Re: [RFC][PATCH v2 02/11] perf: core, add pmu register and lookup functions

Re: [RFC][PATCH v2 02/11] perf: core, add pmu register and lookup functions

[Cyrill Gorcunov]

On Wed, May 19, 2010 at 01:45:02AM +0000, Lin Ming wrote:
...
> +int perf_event_register_pmu(struct pmu *pmu)
> +{
> +	struct pmu *tmp;
> +	int ret = 1;
> +
> +	rcu_read_lock();
> +	list_for_each_entry_rcu(tmp, &pmus, entry) {
> +		if (tmp->id == pmu->id) {
> +			ret = 0;
> +			goto err;

Hi Ming, is it supposed to get out in locked state?
I mean imbalanced rcu_read_lock somehow suspicicous ;)

> +		}
> +	}
> +	rcu_read_unlock();
> +
> +	spin_lock(&pmus_lock);
> +	list_add_tail_rcu(&pmu->entry, &pmus);
> +	spin_unlock(&pmus_lock);
> +
> +err:
> +	return ret;
> +}
> +
...

	-- Cyrill

[RFC][PATCH v2 02/11] perf: core, add pmu register and lookup functions

[Lin Ming]

And add new api pmu::init_event

Changes log
v2: Use RCU for synchronization (Peter Zijlstra)
v1: add pmu register and lookup functions

Signed-off-by: Lin Ming <ming.m.lin@intel.com>
---
 include/linux/perf_event.h |   11 ++++++
 kernel/perf_event.c        |   73 ++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 84 insertions(+), 0 deletions(-)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 6c01c5f..40809f5 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -549,10 +549,16 @@ struct perf_event;
 
 #define PERF_EVENT_TXN_STARTED 1
 
+#define PMU_TYPE_CPU 0
+#define PMU_TYPE_NODE 1
+
 /**
  * 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 +575,8 @@ 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);
 };
 
 /**
@@ -1013,6 +1021,9 @@ 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 int perf_event_register_pmu(struct pmu *pmu);
+extern void perf_event_unregister_pmu(int id);
 #else
 static inline void
 perf_event_task_sched_in(struct task_struct *task)			{ }
diff --git a/kernel/perf_event.c b/kernel/perf_event.c
index ba7a37a..31b032b 100644
--- a/kernel/perf_event.c
+++ b/kernel/perf_event.c
@@ -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 DEFINE_SPINLOCK(pmus_lock);
+
 int perf_max_events __read_mostly = 1;
 static int perf_reserved_percpu __read_mostly;
 static int perf_overcommit __read_mostly = 1;
@@ -4678,6 +4684,34 @@ 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 = NULL;
+	int pmu_id = event->attr.type;
+
+	switch (event->attr.type) {
+		case PERF_TYPE_HARDWARE:
+		case PERF_TYPE_HW_CACHE:
+		case PERF_TYPE_RAW:
+			pmu_id = PMU_TYPE_CPU;
+			break;
+
+		/* TBD: will add other pmu type later */
+
+		default:
+			return NULL;
+	}
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(pmu, &pmus, entry) {
+		if (pmu->id == pmu_id)
+			break;
+	}
+	rcu_read_unlock();
+
+	return pmu;
+}
+
 /*
  * Allocate and initialize a event structure
  */
@@ -5635,6 +5669,8 @@ 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,
@@ -5734,3 +5770,40 @@ static int __init perf_event_sysfs_init(void)
 				  &perfclass_attr_group);
 }
 device_initcall(perf_event_sysfs_init);
+
+int perf_event_register_pmu(struct pmu *pmu)
+{
+	struct pmu *tmp;
+	int ret = 1;
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(tmp, &pmus, entry) {
+		if (tmp->id == pmu->id) {
+			ret = 0;
+			goto err;
+		}
+	}
+	rcu_read_unlock();
+
+	spin_lock(&pmus_lock);
+	list_add_tail_rcu(&pmu->entry, &pmus);
+	spin_unlock(&pmus_lock);
+
+err:
+	return ret;
+}
+
+void perf_event_unregister_pmu(int id)
+{
+	struct pmu *tmp = NULL;
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(tmp, &pmus, entry) {
+		if (tmp->id == id)
+			break;
+	}
+	rcu_read_unlock();
+
+	if (tmp)
+		list_del_rcu(&tmp->entry);
+}

Re: [RFC][PATCH v2 02/11] perf: core, add pmu register and lookup functions

[Lin Ming]

On Wed, 2010-05-19 at 02:38 +0800, Cyrill Gorcunov wrote:
> On Wed, May 19, 2010 at 01:45:02AM +0000, Lin Ming wrote:
> ...
> > +int perf_event_register_pmu(struct pmu *pmu)
> > +{
> > +	struct pmu *tmp;
> > +	int ret = 1;
> > +
> > +	rcu_read_lock();
> > +	list_for_each_entry_rcu(tmp, &pmus, entry) {
> > +		if (tmp->id == pmu->id) {
> > +			ret = 0;
> > +			goto err;
> 
> Hi Ming, is it supposed to get out in locked state?
> I mean imbalanced rcu_read_lock somehow suspicicous ;)

Ah, yes, that's real bug.
I'll fix it.

Thanks for review.

> 
> > +		}
> > +	}
> > +	rcu_read_unlock();
> > +
> > +	spin_lock(&pmus_lock);
> > +	list_add_tail_rcu(&pmu->entry, &pmus);
> > +	spin_unlock(&pmus_lock);
> > +
> > +err:
> > +	return ret;
> > +}
> > +
> ...
> 
> 	-- Cyrill