PEBS and LBR fixes
Subsystems (4)
cover PEBS and LBR fixes
With these patches PEBS also works on my Core2 machines, although sometimes it fails to generate samples for some reason, so there's more errata to read.
PATCH 1/5 perf: Rework the arch CPU-hotplug hooks
Remove the hw_perf_event_*() hotplug hooks in favour of per PMU hotplug notifiers. This has the advantage of reducing the static weak interface as well as exposing all hotplug actions to the PMU.
Use this to fix x86 hotplug usage where we did things in ONLINE which should have been done in UP_PREPARE or STARTING.
--- arch/powerpc/kernel/perf_event.c | 21 +++++++++ arch/sh/kernel/perf_event.c | 20 ++++++++- arch/x86/kernel/cpu/perf_event.c | 66 ++++++++++++++++++------------ arch/x86/kernel/cpu/perf_event_amd.c | 60 +++++++++++---------------- arch/x86/kernel/cpu/perf_event_intel.c | 5 +- arch/x86/kernel/cpu/perf_event_intel_ds.c | 4 - include/linux/perf_event.h | 16 +++++++ kernel/perf_event.c | 29 ------------- 8 files changed, 127 insertions(+), 94 deletions(-)
arch/powerpc/kernel/perf_event.c +20 -1
@@ -1287,7 +1287,7 @@ static void perf_event_interrupt(structirq_exit (); } -void hw_perf_event_setup (int cpu) +static void power_pmu_setup (int cpu) {struct cpu_hw_events *cpuhw = &per_cpu (cpu_hw_events, cpu);
@@ -1297,6 +1297,23 @@ void hw_perf_event_setup(int cpu) cpuhw->mmcr[0 ] =MMCR0_FC ; } +static int __cpuinit +power_pmu_notify (struct notifier_block *self,unsigned long action,void *hcpu) +{ +unsigned int cpu = (long )hcpu; + +switch (action & ~CPU_TASKS_FROZEN ) { +case CPU_UP_PREPARE : +power_pmu_setup (cpu); +break ; + +default : +break ; + } + +return NOTIFY_OK ; +} +int register_power_pmu (struct power_pmu *pmu) {if (ppmu)
@@ -1314,5 +1331,7 @@ int register_power_pmu(struct power_pmu freeze_events_kernel =MMCR0_FCHV ;#endif /* CONFIG_PPC64 */ +perf_cpu_notifier (power_pmu_notifier); +return 0 ; }
arch/sh/kernel/perf_event.c +19 -1
@@ -275,13 +275,30 @@ const struct pmu *hw_perf_event_init(strreturn &pmu; } -void hw_perf_event_setup (int cpu) +static void sh_pmu_setup (int cpu) {struct cpu_hw_events *cpuhw = &per_cpu (cpu_hw_events, cpu);memset (cpuhw,0 ,sizeof (struct cpu_hw_events )); } +static int __cpuinit +sh_pmu_notifier (struct notifier_block *self,unsigned long action,void *hcpu) +{ +unsigned int cpu = (long )hcpu; + +switch (action & ~CPU_TASKS_FROZEN ) { +case CPU_UP_PREPARE : +sh_pmu_setup (cpu); +break ; + +default : +break ; + } + +return NOTIFY_OK ; +} +void hw_perf_enable (void ) {if (!sh_pmu_initialized ())
@@ -308,5 +325,6 @@ int register_sh_pmu(struct sh_pmu *pmu)WARN_ON (pmu->num_events >MAX_HWEVENTS ); +perf_cpu_notifier (sh_pmu_notifier);return 0 ; }
arch/x86/kernel/cpu/perf_event.c +40 -26
87 lines not on this page · raw message
arch/x86/kernel/cpu/perf_event_amd.c +25 -35
82 lines not on this page · raw message
arch/x86/kernel/cpu/perf_event_intel.c +4 -1
@@ -794,7 +794,10 @@ static __initconst struct x86_pmu intel_* the generic event period: */ .max_period = (1ULL << 31) - 1, - .get_event_constraints = intel_get_event_constraints + .get_event_constraints = intel_get_event_constraints, + + .cpu_starting = init_debug_store_on_cpu, + .cpu_dying = fini_debug_store_on_cpu, };static __init int intel_pmu_init (void )
arch/x86/kernel/cpu/perf_event_intel_ds.c +2 -2
@@ -63,7 +63,7 @@ struct debug_store {u64 pebs_event_reset[MAX_PEBS_EVENTS ]; }; -static inline void init_debug_store_on_cpu (int cpu) +static void init_debug_store_on_cpu (int cpu) {struct debug_store *ds =per_cpu (cpu_hw_events, cpu).ds;
@@ -75,7 +75,7 @@ static inline void init_debug_store_on_c (u32)((u64)(unsigned long)ds >> 32)); } -static inline void fini_debug_store_on_cpu (int cpu) +static void fini_debug_store_on_cpu (int cpu) {if (!per_cpu (cpu_hw_events, cpu).ds)return ;
include/linux/perf_event.h +16 -0
@@ -961,5 +961,21 @@ static inline void perf_event_disable(st#define perf_output_put (handle, x) \ perf_output_copy((handle), &(x), sizeof(x)) +/* +* This has to have a higher priority than migration_notifier in sched.c. +*/ +#define perf_cpu_notifier (fn) \ +do { \ + static struct notifier_block fn##_nb __cpuinitdata = \ + { .notifier_call = fn, .priority = 20 }; \ + fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \ + (void *)(unsigned long)smp_processor_id()); \ + fn(&fn##_nb, (unsigned long)CPU_STARTING, \ + (void *)(unsigned long)smp_processor_id()); \ + fn(&fn##_nb, (unsigned long)CPU_ONLINE, \ + (void *)(unsigned long)smp_processor_id()); \ + register_cpu_notifier(&fn##_nb); \ +} while (0) +#endif /* __KERNEL__ */ #endif /* _LINUX_PERF_EVENT_H */
kernel/perf_event.c +1 -28
@@ -81,10 +81,6 @@ extern __weak const struct pmu *hw_perf_void __weak hw_perf_disable (void ) {barrier (); }void __weak hw_perf_enable (void ) {barrier (); } -void __weak hw_perf_event_setup (int cpu) {barrier (); } -void __weak hw_perf_event_setup_online (int cpu) {barrier (); } -void __weak hw_perf_event_setup_offline (int cpu) {barrier (); } - int __weak hw_perf_group_sched_in(struct perf_event *group_leader, struct perf_cpu_context *cpuctx,
@@ -5404,8 +5400,6 @@ static void __cpuinit perf_event_init_cpspin_lock (&perf_resource_lock); cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;spin_unlock (&perf_resource_lock); - -hw_perf_event_setup (cpu); }#ifdef CONFIG_HOTPLUG_CPU
@@ -5445,20 +5439,11 @@ perf_cpu_notify(struct notifier_block *sperf_event_init_cpu (cpu);break ; -case CPU_ONLINE : -case CPU_ONLINE_FROZEN : -hw_perf_event_setup_online (cpu); -break ; -case CPU_DOWN_PREPARE :case CPU_DOWN_PREPARE_FROZEN :perf_event_exit_cpu (cpu);break ; -case CPU_DEAD : -hw_perf_event_setup_offline (cpu); -break ; -default :break ; }
@@ -5466,21 +5451,9 @@ perf_cpu_notify(struct notifier_block *sreturn NOTIFY_OK ; } -/* -* This has to have a higher priority than migration_notifier in sched.c. -*/ -static struct notifier_block __cpuinitdata perf_cpu_nb = { - .notifier_call = perf_cpu_notify, - .priority =20 , -}; -void __init perf_event_init (void ) { -perf_cpu_notify (&perf_cpu_nb, (unsigned long )CPU_UP_PREPARE , - (void *)(long )smp_processor_id ()); -perf_cpu_notify (&perf_cpu_nb, (unsigned long )CPU_ONLINE , - (void *)(long )smp_processor_id ()); -register_cpu_notifier (&perf_cpu_nb); +perf_cpu_notifier (perf_cpu_notify); }static ssize_t perf_show_reserve_percpu (struct sysdev_class *class,char *buf)
[tip:perf/urgent] perf: Rework and fix the arch CPU-hotplug hooks
Commit-ID: 3f6da3905398826d85731247e7fbcf53400c18bd Gitweb: http://git.kernel.org/tip/3f6da3905398826d85731247e7fbcf53400c18bd Author: Peter Zijlstra <[email protected]> AuthorDate: Fri, 5 Mar 2010 13:01:18 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Wed, 10 Mar 2010 13:22:24 +0100
perf: Rework and fix the arch CPU-hotplug hooks
Remove the hw_perf_event_*() hotplug hooks in favour of per PMU hotplug notifiers. This has the advantage of reducing the static weak interface as well as exposing all hotplug actions to the PMU.
Use this to fix x86 hotplug usage where we did things in ONLINE which should have been done in UP_PREPARE or STARTING.
LKML-Reference: <[email protected]>
--- arch/powerpc/kernel/perf_event.c | 21 +++++++++- arch/sh/kernel/perf_event.c | 20 +++++++++- arch/x86/kernel/cpu/perf_event.c | 70 +++++++++++++++++++------------- arch/x86/kernel/cpu/perf_event_amd.c | 60 +++++++++++---------------- arch/x86/kernel/cpu/perf_event_intel.c | 5 ++- include/linux/perf_event.h | 16 +++++++ kernel/perf_event.c | 15 ------- 7 files changed, 126 insertions(+), 81 deletions(-)
arch/powerpc/kernel/perf_event.c +20 -1
@@ -1287,7 +1287,7 @@ static void perf_event_interrupt(struct pt_regs *regs)irq_exit (); } -void hw_perf_event_setup (int cpu) +static void power_pmu_setup (int cpu) {struct cpu_hw_events *cpuhw = &per_cpu (cpu_hw_events, cpu);
@@ -1297,6 +1297,23 @@ void hw_perf_event_setup(int cpu) cpuhw->mmcr[0 ] =MMCR0_FC ; } +static int __cpuinit +power_pmu_notify (struct notifier_block *self,unsigned long action,void *hcpu) +{ +unsigned int cpu = (long )hcpu; + +switch (action & ~CPU_TASKS_FROZEN ) { +case CPU_UP_PREPARE : +power_pmu_setup (cpu); +break ; + +default : +break ; + } + +return NOTIFY_OK ; +} +int register_power_pmu (struct power_pmu *pmu) {if (ppmu)
@@ -1314,5 +1331,7 @@ int register_power_pmu(struct power_pmu *pmu) freeze_events_kernel =MMCR0_FCHV ;#endif /* CONFIG_PPC64 */ +perf_cpu_notifier (power_pmu_notifier); +return 0 ; }
arch/sh/kernel/perf_event.c +19 -1
@@ -275,13 +275,30 @@ const struct pmu *hw_perf_event_init(struct perf_event *event)return &pmu; } -void hw_perf_event_setup (int cpu) +static void sh_pmu_setup (int cpu) {struct cpu_hw_events *cpuhw = &per_cpu (cpu_hw_events, cpu);memset (cpuhw,0 ,sizeof (struct cpu_hw_events )); } +static int __cpuinit +sh_pmu_notifier (struct notifier_block *self,unsigned long action,void *hcpu) +{ +unsigned int cpu = (long )hcpu; + +switch (action & ~CPU_TASKS_FROZEN ) { +case CPU_UP_PREPARE : +sh_pmu_setup (cpu); +break ; + +default : +break ; + } + +return NOTIFY_OK ; +} +void hw_perf_enable (void ) {if (!sh_pmu_initialized ())
@@ -308,5 +325,6 @@ int register_sh_pmu(struct sh_pmu *pmu)WARN_ON (pmu->num_events >MAX_HWEVENTS ); +perf_cpu_notifier (sh_pmu_notifier);return 0 ; }
arch/x86/kernel/cpu/perf_event.c +42 -28
103 lines not on this page · raw message
arch/x86/kernel/cpu/perf_event_amd.c +25 -35
82 lines not on this page · raw message
arch/x86/kernel/cpu/perf_event_intel.c +4 -1
@@ -870,7 +870,10 @@ static __initconst struct x86_pmu intel_pmu = { .max_period = (1ULL << 31) - 1, .enable_bts = intel_pmu_enable_bts, .disable_bts = intel_pmu_disable_bts, - .get_event_constraints = intel_get_event_constraints + .get_event_constraints = intel_get_event_constraints, + + .cpu_starting = init_debug_store_on_cpu, + .cpu_dying = fini_debug_store_on_cpu, };static __init int intel_pmu_init (void )
include/linux/perf_event.h +16 -0
@@ -936,5 +936,21 @@ static inline void perf_event_disable(struct perf_event *event) { }#define perf_output_put (handle, x) \ perf_output_copy((handle), &(x), sizeof(x)) +/* +* This has to have a higher priority than migration_notifier in sched.c. +*/ +#define perf_cpu_notifier (fn) \ +do { \ + static struct notifier_block fn##_nb __cpuinitdata = \ + { .notifier_call = fn, .priority = 20 }; \ + fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \ + (void *)(unsigned long)smp_processor_id()); \ + fn(&fn##_nb, (unsigned long)CPU_STARTING, \ + (void *)(unsigned long)smp_processor_id()); \ + fn(&fn##_nb, (unsigned long)CPU_ONLINE, \ + (void *)(unsigned long)smp_processor_id()); \ + register_cpu_notifier(&fn##_nb); \ +} while (0) +#endif /* __KERNEL__ */ #endif /* _LINUX_PERF_EVENT_H */
kernel/perf_event.c +0 -15
@@ -81,10 +81,6 @@ extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)void __weak hw_perf_disable (void ) {barrier (); }void __weak hw_perf_enable (void ) {barrier (); } -void __weak hw_perf_event_setup (int cpu) {barrier (); } -void __weak hw_perf_event_setup_online (int cpu) {barrier (); } -void __weak hw_perf_event_setup_offline (int cpu) {barrier (); } - int __weak hw_perf_group_sched_in(struct perf_event *group_leader, struct perf_cpu_context *cpuctx,
@@ -5382,8 +5378,6 @@ static void __cpuinit perf_event_init_cpu(int cpu)spin_lock (&perf_resource_lock); cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;spin_unlock (&perf_resource_lock); - -hw_perf_event_setup (cpu); }#ifdef CONFIG_HOTPLUG_CPU
@@ -5423,20 +5417,11 @@ perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)perf_event_init_cpu (cpu);break ; -case CPU_ONLINE : -case CPU_ONLINE_FROZEN : -hw_perf_event_setup_online (cpu); -break ; -case CPU_DOWN_PREPARE :case CPU_DOWN_PREPARE_FROZEN :perf_event_exit_cpu (cpu);break ; -case CPU_DEAD : -hw_perf_event_setup_offline (cpu); -break ; -default :break ; }
PATCH 2/5 perf, x86: Fix silly bug in data store buffer allocation
In reply to Peter ZijlstraFix up the ds allocation error path, where we could free @buffer before we used it.
--- arch/x86/kernel/cpu/perf_event_intel_ds.c | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-)
arch/x86/kernel/cpu/perf_event_intel_ds.c +1 -3
@@ -127,10 +127,8 @@ static int reserve_ds_buffers(void) err = -ENOMEM ; ds =kzalloc (sizeof (*ds),GFP_KERNEL ); -if (unlikely (!ds)) { -kfree (buffer); +if (unlikely (!ds))break ; - }per_cpu (cpu_hw_events, cpu).ds = ds;if (x86_pmu.bts) {
[tip:perf/pebs] perf, x86: Fix silly bug in data store buffer allocation
Commit-ID: 3adaebd69557615c1bf0365ce5e32d93ac7d82af Gitweb: http://git.kernel.org/tip/3adaebd69557615c1bf0365ce5e32d93ac7d82af Author: Peter Zijlstra <[email protected]> AuthorDate: Fri, 5 Mar 2010 12:09:29 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Wed, 10 Mar 2010 13:23:34 +0100
perf, x86: Fix silly bug in data store buffer allocation
Fix up the ds allocation error path, where we could free @buffer before we used it.
LKML-Reference: <[email protected]>
--- arch/x86/kernel/cpu/perf_event_intel_ds.c | 4 +--- 1 files changed, 1 insertions(+), 3 deletions(-)
arch/x86/kernel/cpu/perf_event_intel_ds.c +1 -3
@@ -127,10 +127,8 @@ static int reserve_ds_buffers(void) err = -ENOMEM ; ds =kzalloc (sizeof (*ds),GFP_KERNEL ); -if (unlikely (!ds)) { -kfree (buffer); +if (unlikely (!ds))break ; - }per_cpu (cpu_hw_events, cpu).ds = ds;if (x86_pmu.bts) {
PATCH 3/5 perf, x86: Disable PEBS on clowertown chips
In reply to Peter ZijlstraThis CPU has just too many handycaps to be really useful.
--- arch/x86/kernel/cpu/perf_event.c | 4 ++++ arch/x86/kernel/cpu/perf_event_intel.c | 27 +++++++++++++++++++++++++++ 2 files changed, 31 insertions(+)
arch/x86/kernel/cpu/perf_event.c +4 -0
@@ -197,6 +197,7 @@ struct x86_pmu {void (*put_event_constraints)(struct cpu_hw_events *cpuc,struct perf_event *event);struct event_constraint *event_constraints; +void (*quirks)(void );void (*cpu_prepare)(int cpu);void (*cpu_starting)(int cpu);
@@ -1380,6 +1381,9 @@ void __init init_hw_perf_events(void)pr_cont ("%s PMU driver.\n" , x86_pmu.name); +if (x86_pmu.quirks) + x86_pmu.quirks (); +if (x86_pmu.num_events >X86_PMC_MAX_GENERIC ) {WARN (1 ,KERN_ERR "hw perf events %d > max(%d), clipping!" , x86_pmu.num_events,X86_PMC_MAX_GENERIC );
arch/x86/kernel/cpu/perf_event_intel.c +27 -0
@@ -800,6 +800,32 @@ static __initconst struct x86_pmu intel_ .cpu_dying = fini_debug_store_on_cpu, }; +static void intel_clowertown_quirks (void ) +{ +/* +* PEBS is unreliable due to: +* +* AJ67 - PEBS may experience CPL leaks +* AJ68 - PEBS PMI may be delayed by one event +* AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] +* AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS +* +* AJ67 could be worked around by restricting the OS/USR flags. +* AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. +* +* AJ106 could possibly be worked around by not allowing LBR +* usage from PEBS, including the fixup. +* AJ68 could possibly be worked around by always programming +* a pebs_event_reset[0] value and coping with the lost events. +* +* But taken together it might just make sense to not enable PEBS on +* these chips. +*/ +printk (KERN_WARNING "PEBS disabled due to CPU errata.\n" ); + x86_pmu.pebs =0 ; + x86_pmu.pebs_constraints =NULL ; +} +static __init int intel_pmu_init (void ) {union cpuid10_edx edx;
@@ -864,6 +890,7 @@ static __init int intel_pmu_init(void)break ;case 15 :/* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */ + x86_pmu.quirks = intel_clowertown_quirks;case 22 :/* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */ case 23 :/* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */ case 29 :/* six-core 45 nm xeon "Dunnington" */
Re: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 70 lines · Peter Zijlstra
This CPU has just too many handycaps to be really useful.
Signed-off-by Peter Zijlstra <[email protected]>---
arch/x86/kernel/cpu/perf_event.c | 4 ++++arch/x86/kernel/cpu/perf_event_intel.c| 27 +++++++++++++++++++++++++++2 files changed, 31 insertions(+)Index:
linux-2.6/arch/x86/kernel/cpu/perf_event.c
===================================================================arch/x86/kernel/cpu/perf_event.c +4
-0@@ -197,6 +197,7 @@ struct x86_pmu { void (*put_event_constraints)(struct cpu_hw_events *cpuc, struct perf_event *event); struct event_constraint *event_constraints; + void (*quirks)(void); void (*cpu_prepare)(int cpu); void (*cpu_starting)(int cpu);@@ -1380,6 +1381,9 @@ void __init init_hw_perf_events(void) pr_cont("%s PMU driver.\n", x86_pmu.name); + if (x86_pmu.quirks) + x86_pmu.quirks(); + if (x86_pmu.num_events > X86_PMC_MAX_GENERIC) { WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", x86_pmu.num_events, X86_PMC_MAX_GENERIC);Index:
linux-2.6/arch/x86/kernel/cpu/perf_event_intel.c
===================================================================arch/x86/kernel/cpu/perf_event_intel.c +26
-0@@ -800,6 +800,32 @@ static __initconst struct x86_pmu intel_ .cpu_dying = fini_debug_store_on_cpu, }; +static void intel_clowertown_quirks (void ) +{ + /* + * PEBS is unreliable due to: + * + * AJ67 - PEBS may experience CPL leaks + * AJ68 - PEBS PMI may be delayed by one event + * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] + * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS + * + * AJ67 could be worked around by restricting the OS/USR flags. + * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. + * + * AJ106 could possibly be worked around by not allowing LBR + * usage from PEBS, including the fixup. + * AJ68 could possibly be worked around by always programming + * a pebs_event_reset[0] value and coping with the lost events. + * + * But taken together it might just make sense to not enable PEBS on + * these chips. + */ + printk(KERN_WARNING "PEBS disabled due to CPU errata.\n"); + x86_pmu.pebs = 0; + x86_pmu.pebs_constraints = NULL; +} + static __init int intel_pmu_init(void) { union cpuid10_edx edx;@@ -864,6 +890,7 @@ static __init int intel_pmu_init(void) break;
Quoted 2 lines
case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
+ x86_pmu.quirks =intel_clowertown_quirks;
That's too coarse grain!
It is more subtle than this. Some of the errata are marked as Plan fix. They seem to be
fixed in later models. Your looking at the E5xxx series errata but the E7xxx do not have
the same problems.
As for CPL leaking, you get that also with regular sampling when you execute near
the user/kernel boundary given you have skid. Either the kernel or the user has to
drop samples.
Peter Zijlstra
maintainerRe: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 9 lines · Stephane Eranian
Quoted 2 lines
case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */+ x86_pmu.quirks =
intel_clowertown_quirks;That's too coarse grain!
It is more subtle than this. Some of the errata are marked as Plan fix. They seem to be
fixed in later models. Your looking at the E5xxx series errata but the E7xxx do not have
the same problems.
OK, I'll look at those errata again and try to come up with a stepping test for this errata.
Re: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 11 lines · Peter Zijlstra
Quoted 9 lines · Stephane Eranian
Quoted 2 lines
case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
+ x86_pmu.quirks =intel_clowertown_quirks;That's too coarse grain!
It is more subtle than this. Some of the errata are marked as Plan fix. They seem to be
fixed in later models. Your looking at the E5xxx series errata but the E7xxx do not have
the same problems.
Quoted 3 lines
OK, I'll look at those errata again and try to come up with a stepping test for this errata.
There is erratum which you need to implement the workaround for and this is AAJ91 on Nehalem. It does happen.
Peter Zijlstra
maintainerRe: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 12 lines · Stephane Eranian
Quoted 11 lines · Peter Zijlstra
Quoted 9 lines · Stephane Eranian
Quoted 2 lines
case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */+ x86_pmu.quirks =
intel_clowertown_quirks;That's too coarse grain!
It is more subtle than this. Some of the errata are marked as Plan fix. They seem to be
fixed in later models. Your looking at the E5xxx series errata but the E7xxx do not have
the same problems.
Quoted 5 lines
Quoted 3 lines
OK, I'll look at those errata again and try to come up with a stepping test for this errata.
There is erratum which you need to implement the workaround for and this is AAJ91 on Nehalem. It does happen.
Yes, I found that one too today. Its on my todo list.
Peter Zijlstra
maintainerRe: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
In reply to Peter ZijlstraQuoted 11 lines · Peter Zijlstra
Quoted 9 lines · Stephane Eranian
Quoted 2 lines
case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */+ x86_pmu.quirks =
intel_clowertown_quirks;That's too coarse grain!
It is more subtle than this. Some of the errata are marked as Plan fix. They seem to be
fixed in later models. Your looking at the E5xxx series errata but the E7xxx do not have
the same problems.
Quoted 2 lines
OK, I'll look at those errata again and try to come up with a stepping test for this errata.
The two serious ones, AJ106 and AJ68 are no fix and are listed as such in all errata I can find, including the 7[23]00 series.
I checked the 65nm Core2Duo, Xeon 5200 and Xeon 7[23]00 spec updates. Going by that it seems the full model 15 family is broken and I'll leave the patch as is.
Re: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 12 lines · Peter Zijlstra
Quoted 11 lines · Peter Zijlstra
Quoted 9 lines · Stephane Eranian
Quoted 2 lines
case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
+ x86_pmu.quirks =intel_clowertown_quirks;That's too coarse grain!
It is more subtle than this. Some of the errata are marked as Plan fix. They seem to be
fixed in later models. Your looking at the E5xxx series errata but the E7xxx do not have
the same problems.
Quoted 6 lines
Quoted 2 lines
OK, I'll look at those errata again and try to come up with a stepping test for this errata.
The two serious ones, AJ106 and AJ68 are no fix and are listed as such in all errata I can find, including the 7[23]00 series.
Not E74xx. I think it would be fine to drop LBR with PEBS as the work-around to AJ106.
Quoted 3 lines
I checked the 65nm Core2Duo, Xeon 5200 and Xeon 7[23]00 spec updates. Going by that it seems the full model 15 family is broken and I'll leave the patch as is.
But the E74xx are okay and you are excluding them. Worst case you should provide an override.
Peter Zijlstra
maintainerRe: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 6 lines · Stephane Eranian
Quoted 3 lines
The two serious ones, AJ106 and AJ68 are no fix and are listed as such in all errata I can find, including the 7[23]00 series.
Not E74xx. I think it would be fine to drop LBR with PEBS as the work-around to AJ106.
Quoted 8 lines
Quoted 3 lines
I checked the 65nm Core2Duo, Xeon 5200 and Xeon 7[23]00 spec updates. Going by that it seems the full model 15 family is broken and I'll leave the patch as is.
But the E74xx are okay and you are excluding them. Worst case you should provide an override.
From what I can tell E74xx is model 29, which would be just fine with
this patch, this patch only marks model 15 as broken.
Re: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 11 lines · Peter Zijlstra
Quoted 10 lines · Stephane Eranian
Quoted 3 lines
The two serious ones, AJ106 and AJ68 are no fix and are listed as such in all errata I can find, including the 7[23]00 series.
Not E74xx. I think it would be fine to drop LBR with PEBS as the work-around to AJ106.
Quoted 3 lines
I checked the 65nm Core2Duo, Xeon 5200 and Xeon 7[23]00 spec updates. Going by that it seems the full model 15 family is broken and I'll leave the patch as is.
Quoted 6 lines
Quoted 4 lines
But the E74xx are okay and you are excluding them. Worst case you should provide an override.
From what I can tell E74xx is model 29, which would be just fine with
this patch, this patch only marks model 15 as broken.
True, my bad.
So it would be a matter to provide some ways of disabling LBR with PEBS on model 15.
Peter Zijlstra
maintainerRe: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 15 lines · Stephane Eranian
Quoted 14 lines · Peter Zijlstra
Quoted 12 lines · Stephane Eranian
Quoted 3 lines
The two serious ones, AJ106 and AJ68 are no fix and are listed as such in all errata I can find, including the 7[23]00 series.
Not E74xx. I think it would be fine to drop LBR with PEBS as the work-around to AJ106.
Quoted 3 lines
I checked the 65nm Core2Duo, Xeon 5200 and Xeon 7[23]00 spec updates. Going by that it seems the full model 15 family is broken and I'll leave the patch as is.
But the E74xx are okay and you are excluding them. Worst case you should provide an override.
Quoted 6 lines
Quoted 3 lines
Quoted 1 line
From what I can tell E74xx is model 29, which would be just fine with
this patch, this patch only marks model 15 as broken.
True, my bad.
So it would be a matter to provide some ways of disabling LBR with PEBS on model 15.
And some way of dealing with the flaky PEBS PMI, which requires we program the pebs_event_reset thing otherwise there won't be a next PEBS record to re-trigger the PMI.
If it would have stated it always did the threshold comparison first and then the index increment, so that we'd always trigger on the second record we could simply program half the period and always take two and ignore the first, but it says _MAY_, so its a race and there's no reliable solution.
It might all be possible but I don't see it being worth the effort.
Re: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 27 lines · Peter Zijlstra
Quoted 15 lines · Stephane Eranian
Quoted 14 lines · Peter Zijlstra
Quoted 12 lines · Stephane Eranian
Quoted 3 lines
The two serious ones, AJ106 and AJ68 are no fix and are listed as such in all errata I can find, including the 7[23]00 series.
Not E74xx. I think it would be fine to drop LBR with PEBS as the work-around to AJ106.
Quoted 3 lines
I checked the 65nm Core2Duo, Xeon 5200 and Xeon 7[23]00 spec updates. Going by that it seems the full model 15 family is broken and I'll leave the patch as is.
But the E74xx are okay and you are excluding them. Worst case you should provide an override.
Quoted 6 lines
Quoted 3 lines
Quoted 1 line
From what I can tell E74xx is model 29, which would be just fine with
this patch, this patch only marks model 15 as broken.
True, my bad.
So it would be a matter to provide some ways of disabling LBR with PEBS on model 15.And some way of dealing with the flaky PEBS PMI, which requires we program the
pebs_event_resetthing otherwise there won't be a next PEBS record to re-trigger the PMI.
Quoted 8 lines
If it would have stated it always did the threshold comparison first and then the index increment, so that we'd always trigger on the second record we could simply program half the period and always take two and ignore the first, but it says _MAY_, so its a race and there's no reliable solution.
It might all be possible but I don't see it being worth the effort.
There are lots of model15 out there, like the Q6xxx series.
When I read AJ68, my understanding is that it's not that you do not get the interrupt. It will be delayed by one event. The buffer will become full. You won't overrun the buffer, you will get the interrupt at the next event. On interrupt, you have to reset the PEBS position pointer anyway. There is already a disconnect between the sampling period and the actual instruction sampled. That's not making the situation that much worse, unless I am missing something.
Peter Zijlstra
maintainerRe: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 7 lines · Stephane Eranian
When I read AJ68, my understanding is that it's not that you do not get the interrupt. It will be delayed by one event. The buffer will become full. You won't overrun the buffer, you will get the interrupt at the next event. On interrupt, you have to reset the PEBS position pointer anyway. There is already a disconnect between the sampling period and the actual instruction sampled. That's not making the situation that much worse, unless I am missing something.
The current code doesn't use the buffering at all, it uses single-shot PEBS by keeping pebs_event_reset 0 and setting a threshold of a single entry, so if due to AJ68 we miss a PMI it will never come.
I guess we can fudge something, but at what point does the whole thing stop being useful?
It would end up being something with fuzzy period and fuzzy location, which is a loss-loss situation if you ask me.
Re: [PATCH 3/5] perf, x86: Disable PEBS on clowertown chips
Quoted 10 lines · Peter Zijlstra
Quoted 7 lines · Stephane Eranian
When I read AJ68, my understanding is that it's not that you do not get the interrupt. It will be delayed by one event. The buffer will become full. You won't overrun the buffer, you will get the interrupt at the next event. On interrupt, you have to reset the PEBS position pointer anyway. There is already a disconnect between the sampling period and the actual instruction sampled. That's not making the situation that much worse, unless I am missing something.
Quoted 4 lines
The current code doesn't use the buffering at all, it uses single-shot PEBS by keeping
pebs_event_reset0 and setting a threshold of a single entry, so if due to AJ68 we miss a PMI it will never come.
What stops PEBS is that it crosses the end of the buffer. The buffer is one-deep,
threshold = buffer end, you get a single sample. reset has nothing to do with this.
AJ68 does not say you miss a PMI, it says the PMI comes at the next event. I suspect they mean the next observed event, and not necessarily the next recorded event. But I can check on that.
Quoted 2 lines
I guess we can fudge something, but at what point does the whole thing stop being useful?
What matters is that you get a valid sample.
Quoted 4 lines
It would end up being something with fuzzy period and fuzzy location, which is a loss-loss situation if you ask me.
The period is already fuzzy to begin with. Recall our discussion a couple of weeks back.
[tip:perf/pebs] perf, x86: Disable PEBS on clovertown chips
In reply to Peter ZijlstraCommit-ID: 3c44780b220e876b01e39d4028cd6f4205fbf5d6 Gitweb: http://git.kernel.org/tip/3c44780b220e876b01e39d4028cd6f4205fbf5d6 Author: Peter Zijlstra <[email protected]> AuthorDate: Thu, 4 Mar 2010 21:49:01 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Wed, 10 Mar 2010 13:23:35 +0100
perf, x86: Disable PEBS on clovertown chips
This CPU has just too many handycaps to be really useful.
LKML-Reference: <[email protected]>
--- arch/x86/kernel/cpu/perf_event.c | 4 ++++ arch/x86/kernel/cpu/perf_event_intel.c | 27 +++++++++++++++++++++++++++ 2 files changed, 31 insertions(+), 0 deletions(-)
arch/x86/kernel/cpu/perf_event.c +4 -0
@@ -197,6 +197,7 @@ struct x86_pmu {void (*put_event_constraints)(struct cpu_hw_events *cpuc,struct perf_event *event);struct event_constraint *event_constraints; +void (*quirks)(void );void (*cpu_prepare)(int cpu);void (*cpu_starting)(int cpu);
@@ -1373,6 +1374,9 @@ void __init init_hw_perf_events(void)pr_cont ("%s PMU driver.\n" , x86_pmu.name); +if (x86_pmu.quirks) + x86_pmu.quirks (); +if (x86_pmu.num_events >X86_PMC_MAX_GENERIC ) {WARN (1 ,KERN_ERR "hw perf events %d > max(%d), clipping!" , x86_pmu.num_events,X86_PMC_MAX_GENERIC );
arch/x86/kernel/cpu/perf_event_intel.c +27 -0
@@ -792,6 +792,32 @@ static __initconst struct x86_pmu intel_pmu = { .cpu_dying = fini_debug_store_on_cpu, }; +static void intel_clovertown_quirks (void ) +{ +/* +* PEBS is unreliable due to: +* +* AJ67 - PEBS may experience CPL leaks +* AJ68 - PEBS PMI may be delayed by one event +* AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] +* AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS +* +* AJ67 could be worked around by restricting the OS/USR flags. +* AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. +* +* AJ106 could possibly be worked around by not allowing LBR +* usage from PEBS, including the fixup. +* AJ68 could possibly be worked around by always programming +* a pebs_event_reset[0] value and coping with the lost events. +* +* But taken together it might just make sense to not enable PEBS on +* these chips. +*/ +printk (KERN_WARNING "PEBS disabled due to CPU errata.\n" ); + x86_pmu.pebs =0 ; + x86_pmu.pebs_constraints =NULL ; +} +static __init int intel_pmu_init (void ) {union cpuid10_edx edx;
@@ -856,6 +882,7 @@ static __init int intel_pmu_init(void)break ;case 15 :/* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */ + x86_pmu.quirks = intel_clovertown_quirks;case 22 :/* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */ case 23 :/* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */ case 29 :/* six-core 45 nm xeon "Dunnington" */
PATCH 4/5 perf, x86: Clear the LBRs on init
In reply to Peter ZijlstraSome CPUs have errata where the LBR is not cleared on Power-On. So always clear the LBRs before use.
--- arch/x86/kernel/cpu/perf_event_intel.c | 18 ++++++++++++++++-- arch/x86/kernel/cpu/perf_event_intel_lbr.c | 3 +++ 2 files changed, 19 insertions(+), 2 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +16 -2
@@ -775,6 +775,20 @@ static __initconst struct x86_pmu core_p .event_constraints = intel_core_event_constraints, }; +static void intel_pmu_cpu_starting (int cpu) +{ +init_debug_store_on_cpu (cpu); +/* +* Deal with CPUs that don't clear their LBRs on power-up. +*/ +intel_pmu_lbr_reset (); +} + +static void intel_pmu_cpu_dying (int cpu) +{ +fini_debug_store_on_cpu (cpu); +} +static __initconst struct x86_pmu intel_pmu = { .name ="Intel" , .handle_irq = intel_pmu_handle_irq,
@@ -796,8 +810,8 @@ static __initconst struct x86_pmu intel_ .max_period = (1ULL << 31) - 1, .get_event_constraints = intel_get_event_constraints, - .cpu_starting = init_debug_store_on_cpu, - .cpu_dying = fini_debug_store_on_cpu, + .cpu_starting = intel_pmu_cpu_starting, + .cpu_dying = intel_pmu_cpu_dying, };static void intel_clowertown_quirks (void )
arch/x86/kernel/cpu/perf_event_intel_lbr.c +3 -0
@@ -53,6 +53,9 @@ static void intel_pmu_lbr_reset_64(void)static void intel_pmu_lbr_reset (void ) { +if (!x86_pmu.lbr_nr) +return ; +if (x86_pmu.intel_cap.lbr_format ==LBR_FORMAT_32 )intel_pmu_lbr_reset_32 ();else
[tip:perf/pebs] perf, x86: Clear the LBRs on init
Commit-ID: 74846d35b24b6efd61bb88a0a750b6bb257e6e78 Gitweb: http://git.kernel.org/tip/74846d35b24b6efd61bb88a0a750b6bb257e6e78 Author: Peter Zijlstra <[email protected]> AuthorDate: Fri, 5 Mar 2010 13:49:35 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Wed, 10 Mar 2010 13:23:35 +0100
perf, x86: Clear the LBRs on init
Some CPUs have errata where the LBR is not cleared on Power-On. So always clear the LBRs before use.
LKML-Reference: <[email protected]>
--- arch/x86/kernel/cpu/perf_event_intel.c | 18 ++++++++++++++++-- arch/x86/kernel/cpu/perf_event_intel_lbr.c | 3 +++ 2 files changed, 19 insertions(+), 2 deletions(-)
arch/x86/kernel/cpu/perf_event_intel.c +16 -2
@@ -767,6 +767,20 @@ static __initconst struct x86_pmu core_pmu = { .event_constraints = intel_core_event_constraints, }; +static void intel_pmu_cpu_starting (int cpu) +{ +init_debug_store_on_cpu (cpu); +/* +* Deal with CPUs that don't clear their LBRs on power-up. +*/ +intel_pmu_lbr_reset (); +} + +static void intel_pmu_cpu_dying (int cpu) +{ +fini_debug_store_on_cpu (cpu); +} +static __initconst struct x86_pmu intel_pmu = { .name ="Intel" , .handle_irq = intel_pmu_handle_irq,
@@ -788,8 +802,8 @@ static __initconst struct x86_pmu intel_pmu = { .max_period = (1ULL << 31) - 1, .get_event_constraints = intel_get_event_constraints, - .cpu_starting = init_debug_store_on_cpu, - .cpu_dying = fini_debug_store_on_cpu, + .cpu_starting = intel_pmu_cpu_starting, + .cpu_dying = intel_pmu_cpu_dying, };static void intel_clovertown_quirks (void )
arch/x86/kernel/cpu/perf_event_intel_lbr.c +3 -0
@@ -53,6 +53,9 @@ static void intel_pmu_lbr_reset_64(void)static void intel_pmu_lbr_reset (void ) { +if (!x86_pmu.lbr_nr) +return ; +if (x86_pmu.intel_cap.lbr_format ==LBR_FORMAT_32 )intel_pmu_lbr_reset_32 ();else
PATCH 5/5 perf, x86: Robustify PEBS fixup
In reply to Peter ZijlstraIt turns out the LBR is massively unreliable on certain CPUs, so code the fixup a little more defensive to avoid crashing the kernel.
--- arch/x86/kernel/cpu/perf_event_intel_ds.c | 21 +++++++++++++++++++-- 1 file changed, 19 insertions(+), 2 deletions(-)
arch/x86/kernel/cpu/perf_event_intel_ds.c +19 -2
@@ -399,10 +399,23 @@ static int intel_pmu_pebs_fixup_ip(strucif (!x86_pmu.intel_cap.pebs_trap)return 1 ; +/* +* No LBR entry, no basic block, no rewinding +*/ if (!cpuc->lbr_stack.nr || !from || !to)return 0 ; -if (ip < to) +/* +* Basic blocks should never cross user/kernel boundaries +*/ +if (kernel_ip (ip) !=kernel_ip (to)) +return 0 ; + +/* +* unsigned math, either ip is before the start (impossible) or +* the basic block is larger than 1 page (sanity) +*/ +if ((ip - to) >PAGE_SIZE )return 0 ;/*
@@ -420,7 +433,7 @@ static int intel_pmu_pebs_fixup_ip(struc old_to = to;if (!kernel_ip (ip)) { -int bytes, size =min_t (int,MAX_INSN_SIZE , ip - to); +int bytes, size =MAX_INSN_SIZE ; bytes =copy_from_user_nmi (buf, (void __user *)to, size);if (bytes != size)
@@ -440,6 +453,10 @@ static int intel_pmu_pebs_fixup_ip(strucreturn 1 ; } +/* +* Even though we decoded the basic block, the instruction stream +* never matched the given IP, either the TO or the IP got corrupted. +*/ return 0 ; }
[tip:perf/pebs] perf, x86: Robustify PEBS fixup
Commit-ID: a562b1871f7f7d2f3a835c3c1e07fa58d473cfb7 Gitweb: http://git.kernel.org/tip/a562b1871f7f7d2f3a835c3c1e07fa58d473cfb7 Author: Peter Zijlstra <[email protected]> AuthorDate: Fri, 5 Mar 2010 16:29:14 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Wed, 10 Mar 2010 13:23:35 +0100
perf, x86: Robustify PEBS fixup
It turns out the LBR is massively unreliable on certain CPUs, so code the fixup a little more defensive to avoid crashing the kernel.
LKML-Reference: <[email protected]>
--- arch/x86/kernel/cpu/perf_event_intel_ds.c | 21 +++++++++++++++++++-- 1 files changed, 19 insertions(+), 2 deletions(-)
arch/x86/kernel/cpu/perf_event_intel_ds.c +19 -2
@@ -399,10 +399,23 @@ static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)if (!x86_pmu.intel_cap.pebs_trap)return 1 ; +/* +* No LBR entry, no basic block, no rewinding +*/ if (!cpuc->lbr_stack.nr || !from || !to)return 0 ; -if (ip < to) +/* +* Basic blocks should never cross user/kernel boundaries +*/ +if (kernel_ip (ip) !=kernel_ip (to)) +return 0 ; + +/* +* unsigned math, either ip is before the start (impossible) or +* the basic block is larger than 1 page (sanity) +*/ +if ((ip - to) >PAGE_SIZE )return 0 ;/*
@@ -420,7 +433,7 @@ static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs) old_to = to;if (!kernel_ip (ip)) { -int bytes, size =min_t (int,MAX_INSN_SIZE , ip - to); +int bytes, size =MAX_INSN_SIZE ; bytes =copy_from_user_nmi (buf, (void __user *)to, size);if (bytes != size)
@@ -440,6 +453,10 @@ static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)return 1 ; } +/* +* Even though we decoded the basic block, the instruction stream +* never matched the given IP, either the TO or the IP got corrupted. +*/ return 0 ; }