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>
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>
50
MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
51
MODULE_DESCRIPTION("ACPI Processor P-States Driver");
52
MODULE_LICENSE("GPL");
55
UNDEFINED_CAPABLE = 0,
56
SYSTEM_INTEL_MSR_CAPABLE,
60
#define INTEL_MSR_RANGE (0xffff)
62
struct acpi_cpufreq_data {
63
struct acpi_processor_performance *acpi_data;
64
struct cpufreq_frequency_table *freq_table;
66
unsigned int cpu_feature;
69
static DEFINE_PER_CPU(struct acpi_cpufreq_data *, acfreq_data);
71
/* acpi_perf_data is a pointer to percpu data. */
72
static struct acpi_processor_performance __percpu *acpi_perf_data;
74
static struct cpufreq_driver acpi_cpufreq_driver;
76
static unsigned int acpi_pstate_strict;
78
static int check_est_cpu(unsigned int cpuid)
80
struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
82
return cpu_has(cpu, X86_FEATURE_EST);
85
static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
87
struct acpi_processor_performance *perf;
90
perf = data->acpi_data;
92
for (i = 0; i < perf->state_count; i++) {
93
if (value == perf->states[i].status)
94
return data->freq_table[i].frequency;
99
static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
102
struct acpi_processor_performance *perf;
104
msr &= INTEL_MSR_RANGE;
105
perf = data->acpi_data;
107
for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
108
if (msr == perf->states[data->freq_table[i].index].status)
109
return data->freq_table[i].frequency;
111
return data->freq_table[0].frequency;
114
static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
116
switch (data->cpu_feature) {
117
case SYSTEM_INTEL_MSR_CAPABLE:
118
return extract_msr(val, data);
119
case SYSTEM_IO_CAPABLE:
120
return extract_io(val, data);
137
const struct cpumask *mask;
145
/* Called via smp_call_function_single(), on the target CPU */
146
static void do_drv_read(void *_cmd)
148
struct drv_cmd *cmd = _cmd;
152
case SYSTEM_INTEL_MSR_CAPABLE:
153
rdmsr(cmd->addr.msr.reg, cmd->val, h);
155
case SYSTEM_IO_CAPABLE:
156
acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
158
(u32)cmd->addr.io.bit_width);
165
/* Called via smp_call_function_many(), on the target CPUs */
166
static void do_drv_write(void *_cmd)
168
struct drv_cmd *cmd = _cmd;
172
case SYSTEM_INTEL_MSR_CAPABLE:
173
rdmsr(cmd->addr.msr.reg, lo, hi);
174
lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
175
wrmsr(cmd->addr.msr.reg, lo, hi);
177
case SYSTEM_IO_CAPABLE:
178
acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
180
(u32)cmd->addr.io.bit_width);
187
static void drv_read(struct drv_cmd *cmd)
192
err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
193
WARN_ON_ONCE(err); /* smp_call_function_any() was buggy? */
196
static void drv_write(struct drv_cmd *cmd)
200
this_cpu = get_cpu();
201
if (cpumask_test_cpu(this_cpu, cmd->mask))
203
smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
207
static u32 get_cur_val(const struct cpumask *mask)
209
struct acpi_processor_performance *perf;
212
if (unlikely(cpumask_empty(mask)))
215
switch (per_cpu(acfreq_data, cpumask_first(mask))->cpu_feature) {
216
case SYSTEM_INTEL_MSR_CAPABLE:
217
cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
218
cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
220
case SYSTEM_IO_CAPABLE:
221
cmd.type = SYSTEM_IO_CAPABLE;
222
perf = per_cpu(acfreq_data, cpumask_first(mask))->acpi_data;
223
cmd.addr.io.port = perf->control_register.address;
224
cmd.addr.io.bit_width = perf->control_register.bit_width;
233
pr_debug("get_cur_val = %u\n", cmd.val);
238
static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
240
struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
242
unsigned int cached_freq;
244
pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
246
if (unlikely(data == NULL ||
247
data->acpi_data == NULL || data->freq_table == NULL)) {
251
cached_freq = data->freq_table[data->acpi_data->state].frequency;
252
freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
253
if (freq != cached_freq) {
255
* The dreaded BIOS frequency change behind our back.
256
* Force set the frequency on next target call.
261
pr_debug("cur freq = %u\n", freq);
266
static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
267
struct acpi_cpufreq_data *data)
269
unsigned int cur_freq;
272
for (i = 0; i < 100; i++) {
273
cur_freq = extract_freq(get_cur_val(mask), data);
274
if (cur_freq == freq)
281
static int acpi_cpufreq_target(struct cpufreq_policy *policy,
282
unsigned int target_freq, unsigned int relation)
284
struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
285
struct acpi_processor_performance *perf;
286
struct cpufreq_freqs freqs;
288
unsigned int next_state = 0; /* Index into freq_table */
289
unsigned int next_perf_state = 0; /* Index into perf table */
293
pr_debug("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
295
if (unlikely(data == NULL ||
296
data->acpi_data == NULL || data->freq_table == NULL)) {
300
perf = data->acpi_data;
301
result = cpufreq_frequency_table_target(policy,
304
relation, &next_state);
305
if (unlikely(result)) {
310
next_perf_state = data->freq_table[next_state].index;
311
if (perf->state == next_perf_state) {
312
if (unlikely(data->resume)) {
313
pr_debug("Called after resume, resetting to P%d\n",
317
pr_debug("Already at target state (P%d)\n",
323
switch (data->cpu_feature) {
324
case SYSTEM_INTEL_MSR_CAPABLE:
325
cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
326
cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
327
cmd.val = (u32) perf->states[next_perf_state].control;
329
case SYSTEM_IO_CAPABLE:
330
cmd.type = SYSTEM_IO_CAPABLE;
331
cmd.addr.io.port = perf->control_register.address;
332
cmd.addr.io.bit_width = perf->control_register.bit_width;
333
cmd.val = (u32) perf->states[next_perf_state].control;
340
/* cpufreq holds the hotplug lock, so we are safe from here on */
341
if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
342
cmd.mask = policy->cpus;
344
cmd.mask = cpumask_of(policy->cpu);
346
freqs.old = perf->states[perf->state].core_frequency * 1000;
347
freqs.new = data->freq_table[next_state].frequency;
348
for_each_cpu(i, policy->cpus) {
350
cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
355
if (acpi_pstate_strict) {
356
if (!check_freqs(cmd.mask, freqs.new, data)) {
357
pr_debug("acpi_cpufreq_target failed (%d)\n",
364
for_each_cpu(i, policy->cpus) {
366
cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
368
perf->state = next_perf_state;
374
static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
376
struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
378
pr_debug("acpi_cpufreq_verify\n");
380
return cpufreq_frequency_table_verify(policy, data->freq_table);
384
acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
386
struct acpi_processor_performance *perf = data->acpi_data;
389
/* search the closest match to cpu_khz */
392
unsigned long freqn = perf->states[0].core_frequency * 1000;
394
for (i = 0; i < (perf->state_count-1); i++) {
396
freqn = perf->states[i+1].core_frequency * 1000;
397
if ((2 * cpu_khz) > (freqn + freq)) {
402
perf->state = perf->state_count-1;
405
/* assume CPU is at P0... */
407
return perf->states[0].core_frequency * 1000;
411
static void free_acpi_perf_data(void)
415
/* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
416
for_each_possible_cpu(i)
417
free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
419
free_percpu(acpi_perf_data);
423
* acpi_cpufreq_early_init - initialize ACPI P-States library
425
* Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
426
* in order to determine correct frequency and voltage pairings. We can
427
* do _PDC and _PSD and find out the processor dependency for the
428
* actual init that will happen later...
430
static int __init acpi_cpufreq_early_init(void)
433
pr_debug("acpi_cpufreq_early_init\n");
435
acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
436
if (!acpi_perf_data) {
437
pr_debug("Memory allocation error for acpi_perf_data.\n");
440
for_each_possible_cpu(i) {
441
if (!zalloc_cpumask_var_node(
442
&per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
443
GFP_KERNEL, cpu_to_node(i))) {
445
/* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
446
free_acpi_perf_data();
451
/* Do initialization in ACPI core */
452
acpi_processor_preregister_performance(acpi_perf_data);
458
* Some BIOSes do SW_ANY coordination internally, either set it up in hw
459
* or do it in BIOS firmware and won't inform about it to OS. If not
460
* detected, this has a side effect of making CPU run at a different speed
461
* than OS intended it to run at. Detect it and handle it cleanly.
463
static int bios_with_sw_any_bug;
465
static int sw_any_bug_found(const struct dmi_system_id *d)
467
bios_with_sw_any_bug = 1;
471
static const struct dmi_system_id sw_any_bug_dmi_table[] = {
473
.callback = sw_any_bug_found,
474
.ident = "Supermicro Server X6DLP",
476
DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
477
DMI_MATCH(DMI_BIOS_VERSION, "080010"),
478
DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
484
static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
486
/* Intel Xeon Processor 7100 Series Specification Update
487
* http://www.intel.com/Assets/PDF/specupdate/314554.pdf
488
* AL30: A Machine Check Exception (MCE) Occurring during an
489
* Enhanced Intel SpeedStep Technology Ratio Change May Cause
490
* Both Processor Cores to Lock Up. */
491
if (c->x86_vendor == X86_VENDOR_INTEL) {
492
if ((c->x86 == 15) &&
493
(c->x86_model == 6) &&
494
(c->x86_mask == 8)) {
495
printk(KERN_INFO "acpi-cpufreq: Intel(R) "
496
"Xeon(R) 7100 Errata AL30, processors may "
497
"lock up on frequency changes: disabling "
506
static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
509
unsigned int valid_states = 0;
510
unsigned int cpu = policy->cpu;
511
struct acpi_cpufreq_data *data;
512
unsigned int result = 0;
513
struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
514
struct acpi_processor_performance *perf;
516
static int blacklisted;
519
pr_debug("acpi_cpufreq_cpu_init\n");
524
blacklisted = acpi_cpufreq_blacklist(c);
529
data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
533
data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
534
per_cpu(acfreq_data, cpu) = data;
536
if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
537
acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
539
result = acpi_processor_register_performance(data->acpi_data, cpu);
543
perf = data->acpi_data;
544
policy->shared_type = perf->shared_type;
547
* Will let policy->cpus know about dependency only when software
548
* coordination is required.
550
if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
551
policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
552
cpumask_copy(policy->cpus, perf->shared_cpu_map);
554
cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
557
dmi_check_system(sw_any_bug_dmi_table);
558
if (bios_with_sw_any_bug && cpumask_weight(policy->cpus) == 1) {
559
policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
560
cpumask_copy(policy->cpus, cpu_core_mask(cpu));
564
/* capability check */
565
if (perf->state_count <= 1) {
566
pr_debug("No P-States\n");
571
if (perf->control_register.space_id != perf->status_register.space_id) {
576
switch (perf->control_register.space_id) {
577
case ACPI_ADR_SPACE_SYSTEM_IO:
578
pr_debug("SYSTEM IO addr space\n");
579
data->cpu_feature = SYSTEM_IO_CAPABLE;
581
case ACPI_ADR_SPACE_FIXED_HARDWARE:
582
pr_debug("HARDWARE addr space\n");
583
if (!check_est_cpu(cpu)) {
587
data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
590
pr_debug("Unknown addr space %d\n",
591
(u32) (perf->control_register.space_id));
596
data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
597
(perf->state_count+1), GFP_KERNEL);
598
if (!data->freq_table) {
603
/* detect transition latency */
604
policy->cpuinfo.transition_latency = 0;
605
for (i = 0; i < perf->state_count; i++) {
606
if ((perf->states[i].transition_latency * 1000) >
607
policy->cpuinfo.transition_latency)
608
policy->cpuinfo.transition_latency =
609
perf->states[i].transition_latency * 1000;
612
/* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
613
if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
614
policy->cpuinfo.transition_latency > 20 * 1000) {
615
policy->cpuinfo.transition_latency = 20 * 1000;
616
printk_once(KERN_INFO
617
"P-state transition latency capped at 20 uS\n");
621
for (i = 0; i < perf->state_count; i++) {
622
if (i > 0 && perf->states[i].core_frequency >=
623
data->freq_table[valid_states-1].frequency / 1000)
626
data->freq_table[valid_states].index = i;
627
data->freq_table[valid_states].frequency =
628
perf->states[i].core_frequency * 1000;
631
data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
634
result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
638
if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
639
printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
641
switch (perf->control_register.space_id) {
642
case ACPI_ADR_SPACE_SYSTEM_IO:
643
/* Current speed is unknown and not detectable by IO port */
644
policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
646
case ACPI_ADR_SPACE_FIXED_HARDWARE:
647
acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
648
policy->cur = get_cur_freq_on_cpu(cpu);
654
/* notify BIOS that we exist */
655
acpi_processor_notify_smm(THIS_MODULE);
657
/* Check for APERF/MPERF support in hardware */
658
if (cpu_has(c, X86_FEATURE_APERFMPERF))
659
acpi_cpufreq_driver.getavg = cpufreq_get_measured_perf;
661
pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
662
for (i = 0; i < perf->state_count; i++)
663
pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n",
664
(i == perf->state ? '*' : ' '), i,
665
(u32) perf->states[i].core_frequency,
666
(u32) perf->states[i].power,
667
(u32) perf->states[i].transition_latency);
669
cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
672
* the first call to ->target() should result in us actually
673
* writing something to the appropriate registers.
680
kfree(data->freq_table);
682
acpi_processor_unregister_performance(perf, cpu);
685
per_cpu(acfreq_data, cpu) = NULL;
690
static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
692
struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
694
pr_debug("acpi_cpufreq_cpu_exit\n");
697
cpufreq_frequency_table_put_attr(policy->cpu);
698
per_cpu(acfreq_data, policy->cpu) = NULL;
699
acpi_processor_unregister_performance(data->acpi_data,
701
kfree(data->freq_table);
708
static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
710
struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
712
pr_debug("acpi_cpufreq_resume\n");
719
static struct freq_attr *acpi_cpufreq_attr[] = {
720
&cpufreq_freq_attr_scaling_available_freqs,
724
static struct cpufreq_driver acpi_cpufreq_driver = {
725
.verify = acpi_cpufreq_verify,
726
.target = acpi_cpufreq_target,
727
.bios_limit = acpi_processor_get_bios_limit,
728
.init = acpi_cpufreq_cpu_init,
729
.exit = acpi_cpufreq_cpu_exit,
730
.resume = acpi_cpufreq_resume,
731
.name = "acpi-cpufreq",
732
.owner = THIS_MODULE,
733
.attr = acpi_cpufreq_attr,
736
static int __init acpi_cpufreq_init(void)
743
pr_debug("acpi_cpufreq_init\n");
745
ret = acpi_cpufreq_early_init();
749
ret = cpufreq_register_driver(&acpi_cpufreq_driver);
751
free_acpi_perf_data();
756
static void __exit acpi_cpufreq_exit(void)
758
pr_debug("acpi_cpufreq_exit\n");
760
cpufreq_unregister_driver(&acpi_cpufreq_driver);
762
free_percpu(acpi_perf_data);
765
module_param(acpi_pstate_strict, uint, 0644);
766
MODULE_PARM_DESC(acpi_pstate_strict,
767
"value 0 or non-zero. non-zero -> strict ACPI checks are "
768
"performed during frequency changes.");
770
late_initcall(acpi_cpufreq_init);
771
module_exit(acpi_cpufreq_exit);
773
MODULE_ALIAS("acpi");