xref: /dragonfly/sys/kern/kern_objcache.c (revision b7367ef6)
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 1.20 2007/07/02 06:34:26 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	64
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			*privdata;
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 	int			exhausted;	/* oops */
135 
136 	/* NUMA-cluster level caches */
137 	struct magazinedepot	depot[MAXCLUSTERS];
138 
139 	struct percpu_objcache	cache_percpu[];		/* per-cpu caches */
140 };
141 
142 static struct spinlock objcachelist_spin;
143 static SLIST_HEAD(objcachelist, objcache) allobjcaches;
144 
145 static struct magazine *
146 mag_alloc(int capacity)
147 {
148 	struct magazine *mag;
149 
150 	mag = kmalloc(__offsetof(struct magazine, objects[capacity]),
151 			M_OBJMAG, M_INTWAIT | M_ZERO);
152 	mag->capacity = capacity;
153 	mag->rounds = 0;
154 	mag->cleaning = 0;
155 	return (mag);
156 }
157 
158 /*
159  * Utility routine for objects that don't require any de-construction.
160  */
161 
162 static void
163 null_dtor(void *obj, void *privdata)
164 {
165 	/* do nothing */
166 }
167 
168 static boolean_t
169 null_ctor(void *obj, void *privdata, int ocflags)
170 {
171 	return TRUE;
172 }
173 
174 /*
175  * Create an object cache.
176  */
177 struct objcache *
178 objcache_create(const char *name, int cluster_limit, int mag_capacity,
179 		objcache_ctor_fn *ctor, objcache_dtor_fn *dtor, void *privdata,
180 		objcache_alloc_fn *alloc, objcache_free_fn *free,
181 		void *allocator_args)
182 {
183 	struct objcache *oc;
184 	struct magazinedepot *depot;
185 	int cpuid;
186 	int need;
187 	int factor;
188 
189 	/* allocate object cache structure */
190 	oc = kmalloc(__offsetof(struct objcache, cache_percpu[ncpus]),
191 		    M_OBJCACHE, M_WAITOK | M_ZERO);
192 	oc->name = kstrdup(name, M_TEMP);
193 	oc->ctor = ctor ? ctor : null_ctor;
194 	oc->dtor = dtor ? dtor : null_dtor;
195 	oc->privdata = privdata;
196 	oc->free = free;
197 	oc->allocator_args = allocator_args;
198 
199 	/* initialize depots */
200 	depot = &oc->depot[0];
201 
202 	spin_init(&depot->spin);
203 	SLIST_INIT(&depot->fullmagazines);
204 	SLIST_INIT(&depot->emptymagazines);
205 
206 	if (mag_capacity == 0)
207 		mag_capacity = INITIAL_MAG_CAPACITY;
208 
209 	/*
210 	 * The cluster_limit must be sufficient to have three magazines per
211 	 * cpu.  If we have a lot of cpus the mag_capacity might just be
212 	 * too big, reduce it if necessary.
213 	 *
214 	 * Each cpu can hold up to two magazines, with the remainder in the
215 	 * depot.  If many objects are allocated fewer magazines are
216 	 * available.  We have to make sure that each cpu has access to
217 	 * free objects until the object cache hits 75% of its limit.
218 	 */
219 	if (cluster_limit == 0) {
220 		depot->unallocated_objects = -1;
221 	} else {
222 		factor = 8;
223 		need = mag_capacity * ncpus * factor;
224 		if (cluster_limit < need && mag_capacity > 16) {
225 			kprintf("objcache(%s): too small for ncpus"
226 				", adjusting mag_capacity %d->",
227 				name, mag_capacity);
228 			while (need > cluster_limit && mag_capacity > 16) {
229 				mag_capacity >>= 1;
230 				need = mag_capacity * ncpus * factor;
231 			}
232 			kprintf("%d\n", mag_capacity);
233 		}
234 		if (cluster_limit < need) {
235 			kprintf("objcache(%s): too small for ncpus"
236 				", adjusting cluster_limit %d->%d\n",
237 				name, cluster_limit, need);
238 			cluster_limit = need;
239 		}
240 		depot->unallocated_objects = cluster_limit;
241 	}
242 	depot->magcapacity = mag_capacity;
243 	oc->alloc = alloc;
244 
245 	/* initialize per-cpu caches */
246 	for (cpuid = 0; cpuid < ncpus; cpuid++) {
247 		struct percpu_objcache *cache_percpu = &oc->cache_percpu[cpuid];
248 
249 		cache_percpu->loaded_magazine = mag_alloc(mag_capacity);
250 		cache_percpu->previous_magazine = mag_alloc(mag_capacity);
251 	}
252 	spin_lock_wr(&objcachelist_spin);
253 	SLIST_INSERT_HEAD(&allobjcaches, oc, oc_next);
254 	spin_unlock_wr(&objcachelist_spin);
255 
256 	return (oc);
257 }
258 
259 struct objcache *
260 objcache_create_simple(malloc_type_t mtype, size_t objsize)
261 {
262 	struct objcache_malloc_args *margs;
263 	struct objcache *oc;
264 
265 	margs = kmalloc(sizeof(*margs), M_OBJCACHE, M_WAITOK|M_ZERO);
266 	margs->objsize = objsize;
267 	margs->mtype = mtype;
268 	oc = objcache_create(mtype->ks_shortdesc, 0, 0,
269 			     NULL, NULL, NULL,
270 			     objcache_malloc_alloc, objcache_malloc_free,
271 			     margs);
272 	return (oc);
273 }
274 
275 struct objcache *
276 objcache_create_mbacked(malloc_type_t mtype, size_t objsize,
277 			int cluster_limit, int mag_capacity,
278 			objcache_ctor_fn *ctor, objcache_dtor_fn *dtor,
279 			void *privdata)
280 {
281 	struct objcache_malloc_args *margs;
282 	struct objcache *oc;
283 
284 	margs = kmalloc(sizeof(*margs), M_OBJCACHE, M_WAITOK|M_ZERO);
285 	margs->objsize = objsize;
286 	margs->mtype = mtype;
287 	oc = objcache_create(mtype->ks_shortdesc,
288 			     cluster_limit, mag_capacity,
289 			     ctor, dtor, privdata,
290 			     objcache_malloc_alloc, objcache_malloc_free,
291 			     margs);
292 	return(oc);
293 }
294 
295 
296 #define MAGAZINE_EMPTY(mag)	(mag->rounds == 0)
297 #define MAGAZINE_NOTEMPTY(mag)	(mag->rounds != 0)
298 #define MAGAZINE_FULL(mag)	(mag->rounds == mag->capacity)
299 
300 #define	swap(x, y)	({ struct magazine *t = x; x = y; y = t; })
301 
302 /*
303  * Get an object from the object cache.
304  *
305  * WARNING!  ocflags are only used when we have to go to the underlying
306  * allocator, so we cannot depend on flags such as M_ZERO.
307  */
308 void *
309 objcache_get(struct objcache *oc, int ocflags)
310 {
311 	struct percpu_objcache *cpucache = &oc->cache_percpu[mycpuid];
312 	struct magazine *loadedmag;
313 	struct magazine *emptymag;
314 	void *obj;
315 	struct magazinedepot *depot;
316 
317 	KKASSERT((ocflags & M_ZERO) == 0);
318 	crit_enter();
319 	++cpucache->gets_cumulative;
320 
321 retry:
322 	/*
323 	 * Loaded magazine has an object.  This is the hot path.
324 	 * It is lock-free and uses a critical section to block
325 	 * out interrupt handlers on the same processor.
326 	 */
327 	loadedmag = cpucache->loaded_magazine;
328 	if (MAGAZINE_NOTEMPTY(loadedmag)) {
329 		obj = loadedmag->objects[--loadedmag->rounds];
330 		crit_exit();
331 		return (obj);
332 	}
333 
334 	/* Previous magazine has an object. */
335 	if (MAGAZINE_NOTEMPTY(cpucache->previous_magazine)) {
336 		KKASSERT(cpucache->previous_magazine->cleaning +
337 			 cpucache->loaded_magazine->cleaning == 0);
338 		swap(cpucache->loaded_magazine, cpucache->previous_magazine);
339 		loadedmag = cpucache->loaded_magazine;
340 		obj = loadedmag->objects[--loadedmag->rounds];
341 		crit_exit();
342 		return (obj);
343 	}
344 
345 	/*
346 	 * Both magazines empty.  Get a full magazine from the depot and
347 	 * move one of the empty ones to the depot.
348 	 *
349 	 * Obtain the depot spinlock.
350 	 *
351 	 * NOTE: Beyond this point, M_* flags are handled via oc->alloc()
352 	 */
353 	depot = &oc->depot[myclusterid];
354 	spin_lock_wr(&depot->spin);
355 
356 	/*
357 	 * Recheck the cpucache after obtaining the depot spinlock.  This
358 	 * shouldn't be necessary now but don't take any chances.
359 	 */
360 	if (MAGAZINE_NOTEMPTY(cpucache->loaded_magazine) ||
361 	    MAGAZINE_NOTEMPTY(cpucache->previous_magazine)
362 	) {
363 		spin_unlock_wr(&depot->spin);
364 		goto retry;
365 	}
366 
367 	/* Check if depot has a full magazine. */
368 	if (!SLIST_EMPTY(&depot->fullmagazines)) {
369 		emptymag = cpucache->previous_magazine;
370 		cpucache->previous_magazine = cpucache->loaded_magazine;
371 		cpucache->loaded_magazine = SLIST_FIRST(&depot->fullmagazines);
372 		SLIST_REMOVE_HEAD(&depot->fullmagazines, nextmagazine);
373 
374 		/*
375 		 * Return emptymag to the depot.
376 		 */
377 		KKASSERT(MAGAZINE_EMPTY(emptymag));
378 		SLIST_INSERT_HEAD(&depot->emptymagazines,
379 				  emptymag, nextmagazine);
380 		spin_unlock_wr(&depot->spin);
381 		goto retry;
382 	}
383 
384 	/*
385 	 * The depot does not have any non-empty magazines.  If we have
386 	 * not hit our object limit we can allocate a new object using
387 	 * the back-end allocator.
388 	 *
389 	 * note: unallocated_objects can be initialized to -1, which has
390 	 * the effect of removing any allocation limits.
391 	 */
392 	if (depot->unallocated_objects) {
393 		--depot->unallocated_objects;
394 		spin_unlock_wr(&depot->spin);
395 		crit_exit();
396 
397 		obj = oc->alloc(oc->allocator_args, ocflags);
398 		if (obj) {
399 			if (oc->ctor(obj, oc->privdata, ocflags))
400 				return (obj);
401 			oc->free(obj, oc->allocator_args);
402 			spin_lock_wr(&depot->spin);
403 			++depot->unallocated_objects;
404 			spin_unlock_wr(&depot->spin);
405 			if (depot->waiting)
406 				wakeup(depot);
407 			obj = NULL;
408 		}
409 		if (obj == NULL) {
410 			crit_enter();
411 			/*
412 			 * makes debugging easier when gets_cumulative does
413 			 * not include gets_null.
414 			 */
415 			++cpucache->gets_null;
416 			--cpucache->gets_cumulative;
417 			crit_exit();
418 		}
419 		return(obj);
420 	}
421 	if (oc->exhausted == 0) {
422 		kprintf("Warning, objcache(%s): Exhausted!\n", oc->name);
423 		oc->exhausted = 1;
424 	}
425 
426 	/*
427 	 * Otherwise block if allowed to.
428 	 */
429 	if ((ocflags & (M_WAITOK|M_NULLOK)) == M_WAITOK) {
430 		++cpucache->waiting;
431 		++depot->waiting;
432 		msleep(depot, &depot->spin, 0, "objcache_get", 0);
433 		--cpucache->waiting;
434 		--depot->waiting;
435 		spin_unlock_wr(&depot->spin);
436 		goto retry;
437 	}
438 
439 	/*
440 	 * Otherwise fail
441 	 */
442 	++cpucache->gets_null;
443 	--cpucache->gets_cumulative;
444 	crit_exit();
445 	spin_unlock_wr(&depot->spin);
446 	return (NULL);
447 }
448 
449 /*
450  * Wrapper for malloc allocation routines.
451  */
452 void *
453 objcache_malloc_alloc(void *allocator_args, int ocflags)
454 {
455 	struct objcache_malloc_args *alloc_args = allocator_args;
456 
457 	return (kmalloc(alloc_args->objsize, alloc_args->mtype,
458 		       ocflags & OC_MFLAGS));
459 }
460 
461 void
462 objcache_malloc_free(void *obj, void *allocator_args)
463 {
464 	struct objcache_malloc_args *alloc_args = allocator_args;
465 
466 	kfree(obj, alloc_args->mtype);
467 }
468 
469 /*
470  * Wrapper for allocation policies that pre-allocate at initialization time
471  * and don't do run-time allocation.
472  */
473 void *
474 objcache_nop_alloc(void *allocator_args, int ocflags)
475 {
476 	return (NULL);
477 }
478 
479 void
480 objcache_nop_free(void *obj, void *allocator_args)
481 {
482 }
483 
484 /*
485  * Return an object to the object cache.
486  */
487 void
488 objcache_put(struct objcache *oc, void *obj)
489 {
490 	struct percpu_objcache *cpucache = &oc->cache_percpu[mycpuid];
491 	struct magazine *loadedmag;
492 	struct magazinedepot *depot;
493 
494 	crit_enter();
495 	++cpucache->puts_cumulative;
496 
497 	if (CLUSTER_OF(obj) != myclusterid) {
498 #ifdef notyet
499 		/* use lazy IPI to send object to owning cluster XXX todo */
500 		++cpucache->puts_othercluster;
501 		crit_exit();
502 		return;
503 #endif
504 	}
505 
506 retry:
507 	/*
508 	 * Free slot available in loaded magazine.  This is the hot path.
509 	 * It is lock-free and uses a critical section to block out interrupt
510 	 * handlers on the same processor.
511 	 */
512 	loadedmag = cpucache->loaded_magazine;
513 	if (!MAGAZINE_FULL(loadedmag)) {
514 		loadedmag->objects[loadedmag->rounds++] = obj;
515 		if (cpucache->waiting)
516 			wakeup_mycpu(&oc->depot[myclusterid]);
517 		crit_exit();
518 		return;
519 	}
520 
521 	/*
522 	 * Current magazine full, but previous magazine has room.  XXX
523 	 */
524 	if (!MAGAZINE_FULL(cpucache->previous_magazine)) {
525 		KKASSERT(cpucache->previous_magazine->cleaning +
526 			 cpucache->loaded_magazine->cleaning == 0);
527 		swap(cpucache->loaded_magazine, cpucache->previous_magazine);
528 		loadedmag = cpucache->loaded_magazine;
529 		loadedmag->objects[loadedmag->rounds++] = obj;
530 		if (cpucache->waiting)
531 			wakeup_mycpu(&oc->depot[myclusterid]);
532 		crit_exit();
533 		return;
534 	}
535 
536 	/*
537 	 * Both magazines full.  Get an empty magazine from the depot and
538 	 * move a full loaded magazine to the depot.  Even though the
539 	 * magazine may wind up with space available after we block on
540 	 * the spinlock, we still cycle it through to avoid the non-optimal
541 	 * corner-case.
542 	 *
543 	 * Obtain the depot spinlock.
544 	 */
545 	depot = &oc->depot[myclusterid];
546 	spin_lock_wr(&depot->spin);
547 
548 	/*
549 	 * If an empty magazine is available in the depot, cycle it
550 	 * through and retry.
551 	 */
552 	if (!SLIST_EMPTY(&depot->emptymagazines)) {
553 		KKASSERT(cpucache->previous_magazine->cleaning +
554 			 cpucache->loaded_magazine->cleaning == 0);
555 		loadedmag = cpucache->previous_magazine;
556 		cpucache->previous_magazine = cpucache->loaded_magazine;
557 		cpucache->loaded_magazine = SLIST_FIRST(&depot->emptymagazines);
558 		SLIST_REMOVE_HEAD(&depot->emptymagazines, nextmagazine);
559 
560 		/*
561 		 * Return loadedmag to the depot.  Due to blocking it may
562 		 * not be entirely full and could even be empty.
563 		 */
564 		if (MAGAZINE_EMPTY(loadedmag)) {
565 			SLIST_INSERT_HEAD(&depot->emptymagazines,
566 					  loadedmag, nextmagazine);
567 			spin_unlock_wr(&depot->spin);
568 		} else {
569 			SLIST_INSERT_HEAD(&depot->fullmagazines,
570 					  loadedmag, nextmagazine);
571 			spin_unlock_wr(&depot->spin);
572 			if (depot->waiting)
573 				wakeup(depot);
574 		}
575 		goto retry;
576 	}
577 
578 	/*
579 	 * An empty mag is not available.  This is a corner case which can
580 	 * occur due to cpus holding partially full magazines.  Do not try
581 	 * to allocate a mag, just free the object.
582 	 */
583 	++depot->unallocated_objects;
584 	spin_unlock_wr(&depot->spin);
585 	if (depot->waiting)
586 		wakeup(depot);
587 	crit_exit();
588 	oc->dtor(obj, oc->privdata);
589 	oc->free(obj, oc->allocator_args);
590 }
591 
592 /*
593  * The object is being put back into the cache, but the caller has
594  * indicated that the object is not in any shape to be reused and should
595  * be dtor'd immediately.
596  */
597 void
598 objcache_dtor(struct objcache *oc, void *obj)
599 {
600 	struct magazinedepot *depot;
601 
602 	depot = &oc->depot[myclusterid];
603 	spin_lock_wr(&depot->spin);
604 	++depot->unallocated_objects;
605 	spin_unlock_wr(&depot->spin);
606 	if (depot->waiting)
607 		wakeup(depot);
608 	oc->dtor(obj, oc->privdata);
609 	oc->free(obj, oc->allocator_args);
610 }
611 
612 /*
613  * Deallocate all objects in a magazine and free the magazine if requested.
614  * The magazine must already be disassociated from the depot.
615  *
616  * Must be called with a critical section held when called with a per-cpu
617  * magazine.  The magazine may be indirectly modified during the loop.
618  *
619  * The number of objects freed is returned.
620  */
621 static int
622 mag_purge(struct objcache *oc, struct magazine *mag, int freeit)
623 {
624 	int count;
625 	void *obj;
626 
627 	count = 0;
628 	++mag->cleaning;
629 	while (mag->rounds) {
630 		obj = mag->objects[--mag->rounds];
631 		oc->dtor(obj, oc->privdata);		/* MAY BLOCK */
632 		oc->free(obj, oc->allocator_args);	/* MAY BLOCK */
633 		++count;
634 
635 		/*
636 		 * Cycle for interrupts
637 		 */
638 		if ((count & 15) == 0) {
639 			crit_exit();
640 			crit_enter();
641 		}
642 	}
643 	--mag->cleaning;
644 	if (freeit)
645 		kfree(mag, M_OBJMAG);
646 	return(count);
647 }
648 
649 /*
650  * Disassociate zero or more magazines from a magazine list associated with
651  * the depot, update the depot, and move the magazines to a temporary
652  * list.
653  *
654  * The caller must check the depot for waiters and wake it up, typically
655  * after disposing of the magazines this function loads onto the temporary
656  * list.
657  */
658 static void
659 maglist_disassociate(struct magazinedepot *depot, struct magazinelist *maglist,
660 		     struct magazinelist *tmplist, boolean_t purgeall)
661 {
662 	struct magazine *mag;
663 
664 	while ((mag = SLIST_FIRST(maglist)) != NULL) {
665 		SLIST_REMOVE_HEAD(maglist, nextmagazine);
666 		SLIST_INSERT_HEAD(tmplist, mag, nextmagazine);
667 		depot->unallocated_objects += mag->rounds;
668 	}
669 }
670 
671 /*
672  * Deallocate all magazines and their contents from the passed temporary
673  * list.  The magazines have already been accounted for by their depots.
674  *
675  * The total number of rounds freed is returned.  This number is typically
676  * only used to determine whether a wakeup on the depot is needed or not.
677  */
678 static int
679 maglist_purge(struct objcache *oc, struct magazinelist *maglist)
680 {
681 	struct magazine *mag;
682 	int count = 0;
683 
684 	/*
685 	 * can't use SLIST_FOREACH because blocking releases the depot
686 	 * spinlock
687 	 */
688 	while ((mag = SLIST_FIRST(maglist)) != NULL) {
689 		SLIST_REMOVE_HEAD(maglist, nextmagazine);
690 		count += mag_purge(oc, mag, TRUE);
691 	}
692 	return(count);
693 }
694 
695 /*
696  * De-allocates all magazines on the full and empty magazine lists.
697  *
698  * Because this routine is called with a spinlock held, the magazines
699  * can only be disassociated and moved to a temporary list, not freed.
700  *
701  * The caller is responsible for freeing the magazines.
702  */
703 static void
704 depot_disassociate(struct magazinedepot *depot, struct magazinelist *tmplist)
705 {
706 	maglist_disassociate(depot, &depot->fullmagazines, tmplist, TRUE);
707 	maglist_disassociate(depot, &depot->emptymagazines, tmplist, TRUE);
708 }
709 
710 #ifdef notneeded
711 void
712 objcache_reclaim(struct objcache *oc)
713 {
714 	struct percpu_objcache *cache_percpu = &oc->cache_percpu[myclusterid];
715 	struct magazinedepot *depot = &oc->depot[myclusterid];
716 	struct magazinelist tmplist;
717 	int count;
718 
719 	SLIST_INIT(&tmplist);
720 	crit_enter();
721 	count = mag_purge(oc, cache_percpu->loaded_magazine, FALSE);
722 	count += mag_purge(oc, cache_percpu->previous_magazine, FALSE);
723 	crit_exit();
724 
725 	spin_lock_wr(&depot->spin);
726 	depot->unallocated_objects += count;
727 	depot_disassociate(depot, &tmplist);
728 	spin_unlock_wr(&depot->spin);
729 	count += maglist_purge(oc, &tmplist);
730 	if (count && depot->waiting)
731 		wakeup(depot);
732 }
733 #endif
734 
735 /*
736  * Try to free up some memory.  Return as soon as some free memory is found.
737  * For each object cache on the reclaim list, first try the current per-cpu
738  * cache, then the full magazine depot.
739  */
740 boolean_t
741 objcache_reclaimlist(struct objcache *oclist[], int nlist, int ocflags)
742 {
743 	struct objcache *oc;
744 	struct percpu_objcache *cpucache;
745 	struct magazinedepot *depot;
746 	struct magazinelist tmplist;
747 	int i, count;
748 
749 	SLIST_INIT(&tmplist);
750 
751 	for (i = 0; i < nlist; i++) {
752 		oc = oclist[i];
753 		cpucache = &oc->cache_percpu[mycpuid];
754 		depot = &oc->depot[myclusterid];
755 
756 		crit_enter();
757 		count = mag_purge(oc, cpucache->loaded_magazine, FALSE);
758 		if (count == 0)
759 			count += mag_purge(oc, cpucache->previous_magazine, FALSE);
760 		crit_exit();
761 		if (count > 0) {
762 			spin_lock_wr(&depot->spin);
763 			depot->unallocated_objects += count;
764 			spin_unlock_wr(&depot->spin);
765 			if (depot->waiting)
766 				wakeup(depot);
767 			return (TRUE);
768 		}
769 		spin_lock_wr(&depot->spin);
770 		maglist_disassociate(depot, &depot->fullmagazines,
771 				     &tmplist, FALSE);
772 		spin_unlock_wr(&depot->spin);
773 		count = maglist_purge(oc, &tmplist);
774 		if (count > 0) {
775 			if (depot->waiting)
776 				wakeup(depot);
777 			return (TRUE);
778 		}
779 	}
780 	return (FALSE);
781 }
782 
783 /*
784  * Destroy an object cache.  Must have no existing references.
785  */
786 void
787 objcache_destroy(struct objcache *oc)
788 {
789 	struct percpu_objcache *cache_percpu;
790 	struct magazinedepot *depot;
791 	int clusterid, cpuid;
792 	struct magazinelist tmplist;
793 
794 	SLIST_INIT(&tmplist);
795 	for (clusterid = 0; clusterid < MAXCLUSTERS; clusterid++) {
796 		depot = &oc->depot[clusterid];
797 		spin_lock_wr(&depot->spin);
798 		depot_disassociate(depot, &tmplist);
799 		spin_unlock_wr(&depot->spin);
800 	}
801 	maglist_purge(oc, &tmplist);
802 
803 	for (cpuid = 0; cpuid < ncpus; cpuid++) {
804 		cache_percpu = &oc->cache_percpu[cpuid];
805 
806 		mag_purge(oc, cache_percpu->loaded_magazine, TRUE);
807 		mag_purge(oc, cache_percpu->previous_magazine, TRUE);
808 		cache_percpu->loaded_magazine = NULL;
809 		cache_percpu->previous_magazine = NULL;
810 		/* don't bother adjusting depot->unallocated_objects */
811 	}
812 
813 	kfree(oc->name, M_TEMP);
814 	kfree(oc, M_OBJCACHE);
815 }
816 
817 #if 0
818 /*
819  * Populate the per-cluster depot with elements from a linear block
820  * of memory.  Must be called for individually for each cluster.
821  * Populated depots should not be destroyed.
822  */
823 void
824 objcache_populate_linear(struct objcache *oc, void *base, int nelts, int size)
825 {
826 	char *p = base;
827 	char *end = (char *)base + (nelts * size);
828 	struct magazinedepot *depot = &oc->depot[myclusterid];
829 	struct magazine *emptymag = mag_alloc(depot->magcapcity);
830 
831 	while (p < end) {
832 		emptymag->objects[emptymag->rounds++] = p;
833 		if (MAGAZINE_FULL(emptymag)) {
834 			spin_lock_wr(&depot->spin);
835 			SLIST_INSERT_HEAD(&depot->fullmagazines, emptymag,
836 					  nextmagazine);
837 			depot->unallocated_objects += emptymag->rounds;
838 			spin_unlock_wr(&depot->spin);
839 			if (depot->waiting)
840 				wakeup(depot);
841 			emptymag = mag_alloc(depot->magcapacity);
842 		}
843 		p += size;
844 	}
845 	if (MAGAZINE_EMPTY(emptymag)) {
846 		mag_purge(oc, emptymag, TRUE);
847 	} else {
848 		spin_lock_wr(&depot->spin);
849 		SLIST_INSERT_HEAD(&depot->fullmagazines, emptymag,
850 				  nextmagazine);
851 		depot->unallocated_objects += emptymag->rounds;
852 		spin_unlock_wr(&depot->spin);
853 		if (depot->waiting)
854 			wakeup(depot);
855 		emptymag = mag_alloc(depot->magcapacity);
856 	}
857 }
858 #endif
859 
860 #if 0
861 /*
862  * Check depot contention once a minute.
863  * 2 contested locks per second allowed.
864  */
865 static int objcache_rebalance_period;
866 static const int objcache_contention_rate = 120;
867 static struct callout objcache_callout;
868 
869 #define MAXMAGSIZE 512
870 
871 /*
872  * Check depot contention and increase magazine size if necessary.
873  */
874 static void
875 objcache_timer(void *dummy)
876 {
877 	struct objcache *oc;
878 	struct magazinedepot *depot;
879 	struct magazinelist tmplist;
880 
881 	XXX we need to detect when an objcache is destroyed out from under
882 	    us XXX
883 
884 	SLIST_INIT(&tmplist);
885 
886 	spin_lock_wr(&objcachelist_spin);
887 	SLIST_FOREACH(oc, &allobjcaches, oc_next) {
888 		depot = &oc->depot[myclusterid];
889 		if (depot->magcapacity < MAXMAGSIZE) {
890 			if (depot->contested > objcache_contention_rate) {
891 				spin_lock_wr(&depot->spin);
892 				depot_disassociate(depot, &tmplist);
893 				depot->magcapacity *= 2;
894 				spin_unlock_wr(&depot->spin);
895 				kprintf("objcache_timer: increasing cache %s"
896 				       " magsize to %d, contested %d times\n",
897 				    oc->name, depot->magcapacity,
898 				    depot->contested);
899 			}
900 			depot->contested = 0;
901 		}
902 		spin_unlock_wr(&objcachelist_spin);
903 		if (maglist_purge(oc, &tmplist) > 0 && depot->waiting)
904 			wakeup(depot);
905 		spin_lock_wr(&objcachelist_spin);
906 	}
907 	spin_unlock_wr(&objcachelist_spin);
908 
909 	callout_reset(&objcache_callout, objcache_rebalance_period,
910 		      objcache_timer, NULL);
911 }
912 
913 #endif
914 
915 static void
916 objcache_init(void)
917 {
918 	spin_init(&objcachelist_spin);
919 #if 0
920 	callout_init(&objcache_callout);
921 	objcache_rebalance_period = 60 * hz;
922 	callout_reset(&objcache_callout, objcache_rebalance_period,
923 		      objcache_timer, NULL);
924 #endif
925 }
926 SYSINIT(objcache, SI_BOOT2_OBJCACHE, SI_ORDER_FIRST, objcache_init, 0);
927