1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Common code for Intel Running Average Power Limit (RAPL) support.
4 * Copyright (c) 2019, Intel Corporation.
5 */
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/list.h>
11 #include <linux/types.h>
12 #include <linux/device.h>
13 #include <linux/slab.h>
14 #include <linux/log2.h>
15 #include <linux/bitmap.h>
16 #include <linux/delay.h>
17 #include <linux/sysfs.h>
18 #include <linux/cpu.h>
19 #include <linux/powercap.h>
20 #include <linux/suspend.h>
21 #include <linux/intel_rapl.h>
22 #include <linux/processor.h>
23 #include <linux/platform_device.h>
24
25 #include <asm/iosf_mbi.h>
26 #include <asm/cpu_device_id.h>
27 #include <asm/intel-family.h>
28
29 /* Local defines */
30 #define MSR_PLATFORM_POWER_LIMIT 0x0000065C
31
32 /* bitmasks for RAPL MSRs, used by primitive access functions */
33 #define ENERGY_STATUS_MASK 0xffffffff
34
35 #define POWER_LIMIT1_MASK 0x7FFF
36 #define POWER_LIMIT1_ENABLE BIT(15)
37 #define POWER_LIMIT1_CLAMP BIT(16)
38
39 #define POWER_LIMIT2_MASK (0x7FFFULL<<32)
40 #define POWER_LIMIT2_ENABLE BIT_ULL(47)
41 #define POWER_LIMIT2_CLAMP BIT_ULL(48)
42 #define POWER_HIGH_LOCK BIT_ULL(63)
43 #define POWER_LOW_LOCK BIT(31)
44
45 #define TIME_WINDOW1_MASK (0x7FULL<<17)
46 #define TIME_WINDOW2_MASK (0x7FULL<<49)
47
48 #define POWER_UNIT_OFFSET 0
49 #define POWER_UNIT_MASK 0x0F
50
51 #define ENERGY_UNIT_OFFSET 0x08
52 #define ENERGY_UNIT_MASK 0x1F00
53
54 #define TIME_UNIT_OFFSET 0x10
55 #define TIME_UNIT_MASK 0xF0000
56
57 #define POWER_INFO_MAX_MASK (0x7fffULL<<32)
58 #define POWER_INFO_MIN_MASK (0x7fffULL<<16)
59 #define POWER_INFO_MAX_TIME_WIN_MASK (0x3fULL<<48)
60 #define POWER_INFO_THERMAL_SPEC_MASK 0x7fff
61
62 #define PERF_STATUS_THROTTLE_TIME_MASK 0xffffffff
63 #define PP_POLICY_MASK 0x1F
64
65 /* Non HW constants */
66 #define RAPL_PRIMITIVE_DERIVED BIT(1) /* not from raw data */
67 #define RAPL_PRIMITIVE_DUMMY BIT(2)
68
69 #define TIME_WINDOW_MAX_MSEC 40000
70 #define TIME_WINDOW_MIN_MSEC 250
71 #define ENERGY_UNIT_SCALE 1000 /* scale from driver unit to powercap unit */
72 enum unit_type {
73 ARBITRARY_UNIT, /* no translation */
74 POWER_UNIT,
75 ENERGY_UNIT,
76 TIME_UNIT,
77 };
78
79 /* per domain data, some are optional */
80 #define NR_RAW_PRIMITIVES (NR_RAPL_PRIMITIVES - 2)
81
82 #define DOMAIN_STATE_INACTIVE BIT(0)
83 #define DOMAIN_STATE_POWER_LIMIT_SET BIT(1)
84 #define DOMAIN_STATE_BIOS_LOCKED BIT(2)
85
86 static const char pl1_name[] = "long_term";
87 static const char pl2_name[] = "short_term";
88
89 #define power_zone_to_rapl_domain(_zone) \
90 container_of(_zone, struct rapl_domain, power_zone)
91
92 struct rapl_defaults {
93 u8 floor_freq_reg_addr;
94 int (*check_unit)(struct rapl_package *rp, int cpu);
95 void (*set_floor_freq)(struct rapl_domain *rd, bool mode);
96 u64 (*compute_time_window)(struct rapl_package *rp, u64 val,
97 bool to_raw);
98 unsigned int dram_domain_energy_unit;
99 };
100 static struct rapl_defaults *rapl_defaults;
101
102 /* Sideband MBI registers */
103 #define IOSF_CPU_POWER_BUDGET_CTL_BYT (0x2)
104 #define IOSF_CPU_POWER_BUDGET_CTL_TNG (0xdf)
105
106 #define PACKAGE_PLN_INT_SAVED BIT(0)
107 #define MAX_PRIM_NAME (32)
108
109 /* per domain data. used to describe individual knobs such that access function
110 * can be consolidated into one instead of many inline functions.
111 */
112 struct rapl_primitive_info {
113 const char *name;
114 u64 mask;
115 int shift;
116 enum rapl_domain_reg_id id;
117 enum unit_type unit;
118 u32 flag;
119 };
120
121 #define PRIMITIVE_INFO_INIT(p, m, s, i, u, f) { \
122 .name = #p, \
123 .mask = m, \
124 .shift = s, \
125 .id = i, \
126 .unit = u, \
127 .flag = f \
128 }
129
130 static void rapl_init_domains(struct rapl_package *rp);
131 static int rapl_read_data_raw(struct rapl_domain *rd,
132 enum rapl_primitives prim,
133 bool xlate, u64 *data);
134 static int rapl_write_data_raw(struct rapl_domain *rd,
135 enum rapl_primitives prim,
136 unsigned long long value);
137 static u64 rapl_unit_xlate(struct rapl_domain *rd,
138 enum unit_type type, u64 value, int to_raw);
139 static void package_power_limit_irq_save(struct rapl_package *rp);
140
141 static LIST_HEAD(rapl_packages); /* guarded by CPU hotplug lock */
142
143 static const char *const rapl_domain_names[] = {
144 "package",
145 "core",
146 "uncore",
147 "dram",
148 "psys",
149 };
150
get_energy_counter(struct powercap_zone * power_zone,u64 * energy_raw)151 static int get_energy_counter(struct powercap_zone *power_zone,
152 u64 *energy_raw)
153 {
154 struct rapl_domain *rd;
155 u64 energy_now;
156
157 /* prevent CPU hotplug, make sure the RAPL domain does not go
158 * away while reading the counter.
159 */
160 get_online_cpus();
161 rd = power_zone_to_rapl_domain(power_zone);
162
163 if (!rapl_read_data_raw(rd, ENERGY_COUNTER, true, &energy_now)) {
164 *energy_raw = energy_now;
165 put_online_cpus();
166
167 return 0;
168 }
169 put_online_cpus();
170
171 return -EIO;
172 }
173
get_max_energy_counter(struct powercap_zone * pcd_dev,u64 * energy)174 static int get_max_energy_counter(struct powercap_zone *pcd_dev, u64 *energy)
175 {
176 struct rapl_domain *rd = power_zone_to_rapl_domain(pcd_dev);
177
178 *energy = rapl_unit_xlate(rd, ENERGY_UNIT, ENERGY_STATUS_MASK, 0);
179 return 0;
180 }
181
release_zone(struct powercap_zone * power_zone)182 static int release_zone(struct powercap_zone *power_zone)
183 {
184 struct rapl_domain *rd = power_zone_to_rapl_domain(power_zone);
185 struct rapl_package *rp = rd->rp;
186
187 /* package zone is the last zone of a package, we can free
188 * memory here since all children has been unregistered.
189 */
190 if (rd->id == RAPL_DOMAIN_PACKAGE) {
191 kfree(rd);
192 rp->domains = NULL;
193 }
194
195 return 0;
196
197 }
198
find_nr_power_limit(struct rapl_domain * rd)199 static int find_nr_power_limit(struct rapl_domain *rd)
200 {
201 int i, nr_pl = 0;
202
203 for (i = 0; i < NR_POWER_LIMITS; i++) {
204 if (rd->rpl[i].name)
205 nr_pl++;
206 }
207
208 return nr_pl;
209 }
210
set_domain_enable(struct powercap_zone * power_zone,bool mode)211 static int set_domain_enable(struct powercap_zone *power_zone, bool mode)
212 {
213 struct rapl_domain *rd = power_zone_to_rapl_domain(power_zone);
214
215 if (rd->state & DOMAIN_STATE_BIOS_LOCKED)
216 return -EACCES;
217
218 get_online_cpus();
219 rapl_write_data_raw(rd, PL1_ENABLE, mode);
220 if (rapl_defaults->set_floor_freq)
221 rapl_defaults->set_floor_freq(rd, mode);
222 put_online_cpus();
223
224 return 0;
225 }
226
get_domain_enable(struct powercap_zone * power_zone,bool * mode)227 static int get_domain_enable(struct powercap_zone *power_zone, bool *mode)
228 {
229 struct rapl_domain *rd = power_zone_to_rapl_domain(power_zone);
230 u64 val;
231
232 if (rd->state & DOMAIN_STATE_BIOS_LOCKED) {
233 *mode = false;
234 return 0;
235 }
236 get_online_cpus();
237 if (rapl_read_data_raw(rd, PL1_ENABLE, true, &val)) {
238 put_online_cpus();
239 return -EIO;
240 }
241 *mode = val;
242 put_online_cpus();
243
244 return 0;
245 }
246
247 /* per RAPL domain ops, in the order of rapl_domain_type */
248 static const struct powercap_zone_ops zone_ops[] = {
249 /* RAPL_DOMAIN_PACKAGE */
250 {
251 .get_energy_uj = get_energy_counter,
252 .get_max_energy_range_uj = get_max_energy_counter,
253 .release = release_zone,
254 .set_enable = set_domain_enable,
255 .get_enable = get_domain_enable,
256 },
257 /* RAPL_DOMAIN_PP0 */
258 {
259 .get_energy_uj = get_energy_counter,
260 .get_max_energy_range_uj = get_max_energy_counter,
261 .release = release_zone,
262 .set_enable = set_domain_enable,
263 .get_enable = get_domain_enable,
264 },
265 /* RAPL_DOMAIN_PP1 */
266 {
267 .get_energy_uj = get_energy_counter,
268 .get_max_energy_range_uj = get_max_energy_counter,
269 .release = release_zone,
270 .set_enable = set_domain_enable,
271 .get_enable = get_domain_enable,
272 },
273 /* RAPL_DOMAIN_DRAM */
274 {
275 .get_energy_uj = get_energy_counter,
276 .get_max_energy_range_uj = get_max_energy_counter,
277 .release = release_zone,
278 .set_enable = set_domain_enable,
279 .get_enable = get_domain_enable,
280 },
281 /* RAPL_DOMAIN_PLATFORM */
282 {
283 .get_energy_uj = get_energy_counter,
284 .get_max_energy_range_uj = get_max_energy_counter,
285 .release = release_zone,
286 .set_enable = set_domain_enable,
287 .get_enable = get_domain_enable,
288 },
289 };
290
291 /*
292 * Constraint index used by powercap can be different than power limit (PL)
293 * index in that some PLs maybe missing due to non-existent MSRs. So we
294 * need to convert here by finding the valid PLs only (name populated).
295 */
contraint_to_pl(struct rapl_domain * rd,int cid)296 static int contraint_to_pl(struct rapl_domain *rd, int cid)
297 {
298 int i, j;
299
300 for (i = 0, j = 0; i < NR_POWER_LIMITS; i++) {
301 if ((rd->rpl[i].name) && j++ == cid) {
302 pr_debug("%s: index %d\n", __func__, i);
303 return i;
304 }
305 }
306 pr_err("Cannot find matching power limit for constraint %d\n", cid);
307
308 return -EINVAL;
309 }
310
set_power_limit(struct powercap_zone * power_zone,int cid,u64 power_limit)311 static int set_power_limit(struct powercap_zone *power_zone, int cid,
312 u64 power_limit)
313 {
314 struct rapl_domain *rd;
315 struct rapl_package *rp;
316 int ret = 0;
317 int id;
318
319 get_online_cpus();
320 rd = power_zone_to_rapl_domain(power_zone);
321 id = contraint_to_pl(rd, cid);
322 if (id < 0) {
323 ret = id;
324 goto set_exit;
325 }
326
327 rp = rd->rp;
328
329 if (rd->state & DOMAIN_STATE_BIOS_LOCKED) {
330 dev_warn(&power_zone->dev,
331 "%s locked by BIOS, monitoring only\n", rd->name);
332 ret = -EACCES;
333 goto set_exit;
334 }
335
336 switch (rd->rpl[id].prim_id) {
337 case PL1_ENABLE:
338 rapl_write_data_raw(rd, POWER_LIMIT1, power_limit);
339 break;
340 case PL2_ENABLE:
341 rapl_write_data_raw(rd, POWER_LIMIT2, power_limit);
342 break;
343 default:
344 ret = -EINVAL;
345 }
346 if (!ret)
347 package_power_limit_irq_save(rp);
348 set_exit:
349 put_online_cpus();
350 return ret;
351 }
352
get_current_power_limit(struct powercap_zone * power_zone,int cid,u64 * data)353 static int get_current_power_limit(struct powercap_zone *power_zone, int cid,
354 u64 *data)
355 {
356 struct rapl_domain *rd;
357 u64 val;
358 int prim;
359 int ret = 0;
360 int id;
361
362 get_online_cpus();
363 rd = power_zone_to_rapl_domain(power_zone);
364 id = contraint_to_pl(rd, cid);
365 if (id < 0) {
366 ret = id;
367 goto get_exit;
368 }
369
370 switch (rd->rpl[id].prim_id) {
371 case PL1_ENABLE:
372 prim = POWER_LIMIT1;
373 break;
374 case PL2_ENABLE:
375 prim = POWER_LIMIT2;
376 break;
377 default:
378 put_online_cpus();
379 return -EINVAL;
380 }
381 if (rapl_read_data_raw(rd, prim, true, &val))
382 ret = -EIO;
383 else
384 *data = val;
385
386 get_exit:
387 put_online_cpus();
388
389 return ret;
390 }
391
set_time_window(struct powercap_zone * power_zone,int cid,u64 window)392 static int set_time_window(struct powercap_zone *power_zone, int cid,
393 u64 window)
394 {
395 struct rapl_domain *rd;
396 int ret = 0;
397 int id;
398
399 get_online_cpus();
400 rd = power_zone_to_rapl_domain(power_zone);
401 id = contraint_to_pl(rd, cid);
402 if (id < 0) {
403 ret = id;
404 goto set_time_exit;
405 }
406
407 switch (rd->rpl[id].prim_id) {
408 case PL1_ENABLE:
409 rapl_write_data_raw(rd, TIME_WINDOW1, window);
410 break;
411 case PL2_ENABLE:
412 rapl_write_data_raw(rd, TIME_WINDOW2, window);
413 break;
414 default:
415 ret = -EINVAL;
416 }
417
418 set_time_exit:
419 put_online_cpus();
420 return ret;
421 }
422
get_time_window(struct powercap_zone * power_zone,int cid,u64 * data)423 static int get_time_window(struct powercap_zone *power_zone, int cid,
424 u64 *data)
425 {
426 struct rapl_domain *rd;
427 u64 val;
428 int ret = 0;
429 int id;
430
431 get_online_cpus();
432 rd = power_zone_to_rapl_domain(power_zone);
433 id = contraint_to_pl(rd, cid);
434 if (id < 0) {
435 ret = id;
436 goto get_time_exit;
437 }
438
439 switch (rd->rpl[id].prim_id) {
440 case PL1_ENABLE:
441 ret = rapl_read_data_raw(rd, TIME_WINDOW1, true, &val);
442 break;
443 case PL2_ENABLE:
444 ret = rapl_read_data_raw(rd, TIME_WINDOW2, true, &val);
445 break;
446 default:
447 put_online_cpus();
448 return -EINVAL;
449 }
450 if (!ret)
451 *data = val;
452
453 get_time_exit:
454 put_online_cpus();
455
456 return ret;
457 }
458
get_constraint_name(struct powercap_zone * power_zone,int cid)459 static const char *get_constraint_name(struct powercap_zone *power_zone,
460 int cid)
461 {
462 struct rapl_domain *rd;
463 int id;
464
465 rd = power_zone_to_rapl_domain(power_zone);
466 id = contraint_to_pl(rd, cid);
467 if (id >= 0)
468 return rd->rpl[id].name;
469
470 return NULL;
471 }
472
get_max_power(struct powercap_zone * power_zone,int id,u64 * data)473 static int get_max_power(struct powercap_zone *power_zone, int id, u64 *data)
474 {
475 struct rapl_domain *rd;
476 u64 val;
477 int prim;
478 int ret = 0;
479
480 get_online_cpus();
481 rd = power_zone_to_rapl_domain(power_zone);
482 switch (rd->rpl[id].prim_id) {
483 case PL1_ENABLE:
484 prim = THERMAL_SPEC_POWER;
485 break;
486 case PL2_ENABLE:
487 prim = MAX_POWER;
488 break;
489 default:
490 put_online_cpus();
491 return -EINVAL;
492 }
493 if (rapl_read_data_raw(rd, prim, true, &val))
494 ret = -EIO;
495 else
496 *data = val;
497
498 put_online_cpus();
499
500 return ret;
501 }
502
503 static const struct powercap_zone_constraint_ops constraint_ops = {
504 .set_power_limit_uw = set_power_limit,
505 .get_power_limit_uw = get_current_power_limit,
506 .set_time_window_us = set_time_window,
507 .get_time_window_us = get_time_window,
508 .get_max_power_uw = get_max_power,
509 .get_name = get_constraint_name,
510 };
511
512 /* called after domain detection and package level data are set */
rapl_init_domains(struct rapl_package * rp)513 static void rapl_init_domains(struct rapl_package *rp)
514 {
515 enum rapl_domain_type i;
516 enum rapl_domain_reg_id j;
517 struct rapl_domain *rd = rp->domains;
518
519 for (i = 0; i < RAPL_DOMAIN_MAX; i++) {
520 unsigned int mask = rp->domain_map & (1 << i);
521
522 if (!mask)
523 continue;
524
525 rd->rp = rp;
526 rd->name = rapl_domain_names[i];
527 rd->id = i;
528 rd->rpl[0].prim_id = PL1_ENABLE;
529 rd->rpl[0].name = pl1_name;
530 /* some domain may support two power limits */
531 if (rp->priv->limits[i] == 2) {
532 rd->rpl[1].prim_id = PL2_ENABLE;
533 rd->rpl[1].name = pl2_name;
534 }
535
536 for (j = 0; j < RAPL_DOMAIN_REG_MAX; j++)
537 rd->regs[j] = rp->priv->regs[i][j];
538
539 if (i == RAPL_DOMAIN_DRAM) {
540 rd->domain_energy_unit =
541 rapl_defaults->dram_domain_energy_unit;
542 if (rd->domain_energy_unit)
543 pr_info("DRAM domain energy unit %dpj\n",
544 rd->domain_energy_unit);
545 }
546 rd++;
547 }
548 }
549
rapl_unit_xlate(struct rapl_domain * rd,enum unit_type type,u64 value,int to_raw)550 static u64 rapl_unit_xlate(struct rapl_domain *rd, enum unit_type type,
551 u64 value, int to_raw)
552 {
553 u64 units = 1;
554 struct rapl_package *rp = rd->rp;
555 u64 scale = 1;
556
557 switch (type) {
558 case POWER_UNIT:
559 units = rp->power_unit;
560 break;
561 case ENERGY_UNIT:
562 scale = ENERGY_UNIT_SCALE;
563 /* per domain unit takes precedence */
564 if (rd->domain_energy_unit)
565 units = rd->domain_energy_unit;
566 else
567 units = rp->energy_unit;
568 break;
569 case TIME_UNIT:
570 return rapl_defaults->compute_time_window(rp, value, to_raw);
571 case ARBITRARY_UNIT:
572 default:
573 return value;
574 };
575
576 if (to_raw)
577 return div64_u64(value, units) * scale;
578
579 value *= units;
580
581 return div64_u64(value, scale);
582 }
583
584 /* in the order of enum rapl_primitives */
585 static struct rapl_primitive_info rpi[] = {
586 /* name, mask, shift, msr index, unit divisor */
587 PRIMITIVE_INFO_INIT(ENERGY_COUNTER, ENERGY_STATUS_MASK, 0,
588 RAPL_DOMAIN_REG_STATUS, ENERGY_UNIT, 0),
589 PRIMITIVE_INFO_INIT(POWER_LIMIT1, POWER_LIMIT1_MASK, 0,
590 RAPL_DOMAIN_REG_LIMIT, POWER_UNIT, 0),
591 PRIMITIVE_INFO_INIT(POWER_LIMIT2, POWER_LIMIT2_MASK, 32,
592 RAPL_DOMAIN_REG_LIMIT, POWER_UNIT, 0),
593 PRIMITIVE_INFO_INIT(FW_LOCK, POWER_LOW_LOCK, 31,
594 RAPL_DOMAIN_REG_LIMIT, ARBITRARY_UNIT, 0),
595 PRIMITIVE_INFO_INIT(PL1_ENABLE, POWER_LIMIT1_ENABLE, 15,
596 RAPL_DOMAIN_REG_LIMIT, ARBITRARY_UNIT, 0),
597 PRIMITIVE_INFO_INIT(PL1_CLAMP, POWER_LIMIT1_CLAMP, 16,
598 RAPL_DOMAIN_REG_LIMIT, ARBITRARY_UNIT, 0),
599 PRIMITIVE_INFO_INIT(PL2_ENABLE, POWER_LIMIT2_ENABLE, 47,
600 RAPL_DOMAIN_REG_LIMIT, ARBITRARY_UNIT, 0),
601 PRIMITIVE_INFO_INIT(PL2_CLAMP, POWER_LIMIT2_CLAMP, 48,
602 RAPL_DOMAIN_REG_LIMIT, ARBITRARY_UNIT, 0),
603 PRIMITIVE_INFO_INIT(TIME_WINDOW1, TIME_WINDOW1_MASK, 17,
604 RAPL_DOMAIN_REG_LIMIT, TIME_UNIT, 0),
605 PRIMITIVE_INFO_INIT(TIME_WINDOW2, TIME_WINDOW2_MASK, 49,
606 RAPL_DOMAIN_REG_LIMIT, TIME_UNIT, 0),
607 PRIMITIVE_INFO_INIT(THERMAL_SPEC_POWER, POWER_INFO_THERMAL_SPEC_MASK,
608 0, RAPL_DOMAIN_REG_INFO, POWER_UNIT, 0),
609 PRIMITIVE_INFO_INIT(MAX_POWER, POWER_INFO_MAX_MASK, 32,
610 RAPL_DOMAIN_REG_INFO, POWER_UNIT, 0),
611 PRIMITIVE_INFO_INIT(MIN_POWER, POWER_INFO_MIN_MASK, 16,
612 RAPL_DOMAIN_REG_INFO, POWER_UNIT, 0),
613 PRIMITIVE_INFO_INIT(MAX_TIME_WINDOW, POWER_INFO_MAX_TIME_WIN_MASK, 48,
614 RAPL_DOMAIN_REG_INFO, TIME_UNIT, 0),
615 PRIMITIVE_INFO_INIT(THROTTLED_TIME, PERF_STATUS_THROTTLE_TIME_MASK, 0,
616 RAPL_DOMAIN_REG_PERF, TIME_UNIT, 0),
617 PRIMITIVE_INFO_INIT(PRIORITY_LEVEL, PP_POLICY_MASK, 0,
618 RAPL_DOMAIN_REG_POLICY, ARBITRARY_UNIT, 0),
619 /* non-hardware */
620 PRIMITIVE_INFO_INIT(AVERAGE_POWER, 0, 0, 0, POWER_UNIT,
621 RAPL_PRIMITIVE_DERIVED),
622 {NULL, 0, 0, 0},
623 };
624
625 /* Read primitive data based on its related struct rapl_primitive_info.
626 * if xlate flag is set, return translated data based on data units, i.e.
627 * time, energy, and power.
628 * RAPL MSRs are non-architectual and are laid out not consistently across
629 * domains. Here we use primitive info to allow writing consolidated access
630 * functions.
631 * For a given primitive, it is processed by MSR mask and shift. Unit conversion
632 * is pre-assigned based on RAPL unit MSRs read at init time.
633 * 63-------------------------- 31--------------------------- 0
634 * | xxxxx (mask) |
635 * | |<- shift ----------------|
636 * 63-------------------------- 31--------------------------- 0
637 */
rapl_read_data_raw(struct rapl_domain * rd,enum rapl_primitives prim,bool xlate,u64 * data)638 static int rapl_read_data_raw(struct rapl_domain *rd,
639 enum rapl_primitives prim, bool xlate, u64 *data)
640 {
641 u64 value;
642 struct rapl_primitive_info *rp = &rpi[prim];
643 struct reg_action ra;
644 int cpu;
645
646 if (!rp->name || rp->flag & RAPL_PRIMITIVE_DUMMY)
647 return -EINVAL;
648
649 ra.reg = rd->regs[rp->id];
650 if (!ra.reg)
651 return -EINVAL;
652
653 cpu = rd->rp->lead_cpu;
654
655 /* domain with 2 limits has different bit */
656 if (prim == FW_LOCK && rd->rp->priv->limits[rd->id] == 2) {
657 rp->mask = POWER_HIGH_LOCK;
658 rp->shift = 63;
659 }
660 /* non-hardware data are collected by the polling thread */
661 if (rp->flag & RAPL_PRIMITIVE_DERIVED) {
662 *data = rd->rdd.primitives[prim];
663 return 0;
664 }
665
666 ra.mask = rp->mask;
667
668 if (rd->rp->priv->read_raw(cpu, &ra)) {
669 pr_debug("failed to read reg 0x%llx on cpu %d\n", ra.reg, cpu);
670 return -EIO;
671 }
672
673 value = ra.value >> rp->shift;
674
675 if (xlate)
676 *data = rapl_unit_xlate(rd, rp->unit, value, 0);
677 else
678 *data = value;
679
680 return 0;
681 }
682
683 /* Similar use of primitive info in the read counterpart */
rapl_write_data_raw(struct rapl_domain * rd,enum rapl_primitives prim,unsigned long long value)684 static int rapl_write_data_raw(struct rapl_domain *rd,
685 enum rapl_primitives prim,
686 unsigned long long value)
687 {
688 struct rapl_primitive_info *rp = &rpi[prim];
689 int cpu;
690 u64 bits;
691 struct reg_action ra;
692 int ret;
693
694 cpu = rd->rp->lead_cpu;
695 bits = rapl_unit_xlate(rd, rp->unit, value, 1);
696 bits <<= rp->shift;
697 bits &= rp->mask;
698
699 memset(&ra, 0, sizeof(ra));
700
701 ra.reg = rd->regs[rp->id];
702 ra.mask = rp->mask;
703 ra.value = bits;
704
705 ret = rd->rp->priv->write_raw(cpu, &ra);
706
707 return ret;
708 }
709
710 /*
711 * Raw RAPL data stored in MSRs are in certain scales. We need to
712 * convert them into standard units based on the units reported in
713 * the RAPL unit MSRs. This is specific to CPUs as the method to
714 * calculate units differ on different CPUs.
715 * We convert the units to below format based on CPUs.
716 * i.e.
717 * energy unit: picoJoules : Represented in picoJoules by default
718 * power unit : microWatts : Represented in milliWatts by default
719 * time unit : microseconds: Represented in seconds by default
720 */
rapl_check_unit_core(struct rapl_package * rp,int cpu)721 static int rapl_check_unit_core(struct rapl_package *rp, int cpu)
722 {
723 struct reg_action ra;
724 u32 value;
725
726 ra.reg = rp->priv->reg_unit;
727 ra.mask = ~0;
728 if (rp->priv->read_raw(cpu, &ra)) {
729 pr_err("Failed to read power unit REG 0x%llx on CPU %d, exit.\n",
730 rp->priv->reg_unit, cpu);
731 return -ENODEV;
732 }
733
734 value = (ra.value & ENERGY_UNIT_MASK) >> ENERGY_UNIT_OFFSET;
735 rp->energy_unit = ENERGY_UNIT_SCALE * 1000000 / (1 << value);
736
737 value = (ra.value & POWER_UNIT_MASK) >> POWER_UNIT_OFFSET;
738 rp->power_unit = 1000000 / (1 << value);
739
740 value = (ra.value & TIME_UNIT_MASK) >> TIME_UNIT_OFFSET;
741 rp->time_unit = 1000000 / (1 << value);
742
743 pr_debug("Core CPU %s energy=%dpJ, time=%dus, power=%duW\n",
744 rp->name, rp->energy_unit, rp->time_unit, rp->power_unit);
745
746 return 0;
747 }
748
rapl_check_unit_atom(struct rapl_package * rp,int cpu)749 static int rapl_check_unit_atom(struct rapl_package *rp, int cpu)
750 {
751 struct reg_action ra;
752 u32 value;
753
754 ra.reg = rp->priv->reg_unit;
755 ra.mask = ~0;
756 if (rp->priv->read_raw(cpu, &ra)) {
757 pr_err("Failed to read power unit REG 0x%llx on CPU %d, exit.\n",
758 rp->priv->reg_unit, cpu);
759 return -ENODEV;
760 }
761
762 value = (ra.value & ENERGY_UNIT_MASK) >> ENERGY_UNIT_OFFSET;
763 rp->energy_unit = ENERGY_UNIT_SCALE * 1 << value;
764
765 value = (ra.value & POWER_UNIT_MASK) >> POWER_UNIT_OFFSET;
766 rp->power_unit = (1 << value) * 1000;
767
768 value = (ra.value & TIME_UNIT_MASK) >> TIME_UNIT_OFFSET;
769 rp->time_unit = 1000000 / (1 << value);
770
771 pr_debug("Atom %s energy=%dpJ, time=%dus, power=%duW\n",
772 rp->name, rp->energy_unit, rp->time_unit, rp->power_unit);
773
774 return 0;
775 }
776
power_limit_irq_save_cpu(void * info)777 static void power_limit_irq_save_cpu(void *info)
778 {
779 u32 l, h = 0;
780 struct rapl_package *rp = (struct rapl_package *)info;
781
782 /* save the state of PLN irq mask bit before disabling it */
783 rdmsr_safe(MSR_IA32_PACKAGE_THERM_INTERRUPT, &l, &h);
784 if (!(rp->power_limit_irq & PACKAGE_PLN_INT_SAVED)) {
785 rp->power_limit_irq = l & PACKAGE_THERM_INT_PLN_ENABLE;
786 rp->power_limit_irq |= PACKAGE_PLN_INT_SAVED;
787 }
788 l &= ~PACKAGE_THERM_INT_PLN_ENABLE;
789 wrmsr_safe(MSR_IA32_PACKAGE_THERM_INTERRUPT, l, h);
790 }
791
792 /* REVISIT:
793 * When package power limit is set artificially low by RAPL, LVT
794 * thermal interrupt for package power limit should be ignored
795 * since we are not really exceeding the real limit. The intention
796 * is to avoid excessive interrupts while we are trying to save power.
797 * A useful feature might be routing the package_power_limit interrupt
798 * to userspace via eventfd. once we have a usecase, this is simple
799 * to do by adding an atomic notifier.
800 */
801
package_power_limit_irq_save(struct rapl_package * rp)802 static void package_power_limit_irq_save(struct rapl_package *rp)
803 {
804 if (!boot_cpu_has(X86_FEATURE_PTS) || !boot_cpu_has(X86_FEATURE_PLN))
805 return;
806
807 smp_call_function_single(rp->lead_cpu, power_limit_irq_save_cpu, rp, 1);
808 }
809
810 /*
811 * Restore per package power limit interrupt enable state. Called from cpu
812 * hotplug code on package removal.
813 */
package_power_limit_irq_restore(struct rapl_package * rp)814 static void package_power_limit_irq_restore(struct rapl_package *rp)
815 {
816 u32 l, h;
817
818 if (!boot_cpu_has(X86_FEATURE_PTS) || !boot_cpu_has(X86_FEATURE_PLN))
819 return;
820
821 /* irq enable state not saved, nothing to restore */
822 if (!(rp->power_limit_irq & PACKAGE_PLN_INT_SAVED))
823 return;
824
825 rdmsr_safe(MSR_IA32_PACKAGE_THERM_INTERRUPT, &l, &h);
826
827 if (rp->power_limit_irq & PACKAGE_THERM_INT_PLN_ENABLE)
828 l |= PACKAGE_THERM_INT_PLN_ENABLE;
829 else
830 l &= ~PACKAGE_THERM_INT_PLN_ENABLE;
831
832 wrmsr_safe(MSR_IA32_PACKAGE_THERM_INTERRUPT, l, h);
833 }
834
set_floor_freq_default(struct rapl_domain * rd,bool mode)835 static void set_floor_freq_default(struct rapl_domain *rd, bool mode)
836 {
837 int nr_powerlimit = find_nr_power_limit(rd);
838
839 /* always enable clamp such that p-state can go below OS requested
840 * range. power capping priority over guranteed frequency.
841 */
842 rapl_write_data_raw(rd, PL1_CLAMP, mode);
843
844 /* some domains have pl2 */
845 if (nr_powerlimit > 1) {
846 rapl_write_data_raw(rd, PL2_ENABLE, mode);
847 rapl_write_data_raw(rd, PL2_CLAMP, mode);
848 }
849 }
850
set_floor_freq_atom(struct rapl_domain * rd,bool enable)851 static void set_floor_freq_atom(struct rapl_domain *rd, bool enable)
852 {
853 static u32 power_ctrl_orig_val;
854 u32 mdata;
855
856 if (!rapl_defaults->floor_freq_reg_addr) {
857 pr_err("Invalid floor frequency config register\n");
858 return;
859 }
860
861 if (!power_ctrl_orig_val)
862 iosf_mbi_read(BT_MBI_UNIT_PMC, MBI_CR_READ,
863 rapl_defaults->floor_freq_reg_addr,
864 &power_ctrl_orig_val);
865 mdata = power_ctrl_orig_val;
866 if (enable) {
867 mdata &= ~(0x7f << 8);
868 mdata |= 1 << 8;
869 }
870 iosf_mbi_write(BT_MBI_UNIT_PMC, MBI_CR_WRITE,
871 rapl_defaults->floor_freq_reg_addr, mdata);
872 }
873
rapl_compute_time_window_core(struct rapl_package * rp,u64 value,bool to_raw)874 static u64 rapl_compute_time_window_core(struct rapl_package *rp, u64 value,
875 bool to_raw)
876 {
877 u64 f, y; /* fraction and exp. used for time unit */
878
879 /*
880 * Special processing based on 2^Y*(1+F/4), refer
881 * to Intel Software Developer's manual Vol.3B: CH 14.9.3.
882 */
883 if (!to_raw) {
884 f = (value & 0x60) >> 5;
885 y = value & 0x1f;
886 value = (1 << y) * (4 + f) * rp->time_unit / 4;
887 } else {
888 do_div(value, rp->time_unit);
889 y = ilog2(value);
890 f = div64_u64(4 * (value - (1 << y)), 1 << y);
891 value = (y & 0x1f) | ((f & 0x3) << 5);
892 }
893 return value;
894 }
895
rapl_compute_time_window_atom(struct rapl_package * rp,u64 value,bool to_raw)896 static u64 rapl_compute_time_window_atom(struct rapl_package *rp, u64 value,
897 bool to_raw)
898 {
899 /*
900 * Atom time unit encoding is straight forward val * time_unit,
901 * where time_unit is default to 1 sec. Never 0.
902 */
903 if (!to_raw)
904 return (value) ? value *= rp->time_unit : rp->time_unit;
905
906 value = div64_u64(value, rp->time_unit);
907
908 return value;
909 }
910
911 static const struct rapl_defaults rapl_defaults_core = {
912 .floor_freq_reg_addr = 0,
913 .check_unit = rapl_check_unit_core,
914 .set_floor_freq = set_floor_freq_default,
915 .compute_time_window = rapl_compute_time_window_core,
916 };
917
918 static const struct rapl_defaults rapl_defaults_hsw_server = {
919 .check_unit = rapl_check_unit_core,
920 .set_floor_freq = set_floor_freq_default,
921 .compute_time_window = rapl_compute_time_window_core,
922 .dram_domain_energy_unit = 15300,
923 };
924
925 static const struct rapl_defaults rapl_defaults_byt = {
926 .floor_freq_reg_addr = IOSF_CPU_POWER_BUDGET_CTL_BYT,
927 .check_unit = rapl_check_unit_atom,
928 .set_floor_freq = set_floor_freq_atom,
929 .compute_time_window = rapl_compute_time_window_atom,
930 };
931
932 static const struct rapl_defaults rapl_defaults_tng = {
933 .floor_freq_reg_addr = IOSF_CPU_POWER_BUDGET_CTL_TNG,
934 .check_unit = rapl_check_unit_atom,
935 .set_floor_freq = set_floor_freq_atom,
936 .compute_time_window = rapl_compute_time_window_atom,
937 };
938
939 static const struct rapl_defaults rapl_defaults_ann = {
940 .floor_freq_reg_addr = 0,
941 .check_unit = rapl_check_unit_atom,
942 .set_floor_freq = NULL,
943 .compute_time_window = rapl_compute_time_window_atom,
944 };
945
946 static const struct rapl_defaults rapl_defaults_cht = {
947 .floor_freq_reg_addr = 0,
948 .check_unit = rapl_check_unit_atom,
949 .set_floor_freq = NULL,
950 .compute_time_window = rapl_compute_time_window_atom,
951 };
952
953 static const struct x86_cpu_id rapl_ids[] __initconst = {
954 INTEL_CPU_FAM6(SANDYBRIDGE, rapl_defaults_core),
955 INTEL_CPU_FAM6(SANDYBRIDGE_X, rapl_defaults_core),
956
957 INTEL_CPU_FAM6(IVYBRIDGE, rapl_defaults_core),
958 INTEL_CPU_FAM6(IVYBRIDGE_X, rapl_defaults_core),
959
960 INTEL_CPU_FAM6(HASWELL, rapl_defaults_core),
961 INTEL_CPU_FAM6(HASWELL_L, rapl_defaults_core),
962 INTEL_CPU_FAM6(HASWELL_G, rapl_defaults_core),
963 INTEL_CPU_FAM6(HASWELL_X, rapl_defaults_hsw_server),
964
965 INTEL_CPU_FAM6(BROADWELL, rapl_defaults_core),
966 INTEL_CPU_FAM6(BROADWELL_G, rapl_defaults_core),
967 INTEL_CPU_FAM6(BROADWELL_D, rapl_defaults_core),
968 INTEL_CPU_FAM6(BROADWELL_X, rapl_defaults_hsw_server),
969
970 INTEL_CPU_FAM6(SKYLAKE, rapl_defaults_core),
971 INTEL_CPU_FAM6(SKYLAKE_L, rapl_defaults_core),
972 INTEL_CPU_FAM6(SKYLAKE_X, rapl_defaults_hsw_server),
973 INTEL_CPU_FAM6(KABYLAKE_L, rapl_defaults_core),
974 INTEL_CPU_FAM6(KABYLAKE, rapl_defaults_core),
975 INTEL_CPU_FAM6(CANNONLAKE_L, rapl_defaults_core),
976 INTEL_CPU_FAM6(ICELAKE_L, rapl_defaults_core),
977 INTEL_CPU_FAM6(ICELAKE, rapl_defaults_core),
978 INTEL_CPU_FAM6(ICELAKE_NNPI, rapl_defaults_core),
979 INTEL_CPU_FAM6(ICELAKE_X, rapl_defaults_hsw_server),
980 INTEL_CPU_FAM6(ICELAKE_D, rapl_defaults_hsw_server),
981
982 INTEL_CPU_FAM6(ATOM_SILVERMONT, rapl_defaults_byt),
983 INTEL_CPU_FAM6(ATOM_AIRMONT, rapl_defaults_cht),
984 INTEL_CPU_FAM6(ATOM_SILVERMONT_MID, rapl_defaults_tng),
985 INTEL_CPU_FAM6(ATOM_AIRMONT_MID, rapl_defaults_ann),
986 INTEL_CPU_FAM6(ATOM_GOLDMONT, rapl_defaults_core),
987 INTEL_CPU_FAM6(ATOM_GOLDMONT_PLUS, rapl_defaults_core),
988 INTEL_CPU_FAM6(ATOM_GOLDMONT_D, rapl_defaults_core),
989 INTEL_CPU_FAM6(ATOM_TREMONT_D, rapl_defaults_core),
990
991 INTEL_CPU_FAM6(XEON_PHI_KNL, rapl_defaults_hsw_server),
992 INTEL_CPU_FAM6(XEON_PHI_KNM, rapl_defaults_hsw_server),
993 {}
994 };
995
996 MODULE_DEVICE_TABLE(x86cpu, rapl_ids);
997
998 /* Read once for all raw primitive data for domains */
rapl_update_domain_data(struct rapl_package * rp)999 static void rapl_update_domain_data(struct rapl_package *rp)
1000 {
1001 int dmn, prim;
1002 u64 val;
1003
1004 for (dmn = 0; dmn < rp->nr_domains; dmn++) {
1005 pr_debug("update %s domain %s data\n", rp->name,
1006 rp->domains[dmn].name);
1007 /* exclude non-raw primitives */
1008 for (prim = 0; prim < NR_RAW_PRIMITIVES; prim++) {
1009 if (!rapl_read_data_raw(&rp->domains[dmn], prim,
1010 rpi[prim].unit, &val))
1011 rp->domains[dmn].rdd.primitives[prim] = val;
1012 }
1013 }
1014
1015 }
1016
rapl_package_register_powercap(struct rapl_package * rp)1017 static int rapl_package_register_powercap(struct rapl_package *rp)
1018 {
1019 struct rapl_domain *rd;
1020 struct powercap_zone *power_zone = NULL;
1021 int nr_pl, ret;
1022
1023 /* Update the domain data of the new package */
1024 rapl_update_domain_data(rp);
1025
1026 /* first we register package domain as the parent zone */
1027 for (rd = rp->domains; rd < rp->domains + rp->nr_domains; rd++) {
1028 if (rd->id == RAPL_DOMAIN_PACKAGE) {
1029 nr_pl = find_nr_power_limit(rd);
1030 pr_debug("register package domain %s\n", rp->name);
1031 power_zone = powercap_register_zone(&rd->power_zone,
1032 rp->priv->control_type, rp->name,
1033 NULL, &zone_ops[rd->id], nr_pl,
1034 &constraint_ops);
1035 if (IS_ERR(power_zone)) {
1036 pr_debug("failed to register power zone %s\n",
1037 rp->name);
1038 return PTR_ERR(power_zone);
1039 }
1040 /* track parent zone in per package/socket data */
1041 rp->power_zone = power_zone;
1042 /* done, only one package domain per socket */
1043 break;
1044 }
1045 }
1046 if (!power_zone) {
1047 pr_err("no package domain found, unknown topology!\n");
1048 return -ENODEV;
1049 }
1050 /* now register domains as children of the socket/package */
1051 for (rd = rp->domains; rd < rp->domains + rp->nr_domains; rd++) {
1052 if (rd->id == RAPL_DOMAIN_PACKAGE)
1053 continue;
1054 /* number of power limits per domain varies */
1055 nr_pl = find_nr_power_limit(rd);
1056 power_zone = powercap_register_zone(&rd->power_zone,
1057 rp->priv->control_type,
1058 rd->name, rp->power_zone,
1059 &zone_ops[rd->id], nr_pl,
1060 &constraint_ops);
1061
1062 if (IS_ERR(power_zone)) {
1063 pr_debug("failed to register power_zone, %s:%s\n",
1064 rp->name, rd->name);
1065 ret = PTR_ERR(power_zone);
1066 goto err_cleanup;
1067 }
1068 }
1069 return 0;
1070
1071 err_cleanup:
1072 /*
1073 * Clean up previously initialized domains within the package if we
1074 * failed after the first domain setup.
1075 */
1076 while (--rd >= rp->domains) {
1077 pr_debug("unregister %s domain %s\n", rp->name, rd->name);
1078 powercap_unregister_zone(rp->priv->control_type,
1079 &rd->power_zone);
1080 }
1081
1082 return ret;
1083 }
1084
rapl_add_platform_domain(struct rapl_if_priv * priv)1085 int rapl_add_platform_domain(struct rapl_if_priv *priv)
1086 {
1087 struct rapl_domain *rd;
1088 struct powercap_zone *power_zone;
1089 struct reg_action ra;
1090 int ret;
1091
1092 ra.reg = priv->regs[RAPL_DOMAIN_PLATFORM][RAPL_DOMAIN_REG_STATUS];
1093 ra.mask = ~0;
1094 ret = priv->read_raw(0, &ra);
1095 if (ret || !ra.value)
1096 return -ENODEV;
1097
1098 ra.reg = priv->regs[RAPL_DOMAIN_PLATFORM][RAPL_DOMAIN_REG_LIMIT];
1099 ra.mask = ~0;
1100 ret = priv->read_raw(0, &ra);
1101 if (ret || !ra.value)
1102 return -ENODEV;
1103
1104 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
1105 if (!rd)
1106 return -ENOMEM;
1107
1108 rd->name = rapl_domain_names[RAPL_DOMAIN_PLATFORM];
1109 rd->id = RAPL_DOMAIN_PLATFORM;
1110 rd->regs[RAPL_DOMAIN_REG_LIMIT] =
1111 priv->regs[RAPL_DOMAIN_PLATFORM][RAPL_DOMAIN_REG_LIMIT];
1112 rd->regs[RAPL_DOMAIN_REG_STATUS] =
1113 priv->regs[RAPL_DOMAIN_PLATFORM][RAPL_DOMAIN_REG_STATUS];
1114 rd->rpl[0].prim_id = PL1_ENABLE;
1115 rd->rpl[0].name = pl1_name;
1116 rd->rpl[1].prim_id = PL2_ENABLE;
1117 rd->rpl[1].name = pl2_name;
1118 rd->rp = rapl_find_package_domain(0, priv);
1119
1120 power_zone = powercap_register_zone(&rd->power_zone, priv->control_type,
1121 "psys", NULL,
1122 &zone_ops[RAPL_DOMAIN_PLATFORM],
1123 2, &constraint_ops);
1124
1125 if (IS_ERR(power_zone)) {
1126 kfree(rd);
1127 return PTR_ERR(power_zone);
1128 }
1129
1130 priv->platform_rapl_domain = rd;
1131
1132 return 0;
1133 }
1134 EXPORT_SYMBOL_GPL(rapl_add_platform_domain);
1135
rapl_remove_platform_domain(struct rapl_if_priv * priv)1136 void rapl_remove_platform_domain(struct rapl_if_priv *priv)
1137 {
1138 if (priv->platform_rapl_domain) {
1139 powercap_unregister_zone(priv->control_type,
1140 &priv->platform_rapl_domain->power_zone);
1141 kfree(priv->platform_rapl_domain);
1142 }
1143 }
1144 EXPORT_SYMBOL_GPL(rapl_remove_platform_domain);
1145
rapl_check_domain(int cpu,int domain,struct rapl_package * rp)1146 static int rapl_check_domain(int cpu, int domain, struct rapl_package *rp)
1147 {
1148 struct reg_action ra;
1149
1150 switch (domain) {
1151 case RAPL_DOMAIN_PACKAGE:
1152 case RAPL_DOMAIN_PP0:
1153 case RAPL_DOMAIN_PP1:
1154 case RAPL_DOMAIN_DRAM:
1155 ra.reg = rp->priv->regs[domain][RAPL_DOMAIN_REG_STATUS];
1156 break;
1157 case RAPL_DOMAIN_PLATFORM:
1158 /* PSYS(PLATFORM) is not a CPU domain, so avoid printng error */
1159 return -EINVAL;
1160 default:
1161 pr_err("invalid domain id %d\n", domain);
1162 return -EINVAL;
1163 }
1164 /* make sure domain counters are available and contains non-zero
1165 * values, otherwise skip it.
1166 */
1167
1168 ra.mask = ~0;
1169 if (rp->priv->read_raw(cpu, &ra) || !ra.value)
1170 return -ENODEV;
1171
1172 return 0;
1173 }
1174
1175 /*
1176 * Check if power limits are available. Two cases when they are not available:
1177 * 1. Locked by BIOS, in this case we still provide read-only access so that
1178 * users can see what limit is set by the BIOS.
1179 * 2. Some CPUs make some domains monitoring only which means PLx MSRs may not
1180 * exist at all. In this case, we do not show the constraints in powercap.
1181 *
1182 * Called after domains are detected and initialized.
1183 */
rapl_detect_powerlimit(struct rapl_domain * rd)1184 static void rapl_detect_powerlimit(struct rapl_domain *rd)
1185 {
1186 u64 val64;
1187 int i;
1188
1189 /* check if the domain is locked by BIOS, ignore if MSR doesn't exist */
1190 if (!rapl_read_data_raw(rd, FW_LOCK, false, &val64)) {
1191 if (val64) {
1192 pr_info("RAPL %s domain %s locked by BIOS\n",
1193 rd->rp->name, rd->name);
1194 rd->state |= DOMAIN_STATE_BIOS_LOCKED;
1195 }
1196 }
1197 /* check if power limit MSR exists, otherwise domain is monitoring only */
1198 for (i = 0; i < NR_POWER_LIMITS; i++) {
1199 int prim = rd->rpl[i].prim_id;
1200
1201 if (rapl_read_data_raw(rd, prim, false, &val64))
1202 rd->rpl[i].name = NULL;
1203 }
1204 }
1205
1206 /* Detect active and valid domains for the given CPU, caller must
1207 * ensure the CPU belongs to the targeted package and CPU hotlug is disabled.
1208 */
rapl_detect_domains(struct rapl_package * rp,int cpu)1209 static int rapl_detect_domains(struct rapl_package *rp, int cpu)
1210 {
1211 struct rapl_domain *rd;
1212 int i;
1213
1214 for (i = 0; i < RAPL_DOMAIN_MAX; i++) {
1215 /* use physical package id to read counters */
1216 if (!rapl_check_domain(cpu, i, rp)) {
1217 rp->domain_map |= 1 << i;
1218 pr_info("Found RAPL domain %s\n", rapl_domain_names[i]);
1219 }
1220 }
1221 rp->nr_domains = bitmap_weight(&rp->domain_map, RAPL_DOMAIN_MAX);
1222 if (!rp->nr_domains) {
1223 pr_debug("no valid rapl domains found in %s\n", rp->name);
1224 return -ENODEV;
1225 }
1226 pr_debug("found %d domains on %s\n", rp->nr_domains, rp->name);
1227
1228 rp->domains = kcalloc(rp->nr_domains + 1, sizeof(struct rapl_domain),
1229 GFP_KERNEL);
1230 if (!rp->domains)
1231 return -ENOMEM;
1232
1233 rapl_init_domains(rp);
1234
1235 for (rd = rp->domains; rd < rp->domains + rp->nr_domains; rd++)
1236 rapl_detect_powerlimit(rd);
1237
1238 return 0;
1239 }
1240
1241 /* called from CPU hotplug notifier, hotplug lock held */
rapl_remove_package(struct rapl_package * rp)1242 void rapl_remove_package(struct rapl_package *rp)
1243 {
1244 struct rapl_domain *rd, *rd_package = NULL;
1245
1246 package_power_limit_irq_restore(rp);
1247
1248 for (rd = rp->domains; rd < rp->domains + rp->nr_domains; rd++) {
1249 rapl_write_data_raw(rd, PL1_ENABLE, 0);
1250 rapl_write_data_raw(rd, PL1_CLAMP, 0);
1251 if (find_nr_power_limit(rd) > 1) {
1252 rapl_write_data_raw(rd, PL2_ENABLE, 0);
1253 rapl_write_data_raw(rd, PL2_CLAMP, 0);
1254 }
1255 if (rd->id == RAPL_DOMAIN_PACKAGE) {
1256 rd_package = rd;
1257 continue;
1258 }
1259 pr_debug("remove package, undo power limit on %s: %s\n",
1260 rp->name, rd->name);
1261 powercap_unregister_zone(rp->priv->control_type,
1262 &rd->power_zone);
1263 }
1264 /* do parent zone last */
1265 powercap_unregister_zone(rp->priv->control_type,
1266 &rd_package->power_zone);
1267 list_del(&rp->plist);
1268 kfree(rp);
1269 }
1270 EXPORT_SYMBOL_GPL(rapl_remove_package);
1271
1272 /* caller to ensure CPU hotplug lock is held */
rapl_find_package_domain(int cpu,struct rapl_if_priv * priv)1273 struct rapl_package *rapl_find_package_domain(int cpu, struct rapl_if_priv *priv)
1274 {
1275 int id = topology_logical_die_id(cpu);
1276 struct rapl_package *rp;
1277
1278 list_for_each_entry(rp, &rapl_packages, plist) {
1279 if (rp->id == id
1280 && rp->priv->control_type == priv->control_type)
1281 return rp;
1282 }
1283
1284 return NULL;
1285 }
1286 EXPORT_SYMBOL_GPL(rapl_find_package_domain);
1287
1288 /* called from CPU hotplug notifier, hotplug lock held */
rapl_add_package(int cpu,struct rapl_if_priv * priv)1289 struct rapl_package *rapl_add_package(int cpu, struct rapl_if_priv *priv)
1290 {
1291 int id = topology_logical_die_id(cpu);
1292 struct rapl_package *rp;
1293 struct cpuinfo_x86 *c = &cpu_data(cpu);
1294 int ret;
1295
1296 rp = kzalloc(sizeof(struct rapl_package), GFP_KERNEL);
1297 if (!rp)
1298 return ERR_PTR(-ENOMEM);
1299
1300 /* add the new package to the list */
1301 rp->id = id;
1302 rp->lead_cpu = cpu;
1303 rp->priv = priv;
1304
1305 if (topology_max_die_per_package() > 1)
1306 snprintf(rp->name, PACKAGE_DOMAIN_NAME_LENGTH,
1307 "package-%d-die-%d", c->phys_proc_id, c->cpu_die_id);
1308 else
1309 snprintf(rp->name, PACKAGE_DOMAIN_NAME_LENGTH, "package-%d",
1310 c->phys_proc_id);
1311
1312 /* check if the package contains valid domains */
1313 if (rapl_detect_domains(rp, cpu) || rapl_defaults->check_unit(rp, cpu)) {
1314 ret = -ENODEV;
1315 goto err_free_package;
1316 }
1317 ret = rapl_package_register_powercap(rp);
1318 if (!ret) {
1319 INIT_LIST_HEAD(&rp->plist);
1320 list_add(&rp->plist, &rapl_packages);
1321 return rp;
1322 }
1323
1324 err_free_package:
1325 kfree(rp->domains);
1326 kfree(rp);
1327 return ERR_PTR(ret);
1328 }
1329 EXPORT_SYMBOL_GPL(rapl_add_package);
1330
power_limit_state_save(void)1331 static void power_limit_state_save(void)
1332 {
1333 struct rapl_package *rp;
1334 struct rapl_domain *rd;
1335 int nr_pl, ret, i;
1336
1337 get_online_cpus();
1338 list_for_each_entry(rp, &rapl_packages, plist) {
1339 if (!rp->power_zone)
1340 continue;
1341 rd = power_zone_to_rapl_domain(rp->power_zone);
1342 nr_pl = find_nr_power_limit(rd);
1343 for (i = 0; i < nr_pl; i++) {
1344 switch (rd->rpl[i].prim_id) {
1345 case PL1_ENABLE:
1346 ret = rapl_read_data_raw(rd,
1347 POWER_LIMIT1, true,
1348 &rd->rpl[i].last_power_limit);
1349 if (ret)
1350 rd->rpl[i].last_power_limit = 0;
1351 break;
1352 case PL2_ENABLE:
1353 ret = rapl_read_data_raw(rd,
1354 POWER_LIMIT2, true,
1355 &rd->rpl[i].last_power_limit);
1356 if (ret)
1357 rd->rpl[i].last_power_limit = 0;
1358 break;
1359 }
1360 }
1361 }
1362 put_online_cpus();
1363 }
1364
power_limit_state_restore(void)1365 static void power_limit_state_restore(void)
1366 {
1367 struct rapl_package *rp;
1368 struct rapl_domain *rd;
1369 int nr_pl, i;
1370
1371 get_online_cpus();
1372 list_for_each_entry(rp, &rapl_packages, plist) {
1373 if (!rp->power_zone)
1374 continue;
1375 rd = power_zone_to_rapl_domain(rp->power_zone);
1376 nr_pl = find_nr_power_limit(rd);
1377 for (i = 0; i < nr_pl; i++) {
1378 switch (rd->rpl[i].prim_id) {
1379 case PL1_ENABLE:
1380 if (rd->rpl[i].last_power_limit)
1381 rapl_write_data_raw(rd, POWER_LIMIT1,
1382 rd->rpl[i].last_power_limit);
1383 break;
1384 case PL2_ENABLE:
1385 if (rd->rpl[i].last_power_limit)
1386 rapl_write_data_raw(rd, POWER_LIMIT2,
1387 rd->rpl[i].last_power_limit);
1388 break;
1389 }
1390 }
1391 }
1392 put_online_cpus();
1393 }
1394
rapl_pm_callback(struct notifier_block * nb,unsigned long mode,void * _unused)1395 static int rapl_pm_callback(struct notifier_block *nb,
1396 unsigned long mode, void *_unused)
1397 {
1398 switch (mode) {
1399 case PM_SUSPEND_PREPARE:
1400 power_limit_state_save();
1401 break;
1402 case PM_POST_SUSPEND:
1403 power_limit_state_restore();
1404 break;
1405 }
1406 return NOTIFY_OK;
1407 }
1408
1409 static struct notifier_block rapl_pm_notifier = {
1410 .notifier_call = rapl_pm_callback,
1411 };
1412
1413 static struct platform_device *rapl_msr_platdev;
1414
rapl_init(void)1415 static int __init rapl_init(void)
1416 {
1417 const struct x86_cpu_id *id;
1418 int ret;
1419
1420 id = x86_match_cpu(rapl_ids);
1421 if (!id) {
1422 pr_err("driver does not support CPU family %d model %d\n",
1423 boot_cpu_data.x86, boot_cpu_data.x86_model);
1424
1425 return -ENODEV;
1426 }
1427
1428 rapl_defaults = (struct rapl_defaults *)id->driver_data;
1429
1430 ret = register_pm_notifier(&rapl_pm_notifier);
1431 if (ret)
1432 return ret;
1433
1434 rapl_msr_platdev = platform_device_alloc("intel_rapl_msr", 0);
1435 if (!rapl_msr_platdev) {
1436 ret = -ENOMEM;
1437 goto end;
1438 }
1439
1440 ret = platform_device_add(rapl_msr_platdev);
1441 if (ret)
1442 platform_device_put(rapl_msr_platdev);
1443
1444 end:
1445 if (ret)
1446 unregister_pm_notifier(&rapl_pm_notifier);
1447
1448 return ret;
1449 }
1450
rapl_exit(void)1451 static void __exit rapl_exit(void)
1452 {
1453 platform_device_unregister(rapl_msr_platdev);
1454 unregister_pm_notifier(&rapl_pm_notifier);
1455 }
1456
1457 fs_initcall(rapl_init);
1458 module_exit(rapl_exit);
1459
1460 MODULE_DESCRIPTION("Intel Runtime Average Power Limit (RAPL) common code");
1461 MODULE_AUTHOR("Jacob Pan <jacob.jun.pan@intel.com>");
1462 MODULE_LICENSE("GPL v2");
1463
