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/cpufreq.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
25 #include <linux/cpufreq.h>
26 #include <linux/sysfs.h>
27 #include <linux/types.h>
29 #include <linux/acpi.h>
30 #include <linux/vmalloc.h>
31 #include <trace/events/power.h>
33 #include <asm/div64.h>
35 #include <asm/cpu_device_id.h>
36 #include <asm/cpufeature.h>
37 #include <asm/intel-family.h>
39 #define INTEL_PSTATE_SAMPLING_INTERVAL (10 * NSEC_PER_MSEC)
41 #define INTEL_CPUFREQ_TRANSITION_LATENCY 20000
42 #define INTEL_CPUFREQ_TRANSITION_DELAY 500
45 #include <acpi/processor.h>
46 #include <acpi/cppc_acpi.h>
50 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
51 #define fp_toint(X) ((X) >> FRAC_BITS)
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)
58 static inline int32_t mul_fp(int32_t x
, int32_t y
)
60 return ((int64_t)x
* (int64_t)y
) >> FRAC_BITS
;
63 static inline int32_t div_fp(s64 x
, s64 y
)
65 return div64_s64((int64_t)x
<< FRAC_BITS
, y
);
68 static inline int ceiling_fp(int32_t x
)
73 mask
= (1 << FRAC_BITS
) - 1;
79 static inline int32_t percent_fp(int percent
)
81 return div_fp(percent
, 100);
84 static inline u64
mul_ext_fp(u64 x
, u64 y
)
86 return (x
* y
) >> EXT_FRAC_BITS
;
89 static inline u64
div_ext_fp(u64 x
, u64 y
)
91 return div64_u64(x
<< EXT_FRAC_BITS
, y
);
94 static inline int32_t percent_ext_fp(int percent
)
96 return div_ext_fp(percent
, 100);
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
112 * @time: Current time from scheduler
114 * This structure is used in the cpudata structure to store performance sample
115 * data for choosing next P State.
118 int32_t core_avg_perf
;
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
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
140 * Stores the per cpu model P state limits and current P state.
146 int max_pstate_physical
;
149 unsigned int max_freq
;
150 unsigned int turbo_freq
;
154 * struct vid_data - Stores voltage information data
155 * @min: VID data for this platform corresponding to
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)
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.
174 * struct global_params - Global parameters, mostly tunable via sysfs.
175 * @no_turbo: Whether or not to use turbo P-states.
176 * @turbo_disabled: Whethet or not turbo P-states are available at all,
177 * based on the MSR_IA32_MISC_ENABLE value and whether or
178 * not the maximum reported turbo P-state is different from
179 * the maximum reported non-turbo one.
180 * @min_perf_pct: Minimum capacity limit in percent of the maximum turbo
182 * @max_perf_pct: Maximum capacity limit in percent of the maximum turbo
185 struct global_params
{
193 * struct cpudata - Per CPU instance data storage
194 * @cpu: CPU number for this instance data
195 * @policy: CPUFreq policy value
196 * @update_util: CPUFreq utility callback information
197 * @update_util_set: CPUFreq utility callback is set
198 * @iowait_boost: iowait-related boost fraction
199 * @last_update: Time of the last update.
200 * @pstate: Stores P state limits for this CPU
201 * @vid: Stores VID limits for this CPU
202 * @last_sample_time: Last Sample time
203 * @aperf_mperf_shift: Number of clock cycles after aperf, merf is incremented
204 * This shift is a multiplier to mperf delta to
205 * calculate CPU busy.
206 * @prev_aperf: Last APERF value read from APERF MSR
207 * @prev_mperf: Last MPERF value read from MPERF MSR
208 * @prev_tsc: Last timestamp counter (TSC) value
209 * @prev_cummulative_iowait: IO Wait time difference from last and
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
222 * @epp_saved: Saved EPP/EPB during system suspend or CPU offline
224 * @hwp_req_cached: Cached value of the last HWP Request MSR
225 * @hwp_cap_cached: Cached value of the last HWP Capabilities MSR
226 * @last_io_update: Last time when IO wake flag was set
227 * @sched_flags: Store scheduler flags for possible cross CPU update
228 * @hwp_boost_min: Last HWP boosted min performance
230 * This structure stores per CPU instance data for all CPUs.
236 struct update_util_data update_util
;
237 bool update_util_set
;
239 struct pstate_data pstate
;
243 u64 last_sample_time
;
244 u64 aperf_mperf_shift
;
248 u64 prev_cummulative_iowait
;
249 struct sample sample
;
250 int32_t min_perf_ratio
;
251 int32_t max_perf_ratio
;
253 struct acpi_processor_performance acpi_perf_data
;
254 bool valid_pss_table
;
256 unsigned int iowait_boost
;
264 unsigned int sched_flags
;
268 static struct cpudata
**all_cpu_data
;
271 * struct pstate_funcs - Per CPU model specific callbacks
272 * @get_max: Callback to get maximum non turbo effective P state
273 * @get_max_physical: Callback to get maximum non turbo physical P state
274 * @get_min: Callback to get minimum P state
275 * @get_turbo: Callback to get turbo P state
276 * @get_scaling: Callback to get frequency scaling factor
277 * @get_val: Callback to convert P state to actual MSR write value
278 * @get_vid: Callback to get VID data for Atom platforms
280 * Core and Atom CPU models have different way to get P State limits. This
281 * structure is used to store those callbacks.
283 struct pstate_funcs
{
284 int (*get_max
)(void);
285 int (*get_max_physical
)(void);
286 int (*get_min
)(void);
287 int (*get_turbo
)(void);
288 int (*get_scaling
)(void);
289 int (*get_aperf_mperf_shift
)(void);
290 u64 (*get_val
)(struct cpudata
*, int pstate
);
291 void (*get_vid
)(struct cpudata
*);
294 static struct pstate_funcs pstate_funcs __read_mostly
;
296 static int hwp_active __read_mostly
;
297 static int hwp_mode_bdw __read_mostly
;
298 static bool per_cpu_limits __read_mostly
;
299 static bool hwp_boost __read_mostly
;
301 static struct cpufreq_driver
*intel_pstate_driver __read_mostly
;
304 static bool acpi_ppc
;
307 static struct global_params global
;
309 static DEFINE_MUTEX(intel_pstate_driver_lock
);
310 static DEFINE_MUTEX(intel_pstate_limits_lock
);
314 static bool intel_pstate_get_ppc_enable_status(void)
316 if (acpi_gbl_FADT
.preferred_profile
== PM_ENTERPRISE_SERVER
||
317 acpi_gbl_FADT
.preferred_profile
== PM_PERFORMANCE_SERVER
)
323 #ifdef CONFIG_ACPI_CPPC_LIB
325 /* The work item is needed to avoid CPU hotplug locking issues */
326 static void intel_pstste_sched_itmt_work_fn(struct work_struct
*work
)
328 sched_set_itmt_support();
331 static DECLARE_WORK(sched_itmt_work
, intel_pstste_sched_itmt_work_fn
);
333 static void intel_pstate_set_itmt_prio(int cpu
)
335 struct cppc_perf_caps cppc_perf
;
336 static u32 max_highest_perf
= 0, min_highest_perf
= U32_MAX
;
339 ret
= cppc_get_perf_caps(cpu
, &cppc_perf
);
344 * The priorities can be set regardless of whether or not
345 * sched_set_itmt_support(true) has been called and it is valid to
346 * update them at any time after it has been called.
348 sched_set_itmt_core_prio(cppc_perf
.highest_perf
, cpu
);
350 if (max_highest_perf
<= min_highest_perf
) {
351 if (cppc_perf
.highest_perf
> max_highest_perf
)
352 max_highest_perf
= cppc_perf
.highest_perf
;
354 if (cppc_perf
.highest_perf
< min_highest_perf
)
355 min_highest_perf
= cppc_perf
.highest_perf
;
357 if (max_highest_perf
> min_highest_perf
) {
359 * This code can be run during CPU online under the
360 * CPU hotplug locks, so sched_set_itmt_support()
361 * cannot be called from here. Queue up a work item
364 schedule_work(&sched_itmt_work
);
369 static void intel_pstate_set_itmt_prio(int cpu
)
374 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy
*policy
)
381 intel_pstate_set_itmt_prio(policy
->cpu
);
385 if (!intel_pstate_get_ppc_enable_status())
388 cpu
= all_cpu_data
[policy
->cpu
];
390 ret
= acpi_processor_register_performance(&cpu
->acpi_perf_data
,
396 * Check if the control value in _PSS is for PERF_CTL MSR, which should
397 * guarantee that the states returned by it map to the states in our
400 if (cpu
->acpi_perf_data
.control_register
.space_id
!=
401 ACPI_ADR_SPACE_FIXED_HARDWARE
)
405 * If there is only one entry _PSS, simply ignore _PSS and continue as
406 * usual without taking _PSS into account
408 if (cpu
->acpi_perf_data
.state_count
< 2)
411 pr_debug("CPU%u - ACPI _PSS perf data\n", policy
->cpu
);
412 for (i
= 0; i
< cpu
->acpi_perf_data
.state_count
; i
++) {
413 pr_debug(" %cP%d: %u MHz, %u mW, 0x%x\n",
414 (i
== cpu
->acpi_perf_data
.state
? '*' : ' '), i
,
415 (u32
) cpu
->acpi_perf_data
.states
[i
].core_frequency
,
416 (u32
) cpu
->acpi_perf_data
.states
[i
].power
,
417 (u32
) cpu
->acpi_perf_data
.states
[i
].control
);
421 * The _PSS table doesn't contain whole turbo frequency range.
422 * This just contains +1 MHZ above the max non turbo frequency,
423 * with control value corresponding to max turbo ratio. But
424 * when cpufreq set policy is called, it will call with this
425 * max frequency, which will cause a reduced performance as
426 * this driver uses real max turbo frequency as the max
427 * frequency. So correct this frequency in _PSS table to
428 * correct max turbo frequency based on the turbo state.
429 * Also need to convert to MHz as _PSS freq is in MHz.
431 if (!global
.turbo_disabled
)
432 cpu
->acpi_perf_data
.states
[0].core_frequency
=
433 policy
->cpuinfo
.max_freq
/ 1000;
434 cpu
->valid_pss_table
= true;
435 pr_debug("_PPC limits will be enforced\n");
440 cpu
->valid_pss_table
= false;
441 acpi_processor_unregister_performance(policy
->cpu
);
444 static void intel_pstate_exit_perf_limits(struct cpufreq_policy
*policy
)
448 cpu
= all_cpu_data
[policy
->cpu
];
449 if (!cpu
->valid_pss_table
)
452 acpi_processor_unregister_performance(policy
->cpu
);
455 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy
*policy
)
459 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy
*policy
)
464 static inline void update_turbo_state(void)
469 cpu
= all_cpu_data
[0];
470 rdmsrl(MSR_IA32_MISC_ENABLE
, misc_en
);
471 global
.turbo_disabled
=
472 (misc_en
& MSR_IA32_MISC_ENABLE_TURBO_DISABLE
||
473 cpu
->pstate
.max_pstate
== cpu
->pstate
.turbo_pstate
);
476 static int min_perf_pct_min(void)
478 struct cpudata
*cpu
= all_cpu_data
[0];
479 int turbo_pstate
= cpu
->pstate
.turbo_pstate
;
481 return turbo_pstate
?
482 (cpu
->pstate
.min_pstate
* 100 / turbo_pstate
) : 0;
485 static s16
intel_pstate_get_epb(struct cpudata
*cpu_data
)
490 if (!static_cpu_has(X86_FEATURE_EPB
))
493 ret
= rdmsrl_on_cpu(cpu_data
->cpu
, MSR_IA32_ENERGY_PERF_BIAS
, &epb
);
497 return (s16
)(epb
& 0x0f);
500 static s16
intel_pstate_get_epp(struct cpudata
*cpu_data
, u64 hwp_req_data
)
504 if (static_cpu_has(X86_FEATURE_HWP_EPP
)) {
506 * When hwp_req_data is 0, means that caller didn't read
507 * MSR_HWP_REQUEST, so need to read and get EPP.
510 epp
= rdmsrl_on_cpu(cpu_data
->cpu
, MSR_HWP_REQUEST
,
515 epp
= (hwp_req_data
>> 24) & 0xff;
517 /* When there is no EPP present, HWP uses EPB settings */
518 epp
= intel_pstate_get_epb(cpu_data
);
524 static int intel_pstate_set_epb(int cpu
, s16 pref
)
529 if (!static_cpu_has(X86_FEATURE_EPB
))
532 ret
= rdmsrl_on_cpu(cpu
, MSR_IA32_ENERGY_PERF_BIAS
, &epb
);
536 epb
= (epb
& ~0x0f) | pref
;
537 wrmsrl_on_cpu(cpu
, MSR_IA32_ENERGY_PERF_BIAS
, epb
);
543 * EPP/EPB display strings corresponding to EPP index in the
544 * energy_perf_strings[]
546 *-------------------------------------
549 * 2 balance_performance
553 static const char * const energy_perf_strings
[] = {
556 "balance_performance",
561 static const unsigned int epp_values
[] = {
563 HWP_EPP_BALANCE_PERFORMANCE
,
564 HWP_EPP_BALANCE_POWERSAVE
,
568 static int intel_pstate_get_energy_pref_index(struct cpudata
*cpu_data
)
573 epp
= intel_pstate_get_epp(cpu_data
, 0);
577 if (static_cpu_has(X86_FEATURE_HWP_EPP
)) {
578 if (epp
== HWP_EPP_PERFORMANCE
)
580 if (epp
<= HWP_EPP_BALANCE_PERFORMANCE
)
582 if (epp
<= HWP_EPP_BALANCE_POWERSAVE
)
586 } else if (static_cpu_has(X86_FEATURE_EPB
)) {
589 * 0x00-0x03 : Performance
590 * 0x04-0x07 : Balance performance
591 * 0x08-0x0B : Balance power
593 * The EPB is a 4 bit value, but our ranges restrict the
594 * value which can be set. Here only using top two bits
597 index
= (epp
>> 2) + 1;
603 static int intel_pstate_set_energy_pref_index(struct cpudata
*cpu_data
,
610 epp
= cpu_data
->epp_default
;
612 mutex_lock(&intel_pstate_limits_lock
);
614 if (static_cpu_has(X86_FEATURE_HWP_EPP
)) {
617 ret
= rdmsrl_on_cpu(cpu_data
->cpu
, MSR_HWP_REQUEST
, &value
);
621 value
&= ~GENMASK_ULL(31, 24);
624 epp
= epp_values
[pref_index
- 1];
626 value
|= (u64
)epp
<< 24;
627 ret
= wrmsrl_on_cpu(cpu_data
->cpu
, MSR_HWP_REQUEST
, value
);
630 epp
= (pref_index
- 1) << 2;
631 ret
= intel_pstate_set_epb(cpu_data
->cpu
, epp
);
634 mutex_unlock(&intel_pstate_limits_lock
);
639 static ssize_t
show_energy_performance_available_preferences(
640 struct cpufreq_policy
*policy
, char *buf
)
645 while (energy_perf_strings
[i
] != NULL
)
646 ret
+= sprintf(&buf
[ret
], "%s ", energy_perf_strings
[i
++]);
648 ret
+= sprintf(&buf
[ret
], "\n");
653 cpufreq_freq_attr_ro(energy_performance_available_preferences
);
655 static ssize_t
store_energy_performance_preference(
656 struct cpufreq_policy
*policy
, const char *buf
, size_t count
)
658 struct cpudata
*cpu_data
= all_cpu_data
[policy
->cpu
];
659 char str_preference
[21];
662 ret
= sscanf(buf
, "%20s", str_preference
);
666 while (energy_perf_strings
[i
] != NULL
) {
667 if (!strcmp(str_preference
, energy_perf_strings
[i
])) {
668 intel_pstate_set_energy_pref_index(cpu_data
, i
);
677 static ssize_t
show_energy_performance_preference(
678 struct cpufreq_policy
*policy
, char *buf
)
680 struct cpudata
*cpu_data
= all_cpu_data
[policy
->cpu
];
683 preference
= intel_pstate_get_energy_pref_index(cpu_data
);
687 return sprintf(buf
, "%s\n", energy_perf_strings
[preference
]);
690 cpufreq_freq_attr_rw(energy_performance_preference
);
692 static struct freq_attr
*hwp_cpufreq_attrs
[] = {
693 &energy_performance_preference
,
694 &energy_performance_available_preferences
,
698 static void intel_pstate_get_hwp_max(unsigned int cpu
, int *phy_max
,
703 rdmsrl_on_cpu(cpu
, MSR_HWP_CAPABILITIES
, &cap
);
704 WRITE_ONCE(all_cpu_data
[cpu
]->hwp_cap_cached
, cap
);
706 *current_max
= HWP_GUARANTEED_PERF(cap
);
708 *current_max
= HWP_HIGHEST_PERF(cap
);
710 *phy_max
= HWP_HIGHEST_PERF(cap
);
713 static void intel_pstate_hwp_set(unsigned int cpu
)
715 struct cpudata
*cpu_data
= all_cpu_data
[cpu
];
720 max
= cpu_data
->max_perf_ratio
;
721 min
= cpu_data
->min_perf_ratio
;
723 if (cpu_data
->policy
== CPUFREQ_POLICY_PERFORMANCE
)
726 rdmsrl_on_cpu(cpu
, MSR_HWP_REQUEST
, &value
);
728 value
&= ~HWP_MIN_PERF(~0L);
729 value
|= HWP_MIN_PERF(min
);
731 value
&= ~HWP_MAX_PERF(~0L);
732 value
|= HWP_MAX_PERF(max
);
734 if (cpu_data
->epp_policy
== cpu_data
->policy
)
737 cpu_data
->epp_policy
= cpu_data
->policy
;
739 if (cpu_data
->epp_saved
>= 0) {
740 epp
= cpu_data
->epp_saved
;
741 cpu_data
->epp_saved
= -EINVAL
;
745 if (cpu_data
->policy
== CPUFREQ_POLICY_PERFORMANCE
) {
746 epp
= intel_pstate_get_epp(cpu_data
, value
);
747 cpu_data
->epp_powersave
= epp
;
748 /* If EPP read was failed, then don't try to write */
754 /* skip setting EPP, when saved value is invalid */
755 if (cpu_data
->epp_powersave
< 0)
759 * No need to restore EPP when it is not zero. This
761 * - Policy is not changed
762 * - user has manually changed
763 * - Error reading EPB
765 epp
= intel_pstate_get_epp(cpu_data
, value
);
769 epp
= cpu_data
->epp_powersave
;
772 if (static_cpu_has(X86_FEATURE_HWP_EPP
)) {
773 value
&= ~GENMASK_ULL(31, 24);
774 value
|= (u64
)epp
<< 24;
776 intel_pstate_set_epb(cpu
, epp
);
779 WRITE_ONCE(cpu_data
->hwp_req_cached
, value
);
780 wrmsrl_on_cpu(cpu
, MSR_HWP_REQUEST
, value
);
783 static int intel_pstate_hwp_save_state(struct cpufreq_policy
*policy
)
785 struct cpudata
*cpu_data
= all_cpu_data
[policy
->cpu
];
790 cpu_data
->epp_saved
= intel_pstate_get_epp(cpu_data
, 0);
795 static void intel_pstate_hwp_enable(struct cpudata
*cpudata
);
797 static int intel_pstate_resume(struct cpufreq_policy
*policy
)
802 mutex_lock(&intel_pstate_limits_lock
);
804 if (policy
->cpu
== 0)
805 intel_pstate_hwp_enable(all_cpu_data
[policy
->cpu
]);
807 all_cpu_data
[policy
->cpu
]->epp_policy
= 0;
808 intel_pstate_hwp_set(policy
->cpu
);
810 mutex_unlock(&intel_pstate_limits_lock
);
815 static void intel_pstate_update_policies(void)
819 for_each_possible_cpu(cpu
)
820 cpufreq_update_policy(cpu
);
823 /************************** sysfs begin ************************/
824 #define show_one(file_name, object) \
825 static ssize_t show_##file_name \
826 (struct kobject *kobj, struct attribute *attr, char *buf) \
828 return sprintf(buf, "%u\n", global.object); \
831 static ssize_t
intel_pstate_show_status(char *buf
);
832 static int intel_pstate_update_status(const char *buf
, size_t size
);
834 static ssize_t
show_status(struct kobject
*kobj
,
835 struct attribute
*attr
, char *buf
)
839 mutex_lock(&intel_pstate_driver_lock
);
840 ret
= intel_pstate_show_status(buf
);
841 mutex_unlock(&intel_pstate_driver_lock
);
846 static ssize_t
store_status(struct kobject
*a
, struct attribute
*b
,
847 const char *buf
, size_t count
)
849 char *p
= memchr(buf
, '\n', count
);
852 mutex_lock(&intel_pstate_driver_lock
);
853 ret
= intel_pstate_update_status(buf
, p
? p
- buf
: count
);
854 mutex_unlock(&intel_pstate_driver_lock
);
856 return ret
< 0 ? ret
: count
;
859 static ssize_t
show_turbo_pct(struct kobject
*kobj
,
860 struct attribute
*attr
, char *buf
)
863 int total
, no_turbo
, turbo_pct
;
866 mutex_lock(&intel_pstate_driver_lock
);
868 if (!intel_pstate_driver
) {
869 mutex_unlock(&intel_pstate_driver_lock
);
873 cpu
= all_cpu_data
[0];
875 total
= cpu
->pstate
.turbo_pstate
- cpu
->pstate
.min_pstate
+ 1;
876 no_turbo
= cpu
->pstate
.max_pstate
- cpu
->pstate
.min_pstate
+ 1;
877 turbo_fp
= div_fp(no_turbo
, total
);
878 turbo_pct
= 100 - fp_toint(mul_fp(turbo_fp
, int_tofp(100)));
880 mutex_unlock(&intel_pstate_driver_lock
);
882 return sprintf(buf
, "%u\n", turbo_pct
);
885 static ssize_t
show_num_pstates(struct kobject
*kobj
,
886 struct attribute
*attr
, char *buf
)
891 mutex_lock(&intel_pstate_driver_lock
);
893 if (!intel_pstate_driver
) {
894 mutex_unlock(&intel_pstate_driver_lock
);
898 cpu
= all_cpu_data
[0];
899 total
= cpu
->pstate
.turbo_pstate
- cpu
->pstate
.min_pstate
+ 1;
901 mutex_unlock(&intel_pstate_driver_lock
);
903 return sprintf(buf
, "%u\n", total
);
906 static ssize_t
show_no_turbo(struct kobject
*kobj
,
907 struct attribute
*attr
, char *buf
)
911 mutex_lock(&intel_pstate_driver_lock
);
913 if (!intel_pstate_driver
) {
914 mutex_unlock(&intel_pstate_driver_lock
);
918 update_turbo_state();
919 if (global
.turbo_disabled
)
920 ret
= sprintf(buf
, "%u\n", global
.turbo_disabled
);
922 ret
= sprintf(buf
, "%u\n", global
.no_turbo
);
924 mutex_unlock(&intel_pstate_driver_lock
);
929 static ssize_t
store_no_turbo(struct kobject
*a
, struct attribute
*b
,
930 const char *buf
, size_t count
)
935 ret
= sscanf(buf
, "%u", &input
);
939 mutex_lock(&intel_pstate_driver_lock
);
941 if (!intel_pstate_driver
) {
942 mutex_unlock(&intel_pstate_driver_lock
);
946 mutex_lock(&intel_pstate_limits_lock
);
948 update_turbo_state();
949 if (global
.turbo_disabled
) {
950 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
951 mutex_unlock(&intel_pstate_limits_lock
);
952 mutex_unlock(&intel_pstate_driver_lock
);
956 global
.no_turbo
= clamp_t(int, input
, 0, 1);
958 if (global
.no_turbo
) {
959 struct cpudata
*cpu
= all_cpu_data
[0];
960 int pct
= cpu
->pstate
.max_pstate
* 100 / cpu
->pstate
.turbo_pstate
;
962 /* Squash the global minimum into the permitted range. */
963 if (global
.min_perf_pct
> pct
)
964 global
.min_perf_pct
= pct
;
967 mutex_unlock(&intel_pstate_limits_lock
);
969 intel_pstate_update_policies();
971 mutex_unlock(&intel_pstate_driver_lock
);
976 static ssize_t
store_max_perf_pct(struct kobject
*a
, struct attribute
*b
,
977 const char *buf
, size_t count
)
982 ret
= sscanf(buf
, "%u", &input
);
986 mutex_lock(&intel_pstate_driver_lock
);
988 if (!intel_pstate_driver
) {
989 mutex_unlock(&intel_pstate_driver_lock
);
993 mutex_lock(&intel_pstate_limits_lock
);
995 global
.max_perf_pct
= clamp_t(int, input
, global
.min_perf_pct
, 100);
997 mutex_unlock(&intel_pstate_limits_lock
);
999 intel_pstate_update_policies();
1001 mutex_unlock(&intel_pstate_driver_lock
);
1006 static ssize_t
store_min_perf_pct(struct kobject
*a
, struct attribute
*b
,
1007 const char *buf
, size_t count
)
1012 ret
= sscanf(buf
, "%u", &input
);
1016 mutex_lock(&intel_pstate_driver_lock
);
1018 if (!intel_pstate_driver
) {
1019 mutex_unlock(&intel_pstate_driver_lock
);
1023 mutex_lock(&intel_pstate_limits_lock
);
1025 global
.min_perf_pct
= clamp_t(int, input
,
1026 min_perf_pct_min(), global
.max_perf_pct
);
1028 mutex_unlock(&intel_pstate_limits_lock
);
1030 intel_pstate_update_policies();
1032 mutex_unlock(&intel_pstate_driver_lock
);
1037 static ssize_t
show_hwp_dynamic_boost(struct kobject
*kobj
,
1038 struct attribute
*attr
, char *buf
)
1040 return sprintf(buf
, "%u\n", hwp_boost
);
1043 static ssize_t
store_hwp_dynamic_boost(struct kobject
*a
, struct attribute
*b
,
1044 const char *buf
, size_t count
)
1049 ret
= kstrtouint(buf
, 10, &input
);
1053 mutex_lock(&intel_pstate_driver_lock
);
1054 hwp_boost
= !!input
;
1055 intel_pstate_update_policies();
1056 mutex_unlock(&intel_pstate_driver_lock
);
1061 show_one(max_perf_pct
, max_perf_pct
);
1062 show_one(min_perf_pct
, min_perf_pct
);
1064 define_one_global_rw(status
);
1065 define_one_global_rw(no_turbo
);
1066 define_one_global_rw(max_perf_pct
);
1067 define_one_global_rw(min_perf_pct
);
1068 define_one_global_ro(turbo_pct
);
1069 define_one_global_ro(num_pstates
);
1070 define_one_global_rw(hwp_dynamic_boost
);
1072 static struct attribute
*intel_pstate_attributes
[] = {
1080 static const struct attribute_group intel_pstate_attr_group
= {
1081 .attrs
= intel_pstate_attributes
,
1084 static void __init
intel_pstate_sysfs_expose_params(void)
1086 struct kobject
*intel_pstate_kobject
;
1089 intel_pstate_kobject
= kobject_create_and_add("intel_pstate",
1090 &cpu_subsys
.dev_root
->kobj
);
1091 if (WARN_ON(!intel_pstate_kobject
))
1094 rc
= sysfs_create_group(intel_pstate_kobject
, &intel_pstate_attr_group
);
1099 * If per cpu limits are enforced there are no global limits, so
1100 * return without creating max/min_perf_pct attributes
1105 rc
= sysfs_create_file(intel_pstate_kobject
, &max_perf_pct
.attr
);
1108 rc
= sysfs_create_file(intel_pstate_kobject
, &min_perf_pct
.attr
);
1112 rc
= sysfs_create_file(intel_pstate_kobject
,
1113 &hwp_dynamic_boost
.attr
);
1117 /************************** sysfs end ************************/
1119 static void intel_pstate_hwp_enable(struct cpudata
*cpudata
)
1121 /* First disable HWP notification interrupt as we don't process them */
1122 if (static_cpu_has(X86_FEATURE_HWP_NOTIFY
))
1123 wrmsrl_on_cpu(cpudata
->cpu
, MSR_HWP_INTERRUPT
, 0x00);
1125 wrmsrl_on_cpu(cpudata
->cpu
, MSR_PM_ENABLE
, 0x1);
1126 cpudata
->epp_policy
= 0;
1127 if (cpudata
->epp_default
== -EINVAL
)
1128 cpudata
->epp_default
= intel_pstate_get_epp(cpudata
, 0);
1131 #define MSR_IA32_POWER_CTL_BIT_EE 19
1133 /* Disable energy efficiency optimization */
1134 static void intel_pstate_disable_ee(int cpu
)
1139 ret
= rdmsrl_on_cpu(cpu
, MSR_IA32_POWER_CTL
, &power_ctl
);
1143 if (!(power_ctl
& BIT(MSR_IA32_POWER_CTL_BIT_EE
))) {
1144 pr_info("Disabling energy efficiency optimization\n");
1145 power_ctl
|= BIT(MSR_IA32_POWER_CTL_BIT_EE
);
1146 wrmsrl_on_cpu(cpu
, MSR_IA32_POWER_CTL
, power_ctl
);
1150 static int atom_get_min_pstate(void)
1154 rdmsrl(MSR_ATOM_CORE_RATIOS
, value
);
1155 return (value
>> 8) & 0x7F;
1158 static int atom_get_max_pstate(void)
1162 rdmsrl(MSR_ATOM_CORE_RATIOS
, value
);
1163 return (value
>> 16) & 0x7F;
1166 static int atom_get_turbo_pstate(void)
1170 rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS
, value
);
1171 return value
& 0x7F;
1174 static u64
atom_get_val(struct cpudata
*cpudata
, int pstate
)
1180 val
= (u64
)pstate
<< 8;
1181 if (global
.no_turbo
&& !global
.turbo_disabled
)
1182 val
|= (u64
)1 << 32;
1184 vid_fp
= cpudata
->vid
.min
+ mul_fp(
1185 int_tofp(pstate
- cpudata
->pstate
.min_pstate
),
1186 cpudata
->vid
.ratio
);
1188 vid_fp
= clamp_t(int32_t, vid_fp
, cpudata
->vid
.min
, cpudata
->vid
.max
);
1189 vid
= ceiling_fp(vid_fp
);
1191 if (pstate
> cpudata
->pstate
.max_pstate
)
1192 vid
= cpudata
->vid
.turbo
;
1197 static int silvermont_get_scaling(void)
1201 /* Defined in Table 35-6 from SDM (Sept 2015) */
1202 static int silvermont_freq_table
[] = {
1203 83300, 100000, 133300, 116700, 80000};
1205 rdmsrl(MSR_FSB_FREQ
, value
);
1209 return silvermont_freq_table
[i
];
1212 static int airmont_get_scaling(void)
1216 /* Defined in Table 35-10 from SDM (Sept 2015) */
1217 static int airmont_freq_table
[] = {
1218 83300, 100000, 133300, 116700, 80000,
1219 93300, 90000, 88900, 87500};
1221 rdmsrl(MSR_FSB_FREQ
, value
);
1225 return airmont_freq_table
[i
];
1228 static void atom_get_vid(struct cpudata
*cpudata
)
1232 rdmsrl(MSR_ATOM_CORE_VIDS
, value
);
1233 cpudata
->vid
.min
= int_tofp((value
>> 8) & 0x7f);
1234 cpudata
->vid
.max
= int_tofp((value
>> 16) & 0x7f);
1235 cpudata
->vid
.ratio
= div_fp(
1236 cpudata
->vid
.max
- cpudata
->vid
.min
,
1237 int_tofp(cpudata
->pstate
.max_pstate
-
1238 cpudata
->pstate
.min_pstate
));
1240 rdmsrl(MSR_ATOM_CORE_TURBO_VIDS
, value
);
1241 cpudata
->vid
.turbo
= value
& 0x7f;
1244 static int core_get_min_pstate(void)
1248 rdmsrl(MSR_PLATFORM_INFO
, value
);
1249 return (value
>> 40) & 0xFF;
1252 static int core_get_max_pstate_physical(void)
1256 rdmsrl(MSR_PLATFORM_INFO
, value
);
1257 return (value
>> 8) & 0xFF;
1260 static int core_get_tdp_ratio(u64 plat_info
)
1262 /* Check how many TDP levels present */
1263 if (plat_info
& 0x600000000) {
1269 /* Get the TDP level (0, 1, 2) to get ratios */
1270 err
= rdmsrl_safe(MSR_CONFIG_TDP_CONTROL
, &tdp_ctrl
);
1274 /* TDP MSR are continuous starting at 0x648 */
1275 tdp_msr
= MSR_CONFIG_TDP_NOMINAL
+ (tdp_ctrl
& 0x03);
1276 err
= rdmsrl_safe(tdp_msr
, &tdp_ratio
);
1280 /* For level 1 and 2, bits[23:16] contain the ratio */
1281 if (tdp_ctrl
& 0x03)
1284 tdp_ratio
&= 0xff; /* ratios are only 8 bits long */
1285 pr_debug("tdp_ratio %x\n", (int)tdp_ratio
);
1287 return (int)tdp_ratio
;
1293 static int core_get_max_pstate(void)
1301 rdmsrl(MSR_PLATFORM_INFO
, plat_info
);
1302 max_pstate
= (plat_info
>> 8) & 0xFF;
1304 tdp_ratio
= core_get_tdp_ratio(plat_info
);
1309 /* Turbo activation ratio is not used on HWP platforms */
1313 err
= rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO
, &tar
);
1317 /* Do some sanity checking for safety */
1318 tar_levels
= tar
& 0xff;
1319 if (tdp_ratio
- 1 == tar_levels
) {
1320 max_pstate
= tar_levels
;
1321 pr_debug("max_pstate=TAC %x\n", max_pstate
);
1328 static int core_get_turbo_pstate(void)
1333 rdmsrl(MSR_TURBO_RATIO_LIMIT
, value
);
1334 nont
= core_get_max_pstate();
1335 ret
= (value
) & 255;
1341 static inline int core_get_scaling(void)
1346 static u64
core_get_val(struct cpudata
*cpudata
, int pstate
)
1350 val
= (u64
)pstate
<< 8;
1351 if (global
.no_turbo
&& !global
.turbo_disabled
)
1352 val
|= (u64
)1 << 32;
1357 static int knl_get_aperf_mperf_shift(void)
1362 static int knl_get_turbo_pstate(void)
1367 rdmsrl(MSR_TURBO_RATIO_LIMIT
, value
);
1368 nont
= core_get_max_pstate();
1369 ret
= (((value
) >> 8) & 0xFF);
1375 static int intel_pstate_get_base_pstate(struct cpudata
*cpu
)
1377 return global
.no_turbo
|| global
.turbo_disabled
?
1378 cpu
->pstate
.max_pstate
: cpu
->pstate
.turbo_pstate
;
1381 static void intel_pstate_set_pstate(struct cpudata
*cpu
, int pstate
)
1383 trace_cpu_frequency(pstate
* cpu
->pstate
.scaling
, cpu
->cpu
);
1384 cpu
->pstate
.current_pstate
= pstate
;
1386 * Generally, there is no guarantee that this code will always run on
1387 * the CPU being updated, so force the register update to run on the
1390 wrmsrl_on_cpu(cpu
->cpu
, MSR_IA32_PERF_CTL
,
1391 pstate_funcs
.get_val(cpu
, pstate
));
1394 static void intel_pstate_set_min_pstate(struct cpudata
*cpu
)
1396 intel_pstate_set_pstate(cpu
, cpu
->pstate
.min_pstate
);
1399 static void intel_pstate_max_within_limits(struct cpudata
*cpu
)
1403 update_turbo_state();
1404 pstate
= intel_pstate_get_base_pstate(cpu
);
1405 pstate
= max(cpu
->pstate
.min_pstate
, cpu
->max_perf_ratio
);
1406 intel_pstate_set_pstate(cpu
, pstate
);
1409 static void intel_pstate_get_cpu_pstates(struct cpudata
*cpu
)
1411 cpu
->pstate
.min_pstate
= pstate_funcs
.get_min();
1412 cpu
->pstate
.max_pstate
= pstate_funcs
.get_max();
1413 cpu
->pstate
.max_pstate_physical
= pstate_funcs
.get_max_physical();
1414 cpu
->pstate
.turbo_pstate
= pstate_funcs
.get_turbo();
1415 cpu
->pstate
.scaling
= pstate_funcs
.get_scaling();
1416 cpu
->pstate
.max_freq
= cpu
->pstate
.max_pstate
* cpu
->pstate
.scaling
;
1418 if (hwp_active
&& !hwp_mode_bdw
) {
1419 unsigned int phy_max
, current_max
;
1421 intel_pstate_get_hwp_max(cpu
->cpu
, &phy_max
, ¤t_max
);
1422 cpu
->pstate
.turbo_freq
= phy_max
* cpu
->pstate
.scaling
;
1424 cpu
->pstate
.turbo_freq
= cpu
->pstate
.turbo_pstate
* cpu
->pstate
.scaling
;
1427 if (pstate_funcs
.get_aperf_mperf_shift
)
1428 cpu
->aperf_mperf_shift
= pstate_funcs
.get_aperf_mperf_shift();
1430 if (pstate_funcs
.get_vid
)
1431 pstate_funcs
.get_vid(cpu
);
1433 intel_pstate_set_min_pstate(cpu
);
1437 * Long hold time will keep high perf limits for long time,
1438 * which negatively impacts perf/watt for some workloads,
1439 * like specpower. 3ms is based on experiements on some
1442 static int hwp_boost_hold_time_ns
= 3 * NSEC_PER_MSEC
;
1444 static inline void intel_pstate_hwp_boost_up(struct cpudata
*cpu
)
1446 u64 hwp_req
= READ_ONCE(cpu
->hwp_req_cached
);
1447 u32 max_limit
= (hwp_req
& 0xff00) >> 8;
1448 u32 min_limit
= (hwp_req
& 0xff);
1452 * Cases to consider (User changes via sysfs or boot time):
1453 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
1455 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
1456 * Should result in one level boost only for P0.
1457 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
1458 * Should result in two level boost:
1459 * (min + p1)/2 and P1.
1460 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
1461 * Should result in three level boost:
1462 * (min + p1)/2, P1 and P0.
1465 /* If max and min are equal or already at max, nothing to boost */
1466 if (max_limit
== min_limit
|| cpu
->hwp_boost_min
>= max_limit
)
1469 if (!cpu
->hwp_boost_min
)
1470 cpu
->hwp_boost_min
= min_limit
;
1472 /* level at half way mark between min and guranteed */
1473 boost_level1
= (HWP_GUARANTEED_PERF(cpu
->hwp_cap_cached
) + min_limit
) >> 1;
1475 if (cpu
->hwp_boost_min
< boost_level1
)
1476 cpu
->hwp_boost_min
= boost_level1
;
1477 else if (cpu
->hwp_boost_min
< HWP_GUARANTEED_PERF(cpu
->hwp_cap_cached
))
1478 cpu
->hwp_boost_min
= HWP_GUARANTEED_PERF(cpu
->hwp_cap_cached
);
1479 else if (cpu
->hwp_boost_min
== HWP_GUARANTEED_PERF(cpu
->hwp_cap_cached
) &&
1480 max_limit
!= HWP_GUARANTEED_PERF(cpu
->hwp_cap_cached
))
1481 cpu
->hwp_boost_min
= max_limit
;
1485 hwp_req
= (hwp_req
& ~GENMASK_ULL(7, 0)) | cpu
->hwp_boost_min
;
1486 wrmsrl(MSR_HWP_REQUEST
, hwp_req
);
1487 cpu
->last_update
= cpu
->sample
.time
;
1490 static inline void intel_pstate_hwp_boost_down(struct cpudata
*cpu
)
1492 if (cpu
->hwp_boost_min
) {
1495 /* Check if we are idle for hold time to boost down */
1496 expired
= time_after64(cpu
->sample
.time
, cpu
->last_update
+
1497 hwp_boost_hold_time_ns
);
1499 wrmsrl(MSR_HWP_REQUEST
, cpu
->hwp_req_cached
);
1500 cpu
->hwp_boost_min
= 0;
1503 cpu
->last_update
= cpu
->sample
.time
;
1506 static inline void intel_pstate_update_util_hwp_local(struct cpudata
*cpu
,
1509 cpu
->sample
.time
= time
;
1511 if (cpu
->sched_flags
& SCHED_CPUFREQ_IOWAIT
) {
1514 cpu
->sched_flags
= 0;
1516 * Set iowait_boost flag and update time. Since IO WAIT flag
1517 * is set all the time, we can't just conclude that there is
1518 * some IO bound activity is scheduled on this CPU with just
1519 * one occurrence. If we receive at least two in two
1520 * consecutive ticks, then we treat as boost candidate.
1522 if (time_before64(time
, cpu
->last_io_update
+ 2 * TICK_NSEC
))
1525 cpu
->last_io_update
= time
;
1528 intel_pstate_hwp_boost_up(cpu
);
1531 intel_pstate_hwp_boost_down(cpu
);
1535 static inline void intel_pstate_update_util_hwp(struct update_util_data
*data
,
1536 u64 time
, unsigned int flags
)
1538 struct cpudata
*cpu
= container_of(data
, struct cpudata
, update_util
);
1540 cpu
->sched_flags
|= flags
;
1542 if (smp_processor_id() == cpu
->cpu
)
1543 intel_pstate_update_util_hwp_local(cpu
, time
);
1546 static inline void intel_pstate_calc_avg_perf(struct cpudata
*cpu
)
1548 struct sample
*sample
= &cpu
->sample
;
1550 sample
->core_avg_perf
= div_ext_fp(sample
->aperf
, sample
->mperf
);
1553 static inline bool intel_pstate_sample(struct cpudata
*cpu
, u64 time
)
1556 unsigned long flags
;
1559 local_irq_save(flags
);
1560 rdmsrl(MSR_IA32_APERF
, aperf
);
1561 rdmsrl(MSR_IA32_MPERF
, mperf
);
1563 if (cpu
->prev_mperf
== mperf
|| cpu
->prev_tsc
== tsc
) {
1564 local_irq_restore(flags
);
1567 local_irq_restore(flags
);
1569 cpu
->last_sample_time
= cpu
->sample
.time
;
1570 cpu
->sample
.time
= time
;
1571 cpu
->sample
.aperf
= aperf
;
1572 cpu
->sample
.mperf
= mperf
;
1573 cpu
->sample
.tsc
= tsc
;
1574 cpu
->sample
.aperf
-= cpu
->prev_aperf
;
1575 cpu
->sample
.mperf
-= cpu
->prev_mperf
;
1576 cpu
->sample
.tsc
-= cpu
->prev_tsc
;
1578 cpu
->prev_aperf
= aperf
;
1579 cpu
->prev_mperf
= mperf
;
1580 cpu
->prev_tsc
= tsc
;
1582 * First time this function is invoked in a given cycle, all of the
1583 * previous sample data fields are equal to zero or stale and they must
1584 * be populated with meaningful numbers for things to work, so assume
1585 * that sample.time will always be reset before setting the utilization
1586 * update hook and make the caller skip the sample then.
1588 if (cpu
->last_sample_time
) {
1589 intel_pstate_calc_avg_perf(cpu
);
1595 static inline int32_t get_avg_frequency(struct cpudata
*cpu
)
1597 return mul_ext_fp(cpu
->sample
.core_avg_perf
, cpu_khz
);
1600 static inline int32_t get_avg_pstate(struct cpudata
*cpu
)
1602 return mul_ext_fp(cpu
->pstate
.max_pstate_physical
,
1603 cpu
->sample
.core_avg_perf
);
1606 static inline int32_t get_target_pstate(struct cpudata
*cpu
)
1608 struct sample
*sample
= &cpu
->sample
;
1609 int32_t busy_frac
, boost
;
1610 int target
, avg_pstate
;
1612 busy_frac
= div_fp(sample
->mperf
<< cpu
->aperf_mperf_shift
,
1615 boost
= cpu
->iowait_boost
;
1616 cpu
->iowait_boost
>>= 1;
1618 if (busy_frac
< boost
)
1621 sample
->busy_scaled
= busy_frac
* 100;
1623 target
= global
.no_turbo
|| global
.turbo_disabled
?
1624 cpu
->pstate
.max_pstate
: cpu
->pstate
.turbo_pstate
;
1625 target
+= target
>> 2;
1626 target
= mul_fp(target
, busy_frac
);
1627 if (target
< cpu
->pstate
.min_pstate
)
1628 target
= cpu
->pstate
.min_pstate
;
1631 * If the average P-state during the previous cycle was higher than the
1632 * current target, add 50% of the difference to the target to reduce
1633 * possible performance oscillations and offset possible performance
1634 * loss related to moving the workload from one CPU to another within
1637 avg_pstate
= get_avg_pstate(cpu
);
1638 if (avg_pstate
> target
)
1639 target
+= (avg_pstate
- target
) >> 1;
1644 static int intel_pstate_prepare_request(struct cpudata
*cpu
, int pstate
)
1646 int max_pstate
= intel_pstate_get_base_pstate(cpu
);
1649 min_pstate
= max(cpu
->pstate
.min_pstate
, cpu
->min_perf_ratio
);
1650 max_pstate
= max(min_pstate
, cpu
->max_perf_ratio
);
1651 return clamp_t(int, pstate
, min_pstate
, max_pstate
);
1654 static void intel_pstate_update_pstate(struct cpudata
*cpu
, int pstate
)
1656 if (pstate
== cpu
->pstate
.current_pstate
)
1659 cpu
->pstate
.current_pstate
= pstate
;
1660 wrmsrl(MSR_IA32_PERF_CTL
, pstate_funcs
.get_val(cpu
, pstate
));
1663 static void intel_pstate_adjust_pstate(struct cpudata
*cpu
)
1665 int from
= cpu
->pstate
.current_pstate
;
1666 struct sample
*sample
;
1669 update_turbo_state();
1671 target_pstate
= get_target_pstate(cpu
);
1672 target_pstate
= intel_pstate_prepare_request(cpu
, target_pstate
);
1673 trace_cpu_frequency(target_pstate
* cpu
->pstate
.scaling
, cpu
->cpu
);
1674 intel_pstate_update_pstate(cpu
, target_pstate
);
1676 sample
= &cpu
->sample
;
1677 trace_pstate_sample(mul_ext_fp(100, sample
->core_avg_perf
),
1678 fp_toint(sample
->busy_scaled
),
1680 cpu
->pstate
.current_pstate
,
1684 get_avg_frequency(cpu
),
1685 fp_toint(cpu
->iowait_boost
* 100));
1688 static void intel_pstate_update_util(struct update_util_data
*data
, u64 time
,
1691 struct cpudata
*cpu
= container_of(data
, struct cpudata
, update_util
);
1694 /* Don't allow remote callbacks */
1695 if (smp_processor_id() != cpu
->cpu
)
1698 if (flags
& SCHED_CPUFREQ_IOWAIT
) {
1699 cpu
->iowait_boost
= int_tofp(1);
1700 cpu
->last_update
= time
;
1702 * The last time the busy was 100% so P-state was max anyway
1703 * so avoid overhead of computation.
1705 if (fp_toint(cpu
->sample
.busy_scaled
) == 100)
1709 } else if (cpu
->iowait_boost
) {
1710 /* Clear iowait_boost if the CPU may have been idle. */
1711 delta_ns
= time
- cpu
->last_update
;
1712 if (delta_ns
> TICK_NSEC
)
1713 cpu
->iowait_boost
= 0;
1715 cpu
->last_update
= time
;
1716 delta_ns
= time
- cpu
->sample
.time
;
1717 if ((s64
)delta_ns
< INTEL_PSTATE_SAMPLING_INTERVAL
)
1721 if (intel_pstate_sample(cpu
, time
))
1722 intel_pstate_adjust_pstate(cpu
);
1725 static struct pstate_funcs core_funcs
= {
1726 .get_max
= core_get_max_pstate
,
1727 .get_max_physical
= core_get_max_pstate_physical
,
1728 .get_min
= core_get_min_pstate
,
1729 .get_turbo
= core_get_turbo_pstate
,
1730 .get_scaling
= core_get_scaling
,
1731 .get_val
= core_get_val
,
1734 static const struct pstate_funcs silvermont_funcs
= {
1735 .get_max
= atom_get_max_pstate
,
1736 .get_max_physical
= atom_get_max_pstate
,
1737 .get_min
= atom_get_min_pstate
,
1738 .get_turbo
= atom_get_turbo_pstate
,
1739 .get_val
= atom_get_val
,
1740 .get_scaling
= silvermont_get_scaling
,
1741 .get_vid
= atom_get_vid
,
1744 static const struct pstate_funcs airmont_funcs
= {
1745 .get_max
= atom_get_max_pstate
,
1746 .get_max_physical
= atom_get_max_pstate
,
1747 .get_min
= atom_get_min_pstate
,
1748 .get_turbo
= atom_get_turbo_pstate
,
1749 .get_val
= atom_get_val
,
1750 .get_scaling
= airmont_get_scaling
,
1751 .get_vid
= atom_get_vid
,
1754 static const struct pstate_funcs knl_funcs
= {
1755 .get_max
= core_get_max_pstate
,
1756 .get_max_physical
= core_get_max_pstate_physical
,
1757 .get_min
= core_get_min_pstate
,
1758 .get_turbo
= knl_get_turbo_pstate
,
1759 .get_aperf_mperf_shift
= knl_get_aperf_mperf_shift
,
1760 .get_scaling
= core_get_scaling
,
1761 .get_val
= core_get_val
,
1764 #define ICPU(model, policy) \
1765 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1766 (unsigned long)&policy }
1768 static const struct x86_cpu_id intel_pstate_cpu_ids
[] = {
1769 ICPU(INTEL_FAM6_SANDYBRIDGE
, core_funcs
),
1770 ICPU(INTEL_FAM6_SANDYBRIDGE_X
, core_funcs
),
1771 ICPU(INTEL_FAM6_ATOM_SILVERMONT1
, silvermont_funcs
),
1772 ICPU(INTEL_FAM6_IVYBRIDGE
, core_funcs
),
1773 ICPU(INTEL_FAM6_HASWELL_CORE
, core_funcs
),
1774 ICPU(INTEL_FAM6_BROADWELL_CORE
, core_funcs
),
1775 ICPU(INTEL_FAM6_IVYBRIDGE_X
, core_funcs
),
1776 ICPU(INTEL_FAM6_HASWELL_X
, core_funcs
),
1777 ICPU(INTEL_FAM6_HASWELL_ULT
, core_funcs
),
1778 ICPU(INTEL_FAM6_HASWELL_GT3E
, core_funcs
),
1779 ICPU(INTEL_FAM6_BROADWELL_GT3E
, core_funcs
),
1780 ICPU(INTEL_FAM6_ATOM_AIRMONT
, airmont_funcs
),
1781 ICPU(INTEL_FAM6_SKYLAKE_MOBILE
, core_funcs
),
1782 ICPU(INTEL_FAM6_BROADWELL_X
, core_funcs
),
1783 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP
, core_funcs
),
1784 ICPU(INTEL_FAM6_BROADWELL_XEON_D
, core_funcs
),
1785 ICPU(INTEL_FAM6_XEON_PHI_KNL
, knl_funcs
),
1786 ICPU(INTEL_FAM6_XEON_PHI_KNM
, knl_funcs
),
1787 ICPU(INTEL_FAM6_ATOM_GOLDMONT
, core_funcs
),
1788 ICPU(INTEL_FAM6_ATOM_GEMINI_LAKE
, core_funcs
),
1789 ICPU(INTEL_FAM6_SKYLAKE_X
, core_funcs
),
1792 MODULE_DEVICE_TABLE(x86cpu
, intel_pstate_cpu_ids
);
1794 static const struct x86_cpu_id intel_pstate_cpu_oob_ids
[] __initconst
= {
1795 ICPU(INTEL_FAM6_BROADWELL_XEON_D
, core_funcs
),
1796 ICPU(INTEL_FAM6_BROADWELL_X
, core_funcs
),
1797 ICPU(INTEL_FAM6_SKYLAKE_X
, core_funcs
),
1801 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids
[] = {
1802 ICPU(INTEL_FAM6_KABYLAKE_DESKTOP
, core_funcs
),
1806 static const struct x86_cpu_id intel_pstate_hwp_boost_ids
[] = {
1807 ICPU(INTEL_FAM6_SKYLAKE_X
, core_funcs
),
1808 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP
, core_funcs
),
1812 static int intel_pstate_init_cpu(unsigned int cpunum
)
1814 struct cpudata
*cpu
;
1816 cpu
= all_cpu_data
[cpunum
];
1819 cpu
= kzalloc(sizeof(*cpu
), GFP_KERNEL
);
1823 all_cpu_data
[cpunum
] = cpu
;
1825 cpu
->epp_default
= -EINVAL
;
1826 cpu
->epp_powersave
= -EINVAL
;
1827 cpu
->epp_saved
= -EINVAL
;
1830 cpu
= all_cpu_data
[cpunum
];
1835 const struct x86_cpu_id
*id
;
1837 id
= x86_match_cpu(intel_pstate_cpu_ee_disable_ids
);
1839 intel_pstate_disable_ee(cpunum
);
1841 intel_pstate_hwp_enable(cpu
);
1843 id
= x86_match_cpu(intel_pstate_hwp_boost_ids
);
1848 intel_pstate_get_cpu_pstates(cpu
);
1850 pr_debug("controlling: cpu %d\n", cpunum
);
1855 static void intel_pstate_set_update_util_hook(unsigned int cpu_num
)
1857 struct cpudata
*cpu
= all_cpu_data
[cpu_num
];
1859 if (hwp_active
&& !hwp_boost
)
1862 if (cpu
->update_util_set
)
1865 /* Prevent intel_pstate_update_util() from using stale data. */
1866 cpu
->sample
.time
= 0;
1867 cpufreq_add_update_util_hook(cpu_num
, &cpu
->update_util
,
1869 intel_pstate_update_util_hwp
:
1870 intel_pstate_update_util
));
1871 cpu
->update_util_set
= true;
1874 static void intel_pstate_clear_update_util_hook(unsigned int cpu
)
1876 struct cpudata
*cpu_data
= all_cpu_data
[cpu
];
1878 if (!cpu_data
->update_util_set
)
1881 cpufreq_remove_update_util_hook(cpu
);
1882 cpu_data
->update_util_set
= false;
1883 synchronize_sched();
1886 static int intel_pstate_get_max_freq(struct cpudata
*cpu
)
1888 return global
.turbo_disabled
|| global
.no_turbo
?
1889 cpu
->pstate
.max_freq
: cpu
->pstate
.turbo_freq
;
1892 static void intel_pstate_update_perf_limits(struct cpufreq_policy
*policy
,
1893 struct cpudata
*cpu
)
1895 int max_freq
= intel_pstate_get_max_freq(cpu
);
1896 int32_t max_policy_perf
, min_policy_perf
;
1897 int max_state
, turbo_max
;
1900 * HWP needs some special consideration, because on BDX the
1901 * HWP_REQUEST uses abstract value to represent performance
1902 * rather than pure ratios.
1905 intel_pstate_get_hwp_max(cpu
->cpu
, &turbo_max
, &max_state
);
1907 max_state
= intel_pstate_get_base_pstate(cpu
);
1908 turbo_max
= cpu
->pstate
.turbo_pstate
;
1911 max_policy_perf
= max_state
* policy
->max
/ max_freq
;
1912 if (policy
->max
== policy
->min
) {
1913 min_policy_perf
= max_policy_perf
;
1915 min_policy_perf
= max_state
* policy
->min
/ max_freq
;
1916 min_policy_perf
= clamp_t(int32_t, min_policy_perf
,
1917 0, max_policy_perf
);
1920 pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n",
1921 policy
->cpu
, max_state
,
1922 min_policy_perf
, max_policy_perf
);
1924 /* Normalize user input to [min_perf, max_perf] */
1925 if (per_cpu_limits
) {
1926 cpu
->min_perf_ratio
= min_policy_perf
;
1927 cpu
->max_perf_ratio
= max_policy_perf
;
1929 int32_t global_min
, global_max
;
1931 /* Global limits are in percent of the maximum turbo P-state. */
1932 global_max
= DIV_ROUND_UP(turbo_max
* global
.max_perf_pct
, 100);
1933 global_min
= DIV_ROUND_UP(turbo_max
* global
.min_perf_pct
, 100);
1934 global_min
= clamp_t(int32_t, global_min
, 0, global_max
);
1936 pr_debug("cpu:%d global_min:%d global_max:%d\n", policy
->cpu
,
1937 global_min
, global_max
);
1939 cpu
->min_perf_ratio
= max(min_policy_perf
, global_min
);
1940 cpu
->min_perf_ratio
= min(cpu
->min_perf_ratio
, max_policy_perf
);
1941 cpu
->max_perf_ratio
= min(max_policy_perf
, global_max
);
1942 cpu
->max_perf_ratio
= max(min_policy_perf
, cpu
->max_perf_ratio
);
1944 /* Make sure min_perf <= max_perf */
1945 cpu
->min_perf_ratio
= min(cpu
->min_perf_ratio
,
1946 cpu
->max_perf_ratio
);
1949 pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", policy
->cpu
,
1950 cpu
->max_perf_ratio
,
1951 cpu
->min_perf_ratio
);
1954 static int intel_pstate_set_policy(struct cpufreq_policy
*policy
)
1956 struct cpudata
*cpu
;
1958 if (!policy
->cpuinfo
.max_freq
)
1961 pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
1962 policy
->cpuinfo
.max_freq
, policy
->max
);
1964 cpu
= all_cpu_data
[policy
->cpu
];
1965 cpu
->policy
= policy
->policy
;
1967 mutex_lock(&intel_pstate_limits_lock
);
1969 intel_pstate_update_perf_limits(policy
, cpu
);
1971 if (cpu
->policy
== CPUFREQ_POLICY_PERFORMANCE
) {
1973 * NOHZ_FULL CPUs need this as the governor callback may not
1974 * be invoked on them.
1976 intel_pstate_clear_update_util_hook(policy
->cpu
);
1977 intel_pstate_max_within_limits(cpu
);
1979 intel_pstate_set_update_util_hook(policy
->cpu
);
1984 * When hwp_boost was active before and dynamically it
1985 * was turned off, in that case we need to clear the
1989 intel_pstate_clear_update_util_hook(policy
->cpu
);
1990 intel_pstate_hwp_set(policy
->cpu
);
1993 mutex_unlock(&intel_pstate_limits_lock
);
1998 static void intel_pstate_adjust_policy_max(struct cpufreq_policy
*policy
,
1999 struct cpudata
*cpu
)
2001 if (cpu
->pstate
.max_pstate_physical
> cpu
->pstate
.max_pstate
&&
2002 policy
->max
< policy
->cpuinfo
.max_freq
&&
2003 policy
->max
> cpu
->pstate
.max_freq
) {
2004 pr_debug("policy->max > max non turbo frequency\n");
2005 policy
->max
= policy
->cpuinfo
.max_freq
;
2009 static int intel_pstate_verify_policy(struct cpufreq_policy
*policy
)
2011 struct cpudata
*cpu
= all_cpu_data
[policy
->cpu
];
2013 update_turbo_state();
2014 cpufreq_verify_within_limits(policy
, policy
->cpuinfo
.min_freq
,
2015 intel_pstate_get_max_freq(cpu
));
2017 if (policy
->policy
!= CPUFREQ_POLICY_POWERSAVE
&&
2018 policy
->policy
!= CPUFREQ_POLICY_PERFORMANCE
)
2021 intel_pstate_adjust_policy_max(policy
, cpu
);
2026 static void intel_cpufreq_stop_cpu(struct cpufreq_policy
*policy
)
2028 intel_pstate_set_min_pstate(all_cpu_data
[policy
->cpu
]);
2031 static void intel_pstate_stop_cpu(struct cpufreq_policy
*policy
)
2033 pr_debug("CPU %d exiting\n", policy
->cpu
);
2035 intel_pstate_clear_update_util_hook(policy
->cpu
);
2037 intel_pstate_hwp_save_state(policy
);
2039 intel_cpufreq_stop_cpu(policy
);
2042 static int intel_pstate_cpu_exit(struct cpufreq_policy
*policy
)
2044 intel_pstate_exit_perf_limits(policy
);
2046 policy
->fast_switch_possible
= false;
2051 static int __intel_pstate_cpu_init(struct cpufreq_policy
*policy
)
2053 struct cpudata
*cpu
;
2056 rc
= intel_pstate_init_cpu(policy
->cpu
);
2060 cpu
= all_cpu_data
[policy
->cpu
];
2062 cpu
->max_perf_ratio
= 0xFF;
2063 cpu
->min_perf_ratio
= 0;
2065 policy
->min
= cpu
->pstate
.min_pstate
* cpu
->pstate
.scaling
;
2066 policy
->max
= cpu
->pstate
.turbo_pstate
* cpu
->pstate
.scaling
;
2068 /* cpuinfo and default policy values */
2069 policy
->cpuinfo
.min_freq
= cpu
->pstate
.min_pstate
* cpu
->pstate
.scaling
;
2070 update_turbo_state();
2071 policy
->cpuinfo
.max_freq
= global
.turbo_disabled
?
2072 cpu
->pstate
.max_pstate
: cpu
->pstate
.turbo_pstate
;
2073 policy
->cpuinfo
.max_freq
*= cpu
->pstate
.scaling
;
2075 intel_pstate_init_acpi_perf_limits(policy
);
2077 policy
->fast_switch_possible
= true;
2082 static int intel_pstate_cpu_init(struct cpufreq_policy
*policy
)
2084 int ret
= __intel_pstate_cpu_init(policy
);
2089 if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
))
2090 policy
->policy
= CPUFREQ_POLICY_PERFORMANCE
;
2092 policy
->policy
= CPUFREQ_POLICY_POWERSAVE
;
2097 static struct cpufreq_driver intel_pstate
= {
2098 .flags
= CPUFREQ_CONST_LOOPS
,
2099 .verify
= intel_pstate_verify_policy
,
2100 .setpolicy
= intel_pstate_set_policy
,
2101 .suspend
= intel_pstate_hwp_save_state
,
2102 .resume
= intel_pstate_resume
,
2103 .init
= intel_pstate_cpu_init
,
2104 .exit
= intel_pstate_cpu_exit
,
2105 .stop_cpu
= intel_pstate_stop_cpu
,
2106 .name
= "intel_pstate",
2109 static int intel_cpufreq_verify_policy(struct cpufreq_policy
*policy
)
2111 struct cpudata
*cpu
= all_cpu_data
[policy
->cpu
];
2113 update_turbo_state();
2114 cpufreq_verify_within_limits(policy
, policy
->cpuinfo
.min_freq
,
2115 intel_pstate_get_max_freq(cpu
));
2117 intel_pstate_adjust_policy_max(policy
, cpu
);
2119 intel_pstate_update_perf_limits(policy
, cpu
);
2124 /* Use of trace in passive mode:
2126 * In passive mode the trace core_busy field (also known as the
2127 * performance field, and lablelled as such on the graphs; also known as
2128 * core_avg_perf) is not needed and so is re-assigned to indicate if the
2129 * driver call was via the normal or fast switch path. Various graphs
2130 * output from the intel_pstate_tracer.py utility that include core_busy
2131 * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2132 * so we use 10 to indicate the the normal path through the driver, and
2133 * 90 to indicate the fast switch path through the driver.
2134 * The scaled_busy field is not used, and is set to 0.
2137 #define INTEL_PSTATE_TRACE_TARGET 10
2138 #define INTEL_PSTATE_TRACE_FAST_SWITCH 90
2140 static void intel_cpufreq_trace(struct cpudata
*cpu
, unsigned int trace_type
, int old_pstate
)
2142 struct sample
*sample
;
2144 if (!trace_pstate_sample_enabled())
2147 if (!intel_pstate_sample(cpu
, ktime_get()))
2150 sample
= &cpu
->sample
;
2151 trace_pstate_sample(trace_type
,
2154 cpu
->pstate
.current_pstate
,
2158 get_avg_frequency(cpu
),
2159 fp_toint(cpu
->iowait_boost
* 100));
2162 static int intel_cpufreq_target(struct cpufreq_policy
*policy
,
2163 unsigned int target_freq
,
2164 unsigned int relation
)
2166 struct cpudata
*cpu
= all_cpu_data
[policy
->cpu
];
2167 struct cpufreq_freqs freqs
;
2168 int target_pstate
, old_pstate
;
2170 update_turbo_state();
2172 freqs
.old
= policy
->cur
;
2173 freqs
.new = target_freq
;
2175 cpufreq_freq_transition_begin(policy
, &freqs
);
2177 case CPUFREQ_RELATION_L
:
2178 target_pstate
= DIV_ROUND_UP(freqs
.new, cpu
->pstate
.scaling
);
2180 case CPUFREQ_RELATION_H
:
2181 target_pstate
= freqs
.new / cpu
->pstate
.scaling
;
2184 target_pstate
= DIV_ROUND_CLOSEST(freqs
.new, cpu
->pstate
.scaling
);
2187 target_pstate
= intel_pstate_prepare_request(cpu
, target_pstate
);
2188 old_pstate
= cpu
->pstate
.current_pstate
;
2189 if (target_pstate
!= cpu
->pstate
.current_pstate
) {
2190 cpu
->pstate
.current_pstate
= target_pstate
;
2191 wrmsrl_on_cpu(policy
->cpu
, MSR_IA32_PERF_CTL
,
2192 pstate_funcs
.get_val(cpu
, target_pstate
));
2194 freqs
.new = target_pstate
* cpu
->pstate
.scaling
;
2195 intel_cpufreq_trace(cpu
, INTEL_PSTATE_TRACE_TARGET
, old_pstate
);
2196 cpufreq_freq_transition_end(policy
, &freqs
, false);
2201 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy
*policy
,
2202 unsigned int target_freq
)
2204 struct cpudata
*cpu
= all_cpu_data
[policy
->cpu
];
2205 int target_pstate
, old_pstate
;
2207 update_turbo_state();
2209 target_pstate
= DIV_ROUND_UP(target_freq
, cpu
->pstate
.scaling
);
2210 target_pstate
= intel_pstate_prepare_request(cpu
, target_pstate
);
2211 old_pstate
= cpu
->pstate
.current_pstate
;
2212 intel_pstate_update_pstate(cpu
, target_pstate
);
2213 intel_cpufreq_trace(cpu
, INTEL_PSTATE_TRACE_FAST_SWITCH
, old_pstate
);
2214 return target_pstate
* cpu
->pstate
.scaling
;
2217 static int intel_cpufreq_cpu_init(struct cpufreq_policy
*policy
)
2219 int ret
= __intel_pstate_cpu_init(policy
);
2224 policy
->cpuinfo
.transition_latency
= INTEL_CPUFREQ_TRANSITION_LATENCY
;
2225 policy
->transition_delay_us
= INTEL_CPUFREQ_TRANSITION_DELAY
;
2226 /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2227 policy
->cur
= policy
->cpuinfo
.min_freq
;
2232 static struct cpufreq_driver intel_cpufreq
= {
2233 .flags
= CPUFREQ_CONST_LOOPS
,
2234 .verify
= intel_cpufreq_verify_policy
,
2235 .target
= intel_cpufreq_target
,
2236 .fast_switch
= intel_cpufreq_fast_switch
,
2237 .init
= intel_cpufreq_cpu_init
,
2238 .exit
= intel_pstate_cpu_exit
,
2239 .stop_cpu
= intel_cpufreq_stop_cpu
,
2240 .name
= "intel_cpufreq",
2243 static struct cpufreq_driver
*default_driver
= &intel_pstate
;
2245 static void intel_pstate_driver_cleanup(void)
2250 for_each_online_cpu(cpu
) {
2251 if (all_cpu_data
[cpu
]) {
2252 if (intel_pstate_driver
== &intel_pstate
)
2253 intel_pstate_clear_update_util_hook(cpu
);
2255 kfree(all_cpu_data
[cpu
]);
2256 all_cpu_data
[cpu
] = NULL
;
2260 intel_pstate_driver
= NULL
;
2263 static int intel_pstate_register_driver(struct cpufreq_driver
*driver
)
2267 memset(&global
, 0, sizeof(global
));
2268 global
.max_perf_pct
= 100;
2270 intel_pstate_driver
= driver
;
2271 ret
= cpufreq_register_driver(intel_pstate_driver
);
2273 intel_pstate_driver_cleanup();
2277 global
.min_perf_pct
= min_perf_pct_min();
2282 static int intel_pstate_unregister_driver(void)
2287 cpufreq_unregister_driver(intel_pstate_driver
);
2288 intel_pstate_driver_cleanup();
2293 static ssize_t
intel_pstate_show_status(char *buf
)
2295 if (!intel_pstate_driver
)
2296 return sprintf(buf
, "off\n");
2298 return sprintf(buf
, "%s\n", intel_pstate_driver
== &intel_pstate
?
2299 "active" : "passive");
2302 static int intel_pstate_update_status(const char *buf
, size_t size
)
2306 if (size
== 3 && !strncmp(buf
, "off", size
))
2307 return intel_pstate_driver
?
2308 intel_pstate_unregister_driver() : -EINVAL
;
2310 if (size
== 6 && !strncmp(buf
, "active", size
)) {
2311 if (intel_pstate_driver
) {
2312 if (intel_pstate_driver
== &intel_pstate
)
2315 ret
= intel_pstate_unregister_driver();
2320 return intel_pstate_register_driver(&intel_pstate
);
2323 if (size
== 7 && !strncmp(buf
, "passive", size
)) {
2324 if (intel_pstate_driver
) {
2325 if (intel_pstate_driver
== &intel_cpufreq
)
2328 ret
= intel_pstate_unregister_driver();
2333 return intel_pstate_register_driver(&intel_cpufreq
);
2339 static int no_load __initdata
;
2340 static int no_hwp __initdata
;
2341 static int hwp_only __initdata
;
2342 static unsigned int force_load __initdata
;
2344 static int __init
intel_pstate_msrs_not_valid(void)
2346 if (!pstate_funcs
.get_max() ||
2347 !pstate_funcs
.get_min() ||
2348 !pstate_funcs
.get_turbo())
2354 static void __init
copy_cpu_funcs(struct pstate_funcs
*funcs
)
2356 pstate_funcs
.get_max
= funcs
->get_max
;
2357 pstate_funcs
.get_max_physical
= funcs
->get_max_physical
;
2358 pstate_funcs
.get_min
= funcs
->get_min
;
2359 pstate_funcs
.get_turbo
= funcs
->get_turbo
;
2360 pstate_funcs
.get_scaling
= funcs
->get_scaling
;
2361 pstate_funcs
.get_val
= funcs
->get_val
;
2362 pstate_funcs
.get_vid
= funcs
->get_vid
;
2363 pstate_funcs
.get_aperf_mperf_shift
= funcs
->get_aperf_mperf_shift
;
2368 static bool __init
intel_pstate_no_acpi_pss(void)
2372 for_each_possible_cpu(i
) {
2374 union acpi_object
*pss
;
2375 struct acpi_buffer buffer
= { ACPI_ALLOCATE_BUFFER
, NULL
};
2376 struct acpi_processor
*pr
= per_cpu(processors
, i
);
2381 status
= acpi_evaluate_object(pr
->handle
, "_PSS", NULL
, &buffer
);
2382 if (ACPI_FAILURE(status
))
2385 pss
= buffer
.pointer
;
2386 if (pss
&& pss
->type
== ACPI_TYPE_PACKAGE
) {
2397 static bool __init
intel_pstate_no_acpi_pcch(void)
2402 status
= acpi_get_handle(NULL
, "\\_SB", &handle
);
2403 if (ACPI_FAILURE(status
))
2406 return !acpi_has_method(handle
, "PCCH");
2409 static bool __init
intel_pstate_has_acpi_ppc(void)
2413 for_each_possible_cpu(i
) {
2414 struct acpi_processor
*pr
= per_cpu(processors
, i
);
2418 if (acpi_has_method(pr
->handle
, "_PPC"))
2429 /* Hardware vendor-specific info that has its own power management modes */
2430 static struct acpi_platform_list plat_info
[] __initdata
= {
2431 {"HP ", "ProLiant", 0, ACPI_SIG_FADT
, all_versions
, 0, PSS
},
2432 {"ORACLE", "X4-2 ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2433 {"ORACLE", "X4-2L ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2434 {"ORACLE", "X4-2B ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2435 {"ORACLE", "X3-2 ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2436 {"ORACLE", "X3-2L ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2437 {"ORACLE", "X3-2B ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2438 {"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2439 {"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2440 {"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2441 {"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2442 {"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2443 {"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2444 {"ORACLE", "X6-2 ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2445 {"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT
, all_versions
, 0, PPC
},
2449 static bool __init
intel_pstate_platform_pwr_mgmt_exists(void)
2451 const struct x86_cpu_id
*id
;
2455 id
= x86_match_cpu(intel_pstate_cpu_oob_ids
);
2457 rdmsrl(MSR_MISC_PWR_MGMT
, misc_pwr
);
2458 if ( misc_pwr
& (1 << 8))
2462 idx
= acpi_match_platform_list(plat_info
);
2466 switch (plat_info
[idx
].data
) {
2468 if (!intel_pstate_no_acpi_pss())
2471 return intel_pstate_no_acpi_pcch();
2473 return intel_pstate_has_acpi_ppc() && !force_load
;
2479 static void intel_pstate_request_control_from_smm(void)
2482 * It may be unsafe to request P-states control from SMM if _PPC support
2483 * has not been enabled.
2486 acpi_processor_pstate_control();
2488 #else /* CONFIG_ACPI not enabled */
2489 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2490 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2491 static inline void intel_pstate_request_control_from_smm(void) {}
2492 #endif /* CONFIG_ACPI */
2494 #define INTEL_PSTATE_HWP_BROADWELL 0x01
2496 #define ICPU_HWP(model, hwp_mode) \
2497 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_HWP, hwp_mode }
2499 static const struct x86_cpu_id hwp_support_ids
[] __initconst
= {
2500 ICPU_HWP(INTEL_FAM6_BROADWELL_X
, INTEL_PSTATE_HWP_BROADWELL
),
2501 ICPU_HWP(INTEL_FAM6_BROADWELL_XEON_D
, INTEL_PSTATE_HWP_BROADWELL
),
2502 ICPU_HWP(X86_MODEL_ANY
, 0),
2506 static int __init
intel_pstate_init(void)
2508 const struct x86_cpu_id
*id
;
2514 id
= x86_match_cpu(hwp_support_ids
);
2516 copy_cpu_funcs(&core_funcs
);
2519 hwp_mode_bdw
= id
->driver_data
;
2520 intel_pstate
.attr
= hwp_cpufreq_attrs
;
2521 goto hwp_cpu_matched
;
2524 id
= x86_match_cpu(intel_pstate_cpu_ids
);
2528 copy_cpu_funcs((struct pstate_funcs
*)id
->driver_data
);
2531 if (intel_pstate_msrs_not_valid())
2536 * The Intel pstate driver will be ignored if the platform
2537 * firmware has its own power management modes.
2539 if (intel_pstate_platform_pwr_mgmt_exists())
2542 if (!hwp_active
&& hwp_only
)
2545 pr_info("Intel P-state driver initializing\n");
2547 all_cpu_data
= vzalloc(array_size(sizeof(void *), num_possible_cpus()));
2551 intel_pstate_request_control_from_smm();
2553 intel_pstate_sysfs_expose_params();
2555 mutex_lock(&intel_pstate_driver_lock
);
2556 rc
= intel_pstate_register_driver(default_driver
);
2557 mutex_unlock(&intel_pstate_driver_lock
);
2562 pr_info("HWP enabled\n");
2566 device_initcall(intel_pstate_init
);
2568 static int __init
intel_pstate_setup(char *str
)
2573 if (!strcmp(str
, "disable")) {
2575 } else if (!strcmp(str
, "passive")) {
2576 pr_info("Passive mode enabled\n");
2577 default_driver
= &intel_cpufreq
;
2580 if (!strcmp(str
, "no_hwp")) {
2581 pr_info("HWP disabled\n");
2584 if (!strcmp(str
, "force"))
2586 if (!strcmp(str
, "hwp_only"))
2588 if (!strcmp(str
, "per_cpu_perf_limits"))
2589 per_cpu_limits
= true;
2592 if (!strcmp(str
, "support_acpi_ppc"))
2598 early_param("intel_pstate", intel_pstate_setup
);
2600 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2601 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2602 MODULE_LICENSE("GPL");