Loading...
1/*
2 * Linux performance counter support for MIPS.
3 *
4 * Copyright (C) 2010 MIPS Technologies, Inc.
5 * Copyright (C) 2011 Cavium Networks, Inc.
6 * Author: Deng-Cheng Zhu
7 *
8 * This code is based on the implementation for ARM, which is in turn
9 * based on the sparc64 perf event code and the x86 code. Performance
10 * counter access is based on the MIPS Oprofile code. And the callchain
11 * support references the code of MIPS stacktrace.c.
12 *
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License version 2 as
15 * published by the Free Software Foundation.
16 */
17
18#include <linux/cpumask.h>
19#include <linux/interrupt.h>
20#include <linux/smp.h>
21#include <linux/kernel.h>
22#include <linux/perf_event.h>
23#include <linux/uaccess.h>
24
25#include <asm/irq.h>
26#include <asm/irq_regs.h>
27#include <asm/stacktrace.h>
28#include <asm/time.h> /* For perf_irq */
29
30#define MIPS_MAX_HWEVENTS 4
31#define MIPS_TCS_PER_COUNTER 2
32#define MIPS_CPUID_TO_COUNTER_MASK (MIPS_TCS_PER_COUNTER - 1)
33
34struct cpu_hw_events {
35 /* Array of events on this cpu. */
36 struct perf_event *events[MIPS_MAX_HWEVENTS];
37
38 /*
39 * Set the bit (indexed by the counter number) when the counter
40 * is used for an event.
41 */
42 unsigned long used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)];
43
44 /*
45 * Software copy of the control register for each performance counter.
46 * MIPS CPUs vary in performance counters. They use this differently,
47 * and even may not use it.
48 */
49 unsigned int saved_ctrl[MIPS_MAX_HWEVENTS];
50};
51DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
52 .saved_ctrl = {0},
53};
54
55/* The description of MIPS performance events. */
56struct mips_perf_event {
57 unsigned int event_id;
58 /*
59 * MIPS performance counters are indexed starting from 0.
60 * CNTR_EVEN indicates the indexes of the counters to be used are
61 * even numbers.
62 */
63 unsigned int cntr_mask;
64 #define CNTR_EVEN 0x55555555
65 #define CNTR_ODD 0xaaaaaaaa
66 #define CNTR_ALL 0xffffffff
67#ifdef CONFIG_MIPS_MT_SMP
68 enum {
69 T = 0,
70 V = 1,
71 P = 2,
72 } range;
73#else
74 #define T
75 #define V
76 #define P
77#endif
78};
79
80static struct mips_perf_event raw_event;
81static DEFINE_MUTEX(raw_event_mutex);
82
83#define C(x) PERF_COUNT_HW_CACHE_##x
84
85struct mips_pmu {
86 u64 max_period;
87 u64 valid_count;
88 u64 overflow;
89 const char *name;
90 int irq;
91 u64 (*read_counter)(unsigned int idx);
92 void (*write_counter)(unsigned int idx, u64 val);
93 const struct mips_perf_event *(*map_raw_event)(u64 config);
94 const struct mips_perf_event (*general_event_map)[PERF_COUNT_HW_MAX];
95 const struct mips_perf_event (*cache_event_map)
96 [PERF_COUNT_HW_CACHE_MAX]
97 [PERF_COUNT_HW_CACHE_OP_MAX]
98 [PERF_COUNT_HW_CACHE_RESULT_MAX];
99 unsigned int num_counters;
100};
101
102static struct mips_pmu mipspmu;
103
104#define M_CONFIG1_PC (1 << 4)
105
106#define M_PERFCTL_EXL (1 << 0)
107#define M_PERFCTL_KERNEL (1 << 1)
108#define M_PERFCTL_SUPERVISOR (1 << 2)
109#define M_PERFCTL_USER (1 << 3)
110#define M_PERFCTL_INTERRUPT_ENABLE (1 << 4)
111#define M_PERFCTL_EVENT(event) (((event) & 0x3ff) << 5)
112#define M_PERFCTL_VPEID(vpe) ((vpe) << 16)
113
114#ifdef CONFIG_CPU_BMIPS5000
115#define M_PERFCTL_MT_EN(filter) 0
116#else /* !CONFIG_CPU_BMIPS5000 */
117#define M_PERFCTL_MT_EN(filter) ((filter) << 20)
118#endif /* CONFIG_CPU_BMIPS5000 */
119
120#define M_TC_EN_ALL M_PERFCTL_MT_EN(0)
121#define M_TC_EN_VPE M_PERFCTL_MT_EN(1)
122#define M_TC_EN_TC M_PERFCTL_MT_EN(2)
123#define M_PERFCTL_TCID(tcid) ((tcid) << 22)
124#define M_PERFCTL_WIDE (1 << 30)
125#define M_PERFCTL_MORE (1 << 31)
126#define M_PERFCTL_TC (1 << 30)
127
128#define M_PERFCTL_COUNT_EVENT_WHENEVER (M_PERFCTL_EXL | \
129 M_PERFCTL_KERNEL | \
130 M_PERFCTL_USER | \
131 M_PERFCTL_SUPERVISOR | \
132 M_PERFCTL_INTERRUPT_ENABLE)
133
134#ifdef CONFIG_MIPS_MT_SMP
135#define M_PERFCTL_CONFIG_MASK 0x3fff801f
136#else
137#define M_PERFCTL_CONFIG_MASK 0x1f
138#endif
139#define M_PERFCTL_EVENT_MASK 0xfe0
140
141
142#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
143static int cpu_has_mipsmt_pertccounters;
144
145static DEFINE_RWLOCK(pmuint_rwlock);
146
147#if defined(CONFIG_CPU_BMIPS5000)
148#define vpe_id() (cpu_has_mipsmt_pertccounters ? \
149 0 : (smp_processor_id() & MIPS_CPUID_TO_COUNTER_MASK))
150#else
151/*
152 * FIXME: For VSMP, vpe_id() is redefined for Perf-events, because
153 * cpu_data[cpuid].vpe_id reports 0 for _both_ CPUs.
154 */
155#define vpe_id() (cpu_has_mipsmt_pertccounters ? \
156 0 : smp_processor_id())
157#endif
158
159/* Copied from op_model_mipsxx.c */
160static unsigned int vpe_shift(void)
161{
162 if (num_possible_cpus() > 1)
163 return 1;
164
165 return 0;
166}
167
168static unsigned int counters_total_to_per_cpu(unsigned int counters)
169{
170 return counters >> vpe_shift();
171}
172
173#else /* !CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
174#define vpe_id() 0
175
176#endif /* CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
177
178static void resume_local_counters(void);
179static void pause_local_counters(void);
180static irqreturn_t mipsxx_pmu_handle_irq(int, void *);
181static int mipsxx_pmu_handle_shared_irq(void);
182
183static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx)
184{
185 if (vpe_id() == 1)
186 idx = (idx + 2) & 3;
187 return idx;
188}
189
190static u64 mipsxx_pmu_read_counter(unsigned int idx)
191{
192 idx = mipsxx_pmu_swizzle_perf_idx(idx);
193
194 switch (idx) {
195 case 0:
196 /*
197 * The counters are unsigned, we must cast to truncate
198 * off the high bits.
199 */
200 return (u32)read_c0_perfcntr0();
201 case 1:
202 return (u32)read_c0_perfcntr1();
203 case 2:
204 return (u32)read_c0_perfcntr2();
205 case 3:
206 return (u32)read_c0_perfcntr3();
207 default:
208 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
209 return 0;
210 }
211}
212
213static u64 mipsxx_pmu_read_counter_64(unsigned int idx)
214{
215 idx = mipsxx_pmu_swizzle_perf_idx(idx);
216
217 switch (idx) {
218 case 0:
219 return read_c0_perfcntr0_64();
220 case 1:
221 return read_c0_perfcntr1_64();
222 case 2:
223 return read_c0_perfcntr2_64();
224 case 3:
225 return read_c0_perfcntr3_64();
226 default:
227 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
228 return 0;
229 }
230}
231
232static void mipsxx_pmu_write_counter(unsigned int idx, u64 val)
233{
234 idx = mipsxx_pmu_swizzle_perf_idx(idx);
235
236 switch (idx) {
237 case 0:
238 write_c0_perfcntr0(val);
239 return;
240 case 1:
241 write_c0_perfcntr1(val);
242 return;
243 case 2:
244 write_c0_perfcntr2(val);
245 return;
246 case 3:
247 write_c0_perfcntr3(val);
248 return;
249 }
250}
251
252static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val)
253{
254 idx = mipsxx_pmu_swizzle_perf_idx(idx);
255
256 switch (idx) {
257 case 0:
258 write_c0_perfcntr0_64(val);
259 return;
260 case 1:
261 write_c0_perfcntr1_64(val);
262 return;
263 case 2:
264 write_c0_perfcntr2_64(val);
265 return;
266 case 3:
267 write_c0_perfcntr3_64(val);
268 return;
269 }
270}
271
272static unsigned int mipsxx_pmu_read_control(unsigned int idx)
273{
274 idx = mipsxx_pmu_swizzle_perf_idx(idx);
275
276 switch (idx) {
277 case 0:
278 return read_c0_perfctrl0();
279 case 1:
280 return read_c0_perfctrl1();
281 case 2:
282 return read_c0_perfctrl2();
283 case 3:
284 return read_c0_perfctrl3();
285 default:
286 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
287 return 0;
288 }
289}
290
291static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val)
292{
293 idx = mipsxx_pmu_swizzle_perf_idx(idx);
294
295 switch (idx) {
296 case 0:
297 write_c0_perfctrl0(val);
298 return;
299 case 1:
300 write_c0_perfctrl1(val);
301 return;
302 case 2:
303 write_c0_perfctrl2(val);
304 return;
305 case 3:
306 write_c0_perfctrl3(val);
307 return;
308 }
309}
310
311static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc,
312 struct hw_perf_event *hwc)
313{
314 int i;
315
316 /*
317 * We only need to care the counter mask. The range has been
318 * checked definitely.
319 */
320 unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff;
321
322 for (i = mipspmu.num_counters - 1; i >= 0; i--) {
323 /*
324 * Note that some MIPS perf events can be counted by both
325 * even and odd counters, wheresas many other are only by
326 * even _or_ odd counters. This introduces an issue that
327 * when the former kind of event takes the counter the
328 * latter kind of event wants to use, then the "counter
329 * allocation" for the latter event will fail. In fact if
330 * they can be dynamically swapped, they both feel happy.
331 * But here we leave this issue alone for now.
332 */
333 if (test_bit(i, &cntr_mask) &&
334 !test_and_set_bit(i, cpuc->used_mask))
335 return i;
336 }
337
338 return -EAGAIN;
339}
340
341static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx)
342{
343 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
344
345 WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
346
347 cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) |
348 (evt->config_base & M_PERFCTL_CONFIG_MASK) |
349 /* Make sure interrupt enabled. */
350 M_PERFCTL_INTERRUPT_ENABLE;
351 if (IS_ENABLED(CONFIG_CPU_BMIPS5000))
352 /* enable the counter for the calling thread */
353 cpuc->saved_ctrl[idx] |=
354 (1 << (12 + vpe_id())) | M_PERFCTL_TC;
355
356 /*
357 * We do not actually let the counter run. Leave it until start().
358 */
359}
360
361static void mipsxx_pmu_disable_event(int idx)
362{
363 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
364 unsigned long flags;
365
366 WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
367
368 local_irq_save(flags);
369 cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) &
370 ~M_PERFCTL_COUNT_EVENT_WHENEVER;
371 mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]);
372 local_irq_restore(flags);
373}
374
375static int mipspmu_event_set_period(struct perf_event *event,
376 struct hw_perf_event *hwc,
377 int idx)
378{
379 u64 left = local64_read(&hwc->period_left);
380 u64 period = hwc->sample_period;
381 int ret = 0;
382
383 if (unlikely((left + period) & (1ULL << 63))) {
384 /* left underflowed by more than period. */
385 left = period;
386 local64_set(&hwc->period_left, left);
387 hwc->last_period = period;
388 ret = 1;
389 } else if (unlikely((left + period) <= period)) {
390 /* left underflowed by less than period. */
391 left += period;
392 local64_set(&hwc->period_left, left);
393 hwc->last_period = period;
394 ret = 1;
395 }
396
397 if (left > mipspmu.max_period) {
398 left = mipspmu.max_period;
399 local64_set(&hwc->period_left, left);
400 }
401
402 local64_set(&hwc->prev_count, mipspmu.overflow - left);
403
404 mipspmu.write_counter(idx, mipspmu.overflow - left);
405
406 perf_event_update_userpage(event);
407
408 return ret;
409}
410
411static void mipspmu_event_update(struct perf_event *event,
412 struct hw_perf_event *hwc,
413 int idx)
414{
415 u64 prev_raw_count, new_raw_count;
416 u64 delta;
417
418again:
419 prev_raw_count = local64_read(&hwc->prev_count);
420 new_raw_count = mipspmu.read_counter(idx);
421
422 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
423 new_raw_count) != prev_raw_count)
424 goto again;
425
426 delta = new_raw_count - prev_raw_count;
427
428 local64_add(delta, &event->count);
429 local64_sub(delta, &hwc->period_left);
430}
431
432static void mipspmu_start(struct perf_event *event, int flags)
433{
434 struct hw_perf_event *hwc = &event->hw;
435
436 if (flags & PERF_EF_RELOAD)
437 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
438
439 hwc->state = 0;
440
441 /* Set the period for the event. */
442 mipspmu_event_set_period(event, hwc, hwc->idx);
443
444 /* Enable the event. */
445 mipsxx_pmu_enable_event(hwc, hwc->idx);
446}
447
448static void mipspmu_stop(struct perf_event *event, int flags)
449{
450 struct hw_perf_event *hwc = &event->hw;
451
452 if (!(hwc->state & PERF_HES_STOPPED)) {
453 /* We are working on a local event. */
454 mipsxx_pmu_disable_event(hwc->idx);
455 barrier();
456 mipspmu_event_update(event, hwc, hwc->idx);
457 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
458 }
459}
460
461static int mipspmu_add(struct perf_event *event, int flags)
462{
463 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
464 struct hw_perf_event *hwc = &event->hw;
465 int idx;
466 int err = 0;
467
468 perf_pmu_disable(event->pmu);
469
470 /* To look for a free counter for this event. */
471 idx = mipsxx_pmu_alloc_counter(cpuc, hwc);
472 if (idx < 0) {
473 err = idx;
474 goto out;
475 }
476
477 /*
478 * If there is an event in the counter we are going to use then
479 * make sure it is disabled.
480 */
481 event->hw.idx = idx;
482 mipsxx_pmu_disable_event(idx);
483 cpuc->events[idx] = event;
484
485 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
486 if (flags & PERF_EF_START)
487 mipspmu_start(event, PERF_EF_RELOAD);
488
489 /* Propagate our changes to the userspace mapping. */
490 perf_event_update_userpage(event);
491
492out:
493 perf_pmu_enable(event->pmu);
494 return err;
495}
496
497static void mipspmu_del(struct perf_event *event, int flags)
498{
499 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
500 struct hw_perf_event *hwc = &event->hw;
501 int idx = hwc->idx;
502
503 WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
504
505 mipspmu_stop(event, PERF_EF_UPDATE);
506 cpuc->events[idx] = NULL;
507 clear_bit(idx, cpuc->used_mask);
508
509 perf_event_update_userpage(event);
510}
511
512static void mipspmu_read(struct perf_event *event)
513{
514 struct hw_perf_event *hwc = &event->hw;
515
516 /* Don't read disabled counters! */
517 if (hwc->idx < 0)
518 return;
519
520 mipspmu_event_update(event, hwc, hwc->idx);
521}
522
523static void mipspmu_enable(struct pmu *pmu)
524{
525#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
526 write_unlock(&pmuint_rwlock);
527#endif
528 resume_local_counters();
529}
530
531/*
532 * MIPS performance counters can be per-TC. The control registers can
533 * not be directly accessed across CPUs. Hence if we want to do global
534 * control, we need cross CPU calls. on_each_cpu() can help us, but we
535 * can not make sure this function is called with interrupts enabled. So
536 * here we pause local counters and then grab a rwlock and leave the
537 * counters on other CPUs alone. If any counter interrupt raises while
538 * we own the write lock, simply pause local counters on that CPU and
539 * spin in the handler. Also we know we won't be switched to another
540 * CPU after pausing local counters and before grabbing the lock.
541 */
542static void mipspmu_disable(struct pmu *pmu)
543{
544 pause_local_counters();
545#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
546 write_lock(&pmuint_rwlock);
547#endif
548}
549
550static atomic_t active_events = ATOMIC_INIT(0);
551static DEFINE_MUTEX(pmu_reserve_mutex);
552static int (*save_perf_irq)(void);
553
554static int mipspmu_get_irq(void)
555{
556 int err;
557
558 if (mipspmu.irq >= 0) {
559 /* Request my own irq handler. */
560 err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq,
561 IRQF_PERCPU | IRQF_NOBALANCING |
562 IRQF_NO_THREAD | IRQF_NO_SUSPEND |
563 IRQF_SHARED,
564 "mips_perf_pmu", &mipspmu);
565 if (err) {
566 pr_warn("Unable to request IRQ%d for MIPS performance counters!\n",
567 mipspmu.irq);
568 }
569 } else if (cp0_perfcount_irq < 0) {
570 /*
571 * We are sharing the irq number with the timer interrupt.
572 */
573 save_perf_irq = perf_irq;
574 perf_irq = mipsxx_pmu_handle_shared_irq;
575 err = 0;
576 } else {
577 pr_warn("The platform hasn't properly defined its interrupt controller\n");
578 err = -ENOENT;
579 }
580
581 return err;
582}
583
584static void mipspmu_free_irq(void)
585{
586 if (mipspmu.irq >= 0)
587 free_irq(mipspmu.irq, &mipspmu);
588 else if (cp0_perfcount_irq < 0)
589 perf_irq = save_perf_irq;
590}
591
592/*
593 * mipsxx/rm9000/loongson2 have different performance counters, they have
594 * specific low-level init routines.
595 */
596static void reset_counters(void *arg);
597static int __hw_perf_event_init(struct perf_event *event);
598
599static void hw_perf_event_destroy(struct perf_event *event)
600{
601 if (atomic_dec_and_mutex_lock(&active_events,
602 &pmu_reserve_mutex)) {
603 /*
604 * We must not call the destroy function with interrupts
605 * disabled.
606 */
607 on_each_cpu(reset_counters,
608 (void *)(long)mipspmu.num_counters, 1);
609 mipspmu_free_irq();
610 mutex_unlock(&pmu_reserve_mutex);
611 }
612}
613
614static int mipspmu_event_init(struct perf_event *event)
615{
616 int err = 0;
617
618 /* does not support taken branch sampling */
619 if (has_branch_stack(event))
620 return -EOPNOTSUPP;
621
622 switch (event->attr.type) {
623 case PERF_TYPE_RAW:
624 case PERF_TYPE_HARDWARE:
625 case PERF_TYPE_HW_CACHE:
626 break;
627
628 default:
629 return -ENOENT;
630 }
631
632 if (event->cpu >= nr_cpumask_bits ||
633 (event->cpu >= 0 && !cpu_online(event->cpu)))
634 return -ENODEV;
635
636 if (!atomic_inc_not_zero(&active_events)) {
637 mutex_lock(&pmu_reserve_mutex);
638 if (atomic_read(&active_events) == 0)
639 err = mipspmu_get_irq();
640
641 if (!err)
642 atomic_inc(&active_events);
643 mutex_unlock(&pmu_reserve_mutex);
644 }
645
646 if (err)
647 return err;
648
649 return __hw_perf_event_init(event);
650}
651
652static struct pmu pmu = {
653 .pmu_enable = mipspmu_enable,
654 .pmu_disable = mipspmu_disable,
655 .event_init = mipspmu_event_init,
656 .add = mipspmu_add,
657 .del = mipspmu_del,
658 .start = mipspmu_start,
659 .stop = mipspmu_stop,
660 .read = mipspmu_read,
661};
662
663static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev)
664{
665/*
666 * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for
667 * event_id.
668 */
669#ifdef CONFIG_MIPS_MT_SMP
670 return ((unsigned int)pev->range << 24) |
671 (pev->cntr_mask & 0xffff00) |
672 (pev->event_id & 0xff);
673#else
674 return (pev->cntr_mask & 0xffff00) |
675 (pev->event_id & 0xff);
676#endif
677}
678
679static const struct mips_perf_event *mipspmu_map_general_event(int idx)
680{
681
682 if ((*mipspmu.general_event_map)[idx].cntr_mask == 0)
683 return ERR_PTR(-EOPNOTSUPP);
684 return &(*mipspmu.general_event_map)[idx];
685}
686
687static const struct mips_perf_event *mipspmu_map_cache_event(u64 config)
688{
689 unsigned int cache_type, cache_op, cache_result;
690 const struct mips_perf_event *pev;
691
692 cache_type = (config >> 0) & 0xff;
693 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
694 return ERR_PTR(-EINVAL);
695
696 cache_op = (config >> 8) & 0xff;
697 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
698 return ERR_PTR(-EINVAL);
699
700 cache_result = (config >> 16) & 0xff;
701 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
702 return ERR_PTR(-EINVAL);
703
704 pev = &((*mipspmu.cache_event_map)
705 [cache_type]
706 [cache_op]
707 [cache_result]);
708
709 if (pev->cntr_mask == 0)
710 return ERR_PTR(-EOPNOTSUPP);
711
712 return pev;
713
714}
715
716static int validate_group(struct perf_event *event)
717{
718 struct perf_event *sibling, *leader = event->group_leader;
719 struct cpu_hw_events fake_cpuc;
720
721 memset(&fake_cpuc, 0, sizeof(fake_cpuc));
722
723 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0)
724 return -EINVAL;
725
726 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
727 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0)
728 return -EINVAL;
729 }
730
731 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0)
732 return -EINVAL;
733
734 return 0;
735}
736
737/* This is needed by specific irq handlers in perf_event_*.c */
738static void handle_associated_event(struct cpu_hw_events *cpuc,
739 int idx, struct perf_sample_data *data,
740 struct pt_regs *regs)
741{
742 struct perf_event *event = cpuc->events[idx];
743 struct hw_perf_event *hwc = &event->hw;
744
745 mipspmu_event_update(event, hwc, idx);
746 data->period = event->hw.last_period;
747 if (!mipspmu_event_set_period(event, hwc, idx))
748 return;
749
750 if (perf_event_overflow(event, data, regs))
751 mipsxx_pmu_disable_event(idx);
752}
753
754
755static int __n_counters(void)
756{
757 if (!(read_c0_config1() & M_CONFIG1_PC))
758 return 0;
759 if (!(read_c0_perfctrl0() & M_PERFCTL_MORE))
760 return 1;
761 if (!(read_c0_perfctrl1() & M_PERFCTL_MORE))
762 return 2;
763 if (!(read_c0_perfctrl2() & M_PERFCTL_MORE))
764 return 3;
765
766 return 4;
767}
768
769static int n_counters(void)
770{
771 int counters;
772
773 switch (current_cpu_type()) {
774 case CPU_R10000:
775 counters = 2;
776 break;
777
778 case CPU_R12000:
779 case CPU_R14000:
780 case CPU_R16000:
781 counters = 4;
782 break;
783
784 default:
785 counters = __n_counters();
786 }
787
788 return counters;
789}
790
791static void reset_counters(void *arg)
792{
793 int counters = (int)(long)arg;
794 switch (counters) {
795 case 4:
796 mipsxx_pmu_write_control(3, 0);
797 mipspmu.write_counter(3, 0);
798 case 3:
799 mipsxx_pmu_write_control(2, 0);
800 mipspmu.write_counter(2, 0);
801 case 2:
802 mipsxx_pmu_write_control(1, 0);
803 mipspmu.write_counter(1, 0);
804 case 1:
805 mipsxx_pmu_write_control(0, 0);
806 mipspmu.write_counter(0, 0);
807 }
808}
809
810/* 24K/34K/1004K/interAptiv/loongson1 cores share the same event map. */
811static const struct mips_perf_event mipsxxcore_event_map
812 [PERF_COUNT_HW_MAX] = {
813 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
814 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
815 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T },
816 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
817};
818
819/* 74K/proAptiv core has different branch event code. */
820static const struct mips_perf_event mipsxxcore_event_map2
821 [PERF_COUNT_HW_MAX] = {
822 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
823 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
824 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T },
825 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T },
826};
827
828static const struct mips_perf_event loongson3_event_map[PERF_COUNT_HW_MAX] = {
829 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN },
830 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, CNTR_ODD },
831 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x01, CNTR_EVEN },
832 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x01, CNTR_ODD },
833};
834
835static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = {
836 [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
837 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL },
838 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL },
839 [PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL },
840 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL },
841 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL },
842 [PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL },
843};
844
845static const struct mips_perf_event bmips5000_event_map
846 [PERF_COUNT_HW_MAX] = {
847 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, T },
848 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
849 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
850};
851
852static const struct mips_perf_event xlp_event_map[PERF_COUNT_HW_MAX] = {
853 [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
854 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x18, CNTR_ALL }, /* PAPI_TOT_INS */
855 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */
856 [PERF_COUNT_HW_CACHE_MISSES] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */
857 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x1b, CNTR_ALL }, /* PAPI_BR_CN */
858 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x1c, CNTR_ALL }, /* PAPI_BR_MSP */
859};
860
861/* 24K/34K/1004K/interAptiv/loongson1 cores share the same cache event map. */
862static const struct mips_perf_event mipsxxcore_cache_map
863 [PERF_COUNT_HW_CACHE_MAX]
864 [PERF_COUNT_HW_CACHE_OP_MAX]
865 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
866[C(L1D)] = {
867 /*
868 * Like some other architectures (e.g. ARM), the performance
869 * counters don't differentiate between read and write
870 * accesses/misses, so this isn't strictly correct, but it's the
871 * best we can do. Writes and reads get combined.
872 */
873 [C(OP_READ)] = {
874 [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T },
875 [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T },
876 },
877 [C(OP_WRITE)] = {
878 [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T },
879 [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T },
880 },
881},
882[C(L1I)] = {
883 [C(OP_READ)] = {
884 [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T },
885 [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T },
886 },
887 [C(OP_WRITE)] = {
888 [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T },
889 [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T },
890 },
891 [C(OP_PREFETCH)] = {
892 [C(RESULT_ACCESS)] = { 0x14, CNTR_EVEN, T },
893 /*
894 * Note that MIPS has only "hit" events countable for
895 * the prefetch operation.
896 */
897 },
898},
899[C(LL)] = {
900 [C(OP_READ)] = {
901 [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P },
902 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P },
903 },
904 [C(OP_WRITE)] = {
905 [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P },
906 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P },
907 },
908},
909[C(DTLB)] = {
910 [C(OP_READ)] = {
911 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
912 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
913 },
914 [C(OP_WRITE)] = {
915 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
916 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
917 },
918},
919[C(ITLB)] = {
920 [C(OP_READ)] = {
921 [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T },
922 [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T },
923 },
924 [C(OP_WRITE)] = {
925 [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T },
926 [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T },
927 },
928},
929[C(BPU)] = {
930 /* Using the same code for *HW_BRANCH* */
931 [C(OP_READ)] = {
932 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T },
933 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
934 },
935 [C(OP_WRITE)] = {
936 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T },
937 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
938 },
939},
940};
941
942/* 74K/proAptiv core has completely different cache event map. */
943static const struct mips_perf_event mipsxxcore_cache_map2
944 [PERF_COUNT_HW_CACHE_MAX]
945 [PERF_COUNT_HW_CACHE_OP_MAX]
946 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
947[C(L1D)] = {
948 /*
949 * Like some other architectures (e.g. ARM), the performance
950 * counters don't differentiate between read and write
951 * accesses/misses, so this isn't strictly correct, but it's the
952 * best we can do. Writes and reads get combined.
953 */
954 [C(OP_READ)] = {
955 [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T },
956 [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T },
957 },
958 [C(OP_WRITE)] = {
959 [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T },
960 [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T },
961 },
962},
963[C(L1I)] = {
964 [C(OP_READ)] = {
965 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
966 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
967 },
968 [C(OP_WRITE)] = {
969 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
970 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
971 },
972 [C(OP_PREFETCH)] = {
973 [C(RESULT_ACCESS)] = { 0x34, CNTR_EVEN, T },
974 /*
975 * Note that MIPS has only "hit" events countable for
976 * the prefetch operation.
977 */
978 },
979},
980[C(LL)] = {
981 [C(OP_READ)] = {
982 [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P },
983 [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN, P },
984 },
985 [C(OP_WRITE)] = {
986 [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P },
987 [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN, P },
988 },
989},
990/*
991 * 74K core does not have specific DTLB events. proAptiv core has
992 * "speculative" DTLB events which are numbered 0x63 (even/odd) and
993 * not included here. One can use raw events if really needed.
994 */
995[C(ITLB)] = {
996 [C(OP_READ)] = {
997 [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T },
998 [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T },
999 },
1000 [C(OP_WRITE)] = {
1001 [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T },
1002 [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T },
1003 },
1004},
1005[C(BPU)] = {
1006 /* Using the same code for *HW_BRANCH* */
1007 [C(OP_READ)] = {
1008 [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T },
1009 [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T },
1010 },
1011 [C(OP_WRITE)] = {
1012 [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T },
1013 [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T },
1014 },
1015},
1016};
1017
1018static const struct mips_perf_event loongson3_cache_map
1019 [PERF_COUNT_HW_CACHE_MAX]
1020 [PERF_COUNT_HW_CACHE_OP_MAX]
1021 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1022[C(L1D)] = {
1023 /*
1024 * Like some other architectures (e.g. ARM), the performance
1025 * counters don't differentiate between read and write
1026 * accesses/misses, so this isn't strictly correct, but it's the
1027 * best we can do. Writes and reads get combined.
1028 */
1029 [C(OP_READ)] = {
1030 [C(RESULT_MISS)] = { 0x04, CNTR_ODD },
1031 },
1032 [C(OP_WRITE)] = {
1033 [C(RESULT_MISS)] = { 0x04, CNTR_ODD },
1034 },
1035},
1036[C(L1I)] = {
1037 [C(OP_READ)] = {
1038 [C(RESULT_MISS)] = { 0x04, CNTR_EVEN },
1039 },
1040 [C(OP_WRITE)] = {
1041 [C(RESULT_MISS)] = { 0x04, CNTR_EVEN },
1042 },
1043},
1044[C(DTLB)] = {
1045 [C(OP_READ)] = {
1046 [C(RESULT_MISS)] = { 0x09, CNTR_ODD },
1047 },
1048 [C(OP_WRITE)] = {
1049 [C(RESULT_MISS)] = { 0x09, CNTR_ODD },
1050 },
1051},
1052[C(ITLB)] = {
1053 [C(OP_READ)] = {
1054 [C(RESULT_MISS)] = { 0x0c, CNTR_ODD },
1055 },
1056 [C(OP_WRITE)] = {
1057 [C(RESULT_MISS)] = { 0x0c, CNTR_ODD },
1058 },
1059},
1060[C(BPU)] = {
1061 /* Using the same code for *HW_BRANCH* */
1062 [C(OP_READ)] = {
1063 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN },
1064 [C(RESULT_MISS)] = { 0x02, CNTR_ODD },
1065 },
1066 [C(OP_WRITE)] = {
1067 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN },
1068 [C(RESULT_MISS)] = { 0x02, CNTR_ODD },
1069 },
1070},
1071};
1072
1073/* BMIPS5000 */
1074static const struct mips_perf_event bmips5000_cache_map
1075 [PERF_COUNT_HW_CACHE_MAX]
1076 [PERF_COUNT_HW_CACHE_OP_MAX]
1077 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1078[C(L1D)] = {
1079 /*
1080 * Like some other architectures (e.g. ARM), the performance
1081 * counters don't differentiate between read and write
1082 * accesses/misses, so this isn't strictly correct, but it's the
1083 * best we can do. Writes and reads get combined.
1084 */
1085 [C(OP_READ)] = {
1086 [C(RESULT_ACCESS)] = { 12, CNTR_EVEN, T },
1087 [C(RESULT_MISS)] = { 12, CNTR_ODD, T },
1088 },
1089 [C(OP_WRITE)] = {
1090 [C(RESULT_ACCESS)] = { 12, CNTR_EVEN, T },
1091 [C(RESULT_MISS)] = { 12, CNTR_ODD, T },
1092 },
1093},
1094[C(L1I)] = {
1095 [C(OP_READ)] = {
1096 [C(RESULT_ACCESS)] = { 10, CNTR_EVEN, T },
1097 [C(RESULT_MISS)] = { 10, CNTR_ODD, T },
1098 },
1099 [C(OP_WRITE)] = {
1100 [C(RESULT_ACCESS)] = { 10, CNTR_EVEN, T },
1101 [C(RESULT_MISS)] = { 10, CNTR_ODD, T },
1102 },
1103 [C(OP_PREFETCH)] = {
1104 [C(RESULT_ACCESS)] = { 23, CNTR_EVEN, T },
1105 /*
1106 * Note that MIPS has only "hit" events countable for
1107 * the prefetch operation.
1108 */
1109 },
1110},
1111[C(LL)] = {
1112 [C(OP_READ)] = {
1113 [C(RESULT_ACCESS)] = { 28, CNTR_EVEN, P },
1114 [C(RESULT_MISS)] = { 28, CNTR_ODD, P },
1115 },
1116 [C(OP_WRITE)] = {
1117 [C(RESULT_ACCESS)] = { 28, CNTR_EVEN, P },
1118 [C(RESULT_MISS)] = { 28, CNTR_ODD, P },
1119 },
1120},
1121[C(BPU)] = {
1122 /* Using the same code for *HW_BRANCH* */
1123 [C(OP_READ)] = {
1124 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
1125 },
1126 [C(OP_WRITE)] = {
1127 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
1128 },
1129},
1130};
1131
1132
1133static const struct mips_perf_event octeon_cache_map
1134 [PERF_COUNT_HW_CACHE_MAX]
1135 [PERF_COUNT_HW_CACHE_OP_MAX]
1136 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1137[C(L1D)] = {
1138 [C(OP_READ)] = {
1139 [C(RESULT_ACCESS)] = { 0x2b, CNTR_ALL },
1140 [C(RESULT_MISS)] = { 0x2e, CNTR_ALL },
1141 },
1142 [C(OP_WRITE)] = {
1143 [C(RESULT_ACCESS)] = { 0x30, CNTR_ALL },
1144 },
1145},
1146[C(L1I)] = {
1147 [C(OP_READ)] = {
1148 [C(RESULT_ACCESS)] = { 0x18, CNTR_ALL },
1149 },
1150 [C(OP_PREFETCH)] = {
1151 [C(RESULT_ACCESS)] = { 0x19, CNTR_ALL },
1152 },
1153},
1154[C(DTLB)] = {
1155 /*
1156 * Only general DTLB misses are counted use the same event for
1157 * read and write.
1158 */
1159 [C(OP_READ)] = {
1160 [C(RESULT_MISS)] = { 0x35, CNTR_ALL },
1161 },
1162 [C(OP_WRITE)] = {
1163 [C(RESULT_MISS)] = { 0x35, CNTR_ALL },
1164 },
1165},
1166[C(ITLB)] = {
1167 [C(OP_READ)] = {
1168 [C(RESULT_MISS)] = { 0x37, CNTR_ALL },
1169 },
1170},
1171};
1172
1173static const struct mips_perf_event xlp_cache_map
1174 [PERF_COUNT_HW_CACHE_MAX]
1175 [PERF_COUNT_HW_CACHE_OP_MAX]
1176 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1177[C(L1D)] = {
1178 [C(OP_READ)] = {
1179 [C(RESULT_ACCESS)] = { 0x31, CNTR_ALL }, /* PAPI_L1_DCR */
1180 [C(RESULT_MISS)] = { 0x30, CNTR_ALL }, /* PAPI_L1_LDM */
1181 },
1182 [C(OP_WRITE)] = {
1183 [C(RESULT_ACCESS)] = { 0x2f, CNTR_ALL }, /* PAPI_L1_DCW */
1184 [C(RESULT_MISS)] = { 0x2e, CNTR_ALL }, /* PAPI_L1_STM */
1185 },
1186},
1187[C(L1I)] = {
1188 [C(OP_READ)] = {
1189 [C(RESULT_ACCESS)] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */
1190 [C(RESULT_MISS)] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */
1191 },
1192},
1193[C(LL)] = {
1194 [C(OP_READ)] = {
1195 [C(RESULT_ACCESS)] = { 0x35, CNTR_ALL }, /* PAPI_L2_DCR */
1196 [C(RESULT_MISS)] = { 0x37, CNTR_ALL }, /* PAPI_L2_LDM */
1197 },
1198 [C(OP_WRITE)] = {
1199 [C(RESULT_ACCESS)] = { 0x34, CNTR_ALL }, /* PAPI_L2_DCA */
1200 [C(RESULT_MISS)] = { 0x36, CNTR_ALL }, /* PAPI_L2_DCM */
1201 },
1202},
1203[C(DTLB)] = {
1204 /*
1205 * Only general DTLB misses are counted use the same event for
1206 * read and write.
1207 */
1208 [C(OP_READ)] = {
1209 [C(RESULT_MISS)] = { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */
1210 },
1211 [C(OP_WRITE)] = {
1212 [C(RESULT_MISS)] = { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */
1213 },
1214},
1215[C(ITLB)] = {
1216 [C(OP_READ)] = {
1217 [C(RESULT_MISS)] = { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */
1218 },
1219 [C(OP_WRITE)] = {
1220 [C(RESULT_MISS)] = { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */
1221 },
1222},
1223[C(BPU)] = {
1224 [C(OP_READ)] = {
1225 [C(RESULT_MISS)] = { 0x25, CNTR_ALL },
1226 },
1227},
1228};
1229
1230#ifdef CONFIG_MIPS_MT_SMP
1231static void check_and_calc_range(struct perf_event *event,
1232 const struct mips_perf_event *pev)
1233{
1234 struct hw_perf_event *hwc = &event->hw;
1235
1236 if (event->cpu >= 0) {
1237 if (pev->range > V) {
1238 /*
1239 * The user selected an event that is processor
1240 * wide, while expecting it to be VPE wide.
1241 */
1242 hwc->config_base |= M_TC_EN_ALL;
1243 } else {
1244 /*
1245 * FIXME: cpu_data[event->cpu].vpe_id reports 0
1246 * for both CPUs.
1247 */
1248 hwc->config_base |= M_PERFCTL_VPEID(event->cpu);
1249 hwc->config_base |= M_TC_EN_VPE;
1250 }
1251 } else
1252 hwc->config_base |= M_TC_EN_ALL;
1253}
1254#else
1255static void check_and_calc_range(struct perf_event *event,
1256 const struct mips_perf_event *pev)
1257{
1258}
1259#endif
1260
1261static int __hw_perf_event_init(struct perf_event *event)
1262{
1263 struct perf_event_attr *attr = &event->attr;
1264 struct hw_perf_event *hwc = &event->hw;
1265 const struct mips_perf_event *pev;
1266 int err;
1267
1268 /* Returning MIPS event descriptor for generic perf event. */
1269 if (PERF_TYPE_HARDWARE == event->attr.type) {
1270 if (event->attr.config >= PERF_COUNT_HW_MAX)
1271 return -EINVAL;
1272 pev = mipspmu_map_general_event(event->attr.config);
1273 } else if (PERF_TYPE_HW_CACHE == event->attr.type) {
1274 pev = mipspmu_map_cache_event(event->attr.config);
1275 } else if (PERF_TYPE_RAW == event->attr.type) {
1276 /* We are working on the global raw event. */
1277 mutex_lock(&raw_event_mutex);
1278 pev = mipspmu.map_raw_event(event->attr.config);
1279 } else {
1280 /* The event type is not (yet) supported. */
1281 return -EOPNOTSUPP;
1282 }
1283
1284 if (IS_ERR(pev)) {
1285 if (PERF_TYPE_RAW == event->attr.type)
1286 mutex_unlock(&raw_event_mutex);
1287 return PTR_ERR(pev);
1288 }
1289
1290 /*
1291 * We allow max flexibility on how each individual counter shared
1292 * by the single CPU operates (the mode exclusion and the range).
1293 */
1294 hwc->config_base = M_PERFCTL_INTERRUPT_ENABLE;
1295
1296 /* Calculate range bits and validate it. */
1297 if (num_possible_cpus() > 1)
1298 check_and_calc_range(event, pev);
1299
1300 hwc->event_base = mipspmu_perf_event_encode(pev);
1301 if (PERF_TYPE_RAW == event->attr.type)
1302 mutex_unlock(&raw_event_mutex);
1303
1304 if (!attr->exclude_user)
1305 hwc->config_base |= M_PERFCTL_USER;
1306 if (!attr->exclude_kernel) {
1307 hwc->config_base |= M_PERFCTL_KERNEL;
1308 /* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */
1309 hwc->config_base |= M_PERFCTL_EXL;
1310 }
1311 if (!attr->exclude_hv)
1312 hwc->config_base |= M_PERFCTL_SUPERVISOR;
1313
1314 hwc->config_base &= M_PERFCTL_CONFIG_MASK;
1315 /*
1316 * The event can belong to another cpu. We do not assign a local
1317 * counter for it for now.
1318 */
1319 hwc->idx = -1;
1320 hwc->config = 0;
1321
1322 if (!hwc->sample_period) {
1323 hwc->sample_period = mipspmu.max_period;
1324 hwc->last_period = hwc->sample_period;
1325 local64_set(&hwc->period_left, hwc->sample_period);
1326 }
1327
1328 err = 0;
1329 if (event->group_leader != event)
1330 err = validate_group(event);
1331
1332 event->destroy = hw_perf_event_destroy;
1333
1334 if (err)
1335 event->destroy(event);
1336
1337 return err;
1338}
1339
1340static void pause_local_counters(void)
1341{
1342 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1343 int ctr = mipspmu.num_counters;
1344 unsigned long flags;
1345
1346 local_irq_save(flags);
1347 do {
1348 ctr--;
1349 cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr);
1350 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] &
1351 ~M_PERFCTL_COUNT_EVENT_WHENEVER);
1352 } while (ctr > 0);
1353 local_irq_restore(flags);
1354}
1355
1356static void resume_local_counters(void)
1357{
1358 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1359 int ctr = mipspmu.num_counters;
1360
1361 do {
1362 ctr--;
1363 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]);
1364 } while (ctr > 0);
1365}
1366
1367static int mipsxx_pmu_handle_shared_irq(void)
1368{
1369 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1370 struct perf_sample_data data;
1371 unsigned int counters = mipspmu.num_counters;
1372 u64 counter;
1373 int handled = IRQ_NONE;
1374 struct pt_regs *regs;
1375
1376 if (cpu_has_perf_cntr_intr_bit && !(read_c0_cause() & CAUSEF_PCI))
1377 return handled;
1378 /*
1379 * First we pause the local counters, so that when we are locked
1380 * here, the counters are all paused. When it gets locked due to
1381 * perf_disable(), the timer interrupt handler will be delayed.
1382 *
1383 * See also mipsxx_pmu_start().
1384 */
1385 pause_local_counters();
1386#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1387 read_lock(&pmuint_rwlock);
1388#endif
1389
1390 regs = get_irq_regs();
1391
1392 perf_sample_data_init(&data, 0, 0);
1393
1394 switch (counters) {
1395#define HANDLE_COUNTER(n) \
1396 case n + 1: \
1397 if (test_bit(n, cpuc->used_mask)) { \
1398 counter = mipspmu.read_counter(n); \
1399 if (counter & mipspmu.overflow) { \
1400 handle_associated_event(cpuc, n, &data, regs); \
1401 handled = IRQ_HANDLED; \
1402 } \
1403 }
1404 HANDLE_COUNTER(3)
1405 HANDLE_COUNTER(2)
1406 HANDLE_COUNTER(1)
1407 HANDLE_COUNTER(0)
1408 }
1409
1410 /*
1411 * Do all the work for the pending perf events. We can do this
1412 * in here because the performance counter interrupt is a regular
1413 * interrupt, not NMI.
1414 */
1415 if (handled == IRQ_HANDLED)
1416 irq_work_run();
1417
1418#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1419 read_unlock(&pmuint_rwlock);
1420#endif
1421 resume_local_counters();
1422 return handled;
1423}
1424
1425static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev)
1426{
1427 return mipsxx_pmu_handle_shared_irq();
1428}
1429
1430/* 24K */
1431#define IS_BOTH_COUNTERS_24K_EVENT(b) \
1432 ((b) == 0 || (b) == 1 || (b) == 11)
1433
1434/* 34K */
1435#define IS_BOTH_COUNTERS_34K_EVENT(b) \
1436 ((b) == 0 || (b) == 1 || (b) == 11)
1437#ifdef CONFIG_MIPS_MT_SMP
1438#define IS_RANGE_P_34K_EVENT(r, b) \
1439 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \
1440 (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 || \
1441 (r) == 176 || ((b) >= 50 && (b) <= 55) || \
1442 ((b) >= 64 && (b) <= 67))
1443#define IS_RANGE_V_34K_EVENT(r) ((r) == 47)
1444#endif
1445
1446/* 74K */
1447#define IS_BOTH_COUNTERS_74K_EVENT(b) \
1448 ((b) == 0 || (b) == 1)
1449
1450/* proAptiv */
1451#define IS_BOTH_COUNTERS_PROAPTIV_EVENT(b) \
1452 ((b) == 0 || (b) == 1)
1453/* P5600 */
1454#define IS_BOTH_COUNTERS_P5600_EVENT(b) \
1455 ((b) == 0 || (b) == 1)
1456
1457/* 1004K */
1458#define IS_BOTH_COUNTERS_1004K_EVENT(b) \
1459 ((b) == 0 || (b) == 1 || (b) == 11)
1460#ifdef CONFIG_MIPS_MT_SMP
1461#define IS_RANGE_P_1004K_EVENT(r, b) \
1462 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \
1463 (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 || \
1464 (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) || \
1465 (r) == 188 || (b) == 61 || (b) == 62 || \
1466 ((b) >= 64 && (b) <= 67))
1467#define IS_RANGE_V_1004K_EVENT(r) ((r) == 47)
1468#endif
1469
1470/* interAptiv */
1471#define IS_BOTH_COUNTERS_INTERAPTIV_EVENT(b) \
1472 ((b) == 0 || (b) == 1 || (b) == 11)
1473#ifdef CONFIG_MIPS_MT_SMP
1474/* The P/V/T info is not provided for "(b) == 38" in SUM, assume P. */
1475#define IS_RANGE_P_INTERAPTIV_EVENT(r, b) \
1476 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \
1477 (b) == 25 || (b) == 36 || (b) == 38 || (b) == 39 || \
1478 (r) == 44 || (r) == 174 || (r) == 176 || ((b) >= 50 && \
1479 (b) <= 59) || (r) == 188 || (b) == 61 || (b) == 62 || \
1480 ((b) >= 64 && (b) <= 67))
1481#define IS_RANGE_V_INTERAPTIV_EVENT(r) ((r) == 47 || (r) == 175)
1482#endif
1483
1484/* BMIPS5000 */
1485#define IS_BOTH_COUNTERS_BMIPS5000_EVENT(b) \
1486 ((b) == 0 || (b) == 1)
1487
1488
1489/*
1490 * For most cores the user can use 0-255 raw events, where 0-127 for the events
1491 * of even counters, and 128-255 for odd counters. Note that bit 7 is used to
1492 * indicate the even/odd bank selector. So, for example, when user wants to take
1493 * the Event Num of 15 for odd counters (by referring to the user manual), then
1494 * 128 needs to be added to 15 as the input for the event config, i.e., 143 (0x8F)
1495 * to be used.
1496 *
1497 * Some newer cores have even more events, in which case the user can use raw
1498 * events 0-511, where 0-255 are for the events of even counters, and 256-511
1499 * are for odd counters, so bit 8 is used to indicate the even/odd bank selector.
1500 */
1501static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config)
1502{
1503 /* currently most cores have 7-bit event numbers */
1504 unsigned int raw_id = config & 0xff;
1505 unsigned int base_id = raw_id & 0x7f;
1506
1507 switch (current_cpu_type()) {
1508 case CPU_24K:
1509 if (IS_BOTH_COUNTERS_24K_EVENT(base_id))
1510 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1511 else
1512 raw_event.cntr_mask =
1513 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1514#ifdef CONFIG_MIPS_MT_SMP
1515 /*
1516 * This is actually doing nothing. Non-multithreading
1517 * CPUs will not check and calculate the range.
1518 */
1519 raw_event.range = P;
1520#endif
1521 break;
1522 case CPU_34K:
1523 if (IS_BOTH_COUNTERS_34K_EVENT(base_id))
1524 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1525 else
1526 raw_event.cntr_mask =
1527 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1528#ifdef CONFIG_MIPS_MT_SMP
1529 if (IS_RANGE_P_34K_EVENT(raw_id, base_id))
1530 raw_event.range = P;
1531 else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id)))
1532 raw_event.range = V;
1533 else
1534 raw_event.range = T;
1535#endif
1536 break;
1537 case CPU_74K:
1538 case CPU_1074K:
1539 if (IS_BOTH_COUNTERS_74K_EVENT(base_id))
1540 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1541 else
1542 raw_event.cntr_mask =
1543 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1544#ifdef CONFIG_MIPS_MT_SMP
1545 raw_event.range = P;
1546#endif
1547 break;
1548 case CPU_PROAPTIV:
1549 if (IS_BOTH_COUNTERS_PROAPTIV_EVENT(base_id))
1550 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1551 else
1552 raw_event.cntr_mask =
1553 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1554#ifdef CONFIG_MIPS_MT_SMP
1555 raw_event.range = P;
1556#endif
1557 break;
1558 case CPU_P5600:
1559 case CPU_I6400:
1560 /* 8-bit event numbers */
1561 raw_id = config & 0x1ff;
1562 base_id = raw_id & 0xff;
1563 if (IS_BOTH_COUNTERS_P5600_EVENT(base_id))
1564 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1565 else
1566 raw_event.cntr_mask =
1567 raw_id > 255 ? CNTR_ODD : CNTR_EVEN;
1568#ifdef CONFIG_MIPS_MT_SMP
1569 raw_event.range = P;
1570#endif
1571 break;
1572 case CPU_1004K:
1573 if (IS_BOTH_COUNTERS_1004K_EVENT(base_id))
1574 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1575 else
1576 raw_event.cntr_mask =
1577 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1578#ifdef CONFIG_MIPS_MT_SMP
1579 if (IS_RANGE_P_1004K_EVENT(raw_id, base_id))
1580 raw_event.range = P;
1581 else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id)))
1582 raw_event.range = V;
1583 else
1584 raw_event.range = T;
1585#endif
1586 break;
1587 case CPU_INTERAPTIV:
1588 if (IS_BOTH_COUNTERS_INTERAPTIV_EVENT(base_id))
1589 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1590 else
1591 raw_event.cntr_mask =
1592 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1593#ifdef CONFIG_MIPS_MT_SMP
1594 if (IS_RANGE_P_INTERAPTIV_EVENT(raw_id, base_id))
1595 raw_event.range = P;
1596 else if (unlikely(IS_RANGE_V_INTERAPTIV_EVENT(raw_id)))
1597 raw_event.range = V;
1598 else
1599 raw_event.range = T;
1600#endif
1601 break;
1602 case CPU_BMIPS5000:
1603 if (IS_BOTH_COUNTERS_BMIPS5000_EVENT(base_id))
1604 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1605 else
1606 raw_event.cntr_mask =
1607 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1608 break;
1609 case CPU_LOONGSON3:
1610 raw_event.cntr_mask = raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1611 break;
1612 }
1613
1614 raw_event.event_id = base_id;
1615
1616 return &raw_event;
1617}
1618
1619static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config)
1620{
1621 unsigned int raw_id = config & 0xff;
1622 unsigned int base_id = raw_id & 0x7f;
1623
1624
1625 raw_event.cntr_mask = CNTR_ALL;
1626 raw_event.event_id = base_id;
1627
1628 if (current_cpu_type() == CPU_CAVIUM_OCTEON2) {
1629 if (base_id > 0x42)
1630 return ERR_PTR(-EOPNOTSUPP);
1631 } else {
1632 if (base_id > 0x3a)
1633 return ERR_PTR(-EOPNOTSUPP);
1634 }
1635
1636 switch (base_id) {
1637 case 0x00:
1638 case 0x0f:
1639 case 0x1e:
1640 case 0x1f:
1641 case 0x2f:
1642 case 0x34:
1643 case 0x3b ... 0x3f:
1644 return ERR_PTR(-EOPNOTSUPP);
1645 default:
1646 break;
1647 }
1648
1649 return &raw_event;
1650}
1651
1652static const struct mips_perf_event *xlp_pmu_map_raw_event(u64 config)
1653{
1654 unsigned int raw_id = config & 0xff;
1655
1656 /* Only 1-63 are defined */
1657 if ((raw_id < 0x01) || (raw_id > 0x3f))
1658 return ERR_PTR(-EOPNOTSUPP);
1659
1660 raw_event.cntr_mask = CNTR_ALL;
1661 raw_event.event_id = raw_id;
1662
1663 return &raw_event;
1664}
1665
1666static int __init
1667init_hw_perf_events(void)
1668{
1669 int counters, irq;
1670 int counter_bits;
1671
1672 pr_info("Performance counters: ");
1673
1674 counters = n_counters();
1675 if (counters == 0) {
1676 pr_cont("No available PMU.\n");
1677 return -ENODEV;
1678 }
1679
1680#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1681 cpu_has_mipsmt_pertccounters = read_c0_config7() & (1<<19);
1682 if (!cpu_has_mipsmt_pertccounters)
1683 counters = counters_total_to_per_cpu(counters);
1684#endif
1685
1686 if (get_c0_perfcount_int)
1687 irq = get_c0_perfcount_int();
1688 else if (cp0_perfcount_irq >= 0)
1689 irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq;
1690 else
1691 irq = -1;
1692
1693 mipspmu.map_raw_event = mipsxx_pmu_map_raw_event;
1694
1695 switch (current_cpu_type()) {
1696 case CPU_24K:
1697 mipspmu.name = "mips/24K";
1698 mipspmu.general_event_map = &mipsxxcore_event_map;
1699 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1700 break;
1701 case CPU_34K:
1702 mipspmu.name = "mips/34K";
1703 mipspmu.general_event_map = &mipsxxcore_event_map;
1704 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1705 break;
1706 case CPU_74K:
1707 mipspmu.name = "mips/74K";
1708 mipspmu.general_event_map = &mipsxxcore_event_map2;
1709 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1710 break;
1711 case CPU_PROAPTIV:
1712 mipspmu.name = "mips/proAptiv";
1713 mipspmu.general_event_map = &mipsxxcore_event_map2;
1714 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1715 break;
1716 case CPU_P5600:
1717 mipspmu.name = "mips/P5600";
1718 mipspmu.general_event_map = &mipsxxcore_event_map2;
1719 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1720 break;
1721 case CPU_I6400:
1722 mipspmu.name = "mips/I6400";
1723 mipspmu.general_event_map = &mipsxxcore_event_map2;
1724 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1725 break;
1726 case CPU_1004K:
1727 mipspmu.name = "mips/1004K";
1728 mipspmu.general_event_map = &mipsxxcore_event_map;
1729 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1730 break;
1731 case CPU_1074K:
1732 mipspmu.name = "mips/1074K";
1733 mipspmu.general_event_map = &mipsxxcore_event_map;
1734 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1735 break;
1736 case CPU_INTERAPTIV:
1737 mipspmu.name = "mips/interAptiv";
1738 mipspmu.general_event_map = &mipsxxcore_event_map;
1739 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1740 break;
1741 case CPU_LOONGSON1:
1742 mipspmu.name = "mips/loongson1";
1743 mipspmu.general_event_map = &mipsxxcore_event_map;
1744 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1745 break;
1746 case CPU_LOONGSON3:
1747 mipspmu.name = "mips/loongson3";
1748 mipspmu.general_event_map = &loongson3_event_map;
1749 mipspmu.cache_event_map = &loongson3_cache_map;
1750 break;
1751 case CPU_CAVIUM_OCTEON:
1752 case CPU_CAVIUM_OCTEON_PLUS:
1753 case CPU_CAVIUM_OCTEON2:
1754 mipspmu.name = "octeon";
1755 mipspmu.general_event_map = &octeon_event_map;
1756 mipspmu.cache_event_map = &octeon_cache_map;
1757 mipspmu.map_raw_event = octeon_pmu_map_raw_event;
1758 break;
1759 case CPU_BMIPS5000:
1760 mipspmu.name = "BMIPS5000";
1761 mipspmu.general_event_map = &bmips5000_event_map;
1762 mipspmu.cache_event_map = &bmips5000_cache_map;
1763 break;
1764 case CPU_XLP:
1765 mipspmu.name = "xlp";
1766 mipspmu.general_event_map = &xlp_event_map;
1767 mipspmu.cache_event_map = &xlp_cache_map;
1768 mipspmu.map_raw_event = xlp_pmu_map_raw_event;
1769 break;
1770 default:
1771 pr_cont("Either hardware does not support performance "
1772 "counters, or not yet implemented.\n");
1773 return -ENODEV;
1774 }
1775
1776 mipspmu.num_counters = counters;
1777 mipspmu.irq = irq;
1778
1779 if (read_c0_perfctrl0() & M_PERFCTL_WIDE) {
1780 mipspmu.max_period = (1ULL << 63) - 1;
1781 mipspmu.valid_count = (1ULL << 63) - 1;
1782 mipspmu.overflow = 1ULL << 63;
1783 mipspmu.read_counter = mipsxx_pmu_read_counter_64;
1784 mipspmu.write_counter = mipsxx_pmu_write_counter_64;
1785 counter_bits = 64;
1786 } else {
1787 mipspmu.max_period = (1ULL << 31) - 1;
1788 mipspmu.valid_count = (1ULL << 31) - 1;
1789 mipspmu.overflow = 1ULL << 31;
1790 mipspmu.read_counter = mipsxx_pmu_read_counter;
1791 mipspmu.write_counter = mipsxx_pmu_write_counter;
1792 counter_bits = 32;
1793 }
1794
1795 on_each_cpu(reset_counters, (void *)(long)counters, 1);
1796
1797 pr_cont("%s PMU enabled, %d %d-bit counters available to each "
1798 "CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq,
1799 irq < 0 ? " (share with timer interrupt)" : "");
1800
1801 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1802
1803 return 0;
1804}
1805early_initcall(init_hw_perf_events);
1/*
2 * Linux performance counter support for MIPS.
3 *
4 * Copyright (C) 2010 MIPS Technologies, Inc.
5 * Copyright (C) 2011 Cavium Networks, Inc.
6 * Author: Deng-Cheng Zhu
7 *
8 * This code is based on the implementation for ARM, which is in turn
9 * based on the sparc64 perf event code and the x86 code. Performance
10 * counter access is based on the MIPS Oprofile code. And the callchain
11 * support references the code of MIPS stacktrace.c.
12 *
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License version 2 as
15 * published by the Free Software Foundation.
16 */
17
18#include <linux/cpumask.h>
19#include <linux/interrupt.h>
20#include <linux/smp.h>
21#include <linux/kernel.h>
22#include <linux/perf_event.h>
23#include <linux/uaccess.h>
24
25#include <asm/irq.h>
26#include <asm/irq_regs.h>
27#include <asm/stacktrace.h>
28#include <asm/time.h> /* For perf_irq */
29
30#define MIPS_MAX_HWEVENTS 4
31#define MIPS_TCS_PER_COUNTER 2
32#define MIPS_CPUID_TO_COUNTER_MASK (MIPS_TCS_PER_COUNTER - 1)
33
34struct cpu_hw_events {
35 /* Array of events on this cpu. */
36 struct perf_event *events[MIPS_MAX_HWEVENTS];
37
38 /*
39 * Set the bit (indexed by the counter number) when the counter
40 * is used for an event.
41 */
42 unsigned long used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)];
43
44 /*
45 * Software copy of the control register for each performance counter.
46 * MIPS CPUs vary in performance counters. They use this differently,
47 * and even may not use it.
48 */
49 unsigned int saved_ctrl[MIPS_MAX_HWEVENTS];
50};
51DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
52 .saved_ctrl = {0},
53};
54
55/* The description of MIPS performance events. */
56struct mips_perf_event {
57 unsigned int event_id;
58 /*
59 * MIPS performance counters are indexed starting from 0.
60 * CNTR_EVEN indicates the indexes of the counters to be used are
61 * even numbers.
62 */
63 unsigned int cntr_mask;
64 #define CNTR_EVEN 0x55555555
65 #define CNTR_ODD 0xaaaaaaaa
66 #define CNTR_ALL 0xffffffff
67#ifdef CONFIG_MIPS_MT_SMP
68 enum {
69 T = 0,
70 V = 1,
71 P = 2,
72 } range;
73#else
74 #define T
75 #define V
76 #define P
77#endif
78};
79
80static struct mips_perf_event raw_event;
81static DEFINE_MUTEX(raw_event_mutex);
82
83#define C(x) PERF_COUNT_HW_CACHE_##x
84
85struct mips_pmu {
86 u64 max_period;
87 u64 valid_count;
88 u64 overflow;
89 const char *name;
90 int irq;
91 u64 (*read_counter)(unsigned int idx);
92 void (*write_counter)(unsigned int idx, u64 val);
93 const struct mips_perf_event *(*map_raw_event)(u64 config);
94 const struct mips_perf_event (*general_event_map)[PERF_COUNT_HW_MAX];
95 const struct mips_perf_event (*cache_event_map)
96 [PERF_COUNT_HW_CACHE_MAX]
97 [PERF_COUNT_HW_CACHE_OP_MAX]
98 [PERF_COUNT_HW_CACHE_RESULT_MAX];
99 unsigned int num_counters;
100};
101
102static struct mips_pmu mipspmu;
103
104#define M_PERFCTL_EVENT(event) (((event) << MIPS_PERFCTRL_EVENT_S) & \
105 MIPS_PERFCTRL_EVENT)
106#define M_PERFCTL_VPEID(vpe) ((vpe) << MIPS_PERFCTRL_VPEID_S)
107
108#ifdef CONFIG_CPU_BMIPS5000
109#define M_PERFCTL_MT_EN(filter) 0
110#else /* !CONFIG_CPU_BMIPS5000 */
111#define M_PERFCTL_MT_EN(filter) (filter)
112#endif /* CONFIG_CPU_BMIPS5000 */
113
114#define M_TC_EN_ALL M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_ALL)
115#define M_TC_EN_VPE M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_VPE)
116#define M_TC_EN_TC M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_TC)
117
118#define M_PERFCTL_COUNT_EVENT_WHENEVER (MIPS_PERFCTRL_EXL | \
119 MIPS_PERFCTRL_K | \
120 MIPS_PERFCTRL_U | \
121 MIPS_PERFCTRL_S | \
122 MIPS_PERFCTRL_IE)
123
124#ifdef CONFIG_MIPS_MT_SMP
125#define M_PERFCTL_CONFIG_MASK 0x3fff801f
126#else
127#define M_PERFCTL_CONFIG_MASK 0x1f
128#endif
129
130
131#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
132static int cpu_has_mipsmt_pertccounters;
133
134static DEFINE_RWLOCK(pmuint_rwlock);
135
136#if defined(CONFIG_CPU_BMIPS5000)
137#define vpe_id() (cpu_has_mipsmt_pertccounters ? \
138 0 : (smp_processor_id() & MIPS_CPUID_TO_COUNTER_MASK))
139#else
140/*
141 * FIXME: For VSMP, vpe_id() is redefined for Perf-events, because
142 * cpu_data[cpuid].vpe_id reports 0 for _both_ CPUs.
143 */
144#define vpe_id() (cpu_has_mipsmt_pertccounters ? \
145 0 : smp_processor_id())
146#endif
147
148/* Copied from op_model_mipsxx.c */
149static unsigned int vpe_shift(void)
150{
151 if (num_possible_cpus() > 1)
152 return 1;
153
154 return 0;
155}
156
157static unsigned int counters_total_to_per_cpu(unsigned int counters)
158{
159 return counters >> vpe_shift();
160}
161
162#else /* !CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
163#define vpe_id() 0
164
165#endif /* CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
166
167static void resume_local_counters(void);
168static void pause_local_counters(void);
169static irqreturn_t mipsxx_pmu_handle_irq(int, void *);
170static int mipsxx_pmu_handle_shared_irq(void);
171
172static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx)
173{
174 if (vpe_id() == 1)
175 idx = (idx + 2) & 3;
176 return idx;
177}
178
179static u64 mipsxx_pmu_read_counter(unsigned int idx)
180{
181 idx = mipsxx_pmu_swizzle_perf_idx(idx);
182
183 switch (idx) {
184 case 0:
185 /*
186 * The counters are unsigned, we must cast to truncate
187 * off the high bits.
188 */
189 return (u32)read_c0_perfcntr0();
190 case 1:
191 return (u32)read_c0_perfcntr1();
192 case 2:
193 return (u32)read_c0_perfcntr2();
194 case 3:
195 return (u32)read_c0_perfcntr3();
196 default:
197 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
198 return 0;
199 }
200}
201
202static u64 mipsxx_pmu_read_counter_64(unsigned int idx)
203{
204 idx = mipsxx_pmu_swizzle_perf_idx(idx);
205
206 switch (idx) {
207 case 0:
208 return read_c0_perfcntr0_64();
209 case 1:
210 return read_c0_perfcntr1_64();
211 case 2:
212 return read_c0_perfcntr2_64();
213 case 3:
214 return read_c0_perfcntr3_64();
215 default:
216 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
217 return 0;
218 }
219}
220
221static void mipsxx_pmu_write_counter(unsigned int idx, u64 val)
222{
223 idx = mipsxx_pmu_swizzle_perf_idx(idx);
224
225 switch (idx) {
226 case 0:
227 write_c0_perfcntr0(val);
228 return;
229 case 1:
230 write_c0_perfcntr1(val);
231 return;
232 case 2:
233 write_c0_perfcntr2(val);
234 return;
235 case 3:
236 write_c0_perfcntr3(val);
237 return;
238 }
239}
240
241static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val)
242{
243 idx = mipsxx_pmu_swizzle_perf_idx(idx);
244
245 switch (idx) {
246 case 0:
247 write_c0_perfcntr0_64(val);
248 return;
249 case 1:
250 write_c0_perfcntr1_64(val);
251 return;
252 case 2:
253 write_c0_perfcntr2_64(val);
254 return;
255 case 3:
256 write_c0_perfcntr3_64(val);
257 return;
258 }
259}
260
261static unsigned int mipsxx_pmu_read_control(unsigned int idx)
262{
263 idx = mipsxx_pmu_swizzle_perf_idx(idx);
264
265 switch (idx) {
266 case 0:
267 return read_c0_perfctrl0();
268 case 1:
269 return read_c0_perfctrl1();
270 case 2:
271 return read_c0_perfctrl2();
272 case 3:
273 return read_c0_perfctrl3();
274 default:
275 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
276 return 0;
277 }
278}
279
280static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val)
281{
282 idx = mipsxx_pmu_swizzle_perf_idx(idx);
283
284 switch (idx) {
285 case 0:
286 write_c0_perfctrl0(val);
287 return;
288 case 1:
289 write_c0_perfctrl1(val);
290 return;
291 case 2:
292 write_c0_perfctrl2(val);
293 return;
294 case 3:
295 write_c0_perfctrl3(val);
296 return;
297 }
298}
299
300static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc,
301 struct hw_perf_event *hwc)
302{
303 int i;
304
305 /*
306 * We only need to care the counter mask. The range has been
307 * checked definitely.
308 */
309 unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff;
310
311 for (i = mipspmu.num_counters - 1; i >= 0; i--) {
312 /*
313 * Note that some MIPS perf events can be counted by both
314 * even and odd counters, wheresas many other are only by
315 * even _or_ odd counters. This introduces an issue that
316 * when the former kind of event takes the counter the
317 * latter kind of event wants to use, then the "counter
318 * allocation" for the latter event will fail. In fact if
319 * they can be dynamically swapped, they both feel happy.
320 * But here we leave this issue alone for now.
321 */
322 if (test_bit(i, &cntr_mask) &&
323 !test_and_set_bit(i, cpuc->used_mask))
324 return i;
325 }
326
327 return -EAGAIN;
328}
329
330static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx)
331{
332 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
333
334 WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
335
336 cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) |
337 (evt->config_base & M_PERFCTL_CONFIG_MASK) |
338 /* Make sure interrupt enabled. */
339 MIPS_PERFCTRL_IE;
340 if (IS_ENABLED(CONFIG_CPU_BMIPS5000))
341 /* enable the counter for the calling thread */
342 cpuc->saved_ctrl[idx] |=
343 (1 << (12 + vpe_id())) | BRCM_PERFCTRL_TC;
344
345 /*
346 * We do not actually let the counter run. Leave it until start().
347 */
348}
349
350static void mipsxx_pmu_disable_event(int idx)
351{
352 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
353 unsigned long flags;
354
355 WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
356
357 local_irq_save(flags);
358 cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) &
359 ~M_PERFCTL_COUNT_EVENT_WHENEVER;
360 mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]);
361 local_irq_restore(flags);
362}
363
364static int mipspmu_event_set_period(struct perf_event *event,
365 struct hw_perf_event *hwc,
366 int idx)
367{
368 u64 left = local64_read(&hwc->period_left);
369 u64 period = hwc->sample_period;
370 int ret = 0;
371
372 if (unlikely((left + period) & (1ULL << 63))) {
373 /* left underflowed by more than period. */
374 left = period;
375 local64_set(&hwc->period_left, left);
376 hwc->last_period = period;
377 ret = 1;
378 } else if (unlikely((left + period) <= period)) {
379 /* left underflowed by less than period. */
380 left += period;
381 local64_set(&hwc->period_left, left);
382 hwc->last_period = period;
383 ret = 1;
384 }
385
386 if (left > mipspmu.max_period) {
387 left = mipspmu.max_period;
388 local64_set(&hwc->period_left, left);
389 }
390
391 local64_set(&hwc->prev_count, mipspmu.overflow - left);
392
393 mipspmu.write_counter(idx, mipspmu.overflow - left);
394
395 perf_event_update_userpage(event);
396
397 return ret;
398}
399
400static void mipspmu_event_update(struct perf_event *event,
401 struct hw_perf_event *hwc,
402 int idx)
403{
404 u64 prev_raw_count, new_raw_count;
405 u64 delta;
406
407again:
408 prev_raw_count = local64_read(&hwc->prev_count);
409 new_raw_count = mipspmu.read_counter(idx);
410
411 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
412 new_raw_count) != prev_raw_count)
413 goto again;
414
415 delta = new_raw_count - prev_raw_count;
416
417 local64_add(delta, &event->count);
418 local64_sub(delta, &hwc->period_left);
419}
420
421static void mipspmu_start(struct perf_event *event, int flags)
422{
423 struct hw_perf_event *hwc = &event->hw;
424
425 if (flags & PERF_EF_RELOAD)
426 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
427
428 hwc->state = 0;
429
430 /* Set the period for the event. */
431 mipspmu_event_set_period(event, hwc, hwc->idx);
432
433 /* Enable the event. */
434 mipsxx_pmu_enable_event(hwc, hwc->idx);
435}
436
437static void mipspmu_stop(struct perf_event *event, int flags)
438{
439 struct hw_perf_event *hwc = &event->hw;
440
441 if (!(hwc->state & PERF_HES_STOPPED)) {
442 /* We are working on a local event. */
443 mipsxx_pmu_disable_event(hwc->idx);
444 barrier();
445 mipspmu_event_update(event, hwc, hwc->idx);
446 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
447 }
448}
449
450static int mipspmu_add(struct perf_event *event, int flags)
451{
452 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
453 struct hw_perf_event *hwc = &event->hw;
454 int idx;
455 int err = 0;
456
457 perf_pmu_disable(event->pmu);
458
459 /* To look for a free counter for this event. */
460 idx = mipsxx_pmu_alloc_counter(cpuc, hwc);
461 if (idx < 0) {
462 err = idx;
463 goto out;
464 }
465
466 /*
467 * If there is an event in the counter we are going to use then
468 * make sure it is disabled.
469 */
470 event->hw.idx = idx;
471 mipsxx_pmu_disable_event(idx);
472 cpuc->events[idx] = event;
473
474 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
475 if (flags & PERF_EF_START)
476 mipspmu_start(event, PERF_EF_RELOAD);
477
478 /* Propagate our changes to the userspace mapping. */
479 perf_event_update_userpage(event);
480
481out:
482 perf_pmu_enable(event->pmu);
483 return err;
484}
485
486static void mipspmu_del(struct perf_event *event, int flags)
487{
488 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
489 struct hw_perf_event *hwc = &event->hw;
490 int idx = hwc->idx;
491
492 WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
493
494 mipspmu_stop(event, PERF_EF_UPDATE);
495 cpuc->events[idx] = NULL;
496 clear_bit(idx, cpuc->used_mask);
497
498 perf_event_update_userpage(event);
499}
500
501static void mipspmu_read(struct perf_event *event)
502{
503 struct hw_perf_event *hwc = &event->hw;
504
505 /* Don't read disabled counters! */
506 if (hwc->idx < 0)
507 return;
508
509 mipspmu_event_update(event, hwc, hwc->idx);
510}
511
512static void mipspmu_enable(struct pmu *pmu)
513{
514#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
515 write_unlock(&pmuint_rwlock);
516#endif
517 resume_local_counters();
518}
519
520/*
521 * MIPS performance counters can be per-TC. The control registers can
522 * not be directly accessed across CPUs. Hence if we want to do global
523 * control, we need cross CPU calls. on_each_cpu() can help us, but we
524 * can not make sure this function is called with interrupts enabled. So
525 * here we pause local counters and then grab a rwlock and leave the
526 * counters on other CPUs alone. If any counter interrupt raises while
527 * we own the write lock, simply pause local counters on that CPU and
528 * spin in the handler. Also we know we won't be switched to another
529 * CPU after pausing local counters and before grabbing the lock.
530 */
531static void mipspmu_disable(struct pmu *pmu)
532{
533 pause_local_counters();
534#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
535 write_lock(&pmuint_rwlock);
536#endif
537}
538
539static atomic_t active_events = ATOMIC_INIT(0);
540static DEFINE_MUTEX(pmu_reserve_mutex);
541static int (*save_perf_irq)(void);
542
543static int mipspmu_get_irq(void)
544{
545 int err;
546
547 if (mipspmu.irq >= 0) {
548 /* Request my own irq handler. */
549 err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq,
550 IRQF_PERCPU | IRQF_NOBALANCING |
551 IRQF_NO_THREAD | IRQF_NO_SUSPEND |
552 IRQF_SHARED,
553 "mips_perf_pmu", &mipspmu);
554 if (err) {
555 pr_warn("Unable to request IRQ%d for MIPS performance counters!\n",
556 mipspmu.irq);
557 }
558 } else if (cp0_perfcount_irq < 0) {
559 /*
560 * We are sharing the irq number with the timer interrupt.
561 */
562 save_perf_irq = perf_irq;
563 perf_irq = mipsxx_pmu_handle_shared_irq;
564 err = 0;
565 } else {
566 pr_warn("The platform hasn't properly defined its interrupt controller\n");
567 err = -ENOENT;
568 }
569
570 return err;
571}
572
573static void mipspmu_free_irq(void)
574{
575 if (mipspmu.irq >= 0)
576 free_irq(mipspmu.irq, &mipspmu);
577 else if (cp0_perfcount_irq < 0)
578 perf_irq = save_perf_irq;
579}
580
581/*
582 * mipsxx/rm9000/loongson2 have different performance counters, they have
583 * specific low-level init routines.
584 */
585static void reset_counters(void *arg);
586static int __hw_perf_event_init(struct perf_event *event);
587
588static void hw_perf_event_destroy(struct perf_event *event)
589{
590 if (atomic_dec_and_mutex_lock(&active_events,
591 &pmu_reserve_mutex)) {
592 /*
593 * We must not call the destroy function with interrupts
594 * disabled.
595 */
596 on_each_cpu(reset_counters,
597 (void *)(long)mipspmu.num_counters, 1);
598 mipspmu_free_irq();
599 mutex_unlock(&pmu_reserve_mutex);
600 }
601}
602
603static int mipspmu_event_init(struct perf_event *event)
604{
605 int err = 0;
606
607 /* does not support taken branch sampling */
608 if (has_branch_stack(event))
609 return -EOPNOTSUPP;
610
611 switch (event->attr.type) {
612 case PERF_TYPE_RAW:
613 case PERF_TYPE_HARDWARE:
614 case PERF_TYPE_HW_CACHE:
615 break;
616
617 default:
618 return -ENOENT;
619 }
620
621 if (event->cpu >= 0 && !cpu_online(event->cpu))
622 return -ENODEV;
623
624 if (!atomic_inc_not_zero(&active_events)) {
625 mutex_lock(&pmu_reserve_mutex);
626 if (atomic_read(&active_events) == 0)
627 err = mipspmu_get_irq();
628
629 if (!err)
630 atomic_inc(&active_events);
631 mutex_unlock(&pmu_reserve_mutex);
632 }
633
634 if (err)
635 return err;
636
637 return __hw_perf_event_init(event);
638}
639
640static struct pmu pmu = {
641 .pmu_enable = mipspmu_enable,
642 .pmu_disable = mipspmu_disable,
643 .event_init = mipspmu_event_init,
644 .add = mipspmu_add,
645 .del = mipspmu_del,
646 .start = mipspmu_start,
647 .stop = mipspmu_stop,
648 .read = mipspmu_read,
649};
650
651static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev)
652{
653/*
654 * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for
655 * event_id.
656 */
657#ifdef CONFIG_MIPS_MT_SMP
658 return ((unsigned int)pev->range << 24) |
659 (pev->cntr_mask & 0xffff00) |
660 (pev->event_id & 0xff);
661#else
662 return (pev->cntr_mask & 0xffff00) |
663 (pev->event_id & 0xff);
664#endif
665}
666
667static const struct mips_perf_event *mipspmu_map_general_event(int idx)
668{
669
670 if ((*mipspmu.general_event_map)[idx].cntr_mask == 0)
671 return ERR_PTR(-EOPNOTSUPP);
672 return &(*mipspmu.general_event_map)[idx];
673}
674
675static const struct mips_perf_event *mipspmu_map_cache_event(u64 config)
676{
677 unsigned int cache_type, cache_op, cache_result;
678 const struct mips_perf_event *pev;
679
680 cache_type = (config >> 0) & 0xff;
681 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
682 return ERR_PTR(-EINVAL);
683
684 cache_op = (config >> 8) & 0xff;
685 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
686 return ERR_PTR(-EINVAL);
687
688 cache_result = (config >> 16) & 0xff;
689 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
690 return ERR_PTR(-EINVAL);
691
692 pev = &((*mipspmu.cache_event_map)
693 [cache_type]
694 [cache_op]
695 [cache_result]);
696
697 if (pev->cntr_mask == 0)
698 return ERR_PTR(-EOPNOTSUPP);
699
700 return pev;
701
702}
703
704static int validate_group(struct perf_event *event)
705{
706 struct perf_event *sibling, *leader = event->group_leader;
707 struct cpu_hw_events fake_cpuc;
708
709 memset(&fake_cpuc, 0, sizeof(fake_cpuc));
710
711 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0)
712 return -EINVAL;
713
714 for_each_sibling_event(sibling, leader) {
715 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0)
716 return -EINVAL;
717 }
718
719 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0)
720 return -EINVAL;
721
722 return 0;
723}
724
725/* This is needed by specific irq handlers in perf_event_*.c */
726static void handle_associated_event(struct cpu_hw_events *cpuc,
727 int idx, struct perf_sample_data *data,
728 struct pt_regs *regs)
729{
730 struct perf_event *event = cpuc->events[idx];
731 struct hw_perf_event *hwc = &event->hw;
732
733 mipspmu_event_update(event, hwc, idx);
734 data->period = event->hw.last_period;
735 if (!mipspmu_event_set_period(event, hwc, idx))
736 return;
737
738 if (perf_event_overflow(event, data, regs))
739 mipsxx_pmu_disable_event(idx);
740}
741
742
743static int __n_counters(void)
744{
745 if (!cpu_has_perf)
746 return 0;
747 if (!(read_c0_perfctrl0() & MIPS_PERFCTRL_M))
748 return 1;
749 if (!(read_c0_perfctrl1() & MIPS_PERFCTRL_M))
750 return 2;
751 if (!(read_c0_perfctrl2() & MIPS_PERFCTRL_M))
752 return 3;
753
754 return 4;
755}
756
757static int n_counters(void)
758{
759 int counters;
760
761 switch (current_cpu_type()) {
762 case CPU_R10000:
763 counters = 2;
764 break;
765
766 case CPU_R12000:
767 case CPU_R14000:
768 case CPU_R16000:
769 counters = 4;
770 break;
771
772 default:
773 counters = __n_counters();
774 }
775
776 return counters;
777}
778
779static void reset_counters(void *arg)
780{
781 int counters = (int)(long)arg;
782 switch (counters) {
783 case 4:
784 mipsxx_pmu_write_control(3, 0);
785 mipspmu.write_counter(3, 0);
786 case 3:
787 mipsxx_pmu_write_control(2, 0);
788 mipspmu.write_counter(2, 0);
789 case 2:
790 mipsxx_pmu_write_control(1, 0);
791 mipspmu.write_counter(1, 0);
792 case 1:
793 mipsxx_pmu_write_control(0, 0);
794 mipspmu.write_counter(0, 0);
795 }
796}
797
798/* 24K/34K/1004K/interAptiv/loongson1 cores share the same event map. */
799static const struct mips_perf_event mipsxxcore_event_map
800 [PERF_COUNT_HW_MAX] = {
801 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
802 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
803 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T },
804 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
805};
806
807/* 74K/proAptiv core has different branch event code. */
808static const struct mips_perf_event mipsxxcore_event_map2
809 [PERF_COUNT_HW_MAX] = {
810 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
811 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
812 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T },
813 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T },
814};
815
816static const struct mips_perf_event i6x00_event_map[PERF_COUNT_HW_MAX] = {
817 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD },
818 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD },
819 /* These only count dcache, not icache */
820 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x45, CNTR_EVEN | CNTR_ODD },
821 [PERF_COUNT_HW_CACHE_MISSES] = { 0x48, CNTR_EVEN | CNTR_ODD },
822 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x15, CNTR_EVEN | CNTR_ODD },
823 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x16, CNTR_EVEN | CNTR_ODD },
824};
825
826static const struct mips_perf_event loongson3_event_map[PERF_COUNT_HW_MAX] = {
827 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN },
828 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, CNTR_ODD },
829 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x01, CNTR_EVEN },
830 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x01, CNTR_ODD },
831};
832
833static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = {
834 [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
835 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL },
836 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL },
837 [PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL },
838 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL },
839 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL },
840 [PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL },
841};
842
843static const struct mips_perf_event bmips5000_event_map
844 [PERF_COUNT_HW_MAX] = {
845 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, T },
846 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
847 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
848};
849
850static const struct mips_perf_event xlp_event_map[PERF_COUNT_HW_MAX] = {
851 [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
852 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x18, CNTR_ALL }, /* PAPI_TOT_INS */
853 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */
854 [PERF_COUNT_HW_CACHE_MISSES] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */
855 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x1b, CNTR_ALL }, /* PAPI_BR_CN */
856 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x1c, CNTR_ALL }, /* PAPI_BR_MSP */
857};
858
859/* 24K/34K/1004K/interAptiv/loongson1 cores share the same cache event map. */
860static const struct mips_perf_event mipsxxcore_cache_map
861 [PERF_COUNT_HW_CACHE_MAX]
862 [PERF_COUNT_HW_CACHE_OP_MAX]
863 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
864[C(L1D)] = {
865 /*
866 * Like some other architectures (e.g. ARM), the performance
867 * counters don't differentiate between read and write
868 * accesses/misses, so this isn't strictly correct, but it's the
869 * best we can do. Writes and reads get combined.
870 */
871 [C(OP_READ)] = {
872 [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T },
873 [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T },
874 },
875 [C(OP_WRITE)] = {
876 [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T },
877 [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T },
878 },
879},
880[C(L1I)] = {
881 [C(OP_READ)] = {
882 [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T },
883 [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T },
884 },
885 [C(OP_WRITE)] = {
886 [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T },
887 [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T },
888 },
889 [C(OP_PREFETCH)] = {
890 [C(RESULT_ACCESS)] = { 0x14, CNTR_EVEN, T },
891 /*
892 * Note that MIPS has only "hit" events countable for
893 * the prefetch operation.
894 */
895 },
896},
897[C(LL)] = {
898 [C(OP_READ)] = {
899 [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P },
900 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P },
901 },
902 [C(OP_WRITE)] = {
903 [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P },
904 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P },
905 },
906},
907[C(DTLB)] = {
908 [C(OP_READ)] = {
909 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
910 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
911 },
912 [C(OP_WRITE)] = {
913 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
914 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
915 },
916},
917[C(ITLB)] = {
918 [C(OP_READ)] = {
919 [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T },
920 [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T },
921 },
922 [C(OP_WRITE)] = {
923 [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T },
924 [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T },
925 },
926},
927[C(BPU)] = {
928 /* Using the same code for *HW_BRANCH* */
929 [C(OP_READ)] = {
930 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T },
931 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
932 },
933 [C(OP_WRITE)] = {
934 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T },
935 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
936 },
937},
938};
939
940/* 74K/proAptiv core has completely different cache event map. */
941static const struct mips_perf_event mipsxxcore_cache_map2
942 [PERF_COUNT_HW_CACHE_MAX]
943 [PERF_COUNT_HW_CACHE_OP_MAX]
944 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
945[C(L1D)] = {
946 /*
947 * Like some other architectures (e.g. ARM), the performance
948 * counters don't differentiate between read and write
949 * accesses/misses, so this isn't strictly correct, but it's the
950 * best we can do. Writes and reads get combined.
951 */
952 [C(OP_READ)] = {
953 [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T },
954 [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T },
955 },
956 [C(OP_WRITE)] = {
957 [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T },
958 [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T },
959 },
960},
961[C(L1I)] = {
962 [C(OP_READ)] = {
963 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
964 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
965 },
966 [C(OP_WRITE)] = {
967 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
968 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
969 },
970 [C(OP_PREFETCH)] = {
971 [C(RESULT_ACCESS)] = { 0x34, CNTR_EVEN, T },
972 /*
973 * Note that MIPS has only "hit" events countable for
974 * the prefetch operation.
975 */
976 },
977},
978[C(LL)] = {
979 [C(OP_READ)] = {
980 [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P },
981 [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN, P },
982 },
983 [C(OP_WRITE)] = {
984 [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P },
985 [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN, P },
986 },
987},
988/*
989 * 74K core does not have specific DTLB events. proAptiv core has
990 * "speculative" DTLB events which are numbered 0x63 (even/odd) and
991 * not included here. One can use raw events if really needed.
992 */
993[C(ITLB)] = {
994 [C(OP_READ)] = {
995 [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T },
996 [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T },
997 },
998 [C(OP_WRITE)] = {
999 [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T },
1000 [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T },
1001 },
1002},
1003[C(BPU)] = {
1004 /* Using the same code for *HW_BRANCH* */
1005 [C(OP_READ)] = {
1006 [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T },
1007 [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T },
1008 },
1009 [C(OP_WRITE)] = {
1010 [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T },
1011 [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T },
1012 },
1013},
1014};
1015
1016static const struct mips_perf_event i6x00_cache_map
1017 [PERF_COUNT_HW_CACHE_MAX]
1018 [PERF_COUNT_HW_CACHE_OP_MAX]
1019 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1020[C(L1D)] = {
1021 [C(OP_READ)] = {
1022 [C(RESULT_ACCESS)] = { 0x46, CNTR_EVEN | CNTR_ODD },
1023 [C(RESULT_MISS)] = { 0x49, CNTR_EVEN | CNTR_ODD },
1024 },
1025 [C(OP_WRITE)] = {
1026 [C(RESULT_ACCESS)] = { 0x47, CNTR_EVEN | CNTR_ODD },
1027 [C(RESULT_MISS)] = { 0x4a, CNTR_EVEN | CNTR_ODD },
1028 },
1029},
1030[C(L1I)] = {
1031 [C(OP_READ)] = {
1032 [C(RESULT_ACCESS)] = { 0x84, CNTR_EVEN | CNTR_ODD },
1033 [C(RESULT_MISS)] = { 0x85, CNTR_EVEN | CNTR_ODD },
1034 },
1035},
1036[C(DTLB)] = {
1037 /* Can't distinguish read & write */
1038 [C(OP_READ)] = {
1039 [C(RESULT_ACCESS)] = { 0x40, CNTR_EVEN | CNTR_ODD },
1040 [C(RESULT_MISS)] = { 0x41, CNTR_EVEN | CNTR_ODD },
1041 },
1042 [C(OP_WRITE)] = {
1043 [C(RESULT_ACCESS)] = { 0x40, CNTR_EVEN | CNTR_ODD },
1044 [C(RESULT_MISS)] = { 0x41, CNTR_EVEN | CNTR_ODD },
1045 },
1046},
1047[C(BPU)] = {
1048 /* Conditional branches / mispredicted */
1049 [C(OP_READ)] = {
1050 [C(RESULT_ACCESS)] = { 0x15, CNTR_EVEN | CNTR_ODD },
1051 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN | CNTR_ODD },
1052 },
1053},
1054};
1055
1056static const struct mips_perf_event loongson3_cache_map
1057 [PERF_COUNT_HW_CACHE_MAX]
1058 [PERF_COUNT_HW_CACHE_OP_MAX]
1059 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1060[C(L1D)] = {
1061 /*
1062 * Like some other architectures (e.g. ARM), the performance
1063 * counters don't differentiate between read and write
1064 * accesses/misses, so this isn't strictly correct, but it's the
1065 * best we can do. Writes and reads get combined.
1066 */
1067 [C(OP_READ)] = {
1068 [C(RESULT_MISS)] = { 0x04, CNTR_ODD },
1069 },
1070 [C(OP_WRITE)] = {
1071 [C(RESULT_MISS)] = { 0x04, CNTR_ODD },
1072 },
1073},
1074[C(L1I)] = {
1075 [C(OP_READ)] = {
1076 [C(RESULT_MISS)] = { 0x04, CNTR_EVEN },
1077 },
1078 [C(OP_WRITE)] = {
1079 [C(RESULT_MISS)] = { 0x04, CNTR_EVEN },
1080 },
1081},
1082[C(DTLB)] = {
1083 [C(OP_READ)] = {
1084 [C(RESULT_MISS)] = { 0x09, CNTR_ODD },
1085 },
1086 [C(OP_WRITE)] = {
1087 [C(RESULT_MISS)] = { 0x09, CNTR_ODD },
1088 },
1089},
1090[C(ITLB)] = {
1091 [C(OP_READ)] = {
1092 [C(RESULT_MISS)] = { 0x0c, CNTR_ODD },
1093 },
1094 [C(OP_WRITE)] = {
1095 [C(RESULT_MISS)] = { 0x0c, CNTR_ODD },
1096 },
1097},
1098[C(BPU)] = {
1099 /* Using the same code for *HW_BRANCH* */
1100 [C(OP_READ)] = {
1101 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN },
1102 [C(RESULT_MISS)] = { 0x02, CNTR_ODD },
1103 },
1104 [C(OP_WRITE)] = {
1105 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN },
1106 [C(RESULT_MISS)] = { 0x02, CNTR_ODD },
1107 },
1108},
1109};
1110
1111/* BMIPS5000 */
1112static const struct mips_perf_event bmips5000_cache_map
1113 [PERF_COUNT_HW_CACHE_MAX]
1114 [PERF_COUNT_HW_CACHE_OP_MAX]
1115 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1116[C(L1D)] = {
1117 /*
1118 * Like some other architectures (e.g. ARM), the performance
1119 * counters don't differentiate between read and write
1120 * accesses/misses, so this isn't strictly correct, but it's the
1121 * best we can do. Writes and reads get combined.
1122 */
1123 [C(OP_READ)] = {
1124 [C(RESULT_ACCESS)] = { 12, CNTR_EVEN, T },
1125 [C(RESULT_MISS)] = { 12, CNTR_ODD, T },
1126 },
1127 [C(OP_WRITE)] = {
1128 [C(RESULT_ACCESS)] = { 12, CNTR_EVEN, T },
1129 [C(RESULT_MISS)] = { 12, CNTR_ODD, T },
1130 },
1131},
1132[C(L1I)] = {
1133 [C(OP_READ)] = {
1134 [C(RESULT_ACCESS)] = { 10, CNTR_EVEN, T },
1135 [C(RESULT_MISS)] = { 10, CNTR_ODD, T },
1136 },
1137 [C(OP_WRITE)] = {
1138 [C(RESULT_ACCESS)] = { 10, CNTR_EVEN, T },
1139 [C(RESULT_MISS)] = { 10, CNTR_ODD, T },
1140 },
1141 [C(OP_PREFETCH)] = {
1142 [C(RESULT_ACCESS)] = { 23, CNTR_EVEN, T },
1143 /*
1144 * Note that MIPS has only "hit" events countable for
1145 * the prefetch operation.
1146 */
1147 },
1148},
1149[C(LL)] = {
1150 [C(OP_READ)] = {
1151 [C(RESULT_ACCESS)] = { 28, CNTR_EVEN, P },
1152 [C(RESULT_MISS)] = { 28, CNTR_ODD, P },
1153 },
1154 [C(OP_WRITE)] = {
1155 [C(RESULT_ACCESS)] = { 28, CNTR_EVEN, P },
1156 [C(RESULT_MISS)] = { 28, CNTR_ODD, P },
1157 },
1158},
1159[C(BPU)] = {
1160 /* Using the same code for *HW_BRANCH* */
1161 [C(OP_READ)] = {
1162 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
1163 },
1164 [C(OP_WRITE)] = {
1165 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
1166 },
1167},
1168};
1169
1170
1171static const struct mips_perf_event octeon_cache_map
1172 [PERF_COUNT_HW_CACHE_MAX]
1173 [PERF_COUNT_HW_CACHE_OP_MAX]
1174 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1175[C(L1D)] = {
1176 [C(OP_READ)] = {
1177 [C(RESULT_ACCESS)] = { 0x2b, CNTR_ALL },
1178 [C(RESULT_MISS)] = { 0x2e, CNTR_ALL },
1179 },
1180 [C(OP_WRITE)] = {
1181 [C(RESULT_ACCESS)] = { 0x30, CNTR_ALL },
1182 },
1183},
1184[C(L1I)] = {
1185 [C(OP_READ)] = {
1186 [C(RESULT_ACCESS)] = { 0x18, CNTR_ALL },
1187 },
1188 [C(OP_PREFETCH)] = {
1189 [C(RESULT_ACCESS)] = { 0x19, CNTR_ALL },
1190 },
1191},
1192[C(DTLB)] = {
1193 /*
1194 * Only general DTLB misses are counted use the same event for
1195 * read and write.
1196 */
1197 [C(OP_READ)] = {
1198 [C(RESULT_MISS)] = { 0x35, CNTR_ALL },
1199 },
1200 [C(OP_WRITE)] = {
1201 [C(RESULT_MISS)] = { 0x35, CNTR_ALL },
1202 },
1203},
1204[C(ITLB)] = {
1205 [C(OP_READ)] = {
1206 [C(RESULT_MISS)] = { 0x37, CNTR_ALL },
1207 },
1208},
1209};
1210
1211static const struct mips_perf_event xlp_cache_map
1212 [PERF_COUNT_HW_CACHE_MAX]
1213 [PERF_COUNT_HW_CACHE_OP_MAX]
1214 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1215[C(L1D)] = {
1216 [C(OP_READ)] = {
1217 [C(RESULT_ACCESS)] = { 0x31, CNTR_ALL }, /* PAPI_L1_DCR */
1218 [C(RESULT_MISS)] = { 0x30, CNTR_ALL }, /* PAPI_L1_LDM */
1219 },
1220 [C(OP_WRITE)] = {
1221 [C(RESULT_ACCESS)] = { 0x2f, CNTR_ALL }, /* PAPI_L1_DCW */
1222 [C(RESULT_MISS)] = { 0x2e, CNTR_ALL }, /* PAPI_L1_STM */
1223 },
1224},
1225[C(L1I)] = {
1226 [C(OP_READ)] = {
1227 [C(RESULT_ACCESS)] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */
1228 [C(RESULT_MISS)] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */
1229 },
1230},
1231[C(LL)] = {
1232 [C(OP_READ)] = {
1233 [C(RESULT_ACCESS)] = { 0x35, CNTR_ALL }, /* PAPI_L2_DCR */
1234 [C(RESULT_MISS)] = { 0x37, CNTR_ALL }, /* PAPI_L2_LDM */
1235 },
1236 [C(OP_WRITE)] = {
1237 [C(RESULT_ACCESS)] = { 0x34, CNTR_ALL }, /* PAPI_L2_DCA */
1238 [C(RESULT_MISS)] = { 0x36, CNTR_ALL }, /* PAPI_L2_DCM */
1239 },
1240},
1241[C(DTLB)] = {
1242 /*
1243 * Only general DTLB misses are counted use the same event for
1244 * read and write.
1245 */
1246 [C(OP_READ)] = {
1247 [C(RESULT_MISS)] = { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */
1248 },
1249 [C(OP_WRITE)] = {
1250 [C(RESULT_MISS)] = { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */
1251 },
1252},
1253[C(ITLB)] = {
1254 [C(OP_READ)] = {
1255 [C(RESULT_MISS)] = { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */
1256 },
1257 [C(OP_WRITE)] = {
1258 [C(RESULT_MISS)] = { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */
1259 },
1260},
1261[C(BPU)] = {
1262 [C(OP_READ)] = {
1263 [C(RESULT_MISS)] = { 0x25, CNTR_ALL },
1264 },
1265},
1266};
1267
1268#ifdef CONFIG_MIPS_MT_SMP
1269static void check_and_calc_range(struct perf_event *event,
1270 const struct mips_perf_event *pev)
1271{
1272 struct hw_perf_event *hwc = &event->hw;
1273
1274 if (event->cpu >= 0) {
1275 if (pev->range > V) {
1276 /*
1277 * The user selected an event that is processor
1278 * wide, while expecting it to be VPE wide.
1279 */
1280 hwc->config_base |= M_TC_EN_ALL;
1281 } else {
1282 /*
1283 * FIXME: cpu_data[event->cpu].vpe_id reports 0
1284 * for both CPUs.
1285 */
1286 hwc->config_base |= M_PERFCTL_VPEID(event->cpu);
1287 hwc->config_base |= M_TC_EN_VPE;
1288 }
1289 } else
1290 hwc->config_base |= M_TC_EN_ALL;
1291}
1292#else
1293static void check_and_calc_range(struct perf_event *event,
1294 const struct mips_perf_event *pev)
1295{
1296}
1297#endif
1298
1299static int __hw_perf_event_init(struct perf_event *event)
1300{
1301 struct perf_event_attr *attr = &event->attr;
1302 struct hw_perf_event *hwc = &event->hw;
1303 const struct mips_perf_event *pev;
1304 int err;
1305
1306 /* Returning MIPS event descriptor for generic perf event. */
1307 if (PERF_TYPE_HARDWARE == event->attr.type) {
1308 if (event->attr.config >= PERF_COUNT_HW_MAX)
1309 return -EINVAL;
1310 pev = mipspmu_map_general_event(event->attr.config);
1311 } else if (PERF_TYPE_HW_CACHE == event->attr.type) {
1312 pev = mipspmu_map_cache_event(event->attr.config);
1313 } else if (PERF_TYPE_RAW == event->attr.type) {
1314 /* We are working on the global raw event. */
1315 mutex_lock(&raw_event_mutex);
1316 pev = mipspmu.map_raw_event(event->attr.config);
1317 } else {
1318 /* The event type is not (yet) supported. */
1319 return -EOPNOTSUPP;
1320 }
1321
1322 if (IS_ERR(pev)) {
1323 if (PERF_TYPE_RAW == event->attr.type)
1324 mutex_unlock(&raw_event_mutex);
1325 return PTR_ERR(pev);
1326 }
1327
1328 /*
1329 * We allow max flexibility on how each individual counter shared
1330 * by the single CPU operates (the mode exclusion and the range).
1331 */
1332 hwc->config_base = MIPS_PERFCTRL_IE;
1333
1334 /* Calculate range bits and validate it. */
1335 if (num_possible_cpus() > 1)
1336 check_and_calc_range(event, pev);
1337
1338 hwc->event_base = mipspmu_perf_event_encode(pev);
1339 if (PERF_TYPE_RAW == event->attr.type)
1340 mutex_unlock(&raw_event_mutex);
1341
1342 if (!attr->exclude_user)
1343 hwc->config_base |= MIPS_PERFCTRL_U;
1344 if (!attr->exclude_kernel) {
1345 hwc->config_base |= MIPS_PERFCTRL_K;
1346 /* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */
1347 hwc->config_base |= MIPS_PERFCTRL_EXL;
1348 }
1349 if (!attr->exclude_hv)
1350 hwc->config_base |= MIPS_PERFCTRL_S;
1351
1352 hwc->config_base &= M_PERFCTL_CONFIG_MASK;
1353 /*
1354 * The event can belong to another cpu. We do not assign a local
1355 * counter for it for now.
1356 */
1357 hwc->idx = -1;
1358 hwc->config = 0;
1359
1360 if (!hwc->sample_period) {
1361 hwc->sample_period = mipspmu.max_period;
1362 hwc->last_period = hwc->sample_period;
1363 local64_set(&hwc->period_left, hwc->sample_period);
1364 }
1365
1366 err = 0;
1367 if (event->group_leader != event)
1368 err = validate_group(event);
1369
1370 event->destroy = hw_perf_event_destroy;
1371
1372 if (err)
1373 event->destroy(event);
1374
1375 return err;
1376}
1377
1378static void pause_local_counters(void)
1379{
1380 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1381 int ctr = mipspmu.num_counters;
1382 unsigned long flags;
1383
1384 local_irq_save(flags);
1385 do {
1386 ctr--;
1387 cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr);
1388 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] &
1389 ~M_PERFCTL_COUNT_EVENT_WHENEVER);
1390 } while (ctr > 0);
1391 local_irq_restore(flags);
1392}
1393
1394static void resume_local_counters(void)
1395{
1396 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1397 int ctr = mipspmu.num_counters;
1398
1399 do {
1400 ctr--;
1401 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]);
1402 } while (ctr > 0);
1403}
1404
1405static int mipsxx_pmu_handle_shared_irq(void)
1406{
1407 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1408 struct perf_sample_data data;
1409 unsigned int counters = mipspmu.num_counters;
1410 u64 counter;
1411 int handled = IRQ_NONE;
1412 struct pt_regs *regs;
1413
1414 if (cpu_has_perf_cntr_intr_bit && !(read_c0_cause() & CAUSEF_PCI))
1415 return handled;
1416 /*
1417 * First we pause the local counters, so that when we are locked
1418 * here, the counters are all paused. When it gets locked due to
1419 * perf_disable(), the timer interrupt handler will be delayed.
1420 *
1421 * See also mipsxx_pmu_start().
1422 */
1423 pause_local_counters();
1424#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1425 read_lock(&pmuint_rwlock);
1426#endif
1427
1428 regs = get_irq_regs();
1429
1430 perf_sample_data_init(&data, 0, 0);
1431
1432 switch (counters) {
1433#define HANDLE_COUNTER(n) \
1434 case n + 1: \
1435 if (test_bit(n, cpuc->used_mask)) { \
1436 counter = mipspmu.read_counter(n); \
1437 if (counter & mipspmu.overflow) { \
1438 handle_associated_event(cpuc, n, &data, regs); \
1439 handled = IRQ_HANDLED; \
1440 } \
1441 }
1442 HANDLE_COUNTER(3)
1443 HANDLE_COUNTER(2)
1444 HANDLE_COUNTER(1)
1445 HANDLE_COUNTER(0)
1446 }
1447
1448#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1449 read_unlock(&pmuint_rwlock);
1450#endif
1451 resume_local_counters();
1452
1453 /*
1454 * Do all the work for the pending perf events. We can do this
1455 * in here because the performance counter interrupt is a regular
1456 * interrupt, not NMI.
1457 */
1458 if (handled == IRQ_HANDLED)
1459 irq_work_run();
1460
1461 return handled;
1462}
1463
1464static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev)
1465{
1466 return mipsxx_pmu_handle_shared_irq();
1467}
1468
1469/* 24K */
1470#define IS_BOTH_COUNTERS_24K_EVENT(b) \
1471 ((b) == 0 || (b) == 1 || (b) == 11)
1472
1473/* 34K */
1474#define IS_BOTH_COUNTERS_34K_EVENT(b) \
1475 ((b) == 0 || (b) == 1 || (b) == 11)
1476#ifdef CONFIG_MIPS_MT_SMP
1477#define IS_RANGE_P_34K_EVENT(r, b) \
1478 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \
1479 (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 || \
1480 (r) == 176 || ((b) >= 50 && (b) <= 55) || \
1481 ((b) >= 64 && (b) <= 67))
1482#define IS_RANGE_V_34K_EVENT(r) ((r) == 47)
1483#endif
1484
1485/* 74K */
1486#define IS_BOTH_COUNTERS_74K_EVENT(b) \
1487 ((b) == 0 || (b) == 1)
1488
1489/* proAptiv */
1490#define IS_BOTH_COUNTERS_PROAPTIV_EVENT(b) \
1491 ((b) == 0 || (b) == 1)
1492/* P5600 */
1493#define IS_BOTH_COUNTERS_P5600_EVENT(b) \
1494 ((b) == 0 || (b) == 1)
1495
1496/* 1004K */
1497#define IS_BOTH_COUNTERS_1004K_EVENT(b) \
1498 ((b) == 0 || (b) == 1 || (b) == 11)
1499#ifdef CONFIG_MIPS_MT_SMP
1500#define IS_RANGE_P_1004K_EVENT(r, b) \
1501 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \
1502 (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 || \
1503 (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) || \
1504 (r) == 188 || (b) == 61 || (b) == 62 || \
1505 ((b) >= 64 && (b) <= 67))
1506#define IS_RANGE_V_1004K_EVENT(r) ((r) == 47)
1507#endif
1508
1509/* interAptiv */
1510#define IS_BOTH_COUNTERS_INTERAPTIV_EVENT(b) \
1511 ((b) == 0 || (b) == 1 || (b) == 11)
1512#ifdef CONFIG_MIPS_MT_SMP
1513/* The P/V/T info is not provided for "(b) == 38" in SUM, assume P. */
1514#define IS_RANGE_P_INTERAPTIV_EVENT(r, b) \
1515 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \
1516 (b) == 25 || (b) == 36 || (b) == 38 || (b) == 39 || \
1517 (r) == 44 || (r) == 174 || (r) == 176 || ((b) >= 50 && \
1518 (b) <= 59) || (r) == 188 || (b) == 61 || (b) == 62 || \
1519 ((b) >= 64 && (b) <= 67))
1520#define IS_RANGE_V_INTERAPTIV_EVENT(r) ((r) == 47 || (r) == 175)
1521#endif
1522
1523/* BMIPS5000 */
1524#define IS_BOTH_COUNTERS_BMIPS5000_EVENT(b) \
1525 ((b) == 0 || (b) == 1)
1526
1527
1528/*
1529 * For most cores the user can use 0-255 raw events, where 0-127 for the events
1530 * of even counters, and 128-255 for odd counters. Note that bit 7 is used to
1531 * indicate the even/odd bank selector. So, for example, when user wants to take
1532 * the Event Num of 15 for odd counters (by referring to the user manual), then
1533 * 128 needs to be added to 15 as the input for the event config, i.e., 143 (0x8F)
1534 * to be used.
1535 *
1536 * Some newer cores have even more events, in which case the user can use raw
1537 * events 0-511, where 0-255 are for the events of even counters, and 256-511
1538 * are for odd counters, so bit 8 is used to indicate the even/odd bank selector.
1539 */
1540static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config)
1541{
1542 /* currently most cores have 7-bit event numbers */
1543 unsigned int raw_id = config & 0xff;
1544 unsigned int base_id = raw_id & 0x7f;
1545
1546 switch (current_cpu_type()) {
1547 case CPU_24K:
1548 if (IS_BOTH_COUNTERS_24K_EVENT(base_id))
1549 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1550 else
1551 raw_event.cntr_mask =
1552 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1553#ifdef CONFIG_MIPS_MT_SMP
1554 /*
1555 * This is actually doing nothing. Non-multithreading
1556 * CPUs will not check and calculate the range.
1557 */
1558 raw_event.range = P;
1559#endif
1560 break;
1561 case CPU_34K:
1562 if (IS_BOTH_COUNTERS_34K_EVENT(base_id))
1563 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1564 else
1565 raw_event.cntr_mask =
1566 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1567#ifdef CONFIG_MIPS_MT_SMP
1568 if (IS_RANGE_P_34K_EVENT(raw_id, base_id))
1569 raw_event.range = P;
1570 else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id)))
1571 raw_event.range = V;
1572 else
1573 raw_event.range = T;
1574#endif
1575 break;
1576 case CPU_74K:
1577 case CPU_1074K:
1578 if (IS_BOTH_COUNTERS_74K_EVENT(base_id))
1579 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1580 else
1581 raw_event.cntr_mask =
1582 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1583#ifdef CONFIG_MIPS_MT_SMP
1584 raw_event.range = P;
1585#endif
1586 break;
1587 case CPU_PROAPTIV:
1588 if (IS_BOTH_COUNTERS_PROAPTIV_EVENT(base_id))
1589 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1590 else
1591 raw_event.cntr_mask =
1592 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1593#ifdef CONFIG_MIPS_MT_SMP
1594 raw_event.range = P;
1595#endif
1596 break;
1597 case CPU_P5600:
1598 case CPU_P6600:
1599 /* 8-bit event numbers */
1600 raw_id = config & 0x1ff;
1601 base_id = raw_id & 0xff;
1602 if (IS_BOTH_COUNTERS_P5600_EVENT(base_id))
1603 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1604 else
1605 raw_event.cntr_mask =
1606 raw_id > 255 ? CNTR_ODD : CNTR_EVEN;
1607#ifdef CONFIG_MIPS_MT_SMP
1608 raw_event.range = P;
1609#endif
1610 break;
1611 case CPU_I6400:
1612 case CPU_I6500:
1613 /* 8-bit event numbers */
1614 base_id = config & 0xff;
1615 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1616 break;
1617 case CPU_1004K:
1618 if (IS_BOTH_COUNTERS_1004K_EVENT(base_id))
1619 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1620 else
1621 raw_event.cntr_mask =
1622 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1623#ifdef CONFIG_MIPS_MT_SMP
1624 if (IS_RANGE_P_1004K_EVENT(raw_id, base_id))
1625 raw_event.range = P;
1626 else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id)))
1627 raw_event.range = V;
1628 else
1629 raw_event.range = T;
1630#endif
1631 break;
1632 case CPU_INTERAPTIV:
1633 if (IS_BOTH_COUNTERS_INTERAPTIV_EVENT(base_id))
1634 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1635 else
1636 raw_event.cntr_mask =
1637 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1638#ifdef CONFIG_MIPS_MT_SMP
1639 if (IS_RANGE_P_INTERAPTIV_EVENT(raw_id, base_id))
1640 raw_event.range = P;
1641 else if (unlikely(IS_RANGE_V_INTERAPTIV_EVENT(raw_id)))
1642 raw_event.range = V;
1643 else
1644 raw_event.range = T;
1645#endif
1646 break;
1647 case CPU_BMIPS5000:
1648 if (IS_BOTH_COUNTERS_BMIPS5000_EVENT(base_id))
1649 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1650 else
1651 raw_event.cntr_mask =
1652 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1653 break;
1654 case CPU_LOONGSON3:
1655 raw_event.cntr_mask = raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1656 break;
1657 }
1658
1659 raw_event.event_id = base_id;
1660
1661 return &raw_event;
1662}
1663
1664static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config)
1665{
1666 unsigned int raw_id = config & 0xff;
1667 unsigned int base_id = raw_id & 0x7f;
1668
1669
1670 raw_event.cntr_mask = CNTR_ALL;
1671 raw_event.event_id = base_id;
1672
1673 if (current_cpu_type() == CPU_CAVIUM_OCTEON2) {
1674 if (base_id > 0x42)
1675 return ERR_PTR(-EOPNOTSUPP);
1676 } else {
1677 if (base_id > 0x3a)
1678 return ERR_PTR(-EOPNOTSUPP);
1679 }
1680
1681 switch (base_id) {
1682 case 0x00:
1683 case 0x0f:
1684 case 0x1e:
1685 case 0x1f:
1686 case 0x2f:
1687 case 0x34:
1688 case 0x3b ... 0x3f:
1689 return ERR_PTR(-EOPNOTSUPP);
1690 default:
1691 break;
1692 }
1693
1694 return &raw_event;
1695}
1696
1697static const struct mips_perf_event *xlp_pmu_map_raw_event(u64 config)
1698{
1699 unsigned int raw_id = config & 0xff;
1700
1701 /* Only 1-63 are defined */
1702 if ((raw_id < 0x01) || (raw_id > 0x3f))
1703 return ERR_PTR(-EOPNOTSUPP);
1704
1705 raw_event.cntr_mask = CNTR_ALL;
1706 raw_event.event_id = raw_id;
1707
1708 return &raw_event;
1709}
1710
1711static int __init
1712init_hw_perf_events(void)
1713{
1714 int counters, irq;
1715 int counter_bits;
1716
1717 pr_info("Performance counters: ");
1718
1719 counters = n_counters();
1720 if (counters == 0) {
1721 pr_cont("No available PMU.\n");
1722 return -ENODEV;
1723 }
1724
1725#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1726 cpu_has_mipsmt_pertccounters = read_c0_config7() & (1<<19);
1727 if (!cpu_has_mipsmt_pertccounters)
1728 counters = counters_total_to_per_cpu(counters);
1729#endif
1730
1731 if (get_c0_perfcount_int)
1732 irq = get_c0_perfcount_int();
1733 else if (cp0_perfcount_irq >= 0)
1734 irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq;
1735 else
1736 irq = -1;
1737
1738 mipspmu.map_raw_event = mipsxx_pmu_map_raw_event;
1739
1740 switch (current_cpu_type()) {
1741 case CPU_24K:
1742 mipspmu.name = "mips/24K";
1743 mipspmu.general_event_map = &mipsxxcore_event_map;
1744 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1745 break;
1746 case CPU_34K:
1747 mipspmu.name = "mips/34K";
1748 mipspmu.general_event_map = &mipsxxcore_event_map;
1749 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1750 break;
1751 case CPU_74K:
1752 mipspmu.name = "mips/74K";
1753 mipspmu.general_event_map = &mipsxxcore_event_map2;
1754 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1755 break;
1756 case CPU_PROAPTIV:
1757 mipspmu.name = "mips/proAptiv";
1758 mipspmu.general_event_map = &mipsxxcore_event_map2;
1759 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1760 break;
1761 case CPU_P5600:
1762 mipspmu.name = "mips/P5600";
1763 mipspmu.general_event_map = &mipsxxcore_event_map2;
1764 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1765 break;
1766 case CPU_P6600:
1767 mipspmu.name = "mips/P6600";
1768 mipspmu.general_event_map = &mipsxxcore_event_map2;
1769 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1770 break;
1771 case CPU_I6400:
1772 mipspmu.name = "mips/I6400";
1773 mipspmu.general_event_map = &i6x00_event_map;
1774 mipspmu.cache_event_map = &i6x00_cache_map;
1775 break;
1776 case CPU_I6500:
1777 mipspmu.name = "mips/I6500";
1778 mipspmu.general_event_map = &i6x00_event_map;
1779 mipspmu.cache_event_map = &i6x00_cache_map;
1780 break;
1781 case CPU_1004K:
1782 mipspmu.name = "mips/1004K";
1783 mipspmu.general_event_map = &mipsxxcore_event_map;
1784 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1785 break;
1786 case CPU_1074K:
1787 mipspmu.name = "mips/1074K";
1788 mipspmu.general_event_map = &mipsxxcore_event_map;
1789 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1790 break;
1791 case CPU_INTERAPTIV:
1792 mipspmu.name = "mips/interAptiv";
1793 mipspmu.general_event_map = &mipsxxcore_event_map;
1794 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1795 break;
1796 case CPU_LOONGSON1:
1797 mipspmu.name = "mips/loongson1";
1798 mipspmu.general_event_map = &mipsxxcore_event_map;
1799 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1800 break;
1801 case CPU_LOONGSON3:
1802 mipspmu.name = "mips/loongson3";
1803 mipspmu.general_event_map = &loongson3_event_map;
1804 mipspmu.cache_event_map = &loongson3_cache_map;
1805 break;
1806 case CPU_CAVIUM_OCTEON:
1807 case CPU_CAVIUM_OCTEON_PLUS:
1808 case CPU_CAVIUM_OCTEON2:
1809 mipspmu.name = "octeon";
1810 mipspmu.general_event_map = &octeon_event_map;
1811 mipspmu.cache_event_map = &octeon_cache_map;
1812 mipspmu.map_raw_event = octeon_pmu_map_raw_event;
1813 break;
1814 case CPU_BMIPS5000:
1815 mipspmu.name = "BMIPS5000";
1816 mipspmu.general_event_map = &bmips5000_event_map;
1817 mipspmu.cache_event_map = &bmips5000_cache_map;
1818 break;
1819 case CPU_XLP:
1820 mipspmu.name = "xlp";
1821 mipspmu.general_event_map = &xlp_event_map;
1822 mipspmu.cache_event_map = &xlp_cache_map;
1823 mipspmu.map_raw_event = xlp_pmu_map_raw_event;
1824 break;
1825 default:
1826 pr_cont("Either hardware does not support performance "
1827 "counters, or not yet implemented.\n");
1828 return -ENODEV;
1829 }
1830
1831 mipspmu.num_counters = counters;
1832 mipspmu.irq = irq;
1833
1834 if (read_c0_perfctrl0() & MIPS_PERFCTRL_W) {
1835 mipspmu.max_period = (1ULL << 63) - 1;
1836 mipspmu.valid_count = (1ULL << 63) - 1;
1837 mipspmu.overflow = 1ULL << 63;
1838 mipspmu.read_counter = mipsxx_pmu_read_counter_64;
1839 mipspmu.write_counter = mipsxx_pmu_write_counter_64;
1840 counter_bits = 64;
1841 } else {
1842 mipspmu.max_period = (1ULL << 31) - 1;
1843 mipspmu.valid_count = (1ULL << 31) - 1;
1844 mipspmu.overflow = 1ULL << 31;
1845 mipspmu.read_counter = mipsxx_pmu_read_counter;
1846 mipspmu.write_counter = mipsxx_pmu_write_counter;
1847 counter_bits = 32;
1848 }
1849
1850 on_each_cpu(reset_counters, (void *)(long)counters, 1);
1851
1852 pr_cont("%s PMU enabled, %d %d-bit counters available to each "
1853 "CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq,
1854 irq < 0 ? " (share with timer interrupt)" : "");
1855
1856 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1857
1858 return 0;
1859}
1860early_initcall(init_hw_perf_events);