cpufreq: intel_pstate: Remove max/min fractions to limit performance
[muen/linux.git] / drivers / cpufreq / intel_pstate.c
1 /*
2  * intel_pstate.c: Native P state management for Intel processors
3  *
4  * (C) Copyright 2012 Intel Corporation
5  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public License
9  * as published by the Free Software Foundation; version 2
10  * of the License.
11  */
12
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
15 #include <linux/kernel.h>
16 #include <linux/kernel_stat.h>
17 #include <linux/module.h>
18 #include <linux/ktime.h>
19 #include <linux/hrtimer.h>
20 #include <linux/tick.h>
21 #include <linux/slab.h>
22 #include <linux/sched/cpufreq.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
25 #include <linux/cpufreq.h>
26 #include <linux/sysfs.h>
27 #include <linux/types.h>
28 #include <linux/fs.h>
29 #include <linux/debugfs.h>
30 #include <linux/acpi.h>
31 #include <linux/vmalloc.h>
32 #include <trace/events/power.h>
33
34 #include <asm/div64.h>
35 #include <asm/msr.h>
36 #include <asm/cpu_device_id.h>
37 #include <asm/cpufeature.h>
38 #include <asm/intel-family.h>
39
40 #define INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL  (10 * NSEC_PER_MSEC)
41 #define INTEL_PSTATE_HWP_SAMPLING_INTERVAL      (50 * NSEC_PER_MSEC)
42
43 #define INTEL_CPUFREQ_TRANSITION_LATENCY        20000
44 #define INTEL_CPUFREQ_TRANSITION_DELAY          500
45
46 #ifdef CONFIG_ACPI
47 #include <acpi/processor.h>
48 #include <acpi/cppc_acpi.h>
49 #endif
50
51 #define FRAC_BITS 8
52 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
53 #define fp_toint(X) ((X) >> FRAC_BITS)
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 int32_t percent_fp(int percent)
82 {
83         return div_fp(percent, 100);
84 }
85
86 static inline u64 mul_ext_fp(u64 x, u64 y)
87 {
88         return (x * y) >> EXT_FRAC_BITS;
89 }
90
91 static inline u64 div_ext_fp(u64 x, u64 y)
92 {
93         return div64_u64(x << EXT_FRAC_BITS, y);
94 }
95
96 static inline int32_t percent_ext_fp(int percent)
97 {
98         return div_ext_fp(percent, 100);
99 }
100
101 /**
102  * struct sample -      Store performance sample
103  * @core_avg_perf:      Ratio of APERF/MPERF which is the actual average
104  *                      performance during last sample period
105  * @busy_scaled:        Scaled busy value which is used to calculate next
106  *                      P state. This can be different than core_avg_perf
107  *                      to account for cpu idle period
108  * @aperf:              Difference of actual performance frequency clock count
109  *                      read from APERF MSR between last and current sample
110  * @mperf:              Difference of maximum performance frequency clock count
111  *                      read from MPERF MSR between last and current sample
112  * @tsc:                Difference of time stamp counter between last and
113  *                      current sample
114  * @time:               Current time from scheduler
115  *
116  * This structure is used in the cpudata structure to store performance sample
117  * data for choosing next P State.
118  */
119 struct sample {
120         int32_t core_avg_perf;
121         int32_t busy_scaled;
122         u64 aperf;
123         u64 mperf;
124         u64 tsc;
125         u64 time;
126 };
127
128 /**
129  * struct pstate_data - Store P state data
130  * @current_pstate:     Current requested P state
131  * @min_pstate:         Min P state possible for this platform
132  * @max_pstate:         Max P state possible for this platform
133  * @max_pstate_physical:This is physical Max P state for a processor
134  *                      This can be higher than the max_pstate which can
135  *                      be limited by platform thermal design power limits
136  * @scaling:            Scaling factor to  convert frequency to cpufreq
137  *                      frequency units
138  * @turbo_pstate:       Max Turbo P state possible for this platform
139  * @max_freq:           @max_pstate frequency in cpufreq units
140  * @turbo_freq:         @turbo_pstate frequency in cpufreq units
141  *
142  * Stores the per cpu model P state limits and current P state.
143  */
144 struct pstate_data {
145         int     current_pstate;
146         int     min_pstate;
147         int     max_pstate;
148         int     max_pstate_physical;
149         int     scaling;
150         int     turbo_pstate;
151         unsigned int max_freq;
152         unsigned int turbo_freq;
153 };
154
155 /**
156  * struct vid_data -    Stores voltage information data
157  * @min:                VID data for this platform corresponding to
158  *                      the lowest P state
159  * @max:                VID data corresponding to the highest P State.
160  * @turbo:              VID data for turbo P state
161  * @ratio:              Ratio of (vid max - vid min) /
162  *                      (max P state - Min P State)
163  *
164  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
165  * This data is used in Atom platforms, where in addition to target P state,
166  * the voltage data needs to be specified to select next P State.
167  */
168 struct vid_data {
169         int min;
170         int max;
171         int turbo;
172         int32_t ratio;
173 };
174
175 /**
176  * struct _pid -        Stores PID data
177  * @setpoint:           Target set point for busyness or performance
178  * @integral:           Storage for accumulated error values
179  * @p_gain:             PID proportional gain
180  * @i_gain:             PID integral gain
181  * @d_gain:             PID derivative gain
182  * @deadband:           PID deadband
183  * @last_err:           Last error storage for integral part of PID calculation
184  *
185  * Stores PID coefficients and last error for PID controller.
186  */
187 struct _pid {
188         int setpoint;
189         int32_t integral;
190         int32_t p_gain;
191         int32_t i_gain;
192         int32_t d_gain;
193         int deadband;
194         int32_t last_err;
195 };
196
197 /**
198  * struct global_params - Global parameters, mostly tunable via sysfs.
199  * @no_turbo:           Whether or not to use turbo P-states.
200  * @turbo_disabled:     Whethet or not turbo P-states are available at all,
201  *                      based on the MSR_IA32_MISC_ENABLE value and whether or
202  *                      not the maximum reported turbo P-state is different from
203  *                      the maximum reported non-turbo one.
204  * @min_perf_pct:       Minimum capacity limit in percent of the maximum turbo
205  *                      P-state capacity.
206  * @max_perf_pct:       Maximum capacity limit in percent of the maximum turbo
207  *                      P-state capacity.
208  */
209 struct global_params {
210         bool no_turbo;
211         bool turbo_disabled;
212         int max_perf_pct;
213         int min_perf_pct;
214 };
215
216 /**
217  * struct cpudata -     Per CPU instance data storage
218  * @cpu:                CPU number for this instance data
219  * @policy:             CPUFreq policy value
220  * @update_util:        CPUFreq utility callback information
221  * @update_util_set:    CPUFreq utility callback is set
222  * @iowait_boost:       iowait-related boost fraction
223  * @last_update:        Time of the last update.
224  * @pstate:             Stores P state limits for this CPU
225  * @vid:                Stores VID limits for this CPU
226  * @pid:                Stores PID parameters for this CPU
227  * @last_sample_time:   Last Sample time
228  * @prev_aperf:         Last APERF value read from APERF MSR
229  * @prev_mperf:         Last MPERF value read from MPERF MSR
230  * @prev_tsc:           Last timestamp counter (TSC) value
231  * @prev_cummulative_iowait: IO Wait time difference from last and
232  *                      current sample
233  * @sample:             Storage for storing last Sample data
234  * @min_perf_ratio:     Minimum capacity in terms of PERF or HWP ratios
235  * @max_perf_ratio:     Maximum capacity in terms of PERF or HWP ratios
236  * @acpi_perf_data:     Stores ACPI perf information read from _PSS
237  * @valid_pss_table:    Set to true for valid ACPI _PSS entries found
238  * @epp_powersave:      Last saved HWP energy performance preference
239  *                      (EPP) or energy performance bias (EPB),
240  *                      when policy switched to performance
241  * @epp_policy:         Last saved policy used to set EPP/EPB
242  * @epp_default:        Power on default HWP energy performance
243  *                      preference/bias
244  * @epp_saved:          Saved EPP/EPB during system suspend or CPU offline
245  *                      operation
246  *
247  * This structure stores per CPU instance data for all CPUs.
248  */
249 struct cpudata {
250         int cpu;
251
252         unsigned int policy;
253         struct update_util_data update_util;
254         bool   update_util_set;
255
256         struct pstate_data pstate;
257         struct vid_data vid;
258         struct _pid pid;
259
260         u64     last_update;
261         u64     last_sample_time;
262         u64     prev_aperf;
263         u64     prev_mperf;
264         u64     prev_tsc;
265         u64     prev_cummulative_iowait;
266         struct sample sample;
267         int32_t min_perf_ratio;
268         int32_t max_perf_ratio;
269 #ifdef CONFIG_ACPI
270         struct acpi_processor_performance acpi_perf_data;
271         bool valid_pss_table;
272 #endif
273         unsigned int iowait_boost;
274         s16 epp_powersave;
275         s16 epp_policy;
276         s16 epp_default;
277         s16 epp_saved;
278 };
279
280 static struct cpudata **all_cpu_data;
281
282 /**
283  * struct pstate_adjust_policy - Stores static PID configuration data
284  * @sample_rate_ms:     PID calculation sample rate in ms
285  * @sample_rate_ns:     Sample rate calculation in ns
286  * @deadband:           PID deadband
287  * @setpoint:           PID Setpoint
288  * @p_gain_pct:         PID proportional gain
289  * @i_gain_pct:         PID integral gain
290  * @d_gain_pct:         PID derivative gain
291  *
292  * Stores per CPU model static PID configuration data.
293  */
294 struct pstate_adjust_policy {
295         int sample_rate_ms;
296         s64 sample_rate_ns;
297         int deadband;
298         int setpoint;
299         int p_gain_pct;
300         int d_gain_pct;
301         int i_gain_pct;
302 };
303
304 /**
305  * struct pstate_funcs - Per CPU model specific callbacks
306  * @get_max:            Callback to get maximum non turbo effective P state
307  * @get_max_physical:   Callback to get maximum non turbo physical P state
308  * @get_min:            Callback to get minimum P state
309  * @get_turbo:          Callback to get turbo P state
310  * @get_scaling:        Callback to get frequency scaling factor
311  * @get_val:            Callback to convert P state to actual MSR write value
312  * @get_vid:            Callback to get VID data for Atom platforms
313  * @update_util:        Active mode utilization update callback.
314  *
315  * Core and Atom CPU models have different way to get P State limits. This
316  * structure is used to store those callbacks.
317  */
318 struct pstate_funcs {
319         int (*get_max)(void);
320         int (*get_max_physical)(void);
321         int (*get_min)(void);
322         int (*get_turbo)(void);
323         int (*get_scaling)(void);
324         u64 (*get_val)(struct cpudata*, int pstate);
325         void (*get_vid)(struct cpudata *);
326         void (*update_util)(struct update_util_data *data, u64 time,
327                             unsigned int flags);
328 };
329
330 static struct pstate_funcs pstate_funcs __read_mostly;
331 static struct pstate_adjust_policy pid_params __read_mostly = {
332         .sample_rate_ms = 10,
333         .sample_rate_ns = 10 * NSEC_PER_MSEC,
334         .deadband = 0,
335         .setpoint = 97,
336         .p_gain_pct = 20,
337         .d_gain_pct = 0,
338         .i_gain_pct = 0,
339 };
340
341 static int hwp_active __read_mostly;
342 static bool per_cpu_limits __read_mostly;
343
344 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
345
346 #ifdef CONFIG_ACPI
347 static bool acpi_ppc;
348 #endif
349
350 static struct global_params global;
351
352 static DEFINE_MUTEX(intel_pstate_driver_lock);
353 static DEFINE_MUTEX(intel_pstate_limits_lock);
354
355 #ifdef CONFIG_ACPI
356
357 static bool intel_pstate_get_ppc_enable_status(void)
358 {
359         if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
360             acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
361                 return true;
362
363         return acpi_ppc;
364 }
365
366 #ifdef CONFIG_ACPI_CPPC_LIB
367
368 /* The work item is needed to avoid CPU hotplug locking issues */
369 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
370 {
371         sched_set_itmt_support();
372 }
373
374 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
375
376 static void intel_pstate_set_itmt_prio(int cpu)
377 {
378         struct cppc_perf_caps cppc_perf;
379         static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
380         int ret;
381
382         ret = cppc_get_perf_caps(cpu, &cppc_perf);
383         if (ret)
384                 return;
385
386         /*
387          * The priorities can be set regardless of whether or not
388          * sched_set_itmt_support(true) has been called and it is valid to
389          * update them at any time after it has been called.
390          */
391         sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
392
393         if (max_highest_perf <= min_highest_perf) {
394                 if (cppc_perf.highest_perf > max_highest_perf)
395                         max_highest_perf = cppc_perf.highest_perf;
396
397                 if (cppc_perf.highest_perf < min_highest_perf)
398                         min_highest_perf = cppc_perf.highest_perf;
399
400                 if (max_highest_perf > min_highest_perf) {
401                         /*
402                          * This code can be run during CPU online under the
403                          * CPU hotplug locks, so sched_set_itmt_support()
404                          * cannot be called from here.  Queue up a work item
405                          * to invoke it.
406                          */
407                         schedule_work(&sched_itmt_work);
408                 }
409         }
410 }
411 #else
412 static void intel_pstate_set_itmt_prio(int cpu)
413 {
414 }
415 #endif
416
417 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
418 {
419         struct cpudata *cpu;
420         int ret;
421         int i;
422
423         if (hwp_active) {
424                 intel_pstate_set_itmt_prio(policy->cpu);
425                 return;
426         }
427
428         if (!intel_pstate_get_ppc_enable_status())
429                 return;
430
431         cpu = all_cpu_data[policy->cpu];
432
433         ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
434                                                   policy->cpu);
435         if (ret)
436                 return;
437
438         /*
439          * Check if the control value in _PSS is for PERF_CTL MSR, which should
440          * guarantee that the states returned by it map to the states in our
441          * list directly.
442          */
443         if (cpu->acpi_perf_data.control_register.space_id !=
444                                                 ACPI_ADR_SPACE_FIXED_HARDWARE)
445                 goto err;
446
447         /*
448          * If there is only one entry _PSS, simply ignore _PSS and continue as
449          * usual without taking _PSS into account
450          */
451         if (cpu->acpi_perf_data.state_count < 2)
452                 goto err;
453
454         pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
455         for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
456                 pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
457                          (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
458                          (u32) cpu->acpi_perf_data.states[i].core_frequency,
459                          (u32) cpu->acpi_perf_data.states[i].power,
460                          (u32) cpu->acpi_perf_data.states[i].control);
461         }
462
463         /*
464          * The _PSS table doesn't contain whole turbo frequency range.
465          * This just contains +1 MHZ above the max non turbo frequency,
466          * with control value corresponding to max turbo ratio. But
467          * when cpufreq set policy is called, it will call with this
468          * max frequency, which will cause a reduced performance as
469          * this driver uses real max turbo frequency as the max
470          * frequency. So correct this frequency in _PSS table to
471          * correct max turbo frequency based on the turbo state.
472          * Also need to convert to MHz as _PSS freq is in MHz.
473          */
474         if (!global.turbo_disabled)
475                 cpu->acpi_perf_data.states[0].core_frequency =
476                                         policy->cpuinfo.max_freq / 1000;
477         cpu->valid_pss_table = true;
478         pr_debug("_PPC limits will be enforced\n");
479
480         return;
481
482  err:
483         cpu->valid_pss_table = false;
484         acpi_processor_unregister_performance(policy->cpu);
485 }
486
487 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
488 {
489         struct cpudata *cpu;
490
491         cpu = all_cpu_data[policy->cpu];
492         if (!cpu->valid_pss_table)
493                 return;
494
495         acpi_processor_unregister_performance(policy->cpu);
496 }
497 #else
498 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
499 {
500 }
501
502 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
503 {
504 }
505 #endif
506
507 static signed int pid_calc(struct _pid *pid, int32_t busy)
508 {
509         signed int result;
510         int32_t pterm, dterm, fp_error;
511         int32_t integral_limit;
512
513         fp_error = pid->setpoint - busy;
514
515         if (abs(fp_error) <= pid->deadband)
516                 return 0;
517
518         pterm = mul_fp(pid->p_gain, fp_error);
519
520         pid->integral += fp_error;
521
522         /*
523          * We limit the integral here so that it will never
524          * get higher than 30.  This prevents it from becoming
525          * too large an input over long periods of time and allows
526          * it to get factored out sooner.
527          *
528          * The value of 30 was chosen through experimentation.
529          */
530         integral_limit = int_tofp(30);
531         if (pid->integral > integral_limit)
532                 pid->integral = integral_limit;
533         if (pid->integral < -integral_limit)
534                 pid->integral = -integral_limit;
535
536         dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
537         pid->last_err = fp_error;
538
539         result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
540         result = result + (1 << (FRAC_BITS-1));
541         return (signed int)fp_toint(result);
542 }
543
544 static inline void intel_pstate_pid_reset(struct cpudata *cpu)
545 {
546         struct _pid *pid = &cpu->pid;
547
548         pid->p_gain = percent_fp(pid_params.p_gain_pct);
549         pid->d_gain = percent_fp(pid_params.d_gain_pct);
550         pid->i_gain = percent_fp(pid_params.i_gain_pct);
551         pid->setpoint = int_tofp(pid_params.setpoint);
552         pid->last_err  = pid->setpoint - int_tofp(100);
553         pid->deadband  = int_tofp(pid_params.deadband);
554         pid->integral  = 0;
555 }
556
557 static inline void update_turbo_state(void)
558 {
559         u64 misc_en;
560         struct cpudata *cpu;
561
562         cpu = all_cpu_data[0];
563         rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
564         global.turbo_disabled =
565                 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
566                  cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
567 }
568
569 static int min_perf_pct_min(void)
570 {
571         struct cpudata *cpu = all_cpu_data[0];
572         int turbo_pstate = cpu->pstate.turbo_pstate;
573
574         return turbo_pstate ?
575                 DIV_ROUND_UP(cpu->pstate.min_pstate * 100, turbo_pstate) : 0;
576 }
577
578 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
579 {
580         u64 epb;
581         int ret;
582
583         if (!static_cpu_has(X86_FEATURE_EPB))
584                 return -ENXIO;
585
586         ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
587         if (ret)
588                 return (s16)ret;
589
590         return (s16)(epb & 0x0f);
591 }
592
593 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
594 {
595         s16 epp;
596
597         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
598                 /*
599                  * When hwp_req_data is 0, means that caller didn't read
600                  * MSR_HWP_REQUEST, so need to read and get EPP.
601                  */
602                 if (!hwp_req_data) {
603                         epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
604                                             &hwp_req_data);
605                         if (epp)
606                                 return epp;
607                 }
608                 epp = (hwp_req_data >> 24) & 0xff;
609         } else {
610                 /* When there is no EPP present, HWP uses EPB settings */
611                 epp = intel_pstate_get_epb(cpu_data);
612         }
613
614         return epp;
615 }
616
617 static int intel_pstate_set_epb(int cpu, s16 pref)
618 {
619         u64 epb;
620         int ret;
621
622         if (!static_cpu_has(X86_FEATURE_EPB))
623                 return -ENXIO;
624
625         ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
626         if (ret)
627                 return ret;
628
629         epb = (epb & ~0x0f) | pref;
630         wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
631
632         return 0;
633 }
634
635 /*
636  * EPP/EPB display strings corresponding to EPP index in the
637  * energy_perf_strings[]
638  *      index           String
639  *-------------------------------------
640  *      0               default
641  *      1               performance
642  *      2               balance_performance
643  *      3               balance_power
644  *      4               power
645  */
646 static const char * const energy_perf_strings[] = {
647         "default",
648         "performance",
649         "balance_performance",
650         "balance_power",
651         "power",
652         NULL
653 };
654
655 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
656 {
657         s16 epp;
658         int index = -EINVAL;
659
660         epp = intel_pstate_get_epp(cpu_data, 0);
661         if (epp < 0)
662                 return epp;
663
664         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
665                 /*
666                  * Range:
667                  *      0x00-0x3F       :       Performance
668                  *      0x40-0x7F       :       Balance performance
669                  *      0x80-0xBF       :       Balance power
670                  *      0xC0-0xFF       :       Power
671                  * The EPP is a 8 bit value, but our ranges restrict the
672                  * value which can be set. Here only using top two bits
673                  * effectively.
674                  */
675                 index = (epp >> 6) + 1;
676         } else if (static_cpu_has(X86_FEATURE_EPB)) {
677                 /*
678                  * Range:
679                  *      0x00-0x03       :       Performance
680                  *      0x04-0x07       :       Balance performance
681                  *      0x08-0x0B       :       Balance power
682                  *      0x0C-0x0F       :       Power
683                  * The EPB is a 4 bit value, but our ranges restrict the
684                  * value which can be set. Here only using top two bits
685                  * effectively.
686                  */
687                 index = (epp >> 2) + 1;
688         }
689
690         return index;
691 }
692
693 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
694                                               int pref_index)
695 {
696         int epp = -EINVAL;
697         int ret;
698
699         if (!pref_index)
700                 epp = cpu_data->epp_default;
701
702         mutex_lock(&intel_pstate_limits_lock);
703
704         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
705                 u64 value;
706
707                 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
708                 if (ret)
709                         goto return_pref;
710
711                 value &= ~GENMASK_ULL(31, 24);
712
713                 /*
714                  * If epp is not default, convert from index into
715                  * energy_perf_strings to epp value, by shifting 6
716                  * bits left to use only top two bits in epp.
717                  * The resultant epp need to shifted by 24 bits to
718                  * epp position in MSR_HWP_REQUEST.
719                  */
720                 if (epp == -EINVAL)
721                         epp = (pref_index - 1) << 6;
722
723                 value |= (u64)epp << 24;
724                 ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
725         } else {
726                 if (epp == -EINVAL)
727                         epp = (pref_index - 1) << 2;
728                 ret = intel_pstate_set_epb(cpu_data->cpu, epp);
729         }
730 return_pref:
731         mutex_unlock(&intel_pstate_limits_lock);
732
733         return ret;
734 }
735
736 static ssize_t show_energy_performance_available_preferences(
737                                 struct cpufreq_policy *policy, char *buf)
738 {
739         int i = 0;
740         int ret = 0;
741
742         while (energy_perf_strings[i] != NULL)
743                 ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
744
745         ret += sprintf(&buf[ret], "\n");
746
747         return ret;
748 }
749
750 cpufreq_freq_attr_ro(energy_performance_available_preferences);
751
752 static ssize_t store_energy_performance_preference(
753                 struct cpufreq_policy *policy, const char *buf, size_t count)
754 {
755         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
756         char str_preference[21];
757         int ret, i = 0;
758
759         ret = sscanf(buf, "%20s", str_preference);
760         if (ret != 1)
761                 return -EINVAL;
762
763         while (energy_perf_strings[i] != NULL) {
764                 if (!strcmp(str_preference, energy_perf_strings[i])) {
765                         intel_pstate_set_energy_pref_index(cpu_data, i);
766                         return count;
767                 }
768                 ++i;
769         }
770
771         return -EINVAL;
772 }
773
774 static ssize_t show_energy_performance_preference(
775                                 struct cpufreq_policy *policy, char *buf)
776 {
777         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
778         int preference;
779
780         preference = intel_pstate_get_energy_pref_index(cpu_data);
781         if (preference < 0)
782                 return preference;
783
784         return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
785 }
786
787 cpufreq_freq_attr_rw(energy_performance_preference);
788
789 static struct freq_attr *hwp_cpufreq_attrs[] = {
790         &energy_performance_preference,
791         &energy_performance_available_preferences,
792         NULL,
793 };
794
795 static void intel_pstate_get_hwp_max(unsigned int cpu, int *phy_max,
796                                      int *current_max)
797 {
798         u64 cap;
799
800         rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
801         if (global.no_turbo)
802                 *current_max = HWP_GUARANTEED_PERF(cap);
803         else
804                 *current_max = HWP_HIGHEST_PERF(cap);
805
806         *phy_max = HWP_HIGHEST_PERF(cap);
807 }
808
809 static void intel_pstate_hwp_set(unsigned int cpu)
810 {
811         struct cpudata *cpu_data = all_cpu_data[cpu];
812         int max, min;
813         u64 value;
814         s16 epp;
815
816         max = cpu_data->max_perf_ratio;
817         min = cpu_data->min_perf_ratio;
818
819         if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
820                 min = max;
821
822         rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
823
824         value &= ~HWP_MIN_PERF(~0L);
825         value |= HWP_MIN_PERF(min);
826
827         value &= ~HWP_MAX_PERF(~0L);
828         value |= HWP_MAX_PERF(max);
829
830         if (cpu_data->epp_policy == cpu_data->policy)
831                 goto skip_epp;
832
833         cpu_data->epp_policy = cpu_data->policy;
834
835         if (cpu_data->epp_saved >= 0) {
836                 epp = cpu_data->epp_saved;
837                 cpu_data->epp_saved = -EINVAL;
838                 goto update_epp;
839         }
840
841         if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
842                 epp = intel_pstate_get_epp(cpu_data, value);
843                 cpu_data->epp_powersave = epp;
844                 /* If EPP read was failed, then don't try to write */
845                 if (epp < 0)
846                         goto skip_epp;
847
848                 epp = 0;
849         } else {
850                 /* skip setting EPP, when saved value is invalid */
851                 if (cpu_data->epp_powersave < 0)
852                         goto skip_epp;
853
854                 /*
855                  * No need to restore EPP when it is not zero. This
856                  * means:
857                  *  - Policy is not changed
858                  *  - user has manually changed
859                  *  - Error reading EPB
860                  */
861                 epp = intel_pstate_get_epp(cpu_data, value);
862                 if (epp)
863                         goto skip_epp;
864
865                 epp = cpu_data->epp_powersave;
866         }
867 update_epp:
868         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
869                 value &= ~GENMASK_ULL(31, 24);
870                 value |= (u64)epp << 24;
871         } else {
872                 intel_pstate_set_epb(cpu, epp);
873         }
874 skip_epp:
875         wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
876 }
877
878 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
879 {
880         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
881
882         if (!hwp_active)
883                 return 0;
884
885         cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
886
887         return 0;
888 }
889
890 static int intel_pstate_resume(struct cpufreq_policy *policy)
891 {
892         if (!hwp_active)
893                 return 0;
894
895         mutex_lock(&intel_pstate_limits_lock);
896
897         all_cpu_data[policy->cpu]->epp_policy = 0;
898         intel_pstate_hwp_set(policy->cpu);
899
900         mutex_unlock(&intel_pstate_limits_lock);
901
902         return 0;
903 }
904
905 static void intel_pstate_update_policies(void)
906 {
907         int cpu;
908
909         for_each_possible_cpu(cpu)
910                 cpufreq_update_policy(cpu);
911 }
912
913 /************************** debugfs begin ************************/
914 static int pid_param_set(void *data, u64 val)
915 {
916         unsigned int cpu;
917
918         *(u32 *)data = val;
919         pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
920         for_each_possible_cpu(cpu)
921                 if (all_cpu_data[cpu])
922                         intel_pstate_pid_reset(all_cpu_data[cpu]);
923
924         return 0;
925 }
926
927 static int pid_param_get(void *data, u64 *val)
928 {
929         *val = *(u32 *)data;
930         return 0;
931 }
932 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
933
934 static struct dentry *debugfs_parent;
935
936 struct pid_param {
937         char *name;
938         void *value;
939         struct dentry *dentry;
940 };
941
942 static struct pid_param pid_files[] = {
943         {"sample_rate_ms", &pid_params.sample_rate_ms, },
944         {"d_gain_pct", &pid_params.d_gain_pct, },
945         {"i_gain_pct", &pid_params.i_gain_pct, },
946         {"deadband", &pid_params.deadband, },
947         {"setpoint", &pid_params.setpoint, },
948         {"p_gain_pct", &pid_params.p_gain_pct, },
949         {NULL, NULL, }
950 };
951
952 static void intel_pstate_debug_expose_params(void)
953 {
954         int i;
955
956         debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
957         if (IS_ERR_OR_NULL(debugfs_parent))
958                 return;
959
960         for (i = 0; pid_files[i].name; i++) {
961                 struct dentry *dentry;
962
963                 dentry = debugfs_create_file(pid_files[i].name, 0660,
964                                              debugfs_parent, pid_files[i].value,
965                                              &fops_pid_param);
966                 if (!IS_ERR(dentry))
967                         pid_files[i].dentry = dentry;
968         }
969 }
970
971 static void intel_pstate_debug_hide_params(void)
972 {
973         int i;
974
975         if (IS_ERR_OR_NULL(debugfs_parent))
976                 return;
977
978         for (i = 0; pid_files[i].name; i++) {
979                 debugfs_remove(pid_files[i].dentry);
980                 pid_files[i].dentry = NULL;
981         }
982
983         debugfs_remove(debugfs_parent);
984         debugfs_parent = NULL;
985 }
986
987 /************************** debugfs end ************************/
988
989 /************************** sysfs begin ************************/
990 #define show_one(file_name, object)                                     \
991         static ssize_t show_##file_name                                 \
992         (struct kobject *kobj, struct attribute *attr, char *buf)       \
993         {                                                               \
994                 return sprintf(buf, "%u\n", global.object);             \
995         }
996
997 static ssize_t intel_pstate_show_status(char *buf);
998 static int intel_pstate_update_status(const char *buf, size_t size);
999
1000 static ssize_t show_status(struct kobject *kobj,
1001                            struct attribute *attr, char *buf)
1002 {
1003         ssize_t ret;
1004
1005         mutex_lock(&intel_pstate_driver_lock);
1006         ret = intel_pstate_show_status(buf);
1007         mutex_unlock(&intel_pstate_driver_lock);
1008
1009         return ret;
1010 }
1011
1012 static ssize_t store_status(struct kobject *a, struct attribute *b,
1013                             const char *buf, size_t count)
1014 {
1015         char *p = memchr(buf, '\n', count);
1016         int ret;
1017
1018         mutex_lock(&intel_pstate_driver_lock);
1019         ret = intel_pstate_update_status(buf, p ? p - buf : count);
1020         mutex_unlock(&intel_pstate_driver_lock);
1021
1022         return ret < 0 ? ret : count;
1023 }
1024
1025 static ssize_t show_turbo_pct(struct kobject *kobj,
1026                                 struct attribute *attr, char *buf)
1027 {
1028         struct cpudata *cpu;
1029         int total, no_turbo, turbo_pct;
1030         uint32_t turbo_fp;
1031
1032         mutex_lock(&intel_pstate_driver_lock);
1033
1034         if (!intel_pstate_driver) {
1035                 mutex_unlock(&intel_pstate_driver_lock);
1036                 return -EAGAIN;
1037         }
1038
1039         cpu = all_cpu_data[0];
1040
1041         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1042         no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1043         turbo_fp = div_fp(no_turbo, total);
1044         turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1045
1046         mutex_unlock(&intel_pstate_driver_lock);
1047
1048         return sprintf(buf, "%u\n", turbo_pct);
1049 }
1050
1051 static ssize_t show_num_pstates(struct kobject *kobj,
1052                                 struct attribute *attr, char *buf)
1053 {
1054         struct cpudata *cpu;
1055         int total;
1056
1057         mutex_lock(&intel_pstate_driver_lock);
1058
1059         if (!intel_pstate_driver) {
1060                 mutex_unlock(&intel_pstate_driver_lock);
1061                 return -EAGAIN;
1062         }
1063
1064         cpu = all_cpu_data[0];
1065         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1066
1067         mutex_unlock(&intel_pstate_driver_lock);
1068
1069         return sprintf(buf, "%u\n", total);
1070 }
1071
1072 static ssize_t show_no_turbo(struct kobject *kobj,
1073                              struct attribute *attr, char *buf)
1074 {
1075         ssize_t ret;
1076
1077         mutex_lock(&intel_pstate_driver_lock);
1078
1079         if (!intel_pstate_driver) {
1080                 mutex_unlock(&intel_pstate_driver_lock);
1081                 return -EAGAIN;
1082         }
1083
1084         update_turbo_state();
1085         if (global.turbo_disabled)
1086                 ret = sprintf(buf, "%u\n", global.turbo_disabled);
1087         else
1088                 ret = sprintf(buf, "%u\n", global.no_turbo);
1089
1090         mutex_unlock(&intel_pstate_driver_lock);
1091
1092         return ret;
1093 }
1094
1095 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
1096                               const char *buf, size_t count)
1097 {
1098         unsigned int input;
1099         int ret;
1100
1101         ret = sscanf(buf, "%u", &input);
1102         if (ret != 1)
1103                 return -EINVAL;
1104
1105         mutex_lock(&intel_pstate_driver_lock);
1106
1107         if (!intel_pstate_driver) {
1108                 mutex_unlock(&intel_pstate_driver_lock);
1109                 return -EAGAIN;
1110         }
1111
1112         mutex_lock(&intel_pstate_limits_lock);
1113
1114         update_turbo_state();
1115         if (global.turbo_disabled) {
1116                 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1117                 mutex_unlock(&intel_pstate_limits_lock);
1118                 mutex_unlock(&intel_pstate_driver_lock);
1119                 return -EPERM;
1120         }
1121
1122         global.no_turbo = clamp_t(int, input, 0, 1);
1123
1124         if (global.no_turbo) {
1125                 struct cpudata *cpu = all_cpu_data[0];
1126                 int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1127
1128                 /* Squash the global minimum into the permitted range. */
1129                 if (global.min_perf_pct > pct)
1130                         global.min_perf_pct = pct;
1131         }
1132
1133         mutex_unlock(&intel_pstate_limits_lock);
1134
1135         intel_pstate_update_policies();
1136
1137         mutex_unlock(&intel_pstate_driver_lock);
1138
1139         return count;
1140 }
1141
1142 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
1143                                   const char *buf, size_t count)
1144 {
1145         unsigned int input;
1146         int ret;
1147
1148         ret = sscanf(buf, "%u", &input);
1149         if (ret != 1)
1150                 return -EINVAL;
1151
1152         mutex_lock(&intel_pstate_driver_lock);
1153
1154         if (!intel_pstate_driver) {
1155                 mutex_unlock(&intel_pstate_driver_lock);
1156                 return -EAGAIN;
1157         }
1158
1159         mutex_lock(&intel_pstate_limits_lock);
1160
1161         global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1162
1163         mutex_unlock(&intel_pstate_limits_lock);
1164
1165         intel_pstate_update_policies();
1166
1167         mutex_unlock(&intel_pstate_driver_lock);
1168
1169         return count;
1170 }
1171
1172 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
1173                                   const char *buf, size_t count)
1174 {
1175         unsigned int input;
1176         int ret;
1177
1178         ret = sscanf(buf, "%u", &input);
1179         if (ret != 1)
1180                 return -EINVAL;
1181
1182         mutex_lock(&intel_pstate_driver_lock);
1183
1184         if (!intel_pstate_driver) {
1185                 mutex_unlock(&intel_pstate_driver_lock);
1186                 return -EAGAIN;
1187         }
1188
1189         mutex_lock(&intel_pstate_limits_lock);
1190
1191         global.min_perf_pct = clamp_t(int, input,
1192                                       min_perf_pct_min(), global.max_perf_pct);
1193
1194         mutex_unlock(&intel_pstate_limits_lock);
1195
1196         intel_pstate_update_policies();
1197
1198         mutex_unlock(&intel_pstate_driver_lock);
1199
1200         return count;
1201 }
1202
1203 show_one(max_perf_pct, max_perf_pct);
1204 show_one(min_perf_pct, min_perf_pct);
1205
1206 define_one_global_rw(status);
1207 define_one_global_rw(no_turbo);
1208 define_one_global_rw(max_perf_pct);
1209 define_one_global_rw(min_perf_pct);
1210 define_one_global_ro(turbo_pct);
1211 define_one_global_ro(num_pstates);
1212
1213 static struct attribute *intel_pstate_attributes[] = {
1214         &status.attr,
1215         &no_turbo.attr,
1216         &turbo_pct.attr,
1217         &num_pstates.attr,
1218         NULL
1219 };
1220
1221 static struct attribute_group intel_pstate_attr_group = {
1222         .attrs = intel_pstate_attributes,
1223 };
1224
1225 static void __init intel_pstate_sysfs_expose_params(void)
1226 {
1227         struct kobject *intel_pstate_kobject;
1228         int rc;
1229
1230         intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1231                                                 &cpu_subsys.dev_root->kobj);
1232         if (WARN_ON(!intel_pstate_kobject))
1233                 return;
1234
1235         rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1236         if (WARN_ON(rc))
1237                 return;
1238
1239         /*
1240          * If per cpu limits are enforced there are no global limits, so
1241          * return without creating max/min_perf_pct attributes
1242          */
1243         if (per_cpu_limits)
1244                 return;
1245
1246         rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1247         WARN_ON(rc);
1248
1249         rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1250         WARN_ON(rc);
1251
1252 }
1253 /************************** sysfs end ************************/
1254
1255 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1256 {
1257         /* First disable HWP notification interrupt as we don't process them */
1258         if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1259                 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1260
1261         wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1262         cpudata->epp_policy = 0;
1263         if (cpudata->epp_default == -EINVAL)
1264                 cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1265 }
1266
1267 #define MSR_IA32_POWER_CTL_BIT_EE       19
1268
1269 /* Disable energy efficiency optimization */
1270 static void intel_pstate_disable_ee(int cpu)
1271 {
1272         u64 power_ctl;
1273         int ret;
1274
1275         ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1276         if (ret)
1277                 return;
1278
1279         if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1280                 pr_info("Disabling energy efficiency optimization\n");
1281                 power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1282                 wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1283         }
1284 }
1285
1286 static int atom_get_min_pstate(void)
1287 {
1288         u64 value;
1289
1290         rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1291         return (value >> 8) & 0x7F;
1292 }
1293
1294 static int atom_get_max_pstate(void)
1295 {
1296         u64 value;
1297
1298         rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1299         return (value >> 16) & 0x7F;
1300 }
1301
1302 static int atom_get_turbo_pstate(void)
1303 {
1304         u64 value;
1305
1306         rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1307         return value & 0x7F;
1308 }
1309
1310 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1311 {
1312         u64 val;
1313         int32_t vid_fp;
1314         u32 vid;
1315
1316         val = (u64)pstate << 8;
1317         if (global.no_turbo && !global.turbo_disabled)
1318                 val |= (u64)1 << 32;
1319
1320         vid_fp = cpudata->vid.min + mul_fp(
1321                 int_tofp(pstate - cpudata->pstate.min_pstate),
1322                 cpudata->vid.ratio);
1323
1324         vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1325         vid = ceiling_fp(vid_fp);
1326
1327         if (pstate > cpudata->pstate.max_pstate)
1328                 vid = cpudata->vid.turbo;
1329
1330         return val | vid;
1331 }
1332
1333 static int silvermont_get_scaling(void)
1334 {
1335         u64 value;
1336         int i;
1337         /* Defined in Table 35-6 from SDM (Sept 2015) */
1338         static int silvermont_freq_table[] = {
1339                 83300, 100000, 133300, 116700, 80000};
1340
1341         rdmsrl(MSR_FSB_FREQ, value);
1342         i = value & 0x7;
1343         WARN_ON(i > 4);
1344
1345         return silvermont_freq_table[i];
1346 }
1347
1348 static int airmont_get_scaling(void)
1349 {
1350         u64 value;
1351         int i;
1352         /* Defined in Table 35-10 from SDM (Sept 2015) */
1353         static int airmont_freq_table[] = {
1354                 83300, 100000, 133300, 116700, 80000,
1355                 93300, 90000, 88900, 87500};
1356
1357         rdmsrl(MSR_FSB_FREQ, value);
1358         i = value & 0xF;
1359         WARN_ON(i > 8);
1360
1361         return airmont_freq_table[i];
1362 }
1363
1364 static void atom_get_vid(struct cpudata *cpudata)
1365 {
1366         u64 value;
1367
1368         rdmsrl(MSR_ATOM_CORE_VIDS, value);
1369         cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1370         cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1371         cpudata->vid.ratio = div_fp(
1372                 cpudata->vid.max - cpudata->vid.min,
1373                 int_tofp(cpudata->pstate.max_pstate -
1374                         cpudata->pstate.min_pstate));
1375
1376         rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1377         cpudata->vid.turbo = value & 0x7f;
1378 }
1379
1380 static int core_get_min_pstate(void)
1381 {
1382         u64 value;
1383
1384         rdmsrl(MSR_PLATFORM_INFO, value);
1385         return (value >> 40) & 0xFF;
1386 }
1387
1388 static int core_get_max_pstate_physical(void)
1389 {
1390         u64 value;
1391
1392         rdmsrl(MSR_PLATFORM_INFO, value);
1393         return (value >> 8) & 0xFF;
1394 }
1395
1396 static int core_get_tdp_ratio(u64 plat_info)
1397 {
1398         /* Check how many TDP levels present */
1399         if (plat_info & 0x600000000) {
1400                 u64 tdp_ctrl;
1401                 u64 tdp_ratio;
1402                 int tdp_msr;
1403                 int err;
1404
1405                 /* Get the TDP level (0, 1, 2) to get ratios */
1406                 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1407                 if (err)
1408                         return err;
1409
1410                 /* TDP MSR are continuous starting at 0x648 */
1411                 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1412                 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1413                 if (err)
1414                         return err;
1415
1416                 /* For level 1 and 2, bits[23:16] contain the ratio */
1417                 if (tdp_ctrl & 0x03)
1418                         tdp_ratio >>= 16;
1419
1420                 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1421                 pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1422
1423                 return (int)tdp_ratio;
1424         }
1425
1426         return -ENXIO;
1427 }
1428
1429 static int core_get_max_pstate(void)
1430 {
1431         u64 tar;
1432         u64 plat_info;
1433         int max_pstate;
1434         int tdp_ratio;
1435         int err;
1436
1437         rdmsrl(MSR_PLATFORM_INFO, plat_info);
1438         max_pstate = (plat_info >> 8) & 0xFF;
1439
1440         tdp_ratio = core_get_tdp_ratio(plat_info);
1441         if (tdp_ratio <= 0)
1442                 return max_pstate;
1443
1444         if (hwp_active) {
1445                 /* Turbo activation ratio is not used on HWP platforms */
1446                 return tdp_ratio;
1447         }
1448
1449         err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1450         if (!err) {
1451                 int tar_levels;
1452
1453                 /* Do some sanity checking for safety */
1454                 tar_levels = tar & 0xff;
1455                 if (tdp_ratio - 1 == tar_levels) {
1456                         max_pstate = tar_levels;
1457                         pr_debug("max_pstate=TAC %x\n", max_pstate);
1458                 }
1459         }
1460
1461         return max_pstate;
1462 }
1463
1464 static int core_get_turbo_pstate(void)
1465 {
1466         u64 value;
1467         int nont, ret;
1468
1469         rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1470         nont = core_get_max_pstate();
1471         ret = (value) & 255;
1472         if (ret <= nont)
1473                 ret = nont;
1474         return ret;
1475 }
1476
1477 static inline int core_get_scaling(void)
1478 {
1479         return 100000;
1480 }
1481
1482 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1483 {
1484         u64 val;
1485
1486         val = (u64)pstate << 8;
1487         if (global.no_turbo && !global.turbo_disabled)
1488                 val |= (u64)1 << 32;
1489
1490         return val;
1491 }
1492
1493 static int knl_get_turbo_pstate(void)
1494 {
1495         u64 value;
1496         int nont, ret;
1497
1498         rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1499         nont = core_get_max_pstate();
1500         ret = (((value) >> 8) & 0xFF);
1501         if (ret <= nont)
1502                 ret = nont;
1503         return ret;
1504 }
1505
1506 static int intel_pstate_get_base_pstate(struct cpudata *cpu)
1507 {
1508         return global.no_turbo || global.turbo_disabled ?
1509                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1510 }
1511
1512 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1513 {
1514         trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1515         cpu->pstate.current_pstate = pstate;
1516         /*
1517          * Generally, there is no guarantee that this code will always run on
1518          * the CPU being updated, so force the register update to run on the
1519          * right CPU.
1520          */
1521         wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1522                       pstate_funcs.get_val(cpu, pstate));
1523 }
1524
1525 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1526 {
1527         intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1528 }
1529
1530 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1531 {
1532         int pstate;
1533
1534         update_turbo_state();
1535         pstate = intel_pstate_get_base_pstate(cpu);
1536         pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1537         intel_pstate_set_pstate(cpu, pstate);
1538 }
1539
1540 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1541 {
1542         cpu->pstate.min_pstate = pstate_funcs.get_min();
1543         cpu->pstate.max_pstate = pstate_funcs.get_max();
1544         cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1545         cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1546         cpu->pstate.scaling = pstate_funcs.get_scaling();
1547         cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1548         cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1549
1550         if (pstate_funcs.get_vid)
1551                 pstate_funcs.get_vid(cpu);
1552
1553         intel_pstate_set_min_pstate(cpu);
1554 }
1555
1556 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1557 {
1558         struct sample *sample = &cpu->sample;
1559
1560         sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1561 }
1562
1563 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1564 {
1565         u64 aperf, mperf;
1566         unsigned long flags;
1567         u64 tsc;
1568
1569         local_irq_save(flags);
1570         rdmsrl(MSR_IA32_APERF, aperf);
1571         rdmsrl(MSR_IA32_MPERF, mperf);
1572         tsc = rdtsc();
1573         if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1574                 local_irq_restore(flags);
1575                 return false;
1576         }
1577         local_irq_restore(flags);
1578
1579         cpu->last_sample_time = cpu->sample.time;
1580         cpu->sample.time = time;
1581         cpu->sample.aperf = aperf;
1582         cpu->sample.mperf = mperf;
1583         cpu->sample.tsc =  tsc;
1584         cpu->sample.aperf -= cpu->prev_aperf;
1585         cpu->sample.mperf -= cpu->prev_mperf;
1586         cpu->sample.tsc -= cpu->prev_tsc;
1587
1588         cpu->prev_aperf = aperf;
1589         cpu->prev_mperf = mperf;
1590         cpu->prev_tsc = tsc;
1591         /*
1592          * First time this function is invoked in a given cycle, all of the
1593          * previous sample data fields are equal to zero or stale and they must
1594          * be populated with meaningful numbers for things to work, so assume
1595          * that sample.time will always be reset before setting the utilization
1596          * update hook and make the caller skip the sample then.
1597          */
1598         if (cpu->last_sample_time) {
1599                 intel_pstate_calc_avg_perf(cpu);
1600                 return true;
1601         }
1602         return false;
1603 }
1604
1605 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1606 {
1607         return mul_ext_fp(cpu->sample.core_avg_perf,
1608                           cpu->pstate.max_pstate_physical * cpu->pstate.scaling);
1609 }
1610
1611 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1612 {
1613         return mul_ext_fp(cpu->pstate.max_pstate_physical,
1614                           cpu->sample.core_avg_perf);
1615 }
1616
1617 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1618 {
1619         struct sample *sample = &cpu->sample;
1620         int32_t busy_frac, boost;
1621         int target, avg_pstate;
1622
1623         if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE)
1624                 return cpu->pstate.turbo_pstate;
1625
1626         busy_frac = div_fp(sample->mperf, sample->tsc);
1627
1628         boost = cpu->iowait_boost;
1629         cpu->iowait_boost >>= 1;
1630
1631         if (busy_frac < boost)
1632                 busy_frac = boost;
1633
1634         sample->busy_scaled = busy_frac * 100;
1635
1636         target = global.no_turbo || global.turbo_disabled ?
1637                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1638         target += target >> 2;
1639         target = mul_fp(target, busy_frac);
1640         if (target < cpu->pstate.min_pstate)
1641                 target = cpu->pstate.min_pstate;
1642
1643         /*
1644          * If the average P-state during the previous cycle was higher than the
1645          * current target, add 50% of the difference to the target to reduce
1646          * possible performance oscillations and offset possible performance
1647          * loss related to moving the workload from one CPU to another within
1648          * a package/module.
1649          */
1650         avg_pstate = get_avg_pstate(cpu);
1651         if (avg_pstate > target)
1652                 target += (avg_pstate - target) >> 1;
1653
1654         return target;
1655 }
1656
1657 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1658 {
1659         int32_t perf_scaled, max_pstate, current_pstate, sample_ratio;
1660         u64 duration_ns;
1661
1662         if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE)
1663                 return cpu->pstate.turbo_pstate;
1664
1665         /*
1666          * perf_scaled is the ratio of the average P-state during the last
1667          * sampling period to the P-state requested last time (in percent).
1668          *
1669          * That measures the system's response to the previous P-state
1670          * selection.
1671          */
1672         max_pstate = cpu->pstate.max_pstate_physical;
1673         current_pstate = cpu->pstate.current_pstate;
1674         perf_scaled = mul_ext_fp(cpu->sample.core_avg_perf,
1675                                div_fp(100 * max_pstate, current_pstate));
1676
1677         /*
1678          * Since our utilization update callback will not run unless we are
1679          * in C0, check if the actual elapsed time is significantly greater (3x)
1680          * than our sample interval.  If it is, then we were idle for a long
1681          * enough period of time to adjust our performance metric.
1682          */
1683         duration_ns = cpu->sample.time - cpu->last_sample_time;
1684         if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1685                 sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1686                 perf_scaled = mul_fp(perf_scaled, sample_ratio);
1687         } else {
1688                 sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1689                 if (sample_ratio < int_tofp(1))
1690                         perf_scaled = 0;
1691         }
1692
1693         cpu->sample.busy_scaled = perf_scaled;
1694         return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
1695 }
1696
1697 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1698 {
1699         int max_pstate = intel_pstate_get_base_pstate(cpu);
1700         int min_pstate;
1701
1702         min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
1703         max_pstate = max(min_pstate, cpu->max_perf_ratio);
1704         return clamp_t(int, pstate, min_pstate, max_pstate);
1705 }
1706
1707 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1708 {
1709         if (pstate == cpu->pstate.current_pstate)
1710                 return;
1711
1712         cpu->pstate.current_pstate = pstate;
1713         wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1714 }
1715
1716 static void intel_pstate_adjust_pstate(struct cpudata *cpu, int target_pstate)
1717 {
1718         int from = cpu->pstate.current_pstate;
1719         struct sample *sample;
1720
1721         update_turbo_state();
1722
1723         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1724         trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1725         intel_pstate_update_pstate(cpu, target_pstate);
1726
1727         sample = &cpu->sample;
1728         trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1729                 fp_toint(sample->busy_scaled),
1730                 from,
1731                 cpu->pstate.current_pstate,
1732                 sample->mperf,
1733                 sample->aperf,
1734                 sample->tsc,
1735                 get_avg_frequency(cpu),
1736                 fp_toint(cpu->iowait_boost * 100));
1737 }
1738
1739 static void intel_pstate_update_util_hwp(struct update_util_data *data,
1740                                          u64 time, unsigned int flags)
1741 {
1742         struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1743         u64 delta_ns = time - cpu->sample.time;
1744
1745         if ((s64)delta_ns >= INTEL_PSTATE_HWP_SAMPLING_INTERVAL)
1746                 intel_pstate_sample(cpu, time);
1747 }
1748
1749 static void intel_pstate_update_util_pid(struct update_util_data *data,
1750                                          u64 time, unsigned int flags)
1751 {
1752         struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1753         u64 delta_ns = time - cpu->sample.time;
1754
1755         if ((s64)delta_ns < pid_params.sample_rate_ns)
1756                 return;
1757
1758         if (intel_pstate_sample(cpu, time)) {
1759                 int target_pstate;
1760
1761                 target_pstate = get_target_pstate_use_performance(cpu);
1762                 intel_pstate_adjust_pstate(cpu, target_pstate);
1763         }
1764 }
1765
1766 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1767                                      unsigned int flags)
1768 {
1769         struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1770         u64 delta_ns;
1771
1772         if (flags & SCHED_CPUFREQ_IOWAIT) {
1773                 cpu->iowait_boost = int_tofp(1);
1774         } else if (cpu->iowait_boost) {
1775                 /* Clear iowait_boost if the CPU may have been idle. */
1776                 delta_ns = time - cpu->last_update;
1777                 if (delta_ns > TICK_NSEC)
1778                         cpu->iowait_boost = 0;
1779         }
1780         cpu->last_update = time;
1781         delta_ns = time - cpu->sample.time;
1782         if ((s64)delta_ns < INTEL_PSTATE_DEFAULT_SAMPLING_INTERVAL)
1783                 return;
1784
1785         if (intel_pstate_sample(cpu, time)) {
1786                 int target_pstate;
1787
1788                 target_pstate = get_target_pstate_use_cpu_load(cpu);
1789                 intel_pstate_adjust_pstate(cpu, target_pstate);
1790         }
1791 }
1792
1793 static struct pstate_funcs core_funcs = {
1794         .get_max = core_get_max_pstate,
1795         .get_max_physical = core_get_max_pstate_physical,
1796         .get_min = core_get_min_pstate,
1797         .get_turbo = core_get_turbo_pstate,
1798         .get_scaling = core_get_scaling,
1799         .get_val = core_get_val,
1800         .update_util = intel_pstate_update_util_pid,
1801 };
1802
1803 static const struct pstate_funcs silvermont_funcs = {
1804         .get_max = atom_get_max_pstate,
1805         .get_max_physical = atom_get_max_pstate,
1806         .get_min = atom_get_min_pstate,
1807         .get_turbo = atom_get_turbo_pstate,
1808         .get_val = atom_get_val,
1809         .get_scaling = silvermont_get_scaling,
1810         .get_vid = atom_get_vid,
1811         .update_util = intel_pstate_update_util,
1812 };
1813
1814 static const struct pstate_funcs airmont_funcs = {
1815         .get_max = atom_get_max_pstate,
1816         .get_max_physical = atom_get_max_pstate,
1817         .get_min = atom_get_min_pstate,
1818         .get_turbo = atom_get_turbo_pstate,
1819         .get_val = atom_get_val,
1820         .get_scaling = airmont_get_scaling,
1821         .get_vid = atom_get_vid,
1822         .update_util = intel_pstate_update_util,
1823 };
1824
1825 static const struct pstate_funcs knl_funcs = {
1826         .get_max = core_get_max_pstate,
1827         .get_max_physical = core_get_max_pstate_physical,
1828         .get_min = core_get_min_pstate,
1829         .get_turbo = knl_get_turbo_pstate,
1830         .get_scaling = core_get_scaling,
1831         .get_val = core_get_val,
1832         .update_util = intel_pstate_update_util_pid,
1833 };
1834
1835 static const struct pstate_funcs bxt_funcs = {
1836         .get_max = core_get_max_pstate,
1837         .get_max_physical = core_get_max_pstate_physical,
1838         .get_min = core_get_min_pstate,
1839         .get_turbo = core_get_turbo_pstate,
1840         .get_scaling = core_get_scaling,
1841         .get_val = core_get_val,
1842         .update_util = intel_pstate_update_util,
1843 };
1844
1845 #define ICPU(model, policy) \
1846         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1847                         (unsigned long)&policy }
1848
1849 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1850         ICPU(INTEL_FAM6_SANDYBRIDGE,            core_funcs),
1851         ICPU(INTEL_FAM6_SANDYBRIDGE_X,          core_funcs),
1852         ICPU(INTEL_FAM6_ATOM_SILVERMONT1,       silvermont_funcs),
1853         ICPU(INTEL_FAM6_IVYBRIDGE,              core_funcs),
1854         ICPU(INTEL_FAM6_HASWELL_CORE,           core_funcs),
1855         ICPU(INTEL_FAM6_BROADWELL_CORE,         core_funcs),
1856         ICPU(INTEL_FAM6_IVYBRIDGE_X,            core_funcs),
1857         ICPU(INTEL_FAM6_HASWELL_X,              core_funcs),
1858         ICPU(INTEL_FAM6_HASWELL_ULT,            core_funcs),
1859         ICPU(INTEL_FAM6_HASWELL_GT3E,           core_funcs),
1860         ICPU(INTEL_FAM6_BROADWELL_GT3E,         core_funcs),
1861         ICPU(INTEL_FAM6_ATOM_AIRMONT,           airmont_funcs),
1862         ICPU(INTEL_FAM6_SKYLAKE_MOBILE,         core_funcs),
1863         ICPU(INTEL_FAM6_BROADWELL_X,            core_funcs),
1864         ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,        core_funcs),
1865         ICPU(INTEL_FAM6_BROADWELL_XEON_D,       core_funcs),
1866         ICPU(INTEL_FAM6_XEON_PHI_KNL,           knl_funcs),
1867         ICPU(INTEL_FAM6_XEON_PHI_KNM,           knl_funcs),
1868         ICPU(INTEL_FAM6_ATOM_GOLDMONT,          bxt_funcs),
1869         ICPU(INTEL_FAM6_ATOM_GEMINI_LAKE,       bxt_funcs),
1870         {}
1871 };
1872 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1873
1874 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1875         ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_funcs),
1876         ICPU(INTEL_FAM6_BROADWELL_X, core_funcs),
1877         ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1878         {}
1879 };
1880
1881 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1882         ICPU(INTEL_FAM6_KABYLAKE_DESKTOP, core_funcs),
1883         {}
1884 };
1885
1886 static bool pid_in_use(void);
1887
1888 static int intel_pstate_init_cpu(unsigned int cpunum)
1889 {
1890         struct cpudata *cpu;
1891
1892         cpu = all_cpu_data[cpunum];
1893
1894         if (!cpu) {
1895                 cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
1896                 if (!cpu)
1897                         return -ENOMEM;
1898
1899                 all_cpu_data[cpunum] = cpu;
1900
1901                 cpu->epp_default = -EINVAL;
1902                 cpu->epp_powersave = -EINVAL;
1903                 cpu->epp_saved = -EINVAL;
1904         }
1905
1906         cpu = all_cpu_data[cpunum];
1907
1908         cpu->cpu = cpunum;
1909
1910         if (hwp_active) {
1911                 const struct x86_cpu_id *id;
1912
1913                 id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
1914                 if (id)
1915                         intel_pstate_disable_ee(cpunum);
1916
1917                 intel_pstate_hwp_enable(cpu);
1918         } else if (pid_in_use()) {
1919                 intel_pstate_pid_reset(cpu);
1920         }
1921
1922         intel_pstate_get_cpu_pstates(cpu);
1923
1924         pr_debug("controlling: cpu %d\n", cpunum);
1925
1926         return 0;
1927 }
1928
1929 static unsigned int intel_pstate_get(unsigned int cpu_num)
1930 {
1931         struct cpudata *cpu = all_cpu_data[cpu_num];
1932
1933         return cpu ? get_avg_frequency(cpu) : 0;
1934 }
1935
1936 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1937 {
1938         struct cpudata *cpu = all_cpu_data[cpu_num];
1939
1940         if (cpu->update_util_set)
1941                 return;
1942
1943         /* Prevent intel_pstate_update_util() from using stale data. */
1944         cpu->sample.time = 0;
1945         cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1946                                      pstate_funcs.update_util);
1947         cpu->update_util_set = true;
1948 }
1949
1950 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1951 {
1952         struct cpudata *cpu_data = all_cpu_data[cpu];
1953
1954         if (!cpu_data->update_util_set)
1955                 return;
1956
1957         cpufreq_remove_update_util_hook(cpu);
1958         cpu_data->update_util_set = false;
1959         synchronize_sched();
1960 }
1961
1962 static int intel_pstate_get_max_freq(struct cpudata *cpu)
1963 {
1964         return global.turbo_disabled || global.no_turbo ?
1965                         cpu->pstate.max_freq : cpu->pstate.turbo_freq;
1966 }
1967
1968 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
1969                                             struct cpudata *cpu)
1970 {
1971         int max_freq = intel_pstate_get_max_freq(cpu);
1972         int32_t max_policy_perf, min_policy_perf;
1973         int max_state, turbo_max;
1974
1975         /*
1976          * HWP needs some special consideration, because on BDX the
1977          * HWP_REQUEST uses abstract value to represent performance
1978          * rather than pure ratios.
1979          */
1980         if (hwp_active) {
1981                 intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state);
1982         } else {
1983                 max_state = intel_pstate_get_base_pstate(cpu);
1984                 turbo_max = cpu->pstate.turbo_pstate;
1985         }
1986
1987         max_policy_perf = max_state * policy->max / max_freq;
1988         if (policy->max == policy->min) {
1989                 min_policy_perf = max_policy_perf;
1990         } else {
1991                 min_policy_perf = max_state * policy->min / max_freq;
1992                 min_policy_perf = clamp_t(int32_t, min_policy_perf,
1993                                           0, max_policy_perf);
1994         }
1995
1996         pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n",
1997                  policy->cpu, max_state,
1998                  min_policy_perf, max_policy_perf);
1999
2000         /* Normalize user input to [min_perf, max_perf] */
2001         if (per_cpu_limits) {
2002                 cpu->min_perf_ratio = min_policy_perf;
2003                 cpu->max_perf_ratio = max_policy_perf;
2004         } else {
2005                 int32_t global_min, global_max;
2006
2007                 /* Global limits are in percent of the maximum turbo P-state. */
2008                 global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2009                 global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2010                 global_min = clamp_t(int32_t, global_min, 0, global_max);
2011
2012                 pr_debug("cpu:%d global_min:%d global_max:%d\n", policy->cpu,
2013                          global_min, global_max);
2014
2015                 cpu->min_perf_ratio = max(min_policy_perf, global_min);
2016                 cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2017                 cpu->max_perf_ratio = min(max_policy_perf, global_max);
2018                 cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2019
2020                 /* Make sure min_perf <= max_perf */
2021                 cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2022                                           cpu->max_perf_ratio);
2023
2024         }
2025         pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", policy->cpu,
2026                  cpu->max_perf_ratio,
2027                  cpu->min_perf_ratio);
2028 }
2029
2030 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2031 {
2032         struct cpudata *cpu;
2033
2034         if (!policy->cpuinfo.max_freq)
2035                 return -ENODEV;
2036
2037         pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2038                  policy->cpuinfo.max_freq, policy->max);
2039
2040         cpu = all_cpu_data[policy->cpu];
2041         cpu->policy = policy->policy;
2042
2043         mutex_lock(&intel_pstate_limits_lock);
2044
2045         intel_pstate_update_perf_limits(policy, cpu);
2046
2047         if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2048                 /*
2049                  * NOHZ_FULL CPUs need this as the governor callback may not
2050                  * be invoked on them.
2051                  */
2052                 intel_pstate_clear_update_util_hook(policy->cpu);
2053                 intel_pstate_max_within_limits(cpu);
2054         }
2055
2056         intel_pstate_set_update_util_hook(policy->cpu);
2057
2058         if (hwp_active)
2059                 intel_pstate_hwp_set(policy->cpu);
2060
2061         mutex_unlock(&intel_pstate_limits_lock);
2062
2063         return 0;
2064 }
2065
2066 static void intel_pstate_adjust_policy_max(struct cpufreq_policy *policy,
2067                                          struct cpudata *cpu)
2068 {
2069         if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2070             policy->max < policy->cpuinfo.max_freq &&
2071             policy->max > cpu->pstate.max_freq) {
2072                 pr_debug("policy->max > max non turbo frequency\n");
2073                 policy->max = policy->cpuinfo.max_freq;
2074         }
2075 }
2076
2077 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2078 {
2079         struct cpudata *cpu = all_cpu_data[policy->cpu];
2080
2081         update_turbo_state();
2082         cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2083                                      intel_pstate_get_max_freq(cpu));
2084
2085         if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2086             policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2087                 return -EINVAL;
2088
2089         intel_pstate_adjust_policy_max(policy, cpu);
2090
2091         return 0;
2092 }
2093
2094 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2095 {
2096         intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2097 }
2098
2099 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2100 {
2101         pr_debug("CPU %d exiting\n", policy->cpu);
2102
2103         intel_pstate_clear_update_util_hook(policy->cpu);
2104         if (hwp_active)
2105                 intel_pstate_hwp_save_state(policy);
2106         else
2107                 intel_cpufreq_stop_cpu(policy);
2108 }
2109
2110 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2111 {
2112         intel_pstate_exit_perf_limits(policy);
2113
2114         policy->fast_switch_possible = false;
2115
2116         return 0;
2117 }
2118
2119 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2120 {
2121         struct cpudata *cpu;
2122         int rc;
2123
2124         rc = intel_pstate_init_cpu(policy->cpu);
2125         if (rc)
2126                 return rc;
2127
2128         cpu = all_cpu_data[policy->cpu];
2129
2130         cpu->max_perf_ratio = 0xFF;
2131         cpu->min_perf_ratio = 0;
2132
2133         policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2134         policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2135
2136         /* cpuinfo and default policy values */
2137         policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2138         update_turbo_state();
2139         policy->cpuinfo.max_freq = global.turbo_disabled ?
2140                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2141         policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2142
2143         intel_pstate_init_acpi_perf_limits(policy);
2144         cpumask_set_cpu(policy->cpu, policy->cpus);
2145
2146         policy->fast_switch_possible = true;
2147
2148         return 0;
2149 }
2150
2151 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2152 {
2153         int ret = __intel_pstate_cpu_init(policy);
2154
2155         if (ret)
2156                 return ret;
2157
2158         policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
2159         if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE))
2160                 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2161         else
2162                 policy->policy = CPUFREQ_POLICY_POWERSAVE;
2163
2164         return 0;
2165 }
2166
2167 static struct cpufreq_driver intel_pstate = {
2168         .flags          = CPUFREQ_CONST_LOOPS,
2169         .verify         = intel_pstate_verify_policy,
2170         .setpolicy      = intel_pstate_set_policy,
2171         .suspend        = intel_pstate_hwp_save_state,
2172         .resume         = intel_pstate_resume,
2173         .get            = intel_pstate_get,
2174         .init           = intel_pstate_cpu_init,
2175         .exit           = intel_pstate_cpu_exit,
2176         .stop_cpu       = intel_pstate_stop_cpu,
2177         .name           = "intel_pstate",
2178 };
2179
2180 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2181 {
2182         struct cpudata *cpu = all_cpu_data[policy->cpu];
2183
2184         update_turbo_state();
2185         cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2186                                      intel_pstate_get_max_freq(cpu));
2187
2188         intel_pstate_adjust_policy_max(policy, cpu);
2189
2190         intel_pstate_update_perf_limits(policy, cpu);
2191
2192         return 0;
2193 }
2194
2195 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2196                                 unsigned int target_freq,
2197                                 unsigned int relation)
2198 {
2199         struct cpudata *cpu = all_cpu_data[policy->cpu];
2200         struct cpufreq_freqs freqs;
2201         int target_pstate;
2202
2203         update_turbo_state();
2204
2205         freqs.old = policy->cur;
2206         freqs.new = target_freq;
2207
2208         cpufreq_freq_transition_begin(policy, &freqs);
2209         switch (relation) {
2210         case CPUFREQ_RELATION_L:
2211                 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2212                 break;
2213         case CPUFREQ_RELATION_H:
2214                 target_pstate = freqs.new / cpu->pstate.scaling;
2215                 break;
2216         default:
2217                 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2218                 break;
2219         }
2220         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2221         if (target_pstate != cpu->pstate.current_pstate) {
2222                 cpu->pstate.current_pstate = target_pstate;
2223                 wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2224                               pstate_funcs.get_val(cpu, target_pstate));
2225         }
2226         freqs.new = target_pstate * cpu->pstate.scaling;
2227         cpufreq_freq_transition_end(policy, &freqs, false);
2228
2229         return 0;
2230 }
2231
2232 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2233                                               unsigned int target_freq)
2234 {
2235         struct cpudata *cpu = all_cpu_data[policy->cpu];
2236         int target_pstate;
2237
2238         update_turbo_state();
2239
2240         target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2241         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2242         intel_pstate_update_pstate(cpu, target_pstate);
2243         return target_pstate * cpu->pstate.scaling;
2244 }
2245
2246 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2247 {
2248         int ret = __intel_pstate_cpu_init(policy);
2249
2250         if (ret)
2251                 return ret;
2252
2253         policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2254         policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2255         /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2256         policy->cur = policy->cpuinfo.min_freq;
2257
2258         return 0;
2259 }
2260
2261 static struct cpufreq_driver intel_cpufreq = {
2262         .flags          = CPUFREQ_CONST_LOOPS,
2263         .verify         = intel_cpufreq_verify_policy,
2264         .target         = intel_cpufreq_target,
2265         .fast_switch    = intel_cpufreq_fast_switch,
2266         .init           = intel_cpufreq_cpu_init,
2267         .exit           = intel_pstate_cpu_exit,
2268         .stop_cpu       = intel_cpufreq_stop_cpu,
2269         .name           = "intel_cpufreq",
2270 };
2271
2272 static struct cpufreq_driver *default_driver = &intel_pstate;
2273
2274 static bool pid_in_use(void)
2275 {
2276         return intel_pstate_driver == &intel_pstate &&
2277                 pstate_funcs.update_util == intel_pstate_update_util_pid;
2278 }
2279
2280 static void intel_pstate_driver_cleanup(void)
2281 {
2282         unsigned int cpu;
2283
2284         get_online_cpus();
2285         for_each_online_cpu(cpu) {
2286                 if (all_cpu_data[cpu]) {
2287                         if (intel_pstate_driver == &intel_pstate)
2288                                 intel_pstate_clear_update_util_hook(cpu);
2289
2290                         kfree(all_cpu_data[cpu]);
2291                         all_cpu_data[cpu] = NULL;
2292                 }
2293         }
2294         put_online_cpus();
2295         intel_pstate_driver = NULL;
2296 }
2297
2298 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2299 {
2300         int ret;
2301
2302         memset(&global, 0, sizeof(global));
2303         global.max_perf_pct = 100;
2304
2305         intel_pstate_driver = driver;
2306         ret = cpufreq_register_driver(intel_pstate_driver);
2307         if (ret) {
2308                 intel_pstate_driver_cleanup();
2309                 return ret;
2310         }
2311
2312         global.min_perf_pct = min_perf_pct_min();
2313
2314         if (pid_in_use())
2315                 intel_pstate_debug_expose_params();
2316
2317         return 0;
2318 }
2319
2320 static int intel_pstate_unregister_driver(void)
2321 {
2322         if (hwp_active)
2323                 return -EBUSY;
2324
2325         if (pid_in_use())
2326                 intel_pstate_debug_hide_params();
2327
2328         cpufreq_unregister_driver(intel_pstate_driver);
2329         intel_pstate_driver_cleanup();
2330
2331         return 0;
2332 }
2333
2334 static ssize_t intel_pstate_show_status(char *buf)
2335 {
2336         if (!intel_pstate_driver)
2337                 return sprintf(buf, "off\n");
2338
2339         return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2340                                         "active" : "passive");
2341 }
2342
2343 static int intel_pstate_update_status(const char *buf, size_t size)
2344 {
2345         int ret;
2346
2347         if (size == 3 && !strncmp(buf, "off", size))
2348                 return intel_pstate_driver ?
2349                         intel_pstate_unregister_driver() : -EINVAL;
2350
2351         if (size == 6 && !strncmp(buf, "active", size)) {
2352                 if (intel_pstate_driver) {
2353                         if (intel_pstate_driver == &intel_pstate)
2354                                 return 0;
2355
2356                         ret = intel_pstate_unregister_driver();
2357                         if (ret)
2358                                 return ret;
2359                 }
2360
2361                 return intel_pstate_register_driver(&intel_pstate);
2362         }
2363
2364         if (size == 7 && !strncmp(buf, "passive", size)) {
2365                 if (intel_pstate_driver) {
2366                         if (intel_pstate_driver == &intel_cpufreq)
2367                                 return 0;
2368
2369                         ret = intel_pstate_unregister_driver();
2370                         if (ret)
2371                                 return ret;
2372                 }
2373
2374                 return intel_pstate_register_driver(&intel_cpufreq);
2375         }
2376
2377         return -EINVAL;
2378 }
2379
2380 static int no_load __initdata;
2381 static int no_hwp __initdata;
2382 static int hwp_only __initdata;
2383 static unsigned int force_load __initdata;
2384
2385 static int __init intel_pstate_msrs_not_valid(void)
2386 {
2387         if (!pstate_funcs.get_max() ||
2388             !pstate_funcs.get_min() ||
2389             !pstate_funcs.get_turbo())
2390                 return -ENODEV;
2391
2392         return 0;
2393 }
2394
2395 #ifdef CONFIG_ACPI
2396 static void intel_pstate_use_acpi_profile(void)
2397 {
2398         switch (acpi_gbl_FADT.preferred_profile) {
2399         case PM_MOBILE:
2400         case PM_TABLET:
2401         case PM_APPLIANCE_PC:
2402         case PM_DESKTOP:
2403         case PM_WORKSTATION:
2404                 pstate_funcs.update_util = intel_pstate_update_util;
2405         }
2406 }
2407 #else
2408 static void intel_pstate_use_acpi_profile(void)
2409 {
2410 }
2411 #endif
2412
2413 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2414 {
2415         pstate_funcs.get_max   = funcs->get_max;
2416         pstate_funcs.get_max_physical = funcs->get_max_physical;
2417         pstate_funcs.get_min   = funcs->get_min;
2418         pstate_funcs.get_turbo = funcs->get_turbo;
2419         pstate_funcs.get_scaling = funcs->get_scaling;
2420         pstate_funcs.get_val   = funcs->get_val;
2421         pstate_funcs.get_vid   = funcs->get_vid;
2422         pstate_funcs.update_util = funcs->update_util;
2423
2424         intel_pstate_use_acpi_profile();
2425 }
2426
2427 #ifdef CONFIG_ACPI
2428
2429 static bool __init intel_pstate_no_acpi_pss(void)
2430 {
2431         int i;
2432
2433         for_each_possible_cpu(i) {
2434                 acpi_status status;
2435                 union acpi_object *pss;
2436                 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2437                 struct acpi_processor *pr = per_cpu(processors, i);
2438
2439                 if (!pr)
2440                         continue;
2441
2442                 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2443                 if (ACPI_FAILURE(status))
2444                         continue;
2445
2446                 pss = buffer.pointer;
2447                 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2448                         kfree(pss);
2449                         return false;
2450                 }
2451
2452                 kfree(pss);
2453         }
2454
2455         return true;
2456 }
2457
2458 static bool __init intel_pstate_has_acpi_ppc(void)
2459 {
2460         int i;
2461
2462         for_each_possible_cpu(i) {
2463                 struct acpi_processor *pr = per_cpu(processors, i);
2464
2465                 if (!pr)
2466                         continue;
2467                 if (acpi_has_method(pr->handle, "_PPC"))
2468                         return true;
2469         }
2470         return false;
2471 }
2472
2473 enum {
2474         PSS,
2475         PPC,
2476 };
2477
2478 struct hw_vendor_info {
2479         u16  valid;
2480         char oem_id[ACPI_OEM_ID_SIZE];
2481         char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
2482         int  oem_pwr_table;
2483 };
2484
2485 /* Hardware vendor-specific info that has its own power management modes */
2486 static struct hw_vendor_info vendor_info[] __initdata = {
2487         {1, "HP    ", "ProLiant", PSS},
2488         {1, "ORACLE", "X4-2    ", PPC},
2489         {1, "ORACLE", "X4-2L   ", PPC},
2490         {1, "ORACLE", "X4-2B   ", PPC},
2491         {1, "ORACLE", "X3-2    ", PPC},
2492         {1, "ORACLE", "X3-2L   ", PPC},
2493         {1, "ORACLE", "X3-2B   ", PPC},
2494         {1, "ORACLE", "X4470M2 ", PPC},
2495         {1, "ORACLE", "X4270M3 ", PPC},
2496         {1, "ORACLE", "X4270M2 ", PPC},
2497         {1, "ORACLE", "X4170M2 ", PPC},
2498         {1, "ORACLE", "X4170 M3", PPC},
2499         {1, "ORACLE", "X4275 M3", PPC},
2500         {1, "ORACLE", "X6-2    ", PPC},
2501         {1, "ORACLE", "Sudbury ", PPC},
2502         {0, "", ""},
2503 };
2504
2505 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2506 {
2507         struct acpi_table_header hdr;
2508         struct hw_vendor_info *v_info;
2509         const struct x86_cpu_id *id;
2510         u64 misc_pwr;
2511
2512         id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2513         if (id) {
2514                 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2515                 if ( misc_pwr & (1 << 8))
2516                         return true;
2517         }
2518
2519         if (acpi_disabled ||
2520             ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
2521                 return false;
2522
2523         for (v_info = vendor_info; v_info->valid; v_info++) {
2524                 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
2525                         !strncmp(hdr.oem_table_id, v_info->oem_table_id,
2526                                                 ACPI_OEM_TABLE_ID_SIZE))
2527                         switch (v_info->oem_pwr_table) {
2528                         case PSS:
2529                                 return intel_pstate_no_acpi_pss();
2530                         case PPC:
2531                                 return intel_pstate_has_acpi_ppc() &&
2532                                         (!force_load);
2533                         }
2534         }
2535
2536         return false;
2537 }
2538
2539 static void intel_pstate_request_control_from_smm(void)
2540 {
2541         /*
2542          * It may be unsafe to request P-states control from SMM if _PPC support
2543          * has not been enabled.
2544          */
2545         if (acpi_ppc)
2546                 acpi_processor_pstate_control();
2547 }
2548 #else /* CONFIG_ACPI not enabled */
2549 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2550 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2551 static inline void intel_pstate_request_control_from_smm(void) {}
2552 #endif /* CONFIG_ACPI */
2553
2554 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2555         { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
2556         {}
2557 };
2558
2559 static int __init intel_pstate_init(void)
2560 {
2561         int rc;
2562
2563         if (no_load)
2564                 return -ENODEV;
2565
2566         if (x86_match_cpu(hwp_support_ids)) {
2567                 copy_cpu_funcs(&core_funcs);
2568                 if (no_hwp) {
2569                         pstate_funcs.update_util = intel_pstate_update_util;
2570                 } else {
2571                         hwp_active++;
2572                         intel_pstate.attr = hwp_cpufreq_attrs;
2573                         pstate_funcs.update_util = intel_pstate_update_util_hwp;
2574                         goto hwp_cpu_matched;
2575                 }
2576         } else {
2577                 const struct x86_cpu_id *id;
2578
2579                 id = x86_match_cpu(intel_pstate_cpu_ids);
2580                 if (!id)
2581                         return -ENODEV;
2582
2583                 copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
2584         }
2585
2586         if (intel_pstate_msrs_not_valid())
2587                 return -ENODEV;
2588
2589 hwp_cpu_matched:
2590         /*
2591          * The Intel pstate driver will be ignored if the platform
2592          * firmware has its own power management modes.
2593          */
2594         if (intel_pstate_platform_pwr_mgmt_exists())
2595                 return -ENODEV;
2596
2597         if (!hwp_active && hwp_only)
2598                 return -ENOTSUPP;
2599
2600         pr_info("Intel P-state driver initializing\n");
2601
2602         all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
2603         if (!all_cpu_data)
2604                 return -ENOMEM;
2605
2606         intel_pstate_request_control_from_smm();
2607
2608         intel_pstate_sysfs_expose_params();
2609
2610         mutex_lock(&intel_pstate_driver_lock);
2611         rc = intel_pstate_register_driver(default_driver);
2612         mutex_unlock(&intel_pstate_driver_lock);
2613         if (rc)
2614                 return rc;
2615
2616         if (hwp_active)
2617                 pr_info("HWP enabled\n");
2618
2619         return 0;
2620 }
2621 device_initcall(intel_pstate_init);
2622
2623 static int __init intel_pstate_setup(char *str)
2624 {
2625         if (!str)
2626                 return -EINVAL;
2627
2628         if (!strcmp(str, "disable")) {
2629                 no_load = 1;
2630         } else if (!strcmp(str, "passive")) {
2631                 pr_info("Passive mode enabled\n");
2632                 default_driver = &intel_cpufreq;
2633                 no_hwp = 1;
2634         }
2635         if (!strcmp(str, "no_hwp")) {
2636                 pr_info("HWP disabled\n");
2637                 no_hwp = 1;
2638         }
2639         if (!strcmp(str, "force"))
2640                 force_load = 1;
2641         if (!strcmp(str, "hwp_only"))
2642                 hwp_only = 1;
2643         if (!strcmp(str, "per_cpu_perf_limits"))
2644                 per_cpu_limits = true;
2645
2646 #ifdef CONFIG_ACPI
2647         if (!strcmp(str, "support_acpi_ppc"))
2648                 acpi_ppc = true;
2649 #endif
2650
2651         return 0;
2652 }
2653 early_param("intel_pstate", intel_pstate_setup);
2654
2655 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2656 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2657 MODULE_LICENSE("GPL");