]> Git Repo - J-linux.git/blob - drivers/gpu/drm/i915/i915_pmu.c
Merge tag 'nfsd-6.7-1' of git://git.kernel.org/pub/scm/linux/kernel/git/cel/linux
[J-linux.git] / drivers / gpu / drm / i915 / i915_pmu.c
1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2017-2018 Intel Corporation
5  */
6
7 #include <linux/pm_runtime.h>
8
9 #include "gt/intel_engine.h"
10 #include "gt/intel_engine_pm.h"
11 #include "gt/intel_engine_regs.h"
12 #include "gt/intel_engine_user.h"
13 #include "gt/intel_gt.h"
14 #include "gt/intel_gt_pm.h"
15 #include "gt/intel_gt_regs.h"
16 #include "gt/intel_rc6.h"
17 #include "gt/intel_rps.h"
18
19 #include "i915_drv.h"
20 #include "i915_pmu.h"
21
22 /* Frequency for the sampling timer for events which need it. */
23 #define FREQUENCY 200
24 #define PERIOD max_t(u64, 10000, NSEC_PER_SEC / FREQUENCY)
25
26 #define ENGINE_SAMPLE_MASK \
27         (BIT(I915_SAMPLE_BUSY) | \
28          BIT(I915_SAMPLE_WAIT) | \
29          BIT(I915_SAMPLE_SEMA))
30
31 static cpumask_t i915_pmu_cpumask;
32 static unsigned int i915_pmu_target_cpu = -1;
33
34 static u8 engine_config_sample(u64 config)
35 {
36         return config & I915_PMU_SAMPLE_MASK;
37 }
38
39 static u8 engine_event_sample(struct perf_event *event)
40 {
41         return engine_config_sample(event->attr.config);
42 }
43
44 static u8 engine_event_class(struct perf_event *event)
45 {
46         return (event->attr.config >> I915_PMU_CLASS_SHIFT) & 0xff;
47 }
48
49 static u8 engine_event_instance(struct perf_event *event)
50 {
51         return (event->attr.config >> I915_PMU_SAMPLE_BITS) & 0xff;
52 }
53
54 static bool is_engine_config(const u64 config)
55 {
56         return config < __I915_PMU_OTHER(0);
57 }
58
59 static unsigned int config_gt_id(const u64 config)
60 {
61         return config >> __I915_PMU_GT_SHIFT;
62 }
63
64 static u64 config_counter(const u64 config)
65 {
66         return config & ~(~0ULL << __I915_PMU_GT_SHIFT);
67 }
68
69 static unsigned int other_bit(const u64 config)
70 {
71         unsigned int val;
72
73         switch (config_counter(config)) {
74         case I915_PMU_ACTUAL_FREQUENCY:
75                 val =  __I915_PMU_ACTUAL_FREQUENCY_ENABLED;
76                 break;
77         case I915_PMU_REQUESTED_FREQUENCY:
78                 val = __I915_PMU_REQUESTED_FREQUENCY_ENABLED;
79                 break;
80         case I915_PMU_RC6_RESIDENCY:
81                 val = __I915_PMU_RC6_RESIDENCY_ENABLED;
82                 break;
83         default:
84                 /*
85                  * Events that do not require sampling, or tracking state
86                  * transitions between enabled and disabled can be ignored.
87                  */
88                 return -1;
89         }
90
91         return I915_ENGINE_SAMPLE_COUNT +
92                config_gt_id(config) * __I915_PMU_TRACKED_EVENT_COUNT +
93                val;
94 }
95
96 static unsigned int config_bit(const u64 config)
97 {
98         if (is_engine_config(config))
99                 return engine_config_sample(config);
100         else
101                 return other_bit(config);
102 }
103
104 static u32 config_mask(const u64 config)
105 {
106         unsigned int bit = config_bit(config);
107
108         if (__builtin_constant_p(config))
109                 BUILD_BUG_ON(bit >
110                              BITS_PER_TYPE(typeof_member(struct i915_pmu,
111                                                          enable)) - 1);
112         else
113                 WARN_ON_ONCE(bit >
114                              BITS_PER_TYPE(typeof_member(struct i915_pmu,
115                                                          enable)) - 1);
116
117         return BIT(config_bit(config));
118 }
119
120 static bool is_engine_event(struct perf_event *event)
121 {
122         return is_engine_config(event->attr.config);
123 }
124
125 static unsigned int event_bit(struct perf_event *event)
126 {
127         return config_bit(event->attr.config);
128 }
129
130 static u32 frequency_enabled_mask(void)
131 {
132         unsigned int i;
133         u32 mask = 0;
134
135         for (i = 0; i < I915_PMU_MAX_GT; i++)
136                 mask |= config_mask(__I915_PMU_ACTUAL_FREQUENCY(i)) |
137                         config_mask(__I915_PMU_REQUESTED_FREQUENCY(i));
138
139         return mask;
140 }
141
142 static bool pmu_needs_timer(struct i915_pmu *pmu)
143 {
144         struct drm_i915_private *i915 = container_of(pmu, typeof(*i915), pmu);
145         u32 enable;
146
147         /*
148          * Only some counters need the sampling timer.
149          *
150          * We start with a bitmask of all currently enabled events.
151          */
152         enable = pmu->enable;
153
154         /*
155          * Mask out all the ones which do not need the timer, or in
156          * other words keep all the ones that could need the timer.
157          */
158         enable &= frequency_enabled_mask() | ENGINE_SAMPLE_MASK;
159
160         /*
161          * Also there is software busyness tracking available we do not
162          * need the timer for I915_SAMPLE_BUSY counter.
163          */
164         if (i915->caps.scheduler & I915_SCHEDULER_CAP_ENGINE_BUSY_STATS)
165                 enable &= ~BIT(I915_SAMPLE_BUSY);
166
167         /*
168          * If some bits remain it means we need the sampling timer running.
169          */
170         return enable;
171 }
172
173 static u64 __get_rc6(struct intel_gt *gt)
174 {
175         struct drm_i915_private *i915 = gt->i915;
176         u64 val;
177
178         val = intel_rc6_residency_ns(&gt->rc6, INTEL_RC6_RES_RC6);
179
180         if (HAS_RC6p(i915))
181                 val += intel_rc6_residency_ns(&gt->rc6, INTEL_RC6_RES_RC6p);
182
183         if (HAS_RC6pp(i915))
184                 val += intel_rc6_residency_ns(&gt->rc6, INTEL_RC6_RES_RC6pp);
185
186         return val;
187 }
188
189 static inline s64 ktime_since_raw(const ktime_t kt)
190 {
191         return ktime_to_ns(ktime_sub(ktime_get_raw(), kt));
192 }
193
194 static u64 read_sample(struct i915_pmu *pmu, unsigned int gt_id, int sample)
195 {
196         return pmu->sample[gt_id][sample].cur;
197 }
198
199 static void
200 store_sample(struct i915_pmu *pmu, unsigned int gt_id, int sample, u64 val)
201 {
202         pmu->sample[gt_id][sample].cur = val;
203 }
204
205 static void
206 add_sample_mult(struct i915_pmu *pmu, unsigned int gt_id, int sample, u32 val, u32 mul)
207 {
208         pmu->sample[gt_id][sample].cur += mul_u32_u32(val, mul);
209 }
210
211 static u64 get_rc6(struct intel_gt *gt)
212 {
213         struct drm_i915_private *i915 = gt->i915;
214         const unsigned int gt_id = gt->info.id;
215         struct i915_pmu *pmu = &i915->pmu;
216         unsigned long flags;
217         bool awake = false;
218         u64 val;
219
220         if (intel_gt_pm_get_if_awake(gt)) {
221                 val = __get_rc6(gt);
222                 intel_gt_pm_put_async(gt);
223                 awake = true;
224         }
225
226         spin_lock_irqsave(&pmu->lock, flags);
227
228         if (awake) {
229                 store_sample(pmu, gt_id, __I915_SAMPLE_RC6, val);
230         } else {
231                 /*
232                  * We think we are runtime suspended.
233                  *
234                  * Report the delta from when the device was suspended to now,
235                  * on top of the last known real value, as the approximated RC6
236                  * counter value.
237                  */
238                 val = ktime_since_raw(pmu->sleep_last[gt_id]);
239                 val += read_sample(pmu, gt_id, __I915_SAMPLE_RC6);
240         }
241
242         if (val < read_sample(pmu, gt_id, __I915_SAMPLE_RC6_LAST_REPORTED))
243                 val = read_sample(pmu, gt_id, __I915_SAMPLE_RC6_LAST_REPORTED);
244         else
245                 store_sample(pmu, gt_id, __I915_SAMPLE_RC6_LAST_REPORTED, val);
246
247         spin_unlock_irqrestore(&pmu->lock, flags);
248
249         return val;
250 }
251
252 static void init_rc6(struct i915_pmu *pmu)
253 {
254         struct drm_i915_private *i915 = container_of(pmu, typeof(*i915), pmu);
255         struct intel_gt *gt;
256         unsigned int i;
257
258         for_each_gt(gt, i915, i) {
259                 intel_wakeref_t wakeref;
260
261                 with_intel_runtime_pm(gt->uncore->rpm, wakeref) {
262                         u64 val = __get_rc6(gt);
263
264                         store_sample(pmu, i, __I915_SAMPLE_RC6, val);
265                         store_sample(pmu, i, __I915_SAMPLE_RC6_LAST_REPORTED,
266                                      val);
267                         pmu->sleep_last[i] = ktime_get_raw();
268                 }
269         }
270 }
271
272 static void park_rc6(struct intel_gt *gt)
273 {
274         struct i915_pmu *pmu = &gt->i915->pmu;
275
276         store_sample(pmu, gt->info.id, __I915_SAMPLE_RC6, __get_rc6(gt));
277         pmu->sleep_last[gt->info.id] = ktime_get_raw();
278 }
279
280 static void __i915_pmu_maybe_start_timer(struct i915_pmu *pmu)
281 {
282         if (!pmu->timer_enabled && pmu_needs_timer(pmu)) {
283                 pmu->timer_enabled = true;
284                 pmu->timer_last = ktime_get();
285                 hrtimer_start_range_ns(&pmu->timer,
286                                        ns_to_ktime(PERIOD), 0,
287                                        HRTIMER_MODE_REL_PINNED);
288         }
289 }
290
291 void i915_pmu_gt_parked(struct intel_gt *gt)
292 {
293         struct i915_pmu *pmu = &gt->i915->pmu;
294
295         if (!pmu->base.event_init)
296                 return;
297
298         spin_lock_irq(&pmu->lock);
299
300         park_rc6(gt);
301
302         /*
303          * Signal sampling timer to stop if only engine events are enabled and
304          * GPU went idle.
305          */
306         pmu->unparked &= ~BIT(gt->info.id);
307         if (pmu->unparked == 0)
308                 pmu->timer_enabled = false;
309
310         spin_unlock_irq(&pmu->lock);
311 }
312
313 void i915_pmu_gt_unparked(struct intel_gt *gt)
314 {
315         struct i915_pmu *pmu = &gt->i915->pmu;
316
317         if (!pmu->base.event_init)
318                 return;
319
320         spin_lock_irq(&pmu->lock);
321
322         /*
323          * Re-enable sampling timer when GPU goes active.
324          */
325         if (pmu->unparked == 0)
326                 __i915_pmu_maybe_start_timer(pmu);
327
328         pmu->unparked |= BIT(gt->info.id);
329
330         spin_unlock_irq(&pmu->lock);
331 }
332
333 static void
334 add_sample(struct i915_pmu_sample *sample, u32 val)
335 {
336         sample->cur += val;
337 }
338
339 static bool exclusive_mmio_access(const struct drm_i915_private *i915)
340 {
341         /*
342          * We have to avoid concurrent mmio cache line access on gen7 or
343          * risk a machine hang. For a fun history lesson dig out the old
344          * userspace intel_gpu_top and run it on Ivybridge or Haswell!
345          */
346         return GRAPHICS_VER(i915) == 7;
347 }
348
349 static void engine_sample(struct intel_engine_cs *engine, unsigned int period_ns)
350 {
351         struct intel_engine_pmu *pmu = &engine->pmu;
352         bool busy;
353         u32 val;
354
355         val = ENGINE_READ_FW(engine, RING_CTL);
356         if (val == 0) /* powerwell off => engine idle */
357                 return;
358
359         if (val & RING_WAIT)
360                 add_sample(&pmu->sample[I915_SAMPLE_WAIT], period_ns);
361         if (val & RING_WAIT_SEMAPHORE)
362                 add_sample(&pmu->sample[I915_SAMPLE_SEMA], period_ns);
363
364         /* No need to sample when busy stats are supported. */
365         if (intel_engine_supports_stats(engine))
366                 return;
367
368         /*
369          * While waiting on a semaphore or event, MI_MODE reports the
370          * ring as idle. However, previously using the seqno, and with
371          * execlists sampling, we account for the ring waiting as the
372          * engine being busy. Therefore, we record the sample as being
373          * busy if either waiting or !idle.
374          */
375         busy = val & (RING_WAIT_SEMAPHORE | RING_WAIT);
376         if (!busy) {
377                 val = ENGINE_READ_FW(engine, RING_MI_MODE);
378                 busy = !(val & MODE_IDLE);
379         }
380         if (busy)
381                 add_sample(&pmu->sample[I915_SAMPLE_BUSY], period_ns);
382 }
383
384 static void
385 engines_sample(struct intel_gt *gt, unsigned int period_ns)
386 {
387         struct drm_i915_private *i915 = gt->i915;
388         struct intel_engine_cs *engine;
389         enum intel_engine_id id;
390         unsigned long flags;
391
392         if ((i915->pmu.enable & ENGINE_SAMPLE_MASK) == 0)
393                 return;
394
395         if (!intel_gt_pm_is_awake(gt))
396                 return;
397
398         for_each_engine(engine, gt, id) {
399                 if (!engine->pmu.enable)
400                         continue;
401
402                 if (!intel_engine_pm_get_if_awake(engine))
403                         continue;
404
405                 if (exclusive_mmio_access(i915)) {
406                         spin_lock_irqsave(&engine->uncore->lock, flags);
407                         engine_sample(engine, period_ns);
408                         spin_unlock_irqrestore(&engine->uncore->lock, flags);
409                 } else {
410                         engine_sample(engine, period_ns);
411                 }
412
413                 intel_engine_pm_put_async(engine);
414         }
415 }
416
417 static bool
418 frequency_sampling_enabled(struct i915_pmu *pmu, unsigned int gt)
419 {
420         return pmu->enable &
421                (config_mask(__I915_PMU_ACTUAL_FREQUENCY(gt)) |
422                 config_mask(__I915_PMU_REQUESTED_FREQUENCY(gt)));
423 }
424
425 static void
426 frequency_sample(struct intel_gt *gt, unsigned int period_ns)
427 {
428         struct drm_i915_private *i915 = gt->i915;
429         const unsigned int gt_id = gt->info.id;
430         struct i915_pmu *pmu = &i915->pmu;
431         struct intel_rps *rps = &gt->rps;
432
433         if (!frequency_sampling_enabled(pmu, gt_id))
434                 return;
435
436         /* Report 0/0 (actual/requested) frequency while parked. */
437         if (!intel_gt_pm_get_if_awake(gt))
438                 return;
439
440         if (pmu->enable & config_mask(__I915_PMU_ACTUAL_FREQUENCY(gt_id))) {
441                 u32 val;
442
443                 /*
444                  * We take a quick peek here without using forcewake
445                  * so that we don't perturb the system under observation
446                  * (forcewake => !rc6 => increased power use). We expect
447                  * that if the read fails because it is outside of the
448                  * mmio power well, then it will return 0 -- in which
449                  * case we assume the system is running at the intended
450                  * frequency. Fortunately, the read should rarely fail!
451                  */
452                 val = intel_rps_read_actual_frequency_fw(rps);
453                 if (!val)
454                         val = intel_gpu_freq(rps, rps->cur_freq);
455
456                 add_sample_mult(pmu, gt_id, __I915_SAMPLE_FREQ_ACT,
457                                 val, period_ns / 1000);
458         }
459
460         if (pmu->enable & config_mask(__I915_PMU_REQUESTED_FREQUENCY(gt_id))) {
461                 add_sample_mult(pmu, gt_id, __I915_SAMPLE_FREQ_REQ,
462                                 intel_rps_get_requested_frequency(rps),
463                                 period_ns / 1000);
464         }
465
466         intel_gt_pm_put_async(gt);
467 }
468
469 static enum hrtimer_restart i915_sample(struct hrtimer *hrtimer)
470 {
471         struct drm_i915_private *i915 =
472                 container_of(hrtimer, struct drm_i915_private, pmu.timer);
473         struct i915_pmu *pmu = &i915->pmu;
474         unsigned int period_ns;
475         struct intel_gt *gt;
476         unsigned int i;
477         ktime_t now;
478
479         if (!READ_ONCE(pmu->timer_enabled))
480                 return HRTIMER_NORESTART;
481
482         now = ktime_get();
483         period_ns = ktime_to_ns(ktime_sub(now, pmu->timer_last));
484         pmu->timer_last = now;
485
486         /*
487          * Strictly speaking the passed in period may not be 100% accurate for
488          * all internal calculation, since some amount of time can be spent on
489          * grabbing the forcewake. However the potential error from timer call-
490          * back delay greatly dominates this so we keep it simple.
491          */
492
493         for_each_gt(gt, i915, i) {
494                 if (!(pmu->unparked & BIT(i)))
495                         continue;
496
497                 engines_sample(gt, period_ns);
498                 frequency_sample(gt, period_ns);
499         }
500
501         hrtimer_forward(hrtimer, now, ns_to_ktime(PERIOD));
502
503         return HRTIMER_RESTART;
504 }
505
506 static void i915_pmu_event_destroy(struct perf_event *event)
507 {
508         struct drm_i915_private *i915 =
509                 container_of(event->pmu, typeof(*i915), pmu.base);
510
511         drm_WARN_ON(&i915->drm, event->parent);
512
513         drm_dev_put(&i915->drm);
514 }
515
516 static int
517 engine_event_status(struct intel_engine_cs *engine,
518                     enum drm_i915_pmu_engine_sample sample)
519 {
520         switch (sample) {
521         case I915_SAMPLE_BUSY:
522         case I915_SAMPLE_WAIT:
523                 break;
524         case I915_SAMPLE_SEMA:
525                 if (GRAPHICS_VER(engine->i915) < 6)
526                         return -ENODEV;
527                 break;
528         default:
529                 return -ENOENT;
530         }
531
532         return 0;
533 }
534
535 static int
536 config_status(struct drm_i915_private *i915, u64 config)
537 {
538         struct intel_gt *gt = to_gt(i915);
539
540         unsigned int gt_id = config_gt_id(config);
541         unsigned int max_gt_id = HAS_EXTRA_GT_LIST(i915) ? 1 : 0;
542
543         if (gt_id > max_gt_id)
544                 return -ENOENT;
545
546         switch (config_counter(config)) {
547         case I915_PMU_ACTUAL_FREQUENCY:
548                 if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915))
549                         /* Requires a mutex for sampling! */
550                         return -ENODEV;
551                 fallthrough;
552         case I915_PMU_REQUESTED_FREQUENCY:
553                 if (GRAPHICS_VER(i915) < 6)
554                         return -ENODEV;
555                 break;
556         case I915_PMU_INTERRUPTS:
557                 if (gt_id)
558                         return -ENOENT;
559                 break;
560         case I915_PMU_RC6_RESIDENCY:
561                 if (!gt->rc6.supported)
562                         return -ENODEV;
563                 break;
564         case I915_PMU_SOFTWARE_GT_AWAKE_TIME:
565                 break;
566         default:
567                 return -ENOENT;
568         }
569
570         return 0;
571 }
572
573 static int engine_event_init(struct perf_event *event)
574 {
575         struct drm_i915_private *i915 =
576                 container_of(event->pmu, typeof(*i915), pmu.base);
577         struct intel_engine_cs *engine;
578
579         engine = intel_engine_lookup_user(i915, engine_event_class(event),
580                                           engine_event_instance(event));
581         if (!engine)
582                 return -ENODEV;
583
584         return engine_event_status(engine, engine_event_sample(event));
585 }
586
587 static int i915_pmu_event_init(struct perf_event *event)
588 {
589         struct drm_i915_private *i915 =
590                 container_of(event->pmu, typeof(*i915), pmu.base);
591         struct i915_pmu *pmu = &i915->pmu;
592         int ret;
593
594         if (pmu->closed)
595                 return -ENODEV;
596
597         if (event->attr.type != event->pmu->type)
598                 return -ENOENT;
599
600         /* unsupported modes and filters */
601         if (event->attr.sample_period) /* no sampling */
602                 return -EINVAL;
603
604         if (has_branch_stack(event))
605                 return -EOPNOTSUPP;
606
607         if (event->cpu < 0)
608                 return -EINVAL;
609
610         /* only allow running on one cpu at a time */
611         if (!cpumask_test_cpu(event->cpu, &i915_pmu_cpumask))
612                 return -EINVAL;
613
614         if (is_engine_event(event))
615                 ret = engine_event_init(event);
616         else
617                 ret = config_status(i915, event->attr.config);
618         if (ret)
619                 return ret;
620
621         if (!event->parent) {
622                 drm_dev_get(&i915->drm);
623                 event->destroy = i915_pmu_event_destroy;
624         }
625
626         return 0;
627 }
628
629 static u64 __i915_pmu_event_read(struct perf_event *event)
630 {
631         struct drm_i915_private *i915 =
632                 container_of(event->pmu, typeof(*i915), pmu.base);
633         struct i915_pmu *pmu = &i915->pmu;
634         u64 val = 0;
635
636         if (is_engine_event(event)) {
637                 u8 sample = engine_event_sample(event);
638                 struct intel_engine_cs *engine;
639
640                 engine = intel_engine_lookup_user(i915,
641                                                   engine_event_class(event),
642                                                   engine_event_instance(event));
643
644                 if (drm_WARN_ON_ONCE(&i915->drm, !engine)) {
645                         /* Do nothing */
646                 } else if (sample == I915_SAMPLE_BUSY &&
647                            intel_engine_supports_stats(engine)) {
648                         ktime_t unused;
649
650                         val = ktime_to_ns(intel_engine_get_busy_time(engine,
651                                                                      &unused));
652                 } else {
653                         val = engine->pmu.sample[sample].cur;
654                 }
655         } else {
656                 const unsigned int gt_id = config_gt_id(event->attr.config);
657                 const u64 config = config_counter(event->attr.config);
658
659                 switch (config) {
660                 case I915_PMU_ACTUAL_FREQUENCY:
661                         val =
662                            div_u64(read_sample(pmu, gt_id,
663                                                __I915_SAMPLE_FREQ_ACT),
664                                    USEC_PER_SEC /* to MHz */);
665                         break;
666                 case I915_PMU_REQUESTED_FREQUENCY:
667                         val =
668                            div_u64(read_sample(pmu, gt_id,
669                                                __I915_SAMPLE_FREQ_REQ),
670                                    USEC_PER_SEC /* to MHz */);
671                         break;
672                 case I915_PMU_INTERRUPTS:
673                         val = READ_ONCE(pmu->irq_count);
674                         break;
675                 case I915_PMU_RC6_RESIDENCY:
676                         val = get_rc6(i915->gt[gt_id]);
677                         break;
678                 case I915_PMU_SOFTWARE_GT_AWAKE_TIME:
679                         val = ktime_to_ns(intel_gt_get_awake_time(to_gt(i915)));
680                         break;
681                 }
682         }
683
684         return val;
685 }
686
687 static void i915_pmu_event_read(struct perf_event *event)
688 {
689         struct drm_i915_private *i915 =
690                 container_of(event->pmu, typeof(*i915), pmu.base);
691         struct hw_perf_event *hwc = &event->hw;
692         struct i915_pmu *pmu = &i915->pmu;
693         u64 prev, new;
694
695         if (pmu->closed) {
696                 event->hw.state = PERF_HES_STOPPED;
697                 return;
698         }
699
700         prev = local64_read(&hwc->prev_count);
701         do {
702                 new = __i915_pmu_event_read(event);
703         } while (!local64_try_cmpxchg(&hwc->prev_count, &prev, new));
704
705         local64_add(new - prev, &event->count);
706 }
707
708 static void i915_pmu_enable(struct perf_event *event)
709 {
710         struct drm_i915_private *i915 =
711                 container_of(event->pmu, typeof(*i915), pmu.base);
712         const unsigned int bit = event_bit(event);
713         struct i915_pmu *pmu = &i915->pmu;
714         unsigned long flags;
715
716         if (bit == -1)
717                 goto update;
718
719         spin_lock_irqsave(&pmu->lock, flags);
720
721         /*
722          * Update the bitmask of enabled events and increment
723          * the event reference counter.
724          */
725         BUILD_BUG_ON(ARRAY_SIZE(pmu->enable_count) != I915_PMU_MASK_BITS);
726         GEM_BUG_ON(bit >= ARRAY_SIZE(pmu->enable_count));
727         GEM_BUG_ON(pmu->enable_count[bit] == ~0);
728
729         pmu->enable |= BIT(bit);
730         pmu->enable_count[bit]++;
731
732         /*
733          * Start the sampling timer if needed and not already enabled.
734          */
735         __i915_pmu_maybe_start_timer(pmu);
736
737         /*
738          * For per-engine events the bitmask and reference counting
739          * is stored per engine.
740          */
741         if (is_engine_event(event)) {
742                 u8 sample = engine_event_sample(event);
743                 struct intel_engine_cs *engine;
744
745                 engine = intel_engine_lookup_user(i915,
746                                                   engine_event_class(event),
747                                                   engine_event_instance(event));
748
749                 BUILD_BUG_ON(ARRAY_SIZE(engine->pmu.enable_count) !=
750                              I915_ENGINE_SAMPLE_COUNT);
751                 BUILD_BUG_ON(ARRAY_SIZE(engine->pmu.sample) !=
752                              I915_ENGINE_SAMPLE_COUNT);
753                 GEM_BUG_ON(sample >= ARRAY_SIZE(engine->pmu.enable_count));
754                 GEM_BUG_ON(sample >= ARRAY_SIZE(engine->pmu.sample));
755                 GEM_BUG_ON(engine->pmu.enable_count[sample] == ~0);
756
757                 engine->pmu.enable |= BIT(sample);
758                 engine->pmu.enable_count[sample]++;
759         }
760
761         spin_unlock_irqrestore(&pmu->lock, flags);
762
763 update:
764         /*
765          * Store the current counter value so we can report the correct delta
766          * for all listeners. Even when the event was already enabled and has
767          * an existing non-zero value.
768          */
769         local64_set(&event->hw.prev_count, __i915_pmu_event_read(event));
770 }
771
772 static void i915_pmu_disable(struct perf_event *event)
773 {
774         struct drm_i915_private *i915 =
775                 container_of(event->pmu, typeof(*i915), pmu.base);
776         const unsigned int bit = event_bit(event);
777         struct i915_pmu *pmu = &i915->pmu;
778         unsigned long flags;
779
780         if (bit == -1)
781                 return;
782
783         spin_lock_irqsave(&pmu->lock, flags);
784
785         if (is_engine_event(event)) {
786                 u8 sample = engine_event_sample(event);
787                 struct intel_engine_cs *engine;
788
789                 engine = intel_engine_lookup_user(i915,
790                                                   engine_event_class(event),
791                                                   engine_event_instance(event));
792
793                 GEM_BUG_ON(sample >= ARRAY_SIZE(engine->pmu.enable_count));
794                 GEM_BUG_ON(sample >= ARRAY_SIZE(engine->pmu.sample));
795                 GEM_BUG_ON(engine->pmu.enable_count[sample] == 0);
796
797                 /*
798                  * Decrement the reference count and clear the enabled
799                  * bitmask when the last listener on an event goes away.
800                  */
801                 if (--engine->pmu.enable_count[sample] == 0)
802                         engine->pmu.enable &= ~BIT(sample);
803         }
804
805         GEM_BUG_ON(bit >= ARRAY_SIZE(pmu->enable_count));
806         GEM_BUG_ON(pmu->enable_count[bit] == 0);
807         /*
808          * Decrement the reference count and clear the enabled
809          * bitmask when the last listener on an event goes away.
810          */
811         if (--pmu->enable_count[bit] == 0) {
812                 pmu->enable &= ~BIT(bit);
813                 pmu->timer_enabled &= pmu_needs_timer(pmu);
814         }
815
816         spin_unlock_irqrestore(&pmu->lock, flags);
817 }
818
819 static void i915_pmu_event_start(struct perf_event *event, int flags)
820 {
821         struct drm_i915_private *i915 =
822                 container_of(event->pmu, typeof(*i915), pmu.base);
823         struct i915_pmu *pmu = &i915->pmu;
824
825         if (pmu->closed)
826                 return;
827
828         i915_pmu_enable(event);
829         event->hw.state = 0;
830 }
831
832 static void i915_pmu_event_stop(struct perf_event *event, int flags)
833 {
834         struct drm_i915_private *i915 =
835                 container_of(event->pmu, typeof(*i915), pmu.base);
836         struct i915_pmu *pmu = &i915->pmu;
837
838         if (pmu->closed)
839                 goto out;
840
841         if (flags & PERF_EF_UPDATE)
842                 i915_pmu_event_read(event);
843         i915_pmu_disable(event);
844
845 out:
846         event->hw.state = PERF_HES_STOPPED;
847 }
848
849 static int i915_pmu_event_add(struct perf_event *event, int flags)
850 {
851         struct drm_i915_private *i915 =
852                 container_of(event->pmu, typeof(*i915), pmu.base);
853         struct i915_pmu *pmu = &i915->pmu;
854
855         if (pmu->closed)
856                 return -ENODEV;
857
858         if (flags & PERF_EF_START)
859                 i915_pmu_event_start(event, flags);
860
861         return 0;
862 }
863
864 static void i915_pmu_event_del(struct perf_event *event, int flags)
865 {
866         i915_pmu_event_stop(event, PERF_EF_UPDATE);
867 }
868
869 static int i915_pmu_event_event_idx(struct perf_event *event)
870 {
871         return 0;
872 }
873
874 struct i915_str_attribute {
875         struct device_attribute attr;
876         const char *str;
877 };
878
879 static ssize_t i915_pmu_format_show(struct device *dev,
880                                     struct device_attribute *attr, char *buf)
881 {
882         struct i915_str_attribute *eattr;
883
884         eattr = container_of(attr, struct i915_str_attribute, attr);
885         return sprintf(buf, "%s\n", eattr->str);
886 }
887
888 #define I915_PMU_FORMAT_ATTR(_name, _config) \
889         (&((struct i915_str_attribute[]) { \
890                 { .attr = __ATTR(_name, 0444, i915_pmu_format_show, NULL), \
891                   .str = _config, } \
892         })[0].attr.attr)
893
894 static struct attribute *i915_pmu_format_attrs[] = {
895         I915_PMU_FORMAT_ATTR(i915_eventid, "config:0-20"),
896         NULL,
897 };
898
899 static const struct attribute_group i915_pmu_format_attr_group = {
900         .name = "format",
901         .attrs = i915_pmu_format_attrs,
902 };
903
904 struct i915_ext_attribute {
905         struct device_attribute attr;
906         unsigned long val;
907 };
908
909 static ssize_t i915_pmu_event_show(struct device *dev,
910                                    struct device_attribute *attr, char *buf)
911 {
912         struct i915_ext_attribute *eattr;
913
914         eattr = container_of(attr, struct i915_ext_attribute, attr);
915         return sprintf(buf, "config=0x%lx\n", eattr->val);
916 }
917
918 static ssize_t cpumask_show(struct device *dev,
919                             struct device_attribute *attr, char *buf)
920 {
921         return cpumap_print_to_pagebuf(true, buf, &i915_pmu_cpumask);
922 }
923
924 static DEVICE_ATTR_RO(cpumask);
925
926 static struct attribute *i915_cpumask_attrs[] = {
927         &dev_attr_cpumask.attr,
928         NULL,
929 };
930
931 static const struct attribute_group i915_pmu_cpumask_attr_group = {
932         .attrs = i915_cpumask_attrs,
933 };
934
935 #define __event(__counter, __name, __unit) \
936 { \
937         .counter = (__counter), \
938         .name = (__name), \
939         .unit = (__unit), \
940         .global = false, \
941 }
942
943 #define __global_event(__counter, __name, __unit) \
944 { \
945         .counter = (__counter), \
946         .name = (__name), \
947         .unit = (__unit), \
948         .global = true, \
949 }
950
951 #define __engine_event(__sample, __name) \
952 { \
953         .sample = (__sample), \
954         .name = (__name), \
955 }
956
957 static struct i915_ext_attribute *
958 add_i915_attr(struct i915_ext_attribute *attr, const char *name, u64 config)
959 {
960         sysfs_attr_init(&attr->attr.attr);
961         attr->attr.attr.name = name;
962         attr->attr.attr.mode = 0444;
963         attr->attr.show = i915_pmu_event_show;
964         attr->val = config;
965
966         return ++attr;
967 }
968
969 static struct perf_pmu_events_attr *
970 add_pmu_attr(struct perf_pmu_events_attr *attr, const char *name,
971              const char *str)
972 {
973         sysfs_attr_init(&attr->attr.attr);
974         attr->attr.attr.name = name;
975         attr->attr.attr.mode = 0444;
976         attr->attr.show = perf_event_sysfs_show;
977         attr->event_str = str;
978
979         return ++attr;
980 }
981
982 static struct attribute **
983 create_event_attributes(struct i915_pmu *pmu)
984 {
985         struct drm_i915_private *i915 = container_of(pmu, typeof(*i915), pmu);
986         static const struct {
987                 unsigned int counter;
988                 const char *name;
989                 const char *unit;
990                 bool global;
991         } events[] = {
992                 __event(0, "actual-frequency", "M"),
993                 __event(1, "requested-frequency", "M"),
994                 __global_event(2, "interrupts", NULL),
995                 __event(3, "rc6-residency", "ns"),
996                 __event(4, "software-gt-awake-time", "ns"),
997         };
998         static const struct {
999                 enum drm_i915_pmu_engine_sample sample;
1000                 char *name;
1001         } engine_events[] = {
1002                 __engine_event(I915_SAMPLE_BUSY, "busy"),
1003                 __engine_event(I915_SAMPLE_SEMA, "sema"),
1004                 __engine_event(I915_SAMPLE_WAIT, "wait"),
1005         };
1006         unsigned int count = 0;
1007         struct perf_pmu_events_attr *pmu_attr = NULL, *pmu_iter;
1008         struct i915_ext_attribute *i915_attr = NULL, *i915_iter;
1009         struct attribute **attr = NULL, **attr_iter;
1010         struct intel_engine_cs *engine;
1011         struct intel_gt *gt;
1012         unsigned int i, j;
1013
1014         /* Count how many counters we will be exposing. */
1015         for_each_gt(gt, i915, j) {
1016                 for (i = 0; i < ARRAY_SIZE(events); i++) {
1017                         u64 config = ___I915_PMU_OTHER(j, events[i].counter);
1018
1019                         if (!config_status(i915, config))
1020                                 count++;
1021                 }
1022         }
1023
1024         for_each_uabi_engine(engine, i915) {
1025                 for (i = 0; i < ARRAY_SIZE(engine_events); i++) {
1026                         if (!engine_event_status(engine,
1027                                                  engine_events[i].sample))
1028                                 count++;
1029                 }
1030         }
1031
1032         /* Allocate attribute objects and table. */
1033         i915_attr = kcalloc(count, sizeof(*i915_attr), GFP_KERNEL);
1034         if (!i915_attr)
1035                 goto err_alloc;
1036
1037         pmu_attr = kcalloc(count, sizeof(*pmu_attr), GFP_KERNEL);
1038         if (!pmu_attr)
1039                 goto err_alloc;
1040
1041         /* Max one pointer of each attribute type plus a termination entry. */
1042         attr = kcalloc(count * 2 + 1, sizeof(*attr), GFP_KERNEL);
1043         if (!attr)
1044                 goto err_alloc;
1045
1046         i915_iter = i915_attr;
1047         pmu_iter = pmu_attr;
1048         attr_iter = attr;
1049
1050         /* Initialize supported non-engine counters. */
1051         for_each_gt(gt, i915, j) {
1052                 for (i = 0; i < ARRAY_SIZE(events); i++) {
1053                         u64 config = ___I915_PMU_OTHER(j, events[i].counter);
1054                         char *str;
1055
1056                         if (config_status(i915, config))
1057                                 continue;
1058
1059                         if (events[i].global || !HAS_EXTRA_GT_LIST(i915))
1060                                 str = kstrdup(events[i].name, GFP_KERNEL);
1061                         else
1062                                 str = kasprintf(GFP_KERNEL, "%s-gt%u",
1063                                                 events[i].name, j);
1064                         if (!str)
1065                                 goto err;
1066
1067                         *attr_iter++ = &i915_iter->attr.attr;
1068                         i915_iter = add_i915_attr(i915_iter, str, config);
1069
1070                         if (events[i].unit) {
1071                                 if (events[i].global || !HAS_EXTRA_GT_LIST(i915))
1072                                         str = kasprintf(GFP_KERNEL, "%s.unit",
1073                                                         events[i].name);
1074                                 else
1075                                         str = kasprintf(GFP_KERNEL, "%s-gt%u.unit",
1076                                                         events[i].name, j);
1077                                 if (!str)
1078                                         goto err;
1079
1080                                 *attr_iter++ = &pmu_iter->attr.attr;
1081                                 pmu_iter = add_pmu_attr(pmu_iter, str,
1082                                                         events[i].unit);
1083                         }
1084                 }
1085         }
1086
1087         /* Initialize supported engine counters. */
1088         for_each_uabi_engine(engine, i915) {
1089                 for (i = 0; i < ARRAY_SIZE(engine_events); i++) {
1090                         char *str;
1091
1092                         if (engine_event_status(engine,
1093                                                 engine_events[i].sample))
1094                                 continue;
1095
1096                         str = kasprintf(GFP_KERNEL, "%s-%s",
1097                                         engine->name, engine_events[i].name);
1098                         if (!str)
1099                                 goto err;
1100
1101                         *attr_iter++ = &i915_iter->attr.attr;
1102                         i915_iter =
1103                                 add_i915_attr(i915_iter, str,
1104                                               __I915_PMU_ENGINE(engine->uabi_class,
1105                                                                 engine->uabi_instance,
1106                                                                 engine_events[i].sample));
1107
1108                         str = kasprintf(GFP_KERNEL, "%s-%s.unit",
1109                                         engine->name, engine_events[i].name);
1110                         if (!str)
1111                                 goto err;
1112
1113                         *attr_iter++ = &pmu_iter->attr.attr;
1114                         pmu_iter = add_pmu_attr(pmu_iter, str, "ns");
1115                 }
1116         }
1117
1118         pmu->i915_attr = i915_attr;
1119         pmu->pmu_attr = pmu_attr;
1120
1121         return attr;
1122
1123 err:;
1124         for (attr_iter = attr; *attr_iter; attr_iter++)
1125                 kfree((*attr_iter)->name);
1126
1127 err_alloc:
1128         kfree(attr);
1129         kfree(i915_attr);
1130         kfree(pmu_attr);
1131
1132         return NULL;
1133 }
1134
1135 static void free_event_attributes(struct i915_pmu *pmu)
1136 {
1137         struct attribute **attr_iter = pmu->events_attr_group.attrs;
1138
1139         for (; *attr_iter; attr_iter++)
1140                 kfree((*attr_iter)->name);
1141
1142         kfree(pmu->events_attr_group.attrs);
1143         kfree(pmu->i915_attr);
1144         kfree(pmu->pmu_attr);
1145
1146         pmu->events_attr_group.attrs = NULL;
1147         pmu->i915_attr = NULL;
1148         pmu->pmu_attr = NULL;
1149 }
1150
1151 static int i915_pmu_cpu_online(unsigned int cpu, struct hlist_node *node)
1152 {
1153         struct i915_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), cpuhp.node);
1154
1155         GEM_BUG_ON(!pmu->base.event_init);
1156
1157         /* Select the first online CPU as a designated reader. */
1158         if (cpumask_empty(&i915_pmu_cpumask))
1159                 cpumask_set_cpu(cpu, &i915_pmu_cpumask);
1160
1161         return 0;
1162 }
1163
1164 static int i915_pmu_cpu_offline(unsigned int cpu, struct hlist_node *node)
1165 {
1166         struct i915_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), cpuhp.node);
1167         unsigned int target = i915_pmu_target_cpu;
1168
1169         GEM_BUG_ON(!pmu->base.event_init);
1170
1171         /*
1172          * Unregistering an instance generates a CPU offline event which we must
1173          * ignore to avoid incorrectly modifying the shared i915_pmu_cpumask.
1174          */
1175         if (pmu->closed)
1176                 return 0;
1177
1178         if (cpumask_test_and_clear_cpu(cpu, &i915_pmu_cpumask)) {
1179                 target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu);
1180
1181                 /* Migrate events if there is a valid target */
1182                 if (target < nr_cpu_ids) {
1183                         cpumask_set_cpu(target, &i915_pmu_cpumask);
1184                         i915_pmu_target_cpu = target;
1185                 }
1186         }
1187
1188         if (target < nr_cpu_ids && target != pmu->cpuhp.cpu) {
1189                 perf_pmu_migrate_context(&pmu->base, cpu, target);
1190                 pmu->cpuhp.cpu = target;
1191         }
1192
1193         return 0;
1194 }
1195
1196 static enum cpuhp_state cpuhp_slot = CPUHP_INVALID;
1197
1198 int i915_pmu_init(void)
1199 {
1200         int ret;
1201
1202         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1203                                       "perf/x86/intel/i915:online",
1204                                       i915_pmu_cpu_online,
1205                                       i915_pmu_cpu_offline);
1206         if (ret < 0)
1207                 pr_notice("Failed to setup cpuhp state for i915 PMU! (%d)\n",
1208                           ret);
1209         else
1210                 cpuhp_slot = ret;
1211
1212         return 0;
1213 }
1214
1215 void i915_pmu_exit(void)
1216 {
1217         if (cpuhp_slot != CPUHP_INVALID)
1218                 cpuhp_remove_multi_state(cpuhp_slot);
1219 }
1220
1221 static int i915_pmu_register_cpuhp_state(struct i915_pmu *pmu)
1222 {
1223         if (cpuhp_slot == CPUHP_INVALID)
1224                 return -EINVAL;
1225
1226         return cpuhp_state_add_instance(cpuhp_slot, &pmu->cpuhp.node);
1227 }
1228
1229 static void i915_pmu_unregister_cpuhp_state(struct i915_pmu *pmu)
1230 {
1231         cpuhp_state_remove_instance(cpuhp_slot, &pmu->cpuhp.node);
1232 }
1233
1234 static bool is_igp(struct drm_i915_private *i915)
1235 {
1236         struct pci_dev *pdev = to_pci_dev(i915->drm.dev);
1237
1238         /* IGP is 0000:00:02.0 */
1239         return pci_domain_nr(pdev->bus) == 0 &&
1240                pdev->bus->number == 0 &&
1241                PCI_SLOT(pdev->devfn) == 2 &&
1242                PCI_FUNC(pdev->devfn) == 0;
1243 }
1244
1245 void i915_pmu_register(struct drm_i915_private *i915)
1246 {
1247         struct i915_pmu *pmu = &i915->pmu;
1248         const struct attribute_group *attr_groups[] = {
1249                 &i915_pmu_format_attr_group,
1250                 &pmu->events_attr_group,
1251                 &i915_pmu_cpumask_attr_group,
1252                 NULL
1253         };
1254
1255         int ret = -ENOMEM;
1256
1257         if (GRAPHICS_VER(i915) <= 2) {
1258                 drm_info(&i915->drm, "PMU not supported for this GPU.");
1259                 return;
1260         }
1261
1262         spin_lock_init(&pmu->lock);
1263         hrtimer_init(&pmu->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1264         pmu->timer.function = i915_sample;
1265         pmu->cpuhp.cpu = -1;
1266         init_rc6(pmu);
1267
1268         if (!is_igp(i915)) {
1269                 pmu->name = kasprintf(GFP_KERNEL,
1270                                       "i915_%s",
1271                                       dev_name(i915->drm.dev));
1272                 if (pmu->name) {
1273                         /* tools/perf reserves colons as special. */
1274                         strreplace((char *)pmu->name, ':', '_');
1275                 }
1276         } else {
1277                 pmu->name = "i915";
1278         }
1279         if (!pmu->name)
1280                 goto err;
1281
1282         pmu->events_attr_group.name = "events";
1283         pmu->events_attr_group.attrs = create_event_attributes(pmu);
1284         if (!pmu->events_attr_group.attrs)
1285                 goto err_name;
1286
1287         pmu->base.attr_groups = kmemdup(attr_groups, sizeof(attr_groups),
1288                                         GFP_KERNEL);
1289         if (!pmu->base.attr_groups)
1290                 goto err_attr;
1291
1292         pmu->base.module        = THIS_MODULE;
1293         pmu->base.task_ctx_nr   = perf_invalid_context;
1294         pmu->base.event_init    = i915_pmu_event_init;
1295         pmu->base.add           = i915_pmu_event_add;
1296         pmu->base.del           = i915_pmu_event_del;
1297         pmu->base.start         = i915_pmu_event_start;
1298         pmu->base.stop          = i915_pmu_event_stop;
1299         pmu->base.read          = i915_pmu_event_read;
1300         pmu->base.event_idx     = i915_pmu_event_event_idx;
1301
1302         ret = perf_pmu_register(&pmu->base, pmu->name, -1);
1303         if (ret)
1304                 goto err_groups;
1305
1306         ret = i915_pmu_register_cpuhp_state(pmu);
1307         if (ret)
1308                 goto err_unreg;
1309
1310         return;
1311
1312 err_unreg:
1313         perf_pmu_unregister(&pmu->base);
1314 err_groups:
1315         kfree(pmu->base.attr_groups);
1316 err_attr:
1317         pmu->base.event_init = NULL;
1318         free_event_attributes(pmu);
1319 err_name:
1320         if (!is_igp(i915))
1321                 kfree(pmu->name);
1322 err:
1323         drm_notice(&i915->drm, "Failed to register PMU!\n");
1324 }
1325
1326 void i915_pmu_unregister(struct drm_i915_private *i915)
1327 {
1328         struct i915_pmu *pmu = &i915->pmu;
1329
1330         if (!pmu->base.event_init)
1331                 return;
1332
1333         /*
1334          * "Disconnect" the PMU callbacks - since all are atomic synchronize_rcu
1335          * ensures all currently executing ones will have exited before we
1336          * proceed with unregistration.
1337          */
1338         pmu->closed = true;
1339         synchronize_rcu();
1340
1341         hrtimer_cancel(&pmu->timer);
1342
1343         i915_pmu_unregister_cpuhp_state(pmu);
1344
1345         perf_pmu_unregister(&pmu->base);
1346         pmu->base.event_init = NULL;
1347         kfree(pmu->base.attr_groups);
1348         if (!is_igp(i915))
1349                 kfree(pmu->name);
1350         free_event_attributes(pmu);
1351 }
This page took 0.149842 seconds and 4 git commands to generate.