xref: /openbsd-src/sys/kern/subr_pool.c (revision 8445c53715e7030056b779e8ab40efb7820981f2)
1 /*	$OpenBSD: subr_pool.c,v 1.13 2001/09/19 20:50:58 mickey Exp $	*/
2 /*	$NetBSD: subr_pool.c,v 1.59 2001/06/05 18:51:04 thorpej Exp $	*/
3 
4 /*-
5  * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc.
6  * All rights reserved.
7  *
8  * This code is derived from software contributed to The NetBSD Foundation
9  * by Paul Kranenburg; by Jason R. Thorpe of the Numerical Aerospace
10  * Simulation Facility, NASA Ames Research Center.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the NetBSD
23  *	Foundation, Inc. and its contributors.
24  * 4. Neither the name of The NetBSD Foundation nor the names of its
25  *    contributors may be used to endorse or promote products derived
26  *    from this software without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
29  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
30  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
31  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
32  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
33  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
34  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
35  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
36  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
37  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
38  * POSSIBILITY OF SUCH DAMAGE.
39  */
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/proc.h>
44 #include <sys/errno.h>
45 #include <sys/kernel.h>
46 #include <sys/malloc.h>
47 #include <sys/lock.h>
48 #include <sys/pool.h>
49 #include <sys/syslog.h>
50 #include <sys/sysctl.h>
51 
52 #include <vm/vm.h>
53 #include <uvm/uvm.h>
54 
55 /*
56  * XXX - for now.
57  */
58 #define SIMPLELOCK_INITIALIZER { SLOCK_UNLOCKED }
59 #ifdef LOCKDEBUG
60 #define simple_lock_freecheck(a, s) do { /* nothing */ } while (0)
61 #define simple_lock_only_held(lkp, str) do { /* nothing */ } while (0)
62 #endif
63 #define LOCK_ASSERT(x) /* nothing */
64 
65 /*
66  * Pool resource management utility.
67  *
68  * Memory is allocated in pages which are split into pieces according
69  * to the pool item size. Each page is kept on a list headed by `pr_pagelist'
70  * in the pool structure and the individual pool items are on a linked list
71  * headed by `ph_itemlist' in each page header. The memory for building
72  * the page list is either taken from the allocated pages themselves (for
73  * small pool items) or taken from an internal pool of page headers (`phpool').
74  */
75 
76 /* List of all pools */
77 TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head);
78 
79 /* Private pool for page header structures */
80 static struct pool phpool;
81 
82 /* # of seconds to retain page after last use */
83 int pool_inactive_time = 10;
84 
85 /* Next candidate for drainage (see pool_drain()) */
86 static struct pool	*drainpp;
87 
88 /* This spin lock protects both pool_head and drainpp. */
89 struct simplelock pool_head_slock = SIMPLELOCK_INITIALIZER;
90 
91 struct pool_item_header {
92 	/* Page headers */
93 	TAILQ_ENTRY(pool_item_header)
94 				ph_pagelist;	/* pool page list */
95 	TAILQ_HEAD(,pool_item)	ph_itemlist;	/* chunk list for this page */
96 	LIST_ENTRY(pool_item_header)
97 				ph_hashlist;	/* Off-page page headers */
98 	int			ph_nmissing;	/* # of chunks in use */
99 	caddr_t			ph_page;	/* this page's address */
100 	struct timeval		ph_time;	/* last referenced */
101 };
102 
103 struct pool_item {
104 #ifdef DIAGNOSTIC
105 	int pi_magic;
106 #endif
107 #define	PI_MAGIC 0xdeadbeef
108 	/* Other entries use only this list entry */
109 	TAILQ_ENTRY(pool_item)	pi_list;
110 };
111 
112 
113 #define	PR_HASH_INDEX(pp,addr) \
114 	(((u_long)(addr) >> (pp)->pr_pageshift) & (PR_HASHTABSIZE - 1))
115 
116 #define	POOL_NEEDS_CATCHUP(pp)						\
117 	((pp)->pr_nitems < (pp)->pr_minitems)
118 
119 /*
120  * Every pool get a unique serial number assigned to it. If this counter
121  * wraps, we're screwed, but we shouldn't create so many pools anyway.
122  */
123 unsigned int pool_serial;
124 
125 /*
126  * Pool cache management.
127  *
128  * Pool caches provide a way for constructed objects to be cached by the
129  * pool subsystem.  This can lead to performance improvements by avoiding
130  * needless object construction/destruction; it is deferred until absolutely
131  * necessary.
132  *
133  * Caches are grouped into cache groups.  Each cache group references
134  * up to 16 constructed objects.  When a cache allocates an object
135  * from the pool, it calls the object's constructor and places it into
136  * a cache group.  When a cache group frees an object back to the pool,
137  * it first calls the object's destructor.  This allows the object to
138  * persist in constructed form while freed to the cache.
139  *
140  * Multiple caches may exist for each pool.  This allows a single
141  * object type to have multiple constructed forms.  The pool references
142  * each cache, so that when a pool is drained by the pagedaemon, it can
143  * drain each individual cache as well.  Each time a cache is drained,
144  * the most idle cache group is freed to the pool in its entirety.
145  *
146  * Pool caches are layed on top of pools.  By layering them, we can avoid
147  * the complexity of cache management for pools which would not benefit
148  * from it.
149  */
150 
151 /* The cache group pool. */
152 static struct pool pcgpool;
153 
154 /* The pool cache group. */
155 #define	PCG_NOBJECTS		16
156 struct pool_cache_group {
157 	TAILQ_ENTRY(pool_cache_group)
158 		pcg_list;	/* link in the pool cache's group list */
159 	u_int	pcg_avail;	/* # available objects */
160 				/* pointers to the objects */
161 	void	*pcg_objects[PCG_NOBJECTS];
162 };
163 
164 static void	pool_cache_reclaim(struct pool_cache *);
165 
166 static int	pool_catchup(struct pool *);
167 static void	pool_prime_page(struct pool *, caddr_t,
168 		    struct pool_item_header *);
169 static void	*pool_page_alloc(unsigned long, int, int);
170 static void	pool_page_free(void *, unsigned long, int);
171 
172 static void pool_print1(struct pool *, const char *,
173 	int (*)(const char *, ...));
174 
175 /*
176  * Pool log entry. An array of these is allocated in pool_init().
177  */
178 struct pool_log {
179 	const char	*pl_file;
180 	long		pl_line;
181 	int		pl_action;
182 #define	PRLOG_GET	1
183 #define	PRLOG_PUT	2
184 	void		*pl_addr;
185 };
186 
187 /* Number of entries in pool log buffers */
188 #ifndef POOL_LOGSIZE
189 #define	POOL_LOGSIZE	10
190 #endif
191 
192 int pool_logsize = POOL_LOGSIZE;
193 
194 #ifdef POOL_DIAGNOSTIC
195 static __inline void
196 pr_log(struct pool *pp, void *a, int action, const char *file, long line)
197 {
198 	int n = pp->pr_curlogentry;
199 	struct pool_log *pl;
200 
201 	if ((pp->pr_roflags & PR_LOGGING) == 0)
202 		return;
203 
204 	/*
205 	 * Fill in the current entry. Wrap around and overwrite
206 	 * the oldest entry if necessary.
207 	 */
208 	pl = &pp->pr_log[n];
209 	pl->pl_file = file;
210 	pl->pl_line = line;
211 	pl->pl_action = action;
212 	pl->pl_addr = v;
213 	if (++n >= pp->pr_logsize)
214 		n = 0;
215 	pp->pr_curlogentry = n;
216 }
217 
218 static void
219 pr_printlog(struct pool *pp, struct pool_item *pi,
220     int (*pr)(const char *, ...))
221 {
222 	int i = pp->pr_logsize;
223 	int n = pp->pr_curlogentry;
224 
225 	if ((pp->pr_roflags & PR_LOGGING) == 0)
226 		return;
227 
228 	/*
229 	 * Print all entries in this pool's log.
230 	 */
231 	while (i-- > 0) {
232 		struct pool_log *pl = &pp->pr_log[n];
233 		if (pl->pl_action != 0) {
234 			if (pi == NULL || pi == pl->pl_addr) {
235 				(*pr)("\tlog entry %d:\n", i);
236 				(*pr)("\t\taction = %s, addr = %p\n",
237 				    pl->pl_action == PRLOG_GET ? "get" : "put",
238 				    pl->pl_addr);
239 				(*pr)("\t\tfile: %s at line %lu\n",
240 				    pl->pl_file, pl->pl_line);
241 			}
242 		}
243 		if (++n >= pp->pr_logsize)
244 			n = 0;
245 	}
246 }
247 
248 static __inline void
249 pr_enter(struct pool *pp, const char *file, long line)
250 {
251 
252 	if (__predict_false(pp->pr_entered_file != NULL)) {
253 		printf("pool %s: reentrancy at file %s line %ld\n",
254 		    pp->pr_wchan, file, line);
255 		printf("         previous entry at file %s line %ld\n",
256 		    pp->pr_entered_file, pp->pr_entered_line);
257 		panic("pr_enter");
258 	}
259 
260 	pp->pr_entered_file = file;
261 	pp->pr_entered_line = line;
262 }
263 
264 static __inline void
265 pr_leave(struct pool *pp)
266 {
267 
268 	if (__predict_false(pp->pr_entered_file == NULL)) {
269 		printf("pool %s not entered?\n", pp->pr_wchan);
270 		panic("pr_leave");
271 	}
272 
273 	pp->pr_entered_file = NULL;
274 	pp->pr_entered_line = 0;
275 }
276 
277 static __inline__ void
278 pr_enter_check(struct pool *pp, int (*pr)(const char *, ...))
279 {
280 
281 	if (pp->pr_entered_file != NULL)
282 		(*pr)("\n\tcurrently entered from file %s line %ld\n",
283 		    pp->pr_entered_file, pp->pr_entered_line);
284 }
285 #else
286 #define	pr_log(pp, v, action, file, line)
287 #define	pr_printlog(pp, pi, pr)
288 #define	pr_enter(pp, file, line)
289 #define	pr_leave(pp)
290 #define	pr_enter_check(pp, pr)
291 #endif /* POOL_DIAGNOSTIC */
292 
293 /*
294  * Return the pool page header based on page address.
295  */
296 static __inline struct pool_item_header *
297 pr_find_pagehead(struct pool *pp, caddr_t page)
298 {
299 	struct pool_item_header *ph;
300 
301 	if ((pp->pr_roflags & PR_PHINPAGE) != 0)
302 		return ((struct pool_item_header *)(page + pp->pr_phoffset));
303 
304 	for (ph = LIST_FIRST(&pp->pr_hashtab[PR_HASH_INDEX(pp, page)]);
305 	     ph != NULL;
306 	     ph = LIST_NEXT(ph, ph_hashlist)) {
307 		if (ph->ph_page == page)
308 			return (ph);
309 	}
310 	return (NULL);
311 }
312 
313 /*
314  * Remove a page from the pool.
315  */
316 static __inline void
317 pr_rmpage(struct pool *pp, struct pool_item_header *ph)
318 {
319 
320 	/*
321 	 * If the page was idle, decrement the idle page count.
322 	 */
323 	if (ph->ph_nmissing == 0) {
324 #ifdef DIAGNOSTIC
325 		if (pp->pr_nidle == 0)
326 			panic("pr_rmpage: nidle inconsistent");
327 		if (pp->pr_nitems < pp->pr_itemsperpage)
328 			panic("pr_rmpage: nitems inconsistent");
329 #endif
330 		pp->pr_nidle--;
331 	}
332 
333 	pp->pr_nitems -= pp->pr_itemsperpage;
334 
335 	/*
336 	 * Unlink a page from the pool and release it.
337 	 */
338 	TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
339 	(*pp->pr_free)(ph->ph_page, pp->pr_pagesz, pp->pr_mtype);
340 	pp->pr_npages--;
341 	pp->pr_npagefree++;
342 
343 	if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
344 		int s;
345 		LIST_REMOVE(ph, ph_hashlist);
346 		s = splhigh();
347 		pool_put(&phpool, ph);
348 		splx(s);
349 	}
350 
351 	if (pp->pr_curpage == ph) {
352 		/*
353 		 * Find a new non-empty page header, if any.
354 		 * Start search from the page head, to increase the
355 		 * chance for "high water" pages to be freed.
356 		 */
357 		for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
358 		     ph = TAILQ_NEXT(ph, ph_pagelist))
359 			if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
360 				break;
361 
362 		pp->pr_curpage = ph;
363 	}
364 }
365 
366 /*
367  * Initialize the given pool resource structure.
368  *
369  * We export this routine to allow other kernel parts to declare
370  * static pools that must be initialized before malloc() is available.
371  */
372 void
373 pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
374     const char *wchan, size_t pagesz,
375     void *(*alloc)(unsigned long, int, int),
376     void (*release)(void *, unsigned long, int),
377     int mtype)
378 {
379 	int off, slack, i;
380 
381 #ifdef POOL_DIAGNOSTIC
382 	/*
383 	 * Always log if POOL_DIAGNOSTIC is defined.
384 	 */
385 	if (pool_logsize != 0)
386 		flags |= PR_LOGGING;
387 #endif
388 
389 	/*
390 	 * Check arguments and construct default values.
391 	 */
392 	if (!powerof2(pagesz))
393 		panic("pool_init: page size invalid (%lx)\n", (u_long)pagesz);
394 
395 	if (alloc == NULL && release == NULL) {
396 		alloc = pool_page_alloc;
397 		release = pool_page_free;
398 		pagesz = PAGE_SIZE;	/* Rounds to PAGE_SIZE anyhow. */
399 	} else if ((alloc != NULL && release != NULL) == 0) {
400 		/* If you specifiy one, must specify both. */
401 		panic("pool_init: must specify alloc and release together");
402 	}
403 
404 	if (pagesz == 0)
405 		pagesz = PAGE_SIZE;
406 
407 	if (align == 0)
408 		align = ALIGN(1);
409 
410 	if (size < sizeof(struct pool_item))
411 		size = sizeof(struct pool_item);
412 
413 	size = ALIGN(size);
414 	if (size > pagesz)
415 		panic("pool_init: pool item size (%lu) too large",
416 		      (u_long)size);
417 
418 	/*
419 	 * Initialize the pool structure.
420 	 */
421 	TAILQ_INIT(&pp->pr_pagelist);
422 	TAILQ_INIT(&pp->pr_cachelist);
423 	pp->pr_curpage = NULL;
424 	pp->pr_npages = 0;
425 	pp->pr_minitems = 0;
426 	pp->pr_minpages = 0;
427 	pp->pr_maxpages = UINT_MAX;
428 	pp->pr_roflags = flags;
429 	pp->pr_flags = 0;
430 	pp->pr_size = size;
431 	pp->pr_align = align;
432 	pp->pr_wchan = wchan;
433 	pp->pr_mtype = mtype;
434 	pp->pr_alloc = alloc;
435 	pp->pr_free = release;
436 	pp->pr_pagesz = pagesz;
437 	pp->pr_pagemask = ~(pagesz - 1);
438 	pp->pr_pageshift = ffs(pagesz) - 1;
439 	pp->pr_nitems = 0;
440 	pp->pr_nout = 0;
441 	pp->pr_hardlimit = UINT_MAX;
442 	pp->pr_hardlimit_warning = NULL;
443 	pp->pr_hardlimit_ratecap.tv_sec = 0;
444 	pp->pr_hardlimit_ratecap.tv_usec = 0;
445 	pp->pr_hardlimit_warning_last.tv_sec = 0;
446 	pp->pr_hardlimit_warning_last.tv_usec = 0;
447 	pp->pr_serial = ++pool_serial;
448 	if (pool_serial == 0)
449 		panic("pool_init: too much uptime");
450 
451 	/*
452 	 * Decide whether to put the page header off page to avoid
453 	 * wasting too large a part of the page. Off-page page headers
454 	 * go on a hash table, so we can match a returned item
455 	 * with its header based on the page address.
456 	 * We use 1/16 of the page size as the threshold (XXX: tune)
457 	 */
458 	if (pp->pr_size < pagesz/16) {
459 		/* Use the end of the page for the page header */
460 		pp->pr_roflags |= PR_PHINPAGE;
461 		pp->pr_phoffset = off =
462 			pagesz - ALIGN(sizeof(struct pool_item_header));
463 	} else {
464 		/* The page header will be taken from our page header pool */
465 		pp->pr_phoffset = 0;
466 		off = pagesz;
467 		for (i = 0; i < PR_HASHTABSIZE; i++) {
468 			LIST_INIT(&pp->pr_hashtab[i]);
469 		}
470 	}
471 
472 	/*
473 	 * Alignment is to take place at `ioff' within the item. This means
474 	 * we must reserve up to `align - 1' bytes on the page to allow
475 	 * appropriate positioning of each item.
476 	 *
477 	 * Silently enforce `0 <= ioff < align'.
478 	 */
479 	pp->pr_itemoffset = ioff = ioff % align;
480 	pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
481 	KASSERT(pp->pr_itemsperpage != 0);
482 
483 	/*
484 	 * Use the slack between the chunks and the page header
485 	 * for "cache coloring".
486 	 */
487 	slack = off - pp->pr_itemsperpage * pp->pr_size;
488 	pp->pr_maxcolor = (slack / align) * align;
489 	pp->pr_curcolor = 0;
490 
491 	pp->pr_nget = 0;
492 	pp->pr_nfail = 0;
493 	pp->pr_nput = 0;
494 	pp->pr_npagealloc = 0;
495 	pp->pr_npagefree = 0;
496 	pp->pr_hiwat = 0;
497 	pp->pr_nidle = 0;
498 
499 #ifdef POOL_DIAGNOSTIC
500 	if (flags & PR_LOGGING) {
501 		if (kmem_map == NULL ||
502 		    (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
503 		     M_TEMP, M_NOWAIT)) == NULL)
504 			pp->pr_roflags &= ~PR_LOGGING;
505 		pp->pr_curlogentry = 0;
506 		pp->pr_logsize = pool_logsize;
507 	}
508 #endif
509 
510 	pp->pr_entered_file = NULL;
511 	pp->pr_entered_line = 0;
512 
513 	simple_lock_init(&pp->pr_slock);
514 
515 	/*
516 	 * Initialize private page header pool and cache magazine pool if we
517 	 * haven't done so yet.
518 	 * XXX LOCKING.
519 	 */
520 	if (phpool.pr_size == 0) {
521 		pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
522 		    0, "phpool", 0, 0, 0, 0);
523 		pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
524 		    0, "pcgpool", 0, 0, 0, 0);
525 	}
526 
527 	/* Insert into the list of all pools. */
528 	simple_lock(&pool_head_slock);
529 	TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
530 	simple_unlock(&pool_head_slock);
531 }
532 
533 /*
534  * De-commision a pool resource.
535  */
536 void
537 pool_destroy(struct pool *pp)
538 {
539 	struct pool_item_header *ph;
540 	struct pool_cache *pc;
541 
542 	/* Destroy all caches for this pool. */
543 	while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL)
544 		pool_cache_destroy(pc);
545 
546 #ifdef DIAGNOSTIC
547 	if (pp->pr_nout != 0) {
548 		pr_printlog(pp, NULL, printf);
549 		panic("pool_destroy: pool busy: still out: %u\n",
550 		    pp->pr_nout);
551 	}
552 #endif
553 
554 	/* Remove all pages */
555 	if ((pp->pr_roflags & PR_STATIC) == 0)
556 		while ((ph = pp->pr_pagelist.tqh_first) != NULL)
557 			pr_rmpage(pp, ph);
558 
559 	/* Remove from global pool list */
560 	simple_lock(&pool_head_slock);
561 	TAILQ_REMOVE(&pool_head, pp, pr_poollist);
562 	/* XXX Only clear this if we were drainpp? */
563 	drainpp = NULL;
564 	simple_unlock(&pool_head_slock);
565 
566 #ifdef POOL_DIAGNOSTIC
567 	if ((pp->pr_roflags & PR_LOGGING) != 0)
568 		free(pp->pr_log, M_TEMP);
569 #endif
570 
571 	if (pp->pr_roflags & PR_FREEHEADER)
572 		free(pp, M_POOL);
573 }
574 
575 static __inline struct pool_item_header *
576 pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
577 {
578 	struct pool_item_header *ph;
579 	int s;
580 
581 	LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0);
582 
583 	if ((pp->pr_roflags & PR_PHINPAGE) != 0)
584 		ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
585 	else {
586 		s = splhigh();
587 		ph = pool_get(&phpool, flags);
588 		splx(s);
589 	}
590 
591 	return (ph);
592 }
593 
594 /*
595  * Grab an item from the pool; must be called at appropriate spl level
596  */
597 void *
598 #ifdef POOL_DIAGNOSTIC
599 _pool_get(struct pool *pp, int flags, const char *file, long line)
600 #else
601 pool_get(struct pool *pp, int flags)
602 #endif
603 {
604 	struct pool_item *pi;
605 	struct pool_item_header *ph;
606 	void *v;
607 
608 #ifdef DIAGNOSTIC
609 	if (__predict_false((pp->pr_roflags & PR_STATIC) &&
610 			    (flags & PR_MALLOCOK))) {
611 		pr_printlog(pp, NULL, printf);
612 		panic("pool_get: static");
613 	}
614 
615 	if (__predict_false(curproc == NULL && /* doing_shutdown == 0 && XXX*/
616 			    (flags & PR_WAITOK) != 0))
617 		panic("pool_get: must have NOWAIT");
618 
619 #endif
620 	simple_lock(&pp->pr_slock);
621 	pr_enter(pp, file, line);
622 
623  startover:
624 	/*
625 	 * Check to see if we've reached the hard limit.  If we have,
626 	 * and we can wait, then wait until an item has been returned to
627 	 * the pool.
628 	 */
629 #ifdef DIAGNOSTIC
630 	if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
631 		pr_leave(pp);
632 		simple_unlock(&pp->pr_slock);
633 		panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
634 	}
635 #endif
636 	if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
637 		if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
638 			/*
639 			 * XXX: A warning isn't logged in this case.  Should
640 			 * it be?
641 			 */
642 			pp->pr_flags |= PR_WANTED;
643 			pr_leave(pp);
644 			simple_unlock(&pp->pr_slock);
645 			tsleep((caddr_t)pp, PSWP, (char *)pp->pr_wchan, 0);
646 			simple_lock(&pp->pr_slock);
647 			pr_enter(pp, file, line);
648 			goto startover;
649 		}
650 
651 		/*
652 		 * Log a message that the hard limit has been hit.
653 		 */
654 		if (pp->pr_hardlimit_warning != NULL &&
655 		    ratecheck(&pp->pr_hardlimit_warning_last,
656 			      &pp->pr_hardlimit_ratecap))
657 			log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
658 
659 		if (flags & PR_URGENT)
660 			panic("pool_get: urgent");
661 
662 		pp->pr_nfail++;
663 
664 		pr_leave(pp);
665 		simple_unlock(&pp->pr_slock);
666 		return (NULL);
667 	}
668 
669 	/*
670 	 * The convention we use is that if `curpage' is not NULL, then
671 	 * it points at a non-empty bucket. In particular, `curpage'
672 	 * never points at a page header which has PR_PHINPAGE set and
673 	 * has no items in its bucket.
674 	 */
675 	if ((ph = pp->pr_curpage) == NULL) {
676 #ifdef DIAGNOSTIC
677 		if (pp->pr_nitems != 0) {
678 			simple_unlock(&pp->pr_slock);
679 			printf("pool_get: %s: curpage NULL, nitems %u\n",
680 			    pp->pr_wchan, pp->pr_nitems);
681 			panic("pool_get: nitems inconsistent\n");
682 		}
683 #endif
684 
685 		/*
686 		 * Call the back-end page allocator for more memory.
687 		 * Release the pool lock, as the back-end page allocator
688 		 * may block.
689 		 */
690 		pr_leave(pp);
691 		simple_unlock(&pp->pr_slock);
692 		v = (*pp->pr_alloc)(pp->pr_pagesz, flags, pp->pr_mtype);
693 		if (__predict_true(v != NULL))
694 			ph = pool_alloc_item_header(pp, v, flags);
695 		simple_lock(&pp->pr_slock);
696 		pr_enter(pp, file, line);
697 
698 		if (__predict_false(v == NULL || ph == NULL)) {
699 			if (v != NULL)
700 				(*pp->pr_free)(v, pp->pr_pagesz, pp->pr_mtype);
701 
702 			/*
703 			 * We were unable to allocate a page or item
704 			 * header, but we released the lock during
705 			 * allocation, so perhaps items were freed
706 			 * back to the pool.  Check for this case.
707 			 */
708 			if (pp->pr_curpage != NULL)
709 				goto startover;
710 
711 			if (flags & PR_URGENT)
712 				panic("pool_get: urgent");
713 
714 			if ((flags & PR_WAITOK) == 0) {
715 				pp->pr_nfail++;
716 				pr_leave(pp);
717 				simple_unlock(&pp->pr_slock);
718 				return (NULL);
719 			}
720 
721 			/*
722 			 * Wait for items to be returned to this pool.
723 			 *
724 			 * XXX: we actually want to wait just until
725 			 * the page allocator has memory again. Depending
726 			 * on this pool's usage, we might get stuck here
727 			 * for a long time.
728 			 *
729 			 * XXX: maybe we should wake up once a second and
730 			 * try again?
731 			 */
732 			pp->pr_flags |= PR_WANTED;
733 			pr_leave(pp);
734 			simple_unlock(&pp->pr_slock);
735 			tsleep((caddr_t)pp, PSWP, (char *)pp->pr_wchan, 0);
736 			simple_lock(&pp->pr_slock);
737 			pr_enter(pp, file, line);
738 			goto startover;
739 		}
740 
741 		/* We have more memory; add it to the pool */
742 		pp->pr_npagealloc++;
743 		pool_prime_page(pp, v, ph);
744 
745 		/* Start the allocation process over. */
746 		goto startover;
747 	}
748 
749 	if (__predict_false((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL)) {
750 		pr_leave(pp);
751 		simple_unlock(&pp->pr_slock);
752 		panic("pool_get: %s: page empty", pp->pr_wchan);
753 	}
754 #ifdef DIAGNOSTIC
755 	if (__predict_false(pp->pr_nitems == 0)) {
756 		pr_leave(pp);
757 		simple_unlock(&pp->pr_slock);
758 		printf("pool_get: %s: items on itemlist, nitems %u\n",
759 		    pp->pr_wchan, pp->pr_nitems);
760 		panic("pool_get: nitems inconsistent\n");
761 	}
762 
763 	pr_log(pp, v, PRLOG_GET, file, line);
764 
765 	if (__predict_false(pi->pi_magic != PI_MAGIC)) {
766 		pr_printlog(pp, pi, printf);
767 		panic("pool_get(%s): free list modified: magic=%x; page %p;"
768 		       " item addr %p\n",
769 			pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
770 	}
771 #endif
772 
773 	/*
774 	 * Remove from item list.
775 	 */
776 	TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
777 	pp->pr_nitems--;
778 	pp->pr_nout++;
779 	if (ph->ph_nmissing == 0) {
780 #ifdef DIAGNOSTIC
781 		if (__predict_false(pp->pr_nidle == 0))
782 			panic("pool_get: nidle inconsistent");
783 #endif
784 		pp->pr_nidle--;
785 	}
786 	ph->ph_nmissing++;
787 	if (TAILQ_FIRST(&ph->ph_itemlist) == NULL) {
788 #ifdef DIAGNOSTIC
789 		if (__predict_false(ph->ph_nmissing != pp->pr_itemsperpage)) {
790 			pr_leave(pp);
791 			simple_unlock(&pp->pr_slock);
792 			panic("pool_get: %s: nmissing inconsistent",
793 			    pp->pr_wchan);
794 		}
795 #endif
796 		/*
797 		 * Find a new non-empty page header, if any.
798 		 * Start search from the page head, to increase
799 		 * the chance for "high water" pages to be freed.
800 		 *
801 		 * Migrate empty pages to the end of the list.  This
802 		 * will speed the update of curpage as pages become
803 		 * idle.  Empty pages intermingled with idle pages
804 		 * is no big deal.  As soon as a page becomes un-empty,
805 		 * it will move back to the head of the list.
806 		 */
807 		TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
808 		TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
809 		for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
810 		     ph = TAILQ_NEXT(ph, ph_pagelist))
811 			if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
812 				break;
813 
814 		pp->pr_curpage = ph;
815 	}
816 
817 	pp->pr_nget++;
818 
819 	/*
820 	 * If we have a low water mark and we are now below that low
821 	 * water mark, add more items to the pool.
822 	 */
823 	if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
824 		/*
825 		 * XXX: Should we log a warning?  Should we set up a timeout
826 		 * to try again in a second or so?  The latter could break
827 		 * a caller's assumptions about interrupt protection, etc.
828 		 */
829 	}
830 
831 	pr_leave(pp);
832 	simple_unlock(&pp->pr_slock);
833 	return (v);
834 }
835 
836 /*
837  * Internal version of pool_put().  Pool is already locked/entered.
838  */
839 static void
840 pool_do_put(struct pool *pp, void *v)
841 {
842 	struct pool_item *pi = v;
843 	struct pool_item_header *ph;
844 	caddr_t page;
845 	int s;
846 
847 	page = (caddr_t)((u_long)v & pp->pr_pagemask);
848 
849 #ifdef DIAGNOSTIC
850 	if (__predict_false(pp->pr_nout == 0)) {
851 		printf("pool %s: putting with none out\n",
852 		    pp->pr_wchan);
853 		panic("pool_put");
854 	}
855 #endif
856 
857 	if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
858 		pr_printlog(pp, NULL, printf);
859 		panic("pool_put: %s: page header missing", pp->pr_wchan);
860 	}
861 
862 #ifdef LOCKDEBUG
863 	/*
864 	 * Check if we're freeing a locked simple lock.
865 	 */
866 	simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size);
867 #endif
868 
869 	/*
870 	 * Return to item list.
871 	 */
872 #ifdef DIAGNOSTIC
873 	pi->pi_magic = PI_MAGIC;
874 #endif
875 #ifdef DEBUG
876 	{
877 		int i, *ip = v;
878 
879 		for (i = 0; i < pp->pr_size / sizeof(int); i++) {
880 			*ip++ = PI_MAGIC;
881 		}
882 	}
883 #endif
884 
885 	TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
886 	ph->ph_nmissing--;
887 	pp->pr_nput++;
888 	pp->pr_nitems++;
889 	pp->pr_nout--;
890 
891 	/* Cancel "pool empty" condition if it exists */
892 	if (pp->pr_curpage == NULL)
893 		pp->pr_curpage = ph;
894 
895 	if (pp->pr_flags & PR_WANTED) {
896 		pp->pr_flags &= ~PR_WANTED;
897 		if (ph->ph_nmissing == 0)
898 			pp->pr_nidle++;
899 		wakeup((caddr_t)pp);
900 		return;
901 	}
902 
903 	/*
904 	 * If this page is now complete, do one of two things:
905 	 *
906 	 *	(1) If we have more pages than the page high water
907 	 *	    mark, free the page back to the system.
908 	 *
909 	 *	(2) Move it to the end of the page list, so that
910 	 *	    we minimize our chances of fragmenting the
911 	 *	    pool.  Idle pages migrate to the end (along with
912 	 *	    completely empty pages, so that we find un-empty
913 	 *	    pages more quickly when we update curpage) of the
914 	 *	    list so they can be more easily swept up by
915 	 *	    the pagedaemon when pages are scarce.
916 	 */
917 	if (ph->ph_nmissing == 0) {
918 		pp->pr_nidle++;
919 		if (pp->pr_npages > pp->pr_maxpages) {
920 			pr_rmpage(pp, ph);
921 		} else {
922 			TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
923 			TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
924 
925 			/*
926 			 * Update the timestamp on the page.  A page must
927 			 * be idle for some period of time before it can
928 			 * be reclaimed by the pagedaemon.  This minimizes
929 			 * ping-pong'ing for memory.
930 			 */
931 			s = splclock();
932 			ph->ph_time = mono_time;
933 			splx(s);
934 
935 			/*
936 			 * Update the current page pointer.  Just look for
937 			 * the first page with any free items.
938 			 *
939 			 * XXX: Maybe we want an option to look for the
940 			 * page with the fewest available items, to minimize
941 			 * fragmentation?
942 			 */
943 			for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
944 			     ph = TAILQ_NEXT(ph, ph_pagelist))
945 				if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
946 					break;
947 
948 			pp->pr_curpage = ph;
949 		}
950 	}
951 	/*
952 	 * If the page has just become un-empty, move it to the head of
953 	 * the list, and make it the current page.  The next allocation
954 	 * will get the item from this page, instead of further fragmenting
955 	 * the pool.
956 	 */
957 	else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
958 		TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
959 		TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
960 		pp->pr_curpage = ph;
961 	}
962 }
963 
964 /*
965  * Return resource to the pool; must be called at appropriate spl level
966  */
967 #ifdef POOL_DIAGNOSTIC
968 void
969 _pool_put(struct pool *pp, void *v, const char *file, long line)
970 {
971 
972 	simple_lock(&pp->pr_slock);
973 	pr_enter(pp, file, line);
974 
975 	pr_log(pp, v, PRLOG_PUT, file, line);
976 
977 	pool_do_put(pp, v);
978 
979 	pr_leave(pp);
980 	simple_unlock(&pp->pr_slock);
981 }
982 #undef pool_put
983 #endif /* POOL_DIAGNOSTIC */
984 
985 void
986 pool_put(struct pool *pp, void *v)
987 {
988 
989 	simple_lock(&pp->pr_slock);
990 
991 	pool_do_put(pp, v);
992 
993 	simple_unlock(&pp->pr_slock);
994 }
995 
996 #ifdef POOL_DIAGNOSTIC
997 #define		pool_put(h, v)	_pool_put((h), (v), __FILE__, __LINE__)
998 #endif
999 
1000 /*
1001  * Add N items to the pool.
1002  */
1003 int
1004 pool_prime(struct pool *pp, int n)
1005 {
1006 	struct pool_item_header *ph;
1007 	caddr_t cp;
1008 	int newpages, error = 0;
1009 
1010 	simple_lock(&pp->pr_slock);
1011 
1012 	newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1013 
1014 	while (newpages-- > 0) {
1015 		simple_unlock(&pp->pr_slock);
1016 		cp = (*pp->pr_alloc)(pp->pr_pagesz, PR_NOWAIT, pp->pr_mtype);
1017 		if (__predict_true(cp != NULL))
1018 			ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1019 		simple_lock(&pp->pr_slock);
1020 
1021 		if (__predict_false(cp == NULL || ph == NULL)) {
1022 			error = ENOMEM;
1023 			if (cp != NULL)
1024 				(*pp->pr_free)(cp, pp->pr_pagesz, pp->pr_mtype);
1025 			break;
1026 		}
1027 
1028 		pool_prime_page(pp, cp, ph);
1029 		pp->pr_npagealloc++;
1030 		pp->pr_minpages++;
1031 	}
1032 
1033 	if (pp->pr_minpages >= pp->pr_maxpages)
1034 		pp->pr_maxpages = pp->pr_minpages + 1;	/* XXX */
1035 
1036 	simple_unlock(&pp->pr_slock);
1037 	return (0);
1038 }
1039 
1040 /*
1041  * Add a page worth of items to the pool.
1042  *
1043  * Note, we must be called with the pool descriptor LOCKED.
1044  */
1045 static void
1046 pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
1047 {
1048 	struct pool_item *pi;
1049 	caddr_t cp = storage;
1050 	unsigned int align = pp->pr_align;
1051 	unsigned int ioff = pp->pr_itemoffset;
1052 	int n;
1053 
1054 	if (((u_long)cp & (pp->pr_pagesz - 1)) != 0)
1055 		panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1056 
1057 	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1058 		LIST_INSERT_HEAD(&pp->pr_hashtab[PR_HASH_INDEX(pp, cp)],
1059 		    ph, ph_hashlist);
1060 
1061 	/*
1062 	 * Insert page header.
1063 	 */
1064 	TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
1065 	TAILQ_INIT(&ph->ph_itemlist);
1066 	ph->ph_page = storage;
1067 	ph->ph_nmissing = 0;
1068 	memset(&ph->ph_time, 0, sizeof(ph->ph_time));
1069 
1070 	pp->pr_nidle++;
1071 
1072 	/*
1073 	 * Color this page.
1074 	 */
1075 	cp = (caddr_t)(cp + pp->pr_curcolor);
1076 	if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1077 		pp->pr_curcolor = 0;
1078 
1079 	/*
1080 	 * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1081 	 */
1082 	if (ioff != 0)
1083 		cp = (caddr_t)(cp + (align - ioff));
1084 
1085 	/*
1086 	 * Insert remaining chunks on the bucket list.
1087 	 */
1088 	n = pp->pr_itemsperpage;
1089 	pp->pr_nitems += n;
1090 
1091 	while (n--) {
1092 		pi = (struct pool_item *)cp;
1093 
1094 		/* Insert on page list */
1095 		TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
1096 #ifdef DIAGNOSTIC
1097 		pi->pi_magic = PI_MAGIC;
1098 #endif
1099 		cp = (caddr_t)(cp + pp->pr_size);
1100 	}
1101 
1102 	/*
1103 	 * If the pool was depleted, point at the new page.
1104 	 */
1105 	if (pp->pr_curpage == NULL)
1106 		pp->pr_curpage = ph;
1107 
1108 	if (++pp->pr_npages > pp->pr_hiwat)
1109 		pp->pr_hiwat = pp->pr_npages;
1110 }
1111 
1112 /*
1113  * Used by pool_get() when nitems drops below the low water mark.  This
1114  * is used to catch up nitmes with the low water mark.
1115  *
1116  * Note 1, we never wait for memory here, we let the caller decide what to do.
1117  *
1118  * Note 2, this doesn't work with static pools.
1119  *
1120  * Note 3, we must be called with the pool already locked, and we return
1121  * with it locked.
1122  */
1123 static int
1124 pool_catchup(struct pool *pp)
1125 {
1126 	struct pool_item_header *ph;
1127 	caddr_t cp;
1128 	int error = 0;
1129 
1130 	if (pp->pr_roflags & PR_STATIC) {
1131 		/*
1132 		 * We dropped below the low water mark, and this is not a
1133 		 * good thing.  Log a warning.
1134 		 *
1135 		 * XXX: rate-limit this?
1136 		 */
1137 		printf("WARNING: static pool `%s' dropped below low water "
1138 		    "mark\n", pp->pr_wchan);
1139 		return (0);
1140 	}
1141 
1142 	while (POOL_NEEDS_CATCHUP(pp)) {
1143 		/*
1144 		 * Call the page back-end allocator for more memory.
1145 		 *
1146 		 * XXX: We never wait, so should we bother unlocking
1147 		 * the pool descriptor?
1148 		 */
1149 		simple_unlock(&pp->pr_slock);
1150 		cp = (*pp->pr_alloc)(pp->pr_pagesz, PR_NOWAIT, pp->pr_mtype);
1151 		if (__predict_true(cp != NULL))
1152 			ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1153 		simple_lock(&pp->pr_slock);
1154 		if (__predict_false(cp == NULL || ph == NULL)) {
1155 			if (cp != NULL)
1156 				(*pp->pr_free)(cp, pp->pr_pagesz, pp->pr_mtype);
1157 			error = ENOMEM;
1158 			break;
1159 		}
1160 		pool_prime_page(pp, cp, ph);
1161 		pp->pr_npagealloc++;
1162 	}
1163 
1164 	return (error);
1165 }
1166 
1167 void
1168 pool_setlowat(struct pool *pp, int n)
1169 {
1170 	int error;
1171 
1172 	simple_lock(&pp->pr_slock);
1173 
1174 	pp->pr_minitems = n;
1175 	pp->pr_minpages = (n == 0)
1176 		? 0
1177 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1178 
1179 	/* Make sure we're caught up with the newly-set low water mark. */
1180 	if (POOL_NEEDS_CATCHUP(pp) && (error = pool_catchup(pp) != 0)) {
1181 		/*
1182 		 * XXX: Should we log a warning?  Should we set up a timeout
1183 		 * to try again in a second or so?  The latter could break
1184 		 * a caller's assumptions about interrupt protection, etc.
1185 		 */
1186 	}
1187 
1188 	simple_unlock(&pp->pr_slock);
1189 }
1190 
1191 void
1192 pool_sethiwat(struct pool *pp, int n)
1193 {
1194 
1195 	simple_lock(&pp->pr_slock);
1196 
1197 	pp->pr_maxpages = (n == 0)
1198 		? 0
1199 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1200 
1201 	simple_unlock(&pp->pr_slock);
1202 }
1203 
1204 void
1205 pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1206 {
1207 
1208 	simple_lock(&pp->pr_slock);
1209 
1210 	pp->pr_hardlimit = n;
1211 	pp->pr_hardlimit_warning = warnmess;
1212 	pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1213 	pp->pr_hardlimit_warning_last.tv_sec = 0;
1214 	pp->pr_hardlimit_warning_last.tv_usec = 0;
1215 
1216 	/*
1217 	 * In-line version of pool_sethiwat(), because we don't want to
1218 	 * release the lock.
1219 	 */
1220 	pp->pr_maxpages = (n == 0)
1221 		? 0
1222 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1223 
1224 	simple_unlock(&pp->pr_slock);
1225 }
1226 
1227 /*
1228  * Default page allocator.
1229  */
1230 static void *
1231 pool_page_alloc(unsigned long sz, int flags, int mtype)
1232 {
1233 	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
1234 
1235 	return ((void *)uvm_km_alloc_poolpage(waitok));
1236 }
1237 
1238 static void
1239 pool_page_free(void *v, unsigned long sz, int mtype)
1240 {
1241 	uvm_km_free_poolpage((vaddr_t)v);
1242 }
1243 
1244 /*
1245  * Alternate pool page allocator for pools that know they will
1246  * never be accessed in interrupt context.
1247  */
1248 void *
1249 pool_page_alloc_nointr(unsigned long sz, int flags, int mtype)
1250 {
1251 	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
1252 
1253 	return ((void *)uvm_km_alloc_poolpage1(kernel_map, uvm.kernel_object,
1254 	    waitok));
1255 }
1256 
1257 void
1258 pool_page_free_nointr(void *v, unsigned long sz, int mtype)
1259 {
1260 
1261 	uvm_km_free_poolpage1(kernel_map, (vaddr_t)v);
1262 }
1263 
1264 
1265 /*
1266  * Release all complete pages that have not been used recently.
1267  */
1268 void
1269 #ifdef POOL_DIAGNOSTIC
1270 _pool_reclaim(struct pool *pp, const char *file, long line)
1271 #else
1272 pool_reclaim(struct pool *pp)
1273 #endif
1274 {
1275 	struct pool_item_header *ph, *phnext;
1276 	struct pool_cache *pc;
1277 	struct timeval curtime;
1278 	int s;
1279 
1280 	if (pp->pr_roflags & PR_STATIC)
1281 		return;
1282 
1283 	if (simple_lock_try(&pp->pr_slock) == 0)
1284 		return;
1285 	pr_enter(pp, file, line);
1286 
1287 	/*
1288 	 * Reclaim items from the pool's caches.
1289 	 */
1290 	for (pc = TAILQ_FIRST(&pp->pr_cachelist); pc != NULL;
1291 	     pc = TAILQ_NEXT(pc, pc_poollist))
1292 		pool_cache_reclaim(pc);
1293 
1294 	s = splclock();
1295 	curtime = mono_time;
1296 	splx(s);
1297 
1298 	for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; ph = phnext) {
1299 		phnext = TAILQ_NEXT(ph, ph_pagelist);
1300 
1301 		/* Check our minimum page claim */
1302 		if (pp->pr_npages <= pp->pr_minpages)
1303 			break;
1304 
1305 		if (ph->ph_nmissing == 0) {
1306 			struct timeval diff;
1307 			timersub(&curtime, &ph->ph_time, &diff);
1308 			if (diff.tv_sec < pool_inactive_time)
1309 				continue;
1310 
1311 			/*
1312 			 * If freeing this page would put us below
1313 			 * the low water mark, stop now.
1314 			 */
1315 			if ((pp->pr_nitems - pp->pr_itemsperpage) <
1316 			    pp->pr_minitems)
1317 				break;
1318 
1319 			pr_rmpage(pp, ph);
1320 		}
1321 	}
1322 
1323 	pr_leave(pp);
1324 	simple_unlock(&pp->pr_slock);
1325 }
1326 
1327 
1328 /*
1329  * Drain pools, one at a time.
1330  *
1331  * Note, we must never be called from an interrupt context.
1332  */
1333 void
1334 pool_drain(void *arg)
1335 {
1336 	struct pool *pp;
1337 	int s;
1338 
1339 	s = splvm();
1340 	simple_lock(&pool_head_slock);
1341 
1342 	if (drainpp == NULL && (drainpp = TAILQ_FIRST(&pool_head)) == NULL)
1343 		goto out;
1344 
1345 	pp = drainpp;
1346 	drainpp = TAILQ_NEXT(pp, pr_poollist);
1347 
1348 	pool_reclaim(pp);
1349 
1350  out:
1351 	simple_unlock(&pool_head_slock);
1352 	splx(s);
1353 }
1354 
1355 
1356 /*
1357  * Diagnostic helpers.
1358  */
1359 void
1360 pool_printit(struct pool *pp, const char *modif, int (*pr)(const char *, ...))
1361 {
1362 	int s;
1363 
1364 	s = splvm();
1365 	if (simple_lock_try(&pp->pr_slock) == 0) {
1366 		printf("pool %s is locked; try again later\n",
1367 		    pp->pr_wchan);
1368 		splx(s);
1369 		return;
1370 	}
1371 	pool_print1(pp, modif, printf);
1372 	simple_unlock(&pp->pr_slock);
1373 	splx(s);
1374 }
1375 
1376 static void
1377 pool_print1(struct pool *pp, const char *modif, int (*pr)(const char *, ...))
1378 {
1379 	struct pool_item_header *ph;
1380 	struct pool_cache *pc;
1381 	struct pool_cache_group *pcg;
1382 #ifdef DIAGNOSTIC
1383 	struct pool_item *pi;
1384 #endif
1385 	int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1386 	char c;
1387 
1388 	while ((c = *modif++) != '\0') {
1389 		if (c == 'l')
1390 			print_log = 1;
1391 		if (c == 'p')
1392 			print_pagelist = 1;
1393 		if (c == 'c')
1394 			print_cache = 1;
1395 		modif++;
1396 	}
1397 
1398 	(*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1399 	    pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1400 	    pp->pr_roflags);
1401 	(*pr)("\tpagesz %u, mtype %d\n", pp->pr_pagesz, pp->pr_mtype);
1402 	(*pr)("\talloc %p, release %p\n", pp->pr_alloc, pp->pr_free);
1403 	(*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1404 	    pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1405 	(*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1406 	    pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1407 
1408 	(*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1409 	    pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1410 	(*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1411 	    pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1412 
1413 	if (print_pagelist == 0)
1414 		goto skip_pagelist;
1415 
1416 	if ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL)
1417 		(*pr)("\n\tpage list:\n");
1418 	for (; ph != NULL; ph = TAILQ_NEXT(ph, ph_pagelist)) {
1419 		(*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1420 		    ph->ph_page, ph->ph_nmissing,
1421 		    (u_long)ph->ph_time.tv_sec,
1422 		    (u_long)ph->ph_time.tv_usec);
1423 #ifdef DIAGNOSTIC
1424 		for (pi = TAILQ_FIRST(&ph->ph_itemlist); pi != NULL;
1425 		     pi = TAILQ_NEXT(pi, pi_list)) {
1426 			if (pi->pi_magic != PI_MAGIC) {
1427 				(*pr)("\t\t\titem %p, magic 0x%x\n",
1428 				    pi, pi->pi_magic);
1429 			}
1430 		}
1431 #endif
1432 	}
1433 	if (pp->pr_curpage == NULL)
1434 		(*pr)("\tno current page\n");
1435 	else
1436 		(*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1437 
1438  skip_pagelist:
1439 
1440 	if (print_log == 0)
1441 		goto skip_log;
1442 
1443 	(*pr)("\n");
1444 	if ((pp->pr_roflags & PR_LOGGING) == 0)
1445 		(*pr)("\tno log\n");
1446 	else
1447 		pr_printlog(pp, NULL, pr);
1448 
1449  skip_log:
1450 
1451 	if (print_cache == 0)
1452 		goto skip_cache;
1453 
1454 	for (pc = TAILQ_FIRST(&pp->pr_cachelist); pc != NULL;
1455 	     pc = TAILQ_NEXT(pc, pc_poollist)) {
1456 		(*pr)("\tcache %p: allocfrom %p freeto %p\n", pc,
1457 		    pc->pc_allocfrom, pc->pc_freeto);
1458 		(*pr)("\t    hits %lu misses %lu ngroups %lu nitems %lu\n",
1459 		    pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
1460 		for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1461 		     pcg = TAILQ_NEXT(pcg, pcg_list)) {
1462 			(*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail);
1463 			for (i = 0; i < PCG_NOBJECTS; i++)
1464 				(*pr)("\t\t\t%p\n", pcg->pcg_objects[i]);
1465 		}
1466 	}
1467 
1468  skip_cache:
1469 
1470 	pr_enter_check(pp, pr);
1471 }
1472 
1473 int
1474 pool_chk(struct pool *pp, const char *label)
1475 {
1476 	struct pool_item_header *ph;
1477 	int r = 0;
1478 
1479 	simple_lock(&pp->pr_slock);
1480 
1481 	for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
1482 	     ph = TAILQ_NEXT(ph, ph_pagelist)) {
1483 
1484 		struct pool_item *pi;
1485 		int n;
1486 		caddr_t page;
1487 
1488 		page = (caddr_t)((u_long)ph & pp->pr_pagemask);
1489 		if (page != ph->ph_page &&
1490 		    (pp->pr_roflags & PR_PHINPAGE) != 0) {
1491 			if (label != NULL)
1492 				printf("%s: ", label);
1493 			printf("pool(%p:%s): page inconsistency: page %p;"
1494 			       " at page head addr %p (p %p)\n", pp,
1495 				pp->pr_wchan, ph->ph_page,
1496 				ph, page);
1497 			r++;
1498 			goto out;
1499 		}
1500 
1501 		for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
1502 		     pi != NULL;
1503 		     pi = TAILQ_NEXT(pi,pi_list), n++) {
1504 
1505 #ifdef DIAGNOSTIC
1506 			if (pi->pi_magic != PI_MAGIC) {
1507 				if (label != NULL)
1508 					printf("%s: ", label);
1509 				printf("pool(%s): free list modified: magic=%x;"
1510 				       " page %p; item ordinal %d;"
1511 				       " addr %p (p %p)\n",
1512 					pp->pr_wchan, pi->pi_magic, ph->ph_page,
1513 					n, pi, page);
1514 				panic("pool");
1515 			}
1516 #endif
1517 			page = (caddr_t)((u_long)pi & pp->pr_pagemask);
1518 			if (page == ph->ph_page)
1519 				continue;
1520 
1521 			if (label != NULL)
1522 				printf("%s: ", label);
1523 			printf("pool(%p:%s): page inconsistency: page %p;"
1524 			       " item ordinal %d; addr %p (p %p)\n", pp,
1525 				pp->pr_wchan, ph->ph_page,
1526 				n, pi, page);
1527 			r++;
1528 			goto out;
1529 		}
1530 	}
1531 out:
1532 	simple_unlock(&pp->pr_slock);
1533 	return (r);
1534 }
1535 
1536 /*
1537  * pool_cache_init:
1538  *
1539  *	Initialize a pool cache.
1540  *
1541  *	NOTE: If the pool must be protected from interrupts, we expect
1542  *	to be called at the appropriate interrupt priority level.
1543  */
1544 void
1545 pool_cache_init(struct pool_cache *pc, struct pool *pp,
1546     int (*ctor)(void *, void *, int),
1547     void (*dtor)(void *, void *),
1548     void *arg)
1549 {
1550 
1551 	TAILQ_INIT(&pc->pc_grouplist);
1552 	simple_lock_init(&pc->pc_slock);
1553 
1554 	pc->pc_allocfrom = NULL;
1555 	pc->pc_freeto = NULL;
1556 	pc->pc_pool = pp;
1557 
1558 	pc->pc_ctor = ctor;
1559 	pc->pc_dtor = dtor;
1560 	pc->pc_arg  = arg;
1561 
1562 	pc->pc_hits   = 0;
1563 	pc->pc_misses = 0;
1564 
1565 	pc->pc_ngroups = 0;
1566 
1567 	pc->pc_nitems = 0;
1568 
1569 	simple_lock(&pp->pr_slock);
1570 	TAILQ_INSERT_TAIL(&pp->pr_cachelist, pc, pc_poollist);
1571 	simple_unlock(&pp->pr_slock);
1572 }
1573 
1574 /*
1575  * pool_cache_destroy:
1576  *
1577  *	Destroy a pool cache.
1578  */
1579 void
1580 pool_cache_destroy(struct pool_cache *pc)
1581 {
1582 	struct pool *pp = pc->pc_pool;
1583 
1584 	/* First, invalidate the entire cache. */
1585 	pool_cache_invalidate(pc);
1586 
1587 	/* ...and remove it from the pool's cache list. */
1588 	simple_lock(&pp->pr_slock);
1589 	TAILQ_REMOVE(&pp->pr_cachelist, pc, pc_poollist);
1590 	simple_unlock(&pp->pr_slock);
1591 }
1592 
1593 static __inline void *
1594 pcg_get(struct pool_cache_group *pcg)
1595 {
1596 	void *object;
1597 	u_int idx;
1598 
1599 	KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
1600 	KASSERT(pcg->pcg_avail != 0);
1601 	idx = --pcg->pcg_avail;
1602 
1603 	KASSERT(pcg->pcg_objects[idx] != NULL);
1604 	object = pcg->pcg_objects[idx];
1605 	pcg->pcg_objects[idx] = NULL;
1606 
1607 	return (object);
1608 }
1609 
1610 static __inline void
1611 pcg_put(struct pool_cache_group *pcg, void *object)
1612 {
1613 	u_int idx;
1614 
1615 	KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
1616 	idx = pcg->pcg_avail++;
1617 
1618 	KASSERT(pcg->pcg_objects[idx] == NULL);
1619 	pcg->pcg_objects[idx] = object;
1620 }
1621 
1622 /*
1623  * pool_cache_get:
1624  *
1625  *	Get an object from a pool cache.
1626  */
1627 void *
1628 pool_cache_get(struct pool_cache *pc, int flags)
1629 {
1630 	struct pool_cache_group *pcg;
1631 	void *object;
1632 
1633 #ifdef LOCKDEBUG
1634 	if (flags & PR_WAITOK)
1635 		simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)");
1636 #endif
1637 
1638 	simple_lock(&pc->pc_slock);
1639 
1640 	if ((pcg = pc->pc_allocfrom) == NULL) {
1641 		for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1642 		     pcg = TAILQ_NEXT(pcg, pcg_list)) {
1643 			if (pcg->pcg_avail != 0) {
1644 				pc->pc_allocfrom = pcg;
1645 				goto have_group;
1646 			}
1647 		}
1648 
1649 		/*
1650 		 * No groups with any available objects.  Allocate
1651 		 * a new object, construct it, and return it to
1652 		 * the caller.  We will allocate a group, if necessary,
1653 		 * when the object is freed back to the cache.
1654 		 */
1655 		pc->pc_misses++;
1656 		simple_unlock(&pc->pc_slock);
1657 		object = pool_get(pc->pc_pool, flags);
1658 		if (object != NULL && pc->pc_ctor != NULL) {
1659 			if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
1660 				pool_put(pc->pc_pool, object);
1661 				return (NULL);
1662 			}
1663 		}
1664 		return (object);
1665 	}
1666 
1667  have_group:
1668 	pc->pc_hits++;
1669 	pc->pc_nitems--;
1670 	object = pcg_get(pcg);
1671 
1672 	if (pcg->pcg_avail == 0)
1673 		pc->pc_allocfrom = NULL;
1674 
1675 	simple_unlock(&pc->pc_slock);
1676 
1677 	return (object);
1678 }
1679 
1680 /*
1681  * pool_cache_put:
1682  *
1683  *	Put an object back to the pool cache.
1684  */
1685 void
1686 pool_cache_put(struct pool_cache *pc, void *object)
1687 {
1688 	struct pool_cache_group *pcg;
1689 
1690 	simple_lock(&pc->pc_slock);
1691 
1692 	if ((pcg = pc->pc_freeto) == NULL) {
1693 		for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1694 		     pcg = TAILQ_NEXT(pcg, pcg_list)) {
1695 			if (pcg->pcg_avail != PCG_NOBJECTS) {
1696 				pc->pc_freeto = pcg;
1697 				goto have_group;
1698 			}
1699 		}
1700 
1701 		/*
1702 		 * No empty groups to free the object to.  Attempt to
1703 		 * allocate one.
1704 		 */
1705 		simple_unlock(&pc->pc_slock);
1706 		pcg = pool_get(&pcgpool, PR_NOWAIT);
1707 		if (pcg != NULL) {
1708 			memset(pcg, 0, sizeof(*pcg));
1709 			simple_lock(&pc->pc_slock);
1710 			pc->pc_ngroups++;
1711 			TAILQ_INSERT_TAIL(&pc->pc_grouplist, pcg, pcg_list);
1712 			if (pc->pc_freeto == NULL)
1713 				pc->pc_freeto = pcg;
1714 			goto have_group;
1715 		}
1716 
1717 		/*
1718 		 * Unable to allocate a cache group; destruct the object
1719 		 * and free it back to the pool.
1720 		 */
1721 		pool_cache_destruct_object(pc, object);
1722 		return;
1723 	}
1724 
1725  have_group:
1726 	pc->pc_nitems++;
1727 	pcg_put(pcg, object);
1728 
1729 	if (pcg->pcg_avail == PCG_NOBJECTS)
1730 		pc->pc_freeto = NULL;
1731 
1732 	simple_unlock(&pc->pc_slock);
1733 }
1734 
1735 /*
1736  * pool_cache_destruct_object:
1737  *
1738  *	Force destruction of an object and its release back into
1739  *	the pool.
1740  */
1741 void
1742 pool_cache_destruct_object(struct pool_cache *pc, void *object)
1743 {
1744 
1745 	if (pc->pc_dtor != NULL)
1746 		(*pc->pc_dtor)(pc->pc_arg, object);
1747 	pool_put(pc->pc_pool, object);
1748 }
1749 
1750 /*
1751  * pool_cache_do_invalidate:
1752  *
1753  *	This internal function implements pool_cache_invalidate() and
1754  *	pool_cache_reclaim().
1755  */
1756 static void
1757 pool_cache_do_invalidate(struct pool_cache *pc, int free_groups,
1758     void (*putit)(struct pool *, void *))
1759 {
1760 	struct pool_cache_group *pcg, *npcg;
1761 	void *object;
1762 
1763 	for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1764 	     pcg = npcg) {
1765 		npcg = TAILQ_NEXT(pcg, pcg_list);
1766 		while (pcg->pcg_avail != 0) {
1767 			pc->pc_nitems--;
1768 			object = pcg_get(pcg);
1769 			if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg)
1770 				pc->pc_allocfrom = NULL;
1771 			if (pc->pc_dtor != NULL)
1772 				(*pc->pc_dtor)(pc->pc_arg, object);
1773 			(*putit)(pc->pc_pool, object);
1774 		}
1775 		if (free_groups) {
1776 			pc->pc_ngroups--;
1777 			TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list);
1778 			if (pc->pc_freeto == pcg)
1779 				pc->pc_freeto = NULL;
1780 			pool_put(&pcgpool, pcg);
1781 		}
1782 	}
1783 }
1784 
1785 /*
1786  * pool_cache_invalidate:
1787  *
1788  *	Invalidate a pool cache (destruct and release all of the
1789  *	cached objects).
1790  */
1791 void
1792 pool_cache_invalidate(struct pool_cache *pc)
1793 {
1794 
1795 	simple_lock(&pc->pc_slock);
1796 	pool_cache_do_invalidate(pc, 0, pool_put);
1797 	simple_unlock(&pc->pc_slock);
1798 }
1799 
1800 /*
1801  * pool_cache_reclaim:
1802  *
1803  *	Reclaim a pool cache for pool_reclaim().
1804  */
1805 static void
1806 pool_cache_reclaim(struct pool_cache *pc)
1807 {
1808 
1809 	simple_lock(&pc->pc_slock);
1810 	pool_cache_do_invalidate(pc, 1, pool_do_put);
1811 	simple_unlock(&pc->pc_slock);
1812 }
1813 
1814 /*
1815  * We have three different sysctls.
1816  * kern.pool.npools - the number of pools.
1817  * kern.pool.pool.<pool#> - the pool struct for the pool#.
1818  * kern.pool.name.<pool#> - the name for pool#.[6~
1819  */
1820 int
1821 sysctl_dopool(int *name, u_int namelen, char *where, size_t *sizep)
1822 {
1823 	struct pool *pp, *foundpool = NULL;
1824 	size_t buflen = where != NULL ? *sizep : 0;
1825 	int npools = 0, s;
1826 	unsigned int lookfor;
1827 	size_t len;
1828 
1829 	switch (*name) {
1830 	case KERN_POOL_NPOOLS:
1831 		if (namelen != 1 || buflen != sizeof(int))
1832 			return (EINVAL);
1833 		lookfor = 0;
1834 		break;
1835 	case KERN_POOL_NAME:
1836 		if (namelen != 2 || buflen < 1)
1837 			return (EINVAL);
1838 		lookfor = name[1];
1839 		break;
1840 	case KERN_POOL_POOL:
1841 		if (namelen != 2 || buflen != sizeof(struct pool))
1842 			return (EINVAL);
1843 		lookfor = name[1];
1844 		break;
1845 	default:
1846 		return (EINVAL);
1847 	}
1848 
1849 	s = splvm();
1850 	simple_lock(&pool_head_slock);
1851 
1852 	TAILQ_FOREACH(pp, &pool_head, pr_poollist) {
1853 		npools++;
1854 		if (lookfor == pp->pr_serial) {
1855 			foundpool = pp;
1856 			break;
1857 		}
1858 	}
1859 
1860 	simple_unlock(&pool_head_slock);
1861 	splx(s);
1862 
1863 	if (lookfor != 0 && foundpool == NULL)
1864 		return (ENOENT);
1865 
1866 	switch (*name) {
1867 	case KERN_POOL_NPOOLS:
1868 		return copyout(&npools, where, buflen);
1869 	case KERN_POOL_NAME:
1870 		len = strlen(foundpool->pr_wchan) + 1;
1871 		if (*sizep < len)
1872 			return (ENOMEM);
1873 		*sizep = len;
1874 		return copyout(foundpool->pr_wchan, where, len);
1875 	case KERN_POOL_POOL:
1876 		return copyout(foundpool, where, buflen);
1877 	}
1878 	/* NOTREACHED */
1879 	return (0); /* XXX - Stupid gcc */
1880 }
1881