Skip to contentSkip to conversation

KVM: SVM: Bump asid_generation on CPU online to avoid ASID collision after hotplug

PATCH Applied Nikunj A Dadhania 4 participants 5 messages Jul 15 – Jul 16

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

PATCH 1/1 KVM: SVM: Bump asid_generation on CPU online to avoid ASID collision after hotplug

If a vCPU stays scheduled out (or blocked) while the last pCPU it ran on goes through a hotplug cycle (online->offline->online), and the vCPU then resumes execution on the same pCPU, then it is possible for it to run with an ASID that has now been assigned to a different vCPU, resulting in stale TLB translations being used.

svm_enable_virtualization_cpu() resets asid_generation to 1 and sets next_asid to max_asid + 1 on every CPU online event, including hotplug cycles. Because next_asid starts beyond the pool boundary, the first call to new_asid() after an online event always wraps the pool, incrementing asid_generation to 2 and assigning ASIDs starting from min_asid.

Consider two vCPUs from different VMs, vCPU-A pinned to CPU-X holding asid_generation=2 and ASID=N from before the hotplug event:

1. CPU-X goes offline and back online: asid_generation resets to 1, next_asid = max_asid + 1.

2. One or more vCPUs migrate to CPU-X and call new_asid(), wrapping the pool and consuming ASIDs starting from min_asid. Eventually vCPU-B from a different VM is assigned asid_generation=2, ASID=N — the same ASID that vCPU-A held before the hotplug.

3. vCPU-A enters pre_svm_run() on CPU-X: current_vmcb->cpu is unchanged so the migration branch is skipped. Its saved asid_generation=2 matches sd->asid_generation=2, so the generation check silently passes and vCPU-A continues running with ASID=N — the same ASID just freshly assigned to vCPU-B.

Both vCPUs from different VMs now run on CPU-X with the same ASID, causing them to share NPT TLB entries and producing stale translations.

The collision manifests as a KVM internal error (Suberror: 1, emulation failure). The NPT page fault reports a faulting GPA far outside the VM's physical memory range — a sign of stale TLB translations being used. KVM falls back to instruction emulation, which fails on FPU/XSave instructions (XRSTOR, STMXCSR) that the emulator does not implement.

Fix this by incrementing asid_generation instead of resetting it to 1 in svm_enable_virtualization_cpu(). On module load, asid_generation starts at 0 (memset) and the increment produces 1, identical to the old behaviour. On subsequent hotplug cycles the generation advances beyond any value a vCPU previously observed on this CPU, so the generation check in pre_svm_run() reliably forces new_asid() on every vCPU after every hotplug cycle.

---
 arch/x86/kvm/svm/svm.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
arch/x86/kvm/svm/svm.c +6 -1
index 91286d46d13ad..c2c6487bd5bd1 100644
@@ -571,7 +571,12 @@ static int svm_enable_virtualization_cpu(void)
 		return r;
 
 	sd = per_cpu_ptr(&svm_data, me);
-	sd->asid_generation = 1;
+	/*
+	 * Bump the current asid_generation value to ensure any vCPU that
+	 * previously ran on this CPU sees a stale generation and is forced
+	 * to acquire a new ASID, preventing a latent ASID collision.
+	 */
+	sd->asid_generation++;
 	sd->max_asid = cpuid_ebx(SVM_CPUID_FUNC) - 1;
 	sd->next_asid = sd->max_asid + 1;
 	sd->min_asid = max_sev_asid + 1;
Signature · 1 line
2.48.1
Reviewed-by on v1 1/1
#1 lore raw
Quoted 54 lines · Nikunj A Dadhania

If a vCPU stays scheduled out (or blocked) while the last pCPU it ran on goes through a hotplug cycle (online->offline->online), and the vCPU then resumes execution on the same pCPU, then it is possible for it to run with an ASID that has now been assigned to a different vCPU, resulting in stale TLB translations being used.

svm_enable_virtualization_cpu() resets asid_generation to 1 and sets next_asid to max_asid + 1 on every CPU online event, including hotplug cycles. Because next_asid starts beyond the pool boundary, the first call to new_asid() after an online event always wraps the pool, incrementing asid_generation to 2 and assigning ASIDs starting from min_asid.

Consider two vCPUs from different VMs, vCPU-A pinned to CPU-X holding asid_generation=2 and ASID=N from before the hotplug event:

1. CPU-X goes offline and back online: asid_generation resets to 1, next_asid = max_asid + 1.

2. One or more vCPUs migrate to CPU-X and call new_asid(), wrapping the pool and consuming ASIDs starting from min_asid. Eventually vCPU-B from a different VM is assigned asid_generation=2, ASID=N — the same ASID that vCPU-A held before the hotplug.

3. vCPU-A enters pre_svm_run() on CPU-X: current_vmcb->cpu is unchanged so the migration branch is skipped. Its saved asid_generation=2 matches sd->asid_generation=2, so the generation check silently passes and vCPU-A continues running with ASID=N — the same ASID just freshly assigned to vCPU-B.

Both vCPUs from different VMs now run on CPU-X with the same ASID, causing them to share NPT TLB entries and producing stale translations.

The collision manifests as a KVM internal error (Suberror: 1, emulation failure). The NPT page fault reports a faulting GPA far outside the VM's physical memory range — a sign of stale TLB translations being used. KVM falls back to instruction emulation, which fails on FPU/XSave instructions (XRSTOR, STMXCSR) that the emulator does not implement.

Fix this by incrementing asid_generation instead of resetting it to 1 in svm_enable_virtualization_cpu(). On module load, asid_generation starts at 0 (memset) and the increment produces 1, identical to the old behaviour. On subsequent hotplug cycles the generation advances beyond any value a vCPU previously observed on this CPU, so the generation check in pre_svm_run() reliably forces new_asid() on every vCPU after every hotplug cycle.

Quoted 3 lines
---
 arch/x86/kvm/svm/svm.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

- Naveen

#2 lore raw
Quoted 60 lines · Naveen N Rao
Quoted 54 lines · Nikunj A Dadhania

If a vCPU stays scheduled out (or blocked) while the last pCPU it ran on goes through a hotplug cycle (online->offline->online), and the vCPU then resumes execution on the same pCPU, then it is possible for it to run with an ASID that has now been assigned to a different vCPU, resulting in stale TLB translations being used.

svm_enable_virtualization_cpu() resets asid_generation to 1 and sets next_asid to max_asid + 1 on every CPU online event, including hotplug cycles. Because next_asid starts beyond the pool boundary, the first call to new_asid() after an online event always wraps the pool, incrementing asid_generation to 2 and assigning ASIDs starting from min_asid.

Consider two vCPUs from different VMs, vCPU-A pinned to CPU-X holding asid_generation=2 and ASID=N from before the hotplug event:

1. CPU-X goes offline and back online: asid_generation resets to 1, next_asid = max_asid + 1.

2. One or more vCPUs migrate to CPU-X and call new_asid(), wrapping the pool and consuming ASIDs starting from min_asid. Eventually vCPU-B from a different VM is assigned asid_generation=2, ASID=N — the same ASID that vCPU-A held before the hotplug.

3. vCPU-A enters pre_svm_run() on CPU-X: current_vmcb->cpu is unchanged so the migration branch is skipped. Its saved asid_generation=2 matches sd->asid_generation=2, so the generation check silently passes and vCPU-A continues running with ASID=N — the same ASID just freshly assigned to vCPU-B.

Both vCPUs from different VMs now run on CPU-X with the same ASID, causing them to share NPT TLB entries and producing stale translations.

The collision manifests as a KVM internal error (Suberror: 1, emulation failure). The NPT page fault reports a faulting GPA far outside the VM's physical memory range — a sign of stale TLB translations being used. KVM falls back to instruction emulation, which fails on FPU/XSave instructions (XRSTOR, STMXCSR) that the emulator does not implement.

Fix this by incrementing asid_generation instead of resetting it to 1 in svm_enable_virtualization_cpu(). On module load, asid_generation starts at 0 (memset) and the increment produces 1, identical to the old behaviour. On subsequent hotplug cycles the generation advances beyond any value a vCPU previously observed on this CPU, so the generation check in pre_svm_run() reliably forces new_asid() on every vCPU after every hotplug cycle.

Quoted 3 lines
---
 arch/x86/kvm/svm/svm.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
Quoted 2 lines

Thanks for the review!

On the Cc: stable@ — if a v2 becomes necessary I will add it. Otherwise, Sean, could you please add it when applying?

Regards,
Nikunj

Reviewed-by on v1 1/1
#3 lore raw
In reply to Nikunj A Dadhania
Quoted 49 lines · Nikunj A Dadhania

If a vCPU stays scheduled out (or blocked) while the last pCPU it ran on goes through a hotplug cycle (online->offline->online), and the vCPU then resumes execution on the same pCPU, then it is possible for it to run with an ASID that has now been assigned to a different vCPU, resulting in stale TLB translations being used.

svm_enable_virtualization_cpu() resets asid_generation to 1 and sets next_asid to max_asid + 1 on every CPU online event, including hotplug cycles. Because next_asid starts beyond the pool boundary, the first call to new_asid() after an online event always wraps the pool, incrementing asid_generation to 2 and assigning ASIDs starting from min_asid.

Consider two vCPUs from different VMs, vCPU-A pinned to CPU-X holding asid_generation=2 and ASID=N from before the hotplug event:

1. CPU-X goes offline and back online: asid_generation resets to 1, next_asid = max_asid + 1.

2. One or more vCPUs migrate to CPU-X and call new_asid(), wrapping the pool and consuming ASIDs starting from min_asid. Eventually vCPU-B from a different VM is assigned asid_generation=2, ASID=N — the same ASID that vCPU-A held before the hotplug.

3. vCPU-A enters pre_svm_run() on CPU-X: current_vmcb->cpu is unchanged so the migration branch is skipped. Its saved asid_generation=2 matches sd->asid_generation=2, so the generation check silently passes and vCPU-A continues running with ASID=N — the same ASID just freshly assigned to vCPU-B.

Both vCPUs from different VMs now run on CPU-X with the same ASID, causing them to share NPT TLB entries and producing stale translations.

The collision manifests as a KVM internal error (Suberror: 1, emulation failure). The NPT page fault reports a faulting GPA far outside the VM's physical memory range — a sign of stale TLB translations being used. KVM falls back to instruction emulation, which fails on FPU/XSave instructions (XRSTOR, STMXCSR) that the emulator does not implement.

Fix this by incrementing asid_generation instead of resetting it to 1 in svm_enable_virtualization_cpu(). On module load, asid_generation starts at 0 (memset) and the increment produces 1, identical to the old behaviour. On subsequent hotplug cycles the generation advances beyond any value a vCPU previously observed on this CPU, so the generation check in pre_svm_run() reliably forces new_asid() on every vCPU after every hotplug cycle.

Quoted 6 lines
maintainer Applied
#4 lore raw
In reply to Nikunj A Dadhania
Quoted 48 lines · Nikunj A Dadhania

If a vCPU stays scheduled out (or blocked) while the last pCPU it ran on goes through a hotplug cycle (online->offline->online), and the vCPU then resumes execution on the same pCPU, then it is possible for it to run with an ASID that has now been assigned to a different vCPU, resulting in stale TLB translations being used.

svm_enable_virtualization_cpu() resets asid_generation to 1 and sets next_asid to max_asid + 1 on every CPU online event, including hotplug cycles. Because next_asid starts beyond the pool boundary, the first call to new_asid() after an online event always wraps the pool, incrementing asid_generation to 2 and assigning ASIDs starting from min_asid.

Consider two vCPUs from different VMs, vCPU-A pinned to CPU-X holding asid_generation=2 and ASID=N from before the hotplug event:

1. CPU-X goes offline and back online: asid_generation resets to 1, next_asid = max_asid + 1.

2. One or more vCPUs migrate to CPU-X and call new_asid(), wrapping the pool and consuming ASIDs starting from min_asid. Eventually vCPU-B from a different VM is assigned asid_generation=2, ASID=N — the same ASID that vCPU-A held before the hotplug.

3. vCPU-A enters pre_svm_run() on CPU-X: current_vmcb->cpu is unchanged so the migration branch is skipped. Its saved asid_generation=2 matches sd->asid_generation=2, so the generation check silently passes and vCPU-A continues running with ASID=N — the same ASID just freshly assigned to vCPU-B.

Both vCPUs from different VMs now run on CPU-X with the same ASID, causing them to share NPT TLB entries and producing stale translations.

The collision manifests as a KVM internal error (Suberror: 1, emulation failure). The NPT page fault reports a faulting GPA far outside the VM's physical memory range — a sign of stale TLB translations being used. KVM falls back to instruction emulation, which fails on FPU/XSave instructions (XRSTOR, STMXCSR) that the emulator does not implement.

Fix this by incrementing asid_generation instead of resetting it to 1 in svm_enable_virtualization_cpu(). On module load, asid_generation starts at 0 (memset) and the increment produces 1, identical to the old behaviour. On subsequent hotplug cycles the generation advances beyond any value a vCPU previously observed on this CPU, so the generation check in pre_svm_run() reliably forces new_asid() on every vCPU after every hotplug cycle.

Quoted 6 lines

Queued, thanks.

Paolo

Quoted 22 lines
---
  arch/x86/kvm/svm/svm.c | 7 ++++++-
  1 file changed, 6 insertions(+), 1 deletion(-)
arch/x86/kvm/svm/svm.c +6 -1
index 91286d46d13ad..c2c6487bd5bd1 100644
@@ -571,7 +571,12 @@ static int svm_enable_virtualization_cpu(void)
  		return r;
 
  	sd = per_cpu_ptr(&svm_data, me);
-	sd->asid_generation = 1;
+	/*
+	 * Bump the current asid_generation value to ensure any vCPU that
+	 * previously ran on this CPU sees a stale generation and is forced
+	 * to acquire a new ASID, preventing a latent ASID collision.
+	 */
+	sd->asid_generation++;
  	sd->max_asid = cpuid_ebx(SVM_CPUID_FUNC) - 1;
  	sd->next_asid = sd->max_asid + 1;
  	sd->min_asid = max_sev_asid + 1;