Skip to contentSkip to conversation

powerpc: Fix bug where perf_counters breaks oprofile

PATCH In review Paul Mackerras 4 participants 6 messages Sep 9, 2009 – Sep 15, 2009 PowerPC

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 powerpc: Fix bug where perf_counters breaks oprofile

Currently there is a bug where if you use oprofile on a pSeries machine, then use perf_counters, then use oprofile again, oprofile will not work correctly; it will lose the PMU configuration the next time the hypervisor does a partition context switch, and thereafter won't count anything.

Maynard Johnson identified the sequence causing the problem:
- oprofile setup calls ppc_enable_pmcs(), which calls pseries_lpar_enable_pmcs, which tells the hypervisor that we want to use the PMU, and sets the "PMU in use" flag in the lppaca. This flag tells the hypervisor whether it needs to save and restore the PMU config.
- The perf_counter code sets and clears the "PMU in use" flag directly as it context-switches the PMU between tasks, and leaves it clear when it finishes.
- oprofile setup, called for a new oprofile run, calls ppc_enable_pmcs, which does nothing because it has already been called. In particular it doesn't set the "PMU in use" flag.

This fixes the problem by arranging for ppc_enable_pmcs to always set the "PMU in use" flag. It makes the perf_counter code call ppc_enable_pmcs also rather than calling the lower-level function directly, and removes the setting of the "PMU in use" flag from pseries_lpar_enable_pmcs, since that is now done in its caller.

This also removes the declaration of pasemi_enable_pmcs because it isn't defined anywhere.

---
 arch/powerpc/include/asm/pmc.h         |   16 ++++++++++++++--
 arch/powerpc/kernel/perf_counter.c     |   13 +++----------
 arch/powerpc/kernel/sysfs.c            |    3 +++
 arch/powerpc/platforms/pseries/setup.c |    4 ----
 4 files changed, 20 insertions(+), 16 deletions(-)
arch/powerpc/include/asm/pmc.h +14 -2
index d6a616a..ccc68b5 100644
@@ -27,10 +27,22 @@ extern perf_irq_t perf_irq;
 
 int reserve_pmc_hardware(perf_irq_t new_perf_irq);
 void release_pmc_hardware(void);
+void ppc_enable_pmcs(void);
 
 #ifdef CONFIG_PPC64
-void power4_enable_pmcs(void);
-void pasemi_enable_pmcs(void);
+#include <asm/lppaca.h>
+
+static inline void ppc_set_pmu_inuse(int inuse)
+{
+	get_lppaca()->pmcregs_in_use = inuse;
+}
+
+extern void power4_enable_pmcs(void);
+
+#else /* CONFIG_PPC64 */
+
+static inline void ppc_set_pmu_inuse(int inuse) { }
+
 #endif
 
 #endif /* __KERNEL__ */
arch/powerpc/kernel/perf_counter.c +3 -10
index 70e1f57..ccd6b21 100644
@@ -62,7 +62,6 @@ static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
 {
 	return 0;
 }
-static inline void perf_set_pmu_inuse(int inuse) { }
 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp) { }
 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
 {
@@ -93,11 +92,6 @@ static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
 	return 0;
 }
 
-static inline void perf_set_pmu_inuse(int inuse)
-{
-	get_lppaca()->pmcregs_in_use = inuse;
-}
-
 /*
  * The user wants a data address recorded.
  * If we're not doing instruction sampling, give them the SDAR
@@ -531,8 +525,7 @@ void hw_perf_disable(void)
 		 * Check if we ever enabled the PMU on this cpu.
 		 */
 		if (!cpuhw->pmcs_enabled) {
-			if (ppc_md.enable_pmcs)
-				ppc_md.enable_pmcs();
+			ppc_enable_pmcs();
 			cpuhw->pmcs_enabled = 1;
 		}
 
@@ -594,7 +587,7 @@ void hw_perf_enable(void)
 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
 		mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
 		if (cpuhw->n_counters == 0)
-			perf_set_pmu_inuse(0);
+			ppc_set_pmu_inuse(0);
 		goto out_enable;
 	}
 
@@ -627,7 +620,7 @@ void hw_perf_enable(void)
 	 * bit set and set the hardware counters to their initial values.
 	 * Then unfreeze the counters.
 	 */
-	perf_set_pmu_inuse(1);
+	ppc_set_pmu_inuse(1);
 	mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
 	mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
 	mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
arch/powerpc/kernel/sysfs.c +3 -0
index f41aec8..956ab33 100644
@@ -17,6 +17,7 @@
 #include <asm/prom.h>
 #include <asm/machdep.h>
 #include <asm/smp.h>
+#include <asm/pmc.h>
 
 #include "cacheinfo.h"
 
@@ -123,6 +124,8 @@ static DEFINE_PER_CPU(char, pmcs_enabled);
 
 void ppc_enable_pmcs(void)
 {
+	ppc_set_pmu_inuse(1);
+
 	/* Only need to enable them once */
 	if (__get_cpu_var(pmcs_enabled))
 		return;
arch/powerpc/platforms/pseries/setup.c +0 -4
index 8d75ea2..ca5f2e1 100644
@@ -223,10 +223,6 @@ static void pseries_lpar_enable_pmcs(void)
 	set = 1UL << 63;
 	reset = 0;
 	plpar_hcall_norets(H_PERFMON, set, reset);
-
-	/* instruct hypervisor to maintain PMCs */
-	if (firmware_has_feature(FW_FEATURE_SPLPAR))
-		get_lppaca()->pmcregs_in_use = 1;
 }
 
 static void __init pseries_discover_pic(void)
Signature · 1 line
1.6.0.4
#1 lore raw
Quoted 19 lines · Paul Mackerras

Currently there is a bug where if you use oprofile on a pSeries machine, then use perf_counters, then use oprofile again, oprofile will not work correctly; it will lose the PMU configuration the next time the hypervisor does a partition context switch, and thereafter won't count anything.

Maynard Johnson identified the sequence causing the problem:
- oprofile setup calls ppc_enable_pmcs(), which calls pseries_lpar_enable_pmcs, which tells the hypervisor that we want to use the PMU, and sets the "PMU in use" flag in the lppaca. This flag tells the hypervisor whether it needs to save and restore the PMU config.
- The perf_counter code sets and clears the "PMU in use" flag directly as it context-switches the PMU between tasks, and leaves it clear when it finishes.
- oprofile setup, called for a new oprofile run, calls ppc_enable_pmcs, which does nothing because it has already been called. In particular it doesn't set the "PMU in use" flag.

Quoted 8 lines

This fixes the problem by arranging for ppc_enable_pmcs to always set the "PMU in use" flag. It makes the perf_counter code call ppc_enable_pmcs also rather than calling the lower-level function directly, and removes the setting of the "PMU in use" flag from pseries_lpar_enable_pmcs, since that is now done in its caller.

This also removes the declaration of pasemi_enable_pmcs because it isn't defined anywhere.

Thanks, Paul. I tested the patch, and oprofile and perf now play nicely together.

-Maynard

Quoted 127 lines
---
 arch/powerpc/include/asm/pmc.h         |   16 ++++++++++++++--
 arch/powerpc/kernel/perf_counter.c     |   13 +++----------
 arch/powerpc/kernel/sysfs.c            |    3 +++
 arch/powerpc/platforms/pseries/setup.c |    4 ----
 4 files changed, 20 insertions(+), 16 deletions(-)
arch/powerpc/include/asm/pmc.h +14 -2
index d6a616a..ccc68b5 100644
@@ -27,10 +27,22 @@ extern perf_irq_t perf_irq;
 
 int reserve_pmc_hardware(perf_irq_t new_perf_irq);
 void release_pmc_hardware(void);
+void ppc_enable_pmcs(void);
 
 #ifdef CONFIG_PPC64
-void power4_enable_pmcs(void);
-void pasemi_enable_pmcs(void);
+#include <asm/lppaca.h>
+
+static inline void ppc_set_pmu_inuse(int inuse)
+{
+	get_lppaca()->pmcregs_in_use = inuse;
+}
+
+extern void power4_enable_pmcs(void);
+
+#else /* CONFIG_PPC64 */
+
+static inline void ppc_set_pmu_inuse(int inuse) { }
+
 #endif
 
 #endif /* __KERNEL__ */
arch/powerpc/kernel/perf_counter.c +3 -10
index 70e1f57..ccd6b21 100644
@@ -62,7 +62,6 @@ static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
 {
 	return 0;
 }
-static inline void perf_set_pmu_inuse(int inuse) { }
 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp) { }
 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
 {
@@ -93,11 +92,6 @@ static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
 	return 0;
 }
 
-static inline void perf_set_pmu_inuse(int inuse)
-{
-	get_lppaca()->pmcregs_in_use = inuse;
-}
-
 /*
  * The user wants a data address recorded.
  * If we're not doing instruction sampling, give them the SDAR
@@ -531,8 +525,7 @@ void hw_perf_disable(void)
 		 * Check if we ever enabled the PMU on this cpu.
 		 */
 		if (!cpuhw->pmcs_enabled) {
-			if (ppc_md.enable_pmcs)
-				ppc_md.enable_pmcs();
+			ppc_enable_pmcs();
 			cpuhw->pmcs_enabled = 1;
 		}
 
@@ -594,7 +587,7 @@ void hw_perf_enable(void)
 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
 		mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
 		if (cpuhw->n_counters == 0)
-			perf_set_pmu_inuse(0);
+			ppc_set_pmu_inuse(0);
 		goto out_enable;
 	}
 
@@ -627,7 +620,7 @@ void hw_perf_enable(void)
 	 * bit set and set the hardware counters to their initial values.
 	 * Then unfreeze the counters.
 	 */
-	perf_set_pmu_inuse(1);
+	ppc_set_pmu_inuse(1);
 	mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
 	mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
 	mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
arch/powerpc/kernel/sysfs.c +3 -0
index f41aec8..956ab33 100644
@@ -17,6 +17,7 @@
 #include <asm/prom.h>
 #include <asm/machdep.h>
 #include <asm/smp.h>
+#include <asm/pmc.h>
 
 #include "cacheinfo.h"
 
@@ -123,6 +124,8 @@ static DEFINE_PER_CPU(char, pmcs_enabled);
 
 void ppc_enable_pmcs(void)
 {
+	ppc_set_pmu_inuse(1);
+
 	/* Only need to enable them once */
 	if (__get_cpu_var(pmcs_enabled))
 		return;
arch/powerpc/platforms/pseries/setup.c +0 -4
index 8d75ea2..ca5f2e1 100644
@@ -223,10 +223,6 @@ static void pseries_lpar_enable_pmcs(void)
 	set = 1UL << 63;
 	reset = 0;
 	plpar_hcall_norets(H_PERFMON, set, reset);
-
-	/* instruct hypervisor to maintain PMCs */
-	if (firmware_has_feature(FW_FEATURE_SPLPAR))
-		get_lppaca()->pmcregs_in_use = 1;
 }
 
 static void __init pseries_discover_pic(void)
#2 lore raw
Quoted 6 lines · Maynard Johnson
Quoted 5 lines · Paul Mackerras

Currently there is a bug where if you use oprofile on a pSeries machine, then use perf_counters, then use oprofile again, oprofile will not work correctly; it will lose the PMU configuration the next time the hypervisor does a partition context switch, and thereafter won't count anything.

Ben,
Is there any way to get this bug fix into 2.6.31 or is the window closed? Once the problem occurs, you can't get oprofile to work again without a reboot. Really would be nice (for many reasons) to get this fixed in .31.

Thanks.
-Maynard

Quoted 156 lines
Quoted 14 lines

Maynard Johnson identified the sequence causing the problem:
- oprofile setup calls ppc_enable_pmcs(), which calls pseries_lpar_enable_pmcs, which tells the hypervisor that we want to use the PMU, and sets the "PMU in use" flag in the lppaca. This flag tells the hypervisor whether it needs to save and restore the PMU config.
- The perf_counter code sets and clears the "PMU in use" flag directly as it context-switches the PMU between tasks, and leaves it clear when it finishes.
- oprofile setup, called for a new oprofile run, calls ppc_enable_pmcs, which does nothing because it has already been called. In particular it doesn't set the "PMU in use" flag.

Quoted 8 lines

This fixes the problem by arranging for ppc_enable_pmcs to always set the "PMU in use" flag. It makes the perf_counter code call ppc_enable_pmcs also rather than calling the lower-level function directly, and removes the setting of the "PMU in use" flag from pseries_lpar_enable_pmcs, since that is now done in its caller.

This also removes the declaration of pasemi_enable_pmcs because it isn't defined anywhere.

Thanks, Paul. I tested the patch, and oprofile and perf now play nicely together.

-Maynard

Quoted 116 lines
---
 arch/powerpc/include/asm/pmc.h         |   16 ++++++++++++++--
 arch/powerpc/kernel/perf_counter.c     |   13 +++----------
 arch/powerpc/kernel/sysfs.c            |    3 +++
 arch/powerpc/platforms/pseries/setup.c |    4 ----
 4 files changed, 20 insertions(+), 16 deletions(-)
arch/powerpc/include/asm/pmc.h +14 -2
index d6a616a..ccc68b5 100644
@@ -27,10 +27,22 @@ extern perf_irq_t perf_irq;
 
 int reserve_pmc_hardware(perf_irq_t new_perf_irq);
 void release_pmc_hardware(void);
+void ppc_enable_pmcs(void);
 
 #ifdef CONFIG_PPC64
-void power4_enable_pmcs(void);
-void pasemi_enable_pmcs(void);
+#include <asm/lppaca.h>
+
+static inline void ppc_set_pmu_inuse(int inuse)
+{
+	get_lppaca()->pmcregs_in_use = inuse;
+}
+
+extern void power4_enable_pmcs(void);
+
+#else /* CONFIG_PPC64 */
+
+static inline void ppc_set_pmu_inuse(int inuse) { }
+
 #endif
 
 #endif /* __KERNEL__ */
arch/powerpc/kernel/perf_counter.c +3 -10
index 70e1f57..ccd6b21 100644
@@ -62,7 +62,6 @@ static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
 {
 	return 0;
 }
-static inline void perf_set_pmu_inuse(int inuse) { }
 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp) { }
 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
 {
@@ -93,11 +92,6 @@ static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
 	return 0;
 }
 
-static inline void perf_set_pmu_inuse(int inuse)
-{
-	get_lppaca()->pmcregs_in_use = inuse;
-}
-
 /*
  * The user wants a data address recorded.
  * If we're not doing instruction sampling, give them the SDAR
@@ -531,8 +525,7 @@ void hw_perf_disable(void)
 		 * Check if we ever enabled the PMU on this cpu.
 		 */
 		if (!cpuhw->pmcs_enabled) {
-			if (ppc_md.enable_pmcs)
-				ppc_md.enable_pmcs();
+			ppc_enable_pmcs();
 			cpuhw->pmcs_enabled = 1;
 		}
 
@@ -594,7 +587,7 @@ void hw_perf_enable(void)
 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
 		mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
 		if (cpuhw->n_counters == 0)
-			perf_set_pmu_inuse(0);
+			ppc_set_pmu_inuse(0);
 		goto out_enable;
 	}
 
@@ -627,7 +620,7 @@ void hw_perf_enable(void)
 	 * bit set and set the hardware counters to their initial values.
 	 * Then unfreeze the counters.
 	 */
-	perf_set_pmu_inuse(1);
+	ppc_set_pmu_inuse(1);
 	mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
 	mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
 	mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
arch/powerpc/kernel/sysfs.c +3 -0
index f41aec8..956ab33 100644
@@ -17,6 +17,7 @@
 #include <asm/prom.h>
 #include <asm/machdep.h>
 #include <asm/smp.h>
+#include <asm/pmc.h>
 
 #include "cacheinfo.h"
 
@@ -123,6 +124,8 @@ static DEFINE_PER_CPU(char, pmcs_enabled);
 
 void ppc_enable_pmcs(void)
 {
+	ppc_set_pmu_inuse(1);
+
 	/* Only need to enable them once */
 	if (__get_cpu_var(pmcs_enabled))
 		return;
arch/powerpc/platforms/pseries/setup.c +0 -0
index 8d75ea2..ca5f2e1 100644

arch/powerpc/platforms/pseries/setup.c static void pseries_lpar_enable_pmcs(void) · line 228

Quoted 10 lines
arch/powerpc/platforms/pseries/setup.c +0 -4
@@ -223,10 +223,6 @@ static void pseries_lpar_enable_pmcs(void)
 	set = 1UL << 63;
 	reset = 0;
 	plpar_hcall_norets(H_PERFMON, set, reset);
-
-	/* instruct hypervisor to maintain PMCs */
-	if (firmware_has_feature(FW_FEATURE_SPLPAR))
-		get_lppaca()->pmcregs_in_use = 1;
 }
 
 static void __init pseries_discover_pic(void)

_______________________________________________
Linuxppc-dev mailing list
[email protected]
https://lists.ozlabs.org/listinfo/linuxppc-dev

#3 lore raw
Quoted 9 lines · Maynard Johnson
Quoted 6 lines · Maynard Johnson
Quoted 5 lines · Paul Mackerras

Currently there is a bug where if you use oprofile on a pSeries machine, then use perf_counters, then use oprofile again, oprofile will not work correctly; it will lose the PMU configuration the next time the hypervisor does a partition context switch, and thereafter won't count anything.

Ben,
Is there any way to get this bug fix into 2.6.31 or is the window closed? Once the problem occurs, you can't get oprofile to work again without a reboot. Really would be nice (for many reasons) to get this fixed in .31.

.31 is already released. Could probably get it into a -stable .31 kernel though.

josh

#4 lore raw
Quoted 8 lines · Josh Boyer
Quoted 7 lines · Maynard Johnson
Quoted 6 lines · Maynard Johnson
Quoted 5 lines · Paul Mackerras

Currently there is a bug where if you use oprofile on a pSeries machine, then use perf_counters, then use oprofile again, oprofile will not work correctly; it will lose the PMU configuration the next time the hypervisor does a partition context switch, and thereafter won't count anything.

Quoted 5 lines
Quoted 2 lines

Ben,
Is there any way to get this bug fix into 2.6.31 or is the window closed? Once the problem occurs, you can't get oprofile to work again without a reboot. Really would be nice (for many reasons) to get this fixed in .31.

.31 is already released. Could probably get it into a -stable .31 kernel though.

The fix is already upstream as well no ?

Ben.

Quoted 1 line

josh

#5 lore raw
Quoted 9 lines · Benjamin Herrenschmidt
Quoted 8 lines · Josh Boyer
Quoted 7 lines · Maynard Johnson
Quoted 6 lines · Maynard Johnson
Quoted 5 lines · Paul Mackerras

Currently there is a bug where if you use oprofile on a pSeries machine, then use perf_counters, then use oprofile again, oprofile will not work correctly; it will lose the PMU configuration the next time the hypervisor does a partition context switch, and thereafter won't count anything.

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

Ben,
Is there any way to get this bug fix into 2.6.31 or is the window closed? Once the problem occurs, you can't get oprofile to work again without a reboot. Really would be nice (for many reasons) to get this fixed in .31.

.31 is already released. Could probably get it into a -stable .31 kernel though.

The fix is already upstream as well no ?

Sort of. It's in your next branch, but not in Linus' tree.

Anyway, some people like to have particular versions work so -stable is good regardless :).

josh