Fix the regression created by "set S_DEAD on unlink()..." commit
Al Viro [Fri, 30 Apr 2010 21:17:09 +0000 (17:17 -0400)]
1) i_flags simply doesn't work for mount/unlink race prevention;
we may have many links to file and rm on one of those obviously
shouldn't prevent bind on top of another later on.  To fix it
right way we need to mark _dentry_ as unsuitable for mounting
upon; new flag (DCACHE_CANT_MOUNT) is protected by d_flags and
i_mutex on the inode in question.  Set it (with dont_mount(dentry))
in unlink/rmdir/etc., check (with cant_mount(dentry)) in places
in namespace.c that used to check for S_DEAD.  Setting S_DEAD
is still needed in places where we used to set it (for directories
getting killed), since we rely on it for readdir/rmdir race
prevention.

2) rename()/mount() protection has another bogosity - we unhash
the target before we'd checked that it's not a mountpoint.  Fixed.

3) ancient bogosity in pivot_root() - we locked i_mutex on the
right directory, but checked S_DEAD on the different (and wrong)
one.  Noticed and fixed.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

drivers/usb/core/inode.c
fs/configfs/dir.c
fs/namei.c
fs/namespace.c
include/linux/dcache.h

index 4a6366a..111a01a 100644 (file)
@@ -380,6 +380,7 @@ static int usbfs_rmdir(struct inode *dir, struct dentry *dentry)
        mutex_lock(&inode->i_mutex);
        dentry_unhash(dentry);
        if (usbfs_empty(dentry)) {
+               dont_mount(dentry);
                drop_nlink(dentry->d_inode);
                drop_nlink(dentry->d_inode);
                dput(dentry);
index 8e48b52..0b502f8 100644 (file)
@@ -645,6 +645,7 @@ static void detach_groups(struct config_group *group)
 
                configfs_detach_group(sd->s_element);
                child->d_inode->i_flags |= S_DEAD;
+               dont_mount(child);
 
                mutex_unlock(&child->d_inode->i_mutex);
 
@@ -840,6 +841,7 @@ static int configfs_attach_item(struct config_item *parent_item,
                        mutex_lock(&dentry->d_inode->i_mutex);
                        configfs_remove_dir(item);
                        dentry->d_inode->i_flags |= S_DEAD;
+                       dont_mount(dentry);
                        mutex_unlock(&dentry->d_inode->i_mutex);
                        d_delete(dentry);
                }
@@ -882,6 +884,7 @@ static int configfs_attach_group(struct config_item *parent_item,
                if (ret) {
                        configfs_detach_item(item);
                        dentry->d_inode->i_flags |= S_DEAD;
+                       dont_mount(dentry);
                }
                configfs_adjust_dir_dirent_depth_after_populate(sd);
                mutex_unlock(&dentry->d_inode->i_mutex);
@@ -1725,6 +1728,7 @@ void configfs_unregister_subsystem(struct configfs_subsystem *subsys)
        mutex_unlock(&configfs_symlink_mutex);
        configfs_detach_group(&group->cg_item);
        dentry->d_inode->i_flags |= S_DEAD;
+       dont_mount(dentry);
        mutex_unlock(&dentry->d_inode->i_mutex);
 
        d_delete(dentry);
index 16df727..b86b96f 100644 (file)
@@ -2176,8 +2176,10 @@ int vfs_rmdir(struct inode *dir, struct dentry *dentry)
                error = security_inode_rmdir(dir, dentry);
                if (!error) {
                        error = dir->i_op->rmdir(dir, dentry);
-                       if (!error)
+                       if (!error) {
                                dentry->d_inode->i_flags |= S_DEAD;
+                               dont_mount(dentry);
+                       }
                }
        }
        mutex_unlock(&dentry->d_inode->i_mutex);
@@ -2261,7 +2263,7 @@ int vfs_unlink(struct inode *dir, struct dentry *dentry)
                if (!error) {
                        error = dir->i_op->unlink(dir, dentry);
                        if (!error)
-                               dentry->d_inode->i_flags |= S_DEAD;
+                               dont_mount(dentry);
                }
        }
        mutex_unlock(&dentry->d_inode->i_mutex);
@@ -2572,17 +2574,20 @@ static int vfs_rename_dir(struct inode *old_dir, struct dentry *old_dentry,
                return error;
 
        target = new_dentry->d_inode;
-       if (target) {
+       if (target)
                mutex_lock(&target->i_mutex);
-               dentry_unhash(new_dentry);
-       }
        if (d_mountpoint(old_dentry)||d_mountpoint(new_dentry))
                error = -EBUSY;
-       else 
+       else {
+               if (target)
+                       dentry_unhash(new_dentry);
                error = old_dir->i_op->rename(old_dir, old_dentry, new_dir, new_dentry);
+       }
        if (target) {
-               if (!error)
+               if (!error) {
                        target->i_flags |= S_DEAD;
+                       dont_mount(new_dentry);
+               }
                mutex_unlock(&target->i_mutex);
                if (d_unhashed(new_dentry))
                        d_rehash(new_dentry);
@@ -2614,7 +2619,7 @@ static int vfs_rename_other(struct inode *old_dir, struct dentry *old_dentry,
                error = old_dir->i_op->rename(old_dir, old_dentry, new_dir, new_dentry);
        if (!error) {
                if (target)
-                       target->i_flags |= S_DEAD;
+                       dont_mount(new_dentry);
                if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE))
                        d_move(old_dentry, new_dentry);
        }
index 8174c8a..f20cb57 100644 (file)
@@ -1432,7 +1432,7 @@ static int graft_tree(struct vfsmount *mnt, struct path *path)
 
        err = -ENOENT;
        mutex_lock(&path->dentry->d_inode->i_mutex);
-       if (IS_DEADDIR(path->dentry->d_inode))
+       if (cant_mount(path->dentry))
                goto out_unlock;
 
        err = security_sb_check_sb(mnt, path);
@@ -1623,7 +1623,7 @@ static int do_move_mount(struct path *path, char *old_name)
 
        err = -ENOENT;
        mutex_lock(&path->dentry->d_inode->i_mutex);
-       if (IS_DEADDIR(path->dentry->d_inode))
+       if (cant_mount(path->dentry))
                goto out1;
 
        if (d_unlinked(path->dentry))
@@ -2234,7 +2234,7 @@ SYSCALL_DEFINE2(pivot_root, const char __user *, new_root,
        if (!check_mnt(root.mnt))
                goto out2;
        error = -ENOENT;
-       if (IS_DEADDIR(new.dentry->d_inode))
+       if (cant_mount(old.dentry))
                goto out2;
        if (d_unlinked(new.dentry))
                goto out2;
index 30b93b2..eebb617 100644 (file)
@@ -186,6 +186,8 @@ d_iput:             no              no              no       yes
 
 #define DCACHE_FSNOTIFY_PARENT_WATCHED 0x0080 /* Parent inode is watched by some fsnotify listener */
 
+#define DCACHE_CANT_MOUNT      0x0100
+
 extern spinlock_t dcache_lock;
 extern seqlock_t rename_lock;
 
@@ -358,6 +360,18 @@ static inline int d_unlinked(struct dentry *dentry)
        return d_unhashed(dentry) && !IS_ROOT(dentry);
 }
 
+static inline int cant_mount(struct dentry *dentry)
+{
+       return (dentry->d_flags & DCACHE_CANT_MOUNT);
+}
+
+static inline void dont_mount(struct dentry *dentry)
+{
+       spin_lock(&dentry->d_lock);
+       dentry->d_flags |= DCACHE_CANT_MOUNT;
+       spin_unlock(&dentry->d_lock);
+}
+
 static inline struct dentry *dget_parent(struct dentry *dentry)
 {
        struct dentry *ret;