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 1.110.2.2
retrieving revision 1.129.12.5
diff -u -p -r1.110.2.2 -r1.129.12.5
--- src/sys/kern/subr_pool.c	2006/03/01 09:28:46	1.110.2.2
+++ src/sys/kern/subr_pool.c	2007/11/21 21:56:03	1.129.12.5
@@ -1,12 +1,12 @@
-/*	$NetBSD: subr_pool.c,v 1.110.2.2 2006/03/01 09:28:46 yamt Exp $	*/
+/*	$NetBSD: subr_pool.c,v 1.129.12.5 2007/11/21 21:56:03 joerg Exp $	*/
 
 /*-
- * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc.
+ * Copyright (c) 1997, 1999, 2000, 2002, 2007 The NetBSD Foundation, Inc.
  * All rights reserved.
  *
  * This code is derived from software contributed to The NetBSD Foundation
  * by Paul Kranenburg; by Jason R. Thorpe of the Numerical Aerospace
- * Simulation Facility, NASA Ames Research Center.
+ * Simulation Facility, NASA Ames Research Center, and by Andrew Doran.
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions
@@ -38,7 +38,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.110.2.2 2006/03/01 09:28:46 yamt Exp $");
+__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.129.12.5 2007/11/21 21:56:03 joerg Exp $");
 
 #include "opt_pool.h"
 #include "opt_poollog.h"
@@ -46,6 +46,7 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,
 
 #include <sys/param.h>
 #include <sys/systm.h>
+#include <sys/bitops.h>
 #include <sys/proc.h>
 #include <sys/errno.h>
 #include <sys/kernel.h>
@@ -53,6 +54,10 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,
 #include <sys/lock.h>
 #include <sys/pool.h>
 #include <sys/syslog.h>
+#include <sys/debug.h>
+#include <sys/lockdebug.h>
+#include <sys/xcall.h>
+#include <sys/cpu.h>
 
 #include <uvm/uvm.h>
 
@@ -72,22 +77,31 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,
 /* List of all pools */
 LIST_HEAD(,pool) pool_head = LIST_HEAD_INITIALIZER(pool_head);
 
+/* List of all caches. */
+LIST_HEAD(,pool_cache) pool_cache_head =
+    LIST_HEAD_INITIALIZER(pool_cache_head);
+
 /* Private pool for page header structures */
 #define	PHPOOL_MAX	8
 static struct pool phpool[PHPOOL_MAX];
-#define	PHPOOL_FREELIST_NELEM(idx)	(((idx) == 0) ? 0 : (1 << (idx)))
+#define	PHPOOL_FREELIST_NELEM(idx) \
+	(((idx) == 0) ? 0 : BITMAP_SIZE * (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
+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 */
@@ -96,10 +110,13 @@ int pool_inactive_time = 10;
 /* Next candidate for drainage (see pool_drain()) */
 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;
+/* This lock protects both pool_head and drainpp. */
+static kmutex_t pool_head_lock;
+static kcondvar_t pool_busy;
+
+typedef uint32_t pool_item_bitmap_t;
+#define	BITMAP_SIZE	(CHAR_BIT * sizeof(pool_item_bitmap_t))
+#define	BITMAP_MASK	(BITMAP_SIZE - 1)
 
 struct pool_item_header {
 	/* Page headers */
@@ -107,8 +124,9 @@ struct pool_item_header {
 				ph_pagelist;	/* pool page list */
 	SPLAY_ENTRY(pool_item_header)
 				ph_node;	/* Off-page page headers */
-	caddr_t			ph_page;	/* this page's address */
+	void *			ph_page;	/* this page's address */
 	struct timeval		ph_time;	/* last referenced */
+	uint16_t		ph_nmissing;	/* # of chunks in use */
 	union {
 		/* !PR_NOTOUCH */
 		struct {
@@ -117,27 +135,20 @@ struct pool_item_header {
 		} 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)
-			 */
+			uint16_t phu_off;	/* start offset in page */
+			pool_item_bitmap_t phu_bitmap[];
 		} 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
+#define	ph_bitmap	ph_u.phu_notouch.phu_bitmap
 
 struct pool_item {
 #ifdef DIAGNOSTIC
 	u_int pi_magic;
 #endif
-#define	PI_MAGIC 0xdeadbeefU
+#define	PI_MAGIC 0xdeaddeadU
 	/* Other entries use only this list entry */
 	LIST_ENTRY(pool_item)	pi_list;
 };
@@ -153,38 +164,43 @@ struct pool_item {
  * 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.
+ * Caches are grouped into cache groups.  Each cache group references up
+ * to PCG_NUMOBJECTS 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.
+ *
+ * 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 struct pool cache_pool;
+static struct pool cache_cpu_pool;
 
-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 pool_cache_cpu_t *pool_cache_put_slow(pool_cache_cpu_t *, int *,
+					     void *, paddr_t);
+static pool_cache_cpu_t *pool_cache_get_slow(pool_cache_cpu_t *, int *,
+					     void **, paddr_t *, int);
+static void	pool_cache_cpu_init1(struct cpu_info *, pool_cache_t);
+static void	pool_cache_invalidate_groups(pool_cache_t, pcg_t *);
+static void	pool_cache_xcall(pool_cache_t);
 
 static int	pool_catchup(struct pool *);
-static void	pool_prime_page(struct pool *, caddr_t,
+static void	pool_prime_page(struct pool *, void *,
 		    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 *, ...));
@@ -312,57 +328,83 @@ pr_enter_check(struct pool *pp, void (*p
 #define	pr_enter_check(pp, pr)
 #endif /* POOL_DIAGNOSTIC */
 
-static inline int
+static inline unsigned int
 pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
     const void *v)
 {
 	const char *cp = v;
-	int idx;
+	unsigned int idx;
 
 	KASSERT(pp->pr_roflags & PR_NOTOUCH);
-	idx = (cp - ph->ph_page - ph->ph_off) / pp->pr_size;
+	idx = (cp - (char *)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);
+	unsigned int idx = pr_item_notouch_index(pp, ph, obj);
+	pool_item_bitmap_t *bitmap = ph->ph_bitmap + (idx / BITMAP_SIZE);
+	pool_item_bitmap_t mask = 1 << (idx & BITMAP_MASK);
 
-	KASSERT(freelist[idx] == PR_INDEX_USED);
-	freelist[idx] = ph->ph_firstfree;
-	ph->ph_firstfree = idx;
+	KASSERT((*bitmap & mask) == 0);
+	*bitmap |= mask;
 }
 
 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);
+	pool_item_bitmap_t *bitmap = ph->ph_bitmap;
+	unsigned int idx;
+	int i;
+
+	for (i = 0; ; i++) {
+		int bit;
+
+		KASSERT((i * BITMAP_SIZE) < pp->pr_itemsperpage);
+		bit = ffs32(bitmap[i]);
+		if (bit) {
+			pool_item_bitmap_t mask;
+
+			bit--;
+			idx = (i * BITMAP_SIZE) + bit;
+			mask = 1 << bit;
+			KASSERT((bitmap[i] & mask) != 0);
+			bitmap[i] &= ~mask;
+			break;
+		}
+	}
+	KASSERT(idx < pp->pr_itemsperpage);
+	return (char *)ph->ph_page + ph->ph_off + idx * pp->pr_size;
+}
 
-	KASSERT(freelist[idx] != PR_INDEX_USED);
-	ph->ph_firstfree = freelist[idx];
-	freelist[idx] = PR_INDEX_USED;
+static inline void
+pr_item_notouch_init(const struct pool *pp, struct pool_item_header *ph)
+{
+	pool_item_bitmap_t *bitmap = ph->ph_bitmap;
+	const int n = howmany(pp->pr_itemsperpage, BITMAP_SIZE);
+	int i;
 
-	return ph->ph_page + ph->ph_off + idx * pp->pr_size;
+	for (i = 0; i < n; i++) {
+		bitmap[i] = (pool_item_bitmap_t)-1;
+	}
 }
 
 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 if (a->ph_page > b->ph_page)
+		return (-1);
 	else
 		return (0);
 }
@@ -371,18 +413,38 @@ SPLAY_PROTOTYPE(phtree, pool_item_header
 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)
+pr_find_pagehead(struct pool *pp, void *v)
 {
 	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 = (void *)(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 {
+		void *page =
+		    (void *)((uintptr_t)v & pp->pr_alloc->pa_pagemask);
+
+		if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
+			ph = (struct pool_item_header *)((char *)page + pp->pr_phoffset);
+		} else {
+			tmp.ph_page = page;
+			ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
+		}
+	}
 
-	tmp.ph_page = page;
-	ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
+	KASSERT(ph == NULL || ((pp->pr_roflags & PR_PHINPAGE) != 0) ||
+	    ((char *)ph->ph_page <= (char *)v &&
+	    (char *)v < (char *)ph->ph_page + pp->pr_alloc->pa_pagesz));
 	return ph;
 }
 
@@ -390,16 +452,12 @@ 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();
+		if ((pp->pr_roflags & PR_PHINPAGE) == 0)
 			pool_put(pp->pr_phpool, ph);
-			splx(s);
-		}
 	}
 }
 
@@ -411,7 +469,7 @@ pr_rmpage(struct pool *pp, struct pool_i
      struct pool_pagelist *pq)
 {
 
-	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
+	KASSERT(mutex_owned(&pp->pr_lock));
 
 	/*
 	 * If the page was idle, decrement the idle page count.
@@ -442,19 +500,109 @@ pr_rmpage(struct pool *pp, struct pool_i
 	pool_update_curpage(pp);
 }
 
+static bool
+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
-link_pool_init(void)
+pool_subsystem_init(void)
 {
+	struct pool_allocator *pa;
 	__link_set_decl(pools, struct link_pool_init);
 	struct link_pool_init * const *pi;
 
+	mutex_init(&pool_head_lock, MUTEX_DEFAULT, IPL_NONE);
+	cv_init(&pool_busy, "poolbusy");
+
 	__link_set_foreach(pi, pools)
 		pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
 		    (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
-		    (*pi)->palloc);
+		    (*pi)->palloc, (*pi)->ipl);
+
+	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);
+	}
+
+	pool_init(&cache_pool, sizeof(struct pool_cache), CACHE_LINE_SIZE,
+	    0, 0, "pcache", &pool_allocator_nointr, IPL_NONE);
+
+	pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), CACHE_LINE_SIZE,
+	    0, 0, "pcachecpu", &pool_allocator_nointr, IPL_NONE);
 }
 
 /*
@@ -465,14 +613,25 @@ link_pool_init(void)
  */
 void
 pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
-    const char *wchan, struct pool_allocator *palloc)
+    const char *wchan, struct pool_allocator *palloc, int ipl)
 {
-	int off, slack;
+#ifdef DEBUG
+	struct pool *pp1;
+#endif
 	size_t trysize, phsize;
-	int s;
+	int off, slack;
 
-	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
 	/*
@@ -498,23 +657,26 @@ pool_init(struct pool *pp, size_t size, 
 
 		TAILQ_INIT(&palloc->pa_list);
 
-		simple_lock_init(&palloc->pa_slock);
+		mutex_init(&palloc->pa_lock, MUTEX_DEFAULT, IPL_VM);
 		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
 
 	/*
@@ -523,7 +685,7 @@ pool_init(struct pool *pp, size_t size, 
 	LIST_INIT(&pp->pr_emptypages);
 	LIST_INIT(&pp->pr_fullpages);
 	LIST_INIT(&pp->pr_partpages);
-	LIST_INIT(&pp->pr_cachelist);
+	pp->pr_cache = NULL;
 	pp->pr_curpage = NULL;
 	pp->pr_npages = 0;
 	pp->pr_minitems = 0;
@@ -545,6 +707,7 @@ 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
@@ -563,7 +726,7 @@ pool_init(struct pool *pp, size_t size, 
 	/* 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 &&
+	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 */
@@ -623,6 +786,7 @@ pool_init(struct pool *pp, size_t size, 
 	pp->pr_npagefree = 0;
 	pp->pr_hiwat = 0;
 	pp->pr_nidle = 0;
+	pp->pr_refcnt = 0;
 
 #ifdef POOL_DIAGNOSTIC
 	if (flags & PR_LOGGING) {
@@ -638,7 +802,9 @@ pool_init(struct pool *pp, size_t size, 
 	pp->pr_entered_file = NULL;
 	pp->pr_entered_line = 0;
 
-	simple_lock_init(&pp->pr_slock);
+	mutex_init(&pp->pr_lock, MUTEX_DEFAULT, ipl);
+	cv_init(&pp->pr_cv, wchan);
+	pp->pr_ipl = ipl;
 
 	/*
 	 * Initialize private page header pool and cache magazine pool if we
@@ -657,31 +823,36 @@ pool_init(struct pool *pp, size_t size, 
 			    "phpool-%d", nelem);
 			sz = sizeof(struct pool_item_header);
 			if (nelem) {
-				sz = PR_FREELIST_ALIGN(sz)
-				    + nelem * sizeof(pool_item_freelist_t);
+				sz = offsetof(struct pool_item_header,
+				    ph_bitmap[howmany(nelem, BITMAP_SIZE)]);
 			}
 			pool_init(&phpool[idx], sz, 0, 0, 0,
-			    phpool_names[idx], &pool_allocator_meta);
+			    phpool_names[idx], &pool_allocator_meta, IPL_VM);
 		}
 #ifdef POOL_SUBPAGE
 		pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
-		    PR_RECURSIVE, "psppool", &pool_allocator_meta);
+		    PR_RECURSIVE, "psppool", &pool_allocator_meta, IPL_VM);
 #endif
-		pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
-		    0, "pcgpool", &pool_allocator_meta);
+		pool_init(&pcgpool, sizeof(pcg_t), CACHE_LINE_SIZE, 0, 0,
+		    "cachegrp", &pool_allocator_meta, IPL_VM);
 	}
 
-	/* Insert into the list of all pools. */
-	simple_lock(&pool_head_slock);
-	LIST_INSERT_HEAD(&pool_head, pp, pr_poollist);
-	simple_unlock(&pool_head_slock);
+	if (__predict_true(!cold)) {
+		/* Insert into the list of all pools. */
+		mutex_enter(&pool_head_lock);
+		LIST_INSERT_HEAD(&pool_head, pp, pr_poollist);
+		mutex_exit(&pool_head_lock);
+
+		/* Insert this into the list of pools using this allocator. */
+		mutex_enter(&palloc->pa_lock);
+		TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
+		mutex_exit(&palloc->pa_lock);
+	} else {
+		LIST_INSERT_HEAD(&pool_head, pp, pr_poollist);
+		TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
+	}
 
-	/* 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);
 }
 
 /*
@@ -692,26 +863,25 @@ 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);
+	mutex_enter(&pool_head_lock);
+	while (pp->pr_refcnt != 0)
+		cv_wait(&pool_busy, &pool_head_lock);
 	LIST_REMOVE(pp, pr_poollist);
 	if (drainpp == pp)
 		drainpp = NULL;
-	simple_unlock(&pool_head_slock);
+	mutex_exit(&pool_head_lock);
 
 	/* Remove this pool from its allocator's list of pools. */
-	s = splvm();
-	simple_lock(&pp->pr_alloc->pa_slock);
+	pool_reclaim_unregister(pp);
+	mutex_enter(&pp->pr_alloc->pa_lock);
 	TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
-	simple_unlock(&pp->pr_alloc->pa_slock);
-	splx(s);
+	mutex_exit(&pp->pr_alloc->pa_lock);
 
-	s = splvm();
-	simple_lock(&pp->pr_slock);
+	mutex_enter(&pp->pr_lock);
 
-	KASSERT(LIST_EMPTY(&pp->pr_cachelist));
+	KASSERT(pp->pr_cache == NULL);
 
 #ifdef DIAGNOSTIC
 	if (pp->pr_nout != 0) {
@@ -729,8 +899,7 @@ pool_destroy(struct pool *pp)
 	while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
 		pr_rmpage(pp, ph, &pq);
 
-	simple_unlock(&pp->pr_slock);
-	splx(s);
+	mutex_exit(&pp->pr_lock);
 
 	pr_pagelist_free(pp, &pq);
 
@@ -738,6 +907,9 @@ pool_destroy(struct pool *pp)
 	if ((pp->pr_roflags & PR_LOGGING) != 0)
 		free(pp->pr_log, M_TEMP);
 #endif
+
+	cv_destroy(&pp->pr_cv);
+	mutex_destroy(&pp->pr_lock);
 }
 
 void
@@ -754,26 +926,20 @@ pool_set_drain_hook(struct pool *pp, voi
 }
 
 static struct pool_item_header *
-pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
+pool_alloc_item_header(struct pool *pp, void *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 = (struct pool_item_header *) ((char *)storage + pp->pr_phoffset);
+	else
 		ph = pool_get(pp->pr_phpool, flags);
-		splx(s);
-	}
 
 	return (ph);
 }
 
 /*
- * Grab an item from the pool; must be called at appropriate spl level
+ * Grab an item from the pool.
  */
 void *
 #ifdef POOL_DIAGNOSTIC
@@ -797,11 +963,10 @@ pool_get(struct pool *pp, int flags)
 #endif /* DIAGNOSTIC */
 #ifdef LOCKDEBUG
 	if (flags & PR_WAITOK)
-		simple_lock_only_held(NULL, "pool_get(PR_WAITOK)");
-	SCHED_ASSERT_UNLOCKED();
+		ASSERT_SLEEPABLE(NULL, "pool_get(PR_WAITOK)");
 #endif
 
-	simple_lock(&pp->pr_slock);
+	mutex_enter(&pp->pr_lock);
 	pr_enter(pp, file, line);
 
  startover:
@@ -813,7 +978,7 @@ pool_get(struct pool *pp, int flags)
 #ifdef DIAGNOSTIC
 	if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
 		pr_leave(pp);
-		simple_unlock(&pp->pr_slock);
+		mutex_exit(&pp->pr_lock);
 		panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
 	}
 #endif
@@ -825,9 +990,9 @@ pool_get(struct pool *pp, int flags)
 			 * and check the hardlimit condition again.
 			 */
 			pr_leave(pp);
-			simple_unlock(&pp->pr_slock);
+			mutex_exit(&pp->pr_lock);
 			(*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
-			simple_lock(&pp->pr_slock);
+			mutex_enter(&pp->pr_lock);
 			pr_enter(pp, file, line);
 			if (pp->pr_nout < pp->pr_hardlimit)
 				goto startover;
@@ -840,7 +1005,7 @@ pool_get(struct pool *pp, int flags)
 			 */
 			pp->pr_flags |= PR_WANTED;
 			pr_leave(pp);
-			ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
+			cv_wait(&pp->pr_cv, &pp->pr_lock);
 			pr_enter(pp, file, line);
 			goto startover;
 		}
@@ -856,7 +1021,7 @@ pool_get(struct pool *pp, int flags)
 		pp->pr_nfail++;
 
 		pr_leave(pp);
-		simple_unlock(&pp->pr_slock);
+		mutex_exit(&pp->pr_lock);
 		return (NULL);
 	}
 
@@ -867,9 +1032,11 @@ 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);
+			mutex_exit(&pp->pr_lock);
 			printf("pool_get: %s: curpage NULL, nitems %u\n",
 			    pp->pr_wchan, pp->pr_nitems);
 			panic("pool_get: nitems inconsistent");
@@ -882,18 +1049,9 @@ 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);
-
-		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);
-
+		error = pool_grow(pp, flags);
+		pr_enter(pp, file, line);
+		if (error != 0) {
 			/*
 			 * We were unable to allocate a page or item
 			 * header, but we released the lock during
@@ -903,33 +1061,12 @@ 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.
-			 *
-			 * wake up once a second and try again,
-			 * as the check in pool_cache_put_paddr() is racy.
-			 */
-			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, hz, &pp->pr_slock);
-			pr_enter(pp, file, line);
-			goto startover;
+			mutex_exit(&pp->pr_lock);
+			return (NULL);
 		}
 
-		/* 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++;
-
 		/* Start the allocation process over. */
 		goto startover;
 	}
@@ -937,7 +1074,7 @@ pool_get(struct pool *pp, int flags)
 #ifdef DIAGNOSTIC
 		if (__predict_false(ph->ph_nmissing == pp->pr_itemsperpage)) {
 			pr_leave(pp);
-			simple_unlock(&pp->pr_slock);
+			mutex_exit(&pp->pr_lock);
 			panic("pool_get: %s: page empty", pp->pr_wchan);
 		}
 #endif
@@ -949,13 +1086,13 @@ pool_get(struct pool *pp, int flags)
 		v = pi = LIST_FIRST(&ph->ph_itemlist);
 		if (__predict_false(v == NULL)) {
 			pr_leave(pp);
-			simple_unlock(&pp->pr_slock);
+			mutex_exit(&pp->pr_lock);
 			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);
+			mutex_exit(&pp->pr_lock);
 			printf("pool_get: %s: items on itemlist, nitems %u\n",
 			    pp->pr_wchan, pp->pr_nitems);
 			panic("pool_get: nitems inconsistent");
@@ -1002,7 +1139,7 @@ pool_get(struct pool *pp, int flags)
 		if (__predict_false((pp->pr_roflags & PR_NOTOUCH) == 0 &&
 		    !LIST_EMPTY(&ph->ph_itemlist))) {
 			pr_leave(pp);
-			simple_unlock(&pp->pr_slock);
+			mutex_exit(&pp->pr_lock);
 			panic("pool_get: %s: nmissing inconsistent",
 			    pp->pr_wchan);
 		}
@@ -1031,7 +1168,9 @@ pool_get(struct pool *pp, int flags)
 		 */
 	}
 
-	simple_unlock(&pp->pr_slock);
+	mutex_exit(&pp->pr_lock);
+	KASSERT((((vaddr_t)v + pp->pr_itemoffset) & (pp->pr_align - 1)) == 0);
+	FREECHECK_OUT(&pp->pr_freecheck, v);
 	return (v);
 }
 
@@ -1043,13 +1182,10 @@ pool_do_put(struct pool *pp, void *v, st
 {
 	struct pool_item *pi = v;
 	struct pool_item_header *ph;
-	caddr_t page;
-	int s;
-
-	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
-	SCHED_ASSERT_UNLOCKED();
 
-	page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
+	KASSERT(mutex_owned(&pp->pr_lock));
+	FREECHECK_IN(&pp->pr_freecheck, v);
+	LOCKDEBUG_MEM_CHECK(v, pp->pr_size);
 
 #ifdef DIAGNOSTIC
 	if (__predict_false(pp->pr_nout == 0)) {
@@ -1059,18 +1195,11 @@ pool_do_put(struct pool *pp, void *v, st
 	}
 #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);
 	}
 
-#ifdef LOCKDEBUG
-	/*
-	 * Check if we're freeing a locked simple lock.
-	 */
-	simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size);
-#endif
-
 	/*
 	 * Return to item list.
 	 */
@@ -1106,7 +1235,7 @@ pool_do_put(struct pool *pp, void *v, st
 		pp->pr_flags &= ~PR_WANTED;
 		if (ph->ph_nmissing == 0)
 			pp->pr_nidle++;
-		wakeup((caddr_t)pp);
+		cv_broadcast(&pp->pr_cv);
 		return;
 	}
 
@@ -1127,7 +1256,7 @@ pool_do_put(struct pool *pp, void *v, st
 		pp->pr_nidle++;
 		if (pp->pr_npages > pp->pr_minpages &&
 		    (pp->pr_npages > pp->pr_maxpages ||
-		     (pp->pr_alloc->pa_flags & PA_WANT) != 0)) {
+		     pa_starved_p(pp->pr_alloc))) {
 			pr_rmpage(pp, ph, pq);
 		} else {
 			LIST_REMOVE(ph, ph_pagelist);
@@ -1139,9 +1268,7 @@ pool_do_put(struct pool *pp, void *v, st
 			 * 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);
 	}
@@ -1160,7 +1287,7 @@ pool_do_put(struct pool *pp, void *v, st
 }
 
 /*
- * Return resource to the pool; must be called at appropriate spl level
+ * Return resource to the pool.
  */
 #ifdef POOL_DIAGNOSTIC
 void
@@ -1170,7 +1297,7 @@ _pool_put(struct pool *pp, void *v, cons
 
 	LIST_INIT(&pq);
 
-	simple_lock(&pp->pr_slock);
+	mutex_enter(&pp->pr_lock);
 	pr_enter(pp, file, line);
 
 	pr_log(pp, v, PRLOG_PUT, file, line);
@@ -1178,7 +1305,7 @@ _pool_put(struct pool *pp, void *v, cons
 	pool_do_put(pp, v, &pq);
 
 	pr_leave(pp);
-	simple_unlock(&pp->pr_slock);
+	mutex_exit(&pp->pr_lock);
 
 	pr_pagelist_free(pp, &pq);
 }
@@ -1192,9 +1319,9 @@ pool_put(struct pool *pp, void *v)
 
 	LIST_INIT(&pq);
 
-	simple_lock(&pp->pr_slock);
+	mutex_enter(&pp->pr_lock);
 	pool_do_put(pp, v, &pq);
-	simple_unlock(&pp->pr_slock);
+	mutex_exit(&pp->pr_lock);
 
 	pr_pagelist_free(pp, &pq);
 }
@@ -1204,43 +1331,64 @@ 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;
+
+	mutex_exit(&pp->pr_lock);
+	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);
+		}
+		mutex_enter(&pp->pr_lock);
+		return ENOMEM;
+	}
+
+	mutex_enter(&pp->pr_lock);
+	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 = NULL;
-	caddr_t cp;
 	int newpages;
+	int error = 0;
 
-	simple_lock(&pp->pr_slock);
+	mutex_enter(&pp->pr_lock);
 
 	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);
+		error = pool_grow(pp, PR_NOWAIT);
+		if (error) {
 			break;
 		}
-
-		simple_lock(&pp->pr_slock);
-		pool_prime_page(pp, cp, ph);
-		pp->pr_npagealloc++;
 		pp->pr_minpages++;
 	}
 
 	if (pp->pr_minpages >= pp->pr_maxpages)
 		pp->pr_maxpages = pp->pr_minpages + 1;	/* XXX */
 
-	simple_unlock(&pp->pr_slock);
-	return (0);
+	mutex_exit(&pp->pr_lock);
+	return error;
 }
 
 /*
@@ -1249,19 +1397,19 @@ pool_prime(struct pool *pp, int n)
  * Note, we must be called with the pool descriptor LOCKED.
  */
 static void
-pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
+pool_prime_page(struct pool *pp, void *storage, struct pool_item_header *ph)
 {
 	struct pool_item *pi;
-	caddr_t cp = storage;
-	unsigned int align = pp->pr_align;
-	unsigned int ioff = pp->pr_itemoffset;
+	void *cp = storage;
+	const unsigned int align = pp->pr_align;
+	const unsigned int ioff = pp->pr_itemoffset;
 	int n;
-	int s;
 
-	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
+	KASSERT(mutex_owned(&pp->pr_lock));
 
 #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
 
@@ -1272,9 +1420,7 @@ pool_prime_page(struct pool *pp, caddr_t
 	LIST_INIT(&ph->ph_itemlist);
 	ph->ph_page = storage;
 	ph->ph_nmissing = 0;
-	s = splclock();
-	ph->ph_time = mono_time;
-	splx(s);
+	getmicrotime(&ph->ph_time);
 	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
 		SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
 
@@ -1283,7 +1429,7 @@ pool_prime_page(struct pool *pp, caddr_t
 	/*
 	 * Color this page.
 	 */
-	cp = (caddr_t)(cp + pp->pr_curcolor);
+	cp = (char *)cp + pp->pr_curcolor;
 	if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
 		pp->pr_curcolor = 0;
 
@@ -1291,7 +1437,9 @@ pool_prime_page(struct pool *pp, caddr_t
 	 * Adjust storage to apply aligment to `pr_itemoffset' in each item.
 	 */
 	if (ioff != 0)
-		cp = (caddr_t)(cp + (align - ioff));
+		cp = (char *)cp + align - ioff;
+
+	KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
 
 	/*
 	 * Insert remaining chunks on the bucket list.
@@ -1300,14 +1448,7 @@ pool_prime_page(struct pool *pp, caddr_t
 	pp->pr_nitems += n;
 
 	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;
+		pr_item_notouch_init(pp, ph);
 	} else {
 		while (n--) {
 			pi = (struct pool_item *)cp;
@@ -1319,7 +1460,9 @@ pool_prime_page(struct pool *pp, caddr_t
 #ifdef DIAGNOSTIC
 			pi->pi_magic = PI_MAGIC;
 #endif
-			cp = (caddr_t)(cp + pp->pr_size);
+			cp = (char *)cp + pp->pr_size;
+
+			KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
 		}
 	}
 
@@ -1345,34 +1488,15 @@ pool_prime_page(struct pool *pp, caddr_t
 static int
 pool_catchup(struct pool *pp)
 {
-	struct pool_item_header *ph = NULL;
-	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);
-		if (__predict_false(cp == NULL || ph == NULL)) {
-			if (cp != NULL)
-				pool_allocator_free(pp, cp);
-			error = ENOMEM;
-			simple_lock(&pp->pr_slock);
+		error = pool_grow(pp, PR_NOWAIT);
+		if (error) {
 			break;
 		}
-		simple_lock(&pp->pr_slock);
-		pool_prime_page(pp, cp, ph);
-		pp->pr_npagealloc++;
 	}
-
-	return (error);
+	return error;
 }
 
 static void
@@ -1389,7 +1513,7 @@ void
 pool_setlowat(struct pool *pp, int n)
 {
 
-	simple_lock(&pp->pr_slock);
+	mutex_enter(&pp->pr_lock);
 
 	pp->pr_minitems = n;
 	pp->pr_minpages = (n == 0)
@@ -1405,27 +1529,27 @@ pool_setlowat(struct pool *pp, int n)
 		 */
 	}
 
-	simple_unlock(&pp->pr_slock);
+	mutex_exit(&pp->pr_lock);
 }
 
 void
 pool_sethiwat(struct pool *pp, int n)
 {
 
-	simple_lock(&pp->pr_slock);
+	mutex_enter(&pp->pr_lock);
 
 	pp->pr_maxpages = (n == 0)
 		? 0
 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
 
-	simple_unlock(&pp->pr_slock);
+	mutex_exit(&pp->pr_lock);
 }
 
 void
 pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
 {
 
-	simple_lock(&pp->pr_slock);
+	mutex_enter(&pp->pr_lock);
 
 	pp->pr_hardlimit = n;
 	pp->pr_hardlimit_warning = warnmess;
@@ -1441,7 +1565,7 @@ pool_sethardlimit(struct pool *pp, int n
 		? 0
 		: roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
 
-	simple_unlock(&pp->pr_slock);
+	mutex_exit(&pp->pr_lock);
 }
 
 /*
@@ -1455,11 +1579,10 @@ pool_reclaim(struct pool *pp)
 #endif
 {
 	struct pool_item_header *ph, *phnext;
-	struct pool_cache *pc;
 	struct pool_pagelist pq;
-	struct pool_cache_grouplist pcgl;
 	struct timeval curtime, diff;
-	int s;
+	bool klock;
+	int rv;
 
 	if (pp->pr_drain_hook != NULL) {
 		/*
@@ -1468,22 +1591,33 @@ pool_reclaim(struct pool *pp)
 		(*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
 	}
 
-	if (simple_lock_try(&pp->pr_slock) == 0)
+	/*
+	 * XXXSMP Because mutexes at IPL_SOFTXXX are still spinlocks,
+	 * and we are called from the pagedaemon without kernel_lock.
+	 * Does not apply to IPL_SOFTBIO.
+	 */
+	if (pp->pr_ipl == IPL_SOFTNET || pp->pr_ipl == IPL_SOFTCLOCK ||
+	    pp->pr_ipl == IPL_SOFTSERIAL) {
+		KERNEL_LOCK(1, NULL);
+		klock = true;
+	} else
+		klock = false;
+
+	/* Reclaim items from the pool's cache (if any). */
+	if (pp->pr_cache != NULL)
+		pool_cache_invalidate(pp->pr_cache);
+
+	if (mutex_tryenter(&pp->pr_lock) == 0) {
+		if (klock) {
+			KERNEL_UNLOCK_ONE(NULL);
+		}
 		return (0);
+	}
 	pr_enter(pp, file, line);
 
 	LIST_INIT(&pq);
-	LIST_INIT(&pcgl);
-
-	/*
-	 * Reclaim items from the pool's caches.
-	 */
-	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 = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
 		phnext = LIST_NEXT(ph, ph_pagelist);
@@ -1494,7 +1628,8 @@ pool_reclaim(struct pool *pp)
 
 		KASSERT(ph->ph_nmissing == 0);
 		timersub(&curtime, &ph->ph_time, &diff);
-		if (diff.tv_sec < pool_inactive_time)
+		if (diff.tv_sec < pool_inactive_time
+		    && !pa_starved_p(pp->pr_alloc))
 			continue;
 
 		/*
@@ -1509,39 +1644,87 @@ pool_reclaim(struct pool *pp)
 	}
 
 	pr_leave(pp);
-	simple_unlock(&pp->pr_slock);
-	if (LIST_EMPTY(&pq) && LIST_EMPTY(&pcgl))
-		return 0;
+	mutex_exit(&pp->pr_lock);
 
-	pr_pagelist_free(pp, &pq);
-	pcg_grouplist_free(&pcgl);
-	return (1);
+	if (LIST_EMPTY(&pq))
+		rv = 0;
+	else {
+		pr_pagelist_free(pp, &pq);
+		rv = 1;
+	}
+
+	if (klock) {
+		KERNEL_UNLOCK_ONE(NULL);
+	}
+
+	return (rv);
 }
 
 /*
- * Drain pools, one at a time.
+ * Drain pools, one at a time.  This is a two stage process;
+ * drain_start kicks off a cross call to drain CPU-level caches
+ * if the pool has an associated pool_cache.  drain_end waits
+ * for those cross calls to finish, and then drains the cache
+ * (if any) and pool.
  *
- * Note, we must never be called from an interrupt context.
+ * Note, must never be called from interrupt context.
  */
 void
-pool_drain(void *arg)
+pool_drain_start(struct pool **ppp, uint64_t *wp)
 {
 	struct pool *pp;
-	int s;
+
+	KASSERT(!LIST_EMPTY(&pool_head));
 
 	pp = NULL;
-	s = splvm();
-	simple_lock(&pool_head_slock);
-	if (drainpp == NULL) {
-		drainpp = LIST_FIRST(&pool_head);
-	}
-	if (drainpp) {
-		pp = drainpp;
-		drainpp = LIST_NEXT(pp, pr_poollist);
+
+	/* Find next pool to drain, and add a reference. */
+	mutex_enter(&pool_head_lock);
+	do {
+		if (drainpp == NULL) {
+			drainpp = LIST_FIRST(&pool_head);
+		}
+		if (drainpp != NULL) {
+			pp = drainpp;
+			drainpp = LIST_NEXT(pp, pr_poollist);
+		}
+		/*
+		 * Skip completely idle pools.  We depend on at least
+		 * one pool in the system being active.
+		 */
+	} while (pp == NULL || pp->pr_npages == 0);
+	pp->pr_refcnt++;
+	mutex_exit(&pool_head_lock);
+
+	/* If there is a pool_cache, drain CPU level caches. */
+	*ppp = pp;
+	if (pp->pr_cache != NULL) {
+		*wp = xc_broadcast(0, (xcfunc_t)pool_cache_xcall,
+		    pp->pr_cache, NULL);
 	}
-	simple_unlock(&pool_head_slock);
+}
+
+void
+pool_drain_end(struct pool *pp, uint64_t where)
+{
+
+	if (pp == NULL)
+		return;
+
+	KASSERT(pp->pr_refcnt > 0);
+
+	/* Wait for remote draining to complete. */
+	if (pp->pr_cache != NULL)
+		xc_wait(where);
+
+	/* Drain the cache (if any) and pool.. */
 	pool_reclaim(pp);
-	splx(s);
+
+	/* Finally, unlock the pool. */
+	mutex_enter(&pool_head_lock);
+	pp->pr_refcnt--;
+	cv_broadcast(&pool_busy);
+	mutex_exit(&pool_head_lock);
 }
 
 /*
@@ -1550,18 +1733,8 @@ pool_drain(void *arg)
 void
 pool_print(struct pool *pp, const char *modif)
 {
-	int s;
 
-	s = splvm();
-	if (simple_lock_try(&pp->pr_slock) == 0) {
-		printf("pool %s is locked; try again later\n",
-		    pp->pr_wchan);
-		splx(s);
-		return;
-	}
 	pool_print1(pp, modif, printf);
-	simple_unlock(&pp->pr_slock);
-	splx(s);
 }
 
 void
@@ -1569,12 +1742,6 @@ pool_printall(const char *modif, void (*
 {
 	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);
 	}
@@ -1589,20 +1756,6 @@ pool_printit(struct pool *pp, const char
 		return;
 	}
 
-	/*
-	 * Called from DDB; interrupts should be blocked, and all
-	 * other processors should be paused.  We can skip locking
-	 * the pool in this case.
-	 *
-	 * We do a simple_lock_try() just to print the lock
-	 * status, however.
-	 */
-
-	if (simple_lock_try(&pp->pr_slock) == 0)
-		(*pr)("WARNING: pool %s is locked\n", pp->pr_wchan);
-	else
-		simple_unlock(&pp->pr_slock);
-
 	pool_print1(pp, modif, pr);
 }
 
@@ -1637,8 +1790,10 @@ 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;
+	pool_cache_t pc;
+	pcg_t *pcg;
+	pool_cache_cpu_t *cc;
+	uint64_t cpuhit, cpumiss;
 	int i, print_log = 0, print_pagelist = 0, print_cache = 0;
 	char c;
 
@@ -1651,7 +1806,13 @@ pool_print1(struct pool *pp, const char 
 			print_cache = 1;
 	}
 
-	(*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
+	if ((pc = pp->pr_cache) != NULL) {
+		(*pr)("POOL CACHE");
+	} else {
+		(*pr)("POOL");
+	}
+
+	(*pr)(" %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)("\talloc %p\n", pp->pr_alloc);
@@ -1660,7 +1821,7 @@ pool_print1(struct pool *pp, const char 
 	(*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
 	    pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
 
-	(*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
+	(*pr)("\tnget %lu, nfail %lu, nput %lu\n",
 	    pp->pr_nget, pp->pr_nfail, pp->pr_nput);
 	(*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
 	    pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
@@ -1690,12 +1851,11 @@ pool_print1(struct pool *pp, const char 
 	(*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;
 
 #define PR_GROUPLIST(pcg)						\
 	(*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail);		\
@@ -1712,26 +1872,38 @@ pool_print1(struct pool *pp, const char 
 		}							\
 	}
 
-	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);
-		(*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);
+	if (pc != NULL) {
+		cpuhit = 0;
+		cpumiss = 0;
+		for (i = 0; i < MAXCPUS; i++) {
+			if ((cc = pc->pc_cpus[i]) == NULL)
+				continue;
+			cpuhit += cc->cc_hits;
+			cpumiss += cc->cc_misses;
+		}
+		(*pr)("\tcpu layer hits %llu misses %llu\n", cpuhit, cpumiss);
+		(*pr)("\tcache layer hits %llu misses %llu\n",
+		    pc->pc_hits, pc->pc_misses);
+		(*pr)("\tcache layer entry uncontended %llu contended %llu\n",
+		    pc->pc_hits + pc->pc_misses - pc->pc_contended,
+		    pc->pc_contended);
+		(*pr)("\tcache layer empty groups %u full groups %u\n",
+		    pc->pc_nempty, pc->pc_nfull);
+		if (print_cache) {
+			(*pr)("\tfull cache groups:\n");
+			for (pcg = pc->pc_fullgroups; pcg != NULL;
+			    pcg = pcg->pcg_next) {
+				PR_GROUPLIST(pcg);
+			}
+			(*pr)("\tempty cache groups:\n");
+			for (pcg = pc->pc_emptygroups; pcg != NULL;
+			    pcg = pcg->pcg_next) {
+				PR_GROUPLIST(pcg);
+			}
 		}
 	}
 #undef PR_GROUPLIST
 
- skip_cache:
 	pr_enter_check(pp, pr);
 }
 
@@ -1739,19 +1911,21 @@ static int
 pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
 {
 	struct pool_item *pi;
-	caddr_t page;
+	void *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_NOALIGN) == 0) {
+		page = (void *)((uintptr_t)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)
@@ -1766,15 +1940,16 @@ pool_chk_page(struct pool *pp, const cha
 			if (label != NULL)
 				printf("%s: ", label);
 			printf("pool(%s): free list modified: magic=%x;"
-			       " page %p; item ordinal %d;"
-			       " addr %p (p %p)\n",
+			       " page %p; item ordinal %d; addr %p\n",
 				pp->pr_wchan, pi->pi_magic, ph->ph_page,
-				n, pi, page);
+				n, pi);
 			panic("pool");
 		}
 #endif
-		page =
-		    (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
+		if ((pp->pr_roflags & PR_NOALIGN) != 0) {
+			continue;
+		}
+		page = (void *)((uintptr_t)pi & pp->pr_alloc->pa_pagemask);
 		if (page == ph->ph_page)
 			continue;
 
@@ -1796,7 +1971,7 @@ pool_chk(struct pool *pp, const char *la
 	struct pool_item_header *ph;
 	int r = 0;
 
-	simple_lock(&pp->pr_slock);
+	mutex_enter(&pp->pr_lock);
 	LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
 		r = pool_chk_page(pp, label, ph);
 		if (r) {
@@ -1817,7 +1992,7 @@ pool_chk(struct pool *pp, const char *la
 	}
 
 out:
-	simple_unlock(&pp->pr_slock);
+	mutex_exit(&pp->pr_lock);
 	return (r);
 }
 
@@ -1825,38 +2000,93 @@ out:
  * pool_cache_init:
  *
  *	Initialize a pool cache.
+ */
+pool_cache_t
+pool_cache_init(size_t size, u_int align, u_int align_offset, u_int flags,
+    const char *wchan, struct pool_allocator *palloc, int ipl,
+    int (*ctor)(void *, void *, int), void (*dtor)(void *, void *), void *arg)
+{
+	pool_cache_t pc;
+
+	pc = pool_get(&cache_pool, PR_WAITOK);
+	if (pc == NULL)
+		return NULL;
+
+	pool_cache_bootstrap(pc, size, align, align_offset, flags, wchan,
+	   palloc, ipl, ctor, dtor, arg);
+
+	return pc;
+}
+
+/*
+ * pool_cache_bootstrap:
  *
- *	NOTE: If the pool must be protected from interrupts, we expect
- *	to be called at the appropriate interrupt priority level.
+ *	Kernel-private version of pool_cache_init().  The caller
+ *	provides initial storage.
  */
 void
-pool_cache_init(struct pool_cache *pc, struct pool *pp,
-    int (*ctor)(void *, void *, int),
-    void (*dtor)(void *, void *),
+pool_cache_bootstrap(pool_cache_t pc, size_t size, u_int align,
+    u_int align_offset, u_int flags, const char *wchan,
+    struct pool_allocator *palloc, int ipl,
+    int (*ctor)(void *, void *, int), void (*dtor)(void *, void *),
     void *arg)
 {
+	CPU_INFO_ITERATOR cii;
+	struct cpu_info *ci;
+	struct pool *pp;
 
-	LIST_INIT(&pc->pc_emptygroups);
-	LIST_INIT(&pc->pc_fullgroups);
-	LIST_INIT(&pc->pc_partgroups);
-	simple_lock_init(&pc->pc_slock);
+	pp = &pc->pc_pool;
+	if (palloc == NULL && ipl == IPL_NONE)
+		palloc = &pool_allocator_nointr;
+	pool_init(pp, size, align, align_offset, flags, wchan, palloc, ipl);
 
-	pc->pc_pool = pp;
+	mutex_init(&pc->pc_lock, MUTEX_DEFAULT, pp->pr_ipl);
 
+	if (ctor == NULL) {
+		ctor = (int (*)(void *, void *, int))nullop;
+	}
+	if (dtor == NULL) {
+		dtor = (void (*)(void *, void *))nullop;
+	}
+
+	pc->pc_emptygroups = NULL;
+	pc->pc_fullgroups = NULL;
+	pc->pc_partgroups = NULL;
 	pc->pc_ctor = ctor;
 	pc->pc_dtor = dtor;
 	pc->pc_arg  = arg;
-
-	pc->pc_hits   = 0;
+	pc->pc_hits  = 0;
 	pc->pc_misses = 0;
-
-	pc->pc_ngroups = 0;
-
-	pc->pc_nitems = 0;
-
-	simple_lock(&pp->pr_slock);
-	LIST_INSERT_HEAD(&pp->pr_cachelist, pc, pc_poollist);
-	simple_unlock(&pp->pr_slock);
+	pc->pc_nempty = 0;
+	pc->pc_npart = 0;
+	pc->pc_nfull = 0;
+	pc->pc_contended = 0;
+	pc->pc_refcnt = 0;
+	pc->pc_freecheck = NULL;
+
+	/* Allocate per-CPU caches. */
+	memset(pc->pc_cpus, 0, sizeof(pc->pc_cpus));
+	pc->pc_ncpu = 0;
+	if (ncpu == 0) {
+		/* XXX For sparc: boot CPU is not attached yet. */
+		pool_cache_cpu_init1(curcpu(), pc);
+	} else {
+		for (CPU_INFO_FOREACH(cii, ci)) {
+			pool_cache_cpu_init1(ci, pc);
+		}
+	}
+	
+	if (__predict_true(!cold)) {
+		mutex_enter(&pp->pr_lock);
+		pp->pr_cache = pc;
+		mutex_exit(&pp->pr_lock);
+		mutex_enter(&pool_head_lock);
+		LIST_INSERT_HEAD(&pool_cache_head, pc, pc_cachelist);
+		mutex_exit(&pool_head_lock);
+	} else {
+		pp->pr_cache = pc;
+		LIST_INSERT_HEAD(&pool_cache_head, pc, pc_cachelist);
+	}
 }
 
 /*
@@ -1865,297 +2095,618 @@ pool_cache_init(struct pool_cache *pc, s
  *	Destroy a pool cache.
  */
 void
-pool_cache_destroy(struct pool_cache *pc)
+pool_cache_destroy(pool_cache_t pc)
 {
-	struct pool *pp = pc->pc_pool;
+	struct pool *pp = &pc->pc_pool;
+	pool_cache_cpu_t *cc;
+	pcg_t *pcg;
+	int i;
+
+	/* Remove it from the global list. */
+	mutex_enter(&pool_head_lock);
+	while (pc->pc_refcnt != 0)
+		cv_wait(&pool_busy, &pool_head_lock);
+	LIST_REMOVE(pc, pc_cachelist);
+	mutex_exit(&pool_head_lock);
 
 	/* First, invalidate the entire cache. */
 	pool_cache_invalidate(pc);
 
-	/* ...and remove it from the pool's cache list. */
-	simple_lock(&pp->pr_slock);
-	LIST_REMOVE(pc, pc_poollist);
-	simple_unlock(&pp->pr_slock);
+	/* Disassociate it from the pool. */
+	mutex_enter(&pp->pr_lock);
+	pp->pr_cache = NULL;
+	mutex_exit(&pp->pr_lock);
+
+	/* Destroy per-CPU data */
+	for (i = 0; i < MAXCPUS; i++) {
+		if ((cc = pc->pc_cpus[i]) == NULL)
+			continue;
+		if ((pcg = cc->cc_current) != NULL) {
+			pcg->pcg_next = NULL;
+			pool_cache_invalidate_groups(pc, pcg);
+		}
+		if ((pcg = cc->cc_previous) != NULL) {
+			pcg->pcg_next = NULL;
+			pool_cache_invalidate_groups(pc, pcg);
+		}
+		if (cc != &pc->pc_cpu0)
+			pool_put(&cache_cpu_pool, cc);
+	}
+
+	/* Finally, destroy it. */
+	mutex_destroy(&pc->pc_lock);
+	pool_destroy(pp);
+	pool_put(&cache_pool, pc);
 }
 
-static inline void *
-pcg_get(struct pool_cache_group *pcg, paddr_t *pap)
-{
-	void *object;
-	u_int idx;
+/*
+ * pool_cache_cpu_init1:
+ *
+ *	Called for each pool_cache whenever a new CPU is attached.
+ */
+static void
+pool_cache_cpu_init1(struct cpu_info *ci, pool_cache_t pc)
+{
+	pool_cache_cpu_t *cc;
+	int index;
 
-	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;
+	index = ci->ci_index;
+
+	KASSERT(index < MAXCPUS);
+	KASSERT(((uintptr_t)pc->pc_cpus & (CACHE_LINE_SIZE - 1)) == 0);
+
+	if ((cc = pc->pc_cpus[index]) != NULL) {
+		KASSERT(cc->cc_cpuindex == index);
+		return;
+	}
 
-	return (object);
+	/*
+	 * The first CPU is 'free'.  This needs to be the case for
+	 * bootstrap - we may not be able to allocate yet.
+	 */
+	if (pc->pc_ncpu == 0) {
+		cc = &pc->pc_cpu0;
+		pc->pc_ncpu = 1;
+	} else {
+		mutex_enter(&pc->pc_lock);
+		pc->pc_ncpu++;
+		mutex_exit(&pc->pc_lock);
+		cc = pool_get(&cache_cpu_pool, PR_WAITOK);
+	}
+
+	cc->cc_ipl = pc->pc_pool.pr_ipl;
+	cc->cc_iplcookie = makeiplcookie(cc->cc_ipl);
+	cc->cc_cache = pc;
+	cc->cc_cpuindex = index;
+	cc->cc_hits = 0;
+	cc->cc_misses = 0;
+	cc->cc_current = NULL;
+	cc->cc_previous = NULL;
+
+	pc->pc_cpus[index] = cc;
 }
 
-static inline void
-pcg_put(struct pool_cache_group *pcg, void *object, paddr_t pa)
+/*
+ * pool_cache_cpu_init:
+ *
+ *	Called whenever a new CPU is attached.
+ */
+void
+pool_cache_cpu_init(struct cpu_info *ci)
 {
-	u_int idx;
+	pool_cache_t pc;
+
+	mutex_enter(&pool_head_lock);
+	LIST_FOREACH(pc, &pool_cache_head, pc_cachelist) {
+		pc->pc_refcnt++;
+		mutex_exit(&pool_head_lock);
+
+		pool_cache_cpu_init1(ci, pc);
+
+		mutex_enter(&pool_head_lock);
+		pc->pc_refcnt--;
+		cv_broadcast(&pool_busy);
+	}
+	mutex_exit(&pool_head_lock);
+}
 
-	KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
-	idx = pcg->pcg_avail++;
+/*
+ * pool_cache_reclaim:
+ *
+ *	Reclaim memory from a pool cache.
+ */
+bool
+pool_cache_reclaim(pool_cache_t pc)
+{
 
-	KASSERT(pcg->pcg_objects[idx].pcgo_va == NULL);
-	pcg->pcg_objects[idx].pcgo_va = object;
-	pcg->pcg_objects[idx].pcgo_pa = pa;
+	return pool_reclaim(&pc->pc_pool);
 }
 
 static void
-pcg_grouplist_free(struct pool_cache_grouplist *pcgl)
+pool_cache_destruct_object1(pool_cache_t pc, void *object)
 {
-	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);
+	(*pc->pc_dtor)(pc->pc_arg, object);
+	pool_put(&pc->pc_pool, object);
 }
 
 /*
- * pool_cache_get{,_paddr}:
+ * pool_cache_destruct_object:
  *
- *	Get an object from a pool cache (optionally returning
- *	the physical address of the object).
+ *	Force destruction of an object and its release back into
+ *	the pool.
  */
-void *
-pool_cache_get_paddr(struct pool_cache *pc, int flags, paddr_t *pap)
+void
+pool_cache_destruct_object(pool_cache_t pc, void *object)
 {
-	struct pool_cache_group *pcg;
-	void *object;
 
-#ifdef LOCKDEBUG
-	if (flags & PR_WAITOK)
-		simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)");
-#endif
+	FREECHECK_IN(&pc->pc_freecheck, object);
 
-	simple_lock(&pc->pc_slock);
+	pool_cache_destruct_object1(pc, object);
+}
 
-	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) {
+/*
+ * pool_cache_invalidate_groups:
+ *
+ *	Invalidate a chain of groups and destruct all objects.
+ */
+static void
+pool_cache_invalidate_groups(pool_cache_t pc, pcg_t *pcg)
+{
+	void *object;
+	pcg_t *next;
+	int i;
 
-		/*
-		 * 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
-		}
-		return (object);
-	}
+	for (; pcg != NULL; pcg = next) {
+		next = pcg->pcg_next;
 
-	pc->pc_hits++;
-	pc->pc_nitems--;
-	object = pcg_get(pcg, pap);
+		for (i = 0; i < pcg->pcg_avail; i++) {
+			object = pcg->pcg_objects[i].pcgo_va;
+			pool_cache_destruct_object1(pc, object);
+		}
 
-	if (pcg->pcg_avail == 0) {
-		LIST_REMOVE(pcg, pcg_list);
-		LIST_INSERT_HEAD(&pc->pc_emptygroups, pcg, pcg_list);
+		pool_put(&pcgpool, pcg);
 	}
-	simple_unlock(&pc->pc_slock);
-
-	return (object);
 }
 
 /*
- * pool_cache_put{,_paddr}:
+ * pool_cache_invalidate:
  *
- *	Put an object back to the pool cache (optionally caching the
- *	physical address of the object).
+ *	Invalidate a pool cache (destruct and release all of the
+ *	cached objects).  Does not reclaim objects from the pool.
  */
 void
-pool_cache_put_paddr(struct pool_cache *pc, void *object, paddr_t pa)
+pool_cache_invalidate(pool_cache_t pc)
 {
-	struct pool_cache_group *pcg;
-	int s;
+	pcg_t *full, *empty, *part;
+
+	mutex_enter(&pc->pc_lock);
+	full = pc->pc_fullgroups;
+	empty = pc->pc_emptygroups;
+	part = pc->pc_partgroups;
+	pc->pc_fullgroups = NULL;
+	pc->pc_emptygroups = NULL;
+	pc->pc_partgroups = NULL;
+	pc->pc_nfull = 0;
+	pc->pc_nempty = 0;
+	pc->pc_npart = 0;
+	mutex_exit(&pc->pc_lock);
+
+	pool_cache_invalidate_groups(pc, full);
+	pool_cache_invalidate_groups(pc, empty);
+	pool_cache_invalidate_groups(pc, part);
+}
+
+void
+pool_cache_set_drain_hook(pool_cache_t pc, void (*fn)(void *, int), void *arg)
+{
+
+	pool_set_drain_hook(&pc->pc_pool, fn, arg);
+}
+
+void
+pool_cache_setlowat(pool_cache_t pc, int n)
+{
+
+	pool_setlowat(&pc->pc_pool, n);
+}
+
+void
+pool_cache_sethiwat(pool_cache_t pc, int n)
+{
+
+	pool_sethiwat(&pc->pc_pool, n);
+}
+
+void
+pool_cache_sethardlimit(pool_cache_t pc, int n, const char *warnmess, int ratecap)
+{
+
+	pool_sethardlimit(&pc->pc_pool, n, warnmess, ratecap);
+}
 
-	if (__predict_false((pc->pc_pool->pr_flags & PR_WANTED) != 0)) {
-		goto destruct;
+static inline pool_cache_cpu_t *
+pool_cache_cpu_enter(pool_cache_t pc, int *s)
+{
+	pool_cache_cpu_t *cc;
+
+	/*
+	 * Prevent other users of the cache from accessing our
+	 * CPU-local data.  To avoid touching shared state, we
+	 * pull the neccessary information from CPU local data.
+	 */
+	crit_enter();
+	cc = pc->pc_cpus[curcpu()->ci_index];
+	KASSERT(cc->cc_cache == pc);
+	if (cc->cc_ipl != IPL_NONE) {
+		*s = splraiseipl(cc->cc_iplcookie);
 	}
+	KASSERT(((uintptr_t)cc & (CACHE_LINE_SIZE - 1)) == 0);
 
-	simple_lock(&pc->pc_slock);
+	return cc;
+}
 
-	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);
-		}
+static inline void
+pool_cache_cpu_exit(pool_cache_cpu_t *cc, int *s)
+{
+
+	/* No longer need exclusive access to the per-CPU data. */
+	if (cc->cc_ipl != IPL_NONE) {
+		splx(*s);
 	}
-	if (pcg == NULL) {
+	crit_exit();
+}
+
+#if __GNUC_PREREQ__(3, 0)
+__attribute ((noinline))
+#endif
+pool_cache_cpu_t *
+pool_cache_get_slow(pool_cache_cpu_t *cc, int *s, void **objectp,
+		    paddr_t *pap, int flags)
+{
+	pcg_t *pcg, *cur;
+	uint64_t ncsw;
+	pool_cache_t pc;
+	void *object;
+
+	pc = cc->cc_cache;
+	cc->cc_misses++;
+
+	/*
+	 * Nothing was available locally.  Try and grab a group
+	 * from the cache.
+	 */
+	if (!mutex_tryenter(&pc->pc_lock)) {
+		ncsw = curlwp->l_ncsw;
+		mutex_enter(&pc->pc_lock);
+		pc->pc_contended++;
 
 		/*
-		 * No empty groups to free the object to.  Attempt to
-		 * allocate one.
+		 * If we context switched while locking, then
+		 * our view of the per-CPU data is invalid:
+		 * retry.
 		 */
-		simple_unlock(&pc->pc_slock);
-		s = splvm();
-		pcg = pool_get(&pcgpool, PR_NOWAIT);
-		splx(s);
-		if (pcg == NULL) {
-destruct:
+		if (curlwp->l_ncsw != ncsw) {
+			mutex_exit(&pc->pc_lock);
+			pool_cache_cpu_exit(cc, s);
+			return pool_cache_cpu_enter(pc, s);
+		}
+	}
 
-			/*
-			 * Unable to allocate a cache group; destruct the object
-			 * and free it back to the pool.
-			 */
-			pool_cache_destruct_object(pc, object);
-			return;
+	if ((pcg = pc->pc_fullgroups) != NULL) {
+		/*
+		 * If there's a full group, release our empty
+		 * group back to the cache.  Install the full
+		 * group as cc_current and return.
+		 */
+		if ((cur = cc->cc_current) != NULL) {
+			KASSERT(cur->pcg_avail == 0);
+			cur->pcg_next = pc->pc_emptygroups;
+			pc->pc_emptygroups = cur;
+			pc->pc_nempty++;
 		}
-		memset(pcg, 0, sizeof(*pcg));
-		simple_lock(&pc->pc_slock);
-		pc->pc_ngroups++;
-		LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
+		KASSERT(pcg->pcg_avail == PCG_NOBJECTS);
+		cc->cc_current = pcg;
+		pc->pc_fullgroups = pcg->pcg_next;
+		pc->pc_hits++;
+		pc->pc_nfull--;
+		mutex_exit(&pc->pc_lock);
+		return cc;
+	}
+
+	/*
+	 * Nothing available locally or in cache.  Take the slow
+	 * path: fetch a new object from the pool and construct
+	 * it.
+	 */
+	pc->pc_misses++;
+	mutex_exit(&pc->pc_lock);
+	pool_cache_cpu_exit(cc, s);
+
+	object = pool_get(&pc->pc_pool, flags);
+	*objectp = object;
+	if (object == NULL)
+		return NULL;
+
+	if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
+		pool_put(&pc->pc_pool, object);
+		*objectp = NULL;
+		return NULL;
 	}
 
-	pc->pc_nitems++;
-	pcg_put(pcg, object, pa);
+	KASSERT((((vaddr_t)object + pc->pc_pool.pr_itemoffset) &
+	    (pc->pc_pool.pr_align - 1)) == 0);
 
-	if (pcg->pcg_avail == PCG_NOBJECTS) {
-		LIST_REMOVE(pcg, pcg_list);
-		LIST_INSERT_HEAD(&pc->pc_fullgroups, pcg, pcg_list);
+	if (pap != NULL) {
+#ifdef POOL_VTOPHYS
+		*pap = POOL_VTOPHYS(object);
+#else
+		*pap = POOL_PADDR_INVALID;
+#endif
 	}
-	simple_unlock(&pc->pc_slock);
+
+	FREECHECK_OUT(&pc->pc_freecheck, object);
+	return NULL;
 }
 
 /*
- * pool_cache_destruct_object:
+ * pool_cache_get{,_paddr}:
  *
- *	Force destruction of an object and its release back into
- *	the pool.
+ *	Get an object from a pool cache (optionally returning
+ *	the physical address of the object).
  */
-void
-pool_cache_destruct_object(struct pool_cache *pc, void *object)
+void *
+pool_cache_get_paddr(pool_cache_t pc, int flags, paddr_t *pap)
 {
+	pool_cache_cpu_t *cc;
+	pcg_t *pcg;
+	void *object;
+	int s;
 
-	if (pc->pc_dtor != NULL)
-		(*pc->pc_dtor)(pc->pc_arg, object);
-	pool_put(pc->pc_pool, object);
+#ifdef LOCKDEBUG
+	if (flags & PR_WAITOK)
+		ASSERT_SLEEPABLE(NULL, "pool_cache_get(PR_WAITOK)");
+#endif
+
+	cc = pool_cache_cpu_enter(pc, &s);
+	do {
+		/* Try and allocate an object from the current group. */
+	 	pcg = cc->cc_current;
+		if (pcg != NULL && pcg->pcg_avail > 0) {
+			object = pcg->pcg_objects[--pcg->pcg_avail].pcgo_va;
+			if (pap != NULL)
+				*pap = pcg->pcg_objects[pcg->pcg_avail].pcgo_pa;
+			pcg->pcg_objects[pcg->pcg_avail].pcgo_va = NULL;
+			KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
+			KASSERT(object != NULL);
+			cc->cc_hits++;
+			pool_cache_cpu_exit(cc, &s);
+			FREECHECK_OUT(&pc->pc_freecheck, object);
+			return object;
+		}
+
+		/*
+		 * That failed.  If the previous group isn't empty, swap
+		 * it with the current group and allocate from there.
+		 */
+		pcg = cc->cc_previous;
+		if (pcg != NULL && pcg->pcg_avail > 0) {
+			cc->cc_previous = cc->cc_current;
+			cc->cc_current = pcg;
+			continue;
+		}
+
+		/*
+		 * Can't allocate from either group: try the slow path.
+		 * If get_slow() allocated an object for us, or if
+		 * no more objects are available, it will return NULL.
+		 * Otherwise, we need to retry.
+		 */
+		cc = pool_cache_get_slow(cc, &s, &object, pap, flags);
+	} while (cc != NULL);
+
+	return object;
 }
 
-static void
-pool_do_cache_invalidate_grouplist(struct pool_cache_grouplist *pcgsl,
-    struct pool_cache *pc, struct pool_pagelist *pq,
-    struct pool_cache_grouplist *pcgdl)
+#if __GNUC_PREREQ__(3, 0)
+__attribute ((noinline))
+#endif
+pool_cache_cpu_t *
+pool_cache_put_slow(pool_cache_cpu_t *cc, int *s, void *object, paddr_t pa)
 {
-	struct pool_cache_group *pcg, *npcg;
-	void *object;
+	pcg_t *pcg, *cur;
+	uint64_t ncsw;
+	pool_cache_t pc;
+
+	pc = cc->cc_cache;
+	cc->cc_misses++;
 
-	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, NULL);
-			if (pc->pc_dtor != NULL)
-				(*pc->pc_dtor)(pc->pc_arg, object);
-			pool_do_put(pc->pc_pool, object, pq);
-		}
-		pc->pc_ngroups--;
-		LIST_REMOVE(pcg, pcg_list);
-		LIST_INSERT_HEAD(pcgdl, pcg, pcg_list);
+	/*
+	 * No free slots locally.  Try to grab an empty, unused 
+	 * group from the cache.
+	 */
+	if (!mutex_tryenter(&pc->pc_lock)) {
+		ncsw = curlwp->l_ncsw;
+		mutex_enter(&pc->pc_lock);
+		pc->pc_contended++;
+
+		/*
+		 * If we context switched while locking, then
+		 * our view of the per-CPU data is invalid:
+		 * retry.
+		 */
+		if (curlwp->l_ncsw != ncsw) {
+			mutex_exit(&pc->pc_lock);
+			pool_cache_cpu_exit(cc, s);
+			return pool_cache_cpu_enter(pc, s);
+		}
 	}
-}
 
-static void
-pool_do_cache_invalidate(struct pool_cache *pc, struct pool_pagelist *pq,
-    struct pool_cache_grouplist *pcgl)
-{
+	if ((pcg = pc->pc_emptygroups) != NULL) {
+		/*
+		 * If there's a empty group, release our full
+		 * group back to the cache.  Install the empty
+		 * group as cc_current and return.
+		 */
+		if ((cur = cc->cc_current) != NULL) {
+			KASSERT(cur->pcg_avail == PCG_NOBJECTS);
+			cur->pcg_next = pc->pc_fullgroups;
+			pc->pc_fullgroups = cur;
+			pc->pc_nfull++;
+		}
+		KASSERT(pcg->pcg_avail == 0);
+		cc->cc_current = pcg;
+		pc->pc_emptygroups = pcg->pcg_next;
+		pc->pc_hits++;
+		pc->pc_nempty--;
+		mutex_exit(&pc->pc_lock);
+		return cc;
+	}
 
-	LOCK_ASSERT(simple_lock_held(&pc->pc_slock));
-	LOCK_ASSERT(simple_lock_held(&pc->pc_pool->pr_slock));
+	/*
+	 * Nothing available locally or in cache.  Take the
+	 * slow path and try to allocate a new group that we
+	 * can release to.
+	 */
+	pc->pc_misses++;
+	mutex_exit(&pc->pc_lock);
+	pool_cache_cpu_exit(cc, s);
 
-	pool_do_cache_invalidate_grouplist(&pc->pc_fullgroups, pc, pq, pcgl);
-	pool_do_cache_invalidate_grouplist(&pc->pc_partgroups, pc, pq, pcgl);
+	/*
+	 * If we can't allocate a new group, just throw the
+	 * object away.
+	 */
+	pcg = pool_get(&pcgpool, PR_NOWAIT);
+	if (pcg == NULL) {
+		pool_cache_destruct_object(pc, object);
+		return NULL;
+	}
+#ifdef DIAGNOSTIC
+	memset(pcg, 0, sizeof(*pcg));
+#else
+	pcg->pcg_avail = 0;
+#endif
 
-	KASSERT(LIST_EMPTY(&pc->pc_partgroups));
-	KASSERT(LIST_EMPTY(&pc->pc_fullgroups));
-	KASSERT(pc->pc_nitems == 0);
-}
+	/*
+	 * Add the empty group to the cache and try again.
+	 */
+	mutex_enter(&pc->pc_lock);
+	pcg->pcg_next = pc->pc_emptygroups;
+	pc->pc_emptygroups = pcg;
+	pc->pc_nempty++;
+	mutex_exit(&pc->pc_lock);
+
+	return pool_cache_cpu_enter(pc, s);
+} 
 
 /*
- * pool_cache_invalidate:
+ * pool_cache_put{,_paddr}:
  *
- *	Invalidate a pool cache (destruct and release all of the
- *	cached objects).
+ *	Put an object back to the pool cache (optionally caching the
+ *	physical address of the object).
  */
 void
-pool_cache_invalidate(struct pool_cache *pc)
+pool_cache_put_paddr(pool_cache_t pc, void *object, paddr_t pa)
 {
-	struct pool_pagelist pq;
-	struct pool_cache_grouplist pcgl;
-
-	LIST_INIT(&pq);
-	LIST_INIT(&pcgl);
+	pool_cache_cpu_t *cc;
+	pcg_t *pcg;
+	int s;
 
-	simple_lock(&pc->pc_slock);
-	simple_lock(&pc->pc_pool->pr_slock);
+	FREECHECK_IN(&pc->pc_freecheck, object);
 
-	pool_do_cache_invalidate(pc, &pq, &pcgl);
+	cc = pool_cache_cpu_enter(pc, &s);
+	do {
+		/* If the current group isn't full, release it there. */
+	 	pcg = cc->cc_current;
+		if (pcg != NULL && pcg->pcg_avail < PCG_NOBJECTS) {
+			KASSERT(pcg->pcg_objects[pcg->pcg_avail].pcgo_va
+			    == NULL);
+			pcg->pcg_objects[pcg->pcg_avail].pcgo_va = object;
+			pcg->pcg_objects[pcg->pcg_avail].pcgo_pa = pa;
+			pcg->pcg_avail++;
+			cc->cc_hits++;
+			pool_cache_cpu_exit(cc, &s);
+			return;
+		}
 
-	simple_unlock(&pc->pc_pool->pr_slock);
-	simple_unlock(&pc->pc_slock);
+		/*
+		 * That failed.  If the previous group is empty, swap
+		 * it with the current group and try again.
+		 */
+		pcg = cc->cc_previous;
+		if (pcg != NULL && pcg->pcg_avail == 0) {
+			cc->cc_previous = cc->cc_current;
+			cc->cc_current = pcg;
+			continue;
+		}
 
-	pr_pagelist_free(pc->pc_pool, &pq);
-	pcg_grouplist_free(&pcgl);
+		/*
+		 * Can't free to either group: try the slow path. 
+		 * If put_slow() releases the object for us, it
+		 * will return NULL.  Otherwise we need to retry.
+		 */
+		cc = pool_cache_put_slow(cc, &s, object, pa);
+	} while (cc != NULL);
 }
 
 /*
- * pool_cache_reclaim:
+ * pool_cache_xcall:
  *
- *	Reclaim a pool cache for pool_reclaim().
+ *	Transfer objects from the per-CPU cache to the global cache.
+ *	Run within a cross-call thread.
  */
 static void
-pool_cache_reclaim(struct pool_cache *pc, struct pool_pagelist *pq,
-    struct pool_cache_grouplist *pcgl)
+pool_cache_xcall(pool_cache_t pc)
 {
-
-	/*
-	 * 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.
+	pool_cache_cpu_t *cc;
+	pcg_t *prev, *cur, **list;
+	int s = 0; /* XXXgcc */
+
+	cc = pool_cache_cpu_enter(pc, &s);
+	cur = cc->cc_current;
+	cc->cc_current = NULL;
+	prev = cc->cc_previous;
+	cc->cc_previous = NULL;
+	pool_cache_cpu_exit(cc, &s);
+
+	/*
+	 * XXXSMP Go to splvm to prevent kernel_lock from being taken,
+	 * because locks at IPL_SOFTXXX are still spinlocks.  Does not
+	 * apply to IPL_SOFTBIO.  Cross-call threads do not take the
+	 * kernel_lock.
 	 */
-	if (simple_lock_try(&pc->pc_slock) == 0)
-		return;
-
-	pool_do_cache_invalidate(pc, pq, pcgl);
-
-	simple_unlock(&pc->pc_slock);
+	s = splvm();
+	mutex_enter(&pc->pc_lock);
+	if (cur != NULL) {
+		if (cur->pcg_avail == PCG_NOBJECTS) {
+			list = &pc->pc_fullgroups;
+			pc->pc_nfull++;
+		} else if (cur->pcg_avail == 0) {
+			list = &pc->pc_emptygroups;
+			pc->pc_nempty++;
+		} else {
+			list = &pc->pc_partgroups;
+			pc->pc_npart++;
+		}
+		cur->pcg_next = *list;
+		*list = cur;
+	}
+	if (prev != NULL) {
+		if (prev->pcg_avail == PCG_NOBJECTS) {
+			list = &pc->pc_fullgroups;
+			pc->pc_nfull++;
+		} else if (prev->pcg_avail == 0) {
+			list = &pc->pc_emptygroups;
+			pc->pc_nempty++;
+		} else {
+			list = &pc->pc_partgroups;
+			pc->pc_npart++;
+		}
+		prev->pcg_next = *list;
+		*list = prev;
+	}
+	mutex_exit(&pc->pc_lock);
+	splx(s);
 }
 
 /*
@@ -2177,10 +2728,12 @@ 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
 
@@ -2190,10 +2743,12 @@ void	pool_page_free_nointr(struct pool *
 #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
 
@@ -2203,6 +2758,7 @@ void	pool_subpage_free(struct pool *, vo
 
 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);
@@ -2210,131 +2766,43 @@ void	pool_subpage_free_nointr(struct poo
 
 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;
 
-	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;
-		}
-
+	res = (*pa->pa_alloc)(pp, flags);
+	if (res == NULL && (flags & PR_WAITOK) == 0) {
 		/*
-		 * Drain all pools, 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);
-			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;
-
-	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;
-	}
-
-	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;
+	bool waitok = (flags & PR_WAITOK) ? true : false;
 
 	return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok));
 }
@@ -2349,7 +2817,7 @@ pool_page_free(struct pool *pp, void *v)
 static void *
 pool_page_alloc_meta(struct pool *pp, int flags)
 {
-	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
+	bool waitok = (flags & PR_WAITOK) ? true : false;
 
 	return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok));
 }
@@ -2366,21 +2834,13 @@ pool_page_free_meta(struct pool *pp, voi
 void *
 pool_subpage_alloc(struct pool *pp, int flags)
 {
-	void *v;
-	int s;
-	s = splvm();
-	v = pool_get(&psppool, flags);
-	splx(s);
-	return v;
+	return pool_get(&psppool, flags);
 }
 
 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. */
@@ -2401,7 +2861,7 @@ pool_subpage_free_nointr(struct pool *pp
 void *
 pool_page_alloc_nointr(struct pool *pp, int flags)
 {
-	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
+	bool waitok = (flags & PR_WAITOK) ? true : false;
 
 	return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok));
 }