?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
io_uring: ensure finish_wait() is always called in __io_uring_task_cancel()
[linux/fpc-iii.git] / drivers / cpufreq / intel_pstate.c
blob6e23376548ced1196516f600ddf32529884fda85
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * intel_pstate.c: Native P state management for Intel processors
4 *
5 * (C) Copyright 2012 Intel Corporation
6 * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
7 */
8
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11 #include <linux/kernel.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/module.h>
14 #include <linux/ktime.h>
15 #include <linux/hrtimer.h>
16 #include <linux/tick.h>
17 #include <linux/slab.h>
18 #include <linux/sched/cpufreq.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
21 #include <linux/cpufreq.h>
22 #include <linux/sysfs.h>
23 #include <linux/types.h>
24 #include <linux/fs.h>
25 #include <linux/acpi.h>
26 #include <linux/vmalloc.h>
27 #include <linux/pm_qos.h>
28 #include <trace/events/power.h>
29
30 #include <asm/div64.h>
31 #include <asm/msr.h>
32 #include <asm/cpu_device_id.h>
33 #include <asm/cpufeature.h>
34 #include <asm/intel-family.h>
35
36 #define INTEL_PSTATE_SAMPLING_INTERVAL (10 * NSEC_PER_MSEC)
37
38 #define INTEL_CPUFREQ_TRANSITION_LATENCY 20000
39 #define INTEL_CPUFREQ_TRANSITION_DELAY_HWP 5000
40 #define INTEL_CPUFREQ_TRANSITION_DELAY 500
41
42 #ifdef CONFIG_ACPI
43 #include <acpi/processor.h>
44 #include <acpi/cppc_acpi.h>
45 #endif
46
47 #define FRAC_BITS 8
48 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
49 #define fp_toint(X) ((X) >> FRAC_BITS)
50
51 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
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: Whether 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 * @turbo_disabled_mf: The @turbo_disabled value reflected by cpuinfo.max_freq.
181 * @min_perf_pct: Minimum capacity limit in percent of the maximum turbo
182 * P-state capacity.
183 * @max_perf_pct: Maximum capacity limit in percent of the maximum turbo
184 * P-state capacity.
185 */
186 struct global_params {
187 bool no_turbo;
188 bool turbo_disabled;
189 bool turbo_disabled_mf;
190 int max_perf_pct;
191 int min_perf_pct;
192 };
193
194 /**
195 * struct cpudata - Per CPU instance data storage
196 * @cpu: CPU number for this instance data
197 * @policy: CPUFreq policy value
198 * @update_util: CPUFreq utility callback information
199 * @update_util_set: CPUFreq utility callback is set
200 * @iowait_boost: iowait-related boost fraction
201 * @last_update: Time of the last update.
202 * @pstate: Stores P state limits for this CPU
203 * @vid: Stores VID limits for this CPU
204 * @last_sample_time: Last Sample time
205 * @aperf_mperf_shift: APERF vs MPERF counting frequency difference
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_cached Cached HWP energy-performance preference value
223 * @hwp_req_cached: Cached value of the last HWP Request MSR
224 * @hwp_cap_cached: Cached value of the last HWP Capabilities MSR
225 * @last_io_update: Last time when IO wake flag was set
226 * @sched_flags: Store scheduler flags for possible cross CPU update
227 * @hwp_boost_min: Last HWP boosted min performance
228 * @suspended: Whether or not the driver has been suspended.
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_cached;
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 bool suspended;
267 };
268
269 static struct cpudata **all_cpu_data;
270
271 /**
272 * struct pstate_funcs - Per CPU model specific callbacks
273 * @get_max: Callback to get maximum non turbo effective P state
274 * @get_max_physical: Callback to get maximum non turbo physical P state
275 * @get_min: Callback to get minimum P state
276 * @get_turbo: Callback to get turbo P state
277 * @get_scaling: Callback to get frequency scaling factor
278 * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
279 * @get_val: Callback to convert P state to actual MSR write value
280 * @get_vid: Callback to get VID data for Atom platforms
281 *
282 * Core and Atom CPU models have different way to get P State limits. This
283 * structure is used to store those callbacks.
284 */
285 struct pstate_funcs {
286 int (*get_max)(void);
287 int (*get_max_physical)(void);
288 int (*get_min)(void);
289 int (*get_turbo)(void);
290 int (*get_scaling)(void);
291 int (*get_aperf_mperf_shift)(void);
292 u64 (*get_val)(struct cpudata*, int pstate);
293 void (*get_vid)(struct cpudata *);
294 };
295
296 static struct pstate_funcs pstate_funcs __read_mostly;
297
298 static int hwp_active __read_mostly;
299 static int hwp_mode_bdw __read_mostly;
300 static bool per_cpu_limits __read_mostly;
301 static bool hwp_boost __read_mostly;
302
303 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
304
305 #ifdef CONFIG_ACPI
306 static bool acpi_ppc;
307 #endif
308
309 static struct global_params global;
310
311 static DEFINE_MUTEX(intel_pstate_driver_lock);
312 static DEFINE_MUTEX(intel_pstate_limits_lock);
313
314 #ifdef CONFIG_ACPI
315
316 static bool intel_pstate_acpi_pm_profile_server(void)
317 {
318 if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
319 acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
320 return true;
321
322 return false;
323 }
324
325 static bool intel_pstate_get_ppc_enable_status(void)
326 {
327 if (intel_pstate_acpi_pm_profile_server())
328 return true;
329
330 return acpi_ppc;
331 }
332
333 #ifdef CONFIG_ACPI_CPPC_LIB
334
335 /* The work item is needed to avoid CPU hotplug locking issues */
336 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
337 {
338 sched_set_itmt_support();
339 }
340
341 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
342
343 static void intel_pstate_set_itmt_prio(int cpu)
344 {
345 struct cppc_perf_caps cppc_perf;
346 static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
347 int ret;
348
349 ret = cppc_get_perf_caps(cpu, &cppc_perf);
350 if (ret)
351 return;
352
353 /*
354 * The priorities can be set regardless of whether or not
355 * sched_set_itmt_support(true) has been called and it is valid to
356 * update them at any time after it has been called.
357 */
358 sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
359
360 if (max_highest_perf <= min_highest_perf) {
361 if (cppc_perf.highest_perf > max_highest_perf)
362 max_highest_perf = cppc_perf.highest_perf;
363
364 if (cppc_perf.highest_perf < min_highest_perf)
365 min_highest_perf = cppc_perf.highest_perf;
366
367 if (max_highest_perf > min_highest_perf) {
368 /*
369 * This code can be run during CPU online under the
370 * CPU hotplug locks, so sched_set_itmt_support()
371 * cannot be called from here. Queue up a work item
372 * to invoke it.
373 */
374 schedule_work(&sched_itmt_work);
375 }
376 }
377 }
378
379 static int intel_pstate_get_cppc_guranteed(int cpu)
380 {
381 struct cppc_perf_caps cppc_perf;
382 int ret;
383
384 ret = cppc_get_perf_caps(cpu, &cppc_perf);
385 if (ret)
386 return ret;
387
388 if (cppc_perf.guaranteed_perf)
389 return cppc_perf.guaranteed_perf;
390
391 return cppc_perf.nominal_perf;
392 }
393
394 #else /* CONFIG_ACPI_CPPC_LIB */
395 static void intel_pstate_set_itmt_prio(int cpu)
396 {
397 }
398 #endif /* CONFIG_ACPI_CPPC_LIB */
399
400 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
401 {
402 struct cpudata *cpu;
403 int ret;
404 int i;
405
406 if (hwp_active) {
407 intel_pstate_set_itmt_prio(policy->cpu);
408 return;
409 }
410
411 if (!intel_pstate_get_ppc_enable_status())
412 return;
413
414 cpu = all_cpu_data[policy->cpu];
415
416 ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
417 policy->cpu);
418 if (ret)
419 return;
420
421 /*
422 * Check if the control value in _PSS is for PERF_CTL MSR, which should
423 * guarantee that the states returned by it map to the states in our
424 * list directly.
425 */
426 if (cpu->acpi_perf_data.control_register.space_id !=
427 ACPI_ADR_SPACE_FIXED_HARDWARE)
428 goto err;
429
430 /*
431 * If there is only one entry _PSS, simply ignore _PSS and continue as
432 * usual without taking _PSS into account
433 */
434 if (cpu->acpi_perf_data.state_count < 2)
435 goto err;
436
437 pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
438 for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
439 pr_debug(" %cP%d: %u MHz, %u mW, 0x%x\n",
440 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
441 (u32) cpu->acpi_perf_data.states[i].core_frequency,
442 (u32) cpu->acpi_perf_data.states[i].power,
443 (u32) cpu->acpi_perf_data.states[i].control);
444 }
445
446 /*
447 * The _PSS table doesn't contain whole turbo frequency range.
448 * This just contains +1 MHZ above the max non turbo frequency,
449 * with control value corresponding to max turbo ratio. But
450 * when cpufreq set policy is called, it will call with this
451 * max frequency, which will cause a reduced performance as
452 * this driver uses real max turbo frequency as the max
453 * frequency. So correct this frequency in _PSS table to
454 * correct max turbo frequency based on the turbo state.
455 * Also need to convert to MHz as _PSS freq is in MHz.
456 */
457 if (!global.turbo_disabled)
458 cpu->acpi_perf_data.states[0].core_frequency =
459 policy->cpuinfo.max_freq / 1000;
460 cpu->valid_pss_table = true;
461 pr_debug("_PPC limits will be enforced\n");
462
463 return;
464
465 err:
466 cpu->valid_pss_table = false;
467 acpi_processor_unregister_performance(policy->cpu);
468 }
469
470 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
471 {
472 struct cpudata *cpu;
473
474 cpu = all_cpu_data[policy->cpu];
475 if (!cpu->valid_pss_table)
476 return;
477
478 acpi_processor_unregister_performance(policy->cpu);
479 }
480 #else /* CONFIG_ACPI */
481 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
482 {
483 }
484
485 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
486 {
487 }
488
489 static inline bool intel_pstate_acpi_pm_profile_server(void)
490 {
491 return false;
492 }
493 #endif /* CONFIG_ACPI */
494
495 #ifndef CONFIG_ACPI_CPPC_LIB
496 static int intel_pstate_get_cppc_guranteed(int cpu)
497 {
498 return -ENOTSUPP;
499 }
500 #endif /* CONFIG_ACPI_CPPC_LIB */
501
502 static inline void update_turbo_state(void)
503 {
504 u64 misc_en;
505 struct cpudata *cpu;
506
507 cpu = all_cpu_data[0];
508 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
509 global.turbo_disabled =
510 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
511 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
512 }
513
514 static int min_perf_pct_min(void)
515 {
516 struct cpudata *cpu = all_cpu_data[0];
517 int turbo_pstate = cpu->pstate.turbo_pstate;
518
519 return turbo_pstate ?
520 (cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
521 }
522
523 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
524 {
525 u64 epb;
526 int ret;
527
528 if (!boot_cpu_has(X86_FEATURE_EPB))
529 return -ENXIO;
530
531 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
532 if (ret)
533 return (s16)ret;
534
535 return (s16)(epb & 0x0f);
536 }
537
538 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
539 {
540 s16 epp;
541
542 if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
543 /*
544 * When hwp_req_data is 0, means that caller didn't read
545 * MSR_HWP_REQUEST, so need to read and get EPP.
546 */
547 if (!hwp_req_data) {
548 epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
549 &hwp_req_data);
550 if (epp)
551 return epp;
552 }
553 epp = (hwp_req_data >> 24) & 0xff;
554 } else {
555 /* When there is no EPP present, HWP uses EPB settings */
556 epp = intel_pstate_get_epb(cpu_data);
557 }
558
559 return epp;
560 }
561
562 static int intel_pstate_set_epb(int cpu, s16 pref)
563 {
564 u64 epb;
565 int ret;
566
567 if (!boot_cpu_has(X86_FEATURE_EPB))
568 return -ENXIO;
569
570 ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
571 if (ret)
572 return ret;
573
574 epb = (epb & ~0x0f) | pref;
575 wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
576
577 return 0;
578 }
579
580 /*
581 * EPP/EPB display strings corresponding to EPP index in the
582 * energy_perf_strings[]
583 * index String
584 *-------------------------------------
585 * 0 default
586 * 1 performance
587 * 2 balance_performance
588 * 3 balance_power
589 * 4 power
590 */
591 static const char * const energy_perf_strings[] = {
592 "default",
593 "performance",
594 "balance_performance",
595 "balance_power",
596 "power",
597 NULL
598 };
599 static const unsigned int epp_values[] = {
600 HWP_EPP_PERFORMANCE,
601 HWP_EPP_BALANCE_PERFORMANCE,
602 HWP_EPP_BALANCE_POWERSAVE,
603 HWP_EPP_POWERSAVE
604 };
605
606 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
607 {
608 s16 epp;
609 int index = -EINVAL;
610
611 *raw_epp = 0;
612 epp = intel_pstate_get_epp(cpu_data, 0);
613 if (epp < 0)
614 return epp;
615
616 if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
617 if (epp == HWP_EPP_PERFORMANCE)
618 return 1;
619 if (epp == HWP_EPP_BALANCE_PERFORMANCE)
620 return 2;
621 if (epp == HWP_EPP_BALANCE_POWERSAVE)
622 return 3;
623 if (epp == HWP_EPP_POWERSAVE)
624 return 4;
625 *raw_epp = epp;
626 return 0;
627 } else if (boot_cpu_has(X86_FEATURE_EPB)) {
628 /*
629 * Range:
630 * 0x00-0x03 : Performance
631 * 0x04-0x07 : Balance performance
632 * 0x08-0x0B : Balance power
633 * 0x0C-0x0F : Power
634 * The EPB is a 4 bit value, but our ranges restrict the
635 * value which can be set. Here only using top two bits
636 * effectively.
637 */
638 index = (epp >> 2) + 1;
639 }
640
641 return index;
642 }
643
644 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
645 {
646 int ret;
647
648 /*
649 * Use the cached HWP Request MSR value, because in the active mode the
650 * register itself may be updated by intel_pstate_hwp_boost_up() or
651 * intel_pstate_hwp_boost_down() at any time.
652 */
653 u64 value = READ_ONCE(cpu->hwp_req_cached);
654
655 value &= ~GENMASK_ULL(31, 24);
656 value |= (u64)epp << 24;
657 /*
658 * The only other updater of hwp_req_cached in the active mode,
659 * intel_pstate_hwp_set(), is called under the same lock as this
660 * function, so it cannot run in parallel with the update below.
661 */
662 WRITE_ONCE(cpu->hwp_req_cached, value);
663 ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
664 if (!ret)
665 cpu->epp_cached = epp;
666
667 return ret;
668 }
669
670 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
671 int pref_index, bool use_raw,
672 u32 raw_epp)
673 {
674 int epp = -EINVAL;
675 int ret;
676
677 if (!pref_index)
678 epp = cpu_data->epp_default;
679
680 if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
681 if (use_raw)
682 epp = raw_epp;
683 else if (epp == -EINVAL)
684 epp = epp_values[pref_index - 1];
685
686 /*
687 * To avoid confusion, refuse to set EPP to any values different
688 * from 0 (performance) if the current policy is "performance",
689 * because those values would be overridden.
690 */
691 if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
692 return -EBUSY;
693
694 ret = intel_pstate_set_epp(cpu_data, epp);
695 } else {
696 if (epp == -EINVAL)
697 epp = (pref_index - 1) << 2;
698 ret = intel_pstate_set_epb(cpu_data->cpu, epp);
699 }
700
701 return ret;
702 }
703
704 static ssize_t show_energy_performance_available_preferences(
705 struct cpufreq_policy *policy, char *buf)
706 {
707 int i = 0;
708 int ret = 0;
709
710 while (energy_perf_strings[i] != NULL)
711 ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
712
713 ret += sprintf(&buf[ret], "\n");
714
715 return ret;
716 }
717
718 cpufreq_freq_attr_ro(energy_performance_available_preferences);
719
720 static struct cpufreq_driver intel_pstate;
721
722 static ssize_t store_energy_performance_preference(
723 struct cpufreq_policy *policy, const char *buf, size_t count)
724 {
725 struct cpudata *cpu = all_cpu_data[policy->cpu];
726 char str_preference[21];
727 bool raw = false;
728 ssize_t ret;
729 u32 epp = 0;
730
731 ret = sscanf(buf, "%20s", str_preference);
732 if (ret != 1)
733 return -EINVAL;
734
735 ret = match_string(energy_perf_strings, -1, str_preference);
736 if (ret < 0) {
737 if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
738 return ret;
739
740 ret = kstrtouint(buf, 10, &epp);
741 if (ret)
742 return ret;
743
744 if (epp > 255)
745 return -EINVAL;
746
747 raw = true;
748 }
749
750 /*
751 * This function runs with the policy R/W semaphore held, which
752 * guarantees that the driver pointer will not change while it is
753 * running.
754 */
755 if (!intel_pstate_driver)
756 return -EAGAIN;
757
758 mutex_lock(&intel_pstate_limits_lock);
759
760 if (intel_pstate_driver == &intel_pstate) {
761 ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
762 } else {
763 /*
764 * In the passive mode the governor needs to be stopped on the
765 * target CPU before the EPP update and restarted after it,
766 * which is super-heavy-weight, so make sure it is worth doing
767 * upfront.
768 */
769 if (!raw)
770 epp = ret ? epp_values[ret - 1] : cpu->epp_default;
771
772 if (cpu->epp_cached != epp) {
773 int err;
774
775 cpufreq_stop_governor(policy);
776 ret = intel_pstate_set_epp(cpu, epp);
777 err = cpufreq_start_governor(policy);
778 if (!ret)
779 ret = err;
780 }
781 }
782
783 mutex_unlock(&intel_pstate_limits_lock);
784
785 return ret ?: count;
786 }
787
788 static ssize_t show_energy_performance_preference(
789 struct cpufreq_policy *policy, char *buf)
790 {
791 struct cpudata *cpu_data = all_cpu_data[policy->cpu];
792 int preference, raw_epp;
793
794 preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
795 if (preference < 0)
796 return preference;
797
798 if (raw_epp)
799 return sprintf(buf, "%d\n", raw_epp);
800 else
801 return sprintf(buf, "%s\n", energy_perf_strings[preference]);
802 }
803
804 cpufreq_freq_attr_rw(energy_performance_preference);
805
806 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
807 {
808 struct cpudata *cpu;
809 u64 cap;
810 int ratio;
811
812 ratio = intel_pstate_get_cppc_guranteed(policy->cpu);
813 if (ratio <= 0) {
814 rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
815 ratio = HWP_GUARANTEED_PERF(cap);
816 }
817
818 cpu = all_cpu_data[policy->cpu];
819
820 return sprintf(buf, "%d\n", ratio * cpu->pstate.scaling);
821 }
822
823 cpufreq_freq_attr_ro(base_frequency);
824
825 static struct freq_attr *hwp_cpufreq_attrs[] = {
826 &energy_performance_preference,
827 &energy_performance_available_preferences,
828 &base_frequency,
829 NULL,
830 };
831
832 static void intel_pstate_get_hwp_max(unsigned int cpu, int *phy_max,
833 int *current_max)
834 {
835 u64 cap;
836
837 rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
838 WRITE_ONCE(all_cpu_data[cpu]->hwp_cap_cached, cap);
839 if (global.no_turbo || global.turbo_disabled)
840 *current_max = HWP_GUARANTEED_PERF(cap);
841 else
842 *current_max = HWP_HIGHEST_PERF(cap);
843
844 *phy_max = HWP_HIGHEST_PERF(cap);
845 }
846
847 static void intel_pstate_hwp_set(unsigned int cpu)
848 {
849 struct cpudata *cpu_data = all_cpu_data[cpu];
850 int max, min;
851 u64 value;
852 s16 epp;
853
854 max = cpu_data->max_perf_ratio;
855 min = cpu_data->min_perf_ratio;
856
857 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
858 min = max;
859
860 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
861
862 value &= ~HWP_MIN_PERF(~0L);
863 value |= HWP_MIN_PERF(min);
864
865 value &= ~HWP_MAX_PERF(~0L);
866 value |= HWP_MAX_PERF(max);
867
868 if (cpu_data->epp_policy == cpu_data->policy)
869 goto skip_epp;
870
871 cpu_data->epp_policy = cpu_data->policy;
872
873 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
874 epp = intel_pstate_get_epp(cpu_data, value);
875 cpu_data->epp_powersave = epp;
876 /* If EPP read was failed, then don't try to write */
877 if (epp < 0)
878 goto skip_epp;
879
880 epp = 0;
881 } else {
882 /* skip setting EPP, when saved value is invalid */
883 if (cpu_data->epp_powersave < 0)
884 goto skip_epp;
885
886 /*
887 * No need to restore EPP when it is not zero. This
888 * means:
889 * - Policy is not changed
890 * - user has manually changed
891 * - Error reading EPB
892 */
893 epp = intel_pstate_get_epp(cpu_data, value);
894 if (epp)
895 goto skip_epp;
896
897 epp = cpu_data->epp_powersave;
898 }
899 if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
900 value &= ~GENMASK_ULL(31, 24);
901 value |= (u64)epp << 24;
902 } else {
903 intel_pstate_set_epb(cpu, epp);
904 }
905 skip_epp:
906 WRITE_ONCE(cpu_data->hwp_req_cached, value);
907 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
908 }
909
910 static void intel_pstate_hwp_offline(struct cpudata *cpu)
911 {
912 u64 value = READ_ONCE(cpu->hwp_req_cached);
913 int min_perf;
914
915 if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
916 /*
917 * In case the EPP has been set to "performance" by the
918 * active mode "performance" scaling algorithm, replace that
919 * temporary value with the cached EPP one.
920 */
921 value &= ~GENMASK_ULL(31, 24);
922 value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
923 WRITE_ONCE(cpu->hwp_req_cached, value);
924 }
925
926 value &= ~GENMASK_ULL(31, 0);
927 min_perf = HWP_LOWEST_PERF(cpu->hwp_cap_cached);
928
929 /* Set hwp_max = hwp_min */
930 value |= HWP_MAX_PERF(min_perf);
931 value |= HWP_MIN_PERF(min_perf);
932
933 /* Set EPP to min */
934 if (boot_cpu_has(X86_FEATURE_HWP_EPP))
935 value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
936
937 wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
938 }
939
940 #define POWER_CTL_EE_ENABLE 1
941 #define POWER_CTL_EE_DISABLE 2
942
943 static int power_ctl_ee_state;
944
945 static void set_power_ctl_ee_state(bool input)
946 {
947 u64 power_ctl;
948
949 mutex_lock(&intel_pstate_driver_lock);
950 rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
951 if (input) {
952 power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
953 power_ctl_ee_state = POWER_CTL_EE_ENABLE;
954 } else {
955 power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
956 power_ctl_ee_state = POWER_CTL_EE_DISABLE;
957 }
958 wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
959 mutex_unlock(&intel_pstate_driver_lock);
960 }
961
962 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
963
964 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
965 {
966 intel_pstate_hwp_enable(cpu);
967 wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
968 }
969
970 static int intel_pstate_suspend(struct cpufreq_policy *policy)
971 {
972 struct cpudata *cpu = all_cpu_data[policy->cpu];
973
974 pr_debug("CPU %d suspending\n", cpu->cpu);
975
976 cpu->suspended = true;
977
978 return 0;
979 }
980
981 static int intel_pstate_resume(struct cpufreq_policy *policy)
982 {
983 struct cpudata *cpu = all_cpu_data[policy->cpu];
984
985 pr_debug("CPU %d resuming\n", cpu->cpu);
986
987 /* Only restore if the system default is changed */
988 if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
989 set_power_ctl_ee_state(true);
990 else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
991 set_power_ctl_ee_state(false);
992
993 if (cpu->suspended && hwp_active) {
994 mutex_lock(&intel_pstate_limits_lock);
995
996 /* Re-enable HWP, because "online" has not done that. */
997 intel_pstate_hwp_reenable(cpu);
998
999 mutex_unlock(&intel_pstate_limits_lock);
1000 }
1001
1002 cpu->suspended = false;
1003
1004 return 0;
1005 }
1006
1007 static void intel_pstate_update_policies(void)
1008 {
1009 int cpu;
1010
1011 for_each_possible_cpu(cpu)
1012 cpufreq_update_policy(cpu);
1013 }
1014
1015 static void intel_pstate_update_max_freq(unsigned int cpu)
1016 {
1017 struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1018 struct cpudata *cpudata;
1019
1020 if (!policy)
1021 return;
1022
1023 cpudata = all_cpu_data[cpu];
1024 policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
1025 cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1026
1027 refresh_frequency_limits(policy);
1028
1029 cpufreq_cpu_release(policy);
1030 }
1031
1032 static void intel_pstate_update_limits(unsigned int cpu)
1033 {
1034 mutex_lock(&intel_pstate_driver_lock);
1035
1036 update_turbo_state();
1037 /*
1038 * If turbo has been turned on or off globally, policy limits for
1039 * all CPUs need to be updated to reflect that.
1040 */
1041 if (global.turbo_disabled_mf != global.turbo_disabled) {
1042 global.turbo_disabled_mf = global.turbo_disabled;
1043 arch_set_max_freq_ratio(global.turbo_disabled);
1044 for_each_possible_cpu(cpu)
1045 intel_pstate_update_max_freq(cpu);
1046 } else {
1047 cpufreq_update_policy(cpu);
1048 }
1049
1050 mutex_unlock(&intel_pstate_driver_lock);
1051 }
1052
1053 /************************** sysfs begin ************************/
1054 #define show_one(file_name, object) \
1055 static ssize_t show_##file_name \
1056 (struct kobject *kobj, struct kobj_attribute *attr, char *buf) \
1057 { \
1058 return sprintf(buf, "%u\n", global.object); \
1059 }
1060
1061 static ssize_t intel_pstate_show_status(char *buf);
1062 static int intel_pstate_update_status(const char *buf, size_t size);
1063
1064 static ssize_t show_status(struct kobject *kobj,
1065 struct kobj_attribute *attr, char *buf)
1066 {
1067 ssize_t ret;
1068
1069 mutex_lock(&intel_pstate_driver_lock);
1070 ret = intel_pstate_show_status(buf);
1071 mutex_unlock(&intel_pstate_driver_lock);
1072
1073 return ret;
1074 }
1075
1076 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1077 const char *buf, size_t count)
1078 {
1079 char *p = memchr(buf, '\n', count);
1080 int ret;
1081
1082 mutex_lock(&intel_pstate_driver_lock);
1083 ret = intel_pstate_update_status(buf, p ? p - buf : count);
1084 mutex_unlock(&intel_pstate_driver_lock);
1085
1086 return ret < 0 ? ret : count;
1087 }
1088
1089 static ssize_t show_turbo_pct(struct kobject *kobj,
1090 struct kobj_attribute *attr, char *buf)
1091 {
1092 struct cpudata *cpu;
1093 int total, no_turbo, turbo_pct;
1094 uint32_t turbo_fp;
1095
1096 mutex_lock(&intel_pstate_driver_lock);
1097
1098 if (!intel_pstate_driver) {
1099 mutex_unlock(&intel_pstate_driver_lock);
1100 return -EAGAIN;
1101 }
1102
1103 cpu = all_cpu_data[0];
1104
1105 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1106 no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1107 turbo_fp = div_fp(no_turbo, total);
1108 turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1109
1110 mutex_unlock(&intel_pstate_driver_lock);
1111
1112 return sprintf(buf, "%u\n", turbo_pct);
1113 }
1114
1115 static ssize_t show_num_pstates(struct kobject *kobj,
1116 struct kobj_attribute *attr, char *buf)
1117 {
1118 struct cpudata *cpu;
1119 int total;
1120
1121 mutex_lock(&intel_pstate_driver_lock);
1122
1123 if (!intel_pstate_driver) {
1124 mutex_unlock(&intel_pstate_driver_lock);
1125 return -EAGAIN;
1126 }
1127
1128 cpu = all_cpu_data[0];
1129 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1130
1131 mutex_unlock(&intel_pstate_driver_lock);
1132
1133 return sprintf(buf, "%u\n", total);
1134 }
1135
1136 static ssize_t show_no_turbo(struct kobject *kobj,
1137 struct kobj_attribute *attr, char *buf)
1138 {
1139 ssize_t ret;
1140
1141 mutex_lock(&intel_pstate_driver_lock);
1142
1143 if (!intel_pstate_driver) {
1144 mutex_unlock(&intel_pstate_driver_lock);
1145 return -EAGAIN;
1146 }
1147
1148 update_turbo_state();
1149 if (global.turbo_disabled)
1150 ret = sprintf(buf, "%u\n", global.turbo_disabled);
1151 else
1152 ret = sprintf(buf, "%u\n", global.no_turbo);
1153
1154 mutex_unlock(&intel_pstate_driver_lock);
1155
1156 return ret;
1157 }
1158
1159 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1160 const char *buf, size_t count)
1161 {
1162 unsigned int input;
1163 int ret;
1164
1165 ret = sscanf(buf, "%u", &input);
1166 if (ret != 1)
1167 return -EINVAL;
1168
1169 mutex_lock(&intel_pstate_driver_lock);
1170
1171 if (!intel_pstate_driver) {
1172 mutex_unlock(&intel_pstate_driver_lock);
1173 return -EAGAIN;
1174 }
1175
1176 mutex_lock(&intel_pstate_limits_lock);
1177
1178 update_turbo_state();
1179 if (global.turbo_disabled) {
1180 pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1181 mutex_unlock(&intel_pstate_limits_lock);
1182 mutex_unlock(&intel_pstate_driver_lock);
1183 return -EPERM;
1184 }
1185
1186 global.no_turbo = clamp_t(int, input, 0, 1);
1187
1188 if (global.no_turbo) {
1189 struct cpudata *cpu = all_cpu_data[0];
1190 int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1191
1192 /* Squash the global minimum into the permitted range. */
1193 if (global.min_perf_pct > pct)
1194 global.min_perf_pct = pct;
1195 }
1196
1197 mutex_unlock(&intel_pstate_limits_lock);
1198
1199 intel_pstate_update_policies();
1200
1201 mutex_unlock(&intel_pstate_driver_lock);
1202
1203 return count;
1204 }
1205
1206 static void update_qos_request(enum freq_qos_req_type type)
1207 {
1208 int max_state, turbo_max, freq, i, perf_pct;
1209 struct freq_qos_request *req;
1210 struct cpufreq_policy *policy;
1211
1212 for_each_possible_cpu(i) {
1213 struct cpudata *cpu = all_cpu_data[i];
1214
1215 policy = cpufreq_cpu_get(i);
1216 if (!policy)
1217 continue;
1218
1219 req = policy->driver_data;
1220 cpufreq_cpu_put(policy);
1221
1222 if (!req)
1223 continue;
1224
1225 if (hwp_active)
1226 intel_pstate_get_hwp_max(i, &turbo_max, &max_state);
1227 else
1228 turbo_max = cpu->pstate.turbo_pstate;
1229
1230 if (type == FREQ_QOS_MIN) {
1231 perf_pct = global.min_perf_pct;
1232 } else {
1233 req++;
1234 perf_pct = global.max_perf_pct;
1235 }
1236
1237 freq = DIV_ROUND_UP(turbo_max * perf_pct, 100);
1238 freq *= cpu->pstate.scaling;
1239
1240 if (freq_qos_update_request(req, freq) < 0)
1241 pr_warn("Failed to update freq constraint: CPU%d\n", i);
1242 }
1243 }
1244
1245 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1246 const char *buf, size_t count)
1247 {
1248 unsigned int input;
1249 int ret;
1250
1251 ret = sscanf(buf, "%u", &input);
1252 if (ret != 1)
1253 return -EINVAL;
1254
1255 mutex_lock(&intel_pstate_driver_lock);
1256
1257 if (!intel_pstate_driver) {
1258 mutex_unlock(&intel_pstate_driver_lock);
1259 return -EAGAIN;
1260 }
1261
1262 mutex_lock(&intel_pstate_limits_lock);
1263
1264 global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1265
1266 mutex_unlock(&intel_pstate_limits_lock);
1267
1268 if (intel_pstate_driver == &intel_pstate)
1269 intel_pstate_update_policies();
1270 else
1271 update_qos_request(FREQ_QOS_MAX);
1272
1273 mutex_unlock(&intel_pstate_driver_lock);
1274
1275 return count;
1276 }
1277
1278 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1279 const char *buf, size_t count)
1280 {
1281 unsigned int input;
1282 int ret;
1283
1284 ret = sscanf(buf, "%u", &input);
1285 if (ret != 1)
1286 return -EINVAL;
1287
1288 mutex_lock(&intel_pstate_driver_lock);
1289
1290 if (!intel_pstate_driver) {
1291 mutex_unlock(&intel_pstate_driver_lock);
1292 return -EAGAIN;
1293 }
1294
1295 mutex_lock(&intel_pstate_limits_lock);
1296
1297 global.min_perf_pct = clamp_t(int, input,
1298 min_perf_pct_min(), global.max_perf_pct);
1299
1300 mutex_unlock(&intel_pstate_limits_lock);
1301
1302 if (intel_pstate_driver == &intel_pstate)
1303 intel_pstate_update_policies();
1304 else
1305 update_qos_request(FREQ_QOS_MIN);
1306
1307 mutex_unlock(&intel_pstate_driver_lock);
1308
1309 return count;
1310 }
1311
1312 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1313 struct kobj_attribute *attr, char *buf)
1314 {
1315 return sprintf(buf, "%u\n", hwp_boost);
1316 }
1317
1318 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1319 struct kobj_attribute *b,
1320 const char *buf, size_t count)
1321 {
1322 unsigned int input;
1323 int ret;
1324
1325 ret = kstrtouint(buf, 10, &input);
1326 if (ret)
1327 return ret;
1328
1329 mutex_lock(&intel_pstate_driver_lock);
1330 hwp_boost = !!input;
1331 intel_pstate_update_policies();
1332 mutex_unlock(&intel_pstate_driver_lock);
1333
1334 return count;
1335 }
1336
1337 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1338 char *buf)
1339 {
1340 u64 power_ctl;
1341 int enable;
1342
1343 rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1344 enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1345 return sprintf(buf, "%d\n", !enable);
1346 }
1347
1348 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1349 const char *buf, size_t count)
1350 {
1351 bool input;
1352 int ret;
1353
1354 ret = kstrtobool(buf, &input);
1355 if (ret)
1356 return ret;
1357
1358 set_power_ctl_ee_state(input);
1359
1360 return count;
1361 }
1362
1363 show_one(max_perf_pct, max_perf_pct);
1364 show_one(min_perf_pct, min_perf_pct);
1365
1366 define_one_global_rw(status);
1367 define_one_global_rw(no_turbo);
1368 define_one_global_rw(max_perf_pct);
1369 define_one_global_rw(min_perf_pct);
1370 define_one_global_ro(turbo_pct);
1371 define_one_global_ro(num_pstates);
1372 define_one_global_rw(hwp_dynamic_boost);
1373 define_one_global_rw(energy_efficiency);
1374
1375 static struct attribute *intel_pstate_attributes[] = {
1376 &status.attr,
1377 &no_turbo.attr,
1378 &turbo_pct.attr,
1379 &num_pstates.attr,
1380 NULL
1381 };
1382
1383 static const struct attribute_group intel_pstate_attr_group = {
1384 .attrs = intel_pstate_attributes,
1385 };
1386
1387 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1388
1389 static struct kobject *intel_pstate_kobject;
1390
1391 static void __init intel_pstate_sysfs_expose_params(void)
1392 {
1393 int rc;
1394
1395 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1396 &cpu_subsys.dev_root->kobj);
1397 if (WARN_ON(!intel_pstate_kobject))
1398 return;
1399
1400 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1401 if (WARN_ON(rc))
1402 return;
1403
1404 /*
1405 * If per cpu limits are enforced there are no global limits, so
1406 * return without creating max/min_perf_pct attributes
1407 */
1408 if (per_cpu_limits)
1409 return;
1410
1411 rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1412 WARN_ON(rc);
1413
1414 rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1415 WARN_ON(rc);
1416
1417 if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1418 rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1419 WARN_ON(rc);
1420 }
1421 }
1422
1423 static void __init intel_pstate_sysfs_remove(void)
1424 {
1425 if (!intel_pstate_kobject)
1426 return;
1427
1428 sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1429
1430 if (!per_cpu_limits) {
1431 sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1432 sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1433
1434 if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1435 sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1436 }
1437
1438 kobject_put(intel_pstate_kobject);
1439 }
1440
1441 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1442 {
1443 int rc;
1444
1445 if (!hwp_active)
1446 return;
1447
1448 rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1449 WARN_ON_ONCE(rc);
1450 }
1451
1452 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1453 {
1454 if (!hwp_active)
1455 return;
1456
1457 sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1458 }
1459
1460 /************************** sysfs end ************************/
1461
1462 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1463 {
1464 /* First disable HWP notification interrupt as we don't process them */
1465 if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1466 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1467
1468 wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1469 if (cpudata->epp_default == -EINVAL)
1470 cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1471 }
1472
1473 static int atom_get_min_pstate(void)
1474 {
1475 u64 value;
1476
1477 rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1478 return (value >> 8) & 0x7F;
1479 }
1480
1481 static int atom_get_max_pstate(void)
1482 {
1483 u64 value;
1484
1485 rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1486 return (value >> 16) & 0x7F;
1487 }
1488
1489 static int atom_get_turbo_pstate(void)
1490 {
1491 u64 value;
1492
1493 rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1494 return value & 0x7F;
1495 }
1496
1497 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1498 {
1499 u64 val;
1500 int32_t vid_fp;
1501 u32 vid;
1502
1503 val = (u64)pstate << 8;
1504 if (global.no_turbo && !global.turbo_disabled)
1505 val |= (u64)1 << 32;
1506
1507 vid_fp = cpudata->vid.min + mul_fp(
1508 int_tofp(pstate - cpudata->pstate.min_pstate),
1509 cpudata->vid.ratio);
1510
1511 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1512 vid = ceiling_fp(vid_fp);
1513
1514 if (pstate > cpudata->pstate.max_pstate)
1515 vid = cpudata->vid.turbo;
1516
1517 return val | vid;
1518 }
1519
1520 static int silvermont_get_scaling(void)
1521 {
1522 u64 value;
1523 int i;
1524 /* Defined in Table 35-6 from SDM (Sept 2015) */
1525 static int silvermont_freq_table[] = {
1526 83300, 100000, 133300, 116700, 80000};
1527
1528 rdmsrl(MSR_FSB_FREQ, value);
1529 i = value & 0x7;
1530 WARN_ON(i > 4);
1531
1532 return silvermont_freq_table[i];
1533 }
1534
1535 static int airmont_get_scaling(void)
1536 {
1537 u64 value;
1538 int i;
1539 /* Defined in Table 35-10 from SDM (Sept 2015) */
1540 static int airmont_freq_table[] = {
1541 83300, 100000, 133300, 116700, 80000,
1542 93300, 90000, 88900, 87500};
1543
1544 rdmsrl(MSR_FSB_FREQ, value);
1545 i = value & 0xF;
1546 WARN_ON(i > 8);
1547
1548 return airmont_freq_table[i];
1549 }
1550
1551 static void atom_get_vid(struct cpudata *cpudata)
1552 {
1553 u64 value;
1554
1555 rdmsrl(MSR_ATOM_CORE_VIDS, value);
1556 cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1557 cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1558 cpudata->vid.ratio = div_fp(
1559 cpudata->vid.max - cpudata->vid.min,
1560 int_tofp(cpudata->pstate.max_pstate -
1561 cpudata->pstate.min_pstate));
1562
1563 rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1564 cpudata->vid.turbo = value & 0x7f;
1565 }
1566
1567 static int core_get_min_pstate(void)
1568 {
1569 u64 value;
1570
1571 rdmsrl(MSR_PLATFORM_INFO, value);
1572 return (value >> 40) & 0xFF;
1573 }
1574
1575 static int core_get_max_pstate_physical(void)
1576 {
1577 u64 value;
1578
1579 rdmsrl(MSR_PLATFORM_INFO, value);
1580 return (value >> 8) & 0xFF;
1581 }
1582
1583 static int core_get_tdp_ratio(u64 plat_info)
1584 {
1585 /* Check how many TDP levels present */
1586 if (plat_info & 0x600000000) {
1587 u64 tdp_ctrl;
1588 u64 tdp_ratio;
1589 int tdp_msr;
1590 int err;
1591
1592 /* Get the TDP level (0, 1, 2) to get ratios */
1593 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1594 if (err)
1595 return err;
1596
1597 /* TDP MSR are continuous starting at 0x648 */
1598 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1599 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1600 if (err)
1601 return err;
1602
1603 /* For level 1 and 2, bits[23:16] contain the ratio */
1604 if (tdp_ctrl & 0x03)
1605 tdp_ratio >>= 16;
1606
1607 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1608 pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1609
1610 return (int)tdp_ratio;
1611 }
1612
1613 return -ENXIO;
1614 }
1615
1616 static int core_get_max_pstate(void)
1617 {
1618 u64 tar;
1619 u64 plat_info;
1620 int max_pstate;
1621 int tdp_ratio;
1622 int err;
1623
1624 rdmsrl(MSR_PLATFORM_INFO, plat_info);
1625 max_pstate = (plat_info >> 8) & 0xFF;
1626
1627 tdp_ratio = core_get_tdp_ratio(plat_info);
1628 if (tdp_ratio <= 0)
1629 return max_pstate;
1630
1631 if (hwp_active) {
1632 /* Turbo activation ratio is not used on HWP platforms */
1633 return tdp_ratio;
1634 }
1635
1636 err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1637 if (!err) {
1638 int tar_levels;
1639
1640 /* Do some sanity checking for safety */
1641 tar_levels = tar & 0xff;
1642 if (tdp_ratio - 1 == tar_levels) {
1643 max_pstate = tar_levels;
1644 pr_debug("max_pstate=TAC %x\n", max_pstate);
1645 }
1646 }
1647
1648 return max_pstate;
1649 }
1650
1651 static int core_get_turbo_pstate(void)
1652 {
1653 u64 value;
1654 int nont, ret;
1655
1656 rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1657 nont = core_get_max_pstate();
1658 ret = (value) & 255;
1659 if (ret <= nont)
1660 ret = nont;
1661 return ret;
1662 }
1663
1664 static inline int core_get_scaling(void)
1665 {
1666 return 100000;
1667 }
1668
1669 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1670 {
1671 u64 val;
1672
1673 val = (u64)pstate << 8;
1674 if (global.no_turbo && !global.turbo_disabled)
1675 val |= (u64)1 << 32;
1676
1677 return val;
1678 }
1679
1680 static int knl_get_aperf_mperf_shift(void)
1681 {
1682 return 10;
1683 }
1684
1685 static int knl_get_turbo_pstate(void)
1686 {
1687 u64 value;
1688 int nont, ret;
1689
1690 rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1691 nont = core_get_max_pstate();
1692 ret = (((value) >> 8) & 0xFF);
1693 if (ret <= nont)
1694 ret = nont;
1695 return ret;
1696 }
1697
1698 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1699 {
1700 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1701 cpu->pstate.current_pstate = pstate;
1702 /*
1703 * Generally, there is no guarantee that this code will always run on
1704 * the CPU being updated, so force the register update to run on the
1705 * right CPU.
1706 */
1707 wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1708 pstate_funcs.get_val(cpu, pstate));
1709 }
1710
1711 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1712 {
1713 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1714 }
1715
1716 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1717 {
1718 int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1719
1720 update_turbo_state();
1721 intel_pstate_set_pstate(cpu, pstate);
1722 }
1723
1724 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1725 {
1726 cpu->pstate.min_pstate = pstate_funcs.get_min();
1727 cpu->pstate.max_pstate = pstate_funcs.get_max();
1728 cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1729 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1730 cpu->pstate.scaling = pstate_funcs.get_scaling();
1731 cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1732
1733 if (hwp_active && !hwp_mode_bdw) {
1734 unsigned int phy_max, current_max;
1735
1736 intel_pstate_get_hwp_max(cpu->cpu, &phy_max, &current_max);
1737 cpu->pstate.turbo_freq = phy_max * cpu->pstate.scaling;
1738 cpu->pstate.turbo_pstate = phy_max;
1739 } else {
1740 cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1741 }
1742
1743 if (pstate_funcs.get_aperf_mperf_shift)
1744 cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
1745
1746 if (pstate_funcs.get_vid)
1747 pstate_funcs.get_vid(cpu);
1748
1749 intel_pstate_set_min_pstate(cpu);
1750 }
1751
1752 /*
1753 * Long hold time will keep high perf limits for long time,
1754 * which negatively impacts perf/watt for some workloads,
1755 * like specpower. 3ms is based on experiements on some
1756 * workoads.
1757 */
1758 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
1759
1760 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
1761 {
1762 u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
1763 u32 max_limit = (hwp_req & 0xff00) >> 8;
1764 u32 min_limit = (hwp_req & 0xff);
1765 u32 boost_level1;
1766
1767 /*
1768 * Cases to consider (User changes via sysfs or boot time):
1769 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
1770 * No boost, return.
1771 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
1772 * Should result in one level boost only for P0.
1773 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
1774 * Should result in two level boost:
1775 * (min + p1)/2 and P1.
1776 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
1777 * Should result in three level boost:
1778 * (min + p1)/2, P1 and P0.
1779 */
1780
1781 /* If max and min are equal or already at max, nothing to boost */
1782 if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
1783 return;
1784
1785 if (!cpu->hwp_boost_min)
1786 cpu->hwp_boost_min = min_limit;
1787
1788 /* level at half way mark between min and guranteed */
1789 boost_level1 = (HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) + min_limit) >> 1;
1790
1791 if (cpu->hwp_boost_min < boost_level1)
1792 cpu->hwp_boost_min = boost_level1;
1793 else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1794 cpu->hwp_boost_min = HWP_GUARANTEED_PERF(cpu->hwp_cap_cached);
1795 else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) &&
1796 max_limit != HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1797 cpu->hwp_boost_min = max_limit;
1798 else
1799 return;
1800
1801 hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
1802 wrmsrl(MSR_HWP_REQUEST, hwp_req);
1803 cpu->last_update = cpu->sample.time;
1804 }
1805
1806 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
1807 {
1808 if (cpu->hwp_boost_min) {
1809 bool expired;
1810
1811 /* Check if we are idle for hold time to boost down */
1812 expired = time_after64(cpu->sample.time, cpu->last_update +
1813 hwp_boost_hold_time_ns);
1814 if (expired) {
1815 wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
1816 cpu->hwp_boost_min = 0;
1817 }
1818 }
1819 cpu->last_update = cpu->sample.time;
1820 }
1821
1822 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
1823 u64 time)
1824 {
1825 cpu->sample.time = time;
1826
1827 if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
1828 bool do_io = false;
1829
1830 cpu->sched_flags = 0;
1831 /*
1832 * Set iowait_boost flag and update time. Since IO WAIT flag
1833 * is set all the time, we can't just conclude that there is
1834 * some IO bound activity is scheduled on this CPU with just
1835 * one occurrence. If we receive at least two in two
1836 * consecutive ticks, then we treat as boost candidate.
1837 */
1838 if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
1839 do_io = true;
1840
1841 cpu->last_io_update = time;
1842
1843 if (do_io)
1844 intel_pstate_hwp_boost_up(cpu);
1845
1846 } else {
1847 intel_pstate_hwp_boost_down(cpu);
1848 }
1849 }
1850
1851 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
1852 u64 time, unsigned int flags)
1853 {
1854 struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1855
1856 cpu->sched_flags |= flags;
1857
1858 if (smp_processor_id() == cpu->cpu)
1859 intel_pstate_update_util_hwp_local(cpu, time);
1860 }
1861
1862 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1863 {
1864 struct sample *sample = &cpu->sample;
1865
1866 sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1867 }
1868
1869 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1870 {
1871 u64 aperf, mperf;
1872 unsigned long flags;
1873 u64 tsc;
1874
1875 local_irq_save(flags);
1876 rdmsrl(MSR_IA32_APERF, aperf);
1877 rdmsrl(MSR_IA32_MPERF, mperf);
1878 tsc = rdtsc();
1879 if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1880 local_irq_restore(flags);
1881 return false;
1882 }
1883 local_irq_restore(flags);
1884
1885 cpu->last_sample_time = cpu->sample.time;
1886 cpu->sample.time = time;
1887 cpu->sample.aperf = aperf;
1888 cpu->sample.mperf = mperf;
1889 cpu->sample.tsc = tsc;
1890 cpu->sample.aperf -= cpu->prev_aperf;
1891 cpu->sample.mperf -= cpu->prev_mperf;
1892 cpu->sample.tsc -= cpu->prev_tsc;
1893
1894 cpu->prev_aperf = aperf;
1895 cpu->prev_mperf = mperf;
1896 cpu->prev_tsc = tsc;
1897 /*
1898 * First time this function is invoked in a given cycle, all of the
1899 * previous sample data fields are equal to zero or stale and they must
1900 * be populated with meaningful numbers for things to work, so assume
1901 * that sample.time will always be reset before setting the utilization
1902 * update hook and make the caller skip the sample then.
1903 */
1904 if (cpu->last_sample_time) {
1905 intel_pstate_calc_avg_perf(cpu);
1906 return true;
1907 }
1908 return false;
1909 }
1910
1911 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1912 {
1913 return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
1914 }
1915
1916 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1917 {
1918 return mul_ext_fp(cpu->pstate.max_pstate_physical,
1919 cpu->sample.core_avg_perf);
1920 }
1921
1922 static inline int32_t get_target_pstate(struct cpudata *cpu)
1923 {
1924 struct sample *sample = &cpu->sample;
1925 int32_t busy_frac;
1926 int target, avg_pstate;
1927
1928 busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
1929 sample->tsc);
1930
1931 if (busy_frac < cpu->iowait_boost)
1932 busy_frac = cpu->iowait_boost;
1933
1934 sample->busy_scaled = busy_frac * 100;
1935
1936 target = global.no_turbo || global.turbo_disabled ?
1937 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1938 target += target >> 2;
1939 target = mul_fp(target, busy_frac);
1940 if (target < cpu->pstate.min_pstate)
1941 target = cpu->pstate.min_pstate;
1942
1943 /*
1944 * If the average P-state during the previous cycle was higher than the
1945 * current target, add 50% of the difference to the target to reduce
1946 * possible performance oscillations and offset possible performance
1947 * loss related to moving the workload from one CPU to another within
1948 * a package/module.
1949 */
1950 avg_pstate = get_avg_pstate(cpu);
1951 if (avg_pstate > target)
1952 target += (avg_pstate - target) >> 1;
1953
1954 return target;
1955 }
1956
1957 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1958 {
1959 int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
1960 int max_pstate = max(min_pstate, cpu->max_perf_ratio);
1961
1962 return clamp_t(int, pstate, min_pstate, max_pstate);
1963 }
1964
1965 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1966 {
1967 if (pstate == cpu->pstate.current_pstate)
1968 return;
1969
1970 cpu->pstate.current_pstate = pstate;
1971 wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1972 }
1973
1974 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
1975 {
1976 int from = cpu->pstate.current_pstate;
1977 struct sample *sample;
1978 int target_pstate;
1979
1980 update_turbo_state();
1981
1982 target_pstate = get_target_pstate(cpu);
1983 target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1984 trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1985 intel_pstate_update_pstate(cpu, target_pstate);
1986
1987 sample = &cpu->sample;
1988 trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1989 fp_toint(sample->busy_scaled),
1990 from,
1991 cpu->pstate.current_pstate,
1992 sample->mperf,
1993 sample->aperf,
1994 sample->tsc,
1995 get_avg_frequency(cpu),
1996 fp_toint(cpu->iowait_boost * 100));
1997 }
1998
1999 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2000 unsigned int flags)
2001 {
2002 struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2003 u64 delta_ns;
2004
2005 /* Don't allow remote callbacks */
2006 if (smp_processor_id() != cpu->cpu)
2007 return;
2008
2009 delta_ns = time - cpu->last_update;
2010 if (flags & SCHED_CPUFREQ_IOWAIT) {
2011 /* Start over if the CPU may have been idle. */
2012 if (delta_ns > TICK_NSEC) {
2013 cpu->iowait_boost = ONE_EIGHTH_FP;
2014 } else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2015 cpu->iowait_boost <<= 1;
2016 if (cpu->iowait_boost > int_tofp(1))
2017 cpu->iowait_boost = int_tofp(1);
2018 } else {
2019 cpu->iowait_boost = ONE_EIGHTH_FP;
2020 }
2021 } else if (cpu->iowait_boost) {
2022 /* Clear iowait_boost if the CPU may have been idle. */
2023 if (delta_ns > TICK_NSEC)
2024 cpu->iowait_boost = 0;
2025 else
2026 cpu->iowait_boost >>= 1;
2027 }
2028 cpu->last_update = time;
2029 delta_ns = time - cpu->sample.time;
2030 if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2031 return;
2032
2033 if (intel_pstate_sample(cpu, time))
2034 intel_pstate_adjust_pstate(cpu);
2035 }
2036
2037 static struct pstate_funcs core_funcs = {
2038 .get_max = core_get_max_pstate,
2039 .get_max_physical = core_get_max_pstate_physical,
2040 .get_min = core_get_min_pstate,
2041 .get_turbo = core_get_turbo_pstate,
2042 .get_scaling = core_get_scaling,
2043 .get_val = core_get_val,
2044 };
2045
2046 static const struct pstate_funcs silvermont_funcs = {
2047 .get_max = atom_get_max_pstate,
2048 .get_max_physical = atom_get_max_pstate,
2049 .get_min = atom_get_min_pstate,
2050 .get_turbo = atom_get_turbo_pstate,
2051 .get_val = atom_get_val,
2052 .get_scaling = silvermont_get_scaling,
2053 .get_vid = atom_get_vid,
2054 };
2055
2056 static const struct pstate_funcs airmont_funcs = {
2057 .get_max = atom_get_max_pstate,
2058 .get_max_physical = atom_get_max_pstate,
2059 .get_min = atom_get_min_pstate,
2060 .get_turbo = atom_get_turbo_pstate,
2061 .get_val = atom_get_val,
2062 .get_scaling = airmont_get_scaling,
2063 .get_vid = atom_get_vid,
2064 };
2065
2066 static const struct pstate_funcs knl_funcs = {
2067 .get_max = core_get_max_pstate,
2068 .get_max_physical = core_get_max_pstate_physical,
2069 .get_min = core_get_min_pstate,
2070 .get_turbo = knl_get_turbo_pstate,
2071 .get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2072 .get_scaling = core_get_scaling,
2073 .get_val = core_get_val,
2074 };
2075
2076 #define X86_MATCH(model, policy) \
2077 X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2078 X86_FEATURE_APERFMPERF, &policy)
2079
2080 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2081 X86_MATCH(SANDYBRIDGE, core_funcs),
2082 X86_MATCH(SANDYBRIDGE_X, core_funcs),
2083 X86_MATCH(ATOM_SILVERMONT, silvermont_funcs),
2084 X86_MATCH(IVYBRIDGE, core_funcs),
2085 X86_MATCH(HASWELL, core_funcs),
2086 X86_MATCH(BROADWELL, core_funcs),
2087 X86_MATCH(IVYBRIDGE_X, core_funcs),
2088 X86_MATCH(HASWELL_X, core_funcs),
2089 X86_MATCH(HASWELL_L, core_funcs),
2090 X86_MATCH(HASWELL_G, core_funcs),
2091 X86_MATCH(BROADWELL_G, core_funcs),
2092 X86_MATCH(ATOM_AIRMONT, airmont_funcs),
2093 X86_MATCH(SKYLAKE_L, core_funcs),
2094 X86_MATCH(BROADWELL_X, core_funcs),
2095 X86_MATCH(SKYLAKE, core_funcs),
2096 X86_MATCH(BROADWELL_D, core_funcs),
2097 X86_MATCH(XEON_PHI_KNL, knl_funcs),
2098 X86_MATCH(XEON_PHI_KNM, knl_funcs),
2099 X86_MATCH(ATOM_GOLDMONT, core_funcs),
2100 X86_MATCH(ATOM_GOLDMONT_PLUS, core_funcs),
2101 X86_MATCH(SKYLAKE_X, core_funcs),
2102 {}
2103 };
2104 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2105
2106 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2107 X86_MATCH(BROADWELL_D, core_funcs),
2108 X86_MATCH(BROADWELL_X, core_funcs),
2109 X86_MATCH(SKYLAKE_X, core_funcs),
2110 {}
2111 };
2112
2113 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2114 X86_MATCH(KABYLAKE, core_funcs),
2115 {}
2116 };
2117
2118 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
2119 X86_MATCH(SKYLAKE_X, core_funcs),
2120 X86_MATCH(SKYLAKE, core_funcs),
2121 {}
2122 };
2123
2124 static int intel_pstate_init_cpu(unsigned int cpunum)
2125 {
2126 struct cpudata *cpu;
2127
2128 cpu = all_cpu_data[cpunum];
2129
2130 if (!cpu) {
2131 cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2132 if (!cpu)
2133 return -ENOMEM;
2134
2135 all_cpu_data[cpunum] = cpu;
2136
2137 cpu->cpu = cpunum;
2138
2139 cpu->epp_default = -EINVAL;
2140
2141 if (hwp_active) {
2142 const struct x86_cpu_id *id;
2143
2144 intel_pstate_hwp_enable(cpu);
2145
2146 id = x86_match_cpu(intel_pstate_hwp_boost_ids);
2147 if (id && intel_pstate_acpi_pm_profile_server())
2148 hwp_boost = true;
2149 }
2150 } else if (hwp_active) {
2151 /*
2152 * Re-enable HWP in case this happens after a resume from ACPI
2153 * S3 if the CPU was offline during the whole system/resume
2154 * cycle.
2155 */
2156 intel_pstate_hwp_reenable(cpu);
2157 }
2158
2159 cpu->epp_powersave = -EINVAL;
2160 cpu->epp_policy = 0;
2161
2162 intel_pstate_get_cpu_pstates(cpu);
2163
2164 pr_debug("controlling: cpu %d\n", cpunum);
2165
2166 return 0;
2167 }
2168
2169 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2170 {
2171 struct cpudata *cpu = all_cpu_data[cpu_num];
2172
2173 if (hwp_active && !hwp_boost)
2174 return;
2175
2176 if (cpu->update_util_set)
2177 return;
2178
2179 /* Prevent intel_pstate_update_util() from using stale data. */
2180 cpu->sample.time = 0;
2181 cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2182 (hwp_active ?
2183 intel_pstate_update_util_hwp :
2184 intel_pstate_update_util));
2185 cpu->update_util_set = true;
2186 }
2187
2188 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2189 {
2190 struct cpudata *cpu_data = all_cpu_data[cpu];
2191
2192 if (!cpu_data->update_util_set)
2193 return;
2194
2195 cpufreq_remove_update_util_hook(cpu);
2196 cpu_data->update_util_set = false;
2197 synchronize_rcu();
2198 }
2199
2200 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2201 {
2202 return global.turbo_disabled || global.no_turbo ?
2203 cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2204 }
2205
2206 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2207 unsigned int policy_min,
2208 unsigned int policy_max)
2209 {
2210 int32_t max_policy_perf, min_policy_perf;
2211 int max_state, turbo_max;
2212 int max_freq;
2213
2214 /*
2215 * HWP needs some special consideration, because on BDX the
2216 * HWP_REQUEST uses abstract value to represent performance
2217 * rather than pure ratios.
2218 */
2219 if (hwp_active) {
2220 intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state);
2221 } else {
2222 max_state = global.no_turbo || global.turbo_disabled ?
2223 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2224 turbo_max = cpu->pstate.turbo_pstate;
2225 }
2226 max_freq = max_state * cpu->pstate.scaling;
2227
2228 max_policy_perf = max_state * policy_max / max_freq;
2229 if (policy_max == policy_min) {
2230 min_policy_perf = max_policy_perf;
2231 } else {
2232 min_policy_perf = max_state * policy_min / max_freq;
2233 min_policy_perf = clamp_t(int32_t, min_policy_perf,
2234 0, max_policy_perf);
2235 }
2236
2237 pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n",
2238 cpu->cpu, max_state, min_policy_perf, max_policy_perf);
2239
2240 /* Normalize user input to [min_perf, max_perf] */
2241 if (per_cpu_limits) {
2242 cpu->min_perf_ratio = min_policy_perf;
2243 cpu->max_perf_ratio = max_policy_perf;
2244 } else {
2245 int32_t global_min, global_max;
2246
2247 /* Global limits are in percent of the maximum turbo P-state. */
2248 global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2249 global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2250 global_min = clamp_t(int32_t, global_min, 0, global_max);
2251
2252 pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2253 global_min, global_max);
2254
2255 cpu->min_perf_ratio = max(min_policy_perf, global_min);
2256 cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2257 cpu->max_perf_ratio = min(max_policy_perf, global_max);
2258 cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2259
2260 /* Make sure min_perf <= max_perf */
2261 cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2262 cpu->max_perf_ratio);
2263
2264 }
2265 pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2266 cpu->max_perf_ratio,
2267 cpu->min_perf_ratio);
2268 }
2269
2270 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2271 {
2272 struct cpudata *cpu;
2273
2274 if (!policy->cpuinfo.max_freq)
2275 return -ENODEV;
2276
2277 pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2278 policy->cpuinfo.max_freq, policy->max);
2279
2280 cpu = all_cpu_data[policy->cpu];
2281 cpu->policy = policy->policy;
2282
2283 mutex_lock(&intel_pstate_limits_lock);
2284
2285 intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2286
2287 if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2288 /*
2289 * NOHZ_FULL CPUs need this as the governor callback may not
2290 * be invoked on them.
2291 */
2292 intel_pstate_clear_update_util_hook(policy->cpu);
2293 intel_pstate_max_within_limits(cpu);
2294 } else {
2295 intel_pstate_set_update_util_hook(policy->cpu);
2296 }
2297
2298 if (hwp_active) {
2299 /*
2300 * When hwp_boost was active before and dynamically it
2301 * was turned off, in that case we need to clear the
2302 * update util hook.
2303 */
2304 if (!hwp_boost)
2305 intel_pstate_clear_update_util_hook(policy->cpu);
2306 intel_pstate_hwp_set(policy->cpu);
2307 }
2308
2309 mutex_unlock(&intel_pstate_limits_lock);
2310
2311 return 0;
2312 }
2313
2314 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2315 struct cpufreq_policy_data *policy)
2316 {
2317 if (!hwp_active &&
2318 cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2319 policy->max < policy->cpuinfo.max_freq &&
2320 policy->max > cpu->pstate.max_freq) {
2321 pr_debug("policy->max > max non turbo frequency\n");
2322 policy->max = policy->cpuinfo.max_freq;
2323 }
2324 }
2325
2326 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2327 struct cpufreq_policy_data *policy)
2328 {
2329 int max_freq;
2330
2331 update_turbo_state();
2332 if (hwp_active) {
2333 int max_state, turbo_max;
2334
2335 intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state);
2336 max_freq = max_state * cpu->pstate.scaling;
2337 } else {
2338 max_freq = intel_pstate_get_max_freq(cpu);
2339 }
2340 cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2341
2342 intel_pstate_adjust_policy_max(cpu, policy);
2343 }
2344
2345 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2346 {
2347 intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2348
2349 return 0;
2350 }
2351
2352 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2353 {
2354 struct cpudata *cpu = all_cpu_data[policy->cpu];
2355
2356 pr_debug("CPU %d going offline\n", cpu->cpu);
2357
2358 if (cpu->suspended)
2359 return 0;
2360
2361 /*
2362 * If the CPU is an SMT thread and it goes offline with the performance
2363 * settings different from the minimum, it will prevent its sibling
2364 * from getting to lower performance levels, so force the minimum
2365 * performance on CPU offline to prevent that from happening.
2366 */
2367 if (hwp_active)
2368 intel_pstate_hwp_offline(cpu);
2369 else
2370 intel_pstate_set_min_pstate(cpu);
2371
2372 intel_pstate_exit_perf_limits(policy);
2373
2374 return 0;
2375 }
2376
2377 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2378 {
2379 struct cpudata *cpu = all_cpu_data[policy->cpu];
2380
2381 pr_debug("CPU %d going online\n", cpu->cpu);
2382
2383 intel_pstate_init_acpi_perf_limits(policy);
2384
2385 if (hwp_active) {
2386 /*
2387 * Re-enable HWP and clear the "suspended" flag to let "resume"
2388 * know that it need not do that.
2389 */
2390 intel_pstate_hwp_reenable(cpu);
2391 cpu->suspended = false;
2392 }
2393
2394 return 0;
2395 }
2396
2397 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2398 {
2399 pr_debug("CPU %d stopping\n", policy->cpu);
2400
2401 intel_pstate_clear_update_util_hook(policy->cpu);
2402 }
2403
2404 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2405 {
2406 pr_debug("CPU %d exiting\n", policy->cpu);
2407
2408 policy->fast_switch_possible = false;
2409
2410 return 0;
2411 }
2412
2413 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2414 {
2415 struct cpudata *cpu;
2416 int rc;
2417
2418 rc = intel_pstate_init_cpu(policy->cpu);
2419 if (rc)
2420 return rc;
2421
2422 cpu = all_cpu_data[policy->cpu];
2423
2424 cpu->max_perf_ratio = 0xFF;
2425 cpu->min_perf_ratio = 0;
2426
2427 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2428 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2429
2430 /* cpuinfo and default policy values */
2431 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2432 update_turbo_state();
2433 global.turbo_disabled_mf = global.turbo_disabled;
2434 policy->cpuinfo.max_freq = global.turbo_disabled ?
2435 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2436 policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2437
2438 if (hwp_active) {
2439 unsigned int max_freq;
2440
2441 max_freq = global.turbo_disabled ?
2442 cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2443 if (max_freq < policy->cpuinfo.max_freq)
2444 policy->cpuinfo.max_freq = max_freq;
2445 }
2446
2447 intel_pstate_init_acpi_perf_limits(policy);
2448
2449 policy->fast_switch_possible = true;
2450
2451 return 0;
2452 }
2453
2454 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2455 {
2456 int ret = __intel_pstate_cpu_init(policy);
2457
2458 if (ret)
2459 return ret;
2460
2461 /*
2462 * Set the policy to powersave to provide a valid fallback value in case
2463 * the default cpufreq governor is neither powersave nor performance.
2464 */
2465 policy->policy = CPUFREQ_POLICY_POWERSAVE;
2466
2467 if (hwp_active) {
2468 struct cpudata *cpu = all_cpu_data[policy->cpu];
2469
2470 cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2471 }
2472
2473 return 0;
2474 }
2475
2476 static struct cpufreq_driver intel_pstate = {
2477 .flags = CPUFREQ_CONST_LOOPS,
2478 .verify = intel_pstate_verify_policy,
2479 .setpolicy = intel_pstate_set_policy,
2480 .suspend = intel_pstate_suspend,
2481 .resume = intel_pstate_resume,
2482 .init = intel_pstate_cpu_init,
2483 .exit = intel_pstate_cpu_exit,
2484 .stop_cpu = intel_pstate_stop_cpu,
2485 .offline = intel_pstate_cpu_offline,
2486 .online = intel_pstate_cpu_online,
2487 .update_limits = intel_pstate_update_limits,
2488 .name = "intel_pstate",
2489 };
2490
2491 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2492 {
2493 struct cpudata *cpu = all_cpu_data[policy->cpu];
2494
2495 intel_pstate_verify_cpu_policy(cpu, policy);
2496 intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2497
2498 return 0;
2499 }
2500
2501 /* Use of trace in passive mode:
2502 *
2503 * In passive mode the trace core_busy field (also known as the
2504 * performance field, and lablelled as such on the graphs; also known as
2505 * core_avg_perf) is not needed and so is re-assigned to indicate if the
2506 * driver call was via the normal or fast switch path. Various graphs
2507 * output from the intel_pstate_tracer.py utility that include core_busy
2508 * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2509 * so we use 10 to indicate the the normal path through the driver, and
2510 * 90 to indicate the fast switch path through the driver.
2511 * The scaled_busy field is not used, and is set to 0.
2512 */
2513
2514 #define INTEL_PSTATE_TRACE_TARGET 10
2515 #define INTEL_PSTATE_TRACE_FAST_SWITCH 90
2516
2517 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2518 {
2519 struct sample *sample;
2520
2521 if (!trace_pstate_sample_enabled())
2522 return;
2523
2524 if (!intel_pstate_sample(cpu, ktime_get()))
2525 return;
2526
2527 sample = &cpu->sample;
2528 trace_pstate_sample(trace_type,
2529 0,
2530 old_pstate,
2531 cpu->pstate.current_pstate,
2532 sample->mperf,
2533 sample->aperf,
2534 sample->tsc,
2535 get_avg_frequency(cpu),
2536 fp_toint(cpu->iowait_boost * 100));
2537 }
2538
2539 static void intel_cpufreq_adjust_hwp(struct cpudata *cpu, u32 min, u32 max,
2540 u32 desired, bool fast_switch)
2541 {
2542 u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2543
2544 value &= ~HWP_MIN_PERF(~0L);
2545 value |= HWP_MIN_PERF(min);
2546
2547 value &= ~HWP_MAX_PERF(~0L);
2548 value |= HWP_MAX_PERF(max);
2549
2550 value &= ~HWP_DESIRED_PERF(~0L);
2551 value |= HWP_DESIRED_PERF(desired);
2552
2553 if (value == prev)
2554 return;
2555
2556 WRITE_ONCE(cpu->hwp_req_cached, value);
2557 if (fast_switch)
2558 wrmsrl(MSR_HWP_REQUEST, value);
2559 else
2560 wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2561 }
2562
2563 static void intel_cpufreq_adjust_perf_ctl(struct cpudata *cpu,
2564 u32 target_pstate, bool fast_switch)
2565 {
2566 if (fast_switch)
2567 wrmsrl(MSR_IA32_PERF_CTL,
2568 pstate_funcs.get_val(cpu, target_pstate));
2569 else
2570 wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2571 pstate_funcs.get_val(cpu, target_pstate));
2572 }
2573
2574 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2575 int target_pstate, bool fast_switch)
2576 {
2577 struct cpudata *cpu = all_cpu_data[policy->cpu];
2578 int old_pstate = cpu->pstate.current_pstate;
2579
2580 target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2581 if (hwp_active) {
2582 int max_pstate = policy->strict_target ?
2583 target_pstate : cpu->max_perf_ratio;
2584
2585 intel_cpufreq_adjust_hwp(cpu, target_pstate, max_pstate, 0,
2586 fast_switch);
2587 } else if (target_pstate != old_pstate) {
2588 intel_cpufreq_adjust_perf_ctl(cpu, target_pstate, fast_switch);
2589 }
2590
2591 cpu->pstate.current_pstate = target_pstate;
2592
2593 intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2594 INTEL_PSTATE_TRACE_TARGET, old_pstate);
2595
2596 return target_pstate;
2597 }
2598
2599 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2600 unsigned int target_freq,
2601 unsigned int relation)
2602 {
2603 struct cpudata *cpu = all_cpu_data[policy->cpu];
2604 struct cpufreq_freqs freqs;
2605 int target_pstate;
2606
2607 update_turbo_state();
2608
2609 freqs.old = policy->cur;
2610 freqs.new = target_freq;
2611
2612 cpufreq_freq_transition_begin(policy, &freqs);
2613
2614 switch (relation) {
2615 case CPUFREQ_RELATION_L:
2616 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2617 break;
2618 case CPUFREQ_RELATION_H:
2619 target_pstate = freqs.new / cpu->pstate.scaling;
2620 break;
2621 default:
2622 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2623 break;
2624 }
2625
2626 target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2627
2628 freqs.new = target_pstate * cpu->pstate.scaling;
2629
2630 cpufreq_freq_transition_end(policy, &freqs, false);
2631
2632 return 0;
2633 }
2634
2635 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2636 unsigned int target_freq)
2637 {
2638 struct cpudata *cpu = all_cpu_data[policy->cpu];
2639 int target_pstate;
2640
2641 update_turbo_state();
2642
2643 target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2644
2645 target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2646
2647 return target_pstate * cpu->pstate.scaling;
2648 }
2649
2650 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2651 unsigned long min_perf,
2652 unsigned long target_perf,
2653 unsigned long capacity)
2654 {
2655 struct cpudata *cpu = all_cpu_data[cpunum];
2656 int old_pstate = cpu->pstate.current_pstate;
2657 int cap_pstate, min_pstate, max_pstate, target_pstate;
2658
2659 update_turbo_state();
2660 cap_pstate = global.turbo_disabled ? cpu->pstate.max_pstate :
2661 cpu->pstate.turbo_pstate;
2662
2663 /* Optimization: Avoid unnecessary divisions. */
2664
2665 target_pstate = cap_pstate;
2666 if (target_perf < capacity)
2667 target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2668
2669 min_pstate = cap_pstate;
2670 if (min_perf < capacity)
2671 min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2672
2673 if (min_pstate < cpu->pstate.min_pstate)
2674 min_pstate = cpu->pstate.min_pstate;
2675
2676 if (min_pstate < cpu->min_perf_ratio)
2677 min_pstate = cpu->min_perf_ratio;
2678
2679 max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2680 if (max_pstate < min_pstate)
2681 max_pstate = min_pstate;
2682
2683 target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2684
2685 intel_cpufreq_adjust_hwp(cpu, min_pstate, max_pstate, target_pstate, true);
2686
2687 cpu->pstate.current_pstate = target_pstate;
2688 intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2689 }
2690
2691 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2692 {
2693 int max_state, turbo_max, min_freq, max_freq, ret;
2694 struct freq_qos_request *req;
2695 struct cpudata *cpu;
2696 struct device *dev;
2697
2698 dev = get_cpu_device(policy->cpu);
2699 if (!dev)
2700 return -ENODEV;
2701
2702 ret = __intel_pstate_cpu_init(policy);
2703 if (ret)
2704 return ret;
2705
2706 policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2707 /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2708 policy->cur = policy->cpuinfo.min_freq;
2709
2710 req = kcalloc(2, sizeof(*req), GFP_KERNEL);
2711 if (!req) {
2712 ret = -ENOMEM;
2713 goto pstate_exit;
2714 }
2715
2716 cpu = all_cpu_data[policy->cpu];
2717
2718 if (hwp_active) {
2719 u64 value;
2720
2721 intel_pstate_get_hwp_max(policy->cpu, &turbo_max, &max_state);
2722 policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
2723 rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
2724 WRITE_ONCE(cpu->hwp_req_cached, value);
2725 cpu->epp_cached = intel_pstate_get_epp(cpu, value);
2726 } else {
2727 turbo_max = cpu->pstate.turbo_pstate;
2728 policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2729 }
2730
2731 min_freq = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2732 min_freq *= cpu->pstate.scaling;
2733 max_freq = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2734 max_freq *= cpu->pstate.scaling;
2735
2736 ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
2737 min_freq);
2738 if (ret < 0) {
2739 dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
2740 goto free_req;
2741 }
2742
2743 ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
2744 max_freq);
2745 if (ret < 0) {
2746 dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
2747 goto remove_min_req;
2748 }
2749
2750 policy->driver_data = req;
2751
2752 return 0;
2753
2754 remove_min_req:
2755 freq_qos_remove_request(req);
2756 free_req:
2757 kfree(req);
2758 pstate_exit:
2759 intel_pstate_exit_perf_limits(policy);
2760
2761 return ret;
2762 }
2763
2764 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
2765 {
2766 struct freq_qos_request *req;
2767
2768 req = policy->driver_data;
2769
2770 freq_qos_remove_request(req + 1);
2771 freq_qos_remove_request(req);
2772 kfree(req);
2773
2774 return intel_pstate_cpu_exit(policy);
2775 }
2776
2777 static struct cpufreq_driver intel_cpufreq = {
2778 .flags = CPUFREQ_CONST_LOOPS,
2779 .verify = intel_cpufreq_verify_policy,
2780 .target = intel_cpufreq_target,
2781 .fast_switch = intel_cpufreq_fast_switch,
2782 .init = intel_cpufreq_cpu_init,
2783 .exit = intel_cpufreq_cpu_exit,
2784 .offline = intel_pstate_cpu_offline,
2785 .online = intel_pstate_cpu_online,
2786 .suspend = intel_pstate_suspend,
2787 .resume = intel_pstate_resume,
2788 .update_limits = intel_pstate_update_limits,
2789 .name = "intel_cpufreq",
2790 };
2791
2792 static struct cpufreq_driver *default_driver;
2793
2794 static void intel_pstate_driver_cleanup(void)
2795 {
2796 unsigned int cpu;
2797
2798 get_online_cpus();
2799 for_each_online_cpu(cpu) {
2800 if (all_cpu_data[cpu]) {
2801 if (intel_pstate_driver == &intel_pstate)
2802 intel_pstate_clear_update_util_hook(cpu);
2803
2804 kfree(all_cpu_data[cpu]);
2805 all_cpu_data[cpu] = NULL;
2806 }
2807 }
2808 put_online_cpus();
2809
2810 intel_pstate_driver = NULL;
2811 }
2812
2813 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2814 {
2815 int ret;
2816
2817 if (driver == &intel_pstate)
2818 intel_pstate_sysfs_expose_hwp_dynamic_boost();
2819
2820 memset(&global, 0, sizeof(global));
2821 global.max_perf_pct = 100;
2822
2823 intel_pstate_driver = driver;
2824 ret = cpufreq_register_driver(intel_pstate_driver);
2825 if (ret) {
2826 intel_pstate_driver_cleanup();
2827 return ret;
2828 }
2829
2830 global.min_perf_pct = min_perf_pct_min();
2831
2832 return 0;
2833 }
2834
2835 static ssize_t intel_pstate_show_status(char *buf)
2836 {
2837 if (!intel_pstate_driver)
2838 return sprintf(buf, "off\n");
2839
2840 return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2841 "active" : "passive");
2842 }
2843
2844 static int intel_pstate_update_status(const char *buf, size_t size)
2845 {
2846 if (size == 3 && !strncmp(buf, "off", size)) {
2847 if (!intel_pstate_driver)
2848 return -EINVAL;
2849
2850 if (hwp_active)
2851 return -EBUSY;
2852
2853 cpufreq_unregister_driver(intel_pstate_driver);
2854 intel_pstate_driver_cleanup();
2855 return 0;
2856 }
2857
2858 if (size == 6 && !strncmp(buf, "active", size)) {
2859 if (intel_pstate_driver) {
2860 if (intel_pstate_driver == &intel_pstate)
2861 return 0;
2862
2863 cpufreq_unregister_driver(intel_pstate_driver);
2864 }
2865
2866 return intel_pstate_register_driver(&intel_pstate);
2867 }
2868
2869 if (size == 7 && !strncmp(buf, "passive", size)) {
2870 if (intel_pstate_driver) {
2871 if (intel_pstate_driver == &intel_cpufreq)
2872 return 0;
2873
2874 cpufreq_unregister_driver(intel_pstate_driver);
2875 intel_pstate_sysfs_hide_hwp_dynamic_boost();
2876 }
2877
2878 return intel_pstate_register_driver(&intel_cpufreq);
2879 }
2880
2881 return -EINVAL;
2882 }
2883
2884 static int no_load __initdata;
2885 static int no_hwp __initdata;
2886 static int hwp_only __initdata;
2887 static unsigned int force_load __initdata;
2888
2889 static int __init intel_pstate_msrs_not_valid(void)
2890 {
2891 if (!pstate_funcs.get_max() ||
2892 !pstate_funcs.get_min() ||
2893 !pstate_funcs.get_turbo())
2894 return -ENODEV;
2895
2896 return 0;
2897 }
2898
2899 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2900 {
2901 pstate_funcs.get_max = funcs->get_max;
2902 pstate_funcs.get_max_physical = funcs->get_max_physical;
2903 pstate_funcs.get_min = funcs->get_min;
2904 pstate_funcs.get_turbo = funcs->get_turbo;
2905 pstate_funcs.get_scaling = funcs->get_scaling;
2906 pstate_funcs.get_val = funcs->get_val;
2907 pstate_funcs.get_vid = funcs->get_vid;
2908 pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
2909 }
2910
2911 #ifdef CONFIG_ACPI
2912
2913 static bool __init intel_pstate_no_acpi_pss(void)
2914 {
2915 int i;
2916
2917 for_each_possible_cpu(i) {
2918 acpi_status status;
2919 union acpi_object *pss;
2920 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2921 struct acpi_processor *pr = per_cpu(processors, i);
2922
2923 if (!pr)
2924 continue;
2925
2926 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2927 if (ACPI_FAILURE(status))
2928 continue;
2929
2930 pss = buffer.pointer;
2931 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2932 kfree(pss);
2933 return false;
2934 }
2935
2936 kfree(pss);
2937 }
2938
2939 pr_debug("ACPI _PSS not found\n");
2940 return true;
2941 }
2942
2943 static bool __init intel_pstate_no_acpi_pcch(void)
2944 {
2945 acpi_status status;
2946 acpi_handle handle;
2947
2948 status = acpi_get_handle(NULL, "\\_SB", &handle);
2949 if (ACPI_FAILURE(status))
2950 goto not_found;
2951
2952 if (acpi_has_method(handle, "PCCH"))
2953 return false;
2954
2955 not_found:
2956 pr_debug("ACPI PCCH not found\n");
2957 return true;
2958 }
2959
2960 static bool __init intel_pstate_has_acpi_ppc(void)
2961 {
2962 int i;
2963
2964 for_each_possible_cpu(i) {
2965 struct acpi_processor *pr = per_cpu(processors, i);
2966
2967 if (!pr)
2968 continue;
2969 if (acpi_has_method(pr->handle, "_PPC"))
2970 return true;
2971 }
2972 pr_debug("ACPI _PPC not found\n");
2973 return false;
2974 }
2975
2976 enum {
2977 PSS,
2978 PPC,
2979 };
2980
2981 /* Hardware vendor-specific info that has its own power management modes */
2982 static struct acpi_platform_list plat_info[] __initdata = {
2983 {"HP ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
2984 {"ORACLE", "X4-2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2985 {"ORACLE", "X4-2L ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2986 {"ORACLE", "X4-2B ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2987 {"ORACLE", "X3-2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2988 {"ORACLE", "X3-2L ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2989 {"ORACLE", "X3-2B ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2990 {"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2991 {"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2992 {"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2993 {"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2994 {"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2995 {"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2996 {"ORACLE", "X6-2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2997 {"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2998 { } /* End */
2999 };
3000
3001 #define BITMASK_OOB (BIT(8) | BIT(18))
3002
3003 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3004 {
3005 const struct x86_cpu_id *id;
3006 u64 misc_pwr;
3007 int idx;
3008
3009 id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3010 if (id) {
3011 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3012 if (misc_pwr & BITMASK_OOB) {
3013 pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3014 pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3015 return true;
3016 }
3017 }
3018
3019 idx = acpi_match_platform_list(plat_info);
3020 if (idx < 0)
3021 return false;
3022
3023 switch (plat_info[idx].data) {
3024 case PSS:
3025 if (!intel_pstate_no_acpi_pss())
3026 return false;
3027
3028 return intel_pstate_no_acpi_pcch();
3029 case PPC:
3030 return intel_pstate_has_acpi_ppc() && !force_load;
3031 }
3032
3033 return false;
3034 }
3035
3036 static void intel_pstate_request_control_from_smm(void)
3037 {
3038 /*
3039 * It may be unsafe to request P-states control from SMM if _PPC support
3040 * has not been enabled.
3041 */
3042 if (acpi_ppc)
3043 acpi_processor_pstate_control();
3044 }
3045 #else /* CONFIG_ACPI not enabled */
3046 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3047 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3048 static inline void intel_pstate_request_control_from_smm(void) {}
3049 #endif /* CONFIG_ACPI */
3050
3051 #define INTEL_PSTATE_HWP_BROADWELL 0x01
3052
3053 #define X86_MATCH_HWP(model, hwp_mode) \
3054 X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3055 X86_FEATURE_HWP, hwp_mode)
3056
3057 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3058 X86_MATCH_HWP(BROADWELL_X, INTEL_PSTATE_HWP_BROADWELL),
3059 X86_MATCH_HWP(BROADWELL_D, INTEL_PSTATE_HWP_BROADWELL),
3060 X86_MATCH_HWP(ANY, 0),
3061 {}
3062 };
3063
3064 static int __init intel_pstate_init(void)
3065 {
3066 const struct x86_cpu_id *id;
3067 int rc;
3068
3069 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3070 return -ENODEV;
3071
3072 if (no_load)
3073 return -ENODEV;
3074
3075 id = x86_match_cpu(hwp_support_ids);
3076 if (id) {
3077 copy_cpu_funcs(&core_funcs);
3078 /*
3079 * Avoid enabling HWP for processors without EPP support,
3080 * because that means incomplete HWP implementation which is a
3081 * corner case and supporting it is generally problematic.
3082 */
3083 if (!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) {
3084 hwp_active++;
3085 hwp_mode_bdw = id->driver_data;
3086 intel_pstate.attr = hwp_cpufreq_attrs;
3087 intel_cpufreq.attr = hwp_cpufreq_attrs;
3088 intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3089 intel_cpufreq.fast_switch = NULL;
3090 intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3091 if (!default_driver)
3092 default_driver = &intel_pstate;
3093
3094 goto hwp_cpu_matched;
3095 }
3096 } else {
3097 id = x86_match_cpu(intel_pstate_cpu_ids);
3098 if (!id) {
3099 pr_info("CPU model not supported\n");
3100 return -ENODEV;
3101 }
3102
3103 copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3104 }
3105
3106 if (intel_pstate_msrs_not_valid()) {
3107 pr_info("Invalid MSRs\n");
3108 return -ENODEV;
3109 }
3110 /* Without HWP start in the passive mode. */
3111 if (!default_driver)
3112 default_driver = &intel_cpufreq;
3113
3114 hwp_cpu_matched:
3115 /*
3116 * The Intel pstate driver will be ignored if the platform
3117 * firmware has its own power management modes.
3118 */
3119 if (intel_pstate_platform_pwr_mgmt_exists()) {
3120 pr_info("P-states controlled by the platform\n");
3121 return -ENODEV;
3122 }
3123
3124 if (!hwp_active && hwp_only)
3125 return -ENOTSUPP;
3126
3127 pr_info("Intel P-state driver initializing\n");
3128
3129 all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3130 if (!all_cpu_data)
3131 return -ENOMEM;
3132
3133 intel_pstate_request_control_from_smm();
3134
3135 intel_pstate_sysfs_expose_params();
3136
3137 mutex_lock(&intel_pstate_driver_lock);
3138 rc = intel_pstate_register_driver(default_driver);
3139 mutex_unlock(&intel_pstate_driver_lock);
3140 if (rc) {
3141 intel_pstate_sysfs_remove();
3142 return rc;
3143 }
3144
3145 if (hwp_active) {
3146 const struct x86_cpu_id *id;
3147
3148 id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3149 if (id) {
3150 set_power_ctl_ee_state(false);
3151 pr_info("Disabling energy efficiency optimization\n");
3152 }
3153
3154 pr_info("HWP enabled\n");
3155 }
3156
3157 return 0;
3158 }
3159 device_initcall(intel_pstate_init);
3160
3161 static int __init intel_pstate_setup(char *str)
3162 {
3163 if (!str)
3164 return -EINVAL;
3165
3166 if (!strcmp(str, "disable"))
3167 no_load = 1;
3168 else if (!strcmp(str, "active"))
3169 default_driver = &intel_pstate;
3170 else if (!strcmp(str, "passive"))
3171 default_driver = &intel_cpufreq;
3172
3173 if (!strcmp(str, "no_hwp")) {
3174 pr_info("HWP disabled\n");
3175 no_hwp = 1;
3176 }
3177 if (!strcmp(str, "force"))
3178 force_load = 1;
3179 if (!strcmp(str, "hwp_only"))
3180 hwp_only = 1;
3181 if (!strcmp(str, "per_cpu_perf_limits"))
3182 per_cpu_limits = true;
3183
3184 #ifdef CONFIG_ACPI
3185 if (!strcmp(str, "support_acpi_ppc"))
3186 acpi_ppc = true;
3187 #endif
3188
3189 return 0;
3190 }
3191 early_param("intel_pstate", intel_pstate_setup);
3192
3193 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3194 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3195 MODULE_LICENSE("GPL");
Linux with added FPC-III support