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
#define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, \
53
MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
54
MODULE_DESCRIPTION("ACPI Processor P-States Driver");
55
MODULE_LICENSE("GPL");
58
UNDEFINED_CAPABLE = 0,
59
SYSTEM_INTEL_MSR_CAPABLE,
63
#define INTEL_MSR_RANGE (0xffff)
65
struct acpi_cpufreq_data {
66
struct acpi_processor_performance *acpi_data;
67
struct cpufreq_frequency_table *freq_table;
69
unsigned int cpu_feature;
72
static DEFINE_PER_CPU(struct acpi_cpufreq_data *, acfreq_data);
74
/* acpi_perf_data is a pointer to percpu data. */
75
static struct acpi_processor_performance __percpu *acpi_perf_data;
77
static struct cpufreq_driver acpi_cpufreq_driver;
79
static unsigned int acpi_pstate_strict;
81
static int check_est_cpu(unsigned int cpuid)
83
struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
85
return cpu_has(cpu, X86_FEATURE_EST);
88
static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
90
struct acpi_processor_performance *perf;
93
perf = data->acpi_data;
95
for (i = 0; i < perf->state_count; i++) {
96
if (value == perf->states[i].status)
97
return data->freq_table[i].frequency;
102
static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
105
struct acpi_processor_performance *perf;
107
msr &= INTEL_MSR_RANGE;
108
perf = data->acpi_data;
110
for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
111
if (msr == perf->states[data->freq_table[i].index].status)
112
return data->freq_table[i].frequency;
114
return data->freq_table[0].frequency;
117
static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
119
switch (data->cpu_feature) {
120
case SYSTEM_INTEL_MSR_CAPABLE:
121
return extract_msr(val, data);
122
case SYSTEM_IO_CAPABLE:
123
return extract_io(val, data);
140
const struct cpumask *mask;
148
/* Called via smp_call_function_single(), on the target CPU */
149
static void do_drv_read(void *_cmd)
151
struct drv_cmd *cmd = _cmd;
155
case SYSTEM_INTEL_MSR_CAPABLE:
156
rdmsr(cmd->addr.msr.reg, cmd->val, h);
158
case SYSTEM_IO_CAPABLE:
159
acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
161
(u32)cmd->addr.io.bit_width);
168
/* Called via smp_call_function_many(), on the target CPUs */
169
static void do_drv_write(void *_cmd)
171
struct drv_cmd *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)
195
err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
196
WARN_ON_ONCE(err); /* smp_call_function_any() was buggy? */
199
static void drv_write(struct drv_cmd *cmd)
203
this_cpu = get_cpu();
204
if (cpumask_test_cpu(this_cpu, cmd->mask))
206
smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
210
static u32 get_cur_val(const struct cpumask *mask)
212
struct acpi_processor_performance *perf;
215
if (unlikely(cpumask_empty(mask)))
218
switch (per_cpu(acfreq_data, cpumask_first(mask))->cpu_feature) {
219
case SYSTEM_INTEL_MSR_CAPABLE:
220
cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
221
cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
223
case SYSTEM_IO_CAPABLE:
224
cmd.type = SYSTEM_IO_CAPABLE;
225
perf = per_cpu(acfreq_data, cpumask_first(mask))->acpi_data;
226
cmd.addr.io.port = perf->control_register.address;
227
cmd.addr.io.bit_width = perf->control_register.bit_width;
236
dprintk("get_cur_val = %u\n", cmd.val);
241
static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
243
struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
245
unsigned int cached_freq;
247
dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
249
if (unlikely(data == NULL ||
250
data->acpi_data == NULL || data->freq_table == NULL)) {
254
cached_freq = data->freq_table[data->acpi_data->state].frequency;
255
freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
256
if (freq != cached_freq) {
258
* The dreaded BIOS frequency change behind our back.
259
* Force set the frequency on next target call.
264
dprintk("cur freq = %u\n", freq);
269
static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
270
struct acpi_cpufreq_data *data)
272
unsigned int cur_freq;
275
for (i = 0; i < 100; i++) {
276
cur_freq = extract_freq(get_cur_val(mask), data);
277
if (cur_freq == freq)
284
static int acpi_cpufreq_target(struct cpufreq_policy *policy,
285
unsigned int target_freq, unsigned int relation)
287
struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
288
struct acpi_processor_performance *perf;
289
struct cpufreq_freqs freqs;
291
unsigned int next_state = 0; /* Index into freq_table */
292
unsigned int next_perf_state = 0; /* Index into perf table */
296
dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
298
if (unlikely(data == NULL ||
299
data->acpi_data == NULL || data->freq_table == NULL)) {
303
perf = data->acpi_data;
304
result = cpufreq_frequency_table_target(policy,
307
relation, &next_state);
308
if (unlikely(result)) {
313
next_perf_state = data->freq_table[next_state].index;
314
if (perf->state == next_perf_state) {
315
if (unlikely(data->resume)) {
316
dprintk("Called after resume, resetting to P%d\n",
320
dprintk("Already at target state (P%d)\n",
326
switch (data->cpu_feature) {
327
case SYSTEM_INTEL_MSR_CAPABLE:
328
cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
329
cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
330
cmd.val = (u32) perf->states[next_perf_state].control;
332
case SYSTEM_IO_CAPABLE:
333
cmd.type = SYSTEM_IO_CAPABLE;
334
cmd.addr.io.port = perf->control_register.address;
335
cmd.addr.io.bit_width = perf->control_register.bit_width;
336
cmd.val = (u32) perf->states[next_perf_state].control;
343
/* cpufreq holds the hotplug lock, so we are safe from here on */
344
if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
345
cmd.mask = policy->cpus;
347
cmd.mask = cpumask_of(policy->cpu);
349
freqs.old = perf->states[perf->state].core_frequency * 1000;
350
freqs.new = data->freq_table[next_state].frequency;
351
for_each_cpu(i, policy->cpus) {
353
cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
358
if (acpi_pstate_strict) {
359
if (!check_freqs(cmd.mask, freqs.new, data)) {
360
dprintk("acpi_cpufreq_target failed (%d)\n",
367
for_each_cpu(i, policy->cpus) {
369
cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
371
perf->state = next_perf_state;
377
static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
379
struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
381
dprintk("acpi_cpufreq_verify\n");
383
return cpufreq_frequency_table_verify(policy, data->freq_table);
387
acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
389
struct acpi_processor_performance *perf = data->acpi_data;
392
/* search the closest match to cpu_khz */
395
unsigned long freqn = perf->states[0].core_frequency * 1000;
397
for (i = 0; i < (perf->state_count-1); i++) {
399
freqn = perf->states[i+1].core_frequency * 1000;
400
if ((2 * cpu_khz) > (freqn + freq)) {
405
perf->state = perf->state_count-1;
408
/* assume CPU is at P0... */
410
return perf->states[0].core_frequency * 1000;
414
static void free_acpi_perf_data(void)
418
/* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
419
for_each_possible_cpu(i)
420
free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
422
free_percpu(acpi_perf_data);
426
* acpi_cpufreq_early_init - initialize ACPI P-States library
428
* Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
429
* in order to determine correct frequency and voltage pairings. We can
430
* do _PDC and _PSD and find out the processor dependency for the
431
* actual init that will happen later...
433
static int __init acpi_cpufreq_early_init(void)
436
dprintk("acpi_cpufreq_early_init\n");
438
acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
439
if (!acpi_perf_data) {
440
dprintk("Memory allocation error for acpi_perf_data.\n");
443
for_each_possible_cpu(i) {
444
if (!zalloc_cpumask_var_node(
445
&per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
446
GFP_KERNEL, cpu_to_node(i))) {
448
/* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
449
free_acpi_perf_data();
454
/* Do initialization in ACPI core */
455
acpi_processor_preregister_performance(acpi_perf_data);
461
* Some BIOSes do SW_ANY coordination internally, either set it up in hw
462
* or do it in BIOS firmware and won't inform about it to OS. If not
463
* detected, this has a side effect of making CPU run at a different speed
464
* than OS intended it to run at. Detect it and handle it cleanly.
466
static int bios_with_sw_any_bug;
468
static int sw_any_bug_found(const struct dmi_system_id *d)
470
bios_with_sw_any_bug = 1;
474
static const struct dmi_system_id sw_any_bug_dmi_table[] = {
476
.callback = sw_any_bug_found,
477
.ident = "Supermicro Server X6DLP",
479
DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
480
DMI_MATCH(DMI_BIOS_VERSION, "080010"),
481
DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
487
static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
489
/* Intel Xeon Processor 7100 Series Specification Update
490
* http://www.intel.com/Assets/PDF/specupdate/314554.pdf
491
* AL30: A Machine Check Exception (MCE) Occurring during an
492
* Enhanced Intel SpeedStep Technology Ratio Change May Cause
493
* Both Processor Cores to Lock Up. */
494
if (c->x86_vendor == X86_VENDOR_INTEL) {
495
if ((c->x86 == 15) &&
496
(c->x86_model == 6) &&
497
(c->x86_mask == 8)) {
498
printk(KERN_INFO "acpi-cpufreq: Intel(R) "
499
"Xeon(R) 7100 Errata AL30, processors may "
500
"lock up on frequency changes: disabling "
509
static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
512
unsigned int valid_states = 0;
513
unsigned int cpu = policy->cpu;
514
struct acpi_cpufreq_data *data;
515
unsigned int result = 0;
516
struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
517
struct acpi_processor_performance *perf;
519
static int blacklisted;
522
dprintk("acpi_cpufreq_cpu_init\n");
527
blacklisted = acpi_cpufreq_blacklist(c);
532
data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
536
data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
537
per_cpu(acfreq_data, cpu) = data;
539
if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
540
acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
542
result = acpi_processor_register_performance(data->acpi_data, cpu);
546
perf = data->acpi_data;
547
policy->shared_type = perf->shared_type;
550
* Will let policy->cpus know about dependency only when software
551
* coordination is required.
553
if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
554
policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
555
cpumask_copy(policy->cpus, perf->shared_cpu_map);
557
cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
560
dmi_check_system(sw_any_bug_dmi_table);
561
if (bios_with_sw_any_bug && cpumask_weight(policy->cpus) == 1) {
562
policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
563
cpumask_copy(policy->cpus, cpu_core_mask(cpu));
567
/* capability check */
568
if (perf->state_count <= 1) {
569
dprintk("No P-States\n");
574
if (perf->control_register.space_id != perf->status_register.space_id) {
579
switch (perf->control_register.space_id) {
580
case ACPI_ADR_SPACE_SYSTEM_IO:
581
dprintk("SYSTEM IO addr space\n");
582
data->cpu_feature = SYSTEM_IO_CAPABLE;
584
case ACPI_ADR_SPACE_FIXED_HARDWARE:
585
dprintk("HARDWARE addr space\n");
586
if (!check_est_cpu(cpu)) {
590
data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
593
dprintk("Unknown addr space %d\n",
594
(u32) (perf->control_register.space_id));
599
data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
600
(perf->state_count+1), GFP_KERNEL);
601
if (!data->freq_table) {
606
/* detect transition latency */
607
policy->cpuinfo.transition_latency = 0;
608
for (i = 0; i < perf->state_count; i++) {
609
if ((perf->states[i].transition_latency * 1000) >
610
policy->cpuinfo.transition_latency)
611
policy->cpuinfo.transition_latency =
612
perf->states[i].transition_latency * 1000;
615
/* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
616
if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
617
policy->cpuinfo.transition_latency > 20 * 1000) {
618
policy->cpuinfo.transition_latency = 20 * 1000;
619
printk_once(KERN_INFO
620
"P-state transition latency capped at 20 uS\n");
624
for (i = 0; i < perf->state_count; i++) {
625
if (i > 0 && perf->states[i].core_frequency >=
626
data->freq_table[valid_states-1].frequency / 1000)
629
data->freq_table[valid_states].index = i;
630
data->freq_table[valid_states].frequency =
631
perf->states[i].core_frequency * 1000;
634
data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
637
result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
641
if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
642
printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
644
switch (perf->control_register.space_id) {
645
case ACPI_ADR_SPACE_SYSTEM_IO:
646
/* Current speed is unknown and not detectable by IO port */
647
policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
649
case ACPI_ADR_SPACE_FIXED_HARDWARE:
650
acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
651
policy->cur = get_cur_freq_on_cpu(cpu);
657
/* notify BIOS that we exist */
658
acpi_processor_notify_smm(THIS_MODULE);
660
/* Check for APERF/MPERF support in hardware */
661
if (cpu_has(c, X86_FEATURE_APERFMPERF))
662
acpi_cpufreq_driver.getavg = cpufreq_get_measured_perf;
664
dprintk("CPU%u - ACPI performance management activated.\n", cpu);
665
for (i = 0; i < perf->state_count; i++)
666
dprintk(" %cP%d: %d MHz, %d mW, %d uS\n",
667
(i == perf->state ? '*' : ' '), i,
668
(u32) perf->states[i].core_frequency,
669
(u32) perf->states[i].power,
670
(u32) perf->states[i].transition_latency);
672
cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
675
* the first call to ->target() should result in us actually
676
* writing something to the appropriate registers.
683
kfree(data->freq_table);
685
acpi_processor_unregister_performance(perf, cpu);
688
per_cpu(acfreq_data, cpu) = NULL;
693
static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
695
struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
697
dprintk("acpi_cpufreq_cpu_exit\n");
700
cpufreq_frequency_table_put_attr(policy->cpu);
701
per_cpu(acfreq_data, policy->cpu) = NULL;
702
acpi_processor_unregister_performance(data->acpi_data,
704
kfree(data->freq_table);
711
static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
713
struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
715
dprintk("acpi_cpufreq_resume\n");
722
static struct freq_attr *acpi_cpufreq_attr[] = {
723
&cpufreq_freq_attr_scaling_available_freqs,
727
static struct cpufreq_driver acpi_cpufreq_driver = {
728
.verify = acpi_cpufreq_verify,
729
.target = acpi_cpufreq_target,
730
.bios_limit = acpi_processor_get_bios_limit,
731
.init = acpi_cpufreq_cpu_init,
732
.exit = acpi_cpufreq_cpu_exit,
733
.resume = acpi_cpufreq_resume,
734
.name = "acpi-cpufreq",
735
.owner = THIS_MODULE,
736
.attr = acpi_cpufreq_attr,
739
static int __init acpi_cpufreq_init(void)
746
dprintk("acpi_cpufreq_init\n");
748
ret = acpi_cpufreq_early_init();
752
ret = cpufreq_register_driver(&acpi_cpufreq_driver);
754
free_acpi_perf_data();
759
static void __exit acpi_cpufreq_exit(void)
761
dprintk("acpi_cpufreq_exit\n");
763
cpufreq_unregister_driver(&acpi_cpufreq_driver);
765
free_percpu(acpi_perf_data);
768
module_param(acpi_pstate_strict, uint, 0644);
769
MODULE_PARM_DESC(acpi_pstate_strict,
770
"value 0 or non-zero. non-zero -> strict ACPI checks are "
771
"performed during frequency changes.");
773
late_initcall(acpi_cpufreq_init);
774
module_exit(acpi_cpufreq_exit);
776
MODULE_ALIAS("acpi");