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
23#include <linux/acpi.h>
24#include <linux/io.h>
25#include <linux/delay.h>
26#include <linux/uaccess.h>
27
28#include <acpi/processor.h>
29#include <acpi/cppc_acpi.h>
30
31#include <asm/msr.h>
32#include <asm/processor.h>
33#include <asm/cpufeature.h>
34#include <asm/cpu_device_id.h>
35
36MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
37MODULE_DESCRIPTION("ACPI Processor P-States Driver");
38MODULE_LICENSE("GPL");
39
40enum {
41 UNDEFINED_CAPABLE = 0,
42 SYSTEM_INTEL_MSR_CAPABLE,
43 SYSTEM_AMD_MSR_CAPABLE,
44 SYSTEM_IO_CAPABLE,
45};
46
47#define INTEL_MSR_RANGE (0xffff)
48#define AMD_MSR_RANGE (0x7)
49#define HYGON_MSR_RANGE (0x7)
50
51#define MSR_K7_HWCR_CPB_DIS (1ULL << 25)
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 u32 lo, hi;
77 u64 msr;
78
79 switch (boot_cpu_data.x86_vendor) {
80 case X86_VENDOR_INTEL:
81 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
82 msr = lo | ((u64)hi << 32);
83 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
84 case X86_VENDOR_HYGON:
85 case X86_VENDOR_AMD:
86 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
87 msr = lo | ((u64)hi << 32);
88 return !(msr & MSR_K7_HWCR_CPB_DIS);
89 }
90 return false;
91}
92
93static int boost_set_msr(bool enable)
94{
95 u32 msr_addr;
96 u64 msr_mask, val;
97
98 switch (boot_cpu_data.x86_vendor) {
99 case X86_VENDOR_INTEL:
100 msr_addr = MSR_IA32_MISC_ENABLE;
101 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
102 break;
103 case X86_VENDOR_HYGON:
104 case X86_VENDOR_AMD:
105 msr_addr = MSR_K7_HWCR;
106 msr_mask = MSR_K7_HWCR_CPB_DIS;
107 break;
108 default:
109 return -EINVAL;
110 }
111
112 rdmsrl(msr_addr, val);
113
114 if (enable)
115 val &= ~msr_mask;
116 else
117 val |= msr_mask;
118
119 wrmsrl(msr_addr, val);
120 return 0;
121}
122
123static void boost_set_msr_each(void *p_en)
124{
125 bool enable = (bool) p_en;
126
127 boost_set_msr(enable);
128}
129
130static int set_boost(struct cpufreq_policy *policy, int val)
131{
132 on_each_cpu_mask(policy->cpus, boost_set_msr_each,
133 (void *)(long)val, 1);
134 pr_debug("CPU %*pbl: Core Boosting %sabled.\n",
135 cpumask_pr_args(policy->cpus), val ? "en" : "dis");
136
137 return 0;
138}
139
140static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
141{
142 struct acpi_cpufreq_data *data = policy->driver_data;
143
144 if (unlikely(!data))
145 return -ENODEV;
146
147 return cpufreq_show_cpus(data->freqdomain_cpus, buf);
148}
149
150cpufreq_freq_attr_ro(freqdomain_cpus);
151
152#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
153static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
154 size_t count)
155{
156 int ret;
157 unsigned int val = 0;
158
159 if (!acpi_cpufreq_driver.set_boost)
160 return -EINVAL;
161
162 ret = kstrtouint(buf, 10, &val);
163 if (ret || val > 1)
164 return -EINVAL;
165
166 get_online_cpus();
167 set_boost(policy, val);
168 put_online_cpus();
169
170 return count;
171}
172
173static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
174{
175 return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
176}
177
178cpufreq_freq_attr_rw(cpb);
179#endif
180
181static int check_est_cpu(unsigned int cpuid)
182{
183 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
184
185 return cpu_has(cpu, X86_FEATURE_EST);
186}
187
188static int check_amd_hwpstate_cpu(unsigned int cpuid)
189{
190 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
191
192 return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
193}
194
195static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
196{
197 struct acpi_cpufreq_data *data = policy->driver_data;
198 struct acpi_processor_performance *perf;
199 int i;
200
201 perf = to_perf_data(data);
202
203 for (i = 0; i < perf->state_count; i++) {
204 if (value == perf->states[i].status)
205 return policy->freq_table[i].frequency;
206 }
207 return 0;
208}
209
210static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
211{
212 struct acpi_cpufreq_data *data = policy->driver_data;
213 struct cpufreq_frequency_table *pos;
214 struct acpi_processor_performance *perf;
215
216 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
217 msr &= AMD_MSR_RANGE;
218 else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
219 msr &= HYGON_MSR_RANGE;
220 else
221 msr &= INTEL_MSR_RANGE;
222
223 perf = to_perf_data(data);
224
225 cpufreq_for_each_entry(pos, policy->freq_table)
226 if (msr == perf->states[pos->driver_data].status)
227 return pos->frequency;
228 return policy->freq_table[0].frequency;
229}
230
231static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
232{
233 struct acpi_cpufreq_data *data = policy->driver_data;
234
235 switch (data->cpu_feature) {
236 case SYSTEM_INTEL_MSR_CAPABLE:
237 case SYSTEM_AMD_MSR_CAPABLE:
238 return extract_msr(policy, val);
239 case SYSTEM_IO_CAPABLE:
240 return extract_io(policy, val);
241 default:
242 return 0;
243 }
244}
245
246static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
247{
248 u32 val, dummy __always_unused;
249
250 rdmsr(MSR_IA32_PERF_CTL, val, dummy);
251 return val;
252}
253
254static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
255{
256 u32 lo, hi;
257
258 rdmsr(MSR_IA32_PERF_CTL, lo, hi);
259 lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
260 wrmsr(MSR_IA32_PERF_CTL, lo, hi);
261}
262
263static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
264{
265 u32 val, dummy __always_unused;
266
267 rdmsr(MSR_AMD_PERF_CTL, val, dummy);
268 return val;
269}
270
271static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
272{
273 wrmsr(MSR_AMD_PERF_CTL, val, 0);
274}
275
276static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
277{
278 u32 val;
279
280 acpi_os_read_port(reg->address, &val, reg->bit_width);
281 return val;
282}
283
284static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
285{
286 acpi_os_write_port(reg->address, val, reg->bit_width);
287}
288
289struct drv_cmd {
290 struct acpi_pct_register *reg;
291 u32 val;
292 union {
293 void (*write)(struct acpi_pct_register *reg, u32 val);
294 u32 (*read)(struct acpi_pct_register *reg);
295 } func;
296};
297
298/* Called via smp_call_function_single(), on the target CPU */
299static void do_drv_read(void *_cmd)
300{
301 struct drv_cmd *cmd = _cmd;
302
303 cmd->val = cmd->func.read(cmd->reg);
304}
305
306static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
307{
308 struct acpi_processor_performance *perf = to_perf_data(data);
309 struct drv_cmd cmd = {
310 .reg = &perf->control_register,
311 .func.read = data->cpu_freq_read,
312 };
313 int err;
314
315 err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
316 WARN_ON_ONCE(err); /* smp_call_function_any() was buggy? */
317 return cmd.val;
318}
319
320/* Called via smp_call_function_many(), on the target CPUs */
321static void do_drv_write(void *_cmd)
322{
323 struct drv_cmd *cmd = _cmd;
324
325 cmd->func.write(cmd->reg, cmd->val);
326}
327
328static void drv_write(struct acpi_cpufreq_data *data,
329 const struct cpumask *mask, u32 val)
330{
331 struct acpi_processor_performance *perf = to_perf_data(data);
332 struct drv_cmd cmd = {
333 .reg = &perf->control_register,
334 .val = val,
335 .func.write = data->cpu_freq_write,
336 };
337 int this_cpu;
338
339 this_cpu = get_cpu();
340 if (cpumask_test_cpu(this_cpu, mask))
341 do_drv_write(&cmd);
342
343 smp_call_function_many(mask, do_drv_write, &cmd, 1);
344 put_cpu();
345}
346
347static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
348{
349 u32 val;
350
351 if (unlikely(cpumask_empty(mask)))
352 return 0;
353
354 val = drv_read(data, mask);
355
356 pr_debug("%s = %u\n", __func__, val);
357
358 return val;
359}
360
361static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
362{
363 struct acpi_cpufreq_data *data;
364 struct cpufreq_policy *policy;
365 unsigned int freq;
366 unsigned int cached_freq;
367
368 pr_debug("%s (%d)\n", __func__, cpu);
369
370 policy = cpufreq_cpu_get_raw(cpu);
371 if (unlikely(!policy))
372 return 0;
373
374 data = policy->driver_data;
375 if (unlikely(!data || !policy->freq_table))
376 return 0;
377
378 cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
379 freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
380 if (freq != cached_freq) {
381 /*
382 * The dreaded BIOS frequency change behind our back.
383 * Force set the frequency on next target call.
384 */
385 data->resume = 1;
386 }
387
388 pr_debug("cur freq = %u\n", freq);
389
390 return freq;
391}
392
393static unsigned int check_freqs(struct cpufreq_policy *policy,
394 const struct cpumask *mask, unsigned int freq)
395{
396 struct acpi_cpufreq_data *data = policy->driver_data;
397 unsigned int cur_freq;
398 unsigned int i;
399
400 for (i = 0; i < 100; i++) {
401 cur_freq = extract_freq(policy, get_cur_val(mask, data));
402 if (cur_freq == freq)
403 return 1;
404 udelay(10);
405 }
406 return 0;
407}
408
409static int acpi_cpufreq_target(struct cpufreq_policy *policy,
410 unsigned int index)
411{
412 struct acpi_cpufreq_data *data = policy->driver_data;
413 struct acpi_processor_performance *perf;
414 const struct cpumask *mask;
415 unsigned int next_perf_state = 0; /* Index into perf table */
416 int result = 0;
417
418 if (unlikely(!data)) {
419 return -ENODEV;
420 }
421
422 perf = to_perf_data(data);
423 next_perf_state = policy->freq_table[index].driver_data;
424 if (perf->state == next_perf_state) {
425 if (unlikely(data->resume)) {
426 pr_debug("Called after resume, resetting to P%d\n",
427 next_perf_state);
428 data->resume = 0;
429 } else {
430 pr_debug("Already at target state (P%d)\n",
431 next_perf_state);
432 return 0;
433 }
434 }
435
436 /*
437 * The core won't allow CPUs to go away until the governor has been
438 * stopped, so we can rely on the stability of policy->cpus.
439 */
440 mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
441 cpumask_of(policy->cpu) : policy->cpus;
442
443 drv_write(data, mask, perf->states[next_perf_state].control);
444
445 if (acpi_pstate_strict) {
446 if (!check_freqs(policy, mask,
447 policy->freq_table[index].frequency)) {
448 pr_debug("%s (%d)\n", __func__, policy->cpu);
449 result = -EAGAIN;
450 }
451 }
452
453 if (!result)
454 perf->state = next_perf_state;
455
456 return result;
457}
458
459static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
460 unsigned int target_freq)
461{
462 struct acpi_cpufreq_data *data = policy->driver_data;
463 struct acpi_processor_performance *perf;
464 struct cpufreq_frequency_table *entry;
465 unsigned int next_perf_state, next_freq, index;
466
467 /*
468 * Find the closest frequency above target_freq.
469 */
470 if (policy->cached_target_freq == target_freq)
471 index = policy->cached_resolved_idx;
472 else
473 index = cpufreq_table_find_index_dl(policy, target_freq);
474
475 entry = &policy->freq_table[index];
476 next_freq = entry->frequency;
477 next_perf_state = entry->driver_data;
478
479 perf = to_perf_data(data);
480 if (perf->state == next_perf_state) {
481 if (unlikely(data->resume))
482 data->resume = 0;
483 else
484 return next_freq;
485 }
486
487 data->cpu_freq_write(&perf->control_register,
488 perf->states[next_perf_state].control);
489 perf->state = next_perf_state;
490 return next_freq;
491}
492
493static unsigned long
494acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
495{
496 struct acpi_processor_performance *perf;
497
498 perf = to_perf_data(data);
499 if (cpu_khz) {
500 /* search the closest match to cpu_khz */
501 unsigned int i;
502 unsigned long freq;
503 unsigned long freqn = perf->states[0].core_frequency * 1000;
504
505 for (i = 0; i < (perf->state_count-1); i++) {
506 freq = freqn;
507 freqn = perf->states[i+1].core_frequency * 1000;
508 if ((2 * cpu_khz) > (freqn + freq)) {
509 perf->state = i;
510 return freq;
511 }
512 }
513 perf->state = perf->state_count-1;
514 return freqn;
515 } else {
516 /* assume CPU is at P0... */
517 perf->state = 0;
518 return perf->states[0].core_frequency * 1000;
519 }
520}
521
522static void free_acpi_perf_data(void)
523{
524 unsigned int i;
525
526 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
527 for_each_possible_cpu(i)
528 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
529 ->shared_cpu_map);
530 free_percpu(acpi_perf_data);
531}
532
533static int cpufreq_boost_online(unsigned int cpu)
534{
535 /*
536 * On the CPU_UP path we simply keep the boost-disable flag
537 * in sync with the current global state.
538 */
539 return boost_set_msr(acpi_cpufreq_driver.boost_enabled);
540}
541
542static int cpufreq_boost_down_prep(unsigned int cpu)
543{
544 /*
545 * Clear the boost-disable bit on the CPU_DOWN path so that
546 * this cpu cannot block the remaining ones from boosting.
547 */
548 return boost_set_msr(1);
549}
550
551/*
552 * acpi_cpufreq_early_init - initialize ACPI P-States library
553 *
554 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
555 * in order to determine correct frequency and voltage pairings. We can
556 * do _PDC and _PSD and find out the processor dependency for the
557 * actual init that will happen later...
558 */
559static int __init acpi_cpufreq_early_init(void)
560{
561 unsigned int i;
562 pr_debug("%s\n", __func__);
563
564 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
565 if (!acpi_perf_data) {
566 pr_debug("Memory allocation error for acpi_perf_data.\n");
567 return -ENOMEM;
568 }
569 for_each_possible_cpu(i) {
570 if (!zalloc_cpumask_var_node(
571 &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
572 GFP_KERNEL, cpu_to_node(i))) {
573
574 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
575 free_acpi_perf_data();
576 return -ENOMEM;
577 }
578 }
579
580 /* Do initialization in ACPI core */
581 acpi_processor_preregister_performance(acpi_perf_data);
582 return 0;
583}
584
585#ifdef CONFIG_SMP
586/*
587 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
588 * or do it in BIOS firmware and won't inform about it to OS. If not
589 * detected, this has a side effect of making CPU run at a different speed
590 * than OS intended it to run at. Detect it and handle it cleanly.
591 */
592static int bios_with_sw_any_bug;
593
594static int sw_any_bug_found(const struct dmi_system_id *d)
595{
596 bios_with_sw_any_bug = 1;
597 return 0;
598}
599
600static const struct dmi_system_id sw_any_bug_dmi_table[] = {
601 {
602 .callback = sw_any_bug_found,
603 .ident = "Supermicro Server X6DLP",
604 .matches = {
605 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
606 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
607 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
608 },
609 },
610 { }
611};
612
613static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
614{
615 /* Intel Xeon Processor 7100 Series Specification Update
616 * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
617 * AL30: A Machine Check Exception (MCE) Occurring during an
618 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
619 * Both Processor Cores to Lock Up. */
620 if (c->x86_vendor == X86_VENDOR_INTEL) {
621 if ((c->x86 == 15) &&
622 (c->x86_model == 6) &&
623 (c->x86_stepping == 8)) {
624 pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
625 return -ENODEV;
626 }
627 }
628 return 0;
629}
630#endif
631
632#ifdef CONFIG_ACPI_CPPC_LIB
633static u64 get_max_boost_ratio(unsigned int cpu)
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 highest_perf = amd_get_highest_perf();
651 else
652 highest_perf = perf_caps.highest_perf;
653
654 nominal_perf = perf_caps.nominal_perf;
655
656 if (!highest_perf || !nominal_perf) {
657 pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
658 return 0;
659 }
660
661 if (highest_perf < nominal_perf) {
662 pr_debug("CPU%d: nominal performance above highest\n", cpu);
663 return 0;
664 }
665
666 return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
667}
668#else
669static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
670#endif
671
672static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
673{
674 struct cpufreq_frequency_table *freq_table;
675 struct acpi_processor_performance *perf;
676 struct acpi_cpufreq_data *data;
677 unsigned int cpu = policy->cpu;
678 struct cpuinfo_x86 *c = &cpu_data(cpu);
679 unsigned int valid_states = 0;
680 unsigned int result = 0;
681 u64 max_boost_ratio;
682 unsigned int i;
683#ifdef CONFIG_SMP
684 static int blacklisted;
685#endif
686
687 pr_debug("%s\n", __func__);
688
689#ifdef CONFIG_SMP
690 if (blacklisted)
691 return blacklisted;
692 blacklisted = acpi_cpufreq_blacklist(c);
693 if (blacklisted)
694 return blacklisted;
695#endif
696
697 data = kzalloc(sizeof(*data), GFP_KERNEL);
698 if (!data)
699 return -ENOMEM;
700
701 if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
702 result = -ENOMEM;
703 goto err_free;
704 }
705
706 perf = per_cpu_ptr(acpi_perf_data, cpu);
707 data->acpi_perf_cpu = cpu;
708 policy->driver_data = data;
709
710 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
711 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
712
713 result = acpi_processor_register_performance(perf, cpu);
714 if (result)
715 goto err_free_mask;
716
717 policy->shared_type = perf->shared_type;
718
719 /*
720 * Will let policy->cpus know about dependency only when software
721 * coordination is required.
722 */
723 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
724 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
725 cpumask_copy(policy->cpus, perf->shared_cpu_map);
726 }
727 cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
728
729#ifdef CONFIG_SMP
730 dmi_check_system(sw_any_bug_dmi_table);
731 if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
732 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
733 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
734 }
735
736 if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
737 !acpi_pstate_strict) {
738 cpumask_clear(policy->cpus);
739 cpumask_set_cpu(cpu, policy->cpus);
740 cpumask_copy(data->freqdomain_cpus,
741 topology_sibling_cpumask(cpu));
742 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
743 pr_info_once("overriding BIOS provided _PSD data\n");
744 }
745#endif
746
747 /* capability check */
748 if (perf->state_count <= 1) {
749 pr_debug("No P-States\n");
750 result = -ENODEV;
751 goto err_unreg;
752 }
753
754 if (perf->control_register.space_id != perf->status_register.space_id) {
755 result = -ENODEV;
756 goto err_unreg;
757 }
758
759 switch (perf->control_register.space_id) {
760 case ACPI_ADR_SPACE_SYSTEM_IO:
761 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
762 boot_cpu_data.x86 == 0xf) {
763 pr_debug("AMD K8 systems must use native drivers.\n");
764 result = -ENODEV;
765 goto err_unreg;
766 }
767 pr_debug("SYSTEM IO addr space\n");
768 data->cpu_feature = SYSTEM_IO_CAPABLE;
769 data->cpu_freq_read = cpu_freq_read_io;
770 data->cpu_freq_write = cpu_freq_write_io;
771 break;
772 case ACPI_ADR_SPACE_FIXED_HARDWARE:
773 pr_debug("HARDWARE addr space\n");
774 if (check_est_cpu(cpu)) {
775 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
776 data->cpu_freq_read = cpu_freq_read_intel;
777 data->cpu_freq_write = cpu_freq_write_intel;
778 break;
779 }
780 if (check_amd_hwpstate_cpu(cpu)) {
781 data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
782 data->cpu_freq_read = cpu_freq_read_amd;
783 data->cpu_freq_write = cpu_freq_write_amd;
784 break;
785 }
786 result = -ENODEV;
787 goto err_unreg;
788 default:
789 pr_debug("Unknown addr space %d\n",
790 (u32) (perf->control_register.space_id));
791 result = -ENODEV;
792 goto err_unreg;
793 }
794
795 freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
796 GFP_KERNEL);
797 if (!freq_table) {
798 result = -ENOMEM;
799 goto err_unreg;
800 }
801
802 /* detect transition latency */
803 policy->cpuinfo.transition_latency = 0;
804 for (i = 0; i < perf->state_count; i++) {
805 if ((perf->states[i].transition_latency * 1000) >
806 policy->cpuinfo.transition_latency)
807 policy->cpuinfo.transition_latency =
808 perf->states[i].transition_latency * 1000;
809 }
810
811 /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
812 if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
813 policy->cpuinfo.transition_latency > 20 * 1000) {
814 policy->cpuinfo.transition_latency = 20 * 1000;
815 pr_info_once("P-state transition latency capped at 20 uS\n");
816 }
817
818 /* table init */
819 for (i = 0; i < perf->state_count; i++) {
820 if (i > 0 && perf->states[i].core_frequency >=
821 freq_table[valid_states-1].frequency / 1000)
822 continue;
823
824 freq_table[valid_states].driver_data = i;
825 freq_table[valid_states].frequency =
826 perf->states[i].core_frequency * 1000;
827 valid_states++;
828 }
829 freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
830
831 max_boost_ratio = get_max_boost_ratio(cpu);
832 if (max_boost_ratio) {
833 unsigned int freq = freq_table[0].frequency;
834
835 /*
836 * Because the loop above sorts the freq_table entries in the
837 * descending order, freq is the maximum frequency in the table.
838 * Assume that it corresponds to the CPPC nominal frequency and
839 * use it to set cpuinfo.max_freq.
840 */
841 policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
842 } else {
843 /*
844 * If the maximum "boost" frequency is unknown, ask the arch
845 * scale-invariance code to use the "nominal" performance for
846 * CPU utilization scaling so as to prevent the schedutil
847 * governor from selecting inadequate CPU frequencies.
848 */
849 arch_set_max_freq_ratio(true);
850 }
851
852 policy->freq_table = freq_table;
853 perf->state = 0;
854
855 switch (perf->control_register.space_id) {
856 case ACPI_ADR_SPACE_SYSTEM_IO:
857 /*
858 * The core will not set policy->cur, because
859 * cpufreq_driver->get is NULL, so we need to set it here.
860 * However, we have to guess it, because the current speed is
861 * unknown and not detectable via IO ports.
862 */
863 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
864 break;
865 case ACPI_ADR_SPACE_FIXED_HARDWARE:
866 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
867 break;
868 default:
869 break;
870 }
871
872 /* notify BIOS that we exist */
873 acpi_processor_notify_smm(THIS_MODULE);
874
875 pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
876 for (i = 0; i < perf->state_count; i++)
877 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n",
878 (i == perf->state ? '*' : ' '), i,
879 (u32) perf->states[i].core_frequency,
880 (u32) perf->states[i].power,
881 (u32) perf->states[i].transition_latency);
882
883 /*
884 * the first call to ->target() should result in us actually
885 * writing something to the appropriate registers.
886 */
887 data->resume = 1;
888
889 policy->fast_switch_possible = !acpi_pstate_strict &&
890 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
891
892 return result;
893
894err_unreg:
895 acpi_processor_unregister_performance(cpu);
896err_free_mask:
897 free_cpumask_var(data->freqdomain_cpus);
898err_free:
899 kfree(data);
900 policy->driver_data = NULL;
901
902 return result;
903}
904
905static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
906{
907 struct acpi_cpufreq_data *data = policy->driver_data;
908
909 pr_debug("%s\n", __func__);
910
911 policy->fast_switch_possible = false;
912 policy->driver_data = NULL;
913 acpi_processor_unregister_performance(data->acpi_perf_cpu);
914 free_cpumask_var(data->freqdomain_cpus);
915 kfree(policy->freq_table);
916 kfree(data);
917
918 return 0;
919}
920
921static void acpi_cpufreq_cpu_ready(struct cpufreq_policy *policy)
922{
923 struct acpi_processor_performance *perf = per_cpu_ptr(acpi_perf_data,
924 policy->cpu);
925 unsigned int freq = policy->freq_table[0].frequency;
926
927 if (perf->states[0].core_frequency * 1000 != freq)
928 pr_warn(FW_WARN "P-state 0 is not max freq\n");
929}
930
931static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
932{
933 struct acpi_cpufreq_data *data = policy->driver_data;
934
935 pr_debug("%s\n", __func__);
936
937 data->resume = 1;
938
939 return 0;
940}
941
942static struct freq_attr *acpi_cpufreq_attr[] = {
943 &cpufreq_freq_attr_scaling_available_freqs,
944 &freqdomain_cpus,
945#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
946 &cpb,
947#endif
948 NULL,
949};
950
951static struct cpufreq_driver acpi_cpufreq_driver = {
952 .verify = cpufreq_generic_frequency_table_verify,
953 .target_index = acpi_cpufreq_target,
954 .fast_switch = acpi_cpufreq_fast_switch,
955 .bios_limit = acpi_processor_get_bios_limit,
956 .init = acpi_cpufreq_cpu_init,
957 .exit = acpi_cpufreq_cpu_exit,
958 .ready = acpi_cpufreq_cpu_ready,
959 .resume = acpi_cpufreq_resume,
960 .name = "acpi-cpufreq",
961 .attr = acpi_cpufreq_attr,
962};
963
964static enum cpuhp_state acpi_cpufreq_online;
965
966static void __init acpi_cpufreq_boost_init(void)
967{
968 int ret;
969
970 if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
971 pr_debug("Boost capabilities not present in the processor\n");
972 return;
973 }
974
975 acpi_cpufreq_driver.set_boost = set_boost;
976 acpi_cpufreq_driver.boost_enabled = boost_state(0);
977
978 /*
979 * This calls the online callback on all online cpu and forces all
980 * MSRs to the same value.
981 */
982 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "cpufreq/acpi:online",
983 cpufreq_boost_online, cpufreq_boost_down_prep);
984 if (ret < 0) {
985 pr_err("acpi_cpufreq: failed to register hotplug callbacks\n");
986 return;
987 }
988 acpi_cpufreq_online = ret;
989}
990
991static void acpi_cpufreq_boost_exit(void)
992{
993 if (acpi_cpufreq_online > 0)
994 cpuhp_remove_state_nocalls(acpi_cpufreq_online);
995}
996
997static int __init acpi_cpufreq_init(void)
998{
999 int ret;
1000
1001 if (acpi_disabled)
1002 return -ENODEV;
1003
1004 /* don't keep reloading if cpufreq_driver exists */
1005 if (cpufreq_get_current_driver())
1006 return -EEXIST;
1007
1008 pr_debug("%s\n", __func__);
1009
1010 ret = acpi_cpufreq_early_init();
1011 if (ret)
1012 return ret;
1013
1014#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
1015 /* this is a sysfs file with a strange name and an even stranger
1016 * semantic - per CPU instantiation, but system global effect.
1017 * Lets enable it only on AMD CPUs for compatibility reasons and
1018 * only if configured. This is considered legacy code, which
1019 * will probably be removed at some point in the future.
1020 */
1021 if (!check_amd_hwpstate_cpu(0)) {
1022 struct freq_attr **attr;
1023
1024 pr_debug("CPB unsupported, do not expose it\n");
1025
1026 for (attr = acpi_cpufreq_attr; *attr; attr++)
1027 if (*attr == &cpb) {
1028 *attr = NULL;
1029 break;
1030 }
1031 }
1032#endif
1033 acpi_cpufreq_boost_init();
1034
1035 ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1036 if (ret) {
1037 free_acpi_perf_data();
1038 acpi_cpufreq_boost_exit();
1039 }
1040 return ret;
1041}
1042
1043static void __exit acpi_cpufreq_exit(void)
1044{
1045 pr_debug("%s\n", __func__);
1046
1047 acpi_cpufreq_boost_exit();
1048
1049 cpufreq_unregister_driver(&acpi_cpufreq_driver);
1050
1051 free_acpi_perf_data();
1052}
1053
1054module_param(acpi_pstate_strict, uint, 0644);
1055MODULE_PARM_DESC(acpi_pstate_strict,
1056 "value 0 or non-zero. non-zero -> strict ACPI checks are "
1057 "performed during frequency changes.");
1058
1059late_initcall(acpi_cpufreq_init);
1060module_exit(acpi_cpufreq_exit);
1061
1062static const struct x86_cpu_id __maybe_unused acpi_cpufreq_ids[] = {
1063 X86_MATCH_FEATURE(X86_FEATURE_ACPI, NULL),
1064 X86_MATCH_FEATURE(X86_FEATURE_HW_PSTATE, NULL),
1065 {}
1066};
1067MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1068
1069static const struct acpi_device_id __maybe_unused processor_device_ids[] = {
1070 {ACPI_PROCESSOR_OBJECT_HID, },
1071 {ACPI_PROCESSOR_DEVICE_HID, },
1072 {},
1073};
1074MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1075
1076MODULE_ALIAS("acpi");