xref: /linux/drivers/acpi/processor_idle.c (revision b0148a98ec5151fec82064d95f11eb9efbc628ea)
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_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
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 			break;
192 		}
193 	}
194 
195 	/* Prepare to use new state. */
196 	switch (new->type) {
197 	case ACPI_STATE_C3:
198 		/* Enable bus master reload */
199 		if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
200 			acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
201 		break;
202 	}
203 
204 	pr->power.state = new;
205 
206 	return;
207 }
208 
209 static void acpi_safe_halt(void)
210 {
211 	current_thread_info()->status &= ~TS_POLLING;
212 	/*
213 	 * TS_POLLING-cleared state must be visible before we
214 	 * test NEED_RESCHED:
215 	 */
216 	smp_mb();
217 	if (!need_resched())
218 		safe_halt();
219 	current_thread_info()->status |= TS_POLLING;
220 }
221 
222 static atomic_t c3_cpu_count;
223 
224 /* Common C-state entry for C2, C3, .. */
225 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
226 {
227 	if (cstate->space_id == ACPI_CSTATE_FFH) {
228 		/* Call into architectural FFH based C-state */
229 		acpi_processor_ffh_cstate_enter(cstate);
230 	} else {
231 		int unused;
232 		/* IO port based C-state */
233 		inb(cstate->address);
234 		/* Dummy wait op - must do something useless after P_LVL2 read
235 		   because chipsets cannot guarantee that STPCLK# signal
236 		   gets asserted in time to freeze execution properly. */
237 		unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
238 	}
239 }
240 
241 static void acpi_processor_idle(void)
242 {
243 	struct acpi_processor *pr = NULL;
244 	struct acpi_processor_cx *cx = NULL;
245 	struct acpi_processor_cx *next_state = NULL;
246 	int sleep_ticks = 0;
247 	u32 t1, t2 = 0;
248 
249 	pr = processors[smp_processor_id()];
250 	if (!pr)
251 		return;
252 
253 	/*
254 	 * Interrupts must be disabled during bus mastering calculations and
255 	 * for C2/C3 transitions.
256 	 */
257 	local_irq_disable();
258 
259 	/*
260 	 * Check whether we truly need to go idle, or should
261 	 * reschedule:
262 	 */
263 	if (unlikely(need_resched())) {
264 		local_irq_enable();
265 		return;
266 	}
267 
268 	cx = pr->power.state;
269 	if (!cx) {
270 		if (pm_idle_save)
271 			pm_idle_save();
272 		else
273 			acpi_safe_halt();
274 		return;
275 	}
276 
277 	/*
278 	 * Check BM Activity
279 	 * -----------------
280 	 * Check for bus mastering activity (if required), record, and check
281 	 * for demotion.
282 	 */
283 	if (pr->flags.bm_check) {
284 		u32 bm_status = 0;
285 		unsigned long diff = jiffies - pr->power.bm_check_timestamp;
286 
287 		if (diff > 31)
288 			diff = 31;
289 
290 		pr->power.bm_activity <<= diff;
291 
292 		acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
293 		if (bm_status) {
294 			pr->power.bm_activity |= 0x1;
295 			acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
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_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
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 		/*
349 		 * TS_POLLING-cleared state must be visible before we
350 		 * test NEED_RESCHED:
351 		 */
352 		smp_mb();
353 		if (need_resched()) {
354 			current_thread_info()->status |= TS_POLLING;
355 			local_irq_enable();
356 			return;
357 		}
358 	}
359 
360 	switch (cx->type) {
361 
362 	case ACPI_STATE_C1:
363 		/*
364 		 * Invoke C1.
365 		 * Use the appropriate idle routine, the one that would
366 		 * be used without acpi C-states.
367 		 */
368 		if (pm_idle_save)
369 			pm_idle_save();
370 		else
371 			acpi_safe_halt();
372 
373 		/*
374 		 * TBD: Can't get time duration while in C1, as resumes
375 		 *      go to an ISR rather than here.  Need to instrument
376 		 *      base interrupt handler.
377 		 */
378 		sleep_ticks = 0xFFFFFFFF;
379 		break;
380 
381 	case ACPI_STATE_C2:
382 		/* Get start time (ticks) */
383 		t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
384 		/* Invoke C2 */
385 		acpi_cstate_enter(cx);
386 		/* Get end time (ticks) */
387 		t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
388 
389 #ifdef CONFIG_GENERIC_TIME
390 		/* TSC halts in C2, so notify users */
391 		mark_tsc_unstable();
392 #endif
393 		/* Re-enable interrupts */
394 		local_irq_enable();
395 		current_thread_info()->status |= TS_POLLING;
396 		/* Compute time (ticks) that we were actually asleep */
397 		sleep_ticks =
398 		    ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
399 		break;
400 
401 	case ACPI_STATE_C3:
402 
403 		if (pr->flags.bm_check) {
404 			if (atomic_inc_return(&c3_cpu_count) ==
405 			    num_online_cpus()) {
406 				/*
407 				 * All CPUs are trying to go to C3
408 				 * Disable bus master arbitration
409 				 */
410 				acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
411 			}
412 		} else {
413 			/* SMP with no shared cache... Invalidate cache  */
414 			ACPI_FLUSH_CPU_CACHE();
415 		}
416 
417 		/* Get start time (ticks) */
418 		t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
419 		/* Invoke C3 */
420 		acpi_cstate_enter(cx);
421 		/* Get end time (ticks) */
422 		t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
423 		if (pr->flags.bm_check) {
424 			/* Enable bus master arbitration */
425 			atomic_dec(&c3_cpu_count);
426 			acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
427 		}
428 
429 #ifdef CONFIG_GENERIC_TIME
430 		/* TSC halts in C3, so notify users */
431 		mark_tsc_unstable();
432 #endif
433 		/* Re-enable interrupts */
434 		local_irq_enable();
435 		current_thread_info()->status |= TS_POLLING;
436 		/* Compute time (ticks) that we were actually asleep */
437 		sleep_ticks =
438 		    ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
439 		break;
440 
441 	default:
442 		local_irq_enable();
443 		return;
444 	}
445 	cx->usage++;
446 	if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
447 		cx->time += sleep_ticks;
448 
449 	next_state = pr->power.state;
450 
451 #ifdef CONFIG_HOTPLUG_CPU
452 	/* Don't do promotion/demotion */
453 	if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
454 	    !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
455 		next_state = cx;
456 		goto end;
457 	}
458 #endif
459 
460 	/*
461 	 * Promotion?
462 	 * ----------
463 	 * Track the number of longs (time asleep is greater than threshold)
464 	 * and promote when the count threshold is reached.  Note that bus
465 	 * mastering activity may prevent promotions.
466 	 * Do not promote above max_cstate.
467 	 */
468 	if (cx->promotion.state &&
469 	    ((cx->promotion.state - pr->power.states) <= max_cstate)) {
470 		if (sleep_ticks > cx->promotion.threshold.ticks &&
471 		  cx->promotion.state->latency <= system_latency_constraint()) {
472 			cx->promotion.count++;
473 			cx->demotion.count = 0;
474 			if (cx->promotion.count >=
475 			    cx->promotion.threshold.count) {
476 				if (pr->flags.bm_check) {
477 					if (!
478 					    (pr->power.bm_activity & cx->
479 					     promotion.threshold.bm)) {
480 						next_state =
481 						    cx->promotion.state;
482 						goto end;
483 					}
484 				} else {
485 					next_state = cx->promotion.state;
486 					goto end;
487 				}
488 			}
489 		}
490 	}
491 
492 	/*
493 	 * Demotion?
494 	 * ---------
495 	 * Track the number of shorts (time asleep is less than time threshold)
496 	 * and demote when the usage threshold is reached.
497 	 */
498 	if (cx->demotion.state) {
499 		if (sleep_ticks < cx->demotion.threshold.ticks) {
500 			cx->demotion.count++;
501 			cx->promotion.count = 0;
502 			if (cx->demotion.count >= cx->demotion.threshold.count) {
503 				next_state = cx->demotion.state;
504 				goto end;
505 			}
506 		}
507 	}
508 
509       end:
510 	/*
511 	 * Demote if current state exceeds max_cstate
512 	 * or if the latency of the current state is unacceptable
513 	 */
514 	if ((pr->power.state - pr->power.states) > max_cstate ||
515 		pr->power.state->latency > system_latency_constraint()) {
516 		if (cx->demotion.state)
517 			next_state = cx->demotion.state;
518 	}
519 
520 	/*
521 	 * New Cx State?
522 	 * -------------
523 	 * If we're going to start using a new Cx state we must clean up
524 	 * from the previous and prepare to use the new.
525 	 */
526 	if (next_state != pr->power.state)
527 		acpi_processor_power_activate(pr, next_state);
528 }
529 
530 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
531 {
532 	unsigned int i;
533 	unsigned int state_is_set = 0;
534 	struct acpi_processor_cx *lower = NULL;
535 	struct acpi_processor_cx *higher = NULL;
536 	struct acpi_processor_cx *cx;
537 
538 
539 	if (!pr)
540 		return -EINVAL;
541 
542 	/*
543 	 * This function sets the default Cx state policy (OS idle handler).
544 	 * Our scheme is to promote quickly to C2 but more conservatively
545 	 * to C3.  We're favoring C2  for its characteristics of low latency
546 	 * (quick response), good power savings, and ability to allow bus
547 	 * mastering activity.  Note that the Cx state policy is completely
548 	 * customizable and can be altered dynamically.
549 	 */
550 
551 	/* startup state */
552 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
553 		cx = &pr->power.states[i];
554 		if (!cx->valid)
555 			continue;
556 
557 		if (!state_is_set)
558 			pr->power.state = cx;
559 		state_is_set++;
560 		break;
561 	}
562 
563 	if (!state_is_set)
564 		return -ENODEV;
565 
566 	/* demotion */
567 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
568 		cx = &pr->power.states[i];
569 		if (!cx->valid)
570 			continue;
571 
572 		if (lower) {
573 			cx->demotion.state = lower;
574 			cx->demotion.threshold.ticks = cx->latency_ticks;
575 			cx->demotion.threshold.count = 1;
576 			if (cx->type == ACPI_STATE_C3)
577 				cx->demotion.threshold.bm = bm_history;
578 		}
579 
580 		lower = cx;
581 	}
582 
583 	/* promotion */
584 	for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
585 		cx = &pr->power.states[i];
586 		if (!cx->valid)
587 			continue;
588 
589 		if (higher) {
590 			cx->promotion.state = higher;
591 			cx->promotion.threshold.ticks = cx->latency_ticks;
592 			if (cx->type >= ACPI_STATE_C2)
593 				cx->promotion.threshold.count = 4;
594 			else
595 				cx->promotion.threshold.count = 10;
596 			if (higher->type == ACPI_STATE_C3)
597 				cx->promotion.threshold.bm = bm_history;
598 		}
599 
600 		higher = cx;
601 	}
602 
603 	return 0;
604 }
605 
606 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
607 {
608 
609 	if (!pr)
610 		return -EINVAL;
611 
612 	if (!pr->pblk)
613 		return -ENODEV;
614 
615 	/* if info is obtained from pblk/fadt, type equals state */
616 	pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
617 	pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
618 
619 #ifndef CONFIG_HOTPLUG_CPU
620 	/*
621 	 * Check for P_LVL2_UP flag before entering C2 and above on
622 	 * an SMP system.
623 	 */
624 	if ((num_online_cpus() > 1) &&
625 	    !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
626 		return -ENODEV;
627 #endif
628 
629 	/* determine C2 and C3 address from pblk */
630 	pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
631 	pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
632 
633 	/* determine latencies from FADT */
634 	pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
635 	pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
636 
637 	ACPI_DEBUG_PRINT((ACPI_DB_INFO,
638 			  "lvl2[0x%08x] lvl3[0x%08x]\n",
639 			  pr->power.states[ACPI_STATE_C2].address,
640 			  pr->power.states[ACPI_STATE_C3].address));
641 
642 	return 0;
643 }
644 
645 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
646 {
647 	if (!pr->power.states[ACPI_STATE_C1].valid) {
648 		/* set the first C-State to C1 */
649 		/* all processors need to support C1 */
650 		pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
651 		pr->power.states[ACPI_STATE_C1].valid = 1;
652 	}
653 	/* the C0 state only exists as a filler in our array */
654 	pr->power.states[ACPI_STATE_C0].valid = 1;
655 	return 0;
656 }
657 
658 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
659 {
660 	acpi_status status = 0;
661 	acpi_integer count;
662 	int current_count;
663 	int i;
664 	struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
665 	union acpi_object *cst;
666 
667 
668 	if (nocst)
669 		return -ENODEV;
670 
671 	current_count = 0;
672 
673 	status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
674 	if (ACPI_FAILURE(status)) {
675 		ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
676 		return -ENODEV;
677 	}
678 
679 	cst = buffer.pointer;
680 
681 	/* There must be at least 2 elements */
682 	if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
683 		printk(KERN_ERR PREFIX "not enough elements in _CST\n");
684 		status = -EFAULT;
685 		goto end;
686 	}
687 
688 	count = cst->package.elements[0].integer.value;
689 
690 	/* Validate number of power states. */
691 	if (count < 1 || count != cst->package.count - 1) {
692 		printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
693 		status = -EFAULT;
694 		goto end;
695 	}
696 
697 	/* Tell driver that at least _CST is supported. */
698 	pr->flags.has_cst = 1;
699 
700 	for (i = 1; i <= count; i++) {
701 		union acpi_object *element;
702 		union acpi_object *obj;
703 		struct acpi_power_register *reg;
704 		struct acpi_processor_cx cx;
705 
706 		memset(&cx, 0, sizeof(cx));
707 
708 		element = &(cst->package.elements[i]);
709 		if (element->type != ACPI_TYPE_PACKAGE)
710 			continue;
711 
712 		if (element->package.count != 4)
713 			continue;
714 
715 		obj = &(element->package.elements[0]);
716 
717 		if (obj->type != ACPI_TYPE_BUFFER)
718 			continue;
719 
720 		reg = (struct acpi_power_register *)obj->buffer.pointer;
721 
722 		if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
723 		    (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
724 			continue;
725 
726 		/* There should be an easy way to extract an integer... */
727 		obj = &(element->package.elements[1]);
728 		if (obj->type != ACPI_TYPE_INTEGER)
729 			continue;
730 
731 		cx.type = obj->integer.value;
732 		/*
733 		 * Some buggy BIOSes won't list C1 in _CST -
734 		 * Let acpi_processor_get_power_info_default() handle them later
735 		 */
736 		if (i == 1 && cx.type != ACPI_STATE_C1)
737 			current_count++;
738 
739 		cx.address = reg->address;
740 		cx.index = current_count + 1;
741 
742 		cx.space_id = ACPI_CSTATE_SYSTEMIO;
743 		if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
744 			if (acpi_processor_ffh_cstate_probe
745 					(pr->id, &cx, reg) == 0) {
746 				cx.space_id = ACPI_CSTATE_FFH;
747 			} else if (cx.type != ACPI_STATE_C1) {
748 				/*
749 				 * C1 is a special case where FIXED_HARDWARE
750 				 * can be handled in non-MWAIT way as well.
751 				 * In that case, save this _CST entry info.
752 				 * That is, we retain space_id of SYSTEM_IO for
753 				 * halt based C1.
754 				 * Otherwise, ignore this info and continue.
755 				 */
756 				continue;
757 			}
758 		}
759 
760 		obj = &(element->package.elements[2]);
761 		if (obj->type != ACPI_TYPE_INTEGER)
762 			continue;
763 
764 		cx.latency = obj->integer.value;
765 
766 		obj = &(element->package.elements[3]);
767 		if (obj->type != ACPI_TYPE_INTEGER)
768 			continue;
769 
770 		cx.power = obj->integer.value;
771 
772 		current_count++;
773 		memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
774 
775 		/*
776 		 * We support total ACPI_PROCESSOR_MAX_POWER - 1
777 		 * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
778 		 */
779 		if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
780 			printk(KERN_WARNING
781 			       "Limiting number of power states to max (%d)\n",
782 			       ACPI_PROCESSOR_MAX_POWER);
783 			printk(KERN_WARNING
784 			       "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
785 			break;
786 		}
787 	}
788 
789 	ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
790 			  current_count));
791 
792 	/* Validate number of power states discovered */
793 	if (current_count < 2)
794 		status = -EFAULT;
795 
796       end:
797 	kfree(buffer.pointer);
798 
799 	return status;
800 }
801 
802 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
803 {
804 
805 	if (!cx->address)
806 		return;
807 
808 	/*
809 	 * C2 latency must be less than or equal to 100
810 	 * microseconds.
811 	 */
812 	else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
813 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
814 				  "latency too large [%d]\n", cx->latency));
815 		return;
816 	}
817 
818 	/*
819 	 * Otherwise we've met all of our C2 requirements.
820 	 * Normalize the C2 latency to expidite policy
821 	 */
822 	cx->valid = 1;
823 	cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
824 
825 	return;
826 }
827 
828 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
829 					   struct acpi_processor_cx *cx)
830 {
831 	static int bm_check_flag;
832 
833 
834 	if (!cx->address)
835 		return;
836 
837 	/*
838 	 * C3 latency must be less than or equal to 1000
839 	 * microseconds.
840 	 */
841 	else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
842 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
843 				  "latency too large [%d]\n", cx->latency));
844 		return;
845 	}
846 
847 	/*
848 	 * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
849 	 * DMA transfers are used by any ISA device to avoid livelock.
850 	 * Note that we could disable Type-F DMA (as recommended by
851 	 * the erratum), but this is known to disrupt certain ISA
852 	 * devices thus we take the conservative approach.
853 	 */
854 	else if (errata.piix4.fdma) {
855 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
856 				  "C3 not supported on PIIX4 with Type-F DMA\n"));
857 		return;
858 	}
859 
860 	/* All the logic here assumes flags.bm_check is same across all CPUs */
861 	if (!bm_check_flag) {
862 		/* Determine whether bm_check is needed based on CPU  */
863 		acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
864 		bm_check_flag = pr->flags.bm_check;
865 	} else {
866 		pr->flags.bm_check = bm_check_flag;
867 	}
868 
869 	if (pr->flags.bm_check) {
870 		/* bus mastering control is necessary */
871 		if (!pr->flags.bm_control) {
872 			ACPI_DEBUG_PRINT((ACPI_DB_INFO,
873 					  "C3 support requires bus mastering control\n"));
874 			return;
875 		}
876 	} else {
877 		/*
878 		 * WBINVD should be set in fadt, for C3 state to be
879 		 * supported on when bm_check is not required.
880 		 */
881 		if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
882 			ACPI_DEBUG_PRINT((ACPI_DB_INFO,
883 					  "Cache invalidation should work properly"
884 					  " for C3 to be enabled on SMP systems\n"));
885 			return;
886 		}
887 		acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
888 	}
889 
890 	/*
891 	 * Otherwise we've met all of our C3 requirements.
892 	 * Normalize the C3 latency to expidite policy.  Enable
893 	 * checking of bus mastering status (bm_check) so we can
894 	 * use this in our C3 policy
895 	 */
896 	cx->valid = 1;
897 	cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
898 
899 	return;
900 }
901 
902 static int acpi_processor_power_verify(struct acpi_processor *pr)
903 {
904 	unsigned int i;
905 	unsigned int working = 0;
906 
907 #ifdef ARCH_APICTIMER_STOPS_ON_C3
908 	int timer_broadcast = 0;
909 	cpumask_t mask = cpumask_of_cpu(pr->id);
910 	on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
911 #endif
912 
913 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
914 		struct acpi_processor_cx *cx = &pr->power.states[i];
915 
916 		switch (cx->type) {
917 		case ACPI_STATE_C1:
918 			cx->valid = 1;
919 			break;
920 
921 		case ACPI_STATE_C2:
922 			acpi_processor_power_verify_c2(cx);
923 #ifdef ARCH_APICTIMER_STOPS_ON_C3
924 			/* Some AMD systems fake C3 as C2, but still
925 			   have timer troubles */
926 			if (cx->valid &&
927 				boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
928 				timer_broadcast++;
929 #endif
930 			break;
931 
932 		case ACPI_STATE_C3:
933 			acpi_processor_power_verify_c3(pr, cx);
934 #ifdef ARCH_APICTIMER_STOPS_ON_C3
935 			if (cx->valid)
936 				timer_broadcast++;
937 #endif
938 			break;
939 		}
940 
941 		if (cx->valid)
942 			working++;
943 	}
944 
945 #ifdef ARCH_APICTIMER_STOPS_ON_C3
946 	if (timer_broadcast)
947 		on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
948 #endif
949 
950 	return (working);
951 }
952 
953 static int acpi_processor_get_power_info(struct acpi_processor *pr)
954 {
955 	unsigned int i;
956 	int result;
957 
958 
959 	/* NOTE: the idle thread may not be running while calling
960 	 * this function */
961 
962 	/* Zero initialize all the C-states info. */
963 	memset(pr->power.states, 0, sizeof(pr->power.states));
964 
965 	result = acpi_processor_get_power_info_cst(pr);
966 	if (result == -ENODEV)
967 		result = acpi_processor_get_power_info_fadt(pr);
968 
969 	if (result)
970 		return result;
971 
972 	acpi_processor_get_power_info_default(pr);
973 
974 	pr->power.count = acpi_processor_power_verify(pr);
975 
976 	/*
977 	 * Set Default Policy
978 	 * ------------------
979 	 * Now that we know which states are supported, set the default
980 	 * policy.  Note that this policy can be changed dynamically
981 	 * (e.g. encourage deeper sleeps to conserve battery life when
982 	 * not on AC).
983 	 */
984 	result = acpi_processor_set_power_policy(pr);
985 	if (result)
986 		return result;
987 
988 	/*
989 	 * if one state of type C2 or C3 is available, mark this
990 	 * CPU as being "idle manageable"
991 	 */
992 	for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
993 		if (pr->power.states[i].valid) {
994 			pr->power.count = i;
995 			if (pr->power.states[i].type >= ACPI_STATE_C2)
996 				pr->flags.power = 1;
997 		}
998 	}
999 
1000 	return 0;
1001 }
1002 
1003 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1004 {
1005 	int result = 0;
1006 
1007 
1008 	if (!pr)
1009 		return -EINVAL;
1010 
1011 	if (nocst) {
1012 		return -ENODEV;
1013 	}
1014 
1015 	if (!pr->flags.power_setup_done)
1016 		return -ENODEV;
1017 
1018 	/* Fall back to the default idle loop */
1019 	pm_idle = pm_idle_save;
1020 	synchronize_sched();	/* Relies on interrupts forcing exit from idle. */
1021 
1022 	pr->flags.power = 0;
1023 	result = acpi_processor_get_power_info(pr);
1024 	if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1025 		pm_idle = acpi_processor_idle;
1026 
1027 	return result;
1028 }
1029 
1030 /* proc interface */
1031 
1032 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1033 {
1034 	struct acpi_processor *pr = seq->private;
1035 	unsigned int i;
1036 
1037 
1038 	if (!pr)
1039 		goto end;
1040 
1041 	seq_printf(seq, "active state:            C%zd\n"
1042 		   "max_cstate:              C%d\n"
1043 		   "bus master activity:     %08x\n"
1044 		   "maximum allowed latency: %d usec\n",
1045 		   pr->power.state ? pr->power.state - pr->power.states : 0,
1046 		   max_cstate, (unsigned)pr->power.bm_activity,
1047 		   system_latency_constraint());
1048 
1049 	seq_puts(seq, "states:\n");
1050 
1051 	for (i = 1; i <= pr->power.count; i++) {
1052 		seq_printf(seq, "   %cC%d:                  ",
1053 			   (&pr->power.states[i] ==
1054 			    pr->power.state ? '*' : ' '), i);
1055 
1056 		if (!pr->power.states[i].valid) {
1057 			seq_puts(seq, "<not supported>\n");
1058 			continue;
1059 		}
1060 
1061 		switch (pr->power.states[i].type) {
1062 		case ACPI_STATE_C1:
1063 			seq_printf(seq, "type[C1] ");
1064 			break;
1065 		case ACPI_STATE_C2:
1066 			seq_printf(seq, "type[C2] ");
1067 			break;
1068 		case ACPI_STATE_C3:
1069 			seq_printf(seq, "type[C3] ");
1070 			break;
1071 		default:
1072 			seq_printf(seq, "type[--] ");
1073 			break;
1074 		}
1075 
1076 		if (pr->power.states[i].promotion.state)
1077 			seq_printf(seq, "promotion[C%zd] ",
1078 				   (pr->power.states[i].promotion.state -
1079 				    pr->power.states));
1080 		else
1081 			seq_puts(seq, "promotion[--] ");
1082 
1083 		if (pr->power.states[i].demotion.state)
1084 			seq_printf(seq, "demotion[C%zd] ",
1085 				   (pr->power.states[i].demotion.state -
1086 				    pr->power.states));
1087 		else
1088 			seq_puts(seq, "demotion[--] ");
1089 
1090 		seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1091 			   pr->power.states[i].latency,
1092 			   pr->power.states[i].usage,
1093 			   (unsigned long long)pr->power.states[i].time);
1094 	}
1095 
1096       end:
1097 	return 0;
1098 }
1099 
1100 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1101 {
1102 	return single_open(file, acpi_processor_power_seq_show,
1103 			   PDE(inode)->data);
1104 }
1105 
1106 static const struct file_operations acpi_processor_power_fops = {
1107 	.open = acpi_processor_power_open_fs,
1108 	.read = seq_read,
1109 	.llseek = seq_lseek,
1110 	.release = single_release,
1111 };
1112 
1113 #ifdef CONFIG_SMP
1114 static void smp_callback(void *v)
1115 {
1116 	/* we already woke the CPU up, nothing more to do */
1117 }
1118 
1119 /*
1120  * This function gets called when a part of the kernel has a new latency
1121  * requirement.  This means we need to get all processors out of their C-state,
1122  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1123  * wakes them all right up.
1124  */
1125 static int acpi_processor_latency_notify(struct notifier_block *b,
1126 		unsigned long l, void *v)
1127 {
1128 	smp_call_function(smp_callback, NULL, 0, 1);
1129 	return NOTIFY_OK;
1130 }
1131 
1132 static struct notifier_block acpi_processor_latency_notifier = {
1133 	.notifier_call = acpi_processor_latency_notify,
1134 };
1135 #endif
1136 
1137 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1138 			      struct acpi_device *device)
1139 {
1140 	acpi_status status = 0;
1141 	static int first_run;
1142 	struct proc_dir_entry *entry = NULL;
1143 	unsigned int i;
1144 
1145 
1146 	if (!first_run) {
1147 		dmi_check_system(processor_power_dmi_table);
1148 		if (max_cstate < ACPI_C_STATES_MAX)
1149 			printk(KERN_NOTICE
1150 			       "ACPI: processor limited to max C-state %d\n",
1151 			       max_cstate);
1152 		first_run++;
1153 #ifdef CONFIG_SMP
1154 		register_latency_notifier(&acpi_processor_latency_notifier);
1155 #endif
1156 	}
1157 
1158 	if (!pr)
1159 		return -EINVAL;
1160 
1161 	if (acpi_gbl_FADT.cst_control && !nocst) {
1162 		status =
1163 		    acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1164 		if (ACPI_FAILURE(status)) {
1165 			ACPI_EXCEPTION((AE_INFO, status,
1166 					"Notifying BIOS of _CST ability failed"));
1167 		}
1168 	}
1169 
1170 	acpi_processor_get_power_info(pr);
1171 
1172 	/*
1173 	 * Install the idle handler if processor power management is supported.
1174 	 * Note that we use previously set idle handler will be used on
1175 	 * platforms that only support C1.
1176 	 */
1177 	if ((pr->flags.power) && (!boot_option_idle_override)) {
1178 		printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1179 		for (i = 1; i <= pr->power.count; i++)
1180 			if (pr->power.states[i].valid)
1181 				printk(" C%d[C%d]", i,
1182 				       pr->power.states[i].type);
1183 		printk(")\n");
1184 
1185 		if (pr->id == 0) {
1186 			pm_idle_save = pm_idle;
1187 			pm_idle = acpi_processor_idle;
1188 		}
1189 	}
1190 
1191 	/* 'power' [R] */
1192 	entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1193 				  S_IRUGO, acpi_device_dir(device));
1194 	if (!entry)
1195 		return -EIO;
1196 	else {
1197 		entry->proc_fops = &acpi_processor_power_fops;
1198 		entry->data = acpi_driver_data(device);
1199 		entry->owner = THIS_MODULE;
1200 	}
1201 
1202 	pr->flags.power_setup_done = 1;
1203 
1204 	return 0;
1205 }
1206 
1207 int acpi_processor_power_exit(struct acpi_processor *pr,
1208 			      struct acpi_device *device)
1209 {
1210 
1211 	pr->flags.power_setup_done = 0;
1212 
1213 	if (acpi_device_dir(device))
1214 		remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1215 				  acpi_device_dir(device));
1216 
1217 	/* Unregister the idle handler when processor #0 is removed. */
1218 	if (pr->id == 0) {
1219 		pm_idle = pm_idle_save;
1220 
1221 		/*
1222 		 * We are about to unload the current idle thread pm callback
1223 		 * (pm_idle), Wait for all processors to update cached/local
1224 		 * copies of pm_idle before proceeding.
1225 		 */
1226 		cpu_idle_wait();
1227 #ifdef CONFIG_SMP
1228 		unregister_latency_notifier(&acpi_processor_latency_notifier);
1229 #endif
1230 	}
1231 
1232 	return 0;
1233 }
1234