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// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * acpi-cpufreq.c - ACPI Processor P-States Driver
4 *
5 * Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
6 * Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
7 * Copyright (C) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
8 * Copyright (C) 2006 Denis Sadykov <denis.m.sadykov@intel.com>
9 */
10
11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12
13#include <linux/kernel.h>
14#include <linux/module.h>
15#include <linux/init.h>
16#include <linux/smp.h>
17#include <linux/sched.h>
18#include <linux/cpufreq.h>
19#include <linux/compiler.h>
20#include <linux/dmi.h>
21#include <linux/slab.h>
22#include <linux/string_helpers.h>
23#include <linux/platform_device.h>
24
25#include <linux/acpi.h>
26#include <linux/io.h>
27#include <linux/delay.h>
28#include <linux/uaccess.h>
29
30#include <acpi/processor.h>
31#include <acpi/cppc_acpi.h>
32
33#include <asm/msr.h>
34#include <asm/processor.h>
35#include <asm/cpufeature.h>
36#include <asm/cpu_device_id.h>
37
38MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
39MODULE_DESCRIPTION("ACPI Processor P-States Driver");
40MODULE_LICENSE("GPL");
41
42enum {
43 UNDEFINED_CAPABLE = 0,
44 SYSTEM_INTEL_MSR_CAPABLE,
45 SYSTEM_AMD_MSR_CAPABLE,
46 SYSTEM_IO_CAPABLE,
47};
48
49#define INTEL_MSR_RANGE (0xffff)
50#define AMD_MSR_RANGE (0x7)
51#define HYGON_MSR_RANGE (0x7)
52
53struct acpi_cpufreq_data {
54 unsigned int resume;
55 unsigned int cpu_feature;
56 unsigned int acpi_perf_cpu;
57 cpumask_var_t freqdomain_cpus;
58 void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
59 u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
60};
61
62/* acpi_perf_data is a pointer to percpu data. */
63static struct acpi_processor_performance __percpu *acpi_perf_data;
64
65static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
66{
67 return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
68}
69
70static struct cpufreq_driver acpi_cpufreq_driver;
71
72static unsigned int acpi_pstate_strict;
73
74static bool boost_state(unsigned int cpu)
75{
76 u64 msr;
77
78 switch (boot_cpu_data.x86_vendor) {
79 case X86_VENDOR_INTEL:
80 case X86_VENDOR_CENTAUR:
81 case X86_VENDOR_ZHAOXIN:
82 rdmsrl_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &msr);
83 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
84 case X86_VENDOR_HYGON:
85 case X86_VENDOR_AMD:
86 rdmsrl_on_cpu(cpu, MSR_K7_HWCR, &msr);
87 return !(msr & MSR_K7_HWCR_CPB_DIS);
88 }
89 return false;
90}
91
92static int boost_set_msr(bool enable)
93{
94 u32 msr_addr;
95 u64 msr_mask, val;
96
97 switch (boot_cpu_data.x86_vendor) {
98 case X86_VENDOR_INTEL:
99 case X86_VENDOR_CENTAUR:
100 case X86_VENDOR_ZHAOXIN:
101 msr_addr = MSR_IA32_MISC_ENABLE;
102 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
103 break;
104 case X86_VENDOR_HYGON:
105 case X86_VENDOR_AMD:
106 msr_addr = MSR_K7_HWCR;
107 msr_mask = MSR_K7_HWCR_CPB_DIS;
108 break;
109 default:
110 return -EINVAL;
111 }
112
113 rdmsrl(msr_addr, val);
114
115 if (enable)
116 val &= ~msr_mask;
117 else
118 val |= msr_mask;
119
120 wrmsrl(msr_addr, val);
121 return 0;
122}
123
124static void boost_set_msr_each(void *p_en)
125{
126 bool enable = (bool) p_en;
127
128 boost_set_msr(enable);
129}
130
131static int set_boost(struct cpufreq_policy *policy, int val)
132{
133 on_each_cpu_mask(policy->cpus, boost_set_msr_each,
134 (void *)(long)val, 1);
135 pr_debug("CPU %*pbl: Core Boosting %s.\n",
136 cpumask_pr_args(policy->cpus), str_enabled_disabled(val));
137
138 return 0;
139}
140
141static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
142{
143 struct acpi_cpufreq_data *data = policy->driver_data;
144
145 if (unlikely(!data))
146 return -ENODEV;
147
148 return cpufreq_show_cpus(data->freqdomain_cpus, buf);
149}
150
151cpufreq_freq_attr_ro(freqdomain_cpus);
152
153#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
154static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
155 size_t count)
156{
157 int ret;
158 unsigned int val = 0;
159
160 if (!acpi_cpufreq_driver.set_boost)
161 return -EINVAL;
162
163 ret = kstrtouint(buf, 10, &val);
164 if (ret || val > 1)
165 return -EINVAL;
166
167 cpus_read_lock();
168 set_boost(policy, val);
169 cpus_read_unlock();
170
171 return count;
172}
173
174static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
175{
176 return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
177}
178
179cpufreq_freq_attr_rw(cpb);
180#endif
181
182static int check_est_cpu(unsigned int cpuid)
183{
184 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
185
186 return cpu_has(cpu, X86_FEATURE_EST);
187}
188
189static int check_amd_hwpstate_cpu(unsigned int cpuid)
190{
191 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
192
193 return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
194}
195
196static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
197{
198 struct acpi_cpufreq_data *data = policy->driver_data;
199 struct acpi_processor_performance *perf;
200 int i;
201
202 perf = to_perf_data(data);
203
204 for (i = 0; i < perf->state_count; i++) {
205 if (value == perf->states[i].status)
206 return policy->freq_table[i].frequency;
207 }
208 return 0;
209}
210
211static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
212{
213 struct acpi_cpufreq_data *data = policy->driver_data;
214 struct cpufreq_frequency_table *pos;
215 struct acpi_processor_performance *perf;
216
217 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
218 msr &= AMD_MSR_RANGE;
219 else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
220 msr &= HYGON_MSR_RANGE;
221 else
222 msr &= INTEL_MSR_RANGE;
223
224 perf = to_perf_data(data);
225
226 cpufreq_for_each_entry(pos, policy->freq_table)
227 if (msr == perf->states[pos->driver_data].status)
228 return pos->frequency;
229 return policy->freq_table[0].frequency;
230}
231
232static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
233{
234 struct acpi_cpufreq_data *data = policy->driver_data;
235
236 switch (data->cpu_feature) {
237 case SYSTEM_INTEL_MSR_CAPABLE:
238 case SYSTEM_AMD_MSR_CAPABLE:
239 return extract_msr(policy, val);
240 case SYSTEM_IO_CAPABLE:
241 return extract_io(policy, val);
242 default:
243 return 0;
244 }
245}
246
247static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
248{
249 u32 val, dummy __always_unused;
250
251 rdmsr(MSR_IA32_PERF_CTL, val, dummy);
252 return val;
253}
254
255static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
256{
257 u32 lo, hi;
258
259 rdmsr(MSR_IA32_PERF_CTL, lo, hi);
260 lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
261 wrmsr(MSR_IA32_PERF_CTL, lo, hi);
262}
263
264static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
265{
266 u32 val, dummy __always_unused;
267
268 rdmsr(MSR_AMD_PERF_CTL, val, dummy);
269 return val;
270}
271
272static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
273{
274 wrmsr(MSR_AMD_PERF_CTL, val, 0);
275}
276
277static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
278{
279 u32 val;
280
281 acpi_os_read_port(reg->address, &val, reg->bit_width);
282 return val;
283}
284
285static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
286{
287 acpi_os_write_port(reg->address, val, reg->bit_width);
288}
289
290struct drv_cmd {
291 struct acpi_pct_register *reg;
292 u32 val;
293 union {
294 void (*write)(struct acpi_pct_register *reg, u32 val);
295 u32 (*read)(struct acpi_pct_register *reg);
296 } func;
297};
298
299/* Called via smp_call_function_single(), on the target CPU */
300static void do_drv_read(void *_cmd)
301{
302 struct drv_cmd *cmd = _cmd;
303
304 cmd->val = cmd->func.read(cmd->reg);
305}
306
307static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
308{
309 struct acpi_processor_performance *perf = to_perf_data(data);
310 struct drv_cmd cmd = {
311 .reg = &perf->control_register,
312 .func.read = data->cpu_freq_read,
313 };
314 int err;
315
316 err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
317 WARN_ON_ONCE(err); /* smp_call_function_any() was buggy? */
318 return cmd.val;
319}
320
321/* Called via smp_call_function_many(), on the target CPUs */
322static void do_drv_write(void *_cmd)
323{
324 struct drv_cmd *cmd = _cmd;
325
326 cmd->func.write(cmd->reg, cmd->val);
327}
328
329static void drv_write(struct acpi_cpufreq_data *data,
330 const struct cpumask *mask, u32 val)
331{
332 struct acpi_processor_performance *perf = to_perf_data(data);
333 struct drv_cmd cmd = {
334 .reg = &perf->control_register,
335 .val = val,
336 .func.write = data->cpu_freq_write,
337 };
338 int this_cpu;
339
340 this_cpu = get_cpu();
341 if (cpumask_test_cpu(this_cpu, mask))
342 do_drv_write(&cmd);
343
344 smp_call_function_many(mask, do_drv_write, &cmd, 1);
345 put_cpu();
346}
347
348static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
349{
350 u32 val;
351
352 if (unlikely(cpumask_empty(mask)))
353 return 0;
354
355 val = drv_read(data, mask);
356
357 pr_debug("%s = %u\n", __func__, val);
358
359 return val;
360}
361
362static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
363{
364 struct acpi_cpufreq_data *data;
365 struct cpufreq_policy *policy;
366 unsigned int freq;
367 unsigned int cached_freq;
368
369 pr_debug("%s (%d)\n", __func__, cpu);
370
371 policy = cpufreq_cpu_get_raw(cpu);
372 if (unlikely(!policy))
373 return 0;
374
375 data = policy->driver_data;
376 if (unlikely(!data || !policy->freq_table))
377 return 0;
378
379 cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
380 freq = extract_freq(policy, 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(struct cpufreq_policy *policy,
395 const struct cpumask *mask, unsigned int freq)
396{
397 struct acpi_cpufreq_data *data = policy->driver_data;
398 unsigned int cur_freq;
399 unsigned int i;
400
401 for (i = 0; i < 100; i++) {
402 cur_freq = extract_freq(policy, get_cur_val(mask, data));
403 if (cur_freq == freq)
404 return 1;
405 udelay(10);
406 }
407 return 0;
408}
409
410static int acpi_cpufreq_target(struct cpufreq_policy *policy,
411 unsigned int index)
412{
413 struct acpi_cpufreq_data *data = policy->driver_data;
414 struct acpi_processor_performance *perf;
415 const struct cpumask *mask;
416 unsigned int next_perf_state = 0; /* Index into perf table */
417 int result = 0;
418
419 if (unlikely(!data)) {
420 return -ENODEV;
421 }
422
423 perf = to_perf_data(data);
424 next_perf_state = policy->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 return 0;
434 }
435 }
436
437 /*
438 * The core won't allow CPUs to go away until the governor has been
439 * stopped, so we can rely on the stability of policy->cpus.
440 */
441 mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
442 cpumask_of(policy->cpu) : policy->cpus;
443
444 drv_write(data, mask, perf->states[next_perf_state].control);
445
446 if (acpi_pstate_strict) {
447 if (!check_freqs(policy, mask,
448 policy->freq_table[index].frequency)) {
449 pr_debug("%s (%d)\n", __func__, policy->cpu);
450 result = -EAGAIN;
451 }
452 }
453
454 if (!result)
455 perf->state = next_perf_state;
456
457 return result;
458}
459
460static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
461 unsigned int target_freq)
462{
463 struct acpi_cpufreq_data *data = policy->driver_data;
464 struct acpi_processor_performance *perf;
465 struct cpufreq_frequency_table *entry;
466 unsigned int next_perf_state, next_freq, index;
467
468 /*
469 * Find the closest frequency above target_freq.
470 */
471 if (policy->cached_target_freq == target_freq)
472 index = policy->cached_resolved_idx;
473 else
474 index = cpufreq_table_find_index_dl(policy, target_freq,
475 false);
476
477 entry = &policy->freq_table[index];
478 next_freq = entry->frequency;
479 next_perf_state = entry->driver_data;
480
481 perf = to_perf_data(data);
482 if (perf->state == next_perf_state) {
483 if (unlikely(data->resume))
484 data->resume = 0;
485 else
486 return next_freq;
487 }
488
489 data->cpu_freq_write(&perf->control_register,
490 perf->states[next_perf_state].control);
491 perf->state = next_perf_state;
492 return next_freq;
493}
494
495static unsigned long
496acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
497{
498 struct acpi_processor_performance *perf;
499
500 perf = to_perf_data(data);
501 if (cpu_khz) {
502 /* search the closest match to cpu_khz */
503 unsigned int i;
504 unsigned long freq;
505 unsigned long freqn = perf->states[0].core_frequency * 1000;
506
507 for (i = 0; i < (perf->state_count-1); i++) {
508 freq = freqn;
509 freqn = perf->states[i+1].core_frequency * 1000;
510 if ((2 * cpu_khz) > (freqn + freq)) {
511 perf->state = i;
512 return freq;
513 }
514 }
515 perf->state = perf->state_count-1;
516 return freqn;
517 } else {
518 /* assume CPU is at P0... */
519 perf->state = 0;
520 return perf->states[0].core_frequency * 1000;
521 }
522}
523
524static void free_acpi_perf_data(void)
525{
526 unsigned int i;
527
528 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
529 for_each_possible_cpu(i)
530 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
531 ->shared_cpu_map);
532 free_percpu(acpi_perf_data);
533}
534
535static int cpufreq_boost_down_prep(unsigned int cpu)
536{
537 /*
538 * Clear the boost-disable bit on the CPU_DOWN path so that
539 * this cpu cannot block the remaining ones from boosting.
540 */
541 return boost_set_msr(1);
542}
543
544/*
545 * acpi_cpufreq_early_init - initialize ACPI P-States library
546 *
547 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
548 * in order to determine correct frequency and voltage pairings. We can
549 * do _PDC and _PSD and find out the processor dependency for the
550 * actual init that will happen later...
551 */
552static int __init acpi_cpufreq_early_init(void)
553{
554 unsigned int i;
555 pr_debug("%s\n", __func__);
556
557 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
558 if (!acpi_perf_data) {
559 pr_debug("Memory allocation error for acpi_perf_data.\n");
560 return -ENOMEM;
561 }
562 for_each_possible_cpu(i) {
563 if (!zalloc_cpumask_var_node(
564 &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
565 GFP_KERNEL, cpu_to_node(i))) {
566
567 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
568 free_acpi_perf_data();
569 return -ENOMEM;
570 }
571 }
572
573 /* Do initialization in ACPI core */
574 acpi_processor_preregister_performance(acpi_perf_data);
575 return 0;
576}
577
578#ifdef CONFIG_SMP
579/*
580 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
581 * or do it in BIOS firmware and won't inform about it to OS. If not
582 * detected, this has a side effect of making CPU run at a different speed
583 * than OS intended it to run at. Detect it and handle it cleanly.
584 */
585static int bios_with_sw_any_bug;
586
587static int sw_any_bug_found(const struct dmi_system_id *d)
588{
589 bios_with_sw_any_bug = 1;
590 return 0;
591}
592
593static const struct dmi_system_id sw_any_bug_dmi_table[] = {
594 {
595 .callback = sw_any_bug_found,
596 .ident = "Supermicro Server X6DLP",
597 .matches = {
598 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
599 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
600 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
601 },
602 },
603 { }
604};
605
606static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
607{
608 /* Intel Xeon Processor 7100 Series Specification Update
609 * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
610 * AL30: A Machine Check Exception (MCE) Occurring during an
611 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
612 * Both Processor Cores to Lock Up. */
613 if (c->x86_vendor == X86_VENDOR_INTEL) {
614 if ((c->x86 == 15) &&
615 (c->x86_model == 6) &&
616 (c->x86_stepping == 8)) {
617 pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
618 return -ENODEV;
619 }
620 }
621 return 0;
622}
623#endif
624
625#ifdef CONFIG_ACPI_CPPC_LIB
626/*
627 * get_max_boost_ratio: Computes the max_boost_ratio as the ratio
628 * between the highest_perf and the nominal_perf.
629 *
630 * Returns the max_boost_ratio for @cpu. Returns the CPPC nominal
631 * frequency via @nominal_freq if it is non-NULL pointer.
632 */
633static u64 get_max_boost_ratio(unsigned int cpu, u64 *nominal_freq)
634{
635 struct cppc_perf_caps perf_caps;
636 u64 highest_perf, nominal_perf;
637 int ret;
638
639 if (acpi_pstate_strict)
640 return 0;
641
642 ret = cppc_get_perf_caps(cpu, &perf_caps);
643 if (ret) {
644 pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
645 cpu, ret);
646 return 0;
647 }
648
649 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
650 ret = amd_get_boost_ratio_numerator(cpu, &highest_perf);
651 if (ret) {
652 pr_debug("CPU%d: Unable to get boost ratio numerator (%d)\n",
653 cpu, ret);
654 return 0;
655 }
656 } else {
657 highest_perf = perf_caps.highest_perf;
658 }
659
660 nominal_perf = perf_caps.nominal_perf;
661
662 if (nominal_freq)
663 *nominal_freq = perf_caps.nominal_freq;
664
665 if (!highest_perf || !nominal_perf) {
666 pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
667 return 0;
668 }
669
670 if (highest_perf < nominal_perf) {
671 pr_debug("CPU%d: nominal performance above highest\n", cpu);
672 return 0;
673 }
674
675 return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
676}
677
678#else
679static inline u64 get_max_boost_ratio(unsigned int cpu, u64 *nominal_freq)
680{
681 return 0;
682}
683#endif
684
685static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
686{
687 struct cpufreq_frequency_table *freq_table;
688 struct acpi_processor_performance *perf;
689 struct acpi_cpufreq_data *data;
690 unsigned int cpu = policy->cpu;
691 struct cpuinfo_x86 *c = &cpu_data(cpu);
692 u64 max_boost_ratio, nominal_freq = 0;
693 unsigned int valid_states = 0;
694 unsigned int result = 0;
695 unsigned int i;
696#ifdef CONFIG_SMP
697 static int blacklisted;
698#endif
699
700 pr_debug("%s\n", __func__);
701
702#ifdef CONFIG_SMP
703 if (blacklisted)
704 return blacklisted;
705 blacklisted = acpi_cpufreq_blacklist(c);
706 if (blacklisted)
707 return blacklisted;
708#endif
709
710 data = kzalloc(sizeof(*data), GFP_KERNEL);
711 if (!data)
712 return -ENOMEM;
713
714 if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
715 result = -ENOMEM;
716 goto err_free;
717 }
718
719 perf = per_cpu_ptr(acpi_perf_data, cpu);
720 data->acpi_perf_cpu = cpu;
721 policy->driver_data = data;
722
723 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
724 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
725
726 result = acpi_processor_register_performance(perf, cpu);
727 if (result)
728 goto err_free_mask;
729
730 policy->shared_type = perf->shared_type;
731
732 /*
733 * Will let policy->cpus know about dependency only when software
734 * coordination is required.
735 */
736 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
737 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
738 cpumask_copy(policy->cpus, perf->shared_cpu_map);
739 }
740 cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
741
742#ifdef CONFIG_SMP
743 dmi_check_system(sw_any_bug_dmi_table);
744 if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
745 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
746 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
747 }
748
749 if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
750 !acpi_pstate_strict) {
751 cpumask_clear(policy->cpus);
752 cpumask_set_cpu(cpu, policy->cpus);
753 cpumask_copy(data->freqdomain_cpus,
754 topology_sibling_cpumask(cpu));
755 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
756 pr_info_once("overriding BIOS provided _PSD data\n");
757 }
758#endif
759
760 /* capability check */
761 if (perf->state_count <= 1) {
762 pr_debug("No P-States\n");
763 result = -ENODEV;
764 goto err_unreg;
765 }
766
767 if (perf->control_register.space_id != perf->status_register.space_id) {
768 result = -ENODEV;
769 goto err_unreg;
770 }
771
772 switch (perf->control_register.space_id) {
773 case ACPI_ADR_SPACE_SYSTEM_IO:
774 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
775 boot_cpu_data.x86 == 0xf) {
776 pr_debug("AMD K8 systems must use native drivers.\n");
777 result = -ENODEV;
778 goto err_unreg;
779 }
780 pr_debug("SYSTEM IO addr space\n");
781 data->cpu_feature = SYSTEM_IO_CAPABLE;
782 data->cpu_freq_read = cpu_freq_read_io;
783 data->cpu_freq_write = cpu_freq_write_io;
784 break;
785 case ACPI_ADR_SPACE_FIXED_HARDWARE:
786 pr_debug("HARDWARE addr space\n");
787 if (check_est_cpu(cpu)) {
788 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
789 data->cpu_freq_read = cpu_freq_read_intel;
790 data->cpu_freq_write = cpu_freq_write_intel;
791 break;
792 }
793 if (check_amd_hwpstate_cpu(cpu)) {
794 data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
795 data->cpu_freq_read = cpu_freq_read_amd;
796 data->cpu_freq_write = cpu_freq_write_amd;
797 break;
798 }
799 result = -ENODEV;
800 goto err_unreg;
801 default:
802 pr_debug("Unknown addr space %d\n",
803 (u32) (perf->control_register.space_id));
804 result = -ENODEV;
805 goto err_unreg;
806 }
807
808 freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
809 GFP_KERNEL);
810 if (!freq_table) {
811 result = -ENOMEM;
812 goto err_unreg;
813 }
814
815 /* detect transition latency */
816 policy->cpuinfo.transition_latency = 0;
817 for (i = 0; i < perf->state_count; i++) {
818 if ((perf->states[i].transition_latency * 1000) >
819 policy->cpuinfo.transition_latency)
820 policy->cpuinfo.transition_latency =
821 perf->states[i].transition_latency * 1000;
822 }
823
824 /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
825 if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
826 policy->cpuinfo.transition_latency > 20 * 1000) {
827 policy->cpuinfo.transition_latency = 20 * 1000;
828 pr_info_once("P-state transition latency capped at 20 uS\n");
829 }
830
831 /* table init */
832 for (i = 0; i < perf->state_count; i++) {
833 if (i > 0 && perf->states[i].core_frequency >=
834 freq_table[valid_states-1].frequency / 1000)
835 continue;
836
837 freq_table[valid_states].driver_data = i;
838 freq_table[valid_states].frequency =
839 perf->states[i].core_frequency * 1000;
840 valid_states++;
841 }
842 freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
843
844 max_boost_ratio = get_max_boost_ratio(cpu, &nominal_freq);
845 if (max_boost_ratio) {
846 unsigned int freq = nominal_freq;
847
848 /*
849 * The loop above sorts the freq_table entries in the
850 * descending order. If ACPI CPPC has not advertised
851 * the nominal frequency (this is possible in CPPC
852 * revisions prior to 3), then use the first entry in
853 * the pstate table as a proxy for nominal frequency.
854 */
855 if (!freq)
856 freq = freq_table[0].frequency;
857
858 policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
859 } else {
860 /*
861 * If the maximum "boost" frequency is unknown, ask the arch
862 * scale-invariance code to use the "nominal" performance for
863 * CPU utilization scaling so as to prevent the schedutil
864 * governor from selecting inadequate CPU frequencies.
865 */
866 arch_set_max_freq_ratio(true);
867 }
868
869 policy->freq_table = freq_table;
870 perf->state = 0;
871
872 switch (perf->control_register.space_id) {
873 case ACPI_ADR_SPACE_SYSTEM_IO:
874 /*
875 * The core will not set policy->cur, because
876 * cpufreq_driver->get is NULL, so we need to set it here.
877 * However, we have to guess it, because the current speed is
878 * unknown and not detectable via IO ports.
879 */
880 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
881 break;
882 case ACPI_ADR_SPACE_FIXED_HARDWARE:
883 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
884 break;
885 default:
886 break;
887 }
888
889 /* notify BIOS that we exist */
890 acpi_processor_notify_smm(THIS_MODULE);
891
892 pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
893 for (i = 0; i < perf->state_count; i++)
894 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n",
895 (i == perf->state ? '*' : ' '), i,
896 (u32) perf->states[i].core_frequency,
897 (u32) perf->states[i].power,
898 (u32) perf->states[i].transition_latency);
899
900 /*
901 * the first call to ->target() should result in us actually
902 * writing something to the appropriate registers.
903 */
904 data->resume = 1;
905
906 policy->fast_switch_possible = !acpi_pstate_strict &&
907 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
908
909 if (perf->states[0].core_frequency * 1000 != freq_table[0].frequency)
910 pr_warn(FW_WARN "P-state 0 is not max freq\n");
911
912 if (acpi_cpufreq_driver.set_boost) {
913 set_boost(policy, acpi_cpufreq_driver.boost_enabled);
914 policy->boost_enabled = acpi_cpufreq_driver.boost_enabled;
915 }
916
917 return result;
918
919err_unreg:
920 acpi_processor_unregister_performance(cpu);
921err_free_mask:
922 free_cpumask_var(data->freqdomain_cpus);
923err_free:
924 kfree(data);
925 policy->driver_data = NULL;
926
927 return result;
928}
929
930static void acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
931{
932 struct acpi_cpufreq_data *data = policy->driver_data;
933
934 pr_debug("%s\n", __func__);
935
936 cpufreq_boost_down_prep(policy->cpu);
937 policy->fast_switch_possible = false;
938 policy->driver_data = NULL;
939 acpi_processor_unregister_performance(data->acpi_perf_cpu);
940 free_cpumask_var(data->freqdomain_cpus);
941 kfree(policy->freq_table);
942 kfree(data);
943}
944
945static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
946{
947 struct acpi_cpufreq_data *data = policy->driver_data;
948
949 pr_debug("%s\n", __func__);
950
951 data->resume = 1;
952
953 return 0;
954}
955
956static struct freq_attr *acpi_cpufreq_attr[] = {
957 &cpufreq_freq_attr_scaling_available_freqs,
958 &freqdomain_cpus,
959#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
960 &cpb,
961#endif
962 NULL,
963};
964
965static struct cpufreq_driver acpi_cpufreq_driver = {
966 .verify = cpufreq_generic_frequency_table_verify,
967 .target_index = acpi_cpufreq_target,
968 .fast_switch = acpi_cpufreq_fast_switch,
969 .bios_limit = acpi_processor_get_bios_limit,
970 .init = acpi_cpufreq_cpu_init,
971 .exit = acpi_cpufreq_cpu_exit,
972 .resume = acpi_cpufreq_resume,
973 .name = "acpi-cpufreq",
974 .attr = acpi_cpufreq_attr,
975};
976
977static void __init acpi_cpufreq_boost_init(void)
978{
979 if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
980 pr_debug("Boost capabilities not present in the processor\n");
981 return;
982 }
983
984 acpi_cpufreq_driver.set_boost = set_boost;
985 acpi_cpufreq_driver.boost_enabled = boost_state(0);
986}
987
988static int __init acpi_cpufreq_probe(struct platform_device *pdev)
989{
990 int ret;
991
992 if (acpi_disabled)
993 return -ENODEV;
994
995 /* don't keep reloading if cpufreq_driver exists */
996 if (cpufreq_get_current_driver())
997 return -ENODEV;
998
999 pr_debug("%s\n", __func__);
1000
1001 ret = acpi_cpufreq_early_init();
1002 if (ret)
1003 return ret;
1004
1005#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
1006 /* this is a sysfs file with a strange name and an even stranger
1007 * semantic - per CPU instantiation, but system global effect.
1008 * Lets enable it only on AMD CPUs for compatibility reasons and
1009 * only if configured. This is considered legacy code, which
1010 * will probably be removed at some point in the future.
1011 */
1012 if (!check_amd_hwpstate_cpu(0)) {
1013 struct freq_attr **attr;
1014
1015 pr_debug("CPB unsupported, do not expose it\n");
1016
1017 for (attr = acpi_cpufreq_attr; *attr; attr++)
1018 if (*attr == &cpb) {
1019 *attr = NULL;
1020 break;
1021 }
1022 }
1023#endif
1024 acpi_cpufreq_boost_init();
1025
1026 ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1027 if (ret) {
1028 free_acpi_perf_data();
1029 }
1030 return ret;
1031}
1032
1033static void acpi_cpufreq_remove(struct platform_device *pdev)
1034{
1035 pr_debug("%s\n", __func__);
1036
1037 cpufreq_unregister_driver(&acpi_cpufreq_driver);
1038
1039 free_acpi_perf_data();
1040}
1041
1042static struct platform_driver acpi_cpufreq_platdrv = {
1043 .driver = {
1044 .name = "acpi-cpufreq",
1045 },
1046 .remove = acpi_cpufreq_remove,
1047};
1048
1049static int __init acpi_cpufreq_init(void)
1050{
1051 return platform_driver_probe(&acpi_cpufreq_platdrv, acpi_cpufreq_probe);
1052}
1053
1054static void __exit acpi_cpufreq_exit(void)
1055{
1056 platform_driver_unregister(&acpi_cpufreq_platdrv);
1057}
1058
1059module_param(acpi_pstate_strict, uint, 0644);
1060MODULE_PARM_DESC(acpi_pstate_strict,
1061 "value 0 or non-zero. non-zero -> strict ACPI checks are "
1062 "performed during frequency changes.");
1063
1064late_initcall(acpi_cpufreq_init);
1065module_exit(acpi_cpufreq_exit);
1066
1067MODULE_ALIAS("platform:acpi-cpufreq");