Index: sys/amd64/amd64/pmap.c
===================================================================
--- sys/amd64/amd64/pmap.c
+++ sys/amd64/amd64/pmap.c
@@ -2261,6 +2261,8 @@
 		}
 	}
 	PMAP_UNLOCK(pmap);
+	if (pa == VM_PAGE_TO_PHYS(zero_page))
+		return (0);
 	return (pa);
 }
 
@@ -2304,7 +2306,10 @@
 				    &pa))
 					goto retry;
 				m = PHYS_TO_VM_PAGE(pte & PG_FRAME);
-				vm_page_hold(m);
+				if (m == zero_page)
+					m = NULL;
+				else
+					vm_page_hold(m);
 			}
 		}
 	}
Index: sys/i386/i386/pmap.c
===================================================================
--- sys/i386/i386/pmap.c
+++ sys/i386/i386/pmap.c
@@ -1502,6 +1502,8 @@
 		}
 	}
 	PMAP_UNLOCK(pmap);
+	if (rtval == VM_PAGE_TO_PHYS(zero_page))
+		return (0);
 	return (rtval);
 }
 
@@ -1545,7 +1547,10 @@
 				    &pa))
 					goto retry;
 				m = PHYS_TO_VM_PAGE(pte & PG_FRAME);
-				vm_page_hold(m);
+				if (m == zero_page)
+					m = NULL;
+				else
+					vm_page_hold(m);
 			}
 		}
 	}
Index: sys/vm/vm_fault.c
===================================================================
--- sys/vm/vm_fault.c
+++ sys/vm/vm_fault.c
@@ -730,6 +730,33 @@
 		 * page.  (Default objects are zero-fill, so there is no real
 		 * pager for them.)
 		 */
+
+		if (fs.object == fs.first_object && (fs.object->flags &
+		    (OBJ_ONEMAPPING | OBJ_NOSPLIT)) == OBJ_ONEMAPPING &&
+		    fs.object->shadow_count == 0 &&
+		    fs.object->backing_object == NULL &&
+		    fault_type == VM_PROT_READ && !wired && m_hold == NULL) {
+			/*
+			 * This is a lone nameless anonymous object
+			 * not participating in the shadow chains, and
+			 * the fault is for read.  Install transient
+			 * zero page mapping to avoid allocating real
+			 * physical memory until some user content is
+			 * written there.
+			 *
+			 * Write faults allocate and install the real
+			 * page.  Pmaps must not return zero_page from
+			 * pmap_extract() and pmap_extract_and_hold()
+			 * for this optimization to work.
+			 */
+			rv = pmap_enter(map->pmap, vaddr, zero_page,
+			    VM_PROT_READ, fault_type | PMAP_ENTER_NOSLEEP, 0);
+			if (rv == KERN_SUCCESS) {
+				unlock_and_deallocate(&fs);
+				return (rv);
+			}
+		}
+
 		if (fs.object->type != OBJT_DEFAULT ||
 		    fs.object == fs.first_object) {
 			if (fs.pindex >= fs.object->size) {
Index: sys/vm/vm_kern.c
===================================================================
--- sys/vm/vm_kern.c
+++ sys/vm/vm_kern.c
@@ -610,6 +610,8 @@
 	vm_map_unlock(map);
 }
 
+vm_page_t zero_page;
+
 void
 kmem_init_zero_region(void)
 {
@@ -622,7 +624,7 @@
 	 * zeros, while not using much more physical resources.
 	 */
 	addr = kva_alloc(ZERO_REGION_SIZE);
-	m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
+	zero_page = m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
 	    VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO);
 	if ((m->flags & PG_ZERO) == 0)
 		pmap_zero_page(m);
Index: sys/vm/vm_page.h
===================================================================
--- sys/vm/vm_page.h
+++ sys/vm/vm_page.h
@@ -370,6 +370,7 @@
 extern vm_page_t vm_page_array;		/* First resident page in table */
 extern long vm_page_array_size;		/* number of vm_page_t's */
 extern long first_page;			/* first physical page number */
+extern vm_page_t zero_page;
 
 #define VM_PAGE_TO_PHYS(entry)	((entry)->phys_addr)