acpi-cpufreq: Add quirk to disable _PSD usage on all AMD CPUs
[linux-2.6-block.git] / drivers / cpufreq / acpi-cpufreq.c
1 /*
2  * acpi-cpufreq.c - ACPI Processor P-States Driver
3  *
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) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
7  *  Copyright (C) 2006       Denis Sadykov <denis.m.sadykov@intel.com>
8  *
9  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
10  *
11  *  This program is free software; you can redistribute it and/or modify
12  *  it under the terms of the GNU General Public License as published by
13  *  the Free Software Foundation; either version 2 of the License, or (at
14  *  your option) any later version.
15  *
16  *  This program is distributed in the hope that it will be useful, but
17  *  WITHOUT ANY WARRANTY; without even the implied warranty of
18  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
19  *  General Public License for more details.
20  *
21  *  You should have received a copy of the GNU General Public License along
22  *  with this program; if not, write to the Free Software Foundation, Inc.,
23  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
24  *
25  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
26  */
27
28 #include <linux/kernel.h>
29 #include <linux/module.h>
30 #include <linux/init.h>
31 #include <linux/smp.h>
32 #include <linux/sched.h>
33 #include <linux/cpufreq.h>
34 #include <linux/compiler.h>
35 #include <linux/dmi.h>
36 #include <linux/slab.h>
37
38 #include <linux/acpi.h>
39 #include <linux/io.h>
40 #include <linux/delay.h>
41 #include <linux/uaccess.h>
42
43 #include <acpi/processor.h>
44
45 #include <asm/msr.h>
46 #include <asm/processor.h>
47 #include <asm/cpufeature.h>
48 #include "mperf.h"
49
50 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
51 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
52 MODULE_LICENSE("GPL");
53
54 #define PFX "acpi-cpufreq: "
55
56 enum {
57         UNDEFINED_CAPABLE = 0,
58         SYSTEM_INTEL_MSR_CAPABLE,
59         SYSTEM_AMD_MSR_CAPABLE,
60         SYSTEM_IO_CAPABLE,
61 };
62
63 #define INTEL_MSR_RANGE         (0xffff)
64 #define AMD_MSR_RANGE           (0x7)
65
66 struct acpi_cpufreq_data {
67         struct acpi_processor_performance *acpi_data;
68         struct cpufreq_frequency_table *freq_table;
69         unsigned int resume;
70         unsigned int cpu_feature;
71 };
72
73 static DEFINE_PER_CPU(struct acpi_cpufreq_data *, acfreq_data);
74
75 /* acpi_perf_data is a pointer to percpu data. */
76 static struct acpi_processor_performance __percpu *acpi_perf_data;
77
78 static struct cpufreq_driver acpi_cpufreq_driver;
79
80 static unsigned int acpi_pstate_strict;
81
82 static int check_est_cpu(unsigned int cpuid)
83 {
84         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
85
86         return cpu_has(cpu, X86_FEATURE_EST);
87 }
88
89 static int check_amd_hwpstate_cpu(unsigned int cpuid)
90 {
91         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
92
93         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
94 }
95
96 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
97 {
98         struct acpi_processor_performance *perf;
99         int i;
100
101         perf = data->acpi_data;
102
103         for (i = 0; i < perf->state_count; i++) {
104                 if (value == perf->states[i].status)
105                         return data->freq_table[i].frequency;
106         }
107         return 0;
108 }
109
110 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
111 {
112         int i;
113         struct acpi_processor_performance *perf;
114
115         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
116                 msr &= AMD_MSR_RANGE;
117         else
118                 msr &= INTEL_MSR_RANGE;
119
120         perf = data->acpi_data;
121
122         for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
123                 if (msr == perf->states[data->freq_table[i].index].status)
124                         return data->freq_table[i].frequency;
125         }
126         return data->freq_table[0].frequency;
127 }
128
129 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
130 {
131         switch (data->cpu_feature) {
132         case SYSTEM_INTEL_MSR_CAPABLE:
133         case SYSTEM_AMD_MSR_CAPABLE:
134                 return extract_msr(val, data);
135         case SYSTEM_IO_CAPABLE:
136                 return extract_io(val, data);
137         default:
138                 return 0;
139         }
140 }
141
142 struct msr_addr {
143         u32 reg;
144 };
145
146 struct io_addr {
147         u16 port;
148         u8 bit_width;
149 };
150
151 struct drv_cmd {
152         unsigned int type;
153         const struct cpumask *mask;
154         union {
155                 struct msr_addr msr;
156                 struct io_addr io;
157         } addr;
158         u32 val;
159 };
160
161 /* Called via smp_call_function_single(), on the target CPU */
162 static void do_drv_read(void *_cmd)
163 {
164         struct drv_cmd *cmd = _cmd;
165         u32 h;
166
167         switch (cmd->type) {
168         case SYSTEM_INTEL_MSR_CAPABLE:
169         case SYSTEM_AMD_MSR_CAPABLE:
170                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
171                 break;
172         case SYSTEM_IO_CAPABLE:
173                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
174                                 &cmd->val,
175                                 (u32)cmd->addr.io.bit_width);
176                 break;
177         default:
178                 break;
179         }
180 }
181
182 /* Called via smp_call_function_many(), on the target CPUs */
183 static void do_drv_write(void *_cmd)
184 {
185         struct drv_cmd *cmd = _cmd;
186         u32 lo, hi;
187
188         switch (cmd->type) {
189         case SYSTEM_INTEL_MSR_CAPABLE:
190                 rdmsr(cmd->addr.msr.reg, lo, hi);
191                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
192                 wrmsr(cmd->addr.msr.reg, lo, hi);
193                 break;
194         case SYSTEM_AMD_MSR_CAPABLE:
195                 wrmsr(cmd->addr.msr.reg, cmd->val, 0);
196                 break;
197         case SYSTEM_IO_CAPABLE:
198                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
199                                 cmd->val,
200                                 (u32)cmd->addr.io.bit_width);
201                 break;
202         default:
203                 break;
204         }
205 }
206
207 static void drv_read(struct drv_cmd *cmd)
208 {
209         int err;
210         cmd->val = 0;
211
212         err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
213         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
214 }
215
216 static void drv_write(struct drv_cmd *cmd)
217 {
218         int this_cpu;
219
220         this_cpu = get_cpu();
221         if (cpumask_test_cpu(this_cpu, cmd->mask))
222                 do_drv_write(cmd);
223         smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
224         put_cpu();
225 }
226
227 static u32 get_cur_val(const struct cpumask *mask)
228 {
229         struct acpi_processor_performance *perf;
230         struct drv_cmd cmd;
231
232         if (unlikely(cpumask_empty(mask)))
233                 return 0;
234
235         switch (per_cpu(acfreq_data, cpumask_first(mask))->cpu_feature) {
236         case SYSTEM_INTEL_MSR_CAPABLE:
237                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
238                 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
239                 break;
240         case SYSTEM_AMD_MSR_CAPABLE:
241                 cmd.type = SYSTEM_AMD_MSR_CAPABLE;
242                 cmd.addr.msr.reg = MSR_AMD_PERF_STATUS;
243                 break;
244         case SYSTEM_IO_CAPABLE:
245                 cmd.type = SYSTEM_IO_CAPABLE;
246                 perf = per_cpu(acfreq_data, cpumask_first(mask))->acpi_data;
247                 cmd.addr.io.port = perf->control_register.address;
248                 cmd.addr.io.bit_width = perf->control_register.bit_width;
249                 break;
250         default:
251                 return 0;
252         }
253
254         cmd.mask = mask;
255         drv_read(&cmd);
256
257         pr_debug("get_cur_val = %u\n", cmd.val);
258
259         return cmd.val;
260 }
261
262 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
263 {
264         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
265         unsigned int freq;
266         unsigned int cached_freq;
267
268         pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
269
270         if (unlikely(data == NULL ||
271                      data->acpi_data == NULL || data->freq_table == NULL)) {
272                 return 0;
273         }
274
275         cached_freq = data->freq_table[data->acpi_data->state].frequency;
276         freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
277         if (freq != cached_freq) {
278                 /*
279                  * The dreaded BIOS frequency change behind our back.
280                  * Force set the frequency on next target call.
281                  */
282                 data->resume = 1;
283         }
284
285         pr_debug("cur freq = %u\n", freq);
286
287         return freq;
288 }
289
290 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
291                                 struct acpi_cpufreq_data *data)
292 {
293         unsigned int cur_freq;
294         unsigned int i;
295
296         for (i = 0; i < 100; i++) {
297                 cur_freq = extract_freq(get_cur_val(mask), data);
298                 if (cur_freq == freq)
299                         return 1;
300                 udelay(10);
301         }
302         return 0;
303 }
304
305 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
306                                unsigned int target_freq, unsigned int relation)
307 {
308         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
309         struct acpi_processor_performance *perf;
310         struct cpufreq_freqs freqs;
311         struct drv_cmd cmd;
312         unsigned int next_state = 0; /* Index into freq_table */
313         unsigned int next_perf_state = 0; /* Index into perf table */
314         unsigned int i;
315         int result = 0;
316
317         pr_debug("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
318
319         if (unlikely(data == NULL ||
320              data->acpi_data == NULL || data->freq_table == NULL)) {
321                 return -ENODEV;
322         }
323
324         perf = data->acpi_data;
325         result = cpufreq_frequency_table_target(policy,
326                                                 data->freq_table,
327                                                 target_freq,
328                                                 relation, &next_state);
329         if (unlikely(result)) {
330                 result = -ENODEV;
331                 goto out;
332         }
333
334         next_perf_state = data->freq_table[next_state].index;
335         if (perf->state == next_perf_state) {
336                 if (unlikely(data->resume)) {
337                         pr_debug("Called after resume, resetting to P%d\n",
338                                 next_perf_state);
339                         data->resume = 0;
340                 } else {
341                         pr_debug("Already at target state (P%d)\n",
342                                 next_perf_state);
343                         goto out;
344                 }
345         }
346
347         switch (data->cpu_feature) {
348         case SYSTEM_INTEL_MSR_CAPABLE:
349                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
350                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
351                 cmd.val = (u32) perf->states[next_perf_state].control;
352                 break;
353         case SYSTEM_AMD_MSR_CAPABLE:
354                 cmd.type = SYSTEM_AMD_MSR_CAPABLE;
355                 cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
356                 cmd.val = (u32) perf->states[next_perf_state].control;
357                 break;
358         case SYSTEM_IO_CAPABLE:
359                 cmd.type = SYSTEM_IO_CAPABLE;
360                 cmd.addr.io.port = perf->control_register.address;
361                 cmd.addr.io.bit_width = perf->control_register.bit_width;
362                 cmd.val = (u32) perf->states[next_perf_state].control;
363                 break;
364         default:
365                 result = -ENODEV;
366                 goto out;
367         }
368
369         /* cpufreq holds the hotplug lock, so we are safe from here on */
370         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
371                 cmd.mask = policy->cpus;
372         else
373                 cmd.mask = cpumask_of(policy->cpu);
374
375         freqs.old = perf->states[perf->state].core_frequency * 1000;
376         freqs.new = data->freq_table[next_state].frequency;
377         for_each_cpu(i, policy->cpus) {
378                 freqs.cpu = i;
379                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
380         }
381
382         drv_write(&cmd);
383
384         if (acpi_pstate_strict) {
385                 if (!check_freqs(cmd.mask, freqs.new, data)) {
386                         pr_debug("acpi_cpufreq_target failed (%d)\n",
387                                 policy->cpu);
388                         result = -EAGAIN;
389                         goto out;
390                 }
391         }
392
393         for_each_cpu(i, policy->cpus) {
394                 freqs.cpu = i;
395                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
396         }
397         perf->state = next_perf_state;
398
399 out:
400         return result;
401 }
402
403 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
404 {
405         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
406
407         pr_debug("acpi_cpufreq_verify\n");
408
409         return cpufreq_frequency_table_verify(policy, data->freq_table);
410 }
411
412 static unsigned long
413 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
414 {
415         struct acpi_processor_performance *perf = data->acpi_data;
416
417         if (cpu_khz) {
418                 /* search the closest match to cpu_khz */
419                 unsigned int i;
420                 unsigned long freq;
421                 unsigned long freqn = perf->states[0].core_frequency * 1000;
422
423                 for (i = 0; i < (perf->state_count-1); i++) {
424                         freq = freqn;
425                         freqn = perf->states[i+1].core_frequency * 1000;
426                         if ((2 * cpu_khz) > (freqn + freq)) {
427                                 perf->state = i;
428                                 return freq;
429                         }
430                 }
431                 perf->state = perf->state_count-1;
432                 return freqn;
433         } else {
434                 /* assume CPU is at P0... */
435                 perf->state = 0;
436                 return perf->states[0].core_frequency * 1000;
437         }
438 }
439
440 static void free_acpi_perf_data(void)
441 {
442         unsigned int i;
443
444         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
445         for_each_possible_cpu(i)
446                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
447                                  ->shared_cpu_map);
448         free_percpu(acpi_perf_data);
449 }
450
451 /*
452  * acpi_cpufreq_early_init - initialize ACPI P-States library
453  *
454  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
455  * in order to determine correct frequency and voltage pairings. We can
456  * do _PDC and _PSD and find out the processor dependency for the
457  * actual init that will happen later...
458  */
459 static int __init acpi_cpufreq_early_init(void)
460 {
461         unsigned int i;
462         pr_debug("acpi_cpufreq_early_init\n");
463
464         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
465         if (!acpi_perf_data) {
466                 pr_debug("Memory allocation error for acpi_perf_data.\n");
467                 return -ENOMEM;
468         }
469         for_each_possible_cpu(i) {
470                 if (!zalloc_cpumask_var_node(
471                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
472                         GFP_KERNEL, cpu_to_node(i))) {
473
474                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
475                         free_acpi_perf_data();
476                         return -ENOMEM;
477                 }
478         }
479
480         /* Do initialization in ACPI core */
481         acpi_processor_preregister_performance(acpi_perf_data);
482         return 0;
483 }
484
485 #ifdef CONFIG_SMP
486 /*
487  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
488  * or do it in BIOS firmware and won't inform about it to OS. If not
489  * detected, this has a side effect of making CPU run at a different speed
490  * than OS intended it to run at. Detect it and handle it cleanly.
491  */
492 static int bios_with_sw_any_bug;
493
494 static int sw_any_bug_found(const struct dmi_system_id *d)
495 {
496         bios_with_sw_any_bug = 1;
497         return 0;
498 }
499
500 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
501         {
502                 .callback = sw_any_bug_found,
503                 .ident = "Supermicro Server X6DLP",
504                 .matches = {
505                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
506                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
507                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
508                 },
509         },
510         { }
511 };
512
513 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
514 {
515         /* Intel Xeon Processor 7100 Series Specification Update
516          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
517          * AL30: A Machine Check Exception (MCE) Occurring during an
518          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
519          * Both Processor Cores to Lock Up. */
520         if (c->x86_vendor == X86_VENDOR_INTEL) {
521                 if ((c->x86 == 15) &&
522                     (c->x86_model == 6) &&
523                     (c->x86_mask == 8)) {
524                         printk(KERN_INFO "acpi-cpufreq: Intel(R) "
525                             "Xeon(R) 7100 Errata AL30, processors may "
526                             "lock up on frequency changes: disabling "
527                             "acpi-cpufreq.\n");
528                         return -ENODEV;
529                     }
530                 }
531         return 0;
532 }
533 #endif
534
535 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
536 {
537         unsigned int i;
538         unsigned int valid_states = 0;
539         unsigned int cpu = policy->cpu;
540         struct acpi_cpufreq_data *data;
541         unsigned int result = 0;
542         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
543         struct acpi_processor_performance *perf;
544 #ifdef CONFIG_SMP
545         static int blacklisted;
546 #endif
547
548         pr_debug("acpi_cpufreq_cpu_init\n");
549
550 #ifdef CONFIG_SMP
551         if (blacklisted)
552                 return blacklisted;
553         blacklisted = acpi_cpufreq_blacklist(c);
554         if (blacklisted)
555                 return blacklisted;
556 #endif
557
558         data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
559         if (!data)
560                 return -ENOMEM;
561
562         data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
563         per_cpu(acfreq_data, cpu) = data;
564
565         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
566                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
567
568         result = acpi_processor_register_performance(data->acpi_data, cpu);
569         if (result)
570                 goto err_free;
571
572         perf = data->acpi_data;
573         policy->shared_type = perf->shared_type;
574
575         /*
576          * Will let policy->cpus know about dependency only when software
577          * coordination is required.
578          */
579         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
580             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
581                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
582         }
583         cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
584
585 #ifdef CONFIG_SMP
586         dmi_check_system(sw_any_bug_dmi_table);
587         if (bios_with_sw_any_bug && cpumask_weight(policy->cpus) == 1) {
588                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
589                 cpumask_copy(policy->cpus, cpu_core_mask(cpu));
590         }
591
592         if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
593                 cpumask_clear(policy->cpus);
594                 cpumask_set_cpu(cpu, policy->cpus);
595                 cpumask_copy(policy->related_cpus, cpu_sibling_mask(cpu));
596                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
597                 pr_info_once(PFX "overriding BIOS provided _PSD data\n");
598         }
599 #endif
600
601         /* capability check */
602         if (perf->state_count <= 1) {
603                 pr_debug("No P-States\n");
604                 result = -ENODEV;
605                 goto err_unreg;
606         }
607
608         if (perf->control_register.space_id != perf->status_register.space_id) {
609                 result = -ENODEV;
610                 goto err_unreg;
611         }
612
613         switch (perf->control_register.space_id) {
614         case ACPI_ADR_SPACE_SYSTEM_IO:
615                 pr_debug("SYSTEM IO addr space\n");
616                 data->cpu_feature = SYSTEM_IO_CAPABLE;
617                 break;
618         case ACPI_ADR_SPACE_FIXED_HARDWARE:
619                 pr_debug("HARDWARE addr space\n");
620                 if (check_est_cpu(cpu)) {
621                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
622                         break;
623                 }
624                 if (check_amd_hwpstate_cpu(cpu)) {
625                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
626                         break;
627                 }
628                 result = -ENODEV;
629                 goto err_unreg;
630         default:
631                 pr_debug("Unknown addr space %d\n",
632                         (u32) (perf->control_register.space_id));
633                 result = -ENODEV;
634                 goto err_unreg;
635         }
636
637         data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
638                     (perf->state_count+1), GFP_KERNEL);
639         if (!data->freq_table) {
640                 result = -ENOMEM;
641                 goto err_unreg;
642         }
643
644         /* detect transition latency */
645         policy->cpuinfo.transition_latency = 0;
646         for (i = 0; i < perf->state_count; i++) {
647                 if ((perf->states[i].transition_latency * 1000) >
648                     policy->cpuinfo.transition_latency)
649                         policy->cpuinfo.transition_latency =
650                             perf->states[i].transition_latency * 1000;
651         }
652
653         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
654         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
655             policy->cpuinfo.transition_latency > 20 * 1000) {
656                 policy->cpuinfo.transition_latency = 20 * 1000;
657                 printk_once(KERN_INFO
658                             "P-state transition latency capped at 20 uS\n");
659         }
660
661         /* table init */
662         for (i = 0; i < perf->state_count; i++) {
663                 if (i > 0 && perf->states[i].core_frequency >=
664                     data->freq_table[valid_states-1].frequency / 1000)
665                         continue;
666
667                 data->freq_table[valid_states].index = i;
668                 data->freq_table[valid_states].frequency =
669                     perf->states[i].core_frequency * 1000;
670                 valid_states++;
671         }
672         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
673         perf->state = 0;
674
675         result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
676         if (result)
677                 goto err_freqfree;
678
679         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
680                 printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
681
682         switch (perf->control_register.space_id) {
683         case ACPI_ADR_SPACE_SYSTEM_IO:
684                 /* Current speed is unknown and not detectable by IO port */
685                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
686                 break;
687         case ACPI_ADR_SPACE_FIXED_HARDWARE:
688                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
689                 policy->cur = get_cur_freq_on_cpu(cpu);
690                 break;
691         default:
692                 break;
693         }
694
695         /* notify BIOS that we exist */
696         acpi_processor_notify_smm(THIS_MODULE);
697
698         /* Check for APERF/MPERF support in hardware */
699         if (boot_cpu_has(X86_FEATURE_APERFMPERF))
700                 acpi_cpufreq_driver.getavg = cpufreq_get_measured_perf;
701
702         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
703         for (i = 0; i < perf->state_count; i++)
704                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
705                         (i == perf->state ? '*' : ' '), i,
706                         (u32) perf->states[i].core_frequency,
707                         (u32) perf->states[i].power,
708                         (u32) perf->states[i].transition_latency);
709
710         cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
711
712         /*
713          * the first call to ->target() should result in us actually
714          * writing something to the appropriate registers.
715          */
716         data->resume = 1;
717
718         return result;
719
720 err_freqfree:
721         kfree(data->freq_table);
722 err_unreg:
723         acpi_processor_unregister_performance(perf, cpu);
724 err_free:
725         kfree(data);
726         per_cpu(acfreq_data, cpu) = NULL;
727
728         return result;
729 }
730
731 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
732 {
733         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
734
735         pr_debug("acpi_cpufreq_cpu_exit\n");
736
737         if (data) {
738                 cpufreq_frequency_table_put_attr(policy->cpu);
739                 per_cpu(acfreq_data, policy->cpu) = NULL;
740                 acpi_processor_unregister_performance(data->acpi_data,
741                                                       policy->cpu);
742                 kfree(data->freq_table);
743                 kfree(data);
744         }
745
746         return 0;
747 }
748
749 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
750 {
751         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
752
753         pr_debug("acpi_cpufreq_resume\n");
754
755         data->resume = 1;
756
757         return 0;
758 }
759
760 static struct freq_attr *acpi_cpufreq_attr[] = {
761         &cpufreq_freq_attr_scaling_available_freqs,
762         NULL,
763 };
764
765 static struct cpufreq_driver acpi_cpufreq_driver = {
766         .verify         = acpi_cpufreq_verify,
767         .target         = acpi_cpufreq_target,
768         .bios_limit     = acpi_processor_get_bios_limit,
769         .init           = acpi_cpufreq_cpu_init,
770         .exit           = acpi_cpufreq_cpu_exit,
771         .resume         = acpi_cpufreq_resume,
772         .name           = "acpi-cpufreq",
773         .owner          = THIS_MODULE,
774         .attr           = acpi_cpufreq_attr,
775 };
776
777 static int __init acpi_cpufreq_init(void)
778 {
779         int ret;
780
781         if (acpi_disabled)
782                 return 0;
783
784         pr_debug("acpi_cpufreq_init\n");
785
786         ret = acpi_cpufreq_early_init();
787         if (ret)
788                 return ret;
789
790         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
791         if (ret)
792                 free_acpi_perf_data();
793
794         return ret;
795 }
796
797 static void __exit acpi_cpufreq_exit(void)
798 {
799         pr_debug("acpi_cpufreq_exit\n");
800
801         cpufreq_unregister_driver(&acpi_cpufreq_driver);
802
803         free_acpi_perf_data();
804 }
805
806 module_param(acpi_pstate_strict, uint, 0644);
807 MODULE_PARM_DESC(acpi_pstate_strict,
808         "value 0 or non-zero. non-zero -> strict ACPI checks are "
809         "performed during frequency changes.");
810
811 late_initcall(acpi_cpufreq_init);
812 module_exit(acpi_cpufreq_exit);
813
814 MODULE_ALIAS("acpi");