xref: /linux/drivers/acpi/processor_idle.c (revision 643d1f7fe3aa12c8bdea6fa5b4ba874ff6dd601d)
1 /*
2  * processor_idle - idle state submodule to the ACPI processor driver
3  *
4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
6  *  Copyright (C) 2004, 2005 Dominik Brodowski <linux@brodo.de>
7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
8  *  			- Added processor hotplug support
9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
10  *  			- Added support for C3 on SMP
11  *
12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
13  *
14  *  This program is free software; you can redistribute it and/or modify
15  *  it under the terms of the GNU General Public License as published by
16  *  the Free Software Foundation; either version 2 of the License, or (at
17  *  your option) any later version.
18  *
19  *  This program is distributed in the hope that it will be useful, but
20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
22  *  General Public License for more details.
23  *
24  *  You should have received a copy of the GNU General Public License along
25  *  with this program; if not, write to the Free Software Foundation, Inc.,
26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
27  *
28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
29  */
30 
31 #include <linux/kernel.h>
32 #include <linux/module.h>
33 #include <linux/init.h>
34 #include <linux/cpufreq.h>
35 #include <linux/proc_fs.h>
36 #include <linux/seq_file.h>
37 #include <linux/acpi.h>
38 #include <linux/dmi.h>
39 #include <linux/moduleparam.h>
40 #include <linux/sched.h>	/* need_resched() */
41 #include <linux/latency.h>
42 #include <linux/clockchips.h>
43 #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 #endif
102 
103 /*
104  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
105  * For now disable this. Probably a bug somewhere else.
106  *
107  * To skip this limit, boot/load with a large max_cstate limit.
108  */
109 static int set_max_cstate(const struct dmi_system_id *id)
110 {
111 	if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
112 		return 0;
113 
114 	printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
115 	       " Override with \"processor.max_cstate=%d\"\n", id->ident,
116 	       (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
117 
118 	max_cstate = (long)id->driver_data;
119 
120 	return 0;
121 }
122 
123 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
124    callers to only run once -AK */
125 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
126 	{ set_max_cstate, "IBM ThinkPad R40e", {
127 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
128 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
129 	{ set_max_cstate, "IBM ThinkPad R40e", {
130 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
131 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
132 	{ set_max_cstate, "IBM ThinkPad R40e", {
133 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
134 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
135 	{ set_max_cstate, "IBM ThinkPad R40e", {
136 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
137 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
138 	{ set_max_cstate, "IBM ThinkPad R40e", {
139 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
140 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
141 	{ set_max_cstate, "IBM ThinkPad R40e", {
142 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
143 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
144 	{ set_max_cstate, "IBM ThinkPad R40e", {
145 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
146 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
147 	{ set_max_cstate, "IBM ThinkPad R40e", {
148 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
149 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
150 	{ set_max_cstate, "IBM ThinkPad R40e", {
151 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
152 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
153 	{ set_max_cstate, "IBM ThinkPad R40e", {
154 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
155 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
156 	{ set_max_cstate, "IBM ThinkPad R40e", {
157 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
158 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
159 	{ set_max_cstate, "IBM ThinkPad R40e", {
160 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
161 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
162 	{ set_max_cstate, "IBM ThinkPad R40e", {
163 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
164 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
165 	{ set_max_cstate, "IBM ThinkPad R40e", {
166 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
167 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
168 	{ set_max_cstate, "IBM ThinkPad R40e", {
169 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
170 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
171 	{ set_max_cstate, "IBM ThinkPad R40e", {
172 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
173 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
174 	{ set_max_cstate, "Medion 41700", {
175 	  DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
176 	  DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
177 	{ set_max_cstate, "Clevo 5600D", {
178 	  DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
179 	  DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
180 	 (void *)2},
181 	{},
182 };
183 
184 static inline u32 ticks_elapsed(u32 t1, u32 t2)
185 {
186 	if (t2 >= t1)
187 		return (t2 - t1);
188 	else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
189 		return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
190 	else
191 		return ((0xFFFFFFFF - t1) + t2);
192 }
193 
194 static inline u32 ticks_elapsed_in_us(u32 t1, u32 t2)
195 {
196 	if (t2 >= t1)
197 		return PM_TIMER_TICKS_TO_US(t2 - t1);
198 	else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
199 		return PM_TIMER_TICKS_TO_US(((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
200 	else
201 		return PM_TIMER_TICKS_TO_US((0xFFFFFFFF - t1) + t2);
202 }
203 
204 static void acpi_safe_halt(void)
205 {
206 	current_thread_info()->status &= ~TS_POLLING;
207 	/*
208 	 * TS_POLLING-cleared state must be visible before we
209 	 * test NEED_RESCHED:
210 	 */
211 	smp_mb();
212 	if (!need_resched())
213 		safe_halt();
214 	current_thread_info()->status |= TS_POLLING;
215 }
216 
217 #ifndef CONFIG_CPU_IDLE
218 
219 static void
220 acpi_processor_power_activate(struct acpi_processor *pr,
221 			      struct acpi_processor_cx *new)
222 {
223 	struct acpi_processor_cx *old;
224 
225 	if (!pr || !new)
226 		return;
227 
228 	old = pr->power.state;
229 
230 	if (old)
231 		old->promotion.count = 0;
232 	new->demotion.count = 0;
233 
234 	/* Cleanup from old state. */
235 	if (old) {
236 		switch (old->type) {
237 		case ACPI_STATE_C3:
238 			/* Disable bus master reload */
239 			if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
240 				acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
241 			break;
242 		}
243 	}
244 
245 	/* Prepare to use new state. */
246 	switch (new->type) {
247 	case ACPI_STATE_C3:
248 		/* Enable bus master reload */
249 		if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
250 			acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
251 		break;
252 	}
253 
254 	pr->power.state = new;
255 
256 	return;
257 }
258 
259 static atomic_t c3_cpu_count;
260 
261 /* Common C-state entry for C2, C3, .. */
262 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
263 {
264 	if (cstate->space_id == ACPI_CSTATE_FFH) {
265 		/* Call into architectural FFH based C-state */
266 		acpi_processor_ffh_cstate_enter(cstate);
267 	} else {
268 		int unused;
269 		/* IO port based C-state */
270 		inb(cstate->address);
271 		/* Dummy wait op - must do something useless after P_LVL2 read
272 		   because chipsets cannot guarantee that STPCLK# signal
273 		   gets asserted in time to freeze execution properly. */
274 		unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
275 	}
276 }
277 #endif /* !CONFIG_CPU_IDLE */
278 
279 #ifdef ARCH_APICTIMER_STOPS_ON_C3
280 
281 /*
282  * Some BIOS implementations switch to C3 in the published C2 state.
283  * This seems to be a common problem on AMD boxen, but other vendors
284  * are affected too. We pick the most conservative approach: we assume
285  * that the local APIC stops in both C2 and C3.
286  */
287 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
288 				   struct acpi_processor_cx *cx)
289 {
290 	struct acpi_processor_power *pwr = &pr->power;
291 	u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
292 
293 	/*
294 	 * Check, if one of the previous states already marked the lapic
295 	 * unstable
296 	 */
297 	if (pwr->timer_broadcast_on_state < state)
298 		return;
299 
300 	if (cx->type >= type)
301 		pr->power.timer_broadcast_on_state = state;
302 }
303 
304 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr)
305 {
306 	unsigned long reason;
307 
308 	reason = pr->power.timer_broadcast_on_state < INT_MAX ?
309 		CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
310 
311 	clockevents_notify(reason, &pr->id);
312 }
313 
314 /* Power(C) State timer broadcast control */
315 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
316 				       struct acpi_processor_cx *cx,
317 				       int broadcast)
318 {
319 	int state = cx - pr->power.states;
320 
321 	if (state >= pr->power.timer_broadcast_on_state) {
322 		unsigned long reason;
323 
324 		reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
325 			CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
326 		clockevents_notify(reason, &pr->id);
327 	}
328 }
329 
330 #else
331 
332 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
333 				   struct acpi_processor_cx *cstate) { }
334 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr) { }
335 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
336 				       struct acpi_processor_cx *cx,
337 				       int broadcast)
338 {
339 }
340 
341 #endif
342 
343 /*
344  * Suspend / resume control
345  */
346 static int acpi_idle_suspend;
347 
348 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
349 {
350 	acpi_idle_suspend = 1;
351 	return 0;
352 }
353 
354 int acpi_processor_resume(struct acpi_device * device)
355 {
356 	acpi_idle_suspend = 0;
357 	return 0;
358 }
359 
360 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
361 static int tsc_halts_in_c(int state)
362 {
363 	switch (boot_cpu_data.x86_vendor) {
364 	case X86_VENDOR_AMD:
365 		/*
366 		 * AMD Fam10h TSC will tick in all
367 		 * C/P/S0/S1 states when this bit is set.
368 		 */
369 		if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
370 			return 0;
371 		/*FALL THROUGH*/
372 	case X86_VENDOR_INTEL:
373 		/* Several cases known where TSC halts in C2 too */
374 	default:
375 		return state > ACPI_STATE_C1;
376 	}
377 }
378 #endif
379 
380 #ifndef CONFIG_CPU_IDLE
381 static void acpi_processor_idle(void)
382 {
383 	struct acpi_processor *pr = NULL;
384 	struct acpi_processor_cx *cx = NULL;
385 	struct acpi_processor_cx *next_state = NULL;
386 	int sleep_ticks = 0;
387 	u32 t1, t2 = 0;
388 
389 	/*
390 	 * Interrupts must be disabled during bus mastering calculations and
391 	 * for C2/C3 transitions.
392 	 */
393 	local_irq_disable();
394 
395 	pr = processors[smp_processor_id()];
396 	if (!pr) {
397 		local_irq_enable();
398 		return;
399 	}
400 
401 	/*
402 	 * Check whether we truly need to go idle, or should
403 	 * reschedule:
404 	 */
405 	if (unlikely(need_resched())) {
406 		local_irq_enable();
407 		return;
408 	}
409 
410 	cx = pr->power.state;
411 	if (!cx || acpi_idle_suspend) {
412 		if (pm_idle_save)
413 			pm_idle_save();
414 		else
415 			acpi_safe_halt();
416 		return;
417 	}
418 
419 	/*
420 	 * Check BM Activity
421 	 * -----------------
422 	 * Check for bus mastering activity (if required), record, and check
423 	 * for demotion.
424 	 */
425 	if (pr->flags.bm_check) {
426 		u32 bm_status = 0;
427 		unsigned long diff = jiffies - pr->power.bm_check_timestamp;
428 
429 		if (diff > 31)
430 			diff = 31;
431 
432 		pr->power.bm_activity <<= diff;
433 
434 		acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
435 		if (bm_status) {
436 			pr->power.bm_activity |= 0x1;
437 			acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
438 		}
439 		/*
440 		 * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
441 		 * the true state of bus mastering activity; forcing us to
442 		 * manually check the BMIDEA bit of each IDE channel.
443 		 */
444 		else if (errata.piix4.bmisx) {
445 			if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
446 			    || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
447 				pr->power.bm_activity |= 0x1;
448 		}
449 
450 		pr->power.bm_check_timestamp = jiffies;
451 
452 		/*
453 		 * If bus mastering is or was active this jiffy, demote
454 		 * to avoid a faulty transition.  Note that the processor
455 		 * won't enter a low-power state during this call (to this
456 		 * function) but should upon the next.
457 		 *
458 		 * TBD: A better policy might be to fallback to the demotion
459 		 *      state (use it for this quantum only) istead of
460 		 *      demoting -- and rely on duration as our sole demotion
461 		 *      qualification.  This may, however, introduce DMA
462 		 *      issues (e.g. floppy DMA transfer overrun/underrun).
463 		 */
464 		if ((pr->power.bm_activity & 0x1) &&
465 		    cx->demotion.threshold.bm) {
466 			local_irq_enable();
467 			next_state = cx->demotion.state;
468 			goto end;
469 		}
470 	}
471 
472 #ifdef CONFIG_HOTPLUG_CPU
473 	/*
474 	 * Check for P_LVL2_UP flag before entering C2 and above on
475 	 * an SMP system. We do it here instead of doing it at _CST/P_LVL
476 	 * detection phase, to work cleanly with logical CPU hotplug.
477 	 */
478 	if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
479 	    !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
480 		cx = &pr->power.states[ACPI_STATE_C1];
481 #endif
482 
483 	/*
484 	 * Sleep:
485 	 * ------
486 	 * Invoke the current Cx state to put the processor to sleep.
487 	 */
488 	if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
489 		current_thread_info()->status &= ~TS_POLLING;
490 		/*
491 		 * TS_POLLING-cleared state must be visible before we
492 		 * test NEED_RESCHED:
493 		 */
494 		smp_mb();
495 		if (need_resched()) {
496 			current_thread_info()->status |= TS_POLLING;
497 			local_irq_enable();
498 			return;
499 		}
500 	}
501 
502 	switch (cx->type) {
503 
504 	case ACPI_STATE_C1:
505 		/*
506 		 * Invoke C1.
507 		 * Use the appropriate idle routine, the one that would
508 		 * be used without acpi C-states.
509 		 */
510 		if (pm_idle_save)
511 			pm_idle_save();
512 		else
513 			acpi_safe_halt();
514 
515 		/*
516 		 * TBD: Can't get time duration while in C1, as resumes
517 		 *      go to an ISR rather than here.  Need to instrument
518 		 *      base interrupt handler.
519 		 *
520 		 * Note: the TSC better not stop in C1, sched_clock() will
521 		 *       skew otherwise.
522 		 */
523 		sleep_ticks = 0xFFFFFFFF;
524 		break;
525 
526 	case ACPI_STATE_C2:
527 		/* Get start time (ticks) */
528 		t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
529 		/* Tell the scheduler that we are going deep-idle: */
530 		sched_clock_idle_sleep_event();
531 		/* Invoke C2 */
532 		acpi_state_timer_broadcast(pr, cx, 1);
533 		acpi_cstate_enter(cx);
534 		/* Get end time (ticks) */
535 		t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
536 
537 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
538 		/* TSC halts in C2, so notify users */
539 		if (tsc_halts_in_c(ACPI_STATE_C2))
540 			mark_tsc_unstable("possible TSC halt in C2");
541 #endif
542 		/* Compute time (ticks) that we were actually asleep */
543 		sleep_ticks = ticks_elapsed(t1, t2);
544 
545 		/* Tell the scheduler how much we idled: */
546 		sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
547 
548 		/* Re-enable interrupts */
549 		local_irq_enable();
550 		/* Do not account our idle-switching overhead: */
551 		sleep_ticks -= cx->latency_ticks + C2_OVERHEAD;
552 
553 		current_thread_info()->status |= TS_POLLING;
554 		acpi_state_timer_broadcast(pr, cx, 0);
555 		break;
556 
557 	case ACPI_STATE_C3:
558 		acpi_unlazy_tlb(smp_processor_id());
559 		/*
560 		 * Must be done before busmaster disable as we might
561 		 * need to access HPET !
562 		 */
563 		acpi_state_timer_broadcast(pr, cx, 1);
564 		/*
565 		 * disable bus master
566 		 * bm_check implies we need ARB_DIS
567 		 * !bm_check implies we need cache flush
568 		 * bm_control implies whether we can do ARB_DIS
569 		 *
570 		 * That leaves a case where bm_check is set and bm_control is
571 		 * not set. In that case we cannot do much, we enter C3
572 		 * without doing anything.
573 		 */
574 		if (pr->flags.bm_check && pr->flags.bm_control) {
575 			if (atomic_inc_return(&c3_cpu_count) ==
576 			    num_online_cpus()) {
577 				/*
578 				 * All CPUs are trying to go to C3
579 				 * Disable bus master arbitration
580 				 */
581 				acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
582 			}
583 		} else if (!pr->flags.bm_check) {
584 			/* SMP with no shared cache... Invalidate cache  */
585 			ACPI_FLUSH_CPU_CACHE();
586 		}
587 
588 		/* Get start time (ticks) */
589 		t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
590 		/* Invoke C3 */
591 		/* Tell the scheduler that we are going deep-idle: */
592 		sched_clock_idle_sleep_event();
593 		acpi_cstate_enter(cx);
594 		/* Get end time (ticks) */
595 		t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
596 		if (pr->flags.bm_check && pr->flags.bm_control) {
597 			/* Enable bus master arbitration */
598 			atomic_dec(&c3_cpu_count);
599 			acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
600 		}
601 
602 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
603 		/* TSC halts in C3, so notify users */
604 		if (tsc_halts_in_c(ACPI_STATE_C3))
605 			mark_tsc_unstable("TSC halts in C3");
606 #endif
607 		/* Compute time (ticks) that we were actually asleep */
608 		sleep_ticks = ticks_elapsed(t1, t2);
609 		/* Tell the scheduler how much we idled: */
610 		sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
611 
612 		/* Re-enable interrupts */
613 		local_irq_enable();
614 		/* Do not account our idle-switching overhead: */
615 		sleep_ticks -= cx->latency_ticks + C3_OVERHEAD;
616 
617 		current_thread_info()->status |= TS_POLLING;
618 		acpi_state_timer_broadcast(pr, cx, 0);
619 		break;
620 
621 	default:
622 		local_irq_enable();
623 		return;
624 	}
625 	cx->usage++;
626 	if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
627 		cx->time += sleep_ticks;
628 
629 	next_state = pr->power.state;
630 
631 #ifdef CONFIG_HOTPLUG_CPU
632 	/* Don't do promotion/demotion */
633 	if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
634 	    !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
635 		next_state = cx;
636 		goto end;
637 	}
638 #endif
639 
640 	/*
641 	 * Promotion?
642 	 * ----------
643 	 * Track the number of longs (time asleep is greater than threshold)
644 	 * and promote when the count threshold is reached.  Note that bus
645 	 * mastering activity may prevent promotions.
646 	 * Do not promote above max_cstate.
647 	 */
648 	if (cx->promotion.state &&
649 	    ((cx->promotion.state - pr->power.states) <= max_cstate)) {
650 		if (sleep_ticks > cx->promotion.threshold.ticks &&
651 		  cx->promotion.state->latency <= system_latency_constraint()) {
652 			cx->promotion.count++;
653 			cx->demotion.count = 0;
654 			if (cx->promotion.count >=
655 			    cx->promotion.threshold.count) {
656 				if (pr->flags.bm_check) {
657 					if (!
658 					    (pr->power.bm_activity & cx->
659 					     promotion.threshold.bm)) {
660 						next_state =
661 						    cx->promotion.state;
662 						goto end;
663 					}
664 				} else {
665 					next_state = cx->promotion.state;
666 					goto end;
667 				}
668 			}
669 		}
670 	}
671 
672 	/*
673 	 * Demotion?
674 	 * ---------
675 	 * Track the number of shorts (time asleep is less than time threshold)
676 	 * and demote when the usage threshold is reached.
677 	 */
678 	if (cx->demotion.state) {
679 		if (sleep_ticks < cx->demotion.threshold.ticks) {
680 			cx->demotion.count++;
681 			cx->promotion.count = 0;
682 			if (cx->demotion.count >= cx->demotion.threshold.count) {
683 				next_state = cx->demotion.state;
684 				goto end;
685 			}
686 		}
687 	}
688 
689       end:
690 	/*
691 	 * Demote if current state exceeds max_cstate
692 	 * or if the latency of the current state is unacceptable
693 	 */
694 	if ((pr->power.state - pr->power.states) > max_cstate ||
695 		pr->power.state->latency > system_latency_constraint()) {
696 		if (cx->demotion.state)
697 			next_state = cx->demotion.state;
698 	}
699 
700 	/*
701 	 * New Cx State?
702 	 * -------------
703 	 * If we're going to start using a new Cx state we must clean up
704 	 * from the previous and prepare to use the new.
705 	 */
706 	if (next_state != pr->power.state)
707 		acpi_processor_power_activate(pr, next_state);
708 }
709 
710 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
711 {
712 	unsigned int i;
713 	unsigned int state_is_set = 0;
714 	struct acpi_processor_cx *lower = NULL;
715 	struct acpi_processor_cx *higher = NULL;
716 	struct acpi_processor_cx *cx;
717 
718 
719 	if (!pr)
720 		return -EINVAL;
721 
722 	/*
723 	 * This function sets the default Cx state policy (OS idle handler).
724 	 * Our scheme is to promote quickly to C2 but more conservatively
725 	 * to C3.  We're favoring C2  for its characteristics of low latency
726 	 * (quick response), good power savings, and ability to allow bus
727 	 * mastering activity.  Note that the Cx state policy is completely
728 	 * customizable and can be altered dynamically.
729 	 */
730 
731 	/* startup state */
732 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
733 		cx = &pr->power.states[i];
734 		if (!cx->valid)
735 			continue;
736 
737 		if (!state_is_set)
738 			pr->power.state = cx;
739 		state_is_set++;
740 		break;
741 	}
742 
743 	if (!state_is_set)
744 		return -ENODEV;
745 
746 	/* demotion */
747 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
748 		cx = &pr->power.states[i];
749 		if (!cx->valid)
750 			continue;
751 
752 		if (lower) {
753 			cx->demotion.state = lower;
754 			cx->demotion.threshold.ticks = cx->latency_ticks;
755 			cx->demotion.threshold.count = 1;
756 			if (cx->type == ACPI_STATE_C3)
757 				cx->demotion.threshold.bm = bm_history;
758 		}
759 
760 		lower = cx;
761 	}
762 
763 	/* promotion */
764 	for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
765 		cx = &pr->power.states[i];
766 		if (!cx->valid)
767 			continue;
768 
769 		if (higher) {
770 			cx->promotion.state = higher;
771 			cx->promotion.threshold.ticks = cx->latency_ticks;
772 			if (cx->type >= ACPI_STATE_C2)
773 				cx->promotion.threshold.count = 4;
774 			else
775 				cx->promotion.threshold.count = 10;
776 			if (higher->type == ACPI_STATE_C3)
777 				cx->promotion.threshold.bm = bm_history;
778 		}
779 
780 		higher = cx;
781 	}
782 
783 	return 0;
784 }
785 #endif /* !CONFIG_CPU_IDLE */
786 
787 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
788 {
789 
790 	if (!pr)
791 		return -EINVAL;
792 
793 	if (!pr->pblk)
794 		return -ENODEV;
795 
796 	/* if info is obtained from pblk/fadt, type equals state */
797 	pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
798 	pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
799 
800 #ifndef CONFIG_HOTPLUG_CPU
801 	/*
802 	 * Check for P_LVL2_UP flag before entering C2 and above on
803 	 * an SMP system.
804 	 */
805 	if ((num_online_cpus() > 1) &&
806 	    !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
807 		return -ENODEV;
808 #endif
809 
810 	/* determine C2 and C3 address from pblk */
811 	pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
812 	pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
813 
814 	/* determine latencies from FADT */
815 	pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
816 	pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
817 
818 	ACPI_DEBUG_PRINT((ACPI_DB_INFO,
819 			  "lvl2[0x%08x] lvl3[0x%08x]\n",
820 			  pr->power.states[ACPI_STATE_C2].address,
821 			  pr->power.states[ACPI_STATE_C3].address));
822 
823 	return 0;
824 }
825 
826 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
827 {
828 	if (!pr->power.states[ACPI_STATE_C1].valid) {
829 		/* set the first C-State to C1 */
830 		/* all processors need to support C1 */
831 		pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
832 		pr->power.states[ACPI_STATE_C1].valid = 1;
833 	}
834 	/* the C0 state only exists as a filler in our array */
835 	pr->power.states[ACPI_STATE_C0].valid = 1;
836 	return 0;
837 }
838 
839 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
840 {
841 	acpi_status status = 0;
842 	acpi_integer count;
843 	int current_count;
844 	int i;
845 	struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
846 	union acpi_object *cst;
847 
848 
849 	if (nocst)
850 		return -ENODEV;
851 
852 	current_count = 0;
853 
854 	status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
855 	if (ACPI_FAILURE(status)) {
856 		ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
857 		return -ENODEV;
858 	}
859 
860 	cst = buffer.pointer;
861 
862 	/* There must be at least 2 elements */
863 	if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
864 		printk(KERN_ERR PREFIX "not enough elements in _CST\n");
865 		status = -EFAULT;
866 		goto end;
867 	}
868 
869 	count = cst->package.elements[0].integer.value;
870 
871 	/* Validate number of power states. */
872 	if (count < 1 || count != cst->package.count - 1) {
873 		printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
874 		status = -EFAULT;
875 		goto end;
876 	}
877 
878 	/* Tell driver that at least _CST is supported. */
879 	pr->flags.has_cst = 1;
880 
881 	for (i = 1; i <= count; i++) {
882 		union acpi_object *element;
883 		union acpi_object *obj;
884 		struct acpi_power_register *reg;
885 		struct acpi_processor_cx cx;
886 
887 		memset(&cx, 0, sizeof(cx));
888 
889 		element = &(cst->package.elements[i]);
890 		if (element->type != ACPI_TYPE_PACKAGE)
891 			continue;
892 
893 		if (element->package.count != 4)
894 			continue;
895 
896 		obj = &(element->package.elements[0]);
897 
898 		if (obj->type != ACPI_TYPE_BUFFER)
899 			continue;
900 
901 		reg = (struct acpi_power_register *)obj->buffer.pointer;
902 
903 		if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
904 		    (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
905 			continue;
906 
907 		/* There should be an easy way to extract an integer... */
908 		obj = &(element->package.elements[1]);
909 		if (obj->type != ACPI_TYPE_INTEGER)
910 			continue;
911 
912 		cx.type = obj->integer.value;
913 		/*
914 		 * Some buggy BIOSes won't list C1 in _CST -
915 		 * Let acpi_processor_get_power_info_default() handle them later
916 		 */
917 		if (i == 1 && cx.type != ACPI_STATE_C1)
918 			current_count++;
919 
920 		cx.address = reg->address;
921 		cx.index = current_count + 1;
922 
923 		cx.space_id = ACPI_CSTATE_SYSTEMIO;
924 		if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
925 			if (acpi_processor_ffh_cstate_probe
926 					(pr->id, &cx, reg) == 0) {
927 				cx.space_id = ACPI_CSTATE_FFH;
928 			} else if (cx.type != ACPI_STATE_C1) {
929 				/*
930 				 * C1 is a special case where FIXED_HARDWARE
931 				 * can be handled in non-MWAIT way as well.
932 				 * In that case, save this _CST entry info.
933 				 * That is, we retain space_id of SYSTEM_IO for
934 				 * halt based C1.
935 				 * Otherwise, ignore this info and continue.
936 				 */
937 				continue;
938 			}
939 		}
940 
941 		obj = &(element->package.elements[2]);
942 		if (obj->type != ACPI_TYPE_INTEGER)
943 			continue;
944 
945 		cx.latency = obj->integer.value;
946 
947 		obj = &(element->package.elements[3]);
948 		if (obj->type != ACPI_TYPE_INTEGER)
949 			continue;
950 
951 		cx.power = obj->integer.value;
952 
953 		current_count++;
954 		memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
955 
956 		/*
957 		 * We support total ACPI_PROCESSOR_MAX_POWER - 1
958 		 * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
959 		 */
960 		if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
961 			printk(KERN_WARNING
962 			       "Limiting number of power states to max (%d)\n",
963 			       ACPI_PROCESSOR_MAX_POWER);
964 			printk(KERN_WARNING
965 			       "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
966 			break;
967 		}
968 	}
969 
970 	ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
971 			  current_count));
972 
973 	/* Validate number of power states discovered */
974 	if (current_count < 2)
975 		status = -EFAULT;
976 
977       end:
978 	kfree(buffer.pointer);
979 
980 	return status;
981 }
982 
983 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
984 {
985 
986 	if (!cx->address)
987 		return;
988 
989 	/*
990 	 * C2 latency must be less than or equal to 100
991 	 * microseconds.
992 	 */
993 	else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
994 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
995 				  "latency too large [%d]\n", cx->latency));
996 		return;
997 	}
998 
999 	/*
1000 	 * Otherwise we've met all of our C2 requirements.
1001 	 * Normalize the C2 latency to expidite policy
1002 	 */
1003 	cx->valid = 1;
1004 
1005 #ifndef CONFIG_CPU_IDLE
1006 	cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1007 #else
1008 	cx->latency_ticks = cx->latency;
1009 #endif
1010 
1011 	return;
1012 }
1013 
1014 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
1015 					   struct acpi_processor_cx *cx)
1016 {
1017 	static int bm_check_flag;
1018 
1019 
1020 	if (!cx->address)
1021 		return;
1022 
1023 	/*
1024 	 * C3 latency must be less than or equal to 1000
1025 	 * microseconds.
1026 	 */
1027 	else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
1028 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1029 				  "latency too large [%d]\n", cx->latency));
1030 		return;
1031 	}
1032 
1033 	/*
1034 	 * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
1035 	 * DMA transfers are used by any ISA device to avoid livelock.
1036 	 * Note that we could disable Type-F DMA (as recommended by
1037 	 * the erratum), but this is known to disrupt certain ISA
1038 	 * devices thus we take the conservative approach.
1039 	 */
1040 	else if (errata.piix4.fdma) {
1041 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1042 				  "C3 not supported on PIIX4 with Type-F DMA\n"));
1043 		return;
1044 	}
1045 
1046 	/* All the logic here assumes flags.bm_check is same across all CPUs */
1047 	if (!bm_check_flag) {
1048 		/* Determine whether bm_check is needed based on CPU  */
1049 		acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
1050 		bm_check_flag = pr->flags.bm_check;
1051 	} else {
1052 		pr->flags.bm_check = bm_check_flag;
1053 	}
1054 
1055 	if (pr->flags.bm_check) {
1056 		if (!pr->flags.bm_control) {
1057 			if (pr->flags.has_cst != 1) {
1058 				/* bus mastering control is necessary */
1059 				ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1060 					"C3 support requires BM control\n"));
1061 				return;
1062 			} else {
1063 				/* Here we enter C3 without bus mastering */
1064 				ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1065 					"C3 support without BM control\n"));
1066 			}
1067 		}
1068 	} else {
1069 		/*
1070 		 * WBINVD should be set in fadt, for C3 state to be
1071 		 * supported on when bm_check is not required.
1072 		 */
1073 		if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
1074 			ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1075 					  "Cache invalidation should work properly"
1076 					  " for C3 to be enabled on SMP systems\n"));
1077 			return;
1078 		}
1079 		acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1080 	}
1081 
1082 	/*
1083 	 * Otherwise we've met all of our C3 requirements.
1084 	 * Normalize the C3 latency to expidite policy.  Enable
1085 	 * checking of bus mastering status (bm_check) so we can
1086 	 * use this in our C3 policy
1087 	 */
1088 	cx->valid = 1;
1089 
1090 #ifndef CONFIG_CPU_IDLE
1091 	cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1092 #else
1093 	cx->latency_ticks = cx->latency;
1094 #endif
1095 
1096 	return;
1097 }
1098 
1099 static int acpi_processor_power_verify(struct acpi_processor *pr)
1100 {
1101 	unsigned int i;
1102 	unsigned int working = 0;
1103 
1104 	pr->power.timer_broadcast_on_state = INT_MAX;
1105 
1106 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1107 		struct acpi_processor_cx *cx = &pr->power.states[i];
1108 
1109 		switch (cx->type) {
1110 		case ACPI_STATE_C1:
1111 			cx->valid = 1;
1112 			break;
1113 
1114 		case ACPI_STATE_C2:
1115 			acpi_processor_power_verify_c2(cx);
1116 			if (cx->valid)
1117 				acpi_timer_check_state(i, pr, cx);
1118 			break;
1119 
1120 		case ACPI_STATE_C3:
1121 			acpi_processor_power_verify_c3(pr, cx);
1122 			if (cx->valid)
1123 				acpi_timer_check_state(i, pr, cx);
1124 			break;
1125 		}
1126 
1127 		if (cx->valid)
1128 			working++;
1129 	}
1130 
1131 	acpi_propagate_timer_broadcast(pr);
1132 
1133 	return (working);
1134 }
1135 
1136 static int acpi_processor_get_power_info(struct acpi_processor *pr)
1137 {
1138 	unsigned int i;
1139 	int result;
1140 
1141 
1142 	/* NOTE: the idle thread may not be running while calling
1143 	 * this function */
1144 
1145 	/* Zero initialize all the C-states info. */
1146 	memset(pr->power.states, 0, sizeof(pr->power.states));
1147 
1148 	result = acpi_processor_get_power_info_cst(pr);
1149 	if (result == -ENODEV)
1150 		result = acpi_processor_get_power_info_fadt(pr);
1151 
1152 	if (result)
1153 		return result;
1154 
1155 	acpi_processor_get_power_info_default(pr);
1156 
1157 	pr->power.count = acpi_processor_power_verify(pr);
1158 
1159 #ifndef CONFIG_CPU_IDLE
1160 	/*
1161 	 * Set Default Policy
1162 	 * ------------------
1163 	 * Now that we know which states are supported, set the default
1164 	 * policy.  Note that this policy can be changed dynamically
1165 	 * (e.g. encourage deeper sleeps to conserve battery life when
1166 	 * not on AC).
1167 	 */
1168 	result = acpi_processor_set_power_policy(pr);
1169 	if (result)
1170 		return result;
1171 #endif
1172 
1173 	/*
1174 	 * if one state of type C2 or C3 is available, mark this
1175 	 * CPU as being "idle manageable"
1176 	 */
1177 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1178 		if (pr->power.states[i].valid) {
1179 			pr->power.count = i;
1180 			if (pr->power.states[i].type >= ACPI_STATE_C2)
1181 				pr->flags.power = 1;
1182 		}
1183 	}
1184 
1185 	return 0;
1186 }
1187 
1188 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1189 {
1190 	struct acpi_processor *pr = seq->private;
1191 	unsigned int i;
1192 
1193 
1194 	if (!pr)
1195 		goto end;
1196 
1197 	seq_printf(seq, "active state:            C%zd\n"
1198 		   "max_cstate:              C%d\n"
1199 		   "bus master activity:     %08x\n"
1200 		   "maximum allowed latency: %d usec\n",
1201 		   pr->power.state ? pr->power.state - pr->power.states : 0,
1202 		   max_cstate, (unsigned)pr->power.bm_activity,
1203 		   system_latency_constraint());
1204 
1205 	seq_puts(seq, "states:\n");
1206 
1207 	for (i = 1; i <= pr->power.count; i++) {
1208 		seq_printf(seq, "   %cC%d:                  ",
1209 			   (&pr->power.states[i] ==
1210 			    pr->power.state ? '*' : ' '), i);
1211 
1212 		if (!pr->power.states[i].valid) {
1213 			seq_puts(seq, "<not supported>\n");
1214 			continue;
1215 		}
1216 
1217 		switch (pr->power.states[i].type) {
1218 		case ACPI_STATE_C1:
1219 			seq_printf(seq, "type[C1] ");
1220 			break;
1221 		case ACPI_STATE_C2:
1222 			seq_printf(seq, "type[C2] ");
1223 			break;
1224 		case ACPI_STATE_C3:
1225 			seq_printf(seq, "type[C3] ");
1226 			break;
1227 		default:
1228 			seq_printf(seq, "type[--] ");
1229 			break;
1230 		}
1231 
1232 		if (pr->power.states[i].promotion.state)
1233 			seq_printf(seq, "promotion[C%zd] ",
1234 				   (pr->power.states[i].promotion.state -
1235 				    pr->power.states));
1236 		else
1237 			seq_puts(seq, "promotion[--] ");
1238 
1239 		if (pr->power.states[i].demotion.state)
1240 			seq_printf(seq, "demotion[C%zd] ",
1241 				   (pr->power.states[i].demotion.state -
1242 				    pr->power.states));
1243 		else
1244 			seq_puts(seq, "demotion[--] ");
1245 
1246 		seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1247 			   pr->power.states[i].latency,
1248 			   pr->power.states[i].usage,
1249 			   (unsigned long long)pr->power.states[i].time);
1250 	}
1251 
1252       end:
1253 	return 0;
1254 }
1255 
1256 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1257 {
1258 	return single_open(file, acpi_processor_power_seq_show,
1259 			   PDE(inode)->data);
1260 }
1261 
1262 static const struct file_operations acpi_processor_power_fops = {
1263 	.open = acpi_processor_power_open_fs,
1264 	.read = seq_read,
1265 	.llseek = seq_lseek,
1266 	.release = single_release,
1267 };
1268 
1269 #ifndef CONFIG_CPU_IDLE
1270 
1271 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1272 {
1273 	int result = 0;
1274 
1275 
1276 	if (!pr)
1277 		return -EINVAL;
1278 
1279 	if (nocst) {
1280 		return -ENODEV;
1281 	}
1282 
1283 	if (!pr->flags.power_setup_done)
1284 		return -ENODEV;
1285 
1286 	/* Fall back to the default idle loop */
1287 	pm_idle = pm_idle_save;
1288 	synchronize_sched();	/* Relies on interrupts forcing exit from idle. */
1289 
1290 	pr->flags.power = 0;
1291 	result = acpi_processor_get_power_info(pr);
1292 	if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1293 		pm_idle = acpi_processor_idle;
1294 
1295 	return result;
1296 }
1297 
1298 #ifdef CONFIG_SMP
1299 static void smp_callback(void *v)
1300 {
1301 	/* we already woke the CPU up, nothing more to do */
1302 }
1303 
1304 /*
1305  * This function gets called when a part of the kernel has a new latency
1306  * requirement.  This means we need to get all processors out of their C-state,
1307  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1308  * wakes them all right up.
1309  */
1310 static int acpi_processor_latency_notify(struct notifier_block *b,
1311 		unsigned long l, void *v)
1312 {
1313 	smp_call_function(smp_callback, NULL, 0, 1);
1314 	return NOTIFY_OK;
1315 }
1316 
1317 static struct notifier_block acpi_processor_latency_notifier = {
1318 	.notifier_call = acpi_processor_latency_notify,
1319 };
1320 
1321 #endif
1322 
1323 #else /* CONFIG_CPU_IDLE */
1324 
1325 /**
1326  * acpi_idle_bm_check - checks if bus master activity was detected
1327  */
1328 static int acpi_idle_bm_check(void)
1329 {
1330 	u32 bm_status = 0;
1331 
1332 	acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
1333 	if (bm_status)
1334 		acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
1335 	/*
1336 	 * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
1337 	 * the true state of bus mastering activity; forcing us to
1338 	 * manually check the BMIDEA bit of each IDE channel.
1339 	 */
1340 	else if (errata.piix4.bmisx) {
1341 		if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
1342 		    || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
1343 			bm_status = 1;
1344 	}
1345 	return bm_status;
1346 }
1347 
1348 /**
1349  * acpi_idle_update_bm_rld - updates the BM_RLD bit depending on target state
1350  * @pr: the processor
1351  * @target: the new target state
1352  */
1353 static inline void acpi_idle_update_bm_rld(struct acpi_processor *pr,
1354 					   struct acpi_processor_cx *target)
1355 {
1356 	if (pr->flags.bm_rld_set && target->type != ACPI_STATE_C3) {
1357 		acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1358 		pr->flags.bm_rld_set = 0;
1359 	}
1360 
1361 	if (!pr->flags.bm_rld_set && target->type == ACPI_STATE_C3) {
1362 		acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
1363 		pr->flags.bm_rld_set = 1;
1364 	}
1365 }
1366 
1367 /**
1368  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
1369  * @cx: cstate data
1370  */
1371 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
1372 {
1373 	if (cx->space_id == ACPI_CSTATE_FFH) {
1374 		/* Call into architectural FFH based C-state */
1375 		acpi_processor_ffh_cstate_enter(cx);
1376 	} else {
1377 		int unused;
1378 		/* IO port based C-state */
1379 		inb(cx->address);
1380 		/* Dummy wait op - must do something useless after P_LVL2 read
1381 		   because chipsets cannot guarantee that STPCLK# signal
1382 		   gets asserted in time to freeze execution properly. */
1383 		unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
1384 	}
1385 }
1386 
1387 /**
1388  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
1389  * @dev: the target CPU
1390  * @state: the state data
1391  *
1392  * This is equivalent to the HALT instruction.
1393  */
1394 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
1395 			      struct cpuidle_state *state)
1396 {
1397 	struct acpi_processor *pr;
1398 	struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1399 	pr = processors[smp_processor_id()];
1400 
1401 	if (unlikely(!pr))
1402 		return 0;
1403 
1404 	if (pr->flags.bm_check)
1405 		acpi_idle_update_bm_rld(pr, cx);
1406 
1407 	acpi_safe_halt();
1408 
1409 	cx->usage++;
1410 
1411 	return 0;
1412 }
1413 
1414 /**
1415  * acpi_idle_enter_simple - enters an ACPI state without BM handling
1416  * @dev: the target CPU
1417  * @state: the state data
1418  */
1419 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
1420 				  struct cpuidle_state *state)
1421 {
1422 	struct acpi_processor *pr;
1423 	struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1424 	u32 t1, t2;
1425 	int sleep_ticks = 0;
1426 
1427 	pr = processors[smp_processor_id()];
1428 
1429 	if (unlikely(!pr))
1430 		return 0;
1431 
1432 	if (acpi_idle_suspend)
1433 		return(acpi_idle_enter_c1(dev, state));
1434 
1435 	local_irq_disable();
1436 	current_thread_info()->status &= ~TS_POLLING;
1437 	/*
1438 	 * TS_POLLING-cleared state must be visible before we test
1439 	 * NEED_RESCHED:
1440 	 */
1441 	smp_mb();
1442 
1443 	if (unlikely(need_resched())) {
1444 		current_thread_info()->status |= TS_POLLING;
1445 		local_irq_enable();
1446 		return 0;
1447 	}
1448 
1449 	acpi_unlazy_tlb(smp_processor_id());
1450 	/*
1451 	 * Must be done before busmaster disable as we might need to
1452 	 * access HPET !
1453 	 */
1454 	acpi_state_timer_broadcast(pr, cx, 1);
1455 
1456 	if (pr->flags.bm_check)
1457 		acpi_idle_update_bm_rld(pr, cx);
1458 
1459 	if (cx->type == ACPI_STATE_C3)
1460 		ACPI_FLUSH_CPU_CACHE();
1461 
1462 	t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1463 	/* Tell the scheduler that we are going deep-idle: */
1464 	sched_clock_idle_sleep_event();
1465 	acpi_idle_do_entry(cx);
1466 	t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1467 
1468 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
1469 	/* TSC could halt in idle, so notify users */
1470 	if (tsc_halts_in_c(cx->type))
1471 		mark_tsc_unstable("TSC halts in idle");;
1472 #endif
1473 	sleep_ticks = ticks_elapsed(t1, t2);
1474 
1475 	/* Tell the scheduler how much we idled: */
1476 	sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1477 
1478 	local_irq_enable();
1479 	current_thread_info()->status |= TS_POLLING;
1480 
1481 	cx->usage++;
1482 
1483 	acpi_state_timer_broadcast(pr, cx, 0);
1484 	cx->time += sleep_ticks;
1485 	return ticks_elapsed_in_us(t1, t2);
1486 }
1487 
1488 static int c3_cpu_count;
1489 static DEFINE_SPINLOCK(c3_lock);
1490 
1491 /**
1492  * acpi_idle_enter_bm - enters C3 with proper BM handling
1493  * @dev: the target CPU
1494  * @state: the state data
1495  *
1496  * If BM is detected, the deepest non-C3 idle state is entered instead.
1497  */
1498 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
1499 			      struct cpuidle_state *state)
1500 {
1501 	struct acpi_processor *pr;
1502 	struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1503 	u32 t1, t2;
1504 	int sleep_ticks = 0;
1505 
1506 	pr = processors[smp_processor_id()];
1507 
1508 	if (unlikely(!pr))
1509 		return 0;
1510 
1511 	if (acpi_idle_suspend)
1512 		return(acpi_idle_enter_c1(dev, state));
1513 
1514 	if (acpi_idle_bm_check()) {
1515 		if (dev->safe_state) {
1516 			return dev->safe_state->enter(dev, dev->safe_state);
1517 		} else {
1518 			acpi_safe_halt();
1519 			return 0;
1520 		}
1521 	}
1522 
1523 	local_irq_disable();
1524 	current_thread_info()->status &= ~TS_POLLING;
1525 	/*
1526 	 * TS_POLLING-cleared state must be visible before we test
1527 	 * NEED_RESCHED:
1528 	 */
1529 	smp_mb();
1530 
1531 	if (unlikely(need_resched())) {
1532 		current_thread_info()->status |= TS_POLLING;
1533 		local_irq_enable();
1534 		return 0;
1535 	}
1536 
1537 	/* Tell the scheduler that we are going deep-idle: */
1538 	sched_clock_idle_sleep_event();
1539 	/*
1540 	 * Must be done before busmaster disable as we might need to
1541 	 * access HPET !
1542 	 */
1543 	acpi_state_timer_broadcast(pr, cx, 1);
1544 
1545 	acpi_idle_update_bm_rld(pr, cx);
1546 
1547 	/*
1548 	 * disable bus master
1549 	 * bm_check implies we need ARB_DIS
1550 	 * !bm_check implies we need cache flush
1551 	 * bm_control implies whether we can do ARB_DIS
1552 	 *
1553 	 * That leaves a case where bm_check is set and bm_control is
1554 	 * not set. In that case we cannot do much, we enter C3
1555 	 * without doing anything.
1556 	 */
1557 	if (pr->flags.bm_check && pr->flags.bm_control) {
1558 		spin_lock(&c3_lock);
1559 		c3_cpu_count++;
1560 		/* Disable bus master arbitration when all CPUs are in C3 */
1561 		if (c3_cpu_count == num_online_cpus())
1562 			acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
1563 		spin_unlock(&c3_lock);
1564 	} else if (!pr->flags.bm_check) {
1565 		ACPI_FLUSH_CPU_CACHE();
1566 	}
1567 
1568 	t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1569 	acpi_idle_do_entry(cx);
1570 	t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1571 
1572 	/* Re-enable bus master arbitration */
1573 	if (pr->flags.bm_check && pr->flags.bm_control) {
1574 		spin_lock(&c3_lock);
1575 		acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
1576 		c3_cpu_count--;
1577 		spin_unlock(&c3_lock);
1578 	}
1579 
1580 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
1581 	/* TSC could halt in idle, so notify users */
1582 	if (tsc_halts_in_c(ACPI_STATE_C3))
1583 		mark_tsc_unstable("TSC halts in idle");
1584 #endif
1585 	sleep_ticks = ticks_elapsed(t1, t2);
1586 	/* Tell the scheduler how much we idled: */
1587 	sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1588 
1589 	local_irq_enable();
1590 	current_thread_info()->status |= TS_POLLING;
1591 
1592 	cx->usage++;
1593 
1594 	acpi_state_timer_broadcast(pr, cx, 0);
1595 	cx->time += sleep_ticks;
1596 	return ticks_elapsed_in_us(t1, t2);
1597 }
1598 
1599 struct cpuidle_driver acpi_idle_driver = {
1600 	.name =		"acpi_idle",
1601 	.owner =	THIS_MODULE,
1602 };
1603 
1604 /**
1605  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1606  * @pr: the ACPI processor
1607  */
1608 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1609 {
1610 	int i, count = 0;
1611 	struct acpi_processor_cx *cx;
1612 	struct cpuidle_state *state;
1613 	struct cpuidle_device *dev = &pr->power.dev;
1614 
1615 	if (!pr->flags.power_setup_done)
1616 		return -EINVAL;
1617 
1618 	if (pr->flags.power == 0) {
1619 		return -EINVAL;
1620 	}
1621 
1622 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1623 		cx = &pr->power.states[i];
1624 		state = &dev->states[count];
1625 
1626 		if (!cx->valid)
1627 			continue;
1628 
1629 #ifdef CONFIG_HOTPLUG_CPU
1630 		if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1631 		    !pr->flags.has_cst &&
1632 		    !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1633 			continue;
1634 #endif
1635 		cpuidle_set_statedata(state, cx);
1636 
1637 		snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1638 		state->exit_latency = cx->latency;
1639 		state->target_residency = cx->latency * 6;
1640 		state->power_usage = cx->power;
1641 
1642 		state->flags = 0;
1643 		switch (cx->type) {
1644 			case ACPI_STATE_C1:
1645 			state->flags |= CPUIDLE_FLAG_SHALLOW;
1646 			state->enter = acpi_idle_enter_c1;
1647 			dev->safe_state = state;
1648 			break;
1649 
1650 			case ACPI_STATE_C2:
1651 			state->flags |= CPUIDLE_FLAG_BALANCED;
1652 			state->flags |= CPUIDLE_FLAG_TIME_VALID;
1653 			state->enter = acpi_idle_enter_simple;
1654 			dev->safe_state = state;
1655 			break;
1656 
1657 			case ACPI_STATE_C3:
1658 			state->flags |= CPUIDLE_FLAG_DEEP;
1659 			state->flags |= CPUIDLE_FLAG_TIME_VALID;
1660 			state->flags |= CPUIDLE_FLAG_CHECK_BM;
1661 			state->enter = pr->flags.bm_check ?
1662 					acpi_idle_enter_bm :
1663 					acpi_idle_enter_simple;
1664 			break;
1665 		}
1666 
1667 		count++;
1668 	}
1669 
1670 	dev->state_count = count;
1671 
1672 	if (!count)
1673 		return -EINVAL;
1674 
1675 	return 0;
1676 }
1677 
1678 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1679 {
1680 	int ret;
1681 
1682 	if (!pr)
1683 		return -EINVAL;
1684 
1685 	if (nocst) {
1686 		return -ENODEV;
1687 	}
1688 
1689 	if (!pr->flags.power_setup_done)
1690 		return -ENODEV;
1691 
1692 	cpuidle_pause_and_lock();
1693 	cpuidle_disable_device(&pr->power.dev);
1694 	acpi_processor_get_power_info(pr);
1695 	acpi_processor_setup_cpuidle(pr);
1696 	ret = cpuidle_enable_device(&pr->power.dev);
1697 	cpuidle_resume_and_unlock();
1698 
1699 	return ret;
1700 }
1701 
1702 #endif /* CONFIG_CPU_IDLE */
1703 
1704 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1705 			      struct acpi_device *device)
1706 {
1707 	acpi_status status = 0;
1708 	static int first_run;
1709 	struct proc_dir_entry *entry = NULL;
1710 	unsigned int i;
1711 
1712 
1713 	if (!first_run) {
1714 		dmi_check_system(processor_power_dmi_table);
1715 		max_cstate = acpi_processor_cstate_check(max_cstate);
1716 		if (max_cstate < ACPI_C_STATES_MAX)
1717 			printk(KERN_NOTICE
1718 			       "ACPI: processor limited to max C-state %d\n",
1719 			       max_cstate);
1720 		first_run++;
1721 #if !defined (CONFIG_CPU_IDLE) && defined (CONFIG_SMP)
1722 		register_latency_notifier(&acpi_processor_latency_notifier);
1723 #endif
1724 	}
1725 
1726 	if (!pr)
1727 		return -EINVAL;
1728 
1729 	if (acpi_gbl_FADT.cst_control && !nocst) {
1730 		status =
1731 		    acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1732 		if (ACPI_FAILURE(status)) {
1733 			ACPI_EXCEPTION((AE_INFO, status,
1734 					"Notifying BIOS of _CST ability failed"));
1735 		}
1736 	}
1737 
1738 	acpi_processor_get_power_info(pr);
1739 	pr->flags.power_setup_done = 1;
1740 
1741 	/*
1742 	 * Install the idle handler if processor power management is supported.
1743 	 * Note that we use previously set idle handler will be used on
1744 	 * platforms that only support C1.
1745 	 */
1746 	if ((pr->flags.power) && (!boot_option_idle_override)) {
1747 #ifdef CONFIG_CPU_IDLE
1748 		acpi_processor_setup_cpuidle(pr);
1749 		pr->power.dev.cpu = pr->id;
1750 		if (cpuidle_register_device(&pr->power.dev))
1751 			return -EIO;
1752 #endif
1753 
1754 		printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1755 		for (i = 1; i <= pr->power.count; i++)
1756 			if (pr->power.states[i].valid)
1757 				printk(" C%d[C%d]", i,
1758 				       pr->power.states[i].type);
1759 		printk(")\n");
1760 
1761 #ifndef CONFIG_CPU_IDLE
1762 		if (pr->id == 0) {
1763 			pm_idle_save = pm_idle;
1764 			pm_idle = acpi_processor_idle;
1765 		}
1766 #endif
1767 	}
1768 
1769 	/* 'power' [R] */
1770 	entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1771 				  S_IRUGO, acpi_device_dir(device));
1772 	if (!entry)
1773 		return -EIO;
1774 	else {
1775 		entry->proc_fops = &acpi_processor_power_fops;
1776 		entry->data = acpi_driver_data(device);
1777 		entry->owner = THIS_MODULE;
1778 	}
1779 
1780 	return 0;
1781 }
1782 
1783 int acpi_processor_power_exit(struct acpi_processor *pr,
1784 			      struct acpi_device *device)
1785 {
1786 #ifdef CONFIG_CPU_IDLE
1787 	if ((pr->flags.power) && (!boot_option_idle_override))
1788 		cpuidle_unregister_device(&pr->power.dev);
1789 #endif
1790 	pr->flags.power_setup_done = 0;
1791 
1792 	if (acpi_device_dir(device))
1793 		remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1794 				  acpi_device_dir(device));
1795 
1796 #ifndef CONFIG_CPU_IDLE
1797 
1798 	/* Unregister the idle handler when processor #0 is removed. */
1799 	if (pr->id == 0) {
1800 		pm_idle = pm_idle_save;
1801 
1802 		/*
1803 		 * We are about to unload the current idle thread pm callback
1804 		 * (pm_idle), Wait for all processors to update cached/local
1805 		 * copies of pm_idle before proceeding.
1806 		 */
1807 		cpu_idle_wait();
1808 #ifdef CONFIG_SMP
1809 		unregister_latency_notifier(&acpi_processor_latency_notifier);
1810 #endif
1811 	}
1812 #endif
1813 
1814 	return 0;
1815 }
1816