2 * Performance events 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>
9 * Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
10 * Copyright (C) 2009 Google, Inc., Stephane Eranian
12 * For licencing details see kernel-base/COPYING
15 #include <linux/perf_event.h>
16 #include <linux/capability.h>
17 #include <linux/notifier.h>
18 #include <linux/hardirq.h>
19 #include <linux/kprobes.h>
20 #include <linux/module.h>
21 #include <linux/kdebug.h>
22 #include <linux/sched.h>
23 #include <linux/uaccess.h>
24 #include <linux/slab.h>
25 #include <linux/highmem.h>
26 #include <linux/cpu.h>
27 #include <linux/bitops.h>
30 #include <asm/stacktrace.h>
32 #include <asm/compat.h>
36 #define wrmsrl(msr, val) \
38 trace_printk("wrmsrl(%lx, %lx)\n", (unsigned long)(msr),\
39 (unsigned long)(val)); \
40 native_write_msr((msr), (u32)((u64)(val)), \
41 (u32)((u64)(val) >> 32)); \
46 * best effort, GUP based copy_from_user() that assumes IRQ or NMI context
49 copy_from_user_nmi(void *to
, const void __user
*from
, unsigned long n
)
51 unsigned long offset
, addr
= (unsigned long)from
;
52 int type
= in_nmi() ? KM_NMI
: KM_IRQ0
;
53 unsigned long size
, len
= 0;
59 ret
= __get_user_pages_fast(addr
, 1, 0, &page
);
63 offset
= addr
& (PAGE_SIZE
- 1);
64 size
= min(PAGE_SIZE
- offset
, n
- len
);
66 map
= kmap_atomic(page
, type
);
67 memcpy(to
, map
+offset
, size
);
68 kunmap_atomic(map
, type
);
80 struct event_constraint
{
82 unsigned long idxmsk
[BITS_TO_LONGS(X86_PMC_IDX_MAX
)];
91 int nb_id
; /* NorthBridge id */
92 int refcnt
; /* reference count */
93 struct perf_event
*owners
[X86_PMC_IDX_MAX
];
94 struct event_constraint event_constraints
[X86_PMC_IDX_MAX
];
97 #define MAX_LBR_ENTRIES 16
99 struct cpu_hw_events
{
101 * Generic x86 PMC bits
103 struct perf_event
*events
[X86_PMC_IDX_MAX
]; /* in counter order */
104 unsigned long active_mask
[BITS_TO_LONGS(X86_PMC_IDX_MAX
)];
109 int assign
[X86_PMC_IDX_MAX
]; /* event to counter assignment */
110 u64 tags
[X86_PMC_IDX_MAX
];
111 struct perf_event
*event_list
[X86_PMC_IDX_MAX
]; /* in enabled order */
114 * Intel DebugStore bits
116 struct debug_store
*ds
;
124 struct perf_branch_stack lbr_stack
;
125 struct perf_branch_entry lbr_entries
[MAX_LBR_ENTRIES
];
130 struct amd_nb
*amd_nb
;
133 #define __EVENT_CONSTRAINT(c, n, m, w) {\
134 { .idxmsk64 = (n) }, \
140 #define EVENT_CONSTRAINT(c, n, m) \
141 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n))
144 * Constraint on the Event code.
146 #define INTEL_EVENT_CONSTRAINT(c, n) \
147 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT)
150 * Constraint on the Event code + UMask + fixed-mask
152 * filter mask to validate fixed counter events.
153 * the following filters disqualify for fixed counters:
157 * The other filters are supported by fixed counters.
158 * The any-thread option is supported starting with v3.
160 #define FIXED_EVENT_CONSTRAINT(c, n) \
161 EVENT_CONSTRAINT(c, (1ULL << (32+n)), X86_RAW_EVENT_MASK)
164 * Constraint on the Event code + UMask
166 #define PEBS_EVENT_CONSTRAINT(c, n) \
167 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
169 #define EVENT_CONSTRAINT_END \
170 EVENT_CONSTRAINT(0, 0, 0)
172 #define for_each_event_constraint(e, c) \
173 for ((e) = (c); (e)->weight; (e)++)
175 union perf_capabilities
{
179 u64 pebs_arch_reg
: 1;
187 * struct x86_pmu - generic x86 pmu
191 * Generic x86 PMC bits
195 int (*handle_irq
)(struct pt_regs
*);
196 void (*disable_all
)(void);
197 void (*enable_all
)(int added
);
198 void (*enable
)(struct perf_event
*);
199 void (*disable
)(struct perf_event
*);
200 int (*hw_config
)(struct perf_event
*event
);
201 int (*schedule_events
)(struct cpu_hw_events
*cpuc
, int n
, int *assign
);
204 u64 (*event_map
)(int);
207 int num_counters_fixed
;
212 struct event_constraint
*
213 (*get_event_constraints
)(struct cpu_hw_events
*cpuc
,
214 struct perf_event
*event
);
216 void (*put_event_constraints
)(struct cpu_hw_events
*cpuc
,
217 struct perf_event
*event
);
218 struct event_constraint
*event_constraints
;
219 void (*quirks
)(void);
221 int (*cpu_prepare
)(int cpu
);
222 void (*cpu_starting
)(int cpu
);
223 void (*cpu_dying
)(int cpu
);
224 void (*cpu_dead
)(int cpu
);
227 * Intel Arch Perfmon v2+
230 union perf_capabilities intel_cap
;
233 * Intel DebugStore bits
236 int pebs_record_size
;
237 void (*drain_pebs
)(struct pt_regs
*regs
);
238 struct event_constraint
*pebs_constraints
;
243 unsigned long lbr_tos
, lbr_from
, lbr_to
; /* MSR base regs */
244 int lbr_nr
; /* hardware stack size */
247 static struct x86_pmu x86_pmu __read_mostly
;
249 static DEFINE_PER_CPU(struct cpu_hw_events
, cpu_hw_events
) = {
253 static int x86_perf_event_set_period(struct perf_event
*event
);
256 * Generalized hw caching related hw_event table, filled
257 * in on a per model basis. A value of 0 means
258 * 'not supported', -1 means 'hw_event makes no sense on
259 * this CPU', any other value means the raw hw_event
263 #define C(x) PERF_COUNT_HW_CACHE_##x
265 static u64 __read_mostly hw_cache_event_ids
266 [PERF_COUNT_HW_CACHE_MAX
]
267 [PERF_COUNT_HW_CACHE_OP_MAX
]
268 [PERF_COUNT_HW_CACHE_RESULT_MAX
];
271 * Propagate event elapsed time into the generic event.
272 * Can only be executed on the CPU where the event is active.
273 * Returns the delta events processed.
276 x86_perf_event_update(struct perf_event
*event
)
278 struct hw_perf_event
*hwc
= &event
->hw
;
279 int shift
= 64 - x86_pmu
.cntval_bits
;
280 u64 prev_raw_count
, new_raw_count
;
284 if (idx
== X86_PMC_IDX_FIXED_BTS
)
288 * Careful: an NMI might modify the previous event value.
290 * Our tactic to handle this is to first atomically read and
291 * exchange a new raw count - then add that new-prev delta
292 * count to the generic event atomically:
295 prev_raw_count
= atomic64_read(&hwc
->prev_count
);
296 rdmsrl(hwc
->event_base
+ idx
, new_raw_count
);
298 if (atomic64_cmpxchg(&hwc
->prev_count
, prev_raw_count
,
299 new_raw_count
) != prev_raw_count
)
303 * Now we have the new raw value and have updated the prev
304 * timestamp already. We can now calculate the elapsed delta
305 * (event-)time and add that to the generic event.
307 * Careful, not all hw sign-extends above the physical width
310 delta
= (new_raw_count
<< shift
) - (prev_raw_count
<< shift
);
313 atomic64_add(delta
, &event
->count
);
314 atomic64_sub(delta
, &hwc
->period_left
);
316 return new_raw_count
;
319 static atomic_t active_events
;
320 static DEFINE_MUTEX(pmc_reserve_mutex
);
322 #ifdef CONFIG_X86_LOCAL_APIC
324 static bool reserve_pmc_hardware(void)
328 if (nmi_watchdog
== NMI_LOCAL_APIC
)
329 disable_lapic_nmi_watchdog();
331 for (i
= 0; i
< x86_pmu
.num_counters
; i
++) {
332 if (!reserve_perfctr_nmi(x86_pmu
.perfctr
+ i
))
336 for (i
= 0; i
< x86_pmu
.num_counters
; i
++) {
337 if (!reserve_evntsel_nmi(x86_pmu
.eventsel
+ i
))
344 for (i
--; i
>= 0; i
--)
345 release_evntsel_nmi(x86_pmu
.eventsel
+ i
);
347 i
= x86_pmu
.num_counters
;
350 for (i
--; i
>= 0; i
--)
351 release_perfctr_nmi(x86_pmu
.perfctr
+ i
);
353 if (nmi_watchdog
== NMI_LOCAL_APIC
)
354 enable_lapic_nmi_watchdog();
359 static void release_pmc_hardware(void)
363 for (i
= 0; i
< x86_pmu
.num_counters
; i
++) {
364 release_perfctr_nmi(x86_pmu
.perfctr
+ i
);
365 release_evntsel_nmi(x86_pmu
.eventsel
+ i
);
368 if (nmi_watchdog
== NMI_LOCAL_APIC
)
369 enable_lapic_nmi_watchdog();
374 static bool reserve_pmc_hardware(void) { return true; }
375 static void release_pmc_hardware(void) {}
379 static int reserve_ds_buffers(void);
380 static void release_ds_buffers(void);
382 static void hw_perf_event_destroy(struct perf_event
*event
)
384 if (atomic_dec_and_mutex_lock(&active_events
, &pmc_reserve_mutex
)) {
385 release_pmc_hardware();
386 release_ds_buffers();
387 mutex_unlock(&pmc_reserve_mutex
);
391 static inline int x86_pmu_initialized(void)
393 return x86_pmu
.handle_irq
!= NULL
;
397 set_ext_hw_attr(struct hw_perf_event
*hwc
, struct perf_event_attr
*attr
)
399 unsigned int cache_type
, cache_op
, cache_result
;
402 config
= attr
->config
;
404 cache_type
= (config
>> 0) & 0xff;
405 if (cache_type
>= PERF_COUNT_HW_CACHE_MAX
)
408 cache_op
= (config
>> 8) & 0xff;
409 if (cache_op
>= PERF_COUNT_HW_CACHE_OP_MAX
)
412 cache_result
= (config
>> 16) & 0xff;
413 if (cache_result
>= PERF_COUNT_HW_CACHE_RESULT_MAX
)
416 val
= hw_cache_event_ids
[cache_type
][cache_op
][cache_result
];
429 static int x86_setup_perfctr(struct perf_event
*event
)
431 struct perf_event_attr
*attr
= &event
->attr
;
432 struct hw_perf_event
*hwc
= &event
->hw
;
435 if (!hwc
->sample_period
) {
436 hwc
->sample_period
= x86_pmu
.max_period
;
437 hwc
->last_period
= hwc
->sample_period
;
438 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
441 * If we have a PMU initialized but no APIC
442 * interrupts, we cannot sample hardware
443 * events (user-space has to fall back and
444 * sample via a hrtimer based software event):
450 if (attr
->type
== PERF_TYPE_RAW
)
453 if (attr
->type
== PERF_TYPE_HW_CACHE
)
454 return set_ext_hw_attr(hwc
, attr
);
456 if (attr
->config
>= x86_pmu
.max_events
)
462 config
= x86_pmu
.event_map(attr
->config
);
473 if ((attr
->config
== PERF_COUNT_HW_BRANCH_INSTRUCTIONS
) &&
474 (hwc
->sample_period
== 1)) {
475 /* BTS is not supported by this architecture. */
479 /* BTS is currently only allowed for user-mode. */
480 if (!attr
->exclude_kernel
)
484 hwc
->config
|= config
;
489 static int x86_pmu_hw_config(struct perf_event
*event
)
493 * (keep 'enabled' bit clear for now)
495 event
->hw
.config
= ARCH_PERFMON_EVENTSEL_INT
;
498 * Count user and OS events unless requested not to
500 if (!event
->attr
.exclude_user
)
501 event
->hw
.config
|= ARCH_PERFMON_EVENTSEL_USR
;
502 if (!event
->attr
.exclude_kernel
)
503 event
->hw
.config
|= ARCH_PERFMON_EVENTSEL_OS
;
505 if (event
->attr
.type
== PERF_TYPE_RAW
)
506 event
->hw
.config
|= event
->attr
.config
& X86_RAW_EVENT_MASK
;
508 return x86_setup_perfctr(event
);
512 * Setup the hardware configuration for a given attr_type
514 static int __hw_perf_event_init(struct perf_event
*event
)
518 if (!x86_pmu_initialized())
522 if (!atomic_inc_not_zero(&active_events
)) {
523 mutex_lock(&pmc_reserve_mutex
);
524 if (atomic_read(&active_events
) == 0) {
525 if (!reserve_pmc_hardware())
528 err
= reserve_ds_buffers();
530 release_pmc_hardware();
534 atomic_inc(&active_events
);
535 mutex_unlock(&pmc_reserve_mutex
);
540 event
->destroy
= hw_perf_event_destroy
;
543 event
->hw
.last_cpu
= -1;
544 event
->hw
.last_tag
= ~0ULL;
546 return x86_pmu
.hw_config(event
);
549 static void x86_pmu_disable_all(void)
551 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
554 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
557 if (!test_bit(idx
, cpuc
->active_mask
))
559 rdmsrl(x86_pmu
.eventsel
+ idx
, val
);
560 if (!(val
& ARCH_PERFMON_EVENTSEL_ENABLE
))
562 val
&= ~ARCH_PERFMON_EVENTSEL_ENABLE
;
563 wrmsrl(x86_pmu
.eventsel
+ idx
, val
);
567 void hw_perf_disable(void)
569 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
571 if (!x86_pmu_initialized())
581 x86_pmu
.disable_all();
584 static void x86_pmu_enable_all(int added
)
586 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
589 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
590 struct perf_event
*event
= cpuc
->events
[idx
];
593 if (!test_bit(idx
, cpuc
->active_mask
))
596 val
= event
->hw
.config
;
597 val
|= ARCH_PERFMON_EVENTSEL_ENABLE
;
598 wrmsrl(x86_pmu
.eventsel
+ idx
, val
);
602 static const struct pmu pmu
;
604 static inline int is_x86_event(struct perf_event
*event
)
606 return event
->pmu
== &pmu
;
609 static int x86_schedule_events(struct cpu_hw_events
*cpuc
, int n
, int *assign
)
611 struct event_constraint
*c
, *constraints
[X86_PMC_IDX_MAX
];
612 unsigned long used_mask
[BITS_TO_LONGS(X86_PMC_IDX_MAX
)];
613 int i
, j
, w
, wmax
, num
= 0;
614 struct hw_perf_event
*hwc
;
616 bitmap_zero(used_mask
, X86_PMC_IDX_MAX
);
618 for (i
= 0; i
< n
; i
++) {
619 c
= x86_pmu
.get_event_constraints(cpuc
, cpuc
->event_list
[i
]);
624 * fastpath, try to reuse previous register
626 for (i
= 0; i
< n
; i
++) {
627 hwc
= &cpuc
->event_list
[i
]->hw
;
634 /* constraint still honored */
635 if (!test_bit(hwc
->idx
, c
->idxmsk
))
638 /* not already used */
639 if (test_bit(hwc
->idx
, used_mask
))
642 __set_bit(hwc
->idx
, used_mask
);
644 assign
[i
] = hwc
->idx
;
653 bitmap_zero(used_mask
, X86_PMC_IDX_MAX
);
656 * weight = number of possible counters
658 * 1 = most constrained, only works on one counter
659 * wmax = least constrained, works on any counter
661 * assign events to counters starting with most
662 * constrained events.
664 wmax
= x86_pmu
.num_counters
;
667 * when fixed event counters are present,
668 * wmax is incremented by 1 to account
669 * for one more choice
671 if (x86_pmu
.num_counters_fixed
)
674 for (w
= 1, num
= n
; num
&& w
<= wmax
; w
++) {
676 for (i
= 0; num
&& i
< n
; i
++) {
678 hwc
= &cpuc
->event_list
[i
]->hw
;
683 for_each_set_bit(j
, c
->idxmsk
, X86_PMC_IDX_MAX
) {
684 if (!test_bit(j
, used_mask
))
688 if (j
== X86_PMC_IDX_MAX
)
691 __set_bit(j
, used_mask
);
700 * scheduling failed or is just a simulation,
701 * free resources if necessary
703 if (!assign
|| num
) {
704 for (i
= 0; i
< n
; i
++) {
705 if (x86_pmu
.put_event_constraints
)
706 x86_pmu
.put_event_constraints(cpuc
, cpuc
->event_list
[i
]);
709 return num
? -ENOSPC
: 0;
713 * dogrp: true if must collect siblings events (group)
714 * returns total number of events and error code
716 static int collect_events(struct cpu_hw_events
*cpuc
, struct perf_event
*leader
, bool dogrp
)
718 struct perf_event
*event
;
721 max_count
= x86_pmu
.num_counters
+ x86_pmu
.num_counters_fixed
;
723 /* current number of events already accepted */
726 if (is_x86_event(leader
)) {
729 cpuc
->event_list
[n
] = leader
;
735 list_for_each_entry(event
, &leader
->sibling_list
, group_entry
) {
736 if (!is_x86_event(event
) ||
737 event
->state
<= PERF_EVENT_STATE_OFF
)
743 cpuc
->event_list
[n
] = event
;
749 static inline void x86_assign_hw_event(struct perf_event
*event
,
750 struct cpu_hw_events
*cpuc
, int i
)
752 struct hw_perf_event
*hwc
= &event
->hw
;
754 hwc
->idx
= cpuc
->assign
[i
];
755 hwc
->last_cpu
= smp_processor_id();
756 hwc
->last_tag
= ++cpuc
->tags
[i
];
758 if (hwc
->idx
== X86_PMC_IDX_FIXED_BTS
) {
759 hwc
->config_base
= 0;
761 } else if (hwc
->idx
>= X86_PMC_IDX_FIXED
) {
762 hwc
->config_base
= MSR_ARCH_PERFMON_FIXED_CTR_CTRL
;
764 * We set it so that event_base + idx in wrmsr/rdmsr maps to
765 * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
768 MSR_ARCH_PERFMON_FIXED_CTR0
- X86_PMC_IDX_FIXED
;
770 hwc
->config_base
= x86_pmu
.eventsel
;
771 hwc
->event_base
= x86_pmu
.perfctr
;
775 static inline int match_prev_assignment(struct hw_perf_event
*hwc
,
776 struct cpu_hw_events
*cpuc
,
779 return hwc
->idx
== cpuc
->assign
[i
] &&
780 hwc
->last_cpu
== smp_processor_id() &&
781 hwc
->last_tag
== cpuc
->tags
[i
];
784 static int x86_pmu_start(struct perf_event
*event
);
785 static void x86_pmu_stop(struct perf_event
*event
);
787 void hw_perf_enable(void)
789 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
790 struct perf_event
*event
;
791 struct hw_perf_event
*hwc
;
792 int i
, added
= cpuc
->n_added
;
794 if (!x86_pmu_initialized())
801 int n_running
= cpuc
->n_events
- cpuc
->n_added
;
803 * apply assignment obtained either from
804 * hw_perf_group_sched_in() or x86_pmu_enable()
806 * step1: save events moving to new counters
807 * step2: reprogram moved events into new counters
809 for (i
= 0; i
< n_running
; i
++) {
810 event
= cpuc
->event_list
[i
];
814 * we can avoid reprogramming counter if:
815 * - assigned same counter as last time
816 * - running on same CPU as last time
817 * - no other event has used the counter since
819 if (hwc
->idx
== -1 ||
820 match_prev_assignment(hwc
, cpuc
, i
))
826 for (i
= 0; i
< cpuc
->n_events
; i
++) {
827 event
= cpuc
->event_list
[i
];
830 if (!match_prev_assignment(hwc
, cpuc
, i
))
831 x86_assign_hw_event(event
, cpuc
, i
);
832 else if (i
< n_running
)
835 x86_pmu_start(event
);
838 perf_events_lapic_init();
844 x86_pmu
.enable_all(added
);
847 static inline void __x86_pmu_enable_event(struct hw_perf_event
*hwc
,
850 wrmsrl(hwc
->config_base
+ hwc
->idx
, hwc
->config
| enable_mask
);
853 static inline void x86_pmu_disable_event(struct perf_event
*event
)
855 struct hw_perf_event
*hwc
= &event
->hw
;
857 wrmsrl(hwc
->config_base
+ hwc
->idx
, hwc
->config
);
860 static DEFINE_PER_CPU(u64
[X86_PMC_IDX_MAX
], pmc_prev_left
);
863 * Set the next IRQ period, based on the hwc->period_left value.
864 * To be called with the event disabled in hw:
867 x86_perf_event_set_period(struct perf_event
*event
)
869 struct hw_perf_event
*hwc
= &event
->hw
;
870 s64 left
= atomic64_read(&hwc
->period_left
);
871 s64 period
= hwc
->sample_period
;
872 int ret
= 0, idx
= hwc
->idx
;
874 if (idx
== X86_PMC_IDX_FIXED_BTS
)
878 * If we are way outside a reasonable range then just skip forward:
880 if (unlikely(left
<= -period
)) {
882 atomic64_set(&hwc
->period_left
, left
);
883 hwc
->last_period
= period
;
887 if (unlikely(left
<= 0)) {
889 atomic64_set(&hwc
->period_left
, left
);
890 hwc
->last_period
= period
;
894 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
896 if (unlikely(left
< 2))
899 if (left
> x86_pmu
.max_period
)
900 left
= x86_pmu
.max_period
;
902 per_cpu(pmc_prev_left
[idx
], smp_processor_id()) = left
;
905 * The hw event starts counting from this event offset,
906 * mark it to be able to extra future deltas:
908 atomic64_set(&hwc
->prev_count
, (u64
)-left
);
910 wrmsrl(hwc
->event_base
+ idx
,
911 (u64
)(-left
) & x86_pmu
.cntval_mask
);
913 perf_event_update_userpage(event
);
918 static void x86_pmu_enable_event(struct perf_event
*event
)
920 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
922 __x86_pmu_enable_event(&event
->hw
,
923 ARCH_PERFMON_EVENTSEL_ENABLE
);
927 * activate a single event
929 * The event is added to the group of enabled events
930 * but only if it can be scehduled with existing events.
932 * Called with PMU disabled. If successful and return value 1,
933 * then guaranteed to call perf_enable() and hw_perf_enable()
935 static int x86_pmu_enable(struct perf_event
*event
)
937 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
938 struct hw_perf_event
*hwc
;
939 int assign
[X86_PMC_IDX_MAX
];
945 n
= collect_events(cpuc
, event
, false);
949 ret
= x86_pmu
.schedule_events(cpuc
, n
, assign
);
953 * copy new assignment, now we know it is possible
954 * will be used by hw_perf_enable()
956 memcpy(cpuc
->assign
, assign
, n
*sizeof(int));
959 cpuc
->n_added
+= n
- n0
;
964 static int x86_pmu_start(struct perf_event
*event
)
966 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
967 int idx
= event
->hw
.idx
;
972 x86_perf_event_set_period(event
);
973 cpuc
->events
[idx
] = event
;
974 __set_bit(idx
, cpuc
->active_mask
);
975 x86_pmu
.enable(event
);
976 perf_event_update_userpage(event
);
981 static void x86_pmu_unthrottle(struct perf_event
*event
)
983 int ret
= x86_pmu_start(event
);
987 void perf_event_print_debug(void)
989 u64 ctrl
, status
, overflow
, pmc_ctrl
, pmc_count
, prev_left
, fixed
;
991 struct cpu_hw_events
*cpuc
;
995 if (!x86_pmu
.num_counters
)
998 local_irq_save(flags
);
1000 cpu
= smp_processor_id();
1001 cpuc
= &per_cpu(cpu_hw_events
, cpu
);
1003 if (x86_pmu
.version
>= 2) {
1004 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL
, ctrl
);
1005 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS
, status
);
1006 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL
, overflow
);
1007 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL
, fixed
);
1008 rdmsrl(MSR_IA32_PEBS_ENABLE
, pebs
);
1011 pr_info("CPU#%d: ctrl: %016llx\n", cpu
, ctrl
);
1012 pr_info("CPU#%d: status: %016llx\n", cpu
, status
);
1013 pr_info("CPU#%d: overflow: %016llx\n", cpu
, overflow
);
1014 pr_info("CPU#%d: fixed: %016llx\n", cpu
, fixed
);
1015 pr_info("CPU#%d: pebs: %016llx\n", cpu
, pebs
);
1017 pr_info("CPU#%d: active: %016llx\n", cpu
, *(u64
*)cpuc
->active_mask
);
1019 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
1020 rdmsrl(x86_pmu
.eventsel
+ idx
, pmc_ctrl
);
1021 rdmsrl(x86_pmu
.perfctr
+ idx
, pmc_count
);
1023 prev_left
= per_cpu(pmc_prev_left
[idx
], cpu
);
1025 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1026 cpu
, idx
, pmc_ctrl
);
1027 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1028 cpu
, idx
, pmc_count
);
1029 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1030 cpu
, idx
, prev_left
);
1032 for (idx
= 0; idx
< x86_pmu
.num_counters_fixed
; idx
++) {
1033 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0
+ idx
, pmc_count
);
1035 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1036 cpu
, idx
, pmc_count
);
1038 local_irq_restore(flags
);
1041 static void x86_pmu_stop(struct perf_event
*event
)
1043 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1044 struct hw_perf_event
*hwc
= &event
->hw
;
1047 if (!__test_and_clear_bit(idx
, cpuc
->active_mask
))
1050 x86_pmu
.disable(event
);
1053 * Drain the remaining delta count out of a event
1054 * that we are disabling:
1056 x86_perf_event_update(event
);
1058 cpuc
->events
[idx
] = NULL
;
1061 static void x86_pmu_disable(struct perf_event
*event
)
1063 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1066 x86_pmu_stop(event
);
1068 for (i
= 0; i
< cpuc
->n_events
; i
++) {
1069 if (event
== cpuc
->event_list
[i
]) {
1071 if (x86_pmu
.put_event_constraints
)
1072 x86_pmu
.put_event_constraints(cpuc
, event
);
1074 while (++i
< cpuc
->n_events
)
1075 cpuc
->event_list
[i
-1] = cpuc
->event_list
[i
];
1081 perf_event_update_userpage(event
);
1084 static int x86_pmu_handle_irq(struct pt_regs
*regs
)
1086 struct perf_sample_data data
;
1087 struct cpu_hw_events
*cpuc
;
1088 struct perf_event
*event
;
1089 struct hw_perf_event
*hwc
;
1090 int idx
, handled
= 0;
1093 perf_sample_data_init(&data
, 0);
1095 cpuc
= &__get_cpu_var(cpu_hw_events
);
1097 for (idx
= 0; idx
< x86_pmu
.num_counters
; idx
++) {
1098 if (!test_bit(idx
, cpuc
->active_mask
))
1101 event
= cpuc
->events
[idx
];
1104 val
= x86_perf_event_update(event
);
1105 if (val
& (1ULL << (x86_pmu
.cntval_bits
- 1)))
1112 data
.period
= event
->hw
.last_period
;
1114 if (!x86_perf_event_set_period(event
))
1117 if (perf_event_overflow(event
, 1, &data
, regs
))
1118 x86_pmu_stop(event
);
1122 inc_irq_stat(apic_perf_irqs
);
1127 void smp_perf_pending_interrupt(struct pt_regs
*regs
)
1131 inc_irq_stat(apic_pending_irqs
);
1132 perf_event_do_pending();
1136 void set_perf_event_pending(void)
1138 #ifdef CONFIG_X86_LOCAL_APIC
1139 if (!x86_pmu
.apic
|| !x86_pmu_initialized())
1142 apic
->send_IPI_self(LOCAL_PENDING_VECTOR
);
1146 void perf_events_lapic_init(void)
1148 if (!x86_pmu
.apic
|| !x86_pmu_initialized())
1152 * Always use NMI for PMU
1154 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
1157 static int __kprobes
1158 perf_event_nmi_handler(struct notifier_block
*self
,
1159 unsigned long cmd
, void *__args
)
1161 struct die_args
*args
= __args
;
1162 struct pt_regs
*regs
;
1164 if (!atomic_read(&active_events
))
1178 apic_write(APIC_LVTPC
, APIC_DM_NMI
);
1180 * Can't rely on the handled return value to say it was our NMI, two
1181 * events could trigger 'simultaneously' raising two back-to-back NMIs.
1183 * If the first NMI handles both, the latter will be empty and daze
1186 x86_pmu
.handle_irq(regs
);
1191 static __read_mostly
struct notifier_block perf_event_nmi_notifier
= {
1192 .notifier_call
= perf_event_nmi_handler
,
1197 static struct event_constraint unconstrained
;
1198 static struct event_constraint emptyconstraint
;
1200 static struct event_constraint
*
1201 x86_get_event_constraints(struct cpu_hw_events
*cpuc
, struct perf_event
*event
)
1203 struct event_constraint
*c
;
1205 if (x86_pmu
.event_constraints
) {
1206 for_each_event_constraint(c
, x86_pmu
.event_constraints
) {
1207 if ((event
->hw
.config
& c
->cmask
) == c
->code
)
1212 return &unconstrained
;
1215 static int x86_event_sched_in(struct perf_event
*event
,
1216 struct perf_cpu_context
*cpuctx
)
1220 event
->state
= PERF_EVENT_STATE_ACTIVE
;
1221 event
->oncpu
= smp_processor_id();
1222 event
->tstamp_running
+= event
->ctx
->time
- event
->tstamp_stopped
;
1224 if (!is_x86_event(event
))
1225 ret
= event
->pmu
->enable(event
);
1227 if (!ret
&& !is_software_event(event
))
1228 cpuctx
->active_oncpu
++;
1230 if (!ret
&& event
->attr
.exclusive
)
1231 cpuctx
->exclusive
= 1;
1236 static void x86_event_sched_out(struct perf_event
*event
,
1237 struct perf_cpu_context
*cpuctx
)
1239 event
->state
= PERF_EVENT_STATE_INACTIVE
;
1242 if (!is_x86_event(event
))
1243 event
->pmu
->disable(event
);
1245 event
->tstamp_running
-= event
->ctx
->time
- event
->tstamp_stopped
;
1247 if (!is_software_event(event
))
1248 cpuctx
->active_oncpu
--;
1250 if (event
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
1251 cpuctx
->exclusive
= 0;
1255 * Called to enable a whole group of events.
1256 * Returns 1 if the group was enabled, or -EAGAIN if it could not be.
1257 * Assumes the caller has disabled interrupts and has
1258 * frozen the PMU with hw_perf_save_disable.
1260 * called with PMU disabled. If successful and return value 1,
1261 * then guaranteed to call perf_enable() and hw_perf_enable()
1263 int hw_perf_group_sched_in(struct perf_event
*leader
,
1264 struct perf_cpu_context
*cpuctx
,
1265 struct perf_event_context
*ctx
)
1267 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
1268 struct perf_event
*sub
;
1269 int assign
[X86_PMC_IDX_MAX
];
1272 if (!x86_pmu_initialized())
1275 /* n0 = total number of events */
1276 n0
= collect_events(cpuc
, leader
, true);
1280 ret
= x86_pmu
.schedule_events(cpuc
, n0
, assign
);
1284 ret
= x86_event_sched_in(leader
, cpuctx
);
1289 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
1290 if (sub
->state
> PERF_EVENT_STATE_OFF
) {
1291 ret
= x86_event_sched_in(sub
, cpuctx
);
1298 * copy new assignment, now we know it is possible
1299 * will be used by hw_perf_enable()
1301 memcpy(cpuc
->assign
, assign
, n0
*sizeof(int));
1303 cpuc
->n_events
= n0
;
1304 cpuc
->n_added
+= n1
;
1305 ctx
->nr_active
+= n1
;
1308 * 1 means successful and events are active
1309 * This is not quite true because we defer
1310 * actual activation until hw_perf_enable() but
1311 * this way we* ensure caller won't try to enable
1316 x86_event_sched_out(leader
, cpuctx
);
1318 list_for_each_entry(sub
, &leader
->sibling_list
, group_entry
) {
1319 if (sub
->state
== PERF_EVENT_STATE_ACTIVE
) {
1320 x86_event_sched_out(sub
, cpuctx
);
1328 #include "perf_event_amd.c"
1329 #include "perf_event_p6.c"
1330 #include "perf_event_p4.c"
1331 #include "perf_event_intel_lbr.c"
1332 #include "perf_event_intel_ds.c"
1333 #include "perf_event_intel.c"
1335 static int __cpuinit
1336 x86_pmu_notifier(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
1338 unsigned int cpu
= (long)hcpu
;
1339 int ret
= NOTIFY_OK
;
1341 switch (action
& ~CPU_TASKS_FROZEN
) {
1342 case CPU_UP_PREPARE
:
1343 if (x86_pmu
.cpu_prepare
)
1344 ret
= x86_pmu
.cpu_prepare(cpu
);
1348 if (x86_pmu
.cpu_starting
)
1349 x86_pmu
.cpu_starting(cpu
);
1353 if (x86_pmu
.cpu_dying
)
1354 x86_pmu
.cpu_dying(cpu
);
1357 case CPU_UP_CANCELED
:
1359 if (x86_pmu
.cpu_dead
)
1360 x86_pmu
.cpu_dead(cpu
);
1370 static void __init
pmu_check_apic(void)
1376 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1377 pr_info("no hardware sampling interrupt available.\n");
1380 void __init
init_hw_perf_events(void)
1382 struct event_constraint
*c
;
1385 pr_info("Performance Events: ");
1387 switch (boot_cpu_data
.x86_vendor
) {
1388 case X86_VENDOR_INTEL
:
1389 err
= intel_pmu_init();
1391 case X86_VENDOR_AMD
:
1392 err
= amd_pmu_init();
1398 pr_cont("no PMU driver, software events only.\n");
1404 pr_cont("%s PMU driver.\n", x86_pmu
.name
);
1409 if (x86_pmu
.num_counters
> X86_PMC_MAX_GENERIC
) {
1410 WARN(1, KERN_ERR
"hw perf events %d > max(%d), clipping!",
1411 x86_pmu
.num_counters
, X86_PMC_MAX_GENERIC
);
1412 x86_pmu
.num_counters
= X86_PMC_MAX_GENERIC
;
1414 x86_pmu
.intel_ctrl
= (1 << x86_pmu
.num_counters
) - 1;
1415 perf_max_events
= x86_pmu
.num_counters
;
1417 if (x86_pmu
.num_counters_fixed
> X86_PMC_MAX_FIXED
) {
1418 WARN(1, KERN_ERR
"hw perf events fixed %d > max(%d), clipping!",
1419 x86_pmu
.num_counters_fixed
, X86_PMC_MAX_FIXED
);
1420 x86_pmu
.num_counters_fixed
= X86_PMC_MAX_FIXED
;
1423 x86_pmu
.intel_ctrl
|=
1424 ((1LL << x86_pmu
.num_counters_fixed
)-1) << X86_PMC_IDX_FIXED
;
1426 perf_events_lapic_init();
1427 register_die_notifier(&perf_event_nmi_notifier
);
1429 unconstrained
= (struct event_constraint
)
1430 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu
.num_counters
) - 1,
1431 0, x86_pmu
.num_counters
);
1433 if (x86_pmu
.event_constraints
) {
1434 for_each_event_constraint(c
, x86_pmu
.event_constraints
) {
1435 if (c
->cmask
!= X86_RAW_EVENT_MASK
)
1438 c
->idxmsk64
|= (1ULL << x86_pmu
.num_counters
) - 1;
1439 c
->weight
+= x86_pmu
.num_counters
;
1443 pr_info("... version: %d\n", x86_pmu
.version
);
1444 pr_info("... bit width: %d\n", x86_pmu
.cntval_bits
);
1445 pr_info("... generic registers: %d\n", x86_pmu
.num_counters
);
1446 pr_info("... value mask: %016Lx\n", x86_pmu
.cntval_mask
);
1447 pr_info("... max period: %016Lx\n", x86_pmu
.max_period
);
1448 pr_info("... fixed-purpose events: %d\n", x86_pmu
.num_counters_fixed
);
1449 pr_info("... event mask: %016Lx\n", x86_pmu
.intel_ctrl
);
1451 perf_cpu_notifier(x86_pmu_notifier
);
1454 static inline void x86_pmu_read(struct perf_event
*event
)
1456 x86_perf_event_update(event
);
1459 static const struct pmu pmu
= {
1460 .enable
= x86_pmu_enable
,
1461 .disable
= x86_pmu_disable
,
1462 .start
= x86_pmu_start
,
1463 .stop
= x86_pmu_stop
,
1464 .read
= x86_pmu_read
,
1465 .unthrottle
= x86_pmu_unthrottle
,
1469 * validate that we can schedule this event
1471 static int validate_event(struct perf_event
*event
)
1473 struct cpu_hw_events
*fake_cpuc
;
1474 struct event_constraint
*c
;
1477 fake_cpuc
= kmalloc(sizeof(*fake_cpuc
), GFP_KERNEL
| __GFP_ZERO
);
1481 c
= x86_pmu
.get_event_constraints(fake_cpuc
, event
);
1483 if (!c
|| !c
->weight
)
1486 if (x86_pmu
.put_event_constraints
)
1487 x86_pmu
.put_event_constraints(fake_cpuc
, event
);
1495 * validate a single event group
1497 * validation include:
1498 * - check events are compatible which each other
1499 * - events do not compete for the same counter
1500 * - number of events <= number of counters
1502 * validation ensures the group can be loaded onto the
1503 * PMU if it was the only group available.
1505 static int validate_group(struct perf_event
*event
)
1507 struct perf_event
*leader
= event
->group_leader
;
1508 struct cpu_hw_events
*fake_cpuc
;
1512 fake_cpuc
= kmalloc(sizeof(*fake_cpuc
), GFP_KERNEL
| __GFP_ZERO
);
1517 * the event is not yet connected with its
1518 * siblings therefore we must first collect
1519 * existing siblings, then add the new event
1520 * before we can simulate the scheduling
1523 n
= collect_events(fake_cpuc
, leader
, true);
1527 fake_cpuc
->n_events
= n
;
1528 n
= collect_events(fake_cpuc
, event
, false);
1532 fake_cpuc
->n_events
= n
;
1534 ret
= x86_pmu
.schedule_events(fake_cpuc
, n
, NULL
);
1542 const struct pmu
*hw_perf_event_init(struct perf_event
*event
)
1544 const struct pmu
*tmp
;
1547 err
= __hw_perf_event_init(event
);
1550 * we temporarily connect event to its pmu
1551 * such that validate_group() can classify
1552 * it as an x86 event using is_x86_event()
1557 if (event
->group_leader
!= event
)
1558 err
= validate_group(event
);
1560 err
= validate_event(event
);
1566 event
->destroy(event
);
1567 return ERR_PTR(err
);
1578 void callchain_store(struct perf_callchain_entry
*entry
, u64 ip
)
1580 if (entry
->nr
< PERF_MAX_STACK_DEPTH
)
1581 entry
->ip
[entry
->nr
++] = ip
;
1584 static DEFINE_PER_CPU(struct perf_callchain_entry
, pmc_irq_entry
);
1585 static DEFINE_PER_CPU(struct perf_callchain_entry
, pmc_nmi_entry
);
1589 backtrace_warning_symbol(void *data
, char *msg
, unsigned long symbol
)
1591 /* Ignore warnings */
1594 static void backtrace_warning(void *data
, char *msg
)
1596 /* Ignore warnings */
1599 static int backtrace_stack(void *data
, char *name
)
1604 static void backtrace_address(void *data
, unsigned long addr
, int reliable
)
1606 struct perf_callchain_entry
*entry
= data
;
1608 callchain_store(entry
, addr
);
1611 static const struct stacktrace_ops backtrace_ops
= {
1612 .warning
= backtrace_warning
,
1613 .warning_symbol
= backtrace_warning_symbol
,
1614 .stack
= backtrace_stack
,
1615 .address
= backtrace_address
,
1616 .walk_stack
= print_context_stack_bp
,
1619 #include "../dumpstack.h"
1622 perf_callchain_kernel(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1624 callchain_store(entry
, PERF_CONTEXT_KERNEL
);
1625 callchain_store(entry
, regs
->ip
);
1627 dump_trace(NULL
, regs
, NULL
, regs
->bp
, &backtrace_ops
, entry
);
1630 #ifdef CONFIG_COMPAT
1632 perf_callchain_user32(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1634 /* 32-bit process in 64-bit kernel. */
1635 struct stack_frame_ia32 frame
;
1636 const void __user
*fp
;
1638 if (!test_thread_flag(TIF_IA32
))
1641 fp
= compat_ptr(regs
->bp
);
1642 while (entry
->nr
< PERF_MAX_STACK_DEPTH
) {
1643 unsigned long bytes
;
1644 frame
.next_frame
= 0;
1645 frame
.return_address
= 0;
1647 bytes
= copy_from_user_nmi(&frame
, fp
, sizeof(frame
));
1648 if (bytes
!= sizeof(frame
))
1651 if (fp
< compat_ptr(regs
->sp
))
1654 callchain_store(entry
, frame
.return_address
);
1655 fp
= compat_ptr(frame
.next_frame
);
1661 perf_callchain_user32(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1668 perf_callchain_user(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1670 struct stack_frame frame
;
1671 const void __user
*fp
;
1673 if (!user_mode(regs
))
1674 regs
= task_pt_regs(current
);
1676 fp
= (void __user
*)regs
->bp
;
1678 callchain_store(entry
, PERF_CONTEXT_USER
);
1679 callchain_store(entry
, regs
->ip
);
1681 if (perf_callchain_user32(regs
, entry
))
1684 while (entry
->nr
< PERF_MAX_STACK_DEPTH
) {
1685 unsigned long bytes
;
1686 frame
.next_frame
= NULL
;
1687 frame
.return_address
= 0;
1689 bytes
= copy_from_user_nmi(&frame
, fp
, sizeof(frame
));
1690 if (bytes
!= sizeof(frame
))
1693 if ((unsigned long)fp
< regs
->sp
)
1696 callchain_store(entry
, frame
.return_address
);
1697 fp
= frame
.next_frame
;
1702 perf_do_callchain(struct pt_regs
*regs
, struct perf_callchain_entry
*entry
)
1709 is_user
= user_mode(regs
);
1711 if (is_user
&& current
->state
!= TASK_RUNNING
)
1715 perf_callchain_kernel(regs
, entry
);
1718 perf_callchain_user(regs
, entry
);
1721 struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1723 struct perf_callchain_entry
*entry
;
1725 if (perf_guest_cbs
&& perf_guest_cbs
->is_in_guest()) {
1726 /* TODO: We don't support guest os callchain now */
1731 entry
= &__get_cpu_var(pmc_nmi_entry
);
1733 entry
= &__get_cpu_var(pmc_irq_entry
);
1737 perf_do_callchain(regs
, entry
);
1742 void perf_arch_fetch_caller_regs(struct pt_regs
*regs
, unsigned long ip
, int skip
)
1746 * perf_arch_fetch_caller_regs adds another call, we need to increment
1749 regs
->bp
= rewind_frame_pointer(skip
+ 1);
1750 regs
->cs
= __KERNEL_CS
;
1751 local_save_flags(regs
->flags
);
1754 unsigned long perf_instruction_pointer(struct pt_regs
*regs
)
1758 if (perf_guest_cbs
&& perf_guest_cbs
->is_in_guest())
1759 ip
= perf_guest_cbs
->get_guest_ip();
1761 ip
= instruction_pointer(regs
);
1766 unsigned long perf_misc_flags(struct pt_regs
*regs
)
1770 if (perf_guest_cbs
&& perf_guest_cbs
->is_in_guest()) {
1771 if (perf_guest_cbs
->is_user_mode())
1772 misc
|= PERF_RECORD_MISC_GUEST_USER
;
1774 misc
|= PERF_RECORD_MISC_GUEST_KERNEL
;
1776 if (user_mode(regs
))
1777 misc
|= PERF_RECORD_MISC_USER
;
1779 misc
|= PERF_RECORD_MISC_KERNEL
;
1782 if (regs
->flags
& PERF_EFLAGS_EXACT
)
1783 misc
|= PERF_RECORD_MISC_EXACT
;