xref: /freebsd/sys/dev/acpica/acpi_cpu.c (revision 6af83ee0d2941d18880b6aaa2b4facd1d30c6106)
1 /*-
2  * Copyright (c) 2003-2005 Nate Lawson (SDG)
3  * Copyright (c) 2001 Michael Smith
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_acpi.h"
32 #include <sys/param.h>
33 #include <sys/bus.h>
34 #include <sys/cpu.h>
35 #include <sys/kernel.h>
36 #include <sys/malloc.h>
37 #include <sys/module.h>
38 #include <sys/pcpu.h>
39 #include <sys/power.h>
40 #include <sys/proc.h>
41 #include <sys/sbuf.h>
42 #include <sys/smp.h>
43 
44 #include <dev/pci/pcivar.h>
45 #include <machine/atomic.h>
46 #include <machine/bus.h>
47 #include <sys/rman.h>
48 
49 #include "acpi.h"
50 #include <dev/acpica/acpivar.h>
51 
52 /*
53  * Support for ACPI Processor devices, including C[1-3] sleep states.
54  *
55  * TODO: implement scans of all CPUs to be sure all Cx states are
56  * equivalent.
57  */
58 
59 /* Hooks for the ACPI CA debugging infrastructure */
60 #define _COMPONENT	ACPI_PROCESSOR
61 ACPI_MODULE_NAME("PROCESSOR")
62 
63 struct acpi_cx {
64     struct resource	*p_lvlx;	/* Register to read to enter state. */
65     uint32_t		 type;		/* C1-3 (C4 and up treated as C3). */
66     uint32_t		 trans_lat;	/* Transition latency (usec). */
67     uint32_t		 power;		/* Power consumed (mW). */
68     int			 res_type;	/* Resource type for p_lvlx. */
69 };
70 #define MAX_CX_STATES	 8
71 
72 struct acpi_cpu_softc {
73     device_t		 cpu_dev;
74     ACPI_HANDLE		 cpu_handle;
75     struct pcpu		*cpu_pcpu;
76     uint32_t		 cpu_acpi_id;	/* ACPI processor id */
77     uint32_t		 cpu_p_blk;	/* ACPI P_BLK location */
78     uint32_t		 cpu_p_blk_len;	/* P_BLK length (must be 6). */
79     struct acpi_cx	 cpu_cx_states[MAX_CX_STATES];
80     int			 cpu_cx_count;	/* Number of valid Cx states. */
81     int			 cpu_prev_sleep;/* Last idle sleep duration. */
82 };
83 
84 struct acpi_cpu_device {
85     struct resource_list        ad_rl;
86 };
87 
88 #define CPU_GET_REG(reg, width) 					\
89     (bus_space_read_ ## width(rman_get_bustag((reg)), 			\
90 		      rman_get_bushandle((reg)), 0))
91 #define CPU_SET_REG(reg, width, val)					\
92     (bus_space_write_ ## width(rman_get_bustag((reg)), 			\
93 		       rman_get_bushandle((reg)), 0, (val)))
94 
95 #define PM_USEC(x)	 ((x) >> 2)	/* ~4 clocks per usec (3.57955 Mhz) */
96 
97 #define ACPI_NOTIFY_CX_STATES	0x81	/* _CST changed. */
98 
99 #define CPU_QUIRK_NO_C3		(1<<0)	/* C3-type states are not usable. */
100 #define CPU_QUIRK_NO_BM_CTRL	(1<<2)	/* No bus mastering control. */
101 
102 #define PCI_VENDOR_INTEL	0x8086
103 #define PCI_DEVICE_82371AB_3	0x7113	/* PIIX4 chipset for quirks. */
104 #define PCI_REVISION_A_STEP	0
105 #define PCI_REVISION_B_STEP	1
106 #define PCI_REVISION_4E		2
107 #define PCI_REVISION_4M		3
108 
109 /* Platform hardware resource information. */
110 static uint32_t		 cpu_smi_cmd;	/* Value to write to SMI_CMD. */
111 static uint8_t		 cpu_cst_cnt;	/* Indicate we are _CST aware. */
112 static int		 cpu_rid;	/* Driver-wide resource id. */
113 static int		 cpu_quirks;	/* Indicate any hardware bugs. */
114 
115 /* Runtime state. */
116 static int		 cpu_cx_count;	/* Number of valid states */
117 static int		 cpu_non_c3;	/* Index of lowest non-C3 state. */
118 static u_int		 cpu_cx_stats[MAX_CX_STATES];/* Cx usage history. */
119 
120 /* Values for sysctl. */
121 static struct sysctl_ctx_list acpi_cpu_sysctl_ctx;
122 static struct sysctl_oid *acpi_cpu_sysctl_tree;
123 static int		 cpu_cx_lowest;
124 static char 		 cpu_cx_supported[64];
125 
126 static device_t		*cpu_devices;
127 static int		 cpu_ndevices;
128 static struct acpi_cpu_softc **cpu_softc;
129 ACPI_SERIAL_DECL(cpu, "ACPI CPU");
130 
131 static int	acpi_cpu_probe(device_t dev);
132 static int	acpi_cpu_attach(device_t dev);
133 static int	acpi_pcpu_get_id(uint32_t idx, uint32_t *acpi_id,
134 		    uint32_t *cpu_id);
135 static struct resource_list *acpi_cpu_get_rlist(device_t dev, device_t child);
136 static device_t	acpi_cpu_add_child(device_t dev, int order, const char *name,
137 		    int unit);
138 static int	acpi_cpu_read_ivar(device_t dev, device_t child, int index,
139 		    uintptr_t *result);
140 static int	acpi_cpu_shutdown(device_t dev);
141 static int	acpi_cpu_cx_probe(struct acpi_cpu_softc *sc);
142 static int	acpi_cpu_cx_cst(struct acpi_cpu_softc *sc);
143 static void	acpi_cpu_startup(void *arg);
144 static void	acpi_cpu_startup_cx(void);
145 static void	acpi_cpu_idle(void);
146 static void	acpi_cpu_notify(ACPI_HANDLE h, UINT32 notify, void *context);
147 static int	acpi_cpu_quirks(struct acpi_cpu_softc *sc);
148 static int	acpi_cpu_usage_sysctl(SYSCTL_HANDLER_ARGS);
149 static int	acpi_cpu_cx_lowest_sysctl(SYSCTL_HANDLER_ARGS);
150 
151 static device_method_t acpi_cpu_methods[] = {
152     /* Device interface */
153     DEVMETHOD(device_probe,	acpi_cpu_probe),
154     DEVMETHOD(device_attach,	acpi_cpu_attach),
155     DEVMETHOD(device_detach,	bus_generic_detach),
156     DEVMETHOD(device_shutdown,	acpi_cpu_shutdown),
157     DEVMETHOD(device_suspend,	bus_generic_suspend),
158     DEVMETHOD(device_resume,	bus_generic_resume),
159 
160     /* Bus interface */
161     DEVMETHOD(bus_add_child,	acpi_cpu_add_child),
162     DEVMETHOD(bus_read_ivar,	acpi_cpu_read_ivar),
163     DEVMETHOD(bus_get_resource_list, acpi_cpu_get_rlist),
164     DEVMETHOD(bus_get_resource,	bus_generic_rl_get_resource),
165     DEVMETHOD(bus_set_resource,	bus_generic_rl_set_resource),
166     DEVMETHOD(bus_alloc_resource, bus_generic_rl_alloc_resource),
167     DEVMETHOD(bus_release_resource, bus_generic_rl_release_resource),
168     DEVMETHOD(bus_driver_added,	bus_generic_driver_added),
169     DEVMETHOD(bus_activate_resource, bus_generic_activate_resource),
170     DEVMETHOD(bus_deactivate_resource, bus_generic_deactivate_resource),
171     DEVMETHOD(bus_setup_intr,	bus_generic_setup_intr),
172     DEVMETHOD(bus_teardown_intr, bus_generic_teardown_intr),
173 
174     {0, 0}
175 };
176 
177 static driver_t acpi_cpu_driver = {
178     "cpu",
179     acpi_cpu_methods,
180     sizeof(struct acpi_cpu_softc),
181 };
182 
183 static devclass_t acpi_cpu_devclass;
184 DRIVER_MODULE(cpu, acpi, acpi_cpu_driver, acpi_cpu_devclass, 0, 0);
185 MODULE_DEPEND(cpu, acpi, 1, 1, 1);
186 
187 static int
188 acpi_cpu_probe(device_t dev)
189 {
190     int			   acpi_id, cpu_id, cx_count;
191     ACPI_BUFFER		   buf;
192     ACPI_HANDLE		   handle;
193     char		   msg[32];
194     ACPI_OBJECT		   *obj;
195     ACPI_STATUS		   status;
196 
197     if (acpi_disabled("cpu") || acpi_get_type(dev) != ACPI_TYPE_PROCESSOR)
198 	return (ENXIO);
199 
200     handle = acpi_get_handle(dev);
201     if (cpu_softc == NULL)
202 	cpu_softc = malloc(sizeof(struct acpi_cpu_softc *) *
203 	    (mp_maxid + 1), M_TEMP /* XXX */, M_WAITOK | M_ZERO);
204 
205     /* Get our Processor object. */
206     buf.Pointer = NULL;
207     buf.Length = ACPI_ALLOCATE_BUFFER;
208     status = AcpiEvaluateObject(handle, NULL, NULL, &buf);
209     if (ACPI_FAILURE(status)) {
210 	device_printf(dev, "probe failed to get Processor obj - %s\n",
211 		      AcpiFormatException(status));
212 	return (ENXIO);
213     }
214     obj = (ACPI_OBJECT *)buf.Pointer;
215     if (obj->Type != ACPI_TYPE_PROCESSOR) {
216 	device_printf(dev, "Processor object has bad type %d\n", obj->Type);
217 	AcpiOsFree(obj);
218 	return (ENXIO);
219     }
220 
221     /*
222      * Find the processor associated with our unit.  We could use the
223      * ProcId as a key, however, some boxes do not have the same values
224      * in their Processor object as the ProcId values in the MADT.
225      */
226     acpi_id = obj->Processor.ProcId;
227     AcpiOsFree(obj);
228     if (acpi_pcpu_get_id(device_get_unit(dev), &acpi_id, &cpu_id) != 0)
229 	return (ENXIO);
230 
231     /*
232      * Check if we already probed this processor.  We scan the bus twice
233      * so it's possible we've already seen this one.
234      */
235     if (cpu_softc[cpu_id] != NULL)
236 	return (ENXIO);
237 
238     /* Get a count of Cx states for our device string. */
239     cx_count = 0;
240     buf.Pointer = NULL;
241     buf.Length = ACPI_ALLOCATE_BUFFER;
242     status = AcpiEvaluateObject(handle, "_CST", NULL, &buf);
243     if (ACPI_SUCCESS(status)) {
244 	obj = (ACPI_OBJECT *)buf.Pointer;
245 	if (ACPI_PKG_VALID(obj, 2))
246 	    acpi_PkgInt32(obj, 0, &cx_count);
247 	AcpiOsFree(obj);
248     } else {
249 	if (AcpiGbl_FADT->Plvl2Lat <= 100)
250 	    cx_count++;
251 	if (AcpiGbl_FADT->Plvl3Lat <= 1000)
252 	    cx_count++;
253 	if (cx_count > 0)
254 	    cx_count++;
255     }
256     if (cx_count > 0)
257 	snprintf(msg, sizeof(msg), "ACPI CPU (%d Cx states)", cx_count);
258     else
259 	strlcpy(msg, "ACPI CPU", sizeof(msg));
260     device_set_desc_copy(dev, msg);
261 
262     /* Mark this processor as in-use and save our derived id for attach. */
263     cpu_softc[cpu_id] = (void *)1;
264     acpi_set_magic(dev, cpu_id);
265 
266     return (0);
267 }
268 
269 static int
270 acpi_cpu_attach(device_t dev)
271 {
272     ACPI_BUFFER		   buf;
273     ACPI_OBJECT		   *obj;
274     struct pcpu		   *pcpu_data;
275     struct acpi_cpu_softc *sc;
276     struct acpi_softc	  *acpi_sc;
277     ACPI_STATUS		   status;
278     int			   cpu_id;
279 
280     ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__);
281 
282     sc = device_get_softc(dev);
283     sc->cpu_dev = dev;
284     sc->cpu_handle = acpi_get_handle(dev);
285     cpu_id = acpi_get_magic(dev);
286     cpu_softc[cpu_id] = sc;
287     pcpu_data = pcpu_find(cpu_id);
288     pcpu_data->pc_device = dev;
289     sc->cpu_pcpu = pcpu_data;
290     cpu_smi_cmd = AcpiGbl_FADT->SmiCmd;
291     cpu_cst_cnt = AcpiGbl_FADT->CstCnt;
292 
293     buf.Pointer = NULL;
294     buf.Length = ACPI_ALLOCATE_BUFFER;
295     status = AcpiEvaluateObject(sc->cpu_handle, NULL, NULL, &buf);
296     if (ACPI_FAILURE(status)) {
297 	device_printf(dev, "attach failed to get Processor obj - %s\n",
298 		      AcpiFormatException(status));
299 	return (ENXIO);
300     }
301     obj = (ACPI_OBJECT *)buf.Pointer;
302     sc->cpu_p_blk = obj->Processor.PblkAddress;
303     sc->cpu_p_blk_len = obj->Processor.PblkLength;
304     sc->cpu_acpi_id = obj->Processor.ProcId;
305     AcpiOsFree(obj);
306     ACPI_DEBUG_PRINT((ACPI_DB_INFO, "acpi_cpu%d: P_BLK at %#x/%d\n",
307 		     device_get_unit(dev), sc->cpu_p_blk, sc->cpu_p_blk_len));
308 
309     acpi_sc = acpi_device_get_parent_softc(dev);
310     sysctl_ctx_init(&acpi_cpu_sysctl_ctx);
311     acpi_cpu_sysctl_tree = SYSCTL_ADD_NODE(&acpi_cpu_sysctl_ctx,
312 	SYSCTL_CHILDREN(acpi_sc->acpi_sysctl_tree), OID_AUTO, "cpu",
313 	CTLFLAG_RD, 0, "");
314 
315     /*
316      * Probe for Cx state support.  If it isn't present, free up unused
317      * resources.
318      */
319     if (acpi_cpu_cx_probe(sc) == 0) {
320 	status = AcpiInstallNotifyHandler(sc->cpu_handle, ACPI_DEVICE_NOTIFY,
321 					  acpi_cpu_notify, sc);
322 	if (device_get_unit(dev) == 0)
323 	    AcpiOsQueueForExecution(OSD_PRIORITY_LO, acpi_cpu_startup, NULL);
324     } else
325 	sysctl_ctx_free(&acpi_cpu_sysctl_ctx);
326 
327     /* Call identify and then probe/attach for cpu child drivers. */
328     bus_generic_probe(dev);
329     bus_generic_attach(dev);
330 
331     return (0);
332 }
333 
334 /*
335  * Find the nth present CPU and return its pc_cpuid as well as set the
336  * pc_acpi_id from the most reliable source.
337  */
338 static int
339 acpi_pcpu_get_id(uint32_t idx, uint32_t *acpi_id, uint32_t *cpu_id)
340 {
341     struct pcpu	*pcpu_data;
342     uint32_t	 i;
343 
344     KASSERT(acpi_id != NULL, ("Null acpi_id"));
345     KASSERT(cpu_id != NULL, ("Null cpu_id"));
346     for (i = 0; i <= mp_maxid; i++) {
347 	if (CPU_ABSENT(i))
348 	    continue;
349 	pcpu_data = pcpu_find(i);
350 	KASSERT(pcpu_data != NULL, ("no pcpu data for %d", i));
351 	if (idx-- == 0) {
352 	    /*
353 	     * If pc_acpi_id was not initialized (e.g., a non-APIC UP box)
354 	     * override it with the value from the ASL.  Otherwise, if the
355 	     * two don't match, prefer the MADT-derived value.  Finally,
356 	     * return the pc_cpuid to reference this processor.
357 	     */
358 	    if (pcpu_data->pc_acpi_id == 0xffffffff)
359 		 pcpu_data->pc_acpi_id = *acpi_id;
360 	    else if (pcpu_data->pc_acpi_id != *acpi_id)
361 		*acpi_id = pcpu_data->pc_acpi_id;
362 	    *cpu_id = pcpu_data->pc_cpuid;
363 	    return (0);
364 	}
365     }
366 
367     return (ESRCH);
368 }
369 
370 static struct resource_list *
371 acpi_cpu_get_rlist(device_t dev, device_t child)
372 {
373     struct acpi_cpu_device *ad;
374 
375     ad = device_get_ivars(child);
376     if (ad == NULL)
377 	return (NULL);
378     return (&ad->ad_rl);
379 }
380 
381 static device_t
382 acpi_cpu_add_child(device_t dev, int order, const char *name, int unit)
383 {
384     struct acpi_cpu_device  *ad;
385     device_t            child;
386 
387     if ((ad = malloc(sizeof(*ad), M_TEMP, M_NOWAIT | M_ZERO)) == NULL)
388         return (NULL);
389 
390     resource_list_init(&ad->ad_rl);
391 
392     child = device_add_child_ordered(dev, order, name, unit);
393     if (child != NULL)
394         device_set_ivars(child, ad);
395     return (child);
396 }
397 
398 static int
399 acpi_cpu_read_ivar(device_t dev, device_t child, int index, uintptr_t *result)
400 {
401     struct acpi_cpu_softc *sc;
402 
403     sc = device_get_softc(dev);
404     switch (index) {
405     case ACPI_IVAR_HANDLE:
406 	*result = (uintptr_t)sc->cpu_handle;
407 	break;
408     case CPU_IVAR_PCPU:
409 	*result = (uintptr_t)sc->cpu_pcpu;
410 	break;
411     default:
412 	return (ENOENT);
413     }
414     return (0);
415 }
416 
417 static int
418 acpi_cpu_shutdown(device_t dev)
419 {
420     ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__);
421 
422     /* Allow children to shutdown first. */
423     bus_generic_shutdown(dev);
424 
425     /* Disable any entry to the idle function. */
426     cpu_cx_count = 0;
427 
428     /* Signal and wait for all processors to exit acpi_cpu_idle(). */
429     smp_rendezvous(NULL, NULL, NULL, NULL);
430 
431     return_VALUE (0);
432 }
433 
434 static int
435 acpi_cpu_cx_probe(struct acpi_cpu_softc *sc)
436 {
437     ACPI_GENERIC_ADDRESS gas;
438     struct acpi_cx	*cx_ptr;
439     int			 error;
440 
441     ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__);
442 
443     /*
444      * Bus mastering arbitration control is needed to keep caches coherent
445      * while sleeping in C3.  If it's not present but a working flush cache
446      * instruction is present, flush the caches before entering C3 instead.
447      * Otherwise, just disable C3 completely.
448      */
449     if (AcpiGbl_FADT->V1_Pm2CntBlk == 0 || AcpiGbl_FADT->Pm2CntLen == 0) {
450 	if (AcpiGbl_FADT->WbInvd && AcpiGbl_FADT->WbInvdFlush == 0) {
451 	    cpu_quirks |= CPU_QUIRK_NO_BM_CTRL;
452 	    ACPI_DEBUG_PRINT((ACPI_DB_INFO,
453 		"acpi_cpu%d: no BM control, using flush cache method\n",
454 		device_get_unit(sc->cpu_dev)));
455 	} else {
456 	    cpu_quirks |= CPU_QUIRK_NO_C3;
457 	    ACPI_DEBUG_PRINT((ACPI_DB_INFO,
458 		"acpi_cpu%d: no BM control, C3 not available\n",
459 		device_get_unit(sc->cpu_dev)));
460 	}
461     }
462 
463     /*
464      * First, check for the ACPI 2.0 _CST sleep states object.
465      * If not usable, fall back to the P_BLK's P_LVL2 and P_LVL3.
466      */
467     sc->cpu_cx_count = 0;
468     error = acpi_cpu_cx_cst(sc);
469     if (error != 0) {
470 	cx_ptr = sc->cpu_cx_states;
471 
472 	/* C1 has been required since just after ACPI 1.0 */
473 	cx_ptr->type = ACPI_STATE_C1;
474 	cx_ptr->trans_lat = 0;
475 	cpu_non_c3 = 0;
476 	cx_ptr++;
477 	sc->cpu_cx_count++;
478 
479 	/*
480 	 * The spec says P_BLK must be 6 bytes long.  However, some systems
481 	 * use it to indicate a fractional set of features present so we
482 	 * take 5 as C2.  Some may also have a value of 7 to indicate
483 	 * another C3 but most use _CST for this (as required) and having
484 	 * "only" C1-C3 is not a hardship.
485 	 */
486 	if (sc->cpu_p_blk_len < 5)
487 	    goto done;
488 
489 	/* Validate and allocate resources for C2 (P_LVL2). */
490 	gas.AddressSpaceId = ACPI_ADR_SPACE_SYSTEM_IO;
491 	gas.RegisterBitWidth = 8;
492 	if (AcpiGbl_FADT->Plvl2Lat <= 100) {
493 	    gas.Address = sc->cpu_p_blk + 4;
494 	    acpi_bus_alloc_gas(sc->cpu_dev, &cx_ptr->res_type, &cpu_rid, &gas,
495 		&cx_ptr->p_lvlx);
496 	    if (cx_ptr->p_lvlx != NULL) {
497 		cpu_rid++;
498 		cx_ptr->type = ACPI_STATE_C2;
499 		cx_ptr->trans_lat = AcpiGbl_FADT->Plvl2Lat;
500 		cpu_non_c3 = 1;
501 		cx_ptr++;
502 		sc->cpu_cx_count++;
503 	    }
504 	}
505 	if (sc->cpu_p_blk_len < 6)
506 	    goto done;
507 
508 	/* Validate and allocate resources for C3 (P_LVL3). */
509 	if (AcpiGbl_FADT->Plvl3Lat <= 1000 &&
510 	    (cpu_quirks & CPU_QUIRK_NO_C3) == 0) {
511 	    gas.Address = sc->cpu_p_blk + 5;
512 	    acpi_bus_alloc_gas(sc->cpu_dev, &cx_ptr->res_type, &cpu_rid, &gas,
513 		&cx_ptr->p_lvlx);
514 	    if (cx_ptr->p_lvlx != NULL) {
515 		cpu_rid++;
516 		cx_ptr->type = ACPI_STATE_C3;
517 		cx_ptr->trans_lat = AcpiGbl_FADT->Plvl3Lat;
518 		cx_ptr++;
519 		sc->cpu_cx_count++;
520 	    }
521 	}
522     }
523 
524 done:
525     /* If no valid registers were found, don't attach. */
526     if (sc->cpu_cx_count == 0)
527 	return (ENXIO);
528 
529     /* Use initial sleep value of 1 sec. to start with lowest idle state. */
530     sc->cpu_prev_sleep = 1000000;
531 
532     return (0);
533 }
534 
535 /*
536  * Parse a _CST package and set up its Cx states.  Since the _CST object
537  * can change dynamically, our notify handler may call this function
538  * to clean up and probe the new _CST package.
539  */
540 static int
541 acpi_cpu_cx_cst(struct acpi_cpu_softc *sc)
542 {
543     struct	 acpi_cx *cx_ptr;
544     ACPI_STATUS	 status;
545     ACPI_BUFFER	 buf;
546     ACPI_OBJECT	*top;
547     ACPI_OBJECT	*pkg;
548     uint32_t	 count;
549     int		 i;
550 
551     ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__);
552 
553     buf.Pointer = NULL;
554     buf.Length = ACPI_ALLOCATE_BUFFER;
555     status = AcpiEvaluateObject(sc->cpu_handle, "_CST", NULL, &buf);
556     if (ACPI_FAILURE(status))
557 	return (ENXIO);
558 
559     /* _CST is a package with a count and at least one Cx package. */
560     top = (ACPI_OBJECT *)buf.Pointer;
561     if (!ACPI_PKG_VALID(top, 2) || acpi_PkgInt32(top, 0, &count) != 0) {
562 	device_printf(sc->cpu_dev, "Invalid _CST package\n");
563 	AcpiOsFree(buf.Pointer);
564 	return (ENXIO);
565     }
566     if (count != top->Package.Count - 1) {
567 	device_printf(sc->cpu_dev, "Invalid _CST state count (%d != %d)\n",
568 	       count, top->Package.Count - 1);
569 	count = top->Package.Count - 1;
570     }
571     if (count > MAX_CX_STATES) {
572 	device_printf(sc->cpu_dev, "_CST has too many states (%d)\n", count);
573 	count = MAX_CX_STATES;
574     }
575 
576     /* Set up all valid states. */
577     sc->cpu_cx_count = 0;
578     cx_ptr = sc->cpu_cx_states;
579     for (i = 0; i < count; i++) {
580 	pkg = &top->Package.Elements[i + 1];
581 	if (!ACPI_PKG_VALID(pkg, 4) ||
582 	    acpi_PkgInt32(pkg, 1, &cx_ptr->type) != 0 ||
583 	    acpi_PkgInt32(pkg, 2, &cx_ptr->trans_lat) != 0 ||
584 	    acpi_PkgInt32(pkg, 3, &cx_ptr->power) != 0) {
585 
586 	    device_printf(sc->cpu_dev, "Skipping invalid Cx state package\n");
587 	    continue;
588 	}
589 
590 	/* Validate the state to see if we should use it. */
591 	switch (cx_ptr->type) {
592 	case ACPI_STATE_C1:
593 	    cpu_non_c3 = i;
594 	    cx_ptr++;
595 	    sc->cpu_cx_count++;
596 	    continue;
597 	case ACPI_STATE_C2:
598 	    if (cx_ptr->trans_lat > 100) {
599 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
600 				 "acpi_cpu%d: C2[%d] not available.\n",
601 				 device_get_unit(sc->cpu_dev), i));
602 		continue;
603 	    }
604 	    cpu_non_c3 = i;
605 	    break;
606 	case ACPI_STATE_C3:
607 	default:
608 	    if (cx_ptr->trans_lat > 1000 ||
609 		(cpu_quirks & CPU_QUIRK_NO_C3) != 0) {
610 
611 		ACPI_DEBUG_PRINT((ACPI_DB_INFO,
612 				 "acpi_cpu%d: C3[%d] not available.\n",
613 				 device_get_unit(sc->cpu_dev), i));
614 		continue;
615 	    }
616 	    break;
617 	}
618 
619 #ifdef notyet
620 	/* Free up any previous register. */
621 	if (cx_ptr->p_lvlx != NULL) {
622 	    bus_release_resource(sc->cpu_dev, 0, 0, cx_ptr->p_lvlx);
623 	    cx_ptr->p_lvlx = NULL;
624 	}
625 #endif
626 
627 	/* Allocate the control register for C2 or C3. */
628 	acpi_PkgGas(sc->cpu_dev, pkg, 0, &cx_ptr->res_type, &cpu_rid,
629 	    &cx_ptr->p_lvlx);
630 	if (cx_ptr->p_lvlx) {
631 	    cpu_rid++;
632 	    ACPI_DEBUG_PRINT((ACPI_DB_INFO,
633 			     "acpi_cpu%d: Got C%d - %d latency\n",
634 			     device_get_unit(sc->cpu_dev), cx_ptr->type,
635 			     cx_ptr->trans_lat));
636 	    cx_ptr++;
637 	    sc->cpu_cx_count++;
638 	}
639     }
640     AcpiOsFree(buf.Pointer);
641 
642     return (0);
643 }
644 
645 /*
646  * Call this *after* all CPUs have been attached.
647  */
648 static void
649 acpi_cpu_startup(void *arg)
650 {
651     struct acpi_cpu_softc *sc;
652     int count, i;
653 
654     /* Get set of CPU devices */
655     devclass_get_devices(acpi_cpu_devclass, &cpu_devices, &cpu_ndevices);
656 
657     /* Check for quirks via the first CPU device. */
658     sc = device_get_softc(cpu_devices[0]);
659     acpi_cpu_quirks(sc);
660 
661     /*
662      * Make sure all the processors' Cx counts match.  We should probably
663      * also check the contents of each.  However, no known systems have
664      * non-matching Cx counts so we'll deal with this later.
665      */
666     count = MAX_CX_STATES;
667     for (i = 0; i < cpu_ndevices; i++) {
668 	sc = device_get_softc(cpu_devices[i]);
669 	count = min(sc->cpu_cx_count, count);
670     }
671     cpu_cx_count = count;
672 
673     /* Perform Cx final initialization. */
674     sc = device_get_softc(cpu_devices[0]);
675     if (cpu_cx_count > 0)
676 	acpi_cpu_startup_cx();
677 }
678 
679 static void
680 acpi_cpu_startup_cx()
681 {
682     struct acpi_cpu_softc *sc;
683     struct sbuf sb;
684     int i;
685 
686     /*
687      * Set up the list of Cx states, eliminating C3 states by truncating
688      * cpu_cx_count if quirks indicate C3 is not usable.
689      */
690     sc = device_get_softc(cpu_devices[0]);
691     sbuf_new(&sb, cpu_cx_supported, sizeof(cpu_cx_supported), SBUF_FIXEDLEN);
692     for (i = 0; i < cpu_cx_count; i++) {
693 	if ((cpu_quirks & CPU_QUIRK_NO_C3) == 0 ||
694 	    sc->cpu_cx_states[i].type != ACPI_STATE_C3)
695 	    sbuf_printf(&sb, "C%d/%d ", i + 1, sc->cpu_cx_states[i].trans_lat);
696 	else
697 	    cpu_cx_count = i;
698     }
699     sbuf_trim(&sb);
700     sbuf_finish(&sb);
701     SYSCTL_ADD_STRING(&acpi_cpu_sysctl_ctx,
702 		      SYSCTL_CHILDREN(acpi_cpu_sysctl_tree),
703 		      OID_AUTO, "cx_supported", CTLFLAG_RD, cpu_cx_supported,
704 		      0, "Cx/microsecond values for supported Cx states");
705     SYSCTL_ADD_PROC(&acpi_cpu_sysctl_ctx,
706 		    SYSCTL_CHILDREN(acpi_cpu_sysctl_tree),
707 		    OID_AUTO, "cx_lowest", CTLTYPE_STRING | CTLFLAG_RW,
708 		    NULL, 0, acpi_cpu_cx_lowest_sysctl, "A",
709 		    "lowest Cx sleep state to use");
710     SYSCTL_ADD_PROC(&acpi_cpu_sysctl_ctx,
711 		    SYSCTL_CHILDREN(acpi_cpu_sysctl_tree),
712 		    OID_AUTO, "cx_usage", CTLTYPE_STRING | CTLFLAG_RD,
713 		    NULL, 0, acpi_cpu_usage_sysctl, "A",
714 		    "percent usage for each Cx state");
715 
716 #ifdef notyet
717     /* Signal platform that we can handle _CST notification. */
718     if (cpu_cst_cnt != 0) {
719 	ACPI_LOCK(acpi);
720 	AcpiOsWritePort(cpu_smi_cmd, cpu_cst_cnt, 8);
721 	ACPI_UNLOCK(acpi);
722     }
723 #endif
724 
725     /* Take over idling from cpu_idle_default(). */
726     cpu_idle_hook = acpi_cpu_idle;
727 }
728 
729 /*
730  * Idle the CPU in the lowest state possible.  This function is called with
731  * interrupts disabled.  Note that once it re-enables interrupts, a task
732  * switch can occur so do not access shared data (i.e. the softc) after
733  * interrupts are re-enabled.
734  */
735 static void
736 acpi_cpu_idle()
737 {
738     struct	acpi_cpu_softc *sc;
739     struct	acpi_cx *cx_next;
740     uint32_t	start_time, end_time;
741     int		bm_active, cx_next_idx, i;
742 
743     /* If disabled, return immediately. */
744     if (cpu_cx_count == 0) {
745 	ACPI_ENABLE_IRQS();
746 	return;
747     }
748 
749     /*
750      * Look up our CPU id to get our softc.  If it's NULL, we'll use C1
751      * since there is no ACPI processor object for this CPU.  This occurs
752      * for logical CPUs in the HTT case.
753      */
754     sc = cpu_softc[PCPU_GET(cpuid)];
755     if (sc == NULL) {
756 	acpi_cpu_c1();
757 	return;
758     }
759 
760     /*
761      * If we slept 100 us or more, use the lowest Cx state.  Otherwise,
762      * find the lowest state that has a latency less than or equal to
763      * the length of our last sleep.
764      */
765     cx_next_idx = cpu_cx_lowest;
766     if (sc->cpu_prev_sleep < 100)
767 	for (i = cpu_cx_lowest; i >= 0; i--)
768 	    if (sc->cpu_cx_states[i].trans_lat <= sc->cpu_prev_sleep) {
769 		cx_next_idx = i;
770 		break;
771 	    }
772 
773     /*
774      * Check for bus master activity.  If there was activity, clear
775      * the bit and use the lowest non-C3 state.  Note that the USB
776      * driver polling for new devices keeps this bit set all the
777      * time if USB is loaded.
778      */
779     if ((cpu_quirks & CPU_QUIRK_NO_BM_CTRL) == 0) {
780 	AcpiGetRegister(ACPI_BITREG_BUS_MASTER_STATUS, &bm_active,
781 	    ACPI_MTX_DO_NOT_LOCK);
782 	if (bm_active != 0) {
783 	    AcpiSetRegister(ACPI_BITREG_BUS_MASTER_STATUS, 1,
784 		ACPI_MTX_DO_NOT_LOCK);
785 	    cx_next_idx = min(cx_next_idx, cpu_non_c3);
786 	}
787     }
788 
789     /* Select the next state and update statistics. */
790     cx_next = &sc->cpu_cx_states[cx_next_idx];
791     cpu_cx_stats[cx_next_idx]++;
792     KASSERT(cx_next->type != ACPI_STATE_C0, ("acpi_cpu_idle: C0 sleep"));
793 
794     /*
795      * Execute HLT (or equivalent) and wait for an interrupt.  We can't
796      * calculate the time spent in C1 since the place we wake up is an
797      * ISR.  Assume we slept one quantum and return.
798      */
799     if (cx_next->type == ACPI_STATE_C1) {
800 	sc->cpu_prev_sleep = 1000000 / hz;
801 	acpi_cpu_c1();
802 	return;
803     }
804 
805     /*
806      * For C3, disable bus master arbitration and enable bus master wake
807      * if BM control is available, otherwise flush the CPU cache.
808      */
809     if (cx_next->type == ACPI_STATE_C3) {
810 	if ((cpu_quirks & CPU_QUIRK_NO_BM_CTRL) == 0) {
811 	    AcpiSetRegister(ACPI_BITREG_ARB_DISABLE, 1, ACPI_MTX_DO_NOT_LOCK);
812 	    AcpiSetRegister(ACPI_BITREG_BUS_MASTER_RLD, 1,
813 		ACPI_MTX_DO_NOT_LOCK);
814 	} else
815 	    ACPI_FLUSH_CPU_CACHE();
816     }
817 
818     /*
819      * Read from P_LVLx to enter C2(+), checking time spent asleep.
820      * Use the ACPI timer for measuring sleep time.  Since we need to
821      * get the time very close to the CPU start/stop clock logic, this
822      * is the only reliable time source.
823      */
824     AcpiHwLowLevelRead(32, &start_time, &AcpiGbl_FADT->XPmTmrBlk);
825     CPU_GET_REG(cx_next->p_lvlx, 1);
826 
827     /*
828      * Read the end time twice.  Since it may take an arbitrary time
829      * to enter the idle state, the first read may be executed before
830      * the processor has stopped.  Doing it again provides enough
831      * margin that we are certain to have a correct value.
832      */
833     AcpiHwLowLevelRead(32, &end_time, &AcpiGbl_FADT->XPmTmrBlk);
834     AcpiHwLowLevelRead(32, &end_time, &AcpiGbl_FADT->XPmTmrBlk);
835 
836     /* Enable bus master arbitration and disable bus master wakeup. */
837     if (cx_next->type == ACPI_STATE_C3 &&
838 	(cpu_quirks & CPU_QUIRK_NO_BM_CTRL) == 0) {
839 	AcpiSetRegister(ACPI_BITREG_ARB_DISABLE, 0, ACPI_MTX_DO_NOT_LOCK);
840 	AcpiSetRegister(ACPI_BITREG_BUS_MASTER_RLD, 0, ACPI_MTX_DO_NOT_LOCK);
841     }
842 
843     /* Find the actual time asleep in microseconds, minus overhead. */
844     end_time = acpi_TimerDelta(end_time, start_time);
845     sc->cpu_prev_sleep = PM_USEC(end_time) - cx_next->trans_lat;
846     ACPI_ENABLE_IRQS();
847 }
848 
849 /*
850  * Re-evaluate the _CST object when we are notified that it changed.
851  *
852  * XXX Re-evaluation disabled until locking is done.
853  */
854 static void
855 acpi_cpu_notify(ACPI_HANDLE h, UINT32 notify, void *context)
856 {
857     struct acpi_cpu_softc *sc = (struct acpi_cpu_softc *)context;
858 
859     if (notify != ACPI_NOTIFY_CX_STATES)
860 	return;
861 
862     device_printf(sc->cpu_dev, "Cx states changed\n");
863     /* acpi_cpu_cx_cst(sc); */
864 }
865 
866 static int
867 acpi_cpu_quirks(struct acpi_cpu_softc *sc)
868 {
869     device_t acpi_dev;
870 
871     /*
872      * C3 on multiple CPUs requires using the expensive flush cache
873      * instruction.
874      */
875     if (mp_ncpus > 1)
876 	cpu_quirks |= CPU_QUIRK_NO_BM_CTRL;
877 
878     /* Look for various quirks of the PIIX4 part. */
879     acpi_dev = pci_find_device(PCI_VENDOR_INTEL, PCI_DEVICE_82371AB_3);
880     if (acpi_dev != NULL) {
881 	switch (pci_get_revid(acpi_dev)) {
882 	/*
883 	 * Disable C3 support for all PIIX4 chipsets.  Some of these parts
884 	 * do not report the BMIDE status to the BM status register and
885 	 * others have a livelock bug if Type-F DMA is enabled.  Linux
886 	 * works around the BMIDE bug by reading the BM status directly
887 	 * but we take the simpler approach of disabling C3 for these
888 	 * parts.
889 	 *
890 	 * See erratum #18 ("C3 Power State/BMIDE and Type-F DMA
891 	 * Livelock") from the January 2002 PIIX4 specification update.
892 	 * Applies to all PIIX4 models.
893 	 */
894 	case PCI_REVISION_4E:
895 	case PCI_REVISION_4M:
896 	    cpu_quirks |= CPU_QUIRK_NO_C3;
897 	    break;
898 	default:
899 	    break;
900 	}
901     }
902 
903     return (0);
904 }
905 
906 static int
907 acpi_cpu_usage_sysctl(SYSCTL_HANDLER_ARGS)
908 {
909     struct sbuf	 sb;
910     char	 buf[128];
911     int		 i;
912     uintmax_t	 fract, sum, whole;
913 
914     sum = 0;
915     for (i = 0; i < cpu_cx_count; i++)
916 	sum += cpu_cx_stats[i];
917     sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
918     for (i = 0; i < cpu_cx_count; i++) {
919 	if (sum > 0) {
920 	    whole = (uintmax_t)cpu_cx_stats[i] * 100;
921 	    fract = (whole % sum) * 100;
922 	    sbuf_printf(&sb, "%u.%02u%% ", (u_int)(whole / sum),
923 		(u_int)(fract / sum));
924 	} else
925 	    sbuf_printf(&sb, "0%% ");
926     }
927     sbuf_trim(&sb);
928     sbuf_finish(&sb);
929     sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
930     sbuf_delete(&sb);
931 
932     return (0);
933 }
934 
935 static int
936 acpi_cpu_cx_lowest_sysctl(SYSCTL_HANDLER_ARGS)
937 {
938     struct	 acpi_cpu_softc *sc;
939     char	 state[8];
940     int		 val, error, i;
941 
942     sc = device_get_softc(cpu_devices[0]);
943     snprintf(state, sizeof(state), "C%d", cpu_cx_lowest + 1);
944     error = sysctl_handle_string(oidp, state, sizeof(state), req);
945     if (error != 0 || req->newptr == NULL)
946 	return (error);
947     if (strlen(state) < 2 || toupper(state[0]) != 'C')
948 	return (EINVAL);
949     val = (int) strtol(state + 1, NULL, 10) - 1;
950     if (val < 0 || val > cpu_cx_count - 1)
951 	return (EINVAL);
952 
953     ACPI_SERIAL_BEGIN(cpu);
954     cpu_cx_lowest = val;
955 
956     /* If not disabling, cache the new lowest non-C3 state. */
957     cpu_non_c3 = 0;
958     for (i = cpu_cx_lowest; i >= 0; i--) {
959 	if (sc->cpu_cx_states[i].type < ACPI_STATE_C3) {
960 	    cpu_non_c3 = i;
961 	    break;
962 	}
963     }
964 
965     /* Reset the statistics counters. */
966     bzero(cpu_cx_stats, sizeof(cpu_cx_stats));
967     ACPI_SERIAL_END(cpu);
968 
969     return (0);
970 }
971