1 /* Performance event support for sparc64.
3 * Copyright (C) 2009, 2010 David S. Miller <davem@davemloft.net>
5 * This code is based almost entirely upon the x86 perf event
8 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
9 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
10 * Copyright (C) 2009 Jaswinder Singh Rajput
11 * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
12 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
15 #include <linux/perf_event.h>
16 #include <linux/kprobes.h>
17 #include <linux/ftrace.h>
18 #include <linux/kernel.h>
19 #include <linux/kdebug.h>
20 #include <linux/mutex.h>
22 #include <asm/stacktrace.h>
23 #include <asm/cpudata.h>
24 #include <asm/uaccess.h>
25 #include <asm/atomic.h>
32 /* Sparc64 chips have two performance counters, 32-bits each, with
33 * overflow interrupts generated on transition from 0xffffffff to 0.
34 * The counters are accessed in one go using a 64-bit register.
36 * Both counters are controlled using a single control register. The
37 * only way to stop all sampling is to clear all of the context (user,
38 * supervisor, hypervisor) sampling enable bits. But these bits apply
39 * to both counters, thus the two counters can't be enabled/disabled
42 * The control register has two event fields, one for each of the two
43 * counters. It's thus nearly impossible to have one counter going
44 * while keeping the other one stopped. Therefore it is possible to
45 * get overflow interrupts for counters not currently "in use" and
46 * that condition must be checked in the overflow interrupt handler.
48 * So we use a hack, in that we program inactive counters with the
49 * "sw_count0" and "sw_count1" events. These count how many times
50 * the instruction "sethi %hi(0xfc000), %g0" is executed. It's an
51 * unusual way to encode a NOP and therefore will not trigger in
55 #define MAX_HWEVENTS 2
56 #define MAX_PERIOD ((1UL << 32) - 1)
58 #define PIC_UPPER_INDEX 0
59 #define PIC_LOWER_INDEX 1
60 #define PIC_NO_INDEX -1
62 struct cpu_hw_events {
63 /* Number of events currently scheduled onto this cpu.
64 * This tells how many entries in the arrays below
69 /* Number of new events added since the last hw_perf_disable().
70 * This works because the perf event layer always adds new
71 * events inside of a perf_{disable,enable}() sequence.
75 /* Array of events current scheduled on this cpu. */
76 struct perf_event *event[MAX_HWEVENTS];
78 /* Array of encoded longs, specifying the %pcr register
79 * encoding and the mask of PIC counters this even can
80 * be scheduled on. See perf_event_encode() et al.
82 unsigned long events[MAX_HWEVENTS];
84 /* The current counter index assigned to an event. When the
85 * event hasn't been programmed into the cpu yet, this will
86 * hold PIC_NO_INDEX. The event->hw.idx value tells us where
87 * we ought to schedule the event.
89 int current_idx[MAX_HWEVENTS];
91 /* Software copy of %pcr register on this cpu. */
94 /* Enabled/disable state. */
97 unsigned int group_flag;
99 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { .enabled = 1, };
101 /* An event map describes the characteristics of a performance
102 * counter event. In particular it gives the encoding as well as
103 * a mask telling which counters the event can be measured on.
105 struct perf_event_map {
108 #define PIC_NONE 0x00
109 #define PIC_UPPER 0x01
110 #define PIC_LOWER 0x02
113 /* Encode a perf_event_map entry into a long. */
114 static unsigned long perf_event_encode(const struct perf_event_map *pmap)
116 return ((unsigned long) pmap->encoding << 16) | pmap->pic_mask;
119 static u8 perf_event_get_msk(unsigned long val)
124 static u64 perf_event_get_enc(unsigned long val)
129 #define C(x) PERF_COUNT_HW_CACHE_##x
131 #define CACHE_OP_UNSUPPORTED 0xfffe
132 #define CACHE_OP_NONSENSE 0xffff
134 typedef struct perf_event_map cache_map_t
135 [PERF_COUNT_HW_CACHE_MAX]
136 [PERF_COUNT_HW_CACHE_OP_MAX]
137 [PERF_COUNT_HW_CACHE_RESULT_MAX];
140 const struct perf_event_map *(*event_map)(int);
141 const cache_map_t *cache_map;
152 static const struct perf_event_map ultra3_perfmon_event_map[] = {
153 [PERF_COUNT_HW_CPU_CYCLES] = { 0x0000, PIC_UPPER | PIC_LOWER },
154 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x0001, PIC_UPPER | PIC_LOWER },
155 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0009, PIC_LOWER },
156 [PERF_COUNT_HW_CACHE_MISSES] = { 0x0009, PIC_UPPER },
159 static const struct perf_event_map *ultra3_event_map(int event_id)
161 return &ultra3_perfmon_event_map[event_id];
164 static const cache_map_t ultra3_cache_map = {
167 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
168 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
171 [C(RESULT_ACCESS)] = { 0x0a, PIC_LOWER },
172 [C(RESULT_MISS)] = { 0x0a, PIC_UPPER },
175 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
176 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
181 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
182 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
185 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
186 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
188 [ C(OP_PREFETCH) ] = {
189 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
190 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
195 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER, },
196 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER, },
199 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER },
200 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER },
203 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
204 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
209 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
210 [C(RESULT_MISS)] = { 0x12, PIC_UPPER, },
213 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
214 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
216 [ C(OP_PREFETCH) ] = {
217 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
218 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
223 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
224 [C(RESULT_MISS)] = { 0x11, PIC_UPPER, },
227 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
228 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
230 [ C(OP_PREFETCH) ] = {
231 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
232 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
237 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
238 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
241 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
242 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
244 [ C(OP_PREFETCH) ] = {
245 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
246 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
251 static const struct sparc_pmu ultra3_pmu = {
252 .event_map = ultra3_event_map,
253 .cache_map = &ultra3_cache_map,
254 .max_events = ARRAY_SIZE(ultra3_perfmon_event_map),
262 /* Niagara1 is very limited. The upper PIC is hard-locked to count
263 * only instructions, so it is free running which creates all kinds of
264 * problems. Some hardware designs make one wonder if the creator
265 * even looked at how this stuff gets used by software.
267 static const struct perf_event_map niagara1_perfmon_event_map[] = {
268 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, PIC_UPPER },
269 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, PIC_UPPER },
270 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0, PIC_NONE },
271 [PERF_COUNT_HW_CACHE_MISSES] = { 0x03, PIC_LOWER },
274 static const struct perf_event_map *niagara1_event_map(int event_id)
276 return &niagara1_perfmon_event_map[event_id];
279 static const cache_map_t niagara1_cache_map = {
282 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
283 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
286 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
287 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
290 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
291 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
296 [C(RESULT_ACCESS)] = { 0x00, PIC_UPPER },
297 [C(RESULT_MISS)] = { 0x02, PIC_LOWER, },
300 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
301 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
303 [ C(OP_PREFETCH) ] = {
304 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
305 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
310 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
311 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
314 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
315 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
318 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
319 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
324 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
325 [C(RESULT_MISS)] = { 0x05, PIC_LOWER, },
328 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
329 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
331 [ C(OP_PREFETCH) ] = {
332 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
333 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
338 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
339 [C(RESULT_MISS)] = { 0x04, PIC_LOWER, },
342 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
343 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
345 [ C(OP_PREFETCH) ] = {
346 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
347 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
352 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
353 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
356 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
357 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
359 [ C(OP_PREFETCH) ] = {
360 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
361 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
366 static const struct sparc_pmu niagara1_pmu = {
367 .event_map = niagara1_event_map,
368 .cache_map = &niagara1_cache_map,
369 .max_events = ARRAY_SIZE(niagara1_perfmon_event_map),
377 static const struct perf_event_map niagara2_perfmon_event_map[] = {
378 [PERF_COUNT_HW_CPU_CYCLES] = { 0x02ff, PIC_UPPER | PIC_LOWER },
379 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x02ff, PIC_UPPER | PIC_LOWER },
380 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0208, PIC_UPPER | PIC_LOWER },
381 [PERF_COUNT_HW_CACHE_MISSES] = { 0x0302, PIC_UPPER | PIC_LOWER },
382 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x0201, PIC_UPPER | PIC_LOWER },
383 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x0202, PIC_UPPER | PIC_LOWER },
386 static const struct perf_event_map *niagara2_event_map(int event_id)
388 return &niagara2_perfmon_event_map[event_id];
391 static const cache_map_t niagara2_cache_map = {
394 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
395 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
398 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
399 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
402 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
403 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
408 [C(RESULT_ACCESS)] = { 0x02ff, PIC_UPPER | PIC_LOWER, },
409 [C(RESULT_MISS)] = { 0x0301, PIC_UPPER | PIC_LOWER, },
412 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
413 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
415 [ C(OP_PREFETCH) ] = {
416 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
417 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
422 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
423 [C(RESULT_MISS)] = { 0x0330, PIC_UPPER | PIC_LOWER, },
426 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
427 [C(RESULT_MISS)] = { 0x0320, PIC_UPPER | PIC_LOWER, },
430 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
431 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
436 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
437 [C(RESULT_MISS)] = { 0x0b08, PIC_UPPER | PIC_LOWER, },
440 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
441 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
443 [ C(OP_PREFETCH) ] = {
444 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
445 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
450 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
451 [C(RESULT_MISS)] = { 0xb04, PIC_UPPER | PIC_LOWER, },
454 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
455 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
457 [ C(OP_PREFETCH) ] = {
458 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
459 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
464 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
465 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
468 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
469 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
471 [ C(OP_PREFETCH) ] = {
472 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
473 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
478 static const struct sparc_pmu niagara2_pmu = {
479 .event_map = niagara2_event_map,
480 .cache_map = &niagara2_cache_map,
481 .max_events = ARRAY_SIZE(niagara2_perfmon_event_map),
491 static const struct sparc_pmu *sparc_pmu __read_mostly;
493 static u64 event_encoding(u64 event_id, int idx)
495 if (idx == PIC_UPPER_INDEX)
496 event_id <<= sparc_pmu->upper_shift;
498 event_id <<= sparc_pmu->lower_shift;
502 static u64 mask_for_index(int idx)
504 return event_encoding(sparc_pmu->event_mask, idx);
507 static u64 nop_for_index(int idx)
509 return event_encoding(idx == PIC_UPPER_INDEX ?
510 sparc_pmu->upper_nop :
511 sparc_pmu->lower_nop, idx);
514 static inline void sparc_pmu_enable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
516 u64 enc, val, mask = mask_for_index(idx);
518 enc = perf_event_get_enc(cpuc->events[idx]);
522 val |= event_encoding(enc, idx);
525 pcr_ops->write(cpuc->pcr);
528 static inline void sparc_pmu_disable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
530 u64 mask = mask_for_index(idx);
531 u64 nop = nop_for_index(idx);
539 pcr_ops->write(cpuc->pcr);
542 static u32 read_pmc(int idx)
547 if (idx == PIC_UPPER_INDEX)
550 return val & 0xffffffff;
553 static void write_pmc(int idx, u64 val)
555 u64 shift, mask, pic;
558 if (idx == PIC_UPPER_INDEX)
561 mask = ((u64) 0xffffffff) << shift;
570 static u64 sparc_perf_event_update(struct perf_event *event,
571 struct hw_perf_event *hwc, int idx)
574 u64 prev_raw_count, new_raw_count;
578 prev_raw_count = local64_read(&hwc->prev_count);
579 new_raw_count = read_pmc(idx);
581 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
582 new_raw_count) != prev_raw_count)
585 delta = (new_raw_count << shift) - (prev_raw_count << shift);
588 local64_add(delta, &event->count);
589 local64_sub(delta, &hwc->period_left);
591 return new_raw_count;
594 static int sparc_perf_event_set_period(struct perf_event *event,
595 struct hw_perf_event *hwc, int idx)
597 s64 left = local64_read(&hwc->period_left);
598 s64 period = hwc->sample_period;
601 if (unlikely(left <= -period)) {
603 local64_set(&hwc->period_left, left);
604 hwc->last_period = period;
608 if (unlikely(left <= 0)) {
610 local64_set(&hwc->period_left, left);
611 hwc->last_period = period;
614 if (left > MAX_PERIOD)
617 local64_set(&hwc->prev_count, (u64)-left);
619 write_pmc(idx, (u64)(-left) & 0xffffffff);
621 perf_event_update_userpage(event);
626 /* If performance event entries have been added, move existing
627 * events around (if necessary) and then assign new entries to
630 static u64 maybe_change_configuration(struct cpu_hw_events *cpuc, u64 pcr)
637 /* Read in the counters which are moving. */
638 for (i = 0; i < cpuc->n_events; i++) {
639 struct perf_event *cp = cpuc->event[i];
641 if (cpuc->current_idx[i] != PIC_NO_INDEX &&
642 cpuc->current_idx[i] != cp->hw.idx) {
643 sparc_perf_event_update(cp, &cp->hw,
644 cpuc->current_idx[i]);
645 cpuc->current_idx[i] = PIC_NO_INDEX;
649 /* Assign to counters all unassigned events. */
650 for (i = 0; i < cpuc->n_events; i++) {
651 struct perf_event *cp = cpuc->event[i];
652 struct hw_perf_event *hwc = &cp->hw;
656 if (cpuc->current_idx[i] != PIC_NO_INDEX)
659 sparc_perf_event_set_period(cp, hwc, idx);
660 cpuc->current_idx[i] = idx;
662 enc = perf_event_get_enc(cpuc->events[i]);
663 pcr &= ~mask_for_index(idx);
664 if (hwc->state & PERF_HES_STOPPED)
665 pcr |= nop_for_index(idx);
667 pcr |= event_encoding(enc, idx);
673 static void sparc_pmu_enable(struct pmu *pmu)
675 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
685 if (!cpuc->n_events) {
688 pcr = maybe_change_configuration(cpuc, pcr);
690 /* We require that all of the events have the same
691 * configuration, so just fetch the settings from the
694 cpuc->pcr = pcr | cpuc->event[0]->hw.config_base;
697 pcr_ops->write(cpuc->pcr);
700 static void sparc_pmu_disable(struct pmu *pmu)
702 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
712 val &= ~(PCR_UTRACE | PCR_STRACE |
713 sparc_pmu->hv_bit | sparc_pmu->irq_bit);
716 pcr_ops->write(cpuc->pcr);
719 static int active_event_index(struct cpu_hw_events *cpuc,
720 struct perf_event *event)
724 for (i = 0; i < cpuc->n_events; i++) {
725 if (cpuc->event[i] == event)
728 BUG_ON(i == cpuc->n_events);
729 return cpuc->current_idx[i];
732 static void sparc_pmu_start(struct perf_event *event, int flags)
734 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
735 int idx = active_event_index(cpuc, event);
737 if (flags & PERF_EF_RELOAD) {
738 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
739 sparc_perf_event_set_period(event, &event->hw, idx);
744 sparc_pmu_enable_event(cpuc, &event->hw, idx);
747 static void sparc_pmu_stop(struct perf_event *event, int flags)
749 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
750 int idx = active_event_index(cpuc, event);
752 if (!(event->hw.state & PERF_HES_STOPPED)) {
753 sparc_pmu_disable_event(cpuc, &event->hw, idx);
754 event->hw.state |= PERF_HES_STOPPED;
757 if (!(event->hw.state & PERF_HES_UPTODATE) && (flags & PERF_EF_UPDATE)) {
758 sparc_perf_event_update(event, &event->hw, idx);
759 event->hw.state |= PERF_HES_UPTODATE;
763 static void sparc_pmu_del(struct perf_event *event, int _flags)
765 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
769 local_irq_save(flags);
770 perf_pmu_disable(event->pmu);
772 for (i = 0; i < cpuc->n_events; i++) {
773 if (event == cpuc->event[i]) {
774 /* Absorb the final count and turn off the
777 sparc_pmu_stop(event, PERF_EF_UPDATE);
779 /* Shift remaining entries down into
782 while (++i < cpuc->n_events) {
783 cpuc->event[i - 1] = cpuc->event[i];
784 cpuc->events[i - 1] = cpuc->events[i];
785 cpuc->current_idx[i - 1] =
786 cpuc->current_idx[i];
789 perf_event_update_userpage(event);
796 perf_pmu_enable(event->pmu);
797 local_irq_restore(flags);
800 static void sparc_pmu_read(struct perf_event *event)
802 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
803 int idx = active_event_index(cpuc, event);
804 struct hw_perf_event *hwc = &event->hw;
806 sparc_perf_event_update(event, hwc, idx);
809 static atomic_t active_events = ATOMIC_INIT(0);
810 static DEFINE_MUTEX(pmc_grab_mutex);
812 static void perf_stop_nmi_watchdog(void *unused)
814 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
816 stop_nmi_watchdog(NULL);
817 cpuc->pcr = pcr_ops->read();
820 void perf_event_grab_pmc(void)
822 if (atomic_inc_not_zero(&active_events))
825 mutex_lock(&pmc_grab_mutex);
826 if (atomic_read(&active_events) == 0) {
827 if (atomic_read(&nmi_active) > 0) {
828 on_each_cpu(perf_stop_nmi_watchdog, NULL, 1);
829 BUG_ON(atomic_read(&nmi_active) != 0);
831 atomic_inc(&active_events);
833 mutex_unlock(&pmc_grab_mutex);
836 void perf_event_release_pmc(void)
838 if (atomic_dec_and_mutex_lock(&active_events, &pmc_grab_mutex)) {
839 if (atomic_read(&nmi_active) == 0)
840 on_each_cpu(start_nmi_watchdog, NULL, 1);
841 mutex_unlock(&pmc_grab_mutex);
845 static const struct perf_event_map *sparc_map_cache_event(u64 config)
847 unsigned int cache_type, cache_op, cache_result;
848 const struct perf_event_map *pmap;
850 if (!sparc_pmu->cache_map)
851 return ERR_PTR(-ENOENT);
853 cache_type = (config >> 0) & 0xff;
854 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
855 return ERR_PTR(-EINVAL);
857 cache_op = (config >> 8) & 0xff;
858 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
859 return ERR_PTR(-EINVAL);
861 cache_result = (config >> 16) & 0xff;
862 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
863 return ERR_PTR(-EINVAL);
865 pmap = &((*sparc_pmu->cache_map)[cache_type][cache_op][cache_result]);
867 if (pmap->encoding == CACHE_OP_UNSUPPORTED)
868 return ERR_PTR(-ENOENT);
870 if (pmap->encoding == CACHE_OP_NONSENSE)
871 return ERR_PTR(-EINVAL);
876 static void hw_perf_event_destroy(struct perf_event *event)
878 perf_event_release_pmc();
881 /* Make sure all events can be scheduled into the hardware at
882 * the same time. This is simplified by the fact that we only
883 * need to support 2 simultaneous HW events.
885 * As a side effect, the evts[]->hw.idx values will be assigned
886 * on success. These are pending indexes. When the events are
887 * actually programmed into the chip, these values will propagate
888 * to the per-cpu cpuc->current_idx[] slots, see the code in
889 * maybe_change_configuration() for details.
891 static int sparc_check_constraints(struct perf_event **evts,
892 unsigned long *events, int n_ev)
894 u8 msk0 = 0, msk1 = 0;
897 /* This case is possible when we are invoked from
898 * hw_perf_group_sched_in().
903 if (n_ev > MAX_HWEVENTS)
906 msk0 = perf_event_get_msk(events[0]);
908 if (msk0 & PIC_LOWER)
913 msk1 = perf_event_get_msk(events[1]);
915 /* If both events can go on any counter, OK. */
916 if (msk0 == (PIC_UPPER | PIC_LOWER) &&
917 msk1 == (PIC_UPPER | PIC_LOWER))
920 /* If one event is limited to a specific counter,
921 * and the other can go on both, OK.
923 if ((msk0 == PIC_UPPER || msk0 == PIC_LOWER) &&
924 msk1 == (PIC_UPPER | PIC_LOWER)) {
925 if (msk0 & PIC_LOWER)
930 if ((msk1 == PIC_UPPER || msk1 == PIC_LOWER) &&
931 msk0 == (PIC_UPPER | PIC_LOWER)) {
932 if (msk1 & PIC_UPPER)
937 /* If the events are fixed to different counters, OK. */
938 if ((msk0 == PIC_UPPER && msk1 == PIC_LOWER) ||
939 (msk0 == PIC_LOWER && msk1 == PIC_UPPER)) {
940 if (msk0 & PIC_LOWER)
945 /* Otherwise, there is a conflict. */
949 evts[0]->hw.idx = idx0;
951 evts[1]->hw.idx = idx0 ^ 1;
955 static int check_excludes(struct perf_event **evts, int n_prev, int n_new)
957 int eu = 0, ek = 0, eh = 0;
958 struct perf_event *event;
966 for (i = 0; i < n; i++) {
969 eu = event->attr.exclude_user;
970 ek = event->attr.exclude_kernel;
971 eh = event->attr.exclude_hv;
973 } else if (event->attr.exclude_user != eu ||
974 event->attr.exclude_kernel != ek ||
975 event->attr.exclude_hv != eh) {
983 static int collect_events(struct perf_event *group, int max_count,
984 struct perf_event *evts[], unsigned long *events,
987 struct perf_event *event;
990 if (!is_software_event(group)) {
994 events[n] = group->hw.event_base;
995 current_idx[n++] = PIC_NO_INDEX;
997 list_for_each_entry(event, &group->sibling_list, group_entry) {
998 if (!is_software_event(event) &&
999 event->state != PERF_EVENT_STATE_OFF) {
1003 events[n] = event->hw.event_base;
1004 current_idx[n++] = PIC_NO_INDEX;
1010 static int sparc_pmu_add(struct perf_event *event, int ef_flags)
1012 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1013 int n0, ret = -EAGAIN;
1014 unsigned long flags;
1016 local_irq_save(flags);
1017 perf_pmu_disable(event->pmu);
1019 n0 = cpuc->n_events;
1020 if (n0 >= MAX_HWEVENTS)
1023 cpuc->event[n0] = event;
1024 cpuc->events[n0] = event->hw.event_base;
1025 cpuc->current_idx[n0] = PIC_NO_INDEX;
1027 event->hw.state = PERF_HES_UPTODATE;
1028 if (!(ef_flags & PERF_EF_START))
1029 event->hw.state |= PERF_HES_STOPPED;
1032 * If group events scheduling transaction was started,
1033 * skip the schedulability test here, it will be performed
1034 * at commit time(->commit_txn) as a whole
1036 if (cpuc->group_flag & PERF_EVENT_TXN)
1039 if (check_excludes(cpuc->event, n0, 1))
1041 if (sparc_check_constraints(cpuc->event, cpuc->events, n0 + 1))
1050 perf_pmu_enable(event->pmu);
1051 local_irq_restore(flags);
1055 static int sparc_pmu_event_init(struct perf_event *event)
1057 struct perf_event_attr *attr = &event->attr;
1058 struct perf_event *evts[MAX_HWEVENTS];
1059 struct hw_perf_event *hwc = &event->hw;
1060 unsigned long events[MAX_HWEVENTS];
1061 int current_idx_dmy[MAX_HWEVENTS];
1062 const struct perf_event_map *pmap;
1065 if (atomic_read(&nmi_active) < 0)
1068 switch (attr->type) {
1069 case PERF_TYPE_HARDWARE:
1070 if (attr->config >= sparc_pmu->max_events)
1072 pmap = sparc_pmu->event_map(attr->config);
1075 case PERF_TYPE_HW_CACHE:
1076 pmap = sparc_map_cache_event(attr->config);
1078 return PTR_ERR(pmap);
1091 hwc->event_base = perf_event_encode(pmap);
1094 * User gives us "(encoding << 16) | pic_mask" for
1095 * PERF_TYPE_RAW events.
1097 hwc->event_base = attr->config;
1100 /* We save the enable bits in the config_base. */
1101 hwc->config_base = sparc_pmu->irq_bit;
1102 if (!attr->exclude_user)
1103 hwc->config_base |= PCR_UTRACE;
1104 if (!attr->exclude_kernel)
1105 hwc->config_base |= PCR_STRACE;
1106 if (!attr->exclude_hv)
1107 hwc->config_base |= sparc_pmu->hv_bit;
1110 if (event->group_leader != event) {
1111 n = collect_events(event->group_leader,
1113 evts, events, current_idx_dmy);
1117 events[n] = hwc->event_base;
1120 if (check_excludes(evts, n, 1))
1123 if (sparc_check_constraints(evts, events, n + 1))
1126 hwc->idx = PIC_NO_INDEX;
1128 /* Try to do all error checking before this point, as unwinding
1129 * state after grabbing the PMC is difficult.
1131 perf_event_grab_pmc();
1132 event->destroy = hw_perf_event_destroy;
1134 if (!hwc->sample_period) {
1135 hwc->sample_period = MAX_PERIOD;
1136 hwc->last_period = hwc->sample_period;
1137 local64_set(&hwc->period_left, hwc->sample_period);
1144 * Start group events scheduling transaction
1145 * Set the flag to make pmu::enable() not perform the
1146 * schedulability test, it will be performed at commit time
1148 static void sparc_pmu_start_txn(struct pmu *pmu)
1150 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1152 perf_pmu_disable(pmu);
1153 cpuhw->group_flag |= PERF_EVENT_TXN;
1157 * Stop group events scheduling transaction
1158 * Clear the flag and pmu::enable() will perform the
1159 * schedulability test.
1161 static void sparc_pmu_cancel_txn(struct pmu *pmu)
1163 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1165 cpuhw->group_flag &= ~PERF_EVENT_TXN;
1166 perf_pmu_enable(pmu);
1170 * Commit group events scheduling transaction
1171 * Perform the group schedulability test as a whole
1172 * Return 0 if success
1174 static int sparc_pmu_commit_txn(struct pmu *pmu)
1176 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1182 cpuc = &__get_cpu_var(cpu_hw_events);
1184 if (check_excludes(cpuc->event, 0, n))
1186 if (sparc_check_constraints(cpuc->event, cpuc->events, n))
1189 cpuc->group_flag &= ~PERF_EVENT_TXN;
1190 perf_pmu_enable(pmu);
1194 static struct pmu pmu = {
1195 .pmu_enable = sparc_pmu_enable,
1196 .pmu_disable = sparc_pmu_disable,
1197 .event_init = sparc_pmu_event_init,
1198 .add = sparc_pmu_add,
1199 .del = sparc_pmu_del,
1200 .start = sparc_pmu_start,
1201 .stop = sparc_pmu_stop,
1202 .read = sparc_pmu_read,
1203 .start_txn = sparc_pmu_start_txn,
1204 .cancel_txn = sparc_pmu_cancel_txn,
1205 .commit_txn = sparc_pmu_commit_txn,
1208 void perf_event_print_debug(void)
1210 unsigned long flags;
1217 local_irq_save(flags);
1219 cpu = smp_processor_id();
1221 pcr = pcr_ops->read();
1225 pr_info("CPU#%d: PCR[%016llx] PIC[%016llx]\n",
1228 local_irq_restore(flags);
1231 static int __kprobes perf_event_nmi_handler(struct notifier_block *self,
1232 unsigned long cmd, void *__args)
1234 struct die_args *args = __args;
1235 struct perf_sample_data data;
1236 struct cpu_hw_events *cpuc;
1237 struct pt_regs *regs;
1240 if (!atomic_read(&active_events))
1253 perf_sample_data_init(&data, 0);
1255 cpuc = &__get_cpu_var(cpu_hw_events);
1257 /* If the PMU has the TOE IRQ enable bits, we need to do a
1258 * dummy write to the %pcr to clear the overflow bits and thus
1261 * Do this before we peek at the counters to determine
1262 * overflow so we don't lose any events.
1264 if (sparc_pmu->irq_bit)
1265 pcr_ops->write(cpuc->pcr);
1267 for (i = 0; i < cpuc->n_events; i++) {
1268 struct perf_event *event = cpuc->event[i];
1269 int idx = cpuc->current_idx[i];
1270 struct hw_perf_event *hwc;
1274 val = sparc_perf_event_update(event, hwc, idx);
1275 if (val & (1ULL << 31))
1278 data.period = event->hw.last_period;
1279 if (!sparc_perf_event_set_period(event, hwc, idx))
1282 if (perf_event_overflow(event, 1, &data, regs))
1283 sparc_pmu_stop(event, 0);
1289 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1290 .notifier_call = perf_event_nmi_handler,
1293 static bool __init supported_pmu(void)
1295 if (!strcmp(sparc_pmu_type, "ultra3") ||
1296 !strcmp(sparc_pmu_type, "ultra3+") ||
1297 !strcmp(sparc_pmu_type, "ultra3i") ||
1298 !strcmp(sparc_pmu_type, "ultra4+")) {
1299 sparc_pmu = &ultra3_pmu;
1302 if (!strcmp(sparc_pmu_type, "niagara")) {
1303 sparc_pmu = &niagara1_pmu;
1306 if (!strcmp(sparc_pmu_type, "niagara2") ||
1307 !strcmp(sparc_pmu_type, "niagara3")) {
1308 sparc_pmu = &niagara2_pmu;
1314 int __init init_hw_perf_events(void)
1316 pr_info("Performance events: ");
1318 if (!supported_pmu()) {
1319 pr_cont("No support for PMU type '%s'\n", sparc_pmu_type);
1323 pr_cont("Supported PMU type is '%s'\n", sparc_pmu_type);
1325 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1326 register_die_notifier(&perf_event_nmi_notifier);
1330 early_initcall(init_hw_perf_events);
1332 void perf_callchain_kernel(struct perf_callchain_entry *entry,
1333 struct pt_regs *regs)
1335 unsigned long ksp, fp;
1336 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
1340 stack_trace_flush();
1342 perf_callchain_store(entry, regs->tpc);
1344 ksp = regs->u_regs[UREG_I6];
1345 fp = ksp + STACK_BIAS;
1347 struct sparc_stackf *sf;
1348 struct pt_regs *regs;
1351 if (!kstack_valid(current_thread_info(), fp))
1354 sf = (struct sparc_stackf *) fp;
1355 regs = (struct pt_regs *) (sf + 1);
1357 if (kstack_is_trap_frame(current_thread_info(), regs)) {
1358 if (user_mode(regs))
1361 fp = regs->u_regs[UREG_I6] + STACK_BIAS;
1363 pc = sf->callers_pc;
1364 fp = (unsigned long)sf->fp + STACK_BIAS;
1366 perf_callchain_store(entry, pc);
1367 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
1368 if ((pc + 8UL) == (unsigned long) &return_to_handler) {
1369 int index = current->curr_ret_stack;
1370 if (current->ret_stack && index >= graph) {
1371 pc = current->ret_stack[index - graph].ret;
1372 perf_callchain_store(entry, pc);
1377 } while (entry->nr < PERF_MAX_STACK_DEPTH);
1380 static void perf_callchain_user_64(struct perf_callchain_entry *entry,
1381 struct pt_regs *regs)
1385 ufp = regs->u_regs[UREG_I6] + STACK_BIAS;
1387 struct sparc_stackf *usf, sf;
1390 usf = (struct sparc_stackf *) ufp;
1391 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1395 ufp = (unsigned long)sf.fp + STACK_BIAS;
1396 perf_callchain_store(entry, pc);
1397 } while (entry->nr < PERF_MAX_STACK_DEPTH);
1400 static void perf_callchain_user_32(struct perf_callchain_entry *entry,
1401 struct pt_regs *regs)
1405 ufp = regs->u_regs[UREG_I6] & 0xffffffffUL;
1407 struct sparc_stackf32 *usf, sf;
1410 usf = (struct sparc_stackf32 *) ufp;
1411 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1415 ufp = (unsigned long)sf.fp;
1416 perf_callchain_store(entry, pc);
1417 } while (entry->nr < PERF_MAX_STACK_DEPTH);
1421 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1423 perf_callchain_store(entry, regs->tpc);
1429 if (test_thread_flag(TIF_32BIT))
1430 perf_callchain_user_32(entry, regs);
1432 perf_callchain_user_64(entry, regs);