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 rdmsr(cmd->addr.msr.reg, lo, hi);
175 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
176 wrmsr(cmd->addr.msr.reg, lo, hi);
178 case SYSTEM_IO_CAPABLE:
179 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
181 (u32)cmd->addr.io.bit_width);
188 static void drv_read(struct drv_cmd *cmd)
190 cpumask_t saved_mask = current->cpus_allowed;
193 set_cpus_allowed(current, cmd->mask);
195 set_cpus_allowed(current, saved_mask);
198 static void drv_write(struct drv_cmd *cmd)
200 cpumask_t saved_mask = current->cpus_allowed;
203 for_each_cpu_mask(i, cmd->mask) {
204 set_cpus_allowed(current, cpumask_of_cpu(i));
208 set_cpus_allowed(current, saved_mask);
212 static u32 get_cur_val(cpumask_t mask)
214 struct acpi_processor_performance *perf;
217 if (unlikely(cpus_empty(mask)))
220 switch (drv_data[first_cpu(mask)]->cpu_feature) {
221 case SYSTEM_INTEL_MSR_CAPABLE:
222 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
223 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
225 case SYSTEM_IO_CAPABLE:
226 cmd.type = SYSTEM_IO_CAPABLE;
227 perf = drv_data[first_cpu(mask)]->acpi_data;
228 cmd.addr.io.port = perf->control_register.address;
229 cmd.addr.io.bit_width = perf->control_register.bit_width;
239 dprintk("get_cur_val = %u\n", cmd.val);
245 * Return the measured active (C0) frequency on this CPU since last call
248 * Return: Average CPU frequency in terms of max frequency (zero on error)
250 * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
251 * over a period of time, while CPU is in C0 state.
252 * IA32_MPERF counts at the rate of max advertised frequency
253 * IA32_APERF counts at the rate of actual CPU frequency
254 * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
255 * no meaning should be associated with absolute values of these MSRs.
257 static unsigned int get_measured_perf(unsigned int cpu)
265 } aperf_cur, mperf_cur;
267 cpumask_t saved_mask;
268 unsigned int perf_percent;
271 saved_mask = current->cpus_allowed;
272 set_cpus_allowed(current, cpumask_of_cpu(cpu));
273 if (get_cpu() != cpu) {
274 /* We were not able to run on requested processor */
279 rdmsr(MSR_IA32_APERF, aperf_cur.split.lo, aperf_cur.split.hi);
280 rdmsr(MSR_IA32_MPERF, mperf_cur.split.lo, mperf_cur.split.hi);
282 wrmsr(MSR_IA32_APERF, 0,0);
283 wrmsr(MSR_IA32_MPERF, 0,0);
287 * We dont want to do 64 bit divide with 32 bit kernel
288 * Get an approximate value. Return failure in case we cannot get
289 * an approximate value.
291 if (unlikely(aperf_cur.split.hi || mperf_cur.split.hi)) {
295 h = max_t(u32, aperf_cur.split.hi, mperf_cur.split.hi);
296 shift_count = fls(h);
298 aperf_cur.whole >>= shift_count;
299 mperf_cur.whole >>= shift_count;
302 if (((unsigned long)(-1) / 100) < aperf_cur.split.lo) {
304 aperf_cur.split.lo >>= shift_count;
305 mperf_cur.split.lo >>= shift_count;
308 if (aperf_cur.split.lo && mperf_cur.split.lo)
309 perf_percent = (aperf_cur.split.lo * 100) / mperf_cur.split.lo;
314 if (unlikely(((unsigned long)(-1) / 100) < aperf_cur.whole)) {
316 aperf_cur.whole >>= shift_count;
317 mperf_cur.whole >>= shift_count;
320 if (aperf_cur.whole && mperf_cur.whole)
321 perf_percent = (aperf_cur.whole * 100) / mperf_cur.whole;
327 retval = drv_data[cpu]->max_freq * perf_percent / 100;
330 set_cpus_allowed(current, saved_mask);
332 dprintk("cpu %d: performance percent %d\n", cpu, perf_percent);
336 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
338 struct acpi_cpufreq_data *data = drv_data[cpu];
341 dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
343 if (unlikely(data == NULL ||
344 data->acpi_data == NULL || data->freq_table == NULL)) {
348 freq = extract_freq(get_cur_val(cpumask_of_cpu(cpu)), data);
349 dprintk("cur freq = %u\n", freq);
354 static unsigned int check_freqs(cpumask_t mask, unsigned int freq,
355 struct acpi_cpufreq_data *data)
357 unsigned int cur_freq;
360 for (i=0; i<100; i++) {
361 cur_freq = extract_freq(get_cur_val(mask), data);
362 if (cur_freq == freq)
369 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
370 unsigned int target_freq, unsigned int relation)
372 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
373 struct acpi_processor_performance *perf;
374 struct cpufreq_freqs freqs;
375 cpumask_t online_policy_cpus;
377 unsigned int next_state = 0; /* Index into freq_table */
378 unsigned int next_perf_state = 0; /* Index into perf table */
382 dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
384 if (unlikely(data == NULL ||
385 data->acpi_data == NULL || data->freq_table == NULL)) {
389 perf = data->acpi_data;
390 result = cpufreq_frequency_table_target(policy,
393 relation, &next_state);
394 if (unlikely(result))
397 #ifdef CONFIG_HOTPLUG_CPU
398 /* cpufreq holds the hotplug lock, so we are safe from here on */
399 cpus_and(online_policy_cpus, cpu_online_map, policy->cpus);
401 online_policy_cpus = policy->cpus;
404 next_perf_state = data->freq_table[next_state].index;
405 if (perf->state == next_perf_state) {
406 if (unlikely(data->resume)) {
407 dprintk("Called after resume, resetting to P%d\n",
411 dprintk("Already at target state (P%d)\n",
417 switch (data->cpu_feature) {
418 case SYSTEM_INTEL_MSR_CAPABLE:
419 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
420 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
421 cmd.val = (u32) perf->states[next_perf_state].control;
423 case SYSTEM_IO_CAPABLE:
424 cmd.type = SYSTEM_IO_CAPABLE;
425 cmd.addr.io.port = perf->control_register.address;
426 cmd.addr.io.bit_width = perf->control_register.bit_width;
427 cmd.val = (u32) perf->states[next_perf_state].control;
433 cpus_clear(cmd.mask);
435 if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
436 cmd.mask = online_policy_cpus;
438 cpu_set(policy->cpu, cmd.mask);
440 freqs.old = perf->states[perf->state].core_frequency * 1000;
441 freqs.new = data->freq_table[next_state].frequency;
442 for_each_cpu_mask(i, cmd.mask) {
444 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
449 if (acpi_pstate_strict) {
450 if (!check_freqs(cmd.mask, freqs.new, data)) {
451 dprintk("acpi_cpufreq_target failed (%d)\n",
457 for_each_cpu_mask(i, cmd.mask) {
459 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
461 perf->state = next_perf_state;
466 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
468 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
470 dprintk("acpi_cpufreq_verify\n");
472 return cpufreq_frequency_table_verify(policy, data->freq_table);
476 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
478 struct acpi_processor_performance *perf = data->acpi_data;
481 /* search the closest match to cpu_khz */
484 unsigned long freqn = perf->states[0].core_frequency * 1000;
486 for (i=0; i<(perf->state_count-1); i++) {
488 freqn = perf->states[i+1].core_frequency * 1000;
489 if ((2 * cpu_khz) > (freqn + freq)) {
494 perf->state = perf->state_count-1;
497 /* assume CPU is at P0... */
499 return perf->states[0].core_frequency * 1000;
504 * acpi_cpufreq_early_init - initialize ACPI P-States library
506 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
507 * in order to determine correct frequency and voltage pairings. We can
508 * do _PDC and _PSD and find out the processor dependency for the
509 * actual init that will happen later...
511 static int acpi_cpufreq_early_init(void)
513 struct acpi_processor_performance *data;
517 dprintk("acpi_cpufreq_early_init\n");
519 for_each_possible_cpu(i) {
520 data = kzalloc(sizeof(struct acpi_processor_performance),
523 for_each_cpu_mask(j, covered) {
524 kfree(acpi_perf_data[j]);
525 acpi_perf_data[j] = NULL;
529 acpi_perf_data[i] = data;
533 /* Do initialization in ACPI core */
534 acpi_processor_preregister_performance(acpi_perf_data);
540 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
541 * or do it in BIOS firmware and won't inform about it to OS. If not
542 * detected, this has a side effect of making CPU run at a different speed
543 * than OS intended it to run at. Detect it and handle it cleanly.
545 static int bios_with_sw_any_bug;
547 static int sw_any_bug_found(struct dmi_system_id *d)
549 bios_with_sw_any_bug = 1;
553 static struct dmi_system_id sw_any_bug_dmi_table[] = {
555 .callback = sw_any_bug_found,
556 .ident = "Supermicro Server X6DLP",
558 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
559 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
560 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
567 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
570 unsigned int valid_states = 0;
571 unsigned int cpu = policy->cpu;
572 struct acpi_cpufreq_data *data;
573 unsigned int result = 0;
574 struct cpuinfo_x86 *c = &cpu_data[policy->cpu];
575 struct acpi_processor_performance *perf;
577 dprintk("acpi_cpufreq_cpu_init\n");
579 if (!acpi_perf_data[cpu])
582 data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
586 data->acpi_data = acpi_perf_data[cpu];
587 drv_data[cpu] = data;
589 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
590 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
592 result = acpi_processor_register_performance(data->acpi_data, cpu);
596 perf = data->acpi_data;
597 policy->shared_type = perf->shared_type;
600 * Will let policy->cpus know about dependency only when software
601 * coordination is required.
603 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
604 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
605 policy->cpus = perf->shared_cpu_map;
609 dmi_check_system(sw_any_bug_dmi_table);
610 if (bios_with_sw_any_bug && cpus_weight(policy->cpus) == 1) {
611 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
612 policy->cpus = cpu_core_map[cpu];
616 /* capability check */
617 if (perf->state_count <= 1) {
618 dprintk("No P-States\n");
623 if (perf->control_register.space_id != perf->status_register.space_id) {
628 switch (perf->control_register.space_id) {
629 case ACPI_ADR_SPACE_SYSTEM_IO:
630 dprintk("SYSTEM IO addr space\n");
631 data->cpu_feature = SYSTEM_IO_CAPABLE;
633 case ACPI_ADR_SPACE_FIXED_HARDWARE:
634 dprintk("HARDWARE addr space\n");
635 if (!check_est_cpu(cpu)) {
639 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
642 dprintk("Unknown addr space %d\n",
643 (u32) (perf->control_register.space_id));
648 data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
649 (perf->state_count+1), GFP_KERNEL);
650 if (!data->freq_table) {
655 /* detect transition latency */
656 policy->cpuinfo.transition_latency = 0;
657 for (i=0; i<perf->state_count; i++) {
658 if ((perf->states[i].transition_latency * 1000) >
659 policy->cpuinfo.transition_latency)
660 policy->cpuinfo.transition_latency =
661 perf->states[i].transition_latency * 1000;
663 policy->governor = CPUFREQ_DEFAULT_GOVERNOR;
665 data->max_freq = perf->states[0].core_frequency * 1000;
667 for (i=0; i<perf->state_count; i++) {
668 if (i>0 && perf->states[i].core_frequency ==
669 perf->states[i-1].core_frequency)
672 data->freq_table[valid_states].index = i;
673 data->freq_table[valid_states].frequency =
674 perf->states[i].core_frequency * 1000;
677 data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
680 result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
684 switch (perf->control_register.space_id) {
685 case ACPI_ADR_SPACE_SYSTEM_IO:
686 /* Current speed is unknown and not detectable by IO port */
687 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
689 case ACPI_ADR_SPACE_FIXED_HARDWARE:
690 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
691 policy->cur = get_cur_freq_on_cpu(cpu);
697 /* notify BIOS that we exist */
698 acpi_processor_notify_smm(THIS_MODULE);
700 /* Check for APERF/MPERF support in hardware */
701 if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
704 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
705 acpi_cpufreq_driver.getavg = get_measured_perf;
708 dprintk("CPU%u - ACPI performance management activated.\n", cpu);
709 for (i = 0; i < perf->state_count; i++)
710 dprintk(" %cP%d: %d MHz, %d mW, %d uS\n",
711 (i == perf->state ? '*' : ' '), i,
712 (u32) perf->states[i].core_frequency,
713 (u32) perf->states[i].power,
714 (u32) perf->states[i].transition_latency);
716 cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
719 * the first call to ->target() should result in us actually
720 * writing something to the appropriate registers.
727 kfree(data->freq_table);
729 acpi_processor_unregister_performance(perf, cpu);
732 drv_data[cpu] = NULL;
737 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
739 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
741 dprintk("acpi_cpufreq_cpu_exit\n");
744 cpufreq_frequency_table_put_attr(policy->cpu);
745 drv_data[policy->cpu] = NULL;
746 acpi_processor_unregister_performance(data->acpi_data,
754 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
756 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
758 dprintk("acpi_cpufreq_resume\n");
765 static struct freq_attr *acpi_cpufreq_attr[] = {
766 &cpufreq_freq_attr_scaling_available_freqs,
770 static struct cpufreq_driver acpi_cpufreq_driver = {
771 .verify = acpi_cpufreq_verify,
772 .target = acpi_cpufreq_target,
773 .init = acpi_cpufreq_cpu_init,
774 .exit = acpi_cpufreq_cpu_exit,
775 .resume = acpi_cpufreq_resume,
776 .name = "acpi-cpufreq",
777 .owner = THIS_MODULE,
778 .attr = acpi_cpufreq_attr,
781 static int __init acpi_cpufreq_init(void)
783 dprintk("acpi_cpufreq_init\n");
785 acpi_cpufreq_early_init();
787 return cpufreq_register_driver(&acpi_cpufreq_driver);
790 static void __exit acpi_cpufreq_exit(void)
793 dprintk("acpi_cpufreq_exit\n");
795 cpufreq_unregister_driver(&acpi_cpufreq_driver);
797 for_each_possible_cpu(i) {
798 kfree(acpi_perf_data[i]);
799 acpi_perf_data[i] = NULL;
804 module_param(acpi_pstate_strict, uint, 0644);
805 MODULE_PARM_DESC(acpi_pstate_strict,
806 "value 0 or non-zero. non-zero -> strict ACPI checks are "
807 "performed during frequency changes.");
809 late_initcall(acpi_cpufreq_init);
810 module_exit(acpi_cpufreq_exit);
812 MODULE_ALIAS("acpi");