2 * intel_pstate.c: Native P state management for Intel processors
4 * (C) Copyright 2012 Intel Corporation
5 * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
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
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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.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>
29 #include <linux/debugfs.h>
30 #include <linux/acpi.h>
31 #include <linux/vmalloc.h>
32 #include <trace/events/power.h>
34 #include <asm/div64.h>
36 #include <asm/cpu_device_id.h>
37 #include <asm/cpufeature.h>
38 #include <asm/intel-family.h>
40 #define INTEL_CPUFREQ_TRANSITION_LATENCY 20000
42 #define ATOM_RATIOS 0x66a
43 #define ATOM_VIDS 0x66b
44 #define ATOM_TURBO_RATIOS 0x66c
45 #define ATOM_TURBO_VIDS 0x66d
48 #include <acpi/processor.h>
52 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
53 #define fp_toint(X) ((X) >> FRAC_BITS)
56 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
57 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
58 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
60 static inline int32_t mul_fp(int32_t x
, int32_t y
)
62 return ((int64_t)x
* (int64_t)y
) >> FRAC_BITS
;
65 static inline int32_t div_fp(s64 x
, s64 y
)
67 return div64_s64((int64_t)x
<< FRAC_BITS
, y
);
70 static inline int ceiling_fp(int32_t x
)
75 mask
= (1 << FRAC_BITS
) - 1;
81 static inline u64
mul_ext_fp(u64 x
, u64 y
)
83 return (x
* y
) >> EXT_FRAC_BITS
;
86 static inline u64
div_ext_fp(u64 x
, u64 y
)
88 return div64_u64(x
<< EXT_FRAC_BITS
, y
);
92 * struct sample - Store performance sample
93 * @core_avg_perf: Ratio of APERF/MPERF which is the actual average
94 * performance during last sample period
95 * @busy_scaled: Scaled busy value which is used to calculate next
96 * P state. This can be different than core_avg_perf
97 * to account for cpu idle period
98 * @aperf: Difference of actual performance frequency clock count
99 * read from APERF MSR between last and current sample
100 * @mperf: Difference of maximum performance frequency clock count
101 * read from MPERF MSR between last and current sample
102 * @tsc: Difference of time stamp counter between last and
104 * @time: Current time from scheduler
106 * This structure is used in the cpudata structure to store performance sample
107 * data for choosing next P State.
110 int32_t core_avg_perf
;
119 * struct pstate_data - Store P state data
120 * @current_pstate: Current requested P state
121 * @min_pstate: Min P state possible for this platform
122 * @max_pstate: Max P state possible for this platform
123 * @max_pstate_physical:This is physical Max P state for a processor
124 * This can be higher than the max_pstate which can
125 * be limited by platform thermal design power limits
126 * @scaling: Scaling factor to convert frequency to cpufreq
128 * @turbo_pstate: Max Turbo P state possible for this platform
129 * @max_freq: @max_pstate frequency in cpufreq units
130 * @turbo_freq: @turbo_pstate frequency in cpufreq units
132 * Stores the per cpu model P state limits and current P state.
138 int max_pstate_physical
;
141 unsigned int max_freq
;
142 unsigned int turbo_freq
;
146 * struct vid_data - Stores voltage information data
147 * @min: VID data for this platform corresponding to
149 * @max: VID data corresponding to the highest P State.
150 * @turbo: VID data for turbo P state
151 * @ratio: Ratio of (vid max - vid min) /
152 * (max P state - Min P State)
154 * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
155 * This data is used in Atom platforms, where in addition to target P state,
156 * the voltage data needs to be specified to select next P State.
166 * struct _pid - Stores PID data
167 * @setpoint: Target set point for busyness or performance
168 * @integral: Storage for accumulated error values
169 * @p_gain: PID proportional gain
170 * @i_gain: PID integral gain
171 * @d_gain: PID derivative gain
172 * @deadband: PID deadband
173 * @last_err: Last error storage for integral part of PID calculation
175 * Stores PID coefficients and last error for PID controller.
188 * struct perf_limits - Store user and policy limits
189 * @no_turbo: User requested turbo state from intel_pstate sysfs
190 * @turbo_disabled: Platform turbo status either from msr
191 * MSR_IA32_MISC_ENABLE or when maximum available pstate
192 * matches the maximum turbo pstate
193 * @max_perf_pct: Effective maximum performance limit in percentage, this
194 * is minimum of either limits enforced by cpufreq policy
195 * or limits from user set limits via intel_pstate sysfs
196 * @min_perf_pct: Effective minimum performance limit in percentage, this
197 * is maximum of either limits enforced by cpufreq policy
198 * or limits from user set limits via intel_pstate sysfs
199 * @max_perf: This is a scaled value between 0 to 255 for max_perf_pct
200 * This value is used to limit max pstate
201 * @min_perf: This is a scaled value between 0 to 255 for min_perf_pct
202 * This value is used to limit min pstate
203 * @max_policy_pct: The maximum performance in percentage enforced by
204 * cpufreq setpolicy interface
205 * @max_sysfs_pct: The maximum performance in percentage enforced by
206 * intel pstate sysfs interface, unused when per cpu
207 * controls are enforced
208 * @min_policy_pct: The minimum performance in percentage enforced by
209 * cpufreq setpolicy interface
210 * @min_sysfs_pct: The minimum performance in percentage enforced by
211 * intel pstate sysfs interface, unused when per cpu
212 * controls are enforced
214 * Storage for user and policy defined limits.
230 * struct cpudata - Per CPU instance data storage
231 * @cpu: CPU number for this instance data
232 * @policy: CPUFreq policy value
233 * @update_util: CPUFreq utility callback information
234 * @update_util_set: CPUFreq utility callback is set
235 * @iowait_boost: iowait-related boost fraction
236 * @last_update: Time of the last update.
237 * @pstate: Stores P state limits for this CPU
238 * @vid: Stores VID limits for this CPU
239 * @pid: Stores PID parameters for this CPU
240 * @last_sample_time: Last Sample time
241 * @prev_aperf: Last APERF value read from APERF MSR
242 * @prev_mperf: Last MPERF value read from MPERF MSR
243 * @prev_tsc: Last timestamp counter (TSC) value
244 * @prev_cummulative_iowait: IO Wait time difference from last and
246 * @sample: Storage for storing last Sample data
247 * @perf_limits: Pointer to perf_limit unique to this CPU
248 * Not all field in the structure are applicable
249 * when per cpu controls are enforced
250 * @acpi_perf_data: Stores ACPI perf information read from _PSS
251 * @valid_pss_table: Set to true for valid ACPI _PSS entries found
252 * @epp_powersave: Last saved HWP energy performance preference
253 * (EPP) or energy performance bias (EPB),
254 * when policy switched to performance
255 * @epp_policy: Last saved policy used to set EPP/EPB
256 * @epp_default: Power on default HWP energy performance
258 * @epp_saved: Saved EPP/EPB during system suspend or CPU offline
261 * This structure stores per CPU instance data for all CPUs.
267 struct update_util_data update_util
;
268 bool update_util_set
;
270 struct pstate_data pstate
;
275 u64 last_sample_time
;
279 u64 prev_cummulative_iowait
;
280 struct sample sample
;
281 struct perf_limits
*perf_limits
;
283 struct acpi_processor_performance acpi_perf_data
;
284 bool valid_pss_table
;
286 unsigned int iowait_boost
;
293 static struct cpudata
**all_cpu_data
;
296 * struct pstate_adjust_policy - Stores static PID configuration data
297 * @sample_rate_ms: PID calculation sample rate in ms
298 * @sample_rate_ns: Sample rate calculation in ns
299 * @deadband: PID deadband
300 * @setpoint: PID Setpoint
301 * @p_gain_pct: PID proportional gain
302 * @i_gain_pct: PID integral gain
303 * @d_gain_pct: PID derivative gain
305 * Stores per CPU model static PID configuration data.
307 struct pstate_adjust_policy
{
318 * struct pstate_funcs - Per CPU model specific callbacks
319 * @get_max: Callback to get maximum non turbo effective P state
320 * @get_max_physical: Callback to get maximum non turbo physical P state
321 * @get_min: Callback to get minimum P state
322 * @get_turbo: Callback to get turbo P state
323 * @get_scaling: Callback to get frequency scaling factor
324 * @get_val: Callback to convert P state to actual MSR write value
325 * @get_vid: Callback to get VID data for Atom platforms
326 * @get_target_pstate: Callback to a function to calculate next P state to use
328 * Core and Atom CPU models have different way to get P State limits. This
329 * structure is used to store those callbacks.
331 struct pstate_funcs
{
332 int (*get_max
)(void);
333 int (*get_max_physical
)(void);
334 int (*get_min
)(void);
335 int (*get_turbo
)(void);
336 int (*get_scaling
)(void);
337 u64 (*get_val
)(struct cpudata
*, int pstate
);
338 void (*get_vid
)(struct cpudata
*);
339 int32_t (*get_target_pstate
)(struct cpudata
*);
343 * struct cpu_defaults- Per CPU model default config data
344 * @pid_policy: PID config data
345 * @funcs: Callback function data
347 struct cpu_defaults
{
348 struct pstate_adjust_policy pid_policy
;
349 struct pstate_funcs funcs
;
352 static inline int32_t get_target_pstate_use_performance(struct cpudata
*cpu
);
353 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata
*cpu
);
355 static struct pstate_adjust_policy pid_params __read_mostly
;
356 static struct pstate_funcs pstate_funcs __read_mostly
;
357 static int hwp_active __read_mostly
;
358 static bool per_cpu_limits __read_mostly
;
361 static bool acpi_ppc
;
364 static struct perf_limits performance_limits
= {
368 .max_perf
= int_ext_tofp(1),
370 .min_perf
= int_ext_tofp(1),
371 .max_policy_pct
= 100,
372 .max_sysfs_pct
= 100,
377 static struct perf_limits powersave_limits
= {
381 .max_perf
= int_ext_tofp(1),
384 .max_policy_pct
= 100,
385 .max_sysfs_pct
= 100,
390 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
391 static struct perf_limits
*limits
= &performance_limits
;
393 static struct perf_limits
*limits
= &powersave_limits
;
396 static DEFINE_MUTEX(intel_pstate_limits_lock
);
400 static bool intel_pstate_get_ppc_enable_status(void)
402 if (acpi_gbl_FADT
.preferred_profile
== PM_ENTERPRISE_SERVER
||
403 acpi_gbl_FADT
.preferred_profile
== PM_PERFORMANCE_SERVER
)
409 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy
*policy
)
418 if (!intel_pstate_get_ppc_enable_status())
421 cpu
= all_cpu_data
[policy
->cpu
];
423 ret
= acpi_processor_register_performance(&cpu
->acpi_perf_data
,
429 * Check if the control value in _PSS is for PERF_CTL MSR, which should
430 * guarantee that the states returned by it map to the states in our
433 if (cpu
->acpi_perf_data
.control_register
.space_id
!=
434 ACPI_ADR_SPACE_FIXED_HARDWARE
)
438 * If there is only one entry _PSS, simply ignore _PSS and continue as
439 * usual without taking _PSS into account
441 if (cpu
->acpi_perf_data
.state_count
< 2)
444 pr_debug("CPU%u - ACPI _PSS perf data\n", policy
->cpu
);
445 for (i
= 0; i
< cpu
->acpi_perf_data
.state_count
; i
++) {
446 pr_debug(" %cP%d: %u MHz, %u mW, 0x%x\n",
447 (i
== cpu
->acpi_perf_data
.state
? '*' : ' '), i
,
448 (u32
) cpu
->acpi_perf_data
.states
[i
].core_frequency
,
449 (u32
) cpu
->acpi_perf_data
.states
[i
].power
,
450 (u32
) cpu
->acpi_perf_data
.states
[i
].control
);
454 * The _PSS table doesn't contain whole turbo frequency range.
455 * This just contains +1 MHZ above the max non turbo frequency,
456 * with control value corresponding to max turbo ratio. But
457 * when cpufreq set policy is called, it will call with this
458 * max frequency, which will cause a reduced performance as
459 * this driver uses real max turbo frequency as the max
460 * frequency. So correct this frequency in _PSS table to
461 * correct max turbo frequency based on the turbo state.
462 * Also need to convert to MHz as _PSS freq is in MHz.
464 if (!limits
->turbo_disabled
)
465 cpu
->acpi_perf_data
.states
[0].core_frequency
=
466 policy
->cpuinfo
.max_freq
/ 1000;
467 cpu
->valid_pss_table
= true;
468 pr_debug("_PPC limits will be enforced\n");
473 cpu
->valid_pss_table
= false;
474 acpi_processor_unregister_performance(policy
->cpu
);
477 static void intel_pstate_exit_perf_limits(struct cpufreq_policy
*policy
)
481 cpu
= all_cpu_data
[policy
->cpu
];
482 if (!cpu
->valid_pss_table
)
485 acpi_processor_unregister_performance(policy
->cpu
);
489 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy
*policy
)
493 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy
*policy
)
498 static inline void pid_reset(struct _pid
*pid
, int setpoint
, int busy
,
499 int deadband
, int integral
) {
500 pid
->setpoint
= int_tofp(setpoint
);
501 pid
->deadband
= int_tofp(deadband
);
502 pid
->integral
= int_tofp(integral
);
503 pid
->last_err
= int_tofp(setpoint
) - int_tofp(busy
);
506 static inline void pid_p_gain_set(struct _pid
*pid
, int percent
)
508 pid
->p_gain
= div_fp(percent
, 100);
511 static inline void pid_i_gain_set(struct _pid
*pid
, int percent
)
513 pid
->i_gain
= div_fp(percent
, 100);
516 static inline void pid_d_gain_set(struct _pid
*pid
, int percent
)
518 pid
->d_gain
= div_fp(percent
, 100);
521 static signed int pid_calc(struct _pid
*pid
, int32_t busy
)
524 int32_t pterm
, dterm
, fp_error
;
525 int32_t integral_limit
;
527 fp_error
= pid
->setpoint
- busy
;
529 if (abs(fp_error
) <= pid
->deadband
)
532 pterm
= mul_fp(pid
->p_gain
, fp_error
);
534 pid
->integral
+= fp_error
;
537 * We limit the integral here so that it will never
538 * get higher than 30. This prevents it from becoming
539 * too large an input over long periods of time and allows
540 * it to get factored out sooner.
542 * The value of 30 was chosen through experimentation.
544 integral_limit
= int_tofp(30);
545 if (pid
->integral
> integral_limit
)
546 pid
->integral
= integral_limit
;
547 if (pid
->integral
< -integral_limit
)
548 pid
->integral
= -integral_limit
;
550 dterm
= mul_fp(pid
->d_gain
, fp_error
- pid
->last_err
);
551 pid
->last_err
= fp_error
;
553 result
= pterm
+ mul_fp(pid
->integral
, pid
->i_gain
) + dterm
;
554 result
= result
+ (1 << (FRAC_BITS
-1));
555 return (signed int)fp_toint(result
);
558 static inline void intel_pstate_busy_pid_reset(struct cpudata
*cpu
)
560 pid_p_gain_set(&cpu
->pid
, pid_params
.p_gain_pct
);
561 pid_d_gain_set(&cpu
->pid
, pid_params
.d_gain_pct
);
562 pid_i_gain_set(&cpu
->pid
, pid_params
.i_gain_pct
);
564 pid_reset(&cpu
->pid
, pid_params
.setpoint
, 100, pid_params
.deadband
, 0);
567 static inline void intel_pstate_reset_all_pid(void)
571 for_each_online_cpu(cpu
) {
572 if (all_cpu_data
[cpu
])
573 intel_pstate_busy_pid_reset(all_cpu_data
[cpu
]);
577 static inline void update_turbo_state(void)
582 cpu
= all_cpu_data
[0];
583 rdmsrl(MSR_IA32_MISC_ENABLE
, misc_en
);
584 limits
->turbo_disabled
=
585 (misc_en
& MSR_IA32_MISC_ENABLE_TURBO_DISABLE
||
586 cpu
->pstate
.max_pstate
== cpu
->pstate
.turbo_pstate
);
589 static s16
intel_pstate_get_epb(struct cpudata
*cpu_data
)
594 if (!static_cpu_has(X86_FEATURE_EPB
))
597 ret
= rdmsrl_on_cpu(cpu_data
->cpu
, MSR_IA32_ENERGY_PERF_BIAS
, &epb
);
601 return (s16
)(epb
& 0x0f);
604 static s16
intel_pstate_get_epp(struct cpudata
*cpu_data
, u64 hwp_req_data
)
608 if (static_cpu_has(X86_FEATURE_HWP_EPP
)) {
610 * When hwp_req_data is 0, means that caller didn't read
611 * MSR_HWP_REQUEST, so need to read and get EPP.
614 epp
= rdmsrl_on_cpu(cpu_data
->cpu
, MSR_HWP_REQUEST
,
619 epp
= (hwp_req_data
>> 24) & 0xff;
621 /* When there is no EPP present, HWP uses EPB settings */
622 epp
= intel_pstate_get_epb(cpu_data
);
628 static int intel_pstate_set_epb(int cpu
, s16 pref
)
633 if (!static_cpu_has(X86_FEATURE_EPB
))
636 ret
= rdmsrl_on_cpu(cpu
, MSR_IA32_ENERGY_PERF_BIAS
, &epb
);
640 epb
= (epb
& ~0x0f) | pref
;
641 wrmsrl_on_cpu(cpu
, MSR_IA32_ENERGY_PERF_BIAS
, epb
);
647 * EPP/EPB display strings corresponding to EPP index in the
648 * energy_perf_strings[]
650 *-------------------------------------
653 * 2 balance_performance
657 static const char * const energy_perf_strings
[] = {
660 "balance_performance",
666 static int intel_pstate_get_energy_pref_index(struct cpudata
*cpu_data
)
671 epp
= intel_pstate_get_epp(cpu_data
, 0);
675 if (static_cpu_has(X86_FEATURE_HWP_EPP
)) {
678 * 0x00-0x3F : Performance
679 * 0x40-0x7F : Balance performance
680 * 0x80-0xBF : Balance power
682 * The EPP is a 8 bit value, but our ranges restrict the
683 * value which can be set. Here only using top two bits
686 index
= (epp
>> 6) + 1;
687 } else if (static_cpu_has(X86_FEATURE_EPB
)) {
690 * 0x00-0x03 : Performance
691 * 0x04-0x07 : Balance performance
692 * 0x08-0x0B : Balance power
694 * The EPB is a 4 bit value, but our ranges restrict the
695 * value which can be set. Here only using top two bits
698 index
= (epp
>> 2) + 1;
704 static int intel_pstate_set_energy_pref_index(struct cpudata
*cpu_data
,
711 epp
= cpu_data
->epp_default
;
713 mutex_lock(&intel_pstate_limits_lock
);
715 if (static_cpu_has(X86_FEATURE_HWP_EPP
)) {
718 ret
= rdmsrl_on_cpu(cpu_data
->cpu
, MSR_HWP_REQUEST
, &value
);
722 value
&= ~GENMASK_ULL(31, 24);
725 * If epp is not default, convert from index into
726 * energy_perf_strings to epp value, by shifting 6
727 * bits left to use only top two bits in epp.
728 * The resultant epp need to shifted by 24 bits to
729 * epp position in MSR_HWP_REQUEST.
732 epp
= (pref_index
- 1) << 6;
734 value
|= (u64
)epp
<< 24;
735 ret
= wrmsrl_on_cpu(cpu_data
->cpu
, MSR_HWP_REQUEST
, value
);
738 epp
= (pref_index
- 1) << 2;
739 ret
= intel_pstate_set_epb(cpu_data
->cpu
, epp
);
742 mutex_unlock(&intel_pstate_limits_lock
);
747 static ssize_t
show_energy_performance_available_preferences(
748 struct cpufreq_policy
*policy
, char *buf
)
753 while (energy_perf_strings
[i
] != NULL
)
754 ret
+= sprintf(&buf
[ret
], "%s ", energy_perf_strings
[i
++]);
756 ret
+= sprintf(&buf
[ret
], "\n");
761 cpufreq_freq_attr_ro(energy_performance_available_preferences
);
763 static ssize_t
store_energy_performance_preference(
764 struct cpufreq_policy
*policy
, const char *buf
, size_t count
)
766 struct cpudata
*cpu_data
= all_cpu_data
[policy
->cpu
];
767 char str_preference
[21];
770 ret
= sscanf(buf
, "%20s", str_preference
);
774 while (energy_perf_strings
[i
] != NULL
) {
775 if (!strcmp(str_preference
, energy_perf_strings
[i
])) {
776 intel_pstate_set_energy_pref_index(cpu_data
, i
);
785 static ssize_t
show_energy_performance_preference(
786 struct cpufreq_policy
*policy
, char *buf
)
788 struct cpudata
*cpu_data
= all_cpu_data
[policy
->cpu
];
791 preference
= intel_pstate_get_energy_pref_index(cpu_data
);
795 return sprintf(buf
, "%s\n", energy_perf_strings
[preference
]);
798 cpufreq_freq_attr_rw(energy_performance_preference
);
800 static struct freq_attr
*hwp_cpufreq_attrs
[] = {
801 &energy_performance_preference
,
802 &energy_performance_available_preferences
,
806 static void intel_pstate_hwp_set(const struct cpumask
*cpumask
)
808 int min
, hw_min
, max
, hw_max
, cpu
, range
, adj_range
;
809 struct perf_limits
*perf_limits
= limits
;
812 for_each_cpu(cpu
, cpumask
) {
813 int max_perf_pct
, min_perf_pct
;
814 struct cpudata
*cpu_data
= all_cpu_data
[cpu
];
818 perf_limits
= all_cpu_data
[cpu
]->perf_limits
;
820 rdmsrl_on_cpu(cpu
, MSR_HWP_CAPABILITIES
, &cap
);
821 hw_min
= HWP_LOWEST_PERF(cap
);
822 hw_max
= HWP_HIGHEST_PERF(cap
);
823 range
= hw_max
- hw_min
;
825 max_perf_pct
= perf_limits
->max_perf_pct
;
826 min_perf_pct
= perf_limits
->min_perf_pct
;
828 rdmsrl_on_cpu(cpu
, MSR_HWP_REQUEST
, &value
);
829 adj_range
= min_perf_pct
* range
/ 100;
830 min
= hw_min
+ adj_range
;
831 value
&= ~HWP_MIN_PERF(~0L);
832 value
|= HWP_MIN_PERF(min
);
834 adj_range
= max_perf_pct
* range
/ 100;
835 max
= hw_min
+ adj_range
;
836 if (limits
->no_turbo
) {
837 hw_max
= HWP_GUARANTEED_PERF(cap
);
842 value
&= ~HWP_MAX_PERF(~0L);
843 value
|= HWP_MAX_PERF(max
);
845 if (cpu_data
->epp_policy
== cpu_data
->policy
)
848 cpu_data
->epp_policy
= cpu_data
->policy
;
850 if (cpu_data
->epp_saved
>= 0) {
851 epp
= cpu_data
->epp_saved
;
852 cpu_data
->epp_saved
= -EINVAL
;
856 if (cpu_data
->policy
== CPUFREQ_POLICY_PERFORMANCE
) {
857 epp
= intel_pstate_get_epp(cpu_data
, value
);
858 cpu_data
->epp_powersave
= epp
;
859 /* If EPP read was failed, then don't try to write */
866 /* skip setting EPP, when saved value is invalid */
867 if (cpu_data
->epp_powersave
< 0)
871 * No need to restore EPP when it is not zero. This
873 * - Policy is not changed
874 * - user has manually changed
875 * - Error reading EPB
877 epp
= intel_pstate_get_epp(cpu_data
, value
);
881 epp
= cpu_data
->epp_powersave
;
884 if (static_cpu_has(X86_FEATURE_HWP_EPP
)) {
885 value
&= ~GENMASK_ULL(31, 24);
886 value
|= (u64
)epp
<< 24;
888 intel_pstate_set_epb(cpu
, epp
);
891 wrmsrl_on_cpu(cpu
, MSR_HWP_REQUEST
, value
);
895 static int intel_pstate_hwp_set_policy(struct cpufreq_policy
*policy
)
898 intel_pstate_hwp_set(policy
->cpus
);
903 static int intel_pstate_hwp_save_state(struct cpufreq_policy
*policy
)
905 struct cpudata
*cpu_data
= all_cpu_data
[policy
->cpu
];
910 cpu_data
->epp_saved
= intel_pstate_get_epp(cpu_data
, 0);
915 static int intel_pstate_resume(struct cpufreq_policy
*policy
)
922 mutex_lock(&intel_pstate_limits_lock
);
924 all_cpu_data
[policy
->cpu
]->epp_policy
= 0;
926 ret
= intel_pstate_hwp_set_policy(policy
);
928 mutex_unlock(&intel_pstate_limits_lock
);
933 static void intel_pstate_hwp_set_online_cpus(void)
936 intel_pstate_hwp_set(cpu_online_mask
);
940 /************************** debugfs begin ************************/
941 static int pid_param_set(void *data
, u64 val
)
944 intel_pstate_reset_all_pid();
948 static int pid_param_get(void *data
, u64
*val
)
953 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param
, pid_param_get
, pid_param_set
, "%llu\n");
960 static struct pid_param pid_files
[] = {
961 {"sample_rate_ms", &pid_params
.sample_rate_ms
},
962 {"d_gain_pct", &pid_params
.d_gain_pct
},
963 {"i_gain_pct", &pid_params
.i_gain_pct
},
964 {"deadband", &pid_params
.deadband
},
965 {"setpoint", &pid_params
.setpoint
},
966 {"p_gain_pct", &pid_params
.p_gain_pct
},
970 static void __init
intel_pstate_debug_expose_params(void)
972 struct dentry
*debugfs_parent
;
975 debugfs_parent
= debugfs_create_dir("pstate_snb", NULL
);
976 if (IS_ERR_OR_NULL(debugfs_parent
))
978 while (pid_files
[i
].name
) {
979 debugfs_create_file(pid_files
[i
].name
, 0660,
980 debugfs_parent
, pid_files
[i
].value
,
986 /************************** debugfs end ************************/
988 /************************** sysfs begin ************************/
989 #define show_one(file_name, object) \
990 static ssize_t show_##file_name \
991 (struct kobject *kobj, struct attribute *attr, char *buf) \
993 return sprintf(buf, "%u\n", limits->object); \
996 static ssize_t
show_turbo_pct(struct kobject
*kobj
,
997 struct attribute
*attr
, char *buf
)
1000 int total
, no_turbo
, turbo_pct
;
1003 cpu
= all_cpu_data
[0];
1005 total
= cpu
->pstate
.turbo_pstate
- cpu
->pstate
.min_pstate
+ 1;
1006 no_turbo
= cpu
->pstate
.max_pstate
- cpu
->pstate
.min_pstate
+ 1;
1007 turbo_fp
= div_fp(no_turbo
, total
);
1008 turbo_pct
= 100 - fp_toint(mul_fp(turbo_fp
, int_tofp(100)));
1009 return sprintf(buf
, "%u\n", turbo_pct
);
1012 static ssize_t
show_num_pstates(struct kobject
*kobj
,
1013 struct attribute
*attr
, char *buf
)
1015 struct cpudata
*cpu
;
1018 cpu
= all_cpu_data
[0];
1019 total
= cpu
->pstate
.turbo_pstate
- cpu
->pstate
.min_pstate
+ 1;
1020 return sprintf(buf
, "%u\n", total
);
1023 static ssize_t
show_no_turbo(struct kobject
*kobj
,
1024 struct attribute
*attr
, char *buf
)
1028 update_turbo_state();
1029 if (limits
->turbo_disabled
)
1030 ret
= sprintf(buf
, "%u\n", limits
->turbo_disabled
);
1032 ret
= sprintf(buf
, "%u\n", limits
->no_turbo
);
1037 static ssize_t
store_no_turbo(struct kobject
*a
, struct attribute
*b
,
1038 const char *buf
, size_t count
)
1043 ret
= sscanf(buf
, "%u", &input
);
1047 mutex_lock(&intel_pstate_limits_lock
);
1049 update_turbo_state();
1050 if (limits
->turbo_disabled
) {
1051 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1052 mutex_unlock(&intel_pstate_limits_lock
);
1056 limits
->no_turbo
= clamp_t(int, input
, 0, 1);
1059 intel_pstate_hwp_set_online_cpus();
1061 mutex_unlock(&intel_pstate_limits_lock
);
1066 static ssize_t
store_max_perf_pct(struct kobject
*a
, struct attribute
*b
,
1067 const char *buf
, size_t count
)
1072 ret
= sscanf(buf
, "%u", &input
);
1076 mutex_lock(&intel_pstate_limits_lock
);
1078 limits
->max_sysfs_pct
= clamp_t(int, input
, 0 , 100);
1079 limits
->max_perf_pct
= min(limits
->max_policy_pct
,
1080 limits
->max_sysfs_pct
);
1081 limits
->max_perf_pct
= max(limits
->min_policy_pct
,
1082 limits
->max_perf_pct
);
1083 limits
->max_perf_pct
= max(limits
->min_perf_pct
,
1084 limits
->max_perf_pct
);
1085 limits
->max_perf
= div_ext_fp(limits
->max_perf_pct
, 100);
1088 intel_pstate_hwp_set_online_cpus();
1090 mutex_unlock(&intel_pstate_limits_lock
);
1095 static ssize_t
store_min_perf_pct(struct kobject
*a
, struct attribute
*b
,
1096 const char *buf
, size_t count
)
1101 ret
= sscanf(buf
, "%u", &input
);
1105 mutex_lock(&intel_pstate_limits_lock
);
1107 limits
->min_sysfs_pct
= clamp_t(int, input
, 0 , 100);
1108 limits
->min_perf_pct
= max(limits
->min_policy_pct
,
1109 limits
->min_sysfs_pct
);
1110 limits
->min_perf_pct
= min(limits
->max_policy_pct
,
1111 limits
->min_perf_pct
);
1112 limits
->min_perf_pct
= min(limits
->max_perf_pct
,
1113 limits
->min_perf_pct
);
1114 limits
->min_perf
= div_ext_fp(limits
->min_perf_pct
, 100);
1117 intel_pstate_hwp_set_online_cpus();
1119 mutex_unlock(&intel_pstate_limits_lock
);
1124 show_one(max_perf_pct
, max_perf_pct
);
1125 show_one(min_perf_pct
, min_perf_pct
);
1127 define_one_global_rw(no_turbo
);
1128 define_one_global_rw(max_perf_pct
);
1129 define_one_global_rw(min_perf_pct
);
1130 define_one_global_ro(turbo_pct
);
1131 define_one_global_ro(num_pstates
);
1133 static struct attribute
*intel_pstate_attributes
[] = {
1140 static struct attribute_group intel_pstate_attr_group
= {
1141 .attrs
= intel_pstate_attributes
,
1144 static void __init
intel_pstate_sysfs_expose_params(void)
1146 struct kobject
*intel_pstate_kobject
;
1149 intel_pstate_kobject
= kobject_create_and_add("intel_pstate",
1150 &cpu_subsys
.dev_root
->kobj
);
1151 if (WARN_ON(!intel_pstate_kobject
))
1154 rc
= sysfs_create_group(intel_pstate_kobject
, &intel_pstate_attr_group
);
1159 * If per cpu limits are enforced there are no global limits, so
1160 * return without creating max/min_perf_pct attributes
1165 rc
= sysfs_create_file(intel_pstate_kobject
, &max_perf_pct
.attr
);
1168 rc
= sysfs_create_file(intel_pstate_kobject
, &min_perf_pct
.attr
);
1172 /************************** sysfs end ************************/
1174 static void intel_pstate_hwp_enable(struct cpudata
*cpudata
)
1176 /* First disable HWP notification interrupt as we don't process them */
1177 if (static_cpu_has(X86_FEATURE_HWP_NOTIFY
))
1178 wrmsrl_on_cpu(cpudata
->cpu
, MSR_HWP_INTERRUPT
, 0x00);
1180 wrmsrl_on_cpu(cpudata
->cpu
, MSR_PM_ENABLE
, 0x1);
1181 cpudata
->epp_policy
= 0;
1182 if (cpudata
->epp_default
== -EINVAL
)
1183 cpudata
->epp_default
= intel_pstate_get_epp(cpudata
, 0);
1186 static int atom_get_min_pstate(void)
1190 rdmsrl(ATOM_RATIOS
, value
);
1191 return (value
>> 8) & 0x7F;
1194 static int atom_get_max_pstate(void)
1198 rdmsrl(ATOM_RATIOS
, value
);
1199 return (value
>> 16) & 0x7F;
1202 static int atom_get_turbo_pstate(void)
1206 rdmsrl(ATOM_TURBO_RATIOS
, value
);
1207 return value
& 0x7F;
1210 static u64
atom_get_val(struct cpudata
*cpudata
, int pstate
)
1216 val
= (u64
)pstate
<< 8;
1217 if (limits
->no_turbo
&& !limits
->turbo_disabled
)
1218 val
|= (u64
)1 << 32;
1220 vid_fp
= cpudata
->vid
.min
+ mul_fp(
1221 int_tofp(pstate
- cpudata
->pstate
.min_pstate
),
1222 cpudata
->vid
.ratio
);
1224 vid_fp
= clamp_t(int32_t, vid_fp
, cpudata
->vid
.min
, cpudata
->vid
.max
);
1225 vid
= ceiling_fp(vid_fp
);
1227 if (pstate
> cpudata
->pstate
.max_pstate
)
1228 vid
= cpudata
->vid
.turbo
;
1233 static int silvermont_get_scaling(void)
1237 /* Defined in Table 35-6 from SDM (Sept 2015) */
1238 static int silvermont_freq_table
[] = {
1239 83300, 100000, 133300, 116700, 80000};
1241 rdmsrl(MSR_FSB_FREQ
, value
);
1245 return silvermont_freq_table
[i
];
1248 static int airmont_get_scaling(void)
1252 /* Defined in Table 35-10 from SDM (Sept 2015) */
1253 static int airmont_freq_table
[] = {
1254 83300, 100000, 133300, 116700, 80000,
1255 93300, 90000, 88900, 87500};
1257 rdmsrl(MSR_FSB_FREQ
, value
);
1261 return airmont_freq_table
[i
];
1264 static void atom_get_vid(struct cpudata
*cpudata
)
1268 rdmsrl(ATOM_VIDS
, value
);
1269 cpudata
->vid
.min
= int_tofp((value
>> 8) & 0x7f);
1270 cpudata
->vid
.max
= int_tofp((value
>> 16) & 0x7f);
1271 cpudata
->vid
.ratio
= div_fp(
1272 cpudata
->vid
.max
- cpudata
->vid
.min
,
1273 int_tofp(cpudata
->pstate
.max_pstate
-
1274 cpudata
->pstate
.min_pstate
));
1276 rdmsrl(ATOM_TURBO_VIDS
, value
);
1277 cpudata
->vid
.turbo
= value
& 0x7f;
1280 static int core_get_min_pstate(void)
1284 rdmsrl(MSR_PLATFORM_INFO
, value
);
1285 return (value
>> 40) & 0xFF;
1288 static int core_get_max_pstate_physical(void)
1292 rdmsrl(MSR_PLATFORM_INFO
, value
);
1293 return (value
>> 8) & 0xFF;
1296 static int core_get_max_pstate(void)
1303 rdmsrl(MSR_PLATFORM_INFO
, plat_info
);
1304 max_pstate
= (plat_info
>> 8) & 0xFF;
1306 err
= rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO
, &tar
);
1308 /* Do some sanity checking for safety */
1309 if (plat_info
& 0x600000000) {
1314 err
= rdmsrl_safe(MSR_CONFIG_TDP_CONTROL
, &tdp_ctrl
);
1318 tdp_msr
= MSR_CONFIG_TDP_NOMINAL
+ (tdp_ctrl
& 0x3);
1319 err
= rdmsrl_safe(tdp_msr
, &tdp_ratio
);
1323 /* For level 1 and 2, bits[23:16] contain the ratio */
1327 tdp_ratio
&= 0xff; /* ratios are only 8 bits long */
1328 if (tdp_ratio
- 1 == tar
) {
1330 pr_debug("max_pstate=TAC %x\n", max_pstate
);
1341 static int core_get_turbo_pstate(void)
1346 rdmsrl(MSR_TURBO_RATIO_LIMIT
, value
);
1347 nont
= core_get_max_pstate();
1348 ret
= (value
) & 255;
1354 static inline int core_get_scaling(void)
1359 static u64
core_get_val(struct cpudata
*cpudata
, int pstate
)
1363 val
= (u64
)pstate
<< 8;
1364 if (limits
->no_turbo
&& !limits
->turbo_disabled
)
1365 val
|= (u64
)1 << 32;
1370 static int knl_get_turbo_pstate(void)
1375 rdmsrl(MSR_TURBO_RATIO_LIMIT
, value
);
1376 nont
= core_get_max_pstate();
1377 ret
= (((value
) >> 8) & 0xFF);
1383 static struct cpu_defaults core_params
= {
1385 .sample_rate_ms
= 10,
1393 .get_max
= core_get_max_pstate
,
1394 .get_max_physical
= core_get_max_pstate_physical
,
1395 .get_min
= core_get_min_pstate
,
1396 .get_turbo
= core_get_turbo_pstate
,
1397 .get_scaling
= core_get_scaling
,
1398 .get_val
= core_get_val
,
1399 .get_target_pstate
= get_target_pstate_use_performance
,
1403 static const struct cpu_defaults silvermont_params
= {
1405 .sample_rate_ms
= 10,
1413 .get_max
= atom_get_max_pstate
,
1414 .get_max_physical
= atom_get_max_pstate
,
1415 .get_min
= atom_get_min_pstate
,
1416 .get_turbo
= atom_get_turbo_pstate
,
1417 .get_val
= atom_get_val
,
1418 .get_scaling
= silvermont_get_scaling
,
1419 .get_vid
= atom_get_vid
,
1420 .get_target_pstate
= get_target_pstate_use_cpu_load
,
1424 static const struct cpu_defaults airmont_params
= {
1426 .sample_rate_ms
= 10,
1434 .get_max
= atom_get_max_pstate
,
1435 .get_max_physical
= atom_get_max_pstate
,
1436 .get_min
= atom_get_min_pstate
,
1437 .get_turbo
= atom_get_turbo_pstate
,
1438 .get_val
= atom_get_val
,
1439 .get_scaling
= airmont_get_scaling
,
1440 .get_vid
= atom_get_vid
,
1441 .get_target_pstate
= get_target_pstate_use_cpu_load
,
1445 static const struct cpu_defaults knl_params
= {
1447 .sample_rate_ms
= 10,
1455 .get_max
= core_get_max_pstate
,
1456 .get_max_physical
= core_get_max_pstate_physical
,
1457 .get_min
= core_get_min_pstate
,
1458 .get_turbo
= knl_get_turbo_pstate
,
1459 .get_scaling
= core_get_scaling
,
1460 .get_val
= core_get_val
,
1461 .get_target_pstate
= get_target_pstate_use_performance
,
1465 static const struct cpu_defaults bxt_params
= {
1467 .sample_rate_ms
= 10,
1475 .get_max
= core_get_max_pstate
,
1476 .get_max_physical
= core_get_max_pstate_physical
,
1477 .get_min
= core_get_min_pstate
,
1478 .get_turbo
= core_get_turbo_pstate
,
1479 .get_scaling
= core_get_scaling
,
1480 .get_val
= core_get_val
,
1481 .get_target_pstate
= get_target_pstate_use_cpu_load
,
1485 static void intel_pstate_get_min_max(struct cpudata
*cpu
, int *min
, int *max
)
1487 int max_perf
= cpu
->pstate
.turbo_pstate
;
1490 struct perf_limits
*perf_limits
= limits
;
1492 if (limits
->no_turbo
|| limits
->turbo_disabled
)
1493 max_perf
= cpu
->pstate
.max_pstate
;
1496 perf_limits
= cpu
->perf_limits
;
1499 * performance can be limited by user through sysfs, by cpufreq
1500 * policy, or by cpu specific default values determined through
1503 max_perf_adj
= fp_ext_toint(max_perf
* perf_limits
->max_perf
);
1504 *max
= clamp_t(int, max_perf_adj
,
1505 cpu
->pstate
.min_pstate
, cpu
->pstate
.turbo_pstate
);
1507 min_perf
= fp_ext_toint(max_perf
* perf_limits
->min_perf
);
1508 *min
= clamp_t(int, min_perf
, cpu
->pstate
.min_pstate
, max_perf
);
1511 static void intel_pstate_set_pstate(struct cpudata
*cpu
, int pstate
)
1513 trace_cpu_frequency(pstate
* cpu
->pstate
.scaling
, cpu
->cpu
);
1514 cpu
->pstate
.current_pstate
= pstate
;
1516 * Generally, there is no guarantee that this code will always run on
1517 * the CPU being updated, so force the register update to run on the
1520 wrmsrl_on_cpu(cpu
->cpu
, MSR_IA32_PERF_CTL
,
1521 pstate_funcs
.get_val(cpu
, pstate
));
1524 static void intel_pstate_set_min_pstate(struct cpudata
*cpu
)
1526 intel_pstate_set_pstate(cpu
, cpu
->pstate
.min_pstate
);
1529 static void intel_pstate_max_within_limits(struct cpudata
*cpu
)
1531 int min_pstate
, max_pstate
;
1533 update_turbo_state();
1534 intel_pstate_get_min_max(cpu
, &min_pstate
, &max_pstate
);
1535 intel_pstate_set_pstate(cpu
, max_pstate
);
1538 static void intel_pstate_get_cpu_pstates(struct cpudata
*cpu
)
1540 cpu
->pstate
.min_pstate
= pstate_funcs
.get_min();
1541 cpu
->pstate
.max_pstate
= pstate_funcs
.get_max();
1542 cpu
->pstate
.max_pstate_physical
= pstate_funcs
.get_max_physical();
1543 cpu
->pstate
.turbo_pstate
= pstate_funcs
.get_turbo();
1544 cpu
->pstate
.scaling
= pstate_funcs
.get_scaling();
1545 cpu
->pstate
.max_freq
= cpu
->pstate
.max_pstate
* cpu
->pstate
.scaling
;
1546 cpu
->pstate
.turbo_freq
= cpu
->pstate
.turbo_pstate
* cpu
->pstate
.scaling
;
1548 if (pstate_funcs
.get_vid
)
1549 pstate_funcs
.get_vid(cpu
);
1551 intel_pstate_set_min_pstate(cpu
);
1554 static inline void intel_pstate_calc_avg_perf(struct cpudata
*cpu
)
1556 struct sample
*sample
= &cpu
->sample
;
1558 sample
->core_avg_perf
= div_ext_fp(sample
->aperf
, sample
->mperf
);
1561 static inline bool intel_pstate_sample(struct cpudata
*cpu
, u64 time
)
1564 unsigned long flags
;
1567 local_irq_save(flags
);
1568 rdmsrl(MSR_IA32_APERF
, aperf
);
1569 rdmsrl(MSR_IA32_MPERF
, mperf
);
1571 if (cpu
->prev_mperf
== mperf
|| cpu
->prev_tsc
== tsc
) {
1572 local_irq_restore(flags
);
1575 local_irq_restore(flags
);
1577 cpu
->last_sample_time
= cpu
->sample
.time
;
1578 cpu
->sample
.time
= time
;
1579 cpu
->sample
.aperf
= aperf
;
1580 cpu
->sample
.mperf
= mperf
;
1581 cpu
->sample
.tsc
= tsc
;
1582 cpu
->sample
.aperf
-= cpu
->prev_aperf
;
1583 cpu
->sample
.mperf
-= cpu
->prev_mperf
;
1584 cpu
->sample
.tsc
-= cpu
->prev_tsc
;
1586 cpu
->prev_aperf
= aperf
;
1587 cpu
->prev_mperf
= mperf
;
1588 cpu
->prev_tsc
= tsc
;
1590 * First time this function is invoked in a given cycle, all of the
1591 * previous sample data fields are equal to zero or stale and they must
1592 * be populated with meaningful numbers for things to work, so assume
1593 * that sample.time will always be reset before setting the utilization
1594 * update hook and make the caller skip the sample then.
1596 return !!cpu
->last_sample_time
;
1599 static inline int32_t get_avg_frequency(struct cpudata
*cpu
)
1601 return mul_ext_fp(cpu
->sample
.core_avg_perf
,
1602 cpu
->pstate
.max_pstate_physical
* cpu
->pstate
.scaling
);
1605 static inline int32_t get_avg_pstate(struct cpudata
*cpu
)
1607 return mul_ext_fp(cpu
->pstate
.max_pstate_physical
,
1608 cpu
->sample
.core_avg_perf
);
1611 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata
*cpu
)
1613 struct sample
*sample
= &cpu
->sample
;
1614 int32_t busy_frac
, boost
;
1615 int target
, avg_pstate
;
1617 busy_frac
= div_fp(sample
->mperf
, sample
->tsc
);
1619 boost
= cpu
->iowait_boost
;
1620 cpu
->iowait_boost
>>= 1;
1622 if (busy_frac
< boost
)
1625 sample
->busy_scaled
= busy_frac
* 100;
1627 target
= limits
->no_turbo
|| limits
->turbo_disabled
?
1628 cpu
->pstate
.max_pstate
: cpu
->pstate
.turbo_pstate
;
1629 target
+= target
>> 2;
1630 target
= mul_fp(target
, busy_frac
);
1631 if (target
< cpu
->pstate
.min_pstate
)
1632 target
= cpu
->pstate
.min_pstate
;
1635 * If the average P-state during the previous cycle was higher than the
1636 * current target, add 50% of the difference to the target to reduce
1637 * possible performance oscillations and offset possible performance
1638 * loss related to moving the workload from one CPU to another within
1641 avg_pstate
= get_avg_pstate(cpu
);
1642 if (avg_pstate
> target
)
1643 target
+= (avg_pstate
- target
) >> 1;
1648 static inline int32_t get_target_pstate_use_performance(struct cpudata
*cpu
)
1650 int32_t perf_scaled
, max_pstate
, current_pstate
, sample_ratio
;
1654 * perf_scaled is the ratio of the average P-state during the last
1655 * sampling period to the P-state requested last time (in percent).
1657 * That measures the system's response to the previous P-state
1660 max_pstate
= cpu
->pstate
.max_pstate_physical
;
1661 current_pstate
= cpu
->pstate
.current_pstate
;
1662 perf_scaled
= mul_ext_fp(cpu
->sample
.core_avg_perf
,
1663 div_fp(100 * max_pstate
, current_pstate
));
1666 * Since our utilization update callback will not run unless we are
1667 * in C0, check if the actual elapsed time is significantly greater (3x)
1668 * than our sample interval. If it is, then we were idle for a long
1669 * enough period of time to adjust our performance metric.
1671 duration_ns
= cpu
->sample
.time
- cpu
->last_sample_time
;
1672 if ((s64
)duration_ns
> pid_params
.sample_rate_ns
* 3) {
1673 sample_ratio
= div_fp(pid_params
.sample_rate_ns
, duration_ns
);
1674 perf_scaled
= mul_fp(perf_scaled
, sample_ratio
);
1676 sample_ratio
= div_fp(100 * cpu
->sample
.mperf
, cpu
->sample
.tsc
);
1677 if (sample_ratio
< int_tofp(1))
1681 cpu
->sample
.busy_scaled
= perf_scaled
;
1682 return cpu
->pstate
.current_pstate
- pid_calc(&cpu
->pid
, perf_scaled
);
1685 static int intel_pstate_prepare_request(struct cpudata
*cpu
, int pstate
)
1687 int max_perf
, min_perf
;
1689 intel_pstate_get_min_max(cpu
, &min_perf
, &max_perf
);
1690 pstate
= clamp_t(int, pstate
, min_perf
, max_perf
);
1691 trace_cpu_frequency(pstate
* cpu
->pstate
.scaling
, cpu
->cpu
);
1695 static void intel_pstate_update_pstate(struct cpudata
*cpu
, int pstate
)
1697 pstate
= intel_pstate_prepare_request(cpu
, pstate
);
1698 if (pstate
== cpu
->pstate
.current_pstate
)
1701 cpu
->pstate
.current_pstate
= pstate
;
1702 wrmsrl(MSR_IA32_PERF_CTL
, pstate_funcs
.get_val(cpu
, pstate
));
1705 static inline void intel_pstate_adjust_busy_pstate(struct cpudata
*cpu
)
1707 int from
, target_pstate
;
1708 struct sample
*sample
;
1710 from
= cpu
->pstate
.current_pstate
;
1712 target_pstate
= cpu
->policy
== CPUFREQ_POLICY_PERFORMANCE
?
1713 cpu
->pstate
.turbo_pstate
: pstate_funcs
.get_target_pstate(cpu
);
1715 update_turbo_state();
1717 intel_pstate_update_pstate(cpu
, target_pstate
);
1719 sample
= &cpu
->sample
;
1720 trace_pstate_sample(mul_ext_fp(100, sample
->core_avg_perf
),
1721 fp_toint(sample
->busy_scaled
),
1723 cpu
->pstate
.current_pstate
,
1727 get_avg_frequency(cpu
),
1728 fp_toint(cpu
->iowait_boost
* 100));
1731 static void intel_pstate_update_util(struct update_util_data
*data
, u64 time
,
1734 struct cpudata
*cpu
= container_of(data
, struct cpudata
, update_util
);
1737 if (pstate_funcs
.get_target_pstate
== get_target_pstate_use_cpu_load
) {
1738 if (flags
& SCHED_CPUFREQ_IOWAIT
) {
1739 cpu
->iowait_boost
= int_tofp(1);
1740 } else if (cpu
->iowait_boost
) {
1741 /* Clear iowait_boost if the CPU may have been idle. */
1742 delta_ns
= time
- cpu
->last_update
;
1743 if (delta_ns
> TICK_NSEC
)
1744 cpu
->iowait_boost
= 0;
1746 cpu
->last_update
= time
;
1749 delta_ns
= time
- cpu
->sample
.time
;
1750 if ((s64
)delta_ns
>= pid_params
.sample_rate_ns
) {
1751 bool sample_taken
= intel_pstate_sample(cpu
, time
);
1754 intel_pstate_calc_avg_perf(cpu
);
1756 intel_pstate_adjust_busy_pstate(cpu
);
1761 #define ICPU(model, policy) \
1762 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1763 (unsigned long)&policy }
1765 static const struct x86_cpu_id intel_pstate_cpu_ids
[] = {
1766 ICPU(INTEL_FAM6_SANDYBRIDGE
, core_params
),
1767 ICPU(INTEL_FAM6_SANDYBRIDGE_X
, core_params
),
1768 ICPU(INTEL_FAM6_ATOM_SILVERMONT1
, silvermont_params
),
1769 ICPU(INTEL_FAM6_IVYBRIDGE
, core_params
),
1770 ICPU(INTEL_FAM6_HASWELL_CORE
, core_params
),
1771 ICPU(INTEL_FAM6_BROADWELL_CORE
, core_params
),
1772 ICPU(INTEL_FAM6_IVYBRIDGE_X
, core_params
),
1773 ICPU(INTEL_FAM6_HASWELL_X
, core_params
),
1774 ICPU(INTEL_FAM6_HASWELL_ULT
, core_params
),
1775 ICPU(INTEL_FAM6_HASWELL_GT3E
, core_params
),
1776 ICPU(INTEL_FAM6_BROADWELL_GT3E
, core_params
),
1777 ICPU(INTEL_FAM6_ATOM_AIRMONT
, airmont_params
),
1778 ICPU(INTEL_FAM6_SKYLAKE_MOBILE
, core_params
),
1779 ICPU(INTEL_FAM6_BROADWELL_X
, core_params
),
1780 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP
, core_params
),
1781 ICPU(INTEL_FAM6_BROADWELL_XEON_D
, core_params
),
1782 ICPU(INTEL_FAM6_XEON_PHI_KNL
, knl_params
),
1783 ICPU(INTEL_FAM6_XEON_PHI_KNM
, knl_params
),
1784 ICPU(INTEL_FAM6_ATOM_GOLDMONT
, bxt_params
),
1787 MODULE_DEVICE_TABLE(x86cpu
, intel_pstate_cpu_ids
);
1789 static const struct x86_cpu_id intel_pstate_cpu_oob_ids
[] __initconst
= {
1790 ICPU(INTEL_FAM6_BROADWELL_XEON_D
, core_params
),
1791 ICPU(INTEL_FAM6_BROADWELL_X
, core_params
),
1792 ICPU(INTEL_FAM6_SKYLAKE_X
, core_params
),
1796 static int intel_pstate_init_cpu(unsigned int cpunum
)
1798 struct cpudata
*cpu
;
1800 cpu
= all_cpu_data
[cpunum
];
1803 unsigned int size
= sizeof(struct cpudata
);
1806 size
+= sizeof(struct perf_limits
);
1808 cpu
= kzalloc(size
, GFP_KERNEL
);
1812 all_cpu_data
[cpunum
] = cpu
;
1814 cpu
->perf_limits
= (struct perf_limits
*)(cpu
+ 1);
1816 cpu
->epp_default
= -EINVAL
;
1817 cpu
->epp_powersave
= -EINVAL
;
1818 cpu
->epp_saved
= -EINVAL
;
1821 cpu
= all_cpu_data
[cpunum
];
1826 intel_pstate_hwp_enable(cpu
);
1827 pid_params
.sample_rate_ms
= 50;
1828 pid_params
.sample_rate_ns
= 50 * NSEC_PER_MSEC
;
1831 intel_pstate_get_cpu_pstates(cpu
);
1833 intel_pstate_busy_pid_reset(cpu
);
1835 pr_debug("controlling: cpu %d\n", cpunum
);
1840 static unsigned int intel_pstate_get(unsigned int cpu_num
)
1842 struct cpudata
*cpu
= all_cpu_data
[cpu_num
];
1844 return cpu
? get_avg_frequency(cpu
) : 0;
1847 static void intel_pstate_set_update_util_hook(unsigned int cpu_num
)
1849 struct cpudata
*cpu
= all_cpu_data
[cpu_num
];
1851 if (cpu
->update_util_set
)
1854 /* Prevent intel_pstate_update_util() from using stale data. */
1855 cpu
->sample
.time
= 0;
1856 cpufreq_add_update_util_hook(cpu_num
, &cpu
->update_util
,
1857 intel_pstate_update_util
);
1858 cpu
->update_util_set
= true;
1861 static void intel_pstate_clear_update_util_hook(unsigned int cpu
)
1863 struct cpudata
*cpu_data
= all_cpu_data
[cpu
];
1865 if (!cpu_data
->update_util_set
)
1868 cpufreq_remove_update_util_hook(cpu
);
1869 cpu_data
->update_util_set
= false;
1870 synchronize_sched();
1873 static void intel_pstate_set_performance_limits(struct perf_limits
*limits
)
1875 limits
->no_turbo
= 0;
1876 limits
->turbo_disabled
= 0;
1877 limits
->max_perf_pct
= 100;
1878 limits
->max_perf
= int_ext_tofp(1);
1879 limits
->min_perf_pct
= 100;
1880 limits
->min_perf
= int_ext_tofp(1);
1881 limits
->max_policy_pct
= 100;
1882 limits
->max_sysfs_pct
= 100;
1883 limits
->min_policy_pct
= 0;
1884 limits
->min_sysfs_pct
= 0;
1887 static void intel_pstate_update_perf_limits(struct cpufreq_policy
*policy
,
1888 struct perf_limits
*limits
)
1891 limits
->max_policy_pct
= DIV_ROUND_UP(policy
->max
* 100,
1892 policy
->cpuinfo
.max_freq
);
1893 limits
->max_policy_pct
= clamp_t(int, limits
->max_policy_pct
, 0, 100);
1894 if (policy
->max
== policy
->min
) {
1895 limits
->min_policy_pct
= limits
->max_policy_pct
;
1897 limits
->min_policy_pct
= DIV_ROUND_UP(policy
->min
* 100,
1898 policy
->cpuinfo
.max_freq
);
1899 limits
->min_policy_pct
= clamp_t(int, limits
->min_policy_pct
,
1903 /* Normalize user input to [min_policy_pct, max_policy_pct] */
1904 limits
->min_perf_pct
= max(limits
->min_policy_pct
,
1905 limits
->min_sysfs_pct
);
1906 limits
->min_perf_pct
= min(limits
->max_policy_pct
,
1907 limits
->min_perf_pct
);
1908 limits
->max_perf_pct
= min(limits
->max_policy_pct
,
1909 limits
->max_sysfs_pct
);
1910 limits
->max_perf_pct
= max(limits
->min_policy_pct
,
1911 limits
->max_perf_pct
);
1913 /* Make sure min_perf_pct <= max_perf_pct */
1914 limits
->min_perf_pct
= min(limits
->max_perf_pct
, limits
->min_perf_pct
);
1916 limits
->min_perf
= div_ext_fp(limits
->min_perf_pct
, 100);
1917 limits
->max_perf
= div_ext_fp(limits
->max_perf_pct
, 100);
1918 limits
->max_perf
= round_up(limits
->max_perf
, EXT_FRAC_BITS
);
1919 limits
->min_perf
= round_up(limits
->min_perf
, EXT_FRAC_BITS
);
1921 pr_debug("cpu:%d max_perf_pct:%d min_perf_pct:%d\n", policy
->cpu
,
1922 limits
->max_perf_pct
, limits
->min_perf_pct
);
1925 static int intel_pstate_set_policy(struct cpufreq_policy
*policy
)
1927 struct cpudata
*cpu
;
1928 struct perf_limits
*perf_limits
= NULL
;
1930 if (!policy
->cpuinfo
.max_freq
)
1933 pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
1934 policy
->cpuinfo
.max_freq
, policy
->max
);
1936 cpu
= all_cpu_data
[policy
->cpu
];
1937 cpu
->policy
= policy
->policy
;
1939 if (cpu
->pstate
.max_pstate_physical
> cpu
->pstate
.max_pstate
&&
1940 policy
->max
< policy
->cpuinfo
.max_freq
&&
1941 policy
->max
> cpu
->pstate
.max_pstate
* cpu
->pstate
.scaling
) {
1942 pr_debug("policy->max > max non turbo frequency\n");
1943 policy
->max
= policy
->cpuinfo
.max_freq
;
1947 perf_limits
= cpu
->perf_limits
;
1949 mutex_lock(&intel_pstate_limits_lock
);
1951 if (policy
->policy
== CPUFREQ_POLICY_PERFORMANCE
) {
1953 limits
= &performance_limits
;
1954 perf_limits
= limits
;
1956 if (policy
->max
>= policy
->cpuinfo
.max_freq
) {
1957 pr_debug("set performance\n");
1958 intel_pstate_set_performance_limits(perf_limits
);
1962 pr_debug("set powersave\n");
1964 limits
= &powersave_limits
;
1965 perf_limits
= limits
;
1970 intel_pstate_update_perf_limits(policy
, perf_limits
);
1972 if (cpu
->policy
== CPUFREQ_POLICY_PERFORMANCE
) {
1974 * NOHZ_FULL CPUs need this as the governor callback may not
1975 * be invoked on them.
1977 intel_pstate_clear_update_util_hook(policy
->cpu
);
1978 intel_pstate_max_within_limits(cpu
);
1981 intel_pstate_set_update_util_hook(policy
->cpu
);
1983 intel_pstate_hwp_set_policy(policy
);
1985 mutex_unlock(&intel_pstate_limits_lock
);
1990 static int intel_pstate_verify_policy(struct cpufreq_policy
*policy
)
1992 cpufreq_verify_within_cpu_limits(policy
);
1994 if (policy
->policy
!= CPUFREQ_POLICY_POWERSAVE
&&
1995 policy
->policy
!= CPUFREQ_POLICY_PERFORMANCE
)
2001 static void intel_cpufreq_stop_cpu(struct cpufreq_policy
*policy
)
2003 intel_pstate_set_min_pstate(all_cpu_data
[policy
->cpu
]);
2006 static void intel_pstate_stop_cpu(struct cpufreq_policy
*policy
)
2008 pr_debug("CPU %d exiting\n", policy
->cpu
);
2010 intel_pstate_clear_update_util_hook(policy
->cpu
);
2012 intel_pstate_hwp_save_state(policy
);
2014 intel_cpufreq_stop_cpu(policy
);
2017 static int intel_pstate_cpu_exit(struct cpufreq_policy
*policy
)
2019 intel_pstate_exit_perf_limits(policy
);
2021 policy
->fast_switch_possible
= false;
2026 static int __intel_pstate_cpu_init(struct cpufreq_policy
*policy
)
2028 struct cpudata
*cpu
;
2031 rc
= intel_pstate_init_cpu(policy
->cpu
);
2035 cpu
= all_cpu_data
[policy
->cpu
];
2038 * We need sane value in the cpu->perf_limits, so inherit from global
2039 * perf_limits limits, which are seeded with values based on the
2040 * CONFIG_CPU_FREQ_DEFAULT_GOV_*, during boot up.
2043 memcpy(cpu
->perf_limits
, limits
, sizeof(struct perf_limits
));
2045 policy
->min
= cpu
->pstate
.min_pstate
* cpu
->pstate
.scaling
;
2046 policy
->max
= cpu
->pstate
.turbo_pstate
* cpu
->pstate
.scaling
;
2048 /* cpuinfo and default policy values */
2049 policy
->cpuinfo
.min_freq
= cpu
->pstate
.min_pstate
* cpu
->pstate
.scaling
;
2050 update_turbo_state();
2051 policy
->cpuinfo
.max_freq
= limits
->turbo_disabled
?
2052 cpu
->pstate
.max_pstate
: cpu
->pstate
.turbo_pstate
;
2053 policy
->cpuinfo
.max_freq
*= cpu
->pstate
.scaling
;
2055 intel_pstate_init_acpi_perf_limits(policy
);
2056 cpumask_set_cpu(policy
->cpu
, policy
->cpus
);
2058 policy
->fast_switch_possible
= true;
2063 static int intel_pstate_cpu_init(struct cpufreq_policy
*policy
)
2065 int ret
= __intel_pstate_cpu_init(policy
);
2070 policy
->cpuinfo
.transition_latency
= CPUFREQ_ETERNAL
;
2071 if (limits
->min_perf_pct
== 100 && limits
->max_perf_pct
== 100)
2072 policy
->policy
= CPUFREQ_POLICY_PERFORMANCE
;
2074 policy
->policy
= CPUFREQ_POLICY_POWERSAVE
;
2079 static struct cpufreq_driver intel_pstate
= {
2080 .flags
= CPUFREQ_CONST_LOOPS
,
2081 .verify
= intel_pstate_verify_policy
,
2082 .setpolicy
= intel_pstate_set_policy
,
2083 .suspend
= intel_pstate_hwp_save_state
,
2084 .resume
= intel_pstate_resume
,
2085 .get
= intel_pstate_get
,
2086 .init
= intel_pstate_cpu_init
,
2087 .exit
= intel_pstate_cpu_exit
,
2088 .stop_cpu
= intel_pstate_stop_cpu
,
2089 .name
= "intel_pstate",
2092 static int intel_cpufreq_verify_policy(struct cpufreq_policy
*policy
)
2094 struct cpudata
*cpu
= all_cpu_data
[policy
->cpu
];
2095 struct perf_limits
*perf_limits
= limits
;
2097 update_turbo_state();
2098 policy
->cpuinfo
.max_freq
= limits
->turbo_disabled
?
2099 cpu
->pstate
.max_freq
: cpu
->pstate
.turbo_freq
;
2101 cpufreq_verify_within_cpu_limits(policy
);
2104 perf_limits
= cpu
->perf_limits
;
2106 intel_pstate_update_perf_limits(policy
, perf_limits
);
2111 static unsigned int intel_cpufreq_turbo_update(struct cpudata
*cpu
,
2112 struct cpufreq_policy
*policy
,
2113 unsigned int target_freq
)
2115 unsigned int max_freq
;
2117 update_turbo_state();
2119 max_freq
= limits
->no_turbo
|| limits
->turbo_disabled
?
2120 cpu
->pstate
.max_freq
: cpu
->pstate
.turbo_freq
;
2121 policy
->cpuinfo
.max_freq
= max_freq
;
2122 if (policy
->max
> max_freq
)
2123 policy
->max
= max_freq
;
2125 if (target_freq
> max_freq
)
2126 target_freq
= max_freq
;
2131 static int intel_cpufreq_target(struct cpufreq_policy
*policy
,
2132 unsigned int target_freq
,
2133 unsigned int relation
)
2135 struct cpudata
*cpu
= all_cpu_data
[policy
->cpu
];
2136 struct cpufreq_freqs freqs
;
2139 freqs
.old
= policy
->cur
;
2140 freqs
.new = intel_cpufreq_turbo_update(cpu
, policy
, target_freq
);
2142 cpufreq_freq_transition_begin(policy
, &freqs
);
2144 case CPUFREQ_RELATION_L
:
2145 target_pstate
= DIV_ROUND_UP(freqs
.new, cpu
->pstate
.scaling
);
2147 case CPUFREQ_RELATION_H
:
2148 target_pstate
= freqs
.new / cpu
->pstate
.scaling
;
2151 target_pstate
= DIV_ROUND_CLOSEST(freqs
.new, cpu
->pstate
.scaling
);
2154 target_pstate
= intel_pstate_prepare_request(cpu
, target_pstate
);
2155 if (target_pstate
!= cpu
->pstate
.current_pstate
) {
2156 cpu
->pstate
.current_pstate
= target_pstate
;
2157 wrmsrl_on_cpu(policy
->cpu
, MSR_IA32_PERF_CTL
,
2158 pstate_funcs
.get_val(cpu
, target_pstate
));
2160 cpufreq_freq_transition_end(policy
, &freqs
, false);
2165 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy
*policy
,
2166 unsigned int target_freq
)
2168 struct cpudata
*cpu
= all_cpu_data
[policy
->cpu
];
2171 target_freq
= intel_cpufreq_turbo_update(cpu
, policy
, target_freq
);
2172 target_pstate
= DIV_ROUND_UP(target_freq
, cpu
->pstate
.scaling
);
2173 intel_pstate_update_pstate(cpu
, target_pstate
);
2177 static int intel_cpufreq_cpu_init(struct cpufreq_policy
*policy
)
2179 int ret
= __intel_pstate_cpu_init(policy
);
2184 policy
->cpuinfo
.transition_latency
= INTEL_CPUFREQ_TRANSITION_LATENCY
;
2185 /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2186 policy
->cur
= policy
->cpuinfo
.min_freq
;
2191 static struct cpufreq_driver intel_cpufreq
= {
2192 .flags
= CPUFREQ_CONST_LOOPS
,
2193 .verify
= intel_cpufreq_verify_policy
,
2194 .target
= intel_cpufreq_target
,
2195 .fast_switch
= intel_cpufreq_fast_switch
,
2196 .init
= intel_cpufreq_cpu_init
,
2197 .exit
= intel_pstate_cpu_exit
,
2198 .stop_cpu
= intel_cpufreq_stop_cpu
,
2199 .name
= "intel_cpufreq",
2202 static struct cpufreq_driver
*intel_pstate_driver
= &intel_pstate
;
2204 static int no_load __initdata
;
2205 static int no_hwp __initdata
;
2206 static int hwp_only __initdata
;
2207 static unsigned int force_load __initdata
;
2209 static int __init
intel_pstate_msrs_not_valid(void)
2211 if (!pstate_funcs
.get_max() ||
2212 !pstate_funcs
.get_min() ||
2213 !pstate_funcs
.get_turbo())
2219 static void __init
copy_pid_params(struct pstate_adjust_policy
*policy
)
2221 pid_params
.sample_rate_ms
= policy
->sample_rate_ms
;
2222 pid_params
.sample_rate_ns
= pid_params
.sample_rate_ms
* NSEC_PER_MSEC
;
2223 pid_params
.p_gain_pct
= policy
->p_gain_pct
;
2224 pid_params
.i_gain_pct
= policy
->i_gain_pct
;
2225 pid_params
.d_gain_pct
= policy
->d_gain_pct
;
2226 pid_params
.deadband
= policy
->deadband
;
2227 pid_params
.setpoint
= policy
->setpoint
;
2231 static void intel_pstate_use_acpi_profile(void)
2233 if (acpi_gbl_FADT
.preferred_profile
== PM_MOBILE
)
2234 pstate_funcs
.get_target_pstate
=
2235 get_target_pstate_use_cpu_load
;
2238 static void intel_pstate_use_acpi_profile(void)
2243 static void __init
copy_cpu_funcs(struct pstate_funcs
*funcs
)
2245 pstate_funcs
.get_max
= funcs
->get_max
;
2246 pstate_funcs
.get_max_physical
= funcs
->get_max_physical
;
2247 pstate_funcs
.get_min
= funcs
->get_min
;
2248 pstate_funcs
.get_turbo
= funcs
->get_turbo
;
2249 pstate_funcs
.get_scaling
= funcs
->get_scaling
;
2250 pstate_funcs
.get_val
= funcs
->get_val
;
2251 pstate_funcs
.get_vid
= funcs
->get_vid
;
2252 pstate_funcs
.get_target_pstate
= funcs
->get_target_pstate
;
2254 intel_pstate_use_acpi_profile();
2259 static bool __init
intel_pstate_no_acpi_pss(void)
2263 for_each_possible_cpu(i
) {
2265 union acpi_object
*pss
;
2266 struct acpi_buffer buffer
= { ACPI_ALLOCATE_BUFFER
, NULL
};
2267 struct acpi_processor
*pr
= per_cpu(processors
, i
);
2272 status
= acpi_evaluate_object(pr
->handle
, "_PSS", NULL
, &buffer
);
2273 if (ACPI_FAILURE(status
))
2276 pss
= buffer
.pointer
;
2277 if (pss
&& pss
->type
== ACPI_TYPE_PACKAGE
) {
2288 static bool __init
intel_pstate_has_acpi_ppc(void)
2292 for_each_possible_cpu(i
) {
2293 struct acpi_processor
*pr
= per_cpu(processors
, i
);
2297 if (acpi_has_method(pr
->handle
, "_PPC"))
2308 struct hw_vendor_info
{
2310 char oem_id
[ACPI_OEM_ID_SIZE
];
2311 char oem_table_id
[ACPI_OEM_TABLE_ID_SIZE
];
2315 /* Hardware vendor-specific info that has its own power management modes */
2316 static struct hw_vendor_info vendor_info
[] __initdata
= {
2317 {1, "HP ", "ProLiant", PSS
},
2318 {1, "ORACLE", "X4-2 ", PPC
},
2319 {1, "ORACLE", "X4-2L ", PPC
},
2320 {1, "ORACLE", "X4-2B ", PPC
},
2321 {1, "ORACLE", "X3-2 ", PPC
},
2322 {1, "ORACLE", "X3-2L ", PPC
},
2323 {1, "ORACLE", "X3-2B ", PPC
},
2324 {1, "ORACLE", "X4470M2 ", PPC
},
2325 {1, "ORACLE", "X4270M3 ", PPC
},
2326 {1, "ORACLE", "X4270M2 ", PPC
},
2327 {1, "ORACLE", "X4170M2 ", PPC
},
2328 {1, "ORACLE", "X4170 M3", PPC
},
2329 {1, "ORACLE", "X4275 M3", PPC
},
2330 {1, "ORACLE", "X6-2 ", PPC
},
2331 {1, "ORACLE", "Sudbury ", PPC
},
2335 static bool __init
intel_pstate_platform_pwr_mgmt_exists(void)
2337 struct acpi_table_header hdr
;
2338 struct hw_vendor_info
*v_info
;
2339 const struct x86_cpu_id
*id
;
2342 id
= x86_match_cpu(intel_pstate_cpu_oob_ids
);
2344 rdmsrl(MSR_MISC_PWR_MGMT
, misc_pwr
);
2345 if ( misc_pwr
& (1 << 8))
2349 if (acpi_disabled
||
2350 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT
, 0, &hdr
)))
2353 for (v_info
= vendor_info
; v_info
->valid
; v_info
++) {
2354 if (!strncmp(hdr
.oem_id
, v_info
->oem_id
, ACPI_OEM_ID_SIZE
) &&
2355 !strncmp(hdr
.oem_table_id
, v_info
->oem_table_id
,
2356 ACPI_OEM_TABLE_ID_SIZE
))
2357 switch (v_info
->oem_pwr_table
) {
2359 return intel_pstate_no_acpi_pss();
2361 return intel_pstate_has_acpi_ppc() &&
2369 static void intel_pstate_request_control_from_smm(void)
2372 * It may be unsafe to request P-states control from SMM if _PPC support
2373 * has not been enabled.
2376 acpi_processor_pstate_control();
2378 #else /* CONFIG_ACPI not enabled */
2379 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2380 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2381 static inline void intel_pstate_request_control_from_smm(void) {}
2382 #endif /* CONFIG_ACPI */
2384 static const struct x86_cpu_id hwp_support_ids
[] __initconst
= {
2385 { X86_VENDOR_INTEL
, 6, X86_MODEL_ANY
, X86_FEATURE_HWP
},
2389 static int __init
intel_pstate_init(void)
2392 const struct x86_cpu_id
*id
;
2393 struct cpu_defaults
*cpu_def
;
2398 if (x86_match_cpu(hwp_support_ids
) && !no_hwp
) {
2399 copy_cpu_funcs(&core_params
.funcs
);
2401 intel_pstate
.attr
= hwp_cpufreq_attrs
;
2402 goto hwp_cpu_matched
;
2405 id
= x86_match_cpu(intel_pstate_cpu_ids
);
2409 cpu_def
= (struct cpu_defaults
*)id
->driver_data
;
2411 copy_pid_params(&cpu_def
->pid_policy
);
2412 copy_cpu_funcs(&cpu_def
->funcs
);
2414 if (intel_pstate_msrs_not_valid())
2419 * The Intel pstate driver will be ignored if the platform
2420 * firmware has its own power management modes.
2422 if (intel_pstate_platform_pwr_mgmt_exists())
2425 pr_info("Intel P-state driver initializing\n");
2427 all_cpu_data
= vzalloc(sizeof(void *) * num_possible_cpus());
2431 if (!hwp_active
&& hwp_only
)
2434 intel_pstate_request_control_from_smm();
2436 rc
= cpufreq_register_driver(intel_pstate_driver
);
2440 if (intel_pstate_driver
== &intel_pstate
&& !hwp_active
&&
2441 pstate_funcs
.get_target_pstate
!= get_target_pstate_use_cpu_load
)
2442 intel_pstate_debug_expose_params();
2444 intel_pstate_sysfs_expose_params();
2447 pr_info("HWP enabled\n");
2452 for_each_online_cpu(cpu
) {
2453 if (all_cpu_data
[cpu
]) {
2454 if (intel_pstate_driver
== &intel_pstate
)
2455 intel_pstate_clear_update_util_hook(cpu
);
2457 kfree(all_cpu_data
[cpu
]);
2462 vfree(all_cpu_data
);
2465 device_initcall(intel_pstate_init
);
2467 static int __init
intel_pstate_setup(char *str
)
2472 if (!strcmp(str
, "disable")) {
2474 } else if (!strcmp(str
, "passive")) {
2475 pr_info("Passive mode enabled\n");
2476 intel_pstate_driver
= &intel_cpufreq
;
2479 if (!strcmp(str
, "no_hwp")) {
2480 pr_info("HWP disabled\n");
2483 if (!strcmp(str
, "force"))
2485 if (!strcmp(str
, "hwp_only"))
2487 if (!strcmp(str
, "per_cpu_perf_limits"))
2488 per_cpu_limits
= true;
2491 if (!strcmp(str
, "support_acpi_ppc"))
2497 early_param("intel_pstate", intel_pstate_setup
);
2499 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2500 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2501 MODULE_LICENSE("GPL");