xref: /linux/drivers/cpufreq/intel_pstate.c (revision 957e3facd147510f2cf8780e38606f1d707f0e33)
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 <trace/events/power.h>
30 
31 #include <asm/div64.h>
32 #include <asm/msr.h>
33 #include <asm/cpu_device_id.h>
34 
35 #define BYT_RATIOS		0x66a
36 #define BYT_VIDS		0x66b
37 #define BYT_TURBO_RATIOS	0x66c
38 #define BYT_TURBO_VIDS		0x66d
39 
40 #define FRAC_BITS 8
41 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
42 #define fp_toint(X) ((X) >> FRAC_BITS)
43 
44 
45 static inline int32_t mul_fp(int32_t x, int32_t y)
46 {
47 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
48 }
49 
50 static inline int32_t div_fp(int32_t x, int32_t y)
51 {
52 	return div_s64((int64_t)x << FRAC_BITS, y);
53 }
54 
55 static inline int ceiling_fp(int32_t x)
56 {
57 	int mask, ret;
58 
59 	ret = fp_toint(x);
60 	mask = (1 << FRAC_BITS) - 1;
61 	if (x & mask)
62 		ret += 1;
63 	return ret;
64 }
65 
66 struct sample {
67 	int32_t core_pct_busy;
68 	u64 aperf;
69 	u64 mperf;
70 	int freq;
71 	ktime_t time;
72 };
73 
74 struct pstate_data {
75 	int	current_pstate;
76 	int	min_pstate;
77 	int	max_pstate;
78 	int	scaling;
79 	int	turbo_pstate;
80 };
81 
82 struct vid_data {
83 	int min;
84 	int max;
85 	int turbo;
86 	int32_t ratio;
87 };
88 
89 struct _pid {
90 	int setpoint;
91 	int32_t integral;
92 	int32_t p_gain;
93 	int32_t i_gain;
94 	int32_t d_gain;
95 	int deadband;
96 	int32_t last_err;
97 };
98 
99 struct cpudata {
100 	int cpu;
101 
102 	struct timer_list timer;
103 
104 	struct pstate_data pstate;
105 	struct vid_data vid;
106 	struct _pid pid;
107 
108 	ktime_t last_sample_time;
109 	u64	prev_aperf;
110 	u64	prev_mperf;
111 	struct sample sample;
112 };
113 
114 static struct cpudata **all_cpu_data;
115 struct pstate_adjust_policy {
116 	int sample_rate_ms;
117 	int deadband;
118 	int setpoint;
119 	int p_gain_pct;
120 	int d_gain_pct;
121 	int i_gain_pct;
122 };
123 
124 struct pstate_funcs {
125 	int (*get_max)(void);
126 	int (*get_min)(void);
127 	int (*get_turbo)(void);
128 	int (*get_scaling)(void);
129 	void (*set)(struct cpudata*, int pstate);
130 	void (*get_vid)(struct cpudata *);
131 };
132 
133 struct cpu_defaults {
134 	struct pstate_adjust_policy pid_policy;
135 	struct pstate_funcs funcs;
136 };
137 
138 static struct pstate_adjust_policy pid_params;
139 static struct pstate_funcs pstate_funcs;
140 static int hwp_active;
141 
142 struct perf_limits {
143 	int no_turbo;
144 	int turbo_disabled;
145 	int max_perf_pct;
146 	int min_perf_pct;
147 	int32_t max_perf;
148 	int32_t min_perf;
149 	int max_policy_pct;
150 	int max_sysfs_pct;
151 };
152 
153 static struct perf_limits limits = {
154 	.no_turbo = 0,
155 	.turbo_disabled = 0,
156 	.max_perf_pct = 100,
157 	.max_perf = int_tofp(1),
158 	.min_perf_pct = 0,
159 	.min_perf = 0,
160 	.max_policy_pct = 100,
161 	.max_sysfs_pct = 100,
162 };
163 
164 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
165 			     int deadband, int integral) {
166 	pid->setpoint = setpoint;
167 	pid->deadband  = deadband;
168 	pid->integral  = int_tofp(integral);
169 	pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
170 }
171 
172 static inline void pid_p_gain_set(struct _pid *pid, int percent)
173 {
174 	pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
175 }
176 
177 static inline void pid_i_gain_set(struct _pid *pid, int percent)
178 {
179 	pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
180 }
181 
182 static inline void pid_d_gain_set(struct _pid *pid, int percent)
183 {
184 	pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
185 }
186 
187 static signed int pid_calc(struct _pid *pid, int32_t busy)
188 {
189 	signed int result;
190 	int32_t pterm, dterm, fp_error;
191 	int32_t integral_limit;
192 
193 	fp_error = int_tofp(pid->setpoint) - busy;
194 
195 	if (abs(fp_error) <= int_tofp(pid->deadband))
196 		return 0;
197 
198 	pterm = mul_fp(pid->p_gain, fp_error);
199 
200 	pid->integral += fp_error;
201 
202 	/* limit the integral term */
203 	integral_limit = int_tofp(30);
204 	if (pid->integral > integral_limit)
205 		pid->integral = integral_limit;
206 	if (pid->integral < -integral_limit)
207 		pid->integral = -integral_limit;
208 
209 	dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
210 	pid->last_err = fp_error;
211 
212 	result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
213 	result = result + (1 << (FRAC_BITS-1));
214 	return (signed int)fp_toint(result);
215 }
216 
217 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
218 {
219 	pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
220 	pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
221 	pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
222 
223 	pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
224 }
225 
226 static inline void intel_pstate_reset_all_pid(void)
227 {
228 	unsigned int cpu;
229 
230 	for_each_online_cpu(cpu) {
231 		if (all_cpu_data[cpu])
232 			intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
233 	}
234 }
235 
236 static inline void update_turbo_state(void)
237 {
238 	u64 misc_en;
239 	struct cpudata *cpu;
240 
241 	cpu = all_cpu_data[0];
242 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
243 	limits.turbo_disabled =
244 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
245 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
246 }
247 
248 #define PCT_TO_HWP(x) (x * 255 / 100)
249 static void intel_pstate_hwp_set(void)
250 {
251 	int min, max, cpu;
252 	u64 value, freq;
253 
254 	get_online_cpus();
255 
256 	for_each_online_cpu(cpu) {
257 		rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
258 		min = PCT_TO_HWP(limits.min_perf_pct);
259 		value &= ~HWP_MIN_PERF(~0L);
260 		value |= HWP_MIN_PERF(min);
261 
262 		max = PCT_TO_HWP(limits.max_perf_pct);
263 		if (limits.no_turbo) {
264 			rdmsrl( MSR_HWP_CAPABILITIES, freq);
265 			max = HWP_GUARANTEED_PERF(freq);
266 		}
267 
268 		value &= ~HWP_MAX_PERF(~0L);
269 		value |= HWP_MAX_PERF(max);
270 		wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
271 	}
272 
273 	put_online_cpus();
274 }
275 
276 /************************** debugfs begin ************************/
277 static int pid_param_set(void *data, u64 val)
278 {
279 	*(u32 *)data = val;
280 	intel_pstate_reset_all_pid();
281 	return 0;
282 }
283 
284 static int pid_param_get(void *data, u64 *val)
285 {
286 	*val = *(u32 *)data;
287 	return 0;
288 }
289 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
290 
291 struct pid_param {
292 	char *name;
293 	void *value;
294 };
295 
296 static struct pid_param pid_files[] = {
297 	{"sample_rate_ms", &pid_params.sample_rate_ms},
298 	{"d_gain_pct", &pid_params.d_gain_pct},
299 	{"i_gain_pct", &pid_params.i_gain_pct},
300 	{"deadband", &pid_params.deadband},
301 	{"setpoint", &pid_params.setpoint},
302 	{"p_gain_pct", &pid_params.p_gain_pct},
303 	{NULL, NULL}
304 };
305 
306 static void __init intel_pstate_debug_expose_params(void)
307 {
308 	struct dentry *debugfs_parent;
309 	int i = 0;
310 
311 	if (hwp_active)
312 		return;
313 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
314 	if (IS_ERR_OR_NULL(debugfs_parent))
315 		return;
316 	while (pid_files[i].name) {
317 		debugfs_create_file(pid_files[i].name, 0660,
318 				    debugfs_parent, pid_files[i].value,
319 				    &fops_pid_param);
320 		i++;
321 	}
322 }
323 
324 /************************** debugfs end ************************/
325 
326 /************************** sysfs begin ************************/
327 #define show_one(file_name, object)					\
328 	static ssize_t show_##file_name					\
329 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
330 	{								\
331 		return sprintf(buf, "%u\n", limits.object);		\
332 	}
333 
334 static ssize_t show_no_turbo(struct kobject *kobj,
335 			     struct attribute *attr, char *buf)
336 {
337 	ssize_t ret;
338 
339 	update_turbo_state();
340 	if (limits.turbo_disabled)
341 		ret = sprintf(buf, "%u\n", limits.turbo_disabled);
342 	else
343 		ret = sprintf(buf, "%u\n", limits.no_turbo);
344 
345 	return ret;
346 }
347 
348 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
349 			      const char *buf, size_t count)
350 {
351 	unsigned int input;
352 	int ret;
353 
354 	ret = sscanf(buf, "%u", &input);
355 	if (ret != 1)
356 		return -EINVAL;
357 
358 	update_turbo_state();
359 	if (limits.turbo_disabled) {
360 		pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
361 		return -EPERM;
362 	}
363 
364 	limits.no_turbo = clamp_t(int, input, 0, 1);
365 
366 	if (hwp_active)
367 		intel_pstate_hwp_set();
368 
369 	return count;
370 }
371 
372 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
373 				  const char *buf, size_t count)
374 {
375 	unsigned int input;
376 	int ret;
377 
378 	ret = sscanf(buf, "%u", &input);
379 	if (ret != 1)
380 		return -EINVAL;
381 
382 	limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
383 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
384 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
385 
386 	if (hwp_active)
387 		intel_pstate_hwp_set();
388 	return count;
389 }
390 
391 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
392 				  const char *buf, size_t count)
393 {
394 	unsigned int input;
395 	int ret;
396 
397 	ret = sscanf(buf, "%u", &input);
398 	if (ret != 1)
399 		return -EINVAL;
400 	limits.min_perf_pct = clamp_t(int, input, 0 , 100);
401 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
402 
403 	if (hwp_active)
404 		intel_pstate_hwp_set();
405 	return count;
406 }
407 
408 show_one(max_perf_pct, max_perf_pct);
409 show_one(min_perf_pct, min_perf_pct);
410 
411 define_one_global_rw(no_turbo);
412 define_one_global_rw(max_perf_pct);
413 define_one_global_rw(min_perf_pct);
414 
415 static struct attribute *intel_pstate_attributes[] = {
416 	&no_turbo.attr,
417 	&max_perf_pct.attr,
418 	&min_perf_pct.attr,
419 	NULL
420 };
421 
422 static struct attribute_group intel_pstate_attr_group = {
423 	.attrs = intel_pstate_attributes,
424 };
425 
426 static void __init intel_pstate_sysfs_expose_params(void)
427 {
428 	struct kobject *intel_pstate_kobject;
429 	int rc;
430 
431 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
432 						&cpu_subsys.dev_root->kobj);
433 	BUG_ON(!intel_pstate_kobject);
434 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
435 	BUG_ON(rc);
436 }
437 /************************** sysfs end ************************/
438 
439 static void intel_pstate_hwp_enable(void)
440 {
441 	hwp_active++;
442 	pr_info("intel_pstate HWP enabled\n");
443 
444 	wrmsrl( MSR_PM_ENABLE, 0x1);
445 }
446 
447 static int byt_get_min_pstate(void)
448 {
449 	u64 value;
450 
451 	rdmsrl(BYT_RATIOS, value);
452 	return (value >> 8) & 0x7F;
453 }
454 
455 static int byt_get_max_pstate(void)
456 {
457 	u64 value;
458 
459 	rdmsrl(BYT_RATIOS, value);
460 	return (value >> 16) & 0x7F;
461 }
462 
463 static int byt_get_turbo_pstate(void)
464 {
465 	u64 value;
466 
467 	rdmsrl(BYT_TURBO_RATIOS, value);
468 	return value & 0x7F;
469 }
470 
471 static void byt_set_pstate(struct cpudata *cpudata, int pstate)
472 {
473 	u64 val;
474 	int32_t vid_fp;
475 	u32 vid;
476 
477 	val = pstate << 8;
478 	if (limits.no_turbo && !limits.turbo_disabled)
479 		val |= (u64)1 << 32;
480 
481 	vid_fp = cpudata->vid.min + mul_fp(
482 		int_tofp(pstate - cpudata->pstate.min_pstate),
483 		cpudata->vid.ratio);
484 
485 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
486 	vid = ceiling_fp(vid_fp);
487 
488 	if (pstate > cpudata->pstate.max_pstate)
489 		vid = cpudata->vid.turbo;
490 
491 	val |= vid;
492 
493 	wrmsrl(MSR_IA32_PERF_CTL, val);
494 }
495 
496 #define BYT_BCLK_FREQS 5
497 static int byt_freq_table[BYT_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800};
498 
499 static int byt_get_scaling(void)
500 {
501 	u64 value;
502 	int i;
503 
504 	rdmsrl(MSR_FSB_FREQ, value);
505 	i = value & 0x3;
506 
507 	BUG_ON(i > BYT_BCLK_FREQS);
508 
509 	return byt_freq_table[i] * 100;
510 }
511 
512 static void byt_get_vid(struct cpudata *cpudata)
513 {
514 	u64 value;
515 
516 	rdmsrl(BYT_VIDS, value);
517 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
518 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
519 	cpudata->vid.ratio = div_fp(
520 		cpudata->vid.max - cpudata->vid.min,
521 		int_tofp(cpudata->pstate.max_pstate -
522 			cpudata->pstate.min_pstate));
523 
524 	rdmsrl(BYT_TURBO_VIDS, value);
525 	cpudata->vid.turbo = value & 0x7f;
526 }
527 
528 static int core_get_min_pstate(void)
529 {
530 	u64 value;
531 
532 	rdmsrl(MSR_PLATFORM_INFO, value);
533 	return (value >> 40) & 0xFF;
534 }
535 
536 static int core_get_max_pstate(void)
537 {
538 	u64 value;
539 
540 	rdmsrl(MSR_PLATFORM_INFO, value);
541 	return (value >> 8) & 0xFF;
542 }
543 
544 static int core_get_turbo_pstate(void)
545 {
546 	u64 value;
547 	int nont, ret;
548 
549 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
550 	nont = core_get_max_pstate();
551 	ret = (value) & 255;
552 	if (ret <= nont)
553 		ret = nont;
554 	return ret;
555 }
556 
557 static inline int core_get_scaling(void)
558 {
559 	return 100000;
560 }
561 
562 static void core_set_pstate(struct cpudata *cpudata, int pstate)
563 {
564 	u64 val;
565 
566 	val = pstate << 8;
567 	if (limits.no_turbo && !limits.turbo_disabled)
568 		val |= (u64)1 << 32;
569 
570 	wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
571 }
572 
573 static struct cpu_defaults core_params = {
574 	.pid_policy = {
575 		.sample_rate_ms = 10,
576 		.deadband = 0,
577 		.setpoint = 97,
578 		.p_gain_pct = 20,
579 		.d_gain_pct = 0,
580 		.i_gain_pct = 0,
581 	},
582 	.funcs = {
583 		.get_max = core_get_max_pstate,
584 		.get_min = core_get_min_pstate,
585 		.get_turbo = core_get_turbo_pstate,
586 		.get_scaling = core_get_scaling,
587 		.set = core_set_pstate,
588 	},
589 };
590 
591 static struct cpu_defaults byt_params = {
592 	.pid_policy = {
593 		.sample_rate_ms = 10,
594 		.deadband = 0,
595 		.setpoint = 97,
596 		.p_gain_pct = 14,
597 		.d_gain_pct = 0,
598 		.i_gain_pct = 4,
599 	},
600 	.funcs = {
601 		.get_max = byt_get_max_pstate,
602 		.get_min = byt_get_min_pstate,
603 		.get_turbo = byt_get_turbo_pstate,
604 		.set = byt_set_pstate,
605 		.get_scaling = byt_get_scaling,
606 		.get_vid = byt_get_vid,
607 	},
608 };
609 
610 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
611 {
612 	int max_perf = cpu->pstate.turbo_pstate;
613 	int max_perf_adj;
614 	int min_perf;
615 
616 	if (limits.no_turbo || limits.turbo_disabled)
617 		max_perf = cpu->pstate.max_pstate;
618 
619 	max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
620 	*max = clamp_t(int, max_perf_adj,
621 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
622 
623 	min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
624 	*min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
625 }
626 
627 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
628 {
629 	int max_perf, min_perf;
630 
631 	update_turbo_state();
632 
633 	intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
634 
635 	pstate = clamp_t(int, pstate, min_perf, max_perf);
636 
637 	if (pstate == cpu->pstate.current_pstate)
638 		return;
639 
640 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
641 
642 	cpu->pstate.current_pstate = pstate;
643 
644 	pstate_funcs.set(cpu, pstate);
645 }
646 
647 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
648 {
649 	cpu->pstate.min_pstate = pstate_funcs.get_min();
650 	cpu->pstate.max_pstate = pstate_funcs.get_max();
651 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
652 	cpu->pstate.scaling = pstate_funcs.get_scaling();
653 
654 	if (pstate_funcs.get_vid)
655 		pstate_funcs.get_vid(cpu);
656 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
657 }
658 
659 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
660 {
661 	struct sample *sample = &cpu->sample;
662 	int64_t core_pct;
663 
664 	core_pct = int_tofp(sample->aperf) * int_tofp(100);
665 	core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
666 
667 	sample->freq = fp_toint(
668 		mul_fp(int_tofp(
669 			cpu->pstate.max_pstate * cpu->pstate.scaling / 100),
670 			core_pct));
671 
672 	sample->core_pct_busy = (int32_t)core_pct;
673 }
674 
675 static inline void intel_pstate_sample(struct cpudata *cpu)
676 {
677 	u64 aperf, mperf;
678 	unsigned long flags;
679 
680 	local_irq_save(flags);
681 	rdmsrl(MSR_IA32_APERF, aperf);
682 	rdmsrl(MSR_IA32_MPERF, mperf);
683 	local_irq_restore(flags);
684 
685 	cpu->last_sample_time = cpu->sample.time;
686 	cpu->sample.time = ktime_get();
687 	cpu->sample.aperf = aperf;
688 	cpu->sample.mperf = mperf;
689 	cpu->sample.aperf -= cpu->prev_aperf;
690 	cpu->sample.mperf -= cpu->prev_mperf;
691 
692 	intel_pstate_calc_busy(cpu);
693 
694 	cpu->prev_aperf = aperf;
695 	cpu->prev_mperf = mperf;
696 }
697 
698 static inline void intel_hwp_set_sample_time(struct cpudata *cpu)
699 {
700 	int delay;
701 
702 	delay = msecs_to_jiffies(50);
703 	mod_timer_pinned(&cpu->timer, jiffies + delay);
704 }
705 
706 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
707 {
708 	int delay;
709 
710 	delay = msecs_to_jiffies(pid_params.sample_rate_ms);
711 	mod_timer_pinned(&cpu->timer, jiffies + delay);
712 }
713 
714 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
715 {
716 	int32_t core_busy, max_pstate, current_pstate, sample_ratio;
717 	u32 duration_us;
718 	u32 sample_time;
719 
720 	core_busy = cpu->sample.core_pct_busy;
721 	max_pstate = int_tofp(cpu->pstate.max_pstate);
722 	current_pstate = int_tofp(cpu->pstate.current_pstate);
723 	core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
724 
725 	sample_time = pid_params.sample_rate_ms  * USEC_PER_MSEC;
726 	duration_us = (u32) ktime_us_delta(cpu->sample.time,
727 					   cpu->last_sample_time);
728 	if (duration_us > sample_time * 3) {
729 		sample_ratio = div_fp(int_tofp(sample_time),
730 				      int_tofp(duration_us));
731 		core_busy = mul_fp(core_busy, sample_ratio);
732 	}
733 
734 	return core_busy;
735 }
736 
737 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
738 {
739 	int32_t busy_scaled;
740 	struct _pid *pid;
741 	signed int ctl;
742 
743 	pid = &cpu->pid;
744 	busy_scaled = intel_pstate_get_scaled_busy(cpu);
745 
746 	ctl = pid_calc(pid, busy_scaled);
747 
748 	/* Negative values of ctl increase the pstate and vice versa */
749 	intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl);
750 }
751 
752 static void intel_hwp_timer_func(unsigned long __data)
753 {
754 	struct cpudata *cpu = (struct cpudata *) __data;
755 
756 	intel_pstate_sample(cpu);
757 	intel_hwp_set_sample_time(cpu);
758 }
759 
760 static void intel_pstate_timer_func(unsigned long __data)
761 {
762 	struct cpudata *cpu = (struct cpudata *) __data;
763 	struct sample *sample;
764 
765 	intel_pstate_sample(cpu);
766 
767 	sample = &cpu->sample;
768 
769 	intel_pstate_adjust_busy_pstate(cpu);
770 
771 	trace_pstate_sample(fp_toint(sample->core_pct_busy),
772 			fp_toint(intel_pstate_get_scaled_busy(cpu)),
773 			cpu->pstate.current_pstate,
774 			sample->mperf,
775 			sample->aperf,
776 			sample->freq);
777 
778 	intel_pstate_set_sample_time(cpu);
779 }
780 
781 #define ICPU(model, policy) \
782 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
783 			(unsigned long)&policy }
784 
785 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
786 	ICPU(0x2a, core_params),
787 	ICPU(0x2d, core_params),
788 	ICPU(0x37, byt_params),
789 	ICPU(0x3a, core_params),
790 	ICPU(0x3c, core_params),
791 	ICPU(0x3d, core_params),
792 	ICPU(0x3e, core_params),
793 	ICPU(0x3f, core_params),
794 	ICPU(0x45, core_params),
795 	ICPU(0x46, core_params),
796 	ICPU(0x47, core_params),
797 	ICPU(0x4c, byt_params),
798 	ICPU(0x4f, core_params),
799 	ICPU(0x56, core_params),
800 	{}
801 };
802 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
803 
804 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
805 	ICPU(0x56, core_params),
806 	{}
807 };
808 
809 static int intel_pstate_init_cpu(unsigned int cpunum)
810 {
811 	struct cpudata *cpu;
812 
813 	if (!all_cpu_data[cpunum])
814 		all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
815 					       GFP_KERNEL);
816 	if (!all_cpu_data[cpunum])
817 		return -ENOMEM;
818 
819 	cpu = all_cpu_data[cpunum];
820 
821 	cpu->cpu = cpunum;
822 	intel_pstate_get_cpu_pstates(cpu);
823 
824 	init_timer_deferrable(&cpu->timer);
825 	cpu->timer.data = (unsigned long)cpu;
826 	cpu->timer.expires = jiffies + HZ/100;
827 
828 	if (!hwp_active)
829 		cpu->timer.function = intel_pstate_timer_func;
830 	else
831 		cpu->timer.function = intel_hwp_timer_func;
832 
833 	intel_pstate_busy_pid_reset(cpu);
834 	intel_pstate_sample(cpu);
835 
836 	add_timer_on(&cpu->timer, cpunum);
837 
838 	pr_debug("Intel pstate controlling: cpu %d\n", cpunum);
839 
840 	return 0;
841 }
842 
843 static unsigned int intel_pstate_get(unsigned int cpu_num)
844 {
845 	struct sample *sample;
846 	struct cpudata *cpu;
847 
848 	cpu = all_cpu_data[cpu_num];
849 	if (!cpu)
850 		return 0;
851 	sample = &cpu->sample;
852 	return sample->freq;
853 }
854 
855 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
856 {
857 	if (!policy->cpuinfo.max_freq)
858 		return -ENODEV;
859 
860 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
861 		limits.min_perf_pct = 100;
862 		limits.min_perf = int_tofp(1);
863 		limits.max_policy_pct = 100;
864 		limits.max_perf_pct = 100;
865 		limits.max_perf = int_tofp(1);
866 		limits.no_turbo = 0;
867 		return 0;
868 	}
869 
870 	limits.min_perf_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
871 	limits.min_perf_pct = clamp_t(int, limits.min_perf_pct, 0 , 100);
872 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
873 
874 	limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq;
875 	limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
876 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
877 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
878 
879 	if (hwp_active)
880 		intel_pstate_hwp_set();
881 
882 	return 0;
883 }
884 
885 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
886 {
887 	cpufreq_verify_within_cpu_limits(policy);
888 
889 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
890 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
891 		return -EINVAL;
892 
893 	return 0;
894 }
895 
896 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
897 {
898 	int cpu_num = policy->cpu;
899 	struct cpudata *cpu = all_cpu_data[cpu_num];
900 
901 	pr_info("intel_pstate CPU %d exiting\n", cpu_num);
902 
903 	del_timer_sync(&all_cpu_data[cpu_num]->timer);
904 	if (hwp_active)
905 		return;
906 
907 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
908 }
909 
910 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
911 {
912 	struct cpudata *cpu;
913 	int rc;
914 
915 	rc = intel_pstate_init_cpu(policy->cpu);
916 	if (rc)
917 		return rc;
918 
919 	cpu = all_cpu_data[policy->cpu];
920 
921 	if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
922 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
923 	else
924 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
925 
926 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
927 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
928 
929 	/* cpuinfo and default policy values */
930 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
931 	policy->cpuinfo.max_freq =
932 		cpu->pstate.turbo_pstate * cpu->pstate.scaling;
933 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
934 	cpumask_set_cpu(policy->cpu, policy->cpus);
935 
936 	return 0;
937 }
938 
939 static struct cpufreq_driver intel_pstate_driver = {
940 	.flags		= CPUFREQ_CONST_LOOPS,
941 	.verify		= intel_pstate_verify_policy,
942 	.setpolicy	= intel_pstate_set_policy,
943 	.get		= intel_pstate_get,
944 	.init		= intel_pstate_cpu_init,
945 	.stop_cpu	= intel_pstate_stop_cpu,
946 	.name		= "intel_pstate",
947 };
948 
949 static int __initdata no_load;
950 static int __initdata no_hwp;
951 
952 static int intel_pstate_msrs_not_valid(void)
953 {
954 	/* Check that all the msr's we are using are valid. */
955 	u64 aperf, mperf, tmp;
956 
957 	rdmsrl(MSR_IA32_APERF, aperf);
958 	rdmsrl(MSR_IA32_MPERF, mperf);
959 
960 	if (!pstate_funcs.get_max() ||
961 	    !pstate_funcs.get_min() ||
962 	    !pstate_funcs.get_turbo())
963 		return -ENODEV;
964 
965 	rdmsrl(MSR_IA32_APERF, tmp);
966 	if (!(tmp - aperf))
967 		return -ENODEV;
968 
969 	rdmsrl(MSR_IA32_MPERF, tmp);
970 	if (!(tmp - mperf))
971 		return -ENODEV;
972 
973 	return 0;
974 }
975 
976 static void copy_pid_params(struct pstate_adjust_policy *policy)
977 {
978 	pid_params.sample_rate_ms = policy->sample_rate_ms;
979 	pid_params.p_gain_pct = policy->p_gain_pct;
980 	pid_params.i_gain_pct = policy->i_gain_pct;
981 	pid_params.d_gain_pct = policy->d_gain_pct;
982 	pid_params.deadband = policy->deadband;
983 	pid_params.setpoint = policy->setpoint;
984 }
985 
986 static void copy_cpu_funcs(struct pstate_funcs *funcs)
987 {
988 	pstate_funcs.get_max   = funcs->get_max;
989 	pstate_funcs.get_min   = funcs->get_min;
990 	pstate_funcs.get_turbo = funcs->get_turbo;
991 	pstate_funcs.get_scaling = funcs->get_scaling;
992 	pstate_funcs.set       = funcs->set;
993 	pstate_funcs.get_vid   = funcs->get_vid;
994 }
995 
996 #if IS_ENABLED(CONFIG_ACPI)
997 #include <acpi/processor.h>
998 
999 static bool intel_pstate_no_acpi_pss(void)
1000 {
1001 	int i;
1002 
1003 	for_each_possible_cpu(i) {
1004 		acpi_status status;
1005 		union acpi_object *pss;
1006 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1007 		struct acpi_processor *pr = per_cpu(processors, i);
1008 
1009 		if (!pr)
1010 			continue;
1011 
1012 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1013 		if (ACPI_FAILURE(status))
1014 			continue;
1015 
1016 		pss = buffer.pointer;
1017 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1018 			kfree(pss);
1019 			return false;
1020 		}
1021 
1022 		kfree(pss);
1023 	}
1024 
1025 	return true;
1026 }
1027 
1028 static bool intel_pstate_has_acpi_ppc(void)
1029 {
1030 	int i;
1031 
1032 	for_each_possible_cpu(i) {
1033 		struct acpi_processor *pr = per_cpu(processors, i);
1034 
1035 		if (!pr)
1036 			continue;
1037 		if (acpi_has_method(pr->handle, "_PPC"))
1038 			return true;
1039 	}
1040 	return false;
1041 }
1042 
1043 enum {
1044 	PSS,
1045 	PPC,
1046 };
1047 
1048 struct hw_vendor_info {
1049 	u16  valid;
1050 	char oem_id[ACPI_OEM_ID_SIZE];
1051 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1052 	int  oem_pwr_table;
1053 };
1054 
1055 /* Hardware vendor-specific info that has its own power management modes */
1056 static struct hw_vendor_info vendor_info[] = {
1057 	{1, "HP    ", "ProLiant", PSS},
1058 	{1, "ORACLE", "X4-2    ", PPC},
1059 	{1, "ORACLE", "X4-2L   ", PPC},
1060 	{1, "ORACLE", "X4-2B   ", PPC},
1061 	{1, "ORACLE", "X3-2    ", PPC},
1062 	{1, "ORACLE", "X3-2L   ", PPC},
1063 	{1, "ORACLE", "X3-2B   ", PPC},
1064 	{1, "ORACLE", "X4470M2 ", PPC},
1065 	{1, "ORACLE", "X4270M3 ", PPC},
1066 	{1, "ORACLE", "X4270M2 ", PPC},
1067 	{1, "ORACLE", "X4170M2 ", PPC},
1068 	{0, "", ""},
1069 };
1070 
1071 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1072 {
1073 	struct acpi_table_header hdr;
1074 	struct hw_vendor_info *v_info;
1075 	const struct x86_cpu_id *id;
1076 	u64 misc_pwr;
1077 
1078 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1079 	if (id) {
1080 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1081 		if ( misc_pwr & (1 << 8))
1082 			return true;
1083 	}
1084 
1085 	if (acpi_disabled ||
1086 	    ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1087 		return false;
1088 
1089 	for (v_info = vendor_info; v_info->valid; v_info++) {
1090 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1091 			!strncmp(hdr.oem_table_id, v_info->oem_table_id,
1092 						ACPI_OEM_TABLE_ID_SIZE))
1093 			switch (v_info->oem_pwr_table) {
1094 			case PSS:
1095 				return intel_pstate_no_acpi_pss();
1096 			case PPC:
1097 				return intel_pstate_has_acpi_ppc();
1098 			}
1099 	}
1100 
1101 	return false;
1102 }
1103 #else /* CONFIG_ACPI not enabled */
1104 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1105 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1106 #endif /* CONFIG_ACPI */
1107 
1108 static int __init intel_pstate_init(void)
1109 {
1110 	int cpu, rc = 0;
1111 	const struct x86_cpu_id *id;
1112 	struct cpu_defaults *cpu_info;
1113 	struct cpuinfo_x86 *c = &boot_cpu_data;
1114 
1115 	if (no_load)
1116 		return -ENODEV;
1117 
1118 	id = x86_match_cpu(intel_pstate_cpu_ids);
1119 	if (!id)
1120 		return -ENODEV;
1121 
1122 	/*
1123 	 * The Intel pstate driver will be ignored if the platform
1124 	 * firmware has its own power management modes.
1125 	 */
1126 	if (intel_pstate_platform_pwr_mgmt_exists())
1127 		return -ENODEV;
1128 
1129 	cpu_info = (struct cpu_defaults *)id->driver_data;
1130 
1131 	copy_pid_params(&cpu_info->pid_policy);
1132 	copy_cpu_funcs(&cpu_info->funcs);
1133 
1134 	if (intel_pstate_msrs_not_valid())
1135 		return -ENODEV;
1136 
1137 	pr_info("Intel P-state driver initializing.\n");
1138 
1139 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1140 	if (!all_cpu_data)
1141 		return -ENOMEM;
1142 
1143 	if (cpu_has(c,X86_FEATURE_HWP) && !no_hwp)
1144 		intel_pstate_hwp_enable();
1145 
1146 	rc = cpufreq_register_driver(&intel_pstate_driver);
1147 	if (rc)
1148 		goto out;
1149 
1150 	intel_pstate_debug_expose_params();
1151 	intel_pstate_sysfs_expose_params();
1152 
1153 	return rc;
1154 out:
1155 	get_online_cpus();
1156 	for_each_online_cpu(cpu) {
1157 		if (all_cpu_data[cpu]) {
1158 			del_timer_sync(&all_cpu_data[cpu]->timer);
1159 			kfree(all_cpu_data[cpu]);
1160 		}
1161 	}
1162 
1163 	put_online_cpus();
1164 	vfree(all_cpu_data);
1165 	return -ENODEV;
1166 }
1167 device_initcall(intel_pstate_init);
1168 
1169 static int __init intel_pstate_setup(char *str)
1170 {
1171 	if (!str)
1172 		return -EINVAL;
1173 
1174 	if (!strcmp(str, "disable"))
1175 		no_load = 1;
1176 	if (!strcmp(str, "no_hwp"))
1177 		no_hwp = 1;
1178 	return 0;
1179 }
1180 early_param("intel_pstate", intel_pstate_setup);
1181 
1182 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1183 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1184 MODULE_LICENSE("GPL");
1185