1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * acpi-cpufreq.c - ACPI Processor P-States Driver
11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/init.h>
16 #include <linux/smp.h>
17 #include <linux/sched.h>
18 #include <linux/cpufreq.h>
19 #include <linux/compiler.h>
20 #include <linux/dmi.h>
21 #include <linux/slab.h>
22 #include <linux/string_helpers.h>
23 #include <linux/platform_device.h>
25 #include <linux/acpi.h>
27 #include <linux/delay.h>
28 #include <linux/uaccess.h>
30 #include <acpi/processor.h>
31 #include <acpi/cppc_acpi.h>
34 #include <asm/processor.h>
35 #include <asm/cpufeature.h>
36 #include <asm/cpu_device_id.h>
38 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
39 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
40 MODULE_LICENSE("GPL");
43 UNDEFINED_CAPABLE = 0,
44 SYSTEM_INTEL_MSR_CAPABLE,
45 SYSTEM_AMD_MSR_CAPABLE,
49 #define INTEL_MSR_RANGE (0xffff)
50 #define AMD_MSR_RANGE (0x7)
51 #define HYGON_MSR_RANGE (0x7)
53 struct acpi_cpufreq_data {
55 unsigned int cpu_feature;
56 unsigned int acpi_perf_cpu;
57 cpumask_var_t freqdomain_cpus;
58 void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
59 u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
62 /* acpi_perf_data is a pointer to percpu data. */
63 static struct acpi_processor_performance __percpu *acpi_perf_data;
65 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
67 return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
70 static struct cpufreq_driver acpi_cpufreq_driver;
72 static unsigned int acpi_pstate_strict;
74 static bool boost_state(unsigned int cpu)
79 switch (boot_cpu_data.x86_vendor) {
80 case X86_VENDOR_INTEL:
81 case X86_VENDOR_CENTAUR:
82 case X86_VENDOR_ZHAOXIN:
83 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
84 msr = lo | ((u64)hi << 32);
85 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
86 case X86_VENDOR_HYGON:
88 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
89 msr = lo | ((u64)hi << 32);
90 return !(msr & MSR_K7_HWCR_CPB_DIS);
95 static int boost_set_msr(bool enable)
100 switch (boot_cpu_data.x86_vendor) {
101 case X86_VENDOR_INTEL:
102 case X86_VENDOR_CENTAUR:
103 case X86_VENDOR_ZHAOXIN:
104 msr_addr = MSR_IA32_MISC_ENABLE;
105 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
107 case X86_VENDOR_HYGON:
109 msr_addr = MSR_K7_HWCR;
110 msr_mask = MSR_K7_HWCR_CPB_DIS;
116 rdmsrl(msr_addr, val);
123 wrmsrl(msr_addr, val);
127 static void boost_set_msr_each(void *p_en)
129 bool enable = (bool) p_en;
131 boost_set_msr(enable);
134 static int set_boost(struct cpufreq_policy *policy, int val)
136 on_each_cpu_mask(policy->cpus, boost_set_msr_each,
137 (void *)(long)val, 1);
138 pr_debug("CPU %*pbl: Core Boosting %s.\n",
139 cpumask_pr_args(policy->cpus), str_enabled_disabled(val));
144 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
146 struct acpi_cpufreq_data *data = policy->driver_data;
151 return cpufreq_show_cpus(data->freqdomain_cpus, buf);
154 cpufreq_freq_attr_ro(freqdomain_cpus);
156 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
157 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
161 unsigned int val = 0;
163 if (!acpi_cpufreq_driver.set_boost)
166 ret = kstrtouint(buf, 10, &val);
171 set_boost(policy, val);
177 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
179 return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
182 cpufreq_freq_attr_rw(cpb);
185 static int check_est_cpu(unsigned int cpuid)
187 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
189 return cpu_has(cpu, X86_FEATURE_EST);
192 static int check_amd_hwpstate_cpu(unsigned int cpuid)
194 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
196 return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
199 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
201 struct acpi_cpufreq_data *data = policy->driver_data;
202 struct acpi_processor_performance *perf;
205 perf = to_perf_data(data);
207 for (i = 0; i < perf->state_count; i++) {
208 if (value == perf->states[i].status)
209 return policy->freq_table[i].frequency;
214 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
216 struct acpi_cpufreq_data *data = policy->driver_data;
217 struct cpufreq_frequency_table *pos;
218 struct acpi_processor_performance *perf;
220 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
221 msr &= AMD_MSR_RANGE;
222 else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
223 msr &= HYGON_MSR_RANGE;
225 msr &= INTEL_MSR_RANGE;
227 perf = to_perf_data(data);
229 cpufreq_for_each_entry(pos, policy->freq_table)
230 if (msr == perf->states[pos->driver_data].status)
231 return pos->frequency;
232 return policy->freq_table[0].frequency;
235 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
237 struct acpi_cpufreq_data *data = policy->driver_data;
239 switch (data->cpu_feature) {
240 case SYSTEM_INTEL_MSR_CAPABLE:
241 case SYSTEM_AMD_MSR_CAPABLE:
242 return extract_msr(policy, val);
243 case SYSTEM_IO_CAPABLE:
244 return extract_io(policy, val);
250 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
252 u32 val, dummy __always_unused;
254 rdmsr(MSR_IA32_PERF_CTL, val, dummy);
258 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
262 rdmsr(MSR_IA32_PERF_CTL, lo, hi);
263 lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
264 wrmsr(MSR_IA32_PERF_CTL, lo, hi);
267 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
269 u32 val, dummy __always_unused;
271 rdmsr(MSR_AMD_PERF_CTL, val, dummy);
275 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
277 wrmsr(MSR_AMD_PERF_CTL, val, 0);
280 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
284 acpi_os_read_port(reg->address, &val, reg->bit_width);
288 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
290 acpi_os_write_port(reg->address, val, reg->bit_width);
294 struct acpi_pct_register *reg;
297 void (*write)(struct acpi_pct_register *reg, u32 val);
298 u32 (*read)(struct acpi_pct_register *reg);
302 /* Called via smp_call_function_single(), on the target CPU */
303 static void do_drv_read(void *_cmd)
305 struct drv_cmd *cmd = _cmd;
307 cmd->val = cmd->func.read(cmd->reg);
310 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
312 struct acpi_processor_performance *perf = to_perf_data(data);
313 struct drv_cmd cmd = {
314 .reg = &perf->control_register,
315 .func.read = data->cpu_freq_read,
319 err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
320 WARN_ON_ONCE(err); /* smp_call_function_any() was buggy? */
324 /* Called via smp_call_function_many(), on the target CPUs */
325 static void do_drv_write(void *_cmd)
327 struct drv_cmd *cmd = _cmd;
329 cmd->func.write(cmd->reg, cmd->val);
332 static void drv_write(struct acpi_cpufreq_data *data,
333 const struct cpumask *mask, u32 val)
335 struct acpi_processor_performance *perf = to_perf_data(data);
336 struct drv_cmd cmd = {
337 .reg = &perf->control_register,
339 .func.write = data->cpu_freq_write,
343 this_cpu = get_cpu();
344 if (cpumask_test_cpu(this_cpu, mask))
347 smp_call_function_many(mask, do_drv_write, &cmd, 1);
351 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
355 if (unlikely(cpumask_empty(mask)))
358 val = drv_read(data, mask);
360 pr_debug("%s = %u\n", __func__, val);
365 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
367 struct acpi_cpufreq_data *data;
368 struct cpufreq_policy *policy;
370 unsigned int cached_freq;
372 pr_debug("%s (%d)\n", __func__, cpu);
374 policy = cpufreq_cpu_get_raw(cpu);
375 if (unlikely(!policy))
378 data = policy->driver_data;
379 if (unlikely(!data || !policy->freq_table))
382 cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
383 freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
384 if (freq != cached_freq) {
386 * The dreaded BIOS frequency change behind our back.
387 * Force set the frequency on next target call.
392 pr_debug("cur freq = %u\n", freq);
397 static unsigned int check_freqs(struct cpufreq_policy *policy,
398 const struct cpumask *mask, unsigned int freq)
400 struct acpi_cpufreq_data *data = policy->driver_data;
401 unsigned int cur_freq;
404 for (i = 0; i < 100; i++) {
405 cur_freq = extract_freq(policy, get_cur_val(mask, data));
406 if (cur_freq == freq)
413 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
416 struct acpi_cpufreq_data *data = policy->driver_data;
417 struct acpi_processor_performance *perf;
418 const struct cpumask *mask;
419 unsigned int next_perf_state = 0; /* Index into perf table */
422 if (unlikely(!data)) {
426 perf = to_perf_data(data);
427 next_perf_state = policy->freq_table[index].driver_data;
428 if (perf->state == next_perf_state) {
429 if (unlikely(data->resume)) {
430 pr_debug("Called after resume, resetting to P%d\n",
434 pr_debug("Already at target state (P%d)\n",
441 * The core won't allow CPUs to go away until the governor has been
442 * stopped, so we can rely on the stability of policy->cpus.
444 mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
445 cpumask_of(policy->cpu) : policy->cpus;
447 drv_write(data, mask, perf->states[next_perf_state].control);
449 if (acpi_pstate_strict) {
450 if (!check_freqs(policy, mask,
451 policy->freq_table[index].frequency)) {
452 pr_debug("%s (%d)\n", __func__, policy->cpu);
458 perf->state = next_perf_state;
463 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
464 unsigned int target_freq)
466 struct acpi_cpufreq_data *data = policy->driver_data;
467 struct acpi_processor_performance *perf;
468 struct cpufreq_frequency_table *entry;
469 unsigned int next_perf_state, next_freq, index;
472 * Find the closest frequency above target_freq.
474 if (policy->cached_target_freq == target_freq)
475 index = policy->cached_resolved_idx;
477 index = cpufreq_table_find_index_dl(policy, target_freq,
480 entry = &policy->freq_table[index];
481 next_freq = entry->frequency;
482 next_perf_state = entry->driver_data;
484 perf = to_perf_data(data);
485 if (perf->state == next_perf_state) {
486 if (unlikely(data->resume))
492 data->cpu_freq_write(&perf->control_register,
493 perf->states[next_perf_state].control);
494 perf->state = next_perf_state;
499 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
501 struct acpi_processor_performance *perf;
503 perf = to_perf_data(data);
505 /* search the closest match to cpu_khz */
508 unsigned long freqn = perf->states[0].core_frequency * 1000;
510 for (i = 0; i < (perf->state_count-1); i++) {
512 freqn = perf->states[i+1].core_frequency * 1000;
513 if ((2 * cpu_khz) > (freqn + freq)) {
518 perf->state = perf->state_count-1;
521 /* assume CPU is at P0... */
523 return perf->states[0].core_frequency * 1000;
527 static void free_acpi_perf_data(void)
531 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
532 for_each_possible_cpu(i)
533 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
535 free_percpu(acpi_perf_data);
538 static int cpufreq_boost_down_prep(unsigned int cpu)
541 * Clear the boost-disable bit on the CPU_DOWN path so that
542 * this cpu cannot block the remaining ones from boosting.
544 return boost_set_msr(1);
548 * acpi_cpufreq_early_init - initialize ACPI P-States library
550 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
551 * in order to determine correct frequency and voltage pairings. We can
552 * do _PDC and _PSD and find out the processor dependency for the
553 * actual init that will happen later...
555 static int __init acpi_cpufreq_early_init(void)
558 pr_debug("%s\n", __func__);
560 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
561 if (!acpi_perf_data) {
562 pr_debug("Memory allocation error for acpi_perf_data.\n");
565 for_each_possible_cpu(i) {
566 if (!zalloc_cpumask_var_node(
567 &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
568 GFP_KERNEL, cpu_to_node(i))) {
570 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
571 free_acpi_perf_data();
576 /* Do initialization in ACPI core */
577 acpi_processor_preregister_performance(acpi_perf_data);
583 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
584 * or do it in BIOS firmware and won't inform about it to OS. If not
585 * detected, this has a side effect of making CPU run at a different speed
586 * than OS intended it to run at. Detect it and handle it cleanly.
588 static int bios_with_sw_any_bug;
590 static int sw_any_bug_found(const struct dmi_system_id *d)
592 bios_with_sw_any_bug = 1;
596 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
598 .callback = sw_any_bug_found,
599 .ident = "Supermicro Server X6DLP",
601 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
602 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
603 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
609 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
611 /* Intel Xeon Processor 7100 Series Specification Update
612 * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
613 * AL30: A Machine Check Exception (MCE) Occurring during an
614 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
615 * Both Processor Cores to Lock Up. */
616 if (c->x86_vendor == X86_VENDOR_INTEL) {
617 if ((c->x86 == 15) &&
618 (c->x86_model == 6) &&
619 (c->x86_stepping == 8)) {
620 pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
628 #ifdef CONFIG_ACPI_CPPC_LIB
629 static u64 get_max_boost_ratio(unsigned int cpu)
631 struct cppc_perf_caps perf_caps;
632 u64 highest_perf, nominal_perf;
635 if (acpi_pstate_strict)
638 ret = cppc_get_perf_caps(cpu, &perf_caps);
640 pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
645 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
646 ret = amd_get_boost_ratio_numerator(cpu, &highest_perf);
648 pr_debug("CPU%d: Unable to get boost ratio numerator (%d)\n",
653 highest_perf = perf_caps.highest_perf;
656 nominal_perf = perf_caps.nominal_perf;
658 if (!highest_perf || !nominal_perf) {
659 pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
663 if (highest_perf < nominal_perf) {
664 pr_debug("CPU%d: nominal performance above highest\n", cpu);
668 return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
671 static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
674 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
676 struct cpufreq_frequency_table *freq_table;
677 struct acpi_processor_performance *perf;
678 struct acpi_cpufreq_data *data;
679 unsigned int cpu = policy->cpu;
680 struct cpuinfo_x86 *c = &cpu_data(cpu);
681 unsigned int valid_states = 0;
682 unsigned int result = 0;
686 static int blacklisted;
689 pr_debug("%s\n", __func__);
694 blacklisted = acpi_cpufreq_blacklist(c);
699 data = kzalloc(sizeof(*data), GFP_KERNEL);
703 if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
708 perf = per_cpu_ptr(acpi_perf_data, cpu);
709 data->acpi_perf_cpu = cpu;
710 policy->driver_data = data;
712 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
713 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
715 result = acpi_processor_register_performance(perf, cpu);
719 policy->shared_type = perf->shared_type;
722 * Will let policy->cpus know about dependency only when software
723 * coordination is required.
725 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
726 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
727 cpumask_copy(policy->cpus, perf->shared_cpu_map);
729 cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
732 dmi_check_system(sw_any_bug_dmi_table);
733 if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
734 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
735 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
738 if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
739 !acpi_pstate_strict) {
740 cpumask_clear(policy->cpus);
741 cpumask_set_cpu(cpu, policy->cpus);
742 cpumask_copy(data->freqdomain_cpus,
743 topology_sibling_cpumask(cpu));
744 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
745 pr_info_once("overriding BIOS provided _PSD data\n");
749 /* capability check */
750 if (perf->state_count <= 1) {
751 pr_debug("No P-States\n");
756 if (perf->control_register.space_id != perf->status_register.space_id) {
761 switch (perf->control_register.space_id) {
762 case ACPI_ADR_SPACE_SYSTEM_IO:
763 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
764 boot_cpu_data.x86 == 0xf) {
765 pr_debug("AMD K8 systems must use native drivers.\n");
769 pr_debug("SYSTEM IO addr space\n");
770 data->cpu_feature = SYSTEM_IO_CAPABLE;
771 data->cpu_freq_read = cpu_freq_read_io;
772 data->cpu_freq_write = cpu_freq_write_io;
774 case ACPI_ADR_SPACE_FIXED_HARDWARE:
775 pr_debug("HARDWARE addr space\n");
776 if (check_est_cpu(cpu)) {
777 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
778 data->cpu_freq_read = cpu_freq_read_intel;
779 data->cpu_freq_write = cpu_freq_write_intel;
782 if (check_amd_hwpstate_cpu(cpu)) {
783 data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
784 data->cpu_freq_read = cpu_freq_read_amd;
785 data->cpu_freq_write = cpu_freq_write_amd;
791 pr_debug("Unknown addr space %d\n",
792 (u32) (perf->control_register.space_id));
797 freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
804 /* detect transition latency */
805 policy->cpuinfo.transition_latency = 0;
806 for (i = 0; i < perf->state_count; i++) {
807 if ((perf->states[i].transition_latency * 1000) >
808 policy->cpuinfo.transition_latency)
809 policy->cpuinfo.transition_latency =
810 perf->states[i].transition_latency * 1000;
813 /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
814 if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
815 policy->cpuinfo.transition_latency > 20 * 1000) {
816 policy->cpuinfo.transition_latency = 20 * 1000;
817 pr_info_once("P-state transition latency capped at 20 uS\n");
821 for (i = 0; i < perf->state_count; i++) {
822 if (i > 0 && perf->states[i].core_frequency >=
823 freq_table[valid_states-1].frequency / 1000)
826 freq_table[valid_states].driver_data = i;
827 freq_table[valid_states].frequency =
828 perf->states[i].core_frequency * 1000;
831 freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
833 max_boost_ratio = get_max_boost_ratio(cpu);
834 if (max_boost_ratio) {
835 unsigned int freq = freq_table[0].frequency;
838 * Because the loop above sorts the freq_table entries in the
839 * descending order, freq is the maximum frequency in the table.
840 * Assume that it corresponds to the CPPC nominal frequency and
841 * use it to set cpuinfo.max_freq.
843 policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
846 * If the maximum "boost" frequency is unknown, ask the arch
847 * scale-invariance code to use the "nominal" performance for
848 * CPU utilization scaling so as to prevent the schedutil
849 * governor from selecting inadequate CPU frequencies.
851 arch_set_max_freq_ratio(true);
854 policy->freq_table = freq_table;
857 switch (perf->control_register.space_id) {
858 case ACPI_ADR_SPACE_SYSTEM_IO:
860 * The core will not set policy->cur, because
861 * cpufreq_driver->get is NULL, so we need to set it here.
862 * However, we have to guess it, because the current speed is
863 * unknown and not detectable via IO ports.
865 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
867 case ACPI_ADR_SPACE_FIXED_HARDWARE:
868 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
874 /* notify BIOS that we exist */
875 acpi_processor_notify_smm(THIS_MODULE);
877 pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
878 for (i = 0; i < perf->state_count; i++)
879 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n",
880 (i == perf->state ? '*' : ' '), i,
881 (u32) perf->states[i].core_frequency,
882 (u32) perf->states[i].power,
883 (u32) perf->states[i].transition_latency);
886 * the first call to ->target() should result in us actually
887 * writing something to the appropriate registers.
891 policy->fast_switch_possible = !acpi_pstate_strict &&
892 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
894 if (perf->states[0].core_frequency * 1000 != freq_table[0].frequency)
895 pr_warn(FW_WARN "P-state 0 is not max freq\n");
897 if (acpi_cpufreq_driver.set_boost) {
898 set_boost(policy, acpi_cpufreq_driver.boost_enabled);
899 policy->boost_enabled = acpi_cpufreq_driver.boost_enabled;
905 acpi_processor_unregister_performance(cpu);
907 free_cpumask_var(data->freqdomain_cpus);
910 policy->driver_data = NULL;
915 static void acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
917 struct acpi_cpufreq_data *data = policy->driver_data;
919 pr_debug("%s\n", __func__);
921 cpufreq_boost_down_prep(policy->cpu);
922 policy->fast_switch_possible = false;
923 policy->driver_data = NULL;
924 acpi_processor_unregister_performance(data->acpi_perf_cpu);
925 free_cpumask_var(data->freqdomain_cpus);
926 kfree(policy->freq_table);
930 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
932 struct acpi_cpufreq_data *data = policy->driver_data;
934 pr_debug("%s\n", __func__);
941 static struct freq_attr *acpi_cpufreq_attr[] = {
942 &cpufreq_freq_attr_scaling_available_freqs,
944 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
950 static struct cpufreq_driver acpi_cpufreq_driver = {
951 .verify = cpufreq_generic_frequency_table_verify,
952 .target_index = acpi_cpufreq_target,
953 .fast_switch = acpi_cpufreq_fast_switch,
954 .bios_limit = acpi_processor_get_bios_limit,
955 .init = acpi_cpufreq_cpu_init,
956 .exit = acpi_cpufreq_cpu_exit,
957 .resume = acpi_cpufreq_resume,
958 .name = "acpi-cpufreq",
959 .attr = acpi_cpufreq_attr,
962 static void __init acpi_cpufreq_boost_init(void)
964 if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
965 pr_debug("Boost capabilities not present in the processor\n");
969 acpi_cpufreq_driver.set_boost = set_boost;
970 acpi_cpufreq_driver.boost_enabled = boost_state(0);
973 static int __init acpi_cpufreq_probe(struct platform_device *pdev)
980 /* don't keep reloading if cpufreq_driver exists */
981 if (cpufreq_get_current_driver())
984 pr_debug("%s\n", __func__);
986 ret = acpi_cpufreq_early_init();
990 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
991 /* this is a sysfs file with a strange name and an even stranger
992 * semantic - per CPU instantiation, but system global effect.
993 * Lets enable it only on AMD CPUs for compatibility reasons and
994 * only if configured. This is considered legacy code, which
995 * will probably be removed at some point in the future.
997 if (!check_amd_hwpstate_cpu(0)) {
998 struct freq_attr **attr;
1000 pr_debug("CPB unsupported, do not expose it\n");
1002 for (attr = acpi_cpufreq_attr; *attr; attr++)
1003 if (*attr == &cpb) {
1009 acpi_cpufreq_boost_init();
1011 ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1013 free_acpi_perf_data();
1018 static void acpi_cpufreq_remove(struct platform_device *pdev)
1020 pr_debug("%s\n", __func__);
1022 cpufreq_unregister_driver(&acpi_cpufreq_driver);
1024 free_acpi_perf_data();
1027 static struct platform_driver acpi_cpufreq_platdrv = {
1029 .name = "acpi-cpufreq",
1031 .remove_new = acpi_cpufreq_remove,
1034 static int __init acpi_cpufreq_init(void)
1036 return platform_driver_probe(&acpi_cpufreq_platdrv, acpi_cpufreq_probe);
1039 static void __exit acpi_cpufreq_exit(void)
1041 platform_driver_unregister(&acpi_cpufreq_platdrv);
1044 module_param(acpi_pstate_strict, uint, 0644);
1045 MODULE_PARM_DESC(acpi_pstate_strict,
1046 "value 0 or non-zero. non-zero -> strict ACPI checks are "
1047 "performed during frequency changes.");
1049 late_initcall(acpi_cpufreq_init);
1050 module_exit(acpi_cpufreq_exit);
1052 MODULE_ALIAS("platform:acpi-cpufreq");