1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * intel_pstate.c: Native P state management for Intel processors
4  *
5  * (C) Copyright 2012 Intel Corporation
6  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
7  */
8 
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include <linux/kernel.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/module.h>
14 #include <linux/ktime.h>
15 #include <linux/hrtimer.h>
16 #include <linux/tick.h>
17 #include <linux/slab.h>
18 #include <linux/sched/cpufreq.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
21 #include <linux/cpufreq.h>
22 #include <linux/sysfs.h>
23 #include <linux/types.h>
24 #include <linux/fs.h>
25 #include <linux/acpi.h>
26 #include <linux/vmalloc.h>
27 #include <linux/pm_qos.h>
28 #include <trace/events/power.h>
29 
30 #include <asm/cpu.h>
31 #include <asm/div64.h>
32 #include <asm/msr.h>
33 #include <asm/cpu_device_id.h>
34 #include <asm/cpufeature.h>
35 #include <asm/intel-family.h>
36 #include "../drivers/thermal/intel/thermal_interrupt.h"
37 
38 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
39 
40 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
41 #define INTEL_CPUFREQ_TRANSITION_DELAY_HWP	5000
42 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
43 
44 #ifdef CONFIG_ACPI
45 #include <acpi/processor.h>
46 #include <acpi/cppc_acpi.h>
47 #endif
48 
49 #define FRAC_BITS 8
50 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
51 #define fp_toint(X) ((X) >> FRAC_BITS)
52 
53 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
54 
55 #define EXT_BITS 6
56 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
57 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
58 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
59 
60 static inline int32_t mul_fp(int32_t x, int32_t y)
61 {
62 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
63 }
64 
65 static inline int32_t div_fp(s64 x, s64 y)
66 {
67 	return div64_s64((int64_t)x << FRAC_BITS, y);
68 }
69 
70 static inline int ceiling_fp(int32_t x)
71 {
72 	int mask, ret;
73 
74 	ret = fp_toint(x);
75 	mask = (1 << FRAC_BITS) - 1;
76 	if (x & mask)
77 		ret += 1;
78 	return ret;
79 }
80 
81 static inline u64 mul_ext_fp(u64 x, u64 y)
82 {
83 	return (x * y) >> EXT_FRAC_BITS;
84 }
85 
86 static inline u64 div_ext_fp(u64 x, u64 y)
87 {
88 	return div64_u64(x << EXT_FRAC_BITS, y);
89 }
90 
91 /**
92  * struct sample -	Store performance sample
93  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
94  *			performance during last sample period
95  * @busy_scaled:	Scaled busy value which is used to calculate next
96  *			P state. This can be different than core_avg_perf
97  *			to account for cpu idle period
98  * @aperf:		Difference of actual performance frequency clock count
99  *			read from APERF MSR between last and current sample
100  * @mperf:		Difference of maximum performance frequency clock count
101  *			read from MPERF MSR between last and current sample
102  * @tsc:		Difference of time stamp counter between last and
103  *			current sample
104  * @time:		Current time from scheduler
105  *
106  * This structure is used in the cpudata structure to store performance sample
107  * data for choosing next P State.
108  */
109 struct sample {
110 	int32_t core_avg_perf;
111 	int32_t busy_scaled;
112 	u64 aperf;
113 	u64 mperf;
114 	u64 tsc;
115 	u64 time;
116 };
117 
118 /**
119  * struct pstate_data - Store P state data
120  * @current_pstate:	Current requested P state
121  * @min_pstate:		Min P state possible for this platform
122  * @max_pstate:		Max P state possible for this platform
123  * @max_pstate_physical:This is physical Max P state for a processor
124  *			This can be higher than the max_pstate which can
125  *			be limited by platform thermal design power limits
126  * @perf_ctl_scaling:	PERF_CTL P-state to frequency scaling factor
127  * @scaling:		Scaling factor between performance and frequency
128  * @turbo_pstate:	Max Turbo P state possible for this platform
129  * @min_freq:		@min_pstate frequency in cpufreq units
130  * @max_freq:		@max_pstate frequency in cpufreq units
131  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
132  *
133  * Stores the per cpu model P state limits and current P state.
134  */
135 struct pstate_data {
136 	int	current_pstate;
137 	int	min_pstate;
138 	int	max_pstate;
139 	int	max_pstate_physical;
140 	int	perf_ctl_scaling;
141 	int	scaling;
142 	int	turbo_pstate;
143 	unsigned int min_freq;
144 	unsigned int max_freq;
145 	unsigned int turbo_freq;
146 };
147 
148 /**
149  * struct vid_data -	Stores voltage information data
150  * @min:		VID data for this platform corresponding to
151  *			the lowest P state
152  * @max:		VID data corresponding to the highest P State.
153  * @turbo:		VID data for turbo P state
154  * @ratio:		Ratio of (vid max - vid min) /
155  *			(max P state - Min P State)
156  *
157  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
158  * This data is used in Atom platforms, where in addition to target P state,
159  * the voltage data needs to be specified to select next P State.
160  */
161 struct vid_data {
162 	int min;
163 	int max;
164 	int turbo;
165 	int32_t ratio;
166 };
167 
168 /**
169  * struct global_params - Global parameters, mostly tunable via sysfs.
170  * @no_turbo:		Whether or not to use turbo P-states.
171  * @turbo_disabled:	Whether or not turbo P-states are available at all,
172  *			based on the MSR_IA32_MISC_ENABLE value and whether or
173  *			not the maximum reported turbo P-state is different from
174  *			the maximum reported non-turbo one.
175  * @turbo_disabled_mf:	The @turbo_disabled value reflected by cpuinfo.max_freq.
176  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
177  *			P-state capacity.
178  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
179  *			P-state capacity.
180  */
181 struct global_params {
182 	bool no_turbo;
183 	bool turbo_disabled;
184 	bool turbo_disabled_mf;
185 	int max_perf_pct;
186 	int min_perf_pct;
187 };
188 
189 /**
190  * struct cpudata -	Per CPU instance data storage
191  * @cpu:		CPU number for this instance data
192  * @policy:		CPUFreq policy value
193  * @update_util:	CPUFreq utility callback information
194  * @update_util_set:	CPUFreq utility callback is set
195  * @iowait_boost:	iowait-related boost fraction
196  * @last_update:	Time of the last update.
197  * @pstate:		Stores P state limits for this CPU
198  * @vid:		Stores VID limits for this CPU
199  * @last_sample_time:	Last Sample time
200  * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
201  * @prev_aperf:		Last APERF value read from APERF MSR
202  * @prev_mperf:		Last MPERF value read from MPERF MSR
203  * @prev_tsc:		Last timestamp counter (TSC) value
204  * @prev_cummulative_iowait: IO Wait time difference from last and
205  *			current sample
206  * @sample:		Storage for storing last Sample data
207  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
208  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
209  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
210  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
211  * @epp_powersave:	Last saved HWP energy performance preference
212  *			(EPP) or energy performance bias (EPB),
213  *			when policy switched to performance
214  * @epp_policy:		Last saved policy used to set EPP/EPB
215  * @epp_default:	Power on default HWP energy performance
216  *			preference/bias
217  * @epp_cached		Cached HWP energy-performance preference value
218  * @hwp_req_cached:	Cached value of the last HWP Request MSR
219  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
220  * @last_io_update:	Last time when IO wake flag was set
221  * @sched_flags:	Store scheduler flags for possible cross CPU update
222  * @hwp_boost_min:	Last HWP boosted min performance
223  * @suspended:		Whether or not the driver has been suspended.
224  * @hwp_notify_work:	workqueue for HWP notifications.
225  *
226  * This structure stores per CPU instance data for all CPUs.
227  */
228 struct cpudata {
229 	int cpu;
230 
231 	unsigned int policy;
232 	struct update_util_data update_util;
233 	bool   update_util_set;
234 
235 	struct pstate_data pstate;
236 	struct vid_data vid;
237 
238 	u64	last_update;
239 	u64	last_sample_time;
240 	u64	aperf_mperf_shift;
241 	u64	prev_aperf;
242 	u64	prev_mperf;
243 	u64	prev_tsc;
244 	u64	prev_cummulative_iowait;
245 	struct sample sample;
246 	int32_t	min_perf_ratio;
247 	int32_t	max_perf_ratio;
248 #ifdef CONFIG_ACPI
249 	struct acpi_processor_performance acpi_perf_data;
250 	bool valid_pss_table;
251 #endif
252 	unsigned int iowait_boost;
253 	s16 epp_powersave;
254 	s16 epp_policy;
255 	s16 epp_default;
256 	s16 epp_cached;
257 	u64 hwp_req_cached;
258 	u64 hwp_cap_cached;
259 	u64 last_io_update;
260 	unsigned int sched_flags;
261 	u32 hwp_boost_min;
262 	bool suspended;
263 	struct delayed_work hwp_notify_work;
264 };
265 
266 static struct cpudata **all_cpu_data;
267 
268 /**
269  * struct pstate_funcs - Per CPU model specific callbacks
270  * @get_max:		Callback to get maximum non turbo effective P state
271  * @get_max_physical:	Callback to get maximum non turbo physical P state
272  * @get_min:		Callback to get minimum P state
273  * @get_turbo:		Callback to get turbo P state
274  * @get_scaling:	Callback to get frequency scaling factor
275  * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
276  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
277  * @get_val:		Callback to convert P state to actual MSR write value
278  * @get_vid:		Callback to get VID data for Atom platforms
279  *
280  * Core and Atom CPU models have different way to get P State limits. This
281  * structure is used to store those callbacks.
282  */
283 struct pstate_funcs {
284 	int (*get_max)(int cpu);
285 	int (*get_max_physical)(int cpu);
286 	int (*get_min)(int cpu);
287 	int (*get_turbo)(int cpu);
288 	int (*get_scaling)(void);
289 	int (*get_cpu_scaling)(int cpu);
290 	int (*get_aperf_mperf_shift)(void);
291 	u64 (*get_val)(struct cpudata*, int pstate);
292 	void (*get_vid)(struct cpudata *);
293 };
294 
295 static struct pstate_funcs pstate_funcs __read_mostly;
296 
297 static int hwp_active __read_mostly;
298 static int hwp_mode_bdw __read_mostly;
299 static bool per_cpu_limits __read_mostly;
300 static bool hwp_boost __read_mostly;
301 static bool hwp_forced __read_mostly;
302 
303 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
304 
305 #ifdef CONFIG_ACPI
306 static bool acpi_ppc;
307 #endif
308 
309 static struct global_params global;
310 
311 static DEFINE_MUTEX(intel_pstate_driver_lock);
312 static DEFINE_MUTEX(intel_pstate_limits_lock);
313 
314 #ifdef CONFIG_ACPI
315 
316 static bool intel_pstate_acpi_pm_profile_server(void)
317 {
318 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
319 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
320 		return true;
321 
322 	return false;
323 }
324 
325 static bool intel_pstate_get_ppc_enable_status(void)
326 {
327 	if (intel_pstate_acpi_pm_profile_server())
328 		return true;
329 
330 	return acpi_ppc;
331 }
332 
333 #ifdef CONFIG_ACPI_CPPC_LIB
334 
335 /* The work item is needed to avoid CPU hotplug locking issues */
336 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
337 {
338 	sched_set_itmt_support();
339 }
340 
341 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
342 
343 #define CPPC_MAX_PERF	U8_MAX
344 
345 static void intel_pstate_set_itmt_prio(int cpu)
346 {
347 	struct cppc_perf_caps cppc_perf;
348 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
349 	int ret;
350 
351 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
352 	if (ret)
353 		return;
354 
355 	/*
356 	 * On some systems with overclocking enabled, CPPC.highest_perf is hardcoded to 0xff.
357 	 * In this case we can't use CPPC.highest_perf to enable ITMT.
358 	 * In this case we can look at MSR_HWP_CAPABILITIES bits [8:0] to decide.
359 	 */
360 	if (cppc_perf.highest_perf == CPPC_MAX_PERF)
361 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
362 
363 	/*
364 	 * The priorities can be set regardless of whether or not
365 	 * sched_set_itmt_support(true) has been called and it is valid to
366 	 * update them at any time after it has been called.
367 	 */
368 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
369 
370 	if (max_highest_perf <= min_highest_perf) {
371 		if (cppc_perf.highest_perf > max_highest_perf)
372 			max_highest_perf = cppc_perf.highest_perf;
373 
374 		if (cppc_perf.highest_perf < min_highest_perf)
375 			min_highest_perf = cppc_perf.highest_perf;
376 
377 		if (max_highest_perf > min_highest_perf) {
378 			/*
379 			 * This code can be run during CPU online under the
380 			 * CPU hotplug locks, so sched_set_itmt_support()
381 			 * cannot be called from here.  Queue up a work item
382 			 * to invoke it.
383 			 */
384 			schedule_work(&sched_itmt_work);
385 		}
386 	}
387 }
388 
389 static int intel_pstate_get_cppc_guaranteed(int cpu)
390 {
391 	struct cppc_perf_caps cppc_perf;
392 	int ret;
393 
394 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
395 	if (ret)
396 		return ret;
397 
398 	if (cppc_perf.guaranteed_perf)
399 		return cppc_perf.guaranteed_perf;
400 
401 	return cppc_perf.nominal_perf;
402 }
403 #else /* CONFIG_ACPI_CPPC_LIB */
404 static inline void intel_pstate_set_itmt_prio(int cpu)
405 {
406 }
407 #endif /* CONFIG_ACPI_CPPC_LIB */
408 
409 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
410 {
411 	struct cpudata *cpu;
412 	int ret;
413 	int i;
414 
415 	if (hwp_active) {
416 		intel_pstate_set_itmt_prio(policy->cpu);
417 		return;
418 	}
419 
420 	if (!intel_pstate_get_ppc_enable_status())
421 		return;
422 
423 	cpu = all_cpu_data[policy->cpu];
424 
425 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
426 						  policy->cpu);
427 	if (ret)
428 		return;
429 
430 	/*
431 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
432 	 * guarantee that the states returned by it map to the states in our
433 	 * list directly.
434 	 */
435 	if (cpu->acpi_perf_data.control_register.space_id !=
436 						ACPI_ADR_SPACE_FIXED_HARDWARE)
437 		goto err;
438 
439 	/*
440 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
441 	 * usual without taking _PSS into account
442 	 */
443 	if (cpu->acpi_perf_data.state_count < 2)
444 		goto err;
445 
446 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
447 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
448 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
449 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
450 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
451 			 (u32) cpu->acpi_perf_data.states[i].power,
452 			 (u32) cpu->acpi_perf_data.states[i].control);
453 	}
454 
455 	cpu->valid_pss_table = true;
456 	pr_debug("_PPC limits will be enforced\n");
457 
458 	return;
459 
460  err:
461 	cpu->valid_pss_table = false;
462 	acpi_processor_unregister_performance(policy->cpu);
463 }
464 
465 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
466 {
467 	struct cpudata *cpu;
468 
469 	cpu = all_cpu_data[policy->cpu];
470 	if (!cpu->valid_pss_table)
471 		return;
472 
473 	acpi_processor_unregister_performance(policy->cpu);
474 }
475 #else /* CONFIG_ACPI */
476 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
477 {
478 }
479 
480 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
481 {
482 }
483 
484 static inline bool intel_pstate_acpi_pm_profile_server(void)
485 {
486 	return false;
487 }
488 #endif /* CONFIG_ACPI */
489 
490 #ifndef CONFIG_ACPI_CPPC_LIB
491 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
492 {
493 	return -ENOTSUPP;
494 }
495 #endif /* CONFIG_ACPI_CPPC_LIB */
496 
497 /**
498  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
499  * @cpu: Target CPU.
500  *
501  * On hybrid processors, HWP may expose more performance levels than there are
502  * P-states accessible through the PERF_CTL interface.  If that happens, the
503  * scaling factor between HWP performance levels and CPU frequency will be less
504  * than the scaling factor between P-state values and CPU frequency.
505  *
506  * In that case, adjust the CPU parameters used in computations accordingly.
507  */
508 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
509 {
510 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
511 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
512 	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
513 	int scaling = cpu->pstate.scaling;
514 
515 	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
516 	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
517 	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
518 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
519 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
520 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
521 
522 	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
523 					   perf_ctl_scaling);
524 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
525 					 perf_ctl_scaling);
526 
527 	cpu->pstate.max_pstate_physical =
528 			DIV_ROUND_UP(perf_ctl_max_phys * perf_ctl_scaling,
529 				     scaling);
530 
531 	cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
532 	/*
533 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
534 	 * the effective range of HWP performance levels.
535 	 */
536 	cpu->pstate.min_pstate = DIV_ROUND_UP(cpu->pstate.min_freq, scaling);
537 }
538 
539 static inline void update_turbo_state(void)
540 {
541 	u64 misc_en;
542 	struct cpudata *cpu;
543 
544 	cpu = all_cpu_data[0];
545 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
546 	global.turbo_disabled =
547 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
548 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
549 }
550 
551 static int min_perf_pct_min(void)
552 {
553 	struct cpudata *cpu = all_cpu_data[0];
554 	int turbo_pstate = cpu->pstate.turbo_pstate;
555 
556 	return turbo_pstate ?
557 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
558 }
559 
560 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
561 {
562 	u64 epb;
563 	int ret;
564 
565 	if (!boot_cpu_has(X86_FEATURE_EPB))
566 		return -ENXIO;
567 
568 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
569 	if (ret)
570 		return (s16)ret;
571 
572 	return (s16)(epb & 0x0f);
573 }
574 
575 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
576 {
577 	s16 epp;
578 
579 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
580 		/*
581 		 * When hwp_req_data is 0, means that caller didn't read
582 		 * MSR_HWP_REQUEST, so need to read and get EPP.
583 		 */
584 		if (!hwp_req_data) {
585 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
586 					    &hwp_req_data);
587 			if (epp)
588 				return epp;
589 		}
590 		epp = (hwp_req_data >> 24) & 0xff;
591 	} else {
592 		/* When there is no EPP present, HWP uses EPB settings */
593 		epp = intel_pstate_get_epb(cpu_data);
594 	}
595 
596 	return epp;
597 }
598 
599 static int intel_pstate_set_epb(int cpu, s16 pref)
600 {
601 	u64 epb;
602 	int ret;
603 
604 	if (!boot_cpu_has(X86_FEATURE_EPB))
605 		return -ENXIO;
606 
607 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
608 	if (ret)
609 		return ret;
610 
611 	epb = (epb & ~0x0f) | pref;
612 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
613 
614 	return 0;
615 }
616 
617 /*
618  * EPP/EPB display strings corresponding to EPP index in the
619  * energy_perf_strings[]
620  *	index		String
621  *-------------------------------------
622  *	0		default
623  *	1		performance
624  *	2		balance_performance
625  *	3		balance_power
626  *	4		power
627  */
628 
629 enum energy_perf_value_index {
630 	EPP_INDEX_DEFAULT = 0,
631 	EPP_INDEX_PERFORMANCE,
632 	EPP_INDEX_BALANCE_PERFORMANCE,
633 	EPP_INDEX_BALANCE_POWERSAVE,
634 	EPP_INDEX_POWERSAVE,
635 };
636 
637 static const char * const energy_perf_strings[] = {
638 	[EPP_INDEX_DEFAULT] = "default",
639 	[EPP_INDEX_PERFORMANCE] = "performance",
640 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
641 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
642 	[EPP_INDEX_POWERSAVE] = "power",
643 	NULL
644 };
645 static unsigned int epp_values[] = {
646 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
647 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
648 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
649 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
650 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
651 };
652 
653 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
654 {
655 	s16 epp;
656 	int index = -EINVAL;
657 
658 	*raw_epp = 0;
659 	epp = intel_pstate_get_epp(cpu_data, 0);
660 	if (epp < 0)
661 		return epp;
662 
663 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
664 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
665 			return EPP_INDEX_PERFORMANCE;
666 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
667 			return EPP_INDEX_BALANCE_PERFORMANCE;
668 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
669 			return EPP_INDEX_BALANCE_POWERSAVE;
670 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
671 			return EPP_INDEX_POWERSAVE;
672 		*raw_epp = epp;
673 		return 0;
674 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
675 		/*
676 		 * Range:
677 		 *	0x00-0x03	:	Performance
678 		 *	0x04-0x07	:	Balance performance
679 		 *	0x08-0x0B	:	Balance power
680 		 *	0x0C-0x0F	:	Power
681 		 * The EPB is a 4 bit value, but our ranges restrict the
682 		 * value which can be set. Here only using top two bits
683 		 * effectively.
684 		 */
685 		index = (epp >> 2) + 1;
686 	}
687 
688 	return index;
689 }
690 
691 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
692 {
693 	int ret;
694 
695 	/*
696 	 * Use the cached HWP Request MSR value, because in the active mode the
697 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
698 	 * intel_pstate_hwp_boost_down() at any time.
699 	 */
700 	u64 value = READ_ONCE(cpu->hwp_req_cached);
701 
702 	value &= ~GENMASK_ULL(31, 24);
703 	value |= (u64)epp << 24;
704 	/*
705 	 * The only other updater of hwp_req_cached in the active mode,
706 	 * intel_pstate_hwp_set(), is called under the same lock as this
707 	 * function, so it cannot run in parallel with the update below.
708 	 */
709 	WRITE_ONCE(cpu->hwp_req_cached, value);
710 	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
711 	if (!ret)
712 		cpu->epp_cached = epp;
713 
714 	return ret;
715 }
716 
717 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
718 					      int pref_index, bool use_raw,
719 					      u32 raw_epp)
720 {
721 	int epp = -EINVAL;
722 	int ret;
723 
724 	if (!pref_index)
725 		epp = cpu_data->epp_default;
726 
727 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
728 		if (use_raw)
729 			epp = raw_epp;
730 		else if (epp == -EINVAL)
731 			epp = epp_values[pref_index];
732 
733 		/*
734 		 * To avoid confusion, refuse to set EPP to any values different
735 		 * from 0 (performance) if the current policy is "performance",
736 		 * because those values would be overridden.
737 		 */
738 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
739 			return -EBUSY;
740 
741 		ret = intel_pstate_set_epp(cpu_data, epp);
742 	} else {
743 		if (epp == -EINVAL)
744 			epp = (pref_index - 1) << 2;
745 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
746 	}
747 
748 	return ret;
749 }
750 
751 static ssize_t show_energy_performance_available_preferences(
752 				struct cpufreq_policy *policy, char *buf)
753 {
754 	int i = 0;
755 	int ret = 0;
756 
757 	while (energy_perf_strings[i] != NULL)
758 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
759 
760 	ret += sprintf(&buf[ret], "\n");
761 
762 	return ret;
763 }
764 
765 cpufreq_freq_attr_ro(energy_performance_available_preferences);
766 
767 static struct cpufreq_driver intel_pstate;
768 
769 static ssize_t store_energy_performance_preference(
770 		struct cpufreq_policy *policy, const char *buf, size_t count)
771 {
772 	struct cpudata *cpu = all_cpu_data[policy->cpu];
773 	char str_preference[21];
774 	bool raw = false;
775 	ssize_t ret;
776 	u32 epp = 0;
777 
778 	ret = sscanf(buf, "%20s", str_preference);
779 	if (ret != 1)
780 		return -EINVAL;
781 
782 	ret = match_string(energy_perf_strings, -1, str_preference);
783 	if (ret < 0) {
784 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
785 			return ret;
786 
787 		ret = kstrtouint(buf, 10, &epp);
788 		if (ret)
789 			return ret;
790 
791 		if (epp > 255)
792 			return -EINVAL;
793 
794 		raw = true;
795 	}
796 
797 	/*
798 	 * This function runs with the policy R/W semaphore held, which
799 	 * guarantees that the driver pointer will not change while it is
800 	 * running.
801 	 */
802 	if (!intel_pstate_driver)
803 		return -EAGAIN;
804 
805 	mutex_lock(&intel_pstate_limits_lock);
806 
807 	if (intel_pstate_driver == &intel_pstate) {
808 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
809 	} else {
810 		/*
811 		 * In the passive mode the governor needs to be stopped on the
812 		 * target CPU before the EPP update and restarted after it,
813 		 * which is super-heavy-weight, so make sure it is worth doing
814 		 * upfront.
815 		 */
816 		if (!raw)
817 			epp = ret ? epp_values[ret] : cpu->epp_default;
818 
819 		if (cpu->epp_cached != epp) {
820 			int err;
821 
822 			cpufreq_stop_governor(policy);
823 			ret = intel_pstate_set_epp(cpu, epp);
824 			err = cpufreq_start_governor(policy);
825 			if (!ret)
826 				ret = err;
827 		} else {
828 			ret = 0;
829 		}
830 	}
831 
832 	mutex_unlock(&intel_pstate_limits_lock);
833 
834 	return ret ?: count;
835 }
836 
837 static ssize_t show_energy_performance_preference(
838 				struct cpufreq_policy *policy, char *buf)
839 {
840 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
841 	int preference, raw_epp;
842 
843 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
844 	if (preference < 0)
845 		return preference;
846 
847 	if (raw_epp)
848 		return  sprintf(buf, "%d\n", raw_epp);
849 	else
850 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
851 }
852 
853 cpufreq_freq_attr_rw(energy_performance_preference);
854 
855 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
856 {
857 	struct cpudata *cpu = all_cpu_data[policy->cpu];
858 	int ratio, freq;
859 
860 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
861 	if (ratio <= 0) {
862 		u64 cap;
863 
864 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
865 		ratio = HWP_GUARANTEED_PERF(cap);
866 	}
867 
868 	freq = ratio * cpu->pstate.scaling;
869 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
870 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
871 
872 	return sprintf(buf, "%d\n", freq);
873 }
874 
875 cpufreq_freq_attr_ro(base_frequency);
876 
877 static struct freq_attr *hwp_cpufreq_attrs[] = {
878 	&energy_performance_preference,
879 	&energy_performance_available_preferences,
880 	&base_frequency,
881 	NULL,
882 };
883 
884 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
885 {
886 	u64 cap;
887 
888 	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
889 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
890 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
891 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
892 }
893 
894 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
895 {
896 	int scaling = cpu->pstate.scaling;
897 
898 	__intel_pstate_get_hwp_cap(cpu);
899 
900 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
901 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
902 	if (scaling != cpu->pstate.perf_ctl_scaling) {
903 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
904 
905 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
906 						 perf_ctl_scaling);
907 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
908 						   perf_ctl_scaling);
909 	}
910 }
911 
912 static void intel_pstate_hwp_set(unsigned int cpu)
913 {
914 	struct cpudata *cpu_data = all_cpu_data[cpu];
915 	int max, min;
916 	u64 value;
917 	s16 epp;
918 
919 	max = cpu_data->max_perf_ratio;
920 	min = cpu_data->min_perf_ratio;
921 
922 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
923 		min = max;
924 
925 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
926 
927 	value &= ~HWP_MIN_PERF(~0L);
928 	value |= HWP_MIN_PERF(min);
929 
930 	value &= ~HWP_MAX_PERF(~0L);
931 	value |= HWP_MAX_PERF(max);
932 
933 	if (cpu_data->epp_policy == cpu_data->policy)
934 		goto skip_epp;
935 
936 	cpu_data->epp_policy = cpu_data->policy;
937 
938 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
939 		epp = intel_pstate_get_epp(cpu_data, value);
940 		cpu_data->epp_powersave = epp;
941 		/* If EPP read was failed, then don't try to write */
942 		if (epp < 0)
943 			goto skip_epp;
944 
945 		epp = 0;
946 	} else {
947 		/* skip setting EPP, when saved value is invalid */
948 		if (cpu_data->epp_powersave < 0)
949 			goto skip_epp;
950 
951 		/*
952 		 * No need to restore EPP when it is not zero. This
953 		 * means:
954 		 *  - Policy is not changed
955 		 *  - user has manually changed
956 		 *  - Error reading EPB
957 		 */
958 		epp = intel_pstate_get_epp(cpu_data, value);
959 		if (epp)
960 			goto skip_epp;
961 
962 		epp = cpu_data->epp_powersave;
963 	}
964 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
965 		value &= ~GENMASK_ULL(31, 24);
966 		value |= (u64)epp << 24;
967 	} else {
968 		intel_pstate_set_epb(cpu, epp);
969 	}
970 skip_epp:
971 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
972 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
973 }
974 
975 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
976 
977 static void intel_pstate_hwp_offline(struct cpudata *cpu)
978 {
979 	u64 value = READ_ONCE(cpu->hwp_req_cached);
980 	int min_perf;
981 
982 	intel_pstate_disable_hwp_interrupt(cpu);
983 
984 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
985 		/*
986 		 * In case the EPP has been set to "performance" by the
987 		 * active mode "performance" scaling algorithm, replace that
988 		 * temporary value with the cached EPP one.
989 		 */
990 		value &= ~GENMASK_ULL(31, 24);
991 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
992 		/*
993 		 * However, make sure that EPP will be set to "performance" when
994 		 * the CPU is brought back online again and the "performance"
995 		 * scaling algorithm is still in effect.
996 		 */
997 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
998 	}
999 
1000 	/*
1001 	 * Clear the desired perf field in the cached HWP request value to
1002 	 * prevent nonzero desired values from being leaked into the active
1003 	 * mode.
1004 	 */
1005 	value &= ~HWP_DESIRED_PERF(~0L);
1006 	WRITE_ONCE(cpu->hwp_req_cached, value);
1007 
1008 	value &= ~GENMASK_ULL(31, 0);
1009 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1010 
1011 	/* Set hwp_max = hwp_min */
1012 	value |= HWP_MAX_PERF(min_perf);
1013 	value |= HWP_MIN_PERF(min_perf);
1014 
1015 	/* Set EPP to min */
1016 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1017 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1018 
1019 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1020 }
1021 
1022 #define POWER_CTL_EE_ENABLE	1
1023 #define POWER_CTL_EE_DISABLE	2
1024 
1025 static int power_ctl_ee_state;
1026 
1027 static void set_power_ctl_ee_state(bool input)
1028 {
1029 	u64 power_ctl;
1030 
1031 	mutex_lock(&intel_pstate_driver_lock);
1032 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1033 	if (input) {
1034 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1035 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1036 	} else {
1037 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1038 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1039 	}
1040 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1041 	mutex_unlock(&intel_pstate_driver_lock);
1042 }
1043 
1044 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1045 
1046 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1047 {
1048 	intel_pstate_hwp_enable(cpu);
1049 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1050 }
1051 
1052 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1053 {
1054 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1055 
1056 	pr_debug("CPU %d suspending\n", cpu->cpu);
1057 
1058 	cpu->suspended = true;
1059 
1060 	/* disable HWP interrupt and cancel any pending work */
1061 	intel_pstate_disable_hwp_interrupt(cpu);
1062 
1063 	return 0;
1064 }
1065 
1066 static int intel_pstate_resume(struct cpufreq_policy *policy)
1067 {
1068 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1069 
1070 	pr_debug("CPU %d resuming\n", cpu->cpu);
1071 
1072 	/* Only restore if the system default is changed */
1073 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1074 		set_power_ctl_ee_state(true);
1075 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1076 		set_power_ctl_ee_state(false);
1077 
1078 	if (cpu->suspended && hwp_active) {
1079 		mutex_lock(&intel_pstate_limits_lock);
1080 
1081 		/* Re-enable HWP, because "online" has not done that. */
1082 		intel_pstate_hwp_reenable(cpu);
1083 
1084 		mutex_unlock(&intel_pstate_limits_lock);
1085 	}
1086 
1087 	cpu->suspended = false;
1088 
1089 	return 0;
1090 }
1091 
1092 static void intel_pstate_update_policies(void)
1093 {
1094 	int cpu;
1095 
1096 	for_each_possible_cpu(cpu)
1097 		cpufreq_update_policy(cpu);
1098 }
1099 
1100 static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1101 					   struct cpufreq_policy *policy)
1102 {
1103 	policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
1104 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1105 	refresh_frequency_limits(policy);
1106 }
1107 
1108 static void intel_pstate_update_max_freq(unsigned int cpu)
1109 {
1110 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1111 
1112 	if (!policy)
1113 		return;
1114 
1115 	__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1116 
1117 	cpufreq_cpu_release(policy);
1118 }
1119 
1120 static void intel_pstate_update_limits(unsigned int cpu)
1121 {
1122 	mutex_lock(&intel_pstate_driver_lock);
1123 
1124 	update_turbo_state();
1125 	/*
1126 	 * If turbo has been turned on or off globally, policy limits for
1127 	 * all CPUs need to be updated to reflect that.
1128 	 */
1129 	if (global.turbo_disabled_mf != global.turbo_disabled) {
1130 		global.turbo_disabled_mf = global.turbo_disabled;
1131 		arch_set_max_freq_ratio(global.turbo_disabled);
1132 		for_each_possible_cpu(cpu)
1133 			intel_pstate_update_max_freq(cpu);
1134 	} else {
1135 		cpufreq_update_policy(cpu);
1136 	}
1137 
1138 	mutex_unlock(&intel_pstate_driver_lock);
1139 }
1140 
1141 /************************** sysfs begin ************************/
1142 #define show_one(file_name, object)					\
1143 	static ssize_t show_##file_name					\
1144 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1145 	{								\
1146 		return sprintf(buf, "%u\n", global.object);		\
1147 	}
1148 
1149 static ssize_t intel_pstate_show_status(char *buf);
1150 static int intel_pstate_update_status(const char *buf, size_t size);
1151 
1152 static ssize_t show_status(struct kobject *kobj,
1153 			   struct kobj_attribute *attr, char *buf)
1154 {
1155 	ssize_t ret;
1156 
1157 	mutex_lock(&intel_pstate_driver_lock);
1158 	ret = intel_pstate_show_status(buf);
1159 	mutex_unlock(&intel_pstate_driver_lock);
1160 
1161 	return ret;
1162 }
1163 
1164 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1165 			    const char *buf, size_t count)
1166 {
1167 	char *p = memchr(buf, '\n', count);
1168 	int ret;
1169 
1170 	mutex_lock(&intel_pstate_driver_lock);
1171 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1172 	mutex_unlock(&intel_pstate_driver_lock);
1173 
1174 	return ret < 0 ? ret : count;
1175 }
1176 
1177 static ssize_t show_turbo_pct(struct kobject *kobj,
1178 				struct kobj_attribute *attr, char *buf)
1179 {
1180 	struct cpudata *cpu;
1181 	int total, no_turbo, turbo_pct;
1182 	uint32_t turbo_fp;
1183 
1184 	mutex_lock(&intel_pstate_driver_lock);
1185 
1186 	if (!intel_pstate_driver) {
1187 		mutex_unlock(&intel_pstate_driver_lock);
1188 		return -EAGAIN;
1189 	}
1190 
1191 	cpu = all_cpu_data[0];
1192 
1193 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1194 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1195 	turbo_fp = div_fp(no_turbo, total);
1196 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1197 
1198 	mutex_unlock(&intel_pstate_driver_lock);
1199 
1200 	return sprintf(buf, "%u\n", turbo_pct);
1201 }
1202 
1203 static ssize_t show_num_pstates(struct kobject *kobj,
1204 				struct kobj_attribute *attr, char *buf)
1205 {
1206 	struct cpudata *cpu;
1207 	int total;
1208 
1209 	mutex_lock(&intel_pstate_driver_lock);
1210 
1211 	if (!intel_pstate_driver) {
1212 		mutex_unlock(&intel_pstate_driver_lock);
1213 		return -EAGAIN;
1214 	}
1215 
1216 	cpu = all_cpu_data[0];
1217 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1218 
1219 	mutex_unlock(&intel_pstate_driver_lock);
1220 
1221 	return sprintf(buf, "%u\n", total);
1222 }
1223 
1224 static ssize_t show_no_turbo(struct kobject *kobj,
1225 			     struct kobj_attribute *attr, char *buf)
1226 {
1227 	ssize_t ret;
1228 
1229 	mutex_lock(&intel_pstate_driver_lock);
1230 
1231 	if (!intel_pstate_driver) {
1232 		mutex_unlock(&intel_pstate_driver_lock);
1233 		return -EAGAIN;
1234 	}
1235 
1236 	update_turbo_state();
1237 	if (global.turbo_disabled)
1238 		ret = sprintf(buf, "%u\n", global.turbo_disabled);
1239 	else
1240 		ret = sprintf(buf, "%u\n", global.no_turbo);
1241 
1242 	mutex_unlock(&intel_pstate_driver_lock);
1243 
1244 	return ret;
1245 }
1246 
1247 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1248 			      const char *buf, size_t count)
1249 {
1250 	unsigned int input;
1251 	int ret;
1252 
1253 	ret = sscanf(buf, "%u", &input);
1254 	if (ret != 1)
1255 		return -EINVAL;
1256 
1257 	mutex_lock(&intel_pstate_driver_lock);
1258 
1259 	if (!intel_pstate_driver) {
1260 		mutex_unlock(&intel_pstate_driver_lock);
1261 		return -EAGAIN;
1262 	}
1263 
1264 	mutex_lock(&intel_pstate_limits_lock);
1265 
1266 	update_turbo_state();
1267 	if (global.turbo_disabled) {
1268 		pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1269 		mutex_unlock(&intel_pstate_limits_lock);
1270 		mutex_unlock(&intel_pstate_driver_lock);
1271 		return -EPERM;
1272 	}
1273 
1274 	global.no_turbo = clamp_t(int, input, 0, 1);
1275 
1276 	if (global.no_turbo) {
1277 		struct cpudata *cpu = all_cpu_data[0];
1278 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1279 
1280 		/* Squash the global minimum into the permitted range. */
1281 		if (global.min_perf_pct > pct)
1282 			global.min_perf_pct = pct;
1283 	}
1284 
1285 	mutex_unlock(&intel_pstate_limits_lock);
1286 
1287 	intel_pstate_update_policies();
1288 	arch_set_max_freq_ratio(global.no_turbo);
1289 
1290 	mutex_unlock(&intel_pstate_driver_lock);
1291 
1292 	return count;
1293 }
1294 
1295 static void update_qos_request(enum freq_qos_req_type type)
1296 {
1297 	struct freq_qos_request *req;
1298 	struct cpufreq_policy *policy;
1299 	int i;
1300 
1301 	for_each_possible_cpu(i) {
1302 		struct cpudata *cpu = all_cpu_data[i];
1303 		unsigned int freq, perf_pct;
1304 
1305 		policy = cpufreq_cpu_get(i);
1306 		if (!policy)
1307 			continue;
1308 
1309 		req = policy->driver_data;
1310 		cpufreq_cpu_put(policy);
1311 
1312 		if (!req)
1313 			continue;
1314 
1315 		if (hwp_active)
1316 			intel_pstate_get_hwp_cap(cpu);
1317 
1318 		if (type == FREQ_QOS_MIN) {
1319 			perf_pct = global.min_perf_pct;
1320 		} else {
1321 			req++;
1322 			perf_pct = global.max_perf_pct;
1323 		}
1324 
1325 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1326 
1327 		if (freq_qos_update_request(req, freq) < 0)
1328 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1329 	}
1330 }
1331 
1332 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1333 				  const char *buf, size_t count)
1334 {
1335 	unsigned int input;
1336 	int ret;
1337 
1338 	ret = sscanf(buf, "%u", &input);
1339 	if (ret != 1)
1340 		return -EINVAL;
1341 
1342 	mutex_lock(&intel_pstate_driver_lock);
1343 
1344 	if (!intel_pstate_driver) {
1345 		mutex_unlock(&intel_pstate_driver_lock);
1346 		return -EAGAIN;
1347 	}
1348 
1349 	mutex_lock(&intel_pstate_limits_lock);
1350 
1351 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1352 
1353 	mutex_unlock(&intel_pstate_limits_lock);
1354 
1355 	if (intel_pstate_driver == &intel_pstate)
1356 		intel_pstate_update_policies();
1357 	else
1358 		update_qos_request(FREQ_QOS_MAX);
1359 
1360 	mutex_unlock(&intel_pstate_driver_lock);
1361 
1362 	return count;
1363 }
1364 
1365 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1366 				  const char *buf, size_t count)
1367 {
1368 	unsigned int input;
1369 	int ret;
1370 
1371 	ret = sscanf(buf, "%u", &input);
1372 	if (ret != 1)
1373 		return -EINVAL;
1374 
1375 	mutex_lock(&intel_pstate_driver_lock);
1376 
1377 	if (!intel_pstate_driver) {
1378 		mutex_unlock(&intel_pstate_driver_lock);
1379 		return -EAGAIN;
1380 	}
1381 
1382 	mutex_lock(&intel_pstate_limits_lock);
1383 
1384 	global.min_perf_pct = clamp_t(int, input,
1385 				      min_perf_pct_min(), global.max_perf_pct);
1386 
1387 	mutex_unlock(&intel_pstate_limits_lock);
1388 
1389 	if (intel_pstate_driver == &intel_pstate)
1390 		intel_pstate_update_policies();
1391 	else
1392 		update_qos_request(FREQ_QOS_MIN);
1393 
1394 	mutex_unlock(&intel_pstate_driver_lock);
1395 
1396 	return count;
1397 }
1398 
1399 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1400 				struct kobj_attribute *attr, char *buf)
1401 {
1402 	return sprintf(buf, "%u\n", hwp_boost);
1403 }
1404 
1405 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1406 				       struct kobj_attribute *b,
1407 				       const char *buf, size_t count)
1408 {
1409 	unsigned int input;
1410 	int ret;
1411 
1412 	ret = kstrtouint(buf, 10, &input);
1413 	if (ret)
1414 		return ret;
1415 
1416 	mutex_lock(&intel_pstate_driver_lock);
1417 	hwp_boost = !!input;
1418 	intel_pstate_update_policies();
1419 	mutex_unlock(&intel_pstate_driver_lock);
1420 
1421 	return count;
1422 }
1423 
1424 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1425 				      char *buf)
1426 {
1427 	u64 power_ctl;
1428 	int enable;
1429 
1430 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1431 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1432 	return sprintf(buf, "%d\n", !enable);
1433 }
1434 
1435 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1436 				       const char *buf, size_t count)
1437 {
1438 	bool input;
1439 	int ret;
1440 
1441 	ret = kstrtobool(buf, &input);
1442 	if (ret)
1443 		return ret;
1444 
1445 	set_power_ctl_ee_state(input);
1446 
1447 	return count;
1448 }
1449 
1450 show_one(max_perf_pct, max_perf_pct);
1451 show_one(min_perf_pct, min_perf_pct);
1452 
1453 define_one_global_rw(status);
1454 define_one_global_rw(no_turbo);
1455 define_one_global_rw(max_perf_pct);
1456 define_one_global_rw(min_perf_pct);
1457 define_one_global_ro(turbo_pct);
1458 define_one_global_ro(num_pstates);
1459 define_one_global_rw(hwp_dynamic_boost);
1460 define_one_global_rw(energy_efficiency);
1461 
1462 static struct attribute *intel_pstate_attributes[] = {
1463 	&status.attr,
1464 	&no_turbo.attr,
1465 	NULL
1466 };
1467 
1468 static const struct attribute_group intel_pstate_attr_group = {
1469 	.attrs = intel_pstate_attributes,
1470 };
1471 
1472 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1473 
1474 static struct kobject *intel_pstate_kobject;
1475 
1476 static void __init intel_pstate_sysfs_expose_params(void)
1477 {
1478 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1479 	int rc;
1480 
1481 	if (dev_root) {
1482 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1483 		put_device(dev_root);
1484 	}
1485 	if (WARN_ON(!intel_pstate_kobject))
1486 		return;
1487 
1488 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1489 	if (WARN_ON(rc))
1490 		return;
1491 
1492 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1493 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1494 		WARN_ON(rc);
1495 
1496 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1497 		WARN_ON(rc);
1498 	}
1499 
1500 	/*
1501 	 * If per cpu limits are enforced there are no global limits, so
1502 	 * return without creating max/min_perf_pct attributes
1503 	 */
1504 	if (per_cpu_limits)
1505 		return;
1506 
1507 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1508 	WARN_ON(rc);
1509 
1510 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1511 	WARN_ON(rc);
1512 
1513 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1514 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1515 		WARN_ON(rc);
1516 	}
1517 }
1518 
1519 static void __init intel_pstate_sysfs_remove(void)
1520 {
1521 	if (!intel_pstate_kobject)
1522 		return;
1523 
1524 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1525 
1526 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1527 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1528 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1529 	}
1530 
1531 	if (!per_cpu_limits) {
1532 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1533 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1534 
1535 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1536 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1537 	}
1538 
1539 	kobject_put(intel_pstate_kobject);
1540 }
1541 
1542 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1543 {
1544 	int rc;
1545 
1546 	if (!hwp_active)
1547 		return;
1548 
1549 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1550 	WARN_ON_ONCE(rc);
1551 }
1552 
1553 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1554 {
1555 	if (!hwp_active)
1556 		return;
1557 
1558 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1559 }
1560 
1561 /************************** sysfs end ************************/
1562 
1563 static void intel_pstate_notify_work(struct work_struct *work)
1564 {
1565 	struct cpudata *cpudata =
1566 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1567 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1568 
1569 	if (policy) {
1570 		intel_pstate_get_hwp_cap(cpudata);
1571 		__intel_pstate_update_max_freq(cpudata, policy);
1572 
1573 		cpufreq_cpu_release(policy);
1574 	}
1575 
1576 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1577 }
1578 
1579 static DEFINE_SPINLOCK(hwp_notify_lock);
1580 static cpumask_t hwp_intr_enable_mask;
1581 
1582 void notify_hwp_interrupt(void)
1583 {
1584 	unsigned int this_cpu = smp_processor_id();
1585 	struct cpudata *cpudata;
1586 	unsigned long flags;
1587 	u64 value;
1588 
1589 	if (!READ_ONCE(hwp_active) || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1590 		return;
1591 
1592 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1593 	if (!(value & 0x01))
1594 		return;
1595 
1596 	spin_lock_irqsave(&hwp_notify_lock, flags);
1597 
1598 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1599 		goto ack_intr;
1600 
1601 	/*
1602 	 * Currently we never free all_cpu_data. And we can't reach here
1603 	 * without this allocated. But for safety for future changes, added
1604 	 * check.
1605 	 */
1606 	if (unlikely(!READ_ONCE(all_cpu_data)))
1607 		goto ack_intr;
1608 
1609 	/*
1610 	 * The free is done during cleanup, when cpufreq registry is failed.
1611 	 * We wouldn't be here if it fails on init or switch status. But for
1612 	 * future changes, added check.
1613 	 */
1614 	cpudata = READ_ONCE(all_cpu_data[this_cpu]);
1615 	if (unlikely(!cpudata))
1616 		goto ack_intr;
1617 
1618 	schedule_delayed_work(&cpudata->hwp_notify_work, msecs_to_jiffies(10));
1619 
1620 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1621 
1622 	return;
1623 
1624 ack_intr:
1625 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1626 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1627 }
1628 
1629 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1630 {
1631 	unsigned long flags;
1632 
1633 	if (!boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1634 		return;
1635 
1636 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1637 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1638 
1639 	spin_lock_irqsave(&hwp_notify_lock, flags);
1640 	if (cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask))
1641 		cancel_delayed_work(&cpudata->hwp_notify_work);
1642 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1643 }
1644 
1645 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1646 {
1647 	/* Enable HWP notification interrupt for guaranteed performance change */
1648 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1649 		unsigned long flags;
1650 
1651 		spin_lock_irqsave(&hwp_notify_lock, flags);
1652 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1653 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1654 		spin_unlock_irqrestore(&hwp_notify_lock, flags);
1655 
1656 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1657 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x01);
1658 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1659 	}
1660 }
1661 
1662 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1663 {
1664 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1665 
1666 	/*
1667 	 * If this CPU gen doesn't call for change in balance_perf
1668 	 * EPP return.
1669 	 */
1670 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1671 		return;
1672 
1673 	/*
1674 	 * If the EPP is set by firmware, which means that firmware enabled HWP
1675 	 * - Is equal or less than 0x80 (default balance_perf EPP)
1676 	 * - But less performance oriented than performance EPP
1677 	 *   then use this as new balance_perf EPP.
1678 	 */
1679 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1680 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1681 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1682 		return;
1683 	}
1684 
1685 	/*
1686 	 * Use hard coded value per gen to update the balance_perf
1687 	 * and default EPP.
1688 	 */
1689 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1690 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1691 }
1692 
1693 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1694 {
1695 	/* First disable HWP notification interrupt till we activate again */
1696 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1697 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1698 
1699 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1700 
1701 	intel_pstate_enable_hwp_interrupt(cpudata);
1702 
1703 	if (cpudata->epp_default >= 0)
1704 		return;
1705 
1706 	intel_pstate_update_epp_defaults(cpudata);
1707 }
1708 
1709 static int atom_get_min_pstate(int not_used)
1710 {
1711 	u64 value;
1712 
1713 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1714 	return (value >> 8) & 0x7F;
1715 }
1716 
1717 static int atom_get_max_pstate(int not_used)
1718 {
1719 	u64 value;
1720 
1721 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1722 	return (value >> 16) & 0x7F;
1723 }
1724 
1725 static int atom_get_turbo_pstate(int not_used)
1726 {
1727 	u64 value;
1728 
1729 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1730 	return value & 0x7F;
1731 }
1732 
1733 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1734 {
1735 	u64 val;
1736 	int32_t vid_fp;
1737 	u32 vid;
1738 
1739 	val = (u64)pstate << 8;
1740 	if (global.no_turbo && !global.turbo_disabled)
1741 		val |= (u64)1 << 32;
1742 
1743 	vid_fp = cpudata->vid.min + mul_fp(
1744 		int_tofp(pstate - cpudata->pstate.min_pstate),
1745 		cpudata->vid.ratio);
1746 
1747 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1748 	vid = ceiling_fp(vid_fp);
1749 
1750 	if (pstate > cpudata->pstate.max_pstate)
1751 		vid = cpudata->vid.turbo;
1752 
1753 	return val | vid;
1754 }
1755 
1756 static int silvermont_get_scaling(void)
1757 {
1758 	u64 value;
1759 	int i;
1760 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1761 	static int silvermont_freq_table[] = {
1762 		83300, 100000, 133300, 116700, 80000};
1763 
1764 	rdmsrl(MSR_FSB_FREQ, value);
1765 	i = value & 0x7;
1766 	WARN_ON(i > 4);
1767 
1768 	return silvermont_freq_table[i];
1769 }
1770 
1771 static int airmont_get_scaling(void)
1772 {
1773 	u64 value;
1774 	int i;
1775 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1776 	static int airmont_freq_table[] = {
1777 		83300, 100000, 133300, 116700, 80000,
1778 		93300, 90000, 88900, 87500};
1779 
1780 	rdmsrl(MSR_FSB_FREQ, value);
1781 	i = value & 0xF;
1782 	WARN_ON(i > 8);
1783 
1784 	return airmont_freq_table[i];
1785 }
1786 
1787 static void atom_get_vid(struct cpudata *cpudata)
1788 {
1789 	u64 value;
1790 
1791 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1792 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1793 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1794 	cpudata->vid.ratio = div_fp(
1795 		cpudata->vid.max - cpudata->vid.min,
1796 		int_tofp(cpudata->pstate.max_pstate -
1797 			cpudata->pstate.min_pstate));
1798 
1799 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1800 	cpudata->vid.turbo = value & 0x7f;
1801 }
1802 
1803 static int core_get_min_pstate(int cpu)
1804 {
1805 	u64 value;
1806 
1807 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1808 	return (value >> 40) & 0xFF;
1809 }
1810 
1811 static int core_get_max_pstate_physical(int cpu)
1812 {
1813 	u64 value;
1814 
1815 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1816 	return (value >> 8) & 0xFF;
1817 }
1818 
1819 static int core_get_tdp_ratio(int cpu, u64 plat_info)
1820 {
1821 	/* Check how many TDP levels present */
1822 	if (plat_info & 0x600000000) {
1823 		u64 tdp_ctrl;
1824 		u64 tdp_ratio;
1825 		int tdp_msr;
1826 		int err;
1827 
1828 		/* Get the TDP level (0, 1, 2) to get ratios */
1829 		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1830 		if (err)
1831 			return err;
1832 
1833 		/* TDP MSR are continuous starting at 0x648 */
1834 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1835 		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
1836 		if (err)
1837 			return err;
1838 
1839 		/* For level 1 and 2, bits[23:16] contain the ratio */
1840 		if (tdp_ctrl & 0x03)
1841 			tdp_ratio >>= 16;
1842 
1843 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1844 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1845 
1846 		return (int)tdp_ratio;
1847 	}
1848 
1849 	return -ENXIO;
1850 }
1851 
1852 static int core_get_max_pstate(int cpu)
1853 {
1854 	u64 tar;
1855 	u64 plat_info;
1856 	int max_pstate;
1857 	int tdp_ratio;
1858 	int err;
1859 
1860 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
1861 	max_pstate = (plat_info >> 8) & 0xFF;
1862 
1863 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
1864 	if (tdp_ratio <= 0)
1865 		return max_pstate;
1866 
1867 	if (hwp_active) {
1868 		/* Turbo activation ratio is not used on HWP platforms */
1869 		return tdp_ratio;
1870 	}
1871 
1872 	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
1873 	if (!err) {
1874 		int tar_levels;
1875 
1876 		/* Do some sanity checking for safety */
1877 		tar_levels = tar & 0xff;
1878 		if (tdp_ratio - 1 == tar_levels) {
1879 			max_pstate = tar_levels;
1880 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1881 		}
1882 	}
1883 
1884 	return max_pstate;
1885 }
1886 
1887 static int core_get_turbo_pstate(int cpu)
1888 {
1889 	u64 value;
1890 	int nont, ret;
1891 
1892 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1893 	nont = core_get_max_pstate(cpu);
1894 	ret = (value) & 255;
1895 	if (ret <= nont)
1896 		ret = nont;
1897 	return ret;
1898 }
1899 
1900 static inline int core_get_scaling(void)
1901 {
1902 	return 100000;
1903 }
1904 
1905 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1906 {
1907 	u64 val;
1908 
1909 	val = (u64)pstate << 8;
1910 	if (global.no_turbo && !global.turbo_disabled)
1911 		val |= (u64)1 << 32;
1912 
1913 	return val;
1914 }
1915 
1916 static int knl_get_aperf_mperf_shift(void)
1917 {
1918 	return 10;
1919 }
1920 
1921 static int knl_get_turbo_pstate(int cpu)
1922 {
1923 	u64 value;
1924 	int nont, ret;
1925 
1926 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1927 	nont = core_get_max_pstate(cpu);
1928 	ret = (((value) >> 8) & 0xFF);
1929 	if (ret <= nont)
1930 		ret = nont;
1931 	return ret;
1932 }
1933 
1934 static void hybrid_get_type(void *data)
1935 {
1936 	u8 *cpu_type = data;
1937 
1938 	*cpu_type = get_this_hybrid_cpu_type();
1939 }
1940 
1941 static int hybrid_get_cpu_scaling(int cpu)
1942 {
1943 	u8 cpu_type = 0;
1944 
1945 	smp_call_function_single(cpu, hybrid_get_type, &cpu_type, 1);
1946 	/* P-cores have a smaller perf level-to-freqency scaling factor. */
1947 	if (cpu_type == 0x40)
1948 		return 78741;
1949 
1950 	return core_get_scaling();
1951 }
1952 
1953 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1954 {
1955 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1956 	cpu->pstate.current_pstate = pstate;
1957 	/*
1958 	 * Generally, there is no guarantee that this code will always run on
1959 	 * the CPU being updated, so force the register update to run on the
1960 	 * right CPU.
1961 	 */
1962 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1963 		      pstate_funcs.get_val(cpu, pstate));
1964 }
1965 
1966 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1967 {
1968 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1969 }
1970 
1971 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1972 {
1973 	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1974 
1975 	update_turbo_state();
1976 	intel_pstate_set_pstate(cpu, pstate);
1977 }
1978 
1979 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1980 {
1981 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
1982 	int perf_ctl_scaling = pstate_funcs.get_scaling();
1983 
1984 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
1985 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
1986 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
1987 
1988 	if (hwp_active && !hwp_mode_bdw) {
1989 		__intel_pstate_get_hwp_cap(cpu);
1990 
1991 		if (pstate_funcs.get_cpu_scaling) {
1992 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
1993 			if (cpu->pstate.scaling != perf_ctl_scaling)
1994 				intel_pstate_hybrid_hwp_adjust(cpu);
1995 		} else {
1996 			cpu->pstate.scaling = perf_ctl_scaling;
1997 		}
1998 	} else {
1999 		cpu->pstate.scaling = perf_ctl_scaling;
2000 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2001 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2002 	}
2003 
2004 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2005 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2006 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2007 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2008 	}
2009 
2010 	if (pstate_funcs.get_aperf_mperf_shift)
2011 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2012 
2013 	if (pstate_funcs.get_vid)
2014 		pstate_funcs.get_vid(cpu);
2015 
2016 	intel_pstate_set_min_pstate(cpu);
2017 }
2018 
2019 /*
2020  * Long hold time will keep high perf limits for long time,
2021  * which negatively impacts perf/watt for some workloads,
2022  * like specpower. 3ms is based on experiements on some
2023  * workoads.
2024  */
2025 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2026 
2027 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2028 {
2029 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2030 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2031 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2032 	u32 min_limit = (hwp_req & 0xff);
2033 	u32 boost_level1;
2034 
2035 	/*
2036 	 * Cases to consider (User changes via sysfs or boot time):
2037 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2038 	 *	No boost, return.
2039 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2040 	 *     Should result in one level boost only for P0.
2041 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2042 	 *     Should result in two level boost:
2043 	 *         (min + p1)/2 and P1.
2044 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2045 	 *     Should result in three level boost:
2046 	 *        (min + p1)/2, P1 and P0.
2047 	 */
2048 
2049 	/* If max and min are equal or already at max, nothing to boost */
2050 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2051 		return;
2052 
2053 	if (!cpu->hwp_boost_min)
2054 		cpu->hwp_boost_min = min_limit;
2055 
2056 	/* level at half way mark between min and guranteed */
2057 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2058 
2059 	if (cpu->hwp_boost_min < boost_level1)
2060 		cpu->hwp_boost_min = boost_level1;
2061 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2062 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2063 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2064 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2065 		cpu->hwp_boost_min = max_limit;
2066 	else
2067 		return;
2068 
2069 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2070 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2071 	cpu->last_update = cpu->sample.time;
2072 }
2073 
2074 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2075 {
2076 	if (cpu->hwp_boost_min) {
2077 		bool expired;
2078 
2079 		/* Check if we are idle for hold time to boost down */
2080 		expired = time_after64(cpu->sample.time, cpu->last_update +
2081 				       hwp_boost_hold_time_ns);
2082 		if (expired) {
2083 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2084 			cpu->hwp_boost_min = 0;
2085 		}
2086 	}
2087 	cpu->last_update = cpu->sample.time;
2088 }
2089 
2090 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2091 						      u64 time)
2092 {
2093 	cpu->sample.time = time;
2094 
2095 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2096 		bool do_io = false;
2097 
2098 		cpu->sched_flags = 0;
2099 		/*
2100 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2101 		 * is set all the time, we can't just conclude that there is
2102 		 * some IO bound activity is scheduled on this CPU with just
2103 		 * one occurrence. If we receive at least two in two
2104 		 * consecutive ticks, then we treat as boost candidate.
2105 		 */
2106 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2107 			do_io = true;
2108 
2109 		cpu->last_io_update = time;
2110 
2111 		if (do_io)
2112 			intel_pstate_hwp_boost_up(cpu);
2113 
2114 	} else {
2115 		intel_pstate_hwp_boost_down(cpu);
2116 	}
2117 }
2118 
2119 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2120 						u64 time, unsigned int flags)
2121 {
2122 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2123 
2124 	cpu->sched_flags |= flags;
2125 
2126 	if (smp_processor_id() == cpu->cpu)
2127 		intel_pstate_update_util_hwp_local(cpu, time);
2128 }
2129 
2130 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2131 {
2132 	struct sample *sample = &cpu->sample;
2133 
2134 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2135 }
2136 
2137 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2138 {
2139 	u64 aperf, mperf;
2140 	unsigned long flags;
2141 	u64 tsc;
2142 
2143 	local_irq_save(flags);
2144 	rdmsrl(MSR_IA32_APERF, aperf);
2145 	rdmsrl(MSR_IA32_MPERF, mperf);
2146 	tsc = rdtsc();
2147 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2148 		local_irq_restore(flags);
2149 		return false;
2150 	}
2151 	local_irq_restore(flags);
2152 
2153 	cpu->last_sample_time = cpu->sample.time;
2154 	cpu->sample.time = time;
2155 	cpu->sample.aperf = aperf;
2156 	cpu->sample.mperf = mperf;
2157 	cpu->sample.tsc =  tsc;
2158 	cpu->sample.aperf -= cpu->prev_aperf;
2159 	cpu->sample.mperf -= cpu->prev_mperf;
2160 	cpu->sample.tsc -= cpu->prev_tsc;
2161 
2162 	cpu->prev_aperf = aperf;
2163 	cpu->prev_mperf = mperf;
2164 	cpu->prev_tsc = tsc;
2165 	/*
2166 	 * First time this function is invoked in a given cycle, all of the
2167 	 * previous sample data fields are equal to zero or stale and they must
2168 	 * be populated with meaningful numbers for things to work, so assume
2169 	 * that sample.time will always be reset before setting the utilization
2170 	 * update hook and make the caller skip the sample then.
2171 	 */
2172 	if (cpu->last_sample_time) {
2173 		intel_pstate_calc_avg_perf(cpu);
2174 		return true;
2175 	}
2176 	return false;
2177 }
2178 
2179 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2180 {
2181 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2182 }
2183 
2184 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2185 {
2186 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2187 			  cpu->sample.core_avg_perf);
2188 }
2189 
2190 static inline int32_t get_target_pstate(struct cpudata *cpu)
2191 {
2192 	struct sample *sample = &cpu->sample;
2193 	int32_t busy_frac;
2194 	int target, avg_pstate;
2195 
2196 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2197 			   sample->tsc);
2198 
2199 	if (busy_frac < cpu->iowait_boost)
2200 		busy_frac = cpu->iowait_boost;
2201 
2202 	sample->busy_scaled = busy_frac * 100;
2203 
2204 	target = global.no_turbo || global.turbo_disabled ?
2205 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2206 	target += target >> 2;
2207 	target = mul_fp(target, busy_frac);
2208 	if (target < cpu->pstate.min_pstate)
2209 		target = cpu->pstate.min_pstate;
2210 
2211 	/*
2212 	 * If the average P-state during the previous cycle was higher than the
2213 	 * current target, add 50% of the difference to the target to reduce
2214 	 * possible performance oscillations and offset possible performance
2215 	 * loss related to moving the workload from one CPU to another within
2216 	 * a package/module.
2217 	 */
2218 	avg_pstate = get_avg_pstate(cpu);
2219 	if (avg_pstate > target)
2220 		target += (avg_pstate - target) >> 1;
2221 
2222 	return target;
2223 }
2224 
2225 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2226 {
2227 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2228 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2229 
2230 	return clamp_t(int, pstate, min_pstate, max_pstate);
2231 }
2232 
2233 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2234 {
2235 	if (pstate == cpu->pstate.current_pstate)
2236 		return;
2237 
2238 	cpu->pstate.current_pstate = pstate;
2239 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2240 }
2241 
2242 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2243 {
2244 	int from = cpu->pstate.current_pstate;
2245 	struct sample *sample;
2246 	int target_pstate;
2247 
2248 	update_turbo_state();
2249 
2250 	target_pstate = get_target_pstate(cpu);
2251 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2252 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2253 	intel_pstate_update_pstate(cpu, target_pstate);
2254 
2255 	sample = &cpu->sample;
2256 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2257 		fp_toint(sample->busy_scaled),
2258 		from,
2259 		cpu->pstate.current_pstate,
2260 		sample->mperf,
2261 		sample->aperf,
2262 		sample->tsc,
2263 		get_avg_frequency(cpu),
2264 		fp_toint(cpu->iowait_boost * 100));
2265 }
2266 
2267 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2268 				     unsigned int flags)
2269 {
2270 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2271 	u64 delta_ns;
2272 
2273 	/* Don't allow remote callbacks */
2274 	if (smp_processor_id() != cpu->cpu)
2275 		return;
2276 
2277 	delta_ns = time - cpu->last_update;
2278 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2279 		/* Start over if the CPU may have been idle. */
2280 		if (delta_ns > TICK_NSEC) {
2281 			cpu->iowait_boost = ONE_EIGHTH_FP;
2282 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2283 			cpu->iowait_boost <<= 1;
2284 			if (cpu->iowait_boost > int_tofp(1))
2285 				cpu->iowait_boost = int_tofp(1);
2286 		} else {
2287 			cpu->iowait_boost = ONE_EIGHTH_FP;
2288 		}
2289 	} else if (cpu->iowait_boost) {
2290 		/* Clear iowait_boost if the CPU may have been idle. */
2291 		if (delta_ns > TICK_NSEC)
2292 			cpu->iowait_boost = 0;
2293 		else
2294 			cpu->iowait_boost >>= 1;
2295 	}
2296 	cpu->last_update = time;
2297 	delta_ns = time - cpu->sample.time;
2298 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2299 		return;
2300 
2301 	if (intel_pstate_sample(cpu, time))
2302 		intel_pstate_adjust_pstate(cpu);
2303 }
2304 
2305 static struct pstate_funcs core_funcs = {
2306 	.get_max = core_get_max_pstate,
2307 	.get_max_physical = core_get_max_pstate_physical,
2308 	.get_min = core_get_min_pstate,
2309 	.get_turbo = core_get_turbo_pstate,
2310 	.get_scaling = core_get_scaling,
2311 	.get_val = core_get_val,
2312 };
2313 
2314 static const struct pstate_funcs silvermont_funcs = {
2315 	.get_max = atom_get_max_pstate,
2316 	.get_max_physical = atom_get_max_pstate,
2317 	.get_min = atom_get_min_pstate,
2318 	.get_turbo = atom_get_turbo_pstate,
2319 	.get_val = atom_get_val,
2320 	.get_scaling = silvermont_get_scaling,
2321 	.get_vid = atom_get_vid,
2322 };
2323 
2324 static const struct pstate_funcs airmont_funcs = {
2325 	.get_max = atom_get_max_pstate,
2326 	.get_max_physical = atom_get_max_pstate,
2327 	.get_min = atom_get_min_pstate,
2328 	.get_turbo = atom_get_turbo_pstate,
2329 	.get_val = atom_get_val,
2330 	.get_scaling = airmont_get_scaling,
2331 	.get_vid = atom_get_vid,
2332 };
2333 
2334 static const struct pstate_funcs knl_funcs = {
2335 	.get_max = core_get_max_pstate,
2336 	.get_max_physical = core_get_max_pstate_physical,
2337 	.get_min = core_get_min_pstate,
2338 	.get_turbo = knl_get_turbo_pstate,
2339 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2340 	.get_scaling = core_get_scaling,
2341 	.get_val = core_get_val,
2342 };
2343 
2344 #define X86_MATCH(model, policy)					 \
2345 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2346 					   X86_FEATURE_APERFMPERF, &policy)
2347 
2348 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2349 	X86_MATCH(SANDYBRIDGE,		core_funcs),
2350 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2351 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2352 	X86_MATCH(IVYBRIDGE,		core_funcs),
2353 	X86_MATCH(HASWELL,		core_funcs),
2354 	X86_MATCH(BROADWELL,		core_funcs),
2355 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2356 	X86_MATCH(HASWELL_X,		core_funcs),
2357 	X86_MATCH(HASWELL_L,		core_funcs),
2358 	X86_MATCH(HASWELL_G,		core_funcs),
2359 	X86_MATCH(BROADWELL_G,		core_funcs),
2360 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2361 	X86_MATCH(SKYLAKE_L,		core_funcs),
2362 	X86_MATCH(BROADWELL_X,		core_funcs),
2363 	X86_MATCH(SKYLAKE,		core_funcs),
2364 	X86_MATCH(BROADWELL_D,		core_funcs),
2365 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2366 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2367 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2368 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2369 	X86_MATCH(SKYLAKE_X,		core_funcs),
2370 	X86_MATCH(COMETLAKE,		core_funcs),
2371 	X86_MATCH(ICELAKE_X,		core_funcs),
2372 	X86_MATCH(TIGERLAKE,		core_funcs),
2373 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2374 	{}
2375 };
2376 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2377 
2378 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2379 	X86_MATCH(BROADWELL_D,		core_funcs),
2380 	X86_MATCH(BROADWELL_X,		core_funcs),
2381 	X86_MATCH(SKYLAKE_X,		core_funcs),
2382 	X86_MATCH(ICELAKE_X,		core_funcs),
2383 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2384 	{}
2385 };
2386 
2387 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2388 	X86_MATCH(KABYLAKE,		core_funcs),
2389 	{}
2390 };
2391 
2392 static int intel_pstate_init_cpu(unsigned int cpunum)
2393 {
2394 	struct cpudata *cpu;
2395 
2396 	cpu = all_cpu_data[cpunum];
2397 
2398 	if (!cpu) {
2399 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2400 		if (!cpu)
2401 			return -ENOMEM;
2402 
2403 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2404 
2405 		cpu->cpu = cpunum;
2406 
2407 		cpu->epp_default = -EINVAL;
2408 
2409 		if (hwp_active) {
2410 			intel_pstate_hwp_enable(cpu);
2411 
2412 			if (intel_pstate_acpi_pm_profile_server())
2413 				hwp_boost = true;
2414 		}
2415 	} else if (hwp_active) {
2416 		/*
2417 		 * Re-enable HWP in case this happens after a resume from ACPI
2418 		 * S3 if the CPU was offline during the whole system/resume
2419 		 * cycle.
2420 		 */
2421 		intel_pstate_hwp_reenable(cpu);
2422 	}
2423 
2424 	cpu->epp_powersave = -EINVAL;
2425 	cpu->epp_policy = 0;
2426 
2427 	intel_pstate_get_cpu_pstates(cpu);
2428 
2429 	pr_debug("controlling: cpu %d\n", cpunum);
2430 
2431 	return 0;
2432 }
2433 
2434 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2435 {
2436 	struct cpudata *cpu = all_cpu_data[cpu_num];
2437 
2438 	if (hwp_active && !hwp_boost)
2439 		return;
2440 
2441 	if (cpu->update_util_set)
2442 		return;
2443 
2444 	/* Prevent intel_pstate_update_util() from using stale data. */
2445 	cpu->sample.time = 0;
2446 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2447 				     (hwp_active ?
2448 				      intel_pstate_update_util_hwp :
2449 				      intel_pstate_update_util));
2450 	cpu->update_util_set = true;
2451 }
2452 
2453 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2454 {
2455 	struct cpudata *cpu_data = all_cpu_data[cpu];
2456 
2457 	if (!cpu_data->update_util_set)
2458 		return;
2459 
2460 	cpufreq_remove_update_util_hook(cpu);
2461 	cpu_data->update_util_set = false;
2462 	synchronize_rcu();
2463 }
2464 
2465 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2466 {
2467 	return global.turbo_disabled || global.no_turbo ?
2468 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2469 }
2470 
2471 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2472 					    unsigned int policy_min,
2473 					    unsigned int policy_max)
2474 {
2475 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2476 	int32_t max_policy_perf, min_policy_perf;
2477 
2478 	max_policy_perf = policy_max / perf_ctl_scaling;
2479 	if (policy_max == policy_min) {
2480 		min_policy_perf = max_policy_perf;
2481 	} else {
2482 		min_policy_perf = policy_min / perf_ctl_scaling;
2483 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2484 					  0, max_policy_perf);
2485 	}
2486 
2487 	/*
2488 	 * HWP needs some special consideration, because HWP_REQUEST uses
2489 	 * abstract values to represent performance rather than pure ratios.
2490 	 */
2491 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2492 		int scaling = cpu->pstate.scaling;
2493 		int freq;
2494 
2495 		freq = max_policy_perf * perf_ctl_scaling;
2496 		max_policy_perf = DIV_ROUND_UP(freq, scaling);
2497 		freq = min_policy_perf * perf_ctl_scaling;
2498 		min_policy_perf = DIV_ROUND_UP(freq, scaling);
2499 	}
2500 
2501 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2502 		 cpu->cpu, min_policy_perf, max_policy_perf);
2503 
2504 	/* Normalize user input to [min_perf, max_perf] */
2505 	if (per_cpu_limits) {
2506 		cpu->min_perf_ratio = min_policy_perf;
2507 		cpu->max_perf_ratio = max_policy_perf;
2508 	} else {
2509 		int turbo_max = cpu->pstate.turbo_pstate;
2510 		int32_t global_min, global_max;
2511 
2512 		/* Global limits are in percent of the maximum turbo P-state. */
2513 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2514 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2515 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2516 
2517 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2518 			 global_min, global_max);
2519 
2520 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2521 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2522 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2523 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2524 
2525 		/* Make sure min_perf <= max_perf */
2526 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2527 					  cpu->max_perf_ratio);
2528 
2529 	}
2530 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2531 		 cpu->max_perf_ratio,
2532 		 cpu->min_perf_ratio);
2533 }
2534 
2535 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2536 {
2537 	struct cpudata *cpu;
2538 
2539 	if (!policy->cpuinfo.max_freq)
2540 		return -ENODEV;
2541 
2542 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2543 		 policy->cpuinfo.max_freq, policy->max);
2544 
2545 	cpu = all_cpu_data[policy->cpu];
2546 	cpu->policy = policy->policy;
2547 
2548 	mutex_lock(&intel_pstate_limits_lock);
2549 
2550 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2551 
2552 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2553 		/*
2554 		 * NOHZ_FULL CPUs need this as the governor callback may not
2555 		 * be invoked on them.
2556 		 */
2557 		intel_pstate_clear_update_util_hook(policy->cpu);
2558 		intel_pstate_max_within_limits(cpu);
2559 	} else {
2560 		intel_pstate_set_update_util_hook(policy->cpu);
2561 	}
2562 
2563 	if (hwp_active) {
2564 		/*
2565 		 * When hwp_boost was active before and dynamically it
2566 		 * was turned off, in that case we need to clear the
2567 		 * update util hook.
2568 		 */
2569 		if (!hwp_boost)
2570 			intel_pstate_clear_update_util_hook(policy->cpu);
2571 		intel_pstate_hwp_set(policy->cpu);
2572 	}
2573 
2574 	mutex_unlock(&intel_pstate_limits_lock);
2575 
2576 	return 0;
2577 }
2578 
2579 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2580 					   struct cpufreq_policy_data *policy)
2581 {
2582 	if (!hwp_active &&
2583 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2584 	    policy->max < policy->cpuinfo.max_freq &&
2585 	    policy->max > cpu->pstate.max_freq) {
2586 		pr_debug("policy->max > max non turbo frequency\n");
2587 		policy->max = policy->cpuinfo.max_freq;
2588 	}
2589 }
2590 
2591 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2592 					   struct cpufreq_policy_data *policy)
2593 {
2594 	int max_freq;
2595 
2596 	update_turbo_state();
2597 	if (hwp_active) {
2598 		intel_pstate_get_hwp_cap(cpu);
2599 		max_freq = global.no_turbo || global.turbo_disabled ?
2600 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2601 	} else {
2602 		max_freq = intel_pstate_get_max_freq(cpu);
2603 	}
2604 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2605 
2606 	intel_pstate_adjust_policy_max(cpu, policy);
2607 }
2608 
2609 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2610 {
2611 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2612 
2613 	return 0;
2614 }
2615 
2616 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2617 {
2618 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2619 
2620 	pr_debug("CPU %d going offline\n", cpu->cpu);
2621 
2622 	if (cpu->suspended)
2623 		return 0;
2624 
2625 	/*
2626 	 * If the CPU is an SMT thread and it goes offline with the performance
2627 	 * settings different from the minimum, it will prevent its sibling
2628 	 * from getting to lower performance levels, so force the minimum
2629 	 * performance on CPU offline to prevent that from happening.
2630 	 */
2631 	if (hwp_active)
2632 		intel_pstate_hwp_offline(cpu);
2633 	else
2634 		intel_pstate_set_min_pstate(cpu);
2635 
2636 	intel_pstate_exit_perf_limits(policy);
2637 
2638 	return 0;
2639 }
2640 
2641 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2642 {
2643 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2644 
2645 	pr_debug("CPU %d going online\n", cpu->cpu);
2646 
2647 	intel_pstate_init_acpi_perf_limits(policy);
2648 
2649 	if (hwp_active) {
2650 		/*
2651 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2652 		 * know that it need not do that.
2653 		 */
2654 		intel_pstate_hwp_reenable(cpu);
2655 		cpu->suspended = false;
2656 	}
2657 
2658 	return 0;
2659 }
2660 
2661 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2662 {
2663 	intel_pstate_clear_update_util_hook(policy->cpu);
2664 
2665 	return intel_cpufreq_cpu_offline(policy);
2666 }
2667 
2668 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2669 {
2670 	pr_debug("CPU %d exiting\n", policy->cpu);
2671 
2672 	policy->fast_switch_possible = false;
2673 
2674 	return 0;
2675 }
2676 
2677 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2678 {
2679 	struct cpudata *cpu;
2680 	int rc;
2681 
2682 	rc = intel_pstate_init_cpu(policy->cpu);
2683 	if (rc)
2684 		return rc;
2685 
2686 	cpu = all_cpu_data[policy->cpu];
2687 
2688 	cpu->max_perf_ratio = 0xFF;
2689 	cpu->min_perf_ratio = 0;
2690 
2691 	/* cpuinfo and default policy values */
2692 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2693 	update_turbo_state();
2694 	global.turbo_disabled_mf = global.turbo_disabled;
2695 	policy->cpuinfo.max_freq = global.turbo_disabled ?
2696 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2697 
2698 	policy->min = policy->cpuinfo.min_freq;
2699 	policy->max = policy->cpuinfo.max_freq;
2700 
2701 	intel_pstate_init_acpi_perf_limits(policy);
2702 
2703 	policy->fast_switch_possible = true;
2704 
2705 	return 0;
2706 }
2707 
2708 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2709 {
2710 	int ret = __intel_pstate_cpu_init(policy);
2711 
2712 	if (ret)
2713 		return ret;
2714 
2715 	/*
2716 	 * Set the policy to powersave to provide a valid fallback value in case
2717 	 * the default cpufreq governor is neither powersave nor performance.
2718 	 */
2719 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2720 
2721 	if (hwp_active) {
2722 		struct cpudata *cpu = all_cpu_data[policy->cpu];
2723 
2724 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2725 	}
2726 
2727 	return 0;
2728 }
2729 
2730 static struct cpufreq_driver intel_pstate = {
2731 	.flags		= CPUFREQ_CONST_LOOPS,
2732 	.verify		= intel_pstate_verify_policy,
2733 	.setpolicy	= intel_pstate_set_policy,
2734 	.suspend	= intel_pstate_suspend,
2735 	.resume		= intel_pstate_resume,
2736 	.init		= intel_pstate_cpu_init,
2737 	.exit		= intel_pstate_cpu_exit,
2738 	.offline	= intel_pstate_cpu_offline,
2739 	.online		= intel_pstate_cpu_online,
2740 	.update_limits	= intel_pstate_update_limits,
2741 	.name		= "intel_pstate",
2742 };
2743 
2744 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2745 {
2746 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2747 
2748 	intel_pstate_verify_cpu_policy(cpu, policy);
2749 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2750 
2751 	return 0;
2752 }
2753 
2754 /* Use of trace in passive mode:
2755  *
2756  * In passive mode the trace core_busy field (also known as the
2757  * performance field, and lablelled as such on the graphs; also known as
2758  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2759  * driver call was via the normal or fast switch path. Various graphs
2760  * output from the intel_pstate_tracer.py utility that include core_busy
2761  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2762  * so we use 10 to indicate the normal path through the driver, and
2763  * 90 to indicate the fast switch path through the driver.
2764  * The scaled_busy field is not used, and is set to 0.
2765  */
2766 
2767 #define	INTEL_PSTATE_TRACE_TARGET 10
2768 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2769 
2770 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2771 {
2772 	struct sample *sample;
2773 
2774 	if (!trace_pstate_sample_enabled())
2775 		return;
2776 
2777 	if (!intel_pstate_sample(cpu, ktime_get()))
2778 		return;
2779 
2780 	sample = &cpu->sample;
2781 	trace_pstate_sample(trace_type,
2782 		0,
2783 		old_pstate,
2784 		cpu->pstate.current_pstate,
2785 		sample->mperf,
2786 		sample->aperf,
2787 		sample->tsc,
2788 		get_avg_frequency(cpu),
2789 		fp_toint(cpu->iowait_boost * 100));
2790 }
2791 
2792 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2793 				     u32 desired, bool fast_switch)
2794 {
2795 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2796 
2797 	value &= ~HWP_MIN_PERF(~0L);
2798 	value |= HWP_MIN_PERF(min);
2799 
2800 	value &= ~HWP_MAX_PERF(~0L);
2801 	value |= HWP_MAX_PERF(max);
2802 
2803 	value &= ~HWP_DESIRED_PERF(~0L);
2804 	value |= HWP_DESIRED_PERF(desired);
2805 
2806 	if (value == prev)
2807 		return;
2808 
2809 	WRITE_ONCE(cpu->hwp_req_cached, value);
2810 	if (fast_switch)
2811 		wrmsrl(MSR_HWP_REQUEST, value);
2812 	else
2813 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2814 }
2815 
2816 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2817 					  u32 target_pstate, bool fast_switch)
2818 {
2819 	if (fast_switch)
2820 		wrmsrl(MSR_IA32_PERF_CTL,
2821 		       pstate_funcs.get_val(cpu, target_pstate));
2822 	else
2823 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2824 			      pstate_funcs.get_val(cpu, target_pstate));
2825 }
2826 
2827 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2828 				       int target_pstate, bool fast_switch)
2829 {
2830 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2831 	int old_pstate = cpu->pstate.current_pstate;
2832 
2833 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2834 	if (hwp_active) {
2835 		int max_pstate = policy->strict_target ?
2836 					target_pstate : cpu->max_perf_ratio;
2837 
2838 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2839 					 fast_switch);
2840 	} else if (target_pstate != old_pstate) {
2841 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2842 	}
2843 
2844 	cpu->pstate.current_pstate = target_pstate;
2845 
2846 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2847 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2848 
2849 	return target_pstate;
2850 }
2851 
2852 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2853 				unsigned int target_freq,
2854 				unsigned int relation)
2855 {
2856 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2857 	struct cpufreq_freqs freqs;
2858 	int target_pstate;
2859 
2860 	update_turbo_state();
2861 
2862 	freqs.old = policy->cur;
2863 	freqs.new = target_freq;
2864 
2865 	cpufreq_freq_transition_begin(policy, &freqs);
2866 
2867 	switch (relation) {
2868 	case CPUFREQ_RELATION_L:
2869 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2870 		break;
2871 	case CPUFREQ_RELATION_H:
2872 		target_pstate = freqs.new / cpu->pstate.scaling;
2873 		break;
2874 	default:
2875 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2876 		break;
2877 	}
2878 
2879 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2880 
2881 	freqs.new = target_pstate * cpu->pstate.scaling;
2882 
2883 	cpufreq_freq_transition_end(policy, &freqs, false);
2884 
2885 	return 0;
2886 }
2887 
2888 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2889 					      unsigned int target_freq)
2890 {
2891 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2892 	int target_pstate;
2893 
2894 	update_turbo_state();
2895 
2896 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2897 
2898 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2899 
2900 	return target_pstate * cpu->pstate.scaling;
2901 }
2902 
2903 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2904 				      unsigned long min_perf,
2905 				      unsigned long target_perf,
2906 				      unsigned long capacity)
2907 {
2908 	struct cpudata *cpu = all_cpu_data[cpunum];
2909 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2910 	int old_pstate = cpu->pstate.current_pstate;
2911 	int cap_pstate, min_pstate, max_pstate, target_pstate;
2912 
2913 	update_turbo_state();
2914 	cap_pstate = global.turbo_disabled ? HWP_GUARANTEED_PERF(hwp_cap) :
2915 					     HWP_HIGHEST_PERF(hwp_cap);
2916 
2917 	/* Optimization: Avoid unnecessary divisions. */
2918 
2919 	target_pstate = cap_pstate;
2920 	if (target_perf < capacity)
2921 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2922 
2923 	min_pstate = cap_pstate;
2924 	if (min_perf < capacity)
2925 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2926 
2927 	if (min_pstate < cpu->pstate.min_pstate)
2928 		min_pstate = cpu->pstate.min_pstate;
2929 
2930 	if (min_pstate < cpu->min_perf_ratio)
2931 		min_pstate = cpu->min_perf_ratio;
2932 
2933 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2934 	if (max_pstate < min_pstate)
2935 		max_pstate = min_pstate;
2936 
2937 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2938 
2939 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
2940 
2941 	cpu->pstate.current_pstate = target_pstate;
2942 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2943 }
2944 
2945 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2946 {
2947 	struct freq_qos_request *req;
2948 	struct cpudata *cpu;
2949 	struct device *dev;
2950 	int ret, freq;
2951 
2952 	dev = get_cpu_device(policy->cpu);
2953 	if (!dev)
2954 		return -ENODEV;
2955 
2956 	ret = __intel_pstate_cpu_init(policy);
2957 	if (ret)
2958 		return ret;
2959 
2960 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2961 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2962 	policy->cur = policy->cpuinfo.min_freq;
2963 
2964 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
2965 	if (!req) {
2966 		ret = -ENOMEM;
2967 		goto pstate_exit;
2968 	}
2969 
2970 	cpu = all_cpu_data[policy->cpu];
2971 
2972 	if (hwp_active) {
2973 		u64 value;
2974 
2975 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
2976 
2977 		intel_pstate_get_hwp_cap(cpu);
2978 
2979 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
2980 		WRITE_ONCE(cpu->hwp_req_cached, value);
2981 
2982 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
2983 	} else {
2984 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2985 	}
2986 
2987 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
2988 
2989 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
2990 				   freq);
2991 	if (ret < 0) {
2992 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
2993 		goto free_req;
2994 	}
2995 
2996 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
2997 
2998 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
2999 				   freq);
3000 	if (ret < 0) {
3001 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3002 		goto remove_min_req;
3003 	}
3004 
3005 	policy->driver_data = req;
3006 
3007 	return 0;
3008 
3009 remove_min_req:
3010 	freq_qos_remove_request(req);
3011 free_req:
3012 	kfree(req);
3013 pstate_exit:
3014 	intel_pstate_exit_perf_limits(policy);
3015 
3016 	return ret;
3017 }
3018 
3019 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3020 {
3021 	struct freq_qos_request *req;
3022 
3023 	req = policy->driver_data;
3024 
3025 	freq_qos_remove_request(req + 1);
3026 	freq_qos_remove_request(req);
3027 	kfree(req);
3028 
3029 	return intel_pstate_cpu_exit(policy);
3030 }
3031 
3032 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3033 {
3034 	intel_pstate_suspend(policy);
3035 
3036 	if (hwp_active) {
3037 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3038 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3039 
3040 		/*
3041 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3042 		 * intel_cpufreq_adjust_perf() is in use and the last value
3043 		 * written by it may not be suitable.
3044 		 */
3045 		value &= ~HWP_DESIRED_PERF(~0L);
3046 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3047 		WRITE_ONCE(cpu->hwp_req_cached, value);
3048 	}
3049 
3050 	return 0;
3051 }
3052 
3053 static struct cpufreq_driver intel_cpufreq = {
3054 	.flags		= CPUFREQ_CONST_LOOPS,
3055 	.verify		= intel_cpufreq_verify_policy,
3056 	.target		= intel_cpufreq_target,
3057 	.fast_switch	= intel_cpufreq_fast_switch,
3058 	.init		= intel_cpufreq_cpu_init,
3059 	.exit		= intel_cpufreq_cpu_exit,
3060 	.offline	= intel_cpufreq_cpu_offline,
3061 	.online		= intel_pstate_cpu_online,
3062 	.suspend	= intel_cpufreq_suspend,
3063 	.resume		= intel_pstate_resume,
3064 	.update_limits	= intel_pstate_update_limits,
3065 	.name		= "intel_cpufreq",
3066 };
3067 
3068 static struct cpufreq_driver *default_driver;
3069 
3070 static void intel_pstate_driver_cleanup(void)
3071 {
3072 	unsigned int cpu;
3073 
3074 	cpus_read_lock();
3075 	for_each_online_cpu(cpu) {
3076 		if (all_cpu_data[cpu]) {
3077 			if (intel_pstate_driver == &intel_pstate)
3078 				intel_pstate_clear_update_util_hook(cpu);
3079 
3080 			spin_lock(&hwp_notify_lock);
3081 			kfree(all_cpu_data[cpu]);
3082 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3083 			spin_unlock(&hwp_notify_lock);
3084 		}
3085 	}
3086 	cpus_read_unlock();
3087 
3088 	intel_pstate_driver = NULL;
3089 }
3090 
3091 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3092 {
3093 	int ret;
3094 
3095 	if (driver == &intel_pstate)
3096 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3097 
3098 	memset(&global, 0, sizeof(global));
3099 	global.max_perf_pct = 100;
3100 
3101 	intel_pstate_driver = driver;
3102 	ret = cpufreq_register_driver(intel_pstate_driver);
3103 	if (ret) {
3104 		intel_pstate_driver_cleanup();
3105 		return ret;
3106 	}
3107 
3108 	global.min_perf_pct = min_perf_pct_min();
3109 
3110 	return 0;
3111 }
3112 
3113 static ssize_t intel_pstate_show_status(char *buf)
3114 {
3115 	if (!intel_pstate_driver)
3116 		return sprintf(buf, "off\n");
3117 
3118 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3119 					"active" : "passive");
3120 }
3121 
3122 static int intel_pstate_update_status(const char *buf, size_t size)
3123 {
3124 	if (size == 3 && !strncmp(buf, "off", size)) {
3125 		if (!intel_pstate_driver)
3126 			return -EINVAL;
3127 
3128 		if (hwp_active)
3129 			return -EBUSY;
3130 
3131 		cpufreq_unregister_driver(intel_pstate_driver);
3132 		intel_pstate_driver_cleanup();
3133 		return 0;
3134 	}
3135 
3136 	if (size == 6 && !strncmp(buf, "active", size)) {
3137 		if (intel_pstate_driver) {
3138 			if (intel_pstate_driver == &intel_pstate)
3139 				return 0;
3140 
3141 			cpufreq_unregister_driver(intel_pstate_driver);
3142 		}
3143 
3144 		return intel_pstate_register_driver(&intel_pstate);
3145 	}
3146 
3147 	if (size == 7 && !strncmp(buf, "passive", size)) {
3148 		if (intel_pstate_driver) {
3149 			if (intel_pstate_driver == &intel_cpufreq)
3150 				return 0;
3151 
3152 			cpufreq_unregister_driver(intel_pstate_driver);
3153 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3154 		}
3155 
3156 		return intel_pstate_register_driver(&intel_cpufreq);
3157 	}
3158 
3159 	return -EINVAL;
3160 }
3161 
3162 static int no_load __initdata;
3163 static int no_hwp __initdata;
3164 static int hwp_only __initdata;
3165 static unsigned int force_load __initdata;
3166 
3167 static int __init intel_pstate_msrs_not_valid(void)
3168 {
3169 	if (!pstate_funcs.get_max(0) ||
3170 	    !pstate_funcs.get_min(0) ||
3171 	    !pstate_funcs.get_turbo(0))
3172 		return -ENODEV;
3173 
3174 	return 0;
3175 }
3176 
3177 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3178 {
3179 	pstate_funcs.get_max   = funcs->get_max;
3180 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3181 	pstate_funcs.get_min   = funcs->get_min;
3182 	pstate_funcs.get_turbo = funcs->get_turbo;
3183 	pstate_funcs.get_scaling = funcs->get_scaling;
3184 	pstate_funcs.get_val   = funcs->get_val;
3185 	pstate_funcs.get_vid   = funcs->get_vid;
3186 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3187 }
3188 
3189 #ifdef CONFIG_ACPI
3190 
3191 static bool __init intel_pstate_no_acpi_pss(void)
3192 {
3193 	int i;
3194 
3195 	for_each_possible_cpu(i) {
3196 		acpi_status status;
3197 		union acpi_object *pss;
3198 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3199 		struct acpi_processor *pr = per_cpu(processors, i);
3200 
3201 		if (!pr)
3202 			continue;
3203 
3204 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3205 		if (ACPI_FAILURE(status))
3206 			continue;
3207 
3208 		pss = buffer.pointer;
3209 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3210 			kfree(pss);
3211 			return false;
3212 		}
3213 
3214 		kfree(pss);
3215 	}
3216 
3217 	pr_debug("ACPI _PSS not found\n");
3218 	return true;
3219 }
3220 
3221 static bool __init intel_pstate_no_acpi_pcch(void)
3222 {
3223 	acpi_status status;
3224 	acpi_handle handle;
3225 
3226 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3227 	if (ACPI_FAILURE(status))
3228 		goto not_found;
3229 
3230 	if (acpi_has_method(handle, "PCCH"))
3231 		return false;
3232 
3233 not_found:
3234 	pr_debug("ACPI PCCH not found\n");
3235 	return true;
3236 }
3237 
3238 static bool __init intel_pstate_has_acpi_ppc(void)
3239 {
3240 	int i;
3241 
3242 	for_each_possible_cpu(i) {
3243 		struct acpi_processor *pr = per_cpu(processors, i);
3244 
3245 		if (!pr)
3246 			continue;
3247 		if (acpi_has_method(pr->handle, "_PPC"))
3248 			return true;
3249 	}
3250 	pr_debug("ACPI _PPC not found\n");
3251 	return false;
3252 }
3253 
3254 enum {
3255 	PSS,
3256 	PPC,
3257 };
3258 
3259 /* Hardware vendor-specific info that has its own power management modes */
3260 static struct acpi_platform_list plat_info[] __initdata = {
3261 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3262 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3263 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3264 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3265 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3266 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3267 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3268 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3269 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3270 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3271 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3272 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3273 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3274 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3275 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3276 	{ } /* End */
3277 };
3278 
3279 #define BITMASK_OOB	(BIT(8) | BIT(18))
3280 
3281 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3282 {
3283 	const struct x86_cpu_id *id;
3284 	u64 misc_pwr;
3285 	int idx;
3286 
3287 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3288 	if (id) {
3289 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3290 		if (misc_pwr & BITMASK_OOB) {
3291 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3292 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3293 			return true;
3294 		}
3295 	}
3296 
3297 	idx = acpi_match_platform_list(plat_info);
3298 	if (idx < 0)
3299 		return false;
3300 
3301 	switch (plat_info[idx].data) {
3302 	case PSS:
3303 		if (!intel_pstate_no_acpi_pss())
3304 			return false;
3305 
3306 		return intel_pstate_no_acpi_pcch();
3307 	case PPC:
3308 		return intel_pstate_has_acpi_ppc() && !force_load;
3309 	}
3310 
3311 	return false;
3312 }
3313 
3314 static void intel_pstate_request_control_from_smm(void)
3315 {
3316 	/*
3317 	 * It may be unsafe to request P-states control from SMM if _PPC support
3318 	 * has not been enabled.
3319 	 */
3320 	if (acpi_ppc)
3321 		acpi_processor_pstate_control();
3322 }
3323 #else /* CONFIG_ACPI not enabled */
3324 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3325 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3326 static inline void intel_pstate_request_control_from_smm(void) {}
3327 #endif /* CONFIG_ACPI */
3328 
3329 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3330 
3331 #define X86_MATCH_HWP(model, hwp_mode)					\
3332 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3333 					   X86_FEATURE_HWP, hwp_mode)
3334 
3335 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3336 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3337 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3338 	X86_MATCH_HWP(ANY,		0),
3339 	{}
3340 };
3341 
3342 static bool intel_pstate_hwp_is_enabled(void)
3343 {
3344 	u64 value;
3345 
3346 	rdmsrl(MSR_PM_ENABLE, value);
3347 	return !!(value & 0x1);
3348 }
3349 
3350 static const struct x86_cpu_id intel_epp_balance_perf[] = {
3351 	/*
3352 	 * Set EPP value as 102, this is the max suggested EPP
3353 	 * which can result in one core turbo frequency for
3354 	 * AlderLake Mobile CPUs.
3355 	 */
3356 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L, 102),
3357 	X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, 32),
3358 	{}
3359 };
3360 
3361 static int __init intel_pstate_init(void)
3362 {
3363 	static struct cpudata **_all_cpu_data;
3364 	const struct x86_cpu_id *id;
3365 	int rc;
3366 
3367 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3368 		return -ENODEV;
3369 
3370 	id = x86_match_cpu(hwp_support_ids);
3371 	if (id) {
3372 		hwp_forced = intel_pstate_hwp_is_enabled();
3373 
3374 		if (hwp_forced)
3375 			pr_info("HWP enabled by BIOS\n");
3376 		else if (no_load)
3377 			return -ENODEV;
3378 
3379 		copy_cpu_funcs(&core_funcs);
3380 		/*
3381 		 * Avoid enabling HWP for processors without EPP support,
3382 		 * because that means incomplete HWP implementation which is a
3383 		 * corner case and supporting it is generally problematic.
3384 		 *
3385 		 * If HWP is enabled already, though, there is no choice but to
3386 		 * deal with it.
3387 		 */
3388 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3389 			WRITE_ONCE(hwp_active, 1);
3390 			hwp_mode_bdw = id->driver_data;
3391 			intel_pstate.attr = hwp_cpufreq_attrs;
3392 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3393 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3394 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3395 			if (!default_driver)
3396 				default_driver = &intel_pstate;
3397 
3398 			if (boot_cpu_has(X86_FEATURE_HYBRID_CPU))
3399 				pstate_funcs.get_cpu_scaling = hybrid_get_cpu_scaling;
3400 
3401 			goto hwp_cpu_matched;
3402 		}
3403 		pr_info("HWP not enabled\n");
3404 	} else {
3405 		if (no_load)
3406 			return -ENODEV;
3407 
3408 		id = x86_match_cpu(intel_pstate_cpu_ids);
3409 		if (!id) {
3410 			pr_info("CPU model not supported\n");
3411 			return -ENODEV;
3412 		}
3413 
3414 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3415 	}
3416 
3417 	if (intel_pstate_msrs_not_valid()) {
3418 		pr_info("Invalid MSRs\n");
3419 		return -ENODEV;
3420 	}
3421 	/* Without HWP start in the passive mode. */
3422 	if (!default_driver)
3423 		default_driver = &intel_cpufreq;
3424 
3425 hwp_cpu_matched:
3426 	/*
3427 	 * The Intel pstate driver will be ignored if the platform
3428 	 * firmware has its own power management modes.
3429 	 */
3430 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3431 		pr_info("P-states controlled by the platform\n");
3432 		return -ENODEV;
3433 	}
3434 
3435 	if (!hwp_active && hwp_only)
3436 		return -ENOTSUPP;
3437 
3438 	pr_info("Intel P-state driver initializing\n");
3439 
3440 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3441 	if (!_all_cpu_data)
3442 		return -ENOMEM;
3443 
3444 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3445 
3446 	intel_pstate_request_control_from_smm();
3447 
3448 	intel_pstate_sysfs_expose_params();
3449 
3450 	if (hwp_active) {
3451 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_balance_perf);
3452 
3453 		if (id)
3454 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = id->driver_data;
3455 	}
3456 
3457 	mutex_lock(&intel_pstate_driver_lock);
3458 	rc = intel_pstate_register_driver(default_driver);
3459 	mutex_unlock(&intel_pstate_driver_lock);
3460 	if (rc) {
3461 		intel_pstate_sysfs_remove();
3462 		return rc;
3463 	}
3464 
3465 	if (hwp_active) {
3466 		const struct x86_cpu_id *id;
3467 
3468 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3469 		if (id) {
3470 			set_power_ctl_ee_state(false);
3471 			pr_info("Disabling energy efficiency optimization\n");
3472 		}
3473 
3474 		pr_info("HWP enabled\n");
3475 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3476 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3477 	}
3478 
3479 	return 0;
3480 }
3481 device_initcall(intel_pstate_init);
3482 
3483 static int __init intel_pstate_setup(char *str)
3484 {
3485 	if (!str)
3486 		return -EINVAL;
3487 
3488 	if (!strcmp(str, "disable"))
3489 		no_load = 1;
3490 	else if (!strcmp(str, "active"))
3491 		default_driver = &intel_pstate;
3492 	else if (!strcmp(str, "passive"))
3493 		default_driver = &intel_cpufreq;
3494 
3495 	if (!strcmp(str, "no_hwp"))
3496 		no_hwp = 1;
3497 
3498 	if (!strcmp(str, "force"))
3499 		force_load = 1;
3500 	if (!strcmp(str, "hwp_only"))
3501 		hwp_only = 1;
3502 	if (!strcmp(str, "per_cpu_perf_limits"))
3503 		per_cpu_limits = true;
3504 
3505 #ifdef CONFIG_ACPI
3506 	if (!strcmp(str, "support_acpi_ppc"))
3507 		acpi_ppc = true;
3508 #endif
3509 
3510 	return 0;
3511 }
3512 early_param("intel_pstate", intel_pstate_setup);
3513 
3514 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3515 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3516