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 /* acpi_perf_data is a pointer to percpu data. */
72 static struct acpi_processor_performance *acpi_perf_data;
74 static struct cpufreq_driver acpi_cpufreq_driver;
76 static unsigned int acpi_pstate_strict;
78 static int check_est_cpu(unsigned int cpuid)
80 struct cpuinfo_x86 *cpu = &cpu_data[cpuid];
82 if (cpu->x86_vendor != X86_VENDOR_INTEL ||
83 !cpu_has(cpu, X86_FEATURE_EST))
89 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
91 struct acpi_processor_performance *perf;
94 perf = data->acpi_data;
96 for (i=0; i<perf->state_count; i++) {
97 if (value == perf->states[i].status)
98 return data->freq_table[i].frequency;
103 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
106 struct acpi_processor_performance *perf;
108 msr &= INTEL_MSR_RANGE;
109 perf = data->acpi_data;
111 for (i=0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
112 if (msr == perf->states[data->freq_table[i].index].status)
113 return data->freq_table[i].frequency;
115 return data->freq_table[0].frequency;
118 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
120 switch (data->cpu_feature) {
121 case SYSTEM_INTEL_MSR_CAPABLE:
122 return extract_msr(val, data);
123 case SYSTEM_IO_CAPABLE:
124 return extract_io(val, data);
151 static void do_drv_read(struct drv_cmd *cmd)
156 case SYSTEM_INTEL_MSR_CAPABLE:
157 rdmsr(cmd->addr.msr.reg, cmd->val, h);
159 case SYSTEM_IO_CAPABLE:
160 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
162 (u32)cmd->addr.io.bit_width);
169 static void do_drv_write(struct drv_cmd *cmd)
174 case SYSTEM_INTEL_MSR_CAPABLE:
175 rdmsr(cmd->addr.msr.reg, lo, hi);
176 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
177 wrmsr(cmd->addr.msr.reg, lo, hi);
179 case SYSTEM_IO_CAPABLE:
180 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
182 (u32)cmd->addr.io.bit_width);
189 static void drv_read(struct drv_cmd *cmd)
191 cpumask_t saved_mask = current->cpus_allowed;
194 set_cpus_allowed(current, cmd->mask);
196 set_cpus_allowed(current, saved_mask);
199 static void drv_write(struct drv_cmd *cmd)
201 cpumask_t saved_mask = current->cpus_allowed;
204 for_each_cpu_mask(i, cmd->mask) {
205 set_cpus_allowed(current, cpumask_of_cpu(i));
209 set_cpus_allowed(current, saved_mask);
213 static u32 get_cur_val(cpumask_t mask)
215 struct acpi_processor_performance *perf;
218 if (unlikely(cpus_empty(mask)))
221 switch (drv_data[first_cpu(mask)]->cpu_feature) {
222 case SYSTEM_INTEL_MSR_CAPABLE:
223 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
224 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
226 case SYSTEM_IO_CAPABLE:
227 cmd.type = SYSTEM_IO_CAPABLE;
228 perf = drv_data[first_cpu(mask)]->acpi_data;
229 cmd.addr.io.port = perf->control_register.address;
230 cmd.addr.io.bit_width = perf->control_register.bit_width;
240 dprintk("get_cur_val = %u\n", cmd.val);
246 * Return the measured active (C0) frequency on this CPU since last call
249 * Return: Average CPU frequency in terms of max frequency (zero on error)
251 * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
252 * over a period of time, while CPU is in C0 state.
253 * IA32_MPERF counts at the rate of max advertised frequency
254 * IA32_APERF counts at the rate of actual CPU frequency
255 * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
256 * no meaning should be associated with absolute values of these MSRs.
258 static unsigned int get_measured_perf(unsigned int cpu)
266 } aperf_cur, mperf_cur;
268 cpumask_t saved_mask;
269 unsigned int perf_percent;
272 saved_mask = current->cpus_allowed;
273 set_cpus_allowed(current, cpumask_of_cpu(cpu));
274 if (get_cpu() != cpu) {
275 /* We were not able to run on requested processor */
280 rdmsr(MSR_IA32_APERF, aperf_cur.split.lo, aperf_cur.split.hi);
281 rdmsr(MSR_IA32_MPERF, mperf_cur.split.lo, mperf_cur.split.hi);
283 wrmsr(MSR_IA32_APERF, 0,0);
284 wrmsr(MSR_IA32_MPERF, 0,0);
288 * We dont want to do 64 bit divide with 32 bit kernel
289 * Get an approximate value. Return failure in case we cannot get
290 * an approximate value.
292 if (unlikely(aperf_cur.split.hi || mperf_cur.split.hi)) {
296 h = max_t(u32, aperf_cur.split.hi, mperf_cur.split.hi);
297 shift_count = fls(h);
299 aperf_cur.whole >>= shift_count;
300 mperf_cur.whole >>= shift_count;
303 if (((unsigned long)(-1) / 100) < aperf_cur.split.lo) {
305 aperf_cur.split.lo >>= shift_count;
306 mperf_cur.split.lo >>= shift_count;
309 if (aperf_cur.split.lo && mperf_cur.split.lo)
310 perf_percent = (aperf_cur.split.lo * 100) / mperf_cur.split.lo;
315 if (unlikely(((unsigned long)(-1) / 100) < aperf_cur.whole)) {
317 aperf_cur.whole >>= shift_count;
318 mperf_cur.whole >>= shift_count;
321 if (aperf_cur.whole && mperf_cur.whole)
322 perf_percent = (aperf_cur.whole * 100) / mperf_cur.whole;
328 retval = drv_data[cpu]->max_freq * perf_percent / 100;
331 set_cpus_allowed(current, saved_mask);
333 dprintk("cpu %d: performance percent %d\n", cpu, perf_percent);
337 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
339 struct acpi_cpufreq_data *data = drv_data[cpu];
342 dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
344 if (unlikely(data == NULL ||
345 data->acpi_data == NULL || data->freq_table == NULL)) {
349 freq = extract_freq(get_cur_val(cpumask_of_cpu(cpu)), data);
350 dprintk("cur freq = %u\n", freq);
355 static unsigned int check_freqs(cpumask_t mask, unsigned int freq,
356 struct acpi_cpufreq_data *data)
358 unsigned int cur_freq;
361 for (i=0; i<100; i++) {
362 cur_freq = extract_freq(get_cur_val(mask), data);
363 if (cur_freq == freq)
370 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
371 unsigned int target_freq, unsigned int relation)
373 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
374 struct acpi_processor_performance *perf;
375 struct cpufreq_freqs freqs;
376 cpumask_t online_policy_cpus;
378 unsigned int next_state = 0; /* Index into freq_table */
379 unsigned int next_perf_state = 0; /* Index into perf table */
383 dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
385 if (unlikely(data == NULL ||
386 data->acpi_data == NULL || data->freq_table == NULL)) {
390 perf = data->acpi_data;
391 result = cpufreq_frequency_table_target(policy,
394 relation, &next_state);
395 if (unlikely(result))
398 #ifdef CONFIG_HOTPLUG_CPU
399 /* cpufreq holds the hotplug lock, so we are safe from here on */
400 cpus_and(online_policy_cpus, cpu_online_map, policy->cpus);
402 online_policy_cpus = policy->cpus;
405 next_perf_state = data->freq_table[next_state].index;
406 if (perf->state == next_perf_state) {
407 if (unlikely(data->resume)) {
408 dprintk("Called after resume, resetting to P%d\n",
412 dprintk("Already at target state (P%d)\n",
418 switch (data->cpu_feature) {
419 case SYSTEM_INTEL_MSR_CAPABLE:
420 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
421 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
422 cmd.val = (u32) perf->states[next_perf_state].control;
424 case SYSTEM_IO_CAPABLE:
425 cmd.type = SYSTEM_IO_CAPABLE;
426 cmd.addr.io.port = perf->control_register.address;
427 cmd.addr.io.bit_width = perf->control_register.bit_width;
428 cmd.val = (u32) perf->states[next_perf_state].control;
434 cpus_clear(cmd.mask);
436 if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
437 cmd.mask = online_policy_cpus;
439 cpu_set(policy->cpu, cmd.mask);
441 freqs.old = perf->states[perf->state].core_frequency * 1000;
442 freqs.new = data->freq_table[next_state].frequency;
443 for_each_cpu_mask(i, cmd.mask) {
445 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
450 if (acpi_pstate_strict) {
451 if (!check_freqs(cmd.mask, freqs.new, data)) {
452 dprintk("acpi_cpufreq_target failed (%d)\n",
458 for_each_cpu_mask(i, cmd.mask) {
460 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
462 perf->state = next_perf_state;
467 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
469 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
471 dprintk("acpi_cpufreq_verify\n");
473 return cpufreq_frequency_table_verify(policy, data->freq_table);
477 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
479 struct acpi_processor_performance *perf = data->acpi_data;
482 /* search the closest match to cpu_khz */
485 unsigned long freqn = perf->states[0].core_frequency * 1000;
487 for (i=0; i<(perf->state_count-1); i++) {
489 freqn = perf->states[i+1].core_frequency * 1000;
490 if ((2 * cpu_khz) > (freqn + freq)) {
495 perf->state = perf->state_count-1;
498 /* assume CPU is at P0... */
500 return perf->states[0].core_frequency * 1000;
505 * acpi_cpufreq_early_init - initialize ACPI P-States library
507 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
508 * in order to determine correct frequency and voltage pairings. We can
509 * do _PDC and _PSD and find out the processor dependency for the
510 * actual init that will happen later...
512 static int __init acpi_cpufreq_early_init(void)
514 dprintk("acpi_cpufreq_early_init\n");
516 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
517 if (!acpi_perf_data) {
518 dprintk("Memory allocation error for acpi_perf_data.\n");
522 /* Do initialization in ACPI core */
523 acpi_processor_preregister_performance(acpi_perf_data);
529 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
530 * or do it in BIOS firmware and won't inform about it to OS. If not
531 * detected, this has a side effect of making CPU run at a different speed
532 * than OS intended it to run at. Detect it and handle it cleanly.
534 static int bios_with_sw_any_bug;
536 static int sw_any_bug_found(const struct dmi_system_id *d)
538 bios_with_sw_any_bug = 1;
542 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
544 .callback = sw_any_bug_found,
545 .ident = "Supermicro Server X6DLP",
547 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
548 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
549 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
556 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
559 unsigned int valid_states = 0;
560 unsigned int cpu = policy->cpu;
561 struct acpi_cpufreq_data *data;
562 unsigned int result = 0;
563 struct cpuinfo_x86 *c = &cpu_data[policy->cpu];
564 struct acpi_processor_performance *perf;
566 dprintk("acpi_cpufreq_cpu_init\n");
568 data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
572 data->acpi_data = percpu_ptr(acpi_perf_data, cpu);
573 drv_data[cpu] = data;
575 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
576 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
578 result = acpi_processor_register_performance(data->acpi_data, cpu);
582 perf = data->acpi_data;
583 policy->shared_type = perf->shared_type;
586 * Will let policy->cpus know about dependency only when software
587 * coordination is required.
589 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
590 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
591 policy->cpus = perf->shared_cpu_map;
595 dmi_check_system(sw_any_bug_dmi_table);
596 if (bios_with_sw_any_bug && cpus_weight(policy->cpus) == 1) {
597 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
598 policy->cpus = cpu_core_map[cpu];
602 /* capability check */
603 if (perf->state_count <= 1) {
604 dprintk("No P-States\n");
609 if (perf->control_register.space_id != perf->status_register.space_id) {
614 switch (perf->control_register.space_id) {
615 case ACPI_ADR_SPACE_SYSTEM_IO:
616 dprintk("SYSTEM IO addr space\n");
617 data->cpu_feature = SYSTEM_IO_CAPABLE;
619 case ACPI_ADR_SPACE_FIXED_HARDWARE:
620 dprintk("HARDWARE addr space\n");
621 if (!check_est_cpu(cpu)) {
625 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
628 dprintk("Unknown addr space %d\n",
629 (u32) (perf->control_register.space_id));
634 data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
635 (perf->state_count+1), GFP_KERNEL);
636 if (!data->freq_table) {
641 /* detect transition latency */
642 policy->cpuinfo.transition_latency = 0;
643 for (i=0; i<perf->state_count; i++) {
644 if ((perf->states[i].transition_latency * 1000) >
645 policy->cpuinfo.transition_latency)
646 policy->cpuinfo.transition_latency =
647 perf->states[i].transition_latency * 1000;
650 data->max_freq = perf->states[0].core_frequency * 1000;
652 for (i=0; i<perf->state_count; i++) {
653 if (i>0 && perf->states[i].core_frequency >=
654 data->freq_table[valid_states-1].frequency / 1000)
657 data->freq_table[valid_states].index = i;
658 data->freq_table[valid_states].frequency =
659 perf->states[i].core_frequency * 1000;
662 data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
665 result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
669 switch (perf->control_register.space_id) {
670 case ACPI_ADR_SPACE_SYSTEM_IO:
671 /* Current speed is unknown and not detectable by IO port */
672 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
674 case ACPI_ADR_SPACE_FIXED_HARDWARE:
675 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
676 policy->cur = get_cur_freq_on_cpu(cpu);
682 /* notify BIOS that we exist */
683 acpi_processor_notify_smm(THIS_MODULE);
685 /* Check for APERF/MPERF support in hardware */
686 if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
689 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
690 acpi_cpufreq_driver.getavg = get_measured_perf;
693 dprintk("CPU%u - ACPI performance management activated.\n", cpu);
694 for (i = 0; i < perf->state_count; i++)
695 dprintk(" %cP%d: %d MHz, %d mW, %d uS\n",
696 (i == perf->state ? '*' : ' '), i,
697 (u32) perf->states[i].core_frequency,
698 (u32) perf->states[i].power,
699 (u32) perf->states[i].transition_latency);
701 cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
704 * the first call to ->target() should result in us actually
705 * writing something to the appropriate registers.
712 kfree(data->freq_table);
714 acpi_processor_unregister_performance(perf, cpu);
717 drv_data[cpu] = NULL;
722 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
724 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
726 dprintk("acpi_cpufreq_cpu_exit\n");
729 cpufreq_frequency_table_put_attr(policy->cpu);
730 drv_data[policy->cpu] = NULL;
731 acpi_processor_unregister_performance(data->acpi_data,
739 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
741 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
743 dprintk("acpi_cpufreq_resume\n");
750 static struct freq_attr *acpi_cpufreq_attr[] = {
751 &cpufreq_freq_attr_scaling_available_freqs,
755 static struct cpufreq_driver acpi_cpufreq_driver = {
756 .verify = acpi_cpufreq_verify,
757 .target = acpi_cpufreq_target,
758 .init = acpi_cpufreq_cpu_init,
759 .exit = acpi_cpufreq_cpu_exit,
760 .resume = acpi_cpufreq_resume,
761 .name = "acpi-cpufreq",
762 .owner = THIS_MODULE,
763 .attr = acpi_cpufreq_attr,
766 static int __init acpi_cpufreq_init(void)
770 dprintk("acpi_cpufreq_init\n");
772 ret = acpi_cpufreq_early_init();
776 return cpufreq_register_driver(&acpi_cpufreq_driver);
779 static void __exit acpi_cpufreq_exit(void)
781 dprintk("acpi_cpufreq_exit\n");
783 cpufreq_unregister_driver(&acpi_cpufreq_driver);
785 free_percpu(acpi_perf_data);
790 module_param(acpi_pstate_strict, uint, 0644);
791 MODULE_PARM_DESC(acpi_pstate_strict,
792 "value 0 or non-zero. non-zero -> strict ACPI checks are "
793 "performed during frequency changes.");
795 late_initcall(acpi_cpufreq_init);
796 module_exit(acpi_cpufreq_exit);
798 MODULE_ALIAS("acpi");