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// SPDX-License-Identifier: GPL-2.0
2/*
3 * ARMv5 [xscale] Performance counter handling code.
4 *
5 * Copyright (C) 2010, ARM Ltd., Will Deacon <will.deacon@arm.com>
6 *
7 * Based on the previous xscale OProfile code.
8 *
9 * There are two variants of the xscale PMU that we support:
10 * - xscale1pmu: 2 event counters and a cycle counter
11 * - xscale2pmu: 4 event counters and a cycle counter
12 * The two variants share event definitions, but have different
13 * PMU structures.
14 */
15
16#ifdef CONFIG_CPU_XSCALE
17
18#include <asm/cputype.h>
19#include <asm/irq_regs.h>
20
21#include <linux/of.h>
22#include <linux/perf/arm_pmu.h>
23#include <linux/platform_device.h>
24
25enum xscale_perf_types {
26 XSCALE_PERFCTR_ICACHE_MISS = 0x00,
27 XSCALE_PERFCTR_ICACHE_NO_DELIVER = 0x01,
28 XSCALE_PERFCTR_DATA_STALL = 0x02,
29 XSCALE_PERFCTR_ITLB_MISS = 0x03,
30 XSCALE_PERFCTR_DTLB_MISS = 0x04,
31 XSCALE_PERFCTR_BRANCH = 0x05,
32 XSCALE_PERFCTR_BRANCH_MISS = 0x06,
33 XSCALE_PERFCTR_INSTRUCTION = 0x07,
34 XSCALE_PERFCTR_DCACHE_FULL_STALL = 0x08,
35 XSCALE_PERFCTR_DCACHE_FULL_STALL_CONTIG = 0x09,
36 XSCALE_PERFCTR_DCACHE_ACCESS = 0x0A,
37 XSCALE_PERFCTR_DCACHE_MISS = 0x0B,
38 XSCALE_PERFCTR_DCACHE_WRITE_BACK = 0x0C,
39 XSCALE_PERFCTR_PC_CHANGED = 0x0D,
40 XSCALE_PERFCTR_BCU_REQUEST = 0x10,
41 XSCALE_PERFCTR_BCU_FULL = 0x11,
42 XSCALE_PERFCTR_BCU_DRAIN = 0x12,
43 XSCALE_PERFCTR_BCU_ECC_NO_ELOG = 0x14,
44 XSCALE_PERFCTR_BCU_1_BIT_ERR = 0x15,
45 XSCALE_PERFCTR_RMW = 0x16,
46 /* XSCALE_PERFCTR_CCNT is not hardware defined */
47 XSCALE_PERFCTR_CCNT = 0xFE,
48 XSCALE_PERFCTR_UNUSED = 0xFF,
49};
50
51enum xscale_counters {
52 XSCALE_CYCLE_COUNTER = 0,
53 XSCALE_COUNTER0,
54 XSCALE_COUNTER1,
55 XSCALE_COUNTER2,
56 XSCALE_COUNTER3,
57};
58
59static const unsigned xscale_perf_map[PERF_COUNT_HW_MAX] = {
60 PERF_MAP_ALL_UNSUPPORTED,
61 [PERF_COUNT_HW_CPU_CYCLES] = XSCALE_PERFCTR_CCNT,
62 [PERF_COUNT_HW_INSTRUCTIONS] = XSCALE_PERFCTR_INSTRUCTION,
63 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = XSCALE_PERFCTR_BRANCH,
64 [PERF_COUNT_HW_BRANCH_MISSES] = XSCALE_PERFCTR_BRANCH_MISS,
65 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = XSCALE_PERFCTR_ICACHE_NO_DELIVER,
66};
67
68static const unsigned xscale_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
69 [PERF_COUNT_HW_CACHE_OP_MAX]
70 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
71 PERF_CACHE_MAP_ALL_UNSUPPORTED,
72
73 [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)] = XSCALE_PERFCTR_DCACHE_ACCESS,
74 [C(L1D)][C(OP_READ)][C(RESULT_MISS)] = XSCALE_PERFCTR_DCACHE_MISS,
75 [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = XSCALE_PERFCTR_DCACHE_ACCESS,
76 [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = XSCALE_PERFCTR_DCACHE_MISS,
77
78 [C(L1I)][C(OP_READ)][C(RESULT_MISS)] = XSCALE_PERFCTR_ICACHE_MISS,
79
80 [C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = XSCALE_PERFCTR_DTLB_MISS,
81 [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)] = XSCALE_PERFCTR_DTLB_MISS,
82
83 [C(ITLB)][C(OP_READ)][C(RESULT_MISS)] = XSCALE_PERFCTR_ITLB_MISS,
84 [C(ITLB)][C(OP_WRITE)][C(RESULT_MISS)] = XSCALE_PERFCTR_ITLB_MISS,
85};
86
87#define XSCALE_PMU_ENABLE 0x001
88#define XSCALE_PMN_RESET 0x002
89#define XSCALE_CCNT_RESET 0x004
90#define XSCALE_PMU_RESET (CCNT_RESET | PMN_RESET)
91#define XSCALE_PMU_CNT64 0x008
92
93#define XSCALE1_OVERFLOWED_MASK 0x700
94#define XSCALE1_CCOUNT_OVERFLOW 0x400
95#define XSCALE1_COUNT0_OVERFLOW 0x100
96#define XSCALE1_COUNT1_OVERFLOW 0x200
97#define XSCALE1_CCOUNT_INT_EN 0x040
98#define XSCALE1_COUNT0_INT_EN 0x010
99#define XSCALE1_COUNT1_INT_EN 0x020
100#define XSCALE1_COUNT0_EVT_SHFT 12
101#define XSCALE1_COUNT0_EVT_MASK (0xff << XSCALE1_COUNT0_EVT_SHFT)
102#define XSCALE1_COUNT1_EVT_SHFT 20
103#define XSCALE1_COUNT1_EVT_MASK (0xff << XSCALE1_COUNT1_EVT_SHFT)
104
105static inline u32
106xscale1pmu_read_pmnc(void)
107{
108 u32 val;
109 asm volatile("mrc p14, 0, %0, c0, c0, 0" : "=r" (val));
110 return val;
111}
112
113static inline void
114xscale1pmu_write_pmnc(u32 val)
115{
116 /* upper 4bits and 7, 11 are write-as-0 */
117 val &= 0xffff77f;
118 asm volatile("mcr p14, 0, %0, c0, c0, 0" : : "r" (val));
119}
120
121static inline int
122xscale1_pmnc_counter_has_overflowed(unsigned long pmnc,
123 enum xscale_counters counter)
124{
125 int ret = 0;
126
127 switch (counter) {
128 case XSCALE_CYCLE_COUNTER:
129 ret = pmnc & XSCALE1_CCOUNT_OVERFLOW;
130 break;
131 case XSCALE_COUNTER0:
132 ret = pmnc & XSCALE1_COUNT0_OVERFLOW;
133 break;
134 case XSCALE_COUNTER1:
135 ret = pmnc & XSCALE1_COUNT1_OVERFLOW;
136 break;
137 default:
138 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
139 }
140
141 return ret;
142}
143
144static irqreturn_t
145xscale1pmu_handle_irq(int irq_num, void *dev)
146{
147 unsigned long pmnc;
148 struct perf_sample_data data;
149 struct arm_pmu *cpu_pmu = (struct arm_pmu *)dev;
150 struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
151 struct pt_regs *regs;
152 int idx;
153
154 /*
155 * NOTE: there's an A stepping erratum that states if an overflow
156 * bit already exists and another occurs, the previous
157 * Overflow bit gets cleared. There's no workaround.
158 * Fixed in B stepping or later.
159 */
160 pmnc = xscale1pmu_read_pmnc();
161
162 /*
163 * Write the value back to clear the overflow flags. Overflow
164 * flags remain in pmnc for use below. We also disable the PMU
165 * while we process the interrupt.
166 */
167 xscale1pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
168
169 if (!(pmnc & XSCALE1_OVERFLOWED_MASK))
170 return IRQ_NONE;
171
172 regs = get_irq_regs();
173
174 for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
175 struct perf_event *event = cpuc->events[idx];
176 struct hw_perf_event *hwc;
177
178 if (!event)
179 continue;
180
181 if (!xscale1_pmnc_counter_has_overflowed(pmnc, idx))
182 continue;
183
184 hwc = &event->hw;
185 armpmu_event_update(event);
186 perf_sample_data_init(&data, 0, hwc->last_period);
187 if (!armpmu_event_set_period(event))
188 continue;
189
190 if (perf_event_overflow(event, &data, regs))
191 cpu_pmu->disable(event);
192 }
193
194 irq_work_run();
195
196 /*
197 * Re-enable the PMU.
198 */
199 pmnc = xscale1pmu_read_pmnc() | XSCALE_PMU_ENABLE;
200 xscale1pmu_write_pmnc(pmnc);
201
202 return IRQ_HANDLED;
203}
204
205static void xscale1pmu_enable_event(struct perf_event *event)
206{
207 unsigned long val, mask, evt, flags;
208 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
209 struct hw_perf_event *hwc = &event->hw;
210 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
211 int idx = hwc->idx;
212
213 switch (idx) {
214 case XSCALE_CYCLE_COUNTER:
215 mask = 0;
216 evt = XSCALE1_CCOUNT_INT_EN;
217 break;
218 case XSCALE_COUNTER0:
219 mask = XSCALE1_COUNT0_EVT_MASK;
220 evt = (hwc->config_base << XSCALE1_COUNT0_EVT_SHFT) |
221 XSCALE1_COUNT0_INT_EN;
222 break;
223 case XSCALE_COUNTER1:
224 mask = XSCALE1_COUNT1_EVT_MASK;
225 evt = (hwc->config_base << XSCALE1_COUNT1_EVT_SHFT) |
226 XSCALE1_COUNT1_INT_EN;
227 break;
228 default:
229 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
230 return;
231 }
232
233 raw_spin_lock_irqsave(&events->pmu_lock, flags);
234 val = xscale1pmu_read_pmnc();
235 val &= ~mask;
236 val |= evt;
237 xscale1pmu_write_pmnc(val);
238 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
239}
240
241static void xscale1pmu_disable_event(struct perf_event *event)
242{
243 unsigned long val, mask, evt, flags;
244 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
245 struct hw_perf_event *hwc = &event->hw;
246 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
247 int idx = hwc->idx;
248
249 switch (idx) {
250 case XSCALE_CYCLE_COUNTER:
251 mask = XSCALE1_CCOUNT_INT_EN;
252 evt = 0;
253 break;
254 case XSCALE_COUNTER0:
255 mask = XSCALE1_COUNT0_INT_EN | XSCALE1_COUNT0_EVT_MASK;
256 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT0_EVT_SHFT;
257 break;
258 case XSCALE_COUNTER1:
259 mask = XSCALE1_COUNT1_INT_EN | XSCALE1_COUNT1_EVT_MASK;
260 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT1_EVT_SHFT;
261 break;
262 default:
263 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
264 return;
265 }
266
267 raw_spin_lock_irqsave(&events->pmu_lock, flags);
268 val = xscale1pmu_read_pmnc();
269 val &= ~mask;
270 val |= evt;
271 xscale1pmu_write_pmnc(val);
272 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
273}
274
275static int
276xscale1pmu_get_event_idx(struct pmu_hw_events *cpuc,
277 struct perf_event *event)
278{
279 struct hw_perf_event *hwc = &event->hw;
280 if (XSCALE_PERFCTR_CCNT == hwc->config_base) {
281 if (test_and_set_bit(XSCALE_CYCLE_COUNTER, cpuc->used_mask))
282 return -EAGAIN;
283
284 return XSCALE_CYCLE_COUNTER;
285 } else {
286 if (!test_and_set_bit(XSCALE_COUNTER1, cpuc->used_mask))
287 return XSCALE_COUNTER1;
288
289 if (!test_and_set_bit(XSCALE_COUNTER0, cpuc->used_mask))
290 return XSCALE_COUNTER0;
291
292 return -EAGAIN;
293 }
294}
295
296static void xscale1pmu_start(struct arm_pmu *cpu_pmu)
297{
298 unsigned long flags, val;
299 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
300
301 raw_spin_lock_irqsave(&events->pmu_lock, flags);
302 val = xscale1pmu_read_pmnc();
303 val |= XSCALE_PMU_ENABLE;
304 xscale1pmu_write_pmnc(val);
305 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
306}
307
308static void xscale1pmu_stop(struct arm_pmu *cpu_pmu)
309{
310 unsigned long flags, val;
311 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
312
313 raw_spin_lock_irqsave(&events->pmu_lock, flags);
314 val = xscale1pmu_read_pmnc();
315 val &= ~XSCALE_PMU_ENABLE;
316 xscale1pmu_write_pmnc(val);
317 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
318}
319
320static inline u32 xscale1pmu_read_counter(struct perf_event *event)
321{
322 struct hw_perf_event *hwc = &event->hw;
323 int counter = hwc->idx;
324 u32 val = 0;
325
326 switch (counter) {
327 case XSCALE_CYCLE_COUNTER:
328 asm volatile("mrc p14, 0, %0, c1, c0, 0" : "=r" (val));
329 break;
330 case XSCALE_COUNTER0:
331 asm volatile("mrc p14, 0, %0, c2, c0, 0" : "=r" (val));
332 break;
333 case XSCALE_COUNTER1:
334 asm volatile("mrc p14, 0, %0, c3, c0, 0" : "=r" (val));
335 break;
336 }
337
338 return val;
339}
340
341static inline void xscale1pmu_write_counter(struct perf_event *event, u32 val)
342{
343 struct hw_perf_event *hwc = &event->hw;
344 int counter = hwc->idx;
345
346 switch (counter) {
347 case XSCALE_CYCLE_COUNTER:
348 asm volatile("mcr p14, 0, %0, c1, c0, 0" : : "r" (val));
349 break;
350 case XSCALE_COUNTER0:
351 asm volatile("mcr p14, 0, %0, c2, c0, 0" : : "r" (val));
352 break;
353 case XSCALE_COUNTER1:
354 asm volatile("mcr p14, 0, %0, c3, c0, 0" : : "r" (val));
355 break;
356 }
357}
358
359static int xscale_map_event(struct perf_event *event)
360{
361 return armpmu_map_event(event, &xscale_perf_map,
362 &xscale_perf_cache_map, 0xFF);
363}
364
365static int xscale1pmu_init(struct arm_pmu *cpu_pmu)
366{
367 cpu_pmu->name = "armv5_xscale1";
368 cpu_pmu->handle_irq = xscale1pmu_handle_irq;
369 cpu_pmu->enable = xscale1pmu_enable_event;
370 cpu_pmu->disable = xscale1pmu_disable_event;
371 cpu_pmu->read_counter = xscale1pmu_read_counter;
372 cpu_pmu->write_counter = xscale1pmu_write_counter;
373 cpu_pmu->get_event_idx = xscale1pmu_get_event_idx;
374 cpu_pmu->start = xscale1pmu_start;
375 cpu_pmu->stop = xscale1pmu_stop;
376 cpu_pmu->map_event = xscale_map_event;
377 cpu_pmu->num_events = 3;
378 cpu_pmu->max_period = (1LLU << 32) - 1;
379
380 return 0;
381}
382
383#define XSCALE2_OVERFLOWED_MASK 0x01f
384#define XSCALE2_CCOUNT_OVERFLOW 0x001
385#define XSCALE2_COUNT0_OVERFLOW 0x002
386#define XSCALE2_COUNT1_OVERFLOW 0x004
387#define XSCALE2_COUNT2_OVERFLOW 0x008
388#define XSCALE2_COUNT3_OVERFLOW 0x010
389#define XSCALE2_CCOUNT_INT_EN 0x001
390#define XSCALE2_COUNT0_INT_EN 0x002
391#define XSCALE2_COUNT1_INT_EN 0x004
392#define XSCALE2_COUNT2_INT_EN 0x008
393#define XSCALE2_COUNT3_INT_EN 0x010
394#define XSCALE2_COUNT0_EVT_SHFT 0
395#define XSCALE2_COUNT0_EVT_MASK (0xff << XSCALE2_COUNT0_EVT_SHFT)
396#define XSCALE2_COUNT1_EVT_SHFT 8
397#define XSCALE2_COUNT1_EVT_MASK (0xff << XSCALE2_COUNT1_EVT_SHFT)
398#define XSCALE2_COUNT2_EVT_SHFT 16
399#define XSCALE2_COUNT2_EVT_MASK (0xff << XSCALE2_COUNT2_EVT_SHFT)
400#define XSCALE2_COUNT3_EVT_SHFT 24
401#define XSCALE2_COUNT3_EVT_MASK (0xff << XSCALE2_COUNT3_EVT_SHFT)
402
403static inline u32
404xscale2pmu_read_pmnc(void)
405{
406 u32 val;
407 asm volatile("mrc p14, 0, %0, c0, c1, 0" : "=r" (val));
408 /* bits 1-2 and 4-23 are read-unpredictable */
409 return val & 0xff000009;
410}
411
412static inline void
413xscale2pmu_write_pmnc(u32 val)
414{
415 /* bits 4-23 are write-as-0, 24-31 are write ignored */
416 val &= 0xf;
417 asm volatile("mcr p14, 0, %0, c0, c1, 0" : : "r" (val));
418}
419
420static inline u32
421xscale2pmu_read_overflow_flags(void)
422{
423 u32 val;
424 asm volatile("mrc p14, 0, %0, c5, c1, 0" : "=r" (val));
425 return val;
426}
427
428static inline void
429xscale2pmu_write_overflow_flags(u32 val)
430{
431 asm volatile("mcr p14, 0, %0, c5, c1, 0" : : "r" (val));
432}
433
434static inline u32
435xscale2pmu_read_event_select(void)
436{
437 u32 val;
438 asm volatile("mrc p14, 0, %0, c8, c1, 0" : "=r" (val));
439 return val;
440}
441
442static inline void
443xscale2pmu_write_event_select(u32 val)
444{
445 asm volatile("mcr p14, 0, %0, c8, c1, 0" : : "r"(val));
446}
447
448static inline u32
449xscale2pmu_read_int_enable(void)
450{
451 u32 val;
452 asm volatile("mrc p14, 0, %0, c4, c1, 0" : "=r" (val));
453 return val;
454}
455
456static void
457xscale2pmu_write_int_enable(u32 val)
458{
459 asm volatile("mcr p14, 0, %0, c4, c1, 0" : : "r" (val));
460}
461
462static inline int
463xscale2_pmnc_counter_has_overflowed(unsigned long of_flags,
464 enum xscale_counters counter)
465{
466 int ret = 0;
467
468 switch (counter) {
469 case XSCALE_CYCLE_COUNTER:
470 ret = of_flags & XSCALE2_CCOUNT_OVERFLOW;
471 break;
472 case XSCALE_COUNTER0:
473 ret = of_flags & XSCALE2_COUNT0_OVERFLOW;
474 break;
475 case XSCALE_COUNTER1:
476 ret = of_flags & XSCALE2_COUNT1_OVERFLOW;
477 break;
478 case XSCALE_COUNTER2:
479 ret = of_flags & XSCALE2_COUNT2_OVERFLOW;
480 break;
481 case XSCALE_COUNTER3:
482 ret = of_flags & XSCALE2_COUNT3_OVERFLOW;
483 break;
484 default:
485 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
486 }
487
488 return ret;
489}
490
491static irqreturn_t
492xscale2pmu_handle_irq(int irq_num, void *dev)
493{
494 unsigned long pmnc, of_flags;
495 struct perf_sample_data data;
496 struct arm_pmu *cpu_pmu = (struct arm_pmu *)dev;
497 struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
498 struct pt_regs *regs;
499 int idx;
500
501 /* Disable the PMU. */
502 pmnc = xscale2pmu_read_pmnc();
503 xscale2pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
504
505 /* Check the overflow flag register. */
506 of_flags = xscale2pmu_read_overflow_flags();
507 if (!(of_flags & XSCALE2_OVERFLOWED_MASK))
508 return IRQ_NONE;
509
510 /* Clear the overflow bits. */
511 xscale2pmu_write_overflow_flags(of_flags);
512
513 regs = get_irq_regs();
514
515 for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
516 struct perf_event *event = cpuc->events[idx];
517 struct hw_perf_event *hwc;
518
519 if (!event)
520 continue;
521
522 if (!xscale2_pmnc_counter_has_overflowed(of_flags, idx))
523 continue;
524
525 hwc = &event->hw;
526 armpmu_event_update(event);
527 perf_sample_data_init(&data, 0, hwc->last_period);
528 if (!armpmu_event_set_period(event))
529 continue;
530
531 if (perf_event_overflow(event, &data, regs))
532 cpu_pmu->disable(event);
533 }
534
535 irq_work_run();
536
537 /*
538 * Re-enable the PMU.
539 */
540 pmnc = xscale2pmu_read_pmnc() | XSCALE_PMU_ENABLE;
541 xscale2pmu_write_pmnc(pmnc);
542
543 return IRQ_HANDLED;
544}
545
546static void xscale2pmu_enable_event(struct perf_event *event)
547{
548 unsigned long flags, ien, evtsel;
549 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
550 struct hw_perf_event *hwc = &event->hw;
551 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
552 int idx = hwc->idx;
553
554 ien = xscale2pmu_read_int_enable();
555 evtsel = xscale2pmu_read_event_select();
556
557 switch (idx) {
558 case XSCALE_CYCLE_COUNTER:
559 ien |= XSCALE2_CCOUNT_INT_EN;
560 break;
561 case XSCALE_COUNTER0:
562 ien |= XSCALE2_COUNT0_INT_EN;
563 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
564 evtsel |= hwc->config_base << XSCALE2_COUNT0_EVT_SHFT;
565 break;
566 case XSCALE_COUNTER1:
567 ien |= XSCALE2_COUNT1_INT_EN;
568 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
569 evtsel |= hwc->config_base << XSCALE2_COUNT1_EVT_SHFT;
570 break;
571 case XSCALE_COUNTER2:
572 ien |= XSCALE2_COUNT2_INT_EN;
573 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
574 evtsel |= hwc->config_base << XSCALE2_COUNT2_EVT_SHFT;
575 break;
576 case XSCALE_COUNTER3:
577 ien |= XSCALE2_COUNT3_INT_EN;
578 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
579 evtsel |= hwc->config_base << XSCALE2_COUNT3_EVT_SHFT;
580 break;
581 default:
582 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
583 return;
584 }
585
586 raw_spin_lock_irqsave(&events->pmu_lock, flags);
587 xscale2pmu_write_event_select(evtsel);
588 xscale2pmu_write_int_enable(ien);
589 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
590}
591
592static void xscale2pmu_disable_event(struct perf_event *event)
593{
594 unsigned long flags, ien, evtsel, of_flags;
595 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
596 struct hw_perf_event *hwc = &event->hw;
597 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
598 int idx = hwc->idx;
599
600 ien = xscale2pmu_read_int_enable();
601 evtsel = xscale2pmu_read_event_select();
602
603 switch (idx) {
604 case XSCALE_CYCLE_COUNTER:
605 ien &= ~XSCALE2_CCOUNT_INT_EN;
606 of_flags = XSCALE2_CCOUNT_OVERFLOW;
607 break;
608 case XSCALE_COUNTER0:
609 ien &= ~XSCALE2_COUNT0_INT_EN;
610 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
611 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT0_EVT_SHFT;
612 of_flags = XSCALE2_COUNT0_OVERFLOW;
613 break;
614 case XSCALE_COUNTER1:
615 ien &= ~XSCALE2_COUNT1_INT_EN;
616 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
617 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT1_EVT_SHFT;
618 of_flags = XSCALE2_COUNT1_OVERFLOW;
619 break;
620 case XSCALE_COUNTER2:
621 ien &= ~XSCALE2_COUNT2_INT_EN;
622 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
623 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT2_EVT_SHFT;
624 of_flags = XSCALE2_COUNT2_OVERFLOW;
625 break;
626 case XSCALE_COUNTER3:
627 ien &= ~XSCALE2_COUNT3_INT_EN;
628 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
629 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT3_EVT_SHFT;
630 of_flags = XSCALE2_COUNT3_OVERFLOW;
631 break;
632 default:
633 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
634 return;
635 }
636
637 raw_spin_lock_irqsave(&events->pmu_lock, flags);
638 xscale2pmu_write_event_select(evtsel);
639 xscale2pmu_write_int_enable(ien);
640 xscale2pmu_write_overflow_flags(of_flags);
641 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
642}
643
644static int
645xscale2pmu_get_event_idx(struct pmu_hw_events *cpuc,
646 struct perf_event *event)
647{
648 int idx = xscale1pmu_get_event_idx(cpuc, event);
649 if (idx >= 0)
650 goto out;
651
652 if (!test_and_set_bit(XSCALE_COUNTER3, cpuc->used_mask))
653 idx = XSCALE_COUNTER3;
654 else if (!test_and_set_bit(XSCALE_COUNTER2, cpuc->used_mask))
655 idx = XSCALE_COUNTER2;
656out:
657 return idx;
658}
659
660static void xscale2pmu_start(struct arm_pmu *cpu_pmu)
661{
662 unsigned long flags, val;
663 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
664
665 raw_spin_lock_irqsave(&events->pmu_lock, flags);
666 val = xscale2pmu_read_pmnc() & ~XSCALE_PMU_CNT64;
667 val |= XSCALE_PMU_ENABLE;
668 xscale2pmu_write_pmnc(val);
669 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
670}
671
672static void xscale2pmu_stop(struct arm_pmu *cpu_pmu)
673{
674 unsigned long flags, val;
675 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
676
677 raw_spin_lock_irqsave(&events->pmu_lock, flags);
678 val = xscale2pmu_read_pmnc();
679 val &= ~XSCALE_PMU_ENABLE;
680 xscale2pmu_write_pmnc(val);
681 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
682}
683
684static inline u32 xscale2pmu_read_counter(struct perf_event *event)
685{
686 struct hw_perf_event *hwc = &event->hw;
687 int counter = hwc->idx;
688 u32 val = 0;
689
690 switch (counter) {
691 case XSCALE_CYCLE_COUNTER:
692 asm volatile("mrc p14, 0, %0, c1, c1, 0" : "=r" (val));
693 break;
694 case XSCALE_COUNTER0:
695 asm volatile("mrc p14, 0, %0, c0, c2, 0" : "=r" (val));
696 break;
697 case XSCALE_COUNTER1:
698 asm volatile("mrc p14, 0, %0, c1, c2, 0" : "=r" (val));
699 break;
700 case XSCALE_COUNTER2:
701 asm volatile("mrc p14, 0, %0, c2, c2, 0" : "=r" (val));
702 break;
703 case XSCALE_COUNTER3:
704 asm volatile("mrc p14, 0, %0, c3, c2, 0" : "=r" (val));
705 break;
706 }
707
708 return val;
709}
710
711static inline void xscale2pmu_write_counter(struct perf_event *event, u32 val)
712{
713 struct hw_perf_event *hwc = &event->hw;
714 int counter = hwc->idx;
715
716 switch (counter) {
717 case XSCALE_CYCLE_COUNTER:
718 asm volatile("mcr p14, 0, %0, c1, c1, 0" : : "r" (val));
719 break;
720 case XSCALE_COUNTER0:
721 asm volatile("mcr p14, 0, %0, c0, c2, 0" : : "r" (val));
722 break;
723 case XSCALE_COUNTER1:
724 asm volatile("mcr p14, 0, %0, c1, c2, 0" : : "r" (val));
725 break;
726 case XSCALE_COUNTER2:
727 asm volatile("mcr p14, 0, %0, c2, c2, 0" : : "r" (val));
728 break;
729 case XSCALE_COUNTER3:
730 asm volatile("mcr p14, 0, %0, c3, c2, 0" : : "r" (val));
731 break;
732 }
733}
734
735static int xscale2pmu_init(struct arm_pmu *cpu_pmu)
736{
737 cpu_pmu->name = "armv5_xscale2";
738 cpu_pmu->handle_irq = xscale2pmu_handle_irq;
739 cpu_pmu->enable = xscale2pmu_enable_event;
740 cpu_pmu->disable = xscale2pmu_disable_event;
741 cpu_pmu->read_counter = xscale2pmu_read_counter;
742 cpu_pmu->write_counter = xscale2pmu_write_counter;
743 cpu_pmu->get_event_idx = xscale2pmu_get_event_idx;
744 cpu_pmu->start = xscale2pmu_start;
745 cpu_pmu->stop = xscale2pmu_stop;
746 cpu_pmu->map_event = xscale_map_event;
747 cpu_pmu->num_events = 5;
748 cpu_pmu->max_period = (1LLU << 32) - 1;
749
750 return 0;
751}
752
753static const struct pmu_probe_info xscale_pmu_probe_table[] = {
754 XSCALE_PMU_PROBE(ARM_CPU_XSCALE_ARCH_V1, xscale1pmu_init),
755 XSCALE_PMU_PROBE(ARM_CPU_XSCALE_ARCH_V2, xscale2pmu_init),
756 { /* sentinel value */ }
757};
758
759static int xscale_pmu_device_probe(struct platform_device *pdev)
760{
761 return arm_pmu_device_probe(pdev, NULL, xscale_pmu_probe_table);
762}
763
764static struct platform_driver xscale_pmu_driver = {
765 .driver = {
766 .name = "xscale-pmu",
767 },
768 .probe = xscale_pmu_device_probe,
769};
770
771builtin_platform_driver(xscale_pmu_driver);
772#endif /* CONFIG_CPU_XSCALE */