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};
169 params[0].integer.value = ACPI_PROCESSOR_NOTIFY_PERFORMANCE;
170 params[1].integer.value = status;
172 /* when there is no _OST , skip it */
173 if (ACPI_FAILURE(acpi_get_handle(handle, "_OST", &temp)))
176 acpi_evaluate_object(handle, "_OST", &arg_list, NULL);
180 int acpi_processor_ppc_has_changed(struct acpi_processor *pr, int event_flag)
186 * Only when it is notification event, the _OST object
187 * will be evaluated. Otherwise it is skipped.
190 acpi_processor_ppc_ost(pr->handle, 1);
194 ret = acpi_processor_get_platform_limit(pr);
196 * Only when it is notification event, the _OST object
197 * will be evaluated. Otherwise it is skipped.
201 acpi_processor_ppc_ost(pr->handle, 1);
203 acpi_processor_ppc_ost(pr->handle, 0);
208 return cpufreq_update_policy(pr->id);
211 int acpi_processor_get_bios_limit(int cpu, unsigned int *limit)
213 struct acpi_processor *pr;
215 pr = per_cpu(processors, cpu);
216 if (!pr || !pr->performance || !pr->performance->state_count)
218 *limit = pr->performance->states[pr->performance_platform_limit].
219 core_frequency * 1000;
222 EXPORT_SYMBOL(acpi_processor_get_bios_limit);
224 void acpi_processor_ppc_init(void)
226 if (!cpufreq_register_notifier
227 (&acpi_ppc_notifier_block, CPUFREQ_POLICY_NOTIFIER))
228 acpi_processor_ppc_status |= PPC_REGISTERED;
231 "Warning: Processor Platform Limit not supported.\n");
234 void acpi_processor_ppc_exit(void)
236 if (acpi_processor_ppc_status & PPC_REGISTERED)
237 cpufreq_unregister_notifier(&acpi_ppc_notifier_block,
238 CPUFREQ_POLICY_NOTIFIER);
240 acpi_processor_ppc_status &= ~PPC_REGISTERED;
244 * Do a quick check if the systems looks like it should use ACPI
245 * cpufreq. We look at a _PCT method being available, but don't
246 * do a whole lot of sanity checks.
248 void acpi_processor_load_module(struct acpi_processor *pr)
250 static int requested;
251 acpi_status status = 0;
252 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
254 if (!arch_has_acpi_pdc() || requested)
256 status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
257 if (!ACPI_FAILURE(status)) {
258 printk(KERN_INFO PREFIX "Requesting acpi_cpufreq\n");
259 request_module_nowait("acpi_cpufreq");
262 kfree(buffer.pointer);
265 static int acpi_processor_get_performance_control(struct acpi_processor *pr)
268 acpi_status status = 0;
269 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
270 union acpi_object *pct = NULL;
271 union acpi_object obj = { 0 };
274 status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
275 if (ACPI_FAILURE(status)) {
276 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PCT"));
280 pct = (union acpi_object *)buffer.pointer;
281 if (!pct || (pct->type != ACPI_TYPE_PACKAGE)
282 || (pct->package.count != 2)) {
283 printk(KERN_ERR PREFIX "Invalid _PCT data\n");
292 obj = pct->package.elements[0];
294 if ((obj.type != ACPI_TYPE_BUFFER)
295 || (obj.buffer.length < sizeof(struct acpi_pct_register))
296 || (obj.buffer.pointer == NULL)) {
297 printk(KERN_ERR PREFIX "Invalid _PCT data (control_register)\n");
301 memcpy(&pr->performance->control_register, obj.buffer.pointer,
302 sizeof(struct acpi_pct_register));
308 obj = pct->package.elements[1];
310 if ((obj.type != ACPI_TYPE_BUFFER)
311 || (obj.buffer.length < sizeof(struct acpi_pct_register))
312 || (obj.buffer.pointer == NULL)) {
313 printk(KERN_ERR PREFIX "Invalid _PCT data (status_register)\n");
318 memcpy(&pr->performance->status_register, obj.buffer.pointer,
319 sizeof(struct acpi_pct_register));
322 kfree(buffer.pointer);
329 * Some AMDs have 50MHz frequency multiples, but only provide 100MHz rounding
330 * in their ACPI data. Calculate the real values and fix up the _PSS data.
332 static void amd_fixup_frequency(struct acpi_processor_px *px, int i)
334 u32 hi, lo, fid, did;
335 int index = px->control & 0x00000007;
337 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
340 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
341 || boot_cpu_data.x86 == 0x11) {
342 rdmsr(MSR_AMD_PSTATE_DEF_BASE + index, lo, hi);
345 * Bit 63: PstateEn. Read-write. If set, the P-state is valid.
352 if (boot_cpu_data.x86 == 0x10)
353 px->core_frequency = (100 * (fid + 0x10)) >> did;
355 px->core_frequency = (100 * (fid + 8)) >> did;
359 static void amd_fixup_frequency(struct acpi_processor_px *px, int i) {};
362 static int acpi_processor_get_performance_states(struct acpi_processor *pr)
365 acpi_status status = AE_OK;
366 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
367 struct acpi_buffer format = { sizeof("NNNNNN"), "NNNNNN" };
368 struct acpi_buffer state = { 0, NULL };
369 union acpi_object *pss = NULL;
371 int last_invalid = -1;
374 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
375 if (ACPI_FAILURE(status)) {
376 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PSS"));
380 pss = buffer.pointer;
381 if (!pss || (pss->type != ACPI_TYPE_PACKAGE)) {
382 printk(KERN_ERR PREFIX "Invalid _PSS data\n");
387 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d performance states\n",
388 pss->package.count));
390 pr->performance->state_count = pss->package.count;
391 pr->performance->states =
392 kmalloc(sizeof(struct acpi_processor_px) * pss->package.count,
394 if (!pr->performance->states) {
399 for (i = 0; i < pr->performance->state_count; i++) {
401 struct acpi_processor_px *px = &(pr->performance->states[i]);
403 state.length = sizeof(struct acpi_processor_px);
406 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Extracting state %d\n", i));
408 status = acpi_extract_package(&(pss->package.elements[i]),
410 if (ACPI_FAILURE(status)) {
411 ACPI_EXCEPTION((AE_INFO, status, "Invalid _PSS data"));
413 kfree(pr->performance->states);
417 amd_fixup_frequency(px, i);
419 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
420 "State [%d]: core_frequency[%d] power[%d] transition_latency[%d] bus_master_latency[%d] control[0x%x] status[0x%x]\n",
422 (u32) px->core_frequency,
424 (u32) px->transition_latency,
425 (u32) px->bus_master_latency,
426 (u32) px->control, (u32) px->status));
429 * Check that ACPI's u64 MHz will be valid as u32 KHz in cpufreq
431 if (!px->core_frequency ||
432 ((u32)(px->core_frequency * 1000) !=
433 (px->core_frequency * 1000))) {
434 printk(KERN_ERR FW_BUG PREFIX
435 "Invalid BIOS _PSS frequency found for processor %d: 0x%llx MHz\n",
436 pr->id, px->core_frequency);
437 if (last_invalid == -1)
440 if (last_invalid != -1) {
442 * Copy this valid entry over last_invalid entry
444 memcpy(&(pr->performance->states[last_invalid]),
445 px, sizeof(struct acpi_processor_px));
451 if (last_invalid == 0) {
452 printk(KERN_ERR FW_BUG PREFIX
453 "No valid BIOS _PSS frequency found for processor %d\n", pr->id);
455 kfree(pr->performance->states);
456 pr->performance->states = NULL;
459 if (last_invalid > 0)
460 pr->performance->state_count = last_invalid;
463 kfree(buffer.pointer);
468 static int acpi_processor_get_performance_info(struct acpi_processor *pr)
471 acpi_status status = AE_OK;
472 acpi_handle handle = NULL;
474 if (!pr || !pr->performance || !pr->handle)
477 status = acpi_get_handle(pr->handle, "_PCT", &handle);
478 if (ACPI_FAILURE(status)) {
479 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
480 "ACPI-based processor performance control unavailable\n"));
484 result = acpi_processor_get_performance_control(pr);
488 result = acpi_processor_get_performance_states(pr);
492 /* We need to call _PPC once when cpufreq starts */
494 result = acpi_processor_get_platform_limit(pr);
499 * Having _PPC but missing frequencies (_PSS, _PCT) is a very good hint that
500 * the BIOS is older than the CPU and does not know its frequencies
504 if (ACPI_SUCCESS(acpi_get_handle(pr->handle, "_PPC", &handle))){
505 if(boot_cpu_has(X86_FEATURE_EST))
506 printk(KERN_WARNING FW_BUG "BIOS needs update for CPU "
507 "frequency support\n");
513 int acpi_processor_notify_smm(struct module *calling_module)
516 static int is_done = 0;
519 if (!(acpi_processor_ppc_status & PPC_REGISTERED))
522 if (!try_module_get(calling_module))
525 /* is_done is set to negative if an error occurred,
526 * and to postitive if _no_ error occurred, but SMM
527 * was already notified. This avoids double notification
528 * which might lead to unexpected results...
531 module_put(calling_module);
533 } else if (is_done < 0) {
534 module_put(calling_module);
540 /* Can't write pstate_control to smi_command if either value is zero */
541 if ((!acpi_gbl_FADT.smi_command) || (!acpi_gbl_FADT.pstate_control)) {
542 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No SMI port or pstate_control\n"));
543 module_put(calling_module);
547 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
548 "Writing pstate_control [0x%x] to smi_command [0x%x]\n",
549 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command));
551 status = acpi_os_write_port(acpi_gbl_FADT.smi_command,
552 (u32) acpi_gbl_FADT.pstate_control, 8);
553 if (ACPI_FAILURE(status)) {
554 ACPI_EXCEPTION((AE_INFO, status,
555 "Failed to write pstate_control [0x%x] to "
556 "smi_command [0x%x]", acpi_gbl_FADT.pstate_control,
557 acpi_gbl_FADT.smi_command));
558 module_put(calling_module);
562 /* Success. If there's no _PPC, we need to fear nothing, so
563 * we can allow the cpufreq driver to be rmmod'ed. */
566 if (!(acpi_processor_ppc_status & PPC_IN_USE))
567 module_put(calling_module);
572 EXPORT_SYMBOL(acpi_processor_notify_smm);
574 static int acpi_processor_get_psd(struct acpi_processor *pr)
577 acpi_status status = AE_OK;
578 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
579 struct acpi_buffer format = {sizeof("NNNNN"), "NNNNN"};
580 struct acpi_buffer state = {0, NULL};
581 union acpi_object *psd = NULL;
582 struct acpi_psd_package *pdomain;
584 status = acpi_evaluate_object(pr->handle, "_PSD", NULL, &buffer);
585 if (ACPI_FAILURE(status)) {
589 psd = buffer.pointer;
590 if (!psd || (psd->type != ACPI_TYPE_PACKAGE)) {
591 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
596 if (psd->package.count != 1) {
597 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
602 pdomain = &(pr->performance->domain_info);
604 state.length = sizeof(struct acpi_psd_package);
605 state.pointer = pdomain;
607 status = acpi_extract_package(&(psd->package.elements[0]),
609 if (ACPI_FAILURE(status)) {
610 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
615 if (pdomain->num_entries != ACPI_PSD_REV0_ENTRIES) {
616 printk(KERN_ERR PREFIX "Unknown _PSD:num_entries\n");
621 if (pdomain->revision != ACPI_PSD_REV0_REVISION) {
622 printk(KERN_ERR PREFIX "Unknown _PSD:revision\n");
627 if (pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ALL &&
628 pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ANY &&
629 pdomain->coord_type != DOMAIN_COORD_TYPE_HW_ALL) {
630 printk(KERN_ERR PREFIX "Invalid _PSD:coord_type\n");
635 kfree(buffer.pointer);
639 int acpi_processor_preregister_performance(
640 struct acpi_processor_performance __percpu *performance)
642 int count, count_target;
645 cpumask_var_t covered_cpus;
646 struct acpi_processor *pr;
647 struct acpi_psd_package *pdomain;
648 struct acpi_processor *match_pr;
649 struct acpi_psd_package *match_pdomain;
651 if (!zalloc_cpumask_var(&covered_cpus, GFP_KERNEL))
654 mutex_lock(&performance_mutex);
657 * Check if another driver has already registered, and abort before
658 * changing pr->performance if it has. Check input data as well.
660 for_each_possible_cpu(i) {
661 pr = per_cpu(processors, i);
663 /* Look only at processors in ACPI namespace */
667 if (pr->performance) {
672 if (!performance || !per_cpu_ptr(performance, i)) {
678 /* Call _PSD for all CPUs */
679 for_each_possible_cpu(i) {
680 pr = per_cpu(processors, i);
684 pr->performance = per_cpu_ptr(performance, i);
685 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
686 if (acpi_processor_get_psd(pr)) {
695 * Now that we have _PSD data from all CPUs, lets setup P-state
698 for_each_possible_cpu(i) {
699 pr = per_cpu(processors, i);
703 if (cpumask_test_cpu(i, covered_cpus))
706 pdomain = &(pr->performance->domain_info);
707 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
708 cpumask_set_cpu(i, covered_cpus);
709 if (pdomain->num_processors <= 1)
712 /* Validate the Domain info */
713 count_target = pdomain->num_processors;
715 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
716 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
717 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
718 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_HW;
719 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
720 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ANY;
722 for_each_possible_cpu(j) {
726 match_pr = per_cpu(processors, j);
730 match_pdomain = &(match_pr->performance->domain_info);
731 if (match_pdomain->domain != pdomain->domain)
734 /* Here i and j are in the same domain */
736 if (match_pdomain->num_processors != count_target) {
741 if (pdomain->coord_type != match_pdomain->coord_type) {
746 cpumask_set_cpu(j, covered_cpus);
747 cpumask_set_cpu(j, pr->performance->shared_cpu_map);
751 for_each_possible_cpu(j) {
755 match_pr = per_cpu(processors, j);
759 match_pdomain = &(match_pr->performance->domain_info);
760 if (match_pdomain->domain != pdomain->domain)
763 match_pr->performance->shared_type =
764 pr->performance->shared_type;
765 cpumask_copy(match_pr->performance->shared_cpu_map,
766 pr->performance->shared_cpu_map);
771 for_each_possible_cpu(i) {
772 pr = per_cpu(processors, i);
773 if (!pr || !pr->performance)
776 /* Assume no coordination on any error parsing domain info */
778 cpumask_clear(pr->performance->shared_cpu_map);
779 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
780 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
782 pr->performance = NULL; /* Will be set for real in register */
786 mutex_unlock(&performance_mutex);
787 free_cpumask_var(covered_cpus);
790 EXPORT_SYMBOL(acpi_processor_preregister_performance);
793 acpi_processor_register_performance(struct acpi_processor_performance
794 *performance, unsigned int cpu)
796 struct acpi_processor *pr;
798 if (!(acpi_processor_ppc_status & PPC_REGISTERED))
801 mutex_lock(&performance_mutex);
803 pr = per_cpu(processors, cpu);
805 mutex_unlock(&performance_mutex);
809 if (pr->performance) {
810 mutex_unlock(&performance_mutex);
814 WARN_ON(!performance);
816 pr->performance = performance;
818 if (acpi_processor_get_performance_info(pr)) {
819 pr->performance = NULL;
820 mutex_unlock(&performance_mutex);
824 mutex_unlock(&performance_mutex);
828 EXPORT_SYMBOL(acpi_processor_register_performance);
831 acpi_processor_unregister_performance(struct acpi_processor_performance
832 *performance, unsigned int cpu)
834 struct acpi_processor *pr;
836 mutex_lock(&performance_mutex);
838 pr = per_cpu(processors, cpu);
840 mutex_unlock(&performance_mutex);
845 kfree(pr->performance->states);
846 pr->performance = NULL;
848 mutex_unlock(&performance_mutex);
853 EXPORT_SYMBOL(acpi_processor_unregister_performance);