Skip to contentSkip to conversation

KVM: SVM: Add Page Modification Logging (PML) support

PATCH Superseded Nikunj A Dadhania 5 participants 29 messages Apr 7 – May 14

Subsystems (3)

v6 of 8 Superseded Latest is v8 →

lore
Share & tools
lore
#0 lore raw

v6 cover KVM: SVM: Add Page Modification Logging (PML) support

This series implements Page Modification Logging (PML) for guests on AMD processors, bringing hardware-assisted dirty logging support. PML is designed to track guest modified memory pages. PML enables the hypervisor to identify which pages in a guest's memory have been modified since the last checkpoint or during live migration.

The PML feature uses two new VMCB fields (PML_ADDR and PML_INDEX) and generates a VMEXIT when the 4KB log buffer becomes full.

The feature is enabled by default when hardware support is detected and can be disabled via the 'pml' module parameter.

Changelog:
v6:
* Collect RoBs
* Rebase on latest kvm/master
* Dropped SEV PML selftest (Kai Huang)
* Added guard(svm_vmcb01) helpers to temporarily switch to vmcb01
* Use guard(svm_vmcb01) to ensure vmcb01 is targeted during PML updates when L2 is active, similar to VMX. Dropped patch moving nested CPU dirty logging to common code; upstream now updates vmcs01 on-demand via temporary VMCS switch instead of deferring updates.

v5: https://lore.kernel.org/kvm/[email protected]/
* Rebased on latest kvm/next
* Use EXPORT_SYMBOL_FOR_KVM_INTERNAL (Kai Huang)
* Use cpu_dirty_log_size instead of enable_pml for PML checks, with this patch moving enable_pml to common code is not needed (Kai Huang)
* Added SEV PML self test

Nikunj A Dadhania (7):
KVM: x86: Carve out PML flush routine
KVM: x86: Move PML page to common vcpu arch structure
KVM: VMX: Use cpu_dirty_log_size instead of enable_pml for PML checks x86/cpufeatures: Add Page modification logging
KVM: SVM: Use BIT_ULL for 64-bit nested_ctl bit definitions
KVM: nSVM: Add helpers to temporarily switch to vmcb01
KVM: SVM: Add Page modification logging support

 arch/x86/include/asm/cpufeatures.h |   1 +
 arch/x86/include/asm/kvm_host.h    |   2 +
 arch/x86/include/asm/svm.h         |  12 ++-
 arch/x86/include/uapi/asm/svm.h    |   2 +
 arch/x86/kernel/cpu/scattered.c    |   1 +
 arch/x86/kvm/svm/nested.c          |  13 ++-
 arch/x86/kvm/svm/sev.c             |   2 +-
 arch/x86/kvm/svm/svm.c             | 125 ++++++++++++++++++++++++++++-
 arch/x86/kvm/svm/svm.h             |   3 +
 arch/x86/kvm/vmx/vmx.c             |  50 ++++--------
 arch/x86/kvm/vmx/vmx.h             |   7 --
 arch/x86/kvm/x86.c                 |  31 +++++++
 arch/x86/kvm/x86.h                 |   8 ++
 13 files changed, 206 insertions(+), 51 deletions(-)

base-commit: df83746075778958954aa0460cca55f4b3fc9c02

Signature · 1 line
2.48.1
#1 lore raw

PATCH v6 1/7 KVM: x86: Carve out PML flush routine

Move the PML (Page Modification Logging) buffer flushing logic from VMX-specific code to common x86 KVM code to enable reuse by SVM and avoid code duplication.

The AMD SVM PML implementations share the same behavior as VMX PML:
1) The PML buffer is a 4K page with 512 entries
2) Hardware records dirty GPAs in reverse order (from index 511 to 0)
3) Hardware clears bits 11:0 when recording GPAs

No functional change intended for VMX, except tone down the WARN_ON() to WARN_ON_ONCE() for the page alignment check. If hardware exhibits this behavior once, it's likely to occur repeatedly, so use WARN_ON_ONCE() to avoid log flooding while still capturing the unexpected condition.

The refactoring prepares for SVM to leverage the same PML flushing implementation.

---
 arch/x86/kvm/vmx/vmx.c | 26 ++------------------------
 arch/x86/kvm/vmx/vmx.h |  5 -----
 arch/x86/kvm/x86.c     | 31 +++++++++++++++++++++++++++++++
 arch/x86/kvm/x86.h     |  8 ++++++++
 4 files changed, 41 insertions(+), 29 deletions(-)
arch/x86/kvm/vmx/vmx.c +2 -24
index 8b24e682535b..b77750a2efc2 100644
@@ -6502,37 +6502,15 @@ static void vmx_destroy_pml_buffer(struct vcpu_vmx *vmx)
 static void vmx_flush_pml_buffer(struct kvm_vcpu *vcpu)
 {
 	struct vcpu_vmx *vmx = to_vmx(vcpu);
-	u16 pml_idx, pml_tail_index;
-	u64 *pml_buf;
-	int i;
+	u16 pml_idx;
 
 	pml_idx = vmcs_read16(GUEST_PML_INDEX);
 
 	/* Do nothing if PML buffer is empty */
 	if (pml_idx == PML_HEAD_INDEX)
 		return;
-	/*
-	 * PML index always points to the next available PML buffer entity
-	 * unless PML log has just overflowed.
-	 */
-	pml_tail_index = (pml_idx >= PML_LOG_NR_ENTRIES) ? 0 : pml_idx + 1;
 
-	/*
-	 * PML log is written backwards: the CPU first writes the entry 511
-	 * then the entry 510, and so on.
-	 *
-	 * Read the entries in the same order they were written, to ensure that
-	 * the dirty ring is filled in the same order the CPU wrote them.
-	 */
-	pml_buf = page_address(vmx->pml_pg);
-
-	for (i = PML_HEAD_INDEX; i >= pml_tail_index; i--) {
-		u64 gpa;
-
-		gpa = pml_buf[i];
-		WARN_ON(gpa & (PAGE_SIZE - 1));
-		kvm_vcpu_mark_page_dirty(vcpu, gpa >> PAGE_SHIFT);
-	}
+	kvm_flush_pml_buffer(vcpu, vmx->pml_pg, pml_idx);
 
 	/* reset PML index */
 	vmcs_write16(GUEST_PML_INDEX, PML_HEAD_INDEX);
arch/x86/kvm/vmx/vmx.h +0 -5
index 70bfe81dea54..9b0c5dde9437 100644
@@ -262,11 +262,6 @@ struct vcpu_vmx {
 	unsigned int ple_window;
 	bool ple_window_dirty;
 
-	/* Support for PML */
-#define PML_LOG_NR_ENTRIES	512
-	/* PML is written backwards: this is the first entry written by the CPU */
-#define PML_HEAD_INDEX		(PML_LOG_NR_ENTRIES-1)
-
 	struct page *pml_pg;
 
 	/* apic deadline value in host tsc */
arch/x86/kvm/x86.c +31 -0
index fd1c4a36b593..628b6f51d2be 100644
@@ -6720,6 +6720,37 @@ void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
 		kvm_vcpu_kick(vcpu);
 }
 
+void kvm_flush_pml_buffer(struct kvm_vcpu *vcpu, struct page *pml_page, u16 pml_idx)
+{
+	u16 pml_tail_index;
+	u64 *pml_buf;
+	int i;
+
+	/*
+	 * PML index always points to the next available PML buffer entity
+	 * unless PML log has just overflowed.
+	 */
+	pml_tail_index = (pml_idx >= PML_LOG_NR_ENTRIES) ? 0 : pml_idx + 1;
+
+	/*
+	 * PML log is written backwards: the CPU first writes the entry 511
+	 * then the entry 510, and so on.
+	 *
+	 * Read the entries in the same order they were written, to ensure that
+	 * the dirty ring is filled in the same order the CPU wrote them.
+	 */
+	pml_buf = page_address(pml_page);
+
+	for (i = PML_HEAD_INDEX; i >= pml_tail_index; i--) {
+		u64 gpa;
+
+		gpa = pml_buf[i];
+		WARN_ON_ONCE(gpa & (PAGE_SIZE - 1));
+		kvm_vcpu_mark_page_dirty(vcpu, gpa >> PAGE_SHIFT);
+	}
+}
+EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_flush_pml_buffer);
+
 int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
 			    struct kvm_enable_cap *cap)
 {
arch/x86/kvm/x86.h +8 -0
index 94d4f07aaaa0..3b2cc2756033 100644
@@ -763,4 +763,12 @@ static inline bool kvm_is_valid_u_s_cet(struct kvm_vcpu *vcpu, u64 data)
 
 	return true;
 }
+
+/* Support for PML */
+#define PML_LOG_NR_ENTRIES	512
+/* PML is written backwards: this is the first entry written by the CPU */
+#define PML_HEAD_INDEX		(PML_LOG_NR_ENTRIES-1)
+
+void kvm_flush_pml_buffer(struct kvm_vcpu *vcpu, struct page *pml_pg, u16 pml_idx);
+
 #endif
Signature · 1 line
2.48.1
#2 lore raw

Re: [PATCH v6 1/7] KVM: x86: Carve out PML flush routine

arch/x86/kvm/x86.c void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot) · line 6723

Quoted 9 lines · Nikunj A Dadhania
arch/x86/kvm/x86.c +1 -0
index fd1c4a36b593..628b6f51d2be 100644
@@ -6720,6 +6720,37 @@ void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
 		kvm_vcpu_kick(vcpu);
 }
 
+void kvm_flush_pml_buffer(struct kvm_vcpu *vcpu, struct page *pml_page, u16 pml_idx)

Swap the order of patches 1 and 2, then you don't need to add @pml_page, only to remove it one patch later.

#3 lore raw

Re: [PATCH v6 1/7] KVM: x86: Carve out PML flush routine

Quoted 9 lines · Sean Christopherson
Quoted 8 lines · Nikunj A Dadhania
arch/x86/kvm/x86.c +0 -0
index fd1c4a36b593..628b6f51d2be 100644
@@ -6720,6 +6720,37 @@ void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
 		kvm_vcpu_kick(vcpu);
 }
 
Quoted 4 lines

arch/x86/kvm/x86.c void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot) · line 6723

Quoted 1 line
arch/x86/kvm/x86.c +1 -0
+void kvm_flush_pml_buffer(struct kvm_vcpu *vcpu, struct page *pml_page, u16 pml_idx)

Swap the order of patches 1 and 2, then you don't need to add @pml_page, only to remove it one patch later.

That makes sense—I'll swap the order of patches 1 and 2 and address this in my next revision.

Regards,
Nikunj

#4 lore raw

PATCH v6 2/7 KVM: x86: Move PML page to common vcpu arch structure

In reply to Nikunj A Dadhania

Move the PML page pointer from VMX-specific vcpu_vmx structure to the common kvm_vcpu_arch structure to enable sharing between VMX and SVM implementations. Only the page pointer is moved to x86 common code while keeping allocation logic vendor-specific, since AMD requires snp_safe_alloc_page() for PML buffer allocation.

Update all VMX references accordingly, and simplify the kvm_flush_pml_buffer() interface by removing the page parameter since it can now access the page directly from the vcpu structure.

No functional change, restructuring to prepare for SVM PML support.

---
 arch/x86/include/asm/kvm_host.h |  2 ++
 arch/x86/kvm/vmx/vmx.c          | 24 ++++++++++++------------
 arch/x86/kvm/vmx/vmx.h          |  2 --
 arch/x86/kvm/x86.c              |  4 ++--
 arch/x86/kvm/x86.h              |  2 +-
 5 files changed, 17 insertions(+), 17 deletions(-)
arch/x86/include/asm/kvm_host.h +2 -0
index 6e4e3ef9b8c7..9ed19e99d481 100644
@@ -873,6 +873,8 @@ struct kvm_vcpu_arch {
 	 */
 	struct kvm_mmu_memory_cache mmu_external_spt_cache;
 
+	struct page *pml_page;
+
 	/*
 	 * QEMU userspace and the guest each have their own FPU state.
 	 * In vcpu_run, we switch between the user and guest FPU contexts.
arch/x86/kvm/vmx/vmx.c +12 -12
index b77750a2efc2..909085a5dfb3 100644
@@ -4934,7 +4934,8 @@ int vmx_vcpu_precreate(struct kvm *kvm)
 
 static void init_vmcs(struct vcpu_vmx *vmx)
 {
-	struct kvm *kvm = vmx->vcpu.kvm;
+	struct kvm_vcpu *vcpu = &vmx->vcpu;
+	struct kvm *kvm = vcpu->kvm;
 	struct kvm_vmx *kvm_vmx = to_kvm_vmx(kvm);
 
 	if (nested)
@@ -5026,7 +5027,7 @@ static void init_vmcs(struct vcpu_vmx *vmx)
 		vmcs_write64(XSS_EXIT_BITMAP, VMX_XSS_EXIT_BITMAP);
 
 	if (enable_pml) {
-		vmcs_write64(PML_ADDRESS, page_to_phys(vmx->pml_pg));
+		vmcs_write64(PML_ADDRESS, page_to_phys(vcpu->arch.pml_page));
 		vmcs_write16(GUEST_PML_INDEX, PML_HEAD_INDEX);
 	}
 
@@ -6491,17 +6492,16 @@ void vmx_get_entry_info(struct kvm_vcpu *vcpu, u32 *intr_info, u32 *error_code)
 		*error_code = 0;
 }
 
-static void vmx_destroy_pml_buffer(struct vcpu_vmx *vmx)
+static void vmx_destroy_pml_buffer(struct kvm_vcpu *vcpu)
 {
-	if (vmx->pml_pg) {
-		__free_page(vmx->pml_pg);
-		vmx->pml_pg = NULL;
+	if (vcpu->arch.pml_page) {
+		__free_page(vcpu->arch.pml_page);
+		vcpu->arch.pml_page = NULL;
 	}
 }
 
 static void vmx_flush_pml_buffer(struct kvm_vcpu *vcpu)
 {
-	struct vcpu_vmx *vmx = to_vmx(vcpu);
 	u16 pml_idx;
 
 	pml_idx = vmcs_read16(GUEST_PML_INDEX);
@@ -6510,7 +6510,7 @@ static void vmx_flush_pml_buffer(struct kvm_vcpu *vcpu)
 	if (pml_idx == PML_HEAD_INDEX)
 		return;
 
-	kvm_flush_pml_buffer(vcpu, vmx->pml_pg, pml_idx);
+	kvm_flush_pml_buffer(vcpu, pml_idx);
 
 	/* reset PML index */
 	vmcs_write16(GUEST_PML_INDEX, PML_HEAD_INDEX);
@@ -7738,7 +7738,7 @@ void vmx_vcpu_free(struct kvm_vcpu *vcpu)
 	struct vcpu_vmx *vmx = to_vmx(vcpu);
 
 	if (enable_pml)
-		vmx_destroy_pml_buffer(vmx);
+		vmx_destroy_pml_buffer(vcpu);
 	free_vpid(vmx->vpid);
 	nested_vmx_free_vcpu(vcpu);
 	free_loaded_vmcs(vmx->loaded_vmcs);
@@ -7767,8 +7767,8 @@ int vmx_vcpu_create(struct kvm_vcpu *vcpu)
 	 * for the guest), etc.
 	 */
 	if (enable_pml) {
-		vmx->pml_pg = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
-		if (!vmx->pml_pg)
+		vcpu->arch.pml_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
+		if (!vcpu->arch.pml_page)
 			goto free_vpid;
 	}
 
@@ -7839,7 +7839,7 @@ int vmx_vcpu_create(struct kvm_vcpu *vcpu)
 free_vmcs:
 	free_loaded_vmcs(vmx->loaded_vmcs);
 free_pml:
-	vmx_destroy_pml_buffer(vmx);
+	vmx_destroy_pml_buffer(vcpu);
 free_vpid:
 	free_vpid(vmx->vpid);
 	return err;
arch/x86/kvm/vmx/vmx.h +0 -2
index 9b0c5dde9437..ba04d32a1412 100644
@@ -262,8 +262,6 @@ struct vcpu_vmx {
 	unsigned int ple_window;
 	bool ple_window_dirty;
 
-	struct page *pml_pg;
-
 	/* apic deadline value in host tsc */
 	u64 hv_deadline_tsc;
 
arch/x86/kvm/x86.c +2 -2
index 628b6f51d2be..52c06d095c25 100644
@@ -6720,7 +6720,7 @@ void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
 		kvm_vcpu_kick(vcpu);
 }
 
-void kvm_flush_pml_buffer(struct kvm_vcpu *vcpu, struct page *pml_page, u16 pml_idx)
+void kvm_flush_pml_buffer(struct kvm_vcpu *vcpu, u16 pml_idx)
 {
 	u16 pml_tail_index;
 	u64 *pml_buf;
@@ -6739,7 +6739,7 @@ void kvm_flush_pml_buffer(struct kvm_vcpu *vcpu, struct page *pml_page, u16 pml_
 	 * Read the entries in the same order they were written, to ensure that
 	 * the dirty ring is filled in the same order the CPU wrote them.
 	 */
-	pml_buf = page_address(pml_page);
+	pml_buf = page_address(vcpu->arch.pml_page);
 
 	for (i = PML_HEAD_INDEX; i >= pml_tail_index; i--) {
 		u64 gpa;
arch/x86/kvm/x86.h +1 -1
index 3b2cc2756033..21d88a768d8e 100644
@@ -769,6 +769,6 @@ static inline bool kvm_is_valid_u_s_cet(struct kvm_vcpu *vcpu, u64 data)
 /* PML is written backwards: this is the first entry written by the CPU */
 #define PML_HEAD_INDEX		(PML_LOG_NR_ENTRIES-1)
 
-void kvm_flush_pml_buffer(struct kvm_vcpu *vcpu, struct page *pml_pg, u16 pml_idx);
+void kvm_flush_pml_buffer(struct kvm_vcpu *vcpu, u16 pml_idx);
 
 #endif
Signature · 1 line
2.48.1
#5 lore raw

Re: [PATCH v6 2/7] KVM: x86: Move PML page to common vcpu arch structure

arch/x86/kvm/vmx/vmx.c int vmx_vcpu_precreate(struct kvm *kvm) · line 4938

Quoted 11 lines · Nikunj A Dadhania
arch/x86/kvm/vmx/vmx.c +2 -1
index b77750a2efc2..909085a5dfb3 100644
@@ -4934,7 +4934,8 @@ int vmx_vcpu_precreate(struct kvm *kvm)
 
 static void init_vmcs(struct vcpu_vmx *vmx)
 {
-	struct kvm *kvm = vmx->vcpu.kvm;
+	struct kvm_vcpu *vcpu = &vmx->vcpu;
+	struct kvm *kvm = vcpu->kvm;

There are 9+ existing uses of "vmx->vcpu" in init_vmcs(). I'd rather insert a prep patch to pass @vcpu instead of @vmx, do to_vmx() at the top, and switch all existing vmx->vcpu usage.

From: Sean Christopherson <[email protected]>
Date: Fri, 8 May 2026 15:55:55 -0700
Subject: [PATCH] KVM: VMX: Pass @vcpu, not @vmx to init_vmcs()

Pass @vcpu instead of @vmx to init_vmcs(), and switch all of the vmx->vcpu usage to a simple vcpu.

No functional change intended.

---
 arch/x86/kvm/vmx/vmx.c | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)
arch/x86/kvm/vmx/vmx.c +13 -12
index 5c2c33a5f7dc..b8d517b03bb4 100644
@@ -4823,10 +4823,11 @@ int vmx_vcpu_precreate(struct kvm *kvm)
 
 #define VMX_XSS_EXIT_BITMAP 0
 
-static void init_vmcs(struct vcpu_vmx *vmx)
+static void init_vmcs(struct kvm_vcpu *vcpu)
 {
-	struct kvm *kvm = vmx->vcpu.kvm;
+	struct kvm *kvm = vcpu->kvm;
 	struct kvm_vmx *kvm_vmx = to_kvm_vmx(kvm);
+	struct vcpu_vmx *vmx = to_vmx(vcpu);
 
 	if (nested)
 		nested_vmx_set_vmcs_shadowing_bitmap();
@@ -4851,7 +4852,7 @@ static void init_vmcs(struct vcpu_vmx *vmx)
 	if (cpu_has_tertiary_exec_ctrls())
 		tertiary_exec_controls_set(vmx, vmx_tertiary_exec_control(vmx));
 
-	if (enable_apicv && lapic_in_kernel(&vmx->vcpu)) {
+	if (enable_apicv && lapic_in_kernel(vcpu)) {
 		vmcs_write64(EOI_EXIT_BITMAP0, 0);
 		vmcs_write64(EOI_EXIT_BITMAP1, 0);
 		vmcs_write64(EOI_EXIT_BITMAP2, 0);
@@ -4863,7 +4864,7 @@ static void init_vmcs(struct vcpu_vmx *vmx)
 		vmcs_write64(POSTED_INTR_DESC_ADDR, __pa((&vmx->vt.pi_desc)));
 	}
 
-	if (vmx_can_use_ipiv(&vmx->vcpu)) {
+	if (vmx_can_use_ipiv(vcpu)) {
 		vmcs_write64(PID_POINTER_TABLE, __pa(kvm_vmx->pid_table));
 		vmcs_write16(LAST_PID_POINTER_INDEX, kvm->arch.max_vcpu_ids - 1);
 	}
@@ -4898,15 +4899,15 @@ static void init_vmcs(struct vcpu_vmx *vmx)
 	vmcs_write64(VM_ENTRY_MSR_LOAD_ADDR, __pa(vmx->msr_autoload.guest.val));
 
 	if (vmcs_config.vmentry_ctrl & VM_ENTRY_LOAD_IA32_PAT)
-		vmcs_write64(GUEST_IA32_PAT, vmx->vcpu.arch.pat);
+		vmcs_write64(GUEST_IA32_PAT, vcpu->arch.pat);
 
 	vm_exit_controls_set(vmx, vmx_get_initial_vmexit_ctrl());
 
 	/* 22.2.1, 20.8.1 */
 	vm_entry_controls_set(vmx, vmx_get_initial_vmentry_ctrl());
 
-	vmx->vcpu.arch.cr0_guest_owned_bits = vmx_l1_guest_owned_cr0_bits();
-	vmcs_writel(CR0_GUEST_HOST_MASK, ~vmx->vcpu.arch.cr0_guest_owned_bits);
+	vcpu->arch.cr0_guest_owned_bits = vmx_l1_guest_owned_cr0_bits();
+	vmcs_writel(CR0_GUEST_HOST_MASK, ~vcpu->arch.cr0_guest_owned_bits);
 
 	set_cr4_guest_host_mask(vmx);
 
@@ -4921,7 +4922,7 @@ static void init_vmcs(struct vcpu_vmx *vmx)
 		vmcs_write16(GUEST_PML_INDEX, PML_HEAD_INDEX);
 	}
 
-	vmx_write_encls_bitmap(&vmx->vcpu, NULL);
+	vmx_write_encls_bitmap(vcpu, NULL);
 
 	if (vmx_pt_mode_is_host_guest()) {
 		memset(&vmx->pt_desc, 0, sizeof(vmx->pt_desc));
@@ -4934,13 +4935,13 @@ static void init_vmcs(struct vcpu_vmx *vmx)
 	vmcs_writel(GUEST_SYSENTER_ESP, 0);
 	vmcs_writel(GUEST_SYSENTER_EIP, 0);
 
-	vmx_guest_debugctl_write(&vmx->vcpu, 0);
+	vmx_guest_debugctl_write(vcpu, 0);
 
 	if (cpu_has_vmx_tpr_shadow()) {
 		vmcs_write64(VIRTUAL_APIC_PAGE_ADDR, 0);
-		if (cpu_need_tpr_shadow(&vmx->vcpu))
+		if (cpu_need_tpr_shadow(vcpu))
 			vmcs_write64(VIRTUAL_APIC_PAGE_ADDR,
-				     __pa(vmx->vcpu.arch.apic->regs));
+				     __pa(vcpu->arch.apic->regs));
 		vmcs_write32(TPR_THRESHOLD, 0);
 	}
 
@@ -4951,7 +4952,7 @@ static void __vmx_vcpu_reset(struct kvm_vcpu *vcpu)
 {
 	struct vcpu_vmx *vmx = to_vmx(vcpu);
 
-	init_vmcs(vmx);
+	init_vmcs(vcpu);
 
 	if (nested &&
 	    kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_STUFF_FEATURE_MSRS))

base-commit: 4c88d35f51436169884ce87bcf26f31e10106e71

#6 lore raw

Re: [PATCH v6 2/7] KVM: x86: Move PML page to common vcpu arch structure

Quoted 21 lines · Sean Christopherson

arch/x86/kvm/vmx/vmx.c int vmx_vcpu_precreate(struct kvm *kvm) · line 4938

Quoted 11 lines · Nikunj A Dadhania
arch/x86/kvm/vmx/vmx.c +2 -1
index b77750a2efc2..909085a5dfb3 100644
@@ -4934,7 +4934,8 @@ int vmx_vcpu_precreate(struct kvm *kvm)
 
 static void init_vmcs(struct vcpu_vmx *vmx)
 {
-	struct kvm *kvm = vmx->vcpu.kvm;
+	struct kvm_vcpu *vcpu = &vmx->vcpu;
+	struct kvm *kvm = vcpu->kvm;

There are 9+ existing uses of "vmx->vcpu" in init_vmcs(). I'd rather insert a prep patch to pass @vcpu instead of @vmx, do to_vmx() at the top, and switch all existing vmx->vcpu usage.

From: Sean Christopherson <[email protected]>
Date: Fri, 8 May 2026 15:55:55 -0700
Subject: [PATCH] KVM: VMX: Pass @vcpu, not @vmx to init_vmcs()

Quoted 6 lines

Pass @vcpu instead of @vmx to init_vmcs(), and switch all of the vmx->vcpu usage to a simple vcpu.

No functional change intended.

Ack. I’ll include this as a preparatory patch in my series unless it lands in your tree before I post.

Regards,
Nikunj

#7 lore raw

PATCH v6 3/7 KVM: VMX: Use cpu_dirty_log_size instead of enable_pml for PML checks

In reply to Nikunj A Dadhania

Replace the enable_pml check with cpu_dirty_log_size in VMX PML code to determine whether PML is enabled on a per-VM basis. The enable_pml module parameter is a global setting that doesn't reflect per-VM capabilities, whereas cpu_dirty_log_size accurately indicates whether a specific VM has PML enabled.

For example, TDX VMs don't yet support PML. Using cpu_dirty_log_size ensures the check correctly reflects this, while enable_pml would incorrectly indicate PML is available.

This also improves consistency with kvm_mmu_update_cpu_dirty_logging(), which already uses cpu_dirty_log_size to determine PML enablement.

---
 arch/x86/kvm/vmx/vmx.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kvm/vmx/vmx.c +1 -1
index 909085a5dfb3..a5a8bb5a7020 100644
@@ -8439,7 +8439,7 @@ void vmx_update_cpu_dirty_logging(struct kvm_vcpu *vcpu)
 {
 	struct vcpu_vmx *vmx = to_vmx(vcpu);
 
-	if (WARN_ON_ONCE(!enable_pml))
+	if (WARN_ON_ONCE(!vcpu->kvm->arch.cpu_dirty_log_size))
 		return;
 
 	guard(vmx_vmcs01)(vcpu);
Signature · 1 line
2.48.1
#8 lore raw

PATCH v6 4/7 x86/cpufeatures: Add Page modification logging

In reply to Nikunj A Dadhania

Page modification logging(PML) is a hardware feature designed to track guest modified memory pages. PML enables the hypervisor to identify which pages in a guest's memory have been changed since the last checkpoint or during live migration.

The PML feature is advertised via CPUID leaf 0x8000000A ECX[4] bit.

---
 arch/x86/include/asm/cpufeatures.h | 1 +
 arch/x86/kernel/cpu/scattered.c    | 1 +
 2 files changed, 2 insertions(+)
arch/x86/include/asm/cpufeatures.h +1 -0
index dbe104df339b..8c669c9b11d0 100644
@@ -228,6 +228,7 @@
 #define X86_FEATURE_PVUNLOCK		( 8*32+20) /* PV unlock function */
 #define X86_FEATURE_VCPUPREEMPT		( 8*32+21) /* PV vcpu_is_preempted function */
 #define X86_FEATURE_TDX_GUEST		( 8*32+22) /* "tdx_guest" Intel Trust Domain Extensions Guest */
+#define X86_FEATURE_PML			( 8*32+23) /* AMD Page Modification logging */
 
 /* Intel-defined CPU features, CPUID level 0x00000007:0 (EBX), word 9 */
 #define X86_FEATURE_FSGSBASE		( 9*32+ 0) /* "fsgsbase" RDFSBASE, WRFSBASE, RDGSBASE, WRGSBASE instructions*/
arch/x86/kernel/cpu/scattered.c +1 -0
index 42c7eac0c387..cdda4e72c5e6 100644
@@ -53,6 +53,7 @@ static const struct cpuid_bit cpuid_bits[] = {
 	{ X86_FEATURE_PROC_FEEDBACK,		CPUID_EDX, 11, 0x80000007, 0 },
 	{ X86_FEATURE_AMD_FAST_CPPC,		CPUID_EDX, 15, 0x80000007, 0 },
 	{ X86_FEATURE_MBA,			CPUID_EBX,  6, 0x80000008, 0 },
+	{ X86_FEATURE_PML,                      CPUID_ECX,  4, 0x8000000a, 0 },
 	{ X86_FEATURE_X2AVIC_EXT,		CPUID_ECX,  6, 0x8000000a, 0 },
 	{ X86_FEATURE_COHERENCY_SFW_NO,		CPUID_EBX, 31, 0x8000001f, 0 },
 	{ X86_FEATURE_SMBA,			CPUID_EBX,  2, 0x80000020, 0 },
Signature · 1 line
2.48.1
#9 lore raw

PATCH v6 5/7 KVM: SVM: Use BIT_ULL for 64-bit nested_ctl bit definitions

In reply to Nikunj A Dadhania

Replace BIT() with BIT_ULL() for SVM nested control bit definitions since nested_ctl is a 64-bit field in the VMCB control area structure.

No functional change intended.

---
 arch/x86/include/asm/svm.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
arch/x86/include/asm/svm.h +3 -3
index edde36097ddc..8fe91a0651bc 100644
@@ -239,9 +239,9 @@ struct __attribute__ ((__packed__)) vmcb_control_area {
 #define SVM_IOIO_SIZE_MASK (7 << SVM_IOIO_SIZE_SHIFT)
 #define SVM_IOIO_ASIZE_MASK (7 << SVM_IOIO_ASIZE_SHIFT)
 
-#define SVM_NESTED_CTL_NP_ENABLE	BIT(0)
-#define SVM_NESTED_CTL_SEV_ENABLE	BIT(1)
-#define SVM_NESTED_CTL_SEV_ES_ENABLE	BIT(2)
+#define SVM_NESTED_CTL_NP_ENABLE	BIT_ULL(0)
+#define SVM_NESTED_CTL_SEV_ENABLE	BIT_ULL(1)
+#define SVM_NESTED_CTL_SEV_ES_ENABLE	BIT_ULL(2)
 
 
 #define SVM_TSC_RATIO_RSVD	0xffffff0000000000ULL
Signature · 1 line
2.48.1
#10 lore raw

PATCH v6 6/7 KVM: nSVM: Add helpers to temporarily switch to vmcb01

In reply to Nikunj A Dadhania

If KVM needs to update controls that belong in vmcb01 while L2 is active, it must switch to vmcb01 and update immediately instead of deferring until the next nested VM-Exit. Deferring updates creates ordering and state consistency problems, e.g. KVM thinks a control is enabled while vmcb01 still has stale state.

Add helpers to temporarily switch to vmcb01 to ensure updates happen on-demand.

---
 arch/x86/kvm/svm/svm.c | 24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)
arch/x86/kvm/svm/svm.c +24 -0
index e6477affac9a..32b002f92528 100644
@@ -1251,6 +1251,30 @@ void svm_switch_vmcb(struct vcpu_svm *svm, struct kvm_vmcb_info *target_vmcb)
 	svm->vmcb = target_vmcb->ptr;
 }
 
+static void svm_load_vmcb01(struct kvm_vcpu *vcpu)
+{
+	struct vcpu_svm *svm = to_svm(vcpu);
+
+	if (!is_guest_mode(vcpu)) {
+		WARN_ON_ONCE(svm->vmcb != svm->vmcb01.ptr);
+		return;
+	}
+
+	WARN_ON_ONCE(svm->vmcb != svm->nested.vmcb02.ptr);
+	svm_switch_vmcb(svm, &svm->vmcb01);
+}
+
+static void svm_put_vmcb01(struct kvm_vcpu *vcpu)
+{
+	if (!is_guest_mode(vcpu))
+		return;
+
+	svm_switch_vmcb(to_svm(vcpu), &to_svm(vcpu)->nested.vmcb02);
+}
+
+DEFINE_GUARD(svm_vmcb01, struct kvm_vcpu *,
+	     svm_load_vmcb01(_T), svm_put_vmcb01(_T))
+
 static int svm_vcpu_precreate(struct kvm *kvm)
 {
 	return avic_alloc_physical_id_table(kvm);
Signature · 1 line
2.48.1
#11 lore raw

PATCH v6 7/7 KVM: SVM: Add Page modification logging support

In reply to Nikunj A Dadhania

Currently, dirty logging relies on write protecting guest memory and marking dirty GFNs during subsequent write faults. This method works but incurs overhead due to additional write faults for each dirty GFN.

Implement support for the Page Modification Logging (PML) feature, a hardware-assisted method for efficient dirty logging. PML automatically logs dirty GPA[51:12] to a 4K buffer when the CPU sets NPT D-bits. Two new VMCB fields are utilized: PML_ADDR and PML_INDEX. The PML_INDEX is initialized to 511 (8 bytes per GPA entry), and the CPU decreases the PML_INDEX after logging each GPA. When the PML buffer is full, a VMEXIT(PML_FULL) with exit code 0x407 is generated.

PML operates on guest physical addresses at the NPT level, tracking D-bit updates in page tables rather than memory content. This allows it to work identically for normal and confidential computing guests (SEV/SEV-ES/SEV-SNP):, enabling cpu_dirty_log_size to be set uniformly for all AMD VMs without special-casing encrypted guests

When L2 is active, svm->vmcb points to vmcb02, so updates to PML controls must explicitly target vmcb01 to ensure the L1's state remains correct. Use the svm_vmcb01 guard to ensure vmcb01 is active during the update.

Disable PML for nested guests.

Add a new module parameter to enable/disable PML, and enable it by default when supported

---
 arch/x86/include/asm/svm.h      |   6 +-
 arch/x86/include/uapi/asm/svm.h |   2 +
 arch/x86/kvm/svm/nested.c       |  13 +++-
 arch/x86/kvm/svm/sev.c          |   2 +-
 arch/x86/kvm/svm/svm.c          | 101 +++++++++++++++++++++++++++++++-
 arch/x86/kvm/svm/svm.h          |   3 +
 6 files changed, 122 insertions(+), 5 deletions(-)
arch/x86/include/asm/svm.h +5 -1
index 8fe91a0651bc..d52096d3eaa2 100644
@@ -165,7 +165,10 @@ struct __attribute__ ((__packed__)) vmcb_control_area {
 	u8 reserved_9[22];
 	u64 allowed_sev_features;	/* Offset 0x138 */
 	u64 guest_sev_features;		/* Offset 0x140 */
-	u8 reserved_10[664];
+	u8 reserved_10[128];
+	u64 pml_addr;			/* Offset 0x1c8 */
+	u16 pml_index;			/* Offset 0x1d0 */
+	u8 reserved_11[526];
 	/*
 	 * Offset 0x3e0, 32 bytes reserved
 	 * for use by hypervisor/software.
@@ -242,6 +245,7 @@ struct __attribute__ ((__packed__)) vmcb_control_area {
 #define SVM_NESTED_CTL_NP_ENABLE	BIT_ULL(0)
 #define SVM_NESTED_CTL_SEV_ENABLE	BIT_ULL(1)
 #define SVM_NESTED_CTL_SEV_ES_ENABLE	BIT_ULL(2)
+#define SVM_NESTED_CTL_PML_ENABLE	BIT_ULL(11)
 
 
 #define SVM_TSC_RATIO_RSVD	0xffffff0000000000ULL
arch/x86/include/uapi/asm/svm.h +2 -0
index 010a45c9f614..e80676185092 100644
@@ -101,6 +101,7 @@
 #define SVM_EXIT_AVIC_INCOMPLETE_IPI		0x401
 #define SVM_EXIT_AVIC_UNACCELERATED_ACCESS	0x402
 #define SVM_EXIT_VMGEXIT       0x403
+#define SVM_EXIT_PML_FULL	0x407
 
 /* SEV-ES software-defined VMGEXIT events */
 #define SVM_VMGEXIT_MMIO_READ			0x80000001ull
@@ -236,6 +237,7 @@
 	{ SVM_EXIT_AVIC_INCOMPLETE_IPI,		"avic_incomplete_ipi" }, \
 	{ SVM_EXIT_AVIC_UNACCELERATED_ACCESS,   "avic_unaccelerated_access" }, \
 	{ SVM_EXIT_VMGEXIT,		"vmgexit" }, \
+	{ SVM_EXIT_PML_FULL,		"pml_full" }, \
 	{ SVM_VMGEXIT_MMIO_READ,	"vmgexit_mmio_read" }, \
 	{ SVM_VMGEXIT_MMIO_WRITE,	"vmgexit_mmio_write" }, \
 	{ SVM_VMGEXIT_NMI_COMPLETE,	"vmgexit_nmi_complete" }, \
arch/x86/kvm/svm/nested.c +12 -1
index b36c33255bed..9afe4d86793b 100644
@@ -790,12 +790,23 @@ static void nested_vmcb02_prepare_control(struct vcpu_svm *svm,
 						V_NMI_BLOCKING_MASK);
 	}
 
-	/* Copied from vmcb01.  msrpm_base can be overwritten later.  */
+	/* Copied from vmcb01. msrpm_base/nested_ctl can be overwritten later. */
 	vmcb02->control.nested_ctl = vmcb01->control.nested_ctl;
 	vmcb02->control.iopm_base_pa = vmcb01->control.iopm_base_pa;
 	vmcb02->control.msrpm_base_pa = vmcb01->control.msrpm_base_pa;
 	vmcb_mark_dirty(vmcb02, VMCB_PERM_MAP);
 
+	/*
+	 * Disable PML for nested guests. When L2 runs with PML enabled, the
+	 * CPU logs L2 GPAs rather than L1 GPAs, breaking dirty page tracking
+	 * for the L0 hypervisor.
+	 */
+	if (pml) {
+		vmcb02->control.nested_ctl &= ~SVM_NESTED_CTL_PML_ENABLE;
+		vmcb02->control.pml_addr = 0;
+		vmcb02->control.pml_index = -1;
+	}
+
 	/*
 	 * Stash vmcb02's counter if the guest hasn't moved past the guilty
 	 * instruction; otherwise, reset the counter to '0'.
arch/x86/kvm/svm/sev.c +1 -1
index 3f9c1aa39a0a..427de45843f2 100644
@@ -4820,7 +4820,7 @@ struct page *snp_safe_alloc_page_node(int node, gfp_t gfp)
 	 * Allocate an SNP-safe page to workaround the SNP erratum where
 	 * the CPU will incorrectly signal an RMP violation #PF if a
 	 * hugepage (2MB or 1GB) collides with the RMP entry of a
-	 * 2MB-aligned VMCB, VMSA, or AVIC backing page.
+	 * 2MB-aligned VMCB, VMSA, PML or AVIC backing page.
 	 *
 	 * Allocate one extra page, choose a page which is not
 	 * 2MB-aligned, and free the other.
arch/x86/kvm/svm/svm.c +99 -2

176 lines not on this page · raw message

arch/x86/kvm/svm/svm.h +3 -0
index 6942e6b0eda6..88de4d6cd9bb 100644
@@ -50,6 +50,7 @@ extern int vgif;
 extern bool intercept_smi;
 extern bool vnmi;
 extern int lbrv;
+extern bool pml;
 
 extern int tsc_aux_uret_slot __ro_after_init;
 
@@ -746,6 +747,8 @@ static inline void svm_enable_intercept_for_msr(struct kvm_vcpu *vcpu,
 	svm_set_intercept_for_msr(vcpu, msr, type, true);
 }
 
+void svm_update_cpu_dirty_logging(struct kvm_vcpu *vcpu);
+
 /* nested.c */
 
 #define NESTED_EXIT_HOST	0	/* Exit handled on host level */
Signature · 1 line
2.48.1
#12 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Sashiko reported a couple of issues [1]. Let me address them here:

Quoted 15 lines · Nikunj A Dadhania
arch/x86/kvm/svm/svm.c +10 -0
@@ -1206,6 +1209,16 @@ static void init_vmcb(struct kvm_vcpu *vcpu, bool init_event)
 	if (vcpu->kvm->arch.bus_lock_detection_enabled)
 		svm_set_intercept(svm, INTERCEPT_BUSLOCK);
 
+	if (pml) {
+		/*
+		 * Populate the page address and index here, PML is enabled
+		 * when dirty logging is enabled on the memslot through
+		 * svm_update_cpu_dirty_logging()
+		 */
+		control->pml_addr = (u64)__sme_set(page_to_phys(vcpu->arch.pml_page));
+		control->pml_index = PML_HEAD_INDEX;
+	}
+
 
Quoted 3 lines

If the guest receives an INIT IPI and init_vmcb() is called to reset the vCPU, does unconditionally setting pml_index to PML_HEAD_INDEX discard any un-flushed dirty GPAs logged by the hardware?

There are two scenarios where init_vmcb() is called:

1) During vCPU creation time, where we need to set pml_index to PML_HEAD_INDEX
2) During vCPU reset, when init_event=true

Before vCPU reset:
vcpu_enter_guest()
└─> kvm_x86_call(vcpu_run) [VMRUN]
└─> [guest executes, PML accumulates dirty pages] └─> VMEXIT
└─> svm_handle_exit() --> PML buffer flushed here └─> return to vcpu_run()

vCPU Reset:
vcpu_enter_guest()
├─> kvm_check_request(KVM_REQ_EVENT)
├─> kvm_apic_accept_events()
│ └─> kvm_vcpu_reset(..., true)
│ └─> init_vmcb(..., true)
│ └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
└─> kvm_x86_call(): Next VMRUN

Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Quoted 20 lines
arch/x86/kvm/svm/svm.c +15 -0
+static void svm_flush_pml_buffer(struct kvm_vcpu *vcpu)
+{
+	struct vcpu_svm *svm = to_svm(vcpu);
+	struct vmcb_control_area *control = &svm->vmcb->control;
+
+	/* Do nothing if PML buffer is empty */
+	if (control->pml_index == PML_HEAD_INDEX)
+		return;
+
+	kvm_flush_pml_buffer(vcpu, control->pml_index);
+
+	/* Reset the PML index */
+	control->pml_index = PML_HEAD_INDEX;
+}
+

Since the AMD SVM architecture aggressively caches VMCB control fields across VMRUN instructions using VMCB clean bits, will the hardware recognize this updated pml_index?

Quoted 7 lines

If the VMCB isn't explicitly marked as dirty here, for example by calling vmcb_mark_dirty(svm->vmcb, VMCB_NPT), could the CPU continue using its cached index?

That cached index would be less than 0 after a PML_FULL exit, so would the CPU immediately trigger another PML_FULL exit on the next logged write and cause an infinite VM-Exit loop?

I got clarification from the design team: No clean bits were defined for the PML_ADDR and PML_INDEX fields, so on a VMRUN, microcode always loads them from the VMCB if the feature is enabled.

Regards,
Nikunj

1. https://sashiko.dev/#/patchset/20260407063245.2755579-1-nikunj%40amd.com

#13 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 18 lines · Nikunj A. Dadhania

Sashiko reported a couple of issues [1]. Let me address them here:

Quoted 14 lines · Nikunj A Dadhania
(quoted diff) +10 -0
@@ -1206,6 +1209,16 @@ static void init_vmcb(struct kvm_vcpu *vcpu, bool init_event)
 	if (vcpu->kvm->arch.bus_lock_detection_enabled)
 		svm_set_intercept(svm, INTERCEPT_BUSLOCK);
 
+	if (pml) {
+		/*
+		 * Populate the page address and index here, PML is enabled
+		 * when dirty logging is enabled on the memslot through
+		 * svm_update_cpu_dirty_logging()
+		 */
+		control->pml_addr = (u64)__sme_set(page_to_phys(vcpu->arch.pml_page));
+		control->pml_index = PML_HEAD_INDEX;
+	}
+
Quoted 8 lines
Quoted 3 lines

If the guest receives an INIT IPI and init_vmcb() is called to reset the vCPU, does unconditionally setting pml_index to PML_HEAD_INDEX discard any un-flushed dirty GPAs logged by the hardware?

There are two scenarios where init_vmcb() is called:

1) During vCPU creation time, where we need to set pml_index to PML_HEAD_INDEX
2) During vCPU reset, when init_event=true

Nit, please just call this INIT, even though KVM calls the function "reset". #1 is KVM's one and only true RESET flow.

Quoted 17 lines

Before vCPU reset:
vcpu_enter_guest()
└─> kvm_x86_call(vcpu_run) [VMRUN]
└─> [guest executes, PML accumulates dirty pages] └─> VMEXIT
└─> svm_handle_exit() --> PML buffer flushed here └─> return to vcpu_run()

vCPU Reset:
vcpu_enter_guest()
├─> kvm_check_request(KVM_REQ_EVENT)
├─> kvm_apic_accept_events()
│ └─> kvm_vcpu_reset(..., true)
│ └─> init_vmcb(..., true)
│ └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
└─> kvm_x86_call(): Next VMRUN

Quoted 5 lines
Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Huh. There's a pre-existing bug here. Commit f7f39c50edb9 ("KVM: x86: Exit to userspace if fastpath triggers one on instruction skip") added a path that skips kvm_x86_ops.handle_exit(), and specifically can give userspace control without going through vmx_flush_pml_buffer():

	if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
		return 0;

	r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);

Given that SVM support for PML is (obviously) on its way, it's mildly tempting to add a dedicated kvm_x86_ops hook to flush the buffer on a fastpath userspace exit. But, I dislike one-off kvm_x86_ops hooks, and that only works if there's no other vendor action required. E.g. very theoretically, a fastpath userspace exit could also be coincident with bus_lock_detected.

Yikes! And I think userspace could see a stale CR0 and/or CR3 on AMD. Hmm, but waiting until the full exit path to grab CR0 and CR3 is flawed on its own, e.g. it's a bug waiting to happen if KVM consumes CR3 in the fastpath.

So over two patches, I think the fix for those issues is the below. I'll test and send a mini-series. I don't think there's anything you need to do; I can resolve the resulting conflict easy enough.

x86/kvm/svm/svm.c +8 -7
index e7fdd7a9c280..df0bd132edf7 100644
@@ -3644,13 +3644,8 @@ static int svm_handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
        struct vcpu_svm *svm = to_svm(vcpu);
        struct kvm_run *kvm_run = vcpu->run;
 
-       /* SEV-ES guests must use the CR write traps to track CR registers. */
-       if (!is_sev_es_guest(vcpu)) {
-               if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
-                       vcpu->arch.cr0 = svm->vmcb->save.cr0;
-               if (npt_enabled)
-                       vcpu->arch.cr3 = svm->vmcb->save.cr3;
-       }
+       if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
+               return 0;
 
        if (is_guest_mode(vcpu)) {
                int vmexit;
@@ -4502,11 +4497,17 @@ static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu, u64 run_flags)
        if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
                x86_spec_ctrl_restore_host(svm->virt_spec_ctrl);
 
+       /* SEV-ES guests must use the CR write traps to track CR registers. */
        if (!is_sev_es_guest(vcpu)) {
                vcpu->arch.cr2 = svm->vmcb->save.cr2;
                vcpu->arch.regs[VCPU_REGS_RAX] = svm->vmcb->save.rax;
                vcpu->arch.regs[VCPU_REGS_RSP] = svm->vmcb->save.rsp;
                vcpu->arch.regs[VCPU_REGS_RIP] = svm->vmcb->save.rip;
+
+               if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
+                       vcpu->arch.cr0 = svm->vmcb->save.cr0;
+               if (npt_enabled)
+                       vcpu->arch.cr3 = svm->vmcb->save.cr3;
        }
        vcpu->arch.regs_dirty = 0;
 
x86/kvm/vmx/vmx.c +3 -0
index a29896a9ef14..4cb355ecfe46 100644
@@ -6687,6 +6687,9 @@ static int __vmx_handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
        if (enable_pml && !is_guest_mode(vcpu))
                vmx_flush_pml_buffer(vcpu);
 
+       if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
+               return 0;
+
        /*
         * KVM should never reach this point with a pending nested VM-Enter.
         * More specifically, short-circuiting VM-Entry to emulate L2 due to
x86/kvm/x86.c +0 -3
index 0a1b63c63d1a..9ad7ec3bf0f1 100644
@@ -11602,9 +11602,6 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
        if (vcpu->arch.apic_attention)
                kvm_lapic_sync_from_vapic(vcpu);
 
-       if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
-               return 0;
-
        r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);
        return r;
#14 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 25 lines · Sean Christopherson
Quoted 9 lines

   vCPU Reset:
     vcpu_enter_guest()

       ├─> kvm_check_request(KVM_REQ_EVENT)
       ├─> kvm_apic_accept_events()
       │     └─> kvm_vcpu_reset(..., true)
       │           └─> init_vmcb(..., true)
       │                 └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
       └─> kvm_x86_call(): Next VMRUN
Quoted 5 lines
Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Huh.  There's a pre-existing bug here.  Commit f7f39c50edb9 ("KVM: x86: Exit to userspace if fastpath triggers one on instruction skip") added a path that skips kvm_x86_ops.handle_exit(), and specifically can give userspace control without going through vmx_flush_pml_buffer():

	if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
		return 0;

	r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);
Quoted 5 lines

Given that SVM support for PML is (obviously) on its way, it's mildly tempting to add a dedicated kvm_x86_ops hook to flush the buffer on a fastpath userspace exit.  But, I dislike one-off kvm_x86_ops hooks, and that only works if there's no other vendor action required.  E.g. very theoretically, a fastpath userspace exit could also be coincident with bus_lock_detected.

Seems vmx_vcpu_reset() doesn't reset PML index upon INIT event, which seems to be fine since we are not losing any dirty GPA tracking AFAICT (otherwise we already have a bug for VMX here)?

How about doing the same for SVM?

Alternatively, we can always manually flush PML buffer upon INIT before resetting PML index. This still allows to do the EXIT_FASTPATH_EXIT_USERSPACE check in the common code.

Not sure which way is better, though.

#15 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 32 lines · Kai Huang
Quoted 25 lines · Sean Christopherson
Quoted 9 lines

   vCPU Reset:
     vcpu_enter_guest()

       ├─> kvm_check_request(KVM_REQ_EVENT)
       ├─> kvm_apic_accept_events()
       │     └─> kvm_vcpu_reset(..., true)
       │           └─> init_vmcb(..., true)
       │                 └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
       └─> kvm_x86_call(): Next VMRUN
Quoted 5 lines
Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Huh.  There's a pre-existing bug here.  Commit f7f39c50edb9 ("KVM: x86: Exit to userspace if fastpath triggers one on instruction skip") added a path that skips kvm_x86_ops.handle_exit(), and specifically can give userspace control without going through vmx_flush_pml_buffer():

	if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
		return 0;

	r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);
Quoted 5 lines

Given that SVM support for PML is (obviously) on its way, it's mildly tempting to add a dedicated kvm_x86_ops hook to flush the buffer on a fastpath userspace exit.  But, I dislike one-off kvm_x86_ops hooks, and that only works if there's no other vendor action required.  E.g. very theoretically, a fastpath userspace exit could also be coincident with bus_lock_detected.

Quoted 5 lines

Seems vmx_vcpu_reset() doesn't reset PML index upon INIT event, which seems to be fine since we are not losing any dirty GPA tracking AFAICT (otherwise we already have a bug for VMX here)?

How about doing the same for SVM?

We don't really have that luxury. On SHUTDOWN (even intercepted SHUTDOWN), the state of the VMCB is technically undefined. I.e. KVM needs to write _something_.

Hmm, actually, how is that going to work? Dropping PML entries just because a vCPU hit SHUTDOWN isn't going to fly. E.g. if a VM crashes while it's being migrated by the host, then it could end up with corrupted, incoherent data on the target due to leaving dirtied pages behind.

Quoted 5 lines

Alternatively, we can always manually flush PML buffer upon INIT before resetting PML index. This still allows to do the EXIT_FASTPATH_EXIT_USERSPACE check in the common code.

Not sure which way is better, though.

#16 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 33 lines · Sean Christopherson
Quoted 32 lines · Kai Huang
Quoted 30 lines · Sean Christopherson
Quoted 9 lines

   vCPU Reset:
     vcpu_enter_guest()

       ├─> kvm_check_request(KVM_REQ_EVENT)
       ├─> kvm_apic_accept_events()
       │     └─> kvm_vcpu_reset(..., true)
       │           └─> init_vmcb(..., true)
       │                 └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
       └─> kvm_x86_call(): Next VMRUN
Quoted 5 lines
Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Huh.  There's a pre-existing bug here.  Commit f7f39c50edb9 ("KVM: x86: Exit to userspace if fastpath triggers one on instruction skip") added a path that skips kvm_x86_ops.handle_exit(), and specifically can give userspace control without going through vmx_flush_pml_buffer():

	if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
		return 0;

	r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);

Given that SVM support for PML is (obviously) on its way, it's mildly tempting to add a dedicated kvm_x86_ops hook to flush the buffer on a fastpath userspace exit.  But, I dislike one-off kvm_x86_ops hooks, and that only works if there's no other vendor action required.  E.g. very theoretically, a fastpath userspace exit could also be coincident with bus_lock_detected.

Quoted 8 lines
Quoted 5 lines

Seems vmx_vcpu_reset() doesn't reset PML index upon INIT event, which seems to be fine since we are not losing any dirty GPA tracking AFAICT (otherwise we already have a bug for VMX here)?

How about doing the same for SVM?

We don't really have that luxury. On SHUTDOWN (even intercepted SHUTDOWN), the state of the VMCB is technically undefined. I.e. KVM needs to write _something_.

You mean KVM needs to reset VMCB to reflect the architecturally defined INIT state for a vCPU? Or the hardware itself may reset VMCB thus may reset PML index?

Quoted 4 lines

Hmm, actually, how is that going to work? Dropping PML entries just because a vCPU hit SHUTDOWN isn't going to fly.

Not dropping, but just leave PML index unchanged. The PML buffer itself is still there unchanged, and PML is still working in hardware thus the buffer will eventually get flushed.

This is the case for VMX AFAICT, thus I wonder whether this also works for SVM.

Quoted 3 lines

E.g. if a VM crashes while it's being
migrated by the host, then it could end up with corrupted, incoherent data on the target due to leaving dirtied pages behind.

If we leave PML index unchanged upon INIT as mentioned above, I don't think we will lose any dirty GPA tracking. But maybe I am missing something.

#17 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 23 lines · Kai
Quoted 22 lines · Sean Christopherson
Quoted 21 lines · Kai Huang
Quoted 20 lines · Sean Christopherson
Quoted 9 lines

   vCPU Reset:
     vcpu_enter_guest()

       ├─> kvm_check_request(KVM_REQ_EVENT)
       ├─> kvm_apic_accept_events()
       │     └─> kvm_vcpu_reset(..., true)
       │           └─> init_vmcb(..., true)
       │                 └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
       └─> kvm_x86_call(): Next VMRUN
Quoted 5 lines
Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Huh.  There's a pre-existing bug here.  Commit f7f39c50edb9 ("KVM: x86: Exit to userspace if fastpath triggers one on instruction skip") added a path that skips kvm_x86_ops.handle_exit(), and specifically can give userspace control without going through vmx_flush_pml_buffer():

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
	if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
		return 0;

	r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);

Ah right.

Quoted 24 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

Given that SVM support for PML is (obviously) on its way, it's mildly tempting to add a dedicated kvm_x86_ops hook to flush the buffer on a fastpath userspace exit.  But, I dislike one-off kvm_x86_ops hooks, and that only works if there's no other vendor action required.  E.g. very theoretically, a fastpath userspace exit could also be coincident with bus_lock_detected.

Quoted 8 lines
Quoted 5 lines

Seems vmx_vcpu_reset() doesn't reset PML index upon INIT event, which seems to be fine since we are not losing any dirty GPA tracking AFAICT (otherwise we already have a bug for VMX here)?

How about doing the same for SVM?

We don't really have that luxury. On SHUTDOWN (even intercepted SHUTDOWN), the state of the VMCB is technically undefined. I.e. KVM needs to write _something_.

You mean KVM needs to reset VMCB to reflect the architecturally defined INIT state for a vCPU? Or the hardware itself may reset VMCB thus may reset PML index?

Quoted 4 lines

Hmm, actually, how is that going to work? Dropping PML entries just because a vCPU hit SHUTDOWN isn't going to fly.

Quoted 6 lines

Not dropping, but just leave PML index unchanged. The PML buffer itself is still there unchanged, and PML is still working in hardware thus the buffer will eventually get flushed.

This is the case for VMX AFAICT, thus I wonder whether this also works for SVM.

Theoretically, while the migration is active and we did a fast path user space exit and never returned, the entries remain in PML buffer. If we flush every time on VM exit, as per Sean's suggestion, PML buffer is always flushed before returning to user space.

Quoted 7 lines
Quoted 4 lines

E.g. if a VM crashes while it's being
migrated by the host, then it could end up with corrupted, incoherent data on the target due to leaving dirtied pages behind.

If we leave PML index unchanged upon INIT as mentioned above, I don't think we will lose any dirty GPA tracking. But maybe I am missing something.

Regards
Nikunj

#18 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 60 lines · Nikunj A. Dadhania
Quoted 23 lines · Kai
Quoted 22 lines · Sean Christopherson
Quoted 21 lines · Kai Huang
Quoted 20 lines · Sean Christopherson
Quoted 9 lines

   vCPU Reset:
     vcpu_enter_guest()

       ├─> kvm_check_request(KVM_REQ_EVENT)
       ├─> kvm_apic_accept_events()
       │     └─> kvm_vcpu_reset(..., true)
       │           └─> init_vmcb(..., true)
       │                 └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
       └─> kvm_x86_call(): Next VMRUN
Quoted 5 lines
Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Huh.  There's a pre-existing bug here.  Commit f7f39c50edb9 ("KVM: x86: Exit to userspace if fastpath triggers one on instruction skip") added a path that skips kvm_x86_ops.handle_exit(), and specifically can give userspace control without going through vmx_flush_pml_buffer():

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
	if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
		return 0;

	r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);

Ah right.

Quoted 28 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

Given that SVM support for PML is (obviously) on its way, it's mildly tempting to add a dedicated kvm_x86_ops hook to flush the buffer on a fastpath userspace exit.  But, I dislike one-off kvm_x86_ops hooks, and that only works if there's no other vendor action required.  E.g. very theoretically, a fastpath userspace exit could also be coincident with bus_lock_detected.

Quoted 8 lines
Quoted 5 lines

Seems vmx_vcpu_reset() doesn't reset PML index upon INIT event, which seems to be fine since we are not losing any dirty GPA tracking AFAICT (otherwise we already have a bug for VMX here)?

How about doing the same for SVM?

We don't really have that luxury. On SHUTDOWN (even intercepted SHUTDOWN), the state of the VMCB is technically undefined. I.e. KVM needs to write _something_.

You mean KVM needs to reset VMCB to reflect the architecturally defined INIT state for a vCPU? Or the hardware itself may reset VMCB thus may reset PML index?

Quoted 4 lines

Hmm, actually, how is that going to work? Dropping PML entries just because a vCPU hit SHUTDOWN isn't going to fly.

Not dropping, but just leave PML index unchanged. The PML buffer itself is still there unchanged, and PML is still working in hardware thus the buffer will eventually get flushed.

Quoted 7 lines
Quoted 2 lines

This is the case for VMX AFAICT, thus I wonder whether this also works for SVM.

Theoretically, while the migration is active and we did a fast path user space exit and never returned, the entries remain in PML buffer. If we flush every time on VM exit, as per Sean's suggestion, PML buffer is always flushed before returning to user space.

Oh indeed I missed something here. I assumed the last GET_DIRTY_LOG ioctl can always flush the PML buffer. But I forgot KVM flushes the buffer by kicking all vcpus.

So yeah we have a loophole on the EXIT_FASTPATH_EXIT_USERSPACE, which seems can be controlled by userspace (e.g., single step debugging?).

So what Sean posted makes sense. Sorry for the noise.

Or maybe we could simply flush PML buffer in {vmx|svm}_vcpu_enter_exit()?

#19 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

In reply to Huang, Kai
Quoted 38 lines · Kai Huang
Quoted 37 lines · Sean Christopherson
Quoted 36 lines · Kai Huang
Quoted 25 lines · Sean Christopherson
Quoted 9 lines

   vCPU Reset:
     vcpu_enter_guest()

       ├─> kvm_check_request(KVM_REQ_EVENT)
       ├─> kvm_apic_accept_events()
       │     └─> kvm_vcpu_reset(..., true)
       │           └─> init_vmcb(..., true)
       │                 └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
       └─> kvm_x86_call(): Next VMRUN
Quoted 5 lines
Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Huh.  There's a pre-existing bug here.  Commit f7f39c50edb9 ("KVM: x86: Exit to userspace if fastpath triggers one on instruction skip") added a path that skips kvm_x86_ops.handle_exit(), and specifically can give userspace control without going through vmx_flush_pml_buffer():

	if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
		return 0;

	r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);
Quoted 5 lines

Given that SVM support for PML is (obviously) on its way, it's mildly tempting to add a dedicated kvm_x86_ops hook to flush the buffer on a fastpath userspace exit.  But, I dislike one-off kvm_x86_ops hooks, and that only works if there's no other vendor action required.  E.g. very theoretically, a fastpath userspace exit could also be coincident with bus_lock_detected.

Seems vmx_vcpu_reset() doesn't reset PML index upon INIT event, which seems to be fine since we are not losing any dirty GPA tracking AFAICT (otherwise we already have a bug for VMX here)?

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

How about doing the same for SVM?

We don't really have that luxury. On SHUTDOWN (even intercepted SHUTDOWN), the state of the VMCB is technically undefined. I.e. KVM needs to write _something_.

You mean KVM needs to reset VMCB to reflect the architecturally defined INIT state for a vCPU? Or the hardware itself may reset VMCB thus may reset PML index?

Neither. The APM states that the VMCB is undefined after SHUTDOWN. PML index
could be anything:

15.14.3 Shutdown Intercept
When this intercept occurs, any condition that normally causes a shutdown causes a #VMEXIT to the VMM instead. After an intercepted shutdown, the state saved in the VMCB is undefined.

KVM synthesizes an INIT because it's the least awful option.

static int shutdown_interception(struct kvm_vcpu *vcpu) {
struct kvm_run *kvm_run = vcpu->run; struct vcpu_svm *svm = to_svm(vcpu);

/*
* VMCB is undefined after a SHUTDOWN intercept. INIT the vCPU to put
* the VMCB in a known good state. Unfortuately, KVM doesn't have
* KVM_MP_STATE_SHUTDOWN and can't add it without potentially breaking
* userspace. At a platform view, INIT is acceptable behavior as
* there exist bare metal platforms that automatically INIT the CPU
* in response to shutdown.
*
* The VM save area for SEV-ES guests has already been encrypted so it
* cannot be reinitialized, i.e. synthesizing INIT is futile. */
if (!is_sev_es_guest(vcpu)) {
clear_page(svm->vmcb);

#ifdef CONFIG_KVM_SMM
		if (is_smm(vcpu))
			kvm_smm_changed(vcpu, false);
#endif
		kvm_vcpu_reset(vcpu, true);
	}

	kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
	return 0;

}

Now, maybe the APM is trying to say only the save area is undefined, in which case PML Index is fine and can and should be left alone. But if that's the case, the APM needs to be updated to make explicitly clear what fields in the VMCS are and are not valid after SHUTDOWN.

Quoted 7 lines
Quoted 3 lines

Hmm, actually, how is that going to work? Dropping PML entries just because a vCPU hit SHUTDOWN isn't going to fly.

Not dropping, but just leave PML index unchanged. The PML buffer itself is still there unchanged, and PML is still working in hardware thus the buffer will eventually get flushed.

As above, the PML index could be garbage.

Quoted 2 lines

This is the case for VMX AFAICT, thus I wonder whether this also works for SVM.

No, VMX is fine.

Quoted 6 lines
Quoted 3 lines

E.g. if a VM crashes while it's being
migrated by the host, then it could end up with corrupted, incoherent data on the target due to leaving dirtied pages behind.

If we leave PML index unchanged upon INIT as mentioned above, I don't think we will lose any dirty GPA tracking. But maybe I am missing something.

Correct. The question is whether or not AMD's architecture guarantees PML Index will be valid/accurate after SHUTDOWN.

#20 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 85 lines · Sean Christopherson
Quoted 38 lines · Kai Huang
Quoted 37 lines · Sean Christopherson
Quoted 36 lines · Kai Huang
Quoted 25 lines · Sean Christopherson
Quoted 9 lines

   vCPU Reset:
     vcpu_enter_guest()

       ├─> kvm_check_request(KVM_REQ_EVENT)
       ├─> kvm_apic_accept_events()
       │     └─> kvm_vcpu_reset(..., true)
       │           └─> init_vmcb(..., true)
       │                 └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
       └─> kvm_x86_call(): Next VMRUN
Quoted 5 lines
Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Huh.  There's a pre-existing bug here.  Commit f7f39c50edb9 ("KVM: x86: Exit to userspace if fastpath triggers one on instruction skip") added a path that skips kvm_x86_ops.handle_exit(), and specifically can give userspace control without going through vmx_flush_pml_buffer():

	if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
		return 0;

	r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);
Quoted 5 lines

Given that SVM support for PML is (obviously) on its way, it's mildly tempting to add a dedicated kvm_x86_ops hook to flush the buffer on a fastpath userspace exit.  But, I dislike one-off kvm_x86_ops hooks, and that only works if there's no other vendor action required.  E.g. very theoretically, a fastpath userspace exit could also be coincident with bus_lock_detected.

Seems vmx_vcpu_reset() doesn't reset PML index upon INIT event, which seems to be fine since we are not losing any dirty GPA tracking AFAICT (otherwise we already have a bug for VMX here)?

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

How about doing the same for SVM?

We don't really have that luxury. On SHUTDOWN (even intercepted SHUTDOWN), the state of the VMCB is technically undefined. I.e. KVM needs to write _something_.

You mean KVM needs to reset VMCB to reflect the architecturally defined INIT state for a vCPU? Or the hardware itself may reset VMCB thus may reset PML index?

Neither. The APM states that the VMCB is undefined after SHUTDOWN. PML index
could be anything:

15.14.3 Shutdown Intercept
When this intercept occurs, any condition that normally causes a shutdown causes a #VMEXIT to the VMM instead. After an intercepted shutdown, the state saved in the VMCB is undefined.

KVM synthesizes an INIT because it's the least awful option.

static int shutdown_interception(struct kvm_vcpu *vcpu) {
struct kvm_run *kvm_run = vcpu->run; struct vcpu_svm *svm = to_svm(vcpu);

/*
* VMCB is undefined after a SHUTDOWN intercept. INIT the vCPU to put
* the VMCB in a known good state. Unfortuately, KVM doesn't have
* KVM_MP_STATE_SHUTDOWN and can't add it without potentially breaking
* userspace. At a platform view, INIT is acceptable behavior as
* there exist bare metal platforms that automatically INIT the CPU
* in response to shutdown.
*
* The VM save area for SEV-ES guests has already been encrypted so it
* cannot be reinitialized, i.e. synthesizing INIT is futile. */
if (!is_sev_es_guest(vcpu)) {
clear_page(svm->vmcb);

#ifdef CONFIG_KVM_SMM
		if (is_smm(vcpu))
			kvm_smm_changed(vcpu, false);
#endif
		kvm_vcpu_reset(vcpu, true);
	}

	kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
	return 0;
Quoted 7 lines

}

Now, maybe the APM is trying to say only the save area is undefined, in which case PML Index is fine and can and should be left alone. But if that's the case, the APM needs to be updated to make explicitly clear what fields in the VMCS are and are not valid after SHUTDOWN.

Quoted 1 line
Quoted 1 line

Thanks for all the info! :-)

#21 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

In reply to Sean Christopherson
Quoted 85 lines · Sean Christopherson
Quoted 38 lines · Kai Huang
Quoted 37 lines · Sean Christopherson
Quoted 36 lines · Kai Huang
Quoted 25 lines · Sean Christopherson
Quoted 9 lines

   vCPU Reset:
     vcpu_enter_guest()

       ├─> kvm_check_request(KVM_REQ_EVENT)
       ├─> kvm_apic_accept_events()
       │     └─> kvm_vcpu_reset(..., true)
       │           └─> init_vmcb(..., true)
       │                 └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
       └─> kvm_x86_call(): Next VMRUN
Quoted 5 lines
Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Huh.  There's a pre-existing bug here.  Commit f7f39c50edb9 ("KVM: x86: Exit to userspace if fastpath triggers one on instruction skip") added a path that skips kvm_x86_ops.handle_exit(), and specifically can give userspace control without going through vmx_flush_pml_buffer():

	if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
		return 0;

	r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);
Quoted 5 lines

Given that SVM support for PML is (obviously) on its way, it's mildly tempting to add a dedicated kvm_x86_ops hook to flush the buffer on a fastpath userspace exit.  But, I dislike one-off kvm_x86_ops hooks, and that only works if there's no other vendor action required.  E.g. very theoretically, a fastpath userspace exit could also be coincident with bus_lock_detected.

Seems vmx_vcpu_reset() doesn't reset PML index upon INIT event, which seems to be fine since we are not losing any dirty GPA tracking AFAICT (otherwise we already have a bug for VMX here)?

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

How about doing the same for SVM?

We don't really have that luxury. On SHUTDOWN (even intercepted SHUTDOWN), the state of the VMCB is technically undefined. I.e. KVM needs to write _something_.

You mean KVM needs to reset VMCB to reflect the architecturally defined INIT state for a vCPU? Or the hardware itself may reset VMCB thus may reset PML index?

Neither. The APM states that the VMCB is undefined after SHUTDOWN. PML index
could be anything:

15.14.3 Shutdown Intercept
When this intercept occurs, any condition that normally causes a shutdown causes a #VMEXIT to the VMM instead. After an intercepted shutdown, the state saved in the VMCB is undefined.

KVM synthesizes an INIT because it's the least awful option.

static int shutdown_interception(struct kvm_vcpu *vcpu) {
struct kvm_run *kvm_run = vcpu->run; struct vcpu_svm *svm = to_svm(vcpu);

/*
* VMCB is undefined after a SHUTDOWN intercept. INIT the vCPU to put
* the VMCB in a known good state. Unfortuately, KVM doesn't have
* KVM_MP_STATE_SHUTDOWN and can't add it without potentially breaking
* userspace. At a platform view, INIT is acceptable behavior as
* there exist bare metal platforms that automatically INIT the CPU
* in response to shutdown.
*
* The VM save area for SEV-ES guests has already been encrypted so it
* cannot be reinitialized, i.e. synthesizing INIT is futile. */
if (!is_sev_es_guest(vcpu)) {
clear_page(svm->vmcb);

#ifdef CONFIG_KVM_SMM
		if (is_smm(vcpu))
			kvm_smm_changed(vcpu, false);
#endif
		kvm_vcpu_reset(vcpu, true);
	}

	kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
	return 0;
Quoted 6 lines

}

Now, maybe the APM is trying to say only the save area is undefined, in which case PML Index is fine and can and should be left alone. But if that's the case, the APM needs to be updated to make explicitly clear what fields in the VMCS are and are not valid after SHUTDOWN.

I'm checking to see if I can get clarification on this.

Thanks,
Tom

Quoted 26 lines
Quoted 8 lines
Quoted 4 lines

Hmm, actually, how is that going to work? Dropping PML entries just because a vCPU hit SHUTDOWN isn't going to fly.

Not dropping, but just leave PML index unchanged. The PML buffer itself is still there unchanged, and PML is still working in hardware thus the buffer will eventually get flushed.

As above, the PML index could be garbage.

Quoted 2 lines

This is the case for VMX AFAICT, thus I wonder whether this also works for SVM.

No, VMX is fine.

Quoted 6 lines
Quoted 3 lines

E.g. if a VM crashes while it's being
migrated by the host, then it could end up with corrupted, incoherent data on the target due to leaving dirtied pages behind.

If we leave PML index unchanged upon INIT as mentioned above, I don't think we will lose any dirty GPA tracking. But maybe I am missing something.

Correct. The question is whether or not AMD's architecture guarantees PML Index will be valid/accurate after SHUTDOWN.

#22 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 88 lines · Tom Lendacky
Quoted 87 lines · Sean Christopherson
Quoted 38 lines · Kai Huang
Quoted 37 lines · Sean Christopherson
Quoted 36 lines · Kai Huang
Quoted 25 lines · Sean Christopherson
Quoted 9 lines

   vCPU Reset:
     vcpu_enter_guest()

       ├─> kvm_check_request(KVM_REQ_EVENT)
       ├─> kvm_apic_accept_events()
       │     └─> kvm_vcpu_reset(..., true)
       │           └─> init_vmcb(..., true)
       │                 └─> control->pml_index = PML_HEAD_INDEX -- PML buffer was already flushed
       └─> kvm_x86_call(): Next VMRUN
Quoted 5 lines
Quoted 3 lines

Could this result in the hypervisor losing track of dirty memory during live migration, leading to memory corruption on the destination host, since svm_flush_pml_buffer() isn't called before resetting the index?

AFAIU, no. The PML buffer is always flushed opportunistically at every VM exit.

Huh.  There's a pre-existing bug here.  Commit f7f39c50edb9 ("KVM: x86: Exit to userspace if fastpath triggers one on instruction skip") added a path that skips kvm_x86_ops.handle_exit(), and specifically can give userspace control without going through vmx_flush_pml_buffer():

	if (unlikely(exit_fastpath == EXIT_FASTPATH_EXIT_USERSPACE))
		return 0;

	r = kvm_x86_call(handle_exit)(vcpu, exit_fastpath);
Quoted 5 lines

Given that SVM support for PML is (obviously) on its way, it's mildly tempting to add a dedicated kvm_x86_ops hook to flush the buffer on a fastpath userspace exit.  But, I dislike one-off kvm_x86_ops hooks, and that only works if there's no other vendor action required.  E.g. very theoretically, a fastpath userspace exit could also be coincident with bus_lock_detected.

Seems vmx_vcpu_reset() doesn't reset PML index upon INIT event, which seems to be fine since we are not losing any dirty GPA tracking AFAICT (otherwise we already have a bug for VMX here)?

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

How about doing the same for SVM?

We don't really have that luxury. On SHUTDOWN (even intercepted SHUTDOWN), the state of the VMCB is technically undefined. I.e. KVM needs to write _something_.

You mean KVM needs to reset VMCB to reflect the architecturally defined INIT state for a vCPU? Or the hardware itself may reset VMCB thus may reset PML index?

Neither. The APM states that the VMCB is undefined after SHUTDOWN. PML index
could be anything:

15.14.3 Shutdown Intercept
When this intercept occurs, any condition that normally causes a shutdown causes a #VMEXIT to the VMM instead. After an intercepted shutdown, the state saved in the VMCB is undefined.

KVM synthesizes an INIT because it's the least awful option.

static int shutdown_interception(struct kvm_vcpu *vcpu) {
struct kvm_run *kvm_run = vcpu->run; struct vcpu_svm *svm = to_svm(vcpu);

/*
* VMCB is undefined after a SHUTDOWN intercept. INIT the vCPU to put
* the VMCB in a known good state. Unfortuately, KVM doesn't have
* KVM_MP_STATE_SHUTDOWN and can't add it without potentially breaking
* userspace. At a platform view, INIT is acceptable behavior as
* there exist bare metal platforms that automatically INIT the CPU
* in response to shutdown.
*
* The VM save area for SEV-ES guests has already been encrypted so it
* cannot be reinitialized, i.e. synthesizing INIT is futile. */
if (!is_sev_es_guest(vcpu)) {
clear_page(svm->vmcb);

#ifdef CONFIG_KVM_SMM
		if (is_smm(vcpu))
			kvm_smm_changed(vcpu, false);
#endif
		kvm_vcpu_reset(vcpu, true);
	}

	kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
	return 0;

}

Quoted 6 lines
Quoted 4 lines

Now, maybe the APM is trying to say only the save area is undefined, in which case PML Index is fine and can and should be left alone. But if that's the case, the APM needs to be updated to make explicitly clear what fields in the VMCS are and are not valid after SHUTDOWN.

I'm checking to see if I can get clarification on this.

PML_INDEX will be valid even when a SHUTDOWN is intercepted. I'm working to see if we can update the APM to identify what would be valid and what wouldn't be in this situation.

For a triple-fault SHUTDOWN, the saved x86 state could be inconsistent because the CPU stopped in the middle of a sequence and redirected to shutdown. Everything will be saved, but it might not be valid x86 state. The question then becomes are there fields within the VMCB control area that are based on x86 state that might not be valid (e.g. V_IRQ or nRIP, etc.).

Thanks,
Tom

Quoted 32 lines

Thanks,
Tom

Quoted 26 lines
Quoted 8 lines
Quoted 4 lines

Hmm, actually, how is that going to work? Dropping PML entries just because a vCPU hit SHUTDOWN isn't going to fly.

Not dropping, but just leave PML index unchanged. The PML buffer itself is still there unchanged, and PML is still working in hardware thus the buffer will eventually get flushed.

As above, the PML index could be garbage.

Quoted 2 lines

This is the case for VMX AFAICT, thus I wonder whether this also works for SVM.

No, VMX is fine.

Quoted 6 lines
Quoted 3 lines

E.g. if a VM crashes while it's being
migrated by the host, then it could end up with corrupted, incoherent data on the target due to leaving dirtied pages behind.

If we leave PML index unchanged upon INIT as mentioned above, I don't think we will lose any dirty GPA tracking. But maybe I am missing something.

Correct. The question is whether or not AMD's architecture guarantees PML Index will be valid/accurate after SHUTDOWN.

#23 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 8 lines · Tom Lendacky
Quoted 6 lines · Tom Lendacky
Quoted 4 lines

Now, maybe the APM is trying to say only the save area is undefined, in which case PML Index is fine and can and should be left alone. But if that's the case, the APM needs to be updated to make explicitly clear what fields in the VMCS are and are not valid after SHUTDOWN.

I'm checking to see if I can get clarification on this.

Quoted 7 lines

PML_INDEX will be valid even when a SHUTDOWN is intercepted. I'm working to see if we can update the APM to identify what would be valid and what wouldn't be in this situation.

For a triple-fault SHUTDOWN, the saved x86 state could be inconsistent because the CPU stopped in the middle of a sequence and redirected to shutdown. Everything will be saved, but it might not be valid x86 state.

Yeah, I assumed that the APM wording was an overly-broad description of why guest state would be garbage.

Quoted 3 lines

The question then becomes are there fields within the VMCB control area that are based on x86 state that might not be valid (e.g. V_IRQ or nRIP, etc.).

FWIW, even a conservative iset of "these fields/areas are guaranteed to be sane and consistent" would be super helpful. The big problem right now is that the wording reads like the entire GHCB is suspect.

#24 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 20 lines · Sean Christopherson
Quoted 8 lines · Tom Lendacky
Quoted 6 lines · Tom Lendacky
Quoted 4 lines

Now, maybe the APM is trying to say only the save area is undefined, in which case PML Index is fine and can and should be left alone. But if that's the case, the APM needs to be updated to make explicitly clear what fields in the VMCS are and are not valid after SHUTDOWN.

I'm checking to see if I can get clarification on this.

Quoted 7 lines

PML_INDEX will be valid even when a SHUTDOWN is intercepted. I'm working to see if we can update the APM to identify what would be valid and what wouldn't be in this situation.

For a triple-fault SHUTDOWN, the saved x86 state could be inconsistent because the CPU stopped in the middle of a sequence and redirected to shutdown. Everything will be saved, but it might not be valid x86 state.

Yeah, I assumed that the APM wording was an overly-broad description of why guest state would be garbage.

Quoted 7 lines
Quoted 3 lines

The question then becomes are there fields within the VMCB control area that are based on x86 state that might not be valid (e.g. V_IRQ or nRIP, etc.).

FWIW, even a conservative iset of "these fields/areas are guaranteed to be sane and consistent" would be super helpful. The big problem right now is that the wording reads like the entire GHCB is suspect.

I believe you meant VMCB above, not GHCB. :-)

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

The APM states that the VMCB is undefined after SHUTDOWN. PML index could be anything:

15.14.3 Shutdown Intercept
When this intercept occurs, any condition that normally causes a shutdown causes a #VMEXIT to the VMM instead. After an intercepted shutdown, the state saved in the VMCB is undefined.

The anticipated APM update will clarify the wording as follows:

"After an intercepted shutdown, the VMCB control area is valid (with the
exception of offsets 60h, 61h, and 68h) and the VMCB state save area is undefined."

Based on this clarification, I will initialize the PML index only during reset and leave it unchanged in the case of an INIT event.

arch/x86/kvm/svm/svm.c +3 -1
index d9dc725786e6..f5a17274868e 100644
@@ -1216,7 +1216,9 @@ static void init_vmcb(struct kvm_vcpu *vcpu, bool init_event)
 		 * svm_update_cpu_dirty_logging()
 		 */
 		control->pml_addr = (u64)__sme_set(page_to_phys(vcpu->arch.pml_page));
-		control->pml_index = PML_HEAD_INDEX;
+
+		if (!init_event)
+			control->pml_index = PML_HEAD_INDEX;
 	}
 
 	if (sev_guest(vcpu->kvm))

Regards,
Nikunj

#25 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

In reply to Nikunj A Dadhania

arch/x86/kvm/svm/svm.c static int bus_lock_exit(struct kvm_vcpu *vcpu) · line 3288

Quoted 12 lines
arch/x86/kvm/svm/svm.c +8 -0
@@ -3253,6 +3278,54 @@ static int bus_lock_exit(struct kvm_vcpu *vcpu)
 	return 0;
 }
 
+void svm_update_cpu_dirty_logging(struct kvm_vcpu *vcpu)
+{
+	struct vcpu_svm *svm = to_svm(vcpu);
+
+	if (WARN_ON_ONCE(!vcpu->kvm->arch.cpu_dirty_log_size))
+		return;
+
+	guard(svm_vmcb01)(vcpu);

Why do we need this? Why not update vmcb01 directly below?

IIUC VMX needs this because it updates the loaded VMCS, but the VMCB is just memory.

Quoted 14 lines
arch/x86/kvm/svm/svm.c +14 -0
+
+	/*
+	 * Note, nr_memslots_dirty_logging can be changed concurrent with this
+	 * code, but in that case another update request will be made and so
+	 * the guest will never run with a stale PML value.
+	 */
+	if (atomic_read(&vcpu->kvm->nr_memslots_dirty_logging))
+		svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_PML_ENABLE;
+	else
+		svm->vmcb->control.nested_ctl &= ~SVM_NESTED_CTL_PML_ENABLE;
+
+	vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
+}
+
#26 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 4 lines · Yosry Ahmed
Quoted 4 lines
arch/x86/kvm/svm/svm.c +0 -0
@@ -3253,6 +3278,54 @@ static int bus_lock_exit(struct kvm_vcpu *vcpu)
 	return 0;
 }
 
Quoted 10 lines

arch/x86/kvm/svm/svm.c static int bus_lock_exit(struct kvm_vcpu *vcpu) · line 3288

Quoted 8 lines
arch/x86/kvm/svm/svm.c +8 -0
+void svm_update_cpu_dirty_logging(struct kvm_vcpu *vcpu)
+{
+	struct vcpu_svm *svm = to_svm(vcpu);
+
+	if (WARN_ON_ONCE(!vcpu->kvm->arch.cpu_dirty_log_size))
+		return;
+
+	guard(svm_vmcb01)(vcpu);

Why do we need this? Why not update vmcb01 directly below?

Yeah, I don't see the point of patch 6. guard(svm_vmcb01) doesn't exist in the current code base because it's not needed.

Quoted 2 lines

IIUC VMX needs this because it updates the loaded VMCS, but the VMCB is just memory.

Yep. If VMWRITE could target an arbitrary VMCS pointer, VMX wouldn't need the helper either.

#27 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 5 lines · Sean Christopherson
Quoted 4 lines · Yosry Ahmed
Quoted 4 lines
arch/x86/kvm/svm/svm.c +0 -0
@@ -3253,6 +3278,54 @@ static int bus_lock_exit(struct kvm_vcpu *vcpu)
 	return 0;
 }
 
Quoted 10 lines
Quoted 10 lines

arch/x86/kvm/svm/svm.c static int bus_lock_exit(struct kvm_vcpu *vcpu) · line 3288

Quoted 8 lines
arch/x86/kvm/svm/svm.c +8 -0
+void svm_update_cpu_dirty_logging(struct kvm_vcpu *vcpu)
+{
+	struct vcpu_svm *svm = to_svm(vcpu);
+
+	if (WARN_ON_ONCE(!vcpu->kvm->arch.cpu_dirty_log_size))
+		return;
+
+	guard(svm_vmcb01)(vcpu);

Why do we need this? Why not update vmcb01 directly below?

Are you suggesting something like below:

void svm_update_cpu_dirty_logging(struct kvm_vcpu *vcpu)

{
	struct vcpu_svm *svm = to_svm(vcpu);
	struct vmcb *vmcb01 = svm->vmcb01.ptr;

	if (WARN_ON_ONCE(!vcpu->kvm->arch.cpu_dirty_log_size))
		return;

	/*
	 * Note, nr_memslots_dirty_logging can be changed concurrent with this
	 * code, but in that case another update request will be made and so
	 * the guest will never run with a stale PML value.
	 */
	if (atomic_read(&vcpu->kvm->nr_memslots_dirty_logging))
		vmcb01->control.nested_ctl |= SVM_NESTED_CTL_PML_ENABLE;
	else
		vmcb01->control.nested_ctl &= ~SVM_NESTED_CTL_PML_ENABLE;

	vmcb_mark_dirty(vmcb01, VMCB_NPT);
}
Quoted 3 lines

Yeah, I don't see the point of patch 6. guard(svm_vmcb01) doesn't exist in the current code base because it's not needed.

Sure, I can drop patch6. I wanted to keep the code similar.

Quoted 6 lines
Quoted 3 lines

IIUC VMX needs this because it updates the loaded VMCS, but the VMCB is just memory.

Yep. If VMWRITE could target an arbitrary VMCS pointer, VMX wouldn't need the helper either.

Regards
Nikunj

#28 lore raw

Re: [PATCH v6 7/7] KVM: SVM: Add Page modification logging support

Quoted 19 lines · Nikunj A. Dadhania
Quoted 5 lines · Sean Christopherson
Quoted 4 lines · Yosry Ahmed
Quoted 4 lines
arch/x86/kvm/svm/svm.c +0 -0
@@ -3253,6 +3278,54 @@ static int bus_lock_exit(struct kvm_vcpu *vcpu)
 	return 0;
 }
 
Quoted 10 lines
Quoted 10 lines

arch/x86/kvm/svm/svm.c static int bus_lock_exit(struct kvm_vcpu *vcpu) · line 3288

Quoted 8 lines
arch/x86/kvm/svm/svm.c +8 -0
+void svm_update_cpu_dirty_logging(struct kvm_vcpu *vcpu)
+{
+	struct vcpu_svm *svm = to_svm(vcpu);
+
+	if (WARN_ON_ONCE(!vcpu->kvm->arch.cpu_dirty_log_size))
+		return;
+
+	guard(svm_vmcb01)(vcpu);

Why do we need this? Why not update vmcb01 directly below?

Quoted 1 line

Are you suggesting something like below:

Yes.

Quoted 21 lines

void svm_update_cpu_dirty_logging(struct kvm_vcpu *vcpu)

{
	struct vcpu_svm *svm = to_svm(vcpu);
	struct vmcb *vmcb01 = svm->vmcb01.ptr;

	if (WARN_ON_ONCE(!vcpu->kvm->arch.cpu_dirty_log_size))
		return;

	/*
	 * Note, nr_memslots_dirty_logging can be changed concurrent with this
	 * code, but in that case another update request will be made and so
	 * the guest will never run with a stale PML value.
	 */
	if (atomic_read(&vcpu->kvm->nr_memslots_dirty_logging))
		vmcb01->control.nested_ctl |= SVM_NESTED_CTL_PML_ENABLE;
	else
		vmcb01->control.nested_ctl &= ~SVM_NESTED_CTL_PML_ENABLE;

	vmcb_mark_dirty(vmcb01, VMCB_NPT);
}