Skip to contentSkip to conversation

Updated Broadwell perf patchkit

PATCH In review Andi Kleen 4 participants 19 messages Aug 14, 2014 – Aug 15, 2014

Subsystems (2)
lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

Addressed the earlier review feedback. Also changed Haswell to use the Broadwell cache event list, which is far more correct for it than Sandy Bridge's.

-Andi

author
#1 lore raw

PATCH 1/5 perf, x86: Remove incorrect model number from Haswell perf

From: Andi Kleen <[email protected]>

---
 arch/x86/kernel/cpu/perf_event_intel.c | 1 -
 1 file changed, 1 deletion(-)
arch/x86/kernel/cpu/perf_event_intel.c +0 -1
index 2502d0d..ef6c8b7 100644
@@ -2541,7 +2541,6 @@ __init int intel_pmu_init(void)
 
 	case 60: /* Haswell Client */
 	case 70:
-	case 71:
 	case 63:
 	case 69:
 		x86_pmu.late_ack = true;
Signature · 1 line
1.9.3
author
#2 lore raw

PATCH 2/5 perf, x86: Document all Haswell models

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

Add names for each Haswell model as requested by Peter.

---
 arch/x86/kernel/cpu/perf_event_intel.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +3 -3
index ef6c8b7..03befdd 100644
@@ -2540,9 +2540,9 @@ __init int intel_pmu_init(void)
 
 
 	case 60: /* Haswell Client */
-	case 70:
-	case 63:
-	case 69:
+	case 70: /* Crystall Well */
+	case 63: /* Haswell Server */
+	case 69: /* Haswell ULT */
 		x86_pmu.late_ack = true;
 		memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, sizeof(hw_cache_event_ids));
 		memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
Signature · 1 line
1.9.3
maintainer
#3 lore raw

Re: [PATCH 2/5] perf, x86: Document all Haswell models

Quoted 14 lines · Andi Kleen

From: Andi Kleen <[email protected]>

Add names for each Haswell model as requested by Peter.

---
 arch/x86/kernel/cpu/perf_event_intel.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +0 -0
index ef6c8b7..03befdd 100644

arch/x86/kernel/cpu/perf_event_intel.c __init int intel_pmu_init(void) · line 2545

Quoted 6 lines
arch/x86/kernel/cpu/perf_event_intel.c +3 -0
@@ -2540,9 +2540,9 @@ __init int intel_pmu_init(void)
 
 
 	case 60: /* Haswell Client */
+	case 70: /* Crystall Well */
+	case 63: /* Haswell Server */
+	case 69: /* Haswell ULT */

So I googled Crystalwell, and I'm not sure I understand; is that Haswell-H + GT3e or it that Haswell + GT3e, the distinction being that there are also Desktop parts with Iris Pro 5200, such like the Haswell-R.

Haswell Server, is that the single socket one, or is it like IVB both the EP and EX parts?

And is 69 only the ULT or also the ULX parts?

Would something like this be accurate:?

case 60: /* 22nm Haswell */
case 70: /* 22nm Haswell + GT3e (Iris Pro 5200) */
case 69: /* 22nm Haswell ULT */
case 63: /* 22nm Haswell-EP/EX */

#4 lore raw

Re: [PATCH 2/5] perf, x86: Document all Haswell models

Quoted 1 line

Would something like this be accurate:?

I believe my version is more accurate, but feel free to use yours.

-Andi

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

PATCH 3/5 perf, x86: Add Broadwell core support

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

Add Broadwell support for Broadwell Client to perf. This is very similar to Haswell. It uses a new cache event table, because there were various changes there.

The constraint list has one new event that needs to be handled over Haswell.

The PEBS event list is the same, so we reuse Haswell's.

v2: Remove unnamed model numbers.

---
 arch/x86/kernel/cpu/perf_event_intel.c | 150 +++++++++++++++++++++++++++++++++
 1 file changed, 150 insertions(+)
arch/x86/kernel/cpu/perf_event_intel.c +150 -0
index 03befdd..4bfb0ec 100644
@@ -220,6 +220,15 @@ static struct event_constraint intel_hsw_event_constraints[] = {
 	EVENT_CONSTRAINT_END
 };
 
+struct event_constraint intel_bdw_event_constraints[] = {
+	FIXED_EVENT_CONSTRAINT(0x00c0, 0),	/* INST_RETIRED.ANY */
+	FIXED_EVENT_CONSTRAINT(0x003c, 1),	/* CPU_CLK_UNHALTED.CORE */
+	FIXED_EVENT_CONSTRAINT(0x0300, 2),	/* CPU_CLK_UNHALTED.REF */
+	INTEL_UEVENT_CONSTRAINT(0x148, 0x4),	/* L1D_PEND_MISS.PENDING */
+	INTEL_EVENT_CONSTRAINT(0xa3, 0x4),	/* CYCLE_ACTIVITY.* */
+	EVENT_CONSTRAINT_END
+};
+
 static u64 intel_pmu_event_map(int hw_event)
 {
 	return intel_perfmon_event_map[hw_event];
@@ -415,6 +424,126 @@ static __initconst const u64 snb_hw_cache_event_ids
 
 };
 
+static __initconst const u64 bdw_hw_cache_event_ids
+				[PERF_COUNT_HW_CACHE_MAX]
+				[PERF_COUNT_HW_CACHE_OP_MAX]
+				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
+{
+ [ C(L1D ) ] = {
+	[ C(OP_READ) ] = {
+		[ C(RESULT_ACCESS) ] = 0x81d0, 	/* MEM_UOPS_RETIRED.ALL_LOADS */
+		[ C(RESULT_MISS)   ] = 0x151, 	/* L1D.REPLACEMENT */
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = 0x82d0, 	/* MEM_UOPS_RETIRED.ALL_STORES */
+		[ C(RESULT_MISS)   ] = 0x0,
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = 0x0,
+		[ C(RESULT_MISS)   ] = 0x0,
+	},
+ },
+ [ C(L1I ) ] = {
+	[ C(OP_READ) ] = {
+		[ C(RESULT_ACCESS) ] = 0x0,
+		[ C(RESULT_MISS)   ] = 0x280, 	/* ICACHE.MISSES */
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = -1,
+		[ C(RESULT_MISS)   ] = -1,
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = 0x0,
+		[ C(RESULT_MISS)   ] = 0x0,
+	},
+ },
+ [ C(LL  ) ] = {
+	[ C(OP_READ) ] = {
+		/* OFFCORE_RESPONSE:ALL_DATA_RD|ALL_CODE_RD */
+		[ C(RESULT_ACCESS) ] = 0x1b7,
+		/* OFFCORE_RESPONSE:ALL_DATA_RD|ALL_CODE_RD|SUPPLIER_NONE|
+                   L3_MISS|ANY_SNOOP */
+		[ C(RESULT_MISS)   ] = 0x1b7,
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = 0x1b7, 	/* OFFCORE_RESPONSE:ALL_RFO */
+		/* OFFCORE_RESPONSE:ALL_RFO|SUPPLIER_NONE|L3_MISS|ANY_SNOOP */
+		[ C(RESULT_MISS)   ] = 0x1b7,
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = 0x0,
+		[ C(RESULT_MISS)   ] = 0x0,
+	},
+ },
+ [ C(DTLB) ] = {
+	[ C(OP_READ) ] = {
+		[ C(RESULT_ACCESS) ] = 0x81d0, 	/* MEM_UOPS_RETIRED.ALL_LOADS */
+		[ C(RESULT_MISS)   ] = 0x108, 	/* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = 0x82d0, 	/* MEM_UOPS_RETIRED.ALL_STORES */
+		[ C(RESULT_MISS)   ] = 0x149, 	/* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = 0x0,
+		[ C(RESULT_MISS)   ] = 0x0,
+	},
+ },
+ [ C(ITLB) ] = {
+	[ C(OP_READ) ] = {
+		[ C(RESULT_ACCESS) ] = 0x6085, 	/* ITLB_MISSES.STLB_HIT */
+		[ C(RESULT_MISS)   ] = 0x185, 	/* ITLB_MISSES.MISS_CAUSES_A_WALK */
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = -1,
+		[ C(RESULT_MISS)   ] = -1,
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = -1,
+		[ C(RESULT_MISS)   ] = -1,
+	},
+ },
+ [ C(BPU ) ] = {
+	[ C(OP_READ) ] = {
+		[ C(RESULT_ACCESS) ] = 0xc4, 	/* BR_INST_RETIRED.ALL_BRANCHES */
+		[ C(RESULT_MISS)   ] = 0xc5, 	/* BR_MISP_RETIRED.ALL_BRANCHES */
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = -1,
+		[ C(RESULT_MISS)   ] = -1,
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = -1,
+		[ C(RESULT_MISS)   ] = -1,
+	},
+ },
+};
+
+static __initconst const u64 bdw_hw_cache_extra_regs
+				[PERF_COUNT_HW_CACHE_MAX]
+				[PERF_COUNT_HW_CACHE_OP_MAX]
+				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
+{
+ [ C(LL  ) ] = {
+	[ C(OP_READ) ] = {
+		/* OFFCORE_RESPONSE:ALL_DATA_RD|ALL_CODE_RD */
+		[ C(RESULT_ACCESS) ] = 0x2d5,
+		/* OFFCORE_RESPONSE:ALL_DATA_RD|ALL_CODE_RD|SUPPLIER_NONE|
+                   L3_MISS|ANY_SNOOP */
+		[ C(RESULT_MISS)   ] = 0x3fbc0202d5,
+	},
+	[ C(OP_WRITE) ] = {
+		[ C(RESULT_ACCESS) ] = 0x122, 	/* OFFCORE_RESPONSE:ALL_RFO */
+		/* OFFCORE_RESPONSE:ALL_RFO|SUPPLIER_NONE|L3_MISS|ANY_SNOOP */
+		[ C(RESULT_MISS)   ] = 0x3fbc020122,
+	},
+	[ C(OP_PREFETCH) ] = {
+		[ C(RESULT_ACCESS) ] = 0x0,
+		[ C(RESULT_MISS)   ] = 0x0,
+	},
+ },
+};
+
 static __initconst const u64 westmere_hw_cache_event_ids
 				[PERF_COUNT_HW_CACHE_MAX]
 				[PERF_COUNT_HW_CACHE_OP_MAX]
@@ -2564,6 +2693,27 @@ __init int intel_pmu_init(void)
 		pr_cont("Haswell events, ");
 		break;
 
+	case 61: /* Broadwell Client */
+		x86_pmu.late_ack = true;
+		memcpy(hw_cache_event_ids, bdw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
+		memcpy(hw_cache_extra_regs, bdw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
+
+		intel_pmu_lbr_init_snb();
+
+		x86_pmu.event_constraints = intel_bdw_event_constraints;
+		x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints;
+		x86_pmu.extra_regs = intel_snbep_extra_regs;
+		x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
+		/* all extra regs are per-cpu when HT is on */
+		x86_pmu.er_flags |= ERF_HAS_RSP_1;
+		x86_pmu.er_flags |= ERF_NO_HT_SHARING;
+
+		x86_pmu.hw_config = hsw_hw_config;
+		x86_pmu.get_event_constraints = hsw_get_event_constraints;
+		x86_pmu.cpu_events = hsw_events_attrs;
+		pr_cont("Broadwell events, ");
+		break;
+
 	default:
 		switch (x86_pmu.version) {
 		case 1:
Signature · 1 line
1.9.3
maintainer
#6 lore raw

Re: [PATCH 3/5] perf, x86: Add Broadwell core support

arch/x86/kernel/cpu/perf_event_intel.c __init int intel_pmu_init(void) · line 2696

Quoted 3 lines · Andi Kleen

v2: Remove unnamed model numbers.

arch/x86/kernel/cpu/perf_event_intel.c +1 -0
+	case 61: /* Broadwell Client */

Seriously? So 71 and 79 are fine as a number but you cannot name them?

maintainer
#7 lore raw

Re: [PATCH 3/5] perf, x86: Add Broadwell core support

Quoted 7 lines · Peter Zijlstra

arch/x86/kernel/cpu/perf_event_intel.c __init int intel_pmu_init(void) · line 2696

Quoted 3 lines · Andi Kleen

v2: Remove unnamed model numbers.

arch/x86/kernel/cpu/perf_event_intel.c +1 -0
+	case 61: /* Broadwell Client */

Seriously? So 71 and 79 are fine as a number but you cannot name them?

would something like:

	case 71: /* 14nm Broadwell TBA */

work? Where you then will update the description at the appropriate time?

#8 lore raw

Re: [PATCH 3/5] perf, x86: Add Broadwell core support

Quoted 9 lines · Peter Zijlstra
Quoted 7 lines · Peter Zijlstra

arch/x86/kernel/cpu/perf_event_intel.c __init int intel_pmu_init(void) · line 2696

Quoted 3 lines · Andi Kleen

v2: Remove unnamed model numbers.

arch/x86/kernel/cpu/perf_event_intel.c +1 -0
+	case 61: /* Broadwell Client */

Seriously? So 71 and 79 are fine as a number but you cannot name them?

Quoted 6 lines

would something like:

	case 71: /* 14nm Broadwell TBA */

work? Where you then will update the description at the appropriate time?

There won't be a single description, it will turn into all kinds of products. The names I used are the ones we use.

Really, Peter, it's futile to try to reproduce http://ark.intel.com
in perf comments. If you need information on specific Intel products please look it up there.

-Andi

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

Re: [PATCH 3/5] perf, x86: Add Broadwell core support

Quoted 3 lines · Andi Kleen

Really, Peter, it's futile to try to reproduce http://ark.intel.com
in perf comments. If you need information on specific Intel products please look it up there.

ark.intel.com is useless, I never can find anything there. The best I can find is wikipedia and they stopped listing model numbers long ago :/

author
#10 lore raw

PATCH 4/5 perf, x86: Add INST_RETIRED.ALL workarounds

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

On Broadwell INST_RETIRED.ALL cannot be used with any period that doesn't have the lowest 6 bits cleared. And the period should not be smaller than 128.

Add a new callback to enforce this, and set it for Broadwell.

This is erratum BDM57 and BDM11.

v2: Use correct event name in description. Use EVENT() macro.

---
 arch/x86/kernel/cpu/perf_event.c       |  3 +++
 arch/x86/kernel/cpu/perf_event.h       |  1 +
 arch/x86/kernel/cpu/perf_event_intel.c | 19 +++++++++++++++++++
 3 files changed, 23 insertions(+)
arch/x86/kernel/cpu/perf_event.c +3 -0
index 0adc5e3..a0adf58 100644
@@ -980,6 +980,9 @@ int x86_perf_event_set_period(struct perf_event *event)
 	if (left > x86_pmu.max_period)
 		left = x86_pmu.max_period;
 
+	if (x86_pmu.limit_period)
+		left = x86_pmu.limit_period(event, left);
+
 	per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
 
 	/*
arch/x86/kernel/cpu/perf_event.h +1 -0
index de81627..a46e391 100644
@@ -456,6 +456,7 @@ struct x86_pmu {
 	struct x86_pmu_quirk *quirks;
 	int		perfctr_second_write;
 	bool		late_ack;
+	unsigned	(*limit_period)(struct perf_event *event, unsigned l);
 
 	/*
 	 * sysfs attrs
arch/x86/kernel/cpu/perf_event_intel.c +19 -0
index 4bfb0ec..66260e1 100644
@@ -2034,6 +2034,24 @@ hsw_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
 	return c;
 }
 
+/*
+ * Broadwell:
+ * The INST_RETIRED.ALL period always needs to have lowest
+ * 6bits cleared (BDM57). It shall not use a period smaller
+ * than 100 (BDM11). We combine the two to enforce
+ * a min-period of 128.
+ */
+static unsigned bdw_limit_period(struct perf_event *event, unsigned left)
+{
+	if ((event->hw.config & 0xffff) ==
+			X86_CONFIG(.event=0xc0, .umask=0x01)) {
+		if (left < 128)
+			left = 128;
+		left &= ~0x3fu;
+	}
+	return left;
+}
+
 PMU_FORMAT_ATTR(event,	"config:0-7"	);
 PMU_FORMAT_ATTR(umask,	"config:8-15"	);
 PMU_FORMAT_ATTR(edge,	"config:18"	);
@@ -2711,6 +2729,7 @@ __init int intel_pmu_init(void)
 		x86_pmu.hw_config = hsw_hw_config;
 		x86_pmu.get_event_constraints = hsw_get_event_constraints;
 		x86_pmu.cpu_events = hsw_events_attrs;
+		x86_pmu.limit_period = bdw_limit_period;
 		pr_cont("Broadwell events, ");
 		break;
 
Signature · 1 line
1.9.3
#11 lore raw

Re: [PATCH 4/5] perf, x86: Add INST_RETIRED.ALL workarounds

Quoted 6 lines · Andi Kleen

From: Andi Kleen <[email protected]>

On Broadwell INST_RETIRED.ALL cannot be used with any period that doesn't have the lowest 6 bits cleared. And the period should not be smaller than 128.

If you have frequency mode enabled, then I suspect this works okay. It may be that the kernel will keep trying to set a period lower than 128 but you will correct it.

I am more worried about the case where the user sets a fixed period with some of the bottom 6 bits set. The apps thinks it is sampling at X occurences per sample, when it is in fact at X - 63 (worst case). I think this would be okay, if there was a trace of this in the sampling buffer, i.e., PERF_SAMPLE_PERIOD. But I recall that perf record does not request this flag when using a fixed period. So no way of knowing what the actual period was, at least when using the perf tool. I understand also that for this event, 64 occurrences may not matter as much, maybe except if you use some filters such as cmask.

Quoted 78 lines

Add a new callback to enforce this, and set it for Broadwell.

This is erratum BDM57 and BDM11.

v2: Use correct event name in description. Use EVENT() macro.

---
 arch/x86/kernel/cpu/perf_event.c       |  3 +++
 arch/x86/kernel/cpu/perf_event.h       |  1 +
 arch/x86/kernel/cpu/perf_event_intel.c | 19 +++++++++++++++++++
 3 files changed, 23 insertions(+)
arch/x86/kernel/cpu/perf_event.c +3 -0
index 0adc5e3..a0adf58 100644
@@ -980,6 +980,9 @@ int x86_perf_event_set_period(struct perf_event *event)
        if (left > x86_pmu.max_period)
                left = x86_pmu.max_period;
 
+       if (x86_pmu.limit_period)
+               left = x86_pmu.limit_period(event, left);
+
        per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
 
        /*
arch/x86/kernel/cpu/perf_event.h +1 -0
index de81627..a46e391 100644
@@ -456,6 +456,7 @@ struct x86_pmu {
        struct x86_pmu_quirk *quirks;
        int             perfctr_second_write;
        bool            late_ack;
+       unsigned        (*limit_period)(struct perf_event *event, unsigned l);
 
        /*
         * sysfs attrs
arch/x86/kernel/cpu/perf_event_intel.c +19 -1
index 4bfb0ec..66260e1 100644
@@ -2034,6 +2034,24 @@ hsw_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
        return c;
 }
 
+/*
+ * Broadwell:
+ * The INST_RETIRED.ALL period always needs to have lowest
+ * 6bits cleared (BDM57). It shall not use a period smaller
+ * than 100 (BDM11). We combine the two to enforce
+ * a min-period of 128.
+ */
+static unsigned bdw_limit_period(struct perf_event *event, unsigned left)
+{
+       if ((event->hw.config & 0xffff) ==
+                       X86_CONFIG(.event=0xc0, .umask=0x01)) {
+               if (left < 128)
+                       left = 128;
+               left &= ~0x3fu;
+       }
+       return left;
+}
+
 PMU_FORMAT_ATTR(event, "config:0-7"    );
 PMU_FORMAT_ATTR(umask, "config:8-15"   );
 PMU_FORMAT_ATTR(edge,  "config:18"     );
@@ -2711,6 +2729,7 @@ __init int intel_pmu_init(void)
                x86_pmu.hw_config = hsw_hw_config;
                x86_pmu.get_event_constraints = hsw_get_event_constraints;
                x86_pmu.cpu_events = hsw_events_attrs;
+               x86_pmu.limit_period = bdw_limit_period;
                pr_cont("Broadwell events, ");
                break;
 
--

1.9.3

#12 lore raw

Re: [PATCH 4/5] perf, x86: Add INST_RETIRED.ALL workarounds

I understand all your points, but there's no alternative. The only other way would be to disable INST_RETIRED.ALL.

-Andi

#13 lore raw

Re: [PATCH 4/5] perf, x86: Add INST_RETIRED.ALL workarounds

[+perf tool maintainers]

Quoted 4 lines · Andi Kleen

I understand all your points, but there's no alternative. The only other way would be to disable INST_RETIRED.ALL.

You cannot do that either. INST_RETIRED:ALL is important. I assume the bug applies whether or not the event is used with a filter.

I think we need to ensure that by looking at the perf.data file, one can reconstruct the total number of inst_Retired:all occurrences for the run. With a fixed period, one would do num_samples * fixed_period. I know the Gooda tool does that. It is used to estimate the number of events captured vs. the number of events occurring.

So what I am saying is that on BDW, we need to have perf force PERF_SAMPLE_PERIOD even with a fixed period is passed. The kernel cannot simply add PERF_SAMPLE_PERIOD to the event sample_format because the tool would get confused by the sample record size otherwise.

I understand this is not pretty. We'd need to have a on the evsel just before making the syscall. That callback would be for x86 and it would check the CPUID to force PERF_SAMPLE_PERIOD. That's my thinking but there may be a better way of doing this.

#14 lore raw

Re: [PATCH 4/5] perf, x86: Add INST_RETIRED.ALL workarounds

Quoted 12 lines · Stephane Eranian

[+perf tool maintainers]

Quoted 4 lines · Andi Kleen

I understand all your points, but there's no alternative. The only other way would be to disable INST_RETIRED.ALL.

You cannot do that either. INST_RETIRED:ALL is important. I assume the bug applies whether or not the event is used with a filter.

Quoted 6 lines

I think we need to ensure that by looking at the perf.data file, one can reconstruct the total number of inst_Retired:all occurrences for the run. With a fixed period, one would do num_samples * fixed_period. I know the Gooda tool does that. It is used to estimate the number of events captured vs. the number of events occurring.

Is that really a problem?

Normally periods are not that small, especially not for instruction retired. I don't think you can run such a small period on instruction retired for
any significant time without throttling.

With sensible periods, let's say >10k, the error from losing a few bits is very small. It would surprise me if you can actually measure it.

There will be always much more jitter just from standard system noise.

-Andi

maintainer
#15 lore raw

Re: [PATCH 4/5] perf, x86: Add INST_RETIRED.ALL workarounds

In reply to Stephane Eranian
Quoted 7 lines · Stephane Eranian

[+perf tool maintainers]

Quoted 4 lines · Andi Kleen

I understand all your points, but there's no alternative. The only other way would be to disable INST_RETIRED.ALL.

Quoted 8 lines

You cannot do that either. INST_RETIRED:ALL is important. I assume the bug applies whether or not the event is used with a filter.

I think we need to ensure that by looking at the perf.data file, one can reconstruct the total number of inst_Retired:all occurrences for the run. With a fixed period, one would do num_samples * fixed_period. I know the Gooda tool does that. It is used to estimate the number of events captured vs. the number of events occurring.

OK, I think we can make that work; IFF we guarantee perf_event_attr::sample_period >= 128.

Suppose we start out with sample_period=192; then we'll set period_left to 192, we'll end up with left = 128 (we truncate the lower bits). We get an interrupt, find that period_left = 64 (>0 so we return 0 and don't get an overflow handler), up that to 128. Then we trigger again, at n=256. Then we find period_left = -64 (<=0 so we return 1 and do get an overflow). We increment with sample_period so we get left = 128. We fire again, at n=384, period_left = 0 (<=0 so we return 1 and get an overflow). And on and on.

So while the individual interrupts are 'wrong' we get then with interval=256,128 in exactly the right ratio to average out at 192. And this works for everything >=128.

So the num_samples*fixed_period thing is still entirely correct +- 127, which is good enough I'd say, as you already have that error anyhow.

So no need to 'fix' the tools, al we need to do is refuse to create INST_RETIRED:ALL events with sample_period < 128.

author
#16 lore raw

Re: [PATCH 4/5] perf, x86: Add INST_RETIRED.ALL workarounds

Quoted 2 lines

So no need to 'fix' the tools, al we need to do is refuse to create INST_RETIRED:ALL events with sample_period < 128.

Like this?

arch/x86/kernel/cpu/perf_event.c +6 -0
index a0adf58..c592d4d 100644
@@ -443,6 +443,12 @@ int x86_pmu_hw_config(struct perf_event *event)
 	if (event->attr.type == PERF_TYPE_RAW)
 		event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
 
+	if (event->attr.sample_period && x86_pmu.limit_period) {
+		if (x86_pmu.limit_period(event, event->attr.sample_period) <
+				event->attr.sample_period)
+			return -EINVAL;
+	}
+
 	return x86_setup_perfctr(event);
 }
 

-Andi

maintainer
#17 lore raw

Re: [PATCH 4/5] perf, x86: Add INST_RETIRED.ALL workarounds

In reply to Andi Kleen

arch/x86/kernel/cpu/perf_event_intel.c hsw_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) · line 2046

Quoted 5 lines · Andi Kleen

v2: Use correct event name in description. Use EVENT() macro.

arch/x86/kernel/cpu/perf_event_intel.c +3 -0
+static unsigned bdw_limit_period(struct perf_event *event, unsigned left)
+{
+	if ((event->hw.config & 0xffff) ==

I was thinking you should use INTEL_ARCH_EVENT_MASK or something instead of the raw 0xFFFF there, but that X86_CONFIG() usage is nice too :-)

Quoted 7 lines
arch/x86/kernel/cpu/perf_event_intel.c +7 -0
+			X86_CONFIG(.event=0xc0, .umask=0x01)) {
+		if (left < 128)
+			left = 128;
+		left &= ~0x3fu;
+	}
+	return left;
+}
author
#18 lore raw

PATCH 5/5 perf, x86: Use Broadwell cache event list for Haswell

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

Use the newly added Broadwell cache event list for Haswell too. They are identical, but Haswell is very different from the Sandy Bridge list that was used previously. That fixes a wide range of mis-counting cache events.

The prefetch events are gone now. They way the hardware counts them is very misleading (some prefetches included, others not), so it seemed best to leave them out.

---
 arch/x86/kernel/cpu/perf_event_intel.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +2 -2
index 66260e1..178ddc0 100644
@@ -2691,8 +2691,8 @@ __init int intel_pmu_init(void)
 	case 63: /* Haswell Server */
 	case 69: /* Haswell ULT */
 		x86_pmu.late_ack = true;
-		memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, sizeof(hw_cache_event_ids));
-		memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
+		memcpy(hw_cache_event_ids, bdw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
+		memcpy(hw_cache_extra_regs, bdw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
 
 		intel_pmu_lbr_init_snb();
 
Signature · 1 line
1.9.3