75140dd07037c72694b7d93b7eadb592e6afe28d
[sfrench/cifs-2.6.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/acpi.h>
30 #include <linux/vmalloc.h>
31 #include <trace/events/power.h>
32
33 #include <asm/div64.h>
34 #include <asm/msr.h>
35 #include <asm/cpu_device_id.h>
36 #include <asm/cpufeature.h>
37 #include <asm/intel-family.h>
38
39 #define INTEL_PSTATE_SAMPLING_INTERVAL  (10 * NSEC_PER_MSEC)
40
41 #define INTEL_CPUFREQ_TRANSITION_LATENCY        20000
42 #define INTEL_CPUFREQ_TRANSITION_DELAY          500
43
44 #ifdef CONFIG_ACPI
45 #include <acpi/processor.h>
46 #include <acpi/cppc_acpi.h>
47 #endif
48
49 #define FRAC_BITS 8
50 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
51 #define fp_toint(X) ((X) >> FRAC_BITS)
52
53 #define EXT_BITS 6
54 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
55 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
56 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
57
58 static inline int32_t mul_fp(int32_t x, int32_t y)
59 {
60         return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
61 }
62
63 static inline int32_t div_fp(s64 x, s64 y)
64 {
65         return div64_s64((int64_t)x << FRAC_BITS, y);
66 }
67
68 static inline int ceiling_fp(int32_t x)
69 {
70         int mask, ret;
71
72         ret = fp_toint(x);
73         mask = (1 << FRAC_BITS) - 1;
74         if (x & mask)
75                 ret += 1;
76         return ret;
77 }
78
79 static inline int32_t percent_fp(int percent)
80 {
81         return div_fp(percent, 100);
82 }
83
84 static inline u64 mul_ext_fp(u64 x, u64 y)
85 {
86         return (x * y) >> EXT_FRAC_BITS;
87 }
88
89 static inline u64 div_ext_fp(u64 x, u64 y)
90 {
91         return div64_u64(x << EXT_FRAC_BITS, y);
92 }
93
94 static inline int32_t percent_ext_fp(int percent)
95 {
96         return div_ext_fp(percent, 100);
97 }
98
99 /**
100  * struct sample -      Store performance sample
101  * @core_avg_perf:      Ratio of APERF/MPERF which is the actual average
102  *                      performance during last sample period
103  * @busy_scaled:        Scaled busy value which is used to calculate next
104  *                      P state. This can be different than core_avg_perf
105  *                      to account for cpu idle period
106  * @aperf:              Difference of actual performance frequency clock count
107  *                      read from APERF MSR between last and current sample
108  * @mperf:              Difference of maximum performance frequency clock count
109  *                      read from MPERF MSR between last and current sample
110  * @tsc:                Difference of time stamp counter between last and
111  *                      current sample
112  * @time:               Current time from scheduler
113  *
114  * This structure is used in the cpudata structure to store performance sample
115  * data for choosing next P State.
116  */
117 struct sample {
118         int32_t core_avg_perf;
119         int32_t busy_scaled;
120         u64 aperf;
121         u64 mperf;
122         u64 tsc;
123         u64 time;
124 };
125
126 /**
127  * struct pstate_data - Store P state data
128  * @current_pstate:     Current requested P state
129  * @min_pstate:         Min P state possible for this platform
130  * @max_pstate:         Max P state possible for this platform
131  * @max_pstate_physical:This is physical Max P state for a processor
132  *                      This can be higher than the max_pstate which can
133  *                      be limited by platform thermal design power limits
134  * @scaling:            Scaling factor to  convert frequency to cpufreq
135  *                      frequency units
136  * @turbo_pstate:       Max Turbo P state possible for this platform
137  * @max_freq:           @max_pstate frequency in cpufreq units
138  * @turbo_freq:         @turbo_pstate frequency in cpufreq units
139  *
140  * Stores the per cpu model P state limits and current P state.
141  */
142 struct pstate_data {
143         int     current_pstate;
144         int     min_pstate;
145         int     max_pstate;
146         int     max_pstate_physical;
147         int     scaling;
148         int     turbo_pstate;
149         unsigned int max_freq;
150         unsigned int turbo_freq;
151 };
152
153 /**
154  * struct vid_data -    Stores voltage information data
155  * @min:                VID data for this platform corresponding to
156  *                      the lowest P state
157  * @max:                VID data corresponding to the highest P State.
158  * @turbo:              VID data for turbo P state
159  * @ratio:              Ratio of (vid max - vid min) /
160  *                      (max P state - Min P State)
161  *
162  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
163  * This data is used in Atom platforms, where in addition to target P state,
164  * the voltage data needs to be specified to select next P State.
165  */
166 struct vid_data {
167         int min;
168         int max;
169         int turbo;
170         int32_t ratio;
171 };
172
173 /**
174  * struct global_params - Global parameters, mostly tunable via sysfs.
175  * @no_turbo:           Whether or not to use turbo P-states.
176  * @turbo_disabled:     Whethet or not turbo P-states are available at all,
177  *                      based on the MSR_IA32_MISC_ENABLE value and whether or
178  *                      not the maximum reported turbo P-state is different from
179  *                      the maximum reported non-turbo one.
180  * @min_perf_pct:       Minimum capacity limit in percent of the maximum turbo
181  *                      P-state capacity.
182  * @max_perf_pct:       Maximum capacity limit in percent of the maximum turbo
183  *                      P-state capacity.
184  */
185 struct global_params {
186         bool no_turbo;
187         bool turbo_disabled;
188         int max_perf_pct;
189         int min_perf_pct;
190 };
191
192 /**
193  * struct cpudata -     Per CPU instance data storage
194  * @cpu:                CPU number for this instance data
195  * @policy:             CPUFreq policy value
196  * @update_util:        CPUFreq utility callback information
197  * @update_util_set:    CPUFreq utility callback is set
198  * @iowait_boost:       iowait-related boost fraction
199  * @last_update:        Time of the last update.
200  * @pstate:             Stores P state limits for this CPU
201  * @vid:                Stores VID limits for this CPU
202  * @last_sample_time:   Last Sample time
203  * @aperf_mperf_shift:  Number of clock cycles after aperf, merf is incremented
204  *                      This shift is a multiplier to mperf delta to
205  *                      calculate CPU busy.
206  * @prev_aperf:         Last APERF value read from APERF MSR
207  * @prev_mperf:         Last MPERF value read from MPERF MSR
208  * @prev_tsc:           Last timestamp counter (TSC) value
209  * @prev_cummulative_iowait: IO Wait time difference from last and
210  *                      current sample
211  * @sample:             Storage for storing last Sample data
212  * @min_perf_ratio:     Minimum capacity in terms of PERF or HWP ratios
213  * @max_perf_ratio:     Maximum capacity in terms of PERF or HWP ratios
214  * @acpi_perf_data:     Stores ACPI perf information read from _PSS
215  * @valid_pss_table:    Set to true for valid ACPI _PSS entries found
216  * @epp_powersave:      Last saved HWP energy performance preference
217  *                      (EPP) or energy performance bias (EPB),
218  *                      when policy switched to performance
219  * @epp_policy:         Last saved policy used to set EPP/EPB
220  * @epp_default:        Power on default HWP energy performance
221  *                      preference/bias
222  * @epp_saved:          Saved EPP/EPB during system suspend or CPU offline
223  *                      operation
224  * @hwp_req_cached:     Cached value of the last HWP Request MSR
225  * @hwp_cap_cached:     Cached value of the last HWP Capabilities MSR
226  * @last_io_update:     Last time when IO wake flag was set
227  * @sched_flags:        Store scheduler flags for possible cross CPU update
228  * @hwp_boost_min:      Last HWP boosted min performance
229  *
230  * This structure stores per CPU instance data for all CPUs.
231  */
232 struct cpudata {
233         int cpu;
234
235         unsigned int policy;
236         struct update_util_data update_util;
237         bool   update_util_set;
238
239         struct pstate_data pstate;
240         struct vid_data vid;
241
242         u64     last_update;
243         u64     last_sample_time;
244         u64     aperf_mperf_shift;
245         u64     prev_aperf;
246         u64     prev_mperf;
247         u64     prev_tsc;
248         u64     prev_cummulative_iowait;
249         struct sample sample;
250         int32_t min_perf_ratio;
251         int32_t max_perf_ratio;
252 #ifdef CONFIG_ACPI
253         struct acpi_processor_performance acpi_perf_data;
254         bool valid_pss_table;
255 #endif
256         unsigned int iowait_boost;
257         s16 epp_powersave;
258         s16 epp_policy;
259         s16 epp_default;
260         s16 epp_saved;
261         u64 hwp_req_cached;
262         u64 hwp_cap_cached;
263         u64 last_io_update;
264         unsigned int sched_flags;
265         u32 hwp_boost_min;
266 };
267
268 static struct cpudata **all_cpu_data;
269
270 /**
271  * struct pstate_funcs - Per CPU model specific callbacks
272  * @get_max:            Callback to get maximum non turbo effective P state
273  * @get_max_physical:   Callback to get maximum non turbo physical P state
274  * @get_min:            Callback to get minimum P state
275  * @get_turbo:          Callback to get turbo P state
276  * @get_scaling:        Callback to get frequency scaling factor
277  * @get_val:            Callback to convert P state to actual MSR write value
278  * @get_vid:            Callback to get VID data for Atom platforms
279  *
280  * Core and Atom CPU models have different way to get P State limits. This
281  * structure is used to store those callbacks.
282  */
283 struct pstate_funcs {
284         int (*get_max)(void);
285         int (*get_max_physical)(void);
286         int (*get_min)(void);
287         int (*get_turbo)(void);
288         int (*get_scaling)(void);
289         int (*get_aperf_mperf_shift)(void);
290         u64 (*get_val)(struct cpudata*, int pstate);
291         void (*get_vid)(struct cpudata *);
292 };
293
294 static struct pstate_funcs pstate_funcs __read_mostly;
295
296 static int hwp_active __read_mostly;
297 static int hwp_mode_bdw __read_mostly;
298 static bool per_cpu_limits __read_mostly;
299 static bool hwp_boost __read_mostly;
300
301 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
302
303 #ifdef CONFIG_ACPI
304 static bool acpi_ppc;
305 #endif
306
307 static struct global_params global;
308
309 static DEFINE_MUTEX(intel_pstate_driver_lock);
310 static DEFINE_MUTEX(intel_pstate_limits_lock);
311
312 #ifdef CONFIG_ACPI
313
314 static bool intel_pstate_acpi_pm_profile_server(void)
315 {
316         if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
317             acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
318                 return true;
319
320         return false;
321 }
322
323 static bool intel_pstate_get_ppc_enable_status(void)
324 {
325         if (intel_pstate_acpi_pm_profile_server())
326                 return true;
327
328         return acpi_ppc;
329 }
330
331 #ifdef CONFIG_ACPI_CPPC_LIB
332
333 /* The work item is needed to avoid CPU hotplug locking issues */
334 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
335 {
336         sched_set_itmt_support();
337 }
338
339 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
340
341 static void intel_pstate_set_itmt_prio(int cpu)
342 {
343         struct cppc_perf_caps cppc_perf;
344         static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
345         int ret;
346
347         ret = cppc_get_perf_caps(cpu, &cppc_perf);
348         if (ret)
349                 return;
350
351         /*
352          * The priorities can be set regardless of whether or not
353          * sched_set_itmt_support(true) has been called and it is valid to
354          * update them at any time after it has been called.
355          */
356         sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
357
358         if (max_highest_perf <= min_highest_perf) {
359                 if (cppc_perf.highest_perf > max_highest_perf)
360                         max_highest_perf = cppc_perf.highest_perf;
361
362                 if (cppc_perf.highest_perf < min_highest_perf)
363                         min_highest_perf = cppc_perf.highest_perf;
364
365                 if (max_highest_perf > min_highest_perf) {
366                         /*
367                          * This code can be run during CPU online under the
368                          * CPU hotplug locks, so sched_set_itmt_support()
369                          * cannot be called from here.  Queue up a work item
370                          * to invoke it.
371                          */
372                         schedule_work(&sched_itmt_work);
373                 }
374         }
375 }
376 #else
377 static void intel_pstate_set_itmt_prio(int cpu)
378 {
379 }
380 #endif
381
382 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
383 {
384         struct cpudata *cpu;
385         int ret;
386         int i;
387
388         if (hwp_active) {
389                 intel_pstate_set_itmt_prio(policy->cpu);
390                 return;
391         }
392
393         if (!intel_pstate_get_ppc_enable_status())
394                 return;
395
396         cpu = all_cpu_data[policy->cpu];
397
398         ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
399                                                   policy->cpu);
400         if (ret)
401                 return;
402
403         /*
404          * Check if the control value in _PSS is for PERF_CTL MSR, which should
405          * guarantee that the states returned by it map to the states in our
406          * list directly.
407          */
408         if (cpu->acpi_perf_data.control_register.space_id !=
409                                                 ACPI_ADR_SPACE_FIXED_HARDWARE)
410                 goto err;
411
412         /*
413          * If there is only one entry _PSS, simply ignore _PSS and continue as
414          * usual without taking _PSS into account
415          */
416         if (cpu->acpi_perf_data.state_count < 2)
417                 goto err;
418
419         pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
420         for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
421                 pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
422                          (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
423                          (u32) cpu->acpi_perf_data.states[i].core_frequency,
424                          (u32) cpu->acpi_perf_data.states[i].power,
425                          (u32) cpu->acpi_perf_data.states[i].control);
426         }
427
428         /*
429          * The _PSS table doesn't contain whole turbo frequency range.
430          * This just contains +1 MHZ above the max non turbo frequency,
431          * with control value corresponding to max turbo ratio. But
432          * when cpufreq set policy is called, it will call with this
433          * max frequency, which will cause a reduced performance as
434          * this driver uses real max turbo frequency as the max
435          * frequency. So correct this frequency in _PSS table to
436          * correct max turbo frequency based on the turbo state.
437          * Also need to convert to MHz as _PSS freq is in MHz.
438          */
439         if (!global.turbo_disabled)
440                 cpu->acpi_perf_data.states[0].core_frequency =
441                                         policy->cpuinfo.max_freq / 1000;
442         cpu->valid_pss_table = true;
443         pr_debug("_PPC limits will be enforced\n");
444
445         return;
446
447  err:
448         cpu->valid_pss_table = false;
449         acpi_processor_unregister_performance(policy->cpu);
450 }
451
452 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
453 {
454         struct cpudata *cpu;
455
456         cpu = all_cpu_data[policy->cpu];
457         if (!cpu->valid_pss_table)
458                 return;
459
460         acpi_processor_unregister_performance(policy->cpu);
461 }
462 #else
463 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
464 {
465 }
466
467 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
468 {
469 }
470
471 static inline bool intel_pstate_acpi_pm_profile_server(void)
472 {
473         return false;
474 }
475 #endif
476
477 static inline void update_turbo_state(void)
478 {
479         u64 misc_en;
480         struct cpudata *cpu;
481
482         cpu = all_cpu_data[0];
483         rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
484         global.turbo_disabled =
485                 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
486                  cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
487 }
488
489 static int min_perf_pct_min(void)
490 {
491         struct cpudata *cpu = all_cpu_data[0];
492         int turbo_pstate = cpu->pstate.turbo_pstate;
493
494         return turbo_pstate ?
495                 (cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
496 }
497
498 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
499 {
500         u64 epb;
501         int ret;
502
503         if (!static_cpu_has(X86_FEATURE_EPB))
504                 return -ENXIO;
505
506         ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
507         if (ret)
508                 return (s16)ret;
509
510         return (s16)(epb & 0x0f);
511 }
512
513 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
514 {
515         s16 epp;
516
517         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
518                 /*
519                  * When hwp_req_data is 0, means that caller didn't read
520                  * MSR_HWP_REQUEST, so need to read and get EPP.
521                  */
522                 if (!hwp_req_data) {
523                         epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
524                                             &hwp_req_data);
525                         if (epp)
526                                 return epp;
527                 }
528                 epp = (hwp_req_data >> 24) & 0xff;
529         } else {
530                 /* When there is no EPP present, HWP uses EPB settings */
531                 epp = intel_pstate_get_epb(cpu_data);
532         }
533
534         return epp;
535 }
536
537 static int intel_pstate_set_epb(int cpu, s16 pref)
538 {
539         u64 epb;
540         int ret;
541
542         if (!static_cpu_has(X86_FEATURE_EPB))
543                 return -ENXIO;
544
545         ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
546         if (ret)
547                 return ret;
548
549         epb = (epb & ~0x0f) | pref;
550         wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
551
552         return 0;
553 }
554
555 /*
556  * EPP/EPB display strings corresponding to EPP index in the
557  * energy_perf_strings[]
558  *      index           String
559  *-------------------------------------
560  *      0               default
561  *      1               performance
562  *      2               balance_performance
563  *      3               balance_power
564  *      4               power
565  */
566 static const char * const energy_perf_strings[] = {
567         "default",
568         "performance",
569         "balance_performance",
570         "balance_power",
571         "power",
572         NULL
573 };
574 static const unsigned int epp_values[] = {
575         HWP_EPP_PERFORMANCE,
576         HWP_EPP_BALANCE_PERFORMANCE,
577         HWP_EPP_BALANCE_POWERSAVE,
578         HWP_EPP_POWERSAVE
579 };
580
581 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
582 {
583         s16 epp;
584         int index = -EINVAL;
585
586         epp = intel_pstate_get_epp(cpu_data, 0);
587         if (epp < 0)
588                 return epp;
589
590         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
591                 if (epp == HWP_EPP_PERFORMANCE)
592                         return 1;
593                 if (epp <= HWP_EPP_BALANCE_PERFORMANCE)
594                         return 2;
595                 if (epp <= HWP_EPP_BALANCE_POWERSAVE)
596                         return 3;
597                 else
598                         return 4;
599         } else if (static_cpu_has(X86_FEATURE_EPB)) {
600                 /*
601                  * Range:
602                  *      0x00-0x03       :       Performance
603                  *      0x04-0x07       :       Balance performance
604                  *      0x08-0x0B       :       Balance power
605                  *      0x0C-0x0F       :       Power
606                  * The EPB is a 4 bit value, but our ranges restrict the
607                  * value which can be set. Here only using top two bits
608                  * effectively.
609                  */
610                 index = (epp >> 2) + 1;
611         }
612
613         return index;
614 }
615
616 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
617                                               int pref_index)
618 {
619         int epp = -EINVAL;
620         int ret;
621
622         if (!pref_index)
623                 epp = cpu_data->epp_default;
624
625         mutex_lock(&intel_pstate_limits_lock);
626
627         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
628                 u64 value;
629
630                 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
631                 if (ret)
632                         goto return_pref;
633
634                 value &= ~GENMASK_ULL(31, 24);
635
636                 if (epp == -EINVAL)
637                         epp = epp_values[pref_index - 1];
638
639                 value |= (u64)epp << 24;
640                 ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
641         } else {
642                 if (epp == -EINVAL)
643                         epp = (pref_index - 1) << 2;
644                 ret = intel_pstate_set_epb(cpu_data->cpu, epp);
645         }
646 return_pref:
647         mutex_unlock(&intel_pstate_limits_lock);
648
649         return ret;
650 }
651
652 static ssize_t show_energy_performance_available_preferences(
653                                 struct cpufreq_policy *policy, char *buf)
654 {
655         int i = 0;
656         int ret = 0;
657
658         while (energy_perf_strings[i] != NULL)
659                 ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
660
661         ret += sprintf(&buf[ret], "\n");
662
663         return ret;
664 }
665
666 cpufreq_freq_attr_ro(energy_performance_available_preferences);
667
668 static ssize_t store_energy_performance_preference(
669                 struct cpufreq_policy *policy, const char *buf, size_t count)
670 {
671         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
672         char str_preference[21];
673         int ret;
674
675         ret = sscanf(buf, "%20s", str_preference);
676         if (ret != 1)
677                 return -EINVAL;
678
679         ret = match_string(energy_perf_strings, -1, str_preference);
680         if (ret < 0)
681                 return ret;
682
683         intel_pstate_set_energy_pref_index(cpu_data, ret);
684         return count;
685 }
686
687 static ssize_t show_energy_performance_preference(
688                                 struct cpufreq_policy *policy, char *buf)
689 {
690         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
691         int preference;
692
693         preference = intel_pstate_get_energy_pref_index(cpu_data);
694         if (preference < 0)
695                 return preference;
696
697         return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
698 }
699
700 cpufreq_freq_attr_rw(energy_performance_preference);
701
702 static struct freq_attr *hwp_cpufreq_attrs[] = {
703         &energy_performance_preference,
704         &energy_performance_available_preferences,
705         NULL,
706 };
707
708 static void intel_pstate_get_hwp_max(unsigned int cpu, int *phy_max,
709                                      int *current_max)
710 {
711         u64 cap;
712
713         rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
714         WRITE_ONCE(all_cpu_data[cpu]->hwp_cap_cached, cap);
715         if (global.no_turbo)
716                 *current_max = HWP_GUARANTEED_PERF(cap);
717         else
718                 *current_max = HWP_HIGHEST_PERF(cap);
719
720         *phy_max = HWP_HIGHEST_PERF(cap);
721 }
722
723 static void intel_pstate_hwp_set(unsigned int cpu)
724 {
725         struct cpudata *cpu_data = all_cpu_data[cpu];
726         int max, min;
727         u64 value;
728         s16 epp;
729
730         max = cpu_data->max_perf_ratio;
731         min = cpu_data->min_perf_ratio;
732
733         if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
734                 min = max;
735
736         rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
737
738         value &= ~HWP_MIN_PERF(~0L);
739         value |= HWP_MIN_PERF(min);
740
741         value &= ~HWP_MAX_PERF(~0L);
742         value |= HWP_MAX_PERF(max);
743
744         if (cpu_data->epp_policy == cpu_data->policy)
745                 goto skip_epp;
746
747         cpu_data->epp_policy = cpu_data->policy;
748
749         if (cpu_data->epp_saved >= 0) {
750                 epp = cpu_data->epp_saved;
751                 cpu_data->epp_saved = -EINVAL;
752                 goto update_epp;
753         }
754
755         if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
756                 epp = intel_pstate_get_epp(cpu_data, value);
757                 cpu_data->epp_powersave = epp;
758                 /* If EPP read was failed, then don't try to write */
759                 if (epp < 0)
760                         goto skip_epp;
761
762                 epp = 0;
763         } else {
764                 /* skip setting EPP, when saved value is invalid */
765                 if (cpu_data->epp_powersave < 0)
766                         goto skip_epp;
767
768                 /*
769                  * No need to restore EPP when it is not zero. This
770                  * means:
771                  *  - Policy is not changed
772                  *  - user has manually changed
773                  *  - Error reading EPB
774                  */
775                 epp = intel_pstate_get_epp(cpu_data, value);
776                 if (epp)
777                         goto skip_epp;
778
779                 epp = cpu_data->epp_powersave;
780         }
781 update_epp:
782         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
783                 value &= ~GENMASK_ULL(31, 24);
784                 value |= (u64)epp << 24;
785         } else {
786                 intel_pstate_set_epb(cpu, epp);
787         }
788 skip_epp:
789         WRITE_ONCE(cpu_data->hwp_req_cached, value);
790         wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
791 }
792
793 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
794 {
795         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
796
797         if (!hwp_active)
798                 return 0;
799
800         cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
801
802         return 0;
803 }
804
805 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
806
807 static int intel_pstate_resume(struct cpufreq_policy *policy)
808 {
809         if (!hwp_active)
810                 return 0;
811
812         mutex_lock(&intel_pstate_limits_lock);
813
814         if (policy->cpu == 0)
815                 intel_pstate_hwp_enable(all_cpu_data[policy->cpu]);
816
817         all_cpu_data[policy->cpu]->epp_policy = 0;
818         intel_pstate_hwp_set(policy->cpu);
819
820         mutex_unlock(&intel_pstate_limits_lock);
821
822         return 0;
823 }
824
825 static void intel_pstate_update_policies(void)
826 {
827         int cpu;
828
829         for_each_possible_cpu(cpu)
830                 cpufreq_update_policy(cpu);
831 }
832
833 /************************** sysfs begin ************************/
834 #define show_one(file_name, object)                                     \
835         static ssize_t show_##file_name                                 \
836         (struct kobject *kobj, struct attribute *attr, char *buf)       \
837         {                                                               \
838                 return sprintf(buf, "%u\n", global.object);             \
839         }
840
841 static ssize_t intel_pstate_show_status(char *buf);
842 static int intel_pstate_update_status(const char *buf, size_t size);
843
844 static ssize_t show_status(struct kobject *kobj,
845                            struct attribute *attr, char *buf)
846 {
847         ssize_t ret;
848
849         mutex_lock(&intel_pstate_driver_lock);
850         ret = intel_pstate_show_status(buf);
851         mutex_unlock(&intel_pstate_driver_lock);
852
853         return ret;
854 }
855
856 static ssize_t store_status(struct kobject *a, struct attribute *b,
857                             const char *buf, size_t count)
858 {
859         char *p = memchr(buf, '\n', count);
860         int ret;
861
862         mutex_lock(&intel_pstate_driver_lock);
863         ret = intel_pstate_update_status(buf, p ? p - buf : count);
864         mutex_unlock(&intel_pstate_driver_lock);
865
866         return ret < 0 ? ret : count;
867 }
868
869 static ssize_t show_turbo_pct(struct kobject *kobj,
870                                 struct attribute *attr, char *buf)
871 {
872         struct cpudata *cpu;
873         int total, no_turbo, turbo_pct;
874         uint32_t turbo_fp;
875
876         mutex_lock(&intel_pstate_driver_lock);
877
878         if (!intel_pstate_driver) {
879                 mutex_unlock(&intel_pstate_driver_lock);
880                 return -EAGAIN;
881         }
882
883         cpu = all_cpu_data[0];
884
885         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
886         no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
887         turbo_fp = div_fp(no_turbo, total);
888         turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
889
890         mutex_unlock(&intel_pstate_driver_lock);
891
892         return sprintf(buf, "%u\n", turbo_pct);
893 }
894
895 static ssize_t show_num_pstates(struct kobject *kobj,
896                                 struct attribute *attr, char *buf)
897 {
898         struct cpudata *cpu;
899         int total;
900
901         mutex_lock(&intel_pstate_driver_lock);
902
903         if (!intel_pstate_driver) {
904                 mutex_unlock(&intel_pstate_driver_lock);
905                 return -EAGAIN;
906         }
907
908         cpu = all_cpu_data[0];
909         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
910
911         mutex_unlock(&intel_pstate_driver_lock);
912
913         return sprintf(buf, "%u\n", total);
914 }
915
916 static ssize_t show_no_turbo(struct kobject *kobj,
917                              struct attribute *attr, char *buf)
918 {
919         ssize_t ret;
920
921         mutex_lock(&intel_pstate_driver_lock);
922
923         if (!intel_pstate_driver) {
924                 mutex_unlock(&intel_pstate_driver_lock);
925                 return -EAGAIN;
926         }
927
928         update_turbo_state();
929         if (global.turbo_disabled)
930                 ret = sprintf(buf, "%u\n", global.turbo_disabled);
931         else
932                 ret = sprintf(buf, "%u\n", global.no_turbo);
933
934         mutex_unlock(&intel_pstate_driver_lock);
935
936         return ret;
937 }
938
939 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
940                               const char *buf, size_t count)
941 {
942         unsigned int input;
943         int ret;
944
945         ret = sscanf(buf, "%u", &input);
946         if (ret != 1)
947                 return -EINVAL;
948
949         mutex_lock(&intel_pstate_driver_lock);
950
951         if (!intel_pstate_driver) {
952                 mutex_unlock(&intel_pstate_driver_lock);
953                 return -EAGAIN;
954         }
955
956         mutex_lock(&intel_pstate_limits_lock);
957
958         update_turbo_state();
959         if (global.turbo_disabled) {
960                 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
961                 mutex_unlock(&intel_pstate_limits_lock);
962                 mutex_unlock(&intel_pstate_driver_lock);
963                 return -EPERM;
964         }
965
966         global.no_turbo = clamp_t(int, input, 0, 1);
967
968         if (global.no_turbo) {
969                 struct cpudata *cpu = all_cpu_data[0];
970                 int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
971
972                 /* Squash the global minimum into the permitted range. */
973                 if (global.min_perf_pct > pct)
974                         global.min_perf_pct = pct;
975         }
976
977         mutex_unlock(&intel_pstate_limits_lock);
978
979         intel_pstate_update_policies();
980
981         mutex_unlock(&intel_pstate_driver_lock);
982
983         return count;
984 }
985
986 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
987                                   const char *buf, size_t count)
988 {
989         unsigned int input;
990         int ret;
991
992         ret = sscanf(buf, "%u", &input);
993         if (ret != 1)
994                 return -EINVAL;
995
996         mutex_lock(&intel_pstate_driver_lock);
997
998         if (!intel_pstate_driver) {
999                 mutex_unlock(&intel_pstate_driver_lock);
1000                 return -EAGAIN;
1001         }
1002
1003         mutex_lock(&intel_pstate_limits_lock);
1004
1005         global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1006
1007         mutex_unlock(&intel_pstate_limits_lock);
1008
1009         intel_pstate_update_policies();
1010
1011         mutex_unlock(&intel_pstate_driver_lock);
1012
1013         return count;
1014 }
1015
1016 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
1017                                   const char *buf, size_t count)
1018 {
1019         unsigned int input;
1020         int ret;
1021
1022         ret = sscanf(buf, "%u", &input);
1023         if (ret != 1)
1024                 return -EINVAL;
1025
1026         mutex_lock(&intel_pstate_driver_lock);
1027
1028         if (!intel_pstate_driver) {
1029                 mutex_unlock(&intel_pstate_driver_lock);
1030                 return -EAGAIN;
1031         }
1032
1033         mutex_lock(&intel_pstate_limits_lock);
1034
1035         global.min_perf_pct = clamp_t(int, input,
1036                                       min_perf_pct_min(), global.max_perf_pct);
1037
1038         mutex_unlock(&intel_pstate_limits_lock);
1039
1040         intel_pstate_update_policies();
1041
1042         mutex_unlock(&intel_pstate_driver_lock);
1043
1044         return count;
1045 }
1046
1047 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1048                                 struct attribute *attr, char *buf)
1049 {
1050         return sprintf(buf, "%u\n", hwp_boost);
1051 }
1052
1053 static ssize_t store_hwp_dynamic_boost(struct kobject *a, struct attribute *b,
1054                                        const char *buf, size_t count)
1055 {
1056         unsigned int input;
1057         int ret;
1058
1059         ret = kstrtouint(buf, 10, &input);
1060         if (ret)
1061                 return ret;
1062
1063         mutex_lock(&intel_pstate_driver_lock);
1064         hwp_boost = !!input;
1065         intel_pstate_update_policies();
1066         mutex_unlock(&intel_pstate_driver_lock);
1067
1068         return count;
1069 }
1070
1071 show_one(max_perf_pct, max_perf_pct);
1072 show_one(min_perf_pct, min_perf_pct);
1073
1074 define_one_global_rw(status);
1075 define_one_global_rw(no_turbo);
1076 define_one_global_rw(max_perf_pct);
1077 define_one_global_rw(min_perf_pct);
1078 define_one_global_ro(turbo_pct);
1079 define_one_global_ro(num_pstates);
1080 define_one_global_rw(hwp_dynamic_boost);
1081
1082 static struct attribute *intel_pstate_attributes[] = {
1083         &status.attr,
1084         &no_turbo.attr,
1085         &turbo_pct.attr,
1086         &num_pstates.attr,
1087         NULL
1088 };
1089
1090 static const struct attribute_group intel_pstate_attr_group = {
1091         .attrs = intel_pstate_attributes,
1092 };
1093
1094 static void __init intel_pstate_sysfs_expose_params(void)
1095 {
1096         struct kobject *intel_pstate_kobject;
1097         int rc;
1098
1099         intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1100                                                 &cpu_subsys.dev_root->kobj);
1101         if (WARN_ON(!intel_pstate_kobject))
1102                 return;
1103
1104         rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1105         if (WARN_ON(rc))
1106                 return;
1107
1108         /*
1109          * If per cpu limits are enforced there are no global limits, so
1110          * return without creating max/min_perf_pct attributes
1111          */
1112         if (per_cpu_limits)
1113                 return;
1114
1115         rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1116         WARN_ON(rc);
1117
1118         rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1119         WARN_ON(rc);
1120
1121         if (hwp_active) {
1122                 rc = sysfs_create_file(intel_pstate_kobject,
1123                                        &hwp_dynamic_boost.attr);
1124                 WARN_ON(rc);
1125         }
1126 }
1127 /************************** sysfs end ************************/
1128
1129 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1130 {
1131         /* First disable HWP notification interrupt as we don't process them */
1132         if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1133                 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1134
1135         wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1136         cpudata->epp_policy = 0;
1137         if (cpudata->epp_default == -EINVAL)
1138                 cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1139 }
1140
1141 #define MSR_IA32_POWER_CTL_BIT_EE       19
1142
1143 /* Disable energy efficiency optimization */
1144 static void intel_pstate_disable_ee(int cpu)
1145 {
1146         u64 power_ctl;
1147         int ret;
1148
1149         ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1150         if (ret)
1151                 return;
1152
1153         if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1154                 pr_info("Disabling energy efficiency optimization\n");
1155                 power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1156                 wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1157         }
1158 }
1159
1160 static int atom_get_min_pstate(void)
1161 {
1162         u64 value;
1163
1164         rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1165         return (value >> 8) & 0x7F;
1166 }
1167
1168 static int atom_get_max_pstate(void)
1169 {
1170         u64 value;
1171
1172         rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1173         return (value >> 16) & 0x7F;
1174 }
1175
1176 static int atom_get_turbo_pstate(void)
1177 {
1178         u64 value;
1179
1180         rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1181         return value & 0x7F;
1182 }
1183
1184 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1185 {
1186         u64 val;
1187         int32_t vid_fp;
1188         u32 vid;
1189
1190         val = (u64)pstate << 8;
1191         if (global.no_turbo && !global.turbo_disabled)
1192                 val |= (u64)1 << 32;
1193
1194         vid_fp = cpudata->vid.min + mul_fp(
1195                 int_tofp(pstate - cpudata->pstate.min_pstate),
1196                 cpudata->vid.ratio);
1197
1198         vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1199         vid = ceiling_fp(vid_fp);
1200
1201         if (pstate > cpudata->pstate.max_pstate)
1202                 vid = cpudata->vid.turbo;
1203
1204         return val | vid;
1205 }
1206
1207 static int silvermont_get_scaling(void)
1208 {
1209         u64 value;
1210         int i;
1211         /* Defined in Table 35-6 from SDM (Sept 2015) */
1212         static int silvermont_freq_table[] = {
1213                 83300, 100000, 133300, 116700, 80000};
1214
1215         rdmsrl(MSR_FSB_FREQ, value);
1216         i = value & 0x7;
1217         WARN_ON(i > 4);
1218
1219         return silvermont_freq_table[i];
1220 }
1221
1222 static int airmont_get_scaling(void)
1223 {
1224         u64 value;
1225         int i;
1226         /* Defined in Table 35-10 from SDM (Sept 2015) */
1227         static int airmont_freq_table[] = {
1228                 83300, 100000, 133300, 116700, 80000,
1229                 93300, 90000, 88900, 87500};
1230
1231         rdmsrl(MSR_FSB_FREQ, value);
1232         i = value & 0xF;
1233         WARN_ON(i > 8);
1234
1235         return airmont_freq_table[i];
1236 }
1237
1238 static void atom_get_vid(struct cpudata *cpudata)
1239 {
1240         u64 value;
1241
1242         rdmsrl(MSR_ATOM_CORE_VIDS, value);
1243         cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1244         cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1245         cpudata->vid.ratio = div_fp(
1246                 cpudata->vid.max - cpudata->vid.min,
1247                 int_tofp(cpudata->pstate.max_pstate -
1248                         cpudata->pstate.min_pstate));
1249
1250         rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1251         cpudata->vid.turbo = value & 0x7f;
1252 }
1253
1254 static int core_get_min_pstate(void)
1255 {
1256         u64 value;
1257
1258         rdmsrl(MSR_PLATFORM_INFO, value);
1259         return (value >> 40) & 0xFF;
1260 }
1261
1262 static int core_get_max_pstate_physical(void)
1263 {
1264         u64 value;
1265
1266         rdmsrl(MSR_PLATFORM_INFO, value);
1267         return (value >> 8) & 0xFF;
1268 }
1269
1270 static int core_get_tdp_ratio(u64 plat_info)
1271 {
1272         /* Check how many TDP levels present */
1273         if (plat_info & 0x600000000) {
1274                 u64 tdp_ctrl;
1275                 u64 tdp_ratio;
1276                 int tdp_msr;
1277                 int err;
1278
1279                 /* Get the TDP level (0, 1, 2) to get ratios */
1280                 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1281                 if (err)
1282                         return err;
1283
1284                 /* TDP MSR are continuous starting at 0x648 */
1285                 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1286                 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1287                 if (err)
1288                         return err;
1289
1290                 /* For level 1 and 2, bits[23:16] contain the ratio */
1291                 if (tdp_ctrl & 0x03)
1292                         tdp_ratio >>= 16;
1293
1294                 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1295                 pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1296
1297                 return (int)tdp_ratio;
1298         }
1299
1300         return -ENXIO;
1301 }
1302
1303 static int core_get_max_pstate(void)
1304 {
1305         u64 tar;
1306         u64 plat_info;
1307         int max_pstate;
1308         int tdp_ratio;
1309         int err;
1310
1311         rdmsrl(MSR_PLATFORM_INFO, plat_info);
1312         max_pstate = (plat_info >> 8) & 0xFF;
1313
1314         tdp_ratio = core_get_tdp_ratio(plat_info);
1315         if (tdp_ratio <= 0)
1316                 return max_pstate;
1317
1318         if (hwp_active) {
1319                 /* Turbo activation ratio is not used on HWP platforms */
1320                 return tdp_ratio;
1321         }
1322
1323         err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1324         if (!err) {
1325                 int tar_levels;
1326
1327                 /* Do some sanity checking for safety */
1328                 tar_levels = tar & 0xff;
1329                 if (tdp_ratio - 1 == tar_levels) {
1330                         max_pstate = tar_levels;
1331                         pr_debug("max_pstate=TAC %x\n", max_pstate);
1332                 }
1333         }
1334
1335         return max_pstate;
1336 }
1337
1338 static int core_get_turbo_pstate(void)
1339 {
1340         u64 value;
1341         int nont, ret;
1342
1343         rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1344         nont = core_get_max_pstate();
1345         ret = (value) & 255;
1346         if (ret <= nont)
1347                 ret = nont;
1348         return ret;
1349 }
1350
1351 static inline int core_get_scaling(void)
1352 {
1353         return 100000;
1354 }
1355
1356 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1357 {
1358         u64 val;
1359
1360         val = (u64)pstate << 8;
1361         if (global.no_turbo && !global.turbo_disabled)
1362                 val |= (u64)1 << 32;
1363
1364         return val;
1365 }
1366
1367 static int knl_get_aperf_mperf_shift(void)
1368 {
1369         return 10;
1370 }
1371
1372 static int knl_get_turbo_pstate(void)
1373 {
1374         u64 value;
1375         int nont, ret;
1376
1377         rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1378         nont = core_get_max_pstate();
1379         ret = (((value) >> 8) & 0xFF);
1380         if (ret <= nont)
1381                 ret = nont;
1382         return ret;
1383 }
1384
1385 static int intel_pstate_get_base_pstate(struct cpudata *cpu)
1386 {
1387         return global.no_turbo || global.turbo_disabled ?
1388                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1389 }
1390
1391 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1392 {
1393         trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1394         cpu->pstate.current_pstate = pstate;
1395         /*
1396          * Generally, there is no guarantee that this code will always run on
1397          * the CPU being updated, so force the register update to run on the
1398          * right CPU.
1399          */
1400         wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1401                       pstate_funcs.get_val(cpu, pstate));
1402 }
1403
1404 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1405 {
1406         intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1407 }
1408
1409 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1410 {
1411         int pstate;
1412
1413         update_turbo_state();
1414         pstate = intel_pstate_get_base_pstate(cpu);
1415         pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1416         intel_pstate_set_pstate(cpu, pstate);
1417 }
1418
1419 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1420 {
1421         cpu->pstate.min_pstate = pstate_funcs.get_min();
1422         cpu->pstate.max_pstate = pstate_funcs.get_max();
1423         cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1424         cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1425         cpu->pstate.scaling = pstate_funcs.get_scaling();
1426         cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1427
1428         if (hwp_active && !hwp_mode_bdw) {
1429                 unsigned int phy_max, current_max;
1430
1431                 intel_pstate_get_hwp_max(cpu->cpu, &phy_max, &current_max);
1432                 cpu->pstate.turbo_freq = phy_max * cpu->pstate.scaling;
1433         } else {
1434                 cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1435         }
1436
1437         if (pstate_funcs.get_aperf_mperf_shift)
1438                 cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
1439
1440         if (pstate_funcs.get_vid)
1441                 pstate_funcs.get_vid(cpu);
1442
1443         intel_pstate_set_min_pstate(cpu);
1444 }
1445
1446 /*
1447  * Long hold time will keep high perf limits for long time,
1448  * which negatively impacts perf/watt for some workloads,
1449  * like specpower. 3ms is based on experiements on some
1450  * workoads.
1451  */
1452 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
1453
1454 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
1455 {
1456         u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
1457         u32 max_limit = (hwp_req & 0xff00) >> 8;
1458         u32 min_limit = (hwp_req & 0xff);
1459         u32 boost_level1;
1460
1461         /*
1462          * Cases to consider (User changes via sysfs or boot time):
1463          * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
1464          *      No boost, return.
1465          * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
1466          *     Should result in one level boost only for P0.
1467          * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
1468          *     Should result in two level boost:
1469          *         (min + p1)/2 and P1.
1470          * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
1471          *     Should result in three level boost:
1472          *        (min + p1)/2, P1 and P0.
1473          */
1474
1475         /* If max and min are equal or already at max, nothing to boost */
1476         if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
1477                 return;
1478
1479         if (!cpu->hwp_boost_min)
1480                 cpu->hwp_boost_min = min_limit;
1481
1482         /* level at half way mark between min and guranteed */
1483         boost_level1 = (HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) + min_limit) >> 1;
1484
1485         if (cpu->hwp_boost_min < boost_level1)
1486                 cpu->hwp_boost_min = boost_level1;
1487         else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1488                 cpu->hwp_boost_min = HWP_GUARANTEED_PERF(cpu->hwp_cap_cached);
1489         else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) &&
1490                  max_limit != HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1491                 cpu->hwp_boost_min = max_limit;
1492         else
1493                 return;
1494
1495         hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
1496         wrmsrl(MSR_HWP_REQUEST, hwp_req);
1497         cpu->last_update = cpu->sample.time;
1498 }
1499
1500 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
1501 {
1502         if (cpu->hwp_boost_min) {
1503                 bool expired;
1504
1505                 /* Check if we are idle for hold time to boost down */
1506                 expired = time_after64(cpu->sample.time, cpu->last_update +
1507                                        hwp_boost_hold_time_ns);
1508                 if (expired) {
1509                         wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
1510                         cpu->hwp_boost_min = 0;
1511                 }
1512         }
1513         cpu->last_update = cpu->sample.time;
1514 }
1515
1516 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
1517                                                       u64 time)
1518 {
1519         cpu->sample.time = time;
1520
1521         if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
1522                 bool do_io = false;
1523
1524                 cpu->sched_flags = 0;
1525                 /*
1526                  * Set iowait_boost flag and update time. Since IO WAIT flag
1527                  * is set all the time, we can't just conclude that there is
1528                  * some IO bound activity is scheduled on this CPU with just
1529                  * one occurrence. If we receive at least two in two
1530                  * consecutive ticks, then we treat as boost candidate.
1531                  */
1532                 if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
1533                         do_io = true;
1534
1535                 cpu->last_io_update = time;
1536
1537                 if (do_io)
1538                         intel_pstate_hwp_boost_up(cpu);
1539
1540         } else {
1541                 intel_pstate_hwp_boost_down(cpu);
1542         }
1543 }
1544
1545 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
1546                                                 u64 time, unsigned int flags)
1547 {
1548         struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1549
1550         cpu->sched_flags |= flags;
1551
1552         if (smp_processor_id() == cpu->cpu)
1553                 intel_pstate_update_util_hwp_local(cpu, time);
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, cpu_khz);
1608 }
1609
1610 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1611 {
1612         return mul_ext_fp(cpu->pstate.max_pstate_physical,
1613                           cpu->sample.core_avg_perf);
1614 }
1615
1616 static inline int32_t get_target_pstate(struct cpudata *cpu)
1617 {
1618         struct sample *sample = &cpu->sample;
1619         int32_t busy_frac, boost;
1620         int target, avg_pstate;
1621
1622         busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
1623                            sample->tsc);
1624
1625         boost = cpu->iowait_boost;
1626         cpu->iowait_boost >>= 1;
1627
1628         if (busy_frac < boost)
1629                 busy_frac = boost;
1630
1631         sample->busy_scaled = busy_frac * 100;
1632
1633         target = global.no_turbo || global.turbo_disabled ?
1634                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1635         target += target >> 2;
1636         target = mul_fp(target, busy_frac);
1637         if (target < cpu->pstate.min_pstate)
1638                 target = cpu->pstate.min_pstate;
1639
1640         /*
1641          * If the average P-state during the previous cycle was higher than the
1642          * current target, add 50% of the difference to the target to reduce
1643          * possible performance oscillations and offset possible performance
1644          * loss related to moving the workload from one CPU to another within
1645          * a package/module.
1646          */
1647         avg_pstate = get_avg_pstate(cpu);
1648         if (avg_pstate > target)
1649                 target += (avg_pstate - target) >> 1;
1650
1651         return target;
1652 }
1653
1654 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1655 {
1656         int max_pstate = intel_pstate_get_base_pstate(cpu);
1657         int min_pstate;
1658
1659         min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
1660         max_pstate = max(min_pstate, cpu->max_perf_ratio);
1661         return clamp_t(int, pstate, min_pstate, max_pstate);
1662 }
1663
1664 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1665 {
1666         if (pstate == cpu->pstate.current_pstate)
1667                 return;
1668
1669         cpu->pstate.current_pstate = pstate;
1670         wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1671 }
1672
1673 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
1674 {
1675         int from = cpu->pstate.current_pstate;
1676         struct sample *sample;
1677         int target_pstate;
1678
1679         update_turbo_state();
1680
1681         target_pstate = get_target_pstate(cpu);
1682         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1683         trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1684         intel_pstate_update_pstate(cpu, target_pstate);
1685
1686         sample = &cpu->sample;
1687         trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1688                 fp_toint(sample->busy_scaled),
1689                 from,
1690                 cpu->pstate.current_pstate,
1691                 sample->mperf,
1692                 sample->aperf,
1693                 sample->tsc,
1694                 get_avg_frequency(cpu),
1695                 fp_toint(cpu->iowait_boost * 100));
1696 }
1697
1698 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1699                                      unsigned int flags)
1700 {
1701         struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1702         u64 delta_ns;
1703
1704         /* Don't allow remote callbacks */
1705         if (smp_processor_id() != cpu->cpu)
1706                 return;
1707
1708         if (flags & SCHED_CPUFREQ_IOWAIT) {
1709                 cpu->iowait_boost = int_tofp(1);
1710                 cpu->last_update = time;
1711                 /*
1712                  * The last time the busy was 100% so P-state was max anyway
1713                  * so avoid overhead of computation.
1714                  */
1715                 if (fp_toint(cpu->sample.busy_scaled) == 100)
1716                         return;
1717
1718                 goto set_pstate;
1719         } else if (cpu->iowait_boost) {
1720                 /* Clear iowait_boost if the CPU may have been idle. */
1721                 delta_ns = time - cpu->last_update;
1722                 if (delta_ns > TICK_NSEC)
1723                         cpu->iowait_boost = 0;
1724         }
1725         cpu->last_update = time;
1726         delta_ns = time - cpu->sample.time;
1727         if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
1728                 return;
1729
1730 set_pstate:
1731         if (intel_pstate_sample(cpu, time))
1732                 intel_pstate_adjust_pstate(cpu);
1733 }
1734
1735 static struct pstate_funcs core_funcs = {
1736         .get_max = core_get_max_pstate,
1737         .get_max_physical = core_get_max_pstate_physical,
1738         .get_min = core_get_min_pstate,
1739         .get_turbo = core_get_turbo_pstate,
1740         .get_scaling = core_get_scaling,
1741         .get_val = core_get_val,
1742 };
1743
1744 static const struct pstate_funcs silvermont_funcs = {
1745         .get_max = atom_get_max_pstate,
1746         .get_max_physical = atom_get_max_pstate,
1747         .get_min = atom_get_min_pstate,
1748         .get_turbo = atom_get_turbo_pstate,
1749         .get_val = atom_get_val,
1750         .get_scaling = silvermont_get_scaling,
1751         .get_vid = atom_get_vid,
1752 };
1753
1754 static const struct pstate_funcs airmont_funcs = {
1755         .get_max = atom_get_max_pstate,
1756         .get_max_physical = atom_get_max_pstate,
1757         .get_min = atom_get_min_pstate,
1758         .get_turbo = atom_get_turbo_pstate,
1759         .get_val = atom_get_val,
1760         .get_scaling = airmont_get_scaling,
1761         .get_vid = atom_get_vid,
1762 };
1763
1764 static const struct pstate_funcs knl_funcs = {
1765         .get_max = core_get_max_pstate,
1766         .get_max_physical = core_get_max_pstate_physical,
1767         .get_min = core_get_min_pstate,
1768         .get_turbo = knl_get_turbo_pstate,
1769         .get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
1770         .get_scaling = core_get_scaling,
1771         .get_val = core_get_val,
1772 };
1773
1774 #define ICPU(model, policy) \
1775         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1776                         (unsigned long)&policy }
1777
1778 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1779         ICPU(INTEL_FAM6_SANDYBRIDGE,            core_funcs),
1780         ICPU(INTEL_FAM6_SANDYBRIDGE_X,          core_funcs),
1781         ICPU(INTEL_FAM6_ATOM_SILVERMONT,        silvermont_funcs),
1782         ICPU(INTEL_FAM6_IVYBRIDGE,              core_funcs),
1783         ICPU(INTEL_FAM6_HASWELL_CORE,           core_funcs),
1784         ICPU(INTEL_FAM6_BROADWELL_CORE,         core_funcs),
1785         ICPU(INTEL_FAM6_IVYBRIDGE_X,            core_funcs),
1786         ICPU(INTEL_FAM6_HASWELL_X,              core_funcs),
1787         ICPU(INTEL_FAM6_HASWELL_ULT,            core_funcs),
1788         ICPU(INTEL_FAM6_HASWELL_GT3E,           core_funcs),
1789         ICPU(INTEL_FAM6_BROADWELL_GT3E,         core_funcs),
1790         ICPU(INTEL_FAM6_ATOM_AIRMONT,           airmont_funcs),
1791         ICPU(INTEL_FAM6_SKYLAKE_MOBILE,         core_funcs),
1792         ICPU(INTEL_FAM6_BROADWELL_X,            core_funcs),
1793         ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,        core_funcs),
1794         ICPU(INTEL_FAM6_BROADWELL_XEON_D,       core_funcs),
1795         ICPU(INTEL_FAM6_XEON_PHI_KNL,           knl_funcs),
1796         ICPU(INTEL_FAM6_XEON_PHI_KNM,           knl_funcs),
1797         ICPU(INTEL_FAM6_ATOM_GOLDMONT,          core_funcs),
1798         ICPU(INTEL_FAM6_ATOM_GOLDMONT_PLUS,     core_funcs),
1799         ICPU(INTEL_FAM6_SKYLAKE_X,              core_funcs),
1800         {}
1801 };
1802 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1803
1804 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1805         ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_funcs),
1806         ICPU(INTEL_FAM6_BROADWELL_X, core_funcs),
1807         ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1808         {}
1809 };
1810
1811 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1812         ICPU(INTEL_FAM6_KABYLAKE_DESKTOP, core_funcs),
1813         {}
1814 };
1815
1816 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
1817         ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1818         ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_funcs),
1819         {}
1820 };
1821
1822 static int intel_pstate_init_cpu(unsigned int cpunum)
1823 {
1824         struct cpudata *cpu;
1825
1826         cpu = all_cpu_data[cpunum];
1827
1828         if (!cpu) {
1829                 cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
1830                 if (!cpu)
1831                         return -ENOMEM;
1832
1833                 all_cpu_data[cpunum] = cpu;
1834
1835                 cpu->epp_default = -EINVAL;
1836                 cpu->epp_powersave = -EINVAL;
1837                 cpu->epp_saved = -EINVAL;
1838         }
1839
1840         cpu = all_cpu_data[cpunum];
1841
1842         cpu->cpu = cpunum;
1843
1844         if (hwp_active) {
1845                 const struct x86_cpu_id *id;
1846
1847                 id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
1848                 if (id)
1849                         intel_pstate_disable_ee(cpunum);
1850
1851                 intel_pstate_hwp_enable(cpu);
1852
1853                 id = x86_match_cpu(intel_pstate_hwp_boost_ids);
1854                 if (id && intel_pstate_acpi_pm_profile_server())
1855                         hwp_boost = true;
1856         }
1857
1858         intel_pstate_get_cpu_pstates(cpu);
1859
1860         pr_debug("controlling: cpu %d\n", cpunum);
1861
1862         return 0;
1863 }
1864
1865 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1866 {
1867         struct cpudata *cpu = all_cpu_data[cpu_num];
1868
1869         if (hwp_active && !hwp_boost)
1870                 return;
1871
1872         if (cpu->update_util_set)
1873                 return;
1874
1875         /* Prevent intel_pstate_update_util() from using stale data. */
1876         cpu->sample.time = 0;
1877         cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1878                                      (hwp_active ?
1879                                       intel_pstate_update_util_hwp :
1880                                       intel_pstate_update_util));
1881         cpu->update_util_set = true;
1882 }
1883
1884 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1885 {
1886         struct cpudata *cpu_data = all_cpu_data[cpu];
1887
1888         if (!cpu_data->update_util_set)
1889                 return;
1890
1891         cpufreq_remove_update_util_hook(cpu);
1892         cpu_data->update_util_set = false;
1893         synchronize_sched();
1894 }
1895
1896 static int intel_pstate_get_max_freq(struct cpudata *cpu)
1897 {
1898         return global.turbo_disabled || global.no_turbo ?
1899                         cpu->pstate.max_freq : cpu->pstate.turbo_freq;
1900 }
1901
1902 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
1903                                             struct cpudata *cpu)
1904 {
1905         int max_freq = intel_pstate_get_max_freq(cpu);
1906         int32_t max_policy_perf, min_policy_perf;
1907         int max_state, turbo_max;
1908
1909         /*
1910          * HWP needs some special consideration, because on BDX the
1911          * HWP_REQUEST uses abstract value to represent performance
1912          * rather than pure ratios.
1913          */
1914         if (hwp_active) {
1915                 intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state);
1916         } else {
1917                 max_state = intel_pstate_get_base_pstate(cpu);
1918                 turbo_max = cpu->pstate.turbo_pstate;
1919         }
1920
1921         max_policy_perf = max_state * policy->max / max_freq;
1922         if (policy->max == policy->min) {
1923                 min_policy_perf = max_policy_perf;
1924         } else {
1925                 min_policy_perf = max_state * policy->min / max_freq;
1926                 min_policy_perf = clamp_t(int32_t, min_policy_perf,
1927                                           0, max_policy_perf);
1928         }
1929
1930         pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n",
1931                  policy->cpu, max_state,
1932                  min_policy_perf, max_policy_perf);
1933
1934         /* Normalize user input to [min_perf, max_perf] */
1935         if (per_cpu_limits) {
1936                 cpu->min_perf_ratio = min_policy_perf;
1937                 cpu->max_perf_ratio = max_policy_perf;
1938         } else {
1939                 int32_t global_min, global_max;
1940
1941                 /* Global limits are in percent of the maximum turbo P-state. */
1942                 global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
1943                 global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
1944                 global_min = clamp_t(int32_t, global_min, 0, global_max);
1945
1946                 pr_debug("cpu:%d global_min:%d global_max:%d\n", policy->cpu,
1947                          global_min, global_max);
1948
1949                 cpu->min_perf_ratio = max(min_policy_perf, global_min);
1950                 cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
1951                 cpu->max_perf_ratio = min(max_policy_perf, global_max);
1952                 cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
1953
1954                 /* Make sure min_perf <= max_perf */
1955                 cpu->min_perf_ratio = min(cpu->min_perf_ratio,
1956                                           cpu->max_perf_ratio);
1957
1958         }
1959         pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", policy->cpu,
1960                  cpu->max_perf_ratio,
1961                  cpu->min_perf_ratio);
1962 }
1963
1964 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
1965 {
1966         struct cpudata *cpu;
1967
1968         if (!policy->cpuinfo.max_freq)
1969                 return -ENODEV;
1970
1971         pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
1972                  policy->cpuinfo.max_freq, policy->max);
1973
1974         cpu = all_cpu_data[policy->cpu];
1975         cpu->policy = policy->policy;
1976
1977         mutex_lock(&intel_pstate_limits_lock);
1978
1979         intel_pstate_update_perf_limits(policy, cpu);
1980
1981         if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
1982                 /*
1983                  * NOHZ_FULL CPUs need this as the governor callback may not
1984                  * be invoked on them.
1985                  */
1986                 intel_pstate_clear_update_util_hook(policy->cpu);
1987                 intel_pstate_max_within_limits(cpu);
1988         } else {
1989                 intel_pstate_set_update_util_hook(policy->cpu);
1990         }
1991
1992         if (hwp_active) {
1993                 /*
1994                  * When hwp_boost was active before and dynamically it
1995                  * was turned off, in that case we need to clear the
1996                  * update util hook.
1997                  */
1998                 if (!hwp_boost)
1999                         intel_pstate_clear_update_util_hook(policy->cpu);
2000                 intel_pstate_hwp_set(policy->cpu);
2001         }
2002
2003         mutex_unlock(&intel_pstate_limits_lock);
2004
2005         return 0;
2006 }
2007
2008 static void intel_pstate_adjust_policy_max(struct cpufreq_policy *policy,
2009                                          struct cpudata *cpu)
2010 {
2011         if (!hwp_active &&
2012             cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2013             policy->max < policy->cpuinfo.max_freq &&
2014             policy->max > cpu->pstate.max_freq) {
2015                 pr_debug("policy->max > max non turbo frequency\n");
2016                 policy->max = policy->cpuinfo.max_freq;
2017         }
2018 }
2019
2020 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2021 {
2022         struct cpudata *cpu = all_cpu_data[policy->cpu];
2023
2024         update_turbo_state();
2025         cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2026                                      intel_pstate_get_max_freq(cpu));
2027
2028         if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2029             policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2030                 return -EINVAL;
2031
2032         intel_pstate_adjust_policy_max(policy, cpu);
2033
2034         return 0;
2035 }
2036
2037 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2038 {
2039         intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2040 }
2041
2042 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2043 {
2044         pr_debug("CPU %d exiting\n", policy->cpu);
2045
2046         intel_pstate_clear_update_util_hook(policy->cpu);
2047         if (hwp_active)
2048                 intel_pstate_hwp_save_state(policy);
2049         else
2050                 intel_cpufreq_stop_cpu(policy);
2051 }
2052
2053 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2054 {
2055         intel_pstate_exit_perf_limits(policy);
2056
2057         policy->fast_switch_possible = false;
2058
2059         return 0;
2060 }
2061
2062 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2063 {
2064         struct cpudata *cpu;
2065         int rc;
2066
2067         rc = intel_pstate_init_cpu(policy->cpu);
2068         if (rc)
2069                 return rc;
2070
2071         cpu = all_cpu_data[policy->cpu];
2072
2073         cpu->max_perf_ratio = 0xFF;
2074         cpu->min_perf_ratio = 0;
2075
2076         policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2077         policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2078
2079         /* cpuinfo and default policy values */
2080         policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2081         update_turbo_state();
2082         policy->cpuinfo.max_freq = global.turbo_disabled ?
2083                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2084         policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2085
2086         if (hwp_active) {
2087                 unsigned int max_freq;
2088
2089                 max_freq = global.turbo_disabled ?
2090                         cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2091                 if (max_freq < policy->cpuinfo.max_freq)
2092                         policy->cpuinfo.max_freq = max_freq;
2093         }
2094
2095         intel_pstate_init_acpi_perf_limits(policy);
2096
2097         policy->fast_switch_possible = true;
2098
2099         return 0;
2100 }
2101
2102 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2103 {
2104         int ret = __intel_pstate_cpu_init(policy);
2105
2106         if (ret)
2107                 return ret;
2108
2109         if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE))
2110                 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2111         else
2112                 policy->policy = CPUFREQ_POLICY_POWERSAVE;
2113
2114         return 0;
2115 }
2116
2117 static struct cpufreq_driver intel_pstate = {
2118         .flags          = CPUFREQ_CONST_LOOPS,
2119         .verify         = intel_pstate_verify_policy,
2120         .setpolicy      = intel_pstate_set_policy,
2121         .suspend        = intel_pstate_hwp_save_state,
2122         .resume         = intel_pstate_resume,
2123         .init           = intel_pstate_cpu_init,
2124         .exit           = intel_pstate_cpu_exit,
2125         .stop_cpu       = intel_pstate_stop_cpu,
2126         .name           = "intel_pstate",
2127 };
2128
2129 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2130 {
2131         struct cpudata *cpu = all_cpu_data[policy->cpu];
2132
2133         update_turbo_state();
2134         cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2135                                      intel_pstate_get_max_freq(cpu));
2136
2137         intel_pstate_adjust_policy_max(policy, cpu);
2138
2139         intel_pstate_update_perf_limits(policy, cpu);
2140
2141         return 0;
2142 }
2143
2144 /* Use of trace in passive mode:
2145  *
2146  * In passive mode the trace core_busy field (also known as the
2147  * performance field, and lablelled as such on the graphs; also known as
2148  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2149  * driver call was via the normal or fast switch path. Various graphs
2150  * output from the intel_pstate_tracer.py utility that include core_busy
2151  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2152  * so we use 10 to indicate the the normal path through the driver, and
2153  * 90 to indicate the fast switch path through the driver.
2154  * The scaled_busy field is not used, and is set to 0.
2155  */
2156
2157 #define INTEL_PSTATE_TRACE_TARGET 10
2158 #define INTEL_PSTATE_TRACE_FAST_SWITCH 90
2159
2160 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2161 {
2162         struct sample *sample;
2163
2164         if (!trace_pstate_sample_enabled())
2165                 return;
2166
2167         if (!intel_pstate_sample(cpu, ktime_get()))
2168                 return;
2169
2170         sample = &cpu->sample;
2171         trace_pstate_sample(trace_type,
2172                 0,
2173                 old_pstate,
2174                 cpu->pstate.current_pstate,
2175                 sample->mperf,
2176                 sample->aperf,
2177                 sample->tsc,
2178                 get_avg_frequency(cpu),
2179                 fp_toint(cpu->iowait_boost * 100));
2180 }
2181
2182 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2183                                 unsigned int target_freq,
2184                                 unsigned int relation)
2185 {
2186         struct cpudata *cpu = all_cpu_data[policy->cpu];
2187         struct cpufreq_freqs freqs;
2188         int target_pstate, old_pstate;
2189
2190         update_turbo_state();
2191
2192         freqs.old = policy->cur;
2193         freqs.new = target_freq;
2194
2195         cpufreq_freq_transition_begin(policy, &freqs);
2196         switch (relation) {
2197         case CPUFREQ_RELATION_L:
2198                 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2199                 break;
2200         case CPUFREQ_RELATION_H:
2201                 target_pstate = freqs.new / cpu->pstate.scaling;
2202                 break;
2203         default:
2204                 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2205                 break;
2206         }
2207         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2208         old_pstate = cpu->pstate.current_pstate;
2209         if (target_pstate != cpu->pstate.current_pstate) {
2210                 cpu->pstate.current_pstate = target_pstate;
2211                 wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2212                               pstate_funcs.get_val(cpu, target_pstate));
2213         }
2214         freqs.new = target_pstate * cpu->pstate.scaling;
2215         intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_TARGET, old_pstate);
2216         cpufreq_freq_transition_end(policy, &freqs, false);
2217
2218         return 0;
2219 }
2220
2221 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2222                                               unsigned int target_freq)
2223 {
2224         struct cpudata *cpu = all_cpu_data[policy->cpu];
2225         int target_pstate, old_pstate;
2226
2227         update_turbo_state();
2228
2229         target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2230         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2231         old_pstate = cpu->pstate.current_pstate;
2232         intel_pstate_update_pstate(cpu, target_pstate);
2233         intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2234         return target_pstate * cpu->pstate.scaling;
2235 }
2236
2237 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2238 {
2239         int ret = __intel_pstate_cpu_init(policy);
2240
2241         if (ret)
2242                 return ret;
2243
2244         policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2245         policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2246         /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2247         policy->cur = policy->cpuinfo.min_freq;
2248
2249         return 0;
2250 }
2251
2252 static struct cpufreq_driver intel_cpufreq = {
2253         .flags          = CPUFREQ_CONST_LOOPS,
2254         .verify         = intel_cpufreq_verify_policy,
2255         .target         = intel_cpufreq_target,
2256         .fast_switch    = intel_cpufreq_fast_switch,
2257         .init           = intel_cpufreq_cpu_init,
2258         .exit           = intel_pstate_cpu_exit,
2259         .stop_cpu       = intel_cpufreq_stop_cpu,
2260         .name           = "intel_cpufreq",
2261 };
2262
2263 static struct cpufreq_driver *default_driver = &intel_pstate;
2264
2265 static void intel_pstate_driver_cleanup(void)
2266 {
2267         unsigned int cpu;
2268
2269         get_online_cpus();
2270         for_each_online_cpu(cpu) {
2271                 if (all_cpu_data[cpu]) {
2272                         if (intel_pstate_driver == &intel_pstate)
2273                                 intel_pstate_clear_update_util_hook(cpu);
2274
2275                         kfree(all_cpu_data[cpu]);
2276                         all_cpu_data[cpu] = NULL;
2277                 }
2278         }
2279         put_online_cpus();
2280         intel_pstate_driver = NULL;
2281 }
2282
2283 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2284 {
2285         int ret;
2286
2287         memset(&global, 0, sizeof(global));
2288         global.max_perf_pct = 100;
2289
2290         intel_pstate_driver = driver;
2291         ret = cpufreq_register_driver(intel_pstate_driver);
2292         if (ret) {
2293                 intel_pstate_driver_cleanup();
2294                 return ret;
2295         }
2296
2297         global.min_perf_pct = min_perf_pct_min();
2298
2299         return 0;
2300 }
2301
2302 static int intel_pstate_unregister_driver(void)
2303 {
2304         if (hwp_active)
2305                 return -EBUSY;
2306
2307         cpufreq_unregister_driver(intel_pstate_driver);
2308         intel_pstate_driver_cleanup();
2309
2310         return 0;
2311 }
2312
2313 static ssize_t intel_pstate_show_status(char *buf)
2314 {
2315         if (!intel_pstate_driver)
2316                 return sprintf(buf, "off\n");
2317
2318         return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2319                                         "active" : "passive");
2320 }
2321
2322 static int intel_pstate_update_status(const char *buf, size_t size)
2323 {
2324         int ret;
2325
2326         if (size == 3 && !strncmp(buf, "off", size))
2327                 return intel_pstate_driver ?
2328                         intel_pstate_unregister_driver() : -EINVAL;
2329
2330         if (size == 6 && !strncmp(buf, "active", size)) {
2331                 if (intel_pstate_driver) {
2332                         if (intel_pstate_driver == &intel_pstate)
2333                                 return 0;
2334
2335                         ret = intel_pstate_unregister_driver();
2336                         if (ret)
2337                                 return ret;
2338                 }
2339
2340                 return intel_pstate_register_driver(&intel_pstate);
2341         }
2342
2343         if (size == 7 && !strncmp(buf, "passive", size)) {
2344                 if (intel_pstate_driver) {
2345                         if (intel_pstate_driver == &intel_cpufreq)
2346                                 return 0;
2347
2348                         ret = intel_pstate_unregister_driver();
2349                         if (ret)
2350                                 return ret;
2351                 }
2352
2353                 return intel_pstate_register_driver(&intel_cpufreq);
2354         }
2355
2356         return -EINVAL;
2357 }
2358
2359 static int no_load __initdata;
2360 static int no_hwp __initdata;
2361 static int hwp_only __initdata;
2362 static unsigned int force_load __initdata;
2363
2364 static int __init intel_pstate_msrs_not_valid(void)
2365 {
2366         if (!pstate_funcs.get_max() ||
2367             !pstate_funcs.get_min() ||
2368             !pstate_funcs.get_turbo())
2369                 return -ENODEV;
2370
2371         return 0;
2372 }
2373
2374 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2375 {
2376         pstate_funcs.get_max   = funcs->get_max;
2377         pstate_funcs.get_max_physical = funcs->get_max_physical;
2378         pstate_funcs.get_min   = funcs->get_min;
2379         pstate_funcs.get_turbo = funcs->get_turbo;
2380         pstate_funcs.get_scaling = funcs->get_scaling;
2381         pstate_funcs.get_val   = funcs->get_val;
2382         pstate_funcs.get_vid   = funcs->get_vid;
2383         pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
2384 }
2385
2386 #ifdef CONFIG_ACPI
2387
2388 static bool __init intel_pstate_no_acpi_pss(void)
2389 {
2390         int i;
2391
2392         for_each_possible_cpu(i) {
2393                 acpi_status status;
2394                 union acpi_object *pss;
2395                 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2396                 struct acpi_processor *pr = per_cpu(processors, i);
2397
2398                 if (!pr)
2399                         continue;
2400
2401                 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2402                 if (ACPI_FAILURE(status))
2403                         continue;
2404
2405                 pss = buffer.pointer;
2406                 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2407                         kfree(pss);
2408                         return false;
2409                 }
2410
2411                 kfree(pss);
2412         }
2413
2414         return true;
2415 }
2416
2417 static bool __init intel_pstate_no_acpi_pcch(void)
2418 {
2419         acpi_status status;
2420         acpi_handle handle;
2421
2422         status = acpi_get_handle(NULL, "\\_SB", &handle);
2423         if (ACPI_FAILURE(status))
2424                 return true;
2425
2426         return !acpi_has_method(handle, "PCCH");
2427 }
2428
2429 static bool __init intel_pstate_has_acpi_ppc(void)
2430 {
2431         int i;
2432
2433         for_each_possible_cpu(i) {
2434                 struct acpi_processor *pr = per_cpu(processors, i);
2435
2436                 if (!pr)
2437                         continue;
2438                 if (acpi_has_method(pr->handle, "_PPC"))
2439                         return true;
2440         }
2441         return false;
2442 }
2443
2444 enum {
2445         PSS,
2446         PPC,
2447 };
2448
2449 /* Hardware vendor-specific info that has its own power management modes */
2450 static struct acpi_platform_list plat_info[] __initdata = {
2451         {"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, 0, PSS},
2452         {"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2453         {"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2454         {"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2455         {"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2456         {"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2457         {"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2458         {"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2459         {"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2460         {"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2461         {"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2462         {"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2463         {"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2464         {"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2465         {"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2466         { } /* End */
2467 };
2468
2469 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2470 {
2471         const struct x86_cpu_id *id;
2472         u64 misc_pwr;
2473         int idx;
2474
2475         id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2476         if (id) {
2477                 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2478                 if ( misc_pwr & (1 << 8))
2479                         return true;
2480         }
2481
2482         idx = acpi_match_platform_list(plat_info);
2483         if (idx < 0)
2484                 return false;
2485
2486         switch (plat_info[idx].data) {
2487         case PSS:
2488                 if (!intel_pstate_no_acpi_pss())
2489                         return false;
2490
2491                 return intel_pstate_no_acpi_pcch();
2492         case PPC:
2493                 return intel_pstate_has_acpi_ppc() && !force_load;
2494         }
2495
2496         return false;
2497 }
2498
2499 static void intel_pstate_request_control_from_smm(void)
2500 {
2501         /*
2502          * It may be unsafe to request P-states control from SMM if _PPC support
2503          * has not been enabled.
2504          */
2505         if (acpi_ppc)
2506                 acpi_processor_pstate_control();
2507 }
2508 #else /* CONFIG_ACPI not enabled */
2509 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2510 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2511 static inline void intel_pstate_request_control_from_smm(void) {}
2512 #endif /* CONFIG_ACPI */
2513
2514 #define INTEL_PSTATE_HWP_BROADWELL      0x01
2515
2516 #define ICPU_HWP(model, hwp_mode) \
2517         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_HWP, hwp_mode }
2518
2519 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2520         ICPU_HWP(INTEL_FAM6_BROADWELL_X, INTEL_PSTATE_HWP_BROADWELL),
2521         ICPU_HWP(INTEL_FAM6_BROADWELL_XEON_D, INTEL_PSTATE_HWP_BROADWELL),
2522         ICPU_HWP(X86_MODEL_ANY, 0),
2523         {}
2524 };
2525
2526 static int __init intel_pstate_init(void)
2527 {
2528         const struct x86_cpu_id *id;
2529         int rc;
2530
2531         if (no_load)
2532                 return -ENODEV;
2533
2534         id = x86_match_cpu(hwp_support_ids);
2535         if (id) {
2536                 copy_cpu_funcs(&core_funcs);
2537                 if (!no_hwp) {
2538                         hwp_active++;
2539                         hwp_mode_bdw = id->driver_data;
2540                         intel_pstate.attr = hwp_cpufreq_attrs;
2541                         goto hwp_cpu_matched;
2542                 }
2543         } else {
2544                 id = x86_match_cpu(intel_pstate_cpu_ids);
2545                 if (!id)
2546                         return -ENODEV;
2547
2548                 copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
2549         }
2550
2551         if (intel_pstate_msrs_not_valid())
2552                 return -ENODEV;
2553
2554 hwp_cpu_matched:
2555         /*
2556          * The Intel pstate driver will be ignored if the platform
2557          * firmware has its own power management modes.
2558          */
2559         if (intel_pstate_platform_pwr_mgmt_exists())
2560                 return -ENODEV;
2561
2562         if (!hwp_active && hwp_only)
2563                 return -ENOTSUPP;
2564
2565         pr_info("Intel P-state driver initializing\n");
2566
2567         all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
2568         if (!all_cpu_data)
2569                 return -ENOMEM;
2570
2571         intel_pstate_request_control_from_smm();
2572
2573         intel_pstate_sysfs_expose_params();
2574
2575         mutex_lock(&intel_pstate_driver_lock);
2576         rc = intel_pstate_register_driver(default_driver);
2577         mutex_unlock(&intel_pstate_driver_lock);
2578         if (rc)
2579                 return rc;
2580
2581         if (hwp_active)
2582                 pr_info("HWP enabled\n");
2583
2584         return 0;
2585 }
2586 device_initcall(intel_pstate_init);
2587
2588 static int __init intel_pstate_setup(char *str)
2589 {
2590         if (!str)
2591                 return -EINVAL;
2592
2593         if (!strcmp(str, "disable")) {
2594                 no_load = 1;
2595         } else if (!strcmp(str, "passive")) {
2596                 pr_info("Passive mode enabled\n");
2597                 default_driver = &intel_cpufreq;
2598                 no_hwp = 1;
2599         }
2600         if (!strcmp(str, "no_hwp")) {
2601                 pr_info("HWP disabled\n");
2602                 no_hwp = 1;
2603         }
2604         if (!strcmp(str, "force"))
2605                 force_load = 1;
2606         if (!strcmp(str, "hwp_only"))
2607                 hwp_only = 1;
2608         if (!strcmp(str, "per_cpu_perf_limits"))
2609                 per_cpu_limits = true;
2610
2611 #ifdef CONFIG_ACPI
2612         if (!strcmp(str, "support_acpi_ppc"))
2613                 acpi_ppc = true;
2614 #endif
2615
2616         return 0;
2617 }
2618 early_param("intel_pstate", intel_pstate_setup);
2619
2620 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2621 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2622 MODULE_LICENSE("GPL");