Merge branches 'next/ar7', 'next/ath79', 'next/bcm63xx', 'next/bmips', 'next/cavium...
[linux-2.6.git] / arch / mips / kernel / perf_event_mipsxx.c
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
32 struct cpu_hw_events {
33         /* Array of events on this cpu. */
34         struct perf_event       *events[MIPS_MAX_HWEVENTS];
35
36         /*
37          * Set the bit (indexed by the counter number) when the counter
38          * is used for an event.
39          */
40         unsigned long           used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)];
41
42         /*
43          * Software copy of the control register for each performance counter.
44          * MIPS CPUs vary in performance counters. They use this differently,
45          * and even may not use it.
46          */
47         unsigned int            saved_ctrl[MIPS_MAX_HWEVENTS];
48 };
49 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
50         .saved_ctrl = {0},
51 };
52
53 /* The description of MIPS performance events. */
54 struct mips_perf_event {
55         unsigned int event_id;
56         /*
57          * MIPS performance counters are indexed starting from 0.
58          * CNTR_EVEN indicates the indexes of the counters to be used are
59          * even numbers.
60          */
61         unsigned int cntr_mask;
62         #define CNTR_EVEN       0x55555555
63         #define CNTR_ODD        0xaaaaaaaa
64         #define CNTR_ALL        0xffffffff
65 #ifdef CONFIG_MIPS_MT_SMP
66         enum {
67                 T  = 0,
68                 V  = 1,
69                 P  = 2,
70         } range;
71 #else
72         #define T
73         #define V
74         #define P
75 #endif
76 };
77
78 static struct mips_perf_event raw_event;
79 static DEFINE_MUTEX(raw_event_mutex);
80
81 #define UNSUPPORTED_PERF_EVENT_ID 0xffffffff
82 #define C(x) PERF_COUNT_HW_CACHE_##x
83
84 struct mips_pmu {
85         u64             max_period;
86         u64             valid_count;
87         u64             overflow;
88         const char      *name;
89         int             irq;
90         u64             (*read_counter)(unsigned int idx);
91         void            (*write_counter)(unsigned int idx, u64 val);
92         const struct mips_perf_event *(*map_raw_event)(u64 config);
93         const struct mips_perf_event (*general_event_map)[PERF_COUNT_HW_MAX];
94         const struct mips_perf_event (*cache_event_map)
95                                 [PERF_COUNT_HW_CACHE_MAX]
96                                 [PERF_COUNT_HW_CACHE_OP_MAX]
97                                 [PERF_COUNT_HW_CACHE_RESULT_MAX];
98         unsigned int    num_counters;
99 };
100
101 static struct mips_pmu mipspmu;
102
103 #define M_CONFIG1_PC    (1 << 4)
104
105 #define M_PERFCTL_EXL                   (1      <<  0)
106 #define M_PERFCTL_KERNEL                (1      <<  1)
107 #define M_PERFCTL_SUPERVISOR            (1      <<  2)
108 #define M_PERFCTL_USER                  (1      <<  3)
109 #define M_PERFCTL_INTERRUPT_ENABLE      (1      <<  4)
110 #define M_PERFCTL_EVENT(event)          (((event) & 0x3ff)  << 5)
111 #define M_PERFCTL_VPEID(vpe)            ((vpe)    << 16)
112 #define M_PERFCTL_MT_EN(filter)         ((filter) << 20)
113 #define    M_TC_EN_ALL                  M_PERFCTL_MT_EN(0)
114 #define    M_TC_EN_VPE                  M_PERFCTL_MT_EN(1)
115 #define    M_TC_EN_TC                   M_PERFCTL_MT_EN(2)
116 #define M_PERFCTL_TCID(tcid)            ((tcid)   << 22)
117 #define M_PERFCTL_WIDE                  (1      << 30)
118 #define M_PERFCTL_MORE                  (1      << 31)
119
120 #define M_PERFCTL_COUNT_EVENT_WHENEVER  (M_PERFCTL_EXL |                \
121                                         M_PERFCTL_KERNEL |              \
122                                         M_PERFCTL_USER |                \
123                                         M_PERFCTL_SUPERVISOR |          \
124                                         M_PERFCTL_INTERRUPT_ENABLE)
125
126 #ifdef CONFIG_MIPS_MT_SMP
127 #define M_PERFCTL_CONFIG_MASK           0x3fff801f
128 #else
129 #define M_PERFCTL_CONFIG_MASK           0x1f
130 #endif
131 #define M_PERFCTL_EVENT_MASK            0xfe0
132
133
134 #ifdef CONFIG_MIPS_MT_SMP
135 static int cpu_has_mipsmt_pertccounters;
136
137 static DEFINE_RWLOCK(pmuint_rwlock);
138
139 /*
140  * FIXME: For VSMP, vpe_id() is redefined for Perf-events, because
141  * cpu_data[cpuid].vpe_id reports 0 for _both_ CPUs.
142  */
143 #if defined(CONFIG_HW_PERF_EVENTS)
144 #define vpe_id()        (cpu_has_mipsmt_pertccounters ? \
145                         0 : smp_processor_id())
146 #else
147 #define vpe_id()        (cpu_has_mipsmt_pertccounters ? \
148                         0 : cpu_data[smp_processor_id()].vpe_id)
149 #endif
150
151 /* Copied from op_model_mipsxx.c */
152 static unsigned int vpe_shift(void)
153 {
154         if (num_possible_cpus() > 1)
155                 return 1;
156
157         return 0;
158 }
159
160 static unsigned int counters_total_to_per_cpu(unsigned int counters)
161 {
162         return counters >> vpe_shift();
163 }
164
165 static unsigned int counters_per_cpu_to_total(unsigned int counters)
166 {
167         return counters << vpe_shift();
168 }
169
170 #else /* !CONFIG_MIPS_MT_SMP */
171 #define vpe_id()        0
172
173 #endif /* CONFIG_MIPS_MT_SMP */
174
175 static void resume_local_counters(void);
176 static void pause_local_counters(void);
177 static irqreturn_t mipsxx_pmu_handle_irq(int, void *);
178 static int mipsxx_pmu_handle_shared_irq(void);
179
180 static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx)
181 {
182         if (vpe_id() == 1)
183                 idx = (idx + 2) & 3;
184         return idx;
185 }
186
187 static u64 mipsxx_pmu_read_counter(unsigned int idx)
188 {
189         idx = mipsxx_pmu_swizzle_perf_idx(idx);
190
191         switch (idx) {
192         case 0:
193                 /*
194                  * The counters are unsigned, we must cast to truncate
195                  * off the high bits.
196                  */
197                 return (u32)read_c0_perfcntr0();
198         case 1:
199                 return (u32)read_c0_perfcntr1();
200         case 2:
201                 return (u32)read_c0_perfcntr2();
202         case 3:
203                 return (u32)read_c0_perfcntr3();
204         default:
205                 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
206                 return 0;
207         }
208 }
209
210 static u64 mipsxx_pmu_read_counter_64(unsigned int idx)
211 {
212         idx = mipsxx_pmu_swizzle_perf_idx(idx);
213
214         switch (idx) {
215         case 0:
216                 return read_c0_perfcntr0_64();
217         case 1:
218                 return read_c0_perfcntr1_64();
219         case 2:
220                 return read_c0_perfcntr2_64();
221         case 3:
222                 return read_c0_perfcntr3_64();
223         default:
224                 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
225                 return 0;
226         }
227 }
228
229 static void mipsxx_pmu_write_counter(unsigned int idx, u64 val)
230 {
231         idx = mipsxx_pmu_swizzle_perf_idx(idx);
232
233         switch (idx) {
234         case 0:
235                 write_c0_perfcntr0(val);
236                 return;
237         case 1:
238                 write_c0_perfcntr1(val);
239                 return;
240         case 2:
241                 write_c0_perfcntr2(val);
242                 return;
243         case 3:
244                 write_c0_perfcntr3(val);
245                 return;
246         }
247 }
248
249 static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val)
250 {
251         idx = mipsxx_pmu_swizzle_perf_idx(idx);
252
253         switch (idx) {
254         case 0:
255                 write_c0_perfcntr0_64(val);
256                 return;
257         case 1:
258                 write_c0_perfcntr1_64(val);
259                 return;
260         case 2:
261                 write_c0_perfcntr2_64(val);
262                 return;
263         case 3:
264                 write_c0_perfcntr3_64(val);
265                 return;
266         }
267 }
268
269 static unsigned int mipsxx_pmu_read_control(unsigned int idx)
270 {
271         idx = mipsxx_pmu_swizzle_perf_idx(idx);
272
273         switch (idx) {
274         case 0:
275                 return read_c0_perfctrl0();
276         case 1:
277                 return read_c0_perfctrl1();
278         case 2:
279                 return read_c0_perfctrl2();
280         case 3:
281                 return read_c0_perfctrl3();
282         default:
283                 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
284                 return 0;
285         }
286 }
287
288 static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val)
289 {
290         idx = mipsxx_pmu_swizzle_perf_idx(idx);
291
292         switch (idx) {
293         case 0:
294                 write_c0_perfctrl0(val);
295                 return;
296         case 1:
297                 write_c0_perfctrl1(val);
298                 return;
299         case 2:
300                 write_c0_perfctrl2(val);
301                 return;
302         case 3:
303                 write_c0_perfctrl3(val);
304                 return;
305         }
306 }
307
308 static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc,
309                                     struct hw_perf_event *hwc)
310 {
311         int i;
312
313         /*
314          * We only need to care the counter mask. The range has been
315          * checked definitely.
316          */
317         unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff;
318
319         for (i = mipspmu.num_counters - 1; i >= 0; i--) {
320                 /*
321                  * Note that some MIPS perf events can be counted by both
322                  * even and odd counters, wheresas many other are only by
323                  * even _or_ odd counters. This introduces an issue that
324                  * when the former kind of event takes the counter the
325                  * latter kind of event wants to use, then the "counter
326                  * allocation" for the latter event will fail. In fact if
327                  * they can be dynamically swapped, they both feel happy.
328                  * But here we leave this issue alone for now.
329                  */
330                 if (test_bit(i, &cntr_mask) &&
331                         !test_and_set_bit(i, cpuc->used_mask))
332                         return i;
333         }
334
335         return -EAGAIN;
336 }
337
338 static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx)
339 {
340         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
341
342         WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
343
344         cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) |
345                 (evt->config_base & M_PERFCTL_CONFIG_MASK) |
346                 /* Make sure interrupt enabled. */
347                 M_PERFCTL_INTERRUPT_ENABLE;
348         /*
349          * We do not actually let the counter run. Leave it until start().
350          */
351 }
352
353 static void mipsxx_pmu_disable_event(int idx)
354 {
355         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
356         unsigned long flags;
357
358         WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
359
360         local_irq_save(flags);
361         cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) &
362                 ~M_PERFCTL_COUNT_EVENT_WHENEVER;
363         mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]);
364         local_irq_restore(flags);
365 }
366
367 static int mipspmu_event_set_period(struct perf_event *event,
368                                     struct hw_perf_event *hwc,
369                                     int idx)
370 {
371         u64 left = local64_read(&hwc->period_left);
372         u64 period = hwc->sample_period;
373         int ret = 0;
374
375         if (unlikely((left + period) & (1ULL << 63))) {
376                 /* left underflowed by more than period. */
377                 left = period;
378                 local64_set(&hwc->period_left, left);
379                 hwc->last_period = period;
380                 ret = 1;
381         } else  if (unlikely((left + period) <= period)) {
382                 /* left underflowed by less than period. */
383                 left += period;
384                 local64_set(&hwc->period_left, left);
385                 hwc->last_period = period;
386                 ret = 1;
387         }
388
389         if (left > mipspmu.max_period) {
390                 left = mipspmu.max_period;
391                 local64_set(&hwc->period_left, left);
392         }
393
394         local64_set(&hwc->prev_count, mipspmu.overflow - left);
395
396         mipspmu.write_counter(idx, mipspmu.overflow - left);
397
398         perf_event_update_userpage(event);
399
400         return ret;
401 }
402
403 static void mipspmu_event_update(struct perf_event *event,
404                                  struct hw_perf_event *hwc,
405                                  int idx)
406 {
407         u64 prev_raw_count, new_raw_count;
408         u64 delta;
409
410 again:
411         prev_raw_count = local64_read(&hwc->prev_count);
412         new_raw_count = mipspmu.read_counter(idx);
413
414         if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
415                                 new_raw_count) != prev_raw_count)
416                 goto again;
417
418         delta = new_raw_count - prev_raw_count;
419
420         local64_add(delta, &event->count);
421         local64_sub(delta, &hwc->period_left);
422 }
423
424 static void mipspmu_start(struct perf_event *event, int flags)
425 {
426         struct hw_perf_event *hwc = &event->hw;
427
428         if (flags & PERF_EF_RELOAD)
429                 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
430
431         hwc->state = 0;
432
433         /* Set the period for the event. */
434         mipspmu_event_set_period(event, hwc, hwc->idx);
435
436         /* Enable the event. */
437         mipsxx_pmu_enable_event(hwc, hwc->idx);
438 }
439
440 static void mipspmu_stop(struct perf_event *event, int flags)
441 {
442         struct hw_perf_event *hwc = &event->hw;
443
444         if (!(hwc->state & PERF_HES_STOPPED)) {
445                 /* We are working on a local event. */
446                 mipsxx_pmu_disable_event(hwc->idx);
447                 barrier();
448                 mipspmu_event_update(event, hwc, hwc->idx);
449                 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
450         }
451 }
452
453 static int mipspmu_add(struct perf_event *event, int flags)
454 {
455         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
456         struct hw_perf_event *hwc = &event->hw;
457         int idx;
458         int err = 0;
459
460         perf_pmu_disable(event->pmu);
461
462         /* To look for a free counter for this event. */
463         idx = mipsxx_pmu_alloc_counter(cpuc, hwc);
464         if (idx < 0) {
465                 err = idx;
466                 goto out;
467         }
468
469         /*
470          * If there is an event in the counter we are going to use then
471          * make sure it is disabled.
472          */
473         event->hw.idx = idx;
474         mipsxx_pmu_disable_event(idx);
475         cpuc->events[idx] = event;
476
477         hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
478         if (flags & PERF_EF_START)
479                 mipspmu_start(event, PERF_EF_RELOAD);
480
481         /* Propagate our changes to the userspace mapping. */
482         perf_event_update_userpage(event);
483
484 out:
485         perf_pmu_enable(event->pmu);
486         return err;
487 }
488
489 static void mipspmu_del(struct perf_event *event, int flags)
490 {
491         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
492         struct hw_perf_event *hwc = &event->hw;
493         int idx = hwc->idx;
494
495         WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
496
497         mipspmu_stop(event, PERF_EF_UPDATE);
498         cpuc->events[idx] = NULL;
499         clear_bit(idx, cpuc->used_mask);
500
501         perf_event_update_userpage(event);
502 }
503
504 static void mipspmu_read(struct perf_event *event)
505 {
506         struct hw_perf_event *hwc = &event->hw;
507
508         /* Don't read disabled counters! */
509         if (hwc->idx < 0)
510                 return;
511
512         mipspmu_event_update(event, hwc, hwc->idx);
513 }
514
515 static void mipspmu_enable(struct pmu *pmu)
516 {
517 #ifdef CONFIG_MIPS_MT_SMP
518         write_unlock(&pmuint_rwlock);
519 #endif
520         resume_local_counters();
521 }
522
523 /*
524  * MIPS performance counters can be per-TC. The control registers can
525  * not be directly accessed accross CPUs. Hence if we want to do global
526  * control, we need cross CPU calls. on_each_cpu() can help us, but we
527  * can not make sure this function is called with interrupts enabled. So
528  * here we pause local counters and then grab a rwlock and leave the
529  * counters on other CPUs alone. If any counter interrupt raises while
530  * we own the write lock, simply pause local counters on that CPU and
531  * spin in the handler. Also we know we won't be switched to another
532  * CPU after pausing local counters and before grabbing the lock.
533  */
534 static void mipspmu_disable(struct pmu *pmu)
535 {
536         pause_local_counters();
537 #ifdef CONFIG_MIPS_MT_SMP
538         write_lock(&pmuint_rwlock);
539 #endif
540 }
541
542 static atomic_t active_events = ATOMIC_INIT(0);
543 static DEFINE_MUTEX(pmu_reserve_mutex);
544 static int (*save_perf_irq)(void);
545
546 static int mipspmu_get_irq(void)
547 {
548         int err;
549
550         if (mipspmu.irq >= 0) {
551                 /* Request my own irq handler. */
552                 err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq,
553                         IRQF_PERCPU | IRQF_NOBALANCING,
554                         "mips_perf_pmu", NULL);
555                 if (err) {
556                         pr_warning("Unable to request IRQ%d for MIPS "
557                            "performance counters!\n", mipspmu.irq);
558                 }
559         } else if (cp0_perfcount_irq < 0) {
560                 /*
561                  * We are sharing the irq number with the timer interrupt.
562                  */
563                 save_perf_irq = perf_irq;
564                 perf_irq = mipsxx_pmu_handle_shared_irq;
565                 err = 0;
566         } else {
567                 pr_warning("The platform hasn't properly defined its "
568                         "interrupt controller.\n");
569                 err = -ENOENT;
570         }
571
572         return err;
573 }
574
575 static void mipspmu_free_irq(void)
576 {
577         if (mipspmu.irq >= 0)
578                 free_irq(mipspmu.irq, NULL);
579         else if (cp0_perfcount_irq < 0)
580                 perf_irq = save_perf_irq;
581 }
582
583 /*
584  * mipsxx/rm9000/loongson2 have different performance counters, they have
585  * specific low-level init routines.
586  */
587 static void reset_counters(void *arg);
588 static int __hw_perf_event_init(struct perf_event *event);
589
590 static void hw_perf_event_destroy(struct perf_event *event)
591 {
592         if (atomic_dec_and_mutex_lock(&active_events,
593                                 &pmu_reserve_mutex)) {
594                 /*
595                  * We must not call the destroy function with interrupts
596                  * disabled.
597                  */
598                 on_each_cpu(reset_counters,
599                         (void *)(long)mipspmu.num_counters, 1);
600                 mipspmu_free_irq();
601                 mutex_unlock(&pmu_reserve_mutex);
602         }
603 }
604
605 static int mipspmu_event_init(struct perf_event *event)
606 {
607         int err = 0;
608
609         switch (event->attr.type) {
610         case PERF_TYPE_RAW:
611         case PERF_TYPE_HARDWARE:
612         case PERF_TYPE_HW_CACHE:
613                 break;
614
615         default:
616                 return -ENOENT;
617         }
618
619         if (event->cpu >= nr_cpumask_bits ||
620             (event->cpu >= 0 && !cpu_online(event->cpu)))
621                 return -ENODEV;
622
623         if (!atomic_inc_not_zero(&active_events)) {
624                 mutex_lock(&pmu_reserve_mutex);
625                 if (atomic_read(&active_events) == 0)
626                         err = mipspmu_get_irq();
627
628                 if (!err)
629                         atomic_inc(&active_events);
630                 mutex_unlock(&pmu_reserve_mutex);
631         }
632
633         if (err)
634                 return err;
635
636         return __hw_perf_event_init(event);
637 }
638
639 static struct pmu pmu = {
640         .pmu_enable     = mipspmu_enable,
641         .pmu_disable    = mipspmu_disable,
642         .event_init     = mipspmu_event_init,
643         .add            = mipspmu_add,
644         .del            = mipspmu_del,
645         .start          = mipspmu_start,
646         .stop           = mipspmu_stop,
647         .read           = mipspmu_read,
648 };
649
650 static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev)
651 {
652 /*
653  * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for
654  * event_id.
655  */
656 #ifdef CONFIG_MIPS_MT_SMP
657         return ((unsigned int)pev->range << 24) |
658                 (pev->cntr_mask & 0xffff00) |
659                 (pev->event_id & 0xff);
660 #else
661         return (pev->cntr_mask & 0xffff00) |
662                 (pev->event_id & 0xff);
663 #endif
664 }
665
666 static const struct mips_perf_event *mipspmu_map_general_event(int idx)
667 {
668         const struct mips_perf_event *pev;
669
670         pev = ((*mipspmu.general_event_map)[idx].event_id ==
671                 UNSUPPORTED_PERF_EVENT_ID ? ERR_PTR(-EOPNOTSUPP) :
672                 &(*mipspmu.general_event_map)[idx]);
673
674         return pev;
675 }
676
677 static const struct mips_perf_event *mipspmu_map_cache_event(u64 config)
678 {
679         unsigned int cache_type, cache_op, cache_result;
680         const struct mips_perf_event *pev;
681
682         cache_type = (config >> 0) & 0xff;
683         if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
684                 return ERR_PTR(-EINVAL);
685
686         cache_op = (config >> 8) & 0xff;
687         if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
688                 return ERR_PTR(-EINVAL);
689
690         cache_result = (config >> 16) & 0xff;
691         if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
692                 return ERR_PTR(-EINVAL);
693
694         pev = &((*mipspmu.cache_event_map)
695                                         [cache_type]
696                                         [cache_op]
697                                         [cache_result]);
698
699         if (pev->event_id == UNSUPPORTED_PERF_EVENT_ID)
700                 return ERR_PTR(-EOPNOTSUPP);
701
702         return pev;
703
704 }
705
706 static int validate_group(struct perf_event *event)
707 {
708         struct perf_event *sibling, *leader = event->group_leader;
709         struct cpu_hw_events fake_cpuc;
710
711         memset(&fake_cpuc, 0, sizeof(fake_cpuc));
712
713         if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0)
714                 return -ENOSPC;
715
716         list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
717                 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0)
718                         return -ENOSPC;
719         }
720
721         if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0)
722                 return -ENOSPC;
723
724         return 0;
725 }
726
727 /* This is needed by specific irq handlers in perf_event_*.c */
728 static void handle_associated_event(struct cpu_hw_events *cpuc,
729                                     int idx, struct perf_sample_data *data,
730                                     struct pt_regs *regs)
731 {
732         struct perf_event *event = cpuc->events[idx];
733         struct hw_perf_event *hwc = &event->hw;
734
735         mipspmu_event_update(event, hwc, idx);
736         data->period = event->hw.last_period;
737         if (!mipspmu_event_set_period(event, hwc, idx))
738                 return;
739
740         if (perf_event_overflow(event, data, regs))
741                 mipsxx_pmu_disable_event(idx);
742 }
743
744
745 static int __n_counters(void)
746 {
747         if (!(read_c0_config1() & M_CONFIG1_PC))
748                 return 0;
749         if (!(read_c0_perfctrl0() & M_PERFCTL_MORE))
750                 return 1;
751         if (!(read_c0_perfctrl1() & M_PERFCTL_MORE))
752                 return 2;
753         if (!(read_c0_perfctrl2() & M_PERFCTL_MORE))
754                 return 3;
755
756         return 4;
757 }
758
759 static int n_counters(void)
760 {
761         int counters;
762
763         switch (current_cpu_type()) {
764         case CPU_R10000:
765                 counters = 2;
766                 break;
767
768         case CPU_R12000:
769         case CPU_R14000:
770                 counters = 4;
771                 break;
772
773         default:
774                 counters = __n_counters();
775         }
776
777         return counters;
778 }
779
780 static void reset_counters(void *arg)
781 {
782         int counters = (int)(long)arg;
783         switch (counters) {
784         case 4:
785                 mipsxx_pmu_write_control(3, 0);
786                 mipspmu.write_counter(3, 0);
787         case 3:
788                 mipsxx_pmu_write_control(2, 0);
789                 mipspmu.write_counter(2, 0);
790         case 2:
791                 mipsxx_pmu_write_control(1, 0);
792                 mipspmu.write_counter(1, 0);
793         case 1:
794                 mipsxx_pmu_write_control(0, 0);
795                 mipspmu.write_counter(0, 0);
796         }
797 }
798
799 /* 24K/34K/1004K cores can share the same event map. */
800 static const struct mips_perf_event mipsxxcore_event_map
801                                 [PERF_COUNT_HW_MAX] = {
802         [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
803         [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
804         [PERF_COUNT_HW_CACHE_REFERENCES] = { UNSUPPORTED_PERF_EVENT_ID },
805         [PERF_COUNT_HW_CACHE_MISSES] = { UNSUPPORTED_PERF_EVENT_ID },
806         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T },
807         [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
808         [PERF_COUNT_HW_BUS_CYCLES] = { UNSUPPORTED_PERF_EVENT_ID },
809 };
810
811 /* 74K core has different branch event code. */
812 static const struct mips_perf_event mipsxx74Kcore_event_map
813                                 [PERF_COUNT_HW_MAX] = {
814         [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
815         [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
816         [PERF_COUNT_HW_CACHE_REFERENCES] = { UNSUPPORTED_PERF_EVENT_ID },
817         [PERF_COUNT_HW_CACHE_MISSES] = { UNSUPPORTED_PERF_EVENT_ID },
818         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T },
819         [PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T },
820         [PERF_COUNT_HW_BUS_CYCLES] = { UNSUPPORTED_PERF_EVENT_ID },
821 };
822
823 static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = {
824         [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
825         [PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL },
826         [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL },
827         [PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL  },
828         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL },
829         [PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL },
830         [PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL },
831 };
832
833 /* 24K/34K/1004K cores can share the same cache event map. */
834 static const struct mips_perf_event mipsxxcore_cache_map
835                                 [PERF_COUNT_HW_CACHE_MAX]
836                                 [PERF_COUNT_HW_CACHE_OP_MAX]
837                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
838 [C(L1D)] = {
839         /*
840          * Like some other architectures (e.g. ARM), the performance
841          * counters don't differentiate between read and write
842          * accesses/misses, so this isn't strictly correct, but it's the
843          * best we can do. Writes and reads get combined.
844          */
845         [C(OP_READ)] = {
846                 [C(RESULT_ACCESS)]      = { 0x0a, CNTR_EVEN, T },
847                 [C(RESULT_MISS)]        = { 0x0b, CNTR_EVEN | CNTR_ODD, T },
848         },
849         [C(OP_WRITE)] = {
850                 [C(RESULT_ACCESS)]      = { 0x0a, CNTR_EVEN, T },
851                 [C(RESULT_MISS)]        = { 0x0b, CNTR_EVEN | CNTR_ODD, T },
852         },
853         [C(OP_PREFETCH)] = {
854                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
855                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
856         },
857 },
858 [C(L1I)] = {
859         [C(OP_READ)] = {
860                 [C(RESULT_ACCESS)]      = { 0x09, CNTR_EVEN, T },
861                 [C(RESULT_MISS)]        = { 0x09, CNTR_ODD, T },
862         },
863         [C(OP_WRITE)] = {
864                 [C(RESULT_ACCESS)]      = { 0x09, CNTR_EVEN, T },
865                 [C(RESULT_MISS)]        = { 0x09, CNTR_ODD, T },
866         },
867         [C(OP_PREFETCH)] = {
868                 [C(RESULT_ACCESS)]      = { 0x14, CNTR_EVEN, T },
869                 /*
870                  * Note that MIPS has only "hit" events countable for
871                  * the prefetch operation.
872                  */
873                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
874         },
875 },
876 [C(LL)] = {
877         [C(OP_READ)] = {
878                 [C(RESULT_ACCESS)]      = { 0x15, CNTR_ODD, P },
879                 [C(RESULT_MISS)]        = { 0x16, CNTR_EVEN, P },
880         },
881         [C(OP_WRITE)] = {
882                 [C(RESULT_ACCESS)]      = { 0x15, CNTR_ODD, P },
883                 [C(RESULT_MISS)]        = { 0x16, CNTR_EVEN, P },
884         },
885         [C(OP_PREFETCH)] = {
886                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
887                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
888         },
889 },
890 [C(DTLB)] = {
891         [C(OP_READ)] = {
892                 [C(RESULT_ACCESS)]      = { 0x06, CNTR_EVEN, T },
893                 [C(RESULT_MISS)]        = { 0x06, CNTR_ODD, T },
894         },
895         [C(OP_WRITE)] = {
896                 [C(RESULT_ACCESS)]      = { 0x06, CNTR_EVEN, T },
897                 [C(RESULT_MISS)]        = { 0x06, CNTR_ODD, T },
898         },
899         [C(OP_PREFETCH)] = {
900                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
901                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
902         },
903 },
904 [C(ITLB)] = {
905         [C(OP_READ)] = {
906                 [C(RESULT_ACCESS)]      = { 0x05, CNTR_EVEN, T },
907                 [C(RESULT_MISS)]        = { 0x05, CNTR_ODD, T },
908         },
909         [C(OP_WRITE)] = {
910                 [C(RESULT_ACCESS)]      = { 0x05, CNTR_EVEN, T },
911                 [C(RESULT_MISS)]        = { 0x05, CNTR_ODD, T },
912         },
913         [C(OP_PREFETCH)] = {
914                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
915                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
916         },
917 },
918 [C(BPU)] = {
919         /* Using the same code for *HW_BRANCH* */
920         [C(OP_READ)] = {
921                 [C(RESULT_ACCESS)]      = { 0x02, CNTR_EVEN, T },
922                 [C(RESULT_MISS)]        = { 0x02, CNTR_ODD, T },
923         },
924         [C(OP_WRITE)] = {
925                 [C(RESULT_ACCESS)]      = { 0x02, CNTR_EVEN, T },
926                 [C(RESULT_MISS)]        = { 0x02, CNTR_ODD, T },
927         },
928         [C(OP_PREFETCH)] = {
929                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
930                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
931         },
932 },
933 [C(NODE)] = {
934         [C(OP_READ)] = {
935                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
936                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
937         },
938         [C(OP_WRITE)] = {
939                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
940                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
941         },
942         [C(OP_PREFETCH)] = {
943                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
944                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
945         },
946 },
947 };
948
949 /* 74K core has completely different cache event map. */
950 static const struct mips_perf_event mipsxx74Kcore_cache_map
951                                 [PERF_COUNT_HW_CACHE_MAX]
952                                 [PERF_COUNT_HW_CACHE_OP_MAX]
953                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
954 [C(L1D)] = {
955         /*
956          * Like some other architectures (e.g. ARM), the performance
957          * counters don't differentiate between read and write
958          * accesses/misses, so this isn't strictly correct, but it's the
959          * best we can do. Writes and reads get combined.
960          */
961         [C(OP_READ)] = {
962                 [C(RESULT_ACCESS)]      = { 0x17, CNTR_ODD, T },
963                 [C(RESULT_MISS)]        = { 0x18, CNTR_ODD, T },
964         },
965         [C(OP_WRITE)] = {
966                 [C(RESULT_ACCESS)]      = { 0x17, CNTR_ODD, T },
967                 [C(RESULT_MISS)]        = { 0x18, CNTR_ODD, T },
968         },
969         [C(OP_PREFETCH)] = {
970                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
971                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
972         },
973 },
974 [C(L1I)] = {
975         [C(OP_READ)] = {
976                 [C(RESULT_ACCESS)]      = { 0x06, CNTR_EVEN, T },
977                 [C(RESULT_MISS)]        = { 0x06, CNTR_ODD, T },
978         },
979         [C(OP_WRITE)] = {
980                 [C(RESULT_ACCESS)]      = { 0x06, CNTR_EVEN, T },
981                 [C(RESULT_MISS)]        = { 0x06, CNTR_ODD, T },
982         },
983         [C(OP_PREFETCH)] = {
984                 [C(RESULT_ACCESS)]      = { 0x34, CNTR_EVEN, T },
985                 /*
986                  * Note that MIPS has only "hit" events countable for
987                  * the prefetch operation.
988                  */
989                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
990         },
991 },
992 [C(LL)] = {
993         [C(OP_READ)] = {
994                 [C(RESULT_ACCESS)]      = { 0x1c, CNTR_ODD, P },
995                 [C(RESULT_MISS)]        = { 0x1d, CNTR_EVEN | CNTR_ODD, P },
996         },
997         [C(OP_WRITE)] = {
998                 [C(RESULT_ACCESS)]      = { 0x1c, CNTR_ODD, P },
999                 [C(RESULT_MISS)]        = { 0x1d, CNTR_EVEN | CNTR_ODD, P },
1000         },
1001         [C(OP_PREFETCH)] = {
1002                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1003                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1004         },
1005 },
1006 [C(DTLB)] = {
1007         /* 74K core does not have specific DTLB events. */
1008         [C(OP_READ)] = {
1009                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1010                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1011         },
1012         [C(OP_WRITE)] = {
1013                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1014                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1015         },
1016         [C(OP_PREFETCH)] = {
1017                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1018                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1019         },
1020 },
1021 [C(ITLB)] = {
1022         [C(OP_READ)] = {
1023                 [C(RESULT_ACCESS)]      = { 0x04, CNTR_EVEN, T },
1024                 [C(RESULT_MISS)]        = { 0x04, CNTR_ODD, T },
1025         },
1026         [C(OP_WRITE)] = {
1027                 [C(RESULT_ACCESS)]      = { 0x04, CNTR_EVEN, T },
1028                 [C(RESULT_MISS)]        = { 0x04, CNTR_ODD, T },
1029         },
1030         [C(OP_PREFETCH)] = {
1031                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1032                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1033         },
1034 },
1035 [C(BPU)] = {
1036         /* Using the same code for *HW_BRANCH* */
1037         [C(OP_READ)] = {
1038                 [C(RESULT_ACCESS)]      = { 0x27, CNTR_EVEN, T },
1039                 [C(RESULT_MISS)]        = { 0x27, CNTR_ODD, T },
1040         },
1041         [C(OP_WRITE)] = {
1042                 [C(RESULT_ACCESS)]      = { 0x27, CNTR_EVEN, T },
1043                 [C(RESULT_MISS)]        = { 0x27, CNTR_ODD, T },
1044         },
1045         [C(OP_PREFETCH)] = {
1046                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1047                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1048         },
1049 },
1050 [C(NODE)] = {
1051         [C(OP_READ)] = {
1052                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1053                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1054         },
1055         [C(OP_WRITE)] = {
1056                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1057                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1058         },
1059         [C(OP_PREFETCH)] = {
1060                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1061                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1062         },
1063 },
1064 };
1065
1066
1067 static const struct mips_perf_event octeon_cache_map
1068                                 [PERF_COUNT_HW_CACHE_MAX]
1069                                 [PERF_COUNT_HW_CACHE_OP_MAX]
1070                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1071 [C(L1D)] = {
1072         [C(OP_READ)] = {
1073                 [C(RESULT_ACCESS)]      = { 0x2b, CNTR_ALL },
1074                 [C(RESULT_MISS)]        = { 0x2e, CNTR_ALL },
1075         },
1076         [C(OP_WRITE)] = {
1077                 [C(RESULT_ACCESS)]      = { 0x30, CNTR_ALL },
1078                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1079         },
1080         [C(OP_PREFETCH)] = {
1081                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1082                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1083         },
1084 },
1085 [C(L1I)] = {
1086         [C(OP_READ)] = {
1087                 [C(RESULT_ACCESS)]      = { 0x18, CNTR_ALL },
1088                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1089         },
1090         [C(OP_WRITE)] = {
1091                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1092                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1093         },
1094         [C(OP_PREFETCH)] = {
1095                 [C(RESULT_ACCESS)]      = { 0x19, CNTR_ALL },
1096                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1097         },
1098 },
1099 [C(LL)] = {
1100         [C(OP_READ)] = {
1101                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1102                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1103         },
1104         [C(OP_WRITE)] = {
1105                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1106                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1107         },
1108         [C(OP_PREFETCH)] = {
1109                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1110                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1111         },
1112 },
1113 [C(DTLB)] = {
1114         /*
1115          * Only general DTLB misses are counted use the same event for
1116          * read and write.
1117          */
1118         [C(OP_READ)] = {
1119                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1120                 [C(RESULT_MISS)]        = { 0x35, CNTR_ALL },
1121         },
1122         [C(OP_WRITE)] = {
1123                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1124                 [C(RESULT_MISS)]        = { 0x35, CNTR_ALL },
1125         },
1126         [C(OP_PREFETCH)] = {
1127                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1128                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1129         },
1130 },
1131 [C(ITLB)] = {
1132         [C(OP_READ)] = {
1133                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1134                 [C(RESULT_MISS)]        = { 0x37, CNTR_ALL },
1135         },
1136         [C(OP_WRITE)] = {
1137                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1138                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1139         },
1140         [C(OP_PREFETCH)] = {
1141                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1142                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1143         },
1144 },
1145 [C(BPU)] = {
1146         /* Using the same code for *HW_BRANCH* */
1147         [C(OP_READ)] = {
1148                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1149                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1150         },
1151         [C(OP_WRITE)] = {
1152                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1153                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1154         },
1155         [C(OP_PREFETCH)] = {
1156                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
1157                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
1158         },
1159 },
1160 };
1161
1162 #ifdef CONFIG_MIPS_MT_SMP
1163 static void check_and_calc_range(struct perf_event *event,
1164                                  const struct mips_perf_event *pev)
1165 {
1166         struct hw_perf_event *hwc = &event->hw;
1167
1168         if (event->cpu >= 0) {
1169                 if (pev->range > V) {
1170                         /*
1171                          * The user selected an event that is processor
1172                          * wide, while expecting it to be VPE wide.
1173                          */
1174                         hwc->config_base |= M_TC_EN_ALL;
1175                 } else {
1176                         /*
1177                          * FIXME: cpu_data[event->cpu].vpe_id reports 0
1178                          * for both CPUs.
1179                          */
1180                         hwc->config_base |= M_PERFCTL_VPEID(event->cpu);
1181                         hwc->config_base |= M_TC_EN_VPE;
1182                 }
1183         } else
1184                 hwc->config_base |= M_TC_EN_ALL;
1185 }
1186 #else
1187 static void check_and_calc_range(struct perf_event *event,
1188                                  const struct mips_perf_event *pev)
1189 {
1190 }
1191 #endif
1192
1193 static int __hw_perf_event_init(struct perf_event *event)
1194 {
1195         struct perf_event_attr *attr = &event->attr;
1196         struct hw_perf_event *hwc = &event->hw;
1197         const struct mips_perf_event *pev;
1198         int err;
1199
1200         /* Returning MIPS event descriptor for generic perf event. */
1201         if (PERF_TYPE_HARDWARE == event->attr.type) {
1202                 if (event->attr.config >= PERF_COUNT_HW_MAX)
1203                         return -EINVAL;
1204                 pev = mipspmu_map_general_event(event->attr.config);
1205         } else if (PERF_TYPE_HW_CACHE == event->attr.type) {
1206                 pev = mipspmu_map_cache_event(event->attr.config);
1207         } else if (PERF_TYPE_RAW == event->attr.type) {
1208                 /* We are working on the global raw event. */
1209                 mutex_lock(&raw_event_mutex);
1210                 pev = mipspmu.map_raw_event(event->attr.config);
1211         } else {
1212                 /* The event type is not (yet) supported. */
1213                 return -EOPNOTSUPP;
1214         }
1215
1216         if (IS_ERR(pev)) {
1217                 if (PERF_TYPE_RAW == event->attr.type)
1218                         mutex_unlock(&raw_event_mutex);
1219                 return PTR_ERR(pev);
1220         }
1221
1222         /*
1223          * We allow max flexibility on how each individual counter shared
1224          * by the single CPU operates (the mode exclusion and the range).
1225          */
1226         hwc->config_base = M_PERFCTL_INTERRUPT_ENABLE;
1227
1228         /* Calculate range bits and validate it. */
1229         if (num_possible_cpus() > 1)
1230                 check_and_calc_range(event, pev);
1231
1232         hwc->event_base = mipspmu_perf_event_encode(pev);
1233         if (PERF_TYPE_RAW == event->attr.type)
1234                 mutex_unlock(&raw_event_mutex);
1235
1236         if (!attr->exclude_user)
1237                 hwc->config_base |= M_PERFCTL_USER;
1238         if (!attr->exclude_kernel) {
1239                 hwc->config_base |= M_PERFCTL_KERNEL;
1240                 /* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */
1241                 hwc->config_base |= M_PERFCTL_EXL;
1242         }
1243         if (!attr->exclude_hv)
1244                 hwc->config_base |= M_PERFCTL_SUPERVISOR;
1245
1246         hwc->config_base &= M_PERFCTL_CONFIG_MASK;
1247         /*
1248          * The event can belong to another cpu. We do not assign a local
1249          * counter for it for now.
1250          */
1251         hwc->idx = -1;
1252         hwc->config = 0;
1253
1254         if (!hwc->sample_period) {
1255                 hwc->sample_period  = mipspmu.max_period;
1256                 hwc->last_period    = hwc->sample_period;
1257                 local64_set(&hwc->period_left, hwc->sample_period);
1258         }
1259
1260         err = 0;
1261         if (event->group_leader != event)
1262                 err = validate_group(event);
1263
1264         event->destroy = hw_perf_event_destroy;
1265
1266         if (err)
1267                 event->destroy(event);
1268
1269         return err;
1270 }
1271
1272 static void pause_local_counters(void)
1273 {
1274         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1275         int ctr = mipspmu.num_counters;
1276         unsigned long flags;
1277
1278         local_irq_save(flags);
1279         do {
1280                 ctr--;
1281                 cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr);
1282                 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] &
1283                                          ~M_PERFCTL_COUNT_EVENT_WHENEVER);
1284         } while (ctr > 0);
1285         local_irq_restore(flags);
1286 }
1287
1288 static void resume_local_counters(void)
1289 {
1290         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1291         int ctr = mipspmu.num_counters;
1292
1293         do {
1294                 ctr--;
1295                 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]);
1296         } while (ctr > 0);
1297 }
1298
1299 static int mipsxx_pmu_handle_shared_irq(void)
1300 {
1301         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1302         struct perf_sample_data data;
1303         unsigned int counters = mipspmu.num_counters;
1304         u64 counter;
1305         int handled = IRQ_NONE;
1306         struct pt_regs *regs;
1307
1308         if (cpu_has_mips_r2 && !(read_c0_cause() & (1 << 26)))
1309                 return handled;
1310         /*
1311          * First we pause the local counters, so that when we are locked
1312          * here, the counters are all paused. When it gets locked due to
1313          * perf_disable(), the timer interrupt handler will be delayed.
1314          *
1315          * See also mipsxx_pmu_start().
1316          */
1317         pause_local_counters();
1318 #ifdef CONFIG_MIPS_MT_SMP
1319         read_lock(&pmuint_rwlock);
1320 #endif
1321
1322         regs = get_irq_regs();
1323
1324         perf_sample_data_init(&data, 0);
1325
1326         switch (counters) {
1327 #define HANDLE_COUNTER(n)                                               \
1328         case n + 1:                                                     \
1329                 if (test_bit(n, cpuc->used_mask)) {                     \
1330                         counter = mipspmu.read_counter(n);              \
1331                         if (counter & mipspmu.overflow) {               \
1332                                 handle_associated_event(cpuc, n, &data, regs); \
1333                                 handled = IRQ_HANDLED;                  \
1334                         }                                               \
1335                 }
1336         HANDLE_COUNTER(3)
1337         HANDLE_COUNTER(2)
1338         HANDLE_COUNTER(1)
1339         HANDLE_COUNTER(0)
1340         }
1341
1342         /*
1343          * Do all the work for the pending perf events. We can do this
1344          * in here because the performance counter interrupt is a regular
1345          * interrupt, not NMI.
1346          */
1347         if (handled == IRQ_HANDLED)
1348                 irq_work_run();
1349
1350 #ifdef CONFIG_MIPS_MT_SMP
1351         read_unlock(&pmuint_rwlock);
1352 #endif
1353         resume_local_counters();
1354         return handled;
1355 }
1356
1357 static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev)
1358 {
1359         return mipsxx_pmu_handle_shared_irq();
1360 }
1361
1362 /* 24K */
1363 #define IS_BOTH_COUNTERS_24K_EVENT(b)                                   \
1364         ((b) == 0 || (b) == 1 || (b) == 11)
1365
1366 /* 34K */
1367 #define IS_BOTH_COUNTERS_34K_EVENT(b)                                   \
1368         ((b) == 0 || (b) == 1 || (b) == 11)
1369 #ifdef CONFIG_MIPS_MT_SMP
1370 #define IS_RANGE_P_34K_EVENT(r, b)                                      \
1371         ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||             \
1372          (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 ||           \
1373          (r) == 176 || ((b) >= 50 && (b) <= 55) ||                      \
1374          ((b) >= 64 && (b) <= 67))
1375 #define IS_RANGE_V_34K_EVENT(r) ((r) == 47)
1376 #endif
1377
1378 /* 74K */
1379 #define IS_BOTH_COUNTERS_74K_EVENT(b)                                   \
1380         ((b) == 0 || (b) == 1)
1381
1382 /* 1004K */
1383 #define IS_BOTH_COUNTERS_1004K_EVENT(b)                                 \
1384         ((b) == 0 || (b) == 1 || (b) == 11)
1385 #ifdef CONFIG_MIPS_MT_SMP
1386 #define IS_RANGE_P_1004K_EVENT(r, b)                                    \
1387         ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||             \
1388          (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 ||            \
1389          (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) ||        \
1390          (r) == 188 || (b) == 61 || (b) == 62 ||                        \
1391          ((b) >= 64 && (b) <= 67))
1392 #define IS_RANGE_V_1004K_EVENT(r)       ((r) == 47)
1393 #endif
1394
1395 /*
1396  * User can use 0-255 raw events, where 0-127 for the events of even
1397  * counters, and 128-255 for odd counters. Note that bit 7 is used to
1398  * indicate the parity. So, for example, when user wants to take the
1399  * Event Num of 15 for odd counters (by referring to the user manual),
1400  * then 128 needs to be added to 15 as the input for the event config,
1401  * i.e., 143 (0x8F) to be used.
1402  */
1403 static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config)
1404 {
1405         unsigned int raw_id = config & 0xff;
1406         unsigned int base_id = raw_id & 0x7f;
1407
1408         raw_event.event_id = base_id;
1409
1410         switch (current_cpu_type()) {
1411         case CPU_24K:
1412                 if (IS_BOTH_COUNTERS_24K_EVENT(base_id))
1413                         raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1414                 else
1415                         raw_event.cntr_mask =
1416                                 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1417 #ifdef CONFIG_MIPS_MT_SMP
1418                 /*
1419                  * This is actually doing nothing. Non-multithreading
1420                  * CPUs will not check and calculate the range.
1421                  */
1422                 raw_event.range = P;
1423 #endif
1424                 break;
1425         case CPU_34K:
1426                 if (IS_BOTH_COUNTERS_34K_EVENT(base_id))
1427                         raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1428                 else
1429                         raw_event.cntr_mask =
1430                                 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1431 #ifdef CONFIG_MIPS_MT_SMP
1432                 if (IS_RANGE_P_34K_EVENT(raw_id, base_id))
1433                         raw_event.range = P;
1434                 else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id)))
1435                         raw_event.range = V;
1436                 else
1437                         raw_event.range = T;
1438 #endif
1439                 break;
1440         case CPU_74K:
1441                 if (IS_BOTH_COUNTERS_74K_EVENT(base_id))
1442                         raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1443                 else
1444                         raw_event.cntr_mask =
1445                                 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1446 #ifdef CONFIG_MIPS_MT_SMP
1447                 raw_event.range = P;
1448 #endif
1449                 break;
1450         case CPU_1004K:
1451                 if (IS_BOTH_COUNTERS_1004K_EVENT(base_id))
1452                         raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1453                 else
1454                         raw_event.cntr_mask =
1455                                 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1456 #ifdef CONFIG_MIPS_MT_SMP
1457                 if (IS_RANGE_P_1004K_EVENT(raw_id, base_id))
1458                         raw_event.range = P;
1459                 else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id)))
1460                         raw_event.range = V;
1461                 else
1462                         raw_event.range = T;
1463 #endif
1464                 break;
1465         }
1466
1467         return &raw_event;
1468 }
1469
1470 static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config)
1471 {
1472         unsigned int raw_id = config & 0xff;
1473         unsigned int base_id = raw_id & 0x7f;
1474
1475
1476         raw_event.cntr_mask = CNTR_ALL;
1477         raw_event.event_id = base_id;
1478
1479         if (current_cpu_type() == CPU_CAVIUM_OCTEON2) {
1480                 if (base_id > 0x42)
1481                         return ERR_PTR(-EOPNOTSUPP);
1482         } else {
1483                 if (base_id > 0x3a)
1484                         return ERR_PTR(-EOPNOTSUPP);
1485         }
1486
1487         switch (base_id) {
1488         case 0x00:
1489         case 0x0f:
1490         case 0x1e:
1491         case 0x1f:
1492         case 0x2f:
1493         case 0x34:
1494         case 0x3b ... 0x3f:
1495                 return ERR_PTR(-EOPNOTSUPP);
1496         default:
1497                 break;
1498         }
1499
1500         return &raw_event;
1501 }
1502
1503 static int __init
1504 init_hw_perf_events(void)
1505 {
1506         int counters, irq;
1507         int counter_bits;
1508
1509         pr_info("Performance counters: ");
1510
1511         counters = n_counters();
1512         if (counters == 0) {
1513                 pr_cont("No available PMU.\n");
1514                 return -ENODEV;
1515         }
1516
1517 #ifdef CONFIG_MIPS_MT_SMP
1518         cpu_has_mipsmt_pertccounters = read_c0_config7() & (1<<19);
1519         if (!cpu_has_mipsmt_pertccounters)
1520                 counters = counters_total_to_per_cpu(counters);
1521 #endif
1522
1523 #ifdef MSC01E_INT_BASE
1524         if (cpu_has_veic) {
1525                 /*
1526                  * Using platform specific interrupt controller defines.
1527                  */
1528                 irq = MSC01E_INT_BASE + MSC01E_INT_PERFCTR;
1529         } else {
1530 #endif
1531                 if (cp0_perfcount_irq >= 0)
1532                         irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq;
1533                 else
1534                         irq = -1;
1535 #ifdef MSC01E_INT_BASE
1536         }
1537 #endif
1538
1539         mipspmu.map_raw_event = mipsxx_pmu_map_raw_event;
1540
1541         switch (current_cpu_type()) {
1542         case CPU_24K:
1543                 mipspmu.name = "mips/24K";
1544                 mipspmu.general_event_map = &mipsxxcore_event_map;
1545                 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1546                 break;
1547         case CPU_34K:
1548                 mipspmu.name = "mips/34K";
1549                 mipspmu.general_event_map = &mipsxxcore_event_map;
1550                 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1551                 break;
1552         case CPU_74K:
1553                 mipspmu.name = "mips/74K";
1554                 mipspmu.general_event_map = &mipsxx74Kcore_event_map;
1555                 mipspmu.cache_event_map = &mipsxx74Kcore_cache_map;
1556                 break;
1557         case CPU_1004K:
1558                 mipspmu.name = "mips/1004K";
1559                 mipspmu.general_event_map = &mipsxxcore_event_map;
1560                 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1561                 break;
1562         case CPU_CAVIUM_OCTEON:
1563         case CPU_CAVIUM_OCTEON_PLUS:
1564         case CPU_CAVIUM_OCTEON2:
1565                 mipspmu.name = "octeon";
1566                 mipspmu.general_event_map = &octeon_event_map;
1567                 mipspmu.cache_event_map = &octeon_cache_map;
1568                 mipspmu.map_raw_event = octeon_pmu_map_raw_event;
1569                 break;
1570         default:
1571                 pr_cont("Either hardware does not support performance "
1572                         "counters, or not yet implemented.\n");
1573                 return -ENODEV;
1574         }
1575
1576         mipspmu.num_counters = counters;
1577         mipspmu.irq = irq;
1578
1579         if (read_c0_perfctrl0() & M_PERFCTL_WIDE) {
1580                 mipspmu.max_period = (1ULL << 63) - 1;
1581                 mipspmu.valid_count = (1ULL << 63) - 1;
1582                 mipspmu.overflow = 1ULL << 63;
1583                 mipspmu.read_counter = mipsxx_pmu_read_counter_64;
1584                 mipspmu.write_counter = mipsxx_pmu_write_counter_64;
1585                 counter_bits = 64;
1586         } else {
1587                 mipspmu.max_period = (1ULL << 31) - 1;
1588                 mipspmu.valid_count = (1ULL << 31) - 1;
1589                 mipspmu.overflow = 1ULL << 31;
1590                 mipspmu.read_counter = mipsxx_pmu_read_counter;
1591                 mipspmu.write_counter = mipsxx_pmu_write_counter;
1592                 counter_bits = 32;
1593         }
1594
1595         on_each_cpu(reset_counters, (void *)(long)counters, 1);
1596
1597         pr_cont("%s PMU enabled, %d %d-bit counters available to each "
1598                 "CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq,
1599                 irq < 0 ? " (share with timer interrupt)" : "");
1600
1601         perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1602
1603         return 0;
1604 }
1605 early_initcall(init_hw_perf_events);