[AVR32] Update defconfigs
[sfrench/cifs-2.6.git] / 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, 2005 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 #include <linux/sched.h>        /* need_resched() */
41 #include <linux/latency.h>
42 #include <linux/clockchips.h>
43
44 /*
45  * Include the apic definitions for x86 to have the APIC timer related defines
46  * available also for UP (on SMP it gets magically included via linux/smp.h).
47  * asm/acpi.h is not an option, as it would require more include magic. Also
48  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
49  */
50 #ifdef CONFIG_X86
51 #include <asm/apic.h>
52 #endif
53
54 #include <asm/io.h>
55 #include <asm/uaccess.h>
56
57 #include <acpi/acpi_bus.h>
58 #include <acpi/processor.h>
59
60 #define ACPI_PROCESSOR_COMPONENT        0x01000000
61 #define ACPI_PROCESSOR_CLASS            "processor"
62 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
63 ACPI_MODULE_NAME("processor_idle");
64 #define ACPI_PROCESSOR_FILE_POWER       "power"
65 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
66 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
67 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
68 static void (*pm_idle_save) (void) __read_mostly;
69 module_param(max_cstate, uint, 0644);
70
71 static unsigned int nocst __read_mostly;
72 module_param(nocst, uint, 0000);
73
74 /*
75  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
76  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
77  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
78  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
79  * reduce history for more aggressive entry into C3
80  */
81 static unsigned int bm_history __read_mostly =
82     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
83 module_param(bm_history, uint, 0644);
84 /* --------------------------------------------------------------------------
85                                 Power Management
86    -------------------------------------------------------------------------- */
87
88 /*
89  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
90  * For now disable this. Probably a bug somewhere else.
91  *
92  * To skip this limit, boot/load with a large max_cstate limit.
93  */
94 static int set_max_cstate(struct dmi_system_id *id)
95 {
96         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
97                 return 0;
98
99         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
100                " Override with \"processor.max_cstate=%d\"\n", id->ident,
101                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
102
103         max_cstate = (long)id->driver_data;
104
105         return 0;
106 }
107
108 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
109    callers to only run once -AK */
110 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
111         { set_max_cstate, "IBM ThinkPad R40e", {
112           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
113           DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
114         { set_max_cstate, "IBM ThinkPad R40e", {
115           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
116           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
117         { set_max_cstate, "IBM ThinkPad R40e", {
118           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
119           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
120         { set_max_cstate, "IBM ThinkPad R40e", {
121           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
122           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
123         { set_max_cstate, "IBM ThinkPad R40e", {
124           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
125           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
126         { set_max_cstate, "IBM ThinkPad R40e", {
127           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
128           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
129         { set_max_cstate, "IBM ThinkPad R40e", {
130           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
131           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
132         { set_max_cstate, "IBM ThinkPad R40e", {
133           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
134           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
135         { set_max_cstate, "IBM ThinkPad R40e", {
136           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
137           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
138         { set_max_cstate, "IBM ThinkPad R40e", {
139           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
140           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
141         { set_max_cstate, "IBM ThinkPad R40e", {
142           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
143           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
144         { set_max_cstate, "IBM ThinkPad R40e", {
145           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
146           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
147         { set_max_cstate, "IBM ThinkPad R40e", {
148           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
149           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
150         { set_max_cstate, "IBM ThinkPad R40e", {
151           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
152           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
153         { set_max_cstate, "IBM ThinkPad R40e", {
154           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
155           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
156         { set_max_cstate, "IBM ThinkPad R40e", {
157           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
158           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
159         { set_max_cstate, "Medion 41700", {
160           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
161           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
162         { set_max_cstate, "Clevo 5600D", {
163           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
164           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
165          (void *)2},
166         {},
167 };
168
169 static inline u32 ticks_elapsed(u32 t1, u32 t2)
170 {
171         if (t2 >= t1)
172                 return (t2 - t1);
173         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
174                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
175         else
176                 return ((0xFFFFFFFF - t1) + t2);
177 }
178
179 static void
180 acpi_processor_power_activate(struct acpi_processor *pr,
181                               struct acpi_processor_cx *new)
182 {
183         struct acpi_processor_cx *old;
184
185         if (!pr || !new)
186                 return;
187
188         old = pr->power.state;
189
190         if (old)
191                 old->promotion.count = 0;
192         new->demotion.count = 0;
193
194         /* Cleanup from old state. */
195         if (old) {
196                 switch (old->type) {
197                 case ACPI_STATE_C3:
198                         /* Disable bus master reload */
199                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
200                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
201                         break;
202                 }
203         }
204
205         /* Prepare to use new state. */
206         switch (new->type) {
207         case ACPI_STATE_C3:
208                 /* Enable bus master reload */
209                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
210                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
211                 break;
212         }
213
214         pr->power.state = new;
215
216         return;
217 }
218
219 static void acpi_safe_halt(void)
220 {
221         current_thread_info()->status &= ~TS_POLLING;
222         /*
223          * TS_POLLING-cleared state must be visible before we
224          * test NEED_RESCHED:
225          */
226         smp_mb();
227         if (!need_resched())
228                 safe_halt();
229         current_thread_info()->status |= TS_POLLING;
230 }
231
232 static atomic_t c3_cpu_count;
233
234 /* Common C-state entry for C2, C3, .. */
235 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
236 {
237         if (cstate->space_id == ACPI_CSTATE_FFH) {
238                 /* Call into architectural FFH based C-state */
239                 acpi_processor_ffh_cstate_enter(cstate);
240         } else {
241                 int unused;
242                 /* IO port based C-state */
243                 inb(cstate->address);
244                 /* Dummy wait op - must do something useless after P_LVL2 read
245                    because chipsets cannot guarantee that STPCLK# signal
246                    gets asserted in time to freeze execution properly. */
247                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
248         }
249 }
250
251 #ifdef ARCH_APICTIMER_STOPS_ON_C3
252
253 /*
254  * Some BIOS implementations switch to C3 in the published C2 state.
255  * This seems to be a common problem on AMD boxen, but other vendors
256  * are affected too. We pick the most conservative approach: we assume
257  * that the local APIC stops in both C2 and C3.
258  */
259 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
260                                    struct acpi_processor_cx *cx)
261 {
262         struct acpi_processor_power *pwr = &pr->power;
263         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
264
265         /*
266          * Check, if one of the previous states already marked the lapic
267          * unstable
268          */
269         if (pwr->timer_broadcast_on_state < state)
270                 return;
271
272         if (cx->type >= type)
273                 pr->power.timer_broadcast_on_state = state;
274 }
275
276 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr)
277 {
278 #ifdef CONFIG_GENERIC_CLOCKEVENTS
279         unsigned long reason;
280
281         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
282                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
283
284         clockevents_notify(reason, &pr->id);
285 #else
286         cpumask_t mask = cpumask_of_cpu(pr->id);
287
288         if (pr->power.timer_broadcast_on_state < INT_MAX)
289                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
290         else
291                 on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
292 #endif
293 }
294
295 /* Power(C) State timer broadcast control */
296 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
297                                        struct acpi_processor_cx *cx,
298                                        int broadcast)
299 {
300 #ifdef CONFIG_GENERIC_CLOCKEVENTS
301
302         int state = cx - pr->power.states;
303
304         if (state >= pr->power.timer_broadcast_on_state) {
305                 unsigned long reason;
306
307                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
308                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
309                 clockevents_notify(reason, &pr->id);
310         }
311 #endif
312 }
313
314 #else
315
316 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
317                                    struct acpi_processor_cx *cstate) { }
318 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr) { }
319 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
320                                        struct acpi_processor_cx *cx,
321                                        int broadcast)
322 {
323 }
324
325 #endif
326
327 static void acpi_processor_idle(void)
328 {
329         struct acpi_processor *pr = NULL;
330         struct acpi_processor_cx *cx = NULL;
331         struct acpi_processor_cx *next_state = NULL;
332         int sleep_ticks = 0;
333         u32 t1, t2 = 0;
334
335         pr = processors[smp_processor_id()];
336         if (!pr)
337                 return;
338
339         /*
340          * Interrupts must be disabled during bus mastering calculations and
341          * for C2/C3 transitions.
342          */
343         local_irq_disable();
344
345         /*
346          * Check whether we truly need to go idle, or should
347          * reschedule:
348          */
349         if (unlikely(need_resched())) {
350                 local_irq_enable();
351                 return;
352         }
353
354         cx = pr->power.state;
355         if (!cx) {
356                 if (pm_idle_save)
357                         pm_idle_save();
358                 else
359                         acpi_safe_halt();
360                 return;
361         }
362
363         /*
364          * Check BM Activity
365          * -----------------
366          * Check for bus mastering activity (if required), record, and check
367          * for demotion.
368          */
369         if (pr->flags.bm_check) {
370                 u32 bm_status = 0;
371                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
372
373                 if (diff > 31)
374                         diff = 31;
375
376                 pr->power.bm_activity <<= diff;
377
378                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
379                 if (bm_status) {
380                         pr->power.bm_activity |= 0x1;
381                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
382                 }
383                 /*
384                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
385                  * the true state of bus mastering activity; forcing us to
386                  * manually check the BMIDEA bit of each IDE channel.
387                  */
388                 else if (errata.piix4.bmisx) {
389                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
390                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
391                                 pr->power.bm_activity |= 0x1;
392                 }
393
394                 pr->power.bm_check_timestamp = jiffies;
395
396                 /*
397                  * If bus mastering is or was active this jiffy, demote
398                  * to avoid a faulty transition.  Note that the processor
399                  * won't enter a low-power state during this call (to this
400                  * function) but should upon the next.
401                  *
402                  * TBD: A better policy might be to fallback to the demotion
403                  *      state (use it for this quantum only) istead of
404                  *      demoting -- and rely on duration as our sole demotion
405                  *      qualification.  This may, however, introduce DMA
406                  *      issues (e.g. floppy DMA transfer overrun/underrun).
407                  */
408                 if ((pr->power.bm_activity & 0x1) &&
409                     cx->demotion.threshold.bm) {
410                         local_irq_enable();
411                         next_state = cx->demotion.state;
412                         goto end;
413                 }
414         }
415
416 #ifdef CONFIG_HOTPLUG_CPU
417         /*
418          * Check for P_LVL2_UP flag before entering C2 and above on
419          * an SMP system. We do it here instead of doing it at _CST/P_LVL
420          * detection phase, to work cleanly with logical CPU hotplug.
421          */
422         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) && 
423             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
424                 cx = &pr->power.states[ACPI_STATE_C1];
425 #endif
426
427         /*
428          * Sleep:
429          * ------
430          * Invoke the current Cx state to put the processor to sleep.
431          */
432         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
433                 current_thread_info()->status &= ~TS_POLLING;
434                 /*
435                  * TS_POLLING-cleared state must be visible before we
436                  * test NEED_RESCHED:
437                  */
438                 smp_mb();
439                 if (need_resched()) {
440                         current_thread_info()->status |= TS_POLLING;
441                         local_irq_enable();
442                         return;
443                 }
444         }
445
446         switch (cx->type) {
447
448         case ACPI_STATE_C1:
449                 /*
450                  * Invoke C1.
451                  * Use the appropriate idle routine, the one that would
452                  * be used without acpi C-states.
453                  */
454                 if (pm_idle_save)
455                         pm_idle_save();
456                 else
457                         acpi_safe_halt();
458
459                 /*
460                  * TBD: Can't get time duration while in C1, as resumes
461                  *      go to an ISR rather than here.  Need to instrument
462                  *      base interrupt handler.
463                  */
464                 sleep_ticks = 0xFFFFFFFF;
465                 break;
466
467         case ACPI_STATE_C2:
468                 /* Get start time (ticks) */
469                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
470                 /* Invoke C2 */
471                 acpi_state_timer_broadcast(pr, cx, 1);
472                 acpi_cstate_enter(cx);
473                 /* Get end time (ticks) */
474                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
475
476 #ifdef CONFIG_GENERIC_TIME
477                 /* TSC halts in C2, so notify users */
478                 mark_tsc_unstable("possible TSC halt in C2");
479 #endif
480                 /* Re-enable interrupts */
481                 local_irq_enable();
482                 current_thread_info()->status |= TS_POLLING;
483                 /* Compute time (ticks) that we were actually asleep */
484                 sleep_ticks =
485                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
486                 acpi_state_timer_broadcast(pr, cx, 0);
487                 break;
488
489         case ACPI_STATE_C3:
490
491                 if (pr->flags.bm_check) {
492                         if (atomic_inc_return(&c3_cpu_count) ==
493                             num_online_cpus()) {
494                                 /*
495                                  * All CPUs are trying to go to C3
496                                  * Disable bus master arbitration
497                                  */
498                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
499                         }
500                 } else {
501                         /* SMP with no shared cache... Invalidate cache  */
502                         ACPI_FLUSH_CPU_CACHE();
503                 }
504
505                 /* Get start time (ticks) */
506                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
507                 /* Invoke C3 */
508                 acpi_state_timer_broadcast(pr, cx, 1);
509                 acpi_cstate_enter(cx);
510                 /* Get end time (ticks) */
511                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
512                 if (pr->flags.bm_check) {
513                         /* Enable bus master arbitration */
514                         atomic_dec(&c3_cpu_count);
515                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
516                 }
517
518 #ifdef CONFIG_GENERIC_TIME
519                 /* TSC halts in C3, so notify users */
520                 mark_tsc_unstable("TSC halts in C3");
521 #endif
522                 /* Re-enable interrupts */
523                 local_irq_enable();
524                 current_thread_info()->status |= TS_POLLING;
525                 /* Compute time (ticks) that we were actually asleep */
526                 sleep_ticks =
527                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
528                 acpi_state_timer_broadcast(pr, cx, 0);
529                 break;
530
531         default:
532                 local_irq_enable();
533                 return;
534         }
535         cx->usage++;
536         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
537                 cx->time += sleep_ticks;
538
539         next_state = pr->power.state;
540
541 #ifdef CONFIG_HOTPLUG_CPU
542         /* Don't do promotion/demotion */
543         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
544             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
545                 next_state = cx;
546                 goto end;
547         }
548 #endif
549
550         /*
551          * Promotion?
552          * ----------
553          * Track the number of longs (time asleep is greater than threshold)
554          * and promote when the count threshold is reached.  Note that bus
555          * mastering activity may prevent promotions.
556          * Do not promote above max_cstate.
557          */
558         if (cx->promotion.state &&
559             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
560                 if (sleep_ticks > cx->promotion.threshold.ticks &&
561                   cx->promotion.state->latency <= system_latency_constraint()) {
562                         cx->promotion.count++;
563                         cx->demotion.count = 0;
564                         if (cx->promotion.count >=
565                             cx->promotion.threshold.count) {
566                                 if (pr->flags.bm_check) {
567                                         if (!
568                                             (pr->power.bm_activity & cx->
569                                              promotion.threshold.bm)) {
570                                                 next_state =
571                                                     cx->promotion.state;
572                                                 goto end;
573                                         }
574                                 } else {
575                                         next_state = cx->promotion.state;
576                                         goto end;
577                                 }
578                         }
579                 }
580         }
581
582         /*
583          * Demotion?
584          * ---------
585          * Track the number of shorts (time asleep is less than time threshold)
586          * and demote when the usage threshold is reached.
587          */
588         if (cx->demotion.state) {
589                 if (sleep_ticks < cx->demotion.threshold.ticks) {
590                         cx->demotion.count++;
591                         cx->promotion.count = 0;
592                         if (cx->demotion.count >= cx->demotion.threshold.count) {
593                                 next_state = cx->demotion.state;
594                                 goto end;
595                         }
596                 }
597         }
598
599       end:
600         /*
601          * Demote if current state exceeds max_cstate
602          * or if the latency of the current state is unacceptable
603          */
604         if ((pr->power.state - pr->power.states) > max_cstate ||
605                 pr->power.state->latency > system_latency_constraint()) {
606                 if (cx->demotion.state)
607                         next_state = cx->demotion.state;
608         }
609
610         /*
611          * New Cx State?
612          * -------------
613          * If we're going to start using a new Cx state we must clean up
614          * from the previous and prepare to use the new.
615          */
616         if (next_state != pr->power.state)
617                 acpi_processor_power_activate(pr, next_state);
618 }
619
620 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
621 {
622         unsigned int i;
623         unsigned int state_is_set = 0;
624         struct acpi_processor_cx *lower = NULL;
625         struct acpi_processor_cx *higher = NULL;
626         struct acpi_processor_cx *cx;
627
628
629         if (!pr)
630                 return -EINVAL;
631
632         /*
633          * This function sets the default Cx state policy (OS idle handler).
634          * Our scheme is to promote quickly to C2 but more conservatively
635          * to C3.  We're favoring C2  for its characteristics of low latency
636          * (quick response), good power savings, and ability to allow bus
637          * mastering activity.  Note that the Cx state policy is completely
638          * customizable and can be altered dynamically.
639          */
640
641         /* startup state */
642         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
643                 cx = &pr->power.states[i];
644                 if (!cx->valid)
645                         continue;
646
647                 if (!state_is_set)
648                         pr->power.state = cx;
649                 state_is_set++;
650                 break;
651         }
652
653         if (!state_is_set)
654                 return -ENODEV;
655
656         /* demotion */
657         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
658                 cx = &pr->power.states[i];
659                 if (!cx->valid)
660                         continue;
661
662                 if (lower) {
663                         cx->demotion.state = lower;
664                         cx->demotion.threshold.ticks = cx->latency_ticks;
665                         cx->demotion.threshold.count = 1;
666                         if (cx->type == ACPI_STATE_C3)
667                                 cx->demotion.threshold.bm = bm_history;
668                 }
669
670                 lower = cx;
671         }
672
673         /* promotion */
674         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
675                 cx = &pr->power.states[i];
676                 if (!cx->valid)
677                         continue;
678
679                 if (higher) {
680                         cx->promotion.state = higher;
681                         cx->promotion.threshold.ticks = cx->latency_ticks;
682                         if (cx->type >= ACPI_STATE_C2)
683                                 cx->promotion.threshold.count = 4;
684                         else
685                                 cx->promotion.threshold.count = 10;
686                         if (higher->type == ACPI_STATE_C3)
687                                 cx->promotion.threshold.bm = bm_history;
688                 }
689
690                 higher = cx;
691         }
692
693         return 0;
694 }
695
696 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
697 {
698
699         if (!pr)
700                 return -EINVAL;
701
702         if (!pr->pblk)
703                 return -ENODEV;
704
705         /* if info is obtained from pblk/fadt, type equals state */
706         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
707         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
708
709 #ifndef CONFIG_HOTPLUG_CPU
710         /*
711          * Check for P_LVL2_UP flag before entering C2 and above on
712          * an SMP system. 
713          */
714         if ((num_online_cpus() > 1) &&
715             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
716                 return -ENODEV;
717 #endif
718
719         /* determine C2 and C3 address from pblk */
720         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
721         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
722
723         /* determine latencies from FADT */
724         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
725         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
726
727         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
728                           "lvl2[0x%08x] lvl3[0x%08x]\n",
729                           pr->power.states[ACPI_STATE_C2].address,
730                           pr->power.states[ACPI_STATE_C3].address));
731
732         return 0;
733 }
734
735 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
736 {
737         if (!pr->power.states[ACPI_STATE_C1].valid) {
738                 /* set the first C-State to C1 */
739                 /* all processors need to support C1 */
740                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
741                 pr->power.states[ACPI_STATE_C1].valid = 1;
742         }
743         /* the C0 state only exists as a filler in our array */
744         pr->power.states[ACPI_STATE_C0].valid = 1;
745         return 0;
746 }
747
748 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
749 {
750         acpi_status status = 0;
751         acpi_integer count;
752         int current_count;
753         int i;
754         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
755         union acpi_object *cst;
756
757
758         if (nocst)
759                 return -ENODEV;
760
761         current_count = 0;
762
763         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
764         if (ACPI_FAILURE(status)) {
765                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
766                 return -ENODEV;
767         }
768
769         cst = buffer.pointer;
770
771         /* There must be at least 2 elements */
772         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
773                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
774                 status = -EFAULT;
775                 goto end;
776         }
777
778         count = cst->package.elements[0].integer.value;
779
780         /* Validate number of power states. */
781         if (count < 1 || count != cst->package.count - 1) {
782                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
783                 status = -EFAULT;
784                 goto end;
785         }
786
787         /* Tell driver that at least _CST is supported. */
788         pr->flags.has_cst = 1;
789
790         for (i = 1; i <= count; i++) {
791                 union acpi_object *element;
792                 union acpi_object *obj;
793                 struct acpi_power_register *reg;
794                 struct acpi_processor_cx cx;
795
796                 memset(&cx, 0, sizeof(cx));
797
798                 element = &(cst->package.elements[i]);
799                 if (element->type != ACPI_TYPE_PACKAGE)
800                         continue;
801
802                 if (element->package.count != 4)
803                         continue;
804
805                 obj = &(element->package.elements[0]);
806
807                 if (obj->type != ACPI_TYPE_BUFFER)
808                         continue;
809
810                 reg = (struct acpi_power_register *)obj->buffer.pointer;
811
812                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
813                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
814                         continue;
815
816                 /* There should be an easy way to extract an integer... */
817                 obj = &(element->package.elements[1]);
818                 if (obj->type != ACPI_TYPE_INTEGER)
819                         continue;
820
821                 cx.type = obj->integer.value;
822                 /*
823                  * Some buggy BIOSes won't list C1 in _CST -
824                  * Let acpi_processor_get_power_info_default() handle them later
825                  */
826                 if (i == 1 && cx.type != ACPI_STATE_C1)
827                         current_count++;
828
829                 cx.address = reg->address;
830                 cx.index = current_count + 1;
831
832                 cx.space_id = ACPI_CSTATE_SYSTEMIO;
833                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
834                         if (acpi_processor_ffh_cstate_probe
835                                         (pr->id, &cx, reg) == 0) {
836                                 cx.space_id = ACPI_CSTATE_FFH;
837                         } else if (cx.type != ACPI_STATE_C1) {
838                                 /*
839                                  * C1 is a special case where FIXED_HARDWARE
840                                  * can be handled in non-MWAIT way as well.
841                                  * In that case, save this _CST entry info.
842                                  * That is, we retain space_id of SYSTEM_IO for
843                                  * halt based C1.
844                                  * Otherwise, ignore this info and continue.
845                                  */
846                                 continue;
847                         }
848                 }
849
850                 obj = &(element->package.elements[2]);
851                 if (obj->type != ACPI_TYPE_INTEGER)
852                         continue;
853
854                 cx.latency = obj->integer.value;
855
856                 obj = &(element->package.elements[3]);
857                 if (obj->type != ACPI_TYPE_INTEGER)
858                         continue;
859
860                 cx.power = obj->integer.value;
861
862                 current_count++;
863                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
864
865                 /*
866                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
867                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
868                  */
869                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
870                         printk(KERN_WARNING
871                                "Limiting number of power states to max (%d)\n",
872                                ACPI_PROCESSOR_MAX_POWER);
873                         printk(KERN_WARNING
874                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
875                         break;
876                 }
877         }
878
879         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
880                           current_count));
881
882         /* Validate number of power states discovered */
883         if (current_count < 2)
884                 status = -EFAULT;
885
886       end:
887         kfree(buffer.pointer);
888
889         return status;
890 }
891
892 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
893 {
894
895         if (!cx->address)
896                 return;
897
898         /*
899          * C2 latency must be less than or equal to 100
900          * microseconds.
901          */
902         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
903                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
904                                   "latency too large [%d]\n", cx->latency));
905                 return;
906         }
907
908         /*
909          * Otherwise we've met all of our C2 requirements.
910          * Normalize the C2 latency to expidite policy
911          */
912         cx->valid = 1;
913         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
914
915         return;
916 }
917
918 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
919                                            struct acpi_processor_cx *cx)
920 {
921         static int bm_check_flag;
922
923
924         if (!cx->address)
925                 return;
926
927         /*
928          * C3 latency must be less than or equal to 1000
929          * microseconds.
930          */
931         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
932                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
933                                   "latency too large [%d]\n", cx->latency));
934                 return;
935         }
936
937         /*
938          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
939          * DMA transfers are used by any ISA device to avoid livelock.
940          * Note that we could disable Type-F DMA (as recommended by
941          * the erratum), but this is known to disrupt certain ISA
942          * devices thus we take the conservative approach.
943          */
944         else if (errata.piix4.fdma) {
945                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
946                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
947                 return;
948         }
949
950         /* All the logic here assumes flags.bm_check is same across all CPUs */
951         if (!bm_check_flag) {
952                 /* Determine whether bm_check is needed based on CPU  */
953                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
954                 bm_check_flag = pr->flags.bm_check;
955         } else {
956                 pr->flags.bm_check = bm_check_flag;
957         }
958
959         if (pr->flags.bm_check) {
960                 /* bus mastering control is necessary */
961                 if (!pr->flags.bm_control) {
962                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
963                                           "C3 support requires bus mastering control\n"));
964                         return;
965                 }
966         } else {
967                 /*
968                  * WBINVD should be set in fadt, for C3 state to be
969                  * supported on when bm_check is not required.
970                  */
971                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
972                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
973                                           "Cache invalidation should work properly"
974                                           " for C3 to be enabled on SMP systems\n"));
975                         return;
976                 }
977                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
978         }
979
980         /*
981          * Otherwise we've met all of our C3 requirements.
982          * Normalize the C3 latency to expidite policy.  Enable
983          * checking of bus mastering status (bm_check) so we can
984          * use this in our C3 policy
985          */
986         cx->valid = 1;
987         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
988
989         return;
990 }
991
992 static int acpi_processor_power_verify(struct acpi_processor *pr)
993 {
994         unsigned int i;
995         unsigned int working = 0;
996
997         pr->power.timer_broadcast_on_state = INT_MAX;
998
999         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1000                 struct acpi_processor_cx *cx = &pr->power.states[i];
1001
1002                 switch (cx->type) {
1003                 case ACPI_STATE_C1:
1004                         cx->valid = 1;
1005                         break;
1006
1007                 case ACPI_STATE_C2:
1008                         acpi_processor_power_verify_c2(cx);
1009                         if (cx->valid)
1010                                 acpi_timer_check_state(i, pr, cx);
1011                         break;
1012
1013                 case ACPI_STATE_C3:
1014                         acpi_processor_power_verify_c3(pr, cx);
1015                         if (cx->valid)
1016                                 acpi_timer_check_state(i, pr, cx);
1017                         break;
1018                 }
1019
1020                 if (cx->valid)
1021                         working++;
1022         }
1023
1024         acpi_propagate_timer_broadcast(pr);
1025
1026         return (working);
1027 }
1028
1029 static int acpi_processor_get_power_info(struct acpi_processor *pr)
1030 {
1031         unsigned int i;
1032         int result;
1033
1034
1035         /* NOTE: the idle thread may not be running while calling
1036          * this function */
1037
1038         /* Zero initialize all the C-states info. */
1039         memset(pr->power.states, 0, sizeof(pr->power.states));
1040
1041         result = acpi_processor_get_power_info_cst(pr);
1042         if (result == -ENODEV)
1043                 result = acpi_processor_get_power_info_fadt(pr);
1044
1045         if (result)
1046                 return result;
1047
1048         acpi_processor_get_power_info_default(pr);
1049
1050         pr->power.count = acpi_processor_power_verify(pr);
1051
1052         /*
1053          * Set Default Policy
1054          * ------------------
1055          * Now that we know which states are supported, set the default
1056          * policy.  Note that this policy can be changed dynamically
1057          * (e.g. encourage deeper sleeps to conserve battery life when
1058          * not on AC).
1059          */
1060         result = acpi_processor_set_power_policy(pr);
1061         if (result)
1062                 return result;
1063
1064         /*
1065          * if one state of type C2 or C3 is available, mark this
1066          * CPU as being "idle manageable"
1067          */
1068         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1069                 if (pr->power.states[i].valid) {
1070                         pr->power.count = i;
1071                         if (pr->power.states[i].type >= ACPI_STATE_C2)
1072                                 pr->flags.power = 1;
1073                 }
1074         }
1075
1076         return 0;
1077 }
1078
1079 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1080 {
1081         int result = 0;
1082
1083
1084         if (!pr)
1085                 return -EINVAL;
1086
1087         if (nocst) {
1088                 return -ENODEV;
1089         }
1090
1091         if (!pr->flags.power_setup_done)
1092                 return -ENODEV;
1093
1094         /* Fall back to the default idle loop */
1095         pm_idle = pm_idle_save;
1096         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
1097
1098         pr->flags.power = 0;
1099         result = acpi_processor_get_power_info(pr);
1100         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1101                 pm_idle = acpi_processor_idle;
1102
1103         return result;
1104 }
1105
1106 /* proc interface */
1107
1108 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1109 {
1110         struct acpi_processor *pr = seq->private;
1111         unsigned int i;
1112
1113
1114         if (!pr)
1115                 goto end;
1116
1117         seq_printf(seq, "active state:            C%zd\n"
1118                    "max_cstate:              C%d\n"
1119                    "bus master activity:     %08x\n"
1120                    "maximum allowed latency: %d usec\n",
1121                    pr->power.state ? pr->power.state - pr->power.states : 0,
1122                    max_cstate, (unsigned)pr->power.bm_activity,
1123                    system_latency_constraint());
1124
1125         seq_puts(seq, "states:\n");
1126
1127         for (i = 1; i <= pr->power.count; i++) {
1128                 seq_printf(seq, "   %cC%d:                  ",
1129                            (&pr->power.states[i] ==
1130                             pr->power.state ? '*' : ' '), i);
1131
1132                 if (!pr->power.states[i].valid) {
1133                         seq_puts(seq, "<not supported>\n");
1134                         continue;
1135                 }
1136
1137                 switch (pr->power.states[i].type) {
1138                 case ACPI_STATE_C1:
1139                         seq_printf(seq, "type[C1] ");
1140                         break;
1141                 case ACPI_STATE_C2:
1142                         seq_printf(seq, "type[C2] ");
1143                         break;
1144                 case ACPI_STATE_C3:
1145                         seq_printf(seq, "type[C3] ");
1146                         break;
1147                 default:
1148                         seq_printf(seq, "type[--] ");
1149                         break;
1150                 }
1151
1152                 if (pr->power.states[i].promotion.state)
1153                         seq_printf(seq, "promotion[C%zd] ",
1154                                    (pr->power.states[i].promotion.state -
1155                                     pr->power.states));
1156                 else
1157                         seq_puts(seq, "promotion[--] ");
1158
1159                 if (pr->power.states[i].demotion.state)
1160                         seq_printf(seq, "demotion[C%zd] ",
1161                                    (pr->power.states[i].demotion.state -
1162                                     pr->power.states));
1163                 else
1164                         seq_puts(seq, "demotion[--] ");
1165
1166                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1167                            pr->power.states[i].latency,
1168                            pr->power.states[i].usage,
1169                            (unsigned long long)pr->power.states[i].time);
1170         }
1171
1172       end:
1173         return 0;
1174 }
1175
1176 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1177 {
1178         return single_open(file, acpi_processor_power_seq_show,
1179                            PDE(inode)->data);
1180 }
1181
1182 static const struct file_operations acpi_processor_power_fops = {
1183         .open = acpi_processor_power_open_fs,
1184         .read = seq_read,
1185         .llseek = seq_lseek,
1186         .release = single_release,
1187 };
1188
1189 #ifdef CONFIG_SMP
1190 static void smp_callback(void *v)
1191 {
1192         /* we already woke the CPU up, nothing more to do */
1193 }
1194
1195 /*
1196  * This function gets called when a part of the kernel has a new latency
1197  * requirement.  This means we need to get all processors out of their C-state,
1198  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1199  * wakes them all right up.
1200  */
1201 static int acpi_processor_latency_notify(struct notifier_block *b,
1202                 unsigned long l, void *v)
1203 {
1204         smp_call_function(smp_callback, NULL, 0, 1);
1205         return NOTIFY_OK;
1206 }
1207
1208 static struct notifier_block acpi_processor_latency_notifier = {
1209         .notifier_call = acpi_processor_latency_notify,
1210 };
1211 #endif
1212
1213 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1214                               struct acpi_device *device)
1215 {
1216         acpi_status status = 0;
1217         static int first_run;
1218         struct proc_dir_entry *entry = NULL;
1219         unsigned int i;
1220
1221
1222         if (!first_run) {
1223                 dmi_check_system(processor_power_dmi_table);
1224                 if (max_cstate < ACPI_C_STATES_MAX)
1225                         printk(KERN_NOTICE
1226                                "ACPI: processor limited to max C-state %d\n",
1227                                max_cstate);
1228                 first_run++;
1229 #ifdef CONFIG_SMP
1230                 register_latency_notifier(&acpi_processor_latency_notifier);
1231 #endif
1232         }
1233
1234         if (!pr)
1235                 return -EINVAL;
1236
1237         if (acpi_gbl_FADT.cst_control && !nocst) {
1238                 status =
1239                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1240                 if (ACPI_FAILURE(status)) {
1241                         ACPI_EXCEPTION((AE_INFO, status,
1242                                         "Notifying BIOS of _CST ability failed"));
1243                 }
1244         }
1245
1246         acpi_processor_get_power_info(pr);
1247
1248         /*
1249          * Install the idle handler if processor power management is supported.
1250          * Note that we use previously set idle handler will be used on
1251          * platforms that only support C1.
1252          */
1253         if ((pr->flags.power) && (!boot_option_idle_override)) {
1254                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1255                 for (i = 1; i <= pr->power.count; i++)
1256                         if (pr->power.states[i].valid)
1257                                 printk(" C%d[C%d]", i,
1258                                        pr->power.states[i].type);
1259                 printk(")\n");
1260
1261                 if (pr->id == 0) {
1262                         pm_idle_save = pm_idle;
1263                         pm_idle = acpi_processor_idle;
1264                 }
1265         }
1266
1267         /* 'power' [R] */
1268         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1269                                   S_IRUGO, acpi_device_dir(device));
1270         if (!entry)
1271                 return -EIO;
1272         else {
1273                 entry->proc_fops = &acpi_processor_power_fops;
1274                 entry->data = acpi_driver_data(device);
1275                 entry->owner = THIS_MODULE;
1276         }
1277
1278         pr->flags.power_setup_done = 1;
1279
1280         return 0;
1281 }
1282
1283 int acpi_processor_power_exit(struct acpi_processor *pr,
1284                               struct acpi_device *device)
1285 {
1286
1287         pr->flags.power_setup_done = 0;
1288
1289         if (acpi_device_dir(device))
1290                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1291                                   acpi_device_dir(device));
1292
1293         /* Unregister the idle handler when processor #0 is removed. */
1294         if (pr->id == 0) {
1295                 pm_idle = pm_idle_save;
1296
1297                 /*
1298                  * We are about to unload the current idle thread pm callback
1299                  * (pm_idle), Wait for all processors to update cached/local
1300                  * copies of pm_idle before proceeding.
1301                  */
1302                 cpu_idle_wait();
1303 #ifdef CONFIG_SMP
1304                 unregister_latency_notifier(&acpi_processor_latency_notifier);
1305 #endif
1306         }
1307
1308         return 0;
1309 }