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 <linux/slab.h>
37 #include <trace/events/power.h>
39 #include <linux/acpi.h>
41 #include <linux/delay.h>
42 #include <linux/uaccess.h>
44 #include <acpi/processor.h>
47 #include <asm/processor.h>
48 #include <asm/cpufeature.h>
51 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, \
54 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
55 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
56 MODULE_LICENSE("GPL");
59 UNDEFINED_CAPABLE
= 0,
60 SYSTEM_INTEL_MSR_CAPABLE
,
64 #define INTEL_MSR_RANGE (0xffff)
66 struct acpi_cpufreq_data
{
67 struct acpi_processor_performance
*acpi_data
;
68 struct cpufreq_frequency_table
*freq_table
;
70 unsigned int cpu_feature
;
73 static DEFINE_PER_CPU(struct acpi_cpufreq_data
*, acfreq_data
);
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
)
196 err
= smp_call_function_any(cmd
->mask
, do_drv_read
, cmd
, 1);
197 WARN_ON_ONCE(err
); /* smp_call_function_any() was buggy? */
200 static void drv_write(struct drv_cmd
*cmd
)
204 this_cpu
= get_cpu();
205 if (cpumask_test_cpu(this_cpu
, cmd
->mask
))
207 smp_call_function_many(cmd
->mask
, do_drv_write
, cmd
, 1);
211 static u32
get_cur_val(const struct cpumask
*mask
)
213 struct acpi_processor_performance
*perf
;
216 if (unlikely(cpumask_empty(mask
)))
219 switch (per_cpu(acfreq_data
, cpumask_first(mask
))->cpu_feature
) {
220 case SYSTEM_INTEL_MSR_CAPABLE
:
221 cmd
.type
= SYSTEM_INTEL_MSR_CAPABLE
;
222 cmd
.addr
.msr
.reg
= MSR_IA32_PERF_STATUS
;
224 case SYSTEM_IO_CAPABLE
:
225 cmd
.type
= SYSTEM_IO_CAPABLE
;
226 perf
= per_cpu(acfreq_data
, cpumask_first(mask
))->acpi_data
;
227 cmd
.addr
.io
.port
= perf
->control_register
.address
;
228 cmd
.addr
.io
.bit_width
= perf
->control_register
.bit_width
;
237 dprintk("get_cur_val = %u\n", cmd
.val
);
242 static unsigned int get_cur_freq_on_cpu(unsigned int cpu
)
244 struct acpi_cpufreq_data
*data
= per_cpu(acfreq_data
, cpu
);
246 unsigned int cached_freq
;
248 dprintk("get_cur_freq_on_cpu (%d)\n", cpu
);
250 if (unlikely(data
== NULL
||
251 data
->acpi_data
== NULL
|| data
->freq_table
== NULL
)) {
255 cached_freq
= data
->freq_table
[data
->acpi_data
->state
].frequency
;
256 freq
= extract_freq(get_cur_val(cpumask_of(cpu
)), data
);
257 if (freq
!= cached_freq
) {
259 * The dreaded BIOS frequency change behind our back.
260 * Force set the frequency on next target call.
265 dprintk("cur freq = %u\n", freq
);
270 static unsigned int check_freqs(const struct cpumask
*mask
, unsigned int freq
,
271 struct acpi_cpufreq_data
*data
)
273 unsigned int cur_freq
;
276 for (i
= 0; i
< 100; i
++) {
277 cur_freq
= extract_freq(get_cur_val(mask
), data
);
278 if (cur_freq
== freq
)
285 static int acpi_cpufreq_target(struct cpufreq_policy
*policy
,
286 unsigned int target_freq
, unsigned int relation
)
288 struct acpi_cpufreq_data
*data
= per_cpu(acfreq_data
, policy
->cpu
);
289 struct acpi_processor_performance
*perf
;
290 struct cpufreq_freqs freqs
;
292 unsigned int next_state
= 0; /* Index into freq_table */
293 unsigned int next_perf_state
= 0; /* Index into perf table */
297 dprintk("acpi_cpufreq_target %d (%d)\n", target_freq
, policy
->cpu
);
299 if (unlikely(data
== NULL
||
300 data
->acpi_data
== NULL
|| data
->freq_table
== NULL
)) {
304 perf
= data
->acpi_data
;
305 result
= cpufreq_frequency_table_target(policy
,
308 relation
, &next_state
);
309 if (unlikely(result
)) {
314 next_perf_state
= data
->freq_table
[next_state
].index
;
315 if (perf
->state
== next_perf_state
) {
316 if (unlikely(data
->resume
)) {
317 dprintk("Called after resume, resetting to P%d\n",
321 dprintk("Already at target state (P%d)\n",
327 trace_power_frequency(POWER_PSTATE
, data
->freq_table
[next_state
].frequency
);
329 switch (data
->cpu_feature
) {
330 case SYSTEM_INTEL_MSR_CAPABLE
:
331 cmd
.type
= SYSTEM_INTEL_MSR_CAPABLE
;
332 cmd
.addr
.msr
.reg
= MSR_IA32_PERF_CTL
;
333 cmd
.val
= (u32
) perf
->states
[next_perf_state
].control
;
335 case SYSTEM_IO_CAPABLE
:
336 cmd
.type
= SYSTEM_IO_CAPABLE
;
337 cmd
.addr
.io
.port
= perf
->control_register
.address
;
338 cmd
.addr
.io
.bit_width
= perf
->control_register
.bit_width
;
339 cmd
.val
= (u32
) perf
->states
[next_perf_state
].control
;
346 /* cpufreq holds the hotplug lock, so we are safe from here on */
347 if (policy
->shared_type
!= CPUFREQ_SHARED_TYPE_ANY
)
348 cmd
.mask
= policy
->cpus
;
350 cmd
.mask
= cpumask_of(policy
->cpu
);
352 freqs
.old
= perf
->states
[perf
->state
].core_frequency
* 1000;
353 freqs
.new = data
->freq_table
[next_state
].frequency
;
354 for_each_cpu(i
, policy
->cpus
) {
356 cpufreq_notify_transition(&freqs
, CPUFREQ_PRECHANGE
);
361 if (acpi_pstate_strict
) {
362 if (!check_freqs(cmd
.mask
, freqs
.new, data
)) {
363 dprintk("acpi_cpufreq_target failed (%d)\n",
370 for_each_cpu(i
, policy
->cpus
) {
372 cpufreq_notify_transition(&freqs
, CPUFREQ_POSTCHANGE
);
374 perf
->state
= next_perf_state
;
380 static int acpi_cpufreq_verify(struct cpufreq_policy
*policy
)
382 struct acpi_cpufreq_data
*data
= per_cpu(acfreq_data
, policy
->cpu
);
384 dprintk("acpi_cpufreq_verify\n");
386 return cpufreq_frequency_table_verify(policy
, data
->freq_table
);
390 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data
*data
, unsigned int cpu
)
392 struct acpi_processor_performance
*perf
= data
->acpi_data
;
395 /* search the closest match to cpu_khz */
398 unsigned long freqn
= perf
->states
[0].core_frequency
* 1000;
400 for (i
= 0; i
< (perf
->state_count
-1); i
++) {
402 freqn
= perf
->states
[i
+1].core_frequency
* 1000;
403 if ((2 * cpu_khz
) > (freqn
+ freq
)) {
408 perf
->state
= perf
->state_count
-1;
411 /* assume CPU is at P0... */
413 return perf
->states
[0].core_frequency
* 1000;
417 static void free_acpi_perf_data(void)
421 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
422 for_each_possible_cpu(i
)
423 free_cpumask_var(per_cpu_ptr(acpi_perf_data
, i
)
425 free_percpu(acpi_perf_data
);
429 * acpi_cpufreq_early_init - initialize ACPI P-States library
431 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
432 * in order to determine correct frequency and voltage pairings. We can
433 * do _PDC and _PSD and find out the processor dependency for the
434 * actual init that will happen later...
436 static int __init
acpi_cpufreq_early_init(void)
439 dprintk("acpi_cpufreq_early_init\n");
441 acpi_perf_data
= alloc_percpu(struct acpi_processor_performance
);
442 if (!acpi_perf_data
) {
443 dprintk("Memory allocation error for acpi_perf_data.\n");
446 for_each_possible_cpu(i
) {
447 if (!zalloc_cpumask_var_node(
448 &per_cpu_ptr(acpi_perf_data
, i
)->shared_cpu_map
,
449 GFP_KERNEL
, cpu_to_node(i
))) {
451 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
452 free_acpi_perf_data();
457 /* Do initialization in ACPI core */
458 acpi_processor_preregister_performance(acpi_perf_data
);
464 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
465 * or do it in BIOS firmware and won't inform about it to OS. If not
466 * detected, this has a side effect of making CPU run at a different speed
467 * than OS intended it to run at. Detect it and handle it cleanly.
469 static int bios_with_sw_any_bug
;
471 static int sw_any_bug_found(const struct dmi_system_id
*d
)
473 bios_with_sw_any_bug
= 1;
477 static const struct dmi_system_id sw_any_bug_dmi_table
[] = {
479 .callback
= sw_any_bug_found
,
480 .ident
= "Supermicro Server X6DLP",
482 DMI_MATCH(DMI_SYS_VENDOR
, "Supermicro"),
483 DMI_MATCH(DMI_BIOS_VERSION
, "080010"),
484 DMI_MATCH(DMI_PRODUCT_NAME
, "X6DLP"),
490 static int acpi_cpufreq_blacklist(struct cpuinfo_x86
*c
)
492 /* Intel Xeon Processor 7100 Series Specification Update
493 * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
494 * AL30: A Machine Check Exception (MCE) Occurring during an
495 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
496 * Both Processor Cores to Lock Up. */
497 if (c
->x86_vendor
== X86_VENDOR_INTEL
) {
498 if ((c
->x86
== 15) &&
499 (c
->x86_model
== 6) &&
500 (c
->x86_mask
== 8)) {
501 printk(KERN_INFO
"acpi-cpufreq: Intel(R) "
502 "Xeon(R) 7100 Errata AL30, processors may "
503 "lock up on frequency changes: disabling "
512 static int acpi_cpufreq_cpu_init(struct cpufreq_policy
*policy
)
515 unsigned int valid_states
= 0;
516 unsigned int cpu
= policy
->cpu
;
517 struct acpi_cpufreq_data
*data
;
518 unsigned int result
= 0;
519 struct cpuinfo_x86
*c
= &cpu_data(policy
->cpu
);
520 struct acpi_processor_performance
*perf
;
522 static int blacklisted
;
525 dprintk("acpi_cpufreq_cpu_init\n");
530 blacklisted
= acpi_cpufreq_blacklist(c
);
535 data
= kzalloc(sizeof(struct acpi_cpufreq_data
), GFP_KERNEL
);
539 data
->acpi_data
= per_cpu_ptr(acpi_perf_data
, cpu
);
540 per_cpu(acfreq_data
, cpu
) = data
;
542 if (cpu_has(c
, X86_FEATURE_CONSTANT_TSC
))
543 acpi_cpufreq_driver
.flags
|= CPUFREQ_CONST_LOOPS
;
545 result
= acpi_processor_register_performance(data
->acpi_data
, cpu
);
549 perf
= data
->acpi_data
;
550 policy
->shared_type
= perf
->shared_type
;
553 * Will let policy->cpus know about dependency only when software
554 * coordination is required.
556 if (policy
->shared_type
== CPUFREQ_SHARED_TYPE_ALL
||
557 policy
->shared_type
== CPUFREQ_SHARED_TYPE_ANY
) {
558 cpumask_copy(policy
->cpus
, perf
->shared_cpu_map
);
560 cpumask_copy(policy
->related_cpus
, perf
->shared_cpu_map
);
563 dmi_check_system(sw_any_bug_dmi_table
);
564 if (bios_with_sw_any_bug
&& cpumask_weight(policy
->cpus
) == 1) {
565 policy
->shared_type
= CPUFREQ_SHARED_TYPE_ALL
;
566 cpumask_copy(policy
->cpus
, cpu_core_mask(cpu
));
570 /* capability check */
571 if (perf
->state_count
<= 1) {
572 dprintk("No P-States\n");
577 if (perf
->control_register
.space_id
!= perf
->status_register
.space_id
) {
582 switch (perf
->control_register
.space_id
) {
583 case ACPI_ADR_SPACE_SYSTEM_IO
:
584 dprintk("SYSTEM IO addr space\n");
585 data
->cpu_feature
= SYSTEM_IO_CAPABLE
;
587 case ACPI_ADR_SPACE_FIXED_HARDWARE
:
588 dprintk("HARDWARE addr space\n");
589 if (!check_est_cpu(cpu
)) {
593 data
->cpu_feature
= SYSTEM_INTEL_MSR_CAPABLE
;
596 dprintk("Unknown addr space %d\n",
597 (u32
) (perf
->control_register
.space_id
));
602 data
->freq_table
= kmalloc(sizeof(struct cpufreq_frequency_table
) *
603 (perf
->state_count
+1), GFP_KERNEL
);
604 if (!data
->freq_table
) {
609 /* detect transition latency */
610 policy
->cpuinfo
.transition_latency
= 0;
611 for (i
= 0; i
< perf
->state_count
; i
++) {
612 if ((perf
->states
[i
].transition_latency
* 1000) >
613 policy
->cpuinfo
.transition_latency
)
614 policy
->cpuinfo
.transition_latency
=
615 perf
->states
[i
].transition_latency
* 1000;
618 /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
619 if (perf
->control_register
.space_id
== ACPI_ADR_SPACE_FIXED_HARDWARE
&&
620 policy
->cpuinfo
.transition_latency
> 20 * 1000) {
621 policy
->cpuinfo
.transition_latency
= 20 * 1000;
622 printk_once(KERN_INFO
623 "P-state transition latency capped at 20 uS\n");
627 for (i
= 0; i
< perf
->state_count
; i
++) {
628 if (i
> 0 && perf
->states
[i
].core_frequency
>=
629 data
->freq_table
[valid_states
-1].frequency
/ 1000)
632 data
->freq_table
[valid_states
].index
= i
;
633 data
->freq_table
[valid_states
].frequency
=
634 perf
->states
[i
].core_frequency
* 1000;
637 data
->freq_table
[valid_states
].frequency
= CPUFREQ_TABLE_END
;
640 result
= cpufreq_frequency_table_cpuinfo(policy
, data
->freq_table
);
644 if (perf
->states
[0].core_frequency
* 1000 != policy
->cpuinfo
.max_freq
)
645 printk(KERN_WARNING FW_WARN
"P-state 0 is not max freq\n");
647 switch (perf
->control_register
.space_id
) {
648 case ACPI_ADR_SPACE_SYSTEM_IO
:
649 /* Current speed is unknown and not detectable by IO port */
650 policy
->cur
= acpi_cpufreq_guess_freq(data
, policy
->cpu
);
652 case ACPI_ADR_SPACE_FIXED_HARDWARE
:
653 acpi_cpufreq_driver
.get
= get_cur_freq_on_cpu
;
654 policy
->cur
= get_cur_freq_on_cpu(cpu
);
660 /* notify BIOS that we exist */
661 acpi_processor_notify_smm(THIS_MODULE
);
663 /* Check for APERF/MPERF support in hardware */
664 if (cpu_has(c
, X86_FEATURE_APERFMPERF
))
665 acpi_cpufreq_driver
.getavg
= cpufreq_get_measured_perf
;
667 dprintk("CPU%u - ACPI performance management activated.\n", cpu
);
668 for (i
= 0; i
< perf
->state_count
; i
++)
669 dprintk(" %cP%d: %d MHz, %d mW, %d uS\n",
670 (i
== perf
->state
? '*' : ' '), i
,
671 (u32
) perf
->states
[i
].core_frequency
,
672 (u32
) perf
->states
[i
].power
,
673 (u32
) perf
->states
[i
].transition_latency
);
675 cpufreq_frequency_table_get_attr(data
->freq_table
, policy
->cpu
);
678 * the first call to ->target() should result in us actually
679 * writing something to the appropriate registers.
686 kfree(data
->freq_table
);
688 acpi_processor_unregister_performance(perf
, cpu
);
691 per_cpu(acfreq_data
, cpu
) = NULL
;
696 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy
*policy
)
698 struct acpi_cpufreq_data
*data
= per_cpu(acfreq_data
, policy
->cpu
);
700 dprintk("acpi_cpufreq_cpu_exit\n");
703 cpufreq_frequency_table_put_attr(policy
->cpu
);
704 per_cpu(acfreq_data
, policy
->cpu
) = NULL
;
705 acpi_processor_unregister_performance(data
->acpi_data
,
713 static int acpi_cpufreq_resume(struct cpufreq_policy
*policy
)
715 struct acpi_cpufreq_data
*data
= per_cpu(acfreq_data
, policy
->cpu
);
717 dprintk("acpi_cpufreq_resume\n");
724 static struct freq_attr
*acpi_cpufreq_attr
[] = {
725 &cpufreq_freq_attr_scaling_available_freqs
,
729 static struct cpufreq_driver acpi_cpufreq_driver
= {
730 .verify
= acpi_cpufreq_verify
,
731 .target
= acpi_cpufreq_target
,
732 .bios_limit
= acpi_processor_get_bios_limit
,
733 .init
= acpi_cpufreq_cpu_init
,
734 .exit
= acpi_cpufreq_cpu_exit
,
735 .resume
= acpi_cpufreq_resume
,
736 .name
= "acpi-cpufreq",
737 .owner
= THIS_MODULE
,
738 .attr
= acpi_cpufreq_attr
,
741 static int __init
acpi_cpufreq_init(void)
748 dprintk("acpi_cpufreq_init\n");
750 ret
= acpi_cpufreq_early_init();
754 ret
= cpufreq_register_driver(&acpi_cpufreq_driver
);
756 free_acpi_perf_data();
761 static void __exit
acpi_cpufreq_exit(void)
763 dprintk("acpi_cpufreq_exit\n");
765 cpufreq_unregister_driver(&acpi_cpufreq_driver
);
767 free_percpu(acpi_perf_data
);
770 module_param(acpi_pstate_strict
, uint
, 0644);
771 MODULE_PARM_DESC(acpi_pstate_strict
,
772 "value 0 or non-zero. non-zero -> strict ACPI checks are "
773 "performed during frequency changes.");
775 late_initcall(acpi_cpufreq_init
);
776 module_exit(acpi_cpufreq_exit
);
778 MODULE_ALIAS("acpi");