Rework MAP_ANON|MAP_SHARED and shmat cap support

mmap(..., MAP_ANON|MAP_SHARED, ...):
    Require PROT_CAP explicitly to enable capability support in
    shared, anonymous mappings.  When specified, set the MAP_SHARECAP
    cow flag which causes a backing object to be allocated and the
    OBJ_SHARECAP flag set to allow sharing capabilities across address
    spaces.

shmat:
    Always set OBJ_SHARECAP on SysV shared memory objects.  Use of them
    is straightforwardly auditable.  We might want to add an explict
    SHM_SHARECAP flag at some point rather than making this universal,
    but shmat is probably best left in the dustbin of history.

Author: brooksdavis

sys/kern/sysv_shm.c

@@ -900,6 +900,7 @@ shmget_allocate_segment(struct thread *td, key_t key, size_t size, int mode)
 	}
 
 	vm_object_set_flag(shm_object, OBJ_HASCAP);
+	vm_object_set_flag(shm_object, OBJ_SHARECAP);
 	shmseg->object = shm_object;
 	shmseg->u.shm_perm.cuid = shmseg->u.shm_perm.uid = cred->cr_uid;
 	shmseg->u.shm_perm.cgid = shmseg->u.shm_perm.gid = cred->cr_gid;

sys/vm/vm_map.c

@@ -138,6 +138,7 @@ static uma_zone_t vmspace_zone;
 static int vmspace_zinit(void *mem, int size, int flags);
 static void _vm_map_init(vm_map_t map, pmap_t pmap, vm_pointer_t min,
     vm_pointer_t max);
+static inline void vm_map_entry_back(vm_map_entry_t entry);
 static void vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map);
 static void vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry);
 static void vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry);
@@ -2121,6 +2122,17 @@ vm_map_insert1(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
 	    ("overcommit: vm_map_insert leaks vm_map %p", new_entry));
 	new_entry->cred = cred;
 
+	if ((cow & MAP_SHARECAP) != 0) {
+		KASSERT(new_entry->inheritance == VM_INHERIT_SHARE,
+		    ("MAP_SHARECAP on unshared mapping"));
+		if (new_entry->object.vm_object == NULL)
+			vm_map_entry_back(new_entry);
+		VM_OBJECT_WLOCK(new_entry->object.vm_object);
+		vm_object_set_flag(new_entry->object.vm_object, OBJ_HASCAP);
+		vm_object_set_flag(new_entry->object.vm_object, OBJ_SHARECAP);
+		VM_OBJECT_WUNLOCK(new_entry->object.vm_object);
+	}
+
 	/*
 	 * Insert the new entry into the list
 	 */

sys/vm/vm_map.h

@@ -471,6 +471,7 @@ long vmspace_resident_count(struct vmspace *vmspace);
 #define	MAP_NO_HINT		0x00200000
 #define	MAP_CREATE_SHADOW	0x00400000
 #define	MAP_INHERIT_NONE	0x00800000
+#define	MAP_SHARECAP		0x01000000
 
 #define	MAP_SPLIT_BOUNDARY_SHIFT 19
 

sys/vm/vm_mmap.c

@@ -696,9 +696,12 @@ kern_mmap(struct thread *td, const struct mmap_req *mrp)
 		/*
 		 * Mapping blank space is trivial.
 		 */
+		if ((flags & MAP_SHARED) == 0) {
+			prot = VM_PROT_ADD_CAP(prot);
+			max_prot = VM_PROT_ADD_CAP(max_prot);
+		}
 		error = vm_mmap_object(&vms->vm_map, &addr, max_addr, size,
-		    VM_PROT_ADD_CAP(prot), VM_PROT_ADD_CAP(max_prot), flags,
-		    NULL, pos, FALSE, td);
+		    prot, max_prot, flags, NULL, pos, FALSE, td);
 	} else {
 		/*
 		 * Mapping file, get fp for validation and don't let the
@@ -2080,6 +2083,17 @@ vm_mmap_object(vm_map_t map, vm_pointer_t *addr, vm_offset_t max_addr,
 		docow |= MAP_CHECK_EXCL;
 	if ((flags & MAP_GUARD) != 0)
 		docow |= MAP_CREATE_GUARD;
+	/*
+	 * If we're creating a new shared resevation with anonymous
+	 * backing and capability permissions, let the vm system know so
+	 * the entry can be backed and OBJ_SHARECAP set.
+	 *
+	 * XXX: there should be a permission check...
+	 */
+	if ((flags & (MAP_ANON | MAP_SHARED | MAP_RESERVATION_CREATE)) ==
+	    (MAP_ANON | MAP_SHARED | MAP_RESERVATION_CREATE) &&
+	    (maxprot & (VM_PROT_READ_CAP | VM_PROT_WRITE_CAP)) != 0)
+		docow |= MAP_SHARECAP;
 
 	if (fitit) {
 		if ((flags & MAP_ALIGNMENT_MASK) == MAP_ALIGNED_SUPER)