Loading...
1/*
2 * ARMv6 Performance counter handling code.
3 *
4 * Copyright (C) 2009 picoChip Designs, Ltd., Jamie Iles
5 *
6 * ARMv6 has 2 configurable performance counters and a single cycle counter.
7 * They all share a single reset bit but can be written to zero so we can use
8 * that for a reset.
9 *
10 * The counters can't be individually enabled or disabled so when we remove
11 * one event and replace it with another we could get spurious counts from the
12 * wrong event. However, we can take advantage of the fact that the
13 * performance counters can export events to the event bus, and the event bus
14 * itself can be monitored. This requires that we *don't* export the events to
15 * the event bus. The procedure for disabling a configurable counter is:
16 * - change the counter to count the ETMEXTOUT[0] signal (0x20). This
17 * effectively stops the counter from counting.
18 * - disable the counter's interrupt generation (each counter has it's
19 * own interrupt enable bit).
20 * Once stopped, the counter value can be written as 0 to reset.
21 *
22 * To enable a counter:
23 * - enable the counter's interrupt generation.
24 * - set the new event type.
25 *
26 * Note: the dedicated cycle counter only counts cycles and can't be
27 * enabled/disabled independently of the others. When we want to disable the
28 * cycle counter, we have to just disable the interrupt reporting and start
29 * ignoring that counter. When re-enabling, we have to reset the value and
30 * enable the interrupt.
31 */
32
33#if defined(CONFIG_CPU_V6) || defined(CONFIG_CPU_V6K)
34enum armv6_perf_types {
35 ARMV6_PERFCTR_ICACHE_MISS = 0x0,
36 ARMV6_PERFCTR_IBUF_STALL = 0x1,
37 ARMV6_PERFCTR_DDEP_STALL = 0x2,
38 ARMV6_PERFCTR_ITLB_MISS = 0x3,
39 ARMV6_PERFCTR_DTLB_MISS = 0x4,
40 ARMV6_PERFCTR_BR_EXEC = 0x5,
41 ARMV6_PERFCTR_BR_MISPREDICT = 0x6,
42 ARMV6_PERFCTR_INSTR_EXEC = 0x7,
43 ARMV6_PERFCTR_DCACHE_HIT = 0x9,
44 ARMV6_PERFCTR_DCACHE_ACCESS = 0xA,
45 ARMV6_PERFCTR_DCACHE_MISS = 0xB,
46 ARMV6_PERFCTR_DCACHE_WBACK = 0xC,
47 ARMV6_PERFCTR_SW_PC_CHANGE = 0xD,
48 ARMV6_PERFCTR_MAIN_TLB_MISS = 0xF,
49 ARMV6_PERFCTR_EXPL_D_ACCESS = 0x10,
50 ARMV6_PERFCTR_LSU_FULL_STALL = 0x11,
51 ARMV6_PERFCTR_WBUF_DRAINED = 0x12,
52 ARMV6_PERFCTR_CPU_CYCLES = 0xFF,
53 ARMV6_PERFCTR_NOP = 0x20,
54};
55
56enum armv6_counters {
57 ARMV6_CYCLE_COUNTER = 1,
58 ARMV6_COUNTER0,
59 ARMV6_COUNTER1,
60};
61
62/*
63 * The hardware events that we support. We do support cache operations but
64 * we have harvard caches and no way to combine instruction and data
65 * accesses/misses in hardware.
66 */
67static const unsigned armv6_perf_map[PERF_COUNT_HW_MAX] = {
68 [PERF_COUNT_HW_CPU_CYCLES] = ARMV6_PERFCTR_CPU_CYCLES,
69 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV6_PERFCTR_INSTR_EXEC,
70 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
71 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
72 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV6_PERFCTR_BR_EXEC,
73 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV6_PERFCTR_BR_MISPREDICT,
74 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
75};
76
77static const unsigned armv6_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
78 [PERF_COUNT_HW_CACHE_OP_MAX]
79 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
80 [C(L1D)] = {
81 /*
82 * The performance counters don't differentiate between read
83 * and write accesses/misses so this isn't strictly correct,
84 * but it's the best we can do. Writes and reads get
85 * combined.
86 */
87 [C(OP_READ)] = {
88 [C(RESULT_ACCESS)] = ARMV6_PERFCTR_DCACHE_ACCESS,
89 [C(RESULT_MISS)] = ARMV6_PERFCTR_DCACHE_MISS,
90 },
91 [C(OP_WRITE)] = {
92 [C(RESULT_ACCESS)] = ARMV6_PERFCTR_DCACHE_ACCESS,
93 [C(RESULT_MISS)] = ARMV6_PERFCTR_DCACHE_MISS,
94 },
95 [C(OP_PREFETCH)] = {
96 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
97 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
98 },
99 },
100 [C(L1I)] = {
101 [C(OP_READ)] = {
102 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
103 [C(RESULT_MISS)] = ARMV6_PERFCTR_ICACHE_MISS,
104 },
105 [C(OP_WRITE)] = {
106 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
107 [C(RESULT_MISS)] = ARMV6_PERFCTR_ICACHE_MISS,
108 },
109 [C(OP_PREFETCH)] = {
110 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
111 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
112 },
113 },
114 [C(LL)] = {
115 [C(OP_READ)] = {
116 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
117 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
118 },
119 [C(OP_WRITE)] = {
120 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
121 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
122 },
123 [C(OP_PREFETCH)] = {
124 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
125 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
126 },
127 },
128 [C(DTLB)] = {
129 /*
130 * The ARM performance counters can count micro DTLB misses,
131 * micro ITLB misses and main TLB misses. There isn't an event
132 * for TLB misses, so use the micro misses here and if users
133 * want the main TLB misses they can use a raw counter.
134 */
135 [C(OP_READ)] = {
136 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
137 [C(RESULT_MISS)] = ARMV6_PERFCTR_DTLB_MISS,
138 },
139 [C(OP_WRITE)] = {
140 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
141 [C(RESULT_MISS)] = ARMV6_PERFCTR_DTLB_MISS,
142 },
143 [C(OP_PREFETCH)] = {
144 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
145 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
146 },
147 },
148 [C(ITLB)] = {
149 [C(OP_READ)] = {
150 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
151 [C(RESULT_MISS)] = ARMV6_PERFCTR_ITLB_MISS,
152 },
153 [C(OP_WRITE)] = {
154 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
155 [C(RESULT_MISS)] = ARMV6_PERFCTR_ITLB_MISS,
156 },
157 [C(OP_PREFETCH)] = {
158 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
159 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
160 },
161 },
162 [C(BPU)] = {
163 [C(OP_READ)] = {
164 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
165 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
166 },
167 [C(OP_WRITE)] = {
168 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
169 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
170 },
171 [C(OP_PREFETCH)] = {
172 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
173 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
174 },
175 },
176 [C(NODE)] = {
177 [C(OP_READ)] = {
178 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
179 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
180 },
181 [C(OP_WRITE)] = {
182 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
183 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
184 },
185 [C(OP_PREFETCH)] = {
186 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
187 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
188 },
189 },
190};
191
192enum armv6mpcore_perf_types {
193 ARMV6MPCORE_PERFCTR_ICACHE_MISS = 0x0,
194 ARMV6MPCORE_PERFCTR_IBUF_STALL = 0x1,
195 ARMV6MPCORE_PERFCTR_DDEP_STALL = 0x2,
196 ARMV6MPCORE_PERFCTR_ITLB_MISS = 0x3,
197 ARMV6MPCORE_PERFCTR_DTLB_MISS = 0x4,
198 ARMV6MPCORE_PERFCTR_BR_EXEC = 0x5,
199 ARMV6MPCORE_PERFCTR_BR_NOTPREDICT = 0x6,
200 ARMV6MPCORE_PERFCTR_BR_MISPREDICT = 0x7,
201 ARMV6MPCORE_PERFCTR_INSTR_EXEC = 0x8,
202 ARMV6MPCORE_PERFCTR_DCACHE_RDACCESS = 0xA,
203 ARMV6MPCORE_PERFCTR_DCACHE_RDMISS = 0xB,
204 ARMV6MPCORE_PERFCTR_DCACHE_WRACCESS = 0xC,
205 ARMV6MPCORE_PERFCTR_DCACHE_WRMISS = 0xD,
206 ARMV6MPCORE_PERFCTR_DCACHE_EVICTION = 0xE,
207 ARMV6MPCORE_PERFCTR_SW_PC_CHANGE = 0xF,
208 ARMV6MPCORE_PERFCTR_MAIN_TLB_MISS = 0x10,
209 ARMV6MPCORE_PERFCTR_EXPL_MEM_ACCESS = 0x11,
210 ARMV6MPCORE_PERFCTR_LSU_FULL_STALL = 0x12,
211 ARMV6MPCORE_PERFCTR_WBUF_DRAINED = 0x13,
212 ARMV6MPCORE_PERFCTR_CPU_CYCLES = 0xFF,
213};
214
215/*
216 * The hardware events that we support. We do support cache operations but
217 * we have harvard caches and no way to combine instruction and data
218 * accesses/misses in hardware.
219 */
220static const unsigned armv6mpcore_perf_map[PERF_COUNT_HW_MAX] = {
221 [PERF_COUNT_HW_CPU_CYCLES] = ARMV6MPCORE_PERFCTR_CPU_CYCLES,
222 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV6MPCORE_PERFCTR_INSTR_EXEC,
223 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
224 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
225 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV6MPCORE_PERFCTR_BR_EXEC,
226 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV6MPCORE_PERFCTR_BR_MISPREDICT,
227 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
228};
229
230static const unsigned armv6mpcore_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
231 [PERF_COUNT_HW_CACHE_OP_MAX]
232 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
233 [C(L1D)] = {
234 [C(OP_READ)] = {
235 [C(RESULT_ACCESS)] =
236 ARMV6MPCORE_PERFCTR_DCACHE_RDACCESS,
237 [C(RESULT_MISS)] =
238 ARMV6MPCORE_PERFCTR_DCACHE_RDMISS,
239 },
240 [C(OP_WRITE)] = {
241 [C(RESULT_ACCESS)] =
242 ARMV6MPCORE_PERFCTR_DCACHE_WRACCESS,
243 [C(RESULT_MISS)] =
244 ARMV6MPCORE_PERFCTR_DCACHE_WRMISS,
245 },
246 [C(OP_PREFETCH)] = {
247 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
248 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
249 },
250 },
251 [C(L1I)] = {
252 [C(OP_READ)] = {
253 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
254 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ICACHE_MISS,
255 },
256 [C(OP_WRITE)] = {
257 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
258 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ICACHE_MISS,
259 },
260 [C(OP_PREFETCH)] = {
261 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
262 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
263 },
264 },
265 [C(LL)] = {
266 [C(OP_READ)] = {
267 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
268 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
269 },
270 [C(OP_WRITE)] = {
271 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
272 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
273 },
274 [C(OP_PREFETCH)] = {
275 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
276 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
277 },
278 },
279 [C(DTLB)] = {
280 /*
281 * The ARM performance counters can count micro DTLB misses,
282 * micro ITLB misses and main TLB misses. There isn't an event
283 * for TLB misses, so use the micro misses here and if users
284 * want the main TLB misses they can use a raw counter.
285 */
286 [C(OP_READ)] = {
287 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
288 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_DTLB_MISS,
289 },
290 [C(OP_WRITE)] = {
291 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
292 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_DTLB_MISS,
293 },
294 [C(OP_PREFETCH)] = {
295 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
296 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
297 },
298 },
299 [C(ITLB)] = {
300 [C(OP_READ)] = {
301 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
302 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ITLB_MISS,
303 },
304 [C(OP_WRITE)] = {
305 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
306 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ITLB_MISS,
307 },
308 [C(OP_PREFETCH)] = {
309 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
310 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
311 },
312 },
313 [C(BPU)] = {
314 [C(OP_READ)] = {
315 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
316 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
317 },
318 [C(OP_WRITE)] = {
319 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
320 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
321 },
322 [C(OP_PREFETCH)] = {
323 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
324 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
325 },
326 },
327 [C(NODE)] = {
328 [C(OP_READ)] = {
329 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
330 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
331 },
332 [C(OP_WRITE)] = {
333 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
334 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
335 },
336 [C(OP_PREFETCH)] = {
337 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
338 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
339 },
340 },
341};
342
343static inline unsigned long
344armv6_pmcr_read(void)
345{
346 u32 val;
347 asm volatile("mrc p15, 0, %0, c15, c12, 0" : "=r"(val));
348 return val;
349}
350
351static inline void
352armv6_pmcr_write(unsigned long val)
353{
354 asm volatile("mcr p15, 0, %0, c15, c12, 0" : : "r"(val));
355}
356
357#define ARMV6_PMCR_ENABLE (1 << 0)
358#define ARMV6_PMCR_CTR01_RESET (1 << 1)
359#define ARMV6_PMCR_CCOUNT_RESET (1 << 2)
360#define ARMV6_PMCR_CCOUNT_DIV (1 << 3)
361#define ARMV6_PMCR_COUNT0_IEN (1 << 4)
362#define ARMV6_PMCR_COUNT1_IEN (1 << 5)
363#define ARMV6_PMCR_CCOUNT_IEN (1 << 6)
364#define ARMV6_PMCR_COUNT0_OVERFLOW (1 << 8)
365#define ARMV6_PMCR_COUNT1_OVERFLOW (1 << 9)
366#define ARMV6_PMCR_CCOUNT_OVERFLOW (1 << 10)
367#define ARMV6_PMCR_EVT_COUNT0_SHIFT 20
368#define ARMV6_PMCR_EVT_COUNT0_MASK (0xFF << ARMV6_PMCR_EVT_COUNT0_SHIFT)
369#define ARMV6_PMCR_EVT_COUNT1_SHIFT 12
370#define ARMV6_PMCR_EVT_COUNT1_MASK (0xFF << ARMV6_PMCR_EVT_COUNT1_SHIFT)
371
372#define ARMV6_PMCR_OVERFLOWED_MASK \
373 (ARMV6_PMCR_COUNT0_OVERFLOW | ARMV6_PMCR_COUNT1_OVERFLOW | \
374 ARMV6_PMCR_CCOUNT_OVERFLOW)
375
376static inline int
377armv6_pmcr_has_overflowed(unsigned long pmcr)
378{
379 return pmcr & ARMV6_PMCR_OVERFLOWED_MASK;
380}
381
382static inline int
383armv6_pmcr_counter_has_overflowed(unsigned long pmcr,
384 enum armv6_counters counter)
385{
386 int ret = 0;
387
388 if (ARMV6_CYCLE_COUNTER == counter)
389 ret = pmcr & ARMV6_PMCR_CCOUNT_OVERFLOW;
390 else if (ARMV6_COUNTER0 == counter)
391 ret = pmcr & ARMV6_PMCR_COUNT0_OVERFLOW;
392 else if (ARMV6_COUNTER1 == counter)
393 ret = pmcr & ARMV6_PMCR_COUNT1_OVERFLOW;
394 else
395 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
396
397 return ret;
398}
399
400static inline u32
401armv6pmu_read_counter(int counter)
402{
403 unsigned long value = 0;
404
405 if (ARMV6_CYCLE_COUNTER == counter)
406 asm volatile("mrc p15, 0, %0, c15, c12, 1" : "=r"(value));
407 else if (ARMV6_COUNTER0 == counter)
408 asm volatile("mrc p15, 0, %0, c15, c12, 2" : "=r"(value));
409 else if (ARMV6_COUNTER1 == counter)
410 asm volatile("mrc p15, 0, %0, c15, c12, 3" : "=r"(value));
411 else
412 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
413
414 return value;
415}
416
417static inline void
418armv6pmu_write_counter(int counter,
419 u32 value)
420{
421 if (ARMV6_CYCLE_COUNTER == counter)
422 asm volatile("mcr p15, 0, %0, c15, c12, 1" : : "r"(value));
423 else if (ARMV6_COUNTER0 == counter)
424 asm volatile("mcr p15, 0, %0, c15, c12, 2" : : "r"(value));
425 else if (ARMV6_COUNTER1 == counter)
426 asm volatile("mcr p15, 0, %0, c15, c12, 3" : : "r"(value));
427 else
428 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
429}
430
431static void
432armv6pmu_enable_event(struct hw_perf_event *hwc,
433 int idx)
434{
435 unsigned long val, mask, evt, flags;
436
437 if (ARMV6_CYCLE_COUNTER == idx) {
438 mask = 0;
439 evt = ARMV6_PMCR_CCOUNT_IEN;
440 } else if (ARMV6_COUNTER0 == idx) {
441 mask = ARMV6_PMCR_EVT_COUNT0_MASK;
442 evt = (hwc->config_base << ARMV6_PMCR_EVT_COUNT0_SHIFT) |
443 ARMV6_PMCR_COUNT0_IEN;
444 } else if (ARMV6_COUNTER1 == idx) {
445 mask = ARMV6_PMCR_EVT_COUNT1_MASK;
446 evt = (hwc->config_base << ARMV6_PMCR_EVT_COUNT1_SHIFT) |
447 ARMV6_PMCR_COUNT1_IEN;
448 } else {
449 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
450 return;
451 }
452
453 /*
454 * Mask out the current event and set the counter to count the event
455 * that we're interested in.
456 */
457 raw_spin_lock_irqsave(&pmu_lock, flags);
458 val = armv6_pmcr_read();
459 val &= ~mask;
460 val |= evt;
461 armv6_pmcr_write(val);
462 raw_spin_unlock_irqrestore(&pmu_lock, flags);
463}
464
465static irqreturn_t
466armv6pmu_handle_irq(int irq_num,
467 void *dev)
468{
469 unsigned long pmcr = armv6_pmcr_read();
470 struct perf_sample_data data;
471 struct cpu_hw_events *cpuc;
472 struct pt_regs *regs;
473 int idx;
474
475 if (!armv6_pmcr_has_overflowed(pmcr))
476 return IRQ_NONE;
477
478 regs = get_irq_regs();
479
480 /*
481 * The interrupts are cleared by writing the overflow flags back to
482 * the control register. All of the other bits don't have any effect
483 * if they are rewritten, so write the whole value back.
484 */
485 armv6_pmcr_write(pmcr);
486
487 perf_sample_data_init(&data, 0);
488
489 cpuc = &__get_cpu_var(cpu_hw_events);
490 for (idx = 0; idx <= armpmu->num_events; ++idx) {
491 struct perf_event *event = cpuc->events[idx];
492 struct hw_perf_event *hwc;
493
494 if (!test_bit(idx, cpuc->active_mask))
495 continue;
496
497 /*
498 * We have a single interrupt for all counters. Check that
499 * each counter has overflowed before we process it.
500 */
501 if (!armv6_pmcr_counter_has_overflowed(pmcr, idx))
502 continue;
503
504 hwc = &event->hw;
505 armpmu_event_update(event, hwc, idx, 1);
506 data.period = event->hw.last_period;
507 if (!armpmu_event_set_period(event, hwc, idx))
508 continue;
509
510 if (perf_event_overflow(event, &data, regs))
511 armpmu->disable(hwc, idx);
512 }
513
514 /*
515 * Handle the pending perf events.
516 *
517 * Note: this call *must* be run with interrupts disabled. For
518 * platforms that can have the PMU interrupts raised as an NMI, this
519 * will not work.
520 */
521 irq_work_run();
522
523 return IRQ_HANDLED;
524}
525
526static void
527armv6pmu_start(void)
528{
529 unsigned long flags, val;
530
531 raw_spin_lock_irqsave(&pmu_lock, flags);
532 val = armv6_pmcr_read();
533 val |= ARMV6_PMCR_ENABLE;
534 armv6_pmcr_write(val);
535 raw_spin_unlock_irqrestore(&pmu_lock, flags);
536}
537
538static void
539armv6pmu_stop(void)
540{
541 unsigned long flags, val;
542
543 raw_spin_lock_irqsave(&pmu_lock, flags);
544 val = armv6_pmcr_read();
545 val &= ~ARMV6_PMCR_ENABLE;
546 armv6_pmcr_write(val);
547 raw_spin_unlock_irqrestore(&pmu_lock, flags);
548}
549
550static int
551armv6pmu_get_event_idx(struct cpu_hw_events *cpuc,
552 struct hw_perf_event *event)
553{
554 /* Always place a cycle counter into the cycle counter. */
555 if (ARMV6_PERFCTR_CPU_CYCLES == event->config_base) {
556 if (test_and_set_bit(ARMV6_CYCLE_COUNTER, cpuc->used_mask))
557 return -EAGAIN;
558
559 return ARMV6_CYCLE_COUNTER;
560 } else {
561 /*
562 * For anything other than a cycle counter, try and use
563 * counter0 and counter1.
564 */
565 if (!test_and_set_bit(ARMV6_COUNTER1, cpuc->used_mask))
566 return ARMV6_COUNTER1;
567
568 if (!test_and_set_bit(ARMV6_COUNTER0, cpuc->used_mask))
569 return ARMV6_COUNTER0;
570
571 /* The counters are all in use. */
572 return -EAGAIN;
573 }
574}
575
576static void
577armv6pmu_disable_event(struct hw_perf_event *hwc,
578 int idx)
579{
580 unsigned long val, mask, evt, flags;
581
582 if (ARMV6_CYCLE_COUNTER == idx) {
583 mask = ARMV6_PMCR_CCOUNT_IEN;
584 evt = 0;
585 } else if (ARMV6_COUNTER0 == idx) {
586 mask = ARMV6_PMCR_COUNT0_IEN | ARMV6_PMCR_EVT_COUNT0_MASK;
587 evt = ARMV6_PERFCTR_NOP << ARMV6_PMCR_EVT_COUNT0_SHIFT;
588 } else if (ARMV6_COUNTER1 == idx) {
589 mask = ARMV6_PMCR_COUNT1_IEN | ARMV6_PMCR_EVT_COUNT1_MASK;
590 evt = ARMV6_PERFCTR_NOP << ARMV6_PMCR_EVT_COUNT1_SHIFT;
591 } else {
592 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
593 return;
594 }
595
596 /*
597 * Mask out the current event and set the counter to count the number
598 * of ETM bus signal assertion cycles. The external reporting should
599 * be disabled and so this should never increment.
600 */
601 raw_spin_lock_irqsave(&pmu_lock, flags);
602 val = armv6_pmcr_read();
603 val &= ~mask;
604 val |= evt;
605 armv6_pmcr_write(val);
606 raw_spin_unlock_irqrestore(&pmu_lock, flags);
607}
608
609static void
610armv6mpcore_pmu_disable_event(struct hw_perf_event *hwc,
611 int idx)
612{
613 unsigned long val, mask, flags, evt = 0;
614
615 if (ARMV6_CYCLE_COUNTER == idx) {
616 mask = ARMV6_PMCR_CCOUNT_IEN;
617 } else if (ARMV6_COUNTER0 == idx) {
618 mask = ARMV6_PMCR_COUNT0_IEN;
619 } else if (ARMV6_COUNTER1 == idx) {
620 mask = ARMV6_PMCR_COUNT1_IEN;
621 } else {
622 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
623 return;
624 }
625
626 /*
627 * Unlike UP ARMv6, we don't have a way of stopping the counters. We
628 * simply disable the interrupt reporting.
629 */
630 raw_spin_lock_irqsave(&pmu_lock, flags);
631 val = armv6_pmcr_read();
632 val &= ~mask;
633 val |= evt;
634 armv6_pmcr_write(val);
635 raw_spin_unlock_irqrestore(&pmu_lock, flags);
636}
637
638static const struct arm_pmu armv6pmu = {
639 .id = ARM_PERF_PMU_ID_V6,
640 .name = "v6",
641 .handle_irq = armv6pmu_handle_irq,
642 .enable = armv6pmu_enable_event,
643 .disable = armv6pmu_disable_event,
644 .read_counter = armv6pmu_read_counter,
645 .write_counter = armv6pmu_write_counter,
646 .get_event_idx = armv6pmu_get_event_idx,
647 .start = armv6pmu_start,
648 .stop = armv6pmu_stop,
649 .cache_map = &armv6_perf_cache_map,
650 .event_map = &armv6_perf_map,
651 .raw_event_mask = 0xFF,
652 .num_events = 3,
653 .max_period = (1LLU << 32) - 1,
654};
655
656static const struct arm_pmu *__init armv6pmu_init(void)
657{
658 return &armv6pmu;
659}
660
661/*
662 * ARMv6mpcore is almost identical to single core ARMv6 with the exception
663 * that some of the events have different enumerations and that there is no
664 * *hack* to stop the programmable counters. To stop the counters we simply
665 * disable the interrupt reporting and update the event. When unthrottling we
666 * reset the period and enable the interrupt reporting.
667 */
668static const struct arm_pmu armv6mpcore_pmu = {
669 .id = ARM_PERF_PMU_ID_V6MP,
670 .name = "v6mpcore",
671 .handle_irq = armv6pmu_handle_irq,
672 .enable = armv6pmu_enable_event,
673 .disable = armv6mpcore_pmu_disable_event,
674 .read_counter = armv6pmu_read_counter,
675 .write_counter = armv6pmu_write_counter,
676 .get_event_idx = armv6pmu_get_event_idx,
677 .start = armv6pmu_start,
678 .stop = armv6pmu_stop,
679 .cache_map = &armv6mpcore_perf_cache_map,
680 .event_map = &armv6mpcore_perf_map,
681 .raw_event_mask = 0xFF,
682 .num_events = 3,
683 .max_period = (1LLU << 32) - 1,
684};
685
686static const struct arm_pmu *__init armv6mpcore_pmu_init(void)
687{
688 return &armv6mpcore_pmu;
689}
690#else
691static const struct arm_pmu *__init armv6pmu_init(void)
692{
693 return NULL;
694}
695
696static const struct arm_pmu *__init armv6mpcore_pmu_init(void)
697{
698 return NULL;
699}
700#endif /* CONFIG_CPU_V6 || CONFIG_CPU_V6K */
1/*
2 * ARMv6 Performance counter handling code.
3 *
4 * Copyright (C) 2009 picoChip Designs, Ltd., Jamie Iles
5 *
6 * ARMv6 has 2 configurable performance counters and a single cycle counter.
7 * They all share a single reset bit but can be written to zero so we can use
8 * that for a reset.
9 *
10 * The counters can't be individually enabled or disabled so when we remove
11 * one event and replace it with another we could get spurious counts from the
12 * wrong event. However, we can take advantage of the fact that the
13 * performance counters can export events to the event bus, and the event bus
14 * itself can be monitored. This requires that we *don't* export the events to
15 * the event bus. The procedure for disabling a configurable counter is:
16 * - change the counter to count the ETMEXTOUT[0] signal (0x20). This
17 * effectively stops the counter from counting.
18 * - disable the counter's interrupt generation (each counter has it's
19 * own interrupt enable bit).
20 * Once stopped, the counter value can be written as 0 to reset.
21 *
22 * To enable a counter:
23 * - enable the counter's interrupt generation.
24 * - set the new event type.
25 *
26 * Note: the dedicated cycle counter only counts cycles and can't be
27 * enabled/disabled independently of the others. When we want to disable the
28 * cycle counter, we have to just disable the interrupt reporting and start
29 * ignoring that counter. When re-enabling, we have to reset the value and
30 * enable the interrupt.
31 */
32
33#if defined(CONFIG_CPU_V6) || defined(CONFIG_CPU_V6K)
34enum armv6_perf_types {
35 ARMV6_PERFCTR_ICACHE_MISS = 0x0,
36 ARMV6_PERFCTR_IBUF_STALL = 0x1,
37 ARMV6_PERFCTR_DDEP_STALL = 0x2,
38 ARMV6_PERFCTR_ITLB_MISS = 0x3,
39 ARMV6_PERFCTR_DTLB_MISS = 0x4,
40 ARMV6_PERFCTR_BR_EXEC = 0x5,
41 ARMV6_PERFCTR_BR_MISPREDICT = 0x6,
42 ARMV6_PERFCTR_INSTR_EXEC = 0x7,
43 ARMV6_PERFCTR_DCACHE_HIT = 0x9,
44 ARMV6_PERFCTR_DCACHE_ACCESS = 0xA,
45 ARMV6_PERFCTR_DCACHE_MISS = 0xB,
46 ARMV6_PERFCTR_DCACHE_WBACK = 0xC,
47 ARMV6_PERFCTR_SW_PC_CHANGE = 0xD,
48 ARMV6_PERFCTR_MAIN_TLB_MISS = 0xF,
49 ARMV6_PERFCTR_EXPL_D_ACCESS = 0x10,
50 ARMV6_PERFCTR_LSU_FULL_STALL = 0x11,
51 ARMV6_PERFCTR_WBUF_DRAINED = 0x12,
52 ARMV6_PERFCTR_CPU_CYCLES = 0xFF,
53 ARMV6_PERFCTR_NOP = 0x20,
54};
55
56enum armv6_counters {
57 ARMV6_CYCLE_COUNTER = 0,
58 ARMV6_COUNTER0,
59 ARMV6_COUNTER1,
60};
61
62/*
63 * The hardware events that we support. We do support cache operations but
64 * we have harvard caches and no way to combine instruction and data
65 * accesses/misses in hardware.
66 */
67static const unsigned armv6_perf_map[PERF_COUNT_HW_MAX] = {
68 [PERF_COUNT_HW_CPU_CYCLES] = ARMV6_PERFCTR_CPU_CYCLES,
69 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV6_PERFCTR_INSTR_EXEC,
70 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
71 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
72 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV6_PERFCTR_BR_EXEC,
73 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV6_PERFCTR_BR_MISPREDICT,
74 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
75 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = ARMV6_PERFCTR_IBUF_STALL,
76 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = ARMV6_PERFCTR_LSU_FULL_STALL,
77};
78
79static const unsigned armv6_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
80 [PERF_COUNT_HW_CACHE_OP_MAX]
81 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
82 [C(L1D)] = {
83 /*
84 * The performance counters don't differentiate between read
85 * and write accesses/misses so this isn't strictly correct,
86 * but it's the best we can do. Writes and reads get
87 * combined.
88 */
89 [C(OP_READ)] = {
90 [C(RESULT_ACCESS)] = ARMV6_PERFCTR_DCACHE_ACCESS,
91 [C(RESULT_MISS)] = ARMV6_PERFCTR_DCACHE_MISS,
92 },
93 [C(OP_WRITE)] = {
94 [C(RESULT_ACCESS)] = ARMV6_PERFCTR_DCACHE_ACCESS,
95 [C(RESULT_MISS)] = ARMV6_PERFCTR_DCACHE_MISS,
96 },
97 [C(OP_PREFETCH)] = {
98 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
99 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
100 },
101 },
102 [C(L1I)] = {
103 [C(OP_READ)] = {
104 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
105 [C(RESULT_MISS)] = ARMV6_PERFCTR_ICACHE_MISS,
106 },
107 [C(OP_WRITE)] = {
108 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
109 [C(RESULT_MISS)] = ARMV6_PERFCTR_ICACHE_MISS,
110 },
111 [C(OP_PREFETCH)] = {
112 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
113 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
114 },
115 },
116 [C(LL)] = {
117 [C(OP_READ)] = {
118 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
119 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
120 },
121 [C(OP_WRITE)] = {
122 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
123 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
124 },
125 [C(OP_PREFETCH)] = {
126 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
127 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
128 },
129 },
130 [C(DTLB)] = {
131 /*
132 * The ARM performance counters can count micro DTLB misses,
133 * micro ITLB misses and main TLB misses. There isn't an event
134 * for TLB misses, so use the micro misses here and if users
135 * want the main TLB misses they can use a raw counter.
136 */
137 [C(OP_READ)] = {
138 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
139 [C(RESULT_MISS)] = ARMV6_PERFCTR_DTLB_MISS,
140 },
141 [C(OP_WRITE)] = {
142 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
143 [C(RESULT_MISS)] = ARMV6_PERFCTR_DTLB_MISS,
144 },
145 [C(OP_PREFETCH)] = {
146 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
147 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
148 },
149 },
150 [C(ITLB)] = {
151 [C(OP_READ)] = {
152 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
153 [C(RESULT_MISS)] = ARMV6_PERFCTR_ITLB_MISS,
154 },
155 [C(OP_WRITE)] = {
156 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
157 [C(RESULT_MISS)] = ARMV6_PERFCTR_ITLB_MISS,
158 },
159 [C(OP_PREFETCH)] = {
160 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
161 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
162 },
163 },
164 [C(BPU)] = {
165 [C(OP_READ)] = {
166 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
167 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
168 },
169 [C(OP_WRITE)] = {
170 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
171 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
172 },
173 [C(OP_PREFETCH)] = {
174 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
175 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
176 },
177 },
178 [C(NODE)] = {
179 [C(OP_READ)] = {
180 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
181 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
182 },
183 [C(OP_WRITE)] = {
184 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
185 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
186 },
187 [C(OP_PREFETCH)] = {
188 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
189 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
190 },
191 },
192};
193
194enum armv6mpcore_perf_types {
195 ARMV6MPCORE_PERFCTR_ICACHE_MISS = 0x0,
196 ARMV6MPCORE_PERFCTR_IBUF_STALL = 0x1,
197 ARMV6MPCORE_PERFCTR_DDEP_STALL = 0x2,
198 ARMV6MPCORE_PERFCTR_ITLB_MISS = 0x3,
199 ARMV6MPCORE_PERFCTR_DTLB_MISS = 0x4,
200 ARMV6MPCORE_PERFCTR_BR_EXEC = 0x5,
201 ARMV6MPCORE_PERFCTR_BR_NOTPREDICT = 0x6,
202 ARMV6MPCORE_PERFCTR_BR_MISPREDICT = 0x7,
203 ARMV6MPCORE_PERFCTR_INSTR_EXEC = 0x8,
204 ARMV6MPCORE_PERFCTR_DCACHE_RDACCESS = 0xA,
205 ARMV6MPCORE_PERFCTR_DCACHE_RDMISS = 0xB,
206 ARMV6MPCORE_PERFCTR_DCACHE_WRACCESS = 0xC,
207 ARMV6MPCORE_PERFCTR_DCACHE_WRMISS = 0xD,
208 ARMV6MPCORE_PERFCTR_DCACHE_EVICTION = 0xE,
209 ARMV6MPCORE_PERFCTR_SW_PC_CHANGE = 0xF,
210 ARMV6MPCORE_PERFCTR_MAIN_TLB_MISS = 0x10,
211 ARMV6MPCORE_PERFCTR_EXPL_MEM_ACCESS = 0x11,
212 ARMV6MPCORE_PERFCTR_LSU_FULL_STALL = 0x12,
213 ARMV6MPCORE_PERFCTR_WBUF_DRAINED = 0x13,
214 ARMV6MPCORE_PERFCTR_CPU_CYCLES = 0xFF,
215};
216
217/*
218 * The hardware events that we support. We do support cache operations but
219 * we have harvard caches and no way to combine instruction and data
220 * accesses/misses in hardware.
221 */
222static const unsigned armv6mpcore_perf_map[PERF_COUNT_HW_MAX] = {
223 [PERF_COUNT_HW_CPU_CYCLES] = ARMV6MPCORE_PERFCTR_CPU_CYCLES,
224 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV6MPCORE_PERFCTR_INSTR_EXEC,
225 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
226 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
227 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV6MPCORE_PERFCTR_BR_EXEC,
228 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV6MPCORE_PERFCTR_BR_MISPREDICT,
229 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
230 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = ARMV6MPCORE_PERFCTR_IBUF_STALL,
231 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = ARMV6MPCORE_PERFCTR_LSU_FULL_STALL,
232};
233
234static const unsigned armv6mpcore_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
235 [PERF_COUNT_HW_CACHE_OP_MAX]
236 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
237 [C(L1D)] = {
238 [C(OP_READ)] = {
239 [C(RESULT_ACCESS)] =
240 ARMV6MPCORE_PERFCTR_DCACHE_RDACCESS,
241 [C(RESULT_MISS)] =
242 ARMV6MPCORE_PERFCTR_DCACHE_RDMISS,
243 },
244 [C(OP_WRITE)] = {
245 [C(RESULT_ACCESS)] =
246 ARMV6MPCORE_PERFCTR_DCACHE_WRACCESS,
247 [C(RESULT_MISS)] =
248 ARMV6MPCORE_PERFCTR_DCACHE_WRMISS,
249 },
250 [C(OP_PREFETCH)] = {
251 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
252 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
253 },
254 },
255 [C(L1I)] = {
256 [C(OP_READ)] = {
257 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
258 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ICACHE_MISS,
259 },
260 [C(OP_WRITE)] = {
261 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
262 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ICACHE_MISS,
263 },
264 [C(OP_PREFETCH)] = {
265 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
266 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
267 },
268 },
269 [C(LL)] = {
270 [C(OP_READ)] = {
271 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
272 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
273 },
274 [C(OP_WRITE)] = {
275 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
276 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
277 },
278 [C(OP_PREFETCH)] = {
279 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
280 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
281 },
282 },
283 [C(DTLB)] = {
284 /*
285 * The ARM performance counters can count micro DTLB misses,
286 * micro ITLB misses and main TLB misses. There isn't an event
287 * for TLB misses, so use the micro misses here and if users
288 * want the main TLB misses they can use a raw counter.
289 */
290 [C(OP_READ)] = {
291 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
292 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_DTLB_MISS,
293 },
294 [C(OP_WRITE)] = {
295 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
296 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_DTLB_MISS,
297 },
298 [C(OP_PREFETCH)] = {
299 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
300 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
301 },
302 },
303 [C(ITLB)] = {
304 [C(OP_READ)] = {
305 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
306 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ITLB_MISS,
307 },
308 [C(OP_WRITE)] = {
309 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
310 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ITLB_MISS,
311 },
312 [C(OP_PREFETCH)] = {
313 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
314 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
315 },
316 },
317 [C(BPU)] = {
318 [C(OP_READ)] = {
319 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
320 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
321 },
322 [C(OP_WRITE)] = {
323 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
324 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
325 },
326 [C(OP_PREFETCH)] = {
327 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
328 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
329 },
330 },
331 [C(NODE)] = {
332 [C(OP_READ)] = {
333 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
334 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
335 },
336 [C(OP_WRITE)] = {
337 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
338 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
339 },
340 [C(OP_PREFETCH)] = {
341 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
342 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
343 },
344 },
345};
346
347static inline unsigned long
348armv6_pmcr_read(void)
349{
350 u32 val;
351 asm volatile("mrc p15, 0, %0, c15, c12, 0" : "=r"(val));
352 return val;
353}
354
355static inline void
356armv6_pmcr_write(unsigned long val)
357{
358 asm volatile("mcr p15, 0, %0, c15, c12, 0" : : "r"(val));
359}
360
361#define ARMV6_PMCR_ENABLE (1 << 0)
362#define ARMV6_PMCR_CTR01_RESET (1 << 1)
363#define ARMV6_PMCR_CCOUNT_RESET (1 << 2)
364#define ARMV6_PMCR_CCOUNT_DIV (1 << 3)
365#define ARMV6_PMCR_COUNT0_IEN (1 << 4)
366#define ARMV6_PMCR_COUNT1_IEN (1 << 5)
367#define ARMV6_PMCR_CCOUNT_IEN (1 << 6)
368#define ARMV6_PMCR_COUNT0_OVERFLOW (1 << 8)
369#define ARMV6_PMCR_COUNT1_OVERFLOW (1 << 9)
370#define ARMV6_PMCR_CCOUNT_OVERFLOW (1 << 10)
371#define ARMV6_PMCR_EVT_COUNT0_SHIFT 20
372#define ARMV6_PMCR_EVT_COUNT0_MASK (0xFF << ARMV6_PMCR_EVT_COUNT0_SHIFT)
373#define ARMV6_PMCR_EVT_COUNT1_SHIFT 12
374#define ARMV6_PMCR_EVT_COUNT1_MASK (0xFF << ARMV6_PMCR_EVT_COUNT1_SHIFT)
375
376#define ARMV6_PMCR_OVERFLOWED_MASK \
377 (ARMV6_PMCR_COUNT0_OVERFLOW | ARMV6_PMCR_COUNT1_OVERFLOW | \
378 ARMV6_PMCR_CCOUNT_OVERFLOW)
379
380static inline int
381armv6_pmcr_has_overflowed(unsigned long pmcr)
382{
383 return pmcr & ARMV6_PMCR_OVERFLOWED_MASK;
384}
385
386static inline int
387armv6_pmcr_counter_has_overflowed(unsigned long pmcr,
388 enum armv6_counters counter)
389{
390 int ret = 0;
391
392 if (ARMV6_CYCLE_COUNTER == counter)
393 ret = pmcr & ARMV6_PMCR_CCOUNT_OVERFLOW;
394 else if (ARMV6_COUNTER0 == counter)
395 ret = pmcr & ARMV6_PMCR_COUNT0_OVERFLOW;
396 else if (ARMV6_COUNTER1 == counter)
397 ret = pmcr & ARMV6_PMCR_COUNT1_OVERFLOW;
398 else
399 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
400
401 return ret;
402}
403
404static inline u32
405armv6pmu_read_counter(int counter)
406{
407 unsigned long value = 0;
408
409 if (ARMV6_CYCLE_COUNTER == counter)
410 asm volatile("mrc p15, 0, %0, c15, c12, 1" : "=r"(value));
411 else if (ARMV6_COUNTER0 == counter)
412 asm volatile("mrc p15, 0, %0, c15, c12, 2" : "=r"(value));
413 else if (ARMV6_COUNTER1 == counter)
414 asm volatile("mrc p15, 0, %0, c15, c12, 3" : "=r"(value));
415 else
416 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
417
418 return value;
419}
420
421static inline void
422armv6pmu_write_counter(int counter,
423 u32 value)
424{
425 if (ARMV6_CYCLE_COUNTER == counter)
426 asm volatile("mcr p15, 0, %0, c15, c12, 1" : : "r"(value));
427 else if (ARMV6_COUNTER0 == counter)
428 asm volatile("mcr p15, 0, %0, c15, c12, 2" : : "r"(value));
429 else if (ARMV6_COUNTER1 == counter)
430 asm volatile("mcr p15, 0, %0, c15, c12, 3" : : "r"(value));
431 else
432 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
433}
434
435static void
436armv6pmu_enable_event(struct hw_perf_event *hwc,
437 int idx)
438{
439 unsigned long val, mask, evt, flags;
440 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
441
442 if (ARMV6_CYCLE_COUNTER == idx) {
443 mask = 0;
444 evt = ARMV6_PMCR_CCOUNT_IEN;
445 } else if (ARMV6_COUNTER0 == idx) {
446 mask = ARMV6_PMCR_EVT_COUNT0_MASK;
447 evt = (hwc->config_base << ARMV6_PMCR_EVT_COUNT0_SHIFT) |
448 ARMV6_PMCR_COUNT0_IEN;
449 } else if (ARMV6_COUNTER1 == idx) {
450 mask = ARMV6_PMCR_EVT_COUNT1_MASK;
451 evt = (hwc->config_base << ARMV6_PMCR_EVT_COUNT1_SHIFT) |
452 ARMV6_PMCR_COUNT1_IEN;
453 } else {
454 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
455 return;
456 }
457
458 /*
459 * Mask out the current event and set the counter to count the event
460 * that we're interested in.
461 */
462 raw_spin_lock_irqsave(&events->pmu_lock, flags);
463 val = armv6_pmcr_read();
464 val &= ~mask;
465 val |= evt;
466 armv6_pmcr_write(val);
467 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
468}
469
470static irqreturn_t
471armv6pmu_handle_irq(int irq_num,
472 void *dev)
473{
474 unsigned long pmcr = armv6_pmcr_read();
475 struct perf_sample_data data;
476 struct pmu_hw_events *cpuc;
477 struct pt_regs *regs;
478 int idx;
479
480 if (!armv6_pmcr_has_overflowed(pmcr))
481 return IRQ_NONE;
482
483 regs = get_irq_regs();
484
485 /*
486 * The interrupts are cleared by writing the overflow flags back to
487 * the control register. All of the other bits don't have any effect
488 * if they are rewritten, so write the whole value back.
489 */
490 armv6_pmcr_write(pmcr);
491
492 cpuc = &__get_cpu_var(cpu_hw_events);
493 for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
494 struct perf_event *event = cpuc->events[idx];
495 struct hw_perf_event *hwc;
496
497 /* Ignore if we don't have an event. */
498 if (!event)
499 continue;
500
501 /*
502 * We have a single interrupt for all counters. Check that
503 * each counter has overflowed before we process it.
504 */
505 if (!armv6_pmcr_counter_has_overflowed(pmcr, idx))
506 continue;
507
508 hwc = &event->hw;
509 armpmu_event_update(event, hwc, idx);
510 perf_sample_data_init(&data, 0, hwc->last_period);
511 if (!armpmu_event_set_period(event, hwc, idx))
512 continue;
513
514 if (perf_event_overflow(event, &data, regs))
515 cpu_pmu->disable(hwc, idx);
516 }
517
518 /*
519 * Handle the pending perf events.
520 *
521 * Note: this call *must* be run with interrupts disabled. For
522 * platforms that can have the PMU interrupts raised as an NMI, this
523 * will not work.
524 */
525 irq_work_run();
526
527 return IRQ_HANDLED;
528}
529
530static void
531armv6pmu_start(void)
532{
533 unsigned long flags, val;
534 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
535
536 raw_spin_lock_irqsave(&events->pmu_lock, flags);
537 val = armv6_pmcr_read();
538 val |= ARMV6_PMCR_ENABLE;
539 armv6_pmcr_write(val);
540 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
541}
542
543static void
544armv6pmu_stop(void)
545{
546 unsigned long flags, val;
547 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
548
549 raw_spin_lock_irqsave(&events->pmu_lock, flags);
550 val = armv6_pmcr_read();
551 val &= ~ARMV6_PMCR_ENABLE;
552 armv6_pmcr_write(val);
553 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
554}
555
556static int
557armv6pmu_get_event_idx(struct pmu_hw_events *cpuc,
558 struct hw_perf_event *event)
559{
560 /* Always place a cycle counter into the cycle counter. */
561 if (ARMV6_PERFCTR_CPU_CYCLES == event->config_base) {
562 if (test_and_set_bit(ARMV6_CYCLE_COUNTER, cpuc->used_mask))
563 return -EAGAIN;
564
565 return ARMV6_CYCLE_COUNTER;
566 } else {
567 /*
568 * For anything other than a cycle counter, try and use
569 * counter0 and counter1.
570 */
571 if (!test_and_set_bit(ARMV6_COUNTER1, cpuc->used_mask))
572 return ARMV6_COUNTER1;
573
574 if (!test_and_set_bit(ARMV6_COUNTER0, cpuc->used_mask))
575 return ARMV6_COUNTER0;
576
577 /* The counters are all in use. */
578 return -EAGAIN;
579 }
580}
581
582static void
583armv6pmu_disable_event(struct hw_perf_event *hwc,
584 int idx)
585{
586 unsigned long val, mask, evt, flags;
587 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
588
589 if (ARMV6_CYCLE_COUNTER == idx) {
590 mask = ARMV6_PMCR_CCOUNT_IEN;
591 evt = 0;
592 } else if (ARMV6_COUNTER0 == idx) {
593 mask = ARMV6_PMCR_COUNT0_IEN | ARMV6_PMCR_EVT_COUNT0_MASK;
594 evt = ARMV6_PERFCTR_NOP << ARMV6_PMCR_EVT_COUNT0_SHIFT;
595 } else if (ARMV6_COUNTER1 == idx) {
596 mask = ARMV6_PMCR_COUNT1_IEN | ARMV6_PMCR_EVT_COUNT1_MASK;
597 evt = ARMV6_PERFCTR_NOP << ARMV6_PMCR_EVT_COUNT1_SHIFT;
598 } else {
599 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
600 return;
601 }
602
603 /*
604 * Mask out the current event and set the counter to count the number
605 * of ETM bus signal assertion cycles. The external reporting should
606 * be disabled and so this should never increment.
607 */
608 raw_spin_lock_irqsave(&events->pmu_lock, flags);
609 val = armv6_pmcr_read();
610 val &= ~mask;
611 val |= evt;
612 armv6_pmcr_write(val);
613 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
614}
615
616static void
617armv6mpcore_pmu_disable_event(struct hw_perf_event *hwc,
618 int idx)
619{
620 unsigned long val, mask, flags, evt = 0;
621 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
622
623 if (ARMV6_CYCLE_COUNTER == idx) {
624 mask = ARMV6_PMCR_CCOUNT_IEN;
625 } else if (ARMV6_COUNTER0 == idx) {
626 mask = ARMV6_PMCR_COUNT0_IEN;
627 } else if (ARMV6_COUNTER1 == idx) {
628 mask = ARMV6_PMCR_COUNT1_IEN;
629 } else {
630 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
631 return;
632 }
633
634 /*
635 * Unlike UP ARMv6, we don't have a way of stopping the counters. We
636 * simply disable the interrupt reporting.
637 */
638 raw_spin_lock_irqsave(&events->pmu_lock, flags);
639 val = armv6_pmcr_read();
640 val &= ~mask;
641 val |= evt;
642 armv6_pmcr_write(val);
643 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
644}
645
646static int armv6_map_event(struct perf_event *event)
647{
648 return map_cpu_event(event, &armv6_perf_map,
649 &armv6_perf_cache_map, 0xFF);
650}
651
652static struct arm_pmu armv6pmu = {
653 .id = ARM_PERF_PMU_ID_V6,
654 .name = "v6",
655 .handle_irq = armv6pmu_handle_irq,
656 .enable = armv6pmu_enable_event,
657 .disable = armv6pmu_disable_event,
658 .read_counter = armv6pmu_read_counter,
659 .write_counter = armv6pmu_write_counter,
660 .get_event_idx = armv6pmu_get_event_idx,
661 .start = armv6pmu_start,
662 .stop = armv6pmu_stop,
663 .map_event = armv6_map_event,
664 .num_events = 3,
665 .max_period = (1LLU << 32) - 1,
666};
667
668static struct arm_pmu *__init armv6pmu_init(void)
669{
670 return &armv6pmu;
671}
672
673/*
674 * ARMv6mpcore is almost identical to single core ARMv6 with the exception
675 * that some of the events have different enumerations and that there is no
676 * *hack* to stop the programmable counters. To stop the counters we simply
677 * disable the interrupt reporting and update the event. When unthrottling we
678 * reset the period and enable the interrupt reporting.
679 */
680
681static int armv6mpcore_map_event(struct perf_event *event)
682{
683 return map_cpu_event(event, &armv6mpcore_perf_map,
684 &armv6mpcore_perf_cache_map, 0xFF);
685}
686
687static struct arm_pmu armv6mpcore_pmu = {
688 .id = ARM_PERF_PMU_ID_V6MP,
689 .name = "v6mpcore",
690 .handle_irq = armv6pmu_handle_irq,
691 .enable = armv6pmu_enable_event,
692 .disable = armv6mpcore_pmu_disable_event,
693 .read_counter = armv6pmu_read_counter,
694 .write_counter = armv6pmu_write_counter,
695 .get_event_idx = armv6pmu_get_event_idx,
696 .start = armv6pmu_start,
697 .stop = armv6pmu_stop,
698 .map_event = armv6mpcore_map_event,
699 .num_events = 3,
700 .max_period = (1LLU << 32) - 1,
701};
702
703static struct arm_pmu *__init armv6mpcore_pmu_init(void)
704{
705 return &armv6mpcore_pmu;
706}
707#else
708static struct arm_pmu *__init armv6pmu_init(void)
709{
710 return NULL;
711}
712
713static struct arm_pmu *__init armv6mpcore_pmu_init(void)
714{
715 return NULL;
716}
717#endif /* CONFIG_CPU_V6 || CONFIG_CPU_V6K */