Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm...
[deliverable/linux.git] / arch / x86 / events / intel / rapl.c
1 /*
2 * perf_event_intel_rapl.c: support Intel RAPL energy consumption counters
3 * Copyright (C) 2013 Google, Inc., Stephane Eranian
4 *
5 * Intel RAPL interface is specified in the IA-32 Manual Vol3b
6 * section 14.7.1 (September 2013)
7 *
8 * RAPL provides more controls than just reporting energy consumption
9 * however here we only expose the 3 energy consumption free running
10 * counters (pp0, pkg, dram).
11 *
12 * Each of those counters increments in a power unit defined by the
13 * RAPL_POWER_UNIT MSR. On SandyBridge, this unit is 1/(2^16) Joules
14 * but it can vary.
15 *
16 * Counter to rapl events mappings:
17 *
18 * pp0 counter: consumption of all physical cores (power plane 0)
19 * event: rapl_energy_cores
20 * perf code: 0x1
21 *
22 * pkg counter: consumption of the whole processor package
23 * event: rapl_energy_pkg
24 * perf code: 0x2
25 *
26 * dram counter: consumption of the dram domain (servers only)
27 * event: rapl_energy_dram
28 * perf code: 0x3
29 *
30 * dram counter: consumption of the builtin-gpu domain (client only)
31 * event: rapl_energy_gpu
32 * perf code: 0x4
33 *
34 * We manage those counters as free running (read-only). They may be
35 * use simultaneously by other tools, such as turbostat.
36 *
37 * The events only support system-wide mode counting. There is no
38 * sampling support because it does not make sense and is not
39 * supported by the RAPL hardware.
40 *
41 * Because we want to avoid floating-point operations in the kernel,
42 * the events are all reported in fixed point arithmetic (32.32).
43 * Tools must adjust the counts to convert them to Watts using
44 * the duration of the measurement. Tools may use a function such as
45 * ldexp(raw_count, -32);
46 */
47
48 #define pr_fmt(fmt) "RAPL PMU: " fmt
49
50 #include <linux/module.h>
51 #include <linux/slab.h>
52 #include <linux/perf_event.h>
53 #include <asm/cpu_device_id.h>
54 #include "../perf_event.h"
55
56 /*
57 * RAPL energy status counters
58 */
59 #define RAPL_IDX_PP0_NRG_STAT 0 /* all cores */
60 #define INTEL_RAPL_PP0 0x1 /* pseudo-encoding */
61 #define RAPL_IDX_PKG_NRG_STAT 1 /* entire package */
62 #define INTEL_RAPL_PKG 0x2 /* pseudo-encoding */
63 #define RAPL_IDX_RAM_NRG_STAT 2 /* DRAM */
64 #define INTEL_RAPL_RAM 0x3 /* pseudo-encoding */
65 #define RAPL_IDX_PP1_NRG_STAT 3 /* gpu */
66 #define INTEL_RAPL_PP1 0x4 /* pseudo-encoding */
67
68 #define NR_RAPL_DOMAINS 0x4
69 static const char *const rapl_domain_names[NR_RAPL_DOMAINS] __initconst = {
70 "pp0-core",
71 "package",
72 "dram",
73 "pp1-gpu",
74 };
75
76 /* Clients have PP0, PKG */
77 #define RAPL_IDX_CLN (1<<RAPL_IDX_PP0_NRG_STAT|\
78 1<<RAPL_IDX_PKG_NRG_STAT|\
79 1<<RAPL_IDX_PP1_NRG_STAT)
80
81 /* Servers have PP0, PKG, RAM */
82 #define RAPL_IDX_SRV (1<<RAPL_IDX_PP0_NRG_STAT|\
83 1<<RAPL_IDX_PKG_NRG_STAT|\
84 1<<RAPL_IDX_RAM_NRG_STAT)
85
86 /* Servers have PP0, PKG, RAM, PP1 */
87 #define RAPL_IDX_HSW (1<<RAPL_IDX_PP0_NRG_STAT|\
88 1<<RAPL_IDX_PKG_NRG_STAT|\
89 1<<RAPL_IDX_RAM_NRG_STAT|\
90 1<<RAPL_IDX_PP1_NRG_STAT)
91
92 /* Knights Landing has PKG, RAM */
93 #define RAPL_IDX_KNL (1<<RAPL_IDX_PKG_NRG_STAT|\
94 1<<RAPL_IDX_RAM_NRG_STAT)
95
96 /*
97 * event code: LSB 8 bits, passed in attr->config
98 * any other bit is reserved
99 */
100 #define RAPL_EVENT_MASK 0xFFULL
101
102 #define DEFINE_RAPL_FORMAT_ATTR(_var, _name, _format) \
103 static ssize_t __rapl_##_var##_show(struct kobject *kobj, \
104 struct kobj_attribute *attr, \
105 char *page) \
106 { \
107 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
108 return sprintf(page, _format "\n"); \
109 } \
110 static struct kobj_attribute format_attr_##_var = \
111 __ATTR(_name, 0444, __rapl_##_var##_show, NULL)
112
113 #define RAPL_CNTR_WIDTH 32
114
115 #define RAPL_EVENT_ATTR_STR(_name, v, str) \
116 static struct perf_pmu_events_attr event_attr_##v = { \
117 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
118 .id = 0, \
119 .event_str = str, \
120 };
121
122 struct rapl_pmu {
123 raw_spinlock_t lock;
124 int n_active;
125 int cpu;
126 struct list_head active_list;
127 struct pmu *pmu;
128 ktime_t timer_interval;
129 struct hrtimer hrtimer;
130 };
131
132 struct rapl_pmus {
133 struct pmu pmu;
134 unsigned int maxpkg;
135 struct rapl_pmu *pmus[];
136 };
137
138 /* 1/2^hw_unit Joule */
139 static int rapl_hw_unit[NR_RAPL_DOMAINS] __read_mostly;
140 static struct rapl_pmus *rapl_pmus;
141 static cpumask_t rapl_cpu_mask;
142 static unsigned int rapl_cntr_mask;
143 static u64 rapl_timer_ms;
144
145 static inline struct rapl_pmu *cpu_to_rapl_pmu(unsigned int cpu)
146 {
147 return rapl_pmus->pmus[topology_logical_package_id(cpu)];
148 }
149
150 static inline u64 rapl_read_counter(struct perf_event *event)
151 {
152 u64 raw;
153 rdmsrl(event->hw.event_base, raw);
154 return raw;
155 }
156
157 static inline u64 rapl_scale(u64 v, int cfg)
158 {
159 if (cfg > NR_RAPL_DOMAINS) {
160 pr_warn("Invalid domain %d, failed to scale data\n", cfg);
161 return v;
162 }
163 /*
164 * scale delta to smallest unit (1/2^32)
165 * users must then scale back: count * 1/(1e9*2^32) to get Joules
166 * or use ldexp(count, -32).
167 * Watts = Joules/Time delta
168 */
169 return v << (32 - rapl_hw_unit[cfg - 1]);
170 }
171
172 static u64 rapl_event_update(struct perf_event *event)
173 {
174 struct hw_perf_event *hwc = &event->hw;
175 u64 prev_raw_count, new_raw_count;
176 s64 delta, sdelta;
177 int shift = RAPL_CNTR_WIDTH;
178
179 again:
180 prev_raw_count = local64_read(&hwc->prev_count);
181 rdmsrl(event->hw.event_base, new_raw_count);
182
183 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
184 new_raw_count) != prev_raw_count) {
185 cpu_relax();
186 goto again;
187 }
188
189 /*
190 * Now we have the new raw value and have updated the prev
191 * timestamp already. We can now calculate the elapsed delta
192 * (event-)time and add that to the generic event.
193 *
194 * Careful, not all hw sign-extends above the physical width
195 * of the count.
196 */
197 delta = (new_raw_count << shift) - (prev_raw_count << shift);
198 delta >>= shift;
199
200 sdelta = rapl_scale(delta, event->hw.config);
201
202 local64_add(sdelta, &event->count);
203
204 return new_raw_count;
205 }
206
207 static void rapl_start_hrtimer(struct rapl_pmu *pmu)
208 {
209 hrtimer_start(&pmu->hrtimer, pmu->timer_interval,
210 HRTIMER_MODE_REL_PINNED);
211 }
212
213 static enum hrtimer_restart rapl_hrtimer_handle(struct hrtimer *hrtimer)
214 {
215 struct rapl_pmu *pmu = container_of(hrtimer, struct rapl_pmu, hrtimer);
216 struct perf_event *event;
217 unsigned long flags;
218
219 if (!pmu->n_active)
220 return HRTIMER_NORESTART;
221
222 raw_spin_lock_irqsave(&pmu->lock, flags);
223
224 list_for_each_entry(event, &pmu->active_list, active_entry)
225 rapl_event_update(event);
226
227 raw_spin_unlock_irqrestore(&pmu->lock, flags);
228
229 hrtimer_forward_now(hrtimer, pmu->timer_interval);
230
231 return HRTIMER_RESTART;
232 }
233
234 static void rapl_hrtimer_init(struct rapl_pmu *pmu)
235 {
236 struct hrtimer *hr = &pmu->hrtimer;
237
238 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
239 hr->function = rapl_hrtimer_handle;
240 }
241
242 static void __rapl_pmu_event_start(struct rapl_pmu *pmu,
243 struct perf_event *event)
244 {
245 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
246 return;
247
248 event->hw.state = 0;
249
250 list_add_tail(&event->active_entry, &pmu->active_list);
251
252 local64_set(&event->hw.prev_count, rapl_read_counter(event));
253
254 pmu->n_active++;
255 if (pmu->n_active == 1)
256 rapl_start_hrtimer(pmu);
257 }
258
259 static void rapl_pmu_event_start(struct perf_event *event, int mode)
260 {
261 struct rapl_pmu *pmu = event->pmu_private;
262 unsigned long flags;
263
264 raw_spin_lock_irqsave(&pmu->lock, flags);
265 __rapl_pmu_event_start(pmu, event);
266 raw_spin_unlock_irqrestore(&pmu->lock, flags);
267 }
268
269 static void rapl_pmu_event_stop(struct perf_event *event, int mode)
270 {
271 struct rapl_pmu *pmu = event->pmu_private;
272 struct hw_perf_event *hwc = &event->hw;
273 unsigned long flags;
274
275 raw_spin_lock_irqsave(&pmu->lock, flags);
276
277 /* mark event as deactivated and stopped */
278 if (!(hwc->state & PERF_HES_STOPPED)) {
279 WARN_ON_ONCE(pmu->n_active <= 0);
280 pmu->n_active--;
281 if (pmu->n_active == 0)
282 hrtimer_cancel(&pmu->hrtimer);
283
284 list_del(&event->active_entry);
285
286 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
287 hwc->state |= PERF_HES_STOPPED;
288 }
289
290 /* check if update of sw counter is necessary */
291 if ((mode & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
292 /*
293 * Drain the remaining delta count out of a event
294 * that we are disabling:
295 */
296 rapl_event_update(event);
297 hwc->state |= PERF_HES_UPTODATE;
298 }
299
300 raw_spin_unlock_irqrestore(&pmu->lock, flags);
301 }
302
303 static int rapl_pmu_event_add(struct perf_event *event, int mode)
304 {
305 struct rapl_pmu *pmu = event->pmu_private;
306 struct hw_perf_event *hwc = &event->hw;
307 unsigned long flags;
308
309 raw_spin_lock_irqsave(&pmu->lock, flags);
310
311 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
312
313 if (mode & PERF_EF_START)
314 __rapl_pmu_event_start(pmu, event);
315
316 raw_spin_unlock_irqrestore(&pmu->lock, flags);
317
318 return 0;
319 }
320
321 static void rapl_pmu_event_del(struct perf_event *event, int flags)
322 {
323 rapl_pmu_event_stop(event, PERF_EF_UPDATE);
324 }
325
326 static int rapl_pmu_event_init(struct perf_event *event)
327 {
328 u64 cfg = event->attr.config & RAPL_EVENT_MASK;
329 int bit, msr, ret = 0;
330 struct rapl_pmu *pmu;
331
332 /* only look at RAPL events */
333 if (event->attr.type != rapl_pmus->pmu.type)
334 return -ENOENT;
335
336 /* check only supported bits are set */
337 if (event->attr.config & ~RAPL_EVENT_MASK)
338 return -EINVAL;
339
340 if (event->cpu < 0)
341 return -EINVAL;
342
343 /*
344 * check event is known (determines counter)
345 */
346 switch (cfg) {
347 case INTEL_RAPL_PP0:
348 bit = RAPL_IDX_PP0_NRG_STAT;
349 msr = MSR_PP0_ENERGY_STATUS;
350 break;
351 case INTEL_RAPL_PKG:
352 bit = RAPL_IDX_PKG_NRG_STAT;
353 msr = MSR_PKG_ENERGY_STATUS;
354 break;
355 case INTEL_RAPL_RAM:
356 bit = RAPL_IDX_RAM_NRG_STAT;
357 msr = MSR_DRAM_ENERGY_STATUS;
358 break;
359 case INTEL_RAPL_PP1:
360 bit = RAPL_IDX_PP1_NRG_STAT;
361 msr = MSR_PP1_ENERGY_STATUS;
362 break;
363 default:
364 return -EINVAL;
365 }
366 /* check event supported */
367 if (!(rapl_cntr_mask & (1 << bit)))
368 return -EINVAL;
369
370 /* unsupported modes and filters */
371 if (event->attr.exclude_user ||
372 event->attr.exclude_kernel ||
373 event->attr.exclude_hv ||
374 event->attr.exclude_idle ||
375 event->attr.exclude_host ||
376 event->attr.exclude_guest ||
377 event->attr.sample_period) /* no sampling */
378 return -EINVAL;
379
380 /* must be done before validate_group */
381 pmu = cpu_to_rapl_pmu(event->cpu);
382 event->cpu = pmu->cpu;
383 event->pmu_private = pmu;
384 event->hw.event_base = msr;
385 event->hw.config = cfg;
386 event->hw.idx = bit;
387
388 return ret;
389 }
390
391 static void rapl_pmu_event_read(struct perf_event *event)
392 {
393 rapl_event_update(event);
394 }
395
396 static ssize_t rapl_get_attr_cpumask(struct device *dev,
397 struct device_attribute *attr, char *buf)
398 {
399 return cpumap_print_to_pagebuf(true, buf, &rapl_cpu_mask);
400 }
401
402 static DEVICE_ATTR(cpumask, S_IRUGO, rapl_get_attr_cpumask, NULL);
403
404 static struct attribute *rapl_pmu_attrs[] = {
405 &dev_attr_cpumask.attr,
406 NULL,
407 };
408
409 static struct attribute_group rapl_pmu_attr_group = {
410 .attrs = rapl_pmu_attrs,
411 };
412
413 RAPL_EVENT_ATTR_STR(energy-cores, rapl_cores, "event=0x01");
414 RAPL_EVENT_ATTR_STR(energy-pkg , rapl_pkg, "event=0x02");
415 RAPL_EVENT_ATTR_STR(energy-ram , rapl_ram, "event=0x03");
416 RAPL_EVENT_ATTR_STR(energy-gpu , rapl_gpu, "event=0x04");
417
418 RAPL_EVENT_ATTR_STR(energy-cores.unit, rapl_cores_unit, "Joules");
419 RAPL_EVENT_ATTR_STR(energy-pkg.unit , rapl_pkg_unit, "Joules");
420 RAPL_EVENT_ATTR_STR(energy-ram.unit , rapl_ram_unit, "Joules");
421 RAPL_EVENT_ATTR_STR(energy-gpu.unit , rapl_gpu_unit, "Joules");
422
423 /*
424 * we compute in 0.23 nJ increments regardless of MSR
425 */
426 RAPL_EVENT_ATTR_STR(energy-cores.scale, rapl_cores_scale, "2.3283064365386962890625e-10");
427 RAPL_EVENT_ATTR_STR(energy-pkg.scale, rapl_pkg_scale, "2.3283064365386962890625e-10");
428 RAPL_EVENT_ATTR_STR(energy-ram.scale, rapl_ram_scale, "2.3283064365386962890625e-10");
429 RAPL_EVENT_ATTR_STR(energy-gpu.scale, rapl_gpu_scale, "2.3283064365386962890625e-10");
430
431 static struct attribute *rapl_events_srv_attr[] = {
432 EVENT_PTR(rapl_cores),
433 EVENT_PTR(rapl_pkg),
434 EVENT_PTR(rapl_ram),
435
436 EVENT_PTR(rapl_cores_unit),
437 EVENT_PTR(rapl_pkg_unit),
438 EVENT_PTR(rapl_ram_unit),
439
440 EVENT_PTR(rapl_cores_scale),
441 EVENT_PTR(rapl_pkg_scale),
442 EVENT_PTR(rapl_ram_scale),
443 NULL,
444 };
445
446 static struct attribute *rapl_events_cln_attr[] = {
447 EVENT_PTR(rapl_cores),
448 EVENT_PTR(rapl_pkg),
449 EVENT_PTR(rapl_gpu),
450
451 EVENT_PTR(rapl_cores_unit),
452 EVENT_PTR(rapl_pkg_unit),
453 EVENT_PTR(rapl_gpu_unit),
454
455 EVENT_PTR(rapl_cores_scale),
456 EVENT_PTR(rapl_pkg_scale),
457 EVENT_PTR(rapl_gpu_scale),
458 NULL,
459 };
460
461 static struct attribute *rapl_events_hsw_attr[] = {
462 EVENT_PTR(rapl_cores),
463 EVENT_PTR(rapl_pkg),
464 EVENT_PTR(rapl_gpu),
465 EVENT_PTR(rapl_ram),
466
467 EVENT_PTR(rapl_cores_unit),
468 EVENT_PTR(rapl_pkg_unit),
469 EVENT_PTR(rapl_gpu_unit),
470 EVENT_PTR(rapl_ram_unit),
471
472 EVENT_PTR(rapl_cores_scale),
473 EVENT_PTR(rapl_pkg_scale),
474 EVENT_PTR(rapl_gpu_scale),
475 EVENT_PTR(rapl_ram_scale),
476 NULL,
477 };
478
479 static struct attribute *rapl_events_knl_attr[] = {
480 EVENT_PTR(rapl_pkg),
481 EVENT_PTR(rapl_ram),
482
483 EVENT_PTR(rapl_pkg_unit),
484 EVENT_PTR(rapl_ram_unit),
485
486 EVENT_PTR(rapl_pkg_scale),
487 EVENT_PTR(rapl_ram_scale),
488 NULL,
489 };
490
491 static struct attribute_group rapl_pmu_events_group = {
492 .name = "events",
493 .attrs = NULL, /* patched at runtime */
494 };
495
496 DEFINE_RAPL_FORMAT_ATTR(event, event, "config:0-7");
497 static struct attribute *rapl_formats_attr[] = {
498 &format_attr_event.attr,
499 NULL,
500 };
501
502 static struct attribute_group rapl_pmu_format_group = {
503 .name = "format",
504 .attrs = rapl_formats_attr,
505 };
506
507 const struct attribute_group *rapl_attr_groups[] = {
508 &rapl_pmu_attr_group,
509 &rapl_pmu_format_group,
510 &rapl_pmu_events_group,
511 NULL,
512 };
513
514 static void rapl_cpu_exit(int cpu)
515 {
516 struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
517 int target;
518
519 /* Check if exiting cpu is used for collecting rapl events */
520 if (!cpumask_test_and_clear_cpu(cpu, &rapl_cpu_mask))
521 return;
522
523 pmu->cpu = -1;
524 /* Find a new cpu to collect rapl events */
525 target = cpumask_any_but(topology_core_cpumask(cpu), cpu);
526
527 /* Migrate rapl events to the new target */
528 if (target < nr_cpu_ids) {
529 cpumask_set_cpu(target, &rapl_cpu_mask);
530 pmu->cpu = target;
531 perf_pmu_migrate_context(pmu->pmu, cpu, target);
532 }
533 }
534
535 static void rapl_cpu_init(int cpu)
536 {
537 struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
538 int target;
539
540 /*
541 * Check if there is an online cpu in the package which collects rapl
542 * events already.
543 */
544 target = cpumask_any_and(&rapl_cpu_mask, topology_core_cpumask(cpu));
545 if (target < nr_cpu_ids)
546 return;
547
548 cpumask_set_cpu(cpu, &rapl_cpu_mask);
549 pmu->cpu = cpu;
550 }
551
552 static int rapl_cpu_prepare(int cpu)
553 {
554 struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
555
556 if (pmu)
557 return 0;
558
559 pmu = kzalloc_node(sizeof(*pmu), GFP_KERNEL, cpu_to_node(cpu));
560 if (!pmu)
561 return -ENOMEM;
562
563 raw_spin_lock_init(&pmu->lock);
564 INIT_LIST_HEAD(&pmu->active_list);
565 pmu->pmu = &rapl_pmus->pmu;
566 pmu->timer_interval = ms_to_ktime(rapl_timer_ms);
567 pmu->cpu = -1;
568 rapl_hrtimer_init(pmu);
569 rapl_pmus->pmus[topology_logical_package_id(cpu)] = pmu;
570 return 0;
571 }
572
573 static int rapl_cpu_notifier(struct notifier_block *self,
574 unsigned long action, void *hcpu)
575 {
576 unsigned int cpu = (long)hcpu;
577
578 switch (action & ~CPU_TASKS_FROZEN) {
579 case CPU_UP_PREPARE:
580 rapl_cpu_prepare(cpu);
581 break;
582
583 case CPU_DOWN_FAILED:
584 case CPU_ONLINE:
585 rapl_cpu_init(cpu);
586 break;
587
588 case CPU_DOWN_PREPARE:
589 rapl_cpu_exit(cpu);
590 break;
591 }
592 return NOTIFY_OK;
593 }
594
595 static int rapl_check_hw_unit(bool apply_quirk)
596 {
597 u64 msr_rapl_power_unit_bits;
598 int i;
599
600 /* protect rdmsrl() to handle virtualization */
601 if (rdmsrl_safe(MSR_RAPL_POWER_UNIT, &msr_rapl_power_unit_bits))
602 return -1;
603 for (i = 0; i < NR_RAPL_DOMAINS; i++)
604 rapl_hw_unit[i] = (msr_rapl_power_unit_bits >> 8) & 0x1FULL;
605
606 /*
607 * DRAM domain on HSW server and KNL has fixed energy unit which can be
608 * different than the unit from power unit MSR. See
609 * "Intel Xeon Processor E5-1600 and E5-2600 v3 Product Families, V2
610 * of 2. Datasheet, September 2014, Reference Number: 330784-001 "
611 */
612 if (apply_quirk)
613 rapl_hw_unit[RAPL_IDX_RAM_NRG_STAT] = 16;
614
615 /*
616 * Calculate the timer rate:
617 * Use reference of 200W for scaling the timeout to avoid counter
618 * overflows. 200W = 200 Joules/sec
619 * Divide interval by 2 to avoid lockstep (2 * 100)
620 * if hw unit is 32, then we use 2 ms 1/200/2
621 */
622 rapl_timer_ms = 2;
623 if (rapl_hw_unit[0] < 32) {
624 rapl_timer_ms = (1000 / (2 * 100));
625 rapl_timer_ms *= (1ULL << (32 - rapl_hw_unit[0] - 1));
626 }
627 return 0;
628 }
629
630 static void __init rapl_advertise(void)
631 {
632 int i;
633
634 pr_info("API unit is 2^-32 Joules, %d fixed counters, %llu ms ovfl timer\n",
635 hweight32(rapl_cntr_mask), rapl_timer_ms);
636
637 for (i = 0; i < NR_RAPL_DOMAINS; i++) {
638 if (rapl_cntr_mask & (1 << i)) {
639 pr_info("hw unit of domain %s 2^-%d Joules\n",
640 rapl_domain_names[i], rapl_hw_unit[i]);
641 }
642 }
643 }
644
645 static int __init rapl_prepare_cpus(void)
646 {
647 unsigned int cpu, pkg;
648 int ret;
649
650 for_each_online_cpu(cpu) {
651 pkg = topology_logical_package_id(cpu);
652 if (rapl_pmus->pmus[pkg])
653 continue;
654
655 ret = rapl_cpu_prepare(cpu);
656 if (ret)
657 return ret;
658 rapl_cpu_init(cpu);
659 }
660 return 0;
661 }
662
663 static void __init cleanup_rapl_pmus(void)
664 {
665 int i;
666
667 for (i = 0; i < rapl_pmus->maxpkg; i++)
668 kfree(rapl_pmus->pmus + i);
669 kfree(rapl_pmus);
670 }
671
672 static int __init init_rapl_pmus(void)
673 {
674 int maxpkg = topology_max_packages();
675 size_t size;
676
677 size = sizeof(*rapl_pmus) + maxpkg * sizeof(struct rapl_pmu *);
678 rapl_pmus = kzalloc(size, GFP_KERNEL);
679 if (!rapl_pmus)
680 return -ENOMEM;
681
682 rapl_pmus->maxpkg = maxpkg;
683 rapl_pmus->pmu.attr_groups = rapl_attr_groups;
684 rapl_pmus->pmu.task_ctx_nr = perf_invalid_context;
685 rapl_pmus->pmu.event_init = rapl_pmu_event_init;
686 rapl_pmus->pmu.add = rapl_pmu_event_add;
687 rapl_pmus->pmu.del = rapl_pmu_event_del;
688 rapl_pmus->pmu.start = rapl_pmu_event_start;
689 rapl_pmus->pmu.stop = rapl_pmu_event_stop;
690 rapl_pmus->pmu.read = rapl_pmu_event_read;
691 return 0;
692 }
693
694 static const struct x86_cpu_id rapl_cpu_match[] __initconst = {
695 [0] = { .vendor = X86_VENDOR_INTEL, .family = 6 },
696 [1] = {},
697 };
698
699 static int __init rapl_pmu_init(void)
700 {
701 bool apply_quirk = false;
702 int ret;
703
704 if (!x86_match_cpu(rapl_cpu_match))
705 return -ENODEV;
706
707 switch (boot_cpu_data.x86_model) {
708 case 42: /* Sandy Bridge */
709 case 58: /* Ivy Bridge */
710 rapl_cntr_mask = RAPL_IDX_CLN;
711 rapl_pmu_events_group.attrs = rapl_events_cln_attr;
712 break;
713 case 63: /* Haswell-Server */
714 case 79: /* Broadwell-Server */
715 apply_quirk = true;
716 rapl_cntr_mask = RAPL_IDX_SRV;
717 rapl_pmu_events_group.attrs = rapl_events_srv_attr;
718 break;
719 case 60: /* Haswell */
720 case 69: /* Haswell-Celeron */
721 case 70: /* Haswell GT3e */
722 case 61: /* Broadwell */
723 case 71: /* Broadwell-H */
724 rapl_cntr_mask = RAPL_IDX_HSW;
725 rapl_pmu_events_group.attrs = rapl_events_hsw_attr;
726 break;
727 case 45: /* Sandy Bridge-EP */
728 case 62: /* IvyTown */
729 rapl_cntr_mask = RAPL_IDX_SRV;
730 rapl_pmu_events_group.attrs = rapl_events_srv_attr;
731 break;
732 case 87: /* Knights Landing */
733 apply_quirk = true;
734 rapl_cntr_mask = RAPL_IDX_KNL;
735 rapl_pmu_events_group.attrs = rapl_events_knl_attr;
736 break;
737 default:
738 return -ENODEV;
739 }
740
741 ret = rapl_check_hw_unit(apply_quirk);
742 if (ret)
743 return ret;
744
745 ret = init_rapl_pmus();
746 if (ret)
747 return ret;
748
749 cpu_notifier_register_begin();
750
751 ret = rapl_prepare_cpus();
752 if (ret)
753 goto out;
754
755 ret = perf_pmu_register(&rapl_pmus->pmu, "power", -1);
756 if (ret)
757 goto out;
758
759 __perf_cpu_notifier(rapl_cpu_notifier);
760 cpu_notifier_register_done();
761 rapl_advertise();
762 return 0;
763
764 out:
765 pr_warn("Initialization failed (%d), disabled\n", ret);
766 cleanup_rapl_pmus();
767 cpu_notifier_register_done();
768 return ret;
769 }
770 device_initcall(rapl_pmu_init);
This page took 0.062122 seconds and 5 git commands to generate.