2 * acpi-cpufreq.c - ACPI Processor P-States Driver
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>
36 #include <trace/events/power.h>
38 #include <linux/acpi.h>
40 #include <linux/delay.h>
41 #include <linux/uaccess.h>
43 #include <acpi/processor.h>
46 #include <asm/processor.h>
47 #include <asm/cpufeature.h>
49 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, \
52 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
53 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
54 MODULE_LICENSE("GPL");
57 UNDEFINED_CAPABLE
= 0,
58 SYSTEM_INTEL_MSR_CAPABLE
,
62 #define INTEL_MSR_RANGE (0xffff)
64 struct acpi_cpufreq_data
{
65 struct acpi_processor_performance
*acpi_data
;
66 struct cpufreq_frequency_table
*freq_table
;
68 unsigned int cpu_feature
;
71 static DEFINE_PER_CPU(struct acpi_cpufreq_data
*, drv_data
);
73 static DEFINE_PER_CPU(struct aperfmperf
, old_perf
);
75 /* acpi_perf_data is a pointer to percpu data. */
76 static struct acpi_processor_performance
*acpi_perf_data
;
78 static struct cpufreq_driver acpi_cpufreq_driver
;
80 static unsigned int acpi_pstate_strict
;
82 static int check_est_cpu(unsigned int cpuid
)
84 struct cpuinfo_x86
*cpu
= &cpu_data(cpuid
);
86 return 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
);
141 const struct cpumask
*mask
;
149 /* Called via smp_call_function_single(), on the target CPU */
150 static void do_drv_read(void *_cmd
)
152 struct drv_cmd
*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 /* Called via smp_call_function_many(), on the target CPUs */
170 static void do_drv_write(void *_cmd
)
172 struct drv_cmd
*cmd
= _cmd
;
176 case SYSTEM_INTEL_MSR_CAPABLE
:
177 rdmsr(cmd
->addr
.msr
.reg
, lo
, hi
);
178 lo
= (lo
& ~INTEL_MSR_RANGE
) | (cmd
->val
& INTEL_MSR_RANGE
);
179 wrmsr(cmd
->addr
.msr
.reg
, lo
, hi
);
181 case SYSTEM_IO_CAPABLE
:
182 acpi_os_write_port((acpi_io_address
)cmd
->addr
.io
.port
,
184 (u32
)cmd
->addr
.io
.bit_width
);
191 static void drv_read(struct drv_cmd
*cmd
)
195 smp_call_function_single(cpumask_any(cmd
->mask
), do_drv_read
, cmd
, 1);
198 static void drv_write(struct drv_cmd
*cmd
)
202 this_cpu
= get_cpu();
203 if (cpumask_test_cpu(this_cpu
, cmd
->mask
))
205 smp_call_function_many(cmd
->mask
, do_drv_write
, cmd
, 1);
209 static u32
get_cur_val(const struct cpumask
*mask
)
211 struct acpi_processor_performance
*perf
;
214 if (unlikely(cpumask_empty(mask
)))
217 switch (per_cpu(drv_data
, cpumask_first(mask
))->cpu_feature
) {
218 case SYSTEM_INTEL_MSR_CAPABLE
:
219 cmd
.type
= SYSTEM_INTEL_MSR_CAPABLE
;
220 cmd
.addr
.msr
.reg
= MSR_IA32_PERF_STATUS
;
222 case SYSTEM_IO_CAPABLE
:
223 cmd
.type
= SYSTEM_IO_CAPABLE
;
224 perf
= per_cpu(drv_data
, cpumask_first(mask
))->acpi_data
;
225 cmd
.addr
.io
.port
= perf
->control_register
.address
;
226 cmd
.addr
.io
.bit_width
= perf
->control_register
.bit_width
;
235 dprintk("get_cur_val = %u\n", cmd
.val
);
240 /* Called via smp_call_function_single(), on the target CPU */
241 static void read_measured_perf_ctrs(void *_cur
)
243 struct aperfmperf
*am
= _cur
;
249 * Return the measured active (C0) frequency on this CPU since last call
252 * Return: Average CPU frequency in terms of max frequency (zero on error)
254 * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
255 * over a period of time, while CPU is in C0 state.
256 * IA32_MPERF counts at the rate of max advertised frequency
257 * IA32_APERF counts at the rate of actual CPU frequency
258 * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
259 * no meaning should be associated with absolute values of these MSRs.
261 static unsigned int get_measured_perf(struct cpufreq_policy
*policy
,
264 struct aperfmperf perf
;
268 if (smp_call_function_single(cpu
, read_measured_perf_ctrs
, &perf
, 1))
271 ratio
= calc_aperfmperf_ratio(&per_cpu(old_perf
, cpu
), &perf
);
272 per_cpu(old_perf
, cpu
) = perf
;
274 retval
= (policy
->cpuinfo
.max_freq
* ratio
) >> APERFMPERF_SHIFT
;
279 static unsigned int get_cur_freq_on_cpu(unsigned int cpu
)
281 struct acpi_cpufreq_data
*data
= per_cpu(drv_data
, cpu
);
283 unsigned int cached_freq
;
285 dprintk("get_cur_freq_on_cpu (%d)\n", cpu
);
287 if (unlikely(data
== NULL
||
288 data
->acpi_data
== NULL
|| data
->freq_table
== NULL
)) {
292 cached_freq
= data
->freq_table
[data
->acpi_data
->state
].frequency
;
293 freq
= extract_freq(get_cur_val(cpumask_of(cpu
)), data
);
294 if (freq
!= cached_freq
) {
296 * The dreaded BIOS frequency change behind our back.
297 * Force set the frequency on next target call.
302 dprintk("cur freq = %u\n", freq
);
307 static unsigned int check_freqs(const struct cpumask
*mask
, unsigned int freq
,
308 struct acpi_cpufreq_data
*data
)
310 unsigned int cur_freq
;
313 for (i
= 0; i
< 100; i
++) {
314 cur_freq
= extract_freq(get_cur_val(mask
), data
);
315 if (cur_freq
== freq
)
322 static int acpi_cpufreq_target(struct cpufreq_policy
*policy
,
323 unsigned int target_freq
, unsigned int relation
)
325 struct acpi_cpufreq_data
*data
= per_cpu(drv_data
, policy
->cpu
);
326 struct acpi_processor_performance
*perf
;
327 struct cpufreq_freqs freqs
;
329 unsigned int next_state
= 0; /* Index into freq_table */
330 unsigned int next_perf_state
= 0; /* Index into perf table */
334 dprintk("acpi_cpufreq_target %d (%d)\n", target_freq
, policy
->cpu
);
336 if (unlikely(data
== NULL
||
337 data
->acpi_data
== NULL
|| data
->freq_table
== NULL
)) {
341 perf
= data
->acpi_data
;
342 result
= cpufreq_frequency_table_target(policy
,
345 relation
, &next_state
);
346 if (unlikely(result
)) {
351 next_perf_state
= data
->freq_table
[next_state
].index
;
352 if (perf
->state
== next_perf_state
) {
353 if (unlikely(data
->resume
)) {
354 dprintk("Called after resume, resetting to P%d\n",
358 dprintk("Already at target state (P%d)\n",
364 trace_power_frequency(POWER_PSTATE
, data
->freq_table
[next_state
].frequency
);
366 switch (data
->cpu_feature
) {
367 case SYSTEM_INTEL_MSR_CAPABLE
:
368 cmd
.type
= SYSTEM_INTEL_MSR_CAPABLE
;
369 cmd
.addr
.msr
.reg
= MSR_IA32_PERF_CTL
;
370 cmd
.val
= (u32
) perf
->states
[next_perf_state
].control
;
372 case SYSTEM_IO_CAPABLE
:
373 cmd
.type
= SYSTEM_IO_CAPABLE
;
374 cmd
.addr
.io
.port
= perf
->control_register
.address
;
375 cmd
.addr
.io
.bit_width
= perf
->control_register
.bit_width
;
376 cmd
.val
= (u32
) perf
->states
[next_perf_state
].control
;
383 /* cpufreq holds the hotplug lock, so we are safe from here on */
384 if (policy
->shared_type
!= CPUFREQ_SHARED_TYPE_ANY
)
385 cmd
.mask
= policy
->cpus
;
387 cmd
.mask
= cpumask_of(policy
->cpu
);
389 freqs
.old
= perf
->states
[perf
->state
].core_frequency
* 1000;
390 freqs
.new = data
->freq_table
[next_state
].frequency
;
391 for_each_cpu(i
, cmd
.mask
) {
393 cpufreq_notify_transition(&freqs
, CPUFREQ_PRECHANGE
);
398 if (acpi_pstate_strict
) {
399 if (!check_freqs(cmd
.mask
, freqs
.new, data
)) {
400 dprintk("acpi_cpufreq_target failed (%d)\n",
407 for_each_cpu(i
, cmd
.mask
) {
409 cpufreq_notify_transition(&freqs
, CPUFREQ_POSTCHANGE
);
411 perf
->state
= next_perf_state
;
417 static int acpi_cpufreq_verify(struct cpufreq_policy
*policy
)
419 struct acpi_cpufreq_data
*data
= per_cpu(drv_data
, policy
->cpu
);
421 dprintk("acpi_cpufreq_verify\n");
423 return cpufreq_frequency_table_verify(policy
, data
->freq_table
);
427 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data
*data
, unsigned int cpu
)
429 struct acpi_processor_performance
*perf
= data
->acpi_data
;
432 /* search the closest match to cpu_khz */
435 unsigned long freqn
= perf
->states
[0].core_frequency
* 1000;
437 for (i
= 0; i
< (perf
->state_count
-1); i
++) {
439 freqn
= perf
->states
[i
+1].core_frequency
* 1000;
440 if ((2 * cpu_khz
) > (freqn
+ freq
)) {
445 perf
->state
= perf
->state_count
-1;
448 /* assume CPU is at P0... */
450 return perf
->states
[0].core_frequency
* 1000;
454 static void free_acpi_perf_data(void)
458 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
459 for_each_possible_cpu(i
)
460 free_cpumask_var(per_cpu_ptr(acpi_perf_data
, i
)
462 free_percpu(acpi_perf_data
);
466 * acpi_cpufreq_early_init - initialize ACPI P-States library
468 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
469 * in order to determine correct frequency and voltage pairings. We can
470 * do _PDC and _PSD and find out the processor dependency for the
471 * actual init that will happen later...
473 static int __init
acpi_cpufreq_early_init(void)
476 dprintk("acpi_cpufreq_early_init\n");
478 acpi_perf_data
= alloc_percpu(struct acpi_processor_performance
);
479 if (!acpi_perf_data
) {
480 dprintk("Memory allocation error for acpi_perf_data.\n");
483 for_each_possible_cpu(i
) {
484 if (!zalloc_cpumask_var_node(
485 &per_cpu_ptr(acpi_perf_data
, i
)->shared_cpu_map
,
486 GFP_KERNEL
, cpu_to_node(i
))) {
488 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
489 free_acpi_perf_data();
494 /* Do initialization in ACPI core */
495 acpi_processor_preregister_performance(acpi_perf_data
);
501 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
502 * or do it in BIOS firmware and won't inform about it to OS. If not
503 * detected, this has a side effect of making CPU run at a different speed
504 * than OS intended it to run at. Detect it and handle it cleanly.
506 static int bios_with_sw_any_bug
;
508 static int sw_any_bug_found(const struct dmi_system_id
*d
)
510 bios_with_sw_any_bug
= 1;
514 static const struct dmi_system_id sw_any_bug_dmi_table
[] = {
516 .callback
= sw_any_bug_found
,
517 .ident
= "Supermicro Server X6DLP",
519 DMI_MATCH(DMI_SYS_VENDOR
, "Supermicro"),
520 DMI_MATCH(DMI_BIOS_VERSION
, "080010"),
521 DMI_MATCH(DMI_PRODUCT_NAME
, "X6DLP"),
527 static int acpi_cpufreq_blacklist(struct cpuinfo_x86
*c
)
529 /* http://www.intel.com/Assets/PDF/specupdate/314554.pdf
530 * AL30: A Machine Check Exception (MCE) Occurring during an
531 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
532 * Both Processor Cores to Lock Up when HT is enabled*/
533 if (c
->x86_vendor
== X86_VENDOR_INTEL
) {
534 if ((c
->x86
== 15) &&
535 (c
->x86_model
== 6) &&
536 (c
->x86_mask
== 8) && smt_capable())
543 static int acpi_cpufreq_cpu_init(struct cpufreq_policy
*policy
)
546 unsigned int valid_states
= 0;
547 unsigned int cpu
= policy
->cpu
;
548 struct acpi_cpufreq_data
*data
;
549 unsigned int result
= 0;
550 struct cpuinfo_x86
*c
= &cpu_data(policy
->cpu
);
551 struct acpi_processor_performance
*perf
;
553 dprintk("acpi_cpufreq_cpu_init\n");
556 result
= acpi_cpufreq_blacklist(c
);
561 data
= kzalloc(sizeof(struct acpi_cpufreq_data
), GFP_KERNEL
);
565 data
->acpi_data
= per_cpu_ptr(acpi_perf_data
, cpu
);
566 per_cpu(drv_data
, cpu
) = data
;
568 if (cpu_has(c
, X86_FEATURE_CONSTANT_TSC
))
569 acpi_cpufreq_driver
.flags
|= CPUFREQ_CONST_LOOPS
;
571 result
= acpi_processor_register_performance(data
->acpi_data
, cpu
);
575 perf
= data
->acpi_data
;
576 policy
->shared_type
= perf
->shared_type
;
579 * Will let policy->cpus know about dependency only when software
580 * coordination is required.
582 if (policy
->shared_type
== CPUFREQ_SHARED_TYPE_ALL
||
583 policy
->shared_type
== CPUFREQ_SHARED_TYPE_ANY
) {
584 cpumask_copy(policy
->cpus
, perf
->shared_cpu_map
);
586 cpumask_copy(policy
->related_cpus
, perf
->shared_cpu_map
);
589 dmi_check_system(sw_any_bug_dmi_table
);
590 if (bios_with_sw_any_bug
&& cpumask_weight(policy
->cpus
) == 1) {
591 policy
->shared_type
= CPUFREQ_SHARED_TYPE_ALL
;
592 cpumask_copy(policy
->cpus
, cpu_core_mask(cpu
));
596 /* capability check */
597 if (perf
->state_count
<= 1) {
598 dprintk("No P-States\n");
603 if (perf
->control_register
.space_id
!= perf
->status_register
.space_id
) {
608 switch (perf
->control_register
.space_id
) {
609 case ACPI_ADR_SPACE_SYSTEM_IO
:
610 dprintk("SYSTEM IO addr space\n");
611 data
->cpu_feature
= SYSTEM_IO_CAPABLE
;
613 case ACPI_ADR_SPACE_FIXED_HARDWARE
:
614 dprintk("HARDWARE addr space\n");
615 if (!check_est_cpu(cpu
)) {
619 data
->cpu_feature
= SYSTEM_INTEL_MSR_CAPABLE
;
622 dprintk("Unknown addr space %d\n",
623 (u32
) (perf
->control_register
.space_id
));
628 data
->freq_table
= kmalloc(sizeof(struct cpufreq_frequency_table
) *
629 (perf
->state_count
+1), GFP_KERNEL
);
630 if (!data
->freq_table
) {
635 /* detect transition latency */
636 policy
->cpuinfo
.transition_latency
= 0;
637 for (i
= 0; i
< perf
->state_count
; i
++) {
638 if ((perf
->states
[i
].transition_latency
* 1000) >
639 policy
->cpuinfo
.transition_latency
)
640 policy
->cpuinfo
.transition_latency
=
641 perf
->states
[i
].transition_latency
* 1000;
644 /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
645 if (perf
->control_register
.space_id
== ACPI_ADR_SPACE_FIXED_HARDWARE
&&
646 policy
->cpuinfo
.transition_latency
> 20 * 1000) {
647 policy
->cpuinfo
.transition_latency
= 20 * 1000;
648 printk_once(KERN_INFO
649 "P-state transition latency capped at 20 uS\n");
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 if (perf
->states
[0].core_frequency
* 1000 != policy
->cpuinfo
.max_freq
)
671 printk(KERN_WARNING FW_WARN
"P-state 0 is not max freq\n");
673 switch (perf
->control_register
.space_id
) {
674 case ACPI_ADR_SPACE_SYSTEM_IO
:
675 /* Current speed is unknown and not detectable by IO port */
676 policy
->cur
= acpi_cpufreq_guess_freq(data
, policy
->cpu
);
678 case ACPI_ADR_SPACE_FIXED_HARDWARE
:
679 acpi_cpufreq_driver
.get
= get_cur_freq_on_cpu
;
680 policy
->cur
= get_cur_freq_on_cpu(cpu
);
686 /* notify BIOS that we exist */
687 acpi_processor_notify_smm(THIS_MODULE
);
689 /* Check for APERF/MPERF support in hardware */
690 if (cpu_has(c
, X86_FEATURE_APERFMPERF
))
691 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 per_cpu(drv_data
, cpu
) = NULL
;
722 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy
*policy
)
724 struct acpi_cpufreq_data
*data
= per_cpu(drv_data
, policy
->cpu
);
726 dprintk("acpi_cpufreq_cpu_exit\n");
729 cpufreq_frequency_table_put_attr(policy
->cpu
);
730 per_cpu(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
= per_cpu(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)
773 dprintk("acpi_cpufreq_init\n");
775 ret
= acpi_cpufreq_early_init();
779 ret
= cpufreq_register_driver(&acpi_cpufreq_driver
);
781 free_acpi_perf_data();
786 static void __exit
acpi_cpufreq_exit(void)
788 dprintk("acpi_cpufreq_exit\n");
790 cpufreq_unregister_driver(&acpi_cpufreq_driver
);
792 free_percpu(acpi_perf_data
);
795 module_param(acpi_pstate_strict
, uint
, 0644);
796 MODULE_PARM_DESC(acpi_pstate_strict
,
797 "value 0 or non-zero. non-zero -> strict ACPI checks are "
798 "performed during frequency changes.");
800 late_initcall(acpi_cpufreq_init
);
801 module_exit(acpi_cpufreq_exit
);
803 MODULE_ALIAS("acpi");