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/sched.h> /* current */
36 #include <linux/dmi.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 struct acpi_cpufreq_data *drv_data[NR_CPUS];
72 static struct acpi_processor_performance *acpi_perf_data[NR_CPUS];
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);
130 static void wrport(u16 port, u8 bit_width, u32 value)
132 if (bit_width <= 8) {
134 } else if (bit_width <= 16) {
136 } else if (bit_width <= 32) {
141 static void rdport(u16 port, u8 bit_width, u32 * ret)
144 if (bit_width <= 8) {
146 } else if (bit_width <= 16) {
148 } else if (bit_width <= 32) {
174 static void do_drv_read(struct drv_cmd *cmd)
179 case SYSTEM_INTEL_MSR_CAPABLE:
180 rdmsr(cmd->addr.msr.reg, cmd->val, h);
182 case SYSTEM_IO_CAPABLE:
183 rdport(cmd->addr.io.port, cmd->addr.io.bit_width, &cmd->val);
190 static void do_drv_write(struct drv_cmd *cmd)
195 case SYSTEM_INTEL_MSR_CAPABLE:
196 wrmsr(cmd->addr.msr.reg, cmd->val, h);
198 case SYSTEM_IO_CAPABLE:
199 wrport(cmd->addr.io.port, cmd->addr.io.bit_width, cmd->val);
206 static inline void drv_read(struct drv_cmd *cmd)
208 cpumask_t saved_mask = current->cpus_allowed;
211 set_cpus_allowed(current, cmd->mask);
213 set_cpus_allowed(current, saved_mask);
217 static void drv_write(struct drv_cmd *cmd)
219 cpumask_t saved_mask = current->cpus_allowed;
222 for_each_cpu_mask(i, cmd->mask) {
223 set_cpus_allowed(current, cpumask_of_cpu(i));
227 set_cpus_allowed(current, saved_mask);
231 static u32 get_cur_val(cpumask_t mask)
233 struct acpi_processor_performance *perf;
236 if (unlikely(cpus_empty(mask)))
239 switch (drv_data[first_cpu(mask)]->cpu_feature) {
240 case SYSTEM_INTEL_MSR_CAPABLE:
241 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
242 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
244 case SYSTEM_IO_CAPABLE:
245 cmd.type = SYSTEM_IO_CAPABLE;
246 perf = drv_data[first_cpu(mask)]->acpi_data;
247 cmd.addr.io.port = perf->control_register.address;
248 cmd.addr.io.bit_width = perf->control_register.bit_width;
258 dprintk("get_cur_val = %u\n", cmd.val);
264 * Return the measured active (C0) frequency on this CPU since last call
267 * Return: Average CPU frequency in terms of max frequency (zero on error)
269 * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
270 * over a period of time, while CPU is in C0 state.
271 * IA32_MPERF counts at the rate of max advertised frequency
272 * IA32_APERF counts at the rate of actual CPU frequency
273 * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
274 * no meaning should be associated with absolute values of these MSRs.
276 static unsigned int get_measured_perf(unsigned int cpu)
284 } aperf_cur, mperf_cur;
286 cpumask_t saved_mask;
287 unsigned int perf_percent;
290 saved_mask = current->cpus_allowed;
291 set_cpus_allowed(current, cpumask_of_cpu(cpu));
292 if (get_cpu() != cpu) {
293 /* We were not able to run on requested processor */
298 rdmsr(MSR_IA32_APERF, aperf_cur.split.lo, aperf_cur.split.hi);
299 rdmsr(MSR_IA32_MPERF, mperf_cur.split.lo, mperf_cur.split.hi);
301 wrmsr(MSR_IA32_APERF, 0,0);
302 wrmsr(MSR_IA32_MPERF, 0,0);
306 * We dont want to do 64 bit divide with 32 bit kernel
307 * Get an approximate value. Return failure in case we cannot get
308 * an approximate value.
310 if (unlikely(aperf_cur.split.hi || mperf_cur.split.hi)) {
314 h = max_t(u32, aperf_cur.split.hi, mperf_cur.split.hi);
315 shift_count = fls(h);
317 aperf_cur.whole >>= shift_count;
318 mperf_cur.whole >>= shift_count;
321 if (((unsigned long)(-1) / 100) < aperf_cur.split.lo) {
323 aperf_cur.split.lo >>= shift_count;
324 mperf_cur.split.lo >>= shift_count;
327 if (aperf_cur.split.lo && mperf_cur.split.lo) {
328 perf_percent = (aperf_cur.split.lo * 100) / mperf_cur.split.lo;
334 if (unlikely(((unsigned long)(-1) / 100) < aperf_cur.whole)) {
336 aperf_cur.whole >>= shift_count;
337 mperf_cur.whole >>= shift_count;
340 if (aperf_cur.whole && mperf_cur.whole) {
341 perf_percent = (aperf_cur.whole * 100) / mperf_cur.whole;
348 retval = drv_data[cpu]->max_freq * perf_percent / 100;
351 set_cpus_allowed(current, saved_mask);
353 dprintk("cpu %d: performance percent %d\n", cpu, perf_percent);
357 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
359 struct acpi_cpufreq_data *data = drv_data[cpu];
362 dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
364 if (unlikely(data == NULL ||
365 data->acpi_data == NULL || data->freq_table == NULL)) {
369 freq = extract_freq(get_cur_val(cpumask_of_cpu(cpu)), data);
370 dprintk("cur freq = %u\n", freq);
375 static unsigned int check_freqs(cpumask_t mask, unsigned int freq,
376 struct acpi_cpufreq_data *data)
378 unsigned int cur_freq;
381 for (i = 0; i < 100; i++) {
382 cur_freq = extract_freq(get_cur_val(mask), data);
383 if (cur_freq == freq)
390 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
391 unsigned int target_freq, unsigned int relation)
393 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
394 struct acpi_processor_performance *perf;
395 struct cpufreq_freqs freqs;
396 cpumask_t online_policy_cpus;
399 unsigned int next_state = 0;
400 unsigned int next_perf_state = 0;
404 dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
406 if (unlikely(data == NULL ||
407 data->acpi_data == NULL || data->freq_table == NULL)) {
411 perf = data->acpi_data;
412 result = cpufreq_frequency_table_target(policy,
415 relation, &next_state);
416 if (unlikely(result))
419 #ifdef CONFIG_HOTPLUG_CPU
420 /* cpufreq holds the hotplug lock, so we are safe from here on */
421 cpus_and(online_policy_cpus, cpu_online_map, policy->cpus);
423 online_policy_cpus = policy->cpus;
426 next_perf_state = data->freq_table[next_state].index;
427 if (perf->state == next_perf_state) {
428 if (unlikely(data->resume)) {
429 dprintk("Called after resume, resetting to P%d\n",
433 dprintk("Already at target state (P%d)\n",
439 switch (data->cpu_feature) {
440 case SYSTEM_INTEL_MSR_CAPABLE:
441 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
442 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
444 (u32) perf->states[next_perf_state].
445 control & INTEL_MSR_RANGE;
446 cmd.val = (cmd.val & ~INTEL_MSR_RANGE) | msr;
448 case SYSTEM_IO_CAPABLE:
449 cmd.type = SYSTEM_IO_CAPABLE;
450 cmd.addr.io.port = perf->control_register.address;
451 cmd.addr.io.bit_width = perf->control_register.bit_width;
452 cmd.val = (u32) perf->states[next_perf_state].control;
458 cpus_clear(cmd.mask);
460 if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
461 cmd.mask = online_policy_cpus;
463 cpu_set(policy->cpu, cmd.mask);
465 freqs.old = data->freq_table[perf->state].frequency;
466 freqs.new = data->freq_table[next_perf_state].frequency;
467 for_each_cpu_mask(i, cmd.mask) {
469 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
474 if (acpi_pstate_strict) {
475 if (!check_freqs(cmd.mask, freqs.new, data)) {
476 dprintk("acpi_cpufreq_target failed (%d)\n",
482 for_each_cpu_mask(i, cmd.mask) {
484 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
486 perf->state = next_perf_state;
491 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
493 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
495 dprintk("acpi_cpufreq_verify\n");
497 return cpufreq_frequency_table_verify(policy, data->freq_table);
501 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
503 struct acpi_processor_performance *perf = data->acpi_data;
506 /* search the closest match to cpu_khz */
509 unsigned long freqn = perf->states[0].core_frequency * 1000;
511 for (i = 0; i < (perf->state_count - 1); i++) {
513 freqn = perf->states[i + 1].core_frequency * 1000;
514 if ((2 * cpu_khz) > (freqn + freq)) {
519 perf->state = perf->state_count - 1;
522 /* assume CPU is at P0... */
524 return perf->states[0].core_frequency * 1000;
529 * acpi_cpufreq_early_init - initialize ACPI P-States library
531 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
532 * in order to determine correct frequency and voltage pairings. We can
533 * do _PDC and _PSD and find out the processor dependency for the
534 * actual init that will happen later...
536 static int acpi_cpufreq_early_init(void)
538 struct acpi_processor_performance *data;
542 dprintk("acpi_cpufreq_early_init\n");
544 for_each_possible_cpu(i) {
545 data = kzalloc(sizeof(struct acpi_processor_performance),
548 for_each_cpu_mask(j, covered) {
549 kfree(acpi_perf_data[j]);
550 acpi_perf_data[j] = NULL;
554 acpi_perf_data[i] = data;
558 /* Do initialization in ACPI core */
559 acpi_processor_preregister_performance(acpi_perf_data);
564 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
565 * or do it in BIOS firmware and won't inform about it to OS. If not
566 * detected, this has a side effect of making CPU run at a different speed
567 * than OS intended it to run at. Detect it and handle it cleanly.
569 static int bios_with_sw_any_bug;
571 static int sw_any_bug_found(struct dmi_system_id *d)
573 bios_with_sw_any_bug = 1;
577 static struct dmi_system_id sw_any_bug_dmi_table[] = {
579 .callback = sw_any_bug_found,
580 .ident = "Supermicro Server X6DLP",
582 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
583 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
584 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
590 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
593 unsigned int valid_states = 0;
594 unsigned int cpu = policy->cpu;
595 struct acpi_cpufreq_data *data;
596 unsigned int result = 0;
597 struct cpuinfo_x86 *c = &cpu_data[policy->cpu];
598 struct acpi_processor_performance *perf;
600 dprintk("acpi_cpufreq_cpu_init\n");
602 if (!acpi_perf_data[cpu])
605 data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
609 data->acpi_data = acpi_perf_data[cpu];
610 drv_data[cpu] = data;
612 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
613 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
616 result = acpi_processor_register_performance(data->acpi_data, cpu);
620 perf = data->acpi_data;
621 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 policy->cpus = perf->shared_cpu_map;
632 dmi_check_system(sw_any_bug_dmi_table);
633 if (bios_with_sw_any_bug && cpus_weight(policy->cpus) == 1) {
634 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
635 policy->cpus = cpu_core_map[cpu];
639 /* capability check */
640 if (perf->state_count <= 1) {
641 dprintk("No P-States\n");
646 if (perf->control_register.space_id != perf->status_register.space_id) {
651 switch (perf->control_register.space_id) {
652 case ACPI_ADR_SPACE_SYSTEM_IO:
653 dprintk("SYSTEM IO addr space\n");
654 data->cpu_feature = SYSTEM_IO_CAPABLE;
656 case ACPI_ADR_SPACE_FIXED_HARDWARE:
657 dprintk("HARDWARE addr space\n");
658 if (!check_est_cpu(cpu)) {
662 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
665 dprintk("Unknown addr space %d\n",
666 (u32) (perf->control_register.space_id));
672 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;
687 policy->governor = CPUFREQ_DEFAULT_GOVERNOR;
689 data->max_freq = perf->states[0].core_frequency * 1000;
691 for (i = 0; i < perf->state_count; i++) {
692 if (i > 0 && perf->states[i].core_frequency ==
693 perf->states[i - 1].core_frequency)
696 data->freq_table[valid_states].index = i;
697 data->freq_table[valid_states].frequency =
698 perf->states[i].core_frequency * 1000;
701 data->freq_table[perf->state_count].frequency = CPUFREQ_TABLE_END;
703 result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
708 switch (data->cpu_feature) {
709 case ACPI_ADR_SPACE_SYSTEM_IO:
710 /* Current speed is unknown and not detectable by IO port */
711 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
713 case ACPI_ADR_SPACE_FIXED_HARDWARE:
714 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
715 get_cur_freq_on_cpu(cpu);
721 /* notify BIOS that we exist */
722 acpi_processor_notify_smm(THIS_MODULE);
724 /* Check for APERF/MPERF support in hardware */
725 if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
728 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY) {
729 acpi_cpufreq_driver.getavg = get_measured_perf;
733 dprintk("CPU%u - ACPI performance management activated.\n", cpu);
734 for (i = 0; i < perf->state_count; i++)
735 dprintk(" %cP%d: %d MHz, %d mW, %d uS\n",
736 (i == perf->state ? '*' : ' '), i,
737 (u32) perf->states[i].core_frequency,
738 (u32) perf->states[i].power,
739 (u32) perf->states[i].transition_latency);
741 cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
744 * the first call to ->target() should result in us actually
745 * writing something to the appropriate registers.
752 kfree(data->freq_table);
754 acpi_processor_unregister_performance(perf, cpu);
757 drv_data[cpu] = NULL;
762 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
764 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
766 dprintk("acpi_cpufreq_cpu_exit\n");
769 cpufreq_frequency_table_put_attr(policy->cpu);
770 drv_data[policy->cpu] = NULL;
771 acpi_processor_unregister_performance(data->acpi_data,
779 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
781 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
783 dprintk("acpi_cpufreq_resume\n");
790 static struct freq_attr *acpi_cpufreq_attr[] = {
791 &cpufreq_freq_attr_scaling_available_freqs,
795 static struct cpufreq_driver acpi_cpufreq_driver = {
796 .verify = acpi_cpufreq_verify,
797 .target = acpi_cpufreq_target,
798 .init = acpi_cpufreq_cpu_init,
799 .exit = acpi_cpufreq_cpu_exit,
800 .resume = acpi_cpufreq_resume,
801 .name = "acpi-cpufreq",
802 .owner = THIS_MODULE,
803 .attr = acpi_cpufreq_attr,
806 static int __init acpi_cpufreq_init(void)
808 dprintk("acpi_cpufreq_init\n");
810 acpi_cpufreq_early_init();
812 return cpufreq_register_driver(&acpi_cpufreq_driver);
815 static void __exit acpi_cpufreq_exit(void)
818 dprintk("acpi_cpufreq_exit\n");
820 cpufreq_unregister_driver(&acpi_cpufreq_driver);
822 for_each_possible_cpu(i) {
823 kfree(acpi_perf_data[i]);
824 acpi_perf_data[i] = NULL;
829 module_param(acpi_pstate_strict, uint, 0644);
830 MODULE_PARM_DESC(acpi_pstate_strict,
831 "value 0 or non-zero. non-zero -> strict ACPI checks are performed during frequency changes.");
833 late_initcall(acpi_cpufreq_init);
834 module_exit(acpi_cpufreq_exit);
836 MODULE_ALIAS("acpi");