Skip to contentSkip to conversation

perf offcore patchkit for merge

PATCH In review Andi Kleen 7 participants 19 messages Nov 18, 2010 – Dec 1, 2010

Subsystems (2)
lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

This is the latest version of the perf offcore / extra regs patchkit. I addressed all earlier comments I believe and this should be ready for merge now.

- Allocating multiple registers is supported now
- I kept the percore allocation because the manual kmalloc method turned out to be more complicated and alloc_percpu() really already did what I wanted. However I switched to using the CPU notifier to set up the topology for hotplug, together with a bugfix to make sure x86 sets up the topology before calling it.
- A few other improvements.

-Andi

author
#1 lore raw

PATCH 1/4 x86: set cpu masks before calling CPU_STARTING notifiers

From: Andi Kleen <[email protected]>

When booting up a CPU set the various topology masks before calling the CPU_STARTING notifier. This way the notifier can actually use the masks.

This is needed for a perf change.

---
 arch/x86/kernel/smpboot.c |    8 ++++----
 1 files changed, 4 insertions(+), 4 deletions(-)
arch/x86/kernel/smpboot.c +4 -4
index 083e99d..9d2980e 100644
@@ -281,6 +281,10 @@ static void __cpuinit smp_callin(void)
 	 */
 	smp_store_cpu_info(cpuid);
 
+	/* This must be done before setting cpu_online_mask */
+	set_cpu_sibling_map(raw_smp_processor_id());
+	wmb();
+
 	notify_cpu_starting(cpuid);
 
 	/*
@@ -322,10 +326,6 @@ notrace static void __cpuinit start_secondary(void *unused)
 		legacy_pic->unmask(0);
 	}
 
-	/* This must be done before setting cpu_online_mask */
-	set_cpu_sibling_map(raw_smp_processor_id());
-	wmb();
-
 	/*
 	 * We need to hold call_lock, so there is no inconsistency
 	 * between the time smp_call_function() determines number of
Signature · 1 line
1.7.1
#2 lore raw

Re: [PATCH 1/4] x86: set cpu masks before calling CPU_STARTING notifiers

Quoted 18 lines · Andi Kleen

From: Andi Kleen <[email protected]>

When booting up a CPU set the various topology masks before calling the CPU_STARTING notifier. This way the notifier can actually use the masks.

This is needed for a perf change.

---
 arch/x86/kernel/smpboot.c |    8 ++++----
 1 files changed, 4 insertions(+), 4 deletions(-)
arch/x86/kernel/smpboot.c +0 -0
index 083e99d..9d2980e 100644

arch/x86/kernel/smpboot.c static void __cpuinit smp_callin(void) · line 284

Quoted 4 lines
arch/x86/kernel/smpboot.c +1 -0
@@ -281,6 +281,10 @@ static void __cpuinit smp_callin(void)
 	 */
 	smp_store_cpu_info(cpuid);
 
+	/* This must be done before setting cpu_online_mask */

Can we please fix up that comment to match that change ?

arch/x86/kernel/smpboot.c static void __cpuinit smp_callin(void) · line 287

Quoted 3 lines
arch/x86/kernel/smpboot.c +3 -0
+	set_cpu_sibling_map(raw_smp_processor_id());
+	wmb();
+

Thanks,

tglx

author
#3 lore raw

Re: [PATCH 1/4] x86: set cpu masks before calling CPU_STARTING notifiers

Quoted 23 lines · Thomas Gleixner
Quoted 21 lines · Andi Kleen

From: Andi Kleen <[email protected]>

When booting up a CPU set the various topology masks before calling the CPU_STARTING notifier. This way the notifier can actually use the masks.

This is needed for a perf change.

---
 arch/x86/kernel/smpboot.c |    8 ++++----
 1 files changed, 4 insertions(+), 4 deletions(-)
arch/x86/kernel/smpboot.c +0 -0
index 083e99d..9d2980e 100644
@@ -281,6 +281,10 @@ static void __cpuinit smp_callin(void)
 	 */
 	smp_store_cpu_info(cpuid);
 
Quoted 3 lines

arch/x86/kernel/smpboot.c static void __cpuinit smp_callin(void) · line 284

Quoted 1 line
arch/x86/kernel/smpboot.c +1 -0
+	/* This must be done before setting cpu_online_mask */

Can we please fix up that comment to match that change ?

Here's an updated patch.

x86: set cpu masks before calling CPU_STARTING notifiers

When booting up a CPU set the various topology masks before calling the CPU_STARTING notifier. This way the notifier can actually use the masks.

This is needed for a perf change.

arch/x86/kernel/smpboot.c +7 -4
index 083e99d..c019c03 100644
@@ -281,6 +281,13 @@ static void __cpuinit smp_callin(void)
 	 */
 	smp_store_cpu_info(cpuid);
 
+	/* 
+	 * This must be done before setting cpu_online_mask
+	 * or calling notify_cpu_starting.
+	 */
+	set_cpu_sibling_map(raw_smp_processor_id());
+	wmb();
+
 	notify_cpu_starting(cpuid);
 
 	/*
@@ -322,10 +329,6 @@ notrace static void __cpuinit start_secondary(void *unused)
 		legacy_pic->unmask(0);
 	}
 
-	/* This must be done before setting cpu_online_mask */
-	set_cpu_sibling_map(raw_smp_processor_id());
-	wmb();
-
 	/*
 	 * We need to hold call_lock, so there is no inconsistency
 	 * between the time smp_call_function() determines number of
#4 lore raw

[tip:perf/core] x86: Set cpu masks before calling CPU_STARTING notifiers

In reply to Andi Kleen
Commit-ID:  5ef428c4b5950dddce7311e84321abb3aff7ebb0
Gitweb:     http://git.kernel.org/tip/5ef428c4b5950dddce7311e84321abb3aff7ebb0
Author:     Andi Kleen <[email protected]>
AuthorDate: Thu, 18 Nov 2010 11:47:31 +0100
Committer:  Ingo Molnar <[email protected]>
CommitDate: Fri, 26 Nov 2010 15:14:56 +0100

x86: Set cpu masks before calling CPU_STARTING notifiers

When booting up a CPU set the various topology masks before calling the CPU_STARTING notifier. This way the notifier can actually use the masks.

This is needed for a perf change.

LKML-Reference: <[email protected]>

---
 arch/x86/kernel/smpboot.c |   11 +++++++----
 1 files changed, 7 insertions(+), 4 deletions(-)
arch/x86/kernel/smpboot.c +7 -4
index f0a0624..68f61ac 100644
@@ -281,6 +281,13 @@ static void __cpuinit smp_callin(void)
 	 */
 	smp_store_cpu_info(cpuid);
 
+	/*
+	 * This must be done before setting cpu_online_mask
+	 * or calling notify_cpu_starting.
+	 */
+	set_cpu_sibling_map(raw_smp_processor_id());
+	wmb();
+
 	notify_cpu_starting(cpuid);
 
 	/*
@@ -316,10 +323,6 @@ notrace static void __cpuinit start_secondary(void *unused)
 	 */
 	check_tsc_sync_target();
 
-	/* This must be done before setting cpu_online_mask */
-	set_cpu_sibling_map(raw_smp_processor_id());
-	wmb();
-
 	/*
 	 * We need to hold call_lock, so there is no inconsistency
 	 * between the time smp_call_function() determines number of
author
#5 lore raw

PATCH 2/4 perf: Document enhanced event encoding for OFFCORE_MSR

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

---
 tools/perf/Documentation/perf-list.txt |    7 +++++++
 1 files changed, 7 insertions(+), 0 deletions(-)
tools/perf/Documentation/perf-list.txt +7 -0
index 399751b..700afd7 100644
@@ -58,6 +58,13 @@ raw encoding of 0x1A8 can be used:
  perf stat -e r1a8 -a sleep 1
  perf record -e r1a8 ...
 
+Some special events on x86 encode additional data in the normally unused 
+[32;63] bits of the raw value. This is particularly used for 
+the OFFCORE_RESPONSE events on Intel Core i7.  The 16bit 
+mask in the OFFCORE_RESPONSE register is put into bits [32;48].
+For example the OFFCORE_RESPONSE_0.ANY_DATA.ANY_CACHE_DRAM event
+is encoded as r00007f11004301b7.
+
 You should refer to the processor specific documentation for getting these
 details. Some of them are referenced in the SEE ALSO section below.
 
Signature · 1 line
1.7.1
author
#6 lore raw

PATCH 3/4 perf-events: Add support for supplementary event registers v3

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

Intel Nehalem/Westmere have a special OFFCORE_RESPONSE event that can be used to monitor any offcore accesses from a core. This is a very useful event for various tunings, and it's also needed to implement the generic LLC-* events correctly.

Unfortunately this event requires programming a mask in a separate register. And worse this separate register is per core, not per CPU thread.

This patch adds:
- Teaches perf_events that OFFCORE_RESPONSE needs extra parameters. The extra parameters are passed by user space in the unused upper 32bits of the config word.
- Add support to the Intel perf_event core to schedule per core resources. This adds fairly generic infrastructure that can be also used for other per core resources.
The basic code has is patterned after the similar AMD northbridge constraints code.

Thanks to Stephane Eranian who pointed out some problems in the original version and suggested improvements.

Full git tree:
git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux-misc-2.6.git perf-offcore2
Cc: [email protected]
v2: Lots of updates based on review feedback. Also fixes some issues
v3: Fix hotplug. Handle multiple extra registers. Fix init order. Various improvements.

---
 arch/x86/kernel/cpu/perf_event.c       |   70 ++++++++++++
 arch/x86/kernel/cpu/perf_event_intel.c |  192 ++++++++++++++++++++++++++++++++
 include/linux/perf_event.h             |    2 +
 3 files changed, 264 insertions(+), 0 deletions(-)
arch/x86/kernel/cpu/perf_event.c +70 -0
index ed63101..346006a 100644
@@ -93,6 +93,8 @@ struct amd_nb {
 	struct event_constraint event_constraints[X86_PMC_IDX_MAX];
 };
 
+struct intel_percore;
+
 #define MAX_LBR_ENTRIES		16
 
 struct cpu_hw_events {
@@ -127,6 +129,13 @@ struct cpu_hw_events {
 	struct perf_branch_stack	lbr_stack;
 	struct perf_branch_entry	lbr_entries[MAX_LBR_ENTRIES];
 
+	/* 
+	 * Intel percore register state.
+	 * Coordinate shared resources between HT threads.
+	 */
+	int 				percore_used; /* Used by this CPU? */
+	struct intel_percore		*per_core;
+
 	/*
 	 * AMD specific bits
 	 */
@@ -175,6 +184,32 @@ struct cpu_hw_events {
 #define for_each_event_constraint(e, c)	\
 	for ((e) = (c); (e)->weight; (e)++)
 
+/*
+ * Extra registers for specific events.
+ * Some events need large masks and require external MSRs.
+ * Define a mapping to these extra registers.
+ * The actual contents are still encoded in unused parts of the
+ * original config u64.
+ */
+struct extra_reg {
+	unsigned int		event;
+	unsigned int		msr;
+	unsigned int		extra_shift;
+	u64 			config_mask;
+	u64 			valid_mask;
+};
+
+#define EVENT_EXTRA_REG(e, ms, m, vm, es) {	\
+	.event = (e),    	\
+	.msr = (ms),	     	\
+	.config_mask = (m),  	\
+	.valid_mask = (vm),	\
+	.extra_shift = (es),	\
+	}
+#define INTEL_EVENT_EXTRA_REG(event, msr, vm, es)	\
+	EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm, es)
+#define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0, 0)
+
 union perf_capabilities {
 	struct {
 		u64	lbr_format    : 6;
@@ -219,6 +254,7 @@ struct x86_pmu {
 	void		(*put_event_constraints)(struct cpu_hw_events *cpuc,
 						 struct perf_event *event);
 	struct event_constraint *event_constraints;
+	struct event_constraint *percore_constraints;
 	void		(*quirks)(void);
 	int		perfctr_second_write;
 
@@ -247,6 +283,11 @@ struct x86_pmu {
 	 */
 	unsigned long	lbr_tos, lbr_from, lbr_to; /* MSR base regs       */
 	int		lbr_nr;			   /* hardware stack size */
+
+	/*
+	 * Extra registers for events
+	 */
+	struct extra_reg *extra_regs;
 };
 
 static struct x86_pmu x86_pmu __read_mostly;
@@ -321,6 +362,33 @@ again:
 	return new_raw_count;
 }
 
+/*
+ * Find and validate any extra registers to set up.
+ */
+static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
+{
+	struct extra_reg *er;
+	u64 extra;
+
+	event->hw.extra_reg = 0;
+	event->hw.extra_config = 0;
+
+	if (!x86_pmu.extra_regs)
+		return 0;
+
+	for (er = x86_pmu.extra_regs; er->msr; er++) {
+		if (er->event != (config & er->config_mask))
+			continue;
+		event->hw.extra_reg = er->msr;
+		extra = config >> er->extra_shift;
+		if (extra & ~er->valid_mask)
+			return -EINVAL;
+		event->hw.extra_config = extra;
+		break;
+	}
+	return 0;
+}
+
 static atomic_t active_events;
 static DEFINE_MUTEX(pmc_reserve_mutex);
 
@@ -876,6 +944,8 @@ static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
 					  u64 enable_mask)
 {
 	wrmsrl(hwc->config_base + hwc->idx, hwc->config | enable_mask);
+	if (hwc->extra_reg)
+		wrmsrl(hwc->extra_reg, hwc->extra_config);
 }
 
 static inline void x86_pmu_disable_event(struct perf_event *event)
arch/x86/kernel/cpu/perf_event_intel.c +192 -0

261 lines not on this page · raw message

include/linux/perf_event.h +2 -0
index 057bf22..6124e93 100644
@@ -529,6 +529,8 @@ struct hw_perf_event {
 			unsigned long	event_base;
 			int		idx;
 			int		last_cpu;
+			unsigned int	extra_reg;
+			u64		extra_config;
 		};
 		struct { /* software */
 			struct hrtimer	hrtimer;
Signature · 1 line
1.7.1
#7 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

Quoted 24 lines · Andi Kleen

From: Andi Kleen <[email protected]>

Intel Nehalem/Westmere have a special OFFCORE_RESPONSE event that can be used to monitor any offcore accesses from a core. This is a very useful event for various tunings, and it's also needed to implement the generic LLC-* events correctly.

Unfortunately this event requires programming a mask in a separate register. And worse this separate register is per core, not per CPU thread.

This patch adds:
- Teaches perf_events that OFFCORE_RESPONSE needs extra parameters. The extra parameters are passed by user space in the unused upper 32bits of the config word.
- Add support to the Intel perf_event core to schedule per core resources. This adds fairly generic infrastructure that can be also used for other per core resources.
The basic code has is patterned after the similar AMD northbridge constraints code.

Thanks to Stephane Eranian who pointed out some problems in the original version and suggested improvements.

Quoted 6 lines

Full git tree:
git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux-misc-2.6.git perf-offcore2
Cc: [email protected]
v2: Lots of updates based on review feedback. Also fixes some issues
v3: Fix hotplug. Handle multiple extra registers. Fix init order. Various improvements.

Stuff like that shouldn't be mixed in with the tags and usually goes below the --- separator since its not supposed to end up in the committed changelog.

Quoted 8 lines
---
 arch/x86/kernel/cpu/perf_event.c       |   70 ++++++++++++
 arch/x86/kernel/cpu/perf_event_intel.c |  192 ++++++++++++++++++++++++++++++++
 include/linux/perf_event.h             |    2 +
 3 files changed, 264 insertions(+), 0 deletions(-)

arch/x86/kernel/cpu/perf_event.c static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc, · line 949

Quoted 6 lines
arch/x86/kernel/cpu/perf_event.c +2 -0
@@ -876,6 +944,8 @@ static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
 					  u64 enable_mask)
 {
 	wrmsrl(hwc->config_base + hwc->idx, hwc->config | enable_mask);
+	if (hwc->extra_reg)
+		wrmsrl(hwc->extra_reg, hwc->extra_config);
 }

Just wondering, shouldn't we program the extra msr _before_ we flip the enable bit?

Quoted 9 lines
arch/x86/kernel/cpu/perf_event_intel.c +9 -0
+static __init int init_intel_percore(void)
+{
+	int cpu;
+
+	if (!needs_percore)
+		return 0;
+
+	intel_percore = alloc_percpu(struct intel_percore);
+	if (!intel_percore)

arch/x86/kernel/cpu/perf_event_intel.c static void intel_clovertown_quirks(void) · line 1108

Quoted 10 lines
arch/x86/kernel/cpu/perf_event_intel.c +10 -0
+		return -ENOMEM;
+	for_each_possible_cpu(cpu)
+		raw_spin_lock_init(&per_cpu_ptr(intel_percore, cpu)->lock);
+
+	return 0;
+}
+/* 
+ * Runs later because per cpu allocations don't work early on.
+ */
+__initcall(init_intel_percore);

I've got a patch moving the whole pmu init to early_initcall(), which is after mm_init() so it would actually work.

author
#8 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

Quoted 9 lines · Peter Zijlstra

arch/x86/kernel/cpu/perf_event.c static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc, · line 949

Quoted 6 lines
arch/x86/kernel/cpu/perf_event.c +2 -0
 					  u64 enable_mask)
 {
 	wrmsrl(hwc->config_base + hwc->idx, hwc->config | enable_mask);
+	if (hwc->extra_reg)
+		wrmsrl(hwc->extra_reg, hwc->extra_config);
 }

Just wondering, shouldn't we program the extra msr _before_ we flip the enable bit?

Yes that makes sense.

Quoted 6 lines
Quoted 3 lines

+ * Runs later because per cpu allocations don't work early on.
+ */
+__initcall(init_intel_percore);

I've got a patch moving the whole pmu init to early_initcall(), which is after mm_init() so it would actually work.

So do you want to make this patchkit depend on that patch? Or just integrate it and then change later?

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only.
#9 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

Quoted 10 lines · Andi Kleen
Quoted 7 lines
Quoted 4 lines

+ * Runs later because per cpu allocations don't work early on.
+ */
+__initcall(init_intel_percore);

I've got a patch moving the whole pmu init to early_initcall(), which is after mm_init() so it would actually work.

So do you want to make this patchkit depend on that patch? Or just integrate it and then change later?

Doesn't really matter, I can fix it up afterwards, just wanted to let you know.. I can also flip that enable thing.

I'll take these 4 patches and fixup these things and then see what falls out ;-)

maintainer
#10 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

Quoted 9 lines · Peter Zijlstra
Quoted 8 lines · Andi Kleen
Quoted 7 lines
Quoted 4 lines

+ * Runs later because per cpu allocations don't work early on.
+ */
+__initcall(init_intel_percore);

I've got a patch moving the whole pmu init to early_initcall(), which is after mm_init() so it would actually work.

Quoted 8 lines
Quoted 2 lines

So do you want to make this patchkit depend on that patch? Or just integrate it and then change later?

Doesn't really matter, I can fix it up afterwards, just wanted to let you know.. I can also flip that enable thing.

I'll take these 4 patches and fixup these things and then see what falls out ;-)

OK, that initcall patch hit -tip and Ingo wants that allocation stuff fixed up before committing those patches.

Could you run another version of these patches (you can ignore the checkpatch output that results from the non-std struct initialization layout)?

Just make the allocation path like the AMD NB stuff, use the cpu_prepare and cpu_dead hooks to alloc/free and add a refcount in the percpu object.

maintainer
#11 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

Quoted 22 lines · Peter Zijlstra
Quoted 9 lines · Peter Zijlstra
Quoted 8 lines · Andi Kleen
Quoted 7 lines
Quoted 4 lines

+ * Runs later because per cpu allocations don't work early on.
+ */
+__initcall(init_intel_percore);

I've got a patch moving the whole pmu init to early_initcall(), which is after mm_init() so it would actually work.

Quoted 8 lines
Quoted 2 lines

So do you want to make this patchkit depend on that patch? Or just integrate it and then change later?

Doesn't really matter, I can fix it up afterwards, just wanted to let you know.. I can also flip that enable thing.

I'll take these 4 patches and fixup these things and then see what falls out ;-)

OK, that initcall patch hit -tip and Ingo wants that allocation stuff fixed up before committing those patches.

Quoted 7 lines

Could you run another version of these patches (you can ignore the checkpatch output that results from the non-std struct initialization layout)?

Just make the allocation path like the AMD NB stuff, use the cpu_prepare and cpu_dead hooks to alloc/free and add a refcount in the percpu object.

While looking at the AMD code, I think we can do away with amd_nb_lock, all the hotplug stuff is serialized on the hotplug mutex anyway.

#12 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

In reply to Andi Kleen
Quoted 21 lines · Andi Kleen

From: Andi Kleen <[email protected]>

Intel Nehalem/Westmere have a special OFFCORE_RESPONSE event that can be used to monitor any offcore accesses from a core. This is a very useful event for various tunings, and it's also needed to implement the generic LLC-* events correctly.

Unfortunately this event requires programming a mask in a separate register. And worse this separate register is per core, not per CPU thread.

This patch adds:
- Teaches perf_events that OFFCORE_RESPONSE needs extra parameters. The extra parameters are passed by user space in the unused upper 32bits of the config word.
- Add support to the Intel perf_event core to schedule per core resources. This adds fairly generic infrastructure that can be also used for other per core resources.
The basic code has is patterned after the similar AMD northbridge constraints code.

Quoted 2 lines

Thanks to Stephane Eranian who pointed out some problems in the original version and suggested improvements.

WARNING: please, no space before tabs
#59: FILE: arch/x86/kernel/cpu/perf_event.c:136:
+^Iint ^I^I^I^Ipercore_used; /* Used by this CPU? */$

WARNING: please, no space before tabs
#80: FILE: arch/x86/kernel/cpu/perf_event.c:198:
+^Iu64 ^I^I^Iconfig_mask;$

WARNING: please, no space before tabs
#81: FILE: arch/x86/kernel/cpu/perf_event.c:199:
+^Iu64 ^I^I^Ivalid_mask;$

WARNING: please, no space before tabs
#85: FILE: arch/x86/kernel/cpu/perf_event.c:203:
+^I.event = (e), ^I\$

WARNING: please, no space before tabs
#86: FILE: arch/x86/kernel/cpu/perf_event.c:204:
+^I.msr = (ms),^I ^I\$

WARNING: please, no space before tabs
#87: FILE: arch/x86/kernel/cpu/perf_event.c:205:
+^I.config_mask = (m), ^I\$

WARNING: please, no space before tabs
#177: FILE: arch/x86/kernel/cpu/perf_event_intel.c:12:
+^Iint ^I^Iref;^I^I/* reference count */$

WARNING: please, no space before tabs
#179: FILE: arch/x86/kernel/cpu/perf_event_intel.c:14:
+^Iu64 ^I^Iextra_config;^I/* extra MSR config */$

WARNING: please, no space before tabs
#183: FILE: arch/x86/kernel/cpu/perf_event_intel.c:18:
+^Iraw_spinlock_t ^I^Ilock;^I^I/* protect structure */$

WARNING: please, no space before tabs
#184: FILE: arch/x86/kernel/cpu/perf_event_intel.c:19:
+^Istruct er_account ^Iregs[MAX_EXTRA_REGS];$

WARNING: please use device_initcall() instead of __initcall() #408: FILE: arch/x86/kernel/cpu/perf_event_intel.c:1108:
+__initcall(init_intel_percore);

Fixed those up, please be more careful next time.

#13 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

In reply to Andi Kleen
Quoted 10 lines · Andi Kleen

From: Andi Kleen <[email protected]>

Intel Nehalem/Westmere have a special OFFCORE_RESPONSE event that can be used to monitor any offcore accesses from a core. This is a very useful event for various tunings, and it's also needed to implement the generic LLC-* events correctly.

Unfortunately this event requires programming a mask in a separate register. And worse this separate register is per core, not per CPU thread.

This "separate register" is MSR_OFFCORE_RSP_0, right?
But from the SDM, MSR_OFFCORE_RSP_0 is "thread" scope,
see SDM 3b, Appendix B.4 MSRS IN THE INTEL® MICROARCHITECTURE CODENAME NEHALEM

Or am I missing some obvious thing?

Thanks,
Lin Ming

#14 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

Quoted 12 lines · Lin Ming
Quoted 10 lines · Andi Kleen

From: Andi Kleen <[email protected]>

Intel Nehalem/Westmere have a special OFFCORE_RESPONSE event that can be used to monitor any offcore accesses from a core. This is a very useful event for various tunings, and it's also needed to implement the generic LLC-* events correctly.

Unfortunately this event requires programming a mask in a separate register. And worse this separate register is per core, not per CPU thread.

Quoted 6 lines

This "separate register" is MSR_OFFCORE_RSP_0, right?
But from the SDM,  MSR_OFFCORE_RSP_0 is "thread" scope,
see SDM 3b, Appendix B.4 MSRS IN THE INTEL® MICROARCHITECTURE CODENAME NEHALEM

Or am I missing some obvious thing?

The manual is wrong on this.

Quoted 3 lines

Thanks,
Lin Ming

#15 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

Quoted 13 lines · Stephane Eranian
Quoted 12 lines · Lin Ming
Quoted 10 lines · Andi Kleen

From: Andi Kleen <[email protected]>

Intel Nehalem/Westmere have a special OFFCORE_RESPONSE event that can be used to monitor any offcore accesses from a core. This is a very useful event for various tunings, and it's also needed to implement the generic LLC-* events correctly.

Unfortunately this event requires programming a mask in a separate register. And worse this separate register is per core, not per CPU thread.

Quoted 7 lines
Quoted 6 lines

This "separate register" is MSR_OFFCORE_RSP_0, right?
But from the SDM,  MSR_OFFCORE_RSP_0 is "thread" scope,
see SDM 3b, Appendix B.4 MSRS IN THE INTEL® MICROARCHITECTURE CODENAME NEHALEM

Or am I missing some obvious thing?

The manual is wrong on this.

Well, thanks for the info.

#16 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

In reply to Andi Kleen

arch/x86/kernel/cpu/perf_event.c static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc, · line 949

Quoted 7 lines · Andi Kleen
arch/x86/kernel/cpu/perf_event.c +2 -0
@@ -876,6 +944,8 @@ static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
                                          u64 enable_mask)
 {
        wrmsrl(hwc->config_base + hwc->idx, hwc->config | enable_mask);
+       if (hwc->extra_reg)
+               wrmsrl(hwc->extra_reg, hwc->extra_config);
 } 

I thought we agreed it made more sense to program the extra msr before enabling the counter.

maintainer
#17 lore raw

Re: [PATCH 3/4] perf-events: Add support for supplementary event registers v3

Quoted 11 lines · Peter Zijlstra

arch/x86/kernel/cpu/perf_event.c static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc, · line 949

Quoted 7 lines · Andi Kleen
arch/x86/kernel/cpu/perf_event.c +2 -0
@@ -876,6 +944,8 @@ static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
                                          u64 enable_mask)
 {
        wrmsrl(hwc->config_base + hwc->idx, hwc->config | enable_mask);
+       if (hwc->extra_reg)
+               wrmsrl(hwc->extra_reg, hwc->extra_config);
 } 

I thought we agreed it made more sense to program the extra msr before enabling the counter.

Ah sorry, I seem to have mixed up the various versions.

The latest does indeed do as we agreed. Sorry for the noise.

author
#18 lore raw

PATCH 4/4 perf-events: Fix LLC-* events on Intel Nehalem/Westmere v2

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

The generic perf LLC-* events do count the L2 caches, not the real L3 LLC on Intel Nehalem and Westmere. This lead to quite some confusion.

Fixing this properly requires use of the special OFFCORE_RESPONSE events which need a separate mask register. This has been implemented in a earlier patch.

Now use this infrastructure to set correct events for the LLC-* on Nehalem and Westmere

---
 arch/x86/kernel/cpu/perf_event.c       |   12 ++++----
 arch/x86/kernel/cpu/perf_event_intel.c |   46 +++++++++++++++++++++++---------
 2 files changed, 39 insertions(+), 19 deletions(-)
arch/x86/kernel/cpu/perf_event.c +6 -6
index 346006a..b1abe89 100644
@@ -467,8 +467,9 @@ static inline int x86_pmu_initialized(void)
 }
 
 static inline int
-set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event_attr *attr)
+set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
 {
+	struct perf_event_attr *attr = &event->attr;
 	unsigned int cache_type, cache_op, cache_result;
 	u64 config, val;
 
@@ -494,9 +495,8 @@ set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event_attr *attr)
 	if (val == -1)
 		return -EINVAL;
 
-	hwc->config |= val;
-
-	return 0;
+	hwc->config |= val & X86_RAW_EVENT_MASK;
+	return x86_pmu_extra_regs(val, event);
 }
 
 static int x86_setup_perfctr(struct perf_event *event)
@@ -521,10 +521,10 @@ static int x86_setup_perfctr(struct perf_event *event)
 	}
 
 	if (attr->type == PERF_TYPE_RAW)
-		return 0;
+		return x86_pmu_extra_regs(event->attr.config, event);
 
 	if (attr->type == PERF_TYPE_HW_CACHE)
-		return set_ext_hw_attr(hwc, attr);
+		return set_ext_hw_attr(hwc, event);
 
 	if (attr->config >= x86_pmu.max_events)
 		return -EINVAL;
arch/x86/kernel/cpu/perf_event_intel.c +33 -13

72 lines not on this page · raw message

Signature · 1 line
1.7.1