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);
129 static void wrport(u16 port, u8 bit_width, u32 value)
133 else if (bit_width <= 16)
135 else if (bit_width <= 32)
139 static void rdport(u16 port, u8 bit_width, u32 * ret)
144 else if (bit_width <= 16)
146 else if (bit_width <= 32)
171 static void do_drv_read(struct drv_cmd *cmd)
176 case SYSTEM_INTEL_MSR_CAPABLE:
177 rdmsr(cmd->addr.msr.reg, cmd->val, h);
179 case SYSTEM_IO_CAPABLE:
180 rdport(cmd->addr.io.port, cmd->addr.io.bit_width, &cmd->val);
187 static void do_drv_write(struct drv_cmd *cmd)
192 case SYSTEM_INTEL_MSR_CAPABLE:
193 wrmsr(cmd->addr.msr.reg, cmd->val, h);
195 case SYSTEM_IO_CAPABLE:
196 wrport(cmd->addr.io.port, cmd->addr.io.bit_width, cmd->val);
203 static void drv_read(struct drv_cmd *cmd)
205 cpumask_t saved_mask = current->cpus_allowed;
208 set_cpus_allowed(current, cmd->mask);
210 set_cpus_allowed(current, saved_mask);
213 static void drv_write(struct drv_cmd *cmd)
215 cpumask_t saved_mask = current->cpus_allowed;
218 for_each_cpu_mask(i, cmd->mask) {
219 set_cpus_allowed(current, cpumask_of_cpu(i));
223 set_cpus_allowed(current, saved_mask);
227 static u32 get_cur_val(cpumask_t mask)
229 struct acpi_processor_performance *perf;
232 if (unlikely(cpus_empty(mask)))
235 switch (drv_data[first_cpu(mask)]->cpu_feature) {
236 case SYSTEM_INTEL_MSR_CAPABLE:
237 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
238 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
240 case SYSTEM_IO_CAPABLE:
241 cmd.type = SYSTEM_IO_CAPABLE;
242 perf = drv_data[first_cpu(mask)]->acpi_data;
243 cmd.addr.io.port = perf->control_register.address;
244 cmd.addr.io.bit_width = perf->control_register.bit_width;
254 dprintk("get_cur_val = %u\n", cmd.val);
260 * Return the measured active (C0) frequency on this CPU since last call
263 * Return: Average CPU frequency in terms of max frequency (zero on error)
265 * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
266 * over a period of time, while CPU is in C0 state.
267 * IA32_MPERF counts at the rate of max advertised frequency
268 * IA32_APERF counts at the rate of actual CPU frequency
269 * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
270 * no meaning should be associated with absolute values of these MSRs.
272 static unsigned int get_measured_perf(unsigned int cpu)
280 } aperf_cur, mperf_cur;
282 cpumask_t saved_mask;
283 unsigned int perf_percent;
286 saved_mask = current->cpus_allowed;
287 set_cpus_allowed(current, cpumask_of_cpu(cpu));
288 if (get_cpu() != cpu) {
289 /* We were not able to run on requested processor */
294 rdmsr(MSR_IA32_APERF, aperf_cur.split.lo, aperf_cur.split.hi);
295 rdmsr(MSR_IA32_MPERF, mperf_cur.split.lo, mperf_cur.split.hi);
297 wrmsr(MSR_IA32_APERF, 0,0);
298 wrmsr(MSR_IA32_MPERF, 0,0);
302 * We dont want to do 64 bit divide with 32 bit kernel
303 * Get an approximate value. Return failure in case we cannot get
304 * an approximate value.
306 if (unlikely(aperf_cur.split.hi || mperf_cur.split.hi)) {
310 h = max_t(u32, aperf_cur.split.hi, mperf_cur.split.hi);
311 shift_count = fls(h);
313 aperf_cur.whole >>= shift_count;
314 mperf_cur.whole >>= shift_count;
317 if (((unsigned long)(-1) / 100) < aperf_cur.split.lo) {
319 aperf_cur.split.lo >>= shift_count;
320 mperf_cur.split.lo >>= shift_count;
323 if (aperf_cur.split.lo && mperf_cur.split.lo)
324 perf_percent = (aperf_cur.split.lo * 100) / mperf_cur.split.lo;
329 if (unlikely(((unsigned long)(-1) / 100) < aperf_cur.whole)) {
331 aperf_cur.whole >>= shift_count;
332 mperf_cur.whole >>= shift_count;
335 if (aperf_cur.whole && mperf_cur.whole)
336 perf_percent = (aperf_cur.whole * 100) / mperf_cur.whole;
342 retval = drv_data[cpu]->max_freq * perf_percent / 100;
345 set_cpus_allowed(current, saved_mask);
347 dprintk("cpu %d: performance percent %d\n", cpu, perf_percent);
351 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
353 struct acpi_cpufreq_data *data = drv_data[cpu];
356 dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
358 if (unlikely(data == NULL ||
359 data->acpi_data == NULL || data->freq_table == NULL)) {
363 freq = extract_freq(get_cur_val(cpumask_of_cpu(cpu)), data);
364 dprintk("cur freq = %u\n", freq);
369 static unsigned int check_freqs(cpumask_t mask, unsigned int freq,
370 struct acpi_cpufreq_data *data)
372 unsigned int cur_freq;
375 for (i=0; i<100; i++) {
376 cur_freq = extract_freq(get_cur_val(mask), data);
377 if (cur_freq == freq)
384 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
385 unsigned int target_freq, unsigned int relation)
387 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
388 struct acpi_processor_performance *perf;
389 struct cpufreq_freqs freqs;
390 cpumask_t online_policy_cpus;
393 unsigned int next_state = 0;
394 unsigned int next_perf_state = 0;
398 dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
400 if (unlikely(data == NULL ||
401 data->acpi_data == NULL || data->freq_table == NULL)) {
405 perf = data->acpi_data;
406 result = cpufreq_frequency_table_target(policy,
409 relation, &next_state);
410 if (unlikely(result))
413 #ifdef CONFIG_HOTPLUG_CPU
414 /* cpufreq holds the hotplug lock, so we are safe from here on */
415 cpus_and(online_policy_cpus, cpu_online_map, policy->cpus);
417 online_policy_cpus = policy->cpus;
420 next_perf_state = data->freq_table[next_state].index;
421 if (perf->state == next_perf_state) {
422 if (unlikely(data->resume)) {
423 dprintk("Called after resume, resetting to P%d\n",
427 dprintk("Already at target state (P%d)\n",
433 switch (data->cpu_feature) {
434 case SYSTEM_INTEL_MSR_CAPABLE:
435 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
436 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
438 (u32) perf->states[next_perf_state].
439 control & INTEL_MSR_RANGE;
440 cmd.val = (cmd.val & ~INTEL_MSR_RANGE) | msr;
442 case SYSTEM_IO_CAPABLE:
443 cmd.type = SYSTEM_IO_CAPABLE;
444 cmd.addr.io.port = perf->control_register.address;
445 cmd.addr.io.bit_width = perf->control_register.bit_width;
446 cmd.val = (u32) perf->states[next_perf_state].control;
452 cpus_clear(cmd.mask);
454 if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
455 cmd.mask = online_policy_cpus;
457 cpu_set(policy->cpu, cmd.mask);
459 freqs.old = data->freq_table[perf->state].frequency;
460 freqs.new = data->freq_table[next_perf_state].frequency;
461 for_each_cpu_mask(i, cmd.mask) {
463 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
468 if (acpi_pstate_strict) {
469 if (!check_freqs(cmd.mask, freqs.new, data)) {
470 dprintk("acpi_cpufreq_target failed (%d)\n",
476 for_each_cpu_mask(i, cmd.mask) {
478 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
480 perf->state = next_perf_state;
485 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
487 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
489 dprintk("acpi_cpufreq_verify\n");
491 return cpufreq_frequency_table_verify(policy, data->freq_table);
495 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
497 struct acpi_processor_performance *perf = data->acpi_data;
500 /* search the closest match to cpu_khz */
503 unsigned long freqn = perf->states[0].core_frequency * 1000;
505 for (i=0; i<(perf->state_count-1); i++) {
507 freqn = perf->states[i+1].core_frequency * 1000;
508 if ((2 * cpu_khz) > (freqn + freq)) {
513 perf->state = perf->state_count-1;
516 /* assume CPU is at P0... */
518 return perf->states[0].core_frequency * 1000;
523 * acpi_cpufreq_early_init - initialize ACPI P-States library
525 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
526 * in order to determine correct frequency and voltage pairings. We can
527 * do _PDC and _PSD and find out the processor dependency for the
528 * actual init that will happen later...
530 static int acpi_cpufreq_early_init(void)
532 struct acpi_processor_performance *data;
536 dprintk("acpi_cpufreq_early_init\n");
538 for_each_possible_cpu(i) {
539 data = kzalloc(sizeof(struct acpi_processor_performance),
542 for_each_cpu_mask(j, covered) {
543 kfree(acpi_perf_data[j]);
544 acpi_perf_data[j] = NULL;
548 acpi_perf_data[i] = data;
552 /* Do initialization in ACPI core */
553 acpi_processor_preregister_performance(acpi_perf_data);
559 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
560 * or do it in BIOS firmware and won't inform about it to OS. If not
561 * detected, this has a side effect of making CPU run at a different speed
562 * than OS intended it to run at. Detect it and handle it cleanly.
564 static int bios_with_sw_any_bug;
566 static int sw_any_bug_found(struct dmi_system_id *d)
568 bios_with_sw_any_bug = 1;
572 static struct dmi_system_id sw_any_bug_dmi_table[] = {
574 .callback = sw_any_bug_found,
575 .ident = "Supermicro Server X6DLP",
577 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
578 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
579 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
586 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
589 unsigned int valid_states = 0;
590 unsigned int cpu = policy->cpu;
591 struct acpi_cpufreq_data *data;
592 unsigned int result = 0;
593 struct cpuinfo_x86 *c = &cpu_data[policy->cpu];
594 struct acpi_processor_performance *perf;
596 dprintk("acpi_cpufreq_cpu_init\n");
598 if (!acpi_perf_data[cpu])
601 data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
605 data->acpi_data = acpi_perf_data[cpu];
606 drv_data[cpu] = data;
608 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
609 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
611 result = acpi_processor_register_performance(data->acpi_data, cpu);
615 perf = data->acpi_data;
616 policy->shared_type = perf->shared_type;
619 * Will let policy->cpus know about dependency only when software
620 * coordination is required.
622 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
623 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
624 policy->cpus = perf->shared_cpu_map;
628 dmi_check_system(sw_any_bug_dmi_table);
629 if (bios_with_sw_any_bug && cpus_weight(policy->cpus) == 1) {
630 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
631 policy->cpus = cpu_core_map[cpu];
635 /* capability check */
636 if (perf->state_count <= 1) {
637 dprintk("No P-States\n");
642 if (perf->control_register.space_id != perf->status_register.space_id) {
647 switch (perf->control_register.space_id) {
648 case ACPI_ADR_SPACE_SYSTEM_IO:
649 dprintk("SYSTEM IO addr space\n");
650 data->cpu_feature = SYSTEM_IO_CAPABLE;
652 case ACPI_ADR_SPACE_FIXED_HARDWARE:
653 dprintk("HARDWARE addr space\n");
654 if (!check_est_cpu(cpu)) {
658 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
661 dprintk("Unknown addr space %d\n",
662 (u32) (perf->control_register.space_id));
667 data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
668 (perf->state_count+1), GFP_KERNEL);
669 if (!data->freq_table) {
674 /* detect transition latency */
675 policy->cpuinfo.transition_latency = 0;
676 for (i=0; i<perf->state_count; i++) {
677 if ((perf->states[i].transition_latency * 1000) >
678 policy->cpuinfo.transition_latency)
679 policy->cpuinfo.transition_latency =
680 perf->states[i].transition_latency * 1000;
682 policy->governor = CPUFREQ_DEFAULT_GOVERNOR;
684 data->max_freq = perf->states[0].core_frequency * 1000;
686 for (i=0; i<perf->state_count; i++) {
687 if (i>0 && perf->states[i].core_frequency ==
688 perf->states[i-1].core_frequency)
691 data->freq_table[valid_states].index = i;
692 data->freq_table[valid_states].frequency =
693 perf->states[i].core_frequency * 1000;
696 data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
698 result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
702 switch (data->cpu_feature) {
703 case ACPI_ADR_SPACE_SYSTEM_IO:
704 /* Current speed is unknown and not detectable by IO port */
705 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
707 case ACPI_ADR_SPACE_FIXED_HARDWARE:
708 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
709 get_cur_freq_on_cpu(cpu);
715 /* notify BIOS that we exist */
716 acpi_processor_notify_smm(THIS_MODULE);
718 /* Check for APERF/MPERF support in hardware */
719 if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
722 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
723 acpi_cpufreq_driver.getavg = get_measured_perf;
726 dprintk("CPU%u - ACPI performance management activated.\n", cpu);
727 for (i = 0; i < perf->state_count; i++)
728 dprintk(" %cP%d: %d MHz, %d mW, %d uS\n",
729 (i == perf->state ? '*' : ' '), i,
730 (u32) perf->states[i].core_frequency,
731 (u32) perf->states[i].power,
732 (u32) perf->states[i].transition_latency);
734 cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
737 * the first call to ->target() should result in us actually
738 * writing something to the appropriate registers.
745 kfree(data->freq_table);
747 acpi_processor_unregister_performance(perf, cpu);
750 drv_data[cpu] = NULL;
755 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
757 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
759 dprintk("acpi_cpufreq_cpu_exit\n");
762 cpufreq_frequency_table_put_attr(policy->cpu);
763 drv_data[policy->cpu] = NULL;
764 acpi_processor_unregister_performance(data->acpi_data,
772 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
774 struct acpi_cpufreq_data *data = drv_data[policy->cpu];
776 dprintk("acpi_cpufreq_resume\n");
783 static struct freq_attr *acpi_cpufreq_attr[] = {
784 &cpufreq_freq_attr_scaling_available_freqs,
788 static struct cpufreq_driver acpi_cpufreq_driver = {
789 .verify = acpi_cpufreq_verify,
790 .target = acpi_cpufreq_target,
791 .init = acpi_cpufreq_cpu_init,
792 .exit = acpi_cpufreq_cpu_exit,
793 .resume = acpi_cpufreq_resume,
794 .name = "acpi-cpufreq",
795 .owner = THIS_MODULE,
796 .attr = acpi_cpufreq_attr,
799 static int __init acpi_cpufreq_init(void)
801 dprintk("acpi_cpufreq_init\n");
803 acpi_cpufreq_early_init();
805 return cpufreq_register_driver(&acpi_cpufreq_driver);
808 static void __exit acpi_cpufreq_exit(void)
811 dprintk("acpi_cpufreq_exit\n");
813 cpufreq_unregister_driver(&acpi_cpufreq_driver);
815 for_each_possible_cpu(i) {
816 kfree(acpi_perf_data[i]);
817 acpi_perf_data[i] = NULL;
822 module_param(acpi_pstate_strict, uint, 0644);
823 MODULE_PARM_DESC(acpi_pstate_strict,
824 "value 0 or non-zero. non-zero -> strict ACPI checks are "
825 "performed during frequency changes.");
827 late_initcall(acpi_cpufreq_init);
828 module_exit(acpi_cpufreq_exit);
830 MODULE_ALIAS("acpi");