xref: /openbmc/linux/drivers/cpufreq/intel_pstate.c (revision bbde9fc1824aab58bc78c084163007dd6c03fe5b)
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(s64 x, s64 y)
53 {
54 	return div64_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 	u64 tsc;
73 	int freq;
74 	ktime_t time;
75 };
76 
77 struct pstate_data {
78 	int	current_pstate;
79 	int	min_pstate;
80 	int	max_pstate;
81 	int	scaling;
82 	int	turbo_pstate;
83 };
84 
85 struct vid_data {
86 	int min;
87 	int max;
88 	int turbo;
89 	int32_t ratio;
90 };
91 
92 struct _pid {
93 	int setpoint;
94 	int32_t integral;
95 	int32_t p_gain;
96 	int32_t i_gain;
97 	int32_t d_gain;
98 	int deadband;
99 	int32_t last_err;
100 };
101 
102 struct cpudata {
103 	int cpu;
104 
105 	struct timer_list timer;
106 
107 	struct pstate_data pstate;
108 	struct vid_data vid;
109 	struct _pid pid;
110 
111 	ktime_t last_sample_time;
112 	u64	prev_aperf;
113 	u64	prev_mperf;
114 	u64	prev_tsc;
115 	struct sample sample;
116 };
117 
118 static struct cpudata **all_cpu_data;
119 struct pstate_adjust_policy {
120 	int sample_rate_ms;
121 	int deadband;
122 	int setpoint;
123 	int p_gain_pct;
124 	int d_gain_pct;
125 	int i_gain_pct;
126 };
127 
128 struct pstate_funcs {
129 	int (*get_max)(void);
130 	int (*get_min)(void);
131 	int (*get_turbo)(void);
132 	int (*get_scaling)(void);
133 	void (*set)(struct cpudata*, int pstate);
134 	void (*get_vid)(struct cpudata *);
135 };
136 
137 struct cpu_defaults {
138 	struct pstate_adjust_policy pid_policy;
139 	struct pstate_funcs funcs;
140 };
141 
142 static struct pstate_adjust_policy pid_params;
143 static struct pstate_funcs pstate_funcs;
144 static int hwp_active;
145 
146 struct perf_limits {
147 	int no_turbo;
148 	int turbo_disabled;
149 	int max_perf_pct;
150 	int min_perf_pct;
151 	int32_t max_perf;
152 	int32_t min_perf;
153 	int max_policy_pct;
154 	int max_sysfs_pct;
155 	int min_policy_pct;
156 	int min_sysfs_pct;
157 };
158 
159 static struct perf_limits limits = {
160 	.no_turbo = 0,
161 	.turbo_disabled = 0,
162 	.max_perf_pct = 100,
163 	.max_perf = int_tofp(1),
164 	.min_perf_pct = 0,
165 	.min_perf = 0,
166 	.max_policy_pct = 100,
167 	.max_sysfs_pct = 100,
168 	.min_policy_pct = 0,
169 	.min_sysfs_pct = 0,
170 };
171 
172 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
173 			     int deadband, int integral) {
174 	pid->setpoint = setpoint;
175 	pid->deadband  = deadband;
176 	pid->integral  = int_tofp(integral);
177 	pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
178 }
179 
180 static inline void pid_p_gain_set(struct _pid *pid, int percent)
181 {
182 	pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
183 }
184 
185 static inline void pid_i_gain_set(struct _pid *pid, int percent)
186 {
187 	pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
188 }
189 
190 static inline void pid_d_gain_set(struct _pid *pid, int percent)
191 {
192 	pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
193 }
194 
195 static signed int pid_calc(struct _pid *pid, int32_t busy)
196 {
197 	signed int result;
198 	int32_t pterm, dterm, fp_error;
199 	int32_t integral_limit;
200 
201 	fp_error = int_tofp(pid->setpoint) - busy;
202 
203 	if (abs(fp_error) <= int_tofp(pid->deadband))
204 		return 0;
205 
206 	pterm = mul_fp(pid->p_gain, fp_error);
207 
208 	pid->integral += fp_error;
209 
210 	/*
211 	 * We limit the integral here so that it will never
212 	 * get higher than 30.  This prevents it from becoming
213 	 * too large an input over long periods of time and allows
214 	 * it to get factored out sooner.
215 	 *
216 	 * The value of 30 was chosen through experimentation.
217 	 */
218 	integral_limit = int_tofp(30);
219 	if (pid->integral > integral_limit)
220 		pid->integral = integral_limit;
221 	if (pid->integral < -integral_limit)
222 		pid->integral = -integral_limit;
223 
224 	dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
225 	pid->last_err = fp_error;
226 
227 	result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
228 	result = result + (1 << (FRAC_BITS-1));
229 	return (signed int)fp_toint(result);
230 }
231 
232 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
233 {
234 	pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
235 	pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
236 	pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
237 
238 	pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
239 }
240 
241 static inline void intel_pstate_reset_all_pid(void)
242 {
243 	unsigned int cpu;
244 
245 	for_each_online_cpu(cpu) {
246 		if (all_cpu_data[cpu])
247 			intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
248 	}
249 }
250 
251 static inline void update_turbo_state(void)
252 {
253 	u64 misc_en;
254 	struct cpudata *cpu;
255 
256 	cpu = all_cpu_data[0];
257 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
258 	limits.turbo_disabled =
259 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
260 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
261 }
262 
263 #define PCT_TO_HWP(x) (x * 255 / 100)
264 static void intel_pstate_hwp_set(void)
265 {
266 	int min, max, cpu;
267 	u64 value, freq;
268 
269 	get_online_cpus();
270 
271 	for_each_online_cpu(cpu) {
272 		rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
273 		min = PCT_TO_HWP(limits.min_perf_pct);
274 		value &= ~HWP_MIN_PERF(~0L);
275 		value |= HWP_MIN_PERF(min);
276 
277 		max = PCT_TO_HWP(limits.max_perf_pct);
278 		if (limits.no_turbo) {
279 			rdmsrl( MSR_HWP_CAPABILITIES, freq);
280 			max = HWP_GUARANTEED_PERF(freq);
281 		}
282 
283 		value &= ~HWP_MAX_PERF(~0L);
284 		value |= HWP_MAX_PERF(max);
285 		wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
286 	}
287 
288 	put_online_cpus();
289 }
290 
291 /************************** debugfs begin ************************/
292 static int pid_param_set(void *data, u64 val)
293 {
294 	*(u32 *)data = val;
295 	intel_pstate_reset_all_pid();
296 	return 0;
297 }
298 
299 static int pid_param_get(void *data, u64 *val)
300 {
301 	*val = *(u32 *)data;
302 	return 0;
303 }
304 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
305 
306 struct pid_param {
307 	char *name;
308 	void *value;
309 };
310 
311 static struct pid_param pid_files[] = {
312 	{"sample_rate_ms", &pid_params.sample_rate_ms},
313 	{"d_gain_pct", &pid_params.d_gain_pct},
314 	{"i_gain_pct", &pid_params.i_gain_pct},
315 	{"deadband", &pid_params.deadband},
316 	{"setpoint", &pid_params.setpoint},
317 	{"p_gain_pct", &pid_params.p_gain_pct},
318 	{NULL, NULL}
319 };
320 
321 static void __init intel_pstate_debug_expose_params(void)
322 {
323 	struct dentry *debugfs_parent;
324 	int i = 0;
325 
326 	if (hwp_active)
327 		return;
328 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
329 	if (IS_ERR_OR_NULL(debugfs_parent))
330 		return;
331 	while (pid_files[i].name) {
332 		debugfs_create_file(pid_files[i].name, 0660,
333 				    debugfs_parent, pid_files[i].value,
334 				    &fops_pid_param);
335 		i++;
336 	}
337 }
338 
339 /************************** debugfs end ************************/
340 
341 /************************** sysfs begin ************************/
342 #define show_one(file_name, object)					\
343 	static ssize_t show_##file_name					\
344 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
345 	{								\
346 		return sprintf(buf, "%u\n", limits.object);		\
347 	}
348 
349 static ssize_t show_turbo_pct(struct kobject *kobj,
350 				struct attribute *attr, char *buf)
351 {
352 	struct cpudata *cpu;
353 	int total, no_turbo, turbo_pct;
354 	uint32_t turbo_fp;
355 
356 	cpu = all_cpu_data[0];
357 
358 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
359 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
360 	turbo_fp = div_fp(int_tofp(no_turbo), int_tofp(total));
361 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
362 	return sprintf(buf, "%u\n", turbo_pct);
363 }
364 
365 static ssize_t show_num_pstates(struct kobject *kobj,
366 				struct attribute *attr, char *buf)
367 {
368 	struct cpudata *cpu;
369 	int total;
370 
371 	cpu = all_cpu_data[0];
372 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
373 	return sprintf(buf, "%u\n", total);
374 }
375 
376 static ssize_t show_no_turbo(struct kobject *kobj,
377 			     struct attribute *attr, char *buf)
378 {
379 	ssize_t ret;
380 
381 	update_turbo_state();
382 	if (limits.turbo_disabled)
383 		ret = sprintf(buf, "%u\n", limits.turbo_disabled);
384 	else
385 		ret = sprintf(buf, "%u\n", limits.no_turbo);
386 
387 	return ret;
388 }
389 
390 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
391 			      const char *buf, size_t count)
392 {
393 	unsigned int input;
394 	int ret;
395 
396 	ret = sscanf(buf, "%u", &input);
397 	if (ret != 1)
398 		return -EINVAL;
399 
400 	update_turbo_state();
401 	if (limits.turbo_disabled) {
402 		pr_warn("intel_pstate: Turbo disabled by BIOS or unavailable on processor\n");
403 		return -EPERM;
404 	}
405 
406 	limits.no_turbo = clamp_t(int, input, 0, 1);
407 
408 	if (hwp_active)
409 		intel_pstate_hwp_set();
410 
411 	return count;
412 }
413 
414 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
415 				  const char *buf, size_t count)
416 {
417 	unsigned int input;
418 	int ret;
419 
420 	ret = sscanf(buf, "%u", &input);
421 	if (ret != 1)
422 		return -EINVAL;
423 
424 	limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
425 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
426 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
427 
428 	if (hwp_active)
429 		intel_pstate_hwp_set();
430 	return count;
431 }
432 
433 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
434 				  const char *buf, size_t count)
435 {
436 	unsigned int input;
437 	int ret;
438 
439 	ret = sscanf(buf, "%u", &input);
440 	if (ret != 1)
441 		return -EINVAL;
442 
443 	limits.min_sysfs_pct = clamp_t(int, input, 0 , 100);
444 	limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
445 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
446 
447 	if (hwp_active)
448 		intel_pstate_hwp_set();
449 	return count;
450 }
451 
452 show_one(max_perf_pct, max_perf_pct);
453 show_one(min_perf_pct, min_perf_pct);
454 
455 define_one_global_rw(no_turbo);
456 define_one_global_rw(max_perf_pct);
457 define_one_global_rw(min_perf_pct);
458 define_one_global_ro(turbo_pct);
459 define_one_global_ro(num_pstates);
460 
461 static struct attribute *intel_pstate_attributes[] = {
462 	&no_turbo.attr,
463 	&max_perf_pct.attr,
464 	&min_perf_pct.attr,
465 	&turbo_pct.attr,
466 	&num_pstates.attr,
467 	NULL
468 };
469 
470 static struct attribute_group intel_pstate_attr_group = {
471 	.attrs = intel_pstate_attributes,
472 };
473 
474 static void __init intel_pstate_sysfs_expose_params(void)
475 {
476 	struct kobject *intel_pstate_kobject;
477 	int rc;
478 
479 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
480 						&cpu_subsys.dev_root->kobj);
481 	BUG_ON(!intel_pstate_kobject);
482 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
483 	BUG_ON(rc);
484 }
485 /************************** sysfs end ************************/
486 
487 static void intel_pstate_hwp_enable(void)
488 {
489 	hwp_active++;
490 	pr_info("intel_pstate: HWP enabled\n");
491 
492 	wrmsrl( MSR_PM_ENABLE, 0x1);
493 }
494 
495 static int byt_get_min_pstate(void)
496 {
497 	u64 value;
498 
499 	rdmsrl(BYT_RATIOS, value);
500 	return (value >> 8) & 0x7F;
501 }
502 
503 static int byt_get_max_pstate(void)
504 {
505 	u64 value;
506 
507 	rdmsrl(BYT_RATIOS, value);
508 	return (value >> 16) & 0x7F;
509 }
510 
511 static int byt_get_turbo_pstate(void)
512 {
513 	u64 value;
514 
515 	rdmsrl(BYT_TURBO_RATIOS, value);
516 	return value & 0x7F;
517 }
518 
519 static void byt_set_pstate(struct cpudata *cpudata, int pstate)
520 {
521 	u64 val;
522 	int32_t vid_fp;
523 	u32 vid;
524 
525 	val = pstate << 8;
526 	if (limits.no_turbo && !limits.turbo_disabled)
527 		val |= (u64)1 << 32;
528 
529 	vid_fp = cpudata->vid.min + mul_fp(
530 		int_tofp(pstate - cpudata->pstate.min_pstate),
531 		cpudata->vid.ratio);
532 
533 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
534 	vid = ceiling_fp(vid_fp);
535 
536 	if (pstate > cpudata->pstate.max_pstate)
537 		vid = cpudata->vid.turbo;
538 
539 	val |= vid;
540 
541 	wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
542 }
543 
544 #define BYT_BCLK_FREQS 5
545 static int byt_freq_table[BYT_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800};
546 
547 static int byt_get_scaling(void)
548 {
549 	u64 value;
550 	int i;
551 
552 	rdmsrl(MSR_FSB_FREQ, value);
553 	i = value & 0x3;
554 
555 	BUG_ON(i > BYT_BCLK_FREQS);
556 
557 	return byt_freq_table[i] * 100;
558 }
559 
560 static void byt_get_vid(struct cpudata *cpudata)
561 {
562 	u64 value;
563 
564 	rdmsrl(BYT_VIDS, value);
565 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
566 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
567 	cpudata->vid.ratio = div_fp(
568 		cpudata->vid.max - cpudata->vid.min,
569 		int_tofp(cpudata->pstate.max_pstate -
570 			cpudata->pstate.min_pstate));
571 
572 	rdmsrl(BYT_TURBO_VIDS, value);
573 	cpudata->vid.turbo = value & 0x7f;
574 }
575 
576 static int core_get_min_pstate(void)
577 {
578 	u64 value;
579 
580 	rdmsrl(MSR_PLATFORM_INFO, value);
581 	return (value >> 40) & 0xFF;
582 }
583 
584 static int core_get_max_pstate(void)
585 {
586 	u64 value;
587 
588 	rdmsrl(MSR_PLATFORM_INFO, value);
589 	return (value >> 8) & 0xFF;
590 }
591 
592 static int core_get_turbo_pstate(void)
593 {
594 	u64 value;
595 	int nont, ret;
596 
597 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
598 	nont = core_get_max_pstate();
599 	ret = (value) & 255;
600 	if (ret <= nont)
601 		ret = nont;
602 	return ret;
603 }
604 
605 static inline int core_get_scaling(void)
606 {
607 	return 100000;
608 }
609 
610 static void core_set_pstate(struct cpudata *cpudata, int pstate)
611 {
612 	u64 val;
613 
614 	val = pstate << 8;
615 	if (limits.no_turbo && !limits.turbo_disabled)
616 		val |= (u64)1 << 32;
617 
618 	wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
619 }
620 
621 static int knl_get_turbo_pstate(void)
622 {
623 	u64 value;
624 	int nont, ret;
625 
626 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
627 	nont = core_get_max_pstate();
628 	ret = (((value) >> 8) & 0xFF);
629 	if (ret <= nont)
630 		ret = nont;
631 	return ret;
632 }
633 
634 static struct cpu_defaults core_params = {
635 	.pid_policy = {
636 		.sample_rate_ms = 10,
637 		.deadband = 0,
638 		.setpoint = 97,
639 		.p_gain_pct = 20,
640 		.d_gain_pct = 0,
641 		.i_gain_pct = 0,
642 	},
643 	.funcs = {
644 		.get_max = core_get_max_pstate,
645 		.get_min = core_get_min_pstate,
646 		.get_turbo = core_get_turbo_pstate,
647 		.get_scaling = core_get_scaling,
648 		.set = core_set_pstate,
649 	},
650 };
651 
652 static struct cpu_defaults byt_params = {
653 	.pid_policy = {
654 		.sample_rate_ms = 10,
655 		.deadband = 0,
656 		.setpoint = 60,
657 		.p_gain_pct = 14,
658 		.d_gain_pct = 0,
659 		.i_gain_pct = 4,
660 	},
661 	.funcs = {
662 		.get_max = byt_get_max_pstate,
663 		.get_min = byt_get_min_pstate,
664 		.get_turbo = byt_get_turbo_pstate,
665 		.set = byt_set_pstate,
666 		.get_scaling = byt_get_scaling,
667 		.get_vid = byt_get_vid,
668 	},
669 };
670 
671 static struct cpu_defaults knl_params = {
672 	.pid_policy = {
673 		.sample_rate_ms = 10,
674 		.deadband = 0,
675 		.setpoint = 97,
676 		.p_gain_pct = 20,
677 		.d_gain_pct = 0,
678 		.i_gain_pct = 0,
679 	},
680 	.funcs = {
681 		.get_max = core_get_max_pstate,
682 		.get_min = core_get_min_pstate,
683 		.get_turbo = knl_get_turbo_pstate,
684 		.get_scaling = core_get_scaling,
685 		.set = core_set_pstate,
686 	},
687 };
688 
689 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
690 {
691 	int max_perf = cpu->pstate.turbo_pstate;
692 	int max_perf_adj;
693 	int min_perf;
694 
695 	if (limits.no_turbo || limits.turbo_disabled)
696 		max_perf = cpu->pstate.max_pstate;
697 
698 	/*
699 	 * performance can be limited by user through sysfs, by cpufreq
700 	 * policy, or by cpu specific default values determined through
701 	 * experimentation.
702 	 */
703 	max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
704 	*max = clamp_t(int, max_perf_adj,
705 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
706 
707 	min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
708 	*min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
709 }
710 
711 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate, bool force)
712 {
713 	int max_perf, min_perf;
714 
715 	if (force) {
716 		update_turbo_state();
717 
718 		intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
719 
720 		pstate = clamp_t(int, pstate, min_perf, max_perf);
721 
722 		if (pstate == cpu->pstate.current_pstate)
723 			return;
724 	}
725 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
726 
727 	cpu->pstate.current_pstate = pstate;
728 
729 	pstate_funcs.set(cpu, pstate);
730 }
731 
732 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
733 {
734 	cpu->pstate.min_pstate = pstate_funcs.get_min();
735 	cpu->pstate.max_pstate = pstate_funcs.get_max();
736 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
737 	cpu->pstate.scaling = pstate_funcs.get_scaling();
738 
739 	if (pstate_funcs.get_vid)
740 		pstate_funcs.get_vid(cpu);
741 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate, false);
742 }
743 
744 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
745 {
746 	struct sample *sample = &cpu->sample;
747 	int64_t core_pct;
748 
749 	core_pct = int_tofp(sample->aperf) * int_tofp(100);
750 	core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
751 
752 	sample->freq = fp_toint(
753 		mul_fp(int_tofp(
754 			cpu->pstate.max_pstate * cpu->pstate.scaling / 100),
755 			core_pct));
756 
757 	sample->core_pct_busy = (int32_t)core_pct;
758 }
759 
760 static inline void intel_pstate_sample(struct cpudata *cpu)
761 {
762 	u64 aperf, mperf;
763 	unsigned long flags;
764 	u64 tsc;
765 
766 	local_irq_save(flags);
767 	rdmsrl(MSR_IA32_APERF, aperf);
768 	rdmsrl(MSR_IA32_MPERF, mperf);
769 	tsc = native_read_tsc();
770 	local_irq_restore(flags);
771 
772 	cpu->last_sample_time = cpu->sample.time;
773 	cpu->sample.time = ktime_get();
774 	cpu->sample.aperf = aperf;
775 	cpu->sample.mperf = mperf;
776 	cpu->sample.tsc =  tsc;
777 	cpu->sample.aperf -= cpu->prev_aperf;
778 	cpu->sample.mperf -= cpu->prev_mperf;
779 	cpu->sample.tsc -= cpu->prev_tsc;
780 
781 	intel_pstate_calc_busy(cpu);
782 
783 	cpu->prev_aperf = aperf;
784 	cpu->prev_mperf = mperf;
785 	cpu->prev_tsc = tsc;
786 }
787 
788 static inline void intel_hwp_set_sample_time(struct cpudata *cpu)
789 {
790 	int delay;
791 
792 	delay = msecs_to_jiffies(50);
793 	mod_timer_pinned(&cpu->timer, jiffies + delay);
794 }
795 
796 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
797 {
798 	int delay;
799 
800 	delay = msecs_to_jiffies(pid_params.sample_rate_ms);
801 	mod_timer_pinned(&cpu->timer, jiffies + delay);
802 }
803 
804 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
805 {
806 	int32_t core_busy, max_pstate, current_pstate, sample_ratio;
807 	s64 duration_us;
808 	u32 sample_time;
809 
810 	/*
811 	 * core_busy is the ratio of actual performance to max
812 	 * max_pstate is the max non turbo pstate available
813 	 * current_pstate was the pstate that was requested during
814 	 * 	the last sample period.
815 	 *
816 	 * We normalize core_busy, which was our actual percent
817 	 * performance to what we requested during the last sample
818 	 * period. The result will be a percentage of busy at a
819 	 * specified pstate.
820 	 */
821 	core_busy = cpu->sample.core_pct_busy;
822 	max_pstate = int_tofp(cpu->pstate.max_pstate);
823 	current_pstate = int_tofp(cpu->pstate.current_pstate);
824 	core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
825 
826 	/*
827 	 * Since we have a deferred timer, it will not fire unless
828 	 * we are in C0.  So, determine if the actual elapsed time
829 	 * is significantly greater (3x) than our sample interval.  If it
830 	 * is, then we were idle for a long enough period of time
831 	 * to adjust our busyness.
832 	 */
833 	sample_time = pid_params.sample_rate_ms  * USEC_PER_MSEC;
834 	duration_us = ktime_us_delta(cpu->sample.time,
835 				     cpu->last_sample_time);
836 	if (duration_us > sample_time * 3) {
837 		sample_ratio = div_fp(int_tofp(sample_time),
838 				      int_tofp(duration_us));
839 		core_busy = mul_fp(core_busy, sample_ratio);
840 	}
841 
842 	return core_busy;
843 }
844 
845 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
846 {
847 	int32_t busy_scaled;
848 	struct _pid *pid;
849 	signed int ctl;
850 	int from;
851 	struct sample *sample;
852 
853 	from = cpu->pstate.current_pstate;
854 
855 	pid = &cpu->pid;
856 	busy_scaled = intel_pstate_get_scaled_busy(cpu);
857 
858 	ctl = pid_calc(pid, busy_scaled);
859 
860 	/* Negative values of ctl increase the pstate and vice versa */
861 	intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl, true);
862 
863 	sample = &cpu->sample;
864 	trace_pstate_sample(fp_toint(sample->core_pct_busy),
865 		fp_toint(busy_scaled),
866 		from,
867 		cpu->pstate.current_pstate,
868 		sample->mperf,
869 		sample->aperf,
870 		sample->tsc,
871 		sample->freq);
872 }
873 
874 static void intel_hwp_timer_func(unsigned long __data)
875 {
876 	struct cpudata *cpu = (struct cpudata *) __data;
877 
878 	intel_pstate_sample(cpu);
879 	intel_hwp_set_sample_time(cpu);
880 }
881 
882 static void intel_pstate_timer_func(unsigned long __data)
883 {
884 	struct cpudata *cpu = (struct cpudata *) __data;
885 
886 	intel_pstate_sample(cpu);
887 
888 	intel_pstate_adjust_busy_pstate(cpu);
889 
890 	intel_pstate_set_sample_time(cpu);
891 }
892 
893 #define ICPU(model, policy) \
894 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
895 			(unsigned long)&policy }
896 
897 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
898 	ICPU(0x2a, core_params),
899 	ICPU(0x2d, core_params),
900 	ICPU(0x37, byt_params),
901 	ICPU(0x3a, core_params),
902 	ICPU(0x3c, core_params),
903 	ICPU(0x3d, core_params),
904 	ICPU(0x3e, core_params),
905 	ICPU(0x3f, core_params),
906 	ICPU(0x45, core_params),
907 	ICPU(0x46, core_params),
908 	ICPU(0x47, core_params),
909 	ICPU(0x4c, byt_params),
910 	ICPU(0x4e, core_params),
911 	ICPU(0x4f, core_params),
912 	ICPU(0x56, core_params),
913 	ICPU(0x57, knl_params),
914 	{}
915 };
916 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
917 
918 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
919 	ICPU(0x56, core_params),
920 	{}
921 };
922 
923 static int intel_pstate_init_cpu(unsigned int cpunum)
924 {
925 	struct cpudata *cpu;
926 
927 	if (!all_cpu_data[cpunum])
928 		all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
929 					       GFP_KERNEL);
930 	if (!all_cpu_data[cpunum])
931 		return -ENOMEM;
932 
933 	cpu = all_cpu_data[cpunum];
934 
935 	cpu->cpu = cpunum;
936 	intel_pstate_get_cpu_pstates(cpu);
937 
938 	init_timer_deferrable(&cpu->timer);
939 	cpu->timer.data = (unsigned long)cpu;
940 	cpu->timer.expires = jiffies + HZ/100;
941 
942 	if (!hwp_active)
943 		cpu->timer.function = intel_pstate_timer_func;
944 	else
945 		cpu->timer.function = intel_hwp_timer_func;
946 
947 	intel_pstate_busy_pid_reset(cpu);
948 	intel_pstate_sample(cpu);
949 
950 	add_timer_on(&cpu->timer, cpunum);
951 
952 	pr_debug("intel_pstate: controlling: cpu %d\n", cpunum);
953 
954 	return 0;
955 }
956 
957 static unsigned int intel_pstate_get(unsigned int cpu_num)
958 {
959 	struct sample *sample;
960 	struct cpudata *cpu;
961 
962 	cpu = all_cpu_data[cpu_num];
963 	if (!cpu)
964 		return 0;
965 	sample = &cpu->sample;
966 	return sample->freq;
967 }
968 
969 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
970 {
971 	if (!policy->cpuinfo.max_freq)
972 		return -ENODEV;
973 
974 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE &&
975 	    policy->max >= policy->cpuinfo.max_freq) {
976 		limits.min_policy_pct = 100;
977 		limits.min_perf_pct = 100;
978 		limits.min_perf = int_tofp(1);
979 		limits.max_policy_pct = 100;
980 		limits.max_perf_pct = 100;
981 		limits.max_perf = int_tofp(1);
982 		limits.no_turbo = 0;
983 		return 0;
984 	}
985 
986 	limits.min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
987 	limits.min_policy_pct = clamp_t(int, limits.min_policy_pct, 0 , 100);
988 	limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
989 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
990 
991 	limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq;
992 	limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
993 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
994 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
995 
996 	if (hwp_active)
997 		intel_pstate_hwp_set();
998 
999 	return 0;
1000 }
1001 
1002 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
1003 {
1004 	cpufreq_verify_within_cpu_limits(policy);
1005 
1006 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
1007 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
1008 		return -EINVAL;
1009 
1010 	return 0;
1011 }
1012 
1013 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1014 {
1015 	int cpu_num = policy->cpu;
1016 	struct cpudata *cpu = all_cpu_data[cpu_num];
1017 
1018 	pr_debug("intel_pstate: CPU %d exiting\n", cpu_num);
1019 
1020 	del_timer_sync(&all_cpu_data[cpu_num]->timer);
1021 	if (hwp_active)
1022 		return;
1023 
1024 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate, false);
1025 }
1026 
1027 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1028 {
1029 	struct cpudata *cpu;
1030 	int rc;
1031 
1032 	rc = intel_pstate_init_cpu(policy->cpu);
1033 	if (rc)
1034 		return rc;
1035 
1036 	cpu = all_cpu_data[policy->cpu];
1037 
1038 	if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
1039 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1040 	else
1041 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
1042 
1043 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1044 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1045 
1046 	/* cpuinfo and default policy values */
1047 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1048 	policy->cpuinfo.max_freq =
1049 		cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1050 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1051 	cpumask_set_cpu(policy->cpu, policy->cpus);
1052 
1053 	return 0;
1054 }
1055 
1056 static struct cpufreq_driver intel_pstate_driver = {
1057 	.flags		= CPUFREQ_CONST_LOOPS,
1058 	.verify		= intel_pstate_verify_policy,
1059 	.setpolicy	= intel_pstate_set_policy,
1060 	.get		= intel_pstate_get,
1061 	.init		= intel_pstate_cpu_init,
1062 	.stop_cpu	= intel_pstate_stop_cpu,
1063 	.name		= "intel_pstate",
1064 };
1065 
1066 static int __initdata no_load;
1067 static int __initdata no_hwp;
1068 static int __initdata hwp_only;
1069 static unsigned int force_load;
1070 
1071 static int intel_pstate_msrs_not_valid(void)
1072 {
1073 	if (!pstate_funcs.get_max() ||
1074 	    !pstate_funcs.get_min() ||
1075 	    !pstate_funcs.get_turbo())
1076 		return -ENODEV;
1077 
1078 	return 0;
1079 }
1080 
1081 static void copy_pid_params(struct pstate_adjust_policy *policy)
1082 {
1083 	pid_params.sample_rate_ms = policy->sample_rate_ms;
1084 	pid_params.p_gain_pct = policy->p_gain_pct;
1085 	pid_params.i_gain_pct = policy->i_gain_pct;
1086 	pid_params.d_gain_pct = policy->d_gain_pct;
1087 	pid_params.deadband = policy->deadband;
1088 	pid_params.setpoint = policy->setpoint;
1089 }
1090 
1091 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1092 {
1093 	pstate_funcs.get_max   = funcs->get_max;
1094 	pstate_funcs.get_min   = funcs->get_min;
1095 	pstate_funcs.get_turbo = funcs->get_turbo;
1096 	pstate_funcs.get_scaling = funcs->get_scaling;
1097 	pstate_funcs.set       = funcs->set;
1098 	pstate_funcs.get_vid   = funcs->get_vid;
1099 }
1100 
1101 #if IS_ENABLED(CONFIG_ACPI)
1102 #include <acpi/processor.h>
1103 
1104 static bool intel_pstate_no_acpi_pss(void)
1105 {
1106 	int i;
1107 
1108 	for_each_possible_cpu(i) {
1109 		acpi_status status;
1110 		union acpi_object *pss;
1111 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1112 		struct acpi_processor *pr = per_cpu(processors, i);
1113 
1114 		if (!pr)
1115 			continue;
1116 
1117 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1118 		if (ACPI_FAILURE(status))
1119 			continue;
1120 
1121 		pss = buffer.pointer;
1122 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1123 			kfree(pss);
1124 			return false;
1125 		}
1126 
1127 		kfree(pss);
1128 	}
1129 
1130 	return true;
1131 }
1132 
1133 static bool intel_pstate_has_acpi_ppc(void)
1134 {
1135 	int i;
1136 
1137 	for_each_possible_cpu(i) {
1138 		struct acpi_processor *pr = per_cpu(processors, i);
1139 
1140 		if (!pr)
1141 			continue;
1142 		if (acpi_has_method(pr->handle, "_PPC"))
1143 			return true;
1144 	}
1145 	return false;
1146 }
1147 
1148 enum {
1149 	PSS,
1150 	PPC,
1151 };
1152 
1153 struct hw_vendor_info {
1154 	u16  valid;
1155 	char oem_id[ACPI_OEM_ID_SIZE];
1156 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1157 	int  oem_pwr_table;
1158 };
1159 
1160 /* Hardware vendor-specific info that has its own power management modes */
1161 static struct hw_vendor_info vendor_info[] = {
1162 	{1, "HP    ", "ProLiant", PSS},
1163 	{1, "ORACLE", "X4-2    ", PPC},
1164 	{1, "ORACLE", "X4-2L   ", PPC},
1165 	{1, "ORACLE", "X4-2B   ", PPC},
1166 	{1, "ORACLE", "X3-2    ", PPC},
1167 	{1, "ORACLE", "X3-2L   ", PPC},
1168 	{1, "ORACLE", "X3-2B   ", PPC},
1169 	{1, "ORACLE", "X4470M2 ", PPC},
1170 	{1, "ORACLE", "X4270M3 ", PPC},
1171 	{1, "ORACLE", "X4270M2 ", PPC},
1172 	{1, "ORACLE", "X4170M2 ", PPC},
1173 	{0, "", ""},
1174 };
1175 
1176 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1177 {
1178 	struct acpi_table_header hdr;
1179 	struct hw_vendor_info *v_info;
1180 	const struct x86_cpu_id *id;
1181 	u64 misc_pwr;
1182 
1183 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1184 	if (id) {
1185 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1186 		if ( misc_pwr & (1 << 8))
1187 			return true;
1188 	}
1189 
1190 	if (acpi_disabled ||
1191 	    ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1192 		return false;
1193 
1194 	for (v_info = vendor_info; v_info->valid; v_info++) {
1195 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1196 			!strncmp(hdr.oem_table_id, v_info->oem_table_id,
1197 						ACPI_OEM_TABLE_ID_SIZE))
1198 			switch (v_info->oem_pwr_table) {
1199 			case PSS:
1200 				return intel_pstate_no_acpi_pss();
1201 			case PPC:
1202 				return intel_pstate_has_acpi_ppc() &&
1203 					(!force_load);
1204 			}
1205 	}
1206 
1207 	return false;
1208 }
1209 #else /* CONFIG_ACPI not enabled */
1210 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1211 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1212 #endif /* CONFIG_ACPI */
1213 
1214 static int __init intel_pstate_init(void)
1215 {
1216 	int cpu, rc = 0;
1217 	const struct x86_cpu_id *id;
1218 	struct cpu_defaults *cpu_def;
1219 
1220 	if (no_load)
1221 		return -ENODEV;
1222 
1223 	id = x86_match_cpu(intel_pstate_cpu_ids);
1224 	if (!id)
1225 		return -ENODEV;
1226 
1227 	/*
1228 	 * The Intel pstate driver will be ignored if the platform
1229 	 * firmware has its own power management modes.
1230 	 */
1231 	if (intel_pstate_platform_pwr_mgmt_exists())
1232 		return -ENODEV;
1233 
1234 	cpu_def = (struct cpu_defaults *)id->driver_data;
1235 
1236 	copy_pid_params(&cpu_def->pid_policy);
1237 	copy_cpu_funcs(&cpu_def->funcs);
1238 
1239 	if (intel_pstate_msrs_not_valid())
1240 		return -ENODEV;
1241 
1242 	pr_info("Intel P-state driver initializing.\n");
1243 
1244 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1245 	if (!all_cpu_data)
1246 		return -ENOMEM;
1247 
1248 	if (static_cpu_has_safe(X86_FEATURE_HWP) && !no_hwp)
1249 		intel_pstate_hwp_enable();
1250 
1251 	if (!hwp_active && hwp_only)
1252 		goto out;
1253 
1254 	rc = cpufreq_register_driver(&intel_pstate_driver);
1255 	if (rc)
1256 		goto out;
1257 
1258 	intel_pstate_debug_expose_params();
1259 	intel_pstate_sysfs_expose_params();
1260 
1261 	return rc;
1262 out:
1263 	get_online_cpus();
1264 	for_each_online_cpu(cpu) {
1265 		if (all_cpu_data[cpu]) {
1266 			del_timer_sync(&all_cpu_data[cpu]->timer);
1267 			kfree(all_cpu_data[cpu]);
1268 		}
1269 	}
1270 
1271 	put_online_cpus();
1272 	vfree(all_cpu_data);
1273 	return -ENODEV;
1274 }
1275 device_initcall(intel_pstate_init);
1276 
1277 static int __init intel_pstate_setup(char *str)
1278 {
1279 	if (!str)
1280 		return -EINVAL;
1281 
1282 	if (!strcmp(str, "disable"))
1283 		no_load = 1;
1284 	if (!strcmp(str, "no_hwp"))
1285 		no_hwp = 1;
1286 	if (!strcmp(str, "force"))
1287 		force_load = 1;
1288 	if (!strcmp(str, "hwp_only"))
1289 		hwp_only = 1;
1290 	return 0;
1291 }
1292 early_param("intel_pstate", intel_pstate_setup);
1293 
1294 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1295 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1296 MODULE_LICENSE("GPL");
1297