Skip to contentSkip to conversation

x86/perf: Use RET0 as default for guest_get_msrs to handle "no PMU" case

PATCH Applied Sean Christopherson 4 participants 6 messages Mar 9, 2021 – Mar 11, 2021 1 message not in the archive

Subsystems (4)
lore
Share & tools
lore

Sean Christopherson

author maintainer
#0 lore raw

PATCH v2 1/1 x86/perf: Use RET0 as default for guest_get_msrs to handle "no PMU" case

Initialize x86_pmu.guest_get_msrs to return 0/NULL to handle the "nop" case. Patching in perf_guest_get_msrs_nop() during setup does not work if there is no PMU, as setup bails before updating the static calls, leaving x86_pmu.guest_get_msrs NULL and thus a complete nop. Ultimately, this causes VMX abort on VM-Exit due to KVM putting random garbage from the stack into the MSR load list.

Add a comment in KVM to note that nr_msrs is valid if and only if the return value is non-NULL.

Fixes abd562df94d1 ("x86/perf: Use static_call for x86_pmu.guest_get_msrs")
Cc Like Xu <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Jim Mattson <[email protected]>
Reported-by [email protected]

---

v2:
- Use __static_call_return0 to return NULL instead of manually checking the hook at invocation. [Peter]
- Rebase to tip/sched/core, commit 4117cebf1a9f ("psi: Optimize task switch inside shared cgroups").

 arch/x86/events/core.c | 16 +++++-----------
 arch/x86/kvm/vmx/vmx.c |  2 +-
 2 files changed, 6 insertions(+), 12 deletions(-)
arch/x86/events/core.c +5 -11
index 6ddeed3cd2ac..7bb056151ecc 100644
@@ -81,7 +81,11 @@ DEFINE_STATIC_CALL_NULL(x86_pmu_swap_task_ctx, *x86_pmu.swap_task_ctx);
 DEFINE_STATIC_CALL_NULL(x86_pmu_drain_pebs,   *x86_pmu.drain_pebs);
 DEFINE_STATIC_CALL_NULL(x86_pmu_pebs_aliases, *x86_pmu.pebs_aliases);
 
-DEFINE_STATIC_CALL_NULL(x86_pmu_guest_get_msrs,  *x86_pmu.guest_get_msrs);
+/*
+ * This one is magic, it will get called even when PMU init fails (because
+ * there is no PMU), in which case it should simply return NULL.
+ */
+DEFINE_STATIC_CALL_RET0(x86_pmu_guest_get_msrs, *x86_pmu.guest_get_msrs);
 
 u64 __read_mostly hw_cache_event_ids
 				[PERF_COUNT_HW_CACHE_MAX]
@@ -1944,13 +1948,6 @@ static void _x86_pmu_read(struct perf_event *event)
 	x86_perf_event_update(event);
 }
 
-static inline struct perf_guest_switch_msr *
-perf_guest_get_msrs_nop(int *nr)
-{
-	*nr = 0;
-	return NULL;
-}
-
 static int __init init_hw_perf_events(void)
 {
 	struct x86_pmu_quirk *quirk;
@@ -2024,9 +2021,6 @@ static int __init init_hw_perf_events(void)
 	if (!x86_pmu.read)
 		x86_pmu.read = _x86_pmu_read;
 
-	if (!x86_pmu.guest_get_msrs)
-		x86_pmu.guest_get_msrs = perf_guest_get_msrs_nop;
-
 	x86_pmu_static_call_update();
 
 	/*
arch/x86/kvm/vmx/vmx.c +1 -1
index 50810d471462..32cf8287d4a7 100644
@@ -6580,8 +6580,8 @@ static void atomic_switch_perf_msrs(struct vcpu_vmx *vmx)
 	int i, nr_msrs;
 	struct perf_guest_switch_msr *msrs;
 
+	/* Note, nr_msrs may be garbage if perf_guest_get_msrs() returns NULL. */
 	msrs = perf_guest_get_msrs(&nr_msrs);
-
 	if (!msrs)
 		return;
 
Signature · 1 line
2.30.1.766.gb4fecdf3b7-goog
#1 lore raw
Quoted 27 lines · Sean Christopherson

Initialize x86_pmu.guest_get_msrs to return 0/NULL to handle the "nop" case. Patching in perf_guest_get_msrs_nop() during setup does not work if there is no PMU, as setup bails before updating the static calls, leaving x86_pmu.guest_get_msrs NULL and thus a complete nop. Ultimately, this causes VMX abort on VM-Exit due to KVM putting random garbage from the stack into the MSR load list.

Add a comment in KVM to note that nr_msrs is valid if and only if the return value is non-NULL.

Fixes abd562df94d1 ("x86/perf: Use static_call for x86_pmu.guest_get_msrs")
Cc Like Xu <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Jim Mattson <[email protected]>
Reported-by [email protected]

---

v2:
- Use __static_call_return0 to return NULL instead of manually checking the hook at invocation. [Peter]
- Rebase to tip/sched/core, commit 4117cebf1a9f ("psi: Optimize task switch inside shared cgroups").

...

arch/x86/kvm/vmx/vmx.c static void atomic_switch_perf_msrs(struct vcpu_vmx *vmx) · line 6583

Quoted 9 lines
arch/x86/kvm/vmx/vmx.c +1 -0
index 50810d471462..32cf8287d4a7 100644
@@ -6580,8 +6580,8 @@ static void atomic_switch_perf_msrs(struct vcpu_vmx *vmx)
        int i, nr_msrs;
        struct perf_guest_switch_msr *msrs;
 
+       /* Note, nr_msrs may be garbage if perf_guest_get_msrs() returns NULL. */

You could drop the scary comment with a profligate initialization of nr_msrs to 0.

[Apologies to those seeing this twice. I blame gmail.]

A message that is not in the archive (sent off-list or not yet received): CALMp9eRL4QRq4-DKEveNsJ8b2XSYEhZ-zKGFWsCcoZMnDeRtxg@mail.gmail.com

author maintainer
#3 lore raw
Quoted 11 lines · Jim Mattson
Quoted 9 lines · Sean Christopherson
arch/x86/kvm/vmx/vmx.c +0 -0
index 50810d471462..32cf8287d4a7 100644
@@ -6580,8 +6580,8 @@ static void atomic_switch_perf_msrs(struct vcpu_vmx
 *vmx)
        int i, nr_msrs;
        struct perf_guest_switch_msr *msrs;
 
Quoted 6 lines
Quoted 3 lines
arch/x86/kvm/vmx/vmx.c +1 -0
+       /* Note, nr_msrs may be garbage if perf_guest_get_msrs() returns

NULL. */

You could drop the scary comment with a profligate initialization of nr_msrs to 0.

Yeah, I considered that as well. I opted for the scary comment because I wanted to dissuade future patches from modifying this code without taking into account the non-obvious behavior.

maintainer
#4 lore raw
In reply to Sean Christopherson

Review static int __init init_hw_perf_events(void) · line 2028

Quoted 6 lines · Sean Christopherson
(quoted diff) +0 -2
@@ -2024,9 +2021,6 @@ static int __init init_hw_perf_events(void)
 	if (!x86_pmu.read)
 		x86_pmu.read = _x86_pmu_read;
 
-	if (!x86_pmu.guest_get_msrs)
-		x86_pmu.guest_get_msrs = perf_guest_get_msrs_nop;

I suspect I might've been over eager here and we're now in trouble when *_pmu_init() clears x86_pmu.guest_get_msrs (like for instance on AMD).

When that happens we need to restore __static_call_return0, otherwise the following static_call_update() will patch in a NOP and RAX will be garbage again.

So I've taken the liberty to update the patch as below.

---

Subject: x86/perf: Use RET0 as default for guest_get_msrs to handle "no PMU" case
From: Sean Christopherson <[email protected]>
Date: Tue, 9 Mar 2021 09:10:19 -0800

From: Sean Christopherson <[email protected]>

Initialize x86_pmu.guest_get_msrs to return 0/NULL to handle the "nop" case. Patching in perf_guest_get_msrs_nop() during setup does not work if there is no PMU, as setup bails before updating the static calls, leaving x86_pmu.guest_get_msrs NULL and thus a complete nop. Ultimately, this causes VMX abort on VM-Exit due to KVM putting random garbage from the stack into the MSR load list.

Add a comment in KVM to note that nr_msrs is valid if and only if the return value is non-NULL.

---

v2:
- Use __static_call_return0 to return NULL instead of manually checking the hook at invocation. [Peter]
- Rebase to tip/sched/core, commit 4117cebf1a9f ("psi: Optimize task switch inside shared cgroups").

 arch/x86/events/core.c |   15 ++++++---------
 arch/x86/kvm/vmx/vmx.c |    2 +-
 2 files changed, 7 insertions(+), 10 deletions(-)
arch/x86/events/core.c +6 -9
@@ -81,7 +81,11 @@ DEFINE_STATIC_CALL_NULL(x86_pmu_swap_tas
 DEFINE_STATIC_CALL_NULL(x86_pmu_drain_pebs,   *x86_pmu.drain_pebs);
 DEFINE_STATIC_CALL_NULL(x86_pmu_pebs_aliases, *x86_pmu.pebs_aliases);
 
-DEFINE_STATIC_CALL_NULL(x86_pmu_guest_get_msrs,  *x86_pmu.guest_get_msrs);
+/*
+ * This one is magic, it will get called even when PMU init fails (because
+ * there is no PMU), in which case it should simply return NULL.
+ */
+DEFINE_STATIC_CALL_RET0(x86_pmu_guest_get_msrs, *x86_pmu.guest_get_msrs);
 
 u64 __read_mostly hw_cache_event_ids
 				[PERF_COUNT_HW_CACHE_MAX]
@@ -1944,13 +1948,6 @@ static void _x86_pmu_read(struct perf_ev
 	x86_perf_event_update(event);
 }
 
-static inline struct perf_guest_switch_msr *
-perf_guest_get_msrs_nop(int *nr)
-{
-	*nr = 0;
-	return NULL;
-}
-
 static int __init init_hw_perf_events(void)
 {
 	struct x86_pmu_quirk *quirk;
@@ -2025,7 +2022,7 @@ static int __init init_hw_perf_events(vo
 		x86_pmu.read = _x86_pmu_read;
 
 	if (!x86_pmu.guest_get_msrs)
-		x86_pmu.guest_get_msrs = perf_guest_get_msrs_nop;
+		x86_pmu.guest_get_msrs = (void *)&__static_call_return0;
 
 	x86_pmu_static_call_update();
 
arch/x86/kvm/vmx/vmx.c +1 -1
@@ -6580,8 +6580,8 @@ static void atomic_switch_perf_msrs(stru
 	int i, nr_msrs;
 	struct perf_guest_switch_msr *msrs;
 
+	/* Note, nr_msrs may be garbage if perf_guest_get_msrs() returns NULL. */
 	msrs = perf_guest_get_msrs(&nr_msrs);
-
 	if (!msrs)
 		return;
 
author maintainer
#5 lore raw
Quoted 12 lines · Peter Zijlstra

arch/x86/events/core.c static int __init init_hw_perf_events(void) · line 2028

Quoted 6 lines · Sean Christopherson
arch/x86/events/core.c +0 -2
@@ -2024,9 +2021,6 @@ static int __init init_hw_perf_events(void)
 	if (!x86_pmu.read)
 		x86_pmu.read = _x86_pmu_read;
 
-	if (!x86_pmu.guest_get_msrs)
-		x86_pmu.guest_get_msrs = perf_guest_get_msrs_nop;

I suspect I might've been over eager here and we're now in trouble when *_pmu_init() clears x86_pmu.guest_get_msrs (like for instance on AMD).

Quoted 5 lines

When that happens we need to restore __static_call_return0, otherwise the following static_call_update() will patch in a NOP and RAX will be garbage again.

So I've taken the liberty to update the patch as below.

Doh, I managed to forget about that between v1 and v2, too. Thanks much!

#6 lore raw
In reply to Sean Christopherson
The following commit has been merged into the perf/urgent branch of tip:

The following commit has been merged into the perf/urgent branch of tip:

Commit-ID:     c8e2fe13d1d1f3a02842b7b909d4e4846a4b6a2c
Gitweb:        https://git.kernel.org/tip/c8e2fe13d1d1f3a02842b7b909d4e4846a4b6a2c
Author:        Sean Christopherson <[email protected]>
AuthorDate:    Tue, 09 Mar 2021 09:10:19 -08:00
Committer:     Peter Zijlstra <[email protected]>
CommitterDate: Wed, 10 Mar 2021 16:45:09 +01:00

x86/perf: Use RET0 as default for guest_get_msrs to handle "no PMU" case

Initialize x86_pmu.guest_get_msrs to return 0/NULL to handle the "nop" case. Patching in perf_guest_get_msrs_nop() during setup does not work if there is no PMU, as setup bails before updating the static calls, leaving x86_pmu.guest_get_msrs NULL and thus a complete nop. Ultimately, this causes VMX abort on VM-Exit due to KVM putting random garbage from the stack into the MSR load list.

Add a comment in KVM to note that nr_msrs is valid if and only if the return value is non-NULL.

---
 arch/x86/events/core.c | 15 ++++++---------
 arch/x86/kvm/vmx/vmx.c |  2 +-
 2 files changed, 7 insertions(+), 10 deletions(-)
arch/x86/events/core.c +6 -9
index 6ddeed3..18df171 100644
@@ -81,7 +81,11 @@ DEFINE_STATIC_CALL_NULL(x86_pmu_swap_task_ctx, *x86_pmu.swap_task_ctx);
 DEFINE_STATIC_CALL_NULL(x86_pmu_drain_pebs,   *x86_pmu.drain_pebs);
 DEFINE_STATIC_CALL_NULL(x86_pmu_pebs_aliases, *x86_pmu.pebs_aliases);
 
-DEFINE_STATIC_CALL_NULL(x86_pmu_guest_get_msrs,  *x86_pmu.guest_get_msrs);
+/*
+ * This one is magic, it will get called even when PMU init fails (because
+ * there is no PMU), in which case it should simply return NULL.
+ */
+DEFINE_STATIC_CALL_RET0(x86_pmu_guest_get_msrs, *x86_pmu.guest_get_msrs);
 
 u64 __read_mostly hw_cache_event_ids
 				[PERF_COUNT_HW_CACHE_MAX]
@@ -1944,13 +1948,6 @@ static void _x86_pmu_read(struct perf_event *event)
 	x86_perf_event_update(event);
 }
 
-static inline struct perf_guest_switch_msr *
-perf_guest_get_msrs_nop(int *nr)
-{
-	*nr = 0;
-	return NULL;
-}
-
 static int __init init_hw_perf_events(void)
 {
 	struct x86_pmu_quirk *quirk;
@@ -2025,7 +2022,7 @@ static int __init init_hw_perf_events(void)
 		x86_pmu.read = _x86_pmu_read;
 
 	if (!x86_pmu.guest_get_msrs)
-		x86_pmu.guest_get_msrs = perf_guest_get_msrs_nop;
+		x86_pmu.guest_get_msrs = (void *)&__static_call_return0;
 
 	x86_pmu_static_call_update();
 
arch/x86/kvm/vmx/vmx.c +1 -1
index 50810d4..32cf828 100644
@@ -6580,8 +6580,8 @@ static void atomic_switch_perf_msrs(struct vcpu_vmx *vmx)
 	int i, nr_msrs;
 	struct perf_guest_switch_msr *msrs;
 
+	/* Note, nr_msrs may be garbage if perf_guest_get_msrs() returns NULL. */
 	msrs = perf_guest_get_msrs(&nr_msrs);
-
 	if (!msrs)
 		return;