xref: /openbsd-src/sys/kern/subr_pool.c (revision 8e0c768258d4632c51876b4397034bc3152bf8db)
1 /*	$OpenBSD: subr_pool.c,v 1.228 2019/07/19 09:03:03 bluhm 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/errno.h>
37 #include <sys/kernel.h>
38 #include <sys/malloc.h>
39 #include <sys/pool.h>
40 #include <sys/proc.h>
41 #include <sys/syslog.h>
42 #include <sys/sysctl.h>
43 #include <sys/task.h>
44 #include <sys/timeout.h>
45 #include <sys/percpu.h>
46 
47 #include <uvm/uvm_extern.h>
48 
49 /*
50  * Pool resource management utility.
51  *
52  * Memory is allocated in pages which are split into pieces according to
53  * the pool item size. Each page is kept on one of three lists in the
54  * pool structure: `pr_emptypages', `pr_fullpages' and `pr_partpages',
55  * for empty, full and partially-full pages respectively. The individual
56  * pool items are on a linked list headed by `ph_items' in each page
57  * header. The memory for building the page list is either taken from
58  * the allocated pages themselves (for small pool items) or taken from
59  * an internal pool of page headers (`phpool').
60  */
61 
62 /* List of all pools */
63 SIMPLEQ_HEAD(,pool) pool_head = SIMPLEQ_HEAD_INITIALIZER(pool_head);
64 
65 /*
66  * Every pool gets a unique serial number assigned to it. If this counter
67  * wraps, we're screwed, but we shouldn't create so many pools anyway.
68  */
69 unsigned int pool_serial;
70 unsigned int pool_count;
71 
72 /* Lock the previous variables making up the global pool state */
73 struct rwlock pool_lock = RWLOCK_INITIALIZER("pools");
74 
75 /* Private pool for page header structures */
76 struct pool phpool;
77 
78 struct pool_lock_ops {
79 	void	(*pl_init)(struct pool *, union pool_lock *,
80 		    const struct lock_type *);
81 	void	(*pl_enter)(union pool_lock *);
82 	int	(*pl_enter_try)(union pool_lock *);
83 	void	(*pl_leave)(union pool_lock *);
84 	void	(*pl_assert_locked)(union pool_lock *);
85 	void	(*pl_assert_unlocked)(union pool_lock *);
86 	int	(*pl_sleep)(void *, union pool_lock *, int, const char *, int);
87 };
88 
89 static const struct pool_lock_ops pool_lock_ops_mtx;
90 static const struct pool_lock_ops pool_lock_ops_rw;
91 
92 #ifdef WITNESS
93 #define pl_init(pp, pl) do {						\
94 	static const struct lock_type __lock_type = { .lt_name = #pl };	\
95 	(pp)->pr_lock_ops->pl_init(pp, pl, &__lock_type);		\
96 } while (0)
97 #else /* WITNESS */
98 #define pl_init(pp, pl)		(pp)->pr_lock_ops->pl_init(pp, pl, NULL)
99 #endif /* WITNESS */
100 
101 static inline void
102 pl_enter(struct pool *pp, union pool_lock *pl)
103 {
104 	pp->pr_lock_ops->pl_enter(pl);
105 }
106 static inline int
107 pl_enter_try(struct pool *pp, union pool_lock *pl)
108 {
109 	return pp->pr_lock_ops->pl_enter_try(pl);
110 }
111 static inline void
112 pl_leave(struct pool *pp, union pool_lock *pl)
113 {
114 	pp->pr_lock_ops->pl_leave(pl);
115 }
116 static inline void
117 pl_assert_locked(struct pool *pp, union pool_lock *pl)
118 {
119 	pp->pr_lock_ops->pl_assert_locked(pl);
120 }
121 static inline void
122 pl_assert_unlocked(struct pool *pp, union pool_lock *pl)
123 {
124 	pp->pr_lock_ops->pl_assert_unlocked(pl);
125 }
126 static inline int
127 pl_sleep(struct pool *pp, void *ident, union pool_lock *lock, int priority,
128     const char *wmesg, int timo)
129 {
130 	return pp->pr_lock_ops->pl_sleep(ident, lock, priority, wmesg, timo);
131 }
132 
133 struct pool_item {
134 	u_long				pi_magic;
135 	XSIMPLEQ_ENTRY(pool_item)	pi_list;
136 };
137 #define POOL_IMAGIC(ph, pi) ((u_long)(pi) ^ (ph)->ph_magic)
138 
139 struct pool_page_header {
140 	/* Page headers */
141 	TAILQ_ENTRY(pool_page_header)
142 				ph_entry;	/* pool page list */
143 	XSIMPLEQ_HEAD(, pool_item)
144 				ph_items;	/* free items on the page */
145 	RBT_ENTRY(pool_page_header)
146 				ph_node;	/* off-page page headers */
147 	unsigned int		ph_nmissing;	/* # of chunks in use */
148 	caddr_t			ph_page;	/* this page's address */
149 	caddr_t			ph_colored;	/* page's colored address */
150 	unsigned long		ph_magic;
151 	int			ph_tick;
152 };
153 #define POOL_MAGICBIT (1 << 3) /* keep away from perturbed low bits */
154 #define POOL_PHPOISON(ph) ISSET((ph)->ph_magic, POOL_MAGICBIT)
155 
156 #ifdef MULTIPROCESSOR
157 struct pool_cache_item {
158 	struct pool_cache_item	*ci_next;	/* next item in list */
159 	unsigned long		 ci_nitems;	/* number of items in list */
160 	TAILQ_ENTRY(pool_cache_item)
161 				 ci_nextl;	/* entry in list of lists */
162 };
163 
164 /* we store whether the cached item is poisoned in the high bit of nitems */
165 #define POOL_CACHE_ITEM_NITEMS_MASK	0x7ffffffUL
166 #define POOL_CACHE_ITEM_NITEMS_POISON	0x8000000UL
167 
168 #define POOL_CACHE_ITEM_NITEMS(_ci)					\
169     ((_ci)->ci_nitems & POOL_CACHE_ITEM_NITEMS_MASK)
170 
171 #define POOL_CACHE_ITEM_POISONED(_ci)					\
172     ISSET((_ci)->ci_nitems, POOL_CACHE_ITEM_NITEMS_POISON)
173 
174 struct pool_cache {
175 	struct pool_cache_item	*pc_actv;	/* active list of items */
176 	unsigned long		 pc_nactv;	/* actv head nitems cache */
177 	struct pool_cache_item	*pc_prev;	/* previous list of items */
178 
179 	uint64_t		 pc_gen;	/* generation number */
180 	uint64_t		 pc_nget;	/* # of successful requests */
181 	uint64_t		 pc_nfail;	/* # of unsuccessful reqs */
182 	uint64_t		 pc_nput;	/* # of releases */
183 	uint64_t		 pc_nlget;	/* # of list requests */
184 	uint64_t		 pc_nlfail;	/* # of fails getting a list */
185 	uint64_t		 pc_nlput;	/* # of list releases */
186 
187 	int			 pc_nout;
188 };
189 
190 void	*pool_cache_get(struct pool *);
191 void	 pool_cache_put(struct pool *, void *);
192 void	 pool_cache_destroy(struct pool *);
193 void	 pool_cache_gc(struct pool *);
194 #endif
195 void	 pool_cache_pool_info(struct pool *, struct kinfo_pool *);
196 int	 pool_cache_info(struct pool *, void *, size_t *);
197 int	 pool_cache_cpus_info(struct pool *, void *, size_t *);
198 
199 #ifdef POOL_DEBUG
200 int	pool_debug = 1;
201 #else
202 int	pool_debug = 0;
203 #endif
204 
205 #define POOL_INPGHDR(pp) ((pp)->pr_phoffset != 0)
206 
207 struct pool_page_header *
208 	 pool_p_alloc(struct pool *, int, int *);
209 void	 pool_p_insert(struct pool *, struct pool_page_header *);
210 void	 pool_p_remove(struct pool *, struct pool_page_header *);
211 void	 pool_p_free(struct pool *, struct pool_page_header *);
212 
213 void	 pool_update_curpage(struct pool *);
214 void	*pool_do_get(struct pool *, int, int *);
215 void	 pool_do_put(struct pool *, void *);
216 int	 pool_chk_page(struct pool *, struct pool_page_header *, int);
217 int	 pool_chk(struct pool *);
218 void	 pool_get_done(struct pool *, void *, void *);
219 void	 pool_runqueue(struct pool *, int);
220 
221 void	*pool_allocator_alloc(struct pool *, int, int *);
222 void	 pool_allocator_free(struct pool *, void *);
223 
224 /*
225  * The default pool allocator.
226  */
227 void	*pool_page_alloc(struct pool *, int, int *);
228 void	pool_page_free(struct pool *, void *);
229 
230 /*
231  * safe for interrupts; this is the default allocator
232  */
233 struct pool_allocator pool_allocator_single = {
234 	pool_page_alloc,
235 	pool_page_free,
236 	POOL_ALLOC_SIZE(PAGE_SIZE, POOL_ALLOC_ALIGNED)
237 };
238 
239 void	*pool_multi_alloc(struct pool *, int, int *);
240 void	pool_multi_free(struct pool *, void *);
241 
242 struct pool_allocator pool_allocator_multi = {
243 	pool_multi_alloc,
244 	pool_multi_free,
245 	POOL_ALLOC_SIZES(PAGE_SIZE, (1UL << 31), POOL_ALLOC_ALIGNED)
246 };
247 
248 void	*pool_multi_alloc_ni(struct pool *, int, int *);
249 void	pool_multi_free_ni(struct pool *, void *);
250 
251 struct pool_allocator pool_allocator_multi_ni = {
252 	pool_multi_alloc_ni,
253 	pool_multi_free_ni,
254 	POOL_ALLOC_SIZES(PAGE_SIZE, (1UL << 31), POOL_ALLOC_ALIGNED)
255 };
256 
257 #ifdef DDB
258 void	 pool_print_pagelist(struct pool_pagelist *, int (*)(const char *, ...)
259 	     __attribute__((__format__(__kprintf__,1,2))));
260 void	 pool_print1(struct pool *, const char *, int (*)(const char *, ...)
261 	     __attribute__((__format__(__kprintf__,1,2))));
262 #endif
263 
264 /* stale page garbage collectors */
265 void	pool_gc_sched(void *);
266 struct timeout pool_gc_tick = TIMEOUT_INITIALIZER(pool_gc_sched, NULL);
267 void	pool_gc_pages(void *);
268 struct task pool_gc_task = TASK_INITIALIZER(pool_gc_pages, NULL);
269 int pool_wait_free = 1;
270 int pool_wait_gc = 8;
271 
272 RBT_PROTOTYPE(phtree, pool_page_header, ph_node, phtree_compare);
273 
274 static inline int
275 phtree_compare(const struct pool_page_header *a,
276     const struct pool_page_header *b)
277 {
278 	vaddr_t va = (vaddr_t)a->ph_page;
279 	vaddr_t vb = (vaddr_t)b->ph_page;
280 
281 	/* the compares in this order are important for the NFIND to work */
282 	if (vb < va)
283 		return (-1);
284 	if (vb > va)
285 		return (1);
286 
287 	return (0);
288 }
289 
290 RBT_GENERATE(phtree, pool_page_header, ph_node, phtree_compare);
291 
292 /*
293  * Return the pool page header based on page address.
294  */
295 static inline struct pool_page_header *
296 pr_find_pagehead(struct pool *pp, void *v)
297 {
298 	struct pool_page_header *ph, key;
299 
300 	if (POOL_INPGHDR(pp)) {
301 		caddr_t page;
302 
303 		page = (caddr_t)((vaddr_t)v & pp->pr_pgmask);
304 
305 		return ((struct pool_page_header *)(page + pp->pr_phoffset));
306 	}
307 
308 	key.ph_page = v;
309 	ph = RBT_NFIND(phtree, &pp->pr_phtree, &key);
310 	if (ph == NULL)
311 		panic("%s: %s: page header missing", __func__, pp->pr_wchan);
312 
313 	KASSERT(ph->ph_page <= (caddr_t)v);
314 	if (ph->ph_page + pp->pr_pgsize <= (caddr_t)v)
315 		panic("%s: %s: incorrect page", __func__, pp->pr_wchan);
316 
317 	return (ph);
318 }
319 
320 /*
321  * Initialize the given pool resource structure.
322  *
323  * We export this routine to allow other kernel parts to declare
324  * static pools that must be initialized before malloc() is available.
325  */
326 void
327 pool_init(struct pool *pp, size_t size, u_int align, int ipl, int flags,
328     const char *wchan, struct pool_allocator *palloc)
329 {
330 	int off = 0, space;
331 	unsigned int pgsize = PAGE_SIZE, items;
332 	size_t pa_pagesz;
333 #ifdef DIAGNOSTIC
334 	struct pool *iter;
335 #endif
336 
337 	if (align == 0)
338 		align = ALIGN(1);
339 
340 	if (size < sizeof(struct pool_item))
341 		size = sizeof(struct pool_item);
342 
343 	size = roundup(size, align);
344 
345 	while (size * 8 > pgsize)
346 		pgsize <<= 1;
347 
348 	if (palloc == NULL) {
349 		if (pgsize > PAGE_SIZE) {
350 			palloc = ISSET(flags, PR_WAITOK) ?
351 			    &pool_allocator_multi_ni : &pool_allocator_multi;
352 		} else
353 			palloc = &pool_allocator_single;
354 
355 		pa_pagesz = palloc->pa_pagesz;
356 	} else {
357 		size_t pgsizes;
358 
359 		pa_pagesz = palloc->pa_pagesz;
360 		if (pa_pagesz == 0)
361 			pa_pagesz = POOL_ALLOC_DEFAULT;
362 
363 		pgsizes = pa_pagesz & ~POOL_ALLOC_ALIGNED;
364 
365 		/* make sure the allocator can fit at least one item */
366 		if (size > pgsizes) {
367 			panic("%s: pool %s item size 0x%zx > "
368 			    "allocator %p sizes 0x%zx", __func__, wchan,
369 			    size, palloc, pgsizes);
370 		}
371 
372 		/* shrink pgsize until it fits into the range */
373 		while (!ISSET(pgsizes, pgsize))
374 			pgsize >>= 1;
375 	}
376 	KASSERT(ISSET(pa_pagesz, pgsize));
377 
378 	items = pgsize / size;
379 
380 	/*
381 	 * Decide whether to put the page header off page to avoid
382 	 * wasting too large a part of the page. Off-page page headers
383 	 * go into an RB tree, so we can match a returned item with
384 	 * its header based on the page address.
385 	 */
386 	if (ISSET(pa_pagesz, POOL_ALLOC_ALIGNED)) {
387 		if (pgsize - (size * items) >
388 		    sizeof(struct pool_page_header)) {
389 			off = pgsize - sizeof(struct pool_page_header);
390 		} else if (sizeof(struct pool_page_header) * 2 >= size) {
391 			off = pgsize - sizeof(struct pool_page_header);
392 			items = off / size;
393 		}
394 	}
395 
396 	KASSERT(items > 0);
397 
398 	/*
399 	 * Initialize the pool structure.
400 	 */
401 	memset(pp, 0, sizeof(*pp));
402 	if (ISSET(flags, PR_RWLOCK)) {
403 		KASSERT(flags & PR_WAITOK);
404 		pp->pr_lock_ops = &pool_lock_ops_rw;
405 	} else
406 		pp->pr_lock_ops = &pool_lock_ops_mtx;
407 	TAILQ_INIT(&pp->pr_emptypages);
408 	TAILQ_INIT(&pp->pr_fullpages);
409 	TAILQ_INIT(&pp->pr_partpages);
410 	pp->pr_curpage = NULL;
411 	pp->pr_npages = 0;
412 	pp->pr_minitems = 0;
413 	pp->pr_minpages = 0;
414 	pp->pr_maxpages = 8;
415 	pp->pr_size = size;
416 	pp->pr_pgsize = pgsize;
417 	pp->pr_pgmask = ~0UL ^ (pgsize - 1);
418 	pp->pr_phoffset = off;
419 	pp->pr_itemsperpage = items;
420 	pp->pr_wchan = wchan;
421 	pp->pr_alloc = palloc;
422 	pp->pr_nitems = 0;
423 	pp->pr_nout = 0;
424 	pp->pr_hardlimit = UINT_MAX;
425 	pp->pr_hardlimit_warning = NULL;
426 	pp->pr_hardlimit_ratecap.tv_sec = 0;
427 	pp->pr_hardlimit_ratecap.tv_usec = 0;
428 	pp->pr_hardlimit_warning_last.tv_sec = 0;
429 	pp->pr_hardlimit_warning_last.tv_usec = 0;
430 	RBT_INIT(phtree, &pp->pr_phtree);
431 
432 	/*
433 	 * Use the space between the chunks and the page header
434 	 * for cache coloring.
435 	 */
436 	space = POOL_INPGHDR(pp) ? pp->pr_phoffset : pp->pr_pgsize;
437 	space -= pp->pr_itemsperpage * pp->pr_size;
438 	pp->pr_align = align;
439 	pp->pr_maxcolors = (space / align) + 1;
440 
441 	pp->pr_nget = 0;
442 	pp->pr_nfail = 0;
443 	pp->pr_nput = 0;
444 	pp->pr_npagealloc = 0;
445 	pp->pr_npagefree = 0;
446 	pp->pr_hiwat = 0;
447 	pp->pr_nidle = 0;
448 
449 	pp->pr_ipl = ipl;
450 	pp->pr_flags = flags;
451 
452 	pl_init(pp, &pp->pr_lock);
453 	pl_init(pp, &pp->pr_requests_lock);
454 	TAILQ_INIT(&pp->pr_requests);
455 
456 	if (phpool.pr_size == 0) {
457 		pool_init(&phpool, sizeof(struct pool_page_header), 0,
458 		    IPL_HIGH, 0, "phpool", NULL);
459 
460 		/* make sure phpool wont "recurse" */
461 		KASSERT(POOL_INPGHDR(&phpool));
462 	}
463 
464 	/* pglistalloc/constraint parameters */
465 	pp->pr_crange = &kp_dirty;
466 
467 	/* Insert this into the list of all pools. */
468 	rw_enter_write(&pool_lock);
469 #ifdef DIAGNOSTIC
470 	SIMPLEQ_FOREACH(iter, &pool_head, pr_poollist) {
471 		if (iter == pp)
472 			panic("%s: pool %s already on list", __func__, wchan);
473 	}
474 #endif
475 
476 	pp->pr_serial = ++pool_serial;
477 	if (pool_serial == 0)
478 		panic("%s: too much uptime", __func__);
479 
480 	SIMPLEQ_INSERT_HEAD(&pool_head, pp, pr_poollist);
481 	pool_count++;
482 	rw_exit_write(&pool_lock);
483 }
484 
485 /*
486  * Decommission a pool resource.
487  */
488 void
489 pool_destroy(struct pool *pp)
490 {
491 	struct pool_page_header *ph;
492 	struct pool *prev, *iter;
493 
494 #ifdef MULTIPROCESSOR
495 	if (pp->pr_cache != NULL)
496 		pool_cache_destroy(pp);
497 #endif
498 
499 #ifdef DIAGNOSTIC
500 	if (pp->pr_nout != 0)
501 		panic("%s: pool busy: still out: %u", __func__, pp->pr_nout);
502 #endif
503 
504 	/* Remove from global pool list */
505 	rw_enter_write(&pool_lock);
506 	pool_count--;
507 	if (pp == SIMPLEQ_FIRST(&pool_head))
508 		SIMPLEQ_REMOVE_HEAD(&pool_head, pr_poollist);
509 	else {
510 		prev = SIMPLEQ_FIRST(&pool_head);
511 		SIMPLEQ_FOREACH(iter, &pool_head, pr_poollist) {
512 			if (iter == pp) {
513 				SIMPLEQ_REMOVE_AFTER(&pool_head, prev,
514 				    pr_poollist);
515 				break;
516 			}
517 			prev = iter;
518 		}
519 	}
520 	rw_exit_write(&pool_lock);
521 
522 	/* Remove all pages */
523 	while ((ph = TAILQ_FIRST(&pp->pr_emptypages)) != NULL) {
524 		pl_enter(pp, &pp->pr_lock);
525 		pool_p_remove(pp, ph);
526 		pl_leave(pp, &pp->pr_lock);
527 		pool_p_free(pp, ph);
528 	}
529 	KASSERT(TAILQ_EMPTY(&pp->pr_fullpages));
530 	KASSERT(TAILQ_EMPTY(&pp->pr_partpages));
531 }
532 
533 void
534 pool_request_init(struct pool_request *pr,
535     void (*handler)(struct pool *, void *, void *), void *cookie)
536 {
537 	pr->pr_handler = handler;
538 	pr->pr_cookie = cookie;
539 	pr->pr_item = NULL;
540 }
541 
542 void
543 pool_request(struct pool *pp, struct pool_request *pr)
544 {
545 	pl_enter(pp, &pp->pr_requests_lock);
546 	TAILQ_INSERT_TAIL(&pp->pr_requests, pr, pr_entry);
547 	pool_runqueue(pp, PR_NOWAIT);
548 	pl_leave(pp, &pp->pr_requests_lock);
549 }
550 
551 struct pool_get_memory {
552 	union pool_lock lock;
553 	void * volatile v;
554 };
555 
556 /*
557  * Grab an item from the pool.
558  */
559 void *
560 pool_get(struct pool *pp, int flags)
561 {
562 	void *v = NULL;
563 	int slowdown = 0;
564 
565 	KASSERT(flags & (PR_WAITOK | PR_NOWAIT));
566 	if (pp->pr_flags & PR_RWLOCK)
567 		KASSERT(flags & PR_WAITOK);
568 
569 #ifdef MULTIPROCESSOR
570 	if (pp->pr_cache != NULL) {
571 		v = pool_cache_get(pp);
572 		if (v != NULL)
573 			goto good;
574 	}
575 #endif
576 
577 	pl_enter(pp, &pp->pr_lock);
578 	if (pp->pr_nout >= pp->pr_hardlimit) {
579 		if (ISSET(flags, PR_NOWAIT|PR_LIMITFAIL))
580 			goto fail;
581 	} else if ((v = pool_do_get(pp, flags, &slowdown)) == NULL) {
582 		if (ISSET(flags, PR_NOWAIT))
583 			goto fail;
584 	}
585 	pl_leave(pp, &pp->pr_lock);
586 
587 	if ((slowdown || pool_debug == 2) && ISSET(flags, PR_WAITOK))
588 		yield();
589 
590 	if (v == NULL) {
591 		struct pool_get_memory mem = { .v = NULL };
592 		struct pool_request pr;
593 
594 #ifdef DIAGNOSTIC
595 		if (ISSET(flags, PR_WAITOK) && curproc == &proc0)
596 			panic("%s: cannot sleep for memory during boot",
597 			    __func__);
598 #endif
599 		pl_init(pp, &mem.lock);
600 		pool_request_init(&pr, pool_get_done, &mem);
601 		pool_request(pp, &pr);
602 
603 		pl_enter(pp, &mem.lock);
604 		while (mem.v == NULL)
605 			pl_sleep(pp, &mem, &mem.lock, PSWP, pp->pr_wchan, 0);
606 		pl_leave(pp, &mem.lock);
607 
608 		v = mem.v;
609 	}
610 
611 #ifdef MULTIPROCESSOR
612 good:
613 #endif
614 	if (ISSET(flags, PR_ZERO))
615 		memset(v, 0, pp->pr_size);
616 
617 	return (v);
618 
619 fail:
620 	pp->pr_nfail++;
621 	pl_leave(pp, &pp->pr_lock);
622 	return (NULL);
623 }
624 
625 void
626 pool_get_done(struct pool *pp, void *xmem, void *v)
627 {
628 	struct pool_get_memory *mem = xmem;
629 
630 	pl_enter(pp, &mem->lock);
631 	mem->v = v;
632 	pl_leave(pp, &mem->lock);
633 
634 	wakeup_one(mem);
635 }
636 
637 void
638 pool_runqueue(struct pool *pp, int flags)
639 {
640 	struct pool_requests prl = TAILQ_HEAD_INITIALIZER(prl);
641 	struct pool_request *pr;
642 
643 	pl_assert_unlocked(pp, &pp->pr_lock);
644 	pl_assert_locked(pp, &pp->pr_requests_lock);
645 
646 	if (pp->pr_requesting++)
647 		return;
648 
649 	do {
650 		pp->pr_requesting = 1;
651 
652 		/* no TAILQ_JOIN? :( */
653 		while ((pr = TAILQ_FIRST(&pp->pr_requests)) != NULL) {
654 			TAILQ_REMOVE(&pp->pr_requests, pr, pr_entry);
655 			TAILQ_INSERT_TAIL(&prl, pr, pr_entry);
656 		}
657 		if (TAILQ_EMPTY(&prl))
658 			continue;
659 
660 		pl_leave(pp, &pp->pr_requests_lock);
661 
662 		pl_enter(pp, &pp->pr_lock);
663 		pr = TAILQ_FIRST(&prl);
664 		while (pr != NULL) {
665 			int slowdown = 0;
666 
667 			if (pp->pr_nout >= pp->pr_hardlimit)
668 				break;
669 
670 			pr->pr_item = pool_do_get(pp, flags, &slowdown);
671 			if (pr->pr_item == NULL) /* || slowdown ? */
672 				break;
673 
674 			pr = TAILQ_NEXT(pr, pr_entry);
675 		}
676 		pl_leave(pp, &pp->pr_lock);
677 
678 		while ((pr = TAILQ_FIRST(&prl)) != NULL &&
679 		    pr->pr_item != NULL) {
680 			TAILQ_REMOVE(&prl, pr, pr_entry);
681 			(*pr->pr_handler)(pp, pr->pr_cookie, pr->pr_item);
682 		}
683 
684 		pl_enter(pp, &pp->pr_requests_lock);
685 	} while (--pp->pr_requesting);
686 
687 	/* no TAILQ_JOIN :( */
688 	while ((pr = TAILQ_FIRST(&prl)) != NULL) {
689 		TAILQ_REMOVE(&prl, pr, pr_entry);
690 		TAILQ_INSERT_TAIL(&pp->pr_requests, pr, pr_entry);
691 	}
692 }
693 
694 void *
695 pool_do_get(struct pool *pp, int flags, int *slowdown)
696 {
697 	struct pool_item *pi;
698 	struct pool_page_header *ph;
699 
700 	pl_assert_locked(pp, &pp->pr_lock);
701 
702 	splassert(pp->pr_ipl);
703 
704 	/*
705 	 * Account for this item now to avoid races if we need to give up
706 	 * pr_lock to allocate a page.
707 	 */
708 	pp->pr_nout++;
709 
710 	if (pp->pr_curpage == NULL) {
711 		pl_leave(pp, &pp->pr_lock);
712 		ph = pool_p_alloc(pp, flags, slowdown);
713 		pl_enter(pp, &pp->pr_lock);
714 
715 		if (ph == NULL) {
716 			pp->pr_nout--;
717 			return (NULL);
718 		}
719 
720 		pool_p_insert(pp, ph);
721 	}
722 
723 	ph = pp->pr_curpage;
724 	pi = XSIMPLEQ_FIRST(&ph->ph_items);
725 	if (__predict_false(pi == NULL))
726 		panic("%s: %s: page empty", __func__, pp->pr_wchan);
727 
728 	if (__predict_false(pi->pi_magic != POOL_IMAGIC(ph, pi))) {
729 		panic("%s: %s free list modified: "
730 		    "page %p; item addr %p; offset 0x%x=0x%lx != 0x%lx",
731 		    __func__, pp->pr_wchan, ph->ph_page, pi,
732 		    0, pi->pi_magic, POOL_IMAGIC(ph, pi));
733 	}
734 
735 	XSIMPLEQ_REMOVE_HEAD(&ph->ph_items, pi_list);
736 
737 #ifdef DIAGNOSTIC
738 	if (pool_debug && POOL_PHPOISON(ph)) {
739 		size_t pidx;
740 		uint32_t pval;
741 		if (poison_check(pi + 1, pp->pr_size - sizeof(*pi),
742 		    &pidx, &pval)) {
743 			int *ip = (int *)(pi + 1);
744 			panic("%s: %s free list modified: "
745 			    "page %p; item addr %p; offset 0x%zx=0x%x",
746 			    __func__, pp->pr_wchan, ph->ph_page, pi,
747 			    (pidx * sizeof(int)) + sizeof(*pi), ip[pidx]);
748 		}
749 	}
750 #endif /* DIAGNOSTIC */
751 
752 	if (ph->ph_nmissing++ == 0) {
753 		/*
754 		 * This page was previously empty.  Move it to the list of
755 		 * partially-full pages.  This page is already curpage.
756 		 */
757 		TAILQ_REMOVE(&pp->pr_emptypages, ph, ph_entry);
758 		TAILQ_INSERT_TAIL(&pp->pr_partpages, ph, ph_entry);
759 
760 		pp->pr_nidle--;
761 	}
762 
763 	if (ph->ph_nmissing == pp->pr_itemsperpage) {
764 		/*
765 		 * This page is now full.  Move it to the full list
766 		 * and select a new current page.
767 		 */
768 		TAILQ_REMOVE(&pp->pr_partpages, ph, ph_entry);
769 		TAILQ_INSERT_TAIL(&pp->pr_fullpages, ph, ph_entry);
770 		pool_update_curpage(pp);
771 	}
772 
773 	pp->pr_nget++;
774 
775 	return (pi);
776 }
777 
778 /*
779  * Return resource to the pool.
780  */
781 void
782 pool_put(struct pool *pp, void *v)
783 {
784 	struct pool_page_header *ph, *freeph = NULL;
785 
786 #ifdef DIAGNOSTIC
787 	if (v == NULL)
788 		panic("%s: NULL item", __func__);
789 #endif
790 
791 #ifdef MULTIPROCESSOR
792 	if (pp->pr_cache != NULL && TAILQ_EMPTY(&pp->pr_requests)) {
793 		pool_cache_put(pp, v);
794 		return;
795 	}
796 #endif
797 
798 	pl_enter(pp, &pp->pr_lock);
799 
800 	pool_do_put(pp, v);
801 
802 	pp->pr_nout--;
803 	pp->pr_nput++;
804 
805 	/* is it time to free a page? */
806 	if (pp->pr_nidle > pp->pr_maxpages &&
807 	    (ph = TAILQ_FIRST(&pp->pr_emptypages)) != NULL &&
808 	    (ticks - ph->ph_tick) > (hz * pool_wait_free)) {
809 		freeph = ph;
810 		pool_p_remove(pp, freeph);
811 	}
812 
813 	pl_leave(pp, &pp->pr_lock);
814 
815 	if (freeph != NULL)
816 		pool_p_free(pp, freeph);
817 
818 	pool_wakeup(pp);
819 }
820 
821 void
822 pool_wakeup(struct pool *pp)
823 {
824 	if (!TAILQ_EMPTY(&pp->pr_requests)) {
825 		pl_enter(pp, &pp->pr_requests_lock);
826 		pool_runqueue(pp, PR_NOWAIT);
827 		pl_leave(pp, &pp->pr_requests_lock);
828 	}
829 }
830 
831 void
832 pool_do_put(struct pool *pp, void *v)
833 {
834 	struct pool_item *pi = v;
835 	struct pool_page_header *ph;
836 
837 	splassert(pp->pr_ipl);
838 
839 	ph = pr_find_pagehead(pp, v);
840 
841 #ifdef DIAGNOSTIC
842 	if (pool_debug) {
843 		struct pool_item *qi;
844 		XSIMPLEQ_FOREACH(qi, &ph->ph_items, pi_list) {
845 			if (pi == qi) {
846 				panic("%s: %s: double pool_put: %p", __func__,
847 				    pp->pr_wchan, pi);
848 			}
849 		}
850 	}
851 #endif /* DIAGNOSTIC */
852 
853 	pi->pi_magic = POOL_IMAGIC(ph, pi);
854 	XSIMPLEQ_INSERT_HEAD(&ph->ph_items, pi, pi_list);
855 #ifdef DIAGNOSTIC
856 	if (POOL_PHPOISON(ph))
857 		poison_mem(pi + 1, pp->pr_size - sizeof(*pi));
858 #endif /* DIAGNOSTIC */
859 
860 	if (ph->ph_nmissing-- == pp->pr_itemsperpage) {
861 		/*
862 		 * The page was previously completely full, move it to the
863 		 * partially-full list.
864 		 */
865 		TAILQ_REMOVE(&pp->pr_fullpages, ph, ph_entry);
866 		TAILQ_INSERT_TAIL(&pp->pr_partpages, ph, ph_entry);
867 	}
868 
869 	if (ph->ph_nmissing == 0) {
870 		/*
871 		 * The page is now empty, so move it to the empty page list.
872 		 */
873 		pp->pr_nidle++;
874 
875 		ph->ph_tick = ticks;
876 		TAILQ_REMOVE(&pp->pr_partpages, ph, ph_entry);
877 		TAILQ_INSERT_TAIL(&pp->pr_emptypages, ph, ph_entry);
878 		pool_update_curpage(pp);
879 	}
880 }
881 
882 /*
883  * Add N items to the pool.
884  */
885 int
886 pool_prime(struct pool *pp, int n)
887 {
888 	struct pool_pagelist pl = TAILQ_HEAD_INITIALIZER(pl);
889 	struct pool_page_header *ph;
890 	int newpages;
891 
892 	newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
893 
894 	while (newpages-- > 0) {
895 		int slowdown = 0;
896 
897 		ph = pool_p_alloc(pp, PR_NOWAIT, &slowdown);
898 		if (ph == NULL) /* or slowdown? */
899 			break;
900 
901 		TAILQ_INSERT_TAIL(&pl, ph, ph_entry);
902 	}
903 
904 	pl_enter(pp, &pp->pr_lock);
905 	while ((ph = TAILQ_FIRST(&pl)) != NULL) {
906 		TAILQ_REMOVE(&pl, ph, ph_entry);
907 		pool_p_insert(pp, ph);
908 	}
909 	pl_leave(pp, &pp->pr_lock);
910 
911 	return (0);
912 }
913 
914 struct pool_page_header *
915 pool_p_alloc(struct pool *pp, int flags, int *slowdown)
916 {
917 	struct pool_page_header *ph;
918 	struct pool_item *pi;
919 	caddr_t addr;
920 	unsigned int order;
921 	int o;
922 	int n;
923 
924 	pl_assert_unlocked(pp, &pp->pr_lock);
925 	KASSERT(pp->pr_size >= sizeof(*pi));
926 
927 	addr = pool_allocator_alloc(pp, flags, slowdown);
928 	if (addr == NULL)
929 		return (NULL);
930 
931 	if (POOL_INPGHDR(pp))
932 		ph = (struct pool_page_header *)(addr + pp->pr_phoffset);
933 	else {
934 		ph = pool_get(&phpool, flags);
935 		if (ph == NULL) {
936 			pool_allocator_free(pp, addr);
937 			return (NULL);
938 		}
939 	}
940 
941 	XSIMPLEQ_INIT(&ph->ph_items);
942 	ph->ph_page = addr;
943 	addr += pp->pr_align * (pp->pr_npagealloc % pp->pr_maxcolors);
944 	ph->ph_colored = addr;
945 	ph->ph_nmissing = 0;
946 	arc4random_buf(&ph->ph_magic, sizeof(ph->ph_magic));
947 #ifdef DIAGNOSTIC
948 	/* use a bit in ph_magic to record if we poison page items */
949 	if (pool_debug)
950 		SET(ph->ph_magic, POOL_MAGICBIT);
951 	else
952 		CLR(ph->ph_magic, POOL_MAGICBIT);
953 #endif /* DIAGNOSTIC */
954 
955 	n = pp->pr_itemsperpage;
956 	o = 32;
957 	while (n--) {
958 		pi = (struct pool_item *)addr;
959 		pi->pi_magic = POOL_IMAGIC(ph, pi);
960 
961 		if (o == 32) {
962 			order = arc4random();
963 			o = 0;
964 		}
965 		if (ISSET(order, 1 << o++))
966 			XSIMPLEQ_INSERT_TAIL(&ph->ph_items, pi, pi_list);
967 		else
968 			XSIMPLEQ_INSERT_HEAD(&ph->ph_items, pi, pi_list);
969 
970 #ifdef DIAGNOSTIC
971 		if (POOL_PHPOISON(ph))
972 			poison_mem(pi + 1, pp->pr_size - sizeof(*pi));
973 #endif /* DIAGNOSTIC */
974 
975 		addr += pp->pr_size;
976 	}
977 
978 	return (ph);
979 }
980 
981 void
982 pool_p_free(struct pool *pp, struct pool_page_header *ph)
983 {
984 	struct pool_item *pi;
985 
986 	pl_assert_unlocked(pp, &pp->pr_lock);
987 	KASSERT(ph->ph_nmissing == 0);
988 
989 	XSIMPLEQ_FOREACH(pi, &ph->ph_items, pi_list) {
990 		if (__predict_false(pi->pi_magic != POOL_IMAGIC(ph, pi))) {
991 			panic("%s: %s free list modified: "
992 			    "page %p; item addr %p; offset 0x%x=0x%lx",
993 			    __func__, pp->pr_wchan, ph->ph_page, pi,
994 			    0, pi->pi_magic);
995 		}
996 
997 #ifdef DIAGNOSTIC
998 		if (POOL_PHPOISON(ph)) {
999 			size_t pidx;
1000 			uint32_t pval;
1001 			if (poison_check(pi + 1, pp->pr_size - sizeof(*pi),
1002 			    &pidx, &pval)) {
1003 				int *ip = (int *)(pi + 1);
1004 				panic("%s: %s free list modified: "
1005 				    "page %p; item addr %p; offset 0x%zx=0x%x",
1006 				    __func__, pp->pr_wchan, ph->ph_page, pi,
1007 				    pidx * sizeof(int), ip[pidx]);
1008 			}
1009 		}
1010 #endif
1011 	}
1012 
1013 	pool_allocator_free(pp, ph->ph_page);
1014 
1015 	if (!POOL_INPGHDR(pp))
1016 		pool_put(&phpool, ph);
1017 }
1018 
1019 void
1020 pool_p_insert(struct pool *pp, struct pool_page_header *ph)
1021 {
1022 	pl_assert_locked(pp, &pp->pr_lock);
1023 
1024 	/* If the pool was depleted, point at the new page */
1025 	if (pp->pr_curpage == NULL)
1026 		pp->pr_curpage = ph;
1027 
1028 	TAILQ_INSERT_TAIL(&pp->pr_emptypages, ph, ph_entry);
1029 	if (!POOL_INPGHDR(pp))
1030 		RBT_INSERT(phtree, &pp->pr_phtree, ph);
1031 
1032 	pp->pr_nitems += pp->pr_itemsperpage;
1033 	pp->pr_nidle++;
1034 
1035 	pp->pr_npagealloc++;
1036 	if (++pp->pr_npages > pp->pr_hiwat)
1037 		pp->pr_hiwat = pp->pr_npages;
1038 }
1039 
1040 void
1041 pool_p_remove(struct pool *pp, struct pool_page_header *ph)
1042 {
1043 	pl_assert_locked(pp, &pp->pr_lock);
1044 
1045 	pp->pr_npagefree++;
1046 	pp->pr_npages--;
1047 	pp->pr_nidle--;
1048 	pp->pr_nitems -= pp->pr_itemsperpage;
1049 
1050 	if (!POOL_INPGHDR(pp))
1051 		RBT_REMOVE(phtree, &pp->pr_phtree, ph);
1052 	TAILQ_REMOVE(&pp->pr_emptypages, ph, ph_entry);
1053 
1054 	pool_update_curpage(pp);
1055 }
1056 
1057 void
1058 pool_update_curpage(struct pool *pp)
1059 {
1060 	pp->pr_curpage = TAILQ_LAST(&pp->pr_partpages, pool_pagelist);
1061 	if (pp->pr_curpage == NULL) {
1062 		pp->pr_curpage = TAILQ_LAST(&pp->pr_emptypages, pool_pagelist);
1063 	}
1064 }
1065 
1066 void
1067 pool_setlowat(struct pool *pp, int n)
1068 {
1069 	int prime = 0;
1070 
1071 	pl_enter(pp, &pp->pr_lock);
1072 	pp->pr_minitems = n;
1073 	pp->pr_minpages = (n == 0)
1074 		? 0
1075 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1076 
1077 	if (pp->pr_nitems < n)
1078 		prime = n - pp->pr_nitems;
1079 	pl_leave(pp, &pp->pr_lock);
1080 
1081 	if (prime > 0)
1082 		pool_prime(pp, prime);
1083 }
1084 
1085 void
1086 pool_sethiwat(struct pool *pp, int n)
1087 {
1088 	pp->pr_maxpages = (n == 0)
1089 		? 0
1090 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1091 }
1092 
1093 int
1094 pool_sethardlimit(struct pool *pp, u_int n, const char *warnmsg, int ratecap)
1095 {
1096 	int error = 0;
1097 
1098 	if (n < pp->pr_nout) {
1099 		error = EINVAL;
1100 		goto done;
1101 	}
1102 
1103 	pp->pr_hardlimit = n;
1104 	pp->pr_hardlimit_warning = warnmsg;
1105 	pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1106 	pp->pr_hardlimit_warning_last.tv_sec = 0;
1107 	pp->pr_hardlimit_warning_last.tv_usec = 0;
1108 
1109 done:
1110 	return (error);
1111 }
1112 
1113 void
1114 pool_set_constraints(struct pool *pp, const struct kmem_pa_mode *mode)
1115 {
1116 	pp->pr_crange = mode;
1117 }
1118 
1119 /*
1120  * Release all complete pages that have not been used recently.
1121  *
1122  * Returns non-zero if any pages have been reclaimed.
1123  */
1124 int
1125 pool_reclaim(struct pool *pp)
1126 {
1127 	struct pool_page_header *ph, *phnext;
1128 	struct pool_pagelist pl = TAILQ_HEAD_INITIALIZER(pl);
1129 
1130 	pl_enter(pp, &pp->pr_lock);
1131 	for (ph = TAILQ_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1132 		phnext = TAILQ_NEXT(ph, ph_entry);
1133 
1134 		/* Check our minimum page claim */
1135 		if (pp->pr_npages <= pp->pr_minpages)
1136 			break;
1137 
1138 		/*
1139 		 * If freeing this page would put us below
1140 		 * the low water mark, stop now.
1141 		 */
1142 		if ((pp->pr_nitems - pp->pr_itemsperpage) <
1143 		    pp->pr_minitems)
1144 			break;
1145 
1146 		pool_p_remove(pp, ph);
1147 		TAILQ_INSERT_TAIL(&pl, ph, ph_entry);
1148 	}
1149 	pl_leave(pp, &pp->pr_lock);
1150 
1151 	if (TAILQ_EMPTY(&pl))
1152 		return (0);
1153 
1154 	while ((ph = TAILQ_FIRST(&pl)) != NULL) {
1155 		TAILQ_REMOVE(&pl, ph, ph_entry);
1156 		pool_p_free(pp, ph);
1157 	}
1158 
1159 	return (1);
1160 }
1161 
1162 /*
1163  * Release all complete pages that have not been used recently
1164  * from all pools.
1165  */
1166 void
1167 pool_reclaim_all(void)
1168 {
1169 	struct pool	*pp;
1170 
1171 	rw_enter_read(&pool_lock);
1172 	SIMPLEQ_FOREACH(pp, &pool_head, pr_poollist)
1173 		pool_reclaim(pp);
1174 	rw_exit_read(&pool_lock);
1175 }
1176 
1177 #ifdef DDB
1178 #include <machine/db_machdep.h>
1179 #include <ddb/db_output.h>
1180 
1181 /*
1182  * Diagnostic helpers.
1183  */
1184 void
1185 pool_printit(struct pool *pp, const char *modif,
1186     int (*pr)(const char *, ...) __attribute__((__format__(__kprintf__,1,2))))
1187 {
1188 	pool_print1(pp, modif, pr);
1189 }
1190 
1191 void
1192 pool_print_pagelist(struct pool_pagelist *pl,
1193     int (*pr)(const char *, ...) __attribute__((__format__(__kprintf__,1,2))))
1194 {
1195 	struct pool_page_header *ph;
1196 	struct pool_item *pi;
1197 
1198 	TAILQ_FOREACH(ph, pl, ph_entry) {
1199 		(*pr)("\t\tpage %p, color %p, nmissing %d\n",
1200 		    ph->ph_page, ph->ph_colored, ph->ph_nmissing);
1201 		XSIMPLEQ_FOREACH(pi, &ph->ph_items, pi_list) {
1202 			if (pi->pi_magic != POOL_IMAGIC(ph, pi)) {
1203 				(*pr)("\t\t\titem %p, magic 0x%lx\n",
1204 				    pi, pi->pi_magic);
1205 			}
1206 		}
1207 	}
1208 }
1209 
1210 void
1211 pool_print1(struct pool *pp, const char *modif,
1212     int (*pr)(const char *, ...) __attribute__((__format__(__kprintf__,1,2))))
1213 {
1214 	struct pool_page_header *ph;
1215 	int print_pagelist = 0;
1216 	char c;
1217 
1218 	while ((c = *modif++) != '\0') {
1219 		if (c == 'p')
1220 			print_pagelist = 1;
1221 		modif++;
1222 	}
1223 
1224 	(*pr)("POOL %s: size %u maxcolors %u\n", pp->pr_wchan, pp->pr_size,
1225 	    pp->pr_maxcolors);
1226 	(*pr)("\talloc %p\n", pp->pr_alloc);
1227 	(*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1228 	    pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1229 	(*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1230 	    pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1231 
1232 	(*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1233 	    pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1234 	(*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1235 	    pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1236 
1237 	if (print_pagelist == 0)
1238 		return;
1239 
1240 	if ((ph = TAILQ_FIRST(&pp->pr_emptypages)) != NULL)
1241 		(*pr)("\n\tempty page list:\n");
1242 	pool_print_pagelist(&pp->pr_emptypages, pr);
1243 	if ((ph = TAILQ_FIRST(&pp->pr_fullpages)) != NULL)
1244 		(*pr)("\n\tfull page list:\n");
1245 	pool_print_pagelist(&pp->pr_fullpages, pr);
1246 	if ((ph = TAILQ_FIRST(&pp->pr_partpages)) != NULL)
1247 		(*pr)("\n\tpartial-page list:\n");
1248 	pool_print_pagelist(&pp->pr_partpages, pr);
1249 
1250 	if (pp->pr_curpage == NULL)
1251 		(*pr)("\tno current page\n");
1252 	else
1253 		(*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1254 }
1255 
1256 void
1257 db_show_all_pools(db_expr_t expr, int haddr, db_expr_t count, char *modif)
1258 {
1259 	struct pool *pp;
1260 	char maxp[16];
1261 	int ovflw;
1262 	char mode;
1263 
1264 	mode = modif[0];
1265 	if (mode != '\0' && mode != 'a') {
1266 		db_printf("usage: show all pools [/a]\n");
1267 		return;
1268 	}
1269 
1270 	if (mode == '\0')
1271 		db_printf("%-10s%4s%9s%5s%9s%6s%6s%6s%6s%6s%6s%5s\n",
1272 		    "Name",
1273 		    "Size",
1274 		    "Requests",
1275 		    "Fail",
1276 		    "Releases",
1277 		    "Pgreq",
1278 		    "Pgrel",
1279 		    "Npage",
1280 		    "Hiwat",
1281 		    "Minpg",
1282 		    "Maxpg",
1283 		    "Idle");
1284 	else
1285 		db_printf("%-12s %18s %18s\n",
1286 		    "Name", "Address", "Allocator");
1287 
1288 	SIMPLEQ_FOREACH(pp, &pool_head, pr_poollist) {
1289 		if (mode == 'a') {
1290 			db_printf("%-12s %18p %18p\n", pp->pr_wchan, pp,
1291 			    pp->pr_alloc);
1292 			continue;
1293 		}
1294 
1295 		if (!pp->pr_nget)
1296 			continue;
1297 
1298 		if (pp->pr_maxpages == UINT_MAX)
1299 			snprintf(maxp, sizeof maxp, "inf");
1300 		else
1301 			snprintf(maxp, sizeof maxp, "%u", pp->pr_maxpages);
1302 
1303 #define PRWORD(ovflw, fmt, width, fixed, val) do {	\
1304 	(ovflw) += db_printf((fmt),			\
1305 	    (width) - (fixed) - (ovflw) > 0 ?		\
1306 	    (width) - (fixed) - (ovflw) : 0,		\
1307 	    (val)) - (width);				\
1308 	if ((ovflw) < 0)				\
1309 		(ovflw) = 0;				\
1310 } while (/* CONSTCOND */0)
1311 
1312 		ovflw = 0;
1313 		PRWORD(ovflw, "%-*s", 10, 0, pp->pr_wchan);
1314 		PRWORD(ovflw, " %*u", 4, 1, pp->pr_size);
1315 		PRWORD(ovflw, " %*lu", 9, 1, pp->pr_nget);
1316 		PRWORD(ovflw, " %*lu", 5, 1, pp->pr_nfail);
1317 		PRWORD(ovflw, " %*lu", 9, 1, pp->pr_nput);
1318 		PRWORD(ovflw, " %*lu", 6, 1, pp->pr_npagealloc);
1319 		PRWORD(ovflw, " %*lu", 6, 1, pp->pr_npagefree);
1320 		PRWORD(ovflw, " %*d", 6, 1, pp->pr_npages);
1321 		PRWORD(ovflw, " %*d", 6, 1, pp->pr_hiwat);
1322 		PRWORD(ovflw, " %*d", 6, 1, pp->pr_minpages);
1323 		PRWORD(ovflw, " %*s", 6, 1, maxp);
1324 		PRWORD(ovflw, " %*lu\n", 5, 1, pp->pr_nidle);
1325 
1326 		pool_chk(pp);
1327 	}
1328 }
1329 #endif /* DDB */
1330 
1331 #if defined(POOL_DEBUG) || defined(DDB)
1332 int
1333 pool_chk_page(struct pool *pp, struct pool_page_header *ph, int expected)
1334 {
1335 	struct pool_item *pi;
1336 	caddr_t page;
1337 	int n;
1338 	const char *label = pp->pr_wchan;
1339 
1340 	page = (caddr_t)((u_long)ph & pp->pr_pgmask);
1341 	if (page != ph->ph_page && POOL_INPGHDR(pp)) {
1342 		printf("%s: ", label);
1343 		printf("pool(%p:%s): page inconsistency: page %p; "
1344 		    "at page head addr %p (p %p)\n",
1345 		    pp, pp->pr_wchan, ph->ph_page, ph, page);
1346 		return 1;
1347 	}
1348 
1349 	for (pi = XSIMPLEQ_FIRST(&ph->ph_items), n = 0;
1350 	     pi != NULL;
1351 	     pi = XSIMPLEQ_NEXT(&ph->ph_items, pi, pi_list), n++) {
1352 		if ((caddr_t)pi < ph->ph_page ||
1353 		    (caddr_t)pi >= ph->ph_page + pp->pr_pgsize) {
1354 			printf("%s: ", label);
1355 			printf("pool(%p:%s): page inconsistency: page %p;"
1356 			    " item ordinal %d; addr %p\n", pp,
1357 			    pp->pr_wchan, ph->ph_page, n, pi);
1358 			return (1);
1359 		}
1360 
1361 		if (pi->pi_magic != POOL_IMAGIC(ph, pi)) {
1362 			printf("%s: ", label);
1363 			printf("pool(%p:%s): free list modified: "
1364 			    "page %p; item ordinal %d; addr %p "
1365 			    "(p %p); offset 0x%x=0x%lx\n",
1366 			    pp, pp->pr_wchan, ph->ph_page, n, pi, page,
1367 			    0, pi->pi_magic);
1368 		}
1369 
1370 #ifdef DIAGNOSTIC
1371 		if (POOL_PHPOISON(ph)) {
1372 			size_t pidx;
1373 			uint32_t pval;
1374 			if (poison_check(pi + 1, pp->pr_size - sizeof(*pi),
1375 			    &pidx, &pval)) {
1376 				int *ip = (int *)(pi + 1);
1377 				printf("pool(%s): free list modified: "
1378 				    "page %p; item ordinal %d; addr %p "
1379 				    "(p %p); offset 0x%zx=0x%x\n",
1380 				    pp->pr_wchan, ph->ph_page, n, pi,
1381 				    page, pidx * sizeof(int), ip[pidx]);
1382 			}
1383 		}
1384 #endif /* DIAGNOSTIC */
1385 	}
1386 	if (n + ph->ph_nmissing != pp->pr_itemsperpage) {
1387 		printf("pool(%p:%s): page inconsistency: page %p;"
1388 		    " %d on list, %d missing, %d items per page\n", pp,
1389 		    pp->pr_wchan, ph->ph_page, n, ph->ph_nmissing,
1390 		    pp->pr_itemsperpage);
1391 		return 1;
1392 	}
1393 	if (expected >= 0 && n != expected) {
1394 		printf("pool(%p:%s): page inconsistency: page %p;"
1395 		    " %d on list, %d missing, %d expected\n", pp,
1396 		    pp->pr_wchan, ph->ph_page, n, ph->ph_nmissing,
1397 		    expected);
1398 		return 1;
1399 	}
1400 	return 0;
1401 }
1402 
1403 int
1404 pool_chk(struct pool *pp)
1405 {
1406 	struct pool_page_header *ph;
1407 	int r = 0;
1408 
1409 	TAILQ_FOREACH(ph, &pp->pr_emptypages, ph_entry)
1410 		r += pool_chk_page(pp, ph, pp->pr_itemsperpage);
1411 	TAILQ_FOREACH(ph, &pp->pr_fullpages, ph_entry)
1412 		r += pool_chk_page(pp, ph, 0);
1413 	TAILQ_FOREACH(ph, &pp->pr_partpages, ph_entry)
1414 		r += pool_chk_page(pp, ph, -1);
1415 
1416 	return (r);
1417 }
1418 #endif /* defined(POOL_DEBUG) || defined(DDB) */
1419 
1420 #ifdef DDB
1421 void
1422 pool_walk(struct pool *pp, int full,
1423     int (*pr)(const char *, ...) __attribute__((__format__(__kprintf__,1,2))),
1424     void (*func)(void *, int, int (*)(const char *, ...)
1425 	    __attribute__((__format__(__kprintf__,1,2)))))
1426 {
1427 	struct pool_page_header *ph;
1428 	struct pool_item *pi;
1429 	caddr_t cp;
1430 	int n;
1431 
1432 	TAILQ_FOREACH(ph, &pp->pr_fullpages, ph_entry) {
1433 		cp = ph->ph_colored;
1434 		n = ph->ph_nmissing;
1435 
1436 		while (n--) {
1437 			func(cp, full, pr);
1438 			cp += pp->pr_size;
1439 		}
1440 	}
1441 
1442 	TAILQ_FOREACH(ph, &pp->pr_partpages, ph_entry) {
1443 		cp = ph->ph_colored;
1444 		n = ph->ph_nmissing;
1445 
1446 		do {
1447 			XSIMPLEQ_FOREACH(pi, &ph->ph_items, pi_list) {
1448 				if (cp == (caddr_t)pi)
1449 					break;
1450 			}
1451 			if (cp != (caddr_t)pi) {
1452 				func(cp, full, pr);
1453 				n--;
1454 			}
1455 
1456 			cp += pp->pr_size;
1457 		} while (n > 0);
1458 	}
1459 }
1460 #endif
1461 
1462 /*
1463  * We have three different sysctls.
1464  * kern.pool.npools - the number of pools.
1465  * kern.pool.pool.<pool#> - the pool struct for the pool#.
1466  * kern.pool.name.<pool#> - the name for pool#.
1467  */
1468 int
1469 sysctl_dopool(int *name, u_int namelen, char *oldp, size_t *oldlenp)
1470 {
1471 	struct kinfo_pool pi;
1472 	struct pool *pp;
1473 	int rv = ENOENT;
1474 
1475 	switch (name[0]) {
1476 	case KERN_POOL_NPOOLS:
1477 		if (namelen != 1)
1478 			return (ENOTDIR);
1479 		return (sysctl_rdint(oldp, oldlenp, NULL, pool_count));
1480 
1481 	case KERN_POOL_NAME:
1482 	case KERN_POOL_POOL:
1483 	case KERN_POOL_CACHE:
1484 	case KERN_POOL_CACHE_CPUS:
1485 		break;
1486 	default:
1487 		return (EOPNOTSUPP);
1488 	}
1489 
1490 	if (namelen != 2)
1491 		return (ENOTDIR);
1492 
1493 	rw_enter_read(&pool_lock);
1494 
1495 	SIMPLEQ_FOREACH(pp, &pool_head, pr_poollist) {
1496 		if (name[1] == pp->pr_serial)
1497 			break;
1498 	}
1499 
1500 	if (pp == NULL)
1501 		goto done;
1502 
1503 	switch (name[0]) {
1504 	case KERN_POOL_NAME:
1505 		rv = sysctl_rdstring(oldp, oldlenp, NULL, pp->pr_wchan);
1506 		break;
1507 	case KERN_POOL_POOL:
1508 		memset(&pi, 0, sizeof(pi));
1509 
1510 		pl_enter(pp, &pp->pr_lock);
1511 		pi.pr_size = pp->pr_size;
1512 		pi.pr_pgsize = pp->pr_pgsize;
1513 		pi.pr_itemsperpage = pp->pr_itemsperpage;
1514 		pi.pr_npages = pp->pr_npages;
1515 		pi.pr_minpages = pp->pr_minpages;
1516 		pi.pr_maxpages = pp->pr_maxpages;
1517 		pi.pr_hardlimit = pp->pr_hardlimit;
1518 		pi.pr_nout = pp->pr_nout;
1519 		pi.pr_nitems = pp->pr_nitems;
1520 		pi.pr_nget = pp->pr_nget;
1521 		pi.pr_nput = pp->pr_nput;
1522 		pi.pr_nfail = pp->pr_nfail;
1523 		pi.pr_npagealloc = pp->pr_npagealloc;
1524 		pi.pr_npagefree = pp->pr_npagefree;
1525 		pi.pr_hiwat = pp->pr_hiwat;
1526 		pi.pr_nidle = pp->pr_nidle;
1527 		pl_leave(pp, &pp->pr_lock);
1528 
1529 		pool_cache_pool_info(pp, &pi);
1530 
1531 		rv = sysctl_rdstruct(oldp, oldlenp, NULL, &pi, sizeof(pi));
1532 		break;
1533 
1534 	case KERN_POOL_CACHE:
1535 		rv = pool_cache_info(pp, oldp, oldlenp);
1536 		break;
1537 
1538 	case KERN_POOL_CACHE_CPUS:
1539 		rv = pool_cache_cpus_info(pp, oldp, oldlenp);
1540 		break;
1541 	}
1542 
1543 done:
1544 	rw_exit_read(&pool_lock);
1545 
1546 	return (rv);
1547 }
1548 
1549 void
1550 pool_gc_sched(void *null)
1551 {
1552 	task_add(systqmp, &pool_gc_task);
1553 }
1554 
1555 void
1556 pool_gc_pages(void *null)
1557 {
1558 	struct pool *pp;
1559 	struct pool_page_header *ph, *freeph;
1560 	int s;
1561 
1562 	rw_enter_read(&pool_lock);
1563 	s = splvm(); /* XXX go to splvm until all pools _setipl properly */
1564 	SIMPLEQ_FOREACH(pp, &pool_head, pr_poollist) {
1565 #ifdef MULTIPROCESSOR
1566 		if (pp->pr_cache != NULL)
1567 			pool_cache_gc(pp);
1568 #endif
1569 
1570 		if (pp->pr_nidle <= pp->pr_minpages || /* guess */
1571 		    !pl_enter_try(pp, &pp->pr_lock)) /* try */
1572 			continue;
1573 
1574 		/* is it time to free a page? */
1575 		if (pp->pr_nidle > pp->pr_minpages &&
1576 		    (ph = TAILQ_FIRST(&pp->pr_emptypages)) != NULL &&
1577 		    (ticks - ph->ph_tick) > (hz * pool_wait_gc)) {
1578 			freeph = ph;
1579 			pool_p_remove(pp, freeph);
1580 		} else
1581 			freeph = NULL;
1582 
1583 		pl_leave(pp, &pp->pr_lock);
1584 
1585 		if (freeph != NULL)
1586 			pool_p_free(pp, freeph);
1587 	}
1588 	splx(s);
1589 	rw_exit_read(&pool_lock);
1590 
1591 	timeout_add_sec(&pool_gc_tick, 1);
1592 }
1593 
1594 /*
1595  * Pool backend allocators.
1596  */
1597 
1598 void *
1599 pool_allocator_alloc(struct pool *pp, int flags, int *slowdown)
1600 {
1601 	void *v;
1602 
1603 	v = (*pp->pr_alloc->pa_alloc)(pp, flags, slowdown);
1604 
1605 #ifdef DIAGNOSTIC
1606 	if (v != NULL && POOL_INPGHDR(pp)) {
1607 		vaddr_t addr = (vaddr_t)v;
1608 		if ((addr & pp->pr_pgmask) != addr) {
1609 			panic("%s: %s page address %p isnt aligned to %u",
1610 			    __func__, pp->pr_wchan, v, pp->pr_pgsize);
1611 		}
1612 	}
1613 #endif
1614 
1615 	return (v);
1616 }
1617 
1618 void
1619 pool_allocator_free(struct pool *pp, void *v)
1620 {
1621 	struct pool_allocator *pa = pp->pr_alloc;
1622 
1623 	(*pa->pa_free)(pp, v);
1624 }
1625 
1626 void *
1627 pool_page_alloc(struct pool *pp, int flags, int *slowdown)
1628 {
1629 	struct kmem_dyn_mode kd = KMEM_DYN_INITIALIZER;
1630 
1631 	kd.kd_waitok = ISSET(flags, PR_WAITOK);
1632 	kd.kd_slowdown = slowdown;
1633 
1634 	return (km_alloc(pp->pr_pgsize, &kv_page, pp->pr_crange, &kd));
1635 }
1636 
1637 void
1638 pool_page_free(struct pool *pp, void *v)
1639 {
1640 	km_free(v, pp->pr_pgsize, &kv_page, pp->pr_crange);
1641 }
1642 
1643 void *
1644 pool_multi_alloc(struct pool *pp, int flags, int *slowdown)
1645 {
1646 	struct kmem_va_mode kv = kv_intrsafe;
1647 	struct kmem_dyn_mode kd = KMEM_DYN_INITIALIZER;
1648 	void *v;
1649 	int s;
1650 
1651 	if (POOL_INPGHDR(pp))
1652 		kv.kv_align = pp->pr_pgsize;
1653 
1654 	kd.kd_waitok = ISSET(flags, PR_WAITOK);
1655 	kd.kd_slowdown = slowdown;
1656 
1657 	s = splvm();
1658 	v = km_alloc(pp->pr_pgsize, &kv, pp->pr_crange, &kd);
1659 	splx(s);
1660 
1661 	return (v);
1662 }
1663 
1664 void
1665 pool_multi_free(struct pool *pp, void *v)
1666 {
1667 	struct kmem_va_mode kv = kv_intrsafe;
1668 	int s;
1669 
1670 	if (POOL_INPGHDR(pp))
1671 		kv.kv_align = pp->pr_pgsize;
1672 
1673 	s = splvm();
1674 	km_free(v, pp->pr_pgsize, &kv, pp->pr_crange);
1675 	splx(s);
1676 }
1677 
1678 void *
1679 pool_multi_alloc_ni(struct pool *pp, int flags, int *slowdown)
1680 {
1681 	struct kmem_va_mode kv = kv_any;
1682 	struct kmem_dyn_mode kd = KMEM_DYN_INITIALIZER;
1683 	void *v;
1684 
1685 	if (POOL_INPGHDR(pp))
1686 		kv.kv_align = pp->pr_pgsize;
1687 
1688 	kd.kd_waitok = ISSET(flags, PR_WAITOK);
1689 	kd.kd_slowdown = slowdown;
1690 
1691 	KERNEL_LOCK();
1692 	v = km_alloc(pp->pr_pgsize, &kv, pp->pr_crange, &kd);
1693 	KERNEL_UNLOCK();
1694 
1695 	return (v);
1696 }
1697 
1698 void
1699 pool_multi_free_ni(struct pool *pp, void *v)
1700 {
1701 	struct kmem_va_mode kv = kv_any;
1702 
1703 	if (POOL_INPGHDR(pp))
1704 		kv.kv_align = pp->pr_pgsize;
1705 
1706 	KERNEL_LOCK();
1707 	km_free(v, pp->pr_pgsize, &kv, pp->pr_crange);
1708 	KERNEL_UNLOCK();
1709 }
1710 
1711 #ifdef MULTIPROCESSOR
1712 
1713 struct pool pool_caches; /* per cpu cache entries */
1714 
1715 void
1716 pool_cache_init(struct pool *pp)
1717 {
1718 	struct cpumem *cm;
1719 	struct pool_cache *pc;
1720 	struct cpumem_iter i;
1721 
1722 	if (pool_caches.pr_size == 0) {
1723 		pool_init(&pool_caches, sizeof(struct pool_cache),
1724 		    CACHELINESIZE, IPL_NONE, PR_WAITOK | PR_RWLOCK,
1725 		    "plcache", NULL);
1726 	}
1727 
1728 	/* must be able to use the pool items as cache list items */
1729 	KASSERT(pp->pr_size >= sizeof(struct pool_cache_item));
1730 
1731 	cm = cpumem_get(&pool_caches);
1732 
1733 	pl_init(pp, &pp->pr_cache_lock);
1734 	arc4random_buf(pp->pr_cache_magic, sizeof(pp->pr_cache_magic));
1735 	TAILQ_INIT(&pp->pr_cache_lists);
1736 	pp->pr_cache_nitems = 0;
1737 	pp->pr_cache_tick = ticks;
1738 	pp->pr_cache_items = 8;
1739 	pp->pr_cache_contention = 0;
1740 	pp->pr_cache_ngc = 0;
1741 
1742 	CPUMEM_FOREACH(pc, &i, cm) {
1743 		pc->pc_actv = NULL;
1744 		pc->pc_nactv = 0;
1745 		pc->pc_prev = NULL;
1746 
1747 		pc->pc_nget = 0;
1748 		pc->pc_nfail = 0;
1749 		pc->pc_nput = 0;
1750 		pc->pc_nlget = 0;
1751 		pc->pc_nlfail = 0;
1752 		pc->pc_nlput = 0;
1753 		pc->pc_nout = 0;
1754 	}
1755 
1756 	membar_producer();
1757 
1758 	pp->pr_cache = cm;
1759 }
1760 
1761 static inline void
1762 pool_cache_item_magic(struct pool *pp, struct pool_cache_item *ci)
1763 {
1764 	unsigned long *entry = (unsigned long *)&ci->ci_nextl;
1765 
1766 	entry[0] = pp->pr_cache_magic[0] ^ (u_long)ci;
1767 	entry[1] = pp->pr_cache_magic[1] ^ (u_long)ci->ci_next;
1768 }
1769 
1770 static inline void
1771 pool_cache_item_magic_check(struct pool *pp, struct pool_cache_item *ci)
1772 {
1773 	unsigned long *entry;
1774 	unsigned long val;
1775 
1776 	entry = (unsigned long *)&ci->ci_nextl;
1777 	val = pp->pr_cache_magic[0] ^ (u_long)ci;
1778 	if (*entry != val)
1779 		goto fail;
1780 
1781 	entry++;
1782 	val = pp->pr_cache_magic[1] ^ (u_long)ci->ci_next;
1783 	if (*entry != val)
1784 		goto fail;
1785 
1786 	return;
1787 
1788 fail:
1789 	panic("%s: %s cpu free list modified: item addr %p+%zu 0x%lx!=0x%lx",
1790 	    __func__, pp->pr_wchan, ci, (caddr_t)entry - (caddr_t)ci,
1791 	    *entry, val);
1792 }
1793 
1794 static inline void
1795 pool_list_enter(struct pool *pp)
1796 {
1797 	if (pl_enter_try(pp, &pp->pr_cache_lock) == 0) {
1798 		pl_enter(pp, &pp->pr_cache_lock);
1799 		pp->pr_cache_contention++;
1800 	}
1801 }
1802 
1803 static inline void
1804 pool_list_leave(struct pool *pp)
1805 {
1806 	pl_leave(pp, &pp->pr_cache_lock);
1807 }
1808 
1809 static inline struct pool_cache_item *
1810 pool_cache_list_alloc(struct pool *pp, struct pool_cache *pc)
1811 {
1812 	struct pool_cache_item *pl;
1813 
1814 	pool_list_enter(pp);
1815 	pl = TAILQ_FIRST(&pp->pr_cache_lists);
1816 	if (pl != NULL) {
1817 		TAILQ_REMOVE(&pp->pr_cache_lists, pl, ci_nextl);
1818 		pp->pr_cache_nitems -= POOL_CACHE_ITEM_NITEMS(pl);
1819 
1820 		pool_cache_item_magic(pp, pl);
1821 
1822 		pc->pc_nlget++;
1823 	} else
1824 		pc->pc_nlfail++;
1825 
1826 	/* fold this cpus nout into the global while we have the lock */
1827 	pp->pr_cache_nout += pc->pc_nout;
1828 	pc->pc_nout = 0;
1829 	pool_list_leave(pp);
1830 
1831 	return (pl);
1832 }
1833 
1834 static inline void
1835 pool_cache_list_free(struct pool *pp, struct pool_cache *pc,
1836     struct pool_cache_item *ci)
1837 {
1838 	pool_list_enter(pp);
1839 	if (TAILQ_EMPTY(&pp->pr_cache_lists))
1840 		pp->pr_cache_tick = ticks;
1841 
1842 	pp->pr_cache_nitems += POOL_CACHE_ITEM_NITEMS(ci);
1843 	TAILQ_INSERT_TAIL(&pp->pr_cache_lists, ci, ci_nextl);
1844 
1845 	pc->pc_nlput++;
1846 
1847 	/* fold this cpus nout into the global while we have the lock */
1848 	pp->pr_cache_nout += pc->pc_nout;
1849 	pc->pc_nout = 0;
1850 	pool_list_leave(pp);
1851 }
1852 
1853 static inline struct pool_cache *
1854 pool_cache_enter(struct pool *pp, int *s)
1855 {
1856 	struct pool_cache *pc;
1857 
1858 	pc = cpumem_enter(pp->pr_cache);
1859 	*s = splraise(pp->pr_ipl);
1860 	pc->pc_gen++;
1861 
1862 	return (pc);
1863 }
1864 
1865 static inline void
1866 pool_cache_leave(struct pool *pp, struct pool_cache *pc, int s)
1867 {
1868 	pc->pc_gen++;
1869 	splx(s);
1870 	cpumem_leave(pp->pr_cache, pc);
1871 }
1872 
1873 void *
1874 pool_cache_get(struct pool *pp)
1875 {
1876 	struct pool_cache *pc;
1877 	struct pool_cache_item *ci;
1878 	int s;
1879 
1880 	pc = pool_cache_enter(pp, &s);
1881 
1882 	if (pc->pc_actv != NULL) {
1883 		ci = pc->pc_actv;
1884 	} else if (pc->pc_prev != NULL) {
1885 		ci = pc->pc_prev;
1886 		pc->pc_prev = NULL;
1887 	} else if ((ci = pool_cache_list_alloc(pp, pc)) == NULL) {
1888 		pc->pc_nfail++;
1889 		goto done;
1890 	}
1891 
1892 	pool_cache_item_magic_check(pp, ci);
1893 #ifdef DIAGNOSTIC
1894 	if (pool_debug && POOL_CACHE_ITEM_POISONED(ci)) {
1895 		size_t pidx;
1896 		uint32_t pval;
1897 
1898 		if (poison_check(ci + 1, pp->pr_size - sizeof(*ci),
1899 		    &pidx, &pval)) {
1900 			int *ip = (int *)(ci + 1);
1901 			ip += pidx;
1902 
1903 			panic("%s: %s cpu free list modified: "
1904 			    "item addr %p+%zu 0x%x!=0x%x",
1905 			    __func__, pp->pr_wchan, ci,
1906 			    (caddr_t)ip - (caddr_t)ci, *ip, pval);
1907 		}
1908 	}
1909 #endif
1910 
1911 	pc->pc_actv = ci->ci_next;
1912 	pc->pc_nactv = POOL_CACHE_ITEM_NITEMS(ci) - 1;
1913 	pc->pc_nget++;
1914 	pc->pc_nout++;
1915 
1916 done:
1917 	pool_cache_leave(pp, pc, s);
1918 
1919 	return (ci);
1920 }
1921 
1922 void
1923 pool_cache_put(struct pool *pp, void *v)
1924 {
1925 	struct pool_cache *pc;
1926 	struct pool_cache_item *ci = v;
1927 	unsigned long nitems;
1928 	int s;
1929 #ifdef DIAGNOSTIC
1930 	int poison = pool_debug && pp->pr_size > sizeof(*ci);
1931 
1932 	if (poison)
1933 		poison_mem(ci + 1, pp->pr_size - sizeof(*ci));
1934 #endif
1935 
1936 	pc = pool_cache_enter(pp, &s);
1937 
1938 	nitems = pc->pc_nactv;
1939 	if (nitems >= pp->pr_cache_items) {
1940 		if (pc->pc_prev != NULL)
1941 			pool_cache_list_free(pp, pc, pc->pc_prev);
1942 
1943 		pc->pc_prev = pc->pc_actv;
1944 
1945 		pc->pc_actv = NULL;
1946 		pc->pc_nactv = 0;
1947 		nitems = 0;
1948 	}
1949 
1950 	ci->ci_next = pc->pc_actv;
1951 	ci->ci_nitems = ++nitems;
1952 #ifdef DIAGNOSTIC
1953 	ci->ci_nitems |= poison ? POOL_CACHE_ITEM_NITEMS_POISON : 0;
1954 #endif
1955 	pool_cache_item_magic(pp, ci);
1956 
1957 	pc->pc_actv = ci;
1958 	pc->pc_nactv = nitems;
1959 
1960 	pc->pc_nput++;
1961 	pc->pc_nout--;
1962 
1963 	pool_cache_leave(pp, pc, s);
1964 }
1965 
1966 struct pool_cache_item *
1967 pool_cache_list_put(struct pool *pp, struct pool_cache_item *pl)
1968 {
1969 	struct pool_cache_item *rpl, *next;
1970 
1971 	if (pl == NULL)
1972 		return (NULL);
1973 
1974 	rpl = TAILQ_NEXT(pl, ci_nextl);
1975 
1976 	pl_enter(pp, &pp->pr_lock);
1977 	do {
1978 		next = pl->ci_next;
1979 		pool_do_put(pp, pl);
1980 		pl = next;
1981 	} while (pl != NULL);
1982 	pl_leave(pp, &pp->pr_lock);
1983 
1984 	return (rpl);
1985 }
1986 
1987 void
1988 pool_cache_destroy(struct pool *pp)
1989 {
1990 	struct pool_cache *pc;
1991 	struct pool_cache_item *pl;
1992 	struct cpumem_iter i;
1993 	struct cpumem *cm;
1994 
1995 	rw_enter_write(&pool_lock); /* serialise with the gc */
1996 	cm = pp->pr_cache;
1997 	pp->pr_cache = NULL; /* make pool_put avoid the cache */
1998 	rw_exit_write(&pool_lock);
1999 
2000 	CPUMEM_FOREACH(pc, &i, cm) {
2001 		pool_cache_list_put(pp, pc->pc_actv);
2002 		pool_cache_list_put(pp, pc->pc_prev);
2003 	}
2004 
2005 	cpumem_put(&pool_caches, cm);
2006 
2007 	pl = TAILQ_FIRST(&pp->pr_cache_lists);
2008 	while (pl != NULL)
2009 		pl = pool_cache_list_put(pp, pl);
2010 }
2011 
2012 void
2013 pool_cache_gc(struct pool *pp)
2014 {
2015 	unsigned int contention, delta;
2016 
2017 	if ((ticks - pp->pr_cache_tick) > (hz * pool_wait_gc) &&
2018 	    !TAILQ_EMPTY(&pp->pr_cache_lists) &&
2019 	    pl_enter_try(pp, &pp->pr_cache_lock)) {
2020 		struct pool_cache_item *pl = NULL;
2021 
2022 		pl = TAILQ_FIRST(&pp->pr_cache_lists);
2023 		if (pl != NULL) {
2024 			TAILQ_REMOVE(&pp->pr_cache_lists, pl, ci_nextl);
2025 			pp->pr_cache_nitems -= POOL_CACHE_ITEM_NITEMS(pl);
2026 			pp->pr_cache_tick = ticks;
2027 
2028 			pp->pr_cache_ngc++;
2029 		}
2030 
2031 		pl_leave(pp, &pp->pr_cache_lock);
2032 
2033 		pool_cache_list_put(pp, pl);
2034 	}
2035 
2036 	/*
2037 	 * if there's a lot of contention on the pr_cache_mtx then consider
2038 	 * growing the length of the list to reduce the need to access the
2039 	 * global pool.
2040 	 */
2041 
2042 	contention = pp->pr_cache_contention;
2043 	delta = contention - pp->pr_cache_contention_prev;
2044 	if (delta > 8 /* magic */) {
2045 		if ((ncpusfound * 8 * 2) <= pp->pr_cache_nitems)
2046 			pp->pr_cache_items += 8;
2047 	} else if (delta == 0) {
2048 		if (pp->pr_cache_items > 8)
2049 			pp->pr_cache_items--;
2050 	}
2051 	pp->pr_cache_contention_prev = contention;
2052 }
2053 
2054 void
2055 pool_cache_pool_info(struct pool *pp, struct kinfo_pool *pi)
2056 {
2057 	struct pool_cache *pc;
2058 	struct cpumem_iter i;
2059 
2060 	if (pp->pr_cache == NULL)
2061 		return;
2062 
2063 	/* loop through the caches twice to collect stats */
2064 
2065 	/* once without the lock so we can yield while reading nget/nput */
2066 	CPUMEM_FOREACH(pc, &i, pp->pr_cache) {
2067 		uint64_t gen, nget, nput;
2068 
2069 		do {
2070 			while ((gen = pc->pc_gen) & 1)
2071 				yield();
2072 
2073 			nget = pc->pc_nget;
2074 			nput = pc->pc_nput;
2075 		} while (gen != pc->pc_gen);
2076 
2077 		pi->pr_nget += nget;
2078 		pi->pr_nput += nput;
2079 	}
2080 
2081 	/* and once with the mtx so we can get consistent nout values */
2082 	pl_enter(pp, &pp->pr_cache_lock);
2083 	CPUMEM_FOREACH(pc, &i, pp->pr_cache)
2084 		pi->pr_nout += pc->pc_nout;
2085 
2086 	pi->pr_nout += pp->pr_cache_nout;
2087 	pl_leave(pp, &pp->pr_cache_lock);
2088 }
2089 
2090 int
2091 pool_cache_info(struct pool *pp, void *oldp, size_t *oldlenp)
2092 {
2093 	struct kinfo_pool_cache kpc;
2094 
2095 	if (pp->pr_cache == NULL)
2096 		return (EOPNOTSUPP);
2097 
2098 	memset(&kpc, 0, sizeof(kpc)); /* don't leak padding */
2099 
2100 	pl_enter(pp, &pp->pr_cache_lock);
2101 	kpc.pr_ngc = pp->pr_cache_ngc;
2102 	kpc.pr_len = pp->pr_cache_items;
2103 	kpc.pr_nitems = pp->pr_cache_nitems;
2104 	kpc.pr_contention = pp->pr_cache_contention;
2105 	pl_leave(pp, &pp->pr_cache_lock);
2106 
2107 	return (sysctl_rdstruct(oldp, oldlenp, NULL, &kpc, sizeof(kpc)));
2108 }
2109 
2110 int
2111 pool_cache_cpus_info(struct pool *pp, void *oldp, size_t *oldlenp)
2112 {
2113 	struct pool_cache *pc;
2114 	struct kinfo_pool_cache_cpu *kpcc, *info;
2115 	unsigned int cpu = 0;
2116 	struct cpumem_iter i;
2117 	int error = 0;
2118 	size_t len;
2119 
2120 	if (pp->pr_cache == NULL)
2121 		return (EOPNOTSUPP);
2122 	if (*oldlenp % sizeof(*kpcc))
2123 		return (EINVAL);
2124 
2125 	kpcc = mallocarray(ncpusfound, sizeof(*kpcc), M_TEMP,
2126 	    M_WAITOK|M_CANFAIL|M_ZERO);
2127 	if (kpcc == NULL)
2128 		return (EIO);
2129 
2130 	len = ncpusfound * sizeof(*kpcc);
2131 
2132 	CPUMEM_FOREACH(pc, &i, pp->pr_cache) {
2133 		uint64_t gen;
2134 
2135 		if (cpu >= ncpusfound) {
2136 			error = EIO;
2137 			goto err;
2138 		}
2139 
2140 		info = &kpcc[cpu];
2141 		info->pr_cpu = cpu;
2142 
2143 		do {
2144 			while ((gen = pc->pc_gen) & 1)
2145 				yield();
2146 
2147 			info->pr_nget = pc->pc_nget;
2148 			info->pr_nfail = pc->pc_nfail;
2149 			info->pr_nput = pc->pc_nput;
2150 			info->pr_nlget = pc->pc_nlget;
2151 			info->pr_nlfail = pc->pc_nlfail;
2152 			info->pr_nlput = pc->pc_nlput;
2153 		} while (gen != pc->pc_gen);
2154 
2155 		cpu++;
2156 	}
2157 
2158 	error = sysctl_rdstruct(oldp, oldlenp, NULL, kpcc, len);
2159 err:
2160 	free(kpcc, M_TEMP, len);
2161 
2162 	return (error);
2163 }
2164 #else /* MULTIPROCESSOR */
2165 void
2166 pool_cache_init(struct pool *pp)
2167 {
2168 	/* nop */
2169 }
2170 
2171 void
2172 pool_cache_pool_info(struct pool *pp, struct kinfo_pool *pi)
2173 {
2174 	/* nop */
2175 }
2176 
2177 int
2178 pool_cache_info(struct pool *pp, void *oldp, size_t *oldlenp)
2179 {
2180 	return (EOPNOTSUPP);
2181 }
2182 
2183 int
2184 pool_cache_cpus_info(struct pool *pp, void *oldp, size_t *oldlenp)
2185 {
2186 	return (EOPNOTSUPP);
2187 }
2188 #endif /* MULTIPROCESSOR */
2189 
2190 
2191 void
2192 pool_lock_mtx_init(struct pool *pp, union pool_lock *lock,
2193     const struct lock_type *type)
2194 {
2195 	_mtx_init_flags(&lock->prl_mtx, pp->pr_ipl, pp->pr_wchan, 0, type);
2196 }
2197 
2198 void
2199 pool_lock_mtx_enter(union pool_lock *lock)
2200 {
2201 	mtx_enter(&lock->prl_mtx);
2202 }
2203 
2204 int
2205 pool_lock_mtx_enter_try(union pool_lock *lock)
2206 {
2207 	return (mtx_enter_try(&lock->prl_mtx));
2208 }
2209 
2210 void
2211 pool_lock_mtx_leave(union pool_lock *lock)
2212 {
2213 	mtx_leave(&lock->prl_mtx);
2214 }
2215 
2216 void
2217 pool_lock_mtx_assert_locked(union pool_lock *lock)
2218 {
2219 	MUTEX_ASSERT_LOCKED(&lock->prl_mtx);
2220 }
2221 
2222 void
2223 pool_lock_mtx_assert_unlocked(union pool_lock *lock)
2224 {
2225 	MUTEX_ASSERT_UNLOCKED(&lock->prl_mtx);
2226 }
2227 
2228 int
2229 pool_lock_mtx_sleep(void *ident, union pool_lock *lock, int priority,
2230     const char *wmesg, int timo)
2231 {
2232 	return msleep(ident, &lock->prl_mtx, priority, wmesg, timo);
2233 }
2234 
2235 static const struct pool_lock_ops pool_lock_ops_mtx = {
2236 	pool_lock_mtx_init,
2237 	pool_lock_mtx_enter,
2238 	pool_lock_mtx_enter_try,
2239 	pool_lock_mtx_leave,
2240 	pool_lock_mtx_assert_locked,
2241 	pool_lock_mtx_assert_unlocked,
2242 	pool_lock_mtx_sleep,
2243 };
2244 
2245 void
2246 pool_lock_rw_init(struct pool *pp, union pool_lock *lock,
2247     const struct lock_type *type)
2248 {
2249 	_rw_init_flags(&lock->prl_rwlock, pp->pr_wchan, 0, type);
2250 }
2251 
2252 void
2253 pool_lock_rw_enter(union pool_lock *lock)
2254 {
2255 	rw_enter_write(&lock->prl_rwlock);
2256 }
2257 
2258 int
2259 pool_lock_rw_enter_try(union pool_lock *lock)
2260 {
2261 	return (rw_enter(&lock->prl_rwlock, RW_WRITE | RW_NOSLEEP) == 0);
2262 }
2263 
2264 void
2265 pool_lock_rw_leave(union pool_lock *lock)
2266 {
2267 	rw_exit_write(&lock->prl_rwlock);
2268 }
2269 
2270 void
2271 pool_lock_rw_assert_locked(union pool_lock *lock)
2272 {
2273 	rw_assert_wrlock(&lock->prl_rwlock);
2274 }
2275 
2276 void
2277 pool_lock_rw_assert_unlocked(union pool_lock *lock)
2278 {
2279 	KASSERT(rw_status(&lock->prl_rwlock) != RW_WRITE);
2280 }
2281 
2282 int
2283 pool_lock_rw_sleep(void *ident, union pool_lock *lock, int priority,
2284     const char *wmesg, int timo)
2285 {
2286 	return rwsleep(ident, &lock->prl_rwlock, priority, wmesg, timo);
2287 }
2288 
2289 static const struct pool_lock_ops pool_lock_ops_rw = {
2290 	pool_lock_rw_init,
2291 	pool_lock_rw_enter,
2292 	pool_lock_rw_enter_try,
2293 	pool_lock_rw_leave,
2294 	pool_lock_rw_assert_locked,
2295 	pool_lock_rw_assert_unlocked,
2296 	pool_lock_rw_sleep,
2297 };
2298