2 * processor_perflib.c - ACPI Processor P-States Library ($Revision: 71 $)
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;
238 static int acpi_processor_get_performance_control(struct acpi_processor *pr)
241 acpi_status status = 0;
242 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
243 union acpi_object *pct = NULL;
244 union acpi_object obj = { 0 };
247 status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
248 if (ACPI_FAILURE(status)) {
249 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PCT"));
253 pct = (union acpi_object *)buffer.pointer;
254 if (!pct || (pct->type != ACPI_TYPE_PACKAGE)
255 || (pct->package.count != 2)) {
256 printk(KERN_ERR PREFIX "Invalid _PCT data\n");
265 obj = pct->package.elements[0];
267 if ((obj.type != ACPI_TYPE_BUFFER)
268 || (obj.buffer.length < sizeof(struct acpi_pct_register))
269 || (obj.buffer.pointer == NULL)) {
270 printk(KERN_ERR PREFIX "Invalid _PCT data (control_register)\n");
274 memcpy(&pr->performance->control_register, obj.buffer.pointer,
275 sizeof(struct acpi_pct_register));
281 obj = pct->package.elements[1];
283 if ((obj.type != ACPI_TYPE_BUFFER)
284 || (obj.buffer.length < sizeof(struct acpi_pct_register))
285 || (obj.buffer.pointer == NULL)) {
286 printk(KERN_ERR PREFIX "Invalid _PCT data (status_register)\n");
291 memcpy(&pr->performance->status_register, obj.buffer.pointer,
292 sizeof(struct acpi_pct_register));
295 kfree(buffer.pointer);
302 * Some AMDs have 50MHz frequency multiples, but only provide 100MHz rounding
303 * in their ACPI data. Calculate the real values and fix up the _PSS data.
305 static void amd_fixup_frequency(struct acpi_processor_px *px, int i)
307 u32 hi, lo, fid, did;
308 int index = px->control & 0x00000007;
310 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
313 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
314 || boot_cpu_data.x86 == 0x11) {
315 rdmsr(MSR_AMD_PSTATE_DEF_BASE + index, lo, hi);
318 * Bit 63: PstateEn. Read-write. If set, the P-state is valid.
325 if (boot_cpu_data.x86 == 0x10)
326 px->core_frequency = (100 * (fid + 0x10)) >> did;
328 px->core_frequency = (100 * (fid + 8)) >> did;
332 static void amd_fixup_frequency(struct acpi_processor_px *px, int i) {};
335 static int acpi_processor_get_performance_states(struct acpi_processor *pr)
338 acpi_status status = AE_OK;
339 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
340 struct acpi_buffer format = { sizeof("NNNNNN"), "NNNNNN" };
341 struct acpi_buffer state = { 0, NULL };
342 union acpi_object *pss = NULL;
344 int last_invalid = -1;
347 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
348 if (ACPI_FAILURE(status)) {
349 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PSS"));
353 pss = buffer.pointer;
354 if (!pss || (pss->type != ACPI_TYPE_PACKAGE)) {
355 printk(KERN_ERR PREFIX "Invalid _PSS data\n");
360 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d performance states\n",
361 pss->package.count));
363 pr->performance->state_count = pss->package.count;
364 pr->performance->states =
365 kmalloc(sizeof(struct acpi_processor_px) * pss->package.count,
367 if (!pr->performance->states) {
372 for (i = 0; i < pr->performance->state_count; i++) {
374 struct acpi_processor_px *px = &(pr->performance->states[i]);
376 state.length = sizeof(struct acpi_processor_px);
379 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Extracting state %d\n", i));
381 status = acpi_extract_package(&(pss->package.elements[i]),
383 if (ACPI_FAILURE(status)) {
384 ACPI_EXCEPTION((AE_INFO, status, "Invalid _PSS data"));
386 kfree(pr->performance->states);
390 amd_fixup_frequency(px, i);
392 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
393 "State [%d]: core_frequency[%d] power[%d] transition_latency[%d] bus_master_latency[%d] control[0x%x] status[0x%x]\n",
395 (u32) px->core_frequency,
397 (u32) px->transition_latency,
398 (u32) px->bus_master_latency,
399 (u32) px->control, (u32) px->status));
402 * Check that ACPI's u64 MHz will be valid as u32 KHz in cpufreq
404 if (!px->core_frequency ||
405 ((u32)(px->core_frequency * 1000) !=
406 (px->core_frequency * 1000))) {
407 printk(KERN_ERR FW_BUG PREFIX
408 "Invalid BIOS _PSS frequency found for processor %d: 0x%llx MHz\n",
409 pr->id, px->core_frequency);
410 if (last_invalid == -1)
413 if (last_invalid != -1) {
415 * Copy this valid entry over last_invalid entry
417 memcpy(&(pr->performance->states[last_invalid]),
418 px, sizeof(struct acpi_processor_px));
424 if (last_invalid == 0) {
425 printk(KERN_ERR FW_BUG PREFIX
426 "No valid BIOS _PSS frequency found for processor %d\n", pr->id);
428 kfree(pr->performance->states);
429 pr->performance->states = NULL;
432 if (last_invalid > 0)
433 pr->performance->state_count = last_invalid;
436 kfree(buffer.pointer);
441 int acpi_processor_get_performance_info(struct acpi_processor *pr)
445 if (!pr || !pr->performance || !pr->handle)
448 if (!acpi_has_method(pr->handle, "_PCT")) {
449 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
450 "ACPI-based processor performance control unavailable\n"));
454 result = acpi_processor_get_performance_control(pr);
458 result = acpi_processor_get_performance_states(pr);
462 /* We need to call _PPC once when cpufreq starts */
464 result = acpi_processor_get_platform_limit(pr);
469 * Having _PPC but missing frequencies (_PSS, _PCT) is a very good hint that
470 * the BIOS is older than the CPU and does not know its frequencies
474 if (acpi_has_method(pr->handle, "_PPC")) {
475 if(boot_cpu_has(X86_FEATURE_EST))
476 printk(KERN_WARNING FW_BUG "BIOS needs update for CPU "
477 "frequency support\n");
482 EXPORT_SYMBOL_GPL(acpi_processor_get_performance_info);
483 int acpi_processor_notify_smm(struct module *calling_module)
486 static int is_done = 0;
489 if (!(acpi_processor_ppc_status & PPC_REGISTERED))
492 if (!try_module_get(calling_module))
495 /* is_done is set to negative if an error occurred,
496 * and to postitive if _no_ error occurred, but SMM
497 * was already notified. This avoids double notification
498 * which might lead to unexpected results...
501 module_put(calling_module);
503 } else if (is_done < 0) {
504 module_put(calling_module);
510 /* Can't write pstate_control to smi_command if either value is zero */
511 if ((!acpi_gbl_FADT.smi_command) || (!acpi_gbl_FADT.pstate_control)) {
512 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No SMI port or pstate_control\n"));
513 module_put(calling_module);
517 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
518 "Writing pstate_control [0x%x] to smi_command [0x%x]\n",
519 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command));
521 status = acpi_os_write_port(acpi_gbl_FADT.smi_command,
522 (u32) acpi_gbl_FADT.pstate_control, 8);
523 if (ACPI_FAILURE(status)) {
524 ACPI_EXCEPTION((AE_INFO, status,
525 "Failed to write pstate_control [0x%x] to "
526 "smi_command [0x%x]", acpi_gbl_FADT.pstate_control,
527 acpi_gbl_FADT.smi_command));
528 module_put(calling_module);
532 /* Success. If there's no _PPC, we need to fear nothing, so
533 * we can allow the cpufreq driver to be rmmod'ed. */
536 if (!(acpi_processor_ppc_status & PPC_IN_USE))
537 module_put(calling_module);
542 EXPORT_SYMBOL(acpi_processor_notify_smm);
544 static int acpi_processor_get_psd(struct acpi_processor *pr)
547 acpi_status status = AE_OK;
548 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
549 struct acpi_buffer format = {sizeof("NNNNN"), "NNNNN"};
550 struct acpi_buffer state = {0, NULL};
551 union acpi_object *psd = NULL;
552 struct acpi_psd_package *pdomain;
554 status = acpi_evaluate_object(pr->handle, "_PSD", NULL, &buffer);
555 if (ACPI_FAILURE(status)) {
559 psd = buffer.pointer;
560 if (!psd || (psd->type != ACPI_TYPE_PACKAGE)) {
561 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
566 if (psd->package.count != 1) {
567 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
572 pdomain = &(pr->performance->domain_info);
574 state.length = sizeof(struct acpi_psd_package);
575 state.pointer = pdomain;
577 status = acpi_extract_package(&(psd->package.elements[0]),
579 if (ACPI_FAILURE(status)) {
580 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
585 if (pdomain->num_entries != ACPI_PSD_REV0_ENTRIES) {
586 printk(KERN_ERR PREFIX "Unknown _PSD:num_entries\n");
591 if (pdomain->revision != ACPI_PSD_REV0_REVISION) {
592 printk(KERN_ERR PREFIX "Unknown _PSD:revision\n");
597 if (pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ALL &&
598 pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ANY &&
599 pdomain->coord_type != DOMAIN_COORD_TYPE_HW_ALL) {
600 printk(KERN_ERR PREFIX "Invalid _PSD:coord_type\n");
605 kfree(buffer.pointer);
609 int acpi_processor_preregister_performance(
610 struct acpi_processor_performance __percpu *performance)
615 cpumask_var_t covered_cpus;
616 struct acpi_processor *pr;
617 struct acpi_psd_package *pdomain;
618 struct acpi_processor *match_pr;
619 struct acpi_psd_package *match_pdomain;
621 if (!zalloc_cpumask_var(&covered_cpus, GFP_KERNEL))
624 mutex_lock(&performance_mutex);
627 * Check if another driver has already registered, and abort before
628 * changing pr->performance if it has. Check input data as well.
630 for_each_possible_cpu(i) {
631 pr = per_cpu(processors, i);
633 /* Look only at processors in ACPI namespace */
637 if (pr->performance) {
642 if (!performance || !per_cpu_ptr(performance, i)) {
648 /* Call _PSD for all CPUs */
649 for_each_possible_cpu(i) {
650 pr = per_cpu(processors, i);
654 pr->performance = per_cpu_ptr(performance, i);
655 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
656 if (acpi_processor_get_psd(pr)) {
665 * Now that we have _PSD data from all CPUs, lets setup P-state
668 for_each_possible_cpu(i) {
669 pr = per_cpu(processors, i);
673 if (cpumask_test_cpu(i, covered_cpus))
676 pdomain = &(pr->performance->domain_info);
677 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
678 cpumask_set_cpu(i, covered_cpus);
679 if (pdomain->num_processors <= 1)
682 /* Validate the Domain info */
683 count_target = pdomain->num_processors;
684 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
685 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
686 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
687 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_HW;
688 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
689 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ANY;
691 for_each_possible_cpu(j) {
695 match_pr = per_cpu(processors, j);
699 match_pdomain = &(match_pr->performance->domain_info);
700 if (match_pdomain->domain != pdomain->domain)
703 /* Here i and j are in the same domain */
705 if (match_pdomain->num_processors != count_target) {
710 if (pdomain->coord_type != match_pdomain->coord_type) {
715 cpumask_set_cpu(j, covered_cpus);
716 cpumask_set_cpu(j, pr->performance->shared_cpu_map);
719 for_each_possible_cpu(j) {
723 match_pr = per_cpu(processors, j);
727 match_pdomain = &(match_pr->performance->domain_info);
728 if (match_pdomain->domain != pdomain->domain)
731 match_pr->performance->shared_type =
732 pr->performance->shared_type;
733 cpumask_copy(match_pr->performance->shared_cpu_map,
734 pr->performance->shared_cpu_map);
739 for_each_possible_cpu(i) {
740 pr = per_cpu(processors, i);
741 if (!pr || !pr->performance)
744 /* Assume no coordination on any error parsing domain info */
746 cpumask_clear(pr->performance->shared_cpu_map);
747 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
748 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
750 pr->performance = NULL; /* Will be set for real in register */
754 mutex_unlock(&performance_mutex);
755 free_cpumask_var(covered_cpus);
758 EXPORT_SYMBOL(acpi_processor_preregister_performance);
761 acpi_processor_register_performance(struct acpi_processor_performance
762 *performance, unsigned int cpu)
764 struct acpi_processor *pr;
766 if (!(acpi_processor_ppc_status & PPC_REGISTERED))
769 mutex_lock(&performance_mutex);
771 pr = per_cpu(processors, cpu);
773 mutex_unlock(&performance_mutex);
777 if (pr->performance) {
778 mutex_unlock(&performance_mutex);
782 WARN_ON(!performance);
784 pr->performance = performance;
786 if (acpi_processor_get_performance_info(pr)) {
787 pr->performance = NULL;
788 mutex_unlock(&performance_mutex);
792 mutex_unlock(&performance_mutex);
796 EXPORT_SYMBOL(acpi_processor_register_performance);
799 acpi_processor_unregister_performance(struct acpi_processor_performance
800 *performance, unsigned int cpu)
802 struct acpi_processor *pr;
804 mutex_lock(&performance_mutex);
806 pr = per_cpu(processors, cpu);
808 mutex_unlock(&performance_mutex);
813 kfree(pr->performance->states);
814 pr->performance = NULL;
816 mutex_unlock(&performance_mutex);
821 EXPORT_SYMBOL(acpi_processor_unregister_performance);