xref: /linux/drivers/acpi/processor_idle.c (revision 606d099cdd1080bbb50ea50dc52d98252f8f10a1)
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 
43 #include <asm/io.h>
44 #include <asm/uaccess.h>
45 
46 #include <acpi/acpi_bus.h>
47 #include <acpi/processor.h>
48 
49 #define ACPI_PROCESSOR_COMPONENT        0x01000000
50 #define ACPI_PROCESSOR_CLASS            "processor"
51 #define ACPI_PROCESSOR_DRIVER_NAME      "ACPI Processor Driver"
52 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
53 ACPI_MODULE_NAME("acpi_processor")
54 #define ACPI_PROCESSOR_FILE_POWER	"power"
55 #define US_TO_PM_TIMER_TICKS(t)		((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
56 #define C2_OVERHEAD			4	/* 1us (3.579 ticks per us) */
57 #define C3_OVERHEAD			4	/* 1us (3.579 ticks per us) */
58 static void (*pm_idle_save) (void) __read_mostly;
59 module_param(max_cstate, uint, 0644);
60 
61 static unsigned int nocst __read_mostly;
62 module_param(nocst, uint, 0000);
63 
64 /*
65  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
66  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
67  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
68  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
69  * reduce history for more aggressive entry into C3
70  */
71 static unsigned int bm_history __read_mostly =
72     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
73 module_param(bm_history, uint, 0644);
74 /* --------------------------------------------------------------------------
75                                 Power Management
76    -------------------------------------------------------------------------- */
77 
78 /*
79  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
80  * For now disable this. Probably a bug somewhere else.
81  *
82  * To skip this limit, boot/load with a large max_cstate limit.
83  */
84 static int set_max_cstate(struct dmi_system_id *id)
85 {
86 	if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
87 		return 0;
88 
89 	printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
90 	       " Override with \"processor.max_cstate=%d\"\n", id->ident,
91 	       (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
92 
93 	max_cstate = (long)id->driver_data;
94 
95 	return 0;
96 }
97 
98 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
99    callers to only run once -AK */
100 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
101 	{ set_max_cstate, "IBM ThinkPad R40e", {
102 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
103 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
104 	{ set_max_cstate, "IBM ThinkPad R40e", {
105 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
106 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
107 	{ set_max_cstate, "IBM ThinkPad R40e", {
108 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
109 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
110 	{ set_max_cstate, "IBM ThinkPad R40e", {
111 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
112 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
113 	{ set_max_cstate, "IBM ThinkPad R40e", {
114 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
115 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
116 	{ set_max_cstate, "IBM ThinkPad R40e", {
117 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
118 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
119 	{ set_max_cstate, "IBM ThinkPad R40e", {
120 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
121 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
122 	{ set_max_cstate, "IBM ThinkPad R40e", {
123 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
124 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
125 	{ set_max_cstate, "IBM ThinkPad R40e", {
126 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
127 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
128 	{ set_max_cstate, "IBM ThinkPad R40e", {
129 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
130 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
131 	{ set_max_cstate, "IBM ThinkPad R40e", {
132 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
133 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
134 	{ set_max_cstate, "IBM ThinkPad R40e", {
135 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
136 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
137 	{ set_max_cstate, "IBM ThinkPad R40e", {
138 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
139 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
140 	{ set_max_cstate, "IBM ThinkPad R40e", {
141 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
142 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
143 	{ set_max_cstate, "IBM ThinkPad R40e", {
144 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
145 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
146 	{ set_max_cstate, "IBM ThinkPad R40e", {
147 	  DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
148 	  DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
149 	{ set_max_cstate, "Medion 41700", {
150 	  DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
151 	  DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
152 	{ set_max_cstate, "Clevo 5600D", {
153 	  DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
154 	  DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
155 	 (void *)2},
156 	{},
157 };
158 
159 static inline u32 ticks_elapsed(u32 t1, u32 t2)
160 {
161 	if (t2 >= t1)
162 		return (t2 - t1);
163 	else if (!acpi_fadt.tmr_val_ext)
164 		return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
165 	else
166 		return ((0xFFFFFFFF - t1) + t2);
167 }
168 
169 static void
170 acpi_processor_power_activate(struct acpi_processor *pr,
171 			      struct acpi_processor_cx *new)
172 {
173 	struct acpi_processor_cx *old;
174 
175 	if (!pr || !new)
176 		return;
177 
178 	old = pr->power.state;
179 
180 	if (old)
181 		old->promotion.count = 0;
182 	new->demotion.count = 0;
183 
184 	/* Cleanup from old state. */
185 	if (old) {
186 		switch (old->type) {
187 		case ACPI_STATE_C3:
188 			/* Disable bus master reload */
189 			if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
190 				acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0,
191 						  ACPI_MTX_DO_NOT_LOCK);
192 			break;
193 		}
194 	}
195 
196 	/* Prepare to use new state. */
197 	switch (new->type) {
198 	case ACPI_STATE_C3:
199 		/* Enable bus master reload */
200 		if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
201 			acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1,
202 					  ACPI_MTX_DO_NOT_LOCK);
203 		break;
204 	}
205 
206 	pr->power.state = new;
207 
208 	return;
209 }
210 
211 static void acpi_safe_halt(void)
212 {
213 	current_thread_info()->status &= ~TS_POLLING;
214 	smp_mb__after_clear_bit();
215 	if (!need_resched())
216 		safe_halt();
217 	current_thread_info()->status |= TS_POLLING;
218 }
219 
220 static atomic_t c3_cpu_count;
221 
222 /* Common C-state entry for C2, C3, .. */
223 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
224 {
225 	if (cstate->space_id == ACPI_CSTATE_FFH) {
226 		/* Call into architectural FFH based C-state */
227 		acpi_processor_ffh_cstate_enter(cstate);
228 	} else {
229 		int unused;
230 		/* IO port based C-state */
231 		inb(cstate->address);
232 		/* Dummy wait op - must do something useless after P_LVL2 read
233 		   because chipsets cannot guarantee that STPCLK# signal
234 		   gets asserted in time to freeze execution properly. */
235 		unused = inl(acpi_fadt.xpm_tmr_blk.address);
236 	}
237 }
238 
239 static void acpi_processor_idle(void)
240 {
241 	struct acpi_processor *pr = NULL;
242 	struct acpi_processor_cx *cx = NULL;
243 	struct acpi_processor_cx *next_state = NULL;
244 	int sleep_ticks = 0;
245 	u32 t1, t2 = 0;
246 
247 	pr = processors[smp_processor_id()];
248 	if (!pr)
249 		return;
250 
251 	/*
252 	 * Interrupts must be disabled during bus mastering calculations and
253 	 * for C2/C3 transitions.
254 	 */
255 	local_irq_disable();
256 
257 	/*
258 	 * Check whether we truly need to go idle, or should
259 	 * reschedule:
260 	 */
261 	if (unlikely(need_resched())) {
262 		local_irq_enable();
263 		return;
264 	}
265 
266 	cx = pr->power.state;
267 	if (!cx) {
268 		if (pm_idle_save)
269 			pm_idle_save();
270 		else
271 			acpi_safe_halt();
272 		return;
273 	}
274 
275 	/*
276 	 * Check BM Activity
277 	 * -----------------
278 	 * Check for bus mastering activity (if required), record, and check
279 	 * for demotion.
280 	 */
281 	if (pr->flags.bm_check) {
282 		u32 bm_status = 0;
283 		unsigned long diff = jiffies - pr->power.bm_check_timestamp;
284 
285 		if (diff > 31)
286 			diff = 31;
287 
288 		pr->power.bm_activity <<= diff;
289 
290 		acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS,
291 				  &bm_status, ACPI_MTX_DO_NOT_LOCK);
292 		if (bm_status) {
293 			pr->power.bm_activity |= 0x1;
294 			acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS,
295 					  1, ACPI_MTX_DO_NOT_LOCK);
296 		}
297 		/*
298 		 * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
299 		 * the true state of bus mastering activity; forcing us to
300 		 * manually check the BMIDEA bit of each IDE channel.
301 		 */
302 		else if (errata.piix4.bmisx) {
303 			if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
304 			    || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
305 				pr->power.bm_activity |= 0x1;
306 		}
307 
308 		pr->power.bm_check_timestamp = jiffies;
309 
310 		/*
311 		 * If bus mastering is or was active this jiffy, demote
312 		 * to avoid a faulty transition.  Note that the processor
313 		 * won't enter a low-power state during this call (to this
314 		 * function) but should upon the next.
315 		 *
316 		 * TBD: A better policy might be to fallback to the demotion
317 		 *      state (use it for this quantum only) istead of
318 		 *      demoting -- and rely on duration as our sole demotion
319 		 *      qualification.  This may, however, introduce DMA
320 		 *      issues (e.g. floppy DMA transfer overrun/underrun).
321 		 */
322 		if ((pr->power.bm_activity & 0x1) &&
323 		    cx->demotion.threshold.bm) {
324 			local_irq_enable();
325 			next_state = cx->demotion.state;
326 			goto end;
327 		}
328 	}
329 
330 #ifdef CONFIG_HOTPLUG_CPU
331 	/*
332 	 * Check for P_LVL2_UP flag before entering C2 and above on
333 	 * an SMP system. We do it here instead of doing it at _CST/P_LVL
334 	 * detection phase, to work cleanly with logical CPU hotplug.
335 	 */
336 	if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
337 	    !pr->flags.has_cst && !acpi_fadt.plvl2_up)
338 		cx = &pr->power.states[ACPI_STATE_C1];
339 #endif
340 
341 	/*
342 	 * Sleep:
343 	 * ------
344 	 * Invoke the current Cx state to put the processor to sleep.
345 	 */
346 	if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
347 		current_thread_info()->status &= ~TS_POLLING;
348 		smp_mb__after_clear_bit();
349 		if (need_resched()) {
350 			current_thread_info()->status |= TS_POLLING;
351 			local_irq_enable();
352 			return;
353 		}
354 	}
355 
356 	switch (cx->type) {
357 
358 	case ACPI_STATE_C1:
359 		/*
360 		 * Invoke C1.
361 		 * Use the appropriate idle routine, the one that would
362 		 * be used without acpi C-states.
363 		 */
364 		if (pm_idle_save)
365 			pm_idle_save();
366 		else
367 			acpi_safe_halt();
368 
369 		/*
370 		 * TBD: Can't get time duration while in C1, as resumes
371 		 *      go to an ISR rather than here.  Need to instrument
372 		 *      base interrupt handler.
373 		 */
374 		sleep_ticks = 0xFFFFFFFF;
375 		break;
376 
377 	case ACPI_STATE_C2:
378 		/* Get start time (ticks) */
379 		t1 = inl(acpi_fadt.xpm_tmr_blk.address);
380 		/* Invoke C2 */
381 		acpi_cstate_enter(cx);
382 		/* Get end time (ticks) */
383 		t2 = inl(acpi_fadt.xpm_tmr_blk.address);
384 
385 #ifdef CONFIG_GENERIC_TIME
386 		/* TSC halts in C2, so notify users */
387 		mark_tsc_unstable();
388 #endif
389 		/* Re-enable interrupts */
390 		local_irq_enable();
391 		current_thread_info()->status |= TS_POLLING;
392 		/* Compute time (ticks) that we were actually asleep */
393 		sleep_ticks =
394 		    ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
395 		break;
396 
397 	case ACPI_STATE_C3:
398 
399 		if (pr->flags.bm_check) {
400 			if (atomic_inc_return(&c3_cpu_count) ==
401 			    num_online_cpus()) {
402 				/*
403 				 * All CPUs are trying to go to C3
404 				 * Disable bus master arbitration
405 				 */
406 				acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
407 						  ACPI_MTX_DO_NOT_LOCK);
408 			}
409 		} else {
410 			/* SMP with no shared cache... Invalidate cache  */
411 			ACPI_FLUSH_CPU_CACHE();
412 		}
413 
414 		/* Get start time (ticks) */
415 		t1 = inl(acpi_fadt.xpm_tmr_blk.address);
416 		/* Invoke C3 */
417 		acpi_cstate_enter(cx);
418 		/* Get end time (ticks) */
419 		t2 = inl(acpi_fadt.xpm_tmr_blk.address);
420 		if (pr->flags.bm_check) {
421 			/* Enable bus master arbitration */
422 			atomic_dec(&c3_cpu_count);
423 			acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
424 					  ACPI_MTX_DO_NOT_LOCK);
425 		}
426 
427 #ifdef CONFIG_GENERIC_TIME
428 		/* TSC halts in C3, so notify users */
429 		mark_tsc_unstable();
430 #endif
431 		/* Re-enable interrupts */
432 		local_irq_enable();
433 		current_thread_info()->status |= TS_POLLING;
434 		/* Compute time (ticks) that we were actually asleep */
435 		sleep_ticks =
436 		    ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
437 		break;
438 
439 	default:
440 		local_irq_enable();
441 		return;
442 	}
443 	cx->usage++;
444 	if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
445 		cx->time += sleep_ticks;
446 
447 	next_state = pr->power.state;
448 
449 #ifdef CONFIG_HOTPLUG_CPU
450 	/* Don't do promotion/demotion */
451 	if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
452 	    !pr->flags.has_cst && !acpi_fadt.plvl2_up) {
453 		next_state = cx;
454 		goto end;
455 	}
456 #endif
457 
458 	/*
459 	 * Promotion?
460 	 * ----------
461 	 * Track the number of longs (time asleep is greater than threshold)
462 	 * and promote when the count threshold is reached.  Note that bus
463 	 * mastering activity may prevent promotions.
464 	 * Do not promote above max_cstate.
465 	 */
466 	if (cx->promotion.state &&
467 	    ((cx->promotion.state - pr->power.states) <= max_cstate)) {
468 		if (sleep_ticks > cx->promotion.threshold.ticks &&
469 		  cx->promotion.state->latency <= system_latency_constraint()) {
470 			cx->promotion.count++;
471 			cx->demotion.count = 0;
472 			if (cx->promotion.count >=
473 			    cx->promotion.threshold.count) {
474 				if (pr->flags.bm_check) {
475 					if (!
476 					    (pr->power.bm_activity & cx->
477 					     promotion.threshold.bm)) {
478 						next_state =
479 						    cx->promotion.state;
480 						goto end;
481 					}
482 				} else {
483 					next_state = cx->promotion.state;
484 					goto end;
485 				}
486 			}
487 		}
488 	}
489 
490 	/*
491 	 * Demotion?
492 	 * ---------
493 	 * Track the number of shorts (time asleep is less than time threshold)
494 	 * and demote when the usage threshold is reached.
495 	 */
496 	if (cx->demotion.state) {
497 		if (sleep_ticks < cx->demotion.threshold.ticks) {
498 			cx->demotion.count++;
499 			cx->promotion.count = 0;
500 			if (cx->demotion.count >= cx->demotion.threshold.count) {
501 				next_state = cx->demotion.state;
502 				goto end;
503 			}
504 		}
505 	}
506 
507       end:
508 	/*
509 	 * Demote if current state exceeds max_cstate
510 	 * or if the latency of the current state is unacceptable
511 	 */
512 	if ((pr->power.state - pr->power.states) > max_cstate ||
513 		pr->power.state->latency > system_latency_constraint()) {
514 		if (cx->demotion.state)
515 			next_state = cx->demotion.state;
516 	}
517 
518 	/*
519 	 * New Cx State?
520 	 * -------------
521 	 * If we're going to start using a new Cx state we must clean up
522 	 * from the previous and prepare to use the new.
523 	 */
524 	if (next_state != pr->power.state)
525 		acpi_processor_power_activate(pr, next_state);
526 }
527 
528 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
529 {
530 	unsigned int i;
531 	unsigned int state_is_set = 0;
532 	struct acpi_processor_cx *lower = NULL;
533 	struct acpi_processor_cx *higher = NULL;
534 	struct acpi_processor_cx *cx;
535 
536 
537 	if (!pr)
538 		return -EINVAL;
539 
540 	/*
541 	 * This function sets the default Cx state policy (OS idle handler).
542 	 * Our scheme is to promote quickly to C2 but more conservatively
543 	 * to C3.  We're favoring C2  for its characteristics of low latency
544 	 * (quick response), good power savings, and ability to allow bus
545 	 * mastering activity.  Note that the Cx state policy is completely
546 	 * customizable and can be altered dynamically.
547 	 */
548 
549 	/* startup state */
550 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
551 		cx = &pr->power.states[i];
552 		if (!cx->valid)
553 			continue;
554 
555 		if (!state_is_set)
556 			pr->power.state = cx;
557 		state_is_set++;
558 		break;
559 	}
560 
561 	if (!state_is_set)
562 		return -ENODEV;
563 
564 	/* demotion */
565 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
566 		cx = &pr->power.states[i];
567 		if (!cx->valid)
568 			continue;
569 
570 		if (lower) {
571 			cx->demotion.state = lower;
572 			cx->demotion.threshold.ticks = cx->latency_ticks;
573 			cx->demotion.threshold.count = 1;
574 			if (cx->type == ACPI_STATE_C3)
575 				cx->demotion.threshold.bm = bm_history;
576 		}
577 
578 		lower = cx;
579 	}
580 
581 	/* promotion */
582 	for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
583 		cx = &pr->power.states[i];
584 		if (!cx->valid)
585 			continue;
586 
587 		if (higher) {
588 			cx->promotion.state = higher;
589 			cx->promotion.threshold.ticks = cx->latency_ticks;
590 			if (cx->type >= ACPI_STATE_C2)
591 				cx->promotion.threshold.count = 4;
592 			else
593 				cx->promotion.threshold.count = 10;
594 			if (higher->type == ACPI_STATE_C3)
595 				cx->promotion.threshold.bm = bm_history;
596 		}
597 
598 		higher = cx;
599 	}
600 
601 	return 0;
602 }
603 
604 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
605 {
606 
607 	if (!pr)
608 		return -EINVAL;
609 
610 	if (!pr->pblk)
611 		return -ENODEV;
612 
613 	/* if info is obtained from pblk/fadt, type equals state */
614 	pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
615 	pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
616 
617 #ifndef CONFIG_HOTPLUG_CPU
618 	/*
619 	 * Check for P_LVL2_UP flag before entering C2 and above on
620 	 * an SMP system.
621 	 */
622 	if ((num_online_cpus() > 1) && !acpi_fadt.plvl2_up)
623 		return -ENODEV;
624 #endif
625 
626 	/* determine C2 and C3 address from pblk */
627 	pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
628 	pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
629 
630 	/* determine latencies from FADT */
631 	pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
632 	pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
633 
634 	ACPI_DEBUG_PRINT((ACPI_DB_INFO,
635 			  "lvl2[0x%08x] lvl3[0x%08x]\n",
636 			  pr->power.states[ACPI_STATE_C2].address,
637 			  pr->power.states[ACPI_STATE_C3].address));
638 
639 	return 0;
640 }
641 
642 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
643 {
644 	if (!pr->power.states[ACPI_STATE_C1].valid) {
645 		/* set the first C-State to C1 */
646 		/* all processors need to support C1 */
647 		pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
648 		pr->power.states[ACPI_STATE_C1].valid = 1;
649 	}
650 	/* the C0 state only exists as a filler in our array */
651 	pr->power.states[ACPI_STATE_C0].valid = 1;
652 	return 0;
653 }
654 
655 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
656 {
657 	acpi_status status = 0;
658 	acpi_integer count;
659 	int current_count;
660 	int i;
661 	struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
662 	union acpi_object *cst;
663 
664 
665 	if (nocst)
666 		return -ENODEV;
667 
668 	current_count = 0;
669 
670 	status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
671 	if (ACPI_FAILURE(status)) {
672 		ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
673 		return -ENODEV;
674 	}
675 
676 	cst = (union acpi_object *)buffer.pointer;
677 
678 	/* There must be at least 2 elements */
679 	if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
680 		printk(KERN_ERR PREFIX "not enough elements in _CST\n");
681 		status = -EFAULT;
682 		goto end;
683 	}
684 
685 	count = cst->package.elements[0].integer.value;
686 
687 	/* Validate number of power states. */
688 	if (count < 1 || count != cst->package.count - 1) {
689 		printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
690 		status = -EFAULT;
691 		goto end;
692 	}
693 
694 	/* Tell driver that at least _CST is supported. */
695 	pr->flags.has_cst = 1;
696 
697 	for (i = 1; i <= count; i++) {
698 		union acpi_object *element;
699 		union acpi_object *obj;
700 		struct acpi_power_register *reg;
701 		struct acpi_processor_cx cx;
702 
703 		memset(&cx, 0, sizeof(cx));
704 
705 		element = (union acpi_object *)&(cst->package.elements[i]);
706 		if (element->type != ACPI_TYPE_PACKAGE)
707 			continue;
708 
709 		if (element->package.count != 4)
710 			continue;
711 
712 		obj = (union acpi_object *)&(element->package.elements[0]);
713 
714 		if (obj->type != ACPI_TYPE_BUFFER)
715 			continue;
716 
717 		reg = (struct acpi_power_register *)obj->buffer.pointer;
718 
719 		if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
720 		    (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
721 			continue;
722 
723 		/* There should be an easy way to extract an integer... */
724 		obj = (union acpi_object *)&(element->package.elements[1]);
725 		if (obj->type != ACPI_TYPE_INTEGER)
726 			continue;
727 
728 		cx.type = obj->integer.value;
729 		/*
730 		 * Some buggy BIOSes won't list C1 in _CST -
731 		 * Let acpi_processor_get_power_info_default() handle them later
732 		 */
733 		if (i == 1 && cx.type != ACPI_STATE_C1)
734 			current_count++;
735 
736 		cx.address = reg->address;
737 		cx.index = current_count + 1;
738 
739 		cx.space_id = ACPI_CSTATE_SYSTEMIO;
740 		if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
741 			if (acpi_processor_ffh_cstate_probe
742 					(pr->id, &cx, reg) == 0) {
743 				cx.space_id = ACPI_CSTATE_FFH;
744 			} else if (cx.type != ACPI_STATE_C1) {
745 				/*
746 				 * C1 is a special case where FIXED_HARDWARE
747 				 * can be handled in non-MWAIT way as well.
748 				 * In that case, save this _CST entry info.
749 				 * That is, we retain space_id of SYSTEM_IO for
750 				 * halt based C1.
751 				 * Otherwise, ignore this info and continue.
752 				 */
753 				continue;
754 			}
755 		}
756 
757 		obj = (union acpi_object *)&(element->package.elements[2]);
758 		if (obj->type != ACPI_TYPE_INTEGER)
759 			continue;
760 
761 		cx.latency = obj->integer.value;
762 
763 		obj = (union acpi_object *)&(element->package.elements[3]);
764 		if (obj->type != ACPI_TYPE_INTEGER)
765 			continue;
766 
767 		cx.power = obj->integer.value;
768 
769 		current_count++;
770 		memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
771 
772 		/*
773 		 * We support total ACPI_PROCESSOR_MAX_POWER - 1
774 		 * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
775 		 */
776 		if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
777 			printk(KERN_WARNING
778 			       "Limiting number of power states to max (%d)\n",
779 			       ACPI_PROCESSOR_MAX_POWER);
780 			printk(KERN_WARNING
781 			       "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
782 			break;
783 		}
784 	}
785 
786 	ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
787 			  current_count));
788 
789 	/* Validate number of power states discovered */
790 	if (current_count < 2)
791 		status = -EFAULT;
792 
793       end:
794 	kfree(buffer.pointer);
795 
796 	return status;
797 }
798 
799 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
800 {
801 
802 	if (!cx->address)
803 		return;
804 
805 	/*
806 	 * C2 latency must be less than or equal to 100
807 	 * microseconds.
808 	 */
809 	else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
810 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
811 				  "latency too large [%d]\n", cx->latency));
812 		return;
813 	}
814 
815 	/*
816 	 * Otherwise we've met all of our C2 requirements.
817 	 * Normalize the C2 latency to expidite policy
818 	 */
819 	cx->valid = 1;
820 	cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
821 
822 	return;
823 }
824 
825 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
826 					   struct acpi_processor_cx *cx)
827 {
828 	static int bm_check_flag;
829 
830 
831 	if (!cx->address)
832 		return;
833 
834 	/*
835 	 * C3 latency must be less than or equal to 1000
836 	 * microseconds.
837 	 */
838 	else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
839 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
840 				  "latency too large [%d]\n", cx->latency));
841 		return;
842 	}
843 
844 	/*
845 	 * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
846 	 * DMA transfers are used by any ISA device to avoid livelock.
847 	 * Note that we could disable Type-F DMA (as recommended by
848 	 * the erratum), but this is known to disrupt certain ISA
849 	 * devices thus we take the conservative approach.
850 	 */
851 	else if (errata.piix4.fdma) {
852 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
853 				  "C3 not supported on PIIX4 with Type-F DMA\n"));
854 		return;
855 	}
856 
857 	/* All the logic here assumes flags.bm_check is same across all CPUs */
858 	if (!bm_check_flag) {
859 		/* Determine whether bm_check is needed based on CPU  */
860 		acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
861 		bm_check_flag = pr->flags.bm_check;
862 	} else {
863 		pr->flags.bm_check = bm_check_flag;
864 	}
865 
866 	if (pr->flags.bm_check) {
867 		/* bus mastering control is necessary */
868 		if (!pr->flags.bm_control) {
869 			ACPI_DEBUG_PRINT((ACPI_DB_INFO,
870 					  "C3 support requires bus mastering control\n"));
871 			return;
872 		}
873 	} else {
874 		/*
875 		 * WBINVD should be set in fadt, for C3 state to be
876 		 * supported on when bm_check is not required.
877 		 */
878 		if (acpi_fadt.wb_invd != 1) {
879 			ACPI_DEBUG_PRINT((ACPI_DB_INFO,
880 					  "Cache invalidation should work properly"
881 					  " for C3 to be enabled on SMP systems\n"));
882 			return;
883 		}
884 		acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
885 				  0, ACPI_MTX_DO_NOT_LOCK);
886 	}
887 
888 	/*
889 	 * Otherwise we've met all of our C3 requirements.
890 	 * Normalize the C3 latency to expidite policy.  Enable
891 	 * checking of bus mastering status (bm_check) so we can
892 	 * use this in our C3 policy
893 	 */
894 	cx->valid = 1;
895 	cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
896 
897 	return;
898 }
899 
900 static int acpi_processor_power_verify(struct acpi_processor *pr)
901 {
902 	unsigned int i;
903 	unsigned int working = 0;
904 
905 #ifdef ARCH_APICTIMER_STOPS_ON_C3
906 	int timer_broadcast = 0;
907 	cpumask_t mask = cpumask_of_cpu(pr->id);
908 	on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
909 #endif
910 
911 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
912 		struct acpi_processor_cx *cx = &pr->power.states[i];
913 
914 		switch (cx->type) {
915 		case ACPI_STATE_C1:
916 			cx->valid = 1;
917 			break;
918 
919 		case ACPI_STATE_C2:
920 			acpi_processor_power_verify_c2(cx);
921 #ifdef ARCH_APICTIMER_STOPS_ON_C3
922 			/* Some AMD systems fake C3 as C2, but still
923 			   have timer troubles */
924 			if (cx->valid &&
925 				boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
926 				timer_broadcast++;
927 #endif
928 			break;
929 
930 		case ACPI_STATE_C3:
931 			acpi_processor_power_verify_c3(pr, cx);
932 #ifdef ARCH_APICTIMER_STOPS_ON_C3
933 			if (cx->valid)
934 				timer_broadcast++;
935 #endif
936 			break;
937 		}
938 
939 		if (cx->valid)
940 			working++;
941 	}
942 
943 #ifdef ARCH_APICTIMER_STOPS_ON_C3
944 	if (timer_broadcast)
945 		on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
946 #endif
947 
948 	return (working);
949 }
950 
951 static int acpi_processor_get_power_info(struct acpi_processor *pr)
952 {
953 	unsigned int i;
954 	int result;
955 
956 
957 	/* NOTE: the idle thread may not be running while calling
958 	 * this function */
959 
960 	/* Zero initialize all the C-states info. */
961 	memset(pr->power.states, 0, sizeof(pr->power.states));
962 
963 	result = acpi_processor_get_power_info_cst(pr);
964 	if (result == -ENODEV)
965 		result = acpi_processor_get_power_info_fadt(pr);
966 
967 	if (result)
968 		return result;
969 
970 	acpi_processor_get_power_info_default(pr);
971 
972 	pr->power.count = acpi_processor_power_verify(pr);
973 
974 	/*
975 	 * Set Default Policy
976 	 * ------------------
977 	 * Now that we know which states are supported, set the default
978 	 * policy.  Note that this policy can be changed dynamically
979 	 * (e.g. encourage deeper sleeps to conserve battery life when
980 	 * not on AC).
981 	 */
982 	result = acpi_processor_set_power_policy(pr);
983 	if (result)
984 		return result;
985 
986 	/*
987 	 * if one state of type C2 or C3 is available, mark this
988 	 * CPU as being "idle manageable"
989 	 */
990 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
991 		if (pr->power.states[i].valid) {
992 			pr->power.count = i;
993 			if (pr->power.states[i].type >= ACPI_STATE_C2)
994 				pr->flags.power = 1;
995 		}
996 	}
997 
998 	return 0;
999 }
1000 
1001 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1002 {
1003 	int result = 0;
1004 
1005 
1006 	if (!pr)
1007 		return -EINVAL;
1008 
1009 	if (nocst) {
1010 		return -ENODEV;
1011 	}
1012 
1013 	if (!pr->flags.power_setup_done)
1014 		return -ENODEV;
1015 
1016 	/* Fall back to the default idle loop */
1017 	pm_idle = pm_idle_save;
1018 	synchronize_sched();	/* Relies on interrupts forcing exit from idle. */
1019 
1020 	pr->flags.power = 0;
1021 	result = acpi_processor_get_power_info(pr);
1022 	if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1023 		pm_idle = acpi_processor_idle;
1024 
1025 	return result;
1026 }
1027 
1028 /* proc interface */
1029 
1030 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1031 {
1032 	struct acpi_processor *pr = (struct acpi_processor *)seq->private;
1033 	unsigned int i;
1034 
1035 
1036 	if (!pr)
1037 		goto end;
1038 
1039 	seq_printf(seq, "active state:            C%zd\n"
1040 		   "max_cstate:              C%d\n"
1041 		   "bus master activity:     %08x\n"
1042 		   "maximum allowed latency: %d usec\n",
1043 		   pr->power.state ? pr->power.state - pr->power.states : 0,
1044 		   max_cstate, (unsigned)pr->power.bm_activity,
1045 		   system_latency_constraint());
1046 
1047 	seq_puts(seq, "states:\n");
1048 
1049 	for (i = 1; i <= pr->power.count; i++) {
1050 		seq_printf(seq, "   %cC%d:                  ",
1051 			   (&pr->power.states[i] ==
1052 			    pr->power.state ? '*' : ' '), i);
1053 
1054 		if (!pr->power.states[i].valid) {
1055 			seq_puts(seq, "<not supported>\n");
1056 			continue;
1057 		}
1058 
1059 		switch (pr->power.states[i].type) {
1060 		case ACPI_STATE_C1:
1061 			seq_printf(seq, "type[C1] ");
1062 			break;
1063 		case ACPI_STATE_C2:
1064 			seq_printf(seq, "type[C2] ");
1065 			break;
1066 		case ACPI_STATE_C3:
1067 			seq_printf(seq, "type[C3] ");
1068 			break;
1069 		default:
1070 			seq_printf(seq, "type[--] ");
1071 			break;
1072 		}
1073 
1074 		if (pr->power.states[i].promotion.state)
1075 			seq_printf(seq, "promotion[C%zd] ",
1076 				   (pr->power.states[i].promotion.state -
1077 				    pr->power.states));
1078 		else
1079 			seq_puts(seq, "promotion[--] ");
1080 
1081 		if (pr->power.states[i].demotion.state)
1082 			seq_printf(seq, "demotion[C%zd] ",
1083 				   (pr->power.states[i].demotion.state -
1084 				    pr->power.states));
1085 		else
1086 			seq_puts(seq, "demotion[--] ");
1087 
1088 		seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1089 			   pr->power.states[i].latency,
1090 			   pr->power.states[i].usage,
1091 			   pr->power.states[i].time);
1092 	}
1093 
1094       end:
1095 	return 0;
1096 }
1097 
1098 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1099 {
1100 	return single_open(file, acpi_processor_power_seq_show,
1101 			   PDE(inode)->data);
1102 }
1103 
1104 static const struct file_operations acpi_processor_power_fops = {
1105 	.open = acpi_processor_power_open_fs,
1106 	.read = seq_read,
1107 	.llseek = seq_lseek,
1108 	.release = single_release,
1109 };
1110 
1111 #ifdef CONFIG_SMP
1112 static void smp_callback(void *v)
1113 {
1114 	/* we already woke the CPU up, nothing more to do */
1115 }
1116 
1117 /*
1118  * This function gets called when a part of the kernel has a new latency
1119  * requirement.  This means we need to get all processors out of their C-state,
1120  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1121  * wakes them all right up.
1122  */
1123 static int acpi_processor_latency_notify(struct notifier_block *b,
1124 		unsigned long l, void *v)
1125 {
1126 	smp_call_function(smp_callback, NULL, 0, 1);
1127 	return NOTIFY_OK;
1128 }
1129 
1130 static struct notifier_block acpi_processor_latency_notifier = {
1131 	.notifier_call = acpi_processor_latency_notify,
1132 };
1133 #endif
1134 
1135 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1136 			      struct acpi_device *device)
1137 {
1138 	acpi_status status = 0;
1139 	static int first_run;
1140 	struct proc_dir_entry *entry = NULL;
1141 	unsigned int i;
1142 
1143 
1144 	if (!first_run) {
1145 		dmi_check_system(processor_power_dmi_table);
1146 		if (max_cstate < ACPI_C_STATES_MAX)
1147 			printk(KERN_NOTICE
1148 			       "ACPI: processor limited to max C-state %d\n",
1149 			       max_cstate);
1150 		first_run++;
1151 #ifdef CONFIG_SMP
1152 		register_latency_notifier(&acpi_processor_latency_notifier);
1153 #endif
1154 	}
1155 
1156 	if (!pr)
1157 		return -EINVAL;
1158 
1159 	if (acpi_fadt.cst_cnt && !nocst) {
1160 		status =
1161 		    acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1162 		if (ACPI_FAILURE(status)) {
1163 			ACPI_EXCEPTION((AE_INFO, status,
1164 					"Notifying BIOS of _CST ability failed"));
1165 		}
1166 	}
1167 
1168 	acpi_processor_get_power_info(pr);
1169 
1170 	/*
1171 	 * Install the idle handler if processor power management is supported.
1172 	 * Note that we use previously set idle handler will be used on
1173 	 * platforms that only support C1.
1174 	 */
1175 	if ((pr->flags.power) && (!boot_option_idle_override)) {
1176 		printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1177 		for (i = 1; i <= pr->power.count; i++)
1178 			if (pr->power.states[i].valid)
1179 				printk(" C%d[C%d]", i,
1180 				       pr->power.states[i].type);
1181 		printk(")\n");
1182 
1183 		if (pr->id == 0) {
1184 			pm_idle_save = pm_idle;
1185 			pm_idle = acpi_processor_idle;
1186 		}
1187 	}
1188 
1189 	/* 'power' [R] */
1190 	entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1191 				  S_IRUGO, acpi_device_dir(device));
1192 	if (!entry)
1193 		return -EIO;
1194 	else {
1195 		entry->proc_fops = &acpi_processor_power_fops;
1196 		entry->data = acpi_driver_data(device);
1197 		entry->owner = THIS_MODULE;
1198 	}
1199 
1200 	pr->flags.power_setup_done = 1;
1201 
1202 	return 0;
1203 }
1204 
1205 int acpi_processor_power_exit(struct acpi_processor *pr,
1206 			      struct acpi_device *device)
1207 {
1208 
1209 	pr->flags.power_setup_done = 0;
1210 
1211 	if (acpi_device_dir(device))
1212 		remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1213 				  acpi_device_dir(device));
1214 
1215 	/* Unregister the idle handler when processor #0 is removed. */
1216 	if (pr->id == 0) {
1217 		pm_idle = pm_idle_save;
1218 
1219 		/*
1220 		 * We are about to unload the current idle thread pm callback
1221 		 * (pm_idle), Wait for all processors to update cached/local
1222 		 * copies of pm_idle before proceeding.
1223 		 */
1224 		cpu_idle_wait();
1225 #ifdef CONFIG_SMP
1226 		unregister_latency_notifier(&acpi_processor_latency_notifier);
1227 #endif
1228 	}
1229 
1230 	return 0;
1231 }
1232