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 1.76.6.1
retrieving revision 1.122.2.3
diff -u -p -r1.76.6.1 -r1.122.2.3
--- src/sys/kern/subr_pool.c	2002/11/11 23:31:43	1.76.6.1
+++ src/sys/kern/subr_pool.c	2007/01/19 20:49:54	1.122.2.3
@@ -1,7 +1,7 @@
-/*	$NetBSD: subr_pool.c,v 1.76.6.1 2002/11/11 23:31:43 he Exp $	*/
+/*	$NetBSD: subr_pool.c,v 1.122.2.3 2007/01/19 20:49:54 ad Exp $	*/
 
 /*-
- * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc.
+ * Copyright (c) 1997, 1999, 2000, 2002 The NetBSD Foundation, Inc.
  * All rights reserved.
  *
  * This code is derived from software contributed to The NetBSD Foundation
@@ -38,7 +38,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.76.6.1 2002/11/11 23:31:43 he Exp $");
+__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.122.2.3 2007/01/19 20:49:54 ad Exp $");
 
 #include "opt_pool.h"
 #include "opt_poollog.h"
@@ -53,31 +53,48 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,
 #include <sys/lock.h>
 #include <sys/pool.h>
 #include <sys/syslog.h>
+#include <sys/debug.h>
 
 #include <uvm/uvm.h>
 
 /*
  * 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);
+LIST_HEAD(,pool) pool_head = LIST_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 SLIST_HEAD(, pool_allocator) pa_deferinitq =
+    SLIST_HEAD_INITIALIZER(pa_deferinitq);
+
+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,
+	.pa_backingmapptr = &kmem_map,
+};
+
 /* # of seconds to retain page after last use */
 int pool_inactive_time = 10;
 
@@ -87,32 +104,49 @@ 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 {
+			LIST_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 */
 };
-TAILQ_HEAD(pool_pagelist,pool_item_header);
+#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;
+	LIST_ENTRY(pool_item)	pi_list;
 };
 
-#define	PR_HASH_INDEX(pp,addr) \
-	(((u_long)(addr) >> (pp)->pr_alloc->pa_pageshift) & \
-	 (PR_HASHTABSIZE - 1))
-
 #define	POOL_NEEDS_CATCHUP(pp)						\
 	((pp)->pr_nitems < (pp)->pr_minitems)
 
@@ -145,18 +179,27 @@ struct pool_item {
 /* The cache group pool. */
 static struct pool pcgpool;
 
-static void	pool_cache_reclaim(struct pool_cache *);
+static void	pool_cache_reclaim(struct pool_cache *, struct pool_pagelist *,
+				   struct pool_cache_grouplist *);
+static void	pcg_grouplist_free(struct pool_cache_grouplist *);
 
 static int	pool_catchup(struct pool *);
 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 int	pool_grow(struct pool *, int);
+static void	*pool_allocator_alloc(struct pool *, int);
+static 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_init().
  */
@@ -169,6 +212,7 @@ struct pool_log {
 	void		*pl_addr;
 };
 
+#ifdef POOL_DIAGNOSTIC
 /* Number of entries in pool log buffers */
 #ifndef POOL_LOGSIZE
 #define	POOL_LOGSIZE	10
@@ -176,8 +220,7 @@ struct pool_log {
 
 int pool_logsize = POOL_LOGSIZE;
 
-#ifdef POOL_DIAGNOSTIC
-static __inline void
+static inline void
 pr_log(struct pool *pp, void *v, int action, const char *file, long line)
 {
 	int n = pp->pr_curlogentry;
@@ -230,7 +273,7 @@ pr_printlog(struct pool *pp, struct pool
 	}
 }
 
-static __inline void
+static inline void
 pr_enter(struct pool *pp, const char *file, long line)
 {
 
@@ -246,7 +289,7 @@ pr_enter(struct pool *pp, const char *fi
 	pp->pr_entered_line = line;
 }
 
-static __inline void
+static inline void
 pr_leave(struct pool *pp)
 {
 
@@ -259,7 +302,7 @@ pr_leave(struct pool *pp)
 	pp->pr_entered_line = 0;
 }
 
-static __inline void
+static inline void
 pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
 {
 
@@ -275,34 +318,132 @@ pr_enter_check(struct pool *pp, void (*p
 #define	pr_enter_check(pp, pr)
 #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)
+{
+
+	/*
+	 * we consider pool_item_header with smaller ph_page bigger.
+	 * (this unnatural ordering is for the benefit of pr_find_pagehead.)
+	 */
+
+	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.
+ * Return the pool page header based on item address.
  */
-static __inline struct pool_item_header *
-pr_find_pagehead(struct pool *pp, caddr_t page)
+static inline struct pool_item_header *
+pr_find_pagehead(struct pool *pp, void *v)
 {
-	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));
+	if ((pp->pr_roflags & PR_NOALIGN) != 0) {
+		tmp.ph_page = (caddr_t)(uintptr_t)v;
+		ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
+		if (ph == NULL) {
+			ph = SPLAY_ROOT(&pp->pr_phtree);
+			if (ph != NULL && phtree_compare(&tmp, ph) >= 0) {
+				ph = SPLAY_NEXT(phtree, &pp->pr_phtree, ph);
+			}
+			KASSERT(ph == NULL || phtree_compare(&tmp, ph) < 0);
+		}
+	} else {
+		caddr_t page =
+		    (caddr_t)((uintptr_t)v & pp->pr_alloc->pa_pagemask);
+
+		if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
+			ph = (void *)(page + pp->pr_phoffset);
+		} else {
+			tmp.ph_page = page;
+			ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
+		}
+	}
+
+	KASSERT(ph == NULL || ((pp->pr_roflags & PR_PHINPAGE) != 0) ||
+	    (ph->ph_page <= (char *)v &&
+	    (char *)v < ph->ph_page + pp->pr_alloc->pa_pagesz));
+	return ph;
+}
+
+static void
+pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq)
+{
+	struct pool_item_header *ph;
+	int s;
 
-	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);
+	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
+static inline void
 pr_rmpage(struct pool *pp, struct pool_item_header *ph,
      struct pool_pagelist *pq)
 {
-	int s;
+
+	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
 
 	/*
 	 * If the page was idle, decrement the idle page count.
@@ -320,34 +461,112 @@ pr_rmpage(struct pool *pp, struct pool_i
 	pp->pr_nitems -= pp->pr_itemsperpage;
 
 	/*
-	 * Unlink a page from the pool and release it (or queue it for release).
+	 * Unlink the page from the pool and queue it for release.
 	 */
-	TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
-	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);
-		}
-	}
+	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_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.
-		 */
-		TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist)
-			if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
-				break;
+	pool_update_curpage(pp);
+}
 
-		pp->pr_curpage = ph;
+static boolean_t
+pa_starved_p(struct pool_allocator *pa)
+{
+
+	if (pa->pa_backingmap != NULL) {
+		return vm_map_starved_p(pa->pa_backingmap);
+	}
+	return FALSE;
+}
+
+static int
+pool_reclaim_callback(struct callback_entry *ce, void *obj, void *arg)
+{
+	struct pool *pp = obj;
+	struct pool_allocator *pa = pp->pr_alloc;
+
+	KASSERT(&pp->pr_reclaimerentry == ce);
+	pool_reclaim(pp);
+	if (!pa_starved_p(pa)) {
+		return CALLBACK_CHAIN_ABORT;
+	}
+	return CALLBACK_CHAIN_CONTINUE;
+}
+
+static void
+pool_reclaim_register(struct pool *pp)
+{
+	struct vm_map *map = pp->pr_alloc->pa_backingmap;
+	int s;
+
+	if (map == NULL) {
+		return;
+	}
+
+	s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
+	callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback,
+	    &pp->pr_reclaimerentry, pp, pool_reclaim_callback);
+	splx(s);
+}
+
+static void
+pool_reclaim_unregister(struct pool *pp)
+{
+	struct vm_map *map = pp->pr_alloc->pa_backingmap;
+	int s;
+
+	if (map == NULL) {
+		return;
+	}
+
+	s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
+	callback_unregister(&vm_map_to_kernel(map)->vmk_reclaim_callback,
+	    &pp->pr_reclaimerentry);
+	splx(s);
+}
+
+static void
+pa_reclaim_register(struct pool_allocator *pa)
+{
+	struct vm_map *map = *pa->pa_backingmapptr;
+	struct pool *pp;
+
+	KASSERT(pa->pa_backingmap == NULL);
+	if (map == NULL) {
+		SLIST_INSERT_HEAD(&pa_deferinitq, pa, pa_q);
+		return;
+	}
+	pa->pa_backingmap = map;
+	TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
+		pool_reclaim_register(pp);
+	}
+}
+
+/*
+ * Initialize all the pools listed in the "pools" link set.
+ */
+void
+pool_subsystem_init(void)
+{
+	struct pool_allocator *pa;
+	__link_set_decl(pools, struct link_pool_init);
+	struct link_pool_init * const *pi;
+
+	__link_set_foreach(pi, pools)
+		pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
+		    (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
+		    (*pi)->palloc);
+
+	while ((pa = SLIST_FIRST(&pa_deferinitq)) != NULL) {
+		KASSERT(pa->pa_backingmapptr != NULL);
+		KASSERT(*pa->pa_backingmapptr != NULL);
+		SLIST_REMOVE_HEAD(&pa_deferinitq, pa_q);
+		pa_reclaim_register(pa);
 	}
 }
 
@@ -361,7 +580,26 @@ void
 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;
+#ifdef DEBUG
+	struct pool *pp1;
+#endif
+	size_t trysize, phsize;
+	int off, slack, s;
+
+	KASSERT((1UL << (CHAR_BIT * sizeof(pool_item_freelist_t))) - 2 >=
+	    PHPOOL_FREELIST_NELEM(PHPOOL_MAX - 1));
+
+#ifdef DEBUG
+	/*
+	 * Check that the pool hasn't already been initialised and
+	 * added to the list of all pools.
+	 */
+	LIST_FOREACH(pp1, &pool_head, pr_poollist) {
+		if (pp == pp1)
+			panic("pool_init: pool %s already initialised",
+			    wchan);
+	}
+#endif
 
 #ifdef POOL_DIAGNOSTIC
 	/*
@@ -371,62 +609,51 @@ 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;
-	}
-	/*
-	 * We'll assume any user-specified back-end allocator
-	 * will deal with sub-pages, or simply don't care.
-	 */
-#else
 	if (palloc == NULL)
 		palloc = &pool_allocator_kmem;
+#ifdef POOL_SUBPAGE
+	if (size > palloc->pa_pagesz) {
+		if (palloc == &pool_allocator_kmem)
+			palloc = &pool_allocator_kmem_fullpage;
+		else if (palloc == &pool_allocator_nointr)
+			palloc = &pool_allocator_nointr_fullpage;
+	}		
 #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
+		if (palloc->pa_pagesz == 0)
 			palloc->pa_pagesz = PAGE_SIZE;
-#endif /* POOL_SUBPAGE */
-		}
 
 		TAILQ_INIT(&palloc->pa_list);
 
 		simple_lock_init(&palloc->pa_slock);
 		palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
 		palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
+
+		if (palloc->pa_backingmapptr != NULL) {
+			pa_reclaim_register(palloc);
+		}
 		palloc->pa_flags |= PA_INITIALIZED;
 	}
 
 	if (align == 0)
 		align = ALIGN(1);
 
-	if (size < sizeof(struct pool_item))
+	if ((flags & PR_NOTOUCH) == 0 && size < sizeof(struct pool_item))
 		size = sizeof(struct pool_item);
 
 	size = roundup(size, align);
 #ifdef DIAGNOSTIC
 	if (size > palloc->pa_pagesz)
-		panic("pool_init: pool item size (%lu) too large",
-		      (u_long)size);
+		panic("pool_init: pool item size (%zu) too large", size);
 #endif
 
 	/*
 	 * Initialize the pool structure.
 	 */
-	TAILQ_INIT(&pp->pr_pagelist);
-	TAILQ_INIT(&pp->pr_cachelist);
+	LIST_INIT(&pp->pr_emptypages);
+	LIST_INIT(&pp->pr_fullpages);
+	LIST_INIT(&pp->pr_partpages);
+	LIST_INIT(&pp->pr_cachelist);
 	pp->pr_curpage = NULL;
 	pp->pr_npages = 0;
 	pp->pr_minitems = 0;
@@ -448,38 +675,69 @@ pool_init(struct pool *pp, size_t size, 
 	pp->pr_hardlimit_warning_last.tv_usec = 0;
 	pp->pr_drain_hook = NULL;
 	pp->pr_drain_hook_arg = NULL;
+	pp->pr_freecheck = 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 < palloc->pa_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 | PR_NOALIGN)) == 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 = palloc->pa_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 = palloc->pa_pagesz;
-		for (i = 0; i < PR_HASHTABSIZE; i++) {
-			LIST_INIT(&pp->pr_hashtab[i]);
-		}
+		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
@@ -518,29 +776,44 @@ pool_init(struct pool *pp, size_t size, 
 	 * haven't done so yet.
 	 * XXX LOCKING.
 	 */
-	if (phpool.pr_size == 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(&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", NULL);
+		    PR_RECURSIVE, "psppool", &pool_allocator_meta);
 #endif
 		pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
-		    0, "pcgpool", NULL);
+		    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);
+	LIST_INSERT_HEAD(&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);
+	pool_reclaim_register(pp);
 }
 
 /*
@@ -549,37 +822,50 @@ pool_init(struct pool *pp, size_t size, 
 void
 pool_destroy(struct pool *pp)
 {
+	struct pool_pagelist pq;
 	struct pool_item_header *ph;
-	struct pool_cache *pc;
+	int s;
+
+	/* Remove from global pool list */
+	simple_lock(&pool_head_slock);
+	LIST_REMOVE(pp, pr_poollist);
+	if (drainpp == pp)
+		drainpp = NULL;
+	simple_unlock(&pool_head_slock);
 
-	/* Locking order: pool_allocator -> pool */
+	/* Remove this pool from its allocator's list of pools. */
+	pool_reclaim_unregister(pp);
+	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);
 
-	/* Destroy all caches for this pool. */
-	while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL)
-		pool_cache_destroy(pc);
+	s = splvm();
+	simple_lock(&pp->pr_slock);
+
+	KASSERT(LIST_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 */
-	while ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL)
-		pr_rmpage(pp, ph, NULL);
+	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);
-	if (drainpp == pp) {
-		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)
@@ -600,7 +886,7 @@ pool_set_drain_hook(struct pool *pp, voi
 	pp->pr_drain_hook_arg = arg;
 }
 
-static __inline struct pool_item_header *
+static struct pool_item_header *
 pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
 {
 	struct pool_item_header *ph;
@@ -611,8 +897,8 @@ pool_alloc_item_header(struct pool *pp, 
 	if ((pp->pr_roflags & PR_PHINPAGE) != 0)
 		ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
 	else {
-		s = splhigh();
-		ph = pool_get(&phpool, flags);
+		s = splvm();
+		ph = pool_get(pp->pr_phpool, flags);
 		splx(s);
 	}
 
@@ -634,15 +920,18 @@ pool_get(struct pool *pp, int flags)
 	void *v;
 
 #ifdef DIAGNOSTIC
-	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);
 
+#endif /* DIAGNOSTIC */
 #ifdef LOCKDEBUG
 	if (flags & PR_WAITOK)
-		simple_lock_only_held(NULL, "pool_get(PR_WAITOK)");
+		ASSERT_SLEEPABLE(NULL, "pool_get(PR_WAITOK)");
 #endif
-#endif /* DIAGNOSTIC */
 
 	simple_lock(&pp->pr_slock);
 	pr_enter(pp, file, line);
@@ -710,12 +999,14 @@ pool_get(struct pool *pp, int flags)
 	 * has no items in its bucket.
 	 */
 	if ((ph = pp->pr_curpage) == NULL) {
+		int error;
+
 #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
 
@@ -725,17 +1016,9 @@ pool_get(struct pool *pp, int flags)
 		 * may block.
 		 */
 		pr_leave(pp);
-		simple_unlock(&pp->pr_slock);
-		v = pool_allocator_alloc(pp, flags);
-		if (__predict_true(v != NULL))
-			ph = pool_alloc_item_header(pp, v, flags);
-		simple_lock(&pp->pr_slock);
+		error = pool_grow(pp, flags);
 		pr_enter(pp, file, line);
-
-		if (__predict_false(v == NULL || ph == NULL)) {
-			if (v != NULL)
-				pool_allocator_free(pp, v);
-
+		if (error != 0) {
 			/*
 			 * We were unable to allocate a page or item
 			 * header, but we released the lock during
@@ -745,67 +1028,62 @@ pool_get(struct pool *pp, int flags)
 			if (pp->pr_curpage != NULL)
 				goto startover;
 
-			if ((flags & PR_WAITOK) == 0) {
-				pp->pr_nfail++;
-				pr_leave(pp);
-				simple_unlock(&pp->pr_slock);
-				return (NULL);
-			}
-
-			/*
-			 * Wait for items to be returned to this pool.
-			 *
-			 * 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. */
+			pp->pr_nfail++;
 			pr_leave(pp);
-			ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
-			pr_enter(pp, file, line);
-			goto startover;
+			simple_unlock(&pp->pr_slock);
+			return (NULL);
 		}
 
-		/* We have more memory; add it to the pool */
-		pool_prime_page(pp, v, ph);
-		pp->pr_npagealloc++;
-
 		/* 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 = LIST_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);
+		pr_log(pp, v, PRLOG_GET, file, line);
 #endif
 
 #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.
+		 */
+		LIST_REMOVE(pi, pi_list);
+	}
 	pp->pr_nitems--;
 	pp->pr_nout++;
 	if (ph->ph_nmissing == 0) {
@@ -814,11 +1092,19 @@ pool_get(struct pool *pp, int flags)
 			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 &&
+		    !LIST_EMPTY(&ph->ph_itemlist))) {
 			pr_leave(pp);
 			simple_unlock(&pp->pr_slock);
 			panic("pool_get: %s: nmissing inconsistent",
@@ -826,26 +1112,16 @@ pool_get(struct pool *pp, int flags)
 		}
 #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);
-		TAILQ_FOREACH(ph, &pp->pr_pagelist, 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++;
+	pr_leave(pp);
 
 	/*
 	 * If we have a low water mark and we are now below that low
@@ -859,8 +1135,9 @@ pool_get(struct pool *pp, int flags)
 		 */
 	}
 
-	pr_leave(pp);
 	simple_unlock(&pp->pr_slock);
+	KASSERT((((vaddr_t)v + pp->pr_itemoffset) & (pp->pr_align - 1)) == 0);
+	FREECHECK_OUT(&pp->pr_freecheck, v);
 	return (v);
 }
 
@@ -868,16 +1145,13 @@ pool_get(struct pool *pp, int flags)
  * Internal version of pool_put().  Pool is already locked/entered.
  */
 static void
-pool_do_put(struct pool *pp, void *v)
+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;
 
 	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
-
-	page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
+	FREECHECK_IN(&pp->pr_freecheck, v);
 
 #ifdef DIAGNOSTIC
 	if (__predict_false(pp->pr_nout == 0)) {
@@ -887,7 +1161,7 @@ pool_do_put(struct pool *pp, void *v)
 	}
 #endif
 
-	if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
+	if (__predict_false((ph = pr_find_pagehead(pp, v)) == NULL)) {
 		pr_printlog(pp, NULL, printf);
 		panic("pool_put: %s: page header missing", pp->pr_wchan);
 	}
@@ -902,20 +1176,25 @@ pool_do_put(struct pool *pp, void *v)
 	/*
 	 * 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);
+		LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
+	}
+	KDASSERT(ph->ph_nmissing != 0);
 	ph->ph_nmissing--;
 	pp->pr_nput++;
 	pp->pr_nitems++;
@@ -934,27 +1213,27 @@ pool_do_put(struct pool *pp, void *v)
 	}
 
 	/*
-	 * 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 ||
-		    (pp->pr_alloc->pa_flags & PA_WANT) != 0) {
-			pr_rmpage(pp, ph, NULL);
+		if (pp->pr_npages > pp->pr_minpages &&
+		    (pp->pr_npages > pp->pr_maxpages ||
+		     pa_starved_p(pp->pr_alloc))) {
+			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
@@ -962,34 +1241,20 @@ pool_do_put(struct pool *pp, void *v)
 			 * be reclaimed by the pagedaemon.  This minimizes
 			 * ping-pong'ing for memory.
 			 */
-			s = splclock();
-			ph->ph_time = mono_time;
-			splx(s);
+			getmicrotime(&ph->ph_time);
+		}
+		pool_update_curpage(pp);
+	}
 
-			/*
-			 * 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?
-			 */
-			TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist)
-				if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
-					break;
-
-			pp->pr_curpage = ph;
-		}
-	}
 	/*
-	 * 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;
 	}
 }
@@ -1001,16 +1266,21 @@ pool_do_put(struct pool *pp, void *v)
 void
 _pool_put(struct pool *pp, void *v, const char *file, long line)
 {
+	struct pool_pagelist pq;
+
+	LIST_INIT(&pq);
 
 	simple_lock(&pp->pr_slock);
 	pr_enter(pp, file, line);
 
 	pr_log(pp, v, PRLOG_PUT, file, line);
 
-	pool_do_put(pp, v);
+	pool_do_put(pp, v, &pq);
 
 	pr_leave(pp);
 	simple_unlock(&pp->pr_slock);
+
+	pr_pagelist_free(pp, &pq);
 }
 #undef pool_put
 #endif /* POOL_DIAGNOSTIC */
@@ -1018,12 +1288,15 @@ _pool_put(struct pool *pp, void *v, cons
 void
 pool_put(struct pool *pp, void *v)
 {
+	struct pool_pagelist pq;
 
-	simple_lock(&pp->pr_slock);
-
-	pool_do_put(pp, v);
+	LIST_INIT(&pq);
 
+	simple_lock(&pp->pr_slock);
+	pool_do_put(pp, v, &pq);
 	simple_unlock(&pp->pr_slock);
+
+	pr_pagelist_free(pp, &pq);
 }
 
 #ifdef POOL_DIAGNOSTIC
@@ -1031,34 +1304,56 @@ pool_put(struct pool *pp, void *v)
 #endif
 
 /*
+ * pool_grow: grow a pool by a page.
+ *
+ * => called with pool locked.
+ * => unlock and relock the pool.
+ * => return with pool locked.
+ */
+
+static int
+pool_grow(struct pool *pp, int flags)
+{
+	struct pool_item_header *ph = NULL;
+	char *cp;
+
+	simple_unlock(&pp->pr_slock);
+	cp = pool_allocator_alloc(pp, flags);
+	if (__predict_true(cp != NULL)) {
+		ph = pool_alloc_item_header(pp, cp, flags);
+	}
+	if (__predict_false(cp == NULL || ph == NULL)) {
+		if (cp != NULL) {
+			pool_allocator_free(pp, cp);
+		}
+		simple_lock(&pp->pr_slock);
+		return ENOMEM;
+	}
+
+	simple_lock(&pp->pr_slock);
+	pool_prime_page(pp, cp, ph);
+	pp->pr_npagealloc++;
+	return 0;
+}
+
+/*
  * Add N items to the pool.
  */
 int
 pool_prime(struct pool *pp, int n)
 {
-	struct pool_item_header *ph;
-	caddr_t cp;
 	int newpages;
+	int error = 0;
 
 	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);
-		simple_lock(&pp->pr_slock);
-
-		if (__predict_false(cp == NULL || ph == NULL)) {
-			if (cp != NULL)
-				pool_allocator_free(pp, cp);
+		error = pool_grow(pp, PR_NOWAIT);
+		if (error) {
 			break;
 		}
-
-		pool_prime_page(pp, cp, ph);
-		pp->pr_npagealloc++;
 		pp->pr_minpages++;
 	}
 
@@ -1066,7 +1361,7 @@ pool_prime(struct pool *pp, int n)
 		pp->pr_maxpages = pp->pr_minpages + 1;	/* XXX */
 
 	simple_unlock(&pp->pr_slock);
-	return (0);
+	return error;
 }
 
 /*
@@ -1079,27 +1374,28 @@ pool_prime_page(struct pool *pp, caddr_t
 {
 	struct pool_item *pi;
 	caddr_t cp = storage;
-	unsigned int align = pp->pr_align;
-	unsigned int ioff = pp->pr_itemoffset;
+	const unsigned int align = pp->pr_align;
+	const unsigned int ioff = pp->pr_itemoffset;
 	int n;
 
+	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
+
 #ifdef DIAGNOSTIC
-	if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
+	if ((pp->pr_roflags & PR_NOALIGN) == 0 &&
+	    ((uintptr_t)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)
-		LIST_INSERT_HEAD(&pp->pr_hashtab[PR_HASH_INDEX(pp, cp)],
-		    ph, ph_hashlist);
-
 	/*
 	 * Insert page header.
 	 */
-	TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
-	TAILQ_INIT(&ph->ph_itemlist);
+	LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
+	LIST_INIT(&ph->ph_itemlist);
 	ph->ph_page = storage;
 	ph->ph_nmissing = 0;
-	memset(&ph->ph_time, 0, sizeof(ph->ph_time));
+	getmicrotime(&ph->ph_time);
+	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
+		SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
 
 	pp->pr_nidle++;
 
@@ -1116,23 +1412,38 @@ pool_prime_page(struct pool *pp, caddr_t
 	if (ioff != 0)
 		cp = (caddr_t)(cp + (align - ioff));
 
+	KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
+
 	/*
 	 * Insert remaining chunks on the bucket list.
 	 */
 	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;
 
-		KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
+			KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
 
-		/* Insert on page list */
-		TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
+			/* Insert on page list */
+			LIST_INSERT_HEAD(&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);
+
+			KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
+		}
 	}
 
 	/*
@@ -1147,7 +1458,7 @@ pool_prime_page(struct pool *pp, caddr_t
 
 /*
  * Used by pool_get() when nitems drops below the low water mark.  This
- * is used to catch up nitmes with the low water mark.
+ * 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.
  *
@@ -1157,33 +1468,25 @@ pool_prime_page(struct pool *pp, caddr_t
 static int
 pool_catchup(struct pool *pp)
 {
-	struct pool_item_header *ph;
-	caddr_t cp;
 	int error = 0;
 
 	while (POOL_NEEDS_CATCHUP(pp)) {
-		/*
-		 * Call the page back-end allocator for more memory.
-		 *
-		 * XXX: We never wait, so should we bother unlocking
-		 * the pool descriptor?
-		 */
-		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 (__predict_false(cp == NULL || ph == NULL)) {
-			if (cp != NULL)
-				pool_allocator_free(pp, cp);
-			error = ENOMEM;
+		error = pool_grow(pp, PR_NOWAIT);
+		if (error) {
 			break;
 		}
-		pool_prime_page(pp, cp, ph);
-		pp->pr_npagealloc++;
 	}
+	return error;
+}
+
+static void
+pool_update_curpage(struct pool *pp)
+{
 
-	return (error);
+	pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
+	if (pp->pr_curpage == NULL) {
+		pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
+	}
 }
 
 void
@@ -1257,9 +1560,9 @@ pool_reclaim(struct pool *pp)
 {
 	struct pool_item_header *ph, *phnext;
 	struct pool_cache *pc;
-	struct timeval curtime;
 	struct pool_pagelist pq;
-	int s;
+	struct pool_cache_grouplist pcgl;
+	struct timeval curtime, diff;
 
 	if (pp->pr_drain_hook != NULL) {
 		/*
@@ -1272,60 +1575,48 @@ pool_reclaim(struct pool *pp)
 		return (0);
 	pr_enter(pp, file, line);
 
-	TAILQ_INIT(&pq);
+	LIST_INIT(&pq);
+	LIST_INIT(&pcgl);
 
 	/*
 	 * Reclaim items from the pool's caches.
 	 */
-	TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
-		pool_cache_reclaim(pc);
+	LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
+		pool_cache_reclaim(pc, &pq, &pcgl);
 
-	s = splclock();
-	curtime = mono_time;
-	splx(s);
+	getmicrotime(&curtime);
 
-	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
+		    && !pa_starved_p(pp->pr_alloc))
+			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, &pq);
-		}
+		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);
-	}
+	if (LIST_EMPTY(&pq) && LIST_EMPTY(&pcgl))
+		return 0;
 
+	pr_pagelist_free(pp, &pq);
+	pcg_grouplist_free(&pcgl);
 	return (1);
 }
 
@@ -1344,14 +1635,15 @@ pool_drain(void *arg)
 	s = splvm();
 	simple_lock(&pool_head_slock);
 	if (drainpp == NULL) {
-		drainpp = TAILQ_FIRST(&pool_head);
+		drainpp = LIST_FIRST(&pool_head);
 	}
 	if (drainpp) {
 		pp = drainpp;
-		drainpp = TAILQ_NEXT(pp, pr_poollist);
+		drainpp = LIST_NEXT(pp, pr_poollist);
 	}
 	simple_unlock(&pool_head_slock);
-	pool_reclaim(pp);
+	if (pp)
+		pool_reclaim(pp);
 	splx(s);
 }
 
@@ -1376,9 +1668,24 @@ pool_print(struct pool *pp, const char *
 }
 
 void
+pool_printall(const char *modif, void (*pr)(const char *, ...))
+{
+	struct pool *pp;
+
+	if (simple_lock_try(&pool_head_slock) == 0) {
+		(*pr)("WARNING: pool_head_slock is locked\n");
+	} else {
+		simple_unlock(&pool_head_slock);
+	}
+
+	LIST_FOREACH(pp, &pool_head, pr_poollist) {
+		pool_printit(pp, modif, pr);
+	}
+}
+
+void
 pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
 {
-	int didlock = 0;
 
 	if (pp == NULL) {
 		(*pr)("Must specify a pool to print.\n");
@@ -1397,12 +1704,36 @@ pool_printit(struct pool *pp, const char
 	if (simple_lock_try(&pp->pr_slock) == 0)
 		(*pr)("WARNING: pool %s is locked\n", pp->pr_wchan);
 	else
-		didlock = 1;
+		simple_unlock(&pp->pr_slock);
 
 	pool_print1(pp, modif, pr);
+}
 
-	if (didlock)
-		simple_unlock(&pp->pr_slock);
+static void
+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
+
+	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)) {
+			LIST_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
@@ -1411,9 +1742,6 @@ pool_print1(struct pool *pp, const char 
 	struct pool_item_header *ph;
 	struct pool_cache *pc;
 	struct pool_cache_group *pcg;
-#ifdef DIAGNOSTIC
-	struct pool_item *pi;
-#endif
 	int i, print_log = 0, print_pagelist = 0, print_cache = 0;
 	char c;
 
@@ -1424,7 +1752,6 @@ pool_print1(struct pool *pp, const char 
 			print_pagelist = 1;
 		if (c == 'c')
 			print_cache = 1;
-		modif++;
 	}
 
 	(*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
@@ -1444,74 +1771,83 @@ 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
-		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
-	}
+	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;
 
 	(*pr)("\n");
 	if ((pp->pr_roflags & PR_LOGGING) == 0)
 		(*pr)("\tno log\n");
-	else
+	else {
 		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);
+#define PR_GROUPLIST(pcg)						\
+	(*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);		\
+		}							\
+	}
+
+	LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist) {
+		(*pr)("\tcache %p\n", pc);
 		(*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]);
+		(*pr)("\t    full groups:\n");
+		LIST_FOREACH(pcg, &pc->pc_fullgroups, pcg_list) {
+			PR_GROUPLIST(pcg);
+		}
+		(*pr)("\t    partial groups:\n");
+		LIST_FOREACH(pcg, &pc->pc_partgroups, pcg_list) {
+			PR_GROUPLIST(pcg);
+		}
+		(*pr)("\t    empty groups:\n");
+		LIST_FOREACH(pcg, &pc->pc_emptygroups, pcg_list) {
+			PR_GROUPLIST(pcg);
 		}
 	}
+#undef PR_GROUPLIST
 
  skip_cache:
-
 	pr_enter_check(pp, pr);
 }
 
-int
-pool_chk(struct pool *pp, const char *label)
+static int
+pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
 {
-	struct pool_item_header *ph;
-	int r = 0;
-
-	simple_lock(&pp->pr_slock);
-
-	TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) {
-		struct pool_item *pi;
-		int n;
-		caddr_t page;
+	struct pool_item *pi;
+	caddr_t page;
+	int n;
 
-		page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask);
+	if ((pp->pr_roflags & PR_NOALIGN) == 0) {
+		page = (caddr_t)((uintptr_t)ph & pp->pr_alloc->pa_pagemask);
 		if (page != ph->ph_page &&
 		    (pp->pr_roflags & PR_PHINPAGE) != 0) {
 			if (label != NULL)
@@ -1520,41 +1856,73 @@ pool_chk(struct pool *pp, const char *la
 			       " at page head addr %p (p %p)\n", pp,
 				pp->pr_wchan, ph->ph_page,
 				ph, page);
-			r++;
-			goto out;
+			return 1;
 		}
+	}
 
-		for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
-		     pi != NULL;
-		     pi = TAILQ_NEXT(pi,pi_list), n++) {
+	if ((pp->pr_roflags & PR_NOTOUCH) != 0)
+		return 0;
 
-#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;
+	for (pi = LIST_FIRST(&ph->ph_itemlist), n = 0;
+	     pi != NULL;
+	     pi = LIST_NEXT(pi,pi_list), n++) {
 
+#ifdef DIAGNOSTIC
+		if (pi->pi_magic != PI_MAGIC) {
 			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++;
+			printf("pool(%s): free list modified: magic=%x;"
+			       " page %p; item ordinal %d; addr %p\n",
+				pp->pr_wchan, pi->pi_magic, ph->ph_page,
+				n, pi);
+			panic("pool");
+		}
+#endif
+		if ((pp->pr_roflags & PR_NOALIGN) != 0) {
+			continue;
+		}
+		page = (caddr_t)((uintptr_t)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)
+{
+	struct pool_item_header *ph;
+	int r = 0;
+
+	simple_lock(&pp->pr_slock);
+	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);
@@ -1575,11 +1943,11 @@ pool_cache_init(struct pool_cache *pc, s
     void *arg)
 {
 
-	TAILQ_INIT(&pc->pc_grouplist);
+	LIST_INIT(&pc->pc_emptygroups);
+	LIST_INIT(&pc->pc_fullgroups);
+	LIST_INIT(&pc->pc_partgroups);
 	simple_lock_init(&pc->pc_slock);
 
-	pc->pc_allocfrom = NULL;
-	pc->pc_freeto = NULL;
 	pc->pc_pool = pp;
 
 	pc->pc_ctor = ctor;
@@ -1594,7 +1962,7 @@ pool_cache_init(struct pool_cache *pc, s
 	pc->pc_nitems = 0;
 
 	simple_lock(&pp->pr_slock);
-	TAILQ_INSERT_TAIL(&pp->pr_cachelist, pc, pc_poollist);
+	LIST_INSERT_HEAD(&pp->pr_cachelist, pc, pc_poollist);
 	simple_unlock(&pp->pr_slock);
 }
 
@@ -1613,12 +1981,12 @@ pool_cache_destroy(struct pool_cache *pc
 
 	/* ...and remove it from the pool's cache list. */
 	simple_lock(&pp->pr_slock);
-	TAILQ_REMOVE(&pp->pr_cachelist, pc, pc_poollist);
+	LIST_REMOVE(pc, pc_poollist);
 	simple_unlock(&pp->pr_slock);
 }
 
-static __inline void *
-pcg_get(struct pool_cache_group *pcg)
+static inline void *
+pcg_get(struct pool_cache_group *pcg, paddr_t *pap)
 {
 	void *object;
 	u_int idx;
@@ -1627,50 +1995,70 @@ pcg_get(struct pool_cache_group *pcg)
 	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;
+	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)
+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] == NULL);
-	pcg->pcg_objects[idx] = object;
+	KASSERT(pcg->pcg_objects[idx].pcgo_va == NULL);
+	pcg->pcg_objects[idx].pcgo_va = object;
+	pcg->pcg_objects[idx].pcgo_pa = pa;
+}
+
+static void
+pcg_grouplist_free(struct pool_cache_grouplist *pcgl)
+{
+	struct pool_cache_group *pcg;
+	int s;
+
+	s = splvm();
+	while ((pcg = LIST_FIRST(pcgl)) != NULL) {
+		LIST_REMOVE(pcg, pcg_list);
+		pool_put(&pcgpool, pcg);
+	}
+	splx(s);
 }
 
 /*
- * pool_cache_get:
+ * pool_cache_get{,_paddr}:
  *
- *	Get an object from a pool cache.
+ *	Get an object from a pool cache (optionally returning
+ *	the physical address of the object).
  */
 void *
-pool_cache_get(struct pool_cache *pc, int flags)
+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)");
+		ASSERT_SLEEPABLE(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;
-			}
+	pcg = LIST_FIRST(&pc->pc_partgroups);
+	if (pcg == NULL) {
+		pcg = LIST_FIRST(&pc->pc_fullgroups);
+		if (pcg != NULL) {
+			LIST_REMOVE(pcg, pcg_list);
+			LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
 		}
+	}
+	if (pcg == NULL) {
 
 		/*
 		 * No groups with any available objects.  Allocate
@@ -1687,42 +2075,65 @@ pool_cache_get(struct pool_cache *pc, in
 				return (NULL);
 			}
 		}
+		KASSERT((((vaddr_t)object + pc->pc_pool->pr_itemoffset) &
+		    (pc->pc_pool->pr_align - 1)) == 0);
+		if (object != NULL && pap != NULL) {
+#ifdef POOL_VTOPHYS
+			*pap = POOL_VTOPHYS(object);
+#else
+			*pap = POOL_PADDR_INVALID;
+#endif
+		}
+
+		FREECHECK_OUT(&pc->pc_freecheck, object);
 		return (object);
 	}
 
- have_group:
 	pc->pc_hits++;
 	pc->pc_nitems--;
-	object = pcg_get(pcg);
-
-	if (pcg->pcg_avail == 0)
-		pc->pc_allocfrom = NULL;
+	object = pcg_get(pcg, pap);
 
+	if (pcg->pcg_avail == 0) {
+		LIST_REMOVE(pcg, pcg_list);
+		LIST_INSERT_HEAD(&pc->pc_emptygroups, pcg, pcg_list);
+	}
 	simple_unlock(&pc->pc_slock);
 
+	KASSERT((((vaddr_t)object + pc->pc_pool->pr_itemoffset) &
+	    (pc->pc_pool->pr_align - 1)) == 0);
+	FREECHECK_OUT(&pc->pc_freecheck, object);
 	return (object);
 }
 
 /*
- * pool_cache_put:
+ * pool_cache_put{,_paddr}:
  *
- *	Put an object back to the pool cache.
+ *	Put an object back to the pool cache (optionally caching the
+ *	physical address of the object).
  */
 void
-pool_cache_put(struct pool_cache *pc, void *object)
+pool_cache_put_paddr(struct pool_cache *pc, void *object, paddr_t pa)
 {
 	struct pool_cache_group *pcg;
 	int s;
 
+	FREECHECK_IN(&pc->pc_freecheck, object);
+
+	if (__predict_false((pc->pc_pool->pr_flags & PR_WANTED) != 0)) {
+		goto destruct;
+	}
+
 	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;
-			}
+	pcg = LIST_FIRST(&pc->pc_partgroups);
+	if (pcg == NULL) {
+		pcg = LIST_FIRST(&pc->pc_emptygroups);
+		if (pcg != NULL) {
+			LIST_REMOVE(pcg, pcg_list);
+			LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
 		}
+	}
+	if (pcg == NULL) {
 
 		/*
 		 * No empty groups to free the object to.  Attempt to
@@ -1732,31 +2143,29 @@ pool_cache_put(struct pool_cache *pc, vo
 		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;
-		}
+		if (pcg == NULL) {
+destruct:
 
-		/*
-		 * Unable to allocate a cache group; destruct the object
-		 * and free it back to the pool.
-		 */
-		pool_cache_destruct_object(pc, object);
-		return;
+			/*
+			 * Unable to allocate a cache group; destruct the object
+			 * and free it back to the pool.
+			 */
+			pool_cache_destruct_object(pc, object);
+			return;
+		}
+		memset(pcg, 0, sizeof(*pcg));
+		simple_lock(&pc->pc_slock);
+		pc->pc_ngroups++;
+		LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
 	}
 
- have_group:
 	pc->pc_nitems++;
-	pcg_put(pcg, object);
-
-	if (pcg->pcg_avail == PCG_NOBJECTS)
-		pc->pc_freeto = NULL;
+	pcg_put(pcg, object, pa);
 
+	if (pcg->pcg_avail == PCG_NOBJECTS) {
+		LIST_REMOVE(pcg, pcg_list);
+		LIST_INSERT_HEAD(&pc->pc_fullgroups, pcg, pcg_list);
+	}
 	simple_unlock(&pc->pc_slock);
 }
 
@@ -1775,44 +2184,45 @@ pool_cache_destruct_object(struct pool_c
 	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 *))
+pool_do_cache_invalidate_grouplist(struct pool_cache_grouplist *pcgsl,
+    struct pool_cache *pc, struct pool_pagelist *pq,
+    struct pool_cache_grouplist *pcgdl)
 {
 	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);
+	for (pcg = LIST_FIRST(pcgsl); pcg != NULL; pcg = npcg) {
+		npcg = LIST_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;
+			object = pcg_get(pcg, 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_do_put(pc->pc_pool, object, pq);
 		}
+		pc->pc_ngroups--;
+		LIST_REMOVE(pcg, pcg_list);
+		LIST_INSERT_HEAD(pcgdl, pcg, pcg_list);
 	}
 }
 
+static void
+pool_do_cache_invalidate(struct pool_cache *pc, struct pool_pagelist *pq,
+    struct pool_cache_grouplist *pcgl)
+{
+
+	LOCK_ASSERT(simple_lock_held(&pc->pc_slock));
+	LOCK_ASSERT(simple_lock_held(&pc->pc_pool->pr_slock));
+
+	pool_do_cache_invalidate_grouplist(&pc->pc_fullgroups, pc, pq, pcgl);
+	pool_do_cache_invalidate_grouplist(&pc->pc_partgroups, pc, pq, pcgl);
+
+	KASSERT(LIST_EMPTY(&pc->pc_partgroups));
+	KASSERT(LIST_EMPTY(&pc->pc_fullgroups));
+	KASSERT(pc->pc_nitems == 0);
+}
+
 /*
  * pool_cache_invalidate:
  *
@@ -1822,10 +2232,22 @@ pool_cache_do_invalidate(struct pool_cac
 void
 pool_cache_invalidate(struct pool_cache *pc)
 {
+	struct pool_pagelist pq;
+	struct pool_cache_grouplist pcgl;
+
+	LIST_INIT(&pq);
+	LIST_INIT(&pcgl);
 
 	simple_lock(&pc->pc_slock);
-	pool_cache_do_invalidate(pc, 0, pool_put);
+	simple_lock(&pc->pc_pool->pr_slock);
+
+	pool_do_cache_invalidate(pc, &pq, &pcgl);
+
+	simple_unlock(&pc->pc_pool->pr_slock);
 	simple_unlock(&pc->pc_slock);
+
+	pr_pagelist_free(pc->pc_pool, &pq);
+	pcg_grouplist_free(&pcgl);
 }
 
 /*
@@ -1834,11 +2256,21 @@ pool_cache_invalidate(struct pool_cache 
  *	Reclaim a pool cache for pool_reclaim().
  */
 static void
-pool_cache_reclaim(struct pool_cache *pc)
+pool_cache_reclaim(struct pool_cache *pc, struct pool_pagelist *pq,
+    struct pool_cache_grouplist *pcgl)
 {
 
-	simple_lock(&pc->pc_slock);
-	pool_cache_do_invalidate(pc, 1, pool_do_put);
+	/*
+	 * 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;
+
+	pool_do_cache_invalidate(pc, pq, pcgl);
+
 	simple_unlock(&pc->pc_slock);
 }
 
@@ -1858,140 +2290,82 @@ pool_cache_reclaim(struct pool_cache *pc
 void	*pool_page_alloc(struct pool *, int);
 void	pool_page_free(struct pool *, void *);
 
+#ifdef POOL_SUBPAGE
+struct pool_allocator pool_allocator_kmem_fullpage = {
+	pool_page_alloc, pool_page_free, 0,
+	.pa_backingmapptr = &kmem_map,
+};
+#else
 struct pool_allocator pool_allocator_kmem = {
 	pool_page_alloc, pool_page_free, 0,
+	.pa_backingmapptr = &kmem_map,
 };
+#endif
 
 void	*pool_page_alloc_nointr(struct pool *, int);
 void	pool_page_free_nointr(struct pool *, void *);
 
+#ifdef POOL_SUBPAGE
+struct pool_allocator pool_allocator_nointr_fullpage = {
+	pool_page_alloc_nointr, pool_page_free_nointr, 0,
+	.pa_backingmapptr = &kernel_map,
+};
+#else
 struct pool_allocator pool_allocator_nointr = {
 	pool_page_alloc_nointr, pool_page_free_nointr, 0,
+	.pa_backingmapptr = &kernel_map,
 };
+#endif
 
 #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,
+struct pool_allocator pool_allocator_kmem = {
+	pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
+	.pa_backingmapptr = &kmem_map,
+};
+
+void	*pool_subpage_alloc_nointr(struct pool *, int);
+void	pool_subpage_free_nointr(struct pool *, void *);
+
+struct pool_allocator pool_allocator_nointr = {
+	pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
+	.pa_backingmapptr = &kmem_map,
 };
 #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)
+static void *
+pool_allocator_alloc(struct pool *pp, int flags)
 {
-	struct pool_allocator *pa = org->pr_alloc;
-	struct pool *pp, *start;
-	int s, freed;
+	struct pool_allocator *pa = pp->pr_alloc;
 	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;
-		}
+	LOCK_ASSERT(!simple_lock_held(&pp->pr_slock));
 
+	res = (*pa->pa_alloc)(pp, flags);
+	if (res == NULL && (flags & PR_WAITOK) == 0) {
 		/*
-		 * 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?
+		 * We only run the drain hook here if PR_NOWAIT.
+		 * In other cases, the hook will be run in
+		 * pool_reclaim().
 		 */
-		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;
+		if (pp->pr_drain_hook != NULL) {
+			(*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
+			res = (*pa->pa_alloc)(pp, flags);
 		}
-		simple_unlock(&pa->pa_slock);
-		splx(s);
-	} while (freed);
-	return (NULL);
+	}
+	return res;
 }
 
-void
+static 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;
-	}
+	LOCK_ASSERT(!simple_lock_held(&pp->pr_slock));
 
-	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);
+	(*pa->pa_free)(pp, v);
 }
 
 void *
@@ -1999,14 +2373,29 @@ pool_page_alloc(struct pool *pp, int fla
 {
 	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
 
-	return ((void *) uvm_km_alloc_poolpage(waitok));
+	return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok));
 }
 
 void
 pool_page_free(struct pool *pp, void *v)
 {
 
-	uvm_km_free_poolpage((vaddr_t) 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
@@ -2014,45 +2403,49 @@ pool_page_free(struct pool *pp, void *v)
 void *
 pool_subpage_alloc(struct pool *pp, int flags)
 {
-
-	return (pool_get(&psppool, 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)
+pool_subpage_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_nointr(struct pool *pp, void *v)
 {
 
 	pool_subpage_free(pp, v);
 }
-#else
+#endif /* POOL_SUBPAGE */
 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));
+	return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok));
 }
 
 void
 pool_page_free_nointr(struct pool *pp, void *v)
 {
 
-	uvm_km_free_poolpage1(kernel_map, (vaddr_t) v);
+	uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v);
 }
-#endif /* POOL_SUBPAGE */