cpuidle: Single/Global registration of idle states
[linux-2.6.git] / drivers / acpi / processor_idle.c
1 /*
2  * processor_idle - idle state submodule to the ACPI processor driver
3  *
4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
6  *  Copyright (C) 2004, 2005 Dominik Brodowski <linux@brodo.de>
7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
8  *                      - Added processor hotplug support
9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
10  *                      - Added support for C3 on SMP
11  *
12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
13  *
14  *  This program is free software; you can redistribute it and/or modify
15  *  it under the terms of the GNU General Public License as published by
16  *  the Free Software Foundation; either version 2 of the License, or (at
17  *  your option) any later version.
18  *
19  *  This program is distributed in the hope that it will be useful, but
20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
22  *  General Public License for more details.
23  *
24  *  You should have received a copy of the GNU General Public License along
25  *  with this program; if not, write to the Free Software Foundation, Inc.,
26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
27  *
28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
29  */
30
31 #include <linux/kernel.h>
32 #include <linux/module.h>
33 #include <linux/init.h>
34 #include <linux/cpufreq.h>
35 #include <linux/slab.h>
36 #include <linux/acpi.h>
37 #include <linux/dmi.h>
38 #include <linux/moduleparam.h>
39 #include <linux/sched.h>        /* need_resched() */
40 #include <linux/pm_qos_params.h>
41 #include <linux/clockchips.h>
42 #include <linux/cpuidle.h>
43 #include <linux/irqflags.h>
44
45 /*
46  * Include the apic definitions for x86 to have the APIC timer related defines
47  * available also for UP (on SMP it gets magically included via linux/smp.h).
48  * asm/acpi.h is not an option, as it would require more include magic. Also
49  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
50  */
51 #ifdef CONFIG_X86
52 #include <asm/apic.h>
53 #endif
54
55 #include <asm/io.h>
56 #include <asm/uaccess.h>
57
58 #include <acpi/acpi_bus.h>
59 #include <acpi/processor.h>
60 #include <asm/processor.h>
61
62 #define PREFIX "ACPI: "
63
64 #define ACPI_PROCESSOR_CLASS            "processor"
65 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
66 ACPI_MODULE_NAME("processor_idle");
67 #define PM_TIMER_TICK_NS                (1000000000ULL/PM_TIMER_FREQUENCY)
68 #define C2_OVERHEAD                     1       /* 1us */
69 #define C3_OVERHEAD                     1       /* 1us */
70 #define PM_TIMER_TICKS_TO_US(p)         (((p) * 1000)/(PM_TIMER_FREQUENCY/1000))
71
72 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
73 module_param(max_cstate, uint, 0000);
74 static unsigned int nocst __read_mostly;
75 module_param(nocst, uint, 0000);
76 static int bm_check_disable __read_mostly;
77 module_param(bm_check_disable, uint, 0000);
78
79 static unsigned int latency_factor __read_mostly = 2;
80 module_param(latency_factor, uint, 0644);
81
82 static int disabled_by_idle_boot_param(void)
83 {
84         return boot_option_idle_override == IDLE_POLL ||
85                 boot_option_idle_override == IDLE_FORCE_MWAIT ||
86                 boot_option_idle_override == IDLE_HALT;
87 }
88
89 /*
90  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
91  * For now disable this. Probably a bug somewhere else.
92  *
93  * To skip this limit, boot/load with a large max_cstate limit.
94  */
95 static int set_max_cstate(const struct dmi_system_id *id)
96 {
97         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
98                 return 0;
99
100         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
101                " Override with \"processor.max_cstate=%d\"\n", id->ident,
102                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
103
104         max_cstate = (long)id->driver_data;
105
106         return 0;
107 }
108
109 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
110    callers to only run once -AK */
111 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
112         { set_max_cstate, "Clevo 5600D", {
113           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
114           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
115          (void *)2},
116         { set_max_cstate, "Pavilion zv5000", {
117           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
118           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
119          (void *)1},
120         { set_max_cstate, "Asus L8400B", {
121           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
122           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
123          (void *)1},
124         {},
125 };
126
127
128 /*
129  * Callers should disable interrupts before the call and enable
130  * interrupts after return.
131  */
132 static void acpi_safe_halt(void)
133 {
134         current_thread_info()->status &= ~TS_POLLING;
135         /*
136          * TS_POLLING-cleared state must be visible before we
137          * test NEED_RESCHED:
138          */
139         smp_mb();
140         if (!need_resched()) {
141                 safe_halt();
142                 local_irq_disable();
143         }
144         current_thread_info()->status |= TS_POLLING;
145 }
146
147 #ifdef ARCH_APICTIMER_STOPS_ON_C3
148
149 /*
150  * Some BIOS implementations switch to C3 in the published C2 state.
151  * This seems to be a common problem on AMD boxen, but other vendors
152  * are affected too. We pick the most conservative approach: we assume
153  * that the local APIC stops in both C2 and C3.
154  */
155 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
156                                    struct acpi_processor_cx *cx)
157 {
158         struct acpi_processor_power *pwr = &pr->power;
159         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
160
161         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
162                 return;
163
164         if (amd_e400_c1e_detected)
165                 type = ACPI_STATE_C1;
166
167         /*
168          * Check, if one of the previous states already marked the lapic
169          * unstable
170          */
171         if (pwr->timer_broadcast_on_state < state)
172                 return;
173
174         if (cx->type >= type)
175                 pr->power.timer_broadcast_on_state = state;
176 }
177
178 static void __lapic_timer_propagate_broadcast(void *arg)
179 {
180         struct acpi_processor *pr = (struct acpi_processor *) arg;
181         unsigned long reason;
182
183         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
184                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
185
186         clockevents_notify(reason, &pr->id);
187 }
188
189 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
190 {
191         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
192                                  (void *)pr, 1);
193 }
194
195 /* Power(C) State timer broadcast control */
196 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
197                                        struct acpi_processor_cx *cx,
198                                        int broadcast)
199 {
200         int state = cx - pr->power.states;
201
202         if (state >= pr->power.timer_broadcast_on_state) {
203                 unsigned long reason;
204
205                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
206                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
207                 clockevents_notify(reason, &pr->id);
208         }
209 }
210
211 #else
212
213 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
214                                    struct acpi_processor_cx *cstate) { }
215 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
216 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
217                                        struct acpi_processor_cx *cx,
218                                        int broadcast)
219 {
220 }
221
222 #endif
223
224 /*
225  * Suspend / resume control
226  */
227 static int acpi_idle_suspend;
228 static u32 saved_bm_rld;
229
230 static void acpi_idle_bm_rld_save(void)
231 {
232         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
233 }
234 static void acpi_idle_bm_rld_restore(void)
235 {
236         u32 resumed_bm_rld;
237
238         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
239
240         if (resumed_bm_rld != saved_bm_rld)
241                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
242 }
243
244 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
245 {
246         if (acpi_idle_suspend == 1)
247                 return 0;
248
249         acpi_idle_bm_rld_save();
250         acpi_idle_suspend = 1;
251         return 0;
252 }
253
254 int acpi_processor_resume(struct acpi_device * device)
255 {
256         if (acpi_idle_suspend == 0)
257                 return 0;
258
259         acpi_idle_bm_rld_restore();
260         acpi_idle_suspend = 0;
261         return 0;
262 }
263
264 #if defined(CONFIG_X86)
265 static void tsc_check_state(int state)
266 {
267         switch (boot_cpu_data.x86_vendor) {
268         case X86_VENDOR_AMD:
269         case X86_VENDOR_INTEL:
270                 /*
271                  * AMD Fam10h TSC will tick in all
272                  * C/P/S0/S1 states when this bit is set.
273                  */
274                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
275                         return;
276
277                 /*FALL THROUGH*/
278         default:
279                 /* TSC could halt in idle, so notify users */
280                 if (state > ACPI_STATE_C1)
281                         mark_tsc_unstable("TSC halts in idle");
282         }
283 }
284 #else
285 static void tsc_check_state(int state) { return; }
286 #endif
287
288 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
289 {
290
291         if (!pr)
292                 return -EINVAL;
293
294         if (!pr->pblk)
295                 return -ENODEV;
296
297         /* if info is obtained from pblk/fadt, type equals state */
298         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
299         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
300
301 #ifndef CONFIG_HOTPLUG_CPU
302         /*
303          * Check for P_LVL2_UP flag before entering C2 and above on
304          * an SMP system.
305          */
306         if ((num_online_cpus() > 1) &&
307             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
308                 return -ENODEV;
309 #endif
310
311         /* determine C2 and C3 address from pblk */
312         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
313         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
314
315         /* determine latencies from FADT */
316         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
317         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
318
319         /*
320          * FADT specified C2 latency must be less than or equal to
321          * 100 microseconds.
322          */
323         if (acpi_gbl_FADT.C2latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
324                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
325                         "C2 latency too large [%d]\n", acpi_gbl_FADT.C2latency));
326                 /* invalidate C2 */
327                 pr->power.states[ACPI_STATE_C2].address = 0;
328         }
329
330         /*
331          * FADT supplied C3 latency must be less than or equal to
332          * 1000 microseconds.
333          */
334         if (acpi_gbl_FADT.C3latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
335                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
336                         "C3 latency too large [%d]\n", acpi_gbl_FADT.C3latency));
337                 /* invalidate C3 */
338                 pr->power.states[ACPI_STATE_C3].address = 0;
339         }
340
341         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
342                           "lvl2[0x%08x] lvl3[0x%08x]\n",
343                           pr->power.states[ACPI_STATE_C2].address,
344                           pr->power.states[ACPI_STATE_C3].address));
345
346         return 0;
347 }
348
349 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
350 {
351         if (!pr->power.states[ACPI_STATE_C1].valid) {
352                 /* set the first C-State to C1 */
353                 /* all processors need to support C1 */
354                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
355                 pr->power.states[ACPI_STATE_C1].valid = 1;
356                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
357         }
358         /* the C0 state only exists as a filler in our array */
359         pr->power.states[ACPI_STATE_C0].valid = 1;
360         return 0;
361 }
362
363 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
364 {
365         acpi_status status = 0;
366         u64 count;
367         int current_count;
368         int i;
369         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
370         union acpi_object *cst;
371
372
373         if (nocst)
374                 return -ENODEV;
375
376         current_count = 0;
377
378         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
379         if (ACPI_FAILURE(status)) {
380                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
381                 return -ENODEV;
382         }
383
384         cst = buffer.pointer;
385
386         /* There must be at least 2 elements */
387         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
388                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
389                 status = -EFAULT;
390                 goto end;
391         }
392
393         count = cst->package.elements[0].integer.value;
394
395         /* Validate number of power states. */
396         if (count < 1 || count != cst->package.count - 1) {
397                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
398                 status = -EFAULT;
399                 goto end;
400         }
401
402         /* Tell driver that at least _CST is supported. */
403         pr->flags.has_cst = 1;
404
405         for (i = 1; i <= count; i++) {
406                 union acpi_object *element;
407                 union acpi_object *obj;
408                 struct acpi_power_register *reg;
409                 struct acpi_processor_cx cx;
410
411                 memset(&cx, 0, sizeof(cx));
412
413                 element = &(cst->package.elements[i]);
414                 if (element->type != ACPI_TYPE_PACKAGE)
415                         continue;
416
417                 if (element->package.count != 4)
418                         continue;
419
420                 obj = &(element->package.elements[0]);
421
422                 if (obj->type != ACPI_TYPE_BUFFER)
423                         continue;
424
425                 reg = (struct acpi_power_register *)obj->buffer.pointer;
426
427                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
428                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
429                         continue;
430
431                 /* There should be an easy way to extract an integer... */
432                 obj = &(element->package.elements[1]);
433                 if (obj->type != ACPI_TYPE_INTEGER)
434                         continue;
435
436                 cx.type = obj->integer.value;
437                 /*
438                  * Some buggy BIOSes won't list C1 in _CST -
439                  * Let acpi_processor_get_power_info_default() handle them later
440                  */
441                 if (i == 1 && cx.type != ACPI_STATE_C1)
442                         current_count++;
443
444                 cx.address = reg->address;
445                 cx.index = current_count + 1;
446
447                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
448                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
449                         if (acpi_processor_ffh_cstate_probe
450                                         (pr->id, &cx, reg) == 0) {
451                                 cx.entry_method = ACPI_CSTATE_FFH;
452                         } else if (cx.type == ACPI_STATE_C1) {
453                                 /*
454                                  * C1 is a special case where FIXED_HARDWARE
455                                  * can be handled in non-MWAIT way as well.
456                                  * In that case, save this _CST entry info.
457                                  * Otherwise, ignore this info and continue.
458                                  */
459                                 cx.entry_method = ACPI_CSTATE_HALT;
460                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
461                         } else {
462                                 continue;
463                         }
464                         if (cx.type == ACPI_STATE_C1 &&
465                             (boot_option_idle_override == IDLE_NOMWAIT)) {
466                                 /*
467                                  * In most cases the C1 space_id obtained from
468                                  * _CST object is FIXED_HARDWARE access mode.
469                                  * But when the option of idle=halt is added,
470                                  * the entry_method type should be changed from
471                                  * CSTATE_FFH to CSTATE_HALT.
472                                  * When the option of idle=nomwait is added,
473                                  * the C1 entry_method type should be
474                                  * CSTATE_HALT.
475                                  */
476                                 cx.entry_method = ACPI_CSTATE_HALT;
477                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
478                         }
479                 } else {
480                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
481                                  cx.address);
482                 }
483
484                 if (cx.type == ACPI_STATE_C1) {
485                         cx.valid = 1;
486                 }
487
488                 obj = &(element->package.elements[2]);
489                 if (obj->type != ACPI_TYPE_INTEGER)
490                         continue;
491
492                 cx.latency = obj->integer.value;
493
494                 obj = &(element->package.elements[3]);
495                 if (obj->type != ACPI_TYPE_INTEGER)
496                         continue;
497
498                 cx.power = obj->integer.value;
499
500                 current_count++;
501                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
502
503                 /*
504                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
505                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
506                  */
507                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
508                         printk(KERN_WARNING
509                                "Limiting number of power states to max (%d)\n",
510                                ACPI_PROCESSOR_MAX_POWER);
511                         printk(KERN_WARNING
512                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
513                         break;
514                 }
515         }
516
517         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
518                           current_count));
519
520         /* Validate number of power states discovered */
521         if (current_count < 2)
522                 status = -EFAULT;
523
524       end:
525         kfree(buffer.pointer);
526
527         return status;
528 }
529
530 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
531                                            struct acpi_processor_cx *cx)
532 {
533         static int bm_check_flag = -1;
534         static int bm_control_flag = -1;
535
536
537         if (!cx->address)
538                 return;
539
540         /*
541          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
542          * DMA transfers are used by any ISA device to avoid livelock.
543          * Note that we could disable Type-F DMA (as recommended by
544          * the erratum), but this is known to disrupt certain ISA
545          * devices thus we take the conservative approach.
546          */
547         else if (errata.piix4.fdma) {
548                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
549                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
550                 return;
551         }
552
553         /* All the logic here assumes flags.bm_check is same across all CPUs */
554         if (bm_check_flag == -1) {
555                 /* Determine whether bm_check is needed based on CPU  */
556                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
557                 bm_check_flag = pr->flags.bm_check;
558                 bm_control_flag = pr->flags.bm_control;
559         } else {
560                 pr->flags.bm_check = bm_check_flag;
561                 pr->flags.bm_control = bm_control_flag;
562         }
563
564         if (pr->flags.bm_check) {
565                 if (!pr->flags.bm_control) {
566                         if (pr->flags.has_cst != 1) {
567                                 /* bus mastering control is necessary */
568                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
569                                         "C3 support requires BM control\n"));
570                                 return;
571                         } else {
572                                 /* Here we enter C3 without bus mastering */
573                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
574                                         "C3 support without BM control\n"));
575                         }
576                 }
577         } else {
578                 /*
579                  * WBINVD should be set in fadt, for C3 state to be
580                  * supported on when bm_check is not required.
581                  */
582                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
583                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
584                                           "Cache invalidation should work properly"
585                                           " for C3 to be enabled on SMP systems\n"));
586                         return;
587                 }
588         }
589
590         /*
591          * Otherwise we've met all of our C3 requirements.
592          * Normalize the C3 latency to expidite policy.  Enable
593          * checking of bus mastering status (bm_check) so we can
594          * use this in our C3 policy
595          */
596         cx->valid = 1;
597
598         cx->latency_ticks = cx->latency;
599         /*
600          * On older chipsets, BM_RLD needs to be set
601          * in order for Bus Master activity to wake the
602          * system from C3.  Newer chipsets handle DMA
603          * during C3 automatically and BM_RLD is a NOP.
604          * In either case, the proper way to
605          * handle BM_RLD is to set it and leave it set.
606          */
607         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
608
609         return;
610 }
611
612 static int acpi_processor_power_verify(struct acpi_processor *pr)
613 {
614         unsigned int i;
615         unsigned int working = 0;
616
617         pr->power.timer_broadcast_on_state = INT_MAX;
618
619         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
620                 struct acpi_processor_cx *cx = &pr->power.states[i];
621
622                 switch (cx->type) {
623                 case ACPI_STATE_C1:
624                         cx->valid = 1;
625                         break;
626
627                 case ACPI_STATE_C2:
628                         if (!cx->address)
629                                 break;
630                         cx->valid = 1; 
631                         cx->latency_ticks = cx->latency; /* Normalize latency */
632                         break;
633
634                 case ACPI_STATE_C3:
635                         acpi_processor_power_verify_c3(pr, cx);
636                         break;
637                 }
638                 if (!cx->valid)
639                         continue;
640
641                 lapic_timer_check_state(i, pr, cx);
642                 tsc_check_state(cx->type);
643                 working++;
644         }
645
646         lapic_timer_propagate_broadcast(pr);
647
648         return (working);
649 }
650
651 static int acpi_processor_get_power_info(struct acpi_processor *pr)
652 {
653         unsigned int i;
654         int result;
655
656
657         /* NOTE: the idle thread may not be running while calling
658          * this function */
659
660         /* Zero initialize all the C-states info. */
661         memset(pr->power.states, 0, sizeof(pr->power.states));
662
663         result = acpi_processor_get_power_info_cst(pr);
664         if (result == -ENODEV)
665                 result = acpi_processor_get_power_info_fadt(pr);
666
667         if (result)
668                 return result;
669
670         acpi_processor_get_power_info_default(pr);
671
672         pr->power.count = acpi_processor_power_verify(pr);
673
674         /*
675          * if one state of type C2 or C3 is available, mark this
676          * CPU as being "idle manageable"
677          */
678         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
679                 if (pr->power.states[i].valid) {
680                         pr->power.count = i;
681                         if (pr->power.states[i].type >= ACPI_STATE_C2)
682                                 pr->flags.power = 1;
683                 }
684         }
685
686         return 0;
687 }
688
689 /**
690  * acpi_idle_bm_check - checks if bus master activity was detected
691  */
692 static int acpi_idle_bm_check(void)
693 {
694         u32 bm_status = 0;
695
696         if (bm_check_disable)
697                 return 0;
698
699         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
700         if (bm_status)
701                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
702         /*
703          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
704          * the true state of bus mastering activity; forcing us to
705          * manually check the BMIDEA bit of each IDE channel.
706          */
707         else if (errata.piix4.bmisx) {
708                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
709                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
710                         bm_status = 1;
711         }
712         return bm_status;
713 }
714
715 /**
716  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
717  * @cx: cstate data
718  *
719  * Caller disables interrupt before call and enables interrupt after return.
720  */
721 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
722 {
723         /* Don't trace irqs off for idle */
724         stop_critical_timings();
725         if (cx->entry_method == ACPI_CSTATE_FFH) {
726                 /* Call into architectural FFH based C-state */
727                 acpi_processor_ffh_cstate_enter(cx);
728         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
729                 acpi_safe_halt();
730         } else {
731                 /* IO port based C-state */
732                 inb(cx->address);
733                 /* Dummy wait op - must do something useless after P_LVL2 read
734                    because chipsets cannot guarantee that STPCLK# signal
735                    gets asserted in time to freeze execution properly. */
736                 inl(acpi_gbl_FADT.xpm_timer_block.address);
737         }
738         start_critical_timings();
739 }
740
741 /**
742  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
743  * @dev: the target CPU
744  * @drv: cpuidle driver containing cpuidle state info
745  * @index: index of target state
746  *
747  * This is equivalent to the HALT instruction.
748  */
749 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
750                 struct cpuidle_driver *drv, int index)
751 {
752         ktime_t  kt1, kt2;
753         s64 idle_time;
754         struct acpi_processor *pr;
755         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
756         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
757
758         pr = __this_cpu_read(processors);
759         dev->last_residency = 0;
760
761         if (unlikely(!pr))
762                 return -EINVAL;
763
764         local_irq_disable();
765
766         /* Do not access any ACPI IO ports in suspend path */
767         if (acpi_idle_suspend) {
768                 local_irq_enable();
769                 cpu_relax();
770                 return -EINVAL;
771         }
772
773         lapic_timer_state_broadcast(pr, cx, 1);
774         kt1 = ktime_get_real();
775         acpi_idle_do_entry(cx);
776         kt2 = ktime_get_real();
777         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
778
779         /* Update device last_residency*/
780         dev->last_residency = (int)idle_time;
781
782         local_irq_enable();
783         cx->usage++;
784         lapic_timer_state_broadcast(pr, cx, 0);
785
786         return index;
787 }
788
789 /**
790  * acpi_idle_enter_simple - enters an ACPI state without BM handling
791  * @dev: the target CPU
792  * @drv: cpuidle driver with cpuidle state information
793  * @index: the index of suggested state
794  */
795 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
796                 struct cpuidle_driver *drv, int index)
797 {
798         struct acpi_processor *pr;
799         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
800         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
801         ktime_t  kt1, kt2;
802         s64 idle_time_ns;
803         s64 idle_time;
804
805         pr = __this_cpu_read(processors);
806         dev->last_residency = 0;
807
808         if (unlikely(!pr))
809                 return -EINVAL;
810
811         local_irq_disable();
812
813         if (acpi_idle_suspend) {
814                 local_irq_enable();
815                 cpu_relax();
816                 return -EINVAL;
817         }
818
819
820         if (cx->entry_method != ACPI_CSTATE_FFH) {
821                 current_thread_info()->status &= ~TS_POLLING;
822                 /*
823                  * TS_POLLING-cleared state must be visible before we test
824                  * NEED_RESCHED:
825                  */
826                 smp_mb();
827
828                 if (unlikely(need_resched())) {
829                         current_thread_info()->status |= TS_POLLING;
830                         local_irq_enable();
831                         return -EINVAL;
832                 }
833         }
834
835         /*
836          * Must be done before busmaster disable as we might need to
837          * access HPET !
838          */
839         lapic_timer_state_broadcast(pr, cx, 1);
840
841         if (cx->type == ACPI_STATE_C3)
842                 ACPI_FLUSH_CPU_CACHE();
843
844         kt1 = ktime_get_real();
845         /* Tell the scheduler that we are going deep-idle: */
846         sched_clock_idle_sleep_event();
847         acpi_idle_do_entry(cx);
848         kt2 = ktime_get_real();
849         idle_time_ns = ktime_to_ns(ktime_sub(kt2, kt1));
850         idle_time = idle_time_ns;
851         do_div(idle_time, NSEC_PER_USEC);
852
853         /* Update device last_residency*/
854         dev->last_residency = (int)idle_time;
855
856         /* Tell the scheduler how much we idled: */
857         sched_clock_idle_wakeup_event(idle_time_ns);
858
859         local_irq_enable();
860         if (cx->entry_method != ACPI_CSTATE_FFH)
861                 current_thread_info()->status |= TS_POLLING;
862
863         cx->usage++;
864
865         lapic_timer_state_broadcast(pr, cx, 0);
866         cx->time += idle_time;
867         return index;
868 }
869
870 static int c3_cpu_count;
871 static DEFINE_SPINLOCK(c3_lock);
872
873 /**
874  * acpi_idle_enter_bm - enters C3 with proper BM handling
875  * @dev: the target CPU
876  * @drv: cpuidle driver containing state data
877  * @index: the index of suggested state
878  *
879  * If BM is detected, the deepest non-C3 idle state is entered instead.
880  */
881 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
882                 struct cpuidle_driver *drv, int index)
883 {
884         struct acpi_processor *pr;
885         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
886         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
887         ktime_t  kt1, kt2;
888         s64 idle_time_ns;
889         s64 idle_time;
890
891
892         pr = __this_cpu_read(processors);
893         dev->last_residency = 0;
894
895         if (unlikely(!pr))
896                 return -EINVAL;
897
898
899         if (acpi_idle_suspend) {
900                 cpu_relax();
901                 return -EINVAL;
902         }
903
904         if (!cx->bm_sts_skip && acpi_idle_bm_check()) {
905                 if (drv->safe_state_index >= 0) {
906                         return drv->states[drv->safe_state_index].enter(dev,
907                                                 drv, drv->safe_state_index);
908                 } else {
909                         local_irq_disable();
910                         acpi_safe_halt();
911                         local_irq_enable();
912                         return -EINVAL;
913                 }
914         }
915
916         local_irq_disable();
917
918         if (cx->entry_method != ACPI_CSTATE_FFH) {
919                 current_thread_info()->status &= ~TS_POLLING;
920                 /*
921                  * TS_POLLING-cleared state must be visible before we test
922                  * NEED_RESCHED:
923                  */
924                 smp_mb();
925
926                 if (unlikely(need_resched())) {
927                         current_thread_info()->status |= TS_POLLING;
928                         local_irq_enable();
929                         return -EINVAL;
930                 }
931         }
932
933         acpi_unlazy_tlb(smp_processor_id());
934
935         /* Tell the scheduler that we are going deep-idle: */
936         sched_clock_idle_sleep_event();
937         /*
938          * Must be done before busmaster disable as we might need to
939          * access HPET !
940          */
941         lapic_timer_state_broadcast(pr, cx, 1);
942
943         kt1 = ktime_get_real();
944         /*
945          * disable bus master
946          * bm_check implies we need ARB_DIS
947          * !bm_check implies we need cache flush
948          * bm_control implies whether we can do ARB_DIS
949          *
950          * That leaves a case where bm_check is set and bm_control is
951          * not set. In that case we cannot do much, we enter C3
952          * without doing anything.
953          */
954         if (pr->flags.bm_check && pr->flags.bm_control) {
955                 spin_lock(&c3_lock);
956                 c3_cpu_count++;
957                 /* Disable bus master arbitration when all CPUs are in C3 */
958                 if (c3_cpu_count == num_online_cpus())
959                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
960                 spin_unlock(&c3_lock);
961         } else if (!pr->flags.bm_check) {
962                 ACPI_FLUSH_CPU_CACHE();
963         }
964
965         acpi_idle_do_entry(cx);
966
967         /* Re-enable bus master arbitration */
968         if (pr->flags.bm_check && pr->flags.bm_control) {
969                 spin_lock(&c3_lock);
970                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
971                 c3_cpu_count--;
972                 spin_unlock(&c3_lock);
973         }
974         kt2 = ktime_get_real();
975         idle_time_ns = ktime_to_ns(ktime_sub(kt2, kt1));
976         idle_time = idle_time_ns;
977         do_div(idle_time, NSEC_PER_USEC);
978
979         /* Update device last_residency*/
980         dev->last_residency = (int)idle_time;
981
982         /* Tell the scheduler how much we idled: */
983         sched_clock_idle_wakeup_event(idle_time_ns);
984
985         local_irq_enable();
986         if (cx->entry_method != ACPI_CSTATE_FFH)
987                 current_thread_info()->status |= TS_POLLING;
988
989         cx->usage++;
990
991         lapic_timer_state_broadcast(pr, cx, 0);
992         cx->time += idle_time;
993         return index;
994 }
995
996 struct cpuidle_driver acpi_idle_driver = {
997         .name =         "acpi_idle",
998         .owner =        THIS_MODULE,
999 };
1000
1001 /**
1002  * acpi_processor_setup_cpuidle_cx - prepares and configures CPUIDLE
1003  * device i.e. per-cpu data
1004  *
1005  * @pr: the ACPI processor
1006  */
1007 static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr)
1008 {
1009         int i, count = CPUIDLE_DRIVER_STATE_START;
1010         struct acpi_processor_cx *cx;
1011         struct cpuidle_state_usage *state_usage;
1012         struct cpuidle_device *dev = &pr->power.dev;
1013
1014         if (!pr->flags.power_setup_done)
1015                 return -EINVAL;
1016
1017         if (pr->flags.power == 0) {
1018                 return -EINVAL;
1019         }
1020
1021         dev->cpu = pr->id;
1022
1023         if (max_cstate == 0)
1024                 max_cstate = 1;
1025
1026         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1027                 cx = &pr->power.states[i];
1028                 state_usage = &dev->states_usage[count];
1029
1030                 if (!cx->valid)
1031                         continue;
1032
1033 #ifdef CONFIG_HOTPLUG_CPU
1034                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1035                     !pr->flags.has_cst &&
1036                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1037                         continue;
1038 #endif
1039
1040                 cpuidle_set_statedata(state_usage, cx);
1041
1042                 count++;
1043                 if (count == CPUIDLE_STATE_MAX)
1044                         break;
1045         }
1046
1047         dev->state_count = count;
1048
1049         if (!count)
1050                 return -EINVAL;
1051
1052         return 0;
1053 }
1054
1055 /**
1056  * acpi_processor_setup_cpuidle states- prepares and configures cpuidle
1057  * global state data i.e. idle routines
1058  *
1059  * @pr: the ACPI processor
1060  */
1061 static int acpi_processor_setup_cpuidle_states(struct acpi_processor *pr)
1062 {
1063         int i, count = CPUIDLE_DRIVER_STATE_START;
1064         struct acpi_processor_cx *cx;
1065         struct cpuidle_state *state;
1066         struct cpuidle_driver *drv = &acpi_idle_driver;
1067
1068         if (!pr->flags.power_setup_done)
1069                 return -EINVAL;
1070
1071         if (pr->flags.power == 0)
1072                 return -EINVAL;
1073
1074         drv->safe_state_index = -1;
1075         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1076                 drv->states[i].name[0] = '\0';
1077                 drv->states[i].desc[0] = '\0';
1078         }
1079
1080         if (max_cstate == 0)
1081                 max_cstate = 1;
1082
1083         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1084                 cx = &pr->power.states[i];
1085
1086                 if (!cx->valid)
1087                         continue;
1088
1089 #ifdef CONFIG_HOTPLUG_CPU
1090                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1091                     !pr->flags.has_cst &&
1092                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1093                         continue;
1094 #endif
1095
1096                 state = &drv->states[count];
1097                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1098                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1099                 state->exit_latency = cx->latency;
1100                 state->target_residency = cx->latency * latency_factor;
1101
1102                 state->flags = 0;
1103                 switch (cx->type) {
1104                         case ACPI_STATE_C1:
1105                         if (cx->entry_method == ACPI_CSTATE_FFH)
1106                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1107
1108                         state->enter = acpi_idle_enter_c1;
1109                         drv->safe_state_index = count;
1110                         break;
1111
1112                         case ACPI_STATE_C2:
1113                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1114                         state->enter = acpi_idle_enter_simple;
1115                         drv->safe_state_index = count;
1116                         break;
1117
1118                         case ACPI_STATE_C3:
1119                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1120                         state->enter = pr->flags.bm_check ?
1121                                         acpi_idle_enter_bm :
1122                                         acpi_idle_enter_simple;
1123                         break;
1124                 }
1125
1126                 count++;
1127                 if (count == CPUIDLE_STATE_MAX)
1128                         break;
1129         }
1130
1131         drv->state_count = count;
1132
1133         if (!count)
1134                 return -EINVAL;
1135
1136         return 0;
1137 }
1138
1139 int acpi_processor_hotplug(struct acpi_processor *pr)
1140 {
1141         int ret = 0;
1142
1143         if (disabled_by_idle_boot_param())
1144                 return 0;
1145
1146         if (!pr)
1147                 return -EINVAL;
1148
1149         if (nocst) {
1150                 return -ENODEV;
1151         }
1152
1153         if (!pr->flags.power_setup_done)
1154                 return -ENODEV;
1155
1156         cpuidle_pause_and_lock();
1157         cpuidle_disable_device(&pr->power.dev);
1158         acpi_processor_get_power_info(pr);
1159         if (pr->flags.power) {
1160                 acpi_processor_setup_cpuidle_cx(pr);
1161                 ret = cpuidle_enable_device(&pr->power.dev);
1162         }
1163         cpuidle_resume_and_unlock();
1164
1165         return ret;
1166 }
1167
1168 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1169 {
1170         int cpu;
1171         struct acpi_processor *_pr;
1172
1173         if (disabled_by_idle_boot_param())
1174                 return 0;
1175
1176         if (!pr)
1177                 return -EINVAL;
1178
1179         if (nocst)
1180                 return -ENODEV;
1181
1182         if (!pr->flags.power_setup_done)
1183                 return -ENODEV;
1184
1185         /*
1186          * FIXME:  Design the ACPI notification to make it once per
1187          * system instead of once per-cpu.  This condition is a hack
1188          * to make the code that updates C-States be called once.
1189          */
1190
1191         if (smp_processor_id() == 0 &&
1192                         cpuidle_get_driver() == &acpi_idle_driver) {
1193
1194                 cpuidle_pause_and_lock();
1195                 /* Protect against cpu-hotplug */
1196                 get_online_cpus();
1197
1198                 /* Disable all cpuidle devices */
1199                 for_each_online_cpu(cpu) {
1200                         _pr = per_cpu(processors, cpu);
1201                         if (!_pr || !_pr->flags.power_setup_done)
1202                                 continue;
1203                         cpuidle_disable_device(&_pr->power.dev);
1204                 }
1205
1206                 /* Populate Updated C-state information */
1207                 acpi_processor_setup_cpuidle_states(pr);
1208
1209                 /* Enable all cpuidle devices */
1210                 for_each_online_cpu(cpu) {
1211                         _pr = per_cpu(processors, cpu);
1212                         if (!_pr || !_pr->flags.power_setup_done)
1213                                 continue;
1214                         acpi_processor_get_power_info(_pr);
1215                         if (_pr->flags.power) {
1216                                 acpi_processor_setup_cpuidle_cx(_pr);
1217                                 cpuidle_enable_device(&_pr->power.dev);
1218                         }
1219                 }
1220                 put_online_cpus();
1221                 cpuidle_resume_and_unlock();
1222         }
1223
1224         return 0;
1225 }
1226
1227 static int acpi_processor_registered;
1228
1229 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1230                               struct acpi_device *device)
1231 {
1232         acpi_status status = 0;
1233         int retval;
1234         static int first_run;
1235
1236         if (disabled_by_idle_boot_param())
1237                 return 0;
1238
1239         if (!first_run) {
1240                 dmi_check_system(processor_power_dmi_table);
1241                 max_cstate = acpi_processor_cstate_check(max_cstate);
1242                 if (max_cstate < ACPI_C_STATES_MAX)
1243                         printk(KERN_NOTICE
1244                                "ACPI: processor limited to max C-state %d\n",
1245                                max_cstate);
1246                 first_run++;
1247         }
1248
1249         if (!pr)
1250                 return -EINVAL;
1251
1252         if (acpi_gbl_FADT.cst_control && !nocst) {
1253                 status =
1254                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1255                 if (ACPI_FAILURE(status)) {
1256                         ACPI_EXCEPTION((AE_INFO, status,
1257                                         "Notifying BIOS of _CST ability failed"));
1258                 }
1259         }
1260
1261         acpi_processor_get_power_info(pr);
1262         pr->flags.power_setup_done = 1;
1263
1264         /*
1265          * Install the idle handler if processor power management is supported.
1266          * Note that we use previously set idle handler will be used on
1267          * platforms that only support C1.
1268          */
1269         if (pr->flags.power) {
1270                 /* Register acpi_idle_driver if not already registered */
1271                 if (!acpi_processor_registered) {
1272                         acpi_processor_setup_cpuidle_states(pr);
1273                         retval = cpuidle_register_driver(&acpi_idle_driver);
1274                         if (retval)
1275                                 return retval;
1276                         printk(KERN_DEBUG "ACPI: %s registered with cpuidle\n",
1277                                         acpi_idle_driver.name);
1278                 }
1279                 /* Register per-cpu cpuidle_device. Cpuidle driver
1280                  * must already be registered before registering device
1281                  */
1282                 acpi_processor_setup_cpuidle_cx(pr);
1283                 retval = cpuidle_register_device(&pr->power.dev);
1284                 if (retval) {
1285                         if (acpi_processor_registered == 0)
1286                                 cpuidle_unregister_driver(&acpi_idle_driver);
1287                         return retval;
1288                 }
1289                 acpi_processor_registered++;
1290         }
1291         return 0;
1292 }
1293
1294 int acpi_processor_power_exit(struct acpi_processor *pr,
1295                               struct acpi_device *device)
1296 {
1297         if (disabled_by_idle_boot_param())
1298                 return 0;
1299
1300         if (pr->flags.power) {
1301                 cpuidle_unregister_device(&pr->power.dev);
1302                 acpi_processor_registered--;
1303                 if (acpi_processor_registered == 0)
1304                         cpuidle_unregister_driver(&acpi_idle_driver);
1305         }
1306
1307         pr->flags.power_setup_done = 0;
1308         return 0;
1309 }