2 * processor_perflib.c - ACPI Processor P-States Library ($Revision: 71 $)
4 * Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
5 * Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
6 * Copyright (C) 2004 Dominik Brodowski <linux@brodo.de>
7 * Copyright (C) 2004 Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
8 * - Added processor hotplug support
11 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or (at
16 * your option) any later version.
18 * This program is distributed in the hope that it will be useful, but
19 * WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
21 * General Public License for more details.
23 * You should have received a copy of the GNU General Public License along
24 * with this program; if not, write to the Free Software Foundation, Inc.,
25 * 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
29 #include <linux/kernel.h>
30 #include <linux/module.h>
31 #include <linux/init.h>
32 #include <linux/cpufreq.h>
33 #include <linux/slab.h>
36 #include <asm/cpufeature.h>
39 #include <acpi/acpi_bus.h>
40 #include <acpi/acpi_drivers.h>
41 #include <acpi/processor.h>
43 #define PREFIX "ACPI: "
45 #define ACPI_PROCESSOR_CLASS "processor"
46 #define ACPI_PROCESSOR_FILE_PERFORMANCE "performance"
47 #define _COMPONENT ACPI_PROCESSOR_COMPONENT
48 ACPI_MODULE_NAME("processor_perflib");
50 static DEFINE_MUTEX(performance_mutex);
53 * _PPC support is implemented as a CPUfreq policy notifier:
54 * This means each time a CPUfreq driver registered also with
55 * the ACPI core is asked to change the speed policy, the maximum
56 * value is adjusted so that it is within the platform limit.
58 * Also, when a new platform limit value is detected, the CPUfreq
59 * policy is adjusted accordingly.
63 * -1 -> cpufreq low level drivers not initialized -> _PSS, etc. not called yet
65 * 0 -> cpufreq low level drivers initialized -> consider _PPC values
66 * 1 -> ignore _PPC totally -> forced by user through boot param
68 static int ignore_ppc = -1;
69 module_param(ignore_ppc, int, 0644);
70 MODULE_PARM_DESC(ignore_ppc, "If the frequency of your machine gets wrongly" \
71 "limited by BIOS, this should help");
73 #define PPC_REGISTERED 1
76 static int acpi_processor_ppc_status;
78 static int acpi_processor_ppc_notifier(struct notifier_block *nb,
79 unsigned long event, void *data)
81 struct cpufreq_policy *policy = data;
82 struct acpi_processor *pr;
85 if (event == CPUFREQ_START && ignore_ppc <= 0) {
93 if (event != CPUFREQ_INCOMPATIBLE)
96 mutex_lock(&performance_mutex);
98 pr = per_cpu(processors, policy->cpu);
99 if (!pr || !pr->performance)
102 ppc = (unsigned int)pr->performance_platform_limit;
104 if (ppc >= pr->performance->state_count)
107 cpufreq_verify_within_limits(policy, 0,
108 pr->performance->states[ppc].
109 core_frequency * 1000);
112 mutex_unlock(&performance_mutex);
117 static struct notifier_block acpi_ppc_notifier_block = {
118 .notifier_call = acpi_processor_ppc_notifier,
121 static int acpi_processor_get_platform_limit(struct acpi_processor *pr)
123 acpi_status status = 0;
124 unsigned long long ppc = 0;
131 * _PPC indicates the maximum state currently supported by the platform
132 * (e.g. 0 = states 0..n; 1 = states 1..n; etc.
134 status = acpi_evaluate_integer(pr->handle, "_PPC", NULL, &ppc);
136 if (status != AE_NOT_FOUND)
137 acpi_processor_ppc_status |= PPC_IN_USE;
139 if (ACPI_FAILURE(status) && status != AE_NOT_FOUND) {
140 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PPC"));
144 pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr->id,
145 (int)ppc, ppc ? "" : "not");
147 pr->performance_platform_limit = (int)ppc;
152 #define ACPI_PROCESSOR_NOTIFY_PERFORMANCE 0x80
154 * acpi_processor_ppc_ost: Notify firmware the _PPC evaluation status
155 * @handle: ACPI processor handle
156 * @status: the status code of _PPC evaluation
157 * 0: success. OSPM is now using the performance state specificed.
158 * 1: failure. OSPM has not changed the number of P-states in use
160 static void acpi_processor_ppc_ost(acpi_handle handle, int status)
162 union acpi_object params[2] = {
163 {.type = ACPI_TYPE_INTEGER,},
164 {.type = ACPI_TYPE_INTEGER,},
166 struct acpi_object_list arg_list = {2, params};
168 if (acpi_has_method(handle, "_OST")) {
169 params[0].integer.value = ACPI_PROCESSOR_NOTIFY_PERFORMANCE;
170 params[1].integer.value = status;
171 acpi_evaluate_object(handle, "_OST", &arg_list, NULL);
175 int acpi_processor_ppc_has_changed(struct acpi_processor *pr, int event_flag)
181 * Only when it is notification event, the _OST object
182 * will be evaluated. Otherwise it is skipped.
185 acpi_processor_ppc_ost(pr->handle, 1);
189 ret = acpi_processor_get_platform_limit(pr);
191 * Only when it is notification event, the _OST object
192 * will be evaluated. Otherwise it is skipped.
196 acpi_processor_ppc_ost(pr->handle, 1);
198 acpi_processor_ppc_ost(pr->handle, 0);
203 return cpufreq_update_policy(pr->id);
206 int acpi_processor_get_bios_limit(int cpu, unsigned int *limit)
208 struct acpi_processor *pr;
210 pr = per_cpu(processors, cpu);
211 if (!pr || !pr->performance || !pr->performance->state_count)
213 *limit = pr->performance->states[pr->performance_platform_limit].
214 core_frequency * 1000;
217 EXPORT_SYMBOL(acpi_processor_get_bios_limit);
219 void acpi_processor_ppc_init(void)
221 if (!cpufreq_register_notifier
222 (&acpi_ppc_notifier_block, CPUFREQ_POLICY_NOTIFIER))
223 acpi_processor_ppc_status |= PPC_REGISTERED;
226 "Warning: Processor Platform Limit not supported.\n");
229 void acpi_processor_ppc_exit(void)
231 if (acpi_processor_ppc_status & PPC_REGISTERED)
232 cpufreq_unregister_notifier(&acpi_ppc_notifier_block,
233 CPUFREQ_POLICY_NOTIFIER);
235 acpi_processor_ppc_status &= ~PPC_REGISTERED;
239 * Do a quick check if the systems looks like it should use ACPI
240 * cpufreq. We look at a _PCT method being available, but don't
241 * do a whole lot of sanity checks.
243 void acpi_processor_load_module(struct acpi_processor *pr)
245 static int requested;
246 acpi_status status = 0;
247 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
249 if (!arch_has_acpi_pdc() || requested)
251 status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
252 if (!ACPI_FAILURE(status)) {
253 printk(KERN_INFO PREFIX "Requesting acpi_cpufreq\n");
254 request_module_nowait("acpi_cpufreq");
257 kfree(buffer.pointer);
260 static int acpi_processor_get_performance_control(struct acpi_processor *pr)
263 acpi_status status = 0;
264 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
265 union acpi_object *pct = NULL;
266 union acpi_object obj = { 0 };
269 status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
270 if (ACPI_FAILURE(status)) {
271 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PCT"));
275 pct = (union acpi_object *)buffer.pointer;
276 if (!pct || (pct->type != ACPI_TYPE_PACKAGE)
277 || (pct->package.count != 2)) {
278 printk(KERN_ERR PREFIX "Invalid _PCT data\n");
287 obj = pct->package.elements[0];
289 if ((obj.type != ACPI_TYPE_BUFFER)
290 || (obj.buffer.length < sizeof(struct acpi_pct_register))
291 || (obj.buffer.pointer == NULL)) {
292 printk(KERN_ERR PREFIX "Invalid _PCT data (control_register)\n");
296 memcpy(&pr->performance->control_register, obj.buffer.pointer,
297 sizeof(struct acpi_pct_register));
303 obj = pct->package.elements[1];
305 if ((obj.type != ACPI_TYPE_BUFFER)
306 || (obj.buffer.length < sizeof(struct acpi_pct_register))
307 || (obj.buffer.pointer == NULL)) {
308 printk(KERN_ERR PREFIX "Invalid _PCT data (status_register)\n");
313 memcpy(&pr->performance->status_register, obj.buffer.pointer,
314 sizeof(struct acpi_pct_register));
317 kfree(buffer.pointer);
324 * Some AMDs have 50MHz frequency multiples, but only provide 100MHz rounding
325 * in their ACPI data. Calculate the real values and fix up the _PSS data.
327 static void amd_fixup_frequency(struct acpi_processor_px *px, int i)
329 u32 hi, lo, fid, did;
330 int index = px->control & 0x00000007;
332 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
335 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
336 || boot_cpu_data.x86 == 0x11) {
337 rdmsr(MSR_AMD_PSTATE_DEF_BASE + index, lo, hi);
340 * Bit 63: PstateEn. Read-write. If set, the P-state is valid.
347 if (boot_cpu_data.x86 == 0x10)
348 px->core_frequency = (100 * (fid + 0x10)) >> did;
350 px->core_frequency = (100 * (fid + 8)) >> did;
354 static void amd_fixup_frequency(struct acpi_processor_px *px, int i) {};
357 static int acpi_processor_get_performance_states(struct acpi_processor *pr)
360 acpi_status status = AE_OK;
361 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
362 struct acpi_buffer format = { sizeof("NNNNNN"), "NNNNNN" };
363 struct acpi_buffer state = { 0, NULL };
364 union acpi_object *pss = NULL;
366 int last_invalid = -1;
369 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
370 if (ACPI_FAILURE(status)) {
371 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PSS"));
375 pss = buffer.pointer;
376 if (!pss || (pss->type != ACPI_TYPE_PACKAGE)) {
377 printk(KERN_ERR PREFIX "Invalid _PSS data\n");
382 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d performance states\n",
383 pss->package.count));
385 pr->performance->state_count = pss->package.count;
386 pr->performance->states =
387 kmalloc(sizeof(struct acpi_processor_px) * pss->package.count,
389 if (!pr->performance->states) {
394 for (i = 0; i < pr->performance->state_count; i++) {
396 struct acpi_processor_px *px = &(pr->performance->states[i]);
398 state.length = sizeof(struct acpi_processor_px);
401 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Extracting state %d\n", i));
403 status = acpi_extract_package(&(pss->package.elements[i]),
405 if (ACPI_FAILURE(status)) {
406 ACPI_EXCEPTION((AE_INFO, status, "Invalid _PSS data"));
408 kfree(pr->performance->states);
412 amd_fixup_frequency(px, i);
414 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
415 "State [%d]: core_frequency[%d] power[%d] transition_latency[%d] bus_master_latency[%d] control[0x%x] status[0x%x]\n",
417 (u32) px->core_frequency,
419 (u32) px->transition_latency,
420 (u32) px->bus_master_latency,
421 (u32) px->control, (u32) px->status));
424 * Check that ACPI's u64 MHz will be valid as u32 KHz in cpufreq
426 if (!px->core_frequency ||
427 ((u32)(px->core_frequency * 1000) !=
428 (px->core_frequency * 1000))) {
429 printk(KERN_ERR FW_BUG PREFIX
430 "Invalid BIOS _PSS frequency found for processor %d: 0x%llx MHz\n",
431 pr->id, px->core_frequency);
432 if (last_invalid == -1)
435 if (last_invalid != -1) {
437 * Copy this valid entry over last_invalid entry
439 memcpy(&(pr->performance->states[last_invalid]),
440 px, sizeof(struct acpi_processor_px));
446 if (last_invalid == 0) {
447 printk(KERN_ERR FW_BUG PREFIX
448 "No valid BIOS _PSS frequency found for processor %d\n", pr->id);
450 kfree(pr->performance->states);
451 pr->performance->states = NULL;
454 if (last_invalid > 0)
455 pr->performance->state_count = last_invalid;
458 kfree(buffer.pointer);
463 int acpi_processor_get_performance_info(struct acpi_processor *pr)
467 if (!pr || !pr->performance || !pr->handle)
470 if (!acpi_has_method(pr->handle, "_PCT")) {
471 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
472 "ACPI-based processor performance control unavailable\n"));
476 result = acpi_processor_get_performance_control(pr);
480 result = acpi_processor_get_performance_states(pr);
484 /* We need to call _PPC once when cpufreq starts */
486 result = acpi_processor_get_platform_limit(pr);
491 * Having _PPC but missing frequencies (_PSS, _PCT) is a very good hint that
492 * the BIOS is older than the CPU and does not know its frequencies
496 if (acpi_has_method(pr->handle, "_PPC")) {
497 if(boot_cpu_has(X86_FEATURE_EST))
498 printk(KERN_WARNING FW_BUG "BIOS needs update for CPU "
499 "frequency support\n");
504 EXPORT_SYMBOL_GPL(acpi_processor_get_performance_info);
505 int acpi_processor_notify_smm(struct module *calling_module)
508 static int is_done = 0;
511 if (!(acpi_processor_ppc_status & PPC_REGISTERED))
514 if (!try_module_get(calling_module))
517 /* is_done is set to negative if an error occurred,
518 * and to postitive if _no_ error occurred, but SMM
519 * was already notified. This avoids double notification
520 * which might lead to unexpected results...
523 module_put(calling_module);
525 } else if (is_done < 0) {
526 module_put(calling_module);
532 /* Can't write pstate_control to smi_command if either value is zero */
533 if ((!acpi_gbl_FADT.smi_command) || (!acpi_gbl_FADT.pstate_control)) {
534 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No SMI port or pstate_control\n"));
535 module_put(calling_module);
539 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
540 "Writing pstate_control [0x%x] to smi_command [0x%x]\n",
541 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command));
543 status = acpi_os_write_port(acpi_gbl_FADT.smi_command,
544 (u32) acpi_gbl_FADT.pstate_control, 8);
545 if (ACPI_FAILURE(status)) {
546 ACPI_EXCEPTION((AE_INFO, status,
547 "Failed to write pstate_control [0x%x] to "
548 "smi_command [0x%x]", acpi_gbl_FADT.pstate_control,
549 acpi_gbl_FADT.smi_command));
550 module_put(calling_module);
554 /* Success. If there's no _PPC, we need to fear nothing, so
555 * we can allow the cpufreq driver to be rmmod'ed. */
558 if (!(acpi_processor_ppc_status & PPC_IN_USE))
559 module_put(calling_module);
564 EXPORT_SYMBOL(acpi_processor_notify_smm);
566 static int acpi_processor_get_psd(struct acpi_processor *pr)
569 acpi_status status = AE_OK;
570 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
571 struct acpi_buffer format = {sizeof("NNNNN"), "NNNNN"};
572 struct acpi_buffer state = {0, NULL};
573 union acpi_object *psd = NULL;
574 struct acpi_psd_package *pdomain;
576 status = acpi_evaluate_object(pr->handle, "_PSD", NULL, &buffer);
577 if (ACPI_FAILURE(status)) {
581 psd = buffer.pointer;
582 if (!psd || (psd->type != ACPI_TYPE_PACKAGE)) {
583 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
588 if (psd->package.count != 1) {
589 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
594 pdomain = &(pr->performance->domain_info);
596 state.length = sizeof(struct acpi_psd_package);
597 state.pointer = pdomain;
599 status = acpi_extract_package(&(psd->package.elements[0]),
601 if (ACPI_FAILURE(status)) {
602 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
607 if (pdomain->num_entries != ACPI_PSD_REV0_ENTRIES) {
608 printk(KERN_ERR PREFIX "Unknown _PSD:num_entries\n");
613 if (pdomain->revision != ACPI_PSD_REV0_REVISION) {
614 printk(KERN_ERR PREFIX "Unknown _PSD:revision\n");
619 if (pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ALL &&
620 pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ANY &&
621 pdomain->coord_type != DOMAIN_COORD_TYPE_HW_ALL) {
622 printk(KERN_ERR PREFIX "Invalid _PSD:coord_type\n");
627 kfree(buffer.pointer);
631 int acpi_processor_preregister_performance(
632 struct acpi_processor_performance __percpu *performance)
637 cpumask_var_t covered_cpus;
638 struct acpi_processor *pr;
639 struct acpi_psd_package *pdomain;
640 struct acpi_processor *match_pr;
641 struct acpi_psd_package *match_pdomain;
643 if (!zalloc_cpumask_var(&covered_cpus, GFP_KERNEL))
646 mutex_lock(&performance_mutex);
649 * Check if another driver has already registered, and abort before
650 * changing pr->performance if it has. Check input data as well.
652 for_each_possible_cpu(i) {
653 pr = per_cpu(processors, i);
655 /* Look only at processors in ACPI namespace */
659 if (pr->performance) {
664 if (!performance || !per_cpu_ptr(performance, i)) {
670 /* Call _PSD for all CPUs */
671 for_each_possible_cpu(i) {
672 pr = per_cpu(processors, i);
676 pr->performance = per_cpu_ptr(performance, i);
677 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
678 if (acpi_processor_get_psd(pr)) {
687 * Now that we have _PSD data from all CPUs, lets setup P-state
690 for_each_possible_cpu(i) {
691 pr = per_cpu(processors, i);
695 if (cpumask_test_cpu(i, covered_cpus))
698 pdomain = &(pr->performance->domain_info);
699 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
700 cpumask_set_cpu(i, covered_cpus);
701 if (pdomain->num_processors <= 1)
704 /* Validate the Domain info */
705 count_target = pdomain->num_processors;
706 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
707 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
708 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
709 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_HW;
710 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
711 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ANY;
713 for_each_possible_cpu(j) {
717 match_pr = per_cpu(processors, j);
721 match_pdomain = &(match_pr->performance->domain_info);
722 if (match_pdomain->domain != pdomain->domain)
725 /* Here i and j are in the same domain */
727 if (match_pdomain->num_processors != count_target) {
732 if (pdomain->coord_type != match_pdomain->coord_type) {
737 cpumask_set_cpu(j, covered_cpus);
738 cpumask_set_cpu(j, pr->performance->shared_cpu_map);
741 for_each_possible_cpu(j) {
745 match_pr = per_cpu(processors, j);
749 match_pdomain = &(match_pr->performance->domain_info);
750 if (match_pdomain->domain != pdomain->domain)
753 match_pr->performance->shared_type =
754 pr->performance->shared_type;
755 cpumask_copy(match_pr->performance->shared_cpu_map,
756 pr->performance->shared_cpu_map);
761 for_each_possible_cpu(i) {
762 pr = per_cpu(processors, i);
763 if (!pr || !pr->performance)
766 /* Assume no coordination on any error parsing domain info */
768 cpumask_clear(pr->performance->shared_cpu_map);
769 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
770 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
772 pr->performance = NULL; /* Will be set for real in register */
776 mutex_unlock(&performance_mutex);
777 free_cpumask_var(covered_cpus);
780 EXPORT_SYMBOL(acpi_processor_preregister_performance);
783 acpi_processor_register_performance(struct acpi_processor_performance
784 *performance, unsigned int cpu)
786 struct acpi_processor *pr;
788 if (!(acpi_processor_ppc_status & PPC_REGISTERED))
791 mutex_lock(&performance_mutex);
793 pr = per_cpu(processors, cpu);
795 mutex_unlock(&performance_mutex);
799 if (pr->performance) {
800 mutex_unlock(&performance_mutex);
804 WARN_ON(!performance);
806 pr->performance = performance;
808 if (acpi_processor_get_performance_info(pr)) {
809 pr->performance = NULL;
810 mutex_unlock(&performance_mutex);
814 mutex_unlock(&performance_mutex);
818 EXPORT_SYMBOL(acpi_processor_register_performance);
821 acpi_processor_unregister_performance(struct acpi_processor_performance
822 *performance, unsigned int cpu)
824 struct acpi_processor *pr;
826 mutex_lock(&performance_mutex);
828 pr = per_cpu(processors, cpu);
830 mutex_unlock(&performance_mutex);
835 kfree(pr->performance->states);
836 pr->performance = NULL;
838 mutex_unlock(&performance_mutex);
843 EXPORT_SYMBOL(acpi_processor_unregister_performance);