2 * Copyright (c) 2003 Nate Lawson (SDG)
3 * Copyright (c) 2001 Michael Smith
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
27 * $FreeBSD: src/sys/dev/acpica/acpi_cpu.c,v 1.41 2004/06/24 00:38:51 njl Exp $
28 * $DragonFly: src/sys/dev/acpica5/acpi_cpu.c,v 1.5 2004/07/05 00:07:35 dillon Exp $
32 #include <sys/param.h>
34 #include <sys/kernel.h>
35 #include <sys/malloc.h>
36 #include <sys/globaldata.h>
37 #include <sys/power.h>
41 #include <bus/pci/pcivar.h>
42 #include <machine/atomic.h>
43 #include <machine/bus.h>
44 #include <machine/globaldata.h>
45 #include <machine/smp.h>
52 * Support for ACPI Processor devices, including ACPI 2.0 throttling
53 * and C[1-3] sleep states.
55 * TODO: implement scans of all CPUs to be sure all Cx states are
59 /* Hooks for the ACPI CA debugging infrastructure */
60 #define _COMPONENT ACPI_PROCESSOR
61 ACPI_MODULE_NAME("PROCESSOR")
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). */
69 #define MAX_CX_STATES 8
71 struct acpi_cpu_softc {
73 ACPI_HANDLE cpu_handle;
74 uint32_t acpi_id; /* ACPI processor id */
75 uint32_t cpu_p_blk; /* ACPI P_BLK location */
76 uint32_t cpu_p_blk_len; /* P_BLK length (must be 6). */
77 struct resource *cpu_p_cnt; /* Throttling control register */
78 struct acpi_cx cpu_cx_states[MAX_CX_STATES];
79 int cpu_cx_count; /* Number of valid Cx states. */
80 int cpu_prev_sleep;/* Last idle sleep duration. */
83 #define CPU_GET_REG(reg, width) \
84 (bus_space_read_ ## width(rman_get_bustag((reg)), \
85 rman_get_bushandle((reg)), 0))
86 #define CPU_SET_REG(reg, width, val) \
87 (bus_space_write_ ## width(rman_get_bustag((reg)), \
88 rman_get_bushandle((reg)), 0, (val)))
91 * Speeds are stored in counts, from 1 to CPU_MAX_SPEED, and
92 * reported to the user in tenths of a percent.
94 static uint32_t cpu_duty_offset;
95 static uint32_t cpu_duty_width;
96 #define CPU_MAX_SPEED (1 << cpu_duty_width)
97 #define CPU_SPEED_PERCENT(x) ((1000 * (x)) / CPU_MAX_SPEED)
98 #define CPU_SPEED_PRINTABLE(x) (CPU_SPEED_PERCENT(x) / 10), \
99 (CPU_SPEED_PERCENT(x) % 10)
100 #define CPU_P_CNT_THT_EN (1<<4)
101 #define PM_USEC(x) ((x) >> 2) /* ~4 clocks per usec (3.57955 Mhz) */
103 #define ACPI_CPU_NOTIFY_PERF_STATES 0x80 /* _PSS changed. */
104 #define ACPI_CPU_NOTIFY_CX_STATES 0x81 /* _CST changed. */
106 #define CPU_QUIRK_NO_C3 0x0001 /* C3-type states are not usable. */
107 #define CPU_QUIRK_NO_THROTTLE 0x0002 /* Throttling is not usable. */
109 #define PCI_VENDOR_INTEL 0x8086
110 #define PCI_DEVICE_82371AB_3 0x7113 /* PIIX4 chipset for quirks. */
111 #define PCI_REVISION_A_STEP 0
112 #define PCI_REVISION_B_STEP 1
113 #define PCI_REVISION_4E 2
114 #define PCI_REVISION_4M 3
116 /* Platform hardware resource information. */
117 static uint32_t cpu_smi_cmd; /* Value to write to SMI_CMD. */
118 static uint8_t cpu_pstate_cnt;/* Register to take over throttling. */
119 static uint8_t cpu_cst_cnt; /* Indicate we are _CST aware. */
120 static int cpu_rid; /* Driver-wide resource id. */
121 static int cpu_quirks; /* Indicate any hardware bugs. */
124 static int cpu_cx_count; /* Number of valid states */
125 static int cpu_non_c3; /* Index of lowest non-C3 state. */
126 static u_int cpu_cx_stats[MAX_CX_STATES];/* Cx usage history. */
128 /* Values for sysctl. */
129 static uint32_t cpu_throttle_state;
130 static uint32_t cpu_throttle_max;
131 static int cpu_cx_lowest;
132 static char cpu_cx_supported[64];
134 static device_t *cpu_devices;
135 static int cpu_ndevices;
136 static struct acpi_cpu_softc **cpu_softc;
138 static struct sysctl_ctx_list acpi_cpu_sysctl_ctx;
139 static struct sysctl_oid *acpi_cpu_sysctl_tree;
141 static int acpi_cpu_probe(device_t dev);
142 static int acpi_cpu_attach(device_t dev);
143 static int acpi_pcpu_get_id(uint32_t idx, uint32_t *acpi_id,
145 static int acpi_cpu_shutdown(device_t dev);
146 static int acpi_cpu_throttle_probe(struct acpi_cpu_softc *sc);
147 static int acpi_cpu_cx_probe(struct acpi_cpu_softc *sc);
148 static int acpi_cpu_cx_cst(struct acpi_cpu_softc *sc);
149 static void acpi_cpu_startup(void *arg);
150 static void acpi_cpu_startup_throttling(void);
151 static void acpi_cpu_startup_cx(void);
152 static void acpi_cpu_throttle_set(uint32_t speed);
153 static void acpi_cpu_idle(void);
154 static void acpi_cpu_c1(void);
155 static void acpi_cpu_notify(ACPI_HANDLE h, UINT32 notify, void *context);
156 static int acpi_cpu_quirks(struct acpi_cpu_softc *sc);
157 static int acpi_cpu_throttle_sysctl(SYSCTL_HANDLER_ARGS);
158 static int acpi_cpu_usage_sysctl(SYSCTL_HANDLER_ARGS);
159 static int acpi_cpu_cx_lowest_sysctl(SYSCTL_HANDLER_ARGS);
161 static device_method_t acpi_cpu_methods[] = {
162 /* Device interface */
163 DEVMETHOD(device_probe, acpi_cpu_probe),
164 DEVMETHOD(device_attach, acpi_cpu_attach),
165 DEVMETHOD(device_shutdown, acpi_cpu_shutdown),
170 static driver_t acpi_cpu_driver = {
173 sizeof(struct acpi_cpu_softc),
176 static devclass_t acpi_cpu_devclass;
177 DRIVER_MODULE(cpu, acpi, acpi_cpu_driver, acpi_cpu_devclass, 0, 0);
178 MODULE_DEPEND(cpu, acpi, 1, 1, 1);
181 acpi_cpu_probe(device_t dev)
183 int acpi_id, cpu_id, cx_count;
190 if (acpi_disabled("cpu") || acpi_get_type(dev) != ACPI_TYPE_PROCESSOR)
193 handle = acpi_get_handle(dev);
194 if (cpu_softc == NULL)
195 cpu_softc = malloc(sizeof(struct acpi_cpu_softc *) *
196 SMP_MAXCPU, M_TEMP /* XXX */, M_INTWAIT | M_ZERO);
198 /* Get our Processor object. */
200 buf.Length = ACPI_ALLOCATE_BUFFER;
201 status = AcpiEvaluateObject(handle, NULL, NULL, &buf);
202 if (ACPI_FAILURE(status)) {
203 device_printf(dev, "probe failed to get Processor obj - %s\n",
204 AcpiFormatException(status));
207 obj = (ACPI_OBJECT *)buf.Pointer;
208 if (obj->Type != ACPI_TYPE_PROCESSOR) {
209 device_printf(dev, "Processor object has bad type %d\n", obj->Type);
215 * Find the processor associated with our unit. We could use the
216 * ProcId as a key, however, some boxes do not have the same values
217 * in their Processor object as the ProcId values in the MADT.
219 acpi_id = obj->Processor.ProcId;
221 if (acpi_pcpu_get_id(device_get_unit(dev), &acpi_id, &cpu_id) != 0)
225 * Check if we already probed this processor. We scan the bus twice
226 * so it's possible we've already seen this one.
228 if (cpu_softc[cpu_id] != NULL)
231 /* Get a count of Cx states for our device string. */
234 buf.Length = ACPI_ALLOCATE_BUFFER;
235 status = AcpiEvaluateObject(handle, "_CST", NULL, &buf);
236 if (ACPI_SUCCESS(status)) {
237 obj = (ACPI_OBJECT *)buf.Pointer;
238 if (ACPI_PKG_VALID(obj, 2))
239 acpi_PkgInt32(obj, 0, &cx_count);
242 if (AcpiGbl_FADT->Plvl2Lat <= 100)
244 if (AcpiGbl_FADT->Plvl3Lat <= 1000)
250 snprintf(msg, sizeof(msg), "ACPI CPU (%d Cx states)", cx_count);
252 strlcpy(msg, "ACPI CPU", sizeof(msg));
253 device_set_desc_copy(dev, msg);
255 /* Mark this processor as in-use and save our derived id for attach. */
256 cpu_softc[cpu_id] = (void *)1;
257 acpi_set_magic(dev, cpu_id);
263 acpi_cpu_attach(device_t dev)
267 struct acpi_cpu_softc *sc;
268 struct acpi_softc *acpi_sc;
272 ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__);
276 sc = device_get_softc(dev);
278 sc->cpu_handle = acpi_get_handle(dev);
279 cpu_softc[acpi_get_magic(dev)] = sc;
282 buf.Length = ACPI_ALLOCATE_BUFFER;
283 status = AcpiEvaluateObject(sc->cpu_handle, NULL, NULL, &buf);
284 if (ACPI_FAILURE(status)) {
285 device_printf(dev, "attach failed to get Processor obj - %s\n",
286 AcpiFormatException(status));
289 obj = (ACPI_OBJECT *)buf.Pointer;
290 sc->cpu_p_blk = obj->Processor.PblkAddress;
291 sc->cpu_p_blk_len = obj->Processor.PblkLength;
292 sc->acpi_id = obj->Processor.ProcId;
294 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "acpi_cpu%d: P_BLK at %#x/%d\n",
295 device_get_unit(dev), sc->cpu_p_blk, sc->cpu_p_blk_len));
297 acpi_sc = acpi_device_get_parent_softc(dev);
298 sysctl_ctx_init(&acpi_cpu_sysctl_ctx);
299 acpi_cpu_sysctl_tree = SYSCTL_ADD_NODE(&acpi_cpu_sysctl_ctx,
300 SYSCTL_CHILDREN(acpi_sc->acpi_sysctl_tree),
301 OID_AUTO, "cpu", CTLFLAG_RD, 0, "");
303 /* If this is the first device probed, check for quirks. */
304 if (device_get_unit(dev) == 0)
308 * Probe for throttling and Cx state support.
309 * If none of these is present, free up unused resources.
311 thr_ret = acpi_cpu_throttle_probe(sc);
312 cx_ret = acpi_cpu_cx_probe(sc);
313 if (thr_ret == 0 || cx_ret == 0) {
314 status = AcpiInstallNotifyHandler(sc->cpu_handle, ACPI_DEVICE_NOTIFY,
315 acpi_cpu_notify, sc);
316 if (device_get_unit(dev) == 0)
317 AcpiOsQueueForExecution(OSD_PRIORITY_LO, acpi_cpu_startup, NULL);
319 sysctl_ctx_free(&acpi_cpu_sysctl_ctx);
326 * Find the nth present CPU and return its pc_cpuid as well as set the
327 * pc_acpi_id from the most reliable source.
330 acpi_pcpu_get_id(uint32_t idx, uint32_t *acpi_id, uint32_t *cpu_id)
332 struct mdglobaldata *md;
335 KASSERT(acpi_id != NULL, ("Null acpi_id"));
336 KASSERT(cpu_id != NULL, ("Null cpu_id"));
337 for (i = 0; i <= ncpus; i++) {
338 if ((smp_active_mask & (1 << i)) == 0)
340 md = (struct mdglobaldata *)globaldata_find(i);
341 KASSERT(md != NULL, ("no pcpu data for %d", i));
344 * If pc_acpi_id was not initialized (e.g., a non-APIC UP box)
345 * override it with the value from the ASL. Otherwise, if the
346 * two don't match, prefer the MADT-derived value. Finally,
347 * return the pc_cpuid to reference this processor.
349 if (md->gd_acpi_id == 0xffffffff)
350 md->gd_acpi_id = *acpi_id;
351 else if (md->gd_acpi_id != *acpi_id)
352 *acpi_id = md->gd_acpi_id;
353 *cpu_id = md->mi.gd_cpuid;
362 acpi_cpu_shutdown(device_t dev)
364 ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__);
366 /* Disable any entry to the idle function. */
369 /* Signal and wait for all processors to exit acpi_cpu_idle(). */
371 /*smp_rendezvous(NULL, NULL, NULL, NULL);*/
372 KKASSERT(0); /* XXX use rendezvous */
380 acpi_cpu_throttle_probe(struct acpi_cpu_softc *sc)
385 ACPI_GENERIC_ADDRESS gas;
388 ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__);
392 /* Get throttling parameters from the FADT. 0 means not supported. */
393 if (device_get_unit(sc->cpu_dev) == 0) {
394 cpu_smi_cmd = AcpiGbl_FADT->SmiCmd;
395 cpu_pstate_cnt = AcpiGbl_FADT->PstateCnt;
396 cpu_cst_cnt = AcpiGbl_FADT->CstCnt;
397 cpu_duty_offset = AcpiGbl_FADT->DutyOffset;
398 cpu_duty_width = AcpiGbl_FADT->DutyWidth;
400 if (cpu_duty_width == 0 || (cpu_quirks & CPU_QUIRK_NO_THROTTLE) != 0)
403 /* Validate the duty offset/width. */
404 duty_end = cpu_duty_offset + cpu_duty_width - 1;
406 device_printf(sc->cpu_dev, "CLK_VAL field overflows P_CNT register\n");
409 if (cpu_duty_offset <= 4 && duty_end >= 4) {
410 device_printf(sc->cpu_dev, "CLK_VAL field overlaps THT_EN bit\n");
415 * If not present, fall back to using the processor's P_BLK to find
416 * the P_CNT register.
418 * Note that some systems seem to duplicate the P_BLK pointer
419 * across multiple CPUs, so not getting the resource is not fatal.
422 buf.Length = sizeof(obj);
423 status = AcpiEvaluateObject(sc->cpu_handle, "_PTC", NULL, &buf);
424 if (ACPI_SUCCESS(status)) {
425 if (obj.Buffer.Length < sizeof(ACPI_GENERIC_ADDRESS) + 3) {
426 device_printf(sc->cpu_dev, "_PTC buffer too small\n");
429 memcpy(&gas, obj.Buffer.Pointer + 3, sizeof(gas));
430 sc->cpu_p_cnt = acpi_bus_alloc_gas(sc->cpu_dev, &cpu_rid, &gas);
431 if (sc->cpu_p_cnt != NULL) {
432 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "acpi_cpu%d: P_CNT from _PTC\n",
433 device_get_unit(sc->cpu_dev)));
437 /* If _PTC not present or other failure, try the P_BLK. */
438 if (sc->cpu_p_cnt == NULL) {
440 * The spec says P_BLK must be 6 bytes long. However, some
441 * systems use it to indicate a fractional set of features
442 * present so we take anything >= 4.
444 if (sc->cpu_p_blk_len < 4)
446 gas.Address = sc->cpu_p_blk;
447 gas.AddressSpaceId = ACPI_ADR_SPACE_SYSTEM_IO;
448 gas.RegisterBitWidth = 32;
449 sc->cpu_p_cnt = acpi_bus_alloc_gas(sc->cpu_dev, &cpu_rid, &gas);
450 if (sc->cpu_p_cnt != NULL) {
451 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "acpi_cpu%d: P_CNT from P_BLK\n",
452 device_get_unit(sc->cpu_dev)));
454 device_printf(sc->cpu_dev, "Failed to attach throttling P_CNT\n");
464 acpi_cpu_cx_probe(struct acpi_cpu_softc *sc)
466 ACPI_GENERIC_ADDRESS gas;
467 struct acpi_cx *cx_ptr;
470 ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__);
472 /* Bus mastering arbitration control is needed for C3. */
473 if (AcpiGbl_FADT->V1_Pm2CntBlk == 0 || AcpiGbl_FADT->Pm2CntLen == 0) {
474 cpu_quirks |= CPU_QUIRK_NO_C3;
475 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
476 "acpi_cpu%d: No BM control, C3 disabled\n",
477 device_get_unit(sc->cpu_dev)));
481 * First, check for the ACPI 2.0 _CST sleep states object.
482 * If not usable, fall back to the P_BLK's P_LVL2 and P_LVL3.
484 sc->cpu_cx_count = 0;
485 error = acpi_cpu_cx_cst(sc);
487 cx_ptr = sc->cpu_cx_states;
489 /* C1 has been required since just after ACPI 1.0 */
490 cx_ptr->type = ACPI_STATE_C1;
491 cx_ptr->trans_lat = 0;
497 * The spec says P_BLK must be 6 bytes long. However, some systems
498 * use it to indicate a fractional set of features present so we
499 * take 5 as C2. Some may also have a value of 7 to indicate
500 * another C3 but most use _CST for this (as required) and having
501 * "only" C1-C3 is not a hardship.
503 if (sc->cpu_p_blk_len < 5)
506 /* Validate and allocate resources for C2 (P_LVL2). */
507 gas.AddressSpaceId = ACPI_ADR_SPACE_SYSTEM_IO;
508 gas.RegisterBitWidth = 8;
509 if (AcpiGbl_FADT->Plvl2Lat <= 100) {
510 gas.Address = sc->cpu_p_blk + 4;
511 cx_ptr->p_lvlx = acpi_bus_alloc_gas(sc->cpu_dev, &cpu_rid, &gas);
512 if (cx_ptr->p_lvlx != NULL) {
514 cx_ptr->type = ACPI_STATE_C2;
515 cx_ptr->trans_lat = AcpiGbl_FADT->Plvl2Lat;
521 if (sc->cpu_p_blk_len < 6)
524 /* Validate and allocate resources for C3 (P_LVL3). */
525 if (AcpiGbl_FADT->Plvl3Lat <= 1000 &&
526 (cpu_quirks & CPU_QUIRK_NO_C3) == 0) {
528 gas.Address = sc->cpu_p_blk + 5;
529 cx_ptr->p_lvlx = acpi_bus_alloc_gas(sc->cpu_dev, &cpu_rid, &gas);
530 if (cx_ptr->p_lvlx != NULL) {
532 cx_ptr->type = ACPI_STATE_C3;
533 cx_ptr->trans_lat = AcpiGbl_FADT->Plvl3Lat;
541 /* If no valid registers were found, don't attach. */
542 if (sc->cpu_cx_count == 0)
545 /* Use initial sleep value of 1 sec. to start with lowest idle state. */
546 sc->cpu_prev_sleep = 1000000;
552 * Parse a _CST package and set up its Cx states. Since the _CST object
553 * can change dynamically, our notify handler may call this function
554 * to clean up and probe the new _CST package.
557 acpi_cpu_cx_cst(struct acpi_cpu_softc *sc)
559 struct acpi_cx *cx_ptr;
567 ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__);
570 buf.Length = ACPI_ALLOCATE_BUFFER;
571 status = AcpiEvaluateObject(sc->cpu_handle, "_CST", NULL, &buf);
572 if (ACPI_FAILURE(status))
575 /* _CST is a package with a count and at least one Cx package. */
576 top = (ACPI_OBJECT *)buf.Pointer;
577 if (!ACPI_PKG_VALID(top, 2) || acpi_PkgInt32(top, 0, &count) != 0) {
578 device_printf(sc->cpu_dev, "Invalid _CST package\n");
579 AcpiOsFree(buf.Pointer);
582 if (count != top->Package.Count - 1) {
583 device_printf(sc->cpu_dev, "Invalid _CST state count (%d != %d)\n",
584 count, top->Package.Count - 1);
585 count = top->Package.Count - 1;
587 if (count > MAX_CX_STATES) {
588 device_printf(sc->cpu_dev, "_CST has too many states (%d)\n", count);
589 count = MAX_CX_STATES;
592 /* Set up all valid states. */
593 sc->cpu_cx_count = 0;
594 cx_ptr = sc->cpu_cx_states;
595 for (i = 0; i < count; i++) {
596 pkg = &top->Package.Elements[i + 1];
597 if (!ACPI_PKG_VALID(pkg, 4) ||
598 acpi_PkgInt32(pkg, 1, &cx_ptr->type) != 0 ||
599 acpi_PkgInt32(pkg, 2, &cx_ptr->trans_lat) != 0 ||
600 acpi_PkgInt32(pkg, 3, &cx_ptr->power) != 0) {
602 device_printf(sc->cpu_dev, "Skipping invalid Cx state package\n");
606 /* Validate the state to see if we should use it. */
607 switch (cx_ptr->type) {
614 if (cx_ptr->trans_lat > 100) {
615 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
616 "acpi_cpu%d: C2[%d] not available.\n",
617 device_get_unit(sc->cpu_dev), i));
624 if (cx_ptr->trans_lat > 1000 ||
625 (cpu_quirks & CPU_QUIRK_NO_C3) != 0) {
627 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
628 "acpi_cpu%d: C3[%d] not available.\n",
629 device_get_unit(sc->cpu_dev), i));
636 /* Free up any previous register. */
637 if (cx_ptr->p_lvlx != NULL) {
638 bus_release_resource(sc->cpu_dev, 0, 0, cx_ptr->p_lvlx);
639 cx_ptr->p_lvlx = NULL;
643 /* Allocate the control register for C2 or C3. */
644 acpi_PkgGas(sc->cpu_dev, pkg, 0, &cpu_rid, &cx_ptr->p_lvlx);
645 if (cx_ptr->p_lvlx != NULL) {
647 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
648 "acpi_cpu%d: Got C%d - %d latency\n",
649 device_get_unit(sc->cpu_dev), cx_ptr->type,
655 AcpiOsFree(buf.Pointer);
661 * Call this *after* all CPUs have been attached.
664 acpi_cpu_startup(void *arg)
666 struct acpi_cpu_softc *sc;
669 /* Get set of CPU devices */
670 devclass_get_devices(acpi_cpu_devclass, &cpu_devices, &cpu_ndevices);
673 * Make sure all the processors' Cx counts match. We should probably
674 * also check the contents of each. However, no known systems have
675 * non-matching Cx counts so we'll deal with this later.
677 count = MAX_CX_STATES;
678 for (i = 0; i < cpu_ndevices; i++) {
679 sc = device_get_softc(cpu_devices[i]);
680 count = min(sc->cpu_cx_count, count);
682 cpu_cx_count = count;
684 /* Perform throttling and Cx final initialization. */
685 sc = device_get_softc(cpu_devices[0]);
686 if (sc->cpu_p_cnt != NULL)
687 acpi_cpu_startup_throttling();
688 if (cpu_cx_count > 0)
689 acpi_cpu_startup_cx();
693 * Takes the ACPI lock to avoid fighting anyone over the SMI command
697 acpi_cpu_startup_throttling()
701 /* Initialise throttling states */
702 cpu_throttle_max = CPU_MAX_SPEED;
703 cpu_throttle_state = CPU_MAX_SPEED;
705 SYSCTL_ADD_INT(&acpi_cpu_sysctl_ctx,
706 SYSCTL_CHILDREN(acpi_cpu_sysctl_tree),
707 OID_AUTO, "throttle_max", CTLFLAG_RD,
708 &cpu_throttle_max, 0, "maximum CPU speed");
709 SYSCTL_ADD_PROC(&acpi_cpu_sysctl_ctx,
710 SYSCTL_CHILDREN(acpi_cpu_sysctl_tree),
711 OID_AUTO, "throttle_state",
712 CTLTYPE_INT | CTLFLAG_RW, &cpu_throttle_state,
713 0, acpi_cpu_throttle_sysctl, "I", "current CPU speed");
715 /* If ACPI 2.0+, signal platform that we are taking over throttling. */
717 if (cpu_pstate_cnt != 0)
718 AcpiOsWritePort(cpu_smi_cmd, cpu_pstate_cnt, 8);
720 /* Set initial speed to maximum. */
721 acpi_cpu_throttle_set(cpu_throttle_max);
724 printf("acpi_cpu: throttling enabled, %d steps (100%% to %d.%d%%), "
725 "currently %d.%d%%\n", CPU_MAX_SPEED, CPU_SPEED_PRINTABLE(1),
726 CPU_SPEED_PRINTABLE(cpu_throttle_state));
730 extern void (*cpu_idle_hook)(void);
733 acpi_cpu_startup_cx()
735 struct acpi_cpu_softc *sc;
740 sc = device_get_softc(cpu_devices[0]);
741 sbuf_new(&sb, cpu_cx_supported, sizeof(cpu_cx_supported), SBUF_FIXEDLEN);
742 for (i = 0; i < cpu_cx_count; i++)
743 sbuf_printf(&sb, "C%d/%d ", i + 1, sc->cpu_cx_states[i].trans_lat);
746 SYSCTL_ADD_STRING(&acpi_cpu_sysctl_ctx,
747 SYSCTL_CHILDREN(acpi_cpu_sysctl_tree),
748 OID_AUTO, "cx_supported", CTLFLAG_RD, cpu_cx_supported,
749 0, "Cx/microsecond values for supported Cx states");
750 SYSCTL_ADD_PROC(&acpi_cpu_sysctl_ctx,
751 SYSCTL_CHILDREN(acpi_cpu_sysctl_tree),
752 OID_AUTO, "cx_lowest", CTLTYPE_STRING | CTLFLAG_RW,
753 NULL, 0, acpi_cpu_cx_lowest_sysctl, "A",
754 "lowest Cx sleep state to use");
755 SYSCTL_ADD_PROC(&acpi_cpu_sysctl_ctx,
756 SYSCTL_CHILDREN(acpi_cpu_sysctl_tree),
757 OID_AUTO, "cx_usage", CTLTYPE_STRING | CTLFLAG_RD,
758 NULL, 0, acpi_cpu_usage_sysctl, "A",
759 "percent usage for each Cx state");
762 /* Signal platform that we can handle _CST notification. */
763 if (cpu_cst_cnt != 0) {
765 AcpiOsWritePort(cpu_smi_cmd, cpu_cst_cnt, 8);
770 /* Take over idling from cpu_idle_default_hook(). */
772 /* XXX only set this if ncpus == 1, for now XXX */
773 cpu_idle_hook = acpi_cpu_idle;
777 * Set CPUs to the new state.
779 * Must be called with the ACPI lock held.
782 acpi_cpu_throttle_set(uint32_t speed)
784 struct acpi_cpu_softc *sc;
786 uint32_t p_cnt, clk_val;
790 /* Iterate over processors */
791 for (i = 0; i < cpu_ndevices; i++) {
792 sc = device_get_softc(cpu_devices[i]);
793 if (sc->cpu_p_cnt == NULL)
796 /* Get the current P_CNT value and disable throttling */
797 p_cnt = CPU_GET_REG(sc->cpu_p_cnt, 4);
798 p_cnt &= ~CPU_P_CNT_THT_EN;
799 CPU_SET_REG(sc->cpu_p_cnt, 4, p_cnt);
801 /* If we're at maximum speed, that's all */
802 if (speed < CPU_MAX_SPEED) {
803 /* Mask the old CLK_VAL off and or-in the new value */
804 clk_val = (CPU_MAX_SPEED - 1) << cpu_duty_offset;
806 p_cnt |= (speed << cpu_duty_offset);
808 /* Write the new P_CNT value and then enable throttling */
809 CPU_SET_REG(sc->cpu_p_cnt, 4, p_cnt);
810 p_cnt |= CPU_P_CNT_THT_EN;
811 CPU_SET_REG(sc->cpu_p_cnt, 4, p_cnt);
813 ACPI_VPRINT(sc->cpu_dev, acpi_device_get_parent_softc(sc->cpu_dev),
814 "set speed to %d.%d%%\n", CPU_SPEED_PRINTABLE(speed));
816 cpu_throttle_state = speed;
820 * Idle the CPU in the lowest state possible. This function is called with
821 * interrupts disabled. Note that once it re-enables interrupts, a task
822 * switch can occur so do not access shared data (i.e. the softc) after
823 * interrupts are re-enabled.
828 struct acpi_cpu_softc *sc;
829 struct acpi_cx *cx_next;
830 uint32_t start_time, end_time;
831 int bm_active, cx_next_idx, i;
833 /* If disabled, return immediately. */
834 if (cpu_cx_count == 0) {
840 * Look up our CPU id to get our softc. If it's NULL, we'll use C1
841 * since there is no ACPI processor object for this CPU. This occurs
842 * for logical CPUs in the HTT case.
844 sc = cpu_softc[mdcpu->mi.gd_cpuid];
851 * If we slept 100 us or more, use the lowest Cx state. Otherwise,
852 * find the lowest state that has a latency less than or equal to
853 * the length of our last sleep.
855 cx_next_idx = cpu_cx_lowest;
856 if (sc->cpu_prev_sleep < 100)
857 for (i = cpu_cx_lowest; i >= 0; i--)
858 if (sc->cpu_cx_states[i].trans_lat <= sc->cpu_prev_sleep) {
864 * Check for bus master activity. If there was activity, clear
865 * the bit and use the lowest non-C3 state. Note that the USB
866 * driver polling for new devices keeps this bit set all the
867 * time if USB is loaded.
869 AcpiGetRegister(ACPI_BITREG_BUS_MASTER_STATUS, &bm_active,
870 ACPI_MTX_DO_NOT_LOCK);
871 if (bm_active != 0) {
872 AcpiSetRegister(ACPI_BITREG_BUS_MASTER_STATUS, 1,
873 ACPI_MTX_DO_NOT_LOCK);
874 cx_next_idx = min(cx_next_idx, cpu_non_c3);
877 /* Select the next state and update statistics. */
878 cx_next = &sc->cpu_cx_states[cx_next_idx];
879 cpu_cx_stats[cx_next_idx]++;
880 KASSERT(cx_next->type != ACPI_STATE_C0, ("acpi_cpu_idle: C0 sleep"));
883 * Execute HLT (or equivalent) and wait for an interrupt. We can't
884 * calculate the time spent in C1 since the place we wake up is an
885 * ISR. Assume we slept one quantum and return.
887 if (cx_next->type == ACPI_STATE_C1) {
888 sc->cpu_prev_sleep = 1000000 / hz;
893 /* For C3, disable bus master arbitration and enable bus master wake. */
894 if (cx_next->type == ACPI_STATE_C3) {
895 AcpiSetRegister(ACPI_BITREG_ARB_DISABLE, 1, ACPI_MTX_DO_NOT_LOCK);
896 AcpiSetRegister(ACPI_BITREG_BUS_MASTER_RLD, 1, ACPI_MTX_DO_NOT_LOCK);
900 * Read from P_LVLx to enter C2(+), checking time spent asleep.
901 * Use the ACPI timer for measuring sleep time. Since we need to
902 * get the time very close to the CPU start/stop clock logic, this
903 * is the only reliable time source.
905 AcpiHwLowLevelRead(32, &start_time, &AcpiGbl_FADT->XPmTmrBlk);
906 CPU_GET_REG(cx_next->p_lvlx, 1);
909 * Read the end time twice. Since it may take an arbitrary time
910 * to enter the idle state, the first read may be executed before
911 * the processor has stopped. Doing it again provides enough
912 * margin that we are certain to have a correct value.
914 AcpiHwLowLevelRead(32, &end_time, &AcpiGbl_FADT->XPmTmrBlk);
915 AcpiHwLowLevelRead(32, &end_time, &AcpiGbl_FADT->XPmTmrBlk);
917 /* Enable bus master arbitration and disable bus master wakeup. */
918 if (cx_next->type == ACPI_STATE_C3) {
919 AcpiSetRegister(ACPI_BITREG_ARB_DISABLE, 0, ACPI_MTX_DO_NOT_LOCK);
920 AcpiSetRegister(ACPI_BITREG_BUS_MASTER_RLD, 0, ACPI_MTX_DO_NOT_LOCK);
923 /* Find the actual time asleep in microseconds, minus overhead. */
924 end_time = acpi_TimerDelta(end_time, start_time);
925 sc->cpu_prev_sleep = PM_USEC(end_time) - cx_next->trans_lat;
929 /* Put the CPU in C1 in a machine-dependant way. */
934 ia64_call_pal_static(PAL_HALT_LIGHT, 0, 0, 0);
936 __asm __volatile("sti; hlt");
941 * Re-evaluate the _PSS and _CST objects when we are notified that they
944 * XXX Re-evaluation disabled until locking is done.
947 acpi_cpu_notify(ACPI_HANDLE h, UINT32 notify, void *context)
949 struct acpi_cpu_softc *sc = (struct acpi_cpu_softc *)context;
952 case ACPI_CPU_NOTIFY_PERF_STATES:
953 device_printf(sc->cpu_dev, "Performance states changed\n");
954 /* acpi_cpu_px_available(sc); */
956 case ACPI_CPU_NOTIFY_CX_STATES:
957 device_printf(sc->cpu_dev, "Cx states changed\n");
958 /* acpi_cpu_cx_cst(sc); */
961 device_printf(sc->cpu_dev, "Unknown notify %#x\n", notify);
967 acpi_cpu_quirks(struct acpi_cpu_softc *sc)
971 * C3 is not supported on multiple CPUs since this would require
972 * flushing all caches which is currently too expensive.
975 cpu_quirks |= CPU_QUIRK_NO_C3;
978 /* Look for various quirks of the PIIX4 part. */
979 acpi_dev = pci_find_device(PCI_VENDOR_INTEL, PCI_DEVICE_82371AB_3);
980 if (acpi_dev != NULL) {
981 switch (pci_get_revid(acpi_dev)) {
983 * Disable throttling control on PIIX4 A and B-step.
984 * See specification changes #13 ("Manual Throttle Duty Cycle")
985 * and #14 ("Enabling and Disabling Manual Throttle"), plus
986 * erratum #5 ("STPCLK# Deassertion Time") from the January
987 * 2002 PIIX4 specification update. Note that few (if any)
988 * mobile systems ever used this part.
990 case PCI_REVISION_A_STEP:
991 case PCI_REVISION_B_STEP:
992 cpu_quirks |= CPU_QUIRK_NO_THROTTLE;
995 * Disable C3 support for all PIIX4 chipsets. Some of these parts
996 * do not report the BMIDE status to the BM status register and
997 * others have a livelock bug if Type-F DMA is enabled. Linux
998 * works around the BMIDE bug by reading the BM status directly
999 * but we take the simpler approach of disabling C3 for these
1002 * See erratum #18 ("C3 Power State/BMIDE and Type-F DMA
1003 * Livelock") from the January 2002 PIIX4 specification update.
1004 * Applies to all PIIX4 models.
1006 case PCI_REVISION_4E:
1007 case PCI_REVISION_4M:
1008 cpu_quirks |= CPU_QUIRK_NO_C3;
1019 /* Handle changes in the CPU throttling setting. */
1021 acpi_cpu_throttle_sysctl(SYSCTL_HANDLER_ARGS)
1028 argp = (uint32_t *)oidp->oid_arg1;
1030 error = sysctl_handle_int(oidp, &arg, 0, req);
1032 /* Error or no new value */
1033 if (error != 0 || req->newptr == NULL)
1035 if (arg < 1 || arg > cpu_throttle_max)
1038 /* If throttling changed, notify the BIOS of the new rate. */
1042 acpi_cpu_throttle_set(arg);
1050 acpi_cpu_usage_sysctl(SYSCTL_HANDLER_ARGS)
1055 uintmax_t fract, sum, whole;
1058 for (i = 0; i < cpu_cx_count; i++)
1059 sum += cpu_cx_stats[i];
1060 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
1061 for (i = 0; i < cpu_cx_count; i++) {
1063 whole = (uintmax_t)cpu_cx_stats[i] * 100;
1064 fract = (whole % sum) * 100;
1065 sbuf_printf(&sb, "%u.%02u%% ", (u_int)(whole / sum),
1066 (u_int)(fract / sum));
1068 sbuf_printf(&sb, "0%% ");
1072 sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
1079 acpi_cpu_cx_lowest_sysctl(SYSCTL_HANDLER_ARGS)
1081 struct acpi_cpu_softc *sc;
1085 sc = device_get_softc(cpu_devices[0]);
1086 snprintf(state, sizeof(state), "C%d", cpu_cx_lowest + 1);
1087 error = sysctl_handle_string(oidp, state, sizeof(state), req);
1088 if (error != 0 || req->newptr == NULL)
1090 if (strlen(state) < 2 || toupper(state[0]) != 'C')
1092 val = (int) strtol(state + 1, NULL, 10) - 1;
1093 if (val < 0 || val > cpu_cx_count - 1)
1096 cpu_cx_lowest = val;
1098 /* If not disabling, cache the new lowest non-C3 state. */
1100 for (i = cpu_cx_lowest; i >= 0; i--) {
1101 if (sc->cpu_cx_states[i].type < ACPI_STATE_C3) {
1107 /* Reset the statistics counters. */
1108 bzero(cpu_cx_stats, sizeof(cpu_cx_stats));