xref: /linux/drivers/cpufreq/intel_pstate.c (revision 4407f967441aa1adfc11f739e8e9ec0f38fa839f)
1 /*
2  * intel_pstate.c: Native P state management for Intel processors
3  *
4  * (C) Copyright 2012 Intel Corporation
5  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public License
9  * as published by the Free Software Foundation; version 2
10  * of the License.
11  */
12 
13 #include <linux/kernel.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/module.h>
16 #include <linux/ktime.h>
17 #include <linux/hrtimer.h>
18 #include <linux/tick.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/list.h>
22 #include <linux/cpu.h>
23 #include <linux/cpufreq.h>
24 #include <linux/sysfs.h>
25 #include <linux/types.h>
26 #include <linux/fs.h>
27 #include <linux/debugfs.h>
28 #include <linux/acpi.h>
29 #include <linux/vmalloc.h>
30 #include <trace/events/power.h>
31 
32 #include <asm/div64.h>
33 #include <asm/msr.h>
34 #include <asm/cpu_device_id.h>
35 #include <asm/cpufeature.h>
36 
37 #define BYT_RATIOS		0x66a
38 #define BYT_VIDS		0x66b
39 #define BYT_TURBO_RATIOS	0x66c
40 #define BYT_TURBO_VIDS		0x66d
41 
42 #define FRAC_BITS 8
43 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
44 #define fp_toint(X) ((X) >> FRAC_BITS)
45 
46 
47 static inline int32_t mul_fp(int32_t x, int32_t y)
48 {
49 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
50 }
51 
52 static inline int32_t div_fp(int32_t x, int32_t y)
53 {
54 	return div_s64((int64_t)x << FRAC_BITS, y);
55 }
56 
57 static inline int ceiling_fp(int32_t x)
58 {
59 	int mask, ret;
60 
61 	ret = fp_toint(x);
62 	mask = (1 << FRAC_BITS) - 1;
63 	if (x & mask)
64 		ret += 1;
65 	return ret;
66 }
67 
68 struct sample {
69 	int32_t core_pct_busy;
70 	u64 aperf;
71 	u64 mperf;
72 	int freq;
73 	ktime_t time;
74 };
75 
76 struct pstate_data {
77 	int	current_pstate;
78 	int	min_pstate;
79 	int	max_pstate;
80 	int	scaling;
81 	int	turbo_pstate;
82 };
83 
84 struct vid_data {
85 	int min;
86 	int max;
87 	int turbo;
88 	int32_t ratio;
89 };
90 
91 struct _pid {
92 	int setpoint;
93 	int32_t integral;
94 	int32_t p_gain;
95 	int32_t i_gain;
96 	int32_t d_gain;
97 	int deadband;
98 	int32_t last_err;
99 };
100 
101 struct cpudata {
102 	int cpu;
103 
104 	struct timer_list timer;
105 
106 	struct pstate_data pstate;
107 	struct vid_data vid;
108 	struct _pid pid;
109 
110 	ktime_t last_sample_time;
111 	u64	prev_aperf;
112 	u64	prev_mperf;
113 	struct sample sample;
114 };
115 
116 static struct cpudata **all_cpu_data;
117 struct pstate_adjust_policy {
118 	int sample_rate_ms;
119 	int deadband;
120 	int setpoint;
121 	int p_gain_pct;
122 	int d_gain_pct;
123 	int i_gain_pct;
124 };
125 
126 struct pstate_funcs {
127 	int (*get_max)(void);
128 	int (*get_min)(void);
129 	int (*get_turbo)(void);
130 	int (*get_scaling)(void);
131 	void (*set)(struct cpudata*, int pstate);
132 	void (*get_vid)(struct cpudata *);
133 };
134 
135 struct cpu_defaults {
136 	struct pstate_adjust_policy pid_policy;
137 	struct pstate_funcs funcs;
138 };
139 
140 static struct pstate_adjust_policy pid_params;
141 static struct pstate_funcs pstate_funcs;
142 static int hwp_active;
143 
144 struct perf_limits {
145 	int no_turbo;
146 	int turbo_disabled;
147 	int max_perf_pct;
148 	int min_perf_pct;
149 	int32_t max_perf;
150 	int32_t min_perf;
151 	int max_policy_pct;
152 	int max_sysfs_pct;
153 	int min_policy_pct;
154 	int min_sysfs_pct;
155 };
156 
157 static struct perf_limits limits = {
158 	.no_turbo = 0,
159 	.turbo_disabled = 0,
160 	.max_perf_pct = 100,
161 	.max_perf = int_tofp(1),
162 	.min_perf_pct = 0,
163 	.min_perf = 0,
164 	.max_policy_pct = 100,
165 	.max_sysfs_pct = 100,
166 	.min_policy_pct = 0,
167 	.min_sysfs_pct = 0,
168 };
169 
170 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
171 			     int deadband, int integral) {
172 	pid->setpoint = setpoint;
173 	pid->deadband  = deadband;
174 	pid->integral  = int_tofp(integral);
175 	pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
176 }
177 
178 static inline void pid_p_gain_set(struct _pid *pid, int percent)
179 {
180 	pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
181 }
182 
183 static inline void pid_i_gain_set(struct _pid *pid, int percent)
184 {
185 	pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
186 }
187 
188 static inline void pid_d_gain_set(struct _pid *pid, int percent)
189 {
190 	pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
191 }
192 
193 static signed int pid_calc(struct _pid *pid, int32_t busy)
194 {
195 	signed int result;
196 	int32_t pterm, dterm, fp_error;
197 	int32_t integral_limit;
198 
199 	fp_error = int_tofp(pid->setpoint) - busy;
200 
201 	if (abs(fp_error) <= int_tofp(pid->deadband))
202 		return 0;
203 
204 	pterm = mul_fp(pid->p_gain, fp_error);
205 
206 	pid->integral += fp_error;
207 
208 	/*
209 	 * We limit the integral here so that it will never
210 	 * get higher than 30.  This prevents it from becoming
211 	 * too large an input over long periods of time and allows
212 	 * it to get factored out sooner.
213 	 *
214 	 * The value of 30 was chosen through experimentation.
215 	 */
216 	integral_limit = int_tofp(30);
217 	if (pid->integral > integral_limit)
218 		pid->integral = integral_limit;
219 	if (pid->integral < -integral_limit)
220 		pid->integral = -integral_limit;
221 
222 	dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
223 	pid->last_err = fp_error;
224 
225 	result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
226 	result = result + (1 << (FRAC_BITS-1));
227 	return (signed int)fp_toint(result);
228 }
229 
230 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
231 {
232 	pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
233 	pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
234 	pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
235 
236 	pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
237 }
238 
239 static inline void intel_pstate_reset_all_pid(void)
240 {
241 	unsigned int cpu;
242 
243 	for_each_online_cpu(cpu) {
244 		if (all_cpu_data[cpu])
245 			intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
246 	}
247 }
248 
249 static inline void update_turbo_state(void)
250 {
251 	u64 misc_en;
252 	struct cpudata *cpu;
253 
254 	cpu = all_cpu_data[0];
255 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
256 	limits.turbo_disabled =
257 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
258 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
259 }
260 
261 #define PCT_TO_HWP(x) (x * 255 / 100)
262 static void intel_pstate_hwp_set(void)
263 {
264 	int min, max, cpu;
265 	u64 value, freq;
266 
267 	get_online_cpus();
268 
269 	for_each_online_cpu(cpu) {
270 		rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
271 		min = PCT_TO_HWP(limits.min_perf_pct);
272 		value &= ~HWP_MIN_PERF(~0L);
273 		value |= HWP_MIN_PERF(min);
274 
275 		max = PCT_TO_HWP(limits.max_perf_pct);
276 		if (limits.no_turbo) {
277 			rdmsrl( MSR_HWP_CAPABILITIES, freq);
278 			max = HWP_GUARANTEED_PERF(freq);
279 		}
280 
281 		value &= ~HWP_MAX_PERF(~0L);
282 		value |= HWP_MAX_PERF(max);
283 		wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
284 	}
285 
286 	put_online_cpus();
287 }
288 
289 /************************** debugfs begin ************************/
290 static int pid_param_set(void *data, u64 val)
291 {
292 	*(u32 *)data = val;
293 	intel_pstate_reset_all_pid();
294 	return 0;
295 }
296 
297 static int pid_param_get(void *data, u64 *val)
298 {
299 	*val = *(u32 *)data;
300 	return 0;
301 }
302 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
303 
304 struct pid_param {
305 	char *name;
306 	void *value;
307 };
308 
309 static struct pid_param pid_files[] = {
310 	{"sample_rate_ms", &pid_params.sample_rate_ms},
311 	{"d_gain_pct", &pid_params.d_gain_pct},
312 	{"i_gain_pct", &pid_params.i_gain_pct},
313 	{"deadband", &pid_params.deadband},
314 	{"setpoint", &pid_params.setpoint},
315 	{"p_gain_pct", &pid_params.p_gain_pct},
316 	{NULL, NULL}
317 };
318 
319 static void __init intel_pstate_debug_expose_params(void)
320 {
321 	struct dentry *debugfs_parent;
322 	int i = 0;
323 
324 	if (hwp_active)
325 		return;
326 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
327 	if (IS_ERR_OR_NULL(debugfs_parent))
328 		return;
329 	while (pid_files[i].name) {
330 		debugfs_create_file(pid_files[i].name, 0660,
331 				    debugfs_parent, pid_files[i].value,
332 				    &fops_pid_param);
333 		i++;
334 	}
335 }
336 
337 /************************** debugfs end ************************/
338 
339 /************************** sysfs begin ************************/
340 #define show_one(file_name, object)					\
341 	static ssize_t show_##file_name					\
342 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
343 	{								\
344 		return sprintf(buf, "%u\n", limits.object);		\
345 	}
346 
347 static ssize_t show_turbo_pct(struct kobject *kobj,
348 				struct attribute *attr, char *buf)
349 {
350 	struct cpudata *cpu;
351 	int total, no_turbo, turbo_pct;
352 	uint32_t turbo_fp;
353 
354 	cpu = all_cpu_data[0];
355 
356 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
357 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
358 	turbo_fp = div_fp(int_tofp(no_turbo), int_tofp(total));
359 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
360 	return sprintf(buf, "%u\n", turbo_pct);
361 }
362 
363 static ssize_t show_num_pstates(struct kobject *kobj,
364 				struct attribute *attr, char *buf)
365 {
366 	struct cpudata *cpu;
367 	int total;
368 
369 	cpu = all_cpu_data[0];
370 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
371 	return sprintf(buf, "%u\n", total);
372 }
373 
374 static ssize_t show_no_turbo(struct kobject *kobj,
375 			     struct attribute *attr, char *buf)
376 {
377 	ssize_t ret;
378 
379 	update_turbo_state();
380 	if (limits.turbo_disabled)
381 		ret = sprintf(buf, "%u\n", limits.turbo_disabled);
382 	else
383 		ret = sprintf(buf, "%u\n", limits.no_turbo);
384 
385 	return ret;
386 }
387 
388 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
389 			      const char *buf, size_t count)
390 {
391 	unsigned int input;
392 	int ret;
393 
394 	ret = sscanf(buf, "%u", &input);
395 	if (ret != 1)
396 		return -EINVAL;
397 
398 	update_turbo_state();
399 	if (limits.turbo_disabled) {
400 		pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
401 		return -EPERM;
402 	}
403 
404 	limits.no_turbo = clamp_t(int, input, 0, 1);
405 
406 	if (hwp_active)
407 		intel_pstate_hwp_set();
408 
409 	return count;
410 }
411 
412 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
413 				  const char *buf, size_t count)
414 {
415 	unsigned int input;
416 	int ret;
417 
418 	ret = sscanf(buf, "%u", &input);
419 	if (ret != 1)
420 		return -EINVAL;
421 
422 	limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
423 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
424 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
425 
426 	if (hwp_active)
427 		intel_pstate_hwp_set();
428 	return count;
429 }
430 
431 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
432 				  const char *buf, size_t count)
433 {
434 	unsigned int input;
435 	int ret;
436 
437 	ret = sscanf(buf, "%u", &input);
438 	if (ret != 1)
439 		return -EINVAL;
440 
441 	limits.min_sysfs_pct = clamp_t(int, input, 0 , 100);
442 	limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
443 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
444 
445 	if (hwp_active)
446 		intel_pstate_hwp_set();
447 	return count;
448 }
449 
450 show_one(max_perf_pct, max_perf_pct);
451 show_one(min_perf_pct, min_perf_pct);
452 
453 define_one_global_rw(no_turbo);
454 define_one_global_rw(max_perf_pct);
455 define_one_global_rw(min_perf_pct);
456 define_one_global_ro(turbo_pct);
457 define_one_global_ro(num_pstates);
458 
459 static struct attribute *intel_pstate_attributes[] = {
460 	&no_turbo.attr,
461 	&max_perf_pct.attr,
462 	&min_perf_pct.attr,
463 	&turbo_pct.attr,
464 	&num_pstates.attr,
465 	NULL
466 };
467 
468 static struct attribute_group intel_pstate_attr_group = {
469 	.attrs = intel_pstate_attributes,
470 };
471 
472 static void __init intel_pstate_sysfs_expose_params(void)
473 {
474 	struct kobject *intel_pstate_kobject;
475 	int rc;
476 
477 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
478 						&cpu_subsys.dev_root->kobj);
479 	BUG_ON(!intel_pstate_kobject);
480 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
481 	BUG_ON(rc);
482 }
483 /************************** sysfs end ************************/
484 
485 static void intel_pstate_hwp_enable(void)
486 {
487 	hwp_active++;
488 	pr_info("intel_pstate HWP enabled\n");
489 
490 	wrmsrl( MSR_PM_ENABLE, 0x1);
491 }
492 
493 static int byt_get_min_pstate(void)
494 {
495 	u64 value;
496 
497 	rdmsrl(BYT_RATIOS, value);
498 	return (value >> 8) & 0x7F;
499 }
500 
501 static int byt_get_max_pstate(void)
502 {
503 	u64 value;
504 
505 	rdmsrl(BYT_RATIOS, value);
506 	return (value >> 16) & 0x7F;
507 }
508 
509 static int byt_get_turbo_pstate(void)
510 {
511 	u64 value;
512 
513 	rdmsrl(BYT_TURBO_RATIOS, value);
514 	return value & 0x7F;
515 }
516 
517 static void byt_set_pstate(struct cpudata *cpudata, int pstate)
518 {
519 	u64 val;
520 	int32_t vid_fp;
521 	u32 vid;
522 
523 	val = pstate << 8;
524 	if (limits.no_turbo && !limits.turbo_disabled)
525 		val |= (u64)1 << 32;
526 
527 	vid_fp = cpudata->vid.min + mul_fp(
528 		int_tofp(pstate - cpudata->pstate.min_pstate),
529 		cpudata->vid.ratio);
530 
531 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
532 	vid = ceiling_fp(vid_fp);
533 
534 	if (pstate > cpudata->pstate.max_pstate)
535 		vid = cpudata->vid.turbo;
536 
537 	val |= vid;
538 
539 	wrmsrl(MSR_IA32_PERF_CTL, val);
540 }
541 
542 #define BYT_BCLK_FREQS 5
543 static int byt_freq_table[BYT_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800};
544 
545 static int byt_get_scaling(void)
546 {
547 	u64 value;
548 	int i;
549 
550 	rdmsrl(MSR_FSB_FREQ, value);
551 	i = value & 0x3;
552 
553 	BUG_ON(i > BYT_BCLK_FREQS);
554 
555 	return byt_freq_table[i] * 100;
556 }
557 
558 static void byt_get_vid(struct cpudata *cpudata)
559 {
560 	u64 value;
561 
562 	rdmsrl(BYT_VIDS, value);
563 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
564 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
565 	cpudata->vid.ratio = div_fp(
566 		cpudata->vid.max - cpudata->vid.min,
567 		int_tofp(cpudata->pstate.max_pstate -
568 			cpudata->pstate.min_pstate));
569 
570 	rdmsrl(BYT_TURBO_VIDS, value);
571 	cpudata->vid.turbo = value & 0x7f;
572 }
573 
574 static int core_get_min_pstate(void)
575 {
576 	u64 value;
577 
578 	rdmsrl(MSR_PLATFORM_INFO, value);
579 	return (value >> 40) & 0xFF;
580 }
581 
582 static int core_get_max_pstate(void)
583 {
584 	u64 value;
585 
586 	rdmsrl(MSR_PLATFORM_INFO, value);
587 	return (value >> 8) & 0xFF;
588 }
589 
590 static int core_get_turbo_pstate(void)
591 {
592 	u64 value;
593 	int nont, ret;
594 
595 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
596 	nont = core_get_max_pstate();
597 	ret = (value) & 255;
598 	if (ret <= nont)
599 		ret = nont;
600 	return ret;
601 }
602 
603 static inline int core_get_scaling(void)
604 {
605 	return 100000;
606 }
607 
608 static void core_set_pstate(struct cpudata *cpudata, int pstate)
609 {
610 	u64 val;
611 
612 	val = pstate << 8;
613 	if (limits.no_turbo && !limits.turbo_disabled)
614 		val |= (u64)1 << 32;
615 
616 	wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
617 }
618 
619 static int knl_get_turbo_pstate(void)
620 {
621 	u64 value;
622 	int nont, ret;
623 
624 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
625 	nont = core_get_max_pstate();
626 	ret = (((value) >> 8) & 0xFF);
627 	if (ret <= nont)
628 		ret = nont;
629 	return ret;
630 }
631 
632 static struct cpu_defaults core_params = {
633 	.pid_policy = {
634 		.sample_rate_ms = 10,
635 		.deadband = 0,
636 		.setpoint = 97,
637 		.p_gain_pct = 20,
638 		.d_gain_pct = 0,
639 		.i_gain_pct = 0,
640 	},
641 	.funcs = {
642 		.get_max = core_get_max_pstate,
643 		.get_min = core_get_min_pstate,
644 		.get_turbo = core_get_turbo_pstate,
645 		.get_scaling = core_get_scaling,
646 		.set = core_set_pstate,
647 	},
648 };
649 
650 static struct cpu_defaults byt_params = {
651 	.pid_policy = {
652 		.sample_rate_ms = 10,
653 		.deadband = 0,
654 		.setpoint = 60,
655 		.p_gain_pct = 14,
656 		.d_gain_pct = 0,
657 		.i_gain_pct = 4,
658 	},
659 	.funcs = {
660 		.get_max = byt_get_max_pstate,
661 		.get_min = byt_get_min_pstate,
662 		.get_turbo = byt_get_turbo_pstate,
663 		.set = byt_set_pstate,
664 		.get_scaling = byt_get_scaling,
665 		.get_vid = byt_get_vid,
666 	},
667 };
668 
669 static struct cpu_defaults knl_params = {
670 	.pid_policy = {
671 		.sample_rate_ms = 10,
672 		.deadband = 0,
673 		.setpoint = 97,
674 		.p_gain_pct = 20,
675 		.d_gain_pct = 0,
676 		.i_gain_pct = 0,
677 	},
678 	.funcs = {
679 		.get_max = core_get_max_pstate,
680 		.get_min = core_get_min_pstate,
681 		.get_turbo = knl_get_turbo_pstate,
682 		.set = core_set_pstate,
683 	},
684 };
685 
686 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
687 {
688 	int max_perf = cpu->pstate.turbo_pstate;
689 	int max_perf_adj;
690 	int min_perf;
691 
692 	if (limits.no_turbo || limits.turbo_disabled)
693 		max_perf = cpu->pstate.max_pstate;
694 
695 	/*
696 	 * performance can be limited by user through sysfs, by cpufreq
697 	 * policy, or by cpu specific default values determined through
698 	 * experimentation.
699 	 */
700 	max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
701 	*max = clamp_t(int, max_perf_adj,
702 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
703 
704 	min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
705 	*min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
706 }
707 
708 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
709 {
710 	int max_perf, min_perf;
711 
712 	update_turbo_state();
713 
714 	intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
715 
716 	pstate = clamp_t(int, pstate, min_perf, max_perf);
717 
718 	if (pstate == cpu->pstate.current_pstate)
719 		return;
720 
721 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
722 
723 	cpu->pstate.current_pstate = pstate;
724 
725 	pstate_funcs.set(cpu, pstate);
726 }
727 
728 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
729 {
730 	cpu->pstate.min_pstate = pstate_funcs.get_min();
731 	cpu->pstate.max_pstate = pstate_funcs.get_max();
732 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
733 	cpu->pstate.scaling = pstate_funcs.get_scaling();
734 
735 	if (pstate_funcs.get_vid)
736 		pstate_funcs.get_vid(cpu);
737 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
738 }
739 
740 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
741 {
742 	struct sample *sample = &cpu->sample;
743 	int64_t core_pct;
744 
745 	core_pct = int_tofp(sample->aperf) * int_tofp(100);
746 	core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
747 
748 	sample->freq = fp_toint(
749 		mul_fp(int_tofp(
750 			cpu->pstate.max_pstate * cpu->pstate.scaling / 100),
751 			core_pct));
752 
753 	sample->core_pct_busy = (int32_t)core_pct;
754 }
755 
756 static inline void intel_pstate_sample(struct cpudata *cpu)
757 {
758 	u64 aperf, mperf;
759 	unsigned long flags;
760 
761 	local_irq_save(flags);
762 	rdmsrl(MSR_IA32_APERF, aperf);
763 	rdmsrl(MSR_IA32_MPERF, mperf);
764 	local_irq_restore(flags);
765 
766 	cpu->last_sample_time = cpu->sample.time;
767 	cpu->sample.time = ktime_get();
768 	cpu->sample.aperf = aperf;
769 	cpu->sample.mperf = mperf;
770 	cpu->sample.aperf -= cpu->prev_aperf;
771 	cpu->sample.mperf -= cpu->prev_mperf;
772 
773 	intel_pstate_calc_busy(cpu);
774 
775 	cpu->prev_aperf = aperf;
776 	cpu->prev_mperf = mperf;
777 }
778 
779 static inline void intel_hwp_set_sample_time(struct cpudata *cpu)
780 {
781 	int delay;
782 
783 	delay = msecs_to_jiffies(50);
784 	mod_timer_pinned(&cpu->timer, jiffies + delay);
785 }
786 
787 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
788 {
789 	int delay;
790 
791 	delay = msecs_to_jiffies(pid_params.sample_rate_ms);
792 	mod_timer_pinned(&cpu->timer, jiffies + delay);
793 }
794 
795 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
796 {
797 	int32_t core_busy, max_pstate, current_pstate, sample_ratio;
798 	u32 duration_us;
799 	u32 sample_time;
800 
801 	/*
802 	 * core_busy is the ratio of actual performance to max
803 	 * max_pstate is the max non turbo pstate available
804 	 * current_pstate was the pstate that was requested during
805 	 * 	the last sample period.
806 	 *
807 	 * We normalize core_busy, which was our actual percent
808 	 * performance to what we requested during the last sample
809 	 * period. The result will be a percentage of busy at a
810 	 * specified pstate.
811 	 */
812 	core_busy = cpu->sample.core_pct_busy;
813 	max_pstate = int_tofp(cpu->pstate.max_pstate);
814 	current_pstate = int_tofp(cpu->pstate.current_pstate);
815 	core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
816 
817 	/*
818 	 * Since we have a deferred timer, it will not fire unless
819 	 * we are in C0.  So, determine if the actual elapsed time
820 	 * is significantly greater (3x) than our sample interval.  If it
821 	 * is, then we were idle for a long enough period of time
822 	 * to adjust our busyness.
823 	 */
824 	sample_time = pid_params.sample_rate_ms  * USEC_PER_MSEC;
825 	duration_us = (u32) ktime_us_delta(cpu->sample.time,
826 					   cpu->last_sample_time);
827 	if (duration_us > sample_time * 3) {
828 		sample_ratio = div_fp(int_tofp(sample_time),
829 				      int_tofp(duration_us));
830 		core_busy = mul_fp(core_busy, sample_ratio);
831 	}
832 
833 	return core_busy;
834 }
835 
836 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
837 {
838 	int32_t busy_scaled;
839 	struct _pid *pid;
840 	signed int ctl;
841 
842 	pid = &cpu->pid;
843 	busy_scaled = intel_pstate_get_scaled_busy(cpu);
844 
845 	ctl = pid_calc(pid, busy_scaled);
846 
847 	/* Negative values of ctl increase the pstate and vice versa */
848 	intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl);
849 }
850 
851 static void intel_hwp_timer_func(unsigned long __data)
852 {
853 	struct cpudata *cpu = (struct cpudata *) __data;
854 
855 	intel_pstate_sample(cpu);
856 	intel_hwp_set_sample_time(cpu);
857 }
858 
859 static void intel_pstate_timer_func(unsigned long __data)
860 {
861 	struct cpudata *cpu = (struct cpudata *) __data;
862 	struct sample *sample;
863 
864 	intel_pstate_sample(cpu);
865 
866 	sample = &cpu->sample;
867 
868 	intel_pstate_adjust_busy_pstate(cpu);
869 
870 	trace_pstate_sample(fp_toint(sample->core_pct_busy),
871 			fp_toint(intel_pstate_get_scaled_busy(cpu)),
872 			cpu->pstate.current_pstate,
873 			sample->mperf,
874 			sample->aperf,
875 			sample->freq);
876 
877 	intel_pstate_set_sample_time(cpu);
878 }
879 
880 #define ICPU(model, policy) \
881 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
882 			(unsigned long)&policy }
883 
884 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
885 	ICPU(0x2a, core_params),
886 	ICPU(0x2d, core_params),
887 	ICPU(0x37, byt_params),
888 	ICPU(0x3a, core_params),
889 	ICPU(0x3c, core_params),
890 	ICPU(0x3d, core_params),
891 	ICPU(0x3e, core_params),
892 	ICPU(0x3f, core_params),
893 	ICPU(0x45, core_params),
894 	ICPU(0x46, core_params),
895 	ICPU(0x47, core_params),
896 	ICPU(0x4c, byt_params),
897 	ICPU(0x4e, core_params),
898 	ICPU(0x4f, core_params),
899 	ICPU(0x56, core_params),
900 	ICPU(0x57, knl_params),
901 	{}
902 };
903 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
904 
905 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
906 	ICPU(0x56, core_params),
907 	{}
908 };
909 
910 static int intel_pstate_init_cpu(unsigned int cpunum)
911 {
912 	struct cpudata *cpu;
913 
914 	if (!all_cpu_data[cpunum])
915 		all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
916 					       GFP_KERNEL);
917 	if (!all_cpu_data[cpunum])
918 		return -ENOMEM;
919 
920 	cpu = all_cpu_data[cpunum];
921 
922 	cpu->cpu = cpunum;
923 	intel_pstate_get_cpu_pstates(cpu);
924 
925 	init_timer_deferrable(&cpu->timer);
926 	cpu->timer.data = (unsigned long)cpu;
927 	cpu->timer.expires = jiffies + HZ/100;
928 
929 	if (!hwp_active)
930 		cpu->timer.function = intel_pstate_timer_func;
931 	else
932 		cpu->timer.function = intel_hwp_timer_func;
933 
934 	intel_pstate_busy_pid_reset(cpu);
935 	intel_pstate_sample(cpu);
936 
937 	add_timer_on(&cpu->timer, cpunum);
938 
939 	pr_debug("Intel pstate controlling: cpu %d\n", cpunum);
940 
941 	return 0;
942 }
943 
944 static unsigned int intel_pstate_get(unsigned int cpu_num)
945 {
946 	struct sample *sample;
947 	struct cpudata *cpu;
948 
949 	cpu = all_cpu_data[cpu_num];
950 	if (!cpu)
951 		return 0;
952 	sample = &cpu->sample;
953 	return sample->freq;
954 }
955 
956 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
957 {
958 	if (!policy->cpuinfo.max_freq)
959 		return -ENODEV;
960 
961 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE &&
962 	    policy->max >= policy->cpuinfo.max_freq) {
963 		limits.min_policy_pct = 100;
964 		limits.min_perf_pct = 100;
965 		limits.min_perf = int_tofp(1);
966 		limits.max_policy_pct = 100;
967 		limits.max_perf_pct = 100;
968 		limits.max_perf = int_tofp(1);
969 		limits.no_turbo = 0;
970 		return 0;
971 	}
972 
973 	limits.min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
974 	limits.min_policy_pct = clamp_t(int, limits.min_policy_pct, 0 , 100);
975 	limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
976 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
977 
978 	limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq;
979 	limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
980 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
981 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
982 
983 	if (hwp_active)
984 		intel_pstate_hwp_set();
985 
986 	return 0;
987 }
988 
989 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
990 {
991 	cpufreq_verify_within_cpu_limits(policy);
992 
993 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
994 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
995 		return -EINVAL;
996 
997 	return 0;
998 }
999 
1000 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1001 {
1002 	int cpu_num = policy->cpu;
1003 	struct cpudata *cpu = all_cpu_data[cpu_num];
1004 
1005 	pr_info("intel_pstate CPU %d exiting\n", cpu_num);
1006 
1007 	del_timer_sync(&all_cpu_data[cpu_num]->timer);
1008 	if (hwp_active)
1009 		return;
1010 
1011 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1012 }
1013 
1014 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1015 {
1016 	struct cpudata *cpu;
1017 	int rc;
1018 
1019 	rc = intel_pstate_init_cpu(policy->cpu);
1020 	if (rc)
1021 		return rc;
1022 
1023 	cpu = all_cpu_data[policy->cpu];
1024 
1025 	if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
1026 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1027 	else
1028 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
1029 
1030 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1031 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1032 
1033 	/* cpuinfo and default policy values */
1034 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1035 	policy->cpuinfo.max_freq =
1036 		cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1037 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1038 	cpumask_set_cpu(policy->cpu, policy->cpus);
1039 
1040 	return 0;
1041 }
1042 
1043 static struct cpufreq_driver intel_pstate_driver = {
1044 	.flags		= CPUFREQ_CONST_LOOPS,
1045 	.verify		= intel_pstate_verify_policy,
1046 	.setpolicy	= intel_pstate_set_policy,
1047 	.get		= intel_pstate_get,
1048 	.init		= intel_pstate_cpu_init,
1049 	.stop_cpu	= intel_pstate_stop_cpu,
1050 	.name		= "intel_pstate",
1051 };
1052 
1053 static int __initdata no_load;
1054 static int __initdata no_hwp;
1055 static int __initdata hwp_only;
1056 static unsigned int force_load;
1057 
1058 static int intel_pstate_msrs_not_valid(void)
1059 {
1060 	if (!pstate_funcs.get_max() ||
1061 	    !pstate_funcs.get_min() ||
1062 	    !pstate_funcs.get_turbo())
1063 		return -ENODEV;
1064 
1065 	return 0;
1066 }
1067 
1068 static void copy_pid_params(struct pstate_adjust_policy *policy)
1069 {
1070 	pid_params.sample_rate_ms = policy->sample_rate_ms;
1071 	pid_params.p_gain_pct = policy->p_gain_pct;
1072 	pid_params.i_gain_pct = policy->i_gain_pct;
1073 	pid_params.d_gain_pct = policy->d_gain_pct;
1074 	pid_params.deadband = policy->deadband;
1075 	pid_params.setpoint = policy->setpoint;
1076 }
1077 
1078 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1079 {
1080 	pstate_funcs.get_max   = funcs->get_max;
1081 	pstate_funcs.get_min   = funcs->get_min;
1082 	pstate_funcs.get_turbo = funcs->get_turbo;
1083 	pstate_funcs.get_scaling = funcs->get_scaling;
1084 	pstate_funcs.set       = funcs->set;
1085 	pstate_funcs.get_vid   = funcs->get_vid;
1086 }
1087 
1088 #if IS_ENABLED(CONFIG_ACPI)
1089 #include <acpi/processor.h>
1090 
1091 static bool intel_pstate_no_acpi_pss(void)
1092 {
1093 	int i;
1094 
1095 	for_each_possible_cpu(i) {
1096 		acpi_status status;
1097 		union acpi_object *pss;
1098 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1099 		struct acpi_processor *pr = per_cpu(processors, i);
1100 
1101 		if (!pr)
1102 			continue;
1103 
1104 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1105 		if (ACPI_FAILURE(status))
1106 			continue;
1107 
1108 		pss = buffer.pointer;
1109 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1110 			kfree(pss);
1111 			return false;
1112 		}
1113 
1114 		kfree(pss);
1115 	}
1116 
1117 	return true;
1118 }
1119 
1120 static bool intel_pstate_has_acpi_ppc(void)
1121 {
1122 	int i;
1123 
1124 	for_each_possible_cpu(i) {
1125 		struct acpi_processor *pr = per_cpu(processors, i);
1126 
1127 		if (!pr)
1128 			continue;
1129 		if (acpi_has_method(pr->handle, "_PPC"))
1130 			return true;
1131 	}
1132 	return false;
1133 }
1134 
1135 enum {
1136 	PSS,
1137 	PPC,
1138 };
1139 
1140 struct hw_vendor_info {
1141 	u16  valid;
1142 	char oem_id[ACPI_OEM_ID_SIZE];
1143 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1144 	int  oem_pwr_table;
1145 };
1146 
1147 /* Hardware vendor-specific info that has its own power management modes */
1148 static struct hw_vendor_info vendor_info[] = {
1149 	{1, "HP    ", "ProLiant", PSS},
1150 	{1, "ORACLE", "X4-2    ", PPC},
1151 	{1, "ORACLE", "X4-2L   ", PPC},
1152 	{1, "ORACLE", "X4-2B   ", PPC},
1153 	{1, "ORACLE", "X3-2    ", PPC},
1154 	{1, "ORACLE", "X3-2L   ", PPC},
1155 	{1, "ORACLE", "X3-2B   ", PPC},
1156 	{1, "ORACLE", "X4470M2 ", PPC},
1157 	{1, "ORACLE", "X4270M3 ", PPC},
1158 	{1, "ORACLE", "X4270M2 ", PPC},
1159 	{1, "ORACLE", "X4170M2 ", PPC},
1160 	{0, "", ""},
1161 };
1162 
1163 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1164 {
1165 	struct acpi_table_header hdr;
1166 	struct hw_vendor_info *v_info;
1167 	const struct x86_cpu_id *id;
1168 	u64 misc_pwr;
1169 
1170 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1171 	if (id) {
1172 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1173 		if ( misc_pwr & (1 << 8))
1174 			return true;
1175 	}
1176 
1177 	if (acpi_disabled ||
1178 	    ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1179 		return false;
1180 
1181 	for (v_info = vendor_info; v_info->valid; v_info++) {
1182 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1183 			!strncmp(hdr.oem_table_id, v_info->oem_table_id,
1184 						ACPI_OEM_TABLE_ID_SIZE))
1185 			switch (v_info->oem_pwr_table) {
1186 			case PSS:
1187 				return intel_pstate_no_acpi_pss();
1188 			case PPC:
1189 				return intel_pstate_has_acpi_ppc() &&
1190 					(!force_load);
1191 			}
1192 	}
1193 
1194 	return false;
1195 }
1196 #else /* CONFIG_ACPI not enabled */
1197 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1198 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1199 #endif /* CONFIG_ACPI */
1200 
1201 static int __init intel_pstate_init(void)
1202 {
1203 	int cpu, rc = 0;
1204 	const struct x86_cpu_id *id;
1205 	struct cpu_defaults *cpu_def;
1206 
1207 	if (no_load)
1208 		return -ENODEV;
1209 
1210 	id = x86_match_cpu(intel_pstate_cpu_ids);
1211 	if (!id)
1212 		return -ENODEV;
1213 
1214 	/*
1215 	 * The Intel pstate driver will be ignored if the platform
1216 	 * firmware has its own power management modes.
1217 	 */
1218 	if (intel_pstate_platform_pwr_mgmt_exists())
1219 		return -ENODEV;
1220 
1221 	cpu_def = (struct cpu_defaults *)id->driver_data;
1222 
1223 	copy_pid_params(&cpu_def->pid_policy);
1224 	copy_cpu_funcs(&cpu_def->funcs);
1225 
1226 	if (intel_pstate_msrs_not_valid())
1227 		return -ENODEV;
1228 
1229 	pr_info("Intel P-state driver initializing.\n");
1230 
1231 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1232 	if (!all_cpu_data)
1233 		return -ENOMEM;
1234 
1235 	if (static_cpu_has_safe(X86_FEATURE_HWP) && !no_hwp)
1236 		intel_pstate_hwp_enable();
1237 
1238 	if (!hwp_active && hwp_only)
1239 		goto out;
1240 
1241 	rc = cpufreq_register_driver(&intel_pstate_driver);
1242 	if (rc)
1243 		goto out;
1244 
1245 	intel_pstate_debug_expose_params();
1246 	intel_pstate_sysfs_expose_params();
1247 
1248 	return rc;
1249 out:
1250 	get_online_cpus();
1251 	for_each_online_cpu(cpu) {
1252 		if (all_cpu_data[cpu]) {
1253 			del_timer_sync(&all_cpu_data[cpu]->timer);
1254 			kfree(all_cpu_data[cpu]);
1255 		}
1256 	}
1257 
1258 	put_online_cpus();
1259 	vfree(all_cpu_data);
1260 	return -ENODEV;
1261 }
1262 device_initcall(intel_pstate_init);
1263 
1264 static int __init intel_pstate_setup(char *str)
1265 {
1266 	if (!str)
1267 		return -EINVAL;
1268 
1269 	if (!strcmp(str, "disable"))
1270 		no_load = 1;
1271 	if (!strcmp(str, "no_hwp"))
1272 		no_hwp = 1;
1273 	if (!strcmp(str, "force"))
1274 		force_load = 1;
1275 	if (!strcmp(str, "hwp_only"))
1276 		hwp_only = 1;
1277 	return 0;
1278 }
1279 early_param("intel_pstate", intel_pstate_setup);
1280 
1281 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1282 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1283 MODULE_LICENSE("GPL");
1284