xref: /openbsd-src/sys/kern/subr_pool.c (revision a28daedfc357b214be5c701aa8ba8adb29a7f1c2)
1 /*	$OpenBSD: subr_pool.c,v 1.79 2009/04/22 01:16:11 dlg Exp $	*/
2 /*	$NetBSD: subr_pool.c,v 1.61 2001/09/26 07:14:56 chs 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  *
21  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
23  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
24  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
25  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
26  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
27  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
28  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
29  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
30  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
31  * POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/proc.h>
37 #include <sys/errno.h>
38 #include <sys/kernel.h>
39 #include <sys/malloc.h>
40 #include <sys/pool.h>
41 #include <sys/syslog.h>
42 #include <sys/sysctl.h>
43 
44 #include <uvm/uvm.h>
45 
46 
47 /*
48  * Pool resource management utility.
49  *
50  * Memory is allocated in pages which are split into pieces according to
51  * the pool item size. Each page is kept on one of three lists in the
52  * pool structure: `pr_emptypages', `pr_fullpages' and `pr_partpages',
53  * for empty, full and partially-full pages respectively. The individual
54  * pool items are on a linked list headed by `ph_itemlist' in each page
55  * header. The memory for building the page list is either taken from
56  * the allocated pages themselves (for small pool items) or taken from
57  * an internal pool of page headers (`phpool').
58  */
59 
60 /* List of all pools */
61 TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head);
62 
63 /* Private pool for page header structures */
64 struct pool phpool;
65 
66 struct pool_item_header {
67 	/* Page headers */
68 	LIST_ENTRY(pool_item_header)
69 				ph_pagelist;	/* pool page list */
70 	TAILQ_HEAD(,pool_item)	ph_itemlist;	/* chunk list for this page */
71 	SPLAY_ENTRY(pool_item_header)
72 				ph_node;	/* Off-page page headers */
73 	int			ph_nmissing;	/* # of chunks in use */
74 	caddr_t			ph_page;	/* this page's address */
75 	caddr_t			ph_colored;	/* page's colored address */
76 	int			ph_pagesize;
77 };
78 
79 struct pool_item {
80 #ifdef DIAGNOSTIC
81 	u_int32_t pi_magic;
82 #endif
83 	/* Other entries use only this list entry */
84 	TAILQ_ENTRY(pool_item)	pi_list;
85 };
86 
87 #ifdef DEADBEEF1
88 #define	PI_MAGIC DEADBEEF1
89 #else
90 #define	PI_MAGIC 0xdeafbeef
91 #endif
92 
93 #define	POOL_NEEDS_CATCHUP(pp)						\
94 	((pp)->pr_nitems < (pp)->pr_minitems)
95 
96 /*
97  * Every pool gets a unique serial number assigned to it. If this counter
98  * wraps, we're screwed, but we shouldn't create so many pools anyway.
99  */
100 unsigned int pool_serial;
101 
102 int	 pool_catchup(struct pool *);
103 void	 pool_prime_page(struct pool *, caddr_t, struct pool_item_header *);
104 void	 pool_update_curpage(struct pool *);
105 void	*pool_do_get(struct pool *, int);
106 void	 pool_do_put(struct pool *, void *);
107 void	 pr_rmpage(struct pool *, struct pool_item_header *,
108 	    struct pool_pagelist *);
109 int	pool_chk_page(struct pool *, const char *, struct pool_item_header *);
110 struct pool_item_header *pool_alloc_item_header(struct pool *, caddr_t , int);
111 
112 void	*pool_allocator_alloc(struct pool *, int, int *);
113 void	 pool_allocator_free(struct pool *, void *);
114 
115 /*
116  * XXX - quick hack. For pools with large items we want to use a special
117  *       allocator. For now, instead of having the allocator figure out
118  *       the allocation size from the pool (which can be done trivially
119  *       with round_page(pr_itemsperpage * pr_size)) which would require
120  *	 lots of changes everywhere, we just create allocators for each
121  *	 size. We limit those to 128 pages.
122  */
123 #define POOL_LARGE_MAXPAGES 128
124 struct pool_allocator pool_allocator_large[POOL_LARGE_MAXPAGES];
125 struct pool_allocator pool_allocator_large_ni[POOL_LARGE_MAXPAGES];
126 void	*pool_large_alloc(struct pool *, int, int *);
127 void	pool_large_free(struct pool *, void *);
128 void	*pool_large_alloc_ni(struct pool *, int, int *);
129 void	pool_large_free_ni(struct pool *, void *);
130 
131 
132 #ifdef DDB
133 void	 pool_print_pagelist(struct pool_pagelist *,
134 	    int (*)(const char *, ...));
135 void	 pool_print1(struct pool *, const char *, int (*)(const char *, ...));
136 #endif
137 
138 #define pool_sleep(pl) msleep(pl, &pl->pr_mtx, PSWP, pl->pr_wchan, 0)
139 
140 static __inline int
141 phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
142 {
143 	long diff = (vaddr_t)a->ph_page - (vaddr_t)b->ph_page;
144 	if (diff < 0)
145 		return -(-diff >= a->ph_pagesize);
146 	else if (diff > 0)
147 		return (diff >= b->ph_pagesize);
148 	else
149 		return (0);
150 }
151 
152 SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
153 SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
154 
155 /*
156  * Return the pool page header based on page address.
157  */
158 static __inline struct pool_item_header *
159 pr_find_pagehead(struct pool *pp, void *v)
160 {
161 	struct pool_item_header *ph, tmp;
162 
163 	if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
164 		caddr_t page;
165 
166 		page = (caddr_t)((vaddr_t)v & pp->pr_alloc->pa_pagemask);
167 
168 		return ((struct pool_item_header *)(page + pp->pr_phoffset));
169 	}
170 
171 	/*
172 	 * The trick we're using in the tree compare function is to compare
173 	 * two elements equal when they overlap. We want to return the
174 	 * page header that belongs to the element just before this address.
175 	 * We don't want this element to compare equal to the next element,
176 	 * so the compare function takes the pagesize from the lower element.
177 	 * If this header is the lower, its pagesize is zero, so it can't
178 	 * overlap with the next header. But if the header we're looking for
179 	 * is lower, we'll use its pagesize and it will overlap and return
180 	 * equal.
181 	 */
182 	tmp.ph_page = v;
183 	tmp.ph_pagesize = 0;
184 	ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
185 
186 	if (ph) {
187 		KASSERT(ph->ph_page <= (caddr_t)v);
188 		KASSERT(ph->ph_page + ph->ph_pagesize > (caddr_t)v);
189 	}
190 	return ph;
191 }
192 
193 /*
194  * Remove a page from the pool.
195  */
196 void
197 pr_rmpage(struct pool *pp, struct pool_item_header *ph,
198     struct pool_pagelist *pq)
199 {
200 
201 	/*
202 	 * If the page was idle, decrement the idle page count.
203 	 */
204 	if (ph->ph_nmissing == 0) {
205 #ifdef DIAGNOSTIC
206 		if (pp->pr_nidle == 0)
207 			panic("pr_rmpage: nidle inconsistent");
208 		if (pp->pr_nitems < pp->pr_itemsperpage)
209 			panic("pr_rmpage: nitems inconsistent");
210 #endif
211 		pp->pr_nidle--;
212 	}
213 
214 	pp->pr_nitems -= pp->pr_itemsperpage;
215 
216 	/*
217 	 * Unlink a page from the pool and release it (or queue it for release).
218 	 */
219 	LIST_REMOVE(ph, ph_pagelist);
220 	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
221 		SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
222 	if (pq) {
223 		LIST_INSERT_HEAD(pq, ph, ph_pagelist);
224 	} else {
225 		pool_allocator_free(pp, ph->ph_page);
226 		if ((pp->pr_roflags & PR_PHINPAGE) == 0)
227 			pool_put(&phpool, ph);
228 	}
229 	pp->pr_npages--;
230 	pp->pr_npagefree++;
231 
232 	pool_update_curpage(pp);
233 }
234 
235 /*
236  * Initialize the given pool resource structure.
237  *
238  * We export this routine to allow other kernel parts to declare
239  * static pools that must be initialized before malloc() is available.
240  */
241 void
242 pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
243     const char *wchan, struct pool_allocator *palloc)
244 {
245 	int off, slack;
246 
247 #ifdef MALLOC_DEBUG
248 	if ((flags & PR_DEBUG) && (ioff != 0 || align != 0))
249 		flags &= ~PR_DEBUG;
250 #endif
251 	/*
252 	 * Check arguments and construct default values.
253 	 */
254 	if (palloc == NULL) {
255 		if (size > PAGE_SIZE) {
256 			int psize;
257 
258 			/*
259 			 * XXX - should take align into account as well.
260 			 */
261 			if (size == round_page(size))
262 				psize = size / PAGE_SIZE;
263 			else
264 				psize = PAGE_SIZE / roundup(size % PAGE_SIZE,
265 				    1024);
266 			if (psize > POOL_LARGE_MAXPAGES)
267 				psize = POOL_LARGE_MAXPAGES;
268 			if (flags & PR_WAITOK)
269 				palloc = &pool_allocator_large_ni[psize-1];
270 			else
271 				palloc = &pool_allocator_large[psize-1];
272 			if (palloc->pa_pagesz == 0) {
273 				palloc->pa_pagesz = psize * PAGE_SIZE;
274 				if (flags & PR_WAITOK) {
275 					palloc->pa_alloc = pool_large_alloc_ni;
276 					palloc->pa_free = pool_large_free_ni;
277 				} else {
278 					palloc->pa_alloc = pool_large_alloc;
279 					palloc->pa_free = pool_large_free;
280 				}
281 			}
282 		} else {
283 			palloc = &pool_allocator_nointr;
284 		}
285 	}
286 	if (palloc->pa_pagesz == 0) {
287 		palloc->pa_pagesz = PAGE_SIZE;
288 	}
289 	if (palloc->pa_pagemask == 0) {
290 		palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
291 		palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
292 	}
293 
294 	if (align == 0)
295 		align = ALIGN(1);
296 
297 	if (size < sizeof(struct pool_item))
298 		size = sizeof(struct pool_item);
299 
300 	size = roundup(size, align);
301 #ifdef DIAGNOSTIC
302 	if (size > palloc->pa_pagesz)
303 		panic("pool_init: pool item size (%lu) too large",
304 		    (u_long)size);
305 #endif
306 
307 	/*
308 	 * Initialize the pool structure.
309 	 */
310 	LIST_INIT(&pp->pr_emptypages);
311 	LIST_INIT(&pp->pr_fullpages);
312 	LIST_INIT(&pp->pr_partpages);
313 	pp->pr_curpage = NULL;
314 	pp->pr_npages = 0;
315 	pp->pr_minitems = 0;
316 	pp->pr_minpages = 0;
317 	pp->pr_maxpages = 8;
318 	pp->pr_roflags = flags;
319 	pp->pr_flags = 0;
320 	pp->pr_size = size;
321 	pp->pr_align = align;
322 	pp->pr_wchan = wchan;
323 	pp->pr_alloc = palloc;
324 	pp->pr_nitems = 0;
325 	pp->pr_nout = 0;
326 	pp->pr_hardlimit = UINT_MAX;
327 	pp->pr_hardlimit_warning = NULL;
328 	pp->pr_hardlimit_ratecap.tv_sec = 0;
329 	pp->pr_hardlimit_ratecap.tv_usec = 0;
330 	pp->pr_hardlimit_warning_last.tv_sec = 0;
331 	pp->pr_hardlimit_warning_last.tv_usec = 0;
332 	pp->pr_serial = ++pool_serial;
333 	if (pool_serial == 0)
334 		panic("pool_init: too much uptime");
335 
336         /* constructor, destructor, and arg */
337 	pp->pr_ctor = NULL;
338 	pp->pr_dtor = NULL;
339 	pp->pr_arg = NULL;
340 
341 	/*
342 	 * Decide whether to put the page header off page to avoid
343 	 * wasting too large a part of the page. Off-page page headers
344 	 * go on a hash table, so we can match a returned item
345 	 * with its header based on the page address.
346 	 * We use 1/16 of the page size as the threshold (XXX: tune)
347 	 */
348 	if (pp->pr_size < palloc->pa_pagesz/16 && pp->pr_size < PAGE_SIZE) {
349 		/* Use the end of the page for the page header */
350 		pp->pr_roflags |= PR_PHINPAGE;
351 		pp->pr_phoffset = off = palloc->pa_pagesz -
352 		    ALIGN(sizeof(struct pool_item_header));
353 	} else {
354 		/* The page header will be taken from our page header pool */
355 		pp->pr_phoffset = 0;
356 		off = palloc->pa_pagesz;
357 		SPLAY_INIT(&pp->pr_phtree);
358 	}
359 
360 	/*
361 	 * Alignment is to take place at `ioff' within the item. This means
362 	 * we must reserve up to `align - 1' bytes on the page to allow
363 	 * appropriate positioning of each item.
364 	 *
365 	 * Silently enforce `0 <= ioff < align'.
366 	 */
367 	pp->pr_itemoffset = ioff = ioff % align;
368 	pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
369 	KASSERT(pp->pr_itemsperpage != 0);
370 
371 	/*
372 	 * Use the slack between the chunks and the page header
373 	 * for "cache coloring".
374 	 */
375 	slack = off - pp->pr_itemsperpage * pp->pr_size;
376 	pp->pr_maxcolor = (slack / align) * align;
377 	pp->pr_curcolor = 0;
378 
379 	pp->pr_nget = 0;
380 	pp->pr_nfail = 0;
381 	pp->pr_nput = 0;
382 	pp->pr_npagealloc = 0;
383 	pp->pr_npagefree = 0;
384 	pp->pr_hiwat = 0;
385 	pp->pr_nidle = 0;
386 
387 	pp->pr_ipl = -1;
388 	mtx_init(&pp->pr_mtx, IPL_NONE);
389 
390 	if (phpool.pr_size == 0) {
391 		pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
392 		    0, "phpool", NULL);
393 		pool_setipl(&phpool, IPL_HIGH);
394 	}
395 
396 	/* Insert this into the list of all pools. */
397 	TAILQ_INSERT_HEAD(&pool_head, pp, pr_poollist);
398 }
399 
400 void
401 pool_setipl(struct pool *pp, int ipl)
402 {
403 	pp->pr_ipl = ipl;
404 	mtx_init(&pp->pr_mtx, ipl);
405 }
406 
407 /*
408  * Decommission a pool resource.
409  */
410 void
411 pool_destroy(struct pool *pp)
412 {
413 	struct pool_item_header *ph;
414 
415 #ifdef DIAGNOSTIC
416 	if (pp->pr_nout != 0)
417 		panic("pool_destroy: pool busy: still out: %u", pp->pr_nout);
418 #endif
419 
420 	/* Remove all pages */
421 	while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
422 		pr_rmpage(pp, ph, NULL);
423 	KASSERT(LIST_EMPTY(&pp->pr_fullpages));
424 	KASSERT(LIST_EMPTY(&pp->pr_partpages));
425 
426 	/* Remove from global pool list */
427 	TAILQ_REMOVE(&pool_head, pp, pr_poollist);
428 }
429 
430 struct pool_item_header *
431 pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
432 {
433 	struct pool_item_header *ph;
434 
435 	if ((pp->pr_roflags & PR_PHINPAGE) != 0)
436 		ph = (struct pool_item_header *)(storage + pp->pr_phoffset);
437 	else {
438 		ph = pool_get(&phpool, flags);
439 	}
440 
441 	return (ph);
442 }
443 
444 /*
445  * Grab an item from the pool; must be called at appropriate spl level
446  */
447 void *
448 pool_get(struct pool *pp, int flags)
449 {
450 	void *v;
451 
452 	mtx_enter(&pp->pr_mtx);
453 	v = pool_do_get(pp, flags);
454 	mtx_leave(&pp->pr_mtx);
455 	if (v && pp->pr_ctor && pp->pr_ctor(pp->pr_arg, v, flags)) {
456 		mtx_enter(&pp->pr_mtx);
457 		pool_do_put(pp, v);
458 		mtx_leave(&pp->pr_mtx);
459 		v = NULL;
460 	}
461 	if (v) {
462 		pp->pr_nget++;
463 		if (flags & PR_ZERO)
464 			memset(v, 0, pp->pr_size);
465 	}
466 	return (v);
467 }
468 
469 void *
470 pool_do_get(struct pool *pp, int flags)
471 {
472 	struct pool_item *pi;
473 	struct pool_item_header *ph;
474 	void *v;
475 	int slowdown = 0;
476 #ifdef POOL_DEBUG
477 	int i, *ip;
478 #endif
479 
480 #ifdef DIAGNOSTIC
481 	if ((flags & PR_WAITOK) != 0)
482 		splassert(IPL_NONE);
483 	if (pp->pr_ipl != -1)
484 		splassert(pp->pr_ipl);
485 #endif /* DIAGNOSTIC */
486 
487 #ifdef MALLOC_DEBUG
488 	if (pp->pr_roflags & PR_DEBUG) {
489 		void *addr;
490 
491 		addr = NULL;
492 		debug_malloc(pp->pr_size, M_DEBUG,
493 		    (flags & PR_WAITOK) ? M_WAITOK : M_NOWAIT, &addr);
494 		return (addr);
495 	}
496 #endif
497 
498 startover:
499 	/*
500 	 * Check to see if we've reached the hard limit.  If we have,
501 	 * and we can wait, then wait until an item has been returned to
502 	 * the pool.
503 	 */
504 #ifdef DIAGNOSTIC
505 	if (__predict_false(pp->pr_nout > pp->pr_hardlimit))
506 		panic("pool_do_get: %s: crossed hard limit", pp->pr_wchan);
507 #endif
508 	if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
509 		if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
510 			/*
511 			 * XXX: A warning isn't logged in this case.  Should
512 			 * it be?
513 			 */
514 			pp->pr_flags |= PR_WANTED;
515 			pool_sleep(pp);
516 			goto startover;
517 		}
518 
519 		/*
520 		 * Log a message that the hard limit has been hit.
521 		 */
522 		if (pp->pr_hardlimit_warning != NULL &&
523 		    ratecheck(&pp->pr_hardlimit_warning_last,
524 		    &pp->pr_hardlimit_ratecap))
525 			log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
526 
527 		pp->pr_nfail++;
528 		return (NULL);
529 	}
530 
531 	/*
532 	 * The convention we use is that if `curpage' is not NULL, then
533 	 * it points at a non-empty bucket. In particular, `curpage'
534 	 * never points at a page header which has PR_PHINPAGE set and
535 	 * has no items in its bucket.
536 	 */
537 	if ((ph = pp->pr_curpage) == NULL) {
538 #ifdef DIAGNOSTIC
539 		if (pp->pr_nitems != 0) {
540 			printf("pool_do_get: %s: curpage NULL, nitems %u\n",
541 			    pp->pr_wchan, pp->pr_nitems);
542 			panic("pool_do_get: nitems inconsistent");
543 		}
544 #endif
545 
546 		/*
547 		 * Call the back-end page allocator for more memory.
548 		 */
549 		v = pool_allocator_alloc(pp, flags, &slowdown);
550 		if (__predict_true(v != NULL))
551 			ph = pool_alloc_item_header(pp, v, flags);
552 
553 		if (__predict_false(v == NULL || ph == NULL)) {
554 			if (v != NULL)
555 				pool_allocator_free(pp, v);
556 
557 			if ((flags & PR_WAITOK) == 0) {
558 				pp->pr_nfail++;
559 				return (NULL);
560 			}
561 
562 			/*
563 			 * Wait for items to be returned to this pool.
564 			 *
565 			 * XXX: maybe we should wake up once a second and
566 			 * try again?
567 			 */
568 			pp->pr_flags |= PR_WANTED;
569 			pool_sleep(pp);
570 			goto startover;
571 		}
572 
573 		/* We have more memory; add it to the pool */
574 		pool_prime_page(pp, v, ph);
575 		pp->pr_npagealloc++;
576 
577 		if (slowdown && (flags & PR_WAITOK)) {
578 			mtx_leave(&pp->pr_mtx);
579 			yield();
580 			mtx_enter(&pp->pr_mtx);
581 		}
582 
583 		/* Start the allocation process over. */
584 		goto startover;
585 	}
586 	if (__predict_false((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL)) {
587 		panic("pool_do_get: %s: page empty", pp->pr_wchan);
588 	}
589 #ifdef DIAGNOSTIC
590 	if (__predict_false(pp->pr_nitems == 0)) {
591 		printf("pool_do_get: %s: items on itemlist, nitems %u\n",
592 		    pp->pr_wchan, pp->pr_nitems);
593 		panic("pool_do_get: nitems inconsistent");
594 	}
595 #endif
596 
597 #ifdef DIAGNOSTIC
598 	if (__predict_false(pi->pi_magic != PI_MAGIC))
599 		panic("pool_do_get(%s): free list modified: "
600 		    "page %p; item addr %p; offset 0x%x=0x%x",
601 		    pp->pr_wchan, ph->ph_page, pi, 0, pi->pi_magic);
602 #ifdef POOL_DEBUG
603 	for (ip = (int *)pi, i = sizeof(*pi) / sizeof(int);
604 	    i < pp->pr_size / sizeof(int); i++) {
605 		if (ip[i] != PI_MAGIC) {
606 			panic("pool_do_get(%s): free list modified: "
607 			    "page %p; item addr %p; offset 0x%x=0x%x",
608 			    pp->pr_wchan, ph->ph_page, pi,
609 			    i * sizeof(int), ip[i]);
610 		}
611 	}
612 #endif /* POOL_DEBUG */
613 #endif /* DIAGNOSTIC */
614 
615 	/*
616 	 * Remove from item list.
617 	 */
618 	TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
619 	pp->pr_nitems--;
620 	pp->pr_nout++;
621 	if (ph->ph_nmissing == 0) {
622 #ifdef DIAGNOSTIC
623 		if (__predict_false(pp->pr_nidle == 0))
624 			panic("pool_do_get: nidle inconsistent");
625 #endif
626 		pp->pr_nidle--;
627 
628 		/*
629 		 * This page was previously empty.  Move it to the list of
630 		 * partially-full pages.  This page is already curpage.
631 		 */
632 		LIST_REMOVE(ph, ph_pagelist);
633 		LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
634 	}
635 	ph->ph_nmissing++;
636 	if (TAILQ_EMPTY(&ph->ph_itemlist)) {
637 #ifdef DIAGNOSTIC
638 		if (__predict_false(ph->ph_nmissing != pp->pr_itemsperpage)) {
639 			panic("pool_do_get: %s: nmissing inconsistent",
640 			    pp->pr_wchan);
641 		}
642 #endif
643 		/*
644 		 * This page is now full.  Move it to the full list
645 		 * and select a new current page.
646 		 */
647 		LIST_REMOVE(ph, ph_pagelist);
648 		LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
649 		pool_update_curpage(pp);
650 	}
651 
652 	/*
653 	 * If we have a low water mark and we are now below that low
654 	 * water mark, add more items to the pool.
655 	 */
656 	if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
657 		/*
658 		 * XXX: Should we log a warning?  Should we set up a timeout
659 		 * to try again in a second or so?  The latter could break
660 		 * a caller's assumptions about interrupt protection, etc.
661 		 */
662 	}
663 	return (v);
664 }
665 
666 /*
667  * Return resource to the pool; must be called at appropriate spl level
668  */
669 void
670 pool_put(struct pool *pp, void *v)
671 {
672 	if (pp->pr_dtor)
673 		pp->pr_dtor(pp->pr_arg, v);
674 	mtx_enter(&pp->pr_mtx);
675 	pool_do_put(pp, v);
676 	mtx_leave(&pp->pr_mtx);
677 	pp->pr_nput++;
678 }
679 
680 /*
681  * Internal version of pool_put().
682  */
683 void
684 pool_do_put(struct pool *pp, void *v)
685 {
686 	struct pool_item *pi = v;
687 	struct pool_item_header *ph;
688 #ifdef POOL_DEBUG
689 	int i, *ip;
690 #endif
691 
692 	if (v == NULL)
693 		panic("pool_put of NULL");
694 
695 #ifdef MALLOC_DEBUG
696 	if (pp->pr_roflags & PR_DEBUG) {
697 		debug_free(v, M_DEBUG);
698 		return;
699 	}
700 #endif
701 
702 #ifdef DIAGNOSTIC
703 	if (pp->pr_ipl != -1)
704 		splassert(pp->pr_ipl);
705 
706 	if (__predict_false(pp->pr_nout == 0)) {
707 		printf("pool %s: putting with none out\n",
708 		    pp->pr_wchan);
709 		panic("pool_do_put");
710 	}
711 #endif
712 
713 	if (__predict_false((ph = pr_find_pagehead(pp, v)) == NULL)) {
714 		panic("pool_do_put: %s: page header missing", pp->pr_wchan);
715 	}
716 
717 	/*
718 	 * Return to item list.
719 	 */
720 #ifdef DIAGNOSTIC
721 	pi->pi_magic = PI_MAGIC;
722 #ifdef POOL_DEBUG
723 	for (ip = (int *)pi, i = sizeof(*pi)/sizeof(int);
724 	    i < pp->pr_size / sizeof(int); i++)
725 		ip[i] = PI_MAGIC;
726 #endif /* POOL_DEBUG */
727 #endif /* DIAGNOSTIC */
728 
729 	TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
730 	ph->ph_nmissing--;
731 	pp->pr_nitems++;
732 	pp->pr_nout--;
733 
734 	/* Cancel "pool empty" condition if it exists */
735 	if (pp->pr_curpage == NULL)
736 		pp->pr_curpage = ph;
737 
738 	if (pp->pr_flags & PR_WANTED) {
739 		pp->pr_flags &= ~PR_WANTED;
740 		if (ph->ph_nmissing == 0)
741 			pp->pr_nidle++;
742 		wakeup(pp);
743 		return;
744 	}
745 
746 	/*
747 	 * If this page is now empty, do one of two things:
748 	 *
749 	 *	(1) If we have more pages than the page high water mark,
750 	 *	    free the page back to the system.
751 	 *
752 	 *	(2) Otherwise, move the page to the empty page list.
753 	 *
754 	 * Either way, select a new current page (so we use a partially-full
755 	 * page if one is available).
756 	 */
757 	if (ph->ph_nmissing == 0) {
758 		pp->pr_nidle++;
759 		if (pp->pr_nidle > pp->pr_maxpages) {
760 			pr_rmpage(pp, ph, NULL);
761 		} else {
762 			LIST_REMOVE(ph, ph_pagelist);
763 			LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
764 		}
765 		pool_update_curpage(pp);
766 	}
767 
768 	/*
769 	 * If the page was previously completely full, move it to the
770 	 * partially-full list and make it the current page.  The next
771 	 * allocation will get the item from this page, instead of
772 	 * further fragmenting the pool.
773 	 */
774 	else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
775 		LIST_REMOVE(ph, ph_pagelist);
776 		LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
777 		pp->pr_curpage = ph;
778 	}
779 }
780 
781 /*
782  * Add N items to the pool.
783  */
784 int
785 pool_prime(struct pool *pp, int n)
786 {
787 	struct pool_item_header *ph;
788 	caddr_t cp;
789 	int newpages;
790 	int slowdown;
791 
792 	mtx_enter(&pp->pr_mtx);
793 	newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
794 
795 	while (newpages-- > 0) {
796 		cp = pool_allocator_alloc(pp, PR_NOWAIT, &slowdown);
797 		if (__predict_true(cp != NULL))
798 			ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
799 		if (__predict_false(cp == NULL || ph == NULL)) {
800 			if (cp != NULL)
801 				pool_allocator_free(pp, cp);
802 			break;
803 		}
804 
805 		pool_prime_page(pp, cp, ph);
806 		pp->pr_npagealloc++;
807 		pp->pr_minpages++;
808 	}
809 
810 	if (pp->pr_minpages >= pp->pr_maxpages)
811 		pp->pr_maxpages = pp->pr_minpages + 1;	/* XXX */
812 
813 	mtx_leave(&pp->pr_mtx);
814 	return (0);
815 }
816 
817 /*
818  * Add a page worth of items to the pool.
819  *
820  * Note, we must be called with the pool descriptor LOCKED.
821  */
822 void
823 pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
824 {
825 	struct pool_item *pi;
826 	caddr_t cp = storage;
827 	unsigned int align = pp->pr_align;
828 	unsigned int ioff = pp->pr_itemoffset;
829 	int n;
830 #ifdef POOL_DEBUG
831 	int i, *ip;
832 #endif
833 
834 	/*
835 	 * Insert page header.
836 	 */
837 	LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
838 	TAILQ_INIT(&ph->ph_itemlist);
839 	ph->ph_page = storage;
840 	ph->ph_pagesize = pp->pr_alloc->pa_pagesz;
841 	ph->ph_nmissing = 0;
842 	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
843 		SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
844 
845 	pp->pr_nidle++;
846 
847 	/*
848 	 * Color this page.
849 	 */
850 	cp = (caddr_t)(cp + pp->pr_curcolor);
851 	if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
852 		pp->pr_curcolor = 0;
853 
854 	/*
855 	 * Adjust storage to apply aligment to `pr_itemoffset' in each item.
856 	 */
857 	if (ioff != 0)
858 		cp = (caddr_t)(cp + (align - ioff));
859 	ph->ph_colored = cp;
860 
861 	/*
862 	 * Insert remaining chunks on the bucket list.
863 	 */
864 	n = pp->pr_itemsperpage;
865 	pp->pr_nitems += n;
866 
867 	while (n--) {
868 		pi = (struct pool_item *)cp;
869 
870 		KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
871 
872 		/* Insert on page list */
873 		TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
874 
875 #ifdef DIAGNOSTIC
876 		pi->pi_magic = PI_MAGIC;
877 #ifdef POOL_DEBUG
878 		for (ip = (int *)pi, i = sizeof(*pi)/sizeof(int);
879 		    i < pp->pr_size / sizeof(int); i++)
880 			ip[i] = PI_MAGIC;
881 #endif /* POOL_DEBUG */
882 #endif /* DIAGNOSTIC */
883 		cp = (caddr_t)(cp + pp->pr_size);
884 	}
885 
886 	/*
887 	 * If the pool was depleted, point at the new page.
888 	 */
889 	if (pp->pr_curpage == NULL)
890 		pp->pr_curpage = ph;
891 
892 	if (++pp->pr_npages > pp->pr_hiwat)
893 		pp->pr_hiwat = pp->pr_npages;
894 }
895 
896 /*
897  * Used by pool_get() when nitems drops below the low water mark.  This
898  * is used to catch up pr_nitems with the low water mark.
899  *
900  * Note we never wait for memory here, we let the caller decide what to do.
901  */
902 int
903 pool_catchup(struct pool *pp)
904 {
905 	struct pool_item_header *ph;
906 	caddr_t cp;
907 	int error = 0;
908 	int slowdown;
909 
910 	while (POOL_NEEDS_CATCHUP(pp)) {
911 		/*
912 		 * Call the page back-end allocator for more memory.
913 		 */
914 		cp = pool_allocator_alloc(pp, PR_NOWAIT, &slowdown);
915 		if (__predict_true(cp != NULL))
916 			ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
917 		if (__predict_false(cp == NULL || ph == NULL)) {
918 			if (cp != NULL)
919 				pool_allocator_free(pp, cp);
920 			error = ENOMEM;
921 			break;
922 		}
923 		pool_prime_page(pp, cp, ph);
924 		pp->pr_npagealloc++;
925 	}
926 
927 	return (error);
928 }
929 
930 void
931 pool_update_curpage(struct pool *pp)
932 {
933 
934 	pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
935 	if (pp->pr_curpage == NULL) {
936 		pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
937 	}
938 }
939 
940 void
941 pool_setlowat(struct pool *pp, int n)
942 {
943 
944 	pp->pr_minitems = n;
945 	pp->pr_minpages = (n == 0)
946 		? 0
947 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
948 
949 	mtx_enter(&pp->pr_mtx);
950 	/* Make sure we're caught up with the newly-set low water mark. */
951 	if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
952 		/*
953 		 * XXX: Should we log a warning?  Should we set up a timeout
954 		 * to try again in a second or so?  The latter could break
955 		 * a caller's assumptions about interrupt protection, etc.
956 		 */
957 	}
958 	mtx_leave(&pp->pr_mtx);
959 }
960 
961 void
962 pool_sethiwat(struct pool *pp, int n)
963 {
964 
965 	pp->pr_maxpages = (n == 0)
966 		? 0
967 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
968 }
969 
970 int
971 pool_sethardlimit(struct pool *pp, u_int n, const char *warnmsg, int ratecap)
972 {
973 	int error = 0;
974 
975 	if (n < pp->pr_nout) {
976 		error = EINVAL;
977 		goto done;
978 	}
979 
980 	pp->pr_hardlimit = n;
981 	pp->pr_hardlimit_warning = warnmsg;
982 	pp->pr_hardlimit_ratecap.tv_sec = ratecap;
983 	pp->pr_hardlimit_warning_last.tv_sec = 0;
984 	pp->pr_hardlimit_warning_last.tv_usec = 0;
985 
986 	/*
987 	 * In-line version of pool_sethiwat().
988 	 */
989 	pp->pr_maxpages = (n == 0 || n == UINT_MAX)
990 		? n
991 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
992 
993 done:
994 	return (error);
995 }
996 
997 void
998 pool_set_ctordtor(struct pool *pp, int (*ctor)(void *, void *, int),
999     void (*dtor)(void *, void *), void *arg)
1000 {
1001 	pp->pr_ctor = ctor;
1002 	pp->pr_dtor = dtor;
1003 	pp->pr_arg = arg;
1004 }
1005 /*
1006  * Release all complete pages that have not been used recently.
1007  *
1008  * Returns non-zero if any pages have been reclaimed.
1009  */
1010 int
1011 pool_reclaim(struct pool *pp)
1012 {
1013 	struct pool_item_header *ph, *phnext;
1014 	struct pool_pagelist pq;
1015 
1016 	LIST_INIT(&pq);
1017 
1018 	mtx_enter(&pp->pr_mtx);
1019 	for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1020 		phnext = LIST_NEXT(ph, ph_pagelist);
1021 
1022 		/* Check our minimum page claim */
1023 		if (pp->pr_npages <= pp->pr_minpages)
1024 			break;
1025 
1026 		KASSERT(ph->ph_nmissing == 0);
1027 
1028 		/*
1029 		 * If freeing this page would put us below
1030 		 * the low water mark, stop now.
1031 		 */
1032 		if ((pp->pr_nitems - pp->pr_itemsperpage) <
1033 		    pp->pr_minitems)
1034 			break;
1035 
1036 		pr_rmpage(pp, ph, &pq);
1037 	}
1038 	mtx_leave(&pp->pr_mtx);
1039 
1040 	if (LIST_EMPTY(&pq))
1041 		return (0);
1042 	while ((ph = LIST_FIRST(&pq)) != NULL) {
1043 		LIST_REMOVE(ph, ph_pagelist);
1044 		pool_allocator_free(pp, ph->ph_page);
1045 		if (pp->pr_roflags & PR_PHINPAGE)
1046 			continue;
1047 		pool_put(&phpool, ph);
1048 	}
1049 
1050 	return (1);
1051 }
1052 
1053 #ifdef DDB
1054 #include <machine/db_machdep.h>
1055 #include <ddb/db_interface.h>
1056 #include <ddb/db_output.h>
1057 
1058 /*
1059  * Diagnostic helpers.
1060  */
1061 void
1062 pool_printit(struct pool *pp, const char *modif, int (*pr)(const char *, ...))
1063 {
1064 	pool_print1(pp, modif, pr);
1065 }
1066 
1067 void
1068 pool_print_pagelist(struct pool_pagelist *pl, int (*pr)(const char *, ...))
1069 {
1070 	struct pool_item_header *ph;
1071 #ifdef DIAGNOSTIC
1072 	struct pool_item *pi;
1073 #endif
1074 
1075 	LIST_FOREACH(ph, pl, ph_pagelist) {
1076 		(*pr)("\t\tpage %p, nmissing %d\n",
1077 		    ph->ph_page, ph->ph_nmissing);
1078 #ifdef DIAGNOSTIC
1079 		TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1080 			if (pi->pi_magic != PI_MAGIC) {
1081 				(*pr)("\t\t\titem %p, magic 0x%x\n",
1082 				    pi, pi->pi_magic);
1083 			}
1084 		}
1085 #endif
1086 	}
1087 }
1088 
1089 void
1090 pool_print1(struct pool *pp, const char *modif, int (*pr)(const char *, ...))
1091 {
1092 	struct pool_item_header *ph;
1093 	int print_pagelist = 0;
1094 	char c;
1095 
1096 	while ((c = *modif++) != '\0') {
1097 		if (c == 'p')
1098 			print_pagelist = 1;
1099 		modif++;
1100 	}
1101 
1102 	(*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1103 	    pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1104 	    pp->pr_roflags);
1105 	(*pr)("\talloc %p\n", pp->pr_alloc);
1106 	(*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1107 	    pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1108 	(*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1109 	    pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1110 
1111 	(*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1112 	    pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1113 	(*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1114 	    pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1115 
1116 	if (print_pagelist == 0)
1117 		return;
1118 
1119 	if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1120 		(*pr)("\n\tempty page list:\n");
1121 	pool_print_pagelist(&pp->pr_emptypages, pr);
1122 	if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1123 		(*pr)("\n\tfull page list:\n");
1124 	pool_print_pagelist(&pp->pr_fullpages, pr);
1125 	if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1126 		(*pr)("\n\tpartial-page list:\n");
1127 	pool_print_pagelist(&pp->pr_partpages, pr);
1128 
1129 	if (pp->pr_curpage == NULL)
1130 		(*pr)("\tno current page\n");
1131 	else
1132 		(*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1133 }
1134 
1135 void
1136 db_show_all_pools(db_expr_t expr, int haddr, db_expr_t count, char *modif)
1137 {
1138 	struct pool *pp;
1139 	char maxp[16];
1140 	int ovflw;
1141 	char mode;
1142 
1143 	mode = modif[0];
1144 	if (mode != '\0' && mode != 'a') {
1145 		db_printf("usage: show all pools [/a]\n");
1146 		return;
1147 	}
1148 
1149 	if (mode == '\0')
1150 		db_printf("%-10s%4s%9s%5s%9s%6s%6s%6s%6s%6s%6s%5s\n",
1151 		    "Name",
1152 		    "Size",
1153 		    "Requests",
1154 		    "Fail",
1155 		    "Releases",
1156 		    "Pgreq",
1157 		    "Pgrel",
1158 		    "Npage",
1159 		    "Hiwat",
1160 		    "Minpg",
1161 		    "Maxpg",
1162 		    "Idle");
1163 	else
1164 		db_printf("%-10s %18s %18s\n",
1165 		    "Name", "Address", "Allocator");
1166 
1167 	TAILQ_FOREACH(pp, &pool_head, pr_poollist) {
1168 		if (mode == 'a') {
1169 			db_printf("%-10s %18p %18p\n", pp->pr_wchan, pp,
1170 			    pp->pr_alloc);
1171 			continue;
1172 		}
1173 
1174 		if (!pp->pr_nget)
1175 			continue;
1176 
1177 		if (pp->pr_maxpages == UINT_MAX)
1178 			snprintf(maxp, sizeof maxp, "inf");
1179 		else
1180 			snprintf(maxp, sizeof maxp, "%u", pp->pr_maxpages);
1181 
1182 #define PRWORD(ovflw, fmt, width, fixed, val) do {	\
1183 	(ovflw) += db_printf((fmt),			\
1184 	    (width) - (fixed) - (ovflw) > 0 ?		\
1185 	    (width) - (fixed) - (ovflw) : 0,		\
1186 	    (val)) - (width);				\
1187 	if ((ovflw) < 0)				\
1188 		(ovflw) = 0;				\
1189 } while (/* CONSTCOND */0)
1190 
1191 		ovflw = 0;
1192 		PRWORD(ovflw, "%-*s", 10, 0, pp->pr_wchan);
1193 		PRWORD(ovflw, " %*u", 4, 1, pp->pr_size);
1194 		PRWORD(ovflw, " %*lu", 9, 1, pp->pr_nget);
1195 		PRWORD(ovflw, " %*lu", 5, 1, pp->pr_nfail);
1196 		PRWORD(ovflw, " %*lu", 9, 1, pp->pr_nput);
1197 		PRWORD(ovflw, " %*lu", 6, 1, pp->pr_npagealloc);
1198 		PRWORD(ovflw, " %*lu", 6, 1, pp->pr_npagefree);
1199 		PRWORD(ovflw, " %*d", 6, 1, pp->pr_npages);
1200 		PRWORD(ovflw, " %*d", 6, 1, pp->pr_hiwat);
1201 		PRWORD(ovflw, " %*d", 6, 1, pp->pr_minpages);
1202 		PRWORD(ovflw, " %*s", 6, 1, maxp);
1203 		PRWORD(ovflw, " %*lu\n", 5, 1, pp->pr_nidle);
1204 
1205 		pool_chk(pp, pp->pr_wchan);
1206 	}
1207 }
1208 
1209 int
1210 pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1211 {
1212 	struct pool_item *pi;
1213 	caddr_t page;
1214 	int n;
1215 #ifdef POOL_DEBUG
1216 	int i, *ip;
1217 #endif
1218 
1219 	page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask);
1220 	if (page != ph->ph_page &&
1221 	    (pp->pr_roflags & PR_PHINPAGE) != 0) {
1222 		if (label != NULL)
1223 			printf("%s: ", label);
1224 		printf("pool(%p:%s): page inconsistency: page %p; "
1225 		    "at page head addr %p (p %p)\n",
1226 		    pp, pp->pr_wchan, ph->ph_page, ph, page);
1227 		return 1;
1228 	}
1229 
1230 	for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
1231 	     pi != NULL;
1232 	     pi = TAILQ_NEXT(pi,pi_list), n++) {
1233 
1234 #ifdef DIAGNOSTIC
1235 		if (pi->pi_magic != PI_MAGIC) {
1236 			if (label != NULL)
1237 				printf("%s: ", label);
1238 			printf("pool(%s): free list modified: "
1239 			    "page %p; item ordinal %d; addr %p "
1240 			    "(p %p); offset 0x%x=0x%x\n",
1241 			    pp->pr_wchan, ph->ph_page, n, pi, page,
1242 			    0, pi->pi_magic);
1243 		}
1244 #ifdef POOL_DEBUG
1245 		for (ip = (int *)pi, i = sizeof(*pi) / sizeof(int);
1246 		    i < pp->pr_size / sizeof(int); i++) {
1247 			if (ip[i] != PI_MAGIC) {
1248 				printf("pool(%s): free list modified: "
1249 				    "page %p; item ordinal %d; addr %p "
1250 				    "(p %p); offset 0x%x=0x%x\n",
1251 				    pp->pr_wchan, ph->ph_page, n, pi,
1252 				    page, i * sizeof(int), ip[i]);
1253 			}
1254 		}
1255 
1256 #endif /* POOL_DEBUG */
1257 #endif /* DIAGNOSTIC */
1258 		page =
1259 		    (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
1260 		if (page == ph->ph_page)
1261 			continue;
1262 
1263 		if (label != NULL)
1264 			printf("%s: ", label);
1265 		printf("pool(%p:%s): page inconsistency: page %p;"
1266 		    " item ordinal %d; addr %p (p %p)\n", pp,
1267 		    pp->pr_wchan, ph->ph_page, n, pi, page);
1268 		return 1;
1269 	}
1270 	return 0;
1271 }
1272 
1273 int
1274 pool_chk(struct pool *pp, const char *label)
1275 {
1276 	struct pool_item_header *ph;
1277 	int r = 0;
1278 
1279 	LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist)
1280 		r += pool_chk_page(pp, label, ph);
1281 	LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist)
1282 		r += pool_chk_page(pp, label, ph);
1283 	LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist)
1284 		r += pool_chk_page(pp, label, ph);
1285 
1286 	return (r);
1287 }
1288 
1289 void
1290 pool_walk(struct pool *pp, void (*func)(void *))
1291 {
1292 	struct pool_item_header *ph;
1293 	struct pool_item *pi;
1294 	caddr_t cp;
1295 	int n;
1296 
1297 	LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
1298 		cp = ph->ph_colored;
1299 		n = ph->ph_nmissing;
1300 
1301 		while (n--) {
1302 			func(cp);
1303 			cp += pp->pr_size;
1304 		}
1305 	}
1306 
1307 	LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
1308 		cp = ph->ph_colored;
1309 		n = ph->ph_nmissing;
1310 
1311 		do {
1312 			TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1313 				if (cp == (caddr_t)pi)
1314 					break;
1315 			}
1316 			if (cp != (caddr_t)pi) {
1317 				func(cp);
1318 				n--;
1319 			}
1320 
1321 			cp += pp->pr_size;
1322 		} while (n > 0);
1323 	}
1324 }
1325 #endif
1326 
1327 /*
1328  * We have three different sysctls.
1329  * kern.pool.npools - the number of pools.
1330  * kern.pool.pool.<pool#> - the pool struct for the pool#.
1331  * kern.pool.name.<pool#> - the name for pool#.
1332  */
1333 int
1334 sysctl_dopool(int *name, u_int namelen, char *where, size_t *sizep)
1335 {
1336 	struct pool *pp, *foundpool = NULL;
1337 	size_t buflen = where != NULL ? *sizep : 0;
1338 	int npools = 0, s;
1339 	unsigned int lookfor;
1340 	size_t len;
1341 
1342 	switch (*name) {
1343 	case KERN_POOL_NPOOLS:
1344 		if (namelen != 1 || buflen != sizeof(int))
1345 			return (EINVAL);
1346 		lookfor = 0;
1347 		break;
1348 	case KERN_POOL_NAME:
1349 		if (namelen != 2 || buflen < 1)
1350 			return (EINVAL);
1351 		lookfor = name[1];
1352 		break;
1353 	case KERN_POOL_POOL:
1354 		if (namelen != 2 || buflen != sizeof(struct pool))
1355 			return (EINVAL);
1356 		lookfor = name[1];
1357 		break;
1358 	default:
1359 		return (EINVAL);
1360 	}
1361 
1362 	s = splvm();
1363 
1364 	TAILQ_FOREACH(pp, &pool_head, pr_poollist) {
1365 		npools++;
1366 		if (lookfor == pp->pr_serial) {
1367 			foundpool = pp;
1368 			break;
1369 		}
1370 	}
1371 
1372 	splx(s);
1373 
1374 	if (*name != KERN_POOL_NPOOLS && foundpool == NULL)
1375 		return (ENOENT);
1376 
1377 	switch (*name) {
1378 	case KERN_POOL_NPOOLS:
1379 		return copyout(&npools, where, buflen);
1380 	case KERN_POOL_NAME:
1381 		len = strlen(foundpool->pr_wchan) + 1;
1382 		if (*sizep < len)
1383 			return (ENOMEM);
1384 		*sizep = len;
1385 		return copyout(foundpool->pr_wchan, where, len);
1386 	case KERN_POOL_POOL:
1387 		return copyout(foundpool, where, buflen);
1388 	}
1389 	/* NOTREACHED */
1390 	return (0); /* XXX - Stupid gcc */
1391 }
1392 
1393 /*
1394  * Pool backend allocators.
1395  *
1396  * Each pool has a backend allocator that handles allocation, deallocation
1397  */
1398 void	*pool_page_alloc(struct pool *, int, int *);
1399 void	pool_page_free(struct pool *, void *);
1400 
1401 /*
1402  * safe for interrupts, name preserved for compat this is the default
1403  * allocator
1404  */
1405 struct pool_allocator pool_allocator_nointr = {
1406 	pool_page_alloc, pool_page_free, 0,
1407 };
1408 
1409 /*
1410  * XXX - we have at least three different resources for the same allocation
1411  *  and each resource can be depleted. First we have the ready elements in
1412  *  the pool. Then we have the resource (typically a vm_map) for this
1413  *  allocator, then we have physical memory. Waiting for any of these can
1414  *  be unnecessary when any other is freed, but the kernel doesn't support
1415  *  sleeping on multiple addresses, so we have to fake. The caller sleeps on
1416  *  the pool (so that we can be awakened when an item is returned to the pool),
1417  *  but we set PA_WANT on the allocator. When a page is returned to
1418  *  the allocator and PA_WANT is set pool_allocator_free will wakeup all
1419  *  sleeping pools belonging to this allocator. (XXX - thundering herd).
1420  *  We also wake up the allocator in case someone without a pool (malloc)
1421  *  is sleeping waiting for this allocator.
1422  */
1423 
1424 void *
1425 pool_allocator_alloc(struct pool *pp, int flags, int *slowdown)
1426 {
1427 	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
1428 	void *v;
1429 
1430 	if (waitok)
1431 		mtx_leave(&pp->pr_mtx);
1432 	v = pp->pr_alloc->pa_alloc(pp, flags, slowdown);
1433 	if (waitok)
1434 		mtx_enter(&pp->pr_mtx);
1435 
1436 	return (v);
1437 }
1438 
1439 void
1440 pool_allocator_free(struct pool *pp, void *v)
1441 {
1442 	struct pool_allocator *pa = pp->pr_alloc;
1443 
1444 	(*pa->pa_free)(pp, v);
1445 }
1446 
1447 void *
1448 pool_page_alloc(struct pool *pp, int flags, int *slowdown)
1449 {
1450 	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
1451 
1452 	return (uvm_km_getpage(waitok, slowdown));
1453 }
1454 
1455 void
1456 pool_page_free(struct pool *pp, void *v)
1457 {
1458 
1459 	uvm_km_putpage(v);
1460 }
1461 
1462 void *
1463 pool_large_alloc(struct pool *pp, int flags, int *slowdown)
1464 {
1465 	int kfl = (flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT;
1466 	vaddr_t va;
1467 	int s;
1468 
1469 	s = splvm();
1470 	va = uvm_km_kmemalloc(kmem_map, NULL, pp->pr_alloc->pa_pagesz, kfl);
1471 	splx(s);
1472 
1473 	return ((void *)va);
1474 }
1475 
1476 void
1477 pool_large_free(struct pool *pp, void *v)
1478 {
1479 	int s;
1480 
1481 	s = splvm();
1482 	uvm_km_free(kmem_map, (vaddr_t)v, pp->pr_alloc->pa_pagesz);
1483 	splx(s);
1484 }
1485 
1486 void *
1487 pool_large_alloc_ni(struct pool *pp, int flags, int *slowdown)
1488 {
1489 	int kfl = (flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT;
1490 
1491 	return ((void *)uvm_km_kmemalloc(kernel_map, uvm.kernel_object,
1492 	    pp->pr_alloc->pa_pagesz, kfl));
1493 }
1494 
1495 void
1496 pool_large_free_ni(struct pool *pp, void *v)
1497 {
1498 	uvm_km_free(kernel_map, (vaddr_t)v, pp->pr_alloc->pa_pagesz);
1499 }
1500