2 * acpi-cpufreq.c - ACPI Processor P-States Driver ($Revision: 1.4 $)
4 * Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
5 * Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
6 * Copyright (C) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
7 * Copyright (C) 2006 Denis Sadykov <denis.m.sadykov@intel.com>
9 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2 of the License, or (at
14 * your option) any later version.
16 * This program is distributed in the hope that it will be useful, but
17 * WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * General Public License for more details.
21 * You should have received a copy of the GNU General Public License along
22 * with this program; if not, write to the Free Software Foundation, Inc.,
23 * 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
25 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
28 #include <linux/kernel.h>
29 #include <linux/module.h>
30 #include <linux/init.h>
31 #include <linux/smp.h>
32 #include <linux/sched.h>
33 #include <linux/cpufreq.h>
34 #include <linux/compiler.h>
35 #include <linux/dmi.h>
36 #include <linux/ftrace.h>
38 #include <linux/acpi.h>
39 #include <acpi/processor.h>
43 #include <asm/processor.h>
44 #include <asm/cpufeature.h>
45 #include <asm/delay.h>
46 #include <asm/uaccess.h>
48 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, "acpi-cpufreq", msg)
50 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
51 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
52 MODULE_LICENSE("GPL");
55 UNDEFINED_CAPABLE = 0,
56 SYSTEM_INTEL_MSR_CAPABLE,
60 #define INTEL_MSR_RANGE (0xffff)
61 #define CPUID_6_ECX_APERFMPERF_CAPABILITY (0x1)
63 struct acpi_cpufreq_data {
64 struct acpi_processor_performance *acpi_data;
65 struct cpufreq_frequency_table *freq_table;
66 unsigned int max_freq;
68 unsigned int cpu_feature;
71 static DEFINE_PER_CPU(struct acpi_cpufreq_data *, drv_data);
73 /* acpi_perf_data is a pointer to percpu data. */
74 static struct acpi_processor_performance *acpi_perf_data;
76 static struct cpufreq_driver acpi_cpufreq_driver;
78 static unsigned int acpi_pstate_strict;
80 static int check_est_cpu(unsigned int cpuid)
82 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
84 if (cpu->x86_vendor != X86_VENDOR_INTEL ||
85 !cpu_has(cpu, X86_FEATURE_EST))
91 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
93 struct acpi_processor_performance *perf;
96 perf = data->acpi_data;
98 for (i=0; i<perf->state_count; i++) {
99 if (value == perf->states[i].status)
100 return data->freq_table[i].frequency;
105 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
108 struct acpi_processor_performance *perf;
110 msr &= INTEL_MSR_RANGE;
111 perf = data->acpi_data;
113 for (i=0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
114 if (msr == perf->states[data->freq_table[i].index].status)
115 return data->freq_table[i].frequency;
117 return data->freq_table[0].frequency;
120 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
122 switch (data->cpu_feature) {
123 case SYSTEM_INTEL_MSR_CAPABLE:
124 return extract_msr(val, data);
125 case SYSTEM_IO_CAPABLE:
126 return extract_io(val, data);
153 static void do_drv_read(struct drv_cmd *cmd)
158 case SYSTEM_INTEL_MSR_CAPABLE:
159 rdmsr(cmd->addr.msr.reg, cmd->val, h);
161 case SYSTEM_IO_CAPABLE:
162 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
164 (u32)cmd->addr.io.bit_width);
171 static void do_drv_write(struct drv_cmd *cmd)
176 case SYSTEM_INTEL_MSR_CAPABLE:
177 rdmsr(cmd->addr.msr.reg, lo, hi);
178 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
179 wrmsr(cmd->addr.msr.reg, lo, hi);
181 case SYSTEM_IO_CAPABLE:
182 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
184 (u32)cmd->addr.io.bit_width);
191 static void drv_read(struct drv_cmd *cmd)
193 cpumask_t saved_mask = current->cpus_allowed;
196 set_cpus_allowed_ptr(current, &cmd->mask);
198 set_cpus_allowed_ptr(current, &saved_mask);
201 static void drv_write(struct drv_cmd *cmd)
203 cpumask_t saved_mask = current->cpus_allowed;
206 for_each_cpu_mask_nr(i, cmd->mask) {
207 set_cpus_allowed_ptr(current, &cpumask_of_cpu(i));
211 set_cpus_allowed_ptr(current, &saved_mask);
215 static u32 get_cur_val(const cpumask_t *mask)
217 struct acpi_processor_performance *perf;
220 if (unlikely(cpus_empty(*mask)))
223 switch (per_cpu(drv_data, first_cpu(*mask))->cpu_feature) {
224 case SYSTEM_INTEL_MSR_CAPABLE:
225 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
226 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
228 case SYSTEM_IO_CAPABLE:
229 cmd.type = SYSTEM_IO_CAPABLE;
230 perf = per_cpu(drv_data, first_cpu(*mask))->acpi_data;
231 cmd.addr.io.port = perf->control_register.address;
232 cmd.addr.io.bit_width = perf->control_register.bit_width;
242 dprintk("get_cur_val = %u\n", cmd.val);
248 * Return the measured active (C0) frequency on this CPU since last call
251 * Return: Average CPU frequency in terms of max frequency (zero on error)
253 * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
254 * over a period of time, while CPU is in C0 state.
255 * IA32_MPERF counts at the rate of max advertised frequency
256 * IA32_APERF counts at the rate of actual CPU frequency
257 * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
258 * no meaning should be associated with absolute values of these MSRs.
260 static unsigned int get_measured_perf(struct cpufreq_policy *policy,
269 } aperf_cur, mperf_cur;
271 cpumask_t saved_mask;
272 unsigned int perf_percent;
275 saved_mask = current->cpus_allowed;
276 set_cpus_allowed_ptr(current, &cpumask_of_cpu(cpu));
277 if (get_cpu() != cpu) {
278 /* We were not able to run on requested processor */
283 rdmsr(MSR_IA32_APERF, aperf_cur.split.lo, aperf_cur.split.hi);
284 rdmsr(MSR_IA32_MPERF, mperf_cur.split.lo, mperf_cur.split.hi);
286 wrmsr(MSR_IA32_APERF, 0,0);
287 wrmsr(MSR_IA32_MPERF, 0,0);
291 * We dont want to do 64 bit divide with 32 bit kernel
292 * Get an approximate value. Return failure in case we cannot get
293 * an approximate value.
295 if (unlikely(aperf_cur.split.hi || mperf_cur.split.hi)) {
299 h = max_t(u32, aperf_cur.split.hi, mperf_cur.split.hi);
300 shift_count = fls(h);
302 aperf_cur.whole >>= shift_count;
303 mperf_cur.whole >>= shift_count;
306 if (((unsigned long)(-1) / 100) < aperf_cur.split.lo) {
308 aperf_cur.split.lo >>= shift_count;
309 mperf_cur.split.lo >>= shift_count;
312 if (aperf_cur.split.lo && mperf_cur.split.lo)
313 perf_percent = (aperf_cur.split.lo * 100) / mperf_cur.split.lo;
318 if (unlikely(((unsigned long)(-1) / 100) < aperf_cur.whole)) {
320 aperf_cur.whole >>= shift_count;
321 mperf_cur.whole >>= shift_count;
324 if (aperf_cur.whole && mperf_cur.whole)
325 perf_percent = (aperf_cur.whole * 100) / mperf_cur.whole;
331 retval = per_cpu(drv_data, policy->cpu)->max_freq * perf_percent / 100;
334 set_cpus_allowed_ptr(current, &saved_mask);
336 dprintk("cpu %d: performance percent %d\n", cpu, perf_percent);
340 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
342 struct acpi_cpufreq_data *data = per_cpu(drv_data, cpu);
344 unsigned int cached_freq;
346 dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
348 if (unlikely(data == NULL ||
349 data->acpi_data == NULL || data->freq_table == NULL)) {
353 cached_freq = data->freq_table[data->acpi_data->state].frequency;
354 freq = extract_freq(get_cur_val(&cpumask_of_cpu(cpu)), data);
355 if (freq != cached_freq) {
357 * The dreaded BIOS frequency change behind our back.
358 * Force set the frequency on next target call.
363 dprintk("cur freq = %u\n", freq);
368 static unsigned int check_freqs(const cpumask_t *mask, unsigned int freq,
369 struct acpi_cpufreq_data *data)
371 unsigned int cur_freq;
374 for (i=0; i<100; i++) {
375 cur_freq = extract_freq(get_cur_val(mask), data);
376 if (cur_freq == freq)
383 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
384 unsigned int target_freq, unsigned int relation)
386 struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
387 struct acpi_processor_performance *perf;
388 struct cpufreq_freqs freqs;
389 cpumask_t online_policy_cpus;
391 unsigned int next_state = 0; /* Index into freq_table */
392 unsigned int next_perf_state = 0; /* Index into perf table */
395 struct power_trace it;
397 dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
399 if (unlikely(data == NULL ||
400 data->acpi_data == NULL || data->freq_table == NULL)) {
404 perf = data->acpi_data;
405 result = cpufreq_frequency_table_target(policy,
408 relation, &next_state);
409 if (unlikely(result))
412 #ifdef CONFIG_HOTPLUG_CPU
413 /* cpufreq holds the hotplug lock, so we are safe from here on */
414 cpus_and(online_policy_cpus, cpu_online_map, policy->cpus);
416 online_policy_cpus = policy->cpus;
419 next_perf_state = data->freq_table[next_state].index;
420 if (perf->state == next_perf_state) {
421 if (unlikely(data->resume)) {
422 dprintk("Called after resume, resetting to P%d\n",
426 dprintk("Already at target state (P%d)\n",
432 trace_power_mark(&it, POWER_PSTATE, next_perf_state);
434 switch (data->cpu_feature) {
435 case SYSTEM_INTEL_MSR_CAPABLE:
436 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
437 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
438 cmd.val = (u32) perf->states[next_perf_state].control;
440 case SYSTEM_IO_CAPABLE:
441 cmd.type = SYSTEM_IO_CAPABLE;
442 cmd.addr.io.port = perf->control_register.address;
443 cmd.addr.io.bit_width = perf->control_register.bit_width;
444 cmd.val = (u32) perf->states[next_perf_state].control;
450 cpus_clear(cmd.mask);
452 if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
453 cmd.mask = online_policy_cpus;
455 cpu_set(policy->cpu, cmd.mask);
457 freqs.old = perf->states[perf->state].core_frequency * 1000;
458 freqs.new = data->freq_table[next_state].frequency;
459 for_each_cpu_mask_nr(i, cmd.mask) {
461 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
466 if (acpi_pstate_strict) {
467 if (!check_freqs(&cmd.mask, freqs.new, data)) {
468 dprintk("acpi_cpufreq_target failed (%d)\n",
474 for_each_cpu_mask_nr(i, cmd.mask) {
476 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
478 perf->state = next_perf_state;
483 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
485 struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
487 dprintk("acpi_cpufreq_verify\n");
489 return cpufreq_frequency_table_verify(policy, data->freq_table);
493 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
495 struct acpi_processor_performance *perf = data->acpi_data;
498 /* search the closest match to cpu_khz */
501 unsigned long freqn = perf->states[0].core_frequency * 1000;
503 for (i=0; i<(perf->state_count-1); i++) {
505 freqn = perf->states[i+1].core_frequency * 1000;
506 if ((2 * cpu_khz) > (freqn + freq)) {
511 perf->state = perf->state_count-1;
514 /* assume CPU is at P0... */
516 return perf->states[0].core_frequency * 1000;
520 static void free_acpi_perf_data(void)
524 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
525 for_each_possible_cpu(i)
526 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
528 free_percpu(acpi_perf_data);
532 * acpi_cpufreq_early_init - initialize ACPI P-States library
534 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
535 * in order to determine correct frequency and voltage pairings. We can
536 * do _PDC and _PSD and find out the processor dependency for the
537 * actual init that will happen later...
539 static int __init acpi_cpufreq_early_init(void)
542 dprintk("acpi_cpufreq_early_init\n");
544 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
545 if (!acpi_perf_data) {
546 dprintk("Memory allocation error for acpi_perf_data.\n");
549 for_each_possible_cpu(i) {
550 if (!alloc_cpumask_var(&per_cpu_ptr(acpi_perf_data, i)
551 ->shared_cpu_map, GFP_KERNEL)) {
553 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
554 free_acpi_perf_data();
559 /* Do initialization in ACPI core */
560 acpi_processor_preregister_performance(acpi_perf_data);
566 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
567 * or do it in BIOS firmware and won't inform about it to OS. If not
568 * detected, this has a side effect of making CPU run at a different speed
569 * than OS intended it to run at. Detect it and handle it cleanly.
571 static int bios_with_sw_any_bug;
573 static int sw_any_bug_found(const struct dmi_system_id *d)
575 bios_with_sw_any_bug = 1;
579 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
581 .callback = sw_any_bug_found,
582 .ident = "Supermicro Server X6DLP",
584 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
585 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
586 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
593 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
596 unsigned int valid_states = 0;
597 unsigned int cpu = policy->cpu;
598 struct acpi_cpufreq_data *data;
599 unsigned int result = 0;
600 struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
601 struct acpi_processor_performance *perf;
603 dprintk("acpi_cpufreq_cpu_init\n");
605 data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
609 data->acpi_data = percpu_ptr(acpi_perf_data, cpu);
610 per_cpu(drv_data, cpu) = data;
612 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
613 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
615 result = acpi_processor_register_performance(data->acpi_data, cpu);
619 perf = data->acpi_data;
620 policy->shared_type = perf->shared_type;
623 * Will let policy->cpus know about dependency only when software
624 * coordination is required.
626 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
627 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
628 cpumask_copy(&policy->cpus, perf->shared_cpu_map);
630 cpumask_copy(&policy->related_cpus, perf->shared_cpu_map);
633 dmi_check_system(sw_any_bug_dmi_table);
634 if (bios_with_sw_any_bug && cpus_weight(policy->cpus) == 1) {
635 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
636 policy->cpus = per_cpu(cpu_core_map, cpu);
640 /* capability check */
641 if (perf->state_count <= 1) {
642 dprintk("No P-States\n");
647 if (perf->control_register.space_id != perf->status_register.space_id) {
652 switch (perf->control_register.space_id) {
653 case ACPI_ADR_SPACE_SYSTEM_IO:
654 dprintk("SYSTEM IO addr space\n");
655 data->cpu_feature = SYSTEM_IO_CAPABLE;
657 case ACPI_ADR_SPACE_FIXED_HARDWARE:
658 dprintk("HARDWARE addr space\n");
659 if (!check_est_cpu(cpu)) {
663 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
666 dprintk("Unknown addr space %d\n",
667 (u32) (perf->control_register.space_id));
672 data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
673 (perf->state_count+1), GFP_KERNEL);
674 if (!data->freq_table) {
679 /* detect transition latency */
680 policy->cpuinfo.transition_latency = 0;
681 for (i=0; i<perf->state_count; i++) {
682 if ((perf->states[i].transition_latency * 1000) >
683 policy->cpuinfo.transition_latency)
684 policy->cpuinfo.transition_latency =
685 perf->states[i].transition_latency * 1000;
688 data->max_freq = perf->states[0].core_frequency * 1000;
690 for (i=0; i<perf->state_count; i++) {
691 if (i>0 && perf->states[i].core_frequency >=
692 data->freq_table[valid_states-1].frequency / 1000)
695 data->freq_table[valid_states].index = i;
696 data->freq_table[valid_states].frequency =
697 perf->states[i].core_frequency * 1000;
700 data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
703 result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
707 switch (perf->control_register.space_id) {
708 case ACPI_ADR_SPACE_SYSTEM_IO:
709 /* Current speed is unknown and not detectable by IO port */
710 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
712 case ACPI_ADR_SPACE_FIXED_HARDWARE:
713 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
714 policy->cur = get_cur_freq_on_cpu(cpu);
720 /* notify BIOS that we exist */
721 acpi_processor_notify_smm(THIS_MODULE);
723 /* Check for APERF/MPERF support in hardware */
724 if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
727 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
728 acpi_cpufreq_driver.getavg = get_measured_perf;
731 dprintk("CPU%u - ACPI performance management activated.\n", cpu);
732 for (i = 0; i < perf->state_count; i++)
733 dprintk(" %cP%d: %d MHz, %d mW, %d uS\n",
734 (i == perf->state ? '*' : ' '), i,
735 (u32) perf->states[i].core_frequency,
736 (u32) perf->states[i].power,
737 (u32) perf->states[i].transition_latency);
739 cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
742 * the first call to ->target() should result in us actually
743 * writing something to the appropriate registers.
750 kfree(data->freq_table);
752 acpi_processor_unregister_performance(perf, cpu);
755 per_cpu(drv_data, cpu) = NULL;
760 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
762 struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
764 dprintk("acpi_cpufreq_cpu_exit\n");
767 cpufreq_frequency_table_put_attr(policy->cpu);
768 per_cpu(drv_data, policy->cpu) = NULL;
769 acpi_processor_unregister_performance(data->acpi_data,
777 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
779 struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
781 dprintk("acpi_cpufreq_resume\n");
788 static struct freq_attr *acpi_cpufreq_attr[] = {
789 &cpufreq_freq_attr_scaling_available_freqs,
793 static struct cpufreq_driver acpi_cpufreq_driver = {
794 .verify = acpi_cpufreq_verify,
795 .target = acpi_cpufreq_target,
796 .init = acpi_cpufreq_cpu_init,
797 .exit = acpi_cpufreq_cpu_exit,
798 .resume = acpi_cpufreq_resume,
799 .name = "acpi-cpufreq",
800 .owner = THIS_MODULE,
801 .attr = acpi_cpufreq_attr,
804 static int __init acpi_cpufreq_init(void)
811 dprintk("acpi_cpufreq_init\n");
813 ret = acpi_cpufreq_early_init();
817 ret = cpufreq_register_driver(&acpi_cpufreq_driver);
819 free_acpi_perf_data();
824 static void __exit acpi_cpufreq_exit(void)
826 dprintk("acpi_cpufreq_exit\n");
828 cpufreq_unregister_driver(&acpi_cpufreq_driver);
830 free_percpu(acpi_perf_data);
833 module_param(acpi_pstate_strict, uint, 0644);
834 MODULE_PARM_DESC(acpi_pstate_strict,
835 "value 0 or non-zero. non-zero -> strict ACPI checks are "
836 "performed during frequency changes.");
838 late_initcall(acpi_cpufreq_init);
839 module_exit(acpi_cpufreq_exit);
841 MODULE_ALIAS("acpi");