Skip to contentSkip to conversation

perf, x86: Disable perf if the BIOS got its grubby paws on the PMU

PATCH In review Peter Zijlstra 4 participants 6 messages Sep 3, 2010 – Sep 7, 2010

Subsystems (2)
lore
Share & tools
lore

Peter Zijlstra

author maintainer
#0 lore raw

PATCH 1/1 perf, x86: Disable perf if the BIOS got its grubby paws on the PMU

Disable all of perf if we find any active PMCs on boot.

It has been reported that some BIOSes access the PMU, for obscure things better done by the OS. Come down hard on this practice and fully disable the PMU.

[ Boot tested on a westmere system with a sane BIOS -- I don't actually
have an affected system to test this on. ]

---

Robert, do you know if any AMD system BIOSes carry similar Feat^H^H^H^HFailureAdd?

 arch/x86/kernel/cpu/perf_event_intel.c |   35 ++++++++++++++++++++++++++++++++
 1 files changed, 35 insertions(+), 0 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +35 -0
index ee05c90..6849653 100644
@@ -923,6 +923,19 @@ static void intel_clovertown_quirks(void)
 	x86_pmu.pebs_constraints = NULL;
 }
 
+static void print_BIOS_fail(void)
+{
+	printk(KERN_ERR "\n");
+	printk(KERN_ERR "=============================================\n");
+	printk(KERN_ERR "It appears the BIOS is actively using the PMU\n");
+	printk(KERN_ERR "this avoids Linux from using it, please de-  \n");
+	printk(KERN_ERR "activate this BIOS feature or request a BIOS \n");
+	printk(KERN_ERR "update from your vendor.                     \n");
+	printk(KERN_ERR "=============================================\n");
+
+	memset(&x86_pmu, 0, sizeof(x86_pmu));
+}
+
 static __init int intel_pmu_init(void)
 {
 	union cpuid10_edx edx;
@@ -930,6 +943,8 @@ static __init int intel_pmu_init(void)
 	unsigned int unused;
 	unsigned int ebx;
 	int version;
+	u64 val;
+	int i;
 
 	if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
 		switch (boot_cpu_data.x86) {
@@ -968,6 +983,26 @@ static __init int intel_pmu_init(void)
 		x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3);
 
 	/*
+	 * Check to see if the BIOS enabled any of the counters, if so
+	 * complain and bail.
+	 */
+	for (i = 0; i < x86_pmu.num_counters; i++) {
+		rdmsrl(x86_pmu.eventsel + i, val);
+		if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
+			print_BIOS_fail();
+			return -EBUSY;
+		}
+	}
+
+	for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
+		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, val);
+		if (val & (0x03 << i*4)) {
+			print_BIOS_fail();
+			return -EBUSY;
+		}
+	}
+
+	/*
 	 * v2 and above have a perf capabilities MSR
 	 */
 	if (version > 1) {
#1 lore raw
Quoted 5 lines · Peter Zijlstra
arch/x86/kernel/cpu/perf_event_intel.c +3 -0
+static void print_BIOS_fail(void)
+{
+	printk(KERN_ERR "\n");

arch/x86/kernel/cpu/perf_event_intel.c static void intel_clovertown_quirks(void) · line 938

Quoted 10 lines
arch/x86/kernel/cpu/perf_event_intel.c +10 -0
+	printk(KERN_ERR "=============================================\n");
+	printk(KERN_ERR "It appears the BIOS is actively using the PMU\n");
+	printk(KERN_ERR "this avoids Linux from using it, please de-  \n");
+	printk(KERN_ERR "activate this BIOS feature or request a BIOS \n");
+	printk(KERN_ERR "update from your vendor.                     \n");
+	printk(KERN_ERR "=============================================\n");
+
+	memset(&x86_pmu, 0, sizeof(x86_pmu));
+}
+

tell us how you really feel :-)

do you want to add a phone number of the support line of the vendor (based on DMI data of course) ?

#2 lore raw
Quoted 20 lines · Arjan van de Ven
Quoted 5 lines · Peter Zijlstra
arch/x86/kernel/cpu/perf_event_intel.c +3 -0
+static void print_BIOS_fail(void)
+{
+	printk(KERN_ERR "\n");

arch/x86/kernel/cpu/perf_event_intel.c static void intel_clovertown_quirks(void) · line 938

Quoted 10 lines
arch/x86/kernel/cpu/perf_event_intel.c +10 -0
+	printk(KERN_ERR "=============================================\n");
+	printk(KERN_ERR "It appears the BIOS is actively using the PMU\n");
+	printk(KERN_ERR "this avoids Linux from using it, please de-  \n");
+	printk(KERN_ERR "activate this BIOS feature or request a BIOS \n");
+	printk(KERN_ERR "update from your vendor.                     \n");
+	printk(KERN_ERR "=============================================\n");
+
+	memset(&x86_pmu, 0, sizeof(x86_pmu));
+}
+

tell us how you really feel :-)

Quoted 2 lines

do you want to add a phone number of the support line of the vendor (based on DMI data of course) ?

I must object to messages that repeatedly (at least on every boot) tell system administrators to contact their hardware vendor's support lines, when it's not clear what the BIOS is doing is incorrect. There are plenty of valid reasons why BIOS itself would use PMU counters. Dell PowerEdge server power management, handled by the BIOS, certainly does use one.

My understanding is that there is a mechanism for the OS to request BIOS to release use of PMU counters. Are we doing that? If BIOS does not release the counters when asked, ok, that's something to (potentially) warn about. But blanket "BIOS is using a CPU feature! Bad BIOS! No treat for you!" - that's not helpful to anyone.

Thanks,
Matt

Signature · 3 lines
Matt Domsch
Technology Strategist
Dell | Office of the CTO
author maintainer
#3 lore raw
Quoted 19 lines · Matt Domsch
Quoted 14 lines · Arjan van de Ven

arch/x86/kernel/cpu/perf_event_intel.c static void intel_clovertown_quirks(void) · line 937

Quoted 13 lines · Peter Zijlstra
arch/x86/kernel/cpu/perf_event_intel.c +12 -0
+static void print_BIOS_fail(void)
+{
+	printk(KERN_ERR "\n");
+	printk(KERN_ERR "=============================================\n");
+	printk(KERN_ERR "It appears the BIOS is actively using the PMU\n");
+	printk(KERN_ERR "this avoids Linux from using it, please de-  \n");
+	printk(KERN_ERR "activate this BIOS feature or request a BIOS \n");
+	printk(KERN_ERR "update from your vendor.                     \n");
+	printk(KERN_ERR "=============================================\n");
+
+	memset(&x86_pmu, 0, sizeof(x86_pmu));
+}

I must object to messages that repeatedly (at least on every boot) tell system administrators to contact their hardware vendor's support lines, when it's not clear what the BIOS is doing is incorrect.

It is using the PMU, it should not _ever_ do that.

Quoted 4 lines

There
are plenty of valid reasons why BIOS itself would use PMU counters. Dell PowerEdge server power management, handled by the BIOS, certainly does use one.

Them make them stop doing that, or at least provide a BIOS option to disable this Feat^WFailure-add.

An no, doing power-management from the BIOS is most certainly not a valid reason. That's not what BIOSes are for, a BIOS should bring up the system and then sod off.

Quoted 2 lines

My understanding is that there is a mechanism for the OS to request BIOS to release use of PMU counters. Are we doing that?

I'm not aware of any such thing. The Intel Arch docs most certainly don't specify anything about that.

Quoted 4 lines

If BIOS does
not release the counters when asked, ok, that's something to (potentially) warn about. But blanket "BIOS is using a CPU feature! Bad BIOS! No treat for you!" - that's not helpful to anyone.

Again, I'm not aware there is a spec on how to ask anything of the BIOS, let alone a part pertaining to PMU functionality.

#4 lore raw
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines · Peter Zijlstra

arch/x86/kernel/cpu/perf_event_intel.c static void intel_clovertown_quirks(void) · line 934

Quoted 9 lines
arch/x86/kernel/cpu/perf_event_intel.c +9 -0
+static void print_BIOS_fail(void)
+{
+	printk(KERN_ERR "\n");
+	printk(KERN_ERR "=============================================\n");
+	printk(KERN_ERR "It appears the BIOS is actively using the PMU\n");
+	printk(KERN_ERR "this avoids Linux from using it, please de-  \n");
+	printk(KERN_ERR "activate this BIOS feature or request a BIOS \n");
+	printk(KERN_ERR "update from your vendor.                     \n");
+	printk(KERN_ERR "=============================================\n");

I would rather prefer this:

BIOS bug, cpu 1, invalid <register=value> ...

... which is a much better information on one line, explains the bug and is also better parsable. I intend to implement messages like this. So maybe we could find consensus with this or something similar.

A simple grep of dmesg will then give a list of BIOS bugs.

-Robert

Signature · 2 lines
Advanced Micro Devices, Inc.
Operating System Research Center
#5 lore raw
Quoted 13 lines · Robert Richter
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines · Peter Zijlstra

arch/x86/kernel/cpu/perf_event_intel.c static void intel_clovertown_quirks(void) · line 934

Quoted 9 lines
arch/x86/kernel/cpu/perf_event_intel.c +9 -0
+static void print_BIOS_fail(void)
+{
+	printk(KERN_ERR "\n");
+	printk(KERN_ERR "=============================================\n");
+	printk(KERN_ERR "It appears the BIOS is actively using the PMU\n");
+	printk(KERN_ERR "this avoids Linux from using it, please de-  \n");
+	printk(KERN_ERR "activate this BIOS feature or request a BIOS \n");
+	printk(KERN_ERR "update from your vendor.                     \n");
+	printk(KERN_ERR "=============================================\n");

I would rather prefer this:

Quoted 1 line

BIOS bug, cpu 1, invalid <register=value> ...

I got the note there is already a macro for this in kernel.h:

#define FW_BUG "[Firmware Bug]: "

So, we would have something like:

[Firmware Bug]: cpu 1, invalid <register=value> ...

in dmesg.

-Robert

Quoted 12 lines

... which is a much better information on one line, explains the bug and is also better parsable. I intend to implement messages like this. So maybe we could find consensus with this or something similar.

A simple grep of dmesg will then give a list of BIOS bugs.

-Robert

Signature · 2 lines
Advanced Micro Devices, Inc.
Operating System Research Center
Signature · 2 lines
Advanced Micro Devices, Inc.
Operating System Research Center