Skip to contentSkip to conversation

x86, perf: Make merge_attr global to use from perf_event_intel

PATCH In review Andi Kleen 6 participants 22 messages Jun 29, 2015 – Aug 4, 2015

Subsystems (2)
lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

PATCH 1/3 x86, perf: Make merge_attr global to use from perf_event_intel

From: Andi Kleen <[email protected]>

merge_attr allows to merge two sysfs attribute tables. Export it to be usable by other files too.

Next patch is going to use that to extend the sysfs format attributes for a CPU.

---
 arch/x86/kernel/cpu/perf_event.c | 2 +-
 arch/x86/kernel/cpu/perf_event.h | 3 +++
 2 files changed, 4 insertions(+), 1 deletion(-)
arch/x86/kernel/cpu/perf_event.c +1 -1
index 5801a14..81e2a74 100644
@@ -1561,7 +1561,7 @@ static void __init filter_events(struct attribute **attrs)
 }
 
 /* Merge two pointer arrays */
-static __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
+__init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
 {
 	struct attribute **new;
 	int j, i;
arch/x86/kernel/cpu/perf_event.h +3 -0
index 0db9de4..3598f67 100644
@@ -795,6 +795,8 @@ static inline void set_linear_ip(struct pt_regs *regs, unsigned long ip)
 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event);
 ssize_t intel_event_sysfs_show(char *page, u64 config);
 
+struct attribute **merge_attr(struct attribute **a, struct attribute **b);
+
 #ifdef CONFIG_CPU_SUP_AMD
 
 int amd_pmu_init(void);
@@ -940,6 +942,7 @@ static inline int is_ht_workaround_enabled(void)
 {
 	return !!(x86_pmu.flags & PMU_FL_EXCL_ENABLED);
 }
+
 #else /* CONFIG_CPU_SUP_INTEL */
 
 static inline void reserve_ds_buffers(void)
Signature · 1 line
2.4.2
author
#1 lore raw

PATCH 2/3 x86, perf: Support custom test values for extra_regs

From: Andi Kleen <[email protected]>

Current kernels always test extra registers at boot with RMW cycle, to catch lying virtual machines.

For a new register the standard 0x1ff test value does not work, as 0x1ff is not a valid value and causes an #GP.

Add the ability to add custom test values to an extra_reg into the description tables.

---
 arch/x86/kernel/cpu/perf_event.h                    | 15 +++++++++++----
 arch/x86/kernel/cpu/perf_event_intel.c              |  3 ++-
 arch/x86/kernel/cpu/perf_event_intel_uncore_nhmex.c |  5 +++--
 3 files changed, 16 insertions(+), 7 deletions(-)
arch/x86/kernel/cpu/perf_event.h +11 -4
index 3598f67..2a4701c 100644
@@ -422,23 +422,30 @@ struct extra_reg {
 	u64			valid_mask;
 	int			idx;  /* per_xxx->regs[] reg index */
 	bool			extra_msr_access;
+	u64			test_value;
 };
 
-#define EVENT_EXTRA_REG(e, ms, m, vm, i) {	\
+#define EVENT_EXTRA_REG(e, ms, m, vm, i, t) {	\
 	.event = (e),			\
 	.msr = (ms),			\
 	.config_mask = (m),		\
 	.valid_mask = (vm),		\
 	.idx = EXTRA_REG_##i,		\
 	.extra_msr_access = true,	\
+	.test_value = t,		\
 	}
 
 #define INTEL_EVENT_EXTRA_REG(event, msr, vm, idx)	\
-	EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm, idx)
+	EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm, idx, \
+			0x1ffUL)
 
 #define INTEL_UEVENT_EXTRA_REG(event, msr, vm, idx) \
 	EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT | \
-			ARCH_PERFMON_EVENTSEL_UMASK, vm, idx)
+			ARCH_PERFMON_EVENTSEL_UMASK, vm, idx, 0x1ffUL)
+
+#define INTEL_UEVENT_EXTRA_REG_TVAL(event, msr, vm, idx, tval) \
+	EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT | \
+			ARCH_PERFMON_EVENTSEL_UMASK, vm, idx, tval)
 
 #define INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(c) \
 	INTEL_UEVENT_EXTRA_REG(c, \
@@ -446,7 +453,7 @@ struct extra_reg {
 			       0xffff, \
 			       LDLAT)
 
-#define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0, RSP_0)
+#define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0, RSP_0, 0)
 
 union perf_capabilities {
 	struct {
arch/x86/kernel/cpu/perf_event_intel.c +2 -1
index 7c397e8..4df6783 100644
@@ -3575,7 +3575,8 @@ __init int intel_pmu_init(void)
 	 */
 	if (x86_pmu.extra_regs) {
 		for (er = x86_pmu.extra_regs; er->msr; er++) {
-			er->extra_msr_access = check_msr(er->msr, 0x1ffUL);
+			er->extra_msr_access = check_msr(er->msr,
+							 er->test_value);
 			/* Disable LBR select mapping */
 			if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access)
 				x86_pmu.lbr_sel_map = NULL;
arch/x86/kernel/cpu/perf_event_intel_uncore_nhmex.c +3 -2
index 2749965..5bfa493 100644
@@ -136,11 +136,12 @@
 				NHMEX_M_PMON_CTL_FLAG_MODE)
 #define MBOX_INC_SEL_EXTAR_REG(c, r) \
 		EVENT_EXTRA_REG(MBOX_INC_SEL(c), NHMEX_M0_MSR_PMU_##r, \
-				MBOX_INC_SEL_MASK, (u64)-1, NHMEX_M_##r)
+				MBOX_INC_SEL_MASK, (u64)-1, NHMEX_M_##r, \
+				0x1ffUL)
 #define MBOX_SET_FLAG_SEL_EXTRA_REG(c, r) \
 		EVENT_EXTRA_REG(MBOX_SET_FLAG_SEL(c), NHMEX_M0_MSR_PMU_##r, \
 				MBOX_SET_FLAG_SEL_MASK, \
-				(u64)-1, NHMEX_M_##r)
+				(u64)-1, NHMEX_M_##r, 0x1ffUL)
 
 /* NHM-EX Rbox */
 #define NHMEX_R_MSR_GLOBAL_CTL			0xe00
Signature · 1 line
2.4.2
maintainer
#2 lore raw

Re: [PATCH 2/3] x86, perf: Support custom test values for extra_regs

Quoted 10 lines · Andi Kleen

From: Andi Kleen <[email protected]>

Current kernels always test extra registers at boot with RMW cycle, to catch lying virtual machines.

For a new register the standard 0x1ff test value does not work, as 0x1ff is not a valid value and causes an #GP.

Add the ability to add custom test values to an extra_reg into the description tables.

Just wondering; is there not a test value that works for all registers?

#3 lore raw

Re: [PATCH 2/3] x86, perf: Support custom test values for extra_regs

Quoted 6 lines · Peter Zijlstra
Quoted 5 lines · Andi Kleen

From: Andi Kleen <[email protected]>

Current kernels always test extra registers at boot with RMW cycle, to catch lying virtual machines.

Quoted 7 lines
Quoted 5 lines

For a new register the standard 0x1ff test value does not work, as 0x1ff is not a valid value and causes an #GP.

Add the ability to add custom test values to an extra_reg into the description tables.

Just wondering; is there not a test value that works for all registers?

0x11 will likely work everywhere.

-Andi

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

PATCH 3/3 x86, perf: Add PEBS frontend profiling for Skylake

In reply to Andi Kleen

From: Andi Kleen <[email protected]>

Skylake has a new FRONTEND_LATENCY PEBS event to accurate profile frontend problems (like ITLB or decoding issues)

The new event is configured through a separate MSR, which selects a range of sub events.

Define the extra MSR as a extra reg and export support for it through sysfs. To avoid duplicating the existing tables use a new function to add new entries to existing tables.

---
 arch/x86/include/asm/msr-index.h       |  2 ++
 arch/x86/kernel/cpu/perf_event.h       |  1 +
 arch/x86/kernel/cpu/perf_event_intel.c | 11 ++++++++++-
 3 files changed, 13 insertions(+), 1 deletion(-)
arch/x86/include/asm/msr-index.h +2 -0
index 96a00de..a5371dc 100644
@@ -133,6 +133,8 @@
 #define DEBUGCTLMSR_BTS_OFF_USR		(1UL << 10)
 #define DEBUGCTLMSR_FREEZE_LBRS_ON_PMI	(1UL << 11)
 
+#define MSR_PEBS_FRONTEND		0x000003f7
+
 #define MSR_IA32_POWER_CTL		0x000001fc
 
 #define MSR_IA32_MC0_CTL		0x00000400
arch/x86/kernel/cpu/perf_event.h +1 -0
index 2a4701c..6bd7390 100644
@@ -47,6 +47,7 @@ enum extra_reg_type {
 	EXTRA_REG_RSP_1 = 1,	/* offcore_response_1 */
 	EXTRA_REG_LBR   = 2,	/* lbr_select */
 	EXTRA_REG_LDLAT = 3,	/* ld_lat_threshold */
+	EXTRA_REG_FE    = 4,    /* fe_* */
 
 	EXTRA_REG_MAX		/* number of entries needed */
 };
arch/x86/kernel/cpu/perf_event_intel.c +10 -1
index 4df6783..c6cac61 100644
@@ -205,6 +205,7 @@ static struct extra_reg intel_skl_extra_regs[] __read_mostly = {
 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
 	INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
+	INTEL_UEVENT_EXTRA_REG_TVAL(0x01c6, MSR_PEBS_FRONTEND, 0x3fff17, FE, 0x11UL),
 	EVENT_EXTRA_END
 };
 
@@ -2875,6 +2876,8 @@ PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63");
 
 PMU_FORMAT_ATTR(ldlat, "config1:0-15");
 
+PMU_FORMAT_ATTR(frontend, "config1:0-23");
+
 static struct attribute *intel_arch3_formats_attr[] = {
 	&format_attr_event.attr,
 	&format_attr_umask.attr,
@@ -2891,6 +2894,11 @@ static struct attribute *intel_arch3_formats_attr[] = {
 	NULL,
 };
 
+static struct attribute *skl_format_attr[] = {
+	&format_attr_frontend.attr,
+	NULL,
+};
+
 static __initconst const struct x86_pmu core_pmu = {
 	.name			= "core",
 	.handle_irq		= x86_pmu_handle_irq,
@@ -3500,7 +3508,8 @@ __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.format_attrs = merge_attr(intel_arch3_formats_attr,
+						  skl_format_attr);
 		WARN_ON(!x86_pmu.format_attrs);
 		x86_pmu.cpu_events = hsw_events_attrs;
 		pr_cont("Skylake events, ");
Signature · 1 line
2.4.2
#5 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 49 lines · Andi Kleen

From: Andi Kleen <[email protected]>

Skylake has a new FRONTEND_LATENCY PEBS event to accurate profile frontend problems (like ITLB or decoding issues)

The new event is configured through a separate MSR, which selects a range of sub events.

Define the extra MSR as a extra reg and export support for it through sysfs. To avoid duplicating the existing tables use a new function to add new entries to existing tables.

---
 arch/x86/include/asm/msr-index.h       |  2 ++
 arch/x86/kernel/cpu/perf_event.h       |  1 +
 arch/x86/kernel/cpu/perf_event_intel.c | 11 ++++++++++-
 3 files changed, 13 insertions(+), 1 deletion(-)
arch/x86/include/asm/msr-index.h +2 -0
index 96a00de..a5371dc 100644
@@ -133,6 +133,8 @@
 #define DEBUGCTLMSR_BTS_OFF_USR                (1UL << 10)
 #define DEBUGCTLMSR_FREEZE_LBRS_ON_PMI (1UL << 11)
 
+#define MSR_PEBS_FRONTEND              0x000003f7
+
 #define MSR_IA32_POWER_CTL             0x000001fc
 
 #define MSR_IA32_MC0_CTL               0x00000400
arch/x86/kernel/cpu/perf_event.h +1 -0
index 2a4701c..6bd7390 100644
@@ -47,6 +47,7 @@ enum extra_reg_type {
        EXTRA_REG_RSP_1 = 1,    /* offcore_response_1 */
        EXTRA_REG_LBR   = 2,    /* lbr_select */
        EXTRA_REG_LDLAT = 3,    /* ld_lat_threshold */
+       EXTRA_REG_FE    = 4,    /* fe_* */
 
        EXTRA_REG_MAX           /* number of entries needed */
 };
arch/x86/kernel/cpu/perf_event_intel.c +0 -0
index 4df6783..c6cac61 100644

arch/x86/kernel/cpu/perf_event_intel.c static struct extra_reg intel_skl_extra_regs[] __read_mostly = { · line 208

Quoted 4 lines
arch/x86/kernel/cpu/perf_event_intel.c +1 -0
@@ -205,6 +205,7 @@ static struct extra_reg intel_skl_extra_regs[] __read_mostly = {
        INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
        INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
        INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
+       INTEL_UEVENT_EXTRA_REG_TVAL(0x01c6, MSR_PEBS_FRONTEND, 0x3fff17, FE, 0x11UL),

I believe this mask of 0x3fff17 is wrong and should instead be 0x7fffff based on the description of the FRONTEND
MSR I see in the SDM Table 18-54 (bit 0-22 are valid). Otherwise, some valid latency values may be rejected.

Quoted 37 lines
arch/x86/kernel/cpu/perf_event_intel.c +9 -2
        EVENT_EXTRA_END
 };
 
@@ -2875,6 +2876,8 @@ PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63");
 
 PMU_FORMAT_ATTR(ldlat, "config1:0-15");
 
+PMU_FORMAT_ATTR(frontend, "config1:0-23");
+
 static struct attribute *intel_arch3_formats_attr[] = {
        &format_attr_event.attr,
        &format_attr_umask.attr,
@@ -2891,6 +2894,11 @@ static struct attribute *intel_arch3_formats_attr[] = {
        NULL,
 };
 
+static struct attribute *skl_format_attr[] = {
+       &format_attr_frontend.attr,
+       NULL,
+};
+
 static __initconst const struct x86_pmu core_pmu = {
        .name                   = "core",
        .handle_irq             = x86_pmu_handle_irq,
@@ -3500,7 +3508,8 @@ __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.format_attrs = merge_attr(intel_arch3_formats_attr,
+                                                 skl_format_attr);
                WARN_ON(!x86_pmu.format_attrs);
                x86_pmu.cpu_events = hsw_events_attrs;
                pr_cont("Skylake events, ");
--

2.4.2

#6 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 4 lines

I believe this mask of 0x3fff17 is wrong and should instead be 0x7fffff based on the description of the FRONTEND
MSR I see in the SDM Table 18-54 (bit 0-22 are valid). Otherwise, some valid latency values may be rejected.

No, my mask is correct.

-Andi

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

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 6 lines · Andi Kleen
Quoted 4 lines

I believe this mask of 0x3fff17 is wrong and should instead be 0x7fffff based on the description of the FRONTEND
MSR I see in the SDM Table 18-54 (bit 0-22 are valid). Otherwise, some valid latency values may be rejected.

No, my mask is correct.

Provide a proper argument for that. Just claiming 'my mask is correct' definitely falls not into that category.

Thanks,

tglx

#8 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 11 lines · Thomas Gleixner
Quoted 6 lines · Andi Kleen
Quoted 4 lines

I believe this mask of 0x3fff17 is wrong and should instead be 0x7fffff based on the description of the FRONTEND
MSR I see in the SDM Table 18-54 (bit 0-22 are valid). Otherwise, some valid latency values may be rejected.

No, my mask is correct.

Provide a proper argument for that. Just claiming 'my mask is correct' definitely falls not into that category.

Because I actually tested the code unlike you or Stephane.

# wrmsr 0x3f7 0x3fff17
# wrmsr 0x3f7 0x7fffff
wrmsr: CPU 0 cannot set MSR 0x000003f7 to 0x00000000007fffff
# 

-Andi

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

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 13 lines · Andi Kleen
Quoted 11 lines · Thomas Gleixner
Quoted 6 lines · Andi Kleen
Quoted 4 lines

I believe this mask of 0x3fff17 is wrong and should instead be 0x7fffff based on the description of the FRONTEND
MSR I see in the SDM Table 18-54 (bit 0-22 are valid). Otherwise, some valid latency values may be rejected.

No, my mask is correct.

Provide a proper argument for that. Just claiming 'my mask is correct' definitely falls not into that category.

Quoted 6 lines

Because I actually tested the code unlike you or Stephane.

# wrmsr 0x3f7 0x3fff17
# wrmsr 0x3f7 0x7fffff
wrmsr: CPU 0 cannot set MSR 0x000003f7 to 0x00000000007fffff
#

Fair enough!
But then, the SDM is misleading. It is not describing what's implemented for SKL.

#10 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 2 lines

But then, the SDM is misleading. It is not describing what's implemented for SKL.

Actually it has a list of valid values you can put into the various fields. None of them have the bits set you're trying to set.

-Andi

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

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Andi,

Quoted 6 lines · Andi Kleen
Quoted 2 lines

But then, the SDM is misleading. It is not describing what's implemented for SKL.

Actually it has a list of valid values you can put into the various fields. None of them have the bits set you're trying to set.

You are talking about the events (bit 0-7). I am talking about the bubble thresholds. I am okay with the event list for bits 0-7.

Quoted 4 lines

-Andi

--
[email protected] -- Speaking for myself only

#12 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 11 lines · Stephane Eranian

Andi,

Quoted 6 lines · Andi Kleen
Quoted 2 lines

But then, the SDM is misleading. It is not describing what's implemented for SKL.

Actually it has a list of valid values you can put into the various fields. None of them have the bits set you're trying to set.

You are talking about the events (bit 0-7). I am talking about the bubble thresholds. I am okay with the event list for bits 0-7.

Fair enough. There's a one-off in the MSR table and table 18-54. The IDQ bubble width is only 21:20. I'll ask for that to be fixed in both places that document them.

Thanks.

-Andi

#13 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 10 lines · Andi Kleen
Quoted 9 lines · Stephane Eranian

Andi,

Quoted 6 lines · Andi Kleen
Quoted 2 lines

But then, the SDM is misleading. It is not describing what's implemented for SKL.

Actually it has a list of valid values you can put into the various fields. None of them have the bits set you're trying to set.

Quoted 7 lines
Quoted 2 lines

You are talking about the events (bit 0-7). I am talking about the bubble thresholds. I am okay with the event list for bits 0-7.

Fair enough. There's a one-off in the MSR table and table 18-54. The IDQ bubble width is only 21:20. I'll ask for that to be fixed in both places that document them.

There is still something broken here, if bit 22 is not implemented, then you have
a bunch of frontend events in the SKL event table (download.01.org) which are bogus:

{
    "EventCode": "0xC6",
    "UMask": "0x01",
    "EventName": "FRONTEND_RETIRED.LATENCY_GE_8",
    "BriefDescription": "Retired instructions that are fetched after

an interval where the front-end delivered no uops for a period of 8 cycles which was not interrupted by a back-end stall.",
"PublicDescription": "Retired instructions that are fetched after
an interval where the front-end delivered no uops for a period of 8 cycles which was not interrupted by a back-end stall.",

    "Counter": "0,1,2,3",
    "CounterHTOff": "0,1,2,3",
    "SampleAfterValue": "100007",
    "MSRIndex": "0x3F7",
    "MSRValue": "0x400806",   <=========
    ....
  },
  {
    "EventCode": "0xC6",
    "UMask": "0x01",
    "EventName": "FRONTEND_RETIRED.LATENCY_GE_16",
    "BriefDescription": "Retired instructions that are fetched after

an interval where the front-end delivered no uops for a period of 16 cycles which was not interrupted by a back-end stall.",
"PublicDescription": "Retired instructions that are fetched after
an interval where the front-end delivered no uops for a period of 16 cycles which was not interrupted by a back-end stall.",

    "Counter": "0,1,2,3",
    "CounterHTOff": "0,1,2,3",
    "SampleAfterValue": "100007",
    "MSRIndex": "0x3F7",
    "MSRValue": "0x401006", <=====
    ...
     },
  {
    "EventCode": "0xC6",
    "UMask": "0x01",
    "EventName": "FRONTEND_RETIRED.LATENCY_GE_32",
    "BriefDescription": "Retired instructions that are fetched after

an interval where the front-end delivered no uops for a period of 32 cycles which was not interrupted by a back-end stall.",
"PublicDescription": "Retired instructions that are fetched after
an interval where the front-end delivered no uops for a period of 32 cycles which was not interrupted by a back-end stall.",

    "Counter": "0,1,2,3",
    "CounterHTOff": "0,1,2,3",
    "SampleAfterValue": "100007",
    "MSRIndex": "0x3F7",
    "MSRValue": "0x402006", <=========
    ...

},

#14 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 14 lines · Stephane Eranian
Quoted 13 lines · Andi Kleen
Quoted 11 lines · Stephane Eranian

Andi,

Quoted 6 lines · Andi Kleen
Quoted 2 lines

But then, the SDM is misleading. It is not describing what's implemented for SKL.

Actually it has a list of valid values you can put into the various fields. None of them have the bits set you're trying to set.

You are talking about the events (bit 0-7). I am talking about the bubble thresholds. I am okay with the event list for bits 0-7.

Quoted 8 lines
Quoted 4 lines

Fair enough. There's a one-off in the MSR table and table 18-54. The IDQ bubble width is only 21:20. I'll ask for that to be fixed in both places that document them.

There is still something broken here, if bit 22 is not implemented, then you have
a bunch of frontend events in the SKL event table (download.01.org) which are bogus:

I double checked and bit 22 is actually implemented. I'll send a patch to update the mask in the driver. Thanks for the report.

-Andi

#15 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

In reply to Andi Kleen
Quoted 8 lines · Andi Kleen
Quoted 7 lines · Thomas Gleixner
Quoted 5 lines · Andi Kleen
Quoted 4 lines

I believe this mask of 0x3fff17 is wrong and should instead be 0x7fffff based on the description of the FRONTEND
MSR I see in the SDM Table 18-54 (bit 0-22 are valid). Otherwise, some valid latency values may be rejected.

Quoted 6 lines
Quoted 4 lines
Quoted 1 line

No, my mask is correct.

Provide a proper argument for that. Just claiming 'my mask is correct' definitely falls not into that category.

Because I actually tested the code unlike you or Stephane.

You know fcking well that I cannot test on skylake and probably Stephane can't either.

Quoted 4 lines
# wrmsr 0x3f7 0x3fff17
# wrmsr 0x3f7 0x7fffff
wrmsr: CPU 0 cannot set MSR 0x000003f7 to 0x00000000007fffff
# 

So you think that's an explanation? No, it is NOT. It's merily an observation on some particular piece of HW to which you have access to exclusively.

Nobody asks you to reveal NDA information, but you should at least have the courtesy to provide an explanation which is understandable to others. The minimal explanation you could have provided would have been:

"On my test machine the valid mask is 0x3fff17 contrary to the SDM
which defines the valid mask as 0x7fffff"

'My mask is correct' is definitely not an explanation at all.

Thanks,

tglx

#16 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

In reply to Andi Kleen

Andi,

Quoted 7 lines · Andi Kleen
Quoted 4 lines

I believe this mask of 0x3fff17 is wrong and should instead be 0x7fffff based on the description of the FRONTEND
MSR I see in the SDM Table 18-54 (bit 0-22 are valid). Otherwise, some valid latency values may be rejected.

No, my mask is correct.

Ok, so your event mask (0x17) really only allows what's defined instead the full width of the field.

As for the IDQ_BUBBLE_WIDTH, you only allow 2 bits out of 3, so maximum bubble threshold is 3
instead of 7. I assume this is because you know that it cannot have more than 3 simultaneously then.

Would be good to explain this a bit more in the code.

Quoted 2 lines

--
[email protected] -- Speaking for myself only

author
#17 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 13 lines · Stephane Eranian

Andi,

Quoted 7 lines · Andi Kleen
Quoted 4 lines

I believe this mask of 0x3fff17 is wrong and should instead be 0x7fffff based on the description of the FRONTEND
MSR I see in the SDM Table 18-54 (bit 0-22 are valid). Otherwise, some valid latency values may be rejected.

No, my mask is correct.

Ok, so your event mask (0x17) really only allows what's defined instead the full width of the field.

Quoted 6 lines

As for the IDQ_BUBBLE_WIDTH, you only allow 2 bits out of 3, so maximum bubble threshold is 3
instead of 7. I assume this is because you know that it cannot have more than 3 simultaneously then.

Would be good to explain this a bit more in the code.

I don't think the code is the right place to document such registers. I can ask for the SDM to be clarified. But if you follow the documented events you will never set any undefined bits, so you won't need to care about this. We just have to exclude it at the kernel level to avoid #GP crashes if someone is writing random values.

-Andi

maintainer
#18 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 2 lines · Andi Kleen

I don't think the code is the right place to document such registers.

If the code deviates from the SDM, it _is_ the right place to explain, seeing how its the only place.

#19 lore raw

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 6 lines · Peter Zijlstra
Quoted 2 lines · Andi Kleen

I don't think the code is the right place to document such registers.

If the code deviates from the SDM, it _is_ the right place to explain, seeing how its the only place.

It won't anymore in the next revision.
-Andi

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

Re: [PATCH 3/3] x86, perf: Add PEBS frontend profiling for Skylake

Quoted 9 lines · Andi Kleen
Quoted 6 lines · Peter Zijlstra
Quoted 2 lines · Andi Kleen

I don't think the code is the right place to document such registers.

If the code deviates from the SDM, it _is_ the right place to explain, seeing how its the only place.

It won't anymore in the next revision.

And that justifies to leave completely undocumented crap in the code for the current revision?

Quoted 1 line

[email protected] -- Speaking for myself only

I'm sure Intel appreciates that disclaimer.

Thanks,

tglx

#21 lore raw

[tip:perf/core] perf/x86: Make merge_attr() global to use from perf_event_intel

In reply to Andi Kleen
perf/x86: Make merge_attr() global to use from perf_event_intel
Commit-ID:  47732d886385af769449022a02c7cf0ce45d8a5c
Gitweb:     http://git.kernel.org/tip/47732d886385af769449022a02c7cf0ce45d8a5c
Author:     Andi Kleen <[email protected]>
AuthorDate: Mon, 29 Jun 2015 14:22:13 -0700
Committer:  Ingo Molnar <[email protected]>
CommitDate: Tue, 4 Aug 2015 10:16:59 +0200

perf/x86: Make merge_attr() global to use from perf_event_intel

merge_attr() allows to merge two sysfs attribute tables. Export it to be usable by other files too.

Next patch is going to use that to extend the sysfs format attributes for a CPU.

---
 arch/x86/kernel/cpu/perf_event.c | 2 +-
 arch/x86/kernel/cpu/perf_event.h | 3 +++
 2 files changed, 4 insertions(+), 1 deletion(-)
arch/x86/kernel/cpu/perf_event.c +1 -1
index 3658de4..8bac4bb 100644
@@ -1551,7 +1551,7 @@ static void __init filter_events(struct attribute **attrs)
 }
 
 /* Merge two pointer arrays */
-static __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
+__init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
 {
 	struct attribute **new;
 	int j, i;
arch/x86/kernel/cpu/perf_event.h +3 -0
index 8ad9241..5edf6d8 100644
@@ -795,6 +795,8 @@ static inline void set_linear_ip(struct pt_regs *regs, unsigned long ip)
 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event);
 ssize_t intel_event_sysfs_show(char *page, u64 config);
 
+struct attribute **merge_attr(struct attribute **a, struct attribute **b);
+
 #ifdef CONFIG_CPU_SUP_AMD
 
 int amd_pmu_init(void);
@@ -926,6 +928,7 @@ static inline int is_ht_workaround_enabled(void)
 {
 	return !!(x86_pmu.flags & PMU_FL_EXCL_ENABLED);
 }
+
 #else /* CONFIG_CPU_SUP_INTEL */
 
 static inline void reserve_ds_buffers(void)