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 1.42
retrieving revision 1.101
diff -u -p -r1.42 -r1.101
--- src/sys/kern/subr_pool.c	2000/12/06 18:20:52	1.42
+++ src/sys/kern/subr_pool.c	2005/06/18 01:34:03	1.101
@@ -1,7 +1,7 @@
-/*	$NetBSD: subr_pool.c,v 1.42 2000/12/06 18:20:52 thorpej Exp $	*/
+/*	$NetBSD: subr_pool.c,v 1.101 2005/06/18 01:34:03 thorpej 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 1.101 2005/06/18 01:34:03 thorpej Exp $");
+
 #include "opt_pool.h"
 #include "opt_poollog.h"
 #include "opt_lockdebug.h"
@@ -56,19 +59,36 @@
 /*
  * Pool resource management utility.
  *
- * Memory is allocated in pages which are split into pieces according
- * to the pool item size. Each page is kept on a list headed by `pr_pagelist'
- * in the pool structure and the individual pool items are on a linked list
- * headed by `ph_itemlist' in each page header. The memory for building
- * the page list is either taken from the allocated pages themselves (for
- * small pool items) or taken from an internal pool of page headers (`phpool').
+ * Memory is allocated in pages which are split into pieces according to
+ * the pool item size. Each page is kept on one of three lists in the
+ * pool structure: `pr_emptypages', `pr_fullpages' and `pr_partpages',
+ * for empty, full and partially-full pages respectively. The individual
+ * pool items are on a linked list headed by `ph_itemlist' in each page
+ * header. The memory for building the page list is either taken from
+ * the allocated pages themselves (for small pool items) or taken from
+ * an internal pool of page headers (`phpool').
  */
 
 /* List of all pools */
 TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head);
 
 /* Private pool for page header structures */
-static struct pool phpool;
+#define	PHPOOL_MAX	8
+static struct pool phpool[PHPOOL_MAX];
+#define	PHPOOL_FREELIST_NELEM(idx)	(((idx) == 0) ? 0 : (1 << (idx)))
+
+#ifdef POOL_SUBPAGE
+/* Pool of subpages for use by normal pools. */
+static struct pool psppool;
+#endif
+
+static void *pool_page_alloc_meta(struct pool *, int);
+static void pool_page_free_meta(struct pool *, void *);
+
+/* allocator for pool metadata */
+static struct pool_allocator pool_allocator_meta = {
+	pool_page_alloc_meta, pool_page_free_meta
+};
 
 /* # of seconds to retain page after last use */
 int pool_inactive_time = 10;
@@ -79,43 +99,101 @@ static struct pool	*drainpp;
 /* This spin lock protects both pool_head and drainpp. */
 struct simplelock pool_head_slock = SIMPLELOCK_INITIALIZER;
 
+typedef uint8_t pool_item_freelist_t;
+
 struct pool_item_header {
 	/* Page headers */
-	TAILQ_ENTRY(pool_item_header)
-				ph_pagelist;	/* pool page list */
-	TAILQ_HEAD(,pool_item)	ph_itemlist;	/* chunk list for this page */
 	LIST_ENTRY(pool_item_header)
-				ph_hashlist;	/* Off-page page headers */
-	int			ph_nmissing;	/* # of chunks in use */
+				ph_pagelist;	/* pool page list */
+	SPLAY_ENTRY(pool_item_header)
+				ph_node;	/* Off-page page headers */
 	caddr_t			ph_page;	/* this page's address */
 	struct timeval		ph_time;	/* last referenced */
+	union {
+		/* !PR_NOTOUCH */
+		struct {
+			TAILQ_HEAD(, pool_item)
+				phu_itemlist;	/* chunk list for this page */
+		} phu_normal;
+		/* PR_NOTOUCH */
+		struct {
+			uint16_t
+				phu_off;	/* start offset in page */
+			pool_item_freelist_t
+				phu_firstfree;	/* first free item */
+			/*
+			 * XXX it might be better to use
+			 * a simple bitmap and ffs(3)
+			 */
+		} phu_notouch;
+	} ph_u;
+	uint16_t		ph_nmissing;	/* # of chunks in use */
 };
+#define	ph_itemlist	ph_u.phu_normal.phu_itemlist
+#define	ph_off		ph_u.phu_notouch.phu_off
+#define	ph_firstfree	ph_u.phu_notouch.phu_firstfree
 
 struct pool_item {
 #ifdef DIAGNOSTIC
-	int pi_magic;
+	u_int pi_magic;
 #endif
-#define	PI_MAGIC 0xdeadbeef
+#define	PI_MAGIC 0xdeadbeefU
 	/* Other entries use only this list entry */
 	TAILQ_ENTRY(pool_item)	pi_list;
 };
 
+#define	POOL_NEEDS_CATCHUP(pp)						\
+	((pp)->pr_nitems < (pp)->pr_minitems)
 
-#define	PR_HASH_INDEX(pp,addr) \
-	(((u_long)(addr) >> (pp)->pr_pageshift) & (PR_HASHTABSIZE - 1))
+/*
+ * 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 *, struct pool_pagelist *);
 
 static int	pool_catchup(struct pool *);
-static void	pool_prime_page(struct pool *, caddr_t);
-static void	*pool_page_alloc(unsigned long, int, int);
-static void	pool_page_free(void *, unsigned long, int);
+static void	pool_prime_page(struct pool *, caddr_t,
+		    struct pool_item_header *);
+static void	pool_update_curpage(struct pool *);
+
+void		*pool_allocator_alloc(struct pool *, int);
+void		pool_allocator_free(struct pool *, void *);
 
+static void pool_print_pagelist(struct pool *, struct pool_pagelist *,
+	void (*)(const char *, ...));
 static void pool_print1(struct pool *, const char *,
 	void (*)(const char *, ...));
 
+static int pool_chk_page(struct pool *, const char *,
+			 struct pool_item_header *);
+
 /*
- * 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;
@@ -126,6 +204,7 @@ struct pool_log {
 	void		*pl_addr;
 };
 
+#ifdef POOL_DIAGNOSTIC
 /* Number of entries in pool log buffers */
 #ifndef POOL_LOGSIZE
 #define	POOL_LOGSIZE	10
@@ -133,7 +212,6 @@ struct pool_log {
 
 int pool_logsize = POOL_LOGSIZE;
 
-#ifdef DIAGNOSTIC
 static __inline void
 pr_log(struct pool *pp, void *v, int action, const char *file, long line)
 {
@@ -230,7 +308,65 @@ pr_enter_check(struct pool *pp, void (*p
 #define	pr_enter(pp, file, line)
 #define	pr_leave(pp)
 #define	pr_enter_check(pp, pr)
-#endif /* DIAGNOSTIC */
+#endif /* POOL_DIAGNOSTIC */
+
+static __inline int
+pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
+    const void *v)
+{
+	const char *cp = v;
+	int idx;
+
+	KASSERT(pp->pr_roflags & PR_NOTOUCH);
+	idx = (cp - ph->ph_page - ph->ph_off) / pp->pr_size;
+	KASSERT(idx < pp->pr_itemsperpage);
+	return idx;
+}
+
+#define	PR_FREELIST_ALIGN(p) \
+	roundup((uintptr_t)(p), sizeof(pool_item_freelist_t))
+#define	PR_FREELIST(ph)	((pool_item_freelist_t *)PR_FREELIST_ALIGN((ph) + 1))
+#define	PR_INDEX_USED	((pool_item_freelist_t)-1)
+#define	PR_INDEX_EOL	((pool_item_freelist_t)-2)
+
+static __inline void
+pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph,
+    void *obj)
+{
+	int idx = pr_item_notouch_index(pp, ph, obj);
+	pool_item_freelist_t *freelist = PR_FREELIST(ph);
+
+	KASSERT(freelist[idx] == PR_INDEX_USED);
+	freelist[idx] = ph->ph_firstfree;
+	ph->ph_firstfree = idx;
+}
+
+static __inline void *
+pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph)
+{
+	int idx = ph->ph_firstfree;
+	pool_item_freelist_t *freelist = PR_FREELIST(ph);
+
+	KASSERT(freelist[idx] != PR_INDEX_USED);
+	ph->ph_firstfree = freelist[idx];
+	freelist[idx] = PR_INDEX_USED;
+
+	return ph->ph_page + ph->ph_off + idx * pp->pr_size;
+}
+
+static __inline int
+phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
+{
+	if (a->ph_page < b->ph_page)
+		return (-1);
+	else if (a->ph_page > b->ph_page)
+		return (1);
+	else
+		return (0);
+}
+
+SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
+SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
 
 /*
  * Return the pool page header based on page address.
@@ -238,27 +374,43 @@ pr_enter_check(struct pool *pp, void (*p
 static __inline struct pool_item_header *
 pr_find_pagehead(struct pool *pp, caddr_t page)
 {
-	struct pool_item_header *ph;
+	struct pool_item_header *ph, tmp;
 
 	if ((pp->pr_roflags & PR_PHINPAGE) != 0)
 		return ((struct pool_item_header *)(page + pp->pr_phoffset));
 
-	for (ph = LIST_FIRST(&pp->pr_hashtab[PR_HASH_INDEX(pp, page)]);
-	     ph != NULL;
-	     ph = LIST_NEXT(ph, ph_hashlist)) {
-		if (ph->ph_page == page)
-			return (ph);
+	tmp.ph_page = page;
+	ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
+	return ph;
+}
+
+static void
+pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq)
+{
+	struct pool_item_header *ph;
+	int s;
+
+	while ((ph = LIST_FIRST(pq)) != NULL) {
+		LIST_REMOVE(ph, ph_pagelist);
+		pool_allocator_free(pp, ph->ph_page);
+		if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
+			s = splvm();
+			pool_put(pp->pr_phpool, ph);
+			splx(s);
+		}
 	}
-	return (NULL);
 }
 
 /*
  * Remove a page from the pool.
  */
 static __inline void
-pr_rmpage(struct pool *pp, struct pool_item_header *ph)
+pr_rmpage(struct pool *pp, struct pool_item_header *ph,
+     struct pool_pagelist *pq)
 {
 
+	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
+
 	/*
 	 * If the page was idle, decrement the idle page count.
 	 */
@@ -275,65 +427,32 @@ pr_rmpage(struct pool *pp, struct pool_i
 	pp->pr_nitems -= pp->pr_itemsperpage;
 
 	/*
-	 * Unlink a page from the pool and release it.
+	 * Unlink the page from the pool and queue it for release.
 	 */
-	TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
-	(*pp->pr_free)(ph->ph_page, pp->pr_pagesz, pp->pr_mtype);
+	LIST_REMOVE(ph, ph_pagelist);
+	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
+		SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
+	LIST_INSERT_HEAD(pq, ph, ph_pagelist);
+
 	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))
-			if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
-				break;
-
-		pp->pr_curpage = ph;
-	}
+	pool_update_curpage(pp);
 }
 
 /*
- * Allocate and initialize a pool.
+ * Initialize all the pools listed in the "pools" link set.
  */
-struct pool *
-pool_create(size_t size, u_int align, u_int ioff, int nitems,
-    const char *wchan, size_t pagesz,
-    void *(*alloc)(unsigned long, int, int),
-    void (*release)(void *, unsigned long, int),
-    int mtype)
+void
+link_pool_init(void)
 {
-	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);
-		}
-	}
+	__link_set_decl(pools, struct link_pool_init);
+	struct link_pool_init * const *pi;
 
-	return (pp);
+	__link_set_foreach(pi, pools)
+		pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
+		    (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
+		    (*pi)->palloc);
 }
 
 /*
@@ -344,12 +463,14 @@ pool_create(size_t size, u_int align, u_
  */
 void
 pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
-    const char *wchan, size_t pagesz,
-    void *(*alloc)(unsigned long, int, int),
-    void (*release)(void *, unsigned long, int),
-    int mtype)
+    const char *wchan, struct pool_allocator *palloc)
 {
-	int off, slack, i;
+	int off, slack;
+	size_t trysize, phsize;
+	int s;
+
+	KASSERT((1UL << (CHAR_BIT * sizeof(pool_item_freelist_t))) - 2 >=
+	    PHPOOL_FREELIST_NELEM(PHPOOL_MAX - 1));
 
 #ifdef POOL_DIAGNOSTIC
 	/*
@@ -359,23 +480,43 @@ pool_init(struct pool *pp, size_t size, 
 		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);
@@ -383,15 +524,20 @@ pool_init(struct pool *pp, size_t size, 
 	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);
+	LIST_INIT(&pp->pr_emptypages);
+	LIST_INIT(&pp->pr_fullpages);
+	LIST_INIT(&pp->pr_partpages);
+	TAILQ_INIT(&pp->pr_cachelist);
 	pp->pr_curpage = NULL;
 	pp->pr_npages = 0;
 	pp->pr_minitems = 0;
@@ -402,12 +548,7 @@ pool_init(struct pool *pp, size_t size, 
 	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;
@@ -416,37 +557,70 @@ pool_init(struct pool *pp, size_t size, 
 	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
-	 * wasting too large a part of the page. Off-page page headers
-	 * go on a hash table, so we can match a returned item
-	 * with its header based on the page address.
-	 * We use 1/16 of the page size as the threshold (XXX: tune)
+	 * wasting too large a part of the page or too big item.
+	 * Off-page page headers go on a hash table, so we can match
+	 * a returned item with its header based on the page address.
+	 * We use 1/16 of the page size and about 8 times of the item
+	 * size as the threshold (XXX: tune)
+	 *
+	 * However, we'll put the header into the page if we can put
+	 * it without wasting any items.
+	 *
+	 * Silently enforce `0 <= ioff < align'.
 	 */
-	if (pp->pr_size < pagesz/16) {
+	pp->pr_itemoffset = ioff %= align;
+	/* See the comment below about reserved bytes. */
+	trysize = palloc->pa_pagesz - ((align - ioff) % align);
+	phsize = ALIGN(sizeof(struct pool_item_header));
+	if ((pp->pr_roflags & PR_NOTOUCH) == 0 &&
+	    (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
+	    trysize / pp->pr_size == (trysize - phsize) / pp->pr_size)) {
 		/* 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 - phsize;
 	} else {
 		/* The page header will be taken from our page header pool */
 		pp->pr_phoffset = 0;
-		off = pagesz;
-		for (i = 0; i < PR_HASHTABSIZE; i++) {
-			LIST_INIT(&pp->pr_hashtab[i]);
-		}
+		off = palloc->pa_pagesz;
+		SPLAY_INIT(&pp->pr_phtree);
 	}
 
 	/*
 	 * Alignment is to take place at `ioff' within the item. This means
 	 * we must reserve up to `align - 1' bytes on the page to allow
 	 * appropriate positioning of each item.
-	 *
-	 * Silently enforce `0 <= ioff < align'.
 	 */
-	pp->pr_itemoffset = ioff = ioff % align;
 	pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
+	KASSERT(pp->pr_itemsperpage != 0);
+	if ((pp->pr_roflags & PR_NOTOUCH)) {
+		int idx;
+
+		for (idx = 0; pp->pr_itemsperpage > PHPOOL_FREELIST_NELEM(idx);
+		    idx++) {
+			/* nothing */
+		}
+		if (idx >= PHPOOL_MAX) {
+			/*
+			 * if you see this panic, consider to tweak
+			 * PHPOOL_MAX and PHPOOL_FREELIST_NELEM.
+			 */
+			panic("%s: too large itemsperpage(%d) for PR_NOTOUCH",
+			    pp->pr_wchan, pp->pr_itemsperpage);
+		}
+		pp->pr_phpool = &phpool[idx];
+	} else if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
+		pp->pr_phpool = &phpool[0];
+	}
+#if defined(DIAGNOSTIC)
+	else {
+		pp->pr_phpool = NULL;
+	}
+#endif
 
 	/*
 	 * Use the slack between the chunks and the page header
@@ -464,6 +638,7 @@ pool_init(struct pool *pp, size_t size, 
 	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),
@@ -472,6 +647,7 @@ pool_init(struct pool *pp, size_t size, 
 		pp->pr_curlogentry = 0;
 		pp->pr_logsize = pool_logsize;
 	}
+#endif
 
 	pp->pr_entered_file = NULL;
 	pp->pr_entered_line = 0;
@@ -479,18 +655,47 @@ pool_init(struct pool *pp, size_t size, 
 	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) {
-		pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
-			  0, "phpool", 0, 0, 0, 0);
+	if (phpool[0].pr_size == 0) {
+		int idx;
+		for (idx = 0; idx < PHPOOL_MAX; idx++) {
+			static char phpool_names[PHPOOL_MAX][6+1+6+1];
+			int nelem;
+			size_t sz;
+
+			nelem = PHPOOL_FREELIST_NELEM(idx);
+			snprintf(phpool_names[idx], sizeof(phpool_names[idx]),
+			    "phpool-%d", nelem);
+			sz = sizeof(struct pool_item_header);
+			if (nelem) {
+				sz = PR_FREELIST_ALIGN(sz)
+				    + nelem * sizeof(pool_item_freelist_t);
+			}
+			pool_init(&phpool[idx], sz, 0, 0, 0,
+			    phpool_names[idx], &pool_allocator_meta);
+		}
+#ifdef POOL_SUBPAGE
+		pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
+		    PR_RECURSIVE, "psppool", &pool_allocator_meta);
+#endif
+		pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
+		    0, "pcgpool", &pool_allocator_meta);
 	}
 
 	/* 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. */
+	s = splvm();
+	simple_lock(&palloc->pa_slock);
+	TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
+	simple_unlock(&palloc->pa_slock);
+	splx(s);
 }
 
 /*
@@ -499,57 +704,115 @@ pool_init(struct pool *pp, size_t size, 
 void
 pool_destroy(struct pool *pp)
 {
+	struct pool_pagelist pq;
 	struct pool_item_header *ph;
+	int s;
+
+	/* Remove from global pool list */
+	simple_lock(&pool_head_slock);
+	TAILQ_REMOVE(&pool_head, pp, pr_poollist);
+	if (drainpp == pp)
+		drainpp = NULL;
+	simple_unlock(&pool_head_slock);
+
+	/* Remove this pool from its allocator's list of pools. */
+	s = splvm();
+	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);
+	splx(s);
+
+	s = splvm();
+	simple_lock(&pp->pr_slock);
+
+	KASSERT(TAILQ_EMPTY(&pp->pr_cachelist));
 
 #ifdef DIAGNOSTIC
 	if (pp->pr_nout != 0) {
 		pr_printlog(pp, NULL, printf);
-		panic("pool_destroy: pool busy: still out: %u\n",
+		panic("pool_destroy: pool busy: still out: %u",
 		    pp->pr_nout);
 	}
 #endif
 
+	KASSERT(LIST_EMPTY(&pp->pr_fullpages));
+	KASSERT(LIST_EMPTY(&pp->pr_partpages));
+
 	/* Remove all pages */
-	if ((pp->pr_roflags & PR_STATIC) == 0)
-		while ((ph = pp->pr_pagelist.tqh_first) != NULL)
-			pr_rmpage(pp, ph);
+	LIST_INIT(&pq);
+	while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
+		pr_rmpage(pp, ph, &pq);
 
-	/* 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;
-	simple_unlock(&pool_head_slock);
+	simple_unlock(&pp->pr_slock);
+	splx(s);
 
+	pr_pagelist_free(pp, &pq);
+
+#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 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 = splvm();
+		ph = pool_get(pp->pr_phpool, flags);
+		splx(s);
+	}
+
+	return (ph);
+}
 
 /*
  * Grab an item from the pool; must be called at appropriate spl level
  */
 void *
+#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 &&
+	if (__predict_false(pp->pr_itemsperpage == 0))
+		panic("pool_get: pool %p: pr_itemsperpage is zero, "
+		    "pool not initialized?", pp);
+	if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
 			    (flags & PR_WAITOK) != 0))
-		panic("pool_get: must have NOWAIT");
+		panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
+
+#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);
@@ -568,6 +831,21 @@ _pool_get(struct pool *pp, int flags, co
 	}
 #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
@@ -588,9 +866,6 @@ _pool_get(struct pool *pp, int flags, co
 			      &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);
@@ -605,14 +880,12 @@ _pool_get(struct pool *pp, int flags, co
 	 * 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);
 			printf("pool_get: %s: curpage NULL, nitems %u\n",
 			    pp->pr_wchan, pp->pr_nitems);
-			panic("pool_get: nitems inconsistent\n");
+			panic("pool_get: nitems inconsistent");
 		}
 #endif
 
@@ -623,23 +896,26 @@ _pool_get(struct pool *pp, int flags, co
 		 */
 		pr_leave(pp);
 		simple_unlock(&pp->pr_slock);
-		v = (*pp->pr_alloc)(pp->pr_pagesz, flags, pp->pr_mtype);
-		simple_lock(&pp->pr_slock);
-		pr_enter(pp, file, line);
+		v = pool_allocator_alloc(pp, flags);
+		if (__predict_true(v != NULL))
+			ph = pool_alloc_item_header(pp, v, flags);
+
+		if (__predict_false(v == NULL || ph == NULL)) {
+			if (v != NULL)
+				pool_allocator_free(pp, v);
+
+			simple_lock(&pp->pr_slock);
+			pr_enter(pp, file, line);
 
-		if (v == NULL) {
 			/*
-			 * 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);
@@ -650,15 +926,11 @@ _pool_get(struct pool *pp, int flags, co
 			/*
 			 * 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);
@@ -666,42 +938,61 @@ _pool_get(struct pool *pp, int flags, co
 		}
 
 		/* We have more memory; add it to the pool */
+		simple_lock(&pp->pr_slock);
+		pr_enter(pp, file, line);
+		pool_prime_page(pp, v, ph);
 		pp->pr_npagealloc++;
-		pool_prime_page(pp, v);
 
 		/* Start the allocation process over. */
 		goto startover;
 	}
-
-	if (__predict_false((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL)) {
-		pr_leave(pp);
-		simple_unlock(&pp->pr_slock);
-		panic("pool_get: %s: page empty", pp->pr_wchan);
-	}
+	if (pp->pr_roflags & PR_NOTOUCH) {
 #ifdef DIAGNOSTIC
-	if (__predict_false(pp->pr_nitems == 0)) {
-		pr_leave(pp);
-		simple_unlock(&pp->pr_slock);
-		printf("pool_get: %s: items on itemlist, nitems %u\n",
-		    pp->pr_wchan, pp->pr_nitems);
-		panic("pool_get: nitems inconsistent\n");
-	}
+		if (__predict_false(ph->ph_nmissing == pp->pr_itemsperpage)) {
+			pr_leave(pp);
+			simple_unlock(&pp->pr_slock);
+			panic("pool_get: %s: page empty", pp->pr_wchan);
+		}
+#endif
+		v = pr_item_notouch_get(pp, ph);
+#ifdef POOL_DIAGNOSTIC
+		pr_log(pp, v, PRLOG_GET, file, line);
+#endif
+	} else {
+		v = pi = TAILQ_FIRST(&ph->ph_itemlist);
+		if (__predict_false(v == NULL)) {
+			pr_leave(pp);
+			simple_unlock(&pp->pr_slock);
+			panic("pool_get: %s: page empty", pp->pr_wchan);
+		}
+#ifdef DIAGNOSTIC
+		if (__predict_false(pp->pr_nitems == 0)) {
+			pr_leave(pp);
+			simple_unlock(&pp->pr_slock);
+			printf("pool_get: %s: items on itemlist, nitems %u\n",
+			    pp->pr_wchan, pp->pr_nitems);
+			panic("pool_get: nitems inconsistent");
+		}
+#endif
+
+#ifdef POOL_DIAGNOSTIC
+		pr_log(pp, v, PRLOG_GET, file, line);
 #endif
-	pr_log(pp, v, PRLOG_GET, file, line);
 
 #ifdef DIAGNOSTIC
-	if (__predict_false(pi->pi_magic != PI_MAGIC)) {
-		pr_printlog(pp, pi, printf);
-		panic("pool_get(%s): free list modified: magic=%x; page %p;"
-		       " item addr %p\n",
-			pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
-	}
+		if (__predict_false(pi->pi_magic != PI_MAGIC)) {
+			pr_printlog(pp, pi, printf);
+			panic("pool_get(%s): free list modified: "
+			    "magic=%x; page %p; item addr %p\n",
+			    pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
+		}
 #endif
 
-	/*
-	 * Remove from item list.
-	 */
-	TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
+		/*
+		 * Remove from item list.
+		 */
+		TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
+	}
 	pp->pr_nitems--;
 	pp->pr_nout++;
 	if (ph->ph_nmissing == 0) {
@@ -710,11 +1001,19 @@ _pool_get(struct pool *pp, int flags, co
 			panic("pool_get: nidle inconsistent");
 #endif
 		pp->pr_nidle--;
+
+		/*
+		 * This page was previously empty.  Move it to the list of
+		 * partially-full pages.  This page is already curpage.
+		 */
+		LIST_REMOVE(ph, ph_pagelist);
+		LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
 	}
 	ph->ph_nmissing++;
-	if (TAILQ_FIRST(&ph->ph_itemlist) == NULL) {
+	if (ph->ph_nmissing == pp->pr_itemsperpage) {
 #ifdef DIAGNOSTIC
-		if (__predict_false(ph->ph_nmissing != pp->pr_itemsperpage)) {
+		if (__predict_false((pp->pr_roflags & PR_NOTOUCH) == 0 &&
+		    !TAILQ_EMPTY(&ph->ph_itemlist))) {
 			pr_leave(pp);
 			simple_unlock(&pp->pr_slock);
 			panic("pool_get: %s: nmissing inconsistent",
@@ -722,24 +1021,12 @@ _pool_get(struct pool *pp, int flags, co
 		}
 #endif
 		/*
-		 * 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.
-		 *
-		 * Migrate empty pages to the end of the list.  This
-		 * will speed the update of curpage as pages become
-		 * idle.  Empty pages intermingled with idle pages
-		 * is no big deal.  As soon as a page becomes un-empty,
-		 * it will move back to the head of the list.
+		 * This page is now full.  Move it to the full list
+		 * and select a new current page.
 		 */
-		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))
-			if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
-				break;
-
-		pp->pr_curpage = ph;
+		LIST_REMOVE(ph, ph_pagelist);
+		LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
+		pool_update_curpage(pp);
 	}
 
 	pp->pr_nget++;
@@ -748,7 +1035,7 @@ _pool_get(struct pool *pp, int flags, co
 	 * 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
@@ -762,20 +1049,19 @@ _pool_get(struct pool *pp, int flags, co
 }
 
 /*
- * 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(struct pool *pp, void *v, const char *file, long line)
+static void
+pool_do_put(struct pool *pp, void *v, struct pool_pagelist *pq)
 {
 	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)) {
@@ -785,8 +1071,6 @@ _pool_put(struct pool *pp, void *v, cons
 	}
 #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);
@@ -802,20 +1086,25 @@ _pool_put(struct pool *pp, void *v, cons
 	/*
 	 * Return to item list.
 	 */
+	if (pp->pr_roflags & PR_NOTOUCH) {
+		pr_item_notouch_put(pp, ph, v);
+	} else {
 #ifdef DIAGNOSTIC
-	pi->pi_magic = PI_MAGIC;
+		pi->pi_magic = PI_MAGIC;
 #endif
 #ifdef DEBUG
-	{
-		int i, *ip = v;
+		{
+			int i, *ip = v;
 
-		for (i = 0; i < pp->pr_size / sizeof(int); i++) {
-			*ip++ = PI_MAGIC;
+			for (i = 0; i < pp->pr_size / sizeof(int); i++) {
+				*ip++ = PI_MAGIC;
+			}
 		}
-	}
 #endif
 
-	TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
+		TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
+	}
+	KDASSERT(ph->ph_nmissing != 0);
 	ph->ph_nmissing--;
 	pp->pr_nput++;
 	pp->pr_nitems++;
@@ -829,33 +1118,32 @@ _pool_put(struct pool *pp, void *v, cons
 		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;
 	}
 
 	/*
-	 * If this page is now complete, do one of two things:
+	 * If this page is now empty, do one of two things:
+	 *
+	 *	(1) If we have more pages than the page high water mark,
+	 *	    free the page back to the system.  ONLY CONSIDER
+	 *	    FREEING BACK A PAGE IF WE HAVE MORE THAN OUR MINIMUM PAGE
+	 *	    CLAIM.
 	 *
-	 *	(1) If we have more pages than the page high water
-	 *	    mark, free the page back to the system.
+	 *	(2) Otherwise, move the page to the empty page list.
 	 *
-	 *	(2) Move it to the end of the page list, so that
-	 *	    we minimize our chances of fragmenting the
-	 *	    pool.  Idle pages migrate to the end (along with
-	 *	    completely empty pages, so that we find un-empty
-	 *	    pages more quickly when we update curpage) of the
-	 *	    list so they can be more easily swept up by
-	 *	    the pagedaemon when pages are scarce.
+	 * Either way, select a new current page (so we use a partially-full
+	 * page if one is available).
 	 */
 	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_minpages &&
+		    (pp->pr_npages > pp->pr_maxpages ||
+		     (pp->pr_alloc->pa_flags & PA_WANT) != 0)) {
+			pr_rmpage(pp, ph, pq);
 		} else {
-			TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
-			TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
+			LIST_REMOVE(ph, ph_pagelist);
+			LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
 
 			/*
 			 * Update the timestamp on the page.  A page must
@@ -866,84 +1154,102 @@ _pool_put(struct pool *pp, void *v, cons
 			s = splclock();
 			ph->ph_time = mono_time;
 			splx(s);
-
-			/*
-			 * Update the current page pointer.  Just look for
-			 * the first page with any free items.
-			 *
-			 * XXX: Maybe we want an option to look for the
-			 * page with the fewest available items, to minimize
-			 * fragmentation?
-			 */
-			for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
-			     ph = TAILQ_NEXT(ph, ph_pagelist))
-				if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
-					break;
-
-			pp->pr_curpage = ph;
 		}
+		pool_update_curpage(pp);
 	}
+
 	/*
-	 * If the page has just become un-empty, move it to the head of
-	 * the list, and make it the current page.  The next allocation
-	 * will get the item from this page, instead of further fragmenting
-	 * the pool.
+	 * If the page was previously completely full, move it to the
+	 * partially-full list and make it the current page.  The next
+	 * allocation will get the item from this page, instead of
+	 * further fragmenting the pool.
 	 */
 	else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
-		TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
-		TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
+		LIST_REMOVE(ph, ph_pagelist);
+		LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
 		pp->pr_curpage = ph;
 	}
-
-	pr_leave(pp);
-	simple_unlock(&pp->pr_slock);
-
 }
 
 /*
- * Add N items to the pool.
+ * Return resource to the pool; must be called at appropriate spl level
  */
-int
-pool_prime(struct pool *pp, int n, caddr_t storage)
+#ifdef POOL_DIAGNOSTIC
+void
+_pool_put(struct pool *pp, void *v, const char *file, long line)
 {
-	caddr_t cp;
-	int newnitems, newpages;
+	struct pool_pagelist pq;
 
-#ifdef DIAGNOSTIC
-	if (__predict_false(storage && !(pp->pr_roflags & PR_STATIC)))
-		panic("pool_prime: static");
-	/* !storage && static caught below */
-#endif
+	LIST_INIT(&pq);
 
 	simple_lock(&pp->pr_slock);
+	pr_enter(pp, file, line);
 
-	newnitems = pp->pr_minitems + n;
-	newpages =
-		roundup(newnitems, pp->pr_itemsperpage) / pp->pr_itemsperpage
-		- pp->pr_minpages;
+	pr_log(pp, v, PRLOG_PUT, file, line);
 
-	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);
-		}
+	pool_do_put(pp, v, &pq);
 
-		if (cp == NULL) {
-			simple_unlock(&pp->pr_slock);
-			return (ENOMEM);
-		}
+	pr_leave(pp);
+	simple_unlock(&pp->pr_slock);
 
-		pp->pr_npagealloc++;
-		pool_prime_page(pp, cp);
+	if (! LIST_EMPTY(&pq))
+		pr_pagelist_free(pp, &pq);
+}
+#undef pool_put
+#endif /* POOL_DIAGNOSTIC */
+
+void
+pool_put(struct pool *pp, void *v)
+{
+	struct pool_pagelist pq;
+
+	LIST_INIT(&pq);
+
+	simple_lock(&pp->pr_slock);
+	pool_do_put(pp, v, &pq);
+	simple_unlock(&pp->pr_slock);
+
+	if (! LIST_EMPTY(&pq))
+		pr_pagelist_free(pp, &pq);
+}
+
+#ifdef POOL_DIAGNOSTIC
+#define		pool_put(h, v)	_pool_put((h), (v), __FILE__, __LINE__)
+#endif
+
+/*
+ * Add N items to the pool.
+ */
+int
+pool_prime(struct pool *pp, int n)
+{
+	struct pool_item_header *ph = NULL;
+	caddr_t cp;
+	int newpages;
+
+	simple_lock(&pp->pr_slock);
+
+	newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
+
+	while (newpages-- > 0) {
+		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);
+
+		if (__predict_false(cp == NULL || ph == NULL)) {
+			if (cp != NULL)
+				pool_allocator_free(pp, cp);
+			simple_lock(&pp->pr_slock);
+			break;
+		}
+
+		simple_lock(&pp->pr_slock);
+		pool_prime_page(pp, cp, ph);
+		pp->pr_npagealloc++;
 		pp->pr_minpages++;
 	}
 
-	pp->pr_minitems = newnitems;
-
 	if (pp->pr_minpages >= pp->pr_maxpages)
 		pp->pr_maxpages = pp->pr_minpages + 1;	/* XXX */
 
@@ -957,36 +1263,34 @@ pool_prime(struct pool *pp, int n, caddr
  * Note, we must be called with the pool descriptor LOCKED.
  */
 static void
-pool_prime_page(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;
+	int s;
 
-	if (((u_long)cp & (pp->pr_pagesz - 1)) != 0)
-		panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
+	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
 
-	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);
-		LIST_INSERT_HEAD(&pp->pr_hashtab[PR_HASH_INDEX(pp, cp)],
-				 ph, ph_hashlist);
-	}
+#ifdef DIAGNOSTIC
+	if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
+		panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
+#endif
 
 	/*
 	 * Insert page header.
 	 */
-	TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
+	LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
 	TAILQ_INIT(&ph->ph_itemlist);
 	ph->ph_page = storage;
 	ph->ph_nmissing = 0;
-	memset(&ph->ph_time, 0, sizeof(ph->ph_time));
+	s = splclock();
+	ph->ph_time = mono_time;
+	splx(s);
+	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
+		SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
 
 	pp->pr_nidle++;
 
@@ -1009,15 +1313,28 @@ pool_prime_page(struct pool *pp, caddr_t
 	n = pp->pr_itemsperpage;
 	pp->pr_nitems += n;
 
-	while (n--) {
-		pi = (struct pool_item *)cp;
+	if (pp->pr_roflags & PR_NOTOUCH) {
+		pool_item_freelist_t *freelist = PR_FREELIST(ph);
+		int i;
+
+		ph->ph_off = cp - storage;
+		ph->ph_firstfree = 0;
+		for (i = 0; i < n - 1; i++)
+			freelist[i] = i + 1;
+		freelist[n - 1] = PR_INDEX_EOL;
+	} else {
+		while (n--) {
+			pi = (struct pool_item *)cp;
 
-		/* Insert on page list */
-		TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
+			KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
+
+			/* Insert on page list */
+			TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
 #ifdef DIAGNOSTIC
-		pi->pi_magic = PI_MAGIC;
+			pi->pi_magic = PI_MAGIC;
 #endif
-		cp = (caddr_t)(cp + pp->pr_size);
+			cp = (caddr_t)(cp + pp->pr_size);
+		}
 	}
 
 	/*
@@ -1031,36 +1348,22 @@ pool_prime_page(struct pool *pp, caddr_t
 }
 
 /*
- * 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 pr_nitems 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(struct pool *pp)
 {
+	struct pool_item_header *ph = NULL;
 	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.
 		 *
@@ -1068,23 +1371,37 @@ pool_catchup(struct pool *pp)
 		 * the pool descriptor?
 		 */
 		simple_unlock(&pp->pr_slock);
-		cp = (*pp->pr_alloc)(pp->pr_pagesz, 0, pp->pr_mtype);
-		simple_lock(&pp->pr_slock);
-		if (__predict_false(cp == NULL)) {
+		cp = pool_allocator_alloc(pp, PR_NOWAIT);
+		if (__predict_true(cp != NULL))
+			ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
+		if (__predict_false(cp == NULL || ph == NULL)) {
+			if (cp != NULL)
+				pool_allocator_free(pp, cp);
 			error = ENOMEM;
+			simple_lock(&pp->pr_slock);
 			break;
 		}
+		simple_lock(&pp->pr_slock);
+		pool_prime_page(pp, cp, ph);
 		pp->pr_npagealloc++;
-		pool_prime_page(pp, cp);
 	}
 
 	return (error);
 }
 
+static void
+pool_update_curpage(struct pool *pp)
+{
+
+	pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
+	if (pp->pr_curpage == NULL) {
+		pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
+	}
+}
+
 void
 pool_setlowat(struct pool *pp, int n)
 {
-	int error;
 
 	simple_lock(&pp->pr_slock);
 
@@ -1094,8 +1411,7 @@ pool_setlowat(struct pool *pp, int 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
@@ -1143,94 +1459,76 @@ pool_sethardlimit(struct pool *pp, int n
 }
 
 /*
- * Default page allocator.
- */
-static void *
-pool_page_alloc(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(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(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(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
+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;
+	struct timeval diff;
 	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);
 
+	LIST_INIT(&pq);
+
+	/*
+	 * Reclaim items from the pool's caches.
+	 */
+	TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
+		pool_cache_reclaim(pc, &pq);
+
 	s = splclock();
 	curtime = mono_time;
 	splx(s);
 
-	for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; ph = phnext) {
-		phnext = TAILQ_NEXT(ph, ph_pagelist);
+	for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
+		phnext = LIST_NEXT(ph, ph_pagelist);
 
 		/* Check our minimum page claim */
 		if (pp->pr_npages <= pp->pr_minpages)
 			break;
 
-		if (ph->ph_nmissing == 0) {
-			struct timeval diff;
-			timersub(&curtime, &ph->ph_time, &diff);
-			if (diff.tv_sec < pool_inactive_time)
-				continue;
+		KASSERT(ph->ph_nmissing == 0);
+		timersub(&curtime, &ph->ph_time, &diff);
+		if (diff.tv_sec < pool_inactive_time)
+			continue;
 
-			/*
-			 * If freeing this page would put us below
-			 * the low water mark, stop now.
-			 */
-			if ((pp->pr_nitems - pp->pr_itemsperpage) <
-			    pp->pr_minitems)
-				break;
+		/*
+		 * If freeing this page would put us below
+		 * the low water mark, stop now.
+		 */
+		if ((pp->pr_nitems - pp->pr_itemsperpage) <
+		    pp->pr_minitems)
+			break;
 
-			pr_rmpage(pp, ph);
-		}
+		pr_rmpage(pp, ph, &pq);
 	}
 
 	pr_leave(pp);
 	simple_unlock(&pp->pr_slock);
-}
+	if (LIST_EMPTY(&pq))
+		return (0);
 
+	pr_pagelist_free(pp, &pq);
+	return (1);
+}
 
 /*
  * Drain pools, one at a time.
@@ -1243,23 +1541,21 @@ 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.
  */
@@ -1268,7 +1564,7 @@ pool_print(struct pool *pp, const char *
 {
 	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);
@@ -1311,13 +1607,39 @@ pool_printit(struct pool *pp, const char
 }
 
 static void
-pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
+pool_print_pagelist(struct pool *pp, struct pool_pagelist *pl,
+    void (*pr)(const char *, ...))
 {
 	struct pool_item_header *ph;
 #ifdef DIAGNOSTIC
 	struct pool_item *pi;
 #endif
-	int print_log = 0, print_pagelist = 0;
+
+	LIST_FOREACH(ph, pl, ph_pagelist) {
+		(*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
+		    ph->ph_page, ph->ph_nmissing,
+		    (u_long)ph->ph_time.tv_sec,
+		    (u_long)ph->ph_time.tv_usec);
+#ifdef DIAGNOSTIC
+		if (!(pp->pr_roflags & PR_NOTOUCH)) {
+			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);
+				}
+			}
+		}
+#endif
+	}
+}
+
+static void
+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;
+	int i, print_log = 0, print_pagelist = 0, print_cache = 0;
 	char c;
 
 	while ((c = *modif++) != '\0') {
@@ -1325,14 +1647,14 @@ pool_print1(struct pool *pp, const char 
 			print_log = 1;
 		if (c == 'p')
 			print_pagelist = 1;
-		modif++;
+		if (c == 'c')
+			print_cache = 1;
 	}
 
 	(*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",
@@ -1346,30 +1668,22 @@ pool_print1(struct pool *pp, const char 
 	if (print_pagelist == 0)
 		goto skip_pagelist;
 
-	if ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL)
-		(*pr)("\n\tpage list:\n");
-	for (; ph != NULL; ph = TAILQ_NEXT(ph, ph_pagelist)) {
-		(*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
-		    ph->ph_page, ph->ph_nmissing,
-		    (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)) {
-			if (pi->pi_magic != PI_MAGIC) {
-				(*pr)("\t\t\titem %p, magic 0x%x\n",
-				    pi, pi->pi_magic);
-			}
-		}
-#endif
-	}
+	if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
+		(*pr)("\n\tempty page list:\n");
+	pool_print_pagelist(pp, &pp->pr_emptypages, pr);
+	if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
+		(*pr)("\n\tfull page list:\n");
+	pool_print_pagelist(pp, &pp->pr_fullpages, pr);
+	if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
+		(*pr)("\n\tpartial-page list:\n");
+	pool_print_pagelist(pp, &pp->pr_partpages, pr);
+
 	if (pp->pr_curpage == NULL)
 		(*pr)("\tno current page\n");
 	else
 		(*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
 
  skip_pagelist:
-
 	if (print_log == 0)
 		goto skip_log;
 
@@ -1380,10 +1694,90 @@ pool_print1(struct pool *pp, const char 
 		pr_printlog(pp, NULL, 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++) {
+				if (pcg->pcg_objects[i].pcgo_pa !=
+				    POOL_PADDR_INVALID) {
+					(*pr)("\t\t\t%p, 0x%llx\n",
+					    pcg->pcg_objects[i].pcgo_va,
+					    (unsigned long long)
+					    pcg->pcg_objects[i].pcgo_pa);
+				} else {
+					(*pr)("\t\t\t%p\n",
+					    pcg->pcg_objects[i].pcgo_va);
+				}
+			}
+		}
+	}
+
+ skip_cache:
 	pr_enter_check(pp, pr);
 }
 
+static int
+pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
+{
+	struct pool_item *pi;
+	caddr_t page;
+	int n;
+
+	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)
+			printf("%s: ", label);
+		printf("pool(%p:%s): page inconsistency: page %p;"
+		       " at page head addr %p (p %p)\n", pp,
+			pp->pr_wchan, ph->ph_page,
+			ph, page);
+		return 1;
+	}
+
+	if ((pp->pr_roflags & PR_NOTOUCH) != 0)
+		return 0;
+
+	for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
+	     pi != NULL;
+	     pi = TAILQ_NEXT(pi,pi_list), n++) {
+
+#ifdef DIAGNOSTIC
+		if (pi->pi_magic != PI_MAGIC) {
+			if (label != NULL)
+				printf("%s: ", label);
+			printf("pool(%s): free list modified: magic=%x;"
+			       " page %p; item ordinal %d;"
+			       " addr %p (p %p)\n",
+				pp->pr_wchan, pi->pi_magic, ph->ph_page,
+				n, pi, page);
+			panic("pool");
+		}
+#endif
+		page =
+		    (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
+		if (page == ph->ph_page)
+			continue;
+
+		if (label != NULL)
+			printf("%s: ", label);
+		printf("pool(%p:%s): page inconsistency: page %p;"
+		       " item ordinal %d; addr %p (p %p)\n", pp,
+			pp->pr_wchan, ph->ph_page,
+			n, pi, page);
+		return 1;
+	}
+	return 0;
+}
+
+
 int
 pool_chk(struct pool *pp, const char *label)
 {
@@ -1391,58 +1785,576 @@ pool_chk(struct pool *pp, const char *la
 	int r = 0;
 
 	simple_lock(&pp->pr_slock);
-
-	for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
-	     ph = TAILQ_NEXT(ph, ph_pagelist)) {
-
-		struct pool_item *pi;
-		int n;
-		caddr_t page;
-
-		page = (caddr_t)((u_long)ph & pp->pr_pagemask);
-		if (page != ph->ph_page &&
-		    (pp->pr_roflags & PR_PHINPAGE) != 0) {
-			if (label != NULL)
-				printf("%s: ", label);
-			printf("pool(%p:%s): page inconsistency: page %p;"
-			       " at page head addr %p (p %p)\n", pp,
-				pp->pr_wchan, ph->ph_page,
-				ph, page);
-			r++;
+	LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
+		r = pool_chk_page(pp, label, ph);
+		if (r) {
 			goto out;
 		}
+	}
+	LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
+		r = pool_chk_page(pp, label, ph);
+		if (r) {
+			goto out;
+		}
+	}
+	LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
+		r = pool_chk_page(pp, label, ph);
+		if (r) {
+			goto out;
+		}
+	}
+
+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)
+{
 
-		for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
-		     pi != NULL;
-		     pi = TAILQ_NEXT(pi,pi_list), n++) {
+	TAILQ_INIT(&pc->pc_grouplist);
+	simple_lock_init(&pc->pc_slock);
 
-#ifdef DIAGNOSTIC
-			if (pi->pi_magic != PI_MAGIC) {
-				if (label != NULL)
-					printf("%s: ", label);
-				printf("pool(%s): free list modified: magic=%x;"
-				       " page %p; item ordinal %d;"
-				       " addr %p (p %p)\n",
-					pp->pr_wchan, pi->pi_magic, ph->ph_page,
-					n, pi, page);
-				panic("pool");
+	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, paddr_t *pap)
+{
+	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].pcgo_va != NULL);
+	object = pcg->pcg_objects[idx].pcgo_va;
+	if (pap != NULL)
+		*pap = pcg->pcg_objects[idx].pcgo_pa;
+	pcg->pcg_objects[idx].pcgo_va = NULL;
+
+	return (object);
+}
+
+static __inline void
+pcg_put(struct pool_cache_group *pcg, void *object, paddr_t pa)
+{
+	u_int idx;
+
+	KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
+	idx = pcg->pcg_avail++;
+
+	KASSERT(pcg->pcg_objects[idx].pcgo_va == NULL);
+	pcg->pcg_objects[idx].pcgo_va = object;
+	pcg->pcg_objects[idx].pcgo_pa = pa;
+}
+
+/*
+ * pool_cache_get{,_paddr}:
+ *
+ *	Get an object from a pool cache (optionally returning
+ *	the physical address of the object).
+ */
+void *
+pool_cache_get_paddr(struct pool_cache *pc, int flags, paddr_t *pap)
+{
+	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);
 			}
+		}
+		if (object != NULL && pap != NULL) {
+#ifdef POOL_VTOPHYS
+			*pap = POOL_VTOPHYS(object);
+#else
+			*pap = POOL_PADDR_INVALID;
 #endif
-			page = (caddr_t)((u_long)pi & pp->pr_pagemask);
-			if (page == ph->ph_page)
+		}
+		return (object);
+	}
+
+ have_group:
+	pc->pc_hits++;
+	pc->pc_nitems--;
+	object = pcg_get(pcg, pap);
+
+	if (pcg->pcg_avail == 0)
+		pc->pc_allocfrom = NULL;
+
+	simple_unlock(&pc->pc_slock);
+
+	return (object);
+}
+
+/*
+ * pool_cache_put{,_paddr}:
+ *
+ *	Put an object back to the pool cache (optionally caching the
+ *	physical address of the object).
+ */
+void
+pool_cache_put_paddr(struct pool_cache *pc, void *object, paddr_t pa)
+{
+	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, pa);
+
+	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_invalidate:
+ *
+ *	Invalidate a pool cache (destruct and release all of the
+ *	cached objects).
+ */
+void
+pool_cache_invalidate(struct pool_cache *pc)
+{
+	struct pool_pagelist pq;
+	struct pool_cache_group *pcg, *npcg;
+	void *object;
+
+	LIST_INIT(&pq);
+
+	simple_lock(&pc->pc_slock);
+	simple_lock(&pc->pc_pool->pr_slock);
+
+	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, NULL);
+			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);
+			pool_do_put(pc->pc_pool, object, &pq);
+		}
+	}
+
+	simple_unlock(&pc->pc_pool->pr_slock);
+	simple_unlock(&pc->pc_slock);
+
+	if (! LIST_EMPTY(&pq))
+		pr_pagelist_free(pc->pc_pool, &pq);
+}
+
+/*
+ * pool_cache_reclaim:
+ *
+ *	Reclaim a pool cache for pool_reclaim().
+ */
+static void
+pool_cache_reclaim(struct pool_cache *pc, struct pool_pagelist *pq)
+{
+	struct pool_cache_group *pcg, *npcg;
+	void *object;
+	int s;
+
+	/*
+	 * We're locking in the wrong order (normally pool_cache -> pool,
+	 * but the pool is already locked when we get here), so we have
+	 * to use trylock.  If we can't lock the pool_cache, it's not really
+	 * a big deal here.
+	 */
+	if (simple_lock_try(&pc->pc_slock) == 0)
+		return;
+
+	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, NULL);
+			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);
+			pool_do_put(pc->pc_pool, object, pq);
+		}
+		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);
+	}
+
+	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;
+
+	LOCK_ASSERT(!simple_lock_held(&org->pr_slock));
+
+	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 (label != NULL)
-				printf("%s: ", label);
-			printf("pool(%p:%s): page inconsistency: page %p;"
-			       " item ordinal %d; addr %p (p %p)\n", pp,
-				pp->pr_wchan, ph->ph_page,
-				n, pi, page);
-			r++;
-			goto out;
+		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;
+
+	LOCK_ASSERT(!simple_lock_held(&pp->pr_slock));
+
+	(*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;
 	}
-out:
-	simple_unlock(&pp->pr_slock);
-	return (r);
+
+	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_cache(kmem_map, waitok));
+}
+
+void
+pool_page_free(struct pool *pp, void *v)
+{
+
+	uvm_km_free_poolpage_cache(kmem_map, (vaddr_t) v);
+}
+
+static void *
+pool_page_alloc_meta(struct pool *pp, int flags)
+{
+	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
+
+	return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok));
+}
+
+static void
+pool_page_free_meta(struct pool *pp, void *v)
+{
+
+	uvm_km_free_poolpage(kmem_map, (vaddr_t) v);
+}
+
+#ifdef POOL_SUBPAGE
+/* Sub-page allocator, for machines with large hardware pages. */
+void *
+pool_subpage_alloc(struct pool *pp, int flags)
+{
+	void *v;
+	int s;
+	s = splvm();
+	v = pool_get(&psppool, flags);
+	splx(s);
+	return v;
+}
+
+void
+pool_subpage_free(struct pool *pp, void *v)
+{
+	int s;
+	s = splvm();
+	pool_put(&psppool, v);
+	splx(s);
+}
+
+/* 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_poolpage_cache(kernel_map, waitok));
+}
+
+void
+pool_page_free_nointr(struct pool *pp, void *v)
+{
+
+	uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v);
 }
+#endif /* POOL_SUBPAGE */