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 ATOM_RATIOS 0x66a
41 #define ATOM_VIDS 0x66b
42 #define ATOM_TURBO_RATIOS 0x66c
43 #define ATOM_TURBO_VIDS 0x66d
46 #include <acpi/processor.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)
56 static inline int32_t mul_fp(int32_t x, int32_t y)
58 return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
61 static inline int32_t div_fp(s64 x, s64 y)
63 return div64_s64((int64_t)x << FRAC_BITS, y);
66 static inline int ceiling_fp(int32_t x)
71 mask = (1 << FRAC_BITS) - 1;
77 static inline u64 mul_ext_fp(u64 x, u64 y)
79 return (x * y) >> EXT_FRAC_BITS;
82 static inline u64 div_ext_fp(u64 x, u64 y)
84 return div64_u64(x << EXT_FRAC_BITS, y);
88 * struct sample - Store performance sample
89 * @core_avg_perf: Ratio of APERF/MPERF which is the actual average
90 * performance during last sample period
91 * @busy_scaled: Scaled busy value which is used to calculate next
92 * P state. This can be different than core_avg_perf
93 * to account for cpu idle period
94 * @aperf: Difference of actual performance frequency clock count
95 * read from APERF MSR between last and current sample
96 * @mperf: Difference of maximum performance frequency clock count
97 * read from MPERF MSR between last and current sample
98 * @tsc: Difference of time stamp counter between last and
100 * @freq: Effective frequency calculated from APERF/MPERF
101 * @time: Current time from scheduler
103 * This structure is used in the cpudata structure to store performance sample
104 * data for choosing next P State.
107 int32_t core_avg_perf;
117 * struct pstate_data - Store P state data
118 * @current_pstate: Current requested P state
119 * @min_pstate: Min P state possible for this platform
120 * @max_pstate: Max P state possible for this platform
121 * @max_pstate_physical:This is physical Max P state for a processor
122 * This can be higher than the max_pstate which can
123 * be limited by platform thermal design power limits
124 * @scaling: Scaling factor to convert frequency to cpufreq
126 * @turbo_pstate: Max Turbo P state possible for this platform
128 * Stores the per cpu model P state limits and current P state.
134 int max_pstate_physical;
140 * struct vid_data - Stores voltage information data
141 * @min: VID data for this platform corresponding to
143 * @max: VID data corresponding to the highest P State.
144 * @turbo: VID data for turbo P state
145 * @ratio: Ratio of (vid max - vid min) /
146 * (max P state - Min P State)
148 * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
149 * This data is used in Atom platforms, where in addition to target P state,
150 * the voltage data needs to be specified to select next P State.
160 * struct _pid - Stores PID data
161 * @setpoint: Target set point for busyness or performance
162 * @integral: Storage for accumulated error values
163 * @p_gain: PID proportional gain
164 * @i_gain: PID integral gain
165 * @d_gain: PID derivative gain
166 * @deadband: PID deadband
167 * @last_err: Last error storage for integral part of PID calculation
169 * Stores PID coefficients and last error for PID controller.
182 * struct cpudata - Per CPU instance data storage
183 * @cpu: CPU number for this instance data
184 * @update_util: CPUFreq utility callback information
185 * @update_util_set: CPUFreq utility callback is set
186 * @pstate: Stores P state limits for this CPU
187 * @vid: Stores VID limits for this CPU
188 * @pid: Stores PID parameters for this CPU
189 * @last_sample_time: Last Sample time
190 * @prev_aperf: Last APERF value read from APERF MSR
191 * @prev_mperf: Last MPERF value read from MPERF MSR
192 * @prev_tsc: Last timestamp counter (TSC) value
193 * @prev_cummulative_iowait: IO Wait time difference from last and
195 * @sample: Storage for storing last Sample data
196 * @acpi_perf_data: Stores ACPI perf information read from _PSS
197 * @valid_pss_table: Set to true for valid ACPI _PSS entries found
199 * This structure stores per CPU instance data for all CPUs.
204 struct update_util_data update_util;
205 bool update_util_set;
207 struct pstate_data pstate;
211 u64 last_sample_time;
215 u64 prev_cummulative_iowait;
216 struct sample sample;
218 struct acpi_processor_performance acpi_perf_data;
219 bool valid_pss_table;
223 static struct cpudata **all_cpu_data;
226 * struct pid_adjust_policy - Stores static PID configuration data
227 * @sample_rate_ms: PID calculation sample rate in ms
228 * @sample_rate_ns: Sample rate calculation in ns
229 * @deadband: PID deadband
230 * @setpoint: PID Setpoint
231 * @p_gain_pct: PID proportional gain
232 * @i_gain_pct: PID integral gain
233 * @d_gain_pct: PID derivative gain
235 * Stores per CPU model static PID configuration data.
237 struct pstate_adjust_policy {
248 * struct pstate_funcs - Per CPU model specific callbacks
249 * @get_max: Callback to get maximum non turbo effective P state
250 * @get_max_physical: Callback to get maximum non turbo physical P state
251 * @get_min: Callback to get minimum P state
252 * @get_turbo: Callback to get turbo P state
253 * @get_scaling: Callback to get frequency scaling factor
254 * @get_val: Callback to convert P state to actual MSR write value
255 * @get_vid: Callback to get VID data for Atom platforms
256 * @get_target_pstate: Callback to a function to calculate next P state to use
258 * Core and Atom CPU models have different way to get P State limits. This
259 * structure is used to store those callbacks.
261 struct pstate_funcs {
262 int (*get_max)(void);
263 int (*get_max_physical)(void);
264 int (*get_min)(void);
265 int (*get_turbo)(void);
266 int (*get_scaling)(void);
267 u64 (*get_val)(struct cpudata*, int pstate);
268 void (*get_vid)(struct cpudata *);
269 int32_t (*get_target_pstate)(struct cpudata *);
273 * struct cpu_defaults- Per CPU model default config data
274 * @pid_policy: PID config data
275 * @funcs: Callback function data
277 struct cpu_defaults {
278 struct pstate_adjust_policy pid_policy;
279 struct pstate_funcs funcs;
282 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
283 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
285 static struct pstate_adjust_policy pid_params;
286 static struct pstate_funcs pstate_funcs;
287 static int hwp_active;
290 static bool acpi_ppc;
294 * struct perf_limits - Store user and policy limits
295 * @no_turbo: User requested turbo state from intel_pstate sysfs
296 * @turbo_disabled: Platform turbo status either from msr
297 * MSR_IA32_MISC_ENABLE or when maximum available pstate
298 * matches the maximum turbo pstate
299 * @max_perf_pct: Effective maximum performance limit in percentage, this
300 * is minimum of either limits enforced by cpufreq policy
301 * or limits from user set limits via intel_pstate sysfs
302 * @min_perf_pct: Effective minimum performance limit in percentage, this
303 * is maximum of either limits enforced by cpufreq policy
304 * or limits from user set limits via intel_pstate sysfs
305 * @max_perf: This is a scaled value between 0 to 255 for max_perf_pct
306 * This value is used to limit max pstate
307 * @min_perf: This is a scaled value between 0 to 255 for min_perf_pct
308 * This value is used to limit min pstate
309 * @max_policy_pct: The maximum performance in percentage enforced by
310 * cpufreq setpolicy interface
311 * @max_sysfs_pct: The maximum performance in percentage enforced by
312 * intel pstate sysfs interface
313 * @min_policy_pct: The minimum performance in percentage enforced by
314 * cpufreq setpolicy interface
315 * @min_sysfs_pct: The minimum performance in percentage enforced by
316 * intel pstate sysfs interface
318 * Storage for user and policy defined limits.
333 static struct perf_limits performance_limits = {
337 .max_perf = int_tofp(1),
339 .min_perf = int_tofp(1),
340 .max_policy_pct = 100,
341 .max_sysfs_pct = 100,
346 static struct perf_limits powersave_limits = {
350 .max_perf = int_tofp(1),
353 .max_policy_pct = 100,
354 .max_sysfs_pct = 100,
359 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
360 static struct perf_limits *limits = &performance_limits;
362 static struct perf_limits *limits = &powersave_limits;
367 static bool intel_pstate_get_ppc_enable_status(void)
369 if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
370 acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
377 * The max target pstate ratio is a 8 bit value in both PLATFORM_INFO MSR and
378 * in TURBO_RATIO_LIMIT MSR, which pstate driver stores in max_pstate and
379 * max_turbo_pstate fields. The PERF_CTL MSR contains 16 bit value for P state
380 * ratio, out of it only high 8 bits are used. For example 0x1700 is setting
381 * target ratio 0x17. The _PSS control value stores in a format which can be
382 * directly written to PERF_CTL MSR. But in intel_pstate driver this shift
383 * occurs during write to PERF_CTL (E.g. for cores core_set_pstate()).
384 * This function converts the _PSS control value to intel pstate driver format
385 * for comparison and assignment.
387 static int convert_to_native_pstate_format(struct cpudata *cpu, int index)
389 return cpu->acpi_perf_data.states[index].control >> 8;
392 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
402 if (!intel_pstate_get_ppc_enable_status())
405 cpu = all_cpu_data[policy->cpu];
407 ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
413 * Check if the control value in _PSS is for PERF_CTL MSR, which should
414 * guarantee that the states returned by it map to the states in our
417 if (cpu->acpi_perf_data.control_register.space_id !=
418 ACPI_ADR_SPACE_FIXED_HARDWARE)
422 * If there is only one entry _PSS, simply ignore _PSS and continue as
423 * usual without taking _PSS into account
425 if (cpu->acpi_perf_data.state_count < 2)
428 pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
429 for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
430 pr_debug(" %cP%d: %u MHz, %u mW, 0x%x\n",
431 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
432 (u32) cpu->acpi_perf_data.states[i].core_frequency,
433 (u32) cpu->acpi_perf_data.states[i].power,
434 (u32) cpu->acpi_perf_data.states[i].control);
438 * The _PSS table doesn't contain whole turbo frequency range.
439 * This just contains +1 MHZ above the max non turbo frequency,
440 * with control value corresponding to max turbo ratio. But
441 * when cpufreq set policy is called, it will call with this
442 * max frequency, which will cause a reduced performance as
443 * this driver uses real max turbo frequency as the max
444 * frequency. So correct this frequency in _PSS table to
445 * correct max turbo frequency based on the turbo ratio.
446 * Also need to convert to MHz as _PSS freq is in MHz.
448 turbo_pss_ctl = convert_to_native_pstate_format(cpu, 0);
449 if (turbo_pss_ctl > cpu->pstate.max_pstate)
450 cpu->acpi_perf_data.states[0].core_frequency =
451 policy->cpuinfo.max_freq / 1000;
452 cpu->valid_pss_table = true;
453 pr_debug("_PPC limits will be enforced\n");
458 cpu->valid_pss_table = false;
459 acpi_processor_unregister_performance(policy->cpu);
462 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
466 cpu = all_cpu_data[policy->cpu];
467 if (!cpu->valid_pss_table)
470 acpi_processor_unregister_performance(policy->cpu);
474 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
478 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
483 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
484 int deadband, int integral) {
485 pid->setpoint = int_tofp(setpoint);
486 pid->deadband = int_tofp(deadband);
487 pid->integral = int_tofp(integral);
488 pid->last_err = int_tofp(setpoint) - int_tofp(busy);
491 static inline void pid_p_gain_set(struct _pid *pid, int percent)
493 pid->p_gain = div_fp(percent, 100);
496 static inline void pid_i_gain_set(struct _pid *pid, int percent)
498 pid->i_gain = div_fp(percent, 100);
501 static inline void pid_d_gain_set(struct _pid *pid, int percent)
503 pid->d_gain = div_fp(percent, 100);
506 static signed int pid_calc(struct _pid *pid, int32_t busy)
509 int32_t pterm, dterm, fp_error;
510 int32_t integral_limit;
512 fp_error = pid->setpoint - busy;
514 if (abs(fp_error) <= pid->deadband)
517 pterm = mul_fp(pid->p_gain, fp_error);
519 pid->integral += fp_error;
522 * We limit the integral here so that it will never
523 * get higher than 30. This prevents it from becoming
524 * too large an input over long periods of time and allows
525 * it to get factored out sooner.
527 * The value of 30 was chosen through experimentation.
529 integral_limit = int_tofp(30);
530 if (pid->integral > integral_limit)
531 pid->integral = integral_limit;
532 if (pid->integral < -integral_limit)
533 pid->integral = -integral_limit;
535 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
536 pid->last_err = fp_error;
538 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
539 result = result + (1 << (FRAC_BITS-1));
540 return (signed int)fp_toint(result);
543 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
545 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
546 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
547 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
549 pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
552 static inline void intel_pstate_reset_all_pid(void)
556 for_each_online_cpu(cpu) {
557 if (all_cpu_data[cpu])
558 intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
562 static inline void update_turbo_state(void)
567 cpu = all_cpu_data[0];
568 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
569 limits->turbo_disabled =
570 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
571 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
574 static void intel_pstate_hwp_set(const struct cpumask *cpumask)
576 int min, hw_min, max, hw_max, cpu, range, adj_range;
579 rdmsrl(MSR_HWP_CAPABILITIES, cap);
580 hw_min = HWP_LOWEST_PERF(cap);
581 hw_max = HWP_HIGHEST_PERF(cap);
582 range = hw_max - hw_min;
584 for_each_cpu(cpu, cpumask) {
585 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
586 adj_range = limits->min_perf_pct * range / 100;
587 min = hw_min + adj_range;
588 value &= ~HWP_MIN_PERF(~0L);
589 value |= HWP_MIN_PERF(min);
591 adj_range = limits->max_perf_pct * range / 100;
592 max = hw_min + adj_range;
593 if (limits->no_turbo) {
594 hw_max = HWP_GUARANTEED_PERF(cap);
599 value &= ~HWP_MAX_PERF(~0L);
600 value |= HWP_MAX_PERF(max);
601 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
605 static int intel_pstate_hwp_set_policy(struct cpufreq_policy *policy)
608 intel_pstate_hwp_set(policy->cpus);
613 static void intel_pstate_hwp_set_online_cpus(void)
616 intel_pstate_hwp_set(cpu_online_mask);
620 /************************** debugfs begin ************************/
621 static int pid_param_set(void *data, u64 val)
624 intel_pstate_reset_all_pid();
628 static int pid_param_get(void *data, u64 *val)
633 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
640 static struct pid_param pid_files[] = {
641 {"sample_rate_ms", &pid_params.sample_rate_ms},
642 {"d_gain_pct", &pid_params.d_gain_pct},
643 {"i_gain_pct", &pid_params.i_gain_pct},
644 {"deadband", &pid_params.deadband},
645 {"setpoint", &pid_params.setpoint},
646 {"p_gain_pct", &pid_params.p_gain_pct},
650 static void __init intel_pstate_debug_expose_params(void)
652 struct dentry *debugfs_parent;
657 debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
658 if (IS_ERR_OR_NULL(debugfs_parent))
660 while (pid_files[i].name) {
661 debugfs_create_file(pid_files[i].name, 0660,
662 debugfs_parent, pid_files[i].value,
668 /************************** debugfs end ************************/
670 /************************** sysfs begin ************************/
671 #define show_one(file_name, object) \
672 static ssize_t show_##file_name \
673 (struct kobject *kobj, struct attribute *attr, char *buf) \
675 return sprintf(buf, "%u\n", limits->object); \
678 static ssize_t show_turbo_pct(struct kobject *kobj,
679 struct attribute *attr, char *buf)
682 int total, no_turbo, turbo_pct;
685 cpu = all_cpu_data[0];
687 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
688 no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
689 turbo_fp = div_fp(no_turbo, total);
690 turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
691 return sprintf(buf, "%u\n", turbo_pct);
694 static ssize_t show_num_pstates(struct kobject *kobj,
695 struct attribute *attr, char *buf)
700 cpu = all_cpu_data[0];
701 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
702 return sprintf(buf, "%u\n", total);
705 static ssize_t show_no_turbo(struct kobject *kobj,
706 struct attribute *attr, char *buf)
710 update_turbo_state();
711 if (limits->turbo_disabled)
712 ret = sprintf(buf, "%u\n", limits->turbo_disabled);
714 ret = sprintf(buf, "%u\n", limits->no_turbo);
719 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
720 const char *buf, size_t count)
725 ret = sscanf(buf, "%u", &input);
729 update_turbo_state();
730 if (limits->turbo_disabled) {
731 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
735 limits->no_turbo = clamp_t(int, input, 0, 1);
738 intel_pstate_hwp_set_online_cpus();
743 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
744 const char *buf, size_t count)
749 ret = sscanf(buf, "%u", &input);
753 limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
754 limits->max_perf_pct = min(limits->max_policy_pct,
755 limits->max_sysfs_pct);
756 limits->max_perf_pct = max(limits->min_policy_pct,
757 limits->max_perf_pct);
758 limits->max_perf_pct = max(limits->min_perf_pct,
759 limits->max_perf_pct);
760 limits->max_perf = div_fp(limits->max_perf_pct, 100);
763 intel_pstate_hwp_set_online_cpus();
767 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
768 const char *buf, size_t count)
773 ret = sscanf(buf, "%u", &input);
777 limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
778 limits->min_perf_pct = max(limits->min_policy_pct,
779 limits->min_sysfs_pct);
780 limits->min_perf_pct = min(limits->max_policy_pct,
781 limits->min_perf_pct);
782 limits->min_perf_pct = min(limits->max_perf_pct,
783 limits->min_perf_pct);
784 limits->min_perf = div_fp(limits->min_perf_pct, 100);
787 intel_pstate_hwp_set_online_cpus();
791 show_one(max_perf_pct, max_perf_pct);
792 show_one(min_perf_pct, min_perf_pct);
794 define_one_global_rw(no_turbo);
795 define_one_global_rw(max_perf_pct);
796 define_one_global_rw(min_perf_pct);
797 define_one_global_ro(turbo_pct);
798 define_one_global_ro(num_pstates);
800 static struct attribute *intel_pstate_attributes[] = {
809 static struct attribute_group intel_pstate_attr_group = {
810 .attrs = intel_pstate_attributes,
813 static void __init intel_pstate_sysfs_expose_params(void)
815 struct kobject *intel_pstate_kobject;
818 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
819 &cpu_subsys.dev_root->kobj);
820 BUG_ON(!intel_pstate_kobject);
821 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
824 /************************** sysfs end ************************/
826 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
828 /* First disable HWP notification interrupt as we don't process them */
829 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
831 wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
834 static int atom_get_min_pstate(void)
838 rdmsrl(ATOM_RATIOS, value);
839 return (value >> 8) & 0x7F;
842 static int atom_get_max_pstate(void)
846 rdmsrl(ATOM_RATIOS, value);
847 return (value >> 16) & 0x7F;
850 static int atom_get_turbo_pstate(void)
854 rdmsrl(ATOM_TURBO_RATIOS, value);
858 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
864 val = (u64)pstate << 8;
865 if (limits->no_turbo && !limits->turbo_disabled)
868 vid_fp = cpudata->vid.min + mul_fp(
869 int_tofp(pstate - cpudata->pstate.min_pstate),
872 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
873 vid = ceiling_fp(vid_fp);
875 if (pstate > cpudata->pstate.max_pstate)
876 vid = cpudata->vid.turbo;
881 static int silvermont_get_scaling(void)
885 /* Defined in Table 35-6 from SDM (Sept 2015) */
886 static int silvermont_freq_table[] = {
887 83300, 100000, 133300, 116700, 80000};
889 rdmsrl(MSR_FSB_FREQ, value);
893 return silvermont_freq_table[i];
896 static int airmont_get_scaling(void)
900 /* Defined in Table 35-10 from SDM (Sept 2015) */
901 static int airmont_freq_table[] = {
902 83300, 100000, 133300, 116700, 80000,
903 93300, 90000, 88900, 87500};
905 rdmsrl(MSR_FSB_FREQ, value);
909 return airmont_freq_table[i];
912 static void atom_get_vid(struct cpudata *cpudata)
916 rdmsrl(ATOM_VIDS, value);
917 cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
918 cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
919 cpudata->vid.ratio = div_fp(
920 cpudata->vid.max - cpudata->vid.min,
921 int_tofp(cpudata->pstate.max_pstate -
922 cpudata->pstate.min_pstate));
924 rdmsrl(ATOM_TURBO_VIDS, value);
925 cpudata->vid.turbo = value & 0x7f;
928 static int core_get_min_pstate(void)
932 rdmsrl(MSR_PLATFORM_INFO, value);
933 return (value >> 40) & 0xFF;
936 static int core_get_max_pstate_physical(void)
940 rdmsrl(MSR_PLATFORM_INFO, value);
941 return (value >> 8) & 0xFF;
944 static int core_get_max_pstate(void)
951 rdmsrl(MSR_PLATFORM_INFO, plat_info);
952 max_pstate = (plat_info >> 8) & 0xFF;
954 err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
956 /* Do some sanity checking for safety */
957 if (plat_info & 0x600000000) {
962 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
966 tdp_msr = MSR_CONFIG_TDP_NOMINAL + tdp_ctrl;
967 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
971 /* For level 1 and 2, bits[23:16] contain the ratio */
975 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
976 if (tdp_ratio - 1 == tar) {
978 pr_debug("max_pstate=TAC %x\n", max_pstate);
989 static int core_get_turbo_pstate(void)
994 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
995 nont = core_get_max_pstate();
1002 static inline int core_get_scaling(void)
1007 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1011 val = (u64)pstate << 8;
1012 if (limits->no_turbo && !limits->turbo_disabled)
1013 val |= (u64)1 << 32;
1018 static int knl_get_turbo_pstate(void)
1023 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
1024 nont = core_get_max_pstate();
1025 ret = (((value) >> 8) & 0xFF);
1031 static struct cpu_defaults core_params = {
1033 .sample_rate_ms = 10,
1041 .get_max = core_get_max_pstate,
1042 .get_max_physical = core_get_max_pstate_physical,
1043 .get_min = core_get_min_pstate,
1044 .get_turbo = core_get_turbo_pstate,
1045 .get_scaling = core_get_scaling,
1046 .get_val = core_get_val,
1047 .get_target_pstate = get_target_pstate_use_performance,
1051 static struct cpu_defaults silvermont_params = {
1053 .sample_rate_ms = 10,
1061 .get_max = atom_get_max_pstate,
1062 .get_max_physical = atom_get_max_pstate,
1063 .get_min = atom_get_min_pstate,
1064 .get_turbo = atom_get_turbo_pstate,
1065 .get_val = atom_get_val,
1066 .get_scaling = silvermont_get_scaling,
1067 .get_vid = atom_get_vid,
1068 .get_target_pstate = get_target_pstate_use_cpu_load,
1072 static struct cpu_defaults airmont_params = {
1074 .sample_rate_ms = 10,
1082 .get_max = atom_get_max_pstate,
1083 .get_max_physical = atom_get_max_pstate,
1084 .get_min = atom_get_min_pstate,
1085 .get_turbo = atom_get_turbo_pstate,
1086 .get_val = atom_get_val,
1087 .get_scaling = airmont_get_scaling,
1088 .get_vid = atom_get_vid,
1089 .get_target_pstate = get_target_pstate_use_cpu_load,
1093 static struct cpu_defaults knl_params = {
1095 .sample_rate_ms = 10,
1103 .get_max = core_get_max_pstate,
1104 .get_max_physical = core_get_max_pstate_physical,
1105 .get_min = core_get_min_pstate,
1106 .get_turbo = knl_get_turbo_pstate,
1107 .get_scaling = core_get_scaling,
1108 .get_val = core_get_val,
1109 .get_target_pstate = get_target_pstate_use_performance,
1113 static struct cpu_defaults bxt_params = {
1115 .sample_rate_ms = 10,
1123 .get_max = core_get_max_pstate,
1124 .get_max_physical = core_get_max_pstate_physical,
1125 .get_min = core_get_min_pstate,
1126 .get_turbo = core_get_turbo_pstate,
1127 .get_scaling = core_get_scaling,
1128 .get_val = core_get_val,
1129 .get_target_pstate = get_target_pstate_use_cpu_load,
1133 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
1135 int max_perf = cpu->pstate.turbo_pstate;
1139 if (limits->no_turbo || limits->turbo_disabled)
1140 max_perf = cpu->pstate.max_pstate;
1143 * performance can be limited by user through sysfs, by cpufreq
1144 * policy, or by cpu specific default values determined through
1147 max_perf_adj = fp_toint(max_perf * limits->max_perf);
1148 *max = clamp_t(int, max_perf_adj,
1149 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
1151 min_perf = fp_toint(max_perf * limits->min_perf);
1152 *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
1155 static inline void intel_pstate_record_pstate(struct cpudata *cpu, int pstate)
1157 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1158 cpu->pstate.current_pstate = pstate;
1161 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1163 int pstate = cpu->pstate.min_pstate;
1165 intel_pstate_record_pstate(cpu, pstate);
1167 * Generally, there is no guarantee that this code will always run on
1168 * the CPU being updated, so force the register update to run on the
1171 wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1172 pstate_funcs.get_val(cpu, pstate));
1175 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1177 cpu->pstate.min_pstate = pstate_funcs.get_min();
1178 cpu->pstate.max_pstate = pstate_funcs.get_max();
1179 cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1180 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1181 cpu->pstate.scaling = pstate_funcs.get_scaling();
1183 if (pstate_funcs.get_vid)
1184 pstate_funcs.get_vid(cpu);
1186 intel_pstate_set_min_pstate(cpu);
1189 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1191 struct sample *sample = &cpu->sample;
1193 sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1196 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1199 unsigned long flags;
1202 local_irq_save(flags);
1203 rdmsrl(MSR_IA32_APERF, aperf);
1204 rdmsrl(MSR_IA32_MPERF, mperf);
1206 if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1207 local_irq_restore(flags);
1210 local_irq_restore(flags);
1212 cpu->last_sample_time = cpu->sample.time;
1213 cpu->sample.time = time;
1214 cpu->sample.aperf = aperf;
1215 cpu->sample.mperf = mperf;
1216 cpu->sample.tsc = tsc;
1217 cpu->sample.aperf -= cpu->prev_aperf;
1218 cpu->sample.mperf -= cpu->prev_mperf;
1219 cpu->sample.tsc -= cpu->prev_tsc;
1221 cpu->prev_aperf = aperf;
1222 cpu->prev_mperf = mperf;
1223 cpu->prev_tsc = tsc;
1225 * First time this function is invoked in a given cycle, all of the
1226 * previous sample data fields are equal to zero or stale and they must
1227 * be populated with meaningful numbers for things to work, so assume
1228 * that sample.time will always be reset before setting the utilization
1229 * update hook and make the caller skip the sample then.
1231 return !!cpu->last_sample_time;
1234 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1236 return mul_ext_fp(cpu->sample.core_avg_perf,
1237 cpu->pstate.max_pstate_physical * cpu->pstate.scaling);
1240 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1242 return mul_ext_fp(cpu->pstate.max_pstate_physical,
1243 cpu->sample.core_avg_perf);
1246 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1248 struct sample *sample = &cpu->sample;
1249 u64 cummulative_iowait, delta_iowait_us;
1250 u64 delta_iowait_mperf;
1254 cummulative_iowait = get_cpu_iowait_time_us(cpu->cpu, &now);
1257 * Convert iowait time into number of IO cycles spent at max_freq.
1258 * IO is considered as busy only for the cpu_load algorithm. For
1259 * performance this is not needed since we always try to reach the
1260 * maximum P-State, so we are already boosting the IOs.
1262 delta_iowait_us = cummulative_iowait - cpu->prev_cummulative_iowait;
1263 delta_iowait_mperf = div64_u64(delta_iowait_us * cpu->pstate.scaling *
1264 cpu->pstate.max_pstate, MSEC_PER_SEC);
1266 mperf = cpu->sample.mperf + delta_iowait_mperf;
1267 cpu->prev_cummulative_iowait = cummulative_iowait;
1270 * The load can be estimated as the ratio of the mperf counter
1271 * running at a constant frequency during active periods
1272 * (C0) and the time stamp counter running at the same frequency
1273 * also during C-states.
1275 cpu_load = div64_u64(int_tofp(100) * mperf, sample->tsc);
1276 cpu->sample.busy_scaled = cpu_load;
1278 return get_avg_pstate(cpu) - pid_calc(&cpu->pid, cpu_load);
1281 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1283 int32_t perf_scaled, max_pstate, current_pstate, sample_ratio;
1287 * perf_scaled is the average performance during the last sampling
1288 * period scaled by the ratio of the maximum P-state to the P-state
1289 * requested last time (in percent). That measures the system's
1290 * response to the previous P-state selection.
1292 max_pstate = cpu->pstate.max_pstate_physical;
1293 current_pstate = cpu->pstate.current_pstate;
1294 perf_scaled = mul_ext_fp(cpu->sample.core_avg_perf,
1295 div_fp(100 * max_pstate, current_pstate));
1298 * Since our utilization update callback will not run unless we are
1299 * in C0, check if the actual elapsed time is significantly greater (3x)
1300 * than our sample interval. If it is, then we were idle for a long
1301 * enough period of time to adjust our performance metric.
1303 duration_ns = cpu->sample.time - cpu->last_sample_time;
1304 if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1305 sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1306 perf_scaled = mul_fp(perf_scaled, sample_ratio);
1308 sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1309 if (sample_ratio < int_tofp(1))
1313 cpu->sample.busy_scaled = perf_scaled;
1314 return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
1317 static inline void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1319 int max_perf, min_perf;
1321 update_turbo_state();
1323 intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1324 pstate = clamp_t(int, pstate, min_perf, max_perf);
1325 if (pstate == cpu->pstate.current_pstate)
1328 intel_pstate_record_pstate(cpu, pstate);
1329 wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1332 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1334 int from, target_pstate;
1335 struct sample *sample;
1337 from = cpu->pstate.current_pstate;
1339 target_pstate = pstate_funcs.get_target_pstate(cpu);
1341 intel_pstate_update_pstate(cpu, target_pstate);
1343 sample = &cpu->sample;
1344 trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1345 fp_toint(sample->busy_scaled),
1347 cpu->pstate.current_pstate,
1351 get_avg_frequency(cpu));
1354 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1355 unsigned long util, unsigned long max)
1357 struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1358 u64 delta_ns = time - cpu->sample.time;
1360 if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1361 bool sample_taken = intel_pstate_sample(cpu, time);
1364 intel_pstate_calc_avg_perf(cpu);
1366 intel_pstate_adjust_busy_pstate(cpu);
1371 #define ICPU(model, policy) \
1372 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1373 (unsigned long)&policy }
1375 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1376 ICPU(INTEL_FAM6_SANDYBRIDGE, core_params),
1377 ICPU(INTEL_FAM6_SANDYBRIDGE_X, core_params),
1378 ICPU(INTEL_FAM6_ATOM_SILVERMONT1, silvermont_params),
1379 ICPU(INTEL_FAM6_IVYBRIDGE, core_params),
1380 ICPU(INTEL_FAM6_HASWELL_CORE, core_params),
1381 ICPU(INTEL_FAM6_BROADWELL_CORE, core_params),
1382 ICPU(INTEL_FAM6_IVYBRIDGE_X, core_params),
1383 ICPU(INTEL_FAM6_HASWELL_X, core_params),
1384 ICPU(INTEL_FAM6_HASWELL_ULT, core_params),
1385 ICPU(INTEL_FAM6_HASWELL_GT3E, core_params),
1386 ICPU(INTEL_FAM6_BROADWELL_GT3E, core_params),
1387 ICPU(INTEL_FAM6_ATOM_AIRMONT, airmont_params),
1388 ICPU(INTEL_FAM6_SKYLAKE_MOBILE, core_params),
1389 ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1390 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_params),
1391 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1392 ICPU(INTEL_FAM6_XEON_PHI_KNL, knl_params),
1393 ICPU(INTEL_FAM6_ATOM_GOLDMONT, bxt_params),
1396 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1398 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
1399 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1403 static int intel_pstate_init_cpu(unsigned int cpunum)
1405 struct cpudata *cpu;
1407 if (!all_cpu_data[cpunum])
1408 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
1410 if (!all_cpu_data[cpunum])
1413 cpu = all_cpu_data[cpunum];
1418 intel_pstate_hwp_enable(cpu);
1419 pid_params.sample_rate_ms = 50;
1420 pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
1423 intel_pstate_get_cpu_pstates(cpu);
1425 intel_pstate_busy_pid_reset(cpu);
1427 pr_debug("controlling: cpu %d\n", cpunum);
1432 static unsigned int intel_pstate_get(unsigned int cpu_num)
1434 struct cpudata *cpu = all_cpu_data[cpu_num];
1436 return cpu ? get_avg_frequency(cpu) : 0;
1439 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1441 struct cpudata *cpu = all_cpu_data[cpu_num];
1443 /* Prevent intel_pstate_update_util() from using stale data. */
1444 cpu->sample.time = 0;
1445 cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1446 intel_pstate_update_util);
1447 cpu->update_util_set = true;
1450 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1452 struct cpudata *cpu_data = all_cpu_data[cpu];
1454 if (!cpu_data->update_util_set)
1457 cpufreq_remove_update_util_hook(cpu);
1458 cpu_data->update_util_set = false;
1459 synchronize_sched();
1462 static void intel_pstate_set_performance_limits(struct perf_limits *limits)
1464 limits->no_turbo = 0;
1465 limits->turbo_disabled = 0;
1466 limits->max_perf_pct = 100;
1467 limits->max_perf = int_tofp(1);
1468 limits->min_perf_pct = 100;
1469 limits->min_perf = int_tofp(1);
1470 limits->max_policy_pct = 100;
1471 limits->max_sysfs_pct = 100;
1472 limits->min_policy_pct = 0;
1473 limits->min_sysfs_pct = 0;
1476 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
1478 struct cpudata *cpu;
1480 if (!policy->cpuinfo.max_freq)
1483 intel_pstate_clear_update_util_hook(policy->cpu);
1485 pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
1486 policy->cpuinfo.max_freq, policy->max);
1488 cpu = all_cpu_data[0];
1489 if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
1490 policy->max < policy->cpuinfo.max_freq &&
1491 policy->max > cpu->pstate.max_pstate * cpu->pstate.scaling) {
1492 pr_debug("policy->max > max non turbo frequency\n");
1493 policy->max = policy->cpuinfo.max_freq;
1496 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
1497 limits = &performance_limits;
1498 if (policy->max >= policy->cpuinfo.max_freq) {
1499 pr_debug("set performance\n");
1500 intel_pstate_set_performance_limits(limits);
1504 pr_debug("set powersave\n");
1505 limits = &powersave_limits;
1508 limits->min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
1509 limits->min_policy_pct = clamp_t(int, limits->min_policy_pct, 0 , 100);
1510 limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
1511 policy->cpuinfo.max_freq);
1512 limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 , 100);
1514 /* Normalize user input to [min_policy_pct, max_policy_pct] */
1515 limits->min_perf_pct = max(limits->min_policy_pct,
1516 limits->min_sysfs_pct);
1517 limits->min_perf_pct = min(limits->max_policy_pct,
1518 limits->min_perf_pct);
1519 limits->max_perf_pct = min(limits->max_policy_pct,
1520 limits->max_sysfs_pct);
1521 limits->max_perf_pct = max(limits->min_policy_pct,
1522 limits->max_perf_pct);
1524 /* Make sure min_perf_pct <= max_perf_pct */
1525 limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
1527 limits->min_perf = div_fp(limits->min_perf_pct, 100);
1528 limits->max_perf = div_fp(limits->max_perf_pct, 100);
1529 limits->max_perf = round_up(limits->max_perf, FRAC_BITS);
1532 intel_pstate_set_update_util_hook(policy->cpu);
1534 intel_pstate_hwp_set_policy(policy);
1539 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
1541 cpufreq_verify_within_cpu_limits(policy);
1543 if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
1544 policy->policy != CPUFREQ_POLICY_PERFORMANCE)
1550 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1552 int cpu_num = policy->cpu;
1553 struct cpudata *cpu = all_cpu_data[cpu_num];
1555 pr_debug("CPU %d exiting\n", cpu_num);
1557 intel_pstate_clear_update_util_hook(cpu_num);
1562 intel_pstate_set_min_pstate(cpu);
1565 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1567 struct cpudata *cpu;
1570 rc = intel_pstate_init_cpu(policy->cpu);
1574 cpu = all_cpu_data[policy->cpu];
1576 if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
1577 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1579 policy->policy = CPUFREQ_POLICY_POWERSAVE;
1581 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1582 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1584 /* cpuinfo and default policy values */
1585 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1586 update_turbo_state();
1587 policy->cpuinfo.max_freq = limits->turbo_disabled ?
1588 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1589 policy->cpuinfo.max_freq *= cpu->pstate.scaling;
1591 intel_pstate_init_acpi_perf_limits(policy);
1592 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1593 cpumask_set_cpu(policy->cpu, policy->cpus);
1598 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
1600 intel_pstate_exit_perf_limits(policy);
1605 static struct cpufreq_driver intel_pstate_driver = {
1606 .flags = CPUFREQ_CONST_LOOPS,
1607 .verify = intel_pstate_verify_policy,
1608 .setpolicy = intel_pstate_set_policy,
1609 .resume = intel_pstate_hwp_set_policy,
1610 .get = intel_pstate_get,
1611 .init = intel_pstate_cpu_init,
1612 .exit = intel_pstate_cpu_exit,
1613 .stop_cpu = intel_pstate_stop_cpu,
1614 .name = "intel_pstate",
1617 static int __initdata no_load;
1618 static int __initdata no_hwp;
1619 static int __initdata hwp_only;
1620 static unsigned int force_load;
1622 static int intel_pstate_msrs_not_valid(void)
1624 if (!pstate_funcs.get_max() ||
1625 !pstate_funcs.get_min() ||
1626 !pstate_funcs.get_turbo())
1632 static void copy_pid_params(struct pstate_adjust_policy *policy)
1634 pid_params.sample_rate_ms = policy->sample_rate_ms;
1635 pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
1636 pid_params.p_gain_pct = policy->p_gain_pct;
1637 pid_params.i_gain_pct = policy->i_gain_pct;
1638 pid_params.d_gain_pct = policy->d_gain_pct;
1639 pid_params.deadband = policy->deadband;
1640 pid_params.setpoint = policy->setpoint;
1643 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1645 pstate_funcs.get_max = funcs->get_max;
1646 pstate_funcs.get_max_physical = funcs->get_max_physical;
1647 pstate_funcs.get_min = funcs->get_min;
1648 pstate_funcs.get_turbo = funcs->get_turbo;
1649 pstate_funcs.get_scaling = funcs->get_scaling;
1650 pstate_funcs.get_val = funcs->get_val;
1651 pstate_funcs.get_vid = funcs->get_vid;
1652 pstate_funcs.get_target_pstate = funcs->get_target_pstate;
1658 static bool intel_pstate_no_acpi_pss(void)
1662 for_each_possible_cpu(i) {
1664 union acpi_object *pss;
1665 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1666 struct acpi_processor *pr = per_cpu(processors, i);
1671 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1672 if (ACPI_FAILURE(status))
1675 pss = buffer.pointer;
1676 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1687 static bool intel_pstate_has_acpi_ppc(void)
1691 for_each_possible_cpu(i) {
1692 struct acpi_processor *pr = per_cpu(processors, i);
1696 if (acpi_has_method(pr->handle, "_PPC"))
1707 struct hw_vendor_info {
1709 char oem_id[ACPI_OEM_ID_SIZE];
1710 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1714 /* Hardware vendor-specific info that has its own power management modes */
1715 static struct hw_vendor_info vendor_info[] = {
1716 {1, "HP ", "ProLiant", PSS},
1717 {1, "ORACLE", "X4-2 ", PPC},
1718 {1, "ORACLE", "X4-2L ", PPC},
1719 {1, "ORACLE", "X4-2B ", PPC},
1720 {1, "ORACLE", "X3-2 ", PPC},
1721 {1, "ORACLE", "X3-2L ", PPC},
1722 {1, "ORACLE", "X3-2B ", PPC},
1723 {1, "ORACLE", "X4470M2 ", PPC},
1724 {1, "ORACLE", "X4270M3 ", PPC},
1725 {1, "ORACLE", "X4270M2 ", PPC},
1726 {1, "ORACLE", "X4170M2 ", PPC},
1727 {1, "ORACLE", "X4170 M3", PPC},
1728 {1, "ORACLE", "X4275 M3", PPC},
1729 {1, "ORACLE", "X6-2 ", PPC},
1730 {1, "ORACLE", "Sudbury ", PPC},
1734 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1736 struct acpi_table_header hdr;
1737 struct hw_vendor_info *v_info;
1738 const struct x86_cpu_id *id;
1741 id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1743 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1744 if ( misc_pwr & (1 << 8))
1748 if (acpi_disabled ||
1749 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1752 for (v_info = vendor_info; v_info->valid; v_info++) {
1753 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1754 !strncmp(hdr.oem_table_id, v_info->oem_table_id,
1755 ACPI_OEM_TABLE_ID_SIZE))
1756 switch (v_info->oem_pwr_table) {
1758 return intel_pstate_no_acpi_pss();
1760 return intel_pstate_has_acpi_ppc() &&
1767 #else /* CONFIG_ACPI not enabled */
1768 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1769 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1770 #endif /* CONFIG_ACPI */
1772 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
1773 { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
1777 static int __init intel_pstate_init(void)
1780 const struct x86_cpu_id *id;
1781 struct cpu_defaults *cpu_def;
1786 if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
1787 copy_cpu_funcs(&core_params.funcs);
1789 goto hwp_cpu_matched;
1792 id = x86_match_cpu(intel_pstate_cpu_ids);
1796 cpu_def = (struct cpu_defaults *)id->driver_data;
1798 copy_pid_params(&cpu_def->pid_policy);
1799 copy_cpu_funcs(&cpu_def->funcs);
1801 if (intel_pstate_msrs_not_valid())
1806 * The Intel pstate driver will be ignored if the platform
1807 * firmware has its own power management modes.
1809 if (intel_pstate_platform_pwr_mgmt_exists())
1812 pr_info("Intel P-state driver initializing\n");
1814 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1818 if (!hwp_active && hwp_only)
1821 rc = cpufreq_register_driver(&intel_pstate_driver);
1825 intel_pstate_debug_expose_params();
1826 intel_pstate_sysfs_expose_params();
1829 pr_info("HWP enabled\n");
1834 for_each_online_cpu(cpu) {
1835 if (all_cpu_data[cpu]) {
1836 intel_pstate_clear_update_util_hook(cpu);
1837 kfree(all_cpu_data[cpu]);
1842 vfree(all_cpu_data);
1845 device_initcall(intel_pstate_init);
1847 static int __init intel_pstate_setup(char *str)
1852 if (!strcmp(str, "disable"))
1854 if (!strcmp(str, "no_hwp")) {
1855 pr_info("HWP disabled\n");
1858 if (!strcmp(str, "force"))
1860 if (!strcmp(str, "hwp_only"))
1864 if (!strcmp(str, "support_acpi_ppc"))
1870 early_param("intel_pstate", intel_pstate_setup);
1872 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1873 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1874 MODULE_LICENSE("GPL");