Skip to contentSkip to conversation

KVM: s390: pv: fix clock comparator late after suspend/resume

RFC PATCH Superseded Nico Boehr 3 participants 6 messages Aug 29, 2022 – Oct 5, 2022

Subsystems (3)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Nico Boehr

author
#0 lore raw

v2 cover KVM: s390: pv: fix clock comparator late after suspend/resume

v1->v2:
---
- fix broken migration due to deadlock

After a PV guest in QEMU has been paused and resumed, clock comparator interrupts are delivered to the guest much too late.

This is caused by QEMU's tod-kvm device restoring the guest's TOD clock upon guest resume. This is not possible with PV, since the guest's TOD clock is controlled by the ultravisor.

Even if not allowed under PV, KVM allowed the respective call from userspace (VM attribute KVM_S390_VM_TOD) and updated its internal data structures on this call. This can make the ultravisor's and KVM's view of the guest TOD clock inconsistent. This in turn can lead to the late delivery of clock comparator interrupts when KVM calculates when to wake the guest.

This fixes the kernel portion of the problem by disallowing the vm attr call for the guest TOD clock so userspace cannot mess up KVM's view of the guest TOD. This fix causes an ugly warning in QEMU though, hence another fix is due for QEMU to simply not even attempt to set the guest TOD on resume.

Nico Boehr (1):
KVM: s390: pv: don't allow userspace to set the clock under PV

 arch/s390/kvm/kvm-s390.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
Signature · 1 line
2.36.1
author
#1 lore raw

RFC PATCH v2 1/1 KVM: s390: pv: don't allow userspace to set the clock under PV

When running under PV, the guest's TOD clock is under control of the ultravisor and the hypervisor isn't allowed to change it. Hence, don't allow userspace to change the guest's TOD clock by returning -EOPNOTSUPP.

When userspace changes the guest's TOD clock, KVM updates its kvm.arch.epoch field and, in addition, the epoch field in all state descriptions of all VCPUs.

But, under PV, the ultravisor will ignore the epoch field in the state description and simply overwrite it on next SIE exit with the actual guest epoch. This leads to KVM having an incorrect view of the guest's TOD clock: it has updated its internal kvm.arch.epoch field, but the ultravisor ignores the field in the state description.

Whenever a guest is now waiting for a clock comparator, KVM will incorrectly calculate the time when the guest should wake up, possibly causing the guest to sleep for much longer than expected.

With this change, kvm_s390_set_tod() will now take the kvm->lock to be able to call kvm_s390_pv_is_protected(). Since kvm_s390_set_tod_clock() also takes kvm->lock, use __kvm_s390_set_tod_clock() instead.

Fixes 0f3035047140 ("KVM: s390: protvirt: Do only reset registers that are accessible")
---
 arch/s390/kvm/kvm-s390.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
arch/s390/kvm/kvm-s390.c +13 -2
index edfd4bbd0cba..f4ee88e787fe 100644
@@ -1207,6 +1207,8 @@ static int kvm_s390_vm_get_migration(struct kvm *kvm,
 	return 0;
 }
 
+static void __kvm_s390_set_tod_clock(struct kvm *kvm, const struct kvm_s390_vm_tod_clock *gtod);
+
 static int kvm_s390_set_tod_ext(struct kvm *kvm, struct kvm_device_attr *attr)
 {
 	struct kvm_s390_vm_tod_clock gtod;
@@ -1216,7 +1218,7 @@ static int kvm_s390_set_tod_ext(struct kvm *kvm, struct kvm_device_attr *attr)
 
 	if (!test_kvm_facility(kvm, 139) && gtod.epoch_idx)
 		return -EINVAL;
-	kvm_s390_set_tod_clock(kvm, &gtod);
+	__kvm_s390_set_tod_clock(kvm, &gtod);
 
 	VM_EVENT(kvm, 3, "SET: TOD extension: 0x%x, TOD base: 0x%llx",
 		gtod.epoch_idx, gtod.tod);
@@ -1247,7 +1249,7 @@ static int kvm_s390_set_tod_low(struct kvm *kvm, struct kvm_device_attr *attr)
 			   sizeof(gtod.tod)))
 		return -EFAULT;
 
-	kvm_s390_set_tod_clock(kvm, &gtod);
+	__kvm_s390_set_tod_clock(kvm, &gtod);
 	VM_EVENT(kvm, 3, "SET: TOD base: 0x%llx", gtod.tod);
 	return 0;
 }
@@ -1259,6 +1261,12 @@ static int kvm_s390_set_tod(struct kvm *kvm, struct kvm_device_attr *attr)
 	if (attr->flags)
 		return -EINVAL;
 
+	mutex_lock(&kvm->lock);
+	if (kvm_s390_pv_is_protected(kvm)) {
+		ret = -EOPNOTSUPP;
+		goto out_unlock;
+	}
+
 	switch (attr->attr) {
 	case KVM_S390_VM_TOD_EXT:
 		ret = kvm_s390_set_tod_ext(kvm, attr);
@@ -1273,6 +1281,9 @@ static int kvm_s390_set_tod(struct kvm *kvm, struct kvm_device_attr *attr)
 		ret = -ENXIO;
 		break;
 	}
+
+out_unlock:
+	mutex_unlock(&kvm->lock);
 	return ret;
 }
 
Signature · 1 line
2.36.1
author
#2 lore raw

Re: [RFC PATCH v2 1/1] KVM: s390: pv: don't allow userspace to set the clock under PV

Quoting Nico Boehr (2022-08-29 09:56:02)

Quoted 19 lines

When running under PV, the guest's TOD clock is under control of the ultravisor and the hypervisor isn't allowed to change it. Hence, don't allow userspace to change the guest's TOD clock by returning -EOPNOTSUPP.

When userspace changes the guest's TOD clock, KVM updates its kvm.arch.epoch field and, in addition, the epoch field in all state descriptions of all VCPUs.

But, under PV, the ultravisor will ignore the epoch field in the state description and simply overwrite it on next SIE exit with the actual guest epoch. This leads to KVM having an incorrect view of the guest's TOD clock: it has updated its internal kvm.arch.epoch field, but the ultravisor ignores the field in the state description.

Whenever a guest is now waiting for a clock comparator, KVM will incorrectly calculate the time when the guest should wake up, possibly causing the guest to sleep for much longer than expected.

Quoted 5 lines

With this change, kvm_s390_set_tod() will now take the kvm->lock to be able to call kvm_s390_pv_is_protected(). Since kvm_s390_set_tod_clock() also takes kvm->lock, use __kvm_s390_set_tod_clock() instead.

Fixes 0f3035047140 ("KVM: s390: protvirt: Do only reset registers that are accessible")

I missed a

here. Sorry Marc.

#3 lore raw

Re: [RFC PATCH v2 1/1] KVM: s390: pv: don't allow userspace to set the clock under PV

In reply to Nico Boehr

Am 29.08.22 um 09:56 schrieb Nico Boehr:

Quoted 19 lines

When running under PV, the guest's TOD clock is under control of the ultravisor and the hypervisor isn't allowed to change it. Hence, don't allow userspace to change the guest's TOD clock by returning -EOPNOTSUPP.

When userspace changes the guest's TOD clock, KVM updates its kvm.arch.epoch field and, in addition, the epoch field in all state descriptions of all VCPUs.

But, under PV, the ultravisor will ignore the epoch field in the state description and simply overwrite it on next SIE exit with the actual guest epoch. This leads to KVM having an incorrect view of the guest's TOD clock: it has updated its internal kvm.arch.epoch field, but the ultravisor ignores the field in the state description.

Whenever a guest is now waiting for a clock comparator, KVM will incorrectly calculate the time when the guest should wake up, possibly causing the guest to sleep for much longer than expected.

Quoted 6 lines

With this change, kvm_s390_set_tod() will now take the kvm->lock to be able to call kvm_s390_pv_is_protected(). Since kvm_s390_set_tod_clock() also takes kvm->lock, use __kvm_s390_set_tod_clock() instead.

Fixes 0f3035047140 ("KVM: s390: protvirt: Do only reset registers that are accessible")

Has this patch already been sent to the upstream list yet and have we solved all existing problems of the previous version?

Quoted 58 lines
---
  arch/s390/kvm/kvm-s390.c | 15 +++++++++++++--
  1 file changed, 13 insertions(+), 2 deletions(-)
arch/s390/kvm/kvm-s390.c +13 -2
index edfd4bbd0cba..f4ee88e787fe 100644
@@ -1207,6 +1207,8 @@ static int kvm_s390_vm_get_migration(struct kvm *kvm,
  	return 0;
  }
 
+static void __kvm_s390_set_tod_clock(struct kvm *kvm, const struct kvm_s390_vm_tod_clock *gtod);
+
  static int kvm_s390_set_tod_ext(struct kvm *kvm, struct kvm_device_attr *attr)
  {
  	struct kvm_s390_vm_tod_clock gtod;
@@ -1216,7 +1218,7 @@ static int kvm_s390_set_tod_ext(struct kvm *kvm, struct kvm_device_attr *attr)
 
  	if (!test_kvm_facility(kvm, 139) && gtod.epoch_idx)
  		return -EINVAL;
-	kvm_s390_set_tod_clock(kvm, &gtod);
+	__kvm_s390_set_tod_clock(kvm, &gtod);
 
  	VM_EVENT(kvm, 3, "SET: TOD extension: 0x%x, TOD base: 0x%llx",
  		gtod.epoch_idx, gtod.tod);
@@ -1247,7 +1249,7 @@ static int kvm_s390_set_tod_low(struct kvm *kvm, struct kvm_device_attr *attr)
  			   sizeof(gtod.tod)))
  		return -EFAULT;
 
-	kvm_s390_set_tod_clock(kvm, &gtod);
+	__kvm_s390_set_tod_clock(kvm, &gtod);
  	VM_EVENT(kvm, 3, "SET: TOD base: 0x%llx", gtod.tod);
  	return 0;
  }
@@ -1259,6 +1261,12 @@ static int kvm_s390_set_tod(struct kvm *kvm, struct kvm_device_attr *attr)
  	if (attr->flags)
  		return -EINVAL;
 
+	mutex_lock(&kvm->lock);
+	if (kvm_s390_pv_is_protected(kvm)) {
+		ret = -EOPNOTSUPP;
+		goto out_unlock;
+	}
+
  	switch (attr->attr) {
  	case KVM_S390_VM_TOD_EXT:
  		ret = kvm_s390_set_tod_ext(kvm, attr);
@@ -1273,6 +1281,9 @@ static int kvm_s390_set_tod(struct kvm *kvm, struct kvm_device_attr *attr)
  		ret = -ENXIO;
  		break;
  	}
+
+out_unlock:
+	mutex_unlock(&kvm->lock);
  	return ret;
  }
 
author
#4 lore raw

Re: [RFC PATCH v2 1/1] KVM: s390: pv: don't allow userspace to set the clock under PV

Quoting Christian Borntraeger (2022-09-20 13:07:58)

Quoted 20 lines

Am 29.08.22 um 09:56 schrieb Nico Boehr:

Quoted 19 lines

When running under PV, the guest's TOD clock is under control of the ultravisor and the hypervisor isn't allowed to change it. Hence, don't allow userspace to change the guest's TOD clock by returning -EOPNOTSUPP.

When userspace changes the guest's TOD clock, KVM updates its kvm.arch.epoch field and, in addition, the epoch field in all state descriptions of all VCPUs.

But, under PV, the ultravisor will ignore the epoch field in the state description and simply overwrite it on next SIE exit with the actual guest epoch. This leads to KVM having an incorrect view of the guest's TOD clock: it has updated its internal kvm.arch.epoch field, but the ultravisor ignores the field in the state description.

Whenever a guest is now waiting for a clock comparator, KVM will incorrectly calculate the time when the guest should wake up, possibly causing the guest to sleep for much longer than expected.

Quoted 8 lines
Quoted 6 lines

With this change, kvm_s390_set_tod() will now take the kvm->lock to be able to call kvm_s390_pv_is_protected(). Since kvm_s390_set_tod_clock() also takes kvm->lock, use __kvm_s390_set_tod_clock() instead.

Fixes 0f3035047140 ("KVM: s390: protvirt: Do only reset registers that are accessible")

Has this patch already been sent to the upstream list yet and have we solved all existing problems of the previous version?

Yes, the patch works and the issues I identified should be fixed now. I can resend as a non-RFC if you like.

Note that a QEMU change is also needed to silence this message:

warning: Unable to set KVM guest TOD clock: Operation not supported

But I wanted to get the kernel side done first.

maintainer
#5 lore raw

Re: [RFC PATCH v2 1/1] KVM: s390: pv: don't allow userspace to set the clock under PV

In reply to Nico Boehr
Quoted 34 lines · On Mon, 29 Aug 2022 09:56:02 +0200 Nico Boehr <[email protected]> wrote:

When running under PV, the guest's TOD clock is under control of the ultravisor and the hypervisor isn't allowed to change it. Hence, don't allow userspace to change the guest's TOD clock by returning -EOPNOTSUPP.

When userspace changes the guest's TOD clock, KVM updates its kvm.arch.epoch field and, in addition, the epoch field in all state descriptions of all VCPUs.

But, under PV, the ultravisor will ignore the epoch field in the state description and simply overwrite it on next SIE exit with the actual guest epoch. This leads to KVM having an incorrect view of the guest's TOD clock: it has updated its internal kvm.arch.epoch field, but the ultravisor ignores the field in the state description.

Whenever a guest is now waiting for a clock comparator, KVM will incorrectly calculate the time when the guest should wake up, possibly causing the guest to sleep for much longer than expected.

With this change, kvm_s390_set_tod() will now take the kvm->lock to be able to call kvm_s390_pv_is_protected(). Since kvm_s390_set_tod_clock() also takes kvm->lock, use __kvm_s390_set_tod_clock() instead.

Fixes 0f3035047140 ("KVM: s390: protvirt: Do only reset registers that are accessible")
---
 arch/s390/kvm/kvm-s390.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
arch/s390/kvm/kvm-s390.c +0 -0
index edfd4bbd0cba..f4ee88e787fe 100644

arch/s390/kvm/kvm-s390.c static int kvm_s390_vm_get_migration(struct kvm *kvm, · line 1210

Quoted 4 lines
arch/s390/kvm/kvm-s390.c +1 -0
@@ -1207,6 +1207,8 @@ static int kvm_s390_vm_get_migration(struct kvm *kvm,
 	return 0;
 }
 
+static void __kvm_s390_set_tod_clock(struct kvm *kvm, const struct kvm_s390_vm_tod_clock *gtod);

are there any users of kvm_s390_set_tod_clock left after this patch?

maybe that wrapper can be removed (also from the headers), and we can keep the function with the __

it's perhaps cleaner to have a second patch to remove the unused stuff

Quoted 45 lines
arch/s390/kvm/kvm-s390.c +12 -2
+
 static int kvm_s390_set_tod_ext(struct kvm *kvm, struct kvm_device_attr *attr)
 {
 	struct kvm_s390_vm_tod_clock gtod;
@@ -1216,7 +1218,7 @@ static int kvm_s390_set_tod_ext(struct kvm *kvm, struct kvm_device_attr *attr)
 
 	if (!test_kvm_facility(kvm, 139) && gtod.epoch_idx)
 		return -EINVAL;
-	kvm_s390_set_tod_clock(kvm, &gtod);
+	__kvm_s390_set_tod_clock(kvm, &gtod);
 
 	VM_EVENT(kvm, 3, "SET: TOD extension: 0x%x, TOD base: 0x%llx",
 		gtod.epoch_idx, gtod.tod);
@@ -1247,7 +1249,7 @@ static int kvm_s390_set_tod_low(struct kvm *kvm, struct kvm_device_attr *attr)
 			   sizeof(gtod.tod)))
 		return -EFAULT;
 
-	kvm_s390_set_tod_clock(kvm, &gtod);
+	__kvm_s390_set_tod_clock(kvm, &gtod);
 	VM_EVENT(kvm, 3, "SET: TOD base: 0x%llx", gtod.tod);
 	return 0;
 }
@@ -1259,6 +1261,12 @@ static int kvm_s390_set_tod(struct kvm *kvm, struct kvm_device_attr *attr)
 	if (attr->flags)
 		return -EINVAL;
 
+	mutex_lock(&kvm->lock);
+	if (kvm_s390_pv_is_protected(kvm)) {
+		ret = -EOPNOTSUPP;
+		goto out_unlock;
+	}
+
 	switch (attr->attr) {
 	case KVM_S390_VM_TOD_EXT:
 		ret = kvm_s390_set_tod_ext(kvm, attr);
@@ -1273,6 +1281,9 @@ static int kvm_s390_set_tod(struct kvm *kvm, struct kvm_device_attr *attr)
 		ret = -ENXIO;
 		break;
 	}
+
+out_unlock:
+	mutex_unlock(&kvm->lock);
 	return ret;
 }