xref: /illumos-gate/usr/src/uts/common/os/pg.c (revision b885580b)
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 #include <sys/systm.h>
27 #include <sys/types.h>
28 #include <sys/param.h>
29 #include <sys/thread.h>
30 #include <sys/cpuvar.h>
31 #include <sys/cpupart.h>
32 #include <sys/kmem.h>
33 #include <sys/cmn_err.h>
34 #include <sys/kstat.h>
35 #include <sys/processor.h>
36 #include <sys/disp.h>
37 #include <sys/group.h>
38 #include <sys/pg.h>
39 
40 /*
41  * Processor groups
42  *
43  * With the introduction of Chip Multi-Threaded (CMT) processor architectures,
44  * it is no longer necessarily true that a given physical processor module
45  * will present itself as a single schedulable entity (cpu_t). Rather, each
46  * chip and/or processor core may present itself as one or more "logical" CPUs.
47  *
48  * The logical CPUs presented may share physical components such as caches,
49  * data pipes, execution pipelines, FPUs, etc. It is advantageous to have the
50  * kernel be aware of the relationships existing between logical CPUs so that
51  * the appropriate optmizations may be employed.
52  *
53  * The processor group abstraction represents a set of logical CPUs that
54  * generally share some sort of physical or characteristic relationship.
55  *
56  * In the case of a physical sharing relationship, the CPUs in the group may
57  * share a pipeline, cache or floating point unit. In the case of a logical
58  * relationship, a PG may represent the set of CPUs in a processor set, or the
59  * set of CPUs running at a particular clock speed.
60  *
61  * The generic processor group structure, pg_t, contains the elements generic
62  * to a group of CPUs. Depending on the nature of the CPU relationship
63  * (LOGICAL or PHYSICAL), a pointer to a pg may be recast to a "view" of that
64  * PG where more specific data is represented.
65  *
66  * As an example, a PG representing a PHYSICAL relationship, may be recast to
67  * a pghw_t, where data further describing the hardware sharing relationship
68  * is maintained. See pghw.c and pghw.h for details on physical PGs.
69  *
70  * At this time a more specialized casting of a PG representing a LOGICAL
71  * relationship has not been implemented, but the architecture allows for this
72  * in the future.
73  *
74  * Processor Group Classes
75  *
76  * Processor group consumers may wish to maintain and associate specific
77  * data with the PGs they create. For this reason, a mechanism for creating
78  * class specific PGs exists. Classes may overload the default functions for
79  * creating, destroying, and associating CPUs with PGs, and may also register
80  * class specific callbacks to be invoked when the CPU related system
81  * configuration changes. Class specific data is stored/associated with
82  * PGs by incorporating the pg_t (or pghw_t, as appropriate), as the first
83  * element of a class specific PG object. In memory, such a structure may look
84  * like:
85  *
86  * ----------------------- - - -
87  * | common              | | | |  <--(pg_t *)
88  * ----------------------- | | -
89  * | HW specific         | | | <-----(pghw_t *)
90  * ----------------------- | -
91  * | class specific      | | <-------(pg_cmt_t *)
92  * ----------------------- -
93  *
94  * Access to the PG class specific data can be had by casting a pointer to
95  * it's class specific view.
96  */
97 
98 static pg_t		*pg_alloc_default(pg_class_t);
99 static void		pg_free_default(pg_t *);
100 static void		pg_null_op();
101 
102 /*
103  * Bootstrap CPU specific PG data
104  * See pg_cpu_bootstrap()
105  */
106 static cpu_pg_t		bootstrap_pg_data;
107 
108 /*
109  * Bitset of allocated PG ids (they are sequential)
110  * and the next free id in the set.
111  */
112 static bitset_t		pg_id_set;
113 
114 /*
115  * ID space starts from 1 to assume that root has ID 0;
116  */
117 static pgid_t		pg_id_next = 1;
118 
119 /*
120  * Default and externed PG ops vectors
121  */
122 static struct pg_ops pg_ops_default = {
123 	pg_alloc_default,	/* alloc */
124 	pg_free_default,	/* free */
125 	NULL,			/* cpu_init */
126 	NULL,			/* cpu_fini */
127 	NULL,			/* cpu_active */
128 	NULL,			/* cpu_inactive */
129 	NULL,			/* cpupart_in */
130 	NULL,			/* cpupart_out */
131 	NULL,			/* cpupart_move */
132 	NULL,			/* cpu_belongs */
133 	NULL,			/* policy_name */
134 };
135 
136 static struct pg_cb_ops pg_cb_ops_default = {
137 	pg_null_op,		/* thread_swtch */
138 	pg_null_op,		/* thread_remain */
139 };
140 
141 /*
142  * Class specific PG allocation callbacks
143  */
144 #define	PG_ALLOC(class)							\
145 	(pg_classes[class].pgc_ops->alloc ?				\
146 	    pg_classes[class].pgc_ops->alloc() :			\
147 	    pg_classes[pg_default_cid].pgc_ops->alloc())
148 
149 #define	PG_FREE(pg)							\
150 	((pg)->pg_class->pgc_ops->free ?				\
151 	    (pg)->pg_class->pgc_ops->free(pg) :				\
152 	    pg_classes[pg_default_cid].pgc_ops->free(pg))		\
153 
154 
155 /*
156  * Class specific PG policy name
157  */
158 #define	PG_POLICY_NAME(pg)						\
159 	((pg)->pg_class->pgc_ops->policy_name ?				\
160 	    (pg)->pg_class->pgc_ops->policy_name(pg) : NULL)		\
161 
162 /*
163  * Class specific membership test callback
164  */
165 #define	PG_CPU_BELONGS(pg, cp)						\
166 	((pg)->pg_class->pgc_ops->cpu_belongs ?				\
167 	    (pg)->pg_class->pgc_ops->cpu_belongs(pg, cp) : 0)		\
168 
169 /*
170  * CPU configuration callbacks
171  */
172 #define	PG_CPU_INIT(class, cp, cpu_pg)					\
173 {									\
174 	if (pg_classes[class].pgc_ops->cpu_init)			\
175 		pg_classes[class].pgc_ops->cpu_init(cp, cpu_pg);	\
176 }
177 
178 #define	PG_CPU_FINI(class, cp, cpu_pg)					\
179 {									\
180 	if (pg_classes[class].pgc_ops->cpu_fini)			\
181 		pg_classes[class].pgc_ops->cpu_fini(cp, cpu_pg);	\
182 }
183 
184 #define	PG_CPU_ACTIVE(class, cp)					\
185 {									\
186 	if (pg_classes[class].pgc_ops->cpu_active)			\
187 		pg_classes[class].pgc_ops->cpu_active(cp);		\
188 }
189 
190 #define	PG_CPU_INACTIVE(class, cp)					\
191 {									\
192 	if (pg_classes[class].pgc_ops->cpu_inactive)			\
193 		pg_classes[class].pgc_ops->cpu_inactive(cp);		\
194 }
195 
196 /*
197  * CPU / cpupart configuration callbacks
198  */
199 #define	PG_CPUPART_IN(class, cp, pp)					\
200 {									\
201 	if (pg_classes[class].pgc_ops->cpupart_in)			\
202 		pg_classes[class].pgc_ops->cpupart_in(cp, pp);		\
203 }
204 
205 #define	PG_CPUPART_OUT(class, cp, pp)					\
206 {									\
207 	if (pg_classes[class].pgc_ops->cpupart_out)			\
208 		pg_classes[class].pgc_ops->cpupart_out(cp, pp);		\
209 }
210 
211 #define	PG_CPUPART_MOVE(class, cp, old, new)				\
212 {									\
213 	if (pg_classes[class].pgc_ops->cpupart_move)			\
214 		pg_classes[class].pgc_ops->cpupart_move(cp, old, new);	\
215 }
216 
217 
218 
219 static pg_class_t	*pg_classes;
220 static int		pg_nclasses;
221 
222 static pg_cid_t		pg_default_cid;
223 
224 /*
225  * Initialze common PG subsystem.
226  */
227 void
pg_init(void)228 pg_init(void)
229 {
230 	extern void pg_cmt_class_init();
231 	extern void pg_cmt_cpu_startup();
232 
233 	pg_default_cid =
234 	    pg_class_register("default", &pg_ops_default, PGR_LOGICAL);
235 
236 	/*
237 	 * Initialize classes to allow them to register with the framework
238 	 */
239 	pg_cmt_class_init();
240 
241 	pg_cpu0_init();
242 	pg_cmt_cpu_startup(CPU);
243 }
244 
245 /*
246  * Perform CPU 0 initialization
247  */
248 void
pg_cpu0_init(void)249 pg_cpu0_init(void)
250 {
251 	extern void pghw_physid_create();
252 
253 	/*
254 	 * Create the physical ID cache for the boot CPU
255 	 */
256 	pghw_physid_create(CPU);
257 
258 	/*
259 	 * pg_cpu_* require that cpu_lock be held
260 	 */
261 	mutex_enter(&cpu_lock);
262 
263 	(void) pg_cpu_init(CPU, B_FALSE);
264 	pg_cpupart_in(CPU, &cp_default);
265 	pg_cpu_active(CPU);
266 
267 	mutex_exit(&cpu_lock);
268 }
269 
270 /*
271  * Invoked when topology for CPU0 changes
272  * post pg_cpu0_init().
273  *
274  * Currently happens as a result of null_proc_lpa
275  * on Starcat.
276  */
277 void
pg_cpu0_reinit(void)278 pg_cpu0_reinit(void)
279 {
280 	mutex_enter(&cpu_lock);
281 	pg_cpu_inactive(CPU);
282 	pg_cpupart_out(CPU, &cp_default);
283 	pg_cpu_fini(CPU, NULL);
284 
285 	(void) pg_cpu_init(CPU, B_FALSE);
286 	pg_cpupart_in(CPU, &cp_default);
287 	pg_cpu_active(CPU);
288 	mutex_exit(&cpu_lock);
289 }
290 
291 /*
292  * Register a new PG class
293  */
294 pg_cid_t
pg_class_register(char * name,struct pg_ops * ops,pg_relation_t relation)295 pg_class_register(char *name, struct pg_ops *ops, pg_relation_t relation)
296 {
297 	pg_class_t	*newclass;
298 	pg_class_t	*classes_old;
299 	id_t		cid;
300 
301 	mutex_enter(&cpu_lock);
302 
303 	/*
304 	 * Allocate a new pg_class_t in the pg_classes array
305 	 */
306 	if (pg_nclasses == 0) {
307 		pg_classes = kmem_zalloc(sizeof (pg_class_t), KM_SLEEP);
308 	} else {
309 		classes_old = pg_classes;
310 		pg_classes =
311 		    kmem_zalloc(sizeof (pg_class_t) * (pg_nclasses + 1),
312 		    KM_SLEEP);
313 		(void) kcopy(classes_old, pg_classes,
314 		    sizeof (pg_class_t) * pg_nclasses);
315 		kmem_free(classes_old, sizeof (pg_class_t) * pg_nclasses);
316 	}
317 
318 	cid = pg_nclasses++;
319 	newclass = &pg_classes[cid];
320 
321 	(void) strncpy(newclass->pgc_name, name, PG_CLASS_NAME_MAX);
322 	newclass->pgc_id = cid;
323 	newclass->pgc_ops = ops;
324 	newclass->pgc_relation = relation;
325 
326 	mutex_exit(&cpu_lock);
327 
328 	return (cid);
329 }
330 
331 /*
332  * Try to find an existing pg in set in which to place cp.
333  * Returns the pg if found, and NULL otherwise.
334  * In the event that the CPU could belong to multiple
335  * PGs in the set, the first matching PG will be returned.
336  */
337 pg_t *
pg_cpu_find_pg(cpu_t * cp,group_t * set)338 pg_cpu_find_pg(cpu_t *cp, group_t *set)
339 {
340 	pg_t		*pg;
341 	group_iter_t	i;
342 
343 	group_iter_init(&i);
344 	while ((pg = group_iterate(set, &i)) != NULL) {
345 		/*
346 		 * Ask the class if the CPU belongs here
347 		 */
348 		if (PG_CPU_BELONGS(pg, cp))
349 			return (pg);
350 	}
351 	return (NULL);
352 }
353 
354 /*
355  * Iterate over the CPUs in a PG after initializing
356  * the iterator with PG_CPU_ITR_INIT()
357  */
358 cpu_t *
pg_cpu_next(pg_cpu_itr_t * itr)359 pg_cpu_next(pg_cpu_itr_t *itr)
360 {
361 	cpu_t		*cpu;
362 	pg_t		*pg = itr->pg;
363 
364 	cpu = group_iterate(&pg->pg_cpus, &itr->position);
365 	return (cpu);
366 }
367 
368 /*
369  * Test if a given PG contains a given CPU
370  */
371 boolean_t
pg_cpu_find(pg_t * pg,cpu_t * cp)372 pg_cpu_find(pg_t *pg, cpu_t *cp)
373 {
374 	if (group_find(&pg->pg_cpus, cp) == (uint_t)-1)
375 		return (B_FALSE);
376 
377 	return (B_TRUE);
378 }
379 
380 /*
381  * Set the PGs callbacks to the default
382  */
383 void
pg_callback_set_defaults(pg_t * pg)384 pg_callback_set_defaults(pg_t *pg)
385 {
386 	bcopy(&pg_cb_ops_default, &pg->pg_cb, sizeof (struct pg_cb_ops));
387 }
388 
389 /*
390  * Create a PG of a given class.
391  * This routine may block.
392  */
393 pg_t *
pg_create(pg_cid_t cid)394 pg_create(pg_cid_t cid)
395 {
396 	pg_t	*pg;
397 	pgid_t	id;
398 
399 	ASSERT(MUTEX_HELD(&cpu_lock));
400 
401 	/*
402 	 * Call the class specific PG allocation routine
403 	 */
404 	pg = PG_ALLOC(cid);
405 	pg->pg_class = &pg_classes[cid];
406 	pg->pg_relation = pg->pg_class->pgc_relation;
407 
408 	/*
409 	 * Find the next free sequential pg id
410 	 */
411 	do {
412 		if (pg_id_next >= bitset_capacity(&pg_id_set))
413 			bitset_resize(&pg_id_set, pg_id_next + 1);
414 		id = pg_id_next++;
415 	} while (bitset_in_set(&pg_id_set, id));
416 
417 	pg->pg_id = id;
418 	bitset_add(&pg_id_set, pg->pg_id);
419 
420 	/*
421 	 * Create the PG's CPU group
422 	 */
423 	group_create(&pg->pg_cpus);
424 
425 	/*
426 	 * Initialize the events ops vector
427 	 */
428 	pg_callback_set_defaults(pg);
429 
430 	return (pg);
431 }
432 
433 /*
434  * Destroy a PG.
435  * This routine may block.
436  */
437 void
pg_destroy(pg_t * pg)438 pg_destroy(pg_t *pg)
439 {
440 	ASSERT(MUTEX_HELD(&cpu_lock));
441 
442 	group_destroy(&pg->pg_cpus);
443 
444 	/*
445 	 * Unassign the pg_id
446 	 */
447 	if (pg_id_next > pg->pg_id)
448 		pg_id_next = pg->pg_id;
449 	bitset_del(&pg_id_set, pg->pg_id);
450 
451 	/*
452 	 * Invoke the class specific de-allocation routine
453 	 */
454 	PG_FREE(pg);
455 }
456 
457 /*
458  * Add the CPU "cp" to processor group "pg"
459  * This routine may block.
460  */
461 void
pg_cpu_add(pg_t * pg,cpu_t * cp,cpu_pg_t * cpu_pg)462 pg_cpu_add(pg_t *pg, cpu_t *cp, cpu_pg_t *cpu_pg)
463 {
464 	int	err;
465 
466 	ASSERT(MUTEX_HELD(&cpu_lock));
467 
468 	/* This adds the CPU to the PG's CPU group */
469 	err = group_add(&pg->pg_cpus, cp, GRP_RESIZE);
470 	ASSERT(err == 0);
471 
472 	/*
473 	 * The CPU should be referencing the bootstrap PG data still
474 	 * at this point, since this routine may block causing us to
475 	 * enter the dispatcher.
476 	 */
477 	ASSERT(pg_cpu_is_bootstrapped(cp));
478 
479 	/* This adds the PG to the CPUs PG group */
480 	err = group_add(&cpu_pg->pgs, pg, GRP_RESIZE);
481 	ASSERT(err == 0);
482 }
483 
484 /*
485  * Remove "cp" from "pg".
486  * This routine may block.
487  */
488 void
pg_cpu_delete(pg_t * pg,cpu_t * cp,cpu_pg_t * cpu_pg)489 pg_cpu_delete(pg_t *pg, cpu_t *cp, cpu_pg_t *cpu_pg)
490 {
491 	int	err;
492 
493 	ASSERT(MUTEX_HELD(&cpu_lock));
494 
495 	/* Remove the CPU from the PG */
496 	err = group_remove(&pg->pg_cpus, cp, GRP_RESIZE);
497 	ASSERT(err == 0);
498 
499 	/*
500 	 * The CPU should be referencing the bootstrap PG data still
501 	 * at this point, since this routine may block causing us to
502 	 * enter the dispatcher.
503 	 */
504 	ASSERT(pg_cpu_is_bootstrapped(cp));
505 
506 	/* Remove the PG from the CPU's PG group */
507 	err = group_remove(&cpu_pg->pgs, pg, GRP_RESIZE);
508 	ASSERT(err == 0);
509 }
510 
511 /*
512  * Allocate a CPU's PG data. This hangs off struct cpu at cpu_pg
513  */
514 static cpu_pg_t *
pg_cpu_data_alloc(void)515 pg_cpu_data_alloc(void)
516 {
517 	cpu_pg_t	*pgd;
518 
519 	pgd = kmem_zalloc(sizeof (cpu_pg_t), KM_SLEEP);
520 	group_create(&pgd->pgs);
521 	group_create(&pgd->cmt_pgs);
522 
523 	return (pgd);
524 }
525 
526 /*
527  * Free the CPU's PG data.
528  */
529 static void
pg_cpu_data_free(cpu_pg_t * pgd)530 pg_cpu_data_free(cpu_pg_t *pgd)
531 {
532 	group_destroy(&pgd->pgs);
533 	group_destroy(&pgd->cmt_pgs);
534 	kmem_free(pgd, sizeof (cpu_pg_t));
535 }
536 
537 /*
538  * Called when either a new CPU is coming into the system (either
539  * via booting or DR) or when the CPU's PG data is being recalculated.
540  * Allocate its PG data, and notify all registered classes about
541  * the new CPU.
542  *
543  * If "deferred_init" is B_TRUE, the CPU's PG data will be allocated
544  * and returned, but the "bootstrap" structure will be left in place.
545  * The deferred_init option is used when all CPUs in the system are
546  * using the bootstrap structure as part of the process of recalculating
547  * all PG data. The caller must replace the bootstrap structure with the
548  * allocated PG data before pg_cpu_active is called.
549  *
550  * This routine may block.
551  */
552 cpu_pg_t *
pg_cpu_init(cpu_t * cp,boolean_t deferred_init)553 pg_cpu_init(cpu_t *cp, boolean_t deferred_init)
554 {
555 	pg_cid_t	i;
556 	cpu_pg_t	*cpu_pg;
557 
558 	ASSERT(MUTEX_HELD(&cpu_lock));
559 
560 	/*
561 	 * Allocate and size the per CPU pg data
562 	 *
563 	 * The CPU's PG data will be populated by the various
564 	 * PG classes during the invocation of the PG_CPU_INIT()
565 	 * callback below.
566 	 *
567 	 * Since the we could block and enter the dispatcher during
568 	 * this process, the CPU will continue to reference the bootstrap
569 	 * PG data until all the initialization completes.
570 	 */
571 	ASSERT(pg_cpu_is_bootstrapped(cp));
572 
573 	cpu_pg = pg_cpu_data_alloc();
574 
575 	/*
576 	 * Notify all registered classes about the new CPU
577 	 */
578 	for (i = 0; i < pg_nclasses; i++)
579 		PG_CPU_INIT(i, cp, cpu_pg);
580 
581 	/*
582 	 * The CPU's PG data is now ready to use.
583 	 */
584 	if (deferred_init == B_FALSE)
585 		cp->cpu_pg = cpu_pg;
586 
587 	return (cpu_pg);
588 }
589 
590 /*
591  * Either this CPU is being deleted from the system or its PG data is
592  * being recalculated. Notify the classes and free up the CPU's PG data.
593  *
594  * If "cpu_pg_deferred" is non-NULL, it points to the CPU's PG data and
595  * serves to indicate that this CPU is already using the bootstrap
596  * stucture. Used as part of the process to recalculate the PG data for
597  * all CPUs in the system.
598  */
599 void
pg_cpu_fini(cpu_t * cp,cpu_pg_t * cpu_pg_deferred)600 pg_cpu_fini(cpu_t *cp, cpu_pg_t *cpu_pg_deferred)
601 {
602 	pg_cid_t	i;
603 	cpu_pg_t	*cpu_pg;
604 
605 	ASSERT(MUTEX_HELD(&cpu_lock));
606 
607 	if (cpu_pg_deferred == NULL) {
608 		cpu_pg = cp->cpu_pg;
609 
610 		/*
611 		 * This can happen if the CPU coming into the system
612 		 * failed to power on.
613 		 */
614 		if (cpu_pg == NULL || pg_cpu_is_bootstrapped(cp))
615 			return;
616 
617 		/*
618 		 * Have the CPU reference the bootstrap PG data to survive
619 		 * the dispatcher should it block from here on out.
620 		 */
621 		pg_cpu_bootstrap(cp);
622 	} else {
623 		ASSERT(pg_cpu_is_bootstrapped(cp));
624 		cpu_pg = cpu_pg_deferred;
625 	}
626 
627 	for (i = 0; i < pg_nclasses; i++)
628 		PG_CPU_FINI(i, cp, cpu_pg);
629 
630 	pg_cpu_data_free(cpu_pg);
631 }
632 
633 /*
634  * This CPU is becoming active (online)
635  * This routine may not block as it is called from paused CPUs
636  * context.
637  */
638 void
pg_cpu_active(cpu_t * cp)639 pg_cpu_active(cpu_t *cp)
640 {
641 	pg_cid_t	i;
642 
643 	ASSERT(MUTEX_HELD(&cpu_lock));
644 
645 	/*
646 	 * Notify all registered classes about the new CPU
647 	 */
648 	for (i = 0; i < pg_nclasses; i++)
649 		PG_CPU_ACTIVE(i, cp);
650 }
651 
652 /*
653  * This CPU is going inactive (offline)
654  * This routine may not block, as it is called from paused
655  * CPUs context.
656  */
657 void
pg_cpu_inactive(cpu_t * cp)658 pg_cpu_inactive(cpu_t *cp)
659 {
660 	pg_cid_t	i;
661 
662 	ASSERT(MUTEX_HELD(&cpu_lock));
663 
664 	/*
665 	 * Notify all registered classes about the new CPU
666 	 */
667 	for (i = 0; i < pg_nclasses; i++)
668 		PG_CPU_INACTIVE(i, cp);
669 }
670 
671 /*
672  * Invoked when the CPU is about to move into the partition
673  * This routine may block.
674  */
675 void
pg_cpupart_in(cpu_t * cp,cpupart_t * pp)676 pg_cpupart_in(cpu_t *cp, cpupart_t *pp)
677 {
678 	int	i;
679 
680 	ASSERT(MUTEX_HELD(&cpu_lock));
681 
682 	/*
683 	 * Notify all registered classes that the
684 	 * CPU is about to enter the CPU partition
685 	 */
686 	for (i = 0; i < pg_nclasses; i++)
687 		PG_CPUPART_IN(i, cp, pp);
688 }
689 
690 /*
691  * Invoked when the CPU is about to move out of the partition
692  * This routine may block.
693  */
694 /*ARGSUSED*/
695 void
pg_cpupart_out(cpu_t * cp,cpupart_t * pp)696 pg_cpupart_out(cpu_t *cp, cpupart_t *pp)
697 {
698 	int	i;
699 
700 	ASSERT(MUTEX_HELD(&cpu_lock));
701 
702 	/*
703 	 * Notify all registered classes that the
704 	 * CPU is about to leave the CPU partition
705 	 */
706 	for (i = 0; i < pg_nclasses; i++)
707 		PG_CPUPART_OUT(i, cp, pp);
708 }
709 
710 /*
711  * Invoked when the CPU is *moving* partitions.
712  *
713  * This routine may not block, as it is called from paused CPUs
714  * context.
715  */
716 void
pg_cpupart_move(cpu_t * cp,cpupart_t * oldpp,cpupart_t * newpp)717 pg_cpupart_move(cpu_t *cp, cpupart_t *oldpp, cpupart_t *newpp)
718 {
719 	int	i;
720 
721 	ASSERT(MUTEX_HELD(&cpu_lock));
722 
723 	/*
724 	 * Notify all registered classes that the
725 	 * CPU is about to leave the CPU partition
726 	 */
727 	for (i = 0; i < pg_nclasses; i++)
728 		PG_CPUPART_MOVE(i, cp, oldpp, newpp);
729 }
730 
731 /*
732  * Return a class specific string describing a policy implemented
733  * across this PG
734  */
735 char *
pg_policy_name(pg_t * pg)736 pg_policy_name(pg_t *pg)
737 {
738 	char *str;
739 	if ((str = PG_POLICY_NAME(pg)) != NULL)
740 		return (str);
741 
742 	return ("N/A");
743 }
744 
745 /*
746  * Provide the specified CPU a bootstrap pg
747  * This is needed to allow sane behaviour if any PG consuming
748  * code needs to deal with a partially initialized CPU
749  */
750 void
pg_cpu_bootstrap(cpu_t * cp)751 pg_cpu_bootstrap(cpu_t *cp)
752 {
753 	cp->cpu_pg = &bootstrap_pg_data;
754 }
755 
756 /*
757  * Return non-zero if the specified CPU is bootstrapped,
758  * which means it's CPU specific PG data has not yet been
759  * fully constructed.
760  */
761 int
pg_cpu_is_bootstrapped(cpu_t * cp)762 pg_cpu_is_bootstrapped(cpu_t *cp)
763 {
764 	return (cp->cpu_pg == &bootstrap_pg_data);
765 }
766 
767 /*ARGSUSED*/
768 static pg_t *
pg_alloc_default(pg_class_t class)769 pg_alloc_default(pg_class_t class)
770 {
771 	return (kmem_zalloc(sizeof (pg_t), KM_SLEEP));
772 }
773 
774 /*ARGSUSED*/
775 static void
pg_free_default(struct pg * pg)776 pg_free_default(struct pg *pg)
777 {
778 	kmem_free(pg, sizeof (pg_t));
779 }
780 
781 static void
pg_null_op()782 pg_null_op()
783 {
784 }
785 
786 /*
787  * Invoke the "thread switch" callback for each of the CPU's PGs
788  * This is invoked from the dispatcher swtch() routine, which is called
789  * when a thread running an a CPU should switch to another thread.
790  * "cp" is the CPU on which the thread switch is happening
791  * "now" is an unscaled hrtime_t timestamp taken in swtch()
792  * "old" and "new" are the outgoing and incoming threads, respectively.
793  */
794 void
pg_ev_thread_swtch(struct cpu * cp,hrtime_t now,kthread_t * old,kthread_t * new)795 pg_ev_thread_swtch(struct cpu *cp, hrtime_t now, kthread_t *old, kthread_t *new)
796 {
797 	int	i, sz;
798 	group_t	*grp;
799 	pg_t	*pg;
800 
801 	grp = &cp->cpu_pg->pgs;
802 	sz = GROUP_SIZE(grp);
803 	for (i = 0; i < sz; i++) {
804 		pg = GROUP_ACCESS(grp, i);
805 		pg->pg_cb.thread_swtch(pg, cp, now, old, new);
806 	}
807 }
808 
809 /*
810  * Invoke the "thread remain" callback for each of the CPU's PGs.
811  * This is called from the dispatcher's swtch() routine when a thread
812  * running on the CPU "cp" is switching to itself, which can happen as an
813  * artifact of the thread's timeslice expiring.
814  */
815 void
pg_ev_thread_remain(struct cpu * cp,kthread_t * t)816 pg_ev_thread_remain(struct cpu *cp, kthread_t *t)
817 {
818 	int	i, sz;
819 	group_t	*grp;
820 	pg_t	*pg;
821 
822 	grp = &cp->cpu_pg->pgs;
823 	sz = GROUP_SIZE(grp);
824 	for (i = 0; i < sz; i++) {
825 		pg = GROUP_ACCESS(grp, i);
826 		pg->pg_cb.thread_remain(pg, cp, t);
827 	}
828 }
829