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 1.93.2.1
retrieving revision 1.99.8.1.2.1
diff -u -p -r1.93.2.1 -r1.99.8.1.2.1
--- src/sys/kern/subr_pool.c	2004/06/22 08:58:42	1.93.2.1
+++ src/sys/kern/subr_pool.c	2006/03/10 13:19:48	1.99.8.1.2.1
@@ -1,4 +1,4 @@
-/*	$NetBSD: subr_pool.c,v 1.93.2.1 2004/06/22 08:58:42 tron Exp $	*/
+/*	$NetBSD: subr_pool.c,v 1.99.8.1.2.1 2006/03/10 13:19:48 tron 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 1.93.2.1 2004/06/22 08:58:42 tron Exp $");
+__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.99.8.1.2.1 2006/03/10 13:19:48 tron Exp $");
 
 #include "opt_pool.h"
 #include "opt_poollog.h"
@@ -73,13 +73,23 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,
 TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head);
 
 /* Private pool for page header structures */
-static struct pool phpool;
+#define	PHPOOL_MAX	8
+static struct pool phpool[PHPOOL_MAX];
+#define	PHPOOL_FREELIST_NELEM(idx)	(((idx) == 0) ? 0 : (1 << (idx)))
 
 #ifdef POOL_SUBPAGE
 /* Pool of subpages for use by normal pools. */
 static struct pool psppool;
 #endif
 
+static void *pool_page_alloc_meta(struct pool *, int);
+static void pool_page_free_meta(struct pool *, void *);
+
+/* allocator for pool metadata */
+static struct pool_allocator pool_allocator_meta = {
+	pool_page_alloc_meta, pool_page_free_meta
+};
+
 /* # of seconds to retain page after last use */
 int pool_inactive_time = 10;
 
@@ -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 {
+			TAILQ_HEAD(, pool_item)
+				phu_itemlist;	/* chunk list for this page */
+		} phu_normal;
+		/* PR_NOTOUCH */
+		struct {
+			uint16_t
+				phu_off;	/* start offset in page */
+			pool_item_freelist_t
+				phu_firstfree;	/* first free item */
+			/*
+			 * XXX it might be better to use
+			 * a simple bitmap and ffs(3)
+			 */
+		} phu_notouch;
+	} ph_u;
+	uint16_t		ph_nmissing;	/* # of chunks in use */
 };
+#define	ph_itemlist	ph_u.phu_normal.phu_itemlist
+#define	ph_off		ph_u.phu_notouch.phu_off
+#define	ph_firstfree	ph_u.phu_notouch.phu_firstfree
 
 struct pool_item {
 #ifdef DIAGNOSTIC
@@ -142,7 +174,7 @@ 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 *);
 
 static int	pool_catchup(struct pool *);
 static void	pool_prime_page(struct pool *, caddr_t,
@@ -152,7 +184,7 @@ static void	pool_update_curpage(struct p
 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 *, ...));
@@ -279,6 +311,50 @@ pr_enter_check(struct pool *pp, void (*p
 #endif /* POOL_DIAGNOSTIC */
 
 static __inline int
+pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
+    const void *v)
+{
+	const char *cp = v;
+	int idx;
+
+	KASSERT(pp->pr_roflags & PR_NOTOUCH);
+	idx = (cp - ph->ph_page - ph->ph_off) / pp->pr_size;
+	KASSERT(idx < pp->pr_itemsperpage);
+	return idx;
+}
+
+#define	PR_FREELIST_ALIGN(p) \
+	roundup((uintptr_t)(p), sizeof(pool_item_freelist_t))
+#define	PR_FREELIST(ph)	((pool_item_freelist_t *)PR_FREELIST_ALIGN((ph) + 1))
+#define	PR_INDEX_USED	((pool_item_freelist_t)-1)
+#define	PR_INDEX_EOL	((pool_item_freelist_t)-2)
+
+static __inline void
+pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph,
+    void *obj)
+{
+	int idx = pr_item_notouch_index(pp, ph, obj);
+	pool_item_freelist_t *freelist = PR_FREELIST(ph);
+
+	KASSERT(freelist[idx] == PR_INDEX_USED);
+	freelist[idx] = ph->ph_firstfree;
+	ph->ph_firstfree = idx;
+}
+
+static __inline void *
+pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph)
+{
+	int idx = ph->ph_firstfree;
+	pool_item_freelist_t *freelist = PR_FREELIST(ph);
+
+	KASSERT(freelist[idx] != PR_INDEX_USED);
+	ph->ph_firstfree = freelist[idx];
+	freelist[idx] = PR_INDEX_USED;
+
+	return ph->ph_page + ph->ph_off + idx * pp->pr_size;
+}
+
+static __inline int
 phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
 {
 	if (a->ph_page < b->ph_page)
@@ -308,6 +384,23 @@ 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.
  */
@@ -315,9 +408,8 @@ 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 +427,13 @@ 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++;
 
@@ -357,6 +441,21 @@ pr_rmpage(struct pool *pp, struct pool_i
 }
 
 /*
+ * Initialize all the pools listed in the "pools" link set.
+ */
+void
+link_pool_init(void)
+{
+	__link_set_decl(pools, struct link_pool_init);
+	struct link_pool_init * const *pi;
+
+	__link_set_foreach(pi, pools)
+		pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
+		    (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
+		    (*pi)->palloc);
+}
+
+/*
  * Initialize the given pool resource structure.
  *
  * We export this routine to allow other kernel parts to declare
@@ -370,6 +469,9 @@ 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.
@@ -378,35 +480,19 @@ 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);
 
@@ -475,8 +561,9 @@ 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;
@@ -494,6 +581,30 @@ 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
@@ -532,18 +643,30 @@ 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. */
@@ -565,20 +688,28 @@ 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);
+	TAILQ_REMOVE(&pool_head, pp, pr_poollist);
+	if (drainpp == pp)
+		drainpp = NULL;
+	simple_unlock(&pool_head_slock);
+
+	/* Remove this pool from its allocator's list of pools. */
 	s = splvm();
 	simple_lock(&pp->pr_alloc->pa_slock);
 	TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
 	simple_unlock(&pp->pr_alloc->pa_slock);
 	splx(s);
 
-	/* 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(TAILQ_EMPTY(&pp->pr_cachelist));
 
 #ifdef DIAGNOSTIC
 	if (pp->pr_nout != 0) {
@@ -588,19 +719,18 @@ 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)
@@ -633,7 +763,7 @@ 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);
 	}
 
@@ -655,6 +785,9 @@ pool_get(struct pool *pp, int flags)
 	void *v;
 
 #ifdef DIAGNOSTIC
+	if (__predict_false(pp->pr_itemsperpage == 0))
+		panic("pool_get: pool %p: pr_itemsperpage is zero, "
+		    "pool not initialized?", pp);
 	if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
 			    (flags & PR_WAITOK) != 0))
 		panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
@@ -797,38 +930,53 @@ pool_get(struct pool *pp, int flags)
 		/* 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 = TAILQ_FIRST(&ph->ph_itemlist);
+		if (__predict_false(v == NULL)) {
+			pr_leave(pp);
+			simple_unlock(&pp->pr_slock);
+			panic("pool_get: %s: page empty", pp->pr_wchan);
+		}
+#ifdef DIAGNOSTIC
+		if (__predict_false(pp->pr_nitems == 0)) {
+			pr_leave(pp);
+			simple_unlock(&pp->pr_slock);
+			printf("pool_get: %s: items on itemlist, nitems %u\n",
+			    pp->pr_wchan, pp->pr_nitems);
+			panic("pool_get: nitems inconsistent");
+		}
 #endif
 
 #ifdef POOL_DIAGNOSTIC
-	pr_log(pp, v, PRLOG_GET, file, line);
+		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.
+		 */
+		TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
+	}
 	pp->pr_nitems--;
 	pp->pr_nout++;
 	if (ph->ph_nmissing == 0) {
@@ -846,9 +994,10 @@ 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 &&
+		    !TAILQ_EMPTY(&ph->ph_itemlist))) {
 			pr_leave(pp);
 			simple_unlock(&pp->pr_slock);
 			panic("pool_get: %s: nmissing inconsistent",
@@ -887,7 +1036,7 @@ 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;
@@ -921,20 +1070,24 @@ 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);
+		TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
+	}
 	KDASSERT(ph->ph_nmissing != 0);
 	ph->ph_nmissing--;
 	pp->pr_nput++;
@@ -971,9 +1124,7 @@ 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);
@@ -1011,16 +1162,22 @@ 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);
+
+	if (! LIST_EMPTY(&pq))
+		pr_pagelist_free(pp, &pq);
 }
 #undef pool_put
 #endif /* POOL_DIAGNOSTIC */
@@ -1028,12 +1185,16 @@ _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);
+
+	if (! LIST_EMPTY(&pq))
+		pr_pagelist_free(pp, &pq);
 }
 
 #ifdef POOL_DIAGNOSTIC
@@ -1136,17 +1297,28 @@ 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 */
+			TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
 #ifdef DIAGNOSTIC
-		pi->pi_magic = PI_MAGIC;
+			pi->pi_magic = PI_MAGIC;
 #endif
-		cp = (caddr_t)(cp + pp->pr_size);
+			cp = (caddr_t)(cp + pp->pr_size);
+		}
 	}
 
 	/*
@@ -1304,7 +1476,7 @@ pool_reclaim(struct pool *pp)
 	 * Reclaim items from the pool's caches.
 	 */
 	TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
-		pool_cache_reclaim(pc);
+		pool_cache_reclaim(pc, &pq);
 
 	s = splclock();
 	curtime = mono_time;
@@ -1338,17 +1510,7 @@ pool_reclaim(struct pool *pp)
 	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);
-	}
-
+	pr_pagelist_free(pp, &pq);
 	return (1);
 }
 
@@ -1429,7 +1591,8 @@ pool_printit(struct pool *pp, const char
 }
 
 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
@@ -1442,10 +1605,12 @@ 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)) {
+			TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) {
+				if (pi->pi_magic != PI_MAGIC) {
+					(*pr)("\t\t\titem %p, magic 0x%x\n",
+					    pi, pi->pi_magic);
+				}
 			}
 		}
 #endif
@@ -1489,13 +1654,13 @@ 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");
@@ -1561,6 +1726,9 @@ pool_chk_page(struct pool *pp, const cha
 		return 1;
 	}
 
+	if ((pp->pr_roflags & PR_NOTOUCH) != 0)
+		return 0;
+
 	for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
 	     pi != NULL;
 	     pi = TAILQ_NEXT(pi,pi_list), n++) {
@@ -1853,18 +2021,22 @@ pool_cache_destruct_object(struct pool_c
 }
 
 /*
- * pool_cache_do_invalidate:
+ * pool_cache_invalidate:
  *
- *	This internal function implements pool_cache_invalidate() and
- *	pool_cache_reclaim().
+ *	Invalidate a pool cache (destruct and release all of the
+ *	cached objects).
  */
-static void
-pool_cache_do_invalidate(struct pool_cache *pc, int free_groups,
-    void (*putit)(struct pool *, void *))
+void
+pool_cache_invalidate(struct pool_cache *pc)
 {
+	struct pool_pagelist pq;
 	struct pool_cache_group *pcg, *npcg;
 	void *object;
-	int s;
+
+	LIST_INIT(&pq);
+
+	simple_lock(&pc->pc_slock);
+	simple_lock(&pc->pc_pool->pr_slock);
 
 	for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
 	     pcg = npcg) {
@@ -1876,33 +2048,15 @@ pool_cache_do_invalidate(struct pool_cac
 				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);
 		}
 	}
-}
-
-/*
- * pool_cache_invalidate:
- *
- *	Invalidate a pool cache (destruct and release all of the
- *	cached objects).
- */
-void
-pool_cache_invalidate(struct pool_cache *pc)
-{
 
-	simple_lock(&pc->pc_slock);
-	pool_cache_do_invalidate(pc, 0, pool_put);
+	simple_unlock(&pc->pc_pool->pr_slock);
 	simple_unlock(&pc->pc_slock);
+
+	if (! LIST_EMPTY(&pq))
+		pr_pagelist_free(pc->pc_pool, &pq);
 }
 
 /*
@@ -1911,11 +2065,42 @@ 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_group *pcg, *npcg;
+	void *object;
+	int s;
+
+	/*
+	 * We're locking in the wrong order (normally pool_cache -> pool,
+	 * but the pool is already locked when we get here), so we have
+	 * to use trylock.  If we can't lock the pool_cache, it's not really
+	 * a big deal here.
+	 */
+	if (simple_lock_try(&pc->pc_slock) == 0)
+		return;
+
+	for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
+	     pcg = npcg) {
+		npcg = TAILQ_NEXT(pcg, pcg_list);
+		while (pcg->pcg_avail != 0) {
+			pc->pc_nitems--;
+			object = pcg_get(pcg, NULL);
+			if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg)
+				pc->pc_allocfrom = NULL;
+			if (pc->pc_dtor != NULL)
+				(*pc->pc_dtor)(pc->pc_arg, object);
+			pool_do_put(pc->pc_pool, object, pq);
+		}
+		pc->pc_ngroups--;
+		TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list);
+		if (pc->pc_freeto == pcg)
+			pc->pc_freeto = NULL;
+		s = splvm();
+		pool_put(&pcgpool, pcg);
+		splx(s);
+	}
 
-	simple_lock(&pc->pc_slock);
-	pool_cache_do_invalidate(pc, 1, pool_do_put);
 	simple_unlock(&pc->pc_slock);
 }
 
@@ -1935,23 +2120,42 @@ 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 */
 
@@ -2080,14 +2284,29 @@ 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, NULL, 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_poolpage1(kmem_map, NULL, waitok));
+}
+
+static void
+pool_page_free_meta(struct pool *pp, void *v)
+{
+
+	uvm_km_free_poolpage1(kmem_map, (vaddr_t) v);
 }
 
 #ifdef POOL_SUBPAGE
@@ -2114,25 +2333,25 @@ 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,
+	return ((void *) uvm_km_alloc_poolpage_cache(kernel_map,
 	    uvm.kernel_object, waitok));
 }
 
@@ -2140,6 +2359,5 @@ 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 */