Skip to contentSkip to conversation

perf_counter: Implement more accurate per task statistics

tip:perfcounters/urgent In review tip-bot for Peter Zijlstra 2 participants 4 messages Jun 25, 2009 – Jun 26, 2009

lore
Share & tools
lore
#0 lore raw
Commit-ID:  bfbd3381e63aa2a14c6706afb50ce4630aa0d9a2
Gitweb:     http://git.kernel.org/tip/bfbd3381e63aa2a14c6706afb50ce4630aa0d9a2
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Wed, 24 Jun 2009 21:11:59 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Thu, 25 Jun 2009 21:39:07 +0200

perf_counter: Implement more accurate per task statistics

With the introduction of PERF_EVENT_READ we have the possibility to provide accurate counter values for individual tasks in a task hierarchy.

However, due to the lazy context switching used for similar counter contexts our current per task counts are way off.

In order to maintain some of the lazy switch benefits we don't disable it out-right, but simply iterate the active counters and flip the values between the contexts.

This only reads the counters but does not need to reprogram the full PMU.

LKML-Reference: <new-submission>

---
 include/linux/perf_counter.h |    4 ++-
 kernel/perf_counter.c        |   83 ++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 83 insertions(+), 4 deletions(-)
include/linux/perf_counter.h +3 -1
index 6a384f0..de70a10 100644
@@ -178,8 +178,9 @@ struct perf_counter_attr {
 				mmap           :  1, /* include mmap data     */
 				comm	       :  1, /* include comm data     */
 				freq           :  1, /* use freq, not period  */
+				inherit_stat   :  1, /* per task counts       */
 
-				__reserved_1   : 53;
+				__reserved_1   : 52;
 
 	__u32			wakeup_events;	/* wakeup every n events */
 	__u32			__reserved_2;
@@ -602,6 +603,7 @@ struct perf_counter_context {
 	int				nr_counters;
 	int				nr_active;
 	int				is_active;
+	int				nr_stat;
 	atomic_t			refcount;
 	struct task_struct		*task;
 
kernel/perf_counter.c +80 -3

125 lines not on this page · raw message

#1 lore raw

RFC PATCH 1/1 perf_counter: Complete counter swap.

Quoted 30 lines · tip-bot for Peter Zijlstra
(quoted diff) +30 -0
+static void __perf_counter_sync_stat(struct perf_counter *counter,
+				     struct perf_counter *next_counter)
+{
+	u64 value;
+
+	if (!counter->attr.inherit_stat)
+		return;
+
+	/*
+	 * Update the counter value, we cannot use perf_counter_read()
+	 * because we're in the middle of a context switch and have IRQs
+	 * disabled, which upsets smp_call_function_single(), however
+	 * we know the counter must be on the current CPU, therefore we
+	 * don't need to use it.
+	 */
+	switch (counter->state) {
+	case PERF_COUNTER_STATE_ACTIVE:
+		__perf_counter_read(counter);
+		break;
+
+	case PERF_COUNTER_STATE_INACTIVE:
+		update_counter_times(counter);
+		break;
+
+	default:
+		break;
+	}
+
+	/*
+	 * In order to keep per-task stats reliable we need to flip the counter

Review

Quoted 10 lines
(quoted diff) +10 -0
+	 * values when we flip the contexts.
+	 */
+	value = atomic64_read(&next_counter->count);
+	value = atomic64_xchg(&counter->count, value);
+	atomic64_set(&next_counter->count, value);
+
+	/*
+	 * XXX also sync time_enabled and time_running ?
+	 */
+}

Right, so I convinced myself we indeed want to swap the times as well, and realized we need to update the userpage after modifying these counters.

Then again, since inherited counters tend to wander around self-monitoring mmap() + inherit is bound to be broken.. hmm?

Do we want to fix that or shall we simply say: don't do that then!

Paul?

---
Subject: perf_counter: Complete counter swap.

Complete the counter swap by indeed switching the times too and updateing the userpage after modifying the counter values.

---
 kernel/perf_counter.c |    7 ++++++-
 1 files changed, 6 insertions(+), 1 deletions(-)
kernel/perf_counter.c +6 -1
index f2f2326..66ab1e9 100644
@@ -1048,9 +1048,14 @@ static void __perf_counter_sync_stat(struct perf_counter *counter,
 	value = atomic64_xchg(&counter->count, value);
 	atomic64_set(&next_counter->count, value);
 
+	swap(counter->total_time_enabled, next_counter->total_time_enabled);
+	swap(counter->total_time_running, next_counter->total_time_running);
+
 	/*
-	 * XXX also sync time_enabled and time_running ?
+	 * Since we swizzled the values, update the user visible data too.
 	 */
+	perf_counter_update_userpage(counter);
+	perf_counter_update_userpage(next_counter);
 }
 
 #define list_next_entry(pos, member) \
#2 lore raw

Re: [RFC][PATCH] perf_counter: Complete counter swap.

Quoted 8 lines · Peter Zijlstra

Right, so I convinced myself we indeed want to swap the times as well, and realized we need to update the userpage after modifying these counters.

Then again, since inherited counters tend to wander around self-monitoring mmap() + inherit is bound to be broken.. hmm?

Do we want to fix that or shall we simply say: don't do that then!

We only swap the contexts around if all the counters in both contexts have been inherited, i.e. neither context is a top-level parent context. Now I had been going to say that a counter that had been inherited couldn't be used for self-monitoring, and I think that is actually true, but I guess the problem is that the child could have inherited the fd and the mmap too, but the mmap will be on the parent counter not the inherited child counter. And there's no way for the child (or anyone) to mmap the inherited counter since there's no fd for it.

Currently we don't do anything to stop people trying to read the counters directly when they're not self-monitoring - they just get bogus data. Maybe we should put the tid of the task the counter's on into the first page of the mmap so that userspace can at least check if it's the task being monitored.

Unless you can see some way to change the mmap on fork to point to the child counter rather than the parent... Which would possibly be a bit nasty anyway since then the child's address space wouldn't be clone of the parent's like you would expect after a fork.

Paul.

#3 lore raw

[tip:perfcounters/urgent] perf_counter: Complete counter swap

In reply to Peter Zijlstra
Commit-ID:  19d2e755436054dfc2be640bffc32e427c37ac3d
Gitweb:     http://git.kernel.org/tip/19d2e755436054dfc2be640bffc32e427c37ac3d
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Fri, 26 Jun 2009 13:10:23 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Fri, 26 Jun 2009 17:48:54 +0200

perf_counter: Complete counter swap

Complete the counter swap by indeed switching the times too and updating the userpage after modifying the counter values.

LKML-Reference: <1246014623.31755.195.camel@twins>

---
 kernel/perf_counter.c |    7 ++++++-
 1 files changed, 6 insertions(+), 1 deletions(-)
kernel/perf_counter.c +6 -1
index f2f2326..66ab1e9 100644
@@ -1048,9 +1048,14 @@ static void __perf_counter_sync_stat(struct perf_counter *counter,
 	value = atomic64_xchg(&counter->count, value);
 	atomic64_set(&next_counter->count, value);
 
+	swap(counter->total_time_enabled, next_counter->total_time_enabled);
+	swap(counter->total_time_running, next_counter->total_time_running);
+
 	/*
-	 * XXX also sync time_enabled and time_running ?
+	 * Since we swizzled the values, update the user visible data too.
 	 */
+	perf_counter_update_userpage(counter);
+	perf_counter_update_userpage(next_counter);
 }
 
 #define list_next_entry(pos, member) \