xref: /illumos-gate/usr/src/uts/common/os/sunpm.c (revision c6f039c7)

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */
/*
 * Copyright 2010 Nexenta Systems, Inc.  All rights reserved.
 */

/*
 * sunpm.c builds sunpm.o	"power management framework"
 *	kernel-resident power management code.  Implements power management
 *	policy
 *	Assumes: all backwards compat. device components wake up on &
 *		 the pm_info pointer in dev_info is initially NULL
 *
 * PM - (device) Power Management
 *
 * Each device may have 0 or more components.  If a device has no components,
 * then it can't be power managed.  Each component has 2 or more
 * power states.
 *
 * "Backwards Compatible" (bc) devices:
 * There are two different types of devices from the point of view of this
 * code.  The original type, left over from the original PM implementation on
 * the voyager platform are known in this code as "backwards compatible"
 * devices (PM_ISBC(dip) returns true).
 * They are recognized by the pm code by the lack of a pm-components property
 * and a call made by the driver to pm_create_components(9F).
 * For these devices, component 0 is special, and represents the power state
 * of the device.  If component 0 is to be set to power level 0 (off), then
 * the framework must first call into the driver's detach(9E) routine with
 * DDI_PM_SUSPEND, to get the driver to save the hardware state of the device.
 * After setting component 0 from 0 to a non-zero power level, a call must be
 * made into the driver's attach(9E) routine with DDI_PM_RESUME.
 *
 * Currently, the only way to get a bc device power managed is via a set of
 * ioctls (PM_DIRECT_PM, PM_SET_CURRENT_POWER) issued to /dev/pm.
 *
 * For non-bc devices, the driver describes the components by exporting a
 * pm-components(9P) property that tells how many components there are,
 * tells what each component's power state values are, and provides human
 * readable strings (currently unused) for each component name and power state.
 * Devices which export pm-components(9P) are automatically power managed
 * whenever autopm is enabled (via PM_START_PM ioctl issued by pmconfig(1M)
 * after parsing power.conf(4)). The exception to this rule is that power
 * manageable CPU devices may be automatically managed independently of autopm
 * by either enabling or disabling (via PM_START_CPUPM and PM_STOP_CPUPM
 * ioctls) cpupm. If the CPU devices are not managed independently, then they
 * are managed by autopm. In either case, for automatically power managed
 * devices, all components are considered independent of each other, and it is
 * up to the driver to decide when a transition requires saving or restoring
 * hardware state.
 *
 * Each device component also has a threshold time associated with each power
 * transition (see power.conf(4)), and a busy/idle state maintained by the
 * driver calling pm_idle_component(9F) and pm_busy_component(9F).
 * Components are created idle.
 *
 * The PM framework provides several functions:
 * -implement PM policy as described in power.conf(4)
 *  Policy is set by pmconfig(1M) issuing pm ioctls based on power.conf(4).
 *  Policies consist of:
 *    -set threshold values (defaults if none provided by pmconfig)
 *    -set dependencies among devices
 *    -enable/disable autopm
 *    -enable/disable cpupm
 *    -turn down idle components based on thresholds (if autopm or cpupm is
 *     enabled) (aka scanning)
 *    -maintain power states based on dependencies among devices
 *    -upon request, or when the frame buffer powers off, attempt to turn off
 *     all components that are idle or become idle over the next (10 sec)
 *     period in an attempt to get down to an EnergyStar compliant state
 *    -prevent powering off of a device which exported the
 *     pm-no-involuntary-power-cycles property without active involvement of
 *     the device's driver (so no removing power when the device driver is
 *     not attached)
 * -provide a mechanism for a device driver to request that a device's component
 *  be brought back to the power level necessary for the use of the device
 * -allow a process to directly control the power levels of device components
 *  (via ioctls issued to /dev/pm--see usr/src/uts/common/io/pm.c)
 * -ensure that the console frame buffer is powered up before being referenced
 *  via prom_printf() or other prom calls that might generate console output
 * -maintain implicit dependencies (e.g. parent must be powered up if child is)
 * -provide "backwards compatible" behavior for devices without pm-components
 *  property
 *
 * Scanning:
 * Whenever autopm or cpupm  is enabled, the framework attempts to bring each
 * component of each managed device to its lowest power based on the threshold
 * of idleness associated with each transition and the busy/idle state of the
 * component.
 *
 * The actual work of this is done by pm_scan_dev(), which cycles through each
 * component of a device, checking its idleness against its current threshold,
 * and calling pm_set_power() as appropriate to change the power level.
 * This function also indicates when it would next be profitable to scan the
 * device again, and a new scan is scheduled after that time.
 *
 * Dependencies:
 * It is possible to establish a dependency between the power states of two
 * otherwise unrelated devices.  This is currently done to ensure that the
 * cdrom is always up whenever the console framebuffer is up, so that the user
 * can insert a cdrom and see a popup as a result.
 *
 * The dependency terminology used in power.conf(4) is not easy to understand,
 * so we've adopted a different terminology in the implementation.  We write
 * of a "keeps up" and a "kept up" device.  A relationship can be established
 * where one device keeps up another.  That means that if the keepsup device
 * has any component that is at a non-zero power level, all components of the
 * "kept up" device must be brought to full power.  This relationship is
 * asynchronous.  When the keeping device is powered up, a request is queued
 * to a worker thread to bring up the kept device.  The caller does not wait.
 * Scan will not turn down a kept up device.
 *
 * Direct PM:
 * A device may be directly power managed by a process.  If a device is
 * directly pm'd, then it will not be scanned, and dependencies will not be
 * enforced.  * If a directly pm'd device's driver requests a power change (via
 * pm_raise_power(9F)), then the request is blocked and notification is sent
 * to the controlling process, which must issue the requested power change for
 * the driver to proceed.
 *
 */

#include <sys/types.h>
#include <sys/errno.h>
#include <sys/callb.h>		/* callback registration during CPR */
#include <sys/conf.h>		/* driver flags and functions */
#include <sys/open.h>		/* OTYP_CHR definition */
#include <sys/stat.h>		/* S_IFCHR definition */
#include <sys/pathname.h>	/* name -> dev_info xlation */
#include <sys/ddi_impldefs.h>	/* dev_info node fields */
#include <sys/kmem.h>		/* memory alloc stuff */
#include <sys/debug.h>
#include <sys/archsystm.h>
#include <sys/pm.h>
#include <sys/ddi.h>
#include <sys/sunddi.h>
#include <sys/sunndi.h>
#include <sys/sunpm.h>
#include <sys/epm.h>
#include <sys/vfs.h>
#include <sys/mode.h>
#include <sys/mkdev.h>
#include <sys/promif.h>
#include <sys/consdev.h>
#include <sys/esunddi.h>
#include <sys/modctl.h>
#include <sys/fs/ufs_fs.h>
#include <sys/note.h>
#include <sys/taskq.h>
#include <sys/bootconf.h>
#include <sys/reboot.h>
#include <sys/spl.h>
#include <sys/disp.h>
#include <sys/sobject.h>
#include <sys/sunmdi.h>
#include <sys/systm.h>
#include <sys/cpuvar.h>
#include <sys/cyclic.h>
#include <sys/uadmin.h>
#include <sys/srn.h>


/*
 * PM LOCKING
 *	The list of locks:
 * Global pm mutex locks.
 *
 * pm_scan_lock:
 *		It protects the timeout id of the scan thread, and the value
 *		of autopm_enabled and cpupm.  This lock is not held
 *		concurrently with any other PM locks.
 *
 * pm_clone_lock:	Protects the clone list and count of poll events
 *		pending for the pm driver.
 *		Lock ordering:
 *			pm_clone_lock -> pm_pscc_interest_rwlock,
 *			pm_clone_lock -> pm_pscc_direct_rwlock.
 *
 * pm_rsvp_lock:
 *		Used to synchronize the data structures used for processes
 *		to rendezvous with state change information when doing
 *		direct PM.
 *		Lock ordering:
 *			pm_rsvp_lock -> pm_pscc_interest_rwlock,
 *			pm_rsvp_lock -> pm_pscc_direct_rwlock,
 *			pm_rsvp_lock -> pm_clone_lock.
 *
 * ppm_lock:	protects the list of registered ppm drivers
 *		Lock ordering:
 *			ppm_lock -> ppm driver unit_lock
 *
 * pm_compcnt_lock:
 *		Protects count of components that are not at their lowest
 *		power level.
 *		Lock ordering:
 *			pm_compcnt_lock -> ppm_lock.
 *
 * pm_dep_thread_lock:
 *		Protects work list for pm_dep_thread.  Not taken concurrently
 *		with any other pm lock.
 *
 * pm_remdrv_lock:
 *		Serializes the operation of removing noinvol data structure
 *		entries for a branch of the tree when a driver has been
 *		removed from the system (modctl_rem_major).
 *		Lock ordering:
 *			pm_remdrv_lock -> pm_noinvol_rwlock.
 *
 * pm_cfb_lock: (High level spin lock)
 *		Protects the count of how many components of the console
 *		frame buffer are off (so we know if we have to bring up the
 *		console as a result of a prom_printf, etc.
 *		No other locks are taken while holding this lock.
 *
 * pm_loan_lock:
 *		Protects the lock_loan list.  List is used to record that one
 *		thread has acquired a power lock but has launched another thread
 *		to complete its processing.  An entry in the list indicates that
 *		the worker thread can borrow the lock held by the other thread,
 *		which must block on the completion of the worker.  Use is
 *		specific to module loading.
 *		No other locks are taken while holding this lock.
 *
 * Global PM rwlocks
 *
 * pm_thresh_rwlock:
 *		Protects the list of thresholds recorded for future use (when
 *		devices attach).
 *		Lock ordering:
 *			pm_thresh_rwlock -> devi_pm_lock
 *
 * pm_noinvol_rwlock:
 *		Protects list of detached nodes that had noinvol registered.
 *		No other PM locks are taken while holding pm_noinvol_rwlock.
 *
 * pm_pscc_direct_rwlock:
 *		Protects the list that maps devices being directly power
 *		managed to the processes that manage them.
 *		Lock ordering:
 *			pm_pscc_direct_rwlock -> psce_lock
 *
 * pm_pscc_interest_rwlock;
 *		Protects the list that maps state change events to processes
 *		that want to know about them.
 *		Lock ordering:
 *			pm_pscc_interest_rwlock -> psce_lock
 *
 * per-dip locks:
 *
 * Each node has these per-dip locks, which are only used if the device is
 * a candidate for power management (e.g. has pm components)
 *
 * devi_pm_lock:
 *		Protects all power management state of the node except for
 *		power level, which is protected by ndi_devi_enter().
 *		Encapsulated in macros PM_LOCK_DIP()/PM_UNLOCK_DIP().
 *		Lock ordering:
 *			devi_pm_lock -> pm_rsvp_lock,
 *			devi_pm_lock -> pm_dep_thread_lock,
 *			devi_pm_lock -> pm_noinvol_rwlock,
 *			devi_pm_lock -> power lock
 *
 * power lock (ndi_devi_enter()):
 *		Since changing power level is possibly a slow operation (30
 *		seconds to spin up a disk drive), this is locked separately.
 *		Since a call into the driver to change the power level of one
 *		component may result in a call back into the framework to change
 *		the power level of another, this lock allows re-entrancy by
 *		the same thread (ndi_devi_enter is used for this because
 *		the USB framework uses ndi_devi_enter in its power entry point,
 *		and use of any other lock would produce a deadlock.
 *
 * devi_pm_busy_lock:
 *		This lock protects the integrity of the busy count.  It is
 *		only taken by pm_busy_component() and pm_idle_component and
 *		some code that adjust the busy time after the timer gets set
 *		up or after a CPR operation.  It is per-dip to keep from
 *		single-threading all the disk drivers on a system.
 *		It could be per component instead, but most devices have
 *		only one component.
 *		No other PM locks are taken while holding this lock.
 *
 */

static int stdout_is_framebuffer;
static kmutex_t	e_pm_power_lock;
static kmutex_t pm_loan_lock;
kmutex_t	pm_scan_lock;
callb_id_t	pm_cpr_cb_id;
callb_id_t	pm_panic_cb_id;
callb_id_t	pm_halt_cb_id;
int		pm_comps_notlowest;	/* no. of comps not at lowest power */
int		pm_powering_down;	/* cpr is source of DDI_SUSPEND calls */

clock_t pm_id_ticks = 5;	/* ticks to wait before scan during idle-down */
clock_t pm_default_min_scan = PM_DEFAULT_MIN_SCAN;
clock_t pm_cpu_min_scan = PM_CPU_MIN_SCAN;

#define	PM_MIN_SCAN(dip)	(PM_ISCPU(dip) ? pm_cpu_min_scan : \
				    pm_default_min_scan)

static int pm_busop_set_power(dev_info_t *,
    void *, pm_bus_power_op_t, void *, void *);
static int pm_busop_match_request(dev_info_t *, void *);
static int pm_all_to_normal_nexus(dev_info_t *, pm_canblock_t);
static void e_pm_set_max_power(dev_info_t *, int, int);
static int e_pm_get_max_power(dev_info_t *, int);

/*
 * Dependency Processing is done thru a seperate thread.
 */
kmutex_t	pm_dep_thread_lock;
kcondvar_t	pm_dep_thread_cv;
pm_dep_wk_t	*pm_dep_thread_workq = NULL;
pm_dep_wk_t	*pm_dep_thread_tail = NULL;

/*
 * Autopm  must be turned on by a PM_START_PM ioctl, so we don't end up
 * power managing things in single user mode that have been suppressed via
 * power.conf entries.  Protected by pm_scan_lock.
 */
int		autopm_enabled;

/*
 * cpupm is turned on and off, by the PM_START_CPUPM and PM_STOP_CPUPM ioctls,
 * to define the power management behavior of CPU devices separate from
 * autopm. Protected by pm_scan_lock.
 */
pm_cpupm_t	cpupm = PM_CPUPM_NOTSET;

/*
 * Defines the default mode of operation for CPU power management,
 * either the polling implementation, or the event based dispatcher driven
 * implementation.
 */
pm_cpupm_t	cpupm_default_mode = PM_CPUPM_EVENT;

/*
 * AutoS3 depends on autopm being enabled, and must be enabled by
 * PM_START_AUTOS3 command.
 */
int		autoS3_enabled;

#if !defined(__sparc)
/*
 * on sparc these live in fillsysinfo.c
 *
 * If this variable is non-zero, cpr should return "not supported" when
 * it is queried even though it would normally be supported on this platform.
 */
int cpr_supported_override;

/*
 * Some platforms may need to support CPR even in the absence of
 * having the correct platform id information.  If this
 * variable is non-zero, cpr should proceed even in the absence
 * of otherwise being qualified.
 */
int cpr_platform_enable = 0;

#endif

/*
 * pm_S3_enabled indicates that we believe the platform can support S3,
 * which we get from pmconfig(1M)
 */
int		pm_S3_enabled;

/*
 * This flag is true while processes are stopped for a checkpoint/resume.
 * Controlling processes of direct pm'd devices are not available to
 * participate in power level changes, so we bypass them when this is set.
 */
static int	pm_processes_stopped;

#ifdef	DEBUG

/*
 * see common/sys/epm.h for PMD_* values
 */

uint_t		pm_debug = 0;

/*
 * If pm_divertdebug is set, then no prom_printf calls will be made by
 * PMD(), which will prevent debug output from bringing up the console
 * frame buffer.  Clearing this variable before setting pm_debug will result
 * in PMD output going to the console.
 *
 * pm_divertdebug is incremented in pm_set_power() if dip == cfb_dip to avoid
 * deadlocks and decremented at the end of pm_set_power()
 */
uint_t		pm_divertdebug = 1;
volatile uint_t pm_debug_to_console = 0;
kmutex_t	pm_debug_lock;		/* protects pm_divertdebug */

void prdeps(char *);
#endif

/* Globals */

/*
 * List of recorded thresholds and dependencies
 */
pm_thresh_rec_t *pm_thresh_head;
krwlock_t pm_thresh_rwlock;

pm_pdr_t *pm_dep_head;
static int pm_unresolved_deps = 0;
static int pm_prop_deps = 0;

/*
 * List of devices that exported no-involuntary-power-cycles property
 */
pm_noinvol_t *pm_noinvol_head;

/*
 * Locks used in noinvol processing
 */
krwlock_t pm_noinvol_rwlock;
kmutex_t pm_remdrv_lock;

int pm_default_idle_threshold = PM_DEFAULT_SYS_IDLENESS;
int pm_system_idle_threshold;
int pm_cpu_idle_threshold;

/*
 * By default nexus has 0 threshold, and depends on its children to keep it up
 */
int pm_default_nexus_threshold = 0;

/*
 * Data structures shared with common/io/pm.c
 */
kmutex_t	pm_clone_lock;
kcondvar_t	pm_clones_cv[PM_MAX_CLONE];
uint_t		pm_poll_cnt[PM_MAX_CLONE];	/* count of events for poll */
unsigned char	pm_interest[PM_MAX_CLONE];
struct pollhead	pm_pollhead;

/*
 * Data structures shared with common/io/srn.c
 */
kmutex_t	srn_clone_lock;		/* protects srn_signal, srn_inuse */
void (*srn_signal)(int type, int event);
int srn_inuse;				/* stop srn detach */

extern int	hz;
extern char	*platform_module_list[];

/*
 * Wrappers for use in ddi_walk_devs
 */

static int		pm_set_dev_thr_walk(dev_info_t *, void *);
static int		pm_restore_direct_lvl_walk(dev_info_t *, void *);
static int		pm_save_direct_lvl_walk(dev_info_t *, void *);
static int		pm_discard_dep_walk(dev_info_t *, void *);
#ifdef DEBUG
static int		pm_desc_pwrchk_walk(dev_info_t *, void *);
#endif

/*
 * Routines for managing noinvol devices
 */
int			pm_noinvol_update(int, int, int, char *, dev_info_t *);
void			pm_noinvol_update_node(dev_info_t *,
			    pm_bp_noinvol_t *req);

kmutex_t pm_rsvp_lock;
kmutex_t pm_compcnt_lock;
krwlock_t pm_pscc_direct_rwlock;
krwlock_t pm_pscc_interest_rwlock;

#define	PSC_INTEREST	0	/* belongs to interest psc list */
#define	PSC_DIRECT	1	/* belongs to direct psc list */

pscc_t *pm_pscc_interest;
pscc_t *pm_pscc_direct;

#define	PM_MAJOR(dip) ddi_driver_major(dip)
#define	PM_IS_NEXUS(dip) ((PM_MAJOR(dip) == DDI_MAJOR_T_NONE) ? 0 : \
	NEXUS_DRV(devopsp[PM_MAJOR(dip)]))
#define	POWERING_ON(old, new) ((old) == 0 && (new) != 0)
#define	POWERING_OFF(old, new) ((old) != 0 && (new) == 0)

#define	PM_INCR_NOTLOWEST(dip) {					\
	mutex_enter(&pm_compcnt_lock);					\
	if (!PM_IS_NEXUS(dip) ||					\
	    (DEVI(dip)->devi_pm_flags & (PMC_DEV_THRESH|PMC_COMP_THRESH))) {\
		if (pm_comps_notlowest == 0)				\
			pm_ppm_notify_all_lowest(dip, PM_NOT_ALL_LOWEST);\
		pm_comps_notlowest++;					\
		PMD(PMD_LEVEL, ("%s: %s@%s(%s#%d) incr notlowest->%d\n",\
		    pmf, PM_DEVICE(dip), pm_comps_notlowest))		\
	}								\
	mutex_exit(&pm_compcnt_lock);					\
}
#define	PM_DECR_NOTLOWEST(dip) {					\
	mutex_enter(&pm_compcnt_lock);					\
	if (!PM_IS_NEXUS(dip) ||					\
	    (DEVI(dip)->devi_pm_flags & (PMC_DEV_THRESH|PMC_COMP_THRESH))) {\
		ASSERT(pm_comps_notlowest);				\
		pm_comps_notlowest--;					\
		PMD(PMD_LEVEL, ("%s: %s@%s(%s#%d) decr notlowest to "	\
			    "%d\n", pmf, PM_DEVICE(dip), pm_comps_notlowest))\
		if (pm_comps_notlowest == 0)				\
			pm_ppm_notify_all_lowest(dip, PM_ALL_LOWEST);	\
	}								\
	mutex_exit(&pm_compcnt_lock);					\
}

/*
 * console frame-buffer power-management is not enabled when
 * debugging services are present.  to override, set pm_cfb_override
 * to non-zero.
 */
uint_t pm_cfb_comps_off = 0;	/* PM_LEVEL_UNKNOWN is considered on */
kmutex_t pm_cfb_lock;
int pm_cfb_enabled = 1;		/* non-zero allows pm of console frame buffer */
#ifdef DEBUG
int pm_cfb_override = 1;	/* non-zero allows pm of cfb with debuggers */
#else
int pm_cfb_override = 0;	/* non-zero allows pm of cfb with debuggers */
#endif

static dev_info_t *cfb_dip = 0;
static dev_info_t *cfb_dip_detaching = 0;
uint_t cfb_inuse = 0;
static ddi_softintr_t pm_soft_id;
static boolean_t pm_soft_pending;
int	pm_scans_disabled = 0;

/*
 * A structure to record the fact that one thread has borrowed a lock held
 * by another thread.  The context requires that the lender block on the
 * completion of the borrower.
 */
typedef struct lock_loan {
	struct lock_loan	*pmlk_next;
	kthread_t		*pmlk_borrower;
	kthread_t		*pmlk_lender;
	dev_info_t		*pmlk_dip;
} lock_loan_t;
static lock_loan_t lock_loan_head;	/* list head is a dummy element */

#ifdef	DEBUG
#ifdef	PMDDEBUG
#define	PMD_FUNC(func, name)	char *(func) = (name);
#else	/* !PMDDEBUG */
#define	PMD_FUNC(func, name)
#endif	/* PMDDEBUG */
#else	/* !DEBUG */
#define	PMD_FUNC(func, name)
#endif	/* DEBUG */


/*
 * Must be called before first device (including pseudo) attach
 */
void
pm_init_locks(void)
{
	mutex_init(&pm_scan_lock, NULL, MUTEX_DRIVER, NULL);
	mutex_init(&pm_rsvp_lock, NULL, MUTEX_DRIVER, NULL);
	mutex_init(&pm_compcnt_lock, NULL, MUTEX_DRIVER, NULL);
	mutex_init(&pm_dep_thread_lock, NULL, MUTEX_DRIVER, NULL);
	mutex_init(&pm_remdrv_lock, NULL, MUTEX_DRIVER, NULL);
	mutex_init(&pm_loan_lock, NULL, MUTEX_DRIVER, NULL);
	rw_init(&pm_thresh_rwlock, NULL, RW_DEFAULT, NULL);
	rw_init(&pm_noinvol_rwlock, NULL, RW_DEFAULT, NULL);
	cv_init(&pm_dep_thread_cv, NULL, CV_DEFAULT, NULL);
}

static int pm_reset_timestamps(dev_info_t *, void *);

static boolean_t
pm_cpr_callb(void *arg, int code)
{
	_NOTE(ARGUNUSED(arg))
	static int auto_save;
	static pm_cpupm_t cpupm_save;

	switch (code) {
	case CB_CODE_CPR_CHKPT:
		/*
		 * Cancel scan or wait for scan in progress to finish
		 * Other threads may be trying to restart the scan, so we
		 * have to keep at it unil it sticks
		 */
		mutex_enter(&pm_scan_lock);
		ASSERT(!pm_scans_disabled);
		pm_scans_disabled = 1;
		auto_save = autopm_enabled;
		autopm_enabled = 0;
		cpupm_save = cpupm;
		cpupm = PM_CPUPM_NOTSET;
		mutex_exit(&pm_scan_lock);
		ddi_walk_devs(ddi_root_node(), pm_scan_stop_walk, NULL);
		break;

	case CB_CODE_CPR_RESUME:
		ASSERT(!autopm_enabled);
		ASSERT(cpupm == PM_CPUPM_NOTSET);
		ASSERT(pm_scans_disabled);
		pm_scans_disabled = 0;
		/*
		 * Call pm_reset_timestamps to reset timestamps of each
		 * device to the time when the system is resumed so that their
		 * idleness can be re-calculated. That's to avoid devices from
		 * being powered down right after resume if the system was in
		 * suspended mode long enough.
		 */
		ddi_walk_devs(ddi_root_node(), pm_reset_timestamps, NULL);

		autopm_enabled = auto_save;
		cpupm = cpupm_save;
		/*
		 * If there is any auto-pm device, get the scanning
		 * going. Otherwise don't bother.
		 */
		ddi_walk_devs(ddi_root_node(), pm_rescan_walk, NULL);
		break;
	}
	return (B_TRUE);
}

/*
 * This callback routine is called when there is a system panic.  This function
 * exists for prototype matching.
 */
static boolean_t
pm_panic_callb(void *arg, int code)
{
	_NOTE(ARGUNUSED(arg, code))
	void pm_cfb_check_and_powerup(void);
	PMD(PMD_CFB, ("pm_panic_callb\n"))
	pm_cfb_check_and_powerup();
	return (B_TRUE);
}

static boolean_t
pm_halt_callb(void *arg, int code)
{
	_NOTE(ARGUNUSED(arg, code))
	return (B_TRUE);
}

static void pm_dep_thread(void);

/*
 * This needs to be called after the root and platform drivers are loaded
 * and be single-threaded with respect to driver attach/detach
 */
void
pm_init(void)
{
	PMD_FUNC(pmf, "pm_init")
	char **mod;
	extern pri_t minclsyspri;

	pm_comps_notlowest = 0;
	pm_system_idle_threshold = pm_default_idle_threshold;
	pm_cpu_idle_threshold = 0;

	pm_cpr_cb_id = callb_add(pm_cpr_callb, (void *)NULL,
	    CB_CL_CPR_PM, "pm_cpr");
	pm_panic_cb_id = callb_add(pm_panic_callb, (void *)NULL,
	    CB_CL_PANIC, "pm_panic");
	pm_halt_cb_id = callb_add(pm_halt_callb, (void *)NULL,
	    CB_CL_HALT, "pm_halt");

	/*
	 * Create a thread to do dependency processing.
	 */
	(void) thread_create(NULL, 0, (void (*)())pm_dep_thread, NULL, 0, &p0,
	    TS_RUN, minclsyspri);

	/*
	 * loadrootmodules already loaded these ppm drivers, now get them
	 * attached so they can claim the root drivers as they attach
	 */
	for (mod = platform_module_list; *mod; mod++) {
		if (i_ddi_attach_hw_nodes(*mod) != DDI_SUCCESS) {
			cmn_err(CE_WARN, "!cannot load platform pm driver %s\n",
			    *mod);
		} else {
			PMD(PMD_DHR, ("%s: %s (%s)\n", pmf, *mod,
			    ddi_major_to_name(ddi_name_to_major(*mod))))
		}
	}
}

/*
 * pm_scan_init - create pm scan data structure.  Called (if autopm or cpupm
 * enabled) when device becomes power managed or after a failed detach and
 * when autopm is started via PM_START_PM or PM_START_CPUPM ioctls, and after
 * a CPR resume to get all the devices scanning again.
 */
void
pm_scan_init(dev_info_t *dip)
{
	PMD_FUNC(pmf, "scan_init")
	pm_scan_t	*scanp;

	ASSERT(!PM_ISBC(dip));

	PM_LOCK_DIP(dip);
	scanp = PM_GET_PM_SCAN(dip);
	if (!scanp) {
		PMD(PMD_SCAN, ("%s: %s@%s(%s#%d): create scan data\n",
		    pmf, PM_DEVICE(dip)))
		scanp =  kmem_zalloc(sizeof (pm_scan_t), KM_SLEEP);
		DEVI(dip)->devi_pm_scan = scanp;
	} else if (scanp->ps_scan_flags & PM_SCAN_STOP) {
		PMD(PMD_SCAN, ("%s: %s@%s(%s#%d): "
		    "clear PM_SCAN_STOP flag\n", pmf, PM_DEVICE(dip)))
		scanp->ps_scan_flags &= ~PM_SCAN_STOP;
	}
	PM_UNLOCK_DIP(dip);
}

/*
 * pm_scan_fini - remove pm scan data structure when stopping pm on the device
 */
void
pm_scan_fini(dev_info_t *dip)
{
	PMD_FUNC(pmf, "scan_fini")
	pm_scan_t	*scanp;

	PMD(PMD_SCAN, ("%s: %s@%s(%s#%d)\n", pmf, PM_DEVICE(dip)))
	ASSERT(!PM_ISBC(dip));
	PM_LOCK_DIP(dip);
	scanp = PM_GET_PM_SCAN(dip);
	if (!scanp) {
		PM_UNLOCK_DIP(dip);
		return;
	}

	ASSERT(!scanp->ps_scan_id && !(scanp->ps_scan_flags &
	    (PM_SCANNING | PM_SCAN_DISPATCHED | PM_SCAN_AGAIN)));

	kmem_free(scanp, sizeof (pm_scan_t));
	DEVI(dip)->devi_pm_scan = NULL;
	PM_UNLOCK_DIP(dip);
}

/*
 * Given a pointer to a component struct, return the current power level
 * (struct contains index unless it is a continuous level).
 * Located here in hopes of getting both this and dev_is_needed into the
 * cache together
 */
static int
cur_power(pm_component_t *cp)
{
	if (cp->pmc_cur_pwr == PM_LEVEL_UNKNOWN)
		return (cp->pmc_cur_pwr);

	return (cp->pmc_comp.pmc_lvals[cp->pmc_cur_pwr]);
}

static char *
pm_decode_direction(int direction)
{
	switch (direction) {
	case PM_LEVEL_UPONLY:
		return ("up");

	case PM_LEVEL_EXACT:
		return ("exact");

	case PM_LEVEL_DOWNONLY:
		return ("down");

	default:
		return ("INVALID DIRECTION");
	}
}

char *
pm_decode_op(pm_bus_power_op_t op)
{
	switch (op) {
	case BUS_POWER_CHILD_PWRCHG:
		return ("CHILD_PWRCHG");
	case BUS_POWER_NEXUS_PWRUP:
		return ("NEXUS_PWRUP");
	case BUS_POWER_PRE_NOTIFICATION:
		return ("PRE_NOTIFICATION");
	case BUS_POWER_POST_NOTIFICATION:
		return ("POST_NOTIFICATION");
	case BUS_POWER_HAS_CHANGED:
		return ("HAS_CHANGED");
	case BUS_POWER_NOINVOL:
		return ("NOINVOL");
	default:
		return ("UNKNOWN OP");
	}
}

/*
 * Returns true if level is a possible (valid) power level for component
 */
int
e_pm_valid_power(dev_info_t *dip, int cmpt, int level)
{
	PMD_FUNC(pmf, "e_pm_valid_power")
	pm_component_t *cp = PM_CP(dip, cmpt);
	int i;
	int *ip = cp->pmc_comp.pmc_lvals;
	int limit = cp->pmc_comp.pmc_numlevels;

	if (level < 0)
		return (0);
	for (i = 0; i < limit; i++) {
		if (level == *ip++)
			return (1);
	}
#ifdef DEBUG
	if (pm_debug & PMD_FAIL) {
		ip = cp->pmc_comp.pmc_lvals;

		for (i = 0; i < limit; i++)
			PMD(PMD_FAIL, ("%s: index=%d, level=%d\n",
			    pmf, i, *ip++))
	}
#endif
	return (0);
}

static int pm_start(dev_info_t *dip);
/*
 * Returns true if device is pm'd (after calling pm_start if need be)
 */
int
e_pm_valid_info(dev_info_t *dip, pm_info_t **infop)
{
	pm_info_t *info;

	/*
	 * Check if the device is power managed if not.
	 * To make the common case (device is power managed already)
	 * fast, we check without the lock.  If device is not already
	 * power managed, then we take the lock and the long route through
	 * go get it managed.  Devices never go unmanaged until they
	 * detach.
	 */
	info = PM_GET_PM_INFO(dip);
	if (!info) {
		if (!DEVI_IS_ATTACHING(dip)) {
			return (0);
		}
		if (pm_start(dip) != DDI_SUCCESS) {
			return (0);
		}
		info = PM_GET_PM_INFO(dip);
	}
	ASSERT(info);
	if (infop != NULL)
		*infop = info;
	return (1);
}

int
e_pm_valid_comp(dev_info_t *dip, int cmpt, pm_component_t **cpp)
{
	if (cmpt >= 0 && cmpt < PM_NUMCMPTS(dip)) {
		if (cpp != NULL)
			*cpp = PM_CP(dip, cmpt);
		return (1);
	} else {
		return (0);
	}
}

/*
 * Internal guts of ddi_dev_is_needed and pm_raise/lower_power
 */
static int
dev_is_needed(dev_info_t *dip, int cmpt, int level, int direction)
{
	PMD_FUNC(pmf, "din")
	pm_component_t *cp;
	char *pathbuf;
	int result;

	ASSERT(direction == PM_LEVEL_UPONLY || direction == PM_LEVEL_DOWNONLY);
	if (!e_pm_valid_info(dip, NULL) || !e_pm_valid_comp(dip, cmpt, &cp) ||
	    !e_pm_valid_power(dip, cmpt, level))
		return (DDI_FAILURE);

	PMD(PMD_DIN, ("%s: %s@%s(%s#%d) cmpt=%d, dir=%s, new=%d, cur=%d\n",
	    pmf, PM_DEVICE(dip), cmpt, pm_decode_direction(direction),
	    level, cur_power(cp)))

	if (pm_set_power(dip, cmpt, level,  direction,
	    PM_CANBLOCK_BLOCK, 0, &result) != DDI_SUCCESS) {
		if (direction == PM_LEVEL_UPONLY) {
			pathbuf = kmem_alloc(MAXPATHLEN, KM_SLEEP);
			(void) ddi_pathname(dip, pathbuf);
			cmn_err(CE_WARN, "Device %s failed to power up.",
			    pathbuf);
			kmem_free(pathbuf, MAXPATHLEN);
		}
		PMD(PMD_DIN | PMD_FAIL, ("%s: %s@%s(%s#%d) [%d] %s->%d failed, "
		    "errno %d\n", pmf, PM_DEVICE(dip), cmpt,
		    pm_decode_direction(direction), level, result))
		return (DDI_FAILURE);
	}

	PMD(PMD_RESCAN | PMD_DIN, ("%s: pm_rescan %s@%s(%s#%d)\n", pmf,
	    PM_DEVICE(dip)))
	pm_rescan(dip);
	return (DDI_SUCCESS);
}

/*
 * We can get multiple pm_rescan() threads, if one of them discovers
 * that no scan is running at the moment, it kicks it into action.
 * Otherwise, it tells the current scanning thread to scan again when
 * it is done by asserting the PM_SCAN_AGAIN flag. The PM_SCANNING and
 * PM_SCAN_AGAIN flags are used to regulate scan, to make sure only one
 * thread at a time runs the pm_scan_dev() code.
 */
void
pm_rescan(void *arg)
{
	PMD_FUNC(pmf, "rescan")
	dev_info_t	*dip = (dev_info_t *)arg;
	pm_info_t	*info;
	pm_scan_t	*scanp;
	timeout_id_t	scanid;

	PMD(PMD_SCAN, ("%s: %s@%s(%s#%d)\n", pmf, PM_DEVICE(dip)))
	PM_LOCK_DIP(dip);
	info = PM_GET_PM_INFO(dip);
	scanp = PM_GET_PM_SCAN(dip);
	if (pm_scans_disabled || !PM_SCANABLE(dip) || !info || !scanp ||
	    (scanp->ps_scan_flags & PM_SCAN_STOP)) {
		PM_UNLOCK_DIP(dip);
		return;
	}
	if (scanp->ps_scan_flags & PM_SCANNING) {
		scanp->ps_scan_flags |= PM_SCAN_AGAIN;
		PM_UNLOCK_DIP(dip);
		return;
	} else if (scanp->ps_scan_id) {
		scanid = scanp->ps_scan_id;
		scanp->ps_scan_id = 0;
		PMD(PMD_SCAN, ("%s: %s@%s(%s#%d): cancel timeout scanid %lx\n",
		    pmf, PM_DEVICE(dip), (ulong_t)scanid))
		PM_UNLOCK_DIP(dip);
		(void) untimeout(scanid);
		PM_LOCK_DIP(dip);
	}

	/*
	 * Dispatching pm_scan during attach time is risky due to the fact that
	 * attach might soon fail and dip dissolved, and panic may happen while
	 * attempting to stop scan. So schedule a pm_rescan instead.
	 * (Note that if either of the first two terms are true, taskq_dispatch
	 * will not be invoked).
	 *
	 * Multiple pm_scan dispatching is unecessary and costly to keep track
	 * of. The PM_SCAN_DISPATCHED flag is used between pm_rescan and pm_scan
	 * to regulate the dispatching.
	 *
	 * Scan is stopped before the device is detached (in pm_detaching())
	 * but it may get re-started during the post_detach processing if the
	 * driver fails to detach.
	 */
	if (DEVI_IS_ATTACHING(dip) ||
	    (scanp->ps_scan_flags & PM_SCAN_DISPATCHED) ||
	    taskq_dispatch(system_taskq, pm_scan, (void *)dip, TQ_NOSLEEP) ==
	    TASKQID_INVALID) {
		PMD(PMD_SCAN, ("%s: %s@%s(%s#%d): attaching, pm_scan already "
		    "dispatched or dispatching failed\n", pmf, PM_DEVICE(dip)))
		if (scanp->ps_scan_id) {
			scanid = scanp->ps_scan_id;
			scanp->ps_scan_id = 0;
			PM_UNLOCK_DIP(dip);
			(void) untimeout(scanid);
			PM_LOCK_DIP(dip);
			if (scanp->ps_scan_id) {
				PMD(PMD_SCAN, ("%s: %s@%s(%s#%d): a competing "
				    "thread scheduled pm_rescan, scanid %lx\n",
				    pmf, PM_DEVICE(dip),
				    (ulong_t)scanp->ps_scan_id))
				PM_UNLOCK_DIP(dip);
				return;
			}
		}
		scanp->ps_scan_id = timeout(pm_rescan, (void *)dip,
		    (scanp->ps_idle_down ? pm_id_ticks :
		    (PM_MIN_SCAN(dip) * hz)));
		PMD(PMD_SCAN, ("%s: %s@%s(%s#%d): scheduled next pm_rescan, "
		    "scanid %lx\n", pmf, PM_DEVICE(dip),
		    (ulong_t)scanp->ps_scan_id))
	} else {
		PMD(PMD_SCAN, ("%s: dispatched pm_scan for %s@%s(%s#%d)\n",
		    pmf, PM_DEVICE(dip)))
		scanp->ps_scan_flags |= PM_SCAN_DISPATCHED;
	}
	PM_UNLOCK_DIP(dip);
}

void
pm_scan(void *arg)
{
	PMD_FUNC(pmf, "scan")
	dev_info_t	*dip = (dev_info_t *)arg;
	pm_scan_t	*scanp;
	time_t		nextscan;

	PMD(PMD_SCAN, ("%s: %s@%s(%s#%d)\n", pmf, PM_DEVICE(dip)))

	PM_LOCK_DIP(dip);
	scanp = PM_GET_PM_SCAN(dip);
	ASSERT(scanp && PM_GET_PM_INFO(dip));

	if (pm_scans_disabled || !PM_SCANABLE(dip) ||
	    (scanp->ps_scan_flags & PM_SCAN_STOP)) {
		scanp->ps_scan_flags &= ~(PM_SCAN_AGAIN | PM_SCAN_DISPATCHED);
		PM_UNLOCK_DIP(dip);
		return;
	}

	if (scanp->ps_idle_down) {
		/*
		 * make sure we remember idledown was in affect until
		 * we've completed the scan
		 */
		PMID_SET_SCANS(scanp->ps_idle_down)
		PMD(PMD_IDLEDOWN, ("%s: %s@%s(%s#%d): idledown starts "
		    "(pmid %x)\n", pmf, PM_DEVICE(dip), scanp->ps_idle_down))
	}

	/* possible having two threads running pm_scan() */
	if (scanp->ps_scan_flags & PM_SCANNING) {
		scanp->ps_scan_flags |= PM_SCAN_AGAIN;
		PMD(PMD_SCAN, ("%s: scanning, will scan %s@%s(%s#%d) again\n",
		    pmf, PM_DEVICE(dip)))
		scanp->ps_scan_flags &= ~PM_SCAN_DISPATCHED;
		PM_UNLOCK_DIP(dip);
		return;
	}

	scanp->ps_scan_flags |= PM_SCANNING;
	scanp->ps_scan_flags &= ~PM_SCAN_DISPATCHED;
	do {
		scanp->ps_scan_flags &= ~PM_SCAN_AGAIN;
		PM_UNLOCK_DIP(dip);
		nextscan = pm_scan_dev(dip);
		PM_LOCK_DIP(dip);
	} while (scanp->ps_scan_flags & PM_SCAN_AGAIN);

	ASSERT(scanp->ps_scan_flags & PM_SCANNING);
	scanp->ps_scan_flags &= ~PM_SCANNING;

	if (scanp->ps_idle_down) {
		scanp->ps_idle_down &= ~PMID_SCANS;
		PMD(PMD_IDLEDOWN, ("%s: %s@%s(%s#%d): idledown ends "
		    "(pmid %x)\n", pmf, PM_DEVICE(dip), scanp->ps_idle_down))
	}

	/* schedule for next idle check */
	if (nextscan != LONG_MAX) {
		if (nextscan > (LONG_MAX / hz))
			nextscan = (LONG_MAX - 1) / hz;
		if (scanp->ps_scan_id) {
			PMD(PMD_SCAN, ("%s: %s@%s(%s#%d): while scanning "
			    "another rescan scheduled scanid(%lx)\n", pmf,
			    PM_DEVICE(dip), (ulong_t)scanp->ps_scan_id))
			PM_UNLOCK_DIP(dip);
			return;
		} else if (!(scanp->ps_scan_flags & PM_SCAN_STOP)) {
			scanp->ps_scan_id = timeout(pm_rescan, (void *)dip,
			    (clock_t)(nextscan * hz));
			PMD(PMD_SCAN, ("%s: nextscan for %s@%s(%s#%d) in "
			    "%lx sec, scanid(%lx) \n", pmf, PM_DEVICE(dip),
			    (ulong_t)nextscan, (ulong_t)scanp->ps_scan_id))
		}
	}
	PM_UNLOCK_DIP(dip);
}

void
pm_get_timestamps(dev_info_t *dip, time_t *valuep)
{
	int components = PM_NUMCMPTS(dip);
	int i;

	ASSERT(components > 0);
	PM_LOCK_BUSY(dip);	/* so we get a consistent view */
	for (i = 0; i < components; i++) {
		valuep[i] = PM_CP(dip, i)->pmc_timestamp;
	}
	PM_UNLOCK_BUSY(dip);
}

/*
 * Returns true if device needs to be kept up because it exported the
 * "no-involuntary-power-cycles" property or we're pretending it did (console
 * fb case) or it is an ancestor of such a device and has used up the "one
 * free cycle" allowed when all such leaf nodes have voluntarily powered down
 * upon detach
 */
int
pm_noinvol(dev_info_t *dip)
{
	PMD_FUNC(pmf, "noinvol")

	/*
	 * This doesn't change over the life of a driver, so no locking needed
	 */
	if (PM_IS_CFB(dip)) {
		PMD(PMD_NOINVOL | PMD_CFB, ("%s: inhibits CFB %s@%s(%s#%d)\n",
		    pmf, PM_DEVICE(dip)))
		return (1);
	}
	/*
	 * Not an issue if no such kids
	 */
	if (DEVI(dip)->devi_pm_noinvolpm == 0) {
#ifdef DEBUG
		if (DEVI(dip)->devi_pm_volpmd != 0) {
			dev_info_t *pdip = dip;
			do {
				PMD(PMD_NOINVOL, ("%s: %s@%s(%s#%d) noinvol %d "
				    "volpmd %d\n", pmf, PM_DEVICE(pdip),
				    DEVI(pdip)->devi_pm_noinvolpm,
				    DEVI(pdip)->devi_pm_volpmd))
				pdip = ddi_get_parent(pdip);
			} while (pdip);
		}
#endif
		ASSERT(DEVI(dip)->devi_pm_volpmd == 0);
		return (0);
	}

	/*
	 * Since we now maintain the counts correct at every node, we no longer
	 * need to look up the tree.  An ancestor cannot use up the free cycle
	 * without the children getting their counts adjusted.
	 */

#ifdef	DEBUG
	if (DEVI(dip)->devi_pm_noinvolpm != DEVI(dip)->devi_pm_volpmd)
		PMD(PMD_NOINVOL, ("%s: (%d != %d) inhibits %s@%s(%s#%d)\n", pmf,
		    DEVI(dip)->devi_pm_noinvolpm, DEVI(dip)->devi_pm_volpmd,
		    PM_DEVICE(dip)))
#endif
	return (DEVI(dip)->devi_pm_noinvolpm != DEVI(dip)->devi_pm_volpmd);
}

static int	cur_threshold(dev_info_t *, int);
static int	pm_next_lower_power(pm_component_t *, int);

/*
 * This function performs the actual scanning of the device.
 * It attempts to power off the indicated device's components if they have
 * been idle and other restrictions are met.
 * pm_scan_dev calculates and returns when the next scan should happen for
 * this device.
 */
time_t
pm_scan_dev(dev_info_t *dip)
{
	PMD_FUNC(pmf, "scan_dev")
	pm_scan_t	*scanp;
	time_t		*timestamp, idletime, now, thresh;
	time_t		timeleft = 0;
#ifdef PMDDEBUG
	int		curpwr;
#endif
	int		i, nxtpwr, pwrndx, unused;
	size_t		size;
	pm_component_t	 *cp;
	dev_info_t	*pdip = ddi_get_parent(dip);
	int		circ;
	clock_t		min_scan = pm_default_min_scan;

	/*
	 * skip attaching device
	 */
	if (DEVI_IS_ATTACHING(dip)) {
		PMD(PMD_SCAN, ("%s: %s@%s(%s#%d) is attaching, timeleft(%lx)\n",
		    pmf, PM_DEVICE(dip), min_scan))
		return (min_scan);
	}

	PM_LOCK_DIP(dip);
	scanp = PM_GET_PM_SCAN(dip);
	min_scan = PM_MIN_SCAN(dip);
	ASSERT(scanp && PM_GET_PM_INFO(dip));

	PMD(PMD_SCAN, ("%s: [BEGIN %s@%s(%s#%d)]\n", pmf, PM_DEVICE(dip)))
	PMD(PMD_SCAN, ("%s: %s@%s(%s#%d): kuc is %d\n", pmf, PM_DEVICE(dip),
	    PM_KUC(dip)))

	/* no scan under the following conditions */
	if (pm_scans_disabled || !PM_SCANABLE(dip) ||
	    (scanp->ps_scan_flags & PM_SCAN_STOP) ||
	    (PM_KUC(dip) != 0) ||
	    PM_ISDIRECT(dip) || pm_noinvol(dip)) {
		PM_UNLOCK_DIP(dip);
		PMD(PMD_SCAN, ("%s: [END, %s@%s(%s#%d)] no scan, "
		    "scan_disabled(%d), apm_enabled(%d), cpupm(%d), "
		    "kuc(%d), %s directpm, %s pm_noinvol\n",
		    pmf, PM_DEVICE(dip), pm_scans_disabled, autopm_enabled,
		    cpupm, PM_KUC(dip),
		    PM_ISDIRECT(dip) ? "is" : "is not",
		    pm_noinvol(dip) ? "is" : "is not"))
		return (LONG_MAX);
	}
	PM_UNLOCK_DIP(dip);

	if (!ndi_devi_tryenter(pdip, &circ)) {
		PMD(PMD_SCAN, ("%s: %s@%s(%s#%d) can't hold pdip",
		    pmf, PM_DEVICE(pdip)))
		return ((time_t)1);
	}
	now = gethrestime_sec();
	size = PM_NUMCMPTS(dip) * sizeof (time_t);
	timestamp = kmem_alloc(size, KM_SLEEP);
	pm_get_timestamps(dip, timestamp);

	/*
	 * Since we removed support for backwards compatible devices,
	 * (see big comment at top of file)
	 * it is no longer required to deal with component 0 last.
	 */
	for (i = 0; i < PM_NUMCMPTS(dip); i++) {
		/*
		 * If already off (an optimization, perhaps)
		 */
		cp = PM_CP(dip, i);
		pwrndx = cp->pmc_cur_pwr;
#ifdef PMDDEBUG
		curpwr = (pwrndx == PM_LEVEL_UNKNOWN) ?
		    PM_LEVEL_UNKNOWN :
		    cp->pmc_comp.pmc_lvals[pwrndx];
#endif

		if (pwrndx == 0) {
			PMD(PMD_SCAN, ("%s: %s@%s(%s#%d) comp %d off or "
			    "lowest\n", pmf, PM_DEVICE(dip), i))
			/* skip device if off or at its lowest */
			continue;
		}

		thresh = cur_threshold(dip, i);		/* comp i threshold */
		if ((timestamp[i] == 0) || (cp->pmc_busycount > 0)) {
			/* were busy or newly became busy by another thread */
			if (timeleft == 0)
				timeleft = max(thresh, min_scan);
			else
				timeleft = min(
				    timeleft, max(thresh, min_scan));
			continue;
		}

		idletime = now - timestamp[i];		/* idle time */
		PMD(PMD_SCAN, ("%s: %s@%s(%s#%d) comp %d idle time %lx\n",
		    pmf, PM_DEVICE(dip), i, idletime))
		if (idletime >= thresh || PM_IS_PID(dip)) {
			nxtpwr = pm_next_lower_power(cp, pwrndx);
			PMD(PMD_SCAN, ("%s: %s@%s(%s#%d) comp %d, %d->%d\n",
			    pmf, PM_DEVICE(dip), i, curpwr, nxtpwr))
			if (pm_set_power(dip, i, nxtpwr, PM_LEVEL_DOWNONLY,
			    PM_CANBLOCK_FAIL, 1, &unused) != DDI_SUCCESS &&
			    PM_CURPOWER(dip, i) != nxtpwr) {
				PMD(PMD_SCAN, ("%s: %s@%s(%s#%d) comp %d, "
				    "%d->%d Failed\n", pmf, PM_DEVICE(dip),
				    i, curpwr, nxtpwr))
				timeleft = min_scan;
				continue;
			} else {
				PMD(PMD_SCAN, ("%s: %s@%s(%s#%d) comp %d, "
				    "%d->%d, GOOD curpwr %d\n", pmf,
				    PM_DEVICE(dip), i, curpwr, nxtpwr,
				    cur_power(cp)))

				if (nxtpwr == 0)	/* component went off */
					continue;

				/*
				 * scan to next lower level
				 */
				if (timeleft == 0)
					timeleft = max(
					    1, cur_threshold(dip, i));
				else
					timeleft = min(timeleft,
					    max(1, cur_threshold(dip, i)));
				PMD(PMD_SCAN, ("%s: %s@%s(%s#%d) comp %d, "
				    "timeleft(%lx)\n", pmf, PM_DEVICE(dip),
				    i, timeleft))
			}
		} else {	/* comp not idle long enough */
			if (timeleft == 0)
				timeleft = thresh - idletime;
			else
				timeleft = min(timeleft, (thresh - idletime));
			PMD(PMD_SCAN, ("%s: %s@%s(%s#%d) comp %d, timeleft="
			    "%lx\n", pmf, PM_DEVICE(dip), i, timeleft))
		}
	}
	ndi_devi_exit(pdip, circ);
	kmem_free(timestamp, size);
	PMD(PMD_SCAN, ("%s: [END %s@%s(%s#%d)] timeleft(%lx)\n", pmf,
	    PM_DEVICE(dip), timeleft))

	/*
	 * if components are already at lowest level, timeleft is left 0
	 */
	return ((timeleft == 0) ? LONG_MAX : timeleft);
}

/*
 * pm_scan_stop - cancel scheduled pm_rescan,
 *                wait for termination of dispatched pm_scan thread
 *                     and active pm_scan_dev thread.
 */
void
pm_scan_stop(dev_info_t *dip)
{
	PMD_FUNC(pmf, "scan_stop")
	pm_scan_t	*scanp;
	timeout_id_t	scanid;

	PMD(PMD_SCAN, ("%s: [BEGIN %s@%s(%s#%d)]\n", pmf, PM_DEVICE(dip)))
	PM_LOCK_DIP(dip);
	scanp = PM_GET_PM_SCAN(dip);
	if (!scanp) {
		PMD(PMD_SCAN, ("%s: [END %s@%s(%s#%d)] scan not initialized\n",
		    pmf, PM_DEVICE(dip)))
		PM_UNLOCK_DIP(dip);
		return;
	}
	scanp->ps_scan_flags |= PM_SCAN_STOP;

	/* cancel scheduled scan taskq */
	while (scanp->ps_scan_id) {
		scanid = scanp->ps_scan_id;
		scanp->ps_scan_id = 0;
		PM_UNLOCK_DIP(dip);
		(void) untimeout(scanid);
		PM_LOCK_DIP(dip);
	}

	while (scanp->ps_scan_flags & (PM_SCANNING | PM_SCAN_DISPATCHED)) {
		PM_UNLOCK_DIP(dip);
		delay(1);
		PM_LOCK_DIP(dip);
	}
	PM_UNLOCK_DIP(dip);
	PMD(PMD_SCAN, ("%s: [END %s@%s(%s#%d)]\n", pmf, PM_DEVICE(dip)))
}

int
pm_scan_stop_walk(dev_info_t *dip, void *arg)
{
	_NOTE(ARGUNUSED(arg))

	if (!PM_GET_PM_SCAN(dip))
		return (DDI_WALK_CONTINUE);
	ASSERT(!PM_ISBC(dip));
	pm_scan_stop(dip);
	return (DDI_WALK_CONTINUE);
}

/*
 * Converts a power level value to its index
 */
static int
power_val_to_index(pm_component_t *cp, int val)
{
	int limit, i, *ip;

	ASSERT(val != PM_LEVEL_UPONLY && val != PM_LEVEL_DOWNONLY &&
	    val != PM_LEVEL_EXACT);
	/*  convert power value into index (i) */
	limit = cp->pmc_comp.pmc_numlevels;
	ip = cp->pmc_comp.pmc_lvals;
	for (i = 0; i < limit; i++)
		if (val == *ip++)
			return (i);
	return (-1);
}

/*
 * Converts a numeric power level to a printable string
 */
static char *
power_val_to_string(pm_component_t *cp, int val)
{
	int index;

	if (val == PM_LEVEL_UPONLY)
		return ("<UPONLY>");

	if (val == PM_LEVEL_UNKNOWN ||
	    (index = power_val_to_index(cp, val)) == -1)
		return ("<LEVEL_UNKNOWN>");

	return (cp->pmc_comp.pmc_lnames[index]);
}

/*
 * Return true if this node has been claimed by a ppm.
 */
static int
pm_ppm_claimed(dev_info_t *dip)
{
	return (PPM(dip) != NULL);
}

/*
 * A node which was voluntarily power managed has just used up its "free cycle"
 * and need is volpmd field cleared, and the same done to all its descendents
 */
static void
pm_clear_volpm_dip(dev_info_t *dip)
{
	PMD_FUNC(pmf, "clear_volpm_dip")

	if (dip == NULL)
		return;
	PMD(PMD_NOINVOL, ("%s: clear volpm from %s@%s(%s#%d)\n", pmf,
	    PM_DEVICE(dip)))
	DEVI(dip)->devi_pm_volpmd = 0;
	for (dip = ddi_get_child(dip); dip; dip = ddi_get_next_sibling(dip)) {
		pm_clear_volpm_dip(dip);
	}
}

/*
 * A node which was voluntarily power managed has used up the "free cycles"
 * for the subtree that it is the root of.  Scan through the list of detached
 * nodes and adjust the counts of any that are descendents of the node.
 */
static void
pm_clear_volpm_list(dev_info_t *dip)
{
	PMD_FUNC(pmf, "clear_volpm_list")
	char	*pathbuf;
	size_t	len;
	pm_noinvol_t *ip;

	pathbuf = kmem_alloc(MAXPATHLEN, KM_SLEEP);
	(void) ddi_pathname(dip, pathbuf);
	len = strlen(pathbuf);
	PMD(PMD_NOINVOL, ("%s: clear volpm list %s\n", pmf, pathbuf))
	rw_enter(&pm_noinvol_rwlock, RW_WRITER);
	for (ip = pm_noinvol_head; ip; ip = ip->ni_next) {
		PMD(PMD_NOINVOL, ("%s: clear volpm: ni_path %s\n", pmf,
		    ip->ni_path))
		if (strncmp(pathbuf, ip->ni_path, len) == 0 &&
		    ip->ni_path[len] == '/') {
			PMD(PMD_NOINVOL, ("%s: clear volpm: %s\n", pmf,
			    ip->ni_path))
			ip->ni_volpmd = 0;
			ip->ni_wasvolpmd = 0;
		}
	}
	kmem_free(pathbuf, MAXPATHLEN);
	rw_exit(&pm_noinvol_rwlock);
}

/*
 * Powers a device, suspending or resuming the driver if it is a backward
 * compatible device, calling into ppm to change power level.
 * Called with the component's power lock held.
 */
static int
power_dev(dev_info_t *dip, int comp, int level, int old_level,
    pm_canblock_t canblock, pm_ppm_devlist_t **devlist)
{
	PMD_FUNC(pmf, "power_dev")
	power_req_t power_req;
	int		power_op_ret;	/* DDI_SUCCESS or DDI_FAILURE */
	int		resume_needed = 0;
	int		suspended = 0;
	int		result;
#ifdef PMDDEBUG
	struct pm_component *cp = PM_CP(dip, comp);
#endif
	int		bc = PM_ISBC(dip);
	int pm_all_components_off(dev_info_t *);
	int		clearvolpmd = 0;
	char		pathbuf[MAXNAMELEN];
#ifdef PMDDEBUG
	char *ppmname, *ppmaddr;
#endif
	/*
	 * If this is comp 0 of a backwards compat device and we are
	 * going to take the power away, we need to detach it with
	 * DDI_PM_SUSPEND command.
	 */
	if (bc && comp == 0 && POWERING_OFF(old_level, level)) {
		if (devi_detach(dip, DDI_PM_SUSPEND) != DDI_SUCCESS) {
			/* We could not suspend before turning cmpt zero off */
			PMD(PMD_ERROR, ("%s: could not suspend %s@%s(%s#%d)\n",
			    pmf, PM_DEVICE(dip)))
			return (DDI_FAILURE);
		} else {
			DEVI(dip)->devi_pm_flags |= PMC_SUSPENDED;
			suspended++;
		}
	}
	power_req.request_type = PMR_PPM_SET_POWER;
	power_req.req.ppm_set_power_req.who = dip;
	power_req.req.ppm_set_power_req.cmpt = comp;
	power_req.req.ppm_set_power_req.old_level = old_level;
	power_req.req.ppm_set_power_req.new_level = level;
	power_req.req.ppm_set_power_req.canblock = canblock;
	power_req.req.ppm_set_power_req.cookie = NULL;
#ifdef PMDDEBUG
	if (pm_ppm_claimed(dip)) {
		ppmname = PM_NAME(PPM(dip));
		ppmaddr = PM_ADDR(PPM(dip));

	} else {
		ppmname = "noppm";
		ppmaddr = "0";
	}
	PMD(PMD_PPM, ("%s: %s@%s(%s#%d):%s[%d] %s (%d) -> %s (%d) via %s@%s\n",
	    pmf, PM_DEVICE(dip), cp->pmc_comp.pmc_name, comp,
	    power_val_to_string(cp, old_level), old_level,
	    power_val_to_string(cp, level), level, ppmname, ppmaddr))
#endif
	/*
	 * If non-bc noinvolpm device is turning first comp on, or noinvolpm
	 * bc device comp 0 is powering on, then we count it as a power cycle
	 * against its voluntary count.
	 */
	if (DEVI(dip)->devi_pm_volpmd &&
	    (!bc && pm_all_components_off(dip) && level != 0) ||
	    (bc && comp == 0 && POWERING_ON(old_level, level)))
		clearvolpmd = 1;
	if ((power_op_ret = pm_ctlops(PPM(dip), dip, DDI_CTLOPS_POWER,
	    &power_req, &result)) == DDI_SUCCESS) {
		/*
		 * Now do involuntary pm accounting;  If we've just cycled power
		 * on a voluntarily pm'd node, and by inference on its entire
		 * subtree, we need to set the subtree (including those nodes
		 * already detached) volpmd counts to 0, and subtract out the
		 * value of the current node's volpmd count from the ancestors
		 */
		if (clearvolpmd) {
			int volpmd = DEVI(dip)->devi_pm_volpmd;
			pm_clear_volpm_dip(dip);
			pm_clear_volpm_list(dip);
			if (volpmd) {
				(void) ddi_pathname(dip, pathbuf);
				(void) pm_noinvol_update(PM_BP_NOINVOL_POWER,
				    volpmd, 0, pathbuf, dip);
			}
		}
	} else {
		PMD(PMD_FAIL, ("%s: can't set comp %d (%s) of %s@%s(%s#%d) "
		    "to level %d (%s)\n", pmf, comp, cp->pmc_comp.pmc_name,
		    PM_DEVICE(dip), level, power_val_to_string(cp, level)))
	}
	/*
	 * If some other devices were also powered up (e.g. other cpus in
	 * the same domain) return a pointer to that list
	 */
	if (devlist) {
		*devlist = (pm_ppm_devlist_t *)
		    power_req.req.ppm_set_power_req.cookie;
	}
	/*
	 * We will have to resume the device if the device is backwards compat
	 * device and either of the following is true:
	 * -This is comp 0 and we have successfully powered it up
	 * -This is comp 0 and we have failed to power it down. Resume is
	 *  needed because we have suspended it above
	 */

	if (bc && comp == 0) {
		ASSERT(PM_ISDIRECT(dip) || DEVI_IS_DETACHING(dip));
		if (power_op_ret == DDI_SUCCESS) {
			if (POWERING_ON(old_level, level)) {
				/*
				 * It must be either suspended or resumed
				 * via pm_power_has_changed path
				 */
				ASSERT((DEVI(dip)->devi_pm_flags &
				    PMC_SUSPENDED) ||
				    (PM_CP(dip, comp)->pmc_flags &
				    PM_PHC_WHILE_SET_POWER));

					resume_needed = suspended;
			}
		} else {
			if (POWERING_OFF(old_level, level)) {
				/*
				 * It must be either suspended or resumed
				 * via pm_power_has_changed path
				 */
				ASSERT((DEVI(dip)->devi_pm_flags &
				    PMC_SUSPENDED) ||
				    (PM_CP(dip, comp)->pmc_flags &
				    PM_PHC_WHILE_SET_POWER));

					resume_needed = suspended;
			}
		}
	}
	if (resume_needed) {
		ASSERT(DEVI(dip)->devi_pm_flags & PMC_SUSPENDED);
		/* ppm is not interested in DDI_PM_RESUME */
		if ((power_op_ret = devi_attach(dip, DDI_PM_RESUME)) ==
		    DDI_SUCCESS) {
			DEVI(dip)->devi_pm_flags &= ~PMC_SUSPENDED;
		} else
			cmn_err(CE_WARN, "!pm: Can't resume %s@%s(%s#%d)",
			    PM_DEVICE(dip));
	}
	return (power_op_ret);
}

/*
 * Return true if we are the owner or a borrower of the devi lock.  See
 * pm_lock_power_single() about borrowing the lock.
 */
static int
pm_devi_lock_held(dev_info_t *dip)
{
	lock_loan_t *cur;

	if (DEVI_BUSY_OWNED(dip))
		return (1);

	/* return false if no locks borrowed */
	if (lock_loan_head.pmlk_next == NULL)
		return (0);

	mutex_enter(&pm_loan_lock);
	/* see if our thread is registered as a lock borrower. */
	for (cur = lock_loan_head.pmlk_next; cur; cur = cur->pmlk_next)
		if (cur->pmlk_borrower == curthread)
			break;
	mutex_exit(&pm_loan_lock);

	return (cur != NULL && cur->pmlk_lender == DEVI(dip)->devi_busy_thread);
}

/*
 * pm_set_power: adjusts power level of device.	 Assumes device is power
 * manageable & component exists.
 *
 * Cases which require us to bring up devices we keep up ("wekeepups") for
 * backwards compatible devices:
 *	component 0 is off and we're bringing it up from 0
 *		bring up wekeepup first
 *	and recursively when component 0 is off and we bring some other
 *	component up from 0
 * For devices which are not backward compatible, our dependency notion is much
 * simpler.  Unless all components are off, then wekeeps must be on.
 * We don't treat component 0 differently.
 * Canblock tells how to deal with a direct pm'd device.
 * Scan arg tells us if we were called from scan, in which case we don't need
 * to go back to the root node and walk down to change power.
 */
int
pm_set_power(dev_info_t *dip, int comp, int level, int direction,
    pm_canblock_t canblock, int scan, int *retp)
{
	PMD_FUNC(pmf, "set_power")
	char		*pathbuf;
	pm_bp_child_pwrchg_t bpc;
	pm_sp_misc_t	pspm;
	int		ret = DDI_SUCCESS;
	int		unused = DDI_SUCCESS;
	dev_info_t	*pdip = ddi_get_parent(dip);

#ifdef DEBUG
	int		diverted = 0;

	/*
	 * This prevents operations on the console from calling prom_printf and
	 * either deadlocking or bringing up the console because of debug
	 * output
	 */
	if (dip == cfb_dip) {
		diverted++;
		mutex_enter(&pm_debug_lock);
		pm_divertdebug++;
		mutex_exit(&pm_debug_lock);
	}
#endif
	ASSERT(direction == PM_LEVEL_UPONLY || direction == PM_LEVEL_DOWNONLY ||
	    direction == PM_LEVEL_EXACT);
	PMD(PMD_SET, ("%s: %s@%s(%s#%d), comp=%d, dir=%s, new=%d\n",
	    pmf, PM_DEVICE(dip), comp, pm_decode_direction(direction), level))
	pathbuf = kmem_alloc(MAXPATHLEN, KM_SLEEP);
	(void) ddi_pathname(dip, pathbuf);
	bpc.bpc_dip = dip;
	bpc.bpc_path = pathbuf;
	bpc.bpc_comp = comp;
	bpc.bpc_olevel = PM_CURPOWER(dip, comp);
	bpc.bpc_nlevel = level;
	pspm.pspm_direction = direction;
	pspm.pspm_errnop = retp;
	pspm.pspm_canblock = canblock;
	pspm.pspm_scan = scan;
	bpc.bpc_private = &pspm;

	/*
	 * If a config operation is being done (we've locked the parent) or
	 * we already hold the power lock (we've locked the node)
	 * then we can operate directly on the node because we have already
	 * brought up all the ancestors, otherwise, we have to go back to the
	 * top of the tree.
	 */
	if (pm_devi_lock_held(pdip) || pm_devi_lock_held(dip))
		ret = pm_busop_set_power(dip, NULL, BUS_POWER_CHILD_PWRCHG,
		    (void *)&bpc, (void *)&unused);
	else
		ret = pm_busop_bus_power(ddi_root_node(), NULL,
		    BUS_POWER_CHILD_PWRCHG, (void *)&bpc, (void *)&unused);
#ifdef DEBUG
	if (ret != DDI_SUCCESS || *retp != DDI_SUCCESS) {
		PMD(PMD_ERROR, ("%s: %s@%s(%s#%d) can't change power, ret=%d, "
		    "errno=%d\n", pmf, PM_DEVICE(dip), ret, *retp))
	}
	if (diverted) {
		mutex_enter(&pm_debug_lock);
		pm_divertdebug--;
		mutex_exit(&pm_debug_lock);
	}
#endif
	kmem_free(pathbuf, MAXPATHLEN);
	return (ret);
}

/*
 * If holddip is set, then if a dip is found we return with the node held.
 *
 * This code uses the same locking scheme as e_ddi_hold_devi_by_path
 * (resolve_pathname), but it does not drive attach.
 */
dev_info_t *
pm_name_to_dip(char *pathname, int holddip)
{
	struct pathname pn;
	char		*component;
	dev_info_t	*parent, *child;
	int		circ;

	if ((pathname == NULL) || (*pathname != '/'))
		return (NULL);

	/* setup pathname and allocate component */
	if (pn_get(pathname, UIO_SYSSPACE, &pn))
		return (NULL);
	component = kmem_alloc(MAXNAMELEN, KM_SLEEP);

	/* start at top, process '/' component */
	parent = child = ddi_root_node();
	ndi_hold_devi(parent);
	pn_skipslash(&pn);
	ASSERT(i_ddi_devi_attached(parent));

	/* process components of pathname */
	while (pn_pathleft(&pn)) {
		(void) pn_getcomponent(&pn, component);

		/* enter parent and search for component child */
		ndi_devi_enter(parent, &circ);
		child = ndi_devi_findchild(parent, component);
		if ((child == NULL) || !i_ddi_devi_attached(child)) {
			child = NULL;
			ndi_devi_exit(parent, circ);
			ndi_rele_devi(parent);
			goto out;
		}

		/* attached child found, hold child and release parent */
		ndi_hold_devi(child);
		ndi_devi_exit(parent, circ);
		ndi_rele_devi(parent);

		/* child becomes parent, and process next component */
		parent = child;
		pn_skipslash(&pn);

		/* loop with active ndi_devi_hold of child->parent */
	}

out:
	pn_free(&pn);
	kmem_free(component, MAXNAMELEN);

	/* if we are not asked to return with hold, drop current hold */
	if (child && !holddip)
		ndi_rele_devi(child);
	return (child);
}

/*
 * Search for a dependency and mark it unsatisfied
 */
static void
pm_unsatisfy(char *keeper, char *kept)
{
	PMD_FUNC(pmf, "unsatisfy")
	pm_pdr_t *dp;

	PMD(PMD_KEEPS, ("%s: keeper=%s, kept=%s\n", pmf, keeper, kept))
	for (dp = pm_dep_head; dp; dp = dp->pdr_next) {
		if (!dp->pdr_isprop) {
			if (strcmp(dp->pdr_keeper, keeper) == 0 &&
			    (dp->pdr_kept_count > 0) &&
			    strcmp(dp->pdr_kept_paths[0], kept) == 0) {
				if (dp->pdr_satisfied) {
					dp->pdr_satisfied = 0;
					pm_unresolved_deps++;
					PMD(PMD_KEEPS, ("%s: clear satisfied, "
					    "pm_unresolved_deps now %d\n", pmf,
					    pm_unresolved_deps))
				}
			}
		}
	}
}

/*
 * Device dip is being un power managed, it keeps up count other devices.
 * We need to release any hold we have on the kept devices, and also
 * mark the dependency no longer satisfied.
 */
static void
pm_unkeeps(int count, char *keeper, char **keptpaths, int pwr)
{
	PMD_FUNC(pmf, "unkeeps")
	int i, j;
	dev_info_t *kept;
	dev_info_t *dip;
	struct pm_component *cp;
	int keeper_on = 0, circ;

	PMD(PMD_KEEPS, ("%s: count=%d, keeper=%s, keptpaths=%p\n", pmf, count,
	    keeper, (void *)keptpaths))
	/*
	 * Try to grab keeper. Keeper may have gone away by now,
	 * in this case, used the passed in value pwr
	 */
	dip = pm_name_to_dip(keeper, 1);
	for (i = 0; i < count; i++) {
		/* Release power hold */
		kept = pm_name_to_dip(keptpaths[i], 1);
		if (kept) {
			PMD(PMD_KEEPS, ("%s: %s@%s(%s#%d)[%d]\n", pmf,
			    PM_DEVICE(kept), i))
			/*
			 * We need to check if we skipped a bringup here
			 * because we could have failed the bringup
			 * (ie DIRECT PM device) and have
			 * not increment the count.
			 */
			if ((dip != NULL) && (PM_GET_PM_INFO(dip) != NULL)) {
				keeper_on = 0;
				PM_LOCK_POWER(dip, &circ);
				for (j = 0; j < PM_NUMCMPTS(dip); j++) {
					cp = &DEVI(dip)->devi_pm_components[j];
					if (cur_power(cp)) {
						keeper_on++;
						break;
					}
				}
				if (keeper_on && (PM_SKBU(kept) == 0)) {
					pm_rele_power(kept);
					DEVI(kept)->devi_pm_flags
					    &= ~PMC_SKIP_BRINGUP;
				}
				PM_UNLOCK_POWER(dip, circ);
			} else if (pwr) {
				if (PM_SKBU(kept) == 0) {
					pm_rele_power(kept);
					DEVI(kept)->devi_pm_flags
					    &= ~PMC_SKIP_BRINGUP;
				}
			}
			ddi_release_devi(kept);
		}
		/*
		 * mark this dependency not satisfied
		 */
		pm_unsatisfy(keeper, keptpaths[i]);
	}
	if (dip)
		ddi_release_devi(dip);
}

/*
 * Device kept is being un power managed, it is kept up by keeper.
 * We need to mark the dependency no longer satisfied.
 */
static void
pm_unkepts(char *kept, char *keeper)
{
	PMD_FUNC(pmf, "unkepts")
	PMD(PMD_KEEPS, ("%s: kept=%s, keeper=%s\n", pmf, kept, keeper))
	ASSERT(keeper != NULL);
	/*
	 * mark this dependency not satisfied
	 */
	pm_unsatisfy(keeper, kept);
}

/*
 * Removes dependency information and hold on the kepts, if the path is a
 * path of a keeper.
 */
static void
pm_free_keeper(char *path, int pwr)
{
	pm_pdr_t *dp;
	int i;
	size_t length;

	for (dp = pm_dep_head; dp; dp = dp->pdr_next) {
		if (strcmp(dp->pdr_keeper, path) != 0)
			continue;
		/*
		 * Remove all our kept holds and the dependency records,
		 * then free up the kept lists.
		 */
		pm_unkeeps(dp->pdr_kept_count, path, dp->pdr_kept_paths, pwr);
		if (dp->pdr_kept_count)  {
			for (i = 0; i < dp->pdr_kept_count; i++) {
				length = strlen(dp->pdr_kept_paths[i]);
				kmem_free(dp->pdr_kept_paths[i], length + 1);
			}
			kmem_free(dp->pdr_kept_paths,
			    dp->pdr_kept_count * sizeof (char **));
			dp->pdr_kept_paths = NULL;
			dp->pdr_kept_count = 0;
		}
	}
}

/*
 * Removes the device represented by path from the list of kepts, if the
 * path is a path of a kept
 */
static void
pm_free_kept(char *path)
{
	pm_pdr_t *dp;
	int i;
	int j, count;
	size_t length;
	char **paths;

	paths = NULL;
	for (dp = pm_dep_head; dp; dp = dp->pdr_next) {
		if (dp->pdr_kept_count == 0)
			continue;
		count = dp->pdr_kept_count;
		/* Remove this device from the kept path lists */
		for (i = 0; i < count; i++) {
			if (strcmp(dp->pdr_kept_paths[i], path) == 0) {
				pm_unkepts(path, dp->pdr_keeper);
				length = strlen(dp->pdr_kept_paths[i]) + 1;
				kmem_free(dp->pdr_kept_paths[i], length);
				dp->pdr_kept_paths[i] = NULL;
				dp->pdr_kept_count--;
			}
		}
		/* Compact the kept paths array */
		if (dp->pdr_kept_count) {
			length = dp->pdr_kept_count * sizeof (char **);
			paths = kmem_zalloc(length, KM_SLEEP);
			j = 0;
			for (i = 0; i < count; i++) {
				if (dp->pdr_kept_paths[i] != NULL) {
					paths[j] = dp->pdr_kept_paths[i];
					j++;
				}
			}
			ASSERT(j == dp->pdr_kept_count);
		}
		/* Now free the old array and point to the new one */
		kmem_free(dp->pdr_kept_paths, count * sizeof (char **));
		dp->pdr_kept_paths = paths;
	}
}

/*
 * Free the dependency information for a device.
 */
void
pm_free_keeps(char *path, int pwr)
{
	PMD_FUNC(pmf, "free_keeps")

#ifdef DEBUG
	int doprdeps = 0;
	void prdeps(char *);

	PMD(PMD_KEEPS, ("%s: %s\n", pmf, path))
	if (pm_debug & PMD_KEEPS) {
		doprdeps = 1;
		prdeps("pm_free_keeps before");
	}
#endif
	/*
	 * First assume we are a keeper and remove all our kepts.
	 */
	pm_free_keeper(path, pwr);
	/*
	 * Now assume we a kept device, and remove all our records.
	 */
	pm_free_kept(path);
#ifdef	DEBUG
	if (doprdeps) {
		prdeps("pm_free_keeps after");
	}
#endif
}

static int
pm_is_kept(char *path)
{
	pm_pdr_t *dp;
	int i;

	for (dp = pm_dep_head; dp; dp = dp->pdr_next) {
		if (dp->pdr_kept_count == 0)
			continue;
		for (i = 0; i < dp->pdr_kept_count; i++) {
			if (strcmp(dp->pdr_kept_paths[i], path) == 0)
				return (1);
		}
	}
	return (0);
}

static void
e_pm_hold_rele_power(dev_info_t *dip, int cnt)
{
	PMD_FUNC(pmf, "hold_rele_power")
	int circ;

	if ((dip == NULL) ||
	    (PM_GET_PM_INFO(dip) == NULL) || PM_ISBC(dip))
		return;

	PM_LOCK_POWER(dip, &circ);
	ASSERT(cnt >= 0 && PM_KUC(dip) >= 0 || cnt < 0 && PM_KUC(dip) > 0);
	PMD(PMD_KIDSUP, ("%s: kidsupcnt for %s@%s(%s#%d) %d->%d\n", pmf,
	    PM_DEVICE(dip), PM_KUC(dip), (PM_KUC(dip) + cnt)))

	PM_KUC(dip) += cnt;

	ASSERT(PM_KUC(dip) >= 0);
	PM_UNLOCK_POWER(dip, circ);

	if (cnt < 0 && PM_KUC(dip) == 0)
		pm_rescan(dip);
}

#define	MAX_PPM_HANDLERS	4

kmutex_t ppm_lock;	/* in case we ever do multi-threaded startup */

struct	ppm_callbacks {
	int (*ppmc_func)(dev_info_t *);
	dev_info_t	*ppmc_dip;
} ppm_callbacks[MAX_PPM_HANDLERS + 1];


/*
 * This routine calls into all the registered ppms to notify them
 * that either all components of power-managed devices are at their
 * lowest levels or no longer all are at their lowest levels.
 */
static void
pm_ppm_notify_all_lowest(dev_info_t *dip, int mode)
{
	struct ppm_callbacks *ppmcp;
	power_req_t power_req;
	int result = 0;

	power_req.request_type = PMR_PPM_ALL_LOWEST;
	power_req.req.ppm_all_lowest_req.mode = mode;
	mutex_enter(&ppm_lock);
	for (ppmcp = ppm_callbacks; ppmcp->ppmc_func; ppmcp++)
		(void) pm_ctlops((dev_info_t *)ppmcp->ppmc_dip, dip,
		    DDI_CTLOPS_POWER, &power_req, &result);
	mutex_exit(&ppm_lock);
	if (mode == PM_ALL_LOWEST) {
		if (autoS3_enabled) {
			PMD(PMD_SX, ("pm_ppm_notify_all_lowest triggering "
			    "autos3\n"))
			mutex_enter(&srn_clone_lock);
			if (srn_signal) {
				srn_inuse++;
				PMD(PMD_SX, ("(*srn_signal)(AUTOSX, 3)\n"))
				(*srn_signal)(SRN_TYPE_AUTOSX, 3);
				srn_inuse--;
			} else {
				PMD(PMD_SX, ("srn_signal NULL\n"))
			}
			mutex_exit(&srn_clone_lock);
		} else {
			PMD(PMD_SX, ("pm_ppm_notify_all_lowest autos3 "
			    "disabled\n"));
		}
	}
}

static void
pm_set_pm_info(dev_info_t *dip, void *value)
{
	DEVI(dip)->devi_pm_info = value;
}

pm_rsvp_t *pm_blocked_list;

/*
 * Look up an entry in the blocked list by dip and component
 */
static pm_rsvp_t *
pm_rsvp_lookup(dev_info_t *dip, int comp)
{
	pm_rsvp_t *p;
	ASSERT(MUTEX_HELD(&pm_rsvp_lock));
	for (p = pm_blocked_list; p; p = p->pr_next)
		if (p->pr_dip == dip && p->pr_comp == comp) {
			return (p);
		}
	return (NULL);
}

/*
 * Called when a device which is direct power managed (or the parent or
 * dependent of such a device) changes power, or when a pm clone is closed
 * that was direct power managing a device.  This call results in pm_blocked()
 * (below) returning.
 */
void
pm_proceed(dev_info_t *dip, int cmd, int comp, int newlevel)
{
	PMD_FUNC(pmf, "proceed")
	pm_rsvp_t *found = NULL;
	pm_rsvp_t *p;

	mutex_enter(&pm_rsvp_lock);
	switch (cmd) {
	/*
	 * we're giving up control, let any pending op continue
	 */
	case PMP_RELEASE:
		for (p = pm_blocked_list; p; p = p->pr_next) {
			if (dip == p->pr_dip) {
				p->pr_retval = PMP_RELEASE;
				PMD(PMD_DPM, ("%s: RELEASE %s@%s(%s#%d)\n",
				    pmf, PM_DEVICE(dip)))
				cv_signal(&p->pr_cv);
			}
		}
		break;

	/*
	 * process has done PM_SET_CURRENT_POWER; let a matching request
	 * succeed and a non-matching request for the same device fail
	 */
	case PMP_SETPOWER:
		found = pm_rsvp_lookup(dip, comp);
		if (!found)	/* if driver not waiting */
			break;
		/*
		 * This cannot be pm_lower_power, since that can only happen
		 * during detach or probe
		 */
		if (found->pr_newlevel <= newlevel) {
			found->pr_retval = PMP_SUCCEED;
			PMD(PMD_DPM, ("%s: SUCCEED %s@%s(%s#%d)\n", pmf,
			    PM_DEVICE(dip)))
		} else {
			found->pr_retval = PMP_FAIL;
			PMD(PMD_DPM, ("%s: FAIL %s@%s(%s#%d)\n", pmf,
			    PM_DEVICE(dip)))
		}
		cv_signal(&found->pr_cv);
		break;

	default:
		panic("pm_proceed unknown cmd %d", cmd);
	}
	mutex_exit(&pm_rsvp_lock);
}

/*
 * This routine dispatches new work to the dependency thread. Caller must
 * be prepared to block for memory if necessary.
 */
void
pm_dispatch_to_dep_thread(int cmd, char *keeper, char *kept, int wait,
    int *res, int cached_pwr)
{
	pm_dep_wk_t	*new_work;

	new_work = kmem_zalloc(sizeof (pm_dep_wk_t), KM_SLEEP);
	new_work->pdw_type = cmd;
	new_work->pdw_wait = wait;
	new_work->pdw_done = 0;
	new_work->pdw_ret = 0;
	new_work->pdw_pwr = cached_pwr;
	cv_init(&new_work->pdw_cv, NULL, CV_DEFAULT, NULL);
	if (keeper != NULL) {
		new_work->pdw_keeper = kmem_zalloc(strlen(keeper) + 1,
		    KM_SLEEP);
		(void) strcpy(new_work->pdw_keeper, keeper);
	}
	if (kept != NULL) {
		new_work->pdw_kept = kmem_zalloc(strlen(kept) + 1, KM_SLEEP);
		(void) strcpy(new_work->pdw_kept, kept);
	}
	mutex_enter(&pm_dep_thread_lock);
	if (pm_dep_thread_workq == NULL) {
		pm_dep_thread_workq = new_work;
		pm_dep_thread_tail = new_work;
		new_work->pdw_next = NULL;
	} else {
		pm_dep_thread_tail->pdw_next = new_work;
		pm_dep_thread_tail = new_work;
		new_work->pdw_next = NULL;
	}
	cv_signal(&pm_dep_thread_cv);
	/* If caller asked for it, wait till it is done. */
	if (wait)  {
		while (!new_work->pdw_done)
			cv_wait(&new_work->pdw_cv, &pm_dep_thread_lock);
		/*
		 * Pass return status, if any, back.
		 */
		if (res != NULL)
			*res = new_work->pdw_ret;
		/*
		 * If we asked to wait, it is our job to free the request
		 * structure.
		 */
		if (new_work->pdw_keeper)
			kmem_free(new_work->pdw_keeper,
			    strlen(new_work->pdw_keeper) + 1);
		if (new_work->pdw_kept)
			kmem_free(new_work->pdw_kept,
			    strlen(new_work->pdw_kept) + 1);
		kmem_free(new_work, sizeof (pm_dep_wk_t));
	}
	mutex_exit(&pm_dep_thread_lock);
}

/*
 * Release the pm resource for this device.
 */
void
pm_rem_info(dev_info_t *dip)
{
	PMD_FUNC(pmf, "rem_info")
	int		i, count = 0;
	pm_info_t	*info = PM_GET_PM_INFO(dip);
	dev_info_t	*pdip = ddi_get_parent(dip);
	char		*pathbuf;
	int		work_type = PM_DEP_WK_DETACH;

	ASSERT(info);

	ASSERT(!PM_IAM_LOCKING_DIP(dip));
	if (PM_ISDIRECT(dip)) {
		info->pmi_dev_pm_state &= ~PM_DIRECT;
		ASSERT(info->pmi_clone);
		info->pmi_clone = 0;
		pm_proceed(dip, PMP_RELEASE, -1, -1);
	}
	ASSERT(!PM_GET_PM_SCAN(dip));

	/*
	 * Now adjust parent's kidsupcnt.  BC nodes we check only comp 0,
	 * Others we check all components.  BC node that has already
	 * called pm_destroy_components() has zero component count.
	 * Parents that get notification are not adjusted because their
	 * kidsupcnt is always 0 (or 1 during configuration).
	 */
	PMD(PMD_KEEPS, ("%s: %s@%s(%s#%d) has %d components\n", pmf,
	    PM_DEVICE(dip), PM_NUMCMPTS(dip)))

	/* node is detached, so we can examine power without locking */
	if (PM_ISBC(dip)) {
		count = (PM_CURPOWER(dip, 0) != 0);
	} else {
		for (i = 0; i < PM_NUMCMPTS(dip); i++)
			count += (PM_CURPOWER(dip, i) != 0);
	}

	if (PM_NUMCMPTS(dip) && pdip && !PM_WANTS_NOTIFICATION(pdip))
		e_pm_hold_rele_power(pdip, -count);

	/* Schedule a request to clean up dependency records */
	pathbuf = kmem_zalloc(MAXPATHLEN, KM_SLEEP);
	(void) ddi_pathname(dip, pathbuf);
	pm_dispatch_to_dep_thread(work_type, pathbuf, pathbuf,
	    PM_DEP_NOWAIT, NULL, (count > 0));
	kmem_free(pathbuf, MAXPATHLEN);

	/*
	 * Adjust the pm_comps_notlowest count since this device is
	 * not being power-managed anymore.
	 */
	for (i = 0; i < PM_NUMCMPTS(dip); i++) {
		pm_component_t *cp = PM_CP(dip, i);
		if (cp->pmc_cur_pwr != 0)
			PM_DECR_NOTLOWEST(dip)
	}
	/*
	 * Once we clear the info pointer, it looks like it is not power
	 * managed to everybody else.
	 */
	pm_set_pm_info(dip, NULL);
	kmem_free(info, sizeof (pm_info_t));
}

int
pm_get_norm_pwrs(dev_info_t *dip, int **valuep, size_t *length)
{
	int components = PM_NUMCMPTS(dip);
	int *bufp;
	size_t size;
	int i;

	if (components <= 0) {
		cmn_err(CE_NOTE, "!pm: %s@%s(%s#%d) has no components, "
		    "can't get normal power values\n", PM_DEVICE(dip));
		return (DDI_FAILURE);
	} else {
		size = components * sizeof (int);
		bufp = kmem_alloc(size, KM_SLEEP);
		for (i = 0; i < components; i++) {
			bufp[i] = pm_get_normal_power(dip, i);
		}
	}
	*length = size;
	*valuep = bufp;
	return (DDI_SUCCESS);
}

static int
pm_reset_timestamps(dev_info_t *dip, void *arg)
{
	_NOTE(ARGUNUSED(arg))

	int components;
	int	i;

	if (!PM_GET_PM_INFO(dip))
		return (DDI_WALK_CONTINUE);
	components = PM_NUMCMPTS(dip);
	ASSERT(components > 0);
	PM_LOCK_BUSY(dip);
	for (i = 0; i < components; i++) {
		struct pm_component *cp;
		/*
		 * If the component was not marked as busy,
		 * reset its timestamp to now.
		 */
		cp = PM_CP(dip, i);
		if (cp->pmc_timestamp)
			cp->pmc_timestamp = gethrestime_sec();
	}
	PM_UNLOCK_BUSY(dip);
	return (DDI_WALK_CONTINUE);
}

/*
 * Convert a power level to an index into the levels array (or
 * just PM_LEVEL_UNKNOWN in that special case).
 */
static int
pm_level_to_index(dev_info_t *dip, pm_component_t *cp, int level)
{
	PMD_FUNC(pmf, "level_to_index")
	int i;
	int limit = cp->pmc_comp.pmc_numlevels;
	int *ip = cp->pmc_comp.pmc_lvals;

	if (level == PM_LEVEL_UNKNOWN)
		return (level);

	for (i = 0; i < limit; i++) {
		if (level == *ip++) {
			PMD(PMD_LEVEL, ("%s: %s@%s(%s#%d)[%d] to %x\n",
			    pmf, PM_DEVICE(dip),
			    (int)(cp - DEVI(dip)->devi_pm_components), level))
			return (i);
		}
	}
	panic("pm_level_to_index: level %d not found for device "
	    "%s@%s(%s#%d)", level, PM_DEVICE(dip));
	/*NOTREACHED*/
}

/*
 * Internal function to set current power level
 */
static void
e_pm_set_cur_pwr(dev_info_t *dip, pm_component_t *cp, int level)
{
	PMD_FUNC(pmf, "set_cur_pwr")
	int curpwr = (cp->pmc_flags & PM_PHC_WHILE_SET_POWER ?
	    cp->pmc_phc_pwr : cp->pmc_cur_pwr);

	/*
	 * Nothing to adjust if current & new levels are the same.
	 */
	if (curpwr != PM_LEVEL_UNKNOWN &&
	    level == cp->pmc_comp.pmc_lvals[curpwr])
		return;

	/*
	 * Keep the count for comps doing transition to/from lowest
	 * level.
	 */
	if (curpwr == 0) {
		PM_INCR_NOTLOWEST(dip);
	} else if (level == cp->pmc_comp.pmc_lvals[0]) {
		PM_DECR_NOTLOWEST(dip);
	}
	cp->pmc_phc_pwr = PM_LEVEL_UNKNOWN;
	cp->pmc_cur_pwr = pm_level_to_index(dip, cp, level);
}

static int pm_phc_impl(dev_info_t *, int, int, int);

/*
 * This is the default method of setting the power of a device if no ppm
 * driver has claimed it.
 */
int
pm_power(dev_info_t *dip, int comp, int level)
{
	PMD_FUNC(pmf, "power")
	struct dev_ops	*ops;
	int		(*fn)(dev_info_t *, int, int);
	struct pm_component *cp = PM_CP(dip, comp);
	int retval;
	pm_info_t *info = PM_GET_PM_INFO(dip);

	PMD(PMD_KIDSUP, ("%s: %s@%s(%s#%d), comp=%d, level=%d\n", pmf,
	    PM_DEVICE(dip), comp, level))
	if (!(ops = ddi_get_driver(dip))) {
		PMD(PMD_FAIL, ("%s: %s@%s(%s#%d) has no ops\n", pmf,
		    PM_DEVICE(dip)))
		return (DDI_FAILURE);
	}
	if ((ops->devo_rev < 2) || !(fn = ops->devo_power)) {
		PMD(PMD_FAIL, ("%s: %s%s\n", pmf,
		    (ops->devo_rev < 2 ? " wrong devo_rev" : ""),
		    (!fn ? " devo_power NULL" : "")))
		return (DDI_FAILURE);
	}
	cp->pmc_flags |= PM_POWER_OP;
	retval = (*fn)(dip, comp, level);
	cp->pmc_flags &= ~PM_POWER_OP;
	if (retval == DDI_SUCCESS) {
		e_pm_set_cur_pwr(dip, PM_CP(dip, comp), level);
		return (DDI_SUCCESS);
	}

	/*
	 * If pm_power_has_changed() detected a deadlock with pm_power() it
	 * updated only the power level of the component.  If our attempt to
	 * set the device new to a power level above has failed we sync the
	 * total power state via phc code now.
	 */
	if (cp->pmc_flags & PM_PHC_WHILE_SET_POWER) {
		int phc_lvl =
		    cp->pmc_comp.pmc_lvals[cp->pmc_cur_pwr];

		ASSERT(info);
		(void) pm_phc_impl(dip, comp, phc_lvl, 0);
		PMD(PMD_PHC, ("%s: phc %s@%s(%s#%d) comp=%d level=%d\n",
		    pmf, PM_DEVICE(dip), comp, phc_lvl))
	}

	PMD(PMD_FAIL, ("%s: can't set comp=%d (%s) of %s@%s(%s#%d) to "
	    "level=%d (%s)\n", pmf, comp, cp->pmc_comp.pmc_name, PM_DEVICE(dip),
	    level, power_val_to_string(cp, level)));
	return (DDI_FAILURE);
}

int
pm_unmanage(dev_info_t *dip)
{
	PMD_FUNC(pmf, "unmanage")
	power_req_t power_req;
	int result, retval = 0;

	ASSERT(!PM_IAM_LOCKING_DIP(dip));
	PMD(PMD_REMDEV | PMD_KIDSUP, ("%s: %s@%s(%s#%d)\n", pmf,
	    PM_DEVICE(dip)))
	power_req.request_type = PMR_PPM_UNMANAGE;
	power_req.req.ppm_config_req.who = dip;
	if (pm_ppm_claimed(dip))
		retval = pm_ctlops(PPM(dip), dip, DDI_CTLOPS_POWER,
		    &power_req, &result);
#ifdef DEBUG
	else
		retval = pm_ctlops(PPM(dip), dip, DDI_CTLOPS_POWER,
		    &power_req, &result);
#endif
	ASSERT(retval == DDI_SUCCESS);
	pm_rem_info(dip);
	return (retval);
}

int
pm_raise_power(dev_info_t *dip, int comp, int level)
{
	if (level < 0)
		return (DDI_FAILURE);
	if (!e_pm_valid_info(dip, NULL) || !e_pm_valid_comp(dip, comp, NULL) ||
	    !e_pm_valid_power(dip, comp, level))
		return (DDI_FAILURE);

	return (dev_is_needed(dip, comp, level, PM_LEVEL_UPONLY));
}

int
pm_lower_power(dev_info_t *dip, int comp, int level)
{
	PMD_FUNC(pmf, "pm_lower_power")

	if (!e_pm_valid_info(dip, NULL) || !e_pm_valid_comp(dip, comp, NULL) ||
	    !e_pm_valid_power(dip, comp, level)) {
		PMD(PMD_FAIL, ("%s: validation checks failed for %s@%s(%s#%d) "
		    "comp=%d level=%d\n", pmf, PM_DEVICE(dip), comp, level))
		return (DDI_FAILURE);
	}

	if (!DEVI_IS_DETACHING(dip)) {
		PMD(PMD_FAIL, ("%s: %s@%s(%s#%d) not detaching\n",
		    pmf, PM_DEVICE(dip)))
		return (DDI_FAILURE);
	}

	/*
	 * If we don't care about saving power, or we're treating this node
	 * specially, then this is a no-op
	 */
	if (!PM_SCANABLE(dip) || pm_noinvol(dip)) {
		PMD(PMD_FAIL, ("%s: %s@%s(%s#%d) %s%s%s%s\n",
		    pmf, PM_DEVICE(dip),
		    !autopm_enabled ? "!autopm_enabled " : "",
		    !PM_POLLING_CPUPM ? "!cpupm_polling " : "",
		    PM_CPUPM_DISABLED ? "cpupm_disabled " : "",
		    pm_noinvol(dip) ? "pm_noinvol()" : ""))
		return (DDI_SUCCESS);
	}

	if (dev_is_needed(dip, comp, level, PM_LEVEL_DOWNONLY) != DDI_SUCCESS) {
		PMD(PMD_FAIL, ("%s: %s@%s(%s#%d) dev_is_needed failed\n", pmf,
		    PM_DEVICE(dip)))
		return (DDI_FAILURE);
	}
	return (DDI_SUCCESS);
}

/*
 * Find the entries struct for a given dip in the blocked list, return it locked
 */
static psce_t *
pm_psc_dip_to_direct(dev_info_t *dip, pscc_t **psccp)
{
	pscc_t *p;
	psce_t *psce;

	rw_enter(&pm_pscc_direct_rwlock, RW_READER);
	for (p = pm_pscc_direct; p; p = p->pscc_next) {
		if (p->pscc_dip == dip) {
			*psccp = p;
			psce = p->pscc_entries;
			mutex_enter(&psce->psce_lock);
			ASSERT(psce);
			rw_exit(&pm_pscc_direct_rwlock);
			return (psce);
		}
	}
	rw_exit(&pm_pscc_direct_rwlock);
	panic("sunpm: no entry for dip %p in direct list", (void *)dip);
	/*NOTREACHED*/
}

/*
 * Write an entry indicating a power level change (to be passed to a process
 * later) in the given psce.
 * If we were called in the path that brings up the console fb in the
 * case of entering the prom, we don't want to sleep.  If the alloc fails, then
 * we create a record that has a size of -1, a physaddr of NULL, and that
 * has the overflow flag set.
 */
static int
psc_entry(ushort_t event, psce_t *psce, dev_info_t *dip, int comp, int new,
    int old, int which, pm_canblock_t canblock)
{
	char	buf[MAXNAMELEN];
	pm_state_change_t *p;
	size_t	size;
	caddr_t physpath = NULL;
	int	overrun = 0;

	ASSERT(MUTEX_HELD(&psce->psce_lock));
	(void) ddi_pathname(dip, buf);
	size = strlen(buf) + 1;
	p = psce->psce_in;
	if (canblock == PM_CANBLOCK_BYPASS) {
		physpath = kmem_alloc(size, KM_NOSLEEP);
		if (physpath == NULL) {
			/*
			 * mark current entry as overrun
			 */
			p->flags |= PSC_EVENT_LOST;
			size = (size_t)-1;
		}
	} else
		physpath = kmem_alloc(size, KM_SLEEP);
	if (p->size) {	/* overflow; mark the next entry */
		if (p->size != (size_t)-1)
			kmem_free(p->physpath, p->size);
		ASSERT(psce->psce_out == p);
		if (p == psce->psce_last) {
			psce->psce_first->flags |= PSC_EVENT_LOST;
			psce->psce_out = psce->psce_first;
		} else {
			(p + 1)->flags |= PSC_EVENT_LOST;
			psce->psce_out = (p + 1);
		}
		overrun++;
	} else if (physpath == NULL) {	/* alloc failed, mark this entry */
		p->flags |= PSC_EVENT_LOST;
		p->size = 0;
		p->physpath = NULL;
	}
	if (which == PSC_INTEREST) {
		mutex_enter(&pm_compcnt_lock);
		if (pm_comps_notlowest == 0)
			p->flags |= PSC_ALL_LOWEST;
		else
			p->flags &= ~PSC_ALL_LOWEST;
		mutex_exit(&pm_compcnt_lock);
	}
	p->event = event;
	p->timestamp = gethrestime_sec();
	p->component = comp;
	p->old_level = old;
	p->new_level = new;
	p->physpath = physpath;
	p->size = size;
	if (physpath != NULL)
		(void) strcpy(p->physpath, buf);
	if (p == psce->psce_last)
		psce->psce_in = psce->psce_first;
	else
		psce->psce_in = ++p;
	mutex_exit(&psce->psce_lock);
	return (overrun);
}

/*
 * Find the next entry on the interest list.  We keep a pointer to the item we
 * last returned in the user's cooke.  Returns a locked entries struct.
 */
static psce_t *
psc_interest(void **cookie, pscc_t **psccp)
{
	pscc_t *pscc;
	pscc_t **cookiep = (pscc_t **)cookie;

	if (*cookiep == NULL)
		pscc = pm_pscc_interest;
	else
		pscc = (*cookiep)->pscc_next;
	if (pscc) {
		*cookiep = pscc;
		*psccp = pscc;
		mutex_enter(&pscc->pscc_entries->psce_lock);
		return (pscc->pscc_entries);
	} else {
		return (NULL);
	}
}

/*
 * Create an entry for a process to pick up indicating a power level change.
 */
static void
pm_enqueue_notify(ushort_t cmd, dev_info_t *dip, int comp,
    int newlevel, int oldlevel, pm_canblock_t canblock)
{
	PMD_FUNC(pmf, "enqueue_notify")
	pscc_t	*pscc;
	psce_t	*psce;
	void		*cookie = NULL;
	int	overrun;

	ASSERT(MUTEX_HELD(&pm_rsvp_lock));
	switch (cmd) {
	case PSC_PENDING_CHANGE:	/* only for controlling process */
		PMD(PMD_DPM, ("%s: PENDING %s@%s(%s#%d), comp %d, %d -> %d\n",
		    pmf, PM_DEVICE(dip), comp, oldlevel, newlevel))
		psce = pm_psc_dip_to_direct(dip, &pscc);
		ASSERT(psce);
		PMD(PMD_IOCTL, ("%s: PENDING: %s@%s(%s#%d) pm_poll_cnt[%d] "
		    "%d\n", pmf, PM_DEVICE(dip), pscc->pscc_clone,
		    pm_poll_cnt[pscc->pscc_clone]))
		overrun = psc_entry(cmd, psce, dip, comp, newlevel, oldlevel,
		    PSC_DIRECT, canblock);
		PMD(PMD_DPM, ("%s: sig %d\n", pmf, pscc->pscc_clone))
		mutex_enter(&pm_clone_lock);
		if (!overrun)
			pm_poll_cnt[pscc->pscc_clone]++;
		cv_signal(&pm_clones_cv[pscc->pscc_clone]);
		pollwakeup(&pm_pollhead, (POLLRDNORM | POLLIN));
		mutex_exit(&pm_clone_lock);
		break;
	case PSC_HAS_CHANGED:
		PMD(PMD_DPM, ("%s: HAS %s@%s(%s#%d), comp %d, %d -> %d\n",
		    pmf, PM_DEVICE(dip), comp, oldlevel, newlevel))
		if (PM_ISDIRECT(dip) && canblock != PM_CANBLOCK_BYPASS) {
			psce = pm_psc_dip_to_direct(dip, &pscc);
			PMD(PMD_IOCTL, ("%s: HAS: %s@%s(%s#%d) pm_poll_cnt[%d] "
			    "%d\n", pmf, PM_DEVICE(dip), pscc->pscc_clone,
			    pm_poll_cnt[pscc->pscc_clone]))
			overrun = psc_entry(cmd, psce, dip, comp, newlevel,
			    oldlevel, PSC_DIRECT, canblock);
			PMD(PMD_DPM, ("%s: sig %d\n", pmf, pscc->pscc_clone))
			mutex_enter(&pm_clone_lock);
			if (!overrun)
				pm_poll_cnt[pscc->pscc_clone]++;
			cv_signal(&pm_clones_cv[pscc->pscc_clone]);
			pollwakeup(&pm_pollhead, (POLLRDNORM | POLLIN));
			mutex_exit(&pm_clone_lock);
		}
		mutex_enter(&pm_clone_lock);
		rw_enter(&pm_pscc_interest_rwlock, RW_READER);
		while ((psce = psc_interest(&cookie, &pscc)) != NULL) {
			(void) psc_entry(cmd, psce, dip, comp, newlevel,
			    oldlevel, PSC_INTEREST, canblock);
			cv_signal(&pm_clones_cv[pscc->pscc_clone]);
		}
		rw_exit(&pm_pscc_interest_rwlock);
		mutex_exit(&pm_clone_lock);
		break;
#ifdef DEBUG
	default:
		ASSERT(0);
#endif
	}
}

static void
pm_enqueue_notify_others(pm_ppm_devlist_t **listp, pm_canblock_t canblock)
{
	if (listp) {
		pm_ppm_devlist_t *p, *next = NULL;

		for (p = *listp; p; p = next) {
			next = p->ppd_next;
			pm_enqueue_notify(PSC_HAS_CHANGED, p->ppd_who,
			    p->ppd_cmpt, p->ppd_new_level, p->ppd_old_level,
			    canblock);
			kmem_free(p, sizeof (pm_ppm_devlist_t));
		}
		*listp = NULL;
	}
}

/*
 * Try to get the power locks of the parent node and target (child)
 * node.  Return true if successful (with both locks held) or false
 * (with no locks held).
 */
static int
pm_try_parent_child_locks(dev_info_t *pdip,
    dev_info_t *dip, int *pcircp, int *circp)
{
	if (ndi_devi_tryenter(pdip, pcircp))
		if (PM_TRY_LOCK_POWER(dip, circp)) {
			return (1);
		} else {
			ndi_devi_exit(pdip, *pcircp);
		}
	return (0);
}

/*
 * Determine if the power lock owner is blocked by current thread.
 * returns :
 * 	1 - If the thread owning the effective power lock (the first lock on
 *          which a thread blocks when it does PM_LOCK_POWER) is blocked by
 *          a mutex held by the current thread.
 *
 *	0 - otherwise
 *
 * Note : This function is called by pm_power_has_changed to determine whether
 * it is executing in parallel with pm_set_power.
 */
static int
pm_blocked_by_us(dev_info_t *dip)
{
	power_req_t power_req;
	kthread_t *owner;
	int result;
	kmutex_t *mp;
	dev_info_t *ppm = (dev_info_t *)DEVI(dip)->devi_pm_ppm;

	power_req.request_type = PMR_PPM_POWER_LOCK_OWNER;
	power_req.req.ppm_power_lock_owner_req.who = dip;
	if (pm_ctlops(ppm, dip, DDI_CTLOPS_POWER, &power_req, &result) !=
	    DDI_SUCCESS) {
		/*
		 * It is assumed that if the device is claimed by ppm, ppm
		 * will always implement this request type and it'll always
		 * return success. We panic here, if it fails.
		 */
		panic("pm: Can't determine power lock owner of %s@%s(%s#%d)\n",
		    PM_DEVICE(dip));
		/*NOTREACHED*/
	}

	if ((owner = power_req.req.ppm_power_lock_owner_req.owner) != NULL &&
	    owner->t_state == TS_SLEEP &&
	    owner->t_sobj_ops &&
	    SOBJ_TYPE(owner->t_sobj_ops) == SOBJ_MUTEX &&
	    (mp = (kmutex_t *)owner->t_wchan) &&
	    mutex_owner(mp) == curthread)
		return (1);

	return (0);
}

/*
 * Notify parent which wants to hear about a child's power changes.
 */
static void
pm_notify_parent(dev_info_t *dip,
    dev_info_t *pdip, int comp, int old_level, int level)
{
	pm_bp_has_changed_t bphc;
	pm_sp_misc_t pspm;
	char *pathbuf = kmem_alloc(MAXPATHLEN, KM_SLEEP);
	int result = DDI_SUCCESS;

	bphc.bphc_dip = dip;
	bphc.bphc_path = ddi_pathname(dip, pathbuf);
	bphc.bphc_comp = comp;
	bphc.bphc_olevel = old_level;
	bphc.bphc_nlevel = level;
	pspm.pspm_canblock = PM_CANBLOCK_BLOCK;
	pspm.pspm_scan = 0;
	bphc.bphc_private = &pspm;
	(void) (*PM_BUS_POWER_FUNC(pdip))(pdip, NULL,
	    BUS_POWER_HAS_CHANGED, (void *)&bphc, (void *)&result);
	kmem_free(pathbuf, MAXPATHLEN);
}

/*
 * Check if we need to resume a BC device, and make the attach call as required.
 */
static int
pm_check_and_resume(dev_info_t *dip, int comp, int old_level, int level)
{
	int ret = DDI_SUCCESS;

	if (PM_ISBC(dip) && comp == 0 && old_level == 0 && level != 0) {
		ASSERT(DEVI(dip)->devi_pm_flags & PMC_SUSPENDED);
		/* ppm is not interested in DDI_PM_RESUME */
		if ((ret = devi_attach(dip, DDI_PM_RESUME)) != DDI_SUCCESS)
			/* XXX Should we mark it resumed, */
			/* even though it failed? */
			cmn_err(CE_WARN, "!pm: Can't resume %s@%s",
			    PM_NAME(dip), PM_ADDR(dip));
		DEVI(dip)->devi_pm_flags &= ~PMC_SUSPENDED;
	}

	return (ret);
}

/*
 * Tests outside the lock to see if we should bother to enqueue an entry
 * for any watching process.  If yes, then caller will take the lock and
 * do the full protocol
 */
static int
pm_watchers()
{
	if (pm_processes_stopped)
		return (0);
	return (pm_pscc_direct || pm_pscc_interest);
}

static int pm_phc_impl(dev_info_t *, int, int, int);

/*
 * A driver is reporting that the power of one of its device's components
 * has changed.  Update the power state accordingly.
 */
int
pm_power_has_changed(dev_info_t *dip, int comp, int level)
{
	PMD_FUNC(pmf, "pm_power_has_changed")
	int ret;
	dev_info_t *pdip = ddi_get_parent(dip);
	struct pm_component *cp;
	int blocked, circ, pcirc, old_level;

	if (level < 0) {
		PMD(PMD_FAIL, ("%s: %s@%s(%s#%d): bad level=%d\n", pmf,
		    PM_DEVICE(dip), level))
		return (DDI_FAILURE);
	}

	PMD(PMD_KIDSUP | PMD_DEP, ("%s: %s@%s(%s#%d), comp=%d, level=%d\n", pmf,
	    PM_DEVICE(dip), comp, level))

	if (!e_pm_valid_info(dip, NULL) || !e_pm_valid_comp(dip, comp, &cp) ||
	    !e_pm_valid_power(dip, comp, level))
		return (DDI_FAILURE);

	/*
	 * A driver thread calling pm_power_has_changed and another thread
	 * calling pm_set_power can deadlock.  The problem is not resolvable
	 * by changing lock order, so we use pm_blocked_by_us() to detect
	 * this specific deadlock.  If we can't get the lock immediately
	 * and we are deadlocked, just update the component's level, do
	 * notifications, and return.  We intend to update the total power
	 * state later (if the other thread fails to set power to the
	 * desired level).  If we were called because of a power change on a
	 * component that isn't involved in a set_power op, update all state
	 * immediately.
	 */
	cp = PM_CP(dip, comp);
	while (!pm_try_parent_child_locks(pdip, dip, &pcirc, &circ)) {
		if (((blocked = pm_blocked_by_us(dip)) != 0) &&
		    (cp->pmc_flags & PM_POWER_OP)) {
			if (pm_watchers()) {
				mutex_enter(&pm_rsvp_lock);
				pm_enqueue_notify(PSC_HAS_CHANGED, dip, comp,
				    level, cur_power(cp), PM_CANBLOCK_BLOCK);
				mutex_exit(&pm_rsvp_lock);
			}
			if (pdip && PM_WANTS_NOTIFICATION(pdip))
				pm_notify_parent(dip,
				    pdip, comp, cur_power(cp), level);
			(void) pm_check_and_resume(dip,
			    comp, cur_power(cp), level);

			/*
			 * Stash the old power index, update curpwr, and flag
			 * that the total power state needs to be synched.
			 */
			cp->pmc_flags |= PM_PHC_WHILE_SET_POWER;
			/*
			 * Several pm_power_has_changed calls could arrive
			 * while the set power path remains blocked.  Keep the
			 * oldest old power and the newest new power of any
			 * sequence of phc calls which arrive during deadlock.
			 */
			if (cp->pmc_phc_pwr == PM_LEVEL_UNKNOWN)
				cp->pmc_phc_pwr = cp->pmc_cur_pwr;
			cp->pmc_cur_pwr =
			    pm_level_to_index(dip, cp, level);
			PMD(PMD_PHC, ("%s: deadlock for %s@%s(%s#%d), comp=%d, "
			    "level=%d\n", pmf, PM_DEVICE(dip), comp, level))
			return (DDI_SUCCESS);
		} else
			if (blocked) {	/* blocked, but different cmpt? */
				if (!ndi_devi_tryenter(pdip, &pcirc)) {
					cmn_err(CE_NOTE,
					    "!pm: parent kuc not updated due "
					    "to possible deadlock.\n");
					return (pm_phc_impl(dip,
					    comp, level, 1));
				}
				old_level = cur_power(cp);
				if (pdip && !PM_WANTS_NOTIFICATION(pdip) &&
				    (!PM_ISBC(dip) || comp == 0) &&
				    POWERING_ON(old_level, level))
					pm_hold_power(pdip);
				ret = pm_phc_impl(dip, comp, level, 1);
				if (pdip && !PM_WANTS_NOTIFICATION(pdip)) {
					if ((!PM_ISBC(dip) ||
					    comp == 0) && level == 0 &&
					    old_level != PM_LEVEL_UNKNOWN)
						pm_rele_power(pdip);
				}
				ndi_devi_exit(pdip, pcirc);
				/* child lock not held: deadlock */
				return (ret);
			}
		delay(1);
		PMD(PMD_PHC, ("%s: try lock again\n", pmf))
	}

	/* non-deadlock case */
	old_level = cur_power(cp);
	if (pdip && !PM_WANTS_NOTIFICATION(pdip) &&
	    (!PM_ISBC(dip) || comp == 0) && POWERING_ON(old_level, level))
		pm_hold_power(pdip);
	ret = pm_phc_impl(dip, comp, level, 1);
	if (pdip && !PM_WANTS_NOTIFICATION(pdip)) {
		if ((!PM_ISBC(dip) || comp == 0) && level == 0 &&
		    old_level != PM_LEVEL_UNKNOWN)
			pm_rele_power(pdip);
	}
	PM_UNLOCK_POWER(dip, circ);
	ndi_devi_exit(pdip, pcirc);
	return (ret);
}

/*
 * Account for power changes to a component of the the console frame buffer.
 * If lowering power from full (or "unkown", which is treatd as full)
 * we will increment the "components off" count of the fb device.
 * Subsequent lowering of the same component doesn't affect the count.  If
 * raising a component back to full power, we will decrement the count.
 *
 * Return: the increment value for pm_cfb_comps_off (-1, 0, or 1)
 */
static int
calc_cfb_comps_incr(dev_info_t *dip, int cmpt, int old, int new)
{
	struct pm_component *cp = PM_CP(dip, cmpt);
	int on = (old == PM_LEVEL_UNKNOWN || old == cp->pmc_norm_pwr);
	int want_normal = (new == cp->pmc_norm_pwr);
	int incr = 0;

	if (on && !want_normal)
		incr = 1;
	else if (!on && want_normal)
		incr = -1;
	return (incr);
}

/*
 * Adjust the count of console frame buffer components < full power.
 */
static void
update_comps_off(int incr, dev_info_t *dip)
{
		mutex_enter(&pm_cfb_lock);
		pm_cfb_comps_off += incr;
		ASSERT(pm_cfb_comps_off <= PM_NUMCMPTS(dip));
		mutex_exit(&pm_cfb_lock);
}

/*
 * Update the power state in the framework (via the ppm).  The 'notify'
 * argument tells whether to notify watchers.  Power lock is already held.
 */
static int
pm_phc_impl(dev_info_t *dip, int comp, int level, int notify)
{
	PMD_FUNC(pmf, "phc_impl")
	power_req_t power_req;
	int i, dodeps = 0;
	dev_info_t *pdip = ddi_get_parent(dip);
	int result;
	int old_level;
	struct pm_component *cp;
	int incr = 0;
	dev_info_t *ppm = (dev_info_t *)DEVI(dip)->devi_pm_ppm;
	int work_type = 0;
	char *pathbuf;

	/* Must use "official" power level for this test. */
	cp = PM_CP(dip, comp);
	old_level = (cp->pmc_flags & PM_PHC_WHILE_SET_POWER ?
	    cp->pmc_phc_pwr : cp->pmc_cur_pwr);
	if (old_level != PM_LEVEL_UNKNOWN)
		old_level = cp->pmc_comp.pmc_lvals[old_level];

	if (level == old_level) {
		PMD(PMD_SET, ("%s: %s@%s(%s#%d), comp=%d is already at "
		    "level=%d\n", pmf, PM_DEVICE(dip), comp, level))
		return (DDI_SUCCESS);
	}

	/*
	 * Tell ppm about this.
	 */
	power_req.request_type = PMR_PPM_POWER_CHANGE_NOTIFY;
	power_req.req.ppm_notify_level_req.who = dip;
	power_req.req.ppm_notify_level_req.cmpt = comp;
	power_req.req.ppm_notify_level_req.new_level = level;
	power_req.req.ppm_notify_level_req.old_level = old_level;
	if (pm_ctlops(ppm, dip, DDI_CTLOPS_POWER, &power_req,
	    &result) == DDI_FAILURE) {
		PMD(PMD_FAIL, ("%s: pm_ctlops %s@%s(%s#%d) to %d failed\n",
		    pmf, PM_DEVICE(dip), level))
		return (DDI_FAILURE);
	}

	if (PM_IS_CFB(dip)) {
		incr = calc_cfb_comps_incr(dip, comp, old_level, level);

		if (incr) {
			update_comps_off(incr, dip);
			PMD(PMD_CFB, ("%s: %s@%s(%s#%d) comp=%d %d->%d "
			    "cfb_comps_off->%d\n", pmf, PM_DEVICE(dip),
			    comp, old_level, level, pm_cfb_comps_off))
		}
	}
	e_pm_set_cur_pwr(dip, PM_CP(dip, comp), level);
	result = DDI_SUCCESS;

	if (notify) {
		if (pdip && PM_WANTS_NOTIFICATION(pdip))
			pm_notify_parent(dip, pdip, comp, old_level, level);
		(void) pm_check_and_resume(dip, comp, old_level, level);
	}

	/*
	 * Decrement the dependency kidsup count if we turn a device
	 * off.
	 */
	if (POWERING_OFF(old_level, level)) {
		dodeps = 1;
		for (i = 0; i < PM_NUMCMPTS(dip); i++) {
			cp = PM_CP(dip, i);
			if (cur_power(cp)) {
				dodeps = 0;
				break;
			}
		}
		if (dodeps)
			work_type = PM_DEP_WK_POWER_OFF;
	}

	/*
	 * Increment if we turn it on. Check to see
	 * if other comps are already on, if so,
	 * dont increment.
	 */
	if (POWERING_ON(old_level, level)) {
		dodeps = 1;
		for (i = 0; i < PM_NUMCMPTS(dip); i++) {
			cp = PM_CP(dip, i);
			if (comp == i)
				continue;
			/* -1 also treated as 0 in this case */
			if (cur_power(cp) > 0) {
				dodeps = 0;
				break;
			}
		}
		if (dodeps)
			work_type = PM_DEP_WK_POWER_ON;
	}

	if (dodeps) {
		pathbuf = kmem_zalloc(MAXPATHLEN, KM_SLEEP);
		(void) ddi_pathname(dip, pathbuf);
		pm_dispatch_to_dep_thread(work_type, pathbuf, NULL,
		    PM_DEP_NOWAIT, NULL, 0);
		kmem_free(pathbuf, MAXPATHLEN);
	}

	if (notify && (level != old_level) && pm_watchers()) {
		mutex_enter(&pm_rsvp_lock);
		pm_enqueue_notify(PSC_HAS_CHANGED, dip, comp, level, old_level,
		    PM_CANBLOCK_BLOCK);
		mutex_exit(&pm_rsvp_lock);
	}

	PMD(PMD_RESCAN, ("%s: %s@%s(%s#%d): pm_rescan\n", pmf, PM_DEVICE(dip)))
	pm_rescan(dip);
	return (DDI_SUCCESS);
}

/*
 * This function is called at startup time to notify pm of the existence
 * of any platform power managers for this platform.  As a result of
 * this registration, each function provided will be called each time
 * a device node is attached, until one returns true, and it must claim the
 * device node (by returning non-zero) if it wants to be involved in the
 * node's power management.  If it does claim the node, then it will
 * subsequently be notified of attach and detach events.
 *
 */

int
pm_register_ppm(int (*func)(dev_info_t *), dev_info_t *dip)
{
	PMD_FUNC(pmf, "register_ppm")
	struct ppm_callbacks *ppmcp;
	pm_component_t *cp;
	int i, pwr, result, circ;
	power_req_t power_req;
	struct ppm_notify_level_req *p = &power_req.req.ppm_notify_level_req;
	void pm_ppm_claim(dev_info_t *);

	mutex_enter(&ppm_lock);
	ppmcp = ppm_callbacks;
	for (i = 0; i < MAX_PPM_HANDLERS; i++, ppmcp++) {
		if (ppmcp->ppmc_func == NULL) {
			ppmcp->ppmc_func = func;
			ppmcp->ppmc_dip = dip;
			break;
		}
	}
	mutex_exit(&ppm_lock);

	if (i >= MAX_PPM_HANDLERS)
		return (DDI_FAILURE);
	while ((dip = ddi_get_parent(dip)) != NULL) {
		if (dip != ddi_root_node() && PM_GET_PM_INFO(dip) == NULL)
			continue;
		pm_ppm_claim(dip);
		/* don't bother with the not power-manageable nodes */
		if (pm_ppm_claimed(dip) && PM_GET_PM_INFO(dip)) {
			/*
			 * Tell ppm about this.
			 */
			power_req.request_type = PMR_PPM_POWER_CHANGE_NOTIFY;
			p->old_level = PM_LEVEL_UNKNOWN;
			p->who = dip;
			PM_LOCK_POWER(dip, &circ);
			for (i = 0; i < PM_NUMCMPTS(dip); i++) {
				cp = PM_CP(dip, i);
				pwr = cp->pmc_cur_pwr;
				if (pwr != PM_LEVEL_UNKNOWN) {
					p->cmpt = i;
					p->new_level = cur_power(cp);
					p->old_level = PM_LEVEL_UNKNOWN;
					if (pm_ctlops(PPM(dip), dip,
					    DDI_CTLOPS_POWER, &power_req,
					    &result) == DDI_FAILURE) {
						PMD(PMD_FAIL, ("%s: pc "
						    "%s@%s(%s#%d) to %d "
						    "fails\n", pmf,
						    PM_DEVICE(dip), pwr))
					}
				}
			}
			PM_UNLOCK_POWER(dip, circ);
		}
	}
	return (DDI_SUCCESS);
}

/*
 * Call the ppm's that have registered and adjust the devinfo struct as
 * appropriate.  First one to claim it gets it.  The sets of devices claimed
 * by each ppm are assumed to be disjoint.
 */
void
pm_ppm_claim(dev_info_t *dip)
{
	struct ppm_callbacks *ppmcp;

	if (PPM(dip)) {
		return;
	}
	mutex_enter(&ppm_lock);
	for (ppmcp = ppm_callbacks; ppmcp->ppmc_func; ppmcp++) {
		if ((*ppmcp->ppmc_func)(dip)) {
			DEVI(dip)->devi_pm_ppm =
			    (struct dev_info *)ppmcp->ppmc_dip;
			mutex_exit(&ppm_lock);
			return;
		}
	}
	mutex_exit(&ppm_lock);
}

/*
 * Node is being detached so stop autopm until we see if it succeeds, in which
 * case pm_stop will be called.  For backwards compatible devices we bring the
 * device up to full power on the assumption the detach will succeed.
 */
void
pm_detaching(dev_info_t *dip)
{
	PMD_FUNC(pmf, "detaching")
	pm_info_t *info = PM_GET_PM_INFO(dip);
	int iscons;

	PMD(PMD_REMDEV, ("%s: %s@%s(%s#%d), %d comps\n", pmf, PM_DEVICE(dip),
	    PM_NUMCMPTS(dip)))
	if (info == NULL)
		return;
	ASSERT(DEVI_IS_DETACHING(dip));
	PM_LOCK_DIP(dip);
	info->pmi_dev_pm_state |= PM_DETACHING;
	PM_UNLOCK_DIP(dip);
	if (!PM_ISBC(dip))
		pm_scan_stop(dip);

	/*
	 * console and old-style devices get brought up when detaching.
	 */
	iscons = PM_IS_CFB(dip);
	if (iscons || PM_ISBC(dip)) {
		(void) pm_all_to_normal(dip, PM_CANBLOCK_BYPASS);
		if (iscons) {
			mutex_enter(&pm_cfb_lock);
			while (cfb_inuse) {
				mutex_exit(&pm_cfb_lock);
				PMD(PMD_CFB, ("%s: delay; cfb_inuse\n", pmf))
				delay(1);
				mutex_enter(&pm_cfb_lock);
			}
			ASSERT(cfb_dip_detaching == NULL);
			ASSERT(cfb_dip);
			cfb_dip_detaching = cfb_dip;	/* case detach fails */
			cfb_dip = NULL;
			mutex_exit(&pm_cfb_lock);
		}
	}
}

/*
 * Node failed to detach.  If it used to be autopm'd, make it so again.
 */
void
pm_detach_failed(dev_info_t *dip)
{
	PMD_FUNC(pmf, "detach_failed")
	pm_info_t *info = PM_GET_PM_INFO(dip);
	int pm_all_at_normal(dev_info_t *);

	if (info == NULL)
		return;
	ASSERT(DEVI_IS_DETACHING(dip));
	if (info->pmi_dev_pm_state & PM_DETACHING) {
		info->pmi_dev_pm_state &= ~PM_DETACHING;
		if (info->pmi_dev_pm_state & PM_ALLNORM_DEFERRED) {
			/* Make sure the operation is still needed */
			if (!pm_all_at_normal(dip)) {
				if (pm_all_to_normal(dip,
				    PM_CANBLOCK_FAIL) != DDI_SUCCESS) {
					PMD(PMD_ERROR, ("%s: could not bring "
					    "%s@%s(%s#%d) to normal\n", pmf,
					    PM_DEVICE(dip)))
				}
			}
			info->pmi_dev_pm_state &= ~PM_ALLNORM_DEFERRED;
		}
	}
	if (!PM_ISBC(dip)) {
		mutex_enter(&pm_scan_lock);
		if (PM_SCANABLE(dip))
			pm_scan_init(dip);
		mutex_exit(&pm_scan_lock);
		pm_rescan(dip);
	}
}

/* generic Backwards Compatible component */
static char *bc_names[] = {"off", "on"};

static pm_comp_t bc_comp = {"unknown", 2, NULL, NULL, &bc_names[0]};

static void
e_pm_default_levels(dev_info_t *dip, pm_component_t *cp, int norm)
{
	pm_comp_t *pmc;
	pmc = &cp->pmc_comp;
	pmc->pmc_numlevels = 2;
	pmc->pmc_lvals[0] = 0;
	pmc->pmc_lvals[1] = norm;
	e_pm_set_cur_pwr(dip, cp, norm);
}

static void
e_pm_default_components(dev_info_t *dip, int cmpts)
{
	int i;
	pm_component_t *p = DEVI(dip)->devi_pm_components;

	p = DEVI(dip)->devi_pm_components;
	for (