Skip to contentSkip to conversation

perf: core, remove hw_perf_event_init

RFC PATCH Superseded Lin Ming 4 participants 18 messages May 10, 2010 – May 12, 2010 perf events

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Lin Ming

author
#0 lore raw

RFC PATCH 2/9 perf: core, remove hw_perf_event_init

A new field "pmu_id" is added to struct perf_event_attr.

2 new functions: perf_event_register_pmu, perf_event_lookup_pmu
perf_event_register_pmu: the pmu registration facility
perf_event_lookup_pmu: lookup the pmu via the passed in event->attr.pmu_id.

A new api pmu->init_event to replace hw_perf_event_init

---
 include/linux/perf_event.h |    9 +++++++++
 kernel/perf_event.c        |   39 +++++++++++++++++++++++++++++++--------
 2 files changed, 40 insertions(+), 8 deletions(-)
include/linux/perf_event.h +9 -0
index 5856d3b..6246c99 100644
@@ -225,6 +225,8 @@ struct perf_event_attr {
 	__u32			bp_type;
 	__u64			bp_addr;
 	__u64			bp_len;
+
+	int			pmu_id;
 };
 
 /*
@@ -553,6 +555,9 @@ struct perf_event;
  * struct pmu - generic performance monitoring unit
  */
 struct pmu {
+	int id;
+	struct list_head entry;
+
 	int (*enable)			(struct perf_event *event);
 	void (*disable)			(struct perf_event *event);
 	int (*start)			(struct perf_event *event);
@@ -569,6 +574,8 @@ struct pmu {
 	void (*start_txn)	(struct pmu *pmu);
 	void (*cancel_txn)	(struct pmu *pmu);
 	int  (*commit_txn)	(struct pmu *pmu);
+
+	int  (*init_event)	(struct perf_event *event);
 };
 
 /**
@@ -1014,6 +1021,8 @@ extern int perf_swevent_get_recursion_context(void);
 extern void perf_swevent_put_recursion_context(int rctx);
 extern void perf_event_enable(struct perf_event *event);
 extern void perf_event_disable(struct perf_event *event);
+
+extern void perf_event_register_pmu(struct pmu *pmu);
 #else
 static inline void
 perf_event_task_sched_in(struct task_struct *task)			{ }
kernel/perf_event.c +31 -8
index 36baf85..f19d40e 100644
@@ -40,6 +40,12 @@
  */
 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
 
+/*
+ * The list of multiple hw pmus
+ */
+static struct list_head pmus;
+static int pmu_id_curr;
+
 int perf_max_events __read_mostly = 1;
 static int perf_reserved_percpu __read_mostly;
 static int perf_overcommit __read_mostly = 1;
@@ -75,11 +81,6 @@ static DEFINE_SPINLOCK(perf_resource_lock);
 /*
  * Architecture provided APIs - weak aliases:
  */
-extern __weak struct pmu *hw_perf_event_init(struct perf_event *event)
-{
-	return NULL;
-}
-
 void __weak hw_perf_disable(void)		{ barrier(); }
 void __weak hw_perf_enable(void)		{ barrier(); }
 
@@ -4670,6 +4671,19 @@ static struct pmu *sw_perf_event_init(struct perf_event *event)
 	return pmu;
 }
 
+static struct pmu *perf_event_lookup_pmu(struct perf_event *event)
+{
+	struct pmu *pmu;
+	int pmu_id = event->attr.pmu_id;
+
+	list_for_each_entry(pmu, &pmus, entry) {
+		if (pmu->id == pmu_id)
+			return pmu;
+	}
+
+	return NULL;
+}
+
 /*
  * Allocate and initialize a event structure
  */
@@ -4685,7 +4699,7 @@ perf_event_alloc(struct perf_event_attr *attr,
 	struct pmu *pmu;
 	struct perf_event *event;
 	struct hw_perf_event *hwc;
-	long err;
+	long err = 0;
 
 	event = kzalloc(sizeof(*event), gfpflags);
 	if (!event)
@@ -4750,7 +4764,9 @@ perf_event_alloc(struct perf_event_attr *attr,
 	case PERF_TYPE_RAW:
 	case PERF_TYPE_HARDWARE:
 	case PERF_TYPE_HW_CACHE:
-		pmu = hw_perf_event_init(event);
+		pmu = perf_event_lookup_pmu(event);
+		if (pmu && pmu->init_event)
+			err = pmu->init_event(event);
 		break;
 
 	case PERF_TYPE_SOFTWARE:
@@ -4770,7 +4786,6 @@ perf_event_alloc(struct perf_event_attr *attr,
 		break;
 	}
 done:
-	err = 0;
 	if (!pmu)
 		err = -EINVAL;
 	else if (IS_ERR(pmu))
@@ -5627,6 +5642,8 @@ void __init perf_event_init(void)
 	perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
 			(void *)(long)smp_processor_id());
 	register_cpu_notifier(&perf_cpu_nb);
+
+	INIT_LIST_HEAD(&pmus);
 }
 
 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
@@ -5726,3 +5743,9 @@ static int __init perf_event_sysfs_init(void)
 				  &perfclass_attr_group);
 }
 device_initcall(perf_event_sysfs_init);
+
+void perf_event_register_pmu(struct pmu *pmu)
+{
+	pmu->id = pmu_id_curr++;
+	list_add_tail(&pmu->entry, &pmus);
+}
maintainer
#1 lore raw
Quoted 8 lines · Lin Ming
kernel/perf_event.c +8 -0
+static struct pmu *perf_event_lookup_pmu(struct perf_event *event)
+{
+       struct pmu *pmu;
+       int pmu_id = event->attr.pmu_id;
+
+       list_for_each_entry(pmu, &pmus, entry) {
+               if (pmu->id == pmu_id)
+                       return pmu;

kernel/perf_event.c static struct pmu *sw_perf_event_init(struct perf_event *event)

Quoted 10 lines
kernel/perf_event.c +9 -0
+       }
+
+       return NULL;
+}
 
+void perf_event_register_pmu(struct pmu *pmu)
+{
+       pmu->id = pmu_id_curr++;
+       list_add_tail(&pmu->entry, &pmus);
+} 

That will be wanting some sort of synchronization

author
#2 lore raw
Quoted 14 lines · Peter Zijlstra
Quoted 12 lines · Lin Ming
kernel/perf_event.c +12 -0
+static struct pmu *perf_event_lookup_pmu(struct perf_event *event)
+{
+       struct pmu *pmu;
+       int pmu_id = event->attr.pmu_id;
+
+       list_for_each_entry(pmu, &pmus, entry) {
+               if (pmu->id == pmu_id)
+                       return pmu;
+       }
+
+       return NULL;
+}
Quoted 7 lines

kernel/perf_event.c static int __init perf_event_sysfs_init(void) · line 5751

Quoted 5 lines
kernel/perf_event.c +5 -0
+void perf_event_register_pmu(struct pmu *pmu)
+{
+       pmu->id = pmu_id_curr++;
+       list_add_tail(&pmu->entry, &pmus);
+} 

That will be wanting some sort of synchronization

Will add a mutex to protect the list of pmus.

maintainer
#3 lore raw
Quoted 23 lines · Lin Ming
Quoted 14 lines · Peter Zijlstra
Quoted 12 lines · Lin Ming
kernel/perf_event.c +12 -0
+static struct pmu *perf_event_lookup_pmu(struct perf_event *event)
+{
+       struct pmu *pmu;
+       int pmu_id = event->attr.pmu_id;
+
+       list_for_each_entry(pmu, &pmus, entry) {
+               if (pmu->id == pmu_id)
+                       return pmu;
+       }
+
+       return NULL;
+}
Quoted 7 lines

kernel/perf_event.c static int __init perf_event_sysfs_init(void) · line 5751

Quoted 5 lines
kernel/perf_event.c +5 -0
+void perf_event_register_pmu(struct pmu *pmu)
+{
+       pmu->id = pmu_id_curr++;
+       list_add_tail(&pmu->entry, &pmus);
+} 

That will be wanting some sort of synchronization

Quoted 1 line

Will add a mutex to protect the list of pmus.

I'm thinking RCU might be better suited, a mutex for lookup doesn't sound ideal.

author
#4 lore raw
Quoted 24 lines · Peter Zijlstra
Quoted 23 lines · Lin Ming
Quoted 21 lines · Peter Zijlstra
Quoted 8 lines · Lin Ming
kernel/perf_event.c +8 -0
+static struct pmu *perf_event_lookup_pmu(struct perf_event *event)
+{
+       struct pmu *pmu;
+       int pmu_id = event->attr.pmu_id;
+
+       list_for_each_entry(pmu, &pmus, entry) {
+               if (pmu->id == pmu_id)
+                       return pmu;

kernel/perf_event.c static struct pmu *sw_perf_event_init(struct perf_event *event)

Quoted 10 lines
kernel/perf_event.c +9 -0
+       }
+
+       return NULL;
+}
 
+void perf_event_register_pmu(struct pmu *pmu)
+{
+       pmu->id = pmu_id_curr++;
+       list_add_tail(&pmu->entry, &pmus);
+} 

That will be wanting some sort of synchronization

Quoted 4 lines
Quoted 1 line

Will add a mutex to protect the list of pmus.

I'm thinking RCU might be better suited, a mutex for lookup doesn't sound ideal.

Thanks for the idea, I'll do that.

#5 lore raw
In reply to Peter Zijlstra
Quoted 24 lines · Peter Zijlstra
Quoted 23 lines · Lin Ming
Quoted 21 lines · Peter Zijlstra
Quoted 8 lines · Lin Ming
kernel/perf_event.c +8 -0
+static struct pmu *perf_event_lookup_pmu(struct perf_event *event)
+{
+       struct pmu *pmu;
+       int pmu_id = event->attr.pmu_id;
+
+       list_for_each_entry(pmu, &pmus, entry) {
+               if (pmu->id == pmu_id)
+                       return pmu;

kernel/perf_event.c static struct pmu *sw_perf_event_init(struct perf_event *event)

Quoted 10 lines
kernel/perf_event.c +9 -0
+       }
+
+       return NULL;
+}
 
+void perf_event_register_pmu(struct pmu *pmu)
+{
+       pmu->id = pmu_id_curr++;
+       list_add_tail(&pmu->entry, &pmus);
+} 

That will be wanting some sort of synchronization

Quoted 4 lines
Quoted 1 line

Will add a mutex to protect the list of pmus.

I'm thinking RCU might be better suited, a mutex for lookup doesn't sound ideal.

Is it really needed? I expect this function to be called on boot only.

In fact I would even suggest to tag it as __init.

maintainer
#6 lore raw
Quoted 27 lines · Frederic Weisbecker
Quoted 26 lines · Peter Zijlstra
Quoted 24 lines · Lin Ming
Quoted 14 lines · Peter Zijlstra
Quoted 12 lines · Lin Ming
kernel/perf_event.c +12 -0
+static struct pmu *perf_event_lookup_pmu(struct perf_event *event)
+{
+       struct pmu *pmu;
+       int pmu_id = event->attr.pmu_id;
+
+       list_for_each_entry(pmu, &pmus, entry) {
+               if (pmu->id == pmu_id)
+                       return pmu;
+       }
+
+       return NULL;
+}
Quoted 7 lines

kernel/perf_event.c static int __init perf_event_sysfs_init(void) · line 5751

Quoted 5 lines
kernel/perf_event.c +5 -0
+void perf_event_register_pmu(struct pmu *pmu)
+{
+       pmu->id = pmu_id_curr++;
+       list_add_tail(&pmu->entry, &pmus);
+} 

That will be wanting some sort of synchronization

Will add a mutex to protect the list of pmus.

Quoted 8 lines
Quoted 2 lines

I'm thinking RCU might be better suited, a mutex for lookup doesn't sound ideal.

Is it really needed? I expect this function to be called on boot only.

In fact I would even suggest to tag it as __init.

Loadable modules as well as PCI-Hotplug need supporting.

#7 lore raw
Quoted 32 lines · Peter Zijlstra
Quoted 31 lines · Frederic Weisbecker
Quoted 28 lines · Peter Zijlstra
Quoted 23 lines · Lin Ming
Quoted 14 lines · Peter Zijlstra
Quoted 12 lines · Lin Ming
kernel/perf_event.c +12 -0
+static struct pmu *perf_event_lookup_pmu(struct perf_event *event)
+{
+       struct pmu *pmu;
+       int pmu_id = event->attr.pmu_id;
+
+       list_for_each_entry(pmu, &pmus, entry) {
+               if (pmu->id == pmu_id)
+                       return pmu;
+       }
+
+       return NULL;
+}
Quoted 7 lines

kernel/perf_event.c static int __init perf_event_sysfs_init(void) · line 5751

Quoted 5 lines
kernel/perf_event.c +5 -0
+void perf_event_register_pmu(struct pmu *pmu)
+{
+       pmu->id = pmu_id_curr++;
+       list_add_tail(&pmu->entry, &pmus);
+} 

That will be wanting some sort of synchronization

Quoted 1 line

Will add a mutex to protect the list of pmus.

I'm thinking RCU might be better suited, a mutex for lookup doesn't sound ideal.

Quoted 6 lines
Quoted 4 lines

Is it really needed? I expect this function to be called on boot only.

In fact I would even suggest to tag it as __init.

Loadable modules as well as PCI-Hotplug need supporting.

Which module do you have in mind that could register a pmu? And I don't understand the problem with pci-hotplug.

Thanks.

maintainer
#8 lore raw
Quoted 2 lines · Frederic Weisbecker

Which module do you have in mind that could register a pmu? And I don't understand the problem with pci-hotplug.

Well, the DRM drivers for one, but basically the trend is for every aspect of the machine to include PMUs of some form, this includes bus bridges and fancy devices.

The point about PCI-hotplug is that GPUs live on the PCI bus and once they start adding their PMU drivers we need to be able to unplug them.

Same for when the PCI bridge devices start featuring PMU implementations.

author
#9 lore raw
Quoted 8 lines · Peter Zijlstra
Quoted 2 lines · Frederic Weisbecker

Which module do you have in mind that could register a pmu? And I don't understand the problem with pci-hotplug.

Well, the DRM drivers for one, but basically the trend is for every aspect of the machine to include PMUs of some form, this includes bus bridges and fancy devices.

So the concept of "PMU" is really broadened here? Not just only the performance monitoring hardware in CPU?

Quoted 6 lines

The point about PCI-hotplug is that GPUs live on the PCI bus and once they start adding their PMU drivers we need to be able to unplug them.

Same for when the PCI bridge devices start featuring PMU implementations.

#10 lore raw
Quoted 6 lines · Lin Ming
Quoted 5 lines · Peter Zijlstra
Quoted 2 lines · Frederic Weisbecker

Which module do you have in mind that could register a pmu? And I don't understand the problem with pci-hotplug.

Quoted 6 lines
Quoted 3 lines

Well, the DRM drivers for one, but basically the trend is for every aspect of the machine to include PMUs of some form, this includes bus bridges and fancy devices.

So the concept of "PMU" is really broadened here? Not just only the performance monitoring hardware in CPU?

We _really_ dont want to call it a 'PMU' but 'events coming from an event source'.

The reason is that a PMU is an existing term that is quite attached to a CPU - while many hardware events come not from a PMU. Interrupts, error conditions, hotplug events, etc. etc.

Furthermore, the name 'PMU' is even less correct for software events.

So lets stick with 'events' and with some container that originates them. (event_source) Ok?

Ingo

author
#11 lore raw
Quoted 18 lines · Ingo Molnar
Quoted 6 lines · Lin Ming
Quoted 5 lines · Peter Zijlstra
Quoted 2 lines · Frederic Weisbecker

Which module do you have in mind that could register a pmu? And I don't understand the problem with pci-hotplug.

Quoted 6 lines
Quoted 3 lines

Well, the DRM drivers for one, but basically the trend is for every aspect of the machine to include PMUs of some form, this includes bus bridges and fancy devices.

So the concept of "PMU" is really broadened here? Not just only the performance monitoring hardware in CPU?

We _really_ dont want to call it a 'PMU' but 'events coming from an event source'.

Quoted 8 lines

The reason is that a PMU is an existing term that is quite attached to a CPU - while many hardware events come not from a PMU. Interrupts, error conditions, hotplug events, etc. etc.

Furthermore, the name 'PMU' is even less correct for software events.

So lets stick with 'events' and with some container that originates them. (event_source) Ok?

Yeah, understand.
"events coming from an event source" is a good clarification.

Thanks.

maintainer
#12 lore raw
In reply to Lin Ming
Quoted 6 lines · Lin Ming
Quoted 5 lines · Peter Zijlstra
Quoted 2 lines · Frederic Weisbecker

Which module do you have in mind that could register a pmu? And I don't understand the problem with pci-hotplug.

Quoted 6 lines
Quoted 3 lines

Well, the DRM drivers for one, but basically the trend is for every aspect of the machine to include PMUs of some form, this includes bus bridges and fancy devices.

So the concept of "PMU" is really broadened here? Not just only the performance monitoring hardware in CPU?

Arguably the uncore unit is already outside of what we traditionally call a CPU ;-)

But yeah, PMUs are found in lots of things these days.

#13 lore raw
In reply to Peter Zijlstra
Quoted 9 lines · Peter Zijlstra
Quoted 2 lines · Frederic Weisbecker

Which module do you have in mind that could register a pmu? And I don't understand the problem with pci-hotplug.

Well, the DRM drivers for one, but basically the trend is for every aspect of the machine to include PMUs of some form, this includes bus bridges and fancy devices.

Quoted 5 lines

The point about PCI-hotplug is that GPUs live on the PCI bus and once they start adding their PMU drivers we need to be able to unplug them.

Same for when the PCI bridge devices start featuring PMU implementations.

Ah ok.

Thanks.

maintainer
#14 lore raw
In reply to Lin Ming
Quoted 7 lines · Lin Ming

A new field "pmu_id" is added to struct perf_event_attr.

2 new functions: perf_event_register_pmu, perf_event_lookup_pmu
perf_event_register_pmu: the pmu registration facility
perf_event_lookup_pmu: lookup the pmu via the passed in event->attr.pmu_id.

A new api pmu->init_event to replace hw_perf_event_init

Another problem, this patch breaks all of perf by removing the weak call before there is a replacement.

Please rework such that there is a smooth transition.

Also, for compatibilities sake, the cpu pmu will be required to get pmu_id==0, this doesn't provide that.

author
#15 lore raw
Quoted 7 lines · Peter Zijlstra
Quoted 6 lines · Lin Ming

A new field "pmu_id" is added to struct perf_event_attr.

2 new functions: perf_event_register_pmu, perf_event_lookup_pmu
perf_event_register_pmu: the pmu registration facility
perf_event_lookup_pmu: lookup the pmu via the passed in event->attr.pmu_id.

Quoted 6 lines
Quoted 1 line

A new api pmu->init_event to replace hw_perf_event_init

Another problem, this patch breaks all of perf by removing the weak call before there is a replacement.

Please rework such that there is a smooth transition.

OK, I'll rework

Quoted 4 lines

Also, for compatibilities sake, the cpu pmu will be required to get pmu_id==0, this doesn't provide that.

Will provide that.

Thanks,
Lin Ming

#16 lore raw
In reply to Lin Ming
Quoted 114 lines · Lin Ming

A new field "pmu_id" is added to struct perf_event_attr.

2 new functions: perf_event_register_pmu, perf_event_lookup_pmu
perf_event_register_pmu: the pmu registration facility
perf_event_lookup_pmu: lookup the pmu via the passed in event->attr.pmu_id.

A new api pmu->init_event to replace hw_perf_event_init

---
 include/linux/perf_event.h |    9 +++++++++
 kernel/perf_event.c        |   39 +++++++++++++++++++++++++++++++--------
 2 files changed, 40 insertions(+), 8 deletions(-)
include/linux/perf_event.h +9 -0
index 5856d3b..6246c99 100644
@@ -225,6 +225,8 @@ struct perf_event_attr {
 	__u32			bp_type;
 	__u64			bp_addr;
 	__u64			bp_len;
+
+	int			pmu_id;
 };
 
 /*
@@ -553,6 +555,9 @@ struct perf_event;
  * struct pmu - generic performance monitoring unit
  */
 struct pmu {
+	int id;
+	struct list_head entry;
+
 	int (*enable)			(struct perf_event *event);
 	void (*disable)			(struct perf_event *event);
 	int (*start)			(struct perf_event *event);
@@ -569,6 +574,8 @@ struct pmu {
 	void (*start_txn)	(struct pmu *pmu);
 	void (*cancel_txn)	(struct pmu *pmu);
 	int  (*commit_txn)	(struct pmu *pmu);
+
+	int  (*init_event)	(struct perf_event *event);
 };
 
 /**
@@ -1014,6 +1021,8 @@ extern int perf_swevent_get_recursion_context(void);
 extern void perf_swevent_put_recursion_context(int rctx);
 extern void perf_event_enable(struct perf_event *event);
 extern void perf_event_disable(struct perf_event *event);
+
+extern void perf_event_register_pmu(struct pmu *pmu);
 #else
 static inline void
 perf_event_task_sched_in(struct task_struct *task)			{ }
kernel/perf_event.c +20 -6
index 36baf85..f19d40e 100644
@@ -40,6 +40,12 @@
  */
 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
 
+/*
+ * The list of multiple hw pmus
+ */
+static struct list_head pmus;
+static int pmu_id_curr;
+
 int perf_max_events __read_mostly = 1;
 static int perf_reserved_percpu __read_mostly;
 static int perf_overcommit __read_mostly = 1;
@@ -75,11 +81,6 @@ static DEFINE_SPINLOCK(perf_resource_lock);
 /*
  * Architecture provided APIs - weak aliases:
  */
-extern __weak struct pmu *hw_perf_event_init(struct perf_event *event)
-{
-	return NULL;
-}
-
 void __weak hw_perf_disable(void)		{ barrier(); }
 void __weak hw_perf_enable(void)		{ barrier(); }
 
@@ -4670,6 +4671,19 @@ static struct pmu *sw_perf_event_init(struct perf_event *event)
 	return pmu;
 }
 
+static struct pmu *perf_event_lookup_pmu(struct perf_event *event)
+{
+	struct pmu *pmu;
+	int pmu_id = event->attr.pmu_id;
+
+	list_for_each_entry(pmu, &pmus, entry) {
+		if (pmu->id == pmu_id)
+			return pmu;
+	}
+
+	return NULL;
+}
+
 /*
  * Allocate and initialize a event structure
  */
@@ -4685,7 +4699,7 @@ perf_event_alloc(struct perf_event_attr *attr,
 	struct pmu *pmu;
 	struct perf_event *event;
 	struct hw_perf_event *hwc;
-	long err;
+	long err = 0;
 
 	event = kzalloc(sizeof(*event), gfpflags);
 	if (!event)

kernel/perf_event.c perf_event_alloc(struct perf_event_attr *attr, · line 4769

Quoted 7 lines
kernel/perf_event.c +3 -1
@@ -4750,7 +4764,9 @@ perf_event_alloc(struct perf_event_attr *attr,
 	case PERF_TYPE_RAW:
 	case PERF_TYPE_HARDWARE:
 	case PERF_TYPE_HW_CACHE:
-		pmu = hw_perf_event_init(event);
+		pmu = perf_event_lookup_pmu(event);
+		if (pmu && pmu->init_event)
+			err = pmu->init_event(event);

Having the same for software, tracepoints and breakpoints events would be nice, so that we have a single simple path in perf_event_alloc to initialize the event.

Thanks.

author
#17 lore raw
Quoted 112 lines · Frederic Weisbecker
Quoted 111 lines · Lin Ming

A new field "pmu_id" is added to struct perf_event_attr.

2 new functions: perf_event_register_pmu, perf_event_lookup_pmu
perf_event_register_pmu: the pmu registration facility
perf_event_lookup_pmu: lookup the pmu via the passed in event->attr.pmu_id.

A new api pmu->init_event to replace hw_perf_event_init

---
 include/linux/perf_event.h |    9 +++++++++
 kernel/perf_event.c        |   39 +++++++++++++++++++++++++++++++--------
 2 files changed, 40 insertions(+), 8 deletions(-)
include/linux/perf_event.h +9 -0
index 5856d3b..6246c99 100644
@@ -225,6 +225,8 @@ struct perf_event_attr {
 	__u32			bp_type;
 	__u64			bp_addr;
 	__u64			bp_len;
+
+	int			pmu_id;
 };
 
 /*
@@ -553,6 +555,9 @@ struct perf_event;
  * struct pmu - generic performance monitoring unit
  */
 struct pmu {
+	int id;
+	struct list_head entry;
+
 	int (*enable)			(struct perf_event *event);
 	void (*disable)			(struct perf_event *event);
 	int (*start)			(struct perf_event *event);
@@ -569,6 +574,8 @@ struct pmu {
 	void (*start_txn)	(struct pmu *pmu);
 	void (*cancel_txn)	(struct pmu *pmu);
 	int  (*commit_txn)	(struct pmu *pmu);
+
+	int  (*init_event)	(struct perf_event *event);
 };
 
 /**
@@ -1014,6 +1021,8 @@ extern int perf_swevent_get_recursion_context(void);
 extern void perf_swevent_put_recursion_context(int rctx);
 extern void perf_event_enable(struct perf_event *event);
 extern void perf_event_disable(struct perf_event *event);
+
+extern void perf_event_register_pmu(struct pmu *pmu);
 #else
 static inline void
 perf_event_task_sched_in(struct task_struct *task)			{ }
kernel/perf_event.c +20 -6
index 36baf85..f19d40e 100644
@@ -40,6 +40,12 @@
  */
 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
 
+/*
+ * The list of multiple hw pmus
+ */
+static struct list_head pmus;
+static int pmu_id_curr;
+
 int perf_max_events __read_mostly = 1;
 static int perf_reserved_percpu __read_mostly;
 static int perf_overcommit __read_mostly = 1;
@@ -75,11 +81,6 @@ static DEFINE_SPINLOCK(perf_resource_lock);
 /*
  * Architecture provided APIs - weak aliases:
  */
-extern __weak struct pmu *hw_perf_event_init(struct perf_event *event)
-{
-	return NULL;
-}
-
 void __weak hw_perf_disable(void)		{ barrier(); }
 void __weak hw_perf_enable(void)		{ barrier(); }
 
@@ -4670,6 +4671,19 @@ static struct pmu *sw_perf_event_init(struct perf_event *event)
 	return pmu;
 }
 
+static struct pmu *perf_event_lookup_pmu(struct perf_event *event)
+{
+	struct pmu *pmu;
+	int pmu_id = event->attr.pmu_id;
+
+	list_for_each_entry(pmu, &pmus, entry) {
+		if (pmu->id == pmu_id)
+			return pmu;
+	}
+
+	return NULL;
+}
+
 /*
  * Allocate and initialize a event structure
  */
@@ -4685,7 +4699,7 @@ perf_event_alloc(struct perf_event_attr *attr,
 	struct pmu *pmu;
 	struct perf_event *event;
 	struct hw_perf_event *hwc;
-	long err;
+	long err = 0;
 
Quoted 16 lines

kernel/perf_event.c perf_event_alloc(struct perf_event_attr *attr, · line 4769

Quoted 10 lines
kernel/perf_event.c +3 -1
 	event = kzalloc(sizeof(*event), gfpflags);
 	if (!event)
@@ -4750,7 +4764,9 @@ perf_event_alloc(struct perf_event_attr *attr,
 	case PERF_TYPE_RAW:
 	case PERF_TYPE_HARDWARE:
 	case PERF_TYPE_HW_CACHE:
-		pmu = hw_perf_event_init(event);
+		pmu = perf_event_lookup_pmu(event);
+		if (pmu && pmu->init_event)
+			err = pmu->init_event(event);

Having the same for software, tracepoints and breakpoints events would be nice, so that we have a single simple path in perf_event_alloc to initialize the event.

Will add this later.
For now, let's first work on the hw pmus.

Quoted 3 lines

Thanks.