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)
164 for_each_cpu(cpu, policy->related_cpus) {
165 struct acpi_processor *pr = per_cpu(processors, cpu);
171 ret = freq_qos_add_request(&policy->constraints,
173 FREQ_QOS_MAX, INT_MAX);
175 pr_err("Failed to add freq constraint for CPU%d (%d)\n",
180 void acpi_processor_ppc_exit(struct cpufreq_policy *policy)
184 for_each_cpu(cpu, policy->related_cpus) {
185 struct acpi_processor *pr = per_cpu(processors, cpu);
188 freq_qos_remove_request(&pr->perflib_req);
192 static int acpi_processor_get_performance_control(struct acpi_processor *pr)
195 acpi_status status = 0;
196 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
197 union acpi_object *pct = NULL;
198 union acpi_object obj = { 0 };
201 status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
202 if (ACPI_FAILURE(status)) {
203 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PCT"));
207 pct = (union acpi_object *)buffer.pointer;
208 if (!pct || (pct->type != ACPI_TYPE_PACKAGE)
209 || (pct->package.count != 2)) {
210 printk(KERN_ERR PREFIX "Invalid _PCT data\n");
219 obj = pct->package.elements[0];
221 if ((obj.type != ACPI_TYPE_BUFFER)
222 || (obj.buffer.length < sizeof(struct acpi_pct_register))
223 || (obj.buffer.pointer == NULL)) {
224 printk(KERN_ERR PREFIX "Invalid _PCT data (control_register)\n");
228 memcpy(&pr->performance->control_register, obj.buffer.pointer,
229 sizeof(struct acpi_pct_register));
235 obj = pct->package.elements[1];
237 if ((obj.type != ACPI_TYPE_BUFFER)
238 || (obj.buffer.length < sizeof(struct acpi_pct_register))
239 || (obj.buffer.pointer == NULL)) {
240 printk(KERN_ERR PREFIX "Invalid _PCT data (status_register)\n");
245 memcpy(&pr->performance->status_register, obj.buffer.pointer,
246 sizeof(struct acpi_pct_register));
249 kfree(buffer.pointer);
256 * Some AMDs have 50MHz frequency multiples, but only provide 100MHz rounding
257 * in their ACPI data. Calculate the real values and fix up the _PSS data.
259 static void amd_fixup_frequency(struct acpi_processor_px *px, int i)
261 u32 hi, lo, fid, did;
262 int index = px->control & 0x00000007;
264 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
267 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
268 || boot_cpu_data.x86 == 0x11) {
269 rdmsr(MSR_AMD_PSTATE_DEF_BASE + index, lo, hi);
272 * Bit 63: PstateEn. Read-write. If set, the P-state is valid.
279 if (boot_cpu_data.x86 == 0x10)
280 px->core_frequency = (100 * (fid + 0x10)) >> did;
282 px->core_frequency = (100 * (fid + 8)) >> did;
286 static void amd_fixup_frequency(struct acpi_processor_px *px, int i) {};
289 static int acpi_processor_get_performance_states(struct acpi_processor *pr)
292 acpi_status status = AE_OK;
293 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
294 struct acpi_buffer format = { sizeof("NNNNNN"), "NNNNNN" };
295 struct acpi_buffer state = { 0, NULL };
296 union acpi_object *pss = NULL;
298 int last_invalid = -1;
301 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
302 if (ACPI_FAILURE(status)) {
303 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PSS"));
307 pss = buffer.pointer;
308 if (!pss || (pss->type != ACPI_TYPE_PACKAGE)) {
309 printk(KERN_ERR PREFIX "Invalid _PSS data\n");
314 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d performance states\n",
315 pss->package.count));
317 pr->performance->state_count = pss->package.count;
318 pr->performance->states =
319 kmalloc_array(pss->package.count,
320 sizeof(struct acpi_processor_px),
322 if (!pr->performance->states) {
327 for (i = 0; i < pr->performance->state_count; i++) {
329 struct acpi_processor_px *px = &(pr->performance->states[i]);
331 state.length = sizeof(struct acpi_processor_px);
334 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Extracting state %d\n", i));
336 status = acpi_extract_package(&(pss->package.elements[i]),
338 if (ACPI_FAILURE(status)) {
339 ACPI_EXCEPTION((AE_INFO, status, "Invalid _PSS data"));
341 kfree(pr->performance->states);
345 amd_fixup_frequency(px, i);
347 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
348 "State [%d]: core_frequency[%d] power[%d] transition_latency[%d] bus_master_latency[%d] control[0x%x] status[0x%x]\n",
350 (u32) px->core_frequency,
352 (u32) px->transition_latency,
353 (u32) px->bus_master_latency,
354 (u32) px->control, (u32) px->status));
357 * Check that ACPI's u64 MHz will be valid as u32 KHz in cpufreq
359 if (!px->core_frequency ||
360 ((u32)(px->core_frequency * 1000) !=
361 (px->core_frequency * 1000))) {
362 printk(KERN_ERR FW_BUG PREFIX
363 "Invalid BIOS _PSS frequency found for processor %d: 0x%llx MHz\n",
364 pr->id, px->core_frequency);
365 if (last_invalid == -1)
368 if (last_invalid != -1) {
370 * Copy this valid entry over last_invalid entry
372 memcpy(&(pr->performance->states[last_invalid]),
373 px, sizeof(struct acpi_processor_px));
379 if (last_invalid == 0) {
380 printk(KERN_ERR FW_BUG PREFIX
381 "No valid BIOS _PSS frequency found for processor %d\n", pr->id);
383 kfree(pr->performance->states);
384 pr->performance->states = NULL;
387 if (last_invalid > 0)
388 pr->performance->state_count = last_invalid;
391 kfree(buffer.pointer);
396 int acpi_processor_get_performance_info(struct acpi_processor *pr)
400 if (!pr || !pr->performance || !pr->handle)
403 if (!acpi_has_method(pr->handle, "_PCT")) {
404 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
405 "ACPI-based processor performance control unavailable\n"));
409 result = acpi_processor_get_performance_control(pr);
413 result = acpi_processor_get_performance_states(pr);
417 /* We need to call _PPC once when cpufreq starts */
419 result = acpi_processor_get_platform_limit(pr);
424 * Having _PPC but missing frequencies (_PSS, _PCT) is a very good hint that
425 * the BIOS is older than the CPU and does not know its frequencies
429 if (acpi_has_method(pr->handle, "_PPC")) {
430 if(boot_cpu_has(X86_FEATURE_EST))
431 printk(KERN_WARNING FW_BUG "BIOS needs update for CPU "
432 "frequency support\n");
437 EXPORT_SYMBOL_GPL(acpi_processor_get_performance_info);
439 int acpi_processor_pstate_control(void)
443 if (!acpi_gbl_FADT.smi_command || !acpi_gbl_FADT.pstate_control)
446 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
447 "Writing pstate_control [0x%x] to smi_command [0x%x]\n",
448 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command));
450 status = acpi_os_write_port(acpi_gbl_FADT.smi_command,
451 (u32)acpi_gbl_FADT.pstate_control, 8);
452 if (ACPI_SUCCESS(status))
455 ACPI_EXCEPTION((AE_INFO, status,
456 "Failed to write pstate_control [0x%x] to smi_command [0x%x]",
457 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command));
461 int acpi_processor_notify_smm(struct module *calling_module)
463 static int is_done = 0;
466 if (!acpi_processor_cpufreq_init)
469 if (!try_module_get(calling_module))
472 /* is_done is set to negative if an error occurred,
473 * and to postitive if _no_ error occurred, but SMM
474 * was already notified. This avoids double notification
475 * which might lead to unexpected results...
478 module_put(calling_module);
480 } else if (is_done < 0) {
481 module_put(calling_module);
487 result = acpi_processor_pstate_control();
489 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No SMI port or pstate_control\n"));
490 module_put(calling_module);
494 module_put(calling_module);
498 /* Success. If there's no _PPC, we need to fear nothing, so
499 * we can allow the cpufreq driver to be rmmod'ed. */
502 if (!acpi_processor_ppc_in_use)
503 module_put(calling_module);
508 EXPORT_SYMBOL(acpi_processor_notify_smm);
510 int acpi_processor_get_psd(acpi_handle handle, struct acpi_psd_package *pdomain)
513 acpi_status status = AE_OK;
514 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
515 struct acpi_buffer format = {sizeof("NNNNN"), "NNNNN"};
516 struct acpi_buffer state = {0, NULL};
517 union acpi_object *psd = NULL;
519 status = acpi_evaluate_object(handle, "_PSD", NULL, &buffer);
520 if (ACPI_FAILURE(status)) {
524 psd = buffer.pointer;
525 if (!psd || (psd->type != ACPI_TYPE_PACKAGE)) {
526 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
531 if (psd->package.count != 1) {
532 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
537 state.length = sizeof(struct acpi_psd_package);
538 state.pointer = pdomain;
540 status = acpi_extract_package(&(psd->package.elements[0]),
542 if (ACPI_FAILURE(status)) {
543 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
548 if (pdomain->num_entries != ACPI_PSD_REV0_ENTRIES) {
549 printk(KERN_ERR PREFIX "Unknown _PSD:num_entries\n");
554 if (pdomain->revision != ACPI_PSD_REV0_REVISION) {
555 printk(KERN_ERR PREFIX "Unknown _PSD:revision\n");
560 if (pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ALL &&
561 pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ANY &&
562 pdomain->coord_type != DOMAIN_COORD_TYPE_HW_ALL) {
563 printk(KERN_ERR PREFIX "Invalid _PSD:coord_type\n");
568 kfree(buffer.pointer);
571 EXPORT_SYMBOL(acpi_processor_get_psd);
573 int acpi_processor_preregister_performance(
574 struct acpi_processor_performance __percpu *performance)
579 cpumask_var_t covered_cpus;
580 struct acpi_processor *pr;
581 struct acpi_psd_package *pdomain;
582 struct acpi_processor *match_pr;
583 struct acpi_psd_package *match_pdomain;
585 if (!zalloc_cpumask_var(&covered_cpus, GFP_KERNEL))
588 mutex_lock(&performance_mutex);
591 * Check if another driver has already registered, and abort before
592 * changing pr->performance if it has. Check input data as well.
594 for_each_possible_cpu(i) {
595 pr = per_cpu(processors, i);
597 /* Look only at processors in ACPI namespace */
601 if (pr->performance) {
606 if (!performance || !per_cpu_ptr(performance, i)) {
612 /* Call _PSD for all CPUs */
613 for_each_possible_cpu(i) {
614 pr = per_cpu(processors, i);
618 pr->performance = per_cpu_ptr(performance, i);
619 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
620 pdomain = &(pr->performance->domain_info);
621 if (acpi_processor_get_psd(pr->handle, pdomain)) {
630 * Now that we have _PSD data from all CPUs, lets setup P-state
633 for_each_possible_cpu(i) {
634 pr = per_cpu(processors, i);
638 if (cpumask_test_cpu(i, covered_cpus))
641 pdomain = &(pr->performance->domain_info);
642 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
643 cpumask_set_cpu(i, covered_cpus);
644 if (pdomain->num_processors <= 1)
647 /* Validate the Domain info */
648 count_target = pdomain->num_processors;
649 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
650 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
651 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
652 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_HW;
653 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
654 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ANY;
656 for_each_possible_cpu(j) {
660 match_pr = per_cpu(processors, j);
664 match_pdomain = &(match_pr->performance->domain_info);
665 if (match_pdomain->domain != pdomain->domain)
668 /* Here i and j are in the same domain */
670 if (match_pdomain->num_processors != count_target) {
675 if (pdomain->coord_type != match_pdomain->coord_type) {
680 cpumask_set_cpu(j, covered_cpus);
681 cpumask_set_cpu(j, pr->performance->shared_cpu_map);
684 for_each_possible_cpu(j) {
688 match_pr = per_cpu(processors, j);
692 match_pdomain = &(match_pr->performance->domain_info);
693 if (match_pdomain->domain != pdomain->domain)
696 match_pr->performance->shared_type =
697 pr->performance->shared_type;
698 cpumask_copy(match_pr->performance->shared_cpu_map,
699 pr->performance->shared_cpu_map);
704 for_each_possible_cpu(i) {
705 pr = per_cpu(processors, i);
706 if (!pr || !pr->performance)
709 /* Assume no coordination on any error parsing domain info */
711 cpumask_clear(pr->performance->shared_cpu_map);
712 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
713 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
715 pr->performance = NULL; /* Will be set for real in register */
719 mutex_unlock(&performance_mutex);
720 free_cpumask_var(covered_cpus);
723 EXPORT_SYMBOL(acpi_processor_preregister_performance);
726 acpi_processor_register_performance(struct acpi_processor_performance
727 *performance, unsigned int cpu)
729 struct acpi_processor *pr;
731 if (!acpi_processor_cpufreq_init)
734 mutex_lock(&performance_mutex);
736 pr = per_cpu(processors, cpu);
738 mutex_unlock(&performance_mutex);
742 if (pr->performance) {
743 mutex_unlock(&performance_mutex);
747 WARN_ON(!performance);
749 pr->performance = performance;
751 if (acpi_processor_get_performance_info(pr)) {
752 pr->performance = NULL;
753 mutex_unlock(&performance_mutex);
757 mutex_unlock(&performance_mutex);
761 EXPORT_SYMBOL(acpi_processor_register_performance);
763 void acpi_processor_unregister_performance(unsigned int cpu)
765 struct acpi_processor *pr;
767 mutex_lock(&performance_mutex);
769 pr = per_cpu(processors, cpu);
771 mutex_unlock(&performance_mutex);
776 kfree(pr->performance->states);
777 pr->performance = NULL;
779 mutex_unlock(&performance_mutex);
784 EXPORT_SYMBOL(acpi_processor_unregister_performance);