]> asedeno.scripts.mit.edu Git - linux.git/blob - drivers/gpu/drm/i915/i915_pmu.c
sparc,leon: Select USB_UHCI_BIG_ENDIAN_{MMIO,DESC}
[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 i915_pmu_event_destroy(struct perf_event *event)
289 {
290         WARN_ON(event->parent);
291 }
292
293 static int engine_event_init(struct perf_event *event)
294 {
295         struct drm_i915_private *i915 =
296                 container_of(event->pmu, typeof(*i915), pmu.base);
297
298         if (!intel_engine_lookup_user(i915, engine_event_class(event),
299                                       engine_event_instance(event)))
300                 return -ENODEV;
301
302         switch (engine_event_sample(event)) {
303         case I915_SAMPLE_BUSY:
304         case I915_SAMPLE_WAIT:
305                 break;
306         case I915_SAMPLE_SEMA:
307                 if (INTEL_GEN(i915) < 6)
308                         return -ENODEV;
309                 break;
310         default:
311                 return -ENOENT;
312         }
313
314         return 0;
315 }
316
317 static int i915_pmu_event_init(struct perf_event *event)
318 {
319         struct drm_i915_private *i915 =
320                 container_of(event->pmu, typeof(*i915), pmu.base);
321         int ret;
322
323         if (event->attr.type != event->pmu->type)
324                 return -ENOENT;
325
326         /* unsupported modes and filters */
327         if (event->attr.sample_period) /* no sampling */
328                 return -EINVAL;
329
330         if (has_branch_stack(event))
331                 return -EOPNOTSUPP;
332
333         if (event->cpu < 0)
334                 return -EINVAL;
335
336         /* only allow running on one cpu at a time */
337         if (!cpumask_test_cpu(event->cpu, &i915_pmu_cpumask))
338                 return -EINVAL;
339
340         if (is_engine_event(event)) {
341                 ret = engine_event_init(event);
342         } else {
343                 ret = 0;
344                 switch (event->attr.config) {
345                 case I915_PMU_ACTUAL_FREQUENCY:
346                         if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915))
347                                  /* Requires a mutex for sampling! */
348                                 ret = -ENODEV;
349                 case I915_PMU_REQUESTED_FREQUENCY:
350                         if (INTEL_GEN(i915) < 6)
351                                 ret = -ENODEV;
352                         break;
353                 case I915_PMU_INTERRUPTS:
354                         break;
355                 case I915_PMU_RC6_RESIDENCY:
356                         if (!HAS_RC6(i915))
357                                 ret = -ENODEV;
358                         break;
359                 default:
360                         ret = -ENOENT;
361                         break;
362                 }
363         }
364         if (ret)
365                 return ret;
366
367         if (!event->parent)
368                 event->destroy = i915_pmu_event_destroy;
369
370         return 0;
371 }
372
373 static u64 __i915_pmu_event_read(struct perf_event *event)
374 {
375         struct drm_i915_private *i915 =
376                 container_of(event->pmu, typeof(*i915), pmu.base);
377         u64 val = 0;
378
379         if (is_engine_event(event)) {
380                 u8 sample = engine_event_sample(event);
381                 struct intel_engine_cs *engine;
382
383                 engine = intel_engine_lookup_user(i915,
384                                                   engine_event_class(event),
385                                                   engine_event_instance(event));
386
387                 if (WARN_ON_ONCE(!engine)) {
388                         /* Do nothing */
389                 } else if (sample == I915_SAMPLE_BUSY &&
390                            engine->pmu.busy_stats) {
391                         val = ktime_to_ns(intel_engine_get_busy_time(engine));
392                 } else {
393                         val = engine->pmu.sample[sample].cur;
394                 }
395         } else {
396                 switch (event->attr.config) {
397                 case I915_PMU_ACTUAL_FREQUENCY:
398                         val =
399                            div_u64(i915->pmu.sample[__I915_SAMPLE_FREQ_ACT].cur,
400                                    FREQUENCY);
401                         break;
402                 case I915_PMU_REQUESTED_FREQUENCY:
403                         val =
404                            div_u64(i915->pmu.sample[__I915_SAMPLE_FREQ_REQ].cur,
405                                    FREQUENCY);
406                         break;
407                 case I915_PMU_INTERRUPTS:
408                         val = count_interrupts(i915);
409                         break;
410                 case I915_PMU_RC6_RESIDENCY:
411                         intel_runtime_pm_get(i915);
412                         val = intel_rc6_residency_ns(i915,
413                                                      IS_VALLEYVIEW(i915) ?
414                                                      VLV_GT_RENDER_RC6 :
415                                                      GEN6_GT_GFX_RC6);
416                         if (HAS_RC6p(i915))
417                                 val += intel_rc6_residency_ns(i915,
418                                                               GEN6_GT_GFX_RC6p);
419                         if (HAS_RC6pp(i915))
420                                 val += intel_rc6_residency_ns(i915,
421                                                               GEN6_GT_GFX_RC6pp);
422                         intel_runtime_pm_put(i915);
423                         break;
424                 }
425         }
426
427         return val;
428 }
429
430 static void i915_pmu_event_read(struct perf_event *event)
431 {
432         struct hw_perf_event *hwc = &event->hw;
433         u64 prev, new;
434
435 again:
436         prev = local64_read(&hwc->prev_count);
437         new = __i915_pmu_event_read(event);
438
439         if (local64_cmpxchg(&hwc->prev_count, prev, new) != prev)
440                 goto again;
441
442         local64_add(new - prev, &event->count);
443 }
444
445 static bool engine_needs_busy_stats(struct intel_engine_cs *engine)
446 {
447         return intel_engine_supports_stats(engine) &&
448                (engine->pmu.enable & BIT(I915_SAMPLE_BUSY));
449 }
450
451 static void i915_pmu_enable(struct perf_event *event)
452 {
453         struct drm_i915_private *i915 =
454                 container_of(event->pmu, typeof(*i915), pmu.base);
455         unsigned int bit = event_enabled_bit(event);
456         unsigned long flags;
457
458         spin_lock_irqsave(&i915->pmu.lock, flags);
459
460         /*
461          * Update the bitmask of enabled events and increment
462          * the event reference counter.
463          */
464         GEM_BUG_ON(bit >= I915_PMU_MASK_BITS);
465         GEM_BUG_ON(i915->pmu.enable_count[bit] == ~0);
466         i915->pmu.enable |= BIT_ULL(bit);
467         i915->pmu.enable_count[bit]++;
468
469         /*
470          * Start the sampling timer if needed and not already enabled.
471          */
472         __i915_pmu_maybe_start_timer(i915);
473
474         /*
475          * For per-engine events the bitmask and reference counting
476          * is stored per engine.
477          */
478         if (is_engine_event(event)) {
479                 u8 sample = engine_event_sample(event);
480                 struct intel_engine_cs *engine;
481
482                 engine = intel_engine_lookup_user(i915,
483                                                   engine_event_class(event),
484                                                   engine_event_instance(event));
485                 GEM_BUG_ON(!engine);
486                 engine->pmu.enable |= BIT(sample);
487
488                 GEM_BUG_ON(sample >= I915_PMU_SAMPLE_BITS);
489                 GEM_BUG_ON(engine->pmu.enable_count[sample] == ~0);
490                 if (engine->pmu.enable_count[sample]++ == 0) {
491                         /*
492                          * Enable engine busy stats tracking if needed or
493                          * alternatively cancel the scheduled disable.
494                          *
495                          * If the delayed disable was pending, cancel it and
496                          * in this case do not enable since it already is.
497                          */
498                         if (engine_needs_busy_stats(engine) &&
499                             !engine->pmu.busy_stats) {
500                                 engine->pmu.busy_stats = true;
501                                 if (!cancel_delayed_work(&engine->pmu.disable_busy_stats))
502                                         intel_enable_engine_stats(engine);
503                         }
504                 }
505         }
506
507         /*
508          * Store the current counter value so we can report the correct delta
509          * for all listeners. Even when the event was already enabled and has
510          * an existing non-zero value.
511          */
512         local64_set(&event->hw.prev_count, __i915_pmu_event_read(event));
513
514         spin_unlock_irqrestore(&i915->pmu.lock, flags);
515 }
516
517 static void __disable_busy_stats(struct work_struct *work)
518 {
519         struct intel_engine_cs *engine =
520                container_of(work, typeof(*engine), pmu.disable_busy_stats.work);
521
522         intel_disable_engine_stats(engine);
523 }
524
525 static void i915_pmu_disable(struct perf_event *event)
526 {
527         struct drm_i915_private *i915 =
528                 container_of(event->pmu, typeof(*i915), pmu.base);
529         unsigned int bit = event_enabled_bit(event);
530         unsigned long flags;
531
532         spin_lock_irqsave(&i915->pmu.lock, flags);
533
534         if (is_engine_event(event)) {
535                 u8 sample = engine_event_sample(event);
536                 struct intel_engine_cs *engine;
537
538                 engine = intel_engine_lookup_user(i915,
539                                                   engine_event_class(event),
540                                                   engine_event_instance(event));
541                 GEM_BUG_ON(!engine);
542                 GEM_BUG_ON(sample >= I915_PMU_SAMPLE_BITS);
543                 GEM_BUG_ON(engine->pmu.enable_count[sample] == 0);
544                 /*
545                  * Decrement the reference count and clear the enabled
546                  * bitmask when the last listener on an event goes away.
547                  */
548                 if (--engine->pmu.enable_count[sample] == 0) {
549                         engine->pmu.enable &= ~BIT(sample);
550                         if (!engine_needs_busy_stats(engine) &&
551                             engine->pmu.busy_stats) {
552                                 engine->pmu.busy_stats = false;
553                                 /*
554                                  * We request a delayed disable to handle the
555                                  * rapid on/off cycles on events, which can
556                                  * happen when tools like perf stat start, in a
557                                  * nicer way.
558                                  *
559                                  * In addition, this also helps with busy stats
560                                  * accuracy with background CPU offline/online
561                                  * migration events.
562                                  */
563                                 queue_delayed_work(system_wq,
564                                                    &engine->pmu.disable_busy_stats,
565                                                    round_jiffies_up_relative(HZ));
566                         }
567                 }
568         }
569
570         GEM_BUG_ON(bit >= I915_PMU_MASK_BITS);
571         GEM_BUG_ON(i915->pmu.enable_count[bit] == 0);
572         /*
573          * Decrement the reference count and clear the enabled
574          * bitmask when the last listener on an event goes away.
575          */
576         if (--i915->pmu.enable_count[bit] == 0) {
577                 i915->pmu.enable &= ~BIT_ULL(bit);
578                 i915->pmu.timer_enabled &= pmu_needs_timer(i915, true);
579         }
580
581         spin_unlock_irqrestore(&i915->pmu.lock, flags);
582 }
583
584 static void i915_pmu_event_start(struct perf_event *event, int flags)
585 {
586         i915_pmu_enable(event);
587         event->hw.state = 0;
588 }
589
590 static void i915_pmu_event_stop(struct perf_event *event, int flags)
591 {
592         if (flags & PERF_EF_UPDATE)
593                 i915_pmu_event_read(event);
594         i915_pmu_disable(event);
595         event->hw.state = PERF_HES_STOPPED;
596 }
597
598 static int i915_pmu_event_add(struct perf_event *event, int flags)
599 {
600         if (flags & PERF_EF_START)
601                 i915_pmu_event_start(event, flags);
602
603         return 0;
604 }
605
606 static void i915_pmu_event_del(struct perf_event *event, int flags)
607 {
608         i915_pmu_event_stop(event, PERF_EF_UPDATE);
609 }
610
611 static int i915_pmu_event_event_idx(struct perf_event *event)
612 {
613         return 0;
614 }
615
616 struct i915_str_attribute {
617         struct device_attribute attr;
618         const char *str;
619 };
620
621 static ssize_t i915_pmu_format_show(struct device *dev,
622                                     struct device_attribute *attr, char *buf)
623 {
624         struct i915_str_attribute *eattr;
625
626         eattr = container_of(attr, struct i915_str_attribute, attr);
627         return sprintf(buf, "%s\n", eattr->str);
628 }
629
630 #define I915_PMU_FORMAT_ATTR(_name, _config) \
631         (&((struct i915_str_attribute[]) { \
632                 { .attr = __ATTR(_name, 0444, i915_pmu_format_show, NULL), \
633                   .str = _config, } \
634         })[0].attr.attr)
635
636 static struct attribute *i915_pmu_format_attrs[] = {
637         I915_PMU_FORMAT_ATTR(i915_eventid, "config:0-20"),
638         NULL,
639 };
640
641 static const struct attribute_group i915_pmu_format_attr_group = {
642         .name = "format",
643         .attrs = i915_pmu_format_attrs,
644 };
645
646 struct i915_ext_attribute {
647         struct device_attribute attr;
648         unsigned long val;
649 };
650
651 static ssize_t i915_pmu_event_show(struct device *dev,
652                                    struct device_attribute *attr, char *buf)
653 {
654         struct i915_ext_attribute *eattr;
655
656         eattr = container_of(attr, struct i915_ext_attribute, attr);
657         return sprintf(buf, "config=0x%lx\n", eattr->val);
658 }
659
660 #define I915_EVENT_ATTR(_name, _config) \
661         (&((struct i915_ext_attribute[]) { \
662                 { .attr = __ATTR(_name, 0444, i915_pmu_event_show, NULL), \
663                   .val = _config, } \
664         })[0].attr.attr)
665
666 #define I915_EVENT_STR(_name, _str) \
667         (&((struct perf_pmu_events_attr[]) { \
668                 { .attr      = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
669                   .id        = 0, \
670                   .event_str = _str, } \
671         })[0].attr.attr)
672
673 #define I915_EVENT(_name, _config, _unit) \
674         I915_EVENT_ATTR(_name, _config), \
675         I915_EVENT_STR(_name.unit, _unit)
676
677 #define I915_ENGINE_EVENT(_name, _class, _instance, _sample) \
678         I915_EVENT_ATTR(_name, __I915_PMU_ENGINE(_class, _instance, _sample)), \
679         I915_EVENT_STR(_name.unit, "ns")
680
681 #define I915_ENGINE_EVENTS(_name, _class, _instance) \
682         I915_ENGINE_EVENT(_name##_instance-busy, _class, _instance, I915_SAMPLE_BUSY), \
683         I915_ENGINE_EVENT(_name##_instance-sema, _class, _instance, I915_SAMPLE_SEMA), \
684         I915_ENGINE_EVENT(_name##_instance-wait, _class, _instance, I915_SAMPLE_WAIT)
685
686 static struct attribute *i915_pmu_events_attrs[] = {
687         I915_ENGINE_EVENTS(rcs, I915_ENGINE_CLASS_RENDER, 0),
688         I915_ENGINE_EVENTS(bcs, I915_ENGINE_CLASS_COPY, 0),
689         I915_ENGINE_EVENTS(vcs, I915_ENGINE_CLASS_VIDEO, 0),
690         I915_ENGINE_EVENTS(vcs, I915_ENGINE_CLASS_VIDEO, 1),
691         I915_ENGINE_EVENTS(vecs, I915_ENGINE_CLASS_VIDEO_ENHANCE, 0),
692
693         I915_EVENT(actual-frequency,    I915_PMU_ACTUAL_FREQUENCY,    "MHz"),
694         I915_EVENT(requested-frequency, I915_PMU_REQUESTED_FREQUENCY, "MHz"),
695
696         I915_EVENT_ATTR(interrupts, I915_PMU_INTERRUPTS),
697
698         I915_EVENT(rc6-residency,   I915_PMU_RC6_RESIDENCY,   "ns"),
699
700         NULL,
701 };
702
703 static const struct attribute_group i915_pmu_events_attr_group = {
704         .name = "events",
705         .attrs = i915_pmu_events_attrs,
706 };
707
708 static ssize_t
709 i915_pmu_get_attr_cpumask(struct device *dev,
710                           struct device_attribute *attr,
711                           char *buf)
712 {
713         return cpumap_print_to_pagebuf(true, buf, &i915_pmu_cpumask);
714 }
715
716 static DEVICE_ATTR(cpumask, 0444, i915_pmu_get_attr_cpumask, NULL);
717
718 static struct attribute *i915_cpumask_attrs[] = {
719         &dev_attr_cpumask.attr,
720         NULL,
721 };
722
723 static struct attribute_group i915_pmu_cpumask_attr_group = {
724         .attrs = i915_cpumask_attrs,
725 };
726
727 static const struct attribute_group *i915_pmu_attr_groups[] = {
728         &i915_pmu_format_attr_group,
729         &i915_pmu_events_attr_group,
730         &i915_pmu_cpumask_attr_group,
731         NULL
732 };
733
734 static int i915_pmu_cpu_online(unsigned int cpu, struct hlist_node *node)
735 {
736         struct i915_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), node);
737
738         GEM_BUG_ON(!pmu->base.event_init);
739
740         /* Select the first online CPU as a designated reader. */
741         if (!cpumask_weight(&i915_pmu_cpumask))
742                 cpumask_set_cpu(cpu, &i915_pmu_cpumask);
743
744         return 0;
745 }
746
747 static int i915_pmu_cpu_offline(unsigned int cpu, struct hlist_node *node)
748 {
749         struct i915_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), node);
750         unsigned int target;
751
752         GEM_BUG_ON(!pmu->base.event_init);
753
754         if (cpumask_test_and_clear_cpu(cpu, &i915_pmu_cpumask)) {
755                 target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu);
756                 /* Migrate events if there is a valid target */
757                 if (target < nr_cpu_ids) {
758                         cpumask_set_cpu(target, &i915_pmu_cpumask);
759                         perf_pmu_migrate_context(&pmu->base, cpu, target);
760                 }
761         }
762
763         return 0;
764 }
765
766 static enum cpuhp_state cpuhp_slot = CPUHP_INVALID;
767
768 static int i915_pmu_register_cpuhp_state(struct drm_i915_private *i915)
769 {
770         enum cpuhp_state slot;
771         int ret;
772
773         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
774                                       "perf/x86/intel/i915:online",
775                                       i915_pmu_cpu_online,
776                                       i915_pmu_cpu_offline);
777         if (ret < 0)
778                 return ret;
779
780         slot = ret;
781         ret = cpuhp_state_add_instance(slot, &i915->pmu.node);
782         if (ret) {
783                 cpuhp_remove_multi_state(slot);
784                 return ret;
785         }
786
787         cpuhp_slot = slot;
788         return 0;
789 }
790
791 static void i915_pmu_unregister_cpuhp_state(struct drm_i915_private *i915)
792 {
793         WARN_ON(cpuhp_slot == CPUHP_INVALID);
794         WARN_ON(cpuhp_state_remove_instance(cpuhp_slot, &i915->pmu.node));
795         cpuhp_remove_multi_state(cpuhp_slot);
796 }
797
798 void i915_pmu_register(struct drm_i915_private *i915)
799 {
800         struct intel_engine_cs *engine;
801         enum intel_engine_id id;
802         int ret;
803
804         if (INTEL_GEN(i915) <= 2) {
805                 DRM_INFO("PMU not supported for this GPU.");
806                 return;
807         }
808
809         i915->pmu.base.attr_groups      = i915_pmu_attr_groups;
810         i915->pmu.base.task_ctx_nr      = perf_invalid_context;
811         i915->pmu.base.event_init       = i915_pmu_event_init;
812         i915->pmu.base.add              = i915_pmu_event_add;
813         i915->pmu.base.del              = i915_pmu_event_del;
814         i915->pmu.base.start            = i915_pmu_event_start;
815         i915->pmu.base.stop             = i915_pmu_event_stop;
816         i915->pmu.base.read             = i915_pmu_event_read;
817         i915->pmu.base.event_idx        = i915_pmu_event_event_idx;
818
819         spin_lock_init(&i915->pmu.lock);
820         hrtimer_init(&i915->pmu.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
821         i915->pmu.timer.function = i915_sample;
822
823         for_each_engine(engine, i915, id)
824                 INIT_DELAYED_WORK(&engine->pmu.disable_busy_stats,
825                                   __disable_busy_stats);
826
827         ret = perf_pmu_register(&i915->pmu.base, "i915", -1);
828         if (ret)
829                 goto err;
830
831         ret = i915_pmu_register_cpuhp_state(i915);
832         if (ret)
833                 goto err_unreg;
834
835         return;
836
837 err_unreg:
838         perf_pmu_unregister(&i915->pmu.base);
839 err:
840         i915->pmu.base.event_init = NULL;
841         DRM_NOTE("Failed to register PMU! (err=%d)\n", ret);
842 }
843
844 void i915_pmu_unregister(struct drm_i915_private *i915)
845 {
846         struct intel_engine_cs *engine;
847         enum intel_engine_id id;
848
849         if (!i915->pmu.base.event_init)
850                 return;
851
852         WARN_ON(i915->pmu.enable);
853
854         hrtimer_cancel(&i915->pmu.timer);
855
856         for_each_engine(engine, i915, id) {
857                 GEM_BUG_ON(engine->pmu.busy_stats);
858                 flush_delayed_work(&engine->pmu.disable_busy_stats);
859         }
860
861         i915_pmu_unregister_cpuhp_state(i915);
862
863         perf_pmu_unregister(&i915->pmu.base);
864         i915->pmu.base.event_init = NULL;
865 }