2 * Performance counter x86 architecture code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2009 Jaswinder Singh Rajput
7 * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
8 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
10 * For licencing details see kernel-base/COPYING
13 #include <linux/perf_counter.h>
14 #include <linux/capability.h>
15 #include <linux/notifier.h>
16 #include <linux/hardirq.h>
17 #include <linux/kprobes.h>
18 #include <linux/module.h>
19 #include <linux/kdebug.h>
20 #include <linux/sched.h>
21 #include <linux/uaccess.h>
24 #include <asm/stacktrace.h>
27 static u64 perf_counter_mask __read_mostly
;
29 struct cpu_hw_counters
{
30 struct perf_counter
*counters
[X86_PMC_IDX_MAX
];
31 unsigned long used_mask
[BITS_TO_LONGS(X86_PMC_IDX_MAX
)];
32 unsigned long active_mask
[BITS_TO_LONGS(X86_PMC_IDX_MAX
)];
33 unsigned long interrupts
;
39 * struct x86_pmu - generic x86 pmu
44 int (*handle_irq
)(struct pt_regs
*, int);
45 u64 (*save_disable_all
)(void);
46 void (*restore_all
)(u64
);
47 void (*enable
)(struct hw_perf_counter
*, int);
48 void (*disable
)(struct hw_perf_counter
*, int);
51 u64 (*event_map
)(int);
52 u64 (*raw_event
)(u64
);
55 int num_counters_fixed
;
61 static struct x86_pmu x86_pmu __read_mostly
;
63 static DEFINE_PER_CPU(struct cpu_hw_counters
, cpu_hw_counters
) = {
68 * Intel PerfMon v3. Used on Core2 and later.
70 static const u64 intel_perfmon_event_map
[] =
72 [PERF_COUNT_CPU_CYCLES
] = 0x003c,
73 [PERF_COUNT_INSTRUCTIONS
] = 0x00c0,
74 [PERF_COUNT_CACHE_REFERENCES
] = 0x4f2e,
75 [PERF_COUNT_CACHE_MISSES
] = 0x412e,
76 [PERF_COUNT_BRANCH_INSTRUCTIONS
] = 0x00c4,
77 [PERF_COUNT_BRANCH_MISSES
] = 0x00c5,
78 [PERF_COUNT_BUS_CYCLES
] = 0x013c,
81 static u64
intel_pmu_event_map(int event
)
83 return intel_perfmon_event_map
[event
];
86 static u64
intel_pmu_raw_event(u64 event
)
88 #define CORE_EVNTSEL_EVENT_MASK 0x000000FFULL
89 #define CORE_EVNTSEL_UNIT_MASK 0x0000FF00ULL
90 #define CORE_EVNTSEL_COUNTER_MASK 0xFF000000ULL
92 #define CORE_EVNTSEL_MASK \
93 (CORE_EVNTSEL_EVENT_MASK | \
94 CORE_EVNTSEL_UNIT_MASK | \
95 CORE_EVNTSEL_COUNTER_MASK)
97 return event
& CORE_EVNTSEL_MASK
;
101 * AMD Performance Monitor K7 and later.
103 static const u64 amd_perfmon_event_map
[] =
105 [PERF_COUNT_CPU_CYCLES
] = 0x0076,
106 [PERF_COUNT_INSTRUCTIONS
] = 0x00c0,
107 [PERF_COUNT_CACHE_REFERENCES
] = 0x0080,
108 [PERF_COUNT_CACHE_MISSES
] = 0x0081,
109 [PERF_COUNT_BRANCH_INSTRUCTIONS
] = 0x00c4,
110 [PERF_COUNT_BRANCH_MISSES
] = 0x00c5,
113 static u64
amd_pmu_event_map(int event
)
115 return amd_perfmon_event_map
[event
];
118 static u64
amd_pmu_raw_event(u64 event
)
120 #define K7_EVNTSEL_EVENT_MASK 0x7000000FFULL
121 #define K7_EVNTSEL_UNIT_MASK 0x00000FF00ULL
122 #define K7_EVNTSEL_COUNTER_MASK 0x0FF000000ULL
124 #define K7_EVNTSEL_MASK \
125 (K7_EVNTSEL_EVENT_MASK | \
126 K7_EVNTSEL_UNIT_MASK | \
127 K7_EVNTSEL_COUNTER_MASK)
129 return event
& K7_EVNTSEL_MASK
;
133 * Propagate counter elapsed time into the generic counter.
134 * Can only be executed on the CPU where the counter is active.
135 * Returns the delta events processed.
138 x86_perf_counter_update(struct perf_counter
*counter
,
139 struct hw_perf_counter
*hwc
, int idx
)
141 int shift
= 64 - x86_pmu
.counter_bits
;
142 u64 prev_raw_count
, new_raw_count
;
146 * Careful: an NMI might modify the previous counter value.
148 * Our tactic to handle this is to first atomically read and
149 * exchange a new raw count - then add that new-prev delta
150 * count to the generic counter atomically:
153 prev_raw_count
= atomic64_read(&hwc
->prev_count
);
154 rdmsrl(hwc
->counter_base
+ idx
, new_raw_count
);
156 if (atomic64_cmpxchg(&hwc
->prev_count
, prev_raw_count
,
157 new_raw_count
) != prev_raw_count
)
161 * Now we have the new raw value and have updated the prev
162 * timestamp already. We can now calculate the elapsed delta
163 * (counter-)time and add that to the generic counter.
165 * Careful, not all hw sign-extends above the physical width
168 delta
= (new_raw_count
<< shift
) - (prev_raw_count
<< shift
);
171 atomic64_add(delta
, &counter
->count
);
172 atomic64_sub(delta
, &hwc
->period_left
);
174 return new_raw_count
;
177 static atomic_t active_counters
;
178 static DEFINE_MUTEX(pmc_reserve_mutex
);
180 static bool reserve_pmc_hardware(void)
184 if (nmi_watchdog
== NMI_LOCAL_APIC
)
185 disable_lapic_nmi_watchdog();
187 for (i
= 0; i
< x86_pmu
.num_counters
; i
++) {
188 if (!reserve_perfctr_nmi(x86_pmu
.perfctr
+ i
))
192 for (i
= 0; i
< x86_pmu
.num_counters
; i
++) {
193 if (!reserve_evntsel_nmi(x86_pmu
.eventsel
+ i
))
200 for (i
--; i
>= 0; i
--)
201 release_evntsel_nmi(x86_pmu
.eventsel
+ i
);
203 i
= x86_pmu
.num_counters
;
206 for (i
--; i
>= 0; i
--)
207 release_perfctr_nmi(x86_pmu
.perfctr
+ i
);
209 if (nmi_watchdog
== NMI_LOCAL_APIC
)
210 enable_lapic_nmi_watchdog();
215 static void release_pmc_hardware(void)
219 for (i
= 0; i
< x86_pmu
.num_counters
; i
++) {
220 release_perfctr_nmi(x86_pmu
.perfctr
+ i
);
221 release_evntsel_nmi(x86_pmu
.eventsel
+ i
);
224 if (nmi_watchdog
== NMI_LOCAL_APIC
)
225 enable_lapic_nmi_watchdog();
228 static void hw_perf_counter_destroy(struct perf_counter
*counter
)
230 if (atomic_dec_and_mutex_lock(&active_counters
, &pmc_reserve_mutex
)) {
231 release_pmc_hardware();
232 mutex_unlock(&pmc_reserve_mutex
);
236 static inline int x86_pmu_initialized(void)
238 return x86_pmu
.handle_irq
!= NULL
;
242 * Setup the hardware configuration for a given hw_event_type
244 static int __hw_perf_counter_init(struct perf_counter
*counter
)
246 struct perf_counter_hw_event
*hw_event
= &counter
->hw_event
;
247 struct hw_perf_counter
*hwc
= &counter
->hw
;
250 if (!x86_pmu_initialized())
254 if (!atomic_inc_not_zero(&active_counters
)) {
255 mutex_lock(&pmc_reserve_mutex
);
256 if (atomic_read(&active_counters
) == 0 && !reserve_pmc_hardware())
259 atomic_inc(&active_counters
);
260 mutex_unlock(&pmc_reserve_mutex
);
267 * (keep 'enabled' bit clear for now)
269 hwc
->config
= ARCH_PERFMON_EVENTSEL_INT
;
272 * Count user and OS events unless requested not to.
274 if (!hw_event
->exclude_user
)
275 hwc
->config
|= ARCH_PERFMON_EVENTSEL_USR
;
276 if (!hw_event
->exclude_kernel
)
277 hwc
->config
|= ARCH_PERFMON_EVENTSEL_OS
;
280 * If privileged enough, allow NMI events:
283 if (capable(CAP_SYS_ADMIN
) && hw_event
->nmi
)
286 hwc
->irq_period
= hw_event
->irq_period
;
287 if ((s64
)hwc
->irq_period
<= 0 || hwc
->irq_period
> x86_pmu
.max_period
)
288 hwc
->irq_period
= x86_pmu
.max_period
;
290 atomic64_set(&hwc
->period_left
, hwc
->irq_period
);
293 * Raw event type provide the config in the event structure
295 if (perf_event_raw(hw_event
)) {
296 hwc
->config
|= x86_pmu
.raw_event(perf_event_config(hw_event
));
298 if (perf_event_id(hw_event
) >= x86_pmu
.max_events
)
303 hwc
->config
|= x86_pmu
.event_map(perf_event_id(hw_event
));
306 counter
->destroy
= hw_perf_counter_destroy
;
311 static u64
intel_pmu_save_disable_all(void)
315 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, ctrl
);
316 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, 0);
321 static u64
amd_pmu_save_disable_all(void)
323 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
326 enabled
= cpuc
->enabled
;
329 * ensure we write the disable before we start disabling the
330 * counters proper, so that amd_pmu_enable_counter() does the
335 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
338 if (!test_bit(idx
, cpuc
->active_mask
))
340 rdmsrl(MSR_K7_EVNTSEL0
+ idx
, val
);
341 if (!(val
& ARCH_PERFMON_EVENTSEL0_ENABLE
))
343 val
&= ~ARCH_PERFMON_EVENTSEL0_ENABLE
;
344 wrmsrl(MSR_K7_EVNTSEL0
+ idx
, val
);
350 u64
hw_perf_save_disable(void)
352 if (!x86_pmu_initialized())
354 return x86_pmu
.save_disable_all();
357 * Exported because of ACPI idle
359 EXPORT_SYMBOL_GPL(hw_perf_save_disable
);
361 static void intel_pmu_restore_all(u64 ctrl
)
363 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, ctrl
);
366 static void amd_pmu_restore_all(u64 ctrl
)
368 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
371 cpuc
->enabled
= ctrl
;
376 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
379 if (!test_bit(idx
, cpuc
->active_mask
))
381 rdmsrl(MSR_K7_EVNTSEL0
+ idx
, val
);
382 if (val
& ARCH_PERFMON_EVENTSEL0_ENABLE
)
384 val
|= ARCH_PERFMON_EVENTSEL0_ENABLE
;
385 wrmsrl(MSR_K7_EVNTSEL0
+ idx
, val
);
389 void hw_perf_restore(u64 ctrl
)
391 if (!x86_pmu_initialized())
393 x86_pmu
.restore_all(ctrl
);
396 * Exported because of ACPI idle
398 EXPORT_SYMBOL_GPL(hw_perf_restore
);
400 static inline u64
intel_pmu_get_status(void)
404 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS
, status
);
409 static inline void intel_pmu_ack_status(u64 ack
)
411 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL
, ack
);
414 static inline void x86_pmu_enable_counter(struct hw_perf_counter
*hwc
, int idx
)
417 err
= checking_wrmsrl(hwc
->config_base
+ idx
,
418 hwc
->config
| ARCH_PERFMON_EVENTSEL0_ENABLE
);
421 static inline void x86_pmu_disable_counter(struct hw_perf_counter
*hwc
, int idx
)
424 err
= checking_wrmsrl(hwc
->config_base
+ idx
,
429 intel_pmu_disable_fixed(struct hw_perf_counter
*hwc
, int __idx
)
431 int idx
= __idx
- X86_PMC_IDX_FIXED
;
435 mask
= 0xfULL
<< (idx
* 4);
437 rdmsrl(hwc
->config_base
, ctrl_val
);
439 err
= checking_wrmsrl(hwc
->config_base
, ctrl_val
);
443 intel_pmu_disable_counter(struct hw_perf_counter
*hwc
, int idx
)
445 if (unlikely(hwc
->config_base
== MSR_ARCH_PERFMON_FIXED_CTR_CTRL
)) {
446 intel_pmu_disable_fixed(hwc
, idx
);
450 x86_pmu_disable_counter(hwc
, idx
);
454 amd_pmu_disable_counter(struct hw_perf_counter
*hwc
, int idx
)
456 x86_pmu_disable_counter(hwc
, idx
);
459 static DEFINE_PER_CPU(u64
, prev_left
[X86_PMC_IDX_MAX
]);
462 * Set the next IRQ period, based on the hwc->period_left value.
463 * To be called with the counter disabled in hw:
466 x86_perf_counter_set_period(struct perf_counter
*counter
,
467 struct hw_perf_counter
*hwc
, int idx
)
469 s64 left
= atomic64_read(&hwc
->period_left
);
470 s64 period
= hwc
->irq_period
;
474 * If we are way outside a reasoable range then just skip forward:
476 if (unlikely(left
<= -period
)) {
478 atomic64_set(&hwc
->period_left
, left
);
481 if (unlikely(left
<= 0)) {
483 atomic64_set(&hwc
->period_left
, left
);
486 per_cpu(prev_left
[idx
], smp_processor_id()) = left
;
489 * The hw counter starts counting from this counter offset,
490 * mark it to be able to extra future deltas:
492 atomic64_set(&hwc
->prev_count
, (u64
)-left
);
494 err
= checking_wrmsrl(hwc
->counter_base
+ idx
,
495 (u64
)(-left
) & x86_pmu
.counter_mask
);
499 intel_pmu_enable_fixed(struct hw_perf_counter
*hwc
, int __idx
)
501 int idx
= __idx
- X86_PMC_IDX_FIXED
;
502 u64 ctrl_val
, bits
, mask
;
506 * Enable IRQ generation (0x8),
507 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
511 if (hwc
->config
& ARCH_PERFMON_EVENTSEL_USR
)
513 if (hwc
->config
& ARCH_PERFMON_EVENTSEL_OS
)
516 mask
= 0xfULL
<< (idx
* 4);
518 rdmsrl(hwc
->config_base
, ctrl_val
);
521 err
= checking_wrmsrl(hwc
->config_base
, ctrl_val
);
524 static void intel_pmu_enable_counter(struct hw_perf_counter
*hwc
, int idx
)
526 if (unlikely(hwc
->config_base
== MSR_ARCH_PERFMON_FIXED_CTR_CTRL
)) {
527 intel_pmu_enable_fixed(hwc
, idx
);
531 x86_pmu_enable_counter(hwc
, idx
);
534 static void amd_pmu_enable_counter(struct hw_perf_counter
*hwc
, int idx
)
536 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
539 x86_pmu_enable_counter(hwc
, idx
);
541 x86_pmu_disable_counter(hwc
, idx
);
545 fixed_mode_idx(struct perf_counter
*counter
, struct hw_perf_counter
*hwc
)
549 if (!x86_pmu
.num_counters_fixed
)
552 if (unlikely(hwc
->nmi
))
555 event
= hwc
->config
& ARCH_PERFMON_EVENT_MASK
;
557 if (unlikely(event
== x86_pmu
.event_map(PERF_COUNT_INSTRUCTIONS
)))
558 return X86_PMC_IDX_FIXED_INSTRUCTIONS
;
559 if (unlikely(event
== x86_pmu
.event_map(PERF_COUNT_CPU_CYCLES
)))
560 return X86_PMC_IDX_FIXED_CPU_CYCLES
;
561 if (unlikely(event
== x86_pmu
.event_map(PERF_COUNT_BUS_CYCLES
)))
562 return X86_PMC_IDX_FIXED_BUS_CYCLES
;
568 * Find a PMC slot for the freshly enabled / scheduled in counter:
570 static int x86_pmu_enable(struct perf_counter
*counter
)
572 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
573 struct hw_perf_counter
*hwc
= &counter
->hw
;
576 idx
= fixed_mode_idx(counter
, hwc
);
579 * Try to get the fixed counter, if that is already taken
580 * then try to get a generic counter:
582 if (test_and_set_bit(idx
, cpuc
->used_mask
))
585 hwc
->config_base
= MSR_ARCH_PERFMON_FIXED_CTR_CTRL
;
587 * We set it so that counter_base + idx in wrmsr/rdmsr maps to
588 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
591 MSR_ARCH_PERFMON_FIXED_CTR0
- X86_PMC_IDX_FIXED
;
595 /* Try to get the previous generic counter again */
596 if (test_and_set_bit(idx
, cpuc
->used_mask
)) {
598 idx
= find_first_zero_bit(cpuc
->used_mask
,
599 x86_pmu
.num_counters
);
600 if (idx
== x86_pmu
.num_counters
)
603 set_bit(idx
, cpuc
->used_mask
);
606 hwc
->config_base
= x86_pmu
.eventsel
;
607 hwc
->counter_base
= x86_pmu
.perfctr
;
610 perf_counters_lapic_init(hwc
->nmi
);
612 x86_pmu
.disable(hwc
, idx
);
614 cpuc
->counters
[idx
] = counter
;
615 set_bit(idx
, cpuc
->active_mask
);
617 x86_perf_counter_set_period(counter
, hwc
, idx
);
618 x86_pmu
.enable(hwc
, idx
);
623 void perf_counter_print_debug(void)
625 u64 ctrl
, status
, overflow
, pmc_ctrl
, pmc_count
, prev_left
, fixed
;
626 struct cpu_hw_counters
*cpuc
;
630 if (!x86_pmu
.num_counters
)
633 local_irq_save(flags
);
635 cpu
= smp_processor_id();
636 cpuc
= &per_cpu(cpu_hw_counters
, cpu
);
638 if (x86_pmu
.version
>= 2) {
639 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, ctrl
);
640 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS
, status
);
641 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL
, overflow
);
642 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL
, fixed
);
645 pr_info("CPU#%d: ctrl: %016llx\n", cpu
, ctrl
);
646 pr_info("CPU#%d: status: %016llx\n", cpu
, status
);
647 pr_info("CPU#%d: overflow: %016llx\n", cpu
, overflow
);
648 pr_info("CPU#%d: fixed: %016llx\n", cpu
, fixed
);
650 pr_info("CPU#%d: used: %016llx\n", cpu
, *(u64
*)cpuc
->used_mask
);
652 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
653 rdmsrl(x86_pmu
.eventsel
+ idx
, pmc_ctrl
);
654 rdmsrl(x86_pmu
.perfctr
+ idx
, pmc_count
);
656 prev_left
= per_cpu(prev_left
[idx
], cpu
);
658 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
660 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
661 cpu
, idx
, pmc_count
);
662 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
663 cpu
, idx
, prev_left
);
665 for (idx
= 0; idx
< x86_pmu
.num_counters_fixed
; idx
++) {
666 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0
+ idx
, pmc_count
);
668 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
669 cpu
, idx
, pmc_count
);
671 local_irq_restore(flags
);
674 static void x86_pmu_disable(struct perf_counter
*counter
)
676 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
677 struct hw_perf_counter
*hwc
= &counter
->hw
;
681 * Must be done before we disable, otherwise the nmi handler
682 * could reenable again:
684 clear_bit(idx
, cpuc
->active_mask
);
685 x86_pmu
.disable(hwc
, idx
);
688 * Make sure the cleared pointer becomes visible before we
689 * (potentially) free the counter:
694 * Drain the remaining delta count out of a counter
695 * that we are disabling:
697 x86_perf_counter_update(counter
, hwc
, idx
);
698 cpuc
->counters
[idx
] = NULL
;
699 clear_bit(idx
, cpuc
->used_mask
);
703 * Save and restart an expired counter. Called by NMI contexts,
704 * so it has to be careful about preempting normal counter ops:
706 static void intel_pmu_save_and_restart(struct perf_counter
*counter
)
708 struct hw_perf_counter
*hwc
= &counter
->hw
;
711 x86_perf_counter_update(counter
, hwc
, idx
);
712 x86_perf_counter_set_period(counter
, hwc
, idx
);
714 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
715 intel_pmu_enable_counter(hwc
, idx
);
719 * Maximum interrupt frequency of 100KHz per CPU
721 #define PERFMON_MAX_INTERRUPTS (100000/HZ)
724 * This handler is triggered by the local APIC, so the APIC IRQ handling
727 static int intel_pmu_handle_irq(struct pt_regs
*regs
, int nmi
)
729 int bit
, cpu
= smp_processor_id();
731 struct cpu_hw_counters
*cpuc
= &per_cpu(cpu_hw_counters
, cpu
);
734 cpuc
->throttle_ctrl
= intel_pmu_save_disable_all();
736 status
= intel_pmu_get_status();
742 inc_irq_stat(apic_perf_irqs
);
744 for_each_bit(bit
, (unsigned long *)&status
, X86_PMC_IDX_MAX
) {
745 struct perf_counter
*counter
= cpuc
->counters
[bit
];
747 clear_bit(bit
, (unsigned long *) &status
);
748 if (!test_bit(bit
, cpuc
->active_mask
))
751 intel_pmu_save_and_restart(counter
);
752 if (perf_counter_overflow(counter
, nmi
, regs
, 0))
753 intel_pmu_disable_counter(&counter
->hw
, bit
);
756 intel_pmu_ack_status(ack
);
759 * Repeat if there is more work to be done:
761 status
= intel_pmu_get_status();
766 * Restore - do not reenable when global enable is off or throttled:
768 if (cpuc
->throttle_ctrl
) {
769 if (++cpuc
->interrupts
< PERFMON_MAX_INTERRUPTS
) {
770 intel_pmu_restore_all(cpuc
->throttle_ctrl
);
772 pr_info("CPU#%d: perfcounters: max interrupt rate exceeded! Throttle on.\n", smp_processor_id());
779 static int amd_pmu_handle_irq(struct pt_regs
*regs
, int nmi
)
781 int cpu
= smp_processor_id();
782 struct cpu_hw_counters
*cpuc
= &per_cpu(cpu_hw_counters
, cpu
);
785 struct perf_counter
*counter
;
786 struct hw_perf_counter
*hwc
;
790 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
791 if (!test_bit(idx
, cpuc
->active_mask
))
793 counter
= cpuc
->counters
[idx
];
795 val
= x86_perf_counter_update(counter
, hwc
, idx
);
796 if (val
& (1ULL << (x86_pmu
.counter_bits
- 1)))
798 /* counter overflow */
799 x86_perf_counter_set_period(counter
, hwc
, idx
);
801 inc_irq_stat(apic_perf_irqs
);
802 if (perf_counter_overflow(counter
, nmi
, regs
, 0))
803 amd_pmu_disable_counter(hwc
, idx
);
804 else if (cpuc
->interrupts
>= PERFMON_MAX_INTERRUPTS
)
806 * do not reenable when throttled, but reload
809 amd_pmu_disable_counter(hwc
, idx
);
810 else if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
811 amd_pmu_enable_counter(hwc
, idx
);
816 void perf_counter_unthrottle(void)
818 struct cpu_hw_counters
*cpuc
;
820 if (!x86_pmu_initialized())
823 cpuc
= &__get_cpu_var(cpu_hw_counters
);
824 if (cpuc
->interrupts
>= PERFMON_MAX_INTERRUPTS
) {
825 pr_info("CPU#%d: perfcounters: throttle off.\n", smp_processor_id());
828 * Clear them before re-enabling irqs/NMIs again:
830 cpuc
->interrupts
= 0;
831 hw_perf_restore(cpuc
->throttle_ctrl
);
833 cpuc
->interrupts
= 0;
837 void smp_perf_counter_interrupt(struct pt_regs
*regs
)
840 apic_write(APIC_LVTPC
, LOCAL_PERF_VECTOR
);
842 x86_pmu
.handle_irq(regs
, 0);
846 void smp_perf_pending_interrupt(struct pt_regs
*regs
)
850 inc_irq_stat(apic_pending_irqs
);
851 perf_counter_do_pending();
855 void set_perf_counter_pending(void)
857 apic
->send_IPI_self(LOCAL_PENDING_VECTOR
);
860 void perf_counters_lapic_init(int nmi
)
864 if (!x86_pmu_initialized())
868 * Enable the performance counter vector in the APIC LVT:
870 apic_val
= apic_read(APIC_LVTERR
);
872 apic_write(APIC_LVTERR
, apic_val
| APIC_LVT_MASKED
);
874 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
876 apic_write(APIC_LVTPC
, LOCAL_PERF_VECTOR
);
877 apic_write(APIC_LVTERR
, apic_val
);
881 perf_counter_nmi_handler(struct notifier_block
*self
,
882 unsigned long cmd
, void *__args
)
884 struct die_args
*args
= __args
;
885 struct pt_regs
*regs
;
888 if (!atomic_read(&active_counters
))
902 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
903 ret
= x86_pmu
.handle_irq(regs
, 1);
905 return ret
? NOTIFY_STOP
: NOTIFY_OK
;
908 static __read_mostly
struct notifier_block perf_counter_nmi_notifier
= {
909 .notifier_call
= perf_counter_nmi_handler
,
914 static struct x86_pmu intel_pmu
= {
916 .handle_irq
= intel_pmu_handle_irq
,
917 .save_disable_all
= intel_pmu_save_disable_all
,
918 .restore_all
= intel_pmu_restore_all
,
919 .enable
= intel_pmu_enable_counter
,
920 .disable
= intel_pmu_disable_counter
,
921 .eventsel
= MSR_ARCH_PERFMON_EVENTSEL0
,
922 .perfctr
= MSR_ARCH_PERFMON_PERFCTR0
,
923 .event_map
= intel_pmu_event_map
,
924 .raw_event
= intel_pmu_raw_event
,
925 .max_events
= ARRAY_SIZE(intel_perfmon_event_map
),
927 * Intel PMCs cannot be accessed sanely above 32 bit width,
928 * so we install an artificial 1<<31 period regardless of
929 * the generic counter period:
931 .max_period
= (1ULL << 31) - 1,
934 static struct x86_pmu amd_pmu
= {
936 .handle_irq
= amd_pmu_handle_irq
,
937 .save_disable_all
= amd_pmu_save_disable_all
,
938 .restore_all
= amd_pmu_restore_all
,
939 .enable
= amd_pmu_enable_counter
,
940 .disable
= amd_pmu_disable_counter
,
941 .eventsel
= MSR_K7_EVNTSEL0
,
942 .perfctr
= MSR_K7_PERFCTR0
,
943 .event_map
= amd_pmu_event_map
,
944 .raw_event
= amd_pmu_raw_event
,
945 .max_events
= ARRAY_SIZE(amd_perfmon_event_map
),
948 .counter_mask
= (1ULL << 48) - 1,
949 /* use highest bit to detect overflow */
950 .max_period
= (1ULL << 47) - 1,
953 static int intel_pmu_init(void)
955 union cpuid10_edx edx
;
956 union cpuid10_eax eax
;
961 if (!cpu_has(&boot_cpu_data
, X86_FEATURE_ARCH_PERFMON
))
965 * Check whether the Architectural PerfMon supports
966 * Branch Misses Retired Event or not.
968 cpuid(10, &eax
.full
, &ebx
, &unused
, &edx
.full
);
969 if (eax
.split
.mask_length
<= ARCH_PERFMON_BRANCH_MISSES_RETIRED
)
972 version
= eax
.split
.version_id
;
977 x86_pmu
.version
= version
;
978 x86_pmu
.num_counters
= eax
.split
.num_counters
;
981 * Quirk: v2 perfmon does not report fixed-purpose counters, so
982 * assume at least 3 counters:
984 x86_pmu
.num_counters_fixed
= max((int)edx
.split
.num_counters_fixed
, 3);
986 x86_pmu
.counter_bits
= eax
.split
.bit_width
;
987 x86_pmu
.counter_mask
= (1ULL << eax
.split
.bit_width
) - 1;
992 static int amd_pmu_init(void)
998 void __init
init_hw_perf_counters(void)
1002 switch (boot_cpu_data
.x86_vendor
) {
1003 case X86_VENDOR_INTEL
:
1004 err
= intel_pmu_init();
1006 case X86_VENDOR_AMD
:
1007 err
= amd_pmu_init();
1015 pr_info("%s Performance Monitoring support detected.\n", x86_pmu
.name
);
1016 pr_info("... version: %d\n", x86_pmu
.version
);
1017 pr_info("... bit width: %d\n", x86_pmu
.counter_bits
);
1019 pr_info("... num counters: %d\n", x86_pmu
.num_counters
);
1020 if (x86_pmu
.num_counters
> X86_PMC_MAX_GENERIC
) {
1021 x86_pmu
.num_counters
= X86_PMC_MAX_GENERIC
;
1022 WARN(1, KERN_ERR
"hw perf counters %d > max(%d), clipping!",
1023 x86_pmu
.num_counters
, X86_PMC_MAX_GENERIC
);
1025 perf_counter_mask
= (1 << x86_pmu
.num_counters
) - 1;
1026 perf_max_counters
= x86_pmu
.num_counters
;
1028 pr_info("... value mask: %016Lx\n", x86_pmu
.counter_mask
);
1029 pr_info("... max period: %016Lx\n", x86_pmu
.max_period
);
1031 if (x86_pmu
.num_counters_fixed
> X86_PMC_MAX_FIXED
) {
1032 x86_pmu
.num_counters_fixed
= X86_PMC_MAX_FIXED
;
1033 WARN(1, KERN_ERR
"hw perf counters fixed %d > max(%d), clipping!",
1034 x86_pmu
.num_counters_fixed
, X86_PMC_MAX_FIXED
);
1036 pr_info("... fixed counters: %d\n", x86_pmu
.num_counters_fixed
);
1038 perf_counter_mask
|=
1039 ((1LL << x86_pmu
.num_counters_fixed
)-1) << X86_PMC_IDX_FIXED
;
1041 pr_info("... counter mask: %016Lx\n", perf_counter_mask
);
1043 perf_counters_lapic_init(0);
1044 register_die_notifier(&perf_counter_nmi_notifier
);
1047 static inline void x86_pmu_read(struct perf_counter
*counter
)
1049 x86_perf_counter_update(counter
, &counter
->hw
, counter
->hw
.idx
);
1052 static const struct pmu pmu
= {
1053 .enable
= x86_pmu_enable
,
1054 .disable
= x86_pmu_disable
,
1055 .read
= x86_pmu_read
,
1058 const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
1062 err
= __hw_perf_counter_init(counter
);
1064 return ERR_PTR(err
);
1074 void callchain_store(struct perf_callchain_entry
*entry
, unsigned long ip
)
1076 if (entry
->nr
< MAX_STACK_DEPTH
)
1077 entry
->ip
[entry
->nr
++] = ip
;
1080 static DEFINE_PER_CPU(struct perf_callchain_entry
, irq_entry
);
1081 static DEFINE_PER_CPU(struct perf_callchain_entry
, nmi_entry
);
1085 backtrace_warning_symbol(void *data
, char *msg
, unsigned long symbol
)
1087 /* Ignore warnings */
1090 static void backtrace_warning(void *data
, char *msg
)
1092 /* Ignore warnings */
1095 static int backtrace_stack(void *data
, char *name
)
1097 /* Don't bother with IRQ stacks for now */
1101 static void backtrace_address(void *data
, unsigned long addr
, int reliable
)
1103 struct perf_callchain_entry
*entry
= data
;
1106 callchain_store(entry
, addr
);
1109 static const struct stacktrace_ops backtrace_ops
= {
1110 .warning
= backtrace_warning
,
1111 .warning_symbol
= backtrace_warning_symbol
,
1112 .stack
= backtrace_stack
,
1113 .address
= backtrace_address
,
1117 perf_callchain_kernel(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1123 callchain_store(entry
, instruction_pointer(regs
));
1125 stack
= ((char *)regs
+ sizeof(struct pt_regs
));
1126 #ifdef CONFIG_FRAME_POINTER
1127 bp
= frame_pointer(regs
);
1132 dump_trace(NULL
, regs
, (void *)stack
, bp
, &backtrace_ops
, entry
);
1134 entry
->kernel
= entry
->nr
- nr
;
1138 struct stack_frame
{
1139 const void __user
*next_fp
;
1140 unsigned long return_address
;
1143 static int copy_stack_frame(const void __user
*fp
, struct stack_frame
*frame
)
1147 if (!access_ok(VERIFY_READ
, fp
, sizeof(*frame
)))
1151 pagefault_disable();
1152 if (__copy_from_user_inatomic(frame
, fp
, sizeof(*frame
)))
1160 perf_callchain_user(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1162 struct stack_frame frame
;
1163 const void __user
*fp
;
1166 regs
= (struct pt_regs
*)current
->thread
.sp0
- 1;
1167 fp
= (void __user
*)regs
->bp
;
1169 callchain_store(entry
, regs
->ip
);
1171 while (entry
->nr
< MAX_STACK_DEPTH
) {
1172 frame
.next_fp
= NULL
;
1173 frame
.return_address
= 0;
1175 if (!copy_stack_frame(fp
, &frame
))
1178 if ((unsigned long)fp
< user_stack_pointer(regs
))
1181 callchain_store(entry
, frame
.return_address
);
1185 entry
->user
= entry
->nr
- nr
;
1189 perf_do_callchain(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1196 is_user
= user_mode(regs
);
1198 if (!current
|| current
->pid
== 0)
1201 if (is_user
&& current
->state
!= TASK_RUNNING
)
1205 perf_callchain_kernel(regs
, entry
);
1208 perf_callchain_user(regs
, entry
);
1211 struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1213 struct perf_callchain_entry
*entry
;
1216 entry
= &__get_cpu_var(nmi_entry
);
1218 entry
= &__get_cpu_var(irq_entry
);
1225 perf_do_callchain(regs
, entry
);