Merge branch 'timers-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[pandora-kernel.git] / arch / mips / kernel / perf_event_mipsxx.c
1 #if defined(CONFIG_CPU_MIPS32) || defined(CONFIG_CPU_MIPS64) || \
2     defined(CONFIG_CPU_R10000) || defined(CONFIG_CPU_SB1)
3
4 #define M_CONFIG1_PC    (1 << 4)
5
6 #define M_PERFCTL_EXL                   (1UL      <<  0)
7 #define M_PERFCTL_KERNEL                (1UL      <<  1)
8 #define M_PERFCTL_SUPERVISOR            (1UL      <<  2)
9 #define M_PERFCTL_USER                  (1UL      <<  3)
10 #define M_PERFCTL_INTERRUPT_ENABLE      (1UL      <<  4)
11 #define M_PERFCTL_EVENT(event)          (((event) & 0x3ff)  << 5)
12 #define M_PERFCTL_VPEID(vpe)            ((vpe)    << 16)
13 #define M_PERFCTL_MT_EN(filter)         ((filter) << 20)
14 #define    M_TC_EN_ALL                  M_PERFCTL_MT_EN(0)
15 #define    M_TC_EN_VPE                  M_PERFCTL_MT_EN(1)
16 #define    M_TC_EN_TC                   M_PERFCTL_MT_EN(2)
17 #define M_PERFCTL_TCID(tcid)            ((tcid)   << 22)
18 #define M_PERFCTL_WIDE                  (1UL      << 30)
19 #define M_PERFCTL_MORE                  (1UL      << 31)
20
21 #define M_PERFCTL_COUNT_EVENT_WHENEVER  (M_PERFCTL_EXL |                \
22                                         M_PERFCTL_KERNEL |              \
23                                         M_PERFCTL_USER |                \
24                                         M_PERFCTL_SUPERVISOR |          \
25                                         M_PERFCTL_INTERRUPT_ENABLE)
26
27 #ifdef CONFIG_MIPS_MT_SMP
28 #define M_PERFCTL_CONFIG_MASK           0x3fff801f
29 #else
30 #define M_PERFCTL_CONFIG_MASK           0x1f
31 #endif
32 #define M_PERFCTL_EVENT_MASK            0xfe0
33
34 #define M_COUNTER_OVERFLOW              (1UL      << 31)
35
36 #ifdef CONFIG_MIPS_MT_SMP
37 static int cpu_has_mipsmt_pertccounters;
38
39 /*
40  * FIXME: For VSMP, vpe_id() is redefined for Perf-events, because
41  * cpu_data[cpuid].vpe_id reports 0 for _both_ CPUs.
42  */
43 #if defined(CONFIG_HW_PERF_EVENTS)
44 #define vpe_id()        (cpu_has_mipsmt_pertccounters ? \
45                         0 : smp_processor_id())
46 #else
47 #define vpe_id()        (cpu_has_mipsmt_pertccounters ? \
48                         0 : cpu_data[smp_processor_id()].vpe_id)
49 #endif
50
51 /* Copied from op_model_mipsxx.c */
52 static inline unsigned int vpe_shift(void)
53 {
54         if (num_possible_cpus() > 1)
55                 return 1;
56
57         return 0;
58 }
59 #else /* !CONFIG_MIPS_MT_SMP */
60 #define vpe_id()        0
61
62 static inline unsigned int vpe_shift(void)
63 {
64         return 0;
65 }
66 #endif /* CONFIG_MIPS_MT_SMP */
67
68 static inline unsigned int
69 counters_total_to_per_cpu(unsigned int counters)
70 {
71         return counters >> vpe_shift();
72 }
73
74 static inline unsigned int
75 counters_per_cpu_to_total(unsigned int counters)
76 {
77         return counters << vpe_shift();
78 }
79
80 #define __define_perf_accessors(r, n, np)                               \
81                                                                         \
82 static inline unsigned int r_c0_ ## r ## n(void)                        \
83 {                                                                       \
84         unsigned int cpu = vpe_id();                                    \
85                                                                         \
86         switch (cpu) {                                                  \
87         case 0:                                                         \
88                 return read_c0_ ## r ## n();                            \
89         case 1:                                                         \
90                 return read_c0_ ## r ## np();                           \
91         default:                                                        \
92                 BUG();                                                  \
93         }                                                               \
94         return 0;                                                       \
95 }                                                                       \
96                                                                         \
97 static inline void w_c0_ ## r ## n(unsigned int value)                  \
98 {                                                                       \
99         unsigned int cpu = vpe_id();                                    \
100                                                                         \
101         switch (cpu) {                                                  \
102         case 0:                                                         \
103                 write_c0_ ## r ## n(value);                             \
104                 return;                                                 \
105         case 1:                                                         \
106                 write_c0_ ## r ## np(value);                            \
107                 return;                                                 \
108         default:                                                        \
109                 BUG();                                                  \
110         }                                                               \
111         return;                                                         \
112 }                                                                       \
113
114 __define_perf_accessors(perfcntr, 0, 2)
115 __define_perf_accessors(perfcntr, 1, 3)
116 __define_perf_accessors(perfcntr, 2, 0)
117 __define_perf_accessors(perfcntr, 3, 1)
118
119 __define_perf_accessors(perfctrl, 0, 2)
120 __define_perf_accessors(perfctrl, 1, 3)
121 __define_perf_accessors(perfctrl, 2, 0)
122 __define_perf_accessors(perfctrl, 3, 1)
123
124 static inline int __n_counters(void)
125 {
126         if (!(read_c0_config1() & M_CONFIG1_PC))
127                 return 0;
128         if (!(read_c0_perfctrl0() & M_PERFCTL_MORE))
129                 return 1;
130         if (!(read_c0_perfctrl1() & M_PERFCTL_MORE))
131                 return 2;
132         if (!(read_c0_perfctrl2() & M_PERFCTL_MORE))
133                 return 3;
134
135         return 4;
136 }
137
138 static inline int n_counters(void)
139 {
140         int counters;
141
142         switch (current_cpu_type()) {
143         case CPU_R10000:
144                 counters = 2;
145                 break;
146
147         case CPU_R12000:
148         case CPU_R14000:
149                 counters = 4;
150                 break;
151
152         default:
153                 counters = __n_counters();
154         }
155
156         return counters;
157 }
158
159 static void reset_counters(void *arg)
160 {
161         int counters = (int)(long)arg;
162         switch (counters) {
163         case 4:
164                 w_c0_perfctrl3(0);
165                 w_c0_perfcntr3(0);
166         case 3:
167                 w_c0_perfctrl2(0);
168                 w_c0_perfcntr2(0);
169         case 2:
170                 w_c0_perfctrl1(0);
171                 w_c0_perfcntr1(0);
172         case 1:
173                 w_c0_perfctrl0(0);
174                 w_c0_perfcntr0(0);
175         }
176 }
177
178 static inline u64
179 mipsxx_pmu_read_counter(unsigned int idx)
180 {
181         switch (idx) {
182         case 0:
183                 return r_c0_perfcntr0();
184         case 1:
185                 return r_c0_perfcntr1();
186         case 2:
187                 return r_c0_perfcntr2();
188         case 3:
189                 return r_c0_perfcntr3();
190         default:
191                 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
192                 return 0;
193         }
194 }
195
196 static inline void
197 mipsxx_pmu_write_counter(unsigned int idx, u64 val)
198 {
199         switch (idx) {
200         case 0:
201                 w_c0_perfcntr0(val);
202                 return;
203         case 1:
204                 w_c0_perfcntr1(val);
205                 return;
206         case 2:
207                 w_c0_perfcntr2(val);
208                 return;
209         case 3:
210                 w_c0_perfcntr3(val);
211                 return;
212         }
213 }
214
215 static inline unsigned int
216 mipsxx_pmu_read_control(unsigned int idx)
217 {
218         switch (idx) {
219         case 0:
220                 return r_c0_perfctrl0();
221         case 1:
222                 return r_c0_perfctrl1();
223         case 2:
224                 return r_c0_perfctrl2();
225         case 3:
226                 return r_c0_perfctrl3();
227         default:
228                 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
229                 return 0;
230         }
231 }
232
233 static inline void
234 mipsxx_pmu_write_control(unsigned int idx, unsigned int val)
235 {
236         switch (idx) {
237         case 0:
238                 w_c0_perfctrl0(val);
239                 return;
240         case 1:
241                 w_c0_perfctrl1(val);
242                 return;
243         case 2:
244                 w_c0_perfctrl2(val);
245                 return;
246         case 3:
247                 w_c0_perfctrl3(val);
248                 return;
249         }
250 }
251
252 #ifdef CONFIG_MIPS_MT_SMP
253 static DEFINE_RWLOCK(pmuint_rwlock);
254 #endif
255
256 /* 24K/34K/1004K cores can share the same event map. */
257 static const struct mips_perf_event mipsxxcore_event_map
258                                 [PERF_COUNT_HW_MAX] = {
259         [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
260         [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
261         [PERF_COUNT_HW_CACHE_REFERENCES] = { UNSUPPORTED_PERF_EVENT_ID },
262         [PERF_COUNT_HW_CACHE_MISSES] = { UNSUPPORTED_PERF_EVENT_ID },
263         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T },
264         [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
265         [PERF_COUNT_HW_BUS_CYCLES] = { UNSUPPORTED_PERF_EVENT_ID },
266 };
267
268 /* 74K core has different branch event code. */
269 static const struct mips_perf_event mipsxx74Kcore_event_map
270                                 [PERF_COUNT_HW_MAX] = {
271         [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
272         [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
273         [PERF_COUNT_HW_CACHE_REFERENCES] = { UNSUPPORTED_PERF_EVENT_ID },
274         [PERF_COUNT_HW_CACHE_MISSES] = { UNSUPPORTED_PERF_EVENT_ID },
275         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T },
276         [PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T },
277         [PERF_COUNT_HW_BUS_CYCLES] = { UNSUPPORTED_PERF_EVENT_ID },
278 };
279
280 /* 24K/34K/1004K cores can share the same cache event map. */
281 static const struct mips_perf_event mipsxxcore_cache_map
282                                 [PERF_COUNT_HW_CACHE_MAX]
283                                 [PERF_COUNT_HW_CACHE_OP_MAX]
284                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
285 [C(L1D)] = {
286         /*
287          * Like some other architectures (e.g. ARM), the performance
288          * counters don't differentiate between read and write
289          * accesses/misses, so this isn't strictly correct, but it's the
290          * best we can do. Writes and reads get combined.
291          */
292         [C(OP_READ)] = {
293                 [C(RESULT_ACCESS)]      = { 0x0a, CNTR_EVEN, T },
294                 [C(RESULT_MISS)]        = { 0x0b, CNTR_EVEN | CNTR_ODD, T },
295         },
296         [C(OP_WRITE)] = {
297                 [C(RESULT_ACCESS)]      = { 0x0a, CNTR_EVEN, T },
298                 [C(RESULT_MISS)]        = { 0x0b, CNTR_EVEN | CNTR_ODD, T },
299         },
300         [C(OP_PREFETCH)] = {
301                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
302                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
303         },
304 },
305 [C(L1I)] = {
306         [C(OP_READ)] = {
307                 [C(RESULT_ACCESS)]      = { 0x09, CNTR_EVEN, T },
308                 [C(RESULT_MISS)]        = { 0x09, CNTR_ODD, T },
309         },
310         [C(OP_WRITE)] = {
311                 [C(RESULT_ACCESS)]      = { 0x09, CNTR_EVEN, T },
312                 [C(RESULT_MISS)]        = { 0x09, CNTR_ODD, T },
313         },
314         [C(OP_PREFETCH)] = {
315                 [C(RESULT_ACCESS)]      = { 0x14, CNTR_EVEN, T },
316                 /*
317                  * Note that MIPS has only "hit" events countable for
318                  * the prefetch operation.
319                  */
320                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
321         },
322 },
323 [C(LL)] = {
324         [C(OP_READ)] = {
325                 [C(RESULT_ACCESS)]      = { 0x15, CNTR_ODD, P },
326                 [C(RESULT_MISS)]        = { 0x16, CNTR_EVEN, P },
327         },
328         [C(OP_WRITE)] = {
329                 [C(RESULT_ACCESS)]      = { 0x15, CNTR_ODD, P },
330                 [C(RESULT_MISS)]        = { 0x16, CNTR_EVEN, P },
331         },
332         [C(OP_PREFETCH)] = {
333                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
334                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
335         },
336 },
337 [C(DTLB)] = {
338         [C(OP_READ)] = {
339                 [C(RESULT_ACCESS)]      = { 0x06, CNTR_EVEN, T },
340                 [C(RESULT_MISS)]        = { 0x06, CNTR_ODD, T },
341         },
342         [C(OP_WRITE)] = {
343                 [C(RESULT_ACCESS)]      = { 0x06, CNTR_EVEN, T },
344                 [C(RESULT_MISS)]        = { 0x06, CNTR_ODD, T },
345         },
346         [C(OP_PREFETCH)] = {
347                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
348                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
349         },
350 },
351 [C(ITLB)] = {
352         [C(OP_READ)] = {
353                 [C(RESULT_ACCESS)]      = { 0x05, CNTR_EVEN, T },
354                 [C(RESULT_MISS)]        = { 0x05, CNTR_ODD, T },
355         },
356         [C(OP_WRITE)] = {
357                 [C(RESULT_ACCESS)]      = { 0x05, CNTR_EVEN, T },
358                 [C(RESULT_MISS)]        = { 0x05, CNTR_ODD, T },
359         },
360         [C(OP_PREFETCH)] = {
361                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
362                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
363         },
364 },
365 [C(BPU)] = {
366         /* Using the same code for *HW_BRANCH* */
367         [C(OP_READ)] = {
368                 [C(RESULT_ACCESS)]      = { 0x02, CNTR_EVEN, T },
369                 [C(RESULT_MISS)]        = { 0x02, CNTR_ODD, T },
370         },
371         [C(OP_WRITE)] = {
372                 [C(RESULT_ACCESS)]      = { 0x02, CNTR_EVEN, T },
373                 [C(RESULT_MISS)]        = { 0x02, CNTR_ODD, T },
374         },
375         [C(OP_PREFETCH)] = {
376                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
377                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
378         },
379 },
380 [C(NODE)] = {
381         [C(OP_READ)] = {
382                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
383                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
384         },
385         [C(OP_WRITE)] = {
386                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
387                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
388         },
389         [C(OP_PREFETCH)] = {
390                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
391                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
392         },
393 },
394 };
395
396 /* 74K core has completely different cache event map. */
397 static const struct mips_perf_event mipsxx74Kcore_cache_map
398                                 [PERF_COUNT_HW_CACHE_MAX]
399                                 [PERF_COUNT_HW_CACHE_OP_MAX]
400                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
401 [C(L1D)] = {
402         /*
403          * Like some other architectures (e.g. ARM), the performance
404          * counters don't differentiate between read and write
405          * accesses/misses, so this isn't strictly correct, but it's the
406          * best we can do. Writes and reads get combined.
407          */
408         [C(OP_READ)] = {
409                 [C(RESULT_ACCESS)]      = { 0x17, CNTR_ODD, T },
410                 [C(RESULT_MISS)]        = { 0x18, CNTR_ODD, T },
411         },
412         [C(OP_WRITE)] = {
413                 [C(RESULT_ACCESS)]      = { 0x17, CNTR_ODD, T },
414                 [C(RESULT_MISS)]        = { 0x18, CNTR_ODD, T },
415         },
416         [C(OP_PREFETCH)] = {
417                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
418                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
419         },
420 },
421 [C(L1I)] = {
422         [C(OP_READ)] = {
423                 [C(RESULT_ACCESS)]      = { 0x06, CNTR_EVEN, T },
424                 [C(RESULT_MISS)]        = { 0x06, CNTR_ODD, T },
425         },
426         [C(OP_WRITE)] = {
427                 [C(RESULT_ACCESS)]      = { 0x06, CNTR_EVEN, T },
428                 [C(RESULT_MISS)]        = { 0x06, CNTR_ODD, T },
429         },
430         [C(OP_PREFETCH)] = {
431                 [C(RESULT_ACCESS)]      = { 0x34, CNTR_EVEN, T },
432                 /*
433                  * Note that MIPS has only "hit" events countable for
434                  * the prefetch operation.
435                  */
436                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
437         },
438 },
439 [C(LL)] = {
440         [C(OP_READ)] = {
441                 [C(RESULT_ACCESS)]      = { 0x1c, CNTR_ODD, P },
442                 [C(RESULT_MISS)]        = { 0x1d, CNTR_EVEN | CNTR_ODD, P },
443         },
444         [C(OP_WRITE)] = {
445                 [C(RESULT_ACCESS)]      = { 0x1c, CNTR_ODD, P },
446                 [C(RESULT_MISS)]        = { 0x1d, CNTR_EVEN | CNTR_ODD, P },
447         },
448         [C(OP_PREFETCH)] = {
449                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
450                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
451         },
452 },
453 [C(DTLB)] = {
454         /* 74K core does not have specific DTLB events. */
455         [C(OP_READ)] = {
456                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
457                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
458         },
459         [C(OP_WRITE)] = {
460                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
461                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
462         },
463         [C(OP_PREFETCH)] = {
464                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
465                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
466         },
467 },
468 [C(ITLB)] = {
469         [C(OP_READ)] = {
470                 [C(RESULT_ACCESS)]      = { 0x04, CNTR_EVEN, T },
471                 [C(RESULT_MISS)]        = { 0x04, CNTR_ODD, T },
472         },
473         [C(OP_WRITE)] = {
474                 [C(RESULT_ACCESS)]      = { 0x04, CNTR_EVEN, T },
475                 [C(RESULT_MISS)]        = { 0x04, CNTR_ODD, T },
476         },
477         [C(OP_PREFETCH)] = {
478                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
479                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
480         },
481 },
482 [C(BPU)] = {
483         /* Using the same code for *HW_BRANCH* */
484         [C(OP_READ)] = {
485                 [C(RESULT_ACCESS)]      = { 0x27, CNTR_EVEN, T },
486                 [C(RESULT_MISS)]        = { 0x27, CNTR_ODD, T },
487         },
488         [C(OP_WRITE)] = {
489                 [C(RESULT_ACCESS)]      = { 0x27, CNTR_EVEN, T },
490                 [C(RESULT_MISS)]        = { 0x27, CNTR_ODD, T },
491         },
492         [C(OP_PREFETCH)] = {
493                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
494                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
495         },
496 },
497 [C(NODE)] = {
498         [C(OP_READ)] = {
499                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
500                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
501         },
502         [C(OP_WRITE)] = {
503                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
504                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
505         },
506         [C(OP_PREFETCH)] = {
507                 [C(RESULT_ACCESS)]      = { UNSUPPORTED_PERF_EVENT_ID },
508                 [C(RESULT_MISS)]        = { UNSUPPORTED_PERF_EVENT_ID },
509         },
510 },
511 };
512
513 #ifdef CONFIG_MIPS_MT_SMP
514 static void
515 check_and_calc_range(struct perf_event *event,
516                         const struct mips_perf_event *pev)
517 {
518         struct hw_perf_event *hwc = &event->hw;
519
520         if (event->cpu >= 0) {
521                 if (pev->range > V) {
522                         /*
523                          * The user selected an event that is processor
524                          * wide, while expecting it to be VPE wide.
525                          */
526                         hwc->config_base |= M_TC_EN_ALL;
527                 } else {
528                         /*
529                          * FIXME: cpu_data[event->cpu].vpe_id reports 0
530                          * for both CPUs.
531                          */
532                         hwc->config_base |= M_PERFCTL_VPEID(event->cpu);
533                         hwc->config_base |= M_TC_EN_VPE;
534                 }
535         } else
536                 hwc->config_base |= M_TC_EN_ALL;
537 }
538 #else
539 static void
540 check_and_calc_range(struct perf_event *event,
541                         const struct mips_perf_event *pev)
542 {
543 }
544 #endif
545
546 static int __hw_perf_event_init(struct perf_event *event)
547 {
548         struct perf_event_attr *attr = &event->attr;
549         struct hw_perf_event *hwc = &event->hw;
550         const struct mips_perf_event *pev;
551         int err;
552
553         /* Returning MIPS event descriptor for generic perf event. */
554         if (PERF_TYPE_HARDWARE == event->attr.type) {
555                 if (event->attr.config >= PERF_COUNT_HW_MAX)
556                         return -EINVAL;
557                 pev = mipspmu_map_general_event(event->attr.config);
558         } else if (PERF_TYPE_HW_CACHE == event->attr.type) {
559                 pev = mipspmu_map_cache_event(event->attr.config);
560         } else if (PERF_TYPE_RAW == event->attr.type) {
561                 /* We are working on the global raw event. */
562                 mutex_lock(&raw_event_mutex);
563                 pev = mipspmu->map_raw_event(event->attr.config);
564         } else {
565                 /* The event type is not (yet) supported. */
566                 return -EOPNOTSUPP;
567         }
568
569         if (IS_ERR(pev)) {
570                 if (PERF_TYPE_RAW == event->attr.type)
571                         mutex_unlock(&raw_event_mutex);
572                 return PTR_ERR(pev);
573         }
574
575         /*
576          * We allow max flexibility on how each individual counter shared
577          * by the single CPU operates (the mode exclusion and the range).
578          */
579         hwc->config_base = M_PERFCTL_INTERRUPT_ENABLE;
580
581         /* Calculate range bits and validate it. */
582         if (num_possible_cpus() > 1)
583                 check_and_calc_range(event, pev);
584
585         hwc->event_base = mipspmu_perf_event_encode(pev);
586         if (PERF_TYPE_RAW == event->attr.type)
587                 mutex_unlock(&raw_event_mutex);
588
589         if (!attr->exclude_user)
590                 hwc->config_base |= M_PERFCTL_USER;
591         if (!attr->exclude_kernel) {
592                 hwc->config_base |= M_PERFCTL_KERNEL;
593                 /* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */
594                 hwc->config_base |= M_PERFCTL_EXL;
595         }
596         if (!attr->exclude_hv)
597                 hwc->config_base |= M_PERFCTL_SUPERVISOR;
598
599         hwc->config_base &= M_PERFCTL_CONFIG_MASK;
600         /*
601          * The event can belong to another cpu. We do not assign a local
602          * counter for it for now.
603          */
604         hwc->idx = -1;
605         hwc->config = 0;
606
607         if (!hwc->sample_period) {
608                 hwc->sample_period  = MAX_PERIOD;
609                 hwc->last_period    = hwc->sample_period;
610                 local64_set(&hwc->period_left, hwc->sample_period);
611         }
612
613         err = 0;
614         if (event->group_leader != event) {
615                 err = validate_group(event);
616                 if (err)
617                         return -EINVAL;
618         }
619
620         event->destroy = hw_perf_event_destroy;
621
622         return err;
623 }
624
625 static void pause_local_counters(void)
626 {
627         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
628         int counters = mipspmu->num_counters;
629         unsigned long flags;
630
631         local_irq_save(flags);
632         switch (counters) {
633         case 4:
634                 cpuc->saved_ctrl[3] = r_c0_perfctrl3();
635                 w_c0_perfctrl3(cpuc->saved_ctrl[3] &
636                         ~M_PERFCTL_COUNT_EVENT_WHENEVER);
637         case 3:
638                 cpuc->saved_ctrl[2] = r_c0_perfctrl2();
639                 w_c0_perfctrl2(cpuc->saved_ctrl[2] &
640                         ~M_PERFCTL_COUNT_EVENT_WHENEVER);
641         case 2:
642                 cpuc->saved_ctrl[1] = r_c0_perfctrl1();
643                 w_c0_perfctrl1(cpuc->saved_ctrl[1] &
644                         ~M_PERFCTL_COUNT_EVENT_WHENEVER);
645         case 1:
646                 cpuc->saved_ctrl[0] = r_c0_perfctrl0();
647                 w_c0_perfctrl0(cpuc->saved_ctrl[0] &
648                         ~M_PERFCTL_COUNT_EVENT_WHENEVER);
649         }
650         local_irq_restore(flags);
651 }
652
653 static void resume_local_counters(void)
654 {
655         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
656         int counters = mipspmu->num_counters;
657         unsigned long flags;
658
659         local_irq_save(flags);
660         switch (counters) {
661         case 4:
662                 w_c0_perfctrl3(cpuc->saved_ctrl[3]);
663         case 3:
664                 w_c0_perfctrl2(cpuc->saved_ctrl[2]);
665         case 2:
666                 w_c0_perfctrl1(cpuc->saved_ctrl[1]);
667         case 1:
668                 w_c0_perfctrl0(cpuc->saved_ctrl[0]);
669         }
670         local_irq_restore(flags);
671 }
672
673 static int mipsxx_pmu_handle_shared_irq(void)
674 {
675         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
676         struct perf_sample_data data;
677         unsigned int counters = mipspmu->num_counters;
678         unsigned int counter;
679         int handled = IRQ_NONE;
680         struct pt_regs *regs;
681
682         if (cpu_has_mips_r2 && !(read_c0_cause() & (1 << 26)))
683                 return handled;
684
685         /*
686          * First we pause the local counters, so that when we are locked
687          * here, the counters are all paused. When it gets locked due to
688          * perf_disable(), the timer interrupt handler will be delayed.
689          *
690          * See also mipsxx_pmu_start().
691          */
692         pause_local_counters();
693 #ifdef CONFIG_MIPS_MT_SMP
694         read_lock(&pmuint_rwlock);
695 #endif
696
697         regs = get_irq_regs();
698
699         perf_sample_data_init(&data, 0);
700
701         switch (counters) {
702 #define HANDLE_COUNTER(n)                                               \
703         case n + 1:                                                     \
704                 if (test_bit(n, cpuc->used_mask)) {                     \
705                         counter = r_c0_perfcntr ## n();                 \
706                         if (counter & M_COUNTER_OVERFLOW) {             \
707                                 w_c0_perfcntr ## n(counter &            \
708                                                 VALID_COUNT);           \
709                                 if (test_and_change_bit(n, cpuc->msbs)) \
710                                         handle_associated_event(cpuc,   \
711                                                 n, &data, regs);        \
712                                 handled = IRQ_HANDLED;                  \
713                         }                                               \
714                 }
715         HANDLE_COUNTER(3)
716         HANDLE_COUNTER(2)
717         HANDLE_COUNTER(1)
718         HANDLE_COUNTER(0)
719         }
720
721         /*
722          * Do all the work for the pending perf events. We can do this
723          * in here because the performance counter interrupt is a regular
724          * interrupt, not NMI.
725          */
726         if (handled == IRQ_HANDLED)
727                 irq_work_run();
728
729 #ifdef CONFIG_MIPS_MT_SMP
730         read_unlock(&pmuint_rwlock);
731 #endif
732         resume_local_counters();
733         return handled;
734 }
735
736 static irqreturn_t
737 mipsxx_pmu_handle_irq(int irq, void *dev)
738 {
739         return mipsxx_pmu_handle_shared_irq();
740 }
741
742 static void mipsxx_pmu_start(void)
743 {
744 #ifdef CONFIG_MIPS_MT_SMP
745         write_unlock(&pmuint_rwlock);
746 #endif
747         resume_local_counters();
748 }
749
750 /*
751  * MIPS performance counters can be per-TC. The control registers can
752  * not be directly accessed across CPUs. Hence if we want to do global
753  * control, we need cross CPU calls. on_each_cpu() can help us, but we
754  * can not make sure this function is called with interrupts enabled. So
755  * here we pause local counters and then grab a rwlock and leave the
756  * counters on other CPUs alone. If any counter interrupt raises while
757  * we own the write lock, simply pause local counters on that CPU and
758  * spin in the handler. Also we know we won't be switched to another
759  * CPU after pausing local counters and before grabbing the lock.
760  */
761 static void mipsxx_pmu_stop(void)
762 {
763         pause_local_counters();
764 #ifdef CONFIG_MIPS_MT_SMP
765         write_lock(&pmuint_rwlock);
766 #endif
767 }
768
769 static int
770 mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc,
771                         struct hw_perf_event *hwc)
772 {
773         int i;
774
775         /*
776          * We only need to care the counter mask. The range has been
777          * checked definitely.
778          */
779         unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff;
780
781         for (i = mipspmu->num_counters - 1; i >= 0; i--) {
782                 /*
783                  * Note that some MIPS perf events can be counted by both
784                  * even and odd counters, wheresas many other are only by
785                  * even _or_ odd counters. This introduces an issue that
786                  * when the former kind of event takes the counter the
787                  * latter kind of event wants to use, then the "counter
788                  * allocation" for the latter event will fail. In fact if
789                  * they can be dynamically swapped, they both feel happy.
790                  * But here we leave this issue alone for now.
791                  */
792                 if (test_bit(i, &cntr_mask) &&
793                         !test_and_set_bit(i, cpuc->used_mask))
794                         return i;
795         }
796
797         return -EAGAIN;
798 }
799
800 static void
801 mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx)
802 {
803         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
804         unsigned long flags;
805
806         WARN_ON(idx < 0 || idx >= mipspmu->num_counters);
807
808         local_irq_save(flags);
809         cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) |
810                 (evt->config_base & M_PERFCTL_CONFIG_MASK) |
811                 /* Make sure interrupt enabled. */
812                 M_PERFCTL_INTERRUPT_ENABLE;
813         /*
814          * We do not actually let the counter run. Leave it until start().
815          */
816         local_irq_restore(flags);
817 }
818
819 static void
820 mipsxx_pmu_disable_event(int idx)
821 {
822         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
823         unsigned long flags;
824
825         WARN_ON(idx < 0 || idx >= mipspmu->num_counters);
826
827         local_irq_save(flags);
828         cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) &
829                 ~M_PERFCTL_COUNT_EVENT_WHENEVER;
830         mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]);
831         local_irq_restore(flags);
832 }
833
834 /* 24K */
835 #define IS_UNSUPPORTED_24K_EVENT(r, b)                                  \
836         ((b) == 12 || (r) == 151 || (r) == 152 || (b) == 26 ||          \
837          (b) == 27 || (r) == 28 || (r) == 158 || (b) == 31 ||           \
838          (b) == 32 || (b) == 34 || (b) == 36 || (r) == 168 ||           \
839          (r) == 172 || (b) == 47 || ((b) >= 56 && (b) <= 63) ||         \
840          ((b) >= 68 && (b) <= 127))
841 #define IS_BOTH_COUNTERS_24K_EVENT(b)                                   \
842         ((b) == 0 || (b) == 1 || (b) == 11)
843
844 /* 34K */
845 #define IS_UNSUPPORTED_34K_EVENT(r, b)                                  \
846         ((b) == 12 || (r) == 27 || (r) == 158 || (b) == 36 ||           \
847          (b) == 38 || (r) == 175 || ((b) >= 56 && (b) <= 63) ||         \
848          ((b) >= 68 && (b) <= 127))
849 #define IS_BOTH_COUNTERS_34K_EVENT(b)                                   \
850         ((b) == 0 || (b) == 1 || (b) == 11)
851 #ifdef CONFIG_MIPS_MT_SMP
852 #define IS_RANGE_P_34K_EVENT(r, b)                                      \
853         ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||             \
854          (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 ||           \
855          (r) == 176 || ((b) >= 50 && (b) <= 55) ||                      \
856          ((b) >= 64 && (b) <= 67))
857 #define IS_RANGE_V_34K_EVENT(r) ((r) == 47)
858 #endif
859
860 /* 74K */
861 #define IS_UNSUPPORTED_74K_EVENT(r, b)                                  \
862         ((r) == 5 || ((r) >= 135 && (r) <= 137) ||                      \
863          ((b) >= 10 && (b) <= 12) || (b) == 22 || (b) == 27 ||          \
864          (b) == 33 || (b) == 34 || ((b) >= 47 && (b) <= 49) ||          \
865          (r) == 178 || (b) == 55 || (b) == 57 || (b) == 60 ||           \
866          (b) == 61 || (r) == 62 || (r) == 191 ||                        \
867          ((b) >= 64 && (b) <= 127))
868 #define IS_BOTH_COUNTERS_74K_EVENT(b)                                   \
869         ((b) == 0 || (b) == 1)
870
871 /* 1004K */
872 #define IS_UNSUPPORTED_1004K_EVENT(r, b)                                \
873         ((b) == 12 || (r) == 27 || (r) == 158 || (b) == 38 ||           \
874          (r) == 175 || (b) == 63 || ((b) >= 68 && (b) <= 127))
875 #define IS_BOTH_COUNTERS_1004K_EVENT(b)                                 \
876         ((b) == 0 || (b) == 1 || (b) == 11)
877 #ifdef CONFIG_MIPS_MT_SMP
878 #define IS_RANGE_P_1004K_EVENT(r, b)                                    \
879         ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||             \
880          (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 ||            \
881          (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) ||        \
882          (r) == 188 || (b) == 61 || (b) == 62 ||                        \
883          ((b) >= 64 && (b) <= 67))
884 #define IS_RANGE_V_1004K_EVENT(r)       ((r) == 47)
885 #endif
886
887 /*
888  * User can use 0-255 raw events, where 0-127 for the events of even
889  * counters, and 128-255 for odd counters. Note that bit 7 is used to
890  * indicate the parity. So, for example, when user wants to take the
891  * Event Num of 15 for odd counters (by referring to the user manual),
892  * then 128 needs to be added to 15 as the input for the event config,
893  * i.e., 143 (0x8F) to be used.
894  */
895 static const struct mips_perf_event *
896 mipsxx_pmu_map_raw_event(u64 config)
897 {
898         unsigned int raw_id = config & 0xff;
899         unsigned int base_id = raw_id & 0x7f;
900
901         switch (current_cpu_type()) {
902         case CPU_24K:
903                 if (IS_UNSUPPORTED_24K_EVENT(raw_id, base_id))
904                         return ERR_PTR(-EOPNOTSUPP);
905                 raw_event.event_id = base_id;
906                 if (IS_BOTH_COUNTERS_24K_EVENT(base_id))
907                         raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
908                 else
909                         raw_event.cntr_mask =
910                                 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
911 #ifdef CONFIG_MIPS_MT_SMP
912                 /*
913                  * This is actually doing nothing. Non-multithreading
914                  * CPUs will not check and calculate the range.
915                  */
916                 raw_event.range = P;
917 #endif
918                 break;
919         case CPU_34K:
920                 if (IS_UNSUPPORTED_34K_EVENT(raw_id, base_id))
921                         return ERR_PTR(-EOPNOTSUPP);
922                 raw_event.event_id = base_id;
923                 if (IS_BOTH_COUNTERS_34K_EVENT(base_id))
924                         raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
925                 else
926                         raw_event.cntr_mask =
927                                 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
928 #ifdef CONFIG_MIPS_MT_SMP
929                 if (IS_RANGE_P_34K_EVENT(raw_id, base_id))
930                         raw_event.range = P;
931                 else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id)))
932                         raw_event.range = V;
933                 else
934                         raw_event.range = T;
935 #endif
936                 break;
937         case CPU_74K:
938                 if (IS_UNSUPPORTED_74K_EVENT(raw_id, base_id))
939                         return ERR_PTR(-EOPNOTSUPP);
940                 raw_event.event_id = base_id;
941                 if (IS_BOTH_COUNTERS_74K_EVENT(base_id))
942                         raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
943                 else
944                         raw_event.cntr_mask =
945                                 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
946 #ifdef CONFIG_MIPS_MT_SMP
947                 raw_event.range = P;
948 #endif
949                 break;
950         case CPU_1004K:
951                 if (IS_UNSUPPORTED_1004K_EVENT(raw_id, base_id))
952                         return ERR_PTR(-EOPNOTSUPP);
953                 raw_event.event_id = base_id;
954                 if (IS_BOTH_COUNTERS_1004K_EVENT(base_id))
955                         raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
956                 else
957                         raw_event.cntr_mask =
958                                 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
959 #ifdef CONFIG_MIPS_MT_SMP
960                 if (IS_RANGE_P_1004K_EVENT(raw_id, base_id))
961                         raw_event.range = P;
962                 else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id)))
963                         raw_event.range = V;
964                 else
965                         raw_event.range = T;
966 #endif
967                 break;
968         }
969
970         return &raw_event;
971 }
972
973 static struct mips_pmu mipsxxcore_pmu = {
974         .handle_irq = mipsxx_pmu_handle_irq,
975         .handle_shared_irq = mipsxx_pmu_handle_shared_irq,
976         .start = mipsxx_pmu_start,
977         .stop = mipsxx_pmu_stop,
978         .alloc_counter = mipsxx_pmu_alloc_counter,
979         .read_counter = mipsxx_pmu_read_counter,
980         .write_counter = mipsxx_pmu_write_counter,
981         .enable_event = mipsxx_pmu_enable_event,
982         .disable_event = mipsxx_pmu_disable_event,
983         .map_raw_event = mipsxx_pmu_map_raw_event,
984         .general_event_map = &mipsxxcore_event_map,
985         .cache_event_map = &mipsxxcore_cache_map,
986 };
987
988 static struct mips_pmu mipsxx74Kcore_pmu = {
989         .handle_irq = mipsxx_pmu_handle_irq,
990         .handle_shared_irq = mipsxx_pmu_handle_shared_irq,
991         .start = mipsxx_pmu_start,
992         .stop = mipsxx_pmu_stop,
993         .alloc_counter = mipsxx_pmu_alloc_counter,
994         .read_counter = mipsxx_pmu_read_counter,
995         .write_counter = mipsxx_pmu_write_counter,
996         .enable_event = mipsxx_pmu_enable_event,
997         .disable_event = mipsxx_pmu_disable_event,
998         .map_raw_event = mipsxx_pmu_map_raw_event,
999         .general_event_map = &mipsxx74Kcore_event_map,
1000         .cache_event_map = &mipsxx74Kcore_cache_map,
1001 };
1002
1003 static int __init
1004 init_hw_perf_events(void)
1005 {
1006         int counters, irq;
1007
1008         pr_info("Performance counters: ");
1009
1010         counters = n_counters();
1011         if (counters == 0) {
1012                 pr_cont("No available PMU.\n");
1013                 return -ENODEV;
1014         }
1015
1016 #ifdef CONFIG_MIPS_MT_SMP
1017         cpu_has_mipsmt_pertccounters = read_c0_config7() & (1<<19);
1018         if (!cpu_has_mipsmt_pertccounters)
1019                 counters = counters_total_to_per_cpu(counters);
1020 #endif
1021
1022 #ifdef MSC01E_INT_BASE
1023         if (cpu_has_veic) {
1024                 /*
1025                  * Using platform specific interrupt controller defines.
1026                  */
1027                 irq = MSC01E_INT_BASE + MSC01E_INT_PERFCTR;
1028         } else {
1029 #endif
1030                 if (cp0_perfcount_irq >= 0)
1031                         irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq;
1032                 else
1033                         irq = -1;
1034 #ifdef MSC01E_INT_BASE
1035         }
1036 #endif
1037
1038         on_each_cpu(reset_counters, (void *)(long)counters, 1);
1039
1040         switch (current_cpu_type()) {
1041         case CPU_24K:
1042                 mipsxxcore_pmu.name = "mips/24K";
1043                 mipsxxcore_pmu.num_counters = counters;
1044                 mipsxxcore_pmu.irq = irq;
1045                 mipspmu = &mipsxxcore_pmu;
1046                 break;
1047         case CPU_34K:
1048                 mipsxxcore_pmu.name = "mips/34K";
1049                 mipsxxcore_pmu.num_counters = counters;
1050                 mipsxxcore_pmu.irq = irq;
1051                 mipspmu = &mipsxxcore_pmu;
1052                 break;
1053         case CPU_74K:
1054                 mipsxx74Kcore_pmu.name = "mips/74K";
1055                 mipsxx74Kcore_pmu.num_counters = counters;
1056                 mipsxx74Kcore_pmu.irq = irq;
1057                 mipspmu = &mipsxx74Kcore_pmu;
1058                 break;
1059         case CPU_1004K:
1060                 mipsxxcore_pmu.name = "mips/1004K";
1061                 mipsxxcore_pmu.num_counters = counters;
1062                 mipsxxcore_pmu.irq = irq;
1063                 mipspmu = &mipsxxcore_pmu;
1064                 break;
1065         default:
1066                 pr_cont("Either hardware does not support performance "
1067                         "counters, or not yet implemented.\n");
1068                 return -ENODEV;
1069         }
1070
1071         if (mipspmu)
1072                 pr_cont("%s PMU enabled, %d counters available to each "
1073                         "CPU, irq %d%s\n", mipspmu->name, counters, irq,
1074                         irq < 0 ? " (share with timer interrupt)" : "");
1075
1076         perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1077
1078         return 0;
1079 }
1080 early_initcall(init_hw_perf_events);
1081
1082 #endif /* defined(CONFIG_CPU_MIPS32)... */