arch/arm64/kvm/mmu.c static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn, · line 1084

Quoted 10 lines
arch/arm64/kvm/mmu.c +10 -0
+	bool logging_active = memslot_is_logging(memslot);
+	unsigned long fault_level = kvm_vcpu_trap_get_fault_level(vcpu);
+	unsigned long fault_granule;
+
+	fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
+
+	/* Make sure the fault can be handled in the fast path.
+	 * Only handle write permission fault on non-hugepage during dirty
+	 * logging period.
+	 */

Not the correct comment format.

arch/arm64/kvm/mmu.c static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn, · line 1087

Quoted 3 lines
arch/arm64/kvm/mmu.c +3 -0
+	if (fault_status != FSC_PERM || fault_granule != PAGE_SIZE
+			|| !logging_active || !write_fault)
+		return false;

This is all reinventing the logic that already exists in user_mem_abort(). I'm sympathetic to the effort not to bloat it even more, but code duplication doesn't help either.

arch/arm64/kvm/mmu.c static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn, · line 1096

Quoted 9 lines
arch/arm64/kvm/mmu.c +9 -0
+
+
+	/* Pin the pfn to make sure it couldn't be freed and be resued for
+	 * another gfn.
+	 */
+	pfn = __gfn_to_pfn_memslot(memslot, gfn, true, NULL,
+				   write_fault, &writable, NULL);
+	if (is_error_pfn(pfn) || !writable)
+		return false;

What happens if we hit a non-writable mapping? Don't we leak a page reference?

arch/arm64/kvm/mmu.c static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn, · line 1100

Quoted 4 lines
arch/arm64/kvm/mmu.c +4 -0
+
+	read_lock(&kvm->mmu_lock);
+	ret = kvm_pgtable_stage2_relax_perms(
+			vcpu->arch.hw_mmu->pgt, fault_ipa, PAGE_HYP);

PAGE_HYP? Err... no. KVM_PGTABLE_PROT_RW, more likely. Yes, they expand to the same thing, but you are not dealing with nVHE EL2 S1 page tables here.

Quoted 17 lines
arch/arm64/kvm/mmu.c +13 -0
+
+	if (!ret) {
+		kvm_set_pfn_dirty(pfn);
+		mark_page_dirty_in_slot(kvm, memslot, gfn);
+	}
+	read_unlock(&kvm->mmu_lock);
+
+	kvm_set_pfn_accessed(pfn);
+	kvm_release_pfn_clean(pfn);
+
+	return true;
+}
+
 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 			  struct kvm_memory_slot *memslot, unsigned long hva,
 			  unsigned long fault_status)

arch/arm64/kvm/mmu.c static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, · line 1140

Quoted 8 lines
arch/arm64/kvm/mmu.c +2 -0
@@ -1085,6 +1133,8 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 	enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
 	struct kvm_pgtable *pgt;
 
+	if (fast_mark_writable(vcpu, fault_ipa, memslot, fault_status))
+		return 0;
 	fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
 	write_fault = kvm_is_write_fault(vcpu);
 	exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);

You are bypassing all sort of checks that I want to keep. Please integrate this in user_mem_abort instead of this side hack.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
maintainer
#6 lore raw

Re: [RFC PATCH 2/3] KVM: arm64: Add fast path to handle permission relaxation during dirty logging

In reply to Jing Zhang

Coming back to this, as it does bother me.

Quoted 36 lines · Jing Zhang

To reduce MMU lock contention during dirty logging, all permission relaxation operations would be performed under read lock.

---
 arch/arm64/kvm/mmu.c | 50 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 50 insertions(+)
arch/arm64/kvm/mmu.c +19 -0
index cafd5813c949..dd1f43fba4b0 100644
@@ -1063,6 +1063,54 @@ static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn,
 	return 0;
 }
 
+static bool fast_mark_writable(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
+		struct kvm_memory_slot *memslot, unsigned long fault_status)
+{
+	int ret;
+	bool writable;
+	bool write_fault = kvm_is_write_fault(vcpu);
+	gfn_t gfn = fault_ipa >> PAGE_SHIFT;
+	kvm_pfn_t pfn;
+	struct kvm *kvm = vcpu->kvm;
+	bool logging_active = memslot_is_logging(memslot);
+	unsigned long fault_level = kvm_vcpu_trap_get_fault_level(vcpu);
+	unsigned long fault_granule;
+
+	fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
+
+	/* Make sure the fault can be handled in the fast path.
+	 * Only handle write permission fault on non-hugepage during dirty
+	 * logging period.
+	 */

arch/arm64/kvm/mmu.c static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn, · line 1094

Quoted 10 lines
arch/arm64/kvm/mmu.c +10 -0
+	if (fault_status != FSC_PERM || fault_granule != PAGE_SIZE
+			|| !logging_active || !write_fault)
+		return false;
+
+
+	/* Pin the pfn to make sure it couldn't be freed and be resued for
+	 * another gfn.
+	 */
+	pfn = __gfn_to_pfn_memslot(memslot, gfn, true, NULL,
+				   write_fault, &writable, NULL);

Why the requirement to be atomic? Once this returns, the page will have an elevated refcount, atomic or not. Given that we're not in an environment that requires atomicity (we're fully preemptible at this stage), I wonder what this is achieving.

arch/arm64/kvm/mmu.c static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn, · line 1098

Quoted 4 lines
arch/arm64/kvm/mmu.c +4 -0
+	if (is_error_pfn(pfn) || !writable)
+		return false;
+
+	read_lock(&kvm->mmu_lock);

You also dropped the hazarding against a concurrent MMU notifier. Why is it safe to do so?

Quoted 19 lines
arch/arm64/kvm/mmu.c +15 -0
+	ret = kvm_pgtable_stage2_relax_perms(
+			vcpu->arch.hw_mmu->pgt, fault_ipa, PAGE_HYP);
+
+	if (!ret) {
+		kvm_set_pfn_dirty(pfn);
+		mark_page_dirty_in_slot(kvm, memslot, gfn);
+	}
+	read_unlock(&kvm->mmu_lock);
+
+	kvm_set_pfn_accessed(pfn);
+	kvm_release_pfn_clean(pfn);
+
+	return true;
+}
+
 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 			  struct kvm_memory_slot *memslot, unsigned long hva,
 			  unsigned long fault_status)

arch/arm64/kvm/mmu.c static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, · line 1140

Quoted 8 lines
arch/arm64/kvm/mmu.c +2 -0
@@ -1085,6 +1133,8 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 	enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
 	struct kvm_pgtable *pgt;
 
+	if (fast_mark_writable(vcpu, fault_ipa, memslot, fault_status))
+		return 0;
 	fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
 	write_fault = kvm_is_write_fault(vcpu);
 	exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author
#7 lore raw

Re: [RFC PATCH 2/3] KVM: arm64: Add fast path to handle permission relaxation during dirty logging

Hi Marc,

Quoted 95 lines · Marc Zyngier

Coming back to this, as it does bother me.

Quoted 36 lines · Jing Zhang

To reduce MMU lock contention during dirty logging, all permission relaxation operations would be performed under read lock.

---
 arch/arm64/kvm/mmu.c | 50 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 50 insertions(+)
arch/arm64/kvm/mmu.c +19 -0
index cafd5813c949..dd1f43fba4b0 100644
@@ -1063,6 +1063,54 @@ static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn,
      return 0;
 }
 
+static bool fast_mark_writable(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
+             struct kvm_memory_slot *memslot, unsigned long fault_status)
+{
+     int ret;
+     bool writable;
+     bool write_fault = kvm_is_write_fault(vcpu);
+     gfn_t gfn = fault_ipa >> PAGE_SHIFT;
+     kvm_pfn_t pfn;
+     struct kvm *kvm = vcpu->kvm;
+     bool logging_active = memslot_is_logging(memslot);
+     unsigned long fault_level = kvm_vcpu_trap_get_fault_level(vcpu);
+     unsigned long fault_granule;
+
+     fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
+
+     /* Make sure the fault can be handled in the fast path.
+      * Only handle write permission fault on non-hugepage during dirty
+      * logging period.
+      */

arch/arm64/kvm/mmu.c static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn, · line 1094

Quoted 10 lines
arch/arm64/kvm/mmu.c +10 -0
+     if (fault_status != FSC_PERM || fault_granule != PAGE_SIZE
+                     || !logging_active || !write_fault)
+             return false;
+
+
+     /* Pin the pfn to make sure it couldn't be freed and be resued for
+      * another gfn.
+      */
+     pfn = __gfn_to_pfn_memslot(memslot, gfn, true, NULL,
+                                write_fault, &writable, NULL);

Why the requirement to be atomic? Once this returns, the page will have an elevated refcount, atomic or not. Given that we're not in an environment that requires atomicity (we're fully preemptible at this stage), I wonder what this is achieving.

arch/arm64/kvm/mmu.c static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn, · line 1098

Quoted 4 lines
arch/arm64/kvm/mmu.c +4 -0
+     if (is_error_pfn(pfn) || !writable)
+             return false;
+
+     read_lock(&kvm->mmu_lock);

You also dropped the hazarding against a concurrent MMU notifier. Why is it safe to do so?

Quoted 19 lines
arch/arm64/kvm/mmu.c +15 -0
+     ret = kvm_pgtable_stage2_relax_perms(
+                     vcpu->arch.hw_mmu->pgt, fault_ipa, PAGE_HYP);
+
+     if (!ret) {
+             kvm_set_pfn_dirty(pfn);
+             mark_page_dirty_in_slot(kvm, memslot, gfn);
+     }
+     read_unlock(&kvm->mmu_lock);
+
+     kvm_set_pfn_accessed(pfn);
+     kvm_release_pfn_clean(pfn);
+
+     return true;
+}
+
 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
                        struct kvm_memory_slot *memslot, unsigned long hva,
                        unsigned long fault_status)

arch/arm64/kvm/mmu.c static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, · line 1140

Quoted 8 lines
arch/arm64/kvm/mmu.c +2 -0
@@ -1085,6 +1133,8 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
      enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
      struct kvm_pgtable *pgt;
 
+     if (fast_mark_writable(vcpu, fault_ipa, memslot, fault_status))
+             return 0;
      fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
      write_fault = kvm_is_write_fault(vcpu);
      exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);

Thanks,

Quoted 4 lines
        M.

--
Without deviation from the norm, progress is not possible.

Appreciate all the comments here. I'll refactor the patch to implement the fast path in user_mem_abort and address all the problems you mentioned.
Thanks,
Jing
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm

author
#8 lore raw

RFC PATCH 3/3 KVM: selftests: Add vgic initialization for dirty log perf test for ARM

In reply to Jing Zhang

For ARM64, if no vgic is setup before the dirty log perf test, the userspace irqchip would be used, which would affect the dirty log perf test result.

---
 tools/testing/selftests/kvm/dirty_log_perf_test.c | 10 ++++++++++
 1 file changed, 10 insertions(+)
tools/testing/selftests/kvm/dirty_log_perf_test.c +10 -0
index 1954b964d1cf..b501338d9430 100644
@@ -18,6 +18,12 @@
 #include "test_util.h"
 #include "perf_test_util.h"
 #include "guest_modes.h"
+#ifdef __aarch64__
+#include "aarch64/vgic.h"
+
+#define GICD_BASE_GPA			0x8000000ULL
+#define GICR_BASE_GPA			0x80A0000ULL
+#endif
 
 /* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/
 #define TEST_HOST_LOOP_N		2UL
@@ -200,6 +206,10 @@ static void run_test(enum vm_guest_mode mode, void *arg)
 		vm_enable_cap(vm, &cap);
 	}
 
+#ifdef __aarch64__
+	vgic_v3_setup(vm, nr_vcpus, 64, GICD_BASE_GPA, GICR_BASE_GPA);
+#endif
+
 	/* Start the iterations */
 	iteration = 0;
 	host_quit = false;
Signature · 1 line
2.34.1.575.g55b058a8bb-goog
#9 lore raw

Re: [RFC PATCH 3/3] KVM: selftests: Add vgic initialization for dirty log perf test for ARM

Quoted 27 lines · Jing Zhang

For ARM64, if no vgic is setup before the dirty log perf test, the userspace irqchip would be used, which would affect the dirty log perf test result.

---
 tools/testing/selftests/kvm/dirty_log_perf_test.c | 10 ++++++++++
 1 file changed, 10 insertions(+)
tools/testing/selftests/kvm/dirty_log_perf_test.c +6 -0
index 1954b964d1cf..b501338d9430 100644
@@ -18,6 +18,12 @@
 #include "test_util.h"
 #include "perf_test_util.h"
 #include "guest_modes.h"
+#ifdef __aarch64__
+#include "aarch64/vgic.h"
+
+#define GICD_BASE_GPA			0x8000000ULL
+#define GICR_BASE_GPA			0x80A0000ULL
+#endif
 
 /* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/
 #define TEST_HOST_LOOP_N		2UL

tools/testing/selftests/kvm/dirty_log_perf_test.c static void run_test(enum vm_guest_mode mode, void *arg) · line 210

Quoted 5 lines
tools/testing/selftests/kvm/dirty_log_perf_test.c +2 -0
@@ -200,6 +206,10 @@ static void run_test(enum vm_guest_mode mode, void *arg)
 		vm_enable_cap(vm, &cap);
 	}
 
+#ifdef __aarch64__
+	vgic_v3_setup(vm, nr_vcpus, 64, GICD_BASE_GPA, GICR_BASE_GPA);

^^ extra parameter

Thanks,
drew

Quoted 13 lines
tools/testing/selftests/kvm/dirty_log_perf_test.c +2 -0
+#endif
+
 	/* Start the iterations */
 	iteration = 0;
 	host_quit = false;
Signature · 6 lines
2.34.1.575.g55b058a8bb-goog

_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
author
#10 lore raw

Re: [RFC PATCH 3/3] KVM: selftests: Add vgic initialization for dirty log perf test for ARM

Quoted 32 lines · Andrew Jones
Quoted 30 lines · Jing Zhang

For ARM64, if no vgic is setup before the dirty log perf test, the userspace irqchip would be used, which would affect the dirty log perf test result.

---
 tools/testing/selftests/kvm/dirty_log_perf_test.c | 10 ++++++++++
 1 file changed, 10 insertions(+)
tools/testing/selftests/kvm/dirty_log_perf_test.c +6 -0
index 1954b964d1cf..b501338d9430 100644
@@ -18,6 +18,12 @@
 #include "test_util.h"
 #include "perf_test_util.h"
 #include "guest_modes.h"
+#ifdef __aarch64__
+#include "aarch64/vgic.h"
+
+#define GICD_BASE_GPA                        0x8000000ULL
+#define GICR_BASE_GPA                        0x80A0000ULL
+#endif
 
 /* How many host loops to run by default (one KVM_GET_DIRTY_LOG for each loop)*/
 #define TEST_HOST_LOOP_N             2UL
@@ -200,6 +206,10 @@ static void run_test(enum vm_guest_mode mode, void *arg)
              vm_enable_cap(vm, &cap);
      }
 
Quoted 3 lines

tools/testing/selftests/kvm/dirty_log_perf_test.c static void run_test(enum vm_guest_mode mode, void *arg) · line 210

Quoted 2 lines
tools/testing/selftests/kvm/dirty_log_perf_test.c +2 -0
+#ifdef __aarch64__
+     vgic_v3_setup(vm, nr_vcpus, 64, GICD_BASE_GPA, GICR_BASE_GPA);

^^ extra parameter

The patch is based on kvm/queue, which has a patch adding an extra parameter nr_irqs.

Quoted 12 lines

Thanks,
drew

Quoted 8 lines
tools/testing/selftests/kvm/dirty_log_perf_test.c +2 -1
+#endif
+
      /* Start the iterations */
      iteration = 0;
      host_quit = false;
--

2.34.1.575.g55b058a8bb-goog

Quoted 6 lines
Quoted 5 lines

_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm

Thanks,
Jing
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm

maintainer
#11 lore raw

Re: [RFC PATCH 3/3] KVM: selftests: Add vgic initialization for dirty log perf test for ARM

In reply to Jing Zhang
Quoted 4 lines · Jing Zhang

For ARM64, if no vgic is setup before the dirty log perf test, the userspace irqchip would be used, which would affect the dirty log perf test result.

Doesn't it affect *all* performance tests? How much does this change contributes to the performance numbers you give in the cover letter?

Quoted 11 lines
---
 tools/testing/selftests/kvm/dirty_log_perf_test.c | 10 ++++++++++
 1 file changed, 10 insertions(+)
tools/testing/selftests/kvm/dirty_log_perf_test.c +0 -0
index 1954b964d1cf..b501338d9430 100644

tools/testing/selftests/kvm/dirty_log_perf_test.c line 25

Quoted 8 lines
tools/testing/selftests/kvm/dirty_log_perf_test.c +5 -0
@@ -18,6 +18,12 @@
 #include "test_util.h"
 #include "perf_test_util.h"
 #include "guest_modes.h"
+#ifdef __aarch64__
+#include "aarch64/vgic.h"
+
+#define GICD_BASE_GPA			0x8000000ULL
+#define GICR_BASE_GPA			0x80A0000ULL

How did you pick these values?

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author
#12 lore raw

Re: [RFC PATCH 3/3] KVM: selftests: Add vgic initialization for dirty log perf test for ARM

Quoted 11 lines · Marc Zyngier
Quoted 4 lines · Jing Zhang

For ARM64, if no vgic is setup before the dirty log perf test, the userspace irqchip would be used, which would affect the dirty log perf test result.

Doesn't it affect *all* performance tests? How much does this change contributes to the performance numbers you give in the cover letter?

This bottleneck showed up after adding the fast path patch. I didn't try other performance tests with this, but I think it is a good idea to add a vgic setup for all performance tests. I can post another patch later to make it available for all performance tests after finishing this one and verifying all other performance tests. Below is the test result without adding the vgic setup. It shows 20~30% improvement for the different number of vCPUs.

+-------+------------------------+
    | #vCPU | dirty memory time (ms) |
    +-------+------------------------+
    | 1     | 965                    |
    +-------+------------------------+
    | 2     | 1006                    |
    +-------+------------------------+
    | 4     | 1128                    |
    +-------+------------------------+
    | 8     | 2005                   |
    +-------+------------------------+
    | 16    | 3903                   |
    +-------+------------------------+
    | 32    | 7595                   |
    +-------+------------------------+
    | 64    | 15783                  |
    +-------+------------------------+
Quoted 6 lines
Quoted 6 lines
---
 tools/testing/selftests/kvm/dirty_log_perf_test.c | 10 ++++++++++
 1 file changed, 10 insertions(+)
Quoted 15 lines

tools/testing/selftests/kvm/dirty_log_perf_test.c line 25

Quoted 13 lines
tools/testing/selftests/kvm/dirty_log_perf_test.c +5 -0
index 1954b964d1cf..b501338d9430 100644
@@ -18,6 +18,12 @@
 #include "test_util.h"
 #include "perf_test_util.h"
 #include "guest_modes.h"
+#ifdef __aarch64__
+#include "aarch64/vgic.h"
+
+#define GICD_BASE_GPA                        0x8000000ULL
+#define GICR_BASE_GPA                        0x80A0000ULL

How did you pick these values?

I used the same values from other tests.
Talked with Raghavendra about the values. It could be arbitrary and he chose these values from QEMU's configuration.

Quoted 7 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Thanks,
Jing

maintainer
#13 lore raw

Re: [RFC PATCH 3/3] KVM: selftests: Add vgic initialization for dirty log perf test for ARM

Quoted 29 lines · Jing Zhang
Quoted 11 lines · Marc Zyngier
Quoted 4 lines · Jing Zhang

For ARM64, if no vgic is setup before the dirty log perf test, the userspace irqchip would be used, which would affect the dirty log perf test result.

Doesn't it affect *all* performance tests? How much does this change contributes to the performance numbers you give in the cover letter?

This bottleneck showed up after adding the fast path patch. I didn't try other performance tests with this, but I think it is a good idea to add a vgic setup for all performance tests. I can post another patch later to make it available for all performance tests after finishing this one and verifying all other performance tests. Below is the test result without adding the vgic setup. It shows 20~30% improvement for the different number of vCPUs.

+-------+------------------------+
    | #vCPU | dirty memory time (ms) |
    +-------+------------------------+
    | 1     | 965                    |
    +-------+------------------------+
    | 2     | 1006                    |
    +-------+------------------------+
    | 4     | 1128                    |
    +-------+------------------------+
Quoted 8 lines
    | 8     | 2005                   |
    +-------+------------------------+
    | 16    | 3903                   |
    +-------+------------------------+
    | 32    | 7595                   |
    +-------+------------------------+
    | 64    | 15783                  |
    +-------+------------------------+

So please use these numbers in your cover letter when you repost your series, as the improvement you'd observe on actual workloads is likely to be less than what you claim due to this change in the test itself (in other words, if you are going to benchamark something, don't change the benchmark halfway).

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author
#14 lore raw

Re: [RFC PATCH 3/3] KVM: selftests: Add vgic initialization for dirty log perf test for ARM

Quoted 41 lines · Marc Zyngier
Quoted 29 lines · Jing Zhang
Quoted 11 lines · Marc Zyngier
Quoted 4 lines · Jing Zhang

For ARM64, if no vgic is setup before the dirty log perf test, the userspace irqchip would be used, which would affect the dirty log perf test result.

Doesn't it affect *all* performance tests? How much does this change contributes to the performance numbers you give in the cover letter?

This bottleneck showed up after adding the fast path patch. I didn't try other performance tests with this, but I think it is a good idea to add a vgic setup for all performance tests. I can post another patch later to make it available for all performance tests after finishing this one and verifying all other performance tests. Below is the test result without adding the vgic setup. It shows 20~30% improvement for the different number of vCPUs.

+-------+------------------------+
    | #vCPU | dirty memory time (ms) |
    +-------+------------------------+
    | 1     | 965                    |
    +-------+------------------------+
    | 2     | 1006                    |
    +-------+------------------------+
    | 4     | 1128                    |
    +-------+------------------------+
Quoted 8 lines
    | 8     | 2005                   |
    +-------+------------------------+
    | 16    | 3903                   |
    +-------+------------------------+
    | 32    | 7595                   |
    +-------+------------------------+
    | 64    | 15783                  |
    +-------+------------------------+
Quoted 5 lines

So please use these numbers in your cover letter when you repost your series, as the improvement you'd observe on actual workloads is likely to be less than what you claim due to this change in the test itself (in other words, if you are going to benchamark something, don't change the benchmark halfway).

Sure. Will clarify this in the cover letter in future posts. Thanks,
Jing

Quoted 5 lines
        M.

--
Without deviation from the norm, progress is not possible.

_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm

maintainer
#15 lore raw
In reply to Jing Zhang
Quoted 9 lines · Jing Zhang

This patch is to reduce the performance degradation of guest workload during dirty logging on ARM64. A fast path is added to handle permission relaxation during dirty logging. The MMU lock is replaced with rwlock, by which all permision relaxations on leaf pte can be performed under the read lock. This greatly reduces the MMU lock contention during dirty logging. With this solution, the source guest workload performance degradation can be improved by more than 60%.

Quoted 9 lines

Problem:
* A Google internal live migration test shows that the source guest workload performance has >99% degradation for about 105 seconds, >50% degradation for about 112 seconds, >10% degradation for about 112 seconds on ARM64. This shows that most of the time, the guest workload degradtion is above 99%, which obviously needs some improvement compared to the test result on x86 (>99% for 6s, >50% for 9s, >10% for 27s).
* Tested H/W: Ampere Altra 3GHz, #CPU: 64, #Mem: 256GB
* VM spec: #vCPU: 48, #Mem/vCPU: 4GB

What are the host and guest page sizes?

Quoted 7 lines

Analysis:
* We enabled CONFIG_LOCK_STAT in kernel and used dirty_log_perf_test to get the number of contentions of MMU lock and the "dirty memory time" on various VM spec.
By using test command
./dirty_log_perf_test -b 2G -m 2 -i 2 -s anonymous_hugetlb_2mb -v [#vCPU]

How is this test representative of the internal live migration test you mention above? '-m 2' indicates a mode that varies depending on the HW and revision of the test (I just added a bunch of supported modes). Which one is it?

Quoted 45 lines
    Below are the results:
    +-------+------------------------+-----------------------+
    | #vCPU | dirty memory time (ms) | number of contentions |
    +-------+------------------------+-----------------------+
    | 1     | 926                    | 0                     |
    +-------+------------------------+-----------------------+
    | 2     | 1189                   | 4732558               |
    +-------+------------------------+-----------------------+
    | 4     | 2503                   | 11527185              |
    +-------+------------------------+-----------------------+
    | 8     | 5069                   | 24881677              |
    +-------+------------------------+-----------------------+
    | 16    | 10340                  | 50347956              |
    +-------+------------------------+-----------------------+
    | 32    | 20351                  | 100605720             |
    +-------+------------------------+-----------------------+
    | 64    | 40994                  | 201442478             |
    +-------+------------------------+-----------------------+

* From the test results above, the "dirty memory time" and the number of MMU lock contention scale with the number of vCPUs. That means all the dirty memory operations from all vCPU threads have been serialized by the MMU lock. Further analysis also shows that the permission relaxation during dirty logging is where vCPU threads get serialized.

Solution:
* On ARM64, there is no mechanism as PML (Page Modification Logging) and the dirty-bit solution for dirty logging is much complicated compared to the write-protection solution. The straight way to reduce the guest performance degradation is to enhance the concurrency for the permission fault path during dirty logging.
* In this patch, we only put leaf PTE permission relaxation for dirty

    logging under read lock, all others would go under write lock.
    Below are the results based on the solution:
    +-------+------------------------+
    | #vCPU | dirty memory time (ms) |
    +-------+------------------------+
    | 1     | 803                    |
    +-------+------------------------+
    | 2     | 843                    |
    +-------+------------------------+
    | 4     | 942                    |
    +-------+------------------------+
    | 8     | 1458                   |
    +-------+------------------------+
Quoted 8 lines
    | 16    | 2853                   |
    +-------+------------------------+
    | 32    | 5886                   |
    +-------+------------------------+
    | 64    | 12190                  |
    +-------+------------------------+
    All "dirty memory time" have been reduced by more than 60% when the
    number of vCPU grows.

How does that translate to the original problem statement with your live migration test?

Thanks,

	M.
Signature · 5 lines
Without deviation from the norm, progress is not possible.
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
author
#16 lore raw
Quoted 22 lines · Marc Zyngier
Quoted 9 lines · Jing Zhang

This patch is to reduce the performance degradation of guest workload during dirty logging on ARM64. A fast path is added to handle permission relaxation during dirty logging. The MMU lock is replaced with rwlock, by which all permision relaxations on leaf pte can be performed under the read lock. This greatly reduces the MMU lock contention during dirty logging. With this solution, the source guest workload performance degradation can be improved by more than 60%.

Quoted 9 lines

Problem:
* A Google internal live migration test shows that the source guest workload performance has >99% degradation for about 105 seconds, >50% degradation for about 112 seconds, >10% degradation for about 112 seconds on ARM64. This shows that most of the time, the guest workload degradtion is above 99%, which obviously needs some improvement compared to the test result on x86 (>99% for 6s, >50% for 9s, >10% for 27s).
* Tested H/W: Ampere Altra 3GHz, #CPU: 64, #Mem: 256GB
* VM spec: #vCPU: 48, #Mem/vCPU: 4GB

Quoted 1 line

What are the host and guest page sizes?

Both are 4K and guest mem is 2M hugepage backed. Will add the info for future posts.

Quoted 9 lines
Quoted 8 lines

Analysis:
* We enabled CONFIG_LOCK_STAT in kernel and used dirty_log_perf_test to get the number of contentions of MMU lock and the "dirty memory time" on various VM spec.
By using test command
./dirty_log_perf_test -b 2G -m 2 -i 2 -s anonymous_hugetlb_2mb -v [#vCPU]

Quoted 4 lines

How is this test representative of the internal live migration test you mention above? '-m 2' indicates a mode that varies depending on the HW and revision of the test (I just added a bunch of supported modes). Which one is it?

The "dirty memory time" is the time vCPU threads spent in KVM after fault. Higher "dirty memory time" means higher degradation to guest workload.
'-m 2' indicates mode "PA-bits:48, VA-bits:48, 4K pages". Will add this for future posts.

Quoted 55 lines
Quoted 46 lines
    Below are the results:
    +-------+------------------------+-----------------------+
    | #vCPU | dirty memory time (ms) | number of contentions |
    +-------+------------------------+-----------------------+
    | 1     | 926                    | 0                     |
    +-------+------------------------+-----------------------+
    | 2     | 1189                   | 4732558               |
    +-------+------------------------+-----------------------+
    | 4     | 2503                   | 11527185              |
    +-------+------------------------+-----------------------+
    | 8     | 5069                   | 24881677              |
    +-------+------------------------+-----------------------+
    | 16    | 10340                  | 50347956              |
    +-------+------------------------+-----------------------+
    | 32    | 20351                  | 100605720             |
    +-------+------------------------+-----------------------+
    | 64    | 40994                  | 201442478             |
    +-------+------------------------+-----------------------+

* From the test results above, the "dirty memory time" and the number of MMU lock contention scale with the number of vCPUs. That means all the dirty memory operations from all vCPU threads have been serialized by the MMU lock. Further analysis also shows that the permission relaxation during dirty logging is where vCPU threads get serialized.

Solution:
* On ARM64, there is no mechanism as PML (Page Modification Logging) and the dirty-bit solution for dirty logging is much complicated compared to the write-protection solution. The straight way to reduce the guest performance degradation is to enhance the concurrency for the permission fault path during dirty logging.
* In this patch, we only put leaf PTE permission relaxation for dirty

    logging under read lock, all others would go under write lock.
    Below are the results based on the solution:
    +-------+------------------------+
    | #vCPU | dirty memory time (ms) |
    +-------+------------------------+
    | 1     | 803                    |
    +-------+------------------------+
    | 2     | 843                    |
    +-------+------------------------+
    | 4     | 942                    |
    +-------+------------------------+
    | 8     | 1458                   |
    +-------+------------------------+
Quoted 8 lines
    | 16    | 2853                   |
    +-------+------------------------+
    | 32    | 5886                   |
    +-------+------------------------+
    | 64    | 12190                  |
    +-------+------------------------+
    All "dirty memory time" have been reduced by more than 60% when the
    number of vCPU grows.
Quoted 2 lines

How does that translate to the original problem statement with your live migration test?

Based on the solution, the test results from the Google internal live migration test also shows more than 60% improvement with >99% for 30s,

Quoted 1 line

50% for 58s and >10% for 76s.

Will add this info in to future posts.

Quoted 7 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Thanks,
Jing

#17 lore raw
In reply to Jing Zhang

Hi Jing,

Quoted 67 lines · Jing Zhang

This patch is to reduce the performance degradation of guest workload during dirty logging on ARM64. A fast path is added to handle permission relaxation during dirty logging. The MMU lock is replaced with rwlock, by which all permision relaxations on leaf pte can be performed under the read lock. This greatly reduces the MMU lock contention during dirty logging. With this solution, the source guest workload performance degradation can be improved by more than 60%.

Problem:
* A Google internal live migration test shows that the source guest workload performance has >99% degradation for about 105 seconds, >50% degradation for about 112 seconds, >10% degradation for about 112 seconds on ARM64. This shows that most of the time, the guest workload degradtion is above 99%, which obviously needs some improvement compared to the test result on x86 (>99% for 6s, >50% for 9s, >10% for 27s).
* Tested H/W: Ampere Altra 3GHz, #CPU: 64, #Mem: 256GB
* VM spec: #vCPU: 48, #Mem/vCPU: 4GB

Analysis:
* We enabled CONFIG_LOCK_STAT in kernel and used dirty_log_perf_test to get

    the number of contentions of MMU lock and the "dirty memory time" on
    various VM spec.
    By using test command
    ./dirty_log_perf_test -b 2G -m 2 -i 2 -s anonymous_hugetlb_2mb -v [#vCPU]
    Below are the results:
    +-------+------------------------+-----------------------+
    | #vCPU | dirty memory time (ms) | number of contentions |
    +-------+------------------------+-----------------------+
    | 1     | 926                    | 0                     |
    +-------+------------------------+-----------------------+
    | 2     | 1189                   | 4732558               |
    +-------+------------------------+-----------------------+
    | 4     | 2503                   | 11527185              |
    +-------+------------------------+-----------------------+
    | 8     | 5069                   | 24881677              |
    +-------+------------------------+-----------------------+
    | 16    | 10340                  | 50347956              |
    +-------+------------------------+-----------------------+
    | 32    | 20351                  | 100605720             |
    +-------+------------------------+-----------------------+
    | 64    | 40994                  | 201442478             |
    +-------+------------------------+-----------------------+

* From the test results above, the "dirty memory time" and the number of MMU lock contention scale with the number of vCPUs. That means all the dirty memory operations from all vCPU threads have been serialized by the MMU lock. Further analysis also shows that the permission relaxation during dirty logging is where vCPU threads get serialized.

Solution:
* On ARM64, there is no mechanism as PML (Page Modification Logging) and the dirty-bit solution for dirty logging is much complicated compared to the write-protection solution. The straight way to reduce the guest performance degradation is to enhance the concurrency for the permission fault path during dirty logging.
* In this patch, we only put leaf PTE permission relaxation for dirty

    logging under read lock, all others would go under write lock.
    Below are the results based on the solution:
    +-------+------------------------+
    | #vCPU | dirty memory time (ms) |
    +-------+------------------------+
    | 1     | 803                    |
    +-------+------------------------+
    | 2     | 843                    |
    +-------+------------------------+
    | 4     | 942                    |
    +-------+------------------------+
Quoted 8 lines
    | 8     | 1458                   |
    +-------+------------------------+
    | 16    | 2853                   |
    +-------+------------------------+
    | 32    | 5886                   |
    +-------+------------------------+
    | 64    | 12190                  |
    +-------+------------------------+

Just curious, do yo know why is time still doubling (roughly) with the number of cpus? maybe you performed another experiment or have some guess(es).

Thanks,
Ricardo

Quoted 12 lines

All "dirty memory time" have been reduced by more than 60% when the number of vCPU grows.

---

Jing Zhang (3):
KVM: arm64: Use read/write spin lock for MMU protection
KVM: arm64: Add fast path to handle permission relaxation during dirty
logging
KVM: selftests: Add vgic initialization for dirty log perf test for
ARM

Quoted 14 lines
 arch/arm64/include/asm/kvm_host.h             |  2 +
 arch/arm64/kvm/mmu.c                          | 86 +++++++++++++++----
 .../selftests/kvm/dirty_log_perf_test.c       | 10 +++
 3 files changed, 80 insertions(+), 18 deletions(-)

base-commit: fea31d1690945e6dd6c3e89ec5591490857bc3d4

Signature · 6 lines
2.34.1.575.g55b058a8bb-goog

_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm

_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm

author
#18 lore raw
Quoted 80 lines · Ricardo Koller

Hi Jing,

Quoted 67 lines · Jing Zhang

This patch is to reduce the performance degradation of guest workload during dirty logging on ARM64. A fast path is added to handle permission relaxation during dirty logging. The MMU lock is replaced with rwlock, by which all permision relaxations on leaf pte can be performed under the read lock. This greatly reduces the MMU lock contention during dirty logging. With this solution, the source guest workload performance degradation can be improved by more than 60%.

Problem:
* A Google internal live migration test shows that the source guest workload performance has >99% degradation for about 105 seconds, >50% degradation for about 112 seconds, >10% degradation for about 112 seconds on ARM64. This shows that most of the time, the guest workload degradtion is above 99%, which obviously needs some improvement compared to the test result on x86 (>99% for 6s, >50% for 9s, >10% for 27s).
* Tested H/W: Ampere Altra 3GHz, #CPU: 64, #Mem: 256GB
* VM spec: #vCPU: 48, #Mem/vCPU: 4GB

Analysis:
* We enabled CONFIG_LOCK_STAT in kernel and used dirty_log_perf_test to get

    the number of contentions of MMU lock and the "dirty memory time" on
    various VM spec.
    By using test command
    ./dirty_log_perf_test -b 2G -m 2 -i 2 -s anonymous_hugetlb_2mb -v [#vCPU]
    Below are the results:
    +-------+------------------------+-----------------------+
    | #vCPU | dirty memory time (ms) | number of contentions |
    +-------+------------------------+-----------------------+
    | 1     | 926                    | 0                     |
    +-------+------------------------+-----------------------+
    | 2     | 1189                   | 4732558               |
    +-------+------------------------+-----------------------+
    | 4     | 2503                   | 11527185              |
    +-------+------------------------+-----------------------+
    | 8     | 5069                   | 24881677              |
    +-------+------------------------+-----------------------+
    | 16    | 10340                  | 50347956              |
    +-------+------------------------+-----------------------+
    | 32    | 20351                  | 100605720             |
    +-------+------------------------+-----------------------+
    | 64    | 40994                  | 201442478             |
    +-------+------------------------+-----------------------+

* From the test results above, the "dirty memory time" and the number of MMU lock contention scale with the number of vCPUs. That means all the dirty memory operations from all vCPU threads have been serialized by the MMU lock. Further analysis also shows that the permission relaxation during dirty logging is where vCPU threads get serialized.

Solution:
* On ARM64, there is no mechanism as PML (Page Modification Logging) and the dirty-bit solution for dirty logging is much complicated compared to the write-protection solution. The straight way to reduce the guest performance degradation is to enhance the concurrency for the permission fault path during dirty logging.
* In this patch, we only put leaf PTE permission relaxation for dirty

    logging under read lock, all others would go under write lock.
    Below are the results based on the solution:
    +-------+------------------------+
    | #vCPU | dirty memory time (ms) |
    +-------+------------------------+
    | 1     | 803                    |
    +-------+------------------------+
    | 2     | 843                    |
    +-------+------------------------+
    | 4     | 942                    |
    +-------+------------------------+
Quoted 8 lines
    | 8     | 1458                   |
    +-------+------------------------+
    | 16    | 2853                   |
    +-------+------------------------+
    | 32    | 5886                   |
    +-------+------------------------+
    | 64    | 12190                  |
    +-------+------------------------+
Quoted 3 lines

Just curious, do yo know why is time still doubling (roughly) with the number of cpus? maybe you performed another experiment or have some guess(es).

Yes. it is from the serialization caused by TLB flush whenever the permission is relaxed. I tried test by removing the TLB flushes (of course it shouldn't be removed), the time would be close to a constant no matter the number of vCPUs.

Quoted 30 lines

Thanks,
Ricardo

Quoted 26 lines

All "dirty memory time" have been reduced by more than 60% when the number of vCPU grows.

---

Jing Zhang (3):
KVM: arm64: Use read/write spin lock for MMU protection
KVM: arm64: Add fast path to handle permission relaxation during dirty
logging
KVM: selftests: Add vgic initialization for dirty log perf test for
ARM

 arch/arm64/include/asm/kvm_host.h             |  2 +
 arch/arm64/kvm/mmu.c                          | 86 +++++++++++++++----
 .../selftests/kvm/dirty_log_perf_test.c       | 10 +++
 3 files changed, 80 insertions(+), 18 deletions(-)

base-commit: fea31d1690945e6dd6c3e89ec5591490857bc3d4 --
2.34.1.575.g55b058a8bb-goog

_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm

Thanks,
Jing
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm

#19 lore raw
Quoted 81 lines · Jing Zhang
Quoted 80 lines · Ricardo Koller

Hi Jing,

Quoted 75 lines · Jing Zhang

This patch is to reduce the performance degradation of guest workload during dirty logging on ARM64. A fast path is added to handle permission relaxation during dirty logging. The MMU lock is replaced with rwlock, by which all permision relaxations on leaf pte can be performed under the read lock. This greatly reduces the MMU lock contention during dirty logging. With this solution, the source guest workload performance degradation can be improved by more than 60%.

Problem:
* A Google internal live migration test shows that the source guest workload performance has >99% degradation for about 105 seconds, >50% degradation for about 112 seconds, >10% degradation for about 112 seconds on ARM64. This shows that most of the time, the guest workload degradtion is above 99%, which obviously needs some improvement compared to the test result on x86 (>99% for 6s, >50% for 9s, >10% for 27s).
* Tested H/W: Ampere Altra 3GHz, #CPU: 64, #Mem: 256GB
* VM spec: #vCPU: 48, #Mem/vCPU: 4GB

Analysis:
* We enabled CONFIG_LOCK_STAT in kernel and used dirty_log_perf_test to get

    the number of contentions of MMU lock and the "dirty memory time" on
    various VM spec.
    By using test command
    ./dirty_log_perf_test -b 2G -m 2 -i 2 -s anonymous_hugetlb_2mb -v [#vCPU]
    Below are the results:
    +-------+------------------------+-----------------------+
    | #vCPU | dirty memory time (ms) | number of contentions |
    +-------+------------------------+-----------------------+
    | 1     | 926                    | 0                     |
    +-------+------------------------+-----------------------+
    | 2     | 1189                   | 4732558               |
    +-------+------------------------+-----------------------+
    | 4     | 2503                   | 11527185              |
    +-------+------------------------+-----------------------+
    | 8     | 5069                   | 24881677              |
    +-------+------------------------+-----------------------+
    | 16    | 10340                  | 50347956              |
    +-------+------------------------+-----------------------+
    | 32    | 20351                  | 100605720             |
    +-------+------------------------+-----------------------+
    | 64    | 40994                  | 201442478             |
    +-------+------------------------+-----------------------+

* From the test results above, the "dirty memory time" and the number of MMU lock contention scale with the number of vCPUs. That means all the dirty memory operations from all vCPU threads have been serialized by the MMU lock. Further analysis also shows that the permission relaxation during dirty logging is where vCPU threads get serialized.

Solution:
* On ARM64, there is no mechanism as PML (Page Modification Logging) and the dirty-bit solution for dirty logging is much complicated compared to the write-protection solution. The straight way to reduce the guest performance degradation is to enhance the concurrency for the permission fault path during dirty logging.
* In this patch, we only put leaf PTE permission relaxation for dirty

    logging under read lock, all others would go under write lock.
    Below are the results based on the solution:
    +-------+------------------------+
    | #vCPU | dirty memory time (ms) |
    +-------+------------------------+
    | 1     | 803                    |
    +-------+------------------------+
    | 2     | 843                    |
    +-------+------------------------+
    | 4     | 942                    |
    +-------+------------------------+
    | 8     | 1458                   |
    +-------+------------------------+
    | 16    | 2853                   |
    +-------+------------------------+
    | 32    | 5886                   |
    +-------+------------------------+
    | 64    | 12190                  |
    +-------+------------------------+
Quoted 7 lines
Quoted 3 lines

Just curious, do yo know why is time still doubling (roughly) with the number of cpus? maybe you performed another experiment or have some guess(es).

Yes. it is from the serialization caused by TLB flush whenever the permission is relaxed. I tried test by removing the TLB flushes (of course it shouldn't be removed), the time would be close to a constant no matter the number of vCPUs.

Got it, thanks for the info.

Ricardo

Quoted 32 lines
Quoted 30 lines

Thanks,
Ricardo

Quoted 26 lines

All "dirty memory time" have been reduced by more than 60% when the number of vCPU grows.

---

Jing Zhang (3):
KVM: arm64: Use read/write spin lock for MMU protection
KVM: arm64: Add fast path to handle permission relaxation during dirty
logging
KVM: selftests: Add vgic initialization for dirty log perf test for
ARM

 arch/arm64/include/asm/kvm_host.h             |  2 +
 arch/arm64/kvm/mmu.c                          | 86 +++++++++++++++----
 .../selftests/kvm/dirty_log_perf_test.c       | 10 +++
 3 files changed, 80 insertions(+), 18 deletions(-)

base-commit: fea31d1690945e6dd6c3e89ec5591490857bc3d4 --
2.34.1.575.g55b058a8bb-goog

_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm

Thanks,
Jing