Loading...
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
53#define MSR_K7_HWCR_CPB_DIS (1ULL << 25)
54
55struct acpi_cpufreq_data {
56 unsigned int resume;
57 unsigned int cpu_feature;
58 unsigned int acpi_perf_cpu;
59 cpumask_var_t freqdomain_cpus;
60 void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
61 u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
62};
63
64/* acpi_perf_data is a pointer to percpu data. */
65static struct acpi_processor_performance __percpu *acpi_perf_data;
66
67static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
68{
69 return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
70}
71
72static struct cpufreq_driver acpi_cpufreq_driver;
73
74static unsigned int acpi_pstate_strict;
75
76static bool boost_state(unsigned int cpu)
77{
78 u32 lo, hi;
79 u64 msr;
80
81 switch (boot_cpu_data.x86_vendor) {
82 case X86_VENDOR_INTEL:
83 case X86_VENDOR_CENTAUR:
84 case X86_VENDOR_ZHAOXIN:
85 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
86 msr = lo | ((u64)hi << 32);
87 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
88 case X86_VENDOR_HYGON:
89 case X86_VENDOR_AMD:
90 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
91 msr = lo | ((u64)hi << 32);
92 return !(msr & MSR_K7_HWCR_CPB_DIS);
93 }
94 return false;
95}
96
97static int boost_set_msr(bool enable)
98{
99 u32 msr_addr;
100 u64 msr_mask, val;
101
102 switch (boot_cpu_data.x86_vendor) {
103 case X86_VENDOR_INTEL:
104 case X86_VENDOR_CENTAUR:
105 case X86_VENDOR_ZHAOXIN:
106 msr_addr = MSR_IA32_MISC_ENABLE;
107 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
108 break;
109 case X86_VENDOR_HYGON:
110 case X86_VENDOR_AMD:
111 msr_addr = MSR_K7_HWCR;
112 msr_mask = MSR_K7_HWCR_CPB_DIS;
113 break;
114 default:
115 return -EINVAL;
116 }
117
118 rdmsrl(msr_addr, val);
119
120 if (enable)
121 val &= ~msr_mask;
122 else
123 val |= msr_mask;
124
125 wrmsrl(msr_addr, val);
126 return 0;
127}
128
129static void boost_set_msr_each(void *p_en)
130{
131 bool enable = (bool) p_en;
132
133 boost_set_msr(enable);
134}
135
136static int set_boost(struct cpufreq_policy *policy, int val)
137{
138 on_each_cpu_mask(policy->cpus, boost_set_msr_each,
139 (void *)(long)val, 1);
140 pr_debug("CPU %*pbl: Core Boosting %s.\n",
141 cpumask_pr_args(policy->cpus), str_enabled_disabled(val));
142
143 return 0;
144}
145
146static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
147{
148 struct acpi_cpufreq_data *data = policy->driver_data;
149
150 if (unlikely(!data))
151 return -ENODEV;
152
153 return cpufreq_show_cpus(data->freqdomain_cpus, buf);
154}
155
156cpufreq_freq_attr_ro(freqdomain_cpus);
157
158#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
159static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
160 size_t count)
161{
162 int ret;
163 unsigned int val = 0;
164
165 if (!acpi_cpufreq_driver.set_boost)
166 return -EINVAL;
167
168 ret = kstrtouint(buf, 10, &val);
169 if (ret || val > 1)
170 return -EINVAL;
171
172 cpus_read_lock();
173 set_boost(policy, val);
174 cpus_read_unlock();
175
176 return count;
177}
178
179static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
180{
181 return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
182}
183
184cpufreq_freq_attr_rw(cpb);
185#endif
186
187static int check_est_cpu(unsigned int cpuid)
188{
189 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
190
191 return cpu_has(cpu, X86_FEATURE_EST);
192}
193
194static int check_amd_hwpstate_cpu(unsigned int cpuid)
195{
196 struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
197
198 return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
199}
200
201static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
202{
203 struct acpi_cpufreq_data *data = policy->driver_data;
204 struct acpi_processor_performance *perf;
205 int i;
206
207 perf = to_perf_data(data);
208
209 for (i = 0; i < perf->state_count; i++) {
210 if (value == perf->states[i].status)
211 return policy->freq_table[i].frequency;
212 }
213 return 0;
214}
215
216static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
217{
218 struct acpi_cpufreq_data *data = policy->driver_data;
219 struct cpufreq_frequency_table *pos;
220 struct acpi_processor_performance *perf;
221
222 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
223 msr &= AMD_MSR_RANGE;
224 else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
225 msr &= HYGON_MSR_RANGE;
226 else
227 msr &= INTEL_MSR_RANGE;
228
229 perf = to_perf_data(data);
230
231 cpufreq_for_each_entry(pos, policy->freq_table)
232 if (msr == perf->states[pos->driver_data].status)
233 return pos->frequency;
234 return policy->freq_table[0].frequency;
235}
236
237static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
238{
239 struct acpi_cpufreq_data *data = policy->driver_data;
240
241 switch (data->cpu_feature) {
242 case SYSTEM_INTEL_MSR_CAPABLE:
243 case SYSTEM_AMD_MSR_CAPABLE:
244 return extract_msr(policy, val);
245 case SYSTEM_IO_CAPABLE:
246 return extract_io(policy, val);
247 default:
248 return 0;
249 }
250}
251
252static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
253{
254 u32 val, dummy __always_unused;
255
256 rdmsr(MSR_IA32_PERF_CTL, val, dummy);
257 return val;
258}
259
260static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
261{
262 u32 lo, hi;
263
264 rdmsr(MSR_IA32_PERF_CTL, lo, hi);
265 lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
266 wrmsr(MSR_IA32_PERF_CTL, lo, hi);
267}
268
269static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
270{
271 u32 val, dummy __always_unused;
272
273 rdmsr(MSR_AMD_PERF_CTL, val, dummy);
274 return val;
275}
276
277static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
278{
279 wrmsr(MSR_AMD_PERF_CTL, val, 0);
280}
281
282static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
283{
284 u32 val;
285
286 acpi_os_read_port(reg->address, &val, reg->bit_width);
287 return val;
288}
289
290static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
291{
292 acpi_os_write_port(reg->address, val, reg->bit_width);
293}
294
295struct drv_cmd {
296 struct acpi_pct_register *reg;
297 u32 val;
298 union {
299 void (*write)(struct acpi_pct_register *reg, u32 val);
300 u32 (*read)(struct acpi_pct_register *reg);
301 } func;
302};
303
304/* Called via smp_call_function_single(), on the target CPU */
305static void do_drv_read(void *_cmd)
306{
307 struct drv_cmd *cmd = _cmd;
308
309 cmd->val = cmd->func.read(cmd->reg);
310}
311
312static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
313{
314 struct acpi_processor_performance *perf = to_perf_data(data);
315 struct drv_cmd cmd = {
316 .reg = &perf->control_register,
317 .func.read = data->cpu_freq_read,
318 };
319 int err;
320
321 err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
322 WARN_ON_ONCE(err); /* smp_call_function_any() was buggy? */
323 return cmd.val;
324}
325
326/* Called via smp_call_function_many(), on the target CPUs */
327static void do_drv_write(void *_cmd)
328{
329 struct drv_cmd *cmd = _cmd;
330
331 cmd->func.write(cmd->reg, cmd->val);
332}
333
334static void drv_write(struct acpi_cpufreq_data *data,
335 const struct cpumask *mask, u32 val)
336{
337 struct acpi_processor_performance *perf = to_perf_data(data);
338 struct drv_cmd cmd = {
339 .reg = &perf->control_register,
340 .val = val,
341 .func.write = data->cpu_freq_write,
342 };
343 int this_cpu;
344
345 this_cpu = get_cpu();
346 if (cpumask_test_cpu(this_cpu, mask))
347 do_drv_write(&cmd);
348
349 smp_call_function_many(mask, do_drv_write, &cmd, 1);
350 put_cpu();
351}
352
353static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
354{
355 u32 val;
356
357 if (unlikely(cpumask_empty(mask)))
358 return 0;
359
360 val = drv_read(data, mask);
361
362 pr_debug("%s = %u\n", __func__, val);
363
364 return val;
365}
366
367static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
368{
369 struct acpi_cpufreq_data *data;
370 struct cpufreq_policy *policy;
371 unsigned int freq;
372 unsigned int cached_freq;
373
374 pr_debug("%s (%d)\n", __func__, cpu);
375
376 policy = cpufreq_cpu_get_raw(cpu);
377 if (unlikely(!policy))
378 return 0;
379
380 data = policy->driver_data;
381 if (unlikely(!data || !policy->freq_table))
382 return 0;
383
384 cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
385 freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
386 if (freq != cached_freq) {
387 /*
388 * The dreaded BIOS frequency change behind our back.
389 * Force set the frequency on next target call.
390 */
391 data->resume = 1;
392 }
393
394 pr_debug("cur freq = %u\n", freq);
395
396 return freq;
397}
398
399static unsigned int check_freqs(struct cpufreq_policy *policy,
400 const struct cpumask *mask, unsigned int freq)
401{
402 struct acpi_cpufreq_data *data = policy->driver_data;
403 unsigned int cur_freq;
404 unsigned int i;
405
406 for (i = 0; i < 100; i++) {
407 cur_freq = extract_freq(policy, get_cur_val(mask, data));
408 if (cur_freq == freq)
409 return 1;
410 udelay(10);
411 }
412 return 0;
413}
414
415static int acpi_cpufreq_target(struct cpufreq_policy *policy,
416 unsigned int index)
417{
418 struct acpi_cpufreq_data *data = policy->driver_data;
419 struct acpi_processor_performance *perf;
420 const struct cpumask *mask;
421 unsigned int next_perf_state = 0; /* Index into perf table */
422 int result = 0;
423
424 if (unlikely(!data)) {
425 return -ENODEV;
426 }
427
428 perf = to_perf_data(data);
429 next_perf_state = policy->freq_table[index].driver_data;
430 if (perf->state == next_perf_state) {
431 if (unlikely(data->resume)) {
432 pr_debug("Called after resume, resetting to P%d\n",
433 next_perf_state);
434 data->resume = 0;
435 } else {
436 pr_debug("Already at target state (P%d)\n",
437 next_perf_state);
438 return 0;
439 }
440 }
441
442 /*
443 * The core won't allow CPUs to go away until the governor has been
444 * stopped, so we can rely on the stability of policy->cpus.
445 */
446 mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
447 cpumask_of(policy->cpu) : policy->cpus;
448
449 drv_write(data, mask, perf->states[next_perf_state].control);
450
451 if (acpi_pstate_strict) {
452 if (!check_freqs(policy, mask,
453 policy->freq_table[index].frequency)) {
454 pr_debug("%s (%d)\n", __func__, policy->cpu);
455 result = -EAGAIN;
456 }
457 }
458
459 if (!result)
460 perf->state = next_perf_state;
461
462 return result;
463}
464
465static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
466 unsigned int target_freq)
467{
468 struct acpi_cpufreq_data *data = policy->driver_data;
469 struct acpi_processor_performance *perf;
470 struct cpufreq_frequency_table *entry;
471 unsigned int next_perf_state, next_freq, index;
472
473 /*
474 * Find the closest frequency above target_freq.
475 */
476 if (policy->cached_target_freq == target_freq)
477 index = policy->cached_resolved_idx;
478 else
479 index = cpufreq_table_find_index_dl(policy, target_freq,
480 false);
481
482 entry = &policy->freq_table[index];
483 next_freq = entry->frequency;
484 next_perf_state = entry->driver_data;
485
486 perf = to_perf_data(data);
487 if (perf->state == next_perf_state) {
488 if (unlikely(data->resume))
489 data->resume = 0;
490 else
491 return next_freq;
492 }
493
494 data->cpu_freq_write(&perf->control_register,
495 perf->states[next_perf_state].control);
496 perf->state = next_perf_state;
497 return next_freq;
498}
499
500static unsigned long
501acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
502{
503 struct acpi_processor_performance *perf;
504
505 perf = to_perf_data(data);
506 if (cpu_khz) {
507 /* search the closest match to cpu_khz */
508 unsigned int i;
509 unsigned long freq;
510 unsigned long freqn = perf->states[0].core_frequency * 1000;
511
512 for (i = 0; i < (perf->state_count-1); i++) {
513 freq = freqn;
514 freqn = perf->states[i+1].core_frequency * 1000;
515 if ((2 * cpu_khz) > (freqn + freq)) {
516 perf->state = i;
517 return freq;
518 }
519 }
520 perf->state = perf->state_count-1;
521 return freqn;
522 } else {
523 /* assume CPU is at P0... */
524 perf->state = 0;
525 return perf->states[0].core_frequency * 1000;
526 }
527}
528
529static void free_acpi_perf_data(void)
530{
531 unsigned int i;
532
533 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
534 for_each_possible_cpu(i)
535 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
536 ->shared_cpu_map);
537 free_percpu(acpi_perf_data);
538}
539
540static int cpufreq_boost_down_prep(unsigned int cpu)
541{
542 /*
543 * Clear the boost-disable bit on the CPU_DOWN path so that
544 * this cpu cannot block the remaining ones from boosting.
545 */
546 return boost_set_msr(1);
547}
548
549/*
550 * acpi_cpufreq_early_init - initialize ACPI P-States library
551 *
552 * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
553 * in order to determine correct frequency and voltage pairings. We can
554 * do _PDC and _PSD and find out the processor dependency for the
555 * actual init that will happen later...
556 */
557static int __init acpi_cpufreq_early_init(void)
558{
559 unsigned int i;
560 pr_debug("%s\n", __func__);
561
562 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
563 if (!acpi_perf_data) {
564 pr_debug("Memory allocation error for acpi_perf_data.\n");
565 return -ENOMEM;
566 }
567 for_each_possible_cpu(i) {
568 if (!zalloc_cpumask_var_node(
569 &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
570 GFP_KERNEL, cpu_to_node(i))) {
571
572 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
573 free_acpi_perf_data();
574 return -ENOMEM;
575 }
576 }
577
578 /* Do initialization in ACPI core */
579 acpi_processor_preregister_performance(acpi_perf_data);
580 return 0;
581}
582
583#ifdef CONFIG_SMP
584/*
585 * Some BIOSes do SW_ANY coordination internally, either set it up in hw
586 * or do it in BIOS firmware and won't inform about it to OS. If not
587 * detected, this has a side effect of making CPU run at a different speed
588 * than OS intended it to run at. Detect it and handle it cleanly.
589 */
590static int bios_with_sw_any_bug;
591
592static int sw_any_bug_found(const struct dmi_system_id *d)
593{
594 bios_with_sw_any_bug = 1;
595 return 0;
596}
597
598static const struct dmi_system_id sw_any_bug_dmi_table[] = {
599 {
600 .callback = sw_any_bug_found,
601 .ident = "Supermicro Server X6DLP",
602 .matches = {
603 DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
604 DMI_MATCH(DMI_BIOS_VERSION, "080010"),
605 DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
606 },
607 },
608 { }
609};
610
611static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
612{
613 /* Intel Xeon Processor 7100 Series Specification Update
614 * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
615 * AL30: A Machine Check Exception (MCE) Occurring during an
616 * Enhanced Intel SpeedStep Technology Ratio Change May Cause
617 * Both Processor Cores to Lock Up. */
618 if (c->x86_vendor == X86_VENDOR_INTEL) {
619 if ((c->x86 == 15) &&
620 (c->x86_model == 6) &&
621 (c->x86_stepping == 8)) {
622 pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
623 return -ENODEV;
624 }
625 }
626 return 0;
627}
628#endif
629
630#ifdef CONFIG_ACPI_CPPC_LIB
631static u64 get_max_boost_ratio(unsigned int cpu)
632{
633 struct cppc_perf_caps perf_caps;
634 u64 highest_perf, nominal_perf;
635 int ret;
636
637 if (acpi_pstate_strict)
638 return 0;
639
640 ret = cppc_get_perf_caps(cpu, &perf_caps);
641 if (ret) {
642 pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
643 cpu, ret);
644 return 0;
645 }
646
647 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
648 highest_perf = amd_get_highest_perf();
649 else
650 highest_perf = perf_caps.highest_perf;
651
652 nominal_perf = perf_caps.nominal_perf;
653
654 if (!highest_perf || !nominal_perf) {
655 pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
656 return 0;
657 }
658
659 if (highest_perf < nominal_perf) {
660 pr_debug("CPU%d: nominal performance above highest\n", cpu);
661 return 0;
662 }
663
664 return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
665}
666#else
667static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
668#endif
669
670static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
671{
672 struct cpufreq_frequency_table *freq_table;
673 struct acpi_processor_performance *perf;
674 struct acpi_cpufreq_data *data;
675 unsigned int cpu = policy->cpu;
676 struct cpuinfo_x86 *c = &cpu_data(cpu);
677 unsigned int valid_states = 0;
678 unsigned int result = 0;
679 u64 max_boost_ratio;
680 unsigned int i;
681#ifdef CONFIG_SMP
682 static int blacklisted;
683#endif
684
685 pr_debug("%s\n", __func__);
686
687#ifdef CONFIG_SMP
688 if (blacklisted)
689 return blacklisted;
690 blacklisted = acpi_cpufreq_blacklist(c);
691 if (blacklisted)
692 return blacklisted;
693#endif
694
695 data = kzalloc(sizeof(*data), GFP_KERNEL);
696 if (!data)
697 return -ENOMEM;
698
699 if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
700 result = -ENOMEM;
701 goto err_free;
702 }
703
704 perf = per_cpu_ptr(acpi_perf_data, cpu);
705 data->acpi_perf_cpu = cpu;
706 policy->driver_data = data;
707
708 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
709 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
710
711 result = acpi_processor_register_performance(perf, cpu);
712 if (result)
713 goto err_free_mask;
714
715 policy->shared_type = perf->shared_type;
716
717 /*
718 * Will let policy->cpus know about dependency only when software
719 * coordination is required.
720 */
721 if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
722 policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
723 cpumask_copy(policy->cpus, perf->shared_cpu_map);
724 }
725 cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
726
727#ifdef CONFIG_SMP
728 dmi_check_system(sw_any_bug_dmi_table);
729 if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
730 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
731 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
732 }
733
734 if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
735 !acpi_pstate_strict) {
736 cpumask_clear(policy->cpus);
737 cpumask_set_cpu(cpu, policy->cpus);
738 cpumask_copy(data->freqdomain_cpus,
739 topology_sibling_cpumask(cpu));
740 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
741 pr_info_once("overriding BIOS provided _PSD data\n");
742 }
743#endif
744
745 /* capability check */
746 if (perf->state_count <= 1) {
747 pr_debug("No P-States\n");
748 result = -ENODEV;
749 goto err_unreg;
750 }
751
752 if (perf->control_register.space_id != perf->status_register.space_id) {
753 result = -ENODEV;
754 goto err_unreg;
755 }
756
757 switch (perf->control_register.space_id) {
758 case ACPI_ADR_SPACE_SYSTEM_IO:
759 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
760 boot_cpu_data.x86 == 0xf) {
761 pr_debug("AMD K8 systems must use native drivers.\n");
762 result = -ENODEV;
763 goto err_unreg;
764 }
765 pr_debug("SYSTEM IO addr space\n");
766 data->cpu_feature = SYSTEM_IO_CAPABLE;
767 data->cpu_freq_read = cpu_freq_read_io;
768 data->cpu_freq_write = cpu_freq_write_io;
769 break;
770 case ACPI_ADR_SPACE_FIXED_HARDWARE:
771 pr_debug("HARDWARE addr space\n");
772 if (check_est_cpu(cpu)) {
773 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
774 data->cpu_freq_read = cpu_freq_read_intel;
775 data->cpu_freq_write = cpu_freq_write_intel;
776 break;
777 }
778 if (check_amd_hwpstate_cpu(cpu)) {
779 data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
780 data->cpu_freq_read = cpu_freq_read_amd;
781 data->cpu_freq_write = cpu_freq_write_amd;
782 break;
783 }
784 result = -ENODEV;
785 goto err_unreg;
786 default:
787 pr_debug("Unknown addr space %d\n",
788 (u32) (perf->control_register.space_id));
789 result = -ENODEV;
790 goto err_unreg;
791 }
792
793 freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
794 GFP_KERNEL);
795 if (!freq_table) {
796 result = -ENOMEM;
797 goto err_unreg;
798 }
799
800 /* detect transition latency */
801 policy->cpuinfo.transition_latency = 0;
802 for (i = 0; i < perf->state_count; i++) {
803 if ((perf->states[i].transition_latency * 1000) >
804 policy->cpuinfo.transition_latency)
805 policy->cpuinfo.transition_latency =
806 perf->states[i].transition_latency * 1000;
807 }
808
809 /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
810 if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
811 policy->cpuinfo.transition_latency > 20 * 1000) {
812 policy->cpuinfo.transition_latency = 20 * 1000;
813 pr_info_once("P-state transition latency capped at 20 uS\n");
814 }
815
816 /* table init */
817 for (i = 0; i < perf->state_count; i++) {
818 if (i > 0 && perf->states[i].core_frequency >=
819 freq_table[valid_states-1].frequency / 1000)
820 continue;
821
822 freq_table[valid_states].driver_data = i;
823 freq_table[valid_states].frequency =
824 perf->states[i].core_frequency * 1000;
825 valid_states++;
826 }
827 freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
828
829 max_boost_ratio = get_max_boost_ratio(cpu);
830 if (max_boost_ratio) {
831 unsigned int freq = freq_table[0].frequency;
832
833 /*
834 * Because the loop above sorts the freq_table entries in the
835 * descending order, freq is the maximum frequency in the table.
836 * Assume that it corresponds to the CPPC nominal frequency and
837 * use it to set cpuinfo.max_freq.
838 */
839 policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
840 } else {
841 /*
842 * If the maximum "boost" frequency is unknown, ask the arch
843 * scale-invariance code to use the "nominal" performance for
844 * CPU utilization scaling so as to prevent the schedutil
845 * governor from selecting inadequate CPU frequencies.
846 */
847 arch_set_max_freq_ratio(true);
848 }
849
850 policy->freq_table = freq_table;
851 perf->state = 0;
852
853 switch (perf->control_register.space_id) {
854 case ACPI_ADR_SPACE_SYSTEM_IO:
855 /*
856 * The core will not set policy->cur, because
857 * cpufreq_driver->get is NULL, so we need to set it here.
858 * However, we have to guess it, because the current speed is
859 * unknown and not detectable via IO ports.
860 */
861 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
862 break;
863 case ACPI_ADR_SPACE_FIXED_HARDWARE:
864 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
865 break;
866 default:
867 break;
868 }
869
870 /* notify BIOS that we exist */
871 acpi_processor_notify_smm(THIS_MODULE);
872
873 pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
874 for (i = 0; i < perf->state_count; i++)
875 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n",
876 (i == perf->state ? '*' : ' '), i,
877 (u32) perf->states[i].core_frequency,
878 (u32) perf->states[i].power,
879 (u32) perf->states[i].transition_latency);
880
881 /*
882 * the first call to ->target() should result in us actually
883 * writing something to the appropriate registers.
884 */
885 data->resume = 1;
886
887 policy->fast_switch_possible = !acpi_pstate_strict &&
888 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
889
890 if (perf->states[0].core_frequency * 1000 != freq_table[0].frequency)
891 pr_warn(FW_WARN "P-state 0 is not max freq\n");
892
893 if (acpi_cpufreq_driver.set_boost)
894 set_boost(policy, acpi_cpufreq_driver.boost_enabled);
895
896 return result;
897
898err_unreg:
899 acpi_processor_unregister_performance(cpu);
900err_free_mask:
901 free_cpumask_var(data->freqdomain_cpus);
902err_free:
903 kfree(data);
904 policy->driver_data = NULL;
905
906 return result;
907}
908
909static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
910{
911 struct acpi_cpufreq_data *data = policy->driver_data;
912
913 pr_debug("%s\n", __func__);
914
915 cpufreq_boost_down_prep(policy->cpu);
916 policy->fast_switch_possible = false;
917 policy->driver_data = NULL;
918 acpi_processor_unregister_performance(data->acpi_perf_cpu);
919 free_cpumask_var(data->freqdomain_cpus);
920 kfree(policy->freq_table);
921 kfree(data);
922
923 return 0;
924}
925
926static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
927{
928 struct acpi_cpufreq_data *data = policy->driver_data;
929
930 pr_debug("%s\n", __func__);
931
932 data->resume = 1;
933
934 return 0;
935}
936
937static struct freq_attr *acpi_cpufreq_attr[] = {
938 &cpufreq_freq_attr_scaling_available_freqs,
939 &freqdomain_cpus,
940#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
941 &cpb,
942#endif
943 NULL,
944};
945
946static struct cpufreq_driver acpi_cpufreq_driver = {
947 .verify = cpufreq_generic_frequency_table_verify,
948 .target_index = acpi_cpufreq_target,
949 .fast_switch = acpi_cpufreq_fast_switch,
950 .bios_limit = acpi_processor_get_bios_limit,
951 .init = acpi_cpufreq_cpu_init,
952 .exit = acpi_cpufreq_cpu_exit,
953 .resume = acpi_cpufreq_resume,
954 .name = "acpi-cpufreq",
955 .attr = acpi_cpufreq_attr,
956};
957
958static void __init acpi_cpufreq_boost_init(void)
959{
960 if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
961 pr_debug("Boost capabilities not present in the processor\n");
962 return;
963 }
964
965 acpi_cpufreq_driver.set_boost = set_boost;
966 acpi_cpufreq_driver.boost_enabled = boost_state(0);
967}
968
969static int __init acpi_cpufreq_probe(struct platform_device *pdev)
970{
971 int ret;
972
973 if (acpi_disabled)
974 return -ENODEV;
975
976 /* don't keep reloading if cpufreq_driver exists */
977 if (cpufreq_get_current_driver())
978 return -ENODEV;
979
980 pr_debug("%s\n", __func__);
981
982 ret = acpi_cpufreq_early_init();
983 if (ret)
984 return ret;
985
986#ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
987 /* this is a sysfs file with a strange name and an even stranger
988 * semantic - per CPU instantiation, but system global effect.
989 * Lets enable it only on AMD CPUs for compatibility reasons and
990 * only if configured. This is considered legacy code, which
991 * will probably be removed at some point in the future.
992 */
993 if (!check_amd_hwpstate_cpu(0)) {
994 struct freq_attr **attr;
995
996 pr_debug("CPB unsupported, do not expose it\n");
997
998 for (attr = acpi_cpufreq_attr; *attr; attr++)
999 if (*attr == &cpb) {
1000 *attr = NULL;
1001 break;
1002 }
1003 }
1004#endif
1005 acpi_cpufreq_boost_init();
1006
1007 ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1008 if (ret) {
1009 free_acpi_perf_data();
1010 }
1011 return ret;
1012}
1013
1014static void acpi_cpufreq_remove(struct platform_device *pdev)
1015{
1016 pr_debug("%s\n", __func__);
1017
1018 cpufreq_unregister_driver(&acpi_cpufreq_driver);
1019
1020 free_acpi_perf_data();
1021}
1022
1023static struct platform_driver acpi_cpufreq_platdrv = {
1024 .driver = {
1025 .name = "acpi-cpufreq",
1026 },
1027 .remove_new = acpi_cpufreq_remove,
1028};
1029
1030static int __init acpi_cpufreq_init(void)
1031{
1032 return platform_driver_probe(&acpi_cpufreq_platdrv, acpi_cpufreq_probe);
1033}
1034
1035static void __exit acpi_cpufreq_exit(void)
1036{
1037 platform_driver_unregister(&acpi_cpufreq_platdrv);
1038}
1039
1040module_param(acpi_pstate_strict, uint, 0644);
1041MODULE_PARM_DESC(acpi_pstate_strict,
1042 "value 0 or non-zero. non-zero -> strict ACPI checks are "
1043 "performed during frequency changes.");
1044
1045late_initcall(acpi_cpufreq_init);
1046module_exit(acpi_cpufreq_exit);
1047
1048MODULE_ALIAS("platform:acpi-cpufreq");
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");