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)
34 enum 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
56 enum 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 */
67 static 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
79 static 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
194 enum 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 */
222 static 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
234 static 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
347 static inline unsigned long
armv6_pmcr_read(void)348 armv6_pmcr_read(void)
349 {
350 u32 val;
351 asm volatile("mrc p15, 0, %0, c15, c12, 0" : "=r"(val));
352 return val;
353 }
354
355 static inline void
armv6_pmcr_write(unsigned long val)356 armv6_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
380 static inline int
armv6_pmcr_has_overflowed(unsigned long pmcr)381 armv6_pmcr_has_overflowed(unsigned long pmcr)
382 {
383 return pmcr & ARMV6_PMCR_OVERFLOWED_MASK;
384 }
385
386 static inline int
armv6_pmcr_counter_has_overflowed(unsigned long pmcr,enum armv6_counters counter)387 armv6_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
404 static inline u32
armv6pmu_read_counter(int counter)405 armv6pmu_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
421 static inline void
armv6pmu_write_counter(int counter,u32 value)422 armv6pmu_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
435 static void
armv6pmu_enable_event(struct hw_perf_event * hwc,int idx)436 armv6pmu_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
470 static irqreturn_t
armv6pmu_handle_irq(int irq_num,void * dev)471 armv6pmu_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 perf_sample_data_init(&data, 0);
493
494 cpuc = &__get_cpu_var(cpu_hw_events);
495 for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
496 struct perf_event *event = cpuc->events[idx];
497 struct hw_perf_event *hwc;
498
499 /* Ignore if we don't have an event. */
500 if (!event)
501 continue;
502
503 /*
504 * We have a single interrupt for all counters. Check that
505 * each counter has overflowed before we process it.
506 */
507 if (!armv6_pmcr_counter_has_overflowed(pmcr, idx))
508 continue;
509
510 hwc = &event->hw;
511 armpmu_event_update(event, hwc, idx);
512 data.period = event->hw.last_period;
513 if (!armpmu_event_set_period(event, hwc, idx))
514 continue;
515
516 if (perf_event_overflow(event, &data, regs))
517 cpu_pmu->disable(hwc, idx);
518 }
519
520 /*
521 * Handle the pending perf events.
522 *
523 * Note: this call *must* be run with interrupts disabled. For
524 * platforms that can have the PMU interrupts raised as an NMI, this
525 * will not work.
526 */
527 irq_work_run();
528
529 return IRQ_HANDLED;
530 }
531
532 static void
armv6pmu_start(void)533 armv6pmu_start(void)
534 {
535 unsigned long flags, val;
536 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
537
538 raw_spin_lock_irqsave(&events->pmu_lock, flags);
539 val = armv6_pmcr_read();
540 val |= ARMV6_PMCR_ENABLE;
541 armv6_pmcr_write(val);
542 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
543 }
544
545 static void
armv6pmu_stop(void)546 armv6pmu_stop(void)
547 {
548 unsigned long flags, val;
549 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
550
551 raw_spin_lock_irqsave(&events->pmu_lock, flags);
552 val = armv6_pmcr_read();
553 val &= ~ARMV6_PMCR_ENABLE;
554 armv6_pmcr_write(val);
555 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
556 }
557
558 static int
armv6pmu_get_event_idx(struct pmu_hw_events * cpuc,struct hw_perf_event * event)559 armv6pmu_get_event_idx(struct pmu_hw_events *cpuc,
560 struct hw_perf_event *event)
561 {
562 /* Always place a cycle counter into the cycle counter. */
563 if (ARMV6_PERFCTR_CPU_CYCLES == event->config_base) {
564 if (test_and_set_bit(ARMV6_CYCLE_COUNTER, cpuc->used_mask))
565 return -EAGAIN;
566
567 return ARMV6_CYCLE_COUNTER;
568 } else {
569 /*
570 * For anything other than a cycle counter, try and use
571 * counter0 and counter1.
572 */
573 if (!test_and_set_bit(ARMV6_COUNTER1, cpuc->used_mask))
574 return ARMV6_COUNTER1;
575
576 if (!test_and_set_bit(ARMV6_COUNTER0, cpuc->used_mask))
577 return ARMV6_COUNTER0;
578
579 /* The counters are all in use. */
580 return -EAGAIN;
581 }
582 }
583
584 static void
armv6pmu_disable_event(struct hw_perf_event * hwc,int idx)585 armv6pmu_disable_event(struct hw_perf_event *hwc,
586 int idx)
587 {
588 unsigned long val, mask, evt, flags;
589 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
590
591 if (ARMV6_CYCLE_COUNTER == idx) {
592 mask = ARMV6_PMCR_CCOUNT_IEN;
593 evt = 0;
594 } else if (ARMV6_COUNTER0 == idx) {
595 mask = ARMV6_PMCR_COUNT0_IEN | ARMV6_PMCR_EVT_COUNT0_MASK;
596 evt = ARMV6_PERFCTR_NOP << ARMV6_PMCR_EVT_COUNT0_SHIFT;
597 } else if (ARMV6_COUNTER1 == idx) {
598 mask = ARMV6_PMCR_COUNT1_IEN | ARMV6_PMCR_EVT_COUNT1_MASK;
599 evt = ARMV6_PERFCTR_NOP << ARMV6_PMCR_EVT_COUNT1_SHIFT;
600 } else {
601 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
602 return;
603 }
604
605 /*
606 * Mask out the current event and set the counter to count the number
607 * of ETM bus signal assertion cycles. The external reporting should
608 * be disabled and so this should never increment.
609 */
610 raw_spin_lock_irqsave(&events->pmu_lock, flags);
611 val = armv6_pmcr_read();
612 val &= ~mask;
613 val |= evt;
614 armv6_pmcr_write(val);
615 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
616 }
617
618 static void
armv6mpcore_pmu_disable_event(struct hw_perf_event * hwc,int idx)619 armv6mpcore_pmu_disable_event(struct hw_perf_event *hwc,
620 int idx)
621 {
622 unsigned long val, mask, flags, evt = 0;
623 struct pmu_hw_events *events = cpu_pmu->get_hw_events();
624
625 if (ARMV6_CYCLE_COUNTER == idx) {
626 mask = ARMV6_PMCR_CCOUNT_IEN;
627 } else if (ARMV6_COUNTER0 == idx) {
628 mask = ARMV6_PMCR_COUNT0_IEN;
629 } else if (ARMV6_COUNTER1 == idx) {
630 mask = ARMV6_PMCR_COUNT1_IEN;
631 } else {
632 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
633 return;
634 }
635
636 /*
637 * Unlike UP ARMv6, we don't have a way of stopping the counters. We
638 * simply disable the interrupt reporting.
639 */
640 raw_spin_lock_irqsave(&events->pmu_lock, flags);
641 val = armv6_pmcr_read();
642 val &= ~mask;
643 val |= evt;
644 armv6_pmcr_write(val);
645 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
646 }
647
armv6_map_event(struct perf_event * event)648 static int armv6_map_event(struct perf_event *event)
649 {
650 return map_cpu_event(event, &armv6_perf_map,
651 &armv6_perf_cache_map, 0xFF);
652 }
653
654 static struct arm_pmu armv6pmu = {
655 .id = ARM_PERF_PMU_ID_V6,
656 .name = "v6",
657 .handle_irq = armv6pmu_handle_irq,
658 .enable = armv6pmu_enable_event,
659 .disable = armv6pmu_disable_event,
660 .read_counter = armv6pmu_read_counter,
661 .write_counter = armv6pmu_write_counter,
662 .get_event_idx = armv6pmu_get_event_idx,
663 .start = armv6pmu_start,
664 .stop = armv6pmu_stop,
665 .map_event = armv6_map_event,
666 .num_events = 3,
667 .max_period = (1LLU << 32) - 1,
668 };
669
armv6pmu_init(void)670 static struct arm_pmu *__init armv6pmu_init(void)
671 {
672 return &armv6pmu;
673 }
674
675 /*
676 * ARMv6mpcore is almost identical to single core ARMv6 with the exception
677 * that some of the events have different enumerations and that there is no
678 * *hack* to stop the programmable counters. To stop the counters we simply
679 * disable the interrupt reporting and update the event. When unthrottling we
680 * reset the period and enable the interrupt reporting.
681 */
682
armv6mpcore_map_event(struct perf_event * event)683 static int armv6mpcore_map_event(struct perf_event *event)
684 {
685 return map_cpu_event(event, &armv6mpcore_perf_map,
686 &armv6mpcore_perf_cache_map, 0xFF);
687 }
688
689 static struct arm_pmu armv6mpcore_pmu = {
690 .id = ARM_PERF_PMU_ID_V6MP,
691 .name = "v6mpcore",
692 .handle_irq = armv6pmu_handle_irq,
693 .enable = armv6pmu_enable_event,
694 .disable = armv6mpcore_pmu_disable_event,
695 .read_counter = armv6pmu_read_counter,
696 .write_counter = armv6pmu_write_counter,
697 .get_event_idx = armv6pmu_get_event_idx,
698 .start = armv6pmu_start,
699 .stop = armv6pmu_stop,
700 .map_event = armv6mpcore_map_event,
701 .num_events = 3,
702 .max_period = (1LLU << 32) - 1,
703 };
704
armv6mpcore_pmu_init(void)705 static struct arm_pmu *__init armv6mpcore_pmu_init(void)
706 {
707 return &armv6mpcore_pmu;
708 }
709 #else
armv6pmu_init(void)710 static struct arm_pmu *__init armv6pmu_init(void)
711 {
712 return NULL;
713 }
714
armv6mpcore_pmu_init(void)715 static struct arm_pmu *__init armv6mpcore_pmu_init(void)
716 {
717 return NULL;
718 }
719 #endif /* CONFIG_CPU_V6 || CONFIG_CPU_V6K */
720