xref: /openbmc/linux/drivers/cpufreq/intel_pstate.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 static void intel_pstate_hwp_set(void)
264 {
265 	int min, hw_min, max, hw_max, cpu, range, adj_range;
266 	u64 value, cap;
267 
268 	rdmsrl(MSR_HWP_CAPABILITIES, cap);
269 	hw_min = HWP_LOWEST_PERF(cap);
270 	hw_max = HWP_HIGHEST_PERF(cap);
271 	range = hw_max - hw_min;
272 
273 	get_online_cpus();
274 
275 	for_each_online_cpu(cpu) {
276 		rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
277 		adj_range = limits.min_perf_pct * range / 100;
278 		min = hw_min + adj_range;
279 		value &= ~HWP_MIN_PERF(~0L);
280 		value |= HWP_MIN_PERF(min);
281 
282 		adj_range = limits.max_perf_pct * range / 100;
283 		max = hw_min + adj_range;
284 		if (limits.no_turbo) {
285 			hw_max = HWP_GUARANTEED_PERF(cap);
286 			if (hw_max < max)
287 				max = hw_max;
288 		}
289 
290 		value &= ~HWP_MAX_PERF(~0L);
291 		value |= HWP_MAX_PERF(max);
292 		wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
293 	}
294 
295 	put_online_cpus();
296 }
297 
298 /************************** debugfs begin ************************/
299 static int pid_param_set(void *data, u64 val)
300 {
301 	*(u32 *)data = val;
302 	intel_pstate_reset_all_pid();
303 	return 0;
304 }
305 
306 static int pid_param_get(void *data, u64 *val)
307 {
308 	*val = *(u32 *)data;
309 	return 0;
310 }
311 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
312 
313 struct pid_param {
314 	char *name;
315 	void *value;
316 };
317 
318 static struct pid_param pid_files[] = {
319 	{"sample_rate_ms", &pid_params.sample_rate_ms},
320 	{"d_gain_pct", &pid_params.d_gain_pct},
321 	{"i_gain_pct", &pid_params.i_gain_pct},
322 	{"deadband", &pid_params.deadband},
323 	{"setpoint", &pid_params.setpoint},
324 	{"p_gain_pct", &pid_params.p_gain_pct},
325 	{NULL, NULL}
326 };
327 
328 static void __init intel_pstate_debug_expose_params(void)
329 {
330 	struct dentry *debugfs_parent;
331 	int i = 0;
332 
333 	if (hwp_active)
334 		return;
335 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
336 	if (IS_ERR_OR_NULL(debugfs_parent))
337 		return;
338 	while (pid_files[i].name) {
339 		debugfs_create_file(pid_files[i].name, 0660,
340 				    debugfs_parent, pid_files[i].value,
341 				    &fops_pid_param);
342 		i++;
343 	}
344 }
345 
346 /************************** debugfs end ************************/
347 
348 /************************** sysfs begin ************************/
349 #define show_one(file_name, object)					\
350 	static ssize_t show_##file_name					\
351 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
352 	{								\
353 		return sprintf(buf, "%u\n", limits.object);		\
354 	}
355 
356 static ssize_t show_turbo_pct(struct kobject *kobj,
357 				struct attribute *attr, char *buf)
358 {
359 	struct cpudata *cpu;
360 	int total, no_turbo, turbo_pct;
361 	uint32_t turbo_fp;
362 
363 	cpu = all_cpu_data[0];
364 
365 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
366 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
367 	turbo_fp = div_fp(int_tofp(no_turbo), int_tofp(total));
368 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
369 	return sprintf(buf, "%u\n", turbo_pct);
370 }
371 
372 static ssize_t show_num_pstates(struct kobject *kobj,
373 				struct attribute *attr, char *buf)
374 {
375 	struct cpudata *cpu;
376 	int total;
377 
378 	cpu = all_cpu_data[0];
379 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
380 	return sprintf(buf, "%u\n", total);
381 }
382 
383 static ssize_t show_no_turbo(struct kobject *kobj,
384 			     struct attribute *attr, char *buf)
385 {
386 	ssize_t ret;
387 
388 	update_turbo_state();
389 	if (limits.turbo_disabled)
390 		ret = sprintf(buf, "%u\n", limits.turbo_disabled);
391 	else
392 		ret = sprintf(buf, "%u\n", limits.no_turbo);
393 
394 	return ret;
395 }
396 
397 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
398 			      const char *buf, size_t count)
399 {
400 	unsigned int input;
401 	int ret;
402 
403 	ret = sscanf(buf, "%u", &input);
404 	if (ret != 1)
405 		return -EINVAL;
406 
407 	update_turbo_state();
408 	if (limits.turbo_disabled) {
409 		pr_warn("intel_pstate: Turbo disabled by BIOS or unavailable on processor\n");
410 		return -EPERM;
411 	}
412 
413 	limits.no_turbo = clamp_t(int, input, 0, 1);
414 
415 	if (hwp_active)
416 		intel_pstate_hwp_set();
417 
418 	return count;
419 }
420 
421 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
422 				  const char *buf, size_t count)
423 {
424 	unsigned int input;
425 	int ret;
426 
427 	ret = sscanf(buf, "%u", &input);
428 	if (ret != 1)
429 		return -EINVAL;
430 
431 	limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
432 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
433 	limits.max_perf_pct = max(limits.min_policy_pct, limits.max_perf_pct);
434 	limits.max_perf_pct = max(limits.min_perf_pct, limits.max_perf_pct);
435 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
436 
437 	if (hwp_active)
438 		intel_pstate_hwp_set();
439 	return count;
440 }
441 
442 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
443 				  const char *buf, size_t count)
444 {
445 	unsigned int input;
446 	int ret;
447 
448 	ret = sscanf(buf, "%u", &input);
449 	if (ret != 1)
450 		return -EINVAL;
451 
452 	limits.min_sysfs_pct = clamp_t(int, input, 0 , 100);
453 	limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
454 	limits.min_perf_pct = min(limits.max_policy_pct, limits.min_perf_pct);
455 	limits.min_perf_pct = min(limits.max_perf_pct, limits.min_perf_pct);
456 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
457 
458 	if (hwp_active)
459 		intel_pstate_hwp_set();
460 	return count;
461 }
462 
463 show_one(max_perf_pct, max_perf_pct);
464 show_one(min_perf_pct, min_perf_pct);
465 
466 define_one_global_rw(no_turbo);
467 define_one_global_rw(max_perf_pct);
468 define_one_global_rw(min_perf_pct);
469 define_one_global_ro(turbo_pct);
470 define_one_global_ro(num_pstates);
471 
472 static struct attribute *intel_pstate_attributes[] = {
473 	&no_turbo.attr,
474 	&max_perf_pct.attr,
475 	&min_perf_pct.attr,
476 	&turbo_pct.attr,
477 	&num_pstates.attr,
478 	NULL
479 };
480 
481 static struct attribute_group intel_pstate_attr_group = {
482 	.attrs = intel_pstate_attributes,
483 };
484 
485 static void __init intel_pstate_sysfs_expose_params(void)
486 {
487 	struct kobject *intel_pstate_kobject;
488 	int rc;
489 
490 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
491 						&cpu_subsys.dev_root->kobj);
492 	BUG_ON(!intel_pstate_kobject);
493 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
494 	BUG_ON(rc);
495 }
496 /************************** sysfs end ************************/
497 
498 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
499 {
500 	pr_info("intel_pstate: HWP enabled\n");
501 
502 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
503 }
504 
505 static int byt_get_min_pstate(void)
506 {
507 	u64 value;
508 
509 	rdmsrl(BYT_RATIOS, value);
510 	return (value >> 8) & 0x7F;
511 }
512 
513 static int byt_get_max_pstate(void)
514 {
515 	u64 value;
516 
517 	rdmsrl(BYT_RATIOS, value);
518 	return (value >> 16) & 0x7F;
519 }
520 
521 static int byt_get_turbo_pstate(void)
522 {
523 	u64 value;
524 
525 	rdmsrl(BYT_TURBO_RATIOS, value);
526 	return value & 0x7F;
527 }
528 
529 static void byt_set_pstate(struct cpudata *cpudata, int pstate)
530 {
531 	u64 val;
532 	int32_t vid_fp;
533 	u32 vid;
534 
535 	val = (u64)pstate << 8;
536 	if (limits.no_turbo && !limits.turbo_disabled)
537 		val |= (u64)1 << 32;
538 
539 	vid_fp = cpudata->vid.min + mul_fp(
540 		int_tofp(pstate - cpudata->pstate.min_pstate),
541 		cpudata->vid.ratio);
542 
543 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
544 	vid = ceiling_fp(vid_fp);
545 
546 	if (pstate > cpudata->pstate.max_pstate)
547 		vid = cpudata->vid.turbo;
548 
549 	val |= vid;
550 
551 	wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
552 }
553 
554 #define BYT_BCLK_FREQS 5
555 static int byt_freq_table[BYT_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800};
556 
557 static int byt_get_scaling(void)
558 {
559 	u64 value;
560 	int i;
561 
562 	rdmsrl(MSR_FSB_FREQ, value);
563 	i = value & 0x3;
564 
565 	BUG_ON(i > BYT_BCLK_FREQS);
566 
567 	return byt_freq_table[i] * 100;
568 }
569 
570 static void byt_get_vid(struct cpudata *cpudata)
571 {
572 	u64 value;
573 
574 	rdmsrl(BYT_VIDS, value);
575 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
576 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
577 	cpudata->vid.ratio = div_fp(
578 		cpudata->vid.max - cpudata->vid.min,
579 		int_tofp(cpudata->pstate.max_pstate -
580 			cpudata->pstate.min_pstate));
581 
582 	rdmsrl(BYT_TURBO_VIDS, value);
583 	cpudata->vid.turbo = value & 0x7f;
584 }
585 
586 static int core_get_min_pstate(void)
587 {
588 	u64 value;
589 
590 	rdmsrl(MSR_PLATFORM_INFO, value);
591 	return (value >> 40) & 0xFF;
592 }
593 
594 static int core_get_max_pstate(void)
595 {
596 	u64 value;
597 
598 	rdmsrl(MSR_PLATFORM_INFO, value);
599 	return (value >> 8) & 0xFF;
600 }
601 
602 static int core_get_turbo_pstate(void)
603 {
604 	u64 value;
605 	int nont, ret;
606 
607 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
608 	nont = core_get_max_pstate();
609 	ret = (value) & 255;
610 	if (ret <= nont)
611 		ret = nont;
612 	return ret;
613 }
614 
615 static inline int core_get_scaling(void)
616 {
617 	return 100000;
618 }
619 
620 static void core_set_pstate(struct cpudata *cpudata, int pstate)
621 {
622 	u64 val;
623 
624 	val = (u64)pstate << 8;
625 	if (limits.no_turbo && !limits.turbo_disabled)
626 		val |= (u64)1 << 32;
627 
628 	wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
629 }
630 
631 static int knl_get_turbo_pstate(void)
632 {
633 	u64 value;
634 	int nont, ret;
635 
636 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
637 	nont = core_get_max_pstate();
638 	ret = (((value) >> 8) & 0xFF);
639 	if (ret <= nont)
640 		ret = nont;
641 	return ret;
642 }
643 
644 static struct cpu_defaults core_params = {
645 	.pid_policy = {
646 		.sample_rate_ms = 10,
647 		.deadband = 0,
648 		.setpoint = 97,
649 		.p_gain_pct = 20,
650 		.d_gain_pct = 0,
651 		.i_gain_pct = 0,
652 	},
653 	.funcs = {
654 		.get_max = core_get_max_pstate,
655 		.get_min = core_get_min_pstate,
656 		.get_turbo = core_get_turbo_pstate,
657 		.get_scaling = core_get_scaling,
658 		.set = core_set_pstate,
659 	},
660 };
661 
662 static struct cpu_defaults byt_params = {
663 	.pid_policy = {
664 		.sample_rate_ms = 10,
665 		.deadband = 0,
666 		.setpoint = 60,
667 		.p_gain_pct = 14,
668 		.d_gain_pct = 0,
669 		.i_gain_pct = 4,
670 	},
671 	.funcs = {
672 		.get_max = byt_get_max_pstate,
673 		.get_min = byt_get_min_pstate,
674 		.get_turbo = byt_get_turbo_pstate,
675 		.set = byt_set_pstate,
676 		.get_scaling = byt_get_scaling,
677 		.get_vid = byt_get_vid,
678 	},
679 };
680 
681 static struct cpu_defaults knl_params = {
682 	.pid_policy = {
683 		.sample_rate_ms = 10,
684 		.deadband = 0,
685 		.setpoint = 97,
686 		.p_gain_pct = 20,
687 		.d_gain_pct = 0,
688 		.i_gain_pct = 0,
689 	},
690 	.funcs = {
691 		.get_max = core_get_max_pstate,
692 		.get_min = core_get_min_pstate,
693 		.get_turbo = knl_get_turbo_pstate,
694 		.get_scaling = core_get_scaling,
695 		.set = core_set_pstate,
696 	},
697 };
698 
699 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
700 {
701 	int max_perf = cpu->pstate.turbo_pstate;
702 	int max_perf_adj;
703 	int min_perf;
704 
705 	if (limits.no_turbo || limits.turbo_disabled)
706 		max_perf = cpu->pstate.max_pstate;
707 
708 	/*
709 	 * performance can be limited by user through sysfs, by cpufreq
710 	 * policy, or by cpu specific default values determined through
711 	 * experimentation.
712 	 */
713 	max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
714 	*max = clamp_t(int, max_perf_adj,
715 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
716 
717 	min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
718 	*min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
719 }
720 
721 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate, bool force)
722 {
723 	int max_perf, min_perf;
724 
725 	if (force) {
726 		update_turbo_state();
727 
728 		intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
729 
730 		pstate = clamp_t(int, pstate, min_perf, max_perf);
731 
732 		if (pstate == cpu->pstate.current_pstate)
733 			return;
734 	}
735 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
736 
737 	cpu->pstate.current_pstate = pstate;
738 
739 	pstate_funcs.set(cpu, pstate);
740 }
741 
742 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
743 {
744 	cpu->pstate.min_pstate = pstate_funcs.get_min();
745 	cpu->pstate.max_pstate = pstate_funcs.get_max();
746 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
747 	cpu->pstate.scaling = pstate_funcs.get_scaling();
748 
749 	if (pstate_funcs.get_vid)
750 		pstate_funcs.get_vid(cpu);
751 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate, false);
752 }
753 
754 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
755 {
756 	struct sample *sample = &cpu->sample;
757 	int64_t core_pct;
758 
759 	core_pct = int_tofp(sample->aperf) * int_tofp(100);
760 	core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
761 
762 	sample->freq = fp_toint(
763 		mul_fp(int_tofp(
764 			cpu->pstate.max_pstate * cpu->pstate.scaling / 100),
765 			core_pct));
766 
767 	sample->core_pct_busy = (int32_t)core_pct;
768 }
769 
770 static inline void intel_pstate_sample(struct cpudata *cpu)
771 {
772 	u64 aperf, mperf;
773 	unsigned long flags;
774 	u64 tsc;
775 
776 	local_irq_save(flags);
777 	rdmsrl(MSR_IA32_APERF, aperf);
778 	rdmsrl(MSR_IA32_MPERF, mperf);
779 	if (cpu->prev_mperf == mperf) {
780 		local_irq_restore(flags);
781 		return;
782 	}
783 
784 	tsc = rdtsc();
785 	local_irq_restore(flags);
786 
787 	cpu->last_sample_time = cpu->sample.time;
788 	cpu->sample.time = ktime_get();
789 	cpu->sample.aperf = aperf;
790 	cpu->sample.mperf = mperf;
791 	cpu->sample.tsc =  tsc;
792 	cpu->sample.aperf -= cpu->prev_aperf;
793 	cpu->sample.mperf -= cpu->prev_mperf;
794 	cpu->sample.tsc -= cpu->prev_tsc;
795 
796 	intel_pstate_calc_busy(cpu);
797 
798 	cpu->prev_aperf = aperf;
799 	cpu->prev_mperf = mperf;
800 	cpu->prev_tsc = tsc;
801 }
802 
803 static inline void intel_hwp_set_sample_time(struct cpudata *cpu)
804 {
805 	int delay;
806 
807 	delay = msecs_to_jiffies(50);
808 	mod_timer_pinned(&cpu->timer, jiffies + delay);
809 }
810 
811 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
812 {
813 	int delay;
814 
815 	delay = msecs_to_jiffies(pid_params.sample_rate_ms);
816 	mod_timer_pinned(&cpu->timer, jiffies + delay);
817 }
818 
819 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
820 {
821 	int32_t core_busy, max_pstate, current_pstate, sample_ratio;
822 	s64 duration_us;
823 	u32 sample_time;
824 
825 	/*
826 	 * core_busy is the ratio of actual performance to max
827 	 * max_pstate is the max non turbo pstate available
828 	 * current_pstate was the pstate that was requested during
829 	 * 	the last sample period.
830 	 *
831 	 * We normalize core_busy, which was our actual percent
832 	 * performance to what we requested during the last sample
833 	 * period. The result will be a percentage of busy at a
834 	 * specified pstate.
835 	 */
836 	core_busy = cpu->sample.core_pct_busy;
837 	max_pstate = int_tofp(cpu->pstate.max_pstate);
838 	current_pstate = int_tofp(cpu->pstate.current_pstate);
839 	core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
840 
841 	/*
842 	 * Since we have a deferred timer, it will not fire unless
843 	 * we are in C0.  So, determine if the actual elapsed time
844 	 * is significantly greater (3x) than our sample interval.  If it
845 	 * is, then we were idle for a long enough period of time
846 	 * to adjust our busyness.
847 	 */
848 	sample_time = pid_params.sample_rate_ms  * USEC_PER_MSEC;
849 	duration_us = ktime_us_delta(cpu->sample.time,
850 				     cpu->last_sample_time);
851 	if (duration_us > sample_time * 3) {
852 		sample_ratio = div_fp(int_tofp(sample_time),
853 				      int_tofp(duration_us));
854 		core_busy = mul_fp(core_busy, sample_ratio);
855 	}
856 
857 	return core_busy;
858 }
859 
860 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
861 {
862 	int32_t busy_scaled;
863 	struct _pid *pid;
864 	signed int ctl;
865 	int from;
866 	struct sample *sample;
867 
868 	from = cpu->pstate.current_pstate;
869 
870 	pid = &cpu->pid;
871 	busy_scaled = intel_pstate_get_scaled_busy(cpu);
872 
873 	ctl = pid_calc(pid, busy_scaled);
874 
875 	/* Negative values of ctl increase the pstate and vice versa */
876 	intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl, true);
877 
878 	sample = &cpu->sample;
879 	trace_pstate_sample(fp_toint(sample->core_pct_busy),
880 		fp_toint(busy_scaled),
881 		from,
882 		cpu->pstate.current_pstate,
883 		sample->mperf,
884 		sample->aperf,
885 		sample->tsc,
886 		sample->freq);
887 }
888 
889 static void intel_hwp_timer_func(unsigned long __data)
890 {
891 	struct cpudata *cpu = (struct cpudata *) __data;
892 
893 	intel_pstate_sample(cpu);
894 	intel_hwp_set_sample_time(cpu);
895 }
896 
897 static void intel_pstate_timer_func(unsigned long __data)
898 {
899 	struct cpudata *cpu = (struct cpudata *) __data;
900 
901 	intel_pstate_sample(cpu);
902 
903 	intel_pstate_adjust_busy_pstate(cpu);
904 
905 	intel_pstate_set_sample_time(cpu);
906 }
907 
908 #define ICPU(model, policy) \
909 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
910 			(unsigned long)&policy }
911 
912 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
913 	ICPU(0x2a, core_params),
914 	ICPU(0x2d, core_params),
915 	ICPU(0x37, byt_params),
916 	ICPU(0x3a, core_params),
917 	ICPU(0x3c, core_params),
918 	ICPU(0x3d, core_params),
919 	ICPU(0x3e, core_params),
920 	ICPU(0x3f, core_params),
921 	ICPU(0x45, core_params),
922 	ICPU(0x46, core_params),
923 	ICPU(0x47, core_params),
924 	ICPU(0x4c, byt_params),
925 	ICPU(0x4e, core_params),
926 	ICPU(0x4f, core_params),
927 	ICPU(0x5e, core_params),
928 	ICPU(0x56, core_params),
929 	ICPU(0x57, knl_params),
930 	{}
931 };
932 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
933 
934 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
935 	ICPU(0x56, core_params),
936 	{}
937 };
938 
939 static int intel_pstate_init_cpu(unsigned int cpunum)
940 {
941 	struct cpudata *cpu;
942 
943 	if (!all_cpu_data[cpunum])
944 		all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
945 					       GFP_KERNEL);
946 	if (!all_cpu_data[cpunum])
947 		return -ENOMEM;
948 
949 	cpu = all_cpu_data[cpunum];
950 
951 	cpu->cpu = cpunum;
952 
953 	if (hwp_active)
954 		intel_pstate_hwp_enable(cpu);
955 
956 	intel_pstate_get_cpu_pstates(cpu);
957 
958 	init_timer_deferrable(&cpu->timer);
959 	cpu->timer.data = (unsigned long)cpu;
960 	cpu->timer.expires = jiffies + HZ/100;
961 
962 	if (!hwp_active)
963 		cpu->timer.function = intel_pstate_timer_func;
964 	else
965 		cpu->timer.function = intel_hwp_timer_func;
966 
967 	intel_pstate_busy_pid_reset(cpu);
968 	intel_pstate_sample(cpu);
969 
970 	add_timer_on(&cpu->timer, cpunum);
971 
972 	pr_debug("intel_pstate: controlling: cpu %d\n", cpunum);
973 
974 	return 0;
975 }
976 
977 static unsigned int intel_pstate_get(unsigned int cpu_num)
978 {
979 	struct sample *sample;
980 	struct cpudata *cpu;
981 
982 	cpu = all_cpu_data[cpu_num];
983 	if (!cpu)
984 		return 0;
985 	sample = &cpu->sample;
986 	return sample->freq;
987 }
988 
989 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
990 {
991 	if (!policy->cpuinfo.max_freq)
992 		return -ENODEV;
993 
994 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE &&
995 	    policy->max >= policy->cpuinfo.max_freq) {
996 		limits.min_policy_pct = 100;
997 		limits.min_perf_pct = 100;
998 		limits.min_perf = int_tofp(1);
999 		limits.max_policy_pct = 100;
1000 		limits.max_perf_pct = 100;
1001 		limits.max_perf = int_tofp(1);
1002 		limits.no_turbo = 0;
1003 		return 0;
1004 	}
1005 
1006 	limits.min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
1007 	limits.min_policy_pct = clamp_t(int, limits.min_policy_pct, 0 , 100);
1008 	limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq;
1009 	limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
1010 
1011 	/* Normalize user input to [min_policy_pct, max_policy_pct] */
1012 	limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
1013 	limits.min_perf_pct = min(limits.max_policy_pct, limits.min_perf_pct);
1014 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
1015 	limits.max_perf_pct = max(limits.min_policy_pct, limits.max_perf_pct);
1016 
1017 	/* Make sure min_perf_pct <= max_perf_pct */
1018 	limits.min_perf_pct = min(limits.max_perf_pct, limits.min_perf_pct);
1019 
1020 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
1021 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
1022 
1023 	if (hwp_active)
1024 		intel_pstate_hwp_set();
1025 
1026 	return 0;
1027 }
1028 
1029 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
1030 {
1031 	cpufreq_verify_within_cpu_limits(policy);
1032 
1033 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
1034 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
1035 		return -EINVAL;
1036 
1037 	return 0;
1038 }
1039 
1040 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1041 {
1042 	int cpu_num = policy->cpu;
1043 	struct cpudata *cpu = all_cpu_data[cpu_num];
1044 
1045 	pr_debug("intel_pstate: CPU %d exiting\n", cpu_num);
1046 
1047 	del_timer_sync(&all_cpu_data[cpu_num]->timer);
1048 	if (hwp_active)
1049 		return;
1050 
1051 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate, false);
1052 }
1053 
1054 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1055 {
1056 	struct cpudata *cpu;
1057 	int rc;
1058 
1059 	rc = intel_pstate_init_cpu(policy->cpu);
1060 	if (rc)
1061 		return rc;
1062 
1063 	cpu = all_cpu_data[policy->cpu];
1064 
1065 	if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
1066 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1067 	else
1068 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
1069 
1070 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1071 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1072 
1073 	/* cpuinfo and default policy values */
1074 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1075 	policy->cpuinfo.max_freq =
1076 		cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1077 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1078 	cpumask_set_cpu(policy->cpu, policy->cpus);
1079 
1080 	return 0;
1081 }
1082 
1083 static struct cpufreq_driver intel_pstate_driver = {
1084 	.flags		= CPUFREQ_CONST_LOOPS,
1085 	.verify		= intel_pstate_verify_policy,
1086 	.setpolicy	= intel_pstate_set_policy,
1087 	.get		= intel_pstate_get,
1088 	.init		= intel_pstate_cpu_init,
1089 	.stop_cpu	= intel_pstate_stop_cpu,
1090 	.name		= "intel_pstate",
1091 };
1092 
1093 static int __initdata no_load;
1094 static int __initdata no_hwp;
1095 static int __initdata hwp_only;
1096 static unsigned int force_load;
1097 
1098 static int intel_pstate_msrs_not_valid(void)
1099 {
1100 	if (!pstate_funcs.get_max() ||
1101 	    !pstate_funcs.get_min() ||
1102 	    !pstate_funcs.get_turbo())
1103 		return -ENODEV;
1104 
1105 	return 0;
1106 }
1107 
1108 static void copy_pid_params(struct pstate_adjust_policy *policy)
1109 {
1110 	pid_params.sample_rate_ms = policy->sample_rate_ms;
1111 	pid_params.p_gain_pct = policy->p_gain_pct;
1112 	pid_params.i_gain_pct = policy->i_gain_pct;
1113 	pid_params.d_gain_pct = policy->d_gain_pct;
1114 	pid_params.deadband = policy->deadband;
1115 	pid_params.setpoint = policy->setpoint;
1116 }
1117 
1118 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1119 {
1120 	pstate_funcs.get_max   = funcs->get_max;
1121 	pstate_funcs.get_min   = funcs->get_min;
1122 	pstate_funcs.get_turbo = funcs->get_turbo;
1123 	pstate_funcs.get_scaling = funcs->get_scaling;
1124 	pstate_funcs.set       = funcs->set;
1125 	pstate_funcs.get_vid   = funcs->get_vid;
1126 }
1127 
1128 #if IS_ENABLED(CONFIG_ACPI)
1129 #include <acpi/processor.h>
1130 
1131 static bool intel_pstate_no_acpi_pss(void)
1132 {
1133 	int i;
1134 
1135 	for_each_possible_cpu(i) {
1136 		acpi_status status;
1137 		union acpi_object *pss;
1138 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1139 		struct acpi_processor *pr = per_cpu(processors, i);
1140 
1141 		if (!pr)
1142 			continue;
1143 
1144 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1145 		if (ACPI_FAILURE(status))
1146 			continue;
1147 
1148 		pss = buffer.pointer;
1149 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1150 			kfree(pss);
1151 			return false;
1152 		}
1153 
1154 		kfree(pss);
1155 	}
1156 
1157 	return true;
1158 }
1159 
1160 static bool intel_pstate_has_acpi_ppc(void)
1161 {
1162 	int i;
1163 
1164 	for_each_possible_cpu(i) {
1165 		struct acpi_processor *pr = per_cpu(processors, i);
1166 
1167 		if (!pr)
1168 			continue;
1169 		if (acpi_has_method(pr->handle, "_PPC"))
1170 			return true;
1171 	}
1172 	return false;
1173 }
1174 
1175 enum {
1176 	PSS,
1177 	PPC,
1178 };
1179 
1180 struct hw_vendor_info {
1181 	u16  valid;
1182 	char oem_id[ACPI_OEM_ID_SIZE];
1183 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1184 	int  oem_pwr_table;
1185 };
1186 
1187 /* Hardware vendor-specific info that has its own power management modes */
1188 static struct hw_vendor_info vendor_info[] = {
1189 	{1, "HP    ", "ProLiant", PSS},
1190 	{1, "ORACLE", "X4-2    ", PPC},
1191 	{1, "ORACLE", "X4-2L   ", PPC},
1192 	{1, "ORACLE", "X4-2B   ", PPC},
1193 	{1, "ORACLE", "X3-2    ", PPC},
1194 	{1, "ORACLE", "X3-2L   ", PPC},
1195 	{1, "ORACLE", "X3-2B   ", PPC},
1196 	{1, "ORACLE", "X4470M2 ", PPC},
1197 	{1, "ORACLE", "X4270M3 ", PPC},
1198 	{1, "ORACLE", "X4270M2 ", PPC},
1199 	{1, "ORACLE", "X4170M2 ", PPC},
1200 	{1, "ORACLE", "X4170 M3", PPC},
1201 	{1, "ORACLE", "X4275 M3", PPC},
1202 	{1, "ORACLE", "X6-2    ", PPC},
1203 	{1, "ORACLE", "Sudbury ", PPC},
1204 	{0, "", ""},
1205 };
1206 
1207 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1208 {
1209 	struct acpi_table_header hdr;
1210 	struct hw_vendor_info *v_info;
1211 	const struct x86_cpu_id *id;
1212 	u64 misc_pwr;
1213 
1214 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1215 	if (id) {
1216 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1217 		if ( misc_pwr & (1 << 8))
1218 			return true;
1219 	}
1220 
1221 	if (acpi_disabled ||
1222 	    ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1223 		return false;
1224 
1225 	for (v_info = vendor_info; v_info->valid; v_info++) {
1226 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1227 			!strncmp(hdr.oem_table_id, v_info->oem_table_id,
1228 						ACPI_OEM_TABLE_ID_SIZE))
1229 			switch (v_info->oem_pwr_table) {
1230 			case PSS:
1231 				return intel_pstate_no_acpi_pss();
1232 			case PPC:
1233 				return intel_pstate_has_acpi_ppc() &&
1234 					(!force_load);
1235 			}
1236 	}
1237 
1238 	return false;
1239 }
1240 #else /* CONFIG_ACPI not enabled */
1241 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1242 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1243 #endif /* CONFIG_ACPI */
1244 
1245 static int __init intel_pstate_init(void)
1246 {
1247 	int cpu, rc = 0;
1248 	const struct x86_cpu_id *id;
1249 	struct cpu_defaults *cpu_def;
1250 
1251 	if (no_load)
1252 		return -ENODEV;
1253 
1254 	id = x86_match_cpu(intel_pstate_cpu_ids);
1255 	if (!id)
1256 		return -ENODEV;
1257 
1258 	/*
1259 	 * The Intel pstate driver will be ignored if the platform
1260 	 * firmware has its own power management modes.
1261 	 */
1262 	if (intel_pstate_platform_pwr_mgmt_exists())
1263 		return -ENODEV;
1264 
1265 	cpu_def = (struct cpu_defaults *)id->driver_data;
1266 
1267 	copy_pid_params(&cpu_def->pid_policy);
1268 	copy_cpu_funcs(&cpu_def->funcs);
1269 
1270 	if (intel_pstate_msrs_not_valid())
1271 		return -ENODEV;
1272 
1273 	pr_info("Intel P-state driver initializing.\n");
1274 
1275 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1276 	if (!all_cpu_data)
1277 		return -ENOMEM;
1278 
1279 	if (static_cpu_has_safe(X86_FEATURE_HWP) && !no_hwp)
1280 		hwp_active++;
1281 
1282 	if (!hwp_active && hwp_only)
1283 		goto out;
1284 
1285 	rc = cpufreq_register_driver(&intel_pstate_driver);
1286 	if (rc)
1287 		goto out;
1288 
1289 	intel_pstate_debug_expose_params();
1290 	intel_pstate_sysfs_expose_params();
1291 
1292 	return rc;
1293 out:
1294 	get_online_cpus();
1295 	for_each_online_cpu(cpu) {
1296 		if (all_cpu_data[cpu]) {
1297 			del_timer_sync(&all_cpu_data[cpu]->timer);
1298 			kfree(all_cpu_data[cpu]);
1299 		}
1300 	}
1301 
1302 	put_online_cpus();
1303 	vfree(all_cpu_data);
1304 	return -ENODEV;
1305 }
1306 device_initcall(intel_pstate_init);
1307 
1308 static int __init intel_pstate_setup(char *str)
1309 {
1310 	if (!str)
1311 		return -EINVAL;
1312 
1313 	if (!strcmp(str, "disable"))
1314 		no_load = 1;
1315 	if (!strcmp(str, "no_hwp"))
1316 		no_hwp = 1;
1317 	if (!strcmp(str, "force"))
1318 		force_load = 1;
1319 	if (!strcmp(str, "hwp_only"))
1320 		hwp_only = 1;
1321 	return 0;
1322 }
1323 early_param("intel_pstate", intel_pstate_setup);
1324 
1325 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1326 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1327 MODULE_LICENSE("GPL");
1328