d0a001093b2d52d5583b6eb8fdf552978323e68d
[safe/jmp/linux-2.6] / arch / x86 / kernel / cpu / cpufreq / acpi-cpufreq.c
1 /*
2  * acpi-cpufreq.c - ACPI Processor P-States Driver ($Revision: 1.4 $)
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/ftrace.h>
37
38 #include <linux/acpi.h>
39 #include <acpi/processor.h>
40
41 #include <asm/io.h>
42 #include <asm/msr.h>
43 #include <asm/processor.h>
44 #include <asm/cpufeature.h>
45 #include <asm/delay.h>
46 #include <asm/uaccess.h>
47
48 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, "acpi-cpufreq", msg)
49
50 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
51 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
52 MODULE_LICENSE("GPL");
53
54 enum {
55         UNDEFINED_CAPABLE = 0,
56         SYSTEM_INTEL_MSR_CAPABLE,
57         SYSTEM_IO_CAPABLE,
58 };
59
60 #define INTEL_MSR_RANGE         (0xffff)
61 #define CPUID_6_ECX_APERFMPERF_CAPABILITY       (0x1)
62
63 struct acpi_cpufreq_data {
64         struct acpi_processor_performance *acpi_data;
65         struct cpufreq_frequency_table *freq_table;
66         unsigned int max_freq;
67         unsigned int resume;
68         unsigned int cpu_feature;
69 };
70
71 static DEFINE_PER_CPU(struct acpi_cpufreq_data *, drv_data);
72
73 /* acpi_perf_data is a pointer to percpu data. */
74 static struct acpi_processor_performance *acpi_perf_data;
75
76 static struct cpufreq_driver acpi_cpufreq_driver;
77
78 static unsigned int acpi_pstate_strict;
79
80 static int check_est_cpu(unsigned int cpuid)
81 {
82         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
83
84         if (cpu->x86_vendor != X86_VENDOR_INTEL ||
85             !cpu_has(cpu, X86_FEATURE_EST))
86                 return 0;
87
88         return 1;
89 }
90
91 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
92 {
93         struct acpi_processor_performance *perf;
94         int i;
95
96         perf = data->acpi_data;
97
98         for (i=0; i<perf->state_count; i++) {
99                 if (value == perf->states[i].status)
100                         return data->freq_table[i].frequency;
101         }
102         return 0;
103 }
104
105 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
106 {
107         int i;
108         struct acpi_processor_performance *perf;
109
110         msr &= INTEL_MSR_RANGE;
111         perf = data->acpi_data;
112
113         for (i=0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
114                 if (msr == perf->states[data->freq_table[i].index].status)
115                         return data->freq_table[i].frequency;
116         }
117         return data->freq_table[0].frequency;
118 }
119
120 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
121 {
122         switch (data->cpu_feature) {
123         case SYSTEM_INTEL_MSR_CAPABLE:
124                 return extract_msr(val, data);
125         case SYSTEM_IO_CAPABLE:
126                 return extract_io(val, data);
127         default:
128                 return 0;
129         }
130 }
131
132 struct msr_addr {
133         u32 reg;
134 };
135
136 struct io_addr {
137         u16 port;
138         u8 bit_width;
139 };
140
141 typedef union {
142         struct msr_addr msr;
143         struct io_addr io;
144 } drv_addr_union;
145
146 struct drv_cmd {
147         unsigned int type;
148         cpumask_t mask;
149         drv_addr_union addr;
150         u32 val;
151 };
152
153 static void do_drv_read(struct drv_cmd *cmd)
154 {
155         u32 h;
156
157         switch (cmd->type) {
158         case SYSTEM_INTEL_MSR_CAPABLE:
159                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
160                 break;
161         case SYSTEM_IO_CAPABLE:
162                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
163                                 &cmd->val,
164                                 (u32)cmd->addr.io.bit_width);
165                 break;
166         default:
167                 break;
168         }
169 }
170
171 static void do_drv_write(struct drv_cmd *cmd)
172 {
173         u32 lo, hi;
174
175         switch (cmd->type) {
176         case SYSTEM_INTEL_MSR_CAPABLE:
177                 rdmsr(cmd->addr.msr.reg, lo, hi);
178                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
179                 wrmsr(cmd->addr.msr.reg, lo, hi);
180                 break;
181         case SYSTEM_IO_CAPABLE:
182                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
183                                 cmd->val,
184                                 (u32)cmd->addr.io.bit_width);
185                 break;
186         default:
187                 break;
188         }
189 }
190
191 static void drv_read(struct drv_cmd *cmd)
192 {
193         cpumask_t saved_mask = current->cpus_allowed;
194         cmd->val = 0;
195
196         set_cpus_allowed_ptr(current, &cmd->mask);
197         do_drv_read(cmd);
198         set_cpus_allowed_ptr(current, &saved_mask);
199 }
200
201 static void drv_write(struct drv_cmd *cmd)
202 {
203         cpumask_t saved_mask = current->cpus_allowed;
204         unsigned int i;
205
206         for_each_cpu_mask_nr(i, cmd->mask) {
207                 set_cpus_allowed_ptr(current, &cpumask_of_cpu(i));
208                 do_drv_write(cmd);
209         }
210
211         set_cpus_allowed_ptr(current, &saved_mask);
212         return;
213 }
214
215 static u32 get_cur_val(const cpumask_t *mask)
216 {
217         struct acpi_processor_performance *perf;
218         struct drv_cmd cmd;
219
220         if (unlikely(cpus_empty(*mask)))
221                 return 0;
222
223         switch (per_cpu(drv_data, first_cpu(*mask))->cpu_feature) {
224         case SYSTEM_INTEL_MSR_CAPABLE:
225                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
226                 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
227                 break;
228         case SYSTEM_IO_CAPABLE:
229                 cmd.type = SYSTEM_IO_CAPABLE;
230                 perf = per_cpu(drv_data, first_cpu(*mask))->acpi_data;
231                 cmd.addr.io.port = perf->control_register.address;
232                 cmd.addr.io.bit_width = perf->control_register.bit_width;
233                 break;
234         default:
235                 return 0;
236         }
237
238         cmd.mask = *mask;
239
240         drv_read(&cmd);
241
242         dprintk("get_cur_val = %u\n", cmd.val);
243
244         return cmd.val;
245 }
246
247 /*
248  * Return the measured active (C0) frequency on this CPU since last call
249  * to this function.
250  * Input: cpu number
251  * Return: Average CPU frequency in terms of max frequency (zero on error)
252  *
253  * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
254  * over a period of time, while CPU is in C0 state.
255  * IA32_MPERF counts at the rate of max advertised frequency
256  * IA32_APERF counts at the rate of actual CPU frequency
257  * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
258  * no meaning should be associated with absolute values of these MSRs.
259  */
260 static unsigned int get_measured_perf(struct cpufreq_policy *policy,
261                                       unsigned int cpu)
262 {
263         union {
264                 struct {
265                         u32 lo;
266                         u32 hi;
267                 } split;
268                 u64 whole;
269         } aperf_cur, mperf_cur;
270
271         cpumask_t saved_mask;
272         unsigned int perf_percent;
273         unsigned int retval;
274
275         saved_mask = current->cpus_allowed;
276         set_cpus_allowed_ptr(current, &cpumask_of_cpu(cpu));
277         if (get_cpu() != cpu) {
278                 /* We were not able to run on requested processor */
279                 put_cpu();
280                 return 0;
281         }
282
283         rdmsr(MSR_IA32_APERF, aperf_cur.split.lo, aperf_cur.split.hi);
284         rdmsr(MSR_IA32_MPERF, mperf_cur.split.lo, mperf_cur.split.hi);
285
286         wrmsr(MSR_IA32_APERF, 0,0);
287         wrmsr(MSR_IA32_MPERF, 0,0);
288
289 #ifdef __i386__
290         /*
291          * We dont want to do 64 bit divide with 32 bit kernel
292          * Get an approximate value. Return failure in case we cannot get
293          * an approximate value.
294          */
295         if (unlikely(aperf_cur.split.hi || mperf_cur.split.hi)) {
296                 int shift_count;
297                 u32 h;
298
299                 h = max_t(u32, aperf_cur.split.hi, mperf_cur.split.hi);
300                 shift_count = fls(h);
301
302                 aperf_cur.whole >>= shift_count;
303                 mperf_cur.whole >>= shift_count;
304         }
305
306         if (((unsigned long)(-1) / 100) < aperf_cur.split.lo) {
307                 int shift_count = 7;
308                 aperf_cur.split.lo >>= shift_count;
309                 mperf_cur.split.lo >>= shift_count;
310         }
311
312         if (aperf_cur.split.lo && mperf_cur.split.lo)
313                 perf_percent = (aperf_cur.split.lo * 100) / mperf_cur.split.lo;
314         else
315                 perf_percent = 0;
316
317 #else
318         if (unlikely(((unsigned long)(-1) / 100) < aperf_cur.whole)) {
319                 int shift_count = 7;
320                 aperf_cur.whole >>= shift_count;
321                 mperf_cur.whole >>= shift_count;
322         }
323
324         if (aperf_cur.whole && mperf_cur.whole)
325                 perf_percent = (aperf_cur.whole * 100) / mperf_cur.whole;
326         else
327                 perf_percent = 0;
328
329 #endif
330
331         retval = per_cpu(drv_data, policy->cpu)->max_freq * perf_percent / 100;
332
333         put_cpu();
334         set_cpus_allowed_ptr(current, &saved_mask);
335
336         dprintk("cpu %d: performance percent %d\n", cpu, perf_percent);
337         return retval;
338 }
339
340 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
341 {
342         struct acpi_cpufreq_data *data = per_cpu(drv_data, cpu);
343         unsigned int freq;
344         unsigned int cached_freq;
345
346         dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
347
348         if (unlikely(data == NULL ||
349                      data->acpi_data == NULL || data->freq_table == NULL)) {
350                 return 0;
351         }
352
353         cached_freq = data->freq_table[data->acpi_data->state].frequency;
354         freq = extract_freq(get_cur_val(&cpumask_of_cpu(cpu)), data);
355         if (freq != cached_freq) {
356                 /*
357                  * The dreaded BIOS frequency change behind our back.
358                  * Force set the frequency on next target call.
359                  */
360                 data->resume = 1;
361         }
362
363         dprintk("cur freq = %u\n", freq);
364
365         return freq;
366 }
367
368 static unsigned int check_freqs(const cpumask_t *mask, unsigned int freq,
369                                 struct acpi_cpufreq_data *data)
370 {
371         unsigned int cur_freq;
372         unsigned int i;
373
374         for (i=0; i<100; i++) {
375                 cur_freq = extract_freq(get_cur_val(mask), data);
376                 if (cur_freq == freq)
377                         return 1;
378                 udelay(10);
379         }
380         return 0;
381 }
382
383 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
384                                unsigned int target_freq, unsigned int relation)
385 {
386         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
387         struct acpi_processor_performance *perf;
388         struct cpufreq_freqs freqs;
389         cpumask_t online_policy_cpus;
390         struct drv_cmd cmd;
391         unsigned int next_state = 0; /* Index into freq_table */
392         unsigned int next_perf_state = 0; /* Index into perf table */
393         unsigned int i;
394         int result = 0;
395         struct power_trace it;
396
397         dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
398
399         if (unlikely(data == NULL ||
400              data->acpi_data == NULL || data->freq_table == NULL)) {
401                 return -ENODEV;
402         }
403
404         perf = data->acpi_data;
405         result = cpufreq_frequency_table_target(policy,
406                                                 data->freq_table,
407                                                 target_freq,
408                                                 relation, &next_state);
409         if (unlikely(result))
410                 return -ENODEV;
411
412 #ifdef CONFIG_HOTPLUG_CPU
413         /* cpufreq holds the hotplug lock, so we are safe from here on */
414         cpus_and(online_policy_cpus, cpu_online_map, policy->cpus);
415 #else
416         online_policy_cpus = policy->cpus;
417 #endif
418
419         next_perf_state = data->freq_table[next_state].index;
420         if (perf->state == next_perf_state) {
421                 if (unlikely(data->resume)) {
422                         dprintk("Called after resume, resetting to P%d\n",
423                                 next_perf_state);
424                         data->resume = 0;
425                 } else {
426                         dprintk("Already at target state (P%d)\n",
427                                 next_perf_state);
428                         return 0;
429                 }
430         }
431
432         trace_power_mark(&it, POWER_PSTATE, next_perf_state);
433
434         switch (data->cpu_feature) {
435         case SYSTEM_INTEL_MSR_CAPABLE:
436                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
437                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
438                 cmd.val = (u32) perf->states[next_perf_state].control;
439                 break;
440         case SYSTEM_IO_CAPABLE:
441                 cmd.type = SYSTEM_IO_CAPABLE;
442                 cmd.addr.io.port = perf->control_register.address;
443                 cmd.addr.io.bit_width = perf->control_register.bit_width;
444                 cmd.val = (u32) perf->states[next_perf_state].control;
445                 break;
446         default:
447                 return -ENODEV;
448         }
449
450         cpus_clear(cmd.mask);
451
452         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
453                 cmd.mask = online_policy_cpus;
454         else
455                 cpu_set(policy->cpu, cmd.mask);
456
457         freqs.old = perf->states[perf->state].core_frequency * 1000;
458         freqs.new = data->freq_table[next_state].frequency;
459         for_each_cpu_mask_nr(i, cmd.mask) {
460                 freqs.cpu = i;
461                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
462         }
463
464         drv_write(&cmd);
465
466         if (acpi_pstate_strict) {
467                 if (!check_freqs(&cmd.mask, freqs.new, data)) {
468                         dprintk("acpi_cpufreq_target failed (%d)\n",
469                                 policy->cpu);
470                         return -EAGAIN;
471                 }
472         }
473
474         for_each_cpu_mask_nr(i, cmd.mask) {
475                 freqs.cpu = i;
476                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
477         }
478         perf->state = next_perf_state;
479
480         return result;
481 }
482
483 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
484 {
485         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
486
487         dprintk("acpi_cpufreq_verify\n");
488
489         return cpufreq_frequency_table_verify(policy, data->freq_table);
490 }
491
492 static unsigned long
493 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
494 {
495         struct acpi_processor_performance *perf = data->acpi_data;
496
497         if (cpu_khz) {
498                 /* search the closest match to cpu_khz */
499                 unsigned int i;
500                 unsigned long freq;
501                 unsigned long freqn = perf->states[0].core_frequency * 1000;
502
503                 for (i=0; i<(perf->state_count-1); i++) {
504                         freq = freqn;
505                         freqn = perf->states[i+1].core_frequency * 1000;
506                         if ((2 * cpu_khz) > (freqn + freq)) {
507                                 perf->state = i;
508                                 return freq;
509                         }
510                 }
511                 perf->state = perf->state_count-1;
512                 return freqn;
513         } else {
514                 /* assume CPU is at P0... */
515                 perf->state = 0;
516                 return perf->states[0].core_frequency * 1000;
517         }
518 }
519
520 static void free_acpi_perf_data(void)
521 {
522         unsigned int i;
523
524         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
525         for_each_possible_cpu(i)
526                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
527                                  ->shared_cpu_map);
528         free_percpu(acpi_perf_data);
529 }
530
531 /*
532  * acpi_cpufreq_early_init - initialize ACPI P-States library
533  *
534  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
535  * in order to determine correct frequency and voltage pairings. We can
536  * do _PDC and _PSD and find out the processor dependency for the
537  * actual init that will happen later...
538  */
539 static int __init acpi_cpufreq_early_init(void)
540 {
541         unsigned int i;
542         dprintk("acpi_cpufreq_early_init\n");
543
544         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
545         if (!acpi_perf_data) {
546                 dprintk("Memory allocation error for acpi_perf_data.\n");
547                 return -ENOMEM;
548         }
549         for_each_possible_cpu(i) {
550                 if (!alloc_cpumask_var(&per_cpu_ptr(acpi_perf_data, i)
551                                        ->shared_cpu_map, GFP_KERNEL)) {
552
553                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
554                         free_acpi_perf_data();
555                         return -ENOMEM;
556                 }
557         }
558
559         /* Do initialization in ACPI core */
560         acpi_processor_preregister_performance(acpi_perf_data);
561         return 0;
562 }
563
564 #ifdef CONFIG_SMP
565 /*
566  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
567  * or do it in BIOS firmware and won't inform about it to OS. If not
568  * detected, this has a side effect of making CPU run at a different speed
569  * than OS intended it to run at. Detect it and handle it cleanly.
570  */
571 static int bios_with_sw_any_bug;
572
573 static int sw_any_bug_found(const struct dmi_system_id *d)
574 {
575         bios_with_sw_any_bug = 1;
576         return 0;
577 }
578
579 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
580         {
581                 .callback = sw_any_bug_found,
582                 .ident = "Supermicro Server X6DLP",
583                 .matches = {
584                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
585                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
586                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
587                 },
588         },
589         { }
590 };
591 #endif
592
593 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
594 {
595         unsigned int i;
596         unsigned int valid_states = 0;
597         unsigned int cpu = policy->cpu;
598         struct acpi_cpufreq_data *data;
599         unsigned int result = 0;
600         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
601         struct acpi_processor_performance *perf;
602
603         dprintk("acpi_cpufreq_cpu_init\n");
604
605         data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
606         if (!data)
607                 return -ENOMEM;
608
609         data->acpi_data = percpu_ptr(acpi_perf_data, cpu);
610         per_cpu(drv_data, cpu) = data;
611
612         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
613                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
614
615         result = acpi_processor_register_performance(data->acpi_data, cpu);
616         if (result)
617                 goto err_free;
618
619         perf = data->acpi_data;
620         policy->shared_type = perf->shared_type;
621
622         /*
623          * Will let policy->cpus know about dependency only when software
624          * coordination is required.
625          */
626         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
627             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
628                 cpumask_copy(&policy->cpus, perf->shared_cpu_map);
629         }
630         cpumask_copy(&policy->related_cpus, perf->shared_cpu_map);
631
632 #ifdef CONFIG_SMP
633         dmi_check_system(sw_any_bug_dmi_table);
634         if (bios_with_sw_any_bug && cpus_weight(policy->cpus) == 1) {
635                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
636                 policy->cpus = per_cpu(cpu_core_map, cpu);
637         }
638 #endif
639
640         /* capability check */
641         if (perf->state_count <= 1) {
642                 dprintk("No P-States\n");
643                 result = -ENODEV;
644                 goto err_unreg;
645         }
646
647         if (perf->control_register.space_id != perf->status_register.space_id) {
648                 result = -ENODEV;
649                 goto err_unreg;
650         }
651
652         switch (perf->control_register.space_id) {
653         case ACPI_ADR_SPACE_SYSTEM_IO:
654                 dprintk("SYSTEM IO addr space\n");
655                 data->cpu_feature = SYSTEM_IO_CAPABLE;
656                 break;
657         case ACPI_ADR_SPACE_FIXED_HARDWARE:
658                 dprintk("HARDWARE addr space\n");
659                 if (!check_est_cpu(cpu)) {
660                         result = -ENODEV;
661                         goto err_unreg;
662                 }
663                 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
664                 break;
665         default:
666                 dprintk("Unknown addr space %d\n",
667                         (u32) (perf->control_register.space_id));
668                 result = -ENODEV;
669                 goto err_unreg;
670         }
671
672         data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
673                     (perf->state_count+1), GFP_KERNEL);
674         if (!data->freq_table) {
675                 result = -ENOMEM;
676                 goto err_unreg;
677         }
678
679         /* detect transition latency */
680         policy->cpuinfo.transition_latency = 0;
681         for (i=0; i<perf->state_count; i++) {
682                 if ((perf->states[i].transition_latency * 1000) >
683                     policy->cpuinfo.transition_latency)
684                         policy->cpuinfo.transition_latency =
685                             perf->states[i].transition_latency * 1000;
686         }
687
688         data->max_freq = perf->states[0].core_frequency * 1000;
689         /* table init */
690         for (i=0; i<perf->state_count; i++) {
691                 if (i>0 && perf->states[i].core_frequency >=
692                     data->freq_table[valid_states-1].frequency / 1000)
693                         continue;
694
695                 data->freq_table[valid_states].index = i;
696                 data->freq_table[valid_states].frequency =
697                     perf->states[i].core_frequency * 1000;
698                 valid_states++;
699         }
700         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
701         perf->state = 0;
702
703         result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
704         if (result)
705                 goto err_freqfree;
706
707         switch (perf->control_register.space_id) {
708         case ACPI_ADR_SPACE_SYSTEM_IO:
709                 /* Current speed is unknown and not detectable by IO port */
710                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
711                 break;
712         case ACPI_ADR_SPACE_FIXED_HARDWARE:
713                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
714                 policy->cur = get_cur_freq_on_cpu(cpu);
715                 break;
716         default:
717                 break;
718         }
719
720         /* notify BIOS that we exist */
721         acpi_processor_notify_smm(THIS_MODULE);
722
723         /* Check for APERF/MPERF support in hardware */
724         if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
725                 unsigned int ecx;
726                 ecx = cpuid_ecx(6);
727                 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
728                         acpi_cpufreq_driver.getavg = get_measured_perf;
729         }
730
731         dprintk("CPU%u - ACPI performance management activated.\n", cpu);
732         for (i = 0; i < perf->state_count; i++)
733                 dprintk("     %cP%d: %d MHz, %d mW, %d uS\n",
734                         (i == perf->state ? '*' : ' '), i,
735                         (u32) perf->states[i].core_frequency,
736                         (u32) perf->states[i].power,
737                         (u32) perf->states[i].transition_latency);
738
739         cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
740
741         /*
742          * the first call to ->target() should result in us actually
743          * writing something to the appropriate registers.
744          */
745         data->resume = 1;
746
747         return result;
748
749 err_freqfree:
750         kfree(data->freq_table);
751 err_unreg:
752         acpi_processor_unregister_performance(perf, cpu);
753 err_free:
754         kfree(data);
755         per_cpu(drv_data, cpu) = NULL;
756
757         return result;
758 }
759
760 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
761 {
762         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
763
764         dprintk("acpi_cpufreq_cpu_exit\n");
765
766         if (data) {
767                 cpufreq_frequency_table_put_attr(policy->cpu);
768                 per_cpu(drv_data, policy->cpu) = NULL;
769                 acpi_processor_unregister_performance(data->acpi_data,
770                                                       policy->cpu);
771                 kfree(data);
772         }
773
774         return 0;
775 }
776
777 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
778 {
779         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
780
781         dprintk("acpi_cpufreq_resume\n");
782
783         data->resume = 1;
784
785         return 0;
786 }
787
788 static struct freq_attr *acpi_cpufreq_attr[] = {
789         &cpufreq_freq_attr_scaling_available_freqs,
790         NULL,
791 };
792
793 static struct cpufreq_driver acpi_cpufreq_driver = {
794         .verify = acpi_cpufreq_verify,
795         .target = acpi_cpufreq_target,
796         .init = acpi_cpufreq_cpu_init,
797         .exit = acpi_cpufreq_cpu_exit,
798         .resume = acpi_cpufreq_resume,
799         .name = "acpi-cpufreq",
800         .owner = THIS_MODULE,
801         .attr = acpi_cpufreq_attr,
802 };
803
804 static int __init acpi_cpufreq_init(void)
805 {
806         int ret;
807
808         if (acpi_disabled)
809                 return 0;
810
811         dprintk("acpi_cpufreq_init\n");
812
813         ret = acpi_cpufreq_early_init();
814         if (ret)
815                 return ret;
816
817         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
818         if (ret)
819                 free_acpi_perf_data();
820
821         return ret;
822 }
823
824 static void __exit acpi_cpufreq_exit(void)
825 {
826         dprintk("acpi_cpufreq_exit\n");
827
828         cpufreq_unregister_driver(&acpi_cpufreq_driver);
829
830         free_percpu(acpi_perf_data);
831 }
832
833 module_param(acpi_pstate_strict, uint, 0644);
834 MODULE_PARM_DESC(acpi_pstate_strict,
835         "value 0 or non-zero. non-zero -> strict ACPI checks are "
836         "performed during frequency changes.");
837
838 late_initcall(acpi_cpufreq_init);
839 module_exit(acpi_cpufreq_exit);
840
841 MODULE_ALIAS("acpi");