ACPI: consolidate ACPI_*_COMPONENT definitions in acpi_drivers.h
[linux-3.10.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/pm_qos_params.h>
42 #include <linux/clockchips.h>
43 #include <linux/cpuidle.h>
44
45 /*
46  * Include the apic definitions for x86 to have the APIC timer related defines
47  * available also for UP (on SMP it gets magically included via linux/smp.h).
48  * asm/acpi.h is not an option, as it would require more include magic. Also
49  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
50  */
51 #ifdef CONFIG_X86
52 #include <asm/apic.h>
53 #endif
54
55 #include <asm/io.h>
56 #include <asm/uaccess.h>
57
58 #include <acpi/acpi_bus.h>
59 #include <acpi/processor.h>
60 #include <asm/processor.h>
61
62 #define ACPI_PROCESSOR_CLASS            "processor"
63 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
64 ACPI_MODULE_NAME("processor_idle");
65 #define ACPI_PROCESSOR_FILE_POWER       "power"
66 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
67 #define PM_TIMER_TICK_NS                (1000000000ULL/PM_TIMER_FREQUENCY)
68 #ifndef CONFIG_CPU_IDLE
69 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
70 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
71 static void (*pm_idle_save) (void) __read_mostly;
72 #else
73 #define C2_OVERHEAD                     1       /* 1us */
74 #define C3_OVERHEAD                     1       /* 1us */
75 #endif
76 #define PM_TIMER_TICKS_TO_US(p)         (((p) * 1000)/(PM_TIMER_FREQUENCY/1000))
77
78 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
79 #ifdef CONFIG_CPU_IDLE
80 module_param(max_cstate, uint, 0000);
81 #else
82 module_param(max_cstate, uint, 0644);
83 #endif
84 static unsigned int nocst __read_mostly;
85 module_param(nocst, uint, 0000);
86
87 #ifndef CONFIG_CPU_IDLE
88 /*
89  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
90  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
91  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
92  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
93  * reduce history for more aggressive entry into C3
94  */
95 static unsigned int bm_history __read_mostly =
96     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
97 module_param(bm_history, uint, 0644);
98
99 static int acpi_processor_set_power_policy(struct acpi_processor *pr);
100
101 #else   /* CONFIG_CPU_IDLE */
102 static unsigned int latency_factor __read_mostly = 2;
103 module_param(latency_factor, uint, 0644);
104 #endif
105
106 /*
107  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
108  * For now disable this. Probably a bug somewhere else.
109  *
110  * To skip this limit, boot/load with a large max_cstate limit.
111  */
112 static int set_max_cstate(const struct dmi_system_id *id)
113 {
114         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
115                 return 0;
116
117         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
118                " Override with \"processor.max_cstate=%d\"\n", id->ident,
119                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
120
121         max_cstate = (long)id->driver_data;
122
123         return 0;
124 }
125
126 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
127    callers to only run once -AK */
128 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
129         { set_max_cstate, "IBM ThinkPad R40e", {
130           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
131           DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
132         { set_max_cstate, "IBM ThinkPad R40e", {
133           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
134           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
135         { set_max_cstate, "IBM ThinkPad R40e", {
136           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
137           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
138         { set_max_cstate, "IBM ThinkPad R40e", {
139           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
140           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
141         { set_max_cstate, "IBM ThinkPad R40e", {
142           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
143           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
144         { set_max_cstate, "IBM ThinkPad R40e", {
145           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
146           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
147         { set_max_cstate, "IBM ThinkPad R40e", {
148           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
149           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
150         { set_max_cstate, "IBM ThinkPad R40e", {
151           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
152           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
153         { set_max_cstate, "IBM ThinkPad R40e", {
154           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
155           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
156         { set_max_cstate, "IBM ThinkPad R40e", {
157           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
158           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
159         { set_max_cstate, "IBM ThinkPad R40e", {
160           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
161           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
162         { set_max_cstate, "IBM ThinkPad R40e", {
163           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
164           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
165         { set_max_cstate, "IBM ThinkPad R40e", {
166           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
167           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
168         { set_max_cstate, "IBM ThinkPad R40e", {
169           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
170           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
171         { set_max_cstate, "IBM ThinkPad R40e", {
172           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
173           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
174         { set_max_cstate, "IBM ThinkPad R40e", {
175           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
176           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
177         { set_max_cstate, "Medion 41700", {
178           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
179           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
180         { set_max_cstate, "Clevo 5600D", {
181           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
182           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
183          (void *)2},
184         {},
185 };
186
187 static inline u32 ticks_elapsed(u32 t1, u32 t2)
188 {
189         if (t2 >= t1)
190                 return (t2 - t1);
191         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
192                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
193         else
194                 return ((0xFFFFFFFF - t1) + t2);
195 }
196
197 static inline u32 ticks_elapsed_in_us(u32 t1, u32 t2)
198 {
199         if (t2 >= t1)
200                 return PM_TIMER_TICKS_TO_US(t2 - t1);
201         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
202                 return PM_TIMER_TICKS_TO_US(((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
203         else
204                 return PM_TIMER_TICKS_TO_US((0xFFFFFFFF - t1) + t2);
205 }
206
207 /*
208  * Callers should disable interrupts before the call and enable
209  * interrupts after return.
210  */
211 static void acpi_safe_halt(void)
212 {
213         current_thread_info()->status &= ~TS_POLLING;
214         /*
215          * TS_POLLING-cleared state must be visible before we
216          * test NEED_RESCHED:
217          */
218         smp_mb();
219         if (!need_resched()) {
220                 safe_halt();
221                 local_irq_disable();
222         }
223         current_thread_info()->status |= TS_POLLING;
224 }
225
226 #ifndef CONFIG_CPU_IDLE
227
228 static void
229 acpi_processor_power_activate(struct acpi_processor *pr,
230                               struct acpi_processor_cx *new)
231 {
232         struct acpi_processor_cx *old;
233
234         if (!pr || !new)
235                 return;
236
237         old = pr->power.state;
238
239         if (old)
240                 old->promotion.count = 0;
241         new->demotion.count = 0;
242
243         /* Cleanup from old state. */
244         if (old) {
245                 switch (old->type) {
246                 case ACPI_STATE_C3:
247                         /* Disable bus master reload */
248                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
249                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
250                         break;
251                 }
252         }
253
254         /* Prepare to use new state. */
255         switch (new->type) {
256         case ACPI_STATE_C3:
257                 /* Enable bus master reload */
258                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
259                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
260                 break;
261         }
262
263         pr->power.state = new;
264
265         return;
266 }
267
268 static atomic_t c3_cpu_count;
269
270 /* Common C-state entry for C2, C3, .. */
271 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
272 {
273         /* Don't trace irqs off for idle */
274         stop_critical_timings();
275         if (cstate->entry_method == ACPI_CSTATE_FFH) {
276                 /* Call into architectural FFH based C-state */
277                 acpi_processor_ffh_cstate_enter(cstate);
278         } else {
279                 int unused;
280                 /* IO port based C-state */
281                 inb(cstate->address);
282                 /* Dummy wait op - must do something useless after P_LVL2 read
283                    because chipsets cannot guarantee that STPCLK# signal
284                    gets asserted in time to freeze execution properly. */
285                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
286         }
287         start_critical_timings();
288 }
289 #endif /* !CONFIG_CPU_IDLE */
290
291 #ifdef ARCH_APICTIMER_STOPS_ON_C3
292
293 /*
294  * Some BIOS implementations switch to C3 in the published C2 state.
295  * This seems to be a common problem on AMD boxen, but other vendors
296  * are affected too. We pick the most conservative approach: we assume
297  * that the local APIC stops in both C2 and C3.
298  */
299 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
300                                    struct acpi_processor_cx *cx)
301 {
302         struct acpi_processor_power *pwr = &pr->power;
303         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
304
305         /*
306          * Check, if one of the previous states already marked the lapic
307          * unstable
308          */
309         if (pwr->timer_broadcast_on_state < state)
310                 return;
311
312         if (cx->type >= type)
313                 pr->power.timer_broadcast_on_state = state;
314 }
315
316 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr)
317 {
318         unsigned long reason;
319
320         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
321                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
322
323         clockevents_notify(reason, &pr->id);
324 }
325
326 /* Power(C) State timer broadcast control */
327 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
328                                        struct acpi_processor_cx *cx,
329                                        int broadcast)
330 {
331         int state = cx - pr->power.states;
332
333         if (state >= pr->power.timer_broadcast_on_state) {
334                 unsigned long reason;
335
336                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
337                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
338                 clockevents_notify(reason, &pr->id);
339         }
340 }
341
342 #else
343
344 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
345                                    struct acpi_processor_cx *cstate) { }
346 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr) { }
347 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
348                                        struct acpi_processor_cx *cx,
349                                        int broadcast)
350 {
351 }
352
353 #endif
354
355 /*
356  * Suspend / resume control
357  */
358 static int acpi_idle_suspend;
359
360 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
361 {
362         acpi_idle_suspend = 1;
363         return 0;
364 }
365
366 int acpi_processor_resume(struct acpi_device * device)
367 {
368         acpi_idle_suspend = 0;
369         return 0;
370 }
371
372 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
373 static int tsc_halts_in_c(int state)
374 {
375         switch (boot_cpu_data.x86_vendor) {
376         case X86_VENDOR_AMD:
377                 /*
378                  * AMD Fam10h TSC will tick in all
379                  * C/P/S0/S1 states when this bit is set.
380                  */
381                 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
382                         return 0;
383                 /*FALL THROUGH*/
384         case X86_VENDOR_INTEL:
385                 /* Several cases known where TSC halts in C2 too */
386         default:
387                 return state > ACPI_STATE_C1;
388         }
389 }
390 #endif
391
392 #ifndef CONFIG_CPU_IDLE
393 static void acpi_processor_idle(void)
394 {
395         struct acpi_processor *pr = NULL;
396         struct acpi_processor_cx *cx = NULL;
397         struct acpi_processor_cx *next_state = NULL;
398         int sleep_ticks = 0;
399         u32 t1, t2 = 0;
400
401         /*
402          * Interrupts must be disabled during bus mastering calculations and
403          * for C2/C3 transitions.
404          */
405         local_irq_disable();
406
407         pr = __get_cpu_var(processors);
408         if (!pr) {
409                 local_irq_enable();
410                 return;
411         }
412
413         /*
414          * Check whether we truly need to go idle, or should
415          * reschedule:
416          */
417         if (unlikely(need_resched())) {
418                 local_irq_enable();
419                 return;
420         }
421
422         cx = pr->power.state;
423         if (!cx || acpi_idle_suspend) {
424                 if (pm_idle_save) {
425                         pm_idle_save(); /* enables IRQs */
426                 } else {
427                         acpi_safe_halt();
428                         local_irq_enable();
429                 }
430
431                 return;
432         }
433
434         /*
435          * Check BM Activity
436          * -----------------
437          * Check for bus mastering activity (if required), record, and check
438          * for demotion.
439          */
440         if (pr->flags.bm_check) {
441                 u32 bm_status = 0;
442                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
443
444                 if (diff > 31)
445                         diff = 31;
446
447                 pr->power.bm_activity <<= diff;
448
449                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
450                 if (bm_status) {
451                         pr->power.bm_activity |= 0x1;
452                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
453                 }
454                 /*
455                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
456                  * the true state of bus mastering activity; forcing us to
457                  * manually check the BMIDEA bit of each IDE channel.
458                  */
459                 else if (errata.piix4.bmisx) {
460                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
461                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
462                                 pr->power.bm_activity |= 0x1;
463                 }
464
465                 pr->power.bm_check_timestamp = jiffies;
466
467                 /*
468                  * If bus mastering is or was active this jiffy, demote
469                  * to avoid a faulty transition.  Note that the processor
470                  * won't enter a low-power state during this call (to this
471                  * function) but should upon the next.
472                  *
473                  * TBD: A better policy might be to fallback to the demotion
474                  *      state (use it for this quantum only) istead of
475                  *      demoting -- and rely on duration as our sole demotion
476                  *      qualification.  This may, however, introduce DMA
477                  *      issues (e.g. floppy DMA transfer overrun/underrun).
478                  */
479                 if ((pr->power.bm_activity & 0x1) &&
480                     cx->demotion.threshold.bm) {
481                         local_irq_enable();
482                         next_state = cx->demotion.state;
483                         goto end;
484                 }
485         }
486
487 #ifdef CONFIG_HOTPLUG_CPU
488         /*
489          * Check for P_LVL2_UP flag before entering C2 and above on
490          * an SMP system. We do it here instead of doing it at _CST/P_LVL
491          * detection phase, to work cleanly with logical CPU hotplug.
492          */
493         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
494             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
495                 cx = &pr->power.states[ACPI_STATE_C1];
496 #endif
497
498         /*
499          * Sleep:
500          * ------
501          * Invoke the current Cx state to put the processor to sleep.
502          */
503         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
504                 current_thread_info()->status &= ~TS_POLLING;
505                 /*
506                  * TS_POLLING-cleared state must be visible before we
507                  * test NEED_RESCHED:
508                  */
509                 smp_mb();
510                 if (need_resched()) {
511                         current_thread_info()->status |= TS_POLLING;
512                         local_irq_enable();
513                         return;
514                 }
515         }
516
517         switch (cx->type) {
518
519         case ACPI_STATE_C1:
520                 /*
521                  * Invoke C1.
522                  * Use the appropriate idle routine, the one that would
523                  * be used without acpi C-states.
524                  */
525                 if (pm_idle_save) {
526                         pm_idle_save(); /* enables IRQs */
527                 } else {
528                         acpi_safe_halt();
529                         local_irq_enable();
530                 }
531
532                 /*
533                  * TBD: Can't get time duration while in C1, as resumes
534                  *      go to an ISR rather than here.  Need to instrument
535                  *      base interrupt handler.
536                  *
537                  * Note: the TSC better not stop in C1, sched_clock() will
538                  *       skew otherwise.
539                  */
540                 sleep_ticks = 0xFFFFFFFF;
541
542                 break;
543
544         case ACPI_STATE_C2:
545                 /* Get start time (ticks) */
546                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
547                 /* Tell the scheduler that we are going deep-idle: */
548                 sched_clock_idle_sleep_event();
549                 /* Invoke C2 */
550                 acpi_state_timer_broadcast(pr, cx, 1);
551                 acpi_cstate_enter(cx);
552                 /* Get end time (ticks) */
553                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
554
555 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
556                 /* TSC halts in C2, so notify users */
557                 if (tsc_halts_in_c(ACPI_STATE_C2))
558                         mark_tsc_unstable("possible TSC halt in C2");
559 #endif
560                 /* Compute time (ticks) that we were actually asleep */
561                 sleep_ticks = ticks_elapsed(t1, t2);
562
563                 /* Tell the scheduler how much we idled: */
564                 sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
565
566                 /* Re-enable interrupts */
567                 local_irq_enable();
568                 /* Do not account our idle-switching overhead: */
569                 sleep_ticks -= cx->latency_ticks + C2_OVERHEAD;
570
571                 current_thread_info()->status |= TS_POLLING;
572                 acpi_state_timer_broadcast(pr, cx, 0);
573                 break;
574
575         case ACPI_STATE_C3:
576                 acpi_unlazy_tlb(smp_processor_id());
577                 /*
578                  * Must be done before busmaster disable as we might
579                  * need to access HPET !
580                  */
581                 acpi_state_timer_broadcast(pr, cx, 1);
582                 /*
583                  * disable bus master
584                  * bm_check implies we need ARB_DIS
585                  * !bm_check implies we need cache flush
586                  * bm_control implies whether we can do ARB_DIS
587                  *
588                  * That leaves a case where bm_check is set and bm_control is
589                  * not set. In that case we cannot do much, we enter C3
590                  * without doing anything.
591                  */
592                 if (pr->flags.bm_check && pr->flags.bm_control) {
593                         if (atomic_inc_return(&c3_cpu_count) ==
594                             num_online_cpus()) {
595                                 /*
596                                  * All CPUs are trying to go to C3
597                                  * Disable bus master arbitration
598                                  */
599                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
600                         }
601                 } else if (!pr->flags.bm_check) {
602                         /* SMP with no shared cache... Invalidate cache  */
603                         ACPI_FLUSH_CPU_CACHE();
604                 }
605
606                 /* Get start time (ticks) */
607                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
608                 /* Invoke C3 */
609                 /* Tell the scheduler that we are going deep-idle: */
610                 sched_clock_idle_sleep_event();
611                 acpi_cstate_enter(cx);
612                 /* Get end time (ticks) */
613                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
614                 if (pr->flags.bm_check && pr->flags.bm_control) {
615                         /* Enable bus master arbitration */
616                         atomic_dec(&c3_cpu_count);
617                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
618                 }
619
620 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
621                 /* TSC halts in C3, so notify users */
622                 if (tsc_halts_in_c(ACPI_STATE_C3))
623                         mark_tsc_unstable("TSC halts in C3");
624 #endif
625                 /* Compute time (ticks) that we were actually asleep */
626                 sleep_ticks = ticks_elapsed(t1, t2);
627                 /* Tell the scheduler how much we idled: */
628                 sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
629
630                 /* Re-enable interrupts */
631                 local_irq_enable();
632                 /* Do not account our idle-switching overhead: */
633                 sleep_ticks -= cx->latency_ticks + C3_OVERHEAD;
634
635                 current_thread_info()->status |= TS_POLLING;
636                 acpi_state_timer_broadcast(pr, cx, 0);
637                 break;
638
639         default:
640                 local_irq_enable();
641                 return;
642         }
643         cx->usage++;
644         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
645                 cx->time += sleep_ticks;
646
647         next_state = pr->power.state;
648
649 #ifdef CONFIG_HOTPLUG_CPU
650         /* Don't do promotion/demotion */
651         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
652             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
653                 next_state = cx;
654                 goto end;
655         }
656 #endif
657
658         /*
659          * Promotion?
660          * ----------
661          * Track the number of longs (time asleep is greater than threshold)
662          * and promote when the count threshold is reached.  Note that bus
663          * mastering activity may prevent promotions.
664          * Do not promote above max_cstate.
665          */
666         if (cx->promotion.state &&
667             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
668                 if (sleep_ticks > cx->promotion.threshold.ticks &&
669                   cx->promotion.state->latency <=
670                                 pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY)) {
671                         cx->promotion.count++;
672                         cx->demotion.count = 0;
673                         if (cx->promotion.count >=
674                             cx->promotion.threshold.count) {
675                                 if (pr->flags.bm_check) {
676                                         if (!
677                                             (pr->power.bm_activity & cx->
678                                              promotion.threshold.bm)) {
679                                                 next_state =
680                                                     cx->promotion.state;
681                                                 goto end;
682                                         }
683                                 } else {
684                                         next_state = cx->promotion.state;
685                                         goto end;
686                                 }
687                         }
688                 }
689         }
690
691         /*
692          * Demotion?
693          * ---------
694          * Track the number of shorts (time asleep is less than time threshold)
695          * and demote when the usage threshold is reached.
696          */
697         if (cx->demotion.state) {
698                 if (sleep_ticks < cx->demotion.threshold.ticks) {
699                         cx->demotion.count++;
700                         cx->promotion.count = 0;
701                         if (cx->demotion.count >= cx->demotion.threshold.count) {
702                                 next_state = cx->demotion.state;
703                                 goto end;
704                         }
705                 }
706         }
707
708       end:
709         /*
710          * Demote if current state exceeds max_cstate
711          * or if the latency of the current state is unacceptable
712          */
713         if ((pr->power.state - pr->power.states) > max_cstate ||
714                 pr->power.state->latency >
715                                 pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY)) {
716                 if (cx->demotion.state)
717                         next_state = cx->demotion.state;
718         }
719
720         /*
721          * New Cx State?
722          * -------------
723          * If we're going to start using a new Cx state we must clean up
724          * from the previous and prepare to use the new.
725          */
726         if (next_state != pr->power.state)
727                 acpi_processor_power_activate(pr, next_state);
728 }
729
730 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
731 {
732         unsigned int i;
733         unsigned int state_is_set = 0;
734         struct acpi_processor_cx *lower = NULL;
735         struct acpi_processor_cx *higher = NULL;
736         struct acpi_processor_cx *cx;
737
738
739         if (!pr)
740                 return -EINVAL;
741
742         /*
743          * This function sets the default Cx state policy (OS idle handler).
744          * Our scheme is to promote quickly to C2 but more conservatively
745          * to C3.  We're favoring C2  for its characteristics of low latency
746          * (quick response), good power savings, and ability to allow bus
747          * mastering activity.  Note that the Cx state policy is completely
748          * customizable and can be altered dynamically.
749          */
750
751         /* startup state */
752         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
753                 cx = &pr->power.states[i];
754                 if (!cx->valid)
755                         continue;
756
757                 if (!state_is_set)
758                         pr->power.state = cx;
759                 state_is_set++;
760                 break;
761         }
762
763         if (!state_is_set)
764                 return -ENODEV;
765
766         /* demotion */
767         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
768                 cx = &pr->power.states[i];
769                 if (!cx->valid)
770                         continue;
771
772                 if (lower) {
773                         cx->demotion.state = lower;
774                         cx->demotion.threshold.ticks = cx->latency_ticks;
775                         cx->demotion.threshold.count = 1;
776                         if (cx->type == ACPI_STATE_C3)
777                                 cx->demotion.threshold.bm = bm_history;
778                 }
779
780                 lower = cx;
781         }
782
783         /* promotion */
784         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
785                 cx = &pr->power.states[i];
786                 if (!cx->valid)
787                         continue;
788
789                 if (higher) {
790                         cx->promotion.state = higher;
791                         cx->promotion.threshold.ticks = cx->latency_ticks;
792                         if (cx->type >= ACPI_STATE_C2)
793                                 cx->promotion.threshold.count = 4;
794                         else
795                                 cx->promotion.threshold.count = 10;
796                         if (higher->type == ACPI_STATE_C3)
797                                 cx->promotion.threshold.bm = bm_history;
798                 }
799
800                 higher = cx;
801         }
802
803         return 0;
804 }
805 #endif /* !CONFIG_CPU_IDLE */
806
807 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
808 {
809
810         if (!pr)
811                 return -EINVAL;
812
813         if (!pr->pblk)
814                 return -ENODEV;
815
816         /* if info is obtained from pblk/fadt, type equals state */
817         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
818         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
819
820 #ifndef CONFIG_HOTPLUG_CPU
821         /*
822          * Check for P_LVL2_UP flag before entering C2 and above on
823          * an SMP system.
824          */
825         if ((num_online_cpus() > 1) &&
826             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
827                 return -ENODEV;
828 #endif
829
830         /* determine C2 and C3 address from pblk */
831         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
832         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
833
834         /* determine latencies from FADT */
835         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
836         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
837
838         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
839                           "lvl2[0x%08x] lvl3[0x%08x]\n",
840                           pr->power.states[ACPI_STATE_C2].address,
841                           pr->power.states[ACPI_STATE_C3].address));
842
843         return 0;
844 }
845
846 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
847 {
848         if (!pr->power.states[ACPI_STATE_C1].valid) {
849                 /* set the first C-State to C1 */
850                 /* all processors need to support C1 */
851                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
852                 pr->power.states[ACPI_STATE_C1].valid = 1;
853                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
854         }
855         /* the C0 state only exists as a filler in our array */
856         pr->power.states[ACPI_STATE_C0].valid = 1;
857         return 0;
858 }
859
860 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
861 {
862         acpi_status status = 0;
863         acpi_integer count;
864         int current_count;
865         int i;
866         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
867         union acpi_object *cst;
868
869
870         if (nocst)
871                 return -ENODEV;
872
873         current_count = 0;
874
875         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
876         if (ACPI_FAILURE(status)) {
877                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
878                 return -ENODEV;
879         }
880
881         cst = buffer.pointer;
882
883         /* There must be at least 2 elements */
884         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
885                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
886                 status = -EFAULT;
887                 goto end;
888         }
889
890         count = cst->package.elements[0].integer.value;
891
892         /* Validate number of power states. */
893         if (count < 1 || count != cst->package.count - 1) {
894                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
895                 status = -EFAULT;
896                 goto end;
897         }
898
899         /* Tell driver that at least _CST is supported. */
900         pr->flags.has_cst = 1;
901
902         for (i = 1; i <= count; i++) {
903                 union acpi_object *element;
904                 union acpi_object *obj;
905                 struct acpi_power_register *reg;
906                 struct acpi_processor_cx cx;
907
908                 memset(&cx, 0, sizeof(cx));
909
910                 element = &(cst->package.elements[i]);
911                 if (element->type != ACPI_TYPE_PACKAGE)
912                         continue;
913
914                 if (element->package.count != 4)
915                         continue;
916
917                 obj = &(element->package.elements[0]);
918
919                 if (obj->type != ACPI_TYPE_BUFFER)
920                         continue;
921
922                 reg = (struct acpi_power_register *)obj->buffer.pointer;
923
924                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
925                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
926                         continue;
927
928                 /* There should be an easy way to extract an integer... */
929                 obj = &(element->package.elements[1]);
930                 if (obj->type != ACPI_TYPE_INTEGER)
931                         continue;
932
933                 cx.type = obj->integer.value;
934                 /*
935                  * Some buggy BIOSes won't list C1 in _CST -
936                  * Let acpi_processor_get_power_info_default() handle them later
937                  */
938                 if (i == 1 && cx.type != ACPI_STATE_C1)
939                         current_count++;
940
941                 cx.address = reg->address;
942                 cx.index = current_count + 1;
943
944                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
945                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
946                         if (acpi_processor_ffh_cstate_probe
947                                         (pr->id, &cx, reg) == 0) {
948                                 cx.entry_method = ACPI_CSTATE_FFH;
949                         } else if (cx.type == ACPI_STATE_C1) {
950                                 /*
951                                  * C1 is a special case where FIXED_HARDWARE
952                                  * can be handled in non-MWAIT way as well.
953                                  * In that case, save this _CST entry info.
954                                  * Otherwise, ignore this info and continue.
955                                  */
956                                 cx.entry_method = ACPI_CSTATE_HALT;
957                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
958                         } else {
959                                 continue;
960                         }
961                         if (cx.type == ACPI_STATE_C1 &&
962                                         (idle_halt || idle_nomwait)) {
963                                 /*
964                                  * In most cases the C1 space_id obtained from
965                                  * _CST object is FIXED_HARDWARE access mode.
966                                  * But when the option of idle=halt is added,
967                                  * the entry_method type should be changed from
968                                  * CSTATE_FFH to CSTATE_HALT.
969                                  * When the option of idle=nomwait is added,
970                                  * the C1 entry_method type should be
971                                  * CSTATE_HALT.
972                                  */
973                                 cx.entry_method = ACPI_CSTATE_HALT;
974                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
975                         }
976                 } else {
977                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
978                                  cx.address);
979                 }
980
981                 if (cx.type == ACPI_STATE_C1) {
982                         cx.valid = 1;
983                 }
984
985                 obj = &(element->package.elements[2]);
986                 if (obj->type != ACPI_TYPE_INTEGER)
987                         continue;
988
989                 cx.latency = obj->integer.value;
990
991                 obj = &(element->package.elements[3]);
992                 if (obj->type != ACPI_TYPE_INTEGER)
993                         continue;
994
995                 cx.power = obj->integer.value;
996
997                 current_count++;
998                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
999
1000                 /*
1001                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
1002                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
1003                  */
1004                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
1005                         printk(KERN_WARNING
1006                                "Limiting number of power states to max (%d)\n",
1007                                ACPI_PROCESSOR_MAX_POWER);
1008                         printk(KERN_WARNING
1009                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
1010                         break;
1011                 }
1012         }
1013
1014         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
1015                           current_count));
1016
1017         /* Validate number of power states discovered */
1018         if (current_count < 2)
1019                 status = -EFAULT;
1020
1021       end:
1022         kfree(buffer.pointer);
1023
1024         return status;
1025 }
1026
1027 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
1028 {
1029
1030         if (!cx->address)
1031                 return;
1032
1033         /*
1034          * C2 latency must be less than or equal to 100
1035          * microseconds.
1036          */
1037         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
1038                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1039                                   "latency too large [%d]\n", cx->latency));
1040                 return;
1041         }
1042
1043         /*
1044          * Otherwise we've met all of our C2 requirements.
1045          * Normalize the C2 latency to expidite policy
1046          */
1047         cx->valid = 1;
1048
1049 #ifndef CONFIG_CPU_IDLE
1050         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1051 #else
1052         cx->latency_ticks = cx->latency;
1053 #endif
1054
1055         return;
1056 }
1057
1058 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
1059                                            struct acpi_processor_cx *cx)
1060 {
1061         static int bm_check_flag;
1062
1063
1064         if (!cx->address)
1065                 return;
1066
1067         /*
1068          * C3 latency must be less than or equal to 1000
1069          * microseconds.
1070          */
1071         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
1072                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1073                                   "latency too large [%d]\n", cx->latency));
1074                 return;
1075         }
1076
1077         /*
1078          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
1079          * DMA transfers are used by any ISA device to avoid livelock.
1080          * Note that we could disable Type-F DMA (as recommended by
1081          * the erratum), but this is known to disrupt certain ISA
1082          * devices thus we take the conservative approach.
1083          */
1084         else if (errata.piix4.fdma) {
1085                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1086                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
1087                 return;
1088         }
1089
1090         /* All the logic here assumes flags.bm_check is same across all CPUs */
1091         if (!bm_check_flag) {
1092                 /* Determine whether bm_check is needed based on CPU  */
1093                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
1094                 bm_check_flag = pr->flags.bm_check;
1095         } else {
1096                 pr->flags.bm_check = bm_check_flag;
1097         }
1098
1099         if (pr->flags.bm_check) {
1100                 if (!pr->flags.bm_control) {
1101                         if (pr->flags.has_cst != 1) {
1102                                 /* bus mastering control is necessary */
1103                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1104                                         "C3 support requires BM control\n"));
1105                                 return;
1106                         } else {
1107                                 /* Here we enter C3 without bus mastering */
1108                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1109                                         "C3 support without BM control\n"));
1110                         }
1111                 }
1112         } else {
1113                 /*
1114                  * WBINVD should be set in fadt, for C3 state to be
1115                  * supported on when bm_check is not required.
1116                  */
1117                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
1118                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1119                                           "Cache invalidation should work properly"
1120                                           " for C3 to be enabled on SMP systems\n"));
1121                         return;
1122                 }
1123                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1124         }
1125
1126         /*
1127          * Otherwise we've met all of our C3 requirements.
1128          * Normalize the C3 latency to expidite policy.  Enable
1129          * checking of bus mastering status (bm_check) so we can
1130          * use this in our C3 policy
1131          */
1132         cx->valid = 1;
1133
1134 #ifndef CONFIG_CPU_IDLE
1135         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1136 #else
1137         cx->latency_ticks = cx->latency;
1138 #endif
1139
1140         return;
1141 }
1142
1143 static int acpi_processor_power_verify(struct acpi_processor *pr)
1144 {
1145         unsigned int i;
1146         unsigned int working = 0;
1147
1148         pr->power.timer_broadcast_on_state = INT_MAX;
1149
1150         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1151                 struct acpi_processor_cx *cx = &pr->power.states[i];
1152
1153                 switch (cx->type) {
1154                 case ACPI_STATE_C1:
1155                         cx->valid = 1;
1156                         break;
1157
1158                 case ACPI_STATE_C2:
1159                         acpi_processor_power_verify_c2(cx);
1160                         if (cx->valid)
1161                                 acpi_timer_check_state(i, pr, cx);
1162                         break;
1163
1164                 case ACPI_STATE_C3:
1165                         acpi_processor_power_verify_c3(pr, cx);
1166                         if (cx->valid)
1167                                 acpi_timer_check_state(i, pr, cx);
1168                         break;
1169                 }
1170
1171                 if (cx->valid)
1172                         working++;
1173         }
1174
1175         acpi_propagate_timer_broadcast(pr);
1176
1177         return (working);
1178 }
1179
1180 static int acpi_processor_get_power_info(struct acpi_processor *pr)
1181 {
1182         unsigned int i;
1183         int result;
1184
1185
1186         /* NOTE: the idle thread may not be running while calling
1187          * this function */
1188
1189         /* Zero initialize all the C-states info. */
1190         memset(pr->power.states, 0, sizeof(pr->power.states));
1191
1192         result = acpi_processor_get_power_info_cst(pr);
1193         if (result == -ENODEV)
1194                 result = acpi_processor_get_power_info_fadt(pr);
1195
1196         if (result)
1197                 return result;
1198
1199         acpi_processor_get_power_info_default(pr);
1200
1201         pr->power.count = acpi_processor_power_verify(pr);
1202
1203 #ifndef CONFIG_CPU_IDLE
1204         /*
1205          * Set Default Policy
1206          * ------------------
1207          * Now that we know which states are supported, set the default
1208          * policy.  Note that this policy can be changed dynamically
1209          * (e.g. encourage deeper sleeps to conserve battery life when
1210          * not on AC).
1211          */
1212         result = acpi_processor_set_power_policy(pr);
1213         if (result)
1214                 return result;
1215 #endif
1216
1217         /*
1218          * if one state of type C2 or C3 is available, mark this
1219          * CPU as being "idle manageable"
1220          */
1221         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1222                 if (pr->power.states[i].valid) {
1223                         pr->power.count = i;
1224                         if (pr->power.states[i].type >= ACPI_STATE_C2)
1225                                 pr->flags.power = 1;
1226                 }
1227         }
1228
1229         return 0;
1230 }
1231
1232 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1233 {
1234         struct acpi_processor *pr = seq->private;
1235         unsigned int i;
1236
1237
1238         if (!pr)
1239                 goto end;
1240
1241         seq_printf(seq, "active state:            C%zd\n"
1242                    "max_cstate:              C%d\n"
1243                    "bus master activity:     %08x\n"
1244                    "maximum allowed latency: %d usec\n",
1245                    pr->power.state ? pr->power.state - pr->power.states : 0,
1246                    max_cstate, (unsigned)pr->power.bm_activity,
1247                    pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY));
1248
1249         seq_puts(seq, "states:\n");
1250
1251         for (i = 1; i <= pr->power.count; i++) {
1252                 seq_printf(seq, "   %cC%d:                  ",
1253                            (&pr->power.states[i] ==
1254                             pr->power.state ? '*' : ' '), i);
1255
1256                 if (!pr->power.states[i].valid) {
1257                         seq_puts(seq, "<not supported>\n");
1258                         continue;
1259                 }
1260
1261                 switch (pr->power.states[i].type) {
1262                 case ACPI_STATE_C1:
1263                         seq_printf(seq, "type[C1] ");
1264                         break;
1265                 case ACPI_STATE_C2:
1266                         seq_printf(seq, "type[C2] ");
1267                         break;
1268                 case ACPI_STATE_C3:
1269                         seq_printf(seq, "type[C3] ");
1270                         break;
1271                 default:
1272                         seq_printf(seq, "type[--] ");
1273                         break;
1274                 }
1275
1276                 if (pr->power.states[i].promotion.state)
1277                         seq_printf(seq, "promotion[C%zd] ",
1278                                    (pr->power.states[i].promotion.state -
1279                                     pr->power.states));
1280                 else
1281                         seq_puts(seq, "promotion[--] ");
1282
1283                 if (pr->power.states[i].demotion.state)
1284                         seq_printf(seq, "demotion[C%zd] ",
1285                                    (pr->power.states[i].demotion.state -
1286                                     pr->power.states));
1287                 else
1288                         seq_puts(seq, "demotion[--] ");
1289
1290                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1291                            pr->power.states[i].latency,
1292                            pr->power.states[i].usage,
1293                            (unsigned long long)pr->power.states[i].time);
1294         }
1295
1296       end:
1297         return 0;
1298 }
1299
1300 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1301 {
1302         return single_open(file, acpi_processor_power_seq_show,
1303                            PDE(inode)->data);
1304 }
1305
1306 static const struct file_operations acpi_processor_power_fops = {
1307         .owner = THIS_MODULE,
1308         .open = acpi_processor_power_open_fs,
1309         .read = seq_read,
1310         .llseek = seq_lseek,
1311         .release = single_release,
1312 };
1313
1314 #ifndef CONFIG_CPU_IDLE
1315
1316 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1317 {
1318         int result = 0;
1319
1320         if (boot_option_idle_override)
1321                 return 0;
1322
1323         if (!pr)
1324                 return -EINVAL;
1325
1326         if (nocst) {
1327                 return -ENODEV;
1328         }
1329
1330         if (!pr->flags.power_setup_done)
1331                 return -ENODEV;
1332
1333         /*
1334          * Fall back to the default idle loop, when pm_idle_save had
1335          * been initialized.
1336          */
1337         if (pm_idle_save) {
1338                 pm_idle = pm_idle_save;
1339                 /* Relies on interrupts forcing exit from idle. */
1340                 synchronize_sched();
1341         }
1342
1343         pr->flags.power = 0;
1344         result = acpi_processor_get_power_info(pr);
1345         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1346                 pm_idle = acpi_processor_idle;
1347
1348         return result;
1349 }
1350
1351 #ifdef CONFIG_SMP
1352 static void smp_callback(void *v)
1353 {
1354         /* we already woke the CPU up, nothing more to do */
1355 }
1356
1357 /*
1358  * This function gets called when a part of the kernel has a new latency
1359  * requirement.  This means we need to get all processors out of their C-state,
1360  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1361  * wakes them all right up.
1362  */
1363 static int acpi_processor_latency_notify(struct notifier_block *b,
1364                 unsigned long l, void *v)
1365 {
1366         smp_call_function(smp_callback, NULL, 1);
1367         return NOTIFY_OK;
1368 }
1369
1370 static struct notifier_block acpi_processor_latency_notifier = {
1371         .notifier_call = acpi_processor_latency_notify,
1372 };
1373
1374 #endif
1375
1376 #else /* CONFIG_CPU_IDLE */
1377
1378 /**
1379  * acpi_idle_bm_check - checks if bus master activity was detected
1380  */
1381 static int acpi_idle_bm_check(void)
1382 {
1383         u32 bm_status = 0;
1384
1385         acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
1386         if (bm_status)
1387                 acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
1388         /*
1389          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
1390          * the true state of bus mastering activity; forcing us to
1391          * manually check the BMIDEA bit of each IDE channel.
1392          */
1393         else if (errata.piix4.bmisx) {
1394                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
1395                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
1396                         bm_status = 1;
1397         }
1398         return bm_status;
1399 }
1400
1401 /**
1402  * acpi_idle_update_bm_rld - updates the BM_RLD bit depending on target state
1403  * @pr: the processor
1404  * @target: the new target state
1405  */
1406 static inline void acpi_idle_update_bm_rld(struct acpi_processor *pr,
1407                                            struct acpi_processor_cx *target)
1408 {
1409         if (pr->flags.bm_rld_set && target->type != ACPI_STATE_C3) {
1410                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1411                 pr->flags.bm_rld_set = 0;
1412         }
1413
1414         if (!pr->flags.bm_rld_set && target->type == ACPI_STATE_C3) {
1415                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
1416                 pr->flags.bm_rld_set = 1;
1417         }
1418 }
1419
1420 /**
1421  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
1422  * @cx: cstate data
1423  *
1424  * Caller disables interrupt before call and enables interrupt after return.
1425  */
1426 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
1427 {
1428         /* Don't trace irqs off for idle */
1429         stop_critical_timings();
1430         if (cx->entry_method == ACPI_CSTATE_FFH) {
1431                 /* Call into architectural FFH based C-state */
1432                 acpi_processor_ffh_cstate_enter(cx);
1433         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
1434                 acpi_safe_halt();
1435         } else {
1436                 int unused;
1437                 /* IO port based C-state */
1438                 inb(cx->address);
1439                 /* Dummy wait op - must do something useless after P_LVL2 read
1440                    because chipsets cannot guarantee that STPCLK# signal
1441                    gets asserted in time to freeze execution properly. */
1442                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
1443         }
1444         start_critical_timings();
1445 }
1446
1447 /**
1448  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
1449  * @dev: the target CPU
1450  * @state: the state data
1451  *
1452  * This is equivalent to the HALT instruction.
1453  */
1454 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
1455                               struct cpuidle_state *state)
1456 {
1457         u32 t1, t2;
1458         struct acpi_processor *pr;
1459         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1460
1461         pr = __get_cpu_var(processors);
1462
1463         if (unlikely(!pr))
1464                 return 0;
1465
1466         local_irq_disable();
1467
1468         /* Do not access any ACPI IO ports in suspend path */
1469         if (acpi_idle_suspend) {
1470                 acpi_safe_halt();
1471                 local_irq_enable();
1472                 return 0;
1473         }
1474
1475         if (pr->flags.bm_check)
1476                 acpi_idle_update_bm_rld(pr, cx);
1477
1478         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1479         acpi_idle_do_entry(cx);
1480         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1481
1482         local_irq_enable();
1483         cx->usage++;
1484
1485         return ticks_elapsed_in_us(t1, t2);
1486 }
1487
1488 /**
1489  * acpi_idle_enter_simple - enters an ACPI state without BM handling
1490  * @dev: the target CPU
1491  * @state: the state data
1492  */
1493 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
1494                                   struct cpuidle_state *state)
1495 {
1496         struct acpi_processor *pr;
1497         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1498         u32 t1, t2;
1499         int sleep_ticks = 0;
1500
1501         pr = __get_cpu_var(processors);
1502
1503         if (unlikely(!pr))
1504                 return 0;
1505
1506         if (acpi_idle_suspend)
1507                 return(acpi_idle_enter_c1(dev, state));
1508
1509         local_irq_disable();
1510         current_thread_info()->status &= ~TS_POLLING;
1511         /*
1512          * TS_POLLING-cleared state must be visible before we test
1513          * NEED_RESCHED:
1514          */
1515         smp_mb();
1516
1517         if (unlikely(need_resched())) {
1518                 current_thread_info()->status |= TS_POLLING;
1519                 local_irq_enable();
1520                 return 0;
1521         }
1522
1523         /*
1524          * Must be done before busmaster disable as we might need to
1525          * access HPET !
1526          */
1527         acpi_state_timer_broadcast(pr, cx, 1);
1528
1529         if (pr->flags.bm_check)
1530                 acpi_idle_update_bm_rld(pr, cx);
1531
1532         if (cx->type == ACPI_STATE_C3)
1533                 ACPI_FLUSH_CPU_CACHE();
1534
1535         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1536         /* Tell the scheduler that we are going deep-idle: */
1537         sched_clock_idle_sleep_event();
1538         acpi_idle_do_entry(cx);
1539         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1540
1541 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
1542         /* TSC could halt in idle, so notify users */
1543         if (tsc_halts_in_c(cx->type))
1544                 mark_tsc_unstable("TSC halts in idle");;
1545 #endif
1546         sleep_ticks = ticks_elapsed(t1, t2);
1547
1548         /* Tell the scheduler how much we idled: */
1549         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1550
1551         local_irq_enable();
1552         current_thread_info()->status |= TS_POLLING;
1553
1554         cx->usage++;
1555
1556         acpi_state_timer_broadcast(pr, cx, 0);
1557         cx->time += sleep_ticks;
1558         return ticks_elapsed_in_us(t1, t2);
1559 }
1560
1561 static int c3_cpu_count;
1562 static DEFINE_SPINLOCK(c3_lock);
1563
1564 /**
1565  * acpi_idle_enter_bm - enters C3 with proper BM handling
1566  * @dev: the target CPU
1567  * @state: the state data
1568  *
1569  * If BM is detected, the deepest non-C3 idle state is entered instead.
1570  */
1571 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
1572                               struct cpuidle_state *state)
1573 {
1574         struct acpi_processor *pr;
1575         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1576         u32 t1, t2;
1577         int sleep_ticks = 0;
1578
1579         pr = __get_cpu_var(processors);
1580
1581         if (unlikely(!pr))
1582                 return 0;
1583
1584         if (acpi_idle_suspend)
1585                 return(acpi_idle_enter_c1(dev, state));
1586
1587         if (acpi_idle_bm_check()) {
1588                 if (dev->safe_state) {
1589                         dev->last_state = dev->safe_state;
1590                         return dev->safe_state->enter(dev, dev->safe_state);
1591                 } else {
1592                         local_irq_disable();
1593                         acpi_safe_halt();
1594                         local_irq_enable();
1595                         return 0;
1596                 }
1597         }
1598
1599         local_irq_disable();
1600         current_thread_info()->status &= ~TS_POLLING;
1601         /*
1602          * TS_POLLING-cleared state must be visible before we test
1603          * NEED_RESCHED:
1604          */
1605         smp_mb();
1606
1607         if (unlikely(need_resched())) {
1608                 current_thread_info()->status |= TS_POLLING;
1609                 local_irq_enable();
1610                 return 0;
1611         }
1612
1613         acpi_unlazy_tlb(smp_processor_id());
1614
1615         /* Tell the scheduler that we are going deep-idle: */
1616         sched_clock_idle_sleep_event();
1617         /*
1618          * Must be done before busmaster disable as we might need to
1619          * access HPET !
1620          */
1621         acpi_state_timer_broadcast(pr, cx, 1);
1622
1623         acpi_idle_update_bm_rld(pr, cx);
1624
1625         /*
1626          * disable bus master
1627          * bm_check implies we need ARB_DIS
1628          * !bm_check implies we need cache flush
1629          * bm_control implies whether we can do ARB_DIS
1630          *
1631          * That leaves a case where bm_check is set and bm_control is
1632          * not set. In that case we cannot do much, we enter C3
1633          * without doing anything.
1634          */
1635         if (pr->flags.bm_check && pr->flags.bm_control) {
1636                 spin_lock(&c3_lock);
1637                 c3_cpu_count++;
1638                 /* Disable bus master arbitration when all CPUs are in C3 */
1639                 if (c3_cpu_count == num_online_cpus())
1640                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
1641                 spin_unlock(&c3_lock);
1642         } else if (!pr->flags.bm_check) {
1643                 ACPI_FLUSH_CPU_CACHE();
1644         }
1645
1646         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1647         acpi_idle_do_entry(cx);
1648         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1649
1650         /* Re-enable bus master arbitration */
1651         if (pr->flags.bm_check && pr->flags.bm_control) {
1652                 spin_lock(&c3_lock);
1653                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
1654                 c3_cpu_count--;
1655                 spin_unlock(&c3_lock);
1656         }
1657
1658 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
1659         /* TSC could halt in idle, so notify users */
1660         if (tsc_halts_in_c(ACPI_STATE_C3))
1661                 mark_tsc_unstable("TSC halts in idle");
1662 #endif
1663         sleep_ticks = ticks_elapsed(t1, t2);
1664         /* Tell the scheduler how much we idled: */
1665         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1666
1667         local_irq_enable();
1668         current_thread_info()->status |= TS_POLLING;
1669
1670         cx->usage++;
1671
1672         acpi_state_timer_broadcast(pr, cx, 0);
1673         cx->time += sleep_ticks;
1674         return ticks_elapsed_in_us(t1, t2);
1675 }
1676
1677 struct cpuidle_driver acpi_idle_driver = {
1678         .name =         "acpi_idle",
1679         .owner =        THIS_MODULE,
1680 };
1681
1682 /**
1683  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1684  * @pr: the ACPI processor
1685  */
1686 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1687 {
1688         int i, count = CPUIDLE_DRIVER_STATE_START;
1689         struct acpi_processor_cx *cx;
1690         struct cpuidle_state *state;
1691         struct cpuidle_device *dev = &pr->power.dev;
1692
1693         if (!pr->flags.power_setup_done)
1694                 return -EINVAL;
1695
1696         if (pr->flags.power == 0) {
1697                 return -EINVAL;
1698         }
1699
1700         dev->cpu = pr->id;
1701         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1702                 dev->states[i].name[0] = '\0';
1703                 dev->states[i].desc[0] = '\0';
1704         }
1705
1706         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1707                 cx = &pr->power.states[i];
1708                 state = &dev->states[count];
1709
1710                 if (!cx->valid)
1711                         continue;
1712
1713 #ifdef CONFIG_HOTPLUG_CPU
1714                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1715                     !pr->flags.has_cst &&
1716                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1717                         continue;
1718 #endif
1719                 cpuidle_set_statedata(state, cx);
1720
1721                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1722                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1723                 state->exit_latency = cx->latency;
1724                 state->target_residency = cx->latency * latency_factor;
1725                 state->power_usage = cx->power;
1726
1727                 state->flags = 0;
1728                 switch (cx->type) {
1729                         case ACPI_STATE_C1:
1730                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1731                         if (cx->entry_method == ACPI_CSTATE_FFH)
1732                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1733
1734                         state->enter = acpi_idle_enter_c1;
1735                         dev->safe_state = state;
1736                         break;
1737
1738                         case ACPI_STATE_C2:
1739                         state->flags |= CPUIDLE_FLAG_BALANCED;
1740                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1741                         state->enter = acpi_idle_enter_simple;
1742                         dev->safe_state = state;
1743                         break;
1744
1745                         case ACPI_STATE_C3:
1746                         state->flags |= CPUIDLE_FLAG_DEEP;
1747                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1748                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1749                         state->enter = pr->flags.bm_check ?
1750                                         acpi_idle_enter_bm :
1751                                         acpi_idle_enter_simple;
1752                         break;
1753                 }
1754
1755                 count++;
1756                 if (count == CPUIDLE_STATE_MAX)
1757                         break;
1758         }
1759
1760         dev->state_count = count;
1761
1762         if (!count)
1763                 return -EINVAL;
1764
1765         return 0;
1766 }
1767
1768 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1769 {
1770         int ret = 0;
1771
1772         if (boot_option_idle_override)
1773                 return 0;
1774
1775         if (!pr)
1776                 return -EINVAL;
1777
1778         if (nocst) {
1779                 return -ENODEV;
1780         }
1781
1782         if (!pr->flags.power_setup_done)
1783                 return -ENODEV;
1784
1785         cpuidle_pause_and_lock();
1786         cpuidle_disable_device(&pr->power.dev);
1787         acpi_processor_get_power_info(pr);
1788         if (pr->flags.power) {
1789                 acpi_processor_setup_cpuidle(pr);
1790                 ret = cpuidle_enable_device(&pr->power.dev);
1791         }
1792         cpuidle_resume_and_unlock();
1793
1794         return ret;
1795 }
1796
1797 #endif /* CONFIG_CPU_IDLE */
1798
1799 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1800                               struct acpi_device *device)
1801 {
1802         acpi_status status = 0;
1803         static int first_run;
1804         struct proc_dir_entry *entry = NULL;
1805         unsigned int i;
1806
1807         if (boot_option_idle_override)
1808                 return 0;
1809
1810         if (!first_run) {
1811                 if (idle_halt) {
1812                         /*
1813                          * When the boot option of "idle=halt" is added, halt
1814                          * is used for CPU IDLE.
1815                          * In such case C2/C3 is meaningless. So the max_cstate
1816                          * is set to one.
1817                          */
1818                         max_cstate = 1;
1819                 }
1820                 dmi_check_system(processor_power_dmi_table);
1821                 max_cstate = acpi_processor_cstate_check(max_cstate);
1822                 if (max_cstate < ACPI_C_STATES_MAX)
1823                         printk(KERN_NOTICE
1824                                "ACPI: processor limited to max C-state %d\n",
1825                                max_cstate);
1826                 first_run++;
1827 #if !defined(CONFIG_CPU_IDLE) && defined(CONFIG_SMP)
1828                 pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY,
1829                                 &acpi_processor_latency_notifier);
1830 #endif
1831         }
1832
1833         if (!pr)
1834                 return -EINVAL;
1835
1836         if (acpi_gbl_FADT.cst_control && !nocst) {
1837                 status =
1838                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1839                 if (ACPI_FAILURE(status)) {
1840                         ACPI_EXCEPTION((AE_INFO, status,
1841                                         "Notifying BIOS of _CST ability failed"));
1842                 }
1843         }
1844
1845         acpi_processor_get_power_info(pr);
1846         pr->flags.power_setup_done = 1;
1847
1848         /*
1849          * Install the idle handler if processor power management is supported.
1850          * Note that we use previously set idle handler will be used on
1851          * platforms that only support C1.
1852          */
1853         if (pr->flags.power) {
1854 #ifdef CONFIG_CPU_IDLE
1855                 acpi_processor_setup_cpuidle(pr);
1856                 if (cpuidle_register_device(&pr->power.dev))
1857                         return -EIO;
1858 #endif
1859
1860                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1861                 for (i = 1; i <= pr->power.count; i++)
1862                         if (pr->power.states[i].valid)
1863                                 printk(" C%d[C%d]", i,
1864                                        pr->power.states[i].type);
1865                 printk(")\n");
1866
1867 #ifndef CONFIG_CPU_IDLE
1868                 if (pr->id == 0) {
1869                         pm_idle_save = pm_idle;
1870                         pm_idle = acpi_processor_idle;
1871                 }
1872 #endif
1873         }
1874
1875         /* 'power' [R] */
1876         entry = proc_create_data(ACPI_PROCESSOR_FILE_POWER,
1877                                  S_IRUGO, acpi_device_dir(device),
1878                                  &acpi_processor_power_fops,
1879                                  acpi_driver_data(device));
1880         if (!entry)
1881                 return -EIO;
1882         return 0;
1883 }
1884
1885 int acpi_processor_power_exit(struct acpi_processor *pr,
1886                               struct acpi_device *device)
1887 {
1888         if (boot_option_idle_override)
1889                 return 0;
1890
1891 #ifdef CONFIG_CPU_IDLE
1892         cpuidle_unregister_device(&pr->power.dev);
1893 #endif
1894         pr->flags.power_setup_done = 0;
1895
1896         if (acpi_device_dir(device))
1897                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1898                                   acpi_device_dir(device));
1899
1900 #ifndef CONFIG_CPU_IDLE
1901
1902         /* Unregister the idle handler when processor #0 is removed. */
1903         if (pr->id == 0) {
1904                 if (pm_idle_save)
1905                         pm_idle = pm_idle_save;
1906
1907                 /*
1908                  * We are about to unload the current idle thread pm callback
1909                  * (pm_idle), Wait for all processors to update cached/local
1910                  * copies of pm_idle before proceeding.
1911                  */
1912                 cpu_idle_wait();
1913 #ifdef CONFIG_SMP
1914                 pm_qos_remove_notifier(PM_QOS_CPU_DMA_LATENCY,
1915                                 &acpi_processor_latency_notifier);
1916 #endif
1917         }
1918 #endif
1919
1920         return 0;
1921 }