]> Git Repo - linux.git/blob - drivers/gpu/drm/i915/i915_pmu.c
Merge tag 'staging-4.17-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh...
[linux.git] / drivers / gpu / drm / i915 / i915_pmu.c
1 /*
2  * Copyright © 2017 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  */
24
25 #include <linux/perf_event.h>
26 #include <linux/pm_runtime.h>
27
28 #include "i915_drv.h"
29 #include "i915_pmu.h"
30 #include "intel_ringbuffer.h"
31
32 /* Frequency for the sampling timer for events which need it. */
33 #define FREQUENCY 200
34 #define PERIOD max_t(u64, 10000, NSEC_PER_SEC / FREQUENCY)
35
36 #define ENGINE_SAMPLE_MASK \
37         (BIT(I915_SAMPLE_BUSY) | \
38          BIT(I915_SAMPLE_WAIT) | \
39          BIT(I915_SAMPLE_SEMA))
40
41 #define ENGINE_SAMPLE_BITS (1 << I915_PMU_SAMPLE_BITS)
42
43 static cpumask_t i915_pmu_cpumask;
44
45 static u8 engine_config_sample(u64 config)
46 {
47         return config & I915_PMU_SAMPLE_MASK;
48 }
49
50 static u8 engine_event_sample(struct perf_event *event)
51 {
52         return engine_config_sample(event->attr.config);
53 }
54
55 static u8 engine_event_class(struct perf_event *event)
56 {
57         return (event->attr.config >> I915_PMU_CLASS_SHIFT) & 0xff;
58 }
59
60 static u8 engine_event_instance(struct perf_event *event)
61 {
62         return (event->attr.config >> I915_PMU_SAMPLE_BITS) & 0xff;
63 }
64
65 static bool is_engine_config(u64 config)
66 {
67         return config < __I915_PMU_OTHER(0);
68 }
69
70 static unsigned int config_enabled_bit(u64 config)
71 {
72         if (is_engine_config(config))
73                 return engine_config_sample(config);
74         else
75                 return ENGINE_SAMPLE_BITS + (config - __I915_PMU_OTHER(0));
76 }
77
78 static u64 config_enabled_mask(u64 config)
79 {
80         return BIT_ULL(config_enabled_bit(config));
81 }
82
83 static bool is_engine_event(struct perf_event *event)
84 {
85         return is_engine_config(event->attr.config);
86 }
87
88 static unsigned int event_enabled_bit(struct perf_event *event)
89 {
90         return config_enabled_bit(event->attr.config);
91 }
92
93 static bool pmu_needs_timer(struct drm_i915_private *i915, bool gpu_active)
94 {
95         u64 enable;
96
97         /*
98          * Only some counters need the sampling timer.
99          *
100          * We start with a bitmask of all currently enabled events.
101          */
102         enable = i915->pmu.enable;
103
104         /*
105          * Mask out all the ones which do not need the timer, or in
106          * other words keep all the ones that could need the timer.
107          */
108         enable &= config_enabled_mask(I915_PMU_ACTUAL_FREQUENCY) |
109                   config_enabled_mask(I915_PMU_REQUESTED_FREQUENCY) |
110                   ENGINE_SAMPLE_MASK;
111
112         /*
113          * When the GPU is idle per-engine counters do not need to be
114          * running so clear those bits out.
115          */
116         if (!gpu_active)
117                 enable &= ~ENGINE_SAMPLE_MASK;
118         /*
119          * Also there is software busyness tracking available we do not
120          * need the timer for I915_SAMPLE_BUSY counter.
121          *
122          * Use RCS as proxy for all engines.
123          */
124         else if (intel_engine_supports_stats(i915->engine[RCS]))
125                 enable &= ~BIT(I915_SAMPLE_BUSY);
126
127         /*
128          * If some bits remain it means we need the sampling timer running.
129          */
130         return enable;
131 }
132
133 void i915_pmu_gt_parked(struct drm_i915_private *i915)
134 {
135         if (!i915->pmu.base.event_init)
136                 return;
137
138         spin_lock_irq(&i915->pmu.lock);
139         /*
140          * Signal sampling timer to stop if only engine events are enabled and
141          * GPU went idle.
142          */
143         i915->pmu.timer_enabled = pmu_needs_timer(i915, false);
144         spin_unlock_irq(&i915->pmu.lock);
145 }
146
147 static void __i915_pmu_maybe_start_timer(struct drm_i915_private *i915)
148 {
149         if (!i915->pmu.timer_enabled && pmu_needs_timer(i915, true)) {
150                 i915->pmu.timer_enabled = true;
151                 hrtimer_start_range_ns(&i915->pmu.timer,
152                                        ns_to_ktime(PERIOD), 0,
153                                        HRTIMER_MODE_REL_PINNED);
154         }
155 }
156
157 void i915_pmu_gt_unparked(struct drm_i915_private *i915)
158 {
159         if (!i915->pmu.base.event_init)
160                 return;
161
162         spin_lock_irq(&i915->pmu.lock);
163         /*
164          * Re-enable sampling timer when GPU goes active.
165          */
166         __i915_pmu_maybe_start_timer(i915);
167         spin_unlock_irq(&i915->pmu.lock);
168 }
169
170 static bool grab_forcewake(struct drm_i915_private *i915, bool fw)
171 {
172         if (!fw)
173                 intel_uncore_forcewake_get(i915, FORCEWAKE_ALL);
174
175         return true;
176 }
177
178 static void
179 update_sample(struct i915_pmu_sample *sample, u32 unit, u32 val)
180 {
181         sample->cur += mul_u32_u32(val, unit);
182 }
183
184 static void engines_sample(struct drm_i915_private *dev_priv)
185 {
186         struct intel_engine_cs *engine;
187         enum intel_engine_id id;
188         bool fw = false;
189
190         if ((dev_priv->pmu.enable & ENGINE_SAMPLE_MASK) == 0)
191                 return;
192
193         if (!dev_priv->gt.awake)
194                 return;
195
196         if (!intel_runtime_pm_get_if_in_use(dev_priv))
197                 return;
198
199         for_each_engine(engine, dev_priv, id) {
200                 u32 current_seqno = intel_engine_get_seqno(engine);
201                 u32 last_seqno = intel_engine_last_submit(engine);
202                 u32 val;
203
204                 val = !i915_seqno_passed(current_seqno, last_seqno);
205
206                 update_sample(&engine->pmu.sample[I915_SAMPLE_BUSY],
207                               PERIOD, val);
208
209                 if (val && (engine->pmu.enable &
210                     (BIT(I915_SAMPLE_WAIT) | BIT(I915_SAMPLE_SEMA)))) {
211                         fw = grab_forcewake(dev_priv, fw);
212
213                         val = I915_READ_FW(RING_CTL(engine->mmio_base));
214                 } else {
215                         val = 0;
216                 }
217
218                 update_sample(&engine->pmu.sample[I915_SAMPLE_WAIT],
219                               PERIOD, !!(val & RING_WAIT));
220
221                 update_sample(&engine->pmu.sample[I915_SAMPLE_SEMA],
222                               PERIOD, !!(val & RING_WAIT_SEMAPHORE));
223         }
224
225         if (fw)
226                 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
227
228         intel_runtime_pm_put(dev_priv);
229 }
230
231 static void frequency_sample(struct drm_i915_private *dev_priv)
232 {
233         if (dev_priv->pmu.enable &
234             config_enabled_mask(I915_PMU_ACTUAL_FREQUENCY)) {
235                 u32 val;
236
237                 val = dev_priv->gt_pm.rps.cur_freq;
238                 if (dev_priv->gt.awake &&
239                     intel_runtime_pm_get_if_in_use(dev_priv)) {
240                         val = intel_get_cagf(dev_priv,
241                                              I915_READ_NOTRACE(GEN6_RPSTAT1));
242                         intel_runtime_pm_put(dev_priv);
243                 }
244
245                 update_sample(&dev_priv->pmu.sample[__I915_SAMPLE_FREQ_ACT],
246                               1, intel_gpu_freq(dev_priv, val));
247         }
248
249         if (dev_priv->pmu.enable &
250             config_enabled_mask(I915_PMU_REQUESTED_FREQUENCY)) {
251                 update_sample(&dev_priv->pmu.sample[__I915_SAMPLE_FREQ_REQ], 1,
252                               intel_gpu_freq(dev_priv,
253                                              dev_priv->gt_pm.rps.cur_freq));
254         }
255 }
256
257 static enum hrtimer_restart i915_sample(struct hrtimer *hrtimer)
258 {
259         struct drm_i915_private *i915 =
260                 container_of(hrtimer, struct drm_i915_private, pmu.timer);
261
262         if (!READ_ONCE(i915->pmu.timer_enabled))
263                 return HRTIMER_NORESTART;
264
265         engines_sample(i915);
266         frequency_sample(i915);
267
268         hrtimer_forward_now(hrtimer, ns_to_ktime(PERIOD));
269         return HRTIMER_RESTART;
270 }
271
272 static u64 count_interrupts(struct drm_i915_private *i915)
273 {
274         /* open-coded kstat_irqs() */
275         struct irq_desc *desc = irq_to_desc(i915->drm.pdev->irq);
276         u64 sum = 0;
277         int cpu;
278
279         if (!desc || !desc->kstat_irqs)
280                 return 0;
281
282         for_each_possible_cpu(cpu)
283                 sum += *per_cpu_ptr(desc->kstat_irqs, cpu);
284
285         return sum;
286 }
287
288 static void engine_event_destroy(struct perf_event *event)
289 {
290         struct drm_i915_private *i915 =
291                 container_of(event->pmu, typeof(*i915), pmu.base);
292         struct intel_engine_cs *engine;
293
294         engine = intel_engine_lookup_user(i915,
295                                           engine_event_class(event),
296                                           engine_event_instance(event));
297         if (WARN_ON_ONCE(!engine))
298                 return;
299
300         if (engine_event_sample(event) == I915_SAMPLE_BUSY &&
301             intel_engine_supports_stats(engine))
302                 intel_disable_engine_stats(engine);
303 }
304
305 static void i915_pmu_event_destroy(struct perf_event *event)
306 {
307         WARN_ON(event->parent);
308
309         if (is_engine_event(event))
310                 engine_event_destroy(event);
311 }
312
313 static int
314 engine_event_status(struct intel_engine_cs *engine,
315                     enum drm_i915_pmu_engine_sample sample)
316 {
317         switch (sample) {
318         case I915_SAMPLE_BUSY:
319         case I915_SAMPLE_WAIT:
320                 break;
321         case I915_SAMPLE_SEMA:
322                 if (INTEL_GEN(engine->i915) < 6)
323                         return -ENODEV;
324                 break;
325         default:
326                 return -ENOENT;
327         }
328
329         return 0;
330 }
331
332 static int
333 config_status(struct drm_i915_private *i915, u64 config)
334 {
335         switch (config) {
336         case I915_PMU_ACTUAL_FREQUENCY:
337                 if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915))
338                         /* Requires a mutex for sampling! */
339                         return -ENODEV;
340                 /* Fall-through. */
341         case I915_PMU_REQUESTED_FREQUENCY:
342                 if (INTEL_GEN(i915) < 6)
343                         return -ENODEV;
344                 break;
345         case I915_PMU_INTERRUPTS:
346                 break;
347         case I915_PMU_RC6_RESIDENCY:
348                 if (!HAS_RC6(i915))
349                         return -ENODEV;
350                 break;
351         default:
352                 return -ENOENT;
353         }
354
355         return 0;
356 }
357
358 static int engine_event_init(struct perf_event *event)
359 {
360         struct drm_i915_private *i915 =
361                 container_of(event->pmu, typeof(*i915), pmu.base);
362         struct intel_engine_cs *engine;
363         u8 sample;
364         int ret;
365
366         engine = intel_engine_lookup_user(i915, engine_event_class(event),
367                                           engine_event_instance(event));
368         if (!engine)
369                 return -ENODEV;
370
371         sample = engine_event_sample(event);
372         ret = engine_event_status(engine, sample);
373         if (ret)
374                 return ret;
375
376         if (sample == I915_SAMPLE_BUSY && intel_engine_supports_stats(engine))
377                 ret = intel_enable_engine_stats(engine);
378
379         return ret;
380 }
381
382 static int i915_pmu_event_init(struct perf_event *event)
383 {
384         struct drm_i915_private *i915 =
385                 container_of(event->pmu, typeof(*i915), pmu.base);
386         int ret;
387
388         if (event->attr.type != event->pmu->type)
389                 return -ENOENT;
390
391         /* unsupported modes and filters */
392         if (event->attr.sample_period) /* no sampling */
393                 return -EINVAL;
394
395         if (has_branch_stack(event))
396                 return -EOPNOTSUPP;
397
398         if (event->cpu < 0)
399                 return -EINVAL;
400
401         /* only allow running on one cpu at a time */
402         if (!cpumask_test_cpu(event->cpu, &i915_pmu_cpumask))
403                 return -EINVAL;
404
405         if (is_engine_event(event))
406                 ret = engine_event_init(event);
407         else
408                 ret = config_status(i915, event->attr.config);
409         if (ret)
410                 return ret;
411
412         if (!event->parent)
413                 event->destroy = i915_pmu_event_destroy;
414
415         return 0;
416 }
417
418 static u64 __get_rc6(struct drm_i915_private *i915)
419 {
420         u64 val;
421
422         val = intel_rc6_residency_ns(i915,
423                                      IS_VALLEYVIEW(i915) ?
424                                      VLV_GT_RENDER_RC6 :
425                                      GEN6_GT_GFX_RC6);
426
427         if (HAS_RC6p(i915))
428                 val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
429
430         if (HAS_RC6pp(i915))
431                 val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
432
433         return val;
434 }
435
436 static u64 get_rc6(struct drm_i915_private *i915)
437 {
438 #if IS_ENABLED(CONFIG_PM)
439         unsigned long flags;
440         u64 val;
441
442         if (intel_runtime_pm_get_if_in_use(i915)) {
443                 val = __get_rc6(i915);
444                 intel_runtime_pm_put(i915);
445
446                 /*
447                  * If we are coming back from being runtime suspended we must
448                  * be careful not to report a larger value than returned
449                  * previously.
450                  */
451
452                 spin_lock_irqsave(&i915->pmu.lock, flags);
453
454                 if (val >= i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur) {
455                         i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur = 0;
456                         i915->pmu.sample[__I915_SAMPLE_RC6].cur = val;
457                 } else {
458                         val = i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur;
459                 }
460
461                 spin_unlock_irqrestore(&i915->pmu.lock, flags);
462         } else {
463                 struct pci_dev *pdev = i915->drm.pdev;
464                 struct device *kdev = &pdev->dev;
465
466                 /*
467                  * We are runtime suspended.
468                  *
469                  * Report the delta from when the device was suspended to now,
470                  * on top of the last known real value, as the approximated RC6
471                  * counter value.
472                  */
473                 spin_lock_irqsave(&i915->pmu.lock, flags);
474                 spin_lock(&kdev->power.lock);
475
476                 if (!i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur)
477                         i915->pmu.suspended_jiffies_last =
478                                                 kdev->power.suspended_jiffies;
479
480                 val = kdev->power.suspended_jiffies -
481                       i915->pmu.suspended_jiffies_last;
482                 val += jiffies - kdev->power.accounting_timestamp;
483
484                 spin_unlock(&kdev->power.lock);
485
486                 val = jiffies_to_nsecs(val);
487                 val += i915->pmu.sample[__I915_SAMPLE_RC6].cur;
488                 i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur = val;
489
490                 spin_unlock_irqrestore(&i915->pmu.lock, flags);
491         }
492
493         return val;
494 #else
495         return __get_rc6(i915);
496 #endif
497 }
498
499 static u64 __i915_pmu_event_read(struct perf_event *event)
500 {
501         struct drm_i915_private *i915 =
502                 container_of(event->pmu, typeof(*i915), pmu.base);
503         u64 val = 0;
504
505         if (is_engine_event(event)) {
506                 u8 sample = engine_event_sample(event);
507                 struct intel_engine_cs *engine;
508
509                 engine = intel_engine_lookup_user(i915,
510                                                   engine_event_class(event),
511                                                   engine_event_instance(event));
512
513                 if (WARN_ON_ONCE(!engine)) {
514                         /* Do nothing */
515                 } else if (sample == I915_SAMPLE_BUSY &&
516                            intel_engine_supports_stats(engine)) {
517                         val = ktime_to_ns(intel_engine_get_busy_time(engine));
518                 } else {
519                         val = engine->pmu.sample[sample].cur;
520                 }
521         } else {
522                 switch (event->attr.config) {
523                 case I915_PMU_ACTUAL_FREQUENCY:
524                         val =
525                            div_u64(i915->pmu.sample[__I915_SAMPLE_FREQ_ACT].cur,
526                                    FREQUENCY);
527                         break;
528                 case I915_PMU_REQUESTED_FREQUENCY:
529                         val =
530                            div_u64(i915->pmu.sample[__I915_SAMPLE_FREQ_REQ].cur,
531                                    FREQUENCY);
532                         break;
533                 case I915_PMU_INTERRUPTS:
534                         val = count_interrupts(i915);
535                         break;
536                 case I915_PMU_RC6_RESIDENCY:
537                         val = get_rc6(i915);
538                         break;
539                 }
540         }
541
542         return val;
543 }
544
545 static void i915_pmu_event_read(struct perf_event *event)
546 {
547         struct hw_perf_event *hwc = &event->hw;
548         u64 prev, new;
549
550 again:
551         prev = local64_read(&hwc->prev_count);
552         new = __i915_pmu_event_read(event);
553
554         if (local64_cmpxchg(&hwc->prev_count, prev, new) != prev)
555                 goto again;
556
557         local64_add(new - prev, &event->count);
558 }
559
560 static void i915_pmu_enable(struct perf_event *event)
561 {
562         struct drm_i915_private *i915 =
563                 container_of(event->pmu, typeof(*i915), pmu.base);
564         unsigned int bit = event_enabled_bit(event);
565         unsigned long flags;
566
567         spin_lock_irqsave(&i915->pmu.lock, flags);
568
569         /*
570          * Update the bitmask of enabled events and increment
571          * the event reference counter.
572          */
573         GEM_BUG_ON(bit >= I915_PMU_MASK_BITS);
574         GEM_BUG_ON(i915->pmu.enable_count[bit] == ~0);
575         i915->pmu.enable |= BIT_ULL(bit);
576         i915->pmu.enable_count[bit]++;
577
578         /*
579          * Start the sampling timer if needed and not already enabled.
580          */
581         __i915_pmu_maybe_start_timer(i915);
582
583         /*
584          * For per-engine events the bitmask and reference counting
585          * is stored per engine.
586          */
587         if (is_engine_event(event)) {
588                 u8 sample = engine_event_sample(event);
589                 struct intel_engine_cs *engine;
590
591                 engine = intel_engine_lookup_user(i915,
592                                                   engine_event_class(event),
593                                                   engine_event_instance(event));
594                 GEM_BUG_ON(!engine);
595                 engine->pmu.enable |= BIT(sample);
596
597                 GEM_BUG_ON(sample >= I915_PMU_SAMPLE_BITS);
598                 GEM_BUG_ON(engine->pmu.enable_count[sample] == ~0);
599                 engine->pmu.enable_count[sample]++;
600         }
601
602         spin_unlock_irqrestore(&i915->pmu.lock, flags);
603
604         /*
605          * Store the current counter value so we can report the correct delta
606          * for all listeners. Even when the event was already enabled and has
607          * an existing non-zero value.
608          */
609         local64_set(&event->hw.prev_count, __i915_pmu_event_read(event));
610 }
611
612 static void i915_pmu_disable(struct perf_event *event)
613 {
614         struct drm_i915_private *i915 =
615                 container_of(event->pmu, typeof(*i915), pmu.base);
616         unsigned int bit = event_enabled_bit(event);
617         unsigned long flags;
618
619         spin_lock_irqsave(&i915->pmu.lock, flags);
620
621         if (is_engine_event(event)) {
622                 u8 sample = engine_event_sample(event);
623                 struct intel_engine_cs *engine;
624
625                 engine = intel_engine_lookup_user(i915,
626                                                   engine_event_class(event),
627                                                   engine_event_instance(event));
628                 GEM_BUG_ON(!engine);
629                 GEM_BUG_ON(sample >= I915_PMU_SAMPLE_BITS);
630                 GEM_BUG_ON(engine->pmu.enable_count[sample] == 0);
631                 /*
632                  * Decrement the reference count and clear the enabled
633                  * bitmask when the last listener on an event goes away.
634                  */
635                 if (--engine->pmu.enable_count[sample] == 0)
636                         engine->pmu.enable &= ~BIT(sample);
637         }
638
639         GEM_BUG_ON(bit >= I915_PMU_MASK_BITS);
640         GEM_BUG_ON(i915->pmu.enable_count[bit] == 0);
641         /*
642          * Decrement the reference count and clear the enabled
643          * bitmask when the last listener on an event goes away.
644          */
645         if (--i915->pmu.enable_count[bit] == 0) {
646                 i915->pmu.enable &= ~BIT_ULL(bit);
647                 i915->pmu.timer_enabled &= pmu_needs_timer(i915, true);
648         }
649
650         spin_unlock_irqrestore(&i915->pmu.lock, flags);
651 }
652
653 static void i915_pmu_event_start(struct perf_event *event, int flags)
654 {
655         i915_pmu_enable(event);
656         event->hw.state = 0;
657 }
658
659 static void i915_pmu_event_stop(struct perf_event *event, int flags)
660 {
661         if (flags & PERF_EF_UPDATE)
662                 i915_pmu_event_read(event);
663         i915_pmu_disable(event);
664         event->hw.state = PERF_HES_STOPPED;
665 }
666
667 static int i915_pmu_event_add(struct perf_event *event, int flags)
668 {
669         if (flags & PERF_EF_START)
670                 i915_pmu_event_start(event, flags);
671
672         return 0;
673 }
674
675 static void i915_pmu_event_del(struct perf_event *event, int flags)
676 {
677         i915_pmu_event_stop(event, PERF_EF_UPDATE);
678 }
679
680 static int i915_pmu_event_event_idx(struct perf_event *event)
681 {
682         return 0;
683 }
684
685 struct i915_str_attribute {
686         struct device_attribute attr;
687         const char *str;
688 };
689
690 static ssize_t i915_pmu_format_show(struct device *dev,
691                                     struct device_attribute *attr, char *buf)
692 {
693         struct i915_str_attribute *eattr;
694
695         eattr = container_of(attr, struct i915_str_attribute, attr);
696         return sprintf(buf, "%s\n", eattr->str);
697 }
698
699 #define I915_PMU_FORMAT_ATTR(_name, _config) \
700         (&((struct i915_str_attribute[]) { \
701                 { .attr = __ATTR(_name, 0444, i915_pmu_format_show, NULL), \
702                   .str = _config, } \
703         })[0].attr.attr)
704
705 static struct attribute *i915_pmu_format_attrs[] = {
706         I915_PMU_FORMAT_ATTR(i915_eventid, "config:0-20"),
707         NULL,
708 };
709
710 static const struct attribute_group i915_pmu_format_attr_group = {
711         .name = "format",
712         .attrs = i915_pmu_format_attrs,
713 };
714
715 struct i915_ext_attribute {
716         struct device_attribute attr;
717         unsigned long val;
718 };
719
720 static ssize_t i915_pmu_event_show(struct device *dev,
721                                    struct device_attribute *attr, char *buf)
722 {
723         struct i915_ext_attribute *eattr;
724
725         eattr = container_of(attr, struct i915_ext_attribute, attr);
726         return sprintf(buf, "config=0x%lx\n", eattr->val);
727 }
728
729 static struct attribute_group i915_pmu_events_attr_group = {
730         .name = "events",
731         /* Patch in attrs at runtime. */
732 };
733
734 static ssize_t
735 i915_pmu_get_attr_cpumask(struct device *dev,
736                           struct device_attribute *attr,
737                           char *buf)
738 {
739         return cpumap_print_to_pagebuf(true, buf, &i915_pmu_cpumask);
740 }
741
742 static DEVICE_ATTR(cpumask, 0444, i915_pmu_get_attr_cpumask, NULL);
743
744 static struct attribute *i915_cpumask_attrs[] = {
745         &dev_attr_cpumask.attr,
746         NULL,
747 };
748
749 static const struct attribute_group i915_pmu_cpumask_attr_group = {
750         .attrs = i915_cpumask_attrs,
751 };
752
753 static const struct attribute_group *i915_pmu_attr_groups[] = {
754         &i915_pmu_format_attr_group,
755         &i915_pmu_events_attr_group,
756         &i915_pmu_cpumask_attr_group,
757         NULL
758 };
759
760 #define __event(__config, __name, __unit) \
761 { \
762         .config = (__config), \
763         .name = (__name), \
764         .unit = (__unit), \
765 }
766
767 #define __engine_event(__sample, __name) \
768 { \
769         .sample = (__sample), \
770         .name = (__name), \
771 }
772
773 static struct i915_ext_attribute *
774 add_i915_attr(struct i915_ext_attribute *attr, const char *name, u64 config)
775 {
776         sysfs_attr_init(&attr->attr.attr);
777         attr->attr.attr.name = name;
778         attr->attr.attr.mode = 0444;
779         attr->attr.show = i915_pmu_event_show;
780         attr->val = config;
781
782         return ++attr;
783 }
784
785 static struct perf_pmu_events_attr *
786 add_pmu_attr(struct perf_pmu_events_attr *attr, const char *name,
787              const char *str)
788 {
789         sysfs_attr_init(&attr->attr.attr);
790         attr->attr.attr.name = name;
791         attr->attr.attr.mode = 0444;
792         attr->attr.show = perf_event_sysfs_show;
793         attr->event_str = str;
794
795         return ++attr;
796 }
797
798 static struct attribute **
799 create_event_attributes(struct drm_i915_private *i915)
800 {
801         static const struct {
802                 u64 config;
803                 const char *name;
804                 const char *unit;
805         } events[] = {
806                 __event(I915_PMU_ACTUAL_FREQUENCY, "actual-frequency", "MHz"),
807                 __event(I915_PMU_REQUESTED_FREQUENCY, "requested-frequency", "MHz"),
808                 __event(I915_PMU_INTERRUPTS, "interrupts", NULL),
809                 __event(I915_PMU_RC6_RESIDENCY, "rc6-residency", "ns"),
810         };
811         static const struct {
812                 enum drm_i915_pmu_engine_sample sample;
813                 char *name;
814         } engine_events[] = {
815                 __engine_event(I915_SAMPLE_BUSY, "busy"),
816                 __engine_event(I915_SAMPLE_SEMA, "sema"),
817                 __engine_event(I915_SAMPLE_WAIT, "wait"),
818         };
819         unsigned int count = 0;
820         struct perf_pmu_events_attr *pmu_attr = NULL, *pmu_iter;
821         struct i915_ext_attribute *i915_attr = NULL, *i915_iter;
822         struct attribute **attr = NULL, **attr_iter;
823         struct intel_engine_cs *engine;
824         enum intel_engine_id id;
825         unsigned int i;
826
827         /* Count how many counters we will be exposing. */
828         for (i = 0; i < ARRAY_SIZE(events); i++) {
829                 if (!config_status(i915, events[i].config))
830                         count++;
831         }
832
833         for_each_engine(engine, i915, id) {
834                 for (i = 0; i < ARRAY_SIZE(engine_events); i++) {
835                         if (!engine_event_status(engine,
836                                                  engine_events[i].sample))
837                                 count++;
838                 }
839         }
840
841         /* Allocate attribute objects and table. */
842         i915_attr = kcalloc(count, sizeof(*i915_attr), GFP_KERNEL);
843         if (!i915_attr)
844                 goto err_alloc;
845
846         pmu_attr = kcalloc(count, sizeof(*pmu_attr), GFP_KERNEL);
847         if (!pmu_attr)
848                 goto err_alloc;
849
850         /* Max one pointer of each attribute type plus a termination entry. */
851         attr = kcalloc(count * 2 + 1, sizeof(*attr), GFP_KERNEL);
852         if (!attr)
853                 goto err_alloc;
854
855         i915_iter = i915_attr;
856         pmu_iter = pmu_attr;
857         attr_iter = attr;
858
859         /* Initialize supported non-engine counters. */
860         for (i = 0; i < ARRAY_SIZE(events); i++) {
861                 char *str;
862
863                 if (config_status(i915, events[i].config))
864                         continue;
865
866                 str = kstrdup(events[i].name, GFP_KERNEL);
867                 if (!str)
868                         goto err;
869
870                 *attr_iter++ = &i915_iter->attr.attr;
871                 i915_iter = add_i915_attr(i915_iter, str, events[i].config);
872
873                 if (events[i].unit) {
874                         str = kasprintf(GFP_KERNEL, "%s.unit", events[i].name);
875                         if (!str)
876                                 goto err;
877
878                         *attr_iter++ = &pmu_iter->attr.attr;
879                         pmu_iter = add_pmu_attr(pmu_iter, str, events[i].unit);
880                 }
881         }
882
883         /* Initialize supported engine counters. */
884         for_each_engine(engine, i915, id) {
885                 for (i = 0; i < ARRAY_SIZE(engine_events); i++) {
886                         char *str;
887
888                         if (engine_event_status(engine,
889                                                 engine_events[i].sample))
890                                 continue;
891
892                         str = kasprintf(GFP_KERNEL, "%s-%s",
893                                         engine->name, engine_events[i].name);
894                         if (!str)
895                                 goto err;
896
897                         *attr_iter++ = &i915_iter->attr.attr;
898                         i915_iter =
899                                 add_i915_attr(i915_iter, str,
900                                               __I915_PMU_ENGINE(engine->uabi_class,
901                                                                 engine->instance,
902                                                                 engine_events[i].sample));
903
904                         str = kasprintf(GFP_KERNEL, "%s-%s.unit",
905                                         engine->name, engine_events[i].name);
906                         if (!str)
907                                 goto err;
908
909                         *attr_iter++ = &pmu_iter->attr.attr;
910                         pmu_iter = add_pmu_attr(pmu_iter, str, "ns");
911                 }
912         }
913
914         i915->pmu.i915_attr = i915_attr;
915         i915->pmu.pmu_attr = pmu_attr;
916
917         return attr;
918
919 err:;
920         for (attr_iter = attr; *attr_iter; attr_iter++)
921                 kfree((*attr_iter)->name);
922
923 err_alloc:
924         kfree(attr);
925         kfree(i915_attr);
926         kfree(pmu_attr);
927
928         return NULL;
929 }
930
931 static void free_event_attributes(struct drm_i915_private *i915)
932 {
933         struct attribute **attr_iter = i915_pmu_events_attr_group.attrs;
934
935         for (; *attr_iter; attr_iter++)
936                 kfree((*attr_iter)->name);
937
938         kfree(i915_pmu_events_attr_group.attrs);
939         kfree(i915->pmu.i915_attr);
940         kfree(i915->pmu.pmu_attr);
941
942         i915_pmu_events_attr_group.attrs = NULL;
943         i915->pmu.i915_attr = NULL;
944         i915->pmu.pmu_attr = NULL;
945 }
946
947 static int i915_pmu_cpu_online(unsigned int cpu, struct hlist_node *node)
948 {
949         struct i915_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), node);
950
951         GEM_BUG_ON(!pmu->base.event_init);
952
953         /* Select the first online CPU as a designated reader. */
954         if (!cpumask_weight(&i915_pmu_cpumask))
955                 cpumask_set_cpu(cpu, &i915_pmu_cpumask);
956
957         return 0;
958 }
959
960 static int i915_pmu_cpu_offline(unsigned int cpu, struct hlist_node *node)
961 {
962         struct i915_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), node);
963         unsigned int target;
964
965         GEM_BUG_ON(!pmu->base.event_init);
966
967         if (cpumask_test_and_clear_cpu(cpu, &i915_pmu_cpumask)) {
968                 target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu);
969                 /* Migrate events if there is a valid target */
970                 if (target < nr_cpu_ids) {
971                         cpumask_set_cpu(target, &i915_pmu_cpumask);
972                         perf_pmu_migrate_context(&pmu->base, cpu, target);
973                 }
974         }
975
976         return 0;
977 }
978
979 static enum cpuhp_state cpuhp_slot = CPUHP_INVALID;
980
981 static int i915_pmu_register_cpuhp_state(struct drm_i915_private *i915)
982 {
983         enum cpuhp_state slot;
984         int ret;
985
986         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
987                                       "perf/x86/intel/i915:online",
988                                       i915_pmu_cpu_online,
989                                       i915_pmu_cpu_offline);
990         if (ret < 0)
991                 return ret;
992
993         slot = ret;
994         ret = cpuhp_state_add_instance(slot, &i915->pmu.node);
995         if (ret) {
996                 cpuhp_remove_multi_state(slot);
997                 return ret;
998         }
999
1000         cpuhp_slot = slot;
1001         return 0;
1002 }
1003
1004 static void i915_pmu_unregister_cpuhp_state(struct drm_i915_private *i915)
1005 {
1006         WARN_ON(cpuhp_slot == CPUHP_INVALID);
1007         WARN_ON(cpuhp_state_remove_instance(cpuhp_slot, &i915->pmu.node));
1008         cpuhp_remove_multi_state(cpuhp_slot);
1009 }
1010
1011 void i915_pmu_register(struct drm_i915_private *i915)
1012 {
1013         int ret;
1014
1015         if (INTEL_GEN(i915) <= 2) {
1016                 DRM_INFO("PMU not supported for this GPU.");
1017                 return;
1018         }
1019
1020         i915_pmu_events_attr_group.attrs = create_event_attributes(i915);
1021         if (!i915_pmu_events_attr_group.attrs) {
1022                 ret = -ENOMEM;
1023                 goto err;
1024         }
1025
1026         i915->pmu.base.attr_groups      = i915_pmu_attr_groups;
1027         i915->pmu.base.task_ctx_nr      = perf_invalid_context;
1028         i915->pmu.base.event_init       = i915_pmu_event_init;
1029         i915->pmu.base.add              = i915_pmu_event_add;
1030         i915->pmu.base.del              = i915_pmu_event_del;
1031         i915->pmu.base.start            = i915_pmu_event_start;
1032         i915->pmu.base.stop             = i915_pmu_event_stop;
1033         i915->pmu.base.read             = i915_pmu_event_read;
1034         i915->pmu.base.event_idx        = i915_pmu_event_event_idx;
1035
1036         spin_lock_init(&i915->pmu.lock);
1037         hrtimer_init(&i915->pmu.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1038         i915->pmu.timer.function = i915_sample;
1039
1040         ret = perf_pmu_register(&i915->pmu.base, "i915", -1);
1041         if (ret)
1042                 goto err;
1043
1044         ret = i915_pmu_register_cpuhp_state(i915);
1045         if (ret)
1046                 goto err_unreg;
1047
1048         return;
1049
1050 err_unreg:
1051         perf_pmu_unregister(&i915->pmu.base);
1052 err:
1053         i915->pmu.base.event_init = NULL;
1054         free_event_attributes(i915);
1055         DRM_NOTE("Failed to register PMU! (err=%d)\n", ret);
1056 }
1057
1058 void i915_pmu_unregister(struct drm_i915_private *i915)
1059 {
1060         if (!i915->pmu.base.event_init)
1061                 return;
1062
1063         WARN_ON(i915->pmu.enable);
1064
1065         hrtimer_cancel(&i915->pmu.timer);
1066
1067         i915_pmu_unregister_cpuhp_state(i915);
1068
1069         perf_pmu_unregister(&i915->pmu.base);
1070         i915->pmu.base.event_init = NULL;
1071         free_event_attributes(i915);
1072 }
This page took 0.096911 seconds and 4 git commands to generate.