Skip to contentSkip to conversation

KVM: x86: Backport support for interrupt-based APF page-ready delivery in guest

RFC PATCH New Mancini, Riccardo 2 participants 3 messages Oct 17, 2023 – Oct 18, 2023

Subsystems (5)
x86 KVM KVM x86 x86 entry code +1 more subsystems
lore
Share & tools
lore
#0 lore raw

Hey,

Thank you both for the quick feedback.

Quoted 13 lines
Quoted 4 lines

I've backported the guest-side of the patchset to 4.14.326, could you help us and take a look at the backport?
I only backported the original patchset, I'm not sure if there's any other patch (bug fix) that needs to be included in the backpotrt.

I remember us fixing PV feature enablement/disablement for hibernation/kdump later, see e.g.

commit 8b79feffeca28c5459458fe78676b081e87c93a4

Author: Vitaly Kuznetsov <[email protected]>
Date:   Wed Apr 14 14:35:41 2021 +0200

x86/kvm: Teardown PV features on boot CPU as well

Quoted 7 lines

commit 3d6b84132d2a57b5a74100f6923a8feb679ac2ce

Author: Vitaly Kuznetsov <[email protected]>
Date:   Wed Apr 14 14:35:43 2021 +0200

x86/kvm: Disable all PV features on crash

if you're interested in such use-cases. I don't recall any required fixes for normal operation.

These look like issues already present in 4.14, not introduced by the interrupt-based mechanism, correct?
If so, I wouldn't chase them.
Furthermore, I don't even think we hit those use cases in our scenario.

Quoted 11 lines
Quoted 8 lines · Paolo Bonzini
Quoted 6 lines · Vitaly Kuznetsov

In case keeping legacy mechanism is a must, I would suggest you somehow record the fact that the guest has opted for interrupt-based delivery (e.g. set a global variable or use a static key) and short-circuit
do_async_page_fault() to immediately return and not do anything in this case.

Quoted 4 lines
Quoted 2 lines

I guess you mean "not do anything for KVM_PV_REASON_PAGE_READY in this case"?

Yes, of course: KVM_PV_REASON_PAGE_NOT_PRESENT is always a #PF.

I agree this is a difference with the upstream asyncpf-int implementation and it's theoretically incorrect. I think this shouldn't happen in a normal case, but it's better to keep it consistent.
I'll add a check that asyncpf-int is _not_ enabled before processing KVM_PV_REASON_PAGE_READY. Draft diff below.

Thanks,
Riccardo

arch/x86/kernel/kvm.c +6 -1
index 582a366b82d8..bdfdffd35939 100644
@@ -79,6 +79,8 @@ static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
 static int has_steal_clock = 0;
 
+static DEFINE_PER_CPU(u32, kvm_apf_int_enabled);
+
 /*
  * No need for any "IO delay" on KVM
  */
@@ -277,7 +279,8 @@ do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
                prev_state = exception_enter();
                kvm_async_pf_task_wait((u32)read_cr2(), !user_mode(regs));
                exception_exit(prev_state);
-       } else if (reason & KVM_PV_REASON_PAGE_READY) {
+       } else if (!__this_cpu_read(kvm_apf_int_enabled) && (reason & KVM_PV_REASON_PAGE_READY)) {
+               /* this event is only possible if interrupt-based mechanism is disabled */
                rcu_irq_enter();
                kvm_async_pf_task_wake((u32)read_cr2());
                rcu_irq_exit();
@@ -367,6 +370,7 @@ static void kvm_guest_cpu_init(void)
                if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT)) {
                        pa |= KVM_ASYNC_PF_DELIVERY_AS_INT;
                        wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
+                       __this_cpu_write(kvm_apf_int_enabled, 1);
                }
 
                wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
@@ -396,6 +400,7 @@ static void kvm_pv_disable_apf(void)
 
        wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
        __this_cpu_write(apf_reason.enabled, 0);
+       __this_cpu_write(kvm_apf_int_enabled, 0);
 
        printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
               smp_processor_id());
#1 lore raw
Quoted 69 lines · Riccardo

Hey,

Thank you both for the quick feedback.

Quoted 13 lines
Quoted 4 lines

I've backported the guest-side of the patchset to 4.14.326, could you help us and take a look at the backport?
I only backported the original patchset, I'm not sure if there's any other patch (bug fix) that needs to be included in the backpotrt.

I remember us fixing PV feature enablement/disablement for hibernation/kdump later, see e.g.

commit 8b79feffeca28c5459458fe78676b081e87c93a4

Author: Vitaly Kuznetsov <[email protected]>
Date:   Wed Apr 14 14:35:41 2021 +0200

x86/kvm: Teardown PV features on boot CPU as well

Quoted 7 lines

commit 3d6b84132d2a57b5a74100f6923a8feb679ac2ce

Author: Vitaly Kuznetsov <[email protected]>
Date:   Wed Apr 14 14:35:43 2021 +0200

x86/kvm: Disable all PV features on crash

if you're interested in such use-cases. I don't recall any required fixes for normal operation.

These look like issues already present in 4.14, not introduced by the interrupt-based mechanism, correct?
If so, I wouldn't chase them.
Furthermore, I don't even think we hit those use cases in our scenario.

Quoted 11 lines
Quoted 8 lines · Paolo Bonzini
Quoted 6 lines · Vitaly Kuznetsov

In case keeping legacy mechanism is a must, I would suggest you somehow record the fact that the guest has opted for interrupt-based delivery (e.g. set a global variable or use a static key) and short-circuit
do_async_page_fault() to immediately return and not do anything in this case.

Quoted 4 lines
Quoted 2 lines

I guess you mean "not do anything for KVM_PV_REASON_PAGE_READY in this case"?

Yes, of course: KVM_PV_REASON_PAGE_NOT_PRESENT is always a #PF.

I agree this is a difference with the upstream asyncpf-int implementation and it's theoretically incorrect. I think this shouldn't happen in a normal case, but it's better to keep it consistent.
I'll add a check that asyncpf-int is _not_ enabled before processing KVM_PV_REASON_PAGE_READY. Draft diff below.

Thanks,
Riccardo

arch/x86/kernel/kvm.c +2 -0
index 582a366b82d8..bdfdffd35939 100644
@@ -79,6 +79,8 @@ static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
 static int has_steal_clock = 0;
 
+static DEFINE_PER_CPU(u32, kvm_apf_int_enabled);
+
 /*
  * No need for any "IO delay" on KVM
  */

arch/x86/kernel/kvm.c do_async_page_fault(struct pt_regs *regs, unsigned long error_code) · line 282

Quoted 5 lines
arch/x86/kernel/kvm.c +1 -1
@@ -277,7 +279,8 @@ do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
                prev_state = exception_enter();
                kvm_async_pf_task_wait((u32)read_cr2(), !user_mode(regs));
                exception_exit(prev_state);
-       } else if (reason & KVM_PV_REASON_PAGE_READY) {
+       } else if (!__this_cpu_read(kvm_apf_int_enabled) && (reason & KVM_PV_REASON_PAGE_READY)) {

My bad: I completely forgot KVM_PV_REASON_PAGE_READY is actually not used for interrupt-based delivery:

commit 9ce372b33a2ebbd0b965148879ae169a0015d3f3

Author: Vitaly Kuznetsov <[email protected]>
Date:   Thu May 7 16:36:02 2020 +0200

KVM: x86: drop KVM_PV_REASON_PAGE_READY case from kvm_handle_page_fault()

and in fact there's nothing else but KVM_PV_REASON_PAGE_NOT_PRESENT in apf_reason.flags so I agree that this should never happen (but 'kvm_apf_int_enabled' is a good hardening anyway :-)

Quoted 21 lines
arch/x86/kernel/kvm.c +3 -0
+               /* this event is only possible if interrupt-based mechanism is disabled */
                rcu_irq_enter();
                kvm_async_pf_task_wake((u32)read_cr2());
                rcu_irq_exit();
@@ -367,6 +370,7 @@ static void kvm_guest_cpu_init(void)
                if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT)) {
                        pa |= KVM_ASYNC_PF_DELIVERY_AS_INT;
                        wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
+                       __this_cpu_write(kvm_apf_int_enabled, 1);
                }
 
                wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
@@ -396,6 +400,7 @@ static void kvm_pv_disable_apf(void)
 
        wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
        __this_cpu_write(apf_reason.enabled, 0);
+       __this_cpu_write(kvm_apf_int_enabled, 0);
 
        printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
               smp_processor_id());
Signature · 1 line
Vitaly
#2 lore raw

RFC PATCH v2 1/1 KVM: x86: Backport support for interrupt-based APF page-ready delivery in guest

Hey Vitaly, Paolo,

thank you both for the comments!
For visibility, here is the complete v2 patch with the additional safety check to avoid handling page ready notifications from #PF if async-pf-int is enabled.

Thanks,
Riccardo

Commit follows:

This patch backports support for interrupt-based delivery of Async Page Fault notifications in KVM guests from commit b1d405751cd5 ("KVM: x86:
Switch KVM guest to using interrupts for page ready APF delivery") [1].

Differently from the patchet upstream, this patch does not remove support for the legacy #PF-based delivery, and removes unnecessary refactoring to limit changes to KVM guest code.

v2: add kvm_apf_int_enabled flag to prevent handling of PAGE_READY

[1] https://lore.kernel.org/kvm/[email protected]/

---
 arch/x86/entry/entry_32.S            |  5 +++
 arch/x86/entry/entry_64.S            |  5 +++
 arch/x86/include/asm/hardirq.h       |  2 +-
 arch/x86/include/asm/kvm_para.h      |  7 ++++
 arch/x86/include/uapi/asm/kvm_para.h | 16 +++++++++-
 arch/x86/kernel/irq.c                |  2 +-
 arch/x86/kernel/kvm.c                | 48 ++++++++++++++++++++++++----
 7 files changed, 75 insertions(+), 10 deletions(-)
arch/x86/entry/entry_32.S +5 -0
index 1fdedb2eaef3..460eb7a9981e 100644
@@ -896,6 +896,11 @@ BUILD_INTERRUPT3(hyperv_callback_vector, HYPERVISOR_CALLBACK_VECTOR,
 
 #endif /* CONFIG_HYPERV */
 
+#ifdef CONFIG_KVM_GUEST
+BUILD_INTERRUPT3(kvm_async_pf_vector, HYPERVISOR_CALLBACK_VECTOR,
+		 kvm_async_pf_intr)
+#endif
+
 ENTRY(page_fault)
 	ASM_CLAC
 	pushl	$do_page_fault
arch/x86/entry/entry_64.S +5 -0
index 1804ccf52d9b..545d911a2c9e 100644
@@ -1113,6 +1113,11 @@ apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
 	hyperv_callback_vector hyperv_vector_handler
 #endif /* CONFIG_HYPERV */
 
+#ifdef CONFIG_KVM_GUEST
+apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
+	kvm_async_pf_vector kvm_async_pf_intr
+#endif
+
 idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
 idtentry int3			do_int3			has_error_code=0	create_gap=1
 idtentry stack_segment		do_stack_segment	has_error_code=1
arch/x86/include/asm/hardirq.h +1 -1
index 486c843273c4..75e51b7c9f50 100644
@@ -37,7 +37,7 @@ typedef struct {
 #ifdef CONFIG_X86_MCE_AMD
 	unsigned int irq_deferred_error_count;
 #endif
-#if IS_ENABLED(CONFIG_HYPERV) || defined(CONFIG_XEN)
+#if IS_ENABLED(CONFIG_HYPERV) || defined(CONFIG_XEN) || defined(CONFIG_KVM_GUEST)
 	unsigned int irq_hv_callback_count;
 #endif
 } ____cacheline_aligned irq_cpustat_t;
arch/x86/include/asm/kvm_para.h +7 -0
index c373e44049b1..f4806dd3c38d 100644
@@ -4,6 +4,7 @@
 
 #include <asm/processor.h>
 #include <asm/alternative.h>
+#include <linux/interrupt.h>
 #include <uapi/asm/kvm_para.h>
 
 extern void kvmclock_init(void);
@@ -94,6 +95,12 @@ void kvm_async_pf_task_wake(u32 token);
 u32 kvm_read_and_reset_pf_reason(void);
 extern void kvm_disable_steal_time(void);
 
+extern __visible void kvm_async_pf_vector(void);
+#ifdef CONFIG_TRACING
+#define trace_kvm_async_pf_vector kvm_async_pf_vector
+#endif
+__visible void __irq_entry kvm_async_pf_intr(struct pt_regs *regs);
+
 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 void __init kvm_spinlock_init(void);
 #else /* !CONFIG_PARAVIRT_SPINLOCKS */
arch/x86/include/uapi/asm/kvm_para.h +15 -1
index 341db0462b85..29b86e9adc9a 100644
@@ -26,6 +26,8 @@
 #define KVM_FEATURE_PV_EOI		6
 #define KVM_FEATURE_PV_UNHALT		7
 #define KVM_FEATURE_ASYNC_PF_VMEXIT	10
+#define KVM_FEATURE_ASYNC_PF_INT	14
+#define KVM_FEATURE_MSI_EXT_DEST_ID	15
 
 /* The last 8 bits are used to indicate how to interpret the flags field
  * in pvclock structure. If no bits are set, all flags are ignored.
@@ -42,6 +44,8 @@
 #define MSR_KVM_ASYNC_PF_EN 0x4b564d02
 #define MSR_KVM_STEAL_TIME  0x4b564d03
 #define MSR_KVM_PV_EOI_EN      0x4b564d04
+#define MSR_KVM_ASYNC_PF_INT	0x4b564d06
+#define MSR_KVM_ASYNC_PF_ACK	0x4b564d07
 
 struct kvm_steal_time {
 	__u64 steal;
@@ -70,6 +74,11 @@ struct kvm_clock_pairing {
 #define KVM_ASYNC_PF_ENABLED			(1 << 0)
 #define KVM_ASYNC_PF_SEND_ALWAYS		(1 << 1)
 #define KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT	(1 << 2)
+#define KVM_ASYNC_PF_DELIVERY_AS_INT		(1 << 3)
+
+/* MSR_KVM_ASYNC_PF_INT */
+#define KVM_ASYNC_PF_VEC_MASK			GENMASK(7, 0)
+
 
 /* Operations for KVM_HC_MMU_OP */
 #define KVM_MMU_OP_WRITE_PTE            1
@@ -101,8 +110,13 @@ struct kvm_mmu_op_release_pt {
 #define KVM_PV_REASON_PAGE_READY 2
 
 struct kvm_vcpu_pv_apf_data {
+	/* Used for 'page not present' events delivered via #PF */
 	__u32 reason;
-	__u8 pad[60];
+
+	/* Used for 'page ready' events delivered via interrupt notification */
+	__u32 token;
+
+	__u8 pad[56];
 	__u32 enabled;
 };
 
arch/x86/kernel/irq.c +1 -1
index fbcc303fb1f9..d1def86b66bd 100644
@@ -134,7 +134,7 @@ int arch_show_interrupts(struct seq_file *p, int prec)
 		seq_printf(p, "%10u ", per_cpu(mce_poll_count, j));
 	seq_puts(p, "  Machine check polls\n");
 #endif
-#if IS_ENABLED(CONFIG_HYPERV) || defined(CONFIG_XEN)
+#if IS_ENABLED(CONFIG_HYPERV) || defined(CONFIG_XEN) || defined(CONFIG_KVM_GUEST)
 	if (test_bit(HYPERVISOR_CALLBACK_VECTOR, used_vectors)) {
 		seq_printf(p, "%*s: ", prec, "HYP");
 		for_each_online_cpu(j)
arch/x86/kernel/kvm.c +41 -7

91 lines not on this page · raw message

Signature · 1 line
2.40.1