Merge commit 'v2.6.29-rc1' into timers/urgent
[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_var_t mask;
149         drv_addr_union addr;
150         u32 val;
151 };
152
153 static long do_drv_read(void *_cmd)
154 {
155         struct drv_cmd *cmd = _cmd;
156         u32 h;
157
158         switch (cmd->type) {
159         case SYSTEM_INTEL_MSR_CAPABLE:
160                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
161                 break;
162         case SYSTEM_IO_CAPABLE:
163                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
164                                 &cmd->val,
165                                 (u32)cmd->addr.io.bit_width);
166                 break;
167         default:
168                 break;
169         }
170         return 0;
171 }
172
173 static long do_drv_write(void *_cmd)
174 {
175         struct drv_cmd *cmd = _cmd;
176         u32 lo, hi;
177
178         switch (cmd->type) {
179         case SYSTEM_INTEL_MSR_CAPABLE:
180                 rdmsr(cmd->addr.msr.reg, lo, hi);
181                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
182                 wrmsr(cmd->addr.msr.reg, lo, hi);
183                 break;
184         case SYSTEM_IO_CAPABLE:
185                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
186                                 cmd->val,
187                                 (u32)cmd->addr.io.bit_width);
188                 break;
189         default:
190                 break;
191         }
192         return 0;
193 }
194
195 static void drv_read(struct drv_cmd *cmd)
196 {
197         cmd->val = 0;
198
199         work_on_cpu(cpumask_any(cmd->mask), do_drv_read, cmd);
200 }
201
202 static void drv_write(struct drv_cmd *cmd)
203 {
204         unsigned int i;
205
206         for_each_cpu(i, cmd->mask) {
207                 work_on_cpu(i, do_drv_write, cmd);
208         }
209 }
210
211 static u32 get_cur_val(const struct cpumask *mask)
212 {
213         struct acpi_processor_performance *perf;
214         struct drv_cmd cmd;
215
216         if (unlikely(cpumask_empty(mask)))
217                 return 0;
218
219         switch (per_cpu(drv_data, cpumask_first(mask))->cpu_feature) {
220         case SYSTEM_INTEL_MSR_CAPABLE:
221                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
222                 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
223                 break;
224         case SYSTEM_IO_CAPABLE:
225                 cmd.type = SYSTEM_IO_CAPABLE;
226                 perf = per_cpu(drv_data, cpumask_first(mask))->acpi_data;
227                 cmd.addr.io.port = perf->control_register.address;
228                 cmd.addr.io.bit_width = perf->control_register.bit_width;
229                 break;
230         default:
231                 return 0;
232         }
233
234         if (unlikely(!alloc_cpumask_var(&cmd.mask, GFP_KERNEL)))
235                 return 0;
236
237         cpumask_copy(cmd.mask, mask);
238
239         drv_read(&cmd);
240
241         free_cpumask_var(cmd.mask);
242
243         dprintk("get_cur_val = %u\n", cmd.val);
244
245         return cmd.val;
246 }
247
248 struct perf_cur {
249         union {
250                 struct {
251                         u32 lo;
252                         u32 hi;
253                 } split;
254                 u64 whole;
255         } aperf_cur, mperf_cur;
256 };
257
258
259 static long read_measured_perf_ctrs(void *_cur)
260 {
261         struct perf_cur *cur = _cur;
262
263         rdmsr(MSR_IA32_APERF, cur->aperf_cur.split.lo, cur->aperf_cur.split.hi);
264         rdmsr(MSR_IA32_MPERF, cur->mperf_cur.split.lo, cur->mperf_cur.split.hi);
265
266         wrmsr(MSR_IA32_APERF, 0, 0);
267         wrmsr(MSR_IA32_MPERF, 0, 0);
268
269         return 0;
270 }
271
272 /*
273  * Return the measured active (C0) frequency on this CPU since last call
274  * to this function.
275  * Input: cpu number
276  * Return: Average CPU frequency in terms of max frequency (zero on error)
277  *
278  * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
279  * over a period of time, while CPU is in C0 state.
280  * IA32_MPERF counts at the rate of max advertised frequency
281  * IA32_APERF counts at the rate of actual CPU frequency
282  * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
283  * no meaning should be associated with absolute values of these MSRs.
284  */
285 static unsigned int get_measured_perf(struct cpufreq_policy *policy,
286                                       unsigned int cpu)
287 {
288         struct perf_cur cur;
289         unsigned int perf_percent;
290         unsigned int retval;
291
292         if (!work_on_cpu(cpu, read_measured_perf_ctrs, &cur))
293                 return 0;
294
295 #ifdef __i386__
296         /*
297          * We dont want to do 64 bit divide with 32 bit kernel
298          * Get an approximate value. Return failure in case we cannot get
299          * an approximate value.
300          */
301         if (unlikely(cur.aperf_cur.split.hi || cur.mperf_cur.split.hi)) {
302                 int shift_count;
303                 u32 h;
304
305                 h = max_t(u32, cur.aperf_cur.split.hi, cur.mperf_cur.split.hi);
306                 shift_count = fls(h);
307
308                 cur.aperf_cur.whole >>= shift_count;
309                 cur.mperf_cur.whole >>= shift_count;
310         }
311
312         if (((unsigned long)(-1) / 100) < cur.aperf_cur.split.lo) {
313                 int shift_count = 7;
314                 cur.aperf_cur.split.lo >>= shift_count;
315                 cur.mperf_cur.split.lo >>= shift_count;
316         }
317
318         if (cur.aperf_cur.split.lo && cur.mperf_cur.split.lo)
319                 perf_percent = (cur.aperf_cur.split.lo * 100) /
320                                 cur.mperf_cur.split.lo;
321         else
322                 perf_percent = 0;
323
324 #else
325         if (unlikely(((unsigned long)(-1) / 100) < cur.aperf_cur.whole)) {
326                 int shift_count = 7;
327                 cur.aperf_cur.whole >>= shift_count;
328                 cur.mperf_cur.whole >>= shift_count;
329         }
330
331         if (cur.aperf_cur.whole && cur.mperf_cur.whole)
332                 perf_percent = (cur.aperf_cur.whole * 100) /
333                                 cur.mperf_cur.whole;
334         else
335                 perf_percent = 0;
336
337 #endif
338
339         retval = per_cpu(drv_data, policy->cpu)->max_freq * perf_percent / 100;
340
341         return retval;
342 }
343
344 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
345 {
346         struct acpi_cpufreq_data *data = per_cpu(drv_data, cpu);
347         unsigned int freq;
348         unsigned int cached_freq;
349
350         dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
351
352         if (unlikely(data == NULL ||
353                      data->acpi_data == NULL || data->freq_table == NULL)) {
354                 return 0;
355         }
356
357         cached_freq = data->freq_table[data->acpi_data->state].frequency;
358         freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
359         if (freq != cached_freq) {
360                 /*
361                  * The dreaded BIOS frequency change behind our back.
362                  * Force set the frequency on next target call.
363                  */
364                 data->resume = 1;
365         }
366
367         dprintk("cur freq = %u\n", freq);
368
369         return freq;
370 }
371
372 static unsigned int check_freqs(const cpumask_t *mask, unsigned int freq,
373                                 struct acpi_cpufreq_data *data)
374 {
375         unsigned int cur_freq;
376         unsigned int i;
377
378         for (i=0; i<100; i++) {
379                 cur_freq = extract_freq(get_cur_val(mask), data);
380                 if (cur_freq == freq)
381                         return 1;
382                 udelay(10);
383         }
384         return 0;
385 }
386
387 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
388                                unsigned int target_freq, unsigned int relation)
389 {
390         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
391         struct acpi_processor_performance *perf;
392         struct cpufreq_freqs freqs;
393         struct drv_cmd cmd;
394         unsigned int next_state = 0; /* Index into freq_table */
395         unsigned int next_perf_state = 0; /* Index into perf table */
396         unsigned int i;
397         int result = 0;
398         struct power_trace it;
399
400         dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
401
402         if (unlikely(data == NULL ||
403              data->acpi_data == NULL || data->freq_table == NULL)) {
404                 return -ENODEV;
405         }
406
407         if (unlikely(!alloc_cpumask_var(&cmd.mask, GFP_KERNEL)))
408                 return -ENOMEM;
409
410         perf = data->acpi_data;
411         result = cpufreq_frequency_table_target(policy,
412                                                 data->freq_table,
413                                                 target_freq,
414                                                 relation, &next_state);
415         if (unlikely(result)) {
416                 result = -ENODEV;
417                 goto out;
418         }
419
420         next_perf_state = data->freq_table[next_state].index;
421         if (perf->state == next_perf_state) {
422                 if (unlikely(data->resume)) {
423                         dprintk("Called after resume, resetting to P%d\n",
424                                 next_perf_state);
425                         data->resume = 0;
426                 } else {
427                         dprintk("Already at target state (P%d)\n",
428                                 next_perf_state);
429                         goto out;
430                 }
431         }
432
433         trace_power_mark(&it, POWER_PSTATE, next_perf_state);
434
435         switch (data->cpu_feature) {
436         case SYSTEM_INTEL_MSR_CAPABLE:
437                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
438                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
439                 cmd.val = (u32) perf->states[next_perf_state].control;
440                 break;
441         case SYSTEM_IO_CAPABLE:
442                 cmd.type = SYSTEM_IO_CAPABLE;
443                 cmd.addr.io.port = perf->control_register.address;
444                 cmd.addr.io.bit_width = perf->control_register.bit_width;
445                 cmd.val = (u32) perf->states[next_perf_state].control;
446                 break;
447         default:
448                 result = -ENODEV;
449                 goto out;
450         }
451
452         /* cpufreq holds the hotplug lock, so we are safe from here on */
453         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
454                 cpumask_and(cmd.mask, cpu_online_mask, policy->cpus);
455         else
456                 cpumask_copy(cmd.mask, cpumask_of(policy->cpu));
457
458         freqs.old = perf->states[perf->state].core_frequency * 1000;
459         freqs.new = data->freq_table[next_state].frequency;
460         for_each_cpu(i, cmd.mask) {
461                 freqs.cpu = i;
462                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
463         }
464
465         drv_write(&cmd);
466
467         if (acpi_pstate_strict) {
468                 if (!check_freqs(cmd.mask, freqs.new, data)) {
469                         dprintk("acpi_cpufreq_target failed (%d)\n",
470                                 policy->cpu);
471                         result = -EAGAIN;
472                         goto out;
473                 }
474         }
475
476         for_each_cpu(i, cmd.mask) {
477                 freqs.cpu = i;
478                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
479         }
480         perf->state = next_perf_state;
481
482 out:
483         free_cpumask_var(cmd.mask);
484         return result;
485 }
486
487 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
488 {
489         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
490
491         dprintk("acpi_cpufreq_verify\n");
492
493         return cpufreq_frequency_table_verify(policy, data->freq_table);
494 }
495
496 static unsigned long
497 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
498 {
499         struct acpi_processor_performance *perf = data->acpi_data;
500
501         if (cpu_khz) {
502                 /* search the closest match to cpu_khz */
503                 unsigned int i;
504                 unsigned long freq;
505                 unsigned long freqn = perf->states[0].core_frequency * 1000;
506
507                 for (i=0; i<(perf->state_count-1); i++) {
508                         freq = freqn;
509                         freqn = perf->states[i+1].core_frequency * 1000;
510                         if ((2 * cpu_khz) > (freqn + freq)) {
511                                 perf->state = i;
512                                 return freq;
513                         }
514                 }
515                 perf->state = perf->state_count-1;
516                 return freqn;
517         } else {
518                 /* assume CPU is at P0... */
519                 perf->state = 0;
520                 return perf->states[0].core_frequency * 1000;
521         }
522 }
523
524 static void free_acpi_perf_data(void)
525 {
526         unsigned int i;
527
528         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
529         for_each_possible_cpu(i)
530                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
531                                  ->shared_cpu_map);
532         free_percpu(acpi_perf_data);
533 }
534
535 /*
536  * acpi_cpufreq_early_init - initialize ACPI P-States library
537  *
538  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
539  * in order to determine correct frequency and voltage pairings. We can
540  * do _PDC and _PSD and find out the processor dependency for the
541  * actual init that will happen later...
542  */
543 static int __init acpi_cpufreq_early_init(void)
544 {
545         unsigned int i;
546         dprintk("acpi_cpufreq_early_init\n");
547
548         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
549         if (!acpi_perf_data) {
550                 dprintk("Memory allocation error for acpi_perf_data.\n");
551                 return -ENOMEM;
552         }
553         for_each_possible_cpu(i) {
554                 if (!alloc_cpumask_var_node(
555                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
556                         GFP_KERNEL, cpu_to_node(i))) {
557
558                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
559                         free_acpi_perf_data();
560                         return -ENOMEM;
561                 }
562         }
563
564         /* Do initialization in ACPI core */
565         acpi_processor_preregister_performance(acpi_perf_data);
566         return 0;
567 }
568
569 #ifdef CONFIG_SMP
570 /*
571  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
572  * or do it in BIOS firmware and won't inform about it to OS. If not
573  * detected, this has a side effect of making CPU run at a different speed
574  * than OS intended it to run at. Detect it and handle it cleanly.
575  */
576 static int bios_with_sw_any_bug;
577
578 static int sw_any_bug_found(const struct dmi_system_id *d)
579 {
580         bios_with_sw_any_bug = 1;
581         return 0;
582 }
583
584 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
585         {
586                 .callback = sw_any_bug_found,
587                 .ident = "Supermicro Server X6DLP",
588                 .matches = {
589                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
590                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
591                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
592                 },
593         },
594         { }
595 };
596 #endif
597
598 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
599 {
600         unsigned int i;
601         unsigned int valid_states = 0;
602         unsigned int cpu = policy->cpu;
603         struct acpi_cpufreq_data *data;
604         unsigned int result = 0;
605         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
606         struct acpi_processor_performance *perf;
607
608         dprintk("acpi_cpufreq_cpu_init\n");
609
610         data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
611         if (!data)
612                 return -ENOMEM;
613
614         data->acpi_data = percpu_ptr(acpi_perf_data, cpu);
615         per_cpu(drv_data, cpu) = data;
616
617         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
618                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
619
620         result = acpi_processor_register_performance(data->acpi_data, cpu);
621         if (result)
622                 goto err_free;
623
624         perf = data->acpi_data;
625         policy->shared_type = perf->shared_type;
626
627         /*
628          * Will let policy->cpus know about dependency only when software
629          * coordination is required.
630          */
631         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
632             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
633                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
634         }
635         cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
636
637 #ifdef CONFIG_SMP
638         dmi_check_system(sw_any_bug_dmi_table);
639         if (bios_with_sw_any_bug && cpumask_weight(policy->cpus) == 1) {
640                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
641                 cpumask_copy(policy->cpus, cpu_core_mask(cpu));
642         }
643 #endif
644
645         /* capability check */
646         if (perf->state_count <= 1) {
647                 dprintk("No P-States\n");
648                 result = -ENODEV;
649                 goto err_unreg;
650         }
651
652         if (perf->control_register.space_id != perf->status_register.space_id) {
653                 result = -ENODEV;
654                 goto err_unreg;
655         }
656
657         switch (perf->control_register.space_id) {
658         case ACPI_ADR_SPACE_SYSTEM_IO:
659                 dprintk("SYSTEM IO addr space\n");
660                 data->cpu_feature = SYSTEM_IO_CAPABLE;
661                 break;
662         case ACPI_ADR_SPACE_FIXED_HARDWARE:
663                 dprintk("HARDWARE addr space\n");
664                 if (!check_est_cpu(cpu)) {
665                         result = -ENODEV;
666                         goto err_unreg;
667                 }
668                 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
669                 break;
670         default:
671                 dprintk("Unknown addr space %d\n",
672                         (u32) (perf->control_register.space_id));
673                 result = -ENODEV;
674                 goto err_unreg;
675         }
676
677         data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
678                     (perf->state_count+1), GFP_KERNEL);
679         if (!data->freq_table) {
680                 result = -ENOMEM;
681                 goto err_unreg;
682         }
683
684         /* detect transition latency */
685         policy->cpuinfo.transition_latency = 0;
686         for (i=0; i<perf->state_count; i++) {
687                 if ((perf->states[i].transition_latency * 1000) >
688                     policy->cpuinfo.transition_latency)
689                         policy->cpuinfo.transition_latency =
690                             perf->states[i].transition_latency * 1000;
691         }
692
693         data->max_freq = perf->states[0].core_frequency * 1000;
694         /* table init */
695         for (i=0; i<perf->state_count; i++) {
696                 if (i>0 && perf->states[i].core_frequency >=
697                     data->freq_table[valid_states-1].frequency / 1000)
698                         continue;
699
700                 data->freq_table[valid_states].index = i;
701                 data->freq_table[valid_states].frequency =
702                     perf->states[i].core_frequency * 1000;
703                 valid_states++;
704         }
705         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
706         perf->state = 0;
707
708         result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
709         if (result)
710                 goto err_freqfree;
711
712         switch (perf->control_register.space_id) {
713         case ACPI_ADR_SPACE_SYSTEM_IO:
714                 /* Current speed is unknown and not detectable by IO port */
715                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
716                 break;
717         case ACPI_ADR_SPACE_FIXED_HARDWARE:
718                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
719                 policy->cur = get_cur_freq_on_cpu(cpu);
720                 break;
721         default:
722                 break;
723         }
724
725         /* notify BIOS that we exist */
726         acpi_processor_notify_smm(THIS_MODULE);
727
728         /* Check for APERF/MPERF support in hardware */
729         if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
730                 unsigned int ecx;
731                 ecx = cpuid_ecx(6);
732                 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
733                         acpi_cpufreq_driver.getavg = get_measured_perf;
734         }
735
736         dprintk("CPU%u - ACPI performance management activated.\n", cpu);
737         for (i = 0; i < perf->state_count; i++)
738                 dprintk("     %cP%d: %d MHz, %d mW, %d uS\n",
739                         (i == perf->state ? '*' : ' '), i,
740                         (u32) perf->states[i].core_frequency,
741                         (u32) perf->states[i].power,
742                         (u32) perf->states[i].transition_latency);
743
744         cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
745
746         /*
747          * the first call to ->target() should result in us actually
748          * writing something to the appropriate registers.
749          */
750         data->resume = 1;
751
752         return result;
753
754 err_freqfree:
755         kfree(data->freq_table);
756 err_unreg:
757         acpi_processor_unregister_performance(perf, cpu);
758 err_free:
759         kfree(data);
760         per_cpu(drv_data, cpu) = NULL;
761
762         return result;
763 }
764
765 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
766 {
767         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
768
769         dprintk("acpi_cpufreq_cpu_exit\n");
770
771         if (data) {
772                 cpufreq_frequency_table_put_attr(policy->cpu);
773                 per_cpu(drv_data, policy->cpu) = NULL;
774                 acpi_processor_unregister_performance(data->acpi_data,
775                                                       policy->cpu);
776                 kfree(data);
777         }
778
779         return 0;
780 }
781
782 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
783 {
784         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
785
786         dprintk("acpi_cpufreq_resume\n");
787
788         data->resume = 1;
789
790         return 0;
791 }
792
793 static struct freq_attr *acpi_cpufreq_attr[] = {
794         &cpufreq_freq_attr_scaling_available_freqs,
795         NULL,
796 };
797
798 static struct cpufreq_driver acpi_cpufreq_driver = {
799         .verify = acpi_cpufreq_verify,
800         .target = acpi_cpufreq_target,
801         .init = acpi_cpufreq_cpu_init,
802         .exit = acpi_cpufreq_cpu_exit,
803         .resume = acpi_cpufreq_resume,
804         .name = "acpi-cpufreq",
805         .owner = THIS_MODULE,
806         .attr = acpi_cpufreq_attr,
807 };
808
809 static int __init acpi_cpufreq_init(void)
810 {
811         int ret;
812
813         if (acpi_disabled)
814                 return 0;
815
816         dprintk("acpi_cpufreq_init\n");
817
818         ret = acpi_cpufreq_early_init();
819         if (ret)
820                 return ret;
821
822         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
823         if (ret)
824                 free_acpi_perf_data();
825
826         return ret;
827 }
828
829 static void __exit acpi_cpufreq_exit(void)
830 {
831         dprintk("acpi_cpufreq_exit\n");
832
833         cpufreq_unregister_driver(&acpi_cpufreq_driver);
834
835         free_percpu(acpi_perf_data);
836 }
837
838 module_param(acpi_pstate_strict, uint, 0644);
839 MODULE_PARM_DESC(acpi_pstate_strict,
840         "value 0 or non-zero. non-zero -> strict ACPI checks are "
841         "performed during frequency changes.");
842
843 late_initcall(acpi_cpufreq_init);
844 module_exit(acpi_cpufreq_exit);
845
846 MODULE_ALIAS("acpi");