1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * processor_perflib.c - ACPI Processor P-States Library ($Revision: 71 $)
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) 2004 Dominik Brodowski <linux@brodo.de>
8 * Copyright (C) 2004 Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
9 * - Added processor hotplug support
12 #include <linux/kernel.h>
13 #include <linux/module.h>
14 #include <linux/init.h>
15 #include <linux/cpufreq.h>
16 #include <linux/slab.h>
17 #include <linux/acpi.h>
18 #include <acpi/processor.h>
20 #include <asm/cpufeature.h>
23 #define PREFIX "ACPI: "
25 #define ACPI_PROCESSOR_CLASS "processor"
26 #define ACPI_PROCESSOR_FILE_PERFORMANCE "performance"
27 #define _COMPONENT ACPI_PROCESSOR_COMPONENT
28 ACPI_MODULE_NAME("processor_perflib");
30 static DEFINE_MUTEX(performance_mutex);
33 * _PPC support is implemented as a CPUfreq policy notifier:
34 * This means each time a CPUfreq driver registered also with
35 * the ACPI core is asked to change the speed policy, the maximum
36 * value is adjusted so that it is within the platform limit.
38 * Also, when a new platform limit value is detected, the CPUfreq
39 * policy is adjusted accordingly.
43 * -1 -> cpufreq low level drivers not initialized -> _PSS, etc. not called yet
45 * 0 -> cpufreq low level drivers initialized -> consider _PPC values
46 * 1 -> ignore _PPC totally -> forced by user through boot param
48 static int ignore_ppc = -1;
49 module_param(ignore_ppc, int, 0644);
50 MODULE_PARM_DESC(ignore_ppc, "If the frequency of your machine gets wrongly" \
51 "limited by BIOS, this should help");
53 static bool acpi_processor_ppc_in_use;
55 static int acpi_processor_get_platform_limit(struct acpi_processor *pr)
57 acpi_status status = 0;
58 unsigned long long ppc = 0;
65 * _PPC indicates the maximum state currently supported by the platform
66 * (e.g. 0 = states 0..n; 1 = states 1..n; etc.
68 status = acpi_evaluate_integer(pr->handle, "_PPC", NULL, &ppc);
70 if (status != AE_NOT_FOUND)
71 acpi_processor_ppc_in_use = true;
73 if (ACPI_FAILURE(status) && status != AE_NOT_FOUND) {
74 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PPC"));
78 pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr->id,
79 (int)ppc, ppc ? "" : "not");
81 pr->performance_platform_limit = (int)ppc;
83 if (ppc >= pr->performance->state_count ||
84 unlikely(!freq_qos_request_active(&pr->perflib_req)))
87 ret = freq_qos_update_request(&pr->perflib_req,
88 pr->performance->states[ppc].core_frequency * 1000);
90 pr_warn("Failed to update perflib freq constraint: CPU%d (%d)\n",
97 #define ACPI_PROCESSOR_NOTIFY_PERFORMANCE 0x80
99 * acpi_processor_ppc_ost: Notify firmware the _PPC evaluation status
100 * @handle: ACPI processor handle
101 * @status: the status code of _PPC evaluation
102 * 0: success. OSPM is now using the performance state specificed.
103 * 1: failure. OSPM has not changed the number of P-states in use
105 static void acpi_processor_ppc_ost(acpi_handle handle, int status)
107 if (acpi_has_method(handle, "_OST"))
108 acpi_evaluate_ost(handle, ACPI_PROCESSOR_NOTIFY_PERFORMANCE,
112 void acpi_processor_ppc_has_changed(struct acpi_processor *pr, int event_flag)
116 if (ignore_ppc || !pr->performance) {
118 * Only when it is notification event, the _OST object
119 * will be evaluated. Otherwise it is skipped.
122 acpi_processor_ppc_ost(pr->handle, 1);
126 ret = acpi_processor_get_platform_limit(pr);
128 * Only when it is notification event, the _OST object
129 * will be evaluated. Otherwise it is skipped.
133 acpi_processor_ppc_ost(pr->handle, 1);
135 acpi_processor_ppc_ost(pr->handle, 0);
138 cpufreq_update_limits(pr->id);
141 int acpi_processor_get_bios_limit(int cpu, unsigned int *limit)
143 struct acpi_processor *pr;
145 pr = per_cpu(processors, cpu);
146 if (!pr || !pr->performance || !pr->performance->state_count)
148 *limit = pr->performance->states[pr->performance_platform_limit].
149 core_frequency * 1000;
152 EXPORT_SYMBOL(acpi_processor_get_bios_limit);
154 void acpi_processor_ignore_ppc_init(void)
160 void acpi_processor_ppc_init(struct cpufreq_policy *policy)
162 int cpu = policy->cpu;
163 struct acpi_processor *pr = per_cpu(processors, cpu);
169 ret = freq_qos_add_request(&policy->constraints, &pr->perflib_req,
170 FREQ_QOS_MAX, INT_MAX);
172 pr_err("Failed to add freq constraint for CPU%d (%d)\n", cpu,
176 void acpi_processor_ppc_exit(struct cpufreq_policy *policy)
178 struct acpi_processor *pr = per_cpu(processors, policy->cpu);
181 freq_qos_remove_request(&pr->perflib_req);
184 static int acpi_processor_get_performance_control(struct acpi_processor *pr)
187 acpi_status status = 0;
188 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
189 union acpi_object *pct = NULL;
190 union acpi_object obj = { 0 };
193 status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
194 if (ACPI_FAILURE(status)) {
195 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PCT"));
199 pct = (union acpi_object *)buffer.pointer;
200 if (!pct || (pct->type != ACPI_TYPE_PACKAGE)
201 || (pct->package.count != 2)) {
202 printk(KERN_ERR PREFIX "Invalid _PCT data\n");
211 obj = pct->package.elements[0];
213 if ((obj.type != ACPI_TYPE_BUFFER)
214 || (obj.buffer.length < sizeof(struct acpi_pct_register))
215 || (obj.buffer.pointer == NULL)) {
216 printk(KERN_ERR PREFIX "Invalid _PCT data (control_register)\n");
220 memcpy(&pr->performance->control_register, obj.buffer.pointer,
221 sizeof(struct acpi_pct_register));
227 obj = pct->package.elements[1];
229 if ((obj.type != ACPI_TYPE_BUFFER)
230 || (obj.buffer.length < sizeof(struct acpi_pct_register))
231 || (obj.buffer.pointer == NULL)) {
232 printk(KERN_ERR PREFIX "Invalid _PCT data (status_register)\n");
237 memcpy(&pr->performance->status_register, obj.buffer.pointer,
238 sizeof(struct acpi_pct_register));
241 kfree(buffer.pointer);
248 * Some AMDs have 50MHz frequency multiples, but only provide 100MHz rounding
249 * in their ACPI data. Calculate the real values and fix up the _PSS data.
251 static void amd_fixup_frequency(struct acpi_processor_px *px, int i)
253 u32 hi, lo, fid, did;
254 int index = px->control & 0x00000007;
256 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
259 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
260 || boot_cpu_data.x86 == 0x11) {
261 rdmsr(MSR_AMD_PSTATE_DEF_BASE + index, lo, hi);
264 * Bit 63: PstateEn. Read-write. If set, the P-state is valid.
271 if (boot_cpu_data.x86 == 0x10)
272 px->core_frequency = (100 * (fid + 0x10)) >> did;
274 px->core_frequency = (100 * (fid + 8)) >> did;
278 static void amd_fixup_frequency(struct acpi_processor_px *px, int i) {};
281 static int acpi_processor_get_performance_states(struct acpi_processor *pr)
284 acpi_status status = AE_OK;
285 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
286 struct acpi_buffer format = { sizeof("NNNNNN"), "NNNNNN" };
287 struct acpi_buffer state = { 0, NULL };
288 union acpi_object *pss = NULL;
290 int last_invalid = -1;
293 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
294 if (ACPI_FAILURE(status)) {
295 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PSS"));
299 pss = buffer.pointer;
300 if (!pss || (pss->type != ACPI_TYPE_PACKAGE)) {
301 printk(KERN_ERR PREFIX "Invalid _PSS data\n");
306 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d performance states\n",
307 pss->package.count));
309 pr->performance->state_count = pss->package.count;
310 pr->performance->states =
311 kmalloc_array(pss->package.count,
312 sizeof(struct acpi_processor_px),
314 if (!pr->performance->states) {
319 for (i = 0; i < pr->performance->state_count; i++) {
321 struct acpi_processor_px *px = &(pr->performance->states[i]);
323 state.length = sizeof(struct acpi_processor_px);
326 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Extracting state %d\n", i));
328 status = acpi_extract_package(&(pss->package.elements[i]),
330 if (ACPI_FAILURE(status)) {
331 ACPI_EXCEPTION((AE_INFO, status, "Invalid _PSS data"));
333 kfree(pr->performance->states);
337 amd_fixup_frequency(px, i);
339 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
340 "State [%d]: core_frequency[%d] power[%d] transition_latency[%d] bus_master_latency[%d] control[0x%x] status[0x%x]\n",
342 (u32) px->core_frequency,
344 (u32) px->transition_latency,
345 (u32) px->bus_master_latency,
346 (u32) px->control, (u32) px->status));
349 * Check that ACPI's u64 MHz will be valid as u32 KHz in cpufreq
351 if (!px->core_frequency ||
352 ((u32)(px->core_frequency * 1000) !=
353 (px->core_frequency * 1000))) {
354 printk(KERN_ERR FW_BUG PREFIX
355 "Invalid BIOS _PSS frequency found for processor %d: 0x%llx MHz\n",
356 pr->id, px->core_frequency);
357 if (last_invalid == -1)
360 if (last_invalid != -1) {
362 * Copy this valid entry over last_invalid entry
364 memcpy(&(pr->performance->states[last_invalid]),
365 px, sizeof(struct acpi_processor_px));
371 if (last_invalid == 0) {
372 printk(KERN_ERR FW_BUG PREFIX
373 "No valid BIOS _PSS frequency found for processor %d\n", pr->id);
375 kfree(pr->performance->states);
376 pr->performance->states = NULL;
379 if (last_invalid > 0)
380 pr->performance->state_count = last_invalid;
383 kfree(buffer.pointer);
388 int acpi_processor_get_performance_info(struct acpi_processor *pr)
392 if (!pr || !pr->performance || !pr->handle)
395 if (!acpi_has_method(pr->handle, "_PCT")) {
396 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
397 "ACPI-based processor performance control unavailable\n"));
401 result = acpi_processor_get_performance_control(pr);
405 result = acpi_processor_get_performance_states(pr);
409 /* We need to call _PPC once when cpufreq starts */
411 result = acpi_processor_get_platform_limit(pr);
416 * Having _PPC but missing frequencies (_PSS, _PCT) is a very good hint that
417 * the BIOS is older than the CPU and does not know its frequencies
421 if (acpi_has_method(pr->handle, "_PPC")) {
422 if(boot_cpu_has(X86_FEATURE_EST))
423 printk(KERN_WARNING FW_BUG "BIOS needs update for CPU "
424 "frequency support\n");
429 EXPORT_SYMBOL_GPL(acpi_processor_get_performance_info);
431 int acpi_processor_pstate_control(void)
435 if (!acpi_gbl_FADT.smi_command || !acpi_gbl_FADT.pstate_control)
438 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
439 "Writing pstate_control [0x%x] to smi_command [0x%x]\n",
440 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command));
442 status = acpi_os_write_port(acpi_gbl_FADT.smi_command,
443 (u32)acpi_gbl_FADT.pstate_control, 8);
444 if (ACPI_SUCCESS(status))
447 ACPI_EXCEPTION((AE_INFO, status,
448 "Failed to write pstate_control [0x%x] to smi_command [0x%x]",
449 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command));
453 int acpi_processor_notify_smm(struct module *calling_module)
455 static int is_done = 0;
458 if (!acpi_processor_cpufreq_init)
461 if (!try_module_get(calling_module))
464 /* is_done is set to negative if an error occurred,
465 * and to postitive if _no_ error occurred, but SMM
466 * was already notified. This avoids double notification
467 * which might lead to unexpected results...
470 module_put(calling_module);
472 } else if (is_done < 0) {
473 module_put(calling_module);
479 result = acpi_processor_pstate_control();
481 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No SMI port or pstate_control\n"));
482 module_put(calling_module);
486 module_put(calling_module);
490 /* Success. If there's no _PPC, we need to fear nothing, so
491 * we can allow the cpufreq driver to be rmmod'ed. */
494 if (!acpi_processor_ppc_in_use)
495 module_put(calling_module);
500 EXPORT_SYMBOL(acpi_processor_notify_smm);
502 int acpi_processor_get_psd(acpi_handle handle, struct acpi_psd_package *pdomain)
505 acpi_status status = AE_OK;
506 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
507 struct acpi_buffer format = {sizeof("NNNNN"), "NNNNN"};
508 struct acpi_buffer state = {0, NULL};
509 union acpi_object *psd = NULL;
511 status = acpi_evaluate_object(handle, "_PSD", NULL, &buffer);
512 if (ACPI_FAILURE(status)) {
516 psd = buffer.pointer;
517 if (!psd || (psd->type != ACPI_TYPE_PACKAGE)) {
518 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
523 if (psd->package.count != 1) {
524 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
529 state.length = sizeof(struct acpi_psd_package);
530 state.pointer = pdomain;
532 status = acpi_extract_package(&(psd->package.elements[0]),
534 if (ACPI_FAILURE(status)) {
535 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
540 if (pdomain->num_entries != ACPI_PSD_REV0_ENTRIES) {
541 printk(KERN_ERR PREFIX "Unknown _PSD:num_entries\n");
546 if (pdomain->revision != ACPI_PSD_REV0_REVISION) {
547 printk(KERN_ERR PREFIX "Unknown _PSD:revision\n");
552 if (pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ALL &&
553 pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ANY &&
554 pdomain->coord_type != DOMAIN_COORD_TYPE_HW_ALL) {
555 printk(KERN_ERR PREFIX "Invalid _PSD:coord_type\n");
560 kfree(buffer.pointer);
563 EXPORT_SYMBOL(acpi_processor_get_psd);
565 int acpi_processor_preregister_performance(
566 struct acpi_processor_performance __percpu *performance)
571 cpumask_var_t covered_cpus;
572 struct acpi_processor *pr;
573 struct acpi_psd_package *pdomain;
574 struct acpi_processor *match_pr;
575 struct acpi_psd_package *match_pdomain;
577 if (!zalloc_cpumask_var(&covered_cpus, GFP_KERNEL))
580 mutex_lock(&performance_mutex);
583 * Check if another driver has already registered, and abort before
584 * changing pr->performance if it has. Check input data as well.
586 for_each_possible_cpu(i) {
587 pr = per_cpu(processors, i);
589 /* Look only at processors in ACPI namespace */
593 if (pr->performance) {
598 if (!performance || !per_cpu_ptr(performance, i)) {
604 /* Call _PSD for all CPUs */
605 for_each_possible_cpu(i) {
606 pr = per_cpu(processors, i);
610 pr->performance = per_cpu_ptr(performance, i);
611 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
612 pdomain = &(pr->performance->domain_info);
613 if (acpi_processor_get_psd(pr->handle, pdomain)) {
622 * Now that we have _PSD data from all CPUs, lets setup P-state
625 for_each_possible_cpu(i) {
626 pr = per_cpu(processors, i);
630 if (cpumask_test_cpu(i, covered_cpus))
633 pdomain = &(pr->performance->domain_info);
634 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
635 cpumask_set_cpu(i, covered_cpus);
636 if (pdomain->num_processors <= 1)
639 /* Validate the Domain info */
640 count_target = pdomain->num_processors;
641 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
642 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
643 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
644 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_HW;
645 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
646 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ANY;
648 for_each_possible_cpu(j) {
652 match_pr = per_cpu(processors, j);
656 match_pdomain = &(match_pr->performance->domain_info);
657 if (match_pdomain->domain != pdomain->domain)
660 /* Here i and j are in the same domain */
662 if (match_pdomain->num_processors != count_target) {
667 if (pdomain->coord_type != match_pdomain->coord_type) {
672 cpumask_set_cpu(j, covered_cpus);
673 cpumask_set_cpu(j, pr->performance->shared_cpu_map);
676 for_each_possible_cpu(j) {
680 match_pr = per_cpu(processors, j);
684 match_pdomain = &(match_pr->performance->domain_info);
685 if (match_pdomain->domain != pdomain->domain)
688 match_pr->performance->shared_type =
689 pr->performance->shared_type;
690 cpumask_copy(match_pr->performance->shared_cpu_map,
691 pr->performance->shared_cpu_map);
696 for_each_possible_cpu(i) {
697 pr = per_cpu(processors, i);
698 if (!pr || !pr->performance)
701 /* Assume no coordination on any error parsing domain info */
703 cpumask_clear(pr->performance->shared_cpu_map);
704 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
705 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
707 pr->performance = NULL; /* Will be set for real in register */
711 mutex_unlock(&performance_mutex);
712 free_cpumask_var(covered_cpus);
715 EXPORT_SYMBOL(acpi_processor_preregister_performance);
718 acpi_processor_register_performance(struct acpi_processor_performance
719 *performance, unsigned int cpu)
721 struct acpi_processor *pr;
723 if (!acpi_processor_cpufreq_init)
726 mutex_lock(&performance_mutex);
728 pr = per_cpu(processors, cpu);
730 mutex_unlock(&performance_mutex);
734 if (pr->performance) {
735 mutex_unlock(&performance_mutex);
739 WARN_ON(!performance);
741 pr->performance = performance;
743 if (acpi_processor_get_performance_info(pr)) {
744 pr->performance = NULL;
745 mutex_unlock(&performance_mutex);
749 mutex_unlock(&performance_mutex);
753 EXPORT_SYMBOL(acpi_processor_register_performance);
755 void acpi_processor_unregister_performance(unsigned int cpu)
757 struct acpi_processor *pr;
759 mutex_lock(&performance_mutex);
761 pr = per_cpu(processors, cpu);
763 mutex_unlock(&performance_mutex);
768 kfree(pr->performance->states);
769 pr->performance = NULL;
771 mutex_unlock(&performance_mutex);
776 EXPORT_SYMBOL(acpi_processor_unregister_performance);