Re: [PATCH v6 02/16] perf/x86/intel: Handle guest PEBS overflow PMI for KVM guest

[Peter Zijlstra]

On Tue, May 11, 2021 at 10:42:00AM +0800, Like Xu wrote:
> With PEBS virtualization, the guest PEBS records get delivered to the
> guest DS, and the host pmi handler uses perf_guest_cbs->is_in_guest()
> to distinguish whether the PMI comes from the guest code like Intel PT.
> 
> No matter how many guest PEBS counters are overflowed, only triggering
> one fake event is enough. The fake event causes the KVM PMI callback to
> be called, thereby injecting the PEBS overflow PMI into the guest.
> 
> KVM may inject the PMI with BUFFER_OVF set, even if the guest DS is
> empty. That should really be harmless. Thus guest PEBS handler would
> retrieve the correct information from its own PEBS records buffer.
> 
> Originally-by: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> Co-developed-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> Signed-off-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> Signed-off-by: Like Xu <like.xu@xxxxxxxxxxxxxxx>
> ---
>  arch/x86/events/intel/core.c | 40 ++++++++++++++++++++++++++++++++++++
>  1 file changed, 40 insertions(+)
> 
> diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
> index b6e45ee10e16..092ecacf8345 100644
> --- a/arch/x86/events/intel/core.c
> +++ b/arch/x86/events/intel/core.c
> @@ -2780,6 +2780,43 @@ static void intel_pmu_reset(void)
>  	local_irq_restore(flags);
>  }
>  
> +/*
> + * We may be running with guest PEBS events created by KVM, and the
> + * PEBS records are logged into the guest's DS and invisible to host.
> + *
> + * In the case of guest PEBS overflow, we only trigger a fake event
> + * to emulate the PEBS overflow PMI for guest PBES counters in KVM.
> + * The guest will then vm-entry and check the guest DS area to read
> + * the guest PEBS records.
> + *
> + * The contents and other behavior of the guest event do not matter.
> + */
> +static void x86_pmu_handle_guest_pebs(struct pt_regs *regs,
> +				      struct perf_sample_data *data)
> +{
> +	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
> +	u64 guest_pebs_idxs = cpuc->pebs_enabled & ~cpuc->intel_ctrl_host_mask;
> +	struct perf_event *event = NULL;
> +	int bit;
> +
> +	if (!x86_pmu.pebs_active || !guest_pebs_idxs)
> +		return;
> +
> +	for_each_set_bit(bit, (unsigned long *)&guest_pebs_idxs,
> +			 INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed) {
> +		event = cpuc->events[bit];
> +		if (!event->attr.precise_ip)
> +			continue;
> +
> +		perf_sample_data_init(data, 0, event->hw.last_period);
> +		if (perf_event_overflow(event, data, regs))
> +			x86_pmu_stop(event, 0);
> +
> +		/* Inject one fake event is enough. */
> +		break;
> +	}
> +}
> +
>  static int handle_pmi_common(struct pt_regs *regs, u64 status)
>  {
>  	struct perf_sample_data data;
> @@ -2831,6 +2868,9 @@ static int handle_pmi_common(struct pt_regs *regs, u64 status)
>  		u64 pebs_enabled = cpuc->pebs_enabled;
>  
>  		handled++;
> +		if (x86_pmu.pebs_vmx && perf_guest_cbs &&
> +		    perf_guest_cbs->is_in_guest())
> +			x86_pmu_handle_guest_pebs(regs, &data);
>  		x86_pmu.drain_pebs(regs, &data);
>  		status &= intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;
>  

I'm thinking you have your conditions in the wrong order; would it not
be much cheaper to first check: '!x86_pmu.pebs_active || !guest_pebs_idx'
than to do that horrible indirect ->is_in_guest() call?

After all, if the guest doesn't have PEBS enabled, who cares if we're
currently in a guest or not.

Also, something like the below perhaps (arm64 and xen need fixing up at
the very least) could make all that perf_guest_cbs stuff suck less.


---
diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 8e509325c2c3..c8f8fb7c0536 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -90,6 +90,26 @@ DEFINE_STATIC_CALL_NULL(x86_pmu_pebs_aliases, *x86_pmu.pebs_aliases);
  */
 DEFINE_STATIC_CALL_RET0(x86_pmu_guest_get_msrs, *x86_pmu.guest_get_msrs);
 
+DEFINE_STATIC_CALL_RET0(x86_guest_state, *(perf_guest_cbs->state));
+DEFINE_STATIC_CALL_RET0(x86_guest_get_ip, *(perf_guest_cbs->get_ip));
+DEFINE_STATIC_CALL_RET0(x86_guest_handle_intel_pt_intr, *(perf_guest_cbs->handle_intel_pt_intr));
+
+void arch_perf_update_guest_cbs(void)
+{
+	static_call_update(x86_guest_state, (void *)&__static_call_return0);
+	static_call_update(x86_guest_get_ip, (void *)&__static_call_return0);
+	static_call_update(x86_guest_handle_intel_pt_intr, (void *)&__static_call_return0);
+
+	if (perf_guest_cbs && perf_guest_cbs->state)
+		static_call_update(x86_guest_state, perf_guest_cbs->state);
+
+	if (perf_guest_cbs && perf_guest_cbs->get_ip)
+		static_call_update(x86_guest_get_ip, perf_guest_cbs->get_ip);
+
+	if (perf_guest_cbs && perf_guest_cbs->handle_intel_pt_intr)
+		static_call_update(x86_guest_handle_intel_pt_intr, perf_guest_cbs->handle_intel_pt_intr);
+}
+
 u64 __read_mostly hw_cache_event_ids
 				[PERF_COUNT_HW_CACHE_MAX]
 				[PERF_COUNT_HW_CACHE_OP_MAX]
@@ -2736,7 +2756,7 @@ perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re
 	struct unwind_state state;
 	unsigned long addr;
 
-	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
+	if (static_call(x86_guest_state)()) {
 		/* TODO: We don't support guest os callchain now */
 		return;
 	}
@@ -2839,7 +2859,7 @@ perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs
 	struct stack_frame frame;
 	const struct stack_frame __user *fp;
 
-	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
+	if (static_call(x86_guest_state)()) {
 		/* TODO: We don't support guest os callchain now */
 		return;
 	}
@@ -2916,18 +2936,21 @@ static unsigned long code_segment_base(struct pt_regs *regs)
 
 unsigned long perf_instruction_pointer(struct pt_regs *regs)
 {
-	if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
-		return perf_guest_cbs->get_guest_ip();
+	unsigned long ip = static_call(x86_guest_get_ip)();
+
+	if (likely(!ip))
+		ip = regs->ip + code_segment_base(regs);
 
-	return regs->ip + code_segment_base(regs);
+	return ip;
 }
 
 unsigned long perf_misc_flags(struct pt_regs *regs)
 {
+	unsigned int guest = static_call(x86_guest_state)();
 	int misc = 0;
 
-	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
-		if (perf_guest_cbs->is_user_mode())
+	if (guest) {
+		if (guest & PERF_GUEST_USER)
 			misc |= PERF_RECORD_MISC_GUEST_USER;
 		else
 			misc |= PERF_RECORD_MISC_GUEST_KERNEL;
diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
index 2521d03de5e0..ac422c45f940 100644
--- a/arch/x86/events/intel/core.c
+++ b/arch/x86/events/intel/core.c
@@ -2780,6 +2780,8 @@ static void intel_pmu_reset(void)
 	local_irq_restore(flags);
 }
 
+DECLARE_STATIC_CALL(x86_guest_handle_intel_pt_intr, *(perf_guest_cbs->handle_intel_pt_intr));
+
 static int handle_pmi_common(struct pt_regs *regs, u64 status)
 {
 	struct perf_sample_data data;
@@ -2850,10 +2852,7 @@ static int handle_pmi_common(struct pt_regs *regs, u64 status)
 	 */
 	if (__test_and_clear_bit(GLOBAL_STATUS_TRACE_TOPAPMI_BIT, (unsigned long *)&status)) {
 		handled++;
-		if (unlikely(perf_guest_cbs && perf_guest_cbs->is_in_guest() &&
-			perf_guest_cbs->handle_intel_pt_intr))
-			perf_guest_cbs->handle_intel_pt_intr();
-		else
+		if (!static_call(x86_guest_handle_intel_pt_intr)())
 			intel_pt_interrupt();
 	}
 
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 55efbacfc244..2a24e615fa4a 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -1812,7 +1812,7 @@ int kvm_skip_emulated_instruction(struct kvm_vcpu *vcpu);
 int kvm_complete_insn_gp(struct kvm_vcpu *vcpu, int err);
 void __kvm_request_immediate_exit(struct kvm_vcpu *vcpu);
 
-int kvm_is_in_guest(void);
+unsigned int kvm_guest_state(void);
 
 void __user *__x86_set_memory_region(struct kvm *kvm, int id, gpa_t gpa,
 				     u32 size);
diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
index 827886c12c16..2dcbd1b30004 100644
--- a/arch/x86/kvm/pmu.c
+++ b/arch/x86/kvm/pmu.c
@@ -87,7 +87,7 @@ static void kvm_perf_overflow_intr(struct perf_event *perf_event,
 		 * woken up. So we should wake it, but this is impossible from
 		 * NMI context. Do it from irq work instead.
 		 */
-		if (!kvm_is_in_guest())
+		if (!kvm_guest_state())
 			irq_work_queue(&pmc_to_pmu(pmc)->irq_work);
 		else
 			kvm_make_request(KVM_REQ_PMI, pmc->vcpu);
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index bbc4e04e67ad..88f709b3759c 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -8035,44 +8035,46 @@ static void kvm_timer_init(void)
 DEFINE_PER_CPU(struct kvm_vcpu *, current_vcpu);
 EXPORT_PER_CPU_SYMBOL_GPL(current_vcpu);
 
-int kvm_is_in_guest(void)
+static unsigned int kvm_guest_state(void)
 {
-	return __this_cpu_read(current_vcpu) != NULL;
-}
-
-static int kvm_is_user_mode(void)
-{
-	int user_mode = 3;
+	struct kvm_vcpu *vcpu = __this_cpu_read(current_vcpu);
+	unsigned int state = 0;
 
-	if (__this_cpu_read(current_vcpu))
-		user_mode = static_call(kvm_x86_get_cpl)(__this_cpu_read(current_vcpu));
+	if (vcpu)
+		state |= PERF_GUEST_ACTIVE;
+	if (static_call(kvm_x86_get_cpl)(vcpu))
+		state |= PERF_GUEST_USER;
 
-	return user_mode != 0;
+	return state;
 }
 
-static unsigned long kvm_get_guest_ip(void)
+static unsigned long kvm_guest_get_ip(void)
 {
+	struct kvm_vcpu *vcpu = __this_cpu_read(current_vcpu);
 	unsigned long ip = 0;
 
-	if (__this_cpu_read(current_vcpu))
-		ip = kvm_rip_read(__this_cpu_read(current_vcpu));
+	if (vcpu)
+		ip = kvm_rip_read(vcpu);
 
 	return ip;
 }
 
-static void kvm_handle_intel_pt_intr(void)
+static unsigned int kvm_handle_intel_pt_intr(void)
 {
 	struct kvm_vcpu *vcpu = __this_cpu_read(current_vcpu);
 
+	if (!vcpu)
+		return 0;
+
 	kvm_make_request(KVM_REQ_PMI, vcpu);
 	__set_bit(MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI_BIT,
 			(unsigned long *)&vcpu->arch.pmu.global_status);
+	return 1;
 }
 
 static struct perf_guest_info_callbacks kvm_guest_cbs = {
-	.is_in_guest		= kvm_is_in_guest,
-	.is_user_mode		= kvm_is_user_mode,
-	.get_guest_ip		= kvm_get_guest_ip,
+	.state			= kvm_guest_state,
+	.get_ip			= kvm_guest_get_ip,
 	.handle_intel_pt_intr	= kvm_handle_intel_pt_intr,
 };
 
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index f5a6a2f069ed..7eae1fd22db3 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -26,11 +26,13 @@
 # include <asm/local64.h>
 #endif
 
+#define PERF_GUEST_ACTIVE	0x01
+#define PERF_GUEST_USER		0x02
+
 struct perf_guest_info_callbacks {
-	int				(*is_in_guest)(void);
-	int				(*is_user_mode)(void);
-	unsigned long			(*get_guest_ip)(void);
-	void				(*handle_intel_pt_intr)(void);
+	unsigned int			(*state)(void);
+	unsigned long			(*get_ip)(void);
+	unsigned int			(*handle_intel_pt_intr)(void);
 };
 
 #ifdef CONFIG_HAVE_HW_BREAKPOINT
@@ -1237,6 +1239,8 @@ extern void perf_event_bpf_event(struct bpf_prog *prog,
 				 u16 flags);
 
 extern struct perf_guest_info_callbacks *perf_guest_cbs;
+extern void __weak arch_perf_update_guest_cbs(void);
+
 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 2e947a485898..aec531fc9c90 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -6486,9 +6486,17 @@ static void perf_pending_event(struct irq_work *entry)
  */
 struct perf_guest_info_callbacks *perf_guest_cbs;
 
+void __weak arch_perf_update_guest_cbs(void)
+{
+}
+
 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
 {
+	if (WARN_ON_ONCE(perf_guest_cbs))
+		return -EBUSY;
+
 	perf_guest_cbs = cbs;
+	arch_perf_update_guest_cbs();
 	return 0;
 }
 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);