2 * Performance counter x86 architecture code
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
6 * Copyright(C) 2009 Jaswinder Singh Rajput
7 * Copyright(C) 2009 Advanced Micro Devices, Inc., Robert Richter
9 * For licencing details see kernel-base/COPYING
12 #include <linux/perf_counter.h>
13 #include <linux/capability.h>
14 #include <linux/notifier.h>
15 #include <linux/hardirq.h>
16 #include <linux/kprobes.h>
17 #include <linux/module.h>
18 #include <linux/kdebug.h>
19 #include <linux/sched.h>
20 #include <linux/uaccess.h>
23 #include <asm/stacktrace.h>
26 static bool perf_counters_initialized __read_mostly
;
29 * Number of (generic) HW counters:
31 static int nr_counters_generic __read_mostly
;
32 static u64 perf_counter_mask __read_mostly
;
33 static u64 counter_value_mask __read_mostly
;
34 static int counter_value_bits __read_mostly
;
36 static int nr_counters_fixed __read_mostly
;
38 struct cpu_hw_counters
{
39 struct perf_counter
*counters
[X86_PMC_IDX_MAX
];
40 unsigned long used
[BITS_TO_LONGS(X86_PMC_IDX_MAX
)];
41 unsigned long interrupts
;
43 unsigned long active_mask
[BITS_TO_LONGS(X86_PMC_IDX_MAX
)];
48 * struct x86_pmu - generic x86 pmu
51 int (*handle_irq
)(struct pt_regs
*, int);
52 u64 (*save_disable_all
)(void);
53 void (*restore_all
)(u64
);
54 void (*enable
)(int, u64
);
55 void (*disable
)(int, u64
);
58 u64 (*event_map
)(int);
59 u64 (*raw_event
)(u64
);
63 static struct x86_pmu x86_pmu __read_mostly
;
65 static DEFINE_PER_CPU(struct cpu_hw_counters
, cpu_hw_counters
) = {
69 static __read_mostly
int intel_perfmon_version
;
72 * Intel PerfMon v3. Used on Core2 and later.
74 static const u64 intel_perfmon_event_map
[] =
76 [PERF_COUNT_CPU_CYCLES
] = 0x003c,
77 [PERF_COUNT_INSTRUCTIONS
] = 0x00c0,
78 [PERF_COUNT_CACHE_REFERENCES
] = 0x4f2e,
79 [PERF_COUNT_CACHE_MISSES
] = 0x412e,
80 [PERF_COUNT_BRANCH_INSTRUCTIONS
] = 0x00c4,
81 [PERF_COUNT_BRANCH_MISSES
] = 0x00c5,
82 [PERF_COUNT_BUS_CYCLES
] = 0x013c,
85 static u64
intel_pmu_event_map(int event
)
87 return intel_perfmon_event_map
[event
];
90 static u64
intel_pmu_raw_event(u64 event
)
92 #define CORE_EVNTSEL_EVENT_MASK 0x000000FFULL
93 #define CORE_EVNTSEL_UNIT_MASK 0x0000FF00ULL
94 #define CORE_EVNTSEL_COUNTER_MASK 0xFF000000ULL
96 #define CORE_EVNTSEL_MASK \
97 (CORE_EVNTSEL_EVENT_MASK | \
98 CORE_EVNTSEL_UNIT_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_COUNTER_MASK 0x0FF000000ULL
128 #define K7_EVNTSEL_MASK \
129 (K7_EVNTSEL_EVENT_MASK | \
130 K7_EVNTSEL_UNIT_MASK | \
131 K7_EVNTSEL_COUNTER_MASK)
133 return event
& K7_EVNTSEL_MASK
;
137 * Propagate counter elapsed time into the generic counter.
138 * Can only be executed on the CPU where the counter is active.
139 * Returns the delta events processed.
142 x86_perf_counter_update(struct perf_counter
*counter
,
143 struct hw_perf_counter
*hwc
, int idx
)
145 u64 prev_raw_count
, new_raw_count
, delta
;
148 * Careful: an NMI might modify the previous counter value.
150 * Our tactic to handle this is to first atomically read and
151 * exchange a new raw count - then add that new-prev delta
152 * count to the generic counter atomically:
155 prev_raw_count
= atomic64_read(&hwc
->prev_count
);
156 rdmsrl(hwc
->counter_base
+ idx
, new_raw_count
);
158 if (atomic64_cmpxchg(&hwc
->prev_count
, prev_raw_count
,
159 new_raw_count
) != prev_raw_count
)
163 * Now we have the new raw value and have updated the prev
164 * timestamp already. We can now calculate the elapsed delta
165 * (counter-)time and add that to the generic counter.
167 * Careful, not all hw sign-extends above the physical width
168 * of the count, so we do that by clipping the delta to 32 bits:
170 delta
= (u64
)(u32
)((s32
)new_raw_count
- (s32
)prev_raw_count
);
172 atomic64_add(delta
, &counter
->count
);
173 atomic64_sub(delta
, &hwc
->period_left
);
176 static atomic_t num_counters
;
177 static DEFINE_MUTEX(pmc_reserve_mutex
);
179 static bool reserve_pmc_hardware(void)
183 if (nmi_watchdog
== NMI_LOCAL_APIC
)
184 disable_lapic_nmi_watchdog();
186 for (i
= 0; i
< nr_counters_generic
; i
++) {
187 if (!reserve_perfctr_nmi(x86_pmu
.perfctr
+ i
))
191 for (i
= 0; i
< nr_counters_generic
; i
++) {
192 if (!reserve_evntsel_nmi(x86_pmu
.eventsel
+ i
))
199 for (i
--; i
>= 0; i
--)
200 release_evntsel_nmi(x86_pmu
.eventsel
+ i
);
202 i
= nr_counters_generic
;
205 for (i
--; i
>= 0; i
--)
206 release_perfctr_nmi(x86_pmu
.perfctr
+ i
);
208 if (nmi_watchdog
== NMI_LOCAL_APIC
)
209 enable_lapic_nmi_watchdog();
214 static void release_pmc_hardware(void)
218 for (i
= 0; i
< nr_counters_generic
; i
++) {
219 release_perfctr_nmi(x86_pmu
.perfctr
+ i
);
220 release_evntsel_nmi(x86_pmu
.eventsel
+ i
);
223 if (nmi_watchdog
== NMI_LOCAL_APIC
)
224 enable_lapic_nmi_watchdog();
227 static void hw_perf_counter_destroy(struct perf_counter
*counter
)
229 if (atomic_dec_and_mutex_lock(&num_counters
, &pmc_reserve_mutex
)) {
230 release_pmc_hardware();
231 mutex_unlock(&pmc_reserve_mutex
);
236 * Setup the hardware configuration for a given hw_event_type
238 static int __hw_perf_counter_init(struct perf_counter
*counter
)
240 struct perf_counter_hw_event
*hw_event
= &counter
->hw_event
;
241 struct hw_perf_counter
*hwc
= &counter
->hw
;
244 /* disable temporarily */
245 if (boot_cpu_data
.x86_vendor
== X86_VENDOR_AMD
)
248 if (unlikely(!perf_counters_initialized
))
252 if (atomic_inc_not_zero(&num_counters
)) {
253 mutex_lock(&pmc_reserve_mutex
);
254 if (atomic_read(&num_counters
) == 0 && !reserve_pmc_hardware())
257 atomic_inc(&num_counters
);
258 mutex_unlock(&pmc_reserve_mutex
);
265 * (keep 'enabled' bit clear for now)
267 hwc
->config
= ARCH_PERFMON_EVENTSEL_INT
;
270 * Count user and OS events unless requested not to.
272 if (!hw_event
->exclude_user
)
273 hwc
->config
|= ARCH_PERFMON_EVENTSEL_USR
;
274 if (!hw_event
->exclude_kernel
)
275 hwc
->config
|= ARCH_PERFMON_EVENTSEL_OS
;
278 * If privileged enough, allow NMI events:
281 if (capable(CAP_SYS_ADMIN
) && hw_event
->nmi
)
284 hwc
->irq_period
= hw_event
->irq_period
;
286 * Intel PMCs cannot be accessed sanely above 32 bit width,
287 * so we install an artificial 1<<31 period regardless of
288 * the generic counter period:
290 if (boot_cpu_data
.x86_vendor
== X86_VENDOR_INTEL
)
291 if ((s64
)hwc
->irq_period
<= 0 || hwc
->irq_period
> 0x7FFFFFFF)
292 hwc
->irq_period
= 0x7FFFFFFF;
294 atomic64_set(&hwc
->period_left
, hwc
->irq_period
);
297 * Raw event type provide the config in the event structure
299 if (perf_event_raw(hw_event
)) {
300 hwc
->config
|= x86_pmu
.raw_event(perf_event_config(hw_event
));
302 if (perf_event_id(hw_event
) >= x86_pmu
.max_events
)
307 hwc
->config
|= x86_pmu
.event_map(perf_event_id(hw_event
));
310 counter
->destroy
= hw_perf_counter_destroy
;
315 static u64
intel_pmu_save_disable_all(void)
319 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, ctrl
);
320 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, 0);
325 static u64
amd_pmu_save_disable_all(void)
327 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
330 enabled
= cpuc
->enabled
;
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
< nr_counters_generic
; 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
);
354 u64
hw_perf_save_disable(void)
356 if (unlikely(!perf_counters_initialized
))
359 return x86_pmu
.save_disable_all();
362 * Exported because of ACPI idle
364 EXPORT_SYMBOL_GPL(hw_perf_save_disable
);
366 static void intel_pmu_restore_all(u64 ctrl
)
368 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, ctrl
);
371 static void amd_pmu_restore_all(u64 ctrl
)
373 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
376 cpuc
->enabled
= ctrl
;
381 for (idx
= 0; idx
< nr_counters_generic
; idx
++) {
384 if (!test_bit(idx
, cpuc
->active_mask
))
386 rdmsrl(MSR_K7_EVNTSEL0
+ idx
, val
);
387 if (val
& ARCH_PERFMON_EVENTSEL0_ENABLE
)
389 val
|= ARCH_PERFMON_EVENTSEL0_ENABLE
;
390 wrmsrl(MSR_K7_EVNTSEL0
+ idx
, val
);
394 void hw_perf_restore(u64 ctrl
)
396 if (unlikely(!perf_counters_initialized
))
399 x86_pmu
.restore_all(ctrl
);
402 * Exported because of ACPI idle
404 EXPORT_SYMBOL_GPL(hw_perf_restore
);
406 static inline u64
intel_pmu_get_status(u64 mask
)
410 if (unlikely(!perf_counters_initialized
))
412 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS
, status
);
417 static inline void intel_pmu_ack_status(u64 ack
)
419 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL
, ack
);
422 static void intel_pmu_enable_counter(int idx
, u64 config
)
424 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0
+ idx
,
425 config
| ARCH_PERFMON_EVENTSEL0_ENABLE
);
428 static void amd_pmu_enable_counter(int idx
, u64 config
)
430 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
432 set_bit(idx
, cpuc
->active_mask
);
434 config
|= ARCH_PERFMON_EVENTSEL0_ENABLE
;
436 wrmsrl(MSR_K7_EVNTSEL0
+ idx
, config
);
439 static void hw_perf_enable(int idx
, u64 config
)
441 if (unlikely(!perf_counters_initialized
))
444 x86_pmu
.enable(idx
, config
);
447 static void intel_pmu_disable_counter(int idx
, u64 config
)
449 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0
+ idx
, config
);
452 static void amd_pmu_disable_counter(int idx
, u64 config
)
454 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
456 clear_bit(idx
, cpuc
->active_mask
);
457 wrmsrl(MSR_K7_EVNTSEL0
+ idx
, config
);
461 static void hw_perf_disable(int idx
, u64 config
)
463 if (unlikely(!perf_counters_initialized
))
466 x86_pmu
.disable(idx
, config
);
470 __pmc_fixed_disable(struct perf_counter
*counter
,
471 struct hw_perf_counter
*hwc
, unsigned int __idx
)
473 int idx
= __idx
- X86_PMC_IDX_FIXED
;
477 mask
= 0xfULL
<< (idx
* 4);
479 rdmsrl(hwc
->config_base
, ctrl_val
);
481 err
= checking_wrmsrl(hwc
->config_base
, ctrl_val
);
485 __x86_pmu_disable(struct perf_counter
*counter
,
486 struct hw_perf_counter
*hwc
, unsigned int idx
)
488 if (unlikely(hwc
->config_base
== MSR_ARCH_PERFMON_FIXED_CTR_CTRL
))
489 __pmc_fixed_disable(counter
, hwc
, idx
);
491 hw_perf_disable(idx
, hwc
->config
);
494 static DEFINE_PER_CPU(u64
, prev_left
[X86_PMC_IDX_MAX
]);
497 * Set the next IRQ period, based on the hwc->period_left value.
498 * To be called with the counter disabled in hw:
501 x86_perf_counter_set_period(struct perf_counter
*counter
,
502 struct hw_perf_counter
*hwc
, int idx
)
504 s64 left
= atomic64_read(&hwc
->period_left
);
505 s64 period
= hwc
->irq_period
;
509 * If we are way outside a reasoable range then just skip forward:
511 if (unlikely(left
<= -period
)) {
513 atomic64_set(&hwc
->period_left
, left
);
516 if (unlikely(left
<= 0)) {
518 atomic64_set(&hwc
->period_left
, left
);
521 per_cpu(prev_left
[idx
], smp_processor_id()) = left
;
524 * The hw counter starts counting from this counter offset,
525 * mark it to be able to extra future deltas:
527 atomic64_set(&hwc
->prev_count
, (u64
)-left
);
529 err
= checking_wrmsrl(hwc
->counter_base
+ idx
,
530 (u64
)(-left
) & counter_value_mask
);
534 __pmc_fixed_enable(struct perf_counter
*counter
,
535 struct hw_perf_counter
*hwc
, unsigned int __idx
)
537 int idx
= __idx
- X86_PMC_IDX_FIXED
;
538 u64 ctrl_val
, bits
, mask
;
542 * Enable IRQ generation (0x8),
543 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
547 if (hwc
->config
& ARCH_PERFMON_EVENTSEL_USR
)
549 if (hwc
->config
& ARCH_PERFMON_EVENTSEL_OS
)
552 mask
= 0xfULL
<< (idx
* 4);
554 rdmsrl(hwc
->config_base
, ctrl_val
);
557 err
= checking_wrmsrl(hwc
->config_base
, ctrl_val
);
561 __x86_pmu_enable(struct perf_counter
*counter
,
562 struct hw_perf_counter
*hwc
, int idx
)
564 if (unlikely(hwc
->config_base
== MSR_ARCH_PERFMON_FIXED_CTR_CTRL
))
565 __pmc_fixed_enable(counter
, hwc
, idx
);
567 hw_perf_enable(idx
, hwc
->config
);
571 fixed_mode_idx(struct perf_counter
*counter
, struct hw_perf_counter
*hwc
)
575 if (boot_cpu_data
.x86_vendor
== X86_VENDOR_AMD
)
578 if (unlikely(hwc
->nmi
))
581 event
= hwc
->config
& ARCH_PERFMON_EVENT_MASK
;
583 if (unlikely(event
== x86_pmu
.event_map(PERF_COUNT_INSTRUCTIONS
)))
584 return X86_PMC_IDX_FIXED_INSTRUCTIONS
;
585 if (unlikely(event
== x86_pmu
.event_map(PERF_COUNT_CPU_CYCLES
)))
586 return X86_PMC_IDX_FIXED_CPU_CYCLES
;
587 if (unlikely(event
== x86_pmu
.event_map(PERF_COUNT_BUS_CYCLES
)))
588 return X86_PMC_IDX_FIXED_BUS_CYCLES
;
594 * Find a PMC slot for the freshly enabled / scheduled in counter:
596 static int x86_pmu_enable(struct perf_counter
*counter
)
598 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
599 struct hw_perf_counter
*hwc
= &counter
->hw
;
602 idx
= fixed_mode_idx(counter
, hwc
);
605 * Try to get the fixed counter, if that is already taken
606 * then try to get a generic counter:
608 if (test_and_set_bit(idx
, cpuc
->used
))
611 hwc
->config_base
= MSR_ARCH_PERFMON_FIXED_CTR_CTRL
;
613 * We set it so that counter_base + idx in wrmsr/rdmsr maps to
614 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
617 MSR_ARCH_PERFMON_FIXED_CTR0
- X86_PMC_IDX_FIXED
;
621 /* Try to get the previous generic counter again */
622 if (test_and_set_bit(idx
, cpuc
->used
)) {
624 idx
= find_first_zero_bit(cpuc
->used
, nr_counters_generic
);
625 if (idx
== nr_counters_generic
)
628 set_bit(idx
, cpuc
->used
);
631 hwc
->config_base
= x86_pmu
.eventsel
;
632 hwc
->counter_base
= x86_pmu
.perfctr
;
635 perf_counters_lapic_init(hwc
->nmi
);
637 __x86_pmu_disable(counter
, hwc
, idx
);
639 cpuc
->counters
[idx
] = counter
;
641 * Make it visible before enabling the hw:
645 x86_perf_counter_set_period(counter
, hwc
, idx
);
646 __x86_pmu_enable(counter
, hwc
, idx
);
651 void perf_counter_print_debug(void)
653 u64 ctrl
, status
, overflow
, pmc_ctrl
, pmc_count
, prev_left
, fixed
;
654 struct cpu_hw_counters
*cpuc
;
657 if (!nr_counters_generic
)
662 cpu
= smp_processor_id();
663 cpuc
= &per_cpu(cpu_hw_counters
, cpu
);
665 if (intel_perfmon_version
>= 2) {
666 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, ctrl
);
667 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS
, status
);
668 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL
, overflow
);
669 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL
, fixed
);
672 pr_info("CPU#%d: ctrl: %016llx\n", cpu
, ctrl
);
673 pr_info("CPU#%d: status: %016llx\n", cpu
, status
);
674 pr_info("CPU#%d: overflow: %016llx\n", cpu
, overflow
);
675 pr_info("CPU#%d: fixed: %016llx\n", cpu
, fixed
);
677 pr_info("CPU#%d: used: %016llx\n", cpu
, *(u64
*)cpuc
->used
);
679 for (idx
= 0; idx
< nr_counters_generic
; idx
++) {
680 rdmsrl(x86_pmu
.eventsel
+ idx
, pmc_ctrl
);
681 rdmsrl(x86_pmu
.perfctr
+ idx
, pmc_count
);
683 prev_left
= per_cpu(prev_left
[idx
], cpu
);
685 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
687 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
688 cpu
, idx
, pmc_count
);
689 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
690 cpu
, idx
, prev_left
);
692 for (idx
= 0; idx
< nr_counters_fixed
; idx
++) {
693 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0
+ idx
, pmc_count
);
695 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
696 cpu
, idx
, pmc_count
);
701 static void x86_pmu_disable(struct perf_counter
*counter
)
703 struct cpu_hw_counters
*cpuc
= &__get_cpu_var(cpu_hw_counters
);
704 struct hw_perf_counter
*hwc
= &counter
->hw
;
705 unsigned int idx
= hwc
->idx
;
707 __x86_pmu_disable(counter
, hwc
, idx
);
709 clear_bit(idx
, cpuc
->used
);
710 cpuc
->counters
[idx
] = NULL
;
712 * Make sure the cleared pointer becomes visible before we
713 * (potentially) free the counter:
718 * Drain the remaining delta count out of a counter
719 * that we are disabling:
721 x86_perf_counter_update(counter
, hwc
, idx
);
725 * Save and restart an expired counter. Called by NMI contexts,
726 * so it has to be careful about preempting normal counter ops:
728 static void intel_pmu_save_and_restart(struct perf_counter
*counter
)
730 struct hw_perf_counter
*hwc
= &counter
->hw
;
733 x86_perf_counter_update(counter
, hwc
, idx
);
734 x86_perf_counter_set_period(counter
, hwc
, idx
);
736 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
737 __x86_pmu_enable(counter
, hwc
, idx
);
741 * Maximum interrupt frequency of 100KHz per CPU
743 #define PERFMON_MAX_INTERRUPTS (100000/HZ)
746 * This handler is triggered by the local APIC, so the APIC IRQ handling
749 static int intel_pmu_handle_irq(struct pt_regs
*regs
, int nmi
)
751 int bit
, cpu
= smp_processor_id();
753 struct cpu_hw_counters
*cpuc
= &per_cpu(cpu_hw_counters
, cpu
);
756 cpuc
->throttle_ctrl
= intel_pmu_save_disable_all();
758 status
= intel_pmu_get_status(cpuc
->throttle_ctrl
);
764 inc_irq_stat(apic_perf_irqs
);
766 for_each_bit(bit
, (unsigned long *)&status
, X86_PMC_IDX_MAX
) {
767 struct perf_counter
*counter
= cpuc
->counters
[bit
];
769 clear_bit(bit
, (unsigned long *) &status
);
773 intel_pmu_save_and_restart(counter
);
774 if (perf_counter_overflow(counter
, nmi
, regs
, 0))
775 __x86_pmu_disable(counter
, &counter
->hw
, bit
);
778 intel_pmu_ack_status(ack
);
781 * Repeat if there is more work to be done:
783 status
= intel_pmu_get_status(cpuc
->throttle_ctrl
);
788 * Restore - do not reenable when global enable is off or throttled:
790 if (++cpuc
->interrupts
< PERFMON_MAX_INTERRUPTS
)
791 intel_pmu_restore_all(cpuc
->throttle_ctrl
);
796 static int amd_pmu_handle_irq(struct pt_regs
*regs
, int nmi
) { return 0; }
798 void perf_counter_unthrottle(void)
800 struct cpu_hw_counters
*cpuc
;
802 if (!cpu_has(&boot_cpu_data
, X86_FEATURE_ARCH_PERFMON
))
805 if (unlikely(!perf_counters_initialized
))
808 cpuc
= &__get_cpu_var(cpu_hw_counters
);
809 if (cpuc
->interrupts
>= PERFMON_MAX_INTERRUPTS
) {
810 if (printk_ratelimit())
811 printk(KERN_WARNING
"PERFMON: max interrupts exceeded!\n");
812 hw_perf_restore(cpuc
->throttle_ctrl
);
814 cpuc
->interrupts
= 0;
817 void smp_perf_counter_interrupt(struct pt_regs
*regs
)
820 apic_write(APIC_LVTPC
, LOCAL_PERF_VECTOR
);
822 x86_pmu
.handle_irq(regs
, 0);
826 void smp_perf_pending_interrupt(struct pt_regs
*regs
)
830 inc_irq_stat(apic_pending_irqs
);
831 perf_counter_do_pending();
835 void set_perf_counter_pending(void)
837 apic
->send_IPI_self(LOCAL_PENDING_VECTOR
);
840 void perf_counters_lapic_init(int nmi
)
844 if (!perf_counters_initialized
)
847 * Enable the performance counter vector in the APIC LVT:
849 apic_val
= apic_read(APIC_LVTERR
);
851 apic_write(APIC_LVTERR
, apic_val
| APIC_LVT_MASKED
);
853 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
855 apic_write(APIC_LVTPC
, LOCAL_PERF_VECTOR
);
856 apic_write(APIC_LVTERR
, apic_val
);
860 perf_counter_nmi_handler(struct notifier_block
*self
,
861 unsigned long cmd
, void *__args
)
863 struct die_args
*args
= __args
;
864 struct pt_regs
*regs
;
878 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
879 ret
= x86_pmu
.handle_irq(regs
, 1);
881 return ret
? NOTIFY_STOP
: NOTIFY_OK
;
884 static __read_mostly
struct notifier_block perf_counter_nmi_notifier
= {
885 .notifier_call
= perf_counter_nmi_handler
,
890 static struct x86_pmu intel_pmu
= {
891 .handle_irq
= intel_pmu_handle_irq
,
892 .save_disable_all
= intel_pmu_save_disable_all
,
893 .restore_all
= intel_pmu_restore_all
,
894 .enable
= intel_pmu_enable_counter
,
895 .disable
= intel_pmu_disable_counter
,
896 .eventsel
= MSR_ARCH_PERFMON_EVENTSEL0
,
897 .perfctr
= MSR_ARCH_PERFMON_PERFCTR0
,
898 .event_map
= intel_pmu_event_map
,
899 .raw_event
= intel_pmu_raw_event
,
900 .max_events
= ARRAY_SIZE(intel_perfmon_event_map
),
903 static struct x86_pmu amd_pmu
= {
904 .handle_irq
= amd_pmu_handle_irq
,
905 .save_disable_all
= amd_pmu_save_disable_all
,
906 .restore_all
= amd_pmu_restore_all
,
907 .enable
= amd_pmu_enable_counter
,
908 .disable
= amd_pmu_disable_counter
,
909 .eventsel
= MSR_K7_EVNTSEL0
,
910 .perfctr
= MSR_K7_PERFCTR0
,
911 .event_map
= amd_pmu_event_map
,
912 .raw_event
= amd_pmu_raw_event
,
913 .max_events
= ARRAY_SIZE(amd_perfmon_event_map
),
916 static int intel_pmu_init(void)
918 union cpuid10_edx edx
;
919 union cpuid10_eax eax
;
923 if (!cpu_has(&boot_cpu_data
, X86_FEATURE_ARCH_PERFMON
))
927 * Check whether the Architectural PerfMon supports
928 * Branch Misses Retired Event or not.
930 cpuid(10, &eax
.full
, &ebx
, &unused
, &edx
.full
);
931 if (eax
.split
.mask_length
<= ARCH_PERFMON_BRANCH_MISSES_RETIRED
)
934 intel_perfmon_version
= eax
.split
.version_id
;
935 if (intel_perfmon_version
< 2)
938 pr_info("Intel Performance Monitoring support detected.\n");
939 pr_info("... version: %d\n", intel_perfmon_version
);
940 pr_info("... bit width: %d\n", eax
.split
.bit_width
);
941 pr_info("... mask length: %d\n", eax
.split
.mask_length
);
945 nr_counters_generic
= eax
.split
.num_counters
;
946 nr_counters_fixed
= edx
.split
.num_counters_fixed
;
947 counter_value_mask
= (1ULL << eax
.split
.bit_width
) - 1;
952 static int amd_pmu_init(void)
956 nr_counters_generic
= 4;
957 nr_counters_fixed
= 0;
958 counter_value_mask
= 0x0000FFFFFFFFFFFFULL
;
959 counter_value_bits
= 48;
961 pr_info("AMD Performance Monitoring support detected.\n");
965 void __init
init_hw_perf_counters(void)
969 switch (boot_cpu_data
.x86_vendor
) {
970 case X86_VENDOR_INTEL
:
971 err
= intel_pmu_init();
974 err
= amd_pmu_init();
982 pr_info("... num counters: %d\n", nr_counters_generic
);
983 if (nr_counters_generic
> X86_PMC_MAX_GENERIC
) {
984 nr_counters_generic
= X86_PMC_MAX_GENERIC
;
985 WARN(1, KERN_ERR
"hw perf counters %d > max(%d), clipping!",
986 nr_counters_generic
, X86_PMC_MAX_GENERIC
);
988 perf_counter_mask
= (1 << nr_counters_generic
) - 1;
989 perf_max_counters
= nr_counters_generic
;
991 pr_info("... value mask: %016Lx\n", counter_value_mask
);
993 if (nr_counters_fixed
> X86_PMC_MAX_FIXED
) {
994 nr_counters_fixed
= X86_PMC_MAX_FIXED
;
995 WARN(1, KERN_ERR
"hw perf counters fixed %d > max(%d), clipping!",
996 nr_counters_fixed
, X86_PMC_MAX_FIXED
);
998 pr_info("... fixed counters: %d\n", nr_counters_fixed
);
1000 perf_counter_mask
|= ((1LL << nr_counters_fixed
)-1) << X86_PMC_IDX_FIXED
;
1002 pr_info("... counter mask: %016Lx\n", perf_counter_mask
);
1003 perf_counters_initialized
= true;
1005 perf_counters_lapic_init(0);
1006 register_die_notifier(&perf_counter_nmi_notifier
);
1009 static void x86_pmu_read(struct perf_counter
*counter
)
1011 x86_perf_counter_update(counter
, &counter
->hw
, counter
->hw
.idx
);
1014 static const struct pmu pmu
= {
1015 .enable
= x86_pmu_enable
,
1016 .disable
= x86_pmu_disable
,
1017 .read
= x86_pmu_read
,
1020 const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
1024 err
= __hw_perf_counter_init(counter
);
1026 return ERR_PTR(err
);
1036 void callchain_store(struct perf_callchain_entry
*entry
, unsigned long ip
)
1038 if (entry
->nr
< MAX_STACK_DEPTH
)
1039 entry
->ip
[entry
->nr
++] = ip
;
1042 static DEFINE_PER_CPU(struct perf_callchain_entry
, irq_entry
);
1043 static DEFINE_PER_CPU(struct perf_callchain_entry
, nmi_entry
);
1047 backtrace_warning_symbol(void *data
, char *msg
, unsigned long symbol
)
1049 /* Ignore warnings */
1052 static void backtrace_warning(void *data
, char *msg
)
1054 /* Ignore warnings */
1057 static int backtrace_stack(void *data
, char *name
)
1059 /* Don't bother with IRQ stacks for now */
1063 static void backtrace_address(void *data
, unsigned long addr
, int reliable
)
1065 struct perf_callchain_entry
*entry
= data
;
1068 callchain_store(entry
, addr
);
1071 static const struct stacktrace_ops backtrace_ops
= {
1072 .warning
= backtrace_warning
,
1073 .warning_symbol
= backtrace_warning_symbol
,
1074 .stack
= backtrace_stack
,
1075 .address
= backtrace_address
,
1079 perf_callchain_kernel(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1085 callchain_store(entry
, instruction_pointer(regs
));
1087 stack
= ((char *)regs
+ sizeof(struct pt_regs
));
1088 #ifdef CONFIG_FRAME_POINTER
1089 bp
= frame_pointer(regs
);
1094 dump_trace(NULL
, regs
, (void *)stack
, bp
, &backtrace_ops
, entry
);
1096 entry
->kernel
= entry
->nr
- nr
;
1100 struct stack_frame
{
1101 const void __user
*next_fp
;
1102 unsigned long return_address
;
1105 static int copy_stack_frame(const void __user
*fp
, struct stack_frame
*frame
)
1109 if (!access_ok(VERIFY_READ
, fp
, sizeof(*frame
)))
1113 pagefault_disable();
1114 if (__copy_from_user_inatomic(frame
, fp
, sizeof(*frame
)))
1122 perf_callchain_user(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1124 struct stack_frame frame
;
1125 const void __user
*fp
;
1128 regs
= (struct pt_regs
*)current
->thread
.sp0
- 1;
1129 fp
= (void __user
*)regs
->bp
;
1131 callchain_store(entry
, regs
->ip
);
1133 while (entry
->nr
< MAX_STACK_DEPTH
) {
1134 frame
.next_fp
= NULL
;
1135 frame
.return_address
= 0;
1137 if (!copy_stack_frame(fp
, &frame
))
1140 if ((unsigned long)fp
< user_stack_pointer(regs
))
1143 callchain_store(entry
, frame
.return_address
);
1147 entry
->user
= entry
->nr
- nr
;
1151 perf_do_callchain(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1158 is_user
= user_mode(regs
);
1160 if (!current
|| current
->pid
== 0)
1163 if (is_user
&& current
->state
!= TASK_RUNNING
)
1167 perf_callchain_kernel(regs
, entry
);
1170 perf_callchain_user(regs
, entry
);
1173 struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1175 struct perf_callchain_entry
*entry
;
1178 entry
= &__get_cpu_var(nmi_entry
);
1180 entry
= &__get_cpu_var(irq_entry
);
1187 perf_do_callchain(regs
, entry
);