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
;
38 * struct x86_pmu - generic x86 pmu
43 int (*handle_irq
)(struct pt_regs
*, int);
44 void (*disable_all
)(void);
45 void (*enable_all
)(void);
46 void (*enable
)(struct hw_perf_counter
*, int);
47 void (*disable
)(struct hw_perf_counter
*, int);
50 u64 (*event_map
)(int);
51 u64 (*raw_event
)(u64
);
54 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_EDGE_MASK 0x00040000ULL
91 #define CORE_EVNTSEL_INV_MASK 0x00800000ULL
92 #define CORE_EVNTSEL_COUNTER_MASK 0xFF000000ULL
94 #define CORE_EVNTSEL_MASK \
95 (CORE_EVNTSEL_EVENT_MASK | \
96 CORE_EVNTSEL_UNIT_MASK | \
97 CORE_EVNTSEL_EDGE_MASK | \
98 CORE_EVNTSEL_INV_MASK | \
99 CORE_EVNTSEL_COUNTER_MASK)
101 return event
& CORE_EVNTSEL_MASK
;
105 * AMD Performance Monitor K7 and later.
107 static const u64 amd_perfmon_event_map
[] =
109 [PERF_COUNT_CPU_CYCLES
] = 0x0076,
110 [PERF_COUNT_INSTRUCTIONS
] = 0x00c0,
111 [PERF_COUNT_CACHE_REFERENCES
] = 0x0080,
112 [PERF_COUNT_CACHE_MISSES
] = 0x0081,
113 [PERF_COUNT_BRANCH_INSTRUCTIONS
] = 0x00c4,
114 [PERF_COUNT_BRANCH_MISSES
] = 0x00c5,
117 static u64
amd_pmu_event_map(int event
)
119 return amd_perfmon_event_map
[event
];
122 static u64
amd_pmu_raw_event(u64 event
)
124 #define K7_EVNTSEL_EVENT_MASK 0x7000000FFULL
125 #define K7_EVNTSEL_UNIT_MASK 0x00000FF00ULL
126 #define K7_EVNTSEL_EDGE_MASK 0x000040000ULL
127 #define K7_EVNTSEL_INV_MASK 0x000800000ULL
128 #define K7_EVNTSEL_COUNTER_MASK 0x0FF000000ULL
130 #define K7_EVNTSEL_MASK \
131 (K7_EVNTSEL_EVENT_MASK | \
132 K7_EVNTSEL_UNIT_MASK | \
133 K7_EVNTSEL_EDGE_MASK | \
134 K7_EVNTSEL_INV_MASK | \
135 K7_EVNTSEL_COUNTER_MASK)
137 return event
& K7_EVNTSEL_MASK
;
141 * Propagate counter elapsed time into the generic counter.
142 * Can only be executed on the CPU where the counter is active.
143 * Returns the delta events processed.
146 x86_perf_counter_update(struct perf_counter
*counter
,
147 struct hw_perf_counter
*hwc
, int idx
)
149 int shift
= 64 - x86_pmu
.counter_bits
;
150 u64 prev_raw_count
, new_raw_count
;
154 * Careful: an NMI might modify the previous counter value.
156 * Our tactic to handle this is to first atomically read and
157 * exchange a new raw count - then add that new-prev delta
158 * count to the generic counter atomically:
161 prev_raw_count
= atomic64_read(&hwc
->prev_count
);
162 rdmsrl(hwc
->counter_base
+ idx
, new_raw_count
);
164 if (atomic64_cmpxchg(&hwc
->prev_count
, prev_raw_count
,
165 new_raw_count
) != prev_raw_count
)
169 * Now we have the new raw value and have updated the prev
170 * timestamp already. We can now calculate the elapsed delta
171 * (counter-)time and add that to the generic counter.
173 * Careful, not all hw sign-extends above the physical width
176 delta
= (new_raw_count
<< shift
) - (prev_raw_count
<< shift
);
179 atomic64_add(delta
, &counter
->count
);
180 atomic64_sub(delta
, &hwc
->period_left
);
182 return new_raw_count
;
185 static atomic_t active_counters
;
186 static DEFINE_MUTEX(pmc_reserve_mutex
);
188 static bool reserve_pmc_hardware(void)
192 if (nmi_watchdog
== NMI_LOCAL_APIC
)
193 disable_lapic_nmi_watchdog();
195 for (i
= 0; i
< x86_pmu
.num_counters
; i
++) {
196 if (!reserve_perfctr_nmi(x86_pmu
.perfctr
+ i
))
200 for (i
= 0; i
< x86_pmu
.num_counters
; i
++) {
201 if (!reserve_evntsel_nmi(x86_pmu
.eventsel
+ i
))
208 for (i
--; i
>= 0; i
--)
209 release_evntsel_nmi(x86_pmu
.eventsel
+ i
);
211 i
= x86_pmu
.num_counters
;
214 for (i
--; i
>= 0; i
--)
215 release_perfctr_nmi(x86_pmu
.perfctr
+ i
);
217 if (nmi_watchdog
== NMI_LOCAL_APIC
)
218 enable_lapic_nmi_watchdog();
223 static void release_pmc_hardware(void)
227 for (i
= 0; i
< x86_pmu
.num_counters
; i
++) {
228 release_perfctr_nmi(x86_pmu
.perfctr
+ i
);
229 release_evntsel_nmi(x86_pmu
.eventsel
+ i
);
232 if (nmi_watchdog
== NMI_LOCAL_APIC
)
233 enable_lapic_nmi_watchdog();
236 static void hw_perf_counter_destroy(struct perf_counter
*counter
)
238 if (atomic_dec_and_mutex_lock(&active_counters
, &pmc_reserve_mutex
)) {
239 release_pmc_hardware();
240 mutex_unlock(&pmc_reserve_mutex
);
244 static inline int x86_pmu_initialized(void)
246 return x86_pmu
.handle_irq
!= NULL
;
250 * Setup the hardware configuration for a given hw_event_type
252 static int __hw_perf_counter_init(struct perf_counter
*counter
)
254 struct perf_counter_hw_event
*hw_event
= &counter
->hw_event
;
255 struct hw_perf_counter
*hwc
= &counter
->hw
;
258 if (!x86_pmu_initialized())
262 if (!atomic_inc_not_zero(&active_counters
)) {
263 mutex_lock(&pmc_reserve_mutex
);
264 if (atomic_read(&active_counters
) == 0 && !reserve_pmc_hardware())
267 atomic_inc(&active_counters
);
268 mutex_unlock(&pmc_reserve_mutex
);
275 * (keep 'enabled' bit clear for now)
277 hwc
->config
= ARCH_PERFMON_EVENTSEL_INT
;
280 * Count user and OS events unless requested not to.
282 if (!hw_event
->exclude_user
)
283 hwc
->config
|= ARCH_PERFMON_EVENTSEL_USR
;
284 if (!hw_event
->exclude_kernel
)
285 hwc
->config
|= ARCH_PERFMON_EVENTSEL_OS
;
288 * Use NMI events all the time:
293 if (!hwc
->irq_period
)
294 hwc
->irq_period
= x86_pmu
.max_period
;
296 atomic64_set(&hwc
->period_left
,
297 min(x86_pmu
.max_period
, hwc
->irq_period
));
300 * Raw event type provide the config in the event structure
302 if (perf_event_raw(hw_event
)) {
303 hwc
->config
|= x86_pmu
.raw_event(perf_event_config(hw_event
));
305 if (perf_event_id(hw_event
) >= x86_pmu
.max_events
)
310 hwc
->config
|= x86_pmu
.event_map(perf_event_id(hw_event
));
313 counter
->destroy
= hw_perf_counter_destroy
;
318 static void intel_pmu_disable_all(void)
320 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, 0);
323 static void amd_pmu_disable_all(void)
325 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
333 * ensure we write the disable before we start disabling the
334 * counters proper, so that amd_pmu_enable_counter() does the
339 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
342 if (!test_bit(idx
, cpuc
->active_mask
))
344 rdmsrl(MSR_K7_EVNTSEL0
+ idx
, val
);
345 if (!(val
& ARCH_PERFMON_EVENTSEL0_ENABLE
))
347 val
&= ~ARCH_PERFMON_EVENTSEL0_ENABLE
;
348 wrmsrl(MSR_K7_EVNTSEL0
+ idx
, val
);
352 void hw_perf_disable(void)
354 if (!x86_pmu_initialized())
356 return x86_pmu
.disable_all();
359 static void intel_pmu_enable_all(void)
361 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, x86_pmu
.intel_ctrl
);
364 static void amd_pmu_enable_all(void)
366 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
375 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
378 if (!test_bit(idx
, cpuc
->active_mask
))
380 rdmsrl(MSR_K7_EVNTSEL0
+ idx
, val
);
381 if (val
& ARCH_PERFMON_EVENTSEL0_ENABLE
)
383 val
|= ARCH_PERFMON_EVENTSEL0_ENABLE
;
384 wrmsrl(MSR_K7_EVNTSEL0
+ idx
, val
);
388 void hw_perf_enable(void)
390 if (!x86_pmu_initialized())
392 x86_pmu
.enable_all();
395 static inline u64
intel_pmu_get_status(void)
399 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS
, status
);
404 static inline void intel_pmu_ack_status(u64 ack
)
406 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL
, ack
);
409 static inline void x86_pmu_enable_counter(struct hw_perf_counter
*hwc
, int idx
)
412 err
= checking_wrmsrl(hwc
->config_base
+ idx
,
413 hwc
->config
| ARCH_PERFMON_EVENTSEL0_ENABLE
);
416 static inline void x86_pmu_disable_counter(struct hw_perf_counter
*hwc
, int idx
)
419 err
= checking_wrmsrl(hwc
->config_base
+ idx
,
424 intel_pmu_disable_fixed(struct hw_perf_counter
*hwc
, int __idx
)
426 int idx
= __idx
- X86_PMC_IDX_FIXED
;
430 mask
= 0xfULL
<< (idx
* 4);
432 rdmsrl(hwc
->config_base
, ctrl_val
);
434 err
= checking_wrmsrl(hwc
->config_base
, ctrl_val
);
438 intel_pmu_disable_counter(struct hw_perf_counter
*hwc
, int idx
)
440 if (unlikely(hwc
->config_base
== MSR_ARCH_PERFMON_FIXED_CTR_CTRL
)) {
441 intel_pmu_disable_fixed(hwc
, idx
);
445 x86_pmu_disable_counter(hwc
, idx
);
449 amd_pmu_disable_counter(struct hw_perf_counter
*hwc
, int idx
)
451 x86_pmu_disable_counter(hwc
, idx
);
454 static DEFINE_PER_CPU(u64
, prev_left
[X86_PMC_IDX_MAX
]);
457 * Set the next IRQ period, based on the hwc->period_left value.
458 * To be called with the counter disabled in hw:
461 x86_perf_counter_set_period(struct perf_counter
*counter
,
462 struct hw_perf_counter
*hwc
, int idx
)
464 s64 left
= atomic64_read(&hwc
->period_left
);
465 s64 period
= min(x86_pmu
.max_period
, hwc
->irq_period
);
469 * If we are way outside a reasoable range then just skip forward:
471 if (unlikely(left
<= -period
)) {
473 atomic64_set(&hwc
->period_left
, left
);
476 if (unlikely(left
<= 0)) {
478 atomic64_set(&hwc
->period_left
, left
);
481 * Quirk: certain CPUs dont like it if just 1 event is left:
483 if (unlikely(left
< 2))
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 event
= hwc
->config
& ARCH_PERFMON_EVENT_MASK
;
554 if (unlikely(event
== x86_pmu
.event_map(PERF_COUNT_INSTRUCTIONS
)))
555 return X86_PMC_IDX_FIXED_INSTRUCTIONS
;
556 if (unlikely(event
== x86_pmu
.event_map(PERF_COUNT_CPU_CYCLES
)))
557 return X86_PMC_IDX_FIXED_CPU_CYCLES
;
558 if (unlikely(event
== x86_pmu
.event_map(PERF_COUNT_BUS_CYCLES
)))
559 return X86_PMC_IDX_FIXED_BUS_CYCLES
;
565 * Find a PMC slot for the freshly enabled / scheduled in counter:
567 static int x86_pmu_enable(struct perf_counter
*counter
)
569 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
570 struct hw_perf_counter
*hwc
= &counter
->hw
;
573 idx
= fixed_mode_idx(counter
, hwc
);
576 * Try to get the fixed counter, if that is already taken
577 * then try to get a generic counter:
579 if (test_and_set_bit(idx
, cpuc
->used_mask
))
582 hwc
->config_base
= MSR_ARCH_PERFMON_FIXED_CTR_CTRL
;
584 * We set it so that counter_base + idx in wrmsr/rdmsr maps to
585 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
588 MSR_ARCH_PERFMON_FIXED_CTR0
- X86_PMC_IDX_FIXED
;
592 /* Try to get the previous generic counter again */
593 if (test_and_set_bit(idx
, cpuc
->used_mask
)) {
595 idx
= find_first_zero_bit(cpuc
->used_mask
,
596 x86_pmu
.num_counters
);
597 if (idx
== x86_pmu
.num_counters
)
600 set_bit(idx
, cpuc
->used_mask
);
603 hwc
->config_base
= x86_pmu
.eventsel
;
604 hwc
->counter_base
= x86_pmu
.perfctr
;
607 perf_counters_lapic_init();
609 x86_pmu
.disable(hwc
, idx
);
611 cpuc
->counters
[idx
] = counter
;
612 set_bit(idx
, cpuc
->active_mask
);
614 x86_perf_counter_set_period(counter
, hwc
, idx
);
615 x86_pmu
.enable(hwc
, idx
);
620 static void x86_pmu_unthrottle(struct perf_counter
*counter
)
622 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
623 struct hw_perf_counter
*hwc
= &counter
->hw
;
625 if (WARN_ON_ONCE(hwc
->idx
>= X86_PMC_IDX_MAX
||
626 cpuc
->counters
[hwc
->idx
] != counter
))
629 x86_pmu
.enable(hwc
, hwc
->idx
);
632 void perf_counter_print_debug(void)
634 u64 ctrl
, status
, overflow
, pmc_ctrl
, pmc_count
, prev_left
, fixed
;
635 struct cpu_hw_counters
*cpuc
;
639 if (!x86_pmu
.num_counters
)
642 local_irq_save(flags
);
644 cpu
= smp_processor_id();
645 cpuc
= &per_cpu(cpu_hw_counters
, cpu
);
647 if (x86_pmu
.version
>= 2) {
648 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, ctrl
);
649 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS
, status
);
650 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL
, overflow
);
651 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL
, fixed
);
654 pr_info("CPU#%d: ctrl: %016llx\n", cpu
, ctrl
);
655 pr_info("CPU#%d: status: %016llx\n", cpu
, status
);
656 pr_info("CPU#%d: overflow: %016llx\n", cpu
, overflow
);
657 pr_info("CPU#%d: fixed: %016llx\n", cpu
, fixed
);
659 pr_info("CPU#%d: used: %016llx\n", cpu
, *(u64
*)cpuc
->used_mask
);
661 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
662 rdmsrl(x86_pmu
.eventsel
+ idx
, pmc_ctrl
);
663 rdmsrl(x86_pmu
.perfctr
+ idx
, pmc_count
);
665 prev_left
= per_cpu(prev_left
[idx
], cpu
);
667 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
669 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
670 cpu
, idx
, pmc_count
);
671 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
672 cpu
, idx
, prev_left
);
674 for (idx
= 0; idx
< x86_pmu
.num_counters_fixed
; idx
++) {
675 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0
+ idx
, pmc_count
);
677 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
678 cpu
, idx
, pmc_count
);
680 local_irq_restore(flags
);
683 static void x86_pmu_disable(struct perf_counter
*counter
)
685 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
686 struct hw_perf_counter
*hwc
= &counter
->hw
;
690 * Must be done before we disable, otherwise the nmi handler
691 * could reenable again:
693 clear_bit(idx
, cpuc
->active_mask
);
694 x86_pmu
.disable(hwc
, idx
);
697 * Make sure the cleared pointer becomes visible before we
698 * (potentially) free the counter:
703 * Drain the remaining delta count out of a counter
704 * that we are disabling:
706 x86_perf_counter_update(counter
, hwc
, idx
);
707 cpuc
->counters
[idx
] = NULL
;
708 clear_bit(idx
, cpuc
->used_mask
);
712 * Save and restart an expired counter. Called by NMI contexts,
713 * so it has to be careful about preempting normal counter ops:
715 static void intel_pmu_save_and_restart(struct perf_counter
*counter
)
717 struct hw_perf_counter
*hwc
= &counter
->hw
;
720 x86_perf_counter_update(counter
, hwc
, idx
);
721 x86_perf_counter_set_period(counter
, hwc
, idx
);
723 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
724 intel_pmu_enable_counter(hwc
, idx
);
727 static void intel_pmu_reset(void)
732 if (!x86_pmu
.num_counters
)
735 local_irq_save(flags
);
737 printk("clearing PMU state on CPU#%d\n", smp_processor_id());
739 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
740 checking_wrmsrl(x86_pmu
.eventsel
+ idx
, 0ull);
741 checking_wrmsrl(x86_pmu
.perfctr
+ idx
, 0ull);
743 for (idx
= 0; idx
< x86_pmu
.num_counters_fixed
; idx
++) {
744 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0
+ idx
, 0ull);
747 local_irq_restore(flags
);
752 * This handler is triggered by the local APIC, so the APIC IRQ handling
755 static int intel_pmu_handle_irq(struct pt_regs
*regs
, int nmi
)
757 struct cpu_hw_counters
*cpuc
;
758 struct cpu_hw_counters
;
762 cpu
= smp_processor_id();
763 cpuc
= &per_cpu(cpu_hw_counters
, cpu
);
766 status
= intel_pmu_get_status();
775 WARN_ONCE(1, "perfcounters: irq loop stuck!\n");
776 perf_counter_print_debug();
782 inc_irq_stat(apic_perf_irqs
);
784 for_each_bit(bit
, (unsigned long *)&status
, X86_PMC_IDX_MAX
) {
785 struct perf_counter
*counter
= cpuc
->counters
[bit
];
787 clear_bit(bit
, (unsigned long *) &status
);
788 if (!test_bit(bit
, cpuc
->active_mask
))
791 intel_pmu_save_and_restart(counter
);
792 if (perf_counter_overflow(counter
, nmi
, regs
, 0))
793 intel_pmu_disable_counter(&counter
->hw
, bit
);
796 intel_pmu_ack_status(ack
);
799 * Repeat if there is more work to be done:
801 status
= intel_pmu_get_status();
810 static int amd_pmu_handle_irq(struct pt_regs
*regs
, int nmi
)
812 int cpu
, idx
, handled
= 0;
813 struct cpu_hw_counters
*cpuc
;
814 struct perf_counter
*counter
;
815 struct hw_perf_counter
*hwc
;
818 cpu
= smp_processor_id();
819 cpuc
= &per_cpu(cpu_hw_counters
, cpu
);
821 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
822 if (!test_bit(idx
, cpuc
->active_mask
))
825 counter
= cpuc
->counters
[idx
];
828 val
= x86_perf_counter_update(counter
, hwc
, idx
);
829 if (val
& (1ULL << (x86_pmu
.counter_bits
- 1)))
832 /* counter overflow */
833 x86_perf_counter_set_period(counter
, hwc
, idx
);
835 inc_irq_stat(apic_perf_irqs
);
836 if (perf_counter_overflow(counter
, nmi
, regs
, 0))
837 amd_pmu_disable_counter(hwc
, idx
);
843 void smp_perf_counter_interrupt(struct pt_regs
*regs
)
846 apic_write(APIC_LVTPC
, LOCAL_PERF_VECTOR
);
848 x86_pmu
.handle_irq(regs
, 0);
852 void smp_perf_pending_interrupt(struct pt_regs
*regs
)
856 inc_irq_stat(apic_pending_irqs
);
857 perf_counter_do_pending();
861 void set_perf_counter_pending(void)
863 apic
->send_IPI_self(LOCAL_PENDING_VECTOR
);
866 void perf_counters_lapic_init(void)
868 if (!x86_pmu_initialized())
872 * Always use NMI for PMU
874 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
878 perf_counter_nmi_handler(struct notifier_block
*self
,
879 unsigned long cmd
, void *__args
)
881 struct die_args
*args
= __args
;
882 struct pt_regs
*regs
;
884 if (!atomic_read(&active_counters
))
898 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
900 * Can't rely on the handled return value to say it was our NMI, two
901 * counters could trigger 'simultaneously' raising two back-to-back NMIs.
903 * If the first NMI handles both, the latter will be empty and daze
906 x86_pmu
.handle_irq(regs
, 1);
911 static __read_mostly
struct notifier_block perf_counter_nmi_notifier
= {
912 .notifier_call
= perf_counter_nmi_handler
,
917 static struct x86_pmu intel_pmu
= {
919 .handle_irq
= intel_pmu_handle_irq
,
920 .disable_all
= intel_pmu_disable_all
,
921 .enable_all
= intel_pmu_enable_all
,
922 .enable
= intel_pmu_enable_counter
,
923 .disable
= intel_pmu_disable_counter
,
924 .eventsel
= MSR_ARCH_PERFMON_EVENTSEL0
,
925 .perfctr
= MSR_ARCH_PERFMON_PERFCTR0
,
926 .event_map
= intel_pmu_event_map
,
927 .raw_event
= intel_pmu_raw_event
,
928 .max_events
= ARRAY_SIZE(intel_perfmon_event_map
),
930 * Intel PMCs cannot be accessed sanely above 32 bit width,
931 * so we install an artificial 1<<31 period regardless of
932 * the generic counter period:
934 .max_period
= (1ULL << 31) - 1,
937 static struct x86_pmu amd_pmu
= {
939 .handle_irq
= amd_pmu_handle_irq
,
940 .disable_all
= amd_pmu_disable_all
,
941 .enable_all
= amd_pmu_enable_all
,
942 .enable
= amd_pmu_enable_counter
,
943 .disable
= amd_pmu_disable_counter
,
944 .eventsel
= MSR_K7_EVNTSEL0
,
945 .perfctr
= MSR_K7_PERFCTR0
,
946 .event_map
= amd_pmu_event_map
,
947 .raw_event
= amd_pmu_raw_event
,
948 .max_events
= ARRAY_SIZE(amd_perfmon_event_map
),
951 .counter_mask
= (1ULL << 48) - 1,
952 /* use highest bit to detect overflow */
953 .max_period
= (1ULL << 47) - 1,
956 static int intel_pmu_init(void)
958 union cpuid10_edx edx
;
959 union cpuid10_eax eax
;
964 if (!cpu_has(&boot_cpu_data
, X86_FEATURE_ARCH_PERFMON
))
968 * Check whether the Architectural PerfMon supports
969 * Branch Misses Retired Event or not.
971 cpuid(10, &eax
.full
, &ebx
, &unused
, &edx
.full
);
972 if (eax
.split
.mask_length
<= ARCH_PERFMON_BRANCH_MISSES_RETIRED
)
975 version
= eax
.split
.version_id
;
980 x86_pmu
.version
= version
;
981 x86_pmu
.num_counters
= eax
.split
.num_counters
;
984 * Quirk: v2 perfmon does not report fixed-purpose counters, so
985 * assume at least 3 counters:
987 x86_pmu
.num_counters_fixed
= max((int)edx
.split
.num_counters_fixed
, 3);
989 x86_pmu
.counter_bits
= eax
.split
.bit_width
;
990 x86_pmu
.counter_mask
= (1ULL << eax
.split
.bit_width
) - 1;
992 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, x86_pmu
.intel_ctrl
);
997 static int amd_pmu_init(void)
1003 void __init
init_hw_perf_counters(void)
1007 switch (boot_cpu_data
.x86_vendor
) {
1008 case X86_VENDOR_INTEL
:
1009 err
= intel_pmu_init();
1011 case X86_VENDOR_AMD
:
1012 err
= amd_pmu_init();
1020 pr_info("%s Performance Monitoring support detected.\n", x86_pmu
.name
);
1021 pr_info("... version: %d\n", x86_pmu
.version
);
1022 pr_info("... bit width: %d\n", x86_pmu
.counter_bits
);
1024 pr_info("... num counters: %d\n", x86_pmu
.num_counters
);
1025 if (x86_pmu
.num_counters
> X86_PMC_MAX_GENERIC
) {
1026 x86_pmu
.num_counters
= X86_PMC_MAX_GENERIC
;
1027 WARN(1, KERN_ERR
"hw perf counters %d > max(%d), clipping!",
1028 x86_pmu
.num_counters
, X86_PMC_MAX_GENERIC
);
1030 perf_counter_mask
= (1 << x86_pmu
.num_counters
) - 1;
1031 perf_max_counters
= x86_pmu
.num_counters
;
1033 pr_info("... value mask: %016Lx\n", x86_pmu
.counter_mask
);
1034 pr_info("... max period: %016Lx\n", x86_pmu
.max_period
);
1036 if (x86_pmu
.num_counters_fixed
> X86_PMC_MAX_FIXED
) {
1037 x86_pmu
.num_counters_fixed
= X86_PMC_MAX_FIXED
;
1038 WARN(1, KERN_ERR
"hw perf counters fixed %d > max(%d), clipping!",
1039 x86_pmu
.num_counters_fixed
, X86_PMC_MAX_FIXED
);
1041 pr_info("... fixed counters: %d\n", x86_pmu
.num_counters_fixed
);
1043 perf_counter_mask
|=
1044 ((1LL << x86_pmu
.num_counters_fixed
)-1) << X86_PMC_IDX_FIXED
;
1046 pr_info("... counter mask: %016Lx\n", perf_counter_mask
);
1048 perf_counters_lapic_init();
1049 register_die_notifier(&perf_counter_nmi_notifier
);
1052 static inline void x86_pmu_read(struct perf_counter
*counter
)
1054 x86_perf_counter_update(counter
, &counter
->hw
, counter
->hw
.idx
);
1057 static const struct pmu pmu
= {
1058 .enable
= x86_pmu_enable
,
1059 .disable
= x86_pmu_disable
,
1060 .read
= x86_pmu_read
,
1061 .unthrottle
= x86_pmu_unthrottle
,
1064 const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
1068 err
= __hw_perf_counter_init(counter
);
1070 return ERR_PTR(err
);
1080 void callchain_store(struct perf_callchain_entry
*entry
, unsigned long ip
)
1082 if (entry
->nr
< MAX_STACK_DEPTH
)
1083 entry
->ip
[entry
->nr
++] = ip
;
1086 static DEFINE_PER_CPU(struct perf_callchain_entry
, irq_entry
);
1087 static DEFINE_PER_CPU(struct perf_callchain_entry
, nmi_entry
);
1091 backtrace_warning_symbol(void *data
, char *msg
, unsigned long symbol
)
1093 /* Ignore warnings */
1096 static void backtrace_warning(void *data
, char *msg
)
1098 /* Ignore warnings */
1101 static int backtrace_stack(void *data
, char *name
)
1103 /* Don't bother with IRQ stacks for now */
1107 static void backtrace_address(void *data
, unsigned long addr
, int reliable
)
1109 struct perf_callchain_entry
*entry
= data
;
1112 callchain_store(entry
, addr
);
1115 static const struct stacktrace_ops backtrace_ops
= {
1116 .warning
= backtrace_warning
,
1117 .warning_symbol
= backtrace_warning_symbol
,
1118 .stack
= backtrace_stack
,
1119 .address
= backtrace_address
,
1123 perf_callchain_kernel(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1129 callchain_store(entry
, instruction_pointer(regs
));
1131 stack
= ((char *)regs
+ sizeof(struct pt_regs
));
1132 #ifdef CONFIG_FRAME_POINTER
1133 bp
= frame_pointer(regs
);
1138 dump_trace(NULL
, regs
, (void *)stack
, bp
, &backtrace_ops
, entry
);
1140 entry
->kernel
= entry
->nr
- nr
;
1144 struct stack_frame
{
1145 const void __user
*next_fp
;
1146 unsigned long return_address
;
1149 static int copy_stack_frame(const void __user
*fp
, struct stack_frame
*frame
)
1153 if (!access_ok(VERIFY_READ
, fp
, sizeof(*frame
)))
1157 pagefault_disable();
1158 if (__copy_from_user_inatomic(frame
, fp
, sizeof(*frame
)))
1166 perf_callchain_user(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1168 struct stack_frame frame
;
1169 const void __user
*fp
;
1172 regs
= (struct pt_regs
*)current
->thread
.sp0
- 1;
1173 fp
= (void __user
*)regs
->bp
;
1175 callchain_store(entry
, regs
->ip
);
1177 while (entry
->nr
< MAX_STACK_DEPTH
) {
1178 frame
.next_fp
= NULL
;
1179 frame
.return_address
= 0;
1181 if (!copy_stack_frame(fp
, &frame
))
1184 if ((unsigned long)fp
< user_stack_pointer(regs
))
1187 callchain_store(entry
, frame
.return_address
);
1191 entry
->user
= entry
->nr
- nr
;
1195 perf_do_callchain(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1202 is_user
= user_mode(regs
);
1204 if (!current
|| current
->pid
== 0)
1207 if (is_user
&& current
->state
!= TASK_RUNNING
)
1211 perf_callchain_kernel(regs
, entry
);
1214 perf_callchain_user(regs
, entry
);
1217 struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1219 struct perf_callchain_entry
*entry
;
1222 entry
= &__get_cpu_var(nmi_entry
);
1224 entry
= &__get_cpu_var(irq_entry
);
1231 perf_do_callchain(regs
, entry
);