Skip to contentSkip to conversation

KVM: arm64: Use generic guest entry infrastructure

PATCH Superseded Oliver Upton 3 participants 8 messages Jul 29, 2021 – Jul 30, 2021

Subsystems (3)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

v2 cover KVM: arm64: Use generic guest entry infrastructure

The arm64 kernel doesn't yet support the full generic entry infrastructure. That being said, KVM/arm64 doesn't properly handle TIF_NOTIFY_RESUME and could pick this up by switching to the generic guest entry infrasturture.

Patch 1 adds a missing vCPU stat to ARM64 to record the number of signal exits to userspace.

Patch 2 unhitches entry-kvm from entry-generic, as ARM64 doesn't currently support the generic infrastructure.

Patch 3 replaces the open-coded entry handling with the generic xfer function.

This series was tested on an Ampere Mt. Jade reference system. The series cleanly applies to kvm/queue (note that this is deliberate as the generic kvm stats patches have not yet propagated to kvm-arm/queue) at the following commit:

8ad5e63649ff ("KVM: Don't take mmu_lock for range invalidation unless necessary")

v1 -> v2:
- Address Jing's comment
- Carry Jing's r-b tag

v1: http://lore.kernel.org/r/[email protected]

Oliver Upton (3):
KVM: arm64: Record number of signal exits as a vCPU stat
entry: KVM: Allow use of generic KVM entry w/o full generic support
KVM: arm64: Use generic KVM xfer to guest work function

 arch/arm64/include/asm/kvm_host.h |  1 +
 arch/arm64/kvm/Kconfig            |  1 +
 arch/arm64/kvm/arm.c              | 26 ++++++++++++++------------
 arch/arm64/kvm/guest.c            |  1 +
 include/linux/entry-kvm.h         |  6 +++++-
 5 files changed, 22 insertions(+), 13 deletions(-)
Signature · 6 lines
2.32.0.554.ge1b32706d8-goog

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

PATCH v2 1/3 KVM: arm64: Record number of signal exits as a vCPU stat

Most other architectures that implement KVM record a statistic indicating the number of times a vCPU has exited due to a pending signal. Add support for that stat to arm64.

---
 arch/arm64/include/asm/kvm_host.h | 1 +
 arch/arm64/kvm/arm.c              | 1 +
 arch/arm64/kvm/guest.c            | 1 +
 3 files changed, 3 insertions(+)
arch/arm64/include/asm/kvm_host.h +1 -0
index 41911585ae0c..70e129f2b574 100644
@@ -576,6 +576,7 @@ struct kvm_vcpu_stat {
 	u64 wfi_exit_stat;
 	u64 mmio_exit_user;
 	u64 mmio_exit_kernel;
+	u64 signal_exits;
 	u64 exits;
 };
 
arch/arm64/kvm/arm.c +1 -0
index e9a2b8f27792..60d0a546d7fd 100644
@@ -783,6 +783,7 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 		if (signal_pending(current)) {
 			ret = -EINTR;
 			run->exit_reason = KVM_EXIT_INTR;
+			++vcpu->stat.signal_exits;
 		}
 
 		/*
arch/arm64/kvm/guest.c +1 -0
index 1dfb83578277..853d1e8d2e73 100644
@@ -50,6 +50,7 @@ const struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
 	STATS_DESC_COUNTER(VCPU, wfi_exit_stat),
 	STATS_DESC_COUNTER(VCPU, mmio_exit_user),
 	STATS_DESC_COUNTER(VCPU, mmio_exit_kernel),
+	STATS_DESC_COUNTER(VCPU, signal_exits),
 	STATS_DESC_COUNTER(VCPU, exits)
 };
 static_assert(ARRAY_SIZE(kvm_vcpu_stats_desc) ==
Signature · 6 lines
2.32.0.554.ge1b32706d8-goog

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

PATCH v2 2/3 entry: KVM: Allow use of generic KVM entry w/o full generic support

In reply to Oliver Upton

Some architectures (e.g. arm64) have yet to adopt the generic entry infrastructure. Despite that, it would be nice to use some common plumbing for guest entry/exit handling. For example, KVM/arm64 currently does not handle TIF_NOTIFY_PENDING correctly.

Allow use of only the generic KVM entry code by tightening up the include list. No functional change intended.

---
 include/linux/entry-kvm.h | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
include/linux/entry-kvm.h +5 -1
index 136b8d97d8c0..0d7865a0731c 100644
@@ -2,7 +2,11 @@
 #ifndef __LINUX_ENTRYKVM_H
 #define __LINUX_ENTRYKVM_H
 
-#include <linux/entry-common.h>
+#include <linux/static_call_types.h>
+#include <linux/tracehook.h>
+#include <linux/syscalls.h>
+#include <linux/seccomp.h>
+#include <linux/sched.h>
 #include <linux/tick.h>
 
 /* Transfer to guest mode work */
Signature · 6 lines
2.32.0.554.ge1b32706d8-goog

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

PATCH v2 3/3 KVM: arm64: Use generic KVM xfer to guest work function

In reply to Oliver Upton

Clean up handling of checks for pending work by switching to the generic infrastructure to do so.

We pick up handling for TIF_NOTIFY_RESUME from this switch, meaning that task work will be correctly handled.

---
 arch/arm64/kvm/Kconfig |  1 +
 arch/arm64/kvm/arm.c   | 27 ++++++++++++++-------------
 2 files changed, 15 insertions(+), 13 deletions(-)
arch/arm64/kvm/Kconfig +1 -0
index a4eba0908bfa..8bc1fac5fa26 100644
@@ -26,6 +26,7 @@ menuconfig KVM
 	select HAVE_KVM_ARCH_TLB_FLUSH_ALL
 	select KVM_MMIO
 	select KVM_GENERIC_DIRTYLOG_READ_PROTECT
+	select KVM_XFER_TO_GUEST_WORK
 	select SRCU
 	select KVM_VFIO
 	select HAVE_KVM_EVENTFD
arch/arm64/kvm/arm.c +14 -13
index 60d0a546d7fd..9762e2129813 100644
@@ -6,6 +6,7 @@
 
 #include <linux/bug.h>
 #include <linux/cpu_pm.h>
+#include <linux/entry-kvm.h>
 #include <linux/errno.h>
 #include <linux/err.h>
 #include <linux/kvm_host.h>
@@ -714,6 +715,13 @@ static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
 		static_branch_unlikely(&arm64_mismatched_32bit_el0);
 }
 
+static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu)
+{
+	return kvm_request_pending(vcpu) ||
+			need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
+			xfer_to_guest_mode_work_pending();
+}
+
 /**
  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
  * @vcpu:	The VCPU pointer
@@ -757,7 +765,11 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 		/*
 		 * Check conditions before entering the guest
 		 */
-		cond_resched();
+		if (__xfer_to_guest_mode_work_pending()) {
+			ret = xfer_to_guest_mode_handle_work(vcpu);
+			if (!ret)
+				ret = 1;
+		}
 
 		update_vmid(&vcpu->arch.hw_mmu->vmid);
 
@@ -776,16 +788,6 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 
 		kvm_vgic_flush_hwstate(vcpu);
 
-		/*
-		 * Exit if we have a signal pending so that we can deliver the
-		 * signal to user space.
-		 */
-		if (signal_pending(current)) {
-			ret = -EINTR;
-			run->exit_reason = KVM_EXIT_INTR;
-			++vcpu->stat.signal_exits;
-		}
-
 		/*
 		 * If we're using a userspace irqchip, then check if we need
 		 * to tell a userspace irqchip about timer or PMU level
@@ -809,8 +811,7 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 		 */
 		smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 
-		if (ret <= 0 || need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
-		    kvm_request_pending(vcpu)) {
+		if (ret <= 0 || kvm_vcpu_exit_request(vcpu)) {
 			vcpu->mode = OUTSIDE_GUEST_MODE;
 			isb(); /* Ensure work in x_flush_hwstate is committed */
 			kvm_pmu_sync_hwstate(vcpu);
Signature · 6 lines
2.32.0.554.ge1b32706d8-goog

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

Re: [PATCH v2 3/3] KVM: arm64: Use generic KVM xfer to guest work function

Hi Oliver,

Quoted 38 lines · Oliver Upton

Clean up handling of checks for pending work by switching to the generic infrastructure to do so.

We pick up handling for TIF_NOTIFY_RESUME from this switch, meaning that task work will be correctly handled.

---
 arch/arm64/kvm/Kconfig |  1 +
 arch/arm64/kvm/arm.c   | 27 ++++++++++++++-------------
 2 files changed, 15 insertions(+), 13 deletions(-)
arch/arm64/kvm/Kconfig +1 -0
index a4eba0908bfa..8bc1fac5fa26 100644
@@ -26,6 +26,7 @@ menuconfig KVM
 	select HAVE_KVM_ARCH_TLB_FLUSH_ALL
 	select KVM_MMIO
 	select KVM_GENERIC_DIRTYLOG_READ_PROTECT
+	select KVM_XFER_TO_GUEST_WORK
 	select SRCU
 	select KVM_VFIO
 	select HAVE_KVM_EVENTFD
arch/arm64/kvm/arm.c +1 -0
index 60d0a546d7fd..9762e2129813 100644
@@ -6,6 +6,7 @@
 
 #include <linux/bug.h>
 #include <linux/cpu_pm.h>
+#include <linux/entry-kvm.h>
 #include <linux/errno.h>
 #include <linux/err.h>
 #include <linux/kvm_host.h>

arch/arm64/kvm/arm.c static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu) · line 722

Quoted 8 lines
arch/arm64/kvm/arm.c +5 -0
@@ -714,6 +715,13 @@ static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
 		static_branch_unlikely(&arm64_mismatched_32bit_el0);
 }
 
+static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu)
+{
+	return kvm_request_pending(vcpu) ||
+			need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
+			xfer_to_guest_mode_work_pending();

Here's what xfer_to_guest_mode_work_pending() says:

<quote>
* Has to be invoked with interrupts disabled before the transition to
* guest mode.
</quote>

At the point where you call this, we already are in guest mode, at least in the KVM sense.

arch/arm64/kvm/arm.c int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) · line 769

Quoted 12 lines
arch/arm64/kvm/arm.c +4 -1
+}
+
 /**
  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
  * @vcpu:	The VCPU pointer
@@ -757,7 +765,11 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 		/*
 		 * Check conditions before entering the guest
 		 */
-		cond_resched();
+		if (__xfer_to_guest_mode_work_pending()) {
+			ret = xfer_to_guest_mode_handle_work(vcpu);

xfer_to_guest_mode_handle_work() already does the exact equivalent of __xfer_to_guest_mode_work_pending(). Why do we need to do it twice?

Quoted 24 lines
arch/arm64/kvm/arm.c +3 -10
+			if (!ret)
+				ret = 1;
+		}
 
 		update_vmid(&vcpu->arch.hw_mmu->vmid);
 
@@ -776,16 +788,6 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 
 		kvm_vgic_flush_hwstate(vcpu);
 
-		/*
-		 * Exit if we have a signal pending so that we can deliver the
-		 * signal to user space.
-		 */
-		if (signal_pending(current)) {
-			ret = -EINTR;
-			run->exit_reason = KVM_EXIT_INTR;
-			++vcpu->stat.signal_exits;
-		}
-
 		/*
 		 * If we're using a userspace irqchip, then check if we need
 		 * to tell a userspace irqchip about timer or PMU level

arch/arm64/kvm/arm.c int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) · line 814

Quoted 6 lines
arch/arm64/kvm/arm.c +1 -2
@@ -809,8 +811,7 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 		 */
 		smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 
-		if (ret <= 0 || need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
-		    kvm_request_pending(vcpu)) {
+		if (ret <= 0 || kvm_vcpu_exit_request(vcpu)) {

If you are doing this, please move the userspace irqchip handling into the helper as well, so that we have a single function dealing with collecting exit reasons.

arch/arm64/kvm/arm.c int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) · line 817

Quoted 3 lines
arch/arm64/kvm/arm.c +0 -0
 			vcpu->mode = OUTSIDE_GUEST_MODE;
 			isb(); /* Ensure work in x_flush_hwstate is committed */
 			kvm_pmu_sync_hwstate(vcpu);

Thanks,

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

Re: [PATCH v2 3/3] KVM: arm64: Use generic KVM xfer to guest work function

Marc,

Quoted 54 lines · Marc Zyngier

Hi Oliver,

Quoted 38 lines · Oliver Upton

Clean up handling of checks for pending work by switching to the generic infrastructure to do so.

We pick up handling for TIF_NOTIFY_RESUME from this switch, meaning that task work will be correctly handled.

---
 arch/arm64/kvm/Kconfig |  1 +
 arch/arm64/kvm/arm.c   | 27 ++++++++++++++-------------
 2 files changed, 15 insertions(+), 13 deletions(-)
arch/arm64/kvm/Kconfig +1 -0
index a4eba0908bfa..8bc1fac5fa26 100644
@@ -26,6 +26,7 @@ menuconfig KVM
      select HAVE_KVM_ARCH_TLB_FLUSH_ALL
      select KVM_MMIO
      select KVM_GENERIC_DIRTYLOG_READ_PROTECT
+     select KVM_XFER_TO_GUEST_WORK
      select SRCU
      select KVM_VFIO
      select HAVE_KVM_EVENTFD
arch/arm64/kvm/arm.c +1 -0
index 60d0a546d7fd..9762e2129813 100644
@@ -6,6 +6,7 @@
 
 #include <linux/bug.h>
 #include <linux/cpu_pm.h>
+#include <linux/entry-kvm.h>
 #include <linux/errno.h>
 #include <linux/err.h>
 #include <linux/kvm_host.h>

arch/arm64/kvm/arm.c static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu) · line 722

Quoted 8 lines
arch/arm64/kvm/arm.c +5 -0
@@ -714,6 +715,13 @@ static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
              static_branch_unlikely(&arm64_mismatched_32bit_el0);
 }
 
+static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu)
+{
+     return kvm_request_pending(vcpu) ||
+                     need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
+                     xfer_to_guest_mode_work_pending();

Here's what xfer_to_guest_mode_work_pending() says:

Quoted 7 lines

<quote>
* Has to be invoked with interrupts disabled before the transition to
* guest mode.
</quote>

At the point where you call this, we already are in guest mode, at least in the KVM sense.

I believe the comment is suggestive of guest mode in the hardware sense, not KVM's vcpu->mode designation. I got this from arch/x86/kvm/x86.c:vcpu_enter_guest() to infer the author's intentions.

Quoted 16 lines

arch/arm64/kvm/arm.c int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) · line 769

Quoted 13 lines
arch/arm64/kvm/arm.c +4 -1
+}
+
 /**
  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
  * @vcpu:    The VCPU pointer
@@ -757,7 +765,11 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
              /*
               * Check conditions before entering the guest
               */
-             cond_resched();
+             if (__xfer_to_guest_mode_work_pending()) {
+                     ret = xfer_to_guest_mode_handle_work(vcpu);

xfer_to_guest_mode_handle_work() already does the exact equivalent of __xfer_to_guest_mode_work_pending(). Why do we need to do it twice?

Right, there's no need to do the check twice.

Quoted 28 lines
Quoted 28 lines
arch/arm64/kvm/arm.c +3 -10
+                     if (!ret)
+                             ret = 1;
+             }
 
              update_vmid(&vcpu->arch.hw_mmu->vmid);
 
@@ -776,16 +788,6 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 
              kvm_vgic_flush_hwstate(vcpu);
 
-             /*
-              * Exit if we have a signal pending so that we can deliver the
-              * signal to user space.
-              */
-             if (signal_pending(current)) {
-                     ret = -EINTR;
-                     run->exit_reason = KVM_EXIT_INTR;
-                     ++vcpu->stat.signal_exits;
-             }
-
              /*
               * If we're using a userspace irqchip, then check if we need
               * to tell a userspace irqchip about timer or PMU level
@@ -809,8 +811,7 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
               */
              smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 
Quoted 7 lines

arch/arm64/kvm/arm.c int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) · line 814

Quoted 3 lines
arch/arm64/kvm/arm.c +1 -2
-             if (ret <= 0 || need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
-                 kvm_request_pending(vcpu)) {
+             if (ret <= 0 || kvm_vcpu_exit_request(vcpu)) {

If you are doing this, please move the userspace irqchip handling into the helper as well, so that we have a single function dealing with collecting exit reasons.

Sure thing.

Thanks for the quick review, Marc!

--
Best,
Oliver

#6 lore raw

Re: [PATCH v2 3/3] KVM: arm64: Use generic KVM xfer to guest work function

Quoted 21 lines · Oliver Upton
Quoted 19 lines · Marc Zyngier

arch/arm64/kvm/arm.c static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu) · line 722

Quoted 9 lines · Oliver Upton
arch/arm64/kvm/arm.c +5 -0
@@ -714,6 +715,13 @@ static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
              static_branch_unlikely(&arm64_mismatched_32bit_el0);
 }
 
+static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu)
+{
+     return kvm_request_pending(vcpu) ||
+                     need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
+                     xfer_to_guest_mode_work_pending();

Here's what xfer_to_guest_mode_work_pending() says:

<quote>
* Has to be invoked with interrupts disabled before the transition to
* guest mode.
</quote>

Quoted 7 lines
Quoted 2 lines

At the point where you call this, we already are in guest mode, at least in the KVM sense.

I believe the comment is suggestive of guest mode in the hardware sense, not KVM's vcpu->mode designation. I got this from arch/x86/kvm/x86.c:vcpu_enter_guest() to infer the author's intentions.

Yeah, the comment is referring to hardware guest mode. The intent is to verify there is no work to be done before making the expensive world switch. There's no meaningful interaction with vcpu->mode, on x86 it's simply more convenient from a code perspective to throw it into kvm_vcpu_exit_request().

#7 lore raw

Re: [PATCH v2 3/3] KVM: arm64: Use generic KVM xfer to guest work function

Quoted 31 lines · Sean Christopherson
Quoted 21 lines · Oliver Upton
Quoted 19 lines · Marc Zyngier

arch/arm64/kvm/arm.c static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu) · line 722

Quoted 9 lines · Oliver Upton
arch/arm64/kvm/arm.c +5 -0
@@ -714,6 +715,13 @@ static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
              static_branch_unlikely(&arm64_mismatched_32bit_el0);
 }
 
+static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu)
+{
+     return kvm_request_pending(vcpu) ||
+                     need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
+                     xfer_to_guest_mode_work_pending();

Here's what xfer_to_guest_mode_work_pending() says:

<quote>
* Has to be invoked with interrupts disabled before the transition to
* guest mode.
</quote>

Quoted 7 lines
Quoted 2 lines

At the point where you call this, we already are in guest mode, at least in the KVM sense.

I believe the comment is suggestive of guest mode in the hardware sense, not KVM's vcpu->mode designation. I got this from arch/x86/kvm/x86.c:vcpu_enter_guest() to infer the author's intentions.

Quoted 4 lines

Yeah, the comment is referring to hardware guest mode. The intent is to verify there is no work to be done before making the expensive world switch. There's no meaningful interaction with vcpu->mode, on x86 it's simply more convenient from a code perspective to throw it into kvm_vcpu_exit_request().

Yep, the same is true for ARM as well, doing it the way it appears in this patch allows for the recycling of the block to enable irqs and preemption.

--
Thanks,
Oliver