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 = 0,
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 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = XSCALE_PERFCTR_ICACHE_NO_DELIVER,
59 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = HW_OP_UNSUPPORTED,
60};
61
62static const unsigned xscale_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
63 [PERF_COUNT_HW_CACHE_OP_MAX]
64 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
65 [C(L1D)] = {
66 [C(OP_READ)] = {
67 [C(RESULT_ACCESS)] = XSCALE_PERFCTR_DCACHE_ACCESS,
68 [C(RESULT_MISS)] = XSCALE_PERFCTR_DCACHE_MISS,
69 },
70 [C(OP_WRITE)] = {
71 [C(RESULT_ACCESS)] = XSCALE_PERFCTR_DCACHE_ACCESS,
72 [C(RESULT_MISS)] = XSCALE_PERFCTR_DCACHE_MISS,
73 },
74 [C(OP_PREFETCH)] = {
75 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
76 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
77 },
78 },
79 [C(L1I)] = {
80 [C(OP_READ)] = {
81 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
82 [C(RESULT_MISS)] = XSCALE_PERFCTR_ICACHE_MISS,
83 },
84 [C(OP_WRITE)] = {
85 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
86 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
87 },
88 [C(OP_PREFETCH)] = {
89 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
90 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
91 },
92 },
93 [C(LL)] = {
94 [C(OP_READ)] = {
95 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
96 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
97 },
98 [C(OP_WRITE)] = {
99 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
100 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
101 },
102 [C(OP_PREFETCH)] = {
103 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
104 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
105 },
106 },
107 [C(DTLB)] = {
108 [C(OP_READ)] = {
109 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
110 [C(RESULT_MISS)] = XSCALE_PERFCTR_DTLB_MISS,
111 },
112 [C(OP_WRITE)] = {
113 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
114 [C(RESULT_MISS)] = XSCALE_PERFCTR_DTLB_MISS,
115 },
116 [C(OP_PREFETCH)] = {
117 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
118 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
119 },
120 },
121 [C(ITLB)] = {
122 [C(OP_READ)] = {
123 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
124 [C(RESULT_MISS)] = XSCALE_PERFCTR_ITLB_MISS,
125 },
126 [C(OP_WRITE)] = {
127 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
128 [C(RESULT_MISS)] = XSCALE_PERFCTR_ITLB_MISS,
129 },
130 [C(OP_PREFETCH)] = {
131 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
132 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
133 },
134 },
135 [C(BPU)] = {
136 [C(OP_READ)] = {
137 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
138 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
139 },
140 [C(OP_WRITE)] = {
141 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
142 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
143 },
144 [C(OP_PREFETCH)] = {
145 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
146 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
147 },
148 },
149 [C(NODE)] = {
150 [C(OP_READ)] = {
151 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
152 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
153 },
154 [C(OP_WRITE)] = {
155 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
156 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
157 },
158 [C(OP_PREFETCH)] = {
159 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
160 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
161 },
162 },
163};
164
165#define XSCALE_PMU_ENABLE 0x001
166#define XSCALE_PMN_RESET 0x002
167#define XSCALE_CCNT_RESET 0x004
168#define XSCALE_PMU_RESET (CCNT_RESET | PMN_RESET)
169#define XSCALE_PMU_CNT64 0x008
170
171#define XSCALE1_OVERFLOWED_MASK 0x700
172#define XSCALE1_CCOUNT_OVERFLOW 0x400
173#define XSCALE1_COUNT0_OVERFLOW 0x100
174#define XSCALE1_COUNT1_OVERFLOW 0x200
175#define XSCALE1_CCOUNT_INT_EN 0x040
176#define XSCALE1_COUNT0_INT_EN 0x010
177#define XSCALE1_COUNT1_INT_EN 0x020
178#define XSCALE1_COUNT0_EVT_SHFT 12
179#define XSCALE1_COUNT0_EVT_MASK (0xff << XSCALE1_COUNT0_EVT_SHFT)
180#define XSCALE1_COUNT1_EVT_SHFT 20
181#define XSCALE1_COUNT1_EVT_MASK (0xff << XSCALE1_COUNT1_EVT_SHFT)
182
183static inline u32
184xscale1pmu_read_pmnc(void)
185{
186 u32 val;
187 asm volatile("mrc p14, 0, %0, c0, c0, 0" : "=r" (val));
188 return val;
189}
190
191static inline void
192xscale1pmu_write_pmnc(u32 val)
193{
194 /* upper 4bits and 7, 11 are write-as-0 */
195 val &= 0xffff77f;
196 asm volatile("mcr p14, 0, %0, c0, c0, 0" : : "r" (val));
197}
198
199static inline int
200xscale1_pmnc_counter_has_overflowed(unsigned long pmnc,
201 enum xscale_counters counter)
202{
203 int ret = 0;
204
205 switch (counter) {
206 case XSCALE_CYCLE_COUNTER:
207 ret = pmnc & XSCALE1_CCOUNT_OVERFLOW;
208 break;
209 case XSCALE_COUNTER0:
210 ret = pmnc & XSCALE1_COUNT0_OVERFLOW;
211 break;
212 case XSCALE_COUNTER1:
213 ret = pmnc & XSCALE1_COUNT1_OVERFLOW;
214 break;
215 default:
216 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
217 }
218
219 return ret;
220}
221
222static irqreturn_t
223xscale1pmu_handle_irq(int irq_num, void *dev)
224{
225 unsigned long pmnc;
226 struct perf_sample_data data;
227 struct arm_pmu *cpu_pmu = (struct arm_pmu *)dev;
228 struct pmu_hw_events *cpuc = cpu_pmu->get_hw_events();
229 struct pt_regs *regs;
230 int idx;
231
232 /*
233 * NOTE: there's an A stepping erratum that states if an overflow
234 * bit already exists and another occurs, the previous
235 * Overflow bit gets cleared. There's no workaround.
236 * Fixed in B stepping or later.
237 */
238 pmnc = xscale1pmu_read_pmnc();
239
240 /*
241 * Write the value back to clear the overflow flags. Overflow
242 * flags remain in pmnc for use below. We also disable the PMU
243 * while we process the interrupt.
244 */
245 xscale1pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
246
247 if (!(pmnc & XSCALE1_OVERFLOWED_MASK))
248 return IRQ_NONE;
249
250 regs = get_irq_regs();
251
252 for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
253 struct perf_event *event = cpuc->events[idx];
254 struct hw_perf_event *hwc;
255
256 if (!event)
257 continue;
258
259 if (!xscale1_pmnc_counter_has_overflowed(pmnc, idx))
260 continue;
261
262 hwc = &event->hw;
263 armpmu_event_update(event);
264 perf_sample_data_init(&data, 0, hwc->last_period);
265 if (!armpmu_event_set_period(event))
266 continue;
267
268 if (perf_event_overflow(event, &data, regs))
269 cpu_pmu->disable(event);
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 xscale1pmu_enable_event(struct perf_event *event)
284{
285 unsigned long val, mask, evt, flags;
286 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
287 struct hw_perf_event *hwc = &event->hw;
288 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
289 int idx = hwc->idx;
290
291 switch (idx) {
292 case XSCALE_CYCLE_COUNTER:
293 mask = 0;
294 evt = XSCALE1_CCOUNT_INT_EN;
295 break;
296 case XSCALE_COUNTER0:
297 mask = XSCALE1_COUNT0_EVT_MASK;
298 evt = (hwc->config_base << XSCALE1_COUNT0_EVT_SHFT) |
299 XSCALE1_COUNT0_INT_EN;
300 break;
301 case XSCALE_COUNTER1:
302 mask = XSCALE1_COUNT1_EVT_MASK;
303 evt = (hwc->config_base << XSCALE1_COUNT1_EVT_SHFT) |
304 XSCALE1_COUNT1_INT_EN;
305 break;
306 default:
307 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
308 return;
309 }
310
311 raw_spin_lock_irqsave(&events->pmu_lock, flags);
312 val = xscale1pmu_read_pmnc();
313 val &= ~mask;
314 val |= evt;
315 xscale1pmu_write_pmnc(val);
316 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
317}
318
319static void xscale1pmu_disable_event(struct perf_event *event)
320{
321 unsigned long val, mask, evt, flags;
322 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
323 struct hw_perf_event *hwc = &event->hw;
324 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
325 int idx = hwc->idx;
326
327 switch (idx) {
328 case XSCALE_CYCLE_COUNTER:
329 mask = XSCALE1_CCOUNT_INT_EN;
330 evt = 0;
331 break;
332 case XSCALE_COUNTER0:
333 mask = XSCALE1_COUNT0_INT_EN | XSCALE1_COUNT0_EVT_MASK;
334 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT0_EVT_SHFT;
335 break;
336 case XSCALE_COUNTER1:
337 mask = XSCALE1_COUNT1_INT_EN | XSCALE1_COUNT1_EVT_MASK;
338 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT1_EVT_SHFT;
339 break;
340 default:
341 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
342 return;
343 }
344
345 raw_spin_lock_irqsave(&events->pmu_lock, flags);
346 val = xscale1pmu_read_pmnc();
347 val &= ~mask;
348 val |= evt;
349 xscale1pmu_write_pmnc(val);
350 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
351}
352
353static int
354xscale1pmu_get_event_idx(struct pmu_hw_events *cpuc,
355 struct perf_event *event)
356{
357 struct hw_perf_event *hwc = &event->hw;
358 if (XSCALE_PERFCTR_CCNT == hwc->config_base) {
359 if (test_and_set_bit(XSCALE_CYCLE_COUNTER, cpuc->used_mask))
360 return -EAGAIN;
361
362 return XSCALE_CYCLE_COUNTER;
363 } else {
364 if (!test_and_set_bit(XSCALE_COUNTER1, cpuc->used_mask))
365 return XSCALE_COUNTER1;
366
367 if (!test_and_set_bit(XSCALE_COUNTER0, cpuc->used_mask))
368 return XSCALE_COUNTER0;
369
370 return -EAGAIN;
371 }
372}
373
374static void xscale1pmu_start(struct arm_pmu *cpu_pmu)
375{
376 unsigned long flags, val;
377 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
378
379 raw_spin_lock_irqsave(&events->pmu_lock, flags);
380 val = xscale1pmu_read_pmnc();
381 val |= XSCALE_PMU_ENABLE;
382 xscale1pmu_write_pmnc(val);
383 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
384}
385
386static void xscale1pmu_stop(struct arm_pmu *cpu_pmu)
387{
388 unsigned long flags, val;
389 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
390
391 raw_spin_lock_irqsave(&events->pmu_lock, flags);
392 val = xscale1pmu_read_pmnc();
393 val &= ~XSCALE_PMU_ENABLE;
394 xscale1pmu_write_pmnc(val);
395 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
396}
397
398static inline u32 xscale1pmu_read_counter(struct perf_event *event)
399{
400 struct hw_perf_event *hwc = &event->hw;
401 int counter = hwc->idx;
402 u32 val = 0;
403
404 switch (counter) {
405 case XSCALE_CYCLE_COUNTER:
406 asm volatile("mrc p14, 0, %0, c1, c0, 0" : "=r" (val));
407 break;
408 case XSCALE_COUNTER0:
409 asm volatile("mrc p14, 0, %0, c2, c0, 0" : "=r" (val));
410 break;
411 case XSCALE_COUNTER1:
412 asm volatile("mrc p14, 0, %0, c3, c0, 0" : "=r" (val));
413 break;
414 }
415
416 return val;
417}
418
419static inline void xscale1pmu_write_counter(struct perf_event *event, u32 val)
420{
421 struct hw_perf_event *hwc = &event->hw;
422 int counter = hwc->idx;
423
424 switch (counter) {
425 case XSCALE_CYCLE_COUNTER:
426 asm volatile("mcr p14, 0, %0, c1, c0, 0" : : "r" (val));
427 break;
428 case XSCALE_COUNTER0:
429 asm volatile("mcr p14, 0, %0, c2, c0, 0" : : "r" (val));
430 break;
431 case XSCALE_COUNTER1:
432 asm volatile("mcr p14, 0, %0, c3, c0, 0" : : "r" (val));
433 break;
434 }
435}
436
437static int xscale_map_event(struct perf_event *event)
438{
439 return armpmu_map_event(event, &xscale_perf_map,
440 &xscale_perf_cache_map, 0xFF);
441}
442
443static int xscale1pmu_init(struct arm_pmu *cpu_pmu)
444{
445 cpu_pmu->name = "xscale1";
446 cpu_pmu->handle_irq = xscale1pmu_handle_irq;
447 cpu_pmu->enable = xscale1pmu_enable_event;
448 cpu_pmu->disable = xscale1pmu_disable_event;
449 cpu_pmu->read_counter = xscale1pmu_read_counter;
450 cpu_pmu->write_counter = xscale1pmu_write_counter;
451 cpu_pmu->get_event_idx = xscale1pmu_get_event_idx;
452 cpu_pmu->start = xscale1pmu_start;
453 cpu_pmu->stop = xscale1pmu_stop;
454 cpu_pmu->map_event = xscale_map_event;
455 cpu_pmu->num_events = 3;
456 cpu_pmu->max_period = (1LLU << 32) - 1;
457
458 return 0;
459}
460
461#define XSCALE2_OVERFLOWED_MASK 0x01f
462#define XSCALE2_CCOUNT_OVERFLOW 0x001
463#define XSCALE2_COUNT0_OVERFLOW 0x002
464#define XSCALE2_COUNT1_OVERFLOW 0x004
465#define XSCALE2_COUNT2_OVERFLOW 0x008
466#define XSCALE2_COUNT3_OVERFLOW 0x010
467#define XSCALE2_CCOUNT_INT_EN 0x001
468#define XSCALE2_COUNT0_INT_EN 0x002
469#define XSCALE2_COUNT1_INT_EN 0x004
470#define XSCALE2_COUNT2_INT_EN 0x008
471#define XSCALE2_COUNT3_INT_EN 0x010
472#define XSCALE2_COUNT0_EVT_SHFT 0
473#define XSCALE2_COUNT0_EVT_MASK (0xff << XSCALE2_COUNT0_EVT_SHFT)
474#define XSCALE2_COUNT1_EVT_SHFT 8
475#define XSCALE2_COUNT1_EVT_MASK (0xff << XSCALE2_COUNT1_EVT_SHFT)
476#define XSCALE2_COUNT2_EVT_SHFT 16
477#define XSCALE2_COUNT2_EVT_MASK (0xff << XSCALE2_COUNT2_EVT_SHFT)
478#define XSCALE2_COUNT3_EVT_SHFT 24
479#define XSCALE2_COUNT3_EVT_MASK (0xff << XSCALE2_COUNT3_EVT_SHFT)
480
481static inline u32
482xscale2pmu_read_pmnc(void)
483{
484 u32 val;
485 asm volatile("mrc p14, 0, %0, c0, c1, 0" : "=r" (val));
486 /* bits 1-2 and 4-23 are read-unpredictable */
487 return val & 0xff000009;
488}
489
490static inline void
491xscale2pmu_write_pmnc(u32 val)
492{
493 /* bits 4-23 are write-as-0, 24-31 are write ignored */
494 val &= 0xf;
495 asm volatile("mcr p14, 0, %0, c0, c1, 0" : : "r" (val));
496}
497
498static inline u32
499xscale2pmu_read_overflow_flags(void)
500{
501 u32 val;
502 asm volatile("mrc p14, 0, %0, c5, c1, 0" : "=r" (val));
503 return val;
504}
505
506static inline void
507xscale2pmu_write_overflow_flags(u32 val)
508{
509 asm volatile("mcr p14, 0, %0, c5, c1, 0" : : "r" (val));
510}
511
512static inline u32
513xscale2pmu_read_event_select(void)
514{
515 u32 val;
516 asm volatile("mrc p14, 0, %0, c8, c1, 0" : "=r" (val));
517 return val;
518}
519
520static inline void
521xscale2pmu_write_event_select(u32 val)
522{
523 asm volatile("mcr p14, 0, %0, c8, c1, 0" : : "r"(val));
524}
525
526static inline u32
527xscale2pmu_read_int_enable(void)
528{
529 u32 val;
530 asm volatile("mrc p14, 0, %0, c4, c1, 0" : "=r" (val));
531 return val;
532}
533
534static void
535xscale2pmu_write_int_enable(u32 val)
536{
537 asm volatile("mcr p14, 0, %0, c4, c1, 0" : : "r" (val));
538}
539
540static inline int
541xscale2_pmnc_counter_has_overflowed(unsigned long of_flags,
542 enum xscale_counters counter)
543{
544 int ret = 0;
545
546 switch (counter) {
547 case XSCALE_CYCLE_COUNTER:
548 ret = of_flags & XSCALE2_CCOUNT_OVERFLOW;
549 break;
550 case XSCALE_COUNTER0:
551 ret = of_flags & XSCALE2_COUNT0_OVERFLOW;
552 break;
553 case XSCALE_COUNTER1:
554 ret = of_flags & XSCALE2_COUNT1_OVERFLOW;
555 break;
556 case XSCALE_COUNTER2:
557 ret = of_flags & XSCALE2_COUNT2_OVERFLOW;
558 break;
559 case XSCALE_COUNTER3:
560 ret = of_flags & XSCALE2_COUNT3_OVERFLOW;
561 break;
562 default:
563 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
564 }
565
566 return ret;
567}
568
569static irqreturn_t
570xscale2pmu_handle_irq(int irq_num, void *dev)
571{
572 unsigned long pmnc, of_flags;
573 struct perf_sample_data data;
574 struct arm_pmu *cpu_pmu = (struct arm_pmu *)dev;
575 struct pmu_hw_events *cpuc = cpu_pmu->get_hw_events();
576 struct pt_regs *regs;
577 int idx;
578
579 /* Disable the PMU. */
580 pmnc = xscale2pmu_read_pmnc();
581 xscale2pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
582
583 /* Check the overflow flag register. */
584 of_flags = xscale2pmu_read_overflow_flags();
585 if (!(of_flags & XSCALE2_OVERFLOWED_MASK))
586 return IRQ_NONE;
587
588 /* Clear the overflow bits. */
589 xscale2pmu_write_overflow_flags(of_flags);
590
591 regs = get_irq_regs();
592
593 for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
594 struct perf_event *event = cpuc->events[idx];
595 struct hw_perf_event *hwc;
596
597 if (!event)
598 continue;
599
600 if (!xscale2_pmnc_counter_has_overflowed(of_flags, idx))
601 continue;
602
603 hwc = &event->hw;
604 armpmu_event_update(event);
605 perf_sample_data_init(&data, 0, hwc->last_period);
606 if (!armpmu_event_set_period(event))
607 continue;
608
609 if (perf_event_overflow(event, &data, regs))
610 cpu_pmu->disable(event);
611 }
612
613 irq_work_run();
614
615 /*
616 * Re-enable the PMU.
617 */
618 pmnc = xscale2pmu_read_pmnc() | XSCALE_PMU_ENABLE;
619 xscale2pmu_write_pmnc(pmnc);
620
621 return IRQ_HANDLED;
622}
623
624static void xscale2pmu_enable_event(struct perf_event *event)
625{
626 unsigned long flags, ien, evtsel;
627 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
628 struct hw_perf_event *hwc = &event->hw;
629 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
630 int idx = hwc->idx;
631
632 ien = xscale2pmu_read_int_enable();
633 evtsel = xscale2pmu_read_event_select();
634
635 switch (idx) {
636 case XSCALE_CYCLE_COUNTER:
637 ien |= XSCALE2_CCOUNT_INT_EN;
638 break;
639 case XSCALE_COUNTER0:
640 ien |= XSCALE2_COUNT0_INT_EN;
641 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
642 evtsel |= hwc->config_base << XSCALE2_COUNT0_EVT_SHFT;
643 break;
644 case XSCALE_COUNTER1:
645 ien |= XSCALE2_COUNT1_INT_EN;
646 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
647 evtsel |= hwc->config_base << XSCALE2_COUNT1_EVT_SHFT;
648 break;
649 case XSCALE_COUNTER2:
650 ien |= XSCALE2_COUNT2_INT_EN;
651 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
652 evtsel |= hwc->config_base << XSCALE2_COUNT2_EVT_SHFT;
653 break;
654 case XSCALE_COUNTER3:
655 ien |= XSCALE2_COUNT3_INT_EN;
656 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
657 evtsel |= hwc->config_base << XSCALE2_COUNT3_EVT_SHFT;
658 break;
659 default:
660 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
661 return;
662 }
663
664 raw_spin_lock_irqsave(&events->pmu_lock, flags);
665 xscale2pmu_write_event_select(evtsel);
666 xscale2pmu_write_int_enable(ien);
667 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
668}
669
670static void xscale2pmu_disable_event(struct perf_event *event)
671{
672 unsigned long flags, ien, evtsel, of_flags;
673 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
674 struct hw_perf_event *hwc = &event->hw;
675 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
676 int idx = hwc->idx;
677
678 ien = xscale2pmu_read_int_enable();
679 evtsel = xscale2pmu_read_event_select();
680
681 switch (idx) {
682 case XSCALE_CYCLE_COUNTER:
683 ien &= ~XSCALE2_CCOUNT_INT_EN;
684 of_flags = XSCALE2_CCOUNT_OVERFLOW;
685 break;
686 case XSCALE_COUNTER0:
687 ien &= ~XSCALE2_COUNT0_INT_EN;
688 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
689 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT0_EVT_SHFT;
690 of_flags = XSCALE2_COUNT0_OVERFLOW;
691 break;
692 case XSCALE_COUNTER1:
693 ien &= ~XSCALE2_COUNT1_INT_EN;
694 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
695 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT1_EVT_SHFT;
696 of_flags = XSCALE2_COUNT1_OVERFLOW;
697 break;
698 case XSCALE_COUNTER2:
699 ien &= ~XSCALE2_COUNT2_INT_EN;
700 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
701 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT2_EVT_SHFT;
702 of_flags = XSCALE2_COUNT2_OVERFLOW;
703 break;
704 case XSCALE_COUNTER3:
705 ien &= ~XSCALE2_COUNT3_INT_EN;
706 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
707 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT3_EVT_SHFT;
708 of_flags = XSCALE2_COUNT3_OVERFLOW;
709 break;
710 default:
711 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
712 return;
713 }
714
715 raw_spin_lock_irqsave(&events->pmu_lock, flags);
716 xscale2pmu_write_event_select(evtsel);
717 xscale2pmu_write_int_enable(ien);
718 xscale2pmu_write_overflow_flags(of_flags);
719 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
720}
721
722static int
723xscale2pmu_get_event_idx(struct pmu_hw_events *cpuc,
724 struct perf_event *event)
725{
726 int idx = xscale1pmu_get_event_idx(cpuc, event);
727 if (idx >= 0)
728 goto out;
729
730 if (!test_and_set_bit(XSCALE_COUNTER3, cpuc->used_mask))
731 idx = XSCALE_COUNTER3;
732 else if (!test_and_set_bit(XSCALE_COUNTER2, cpuc->used_mask))
733 idx = XSCALE_COUNTER2;
734out:
735 return idx;
736}
737
738static void xscale2pmu_start(struct arm_pmu *cpu_pmu)
739{
740 unsigned long flags, val;
741 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
742
743 raw_spin_lock_irqsave(&events->pmu_lock, flags);
744 val = xscale2pmu_read_pmnc() & ~XSCALE_PMU_CNT64;
745 val |= XSCALE_PMU_ENABLE;
746 xscale2pmu_write_pmnc(val);
747 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
748}
749
750static void xscale2pmu_stop(struct arm_pmu *cpu_pmu)
751{
752 unsigned long flags, val;
753 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
754
755 raw_spin_lock_irqsave(&events->pmu_lock, flags);
756 val = xscale2pmu_read_pmnc();
757 val &= ~XSCALE_PMU_ENABLE;
758 xscale2pmu_write_pmnc(val);
759 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
760}
761
762static inline u32 xscale2pmu_read_counter(struct perf_event *event)
763{
764 struct hw_perf_event *hwc = &event->hw;
765 int counter = hwc->idx;
766 u32 val = 0;
767
768 switch (counter) {
769 case XSCALE_CYCLE_COUNTER:
770 asm volatile("mrc p14, 0, %0, c1, c1, 0" : "=r" (val));
771 break;
772 case XSCALE_COUNTER0:
773 asm volatile("mrc p14, 0, %0, c0, c2, 0" : "=r" (val));
774 break;
775 case XSCALE_COUNTER1:
776 asm volatile("mrc p14, 0, %0, c1, c2, 0" : "=r" (val));
777 break;
778 case XSCALE_COUNTER2:
779 asm volatile("mrc p14, 0, %0, c2, c2, 0" : "=r" (val));
780 break;
781 case XSCALE_COUNTER3:
782 asm volatile("mrc p14, 0, %0, c3, c2, 0" : "=r" (val));
783 break;
784 }
785
786 return val;
787}
788
789static inline void xscale2pmu_write_counter(struct perf_event *event, u32 val)
790{
791 struct hw_perf_event *hwc = &event->hw;
792 int counter = hwc->idx;
793
794 switch (counter) {
795 case XSCALE_CYCLE_COUNTER:
796 asm volatile("mcr p14, 0, %0, c1, c1, 0" : : "r" (val));
797 break;
798 case XSCALE_COUNTER0:
799 asm volatile("mcr p14, 0, %0, c0, c2, 0" : : "r" (val));
800 break;
801 case XSCALE_COUNTER1:
802 asm volatile("mcr p14, 0, %0, c1, c2, 0" : : "r" (val));
803 break;
804 case XSCALE_COUNTER2:
805 asm volatile("mcr p14, 0, %0, c2, c2, 0" : : "r" (val));
806 break;
807 case XSCALE_COUNTER3:
808 asm volatile("mcr p14, 0, %0, c3, c2, 0" : : "r" (val));
809 break;
810 }
811}
812
813static int xscale2pmu_init(struct arm_pmu *cpu_pmu)
814{
815 cpu_pmu->name = "xscale2";
816 cpu_pmu->handle_irq = xscale2pmu_handle_irq;
817 cpu_pmu->enable = xscale2pmu_enable_event;
818 cpu_pmu->disable = xscale2pmu_disable_event;
819 cpu_pmu->read_counter = xscale2pmu_read_counter;
820 cpu_pmu->write_counter = xscale2pmu_write_counter;
821 cpu_pmu->get_event_idx = xscale2pmu_get_event_idx;
822 cpu_pmu->start = xscale2pmu_start;
823 cpu_pmu->stop = xscale2pmu_stop;
824 cpu_pmu->map_event = xscale_map_event;
825 cpu_pmu->num_events = 5;
826 cpu_pmu->max_period = (1LLU << 32) - 1;
827
828 return 0;
829}
830#else
831static inline int xscale1pmu_init(struct arm_pmu *cpu_pmu)
832{
833 return -ENODEV;
834}
835
836static inline int xscale2pmu_init(struct arm_pmu *cpu_pmu)
837{
838 return -ENODEV;
839}
840#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(struct arm_pmu *cpu_pmu)
146{
147 unsigned long pmnc;
148 struct perf_sample_data data;
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;
207 struct hw_perf_event *hwc = &event->hw;
208 int idx = hwc->idx;
209
210 switch (idx) {
211 case XSCALE_CYCLE_COUNTER:
212 mask = 0;
213 evt = XSCALE1_CCOUNT_INT_EN;
214 break;
215 case XSCALE_COUNTER0:
216 mask = XSCALE1_COUNT0_EVT_MASK;
217 evt = (hwc->config_base << XSCALE1_COUNT0_EVT_SHFT) |
218 XSCALE1_COUNT0_INT_EN;
219 break;
220 case XSCALE_COUNTER1:
221 mask = XSCALE1_COUNT1_EVT_MASK;
222 evt = (hwc->config_base << XSCALE1_COUNT1_EVT_SHFT) |
223 XSCALE1_COUNT1_INT_EN;
224 break;
225 default:
226 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
227 return;
228 }
229
230 val = xscale1pmu_read_pmnc();
231 val &= ~mask;
232 val |= evt;
233 xscale1pmu_write_pmnc(val);
234}
235
236static void xscale1pmu_disable_event(struct perf_event *event)
237{
238 unsigned long val, mask, evt;
239 struct hw_perf_event *hwc = &event->hw;
240 int idx = hwc->idx;
241
242 switch (idx) {
243 case XSCALE_CYCLE_COUNTER:
244 mask = XSCALE1_CCOUNT_INT_EN;
245 evt = 0;
246 break;
247 case XSCALE_COUNTER0:
248 mask = XSCALE1_COUNT0_INT_EN | XSCALE1_COUNT0_EVT_MASK;
249 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT0_EVT_SHFT;
250 break;
251 case XSCALE_COUNTER1:
252 mask = XSCALE1_COUNT1_INT_EN | XSCALE1_COUNT1_EVT_MASK;
253 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT1_EVT_SHFT;
254 break;
255 default:
256 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
257 return;
258 }
259
260 val = xscale1pmu_read_pmnc();
261 val &= ~mask;
262 val |= evt;
263 xscale1pmu_write_pmnc(val);
264}
265
266static int
267xscale1pmu_get_event_idx(struct pmu_hw_events *cpuc,
268 struct perf_event *event)
269{
270 struct hw_perf_event *hwc = &event->hw;
271 if (XSCALE_PERFCTR_CCNT == hwc->config_base) {
272 if (test_and_set_bit(XSCALE_CYCLE_COUNTER, cpuc->used_mask))
273 return -EAGAIN;
274
275 return XSCALE_CYCLE_COUNTER;
276 } else {
277 if (!test_and_set_bit(XSCALE_COUNTER1, cpuc->used_mask))
278 return XSCALE_COUNTER1;
279
280 if (!test_and_set_bit(XSCALE_COUNTER0, cpuc->used_mask))
281 return XSCALE_COUNTER0;
282
283 return -EAGAIN;
284 }
285}
286
287static void xscalepmu_clear_event_idx(struct pmu_hw_events *cpuc,
288 struct perf_event *event)
289{
290 clear_bit(event->hw.idx, cpuc->used_mask);
291}
292
293static void xscale1pmu_start(struct arm_pmu *cpu_pmu)
294{
295 unsigned long val;
296
297 val = xscale1pmu_read_pmnc();
298 val |= XSCALE_PMU_ENABLE;
299 xscale1pmu_write_pmnc(val);
300}
301
302static void xscale1pmu_stop(struct arm_pmu *cpu_pmu)
303{
304 unsigned long val;
305
306 val = xscale1pmu_read_pmnc();
307 val &= ~XSCALE_PMU_ENABLE;
308 xscale1pmu_write_pmnc(val);
309}
310
311static inline u64 xscale1pmu_read_counter(struct perf_event *event)
312{
313 struct hw_perf_event *hwc = &event->hw;
314 int counter = hwc->idx;
315 u32 val = 0;
316
317 switch (counter) {
318 case XSCALE_CYCLE_COUNTER:
319 asm volatile("mrc p14, 0, %0, c1, c0, 0" : "=r" (val));
320 break;
321 case XSCALE_COUNTER0:
322 asm volatile("mrc p14, 0, %0, c2, c0, 0" : "=r" (val));
323 break;
324 case XSCALE_COUNTER1:
325 asm volatile("mrc p14, 0, %0, c3, c0, 0" : "=r" (val));
326 break;
327 }
328
329 return val;
330}
331
332static inline void xscale1pmu_write_counter(struct perf_event *event, u64 val)
333{
334 struct hw_perf_event *hwc = &event->hw;
335 int counter = hwc->idx;
336
337 switch (counter) {
338 case XSCALE_CYCLE_COUNTER:
339 asm volatile("mcr p14, 0, %0, c1, c0, 0" : : "r" (val));
340 break;
341 case XSCALE_COUNTER0:
342 asm volatile("mcr p14, 0, %0, c2, c0, 0" : : "r" (val));
343 break;
344 case XSCALE_COUNTER1:
345 asm volatile("mcr p14, 0, %0, c3, c0, 0" : : "r" (val));
346 break;
347 }
348}
349
350static int xscale_map_event(struct perf_event *event)
351{
352 return armpmu_map_event(event, &xscale_perf_map,
353 &xscale_perf_cache_map, 0xFF);
354}
355
356static int xscale1pmu_init(struct arm_pmu *cpu_pmu)
357{
358 cpu_pmu->name = "armv5_xscale1";
359 cpu_pmu->handle_irq = xscale1pmu_handle_irq;
360 cpu_pmu->enable = xscale1pmu_enable_event;
361 cpu_pmu->disable = xscale1pmu_disable_event;
362 cpu_pmu->read_counter = xscale1pmu_read_counter;
363 cpu_pmu->write_counter = xscale1pmu_write_counter;
364 cpu_pmu->get_event_idx = xscale1pmu_get_event_idx;
365 cpu_pmu->clear_event_idx = xscalepmu_clear_event_idx;
366 cpu_pmu->start = xscale1pmu_start;
367 cpu_pmu->stop = xscale1pmu_stop;
368 cpu_pmu->map_event = xscale_map_event;
369 cpu_pmu->num_events = 3;
370
371 return 0;
372}
373
374#define XSCALE2_OVERFLOWED_MASK 0x01f
375#define XSCALE2_CCOUNT_OVERFLOW 0x001
376#define XSCALE2_COUNT0_OVERFLOW 0x002
377#define XSCALE2_COUNT1_OVERFLOW 0x004
378#define XSCALE2_COUNT2_OVERFLOW 0x008
379#define XSCALE2_COUNT3_OVERFLOW 0x010
380#define XSCALE2_CCOUNT_INT_EN 0x001
381#define XSCALE2_COUNT0_INT_EN 0x002
382#define XSCALE2_COUNT1_INT_EN 0x004
383#define XSCALE2_COUNT2_INT_EN 0x008
384#define XSCALE2_COUNT3_INT_EN 0x010
385#define XSCALE2_COUNT0_EVT_SHFT 0
386#define XSCALE2_COUNT0_EVT_MASK (0xff << XSCALE2_COUNT0_EVT_SHFT)
387#define XSCALE2_COUNT1_EVT_SHFT 8
388#define XSCALE2_COUNT1_EVT_MASK (0xff << XSCALE2_COUNT1_EVT_SHFT)
389#define XSCALE2_COUNT2_EVT_SHFT 16
390#define XSCALE2_COUNT2_EVT_MASK (0xff << XSCALE2_COUNT2_EVT_SHFT)
391#define XSCALE2_COUNT3_EVT_SHFT 24
392#define XSCALE2_COUNT3_EVT_MASK (0xff << XSCALE2_COUNT3_EVT_SHFT)
393
394static inline u32
395xscale2pmu_read_pmnc(void)
396{
397 u32 val;
398 asm volatile("mrc p14, 0, %0, c0, c1, 0" : "=r" (val));
399 /* bits 1-2 and 4-23 are read-unpredictable */
400 return val & 0xff000009;
401}
402
403static inline void
404xscale2pmu_write_pmnc(u32 val)
405{
406 /* bits 4-23 are write-as-0, 24-31 are write ignored */
407 val &= 0xf;
408 asm volatile("mcr p14, 0, %0, c0, c1, 0" : : "r" (val));
409}
410
411static inline u32
412xscale2pmu_read_overflow_flags(void)
413{
414 u32 val;
415 asm volatile("mrc p14, 0, %0, c5, c1, 0" : "=r" (val));
416 return val;
417}
418
419static inline void
420xscale2pmu_write_overflow_flags(u32 val)
421{
422 asm volatile("mcr p14, 0, %0, c5, c1, 0" : : "r" (val));
423}
424
425static inline u32
426xscale2pmu_read_event_select(void)
427{
428 u32 val;
429 asm volatile("mrc p14, 0, %0, c8, c1, 0" : "=r" (val));
430 return val;
431}
432
433static inline void
434xscale2pmu_write_event_select(u32 val)
435{
436 asm volatile("mcr p14, 0, %0, c8, c1, 0" : : "r"(val));
437}
438
439static inline u32
440xscale2pmu_read_int_enable(void)
441{
442 u32 val;
443 asm volatile("mrc p14, 0, %0, c4, c1, 0" : "=r" (val));
444 return val;
445}
446
447static void
448xscale2pmu_write_int_enable(u32 val)
449{
450 asm volatile("mcr p14, 0, %0, c4, c1, 0" : : "r" (val));
451}
452
453static inline int
454xscale2_pmnc_counter_has_overflowed(unsigned long of_flags,
455 enum xscale_counters counter)
456{
457 int ret = 0;
458
459 switch (counter) {
460 case XSCALE_CYCLE_COUNTER:
461 ret = of_flags & XSCALE2_CCOUNT_OVERFLOW;
462 break;
463 case XSCALE_COUNTER0:
464 ret = of_flags & XSCALE2_COUNT0_OVERFLOW;
465 break;
466 case XSCALE_COUNTER1:
467 ret = of_flags & XSCALE2_COUNT1_OVERFLOW;
468 break;
469 case XSCALE_COUNTER2:
470 ret = of_flags & XSCALE2_COUNT2_OVERFLOW;
471 break;
472 case XSCALE_COUNTER3:
473 ret = of_flags & XSCALE2_COUNT3_OVERFLOW;
474 break;
475 default:
476 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
477 }
478
479 return ret;
480}
481
482static irqreturn_t
483xscale2pmu_handle_irq(struct arm_pmu *cpu_pmu)
484{
485 unsigned long pmnc, of_flags;
486 struct perf_sample_data data;
487 struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
488 struct pt_regs *regs;
489 int idx;
490
491 /* Disable the PMU. */
492 pmnc = xscale2pmu_read_pmnc();
493 xscale2pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
494
495 /* Check the overflow flag register. */
496 of_flags = xscale2pmu_read_overflow_flags();
497 if (!(of_flags & XSCALE2_OVERFLOWED_MASK))
498 return IRQ_NONE;
499
500 /* Clear the overflow bits. */
501 xscale2pmu_write_overflow_flags(of_flags);
502
503 regs = get_irq_regs();
504
505 for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
506 struct perf_event *event = cpuc->events[idx];
507 struct hw_perf_event *hwc;
508
509 if (!event)
510 continue;
511
512 if (!xscale2_pmnc_counter_has_overflowed(of_flags, idx))
513 continue;
514
515 hwc = &event->hw;
516 armpmu_event_update(event);
517 perf_sample_data_init(&data, 0, hwc->last_period);
518 if (!armpmu_event_set_period(event))
519 continue;
520
521 if (perf_event_overflow(event, &data, regs))
522 cpu_pmu->disable(event);
523 }
524
525 irq_work_run();
526
527 /*
528 * Re-enable the PMU.
529 */
530 pmnc = xscale2pmu_read_pmnc() | XSCALE_PMU_ENABLE;
531 xscale2pmu_write_pmnc(pmnc);
532
533 return IRQ_HANDLED;
534}
535
536static void xscale2pmu_enable_event(struct perf_event *event)
537{
538 unsigned long ien, evtsel;
539 struct hw_perf_event *hwc = &event->hw;
540 int idx = hwc->idx;
541
542 ien = xscale2pmu_read_int_enable();
543 evtsel = xscale2pmu_read_event_select();
544
545 switch (idx) {
546 case XSCALE_CYCLE_COUNTER:
547 ien |= XSCALE2_CCOUNT_INT_EN;
548 break;
549 case XSCALE_COUNTER0:
550 ien |= XSCALE2_COUNT0_INT_EN;
551 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
552 evtsel |= hwc->config_base << XSCALE2_COUNT0_EVT_SHFT;
553 break;
554 case XSCALE_COUNTER1:
555 ien |= XSCALE2_COUNT1_INT_EN;
556 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
557 evtsel |= hwc->config_base << XSCALE2_COUNT1_EVT_SHFT;
558 break;
559 case XSCALE_COUNTER2:
560 ien |= XSCALE2_COUNT2_INT_EN;
561 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
562 evtsel |= hwc->config_base << XSCALE2_COUNT2_EVT_SHFT;
563 break;
564 case XSCALE_COUNTER3:
565 ien |= XSCALE2_COUNT3_INT_EN;
566 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
567 evtsel |= hwc->config_base << XSCALE2_COUNT3_EVT_SHFT;
568 break;
569 default:
570 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
571 return;
572 }
573
574 xscale2pmu_write_event_select(evtsel);
575 xscale2pmu_write_int_enable(ien);
576}
577
578static void xscale2pmu_disable_event(struct perf_event *event)
579{
580 unsigned long ien, evtsel, of_flags;
581 struct hw_perf_event *hwc = &event->hw;
582 int idx = hwc->idx;
583
584 ien = xscale2pmu_read_int_enable();
585 evtsel = xscale2pmu_read_event_select();
586
587 switch (idx) {
588 case XSCALE_CYCLE_COUNTER:
589 ien &= ~XSCALE2_CCOUNT_INT_EN;
590 of_flags = XSCALE2_CCOUNT_OVERFLOW;
591 break;
592 case XSCALE_COUNTER0:
593 ien &= ~XSCALE2_COUNT0_INT_EN;
594 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
595 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT0_EVT_SHFT;
596 of_flags = XSCALE2_COUNT0_OVERFLOW;
597 break;
598 case XSCALE_COUNTER1:
599 ien &= ~XSCALE2_COUNT1_INT_EN;
600 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
601 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT1_EVT_SHFT;
602 of_flags = XSCALE2_COUNT1_OVERFLOW;
603 break;
604 case XSCALE_COUNTER2:
605 ien &= ~XSCALE2_COUNT2_INT_EN;
606 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
607 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT2_EVT_SHFT;
608 of_flags = XSCALE2_COUNT2_OVERFLOW;
609 break;
610 case XSCALE_COUNTER3:
611 ien &= ~XSCALE2_COUNT3_INT_EN;
612 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
613 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT3_EVT_SHFT;
614 of_flags = XSCALE2_COUNT3_OVERFLOW;
615 break;
616 default:
617 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
618 return;
619 }
620
621 xscale2pmu_write_event_select(evtsel);
622 xscale2pmu_write_int_enable(ien);
623 xscale2pmu_write_overflow_flags(of_flags);
624}
625
626static int
627xscale2pmu_get_event_idx(struct pmu_hw_events *cpuc,
628 struct perf_event *event)
629{
630 int idx = xscale1pmu_get_event_idx(cpuc, event);
631 if (idx >= 0)
632 goto out;
633
634 if (!test_and_set_bit(XSCALE_COUNTER3, cpuc->used_mask))
635 idx = XSCALE_COUNTER3;
636 else if (!test_and_set_bit(XSCALE_COUNTER2, cpuc->used_mask))
637 idx = XSCALE_COUNTER2;
638out:
639 return idx;
640}
641
642static void xscale2pmu_start(struct arm_pmu *cpu_pmu)
643{
644 unsigned long val;
645
646 val = xscale2pmu_read_pmnc() & ~XSCALE_PMU_CNT64;
647 val |= XSCALE_PMU_ENABLE;
648 xscale2pmu_write_pmnc(val);
649}
650
651static void xscale2pmu_stop(struct arm_pmu *cpu_pmu)
652{
653 unsigned long val;
654
655 val = xscale2pmu_read_pmnc();
656 val &= ~XSCALE_PMU_ENABLE;
657 xscale2pmu_write_pmnc(val);
658}
659
660static inline u64 xscale2pmu_read_counter(struct perf_event *event)
661{
662 struct hw_perf_event *hwc = &event->hw;
663 int counter = hwc->idx;
664 u32 val = 0;
665
666 switch (counter) {
667 case XSCALE_CYCLE_COUNTER:
668 asm volatile("mrc p14, 0, %0, c1, c1, 0" : "=r" (val));
669 break;
670 case XSCALE_COUNTER0:
671 asm volatile("mrc p14, 0, %0, c0, c2, 0" : "=r" (val));
672 break;
673 case XSCALE_COUNTER1:
674 asm volatile("mrc p14, 0, %0, c1, c2, 0" : "=r" (val));
675 break;
676 case XSCALE_COUNTER2:
677 asm volatile("mrc p14, 0, %0, c2, c2, 0" : "=r" (val));
678 break;
679 case XSCALE_COUNTER3:
680 asm volatile("mrc p14, 0, %0, c3, c2, 0" : "=r" (val));
681 break;
682 }
683
684 return val;
685}
686
687static inline void xscale2pmu_write_counter(struct perf_event *event, u64 val)
688{
689 struct hw_perf_event *hwc = &event->hw;
690 int counter = hwc->idx;
691
692 switch (counter) {
693 case XSCALE_CYCLE_COUNTER:
694 asm volatile("mcr p14, 0, %0, c1, c1, 0" : : "r" (val));
695 break;
696 case XSCALE_COUNTER0:
697 asm volatile("mcr p14, 0, %0, c0, c2, 0" : : "r" (val));
698 break;
699 case XSCALE_COUNTER1:
700 asm volatile("mcr p14, 0, %0, c1, c2, 0" : : "r" (val));
701 break;
702 case XSCALE_COUNTER2:
703 asm volatile("mcr p14, 0, %0, c2, c2, 0" : : "r" (val));
704 break;
705 case XSCALE_COUNTER3:
706 asm volatile("mcr p14, 0, %0, c3, c2, 0" : : "r" (val));
707 break;
708 }
709}
710
711static int xscale2pmu_init(struct arm_pmu *cpu_pmu)
712{
713 cpu_pmu->name = "armv5_xscale2";
714 cpu_pmu->handle_irq = xscale2pmu_handle_irq;
715 cpu_pmu->enable = xscale2pmu_enable_event;
716 cpu_pmu->disable = xscale2pmu_disable_event;
717 cpu_pmu->read_counter = xscale2pmu_read_counter;
718 cpu_pmu->write_counter = xscale2pmu_write_counter;
719 cpu_pmu->get_event_idx = xscale2pmu_get_event_idx;
720 cpu_pmu->clear_event_idx = xscalepmu_clear_event_idx;
721 cpu_pmu->start = xscale2pmu_start;
722 cpu_pmu->stop = xscale2pmu_stop;
723 cpu_pmu->map_event = xscale_map_event;
724 cpu_pmu->num_events = 5;
725
726 return 0;
727}
728
729static const struct pmu_probe_info xscale_pmu_probe_table[] = {
730 XSCALE_PMU_PROBE(ARM_CPU_XSCALE_ARCH_V1, xscale1pmu_init),
731 XSCALE_PMU_PROBE(ARM_CPU_XSCALE_ARCH_V2, xscale2pmu_init),
732 { /* sentinel value */ }
733};
734
735static int xscale_pmu_device_probe(struct platform_device *pdev)
736{
737 return arm_pmu_device_probe(pdev, NULL, xscale_pmu_probe_table);
738}
739
740static struct platform_driver xscale_pmu_driver = {
741 .driver = {
742 .name = "xscale-pmu",
743 },
744 .probe = xscale_pmu_device_probe,
745};
746
747builtin_platform_driver(xscale_pmu_driver);
748#endif /* CONFIG_CPU_XSCALE */