/home/lenb/linux-2.6 branch 'acpi-2.6.12'
[linux-2.6] / drivers / acpi / processor_idle.c
1 /*
2  * processor_idle - idle state submodule to the ACPI processor 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) 2004       Dominik Brodowski <linux@brodo.de>
7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
8  *                      - Added processor hotplug support
9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
10  *                      - Added support for C3 on SMP
11  *
12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
13  *
14  *  This program is free software; you can redistribute it and/or modify
15  *  it under the terms of the GNU General Public License as published by
16  *  the Free Software Foundation; either version 2 of the License, or (at
17  *  your option) any later version.
18  *
19  *  This program is distributed in the hope that it will be useful, but
20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
22  *  General Public License for more details.
23  *
24  *  You should have received a copy of the GNU General Public License along
25  *  with this program; if not, write to the Free Software Foundation, Inc.,
26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
27  *
28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
29  */
30
31 #include <linux/kernel.h>
32 #include <linux/module.h>
33 #include <linux/init.h>
34 #include <linux/cpufreq.h>
35 #include <linux/proc_fs.h>
36 #include <linux/seq_file.h>
37 #include <linux/acpi.h>
38 #include <linux/dmi.h>
39 #include <linux/moduleparam.h>
40
41 #include <asm/io.h>
42 #include <asm/uaccess.h>
43
44 #include <acpi/acpi_bus.h>
45 #include <acpi/processor.h>
46
47 #define ACPI_PROCESSOR_COMPONENT        0x01000000
48 #define ACPI_PROCESSOR_CLASS            "processor"
49 #define ACPI_PROCESSOR_DRIVER_NAME      "ACPI Processor Driver"
50 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
51 ACPI_MODULE_NAME                ("acpi_processor")
52
53 #define ACPI_PROCESSOR_FILE_POWER       "power"
54
55 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
56 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
57 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
58
59 static void (*pm_idle_save)(void);
60 module_param(max_cstate, uint, 0644);
61
62 static unsigned int nocst = 0;
63 module_param(nocst, uint, 0000);
64
65 /*
66  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
67  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
68  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
69  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
70  * reduce history for more aggressive entry into C3
71  */
72 static unsigned int bm_history = (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
73 module_param(bm_history, uint, 0644);
74 /* --------------------------------------------------------------------------
75                                 Power Management
76    -------------------------------------------------------------------------- */
77
78 /*
79  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
80  * For now disable this. Probably a bug somewhere else.
81  *
82  * To skip this limit, boot/load with a large max_cstate limit.
83  */
84 static int no_c2c3(struct dmi_system_id *id)
85 {
86         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
87                 return 0;
88
89         printk(KERN_NOTICE PREFIX "%s detected - C2,C3 disabled."
90                 " Override with \"processor.max_cstate=%d\"\n", id->ident,
91                ACPI_PROCESSOR_MAX_POWER + 1);
92
93         max_cstate = 1;
94
95         return 0;
96 }
97
98
99
100
101 static struct dmi_system_id __initdata processor_power_dmi_table[] = {
102         { no_c2c3, "IBM ThinkPad R40e", {
103           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
104           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }},
105         { no_c2c3, "Medion 41700", {
106           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
107           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J") }},
108         {},
109 };
110
111
112 static inline u32
113 ticks_elapsed (
114         u32                     t1,
115         u32                     t2)
116 {
117         if (t2 >= t1)
118                 return (t2 - t1);
119         else if (!acpi_fadt.tmr_val_ext)
120                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
121         else
122                 return ((0xFFFFFFFF - t1) + t2);
123 }
124
125
126 static void
127 acpi_processor_power_activate (
128         struct acpi_processor   *pr,
129         struct acpi_processor_cx  *new)
130 {
131         struct acpi_processor_cx  *old;
132
133         if (!pr || !new)
134                 return;
135
136         old = pr->power.state;
137
138         if (old)
139                 old->promotion.count = 0;
140         new->demotion.count = 0;
141
142         /* Cleanup from old state. */
143         if (old) {
144                 switch (old->type) {
145                 case ACPI_STATE_C3:
146                         /* Disable bus master reload */
147                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
148                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0, ACPI_MTX_DO_NOT_LOCK);
149                         break;
150                 }
151         }
152
153         /* Prepare to use new state. */
154         switch (new->type) {
155         case ACPI_STATE_C3:
156                 /* Enable bus master reload */
157                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
158                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1, ACPI_MTX_DO_NOT_LOCK);
159                 break;
160         }
161
162         pr->power.state = new;
163
164         return;
165 }
166
167
168 static atomic_t         c3_cpu_count;
169
170
171 static void acpi_processor_idle (void)
172 {
173         struct acpi_processor   *pr = NULL;
174         struct acpi_processor_cx *cx = NULL;
175         struct acpi_processor_cx *next_state = NULL;
176         int                     sleep_ticks = 0;
177         u32                     t1, t2 = 0;
178
179         pr = processors[raw_smp_processor_id()];
180         if (!pr)
181                 return;
182
183         /*
184          * Interrupts must be disabled during bus mastering calculations and
185          * for C2/C3 transitions.
186          */
187         local_irq_disable();
188
189         /*
190          * Check whether we truly need to go idle, or should
191          * reschedule:
192          */
193         if (unlikely(need_resched())) {
194                 local_irq_enable();
195                 return;
196         }
197
198         cx = pr->power.state;
199         if (!cx)
200                 goto easy_out;
201
202         /*
203          * Check BM Activity
204          * -----------------
205          * Check for bus mastering activity (if required), record, and check
206          * for demotion.
207          */
208         if (pr->flags.bm_check) {
209                 u32             bm_status = 0;
210                 unsigned long   diff = jiffies - pr->power.bm_check_timestamp;
211
212                 if (diff > 32)
213                         diff = 32;
214
215                 while (diff) {
216                         /* if we didn't get called, assume there was busmaster activity */
217                         diff--;
218                         if (diff)
219                                 pr->power.bm_activity |= 0x1;
220                         pr->power.bm_activity <<= 1;
221                 }
222
223                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS,
224                         &bm_status, ACPI_MTX_DO_NOT_LOCK);
225                 if (bm_status) {
226                         pr->power.bm_activity++;
227                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS,
228                                 1, ACPI_MTX_DO_NOT_LOCK);
229                 }
230                 /*
231                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
232                  * the true state of bus mastering activity; forcing us to
233                  * manually check the BMIDEA bit of each IDE channel.
234                  */
235                 else if (errata.piix4.bmisx) {
236                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
237                                 || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
238                                 pr->power.bm_activity++;
239                 }
240
241                 pr->power.bm_check_timestamp = jiffies;
242
243                 /*
244                  * Apply bus mastering demotion policy.  Automatically demote
245                  * to avoid a faulty transition.  Note that the processor
246                  * won't enter a low-power state during this call (to this
247                  * funciton) but should upon the next.
248                  *
249                  * TBD: A better policy might be to fallback to the demotion
250                  *      state (use it for this quantum only) istead of
251                  *      demoting -- and rely on duration as our sole demotion
252                  *      qualification.  This may, however, introduce DMA
253                  *      issues (e.g. floppy DMA transfer overrun/underrun).
254                  */
255                 if (pr->power.bm_activity & cx->demotion.threshold.bm) {
256                         local_irq_enable();
257                         next_state = cx->demotion.state;
258                         goto end;
259                 }
260         }
261
262         cx->usage++;
263
264         /*
265          * Sleep:
266          * ------
267          * Invoke the current Cx state to put the processor to sleep.
268          */
269         switch (cx->type) {
270
271         case ACPI_STATE_C1:
272                 /*
273                  * Invoke C1.
274                  * Use the appropriate idle routine, the one that would
275                  * be used without acpi C-states.
276                  */
277                 if (pm_idle_save)
278                         pm_idle_save();
279                 else
280                         safe_halt();
281                 /*
282                  * TBD: Can't get time duration while in C1, as resumes
283                  *      go to an ISR rather than here.  Need to instrument
284                  *      base interrupt handler.
285                  */
286                 sleep_ticks = 0xFFFFFFFF;
287                 break;
288
289         case ACPI_STATE_C2:
290                 /* Get start time (ticks) */
291                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
292                 /* Invoke C2 */
293                 inb(cx->address);
294                 /* Dummy op - must do something useless after P_LVL2 read */
295                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
296                 /* Get end time (ticks) */
297                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
298                 /* Re-enable interrupts */
299                 local_irq_enable();
300                 /* Compute time (ticks) that we were actually asleep */
301                 sleep_ticks = ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
302                 break;
303
304         case ACPI_STATE_C3:
305                 
306                 if (pr->flags.bm_check) {
307                         if (atomic_inc_return(&c3_cpu_count) ==
308                                         num_online_cpus()) {
309                                 /*
310                                  * All CPUs are trying to go to C3
311                                  * Disable bus master arbitration
312                                  */
313                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
314                                         ACPI_MTX_DO_NOT_LOCK);
315                         }
316                 } else {
317                         /* SMP with no shared cache... Invalidate cache  */
318                         ACPI_FLUSH_CPU_CACHE();
319                 }
320                 
321                 /* Get start time (ticks) */
322                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
323                 /* Invoke C3 */
324                 inb(cx->address);
325                 /* Dummy op - must do something useless after P_LVL3 read */
326                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
327                 /* Get end time (ticks) */
328                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
329                 if (pr->flags.bm_check) {
330                         /* Enable bus master arbitration */
331                         atomic_dec(&c3_cpu_count);
332                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0, ACPI_MTX_DO_NOT_LOCK);
333                 }
334
335                 /* Re-enable interrupts */
336                 local_irq_enable();
337                 /* Compute time (ticks) that we were actually asleep */
338                 sleep_ticks = ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
339                 break;
340
341         default:
342                 local_irq_enable();
343                 return;
344         }
345
346         next_state = pr->power.state;
347
348         /*
349          * Promotion?
350          * ----------
351          * Track the number of longs (time asleep is greater than threshold)
352          * and promote when the count threshold is reached.  Note that bus
353          * mastering activity may prevent promotions.
354          * Do not promote above max_cstate.
355          */
356         if (cx->promotion.state &&
357             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
358                 if (sleep_ticks > cx->promotion.threshold.ticks) {
359                         cx->promotion.count++;
360                         cx->demotion.count = 0;
361                         if (cx->promotion.count >= cx->promotion.threshold.count) {
362                                 if (pr->flags.bm_check) {
363                                         if (!(pr->power.bm_activity & cx->promotion.threshold.bm)) {
364                                                 next_state = cx->promotion.state;
365                                                 goto end;
366                                         }
367                                 }
368                                 else {
369                                         next_state = cx->promotion.state;
370                                         goto end;
371                                 }
372                         }
373                 }
374         }
375
376         /*
377          * Demotion?
378          * ---------
379          * Track the number of shorts (time asleep is less than time threshold)
380          * and demote when the usage threshold is reached.
381          */
382         if (cx->demotion.state) {
383                 if (sleep_ticks < cx->demotion.threshold.ticks) {
384                         cx->demotion.count++;
385                         cx->promotion.count = 0;
386                         if (cx->demotion.count >= cx->demotion.threshold.count) {
387                                 next_state = cx->demotion.state;
388                                 goto end;
389                         }
390                 }
391         }
392
393 end:
394         /*
395          * Demote if current state exceeds max_cstate
396          */
397         if ((pr->power.state - pr->power.states) > max_cstate) {
398                 if (cx->demotion.state)
399                         next_state = cx->demotion.state;
400         }
401
402         /*
403          * New Cx State?
404          * -------------
405          * If we're going to start using a new Cx state we must clean up
406          * from the previous and prepare to use the new.
407          */
408         if (next_state != pr->power.state)
409                 acpi_processor_power_activate(pr, next_state);
410
411         return;
412
413  easy_out:
414         /* do C1 instead of busy loop */
415         if (pm_idle_save)
416                 pm_idle_save();
417         else
418                 safe_halt();
419         return;
420 }
421
422
423 static int
424 acpi_processor_set_power_policy (
425         struct acpi_processor   *pr)
426 {
427         unsigned int i;
428         unsigned int state_is_set = 0;
429         struct acpi_processor_cx *lower = NULL;
430         struct acpi_processor_cx *higher = NULL;
431         struct acpi_processor_cx *cx;
432
433         ACPI_FUNCTION_TRACE("acpi_processor_set_power_policy");
434
435         if (!pr)
436                 return_VALUE(-EINVAL);
437
438         /*
439          * This function sets the default Cx state policy (OS idle handler).
440          * Our scheme is to promote quickly to C2 but more conservatively
441          * to C3.  We're favoring C2  for its characteristics of low latency
442          * (quick response), good power savings, and ability to allow bus
443          * mastering activity.  Note that the Cx state policy is completely
444          * customizable and can be altered dynamically.
445          */
446
447         /* startup state */
448         for (i=1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
449                 cx = &pr->power.states[i];
450                 if (!cx->valid)
451                         continue;
452
453                 if (!state_is_set)
454                         pr->power.state = cx;
455                 state_is_set++;
456                 break;
457         }
458
459         if (!state_is_set)
460                 return_VALUE(-ENODEV);
461
462         /* demotion */
463         for (i=1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
464                 cx = &pr->power.states[i];
465                 if (!cx->valid)
466                         continue;
467
468                 if (lower) {
469                         cx->demotion.state = lower;
470                         cx->demotion.threshold.ticks = cx->latency_ticks;
471                         cx->demotion.threshold.count = 1;
472                         if (cx->type == ACPI_STATE_C3)
473                                 cx->demotion.threshold.bm = bm_history;
474                 }
475
476                 lower = cx;
477         }
478
479         /* promotion */
480         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
481                 cx = &pr->power.states[i];
482                 if (!cx->valid)
483                         continue;
484
485                 if (higher) {
486                         cx->promotion.state  = higher;
487                         cx->promotion.threshold.ticks = cx->latency_ticks;
488                         if (cx->type >= ACPI_STATE_C2)
489                                 cx->promotion.threshold.count = 4;
490                         else
491                                 cx->promotion.threshold.count = 10;
492                         if (higher->type == ACPI_STATE_C3)
493                                 cx->promotion.threshold.bm = bm_history;
494                 }
495
496                 higher = cx;
497         }
498
499         return_VALUE(0);
500 }
501
502
503 static int acpi_processor_get_power_info_fadt (struct acpi_processor *pr)
504 {
505         int i;
506
507         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_fadt");
508
509         if (!pr)
510                 return_VALUE(-EINVAL);
511
512         if (!pr->pblk)
513                 return_VALUE(-ENODEV);
514
515         for (i = 0; i < ACPI_PROCESSOR_MAX_POWER; i++)
516                 memset(pr->power.states, 0, sizeof(struct acpi_processor_cx));
517
518         /* if info is obtained from pblk/fadt, type equals state */
519         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
520         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
521         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
522
523         /* the C0 state only exists as a filler in our array,
524          * and all processors need to support C1 */
525         pr->power.states[ACPI_STATE_C0].valid = 1;
526         pr->power.states[ACPI_STATE_C1].valid = 1;
527
528         /* determine C2 and C3 address from pblk */
529         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
530         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
531
532         /* determine latencies from FADT */
533         pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
534         pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
535
536         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
537                           "lvl2[0x%08x] lvl3[0x%08x]\n",
538                           pr->power.states[ACPI_STATE_C2].address,
539                           pr->power.states[ACPI_STATE_C3].address));
540
541         return_VALUE(0);
542 }
543
544
545 static int acpi_processor_get_power_info_default_c1 (struct acpi_processor *pr)
546 {
547         int i;
548
549         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_default_c1");
550
551         for (i = 0; i < ACPI_PROCESSOR_MAX_POWER; i++)
552                 memset(pr->power.states, 0, sizeof(struct acpi_processor_cx));
553
554         /* if info is obtained from pblk/fadt, type equals state */
555         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
556         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
557         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
558
559         /* the C0 state only exists as a filler in our array,
560          * and all processors need to support C1 */
561         pr->power.states[ACPI_STATE_C0].valid = 1;
562         pr->power.states[ACPI_STATE_C1].valid = 1;
563
564         return_VALUE(0);
565 }
566
567
568 static int acpi_processor_get_power_info_cst (struct acpi_processor *pr)
569 {
570         acpi_status             status = 0;
571         acpi_integer            count;
572         int                     i;
573         struct acpi_buffer      buffer = {ACPI_ALLOCATE_BUFFER, NULL};
574         union acpi_object       *cst;
575
576         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_cst");
577
578         if (nocst)
579                 return_VALUE(-ENODEV);
580
581         pr->power.count = 0;
582         for (i = 0; i < ACPI_PROCESSOR_MAX_POWER; i++)
583                 memset(pr->power.states, 0, sizeof(struct acpi_processor_cx));
584
585         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
586         if (ACPI_FAILURE(status)) {
587                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
588                 return_VALUE(-ENODEV);
589         }
590
591         cst = (union acpi_object *) buffer.pointer;
592
593         /* There must be at least 2 elements */
594         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
595                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR, "not enough elements in _CST\n"));
596                 status = -EFAULT;
597                 goto end;
598         }
599
600         count = cst->package.elements[0].integer.value;
601
602         /* Validate number of power states. */
603         if (count < 1 || count != cst->package.count - 1) {
604                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR, "count given by _CST is not valid\n"));
605                 status = -EFAULT;
606                 goto end;
607         }
608
609         /* We support up to ACPI_PROCESSOR_MAX_POWER. */
610         if (count > ACPI_PROCESSOR_MAX_POWER) {
611                 printk(KERN_WARNING "Limiting number of power states to max (%d)\n", ACPI_PROCESSOR_MAX_POWER);
612                 printk(KERN_WARNING "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
613                 count = ACPI_PROCESSOR_MAX_POWER;
614         }
615
616         /* Tell driver that at least _CST is supported. */
617         pr->flags.has_cst = 1;
618
619         for (i = 1; i <= count; i++) {
620                 union acpi_object *element;
621                 union acpi_object *obj;
622                 struct acpi_power_register *reg;
623                 struct acpi_processor_cx cx;
624
625                 memset(&cx, 0, sizeof(cx));
626
627                 element = (union acpi_object *) &(cst->package.elements[i]);
628                 if (element->type != ACPI_TYPE_PACKAGE)
629                         continue;
630
631                 if (element->package.count != 4)
632                         continue;
633
634                 obj = (union acpi_object *) &(element->package.elements[0]);
635
636                 if (obj->type != ACPI_TYPE_BUFFER)
637                         continue;
638
639                 reg = (struct acpi_power_register *) obj->buffer.pointer;
640
641                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
642                         (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
643                         continue;
644
645                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
646                         0 : reg->address;
647
648                 /* There should be an easy way to extract an integer... */
649                 obj = (union acpi_object *) &(element->package.elements[1]);
650                 if (obj->type != ACPI_TYPE_INTEGER)
651                         continue;
652
653                 cx.type = obj->integer.value;
654
655                 if ((cx.type != ACPI_STATE_C1) &&
656                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
657                         continue;
658
659                 if ((cx.type < ACPI_STATE_C1) ||
660                     (cx.type > ACPI_STATE_C3))
661                         continue;
662
663                 obj = (union acpi_object *) &(element->package.elements[2]);
664                 if (obj->type != ACPI_TYPE_INTEGER)
665                         continue;
666
667                 cx.latency = obj->integer.value;
668
669                 obj = (union acpi_object *) &(element->package.elements[3]);
670                 if (obj->type != ACPI_TYPE_INTEGER)
671                         continue;
672
673                 cx.power = obj->integer.value;
674
675                 (pr->power.count)++;
676                 memcpy(&(pr->power.states[pr->power.count]), &cx, sizeof(cx));
677         }
678
679         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n", pr->power.count));
680
681         /* Validate number of power states discovered */
682         if (pr->power.count < 2)
683                 status = -ENODEV;
684
685 end:
686         acpi_os_free(buffer.pointer);
687
688         return_VALUE(status);
689 }
690
691
692 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
693 {
694         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c2");
695
696         if (!cx->address)
697                 return_VOID;
698
699         /*
700          * C2 latency must be less than or equal to 100
701          * microseconds.
702          */
703         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
704                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
705                                   "latency too large [%d]\n",
706                                   cx->latency));
707                 return_VOID;
708         }
709
710         /*
711          * Otherwise we've met all of our C2 requirements.
712          * Normalize the C2 latency to expidite policy
713          */
714         cx->valid = 1;
715         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
716
717         return_VOID;
718 }
719
720
721 static void acpi_processor_power_verify_c3(
722         struct acpi_processor *pr,
723         struct acpi_processor_cx *cx)
724 {
725         static int bm_check_flag;
726
727         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c3");
728
729         if (!cx->address)
730                 return_VOID;
731
732         /*
733          * C3 latency must be less than or equal to 1000
734          * microseconds.
735          */
736         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
737                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
738                                   "latency too large [%d]\n",
739                                   cx->latency));
740                 return_VOID;
741         }
742
743         /*
744          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
745          * DMA transfers are used by any ISA device to avoid livelock.
746          * Note that we could disable Type-F DMA (as recommended by
747          * the erratum), but this is known to disrupt certain ISA
748          * devices thus we take the conservative approach.
749          */
750         else if (errata.piix4.fdma) {
751                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
752                         "C3 not supported on PIIX4 with Type-F DMA\n"));
753                 return_VOID;
754         }
755
756         /* All the logic here assumes flags.bm_check is same across all CPUs */
757         if (!bm_check_flag) {
758                 /* Determine whether bm_check is needed based on CPU  */
759                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
760                 bm_check_flag = pr->flags.bm_check;
761         } else {
762                 pr->flags.bm_check = bm_check_flag;
763         }
764
765         if (pr->flags.bm_check) {
766                 printk("Disabling BM access before entering C3\n");
767                 /* bus mastering control is necessary */
768                 if (!pr->flags.bm_control) {
769                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
770                           "C3 support requires bus mastering control\n"));
771                         return_VOID;
772                 }
773         } else {
774                 printk("Invalidating cache before entering C3\n");
775                 /*
776                  * WBINVD should be set in fadt, for C3 state to be
777                  * supported on when bm_check is not required.
778                  */
779                 if (acpi_fadt.wb_invd != 1) {
780                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
781                           "Cache invalidation should work properly"
782                           " for C3 to be enabled on SMP systems\n"));
783                         return_VOID;
784                 }
785                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
786                                 0, ACPI_MTX_DO_NOT_LOCK);
787         }
788
789         /*
790          * Otherwise we've met all of our C3 requirements.
791          * Normalize the C3 latency to expidite policy.  Enable
792          * checking of bus mastering status (bm_check) so we can
793          * use this in our C3 policy
794          */
795         cx->valid = 1;
796         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
797
798         return_VOID;
799 }
800
801
802 static int acpi_processor_power_verify(struct acpi_processor *pr)
803 {
804         unsigned int i;
805         unsigned int working = 0;
806
807         for (i=1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
808                 struct acpi_processor_cx *cx = &pr->power.states[i];
809
810                 switch (cx->type) {
811                 case ACPI_STATE_C1:
812                         cx->valid = 1;
813                         break;
814
815                 case ACPI_STATE_C2:
816                         acpi_processor_power_verify_c2(cx);
817                         break;
818
819                 case ACPI_STATE_C3:
820                         acpi_processor_power_verify_c3(pr, cx);
821                         break;
822                 }
823
824                 if (cx->valid)
825                         working++;
826         }
827
828         return (working);
829 }
830
831 static int acpi_processor_get_power_info (
832         struct acpi_processor   *pr)
833 {
834         unsigned int i;
835         int result;
836
837         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info");
838
839         /* NOTE: the idle thread may not be running while calling
840          * this function */
841
842         result = acpi_processor_get_power_info_cst(pr);
843         if ((result) || (acpi_processor_power_verify(pr) < 2)) {
844                 result = acpi_processor_get_power_info_fadt(pr);
845                 if (result)
846                         result = acpi_processor_get_power_info_default_c1(pr);
847         }
848
849         /*
850          * Set Default Policy
851          * ------------------
852          * Now that we know which states are supported, set the default
853          * policy.  Note that this policy can be changed dynamically
854          * (e.g. encourage deeper sleeps to conserve battery life when
855          * not on AC).
856          */
857         result = acpi_processor_set_power_policy(pr);
858         if (result)
859                 return_VALUE(result);
860
861         /*
862          * if one state of type C2 or C3 is available, mark this
863          * CPU as being "idle manageable"
864          */
865         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
866                 if (pr->power.states[i].valid) {
867                         pr->power.count = i;
868                         pr->flags.power = 1;
869                 }
870         }
871
872         return_VALUE(0);
873 }
874
875 int acpi_processor_cst_has_changed (struct acpi_processor *pr)
876 {
877         int                     result = 0;
878
879         ACPI_FUNCTION_TRACE("acpi_processor_cst_has_changed");
880
881         if (!pr)
882                 return_VALUE(-EINVAL);
883
884         if ( nocst) {
885                 return_VALUE(-ENODEV);
886         }
887
888         if (!pr->flags.power_setup_done)
889                 return_VALUE(-ENODEV);
890
891         /* Fall back to the default idle loop */
892         pm_idle = pm_idle_save;
893         synchronize_sched();  /* Relies on interrupts forcing exit from idle. */
894
895         pr->flags.power = 0;
896         result = acpi_processor_get_power_info(pr);
897         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
898                 pm_idle = acpi_processor_idle;
899
900         return_VALUE(result);
901 }
902
903 /* proc interface */
904
905 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
906 {
907         struct acpi_processor   *pr = (struct acpi_processor *)seq->private;
908         unsigned int            i;
909
910         ACPI_FUNCTION_TRACE("acpi_processor_power_seq_show");
911
912         if (!pr)
913                 goto end;
914
915         seq_printf(seq, "active state:            C%zd\n"
916                         "max_cstate:              C%d\n"
917                         "bus master activity:     %08x\n",
918                         pr->power.state ? pr->power.state - pr->power.states : 0,
919                         max_cstate,
920                         (unsigned)pr->power.bm_activity);
921
922         seq_puts(seq, "states:\n");
923
924         for (i = 1; i <= pr->power.count; i++) {
925                 seq_printf(seq, "   %cC%d:                  ",
926                         (&pr->power.states[i] == pr->power.state?'*':' '), i);
927
928                 if (!pr->power.states[i].valid) {
929                         seq_puts(seq, "<not supported>\n");
930                         continue;
931                 }
932
933                 switch (pr->power.states[i].type) {
934                 case ACPI_STATE_C1:
935                         seq_printf(seq, "type[C1] ");
936                         break;
937                 case ACPI_STATE_C2:
938                         seq_printf(seq, "type[C2] ");
939                         break;
940                 case ACPI_STATE_C3:
941                         seq_printf(seq, "type[C3] ");
942                         break;
943                 default:
944                         seq_printf(seq, "type[--] ");
945                         break;
946                 }
947
948                 if (pr->power.states[i].promotion.state)
949                         seq_printf(seq, "promotion[C%zd] ",
950                                 (pr->power.states[i].promotion.state -
951                                  pr->power.states));
952                 else
953                         seq_puts(seq, "promotion[--] ");
954
955                 if (pr->power.states[i].demotion.state)
956                         seq_printf(seq, "demotion[C%zd] ",
957                                 (pr->power.states[i].demotion.state -
958                                  pr->power.states));
959                 else
960                         seq_puts(seq, "demotion[--] ");
961
962                 seq_printf(seq, "latency[%03d] usage[%08d]\n",
963                         pr->power.states[i].latency,
964                         pr->power.states[i].usage);
965         }
966
967 end:
968         return_VALUE(0);
969 }
970
971 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
972 {
973         return single_open(file, acpi_processor_power_seq_show,
974                                                 PDE(inode)->data);
975 }
976
977 static struct file_operations acpi_processor_power_fops = {
978         .open           = acpi_processor_power_open_fs,
979         .read           = seq_read,
980         .llseek         = seq_lseek,
981         .release        = single_release,
982 };
983
984 int acpi_processor_power_init(struct acpi_processor *pr, struct acpi_device *device)
985 {
986         acpi_status             status = 0;
987         static int              first_run = 0;
988         struct proc_dir_entry   *entry = NULL;
989         unsigned int i;
990
991         ACPI_FUNCTION_TRACE("acpi_processor_power_init");
992
993         if (!first_run) {
994                 dmi_check_system(processor_power_dmi_table);
995                 if (max_cstate < ACPI_C_STATES_MAX)
996                         printk(KERN_NOTICE "ACPI: processor limited to max C-state %d\n", max_cstate);
997                 first_run++;
998         }
999
1000         if (!pr)
1001                 return_VALUE(-EINVAL);
1002
1003         if (acpi_fadt.cst_cnt && !nocst) {
1004                 status = acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1005                 if (ACPI_FAILURE(status)) {
1006                         ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1007                                           "Notifying BIOS of _CST ability failed\n"));
1008                 }
1009         }
1010
1011         acpi_processor_power_init_pdc(&(pr->power), pr->id);
1012         acpi_processor_set_pdc(pr, pr->power.pdc);
1013         acpi_processor_get_power_info(pr);
1014
1015         /*
1016          * Install the idle handler if processor power management is supported.
1017          * Note that we use previously set idle handler will be used on
1018          * platforms that only support C1.
1019          */
1020         if ((pr->flags.power) && (!boot_option_idle_override)) {
1021                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1022                 for (i = 1; i <= pr->power.count; i++)
1023                         if (pr->power.states[i].valid)
1024                                 printk(" C%d[C%d]", i, pr->power.states[i].type);
1025                 printk(")\n");
1026
1027                 if (pr->id == 0) {
1028                         pm_idle_save = pm_idle;
1029                         pm_idle = acpi_processor_idle;
1030                 }
1031         }
1032
1033         /* 'power' [R] */
1034         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1035                 S_IRUGO, acpi_device_dir(device));
1036         if (!entry)
1037                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1038                         "Unable to create '%s' fs entry\n",
1039                         ACPI_PROCESSOR_FILE_POWER));
1040         else {
1041                 entry->proc_fops = &acpi_processor_power_fops;
1042                 entry->data = acpi_driver_data(device);
1043                 entry->owner = THIS_MODULE;
1044         }
1045
1046         pr->flags.power_setup_done = 1;
1047
1048         return_VALUE(0);
1049 }
1050
1051 int acpi_processor_power_exit(struct acpi_processor *pr, struct acpi_device *device)
1052 {
1053         ACPI_FUNCTION_TRACE("acpi_processor_power_exit");
1054
1055         pr->flags.power_setup_done = 0;
1056
1057         if (acpi_device_dir(device))
1058                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,acpi_device_dir(device));
1059
1060         /* Unregister the idle handler when processor #0 is removed. */
1061         if (pr->id == 0) {
1062                 pm_idle = pm_idle_save;
1063
1064                 /*
1065                  * We are about to unload the current idle thread pm callback
1066                  * (pm_idle), Wait for all processors to update cached/local
1067                  * copies of pm_idle before proceeding.
1068                  */
1069                 cpu_idle_wait();
1070         }
1071
1072         return_VALUE(0);
1073 }