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 #define pr_fmt(fmt) "ACPI: " fmt
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/init.h>
17 #include <linux/cpufreq.h>
18 #include <linux/slab.h>
19 #include <linux/acpi.h>
20 #include <acpi/processor.h>
22 #include <asm/cpufeature.h>
25 #define ACPI_PROCESSOR_FILE_PERFORMANCE "performance"
27 static DEFINE_MUTEX(performance_mutex);
30 * _PPC support is implemented as a CPUfreq policy notifier:
31 * This means each time a CPUfreq driver registered also with
32 * the ACPI core is asked to change the speed policy, the maximum
33 * value is adjusted so that it is within the platform limit.
35 * Also, when a new platform limit value is detected, the CPUfreq
36 * policy is adjusted accordingly.
40 * -1 -> cpufreq low level drivers not initialized -> _PSS, etc. not called yet
42 * 0 -> cpufreq low level drivers initialized -> consider _PPC values
43 * 1 -> ignore _PPC totally -> forced by user through boot param
45 static int ignore_ppc = -1;
46 module_param(ignore_ppc, int, 0644);
47 MODULE_PARM_DESC(ignore_ppc, "If the frequency of your machine gets wrongly" \
48 "limited by BIOS, this should help");
50 static bool acpi_processor_ppc_in_use;
52 static int acpi_processor_get_platform_limit(struct acpi_processor *pr)
54 acpi_status status = 0;
55 unsigned long long ppc = 0;
64 * _PPC indicates the maximum state currently supported by the platform
65 * (e.g. 0 = states 0..n; 1 = states 1..n; etc.
67 status = acpi_evaluate_integer(pr->handle, "_PPC", NULL, &ppc);
68 if (status != AE_NOT_FOUND) {
69 acpi_processor_ppc_in_use = true;
71 if (ACPI_FAILURE(status)) {
72 acpi_evaluation_failure_warn(pr->handle, "_PPC", status);
79 if (pr->performance_platform_limit == index ||
80 ppc >= pr->performance->state_count)
83 pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr->id,
84 index, index ? "is" : "is not");
86 pr->performance_platform_limit = index;
88 if (unlikely(!freq_qos_request_active(&pr->perflib_req)))
92 * If _PPC returns 0, it means that all of the available states can be
96 qos_value = FREQ_QOS_MAX_DEFAULT_VALUE;
98 qos_value = pr->performance->states[index].core_frequency * 1000;
100 ret = freq_qos_update_request(&pr->perflib_req, qos_value);
102 pr_warn("Failed to update perflib freq constraint: CPU%d (%d)\n",
109 #define ACPI_PROCESSOR_NOTIFY_PERFORMANCE 0x80
111 * acpi_processor_ppc_ost: Notify firmware the _PPC evaluation status
112 * @handle: ACPI processor handle
113 * @status: the status code of _PPC evaluation
114 * 0: success. OSPM is now using the performance state specified.
115 * 1: failure. OSPM has not changed the number of P-states in use
117 static void acpi_processor_ppc_ost(acpi_handle handle, int status)
119 if (acpi_has_method(handle, "_OST"))
120 acpi_evaluate_ost(handle, ACPI_PROCESSOR_NOTIFY_PERFORMANCE,
124 void acpi_processor_ppc_has_changed(struct acpi_processor *pr, int event_flag)
128 if (ignore_ppc || !pr->performance) {
130 * Only when it is notification event, the _OST object
131 * will be evaluated. Otherwise it is skipped.
134 acpi_processor_ppc_ost(pr->handle, 1);
138 ret = acpi_processor_get_platform_limit(pr);
140 * Only when it is notification event, the _OST object
141 * will be evaluated. Otherwise it is skipped.
145 acpi_processor_ppc_ost(pr->handle, 1);
147 acpi_processor_ppc_ost(pr->handle, 0);
150 cpufreq_update_limits(pr->id);
153 int acpi_processor_get_bios_limit(int cpu, unsigned int *limit)
155 struct acpi_processor *pr;
157 pr = per_cpu(processors, cpu);
158 if (!pr || !pr->performance || !pr->performance->state_count)
161 *limit = pr->performance->states[pr->performance_platform_limit].
162 core_frequency * 1000;
165 EXPORT_SYMBOL(acpi_processor_get_bios_limit);
167 void acpi_processor_ignore_ppc_init(void)
173 void acpi_processor_ppc_init(struct cpufreq_policy *policy)
177 for_each_cpu(cpu, policy->related_cpus) {
178 struct acpi_processor *pr = per_cpu(processors, cpu);
185 * Reset performance_platform_limit in case there is a stale
186 * value in it, so as to make it match the "no limit" QoS value
189 pr->performance_platform_limit = 0;
191 ret = freq_qos_add_request(&policy->constraints,
192 &pr->perflib_req, FREQ_QOS_MAX,
193 FREQ_QOS_MAX_DEFAULT_VALUE);
195 pr_err("Failed to add freq constraint for CPU%d (%d)\n",
200 void acpi_processor_ppc_exit(struct cpufreq_policy *policy)
204 for_each_cpu(cpu, policy->related_cpus) {
205 struct acpi_processor *pr = per_cpu(processors, cpu);
208 freq_qos_remove_request(&pr->perflib_req);
212 static int acpi_processor_get_performance_control(struct acpi_processor *pr)
215 acpi_status status = 0;
216 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
217 union acpi_object *pct = NULL;
218 union acpi_object obj = { 0 };
220 status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
221 if (ACPI_FAILURE(status)) {
222 acpi_evaluation_failure_warn(pr->handle, "_PCT", status);
226 pct = (union acpi_object *)buffer.pointer;
227 if (!pct || pct->type != ACPI_TYPE_PACKAGE || pct->package.count != 2) {
228 pr_err("Invalid _PCT data\n");
237 obj = pct->package.elements[0];
239 if (!obj.buffer.pointer || obj.type != ACPI_TYPE_BUFFER ||
240 obj.buffer.length < sizeof(struct acpi_pct_register)) {
241 pr_err("Invalid _PCT data (control_register)\n");
245 memcpy(&pr->performance->control_register, obj.buffer.pointer,
246 sizeof(struct acpi_pct_register));
252 obj = pct->package.elements[1];
254 if (!obj.buffer.pointer || obj.type != ACPI_TYPE_BUFFER ||
255 obj.buffer.length < sizeof(struct acpi_pct_register)) {
256 pr_err("Invalid _PCT data (status_register)\n");
261 memcpy(&pr->performance->status_register, obj.buffer.pointer,
262 sizeof(struct acpi_pct_register));
265 kfree(buffer.pointer);
272 * Some AMDs have 50MHz frequency multiples, but only provide 100MHz rounding
273 * in their ACPI data. Calculate the real values and fix up the _PSS data.
275 static void amd_fixup_frequency(struct acpi_processor_px *px, int i)
277 u32 hi, lo, fid, did;
278 int index = px->control & 0x00000007;
280 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
283 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10) ||
284 boot_cpu_data.x86 == 0x11) {
285 rdmsr(MSR_AMD_PSTATE_DEF_BASE + index, lo, hi);
288 * Bit 63: PstateEn. Read-write. If set, the P-state is valid.
295 if (boot_cpu_data.x86 == 0x10)
296 px->core_frequency = (100 * (fid + 0x10)) >> did;
298 px->core_frequency = (100 * (fid + 8)) >> did;
302 static void amd_fixup_frequency(struct acpi_processor_px *px, int i) {};
305 static int acpi_processor_get_performance_states(struct acpi_processor *pr)
308 acpi_status status = AE_OK;
309 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
310 struct acpi_buffer format = { sizeof("NNNNNN"), "NNNNNN" };
311 struct acpi_buffer state = { 0, NULL };
312 union acpi_object *pss = NULL;
314 int last_invalid = -1;
316 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
317 if (ACPI_FAILURE(status)) {
318 acpi_evaluation_failure_warn(pr->handle, "_PSS", status);
322 pss = buffer.pointer;
323 if (!pss || pss->type != ACPI_TYPE_PACKAGE) {
324 pr_err("Invalid _PSS data\n");
329 acpi_handle_debug(pr->handle, "Found %d performance states\n",
332 pr->performance->state_count = pss->package.count;
333 pr->performance->states =
334 kmalloc_array(pss->package.count,
335 sizeof(struct acpi_processor_px),
337 if (!pr->performance->states) {
342 for (i = 0; i < pr->performance->state_count; i++) {
344 struct acpi_processor_px *px = &(pr->performance->states[i]);
346 state.length = sizeof(struct acpi_processor_px);
349 acpi_handle_debug(pr->handle, "Extracting state %d\n", i);
351 status = acpi_extract_package(&(pss->package.elements[i]),
353 if (ACPI_FAILURE(status)) {
354 acpi_handle_warn(pr->handle, "Invalid _PSS data: %s\n",
355 acpi_format_exception(status));
357 kfree(pr->performance->states);
361 amd_fixup_frequency(px, i);
363 acpi_handle_debug(pr->handle,
364 "State [%d]: core_frequency[%d] power[%d] transition_latency[%d] bus_master_latency[%d] control[0x%x] status[0x%x]\n",
366 (u32) px->core_frequency,
368 (u32) px->transition_latency,
369 (u32) px->bus_master_latency,
370 (u32) px->control, (u32) px->status);
373 * Check that ACPI's u64 MHz will be valid as u32 KHz in cpufreq
375 if (!px->core_frequency ||
376 (u32)(px->core_frequency * 1000) != px->core_frequency * 1000) {
378 "Invalid BIOS _PSS frequency found for processor %d: 0x%llx MHz\n",
379 pr->id, px->core_frequency);
380 if (last_invalid == -1)
383 if (last_invalid != -1) {
385 * Copy this valid entry over last_invalid entry
387 memcpy(&(pr->performance->states[last_invalid]),
388 px, sizeof(struct acpi_processor_px));
394 if (last_invalid == 0) {
396 "No valid BIOS _PSS frequency found for processor %d\n", pr->id);
398 kfree(pr->performance->states);
399 pr->performance->states = NULL;
402 if (last_invalid > 0)
403 pr->performance->state_count = last_invalid;
406 kfree(buffer.pointer);
411 int acpi_processor_get_performance_info(struct acpi_processor *pr)
415 if (!pr || !pr->performance || !pr->handle)
418 if (!acpi_has_method(pr->handle, "_PCT")) {
419 acpi_handle_debug(pr->handle,
420 "ACPI-based processor performance control unavailable\n");
424 result = acpi_processor_get_performance_control(pr);
428 result = acpi_processor_get_performance_states(pr);
432 /* We need to call _PPC once when cpufreq starts */
434 result = acpi_processor_get_platform_limit(pr);
439 * Having _PPC but missing frequencies (_PSS, _PCT) is a very good hint that
440 * the BIOS is older than the CPU and does not know its frequencies
444 if (acpi_has_method(pr->handle, "_PPC")) {
445 if(boot_cpu_has(X86_FEATURE_EST))
446 pr_warn(FW_BUG "BIOS needs update for CPU "
447 "frequency support\n");
452 EXPORT_SYMBOL_GPL(acpi_processor_get_performance_info);
454 int acpi_processor_pstate_control(void)
458 if (!acpi_gbl_FADT.smi_command || !acpi_gbl_FADT.pstate_control)
461 pr_debug("Writing pstate_control [0x%x] to smi_command [0x%x]\n",
462 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command);
464 status = acpi_os_write_port(acpi_gbl_FADT.smi_command,
465 (u32)acpi_gbl_FADT.pstate_control, 8);
466 if (ACPI_SUCCESS(status))
469 pr_warn("Failed to write pstate_control [0x%x] to smi_command [0x%x]: %s\n",
470 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command,
471 acpi_format_exception(status));
475 int acpi_processor_notify_smm(struct module *calling_module)
480 if (!acpi_processor_cpufreq_init)
483 if (!try_module_get(calling_module))
487 * is_done is set to negative if an error occurs and to 1 if no error
488 * occurrs, but SMM has been notified already. This avoids repeated
489 * notification which might lead to unexpected results.
498 result = acpi_processor_pstate_control();
503 pr_debug("No SMI port or pstate_control\n");
511 * Success. If there _PPC, unloading the cpufreq driver would be risky,
512 * so disallow it in that case.
514 if (acpi_processor_ppc_in_use)
518 module_put(calling_module);
521 EXPORT_SYMBOL(acpi_processor_notify_smm);
523 int acpi_processor_get_psd(acpi_handle handle, struct acpi_psd_package *pdomain)
526 acpi_status status = AE_OK;
527 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
528 struct acpi_buffer format = {sizeof("NNNNN"), "NNNNN"};
529 struct acpi_buffer state = {0, NULL};
530 union acpi_object *psd = NULL;
532 status = acpi_evaluate_object(handle, "_PSD", NULL, &buffer);
533 if (ACPI_FAILURE(status)) {
537 psd = buffer.pointer;
538 if (!psd || psd->type != ACPI_TYPE_PACKAGE) {
539 pr_err("Invalid _PSD data\n");
544 if (psd->package.count != 1) {
545 pr_err("Invalid _PSD data\n");
550 state.length = sizeof(struct acpi_psd_package);
551 state.pointer = pdomain;
553 status = acpi_extract_package(&(psd->package.elements[0]), &format, &state);
554 if (ACPI_FAILURE(status)) {
555 pr_err("Invalid _PSD data\n");
560 if (pdomain->num_entries != ACPI_PSD_REV0_ENTRIES) {
561 pr_err("Unknown _PSD:num_entries\n");
566 if (pdomain->revision != ACPI_PSD_REV0_REVISION) {
567 pr_err("Unknown _PSD:revision\n");
572 if (pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ALL &&
573 pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ANY &&
574 pdomain->coord_type != DOMAIN_COORD_TYPE_HW_ALL) {
575 pr_err("Invalid _PSD:coord_type\n");
580 kfree(buffer.pointer);
583 EXPORT_SYMBOL(acpi_processor_get_psd);
585 int acpi_processor_preregister_performance(
586 struct acpi_processor_performance __percpu *performance)
591 cpumask_var_t covered_cpus;
592 struct acpi_processor *pr;
593 struct acpi_psd_package *pdomain;
594 struct acpi_processor *match_pr;
595 struct acpi_psd_package *match_pdomain;
597 if (!zalloc_cpumask_var(&covered_cpus, GFP_KERNEL))
600 mutex_lock(&performance_mutex);
603 * Check if another driver has already registered, and abort before
604 * changing pr->performance if it has. Check input data as well.
606 for_each_possible_cpu(i) {
607 pr = per_cpu(processors, i);
609 /* Look only at processors in ACPI namespace */
613 if (pr->performance) {
618 if (!performance || !per_cpu_ptr(performance, i)) {
624 /* Call _PSD for all CPUs */
625 for_each_possible_cpu(i) {
626 pr = per_cpu(processors, i);
630 pr->performance = per_cpu_ptr(performance, i);
631 pdomain = &(pr->performance->domain_info);
632 if (acpi_processor_get_psd(pr->handle, pdomain)) {
641 * Now that we have _PSD data from all CPUs, lets setup P-state
644 for_each_possible_cpu(i) {
645 pr = per_cpu(processors, i);
649 if (cpumask_test_cpu(i, covered_cpus))
652 pdomain = &(pr->performance->domain_info);
653 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
654 cpumask_set_cpu(i, covered_cpus);
655 if (pdomain->num_processors <= 1)
658 /* Validate the Domain info */
659 count_target = pdomain->num_processors;
660 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
661 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
662 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
663 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_HW;
664 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
665 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ANY;
667 for_each_possible_cpu(j) {
671 match_pr = per_cpu(processors, j);
675 match_pdomain = &(match_pr->performance->domain_info);
676 if (match_pdomain->domain != pdomain->domain)
679 /* Here i and j are in the same domain */
681 if (match_pdomain->num_processors != count_target) {
686 if (pdomain->coord_type != match_pdomain->coord_type) {
691 cpumask_set_cpu(j, covered_cpus);
692 cpumask_set_cpu(j, pr->performance->shared_cpu_map);
695 for_each_possible_cpu(j) {
699 match_pr = per_cpu(processors, j);
703 match_pdomain = &(match_pr->performance->domain_info);
704 if (match_pdomain->domain != pdomain->domain)
707 match_pr->performance->shared_type =
708 pr->performance->shared_type;
709 cpumask_copy(match_pr->performance->shared_cpu_map,
710 pr->performance->shared_cpu_map);
715 for_each_possible_cpu(i) {
716 pr = per_cpu(processors, i);
717 if (!pr || !pr->performance)
720 /* Assume no coordination on any error parsing domain info */
722 cpumask_clear(pr->performance->shared_cpu_map);
723 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
724 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_NONE;
726 pr->performance = NULL; /* Will be set for real in register */
730 mutex_unlock(&performance_mutex);
731 free_cpumask_var(covered_cpus);
734 EXPORT_SYMBOL(acpi_processor_preregister_performance);
736 int acpi_processor_register_performance(struct acpi_processor_performance
737 *performance, unsigned int cpu)
739 struct acpi_processor *pr;
741 if (!acpi_processor_cpufreq_init)
744 mutex_lock(&performance_mutex);
746 pr = per_cpu(processors, cpu);
748 mutex_unlock(&performance_mutex);
752 if (pr->performance) {
753 mutex_unlock(&performance_mutex);
757 WARN_ON(!performance);
759 pr->performance = performance;
761 if (acpi_processor_get_performance_info(pr)) {
762 pr->performance = NULL;
763 mutex_unlock(&performance_mutex);
767 mutex_unlock(&performance_mutex);
770 EXPORT_SYMBOL(acpi_processor_register_performance);
772 void acpi_processor_unregister_performance(unsigned int cpu)
774 struct acpi_processor *pr;
776 mutex_lock(&performance_mutex);
778 pr = per_cpu(processors, cpu);
783 kfree(pr->performance->states);
785 pr->performance = NULL;
788 mutex_unlock(&performance_mutex);
790 EXPORT_SYMBOL(acpi_processor_unregister_performance);