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/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 DEFINE_TRACE(power_mark
);
77 /* acpi_perf_data is a pointer to percpu data. */
78 static struct acpi_processor_performance
*acpi_perf_data
;
80 static struct cpufreq_driver acpi_cpufreq_driver
;
82 static unsigned int acpi_pstate_strict
;
84 static int check_est_cpu(unsigned int cpuid
)
86 struct cpuinfo_x86
*cpu
= &cpu_data(cpuid
);
88 return cpu_has(cpu
, X86_FEATURE_EST
);
91 static unsigned extract_io(u32 value
, struct acpi_cpufreq_data
*data
)
93 struct acpi_processor_performance
*perf
;
96 perf
= data
->acpi_data
;
98 for (i
= 0; i
< perf
->state_count
; i
++) {
99 if (value
== perf
->states
[i
].status
)
100 return data
->freq_table
[i
].frequency
;
105 static unsigned extract_msr(u32 msr
, struct acpi_cpufreq_data
*data
)
108 struct acpi_processor_performance
*perf
;
110 msr
&= INTEL_MSR_RANGE
;
111 perf
= data
->acpi_data
;
113 for (i
= 0; data
->freq_table
[i
].frequency
!= CPUFREQ_TABLE_END
; i
++) {
114 if (msr
== perf
->states
[data
->freq_table
[i
].index
].status
)
115 return data
->freq_table
[i
].frequency
;
117 return data
->freq_table
[0].frequency
;
120 static unsigned extract_freq(u32 val
, struct acpi_cpufreq_data
*data
)
122 switch (data
->cpu_feature
) {
123 case SYSTEM_INTEL_MSR_CAPABLE
:
124 return extract_msr(val
, data
);
125 case SYSTEM_IO_CAPABLE
:
126 return extract_io(val
, data
);
143 const struct cpumask
*mask
;
151 /* Called via smp_call_function_single(), on the target CPU */
152 static void do_drv_read(void *_cmd
)
154 struct drv_cmd
*cmd
= _cmd
;
158 case SYSTEM_INTEL_MSR_CAPABLE
:
159 rdmsr(cmd
->addr
.msr
.reg
, cmd
->val
, h
);
161 case SYSTEM_IO_CAPABLE
:
162 acpi_os_read_port((acpi_io_address
)cmd
->addr
.io
.port
,
164 (u32
)cmd
->addr
.io
.bit_width
);
171 /* Called via smp_call_function_many(), on the target CPUs */
172 static void do_drv_write(void *_cmd
)
174 struct drv_cmd
*cmd
= _cmd
;
178 case SYSTEM_INTEL_MSR_CAPABLE
:
179 rdmsr(cmd
->addr
.msr
.reg
, lo
, hi
);
180 lo
= (lo
& ~INTEL_MSR_RANGE
) | (cmd
->val
& INTEL_MSR_RANGE
);
181 wrmsr(cmd
->addr
.msr
.reg
, lo
, hi
);
183 case SYSTEM_IO_CAPABLE
:
184 acpi_os_write_port((acpi_io_address
)cmd
->addr
.io
.port
,
186 (u32
)cmd
->addr
.io
.bit_width
);
193 static void drv_read(struct drv_cmd
*cmd
)
197 smp_call_function_single(cpumask_any(cmd
->mask
), do_drv_read
, cmd
, 1);
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(drv_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(drv_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 /* Called via smp_call_function_single(), on the target CPU */
243 static void read_measured_perf_ctrs(void *_cur
)
245 struct aperfmperf
*am
= _cur
;
251 * Return the measured active (C0) frequency on this CPU since last call
254 * Return: Average CPU frequency in terms of max frequency (zero on error)
256 * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
257 * over a period of time, while CPU is in C0 state.
258 * IA32_MPERF counts at the rate of max advertised frequency
259 * IA32_APERF counts at the rate of actual CPU frequency
260 * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
261 * no meaning should be associated with absolute values of these MSRs.
263 static unsigned int get_measured_perf(struct cpufreq_policy
*policy
,
266 struct aperfmperf perf
;
270 if (smp_call_function_single(cpu
, read_measured_perf_ctrs
, &perf
, 1))
273 ratio
= calc_aperfmperf_ratio(&per_cpu(old_perf
, cpu
), &perf
);
274 per_cpu(old_perf
, cpu
) = perf
;
276 retval
= (policy
->cpuinfo
.max_freq
* ratio
) >> APERFMPERF_SHIFT
;
281 static unsigned int get_cur_freq_on_cpu(unsigned int cpu
)
283 struct acpi_cpufreq_data
*data
= per_cpu(drv_data
, cpu
);
285 unsigned int cached_freq
;
287 dprintk("get_cur_freq_on_cpu (%d)\n", cpu
);
289 if (unlikely(data
== NULL
||
290 data
->acpi_data
== NULL
|| data
->freq_table
== NULL
)) {
294 cached_freq
= data
->freq_table
[data
->acpi_data
->state
].frequency
;
295 freq
= extract_freq(get_cur_val(cpumask_of(cpu
)), data
);
296 if (freq
!= cached_freq
) {
298 * The dreaded BIOS frequency change behind our back.
299 * Force set the frequency on next target call.
304 dprintk("cur freq = %u\n", freq
);
309 static unsigned int check_freqs(const struct cpumask
*mask
, unsigned int freq
,
310 struct acpi_cpufreq_data
*data
)
312 unsigned int cur_freq
;
315 for (i
= 0; i
< 100; i
++) {
316 cur_freq
= extract_freq(get_cur_val(mask
), data
);
317 if (cur_freq
== freq
)
324 static int acpi_cpufreq_target(struct cpufreq_policy
*policy
,
325 unsigned int target_freq
, unsigned int relation
)
327 struct acpi_cpufreq_data
*data
= per_cpu(drv_data
, policy
->cpu
);
328 struct acpi_processor_performance
*perf
;
329 struct cpufreq_freqs freqs
;
331 unsigned int next_state
= 0; /* Index into freq_table */
332 unsigned int next_perf_state
= 0; /* Index into perf table */
335 struct power_trace it
;
337 dprintk("acpi_cpufreq_target %d (%d)\n", target_freq
, policy
->cpu
);
339 if (unlikely(data
== NULL
||
340 data
->acpi_data
== NULL
|| data
->freq_table
== NULL
)) {
344 perf
= data
->acpi_data
;
345 result
= cpufreq_frequency_table_target(policy
,
348 relation
, &next_state
);
349 if (unlikely(result
)) {
354 next_perf_state
= data
->freq_table
[next_state
].index
;
355 if (perf
->state
== next_perf_state
) {
356 if (unlikely(data
->resume
)) {
357 dprintk("Called after resume, resetting to P%d\n",
361 dprintk("Already at target state (P%d)\n",
367 trace_power_mark(&it
, POWER_PSTATE
, next_perf_state
);
369 switch (data
->cpu_feature
) {
370 case SYSTEM_INTEL_MSR_CAPABLE
:
371 cmd
.type
= SYSTEM_INTEL_MSR_CAPABLE
;
372 cmd
.addr
.msr
.reg
= MSR_IA32_PERF_CTL
;
373 cmd
.val
= (u32
) perf
->states
[next_perf_state
].control
;
375 case SYSTEM_IO_CAPABLE
:
376 cmd
.type
= SYSTEM_IO_CAPABLE
;
377 cmd
.addr
.io
.port
= perf
->control_register
.address
;
378 cmd
.addr
.io
.bit_width
= perf
->control_register
.bit_width
;
379 cmd
.val
= (u32
) perf
->states
[next_perf_state
].control
;
386 /* cpufreq holds the hotplug lock, so we are safe from here on */
387 if (policy
->shared_type
!= CPUFREQ_SHARED_TYPE_ANY
)
388 cmd
.mask
= policy
->cpus
;
390 cmd
.mask
= cpumask_of(policy
->cpu
);
392 freqs
.old
= perf
->states
[perf
->state
].core_frequency
* 1000;
393 freqs
.new = data
->freq_table
[next_state
].frequency
;
394 for_each_cpu(i
, cmd
.mask
) {
396 cpufreq_notify_transition(&freqs
, CPUFREQ_PRECHANGE
);
401 if (acpi_pstate_strict
) {
402 if (!check_freqs(cmd
.mask
, freqs
.new, data
)) {
403 dprintk("acpi_cpufreq_target failed (%d)\n",
410 for_each_cpu(i
, cmd
.mask
) {
412 cpufreq_notify_transition(&freqs
, CPUFREQ_POSTCHANGE
);
414 perf
->state
= next_perf_state
;
420 static int acpi_cpufreq_verify(struct cpufreq_policy
*policy
)
422 struct acpi_cpufreq_data
*data
= per_cpu(drv_data
, policy
->cpu
);
424 dprintk("acpi_cpufreq_verify\n");
426 return cpufreq_frequency_table_verify(policy
, data
->freq_table
);
430 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data
*data
, unsigned int cpu
)
432 struct acpi_processor_performance
*perf
= data
->acpi_data
;
435 /* search the closest match to cpu_khz */
438 unsigned long freqn
= perf
->states
[0].core_frequency
* 1000;
440 for (i
= 0; i
< (perf
->state_count
-1); i
++) {
442 freqn
= perf
->states
[i
+1].core_frequency
* 1000;
443 if ((2 * cpu_khz
) > (freqn
+ freq
)) {
448 perf
->state
= perf
->state_count
-1;
451 /* assume CPU is at P0... */
453 return perf
->states
[0].core_frequency
* 1000;
457 static void free_acpi_perf_data(void)
461 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
462 for_each_possible_cpu(i
)
463 free_cpumask_var(per_cpu_ptr(acpi_perf_data
, i
)
465 free_percpu(acpi_perf_data
);
469 * acpi_cpufreq_early_init - initialize ACPI P-States library
471 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
472 * in order to determine correct frequency and voltage pairings. We can
473 * do _PDC and _PSD and find out the processor dependency for the
474 * actual init that will happen later...
476 static int __init
acpi_cpufreq_early_init(void)
479 dprintk("acpi_cpufreq_early_init\n");
481 acpi_perf_data
= alloc_percpu(struct acpi_processor_performance
);
482 if (!acpi_perf_data
) {
483 dprintk("Memory allocation error for acpi_perf_data.\n");
486 for_each_possible_cpu(i
) {
487 if (!zalloc_cpumask_var_node(
488 &per_cpu_ptr(acpi_perf_data
, i
)->shared_cpu_map
,
489 GFP_KERNEL
, cpu_to_node(i
))) {
491 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
492 free_acpi_perf_data();
497 /* Do initialization in ACPI core */
498 acpi_processor_preregister_performance(acpi_perf_data
);
504 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
505 * or do it in BIOS firmware and won't inform about it to OS. If not
506 * detected, this has a side effect of making CPU run at a different speed
507 * than OS intended it to run at. Detect it and handle it cleanly.
509 static int bios_with_sw_any_bug
;
511 static int sw_any_bug_found(const struct dmi_system_id
*d
)
513 bios_with_sw_any_bug
= 1;
517 static const struct dmi_system_id sw_any_bug_dmi_table
[] = {
519 .callback
= sw_any_bug_found
,
520 .ident
= "Supermicro Server X6DLP",
522 DMI_MATCH(DMI_SYS_VENDOR
, "Supermicro"),
523 DMI_MATCH(DMI_BIOS_VERSION
, "080010"),
524 DMI_MATCH(DMI_PRODUCT_NAME
, "X6DLP"),
531 static int acpi_cpufreq_cpu_init(struct cpufreq_policy
*policy
)
534 unsigned int valid_states
= 0;
535 unsigned int cpu
= policy
->cpu
;
536 struct acpi_cpufreq_data
*data
;
537 unsigned int result
= 0;
538 struct cpuinfo_x86
*c
= &cpu_data(policy
->cpu
);
539 struct acpi_processor_performance
*perf
;
541 dprintk("acpi_cpufreq_cpu_init\n");
543 data
= kzalloc(sizeof(struct acpi_cpufreq_data
), GFP_KERNEL
);
547 data
->acpi_data
= per_cpu_ptr(acpi_perf_data
, cpu
);
548 per_cpu(drv_data
, cpu
) = data
;
550 if (cpu_has(c
, X86_FEATURE_CONSTANT_TSC
))
551 acpi_cpufreq_driver
.flags
|= CPUFREQ_CONST_LOOPS
;
553 result
= acpi_processor_register_performance(data
->acpi_data
, cpu
);
557 perf
= data
->acpi_data
;
558 policy
->shared_type
= perf
->shared_type
;
561 * Will let policy->cpus know about dependency only when software
562 * coordination is required.
564 if (policy
->shared_type
== CPUFREQ_SHARED_TYPE_ALL
||
565 policy
->shared_type
== CPUFREQ_SHARED_TYPE_ANY
) {
566 cpumask_copy(policy
->cpus
, perf
->shared_cpu_map
);
568 cpumask_copy(policy
->related_cpus
, perf
->shared_cpu_map
);
571 dmi_check_system(sw_any_bug_dmi_table
);
572 if (bios_with_sw_any_bug
&& cpumask_weight(policy
->cpus
) == 1) {
573 policy
->shared_type
= CPUFREQ_SHARED_TYPE_ALL
;
574 cpumask_copy(policy
->cpus
, cpu_core_mask(cpu
));
578 /* capability check */
579 if (perf
->state_count
<= 1) {
580 dprintk("No P-States\n");
585 if (perf
->control_register
.space_id
!= perf
->status_register
.space_id
) {
590 switch (perf
->control_register
.space_id
) {
591 case ACPI_ADR_SPACE_SYSTEM_IO
:
592 dprintk("SYSTEM IO addr space\n");
593 data
->cpu_feature
= SYSTEM_IO_CAPABLE
;
595 case ACPI_ADR_SPACE_FIXED_HARDWARE
:
596 dprintk("HARDWARE addr space\n");
597 if (!check_est_cpu(cpu
)) {
601 data
->cpu_feature
= SYSTEM_INTEL_MSR_CAPABLE
;
604 dprintk("Unknown addr space %d\n",
605 (u32
) (perf
->control_register
.space_id
));
610 data
->freq_table
= kmalloc(sizeof(struct cpufreq_frequency_table
) *
611 (perf
->state_count
+1), GFP_KERNEL
);
612 if (!data
->freq_table
) {
617 /* detect transition latency */
618 policy
->cpuinfo
.transition_latency
= 0;
619 for (i
= 0; i
< perf
->state_count
; i
++) {
620 if ((perf
->states
[i
].transition_latency
* 1000) >
621 policy
->cpuinfo
.transition_latency
)
622 policy
->cpuinfo
.transition_latency
=
623 perf
->states
[i
].transition_latency
* 1000;
626 /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
627 if (perf
->control_register
.space_id
== ACPI_ADR_SPACE_FIXED_HARDWARE
&&
628 policy
->cpuinfo
.transition_latency
> 20 * 1000) {
629 policy
->cpuinfo
.transition_latency
= 20 * 1000;
630 printk_once(KERN_INFO
631 "P-state transition latency capped at 20 uS\n");
635 for (i
= 0; i
< perf
->state_count
; i
++) {
636 if (i
> 0 && perf
->states
[i
].core_frequency
>=
637 data
->freq_table
[valid_states
-1].frequency
/ 1000)
640 data
->freq_table
[valid_states
].index
= i
;
641 data
->freq_table
[valid_states
].frequency
=
642 perf
->states
[i
].core_frequency
* 1000;
645 data
->freq_table
[valid_states
].frequency
= CPUFREQ_TABLE_END
;
648 result
= cpufreq_frequency_table_cpuinfo(policy
, data
->freq_table
);
652 if (perf
->states
[0].core_frequency
* 1000 != policy
->cpuinfo
.max_freq
)
653 printk(KERN_WARNING FW_WARN
"P-state 0 is not max freq\n");
655 switch (perf
->control_register
.space_id
) {
656 case ACPI_ADR_SPACE_SYSTEM_IO
:
657 /* Current speed is unknown and not detectable by IO port */
658 policy
->cur
= acpi_cpufreq_guess_freq(data
, policy
->cpu
);
660 case ACPI_ADR_SPACE_FIXED_HARDWARE
:
661 acpi_cpufreq_driver
.get
= get_cur_freq_on_cpu
;
662 policy
->cur
= get_cur_freq_on_cpu(cpu
);
668 /* notify BIOS that we exist */
669 acpi_processor_notify_smm(THIS_MODULE
);
671 /* Check for APERF/MPERF support in hardware */
672 if (cpu_has(c
, X86_FEATURE_APERFMPERF
))
673 acpi_cpufreq_driver
.getavg
= get_measured_perf
;
675 dprintk("CPU%u - ACPI performance management activated.\n", cpu
);
676 for (i
= 0; i
< perf
->state_count
; i
++)
677 dprintk(" %cP%d: %d MHz, %d mW, %d uS\n",
678 (i
== perf
->state
? '*' : ' '), i
,
679 (u32
) perf
->states
[i
].core_frequency
,
680 (u32
) perf
->states
[i
].power
,
681 (u32
) perf
->states
[i
].transition_latency
);
683 cpufreq_frequency_table_get_attr(data
->freq_table
, policy
->cpu
);
686 * the first call to ->target() should result in us actually
687 * writing something to the appropriate registers.
694 kfree(data
->freq_table
);
696 acpi_processor_unregister_performance(perf
, cpu
);
699 per_cpu(drv_data
, cpu
) = NULL
;
704 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy
*policy
)
706 struct acpi_cpufreq_data
*data
= per_cpu(drv_data
, policy
->cpu
);
708 dprintk("acpi_cpufreq_cpu_exit\n");
711 cpufreq_frequency_table_put_attr(policy
->cpu
);
712 per_cpu(drv_data
, policy
->cpu
) = NULL
;
713 acpi_processor_unregister_performance(data
->acpi_data
,
721 static int acpi_cpufreq_resume(struct cpufreq_policy
*policy
)
723 struct acpi_cpufreq_data
*data
= per_cpu(drv_data
, policy
->cpu
);
725 dprintk("acpi_cpufreq_resume\n");
732 static struct freq_attr
*acpi_cpufreq_attr
[] = {
733 &cpufreq_freq_attr_scaling_available_freqs
,
737 static struct cpufreq_driver acpi_cpufreq_driver
= {
738 .verify
= acpi_cpufreq_verify
,
739 .target
= acpi_cpufreq_target
,
740 .init
= acpi_cpufreq_cpu_init
,
741 .exit
= acpi_cpufreq_cpu_exit
,
742 .resume
= acpi_cpufreq_resume
,
743 .name
= "acpi-cpufreq",
744 .owner
= THIS_MODULE
,
745 .attr
= acpi_cpufreq_attr
,
748 static int __init
acpi_cpufreq_init(void)
755 dprintk("acpi_cpufreq_init\n");
757 ret
= acpi_cpufreq_early_init();
761 ret
= cpufreq_register_driver(&acpi_cpufreq_driver
);
763 free_acpi_perf_data();
768 static void __exit
acpi_cpufreq_exit(void)
770 dprintk("acpi_cpufreq_exit\n");
772 cpufreq_unregister_driver(&acpi_cpufreq_driver
);
774 free_percpu(acpi_perf_data
);
777 module_param(acpi_pstate_strict
, uint
, 0644);
778 MODULE_PARM_DESC(acpi_pstate_strict
,
779 "value 0 or non-zero. non-zero -> strict ACPI checks are "
780 "performed during frequency changes.");
782 late_initcall(acpi_cpufreq_init
);
783 module_exit(acpi_cpufreq_exit
);
785 MODULE_ALIAS("acpi");