acpi: use non-racy method for proc entries creation
[linux-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/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
61 #define ACPI_PROCESSOR_COMPONENT        0x01000000
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         if (cstate->entry_method == ACPI_CSTATE_FFH) {
274                 /* Call into architectural FFH based C-state */
275                 acpi_processor_ffh_cstate_enter(cstate);
276         } else {
277                 int unused;
278                 /* IO port based C-state */
279                 inb(cstate->address);
280                 /* Dummy wait op - must do something useless after P_LVL2 read
281                    because chipsets cannot guarantee that STPCLK# signal
282                    gets asserted in time to freeze execution properly. */
283                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
284         }
285 }
286 #endif /* !CONFIG_CPU_IDLE */
287
288 #ifdef ARCH_APICTIMER_STOPS_ON_C3
289
290 /*
291  * Some BIOS implementations switch to C3 in the published C2 state.
292  * This seems to be a common problem on AMD boxen, but other vendors
293  * are affected too. We pick the most conservative approach: we assume
294  * that the local APIC stops in both C2 and C3.
295  */
296 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
297                                    struct acpi_processor_cx *cx)
298 {
299         struct acpi_processor_power *pwr = &pr->power;
300         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
301
302         /*
303          * Check, if one of the previous states already marked the lapic
304          * unstable
305          */
306         if (pwr->timer_broadcast_on_state < state)
307                 return;
308
309         if (cx->type >= type)
310                 pr->power.timer_broadcast_on_state = state;
311 }
312
313 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr)
314 {
315         unsigned long reason;
316
317         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
318                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
319
320         clockevents_notify(reason, &pr->id);
321 }
322
323 /* Power(C) State timer broadcast control */
324 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
325                                        struct acpi_processor_cx *cx,
326                                        int broadcast)
327 {
328         int state = cx - pr->power.states;
329
330         if (state >= pr->power.timer_broadcast_on_state) {
331                 unsigned long reason;
332
333                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
334                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
335                 clockevents_notify(reason, &pr->id);
336         }
337 }
338
339 #else
340
341 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
342                                    struct acpi_processor_cx *cstate) { }
343 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr) { }
344 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
345                                        struct acpi_processor_cx *cx,
346                                        int broadcast)
347 {
348 }
349
350 #endif
351
352 /*
353  * Suspend / resume control
354  */
355 static int acpi_idle_suspend;
356
357 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
358 {
359         acpi_idle_suspend = 1;
360         return 0;
361 }
362
363 int acpi_processor_resume(struct acpi_device * device)
364 {
365         acpi_idle_suspend = 0;
366         return 0;
367 }
368
369 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
370 static int tsc_halts_in_c(int state)
371 {
372         switch (boot_cpu_data.x86_vendor) {
373         case X86_VENDOR_AMD:
374                 /*
375                  * AMD Fam10h TSC will tick in all
376                  * C/P/S0/S1 states when this bit is set.
377                  */
378                 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
379                         return 0;
380                 /*FALL THROUGH*/
381         case X86_VENDOR_INTEL:
382                 /* Several cases known where TSC halts in C2 too */
383         default:
384                 return state > ACPI_STATE_C1;
385         }
386 }
387 #endif
388
389 #ifndef CONFIG_CPU_IDLE
390 static void acpi_processor_idle(void)
391 {
392         struct acpi_processor *pr = NULL;
393         struct acpi_processor_cx *cx = NULL;
394         struct acpi_processor_cx *next_state = NULL;
395         int sleep_ticks = 0;
396         u32 t1, t2 = 0;
397
398         /*
399          * Interrupts must be disabled during bus mastering calculations and
400          * for C2/C3 transitions.
401          */
402         local_irq_disable();
403
404         pr = processors[smp_processor_id()];
405         if (!pr) {
406                 local_irq_enable();
407                 return;
408         }
409
410         /*
411          * Check whether we truly need to go idle, or should
412          * reschedule:
413          */
414         if (unlikely(need_resched())) {
415                 local_irq_enable();
416                 return;
417         }
418
419         cx = pr->power.state;
420         if (!cx || acpi_idle_suspend) {
421                 if (pm_idle_save) {
422                         pm_idle_save(); /* enables IRQs */
423                 } else {
424                         acpi_safe_halt();
425                         local_irq_enable();
426                 }
427
428                 return;
429         }
430
431         /*
432          * Check BM Activity
433          * -----------------
434          * Check for bus mastering activity (if required), record, and check
435          * for demotion.
436          */
437         if (pr->flags.bm_check) {
438                 u32 bm_status = 0;
439                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
440
441                 if (diff > 31)
442                         diff = 31;
443
444                 pr->power.bm_activity <<= diff;
445
446                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
447                 if (bm_status) {
448                         pr->power.bm_activity |= 0x1;
449                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
450                 }
451                 /*
452                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
453                  * the true state of bus mastering activity; forcing us to
454                  * manually check the BMIDEA bit of each IDE channel.
455                  */
456                 else if (errata.piix4.bmisx) {
457                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
458                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
459                                 pr->power.bm_activity |= 0x1;
460                 }
461
462                 pr->power.bm_check_timestamp = jiffies;
463
464                 /*
465                  * If bus mastering is or was active this jiffy, demote
466                  * to avoid a faulty transition.  Note that the processor
467                  * won't enter a low-power state during this call (to this
468                  * function) but should upon the next.
469                  *
470                  * TBD: A better policy might be to fallback to the demotion
471                  *      state (use it for this quantum only) istead of
472                  *      demoting -- and rely on duration as our sole demotion
473                  *      qualification.  This may, however, introduce DMA
474                  *      issues (e.g. floppy DMA transfer overrun/underrun).
475                  */
476                 if ((pr->power.bm_activity & 0x1) &&
477                     cx->demotion.threshold.bm) {
478                         local_irq_enable();
479                         next_state = cx->demotion.state;
480                         goto end;
481                 }
482         }
483
484 #ifdef CONFIG_HOTPLUG_CPU
485         /*
486          * Check for P_LVL2_UP flag before entering C2 and above on
487          * an SMP system. We do it here instead of doing it at _CST/P_LVL
488          * detection phase, to work cleanly with logical CPU hotplug.
489          */
490         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
491             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
492                 cx = &pr->power.states[ACPI_STATE_C1];
493 #endif
494
495         /*
496          * Sleep:
497          * ------
498          * Invoke the current Cx state to put the processor to sleep.
499          */
500         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
501                 current_thread_info()->status &= ~TS_POLLING;
502                 /*
503                  * TS_POLLING-cleared state must be visible before we
504                  * test NEED_RESCHED:
505                  */
506                 smp_mb();
507                 if (need_resched()) {
508                         current_thread_info()->status |= TS_POLLING;
509                         local_irq_enable();
510                         return;
511                 }
512         }
513
514         switch (cx->type) {
515
516         case ACPI_STATE_C1:
517                 /*
518                  * Invoke C1.
519                  * Use the appropriate idle routine, the one that would
520                  * be used without acpi C-states.
521                  */
522                 if (pm_idle_save) {
523                         pm_idle_save(); /* enables IRQs */
524                 } else {
525                         acpi_safe_halt();
526                         local_irq_enable();
527                 }
528
529                 /*
530                  * TBD: Can't get time duration while in C1, as resumes
531                  *      go to an ISR rather than here.  Need to instrument
532                  *      base interrupt handler.
533                  *
534                  * Note: the TSC better not stop in C1, sched_clock() will
535                  *       skew otherwise.
536                  */
537                 sleep_ticks = 0xFFFFFFFF;
538
539                 break;
540
541         case ACPI_STATE_C2:
542                 /* Get start time (ticks) */
543                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
544                 /* Tell the scheduler that we are going deep-idle: */
545                 sched_clock_idle_sleep_event();
546                 /* Invoke C2 */
547                 acpi_state_timer_broadcast(pr, cx, 1);
548                 acpi_cstate_enter(cx);
549                 /* Get end time (ticks) */
550                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
551
552 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
553                 /* TSC halts in C2, so notify users */
554                 if (tsc_halts_in_c(ACPI_STATE_C2))
555                         mark_tsc_unstable("possible TSC halt in C2");
556 #endif
557                 /* Compute time (ticks) that we were actually asleep */
558                 sleep_ticks = ticks_elapsed(t1, t2);
559
560                 /* Tell the scheduler how much we idled: */
561                 sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
562
563                 /* Re-enable interrupts */
564                 local_irq_enable();
565                 /* Do not account our idle-switching overhead: */
566                 sleep_ticks -= cx->latency_ticks + C2_OVERHEAD;
567
568                 current_thread_info()->status |= TS_POLLING;
569                 acpi_state_timer_broadcast(pr, cx, 0);
570                 break;
571
572         case ACPI_STATE_C3:
573                 acpi_unlazy_tlb(smp_processor_id());
574                 /*
575                  * Must be done before busmaster disable as we might
576                  * need to access HPET !
577                  */
578                 acpi_state_timer_broadcast(pr, cx, 1);
579                 /*
580                  * disable bus master
581                  * bm_check implies we need ARB_DIS
582                  * !bm_check implies we need cache flush
583                  * bm_control implies whether we can do ARB_DIS
584                  *
585                  * That leaves a case where bm_check is set and bm_control is
586                  * not set. In that case we cannot do much, we enter C3
587                  * without doing anything.
588                  */
589                 if (pr->flags.bm_check && pr->flags.bm_control) {
590                         if (atomic_inc_return(&c3_cpu_count) ==
591                             num_online_cpus()) {
592                                 /*
593                                  * All CPUs are trying to go to C3
594                                  * Disable bus master arbitration
595                                  */
596                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
597                         }
598                 } else if (!pr->flags.bm_check) {
599                         /* SMP with no shared cache... Invalidate cache  */
600                         ACPI_FLUSH_CPU_CACHE();
601                 }
602
603                 /* Get start time (ticks) */
604                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
605                 /* Invoke C3 */
606                 /* Tell the scheduler that we are going deep-idle: */
607                 sched_clock_idle_sleep_event();
608                 acpi_cstate_enter(cx);
609                 /* Get end time (ticks) */
610                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
611                 if (pr->flags.bm_check && pr->flags.bm_control) {
612                         /* Enable bus master arbitration */
613                         atomic_dec(&c3_cpu_count);
614                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
615                 }
616
617 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
618                 /* TSC halts in C3, so notify users */
619                 if (tsc_halts_in_c(ACPI_STATE_C3))
620                         mark_tsc_unstable("TSC halts in C3");
621 #endif
622                 /* Compute time (ticks) that we were actually asleep */
623                 sleep_ticks = ticks_elapsed(t1, t2);
624                 /* Tell the scheduler how much we idled: */
625                 sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
626
627                 /* Re-enable interrupts */
628                 local_irq_enable();
629                 /* Do not account our idle-switching overhead: */
630                 sleep_ticks -= cx->latency_ticks + C3_OVERHEAD;
631
632                 current_thread_info()->status |= TS_POLLING;
633                 acpi_state_timer_broadcast(pr, cx, 0);
634                 break;
635
636         default:
637                 local_irq_enable();
638                 return;
639         }
640         cx->usage++;
641         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
642                 cx->time += sleep_ticks;
643
644         next_state = pr->power.state;
645
646 #ifdef CONFIG_HOTPLUG_CPU
647         /* Don't do promotion/demotion */
648         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
649             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
650                 next_state = cx;
651                 goto end;
652         }
653 #endif
654
655         /*
656          * Promotion?
657          * ----------
658          * Track the number of longs (time asleep is greater than threshold)
659          * and promote when the count threshold is reached.  Note that bus
660          * mastering activity may prevent promotions.
661          * Do not promote above max_cstate.
662          */
663         if (cx->promotion.state &&
664             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
665                 if (sleep_ticks > cx->promotion.threshold.ticks &&
666                   cx->promotion.state->latency <=
667                                 pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY)) {
668                         cx->promotion.count++;
669                         cx->demotion.count = 0;
670                         if (cx->promotion.count >=
671                             cx->promotion.threshold.count) {
672                                 if (pr->flags.bm_check) {
673                                         if (!
674                                             (pr->power.bm_activity & cx->
675                                              promotion.threshold.bm)) {
676                                                 next_state =
677                                                     cx->promotion.state;
678                                                 goto end;
679                                         }
680                                 } else {
681                                         next_state = cx->promotion.state;
682                                         goto end;
683                                 }
684                         }
685                 }
686         }
687
688         /*
689          * Demotion?
690          * ---------
691          * Track the number of shorts (time asleep is less than time threshold)
692          * and demote when the usage threshold is reached.
693          */
694         if (cx->demotion.state) {
695                 if (sleep_ticks < cx->demotion.threshold.ticks) {
696                         cx->demotion.count++;
697                         cx->promotion.count = 0;
698                         if (cx->demotion.count >= cx->demotion.threshold.count) {
699                                 next_state = cx->demotion.state;
700                                 goto end;
701                         }
702                 }
703         }
704
705       end:
706         /*
707          * Demote if current state exceeds max_cstate
708          * or if the latency of the current state is unacceptable
709          */
710         if ((pr->power.state - pr->power.states) > max_cstate ||
711                 pr->power.state->latency >
712                                 pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY)) {
713                 if (cx->demotion.state)
714                         next_state = cx->demotion.state;
715         }
716
717         /*
718          * New Cx State?
719          * -------------
720          * If we're going to start using a new Cx state we must clean up
721          * from the previous and prepare to use the new.
722          */
723         if (next_state != pr->power.state)
724                 acpi_processor_power_activate(pr, next_state);
725 }
726
727 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
728 {
729         unsigned int i;
730         unsigned int state_is_set = 0;
731         struct acpi_processor_cx *lower = NULL;
732         struct acpi_processor_cx *higher = NULL;
733         struct acpi_processor_cx *cx;
734
735
736         if (!pr)
737                 return -EINVAL;
738
739         /*
740          * This function sets the default Cx state policy (OS idle handler).
741          * Our scheme is to promote quickly to C2 but more conservatively
742          * to C3.  We're favoring C2  for its characteristics of low latency
743          * (quick response), good power savings, and ability to allow bus
744          * mastering activity.  Note that the Cx state policy is completely
745          * customizable and can be altered dynamically.
746          */
747
748         /* startup state */
749         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
750                 cx = &pr->power.states[i];
751                 if (!cx->valid)
752                         continue;
753
754                 if (!state_is_set)
755                         pr->power.state = cx;
756                 state_is_set++;
757                 break;
758         }
759
760         if (!state_is_set)
761                 return -ENODEV;
762
763         /* demotion */
764         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
765                 cx = &pr->power.states[i];
766                 if (!cx->valid)
767                         continue;
768
769                 if (lower) {
770                         cx->demotion.state = lower;
771                         cx->demotion.threshold.ticks = cx->latency_ticks;
772                         cx->demotion.threshold.count = 1;
773                         if (cx->type == ACPI_STATE_C3)
774                                 cx->demotion.threshold.bm = bm_history;
775                 }
776
777                 lower = cx;
778         }
779
780         /* promotion */
781         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
782                 cx = &pr->power.states[i];
783                 if (!cx->valid)
784                         continue;
785
786                 if (higher) {
787                         cx->promotion.state = higher;
788                         cx->promotion.threshold.ticks = cx->latency_ticks;
789                         if (cx->type >= ACPI_STATE_C2)
790                                 cx->promotion.threshold.count = 4;
791                         else
792                                 cx->promotion.threshold.count = 10;
793                         if (higher->type == ACPI_STATE_C3)
794                                 cx->promotion.threshold.bm = bm_history;
795                 }
796
797                 higher = cx;
798         }
799
800         return 0;
801 }
802 #endif /* !CONFIG_CPU_IDLE */
803
804 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
805 {
806
807         if (!pr)
808                 return -EINVAL;
809
810         if (!pr->pblk)
811                 return -ENODEV;
812
813         /* if info is obtained from pblk/fadt, type equals state */
814         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
815         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
816
817 #ifndef CONFIG_HOTPLUG_CPU
818         /*
819          * Check for P_LVL2_UP flag before entering C2 and above on
820          * an SMP system.
821          */
822         if ((num_online_cpus() > 1) &&
823             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
824                 return -ENODEV;
825 #endif
826
827         /* determine C2 and C3 address from pblk */
828         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
829         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
830
831         /* determine latencies from FADT */
832         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
833         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
834
835         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
836                           "lvl2[0x%08x] lvl3[0x%08x]\n",
837                           pr->power.states[ACPI_STATE_C2].address,
838                           pr->power.states[ACPI_STATE_C3].address));
839
840         return 0;
841 }
842
843 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
844 {
845         if (!pr->power.states[ACPI_STATE_C1].valid) {
846                 /* set the first C-State to C1 */
847                 /* all processors need to support C1 */
848                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
849                 pr->power.states[ACPI_STATE_C1].valid = 1;
850         }
851         /* the C0 state only exists as a filler in our array */
852         pr->power.states[ACPI_STATE_C0].valid = 1;
853         return 0;
854 }
855
856 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
857 {
858         acpi_status status = 0;
859         acpi_integer count;
860         int current_count;
861         int i;
862         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
863         union acpi_object *cst;
864
865
866         if (nocst)
867                 return -ENODEV;
868
869         current_count = 0;
870
871         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
872         if (ACPI_FAILURE(status)) {
873                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
874                 return -ENODEV;
875         }
876
877         cst = buffer.pointer;
878
879         /* There must be at least 2 elements */
880         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
881                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
882                 status = -EFAULT;
883                 goto end;
884         }
885
886         count = cst->package.elements[0].integer.value;
887
888         /* Validate number of power states. */
889         if (count < 1 || count != cst->package.count - 1) {
890                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
891                 status = -EFAULT;
892                 goto end;
893         }
894
895         /* Tell driver that at least _CST is supported. */
896         pr->flags.has_cst = 1;
897
898         for (i = 1; i <= count; i++) {
899                 union acpi_object *element;
900                 union acpi_object *obj;
901                 struct acpi_power_register *reg;
902                 struct acpi_processor_cx cx;
903
904                 memset(&cx, 0, sizeof(cx));
905
906                 element = &(cst->package.elements[i]);
907                 if (element->type != ACPI_TYPE_PACKAGE)
908                         continue;
909
910                 if (element->package.count != 4)
911                         continue;
912
913                 obj = &(element->package.elements[0]);
914
915                 if (obj->type != ACPI_TYPE_BUFFER)
916                         continue;
917
918                 reg = (struct acpi_power_register *)obj->buffer.pointer;
919
920                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
921                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
922                         continue;
923
924                 /* There should be an easy way to extract an integer... */
925                 obj = &(element->package.elements[1]);
926                 if (obj->type != ACPI_TYPE_INTEGER)
927                         continue;
928
929                 cx.type = obj->integer.value;
930                 /*
931                  * Some buggy BIOSes won't list C1 in _CST -
932                  * Let acpi_processor_get_power_info_default() handle them later
933                  */
934                 if (i == 1 && cx.type != ACPI_STATE_C1)
935                         current_count++;
936
937                 cx.address = reg->address;
938                 cx.index = current_count + 1;
939
940                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
941                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
942                         if (acpi_processor_ffh_cstate_probe
943                                         (pr->id, &cx, reg) == 0) {
944                                 cx.entry_method = ACPI_CSTATE_FFH;
945                         } else if (cx.type == ACPI_STATE_C1) {
946                                 /*
947                                  * C1 is a special case where FIXED_HARDWARE
948                                  * can be handled in non-MWAIT way as well.
949                                  * In that case, save this _CST entry info.
950                                  * Otherwise, ignore this info and continue.
951                                  */
952                                 cx.entry_method = ACPI_CSTATE_HALT;
953                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
954                         } else {
955                                 continue;
956                         }
957                 } else {
958                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
959                                  cx.address);
960                 }
961
962
963                 obj = &(element->package.elements[2]);
964                 if (obj->type != ACPI_TYPE_INTEGER)
965                         continue;
966
967                 cx.latency = obj->integer.value;
968
969                 obj = &(element->package.elements[3]);
970                 if (obj->type != ACPI_TYPE_INTEGER)
971                         continue;
972
973                 cx.power = obj->integer.value;
974
975                 current_count++;
976                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
977
978                 /*
979                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
980                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
981                  */
982                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
983                         printk(KERN_WARNING
984                                "Limiting number of power states to max (%d)\n",
985                                ACPI_PROCESSOR_MAX_POWER);
986                         printk(KERN_WARNING
987                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
988                         break;
989                 }
990         }
991
992         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
993                           current_count));
994
995         /* Validate number of power states discovered */
996         if (current_count < 2)
997                 status = -EFAULT;
998
999       end:
1000         kfree(buffer.pointer);
1001
1002         return status;
1003 }
1004
1005 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
1006 {
1007
1008         if (!cx->address)
1009                 return;
1010
1011         /*
1012          * C2 latency must be less than or equal to 100
1013          * microseconds.
1014          */
1015         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
1016                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1017                                   "latency too large [%d]\n", cx->latency));
1018                 return;
1019         }
1020
1021         /*
1022          * Otherwise we've met all of our C2 requirements.
1023          * Normalize the C2 latency to expidite policy
1024          */
1025         cx->valid = 1;
1026
1027 #ifndef CONFIG_CPU_IDLE
1028         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1029 #else
1030         cx->latency_ticks = cx->latency;
1031 #endif
1032
1033         return;
1034 }
1035
1036 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
1037                                            struct acpi_processor_cx *cx)
1038 {
1039         static int bm_check_flag;
1040
1041
1042         if (!cx->address)
1043                 return;
1044
1045         /*
1046          * C3 latency must be less than or equal to 1000
1047          * microseconds.
1048          */
1049         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
1050                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1051                                   "latency too large [%d]\n", cx->latency));
1052                 return;
1053         }
1054
1055         /*
1056          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
1057          * DMA transfers are used by any ISA device to avoid livelock.
1058          * Note that we could disable Type-F DMA (as recommended by
1059          * the erratum), but this is known to disrupt certain ISA
1060          * devices thus we take the conservative approach.
1061          */
1062         else if (errata.piix4.fdma) {
1063                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1064                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
1065                 return;
1066         }
1067
1068         /* All the logic here assumes flags.bm_check is same across all CPUs */
1069         if (!bm_check_flag) {
1070                 /* Determine whether bm_check is needed based on CPU  */
1071                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
1072                 bm_check_flag = pr->flags.bm_check;
1073         } else {
1074                 pr->flags.bm_check = bm_check_flag;
1075         }
1076
1077         if (pr->flags.bm_check) {
1078                 if (!pr->flags.bm_control) {
1079                         if (pr->flags.has_cst != 1) {
1080                                 /* bus mastering control is necessary */
1081                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1082                                         "C3 support requires BM control\n"));
1083                                 return;
1084                         } else {
1085                                 /* Here we enter C3 without bus mastering */
1086                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1087                                         "C3 support without BM control\n"));
1088                         }
1089                 }
1090         } else {
1091                 /*
1092                  * WBINVD should be set in fadt, for C3 state to be
1093                  * supported on when bm_check is not required.
1094                  */
1095                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
1096                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1097                                           "Cache invalidation should work properly"
1098                                           " for C3 to be enabled on SMP systems\n"));
1099                         return;
1100                 }
1101                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1102         }
1103
1104         /*
1105          * Otherwise we've met all of our C3 requirements.
1106          * Normalize the C3 latency to expidite policy.  Enable
1107          * checking of bus mastering status (bm_check) so we can
1108          * use this in our C3 policy
1109          */
1110         cx->valid = 1;
1111
1112 #ifndef CONFIG_CPU_IDLE
1113         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1114 #else
1115         cx->latency_ticks = cx->latency;
1116 #endif
1117
1118         return;
1119 }
1120
1121 static int acpi_processor_power_verify(struct acpi_processor *pr)
1122 {
1123         unsigned int i;
1124         unsigned int working = 0;
1125
1126         pr->power.timer_broadcast_on_state = INT_MAX;
1127
1128         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1129                 struct acpi_processor_cx *cx = &pr->power.states[i];
1130
1131                 switch (cx->type) {
1132                 case ACPI_STATE_C1:
1133                         cx->valid = 1;
1134                         break;
1135
1136                 case ACPI_STATE_C2:
1137                         acpi_processor_power_verify_c2(cx);
1138                         if (cx->valid)
1139                                 acpi_timer_check_state(i, pr, cx);
1140                         break;
1141
1142                 case ACPI_STATE_C3:
1143                         acpi_processor_power_verify_c3(pr, cx);
1144                         if (cx->valid)
1145                                 acpi_timer_check_state(i, pr, cx);
1146                         break;
1147                 }
1148
1149                 if (cx->valid)
1150                         working++;
1151         }
1152
1153         acpi_propagate_timer_broadcast(pr);
1154
1155         return (working);
1156 }
1157
1158 static int acpi_processor_get_power_info(struct acpi_processor *pr)
1159 {
1160         unsigned int i;
1161         int result;
1162
1163
1164         /* NOTE: the idle thread may not be running while calling
1165          * this function */
1166
1167         /* Zero initialize all the C-states info. */
1168         memset(pr->power.states, 0, sizeof(pr->power.states));
1169
1170         result = acpi_processor_get_power_info_cst(pr);
1171         if (result == -ENODEV)
1172                 result = acpi_processor_get_power_info_fadt(pr);
1173
1174         if (result)
1175                 return result;
1176
1177         acpi_processor_get_power_info_default(pr);
1178
1179         pr->power.count = acpi_processor_power_verify(pr);
1180
1181 #ifndef CONFIG_CPU_IDLE
1182         /*
1183          * Set Default Policy
1184          * ------------------
1185          * Now that we know which states are supported, set the default
1186          * policy.  Note that this policy can be changed dynamically
1187          * (e.g. encourage deeper sleeps to conserve battery life when
1188          * not on AC).
1189          */
1190         result = acpi_processor_set_power_policy(pr);
1191         if (result)
1192                 return result;
1193 #endif
1194
1195         /*
1196          * if one state of type C2 or C3 is available, mark this
1197          * CPU as being "idle manageable"
1198          */
1199         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1200                 if (pr->power.states[i].valid) {
1201                         pr->power.count = i;
1202                         if (pr->power.states[i].type >= ACPI_STATE_C2)
1203                                 pr->flags.power = 1;
1204                 }
1205         }
1206
1207         return 0;
1208 }
1209
1210 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1211 {
1212         struct acpi_processor *pr = seq->private;
1213         unsigned int i;
1214
1215
1216         if (!pr)
1217                 goto end;
1218
1219         seq_printf(seq, "active state:            C%zd\n"
1220                    "max_cstate:              C%d\n"
1221                    "bus master activity:     %08x\n"
1222                    "maximum allowed latency: %d usec\n",
1223                    pr->power.state ? pr->power.state - pr->power.states : 0,
1224                    max_cstate, (unsigned)pr->power.bm_activity,
1225                    pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY));
1226
1227         seq_puts(seq, "states:\n");
1228
1229         for (i = 1; i <= pr->power.count; i++) {
1230                 seq_printf(seq, "   %cC%d:                  ",
1231                            (&pr->power.states[i] ==
1232                             pr->power.state ? '*' : ' '), i);
1233
1234                 if (!pr->power.states[i].valid) {
1235                         seq_puts(seq, "<not supported>\n");
1236                         continue;
1237                 }
1238
1239                 switch (pr->power.states[i].type) {
1240                 case ACPI_STATE_C1:
1241                         seq_printf(seq, "type[C1] ");
1242                         break;
1243                 case ACPI_STATE_C2:
1244                         seq_printf(seq, "type[C2] ");
1245                         break;
1246                 case ACPI_STATE_C3:
1247                         seq_printf(seq, "type[C3] ");
1248                         break;
1249                 default:
1250                         seq_printf(seq, "type[--] ");
1251                         break;
1252                 }
1253
1254                 if (pr->power.states[i].promotion.state)
1255                         seq_printf(seq, "promotion[C%zd] ",
1256                                    (pr->power.states[i].promotion.state -
1257                                     pr->power.states));
1258                 else
1259                         seq_puts(seq, "promotion[--] ");
1260
1261                 if (pr->power.states[i].demotion.state)
1262                         seq_printf(seq, "demotion[C%zd] ",
1263                                    (pr->power.states[i].demotion.state -
1264                                     pr->power.states));
1265                 else
1266                         seq_puts(seq, "demotion[--] ");
1267
1268                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1269                            pr->power.states[i].latency,
1270                            pr->power.states[i].usage,
1271                            (unsigned long long)pr->power.states[i].time);
1272         }
1273
1274       end:
1275         return 0;
1276 }
1277
1278 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1279 {
1280         return single_open(file, acpi_processor_power_seq_show,
1281                            PDE(inode)->data);
1282 }
1283
1284 static const struct file_operations acpi_processor_power_fops = {
1285         .owner = THIS_MODULE,
1286         .open = acpi_processor_power_open_fs,
1287         .read = seq_read,
1288         .llseek = seq_lseek,
1289         .release = single_release,
1290 };
1291
1292 #ifndef CONFIG_CPU_IDLE
1293
1294 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1295 {
1296         int result = 0;
1297
1298
1299         if (!pr)
1300                 return -EINVAL;
1301
1302         if (nocst) {
1303                 return -ENODEV;
1304         }
1305
1306         if (!pr->flags.power_setup_done)
1307                 return -ENODEV;
1308
1309         /* Fall back to the default idle loop */
1310         pm_idle = pm_idle_save;
1311         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
1312
1313         pr->flags.power = 0;
1314         result = acpi_processor_get_power_info(pr);
1315         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1316                 pm_idle = acpi_processor_idle;
1317
1318         return result;
1319 }
1320
1321 #ifdef CONFIG_SMP
1322 static void smp_callback(void *v)
1323 {
1324         /* we already woke the CPU up, nothing more to do */
1325 }
1326
1327 /*
1328  * This function gets called when a part of the kernel has a new latency
1329  * requirement.  This means we need to get all processors out of their C-state,
1330  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1331  * wakes them all right up.
1332  */
1333 static int acpi_processor_latency_notify(struct notifier_block *b,
1334                 unsigned long l, void *v)
1335 {
1336         smp_call_function(smp_callback, NULL, 0, 1);
1337         return NOTIFY_OK;
1338 }
1339
1340 static struct notifier_block acpi_processor_latency_notifier = {
1341         .notifier_call = acpi_processor_latency_notify,
1342 };
1343
1344 #endif
1345
1346 #else /* CONFIG_CPU_IDLE */
1347
1348 /**
1349  * acpi_idle_bm_check - checks if bus master activity was detected
1350  */
1351 static int acpi_idle_bm_check(void)
1352 {
1353         u32 bm_status = 0;
1354
1355         acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
1356         if (bm_status)
1357                 acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
1358         /*
1359          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
1360          * the true state of bus mastering activity; forcing us to
1361          * manually check the BMIDEA bit of each IDE channel.
1362          */
1363         else if (errata.piix4.bmisx) {
1364                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
1365                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
1366                         bm_status = 1;
1367         }
1368         return bm_status;
1369 }
1370
1371 /**
1372  * acpi_idle_update_bm_rld - updates the BM_RLD bit depending on target state
1373  * @pr: the processor
1374  * @target: the new target state
1375  */
1376 static inline void acpi_idle_update_bm_rld(struct acpi_processor *pr,
1377                                            struct acpi_processor_cx *target)
1378 {
1379         if (pr->flags.bm_rld_set && target->type != ACPI_STATE_C3) {
1380                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1381                 pr->flags.bm_rld_set = 0;
1382         }
1383
1384         if (!pr->flags.bm_rld_set && target->type == ACPI_STATE_C3) {
1385                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
1386                 pr->flags.bm_rld_set = 1;
1387         }
1388 }
1389
1390 /**
1391  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
1392  * @cx: cstate data
1393  *
1394  * Caller disables interrupt before call and enables interrupt after return.
1395  */
1396 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
1397 {
1398         if (cx->entry_method == ACPI_CSTATE_FFH) {
1399                 /* Call into architectural FFH based C-state */
1400                 acpi_processor_ffh_cstate_enter(cx);
1401         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
1402                 acpi_safe_halt();
1403         } else {
1404                 int unused;
1405                 /* IO port based C-state */
1406                 inb(cx->address);
1407                 /* Dummy wait op - must do something useless after P_LVL2 read
1408                    because chipsets cannot guarantee that STPCLK# signal
1409                    gets asserted in time to freeze execution properly. */
1410                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
1411         }
1412 }
1413
1414 /**
1415  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
1416  * @dev: the target CPU
1417  * @state: the state data
1418  *
1419  * This is equivalent to the HALT instruction.
1420  */
1421 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
1422                               struct cpuidle_state *state)
1423 {
1424         u32 t1, t2;
1425         struct acpi_processor *pr;
1426         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1427
1428         pr = processors[smp_processor_id()];
1429
1430         if (unlikely(!pr))
1431                 return 0;
1432
1433         local_irq_disable();
1434
1435         /* Do not access any ACPI IO ports in suspend path */
1436         if (acpi_idle_suspend) {
1437                 acpi_safe_halt();
1438                 local_irq_enable();
1439                 return 0;
1440         }
1441
1442         if (pr->flags.bm_check)
1443                 acpi_idle_update_bm_rld(pr, cx);
1444
1445         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1446         acpi_idle_do_entry(cx);
1447         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1448
1449         local_irq_enable();
1450         cx->usage++;
1451
1452         return ticks_elapsed_in_us(t1, t2);
1453 }
1454
1455 /**
1456  * acpi_idle_enter_simple - enters an ACPI state without BM handling
1457  * @dev: the target CPU
1458  * @state: the state data
1459  */
1460 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
1461                                   struct cpuidle_state *state)
1462 {
1463         struct acpi_processor *pr;
1464         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1465         u32 t1, t2;
1466         int sleep_ticks = 0;
1467
1468         pr = processors[smp_processor_id()];
1469
1470         if (unlikely(!pr))
1471                 return 0;
1472
1473         if (acpi_idle_suspend)
1474                 return(acpi_idle_enter_c1(dev, state));
1475
1476         local_irq_disable();
1477         current_thread_info()->status &= ~TS_POLLING;
1478         /*
1479          * TS_POLLING-cleared state must be visible before we test
1480          * NEED_RESCHED:
1481          */
1482         smp_mb();
1483
1484         if (unlikely(need_resched())) {
1485                 current_thread_info()->status |= TS_POLLING;
1486                 local_irq_enable();
1487                 return 0;
1488         }
1489
1490         /*
1491          * Must be done before busmaster disable as we might need to
1492          * access HPET !
1493          */
1494         acpi_state_timer_broadcast(pr, cx, 1);
1495
1496         if (pr->flags.bm_check)
1497                 acpi_idle_update_bm_rld(pr, cx);
1498
1499         if (cx->type == ACPI_STATE_C3)
1500                 ACPI_FLUSH_CPU_CACHE();
1501
1502         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1503         /* Tell the scheduler that we are going deep-idle: */
1504         sched_clock_idle_sleep_event();
1505         acpi_idle_do_entry(cx);
1506         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1507
1508 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
1509         /* TSC could halt in idle, so notify users */
1510         if (tsc_halts_in_c(cx->type))
1511                 mark_tsc_unstable("TSC halts in idle");;
1512 #endif
1513         sleep_ticks = ticks_elapsed(t1, t2);
1514
1515         /* Tell the scheduler how much we idled: */
1516         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1517
1518         local_irq_enable();
1519         current_thread_info()->status |= TS_POLLING;
1520
1521         cx->usage++;
1522
1523         acpi_state_timer_broadcast(pr, cx, 0);
1524         cx->time += sleep_ticks;
1525         return ticks_elapsed_in_us(t1, t2);
1526 }
1527
1528 static int c3_cpu_count;
1529 static DEFINE_SPINLOCK(c3_lock);
1530
1531 /**
1532  * acpi_idle_enter_bm - enters C3 with proper BM handling
1533  * @dev: the target CPU
1534  * @state: the state data
1535  *
1536  * If BM is detected, the deepest non-C3 idle state is entered instead.
1537  */
1538 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
1539                               struct cpuidle_state *state)
1540 {
1541         struct acpi_processor *pr;
1542         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1543         u32 t1, t2;
1544         int sleep_ticks = 0;
1545
1546         pr = processors[smp_processor_id()];
1547
1548         if (unlikely(!pr))
1549                 return 0;
1550
1551         if (acpi_idle_suspend)
1552                 return(acpi_idle_enter_c1(dev, state));
1553
1554         if (acpi_idle_bm_check()) {
1555                 if (dev->safe_state) {
1556                         return dev->safe_state->enter(dev, dev->safe_state);
1557                 } else {
1558                         local_irq_disable();
1559                         acpi_safe_halt();
1560                         local_irq_enable();
1561                         return 0;
1562                 }
1563         }
1564
1565         local_irq_disable();
1566         current_thread_info()->status &= ~TS_POLLING;
1567         /*
1568          * TS_POLLING-cleared state must be visible before we test
1569          * NEED_RESCHED:
1570          */
1571         smp_mb();
1572
1573         if (unlikely(need_resched())) {
1574                 current_thread_info()->status |= TS_POLLING;
1575                 local_irq_enable();
1576                 return 0;
1577         }
1578
1579         acpi_unlazy_tlb(smp_processor_id());
1580
1581         /* Tell the scheduler that we are going deep-idle: */
1582         sched_clock_idle_sleep_event();
1583         /*
1584          * Must be done before busmaster disable as we might need to
1585          * access HPET !
1586          */
1587         acpi_state_timer_broadcast(pr, cx, 1);
1588
1589         acpi_idle_update_bm_rld(pr, cx);
1590
1591         /*
1592          * disable bus master
1593          * bm_check implies we need ARB_DIS
1594          * !bm_check implies we need cache flush
1595          * bm_control implies whether we can do ARB_DIS
1596          *
1597          * That leaves a case where bm_check is set and bm_control is
1598          * not set. In that case we cannot do much, we enter C3
1599          * without doing anything.
1600          */
1601         if (pr->flags.bm_check && pr->flags.bm_control) {
1602                 spin_lock(&c3_lock);
1603                 c3_cpu_count++;
1604                 /* Disable bus master arbitration when all CPUs are in C3 */
1605                 if (c3_cpu_count == num_online_cpus())
1606                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
1607                 spin_unlock(&c3_lock);
1608         } else if (!pr->flags.bm_check) {
1609                 ACPI_FLUSH_CPU_CACHE();
1610         }
1611
1612         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1613         acpi_idle_do_entry(cx);
1614         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1615
1616         /* Re-enable bus master arbitration */
1617         if (pr->flags.bm_check && pr->flags.bm_control) {
1618                 spin_lock(&c3_lock);
1619                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
1620                 c3_cpu_count--;
1621                 spin_unlock(&c3_lock);
1622         }
1623
1624 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
1625         /* TSC could halt in idle, so notify users */
1626         if (tsc_halts_in_c(ACPI_STATE_C3))
1627                 mark_tsc_unstable("TSC halts in idle");
1628 #endif
1629         sleep_ticks = ticks_elapsed(t1, t2);
1630         /* Tell the scheduler how much we idled: */
1631         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1632
1633         local_irq_enable();
1634         current_thread_info()->status |= TS_POLLING;
1635
1636         cx->usage++;
1637
1638         acpi_state_timer_broadcast(pr, cx, 0);
1639         cx->time += sleep_ticks;
1640         return ticks_elapsed_in_us(t1, t2);
1641 }
1642
1643 struct cpuidle_driver acpi_idle_driver = {
1644         .name =         "acpi_idle",
1645         .owner =        THIS_MODULE,
1646 };
1647
1648 /**
1649  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1650  * @pr: the ACPI processor
1651  */
1652 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1653 {
1654         int i, count = CPUIDLE_DRIVER_STATE_START;
1655         struct acpi_processor_cx *cx;
1656         struct cpuidle_state *state;
1657         struct cpuidle_device *dev = &pr->power.dev;
1658
1659         if (!pr->flags.power_setup_done)
1660                 return -EINVAL;
1661
1662         if (pr->flags.power == 0) {
1663                 return -EINVAL;
1664         }
1665
1666         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1667                 dev->states[i].name[0] = '\0';
1668                 dev->states[i].desc[0] = '\0';
1669         }
1670
1671         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1672                 cx = &pr->power.states[i];
1673                 state = &dev->states[count];
1674
1675                 if (!cx->valid)
1676                         continue;
1677
1678 #ifdef CONFIG_HOTPLUG_CPU
1679                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1680                     !pr->flags.has_cst &&
1681                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1682                         continue;
1683 #endif
1684                 cpuidle_set_statedata(state, cx);
1685
1686                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1687                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1688                 state->exit_latency = cx->latency;
1689                 state->target_residency = cx->latency * latency_factor;
1690                 state->power_usage = cx->power;
1691
1692                 state->flags = 0;
1693                 switch (cx->type) {
1694                         case ACPI_STATE_C1:
1695                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1696                         if (cx->entry_method == ACPI_CSTATE_FFH)
1697                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1698
1699                         state->enter = acpi_idle_enter_c1;
1700                         dev->safe_state = state;
1701                         break;
1702
1703                         case ACPI_STATE_C2:
1704                         state->flags |= CPUIDLE_FLAG_BALANCED;
1705                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1706                         state->enter = acpi_idle_enter_simple;
1707                         dev->safe_state = state;
1708                         break;
1709
1710                         case ACPI_STATE_C3:
1711                         state->flags |= CPUIDLE_FLAG_DEEP;
1712                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1713                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1714                         state->enter = pr->flags.bm_check ?
1715                                         acpi_idle_enter_bm :
1716                                         acpi_idle_enter_simple;
1717                         break;
1718                 }
1719
1720                 count++;
1721                 if (count == CPUIDLE_STATE_MAX)
1722                         break;
1723         }
1724
1725         dev->state_count = count;
1726
1727         if (!count)
1728                 return -EINVAL;
1729
1730         return 0;
1731 }
1732
1733 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1734 {
1735         int ret;
1736
1737         if (!pr)
1738                 return -EINVAL;
1739
1740         if (nocst) {
1741                 return -ENODEV;
1742         }
1743
1744         if (!pr->flags.power_setup_done)
1745                 return -ENODEV;
1746
1747         cpuidle_pause_and_lock();
1748         cpuidle_disable_device(&pr->power.dev);
1749         acpi_processor_get_power_info(pr);
1750         acpi_processor_setup_cpuidle(pr);
1751         ret = cpuidle_enable_device(&pr->power.dev);
1752         cpuidle_resume_and_unlock();
1753
1754         return ret;
1755 }
1756
1757 #endif /* CONFIG_CPU_IDLE */
1758
1759 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1760                               struct acpi_device *device)
1761 {
1762         acpi_status status = 0;
1763         static int first_run;
1764         struct proc_dir_entry *entry = NULL;
1765         unsigned int i;
1766
1767
1768         if (!first_run) {
1769                 dmi_check_system(processor_power_dmi_table);
1770                 max_cstate = acpi_processor_cstate_check(max_cstate);
1771                 if (max_cstate < ACPI_C_STATES_MAX)
1772                         printk(KERN_NOTICE
1773                                "ACPI: processor limited to max C-state %d\n",
1774                                max_cstate);
1775                 first_run++;
1776 #if !defined(CONFIG_CPU_IDLE) && defined(CONFIG_SMP)
1777                 pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY,
1778                                 &acpi_processor_latency_notifier);
1779 #endif
1780         }
1781
1782         if (!pr)
1783                 return -EINVAL;
1784
1785         if (acpi_gbl_FADT.cst_control && !nocst) {
1786                 status =
1787                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1788                 if (ACPI_FAILURE(status)) {
1789                         ACPI_EXCEPTION((AE_INFO, status,
1790                                         "Notifying BIOS of _CST ability failed"));
1791                 }
1792         }
1793
1794         acpi_processor_get_power_info(pr);
1795         pr->flags.power_setup_done = 1;
1796
1797         /*
1798          * Install the idle handler if processor power management is supported.
1799          * Note that we use previously set idle handler will be used on
1800          * platforms that only support C1.
1801          */
1802         if ((pr->flags.power) && (!boot_option_idle_override)) {
1803 #ifdef CONFIG_CPU_IDLE
1804                 acpi_processor_setup_cpuidle(pr);
1805                 pr->power.dev.cpu = pr->id;
1806                 if (cpuidle_register_device(&pr->power.dev))
1807                         return -EIO;
1808 #endif
1809
1810                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1811                 for (i = 1; i <= pr->power.count; i++)
1812                         if (pr->power.states[i].valid)
1813                                 printk(" C%d[C%d]", i,
1814                                        pr->power.states[i].type);
1815                 printk(")\n");
1816
1817 #ifndef CONFIG_CPU_IDLE
1818                 if (pr->id == 0) {
1819                         pm_idle_save = pm_idle;
1820                         pm_idle = acpi_processor_idle;
1821                 }
1822 #endif
1823         }
1824
1825         /* 'power' [R] */
1826         entry = proc_create_data(ACPI_PROCESSOR_FILE_POWER,
1827                                  S_IRUGO, acpi_device_dir(device),
1828                                  &acpi_processor_power_fops,
1829                                  acpi_driver_data(device));
1830         if (!entry)
1831                 return -EIO;
1832         return 0;
1833 }
1834
1835 int acpi_processor_power_exit(struct acpi_processor *pr,
1836                               struct acpi_device *device)
1837 {
1838 #ifdef CONFIG_CPU_IDLE
1839         if ((pr->flags.power) && (!boot_option_idle_override))
1840                 cpuidle_unregister_device(&pr->power.dev);
1841 #endif
1842         pr->flags.power_setup_done = 0;
1843
1844         if (acpi_device_dir(device))
1845                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1846                                   acpi_device_dir(device));
1847
1848 #ifndef CONFIG_CPU_IDLE
1849
1850         /* Unregister the idle handler when processor #0 is removed. */
1851         if (pr->id == 0) {
1852                 pm_idle = pm_idle_save;
1853
1854                 /*
1855                  * We are about to unload the current idle thread pm callback
1856                  * (pm_idle), Wait for all processors to update cached/local
1857                  * copies of pm_idle before proceeding.
1858                  */
1859                 cpu_idle_wait();
1860 #ifdef CONFIG_SMP
1861                 pm_qos_remove_notifier(PM_QOS_CPU_DMA_LATENCY,
1862                                 &acpi_processor_latency_notifier);
1863 #endif
1864         }
1865 #endif
1866
1867         return 0;
1868 }