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#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
29
30#include <linux/kernel.h>
31#include <linux/module.h>
32#include <linux/init.h>
33#include <linux/smp.h>
34#include <linux/sched.h>
35#include <linux/cpufreq.h>
36#include <linux/compiler.h>
37#include <linux/dmi.h>
38#include <linux/slab.h>
39
40#include <linux/acpi.h>
41#include <linux/io.h>
42#include <linux/delay.h>
43#include <linux/uaccess.h>
44
45#include <acpi/processor.h>
46
47#include <asm/msr.h>
48#include <asm/processor.h>
49#include <asm/cpufeature.h>
50
51MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
52MODULE_DESCRIPTION("ACPI Processor P-States Driver");
53MODULE_LICENSE("GPL");
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 unsigned int resume;
69 unsigned int cpu_feature;
70 unsigned int acpi_perf_cpu;
71 cpumask_var_t freqdomain_cpus;
72 void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
73 u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
74};
75
76/* acpi_perf_data is a pointer to percpu data. */
77static struct acpi_processor_performance __percpu *acpi_perf_data;
78
79static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
80{
81 return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
82}
83
84static struct cpufreq_driver acpi_cpufreq_driver;
85
86static unsigned int acpi_pstate_strict;
87
88static bool boost_state(unsigned int cpu)
89{
90 u32 lo, hi;
91 u64 msr;
92
93 switch (boot_cpu_data.x86_vendor) {
94 case X86_VENDOR_INTEL:
95 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
96 msr = lo | ((u64)hi << 32);
97 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
98 case X86_VENDOR_AMD:
99 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
100 msr = lo | ((u64)hi << 32);
101 return !(msr & MSR_K7_HWCR_CPB_DIS);
102 }
103 return false;
104}
105
106static int boost_set_msr(bool enable)
107{
108 u32 msr_addr;
109 u64 msr_mask, val;
110
111 switch (boot_cpu_data.x86_vendor) {
112 case X86_VENDOR_INTEL:
113 msr_addr = MSR_IA32_MISC_ENABLE;
114 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
115 break;
116 case X86_VENDOR_AMD:
117 msr_addr = MSR_K7_HWCR;
118 msr_mask = MSR_K7_HWCR_CPB_DIS;
119 break;
120 default:
121 return -EINVAL;
122 }
123
124 rdmsrl(msr_addr, val);
125
126 if (enable)
127 val &= ~msr_mask;
128 else
129 val |= msr_mask;
130
131 wrmsrl(msr_addr, val);
132 return 0;
133}
134
135static void boost_set_msr_each(void *p_en)
136{
137 bool enable = (bool) p_en;
138
139 boost_set_msr(enable);
140}
141
142static int set_boost(int val)
143{
144 get_online_cpus();
145 on_each_cpu(boost_set_msr_each, (void *)(long)val, 1);
146 put_online_cpus();
147 pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
148
149 return 0;
150}
151
152static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
153{
154 struct acpi_cpufreq_data *data = policy->driver_data;
155
156 if (unlikely(!data))
157 return -ENODEV;
158
159 return cpufreq_show_cpus(data->freqdomain_cpus, buf);
160}
161
162cpufreq_freq_attr_ro(freqdomain_cpus);
163
164#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
165static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
166 size_t count)
167{
168 int ret;
169 unsigned int val = 0;
170
171 if (!acpi_cpufreq_driver.set_boost)
172 return -EINVAL;
173
174 ret = kstrtouint(buf, 10, &val);
175 if (ret || val > 1)
176 return -EINVAL;
177
178 set_boost(val);
179
180 return count;
181}
182
183static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
184{
185 return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
186}
187
188cpufreq_freq_attr_rw(cpb);
189#endif
190
191static int check_est_cpu(unsigned int cpuid)
192{
193 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
194
195 return cpu_has(cpu, X86_FEATURE_EST);
196}
197
198static int check_amd_hwpstate_cpu(unsigned int cpuid)
199{
200 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
201
202 return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
203}
204
205static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
206{
207 struct acpi_cpufreq_data *data = policy->driver_data;
208 struct acpi_processor_performance *perf;
209 int i;
210
211 perf = to_perf_data(data);
212
213 for (i = 0; i < perf->state_count; i++) {
214 if (value == perf->states[i].status)
215 return policy->freq_table[i].frequency;
216 }
217 return 0;
218}
219
220static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
221{
222 struct acpi_cpufreq_data *data = policy->driver_data;
223 struct cpufreq_frequency_table *pos;
224 struct acpi_processor_performance *perf;
225
226 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
227 msr &= AMD_MSR_RANGE;
228 else
229 msr &= INTEL_MSR_RANGE;
230
231 perf = to_perf_data(data);
232
233 cpufreq_for_each_entry(pos, policy->freq_table)
234 if (msr == perf->states[pos->driver_data].status)
235 return pos->frequency;
236 return policy->freq_table[0].frequency;
237}
238
239static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
240{
241 struct acpi_cpufreq_data *data = policy->driver_data;
242
243 switch (data->cpu_feature) {
244 case SYSTEM_INTEL_MSR_CAPABLE:
245 case SYSTEM_AMD_MSR_CAPABLE:
246 return extract_msr(policy, val);
247 case SYSTEM_IO_CAPABLE:
248 return extract_io(policy, val);
249 default:
250 return 0;
251 }
252}
253
254static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
255{
256 u32 val, dummy;
257
258 rdmsr(MSR_IA32_PERF_CTL, val, dummy);
259 return val;
260}
261
262static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
263{
264 u32 lo, hi;
265
266 rdmsr(MSR_IA32_PERF_CTL, lo, hi);
267 lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
268 wrmsr(MSR_IA32_PERF_CTL, lo, hi);
269}
270
271static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
272{
273 u32 val, dummy;
274
275 rdmsr(MSR_AMD_PERF_CTL, val, dummy);
276 return val;
277}
278
279static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
280{
281 wrmsr(MSR_AMD_PERF_CTL, val, 0);
282}
283
284static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
285{
286 u32 val;
287
288 acpi_os_read_port(reg->address, &val, reg->bit_width);
289 return val;
290}
291
292static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
293{
294 acpi_os_write_port(reg->address, val, reg->bit_width);
295}
296
297struct drv_cmd {
298 struct acpi_pct_register *reg;
299 u32 val;
300 union {
301 void (*write)(struct acpi_pct_register *reg, u32 val);
302 u32 (*read)(struct acpi_pct_register *reg);
303 } func;
304};
305
306/* Called via smp_call_function_single(), on the target CPU */
307static void do_drv_read(void *_cmd)
308{
309 struct drv_cmd *cmd = _cmd;
310
311 cmd->val = cmd->func.read(cmd->reg);
312}
313
314static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
315{
316 struct acpi_processor_performance *perf = to_perf_data(data);
317 struct drv_cmd cmd = {
318 .reg = &perf->control_register,
319 .func.read = data->cpu_freq_read,
320 };
321 int err;
322
323 err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
324 WARN_ON_ONCE(err); /* smp_call_function_any() was buggy? */
325 return cmd.val;
326}
327
328/* Called via smp_call_function_many(), on the target CPUs */
329static void do_drv_write(void *_cmd)
330{
331 struct drv_cmd *cmd = _cmd;
332
333 cmd->func.write(cmd->reg, cmd->val);
334}
335
336static void drv_write(struct acpi_cpufreq_data *data,
337 const struct cpumask *mask, u32 val)
338{
339 struct acpi_processor_performance *perf = to_perf_data(data);
340 struct drv_cmd cmd = {
341 .reg = &perf->control_register,
342 .val = val,
343 .func.write = data->cpu_freq_write,
344 };
345 int this_cpu;
346
347 this_cpu = get_cpu();
348 if (cpumask_test_cpu(this_cpu, mask))
349 do_drv_write(&cmd);
350
351 smp_call_function_many(mask, do_drv_write, &cmd, 1);
352 put_cpu();
353}
354
355static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
356{
357 u32 val;
358
359 if (unlikely(cpumask_empty(mask)))
360 return 0;
361
362 val = drv_read(data, mask);
363
364 pr_debug("get_cur_val = %u\n", val);
365
366 return val;
367}
368
369static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
370{
371 struct acpi_cpufreq_data *data;
372 struct cpufreq_policy *policy;
373 unsigned int freq;
374 unsigned int cached_freq;
375
376 pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
377
378 policy = cpufreq_cpu_get_raw(cpu);
379 if (unlikely(!policy))
380 return 0;
381
382 data = policy->driver_data;
383 if (unlikely(!data || !policy->freq_table))
384 return 0;
385
386 cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
387 freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
388 if (freq != cached_freq) {
389 /*
390 * The dreaded BIOS frequency change behind our back.
391 * Force set the frequency on next target call.
392 */
393 data->resume = 1;
394 }
395
396 pr_debug("cur freq = %u\n", freq);
397
398 return freq;
399}
400
401static unsigned int check_freqs(struct cpufreq_policy *policy,
402 const struct cpumask *mask, unsigned int freq)
403{
404 struct acpi_cpufreq_data *data = policy->driver_data;
405 unsigned int cur_freq;
406 unsigned int i;
407
408 for (i = 0; i < 100; i++) {
409 cur_freq = extract_freq(policy, get_cur_val(mask, data));
410 if (cur_freq == freq)
411 return 1;
412 udelay(10);
413 }
414 return 0;
415}
416
417static int acpi_cpufreq_target(struct cpufreq_policy *policy,
418 unsigned int index)
419{
420 struct acpi_cpufreq_data *data = policy->driver_data;
421 struct acpi_processor_performance *perf;
422 const struct cpumask *mask;
423 unsigned int next_perf_state = 0; /* Index into perf table */
424 int result = 0;
425
426 if (unlikely(!data)) {
427 return -ENODEV;
428 }
429
430 perf = to_perf_data(data);
431 next_perf_state = policy->freq_table[index].driver_data;
432 if (perf->state == next_perf_state) {
433 if (unlikely(data->resume)) {
434 pr_debug("Called after resume, resetting to P%d\n",
435 next_perf_state);
436 data->resume = 0;
437 } else {
438 pr_debug("Already at target state (P%d)\n",
439 next_perf_state);
440 return 0;
441 }
442 }
443
444 /*
445 * The core won't allow CPUs to go away until the governor has been
446 * stopped, so we can rely on the stability of policy->cpus.
447 */
448 mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
449 cpumask_of(policy->cpu) : policy->cpus;
450
451 drv_write(data, mask, perf->states[next_perf_state].control);
452
453 if (acpi_pstate_strict) {
454 if (!check_freqs(policy, mask,
455 policy->freq_table[index].frequency)) {
456 pr_debug("acpi_cpufreq_target failed (%d)\n",
457 policy->cpu);
458 result = -EAGAIN;
459 }
460 }
461
462 if (!result)
463 perf->state = next_perf_state;
464
465 return result;
466}
467
468unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
469 unsigned int target_freq)
470{
471 struct acpi_cpufreq_data *data = policy->driver_data;
472 struct acpi_processor_performance *perf;
473 struct cpufreq_frequency_table *entry;
474 unsigned int next_perf_state, next_freq, index;
475
476 /*
477 * Find the closest frequency above target_freq.
478 */
479 if (policy->cached_target_freq == target_freq)
480 index = policy->cached_resolved_idx;
481 else
482 index = cpufreq_table_find_index_dl(policy, target_freq);
483
484 entry = &policy->freq_table[index];
485 next_freq = entry->frequency;
486 next_perf_state = entry->driver_data;
487
488 perf = to_perf_data(data);
489 if (perf->state == next_perf_state) {
490 if (unlikely(data->resume))
491 data->resume = 0;
492 else
493 return next_freq;
494 }
495
496 data->cpu_freq_write(&perf->control_register,
497 perf->states[next_perf_state].control);
498 perf->state = next_perf_state;
499 return next_freq;
500}
501
502static unsigned long
503acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
504{
505 struct acpi_processor_performance *perf;
506
507 perf = to_perf_data(data);
508 if (cpu_khz) {
509 /* search the closest match to cpu_khz */
510 unsigned int i;
511 unsigned long freq;
512 unsigned long freqn = perf->states[0].core_frequency * 1000;
513
514 for (i = 0; i < (perf->state_count-1); i++) {
515 freq = freqn;
516 freqn = perf->states[i+1].core_frequency * 1000;
517 if ((2 * cpu_khz) > (freqn + freq)) {
518 perf->state = i;
519 return freq;
520 }
521 }
522 perf->state = perf->state_count-1;
523 return freqn;
524 } else {
525 /* assume CPU is at P0... */
526 perf->state = 0;
527 return perf->states[0].core_frequency * 1000;
528 }
529}
530
531static void free_acpi_perf_data(void)
532{
533 unsigned int i;
534
535 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
536 for_each_possible_cpu(i)
537 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
538 ->shared_cpu_map);
539 free_percpu(acpi_perf_data);
540}
541
542static int cpufreq_boost_online(unsigned int cpu)
543{
544 /*
545 * On the CPU_UP path we simply keep the boost-disable flag
546 * in sync with the current global state.
547 */
548 return boost_set_msr(acpi_cpufreq_driver.boost_enabled);
549}
550
551static int cpufreq_boost_down_prep(unsigned int cpu)
552{
553 /*
554 * Clear the boost-disable bit on the CPU_DOWN path so that
555 * this cpu cannot block the remaining ones from boosting.
556 */
557 return boost_set_msr(1);
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 pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
634 return -ENODEV;
635 }
636 }
637 return 0;
638}
639#endif
640
641static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
642{
643 unsigned int i;
644 unsigned int valid_states = 0;
645 unsigned int cpu = policy->cpu;
646 struct acpi_cpufreq_data *data;
647 unsigned int result = 0;
648 struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
649 struct acpi_processor_performance *perf;
650 struct cpufreq_frequency_table *freq_table;
651#ifdef CONFIG_SMP
652 static int blacklisted;
653#endif
654
655 pr_debug("acpi_cpufreq_cpu_init\n");
656
657#ifdef CONFIG_SMP
658 if (blacklisted)
659 return blacklisted;
660 blacklisted = acpi_cpufreq_blacklist(c);
661 if (blacklisted)
662 return blacklisted;
663#endif
664
665 data = kzalloc(sizeof(*data), GFP_KERNEL);
666 if (!data)
667 return -ENOMEM;
668
669 if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
670 result = -ENOMEM;
671 goto err_free;
672 }
673
674 perf = per_cpu_ptr(acpi_perf_data, cpu);
675 data->acpi_perf_cpu = cpu;
676 policy->driver_data = data;
677
678 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
679 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
680
681 result = acpi_processor_register_performance(perf, cpu);
682 if (result)
683 goto err_free_mask;
684
685 policy->shared_type = perf->shared_type;
686
687 /*
688 * Will let policy->cpus know about dependency only when software
689 * coordination is required.
690 */
691 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
692 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
693 cpumask_copy(policy->cpus, perf->shared_cpu_map);
694 }
695 cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
696
697#ifdef CONFIG_SMP
698 dmi_check_system(sw_any_bug_dmi_table);
699 if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
700 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
701 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
702 }
703
704 if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
705 cpumask_clear(policy->cpus);
706 cpumask_set_cpu(cpu, policy->cpus);
707 cpumask_copy(data->freqdomain_cpus,
708 topology_sibling_cpumask(cpu));
709 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
710 pr_info_once("overriding BIOS provided _PSD data\n");
711 }
712#endif
713
714 /* capability check */
715 if (perf->state_count <= 1) {
716 pr_debug("No P-States\n");
717 result = -ENODEV;
718 goto err_unreg;
719 }
720
721 if (perf->control_register.space_id != perf->status_register.space_id) {
722 result = -ENODEV;
723 goto err_unreg;
724 }
725
726 switch (perf->control_register.space_id) {
727 case ACPI_ADR_SPACE_SYSTEM_IO:
728 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
729 boot_cpu_data.x86 == 0xf) {
730 pr_debug("AMD K8 systems must use native drivers.\n");
731 result = -ENODEV;
732 goto err_unreg;
733 }
734 pr_debug("SYSTEM IO addr space\n");
735 data->cpu_feature = SYSTEM_IO_CAPABLE;
736 data->cpu_freq_read = cpu_freq_read_io;
737 data->cpu_freq_write = cpu_freq_write_io;
738 break;
739 case ACPI_ADR_SPACE_FIXED_HARDWARE:
740 pr_debug("HARDWARE addr space\n");
741 if (check_est_cpu(cpu)) {
742 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
743 data->cpu_freq_read = cpu_freq_read_intel;
744 data->cpu_freq_write = cpu_freq_write_intel;
745 break;
746 }
747 if (check_amd_hwpstate_cpu(cpu)) {
748 data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
749 data->cpu_freq_read = cpu_freq_read_amd;
750 data->cpu_freq_write = cpu_freq_write_amd;
751 break;
752 }
753 result = -ENODEV;
754 goto err_unreg;
755 default:
756 pr_debug("Unknown addr space %d\n",
757 (u32) (perf->control_register.space_id));
758 result = -ENODEV;
759 goto err_unreg;
760 }
761
762 freq_table = kzalloc(sizeof(*freq_table) *
763 (perf->state_count+1), GFP_KERNEL);
764 if (!freq_table) {
765 result = -ENOMEM;
766 goto err_unreg;
767 }
768
769 /* detect transition latency */
770 policy->cpuinfo.transition_latency = 0;
771 for (i = 0; i < perf->state_count; i++) {
772 if ((perf->states[i].transition_latency * 1000) >
773 policy->cpuinfo.transition_latency)
774 policy->cpuinfo.transition_latency =
775 perf->states[i].transition_latency * 1000;
776 }
777
778 /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
779 if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
780 policy->cpuinfo.transition_latency > 20 * 1000) {
781 policy->cpuinfo.transition_latency = 20 * 1000;
782 pr_info_once("P-state transition latency capped at 20 uS\n");
783 }
784
785 /* table init */
786 for (i = 0; i < perf->state_count; i++) {
787 if (i > 0 && perf->states[i].core_frequency >=
788 freq_table[valid_states-1].frequency / 1000)
789 continue;
790
791 freq_table[valid_states].driver_data = i;
792 freq_table[valid_states].frequency =
793 perf->states[i].core_frequency * 1000;
794 valid_states++;
795 }
796 freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
797 perf->state = 0;
798
799 result = cpufreq_table_validate_and_show(policy, freq_table);
800 if (result)
801 goto err_freqfree;
802
803 if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
804 pr_warn(FW_WARN "P-state 0 is not max freq\n");
805
806 switch (perf->control_register.space_id) {
807 case ACPI_ADR_SPACE_SYSTEM_IO:
808 /*
809 * The core will not set policy->cur, because
810 * cpufreq_driver->get is NULL, so we need to set it here.
811 * However, we have to guess it, because the current speed is
812 * unknown and not detectable via IO ports.
813 */
814 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
815 break;
816 case ACPI_ADR_SPACE_FIXED_HARDWARE:
817 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
818 break;
819 default:
820 break;
821 }
822
823 /* notify BIOS that we exist */
824 acpi_processor_notify_smm(THIS_MODULE);
825
826 pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
827 for (i = 0; i < perf->state_count; i++)
828 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n",
829 (i == perf->state ? '*' : ' '), i,
830 (u32) perf->states[i].core_frequency,
831 (u32) perf->states[i].power,
832 (u32) perf->states[i].transition_latency);
833
834 /*
835 * the first call to ->target() should result in us actually
836 * writing something to the appropriate registers.
837 */
838 data->resume = 1;
839
840 policy->fast_switch_possible = !acpi_pstate_strict &&
841 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
842
843 return result;
844
845err_freqfree:
846 kfree(freq_table);
847err_unreg:
848 acpi_processor_unregister_performance(cpu);
849err_free_mask:
850 free_cpumask_var(data->freqdomain_cpus);
851err_free:
852 kfree(data);
853 policy->driver_data = NULL;
854
855 return result;
856}
857
858static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
859{
860 struct acpi_cpufreq_data *data = policy->driver_data;
861
862 pr_debug("acpi_cpufreq_cpu_exit\n");
863
864 policy->fast_switch_possible = false;
865 policy->driver_data = NULL;
866 acpi_processor_unregister_performance(data->acpi_perf_cpu);
867 free_cpumask_var(data->freqdomain_cpus);
868 kfree(policy->freq_table);
869 kfree(data);
870
871 return 0;
872}
873
874static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
875{
876 struct acpi_cpufreq_data *data = policy->driver_data;
877
878 pr_debug("acpi_cpufreq_resume\n");
879
880 data->resume = 1;
881
882 return 0;
883}
884
885static struct freq_attr *acpi_cpufreq_attr[] = {
886 &cpufreq_freq_attr_scaling_available_freqs,
887 &freqdomain_cpus,
888#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
889 &cpb,
890#endif
891 NULL,
892};
893
894static struct cpufreq_driver acpi_cpufreq_driver = {
895 .verify = cpufreq_generic_frequency_table_verify,
896 .target_index = acpi_cpufreq_target,
897 .fast_switch = acpi_cpufreq_fast_switch,
898 .bios_limit = acpi_processor_get_bios_limit,
899 .init = acpi_cpufreq_cpu_init,
900 .exit = acpi_cpufreq_cpu_exit,
901 .resume = acpi_cpufreq_resume,
902 .name = "acpi-cpufreq",
903 .attr = acpi_cpufreq_attr,
904};
905
906static enum cpuhp_state acpi_cpufreq_online;
907
908static void __init acpi_cpufreq_boost_init(void)
909{
910 int ret;
911
912 if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA)))
913 return;
914
915 acpi_cpufreq_driver.set_boost = set_boost;
916 acpi_cpufreq_driver.boost_enabled = boost_state(0);
917
918 /*
919 * This calls the online callback on all online cpu and forces all
920 * MSRs to the same value.
921 */
922 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "cpufreq/acpi:online",
923 cpufreq_boost_online, cpufreq_boost_down_prep);
924 if (ret < 0) {
925 pr_err("acpi_cpufreq: failed to register hotplug callbacks\n");
926 return;
927 }
928 acpi_cpufreq_online = ret;
929}
930
931static void acpi_cpufreq_boost_exit(void)
932{
933 if (acpi_cpufreq_online > 0)
934 cpuhp_remove_state_nocalls(acpi_cpufreq_online);
935}
936
937static int __init acpi_cpufreq_init(void)
938{
939 int ret;
940
941 if (acpi_disabled)
942 return -ENODEV;
943
944 /* don't keep reloading if cpufreq_driver exists */
945 if (cpufreq_get_current_driver())
946 return -EEXIST;
947
948 pr_debug("acpi_cpufreq_init\n");
949
950 ret = acpi_cpufreq_early_init();
951 if (ret)
952 return ret;
953
954#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
955 /* this is a sysfs file with a strange name and an even stranger
956 * semantic - per CPU instantiation, but system global effect.
957 * Lets enable it only on AMD CPUs for compatibility reasons and
958 * only if configured. This is considered legacy code, which
959 * will probably be removed at some point in the future.
960 */
961 if (!check_amd_hwpstate_cpu(0)) {
962 struct freq_attr **attr;
963
964 pr_debug("CPB unsupported, do not expose it\n");
965
966 for (attr = acpi_cpufreq_attr; *attr; attr++)
967 if (*attr == &cpb) {
968 *attr = NULL;
969 break;
970 }
971 }
972#endif
973 acpi_cpufreq_boost_init();
974
975 ret = cpufreq_register_driver(&acpi_cpufreq_driver);
976 if (ret) {
977 free_acpi_perf_data();
978 acpi_cpufreq_boost_exit();
979 }
980 return ret;
981}
982
983static void __exit acpi_cpufreq_exit(void)
984{
985 pr_debug("acpi_cpufreq_exit\n");
986
987 acpi_cpufreq_boost_exit();
988
989 cpufreq_unregister_driver(&acpi_cpufreq_driver);
990
991 free_acpi_perf_data();
992}
993
994module_param(acpi_pstate_strict, uint, 0644);
995MODULE_PARM_DESC(acpi_pstate_strict,
996 "value 0 or non-zero. non-zero -> strict ACPI checks are "
997 "performed during frequency changes.");
998
999late_initcall(acpi_cpufreq_init);
1000module_exit(acpi_cpufreq_exit);
1001
1002static const struct x86_cpu_id acpi_cpufreq_ids[] = {
1003 X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1004 X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1005 {}
1006};
1007MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1008
1009static const struct acpi_device_id processor_device_ids[] = {
1010 {ACPI_PROCESSOR_OBJECT_HID, },
1011 {ACPI_PROCESSOR_DEVICE_HID, },
1012 {},
1013};
1014MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1015
1016MODULE_ALIAS("acpi");