xref: /titanic_41/usr/src/uts/common/io/cpudrv.c (revision 8f588c8376a0fa02470fc4554ec83442b305020f)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 /*
26  * Copyright (c) 2009,  Intel Corporation.
27  * All Rights Reserved.
28  */
29 
30 /*
31  * CPU Device driver. The driver is not DDI-compliant.
32  *
33  * The driver supports following features:
34  *	- Power management.
35  */
36 
37 #include <sys/types.h>
38 #include <sys/param.h>
39 #include <sys/errno.h>
40 #include <sys/modctl.h>
41 #include <sys/kmem.h>
42 #include <sys/conf.h>
43 #include <sys/cmn_err.h>
44 #include <sys/stat.h>
45 #include <sys/debug.h>
46 #include <sys/systm.h>
47 #include <sys/ddi.h>
48 #include <sys/sunddi.h>
49 #include <sys/sdt.h>
50 #include <sys/epm.h>
51 #include <sys/machsystm.h>
52 #include <sys/x_call.h>
53 #include <sys/cpudrv_mach.h>
54 #include <sys/msacct.h>
55 
56 /*
57  * CPU power management
58  *
59  * The supported power saving model is to slow down the CPU (on SPARC by
60  * dividing the CPU clock and on x86 by dropping down a P-state).
61  * Periodically we determine the amount of time the CPU is running
62  * idle thread and threads in user mode during the last quantum.  If the idle
63  * thread was running less than its low water mark for current speed for
64  * number of consecutive sampling periods, or number of running threads in
65  * user mode are above its high water mark, we arrange to go to the higher
66  * speed.  If the idle thread was running more than its high water mark without
67  * dropping a number of consecutive times below the mark, and number of threads
68  * running in user mode are below its low water mark, we arrange to go to the
69  * next lower speed.  While going down, we go through all the speeds.  While
70  * going up we go to the maximum speed to minimize impact on the user, but have
71  * provisions in the driver to go to other speeds.
72  *
73  * The driver does not have knowledge of a particular implementation of this
74  * scheme and will work with all CPUs supporting this model. On SPARC, the
75  * driver determines supported speeds by looking at 'clock-divisors' property
76  * created by OBP. On x86, the driver retrieves the supported speeds from
77  * ACPI.
78  */
79 
80 /*
81  * Configuration function prototypes and data structures
82  */
83 static int cpudrv_attach(dev_info_t *dip, ddi_attach_cmd_t cmd);
84 static int cpudrv_detach(dev_info_t *dip, ddi_detach_cmd_t cmd);
85 static int cpudrv_power(dev_info_t *dip, int comp, int level);
86 
87 struct dev_ops cpudrv_ops = {
88 	DEVO_REV,		/* rev */
89 	0,			/* refcnt */
90 	nodev,			/* getinfo */
91 	nulldev,		/* identify */
92 	nulldev,		/* probe */
93 	cpudrv_attach,		/* attach */
94 	cpudrv_detach,		/* detach */
95 	nodev,			/* reset */
96 	(struct cb_ops *)NULL,	/* cb_ops */
97 	(struct bus_ops *)NULL,	/* bus_ops */
98 	cpudrv_power,		/* power */
99 	ddi_quiesce_not_needed,		/* quiesce */
100 };
101 
102 static struct modldrv modldrv = {
103 	&mod_driverops,			/* modops */
104 	"CPU Driver",			/* linkinfo */
105 	&cpudrv_ops,			/* dev_ops */
106 };
107 
108 static struct modlinkage modlinkage = {
109 	MODREV_1,		/* rev */
110 	&modldrv,		/* linkage */
111 	NULL
112 };
113 
114 /*
115  * Function prototypes
116  */
117 static int cpudrv_init(cpudrv_devstate_t *cpudsp);
118 static void cpudrv_free(cpudrv_devstate_t *cpudsp);
119 static int cpudrv_comp_create(cpudrv_devstate_t *cpudsp);
120 static void cpudrv_monitor_disp(void *arg);
121 static void cpudrv_monitor(void *arg);
122 
123 /*
124  * Driver global variables
125  */
126 uint_t cpudrv_debug = 0;
127 void *cpudrv_state;
128 static uint_t cpudrv_idle_hwm = CPUDRV_IDLE_HWM;
129 static uint_t cpudrv_idle_lwm = CPUDRV_IDLE_LWM;
130 static uint_t cpudrv_idle_buf_zone = CPUDRV_IDLE_BUF_ZONE;
131 static uint_t cpudrv_idle_bhwm_cnt_max = CPUDRV_IDLE_BHWM_CNT_MAX;
132 static uint_t cpudrv_idle_blwm_cnt_max = CPUDRV_IDLE_BLWM_CNT_MAX;
133 static uint_t cpudrv_user_hwm = CPUDRV_USER_HWM;
134 
135 boolean_t cpudrv_enabled = B_TRUE;
136 
137 /*
138  * cpudrv_direct_pm allows user applications to directly control the
139  * power state transitions (direct pm) without following the normal
140  * direct pm protocol. This is needed because the normal protocol
141  * requires that a device only be lowered when it is idle, and be
142  * brought up when it request to do so by calling pm_raise_power().
143  * Ignoring this protocol is harmless for CPU (other than speed).
144  * Moreover it might be the case that CPU is never idle or wants
145  * to be at higher speed because of the addition CPU cycles required
146  * to run the user application.
147  *
148  * The driver will still report idle/busy status to the framework. Although
149  * framework will ignore this information for direct pm devices and not
150  * try to bring them down when idle, user applications can still use this
151  * information if they wants.
152  *
153  * In the future, provide an ioctl to control setting of this mode. In
154  * that case, this variable should move to the state structure and
155  * be protected by the lock in the state structure.
156  */
157 int cpudrv_direct_pm = 0;
158 
159 /*
160  * Arranges for the handler function to be called at the interval suitable
161  * for current speed.
162  */
163 #define	CPUDRV_MONITOR_INIT(cpudsp) { \
164     if (cpudrv_is_enabled(cpudsp)) {	      \
165 		ASSERT(mutex_owned(&(cpudsp)->lock)); \
166 		(cpudsp)->cpudrv_pm.timeout_id = \
167 		    timeout(cpudrv_monitor_disp, \
168 		    (cpudsp), (((cpudsp)->cpudrv_pm.cur_spd == NULL) ? \
169 		    CPUDRV_QUANT_CNT_OTHR : \
170 		    (cpudsp)->cpudrv_pm.cur_spd->quant_cnt)); \
171 	} \
172 }
173 
174 /*
175  * Arranges for the handler function not to be called back.
176  */
177 #define	CPUDRV_MONITOR_FINI(cpudsp) { \
178 	timeout_id_t tmp_tid; \
179 	ASSERT(mutex_owned(&(cpudsp)->lock)); \
180 	tmp_tid = (cpudsp)->cpudrv_pm.timeout_id; \
181 	(cpudsp)->cpudrv_pm.timeout_id = 0; \
182 	mutex_exit(&(cpudsp)->lock); \
183 	if (tmp_tid != 0) { \
184 		(void) untimeout(tmp_tid); \
185 		mutex_enter(&(cpudsp)->cpudrv_pm.timeout_lock); \
186 		while ((cpudsp)->cpudrv_pm.timeout_count != 0) \
187 			cv_wait(&(cpudsp)->cpudrv_pm.timeout_cv, \
188 			    &(cpudsp)->cpudrv_pm.timeout_lock); \
189 		mutex_exit(&(cpudsp)->cpudrv_pm.timeout_lock); \
190 	} \
191 	mutex_enter(&(cpudsp)->lock); \
192 }
193 
194 int
195 _init(void)
196 {
197 	int	error;
198 
199 	DPRINTF(D_INIT, (" _init: function called\n"));
200 	if ((error = ddi_soft_state_init(&cpudrv_state,
201 	    sizeof (cpudrv_devstate_t), 0)) != 0) {
202 		return (error);
203 	}
204 
205 	if ((error = mod_install(&modlinkage)) != 0)  {
206 		ddi_soft_state_fini(&cpudrv_state);
207 	}
208 
209 	/*
210 	 * Callbacks used by the PPM driver.
211 	 */
212 	CPUDRV_SET_PPM_CALLBACKS();
213 	return (error);
214 }
215 
216 int
217 _fini(void)
218 {
219 	int	error;
220 
221 	DPRINTF(D_FINI, (" _fini: function called\n"));
222 	if ((error = mod_remove(&modlinkage)) == 0) {
223 		ddi_soft_state_fini(&cpudrv_state);
224 	}
225 
226 	return (error);
227 }
228 
229 int
230 _info(struct modinfo *modinfop)
231 {
232 	return (mod_info(&modlinkage, modinfop));
233 }
234 
235 /*
236  * Driver attach(9e) entry point.
237  */
238 static int
239 cpudrv_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
240 {
241 	int			instance;
242 	cpudrv_devstate_t	*cpudsp;
243 
244 	instance = ddi_get_instance(dip);
245 
246 	switch (cmd) {
247 	case DDI_ATTACH:
248 		DPRINTF(D_ATTACH, ("cpudrv_attach: instance %d: "
249 		    "DDI_ATTACH called\n", instance));
250 		if (!cpudrv_is_enabled(NULL))
251 			return (DDI_FAILURE);
252 		if (ddi_soft_state_zalloc(cpudrv_state, instance) !=
253 		    DDI_SUCCESS) {
254 			cmn_err(CE_WARN, "cpudrv_attach: instance %d: "
255 			    "can't allocate state", instance);
256 			cpudrv_enabled = B_FALSE;
257 			return (DDI_FAILURE);
258 		}
259 		if ((cpudsp = ddi_get_soft_state(cpudrv_state, instance)) ==
260 		    NULL) {
261 			cmn_err(CE_WARN, "cpudrv_attach: instance %d: "
262 			    "can't get state", instance);
263 			ddi_soft_state_free(cpudrv_state, instance);
264 			cpudrv_enabled = B_FALSE;
265 			return (DDI_FAILURE);
266 		}
267 		cpudsp->dip = dip;
268 
269 		/*
270 		 * Find CPU number for this dev_info node.
271 		 */
272 		if (!cpudrv_get_cpu_id(dip, &(cpudsp->cpu_id))) {
273 			cmn_err(CE_WARN, "cpudrv_attach: instance %d: "
274 			    "can't convert dip to cpu_id", instance);
275 			ddi_soft_state_free(cpudrv_state, instance);
276 			cpudrv_enabled = B_FALSE;
277 			return (DDI_FAILURE);
278 		}
279 
280 		mutex_enter(&cpu_lock);
281 		cpudsp->cp = cpu_get(cpudsp->cpu_id);
282 		mutex_exit(&cpu_lock);
283 		if (cpudsp->cp == NULL) {
284 			cmn_err(CE_WARN, "cpudrv_attach: instance %d: "
285 			    "can't get cpu_t", ddi_get_instance(cpudsp->dip));
286 			ddi_soft_state_free(cpudrv_state, instance);
287 			cpudrv_enabled = B_FALSE;
288 			return (DDI_FAILURE);
289 		}
290 
291 		mutex_init(&cpudsp->lock, NULL, MUTEX_DRIVER, NULL);
292 		if (cpudrv_is_enabled(cpudsp)) {
293 			if (cpudrv_init(cpudsp) != DDI_SUCCESS) {
294 				cpudrv_enabled = B_FALSE;
295 				cpudrv_free(cpudsp);
296 				ddi_soft_state_free(cpudrv_state, instance);
297 				return (DDI_FAILURE);
298 			}
299 			if (cpudrv_comp_create(cpudsp) != DDI_SUCCESS) {
300 				cpudrv_enabled = B_FALSE;
301 				cpudrv_free(cpudsp);
302 				ddi_soft_state_free(cpudrv_state, instance);
303 				return (DDI_FAILURE);
304 			}
305 			if (ddi_prop_update_string(DDI_DEV_T_NONE,
306 			    dip, "pm-class", "CPU") != DDI_PROP_SUCCESS) {
307 				cpudrv_enabled = B_FALSE;
308 				cpudrv_free(cpudsp);
309 				ddi_soft_state_free(cpudrv_state, instance);
310 				return (DDI_FAILURE);
311 			}
312 
313 			/*
314 			 * Taskq is used to dispatch routine to monitor CPU
315 			 * activities.
316 			 */
317 			cpudsp->cpudrv_pm.tq = ddi_taskq_create(dip,
318 			    "cpudrv_monitor", CPUDRV_TASKQ_THREADS,
319 			    TASKQ_DEFAULTPRI, 0);
320 
321 			mutex_init(&cpudsp->cpudrv_pm.timeout_lock, NULL,
322 			    MUTEX_DRIVER, NULL);
323 			cv_init(&cpudsp->cpudrv_pm.timeout_cv, NULL,
324 			    CV_DEFAULT, NULL);
325 
326 			/*
327 			 * Driver needs to assume that CPU is running at
328 			 * unknown speed at DDI_ATTACH and switch it to the
329 			 * needed speed. We assume that initial needed speed
330 			 * is full speed for us.
331 			 */
332 			/*
333 			 * We need to take the lock because cpudrv_monitor()
334 			 * will start running in parallel with attach().
335 			 */
336 			mutex_enter(&cpudsp->lock);
337 			cpudsp->cpudrv_pm.cur_spd = NULL;
338 			cpudsp->cpudrv_pm.pm_started = B_FALSE;
339 			/*
340 			 * We don't call pm_raise_power() directly from attach
341 			 * because driver attach for a slave CPU node can
342 			 * happen before the CPU is even initialized. We just
343 			 * start the monitoring system which understands
344 			 * unknown speed and moves CPU to top speed when it
345 			 * has been initialized.
346 			 */
347 			CPUDRV_MONITOR_INIT(cpudsp);
348 			mutex_exit(&cpudsp->lock);
349 
350 		}
351 
352 		if (!cpudrv_mach_init(cpudsp)) {
353 			cmn_err(CE_WARN, "cpudrv_attach: instance %d: "
354 			    "cpudrv_mach_init failed", instance);
355 			cpudrv_enabled = B_FALSE;
356 			cpudrv_free(cpudsp);
357 			ddi_soft_state_free(cpudrv_state, instance);
358 			return (DDI_FAILURE);
359 		}
360 
361 		CPUDRV_INSTALL_MAX_CHANGE_HANDLER(cpudsp);
362 
363 		(void) ddi_prop_update_int(DDI_DEV_T_NONE, dip,
364 		    DDI_NO_AUTODETACH, 1);
365 		ddi_report_dev(dip);
366 		return (DDI_SUCCESS);
367 
368 	case DDI_RESUME:
369 		DPRINTF(D_ATTACH, ("cpudrv_attach: instance %d: "
370 		    "DDI_RESUME called\n", instance));
371 
372 		cpudsp = ddi_get_soft_state(cpudrv_state, instance);
373 		ASSERT(cpudsp != NULL);
374 
375 		/*
376 		 * Nothing to do for resume, if not doing active PM.
377 		 */
378 		if (!cpudrv_is_enabled(cpudsp))
379 			return (DDI_SUCCESS);
380 
381 		mutex_enter(&cpudsp->lock);
382 		/*
383 		 * Driver needs to assume that CPU is running at unknown speed
384 		 * at DDI_RESUME and switch it to the needed speed. We assume
385 		 * that the needed speed is full speed for us.
386 		 */
387 		cpudsp->cpudrv_pm.cur_spd = NULL;
388 		CPUDRV_MONITOR_INIT(cpudsp);
389 		mutex_exit(&cpudsp->lock);
390 		CPUDRV_REDEFINE_TOPSPEED(dip);
391 		return (DDI_SUCCESS);
392 
393 	default:
394 		return (DDI_FAILURE);
395 	}
396 }
397 
398 /*
399  * Driver detach(9e) entry point.
400  */
401 static int
402 cpudrv_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
403 {
404 	int			instance;
405 	cpudrv_devstate_t	*cpudsp;
406 	cpudrv_pm_t		*cpupm;
407 
408 	instance = ddi_get_instance(dip);
409 
410 	switch (cmd) {
411 	case DDI_DETACH:
412 		DPRINTF(D_DETACH, ("cpudrv_detach: instance %d: "
413 		    "DDI_DETACH called\n", instance));
414 
415 #if defined(__x86)
416 		cpudsp = ddi_get_soft_state(cpudrv_state, instance);
417 		ASSERT(cpudsp != NULL);
418 
419 		/*
420 		 * Nothing to do for detach, if no doing active PM.
421 		 */
422 		if (!cpudrv_is_enabled(cpudsp))
423 			return (DDI_SUCCESS);
424 
425 		/*
426 		 * uninstall PPC/_TPC change notification handler
427 		 */
428 		CPUDRV_UNINSTALL_MAX_CHANGE_HANDLER(cpudsp);
429 
430 		/*
431 		 * destruct platform specific resource
432 		 */
433 		if (!cpudrv_mach_fini(cpudsp))
434 			return (DDI_FAILURE);
435 
436 		mutex_enter(&cpudsp->lock);
437 		CPUDRV_MONITOR_FINI(cpudsp);
438 		cv_destroy(&cpudsp->cpudrv_pm.timeout_cv);
439 		mutex_destroy(&cpudsp->cpudrv_pm.timeout_lock);
440 		ddi_taskq_destroy(cpudsp->cpudrv_pm.tq);
441 		cpudrv_free(cpudsp);
442 		mutex_exit(&cpudsp->lock);
443 		mutex_destroy(&cpudsp->lock);
444 		ddi_soft_state_free(cpudrv_state, instance);
445 		(void) ddi_prop_update_int(DDI_DEV_T_NONE, dip,
446 		    DDI_NO_AUTODETACH, 0);
447 		return (DDI_SUCCESS);
448 
449 #else
450 		/*
451 		 * If the only thing supported by the driver is power
452 		 * management, we can in future enhance the driver and
453 		 * framework that loads it to unload the driver when
454 		 * user has disabled CPU power management.
455 		 */
456 		return (DDI_FAILURE);
457 #endif
458 
459 	case DDI_SUSPEND:
460 		DPRINTF(D_DETACH, ("cpudrv_detach: instance %d: "
461 		    "DDI_SUSPEND called\n", instance));
462 
463 		cpudsp = ddi_get_soft_state(cpudrv_state, instance);
464 		ASSERT(cpudsp != NULL);
465 
466 		/*
467 		 * Nothing to do for suspend, if not doing active PM.
468 		 */
469 		if (!cpudrv_is_enabled(cpudsp))
470 			return (DDI_SUCCESS);
471 
472 		/*
473 		 * During a checkpoint-resume sequence, framework will
474 		 * stop interrupts to quiesce kernel activity. This will
475 		 * leave our monitoring system ineffective. Handle this
476 		 * by stopping our monitoring system and bringing CPU
477 		 * to full speed. In case we are in special direct pm
478 		 * mode, we leave the CPU at whatever speed it is. This
479 		 * is harmless other than speed.
480 		 */
481 		mutex_enter(&cpudsp->lock);
482 		cpupm = &(cpudsp->cpudrv_pm);
483 
484 		DPRINTF(D_DETACH, ("cpudrv_detach: instance %d: DDI_SUSPEND - "
485 		    "cur_spd %d, topspeed %d\n", instance,
486 		    cpupm->cur_spd->pm_level,
487 		    CPUDRV_TOPSPEED(cpupm)->pm_level));
488 
489 		CPUDRV_MONITOR_FINI(cpudsp);
490 
491 		if (!cpudrv_direct_pm && (cpupm->cur_spd !=
492 		    CPUDRV_TOPSPEED(cpupm))) {
493 			if (cpupm->pm_busycnt < 1) {
494 				if ((pm_busy_component(dip, CPUDRV_COMP_NUM)
495 				    == DDI_SUCCESS)) {
496 					cpupm->pm_busycnt++;
497 				} else {
498 					CPUDRV_MONITOR_INIT(cpudsp);
499 					mutex_exit(&cpudsp->lock);
500 					cmn_err(CE_WARN, "cpudrv_detach: "
501 					    "instance %d: can't busy CPU "
502 					    "component", instance);
503 					return (DDI_FAILURE);
504 				}
505 			}
506 			mutex_exit(&cpudsp->lock);
507 			if (pm_raise_power(dip, CPUDRV_COMP_NUM,
508 			    CPUDRV_TOPSPEED(cpupm)->pm_level) !=
509 			    DDI_SUCCESS) {
510 				mutex_enter(&cpudsp->lock);
511 				CPUDRV_MONITOR_INIT(cpudsp);
512 				mutex_exit(&cpudsp->lock);
513 				cmn_err(CE_WARN, "cpudrv_detach: instance %d: "
514 				    "can't raise CPU power level to %d",
515 				    instance,
516 				    CPUDRV_TOPSPEED(cpupm)->pm_level);
517 				return (DDI_FAILURE);
518 			} else {
519 				return (DDI_SUCCESS);
520 			}
521 		} else {
522 			mutex_exit(&cpudsp->lock);
523 			return (DDI_SUCCESS);
524 		}
525 
526 	default:
527 		return (DDI_FAILURE);
528 	}
529 }
530 
531 /*
532  * Driver power(9e) entry point.
533  *
534  * Driver's notion of current power is set *only* in power(9e) entry point
535  * after actual power change operation has been successfully completed.
536  */
537 /* ARGSUSED */
538 static int
539 cpudrv_power(dev_info_t *dip, int comp, int level)
540 {
541 	int			instance;
542 	cpudrv_devstate_t	*cpudsp;
543 	cpudrv_pm_t 		*cpudrvpm;
544 	cpudrv_pm_spd_t		*new_spd;
545 	boolean_t		is_ready;
546 	int			ret;
547 
548 	instance = ddi_get_instance(dip);
549 
550 	DPRINTF(D_POWER, ("cpudrv_power: instance %d: level %d\n",
551 	    instance, level));
552 
553 	if ((cpudsp = ddi_get_soft_state(cpudrv_state, instance)) == NULL) {
554 		cmn_err(CE_WARN, "cpudrv_power: instance %d: can't "
555 		    "get state", instance);
556 		return (DDI_FAILURE);
557 	}
558 
559 	mutex_enter(&cpudsp->lock);
560 	cpudrvpm = &(cpudsp->cpudrv_pm);
561 
562 	/*
563 	 * In normal operation, we fail if we are busy and request is
564 	 * to lower the power level. We let this go through if the driver
565 	 * is in special direct pm mode. On x86, we also let this through
566 	 * if the change is due to a request to govern the max speed.
567 	 */
568 	if (!cpudrv_direct_pm && (cpudrvpm->pm_busycnt >= 1) &&
569 	    !cpudrv_is_governor_thread(cpudrvpm)) {
570 		if ((cpudrvpm->cur_spd != NULL) &&
571 		    (level < cpudrvpm->cur_spd->pm_level)) {
572 			mutex_exit(&cpudsp->lock);
573 			return (DDI_FAILURE);
574 		}
575 	}
576 
577 	for (new_spd = cpudrvpm->head_spd; new_spd; new_spd =
578 	    new_spd->down_spd) {
579 		if (new_spd->pm_level == level)
580 			break;
581 	}
582 	if (!new_spd) {
583 		CPUDRV_RESET_GOVERNOR_THREAD(cpudrvpm);
584 		mutex_exit(&cpudsp->lock);
585 		cmn_err(CE_WARN, "cpudrv_power: instance %d: "
586 		    "can't locate new CPU speed", instance);
587 		return (DDI_FAILURE);
588 	}
589 
590 	/*
591 	 * We currently refuse to power manage if the CPU is not ready to
592 	 * take cross calls (cross calls fail silently if CPU is not ready
593 	 * for it).
594 	 *
595 	 * Additionally, for x86 platforms we cannot power manage an instance,
596 	 * until it has been initialized.
597 	 */
598 	ASSERT(cpudsp->cp);
599 	is_ready = CPUDRV_XCALL_IS_READY(cpudsp->cpu_id);
600 	if (!is_ready) {
601 		DPRINTF(D_POWER, ("cpudrv_power: instance %d: "
602 		    "CPU not ready for x-calls\n", instance));
603 	} else if (!(is_ready = cpudrv_power_ready(cpudsp->cp))) {
604 		DPRINTF(D_POWER, ("cpudrv_power: instance %d: "
605 		    "waiting for all CPUs to be power manageable\n",
606 		    instance));
607 	}
608 	if (!is_ready) {
609 		CPUDRV_RESET_GOVERNOR_THREAD(cpudrvpm);
610 		mutex_exit(&cpudsp->lock);
611 		return (DDI_FAILURE);
612 	}
613 
614 	/*
615 	 * Execute CPU specific routine on the requested CPU to
616 	 * change its speed to normal-speed/divisor.
617 	 */
618 	if ((ret = cpudrv_change_speed(cpudsp, new_spd)) != DDI_SUCCESS) {
619 		cmn_err(CE_WARN, "cpudrv_power: "
620 		    "cpudrv_change_speed() return = %d", ret);
621 		mutex_exit(&cpudsp->lock);
622 		return (DDI_FAILURE);
623 	}
624 
625 	/*
626 	 * Reset idle threshold time for the new power level.
627 	 */
628 	if ((cpudrvpm->cur_spd != NULL) && (level <
629 	    cpudrvpm->cur_spd->pm_level)) {
630 		if (pm_idle_component(dip, CPUDRV_COMP_NUM) ==
631 		    DDI_SUCCESS) {
632 			if (cpudrvpm->pm_busycnt >= 1)
633 				cpudrvpm->pm_busycnt--;
634 		} else {
635 			cmn_err(CE_WARN, "cpudrv_power: instance %d: "
636 			    "can't idle CPU component",
637 			    ddi_get_instance(dip));
638 		}
639 	}
640 	/*
641 	 * Reset various parameters because we are now running at new speed.
642 	 */
643 	cpudrvpm->lastquan_mstate[CMS_IDLE] = 0;
644 	cpudrvpm->lastquan_mstate[CMS_SYSTEM] = 0;
645 	cpudrvpm->lastquan_mstate[CMS_USER] = 0;
646 	cpudrvpm->lastquan_ticks = 0;
647 	cpudrvpm->cur_spd = new_spd;
648 	CPUDRV_RESET_GOVERNOR_THREAD(cpudrvpm);
649 	mutex_exit(&cpudsp->lock);
650 
651 	return (DDI_SUCCESS);
652 }
653 
654 /*
655  * Initialize power management data.
656  */
657 static int
658 cpudrv_init(cpudrv_devstate_t *cpudsp)
659 {
660 	cpudrv_pm_t 	*cpupm = &(cpudsp->cpudrv_pm);
661 	cpudrv_pm_spd_t	*cur_spd;
662 	cpudrv_pm_spd_t	*prev_spd = NULL;
663 	int		*speeds;
664 	uint_t		nspeeds;
665 	int		idle_cnt_percent;
666 	int		user_cnt_percent;
667 	int		i;
668 
669 	CPUDRV_GET_SPEEDS(cpudsp, speeds, nspeeds);
670 	if (nspeeds < 2) {
671 		/* Need at least two speeds to power manage */
672 		CPUDRV_FREE_SPEEDS(speeds, nspeeds);
673 		return (DDI_FAILURE);
674 	}
675 	cpupm->num_spd = nspeeds;
676 
677 	/*
678 	 * Calculate the watermarks and other parameters based on the
679 	 * supplied speeds.
680 	 *
681 	 * One of the basic assumption is that for X amount of CPU work,
682 	 * if CPU is slowed down by a factor of N, the time it takes to
683 	 * do the same work will be N * X.
684 	 *
685 	 * The driver declares that a CPU is idle and ready for slowed down,
686 	 * if amount of idle thread is more than the current speed idle_hwm
687 	 * without dropping below idle_hwm a number of consecutive sampling
688 	 * intervals and number of running threads in user mode are below
689 	 * user_lwm.  We want to set the current user_lwm such that if we
690 	 * just switched to the next slower speed with no change in real work
691 	 * load, the amount of user threads at the slower speed will be such
692 	 * that it falls below the slower speed's user_hwm.  If we didn't do
693 	 * that then we will just come back to the higher speed as soon as we
694 	 * go down even with no change in work load.
695 	 * The user_hwm is a fixed precentage and not calculated dynamically.
696 	 *
697 	 * We bring the CPU up if idle thread at current speed is less than
698 	 * the current speed idle_lwm for a number of consecutive sampling
699 	 * intervals or user threads are above the user_hwm for the current
700 	 * speed.
701 	 */
702 	for (i = 0; i < nspeeds; i++) {
703 		cur_spd = kmem_zalloc(sizeof (cpudrv_pm_spd_t), KM_SLEEP);
704 		cur_spd->speed = speeds[i];
705 		if (i == 0) {	/* normal speed */
706 			cpupm->head_spd = cur_spd;
707 			CPUDRV_TOPSPEED(cpupm) = cur_spd;
708 			cur_spd->quant_cnt = CPUDRV_QUANT_CNT_NORMAL;
709 			cur_spd->idle_hwm =
710 			    (cpudrv_idle_hwm * cur_spd->quant_cnt) / 100;
711 			/* can't speed anymore */
712 			cur_spd->idle_lwm = 0;
713 			cur_spd->user_hwm = UINT_MAX;
714 		} else {
715 			cur_spd->quant_cnt = CPUDRV_QUANT_CNT_OTHR;
716 			ASSERT(prev_spd != NULL);
717 			prev_spd->down_spd = cur_spd;
718 			cur_spd->up_spd = cpupm->head_spd;
719 
720 			/*
721 			 * Let's assume CPU is considered idle at full speed
722 			 * when it is spending I% of time in running the idle
723 			 * thread.  At full speed, CPU will be busy (100 - I) %
724 			 * of times.  This % of busyness increases by factor of
725 			 * N as CPU slows down.  CPU that is idle I% of times
726 			 * in full speed, it is idle (100 - ((100 - I) * N)) %
727 			 * of times in N speed.  The idle_lwm is a fixed
728 			 * percentage.  A large value of N may result in
729 			 * idle_hwm to go below idle_lwm.  We need to make sure
730 			 * that there is at least a buffer zone seperation
731 			 * between the idle_lwm and idle_hwm values.
732 			 */
733 			idle_cnt_percent = CPUDRV_IDLE_CNT_PERCENT(
734 			    cpudrv_idle_hwm, speeds, i);
735 			idle_cnt_percent = max(idle_cnt_percent,
736 			    (cpudrv_idle_lwm + cpudrv_idle_buf_zone));
737 			cur_spd->idle_hwm =
738 			    (idle_cnt_percent * cur_spd->quant_cnt) / 100;
739 			cur_spd->idle_lwm =
740 			    (cpudrv_idle_lwm * cur_spd->quant_cnt) / 100;
741 
742 			/*
743 			 * The lwm for user threads are determined such that
744 			 * if CPU slows down, the load of work in the
745 			 * new speed would still keep the CPU at or below the
746 			 * user_hwm in the new speed.  This is to prevent
747 			 * the quick jump back up to higher speed.
748 			 */
749 			cur_spd->user_hwm = (cpudrv_user_hwm *
750 			    cur_spd->quant_cnt) / 100;
751 			user_cnt_percent = CPUDRV_USER_CNT_PERCENT(
752 			    cpudrv_user_hwm, speeds, i);
753 			prev_spd->user_lwm =
754 			    (user_cnt_percent * prev_spd->quant_cnt) / 100;
755 		}
756 		prev_spd = cur_spd;
757 	}
758 	/* Slowest speed. Can't slow down anymore */
759 	cur_spd->idle_hwm = UINT_MAX;
760 	cur_spd->user_lwm = -1;
761 #ifdef	DEBUG
762 	DPRINTF(D_PM_INIT, ("cpudrv_init: instance %d: head_spd spd %d, "
763 	    "num_spd %d\n", ddi_get_instance(cpudsp->dip),
764 	    cpupm->head_spd->speed, cpupm->num_spd));
765 	for (cur_spd = cpupm->head_spd; cur_spd; cur_spd = cur_spd->down_spd) {
766 		DPRINTF(D_PM_INIT, ("cpudrv_init: instance %d: speed %d, "
767 		    "down_spd spd %d, idle_hwm %d, user_lwm %d, "
768 		    "up_spd spd %d, idle_lwm %d, user_hwm %d, "
769 		    "quant_cnt %d\n", ddi_get_instance(cpudsp->dip),
770 		    cur_spd->speed,
771 		    (cur_spd->down_spd ? cur_spd->down_spd->speed : 0),
772 		    cur_spd->idle_hwm, cur_spd->user_lwm,
773 		    (cur_spd->up_spd ? cur_spd->up_spd->speed : 0),
774 		    cur_spd->idle_lwm, cur_spd->user_hwm,
775 		    cur_spd->quant_cnt));
776 	}
777 #endif	/* DEBUG */
778 	CPUDRV_FREE_SPEEDS(speeds, nspeeds);
779 	return (DDI_SUCCESS);
780 }
781 
782 /*
783  * Free CPU power management data.
784  */
785 static void
786 cpudrv_free(cpudrv_devstate_t *cpudsp)
787 {
788 	cpudrv_pm_t 	*cpupm = &(cpudsp->cpudrv_pm);
789 	cpudrv_pm_spd_t	*cur_spd, *next_spd;
790 
791 	cur_spd = cpupm->head_spd;
792 	while (cur_spd) {
793 		next_spd = cur_spd->down_spd;
794 		kmem_free(cur_spd, sizeof (cpudrv_pm_spd_t));
795 		cur_spd = next_spd;
796 	}
797 	bzero(cpupm, sizeof (cpudrv_pm_t));
798 }
799 
800 /*
801  * Create pm-components property.
802  */
803 static int
804 cpudrv_comp_create(cpudrv_devstate_t *cpudsp)
805 {
806 	cpudrv_pm_t 	*cpupm = &(cpudsp->cpudrv_pm);
807 	cpudrv_pm_spd_t	*cur_spd;
808 	char		**pmc;
809 	int		size;
810 	char		name[] = "NAME=CPU Speed";
811 	int		i, j;
812 	uint_t		comp_spd;
813 	int		result = DDI_FAILURE;
814 
815 	pmc = kmem_zalloc((cpupm->num_spd + 1) * sizeof (char *), KM_SLEEP);
816 	size = CPUDRV_COMP_SIZE();
817 	if (cpupm->num_spd > CPUDRV_COMP_MAX_VAL) {
818 		cmn_err(CE_WARN, "cpudrv_comp_create: instance %d: "
819 		    "number of speeds exceeded limits",
820 		    ddi_get_instance(cpudsp->dip));
821 		kmem_free(pmc, (cpupm->num_spd + 1) * sizeof (char *));
822 		return (result);
823 	}
824 
825 	for (i = cpupm->num_spd, cur_spd = cpupm->head_spd; i > 0;
826 	    i--, cur_spd = cur_spd->down_spd) {
827 		cur_spd->pm_level = i;
828 		pmc[i] = kmem_zalloc((size * sizeof (char)), KM_SLEEP);
829 		comp_spd = CPUDRV_COMP_SPEED(cpupm, cur_spd);
830 		if (comp_spd > CPUDRV_COMP_MAX_VAL) {
831 			cmn_err(CE_WARN, "cpudrv_comp_create: "
832 			    "instance %d: speed exceeded limits",
833 			    ddi_get_instance(cpudsp->dip));
834 			for (j = cpupm->num_spd; j >= i; j--) {
835 				kmem_free(pmc[j], size * sizeof (char));
836 			}
837 			kmem_free(pmc, (cpupm->num_spd + 1) *
838 			    sizeof (char *));
839 			return (result);
840 		}
841 		CPUDRV_COMP_SPRINT(pmc[i], cpupm, cur_spd, comp_spd)
842 		DPRINTF(D_PM_COMP_CREATE, ("cpudrv_comp_create: "
843 		    "instance %d: pm-components power level %d string '%s'\n",
844 		    ddi_get_instance(cpudsp->dip), i, pmc[i]));
845 	}
846 	pmc[0] = kmem_zalloc(sizeof (name), KM_SLEEP);
847 	(void) strcat(pmc[0], name);
848 	DPRINTF(D_PM_COMP_CREATE, ("cpudrv_comp_create: instance %d: "
849 	    "pm-components component name '%s'\n",
850 	    ddi_get_instance(cpudsp->dip), pmc[0]));
851 
852 	if (ddi_prop_update_string_array(DDI_DEV_T_NONE, cpudsp->dip,
853 	    "pm-components", pmc, cpupm->num_spd + 1) == DDI_PROP_SUCCESS) {
854 		result = DDI_SUCCESS;
855 	} else {
856 		cmn_err(CE_WARN, "cpudrv_comp_create: instance %d: "
857 		    "can't create pm-components property",
858 		    ddi_get_instance(cpudsp->dip));
859 	}
860 
861 	for (i = cpupm->num_spd; i > 0; i--) {
862 		kmem_free(pmc[i], size * sizeof (char));
863 	}
864 	kmem_free(pmc[0], sizeof (name));
865 	kmem_free(pmc, (cpupm->num_spd + 1) * sizeof (char *));
866 	return (result);
867 }
868 
869 /*
870  * Mark a component idle.
871  */
872 #define	CPUDRV_MONITOR_PM_IDLE_COMP(dip, cpupm) { \
873 	if ((cpupm)->pm_busycnt >= 1) { \
874 		if (pm_idle_component((dip), CPUDRV_COMP_NUM) == \
875 		    DDI_SUCCESS) { \
876 			DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: " \
877 			    "instance %d: pm_idle_component called\n", \
878 			    ddi_get_instance((dip)))); \
879 			(cpupm)->pm_busycnt--; \
880 		} else { \
881 			cmn_err(CE_WARN, "cpudrv_monitor: instance %d: " \
882 			    "can't idle CPU component", \
883 			    ddi_get_instance((dip))); \
884 		} \
885 	} \
886 }
887 
888 /*
889  * Marks a component busy in both PM framework and driver state structure.
890  */
891 #define	CPUDRV_MONITOR_PM_BUSY_COMP(dip, cpupm) { \
892 	if ((cpupm)->pm_busycnt < 1) { \
893 		if (pm_busy_component((dip), CPUDRV_COMP_NUM) == \
894 		    DDI_SUCCESS) { \
895 			DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: " \
896 			    "instance %d: pm_busy_component called\n", \
897 			    ddi_get_instance((dip)))); \
898 			(cpupm)->pm_busycnt++; \
899 		} else { \
900 			cmn_err(CE_WARN, "cpudrv_monitor: instance %d: " \
901 			    "can't busy CPU component", \
902 			    ddi_get_instance((dip))); \
903 		} \
904 	} \
905 }
906 
907 /*
908  * Marks a component busy and calls pm_raise_power().
909  */
910 #define	CPUDRV_MONITOR_PM_BUSY_AND_RAISE(dip, cpudsp, cpupm, new_spd) { \
911 	int ret; \
912 	/* \
913 	 * Mark driver and PM framework busy first so framework doesn't try \
914 	 * to bring CPU to lower speed when we need to be at higher speed. \
915 	 */ \
916 	CPUDRV_MONITOR_PM_BUSY_COMP((dip), (cpupm)); \
917 	mutex_exit(&(cpudsp)->lock); \
918 	DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: instance %d: " \
919 	    "pm_raise_power called to %d\n", ddi_get_instance((dip)), \
920 		(new_spd->pm_level))); \
921 	ret = pm_raise_power((dip), CPUDRV_COMP_NUM, (new_spd->pm_level)); \
922 	if (ret != DDI_SUCCESS) { \
923 		cmn_err(CE_WARN, "cpudrv_monitor: instance %d: can't " \
924 		    "raise CPU power level", ddi_get_instance((dip))); \
925 	} \
926 	mutex_enter(&(cpudsp)->lock); \
927 	if (ret == DDI_SUCCESS && cpudsp->cpudrv_pm.cur_spd == NULL) { \
928 		cpudsp->cpudrv_pm.cur_spd = new_spd; \
929 	} \
930 }
931 
932 /*
933  * In order to monitor a CPU, we need to hold cpu_lock to access CPU
934  * statistics. Holding cpu_lock is not allowed from a callout routine.
935  * We dispatch a taskq to do that job.
936  */
937 static void
938 cpudrv_monitor_disp(void *arg)
939 {
940 	cpudrv_devstate_t	*cpudsp = (cpudrv_devstate_t *)arg;
941 
942 	/*
943 	 * We are here because the last task has scheduled a timeout.
944 	 * The queue should be empty at this time.
945 	 */
946 	mutex_enter(&cpudsp->cpudrv_pm.timeout_lock);
947 	if ((ddi_taskq_dispatch(cpudsp->cpudrv_pm.tq, cpudrv_monitor, arg,
948 	    DDI_NOSLEEP)) != DDI_SUCCESS) {
949 		mutex_exit(&cpudsp->cpudrv_pm.timeout_lock);
950 		DPRINTF(D_PM_MONITOR, ("cpudrv_monitor_disp: failed to "
951 		    "dispatch the cpudrv_monitor taskq\n"));
952 		mutex_enter(&cpudsp->lock);
953 		CPUDRV_MONITOR_INIT(cpudsp);
954 		mutex_exit(&cpudsp->lock);
955 		return;
956 	}
957 	cpudsp->cpudrv_pm.timeout_count++;
958 	mutex_exit(&cpudsp->cpudrv_pm.timeout_lock);
959 }
960 
961 /*
962  * Monitors each CPU for the amount of time idle thread was running in the
963  * last quantum and arranges for the CPU to go to the lower or higher speed.
964  * Called at the time interval appropriate for the current speed. The
965  * time interval for normal speed is CPUDRV_QUANT_CNT_NORMAL. The time
966  * interval for other speeds (including unknown speed) is
967  * CPUDRV_QUANT_CNT_OTHR.
968  */
969 static void
970 cpudrv_monitor(void *arg)
971 {
972 	cpudrv_devstate_t	*cpudsp = (cpudrv_devstate_t *)arg;
973 	cpudrv_pm_t		*cpupm;
974 	cpudrv_pm_spd_t		*cur_spd, *new_spd;
975 	dev_info_t		*dip;
976 	uint_t			idle_cnt, user_cnt, system_cnt;
977 	clock_t			ticks;
978 	uint_t			tick_cnt;
979 	hrtime_t		msnsecs[NCMSTATES];
980 	boolean_t		is_ready;
981 
982 #define	GET_CPU_MSTATE_CNT(state, cnt) \
983 	msnsecs[state] = NSEC_TO_TICK(msnsecs[state]); \
984 	if (cpupm->lastquan_mstate[state] > msnsecs[state]) \
985 		msnsecs[state] = cpupm->lastquan_mstate[state]; \
986 	cnt = msnsecs[state] - cpupm->lastquan_mstate[state]; \
987 	cpupm->lastquan_mstate[state] = msnsecs[state]
988 
989 	mutex_enter(&cpudsp->lock);
990 	cpupm = &(cpudsp->cpudrv_pm);
991 	if (cpupm->timeout_id == 0) {
992 		mutex_exit(&cpudsp->lock);
993 		goto do_return;
994 	}
995 	cur_spd = cpupm->cur_spd;
996 	dip = cpudsp->dip;
997 
998 	/*
999 	 * We assume that a CPU is initialized and has a valid cpu_t
1000 	 * structure, if it is ready for cross calls. If this changes,
1001 	 * additional checks might be needed.
1002 	 *
1003 	 * Additionally, for x86 platforms we cannot power manage an
1004 	 * instance, until it has been initialized.
1005 	 */
1006 	ASSERT(cpudsp->cp);
1007 	is_ready = CPUDRV_XCALL_IS_READY(cpudsp->cpu_id);
1008 	if (!is_ready) {
1009 		DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: instance %d: "
1010 		    "CPU not ready for x-calls\n", ddi_get_instance(dip)));
1011 	} else if (!(is_ready = cpudrv_power_ready(cpudsp->cp))) {
1012 		DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: instance %d: "
1013 		    "waiting for all CPUs to be power manageable\n",
1014 		    ddi_get_instance(dip)));
1015 	}
1016 	if (!is_ready) {
1017 		/*
1018 		 * Make sure that we are busy so that framework doesn't
1019 		 * try to bring us down in this situation.
1020 		 */
1021 		CPUDRV_MONITOR_PM_BUSY_COMP(dip, cpupm);
1022 		CPUDRV_MONITOR_INIT(cpudsp);
1023 		mutex_exit(&cpudsp->lock);
1024 		goto do_return;
1025 	}
1026 
1027 	/*
1028 	 * Make sure that we are still not at unknown power level.
1029 	 */
1030 	if (cur_spd == NULL) {
1031 		DPRINTF(D_PM_MONITOR, ("cpudrv_monitor: instance %d: "
1032 		    "cur_spd is unknown\n", ddi_get_instance(dip)));
1033 		CPUDRV_MONITOR_PM_BUSY_AND_RAISE(dip, cpudsp, cpupm,
1034 		    CPUDRV_TOPSPEED(cpupm));
1035 		/*
1036 		 * We just changed the speed. Wait till at least next
1037 		 * call to this routine before proceeding ahead.
1038 		 */
1039 		CPUDRV_MONITOR_INIT(cpudsp);
1040 		mutex_exit(&cpudsp->lock);
1041 		goto do_return;
1042 	}
1043 
1044 	mutex_enter(&cpu_lock);
1045 	if (cpudsp->cp == NULL &&
1046 	    (cpudsp->cp = cpu_get(cpudsp->cpu_id)) == NULL) {
1047 		mutex_exit(&cpu_lock);
1048 		CPUDRV_MONITOR_INIT(cpudsp);
1049 		mutex_exit(&cpudsp->lock);
1050 		cmn_err(CE_WARN, "cpudrv_monitor: instance %d: can't get "
1051 		    "cpu_t", ddi_get_instance(dip));
1052 		goto do_return;
1053 	}
1054 
1055 	if (!cpupm->pm_started) {
1056 		cpupm->pm_started = B_TRUE;
1057 		cpudrv_set_supp_freqs(cpudsp);
1058 	}
1059 
1060 	get_cpu_mstate(cpudsp->cp, msnsecs);
1061 	GET_CPU_MSTATE_CNT(CMS_IDLE, idle_cnt);
1062 	GET_CPU_MSTATE_CNT(CMS_USER, user_cnt);
1063 	GET_CPU_MSTATE_CNT(CMS_SYSTEM, system_cnt);
1064 
1065 	/*
1066 	 * We can't do anything when we have just switched to a state
1067 	 * because there is no valid timestamp.
1068 	 */
1069 	if (cpupm->lastquan_ticks == 0) {
1070 		cpupm->lastquan_ticks = NSEC_TO_TICK(gethrtime());
1071 		mutex_exit(&cpu_lock);
1072 		CPUDRV_MONITOR_INIT(cpudsp);
1073 		mutex_exit(&cpudsp->lock);
1074 		goto do_return;
1075 	}
1076 
1077 	/*
1078 	 * Various watermarks are based on this routine being called back
1079 	 * exactly at the requested period. This is not guaranteed
1080 	 * because this routine is called from a taskq that is dispatched
1081 	 * from a timeout routine.  Handle this by finding out how many
1082 	 * ticks have elapsed since the last call and adjusting
1083 	 * the idle_cnt based on the delay added to the requested period
1084 	 * by timeout and taskq.
1085 	 */
1086 	ticks = NSEC_TO_TICK(gethrtime());
1087 	tick_cnt = ticks - cpupm->lastquan_ticks;
1088 	ASSERT(tick_cnt != 0);
1089 	cpupm->lastquan_ticks = ticks;
1090 	mutex_exit(&cpu_lock);
1091 	/*
1092 	 * Time taken between recording the current counts and
1093 	 * arranging the next call of this routine is an error in our
1094 	 * calculation. We minimize the error by calling
1095 	 * CPUDRV_MONITOR_INIT() here instead of end of this routine.
1096 	 */
1097 	CPUDRV_MONITOR_INIT(cpudsp);
1098 	DPRINTF(D_PM_MONITOR_VERBOSE, ("cpudrv_monitor: instance %d: "
1099 	    "idle count %d, user count %d, system count %d, pm_level %d, "
1100 	    "pm_busycnt %d\n", ddi_get_instance(dip), idle_cnt, user_cnt,
1101 	    system_cnt, cur_spd->pm_level, cpupm->pm_busycnt));
1102 
1103 #ifdef	DEBUG
1104 	/*
1105 	 * Notify that timeout and taskq has caused delays and we need to
1106 	 * scale our parameters accordingly.
1107 	 *
1108 	 * To get accurate result, don't turn on other DPRINTFs with
1109 	 * the following DPRINTF. PROM calls generated by other
1110 	 * DPRINTFs changes the timing.
1111 	 */
1112 	if (tick_cnt > cur_spd->quant_cnt) {
1113 		DPRINTF(D_PM_MONITOR_DELAY, ("cpudrv_monitor: instance %d: "
1114 		    "tick count %d > quantum_count %u\n",
1115 		    ddi_get_instance(dip), tick_cnt, cur_spd->quant_cnt));
1116 	}
1117 #endif	/* DEBUG */
1118 
1119 	/*
1120 	 * Adjust counts based on the delay added by timeout and taskq.
1121 	 */
1122 	idle_cnt = (idle_cnt * cur_spd->quant_cnt) / tick_cnt;
1123 	user_cnt = (user_cnt * cur_spd->quant_cnt) / tick_cnt;
1124 
1125 	if ((user_cnt > cur_spd->user_hwm) || (idle_cnt < cur_spd->idle_lwm &&
1126 	    cur_spd->idle_blwm_cnt >= cpudrv_idle_blwm_cnt_max)) {
1127 		cur_spd->idle_blwm_cnt = 0;
1128 		cur_spd->idle_bhwm_cnt = 0;
1129 		/*
1130 		 * In normal situation, arrange to go to next higher speed.
1131 		 * If we are running in special direct pm mode, we just stay
1132 		 * at the current speed.
1133 		 */
1134 		if (cur_spd == cur_spd->up_spd || cpudrv_direct_pm) {
1135 			CPUDRV_MONITOR_PM_BUSY_COMP(dip, cpupm);
1136 		} else {
1137 			new_spd = cur_spd->up_spd;
1138 			CPUDRV_MONITOR_PM_BUSY_AND_RAISE(dip, cpudsp, cpupm,
1139 			    new_spd);
1140 		}
1141 	} else if ((user_cnt <= cur_spd->user_lwm) &&
1142 	    (idle_cnt >= cur_spd->idle_hwm) || !CPU_ACTIVE(cpudsp->cp)) {
1143 		cur_spd->idle_blwm_cnt = 0;
1144 		cur_spd->idle_bhwm_cnt = 0;
1145 		/*
1146 		 * Arrange to go to next lower speed by informing our idle
1147 		 * status to the power management framework.
1148 		 */
1149 		CPUDRV_MONITOR_PM_IDLE_COMP(dip, cpupm);
1150 	} else {
1151 		/*
1152 		 * If we are between the idle water marks and have not
1153 		 * been here enough consecutive times to be considered
1154 		 * busy, just increment the count and return.
1155 		 */
1156 		if ((idle_cnt < cur_spd->idle_hwm) &&
1157 		    (idle_cnt >= cur_spd->idle_lwm) &&
1158 		    (cur_spd->idle_bhwm_cnt < cpudrv_idle_bhwm_cnt_max)) {
1159 			cur_spd->idle_blwm_cnt = 0;
1160 			cur_spd->idle_bhwm_cnt++;
1161 			mutex_exit(&cpudsp->lock);
1162 			goto do_return;
1163 		}
1164 		if (idle_cnt < cur_spd->idle_lwm) {
1165 			cur_spd->idle_blwm_cnt++;
1166 			cur_spd->idle_bhwm_cnt = 0;
1167 		}
1168 		/*
1169 		 * Arranges to stay at the current speed.
1170 		 */
1171 		CPUDRV_MONITOR_PM_BUSY_COMP(dip, cpupm);
1172 	}
1173 	mutex_exit(&cpudsp->lock);
1174 do_return:
1175 	mutex_enter(&cpupm->timeout_lock);
1176 	ASSERT(cpupm->timeout_count > 0);
1177 	cpupm->timeout_count--;
1178 	cv_signal(&cpupm->timeout_cv);
1179 	mutex_exit(&cpupm->timeout_lock);
1180 }
1181