Loading...
1/*
2 * ARMv5 [xscale] Performance counter handling code.
3 *
4 * Copyright (C) 2010, ARM Ltd., Will Deacon <will.deacon@arm.com>
5 *
6 * Based on the previous xscale OProfile code.
7 *
8 * There are two variants of the xscale PMU that we support:
9 * - xscale1pmu: 2 event counters and a cycle counter
10 * - xscale2pmu: 4 event counters and a cycle counter
11 * The two variants share event definitions, but have different
12 * PMU structures.
13 */
14
15#ifdef CONFIG_CPU_XSCALE
16enum xscale_perf_types {
17 XSCALE_PERFCTR_ICACHE_MISS = 0x00,
18 XSCALE_PERFCTR_ICACHE_NO_DELIVER = 0x01,
19 XSCALE_PERFCTR_DATA_STALL = 0x02,
20 XSCALE_PERFCTR_ITLB_MISS = 0x03,
21 XSCALE_PERFCTR_DTLB_MISS = 0x04,
22 XSCALE_PERFCTR_BRANCH = 0x05,
23 XSCALE_PERFCTR_BRANCH_MISS = 0x06,
24 XSCALE_PERFCTR_INSTRUCTION = 0x07,
25 XSCALE_PERFCTR_DCACHE_FULL_STALL = 0x08,
26 XSCALE_PERFCTR_DCACHE_FULL_STALL_CONTIG = 0x09,
27 XSCALE_PERFCTR_DCACHE_ACCESS = 0x0A,
28 XSCALE_PERFCTR_DCACHE_MISS = 0x0B,
29 XSCALE_PERFCTR_DCACHE_WRITE_BACK = 0x0C,
30 XSCALE_PERFCTR_PC_CHANGED = 0x0D,
31 XSCALE_PERFCTR_BCU_REQUEST = 0x10,
32 XSCALE_PERFCTR_BCU_FULL = 0x11,
33 XSCALE_PERFCTR_BCU_DRAIN = 0x12,
34 XSCALE_PERFCTR_BCU_ECC_NO_ELOG = 0x14,
35 XSCALE_PERFCTR_BCU_1_BIT_ERR = 0x15,
36 XSCALE_PERFCTR_RMW = 0x16,
37 /* XSCALE_PERFCTR_CCNT is not hardware defined */
38 XSCALE_PERFCTR_CCNT = 0xFE,
39 XSCALE_PERFCTR_UNUSED = 0xFF,
40};
41
42enum xscale_counters {
43 XSCALE_CYCLE_COUNTER = 1,
44 XSCALE_COUNTER0,
45 XSCALE_COUNTER1,
46 XSCALE_COUNTER2,
47 XSCALE_COUNTER3,
48};
49
50static const unsigned xscale_perf_map[PERF_COUNT_HW_MAX] = {
51 [PERF_COUNT_HW_CPU_CYCLES] = XSCALE_PERFCTR_CCNT,
52 [PERF_COUNT_HW_INSTRUCTIONS] = XSCALE_PERFCTR_INSTRUCTION,
53 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
54 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
55 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = XSCALE_PERFCTR_BRANCH,
56 [PERF_COUNT_HW_BRANCH_MISSES] = XSCALE_PERFCTR_BRANCH_MISS,
57 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
58};
59
60static const unsigned xscale_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
61 [PERF_COUNT_HW_CACHE_OP_MAX]
62 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
63 [C(L1D)] = {
64 [C(OP_READ)] = {
65 [C(RESULT_ACCESS)] = XSCALE_PERFCTR_DCACHE_ACCESS,
66 [C(RESULT_MISS)] = XSCALE_PERFCTR_DCACHE_MISS,
67 },
68 [C(OP_WRITE)] = {
69 [C(RESULT_ACCESS)] = XSCALE_PERFCTR_DCACHE_ACCESS,
70 [C(RESULT_MISS)] = XSCALE_PERFCTR_DCACHE_MISS,
71 },
72 [C(OP_PREFETCH)] = {
73 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
74 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
75 },
76 },
77 [C(L1I)] = {
78 [C(OP_READ)] = {
79 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
80 [C(RESULT_MISS)] = XSCALE_PERFCTR_ICACHE_MISS,
81 },
82 [C(OP_WRITE)] = {
83 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
84 [C(RESULT_MISS)] = XSCALE_PERFCTR_ICACHE_MISS,
85 },
86 [C(OP_PREFETCH)] = {
87 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
88 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
89 },
90 },
91 [C(LL)] = {
92 [C(OP_READ)] = {
93 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
94 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
95 },
96 [C(OP_WRITE)] = {
97 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
98 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
99 },
100 [C(OP_PREFETCH)] = {
101 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
102 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
103 },
104 },
105 [C(DTLB)] = {
106 [C(OP_READ)] = {
107 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
108 [C(RESULT_MISS)] = XSCALE_PERFCTR_DTLB_MISS,
109 },
110 [C(OP_WRITE)] = {
111 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
112 [C(RESULT_MISS)] = XSCALE_PERFCTR_DTLB_MISS,
113 },
114 [C(OP_PREFETCH)] = {
115 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
116 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
117 },
118 },
119 [C(ITLB)] = {
120 [C(OP_READ)] = {
121 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
122 [C(RESULT_MISS)] = XSCALE_PERFCTR_ITLB_MISS,
123 },
124 [C(OP_WRITE)] = {
125 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
126 [C(RESULT_MISS)] = XSCALE_PERFCTR_ITLB_MISS,
127 },
128 [C(OP_PREFETCH)] = {
129 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
130 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
131 },
132 },
133 [C(BPU)] = {
134 [C(OP_READ)] = {
135 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
136 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
137 },
138 [C(OP_WRITE)] = {
139 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
140 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
141 },
142 [C(OP_PREFETCH)] = {
143 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
144 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
145 },
146 },
147 [C(NODE)] = {
148 [C(OP_READ)] = {
149 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
150 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
151 },
152 [C(OP_WRITE)] = {
153 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
154 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
155 },
156 [C(OP_PREFETCH)] = {
157 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
158 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
159 },
160 },
161};
162
163#define XSCALE_PMU_ENABLE 0x001
164#define XSCALE_PMN_RESET 0x002
165#define XSCALE_CCNT_RESET 0x004
166#define XSCALE_PMU_RESET (CCNT_RESET | PMN_RESET)
167#define XSCALE_PMU_CNT64 0x008
168
169#define XSCALE1_OVERFLOWED_MASK 0x700
170#define XSCALE1_CCOUNT_OVERFLOW 0x400
171#define XSCALE1_COUNT0_OVERFLOW 0x100
172#define XSCALE1_COUNT1_OVERFLOW 0x200
173#define XSCALE1_CCOUNT_INT_EN 0x040
174#define XSCALE1_COUNT0_INT_EN 0x010
175#define XSCALE1_COUNT1_INT_EN 0x020
176#define XSCALE1_COUNT0_EVT_SHFT 12
177#define XSCALE1_COUNT0_EVT_MASK (0xff << XSCALE1_COUNT0_EVT_SHFT)
178#define XSCALE1_COUNT1_EVT_SHFT 20
179#define XSCALE1_COUNT1_EVT_MASK (0xff << XSCALE1_COUNT1_EVT_SHFT)
180
181static inline u32
182xscale1pmu_read_pmnc(void)
183{
184 u32 val;
185 asm volatile("mrc p14, 0, %0, c0, c0, 0" : "=r" (val));
186 return val;
187}
188
189static inline void
190xscale1pmu_write_pmnc(u32 val)
191{
192 /* upper 4bits and 7, 11 are write-as-0 */
193 val &= 0xffff77f;
194 asm volatile("mcr p14, 0, %0, c0, c0, 0" : : "r" (val));
195}
196
197static inline int
198xscale1_pmnc_counter_has_overflowed(unsigned long pmnc,
199 enum xscale_counters counter)
200{
201 int ret = 0;
202
203 switch (counter) {
204 case XSCALE_CYCLE_COUNTER:
205 ret = pmnc & XSCALE1_CCOUNT_OVERFLOW;
206 break;
207 case XSCALE_COUNTER0:
208 ret = pmnc & XSCALE1_COUNT0_OVERFLOW;
209 break;
210 case XSCALE_COUNTER1:
211 ret = pmnc & XSCALE1_COUNT1_OVERFLOW;
212 break;
213 default:
214 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
215 }
216
217 return ret;
218}
219
220static irqreturn_t
221xscale1pmu_handle_irq(int irq_num, void *dev)
222{
223 unsigned long pmnc;
224 struct perf_sample_data data;
225 struct cpu_hw_events *cpuc;
226 struct pt_regs *regs;
227 int idx;
228
229 /*
230 * NOTE: there's an A stepping erratum that states if an overflow
231 * bit already exists and another occurs, the previous
232 * Overflow bit gets cleared. There's no workaround.
233 * Fixed in B stepping or later.
234 */
235 pmnc = xscale1pmu_read_pmnc();
236
237 /*
238 * Write the value back to clear the overflow flags. Overflow
239 * flags remain in pmnc for use below. We also disable the PMU
240 * while we process the interrupt.
241 */
242 xscale1pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
243
244 if (!(pmnc & XSCALE1_OVERFLOWED_MASK))
245 return IRQ_NONE;
246
247 regs = get_irq_regs();
248
249 perf_sample_data_init(&data, 0);
250
251 cpuc = &__get_cpu_var(cpu_hw_events);
252 for (idx = 0; idx <= armpmu->num_events; ++idx) {
253 struct perf_event *event = cpuc->events[idx];
254 struct hw_perf_event *hwc;
255
256 if (!test_bit(idx, cpuc->active_mask))
257 continue;
258
259 if (!xscale1_pmnc_counter_has_overflowed(pmnc, idx))
260 continue;
261
262 hwc = &event->hw;
263 armpmu_event_update(event, hwc, idx, 1);
264 data.period = event->hw.last_period;
265 if (!armpmu_event_set_period(event, hwc, idx))
266 continue;
267
268 if (perf_event_overflow(event, &data, regs))
269 armpmu->disable(hwc, idx);
270 }
271
272 irq_work_run();
273
274 /*
275 * Re-enable the PMU.
276 */
277 pmnc = xscale1pmu_read_pmnc() | XSCALE_PMU_ENABLE;
278 xscale1pmu_write_pmnc(pmnc);
279
280 return IRQ_HANDLED;
281}
282
283static void
284xscale1pmu_enable_event(struct hw_perf_event *hwc, int idx)
285{
286 unsigned long val, mask, evt, flags;
287
288 switch (idx) {
289 case XSCALE_CYCLE_COUNTER:
290 mask = 0;
291 evt = XSCALE1_CCOUNT_INT_EN;
292 break;
293 case XSCALE_COUNTER0:
294 mask = XSCALE1_COUNT0_EVT_MASK;
295 evt = (hwc->config_base << XSCALE1_COUNT0_EVT_SHFT) |
296 XSCALE1_COUNT0_INT_EN;
297 break;
298 case XSCALE_COUNTER1:
299 mask = XSCALE1_COUNT1_EVT_MASK;
300 evt = (hwc->config_base << XSCALE1_COUNT1_EVT_SHFT) |
301 XSCALE1_COUNT1_INT_EN;
302 break;
303 default:
304 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
305 return;
306 }
307
308 raw_spin_lock_irqsave(&pmu_lock, flags);
309 val = xscale1pmu_read_pmnc();
310 val &= ~mask;
311 val |= evt;
312 xscale1pmu_write_pmnc(val);
313 raw_spin_unlock_irqrestore(&pmu_lock, flags);
314}
315
316static void
317xscale1pmu_disable_event(struct hw_perf_event *hwc, int idx)
318{
319 unsigned long val, mask, evt, flags;
320
321 switch (idx) {
322 case XSCALE_CYCLE_COUNTER:
323 mask = XSCALE1_CCOUNT_INT_EN;
324 evt = 0;
325 break;
326 case XSCALE_COUNTER0:
327 mask = XSCALE1_COUNT0_INT_EN | XSCALE1_COUNT0_EVT_MASK;
328 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT0_EVT_SHFT;
329 break;
330 case XSCALE_COUNTER1:
331 mask = XSCALE1_COUNT1_INT_EN | XSCALE1_COUNT1_EVT_MASK;
332 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT1_EVT_SHFT;
333 break;
334 default:
335 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
336 return;
337 }
338
339 raw_spin_lock_irqsave(&pmu_lock, flags);
340 val = xscale1pmu_read_pmnc();
341 val &= ~mask;
342 val |= evt;
343 xscale1pmu_write_pmnc(val);
344 raw_spin_unlock_irqrestore(&pmu_lock, flags);
345}
346
347static int
348xscale1pmu_get_event_idx(struct cpu_hw_events *cpuc,
349 struct hw_perf_event *event)
350{
351 if (XSCALE_PERFCTR_CCNT == event->config_base) {
352 if (test_and_set_bit(XSCALE_CYCLE_COUNTER, cpuc->used_mask))
353 return -EAGAIN;
354
355 return XSCALE_CYCLE_COUNTER;
356 } else {
357 if (!test_and_set_bit(XSCALE_COUNTER1, cpuc->used_mask))
358 return XSCALE_COUNTER1;
359
360 if (!test_and_set_bit(XSCALE_COUNTER0, cpuc->used_mask))
361 return XSCALE_COUNTER0;
362
363 return -EAGAIN;
364 }
365}
366
367static void
368xscale1pmu_start(void)
369{
370 unsigned long flags, val;
371
372 raw_spin_lock_irqsave(&pmu_lock, flags);
373 val = xscale1pmu_read_pmnc();
374 val |= XSCALE_PMU_ENABLE;
375 xscale1pmu_write_pmnc(val);
376 raw_spin_unlock_irqrestore(&pmu_lock, flags);
377}
378
379static void
380xscale1pmu_stop(void)
381{
382 unsigned long flags, val;
383
384 raw_spin_lock_irqsave(&pmu_lock, flags);
385 val = xscale1pmu_read_pmnc();
386 val &= ~XSCALE_PMU_ENABLE;
387 xscale1pmu_write_pmnc(val);
388 raw_spin_unlock_irqrestore(&pmu_lock, flags);
389}
390
391static inline u32
392xscale1pmu_read_counter(int counter)
393{
394 u32 val = 0;
395
396 switch (counter) {
397 case XSCALE_CYCLE_COUNTER:
398 asm volatile("mrc p14, 0, %0, c1, c0, 0" : "=r" (val));
399 break;
400 case XSCALE_COUNTER0:
401 asm volatile("mrc p14, 0, %0, c2, c0, 0" : "=r" (val));
402 break;
403 case XSCALE_COUNTER1:
404 asm volatile("mrc p14, 0, %0, c3, c0, 0" : "=r" (val));
405 break;
406 }
407
408 return val;
409}
410
411static inline void
412xscale1pmu_write_counter(int counter, u32 val)
413{
414 switch (counter) {
415 case XSCALE_CYCLE_COUNTER:
416 asm volatile("mcr p14, 0, %0, c1, c0, 0" : : "r" (val));
417 break;
418 case XSCALE_COUNTER0:
419 asm volatile("mcr p14, 0, %0, c2, c0, 0" : : "r" (val));
420 break;
421 case XSCALE_COUNTER1:
422 asm volatile("mcr p14, 0, %0, c3, c0, 0" : : "r" (val));
423 break;
424 }
425}
426
427static const struct arm_pmu xscale1pmu = {
428 .id = ARM_PERF_PMU_ID_XSCALE1,
429 .name = "xscale1",
430 .handle_irq = xscale1pmu_handle_irq,
431 .enable = xscale1pmu_enable_event,
432 .disable = xscale1pmu_disable_event,
433 .read_counter = xscale1pmu_read_counter,
434 .write_counter = xscale1pmu_write_counter,
435 .get_event_idx = xscale1pmu_get_event_idx,
436 .start = xscale1pmu_start,
437 .stop = xscale1pmu_stop,
438 .cache_map = &xscale_perf_cache_map,
439 .event_map = &xscale_perf_map,
440 .raw_event_mask = 0xFF,
441 .num_events = 3,
442 .max_period = (1LLU << 32) - 1,
443};
444
445static const struct arm_pmu *__init xscale1pmu_init(void)
446{
447 return &xscale1pmu;
448}
449
450#define XSCALE2_OVERFLOWED_MASK 0x01f
451#define XSCALE2_CCOUNT_OVERFLOW 0x001
452#define XSCALE2_COUNT0_OVERFLOW 0x002
453#define XSCALE2_COUNT1_OVERFLOW 0x004
454#define XSCALE2_COUNT2_OVERFLOW 0x008
455#define XSCALE2_COUNT3_OVERFLOW 0x010
456#define XSCALE2_CCOUNT_INT_EN 0x001
457#define XSCALE2_COUNT0_INT_EN 0x002
458#define XSCALE2_COUNT1_INT_EN 0x004
459#define XSCALE2_COUNT2_INT_EN 0x008
460#define XSCALE2_COUNT3_INT_EN 0x010
461#define XSCALE2_COUNT0_EVT_SHFT 0
462#define XSCALE2_COUNT0_EVT_MASK (0xff << XSCALE2_COUNT0_EVT_SHFT)
463#define XSCALE2_COUNT1_EVT_SHFT 8
464#define XSCALE2_COUNT1_EVT_MASK (0xff << XSCALE2_COUNT1_EVT_SHFT)
465#define XSCALE2_COUNT2_EVT_SHFT 16
466#define XSCALE2_COUNT2_EVT_MASK (0xff << XSCALE2_COUNT2_EVT_SHFT)
467#define XSCALE2_COUNT3_EVT_SHFT 24
468#define XSCALE2_COUNT3_EVT_MASK (0xff << XSCALE2_COUNT3_EVT_SHFT)
469
470static inline u32
471xscale2pmu_read_pmnc(void)
472{
473 u32 val;
474 asm volatile("mrc p14, 0, %0, c0, c1, 0" : "=r" (val));
475 /* bits 1-2 and 4-23 are read-unpredictable */
476 return val & 0xff000009;
477}
478
479static inline void
480xscale2pmu_write_pmnc(u32 val)
481{
482 /* bits 4-23 are write-as-0, 24-31 are write ignored */
483 val &= 0xf;
484 asm volatile("mcr p14, 0, %0, c0, c1, 0" : : "r" (val));
485}
486
487static inline u32
488xscale2pmu_read_overflow_flags(void)
489{
490 u32 val;
491 asm volatile("mrc p14, 0, %0, c5, c1, 0" : "=r" (val));
492 return val;
493}
494
495static inline void
496xscale2pmu_write_overflow_flags(u32 val)
497{
498 asm volatile("mcr p14, 0, %0, c5, c1, 0" : : "r" (val));
499}
500
501static inline u32
502xscale2pmu_read_event_select(void)
503{
504 u32 val;
505 asm volatile("mrc p14, 0, %0, c8, c1, 0" : "=r" (val));
506 return val;
507}
508
509static inline void
510xscale2pmu_write_event_select(u32 val)
511{
512 asm volatile("mcr p14, 0, %0, c8, c1, 0" : : "r"(val));
513}
514
515static inline u32
516xscale2pmu_read_int_enable(void)
517{
518 u32 val;
519 asm volatile("mrc p14, 0, %0, c4, c1, 0" : "=r" (val));
520 return val;
521}
522
523static void
524xscale2pmu_write_int_enable(u32 val)
525{
526 asm volatile("mcr p14, 0, %0, c4, c1, 0" : : "r" (val));
527}
528
529static inline int
530xscale2_pmnc_counter_has_overflowed(unsigned long of_flags,
531 enum xscale_counters counter)
532{
533 int ret = 0;
534
535 switch (counter) {
536 case XSCALE_CYCLE_COUNTER:
537 ret = of_flags & XSCALE2_CCOUNT_OVERFLOW;
538 break;
539 case XSCALE_COUNTER0:
540 ret = of_flags & XSCALE2_COUNT0_OVERFLOW;
541 break;
542 case XSCALE_COUNTER1:
543 ret = of_flags & XSCALE2_COUNT1_OVERFLOW;
544 break;
545 case XSCALE_COUNTER2:
546 ret = of_flags & XSCALE2_COUNT2_OVERFLOW;
547 break;
548 case XSCALE_COUNTER3:
549 ret = of_flags & XSCALE2_COUNT3_OVERFLOW;
550 break;
551 default:
552 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
553 }
554
555 return ret;
556}
557
558static irqreturn_t
559xscale2pmu_handle_irq(int irq_num, void *dev)
560{
561 unsigned long pmnc, of_flags;
562 struct perf_sample_data data;
563 struct cpu_hw_events *cpuc;
564 struct pt_regs *regs;
565 int idx;
566
567 /* Disable the PMU. */
568 pmnc = xscale2pmu_read_pmnc();
569 xscale2pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
570
571 /* Check the overflow flag register. */
572 of_flags = xscale2pmu_read_overflow_flags();
573 if (!(of_flags & XSCALE2_OVERFLOWED_MASK))
574 return IRQ_NONE;
575
576 /* Clear the overflow bits. */
577 xscale2pmu_write_overflow_flags(of_flags);
578
579 regs = get_irq_regs();
580
581 perf_sample_data_init(&data, 0);
582
583 cpuc = &__get_cpu_var(cpu_hw_events);
584 for (idx = 0; idx <= armpmu->num_events; ++idx) {
585 struct perf_event *event = cpuc->events[idx];
586 struct hw_perf_event *hwc;
587
588 if (!test_bit(idx, cpuc->active_mask))
589 continue;
590
591 if (!xscale2_pmnc_counter_has_overflowed(pmnc, idx))
592 continue;
593
594 hwc = &event->hw;
595 armpmu_event_update(event, hwc, idx, 1);
596 data.period = event->hw.last_period;
597 if (!armpmu_event_set_period(event, hwc, idx))
598 continue;
599
600 if (perf_event_overflow(event, &data, regs))
601 armpmu->disable(hwc, idx);
602 }
603
604 irq_work_run();
605
606 /*
607 * Re-enable the PMU.
608 */
609 pmnc = xscale2pmu_read_pmnc() | XSCALE_PMU_ENABLE;
610 xscale2pmu_write_pmnc(pmnc);
611
612 return IRQ_HANDLED;
613}
614
615static void
616xscale2pmu_enable_event(struct hw_perf_event *hwc, int idx)
617{
618 unsigned long flags, ien, evtsel;
619
620 ien = xscale2pmu_read_int_enable();
621 evtsel = xscale2pmu_read_event_select();
622
623 switch (idx) {
624 case XSCALE_CYCLE_COUNTER:
625 ien |= XSCALE2_CCOUNT_INT_EN;
626 break;
627 case XSCALE_COUNTER0:
628 ien |= XSCALE2_COUNT0_INT_EN;
629 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
630 evtsel |= hwc->config_base << XSCALE2_COUNT0_EVT_SHFT;
631 break;
632 case XSCALE_COUNTER1:
633 ien |= XSCALE2_COUNT1_INT_EN;
634 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
635 evtsel |= hwc->config_base << XSCALE2_COUNT1_EVT_SHFT;
636 break;
637 case XSCALE_COUNTER2:
638 ien |= XSCALE2_COUNT2_INT_EN;
639 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
640 evtsel |= hwc->config_base << XSCALE2_COUNT2_EVT_SHFT;
641 break;
642 case XSCALE_COUNTER3:
643 ien |= XSCALE2_COUNT3_INT_EN;
644 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
645 evtsel |= hwc->config_base << XSCALE2_COUNT3_EVT_SHFT;
646 break;
647 default:
648 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
649 return;
650 }
651
652 raw_spin_lock_irqsave(&pmu_lock, flags);
653 xscale2pmu_write_event_select(evtsel);
654 xscale2pmu_write_int_enable(ien);
655 raw_spin_unlock_irqrestore(&pmu_lock, flags);
656}
657
658static void
659xscale2pmu_disable_event(struct hw_perf_event *hwc, int idx)
660{
661 unsigned long flags, ien, evtsel;
662
663 ien = xscale2pmu_read_int_enable();
664 evtsel = xscale2pmu_read_event_select();
665
666 switch (idx) {
667 case XSCALE_CYCLE_COUNTER:
668 ien &= ~XSCALE2_CCOUNT_INT_EN;
669 break;
670 case XSCALE_COUNTER0:
671 ien &= ~XSCALE2_COUNT0_INT_EN;
672 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
673 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT0_EVT_SHFT;
674 break;
675 case XSCALE_COUNTER1:
676 ien &= ~XSCALE2_COUNT1_INT_EN;
677 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
678 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT1_EVT_SHFT;
679 break;
680 case XSCALE_COUNTER2:
681 ien &= ~XSCALE2_COUNT2_INT_EN;
682 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
683 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT2_EVT_SHFT;
684 break;
685 case XSCALE_COUNTER3:
686 ien &= ~XSCALE2_COUNT3_INT_EN;
687 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
688 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT3_EVT_SHFT;
689 break;
690 default:
691 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
692 return;
693 }
694
695 raw_spin_lock_irqsave(&pmu_lock, flags);
696 xscale2pmu_write_event_select(evtsel);
697 xscale2pmu_write_int_enable(ien);
698 raw_spin_unlock_irqrestore(&pmu_lock, flags);
699}
700
701static int
702xscale2pmu_get_event_idx(struct cpu_hw_events *cpuc,
703 struct hw_perf_event *event)
704{
705 int idx = xscale1pmu_get_event_idx(cpuc, event);
706 if (idx >= 0)
707 goto out;
708
709 if (!test_and_set_bit(XSCALE_COUNTER3, cpuc->used_mask))
710 idx = XSCALE_COUNTER3;
711 else if (!test_and_set_bit(XSCALE_COUNTER2, cpuc->used_mask))
712 idx = XSCALE_COUNTER2;
713out:
714 return idx;
715}
716
717static void
718xscale2pmu_start(void)
719{
720 unsigned long flags, val;
721
722 raw_spin_lock_irqsave(&pmu_lock, flags);
723 val = xscale2pmu_read_pmnc() & ~XSCALE_PMU_CNT64;
724 val |= XSCALE_PMU_ENABLE;
725 xscale2pmu_write_pmnc(val);
726 raw_spin_unlock_irqrestore(&pmu_lock, flags);
727}
728
729static void
730xscale2pmu_stop(void)
731{
732 unsigned long flags, val;
733
734 raw_spin_lock_irqsave(&pmu_lock, flags);
735 val = xscale2pmu_read_pmnc();
736 val &= ~XSCALE_PMU_ENABLE;
737 xscale2pmu_write_pmnc(val);
738 raw_spin_unlock_irqrestore(&pmu_lock, flags);
739}
740
741static inline u32
742xscale2pmu_read_counter(int counter)
743{
744 u32 val = 0;
745
746 switch (counter) {
747 case XSCALE_CYCLE_COUNTER:
748 asm volatile("mrc p14, 0, %0, c1, c1, 0" : "=r" (val));
749 break;
750 case XSCALE_COUNTER0:
751 asm volatile("mrc p14, 0, %0, c0, c2, 0" : "=r" (val));
752 break;
753 case XSCALE_COUNTER1:
754 asm volatile("mrc p14, 0, %0, c1, c2, 0" : "=r" (val));
755 break;
756 case XSCALE_COUNTER2:
757 asm volatile("mrc p14, 0, %0, c2, c2, 0" : "=r" (val));
758 break;
759 case XSCALE_COUNTER3:
760 asm volatile("mrc p14, 0, %0, c3, c2, 0" : "=r" (val));
761 break;
762 }
763
764 return val;
765}
766
767static inline void
768xscale2pmu_write_counter(int counter, u32 val)
769{
770 switch (counter) {
771 case XSCALE_CYCLE_COUNTER:
772 asm volatile("mcr p14, 0, %0, c1, c1, 0" : : "r" (val));
773 break;
774 case XSCALE_COUNTER0:
775 asm volatile("mcr p14, 0, %0, c0, c2, 0" : : "r" (val));
776 break;
777 case XSCALE_COUNTER1:
778 asm volatile("mcr p14, 0, %0, c1, c2, 0" : : "r" (val));
779 break;
780 case XSCALE_COUNTER2:
781 asm volatile("mcr p14, 0, %0, c2, c2, 0" : : "r" (val));
782 break;
783 case XSCALE_COUNTER3:
784 asm volatile("mcr p14, 0, %0, c3, c2, 0" : : "r" (val));
785 break;
786 }
787}
788
789static const struct arm_pmu xscale2pmu = {
790 .id = ARM_PERF_PMU_ID_XSCALE2,
791 .name = "xscale2",
792 .handle_irq = xscale2pmu_handle_irq,
793 .enable = xscale2pmu_enable_event,
794 .disable = xscale2pmu_disable_event,
795 .read_counter = xscale2pmu_read_counter,
796 .write_counter = xscale2pmu_write_counter,
797 .get_event_idx = xscale2pmu_get_event_idx,
798 .start = xscale2pmu_start,
799 .stop = xscale2pmu_stop,
800 .cache_map = &xscale_perf_cache_map,
801 .event_map = &xscale_perf_map,
802 .raw_event_mask = 0xFF,
803 .num_events = 5,
804 .max_period = (1LLU << 32) - 1,
805};
806
807static const struct arm_pmu *__init xscale2pmu_init(void)
808{
809 return &xscale2pmu;
810}
811#else
812static const struct arm_pmu *__init xscale1pmu_init(void)
813{
814 return NULL;
815}
816
817static const struct arm_pmu *__init xscale2pmu_init(void)
818{
819 return NULL;
820}
821#endif /* CONFIG_CPU_XSCALE */
1/*
2 * ARMv5 [xscale] Performance counter handling code.
3 *
4 * Copyright (C) 2010, ARM Ltd., Will Deacon <will.deacon@arm.com>
5 *
6 * Based on the previous xscale OProfile code.
7 *
8 * There are two variants of the xscale PMU that we support:
9 * - xscale1pmu: 2 event counters and a cycle counter
10 * - xscale2pmu: 4 event counters and a cycle counter
11 * The two variants share event definitions, but have different
12 * PMU structures.
13 */
14
15#ifdef CONFIG_CPU_XSCALE
16
17#include <asm/cputype.h>
18#include <asm/irq_regs.h>
19
20#include <linux/of.h>
21#include <linux/perf/arm_pmu.h>
22#include <linux/platform_device.h>
23
24enum xscale_perf_types {
25 XSCALE_PERFCTR_ICACHE_MISS = 0x00,
26 XSCALE_PERFCTR_ICACHE_NO_DELIVER = 0x01,
27 XSCALE_PERFCTR_DATA_STALL = 0x02,
28 XSCALE_PERFCTR_ITLB_MISS = 0x03,
29 XSCALE_PERFCTR_DTLB_MISS = 0x04,
30 XSCALE_PERFCTR_BRANCH = 0x05,
31 XSCALE_PERFCTR_BRANCH_MISS = 0x06,
32 XSCALE_PERFCTR_INSTRUCTION = 0x07,
33 XSCALE_PERFCTR_DCACHE_FULL_STALL = 0x08,
34 XSCALE_PERFCTR_DCACHE_FULL_STALL_CONTIG = 0x09,
35 XSCALE_PERFCTR_DCACHE_ACCESS = 0x0A,
36 XSCALE_PERFCTR_DCACHE_MISS = 0x0B,
37 XSCALE_PERFCTR_DCACHE_WRITE_BACK = 0x0C,
38 XSCALE_PERFCTR_PC_CHANGED = 0x0D,
39 XSCALE_PERFCTR_BCU_REQUEST = 0x10,
40 XSCALE_PERFCTR_BCU_FULL = 0x11,
41 XSCALE_PERFCTR_BCU_DRAIN = 0x12,
42 XSCALE_PERFCTR_BCU_ECC_NO_ELOG = 0x14,
43 XSCALE_PERFCTR_BCU_1_BIT_ERR = 0x15,
44 XSCALE_PERFCTR_RMW = 0x16,
45 /* XSCALE_PERFCTR_CCNT is not hardware defined */
46 XSCALE_PERFCTR_CCNT = 0xFE,
47 XSCALE_PERFCTR_UNUSED = 0xFF,
48};
49
50enum xscale_counters {
51 XSCALE_CYCLE_COUNTER = 0,
52 XSCALE_COUNTER0,
53 XSCALE_COUNTER1,
54 XSCALE_COUNTER2,
55 XSCALE_COUNTER3,
56};
57
58static const unsigned xscale_perf_map[PERF_COUNT_HW_MAX] = {
59 PERF_MAP_ALL_UNSUPPORTED,
60 [PERF_COUNT_HW_CPU_CYCLES] = XSCALE_PERFCTR_CCNT,
61 [PERF_COUNT_HW_INSTRUCTIONS] = XSCALE_PERFCTR_INSTRUCTION,
62 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = XSCALE_PERFCTR_BRANCH,
63 [PERF_COUNT_HW_BRANCH_MISSES] = XSCALE_PERFCTR_BRANCH_MISS,
64 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = XSCALE_PERFCTR_ICACHE_NO_DELIVER,
65};
66
67static const unsigned xscale_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
68 [PERF_COUNT_HW_CACHE_OP_MAX]
69 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
70 PERF_CACHE_MAP_ALL_UNSUPPORTED,
71
72 [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)] = XSCALE_PERFCTR_DCACHE_ACCESS,
73 [C(L1D)][C(OP_READ)][C(RESULT_MISS)] = XSCALE_PERFCTR_DCACHE_MISS,
74 [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = XSCALE_PERFCTR_DCACHE_ACCESS,
75 [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = XSCALE_PERFCTR_DCACHE_MISS,
76
77 [C(L1I)][C(OP_READ)][C(RESULT_MISS)] = XSCALE_PERFCTR_ICACHE_MISS,
78
79 [C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = XSCALE_PERFCTR_DTLB_MISS,
80 [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)] = XSCALE_PERFCTR_DTLB_MISS,
81
82 [C(ITLB)][C(OP_READ)][C(RESULT_MISS)] = XSCALE_PERFCTR_ITLB_MISS,
83 [C(ITLB)][C(OP_WRITE)][C(RESULT_MISS)] = XSCALE_PERFCTR_ITLB_MISS,
84};
85
86#define XSCALE_PMU_ENABLE 0x001
87#define XSCALE_PMN_RESET 0x002
88#define XSCALE_CCNT_RESET 0x004
89#define XSCALE_PMU_RESET (CCNT_RESET | PMN_RESET)
90#define XSCALE_PMU_CNT64 0x008
91
92#define XSCALE1_OVERFLOWED_MASK 0x700
93#define XSCALE1_CCOUNT_OVERFLOW 0x400
94#define XSCALE1_COUNT0_OVERFLOW 0x100
95#define XSCALE1_COUNT1_OVERFLOW 0x200
96#define XSCALE1_CCOUNT_INT_EN 0x040
97#define XSCALE1_COUNT0_INT_EN 0x010
98#define XSCALE1_COUNT1_INT_EN 0x020
99#define XSCALE1_COUNT0_EVT_SHFT 12
100#define XSCALE1_COUNT0_EVT_MASK (0xff << XSCALE1_COUNT0_EVT_SHFT)
101#define XSCALE1_COUNT1_EVT_SHFT 20
102#define XSCALE1_COUNT1_EVT_MASK (0xff << XSCALE1_COUNT1_EVT_SHFT)
103
104static inline u32
105xscale1pmu_read_pmnc(void)
106{
107 u32 val;
108 asm volatile("mrc p14, 0, %0, c0, c0, 0" : "=r" (val));
109 return val;
110}
111
112static inline void
113xscale1pmu_write_pmnc(u32 val)
114{
115 /* upper 4bits and 7, 11 are write-as-0 */
116 val &= 0xffff77f;
117 asm volatile("mcr p14, 0, %0, c0, c0, 0" : : "r" (val));
118}
119
120static inline int
121xscale1_pmnc_counter_has_overflowed(unsigned long pmnc,
122 enum xscale_counters counter)
123{
124 int ret = 0;
125
126 switch (counter) {
127 case XSCALE_CYCLE_COUNTER:
128 ret = pmnc & XSCALE1_CCOUNT_OVERFLOW;
129 break;
130 case XSCALE_COUNTER0:
131 ret = pmnc & XSCALE1_COUNT0_OVERFLOW;
132 break;
133 case XSCALE_COUNTER1:
134 ret = pmnc & XSCALE1_COUNT1_OVERFLOW;
135 break;
136 default:
137 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
138 }
139
140 return ret;
141}
142
143static irqreturn_t
144xscale1pmu_handle_irq(int irq_num, void *dev)
145{
146 unsigned long pmnc;
147 struct perf_sample_data data;
148 struct arm_pmu *cpu_pmu = (struct arm_pmu *)dev;
149 struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
150 struct pt_regs *regs;
151 int idx;
152
153 /*
154 * NOTE: there's an A stepping erratum that states if an overflow
155 * bit already exists and another occurs, the previous
156 * Overflow bit gets cleared. There's no workaround.
157 * Fixed in B stepping or later.
158 */
159 pmnc = xscale1pmu_read_pmnc();
160
161 /*
162 * Write the value back to clear the overflow flags. Overflow
163 * flags remain in pmnc for use below. We also disable the PMU
164 * while we process the interrupt.
165 */
166 xscale1pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
167
168 if (!(pmnc & XSCALE1_OVERFLOWED_MASK))
169 return IRQ_NONE;
170
171 regs = get_irq_regs();
172
173 for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
174 struct perf_event *event = cpuc->events[idx];
175 struct hw_perf_event *hwc;
176
177 if (!event)
178 continue;
179
180 if (!xscale1_pmnc_counter_has_overflowed(pmnc, idx))
181 continue;
182
183 hwc = &event->hw;
184 armpmu_event_update(event);
185 perf_sample_data_init(&data, 0, hwc->last_period);
186 if (!armpmu_event_set_period(event))
187 continue;
188
189 if (perf_event_overflow(event, &data, regs))
190 cpu_pmu->disable(event);
191 }
192
193 irq_work_run();
194
195 /*
196 * Re-enable the PMU.
197 */
198 pmnc = xscale1pmu_read_pmnc() | XSCALE_PMU_ENABLE;
199 xscale1pmu_write_pmnc(pmnc);
200
201 return IRQ_HANDLED;
202}
203
204static void xscale1pmu_enable_event(struct perf_event *event)
205{
206 unsigned long val, mask, evt, flags;
207 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
208 struct hw_perf_event *hwc = &event->hw;
209 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
210 int idx = hwc->idx;
211
212 switch (idx) {
213 case XSCALE_CYCLE_COUNTER:
214 mask = 0;
215 evt = XSCALE1_CCOUNT_INT_EN;
216 break;
217 case XSCALE_COUNTER0:
218 mask = XSCALE1_COUNT0_EVT_MASK;
219 evt = (hwc->config_base << XSCALE1_COUNT0_EVT_SHFT) |
220 XSCALE1_COUNT0_INT_EN;
221 break;
222 case XSCALE_COUNTER1:
223 mask = XSCALE1_COUNT1_EVT_MASK;
224 evt = (hwc->config_base << XSCALE1_COUNT1_EVT_SHFT) |
225 XSCALE1_COUNT1_INT_EN;
226 break;
227 default:
228 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
229 return;
230 }
231
232 raw_spin_lock_irqsave(&events->pmu_lock, flags);
233 val = xscale1pmu_read_pmnc();
234 val &= ~mask;
235 val |= evt;
236 xscale1pmu_write_pmnc(val);
237 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
238}
239
240static void xscale1pmu_disable_event(struct perf_event *event)
241{
242 unsigned long val, mask, evt, flags;
243 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
244 struct hw_perf_event *hwc = &event->hw;
245 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
246 int idx = hwc->idx;
247
248 switch (idx) {
249 case XSCALE_CYCLE_COUNTER:
250 mask = XSCALE1_CCOUNT_INT_EN;
251 evt = 0;
252 break;
253 case XSCALE_COUNTER0:
254 mask = XSCALE1_COUNT0_INT_EN | XSCALE1_COUNT0_EVT_MASK;
255 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT0_EVT_SHFT;
256 break;
257 case XSCALE_COUNTER1:
258 mask = XSCALE1_COUNT1_INT_EN | XSCALE1_COUNT1_EVT_MASK;
259 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT1_EVT_SHFT;
260 break;
261 default:
262 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
263 return;
264 }
265
266 raw_spin_lock_irqsave(&events->pmu_lock, flags);
267 val = xscale1pmu_read_pmnc();
268 val &= ~mask;
269 val |= evt;
270 xscale1pmu_write_pmnc(val);
271 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
272}
273
274static int
275xscale1pmu_get_event_idx(struct pmu_hw_events *cpuc,
276 struct perf_event *event)
277{
278 struct hw_perf_event *hwc = &event->hw;
279 if (XSCALE_PERFCTR_CCNT == hwc->config_base) {
280 if (test_and_set_bit(XSCALE_CYCLE_COUNTER, cpuc->used_mask))
281 return -EAGAIN;
282
283 return XSCALE_CYCLE_COUNTER;
284 } else {
285 if (!test_and_set_bit(XSCALE_COUNTER1, cpuc->used_mask))
286 return XSCALE_COUNTER1;
287
288 if (!test_and_set_bit(XSCALE_COUNTER0, cpuc->used_mask))
289 return XSCALE_COUNTER0;
290
291 return -EAGAIN;
292 }
293}
294
295static void xscale1pmu_start(struct arm_pmu *cpu_pmu)
296{
297 unsigned long flags, val;
298 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
299
300 raw_spin_lock_irqsave(&events->pmu_lock, flags);
301 val = xscale1pmu_read_pmnc();
302 val |= XSCALE_PMU_ENABLE;
303 xscale1pmu_write_pmnc(val);
304 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
305}
306
307static void xscale1pmu_stop(struct arm_pmu *cpu_pmu)
308{
309 unsigned long flags, val;
310 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
311
312 raw_spin_lock_irqsave(&events->pmu_lock, flags);
313 val = xscale1pmu_read_pmnc();
314 val &= ~XSCALE_PMU_ENABLE;
315 xscale1pmu_write_pmnc(val);
316 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
317}
318
319static inline u32 xscale1pmu_read_counter(struct perf_event *event)
320{
321 struct hw_perf_event *hwc = &event->hw;
322 int counter = hwc->idx;
323 u32 val = 0;
324
325 switch (counter) {
326 case XSCALE_CYCLE_COUNTER:
327 asm volatile("mrc p14, 0, %0, c1, c0, 0" : "=r" (val));
328 break;
329 case XSCALE_COUNTER0:
330 asm volatile("mrc p14, 0, %0, c2, c0, 0" : "=r" (val));
331 break;
332 case XSCALE_COUNTER1:
333 asm volatile("mrc p14, 0, %0, c3, c0, 0" : "=r" (val));
334 break;
335 }
336
337 return val;
338}
339
340static inline void xscale1pmu_write_counter(struct perf_event *event, u32 val)
341{
342 struct hw_perf_event *hwc = &event->hw;
343 int counter = hwc->idx;
344
345 switch (counter) {
346 case XSCALE_CYCLE_COUNTER:
347 asm volatile("mcr p14, 0, %0, c1, c0, 0" : : "r" (val));
348 break;
349 case XSCALE_COUNTER0:
350 asm volatile("mcr p14, 0, %0, c2, c0, 0" : : "r" (val));
351 break;
352 case XSCALE_COUNTER1:
353 asm volatile("mcr p14, 0, %0, c3, c0, 0" : : "r" (val));
354 break;
355 }
356}
357
358static int xscale_map_event(struct perf_event *event)
359{
360 return armpmu_map_event(event, &xscale_perf_map,
361 &xscale_perf_cache_map, 0xFF);
362}
363
364static int xscale1pmu_init(struct arm_pmu *cpu_pmu)
365{
366 cpu_pmu->name = "armv5_xscale1";
367 cpu_pmu->handle_irq = xscale1pmu_handle_irq;
368 cpu_pmu->enable = xscale1pmu_enable_event;
369 cpu_pmu->disable = xscale1pmu_disable_event;
370 cpu_pmu->read_counter = xscale1pmu_read_counter;
371 cpu_pmu->write_counter = xscale1pmu_write_counter;
372 cpu_pmu->get_event_idx = xscale1pmu_get_event_idx;
373 cpu_pmu->start = xscale1pmu_start;
374 cpu_pmu->stop = xscale1pmu_stop;
375 cpu_pmu->map_event = xscale_map_event;
376 cpu_pmu->num_events = 3;
377 cpu_pmu->max_period = (1LLU << 32) - 1;
378
379 return 0;
380}
381
382#define XSCALE2_OVERFLOWED_MASK 0x01f
383#define XSCALE2_CCOUNT_OVERFLOW 0x001
384#define XSCALE2_COUNT0_OVERFLOW 0x002
385#define XSCALE2_COUNT1_OVERFLOW 0x004
386#define XSCALE2_COUNT2_OVERFLOW 0x008
387#define XSCALE2_COUNT3_OVERFLOW 0x010
388#define XSCALE2_CCOUNT_INT_EN 0x001
389#define XSCALE2_COUNT0_INT_EN 0x002
390#define XSCALE2_COUNT1_INT_EN 0x004
391#define XSCALE2_COUNT2_INT_EN 0x008
392#define XSCALE2_COUNT3_INT_EN 0x010
393#define XSCALE2_COUNT0_EVT_SHFT 0
394#define XSCALE2_COUNT0_EVT_MASK (0xff << XSCALE2_COUNT0_EVT_SHFT)
395#define XSCALE2_COUNT1_EVT_SHFT 8
396#define XSCALE2_COUNT1_EVT_MASK (0xff << XSCALE2_COUNT1_EVT_SHFT)
397#define XSCALE2_COUNT2_EVT_SHFT 16
398#define XSCALE2_COUNT2_EVT_MASK (0xff << XSCALE2_COUNT2_EVT_SHFT)
399#define XSCALE2_COUNT3_EVT_SHFT 24
400#define XSCALE2_COUNT3_EVT_MASK (0xff << XSCALE2_COUNT3_EVT_SHFT)
401
402static inline u32
403xscale2pmu_read_pmnc(void)
404{
405 u32 val;
406 asm volatile("mrc p14, 0, %0, c0, c1, 0" : "=r" (val));
407 /* bits 1-2 and 4-23 are read-unpredictable */
408 return val & 0xff000009;
409}
410
411static inline void
412xscale2pmu_write_pmnc(u32 val)
413{
414 /* bits 4-23 are write-as-0, 24-31 are write ignored */
415 val &= 0xf;
416 asm volatile("mcr p14, 0, %0, c0, c1, 0" : : "r" (val));
417}
418
419static inline u32
420xscale2pmu_read_overflow_flags(void)
421{
422 u32 val;
423 asm volatile("mrc p14, 0, %0, c5, c1, 0" : "=r" (val));
424 return val;
425}
426
427static inline void
428xscale2pmu_write_overflow_flags(u32 val)
429{
430 asm volatile("mcr p14, 0, %0, c5, c1, 0" : : "r" (val));
431}
432
433static inline u32
434xscale2pmu_read_event_select(void)
435{
436 u32 val;
437 asm volatile("mrc p14, 0, %0, c8, c1, 0" : "=r" (val));
438 return val;
439}
440
441static inline void
442xscale2pmu_write_event_select(u32 val)
443{
444 asm volatile("mcr p14, 0, %0, c8, c1, 0" : : "r"(val));
445}
446
447static inline u32
448xscale2pmu_read_int_enable(void)
449{
450 u32 val;
451 asm volatile("mrc p14, 0, %0, c4, c1, 0" : "=r" (val));
452 return val;
453}
454
455static void
456xscale2pmu_write_int_enable(u32 val)
457{
458 asm volatile("mcr p14, 0, %0, c4, c1, 0" : : "r" (val));
459}
460
461static inline int
462xscale2_pmnc_counter_has_overflowed(unsigned long of_flags,
463 enum xscale_counters counter)
464{
465 int ret = 0;
466
467 switch (counter) {
468 case XSCALE_CYCLE_COUNTER:
469 ret = of_flags & XSCALE2_CCOUNT_OVERFLOW;
470 break;
471 case XSCALE_COUNTER0:
472 ret = of_flags & XSCALE2_COUNT0_OVERFLOW;
473 break;
474 case XSCALE_COUNTER1:
475 ret = of_flags & XSCALE2_COUNT1_OVERFLOW;
476 break;
477 case XSCALE_COUNTER2:
478 ret = of_flags & XSCALE2_COUNT2_OVERFLOW;
479 break;
480 case XSCALE_COUNTER3:
481 ret = of_flags & XSCALE2_COUNT3_OVERFLOW;
482 break;
483 default:
484 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
485 }
486
487 return ret;
488}
489
490static irqreturn_t
491xscale2pmu_handle_irq(int irq_num, void *dev)
492{
493 unsigned long pmnc, of_flags;
494 struct perf_sample_data data;
495 struct arm_pmu *cpu_pmu = (struct arm_pmu *)dev;
496 struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
497 struct pt_regs *regs;
498 int idx;
499
500 /* Disable the PMU. */
501 pmnc = xscale2pmu_read_pmnc();
502 xscale2pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
503
504 /* Check the overflow flag register. */
505 of_flags = xscale2pmu_read_overflow_flags();
506 if (!(of_flags & XSCALE2_OVERFLOWED_MASK))
507 return IRQ_NONE;
508
509 /* Clear the overflow bits. */
510 xscale2pmu_write_overflow_flags(of_flags);
511
512 regs = get_irq_regs();
513
514 for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
515 struct perf_event *event = cpuc->events[idx];
516 struct hw_perf_event *hwc;
517
518 if (!event)
519 continue;
520
521 if (!xscale2_pmnc_counter_has_overflowed(of_flags, idx))
522 continue;
523
524 hwc = &event->hw;
525 armpmu_event_update(event);
526 perf_sample_data_init(&data, 0, hwc->last_period);
527 if (!armpmu_event_set_period(event))
528 continue;
529
530 if (perf_event_overflow(event, &data, regs))
531 cpu_pmu->disable(event);
532 }
533
534 irq_work_run();
535
536 /*
537 * Re-enable the PMU.
538 */
539 pmnc = xscale2pmu_read_pmnc() | XSCALE_PMU_ENABLE;
540 xscale2pmu_write_pmnc(pmnc);
541
542 return IRQ_HANDLED;
543}
544
545static void xscale2pmu_enable_event(struct perf_event *event)
546{
547 unsigned long flags, ien, evtsel;
548 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
549 struct hw_perf_event *hwc = &event->hw;
550 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
551 int idx = hwc->idx;
552
553 ien = xscale2pmu_read_int_enable();
554 evtsel = xscale2pmu_read_event_select();
555
556 switch (idx) {
557 case XSCALE_CYCLE_COUNTER:
558 ien |= XSCALE2_CCOUNT_INT_EN;
559 break;
560 case XSCALE_COUNTER0:
561 ien |= XSCALE2_COUNT0_INT_EN;
562 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
563 evtsel |= hwc->config_base << XSCALE2_COUNT0_EVT_SHFT;
564 break;
565 case XSCALE_COUNTER1:
566 ien |= XSCALE2_COUNT1_INT_EN;
567 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
568 evtsel |= hwc->config_base << XSCALE2_COUNT1_EVT_SHFT;
569 break;
570 case XSCALE_COUNTER2:
571 ien |= XSCALE2_COUNT2_INT_EN;
572 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
573 evtsel |= hwc->config_base << XSCALE2_COUNT2_EVT_SHFT;
574 break;
575 case XSCALE_COUNTER3:
576 ien |= XSCALE2_COUNT3_INT_EN;
577 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
578 evtsel |= hwc->config_base << XSCALE2_COUNT3_EVT_SHFT;
579 break;
580 default:
581 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
582 return;
583 }
584
585 raw_spin_lock_irqsave(&events->pmu_lock, flags);
586 xscale2pmu_write_event_select(evtsel);
587 xscale2pmu_write_int_enable(ien);
588 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
589}
590
591static void xscale2pmu_disable_event(struct perf_event *event)
592{
593 unsigned long flags, ien, evtsel, of_flags;
594 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
595 struct hw_perf_event *hwc = &event->hw;
596 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
597 int idx = hwc->idx;
598
599 ien = xscale2pmu_read_int_enable();
600 evtsel = xscale2pmu_read_event_select();
601
602 switch (idx) {
603 case XSCALE_CYCLE_COUNTER:
604 ien &= ~XSCALE2_CCOUNT_INT_EN;
605 of_flags = XSCALE2_CCOUNT_OVERFLOW;
606 break;
607 case XSCALE_COUNTER0:
608 ien &= ~XSCALE2_COUNT0_INT_EN;
609 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
610 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT0_EVT_SHFT;
611 of_flags = XSCALE2_COUNT0_OVERFLOW;
612 break;
613 case XSCALE_COUNTER1:
614 ien &= ~XSCALE2_COUNT1_INT_EN;
615 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
616 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT1_EVT_SHFT;
617 of_flags = XSCALE2_COUNT1_OVERFLOW;
618 break;
619 case XSCALE_COUNTER2:
620 ien &= ~XSCALE2_COUNT2_INT_EN;
621 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
622 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT2_EVT_SHFT;
623 of_flags = XSCALE2_COUNT2_OVERFLOW;
624 break;
625 case XSCALE_COUNTER3:
626 ien &= ~XSCALE2_COUNT3_INT_EN;
627 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
628 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT3_EVT_SHFT;
629 of_flags = XSCALE2_COUNT3_OVERFLOW;
630 break;
631 default:
632 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
633 return;
634 }
635
636 raw_spin_lock_irqsave(&events->pmu_lock, flags);
637 xscale2pmu_write_event_select(evtsel);
638 xscale2pmu_write_int_enable(ien);
639 xscale2pmu_write_overflow_flags(of_flags);
640 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
641}
642
643static int
644xscale2pmu_get_event_idx(struct pmu_hw_events *cpuc,
645 struct perf_event *event)
646{
647 int idx = xscale1pmu_get_event_idx(cpuc, event);
648 if (idx >= 0)
649 goto out;
650
651 if (!test_and_set_bit(XSCALE_COUNTER3, cpuc->used_mask))
652 idx = XSCALE_COUNTER3;
653 else if (!test_and_set_bit(XSCALE_COUNTER2, cpuc->used_mask))
654 idx = XSCALE_COUNTER2;
655out:
656 return idx;
657}
658
659static void xscale2pmu_start(struct arm_pmu *cpu_pmu)
660{
661 unsigned long flags, val;
662 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
663
664 raw_spin_lock_irqsave(&events->pmu_lock, flags);
665 val = xscale2pmu_read_pmnc() & ~XSCALE_PMU_CNT64;
666 val |= XSCALE_PMU_ENABLE;
667 xscale2pmu_write_pmnc(val);
668 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
669}
670
671static void xscale2pmu_stop(struct arm_pmu *cpu_pmu)
672{
673 unsigned long flags, val;
674 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
675
676 raw_spin_lock_irqsave(&events->pmu_lock, flags);
677 val = xscale2pmu_read_pmnc();
678 val &= ~XSCALE_PMU_ENABLE;
679 xscale2pmu_write_pmnc(val);
680 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
681}
682
683static inline u32 xscale2pmu_read_counter(struct perf_event *event)
684{
685 struct hw_perf_event *hwc = &event->hw;
686 int counter = hwc->idx;
687 u32 val = 0;
688
689 switch (counter) {
690 case XSCALE_CYCLE_COUNTER:
691 asm volatile("mrc p14, 0, %0, c1, c1, 0" : "=r" (val));
692 break;
693 case XSCALE_COUNTER0:
694 asm volatile("mrc p14, 0, %0, c0, c2, 0" : "=r" (val));
695 break;
696 case XSCALE_COUNTER1:
697 asm volatile("mrc p14, 0, %0, c1, c2, 0" : "=r" (val));
698 break;
699 case XSCALE_COUNTER2:
700 asm volatile("mrc p14, 0, %0, c2, c2, 0" : "=r" (val));
701 break;
702 case XSCALE_COUNTER3:
703 asm volatile("mrc p14, 0, %0, c3, c2, 0" : "=r" (val));
704 break;
705 }
706
707 return val;
708}
709
710static inline void xscale2pmu_write_counter(struct perf_event *event, u32 val)
711{
712 struct hw_perf_event *hwc = &event->hw;
713 int counter = hwc->idx;
714
715 switch (counter) {
716 case XSCALE_CYCLE_COUNTER:
717 asm volatile("mcr p14, 0, %0, c1, c1, 0" : : "r" (val));
718 break;
719 case XSCALE_COUNTER0:
720 asm volatile("mcr p14, 0, %0, c0, c2, 0" : : "r" (val));
721 break;
722 case XSCALE_COUNTER1:
723 asm volatile("mcr p14, 0, %0, c1, c2, 0" : : "r" (val));
724 break;
725 case XSCALE_COUNTER2:
726 asm volatile("mcr p14, 0, %0, c2, c2, 0" : : "r" (val));
727 break;
728 case XSCALE_COUNTER3:
729 asm volatile("mcr p14, 0, %0, c3, c2, 0" : : "r" (val));
730 break;
731 }
732}
733
734static int xscale2pmu_init(struct arm_pmu *cpu_pmu)
735{
736 cpu_pmu->name = "armv5_xscale2";
737 cpu_pmu->handle_irq = xscale2pmu_handle_irq;
738 cpu_pmu->enable = xscale2pmu_enable_event;
739 cpu_pmu->disable = xscale2pmu_disable_event;
740 cpu_pmu->read_counter = xscale2pmu_read_counter;
741 cpu_pmu->write_counter = xscale2pmu_write_counter;
742 cpu_pmu->get_event_idx = xscale2pmu_get_event_idx;
743 cpu_pmu->start = xscale2pmu_start;
744 cpu_pmu->stop = xscale2pmu_stop;
745 cpu_pmu->map_event = xscale_map_event;
746 cpu_pmu->num_events = 5;
747 cpu_pmu->max_period = (1LLU << 32) - 1;
748
749 return 0;
750}
751
752static const struct pmu_probe_info xscale_pmu_probe_table[] = {
753 XSCALE_PMU_PROBE(ARM_CPU_XSCALE_ARCH_V1, xscale1pmu_init),
754 XSCALE_PMU_PROBE(ARM_CPU_XSCALE_ARCH_V2, xscale2pmu_init),
755 { /* sentinel value */ }
756};
757
758static int xscale_pmu_device_probe(struct platform_device *pdev)
759{
760 return arm_pmu_device_probe(pdev, NULL, xscale_pmu_probe_table);
761}
762
763static struct platform_driver xscale_pmu_driver = {
764 .driver = {
765 .name = "xscale-pmu",
766 },
767 .probe = xscale_pmu_device_probe,
768};
769
770static int __init register_xscale_pmu_driver(void)
771{
772 return platform_driver_register(&xscale_pmu_driver);
773}
774device_initcall(register_xscale_pmu_driver);
775#endif /* CONFIG_CPU_XSCALE */