xref: /dragonfly/sys/kern/kern_objcache.c (revision 3f5e28f4)
1 /*
2  * Copyright (c) 2005 Jeffrey M. Hsu.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Jeffrey M. Hsu.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of The DragonFly Project nor the names of its
16  *    contributors may be used to endorse or promote products derived
17  *    from this software without specific, prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
21  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
22  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
23  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
24  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
25  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
26  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
27  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
28  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
29  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  * $DragonFly: src/sys/kern/kern_objcache.c,v 1.18 2007/04/30 07:18:53 dillon Exp $
33  */
34 
35 #include <sys/param.h>
36 #include <sys/kernel.h>
37 #include <sys/systm.h>
38 #include <sys/callout.h>
39 #include <sys/globaldata.h>
40 #include <sys/malloc.h>
41 #include <sys/queue.h>
42 #include <sys/objcache.h>
43 #include <sys/spinlock.h>
44 #include <sys/thread.h>
45 #include <sys/thread2.h>
46 #include <sys/spinlock2.h>
47 
48 static MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
49 static MALLOC_DEFINE(M_OBJMAG, "objcache magazine", "Object Cache Magazine");
50 
51 #define	INITIAL_MAG_CAPACITY	256
52 
53 struct magazine {
54 	int			 rounds;
55 	int			 capacity;
56 	int			 cleaning;
57 	SLIST_ENTRY(magazine)	 nextmagazine;
58 	void			*objects[];
59 };
60 
61 SLIST_HEAD(magazinelist, magazine);
62 
63 /*
64  * per-cluster cache of magazines
65  *
66  * All fields in this structure are protected by the spinlock.
67  */
68 struct magazinedepot {
69 	/*
70 	 * The per-cpu object caches only exchanges completely full or
71 	 * completely empty magazines with the depot layer, so only have
72 	 * to cache these two types of magazines.
73 	 */
74 	struct magazinelist	fullmagazines;
75 	struct magazinelist	emptymagazines;
76 	int			magcapacity;
77 
78 	/* protect this structure */
79 	struct spinlock		spin;
80 
81 	/* magazines not yet allocated towards limit */
82 	int			unallocated_objects;
83 
84 	/* infrequently used fields */
85 	int			waiting;	/* waiting for another cpu to
86 						 * return a full magazine to
87 						 * the depot */
88 	int			contested;	/* depot contention count */
89 };
90 
91 /*
92  * per-cpu object cache
93  * All fields in this structure are protected by crit_enter().
94  */
95 struct percpu_objcache {
96 	struct magazine	*loaded_magazine;	/* active magazine */
97 	struct magazine	*previous_magazine;	/* backup magazine */
98 
99 	/* statistics */
100 	int		gets_cumulative;	/* total calls to get */
101 	int		gets_null;		/* objcache_get returned NULL */
102 	int		puts_cumulative;	/* total calls to put */
103 	int		puts_othercluster;	/* returned to other cluster */
104 
105 	/* infrequently used fields */
106 	int		waiting;	/* waiting for a thread on this cpu to
107 					 * return an obj to the per-cpu cache */
108 };
109 
110 /* only until we have NUMA cluster topology information XXX */
111 #define MAXCLUSTERS 1
112 #define myclusterid 0
113 #define CLUSTER_OF(obj) 0
114 
115 /*
116  * Two-level object cache consisting of NUMA cluster-level depots of
117  * fully loaded or completely empty magazines and cpu-level caches of
118  * individual objects.
119  */
120 struct objcache {
121 	char			*name;
122 
123 	/* object constructor and destructor from blank storage */
124 	objcache_ctor_fn	*ctor;
125 	objcache_dtor_fn	*dtor;
126 	void			*private;
127 
128 	/* interface to underlying allocator */
129 	objcache_alloc_fn	*alloc;
130 	objcache_free_fn	*free;
131 	void			*allocator_args;
132 
133 	SLIST_ENTRY(objcache)	oc_next;
134 
135 	/* NUMA-cluster level caches */
136 	struct magazinedepot	depot[MAXCLUSTERS];
137 
138 	struct percpu_objcache	cache_percpu[];		/* per-cpu caches */
139 };
140 
141 static struct spinlock objcachelist_spin;
142 static SLIST_HEAD(objcachelist, objcache) allobjcaches;
143 
144 static struct magazine *
145 mag_alloc(int capacity)
146 {
147 	struct magazine *mag;
148 
149 	mag = kmalloc(__offsetof(struct magazine, objects[capacity]),
150 			M_OBJMAG, M_INTWAIT | M_ZERO);
151 	mag->capacity = capacity;
152 	mag->rounds = 0;
153 	mag->cleaning = 0;
154 	return (mag);
155 }
156 
157 /*
158  * Utility routine for objects that don't require any de-construction.
159  */
160 
161 static void
162 null_dtor(void *obj, void *private)
163 {
164 	/* do nothing */
165 }
166 
167 static boolean_t
168 null_ctor(void *obj, void *private, int ocflags)
169 {
170 	return TRUE;
171 }
172 
173 /*
174  * Create an object cache.
175  */
176 struct objcache *
177 objcache_create(const char *name, int cluster_limit, int mag_capacity,
178 		objcache_ctor_fn *ctor, objcache_dtor_fn *dtor, void *private,
179 		objcache_alloc_fn *alloc, objcache_free_fn *free,
180 		void *allocator_args)
181 {
182 	struct objcache *oc;
183 	struct magazinedepot *depot;
184 	int cpuid;
185 
186 	/* allocate object cache structure */
187 	oc = kmalloc(__offsetof(struct objcache, cache_percpu[ncpus]),
188 		    M_OBJCACHE, M_WAITOK | M_ZERO);
189 	oc->name = kstrdup(name, M_TEMP);
190 	oc->ctor = ctor ? ctor : null_ctor;
191 	oc->dtor = dtor ? dtor : null_dtor;
192 	oc->private = private;
193 	oc->free = free;
194 	oc->allocator_args = allocator_args;
195 
196 	/* initialize depots */
197 	depot = &oc->depot[0];
198 
199 	spin_init(&depot->spin);
200 	SLIST_INIT(&depot->fullmagazines);
201 	SLIST_INIT(&depot->emptymagazines);
202 
203 	if (mag_capacity == 0)
204 		mag_capacity = INITIAL_MAG_CAPACITY;
205 	depot->magcapacity = mag_capacity;
206 
207 	/*
208 	 * The cluster_limit must be sufficient to have three magazines per
209 	 * cpu.
210 	 */
211 	if (cluster_limit == 0) {
212 		depot->unallocated_objects = -1;
213 	} else {
214 		if (cluster_limit < mag_capacity * ncpus * 3)
215 			cluster_limit = mag_capacity * ncpus * 3;
216 		depot->unallocated_objects = cluster_limit;
217 	}
218 	oc->alloc = alloc;
219 
220 	/* initialize per-cpu caches */
221 	for (cpuid = 0; cpuid < ncpus; cpuid++) {
222 		struct percpu_objcache *cache_percpu = &oc->cache_percpu[cpuid];
223 
224 		cache_percpu->loaded_magazine = mag_alloc(mag_capacity);
225 		cache_percpu->previous_magazine = mag_alloc(mag_capacity);
226 	}
227 	spin_lock_wr(&objcachelist_spin);
228 	SLIST_INSERT_HEAD(&allobjcaches, oc, oc_next);
229 	spin_unlock_wr(&objcachelist_spin);
230 
231 	return (oc);
232 }
233 
234 struct objcache *
235 objcache_create_simple(malloc_type_t mtype, size_t objsize)
236 {
237 	struct objcache_malloc_args *margs;
238 	struct objcache *oc;
239 
240 	margs = kmalloc(sizeof(*margs), M_OBJCACHE, M_WAITOK|M_ZERO);
241 	margs->objsize = objsize;
242 	margs->mtype = mtype;
243 	oc = objcache_create(mtype->ks_shortdesc, 0, 0,
244 			     NULL, NULL, NULL,
245 			     objcache_malloc_alloc, objcache_malloc_free,
246 			     margs);
247 	return (oc);
248 }
249 
250 struct objcache *
251 objcache_create_mbacked(malloc_type_t mtype, size_t objsize,
252 			int cluster_limit, int mag_capacity,
253 			objcache_ctor_fn *ctor, objcache_dtor_fn *dtor,
254 			void *private)
255 {
256 	struct objcache_malloc_args *margs;
257 	struct objcache *oc;
258 
259 	margs = kmalloc(sizeof(*margs), M_OBJCACHE, M_WAITOK|M_ZERO);
260 	margs->objsize = objsize;
261 	margs->mtype = mtype;
262 	oc = objcache_create(mtype->ks_shortdesc,
263 			     cluster_limit, mag_capacity,
264 			     ctor, dtor, private,
265 			     objcache_malloc_alloc, objcache_malloc_free,
266 			     margs);
267 	return(oc);
268 }
269 
270 
271 #define MAGAZINE_EMPTY(mag)	(mag->rounds == 0)
272 #define MAGAZINE_NOTEMPTY(mag)	(mag->rounds != 0)
273 #define MAGAZINE_FULL(mag)	(mag->rounds == mag->capacity)
274 
275 #define	swap(x, y)	({ struct magazine *t = x; x = y; y = t; })
276 
277 /*
278  * Get an object from the object cache.
279  *
280  * WARNING!  ocflags are only used when we have to go to the underlying
281  * allocator, so we cannot depend on flags such as M_ZERO.
282  */
283 void *
284 objcache_get(struct objcache *oc, int ocflags)
285 {
286 	struct percpu_objcache *cpucache = &oc->cache_percpu[mycpuid];
287 	struct magazine *loadedmag;
288 	struct magazine *emptymag;
289 	void *obj;
290 	struct magazinedepot *depot;
291 
292 	KKASSERT((ocflags & M_ZERO) == 0);
293 	crit_enter();
294 	++cpucache->gets_cumulative;
295 
296 retry:
297 	/*
298 	 * Loaded magazine has an object.  This is the hot path.
299 	 * It is lock-free and uses a critical section to block
300 	 * out interrupt handlers on the same processor.
301 	 */
302 	loadedmag = cpucache->loaded_magazine;
303 	if (MAGAZINE_NOTEMPTY(loadedmag)) {
304 		obj = loadedmag->objects[--loadedmag->rounds];
305 		crit_exit();
306 		return (obj);
307 	}
308 
309 	/* Previous magazine has an object. */
310 	if (MAGAZINE_NOTEMPTY(cpucache->previous_magazine)) {
311 		KKASSERT(cpucache->previous_magazine->cleaning +
312 			 cpucache->loaded_magazine->cleaning == 0);
313 		swap(cpucache->loaded_magazine, cpucache->previous_magazine);
314 		loadedmag = cpucache->loaded_magazine;
315 		obj = loadedmag->objects[--loadedmag->rounds];
316 		crit_exit();
317 		return (obj);
318 	}
319 
320 	/*
321 	 * Both magazines empty.  Get a full magazine from the depot and
322 	 * move one of the empty ones to the depot.
323 	 *
324 	 * Obtain the depot spinlock.
325 	 *
326 	 * NOTE: Beyond this point, M_* flags are handled via oc->alloc()
327 	 */
328 	depot = &oc->depot[myclusterid];
329 	spin_lock_wr(&depot->spin);
330 
331 	/*
332 	 * Recheck the cpucache after obtaining the depot spinlock.  This
333 	 * shouldn't be necessary now but don't take any chances.
334 	 */
335 	if (MAGAZINE_NOTEMPTY(cpucache->loaded_magazine) ||
336 	    MAGAZINE_NOTEMPTY(cpucache->previous_magazine)
337 	) {
338 		spin_unlock_wr(&depot->spin);
339 		goto retry;
340 	}
341 
342 	/* Check if depot has a full magazine. */
343 	if (!SLIST_EMPTY(&depot->fullmagazines)) {
344 		emptymag = cpucache->previous_magazine;
345 		cpucache->previous_magazine = cpucache->loaded_magazine;
346 		cpucache->loaded_magazine = SLIST_FIRST(&depot->fullmagazines);
347 		SLIST_REMOVE_HEAD(&depot->fullmagazines, nextmagazine);
348 
349 		/*
350 		 * Return emptymag to the depot.
351 		 */
352 		KKASSERT(MAGAZINE_EMPTY(emptymag));
353 		SLIST_INSERT_HEAD(&depot->emptymagazines,
354 				  emptymag, nextmagazine);
355 		spin_unlock_wr(&depot->spin);
356 		goto retry;
357 	}
358 
359 	/*
360 	 * The depot does not have any non-empty magazines.  If we have
361 	 * not hit our object limit we can allocate a new object using
362 	 * the back-end allocator.
363 	 *
364 	 * note: unallocated_objects can be initialized to -1, which has
365 	 * the effect of removing any allocation limits.
366 	 */
367 	if (depot->unallocated_objects) {
368 		--depot->unallocated_objects;
369 		spin_unlock_wr(&depot->spin);
370 		crit_exit();
371 
372 		obj = oc->alloc(oc->allocator_args, ocflags);
373 		if (obj) {
374 			if (oc->ctor(obj, oc->private, ocflags))
375 				return (obj);
376 			oc->free(obj, oc->allocator_args);
377 			spin_lock_wr(&depot->spin);
378 			++depot->unallocated_objects;
379 			spin_unlock_wr(&depot->spin);
380 			if (depot->waiting)
381 				wakeup(depot);
382 			obj = NULL;
383 		}
384 		if (obj == NULL) {
385 			crit_enter();
386 			/*
387 			 * makes debugging easier when gets_cumulative does
388 			 * not include gets_null.
389 			 */
390 			++cpucache->gets_null;
391 			--cpucache->gets_cumulative;
392 			crit_exit();
393 		}
394 		return(obj);
395 	}
396 
397 	/*
398 	 * Otherwise block if allowed to.
399 	 */
400 	if ((ocflags & (M_WAITOK|M_NULLOK)) == M_WAITOK) {
401 		++cpucache->waiting;
402 		++depot->waiting;
403 		msleep(depot, &depot->spin, 0, "objcache_get", 0);
404 		--cpucache->waiting;
405 		--depot->waiting;
406 		spin_unlock_wr(&depot->spin);
407 		goto retry;
408 	}
409 
410 	/*
411 	 * Otherwise fail
412 	 */
413 	++cpucache->gets_null;
414 	--cpucache->gets_cumulative;
415 	crit_exit();
416 	spin_unlock_wr(&depot->spin);
417 	return (NULL);
418 }
419 
420 /*
421  * Wrapper for malloc allocation routines.
422  */
423 void *
424 objcache_malloc_alloc(void *allocator_args, int ocflags)
425 {
426 	struct objcache_malloc_args *alloc_args = allocator_args;
427 
428 	return (kmalloc(alloc_args->objsize, alloc_args->mtype,
429 		       ocflags & OC_MFLAGS));
430 }
431 
432 void
433 objcache_malloc_free(void *obj, void *allocator_args)
434 {
435 	struct objcache_malloc_args *alloc_args = allocator_args;
436 
437 	kfree(obj, alloc_args->mtype);
438 }
439 
440 /*
441  * Wrapper for allocation policies that pre-allocate at initialization time
442  * and don't do run-time allocation.
443  */
444 void *
445 objcache_nop_alloc(void *allocator_args, int ocflags)
446 {
447 	return (NULL);
448 }
449 
450 void
451 objcache_nop_free(void *obj, void *allocator_args)
452 {
453 }
454 
455 /*
456  * Return an object to the object cache.
457  */
458 void
459 objcache_put(struct objcache *oc, void *obj)
460 {
461 	struct percpu_objcache *cpucache = &oc->cache_percpu[mycpuid];
462 	struct magazine *loadedmag;
463 	struct magazinedepot *depot;
464 
465 	crit_enter();
466 	++cpucache->puts_cumulative;
467 
468 	if (CLUSTER_OF(obj) != myclusterid) {
469 #ifdef notyet
470 		/* use lazy IPI to send object to owning cluster XXX todo */
471 		++cpucache->puts_othercluster;
472 		crit_exit();
473 		return;
474 #endif
475 	}
476 
477 retry:
478 	/*
479 	 * Free slot available in loaded magazine.  This is the hot path.
480 	 * It is lock-free and uses a critical section to block out interrupt
481 	 * handlers on the same processor.
482 	 */
483 	loadedmag = cpucache->loaded_magazine;
484 	if (!MAGAZINE_FULL(loadedmag)) {
485 		loadedmag->objects[loadedmag->rounds++] = obj;
486 		if (cpucache->waiting)
487 			wakeup_mycpu(&oc->depot[myclusterid]);
488 		crit_exit();
489 		return;
490 	}
491 
492 	/*
493 	 * Current magazine full, but previous magazine has room.  XXX
494 	 */
495 	if (!MAGAZINE_FULL(cpucache->previous_magazine)) {
496 		KKASSERT(cpucache->previous_magazine->cleaning +
497 			 cpucache->loaded_magazine->cleaning == 0);
498 		swap(cpucache->loaded_magazine, cpucache->previous_magazine);
499 		loadedmag = cpucache->loaded_magazine;
500 		loadedmag->objects[loadedmag->rounds++] = obj;
501 		if (cpucache->waiting)
502 			wakeup_mycpu(&oc->depot[myclusterid]);
503 		crit_exit();
504 		return;
505 	}
506 
507 	/*
508 	 * Both magazines full.  Get an empty magazine from the depot and
509 	 * move a full loaded magazine to the depot.  Even though the
510 	 * magazine may wind up with space available after we block on
511 	 * the spinlock, we still cycle it through to avoid the non-optimal
512 	 * corner-case.
513 	 *
514 	 * Obtain the depot spinlock.
515 	 */
516 	depot = &oc->depot[myclusterid];
517 	spin_lock_wr(&depot->spin);
518 
519 	/*
520 	 * If an empty magazine is available in the depot, cycle it
521 	 * through and retry.
522 	 */
523 	if (!SLIST_EMPTY(&depot->emptymagazines)) {
524 		KKASSERT(cpucache->previous_magazine->cleaning +
525 			 cpucache->loaded_magazine->cleaning == 0);
526 		loadedmag = cpucache->previous_magazine;
527 		cpucache->previous_magazine = cpucache->loaded_magazine;
528 		cpucache->loaded_magazine = SLIST_FIRST(&depot->emptymagazines);
529 		SLIST_REMOVE_HEAD(&depot->emptymagazines, nextmagazine);
530 
531 		/*
532 		 * Return loadedmag to the depot.  Due to blocking it may
533 		 * not be entirely full and could even be empty.
534 		 */
535 		if (MAGAZINE_EMPTY(loadedmag)) {
536 			SLIST_INSERT_HEAD(&depot->emptymagazines,
537 					  loadedmag, nextmagazine);
538 			spin_unlock_wr(&depot->spin);
539 		} else {
540 			SLIST_INSERT_HEAD(&depot->fullmagazines,
541 					  loadedmag, nextmagazine);
542 			spin_unlock_wr(&depot->spin);
543 			if (depot->waiting)
544 				wakeup(depot);
545 		}
546 		goto retry;
547 	}
548 
549 	/*
550 	 * An empty mag is not available.  This is a corner case which can
551 	 * occur due to cpus holding partially full magazines.  Do not try
552 	 * to allocate a mag, just free the object.
553 	 */
554 	++depot->unallocated_objects;
555 	spin_unlock_wr(&depot->spin);
556 	if (depot->waiting)
557 		wakeup(depot);
558 	crit_exit();
559 	oc->dtor(obj, oc->private);
560 	oc->free(obj, oc->allocator_args);
561 }
562 
563 /*
564  * The object is being put back into the cache, but the caller has
565  * indicated that the object is not in any shape to be reused and should
566  * be dtor'd immediately.
567  */
568 void
569 objcache_dtor(struct objcache *oc, void *obj)
570 {
571 	struct magazinedepot *depot;
572 
573 	depot = &oc->depot[myclusterid];
574 	spin_lock_wr(&depot->spin);
575 	++depot->unallocated_objects;
576 	spin_unlock_wr(&depot->spin);
577 	if (depot->waiting)
578 		wakeup(depot);
579 	oc->dtor(obj, oc->private);
580 	oc->free(obj, oc->allocator_args);
581 }
582 
583 /*
584  * Deallocate all objects in a magazine and free the magazine if requested.
585  * The magazine must already be disassociated from the depot.
586  *
587  * Must be called with a critical section held when called with a per-cpu
588  * magazine.  The magazine may be indirectly modified during the loop.
589  *
590  * The number of objects freed is returned.
591  */
592 static int
593 mag_purge(struct objcache *oc, struct magazine *mag, int freeit)
594 {
595 	int count;
596 	void *obj;
597 
598 	count = 0;
599 	++mag->cleaning;
600 	while (mag->rounds) {
601 		obj = mag->objects[--mag->rounds];
602 		oc->dtor(obj, oc->private);		/* MAY BLOCK */
603 		oc->free(obj, oc->allocator_args);	/* MAY BLOCK */
604 		++count;
605 
606 		/*
607 		 * Cycle for interrupts
608 		 */
609 		if ((count & 15) == 0) {
610 			crit_exit();
611 			crit_enter();
612 		}
613 	}
614 	--mag->cleaning;
615 	if (freeit)
616 		kfree(mag, M_OBJMAG);
617 	return(count);
618 }
619 
620 /*
621  * Disassociate zero or more magazines from a magazine list associated with
622  * the depot, update the depot, and move the magazines to a temporary
623  * list.
624  *
625  * The caller must check the depot for waiters and wake it up, typically
626  * after disposing of the magazines this function loads onto the temporary
627  * list.
628  */
629 static void
630 maglist_disassociate(struct magazinedepot *depot, struct magazinelist *maglist,
631 		     struct magazinelist *tmplist, boolean_t purgeall)
632 {
633 	struct magazine *mag;
634 
635 	while ((mag = SLIST_FIRST(maglist)) != NULL) {
636 		SLIST_REMOVE_HEAD(maglist, nextmagazine);
637 		SLIST_INSERT_HEAD(tmplist, mag, nextmagazine);
638 		depot->unallocated_objects += mag->rounds;
639 	}
640 }
641 
642 /*
643  * Deallocate all magazines and their contents from the passed temporary
644  * list.  The magazines have already been accounted for by their depots.
645  *
646  * The total number of rounds freed is returned.  This number is typically
647  * only used to determine whether a wakeup on the depot is needed or not.
648  */
649 static int
650 maglist_purge(struct objcache *oc, struct magazinelist *maglist)
651 {
652 	struct magazine *mag;
653 	int count = 0;
654 
655 	/*
656 	 * can't use SLIST_FOREACH because blocking releases the depot
657 	 * spinlock
658 	 */
659 	while ((mag = SLIST_FIRST(maglist)) != NULL) {
660 		SLIST_REMOVE_HEAD(maglist, nextmagazine);
661 		count += mag_purge(oc, mag, TRUE);
662 	}
663 	return(count);
664 }
665 
666 /*
667  * De-allocates all magazines on the full and empty magazine lists.
668  *
669  * Because this routine is called with a spinlock held, the magazines
670  * can only be disassociated and moved to a temporary list, not freed.
671  *
672  * The caller is responsible for freeing the magazines.
673  */
674 static void
675 depot_disassociate(struct magazinedepot *depot, struct magazinelist *tmplist)
676 {
677 	maglist_disassociate(depot, &depot->fullmagazines, tmplist, TRUE);
678 	maglist_disassociate(depot, &depot->emptymagazines, tmplist, TRUE);
679 }
680 
681 #ifdef notneeded
682 void
683 objcache_reclaim(struct objcache *oc)
684 {
685 	struct percpu_objcache *cache_percpu = &oc->cache_percpu[myclusterid];
686 	struct magazinedepot *depot = &oc->depot[myclusterid];
687 	struct magazinelist tmplist;
688 	int count;
689 
690 	SLIST_INIT(&tmplist);
691 	crit_enter();
692 	count = mag_purge(oc, cache_percpu->loaded_magazine, FALSE);
693 	count += mag_purge(oc, cache_percpu->previous_magazine, FALSE);
694 	crit_exit();
695 
696 	spin_lock_wr(&depot->spin);
697 	depot->unallocated_objects += count;
698 	depot_disassociate(depot, &tmplist);
699 	spin_unlock_wr(&depot->spin);
700 	count += maglist_purge(oc, &tmplist);
701 	if (count && depot->waiting)
702 		wakeup(depot);
703 }
704 #endif
705 
706 /*
707  * Try to free up some memory.  Return as soon as some free memory is found.
708  * For each object cache on the reclaim list, first try the current per-cpu
709  * cache, then the full magazine depot.
710  */
711 boolean_t
712 objcache_reclaimlist(struct objcache *oclist[], int nlist, int ocflags)
713 {
714 	struct objcache *oc;
715 	struct percpu_objcache *cpucache;
716 	struct magazinedepot *depot;
717 	struct magazinelist tmplist;
718 	int i, count;
719 
720 	SLIST_INIT(&tmplist);
721 
722 	for (i = 0; i < nlist; i++) {
723 		oc = oclist[i];
724 		cpucache = &oc->cache_percpu[mycpuid];
725 		depot = &oc->depot[myclusterid];
726 
727 		crit_enter();
728 		count = mag_purge(oc, cpucache->loaded_magazine, FALSE);
729 		if (count == 0)
730 			count += mag_purge(oc, cpucache->previous_magazine, FALSE);
731 		crit_exit();
732 		if (count > 0) {
733 			spin_lock_wr(&depot->spin);
734 			depot->unallocated_objects += count;
735 			spin_unlock_wr(&depot->spin);
736 			if (depot->waiting)
737 				wakeup(depot);
738 			return (TRUE);
739 		}
740 		spin_lock_wr(&depot->spin);
741 		maglist_disassociate(depot, &depot->fullmagazines,
742 				     &tmplist, FALSE);
743 		spin_unlock_wr(&depot->spin);
744 		count = maglist_purge(oc, &tmplist);
745 		if (count > 0) {
746 			if (depot->waiting)
747 				wakeup(depot);
748 			return (TRUE);
749 		}
750 	}
751 	return (FALSE);
752 }
753 
754 /*
755  * Destroy an object cache.  Must have no existing references.
756  */
757 void
758 objcache_destroy(struct objcache *oc)
759 {
760 	struct percpu_objcache *cache_percpu;
761 	struct magazinedepot *depot;
762 	int clusterid, cpuid;
763 	struct magazinelist tmplist;
764 
765 	SLIST_INIT(&tmplist);
766 	for (clusterid = 0; clusterid < MAXCLUSTERS; clusterid++) {
767 		depot = &oc->depot[clusterid];
768 		spin_lock_wr(&depot->spin);
769 		depot_disassociate(depot, &tmplist);
770 		spin_unlock_wr(&depot->spin);
771 	}
772 	maglist_purge(oc, &tmplist);
773 
774 	for (cpuid = 0; cpuid < ncpus; cpuid++) {
775 		cache_percpu = &oc->cache_percpu[cpuid];
776 
777 		mag_purge(oc, cache_percpu->loaded_magazine, TRUE);
778 		mag_purge(oc, cache_percpu->previous_magazine, TRUE);
779 		cache_percpu->loaded_magazine = NULL;
780 		cache_percpu->previous_magazine = NULL;
781 		/* don't bother adjusting depot->unallocated_objects */
782 	}
783 
784 	kfree(oc->name, M_TEMP);
785 	kfree(oc, M_OBJCACHE);
786 }
787 
788 #if 0
789 /*
790  * Populate the per-cluster depot with elements from a linear block
791  * of memory.  Must be called for individually for each cluster.
792  * Populated depots should not be destroyed.
793  */
794 void
795 objcache_populate_linear(struct objcache *oc, void *base, int nelts, int size)
796 {
797 	char *p = base;
798 	char *end = (char *)base + (nelts * size);
799 	struct magazinedepot *depot = &oc->depot[myclusterid];
800 	struct magazine *emptymag = mag_alloc(depot->magcapcity);
801 
802 	while (p < end) {
803 		emptymag->objects[emptymag->rounds++] = p;
804 		if (MAGAZINE_FULL(emptymag)) {
805 			spin_lock_wr(&depot->spin);
806 			SLIST_INSERT_HEAD(&depot->fullmagazines, emptymag,
807 					  nextmagazine);
808 			depot->unallocated_objects += emptymag->rounds;
809 			spin_unlock_wr(&depot->spin);
810 			if (depot->waiting)
811 				wakeup(depot);
812 			emptymag = mag_alloc(depot->magcapacity);
813 		}
814 		p += size;
815 	}
816 	if (MAGAZINE_EMPTY(emptymag)) {
817 		mag_purge(oc, emptymag, TRUE);
818 	} else {
819 		spin_lock_wr(&depot->spin);
820 		SLIST_INSERT_HEAD(&depot->fullmagazines, emptymag,
821 				  nextmagazine);
822 		depot->unallocated_objects += emptymag->rounds;
823 		spin_unlock_wr(&depot->spin);
824 		if (depot->waiting)
825 			wakeup(depot);
826 		emptymag = mag_alloc(depot->magcapacity);
827 	}
828 }
829 #endif
830 
831 #if 0
832 /*
833  * Check depot contention once a minute.
834  * 2 contested locks per second allowed.
835  */
836 static int objcache_rebalance_period;
837 static const int objcache_contention_rate = 120;
838 static struct callout objcache_callout;
839 
840 #define MAXMAGSIZE 512
841 
842 /*
843  * Check depot contention and increase magazine size if necessary.
844  */
845 static void
846 objcache_timer(void *dummy)
847 {
848 	struct objcache *oc;
849 	struct magazinedepot *depot;
850 	struct magazinelist tmplist;
851 
852 	XXX we need to detect when an objcache is destroyed out from under
853 	    us XXX
854 
855 	SLIST_INIT(&tmplist);
856 
857 	spin_lock_wr(&objcachelist_spin);
858 	SLIST_FOREACH(oc, &allobjcaches, oc_next) {
859 		depot = &oc->depot[myclusterid];
860 		if (depot->magcapacity < MAXMAGSIZE) {
861 			if (depot->contested > objcache_contention_rate) {
862 				spin_lock_wr(&depot->spin);
863 				depot_disassociate(depot, &tmplist);
864 				depot->magcapacity *= 2;
865 				spin_unlock_wr(&depot->spin);
866 				kprintf("objcache_timer: increasing cache %s"
867 				       " magsize to %d, contested %d times\n",
868 				    oc->name, depot->magcapacity,
869 				    depot->contested);
870 			}
871 			depot->contested = 0;
872 		}
873 		spin_unlock_wr(&objcachelist_spin);
874 		if (maglist_purge(oc, &tmplist) > 0 && depot->waiting)
875 			wakeup(depot);
876 		spin_lock_wr(&objcachelist_spin);
877 	}
878 	spin_unlock_wr(&objcachelist_spin);
879 
880 	callout_reset(&objcache_callout, objcache_rebalance_period,
881 		      objcache_timer, NULL);
882 }
883 
884 #endif
885 
886 static void
887 objcache_init(void)
888 {
889 	spin_init(&objcachelist_spin);
890 #if 0
891 	callout_init(&objcache_callout);
892 	objcache_rebalance_period = 60 * hz;
893 	callout_reset(&objcache_callout, objcache_rebalance_period,
894 		      objcache_timer, NULL);
895 #endif
896 }
897 SYSINIT(objcache, SI_BOOT2_OBJCACHE, SI_ORDER_FIRST, objcache_init, 0);
898