Add support for MAP_INHERIT_ZERO.

This provides a way for a process to designate pages in its address
space that should be replaced by fresh, zero-initialized anonymous
memory in forked child processes, rather than being copied or shared.

ok jmc, kettenis, tedu, deraadt; positive feedback from many more

4 files changed, 96 insertions, 46 deletions

diff --git a/lib/libc/sys/minherit.2 b/lib/libc/sys/minherit.2
index 05ea31c61a4..d6d0784e60c 100644
--- a/lib/libc/sys/minherit.2
+++ b/lib/libc/sys/minherit.2
@@ -1,4 +1,4 @@
-.\"	$OpenBSD: minherit.2,v 1.13 2007/05/31 19:19:33 jmc Exp $
+.\"	$OpenBSD: minherit.2,v 1.14 2014/06/13 01:48:51 matthew Exp $
 .\"
 .\" Copyright (c) 1991, 1993
 .\"	The Regents of the University of California.  All rights reserved.
@@ -29,7 +29,7 @@
 .\"
 .\"	@(#)minherit.2	8.1 (Berkeley) 6/9/93
 .\"
-.Dd $Mdocdate: May 31 2007 $
+.Dd $Mdocdate: June 13 2014 $
 .Dt MINHERIT 2
 .Os
 .Sh NAME
@@ -45,12 +45,25 @@ The
 .Fn minherit
 system call
 changes the specified pages to have the inheritance characteristic
-.Fa inherit ,
-which can be set to
-.Dv MAP_INHERIT_NONE ,
-.Dv MAP_INHERIT_COPY ,
-or
-.Dv MAP_INHERIT_SHARE .
+.Fa inherit .
+A page's inheritance characteristic controls how it will be mapped
+in child processes as created by
+.Xr fork 2 .
+.Pp
+The possible inheritance characteristics are:
+.Pp
+.Bl -tag -width MAP_INHERIT_SHARE -offset indent -compact
+.It Dv MAP_INHERIT_NONE
+Pages are not mapped in the child process.
+.It Dv MAP_INHERIT_COPY
+Private copy of pages are mapped in the child process.
+.It Dv MAP_INHERIT_SHARE
+Mapped pages are shared between the parent and child processes.
+.It Dv MAP_INHERIT_ZERO
+New anonymous pages (initialized to all zero bytes)
+are mapped in the child process.
+.El
+.Pp
 Not all implementations will guarantee that the inheritance characteristic
 can be set on a page basis;
 the granularity of changes may be as large as an entire region.
@@ -72,15 +85,10 @@ The virtual address range specified by the
 and
 .Fa len
 arguments is not valid.
-.It Bq Er EACCES
-The flags specified by the
+.It Bq Er EINVAL
+The
 .Fa inherit
-argument were not valid for the pages specified
-by the
-.Fa addr
-and
-.Fa len
-arguments.
+argument is invalid.
 .El
 .Sh SEE ALSO
 .Xr madvise 2 ,
diff --git a/sys/sys/mman.h b/sys/sys/mman.h
index cc4fca178a8..899fb7e8af3 100644
--- a/sys/sys/mman.h
+++ b/sys/sys/mman.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: mman.h,v 1.23 2013/05/31 18:41:26 tedu Exp $	*/
+/*	$OpenBSD: mman.h,v 1.24 2014/06/13 01:48:52 matthew Exp $	*/
 /*	$NetBSD: mman.h,v 1.11 1995/03/26 20:24:23 jtc Exp $	*/
 
 /*-
@@ -107,8 +107,7 @@
 #define MAP_INHERIT_SHARE	0	/* share with child */
 #define MAP_INHERIT_COPY	1	/* copy into child */
 #define MAP_INHERIT_NONE	2	/* absent from child */
-#define MAP_INHERIT_DONATE_COPY	3	/* copy and delete -- not
-					   implemented in UVM */
+#define MAP_INHERIT_ZERO	3	/* zero in child */
 
 /*
  * Flags to msync
diff --git a/sys/uvm/uvm_extern.h b/sys/uvm/uvm_extern.h
index 76b5946352f..42329bc9916 100644
--- a/sys/uvm/uvm_extern.h
+++ b/sys/uvm/uvm_extern.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: uvm_extern.h,v 1.115 2014/05/15 03:52:25 guenther Exp $	*/
+/*	$OpenBSD: uvm_extern.h,v 1.116 2014/06/13 01:48:52 matthew Exp $	*/
 /*	$NetBSD: uvm_extern.h,v 1.57 2001/03/09 01:02:12 chs Exp $	*/
 
 /*
@@ -134,7 +134,7 @@ typedef struct vm_page  *vm_page_t;
 #define UVM_INH_SHARE	0x00	/* "share" */
 #define UVM_INH_COPY	0x10	/* "copy" */
 #define UVM_INH_NONE	0x20	/* "none" */
-#define UVM_INH_DONATE	0x30	/* "donate" << not used */
+#define UVM_INH_ZERO	0x30	/* "zero" */
 
 /* 0x40, 0x80: not used */
 
diff --git a/sys/uvm/uvm_map.c b/sys/uvm/uvm_map.c
index 1648d40ec53..90cff04531b 100644
--- a/sys/uvm/uvm_map.c
+++ b/sys/uvm/uvm_map.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: uvm_map.c,v 1.168 2014/05/15 03:52:25 guenther Exp $	*/
+/*	$OpenBSD: uvm_map.c,v 1.169 2014/06/13 01:48:52 matthew Exp $	*/
 /*	$NetBSD: uvm_map.c,v 1.86 2000/11/27 08:40:03 chs Exp $	*/
 
 /*
@@ -190,10 +190,13 @@ int			 uvm_mapent_bias(struct vm_map*, struct vm_map_entry*);
 struct vm_map_entry	*uvm_mapent_clone(struct vm_map*, vaddr_t, vsize_t,
 			    vsize_t, struct vm_map_entry*,
 			    struct uvm_map_deadq*, int, int);
-void			 uvm_mapent_forkshared(struct vmspace*, struct vm_map*,
+struct vm_map_entry	*uvm_mapent_forkshared(struct vmspace*, struct vm_map*,
 			    struct vm_map*, struct vm_map_entry*,
 			    struct uvm_map_deadq*);
-void			 uvm_mapent_forkcopy(struct vmspace*, struct vm_map*,
+struct vm_map_entry	*uvm_mapent_forkcopy(struct vmspace*, struct vm_map*,
+			    struct vm_map*, struct vm_map_entry*,
+			    struct uvm_map_deadq*);
+struct vm_map_entry	*uvm_mapent_forkzero(struct vmspace*, struct vm_map*,
 			    struct vm_map*, struct vm_map_entry*,
 			    struct uvm_map_deadq*);
 
@@ -3158,7 +3161,7 @@ uvmspace_free(struct vmspace *vm)
  * Space must be available.
  * Reference counters are incremented.
  */
-struct vm_map_entry*
+struct vm_map_entry *
 uvm_mapent_clone(struct vm_map *dstmap, vaddr_t dstaddr, vsize_t dstlen,
     vsize_t off, struct vm_map_entry *old_entry, struct uvm_map_deadq *dead,
     int mapent_flags, int amap_share_flags)
@@ -3209,7 +3212,7 @@ uvm_mapent_clone(struct vm_map *dstmap, vaddr_t dstaddr, vsize_t dstlen,
  * share the mapping: this means we want the old and
  * new entries to share amaps and backing objects.
  */
-void
+struct vm_map_entry *
 uvm_mapent_forkshared(struct vmspace *new_vm, struct vm_map *new_map,
     struct vm_map *old_map,
     struct vm_map_entry *old_entry, struct uvm_map_deadq *dead)
@@ -3243,13 +3246,7 @@ uvm_mapent_forkshared(struct vmspace *new_vm, struct vm_map *new_map,
 		pmap_copy(new_map->pmap, old_map->pmap, new_entry->start,
 		    (new_entry->end - new_entry->start), new_entry->start);
 
-	/* Update process statistics. */
-	if (!UVM_ET_ISHOLE(new_entry))
-		new_map->size += new_entry->end - new_entry->start;
-	if (!UVM_ET_ISOBJ(new_entry) && !UVM_ET_ISHOLE(new_entry)) {
-		new_vm->vm_dused +=
-		    uvmspace_dused(new_map, new_entry->start, new_entry->end);
-	}
+	return (new_entry);
 }
 
 /*
@@ -3259,7 +3256,7 @@ uvm_mapent_forkshared(struct vmspace *new_vm, struct vm_map *new_map,
  * allocate new_entry, adjust reference counts.  
  * (note that new references are read-only).
  */
-void
+struct vm_map_entry *
 uvm_mapent_forkcopy(struct vmspace *new_vm, struct vm_map *new_map,
     struct vm_map *old_map,
     struct vm_map_entry *old_entry, struct uvm_map_deadq *dead)
@@ -3396,13 +3393,42 @@ uvm_mapent_forkcopy(struct vmspace *new_vm, struct vm_map *new_map,
 		}
 	}
 
-	/* Update process statistics. */
-	if (!UVM_ET_ISHOLE(new_entry))
-		new_map->size += new_entry->end - new_entry->start;
-	if (!UVM_ET_ISOBJ(new_entry) && !UVM_ET_ISHOLE(new_entry)) {
-		new_vm->vm_dused +=
-		    uvmspace_dused(new_map, new_entry->start, new_entry->end);
+	return (new_entry);
+}
+
+/*
+ * zero the mapping: the new entry will be zero initialized
+ */
+struct vm_map_entry *
+uvm_mapent_forkzero(struct vmspace *new_vm, struct vm_map *new_map,
+    struct vm_map *old_map,
+    struct vm_map_entry *old_entry, struct uvm_map_deadq *dead)
+{
+	struct vm_map_entry *new_entry;
+
+	new_entry = uvm_mapent_clone(new_map, old_entry->start,
+	    old_entry->end - old_entry->start, 0, old_entry,
+	    dead, 0, 0);
+
+	new_entry->etype |=
+	    (UVM_ET_COPYONWRITE|UVM_ET_NEEDSCOPY);
+
+	if (new_entry->aref.ar_amap) {
+		amap_unref(new_entry->aref.ar_amap, new_entry->aref.ar_pageoff,
+		    atop(new_entry->end - new_entry->start), 0);
+		new_entry->aref.ar_amap = NULL;
+		new_entry->aref.ar_pageoff = 0;
 	}
+
+	if (UVM_ET_ISOBJ(new_entry)) {
+		if (new_entry->object.uvm_obj->pgops->pgo_detach)
+			new_entry->object.uvm_obj->pgops->pgo_detach(
+			    new_entry->object.uvm_obj);
+		new_entry->object.uvm_obj = NULL;
+		new_entry->etype &= ~UVM_ET_OBJ;
+	}
+
+	return (new_entry);
 }
 
 /*
@@ -3418,7 +3444,7 @@ uvmspace_fork(struct process *pr)
 	struct vmspace *vm2;
 	struct vm_map *old_map = &vm1->vm_map;
 	struct vm_map *new_map;
-	struct vm_map_entry *old_entry;
+	struct vm_map_entry *old_entry, *new_entry;
 	struct uvm_map_deadq dead;
 
 	vm_map_lock(old_map);
@@ -3450,13 +3476,29 @@ uvmspace_fork(struct process *pr)
 		}
 
 		/* Apply inheritance. */
-		if (old_entry->inheritance == MAP_INHERIT_SHARE) {
-			uvm_mapent_forkshared(vm2, new_map,
+		switch (old_entry->inheritance) {
+		case MAP_INHERIT_SHARE:
+			new_entry = uvm_mapent_forkshared(vm2, new_map,
 			    old_map, old_entry, &dead);
-		}
-		if (old_entry->inheritance == MAP_INHERIT_COPY) {
-			uvm_mapent_forkcopy(vm2, new_map,
+			break;
+		case MAP_INHERIT_COPY:
+			new_entry = uvm_mapent_forkcopy(vm2, new_map,
+			    old_map, old_entry, &dead);
+			break;
+		case MAP_INHERIT_ZERO:
+			new_entry = uvm_mapent_forkzero(vm2, new_map,
 			    old_map, old_entry, &dead);
+			break;
+		default:
+			continue;
+		}
+
+	 	/* Update process statistics. */
+		if (!UVM_ET_ISHOLE(new_entry))
+			new_map->size += new_entry->end - new_entry->start;
+		if (!UVM_ET_ISOBJ(new_entry) && !UVM_ET_ISHOLE(new_entry)) {
+			vm2->vm_dused += uvmspace_dused(
+			    new_map, new_entry->start, new_entry->end);
 		}
 	}
 
@@ -3670,6 +3712,7 @@ uvm_map_inherit(struct vm_map *map, vaddr_t start, vaddr_t end,
 	case MAP_INHERIT_NONE:
 	case MAP_INHERIT_COPY:
 	case MAP_INHERIT_SHARE:
+	case MAP_INHERIT_ZERO:
 		break;
 	default:
 		return (EINVAL);