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 1.87.2.2
retrieving revision 1.115
diff -u -p -r1.87.2.2 -r1.115
--- src/sys/kern/subr_pool.c	2004/09/18 14:53:03	1.87.2.2
+++ src/sys/kern/subr_pool.c	2006/04/15 04:47:11	1.115
@@ -1,4 +1,4 @@
-/*	$NetBSD: subr_pool.c,v 1.87.2.2 2004/09/18 14:53:03 skrll Exp $	*/
+/*	$NetBSD: subr_pool.c,v 1.115 2006/04/15 04:47:11 christos Exp $	*/
 
 /*-
  * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc.
@@ -38,7 +38,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.87.2.2 2004/09/18 14:53:03 skrll Exp $");
+__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.115 2006/04/15 04:47:11 christos Exp $");
 
 #include "opt_pool.h"
 #include "opt_poollog.h"
@@ -70,16 +70,26 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,
  */
 
 /* 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 void *pool_page_alloc_meta(struct pool *, int);
+static void pool_page_free_meta(struct pool *, void *);
+
+/* allocator for pool metadata */
+static struct pool_allocator pool_allocator_meta = {
+	pool_page_alloc_meta, pool_page_free_meta
+};
+
 /* # of seconds to retain page after last use */
 int pool_inactive_time = 10;
 
@@ -89,17 +99,39 @@ 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 */
 	LIST_ENTRY(pool_item_header)
 				ph_pagelist;	/* pool page list */
-	TAILQ_HEAD(,pool_item)	ph_itemlist;	/* chunk list for this page */
 	SPLAY_ENTRY(pool_item_header)
 				ph_node;	/* Off-page page headers */
-	unsigned int		ph_nmissing;	/* # of chunks in use */
 	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 */
 };
+#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
@@ -107,7 +139,7 @@ struct pool_item {
 #endif
 #define	PI_MAGIC 0xdeadbeefU
 	/* Other entries use only this list entry */
-	TAILQ_ENTRY(pool_item)	pi_list;
+	LIST_ENTRY(pool_item)	pi_list;
 };
 
 #define	POOL_NEEDS_CATCHUP(pp)						\
@@ -142,17 +174,20 @@ 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 *);
 
+static int	pool_grow(struct pool *, int);
 void		*pool_allocator_alloc(struct pool *, int);
 void		pool_allocator_free(struct pool *, void *);
 
-static void pool_print_pagelist(struct pool_pagelist *,
+static void pool_print_pagelist(struct pool *, struct pool_pagelist *,
 	void (*)(const char *, ...));
 static void pool_print1(struct pool *, const char *,
 	void (*)(const char *, ...));
@@ -180,7 +215,7 @@ struct pool_log {
 
 int pool_logsize = POOL_LOGSIZE;
 
-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;
@@ -233,7 +268,7 @@ pr_printlog(struct pool *pp, struct pool
 	}
 }
 
-static __inline void
+static inline void
 pr_enter(struct pool *pp, const char *file, long line)
 {
 
@@ -249,7 +284,7 @@ pr_enter(struct pool *pp, const char *fi
 	pp->pr_entered_line = line;
 }
 
-static __inline void
+static inline void
 pr_leave(struct pool *pp)
 {
 
@@ -262,7 +297,7 @@ 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 *, ...))
 {
 
@@ -278,7 +313,51 @@ pr_enter_check(struct pool *pp, void (*p
 #define	pr_enter_check(pp, pr)
 #endif /* POOL_DIAGNOSTIC */
 
-static __inline int
+static inline int
+pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
+    const void *v)
+{
+	const char *cp = v;
+	int idx;
+
+	KASSERT(pp->pr_roflags & PR_NOTOUCH);
+	idx = (cp - ph->ph_page - ph->ph_off) / pp->pr_size;
+	KASSERT(idx < pp->pr_itemsperpage);
+	return idx;
+}
+
+#define	PR_FREELIST_ALIGN(p) \
+	roundup((uintptr_t)(p), sizeof(pool_item_freelist_t))
+#define	PR_FREELIST(ph)	((pool_item_freelist_t *)PR_FREELIST_ALIGN((ph) + 1))
+#define	PR_INDEX_USED	((pool_item_freelist_t)-1)
+#define	PR_INDEX_EOL	((pool_item_freelist_t)-2)
+
+static inline void
+pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph,
+    void *obj)
+{
+	int idx = pr_item_notouch_index(pp, ph, obj);
+	pool_item_freelist_t *freelist = PR_FREELIST(ph);
+
+	KASSERT(freelist[idx] == PR_INDEX_USED);
+	freelist[idx] = ph->ph_firstfree;
+	ph->ph_firstfree = idx;
+}
+
+static inline void *
+pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph)
+{
+	int idx = ph->ph_firstfree;
+	pool_item_freelist_t *freelist = PR_FREELIST(ph);
+
+	KASSERT(freelist[idx] != PR_INDEX_USED);
+	ph->ph_firstfree = freelist[idx];
+	freelist[idx] = PR_INDEX_USED;
+
+	return ph->ph_page + ph->ph_off + idx * pp->pr_size;
+}
+
+static inline int
 phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
 {
 	if (a->ph_page < b->ph_page)
@@ -295,7 +374,7 @@ SPLAY_GENERATE(phtree, pool_item_header,
 /*
  * Return the pool page header based on page address.
  */
-static __inline struct pool_item_header *
+static inline struct pool_item_header *
 pr_find_pagehead(struct pool *pp, caddr_t page)
 {
 	struct pool_item_header *ph, tmp;
@@ -308,16 +387,32 @@ pr_find_pagehead(struct pool *pp, caddr_
 	return ph;
 }
 
+static void
+pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq)
+{
+	struct pool_item_header *ph;
+	int s;
+
+	while ((ph = LIST_FIRST(pq)) != NULL) {
+		LIST_REMOVE(ph, ph_pagelist);
+		pool_allocator_free(pp, ph->ph_page);
+		if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
+			s = splvm();
+			pool_put(pp->pr_phpool, ph);
+			splx(s);
+		}
+	}
+}
+
 /*
  * 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) || pq != NULL);
+	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
 
 	/*
 	 * If the page was idle, decrement the idle page count.
@@ -335,21 +430,13 @@ 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.
 	 */
 	LIST_REMOVE(ph, ph_pagelist);
 	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
 		SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
-	if (pq) {
-		LIST_INSERT_HEAD(pq, ph, ph_pagelist);
-	} else {
-		pool_allocator_free(pp, ph->ph_page);
-		if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
-			s = splvm();
-			pool_put(&phpool, ph);
-			splx(s);
-		}
-	}
+	LIST_INSERT_HEAD(pq, ph, ph_pagelist);
+
 	pp->pr_npages--;
 	pp->pr_npagefree++;
 
@@ -385,6 +472,9 @@ pool_init(struct pool *pp, size_t size, 
 	size_t trysize, phsize;
 	int s;
 
+	KASSERT((1UL << (CHAR_BIT * sizeof(pool_item_freelist_t))) - 2 >=
+	    PHPOOL_FREELIST_NELEM(PHPOOL_MAX - 1));
+
 #ifdef POOL_DIAGNOSTIC
 	/*
 	 * Always log if POOL_DIAGNOSTIC is defined.
@@ -393,35 +483,19 @@ 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);
 
@@ -450,7 +524,7 @@ pool_init(struct pool *pp, size_t size, 
 	LIST_INIT(&pp->pr_emptypages);
 	LIST_INIT(&pp->pr_fullpages);
 	LIST_INIT(&pp->pr_partpages);
-	TAILQ_INIT(&pp->pr_cachelist);
+	LIST_INIT(&pp->pr_cachelist);
 	pp->pr_curpage = NULL;
 	pp->pr_npages = 0;
 	pp->pr_minitems = 0;
@@ -490,8 +564,9 @@ 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_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
-	    trysize / pp->pr_size == (trysize - phsize) / pp->pr_size) {
+	if ((pp->pr_roflags & PR_NOTOUCH) == 0 &&
+	    (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
+	    trysize / pp->pr_size == (trysize - phsize) / pp->pr_size)) {
 		/* Use the end of the page for the page header */
 		pp->pr_roflags |= PR_PHINPAGE;
 		pp->pr_phoffset = off = palloc->pa_pagesz - phsize;
@@ -509,6 +584,30 @@ pool_init(struct pool *pp, size_t size, 
 	 */
 	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
@@ -547,23 +646,35 @@ 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. */
@@ -580,20 +691,28 @@ 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;
 
-	/* Locking order: pool_allocator -> pool */
+	/* 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);
+
+	/* Remove this pool from its allocator's list of pools. */
 	s = splvm();
 	simple_lock(&pp->pr_alloc->pa_slock);
 	TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
 	simple_unlock(&pp->pr_alloc->pa_slock);
 	splx(s);
 
-	/* 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) {
@@ -603,19 +722,18 @@ pool_destroy(struct pool *pp)
 	}
 #endif
 
-	/* Remove all pages */
-	while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
-		pr_rmpage(pp, ph, NULL);
 	KASSERT(LIST_EMPTY(&pp->pr_fullpages));
 	KASSERT(LIST_EMPTY(&pp->pr_partpages));
 
-	/* Remove from global pool list */
-	simple_lock(&pool_head_slock);
-	TAILQ_REMOVE(&pool_head, pp, pr_poollist);
-	if (drainpp == pp) {
-		drainpp = NULL;
-	}
-	simple_unlock(&pool_head_slock);
+	/* Remove all pages */
+	LIST_INIT(&pq);
+	while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
+		pr_rmpage(pp, ph, &pq);
+
+	simple_unlock(&pp->pr_slock);
+	splx(s);
+
+	pr_pagelist_free(pp, &pq);
 
 #ifdef POOL_DIAGNOSTIC
 	if ((pp->pr_roflags & PR_LOGGING) != 0)
@@ -648,7 +766,7 @@ pool_alloc_item_header(struct pool *pp, 
 		ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
 	else {
 		s = splvm();
-		ph = pool_get(&phpool, flags);
+		ph = pool_get(pp->pr_phpool, flags);
 		splx(s);
 	}
 
@@ -677,11 +795,12 @@ pool_get(struct pool *pp, int flags)
 			    (flags & PR_WAITOK) != 0))
 		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)");
+	SCHED_ASSERT_UNLOCKED();
 #endif
-#endif /* DIAGNOSTIC */
 
 	simple_lock(&pp->pr_slock);
 	pr_enter(pp, file, line);
@@ -749,6 +868,8 @@ 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);
@@ -764,18 +885,9 @@ 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
@@ -795,58 +907,66 @@ pool_get(struct pool *pp, int flags)
 			/*
 			 * Wait for items to be returned to this pool.
 			 *
-			 * XXX: maybe we should wake up once a second and
-			 * try again?
+			 * 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. */
 			pr_leave(pp);
-			ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
+			ltsleep(pp, PSWP, pp->pr_wchan, hz, &pp->pr_slock);
 			pr_enter(pp, file, line);
-			goto startover;
 		}
 
-		/* 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;
 	}
-	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");
-	}
+		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) {
@@ -864,9 +984,10 @@ pool_get(struct pool *pp, int flags)
 		LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
 	}
 	ph->ph_nmissing++;
-	if (TAILQ_EMPTY(&ph->ph_itemlist)) {
+	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",
@@ -883,6 +1004,7 @@ pool_get(struct pool *pp, int flags)
 	}
 
 	pp->pr_nget++;
+	pr_leave(pp);
 
 	/*
 	 * If we have a low water mark and we are now below that low
@@ -896,7 +1018,6 @@ pool_get(struct pool *pp, int flags)
 		 */
 	}
 
-	pr_leave(pp);
 	simple_unlock(&pp->pr_slock);
 	return (v);
 }
@@ -905,7 +1026,7 @@ 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;
@@ -913,6 +1034,7 @@ pool_do_put(struct pool *pp, void *v)
 	int s;
 
 	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
+	SCHED_ASSERT_UNLOCKED();
 
 	page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
 
@@ -939,20 +1061,24 @@ 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++;
@@ -989,9 +1115,7 @@ pool_do_put(struct pool *pp, void *v)
 		if (pp->pr_npages > pp->pr_minpages &&
 		    (pp->pr_npages > pp->pr_maxpages ||
 		     (pp->pr_alloc->pa_flags & PA_WANT) != 0)) {
-			simple_unlock(&pp->pr_slock);
-			pr_rmpage(pp, ph, NULL);
-			simple_lock(&pp->pr_slock);
+			pr_rmpage(pp, ph, pq);
 		} else {
 			LIST_REMOVE(ph, ph_pagelist);
 			LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
@@ -1029,16 +1153,21 @@ 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 */
@@ -1046,12 +1175,15 @@ _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
@@ -1059,35 +1191,56 @@ 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 = NULL;
-	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);
-
-		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++;
 	}
 
@@ -1095,7 +1248,7 @@ pool_prime(struct pool *pp, int n)
 		pp->pr_maxpages = pp->pr_minpages + 1;	/* XXX */
 
 	simple_unlock(&pp->pr_slock);
-	return (0);
+	return error;
 }
 
 /*
@@ -1124,7 +1277,7 @@ pool_prime_page(struct pool *pp, caddr_t
 	 * Insert page header.
 	 */
 	LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
-	TAILQ_INIT(&ph->ph_itemlist);
+	LIST_INIT(&ph->ph_itemlist);
 	ph->ph_page = storage;
 	ph->ph_nmissing = 0;
 	s = splclock();
@@ -1154,17 +1307,28 @@ pool_prime_page(struct pool *pp, caddr_t
 	n = pp->pr_itemsperpage;
 	pp->pr_nitems += n;
 
-	while (n--) {
-		pi = (struct pool_item *)cp;
+	if (pp->pr_roflags & PR_NOTOUCH) {
+		pool_item_freelist_t *freelist = PR_FREELIST(ph);
+		int i;
+
+		ph->ph_off = cp - storage;
+		ph->ph_firstfree = 0;
+		for (i = 0; i < n - 1; i++)
+			freelist[i] = i + 1;
+		freelist[n - 1] = PR_INDEX_EOL;
+	} else {
+		while (n--) {
+			pi = (struct pool_item *)cp;
 
-		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);
+		}
 	}
 
 	/*
@@ -1189,34 +1353,15 @@ 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
@@ -1300,9 +1445,9 @@ pool_reclaim(struct pool *pp)
 {
 	struct pool_item_header *ph, *phnext;
 	struct pool_cache *pc;
-	struct timeval curtime;
 	struct pool_pagelist pq;
-	struct timeval diff;
+	struct pool_cache_grouplist pcgl;
+	struct timeval curtime, diff;
 	int s;
 
 	if (pp->pr_drain_hook != NULL) {
@@ -1317,12 +1462,13 @@ pool_reclaim(struct pool *pp)
 	pr_enter(pp, file, line);
 
 	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;
@@ -1353,20 +1499,11 @@ pool_reclaim(struct pool *pp)
 
 	pr_leave(pp);
 	simple_unlock(&pp->pr_slock);
-	if (LIST_EMPTY(&pq))
-		return (0);
-
-	while ((ph = LIST_FIRST(&pq)) != NULL) {
-		LIST_REMOVE(ph, ph_pagelist);
-		pool_allocator_free(pp, ph->ph_page);
-		if (pp->pr_roflags & PR_PHINPAGE) {
-			continue;
-		}
-		s = splvm();
-		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);
 }
 
@@ -1385,14 +1522,15 @@ 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);
 }
 
@@ -1417,9 +1555,24 @@ 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");
@@ -1438,16 +1591,14 @@ 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_pagelist *pl, void (*pr)(const char *, ...))
+pool_print_pagelist(struct pool *pp, struct pool_pagelist *pl,
+    void (*pr)(const char *, ...))
 {
 	struct pool_item_header *ph;
 #ifdef DIAGNOSTIC
@@ -1460,10 +1611,12 @@ pool_print_pagelist(struct pool_pagelist
 		    (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);
+		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
@@ -1507,13 +1660,13 @@ pool_print1(struct pool *pp, const char 
 
 	if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
 		(*pr)("\n\tempty page list:\n");
-	pool_print_pagelist(&pp->pr_emptypages, pr);
+	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->pr_fullpages, pr);
+	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->pr_partpages, pr);
+	pool_print_pagelist(pp, &pp->pr_partpages, pr);
 
 	if (pp->pr_curpage == NULL)
 		(*pr)("\tno current page\n");
@@ -1534,27 +1687,39 @@ pool_print1(struct pool *pp, const char 
 	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++) {
-				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);
-				}
-			}
+		(*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);
@@ -1579,9 +1744,12 @@ pool_chk_page(struct pool *pp, const cha
 		return 1;
 	}
 
-	for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
+	if ((pp->pr_roflags & PR_NOTOUCH) != 0)
+		return 0;
+
+	for (pi = LIST_FIRST(&ph->ph_itemlist), n = 0;
 	     pi != NULL;
-	     pi = TAILQ_NEXT(pi,pi_list), n++) {
+	     pi = LIST_NEXT(pi,pi_list), n++) {
 
 #ifdef DIAGNOSTIC
 		if (pi->pi_magic != PI_MAGIC) {
@@ -1658,11 +1826,11 @@ 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;
@@ -1677,7 +1845,7 @@ 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);
 }
 
@@ -1696,11 +1864,11 @@ 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 *
+static inline void *
 pcg_get(struct pool_cache_group *pcg, paddr_t *pap)
 {
 	void *object;
@@ -1719,7 +1887,7 @@ pcg_get(struct pool_cache_group *pcg, pa
 	return (object);
 }
 
-static __inline void
+static inline void
 pcg_put(struct pool_cache_group *pcg, void *object, paddr_t pa)
 {
 	u_int idx;
@@ -1732,6 +1900,20 @@ pcg_put(struct pool_cache_group *pcg, vo
 	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{,_paddr}:
  *
@@ -1751,13 +1933,15 @@ pool_cache_get_paddr(struct pool_cache *
 
 	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
@@ -1784,14 +1968,14 @@ pool_cache_get_paddr(struct pool_cache *
 		return (object);
 	}
 
- have_group:
 	pc->pc_hits++;
 	pc->pc_nitems--;
 	object = pcg_get(pcg, pap);
 
-	if (pcg->pcg_avail == 0)
-		pc->pc_allocfrom = NULL;
-
+	if (pcg->pcg_avail == 0) {
+		LIST_REMOVE(pcg, pcg_list);
+		LIST_INSERT_HEAD(&pc->pc_emptygroups, pcg, pcg_list);
+	}
 	simple_unlock(&pc->pc_slock);
 
 	return (object);
@@ -1809,15 +1993,21 @@ pool_cache_put_paddr(struct pool_cache *
 	struct pool_cache_group *pcg;
 	int s;
 
+	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
@@ -1827,31 +2017,29 @@ pool_cache_put_paddr(struct pool_cache *
 		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, pa);
 
-	if (pcg->pcg_avail == PCG_NOBJECTS)
-		pc->pc_freeto = NULL;
-
+	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);
 }
 
@@ -1870,44 +2058,45 @@ 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, NULL);
-			if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg)
-				pc->pc_allocfrom = NULL;
 			if (pc->pc_dtor != NULL)
 				(*pc->pc_dtor)(pc->pc_arg, object);
-			(*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:
  *
@@ -1917,10 +2106,22 @@ 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);
 }
 
 /*
@@ -1929,11 +2130,21 @@ 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);
 }
 
@@ -1953,23 +2164,42 @@ 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,
+};
+#else
 struct pool_allocator pool_allocator_kmem = {
 	pool_page_alloc, pool_page_free, 0,
 };
+#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,
+};
+#else
 struct pool_allocator pool_allocator_nointr = {
 	pool_page_alloc_nointr, pool_page_free_nointr, 0,
 };
+#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,
+};
+
+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,
 };
 #endif /* POOL_SUBPAGE */
 
@@ -2017,8 +2247,8 @@ pool_allocator_alloc(struct pool *org, i
 		}
 
 		/*
-		 * Drain all pools, except "org", that use this
-		 * allocator.  We do this to reclaim VA space.
+		 * Drain all pools, that use this allocator.
+		 * We do this to reclaim VA space.
 		 * pa_alloc is responsible for waiting for
 		 * physical memory.
 		 *
@@ -2039,8 +2269,6 @@ pool_allocator_alloc(struct pool *org, i
 		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);
@@ -2098,14 +2326,29 @@ 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
@@ -2132,32 +2375,30 @@ pool_subpage_free(struct pool *pp, void 
 
 /* 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 */