Index: vm/uma_core.c
===================================================================
--- vm/uma_core.c
+++ vm/uma_core.c
@@ -1197,8 +1197,6 @@
 keg_small_init(uma_keg_t keg)
 {
 	u_int rsize;
-	u_int memused;
-	u_int wastedspace;
 	u_int shsize;
 	u_int slabsize;
 
@@ -1231,16 +1229,13 @@
 
 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
 		shsize = 0;
-	else 
+	else {
 		shsize = sizeof(struct uma_slab);
+		if (shsize & UMA_ALIGN_PTR)
+			shsize = (shsize & ~UMA_ALIGN_PTR) +
+			    (UMA_ALIGN_PTR + 1);
+	}
 
-	keg->uk_ipers = (slabsize - shsize) / rsize;
-	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
-	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
-
-	memused = keg->uk_ipers * rsize + shsize;
-	wastedspace = slabsize - memused;
-
 	/*
 	 * We can't do OFFPAGE if we're internal or if we've been
 	 * asked to not go to the VM for buckets.  If we do this we
@@ -1249,35 +1244,43 @@
 	 * of UMA_ZONE_VM, which clearly forbids it.
 	 */
 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
-	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
-		return;
-
-	/*
-	 * See if using an OFFPAGE slab will limit our waste.  Only do
-	 * this if it permits more items per-slab.
-	 *
-	 * XXX We could try growing slabsize to limit max waste as well.
-	 * Historically this was not done because the VM could not
-	 * efficiently handle contiguous allocations.
-	 */
-	if ((wastedspace >= slabsize / UMA_MAX_WASTE) &&
-	    (keg->uk_ipers < (slabsize / keg->uk_rsize))) {
-		keg->uk_ipers = slabsize / keg->uk_rsize;
+	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) {
+		keg->uk_ipers = (slabsize - shsize) / rsize;
 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
-		CTR6(KTR_UMA, "UMA decided we need offpage slab headers for "
-		    "keg: %s(%p), calculated wastedspace = %d, "
-		    "maximum wasted space allowed = %d, "
-		    "calculated ipers = %d, "
-		    "new wasted space = %d\n", keg->uk_name, keg, wastedspace,
-		    slabsize / UMA_MAX_WASTE, keg->uk_ipers,
-		    slabsize - keg->uk_ipers * keg->uk_rsize);
-		keg->uk_flags |= UMA_ZONE_OFFPAGE;
+		return;
 	}
 
+	if (keg->uk_flags & UMA_ZONE_PCPU) {
+		KASSERT(keg->uk_flags & UMA_ZONE_OFFPAGE,
+		    ("%s: keg->uk_flags %u", __func__, keg->uk_flags));
+	} else {
+		/*
+		 * Grow the slab size to limit wasted memory.
+		 */
+		while (slabsize - rsize * (slabsize / rsize) > slabsize /
+		    UMA_MAX_WASTE) {
+			slabsize += UMA_SLAB_SIZE;
+			keg->uk_ppera++;
+		}
+		if (keg->uk_ppera > 1)
+			keg->uk_flags |= UMA_ZONE_VTOSLAB;
+	}
+	keg->uk_ipers = slabsize / rsize;
+	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
+	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
+
+	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0 &&
+	    slabsize - (keg->uk_rsize * keg->uk_ipers) < shsize)
+		keg->uk_flags |= UMA_ZONE_OFFPAGE;
+
 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
 		keg->uk_flags |= UMA_ZONE_HASH;
+
+	if (keg->uk_ppera > 1 && (keg->uk_flags & UMA_ZONE_PCPU) == 0)
+printf(" --> %s: uk_ppera=%d, uk_ipers=%d, uk_size=%d, uk_rsize=%d\n",
+    keg->uk_name, keg->uk_ppera, keg->uk_ipers, keg->uk_size, keg->uk_rsize);
 }
 
 /*
Index: vm/vm_page.c
===================================================================
--- vm/vm_page.c
+++ vm/vm_page.c
@@ -496,6 +496,14 @@
 	bzero((void *)mapped, end - new_end);
 	uma_startup((void *)mapped, boot_pages);
 
+	/*
+	 * Add a physical memory segment (vm_phys_seg) corresponding to the
+	 * preallocated boot pages so that vm_page structures representing
+	 * these pages will be created.  The vm_page structures are required
+	 * when a zone uses UMA_ZONE_VTOSLAB to find its slab headers.
+	 */
+	vm_phys_add_seg(new_end, end);
+
 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
     defined(__i386__) || defined(__mips__)
 	/*