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
;
84 if (event
!= CPUFREQ_ADJUST
)
87 mutex_lock(&performance_mutex
);
89 pr
= per_cpu(processors
, policy
->cpu
);
90 if (!pr
|| !pr
->performance
)
93 ppc
= (unsigned int)pr
->performance_platform_limit
;
95 if (ppc
>= pr
->performance
->state_count
)
98 cpufreq_verify_within_limits(policy
, 0,
99 pr
->performance
->states
[ppc
].
100 core_frequency
* 1000);
103 mutex_unlock(&performance_mutex
);
108 static struct notifier_block acpi_ppc_notifier_block
= {
109 .notifier_call
= acpi_processor_ppc_notifier
,
112 static int acpi_processor_get_platform_limit(struct acpi_processor
*pr
)
114 acpi_status status
= 0;
115 unsigned long long ppc
= 0;
122 * _PPC indicates the maximum state currently supported by the platform
123 * (e.g. 0 = states 0..n; 1 = states 1..n; etc.
125 status
= acpi_evaluate_integer(pr
->handle
, "_PPC", NULL
, &ppc
);
127 if (status
!= AE_NOT_FOUND
)
128 acpi_processor_ppc_status
|= PPC_IN_USE
;
130 if (ACPI_FAILURE(status
) && status
!= AE_NOT_FOUND
) {
131 ACPI_EXCEPTION((AE_INFO
, status
, "Evaluating _PPC"));
135 pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr
->id
,
136 (int)ppc
, ppc
? "" : "not");
138 pr
->performance_platform_limit
= (int)ppc
;
143 #define ACPI_PROCESSOR_NOTIFY_PERFORMANCE 0x80
145 * acpi_processor_ppc_ost: Notify firmware the _PPC evaluation status
146 * @handle: ACPI processor handle
147 * @status: the status code of _PPC evaluation
148 * 0: success. OSPM is now using the performance state specificed.
149 * 1: failure. OSPM has not changed the number of P-states in use
151 static void acpi_processor_ppc_ost(acpi_handle handle
, int status
)
153 if (acpi_has_method(handle
, "_OST"))
154 acpi_evaluate_ost(handle
, ACPI_PROCESSOR_NOTIFY_PERFORMANCE
,
158 void acpi_processor_ppc_has_changed(struct acpi_processor
*pr
, int event_flag
)
162 if (ignore_ppc
|| !pr
->performance
) {
164 * Only when it is notification event, the _OST object
165 * will be evaluated. Otherwise it is skipped.
168 acpi_processor_ppc_ost(pr
->handle
, 1);
172 ret
= acpi_processor_get_platform_limit(pr
);
174 * Only when it is notification event, the _OST object
175 * will be evaluated. Otherwise it is skipped.
179 acpi_processor_ppc_ost(pr
->handle
, 1);
181 acpi_processor_ppc_ost(pr
->handle
, 0);
184 cpufreq_update_policy(pr
->id
);
187 int acpi_processor_get_bios_limit(int cpu
, unsigned int *limit
)
189 struct acpi_processor
*pr
;
191 pr
= per_cpu(processors
, cpu
);
192 if (!pr
|| !pr
->performance
|| !pr
->performance
->state_count
)
194 *limit
= pr
->performance
->states
[pr
->performance_platform_limit
].
195 core_frequency
* 1000;
198 EXPORT_SYMBOL(acpi_processor_get_bios_limit
);
200 void acpi_processor_ppc_init(void)
202 if (!cpufreq_register_notifier
203 (&acpi_ppc_notifier_block
, CPUFREQ_POLICY_NOTIFIER
))
204 acpi_processor_ppc_status
|= PPC_REGISTERED
;
207 "Warning: Processor Platform Limit not supported.\n");
210 void acpi_processor_ppc_exit(void)
212 if (acpi_processor_ppc_status
& PPC_REGISTERED
)
213 cpufreq_unregister_notifier(&acpi_ppc_notifier_block
,
214 CPUFREQ_POLICY_NOTIFIER
);
216 acpi_processor_ppc_status
&= ~PPC_REGISTERED
;
219 static int acpi_processor_get_performance_control(struct acpi_processor
*pr
)
222 acpi_status status
= 0;
223 struct acpi_buffer buffer
= { ACPI_ALLOCATE_BUFFER
, NULL
};
224 union acpi_object
*pct
= NULL
;
225 union acpi_object obj
= { 0 };
228 status
= acpi_evaluate_object(pr
->handle
, "_PCT", NULL
, &buffer
);
229 if (ACPI_FAILURE(status
)) {
230 ACPI_EXCEPTION((AE_INFO
, status
, "Evaluating _PCT"));
234 pct
= (union acpi_object
*)buffer
.pointer
;
235 if (!pct
|| (pct
->type
!= ACPI_TYPE_PACKAGE
)
236 || (pct
->package
.count
!= 2)) {
237 printk(KERN_ERR PREFIX
"Invalid _PCT data\n");
246 obj
= pct
->package
.elements
[0];
248 if ((obj
.type
!= ACPI_TYPE_BUFFER
)
249 || (obj
.buffer
.length
< sizeof(struct acpi_pct_register
))
250 || (obj
.buffer
.pointer
== NULL
)) {
251 printk(KERN_ERR PREFIX
"Invalid _PCT data (control_register)\n");
255 memcpy(&pr
->performance
->control_register
, obj
.buffer
.pointer
,
256 sizeof(struct acpi_pct_register
));
262 obj
= pct
->package
.elements
[1];
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 (status_register)\n");
272 memcpy(&pr
->performance
->status_register
, obj
.buffer
.pointer
,
273 sizeof(struct acpi_pct_register
));
276 kfree(buffer
.pointer
);
283 * Some AMDs have 50MHz frequency multiples, but only provide 100MHz rounding
284 * in their ACPI data. Calculate the real values and fix up the _PSS data.
286 static void amd_fixup_frequency(struct acpi_processor_px
*px
, int i
)
288 u32 hi
, lo
, fid
, did
;
289 int index
= px
->control
& 0x00000007;
291 if (boot_cpu_data
.x86_vendor
!= X86_VENDOR_AMD
)
294 if ((boot_cpu_data
.x86
== 0x10 && boot_cpu_data
.x86_model
< 10)
295 || boot_cpu_data
.x86
== 0x11) {
296 rdmsr(MSR_AMD_PSTATE_DEF_BASE
+ index
, lo
, hi
);
299 * Bit 63: PstateEn. Read-write. If set, the P-state is valid.
306 if (boot_cpu_data
.x86
== 0x10)
307 px
->core_frequency
= (100 * (fid
+ 0x10)) >> did
;
309 px
->core_frequency
= (100 * (fid
+ 8)) >> did
;
313 static void amd_fixup_frequency(struct acpi_processor_px
*px
, int i
) {};
316 static int acpi_processor_get_performance_states(struct acpi_processor
*pr
)
319 acpi_status status
= AE_OK
;
320 struct acpi_buffer buffer
= { ACPI_ALLOCATE_BUFFER
, NULL
};
321 struct acpi_buffer format
= { sizeof("NNNNNN"), "NNNNNN" };
322 struct acpi_buffer state
= { 0, NULL
};
323 union acpi_object
*pss
= NULL
;
325 int last_invalid
= -1;
328 status
= acpi_evaluate_object(pr
->handle
, "_PSS", NULL
, &buffer
);
329 if (ACPI_FAILURE(status
)) {
330 ACPI_EXCEPTION((AE_INFO
, status
, "Evaluating _PSS"));
334 pss
= buffer
.pointer
;
335 if (!pss
|| (pss
->type
!= ACPI_TYPE_PACKAGE
)) {
336 printk(KERN_ERR PREFIX
"Invalid _PSS data\n");
341 ACPI_DEBUG_PRINT((ACPI_DB_INFO
, "Found %d performance states\n",
342 pss
->package
.count
));
344 pr
->performance
->state_count
= pss
->package
.count
;
345 pr
->performance
->states
=
346 kmalloc(sizeof(struct acpi_processor_px
) * pss
->package
.count
,
348 if (!pr
->performance
->states
) {
353 for (i
= 0; i
< pr
->performance
->state_count
; i
++) {
355 struct acpi_processor_px
*px
= &(pr
->performance
->states
[i
]);
357 state
.length
= sizeof(struct acpi_processor_px
);
360 ACPI_DEBUG_PRINT((ACPI_DB_INFO
, "Extracting state %d\n", i
));
362 status
= acpi_extract_package(&(pss
->package
.elements
[i
]),
364 if (ACPI_FAILURE(status
)) {
365 ACPI_EXCEPTION((AE_INFO
, status
, "Invalid _PSS data"));
367 kfree(pr
->performance
->states
);
371 amd_fixup_frequency(px
, i
);
373 ACPI_DEBUG_PRINT((ACPI_DB_INFO
,
374 "State [%d]: core_frequency[%d] power[%d] transition_latency[%d] bus_master_latency[%d] control[0x%x] status[0x%x]\n",
376 (u32
) px
->core_frequency
,
378 (u32
) px
->transition_latency
,
379 (u32
) px
->bus_master_latency
,
380 (u32
) px
->control
, (u32
) px
->status
));
383 * Check that ACPI's u64 MHz will be valid as u32 KHz in cpufreq
385 if (!px
->core_frequency
||
386 ((u32
)(px
->core_frequency
* 1000) !=
387 (px
->core_frequency
* 1000))) {
388 printk(KERN_ERR FW_BUG PREFIX
389 "Invalid BIOS _PSS frequency found for processor %d: 0x%llx MHz\n",
390 pr
->id
, px
->core_frequency
);
391 if (last_invalid
== -1)
394 if (last_invalid
!= -1) {
396 * Copy this valid entry over last_invalid entry
398 memcpy(&(pr
->performance
->states
[last_invalid
]),
399 px
, sizeof(struct acpi_processor_px
));
405 if (last_invalid
== 0) {
406 printk(KERN_ERR FW_BUG PREFIX
407 "No valid BIOS _PSS frequency found for processor %d\n", pr
->id
);
409 kfree(pr
->performance
->states
);
410 pr
->performance
->states
= NULL
;
413 if (last_invalid
> 0)
414 pr
->performance
->state_count
= last_invalid
;
417 kfree(buffer
.pointer
);
422 int acpi_processor_get_performance_info(struct acpi_processor
*pr
)
426 if (!pr
|| !pr
->performance
|| !pr
->handle
)
429 if (!acpi_has_method(pr
->handle
, "_PCT")) {
430 ACPI_DEBUG_PRINT((ACPI_DB_INFO
,
431 "ACPI-based processor performance control unavailable\n"));
435 result
= acpi_processor_get_performance_control(pr
);
439 result
= acpi_processor_get_performance_states(pr
);
443 /* We need to call _PPC once when cpufreq starts */
445 result
= acpi_processor_get_platform_limit(pr
);
450 * Having _PPC but missing frequencies (_PSS, _PCT) is a very good hint that
451 * the BIOS is older than the CPU and does not know its frequencies
455 if (acpi_has_method(pr
->handle
, "_PPC")) {
456 if(boot_cpu_has(X86_FEATURE_EST
))
457 printk(KERN_WARNING FW_BUG
"BIOS needs update for CPU "
458 "frequency support\n");
463 EXPORT_SYMBOL_GPL(acpi_processor_get_performance_info
);
465 int acpi_processor_pstate_control(void)
469 if (!acpi_gbl_FADT
.smi_command
|| !acpi_gbl_FADT
.pstate_control
)
472 ACPI_DEBUG_PRINT((ACPI_DB_INFO
,
473 "Writing pstate_control [0x%x] to smi_command [0x%x]\n",
474 acpi_gbl_FADT
.pstate_control
, acpi_gbl_FADT
.smi_command
));
476 status
= acpi_os_write_port(acpi_gbl_FADT
.smi_command
,
477 (u32
)acpi_gbl_FADT
.pstate_control
, 8);
478 if (ACPI_SUCCESS(status
))
481 ACPI_EXCEPTION((AE_INFO
, status
,
482 "Failed to write pstate_control [0x%x] to smi_command [0x%x]",
483 acpi_gbl_FADT
.pstate_control
, acpi_gbl_FADT
.smi_command
));
487 int acpi_processor_notify_smm(struct module
*calling_module
)
489 static int is_done
= 0;
492 if (!(acpi_processor_ppc_status
& PPC_REGISTERED
))
495 if (!try_module_get(calling_module
))
498 /* is_done is set to negative if an error occurred,
499 * and to postitive if _no_ error occurred, but SMM
500 * was already notified. This avoids double notification
501 * which might lead to unexpected results...
504 module_put(calling_module
);
506 } else if (is_done
< 0) {
507 module_put(calling_module
);
513 result
= acpi_processor_pstate_control();
515 ACPI_DEBUG_PRINT((ACPI_DB_INFO
, "No SMI port or pstate_control\n"));
516 module_put(calling_module
);
520 module_put(calling_module
);
524 /* Success. If there's no _PPC, we need to fear nothing, so
525 * we can allow the cpufreq driver to be rmmod'ed. */
528 if (!(acpi_processor_ppc_status
& PPC_IN_USE
))
529 module_put(calling_module
);
534 EXPORT_SYMBOL(acpi_processor_notify_smm
);
536 int acpi_processor_get_psd(acpi_handle handle
, struct acpi_psd_package
*pdomain
)
539 acpi_status status
= AE_OK
;
540 struct acpi_buffer buffer
= {ACPI_ALLOCATE_BUFFER
, NULL
};
541 struct acpi_buffer format
= {sizeof("NNNNN"), "NNNNN"};
542 struct acpi_buffer state
= {0, NULL
};
543 union acpi_object
*psd
= NULL
;
545 status
= acpi_evaluate_object(handle
, "_PSD", NULL
, &buffer
);
546 if (ACPI_FAILURE(status
)) {
550 psd
= buffer
.pointer
;
551 if (!psd
|| (psd
->type
!= ACPI_TYPE_PACKAGE
)) {
552 printk(KERN_ERR PREFIX
"Invalid _PSD data\n");
557 if (psd
->package
.count
!= 1) {
558 printk(KERN_ERR PREFIX
"Invalid _PSD data\n");
563 state
.length
= sizeof(struct acpi_psd_package
);
564 state
.pointer
= pdomain
;
566 status
= acpi_extract_package(&(psd
->package
.elements
[0]),
568 if (ACPI_FAILURE(status
)) {
569 printk(KERN_ERR PREFIX
"Invalid _PSD data\n");
574 if (pdomain
->num_entries
!= ACPI_PSD_REV0_ENTRIES
) {
575 printk(KERN_ERR PREFIX
"Unknown _PSD:num_entries\n");
580 if (pdomain
->revision
!= ACPI_PSD_REV0_REVISION
) {
581 printk(KERN_ERR PREFIX
"Unknown _PSD:revision\n");
586 if (pdomain
->coord_type
!= DOMAIN_COORD_TYPE_SW_ALL
&&
587 pdomain
->coord_type
!= DOMAIN_COORD_TYPE_SW_ANY
&&
588 pdomain
->coord_type
!= DOMAIN_COORD_TYPE_HW_ALL
) {
589 printk(KERN_ERR PREFIX
"Invalid _PSD:coord_type\n");
594 kfree(buffer
.pointer
);
597 EXPORT_SYMBOL(acpi_processor_get_psd
);
599 int acpi_processor_preregister_performance(
600 struct acpi_processor_performance __percpu
*performance
)
605 cpumask_var_t covered_cpus
;
606 struct acpi_processor
*pr
;
607 struct acpi_psd_package
*pdomain
;
608 struct acpi_processor
*match_pr
;
609 struct acpi_psd_package
*match_pdomain
;
611 if (!zalloc_cpumask_var(&covered_cpus
, GFP_KERNEL
))
614 mutex_lock(&performance_mutex
);
617 * Check if another driver has already registered, and abort before
618 * changing pr->performance if it has. Check input data as well.
620 for_each_possible_cpu(i
) {
621 pr
= per_cpu(processors
, i
);
623 /* Look only at processors in ACPI namespace */
627 if (pr
->performance
) {
632 if (!performance
|| !per_cpu_ptr(performance
, i
)) {
638 /* Call _PSD for all CPUs */
639 for_each_possible_cpu(i
) {
640 pr
= per_cpu(processors
, i
);
644 pr
->performance
= per_cpu_ptr(performance
, i
);
645 cpumask_set_cpu(i
, pr
->performance
->shared_cpu_map
);
646 pdomain
= &(pr
->performance
->domain_info
);
647 if (acpi_processor_get_psd(pr
->handle
, pdomain
)) {
656 * Now that we have _PSD data from all CPUs, lets setup P-state
659 for_each_possible_cpu(i
) {
660 pr
= per_cpu(processors
, i
);
664 if (cpumask_test_cpu(i
, covered_cpus
))
667 pdomain
= &(pr
->performance
->domain_info
);
668 cpumask_set_cpu(i
, pr
->performance
->shared_cpu_map
);
669 cpumask_set_cpu(i
, covered_cpus
);
670 if (pdomain
->num_processors
<= 1)
673 /* Validate the Domain info */
674 count_target
= pdomain
->num_processors
;
675 if (pdomain
->coord_type
== DOMAIN_COORD_TYPE_SW_ALL
)
676 pr
->performance
->shared_type
= CPUFREQ_SHARED_TYPE_ALL
;
677 else if (pdomain
->coord_type
== DOMAIN_COORD_TYPE_HW_ALL
)
678 pr
->performance
->shared_type
= CPUFREQ_SHARED_TYPE_HW
;
679 else if (pdomain
->coord_type
== DOMAIN_COORD_TYPE_SW_ANY
)
680 pr
->performance
->shared_type
= CPUFREQ_SHARED_TYPE_ANY
;
682 for_each_possible_cpu(j
) {
686 match_pr
= per_cpu(processors
, j
);
690 match_pdomain
= &(match_pr
->performance
->domain_info
);
691 if (match_pdomain
->domain
!= pdomain
->domain
)
694 /* Here i and j are in the same domain */
696 if (match_pdomain
->num_processors
!= count_target
) {
701 if (pdomain
->coord_type
!= match_pdomain
->coord_type
) {
706 cpumask_set_cpu(j
, covered_cpus
);
707 cpumask_set_cpu(j
, pr
->performance
->shared_cpu_map
);
710 for_each_possible_cpu(j
) {
714 match_pr
= per_cpu(processors
, j
);
718 match_pdomain
= &(match_pr
->performance
->domain_info
);
719 if (match_pdomain
->domain
!= pdomain
->domain
)
722 match_pr
->performance
->shared_type
=
723 pr
->performance
->shared_type
;
724 cpumask_copy(match_pr
->performance
->shared_cpu_map
,
725 pr
->performance
->shared_cpu_map
);
730 for_each_possible_cpu(i
) {
731 pr
= per_cpu(processors
, i
);
732 if (!pr
|| !pr
->performance
)
735 /* Assume no coordination on any error parsing domain info */
737 cpumask_clear(pr
->performance
->shared_cpu_map
);
738 cpumask_set_cpu(i
, pr
->performance
->shared_cpu_map
);
739 pr
->performance
->shared_type
= CPUFREQ_SHARED_TYPE_ALL
;
741 pr
->performance
= NULL
; /* Will be set for real in register */
745 mutex_unlock(&performance_mutex
);
746 free_cpumask_var(covered_cpus
);
749 EXPORT_SYMBOL(acpi_processor_preregister_performance
);
752 acpi_processor_register_performance(struct acpi_processor_performance
753 *performance
, unsigned int cpu
)
755 struct acpi_processor
*pr
;
757 if (!(acpi_processor_ppc_status
& PPC_REGISTERED
))
760 mutex_lock(&performance_mutex
);
762 pr
= per_cpu(processors
, cpu
);
764 mutex_unlock(&performance_mutex
);
768 if (pr
->performance
) {
769 mutex_unlock(&performance_mutex
);
773 WARN_ON(!performance
);
775 pr
->performance
= performance
;
777 if (acpi_processor_get_performance_info(pr
)) {
778 pr
->performance
= NULL
;
779 mutex_unlock(&performance_mutex
);
783 mutex_unlock(&performance_mutex
);
787 EXPORT_SYMBOL(acpi_processor_register_performance
);
789 void acpi_processor_unregister_performance(unsigned int cpu
)
791 struct acpi_processor
*pr
;
793 mutex_lock(&performance_mutex
);
795 pr
= per_cpu(processors
, cpu
);
797 mutex_unlock(&performance_mutex
);
802 kfree(pr
->performance
->states
);
803 pr
->performance
= NULL
;
805 mutex_unlock(&performance_mutex
);
810 EXPORT_SYMBOL(acpi_processor_unregister_performance
);