[DRAFT PATCH 0/3] perf: Add Intel Nehalem uncore pmu support

[DRAFT PATCH 0/3] perf: Add Intel Nehalem uncore pmu support

[Lin Ming]

Hi, all

Here is the draft patch to add Intel Nehalem uncore pmu support.
It's not fully functional, but I threw it out early to get comments.

For the background of Nehalem uncore pmu, see Intel SDM Volume 3B
"30.6.2 Performance Monitoring Facility in the Uncore"

1. data structure

struct node_hw_events {
        struct perf_event *events[UNCORE_NUM_COUNTERS];
        int n_events;
        struct spinlock lock;
        int enabled;
};

struct node_hw_events is the per node structure.
"lock" protects add/delete events to uncore pmu.

struct uncore_cpu_hw_events {
        unsigned long active_mask[BITS_TO_LONGS(UNCORE_NUM_COUNTERS)];
};

struct uncore_cpu_hw_events is the per logical cpu structure.
"active_mask" represents the counters used by the cpu.
For example, if bit 3, 6 are set for cpuX, then it means uncore counter
3 and 6 are used by cpuX.

2. Uncore pmu NMI handling

Every core in the socket can be programmed to receive uncore counter
overflow interrupt.

In this draft implementation, each core handles the overflow interrupt
caused by the counters with bit set in "active_mask".

3. perf tool update

In this draft, the uncore events are monitored with raw events with "ru"
prefix("u" for uncore).

./perf stat -e ru0101 -- ls

Performance counter stats for 'ls':

             795920  raw 0x101               

        0.002110130  seconds time elapsed

4. Issues

How to eliminate the duplicate counter values accumulated by multi child
processes on the same socket?

perf stat -e ru0101 -- make -j4

Assume the 4 "make" child processes are running on the same socket and
counting uncore raw event "0101", and the counter value read by them are
val0, val1, val2, val3.

Then the final counter result given by "perf stat" will be "val0 + val1
+ val2 + val3".

But this is obvious wrong, because the uncore counter is shared by all
cores in the socket, so the final result should not be accumulated.

Any comment is very appreciated.

 arch/x86/include/asm/msr-index.h              |    1 +
 arch/x86/kernel/cpu/perf_event.c              |   30 ++-
 arch/x86/kernel/cpu/perf_event_intel.c        |    4 +-
 arch/x86/kernel/cpu/perf_event_intel_uncore.c |  280 +++++++++++++++++++++++++
 arch/x86/kernel/cpu/perf_event_intel_uncore.h |   80 +++++++
 arch/x86/kernel/cpu/perf_event_p4.c           |    2 +-
 include/linux/perf_event.h                    |    1 +
 tools/perf/util/parse-events.c                |   14 +-
 8 files changed, 394 insertions(+), 18 deletions(-)


Thanks,
Lin Ming

Re: [DRAFT PATCH 0/3] perf: Add Intel Nehalem uncore pmu support

[Peter Zijlstra]

On Tue, 2010-11-02 at 15:27 +0800, Lin Ming wrote:
> Any comment is very appreciated.

Right, so I was hoping to use the sysfs bits to expose things, I'll try
and get around to looking at your latest effort in that area soonish.
I'll try and sit down with gregkh one of these days to talk it over.

I'm not too sure about 1/3's change to x86_perf_event_update(), but its
not too aweful, the change to x86_perf_event_set_period() however does
look quite gruesome.

It might make sense to simple duplicate that code in the uncore bits,.
dunno.

2/3 looks ok, but I think it would be nice if it would be more self
contained, that is, not be part of the include mess and possibly have
its own NMI_DIE notifier entry.

All in all, Thanks for doing this, its a good start!

Re: [DRAFT PATCH 0/3] perf: Add Intel Nehalem uncore pmu support

[Lin Ming]

On Tue, 2010-11-02 at 20:29 +0800, Peter Zijlstra wrote:
> On Tue, 2010-11-02 at 15:27 +0800, Lin Ming wrote:
> > Any comment is very appreciated.
> 
> Right, so I was hoping to use the sysfs bits to expose things, I'll try
> and get around to looking at your latest effort in that area soonish.
> I'll try and sit down with gregkh one of these days to talk it over.
> 
> I'm not too sure about 1/3's change to x86_perf_event_update(), but its
> not too aweful, the change to x86_perf_event_set_period() however does
> look quite gruesome.
> 
> It might make sense to simple duplicate that code in the uncore bits,.
> dunno.

How about below?

diff --git a/arch/x86/kernel/cpu/perf_event_intel_uncore.c b/arch/x86/kernel/cpu/perf_event_intel_uncore.c
index fafa0f9..b22aa95 100644
--- a/arch/x86/kernel/cpu/perf_event_intel_uncore.c
+++ b/arch/x86/kernel/cpu/perf_event_intel_uncore.c
@@ -80,10 +80,52 @@ static int uncore_pmu_event_init(struct perf_event *event)
 	return 0;
 }
 
+static int
+uncore_perf_event_set_period(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	s64 left = local64_read(&hwc->period_left);
+	s64 period = hwc->sample_period;
+	u64 max_period = (1ULL << UNCORE_NUM_COUNTERS) - 1;
+	int ret = 0, idx = hwc->idx;
+
+	/*
+	 * If we are way outside a reasonable range then just skip forward:
+	 */
+	if (unlikely(left <= -period)) {
+		left = period;
+		local64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+
+	if (unlikely(left <= 0)) {
+		left += period;
+		local64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+
+	if (left > max_period)
+		left = max_period;
+
+	/*
+	 * The hw event starts counting from this event offset,
+	 * mark it to be able to extra future deltas:
+	 */
+	local64_set(&hwc->prev_count, (u64)-left);
+
+	wrmsrl(hwc->event_base + idx, (u64)(-left) & max_period);
+
+	perf_event_update_userpage(event);
+
+	return ret;
+}
+
 static void uncore_pmu_start(struct perf_event *event, int flags)
 {
 	if (flags & PERF_EF_RELOAD)
-		x86_perf_event_set_period(event);
+		uncore_perf_event_set_period(event);
 
 	uncore_pmu_enable_event(event);
 
@@ -200,7 +242,7 @@ static inline void uncore_pmu_ack_status(u64 ack)
 static int uncore_pmu_save_and_restart(struct perf_event *event)
 {
 	x86_perf_event_update(event, UNCORE_CNTVAL_BITS);
-	return x86_perf_event_set_period(event);
+	return uncore_perf_event_set_period(event);
 }
 
 int uncore_pmu_handle_irq(struct pt_regs *regs)


> 
> 2/3 looks ok, but I think it would be nice if it would be more self
> contained, that is, not be part of the include mess and possibly have
> its own NMI_DIE notifier entry.

How about below to make uncore code more self contained?
I'll look at the NMI DIE notifier thing later.

diff --git a/arch/x86/include/asm/perf_event.h b/arch/x86/include/asm/perf_event.h
index 550e26b..8df4e13 100644
--- a/arch/x86/include/asm/perf_event.h
+++ b/arch/x86/include/asm/perf_event.h
@@ -127,6 +127,7 @@ union cpuid10_edx {
 #ifdef CONFIG_PERF_EVENTS
 extern void init_hw_perf_events(void);
 extern void perf_events_lapic_init(void);
+extern void init_uncore_pmu(void);
 
 #define PERF_EVENT_INDEX_OFFSET			0
 
@@ -138,6 +139,7 @@ extern void perf_events_lapic_init(void);
 #define PERF_EFLAGS_EXACT	(1UL << 3)
 
 struct pt_regs;
+extern int uncore_pmu_handle_irq(struct pt_regs *regs);
 extern unsigned long perf_instruction_pointer(struct pt_regs *regs);
 extern unsigned long perf_misc_flags(struct pt_regs *regs);
 #define perf_misc_flags(regs)	perf_misc_flags(regs)
diff --git a/arch/x86/kernel/cpu/Makefile b/arch/x86/kernel/cpu/Makefile
index 3f0ebe4..db4bf99 100644
--- a/arch/x86/kernel/cpu/Makefile
+++ b/arch/x86/kernel/cpu/Makefile
@@ -27,6 +27,7 @@ obj-$(CONFIG_CPU_SUP_TRANSMETA_32)	+= transmeta.o
 obj-$(CONFIG_CPU_SUP_UMC_32)		+= umc.o
 
 obj-$(CONFIG_PERF_EVENTS)		+= perf_event.o
+obj-$(CONFIG_PERF_EVENTS)		+= perf_event_intel_uncore.o
 
 obj-$(CONFIG_X86_MCE)			+= mcheck/
 obj-$(CONFIG_MTRR)			+= mtrr/
diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
index cca07b4..330e4f4 100644
--- a/arch/x86/kernel/cpu/perf_event.c
+++ b/arch/x86/kernel/cpu/perf_event.c
@@ -1215,8 +1215,6 @@ struct pmu_nmi_state {
 
 static DEFINE_PER_CPU(struct pmu_nmi_state, pmu_nmi);
 
-static int uncore_pmu_handle_irq(struct pt_regs *regs);
-
 static int __kprobes
 perf_event_nmi_handler(struct notifier_block *self,
 			 unsigned long cmd, void *__args)
@@ -1308,7 +1306,6 @@ x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
 #include "perf_event_intel_lbr.c"
 #include "perf_event_intel_ds.c"
 #include "perf_event_intel.c"
-#include "perf_event_intel_uncore.c"
 
 static int __cpuinit
 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
diff --git a/arch/x86/kernel/cpu/perf_event_intel_uncore.c b/arch/x86/kernel/cpu/perf_event_intel_uncore.c
index b22aa95..b9f15f2 100644
--- a/arch/x86/kernel/cpu/perf_event_intel_uncore.c
+++ b/arch/x86/kernel/cpu/perf_event_intel_uncore.c
@@ -3,6 +3,7 @@
 static struct node_hw_events uncore_events[MAX_NUMNODES];
 static DEFINE_PER_CPU(struct uncore_cpu_hw_events, uncore_cpu_hw_events);
 static bool uncore_pmu_initialized;
+static atomic_t active_events;
 
 static void uncore_pmu_enable_event(struct perf_event *event)
 {
diff --git a/arch/x86/kernel/cpu/perf_event_intel_uncore.h b/arch/x86/kernel/cpu/perf_event_intel_uncore.h
index 33b9b5e..0a5e6d4 100644
--- a/arch/x86/kernel/cpu/perf_event_intel_uncore.h
+++ b/arch/x86/kernel/cpu/perf_event_intel_uncore.h
@@ -78,3 +78,4 @@ struct uncore_cpu_hw_events {
 	unsigned long active_mask[BITS_TO_LONGS(UNCORE_NUM_COUNTERS)];
 };
 
+extern u64 x86_perf_event_update(struct perf_event *event, int cntval_bits);


> 
> All in all, Thanks for doing this, its a good start!

Thanks for the quick response.

Lin Ming

Re: [DRAFT PATCH 0/3] perf: Add Intel Nehalem uncore pmu support

[Cyrill Gorcunov]

On Tue, Nov 02, 2010 at 03:27:38PM +0800, Lin Ming wrote:
> Hi, all
> 
> Here is the draft patch to add Intel Nehalem uncore pmu support.
> It's not fully functional, but I threw it out early to get comments.
> 

 Hi Ming, sorry for late response, though there are already a lot
of comments anyway :) Great job!

 Cyrill