Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.

===================================================================
RCS file: /ftp/cvs/cvsroot/src/sys/kern/subr_pool.c,v
rcsdiff: /ftp/cvs/cvsroot/src/sys/kern/subr_pool.c,v: warning: Unknown phrases like `commitid ...;' are present.
retrieving revision 1.30.2.2
retrieving revision 1.76.6.1
diff -u -p -r1.30.2.2 -r1.76.6.1
--- src/sys/kern/subr_pool.c	2000/11/22 16:05:24	1.30.2.2
+++ src/sys/kern/subr_pool.c	2002/11/11 23:31:43	1.76.6.1
@@ -1,7 +1,7 @@
-/*	$NetBSD: subr_pool.c,v 1.30.2.2 2000/11/22 16:05:24 bouyer Exp $	*/
+/*	$NetBSD: subr_pool.c,v 1.76.6.1 2002/11/11 23:31:43 he Exp $	*/
 
 /*-
- * Copyright (c) 1997, 1999 The NetBSD Foundation, Inc.
+ * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc.
  * All rights reserved.
  *
  * This code is derived from software contributed to The NetBSD Foundation
@@ -37,6 +37,9 @@
  * POSSIBILITY OF SUCH DAMAGE.
  */
 
+#include <sys/cdefs.h>
+__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.76.6.1 2002/11/11 23:31:43 he Exp $");
+
 #include "opt_pool.h"
 #include "opt_poollog.h"
 #include "opt_lockdebug.h"
@@ -70,6 +73,11 @@ TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD
 /* Private pool for page header structures */
 static struct pool phpool;
 
+#ifdef POOL_SUBPAGE
+/* Pool of subpages for use by normal pools. */
+static struct pool psppool;
+#endif
+
 /* # of seconds to retain page after last use */
 int pool_inactive_time = 10;
 
@@ -90,6 +98,7 @@ struct pool_item_header {
 	caddr_t			ph_page;	/* this page's address */
 	struct timeval		ph_time;	/* last referenced */
 };
+TAILQ_HEAD(pool_pagelist,pool_item_header);
 
 struct pool_item {
 #ifdef DIAGNOSTIC
@@ -100,25 +109,56 @@ struct pool_item {
 	TAILQ_ENTRY(pool_item)	pi_list;
 };
 
-
 #define	PR_HASH_INDEX(pp,addr) \
-	(((u_long)(addr) >> (pp)->pr_pageshift) & (PR_HASHTABSIZE - 1))
+	(((u_long)(addr) >> (pp)->pr_alloc->pa_pageshift) & \
+	 (PR_HASHTABSIZE - 1))
+
+#define	POOL_NEEDS_CATCHUP(pp)						\
+	((pp)->pr_nitems < (pp)->pr_minitems)
+
+/*
+ * Pool cache management.
+ *
+ * Pool caches provide a way for constructed objects to be cached by the
+ * pool subsystem.  This can lead to performance improvements by avoiding
+ * needless object construction/destruction; it is deferred until absolutely
+ * necessary.
+ *
+ * Caches are grouped into cache groups.  Each cache group references
+ * up to 16 constructed objects.  When a cache allocates an object
+ * from the pool, it calls the object's constructor and places it into
+ * a cache group.  When a cache group frees an object back to the pool,
+ * it first calls the object's destructor.  This allows the object to
+ * persist in constructed form while freed to the cache.
+ *
+ * Multiple caches may exist for each pool.  This allows a single
+ * object type to have multiple constructed forms.  The pool references
+ * each cache, so that when a pool is drained by the pagedaemon, it can
+ * drain each individual cache as well.  Each time a cache is drained,
+ * the most idle cache group is freed to the pool in its entirety.
+ *
+ * Pool caches are layed on top of pools.  By layering them, we can avoid
+ * the complexity of cache management for pools which would not benefit
+ * from it.
+ */
 
+/* The cache group pool. */
+static struct pool pcgpool;
 
+static void	pool_cache_reclaim(struct pool_cache *);
 
-static struct pool_item_header
-		*pr_find_pagehead __P((struct pool *, caddr_t));
-static void	pr_rmpage __P((struct pool *, struct pool_item_header *));
-static int	pool_catchup __P((struct pool *));
-static void	pool_prime_page __P((struct pool *, caddr_t));
-static void	*pool_page_alloc __P((unsigned long, int, int));
-static void	pool_page_free __P((void *, unsigned long, int));
+static int	pool_catchup(struct pool *);
+static void	pool_prime_page(struct pool *, caddr_t,
+		    struct pool_item_header *);
 
-static void pool_print1 __P((struct pool *, const char *,
-	void (*)(const char *, ...)));
+void		*pool_allocator_alloc(struct pool *, int);
+void		pool_allocator_free(struct pool *, void *);
+
+static void pool_print1(struct pool *, const char *,
+	void (*)(const char *, ...));
 
 /*
- * Pool log entry. An array of these is allocated in pool_create().
+ * Pool log entry. An array of these is allocated in pool_init().
  */
 struct pool_log {
 	const char	*pl_file;
@@ -136,22 +176,9 @@ struct pool_log {
 
 int pool_logsize = POOL_LOGSIZE;
 
-#ifdef DIAGNOSTIC
-static void	pr_log __P((struct pool *, void *, int, const char *, long));
-static void	pr_printlog __P((struct pool *, struct pool_item *,
-		    void (*)(const char *, ...)));
-static void	pr_enter __P((struct pool *, const char *, long));
-static void	pr_leave __P((struct pool *));
-static void	pr_enter_check __P((struct pool *,
-		    void (*)(const char *, ...)));
-
-static __inline__ void
-pr_log(pp, v, action, file, line)
-	struct pool	*pp;
-	void		*v;
-	int		action;
-	const char	*file;
-	long		line;
+#ifdef POOL_DIAGNOSTIC
+static __inline void
+pr_log(struct pool *pp, void *v, int action, const char *file, long line)
 {
 	int n = pp->pr_curlogentry;
 	struct pool_log *pl;
@@ -174,10 +201,8 @@ pr_log(pp, v, action, file, line)
 }
 
 static void
-pr_printlog(pp, pi, pr)
-	struct pool *pp;
-	struct pool_item *pi;
-	void (*pr) __P((const char *, ...));
+pr_printlog(struct pool *pp, struct pool_item *pi,
+    void (*pr)(const char *, ...))
 {
 	int i = pp->pr_logsize;
 	int n = pp->pr_curlogentry;
@@ -205,11 +230,8 @@ pr_printlog(pp, pi, pr)
 	}
 }
 
-static __inline__ void
-pr_enter(pp, file, line)
-	struct pool *pp;
-	const char *file;
-	long line;
+static __inline void
+pr_enter(struct pool *pp, const char *file, long line)
 {
 
 	if (__predict_false(pp->pr_entered_file != NULL)) {
@@ -224,9 +246,8 @@ pr_enter(pp, file, line)
 	pp->pr_entered_line = line;
 }
 
-static __inline__ void
-pr_leave(pp)
-	struct pool *pp;
+static __inline void
+pr_leave(struct pool *pp)
 {
 
 	if (__predict_false(pp->pr_entered_file == NULL)) {
@@ -238,10 +259,8 @@ pr_leave(pp)
 	pp->pr_entered_line = 0;
 }
 
-static __inline__ void
-pr_enter_check(pp, pr)
-	struct pool *pp;
-	void (*pr) __P((const char *, ...));
+static __inline void
+pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
 {
 
 	if (pp->pr_entered_file != NULL)
@@ -254,15 +273,13 @@ pr_enter_check(pp, pr)
 #define	pr_enter(pp, file, line)
 #define	pr_leave(pp)
 #define	pr_enter_check(pp, pr)
-#endif /* DIAGNOSTIC */
+#endif /* POOL_DIAGNOSTIC */
 
 /*
  * Return the pool page header based on page address.
  */
-static __inline__ struct pool_item_header *
-pr_find_pagehead(pp, page)
-	struct pool *pp;
-	caddr_t page;
+static __inline struct pool_item_header *
+pr_find_pagehead(struct pool *pp, caddr_t page)
 {
 	struct pool_item_header *ph;
 
@@ -281,11 +298,11 @@ pr_find_pagehead(pp, page)
 /*
  * Remove a page from the pool.
  */
-static __inline__ void
-pr_rmpage(pp, ph)
-	struct pool *pp;
-	struct pool_item_header *ph;
+static __inline void
+pr_rmpage(struct pool *pp, struct pool_item_header *ph,
+     struct pool_pagelist *pq)
 {
+	int s;
 
 	/*
 	 * If the page was idle, decrement the idle page count.
@@ -303,29 +320,30 @@ pr_rmpage(pp, ph)
 	pp->pr_nitems -= pp->pr_itemsperpage;
 
 	/*
-	 * Unlink a page from the pool and release it.
+	 * Unlink a page from the pool and release it (or queue it for release).
 	 */
 	TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
-	(*pp->pr_free)(ph->ph_page, pp->pr_pagesz, pp->pr_mtype);
+	if (pq) {
+		TAILQ_INSERT_HEAD(pq, ph, ph_pagelist);
+	} else {
+		pool_allocator_free(pp, ph->ph_page);
+		if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
+			LIST_REMOVE(ph, ph_hashlist);
+			s = splhigh();
+			pool_put(&phpool, ph);
+			splx(s);
+		}
+	}
 	pp->pr_npages--;
 	pp->pr_npagefree++;
 
-	if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
-		int s;
-		LIST_REMOVE(ph, ph_hashlist);
-		s = splhigh();
-		pool_put(&phpool, ph);
-		splx(s);
-	}
-
 	if (pp->pr_curpage == ph) {
 		/*
 		 * Find a new non-empty page header, if any.
 		 * Start search from the page head, to increase the
 		 * chance for "high water" pages to be freed.
 		 */
-		for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
-		     ph = TAILQ_NEXT(ph, ph_pagelist))
+		TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist)
 			if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
 				break;
 
@@ -334,59 +352,14 @@ pr_rmpage(pp, ph)
 }
 
 /*
- * Allocate and initialize a pool.
- */
-struct pool *
-pool_create(size, align, ioff, nitems, wchan, pagesz, alloc, release, mtype)
-	size_t	size;
-	u_int	align;
-	u_int	ioff;
-	int	nitems;
-	const char *wchan;
-	size_t	pagesz;
-	void	*(*alloc) __P((unsigned long, int, int));
-	void	(*release) __P((void *, unsigned long, int));
-	int	mtype;
-{
-	struct pool *pp;
-	int flags;
-
-	pp = (struct pool *)malloc(sizeof(*pp), M_POOL, M_NOWAIT);
-	if (pp == NULL)
-		return (NULL);
-
-	flags = PR_FREEHEADER;
-	pool_init(pp, size, align, ioff, flags, wchan, pagesz,
-		  alloc, release, mtype);
-
-	if (nitems != 0) {
-		if (pool_prime(pp, nitems, NULL) != 0) {
-			pool_destroy(pp);
-			return (NULL);
-		}
-	}
-
-	return (pp);
-}
-
-/*
  * Initialize the given pool resource structure.
  *
  * We export this routine to allow other kernel parts to declare
  * static pools that must be initialized before malloc() is available.
  */
 void
-pool_init(pp, size, align, ioff, flags, wchan, pagesz, alloc, release, mtype)
-	struct pool	*pp;
-	size_t		size;
-	u_int		align;
-	u_int		ioff;
-	int		flags;
-	const char	*wchan;
-	size_t		pagesz;
-	void		*(*alloc) __P((unsigned long, int, int));
-	void		(*release) __P((void *, unsigned long, int));
-	int		mtype;
+pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
+    const char *wchan, struct pool_allocator *palloc)
 {
 	int off, slack, i;
 
@@ -398,23 +371,43 @@ pool_init(pp, size, align, ioff, flags, 
 		flags |= PR_LOGGING;
 #endif
 
+#ifdef POOL_SUBPAGE
+	/*
+	 * XXX We don't provide a real `nointr' back-end
+	 * yet; all sub-pages come from a kmem back-end.
+	 * maybe some day...
+	 */
+	if (palloc == NULL) {
+		extern struct pool_allocator pool_allocator_kmem_subpage;
+		palloc = &pool_allocator_kmem_subpage;
+	}
 	/*
-	 * Check arguments and construct default values.
+	 * We'll assume any user-specified back-end allocator
+	 * will deal with sub-pages, or simply don't care.
 	 */
-	if (!powerof2(pagesz))
-		panic("pool_init: page size invalid (%lx)\n", (u_long)pagesz);
+#else
+	if (palloc == NULL)
+		palloc = &pool_allocator_kmem;
+#endif /* POOL_SUBPAGE */
+	if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
+		if (palloc->pa_pagesz == 0) {
+#ifdef POOL_SUBPAGE
+			if (palloc == &pool_allocator_kmem)
+				palloc->pa_pagesz = PAGE_SIZE;
+			else
+				palloc->pa_pagesz = POOL_SUBPAGE;
+#else
+			palloc->pa_pagesz = PAGE_SIZE;
+#endif /* POOL_SUBPAGE */
+		}
+
+		TAILQ_INIT(&palloc->pa_list);
 
-	if (alloc == NULL && release == NULL) {
-		alloc = pool_page_alloc;
-		release = pool_page_free;
-		pagesz = PAGE_SIZE;	/* Rounds to PAGE_SIZE anyhow. */
-	} else if ((alloc != NULL && release != NULL) == 0) {
-		/* If you specifiy one, must specify both. */
-		panic("pool_init: must specify alloc and release together");
-	}
-			
-	if (pagesz == 0)
-		pagesz = PAGE_SIZE;
+		simple_lock_init(&palloc->pa_slock);
+		palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
+		palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
+		palloc->pa_flags |= PA_INITIALIZED;
+	}
 
 	if (align == 0)
 		align = ALIGN(1);
@@ -422,15 +415,18 @@ pool_init(pp, size, align, ioff, flags, 
 	if (size < sizeof(struct pool_item))
 		size = sizeof(struct pool_item);
 
-	size = ALIGN(size);
-	if (size >= pagesz)
+	size = roundup(size, align);
+#ifdef DIAGNOSTIC
+	if (size > palloc->pa_pagesz)
 		panic("pool_init: pool item size (%lu) too large",
 		      (u_long)size);
+#endif
 
 	/*
 	 * Initialize the pool structure.
 	 */
 	TAILQ_INIT(&pp->pr_pagelist);
+	TAILQ_INIT(&pp->pr_cachelist);
 	pp->pr_curpage = NULL;
 	pp->pr_npages = 0;
 	pp->pr_minitems = 0;
@@ -441,12 +437,7 @@ pool_init(pp, size, align, ioff, flags, 
 	pp->pr_size = size;
 	pp->pr_align = align;
 	pp->pr_wchan = wchan;
-	pp->pr_mtype = mtype;
-	pp->pr_alloc = alloc;
-	pp->pr_free = release;
-	pp->pr_pagesz = pagesz;
-	pp->pr_pagemask = ~(pagesz - 1);
-	pp->pr_pageshift = ffs(pagesz) - 1;
+	pp->pr_alloc = palloc;
 	pp->pr_nitems = 0;
 	pp->pr_nout = 0;
 	pp->pr_hardlimit = UINT_MAX;
@@ -455,6 +446,8 @@ pool_init(pp, size, align, ioff, flags, 
 	pp->pr_hardlimit_ratecap.tv_usec = 0;
 	pp->pr_hardlimit_warning_last.tv_sec = 0;
 	pp->pr_hardlimit_warning_last.tv_usec = 0;
+	pp->pr_drain_hook = NULL;
+	pp->pr_drain_hook_arg = NULL;
 
 	/*
 	 * Decide whether to put the page header off page to avoid
@@ -463,15 +456,15 @@ pool_init(pp, size, align, ioff, flags, 
 	 * with its header based on the page address.
 	 * We use 1/16 of the page size as the threshold (XXX: tune)
 	 */
-	if (pp->pr_size < pagesz/16) {
+	if (pp->pr_size < palloc->pa_pagesz/16) {
 		/* Use the end of the page for the page header */
 		pp->pr_roflags |= PR_PHINPAGE;
-		pp->pr_phoffset = off =
-			pagesz - ALIGN(sizeof(struct pool_item_header));
+		pp->pr_phoffset = off = palloc->pa_pagesz -
+		    ALIGN(sizeof(struct pool_item_header));
 	} else {
 		/* The page header will be taken from our page header pool */
 		pp->pr_phoffset = 0;
-		off = pagesz;
+		off = palloc->pa_pagesz;
 		for (i = 0; i < PR_HASHTABSIZE; i++) {
 			LIST_INIT(&pp->pr_hashtab[i]);
 		}
@@ -486,6 +479,7 @@ pool_init(pp, size, align, ioff, flags, 
 	 */
 	pp->pr_itemoffset = ioff = ioff % align;
 	pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
+	KASSERT(pp->pr_itemsperpage != 0);
 
 	/*
 	 * Use the slack between the chunks and the page header
@@ -503,6 +497,7 @@ pool_init(pp, size, align, ioff, flags, 
 	pp->pr_hiwat = 0;
 	pp->pr_nidle = 0;
 
+#ifdef POOL_DIAGNOSTIC
 	if (flags & PR_LOGGING) {
 		if (kmem_map == NULL ||
 		    (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
@@ -511,6 +506,7 @@ pool_init(pp, size, align, ioff, flags, 
 		pp->pr_curlogentry = 0;
 		pp->pr_logsize = pool_logsize;
 	}
+#endif
 
 	pp->pr_entered_file = NULL;
 	pp->pr_entered_line = 0;
@@ -518,28 +514,52 @@ pool_init(pp, size, align, ioff, flags, 
 	simple_lock_init(&pp->pr_slock);
 
 	/*
-	 * Initialize private page header pool if we haven't done so yet.
+	 * Initialize private page header pool and cache magazine pool if we
+	 * haven't done so yet.
 	 * XXX LOCKING.
 	 */
 	if (phpool.pr_size == 0) {
+#ifdef POOL_SUBPAGE
+		pool_init(&phpool, sizeof(struct pool_item_header), 0, 0, 0,
+		    "phpool", &pool_allocator_kmem);
+		pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
+		    PR_RECURSIVE, "psppool", &pool_allocator_kmem);
+#else
 		pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
-			  0, "phpool", 0, 0, 0, 0);
+		    0, "phpool", NULL);
+#endif
+		pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
+		    0, "pcgpool", NULL);
 	}
 
 	/* Insert into the list of all pools. */
 	simple_lock(&pool_head_slock);
 	TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
 	simple_unlock(&pool_head_slock);
+
+	/* Insert this into the list of pools using this allocator. */
+	simple_lock(&palloc->pa_slock);
+	TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
+	simple_unlock(&palloc->pa_slock);
 }
 
 /*
  * De-commision a pool resource.
  */
 void
-pool_destroy(pp)
-	struct pool *pp;
+pool_destroy(struct pool *pp)
 {
 	struct pool_item_header *ph;
+	struct pool_cache *pc;
+
+	/* Locking order: pool_allocator -> pool */
+	simple_lock(&pp->pr_alloc->pa_slock);
+	TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
+	simple_unlock(&pp->pr_alloc->pa_slock);
+
+	/* Destroy all caches for this pool. */
+	while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL)
+		pool_cache_destroy(pc);
 
 #ifdef DIAGNOSTIC
 	if (pp->pr_nout != 0) {
@@ -550,51 +570,80 @@ pool_destroy(pp)
 #endif
 
 	/* Remove all pages */
-	if ((pp->pr_roflags & PR_STATIC) == 0)
-		while ((ph = pp->pr_pagelist.tqh_first) != NULL)
-			pr_rmpage(pp, ph);
+	while ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL)
+		pr_rmpage(pp, ph, NULL);
 
 	/* Remove from global pool list */
 	simple_lock(&pool_head_slock);
 	TAILQ_REMOVE(&pool_head, pp, pr_poollist);
-	/* XXX Only clear this if we were drainpp? */
-	drainpp = NULL;
+	if (drainpp == pp) {
+		drainpp = NULL;
+	}
 	simple_unlock(&pool_head_slock);
 
+#ifdef POOL_DIAGNOSTIC
 	if ((pp->pr_roflags & PR_LOGGING) != 0)
 		free(pp->pr_log, M_TEMP);
+#endif
+}
+
+void
+pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
+{
 
-	if (pp->pr_roflags & PR_FREEHEADER)
-		free(pp, M_POOL);
+	/* XXX no locking -- must be used just after pool_init() */
+#ifdef DIAGNOSTIC
+	if (pp->pr_drain_hook != NULL)
+		panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
+#endif
+	pp->pr_drain_hook = fn;
+	pp->pr_drain_hook_arg = arg;
 }
 
+static __inline struct pool_item_header *
+pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
+{
+	struct pool_item_header *ph;
+	int s;
+
+	LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0);
+
+	if ((pp->pr_roflags & PR_PHINPAGE) != 0)
+		ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
+	else {
+		s = splhigh();
+		ph = pool_get(&phpool, flags);
+		splx(s);
+	}
+
+	return (ph);
+}
 
 /*
  * Grab an item from the pool; must be called at appropriate spl level
  */
 void *
-_pool_get(pp, flags, file, line)
-	struct pool *pp;
-	int flags;
-	const char *file;
-	long line;
+#ifdef POOL_DIAGNOSTIC
+_pool_get(struct pool *pp, int flags, const char *file, long line)
+#else
+pool_get(struct pool *pp, int flags)
+#endif
 {
-	void *v;
 	struct pool_item *pi;
 	struct pool_item_header *ph;
+	void *v;
 
 #ifdef DIAGNOSTIC
-	if (__predict_false((pp->pr_roflags & PR_STATIC) &&
-			    (flags & PR_MALLOCOK))) {
-		pr_printlog(pp, NULL, printf);
-		panic("pool_get: static");
-	}
-#endif
-
 	if (__predict_false(curproc == NULL && doing_shutdown == 0 &&
 			    (flags & PR_WAITOK) != 0))
 		panic("pool_get: must have NOWAIT");
 
+#ifdef LOCKDEBUG
+	if (flags & PR_WAITOK)
+		simple_lock_only_held(NULL, "pool_get(PR_WAITOK)");
+#endif
+#endif /* DIAGNOSTIC */
+
 	simple_lock(&pp->pr_slock);
 	pr_enter(pp, file, line);
 
@@ -612,6 +661,21 @@ _pool_get(pp, flags, file, line)
 	}
 #endif
 	if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
+		if (pp->pr_drain_hook != NULL) {
+			/*
+			 * Since the drain hook is going to free things
+			 * back to the pool, unlock, call the hook, re-lock,
+			 * and check the hardlimit condition again.
+			 */
+			pr_leave(pp);
+			simple_unlock(&pp->pr_slock);
+			(*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
+			simple_lock(&pp->pr_slock);
+			pr_enter(pp, file, line);
+			if (pp->pr_nout < pp->pr_hardlimit)
+				goto startover;
+		}
+
 		if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
 			/*
 			 * XXX: A warning isn't logged in this case.  Should
@@ -632,9 +696,6 @@ _pool_get(pp, flags, file, line)
 			      &pp->pr_hardlimit_ratecap))
 			log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
 
-		if (flags & PR_URGENT)
-			panic("pool_get: urgent");
-
 		pp->pr_nfail++;
 
 		pr_leave(pp);
@@ -649,8 +710,6 @@ _pool_get(pp, flags, file, line)
 	 * has no items in its bucket.
 	 */
 	if ((ph = pp->pr_curpage) == NULL) {
-		void *v;
-
 #ifdef DIAGNOSTIC
 		if (pp->pr_nitems != 0) {
 			simple_unlock(&pp->pr_slock);
@@ -667,23 +726,25 @@ _pool_get(pp, flags, file, line)
 		 */
 		pr_leave(pp);
 		simple_unlock(&pp->pr_slock);
-		v = (*pp->pr_alloc)(pp->pr_pagesz, flags, pp->pr_mtype);
+		v = pool_allocator_alloc(pp, flags);
+		if (__predict_true(v != NULL))
+			ph = pool_alloc_item_header(pp, v, flags);
 		simple_lock(&pp->pr_slock);
 		pr_enter(pp, file, line);
 
-		if (v == NULL) {
+		if (__predict_false(v == NULL || ph == NULL)) {
+			if (v != NULL)
+				pool_allocator_free(pp, v);
+
 			/*
-			 * We were unable to allocate a page, but
-			 * we released the lock during allocation,
-			 * so perhaps items were freed back to the
-			 * pool.  Check for this case.
+			 * We were unable to allocate a page or item
+			 * header, but we released the lock during
+			 * allocation, so perhaps items were freed
+			 * back to the pool.  Check for this case.
 			 */
 			if (pp->pr_curpage != NULL)
 				goto startover;
 
-			if (flags & PR_URGENT)
-				panic("pool_get: urgent");
-
 			if ((flags & PR_WAITOK) == 0) {
 				pp->pr_nfail++;
 				pr_leave(pp);
@@ -694,15 +755,11 @@ _pool_get(pp, flags, file, line)
 			/*
 			 * Wait for items to be returned to this pool.
 			 *
-			 * XXX: we actually want to wait just until
-			 * the page allocator has memory again. Depending
-			 * on this pool's usage, we might get stuck here
-			 * for a long time.
-			 *
 			 * XXX: maybe we should wake up once a second and
 			 * try again?
 			 */
 			pp->pr_flags |= PR_WANTED;
+			/* PA_WANTED is already set on the allocator. */
 			pr_leave(pp);
 			ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
 			pr_enter(pp, file, line);
@@ -710,8 +767,8 @@ _pool_get(pp, flags, file, line)
 		}
 
 		/* We have more memory; add it to the pool */
+		pool_prime_page(pp, v, ph);
 		pp->pr_npagealloc++;
-		pool_prime_page(pp, v);
 
 		/* Start the allocation process over. */
 		goto startover;
@@ -731,7 +788,10 @@ _pool_get(pp, flags, file, line)
 		panic("pool_get: nitems inconsistent\n");
 	}
 #endif
+
+#ifdef POOL_DIAGNOSTIC
 	pr_log(pp, v, PRLOG_GET, file, line);
+#endif
 
 #ifdef DIAGNOSTIC
 	if (__predict_false(pi->pi_magic != PI_MAGIC)) {
@@ -778,8 +838,7 @@ _pool_get(pp, flags, file, line)
 		 */
 		TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
 		TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
-		for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
-		     ph = TAILQ_NEXT(ph, ph_pagelist))
+		TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist)
 			if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
 				break;
 
@@ -792,7 +851,7 @@ _pool_get(pp, flags, file, line)
 	 * If we have a low water mark and we are now below that low
 	 * water mark, add more items to the pool.
 	 */
-	if (pp->pr_nitems < pp->pr_minitems && pool_catchup(pp) != 0) {
+	if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
 		/*
 		 * XXX: Should we log a warning?  Should we set up a timeout
 		 * to try again in a second or so?  The latter could break
@@ -806,24 +865,19 @@ _pool_get(pp, flags, file, line)
 }
 
 /*
- * Return resource to the pool; must be called at appropriate spl level
+ * Internal version of pool_put().  Pool is already locked/entered.
  */
-void
-_pool_put(pp, v, file, line)
-	struct pool *pp;
-	void *v;
-	const char *file;
-	long line;
+static void
+pool_do_put(struct pool *pp, void *v)
 {
 	struct pool_item *pi = v;
 	struct pool_item_header *ph;
 	caddr_t page;
 	int s;
 
-	page = (caddr_t)((u_long)v & pp->pr_pagemask);
+	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
 
-	simple_lock(&pp->pr_slock);
-	pr_enter(pp, file, line);
+	page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
 
 #ifdef DIAGNOSTIC
 	if (__predict_false(pp->pr_nout == 0)) {
@@ -833,8 +887,6 @@ _pool_put(pp, v, file, line)
 	}
 #endif
 
-	pr_log(pp, v, PRLOG_PUT, file, line);
-
 	if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
 		pr_printlog(pp, NULL, printf);
 		panic("pool_put: %s: page header missing", pp->pr_wchan);
@@ -877,8 +929,6 @@ _pool_put(pp, v, file, line)
 		pp->pr_flags &= ~PR_WANTED;
 		if (ph->ph_nmissing == 0)
 			pp->pr_nidle++;
-		pr_leave(pp);
-		simple_unlock(&pp->pr_slock);
 		wakeup((caddr_t)pp);
 		return;
 	}
@@ -899,8 +949,9 @@ _pool_put(pp, v, file, line)
 	 */
 	if (ph->ph_nmissing == 0) {
 		pp->pr_nidle++;
-		if (pp->pr_npages > pp->pr_maxpages) {
-			pr_rmpage(pp, ph);
+		if (pp->pr_npages > pp->pr_maxpages ||
+		    (pp->pr_alloc->pa_flags & PA_WANT) != 0) {
+			pr_rmpage(pp, ph, NULL);
 		} else {
 			TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
 			TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
@@ -923,8 +974,7 @@ _pool_put(pp, v, file, line)
 			 * page with the fewest available items, to minimize
 			 * fragmentation?
 			 */
-			for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
-			     ph = TAILQ_NEXT(ph, ph_pagelist))
+			TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist)
 				if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
 					break;
 
@@ -942,59 +992,76 @@ _pool_put(pp, v, file, line)
 		TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
 		pp->pr_curpage = ph;
 	}
+}
+
+/*
+ * Return resource to the pool; must be called at appropriate spl level
+ */
+#ifdef POOL_DIAGNOSTIC
+void
+_pool_put(struct pool *pp, void *v, const char *file, long line)
+{
+
+	simple_lock(&pp->pr_slock);
+	pr_enter(pp, file, line);
+
+	pr_log(pp, v, PRLOG_PUT, file, line);
+
+	pool_do_put(pp, v);
 
 	pr_leave(pp);
 	simple_unlock(&pp->pr_slock);
+}
+#undef pool_put
+#endif /* POOL_DIAGNOSTIC */
+
+void
+pool_put(struct pool *pp, void *v)
+{
 
+	simple_lock(&pp->pr_slock);
+
+	pool_do_put(pp, v);
+
+	simple_unlock(&pp->pr_slock);
 }
 
+#ifdef POOL_DIAGNOSTIC
+#define		pool_put(h, v)	_pool_put((h), (v), __FILE__, __LINE__)
+#endif
+
 /*
  * Add N items to the pool.
  */
 int
-pool_prime(pp, n, storage)
-	struct pool *pp;
-	int n;
-	caddr_t storage;
+pool_prime(struct pool *pp, int n)
 {
+	struct pool_item_header *ph;
 	caddr_t cp;
-	int newnitems, newpages;
-
-#ifdef DIAGNOSTIC
-	if (__predict_false(storage && !(pp->pr_roflags & PR_STATIC)))
-		panic("pool_prime: static");
-	/* !storage && static caught below */
-#endif
+	int newpages;
 
 	simple_lock(&pp->pr_slock);
 
-	newnitems = pp->pr_minitems + n;
-	newpages =
-		roundup(newnitems, pp->pr_itemsperpage) / pp->pr_itemsperpage
-		- pp->pr_minpages;
+	newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
 
 	while (newpages-- > 0) {
-		if (pp->pr_roflags & PR_STATIC) {
-			cp = storage;
-			storage += pp->pr_pagesz;
-		} else {
-			simple_unlock(&pp->pr_slock);
-			cp = (*pp->pr_alloc)(pp->pr_pagesz, 0, pp->pr_mtype);
-			simple_lock(&pp->pr_slock);
-		}
+		simple_unlock(&pp->pr_slock);
+		cp = pool_allocator_alloc(pp, PR_NOWAIT);
+		if (__predict_true(cp != NULL))
+			ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
+		simple_lock(&pp->pr_slock);
 
-		if (cp == NULL) {
-			simple_unlock(&pp->pr_slock);
-			return (ENOMEM);
+		if (__predict_false(cp == NULL || ph == NULL)) {
+			if (cp != NULL)
+				pool_allocator_free(pp, cp);
+			break;
 		}
 
+		pool_prime_page(pp, cp, ph);
 		pp->pr_npagealloc++;
-		pool_prime_page(pp, cp);
 		pp->pr_minpages++;
 	}
 
-	pp->pr_minitems = newnitems;
-
 	if (pp->pr_minpages >= pp->pr_maxpages)
 		pp->pr_maxpages = pp->pr_minpages + 1;	/* XXX */
 
@@ -1008,29 +1075,22 @@ pool_prime(pp, n, storage)
  * Note, we must be called with the pool descriptor LOCKED.
  */
 static void
-pool_prime_page(pp, storage)
-	struct pool *pp;
-	caddr_t storage;
+pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
 {
 	struct pool_item *pi;
-	struct pool_item_header *ph;
 	caddr_t cp = storage;
 	unsigned int align = pp->pr_align;
 	unsigned int ioff = pp->pr_itemoffset;
-	int s, n;
+	int n;
 
-	if (((u_long)cp & (pp->pr_pagesz - 1)) != 0)
+#ifdef DIAGNOSTIC
+	if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
 		panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
+#endif
 
-	if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
-		ph = (struct pool_item_header *)(cp + pp->pr_phoffset);
-	} else {
-		s = splhigh();
-		ph = pool_get(&phpool, PR_URGENT);
-		splx(s);
+	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
 		LIST_INSERT_HEAD(&pp->pr_hashtab[PR_HASH_INDEX(pp, cp)],
-				 ph, ph_hashlist);
-	}
+		    ph, ph_hashlist);
 
 	/*
 	 * Insert page header.
@@ -1065,6 +1125,8 @@ pool_prime_page(pp, storage)
 	while (n--) {
 		pi = (struct pool_item *)cp;
 
+		KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
+
 		/* Insert on page list */
 		TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
 #ifdef DIAGNOSTIC
@@ -1084,37 +1146,22 @@ pool_prime_page(pp, storage)
 }
 
 /*
- * Like pool_prime(), except this is used by pool_get() when nitems
- * drops below the low water mark.  This is used to catch up nitmes
- * with the low water mark.
+ * Used by pool_get() when nitems drops below the low water mark.  This
+ * is used to catch up nitmes with the low water mark.
  *
  * Note 1, we never wait for memory here, we let the caller decide what to do.
  *
- * Note 2, this doesn't work with static pools.
- *
- * Note 3, we must be called with the pool already locked, and we return
+ * Note 2, we must be called with the pool already locked, and we return
  * with it locked.
  */
 static int
-pool_catchup(pp)
-	struct pool *pp;
+pool_catchup(struct pool *pp)
 {
+	struct pool_item_header *ph;
 	caddr_t cp;
 	int error = 0;
 
-	if (pp->pr_roflags & PR_STATIC) {
-		/*
-		 * We dropped below the low water mark, and this is not a
-		 * good thing.  Log a warning.
-		 *
-		 * XXX: rate-limit this?
-		 */
-		printf("WARNING: static pool `%s' dropped below low water "
-		    "mark\n", pp->pr_wchan);
-		return (0);
-	}
-
-	while (pp->pr_nitems < pp->pr_minitems) {
+	while (POOL_NEEDS_CATCHUP(pp)) {
 		/*
 		 * Call the page back-end allocator for more memory.
 		 *
@@ -1122,25 +1169,26 @@ pool_catchup(pp)
 		 * the pool descriptor?
 		 */
 		simple_unlock(&pp->pr_slock);
-		cp = (*pp->pr_alloc)(pp->pr_pagesz, 0, pp->pr_mtype);
+		cp = pool_allocator_alloc(pp, PR_NOWAIT);
+		if (__predict_true(cp != NULL))
+			ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
 		simple_lock(&pp->pr_slock);
-		if (__predict_false(cp == NULL)) {
+		if (__predict_false(cp == NULL || ph == NULL)) {
+			if (cp != NULL)
+				pool_allocator_free(pp, cp);
 			error = ENOMEM;
 			break;
 		}
+		pool_prime_page(pp, cp, ph);
 		pp->pr_npagealloc++;
-		pool_prime_page(pp, cp);
 	}
 
 	return (error);
 }
 
 void
-pool_setlowat(pp, n)
-	pool_handle_t	pp;
-	int n;
+pool_setlowat(struct pool *pp, int n)
 {
-	int error;
 
 	simple_lock(&pp->pr_slock);
 
@@ -1150,8 +1198,7 @@ pool_setlowat(pp, n)
 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
 
 	/* Make sure we're caught up with the newly-set low water mark. */
-	if ((pp->pr_nitems < pp->pr_minitems) &&
-	    (error = pool_catchup(pp)) != 0) {
+	if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
 		/*
 		 * XXX: Should we log a warning?  Should we set up a timeout
 		 * to try again in a second or so?  The latter could break
@@ -1163,9 +1210,7 @@ pool_setlowat(pp, n)
 }
 
 void
-pool_sethiwat(pp, n)
-	pool_handle_t	pp;
-	int n;
+pool_sethiwat(struct pool *pp, int n)
 {
 
 	simple_lock(&pp->pr_slock);
@@ -1178,11 +1223,7 @@ pool_sethiwat(pp, n)
 }
 
 void
-pool_sethardlimit(pp, n, warnmess, ratecap)
-	pool_handle_t pp;
-	int n;
-	const char *warnmess;
-	int ratecap;
+pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
 {
 
 	simple_lock(&pp->pr_slock);
@@ -1205,76 +1246,40 @@ pool_sethardlimit(pp, n, warnmess, ratec
 }
 
 /*
- * Default page allocator.
- */
-static void *
-pool_page_alloc(sz, flags, mtype)
-	unsigned long sz;
-	int flags;
-	int mtype;
-{
-	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
-
-	return ((void *)uvm_km_alloc_poolpage(waitok));
-}
-
-static void
-pool_page_free(v, sz, mtype)
-	void *v;
-	unsigned long sz;
-	int mtype;
-{
-
-	uvm_km_free_poolpage((vaddr_t)v);
-}
-
-/*
- * Alternate pool page allocator for pools that know they will
- * never be accessed in interrupt context.
- */
-void *
-pool_page_alloc_nointr(sz, flags, mtype)
-	unsigned long sz;
-	int flags;
-	int mtype;
-{
-	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
-
-	return ((void *)uvm_km_alloc_poolpage1(kernel_map, uvm.kernel_object,
-	    waitok));
-}
-
-void
-pool_page_free_nointr(v, sz, mtype)
-	void *v;
-	unsigned long sz;
-	int mtype;
-{
-
-	uvm_km_free_poolpage1(kernel_map, (vaddr_t)v);
-}
-
-
-/*
  * Release all complete pages that have not been used recently.
  */
-void
-_pool_reclaim(pp, file, line)
-	pool_handle_t pp;
-	const char *file;
-	long line;
+int
+#ifdef POOL_DIAGNOSTIC
+_pool_reclaim(struct pool *pp, const char *file, long line)
+#else
+pool_reclaim(struct pool *pp)
+#endif
 {
 	struct pool_item_header *ph, *phnext;
+	struct pool_cache *pc;
 	struct timeval curtime;
+	struct pool_pagelist pq;
 	int s;
 
-	if (pp->pr_roflags & PR_STATIC)
-		return;
+	if (pp->pr_drain_hook != NULL) {
+		/*
+		 * The drain hook must be called with the pool unlocked.
+		 */
+		(*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
+	}
 
 	if (simple_lock_try(&pp->pr_slock) == 0)
-		return;
+		return (0);
 	pr_enter(pp, file, line);
 
+	TAILQ_INIT(&pq);
+
+	/*
+	 * Reclaim items from the pool's caches.
+	 */
+	TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
+		pool_cache_reclaim(pc);
+
 	s = splclock();
 	curtime = mono_time;
 	splx(s);
@@ -1300,14 +1305,29 @@ _pool_reclaim(pp, file, line)
 			    pp->pr_minitems)
 				break;
 
-			pr_rmpage(pp, ph);
+			pr_rmpage(pp, ph, &pq);
 		}
 	}
 
 	pr_leave(pp);
 	simple_unlock(&pp->pr_slock);
-}
+	if (TAILQ_EMPTY(&pq))
+		return (0);
 
+	while ((ph = TAILQ_FIRST(&pq)) != NULL) {
+		TAILQ_REMOVE(&pq, ph, ph_pagelist);
+		pool_allocator_free(pp, ph->ph_page);
+		if (pp->pr_roflags & PR_PHINPAGE) {
+			continue;
+		}
+		LIST_REMOVE(ph, ph_hashlist);
+		s = splhigh();
+		pool_put(&phpool, ph);
+		splx(s);
+	}
+
+	return (1);
+}
 
 /*
  * Drain pools, one at a time.
@@ -1315,40 +1335,35 @@ _pool_reclaim(pp, file, line)
  * Note, we must never be called from an interrupt context.
  */
 void
-pool_drain(arg)
-	void *arg;
+pool_drain(void *arg)
 {
 	struct pool *pp;
 	int s;
 
-	s = splimp();
+	pp = NULL;
+	s = splvm();
 	simple_lock(&pool_head_slock);
-
-	if (drainpp == NULL && (drainpp = TAILQ_FIRST(&pool_head)) == NULL)
-		goto out;
-
-	pp = drainpp;
-	drainpp = TAILQ_NEXT(pp, pr_poollist);
-
-	pool_reclaim(pp);
-
- out:
+	if (drainpp == NULL) {
+		drainpp = TAILQ_FIRST(&pool_head);
+	}
+	if (drainpp) {
+		pp = drainpp;
+		drainpp = TAILQ_NEXT(pp, pr_poollist);
+	}
 	simple_unlock(&pool_head_slock);
+	pool_reclaim(pp);
 	splx(s);
 }
 
-
 /*
  * Diagnostic helpers.
  */
 void
-pool_print(pp, modif)
-	struct pool *pp;
-	const char *modif;
+pool_print(struct pool *pp, const char *modif)
 {
 	int s;
 
-	s = splimp();
+	s = splvm();
 	if (simple_lock_try(&pp->pr_slock) == 0) {
 		printf("pool %s is locked; try again later\n",
 		    pp->pr_wchan);
@@ -1361,10 +1376,7 @@ pool_print(pp, modif)
 }
 
 void
-pool_printit(pp, modif, pr)
-	struct pool *pp;
-	const char *modif;
-	void (*pr) __P((const char *, ...));
+pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
 {
 	int didlock = 0;
 
@@ -1394,16 +1406,15 @@ pool_printit(pp, modif, pr)
 }
 
 static void
-pool_print1(pp, modif, pr)
-	struct pool *pp;
-	const char *modif;
-	void (*pr) __P((const char *, ...));
+pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
 {
 	struct pool_item_header *ph;
+	struct pool_cache *pc;
+	struct pool_cache_group *pcg;
 #ifdef DIAGNOSTIC
 	struct pool_item *pi;
 #endif
-	int print_log = 0, print_pagelist = 0;
+	int i, print_log = 0, print_pagelist = 0, print_cache = 0;
 	char c;
 
 	while ((c = *modif++) != '\0') {
@@ -1411,14 +1422,15 @@ pool_print1(pp, modif, pr)
 			print_log = 1;
 		if (c == 'p')
 			print_pagelist = 1;
+		if (c == 'c')
+			print_cache = 1;
 		modif++;
 	}
 
 	(*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
 	    pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
 	    pp->pr_roflags);
-	(*pr)("\tpagesz %u, mtype %d\n", pp->pr_pagesz, pp->pr_mtype);
-	(*pr)("\talloc %p, release %p\n", pp->pr_alloc, pp->pr_free);
+	(*pr)("\talloc %p\n", pp->pr_alloc);
 	(*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
 	    pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
 	(*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
@@ -1440,8 +1452,7 @@ pool_print1(pp, modif, pr)
 		    (u_long)ph->ph_time.tv_sec,
 		    (u_long)ph->ph_time.tv_usec);
 #ifdef DIAGNOSTIC
-		for (pi = TAILQ_FIRST(&ph->ph_itemlist); pi != NULL;
-		     pi = TAILQ_NEXT(pi, pi_list)) {
+		TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) {
 			if (pi->pi_magic != PI_MAGIC) {
 				(*pr)("\t\t\titem %p, magic 0x%x\n",
 				    pi, pi->pi_magic);
@@ -1467,27 +1478,40 @@ pool_print1(pp, modif, pr)
 
  skip_log:
 
+	if (print_cache == 0)
+		goto skip_cache;
+
+	TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist) {
+		(*pr)("\tcache %p: allocfrom %p freeto %p\n", pc,
+		    pc->pc_allocfrom, pc->pc_freeto);
+		(*pr)("\t    hits %lu misses %lu ngroups %lu nitems %lu\n",
+		    pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
+		TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
+			(*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail);
+			for (i = 0; i < PCG_NOBJECTS; i++)
+				(*pr)("\t\t\t%p\n", pcg->pcg_objects[i]);
+		}
+	}
+
+ skip_cache:
+
 	pr_enter_check(pp, pr);
 }
 
 int
-pool_chk(pp, label)
-	struct pool *pp;
-	char *label;
+pool_chk(struct pool *pp, const char *label)
 {
 	struct pool_item_header *ph;
 	int r = 0;
 
 	simple_lock(&pp->pr_slock);
 
-	for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
-	     ph = TAILQ_NEXT(ph, ph_pagelist)) {
-
+	TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) {
 		struct pool_item *pi;
 		int n;
 		caddr_t page;
 
-		page = (caddr_t)((u_long)ph & pp->pr_pagemask);
+		page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask);
 		if (page != ph->ph_page &&
 		    (pp->pr_roflags & PR_PHINPAGE) != 0) {
 			if (label != NULL)
@@ -1516,7 +1540,8 @@ pool_chk(pp, label)
 				panic("pool");
 			}
 #endif
-			page = (caddr_t)((u_long)pi & pp->pr_pagemask);
+			page =
+			    (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
 			if (page == ph->ph_page)
 				continue;
 
@@ -1534,3 +1559,500 @@ out:
 	simple_unlock(&pp->pr_slock);
 	return (r);
 }
+
+/*
+ * pool_cache_init:
+ *
+ *	Initialize a pool cache.
+ *
+ *	NOTE: If the pool must be protected from interrupts, we expect
+ *	to be called at the appropriate interrupt priority level.
+ */
+void
+pool_cache_init(struct pool_cache *pc, struct pool *pp,
+    int (*ctor)(void *, void *, int),
+    void (*dtor)(void *, void *),
+    void *arg)
+{
+
+	TAILQ_INIT(&pc->pc_grouplist);
+	simple_lock_init(&pc->pc_slock);
+
+	pc->pc_allocfrom = NULL;
+	pc->pc_freeto = NULL;
+	pc->pc_pool = pp;
+
+	pc->pc_ctor = ctor;
+	pc->pc_dtor = dtor;
+	pc->pc_arg  = arg;
+
+	pc->pc_hits   = 0;
+	pc->pc_misses = 0;
+
+	pc->pc_ngroups = 0;
+
+	pc->pc_nitems = 0;
+
+	simple_lock(&pp->pr_slock);
+	TAILQ_INSERT_TAIL(&pp->pr_cachelist, pc, pc_poollist);
+	simple_unlock(&pp->pr_slock);
+}
+
+/*
+ * pool_cache_destroy:
+ *
+ *	Destroy a pool cache.
+ */
+void
+pool_cache_destroy(struct pool_cache *pc)
+{
+	struct pool *pp = pc->pc_pool;
+
+	/* First, invalidate the entire cache. */
+	pool_cache_invalidate(pc);
+
+	/* ...and remove it from the pool's cache list. */
+	simple_lock(&pp->pr_slock);
+	TAILQ_REMOVE(&pp->pr_cachelist, pc, pc_poollist);
+	simple_unlock(&pp->pr_slock);
+}
+
+static __inline void *
+pcg_get(struct pool_cache_group *pcg)
+{
+	void *object;
+	u_int idx;
+
+	KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
+	KASSERT(pcg->pcg_avail != 0);
+	idx = --pcg->pcg_avail;
+
+	KASSERT(pcg->pcg_objects[idx] != NULL);
+	object = pcg->pcg_objects[idx];
+	pcg->pcg_objects[idx] = NULL;
+
+	return (object);
+}
+
+static __inline void
+pcg_put(struct pool_cache_group *pcg, void *object)
+{
+	u_int idx;
+
+	KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
+	idx = pcg->pcg_avail++;
+
+	KASSERT(pcg->pcg_objects[idx] == NULL);
+	pcg->pcg_objects[idx] = object;
+}
+
+/*
+ * pool_cache_get:
+ *
+ *	Get an object from a pool cache.
+ */
+void *
+pool_cache_get(struct pool_cache *pc, int flags)
+{
+	struct pool_cache_group *pcg;
+	void *object;
+
+#ifdef LOCKDEBUG
+	if (flags & PR_WAITOK)
+		simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)");
+#endif
+
+	simple_lock(&pc->pc_slock);
+
+	if ((pcg = pc->pc_allocfrom) == NULL) {
+		TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
+			if (pcg->pcg_avail != 0) {
+				pc->pc_allocfrom = pcg;
+				goto have_group;
+			}
+		}
+
+		/*
+		 * No groups with any available objects.  Allocate
+		 * a new object, construct it, and return it to
+		 * the caller.  We will allocate a group, if necessary,
+		 * when the object is freed back to the cache.
+		 */
+		pc->pc_misses++;
+		simple_unlock(&pc->pc_slock);
+		object = pool_get(pc->pc_pool, flags);
+		if (object != NULL && pc->pc_ctor != NULL) {
+			if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
+				pool_put(pc->pc_pool, object);
+				return (NULL);
+			}
+		}
+		return (object);
+	}
+
+ have_group:
+	pc->pc_hits++;
+	pc->pc_nitems--;
+	object = pcg_get(pcg);
+
+	if (pcg->pcg_avail == 0)
+		pc->pc_allocfrom = NULL;
+
+	simple_unlock(&pc->pc_slock);
+
+	return (object);
+}
+
+/*
+ * pool_cache_put:
+ *
+ *	Put an object back to the pool cache.
+ */
+void
+pool_cache_put(struct pool_cache *pc, void *object)
+{
+	struct pool_cache_group *pcg;
+	int s;
+
+	simple_lock(&pc->pc_slock);
+
+	if ((pcg = pc->pc_freeto) == NULL) {
+		TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
+			if (pcg->pcg_avail != PCG_NOBJECTS) {
+				pc->pc_freeto = pcg;
+				goto have_group;
+			}
+		}
+
+		/*
+		 * No empty groups to free the object to.  Attempt to
+		 * allocate one.
+		 */
+		simple_unlock(&pc->pc_slock);
+		s = splvm();
+		pcg = pool_get(&pcgpool, PR_NOWAIT);
+		splx(s);
+		if (pcg != NULL) {
+			memset(pcg, 0, sizeof(*pcg));
+			simple_lock(&pc->pc_slock);
+			pc->pc_ngroups++;
+			TAILQ_INSERT_TAIL(&pc->pc_grouplist, pcg, pcg_list);
+			if (pc->pc_freeto == NULL)
+				pc->pc_freeto = pcg;
+			goto have_group;
+		}
+
+		/*
+		 * Unable to allocate a cache group; destruct the object
+		 * and free it back to the pool.
+		 */
+		pool_cache_destruct_object(pc, object);
+		return;
+	}
+
+ have_group:
+	pc->pc_nitems++;
+	pcg_put(pcg, object);
+
+	if (pcg->pcg_avail == PCG_NOBJECTS)
+		pc->pc_freeto = NULL;
+
+	simple_unlock(&pc->pc_slock);
+}
+
+/*
+ * pool_cache_destruct_object:
+ *
+ *	Force destruction of an object and its release back into
+ *	the pool.
+ */
+void
+pool_cache_destruct_object(struct pool_cache *pc, void *object)
+{
+
+	if (pc->pc_dtor != NULL)
+		(*pc->pc_dtor)(pc->pc_arg, object);
+	pool_put(pc->pc_pool, object);
+}
+
+/*
+ * pool_cache_do_invalidate:
+ *
+ *	This internal function implements pool_cache_invalidate() and
+ *	pool_cache_reclaim().
+ */
+static void
+pool_cache_do_invalidate(struct pool_cache *pc, int free_groups,
+    void (*putit)(struct pool *, void *))
+{
+	struct pool_cache_group *pcg, *npcg;
+	void *object;
+	int s;
+
+	for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
+	     pcg = npcg) {
+		npcg = TAILQ_NEXT(pcg, pcg_list);
+		while (pcg->pcg_avail != 0) {
+			pc->pc_nitems--;
+			object = pcg_get(pcg);
+			if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg)
+				pc->pc_allocfrom = NULL;
+			if (pc->pc_dtor != NULL)
+				(*pc->pc_dtor)(pc->pc_arg, object);
+			(*putit)(pc->pc_pool, object);
+		}
+		if (free_groups) {
+			pc->pc_ngroups--;
+			TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list);
+			if (pc->pc_freeto == pcg)
+				pc->pc_freeto = NULL;
+			s = splvm();
+			pool_put(&pcgpool, pcg);
+			splx(s);
+		}
+	}
+}
+
+/*
+ * pool_cache_invalidate:
+ *
+ *	Invalidate a pool cache (destruct and release all of the
+ *	cached objects).
+ */
+void
+pool_cache_invalidate(struct pool_cache *pc)
+{
+
+	simple_lock(&pc->pc_slock);
+	pool_cache_do_invalidate(pc, 0, pool_put);
+	simple_unlock(&pc->pc_slock);
+}
+
+/*
+ * pool_cache_reclaim:
+ *
+ *	Reclaim a pool cache for pool_reclaim().
+ */
+static void
+pool_cache_reclaim(struct pool_cache *pc)
+{
+
+	simple_lock(&pc->pc_slock);
+	pool_cache_do_invalidate(pc, 1, pool_do_put);
+	simple_unlock(&pc->pc_slock);
+}
+
+/*
+ * Pool backend allocators.
+ *
+ * Each pool has a backend allocator that handles allocation, deallocation,
+ * and any additional draining that might be needed.
+ *
+ * We provide two standard allocators:
+ *
+ *	pool_allocator_kmem - the default when no allocator is specified
+ *
+ *	pool_allocator_nointr - used for pools that will not be accessed
+ *	in interrupt context.
+ */
+void	*pool_page_alloc(struct pool *, int);
+void	pool_page_free(struct pool *, void *);
+
+struct pool_allocator pool_allocator_kmem = {
+	pool_page_alloc, pool_page_free, 0,
+};
+
+void	*pool_page_alloc_nointr(struct pool *, int);
+void	pool_page_free_nointr(struct pool *, void *);
+
+struct pool_allocator pool_allocator_nointr = {
+	pool_page_alloc_nointr, pool_page_free_nointr, 0,
+};
+
+#ifdef POOL_SUBPAGE
+void	*pool_subpage_alloc(struct pool *, int);
+void	pool_subpage_free(struct pool *, void *);
+
+struct pool_allocator pool_allocator_kmem_subpage = {
+	pool_subpage_alloc, pool_subpage_free, 0,
+};
+#endif /* POOL_SUBPAGE */
+
+/*
+ * We have at least three different resources for the same allocation and
+ * each resource can be depleted.  First, we have the ready elements in the
+ * pool.  Then we have the resource (typically a vm_map) for this allocator.
+ * Finally, we have physical memory.  Waiting for any of these can be
+ * unnecessary when any other is freed, but the kernel doesn't support
+ * sleeping on multiple wait channels, so we have to employ another strategy.
+ *
+ * The caller sleeps on the pool (so that it can be awakened when an item
+ * is returned to the pool), but we set PA_WANT on the allocator.  When a
+ * page is returned to the allocator and PA_WANT is set, pool_allocator_free
+ * will wake up all sleeping pools belonging to this allocator.
+ *
+ * XXX Thundering herd.
+ */
+void *
+pool_allocator_alloc(struct pool *org, int flags)
+{
+	struct pool_allocator *pa = org->pr_alloc;
+	struct pool *pp, *start;
+	int s, freed;
+	void *res;
+
+	do {
+		if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
+			return (res);
+		if ((flags & PR_WAITOK) == 0) {
+			/*
+			 * We only run the drain hookhere if PR_NOWAIT.
+			 * In other cases, the hook will be run in
+			 * pool_reclaim().
+			 */
+			if (org->pr_drain_hook != NULL) {
+				(*org->pr_drain_hook)(org->pr_drain_hook_arg,
+				    flags);
+				if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
+					return (res);
+			}
+			break;
+		}
+
+		/*
+		 * Drain all pools, except "org", that use this
+		 * allocator.  We do this to reclaim VA space.
+		 * pa_alloc is responsible for waiting for
+		 * physical memory.
+		 *
+		 * XXX We risk looping forever if start if someone
+		 * calls pool_destroy on "start".  But there is no
+		 * other way to have potentially sleeping pool_reclaim,
+		 * non-sleeping locks on pool_allocator, and some
+		 * stirring of drained pools in the allocator.
+		 *
+		 * XXX Maybe we should use pool_head_slock for locking
+		 * the allocators?
+		 */
+		freed = 0;
+
+		s = splvm();
+		simple_lock(&pa->pa_slock);
+		pp = start = TAILQ_FIRST(&pa->pa_list);
+		do {
+			TAILQ_REMOVE(&pa->pa_list, pp, pr_alloc_list);
+			TAILQ_INSERT_TAIL(&pa->pa_list, pp, pr_alloc_list);
+			if (pp == org)
+				continue;
+			simple_unlock(&pa->pa_slock);
+			freed = pool_reclaim(pp);
+			simple_lock(&pa->pa_slock);
+		} while ((pp = TAILQ_FIRST(&pa->pa_list)) != start &&
+			 freed == 0);
+
+		if (freed == 0) {
+			/*
+			 * We set PA_WANT here, the caller will most likely
+			 * sleep waiting for pages (if not, this won't hurt
+			 * that much), and there is no way to set this in
+			 * the caller without violating locking order.
+			 */
+			pa->pa_flags |= PA_WANT;
+		}
+		simple_unlock(&pa->pa_slock);
+		splx(s);
+	} while (freed);
+	return (NULL);
+}
+
+void
+pool_allocator_free(struct pool *pp, void *v)
+{
+	struct pool_allocator *pa = pp->pr_alloc;
+	int s;
+
+	(*pa->pa_free)(pp, v);
+
+	s = splvm();
+	simple_lock(&pa->pa_slock);
+	if ((pa->pa_flags & PA_WANT) == 0) {
+		simple_unlock(&pa->pa_slock);
+		splx(s);
+		return;
+	}
+
+	TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
+		simple_lock(&pp->pr_slock);
+		if ((pp->pr_flags & PR_WANTED) != 0) {
+			pp->pr_flags &= ~PR_WANTED;
+			wakeup(pp);
+		}
+		simple_unlock(&pp->pr_slock);
+	}
+	pa->pa_flags &= ~PA_WANT;
+	simple_unlock(&pa->pa_slock);
+	splx(s);
+}
+
+void *
+pool_page_alloc(struct pool *pp, int flags)
+{
+	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
+
+	return ((void *) uvm_km_alloc_poolpage(waitok));
+}
+
+void
+pool_page_free(struct pool *pp, void *v)
+{
+
+	uvm_km_free_poolpage((vaddr_t) v);
+}
+
+#ifdef POOL_SUBPAGE
+/* Sub-page allocator, for machines with large hardware pages. */
+void *
+pool_subpage_alloc(struct pool *pp, int flags)
+{
+
+	return (pool_get(&psppool, flags));
+}
+
+void
+pool_subpage_free(struct pool *pp, void *v)
+{
+
+	pool_put(&psppool, v);
+}
+
+/* We don't provide a real nointr allocator.  Maybe later. */
+void *
+pool_page_alloc_nointr(struct pool *pp, int flags)
+{
+
+	return (pool_subpage_alloc(pp, flags));
+}
+
+void
+pool_page_free_nointr(struct pool *pp, void *v)
+{
+
+	pool_subpage_free(pp, v);
+}
+#else
+void *
+pool_page_alloc_nointr(struct pool *pp, int flags)
+{
+	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
+
+	return ((void *) uvm_km_alloc_poolpage1(kernel_map,
+	    uvm.kernel_object, waitok));
+}
+
+void
+pool_page_free_nointr(struct pool *pp, void *v)
+{
+
+	uvm_km_free_poolpage1(kernel_map, (vaddr_t) v);
+}
+#endif /* POOL_SUBPAGE */