]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - arch/x86/events/intel/rapl.c
Merge tag 'pnp-4.8-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm
[mirror_ubuntu-artful-kernel.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 * gpu counter: consumption of the builtin-gpu domain (client only)
31 * event: rapl_energy_gpu
32 * perf code: 0x4
33 *
34 * psys counter: consumption of the builtin-psys domain (client only)
35 * event: rapl_energy_psys
36 * perf code: 0x5
37 *
38 * We manage those counters as free running (read-only). They may be
39 * use simultaneously by other tools, such as turbostat.
40 *
41 * The events only support system-wide mode counting. There is no
42 * sampling support because it does not make sense and is not
43 * supported by the RAPL hardware.
44 *
45 * Because we want to avoid floating-point operations in the kernel,
46 * the events are all reported in fixed point arithmetic (32.32).
47 * Tools must adjust the counts to convert them to Watts using
48 * the duration of the measurement. Tools may use a function such as
49 * ldexp(raw_count, -32);
50 */
51
52 #define pr_fmt(fmt) "RAPL PMU: " fmt
53
54 #include <linux/module.h>
55 #include <linux/slab.h>
56 #include <linux/perf_event.h>
57 #include <asm/cpu_device_id.h>
58 #include <asm/intel-family.h>
59 #include "../perf_event.h"
60
61 MODULE_LICENSE("GPL");
62
63 /*
64 * RAPL energy status counters
65 */
66 #define RAPL_IDX_PP0_NRG_STAT 0 /* all cores */
67 #define INTEL_RAPL_PP0 0x1 /* pseudo-encoding */
68 #define RAPL_IDX_PKG_NRG_STAT 1 /* entire package */
69 #define INTEL_RAPL_PKG 0x2 /* pseudo-encoding */
70 #define RAPL_IDX_RAM_NRG_STAT 2 /* DRAM */
71 #define INTEL_RAPL_RAM 0x3 /* pseudo-encoding */
72 #define RAPL_IDX_PP1_NRG_STAT 3 /* gpu */
73 #define INTEL_RAPL_PP1 0x4 /* pseudo-encoding */
74 #define RAPL_IDX_PSYS_NRG_STAT 4 /* psys */
75 #define INTEL_RAPL_PSYS 0x5 /* pseudo-encoding */
76
77 #define NR_RAPL_DOMAINS 0x5
78 static const char *const rapl_domain_names[NR_RAPL_DOMAINS] __initconst = {
79 "pp0-core",
80 "package",
81 "dram",
82 "pp1-gpu",
83 "psys",
84 };
85
86 /* Clients have PP0, PKG */
87 #define RAPL_IDX_CLN (1<<RAPL_IDX_PP0_NRG_STAT|\
88 1<<RAPL_IDX_PKG_NRG_STAT|\
89 1<<RAPL_IDX_PP1_NRG_STAT)
90
91 /* Servers have PP0, PKG, RAM */
92 #define RAPL_IDX_SRV (1<<RAPL_IDX_PP0_NRG_STAT|\
93 1<<RAPL_IDX_PKG_NRG_STAT|\
94 1<<RAPL_IDX_RAM_NRG_STAT)
95
96 /* Servers have PP0, PKG, RAM, PP1 */
97 #define RAPL_IDX_HSW (1<<RAPL_IDX_PP0_NRG_STAT|\
98 1<<RAPL_IDX_PKG_NRG_STAT|\
99 1<<RAPL_IDX_RAM_NRG_STAT|\
100 1<<RAPL_IDX_PP1_NRG_STAT)
101
102 /* SKL clients have PP0, PKG, RAM, PP1, PSYS */
103 #define RAPL_IDX_SKL_CLN (1<<RAPL_IDX_PP0_NRG_STAT|\
104 1<<RAPL_IDX_PKG_NRG_STAT|\
105 1<<RAPL_IDX_RAM_NRG_STAT|\
106 1<<RAPL_IDX_PP1_NRG_STAT|\
107 1<<RAPL_IDX_PSYS_NRG_STAT)
108
109 /* Knights Landing has PKG, RAM */
110 #define RAPL_IDX_KNL (1<<RAPL_IDX_PKG_NRG_STAT|\
111 1<<RAPL_IDX_RAM_NRG_STAT)
112
113 /*
114 * event code: LSB 8 bits, passed in attr->config
115 * any other bit is reserved
116 */
117 #define RAPL_EVENT_MASK 0xFFULL
118
119 #define DEFINE_RAPL_FORMAT_ATTR(_var, _name, _format) \
120 static ssize_t __rapl_##_var##_show(struct kobject *kobj, \
121 struct kobj_attribute *attr, \
122 char *page) \
123 { \
124 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
125 return sprintf(page, _format "\n"); \
126 } \
127 static struct kobj_attribute format_attr_##_var = \
128 __ATTR(_name, 0444, __rapl_##_var##_show, NULL)
129
130 #define RAPL_CNTR_WIDTH 32
131
132 #define RAPL_EVENT_ATTR_STR(_name, v, str) \
133 static struct perf_pmu_events_attr event_attr_##v = { \
134 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
135 .id = 0, \
136 .event_str = str, \
137 };
138
139 struct rapl_pmu {
140 raw_spinlock_t lock;
141 int n_active;
142 int cpu;
143 struct list_head active_list;
144 struct pmu *pmu;
145 ktime_t timer_interval;
146 struct hrtimer hrtimer;
147 };
148
149 struct rapl_pmus {
150 struct pmu pmu;
151 unsigned int maxpkg;
152 struct rapl_pmu *pmus[];
153 };
154
155 /* 1/2^hw_unit Joule */
156 static int rapl_hw_unit[NR_RAPL_DOMAINS] __read_mostly;
157 static struct rapl_pmus *rapl_pmus;
158 static cpumask_t rapl_cpu_mask;
159 static unsigned int rapl_cntr_mask;
160 static u64 rapl_timer_ms;
161
162 static inline struct rapl_pmu *cpu_to_rapl_pmu(unsigned int cpu)
163 {
164 return rapl_pmus->pmus[topology_logical_package_id(cpu)];
165 }
166
167 static inline u64 rapl_read_counter(struct perf_event *event)
168 {
169 u64 raw;
170 rdmsrl(event->hw.event_base, raw);
171 return raw;
172 }
173
174 static inline u64 rapl_scale(u64 v, int cfg)
175 {
176 if (cfg > NR_RAPL_DOMAINS) {
177 pr_warn("Invalid domain %d, failed to scale data\n", cfg);
178 return v;
179 }
180 /*
181 * scale delta to smallest unit (1/2^32)
182 * users must then scale back: count * 1/(1e9*2^32) to get Joules
183 * or use ldexp(count, -32).
184 * Watts = Joules/Time delta
185 */
186 return v << (32 - rapl_hw_unit[cfg - 1]);
187 }
188
189 static u64 rapl_event_update(struct perf_event *event)
190 {
191 struct hw_perf_event *hwc = &event->hw;
192 u64 prev_raw_count, new_raw_count;
193 s64 delta, sdelta;
194 int shift = RAPL_CNTR_WIDTH;
195
196 again:
197 prev_raw_count = local64_read(&hwc->prev_count);
198 rdmsrl(event->hw.event_base, new_raw_count);
199
200 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
201 new_raw_count) != prev_raw_count) {
202 cpu_relax();
203 goto again;
204 }
205
206 /*
207 * Now we have the new raw value and have updated the prev
208 * timestamp already. We can now calculate the elapsed delta
209 * (event-)time and add that to the generic event.
210 *
211 * Careful, not all hw sign-extends above the physical width
212 * of the count.
213 */
214 delta = (new_raw_count << shift) - (prev_raw_count << shift);
215 delta >>= shift;
216
217 sdelta = rapl_scale(delta, event->hw.config);
218
219 local64_add(sdelta, &event->count);
220
221 return new_raw_count;
222 }
223
224 static void rapl_start_hrtimer(struct rapl_pmu *pmu)
225 {
226 hrtimer_start(&pmu->hrtimer, pmu->timer_interval,
227 HRTIMER_MODE_REL_PINNED);
228 }
229
230 static enum hrtimer_restart rapl_hrtimer_handle(struct hrtimer *hrtimer)
231 {
232 struct rapl_pmu *pmu = container_of(hrtimer, struct rapl_pmu, hrtimer);
233 struct perf_event *event;
234 unsigned long flags;
235
236 if (!pmu->n_active)
237 return HRTIMER_NORESTART;
238
239 raw_spin_lock_irqsave(&pmu->lock, flags);
240
241 list_for_each_entry(event, &pmu->active_list, active_entry)
242 rapl_event_update(event);
243
244 raw_spin_unlock_irqrestore(&pmu->lock, flags);
245
246 hrtimer_forward_now(hrtimer, pmu->timer_interval);
247
248 return HRTIMER_RESTART;
249 }
250
251 static void rapl_hrtimer_init(struct rapl_pmu *pmu)
252 {
253 struct hrtimer *hr = &pmu->hrtimer;
254
255 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
256 hr->function = rapl_hrtimer_handle;
257 }
258
259 static void __rapl_pmu_event_start(struct rapl_pmu *pmu,
260 struct perf_event *event)
261 {
262 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
263 return;
264
265 event->hw.state = 0;
266
267 list_add_tail(&event->active_entry, &pmu->active_list);
268
269 local64_set(&event->hw.prev_count, rapl_read_counter(event));
270
271 pmu->n_active++;
272 if (pmu->n_active == 1)
273 rapl_start_hrtimer(pmu);
274 }
275
276 static void rapl_pmu_event_start(struct perf_event *event, int mode)
277 {
278 struct rapl_pmu *pmu = event->pmu_private;
279 unsigned long flags;
280
281 raw_spin_lock_irqsave(&pmu->lock, flags);
282 __rapl_pmu_event_start(pmu, event);
283 raw_spin_unlock_irqrestore(&pmu->lock, flags);
284 }
285
286 static void rapl_pmu_event_stop(struct perf_event *event, int mode)
287 {
288 struct rapl_pmu *pmu = event->pmu_private;
289 struct hw_perf_event *hwc = &event->hw;
290 unsigned long flags;
291
292 raw_spin_lock_irqsave(&pmu->lock, flags);
293
294 /* mark event as deactivated and stopped */
295 if (!(hwc->state & PERF_HES_STOPPED)) {
296 WARN_ON_ONCE(pmu->n_active <= 0);
297 pmu->n_active--;
298 if (pmu->n_active == 0)
299 hrtimer_cancel(&pmu->hrtimer);
300
301 list_del(&event->active_entry);
302
303 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
304 hwc->state |= PERF_HES_STOPPED;
305 }
306
307 /* check if update of sw counter is necessary */
308 if ((mode & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
309 /*
310 * Drain the remaining delta count out of a event
311 * that we are disabling:
312 */
313 rapl_event_update(event);
314 hwc->state |= PERF_HES_UPTODATE;
315 }
316
317 raw_spin_unlock_irqrestore(&pmu->lock, flags);
318 }
319
320 static int rapl_pmu_event_add(struct perf_event *event, int mode)
321 {
322 struct rapl_pmu *pmu = event->pmu_private;
323 struct hw_perf_event *hwc = &event->hw;
324 unsigned long flags;
325
326 raw_spin_lock_irqsave(&pmu->lock, flags);
327
328 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
329
330 if (mode & PERF_EF_START)
331 __rapl_pmu_event_start(pmu, event);
332
333 raw_spin_unlock_irqrestore(&pmu->lock, flags);
334
335 return 0;
336 }
337
338 static void rapl_pmu_event_del(struct perf_event *event, int flags)
339 {
340 rapl_pmu_event_stop(event, PERF_EF_UPDATE);
341 }
342
343 static int rapl_pmu_event_init(struct perf_event *event)
344 {
345 u64 cfg = event->attr.config & RAPL_EVENT_MASK;
346 int bit, msr, ret = 0;
347 struct rapl_pmu *pmu;
348
349 /* only look at RAPL events */
350 if (event->attr.type != rapl_pmus->pmu.type)
351 return -ENOENT;
352
353 /* check only supported bits are set */
354 if (event->attr.config & ~RAPL_EVENT_MASK)
355 return -EINVAL;
356
357 if (event->cpu < 0)
358 return -EINVAL;
359
360 /*
361 * check event is known (determines counter)
362 */
363 switch (cfg) {
364 case INTEL_RAPL_PP0:
365 bit = RAPL_IDX_PP0_NRG_STAT;
366 msr = MSR_PP0_ENERGY_STATUS;
367 break;
368 case INTEL_RAPL_PKG:
369 bit = RAPL_IDX_PKG_NRG_STAT;
370 msr = MSR_PKG_ENERGY_STATUS;
371 break;
372 case INTEL_RAPL_RAM:
373 bit = RAPL_IDX_RAM_NRG_STAT;
374 msr = MSR_DRAM_ENERGY_STATUS;
375 break;
376 case INTEL_RAPL_PP1:
377 bit = RAPL_IDX_PP1_NRG_STAT;
378 msr = MSR_PP1_ENERGY_STATUS;
379 break;
380 case INTEL_RAPL_PSYS:
381 bit = RAPL_IDX_PSYS_NRG_STAT;
382 msr = MSR_PLATFORM_ENERGY_STATUS;
383 break;
384 default:
385 return -EINVAL;
386 }
387 /* check event supported */
388 if (!(rapl_cntr_mask & (1 << bit)))
389 return -EINVAL;
390
391 /* unsupported modes and filters */
392 if (event->attr.exclude_user ||
393 event->attr.exclude_kernel ||
394 event->attr.exclude_hv ||
395 event->attr.exclude_idle ||
396 event->attr.exclude_host ||
397 event->attr.exclude_guest ||
398 event->attr.sample_period) /* no sampling */
399 return -EINVAL;
400
401 /* must be done before validate_group */
402 pmu = cpu_to_rapl_pmu(event->cpu);
403 event->cpu = pmu->cpu;
404 event->pmu_private = pmu;
405 event->hw.event_base = msr;
406 event->hw.config = cfg;
407 event->hw.idx = bit;
408
409 return ret;
410 }
411
412 static void rapl_pmu_event_read(struct perf_event *event)
413 {
414 rapl_event_update(event);
415 }
416
417 static ssize_t rapl_get_attr_cpumask(struct device *dev,
418 struct device_attribute *attr, char *buf)
419 {
420 return cpumap_print_to_pagebuf(true, buf, &rapl_cpu_mask);
421 }
422
423 static DEVICE_ATTR(cpumask, S_IRUGO, rapl_get_attr_cpumask, NULL);
424
425 static struct attribute *rapl_pmu_attrs[] = {
426 &dev_attr_cpumask.attr,
427 NULL,
428 };
429
430 static struct attribute_group rapl_pmu_attr_group = {
431 .attrs = rapl_pmu_attrs,
432 };
433
434 RAPL_EVENT_ATTR_STR(energy-cores, rapl_cores, "event=0x01");
435 RAPL_EVENT_ATTR_STR(energy-pkg , rapl_pkg, "event=0x02");
436 RAPL_EVENT_ATTR_STR(energy-ram , rapl_ram, "event=0x03");
437 RAPL_EVENT_ATTR_STR(energy-gpu , rapl_gpu, "event=0x04");
438 RAPL_EVENT_ATTR_STR(energy-psys, rapl_psys, "event=0x05");
439
440 RAPL_EVENT_ATTR_STR(energy-cores.unit, rapl_cores_unit, "Joules");
441 RAPL_EVENT_ATTR_STR(energy-pkg.unit , rapl_pkg_unit, "Joules");
442 RAPL_EVENT_ATTR_STR(energy-ram.unit , rapl_ram_unit, "Joules");
443 RAPL_EVENT_ATTR_STR(energy-gpu.unit , rapl_gpu_unit, "Joules");
444 RAPL_EVENT_ATTR_STR(energy-psys.unit, rapl_psys_unit, "Joules");
445
446 /*
447 * we compute in 0.23 nJ increments regardless of MSR
448 */
449 RAPL_EVENT_ATTR_STR(energy-cores.scale, rapl_cores_scale, "2.3283064365386962890625e-10");
450 RAPL_EVENT_ATTR_STR(energy-pkg.scale, rapl_pkg_scale, "2.3283064365386962890625e-10");
451 RAPL_EVENT_ATTR_STR(energy-ram.scale, rapl_ram_scale, "2.3283064365386962890625e-10");
452 RAPL_EVENT_ATTR_STR(energy-gpu.scale, rapl_gpu_scale, "2.3283064365386962890625e-10");
453 RAPL_EVENT_ATTR_STR(energy-psys.scale, rapl_psys_scale, "2.3283064365386962890625e-10");
454
455 static struct attribute *rapl_events_srv_attr[] = {
456 EVENT_PTR(rapl_cores),
457 EVENT_PTR(rapl_pkg),
458 EVENT_PTR(rapl_ram),
459
460 EVENT_PTR(rapl_cores_unit),
461 EVENT_PTR(rapl_pkg_unit),
462 EVENT_PTR(rapl_ram_unit),
463
464 EVENT_PTR(rapl_cores_scale),
465 EVENT_PTR(rapl_pkg_scale),
466 EVENT_PTR(rapl_ram_scale),
467 NULL,
468 };
469
470 static struct attribute *rapl_events_cln_attr[] = {
471 EVENT_PTR(rapl_cores),
472 EVENT_PTR(rapl_pkg),
473 EVENT_PTR(rapl_gpu),
474
475 EVENT_PTR(rapl_cores_unit),
476 EVENT_PTR(rapl_pkg_unit),
477 EVENT_PTR(rapl_gpu_unit),
478
479 EVENT_PTR(rapl_cores_scale),
480 EVENT_PTR(rapl_pkg_scale),
481 EVENT_PTR(rapl_gpu_scale),
482 NULL,
483 };
484
485 static struct attribute *rapl_events_hsw_attr[] = {
486 EVENT_PTR(rapl_cores),
487 EVENT_PTR(rapl_pkg),
488 EVENT_PTR(rapl_gpu),
489 EVENT_PTR(rapl_ram),
490
491 EVENT_PTR(rapl_cores_unit),
492 EVENT_PTR(rapl_pkg_unit),
493 EVENT_PTR(rapl_gpu_unit),
494 EVENT_PTR(rapl_ram_unit),
495
496 EVENT_PTR(rapl_cores_scale),
497 EVENT_PTR(rapl_pkg_scale),
498 EVENT_PTR(rapl_gpu_scale),
499 EVENT_PTR(rapl_ram_scale),
500 NULL,
501 };
502
503 static struct attribute *rapl_events_skl_attr[] = {
504 EVENT_PTR(rapl_cores),
505 EVENT_PTR(rapl_pkg),
506 EVENT_PTR(rapl_gpu),
507 EVENT_PTR(rapl_ram),
508 EVENT_PTR(rapl_psys),
509
510 EVENT_PTR(rapl_cores_unit),
511 EVENT_PTR(rapl_pkg_unit),
512 EVENT_PTR(rapl_gpu_unit),
513 EVENT_PTR(rapl_ram_unit),
514 EVENT_PTR(rapl_psys_unit),
515
516 EVENT_PTR(rapl_cores_scale),
517 EVENT_PTR(rapl_pkg_scale),
518 EVENT_PTR(rapl_gpu_scale),
519 EVENT_PTR(rapl_ram_scale),
520 EVENT_PTR(rapl_psys_scale),
521 NULL,
522 };
523
524 static struct attribute *rapl_events_knl_attr[] = {
525 EVENT_PTR(rapl_pkg),
526 EVENT_PTR(rapl_ram),
527
528 EVENT_PTR(rapl_pkg_unit),
529 EVENT_PTR(rapl_ram_unit),
530
531 EVENT_PTR(rapl_pkg_scale),
532 EVENT_PTR(rapl_ram_scale),
533 NULL,
534 };
535
536 static struct attribute_group rapl_pmu_events_group = {
537 .name = "events",
538 .attrs = NULL, /* patched at runtime */
539 };
540
541 DEFINE_RAPL_FORMAT_ATTR(event, event, "config:0-7");
542 static struct attribute *rapl_formats_attr[] = {
543 &format_attr_event.attr,
544 NULL,
545 };
546
547 static struct attribute_group rapl_pmu_format_group = {
548 .name = "format",
549 .attrs = rapl_formats_attr,
550 };
551
552 const struct attribute_group *rapl_attr_groups[] = {
553 &rapl_pmu_attr_group,
554 &rapl_pmu_format_group,
555 &rapl_pmu_events_group,
556 NULL,
557 };
558
559 static void rapl_cpu_exit(int cpu)
560 {
561 struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
562 int target;
563
564 /* Check if exiting cpu is used for collecting rapl events */
565 if (!cpumask_test_and_clear_cpu(cpu, &rapl_cpu_mask))
566 return;
567
568 pmu->cpu = -1;
569 /* Find a new cpu to collect rapl events */
570 target = cpumask_any_but(topology_core_cpumask(cpu), cpu);
571
572 /* Migrate rapl events to the new target */
573 if (target < nr_cpu_ids) {
574 cpumask_set_cpu(target, &rapl_cpu_mask);
575 pmu->cpu = target;
576 perf_pmu_migrate_context(pmu->pmu, cpu, target);
577 }
578 }
579
580 static void rapl_cpu_init(int cpu)
581 {
582 struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
583 int target;
584
585 /*
586 * Check if there is an online cpu in the package which collects rapl
587 * events already.
588 */
589 target = cpumask_any_and(&rapl_cpu_mask, topology_core_cpumask(cpu));
590 if (target < nr_cpu_ids)
591 return;
592
593 cpumask_set_cpu(cpu, &rapl_cpu_mask);
594 pmu->cpu = cpu;
595 }
596
597 static int rapl_cpu_prepare(int cpu)
598 {
599 struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
600
601 if (pmu)
602 return 0;
603
604 pmu = kzalloc_node(sizeof(*pmu), GFP_KERNEL, cpu_to_node(cpu));
605 if (!pmu)
606 return -ENOMEM;
607
608 raw_spin_lock_init(&pmu->lock);
609 INIT_LIST_HEAD(&pmu->active_list);
610 pmu->pmu = &rapl_pmus->pmu;
611 pmu->timer_interval = ms_to_ktime(rapl_timer_ms);
612 pmu->cpu = -1;
613 rapl_hrtimer_init(pmu);
614 rapl_pmus->pmus[topology_logical_package_id(cpu)] = pmu;
615 return 0;
616 }
617
618 static int rapl_cpu_notifier(struct notifier_block *self,
619 unsigned long action, void *hcpu)
620 {
621 unsigned int cpu = (long)hcpu;
622
623 switch (action & ~CPU_TASKS_FROZEN) {
624 case CPU_UP_PREPARE:
625 rapl_cpu_prepare(cpu);
626 break;
627
628 case CPU_DOWN_FAILED:
629 case CPU_ONLINE:
630 rapl_cpu_init(cpu);
631 break;
632
633 case CPU_DOWN_PREPARE:
634 rapl_cpu_exit(cpu);
635 break;
636 }
637 return NOTIFY_OK;
638 }
639
640 static struct notifier_block rapl_cpu_nb = {
641 .notifier_call = rapl_cpu_notifier,
642 .priority = CPU_PRI_PERF + 1,
643 };
644
645 static int rapl_check_hw_unit(bool apply_quirk)
646 {
647 u64 msr_rapl_power_unit_bits;
648 int i;
649
650 /* protect rdmsrl() to handle virtualization */
651 if (rdmsrl_safe(MSR_RAPL_POWER_UNIT, &msr_rapl_power_unit_bits))
652 return -1;
653 for (i = 0; i < NR_RAPL_DOMAINS; i++)
654 rapl_hw_unit[i] = (msr_rapl_power_unit_bits >> 8) & 0x1FULL;
655
656 /*
657 * DRAM domain on HSW server and KNL has fixed energy unit which can be
658 * different than the unit from power unit MSR. See
659 * "Intel Xeon Processor E5-1600 and E5-2600 v3 Product Families, V2
660 * of 2. Datasheet, September 2014, Reference Number: 330784-001 "
661 */
662 if (apply_quirk)
663 rapl_hw_unit[RAPL_IDX_RAM_NRG_STAT] = 16;
664
665 /*
666 * Calculate the timer rate:
667 * Use reference of 200W for scaling the timeout to avoid counter
668 * overflows. 200W = 200 Joules/sec
669 * Divide interval by 2 to avoid lockstep (2 * 100)
670 * if hw unit is 32, then we use 2 ms 1/200/2
671 */
672 rapl_timer_ms = 2;
673 if (rapl_hw_unit[0] < 32) {
674 rapl_timer_ms = (1000 / (2 * 100));
675 rapl_timer_ms *= (1ULL << (32 - rapl_hw_unit[0] - 1));
676 }
677 return 0;
678 }
679
680 static void __init rapl_advertise(void)
681 {
682 int i;
683
684 pr_info("API unit is 2^-32 Joules, %d fixed counters, %llu ms ovfl timer\n",
685 hweight32(rapl_cntr_mask), rapl_timer_ms);
686
687 for (i = 0; i < NR_RAPL_DOMAINS; i++) {
688 if (rapl_cntr_mask & (1 << i)) {
689 pr_info("hw unit of domain %s 2^-%d Joules\n",
690 rapl_domain_names[i], rapl_hw_unit[i]);
691 }
692 }
693 }
694
695 static int __init rapl_prepare_cpus(void)
696 {
697 unsigned int cpu, pkg;
698 int ret;
699
700 for_each_online_cpu(cpu) {
701 pkg = topology_logical_package_id(cpu);
702 if (rapl_pmus->pmus[pkg])
703 continue;
704
705 ret = rapl_cpu_prepare(cpu);
706 if (ret)
707 return ret;
708 rapl_cpu_init(cpu);
709 }
710 return 0;
711 }
712
713 static void cleanup_rapl_pmus(void)
714 {
715 int i;
716
717 for (i = 0; i < rapl_pmus->maxpkg; i++)
718 kfree(rapl_pmus->pmus[i]);
719 kfree(rapl_pmus);
720 }
721
722 static int __init init_rapl_pmus(void)
723 {
724 int maxpkg = topology_max_packages();
725 size_t size;
726
727 size = sizeof(*rapl_pmus) + maxpkg * sizeof(struct rapl_pmu *);
728 rapl_pmus = kzalloc(size, GFP_KERNEL);
729 if (!rapl_pmus)
730 return -ENOMEM;
731
732 rapl_pmus->maxpkg = maxpkg;
733 rapl_pmus->pmu.attr_groups = rapl_attr_groups;
734 rapl_pmus->pmu.task_ctx_nr = perf_invalid_context;
735 rapl_pmus->pmu.event_init = rapl_pmu_event_init;
736 rapl_pmus->pmu.add = rapl_pmu_event_add;
737 rapl_pmus->pmu.del = rapl_pmu_event_del;
738 rapl_pmus->pmu.start = rapl_pmu_event_start;
739 rapl_pmus->pmu.stop = rapl_pmu_event_stop;
740 rapl_pmus->pmu.read = rapl_pmu_event_read;
741 return 0;
742 }
743
744 #define X86_RAPL_MODEL_MATCH(model, init) \
745 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_ANY, (unsigned long)&init }
746
747 struct intel_rapl_init_fun {
748 bool apply_quirk;
749 int cntr_mask;
750 struct attribute **attrs;
751 };
752
753 static const struct intel_rapl_init_fun snb_rapl_init __initconst = {
754 .apply_quirk = false,
755 .cntr_mask = RAPL_IDX_CLN,
756 .attrs = rapl_events_cln_attr,
757 };
758
759 static const struct intel_rapl_init_fun hsx_rapl_init __initconst = {
760 .apply_quirk = true,
761 .cntr_mask = RAPL_IDX_SRV,
762 .attrs = rapl_events_srv_attr,
763 };
764
765 static const struct intel_rapl_init_fun hsw_rapl_init __initconst = {
766 .apply_quirk = false,
767 .cntr_mask = RAPL_IDX_HSW,
768 .attrs = rapl_events_hsw_attr,
769 };
770
771 static const struct intel_rapl_init_fun snbep_rapl_init __initconst = {
772 .apply_quirk = false,
773 .cntr_mask = RAPL_IDX_SRV,
774 .attrs = rapl_events_srv_attr,
775 };
776
777 static const struct intel_rapl_init_fun knl_rapl_init __initconst = {
778 .apply_quirk = true,
779 .cntr_mask = RAPL_IDX_KNL,
780 .attrs = rapl_events_knl_attr,
781 };
782
783 static const struct intel_rapl_init_fun skl_rapl_init __initconst = {
784 .apply_quirk = false,
785 .cntr_mask = RAPL_IDX_SKL_CLN,
786 .attrs = rapl_events_skl_attr,
787 };
788
789 static const struct x86_cpu_id rapl_cpu_match[] __initconst = {
790 X86_RAPL_MODEL_MATCH(INTEL_FAM6_SANDYBRIDGE, snb_rapl_init),
791 X86_RAPL_MODEL_MATCH(INTEL_FAM6_SANDYBRIDGE_X, snbep_rapl_init),
792
793 X86_RAPL_MODEL_MATCH(INTEL_FAM6_IVYBRIDGE, snb_rapl_init),
794 X86_RAPL_MODEL_MATCH(INTEL_FAM6_IVYBRIDGE_X, snbep_rapl_init),
795
796 X86_RAPL_MODEL_MATCH(INTEL_FAM6_HASWELL_CORE, hsw_rapl_init),
797 X86_RAPL_MODEL_MATCH(INTEL_FAM6_HASWELL_X, hsw_rapl_init),
798 X86_RAPL_MODEL_MATCH(INTEL_FAM6_HASWELL_ULT, hsw_rapl_init),
799 X86_RAPL_MODEL_MATCH(INTEL_FAM6_HASWELL_GT3E, hsw_rapl_init),
800
801 X86_RAPL_MODEL_MATCH(INTEL_FAM6_BROADWELL_CORE, hsw_rapl_init),
802 X86_RAPL_MODEL_MATCH(INTEL_FAM6_BROADWELL_GT3E, hsw_rapl_init),
803 X86_RAPL_MODEL_MATCH(INTEL_FAM6_BROADWELL_X, hsw_rapl_init),
804 X86_RAPL_MODEL_MATCH(INTEL_FAM6_BROADWELL_XEON_D, hsw_rapl_init),
805
806 X86_RAPL_MODEL_MATCH(INTEL_FAM6_XEON_PHI_KNL, knl_rapl_init),
807
808 X86_RAPL_MODEL_MATCH(INTEL_FAM6_SKYLAKE_MOBILE, skl_rapl_init),
809 X86_RAPL_MODEL_MATCH(INTEL_FAM6_SKYLAKE_DESKTOP, skl_rapl_init),
810 X86_RAPL_MODEL_MATCH(INTEL_FAM6_SKYLAKE_X, hsx_rapl_init),
811 {},
812 };
813
814 MODULE_DEVICE_TABLE(x86cpu, rapl_cpu_match);
815
816 static int __init rapl_pmu_init(void)
817 {
818 const struct x86_cpu_id *id;
819 struct intel_rapl_init_fun *rapl_init;
820 bool apply_quirk;
821 int ret;
822
823 id = x86_match_cpu(rapl_cpu_match);
824 if (!id)
825 return -ENODEV;
826
827 rapl_init = (struct intel_rapl_init_fun *)id->driver_data;
828 apply_quirk = rapl_init->apply_quirk;
829 rapl_cntr_mask = rapl_init->cntr_mask;
830 rapl_pmu_events_group.attrs = rapl_init->attrs;
831
832 ret = rapl_check_hw_unit(apply_quirk);
833 if (ret)
834 return ret;
835
836 ret = init_rapl_pmus();
837 if (ret)
838 return ret;
839
840 cpu_notifier_register_begin();
841
842 ret = rapl_prepare_cpus();
843 if (ret)
844 goto out;
845
846 ret = perf_pmu_register(&rapl_pmus->pmu, "power", -1);
847 if (ret)
848 goto out;
849
850 __register_cpu_notifier(&rapl_cpu_nb);
851 cpu_notifier_register_done();
852 rapl_advertise();
853 return 0;
854
855 out:
856 pr_warn("Initialization failed (%d), disabled\n", ret);
857 cleanup_rapl_pmus();
858 cpu_notifier_register_done();
859 return ret;
860 }
861 module_init(rapl_pmu_init);
862
863 static void __exit intel_rapl_exit(void)
864 {
865 cpu_notifier_register_begin();
866 __unregister_cpu_notifier(&rapl_cpu_nb);
867 perf_pmu_unregister(&rapl_pmus->pmu);
868 cleanup_rapl_pmus();
869 cpu_notifier_register_done();
870 }
871 module_exit(intel_rapl_exit);