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.
25 #include <linux/kernel.h>
26 #include <linux/module.h>
27 #include <linux/init.h>
28 #include <linux/cpufreq.h>
29 #include <linux/slab.h>
30 #include <linux/acpi.h>
31 #include <acpi/processor.h>
33 #include <asm/cpufeature.h>
36 #define PREFIX "ACPI: "
38 #define ACPI_PROCESSOR_CLASS "processor"
39 #define ACPI_PROCESSOR_FILE_PERFORMANCE "performance"
40 #define _COMPONENT ACPI_PROCESSOR_COMPONENT
41 ACPI_MODULE_NAME("processor_perflib");
43 static DEFINE_MUTEX(performance_mutex
);
46 * _PPC support is implemented as a CPUfreq policy notifier:
47 * This means each time a CPUfreq driver registered also with
48 * the ACPI core is asked to change the speed policy, the maximum
49 * value is adjusted so that it is within the platform limit.
51 * Also, when a new platform limit value is detected, the CPUfreq
52 * policy is adjusted accordingly.
56 * -1 -> cpufreq low level drivers not initialized -> _PSS, etc. not called yet
58 * 0 -> cpufreq low level drivers initialized -> consider _PPC values
59 * 1 -> ignore _PPC totally -> forced by user through boot param
61 static int ignore_ppc
= -1;
62 module_param(ignore_ppc
, int, 0644);
63 MODULE_PARM_DESC(ignore_ppc
, "If the frequency of your machine gets wrongly" \
64 "limited by BIOS, this should help");
66 #define PPC_REGISTERED 1
69 static int acpi_processor_ppc_status
;
71 static int acpi_processor_ppc_notifier(struct notifier_block
*nb
,
72 unsigned long event
, void *data
)
74 struct cpufreq_policy
*policy
= data
;
75 struct acpi_processor
*pr
;
78 if (event
== CPUFREQ_START
&& ignore_ppc
<= 0) {
86 if (event
!= CPUFREQ_ADJUST
)
89 mutex_lock(&performance_mutex
);
91 pr
= per_cpu(processors
, policy
->cpu
);
92 if (!pr
|| !pr
->performance
)
95 ppc
= (unsigned int)pr
->performance_platform_limit
;
97 if (ppc
>= pr
->performance
->state_count
)
100 cpufreq_verify_within_limits(policy
, 0,
101 pr
->performance
->states
[ppc
].
102 core_frequency
* 1000);
105 mutex_unlock(&performance_mutex
);
110 static struct notifier_block acpi_ppc_notifier_block
= {
111 .notifier_call
= acpi_processor_ppc_notifier
,
114 static int acpi_processor_get_platform_limit(struct acpi_processor
*pr
)
116 acpi_status status
= 0;
117 unsigned long long ppc
= 0;
124 * _PPC indicates the maximum state currently supported by the platform
125 * (e.g. 0 = states 0..n; 1 = states 1..n; etc.
127 status
= acpi_evaluate_integer(pr
->handle
, "_PPC", NULL
, &ppc
);
129 if (status
!= AE_NOT_FOUND
)
130 acpi_processor_ppc_status
|= PPC_IN_USE
;
132 if (ACPI_FAILURE(status
) && status
!= AE_NOT_FOUND
) {
133 ACPI_EXCEPTION((AE_INFO
, status
, "Evaluating _PPC"));
137 pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr
->id
,
138 (int)ppc
, ppc
? "" : "not");
140 pr
->performance_platform_limit
= (int)ppc
;
145 #define ACPI_PROCESSOR_NOTIFY_PERFORMANCE 0x80
147 * acpi_processor_ppc_ost: Notify firmware the _PPC evaluation status
148 * @handle: ACPI processor handle
149 * @status: the status code of _PPC evaluation
150 * 0: success. OSPM is now using the performance state specificed.
151 * 1: failure. OSPM has not changed the number of P-states in use
153 static void acpi_processor_ppc_ost(acpi_handle handle
, int status
)
155 if (acpi_has_method(handle
, "_OST"))
156 acpi_evaluate_ost(handle
, ACPI_PROCESSOR_NOTIFY_PERFORMANCE
,
160 int acpi_processor_ppc_has_changed(struct acpi_processor
*pr
, int event_flag
)
166 * Only when it is notification event, the _OST object
167 * will be evaluated. Otherwise it is skipped.
170 acpi_processor_ppc_ost(pr
->handle
, 1);
174 ret
= acpi_processor_get_platform_limit(pr
);
176 * Only when it is notification event, the _OST object
177 * will be evaluated. Otherwise it is skipped.
181 acpi_processor_ppc_ost(pr
->handle
, 1);
183 acpi_processor_ppc_ost(pr
->handle
, 0);
188 return cpufreq_update_policy(pr
->id
);
191 int acpi_processor_get_bios_limit(int cpu
, unsigned int *limit
)
193 struct acpi_processor
*pr
;
195 pr
= per_cpu(processors
, cpu
);
196 if (!pr
|| !pr
->performance
|| !pr
->performance
->state_count
)
198 *limit
= pr
->performance
->states
[pr
->performance_platform_limit
].
199 core_frequency
* 1000;
202 EXPORT_SYMBOL(acpi_processor_get_bios_limit
);
204 void acpi_processor_ppc_init(void)
206 if (!cpufreq_register_notifier
207 (&acpi_ppc_notifier_block
, CPUFREQ_POLICY_NOTIFIER
))
208 acpi_processor_ppc_status
|= PPC_REGISTERED
;
211 "Warning: Processor Platform Limit not supported.\n");
214 void acpi_processor_ppc_exit(void)
216 if (acpi_processor_ppc_status
& PPC_REGISTERED
)
217 cpufreq_unregister_notifier(&acpi_ppc_notifier_block
,
218 CPUFREQ_POLICY_NOTIFIER
);
220 acpi_processor_ppc_status
&= ~PPC_REGISTERED
;
223 static int acpi_processor_get_performance_control(struct acpi_processor
*pr
)
226 acpi_status status
= 0;
227 struct acpi_buffer buffer
= { ACPI_ALLOCATE_BUFFER
, NULL
};
228 union acpi_object
*pct
= NULL
;
229 union acpi_object obj
= { 0 };
232 status
= acpi_evaluate_object(pr
->handle
, "_PCT", NULL
, &buffer
);
233 if (ACPI_FAILURE(status
)) {
234 ACPI_EXCEPTION((AE_INFO
, status
, "Evaluating _PCT"));
238 pct
= (union acpi_object
*)buffer
.pointer
;
239 if (!pct
|| (pct
->type
!= ACPI_TYPE_PACKAGE
)
240 || (pct
->package
.count
!= 2)) {
241 printk(KERN_ERR PREFIX
"Invalid _PCT data\n");
250 obj
= pct
->package
.elements
[0];
252 if ((obj
.type
!= ACPI_TYPE_BUFFER
)
253 || (obj
.buffer
.length
< sizeof(struct acpi_pct_register
))
254 || (obj
.buffer
.pointer
== NULL
)) {
255 printk(KERN_ERR PREFIX
"Invalid _PCT data (control_register)\n");
259 memcpy(&pr
->performance
->control_register
, obj
.buffer
.pointer
,
260 sizeof(struct acpi_pct_register
));
266 obj
= pct
->package
.elements
[1];
268 if ((obj
.type
!= ACPI_TYPE_BUFFER
)
269 || (obj
.buffer
.length
< sizeof(struct acpi_pct_register
))
270 || (obj
.buffer
.pointer
== NULL
)) {
271 printk(KERN_ERR PREFIX
"Invalid _PCT data (status_register)\n");
276 memcpy(&pr
->performance
->status_register
, obj
.buffer
.pointer
,
277 sizeof(struct acpi_pct_register
));
280 kfree(buffer
.pointer
);
287 * Some AMDs have 50MHz frequency multiples, but only provide 100MHz rounding
288 * in their ACPI data. Calculate the real values and fix up the _PSS data.
290 static void amd_fixup_frequency(struct acpi_processor_px
*px
, int i
)
292 u32 hi
, lo
, fid
, did
;
293 int index
= px
->control
& 0x00000007;
295 if (boot_cpu_data
.x86_vendor
!= X86_VENDOR_AMD
)
298 if ((boot_cpu_data
.x86
== 0x10 && boot_cpu_data
.x86_model
< 10)
299 || boot_cpu_data
.x86
== 0x11) {
300 rdmsr(MSR_AMD_PSTATE_DEF_BASE
+ index
, lo
, hi
);
303 * Bit 63: PstateEn. Read-write. If set, the P-state is valid.
310 if (boot_cpu_data
.x86
== 0x10)
311 px
->core_frequency
= (100 * (fid
+ 0x10)) >> did
;
313 px
->core_frequency
= (100 * (fid
+ 8)) >> did
;
317 static void amd_fixup_frequency(struct acpi_processor_px
*px
, int i
) {};
320 static int acpi_processor_get_performance_states(struct acpi_processor
*pr
)
323 acpi_status status
= AE_OK
;
324 struct acpi_buffer buffer
= { ACPI_ALLOCATE_BUFFER
, NULL
};
325 struct acpi_buffer format
= { sizeof("NNNNNN"), "NNNNNN" };
326 struct acpi_buffer state
= { 0, NULL
};
327 union acpi_object
*pss
= NULL
;
329 int last_invalid
= -1;
332 status
= acpi_evaluate_object(pr
->handle
, "_PSS", NULL
, &buffer
);
333 if (ACPI_FAILURE(status
)) {
334 ACPI_EXCEPTION((AE_INFO
, status
, "Evaluating _PSS"));
338 pss
= buffer
.pointer
;
339 if (!pss
|| (pss
->type
!= ACPI_TYPE_PACKAGE
)) {
340 printk(KERN_ERR PREFIX
"Invalid _PSS data\n");
345 ACPI_DEBUG_PRINT((ACPI_DB_INFO
, "Found %d performance states\n",
346 pss
->package
.count
));
348 pr
->performance
->state_count
= pss
->package
.count
;
349 pr
->performance
->states
=
350 kmalloc(sizeof(struct acpi_processor_px
) * pss
->package
.count
,
352 if (!pr
->performance
->states
) {
357 for (i
= 0; i
< pr
->performance
->state_count
; i
++) {
359 struct acpi_processor_px
*px
= &(pr
->performance
->states
[i
]);
361 state
.length
= sizeof(struct acpi_processor_px
);
364 ACPI_DEBUG_PRINT((ACPI_DB_INFO
, "Extracting state %d\n", i
));
366 status
= acpi_extract_package(&(pss
->package
.elements
[i
]),
368 if (ACPI_FAILURE(status
)) {
369 ACPI_EXCEPTION((AE_INFO
, status
, "Invalid _PSS data"));
371 kfree(pr
->performance
->states
);
375 amd_fixup_frequency(px
, i
);
377 ACPI_DEBUG_PRINT((ACPI_DB_INFO
,
378 "State [%d]: core_frequency[%d] power[%d] transition_latency[%d] bus_master_latency[%d] control[0x%x] status[0x%x]\n",
380 (u32
) px
->core_frequency
,
382 (u32
) px
->transition_latency
,
383 (u32
) px
->bus_master_latency
,
384 (u32
) px
->control
, (u32
) px
->status
));
387 * Check that ACPI's u64 MHz will be valid as u32 KHz in cpufreq
389 if (!px
->core_frequency
||
390 ((u32
)(px
->core_frequency
* 1000) !=
391 (px
->core_frequency
* 1000))) {
392 printk(KERN_ERR FW_BUG PREFIX
393 "Invalid BIOS _PSS frequency found for processor %d: 0x%llx MHz\n",
394 pr
->id
, px
->core_frequency
);
395 if (last_invalid
== -1)
398 if (last_invalid
!= -1) {
400 * Copy this valid entry over last_invalid entry
402 memcpy(&(pr
->performance
->states
[last_invalid
]),
403 px
, sizeof(struct acpi_processor_px
));
409 if (last_invalid
== 0) {
410 printk(KERN_ERR FW_BUG PREFIX
411 "No valid BIOS _PSS frequency found for processor %d\n", pr
->id
);
413 kfree(pr
->performance
->states
);
414 pr
->performance
->states
= NULL
;
417 if (last_invalid
> 0)
418 pr
->performance
->state_count
= last_invalid
;
421 kfree(buffer
.pointer
);
426 int acpi_processor_get_performance_info(struct acpi_processor
*pr
)
430 if (!pr
|| !pr
->performance
|| !pr
->handle
)
433 if (!acpi_has_method(pr
->handle
, "_PCT")) {
434 ACPI_DEBUG_PRINT((ACPI_DB_INFO
,
435 "ACPI-based processor performance control unavailable\n"));
439 result
= acpi_processor_get_performance_control(pr
);
443 result
= acpi_processor_get_performance_states(pr
);
447 /* We need to call _PPC once when cpufreq starts */
449 result
= acpi_processor_get_platform_limit(pr
);
454 * Having _PPC but missing frequencies (_PSS, _PCT) is a very good hint that
455 * the BIOS is older than the CPU and does not know its frequencies
459 if (acpi_has_method(pr
->handle
, "_PPC")) {
460 if(boot_cpu_has(X86_FEATURE_EST
))
461 printk(KERN_WARNING FW_BUG
"BIOS needs update for CPU "
462 "frequency support\n");
467 EXPORT_SYMBOL_GPL(acpi_processor_get_performance_info
);
468 int acpi_processor_notify_smm(struct module
*calling_module
)
471 static int is_done
= 0;
474 if (!(acpi_processor_ppc_status
& PPC_REGISTERED
))
477 if (!try_module_get(calling_module
))
480 /* is_done is set to negative if an error occurred,
481 * and to postitive if _no_ error occurred, but SMM
482 * was already notified. This avoids double notification
483 * which might lead to unexpected results...
486 module_put(calling_module
);
488 } else if (is_done
< 0) {
489 module_put(calling_module
);
495 /* Can't write pstate_control to smi_command if either value is zero */
496 if ((!acpi_gbl_FADT
.smi_command
) || (!acpi_gbl_FADT
.pstate_control
)) {
497 ACPI_DEBUG_PRINT((ACPI_DB_INFO
, "No SMI port or pstate_control\n"));
498 module_put(calling_module
);
502 ACPI_DEBUG_PRINT((ACPI_DB_INFO
,
503 "Writing pstate_control [0x%x] to smi_command [0x%x]\n",
504 acpi_gbl_FADT
.pstate_control
, acpi_gbl_FADT
.smi_command
));
506 status
= acpi_os_write_port(acpi_gbl_FADT
.smi_command
,
507 (u32
) acpi_gbl_FADT
.pstate_control
, 8);
508 if (ACPI_FAILURE(status
)) {
509 ACPI_EXCEPTION((AE_INFO
, status
,
510 "Failed to write pstate_control [0x%x] to "
511 "smi_command [0x%x]", acpi_gbl_FADT
.pstate_control
,
512 acpi_gbl_FADT
.smi_command
));
513 module_put(calling_module
);
517 /* Success. If there's no _PPC, we need to fear nothing, so
518 * we can allow the cpufreq driver to be rmmod'ed. */
521 if (!(acpi_processor_ppc_status
& PPC_IN_USE
))
522 module_put(calling_module
);
527 EXPORT_SYMBOL(acpi_processor_notify_smm
);
529 static int acpi_processor_get_psd(struct acpi_processor
*pr
)
532 acpi_status status
= AE_OK
;
533 struct acpi_buffer buffer
= {ACPI_ALLOCATE_BUFFER
, NULL
};
534 struct acpi_buffer format
= {sizeof("NNNNN"), "NNNNN"};
535 struct acpi_buffer state
= {0, NULL
};
536 union acpi_object
*psd
= NULL
;
537 struct acpi_psd_package
*pdomain
;
539 status
= acpi_evaluate_object(pr
->handle
, "_PSD", NULL
, &buffer
);
540 if (ACPI_FAILURE(status
)) {
544 psd
= buffer
.pointer
;
545 if (!psd
|| (psd
->type
!= ACPI_TYPE_PACKAGE
)) {
546 printk(KERN_ERR PREFIX
"Invalid _PSD data\n");
551 if (psd
->package
.count
!= 1) {
552 printk(KERN_ERR PREFIX
"Invalid _PSD data\n");
557 pdomain
= &(pr
->performance
->domain_info
);
559 state
.length
= sizeof(struct acpi_psd_package
);
560 state
.pointer
= pdomain
;
562 status
= acpi_extract_package(&(psd
->package
.elements
[0]),
564 if (ACPI_FAILURE(status
)) {
565 printk(KERN_ERR PREFIX
"Invalid _PSD data\n");
570 if (pdomain
->num_entries
!= ACPI_PSD_REV0_ENTRIES
) {
571 printk(KERN_ERR PREFIX
"Unknown _PSD:num_entries\n");
576 if (pdomain
->revision
!= ACPI_PSD_REV0_REVISION
) {
577 printk(KERN_ERR PREFIX
"Unknown _PSD:revision\n");
582 if (pdomain
->coord_type
!= DOMAIN_COORD_TYPE_SW_ALL
&&
583 pdomain
->coord_type
!= DOMAIN_COORD_TYPE_SW_ANY
&&
584 pdomain
->coord_type
!= DOMAIN_COORD_TYPE_HW_ALL
) {
585 printk(KERN_ERR PREFIX
"Invalid _PSD:coord_type\n");
590 kfree(buffer
.pointer
);
594 int acpi_processor_preregister_performance(
595 struct acpi_processor_performance __percpu
*performance
)
600 cpumask_var_t covered_cpus
;
601 struct acpi_processor
*pr
;
602 struct acpi_psd_package
*pdomain
;
603 struct acpi_processor
*match_pr
;
604 struct acpi_psd_package
*match_pdomain
;
606 if (!zalloc_cpumask_var(&covered_cpus
, GFP_KERNEL
))
609 mutex_lock(&performance_mutex
);
612 * Check if another driver has already registered, and abort before
613 * changing pr->performance if it has. Check input data as well.
615 for_each_possible_cpu(i
) {
616 pr
= per_cpu(processors
, i
);
618 /* Look only at processors in ACPI namespace */
622 if (pr
->performance
) {
627 if (!performance
|| !per_cpu_ptr(performance
, i
)) {
633 /* Call _PSD for all CPUs */
634 for_each_possible_cpu(i
) {
635 pr
= per_cpu(processors
, i
);
639 pr
->performance
= per_cpu_ptr(performance
, i
);
640 cpumask_set_cpu(i
, pr
->performance
->shared_cpu_map
);
641 if (acpi_processor_get_psd(pr
)) {
650 * Now that we have _PSD data from all CPUs, lets setup P-state
653 for_each_possible_cpu(i
) {
654 pr
= per_cpu(processors
, i
);
658 if (cpumask_test_cpu(i
, covered_cpus
))
661 pdomain
= &(pr
->performance
->domain_info
);
662 cpumask_set_cpu(i
, pr
->performance
->shared_cpu_map
);
663 cpumask_set_cpu(i
, covered_cpus
);
664 if (pdomain
->num_processors
<= 1)
667 /* Validate the Domain info */
668 count_target
= pdomain
->num_processors
;
669 if (pdomain
->coord_type
== DOMAIN_COORD_TYPE_SW_ALL
)
670 pr
->performance
->shared_type
= CPUFREQ_SHARED_TYPE_ALL
;
671 else if (pdomain
->coord_type
== DOMAIN_COORD_TYPE_HW_ALL
)
672 pr
->performance
->shared_type
= CPUFREQ_SHARED_TYPE_HW
;
673 else if (pdomain
->coord_type
== DOMAIN_COORD_TYPE_SW_ANY
)
674 pr
->performance
->shared_type
= CPUFREQ_SHARED_TYPE_ANY
;
676 for_each_possible_cpu(j
) {
680 match_pr
= per_cpu(processors
, j
);
684 match_pdomain
= &(match_pr
->performance
->domain_info
);
685 if (match_pdomain
->domain
!= pdomain
->domain
)
688 /* Here i and j are in the same domain */
690 if (match_pdomain
->num_processors
!= count_target
) {
695 if (pdomain
->coord_type
!= match_pdomain
->coord_type
) {
700 cpumask_set_cpu(j
, covered_cpus
);
701 cpumask_set_cpu(j
, pr
->performance
->shared_cpu_map
);
704 for_each_possible_cpu(j
) {
708 match_pr
= per_cpu(processors
, j
);
712 match_pdomain
= &(match_pr
->performance
->domain_info
);
713 if (match_pdomain
->domain
!= pdomain
->domain
)
716 match_pr
->performance
->shared_type
=
717 pr
->performance
->shared_type
;
718 cpumask_copy(match_pr
->performance
->shared_cpu_map
,
719 pr
->performance
->shared_cpu_map
);
724 for_each_possible_cpu(i
) {
725 pr
= per_cpu(processors
, i
);
726 if (!pr
|| !pr
->performance
)
729 /* Assume no coordination on any error parsing domain info */
731 cpumask_clear(pr
->performance
->shared_cpu_map
);
732 cpumask_set_cpu(i
, pr
->performance
->shared_cpu_map
);
733 pr
->performance
->shared_type
= CPUFREQ_SHARED_TYPE_ALL
;
735 pr
->performance
= NULL
; /* Will be set for real in register */
739 mutex_unlock(&performance_mutex
);
740 free_cpumask_var(covered_cpus
);
743 EXPORT_SYMBOL(acpi_processor_preregister_performance
);
746 acpi_processor_register_performance(struct acpi_processor_performance
747 *performance
, unsigned int cpu
)
749 struct acpi_processor
*pr
;
751 if (!(acpi_processor_ppc_status
& PPC_REGISTERED
))
754 mutex_lock(&performance_mutex
);
756 pr
= per_cpu(processors
, cpu
);
758 mutex_unlock(&performance_mutex
);
762 if (pr
->performance
) {
763 mutex_unlock(&performance_mutex
);
767 WARN_ON(!performance
);
769 pr
->performance
= performance
;
771 if (acpi_processor_get_performance_info(pr
)) {
772 pr
->performance
= NULL
;
773 mutex_unlock(&performance_mutex
);
777 mutex_unlock(&performance_mutex
);
781 EXPORT_SYMBOL(acpi_processor_register_performance
);
783 void acpi_processor_unregister_performance(unsigned int cpu
)
785 struct acpi_processor
*pr
;
787 mutex_lock(&performance_mutex
);
789 pr
= per_cpu(processors
, cpu
);
791 mutex_unlock(&performance_mutex
);
796 kfree(pr
->performance
->states
);
797 pr
->performance
= NULL
;
799 mutex_unlock(&performance_mutex
);
804 EXPORT_SYMBOL(acpi_processor_unregister_performance
);