Pool deals fairly well with physical memory shortage, but it doesn't deal

well (not at all) with shortages of the vm_map where the pages are mapped
(usually kmem_map).

Try to deal with it:
 - group all information the backend allocator for a pool in a separate
   struct. The pool will only have a pointer to that struct.
 - change the pool_init API to reflect that.
 - link all pools allocating from the same allocator on a linked list.
 - Since an allocator is responsible to wait for physical memory it will
   only fail (waitok) when it runs out of its backing vm_map, carefully
   drain pools using the same allocator so that va space is freed.
   (see comments in code for caveats and details).
 - change pool_reclaim to return if it actually succeeded to free some
   memory, use that information to make draining easier and more efficient.
 - get rid of PR_URGENT, noone uses it.

1 files changed, 206 insertions, 102 deletions

diff --git a/sys/kern/subr_pool.c b/sys/kern/subr_pool.c
index 18d1ca0a1a4..f8a42136896 100644
--- a/sys/kern/subr_pool.c
+++ b/sys/kern/subr_pool.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: subr_pool.c,v 1.19 2002/01/10 18:56:03 art Exp $	*/
+/*	$OpenBSD: subr_pool.c,v 1.20 2002/01/23 00:39:47 art Exp $	*/
 /*	$NetBSD: subr_pool.c,v 1.61 2001/09/26 07:14:56 chs Exp $	*/
 
 /*-
@@ -109,7 +109,7 @@ struct pool_item {
 };
 
 #define	PR_HASH_INDEX(pp,addr) \
-	(((u_long)(addr) >> (pp)->pr_pageshift) & (PR_HASHTABSIZE - 1))
+	(((u_long)(addr) >> (pp)->pr_alloc->pa_pageshift) & (PR_HASHTABSIZE - 1))
 
 #define	POOL_NEEDS_CATCHUP(pp)						\
 	((pp)->pr_nitems < (pp)->pr_minitems)
@@ -164,8 +164,9 @@ static void	pool_cache_reclaim(struct pool_cache *);
 static int	pool_catchup(struct pool *);
 static void	pool_prime_page(struct pool *, caddr_t,
 		    struct pool_item_header *);
-static void	*pool_page_alloc(unsigned long, int, int);
-static void	pool_page_free(void *, unsigned long, int);
+
+void *pool_allocator_alloc(struct pool *, int);
+void pool_allocator_free(struct pool *, void *);
 
 static void pool_print1(struct pool *, const char *,
 	int (*)(const char *, ...));
@@ -339,7 +340,7 @@ pr_rmpage(struct pool *pp, struct pool_item_header *ph,
 	if (pq) {
 		TAILQ_INSERT_HEAD(pq, ph, ph_pagelist);
 	} else {
-		(*pp->pr_free)(ph->ph_page, pp->pr_pagesz, pp->pr_mtype);
+		pool_allocator_free(pp, ph->ph_page);
 		if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
 			LIST_REMOVE(ph, ph_hashlist);
 			s = splhigh();
@@ -372,10 +373,7 @@ pr_rmpage(struct pool *pp, struct pool_item_header *ph,
  */
 void
 pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
-    const char *wchan, size_t pagesz,
-    void *(*alloc)(unsigned long, int, int),
-    void (*release)(void *, unsigned long, int),
-    int mtype)
+    const char *wchan, struct pool_allocator *palloc)
 {
 	int off, slack, i;
 
@@ -390,20 +388,19 @@ pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
 	/*
 	 * Check arguments and construct default values.
 	 */
-	if (!powerof2(pagesz))
-		panic("pool_init: page size invalid (%lx)\n", (u_long)pagesz);
-
-	if (alloc == NULL && release == NULL) {
-		alloc = pool_page_alloc;
-		release = pool_page_free;
-		pagesz = PAGE_SIZE;	/* Rounds to PAGE_SIZE anyhow. */
-	} else if ((alloc != NULL && release != NULL) == 0) {
-		/* If you specifiy one, must specify both. */
-		panic("pool_init: must specify alloc and release together");
+	if (palloc == NULL)
+		palloc = &pool_allocator_kmem;
+	if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
+		if (palloc->pa_pagesz == 0)
+			palloc->pa_pagesz = PAGE_SIZE;
+
+		TAILQ_INIT(&palloc->pa_list);
+		
+		simple_lock_init(&palloc->pa_slock);
+		palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
+		palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
+		palloc->pa_flags |= PA_INITIALIZED;
 	}
-			
-	if (pagesz == 0)
-		pagesz = PAGE_SIZE;
 
 	if (align == 0)
 		align = ALIGN(1);
@@ -412,9 +409,11 @@ pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
 		size = sizeof(struct pool_item);
 
 	size = ALIGN(size);
-	if (size > pagesz)
+#ifdef DIAGNOSTIC
+	if (size > palloc->pa_pagesz)
 		panic("pool_init: pool item size (%lu) too large",
 		      (u_long)size);
+#endif
 
 	/*
 	 * Initialize the pool structure.
@@ -431,12 +430,7 @@ pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
 	pp->pr_size = size;
 	pp->pr_align = align;
 	pp->pr_wchan = wchan;
-	pp->pr_mtype = mtype;
-	pp->pr_alloc = alloc;
-	pp->pr_free = release;
-	pp->pr_pagesz = pagesz;
-	pp->pr_pagemask = ~(pagesz - 1);
-	pp->pr_pageshift = ffs(pagesz) - 1;
+	pp->pr_alloc = palloc;
 	pp->pr_nitems = 0;
 	pp->pr_nout = 0;
 	pp->pr_hardlimit = UINT_MAX;
@@ -456,15 +450,15 @@ pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
 	 * with its header based on the page address.
 	 * We use 1/16 of the page size as the threshold (XXX: tune)
 	 */
-	if (pp->pr_size < pagesz/16) {
+	if (pp->pr_size < palloc->pa_pagesz/16) {
 		/* Use the end of the page for the page header */
 		pp->pr_roflags |= PR_PHINPAGE;
 		pp->pr_phoffset = off =
-			pagesz - ALIGN(sizeof(struct pool_item_header));
+			palloc->pa_pagesz - ALIGN(sizeof(struct pool_item_header));
 	} else {
 		/* The page header will be taken from our page header pool */
 		pp->pr_phoffset = 0;
-		off = pagesz;
+		off = palloc->pa_pagesz;
 		for (i = 0; i < PR_HASHTABSIZE; i++) {
 			LIST_INIT(&pp->pr_hashtab[i]);
 		}
@@ -520,15 +514,20 @@ pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
 	 */
 	if (phpool.pr_size == 0) {
 		pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
-		    0, "phpool", 0, 0, 0, 0);
+		    0, "phpool", NULL);
 		pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
-		    0, "pcgpool", 0, 0, 0, 0);
+		    0, "pcgpool", NULL);
 	}
 
 	/* Insert into the list of all pools. */
 	simple_lock(&pool_head_slock);
 	TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
 	simple_unlock(&pool_head_slock);
+
+	/* Insert into the list of pools using this allocator. */
+	simple_lock(&palloc->pa_slock);
+	TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
+	simple_unlock(&palloc->pa_slock);
 }
 
 /*
@@ -540,6 +539,13 @@ pool_destroy(struct pool *pp)
 	struct pool_item_header *ph;
 	struct pool_cache *pc;
 
+	/*
+	 * Locking order: pool_allocator -> pool
+	 */
+	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);
+
 	/* Destroy all caches for this pool. */
 	while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL)
 		pool_cache_destroy(pc);
@@ -661,9 +667,6 @@ pool_get(struct pool *pp, int flags)
 			      &pp->pr_hardlimit_ratecap))
 			log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
 
-		if (flags & PR_URGENT)
-			panic("pool_get: urgent");
-
 		pp->pr_nfail++;
 
 		pr_leave(pp);
@@ -694,7 +697,7 @@ pool_get(struct pool *pp, int flags)
 		 */
 		pr_leave(pp);
 		simple_unlock(&pp->pr_slock);
-		v = (*pp->pr_alloc)(pp->pr_pagesz, flags, pp->pr_mtype);
+		v = pool_allocator_alloc(pp, flags);
 		if (__predict_true(v != NULL))
 			ph = pool_alloc_item_header(pp, v, flags);
 		simple_lock(&pp->pr_slock);
@@ -702,7 +705,7 @@ pool_get(struct pool *pp, int flags)
 
 		if (__predict_false(v == NULL || ph == NULL)) {
 			if (v != NULL)
-				(*pp->pr_free)(v, pp->pr_pagesz, pp->pr_mtype);
+				pool_allocator_free(pp, v);
 
 			/*
 			 * We were unable to allocate a page or item
@@ -713,9 +716,6 @@ pool_get(struct pool *pp, int flags)
 			if (pp->pr_curpage != NULL)
 				goto startover;
 
-			if (flags & PR_URGENT)
-				panic("pool_get: urgent");
-
 			if ((flags & PR_WAITOK) == 0) {
 				pp->pr_nfail++;
 				pr_leave(pp);
@@ -726,15 +726,11 @@ pool_get(struct pool *pp, int flags)
 			/*
 			 * Wait for items to be returned to this pool.
 			 *
-			 * XXX: we actually want to wait just until
-			 * the page allocator has memory again. Depending
-			 * on this pool's usage, we might get stuck here
-			 * for a long time.
-			 *
 			 * XXX: maybe we should wake up once a second and
 			 * try again?
 			 */
 			pp->pr_flags |= PR_WANTED;
+			/* PA_WANTED is already set on the allocator */
 			pr_leave(pp);
 			ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
 			pr_enter(pp, file, line);
@@ -852,7 +848,7 @@ pool_do_put(struct pool *pp, void *v)
 
 	LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
 
-	page = (caddr_t)((u_long)v & pp->pr_pagemask);
+	page = (caddr_t)((vaddr_t)v & pp->pr_alloc->pa_pagemask);
 
 #ifdef DIAGNOSTIC
 	if (__predict_false(pp->pr_nout == 0)) {
@@ -1020,7 +1016,7 @@ pool_prime(struct pool *pp, int n)
 
 	while (newpages-- > 0) {
 		simple_unlock(&pp->pr_slock);
-		cp = (*pp->pr_alloc)(pp->pr_pagesz, PR_NOWAIT, pp->pr_mtype);
+		cp = pool_allocator_alloc(pp, PR_NOWAIT);
 		if (__predict_true(cp != NULL))
 			ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
 		simple_lock(&pp->pr_slock);
@@ -1028,7 +1024,7 @@ pool_prime(struct pool *pp, int n)
 		if (__predict_false(cp == NULL || ph == NULL)) {
 			error = ENOMEM;
 			if (cp != NULL)
-				(*pp->pr_free)(cp, pp->pr_pagesz, pp->pr_mtype);
+				pool_allocator_free(pp, cp);
 			break;
 		}
 
@@ -1058,8 +1054,10 @@ pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
 	unsigned int ioff = pp->pr_itemoffset;
 	int n;
 
-	if (((u_long)cp & (pp->pr_pagesz - 1)) != 0)
+#ifdef DIAGNOSTIC
+	if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
 		panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
+#endif
 
 	if ((pp->pr_roflags & PR_PHINPAGE) == 0)
 		LIST_INSERT_HEAD(&pp->pr_hashtab[PR_HASH_INDEX(pp, cp)],
@@ -1154,13 +1152,13 @@ pool_catchup(struct pool *pp)
 		 * the pool descriptor?
 		 */
 		simple_unlock(&pp->pr_slock);
-		cp = (*pp->pr_alloc)(pp->pr_pagesz, PR_NOWAIT, pp->pr_mtype);
+		cp = pool_allocator_alloc(pp, PR_NOWAIT);
 		if (__predict_true(cp != NULL))
 			ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
 		simple_lock(&pp->pr_slock);
 		if (__predict_false(cp == NULL || ph == NULL)) {
 			if (cp != NULL)
-				(*pp->pr_free)(cp, pp->pr_pagesz, pp->pr_mtype);
+				pool_allocator_free(pp, cp);
 			error = ENOMEM;
 			break;
 		}
@@ -1232,48 +1230,11 @@ pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
 }
 
 /*
- * Default page allocator.
- */
-static void *
-pool_page_alloc(unsigned long sz, int flags, int mtype)
-{
-	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
-
-	return ((void *)uvm_km_alloc_poolpage(waitok));
-}
-
-static void
-pool_page_free(void *v, unsigned long sz, int mtype)
-{
-
-	uvm_km_free_poolpage((vaddr_t)v);
-}
-
-/*
- * Alternate pool page allocator for pools that know they will
- * never be accessed in interrupt context.
- */
-void *
-pool_page_alloc_nointr(unsigned long sz, int flags, int mtype)
-{
-	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
-
-	return ((void *)uvm_km_alloc_poolpage1(kernel_map, uvm.kernel_object,
-	    waitok));
-}
-
-void
-pool_page_free_nointr(void *v, unsigned long sz, int mtype)
-{
-
-	uvm_km_free_poolpage1(kernel_map, (vaddr_t)v);
-}
-
-
-/*
  * Release all complete pages that have not been used recently.
+ *
+ * Returns non-zero if any pages have been reclaimed.
  */
-void
+int
 #ifdef POOL_DIAGNOSTIC
 _pool_reclaim(struct pool *pp, const char *file, long line)
 #else
@@ -1287,10 +1248,10 @@ pool_reclaim(struct pool *pp)
 	int s;
 
 	if (pp->pr_roflags & PR_STATIC)
-		return;
+		return 0;
 
 	if (simple_lock_try(&pp->pr_slock) == 0)
-		return;
+		return 0;
 	pr_enter(pp, file, line);
 	TAILQ_INIT(&pq);
 
@@ -1332,11 +1293,11 @@ pool_reclaim(struct pool *pp)
 	pr_leave(pp);
 	simple_unlock(&pp->pr_slock);
 	if (TAILQ_EMPTY(&pq)) {
-		return;
+		return 0;
 	}
 	while ((ph = TAILQ_FIRST(&pq)) != NULL) {
 		TAILQ_REMOVE(&pq, ph, ph_pagelist);
-		(*pp->pr_free)(ph->ph_page, pp->pr_pagesz, pp->pr_mtype);
+		pool_allocator_free(pp, ph->ph_page);
 		if (pp->pr_roflags & PR_PHINPAGE) {
 			continue;
 		}
@@ -1345,6 +1306,8 @@ pool_reclaim(struct pool *pp)
 		pool_put(&phpool, ph);
 		splx(s);
 	}
+
+	return 1;
 }
 
 
@@ -1374,7 +1337,6 @@ pool_drain(void *arg)
 	splx(s);
 }
 
-
 /*
  * Diagnostic helpers.
  */
@@ -1420,8 +1382,7 @@ pool_print1(struct pool *pp, const char *modif, int (*pr)(const char *, ...))
 	(*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
 	    pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
 	    pp->pr_roflags);
-	(*pr)("\tpagesz %u, mtype %d\n", pp->pr_pagesz, pp->pr_mtype);
-	(*pr)("\talloc %p, release %p\n", pp->pr_alloc, pp->pr_free);
+	(*pr)("\talloc %p\n", pp->pr_alloc);
 	(*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
 	    pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
 	(*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
@@ -1502,7 +1463,7 @@ pool_chk(struct pool *pp, const char *label)
 		int n;
 		caddr_t page;
 
-		page = (caddr_t)((u_long)ph & pp->pr_pagemask);
+		page = (caddr_t)((vaddr_t)ph & pp->pr_alloc->pa_pagemask);
 		if (page != ph->ph_page &&
 		    (pp->pr_roflags & PR_PHINPAGE) != 0) {
 			if (label != NULL)
@@ -1531,7 +1492,7 @@ pool_chk(struct pool *pp, const char *label)
 				panic("pool");
 			}
 #endif
-			page = (caddr_t)((u_long)pi & pp->pr_pagemask);
+			page = (caddr_t)((vaddr_t)pi & pp->pr_alloc->pa_pagemask);
 			if (page == ph->ph_page)
 				continue;
 
@@ -1899,3 +1860,146 @@ sysctl_dopool(int *name, u_int namelen, char *where, size_t *sizep)
 	/* NOTREACHED */
 	return (0); /* XXX - Stupid gcc */
 }
+
+/*
+ * Pool backend allocators.
+ *
+ * Each pool has a backend allocator that handles allocation, deallocation
+ * and any additional draining that might be needed.
+ *
+ * We provide two standard allocators.
+ *  pool_alloc_kmem - the default used when no allocator is specified.
+ *  pool_alloc_nointr - used for pools that will not be accessed in
+ *   interrupt context.
+ */
+void	*pool_page_alloc(struct pool *, int);
+void	pool_page_free(struct pool *, void *);
+void	*pool_page_alloc_nointr(struct pool *, int);
+void	pool_page_free_nointr(struct pool *, void *);
+
+struct pool_allocator pool_allocator_kmem = {
+	pool_page_alloc, pool_page_free, 0,
+};
+struct pool_allocator pool_allocator_nointr = {
+	pool_page_alloc_nointr, pool_page_free_nointr, 0,
+};
+
+/*
+ * XXX - 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, then 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 addresses, so we have to fake. The caller sleeps on
+ *  the pool (so that we 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 wakeup all
+ *  sleeping pools belonging to this allocator. (XXX - thundering herd).
+ */
+
+void *
+pool_allocator_alloc(struct pool *org, int flags)
+{
+	struct pool_allocator *pa = org->pr_alloc;
+	struct pool *pp, *start;
+	int s, freed;
+	void *res;
+
+	do {
+		if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
+			return (res);
+		if ((flags & PR_WAITOK) == 0)
+			break;
+
+		/*
+		 * Drain all pools, except 'org', that use this allocator.
+		 * We do this to reclaim va space. pa_alloc is responsible
+		 * for waiting for physical memory.
+		 * XXX - we risk looping forever if start if someone calls
+		 *  pool_destroy on 'start'. But there is no other way to
+		 *  have potentially sleeping pool_reclaim, non-sleeping
+		 *  locks on pool_allocator and some stirring of drained
+		 *  pools in the allocator.
+		 */
+		freed = 0;
+
+		s = splvm();
+		simple_lock(&pa->pa_slock);
+		pp = start = TAILQ_FIRST(&pa->pa_list);
+		do {
+			TAILQ_REMOVE(&pa->pa_list, pp, pr_alloc_list);
+			TAILQ_INSERT_TAIL(&pa->pa_list, pp, pr_alloc_list);
+			if (pp == org)
+				continue;
+			simple_unlock(&pa->pa_list);
+			freed = pool_reclaim(pp)
+			simple_lock(&pa->pa_list);
+		} while ((pp = TAILQ_FIRST(&pa->pa_list)) != start && !freed);
+
+		if (!freed) {
+			/*
+			 * 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;
+		}
+		simple_unlock(&pa->pa_slock);
+		splx(s);
+	} while (freed);
+	return (NULL);
+}
+
+void
+pool_allocator_free(struct pool *pp, void *v)
+{
+	struct pool_allocator *pa = pp->pr_alloc;
+
+	(*pa->pa_free)(pp, v);
+
+	simple_lock(&pa->pa_slock);
+	if ((pa->pa_flags & PA_WANT) == 0) {
+		simple_unlock(&pa->pa_slock);
+		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);
+		}
+	}
+	pa->pa_flags &= ~PA_WANT;
+	simple_unlock(&pa->pa_slock);
+}
+
+void *
+pool_page_alloc(struct pool *pp, int flags)
+{
+	boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
+
+	return ((void *)uvm_km_alloc_poolpage(waitok));
+}
+
+void
+pool_page_free(struct pool *pp, void *v)
+{
+	uvm_km_free_poolpage((vaddr_t)v);
+}
+
+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));
+}
+
+void
+pool_page_free_nointr(struct pool *pp, void *v)
+{
+	uvm_km_free_poolpage1(kernel_map, (vaddr_t)v);
+}