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>
37 #include <linux/acpi.h>
38 #include <acpi/processor.h>
42 #include <asm/processor.h>
43 #include <asm/cpufeature.h>
44 #include <asm/delay.h>
45 #include <asm/uaccess.h>
47 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, "acpi-cpufreq", msg)
49 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
50 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
51 MODULE_LICENSE("GPL");
54 UNDEFINED_CAPABLE = 0,
55 SYSTEM_INTEL_MSR_CAPABLE,
59 #define INTEL_MSR_RANGE (0xffff)
60 #define CPUID_6_ECX_APERFMPERF_CAPABILITY (0x1)
62 struct acpi_cpufreq_data {
63 struct acpi_processor_performance *acpi_data;
64 struct cpufreq_frequency_table *freq_table;
65 unsigned int max_freq;
67 unsigned int cpu_feature;
70 static struct acpi_cpufreq_data *drv_data[NR_CPUS];
71 static struct acpi_processor_performance *acpi_perf_data[NR_CPUS];
73 static struct cpufreq_driver acpi_cpufreq_driver;
75 static unsigned int acpi_pstate_strict;
77 static int check_est_cpu(unsigned int cpuid)
79 struct cpuinfo_x86 *cpu = &cpu_data[cpuid];
81 if (cpu->x86_vendor != X86_VENDOR_INTEL ||
82 !cpu_has(cpu, X86_FEATURE_EST))
88 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
90 struct acpi_processor_performance *perf;
93 perf = data->acpi_data;
95 for (i=0; i<perf->state_count; i++) {
96 if (value == perf->states[i].status)
97 return data->freq_table[i].frequency;
102 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
105 struct acpi_processor_performance *perf;
107 msr &= INTEL_MSR_RANGE;
108 perf = data->acpi_data;
110 for (i=0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
111 if (msr == perf->states[data->freq_table[i].index].status)
112 return data->freq_table[i].frequency;
114 return data->freq_table[0].frequency;
117 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
119 switch (data->cpu_feature) {
120 case SYSTEM_INTEL_MSR_CAPABLE:
121 return extract_msr(val, data);
122 case SYSTEM_IO_CAPABLE:
123 return extract_io(val, data);
150 static void do_drv_read(struct drv_cmd *cmd)
155 case SYSTEM_INTEL_MSR_CAPABLE:
156 rdmsr(cmd->addr.msr.reg, cmd->val, h);
158 case SYSTEM_IO_CAPABLE:
159 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
161 (u32)cmd->addr.io.bit_width);
168 static void do_drv_write(struct drv_cmd *cmd)
173 case SYSTEM_INTEL_MSR_CAPABLE:
174 wrmsr(cmd->addr.msr.reg, cmd->val, h);
176 case SYSTEM_IO_CAPABLE:
177 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
179 (u32)cmd->addr.io.bit_width);
186 static void drv_read(struct drv_cmd *cmd)
188 cpumask_t saved_mask = current->cpus_allowed;
191 set_cpus_allowed(current, cmd->mask);
193 set_cpus_allowed(current, saved_mask);
196 static void drv_write(struct drv_cmd *cmd)
198 cpumask_t saved_mask = current->cpus_allowed;
201 for_each_cpu_mask(i, cmd->mask) {
202 set_cpus_allowed(current, cpumask_of_cpu(i));
206 set_cpus_allowed(current, saved_mask);
210 static u32 get_cur_val(cpumask_t mask)
212 struct acpi_processor_performance *perf;
215 if (unlikely(cpus_empty(mask)))
218 switch (drv_data[first_cpu(mask)]->cpu_feature) {
219 case SYSTEM_INTEL_MSR_CAPABLE:
220 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
221 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
223 case SYSTEM_IO_CAPABLE:
224 cmd.type = SYSTEM_IO_CAPABLE;
225 perf = drv_data[first_cpu(mask)]->acpi_data;
226 cmd.addr.io.port = perf->control_register.address;
227 cmd.addr.io.bit_width = perf->control_register.bit_width;
237 dprintk("get_cur_val = %u\n", cmd.val);
243 * Return the measured active (C0) frequency on this CPU since last call
246 * Return: Average CPU frequency in terms of max frequency (zero on error)
248 * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
249 * over a period of time, while CPU is in C0 state.
250 * IA32_MPERF counts at the rate of max advertised frequency
251 * IA32_APERF counts at the rate of actual CPU frequency
252 * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
253 * no meaning should be associated with absolute values of these MSRs.
255 static unsigned int get_measured_perf(unsigned int cpu)
263 } aperf_cur, mperf_cur;
265 cpumask_t saved_mask;
266 unsigned int perf_percent;
269 saved_mask = current->cpus_allowed;
270 set_cpus_allowed(current, cpumask_of_cpu(cpu));
271 if (get_cpu() != cpu) {
272 /* We were not able to run on requested processor */
277 rdmsr(MSR_IA32_APERF, aperf_cur.split.lo, aperf_cur.split.hi);
278 rdmsr(MSR_IA32_MPERF, mperf_cur.split.lo, mperf_cur.split.hi);
280 wrmsr(MSR_IA32_APERF, 0,0);
281 wrmsr(MSR_IA32_MPERF, 0,0);
285 * We dont want to do 64 bit divide with 32 bit kernel
286 * Get an approximate value. Return failure in case we cannot get
287 * an approximate value.
289 if (unlikely(aperf_cur.split.hi || mperf_cur.split.hi)) {
293 h = max_t(u32, aperf_cur.split.hi, mperf_cur.split.hi);
294 shift_count = fls(h);
296 aperf_cur.whole >>= shift_count;
297 mperf_cur.whole >>= shift_count;
300 if (((unsigned long)(-1) / 100) < aperf_cur.split.lo) {
302 aperf_cur.split.lo >>= shift_count;
303 mperf_cur.split.lo >>= shift_count;
306 if (aperf_cur.split.lo && mperf_cur.split.lo)
307 perf_percent = (aperf_cur.split.lo * 100) / mperf_cur.split.lo;
312 if (unlikely(((unsigned long)(-1) / 100) < aperf_cur.whole)) {
314 aperf_cur.whole >>= shift_count;
315 mperf_cur.whole >>= shift_count;
318 if (aperf_cur.whole && mperf_cur.whole)
319 perf_percent = (aperf_cur.whole * 100) / mperf_cur.whole;
325 retval = drv_data[cpu]->max_freq * perf_percent / 100;
328 set_cpus_allowed(current, saved_mask);
330 dprintk("cpu %d: performance percent %d\n", cpu, perf_percent);
334 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
336 struct acpi_cpufreq_data *data = drv_data[cpu];
339 dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
341 if (unlikely(data == NULL ||
342 data->acpi_data == NULL || data->freq_table == NULL)) {
346 freq = extract_freq(get_cur_val(cpumask_of_cpu(cpu)), data);
347 dprintk("cur freq = %u\n", freq);
352 static unsigned int check_freqs(cpumask_t mask, unsigned int freq,
353 struct acpi_cpufreq_data *data)
355 unsigned int cur_freq;
358 for (i=0; i<100; i++) {
359 cur_freq = extract_freq(get_cur_val(mask), data);
360 if (cur_freq == freq)
367 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
368 unsigned int target_freq, unsigned int relation)
370 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
371 struct acpi_processor_performance *perf;
372 struct cpufreq_freqs freqs;
373 cpumask_t online_policy_cpus;
376 unsigned int next_state = 0; /* Index into freq_table */
377 unsigned int next_perf_state = 0; /* Index into perf table */
381 dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
383 if (unlikely(data == NULL ||
384 data->acpi_data == NULL || data->freq_table == NULL)) {
388 perf = data->acpi_data;
389 result = cpufreq_frequency_table_target(policy,
392 relation, &next_state);
393 if (unlikely(result))
396 #ifdef CONFIG_HOTPLUG_CPU
397 /* cpufreq holds the hotplug lock, so we are safe from here on */
398 cpus_and(online_policy_cpus, cpu_online_map, policy->cpus);
400 online_policy_cpus = policy->cpus;
403 next_perf_state = data->freq_table[next_state].index;
404 if (perf->state == next_perf_state) {
405 if (unlikely(data->resume)) {
406 dprintk("Called after resume, resetting to P%d\n",
410 dprintk("Already at target state (P%d)\n",
416 switch (data->cpu_feature) {
417 case SYSTEM_INTEL_MSR_CAPABLE:
418 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
419 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
421 (u32) perf->states[next_perf_state].
422 control & INTEL_MSR_RANGE;
423 cmd.val = get_cur_val(online_policy_cpus);
424 cmd.val = (cmd.val & ~INTEL_MSR_RANGE) | msr;
426 case SYSTEM_IO_CAPABLE:
427 cmd.type = SYSTEM_IO_CAPABLE;
428 cmd.addr.io.port = perf->control_register.address;
429 cmd.addr.io.bit_width = perf->control_register.bit_width;
430 cmd.val = (u32) perf->states[next_perf_state].control;
436 cpus_clear(cmd.mask);
438 if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
439 cmd.mask = online_policy_cpus;
441 cpu_set(policy->cpu, cmd.mask);
443 freqs.old = perf->states[perf->state].core_frequency * 1000;
444 freqs.new = data->freq_table[next_state].frequency;
445 for_each_cpu_mask(i, cmd.mask) {
447 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
452 if (acpi_pstate_strict) {
453 if (!check_freqs(cmd.mask, freqs.new, data)) {
454 dprintk("acpi_cpufreq_target failed (%d)\n",
460 for_each_cpu_mask(i, cmd.mask) {
462 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
464 perf->state = next_perf_state;
469 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
471 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
473 dprintk("acpi_cpufreq_verify\n");
475 return cpufreq_frequency_table_verify(policy, data->freq_table);
479 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
481 struct acpi_processor_performance *perf = data->acpi_data;
484 /* search the closest match to cpu_khz */
487 unsigned long freqn = perf->states[0].core_frequency * 1000;
489 for (i=0; i<(perf->state_count-1); i++) {
491 freqn = perf->states[i+1].core_frequency * 1000;
492 if ((2 * cpu_khz) > (freqn + freq)) {
497 perf->state = perf->state_count-1;
500 /* assume CPU is at P0... */
502 return perf->states[0].core_frequency * 1000;
507 * acpi_cpufreq_early_init - initialize ACPI P-States library
509 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
510 * in order to determine correct frequency and voltage pairings. We can
511 * do _PDC and _PSD and find out the processor dependency for the
512 * actual init that will happen later...
514 static int acpi_cpufreq_early_init(void)
516 struct acpi_processor_performance *data;
520 dprintk("acpi_cpufreq_early_init\n");
522 for_each_possible_cpu(i) {
523 data = kzalloc(sizeof(struct acpi_processor_performance),
526 for_each_cpu_mask(j, covered) {
527 kfree(acpi_perf_data[j]);
528 acpi_perf_data[j] = NULL;
532 acpi_perf_data[i] = data;
536 /* Do initialization in ACPI core */
537 acpi_processor_preregister_performance(acpi_perf_data);
543 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
544 * or do it in BIOS firmware and won't inform about it to OS. If not
545 * detected, this has a side effect of making CPU run at a different speed
546 * than OS intended it to run at. Detect it and handle it cleanly.
548 static int bios_with_sw_any_bug;
550 static int sw_any_bug_found(struct dmi_system_id *d)
552 bios_with_sw_any_bug = 1;
556 static struct dmi_system_id sw_any_bug_dmi_table[] = {
558 .callback = sw_any_bug_found,
559 .ident = "Supermicro Server X6DLP",
561 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
562 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
563 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
570 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
573 unsigned int valid_states = 0;
574 unsigned int cpu = policy->cpu;
575 struct acpi_cpufreq_data *data;
576 unsigned int result = 0;
577 struct cpuinfo_x86 *c = &cpu_data[policy->cpu];
578 struct acpi_processor_performance *perf;
580 dprintk("acpi_cpufreq_cpu_init\n");
582 if (!acpi_perf_data[cpu])
585 data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
589 data->acpi_data = acpi_perf_data[cpu];
590 drv_data[cpu] = data;
592 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
593 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
595 result = acpi_processor_register_performance(data->acpi_data, cpu);
599 perf = data->acpi_data;
600 policy->shared_type = perf->shared_type;
603 * Will let policy->cpus know about dependency only when software
604 * coordination is required.
606 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
607 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
608 policy->cpus = perf->shared_cpu_map;
612 dmi_check_system(sw_any_bug_dmi_table);
613 if (bios_with_sw_any_bug && cpus_weight(policy->cpus) == 1) {
614 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
615 policy->cpus = cpu_core_map[cpu];
619 /* capability check */
620 if (perf->state_count <= 1) {
621 dprintk("No P-States\n");
626 if (perf->control_register.space_id != perf->status_register.space_id) {
631 switch (perf->control_register.space_id) {
632 case ACPI_ADR_SPACE_SYSTEM_IO:
633 dprintk("SYSTEM IO addr space\n");
634 data->cpu_feature = SYSTEM_IO_CAPABLE;
636 case ACPI_ADR_SPACE_FIXED_HARDWARE:
637 dprintk("HARDWARE addr space\n");
638 if (!check_est_cpu(cpu)) {
642 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
645 dprintk("Unknown addr space %d\n",
646 (u32) (perf->control_register.space_id));
651 data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
652 (perf->state_count+1), GFP_KERNEL);
653 if (!data->freq_table) {
658 /* detect transition latency */
659 policy->cpuinfo.transition_latency = 0;
660 for (i=0; i<perf->state_count; i++) {
661 if ((perf->states[i].transition_latency * 1000) >
662 policy->cpuinfo.transition_latency)
663 policy->cpuinfo.transition_latency =
664 perf->states[i].transition_latency * 1000;
666 policy->governor = CPUFREQ_DEFAULT_GOVERNOR;
668 data->max_freq = perf->states[0].core_frequency * 1000;
670 for (i=0; i<perf->state_count; i++) {
671 if (i>0 && perf->states[i].core_frequency ==
672 perf->states[i-1].core_frequency)
675 data->freq_table[valid_states].index = i;
676 data->freq_table[valid_states].frequency =
677 perf->states[i].core_frequency * 1000;
680 data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
683 result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
687 switch (perf->control_register.space_id) {
688 case ACPI_ADR_SPACE_SYSTEM_IO:
689 /* Current speed is unknown and not detectable by IO port */
690 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
692 case ACPI_ADR_SPACE_FIXED_HARDWARE:
693 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
694 policy->cur = get_cur_freq_on_cpu(cpu);
700 /* notify BIOS that we exist */
701 acpi_processor_notify_smm(THIS_MODULE);
703 /* Check for APERF/MPERF support in hardware */
704 if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
707 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
708 acpi_cpufreq_driver.getavg = get_measured_perf;
711 dprintk("CPU%u - ACPI performance management activated.\n", cpu);
712 for (i = 0; i < perf->state_count; i++)
713 dprintk(" %cP%d: %d MHz, %d mW, %d uS\n",
714 (i == perf->state ? '*' : ' '), i,
715 (u32) perf->states[i].core_frequency,
716 (u32) perf->states[i].power,
717 (u32) perf->states[i].transition_latency);
719 cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
722 * the first call to ->target() should result in us actually
723 * writing something to the appropriate registers.
730 kfree(data->freq_table);
732 acpi_processor_unregister_performance(perf, cpu);
735 drv_data[cpu] = NULL;
740 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
742 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
744 dprintk("acpi_cpufreq_cpu_exit\n");
747 cpufreq_frequency_table_put_attr(policy->cpu);
748 drv_data[policy->cpu] = NULL;
749 acpi_processor_unregister_performance(data->acpi_data,
757 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
759 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
761 dprintk("acpi_cpufreq_resume\n");
768 static struct freq_attr *acpi_cpufreq_attr[] = {
769 &cpufreq_freq_attr_scaling_available_freqs,
773 static struct cpufreq_driver acpi_cpufreq_driver = {
774 .verify = acpi_cpufreq_verify,
775 .target = acpi_cpufreq_target,
776 .init = acpi_cpufreq_cpu_init,
777 .exit = acpi_cpufreq_cpu_exit,
778 .resume = acpi_cpufreq_resume,
779 .name = "acpi-cpufreq",
780 .owner = THIS_MODULE,
781 .attr = acpi_cpufreq_attr,
784 static int __init acpi_cpufreq_init(void)
786 dprintk("acpi_cpufreq_init\n");
788 acpi_cpufreq_early_init();
790 return cpufreq_register_driver(&acpi_cpufreq_driver);
793 static void __exit acpi_cpufreq_exit(void)
796 dprintk("acpi_cpufreq_exit\n");
798 cpufreq_unregister_driver(&acpi_cpufreq_driver);
800 for_each_possible_cpu(i) {
801 kfree(acpi_perf_data[i]);
802 acpi_perf_data[i] = NULL;
807 module_param(acpi_pstate_strict, uint, 0644);
808 MODULE_PARM_DESC(acpi_pstate_strict,
809 "value 0 or non-zero. non-zero -> strict ACPI checks are "
810 "performed during frequency changes.");
812 late_initcall(acpi_cpufreq_init);
813 module_exit(acpi_cpufreq_exit);
815 MODULE_ALIAS("acpi");