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 DEFINE_PER_CPU(struct acpi_cpufreq_data *, drv_data);
72 /* acpi_perf_data is a pointer to percpu data. */
73 static struct acpi_processor_performance *acpi_perf_data;
75 static struct cpufreq_driver acpi_cpufreq_driver;
77 static unsigned int acpi_pstate_strict;
79 static int check_est_cpu(unsigned int cpuid)
81 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
83 if (cpu->x86_vendor != X86_VENDOR_INTEL ||
84 !cpu_has(cpu, X86_FEATURE_EST))
90 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
92 struct acpi_processor_performance *perf;
95 perf = data->acpi_data;
97 for (i=0; i<perf->state_count; i++) {
98 if (value == perf->states[i].status)
99 return data->freq_table[i].frequency;
104 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
107 struct acpi_processor_performance *perf;
109 msr &= INTEL_MSR_RANGE;
110 perf = data->acpi_data;
112 for (i=0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
113 if (msr == perf->states[data->freq_table[i].index].status)
114 return data->freq_table[i].frequency;
116 return data->freq_table[0].frequency;
119 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
121 switch (data->cpu_feature) {
122 case SYSTEM_INTEL_MSR_CAPABLE:
123 return extract_msr(val, data);
124 case SYSTEM_IO_CAPABLE:
125 return extract_io(val, data);
152 static void do_drv_read(struct drv_cmd *cmd)
157 case SYSTEM_INTEL_MSR_CAPABLE:
158 rdmsr(cmd->addr.msr.reg, cmd->val, h);
160 case SYSTEM_IO_CAPABLE:
161 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
163 (u32)cmd->addr.io.bit_width);
170 static void do_drv_write(struct drv_cmd *cmd)
175 case SYSTEM_INTEL_MSR_CAPABLE:
176 rdmsr(cmd->addr.msr.reg, lo, hi);
177 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
178 wrmsr(cmd->addr.msr.reg, lo, hi);
180 case SYSTEM_IO_CAPABLE:
181 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
183 (u32)cmd->addr.io.bit_width);
190 static void drv_read(struct drv_cmd *cmd)
192 cpumask_t saved_mask = current->cpus_allowed;
195 set_cpus_allowed(current, cmd->mask);
197 set_cpus_allowed(current, saved_mask);
200 static void drv_write(struct drv_cmd *cmd)
202 cpumask_t saved_mask = current->cpus_allowed;
205 for_each_cpu_mask(i, cmd->mask) {
206 set_cpus_allowed(current, cpumask_of_cpu(i));
210 set_cpus_allowed(current, saved_mask);
214 static u32 get_cur_val(cpumask_t mask)
216 struct acpi_processor_performance *perf;
219 if (unlikely(cpus_empty(mask)))
222 switch (per_cpu(drv_data, first_cpu(mask))->cpu_feature) {
223 case SYSTEM_INTEL_MSR_CAPABLE:
224 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
225 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
227 case SYSTEM_IO_CAPABLE:
228 cmd.type = SYSTEM_IO_CAPABLE;
229 perf = per_cpu(drv_data, first_cpu(mask))->acpi_data;
230 cmd.addr.io.port = perf->control_register.address;
231 cmd.addr.io.bit_width = perf->control_register.bit_width;
241 dprintk("get_cur_val = %u\n", cmd.val);
247 * Return the measured active (C0) frequency on this CPU since last call
250 * Return: Average CPU frequency in terms of max frequency (zero on error)
252 * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
253 * over a period of time, while CPU is in C0 state.
254 * IA32_MPERF counts at the rate of max advertised frequency
255 * IA32_APERF counts at the rate of actual CPU frequency
256 * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
257 * no meaning should be associated with absolute values of these MSRs.
259 static unsigned int get_measured_perf(unsigned int cpu)
267 } aperf_cur, mperf_cur;
269 cpumask_t saved_mask;
270 unsigned int perf_percent;
273 saved_mask = current->cpus_allowed;
274 set_cpus_allowed(current, cpumask_of_cpu(cpu));
275 if (get_cpu() != cpu) {
276 /* We were not able to run on requested processor */
281 rdmsr(MSR_IA32_APERF, aperf_cur.split.lo, aperf_cur.split.hi);
282 rdmsr(MSR_IA32_MPERF, mperf_cur.split.lo, mperf_cur.split.hi);
284 wrmsr(MSR_IA32_APERF, 0,0);
285 wrmsr(MSR_IA32_MPERF, 0,0);
289 * We dont want to do 64 bit divide with 32 bit kernel
290 * Get an approximate value. Return failure in case we cannot get
291 * an approximate value.
293 if (unlikely(aperf_cur.split.hi || mperf_cur.split.hi)) {
297 h = max_t(u32, aperf_cur.split.hi, mperf_cur.split.hi);
298 shift_count = fls(h);
300 aperf_cur.whole >>= shift_count;
301 mperf_cur.whole >>= shift_count;
304 if (((unsigned long)(-1) / 100) < aperf_cur.split.lo) {
306 aperf_cur.split.lo >>= shift_count;
307 mperf_cur.split.lo >>= shift_count;
310 if (aperf_cur.split.lo && mperf_cur.split.lo)
311 perf_percent = (aperf_cur.split.lo * 100) / mperf_cur.split.lo;
316 if (unlikely(((unsigned long)(-1) / 100) < aperf_cur.whole)) {
318 aperf_cur.whole >>= shift_count;
319 mperf_cur.whole >>= shift_count;
322 if (aperf_cur.whole && mperf_cur.whole)
323 perf_percent = (aperf_cur.whole * 100) / mperf_cur.whole;
329 retval = per_cpu(drv_data, cpu)->max_freq * perf_percent / 100;
332 set_cpus_allowed(current, saved_mask);
334 dprintk("cpu %d: performance percent %d\n", cpu, perf_percent);
338 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
340 struct acpi_cpufreq_data *data = per_cpu(drv_data, cpu);
343 dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
345 if (unlikely(data == NULL ||
346 data->acpi_data == NULL || data->freq_table == NULL)) {
350 freq = extract_freq(get_cur_val(cpumask_of_cpu(cpu)), data);
351 dprintk("cur freq = %u\n", freq);
356 static unsigned int check_freqs(cpumask_t mask, unsigned int freq,
357 struct acpi_cpufreq_data *data)
359 unsigned int cur_freq;
362 for (i=0; i<100; i++) {
363 cur_freq = extract_freq(get_cur_val(mask), data);
364 if (cur_freq == freq)
371 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
372 unsigned int target_freq, unsigned int relation)
374 struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
375 struct acpi_processor_performance *perf;
376 struct cpufreq_freqs freqs;
377 cpumask_t online_policy_cpus;
379 unsigned int next_state = 0; /* Index into freq_table */
380 unsigned int next_perf_state = 0; /* Index into perf table */
384 dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
386 if (unlikely(data == NULL ||
387 data->acpi_data == NULL || data->freq_table == NULL)) {
391 perf = data->acpi_data;
392 result = cpufreq_frequency_table_target(policy,
395 relation, &next_state);
396 if (unlikely(result))
399 #ifdef CONFIG_HOTPLUG_CPU
400 /* cpufreq holds the hotplug lock, so we are safe from here on */
401 cpus_and(online_policy_cpus, cpu_online_map, policy->cpus);
403 online_policy_cpus = policy->cpus;
406 next_perf_state = data->freq_table[next_state].index;
407 if (perf->state == next_perf_state) {
408 if (unlikely(data->resume)) {
409 dprintk("Called after resume, resetting to P%d\n",
413 dprintk("Already at target state (P%d)\n",
419 switch (data->cpu_feature) {
420 case SYSTEM_INTEL_MSR_CAPABLE:
421 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
422 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
423 cmd.val = (u32) perf->states[next_perf_state].control;
425 case SYSTEM_IO_CAPABLE:
426 cmd.type = SYSTEM_IO_CAPABLE;
427 cmd.addr.io.port = perf->control_register.address;
428 cmd.addr.io.bit_width = perf->control_register.bit_width;
429 cmd.val = (u32) perf->states[next_perf_state].control;
435 cpus_clear(cmd.mask);
437 if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
438 cmd.mask = online_policy_cpus;
440 cpu_set(policy->cpu, cmd.mask);
442 freqs.old = perf->states[perf->state].core_frequency * 1000;
443 freqs.new = data->freq_table[next_state].frequency;
444 for_each_cpu_mask(i, cmd.mask) {
446 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
451 if (acpi_pstate_strict) {
452 if (!check_freqs(cmd.mask, freqs.new, data)) {
453 dprintk("acpi_cpufreq_target failed (%d)\n",
459 for_each_cpu_mask(i, cmd.mask) {
461 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
463 perf->state = next_perf_state;
468 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
470 struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
472 dprintk("acpi_cpufreq_verify\n");
474 return cpufreq_frequency_table_verify(policy, data->freq_table);
478 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
480 struct acpi_processor_performance *perf = data->acpi_data;
483 /* search the closest match to cpu_khz */
486 unsigned long freqn = perf->states[0].core_frequency * 1000;
488 for (i=0; i<(perf->state_count-1); i++) {
490 freqn = perf->states[i+1].core_frequency * 1000;
491 if ((2 * cpu_khz) > (freqn + freq)) {
496 perf->state = perf->state_count-1;
499 /* assume CPU is at P0... */
501 return perf->states[0].core_frequency * 1000;
506 * acpi_cpufreq_early_init - initialize ACPI P-States library
508 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
509 * in order to determine correct frequency and voltage pairings. We can
510 * do _PDC and _PSD and find out the processor dependency for the
511 * actual init that will happen later...
513 static int __init acpi_cpufreq_early_init(void)
515 dprintk("acpi_cpufreq_early_init\n");
517 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
518 if (!acpi_perf_data) {
519 dprintk("Memory allocation error for acpi_perf_data.\n");
523 /* Do initialization in ACPI core */
524 acpi_processor_preregister_performance(acpi_perf_data);
530 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
531 * or do it in BIOS firmware and won't inform about it to OS. If not
532 * detected, this has a side effect of making CPU run at a different speed
533 * than OS intended it to run at. Detect it and handle it cleanly.
535 static int bios_with_sw_any_bug;
537 static int sw_any_bug_found(const struct dmi_system_id *d)
539 bios_with_sw_any_bug = 1;
543 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
545 .callback = sw_any_bug_found,
546 .ident = "Supermicro Server X6DLP",
548 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
549 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
550 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
557 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
560 unsigned int valid_states = 0;
561 unsigned int cpu = policy->cpu;
562 struct acpi_cpufreq_data *data;
563 unsigned int result = 0;
564 struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
565 struct acpi_processor_performance *perf;
567 dprintk("acpi_cpufreq_cpu_init\n");
569 data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
573 data->acpi_data = percpu_ptr(acpi_perf_data, cpu);
574 per_cpu(drv_data, cpu) = data;
576 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
577 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
579 result = acpi_processor_register_performance(data->acpi_data, cpu);
583 perf = data->acpi_data;
584 policy->shared_type = perf->shared_type;
587 * Will let policy->cpus know about dependency only when software
588 * coordination is required.
590 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
591 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
592 policy->cpus = perf->shared_cpu_map;
596 dmi_check_system(sw_any_bug_dmi_table);
597 if (bios_with_sw_any_bug && cpus_weight(policy->cpus) == 1) {
598 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
599 policy->cpus = per_cpu(cpu_core_map, cpu);
603 /* capability check */
604 if (perf->state_count <= 1) {
605 dprintk("No P-States\n");
610 if (perf->control_register.space_id != perf->status_register.space_id) {
615 switch (perf->control_register.space_id) {
616 case ACPI_ADR_SPACE_SYSTEM_IO:
617 dprintk("SYSTEM IO addr space\n");
618 data->cpu_feature = SYSTEM_IO_CAPABLE;
620 case ACPI_ADR_SPACE_FIXED_HARDWARE:
621 dprintk("HARDWARE addr space\n");
622 if (!check_est_cpu(cpu)) {
626 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
629 dprintk("Unknown addr space %d\n",
630 (u32) (perf->control_register.space_id));
635 data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
636 (perf->state_count+1), GFP_KERNEL);
637 if (!data->freq_table) {
642 /* detect transition latency */
643 policy->cpuinfo.transition_latency = 0;
644 for (i=0; i<perf->state_count; i++) {
645 if ((perf->states[i].transition_latency * 1000) >
646 policy->cpuinfo.transition_latency)
647 policy->cpuinfo.transition_latency =
648 perf->states[i].transition_latency * 1000;
651 data->max_freq = perf->states[0].core_frequency * 1000;
653 for (i=0; i<perf->state_count; i++) {
654 if (i>0 && perf->states[i].core_frequency >=
655 data->freq_table[valid_states-1].frequency / 1000)
658 data->freq_table[valid_states].index = i;
659 data->freq_table[valid_states].frequency =
660 perf->states[i].core_frequency * 1000;
663 data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
666 result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
670 switch (perf->control_register.space_id) {
671 case ACPI_ADR_SPACE_SYSTEM_IO:
672 /* Current speed is unknown and not detectable by IO port */
673 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
675 case ACPI_ADR_SPACE_FIXED_HARDWARE:
676 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
677 policy->cur = get_cur_freq_on_cpu(cpu);
683 /* notify BIOS that we exist */
684 acpi_processor_notify_smm(THIS_MODULE);
686 /* Check for APERF/MPERF support in hardware */
687 if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
690 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
691 acpi_cpufreq_driver.getavg = get_measured_perf;
694 dprintk("CPU%u - ACPI performance management activated.\n", cpu);
695 for (i = 0; i < perf->state_count; i++)
696 dprintk(" %cP%d: %d MHz, %d mW, %d uS\n",
697 (i == perf->state ? '*' : ' '), i,
698 (u32) perf->states[i].core_frequency,
699 (u32) perf->states[i].power,
700 (u32) perf->states[i].transition_latency);
702 cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
705 * the first call to ->target() should result in us actually
706 * writing something to the appropriate registers.
713 kfree(data->freq_table);
715 acpi_processor_unregister_performance(perf, cpu);
718 per_cpu(drv_data, cpu) = NULL;
723 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
725 struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
727 dprintk("acpi_cpufreq_cpu_exit\n");
730 cpufreq_frequency_table_put_attr(policy->cpu);
731 per_cpu(drv_data, policy->cpu) = NULL;
732 acpi_processor_unregister_performance(data->acpi_data,
740 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
742 struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
744 dprintk("acpi_cpufreq_resume\n");
751 static struct freq_attr *acpi_cpufreq_attr[] = {
752 &cpufreq_freq_attr_scaling_available_freqs,
756 static struct cpufreq_driver acpi_cpufreq_driver = {
757 .verify = acpi_cpufreq_verify,
758 .target = acpi_cpufreq_target,
759 .init = acpi_cpufreq_cpu_init,
760 .exit = acpi_cpufreq_cpu_exit,
761 .resume = acpi_cpufreq_resume,
762 .name = "acpi-cpufreq",
763 .owner = THIS_MODULE,
764 .attr = acpi_cpufreq_attr,
767 static int __init acpi_cpufreq_init(void)
771 dprintk("acpi_cpufreq_init\n");
773 ret = acpi_cpufreq_early_init();
777 return cpufreq_register_driver(&acpi_cpufreq_driver);
780 static void __exit acpi_cpufreq_exit(void)
782 dprintk("acpi_cpufreq_exit\n");
784 cpufreq_unregister_driver(&acpi_cpufreq_driver);
786 free_percpu(acpi_perf_data);
791 module_param(acpi_pstate_strict, uint, 0644);
792 MODULE_PARM_DESC(acpi_pstate_strict,
793 "value 0 or non-zero. non-zero -> strict ACPI checks are "
794 "performed during frequency changes.");
796 late_initcall(acpi_cpufreq_init);
797 module_exit(acpi_cpufreq_exit);
799 MODULE_ALIAS("acpi");