Commit | Line | Data |
---|---|---|
e5dcb58a DD |
1 | /* |
2 | * Linux performance counter support for MIPS. | |
3 | * | |
4 | * Copyright (C) 2010 MIPS Technologies, Inc. | |
82091564 | 5 | * Copyright (C) 2011 Cavium Networks, Inc. |
e5dcb58a DD |
6 | * Author: Deng-Cheng Zhu |
7 | * | |
8 | * This code is based on the implementation for ARM, which is in turn | |
9 | * based on the sparc64 perf event code and the x86 code. Performance | |
10 | * counter access is based on the MIPS Oprofile code. And the callchain | |
11 | * support references the code of MIPS stacktrace.c. | |
12 | * | |
13 | * This program is free software; you can redistribute it and/or modify | |
14 | * it under the terms of the GNU General Public License version 2 as | |
15 | * published by the Free Software Foundation. | |
16 | */ | |
17 | ||
18 | #include <linux/cpumask.h> | |
19 | #include <linux/interrupt.h> | |
20 | #include <linux/smp.h> | |
21 | #include <linux/kernel.h> | |
22 | #include <linux/perf_event.h> | |
23 | #include <linux/uaccess.h> | |
24 | ||
25 | #include <asm/irq.h> | |
26 | #include <asm/irq_regs.h> | |
27 | #include <asm/stacktrace.h> | |
28 | #include <asm/time.h> /* For perf_irq */ | |
29 | ||
e5dcb58a | 30 | #define MIPS_MAX_HWEVENTS 4 |
a7911a8f AC |
31 | #define MIPS_TCS_PER_COUNTER 2 |
32 | #define MIPS_CPUID_TO_COUNTER_MASK (MIPS_TCS_PER_COUNTER - 1) | |
e5dcb58a DD |
33 | |
34 | struct cpu_hw_events { | |
35 | /* Array of events on this cpu. */ | |
36 | struct perf_event *events[MIPS_MAX_HWEVENTS]; | |
37 | ||
38 | /* | |
39 | * Set the bit (indexed by the counter number) when the counter | |
40 | * is used for an event. | |
41 | */ | |
42 | unsigned long used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)]; | |
43 | ||
e5dcb58a DD |
44 | /* |
45 | * Software copy of the control register for each performance counter. | |
46 | * MIPS CPUs vary in performance counters. They use this differently, | |
47 | * and even may not use it. | |
48 | */ | |
49 | unsigned int saved_ctrl[MIPS_MAX_HWEVENTS]; | |
50 | }; | |
51 | DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { | |
52 | .saved_ctrl = {0}, | |
53 | }; | |
54 | ||
55 | /* The description of MIPS performance events. */ | |
56 | struct mips_perf_event { | |
57 | unsigned int event_id; | |
58 | /* | |
59 | * MIPS performance counters are indexed starting from 0. | |
60 | * CNTR_EVEN indicates the indexes of the counters to be used are | |
61 | * even numbers. | |
62 | */ | |
63 | unsigned int cntr_mask; | |
64 | #define CNTR_EVEN 0x55555555 | |
65 | #define CNTR_ODD 0xaaaaaaaa | |
82091564 | 66 | #define CNTR_ALL 0xffffffff |
e5dcb58a DD |
67 | #ifdef CONFIG_MIPS_MT_SMP |
68 | enum { | |
69 | T = 0, | |
70 | V = 1, | |
71 | P = 2, | |
72 | } range; | |
73 | #else | |
74 | #define T | |
75 | #define V | |
76 | #define P | |
77 | #endif | |
78 | }; | |
79 | ||
80 | static struct mips_perf_event raw_event; | |
81 | static DEFINE_MUTEX(raw_event_mutex); | |
82 | ||
e5dcb58a DD |
83 | #define C(x) PERF_COUNT_HW_CACHE_##x |
84 | ||
85 | struct mips_pmu { | |
82091564 DD |
86 | u64 max_period; |
87 | u64 valid_count; | |
88 | u64 overflow; | |
e5dcb58a DD |
89 | const char *name; |
90 | int irq; | |
e5dcb58a DD |
91 | u64 (*read_counter)(unsigned int idx); |
92 | void (*write_counter)(unsigned int idx, u64 val); | |
e5dcb58a DD |
93 | const struct mips_perf_event *(*map_raw_event)(u64 config); |
94 | const struct mips_perf_event (*general_event_map)[PERF_COUNT_HW_MAX]; | |
95 | const struct mips_perf_event (*cache_event_map) | |
96 | [PERF_COUNT_HW_CACHE_MAX] | |
97 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
98 | [PERF_COUNT_HW_CACHE_RESULT_MAX]; | |
99 | unsigned int num_counters; | |
100 | }; | |
101 | ||
82091564 DD |
102 | static struct mips_pmu mipspmu; |
103 | ||
70342287 RB |
104 | #define M_PERFCTL_EXL (1 << 0) |
105 | #define M_PERFCTL_KERNEL (1 << 1) | |
106 | #define M_PERFCTL_SUPERVISOR (1 << 2) | |
107 | #define M_PERFCTL_USER (1 << 3) | |
108 | #define M_PERFCTL_INTERRUPT_ENABLE (1 << 4) | |
82091564 | 109 | #define M_PERFCTL_EVENT(event) (((event) & 0x3ff) << 5) |
70342287 | 110 | #define M_PERFCTL_VPEID(vpe) ((vpe) << 16) |
a7911a8f AC |
111 | |
112 | #ifdef CONFIG_CPU_BMIPS5000 | |
113 | #define M_PERFCTL_MT_EN(filter) 0 | |
114 | #else /* !CONFIG_CPU_BMIPS5000 */ | |
82091564 | 115 | #define M_PERFCTL_MT_EN(filter) ((filter) << 20) |
a7911a8f AC |
116 | #endif /* CONFIG_CPU_BMIPS5000 */ |
117 | ||
70342287 RB |
118 | #define M_TC_EN_ALL M_PERFCTL_MT_EN(0) |
119 | #define M_TC_EN_VPE M_PERFCTL_MT_EN(1) | |
120 | #define M_TC_EN_TC M_PERFCTL_MT_EN(2) | |
121 | #define M_PERFCTL_TCID(tcid) ((tcid) << 22) | |
122 | #define M_PERFCTL_WIDE (1 << 30) | |
123 | #define M_PERFCTL_MORE (1 << 31) | |
124 | #define M_PERFCTL_TC (1 << 30) | |
82091564 DD |
125 | |
126 | #define M_PERFCTL_COUNT_EVENT_WHENEVER (M_PERFCTL_EXL | \ | |
127 | M_PERFCTL_KERNEL | \ | |
128 | M_PERFCTL_USER | \ | |
129 | M_PERFCTL_SUPERVISOR | \ | |
130 | M_PERFCTL_INTERRUPT_ENABLE) | |
131 | ||
132 | #ifdef CONFIG_MIPS_MT_SMP | |
133 | #define M_PERFCTL_CONFIG_MASK 0x3fff801f | |
134 | #else | |
135 | #define M_PERFCTL_CONFIG_MASK 0x1f | |
136 | #endif | |
137 | #define M_PERFCTL_EVENT_MASK 0xfe0 | |
138 | ||
139 | ||
399aaa25 | 140 | #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS |
82091564 DD |
141 | static int cpu_has_mipsmt_pertccounters; |
142 | ||
143 | static DEFINE_RWLOCK(pmuint_rwlock); | |
144 | ||
a7911a8f AC |
145 | #if defined(CONFIG_CPU_BMIPS5000) |
146 | #define vpe_id() (cpu_has_mipsmt_pertccounters ? \ | |
147 | 0 : (smp_processor_id() & MIPS_CPUID_TO_COUNTER_MASK)) | |
148 | #else | |
82091564 DD |
149 | /* |
150 | * FIXME: For VSMP, vpe_id() is redefined for Perf-events, because | |
151 | * cpu_data[cpuid].vpe_id reports 0 for _both_ CPUs. | |
152 | */ | |
82091564 | 153 | #define vpe_id() (cpu_has_mipsmt_pertccounters ? \ |
a7911a8f AC |
154 | 0 : smp_processor_id()) |
155 | #endif | |
82091564 DD |
156 | |
157 | /* Copied from op_model_mipsxx.c */ | |
158 | static unsigned int vpe_shift(void) | |
159 | { | |
160 | if (num_possible_cpus() > 1) | |
161 | return 1; | |
162 | ||
163 | return 0; | |
164 | } | |
165 | ||
166 | static unsigned int counters_total_to_per_cpu(unsigned int counters) | |
167 | { | |
168 | return counters >> vpe_shift(); | |
169 | } | |
170 | ||
399aaa25 | 171 | #else /* !CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */ |
82091564 DD |
172 | #define vpe_id() 0 |
173 | ||
399aaa25 | 174 | #endif /* CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */ |
82091564 DD |
175 | |
176 | static void resume_local_counters(void); | |
177 | static void pause_local_counters(void); | |
178 | static irqreturn_t mipsxx_pmu_handle_irq(int, void *); | |
179 | static int mipsxx_pmu_handle_shared_irq(void); | |
180 | ||
181 | static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx) | |
182 | { | |
183 | if (vpe_id() == 1) | |
184 | idx = (idx + 2) & 3; | |
185 | return idx; | |
186 | } | |
187 | ||
188 | static u64 mipsxx_pmu_read_counter(unsigned int idx) | |
189 | { | |
190 | idx = mipsxx_pmu_swizzle_perf_idx(idx); | |
191 | ||
192 | switch (idx) { | |
193 | case 0: | |
194 | /* | |
195 | * The counters are unsigned, we must cast to truncate | |
196 | * off the high bits. | |
197 | */ | |
198 | return (u32)read_c0_perfcntr0(); | |
199 | case 1: | |
200 | return (u32)read_c0_perfcntr1(); | |
201 | case 2: | |
202 | return (u32)read_c0_perfcntr2(); | |
203 | case 3: | |
204 | return (u32)read_c0_perfcntr3(); | |
205 | default: | |
206 | WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx); | |
207 | return 0; | |
208 | } | |
209 | } | |
210 | ||
211 | static u64 mipsxx_pmu_read_counter_64(unsigned int idx) | |
212 | { | |
213 | idx = mipsxx_pmu_swizzle_perf_idx(idx); | |
214 | ||
215 | switch (idx) { | |
216 | case 0: | |
217 | return read_c0_perfcntr0_64(); | |
218 | case 1: | |
219 | return read_c0_perfcntr1_64(); | |
220 | case 2: | |
221 | return read_c0_perfcntr2_64(); | |
222 | case 3: | |
223 | return read_c0_perfcntr3_64(); | |
224 | default: | |
225 | WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx); | |
226 | return 0; | |
227 | } | |
228 | } | |
229 | ||
230 | static void mipsxx_pmu_write_counter(unsigned int idx, u64 val) | |
231 | { | |
232 | idx = mipsxx_pmu_swizzle_perf_idx(idx); | |
233 | ||
234 | switch (idx) { | |
235 | case 0: | |
236 | write_c0_perfcntr0(val); | |
237 | return; | |
238 | case 1: | |
239 | write_c0_perfcntr1(val); | |
240 | return; | |
241 | case 2: | |
242 | write_c0_perfcntr2(val); | |
243 | return; | |
244 | case 3: | |
245 | write_c0_perfcntr3(val); | |
246 | return; | |
247 | } | |
248 | } | |
249 | ||
250 | static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val) | |
251 | { | |
252 | idx = mipsxx_pmu_swizzle_perf_idx(idx); | |
253 | ||
254 | switch (idx) { | |
255 | case 0: | |
256 | write_c0_perfcntr0_64(val); | |
257 | return; | |
258 | case 1: | |
259 | write_c0_perfcntr1_64(val); | |
260 | return; | |
261 | case 2: | |
262 | write_c0_perfcntr2_64(val); | |
263 | return; | |
264 | case 3: | |
265 | write_c0_perfcntr3_64(val); | |
266 | return; | |
267 | } | |
268 | } | |
269 | ||
270 | static unsigned int mipsxx_pmu_read_control(unsigned int idx) | |
271 | { | |
272 | idx = mipsxx_pmu_swizzle_perf_idx(idx); | |
273 | ||
274 | switch (idx) { | |
275 | case 0: | |
276 | return read_c0_perfctrl0(); | |
277 | case 1: | |
278 | return read_c0_perfctrl1(); | |
279 | case 2: | |
280 | return read_c0_perfctrl2(); | |
281 | case 3: | |
282 | return read_c0_perfctrl3(); | |
283 | default: | |
284 | WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx); | |
285 | return 0; | |
286 | } | |
287 | } | |
288 | ||
289 | static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val) | |
290 | { | |
291 | idx = mipsxx_pmu_swizzle_perf_idx(idx); | |
292 | ||
293 | switch (idx) { | |
294 | case 0: | |
295 | write_c0_perfctrl0(val); | |
296 | return; | |
297 | case 1: | |
298 | write_c0_perfctrl1(val); | |
299 | return; | |
300 | case 2: | |
301 | write_c0_perfctrl2(val); | |
302 | return; | |
303 | case 3: | |
304 | write_c0_perfctrl3(val); | |
305 | return; | |
306 | } | |
307 | } | |
308 | ||
309 | static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc, | |
310 | struct hw_perf_event *hwc) | |
311 | { | |
312 | int i; | |
313 | ||
314 | /* | |
315 | * We only need to care the counter mask. The range has been | |
316 | * checked definitely. | |
317 | */ | |
318 | unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff; | |
319 | ||
320 | for (i = mipspmu.num_counters - 1; i >= 0; i--) { | |
321 | /* | |
322 | * Note that some MIPS perf events can be counted by both | |
323 | * even and odd counters, wheresas many other are only by | |
324 | * even _or_ odd counters. This introduces an issue that | |
325 | * when the former kind of event takes the counter the | |
326 | * latter kind of event wants to use, then the "counter | |
327 | * allocation" for the latter event will fail. In fact if | |
328 | * they can be dynamically swapped, they both feel happy. | |
329 | * But here we leave this issue alone for now. | |
330 | */ | |
331 | if (test_bit(i, &cntr_mask) && | |
332 | !test_and_set_bit(i, cpuc->used_mask)) | |
333 | return i; | |
334 | } | |
335 | ||
336 | return -EAGAIN; | |
337 | } | |
338 | ||
339 | static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx) | |
340 | { | |
35898716 | 341 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
82091564 DD |
342 | |
343 | WARN_ON(idx < 0 || idx >= mipspmu.num_counters); | |
344 | ||
345 | cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) | | |
346 | (evt->config_base & M_PERFCTL_CONFIG_MASK) | | |
347 | /* Make sure interrupt enabled. */ | |
348 | M_PERFCTL_INTERRUPT_ENABLE; | |
a7911a8f AC |
349 | if (IS_ENABLED(CONFIG_CPU_BMIPS5000)) |
350 | /* enable the counter for the calling thread */ | |
351 | cpuc->saved_ctrl[idx] |= | |
352 | (1 << (12 + vpe_id())) | M_PERFCTL_TC; | |
353 | ||
82091564 DD |
354 | /* |
355 | * We do not actually let the counter run. Leave it until start(). | |
356 | */ | |
357 | } | |
358 | ||
359 | static void mipsxx_pmu_disable_event(int idx) | |
360 | { | |
35898716 | 361 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
82091564 DD |
362 | unsigned long flags; |
363 | ||
364 | WARN_ON(idx < 0 || idx >= mipspmu.num_counters); | |
365 | ||
366 | local_irq_save(flags); | |
367 | cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) & | |
368 | ~M_PERFCTL_COUNT_EVENT_WHENEVER; | |
369 | mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]); | |
370 | local_irq_restore(flags); | |
371 | } | |
e5dcb58a DD |
372 | |
373 | static int mipspmu_event_set_period(struct perf_event *event, | |
374 | struct hw_perf_event *hwc, | |
375 | int idx) | |
376 | { | |
82091564 DD |
377 | u64 left = local64_read(&hwc->period_left); |
378 | u64 period = hwc->sample_period; | |
e5dcb58a | 379 | int ret = 0; |
e5dcb58a | 380 | |
82091564 DD |
381 | if (unlikely((left + period) & (1ULL << 63))) { |
382 | /* left underflowed by more than period. */ | |
e5dcb58a DD |
383 | left = period; |
384 | local64_set(&hwc->period_left, left); | |
385 | hwc->last_period = period; | |
386 | ret = 1; | |
82091564 DD |
387 | } else if (unlikely((left + period) <= period)) { |
388 | /* left underflowed by less than period. */ | |
e5dcb58a DD |
389 | left += period; |
390 | local64_set(&hwc->period_left, left); | |
391 | hwc->last_period = period; | |
392 | ret = 1; | |
393 | } | |
394 | ||
82091564 DD |
395 | if (left > mipspmu.max_period) { |
396 | left = mipspmu.max_period; | |
397 | local64_set(&hwc->period_left, left); | |
398 | } | |
e5dcb58a | 399 | |
82091564 | 400 | local64_set(&hwc->prev_count, mipspmu.overflow - left); |
e5dcb58a | 401 | |
82091564 | 402 | mipspmu.write_counter(idx, mipspmu.overflow - left); |
e5dcb58a DD |
403 | |
404 | perf_event_update_userpage(event); | |
405 | ||
406 | return ret; | |
407 | } | |
408 | ||
409 | static void mipspmu_event_update(struct perf_event *event, | |
410 | struct hw_perf_event *hwc, | |
411 | int idx) | |
412 | { | |
82091564 | 413 | u64 prev_raw_count, new_raw_count; |
e5dcb58a DD |
414 | u64 delta; |
415 | ||
416 | again: | |
417 | prev_raw_count = local64_read(&hwc->prev_count); | |
82091564 | 418 | new_raw_count = mipspmu.read_counter(idx); |
e5dcb58a DD |
419 | |
420 | if (local64_cmpxchg(&hwc->prev_count, prev_raw_count, | |
421 | new_raw_count) != prev_raw_count) | |
422 | goto again; | |
423 | ||
82091564 | 424 | delta = new_raw_count - prev_raw_count; |
e5dcb58a DD |
425 | |
426 | local64_add(delta, &event->count); | |
427 | local64_sub(delta, &hwc->period_left); | |
428 | } | |
429 | ||
430 | static void mipspmu_start(struct perf_event *event, int flags) | |
431 | { | |
432 | struct hw_perf_event *hwc = &event->hw; | |
433 | ||
e5dcb58a DD |
434 | if (flags & PERF_EF_RELOAD) |
435 | WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE)); | |
436 | ||
437 | hwc->state = 0; | |
438 | ||
439 | /* Set the period for the event. */ | |
440 | mipspmu_event_set_period(event, hwc, hwc->idx); | |
441 | ||
442 | /* Enable the event. */ | |
82091564 | 443 | mipsxx_pmu_enable_event(hwc, hwc->idx); |
e5dcb58a DD |
444 | } |
445 | ||
446 | static void mipspmu_stop(struct perf_event *event, int flags) | |
447 | { | |
448 | struct hw_perf_event *hwc = &event->hw; | |
449 | ||
e5dcb58a DD |
450 | if (!(hwc->state & PERF_HES_STOPPED)) { |
451 | /* We are working on a local event. */ | |
82091564 | 452 | mipsxx_pmu_disable_event(hwc->idx); |
e5dcb58a DD |
453 | barrier(); |
454 | mipspmu_event_update(event, hwc, hwc->idx); | |
455 | hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE; | |
456 | } | |
457 | } | |
458 | ||
459 | static int mipspmu_add(struct perf_event *event, int flags) | |
460 | { | |
35898716 | 461 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
e5dcb58a DD |
462 | struct hw_perf_event *hwc = &event->hw; |
463 | int idx; | |
464 | int err = 0; | |
465 | ||
466 | perf_pmu_disable(event->pmu); | |
467 | ||
468 | /* To look for a free counter for this event. */ | |
82091564 | 469 | idx = mipsxx_pmu_alloc_counter(cpuc, hwc); |
e5dcb58a DD |
470 | if (idx < 0) { |
471 | err = idx; | |
472 | goto out; | |
473 | } | |
474 | ||
475 | /* | |
476 | * If there is an event in the counter we are going to use then | |
477 | * make sure it is disabled. | |
478 | */ | |
479 | event->hw.idx = idx; | |
82091564 | 480 | mipsxx_pmu_disable_event(idx); |
e5dcb58a DD |
481 | cpuc->events[idx] = event; |
482 | ||
483 | hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE; | |
484 | if (flags & PERF_EF_START) | |
485 | mipspmu_start(event, PERF_EF_RELOAD); | |
486 | ||
487 | /* Propagate our changes to the userspace mapping. */ | |
488 | perf_event_update_userpage(event); | |
489 | ||
490 | out: | |
491 | perf_pmu_enable(event->pmu); | |
492 | return err; | |
493 | } | |
494 | ||
495 | static void mipspmu_del(struct perf_event *event, int flags) | |
496 | { | |
35898716 | 497 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
e5dcb58a DD |
498 | struct hw_perf_event *hwc = &event->hw; |
499 | int idx = hwc->idx; | |
500 | ||
82091564 | 501 | WARN_ON(idx < 0 || idx >= mipspmu.num_counters); |
e5dcb58a DD |
502 | |
503 | mipspmu_stop(event, PERF_EF_UPDATE); | |
504 | cpuc->events[idx] = NULL; | |
505 | clear_bit(idx, cpuc->used_mask); | |
506 | ||
507 | perf_event_update_userpage(event); | |
508 | } | |
509 | ||
510 | static void mipspmu_read(struct perf_event *event) | |
511 | { | |
512 | struct hw_perf_event *hwc = &event->hw; | |
513 | ||
514 | /* Don't read disabled counters! */ | |
515 | if (hwc->idx < 0) | |
516 | return; | |
517 | ||
518 | mipspmu_event_update(event, hwc, hwc->idx); | |
519 | } | |
520 | ||
521 | static void mipspmu_enable(struct pmu *pmu) | |
522 | { | |
399aaa25 | 523 | #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS |
82091564 DD |
524 | write_unlock(&pmuint_rwlock); |
525 | #endif | |
526 | resume_local_counters(); | |
e5dcb58a DD |
527 | } |
528 | ||
82091564 DD |
529 | /* |
530 | * MIPS performance counters can be per-TC. The control registers can | |
92a76f6d | 531 | * not be directly accessed across CPUs. Hence if we want to do global |
82091564 DD |
532 | * control, we need cross CPU calls. on_each_cpu() can help us, but we |
533 | * can not make sure this function is called with interrupts enabled. So | |
534 | * here we pause local counters and then grab a rwlock and leave the | |
535 | * counters on other CPUs alone. If any counter interrupt raises while | |
536 | * we own the write lock, simply pause local counters on that CPU and | |
537 | * spin in the handler. Also we know we won't be switched to another | |
538 | * CPU after pausing local counters and before grabbing the lock. | |
539 | */ | |
e5dcb58a DD |
540 | static void mipspmu_disable(struct pmu *pmu) |
541 | { | |
82091564 | 542 | pause_local_counters(); |
399aaa25 | 543 | #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS |
82091564 DD |
544 | write_lock(&pmuint_rwlock); |
545 | #endif | |
e5dcb58a DD |
546 | } |
547 | ||
548 | static atomic_t active_events = ATOMIC_INIT(0); | |
549 | static DEFINE_MUTEX(pmu_reserve_mutex); | |
550 | static int (*save_perf_irq)(void); | |
551 | ||
552 | static int mipspmu_get_irq(void) | |
553 | { | |
554 | int err; | |
555 | ||
82091564 | 556 | if (mipspmu.irq >= 0) { |
e5dcb58a | 557 | /* Request my own irq handler. */ |
82091564 | 558 | err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq, |
a1ec0e18 JH |
559 | IRQF_PERCPU | IRQF_NOBALANCING | |
560 | IRQF_NO_THREAD | IRQF_NO_SUSPEND | | |
561 | IRQF_SHARED, | |
562 | "mips_perf_pmu", &mipspmu); | |
e5dcb58a | 563 | if (err) { |
7178d2cd JP |
564 | pr_warn("Unable to request IRQ%d for MIPS performance counters!\n", |
565 | mipspmu.irq); | |
e5dcb58a DD |
566 | } |
567 | } else if (cp0_perfcount_irq < 0) { | |
568 | /* | |
569 | * We are sharing the irq number with the timer interrupt. | |
570 | */ | |
571 | save_perf_irq = perf_irq; | |
82091564 | 572 | perf_irq = mipsxx_pmu_handle_shared_irq; |
e5dcb58a DD |
573 | err = 0; |
574 | } else { | |
7178d2cd | 575 | pr_warn("The platform hasn't properly defined its interrupt controller\n"); |
e5dcb58a DD |
576 | err = -ENOENT; |
577 | } | |
578 | ||
579 | return err; | |
580 | } | |
581 | ||
582 | static void mipspmu_free_irq(void) | |
583 | { | |
82091564 | 584 | if (mipspmu.irq >= 0) |
a1ec0e18 | 585 | free_irq(mipspmu.irq, &mipspmu); |
e5dcb58a DD |
586 | else if (cp0_perfcount_irq < 0) |
587 | perf_irq = save_perf_irq; | |
588 | } | |
589 | ||
590 | /* | |
591 | * mipsxx/rm9000/loongson2 have different performance counters, they have | |
592 | * specific low-level init routines. | |
593 | */ | |
594 | static void reset_counters(void *arg); | |
595 | static int __hw_perf_event_init(struct perf_event *event); | |
596 | ||
597 | static void hw_perf_event_destroy(struct perf_event *event) | |
598 | { | |
599 | if (atomic_dec_and_mutex_lock(&active_events, | |
600 | &pmu_reserve_mutex)) { | |
601 | /* | |
602 | * We must not call the destroy function with interrupts | |
603 | * disabled. | |
604 | */ | |
605 | on_each_cpu(reset_counters, | |
82091564 | 606 | (void *)(long)mipspmu.num_counters, 1); |
e5dcb58a DD |
607 | mipspmu_free_irq(); |
608 | mutex_unlock(&pmu_reserve_mutex); | |
609 | } | |
610 | } | |
611 | ||
612 | static int mipspmu_event_init(struct perf_event *event) | |
613 | { | |
614 | int err = 0; | |
615 | ||
2481c5fa SE |
616 | /* does not support taken branch sampling */ |
617 | if (has_branch_stack(event)) | |
618 | return -EOPNOTSUPP; | |
619 | ||
e5dcb58a DD |
620 | switch (event->attr.type) { |
621 | case PERF_TYPE_RAW: | |
622 | case PERF_TYPE_HARDWARE: | |
623 | case PERF_TYPE_HW_CACHE: | |
624 | break; | |
625 | ||
626 | default: | |
627 | return -ENOENT; | |
628 | } | |
629 | ||
82091564 DD |
630 | if (event->cpu >= nr_cpumask_bits || |
631 | (event->cpu >= 0 && !cpu_online(event->cpu))) | |
e5dcb58a DD |
632 | return -ENODEV; |
633 | ||
634 | if (!atomic_inc_not_zero(&active_events)) { | |
e5dcb58a DD |
635 | mutex_lock(&pmu_reserve_mutex); |
636 | if (atomic_read(&active_events) == 0) | |
637 | err = mipspmu_get_irq(); | |
638 | ||
639 | if (!err) | |
640 | atomic_inc(&active_events); | |
641 | mutex_unlock(&pmu_reserve_mutex); | |
642 | } | |
643 | ||
644 | if (err) | |
645 | return err; | |
646 | ||
ff5d7265 | 647 | return __hw_perf_event_init(event); |
e5dcb58a DD |
648 | } |
649 | ||
650 | static struct pmu pmu = { | |
651 | .pmu_enable = mipspmu_enable, | |
652 | .pmu_disable = mipspmu_disable, | |
653 | .event_init = mipspmu_event_init, | |
654 | .add = mipspmu_add, | |
655 | .del = mipspmu_del, | |
656 | .start = mipspmu_start, | |
657 | .stop = mipspmu_stop, | |
658 | .read = mipspmu_read, | |
659 | }; | |
660 | ||
661 | static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev) | |
662 | { | |
663 | /* | |
664 | * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for | |
665 | * event_id. | |
666 | */ | |
667 | #ifdef CONFIG_MIPS_MT_SMP | |
668 | return ((unsigned int)pev->range << 24) | | |
669 | (pev->cntr_mask & 0xffff00) | | |
670 | (pev->event_id & 0xff); | |
671 | #else | |
672 | return (pev->cntr_mask & 0xffff00) | | |
673 | (pev->event_id & 0xff); | |
674 | #endif | |
675 | } | |
676 | ||
677 | static const struct mips_perf_event *mipspmu_map_general_event(int idx) | |
678 | { | |
e5dcb58a | 679 | |
c5600b2d AC |
680 | if ((*mipspmu.general_event_map)[idx].cntr_mask == 0) |
681 | return ERR_PTR(-EOPNOTSUPP); | |
682 | return &(*mipspmu.general_event_map)[idx]; | |
e5dcb58a DD |
683 | } |
684 | ||
685 | static const struct mips_perf_event *mipspmu_map_cache_event(u64 config) | |
686 | { | |
687 | unsigned int cache_type, cache_op, cache_result; | |
688 | const struct mips_perf_event *pev; | |
689 | ||
690 | cache_type = (config >> 0) & 0xff; | |
691 | if (cache_type >= PERF_COUNT_HW_CACHE_MAX) | |
692 | return ERR_PTR(-EINVAL); | |
693 | ||
694 | cache_op = (config >> 8) & 0xff; | |
695 | if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX) | |
696 | return ERR_PTR(-EINVAL); | |
697 | ||
698 | cache_result = (config >> 16) & 0xff; | |
699 | if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX) | |
700 | return ERR_PTR(-EINVAL); | |
701 | ||
82091564 | 702 | pev = &((*mipspmu.cache_event_map) |
e5dcb58a DD |
703 | [cache_type] |
704 | [cache_op] | |
705 | [cache_result]); | |
706 | ||
c5600b2d | 707 | if (pev->cntr_mask == 0) |
e5dcb58a DD |
708 | return ERR_PTR(-EOPNOTSUPP); |
709 | ||
710 | return pev; | |
711 | ||
712 | } | |
713 | ||
e5dcb58a DD |
714 | static int validate_group(struct perf_event *event) |
715 | { | |
716 | struct perf_event *sibling, *leader = event->group_leader; | |
717 | struct cpu_hw_events fake_cpuc; | |
718 | ||
719 | memset(&fake_cpuc, 0, sizeof(fake_cpuc)); | |
720 | ||
266623b7 | 721 | if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0) |
aa2bc1ad | 722 | return -EINVAL; |
e5dcb58a DD |
723 | |
724 | list_for_each_entry(sibling, &leader->sibling_list, group_entry) { | |
266623b7 | 725 | if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0) |
aa2bc1ad | 726 | return -EINVAL; |
e5dcb58a DD |
727 | } |
728 | ||
266623b7 | 729 | if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0) |
aa2bc1ad | 730 | return -EINVAL; |
e5dcb58a DD |
731 | |
732 | return 0; | |
733 | } | |
734 | ||
735 | /* This is needed by specific irq handlers in perf_event_*.c */ | |
736 | static void handle_associated_event(struct cpu_hw_events *cpuc, | |
737 | int idx, struct perf_sample_data *data, | |
738 | struct pt_regs *regs) | |
739 | { | |
740 | struct perf_event *event = cpuc->events[idx]; | |
741 | struct hw_perf_event *hwc = &event->hw; | |
742 | ||
743 | mipspmu_event_update(event, hwc, idx); | |
744 | data->period = event->hw.last_period; | |
745 | if (!mipspmu_event_set_period(event, hwc, idx)) | |
746 | return; | |
747 | ||
748 | if (perf_event_overflow(event, data, regs)) | |
82091564 | 749 | mipsxx_pmu_disable_event(idx); |
e5dcb58a | 750 | } |
3a9ab99e | 751 | |
3a9ab99e | 752 | |
4409af37 | 753 | static int __n_counters(void) |
3a9ab99e | 754 | { |
30228c40 | 755 | if (!cpu_has_perf) |
3a9ab99e DCZ |
756 | return 0; |
757 | if (!(read_c0_perfctrl0() & M_PERFCTL_MORE)) | |
758 | return 1; | |
759 | if (!(read_c0_perfctrl1() & M_PERFCTL_MORE)) | |
760 | return 2; | |
761 | if (!(read_c0_perfctrl2() & M_PERFCTL_MORE)) | |
762 | return 3; | |
763 | ||
764 | return 4; | |
765 | } | |
766 | ||
4409af37 | 767 | static int n_counters(void) |
3a9ab99e DCZ |
768 | { |
769 | int counters; | |
770 | ||
771 | switch (current_cpu_type()) { | |
772 | case CPU_R10000: | |
773 | counters = 2; | |
774 | break; | |
775 | ||
776 | case CPU_R12000: | |
777 | case CPU_R14000: | |
30577391 | 778 | case CPU_R16000: |
3a9ab99e DCZ |
779 | counters = 4; |
780 | break; | |
781 | ||
782 | default: | |
783 | counters = __n_counters(); | |
784 | } | |
785 | ||
786 | return counters; | |
787 | } | |
788 | ||
789 | static void reset_counters(void *arg) | |
790 | { | |
791 | int counters = (int)(long)arg; | |
792 | switch (counters) { | |
793 | case 4: | |
82091564 DD |
794 | mipsxx_pmu_write_control(3, 0); |
795 | mipspmu.write_counter(3, 0); | |
3a9ab99e | 796 | case 3: |
82091564 DD |
797 | mipsxx_pmu_write_control(2, 0); |
798 | mipspmu.write_counter(2, 0); | |
3a9ab99e | 799 | case 2: |
82091564 DD |
800 | mipsxx_pmu_write_control(1, 0); |
801 | mipspmu.write_counter(1, 0); | |
3a9ab99e | 802 | case 1: |
82091564 DD |
803 | mipsxx_pmu_write_control(0, 0); |
804 | mipspmu.write_counter(0, 0); | |
3a9ab99e DCZ |
805 | } |
806 | } | |
807 | ||
9597e432 | 808 | /* 24K/34K/1004K/interAptiv/loongson1 cores share the same event map. */ |
3a9ab99e DCZ |
809 | static const struct mips_perf_event mipsxxcore_event_map |
810 | [PERF_COUNT_HW_MAX] = { | |
811 | [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P }, | |
812 | [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T }, | |
3a9ab99e DCZ |
813 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T }, |
814 | [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T }, | |
3a9ab99e DCZ |
815 | }; |
816 | ||
c52068bd | 817 | /* 74K/proAptiv core has different branch event code. */ |
6b0b8429 | 818 | static const struct mips_perf_event mipsxxcore_event_map2 |
3a9ab99e DCZ |
819 | [PERF_COUNT_HW_MAX] = { |
820 | [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P }, | |
821 | [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T }, | |
3a9ab99e DCZ |
822 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T }, |
823 | [PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T }, | |
3a9ab99e DCZ |
824 | }; |
825 | ||
fd716fca JH |
826 | static const struct mips_perf_event i6400_event_map[PERF_COUNT_HW_MAX] = { |
827 | [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD }, | |
828 | [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD }, | |
829 | /* These only count dcache, not icache */ | |
830 | [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x45, CNTR_EVEN | CNTR_ODD }, | |
831 | [PERF_COUNT_HW_CACHE_MISSES] = { 0x48, CNTR_EVEN | CNTR_ODD }, | |
832 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x15, CNTR_EVEN | CNTR_ODD }, | |
833 | [PERF_COUNT_HW_BRANCH_MISSES] = { 0x16, CNTR_EVEN | CNTR_ODD }, | |
834 | }; | |
835 | ||
f14ceff7 HC |
836 | static const struct mips_perf_event loongson3_event_map[PERF_COUNT_HW_MAX] = { |
837 | [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN }, | |
838 | [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, CNTR_ODD }, | |
839 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x01, CNTR_EVEN }, | |
840 | [PERF_COUNT_HW_BRANCH_MISSES] = { 0x01, CNTR_ODD }, | |
841 | }; | |
842 | ||
939991cf DD |
843 | static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = { |
844 | [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL }, | |
845 | [PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL }, | |
846 | [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL }, | |
70342287 | 847 | [PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL }, |
939991cf DD |
848 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL }, |
849 | [PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL }, | |
850 | [PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL }, | |
851 | }; | |
852 | ||
a7911a8f AC |
853 | static const struct mips_perf_event bmips5000_event_map |
854 | [PERF_COUNT_HW_MAX] = { | |
855 | [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, T }, | |
856 | [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T }, | |
857 | [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T }, | |
858 | }; | |
859 | ||
4be3d2f3 ZSL |
860 | static const struct mips_perf_event xlp_event_map[PERF_COUNT_HW_MAX] = { |
861 | [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL }, | |
862 | [PERF_COUNT_HW_INSTRUCTIONS] = { 0x18, CNTR_ALL }, /* PAPI_TOT_INS */ | |
863 | [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */ | |
864 | [PERF_COUNT_HW_CACHE_MISSES] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */ | |
865 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x1b, CNTR_ALL }, /* PAPI_BR_CN */ | |
866 | [PERF_COUNT_HW_BRANCH_MISSES] = { 0x1c, CNTR_ALL }, /* PAPI_BR_MSP */ | |
4be3d2f3 ZSL |
867 | }; |
868 | ||
9597e432 | 869 | /* 24K/34K/1004K/interAptiv/loongson1 cores share the same cache event map. */ |
3a9ab99e DCZ |
870 | static const struct mips_perf_event mipsxxcore_cache_map |
871 | [PERF_COUNT_HW_CACHE_MAX] | |
872 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
873 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = { | |
874 | [C(L1D)] = { | |
875 | /* | |
876 | * Like some other architectures (e.g. ARM), the performance | |
877 | * counters don't differentiate between read and write | |
878 | * accesses/misses, so this isn't strictly correct, but it's the | |
879 | * best we can do. Writes and reads get combined. | |
880 | */ | |
881 | [C(OP_READ)] = { | |
882 | [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T }, | |
883 | [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T }, | |
884 | }, | |
885 | [C(OP_WRITE)] = { | |
886 | [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T }, | |
887 | [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T }, | |
888 | }, | |
3a9ab99e DCZ |
889 | }, |
890 | [C(L1I)] = { | |
891 | [C(OP_READ)] = { | |
892 | [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T }, | |
893 | [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T }, | |
894 | }, | |
895 | [C(OP_WRITE)] = { | |
896 | [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T }, | |
897 | [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T }, | |
898 | }, | |
899 | [C(OP_PREFETCH)] = { | |
900 | [C(RESULT_ACCESS)] = { 0x14, CNTR_EVEN, T }, | |
901 | /* | |
902 | * Note that MIPS has only "hit" events countable for | |
903 | * the prefetch operation. | |
904 | */ | |
3a9ab99e DCZ |
905 | }, |
906 | }, | |
907 | [C(LL)] = { | |
908 | [C(OP_READ)] = { | |
909 | [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P }, | |
910 | [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P }, | |
911 | }, | |
912 | [C(OP_WRITE)] = { | |
913 | [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P }, | |
914 | [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P }, | |
915 | }, | |
3a9ab99e DCZ |
916 | }, |
917 | [C(DTLB)] = { | |
918 | [C(OP_READ)] = { | |
919 | [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, | |
920 | [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, | |
921 | }, | |
922 | [C(OP_WRITE)] = { | |
923 | [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, | |
924 | [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, | |
925 | }, | |
3a9ab99e DCZ |
926 | }, |
927 | [C(ITLB)] = { | |
928 | [C(OP_READ)] = { | |
929 | [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T }, | |
930 | [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T }, | |
931 | }, | |
932 | [C(OP_WRITE)] = { | |
933 | [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T }, | |
934 | [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T }, | |
935 | }, | |
3a9ab99e DCZ |
936 | }, |
937 | [C(BPU)] = { | |
938 | /* Using the same code for *HW_BRANCH* */ | |
939 | [C(OP_READ)] = { | |
940 | [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T }, | |
941 | [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T }, | |
942 | }, | |
943 | [C(OP_WRITE)] = { | |
944 | [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T }, | |
945 | [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T }, | |
946 | }, | |
89d6c0b5 | 947 | }, |
3a9ab99e DCZ |
948 | }; |
949 | ||
c52068bd | 950 | /* 74K/proAptiv core has completely different cache event map. */ |
6b0b8429 | 951 | static const struct mips_perf_event mipsxxcore_cache_map2 |
3a9ab99e DCZ |
952 | [PERF_COUNT_HW_CACHE_MAX] |
953 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
954 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = { | |
955 | [C(L1D)] = { | |
956 | /* | |
957 | * Like some other architectures (e.g. ARM), the performance | |
958 | * counters don't differentiate between read and write | |
959 | * accesses/misses, so this isn't strictly correct, but it's the | |
960 | * best we can do. Writes and reads get combined. | |
961 | */ | |
962 | [C(OP_READ)] = { | |
963 | [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T }, | |
964 | [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T }, | |
965 | }, | |
966 | [C(OP_WRITE)] = { | |
967 | [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T }, | |
968 | [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T }, | |
969 | }, | |
3a9ab99e DCZ |
970 | }, |
971 | [C(L1I)] = { | |
972 | [C(OP_READ)] = { | |
973 | [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, | |
974 | [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, | |
975 | }, | |
976 | [C(OP_WRITE)] = { | |
977 | [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, | |
978 | [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, | |
979 | }, | |
980 | [C(OP_PREFETCH)] = { | |
981 | [C(RESULT_ACCESS)] = { 0x34, CNTR_EVEN, T }, | |
982 | /* | |
983 | * Note that MIPS has only "hit" events countable for | |
984 | * the prefetch operation. | |
985 | */ | |
3a9ab99e DCZ |
986 | }, |
987 | }, | |
988 | [C(LL)] = { | |
989 | [C(OP_READ)] = { | |
990 | [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P }, | |
7f081f17 | 991 | [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN, P }, |
3a9ab99e DCZ |
992 | }, |
993 | [C(OP_WRITE)] = { | |
994 | [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P }, | |
7f081f17 | 995 | [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN, P }, |
3a9ab99e | 996 | }, |
3a9ab99e | 997 | }, |
c52068bd DCZ |
998 | /* |
999 | * 74K core does not have specific DTLB events. proAptiv core has | |
1000 | * "speculative" DTLB events which are numbered 0x63 (even/odd) and | |
1001 | * not included here. One can use raw events if really needed. | |
1002 | */ | |
3a9ab99e DCZ |
1003 | [C(ITLB)] = { |
1004 | [C(OP_READ)] = { | |
1005 | [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T }, | |
1006 | [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T }, | |
1007 | }, | |
1008 | [C(OP_WRITE)] = { | |
1009 | [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T }, | |
1010 | [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T }, | |
1011 | }, | |
3a9ab99e DCZ |
1012 | }, |
1013 | [C(BPU)] = { | |
1014 | /* Using the same code for *HW_BRANCH* */ | |
1015 | [C(OP_READ)] = { | |
1016 | [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T }, | |
1017 | [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T }, | |
1018 | }, | |
1019 | [C(OP_WRITE)] = { | |
1020 | [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T }, | |
1021 | [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T }, | |
1022 | }, | |
89d6c0b5 | 1023 | }, |
3a9ab99e DCZ |
1024 | }; |
1025 | ||
fd716fca JH |
1026 | static const struct mips_perf_event i6400_cache_map |
1027 | [PERF_COUNT_HW_CACHE_MAX] | |
1028 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1029 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = { | |
1030 | [C(L1D)] = { | |
1031 | [C(OP_READ)] = { | |
1032 | [C(RESULT_ACCESS)] = { 0x46, CNTR_EVEN | CNTR_ODD }, | |
1033 | [C(RESULT_MISS)] = { 0x49, CNTR_EVEN | CNTR_ODD }, | |
1034 | }, | |
1035 | [C(OP_WRITE)] = { | |
1036 | [C(RESULT_ACCESS)] = { 0x47, CNTR_EVEN | CNTR_ODD }, | |
1037 | [C(RESULT_MISS)] = { 0x4a, CNTR_EVEN | CNTR_ODD }, | |
1038 | }, | |
1039 | }, | |
1040 | [C(L1I)] = { | |
1041 | [C(OP_READ)] = { | |
1042 | [C(RESULT_ACCESS)] = { 0x84, CNTR_EVEN | CNTR_ODD }, | |
1043 | [C(RESULT_MISS)] = { 0x85, CNTR_EVEN | CNTR_ODD }, | |
1044 | }, | |
1045 | }, | |
1046 | [C(DTLB)] = { | |
1047 | /* Can't distinguish read & write */ | |
1048 | [C(OP_READ)] = { | |
1049 | [C(RESULT_ACCESS)] = { 0x40, CNTR_EVEN | CNTR_ODD }, | |
1050 | [C(RESULT_MISS)] = { 0x41, CNTR_EVEN | CNTR_ODD }, | |
1051 | }, | |
1052 | [C(OP_WRITE)] = { | |
1053 | [C(RESULT_ACCESS)] = { 0x40, CNTR_EVEN | CNTR_ODD }, | |
1054 | [C(RESULT_MISS)] = { 0x41, CNTR_EVEN | CNTR_ODD }, | |
1055 | }, | |
1056 | }, | |
1057 | [C(BPU)] = { | |
1058 | /* Conditional branches / mispredicted */ | |
1059 | [C(OP_READ)] = { | |
1060 | [C(RESULT_ACCESS)] = { 0x15, CNTR_EVEN | CNTR_ODD }, | |
1061 | [C(RESULT_MISS)] = { 0x16, CNTR_EVEN | CNTR_ODD }, | |
1062 | }, | |
1063 | }, | |
1064 | }; | |
1065 | ||
f14ceff7 HC |
1066 | static const struct mips_perf_event loongson3_cache_map |
1067 | [PERF_COUNT_HW_CACHE_MAX] | |
1068 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1069 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = { | |
1070 | [C(L1D)] = { | |
1071 | /* | |
1072 | * Like some other architectures (e.g. ARM), the performance | |
1073 | * counters don't differentiate between read and write | |
1074 | * accesses/misses, so this isn't strictly correct, but it's the | |
1075 | * best we can do. Writes and reads get combined. | |
1076 | */ | |
1077 | [C(OP_READ)] = { | |
1078 | [C(RESULT_MISS)] = { 0x04, CNTR_ODD }, | |
1079 | }, | |
1080 | [C(OP_WRITE)] = { | |
1081 | [C(RESULT_MISS)] = { 0x04, CNTR_ODD }, | |
1082 | }, | |
1083 | }, | |
1084 | [C(L1I)] = { | |
1085 | [C(OP_READ)] = { | |
1086 | [C(RESULT_MISS)] = { 0x04, CNTR_EVEN }, | |
1087 | }, | |
1088 | [C(OP_WRITE)] = { | |
1089 | [C(RESULT_MISS)] = { 0x04, CNTR_EVEN }, | |
1090 | }, | |
1091 | }, | |
1092 | [C(DTLB)] = { | |
1093 | [C(OP_READ)] = { | |
1094 | [C(RESULT_MISS)] = { 0x09, CNTR_ODD }, | |
1095 | }, | |
1096 | [C(OP_WRITE)] = { | |
1097 | [C(RESULT_MISS)] = { 0x09, CNTR_ODD }, | |
1098 | }, | |
1099 | }, | |
1100 | [C(ITLB)] = { | |
1101 | [C(OP_READ)] = { | |
1102 | [C(RESULT_MISS)] = { 0x0c, CNTR_ODD }, | |
1103 | }, | |
1104 | [C(OP_WRITE)] = { | |
1105 | [C(RESULT_MISS)] = { 0x0c, CNTR_ODD }, | |
1106 | }, | |
1107 | }, | |
1108 | [C(BPU)] = { | |
1109 | /* Using the same code for *HW_BRANCH* */ | |
1110 | [C(OP_READ)] = { | |
1111 | [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN }, | |
1112 | [C(RESULT_MISS)] = { 0x02, CNTR_ODD }, | |
1113 | }, | |
1114 | [C(OP_WRITE)] = { | |
1115 | [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN }, | |
1116 | [C(RESULT_MISS)] = { 0x02, CNTR_ODD }, | |
1117 | }, | |
1118 | }, | |
1119 | }; | |
1120 | ||
a7911a8f AC |
1121 | /* BMIPS5000 */ |
1122 | static const struct mips_perf_event bmips5000_cache_map | |
1123 | [PERF_COUNT_HW_CACHE_MAX] | |
1124 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1125 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = { | |
1126 | [C(L1D)] = { | |
1127 | /* | |
1128 | * Like some other architectures (e.g. ARM), the performance | |
1129 | * counters don't differentiate between read and write | |
1130 | * accesses/misses, so this isn't strictly correct, but it's the | |
1131 | * best we can do. Writes and reads get combined. | |
1132 | */ | |
1133 | [C(OP_READ)] = { | |
1134 | [C(RESULT_ACCESS)] = { 12, CNTR_EVEN, T }, | |
1135 | [C(RESULT_MISS)] = { 12, CNTR_ODD, T }, | |
1136 | }, | |
1137 | [C(OP_WRITE)] = { | |
1138 | [C(RESULT_ACCESS)] = { 12, CNTR_EVEN, T }, | |
1139 | [C(RESULT_MISS)] = { 12, CNTR_ODD, T }, | |
1140 | }, | |
1141 | }, | |
1142 | [C(L1I)] = { | |
1143 | [C(OP_READ)] = { | |
1144 | [C(RESULT_ACCESS)] = { 10, CNTR_EVEN, T }, | |
1145 | [C(RESULT_MISS)] = { 10, CNTR_ODD, T }, | |
1146 | }, | |
1147 | [C(OP_WRITE)] = { | |
1148 | [C(RESULT_ACCESS)] = { 10, CNTR_EVEN, T }, | |
1149 | [C(RESULT_MISS)] = { 10, CNTR_ODD, T }, | |
1150 | }, | |
1151 | [C(OP_PREFETCH)] = { | |
1152 | [C(RESULT_ACCESS)] = { 23, CNTR_EVEN, T }, | |
1153 | /* | |
1154 | * Note that MIPS has only "hit" events countable for | |
1155 | * the prefetch operation. | |
1156 | */ | |
1157 | }, | |
1158 | }, | |
1159 | [C(LL)] = { | |
1160 | [C(OP_READ)] = { | |
1161 | [C(RESULT_ACCESS)] = { 28, CNTR_EVEN, P }, | |
1162 | [C(RESULT_MISS)] = { 28, CNTR_ODD, P }, | |
1163 | }, | |
1164 | [C(OP_WRITE)] = { | |
1165 | [C(RESULT_ACCESS)] = { 28, CNTR_EVEN, P }, | |
1166 | [C(RESULT_MISS)] = { 28, CNTR_ODD, P }, | |
1167 | }, | |
1168 | }, | |
1169 | [C(BPU)] = { | |
1170 | /* Using the same code for *HW_BRANCH* */ | |
1171 | [C(OP_READ)] = { | |
1172 | [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T }, | |
1173 | }, | |
1174 | [C(OP_WRITE)] = { | |
1175 | [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T }, | |
1176 | }, | |
1177 | }, | |
1178 | }; | |
1179 | ||
939991cf DD |
1180 | |
1181 | static const struct mips_perf_event octeon_cache_map | |
1182 | [PERF_COUNT_HW_CACHE_MAX] | |
1183 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1184 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = { | |
1185 | [C(L1D)] = { | |
1186 | [C(OP_READ)] = { | |
1187 | [C(RESULT_ACCESS)] = { 0x2b, CNTR_ALL }, | |
1188 | [C(RESULT_MISS)] = { 0x2e, CNTR_ALL }, | |
1189 | }, | |
1190 | [C(OP_WRITE)] = { | |
1191 | [C(RESULT_ACCESS)] = { 0x30, CNTR_ALL }, | |
939991cf DD |
1192 | }, |
1193 | }, | |
1194 | [C(L1I)] = { | |
1195 | [C(OP_READ)] = { | |
1196 | [C(RESULT_ACCESS)] = { 0x18, CNTR_ALL }, | |
939991cf DD |
1197 | }, |
1198 | [C(OP_PREFETCH)] = { | |
1199 | [C(RESULT_ACCESS)] = { 0x19, CNTR_ALL }, | |
939991cf DD |
1200 | }, |
1201 | }, | |
1202 | [C(DTLB)] = { | |
1203 | /* | |
1204 | * Only general DTLB misses are counted use the same event for | |
1205 | * read and write. | |
1206 | */ | |
1207 | [C(OP_READ)] = { | |
939991cf DD |
1208 | [C(RESULT_MISS)] = { 0x35, CNTR_ALL }, |
1209 | }, | |
1210 | [C(OP_WRITE)] = { | |
939991cf DD |
1211 | [C(RESULT_MISS)] = { 0x35, CNTR_ALL }, |
1212 | }, | |
939991cf DD |
1213 | }, |
1214 | [C(ITLB)] = { | |
1215 | [C(OP_READ)] = { | |
939991cf DD |
1216 | [C(RESULT_MISS)] = { 0x37, CNTR_ALL }, |
1217 | }, | |
939991cf DD |
1218 | }, |
1219 | }; | |
1220 | ||
4be3d2f3 ZSL |
1221 | static const struct mips_perf_event xlp_cache_map |
1222 | [PERF_COUNT_HW_CACHE_MAX] | |
1223 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
1224 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = { | |
1225 | [C(L1D)] = { | |
1226 | [C(OP_READ)] = { | |
1227 | [C(RESULT_ACCESS)] = { 0x31, CNTR_ALL }, /* PAPI_L1_DCR */ | |
1228 | [C(RESULT_MISS)] = { 0x30, CNTR_ALL }, /* PAPI_L1_LDM */ | |
1229 | }, | |
1230 | [C(OP_WRITE)] = { | |
1231 | [C(RESULT_ACCESS)] = { 0x2f, CNTR_ALL }, /* PAPI_L1_DCW */ | |
1232 | [C(RESULT_MISS)] = { 0x2e, CNTR_ALL }, /* PAPI_L1_STM */ | |
1233 | }, | |
4be3d2f3 ZSL |
1234 | }, |
1235 | [C(L1I)] = { | |
1236 | [C(OP_READ)] = { | |
1237 | [C(RESULT_ACCESS)] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */ | |
1238 | [C(RESULT_MISS)] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */ | |
1239 | }, | |
4be3d2f3 ZSL |
1240 | }, |
1241 | [C(LL)] = { | |
1242 | [C(OP_READ)] = { | |
1243 | [C(RESULT_ACCESS)] = { 0x35, CNTR_ALL }, /* PAPI_L2_DCR */ | |
1244 | [C(RESULT_MISS)] = { 0x37, CNTR_ALL }, /* PAPI_L2_LDM */ | |
1245 | }, | |
1246 | [C(OP_WRITE)] = { | |
1247 | [C(RESULT_ACCESS)] = { 0x34, CNTR_ALL }, /* PAPI_L2_DCA */ | |
1248 | [C(RESULT_MISS)] = { 0x36, CNTR_ALL }, /* PAPI_L2_DCM */ | |
1249 | }, | |
4be3d2f3 ZSL |
1250 | }, |
1251 | [C(DTLB)] = { | |
1252 | /* | |
1253 | * Only general DTLB misses are counted use the same event for | |
1254 | * read and write. | |
1255 | */ | |
1256 | [C(OP_READ)] = { | |
4be3d2f3 ZSL |
1257 | [C(RESULT_MISS)] = { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */ |
1258 | }, | |
1259 | [C(OP_WRITE)] = { | |
4be3d2f3 ZSL |
1260 | [C(RESULT_MISS)] = { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */ |
1261 | }, | |
4be3d2f3 ZSL |
1262 | }, |
1263 | [C(ITLB)] = { | |
1264 | [C(OP_READ)] = { | |
4be3d2f3 ZSL |
1265 | [C(RESULT_MISS)] = { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */ |
1266 | }, | |
1267 | [C(OP_WRITE)] = { | |
4be3d2f3 ZSL |
1268 | [C(RESULT_MISS)] = { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */ |
1269 | }, | |
4be3d2f3 ZSL |
1270 | }, |
1271 | [C(BPU)] = { | |
1272 | [C(OP_READ)] = { | |
4be3d2f3 ZSL |
1273 | [C(RESULT_MISS)] = { 0x25, CNTR_ALL }, |
1274 | }, | |
4be3d2f3 ZSL |
1275 | }, |
1276 | }; | |
1277 | ||
3a9ab99e | 1278 | #ifdef CONFIG_MIPS_MT_SMP |
4409af37 DD |
1279 | static void check_and_calc_range(struct perf_event *event, |
1280 | const struct mips_perf_event *pev) | |
3a9ab99e DCZ |
1281 | { |
1282 | struct hw_perf_event *hwc = &event->hw; | |
1283 | ||
1284 | if (event->cpu >= 0) { | |
1285 | if (pev->range > V) { | |
1286 | /* | |
1287 | * The user selected an event that is processor | |
1288 | * wide, while expecting it to be VPE wide. | |
1289 | */ | |
1290 | hwc->config_base |= M_TC_EN_ALL; | |
1291 | } else { | |
1292 | /* | |
1293 | * FIXME: cpu_data[event->cpu].vpe_id reports 0 | |
1294 | * for both CPUs. | |
1295 | */ | |
1296 | hwc->config_base |= M_PERFCTL_VPEID(event->cpu); | |
1297 | hwc->config_base |= M_TC_EN_VPE; | |
1298 | } | |
1299 | } else | |
1300 | hwc->config_base |= M_TC_EN_ALL; | |
1301 | } | |
1302 | #else | |
4409af37 DD |
1303 | static void check_and_calc_range(struct perf_event *event, |
1304 | const struct mips_perf_event *pev) | |
3a9ab99e DCZ |
1305 | { |
1306 | } | |
1307 | #endif | |
1308 | ||
1309 | static int __hw_perf_event_init(struct perf_event *event) | |
1310 | { | |
1311 | struct perf_event_attr *attr = &event->attr; | |
1312 | struct hw_perf_event *hwc = &event->hw; | |
1313 | const struct mips_perf_event *pev; | |
1314 | int err; | |
1315 | ||
1316 | /* Returning MIPS event descriptor for generic perf event. */ | |
1317 | if (PERF_TYPE_HARDWARE == event->attr.type) { | |
1318 | if (event->attr.config >= PERF_COUNT_HW_MAX) | |
1319 | return -EINVAL; | |
1320 | pev = mipspmu_map_general_event(event->attr.config); | |
1321 | } else if (PERF_TYPE_HW_CACHE == event->attr.type) { | |
1322 | pev = mipspmu_map_cache_event(event->attr.config); | |
1323 | } else if (PERF_TYPE_RAW == event->attr.type) { | |
1324 | /* We are working on the global raw event. */ | |
1325 | mutex_lock(&raw_event_mutex); | |
82091564 | 1326 | pev = mipspmu.map_raw_event(event->attr.config); |
3a9ab99e DCZ |
1327 | } else { |
1328 | /* The event type is not (yet) supported. */ | |
1329 | return -EOPNOTSUPP; | |
1330 | } | |
1331 | ||
1332 | if (IS_ERR(pev)) { | |
1333 | if (PERF_TYPE_RAW == event->attr.type) | |
1334 | mutex_unlock(&raw_event_mutex); | |
1335 | return PTR_ERR(pev); | |
1336 | } | |
1337 | ||
1338 | /* | |
1339 | * We allow max flexibility on how each individual counter shared | |
1340 | * by the single CPU operates (the mode exclusion and the range). | |
1341 | */ | |
1342 | hwc->config_base = M_PERFCTL_INTERRUPT_ENABLE; | |
1343 | ||
1344 | /* Calculate range bits and validate it. */ | |
1345 | if (num_possible_cpus() > 1) | |
1346 | check_and_calc_range(event, pev); | |
1347 | ||
1348 | hwc->event_base = mipspmu_perf_event_encode(pev); | |
1349 | if (PERF_TYPE_RAW == event->attr.type) | |
1350 | mutex_unlock(&raw_event_mutex); | |
1351 | ||
1352 | if (!attr->exclude_user) | |
1353 | hwc->config_base |= M_PERFCTL_USER; | |
1354 | if (!attr->exclude_kernel) { | |
1355 | hwc->config_base |= M_PERFCTL_KERNEL; | |
1356 | /* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */ | |
1357 | hwc->config_base |= M_PERFCTL_EXL; | |
1358 | } | |
1359 | if (!attr->exclude_hv) | |
1360 | hwc->config_base |= M_PERFCTL_SUPERVISOR; | |
1361 | ||
1362 | hwc->config_base &= M_PERFCTL_CONFIG_MASK; | |
1363 | /* | |
1364 | * The event can belong to another cpu. We do not assign a local | |
1365 | * counter for it for now. | |
1366 | */ | |
1367 | hwc->idx = -1; | |
1368 | hwc->config = 0; | |
1369 | ||
1370 | if (!hwc->sample_period) { | |
82091564 | 1371 | hwc->sample_period = mipspmu.max_period; |
3a9ab99e DCZ |
1372 | hwc->last_period = hwc->sample_period; |
1373 | local64_set(&hwc->period_left, hwc->sample_period); | |
1374 | } | |
1375 | ||
1376 | err = 0; | |
ff5d7265 | 1377 | if (event->group_leader != event) |
3a9ab99e | 1378 | err = validate_group(event); |
3a9ab99e DCZ |
1379 | |
1380 | event->destroy = hw_perf_event_destroy; | |
ff5d7265 DCZ |
1381 | |
1382 | if (err) | |
1383 | event->destroy(event); | |
1384 | ||
3a9ab99e DCZ |
1385 | return err; |
1386 | } | |
1387 | ||
1388 | static void pause_local_counters(void) | |
1389 | { | |
35898716 | 1390 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
82091564 | 1391 | int ctr = mipspmu.num_counters; |
3a9ab99e DCZ |
1392 | unsigned long flags; |
1393 | ||
1394 | local_irq_save(flags); | |
82091564 DD |
1395 | do { |
1396 | ctr--; | |
1397 | cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr); | |
1398 | mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] & | |
1399 | ~M_PERFCTL_COUNT_EVENT_WHENEVER); | |
1400 | } while (ctr > 0); | |
3a9ab99e DCZ |
1401 | local_irq_restore(flags); |
1402 | } | |
1403 | ||
1404 | static void resume_local_counters(void) | |
1405 | { | |
35898716 | 1406 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
82091564 | 1407 | int ctr = mipspmu.num_counters; |
3a9ab99e | 1408 | |
82091564 DD |
1409 | do { |
1410 | ctr--; | |
1411 | mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]); | |
1412 | } while (ctr > 0); | |
3a9ab99e DCZ |
1413 | } |
1414 | ||
1415 | static int mipsxx_pmu_handle_shared_irq(void) | |
1416 | { | |
35898716 | 1417 | struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); |
3a9ab99e | 1418 | struct perf_sample_data data; |
82091564 DD |
1419 | unsigned int counters = mipspmu.num_counters; |
1420 | u64 counter; | |
3a9ab99e DCZ |
1421 | int handled = IRQ_NONE; |
1422 | struct pt_regs *regs; | |
1423 | ||
da4b62cd | 1424 | if (cpu_has_perf_cntr_intr_bit && !(read_c0_cause() & CAUSEF_PCI)) |
3a9ab99e | 1425 | return handled; |
3a9ab99e DCZ |
1426 | /* |
1427 | * First we pause the local counters, so that when we are locked | |
1428 | * here, the counters are all paused. When it gets locked due to | |
1429 | * perf_disable(), the timer interrupt handler will be delayed. | |
1430 | * | |
1431 | * See also mipsxx_pmu_start(). | |
1432 | */ | |
1433 | pause_local_counters(); | |
399aaa25 | 1434 | #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS |
3a9ab99e DCZ |
1435 | read_lock(&pmuint_rwlock); |
1436 | #endif | |
1437 | ||
1438 | regs = get_irq_regs(); | |
1439 | ||
fd0d000b | 1440 | perf_sample_data_init(&data, 0, 0); |
3a9ab99e DCZ |
1441 | |
1442 | switch (counters) { | |
1443 | #define HANDLE_COUNTER(n) \ | |
1444 | case n + 1: \ | |
1445 | if (test_bit(n, cpuc->used_mask)) { \ | |
82091564 DD |
1446 | counter = mipspmu.read_counter(n); \ |
1447 | if (counter & mipspmu.overflow) { \ | |
1448 | handle_associated_event(cpuc, n, &data, regs); \ | |
3a9ab99e DCZ |
1449 | handled = IRQ_HANDLED; \ |
1450 | } \ | |
1451 | } | |
1452 | HANDLE_COUNTER(3) | |
1453 | HANDLE_COUNTER(2) | |
1454 | HANDLE_COUNTER(1) | |
1455 | HANDLE_COUNTER(0) | |
1456 | } | |
1457 | ||
1458 | /* | |
1459 | * Do all the work for the pending perf events. We can do this | |
1460 | * in here because the performance counter interrupt is a regular | |
1461 | * interrupt, not NMI. | |
1462 | */ | |
1463 | if (handled == IRQ_HANDLED) | |
91f01737 | 1464 | irq_work_run(); |
3a9ab99e | 1465 | |
399aaa25 | 1466 | #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS |
3a9ab99e DCZ |
1467 | read_unlock(&pmuint_rwlock); |
1468 | #endif | |
1469 | resume_local_counters(); | |
1470 | return handled; | |
1471 | } | |
1472 | ||
4409af37 | 1473 | static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev) |
3a9ab99e DCZ |
1474 | { |
1475 | return mipsxx_pmu_handle_shared_irq(); | |
1476 | } | |
1477 | ||
3a9ab99e | 1478 | /* 24K */ |
3a9ab99e DCZ |
1479 | #define IS_BOTH_COUNTERS_24K_EVENT(b) \ |
1480 | ((b) == 0 || (b) == 1 || (b) == 11) | |
1481 | ||
1482 | /* 34K */ | |
3a9ab99e DCZ |
1483 | #define IS_BOTH_COUNTERS_34K_EVENT(b) \ |
1484 | ((b) == 0 || (b) == 1 || (b) == 11) | |
1485 | #ifdef CONFIG_MIPS_MT_SMP | |
1486 | #define IS_RANGE_P_34K_EVENT(r, b) \ | |
1487 | ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \ | |
1488 | (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 || \ | |
1489 | (r) == 176 || ((b) >= 50 && (b) <= 55) || \ | |
1490 | ((b) >= 64 && (b) <= 67)) | |
70342287 | 1491 | #define IS_RANGE_V_34K_EVENT(r) ((r) == 47) |
3a9ab99e DCZ |
1492 | #endif |
1493 | ||
1494 | /* 74K */ | |
3a9ab99e DCZ |
1495 | #define IS_BOTH_COUNTERS_74K_EVENT(b) \ |
1496 | ((b) == 0 || (b) == 1) | |
1497 | ||
c52068bd DCZ |
1498 | /* proAptiv */ |
1499 | #define IS_BOTH_COUNTERS_PROAPTIV_EVENT(b) \ | |
1500 | ((b) == 0 || (b) == 1) | |
560b461b JH |
1501 | /* P5600 */ |
1502 | #define IS_BOTH_COUNTERS_P5600_EVENT(b) \ | |
1503 | ((b) == 0 || (b) == 1) | |
c52068bd | 1504 | |
3a9ab99e | 1505 | /* 1004K */ |
3a9ab99e DCZ |
1506 | #define IS_BOTH_COUNTERS_1004K_EVENT(b) \ |
1507 | ((b) == 0 || (b) == 1 || (b) == 11) | |
1508 | #ifdef CONFIG_MIPS_MT_SMP | |
1509 | #define IS_RANGE_P_1004K_EVENT(r, b) \ | |
1510 | ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \ | |
1511 | (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 || \ | |
1512 | (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) || \ | |
1513 | (r) == 188 || (b) == 61 || (b) == 62 || \ | |
1514 | ((b) >= 64 && (b) <= 67)) | |
1515 | #define IS_RANGE_V_1004K_EVENT(r) ((r) == 47) | |
1516 | #endif | |
1517 | ||
9597e432 DCZ |
1518 | /* interAptiv */ |
1519 | #define IS_BOTH_COUNTERS_INTERAPTIV_EVENT(b) \ | |
1520 | ((b) == 0 || (b) == 1 || (b) == 11) | |
1521 | #ifdef CONFIG_MIPS_MT_SMP | |
1522 | /* The P/V/T info is not provided for "(b) == 38" in SUM, assume P. */ | |
1523 | #define IS_RANGE_P_INTERAPTIV_EVENT(r, b) \ | |
1524 | ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \ | |
1525 | (b) == 25 || (b) == 36 || (b) == 38 || (b) == 39 || \ | |
1526 | (r) == 44 || (r) == 174 || (r) == 176 || ((b) >= 50 && \ | |
1527 | (b) <= 59) || (r) == 188 || (b) == 61 || (b) == 62 || \ | |
1528 | ((b) >= 64 && (b) <= 67)) | |
1529 | #define IS_RANGE_V_INTERAPTIV_EVENT(r) ((r) == 47 || (r) == 175) | |
1530 | #endif | |
1531 | ||
a7911a8f AC |
1532 | /* BMIPS5000 */ |
1533 | #define IS_BOTH_COUNTERS_BMIPS5000_EVENT(b) \ | |
1534 | ((b) == 0 || (b) == 1) | |
1535 | ||
1536 | ||
3a9ab99e | 1537 | /* |
67dca667 JH |
1538 | * For most cores the user can use 0-255 raw events, where 0-127 for the events |
1539 | * of even counters, and 128-255 for odd counters. Note that bit 7 is used to | |
1540 | * indicate the even/odd bank selector. So, for example, when user wants to take | |
1541 | * the Event Num of 15 for odd counters (by referring to the user manual), then | |
1542 | * 128 needs to be added to 15 as the input for the event config, i.e., 143 (0x8F) | |
1543 | * to be used. | |
1544 | * | |
1545 | * Some newer cores have even more events, in which case the user can use raw | |
1546 | * events 0-511, where 0-255 are for the events of even counters, and 256-511 | |
1547 | * are for odd counters, so bit 8 is used to indicate the even/odd bank selector. | |
3a9ab99e | 1548 | */ |
4409af37 | 1549 | static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config) |
3a9ab99e | 1550 | { |
67dca667 | 1551 | /* currently most cores have 7-bit event numbers */ |
3a9ab99e DCZ |
1552 | unsigned int raw_id = config & 0xff; |
1553 | unsigned int base_id = raw_id & 0x7f; | |
1554 | ||
1555 | switch (current_cpu_type()) { | |
1556 | case CPU_24K: | |
3a9ab99e DCZ |
1557 | if (IS_BOTH_COUNTERS_24K_EVENT(base_id)) |
1558 | raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; | |
1559 | else | |
1560 | raw_event.cntr_mask = | |
1561 | raw_id > 127 ? CNTR_ODD : CNTR_EVEN; | |
1562 | #ifdef CONFIG_MIPS_MT_SMP | |
1563 | /* | |
1564 | * This is actually doing nothing. Non-multithreading | |
1565 | * CPUs will not check and calculate the range. | |
1566 | */ | |
1567 | raw_event.range = P; | |
1568 | #endif | |
1569 | break; | |
1570 | case CPU_34K: | |
3a9ab99e DCZ |
1571 | if (IS_BOTH_COUNTERS_34K_EVENT(base_id)) |
1572 | raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; | |
1573 | else | |
1574 | raw_event.cntr_mask = | |
1575 | raw_id > 127 ? CNTR_ODD : CNTR_EVEN; | |
1576 | #ifdef CONFIG_MIPS_MT_SMP | |
1577 | if (IS_RANGE_P_34K_EVENT(raw_id, base_id)) | |
1578 | raw_event.range = P; | |
1579 | else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id))) | |
1580 | raw_event.range = V; | |
1581 | else | |
1582 | raw_event.range = T; | |
1583 | #endif | |
1584 | break; | |
1585 | case CPU_74K: | |
442e14a2 | 1586 | case CPU_1074K: |
3a9ab99e DCZ |
1587 | if (IS_BOTH_COUNTERS_74K_EVENT(base_id)) |
1588 | raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; | |
1589 | else | |
1590 | raw_event.cntr_mask = | |
1591 | raw_id > 127 ? CNTR_ODD : CNTR_EVEN; | |
1592 | #ifdef CONFIG_MIPS_MT_SMP | |
1593 | raw_event.range = P; | |
c52068bd DCZ |
1594 | #endif |
1595 | break; | |
1596 | case CPU_PROAPTIV: | |
1597 | if (IS_BOTH_COUNTERS_PROAPTIV_EVENT(base_id)) | |
1598 | raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; | |
1599 | else | |
1600 | raw_event.cntr_mask = | |
1601 | raw_id > 127 ? CNTR_ODD : CNTR_EVEN; | |
1602 | #ifdef CONFIG_MIPS_MT_SMP | |
1603 | raw_event.range = P; | |
560b461b JH |
1604 | #endif |
1605 | break; | |
1606 | case CPU_P5600: | |
1091bfa2 | 1607 | case CPU_P6600: |
4e88a862 | 1608 | case CPU_I6400: |
560b461b JH |
1609 | /* 8-bit event numbers */ |
1610 | raw_id = config & 0x1ff; | |
1611 | base_id = raw_id & 0xff; | |
1612 | if (IS_BOTH_COUNTERS_P5600_EVENT(base_id)) | |
1613 | raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; | |
1614 | else | |
1615 | raw_event.cntr_mask = | |
1616 | raw_id > 255 ? CNTR_ODD : CNTR_EVEN; | |
1617 | #ifdef CONFIG_MIPS_MT_SMP | |
1618 | raw_event.range = P; | |
3a9ab99e DCZ |
1619 | #endif |
1620 | break; | |
1621 | case CPU_1004K: | |
3a9ab99e DCZ |
1622 | if (IS_BOTH_COUNTERS_1004K_EVENT(base_id)) |
1623 | raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; | |
1624 | else | |
1625 | raw_event.cntr_mask = | |
1626 | raw_id > 127 ? CNTR_ODD : CNTR_EVEN; | |
1627 | #ifdef CONFIG_MIPS_MT_SMP | |
1628 | if (IS_RANGE_P_1004K_EVENT(raw_id, base_id)) | |
1629 | raw_event.range = P; | |
1630 | else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id))) | |
1631 | raw_event.range = V; | |
1632 | else | |
1633 | raw_event.range = T; | |
9597e432 DCZ |
1634 | #endif |
1635 | break; | |
1636 | case CPU_INTERAPTIV: | |
1637 | if (IS_BOTH_COUNTERS_INTERAPTIV_EVENT(base_id)) | |
1638 | raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; | |
1639 | else | |
1640 | raw_event.cntr_mask = | |
1641 | raw_id > 127 ? CNTR_ODD : CNTR_EVEN; | |
1642 | #ifdef CONFIG_MIPS_MT_SMP | |
1643 | if (IS_RANGE_P_INTERAPTIV_EVENT(raw_id, base_id)) | |
1644 | raw_event.range = P; | |
1645 | else if (unlikely(IS_RANGE_V_INTERAPTIV_EVENT(raw_id))) | |
1646 | raw_event.range = V; | |
1647 | else | |
1648 | raw_event.range = T; | |
3a9ab99e DCZ |
1649 | #endif |
1650 | break; | |
a7911a8f AC |
1651 | case CPU_BMIPS5000: |
1652 | if (IS_BOTH_COUNTERS_BMIPS5000_EVENT(base_id)) | |
1653 | raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; | |
1654 | else | |
1655 | raw_event.cntr_mask = | |
1656 | raw_id > 127 ? CNTR_ODD : CNTR_EVEN; | |
f14ceff7 HC |
1657 | break; |
1658 | case CPU_LOONGSON3: | |
1659 | raw_event.cntr_mask = raw_id > 127 ? CNTR_ODD : CNTR_EVEN; | |
1660 | break; | |
3a9ab99e DCZ |
1661 | } |
1662 | ||
67dca667 JH |
1663 | raw_event.event_id = base_id; |
1664 | ||
3a9ab99e DCZ |
1665 | return &raw_event; |
1666 | } | |
1667 | ||
939991cf DD |
1668 | static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config) |
1669 | { | |
1670 | unsigned int raw_id = config & 0xff; | |
1671 | unsigned int base_id = raw_id & 0x7f; | |
1672 | ||
1673 | ||
1674 | raw_event.cntr_mask = CNTR_ALL; | |
1675 | raw_event.event_id = base_id; | |
1676 | ||
1677 | if (current_cpu_type() == CPU_CAVIUM_OCTEON2) { | |
1678 | if (base_id > 0x42) | |
1679 | return ERR_PTR(-EOPNOTSUPP); | |
1680 | } else { | |
1681 | if (base_id > 0x3a) | |
1682 | return ERR_PTR(-EOPNOTSUPP); | |
1683 | } | |
1684 | ||
1685 | switch (base_id) { | |
1686 | case 0x00: | |
1687 | case 0x0f: | |
1688 | case 0x1e: | |
1689 | case 0x1f: | |
1690 | case 0x2f: | |
1691 | case 0x34: | |
1692 | case 0x3b ... 0x3f: | |
1693 | return ERR_PTR(-EOPNOTSUPP); | |
1694 | default: | |
1695 | break; | |
1696 | } | |
1697 | ||
1698 | return &raw_event; | |
1699 | } | |
1700 | ||
4be3d2f3 ZSL |
1701 | static const struct mips_perf_event *xlp_pmu_map_raw_event(u64 config) |
1702 | { | |
1703 | unsigned int raw_id = config & 0xff; | |
1704 | ||
1705 | /* Only 1-63 are defined */ | |
1706 | if ((raw_id < 0x01) || (raw_id > 0x3f)) | |
1707 | return ERR_PTR(-EOPNOTSUPP); | |
1708 | ||
1709 | raw_event.cntr_mask = CNTR_ALL; | |
1710 | raw_event.event_id = raw_id; | |
1711 | ||
1712 | return &raw_event; | |
1713 | } | |
1714 | ||
3a9ab99e DCZ |
1715 | static int __init |
1716 | init_hw_perf_events(void) | |
1717 | { | |
1718 | int counters, irq; | |
82091564 | 1719 | int counter_bits; |
3a9ab99e DCZ |
1720 | |
1721 | pr_info("Performance counters: "); | |
1722 | ||
1723 | counters = n_counters(); | |
1724 | if (counters == 0) { | |
1725 | pr_cont("No available PMU.\n"); | |
1726 | return -ENODEV; | |
1727 | } | |
1728 | ||
399aaa25 | 1729 | #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS |
3a9ab99e DCZ |
1730 | cpu_has_mipsmt_pertccounters = read_c0_config7() & (1<<19); |
1731 | if (!cpu_has_mipsmt_pertccounters) | |
1732 | counters = counters_total_to_per_cpu(counters); | |
1733 | #endif | |
1734 | ||
a669efc4 AB |
1735 | if (get_c0_perfcount_int) |
1736 | irq = get_c0_perfcount_int(); | |
7eca5b14 | 1737 | else if (cp0_perfcount_irq >= 0) |
a669efc4 AB |
1738 | irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq; |
1739 | else | |
1740 | irq = -1; | |
3a9ab99e | 1741 | |
82091564 | 1742 | mipspmu.map_raw_event = mipsxx_pmu_map_raw_event; |
3a9ab99e DCZ |
1743 | |
1744 | switch (current_cpu_type()) { | |
1745 | case CPU_24K: | |
82091564 DD |
1746 | mipspmu.name = "mips/24K"; |
1747 | mipspmu.general_event_map = &mipsxxcore_event_map; | |
1748 | mipspmu.cache_event_map = &mipsxxcore_cache_map; | |
3a9ab99e DCZ |
1749 | break; |
1750 | case CPU_34K: | |
82091564 DD |
1751 | mipspmu.name = "mips/34K"; |
1752 | mipspmu.general_event_map = &mipsxxcore_event_map; | |
1753 | mipspmu.cache_event_map = &mipsxxcore_cache_map; | |
3a9ab99e DCZ |
1754 | break; |
1755 | case CPU_74K: | |
82091564 | 1756 | mipspmu.name = "mips/74K"; |
6b0b8429 DCZ |
1757 | mipspmu.general_event_map = &mipsxxcore_event_map2; |
1758 | mipspmu.cache_event_map = &mipsxxcore_cache_map2; | |
3a9ab99e | 1759 | break; |
c52068bd DCZ |
1760 | case CPU_PROAPTIV: |
1761 | mipspmu.name = "mips/proAptiv"; | |
1762 | mipspmu.general_event_map = &mipsxxcore_event_map2; | |
1763 | mipspmu.cache_event_map = &mipsxxcore_cache_map2; | |
1764 | break; | |
560b461b JH |
1765 | case CPU_P5600: |
1766 | mipspmu.name = "mips/P5600"; | |
1767 | mipspmu.general_event_map = &mipsxxcore_event_map2; | |
1768 | mipspmu.cache_event_map = &mipsxxcore_cache_map2; | |
1769 | break; | |
1091bfa2 PB |
1770 | case CPU_P6600: |
1771 | mipspmu.name = "mips/P6600"; | |
1772 | mipspmu.general_event_map = &mipsxxcore_event_map2; | |
1773 | mipspmu.cache_event_map = &mipsxxcore_cache_map2; | |
1774 | break; | |
4e88a862 MC |
1775 | case CPU_I6400: |
1776 | mipspmu.name = "mips/I6400"; | |
fd716fca JH |
1777 | mipspmu.general_event_map = &i6400_event_map; |
1778 | mipspmu.cache_event_map = &i6400_cache_map; | |
4e88a862 | 1779 | break; |
3a9ab99e | 1780 | case CPU_1004K: |
82091564 DD |
1781 | mipspmu.name = "mips/1004K"; |
1782 | mipspmu.general_event_map = &mipsxxcore_event_map; | |
1783 | mipspmu.cache_event_map = &mipsxxcore_cache_map; | |
3a9ab99e | 1784 | break; |
442e14a2 SH |
1785 | case CPU_1074K: |
1786 | mipspmu.name = "mips/1074K"; | |
1787 | mipspmu.general_event_map = &mipsxxcore_event_map; | |
1788 | mipspmu.cache_event_map = &mipsxxcore_cache_map; | |
1789 | break; | |
9597e432 DCZ |
1790 | case CPU_INTERAPTIV: |
1791 | mipspmu.name = "mips/interAptiv"; | |
1792 | mipspmu.general_event_map = &mipsxxcore_event_map; | |
1793 | mipspmu.cache_event_map = &mipsxxcore_cache_map; | |
1794 | break; | |
2fa36399 KC |
1795 | case CPU_LOONGSON1: |
1796 | mipspmu.name = "mips/loongson1"; | |
1797 | mipspmu.general_event_map = &mipsxxcore_event_map; | |
1798 | mipspmu.cache_event_map = &mipsxxcore_cache_map; | |
1799 | break; | |
f14ceff7 HC |
1800 | case CPU_LOONGSON3: |
1801 | mipspmu.name = "mips/loongson3"; | |
1802 | mipspmu.general_event_map = &loongson3_event_map; | |
1803 | mipspmu.cache_event_map = &loongson3_cache_map; | |
1804 | break; | |
939991cf DD |
1805 | case CPU_CAVIUM_OCTEON: |
1806 | case CPU_CAVIUM_OCTEON_PLUS: | |
1807 | case CPU_CAVIUM_OCTEON2: | |
1808 | mipspmu.name = "octeon"; | |
1809 | mipspmu.general_event_map = &octeon_event_map; | |
1810 | mipspmu.cache_event_map = &octeon_cache_map; | |
1811 | mipspmu.map_raw_event = octeon_pmu_map_raw_event; | |
1812 | break; | |
a7911a8f AC |
1813 | case CPU_BMIPS5000: |
1814 | mipspmu.name = "BMIPS5000"; | |
1815 | mipspmu.general_event_map = &bmips5000_event_map; | |
1816 | mipspmu.cache_event_map = &bmips5000_cache_map; | |
1817 | break; | |
4be3d2f3 ZSL |
1818 | case CPU_XLP: |
1819 | mipspmu.name = "xlp"; | |
1820 | mipspmu.general_event_map = &xlp_event_map; | |
1821 | mipspmu.cache_event_map = &xlp_cache_map; | |
1822 | mipspmu.map_raw_event = xlp_pmu_map_raw_event; | |
1823 | break; | |
3a9ab99e DCZ |
1824 | default: |
1825 | pr_cont("Either hardware does not support performance " | |
1826 | "counters, or not yet implemented.\n"); | |
1827 | return -ENODEV; | |
1828 | } | |
1829 | ||
82091564 DD |
1830 | mipspmu.num_counters = counters; |
1831 | mipspmu.irq = irq; | |
1832 | ||
1833 | if (read_c0_perfctrl0() & M_PERFCTL_WIDE) { | |
1834 | mipspmu.max_period = (1ULL << 63) - 1; | |
1835 | mipspmu.valid_count = (1ULL << 63) - 1; | |
1836 | mipspmu.overflow = 1ULL << 63; | |
1837 | mipspmu.read_counter = mipsxx_pmu_read_counter_64; | |
1838 | mipspmu.write_counter = mipsxx_pmu_write_counter_64; | |
1839 | counter_bits = 64; | |
1840 | } else { | |
1841 | mipspmu.max_period = (1ULL << 31) - 1; | |
1842 | mipspmu.valid_count = (1ULL << 31) - 1; | |
1843 | mipspmu.overflow = 1ULL << 31; | |
1844 | mipspmu.read_counter = mipsxx_pmu_read_counter; | |
1845 | mipspmu.write_counter = mipsxx_pmu_write_counter; | |
1846 | counter_bits = 32; | |
1847 | } | |
1848 | ||
1849 | on_each_cpu(reset_counters, (void *)(long)counters, 1); | |
1850 | ||
1851 | pr_cont("%s PMU enabled, %d %d-bit counters available to each " | |
1852 | "CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq, | |
1853 | irq < 0 ? " (share with timer interrupt)" : ""); | |
3a9ab99e | 1854 | |
404ff638 DCZ |
1855 | perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW); |
1856 | ||
3a9ab99e DCZ |
1857 | return 0; |
1858 | } | |
004417a6 | 1859 | early_initcall(init_hw_perf_events); |