Skip to contentSkip to conversation

perf,x86,p6: Add userspace RDPMC quirk for P6

PATCH In review Peter Zijlstra 3 participants 14 messages Feb 5, 2014 – Feb 7, 2014

Subsystems (2)
lore
Share & tools
lore

Peter Zijlstra

author maintainer
#0 lore raw

PATCH 1/1 perf,x86,p6: Add userspace RDPMC quirk for P6

Subject: perf,x86,p6: Add userspace RDPMC quirk for P6
From: Peter Zijlstra <[email protected]>
Date: Wed Feb 5 11:17:46 CET 2014

P6 class machines can die hard when PCE gets enabled due to CPU errata. The safe way it so disable it by default and keep it disabled.

See errata 26:

http://download.intel.com/design/archives/processors/pro/docs/24268935.pdf

---
 arch/x86/kernel/cpu/perf_event.c    |    6 +++++-
 arch/x86/kernel/cpu/perf_event.h    |    1 +
 arch/x86/kernel/cpu/perf_event_p6.c |    8 ++++++++
 3 files changed, 14 insertions(+), 1 deletion(-)
arch/x86/kernel/cpu/perf_event.c +5 -1
@@ -1521,6 +1521,8 @@ static int __init init_hw_perf_events(vo
 
 	pr_cont("%s PMU driver.\n", x86_pmu.name);
 
+	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
+
 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
 		quirk->func();
 
@@ -1534,7 +1536,6 @@ static int __init init_hw_perf_events(vo
 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
 				   0, x86_pmu.num_counters, 0, 0);
 
-	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
 
 	if (x86_pmu.event_attrs)
@@ -1820,6 +1821,9 @@ static ssize_t set_attr_rdpmc(struct dev
 	if (ret)
 		return ret;
 
+	if (x86_pmu.attr_rdpmc_broken)
+		return -ENOTSUPP;
+
 	if (!!val != !!x86_pmu.attr_rdpmc) {
 		x86_pmu.attr_rdpmc = !!val;
 		smp_call_function(change_rdpmc, (void *)val, 1);
arch/x86/kernel/cpu/perf_event.h +1 -0
@@ -409,6 +409,7 @@ struct x86_pmu {
 	/*
 	 * sysfs attrs
 	 */
+	int		attr_rdpmc_broken;
 	int		attr_rdpmc;
 	struct attribute **format_attrs;
 	struct attribute **event_attrs;
arch/x86/kernel/cpu/perf_event_p6.c +8 -0
@@ -231,6 +231,13 @@ static __initconst const struct x86_pmu
 
 };
 
+static __init void p6_pmu_rdpmc_quirk(void)
+{
+	pr_warn("Userspace RDPMC support disabled due to CPU errata\n");
+	x86_pmu.attr_rdpmc_broken = 1;
+	x86_pmu.attr_rdpmc = 0;
+}
+
 __init int p6_pmu_init(void)
 {
 	switch (boot_cpu_data.x86_model) {
@@ -256,6 +263,7 @@ __init int p6_pmu_init(void)
 	memcpy(hw_cache_event_ids, p6_hw_cache_event_ids,
 		sizeof(hw_cache_event_ids));
 
+	x86_add_quirk(p6_pmu_rdpmc_quirk);
 
 	return 0;
 }
#1 lore raw
Quoted 4 lines · Peter Zijlstra

P6 class machines can die hard when PCE gets enabled due to CPU errata. The safe way it so disable it by default and keep it disabled.

See errata 26:

so you take an errata applicable to a subset of Pentium Pro steppings and apply it to the whole p6 line (which includes the whole way up to Pentium III)?

Or am I missing something in the documentation you link to?

Vince

author maintainer
#2 lore raw
Quoted 5 lines · Vince Weaver
Quoted 3 lines · Peter Zijlstra

P6 class machines can die hard when PCE gets enabled due to CPU errata. The safe way it so disable it by default and keep it disabled.

Quoted 7 lines
Quoted 1 line

See errata 26:

so you take an errata applicable to a subset of Pentium Pro steppings and apply it to the whole p6 line (which includes the whole way up to Pentium III)?

Or am I missing something in the documentation you link to?

I forgot it went all the way to P-III, let me try and find the other erratas.

#3 lore raw
Quoted 8 lines · Peter Zijlstra
Quoted 7 lines · Vince Weaver
Quoted 4 lines · Peter Zijlstra

P6 class machines can die hard when PCE gets enabled due to CPU errata. The safe way it so disable it by default and keep it disabled.

See errata 26:

Quoted 8 lines
Quoted 5 lines

so you take an errata applicable to a subset of Pentium Pro steppings and apply it to the whole p6 line (which includes the whole way up to Pentium III)?

Or am I missing something in the documentation you link to?

I forgot it went all the way to P-III, let me try and find the other erratas.

from the document you linked it looks like it was fixed by the Pentium Pro stepping sB1 (6.1.9).

Has someone actually hit this bug? People used PAPI/perfctr on these machines for years, and it uses rdpmc by default so I would think it would show up if it were an issue. I should check the perfctr code to see if it had some sort of check.

Vince

#4 lore raw
Quoted 4 lines · Vince Weaver

Has someone actually hit this bug? People used PAPI/perfctr on these machines for years, and it uses rdpmc by default so I would think it would show up if it were an issue. I should check the perfctr code to see if it had some sort of check.

I can verify the perfctr code does have a case for PPRO errata 26, but it only disables things on pentium pros with stepping less than 9.

It also disables rdpmc on original Pentiums due to errata 74 but I don't think we have to worry about that one.

Vince

author maintainer
#5 lore raw
Quoted 9 lines · Vince Weaver
Quoted 4 lines · Vince Weaver

Has someone actually hit this bug? People used PAPI/perfctr on these machines for years, and it uses rdpmc by default so I would think it would show up if it were an issue. I should check the perfctr code to see if it had some sort of check.

I can verify the perfctr code does have a case for PPRO errata 26, but it only disables things on pentium pros with stepping less than 9.

The doc said 0xB1, right? Should we keep it at that? Its not like there's a lot of PPro machines left out there, so better safe than sorry.

Quoted 2 lines

It also disables rdpmc on original Pentiums due to errata 74 but I don't think we have to worry about that one.

We don't have a Family 5 driver, so no.

#6 lore raw
Quoted 8 lines · Peter Zijlstra
Quoted 7 lines · Vince Weaver
Quoted 4 lines · Vince Weaver

Has someone actually hit this bug? People used PAPI/perfctr on these machines for years, and it uses rdpmc by default so I would think it would show up if it were an issue. I should check the perfctr code to see if it had some sort of check.

Quoted 6 lines
Quoted 2 lines

I can verify the perfctr code does have a case for PPRO errata 26, but it only disables things on pentium pros with stepping less than 9.

The doc said 0xB1, right? Should we keep it at that? Its not like there's a lot of PPro machines left out there, so better safe than sorry.

mfg steping "sB1" which means processor stepping 9. See page 15 of the document linked in the original patch.

I used to have a Pentium Pro that was apparently affected by this erratum but not any more. The oldest machine I test perf stuff on is a Pentium II.

Vince

author maintainer
#7 lore raw
Quoted 2 lines · Vince Weaver

mfg steping "sB1" which means processor stepping 9. See page 15 of the document linked in the original patch.

Might confusing way to write 9. ok, I'll make it that.

author maintainer
#8 lore raw
In reply to Vince Weaver
Quoted 2 lines · Vince Weaver

from the document you linked it looks like it was fixed by the Pentium Pro stepping sB1 (6.1.9).

OK, I'll add a boot_cpu_data.x86_mask >= 0xB1 test.

Quoted 4 lines

Has someone actually hit this bug? People used PAPI/perfctr on these machines for years, and it uses rdpmc by default so I would think it would show up if it were an issue. I should check the perfctr code to see if it had some sort of check.

Yeah, the reporter had a boot fail due to this.

author maintainer
#9 lore raw
In reply to Peter Zijlstra
Quoted 8 lines · Peter Zijlstra
Quoted 7 lines · Vince Weaver
Quoted 4 lines · Peter Zijlstra

P6 class machines can die hard when PCE gets enabled due to CPU errata. The safe way it so disable it by default and keep it disabled.

See errata 26:

Quoted 8 lines
Quoted 5 lines

so you take an errata applicable to a subset of Pentium Pro steppings and apply it to the whole p6 line (which includes the whole way up to Pentium III)?

Or am I missing something in the documentation you link to?

I forgot it went all the way to P-III, let me try and find the other erratas.

OK, Just checked the P-II errata and its not there anymore, so I suppose only Family 6 model 1 should have this quirk.

Updated patch below.

---
Subject: perf,x86,p6: Add userspace RDPMC quirk for PPro
From: Peter Zijlstra <[email protected]>
Date: Wed Feb 5 11:17:46 CET 2014

PPro machines can die hard when PCE gets enabled due to CPU errata. The safe way it so disable it by default and keep it disabled.

See errata 26:

http://download.intel.com/design/archives/processors/pro/docs/24268935.pdf

---
 arch/x86/kernel/cpu/perf_event.c    |    6 +++++-
 arch/x86/kernel/cpu/perf_event.h    |    1 +
 arch/x86/kernel/cpu/perf_event_p6.c |   35 +++++++++++++++++++++++++----------
 3 files changed, 31 insertions(+), 11 deletions(-)
arch/x86/kernel/cpu/perf_event.c +5 -1
@@ -1521,6 +1521,8 @@ static int __init init_hw_perf_events(vo
 
 	pr_cont("%s PMU driver.\n", x86_pmu.name);
 
+	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
+
 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
 		quirk->func();
 
@@ -1534,7 +1536,6 @@ static int __init init_hw_perf_events(vo
 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
 				   0, x86_pmu.num_counters, 0, 0);
 
-	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
 
 	if (x86_pmu.event_attrs)
@@ -1820,6 +1821,9 @@ static ssize_t set_attr_rdpmc(struct dev
 	if (ret)
 		return ret;
 
+	if (x86_pmu.attr_rdpmc_broken)
+		return -ENOTSUPP;
+
 	if (!!val != !!x86_pmu.attr_rdpmc) {
 		x86_pmu.attr_rdpmc = !!val;
 		smp_call_function(change_rdpmc, (void *)val, 1);
arch/x86/kernel/cpu/perf_event.h +1 -0
@@ -409,6 +409,7 @@ struct x86_pmu {
 	/*
 	 * sysfs attrs
 	 */
+	int		attr_rdpmc_broken;
 	int		attr_rdpmc;
 	struct attribute **format_attrs;
 	struct attribute **event_attrs;
arch/x86/kernel/cpu/perf_event_p6.c +25 -10
@@ -231,20 +231,35 @@ static __initconst const struct x86_pmu
 
 };
 
+static __init void p6_pmu_rdpmc_quirk(void)
+{
+	pr_warn("Userspace RDPMC support disabled due to CPU errata\n");
+	x86_pmu.attr_rdpmc_broken = 1;
+	x86_pmu.attr_rdpmc = 0;
+}
+
 __init int p6_pmu_init(void)
 {
 	switch (boot_cpu_data.x86_model) {
-	case 1:
-	case 3:  /* Pentium Pro */
-	case 5:
-	case 6:  /* Pentium II */
-	case 7:
-	case 8:
-	case 11: /* Pentium III */
-	case 9:
-	case 13:
-		/* Pentium M */
+	case  1: /* Pentium Pro */
+		x86_add_quirk(p6_pmu_rdpmc_quirk);
 		break;
+
+	case  3: /* Pentium II - Klamath */
+	case  5: /* Pentium II - Deschutes */
+	case  6: /* Pentium II - Mendocino */
+		break;
+
+	case  7: /* Pentium III - Katmai */
+	case  8: /* Pentium III - Coppermine */
+	case 10: /* Pentium III Xeon */
+	case 11: /* Pentium III - Tualatin */
+		break;
+
+	case  9: /* Pentium M - Banias */
+	case 13: /* Pentium M - Dothan */
+		break;
+
 	default:
 		pr_cont("unsupported p6 CPU model %d ",
 			boot_cpu_data.x86_model);
#10 lore raw

Peter

As I reported the problem I thought I better contribute by testing the fix.

I took the patch below and managed to change the 3.4.78 code to match and tried to boot my Pentium Pro system but it hung as before. After some investigation I discovered the quirk is not being run, in fact no quirks were being run in init_hw_perf_events as x86_pmu.quirks was NULL.

The problem as far as I can see is that the quirk is added to x86_pmu in p6_pmu_init but then x86_pmu is reassigned at the end of that function (x86_pmu = p6_pmu) resetting x86_pmu.quirks to NULL. x86_pmu looks like it is reassigned again in intel_pmu_init to either core_pmu or intel_pmu after the call to p6_pmu_init so I am not sure which x86_pmu.quirks init_hw_perf_events tries to use.

Hopefully that makes sense.

Mark Davies

Quoted 132 lines · Peter Zijlstra
Quoted 8 lines · Peter Zijlstra
Quoted 7 lines · Vince Weaver
Quoted 4 lines · Peter Zijlstra

P6 class machines can die hard when PCE gets enabled due to CPU errata. The safe way it so disable it by default and keep it disabled.

See errata 26:

Quoted 8 lines
Quoted 5 lines

so you take an errata applicable to a subset of Pentium Pro steppings and apply it to the whole p6 line (which includes the whole way up to Pentium III)?

Or am I missing something in the documentation you link to?

I forgot it went all the way to P-III, let me try and find the other erratas.

OK, Just checked the P-II errata and its not there anymore, so I suppose only Family 6 model 1 should have this quirk.

Updated patch below.

---
Subject: perf,x86,p6: Add userspace RDPMC quirk for PPro
From: Peter Zijlstra <[email protected]>
Date: Wed Feb 5 11:17:46 CET 2014

PPro machines can die hard when PCE gets enabled due to CPU errata. The safe way it so disable it by default and keep it disabled.

See errata 26:

http://download.intel.com/design/archives/processors/pro/docs/24268935.pdf

---
 arch/x86/kernel/cpu/perf_event.c    |    6 +++++-
 arch/x86/kernel/cpu/perf_event.h    |    1 +
 arch/x86/kernel/cpu/perf_event_p6.c |   35 +++++++++++++++++++++++++----------
 3 files changed, 31 insertions(+), 11 deletions(-)
arch/x86/kernel/cpu/perf_event.c +5 -1
@@ -1521,6 +1521,8 @@ static int __init init_hw_perf_events(vo
 
 	pr_cont("%s PMU driver.\n", x86_pmu.name);
 
+	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
+
 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
 		quirk->func();
 
@@ -1534,7 +1536,6 @@ static int __init init_hw_perf_events(vo
 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
 				   0, x86_pmu.num_counters, 0, 0);
 
-	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
 
 	if (x86_pmu.event_attrs)
@@ -1820,6 +1821,9 @@ static ssize_t set_attr_rdpmc(struct dev
 	if (ret)
 		return ret;
 
+	if (x86_pmu.attr_rdpmc_broken)
+		return -ENOTSUPP;
+
 	if (!!val != !!x86_pmu.attr_rdpmc) {
 		x86_pmu.attr_rdpmc = !!val;
 		smp_call_function(change_rdpmc, (void *)val, 1);
arch/x86/kernel/cpu/perf_event.h +1 -0
@@ -409,6 +409,7 @@ struct x86_pmu {
 	/*
 	 * sysfs attrs
 	 */
+	int		attr_rdpmc_broken;
 	int		attr_rdpmc;
 	struct attribute **format_attrs;
 	struct attribute **event_attrs;
arch/x86/kernel/cpu/perf_event_p6.c +25 -10
@@ -231,20 +231,35 @@ static __initconst const struct x86_pmu
 
 };
 
+static __init void p6_pmu_rdpmc_quirk(void)
+{
+	pr_warn("Userspace RDPMC support disabled due to CPU errata\n");
+	x86_pmu.attr_rdpmc_broken = 1;
+	x86_pmu.attr_rdpmc = 0;
+}
+
 __init int p6_pmu_init(void)
 {
 	switch (boot_cpu_data.x86_model) {
-	case 1:
-	case 3:  /* Pentium Pro */
-	case 5:
-	case 6:  /* Pentium II */
-	case 7:
-	case 8:
-	case 11: /* Pentium III */
-	case 9:
-	case 13:
-		/* Pentium M */
+	case  1: /* Pentium Pro */
+		x86_add_quirk(p6_pmu_rdpmc_quirk);
 		break;
+
+	case  3: /* Pentium II - Klamath */
+	case  5: /* Pentium II - Deschutes */
+	case  6: /* Pentium II - Mendocino */
+		break;
+
+	case  7: /* Pentium III - Katmai */
+	case  8: /* Pentium III - Coppermine */
+	case 10: /* Pentium III Xeon */
+	case 11: /* Pentium III - Tualatin */
+		break;
+
+	case  9: /* Pentium M - Banias */
+	case 13: /* Pentium M - Dothan */
+		break;
+
 	default:
 		pr_cont("unsupported p6 CPU model %d ",
 			boot_cpu_data.x86_model);
author maintainer
#11 lore raw
Quoted 5 lines · Mark Davies

Peter

As I reported the problem I thought I better contribute by testing the fix.

Quoted 8 lines

I took the patch below and managed to change the 3.4.78 code to match and tried to boot my Pentium Pro system but it hung as before. After some investigation I discovered the quirk is not being run, in fact no quirks were being run in init_hw_perf_events as x86_pmu.quirks was NULL.

The problem as far as I can see is that the quirk is added to x86_pmu in p6_pmu_init but then x86_pmu is reassigned at the end of that function (x86_pmu = p6_pmu) resetting x86_pmu.quirks to NULL.

Right so far.

Quoted 4 lines

x86_pmu looks like
it is reassigned again in intel_pmu_init to either core_pmu or intel_pmu after the call to p6_pmu_init so I am not sure which x86_pmu.quirks init_hw_perf_events tries to use.

Note how intel_pmu_init() does: return p6_pmu_init(), so it will never get to the x86_pmu assignment for the ARCH_PERFMON drivers.

Yep, thanks for testing and you're (mostly) quite right.

We can do the copy early though, if we return -ENODEV nothing will end up using the x86_pmu anyway, so it doesn't matter what lives in there.

---
Subject: perf,x86,p6: Add userspace RDPMC quirk for PPro
From: Peter Zijlstra <[email protected]>
Date: Wed, 5 Feb 2014 20:48:51 +0100

PPro machines can die hard when PCE gets enabled due to a CPU erratum. The safe way it so disable it by default and keep it disabled.

See erratum 26 in:

http://download.intel.com/design/archives/processors/pro/docs/24268935.pdf

Cc Alan Cox <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Stephane Eranian <[email protected]>
Cc Vince Weaver <[email protected]>
---
 arch/x86/kernel/cpu/perf_event.c    |    6 +++-
 arch/x86/kernel/cpu/perf_event.h    |    1 
 arch/x86/kernel/cpu/perf_event_p6.c |   48 ++++++++++++++++++++++++------------
 3 files changed, 39 insertions(+), 16 deletions(-)
arch/x86/kernel/cpu/perf_event.c +5 -1
@@ -1521,6 +1521,8 @@ static int __init init_hw_perf_events(vo
 
 	pr_cont("%s PMU driver.\n", x86_pmu.name);
 
+	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
+
 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
 		quirk->func();
 
@@ -1534,7 +1536,6 @@ static int __init init_hw_perf_events(vo
 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
 				   0, x86_pmu.num_counters, 0, 0);
 
-	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
 
 	if (x86_pmu.event_attrs)
@@ -1820,6 +1821,9 @@ static ssize_t set_attr_rdpmc(struct dev
 	if (ret)
 		return ret;
 
+	if (x86_pmu.attr_rdpmc_broken)
+		return -ENOTSUPP;
+
 	if (!!val != !!x86_pmu.attr_rdpmc) {
 		x86_pmu.attr_rdpmc = !!val;
 		smp_call_function(change_rdpmc, (void *)val, 1);
arch/x86/kernel/cpu/perf_event.h +1 -0
@@ -409,6 +409,7 @@ struct x86_pmu {
 	/*
 	 * sysfs attrs
 	 */
+	int		attr_rdpmc_broken;
 	int		attr_rdpmc;
 	struct attribute **format_attrs;
 	struct attribute **event_attrs;
arch/x86/kernel/cpu/perf_event_p6.c +33 -15

64 lines not on this page · raw message

#12 lore raw

That fixed it. Booted patched 3.4.78 with CONFIG_APM=y with no problem.

dmesg contains the "Userspace RDPMC support disabled due to a CPU erratum" message

"cat /sys/devices/cpu/rdpmc" gives "0"

"echo 1 >/sys/devices/cpu/rdpmc" gives "bash: echo: write error: Unknown error 524"

Thanks for fixing it. How I missed the return statements in intel_pmu_init I don't know, probably got a bit over excited when I spotted what the problem was.

Mark Davies

Quoted 160 lines · Peter Zijlstra
Quoted 5 lines · Mark Davies

Peter

As I reported the problem I thought I better contribute by testing the fix.

Quoted 8 lines

I took the patch below and managed to change the 3.4.78 code to match and tried to boot my Pentium Pro system but it hung as before. After some investigation I discovered the quirk is not being run, in fact no quirks were being run in init_hw_perf_events as x86_pmu.quirks was NULL.

The problem as far as I can see is that the quirk is added to x86_pmu in p6_pmu_init but then x86_pmu is reassigned at the end of that function (x86_pmu = p6_pmu) resetting x86_pmu.quirks to NULL.

Right so far.

Quoted 4 lines

x86_pmu looks like
it is reassigned again in intel_pmu_init to either core_pmu or intel_pmu after the call to p6_pmu_init so I am not sure which x86_pmu.quirks init_hw_perf_events tries to use.

Note how intel_pmu_init() does: return p6_pmu_init(), so it will never get to the x86_pmu assignment for the ARCH_PERFMON drivers.

Yep, thanks for testing and you're (mostly) quite right.

We can do the copy early though, if we return -ENODEV nothing will end up using the x86_pmu anyway, so it doesn't matter what lives in there.

---
Subject: perf,x86,p6: Add userspace RDPMC quirk for PPro
From: Peter Zijlstra <[email protected]>
Date: Wed, 5 Feb 2014 20:48:51 +0100

PPro machines can die hard when PCE gets enabled due to a CPU erratum. The safe way it so disable it by default and keep it disabled.

See erratum 26 in:

http://download.intel.com/design/archives/processors/pro/docs/24268935.pdf

Cc Alan Cox <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Stephane Eranian <[email protected]>
Cc Vince Weaver <[email protected]>
---
 arch/x86/kernel/cpu/perf_event.c    |    6 +++-
 arch/x86/kernel/cpu/perf_event.h    |    1 
 arch/x86/kernel/cpu/perf_event_p6.c |   48 ++++++++++++++++++++++++------------
 3 files changed, 39 insertions(+), 16 deletions(-)
arch/x86/kernel/cpu/perf_event.c +5 -1
@@ -1521,6 +1521,8 @@ static int __init init_hw_perf_events(vo
 
 	pr_cont("%s PMU driver.\n", x86_pmu.name);
 
+	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
+
 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
 		quirk->func();
 
@@ -1534,7 +1536,6 @@ static int __init init_hw_perf_events(vo
 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
 				   0, x86_pmu.num_counters, 0, 0);
 
-	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
 
 	if (x86_pmu.event_attrs)
@@ -1820,6 +1821,9 @@ static ssize_t set_attr_rdpmc(struct dev
 	if (ret)
 		return ret;
 
+	if (x86_pmu.attr_rdpmc_broken)
+		return -ENOTSUPP;
+
 	if (!!val != !!x86_pmu.attr_rdpmc) {
 		x86_pmu.attr_rdpmc = !!val;
 		smp_call_function(change_rdpmc, (void *)val, 1);
arch/x86/kernel/cpu/perf_event.h +1 -0
@@ -409,6 +409,7 @@ struct x86_pmu {
 	/*
 	 * sysfs attrs
 	 */
+	int		attr_rdpmc_broken;
 	int		attr_rdpmc;
 	struct attribute **format_attrs;
 	struct attribute **event_attrs;
arch/x86/kernel/cpu/perf_event_p6.c +33 -15
@@ -231,31 +231,49 @@ static __initconst const struct x86_pmu
 
 };
 
+static __init void p6_pmu_rdpmc_quirk(void)
+{
+	if (boot_cpu_data.x86_mask < 9) {
+		/*
+		 * PPro erratum 26; fixed in stepping 9 and above.
+		 */
+		pr_warn("Userspace RDPMC support disabled due to a CPU erratum\n");
+		x86_pmu.attr_rdpmc_broken = 1;
+		x86_pmu.attr_rdpmc = 0;
+	}
+}
+
 __init int p6_pmu_init(void)
 {
+	x86_pmu = p6_pmu;
+
 	switch (boot_cpu_data.x86_model) {
-	case 1:
-	case 3:  /* Pentium Pro */
-	case 5:
-	case 6:  /* Pentium II */
-	case 7:
-	case 8:
-	case 11: /* Pentium III */
-	case 9:
-	case 13:
-		/* Pentium M */
+	case  1: /* Pentium Pro */
+		x86_add_quirk(p6_pmu_rdpmc_quirk);
+		break;
+
+	case  3: /* Pentium II - Klamath */
+	case  5: /* Pentium II - Deschutes */
+	case  6: /* Pentium II - Mendocino */
 		break;
+
+	case  7: /* Pentium III - Katmai */
+	case  8: /* Pentium III - Coppermine */
+	case 10: /* Pentium III Xeon */
+	case 11: /* Pentium III - Tualatin */
+		break;
+
+	case  9: /* Pentium M - Banias */
+	case 13: /* Pentium M - Dothan */
+		break;
+
 	default:
-		pr_cont("unsupported p6 CPU model %d ",
-			boot_cpu_data.x86_model);
+		pr_cont("unsupported p6 CPU model %d ", boot_cpu_data.x86_model);
 		return -ENODEV;
 	}
 
-	x86_pmu = p6_pmu;
-
 	memcpy(hw_cache_event_ids, p6_hw_cache_event_ids,
 		sizeof(hw_cache_event_ids));
 
-
 	return 0;
 }
author maintainer
#13 lore raw
Quoted 1 line · Mark Davies

That fixed it. Booted patched 3.4.78 with CONFIG_APM=y with no problem.

Awesome, thanks!