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>
34 #include <linux/acpi.h>
35 #include <acpi/processor.h>
37 #include <asm/cpufeature.h>
40 #define PREFIX "ACPI: "
42 #define ACPI_PROCESSOR_CLASS "processor"
43 #define ACPI_PROCESSOR_FILE_PERFORMANCE "performance"
44 #define _COMPONENT ACPI_PROCESSOR_COMPONENT
45 ACPI_MODULE_NAME("processor_perflib");
47 static DEFINE_MUTEX(performance_mutex);
50 * _PPC support is implemented as a CPUfreq policy notifier:
51 * This means each time a CPUfreq driver registered also with
52 * the ACPI core is asked to change the speed policy, the maximum
53 * value is adjusted so that it is within the platform limit.
55 * Also, when a new platform limit value is detected, the CPUfreq
56 * policy is adjusted accordingly.
60 * -1 -> cpufreq low level drivers not initialized -> _PSS, etc. not called yet
62 * 0 -> cpufreq low level drivers initialized -> consider _PPC values
63 * 1 -> ignore _PPC totally -> forced by user through boot param
65 static int ignore_ppc = -1;
66 module_param(ignore_ppc, int, 0644);
67 MODULE_PARM_DESC(ignore_ppc, "If the frequency of your machine gets wrongly" \
68 "limited by BIOS, this should help");
70 #define PPC_REGISTERED 1
73 static int acpi_processor_ppc_status;
75 static int acpi_processor_ppc_notifier(struct notifier_block *nb,
76 unsigned long event, void *data)
78 struct cpufreq_policy *policy = data;
79 struct acpi_processor *pr;
82 if (event == CPUFREQ_START && ignore_ppc <= 0) {
90 if (event != CPUFREQ_INCOMPATIBLE)
93 mutex_lock(&performance_mutex);
95 pr = per_cpu(processors, policy->cpu);
96 if (!pr || !pr->performance)
99 ppc = (unsigned int)pr->performance_platform_limit;
101 if (ppc >= pr->performance->state_count)
104 cpufreq_verify_within_limits(policy, 0,
105 pr->performance->states[ppc].
106 core_frequency * 1000);
109 mutex_unlock(&performance_mutex);
114 static struct notifier_block acpi_ppc_notifier_block = {
115 .notifier_call = acpi_processor_ppc_notifier,
118 static int acpi_processor_get_platform_limit(struct acpi_processor *pr)
120 acpi_status status = 0;
121 unsigned long long ppc = 0;
128 * _PPC indicates the maximum state currently supported by the platform
129 * (e.g. 0 = states 0..n; 1 = states 1..n; etc.
131 status = acpi_evaluate_integer(pr->handle, "_PPC", NULL, &ppc);
133 if (status != AE_NOT_FOUND)
134 acpi_processor_ppc_status |= PPC_IN_USE;
136 if (ACPI_FAILURE(status) && status != AE_NOT_FOUND) {
137 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PPC"));
141 pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr->id,
142 (int)ppc, ppc ? "" : "not");
144 pr->performance_platform_limit = (int)ppc;
149 #define ACPI_PROCESSOR_NOTIFY_PERFORMANCE 0x80
151 * acpi_processor_ppc_ost: Notify firmware the _PPC evaluation status
152 * @handle: ACPI processor handle
153 * @status: the status code of _PPC evaluation
154 * 0: success. OSPM is now using the performance state specificed.
155 * 1: failure. OSPM has not changed the number of P-states in use
157 static void acpi_processor_ppc_ost(acpi_handle handle, int status)
159 union acpi_object params[2] = {
160 {.type = ACPI_TYPE_INTEGER,},
161 {.type = ACPI_TYPE_INTEGER,},
163 struct acpi_object_list arg_list = {2, params};
165 if (acpi_has_method(handle, "_OST")) {
166 params[0].integer.value = ACPI_PROCESSOR_NOTIFY_PERFORMANCE;
167 params[1].integer.value = status;
168 acpi_evaluate_object(handle, "_OST", &arg_list, NULL);
172 int acpi_processor_ppc_has_changed(struct acpi_processor *pr, int event_flag)
178 * Only when it is notification event, the _OST object
179 * will be evaluated. Otherwise it is skipped.
182 acpi_processor_ppc_ost(pr->handle, 1);
186 ret = acpi_processor_get_platform_limit(pr);
188 * Only when it is notification event, the _OST object
189 * will be evaluated. Otherwise it is skipped.
193 acpi_processor_ppc_ost(pr->handle, 1);
195 acpi_processor_ppc_ost(pr->handle, 0);
200 return cpufreq_update_policy(pr->id);
203 int acpi_processor_get_bios_limit(int cpu, unsigned int *limit)
205 struct acpi_processor *pr;
207 pr = per_cpu(processors, cpu);
208 if (!pr || !pr->performance || !pr->performance->state_count)
210 *limit = pr->performance->states[pr->performance_platform_limit].
211 core_frequency * 1000;
214 EXPORT_SYMBOL(acpi_processor_get_bios_limit);
216 void acpi_processor_ppc_init(void)
218 if (!cpufreq_register_notifier
219 (&acpi_ppc_notifier_block, CPUFREQ_POLICY_NOTIFIER))
220 acpi_processor_ppc_status |= PPC_REGISTERED;
223 "Warning: Processor Platform Limit not supported.\n");
226 void acpi_processor_ppc_exit(void)
228 if (acpi_processor_ppc_status & PPC_REGISTERED)
229 cpufreq_unregister_notifier(&acpi_ppc_notifier_block,
230 CPUFREQ_POLICY_NOTIFIER);
232 acpi_processor_ppc_status &= ~PPC_REGISTERED;
235 static int acpi_processor_get_performance_control(struct acpi_processor *pr)
238 acpi_status status = 0;
239 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
240 union acpi_object *pct = NULL;
241 union acpi_object obj = { 0 };
244 status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
245 if (ACPI_FAILURE(status)) {
246 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PCT"));
250 pct = (union acpi_object *)buffer.pointer;
251 if (!pct || (pct->type != ACPI_TYPE_PACKAGE)
252 || (pct->package.count != 2)) {
253 printk(KERN_ERR PREFIX "Invalid _PCT data\n");
262 obj = pct->package.elements[0];
264 if ((obj.type != ACPI_TYPE_BUFFER)
265 || (obj.buffer.length < sizeof(struct acpi_pct_register))
266 || (obj.buffer.pointer == NULL)) {
267 printk(KERN_ERR PREFIX "Invalid _PCT data (control_register)\n");
271 memcpy(&pr->performance->control_register, obj.buffer.pointer,
272 sizeof(struct acpi_pct_register));
278 obj = pct->package.elements[1];
280 if ((obj.type != ACPI_TYPE_BUFFER)
281 || (obj.buffer.length < sizeof(struct acpi_pct_register))
282 || (obj.buffer.pointer == NULL)) {
283 printk(KERN_ERR PREFIX "Invalid _PCT data (status_register)\n");
288 memcpy(&pr->performance->status_register, obj.buffer.pointer,
289 sizeof(struct acpi_pct_register));
292 kfree(buffer.pointer);
299 * Some AMDs have 50MHz frequency multiples, but only provide 100MHz rounding
300 * in their ACPI data. Calculate the real values and fix up the _PSS data.
302 static void amd_fixup_frequency(struct acpi_processor_px *px, int i)
304 u32 hi, lo, fid, did;
305 int index = px->control & 0x00000007;
307 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
310 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
311 || boot_cpu_data.x86 == 0x11) {
312 rdmsr(MSR_AMD_PSTATE_DEF_BASE + index, lo, hi);
315 * Bit 63: PstateEn. Read-write. If set, the P-state is valid.
322 if (boot_cpu_data.x86 == 0x10)
323 px->core_frequency = (100 * (fid + 0x10)) >> did;
325 px->core_frequency = (100 * (fid + 8)) >> did;
329 static void amd_fixup_frequency(struct acpi_processor_px *px, int i) {};
332 static int acpi_processor_get_performance_states(struct acpi_processor *pr)
335 acpi_status status = AE_OK;
336 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
337 struct acpi_buffer format = { sizeof("NNNNNN"), "NNNNNN" };
338 struct acpi_buffer state = { 0, NULL };
339 union acpi_object *pss = NULL;
341 int last_invalid = -1;
344 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
345 if (ACPI_FAILURE(status)) {
346 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PSS"));
350 pss = buffer.pointer;
351 if (!pss || (pss->type != ACPI_TYPE_PACKAGE)) {
352 printk(KERN_ERR PREFIX "Invalid _PSS data\n");
357 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d performance states\n",
358 pss->package.count));
360 pr->performance->state_count = pss->package.count;
361 pr->performance->states =
362 kmalloc(sizeof(struct acpi_processor_px) * pss->package.count,
364 if (!pr->performance->states) {
369 for (i = 0; i < pr->performance->state_count; i++) {
371 struct acpi_processor_px *px = &(pr->performance->states[i]);
373 state.length = sizeof(struct acpi_processor_px);
376 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Extracting state %d\n", i));
378 status = acpi_extract_package(&(pss->package.elements[i]),
380 if (ACPI_FAILURE(status)) {
381 ACPI_EXCEPTION((AE_INFO, status, "Invalid _PSS data"));
383 kfree(pr->performance->states);
387 amd_fixup_frequency(px, i);
389 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
390 "State [%d]: core_frequency[%d] power[%d] transition_latency[%d] bus_master_latency[%d] control[0x%x] status[0x%x]\n",
392 (u32) px->core_frequency,
394 (u32) px->transition_latency,
395 (u32) px->bus_master_latency,
396 (u32) px->control, (u32) px->status));
399 * Check that ACPI's u64 MHz will be valid as u32 KHz in cpufreq
401 if (!px->core_frequency ||
402 ((u32)(px->core_frequency * 1000) !=
403 (px->core_frequency * 1000))) {
404 printk(KERN_ERR FW_BUG PREFIX
405 "Invalid BIOS _PSS frequency found for processor %d: 0x%llx MHz\n",
406 pr->id, px->core_frequency);
407 if (last_invalid == -1)
410 if (last_invalid != -1) {
412 * Copy this valid entry over last_invalid entry
414 memcpy(&(pr->performance->states[last_invalid]),
415 px, sizeof(struct acpi_processor_px));
421 if (last_invalid == 0) {
422 printk(KERN_ERR FW_BUG PREFIX
423 "No valid BIOS _PSS frequency found for processor %d\n", pr->id);
425 kfree(pr->performance->states);
426 pr->performance->states = NULL;
429 if (last_invalid > 0)
430 pr->performance->state_count = last_invalid;
433 kfree(buffer.pointer);
438 int acpi_processor_get_performance_info(struct acpi_processor *pr)
442 if (!pr || !pr->performance || !pr->handle)
445 if (!acpi_has_method(pr->handle, "_PCT")) {
446 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
447 "ACPI-based processor performance control unavailable\n"));
451 result = acpi_processor_get_performance_control(pr);
455 result = acpi_processor_get_performance_states(pr);
459 /* We need to call _PPC once when cpufreq starts */
461 result = acpi_processor_get_platform_limit(pr);
466 * Having _PPC but missing frequencies (_PSS, _PCT) is a very good hint that
467 * the BIOS is older than the CPU and does not know its frequencies
471 if (acpi_has_method(pr->handle, "_PPC")) {
472 if(boot_cpu_has(X86_FEATURE_EST))
473 printk(KERN_WARNING FW_BUG "BIOS needs update for CPU "
474 "frequency support\n");
479 EXPORT_SYMBOL_GPL(acpi_processor_get_performance_info);
480 int acpi_processor_notify_smm(struct module *calling_module)
483 static int is_done = 0;
486 if (!(acpi_processor_ppc_status & PPC_REGISTERED))
489 if (!try_module_get(calling_module))
492 /* is_done is set to negative if an error occurred,
493 * and to postitive if _no_ error occurred, but SMM
494 * was already notified. This avoids double notification
495 * which might lead to unexpected results...
498 module_put(calling_module);
500 } else if (is_done < 0) {
501 module_put(calling_module);
507 /* Can't write pstate_control to smi_command if either value is zero */
508 if ((!acpi_gbl_FADT.smi_command) || (!acpi_gbl_FADT.pstate_control)) {
509 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No SMI port or pstate_control\n"));
510 module_put(calling_module);
514 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
515 "Writing pstate_control [0x%x] to smi_command [0x%x]\n",
516 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command));
518 status = acpi_os_write_port(acpi_gbl_FADT.smi_command,
519 (u32) acpi_gbl_FADT.pstate_control, 8);
520 if (ACPI_FAILURE(status)) {
521 ACPI_EXCEPTION((AE_INFO, status,
522 "Failed to write pstate_control [0x%x] to "
523 "smi_command [0x%x]", acpi_gbl_FADT.pstate_control,
524 acpi_gbl_FADT.smi_command));
525 module_put(calling_module);
529 /* Success. If there's no _PPC, we need to fear nothing, so
530 * we can allow the cpufreq driver to be rmmod'ed. */
533 if (!(acpi_processor_ppc_status & PPC_IN_USE))
534 module_put(calling_module);
539 EXPORT_SYMBOL(acpi_processor_notify_smm);
541 static int acpi_processor_get_psd(struct acpi_processor *pr)
544 acpi_status status = AE_OK;
545 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
546 struct acpi_buffer format = {sizeof("NNNNN"), "NNNNN"};
547 struct acpi_buffer state = {0, NULL};
548 union acpi_object *psd = NULL;
549 struct acpi_psd_package *pdomain;
551 status = acpi_evaluate_object(pr->handle, "_PSD", NULL, &buffer);
552 if (ACPI_FAILURE(status)) {
556 psd = buffer.pointer;
557 if (!psd || (psd->type != ACPI_TYPE_PACKAGE)) {
558 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
563 if (psd->package.count != 1) {
564 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
569 pdomain = &(pr->performance->domain_info);
571 state.length = sizeof(struct acpi_psd_package);
572 state.pointer = pdomain;
574 status = acpi_extract_package(&(psd->package.elements[0]),
576 if (ACPI_FAILURE(status)) {
577 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
582 if (pdomain->num_entries != ACPI_PSD_REV0_ENTRIES) {
583 printk(KERN_ERR PREFIX "Unknown _PSD:num_entries\n");
588 if (pdomain->revision != ACPI_PSD_REV0_REVISION) {
589 printk(KERN_ERR PREFIX "Unknown _PSD:revision\n");
594 if (pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ALL &&
595 pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ANY &&
596 pdomain->coord_type != DOMAIN_COORD_TYPE_HW_ALL) {
597 printk(KERN_ERR PREFIX "Invalid _PSD:coord_type\n");
602 kfree(buffer.pointer);
606 int acpi_processor_preregister_performance(
607 struct acpi_processor_performance __percpu *performance)
612 cpumask_var_t covered_cpus;
613 struct acpi_processor *pr;
614 struct acpi_psd_package *pdomain;
615 struct acpi_processor *match_pr;
616 struct acpi_psd_package *match_pdomain;
618 if (!zalloc_cpumask_var(&covered_cpus, GFP_KERNEL))
621 mutex_lock(&performance_mutex);
624 * Check if another driver has already registered, and abort before
625 * changing pr->performance if it has. Check input data as well.
627 for_each_possible_cpu(i) {
628 pr = per_cpu(processors, i);
630 /* Look only at processors in ACPI namespace */
634 if (pr->performance) {
639 if (!performance || !per_cpu_ptr(performance, i)) {
645 /* Call _PSD for all CPUs */
646 for_each_possible_cpu(i) {
647 pr = per_cpu(processors, i);
651 pr->performance = per_cpu_ptr(performance, i);
652 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
653 if (acpi_processor_get_psd(pr)) {
662 * Now that we have _PSD data from all CPUs, lets setup P-state
665 for_each_possible_cpu(i) {
666 pr = per_cpu(processors, i);
670 if (cpumask_test_cpu(i, covered_cpus))
673 pdomain = &(pr->performance->domain_info);
674 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
675 cpumask_set_cpu(i, covered_cpus);
676 if (pdomain->num_processors <= 1)
679 /* Validate the Domain info */
680 count_target = pdomain->num_processors;
681 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
682 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
683 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
684 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_HW;
685 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
686 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ANY;
688 for_each_possible_cpu(j) {
692 match_pr = per_cpu(processors, j);
696 match_pdomain = &(match_pr->performance->domain_info);
697 if (match_pdomain->domain != pdomain->domain)
700 /* Here i and j are in the same domain */
702 if (match_pdomain->num_processors != count_target) {
707 if (pdomain->coord_type != match_pdomain->coord_type) {
712 cpumask_set_cpu(j, covered_cpus);
713 cpumask_set_cpu(j, pr->performance->shared_cpu_map);
716 for_each_possible_cpu(j) {
720 match_pr = per_cpu(processors, j);
724 match_pdomain = &(match_pr->performance->domain_info);
725 if (match_pdomain->domain != pdomain->domain)
728 match_pr->performance->shared_type =
729 pr->performance->shared_type;
730 cpumask_copy(match_pr->performance->shared_cpu_map,
731 pr->performance->shared_cpu_map);
736 for_each_possible_cpu(i) {
737 pr = per_cpu(processors, i);
738 if (!pr || !pr->performance)
741 /* Assume no coordination on any error parsing domain info */
743 cpumask_clear(pr->performance->shared_cpu_map);
744 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
745 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
747 pr->performance = NULL; /* Will be set for real in register */
751 mutex_unlock(&performance_mutex);
752 free_cpumask_var(covered_cpus);
755 EXPORT_SYMBOL(acpi_processor_preregister_performance);
758 acpi_processor_register_performance(struct acpi_processor_performance
759 *performance, unsigned int cpu)
761 struct acpi_processor *pr;
763 if (!(acpi_processor_ppc_status & PPC_REGISTERED))
766 mutex_lock(&performance_mutex);
768 pr = per_cpu(processors, cpu);
770 mutex_unlock(&performance_mutex);
774 if (pr->performance) {
775 mutex_unlock(&performance_mutex);
779 WARN_ON(!performance);
781 pr->performance = performance;
783 if (acpi_processor_get_performance_info(pr)) {
784 pr->performance = NULL;
785 mutex_unlock(&performance_mutex);
789 mutex_unlock(&performance_mutex);
793 EXPORT_SYMBOL(acpi_processor_register_performance);
796 acpi_processor_unregister_performance(struct acpi_processor_performance
797 *performance, unsigned int cpu)
799 struct acpi_processor *pr;
801 mutex_lock(&performance_mutex);
803 pr = per_cpu(processors, cpu);
805 mutex_unlock(&performance_mutex);
810 kfree(pr->performance->states);
811 pr->performance = NULL;
813 mutex_unlock(&performance_mutex);
818 EXPORT_SYMBOL(acpi_processor_unregister_performance);