Loading...
1/*
2 * acpi-cpufreq.c - ACPI Processor P-States Driver
3 *
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>
8 *
9 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
10 *
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.
15 *
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.
20 *
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.
24 *
25 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
26 */
27
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
38#include <linux/acpi.h>
39#include <linux/io.h>
40#include <linux/delay.h>
41#include <linux/uaccess.h>
42
43#include <acpi/processor.h>
44
45#include <asm/msr.h>
46#include <asm/processor.h>
47#include <asm/cpufeature.h>
48#include "mperf.h"
49
50MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
51MODULE_DESCRIPTION("ACPI Processor P-States Driver");
52MODULE_LICENSE("GPL");
53
54enum {
55 UNDEFINED_CAPABLE = 0,
56 SYSTEM_INTEL_MSR_CAPABLE,
57 SYSTEM_IO_CAPABLE,
58};
59
60#define INTEL_MSR_RANGE (0xffff)
61
62struct acpi_cpufreq_data {
63 struct acpi_processor_performance *acpi_data;
64 struct cpufreq_frequency_table *freq_table;
65 unsigned int resume;
66 unsigned int cpu_feature;
67};
68
69static DEFINE_PER_CPU(struct acpi_cpufreq_data *, acfreq_data);
70
71/* acpi_perf_data is a pointer to percpu data. */
72static struct acpi_processor_performance __percpu *acpi_perf_data;
73
74static struct cpufreq_driver acpi_cpufreq_driver;
75
76static unsigned int acpi_pstate_strict;
77
78static int check_est_cpu(unsigned int cpuid)
79{
80 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
81
82 return cpu_has(cpu, X86_FEATURE_EST);
83}
84
85static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
86{
87 struct acpi_processor_performance *perf;
88 int i;
89
90 perf = data->acpi_data;
91
92 for (i = 0; i < perf->state_count; i++) {
93 if (value == perf->states[i].status)
94 return data->freq_table[i].frequency;
95 }
96 return 0;
97}
98
99static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
100{
101 int i;
102 struct acpi_processor_performance *perf;
103
104 msr &= INTEL_MSR_RANGE;
105 perf = data->acpi_data;
106
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;
110 }
111 return data->freq_table[0].frequency;
112}
113
114static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
115{
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);
121 default:
122 return 0;
123 }
124}
125
126struct msr_addr {
127 u32 reg;
128};
129
130struct io_addr {
131 u16 port;
132 u8 bit_width;
133};
134
135struct drv_cmd {
136 unsigned int type;
137 const struct cpumask *mask;
138 union {
139 struct msr_addr msr;
140 struct io_addr io;
141 } addr;
142 u32 val;
143};
144
145/* Called via smp_call_function_single(), on the target CPU */
146static void do_drv_read(void *_cmd)
147{
148 struct drv_cmd *cmd = _cmd;
149 u32 h;
150
151 switch (cmd->type) {
152 case SYSTEM_INTEL_MSR_CAPABLE:
153 rdmsr(cmd->addr.msr.reg, cmd->val, h);
154 break;
155 case SYSTEM_IO_CAPABLE:
156 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
157 &cmd->val,
158 (u32)cmd->addr.io.bit_width);
159 break;
160 default:
161 break;
162 }
163}
164
165/* Called via smp_call_function_many(), on the target CPUs */
166static void do_drv_write(void *_cmd)
167{
168 struct drv_cmd *cmd = _cmd;
169 u32 lo, hi;
170
171 switch (cmd->type) {
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);
176 break;
177 case SYSTEM_IO_CAPABLE:
178 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
179 cmd->val,
180 (u32)cmd->addr.io.bit_width);
181 break;
182 default:
183 break;
184 }
185}
186
187static void drv_read(struct drv_cmd *cmd)
188{
189 int err;
190 cmd->val = 0;
191
192 err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
193 WARN_ON_ONCE(err); /* smp_call_function_any() was buggy? */
194}
195
196static void drv_write(struct drv_cmd *cmd)
197{
198 int this_cpu;
199
200 this_cpu = get_cpu();
201 if (cpumask_test_cpu(this_cpu, cmd->mask))
202 do_drv_write(cmd);
203 smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
204 put_cpu();
205}
206
207static u32 get_cur_val(const struct cpumask *mask)
208{
209 struct acpi_processor_performance *perf;
210 struct drv_cmd cmd;
211
212 if (unlikely(cpumask_empty(mask)))
213 return 0;
214
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;
219 break;
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;
225 break;
226 default:
227 return 0;
228 }
229
230 cmd.mask = mask;
231 drv_read(&cmd);
232
233 pr_debug("get_cur_val = %u\n", cmd.val);
234
235 return cmd.val;
236}
237
238static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
239{
240 struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
241 unsigned int freq;
242 unsigned int cached_freq;
243
244 pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
245
246 if (unlikely(data == NULL ||
247 data->acpi_data == NULL || data->freq_table == NULL)) {
248 return 0;
249 }
250
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) {
254 /*
255 * The dreaded BIOS frequency change behind our back.
256 * Force set the frequency on next target call.
257 */
258 data->resume = 1;
259 }
260
261 pr_debug("cur freq = %u\n", freq);
262
263 return freq;
264}
265
266static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
267 struct acpi_cpufreq_data *data)
268{
269 unsigned int cur_freq;
270 unsigned int i;
271
272 for (i = 0; i < 100; i++) {
273 cur_freq = extract_freq(get_cur_val(mask), data);
274 if (cur_freq == freq)
275 return 1;
276 udelay(10);
277 }
278 return 0;
279}
280
281static int acpi_cpufreq_target(struct cpufreq_policy *policy,
282 unsigned int target_freq, unsigned int relation)
283{
284 struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
285 struct acpi_processor_performance *perf;
286 struct cpufreq_freqs freqs;
287 struct drv_cmd cmd;
288 unsigned int next_state = 0; /* Index into freq_table */
289 unsigned int next_perf_state = 0; /* Index into perf table */
290 unsigned int i;
291 int result = 0;
292
293 pr_debug("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
294
295 if (unlikely(data == NULL ||
296 data->acpi_data == NULL || data->freq_table == NULL)) {
297 return -ENODEV;
298 }
299
300 perf = data->acpi_data;
301 result = cpufreq_frequency_table_target(policy,
302 data->freq_table,
303 target_freq,
304 relation, &next_state);
305 if (unlikely(result)) {
306 result = -ENODEV;
307 goto out;
308 }
309
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",
314 next_perf_state);
315 data->resume = 0;
316 } else {
317 pr_debug("Already at target state (P%d)\n",
318 next_perf_state);
319 goto out;
320 }
321 }
322
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;
328 break;
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;
334 break;
335 default:
336 result = -ENODEV;
337 goto out;
338 }
339
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;
343 else
344 cmd.mask = cpumask_of(policy->cpu);
345
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) {
349 freqs.cpu = i;
350 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
351 }
352
353 drv_write(&cmd);
354
355 if (acpi_pstate_strict) {
356 if (!check_freqs(cmd.mask, freqs.new, data)) {
357 pr_debug("acpi_cpufreq_target failed (%d)\n",
358 policy->cpu);
359 result = -EAGAIN;
360 goto out;
361 }
362 }
363
364 for_each_cpu(i, policy->cpus) {
365 freqs.cpu = i;
366 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
367 }
368 perf->state = next_perf_state;
369
370out:
371 return result;
372}
373
374static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
375{
376 struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
377
378 pr_debug("acpi_cpufreq_verify\n");
379
380 return cpufreq_frequency_table_verify(policy, data->freq_table);
381}
382
383static unsigned long
384acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
385{
386 struct acpi_processor_performance *perf = data->acpi_data;
387
388 if (cpu_khz) {
389 /* search the closest match to cpu_khz */
390 unsigned int i;
391 unsigned long freq;
392 unsigned long freqn = perf->states[0].core_frequency * 1000;
393
394 for (i = 0; i < (perf->state_count-1); i++) {
395 freq = freqn;
396 freqn = perf->states[i+1].core_frequency * 1000;
397 if ((2 * cpu_khz) > (freqn + freq)) {
398 perf->state = i;
399 return freq;
400 }
401 }
402 perf->state = perf->state_count-1;
403 return freqn;
404 } else {
405 /* assume CPU is at P0... */
406 perf->state = 0;
407 return perf->states[0].core_frequency * 1000;
408 }
409}
410
411static void free_acpi_perf_data(void)
412{
413 unsigned int i;
414
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)
418 ->shared_cpu_map);
419 free_percpu(acpi_perf_data);
420}
421
422/*
423 * acpi_cpufreq_early_init - initialize ACPI P-States library
424 *
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...
429 */
430static int __init acpi_cpufreq_early_init(void)
431{
432 unsigned int i;
433 pr_debug("acpi_cpufreq_early_init\n");
434
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");
438 return -ENOMEM;
439 }
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))) {
444
445 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
446 free_acpi_perf_data();
447 return -ENOMEM;
448 }
449 }
450
451 /* Do initialization in ACPI core */
452 acpi_processor_preregister_performance(acpi_perf_data);
453 return 0;
454}
455
456#ifdef CONFIG_SMP
457/*
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.
462 */
463static int bios_with_sw_any_bug;
464
465static int sw_any_bug_found(const struct dmi_system_id *d)
466{
467 bios_with_sw_any_bug = 1;
468 return 0;
469}
470
471static const struct dmi_system_id sw_any_bug_dmi_table[] = {
472 {
473 .callback = sw_any_bug_found,
474 .ident = "Supermicro Server X6DLP",
475 .matches = {
476 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
477 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
478 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
479 },
480 },
481 { }
482};
483
484static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
485{
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 "
498 "acpi-cpufreq.\n");
499 return -ENODEV;
500 }
501 }
502 return 0;
503}
504#endif
505
506static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
507{
508 unsigned int i;
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;
515#ifdef CONFIG_SMP
516 static int blacklisted;
517#endif
518
519 pr_debug("acpi_cpufreq_cpu_init\n");
520
521#ifdef CONFIG_SMP
522 if (blacklisted)
523 return blacklisted;
524 blacklisted = acpi_cpufreq_blacklist(c);
525 if (blacklisted)
526 return blacklisted;
527#endif
528
529 data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
530 if (!data)
531 return -ENOMEM;
532
533 data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
534 per_cpu(acfreq_data, cpu) = data;
535
536 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
537 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
538
539 result = acpi_processor_register_performance(data->acpi_data, cpu);
540 if (result)
541 goto err_free;
542
543 perf = data->acpi_data;
544 policy->shared_type = perf->shared_type;
545
546 /*
547 * Will let policy->cpus know about dependency only when software
548 * coordination is required.
549 */
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);
553 }
554 cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
555
556#ifdef CONFIG_SMP
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));
561 }
562#endif
563
564 /* capability check */
565 if (perf->state_count <= 1) {
566 pr_debug("No P-States\n");
567 result = -ENODEV;
568 goto err_unreg;
569 }
570
571 if (perf->control_register.space_id != perf->status_register.space_id) {
572 result = -ENODEV;
573 goto err_unreg;
574 }
575
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;
580 break;
581 case ACPI_ADR_SPACE_FIXED_HARDWARE:
582 pr_debug("HARDWARE addr space\n");
583 if (!check_est_cpu(cpu)) {
584 result = -ENODEV;
585 goto err_unreg;
586 }
587 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
588 break;
589 default:
590 pr_debug("Unknown addr space %d\n",
591 (u32) (perf->control_register.space_id));
592 result = -ENODEV;
593 goto err_unreg;
594 }
595
596 data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
597 (perf->state_count+1), GFP_KERNEL);
598 if (!data->freq_table) {
599 result = -ENOMEM;
600 goto err_unreg;
601 }
602
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;
610 }
611
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");
618 }
619
620 /* table init */
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)
624 continue;
625
626 data->freq_table[valid_states].index = i;
627 data->freq_table[valid_states].frequency =
628 perf->states[i].core_frequency * 1000;
629 valid_states++;
630 }
631 data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
632 perf->state = 0;
633
634 result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
635 if (result)
636 goto err_freqfree;
637
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");
640
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);
645 break;
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);
649 break;
650 default:
651 break;
652 }
653
654 /* notify BIOS that we exist */
655 acpi_processor_notify_smm(THIS_MODULE);
656
657 /* Check for APERF/MPERF support in hardware */
658 if (boot_cpu_has(X86_FEATURE_APERFMPERF))
659 acpi_cpufreq_driver.getavg = cpufreq_get_measured_perf;
660
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);
668
669 cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
670
671 /*
672 * the first call to ->target() should result in us actually
673 * writing something to the appropriate registers.
674 */
675 data->resume = 1;
676
677 return result;
678
679err_freqfree:
680 kfree(data->freq_table);
681err_unreg:
682 acpi_processor_unregister_performance(perf, cpu);
683err_free:
684 kfree(data);
685 per_cpu(acfreq_data, cpu) = NULL;
686
687 return result;
688}
689
690static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
691{
692 struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
693
694 pr_debug("acpi_cpufreq_cpu_exit\n");
695
696 if (data) {
697 cpufreq_frequency_table_put_attr(policy->cpu);
698 per_cpu(acfreq_data, policy->cpu) = NULL;
699 acpi_processor_unregister_performance(data->acpi_data,
700 policy->cpu);
701 kfree(data->freq_table);
702 kfree(data);
703 }
704
705 return 0;
706}
707
708static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
709{
710 struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
711
712 pr_debug("acpi_cpufreq_resume\n");
713
714 data->resume = 1;
715
716 return 0;
717}
718
719static struct freq_attr *acpi_cpufreq_attr[] = {
720 &cpufreq_freq_attr_scaling_available_freqs,
721 NULL,
722};
723
724static 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,
734};
735
736static int __init acpi_cpufreq_init(void)
737{
738 int ret;
739
740 if (acpi_disabled)
741 return 0;
742
743 pr_debug("acpi_cpufreq_init\n");
744
745 ret = acpi_cpufreq_early_init();
746 if (ret)
747 return ret;
748
749 ret = cpufreq_register_driver(&acpi_cpufreq_driver);
750 if (ret)
751 free_acpi_perf_data();
752
753 return ret;
754}
755
756static void __exit acpi_cpufreq_exit(void)
757{
758 pr_debug("acpi_cpufreq_exit\n");
759
760 cpufreq_unregister_driver(&acpi_cpufreq_driver);
761
762 free_acpi_perf_data();
763}
764
765module_param(acpi_pstate_strict, uint, 0644);
766MODULE_PARM_DESC(acpi_pstate_strict,
767 "value 0 or non-zero. non-zero -> strict ACPI checks are "
768 "performed during frequency changes.");
769
770late_initcall(acpi_cpufreq_init);
771module_exit(acpi_cpufreq_exit);
772
773MODULE_ALIAS("acpi");
1/*
2 * acpi-cpufreq.c - ACPI Processor P-States Driver
3 *
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>
8 *
9 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
10 *
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.
15 *
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.
20 *
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.
24 *
25 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
26 */
27
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
38#include <linux/acpi.h>
39#include <linux/io.h>
40#include <linux/delay.h>
41#include <linux/uaccess.h>
42
43#include <acpi/processor.h>
44
45#include <asm/msr.h>
46#include <asm/processor.h>
47#include <asm/cpufeature.h>
48
49MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
50MODULE_DESCRIPTION("ACPI Processor P-States Driver");
51MODULE_LICENSE("GPL");
52
53#define PFX "acpi-cpufreq: "
54
55enum {
56 UNDEFINED_CAPABLE = 0,
57 SYSTEM_INTEL_MSR_CAPABLE,
58 SYSTEM_AMD_MSR_CAPABLE,
59 SYSTEM_IO_CAPABLE,
60};
61
62#define INTEL_MSR_RANGE (0xffff)
63#define AMD_MSR_RANGE (0x7)
64
65#define MSR_K7_HWCR_CPB_DIS (1ULL << 25)
66
67struct acpi_cpufreq_data {
68 struct acpi_processor_performance *acpi_data;
69 struct cpufreq_frequency_table *freq_table;
70 unsigned int resume;
71 unsigned int cpu_feature;
72 cpumask_var_t freqdomain_cpus;
73};
74
75static DEFINE_PER_CPU(struct acpi_cpufreq_data *, acfreq_data);
76
77/* acpi_perf_data is a pointer to percpu data. */
78static struct acpi_processor_performance __percpu *acpi_perf_data;
79
80static struct cpufreq_driver acpi_cpufreq_driver;
81
82static unsigned int acpi_pstate_strict;
83static struct msr __percpu *msrs;
84
85static bool boost_state(unsigned int cpu)
86{
87 u32 lo, hi;
88 u64 msr;
89
90 switch (boot_cpu_data.x86_vendor) {
91 case X86_VENDOR_INTEL:
92 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
93 msr = lo | ((u64)hi << 32);
94 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
95 case X86_VENDOR_AMD:
96 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
97 msr = lo | ((u64)hi << 32);
98 return !(msr & MSR_K7_HWCR_CPB_DIS);
99 }
100 return false;
101}
102
103static void boost_set_msrs(bool enable, const struct cpumask *cpumask)
104{
105 u32 cpu;
106 u32 msr_addr;
107 u64 msr_mask;
108
109 switch (boot_cpu_data.x86_vendor) {
110 case X86_VENDOR_INTEL:
111 msr_addr = MSR_IA32_MISC_ENABLE;
112 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
113 break;
114 case X86_VENDOR_AMD:
115 msr_addr = MSR_K7_HWCR;
116 msr_mask = MSR_K7_HWCR_CPB_DIS;
117 break;
118 default:
119 return;
120 }
121
122 rdmsr_on_cpus(cpumask, msr_addr, msrs);
123
124 for_each_cpu(cpu, cpumask) {
125 struct msr *reg = per_cpu_ptr(msrs, cpu);
126 if (enable)
127 reg->q &= ~msr_mask;
128 else
129 reg->q |= msr_mask;
130 }
131
132 wrmsr_on_cpus(cpumask, msr_addr, msrs);
133}
134
135static int _store_boost(int val)
136{
137 get_online_cpus();
138 boost_set_msrs(val, cpu_online_mask);
139 put_online_cpus();
140 pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
141
142 return 0;
143}
144
145static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
146{
147 struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
148
149 return cpufreq_show_cpus(data->freqdomain_cpus, buf);
150}
151
152cpufreq_freq_attr_ro(freqdomain_cpus);
153
154#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
155static ssize_t store_boost(const char *buf, size_t count)
156{
157 int ret;
158 unsigned long val = 0;
159
160 if (!acpi_cpufreq_driver.boost_supported)
161 return -EINVAL;
162
163 ret = kstrtoul(buf, 10, &val);
164 if (ret || (val > 1))
165 return -EINVAL;
166
167 _store_boost((int) val);
168
169 return count;
170}
171
172static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
173 size_t count)
174{
175 return store_boost(buf, count);
176}
177
178static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
179{
180 return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
181}
182
183cpufreq_freq_attr_rw(cpb);
184#endif
185
186static int check_est_cpu(unsigned int cpuid)
187{
188 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
189
190 return cpu_has(cpu, X86_FEATURE_EST);
191}
192
193static int check_amd_hwpstate_cpu(unsigned int cpuid)
194{
195 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
196
197 return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
198}
199
200static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
201{
202 struct acpi_processor_performance *perf;
203 int i;
204
205 perf = data->acpi_data;
206
207 for (i = 0; i < perf->state_count; i++) {
208 if (value == perf->states[i].status)
209 return data->freq_table[i].frequency;
210 }
211 return 0;
212}
213
214static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
215{
216 int i;
217 struct acpi_processor_performance *perf;
218
219 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
220 msr &= AMD_MSR_RANGE;
221 else
222 msr &= INTEL_MSR_RANGE;
223
224 perf = data->acpi_data;
225
226 for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
227 if (msr == perf->states[data->freq_table[i].driver_data].status)
228 return data->freq_table[i].frequency;
229 }
230 return data->freq_table[0].frequency;
231}
232
233static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
234{
235 switch (data->cpu_feature) {
236 case SYSTEM_INTEL_MSR_CAPABLE:
237 case SYSTEM_AMD_MSR_CAPABLE:
238 return extract_msr(val, data);
239 case SYSTEM_IO_CAPABLE:
240 return extract_io(val, data);
241 default:
242 return 0;
243 }
244}
245
246struct msr_addr {
247 u32 reg;
248};
249
250struct io_addr {
251 u16 port;
252 u8 bit_width;
253};
254
255struct drv_cmd {
256 unsigned int type;
257 const struct cpumask *mask;
258 union {
259 struct msr_addr msr;
260 struct io_addr io;
261 } addr;
262 u32 val;
263};
264
265/* Called via smp_call_function_single(), on the target CPU */
266static void do_drv_read(void *_cmd)
267{
268 struct drv_cmd *cmd = _cmd;
269 u32 h;
270
271 switch (cmd->type) {
272 case SYSTEM_INTEL_MSR_CAPABLE:
273 case SYSTEM_AMD_MSR_CAPABLE:
274 rdmsr(cmd->addr.msr.reg, cmd->val, h);
275 break;
276 case SYSTEM_IO_CAPABLE:
277 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
278 &cmd->val,
279 (u32)cmd->addr.io.bit_width);
280 break;
281 default:
282 break;
283 }
284}
285
286/* Called via smp_call_function_many(), on the target CPUs */
287static void do_drv_write(void *_cmd)
288{
289 struct drv_cmd *cmd = _cmd;
290 u32 lo, hi;
291
292 switch (cmd->type) {
293 case SYSTEM_INTEL_MSR_CAPABLE:
294 rdmsr(cmd->addr.msr.reg, lo, hi);
295 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
296 wrmsr(cmd->addr.msr.reg, lo, hi);
297 break;
298 case SYSTEM_AMD_MSR_CAPABLE:
299 wrmsr(cmd->addr.msr.reg, cmd->val, 0);
300 break;
301 case SYSTEM_IO_CAPABLE:
302 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
303 cmd->val,
304 (u32)cmd->addr.io.bit_width);
305 break;
306 default:
307 break;
308 }
309}
310
311static void drv_read(struct drv_cmd *cmd)
312{
313 int err;
314 cmd->val = 0;
315
316 err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
317 WARN_ON_ONCE(err); /* smp_call_function_any() was buggy? */
318}
319
320static void drv_write(struct drv_cmd *cmd)
321{
322 int this_cpu;
323
324 this_cpu = get_cpu();
325 if (cpumask_test_cpu(this_cpu, cmd->mask))
326 do_drv_write(cmd);
327 smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
328 put_cpu();
329}
330
331static u32 get_cur_val(const struct cpumask *mask)
332{
333 struct acpi_processor_performance *perf;
334 struct drv_cmd cmd;
335
336 if (unlikely(cpumask_empty(mask)))
337 return 0;
338
339 switch (per_cpu(acfreq_data, cpumask_first(mask))->cpu_feature) {
340 case SYSTEM_INTEL_MSR_CAPABLE:
341 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
342 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
343 break;
344 case SYSTEM_AMD_MSR_CAPABLE:
345 cmd.type = SYSTEM_AMD_MSR_CAPABLE;
346 cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
347 break;
348 case SYSTEM_IO_CAPABLE:
349 cmd.type = SYSTEM_IO_CAPABLE;
350 perf = per_cpu(acfreq_data, cpumask_first(mask))->acpi_data;
351 cmd.addr.io.port = perf->control_register.address;
352 cmd.addr.io.bit_width = perf->control_register.bit_width;
353 break;
354 default:
355 return 0;
356 }
357
358 cmd.mask = mask;
359 drv_read(&cmd);
360
361 pr_debug("get_cur_val = %u\n", cmd.val);
362
363 return cmd.val;
364}
365
366static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
367{
368 struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
369 unsigned int freq;
370 unsigned int cached_freq;
371
372 pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
373
374 if (unlikely(data == NULL ||
375 data->acpi_data == NULL || data->freq_table == NULL)) {
376 return 0;
377 }
378
379 cached_freq = data->freq_table[data->acpi_data->state].frequency;
380 freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
381 if (freq != cached_freq) {
382 /*
383 * The dreaded BIOS frequency change behind our back.
384 * Force set the frequency on next target call.
385 */
386 data->resume = 1;
387 }
388
389 pr_debug("cur freq = %u\n", freq);
390
391 return freq;
392}
393
394static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
395 struct acpi_cpufreq_data *data)
396{
397 unsigned int cur_freq;
398 unsigned int i;
399
400 for (i = 0; i < 100; i++) {
401 cur_freq = extract_freq(get_cur_val(mask), data);
402 if (cur_freq == freq)
403 return 1;
404 udelay(10);
405 }
406 return 0;
407}
408
409static int acpi_cpufreq_target(struct cpufreq_policy *policy,
410 unsigned int index)
411{
412 struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
413 struct acpi_processor_performance *perf;
414 struct drv_cmd cmd;
415 unsigned int next_perf_state = 0; /* Index into perf table */
416 int result = 0;
417
418 if (unlikely(data == NULL ||
419 data->acpi_data == NULL || data->freq_table == NULL)) {
420 return -ENODEV;
421 }
422
423 perf = data->acpi_data;
424 next_perf_state = data->freq_table[index].driver_data;
425 if (perf->state == next_perf_state) {
426 if (unlikely(data->resume)) {
427 pr_debug("Called after resume, resetting to P%d\n",
428 next_perf_state);
429 data->resume = 0;
430 } else {
431 pr_debug("Already at target state (P%d)\n",
432 next_perf_state);
433 goto out;
434 }
435 }
436
437 switch (data->cpu_feature) {
438 case SYSTEM_INTEL_MSR_CAPABLE:
439 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
440 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
441 cmd.val = (u32) perf->states[next_perf_state].control;
442 break;
443 case SYSTEM_AMD_MSR_CAPABLE:
444 cmd.type = SYSTEM_AMD_MSR_CAPABLE;
445 cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
446 cmd.val = (u32) perf->states[next_perf_state].control;
447 break;
448 case SYSTEM_IO_CAPABLE:
449 cmd.type = SYSTEM_IO_CAPABLE;
450 cmd.addr.io.port = perf->control_register.address;
451 cmd.addr.io.bit_width = perf->control_register.bit_width;
452 cmd.val = (u32) perf->states[next_perf_state].control;
453 break;
454 default:
455 result = -ENODEV;
456 goto out;
457 }
458
459 /* cpufreq holds the hotplug lock, so we are safe from here on */
460 if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
461 cmd.mask = policy->cpus;
462 else
463 cmd.mask = cpumask_of(policy->cpu);
464
465 drv_write(&cmd);
466
467 if (acpi_pstate_strict) {
468 if (!check_freqs(cmd.mask, data->freq_table[index].frequency,
469 data)) {
470 pr_debug("acpi_cpufreq_target failed (%d)\n",
471 policy->cpu);
472 result = -EAGAIN;
473 }
474 }
475
476 if (!result)
477 perf->state = next_perf_state;
478
479out:
480 return result;
481}
482
483static unsigned long
484acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
485{
486 struct acpi_processor_performance *perf = data->acpi_data;
487
488 if (cpu_khz) {
489 /* search the closest match to cpu_khz */
490 unsigned int i;
491 unsigned long freq;
492 unsigned long freqn = perf->states[0].core_frequency * 1000;
493
494 for (i = 0; i < (perf->state_count-1); i++) {
495 freq = freqn;
496 freqn = perf->states[i+1].core_frequency * 1000;
497 if ((2 * cpu_khz) > (freqn + freq)) {
498 perf->state = i;
499 return freq;
500 }
501 }
502 perf->state = perf->state_count-1;
503 return freqn;
504 } else {
505 /* assume CPU is at P0... */
506 perf->state = 0;
507 return perf->states[0].core_frequency * 1000;
508 }
509}
510
511static void free_acpi_perf_data(void)
512{
513 unsigned int i;
514
515 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
516 for_each_possible_cpu(i)
517 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
518 ->shared_cpu_map);
519 free_percpu(acpi_perf_data);
520}
521
522static int boost_notify(struct notifier_block *nb, unsigned long action,
523 void *hcpu)
524{
525 unsigned cpu = (long)hcpu;
526 const struct cpumask *cpumask;
527
528 cpumask = get_cpu_mask(cpu);
529
530 /*
531 * Clear the boost-disable bit on the CPU_DOWN path so that
532 * this cpu cannot block the remaining ones from boosting. On
533 * the CPU_UP path we simply keep the boost-disable flag in
534 * sync with the current global state.
535 */
536
537 switch (action) {
538 case CPU_UP_PREPARE:
539 case CPU_UP_PREPARE_FROZEN:
540 boost_set_msrs(acpi_cpufreq_driver.boost_enabled, cpumask);
541 break;
542
543 case CPU_DOWN_PREPARE:
544 case CPU_DOWN_PREPARE_FROZEN:
545 boost_set_msrs(1, cpumask);
546 break;
547
548 default:
549 break;
550 }
551
552 return NOTIFY_OK;
553}
554
555
556static struct notifier_block boost_nb = {
557 .notifier_call = boost_notify,
558};
559
560/*
561 * acpi_cpufreq_early_init - initialize ACPI P-States library
562 *
563 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
564 * in order to determine correct frequency and voltage pairings. We can
565 * do _PDC and _PSD and find out the processor dependency for the
566 * actual init that will happen later...
567 */
568static int __init acpi_cpufreq_early_init(void)
569{
570 unsigned int i;
571 pr_debug("acpi_cpufreq_early_init\n");
572
573 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
574 if (!acpi_perf_data) {
575 pr_debug("Memory allocation error for acpi_perf_data.\n");
576 return -ENOMEM;
577 }
578 for_each_possible_cpu(i) {
579 if (!zalloc_cpumask_var_node(
580 &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
581 GFP_KERNEL, cpu_to_node(i))) {
582
583 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
584 free_acpi_perf_data();
585 return -ENOMEM;
586 }
587 }
588
589 /* Do initialization in ACPI core */
590 acpi_processor_preregister_performance(acpi_perf_data);
591 return 0;
592}
593
594#ifdef CONFIG_SMP
595/*
596 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
597 * or do it in BIOS firmware and won't inform about it to OS. If not
598 * detected, this has a side effect of making CPU run at a different speed
599 * than OS intended it to run at. Detect it and handle it cleanly.
600 */
601static int bios_with_sw_any_bug;
602
603static int sw_any_bug_found(const struct dmi_system_id *d)
604{
605 bios_with_sw_any_bug = 1;
606 return 0;
607}
608
609static const struct dmi_system_id sw_any_bug_dmi_table[] = {
610 {
611 .callback = sw_any_bug_found,
612 .ident = "Supermicro Server X6DLP",
613 .matches = {
614 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
615 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
616 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
617 },
618 },
619 { }
620};
621
622static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
623{
624 /* Intel Xeon Processor 7100 Series Specification Update
625 * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
626 * AL30: A Machine Check Exception (MCE) Occurring during an
627 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
628 * Both Processor Cores to Lock Up. */
629 if (c->x86_vendor == X86_VENDOR_INTEL) {
630 if ((c->x86 == 15) &&
631 (c->x86_model == 6) &&
632 (c->x86_mask == 8)) {
633 printk(KERN_INFO "acpi-cpufreq: Intel(R) "
634 "Xeon(R) 7100 Errata AL30, processors may "
635 "lock up on frequency changes: disabling "
636 "acpi-cpufreq.\n");
637 return -ENODEV;
638 }
639 }
640 return 0;
641}
642#endif
643
644static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
645{
646 unsigned int i;
647 unsigned int valid_states = 0;
648 unsigned int cpu = policy->cpu;
649 struct acpi_cpufreq_data *data;
650 unsigned int result = 0;
651 struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
652 struct acpi_processor_performance *perf;
653#ifdef CONFIG_SMP
654 static int blacklisted;
655#endif
656
657 pr_debug("acpi_cpufreq_cpu_init\n");
658
659#ifdef CONFIG_SMP
660 if (blacklisted)
661 return blacklisted;
662 blacklisted = acpi_cpufreq_blacklist(c);
663 if (blacklisted)
664 return blacklisted;
665#endif
666
667 data = kzalloc(sizeof(*data), GFP_KERNEL);
668 if (!data)
669 return -ENOMEM;
670
671 if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
672 result = -ENOMEM;
673 goto err_free;
674 }
675
676 data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
677 per_cpu(acfreq_data, cpu) = data;
678
679 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
680 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
681
682 result = acpi_processor_register_performance(data->acpi_data, cpu);
683 if (result)
684 goto err_free_mask;
685
686 perf = data->acpi_data;
687 policy->shared_type = perf->shared_type;
688
689 /*
690 * Will let policy->cpus know about dependency only when software
691 * coordination is required.
692 */
693 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
694 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
695 cpumask_copy(policy->cpus, perf->shared_cpu_map);
696 }
697 cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
698
699#ifdef CONFIG_SMP
700 dmi_check_system(sw_any_bug_dmi_table);
701 if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
702 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
703 cpumask_copy(policy->cpus, cpu_core_mask(cpu));
704 }
705
706 if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
707 cpumask_clear(policy->cpus);
708 cpumask_set_cpu(cpu, policy->cpus);
709 cpumask_copy(data->freqdomain_cpus, cpu_sibling_mask(cpu));
710 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
711 pr_info_once(PFX "overriding BIOS provided _PSD data\n");
712 }
713#endif
714
715 /* capability check */
716 if (perf->state_count <= 1) {
717 pr_debug("No P-States\n");
718 result = -ENODEV;
719 goto err_unreg;
720 }
721
722 if (perf->control_register.space_id != perf->status_register.space_id) {
723 result = -ENODEV;
724 goto err_unreg;
725 }
726
727 switch (perf->control_register.space_id) {
728 case ACPI_ADR_SPACE_SYSTEM_IO:
729 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
730 boot_cpu_data.x86 == 0xf) {
731 pr_debug("AMD K8 systems must use native drivers.\n");
732 result = -ENODEV;
733 goto err_unreg;
734 }
735 pr_debug("SYSTEM IO addr space\n");
736 data->cpu_feature = SYSTEM_IO_CAPABLE;
737 break;
738 case ACPI_ADR_SPACE_FIXED_HARDWARE:
739 pr_debug("HARDWARE addr space\n");
740 if (check_est_cpu(cpu)) {
741 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
742 break;
743 }
744 if (check_amd_hwpstate_cpu(cpu)) {
745 data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
746 break;
747 }
748 result = -ENODEV;
749 goto err_unreg;
750 default:
751 pr_debug("Unknown addr space %d\n",
752 (u32) (perf->control_register.space_id));
753 result = -ENODEV;
754 goto err_unreg;
755 }
756
757 data->freq_table = kzalloc(sizeof(*data->freq_table) *
758 (perf->state_count+1), GFP_KERNEL);
759 if (!data->freq_table) {
760 result = -ENOMEM;
761 goto err_unreg;
762 }
763
764 /* detect transition latency */
765 policy->cpuinfo.transition_latency = 0;
766 for (i = 0; i < perf->state_count; i++) {
767 if ((perf->states[i].transition_latency * 1000) >
768 policy->cpuinfo.transition_latency)
769 policy->cpuinfo.transition_latency =
770 perf->states[i].transition_latency * 1000;
771 }
772
773 /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
774 if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
775 policy->cpuinfo.transition_latency > 20 * 1000) {
776 policy->cpuinfo.transition_latency = 20 * 1000;
777 printk_once(KERN_INFO
778 "P-state transition latency capped at 20 uS\n");
779 }
780
781 /* table init */
782 for (i = 0; i < perf->state_count; i++) {
783 if (i > 0 && perf->states[i].core_frequency >=
784 data->freq_table[valid_states-1].frequency / 1000)
785 continue;
786
787 data->freq_table[valid_states].driver_data = i;
788 data->freq_table[valid_states].frequency =
789 perf->states[i].core_frequency * 1000;
790 valid_states++;
791 }
792 data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
793 perf->state = 0;
794
795 result = cpufreq_table_validate_and_show(policy, data->freq_table);
796 if (result)
797 goto err_freqfree;
798
799 if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
800 printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
801
802 switch (perf->control_register.space_id) {
803 case ACPI_ADR_SPACE_SYSTEM_IO:
804 /*
805 * The core will not set policy->cur, because
806 * cpufreq_driver->get is NULL, so we need to set it here.
807 * However, we have to guess it, because the current speed is
808 * unknown and not detectable via IO ports.
809 */
810 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
811 break;
812 case ACPI_ADR_SPACE_FIXED_HARDWARE:
813 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
814 break;
815 default:
816 break;
817 }
818
819 /* notify BIOS that we exist */
820 acpi_processor_notify_smm(THIS_MODULE);
821
822 pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
823 for (i = 0; i < perf->state_count; i++)
824 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n",
825 (i == perf->state ? '*' : ' '), i,
826 (u32) perf->states[i].core_frequency,
827 (u32) perf->states[i].power,
828 (u32) perf->states[i].transition_latency);
829
830 /*
831 * the first call to ->target() should result in us actually
832 * writing something to the appropriate registers.
833 */
834 data->resume = 1;
835
836 return result;
837
838err_freqfree:
839 kfree(data->freq_table);
840err_unreg:
841 acpi_processor_unregister_performance(perf, cpu);
842err_free_mask:
843 free_cpumask_var(data->freqdomain_cpus);
844err_free:
845 kfree(data);
846 per_cpu(acfreq_data, cpu) = NULL;
847
848 return result;
849}
850
851static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
852{
853 struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
854
855 pr_debug("acpi_cpufreq_cpu_exit\n");
856
857 if (data) {
858 per_cpu(acfreq_data, policy->cpu) = NULL;
859 acpi_processor_unregister_performance(data->acpi_data,
860 policy->cpu);
861 free_cpumask_var(data->freqdomain_cpus);
862 kfree(data->freq_table);
863 kfree(data);
864 }
865
866 return 0;
867}
868
869static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
870{
871 struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
872
873 pr_debug("acpi_cpufreq_resume\n");
874
875 data->resume = 1;
876
877 return 0;
878}
879
880static struct freq_attr *acpi_cpufreq_attr[] = {
881 &cpufreq_freq_attr_scaling_available_freqs,
882 &freqdomain_cpus,
883 NULL, /* this is a placeholder for cpb, do not remove */
884 NULL,
885};
886
887static struct cpufreq_driver acpi_cpufreq_driver = {
888 .verify = cpufreq_generic_frequency_table_verify,
889 .target_index = acpi_cpufreq_target,
890 .bios_limit = acpi_processor_get_bios_limit,
891 .init = acpi_cpufreq_cpu_init,
892 .exit = acpi_cpufreq_cpu_exit,
893 .resume = acpi_cpufreq_resume,
894 .name = "acpi-cpufreq",
895 .attr = acpi_cpufreq_attr,
896 .set_boost = _store_boost,
897};
898
899static void __init acpi_cpufreq_boost_init(void)
900{
901 if (boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA)) {
902 msrs = msrs_alloc();
903
904 if (!msrs)
905 return;
906
907 acpi_cpufreq_driver.boost_supported = true;
908 acpi_cpufreq_driver.boost_enabled = boost_state(0);
909
910 cpu_notifier_register_begin();
911
912 /* Force all MSRs to the same value */
913 boost_set_msrs(acpi_cpufreq_driver.boost_enabled,
914 cpu_online_mask);
915
916 __register_cpu_notifier(&boost_nb);
917
918 cpu_notifier_register_done();
919 }
920}
921
922static void acpi_cpufreq_boost_exit(void)
923{
924 if (msrs) {
925 unregister_cpu_notifier(&boost_nb);
926
927 msrs_free(msrs);
928 msrs = NULL;
929 }
930}
931
932static int __init acpi_cpufreq_init(void)
933{
934 int ret;
935
936 if (acpi_disabled)
937 return -ENODEV;
938
939 /* don't keep reloading if cpufreq_driver exists */
940 if (cpufreq_get_current_driver())
941 return -EEXIST;
942
943 pr_debug("acpi_cpufreq_init\n");
944
945 ret = acpi_cpufreq_early_init();
946 if (ret)
947 return ret;
948
949#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
950 /* this is a sysfs file with a strange name and an even stranger
951 * semantic - per CPU instantiation, but system global effect.
952 * Lets enable it only on AMD CPUs for compatibility reasons and
953 * only if configured. This is considered legacy code, which
954 * will probably be removed at some point in the future.
955 */
956 if (check_amd_hwpstate_cpu(0)) {
957 struct freq_attr **iter;
958
959 pr_debug("adding sysfs entry for cpb\n");
960
961 for (iter = acpi_cpufreq_attr; *iter != NULL; iter++)
962 ;
963
964 /* make sure there is a terminator behind it */
965 if (iter[1] == NULL)
966 *iter = &cpb;
967 }
968#endif
969 acpi_cpufreq_boost_init();
970
971 ret = cpufreq_register_driver(&acpi_cpufreq_driver);
972 if (ret) {
973 free_acpi_perf_data();
974 acpi_cpufreq_boost_exit();
975 }
976 return ret;
977}
978
979static void __exit acpi_cpufreq_exit(void)
980{
981 pr_debug("acpi_cpufreq_exit\n");
982
983 acpi_cpufreq_boost_exit();
984
985 cpufreq_unregister_driver(&acpi_cpufreq_driver);
986
987 free_acpi_perf_data();
988}
989
990module_param(acpi_pstate_strict, uint, 0644);
991MODULE_PARM_DESC(acpi_pstate_strict,
992 "value 0 or non-zero. non-zero -> strict ACPI checks are "
993 "performed during frequency changes.");
994
995late_initcall(acpi_cpufreq_init);
996module_exit(acpi_cpufreq_exit);
997
998static const struct x86_cpu_id acpi_cpufreq_ids[] = {
999 X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1000 X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1001 {}
1002};
1003MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1004
1005static const struct acpi_device_id processor_device_ids[] = {
1006 {ACPI_PROCESSOR_OBJECT_HID, },
1007 {ACPI_PROCESSOR_DEVICE_HID, },
1008 {},
1009};
1010MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1011
1012MODULE_ALIAS("acpi");