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 1.101.2.3
retrieving revision 1.116
diff -u -p -r1.101.2.3 -r1.116
--- src/sys/kern/subr_pool.c	2007/02/26 09:11:15	1.101.2.3
+++ src/sys/kern/subr_pool.c	2006/04/15 14:23:11	1.116
@@ -1,7 +1,7 @@
-/*	$NetBSD: subr_pool.c,v 1.101.2.3 2007/02/26 09:11:15 yamt Exp $	*/
+/*	$NetBSD: subr_pool.c,v 1.116 2006/04/15 14:23:11 simonb Exp $	*/
 
 /*-
- * Copyright (c) 1997, 1999, 2000, 2002 The NetBSD Foundation, Inc.
+ * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc.
  * All rights reserved.
  *
  * This code is derived from software contributed to The NetBSD Foundation
@@ -38,7 +38,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.101.2.3 2007/02/26 09:11:15 yamt Exp $");
+__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.116 2006/04/15 14:23:11 simonb Exp $");
 
 #include "opt_pool.h"
 #include "opt_poollog.h"
@@ -53,7 +53,6 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,
 #include <sys/lock.h>
 #include <sys/pool.h>
 #include <sys/syslog.h>
-#include <sys/debug.h>
 
 #include <uvm/uvm.h>
 
@@ -83,16 +82,12 @@ static struct pool phpool[PHPOOL_MAX];
 static struct pool psppool;
 #endif
 
-static SLIST_HEAD(, pool_allocator) pa_deferinitq =
-    SLIST_HEAD_INITIALIZER(pa_deferinitq);
-
 static void *pool_page_alloc_meta(struct pool *, int);
 static void pool_page_free_meta(struct pool *, void *);
 
 /* allocator for pool metadata */
 static struct pool_allocator pool_allocator_meta = {
-	pool_page_alloc_meta, pool_page_free_meta,
-	.pa_backingmapptr = &kmem_map,
+	pool_page_alloc_meta, pool_page_free_meta
 };
 
 /* # of seconds to retain page after last use */
@@ -189,8 +184,8 @@ static void	pool_prime_page(struct pool 
 static void	pool_update_curpage(struct pool *);
 
 static int	pool_grow(struct pool *, int);
-static void	*pool_allocator_alloc(struct pool *, int);
-static void	pool_allocator_free(struct pool *, void *);
+void		*pool_allocator_alloc(struct pool *, int);
+void		pool_allocator_free(struct pool *, void *);
 
 static void pool_print_pagelist(struct pool *, struct pool_pagelist *,
 	void (*)(const char *, ...));
@@ -365,16 +360,10 @@ pr_item_notouch_get(const struct pool *p
 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);
 }
@@ -383,38 +372,18 @@ SPLAY_PROTOTYPE(phtree, pool_item_header
 SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
 
 /*
- * Return the pool page header based on item address.
+ * Return the pool page header based on page address.
  */
 static inline struct pool_item_header *
-pr_find_pagehead(struct pool *pp, void *v)
+pr_find_pagehead(struct pool *pp, caddr_t page)
 {
 	struct pool_item_header *ph, tmp;
 
-	if ((pp->pr_roflags & PR_NOALIGN) != 0) {
-		tmp.ph_page = (caddr_t)(uintptr_t)v;
-		ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
-		if (ph == NULL) {
-			ph = SPLAY_ROOT(&pp->pr_phtree);
-			if (ph != NULL && phtree_compare(&tmp, ph) >= 0) {
-				ph = SPLAY_NEXT(phtree, &pp->pr_phtree, ph);
-			}
-			KASSERT(ph == NULL || phtree_compare(&tmp, ph) < 0);
-		}
-	} else {
-		caddr_t page =
-		    (caddr_t)((uintptr_t)v & pp->pr_alloc->pa_pagemask);
-
-		if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
-			ph = (void *)(page + pp->pr_phoffset);
-		} else {
-			tmp.ph_page = page;
-			ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
-		}
-	}
+	if ((pp->pr_roflags & PR_PHINPAGE) != 0)
+		return ((struct pool_item_header *)(page + pp->pr_phoffset));
 
-	KASSERT(ph == NULL || ((pp->pr_roflags & PR_PHINPAGE) != 0) ||
-	    (ph->ph_page <= (char *)v &&
-	    (char *)v < ph->ph_page + pp->pr_alloc->pa_pagesz));
+	tmp.ph_page = page;
+	ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
 	return ph;
 }
 
@@ -474,86 +443,12 @@ 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
-pool_subsystem_init(void)
+link_pool_init(void)
 {
-	struct pool_allocator *pa;
 	__link_set_decl(pools, struct link_pool_init);
 	struct link_pool_init * const *pi;
 
@@ -561,13 +456,6 @@ pool_subsystem_init(void)
 		pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
 		    (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
 		    (*pi)->palloc);
-
-	while ((pa = SLIST_FIRST(&pa_deferinitq)) != NULL) {
-		KASSERT(pa->pa_backingmapptr != NULL);
-		KASSERT(*pa->pa_backingmapptr != NULL);
-		SLIST_REMOVE_HEAD(&pa_deferinitq, pa_q);
-		pa_reclaim_register(pa);
-	}
 }
 
 /*
@@ -628,23 +516,20 @@ pool_init(struct pool *pp, size_t size, 
 		simple_lock_init(&palloc->pa_slock);
 		palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
 		palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
-
-		if (palloc->pa_backingmapptr != NULL) {
-			pa_reclaim_register(palloc);
-		}
 		palloc->pa_flags |= PA_INITIALIZED;
 	}
 
 	if (align == 0)
 		align = ALIGN(1);
 
-	if ((flags & PR_NOTOUCH) == 0 && size < sizeof(struct pool_item))
+	if (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 (%zu) too large", size);
+		panic("pool_init: pool item size (%lu) too large",
+		      (u_long)size);
 #endif
 
 	/*
@@ -675,7 +560,6 @@ 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
@@ -694,7 +578,7 @@ 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 | PR_NOALIGN)) == 0 &&
+	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 */
@@ -813,7 +697,6 @@ pool_init(struct pool *pp, size_t size, 
 	TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
 	simple_unlock(&palloc->pa_slock);
 	splx(s);
-	pool_reclaim_register(pp);
 }
 
 /*
@@ -834,7 +717,6 @@ pool_destroy(struct pool *pp)
 	simple_unlock(&pool_head_slock);
 
 	/* Remove this pool from its allocator's list of pools. */
-	pool_reclaim_unregister(pp);
 	s = splvm();
 	simple_lock(&pp->pr_alloc->pa_slock);
 	TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
@@ -930,7 +812,8 @@ pool_get(struct pool *pp, int flags)
 #endif /* DIAGNOSTIC */
 #ifdef LOCKDEBUG
 	if (flags & PR_WAITOK)
-		ASSERT_SLEEPABLE(NULL, "pool_get(PR_WAITOK)");
+		simple_lock_only_held(NULL, "pool_get(PR_WAITOK)");
+	SCHED_ASSERT_UNLOCKED();
 #endif
 
 	simple_lock(&pp->pr_slock);
@@ -1028,10 +911,24 @@ pool_get(struct pool *pp, int flags)
 			if (pp->pr_curpage != NULL)
 				goto startover;
 
-			pp->pr_nfail++;
+			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. */
 			pr_leave(pp);
-			simple_unlock(&pp->pr_slock);
-			return (NULL);
+			ltsleep(pp, PSWP, pp->pr_wchan, hz, &pp->pr_slock);
+			pr_enter(pp, file, line);
 		}
 
 		/* Start the allocation process over. */
@@ -1136,8 +1033,6 @@ pool_get(struct pool *pp, int flags)
 	}
 
 	simple_unlock(&pp->pr_slock);
-	KASSERT((((vaddr_t)v + pp->pr_itemoffset) & (pp->pr_align - 1)) == 0);
-	FREECHECK_OUT(&pp->pr_freecheck, v);
 	return (v);
 }
 
@@ -1149,9 +1044,13 @@ 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));
-	FREECHECK_IN(&pp->pr_freecheck, v);
+	SCHED_ASSERT_UNLOCKED();
+
+	page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
 
 #ifdef DIAGNOSTIC
 	if (__predict_false(pp->pr_nout == 0)) {
@@ -1161,7 +1060,7 @@ pool_do_put(struct pool *pp, void *v, st
 	}
 #endif
 
-	if (__predict_false((ph = pr_find_pagehead(pp, v)) == NULL)) {
+	if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
 		pr_printlog(pp, NULL, printf);
 		panic("pool_put: %s: page header missing", pp->pr_wchan);
 	}
@@ -1229,7 +1128,7 @@ 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 ||
-		     pa_starved_p(pp->pr_alloc))) {
+		     (pp->pr_alloc->pa_flags & PA_WANT) != 0)) {
 			pr_rmpage(pp, ph, pq);
 		} else {
 			LIST_REMOVE(ph, ph_pagelist);
@@ -1241,7 +1140,9 @@ pool_do_put(struct pool *pp, void *v, st
 			 * be reclaimed by the pagedaemon.  This minimizes
 			 * ping-pong'ing for memory.
 			 */
-			getmicrotime(&ph->ph_time);
+			s = splclock();
+			ph->ph_time = mono_time;
+			splx(s);
 		}
 		pool_update_curpage(pp);
 	}
@@ -1374,15 +1275,15 @@ pool_prime_page(struct pool *pp, caddr_t
 {
 	struct pool_item *pi;
 	caddr_t cp = storage;
-	const unsigned int align = pp->pr_align;
-	const unsigned int ioff = pp->pr_itemoffset;
+	unsigned int align = pp->pr_align;
+	unsigned int ioff = pp->pr_itemoffset;
 	int n;
+	int s;
 
 	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
 
 #ifdef DIAGNOSTIC
-	if ((pp->pr_roflags & PR_NOALIGN) == 0 &&
-	    ((uintptr_t)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
+	if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
 		panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
 #endif
 
@@ -1393,7 +1294,9 @@ pool_prime_page(struct pool *pp, caddr_t
 	LIST_INIT(&ph->ph_itemlist);
 	ph->ph_page = storage;
 	ph->ph_nmissing = 0;
-	getmicrotime(&ph->ph_time);
+	s = splclock();
+	ph->ph_time = mono_time;
+	splx(s);
 	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
 		SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
 
@@ -1412,8 +1315,6 @@ pool_prime_page(struct pool *pp, caddr_t
 	if (ioff != 0)
 		cp = (caddr_t)(cp + (align - ioff));
 
-	KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
-
 	/*
 	 * Insert remaining chunks on the bucket list.
 	 */
@@ -1441,8 +1342,6 @@ pool_prime_page(struct pool *pp, caddr_t
 			pi->pi_magic = PI_MAGIC;
 #endif
 			cp = (caddr_t)(cp + pp->pr_size);
-
-			KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
 		}
 	}
 
@@ -1563,6 +1462,7 @@ pool_reclaim(struct pool *pp)
 	struct pool_pagelist pq;
 	struct pool_cache_grouplist pcgl;
 	struct timeval curtime, diff;
+	int s;
 
 	if (pp->pr_drain_hook != NULL) {
 		/*
@@ -1584,7 +1484,9 @@ pool_reclaim(struct pool *pp)
 	LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
 		pool_cache_reclaim(pc, &pq, &pcgl);
 
-	getmicrotime(&curtime);
+	s = splclock();
+	curtime = mono_time;
+	splx(s);
 
 	for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
 		phnext = LIST_NEXT(ph, ph_pagelist);
@@ -1595,8 +1497,7 @@ pool_reclaim(struct pool *pp)
 
 		KASSERT(ph->ph_nmissing == 0);
 		timersub(&curtime, &ph->ph_time, &diff);
-		if (diff.tv_sec < pool_inactive_time
-		    && !pa_starved_p(pp->pr_alloc))
+		if (diff.tv_sec < pool_inactive_time)
 			continue;
 
 		/*
@@ -1793,9 +1694,8 @@ 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)
@@ -1846,18 +1746,16 @@ pool_chk_page(struct pool *pp, const cha
 	caddr_t page;
 	int n;
 
-	if ((pp->pr_roflags & PR_NOALIGN) == 0) {
-		page = (caddr_t)((uintptr_t)ph & pp->pr_alloc->pa_pagemask);
-		if (page != ph->ph_page &&
-		    (pp->pr_roflags & PR_PHINPAGE) != 0) {
-			if (label != NULL)
-				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;
-		}
+	page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask);
+	if (page != ph->ph_page &&
+	    (pp->pr_roflags & PR_PHINPAGE) != 0) {
+		if (label != NULL)
+			printf("%s: ", label);
+		printf("pool(%p:%s): page inconsistency: page %p;"
+		       " at page head addr %p (p %p)\n", pp,
+			pp->pr_wchan, ph->ph_page,
+			ph, page);
+		return 1;
 	}
 
 	if ((pp->pr_roflags & PR_NOTOUCH) != 0)
@@ -1872,16 +1770,15 @@ 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\n",
+			       " page %p; item ordinal %d;"
+			       " addr %p (p %p)\n",
 				pp->pr_wchan, pi->pi_magic, ph->ph_page,
-				n, pi);
+				n, pi, page);
 			panic("pool");
 		}
 #endif
-		if ((pp->pr_roflags & PR_NOALIGN) != 0) {
-			continue;
-		}
-		page = (caddr_t)((uintptr_t)pi & pp->pr_alloc->pa_pagemask);
+		page =
+		    (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
 		if (page == ph->ph_page)
 			continue;
 
@@ -2045,7 +1942,7 @@ pool_cache_get_paddr(struct pool_cache *
 
 #ifdef LOCKDEBUG
 	if (flags & PR_WAITOK)
-		ASSERT_SLEEPABLE(NULL, "pool_cache_get(PR_WAITOK)");
+		simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)");
 #endif
 
 	simple_lock(&pc->pc_slock);
@@ -2075,8 +1972,6 @@ pool_cache_get_paddr(struct pool_cache *
 				return (NULL);
 			}
 		}
-		KASSERT((((vaddr_t)object + pc->pc_pool->pr_itemoffset) &
-		    (pc->pc_pool->pr_align - 1)) == 0);
 		if (object != NULL && pap != NULL) {
 #ifdef POOL_VTOPHYS
 			*pap = POOL_VTOPHYS(object);
@@ -2084,8 +1979,6 @@ pool_cache_get_paddr(struct pool_cache *
 			*pap = POOL_PADDR_INVALID;
 #endif
 		}
-
-		FREECHECK_OUT(&pc->pc_freecheck, object);
 		return (object);
 	}
 
@@ -2099,9 +1992,6 @@ pool_cache_get_paddr(struct pool_cache *
 	}
 	simple_unlock(&pc->pc_slock);
 
-	KASSERT((((vaddr_t)object + pc->pc_pool->pr_itemoffset) &
-	    (pc->pc_pool->pr_align - 1)) == 0);
-	FREECHECK_OUT(&pc->pc_freecheck, object);
 	return (object);
 }
 
@@ -2117,8 +2007,6 @@ pool_cache_put_paddr(struct pool_cache *
 	struct pool_cache_group *pcg;
 	int s;
 
-	FREECHECK_IN(&pc->pc_freecheck, object);
-
 	if (__predict_false((pc->pc_pool->pr_flags & PR_WANTED) != 0)) {
 		goto destruct;
 	}
@@ -2293,12 +2181,10 @@ 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
 
@@ -2308,12 +2194,10 @@ 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
 
@@ -2323,7 +2207,6 @@ 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);
@@ -2331,47 +2214,131 @@ 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 */
 
-static void *
-pool_allocator_alloc(struct pool *pp, int flags)
+/*
+ * We have at least three different resources for the same allocation and
+ * each resource can be depleted.  First, we have the ready elements in the
+ * pool.  Then we have the resource (typically a vm_map) for this allocator.
+ * Finally, we have physical memory.  Waiting for any of these can be
+ * unnecessary when any other is freed, but the kernel doesn't support
+ * sleeping on multiple wait channels, so we have to employ another strategy.
+ *
+ * The caller sleeps on the pool (so that it can be awakened when an item
+ * is returned to the pool), but we set PA_WANT on the allocator.  When a
+ * page is returned to the allocator and PA_WANT is set, pool_allocator_free
+ * will wake up all sleeping pools belonging to this allocator.
+ *
+ * XXX Thundering herd.
+ */
+void *
+pool_allocator_alloc(struct pool *org, int flags)
 {
-	struct pool_allocator *pa = pp->pr_alloc;
+	struct pool_allocator *pa = org->pr_alloc;
+	struct pool *pp, *start;
+	int s, freed;
 	void *res;
 
-	LOCK_ASSERT(!simple_lock_held(&pp->pr_slock));
+	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) {
 		/*
-		 * We only run the drain hook here if PR_NOWAIT.
-		 * In other cases, the hook will be run in
-		 * pool_reclaim().
+		 * 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?
 		 */
-		if (pp->pr_drain_hook != NULL) {
-			(*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
-			res = (*pa->pa_alloc)(pp, flags);
+		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;
 		}
-	}
-	return res;
+		simple_unlock(&pa->pa_slock);
+		splx(s);
+	} while (freed);
+	return (NULL);
 }
 
-static void
+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)
 {
-	bool waitok = (flags & PR_WAITOK) ? true : false;
+	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
 
 	return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok));
 }
@@ -2386,7 +2353,7 @@ pool_page_free(struct pool *pp, void *v)
 static void *
 pool_page_alloc_meta(struct pool *pp, int flags)
 {
-	bool waitok = (flags & PR_WAITOK) ? true : false;
+	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
 
 	return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok));
 }
@@ -2438,7 +2405,7 @@ pool_subpage_free_nointr(struct pool *pp
 void *
 pool_page_alloc_nointr(struct pool *pp, int flags)
 {
-	bool waitok = (flags & PR_WAITOK) ? true : false;
+	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
 
 	return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok));
 }