[patch] add two inotify_add_watch flags

[patch] add two inotify_add_watch flags

[John McCutchan]

Yo,

The below patch lets user space have more control over the inodes that
inotify will watch. It introduces two new flags. 

	IN_ONLYDIR -- only watch the inode if it is a directory
	This is needed to avoid the race that can occur when we want to be sure
that we are watching a directory.

	IN_DONT_FOLLOW -- don't follow a symlink. In combination with
IN_ONLYDIR we can make sure that we don't watch the target of symlinks.

The issues the flags fix came up when writing the gnome-vfs inotify
backend. Default behaviour is unchanged.


Signed-off-by: John McCutchan <ttb@tentacle.dhs.org>
Index: linux-2.6.14-rc5/fs/inotify.c
===================================================================
--- linux-2.6.14-rc5.orig/fs/inotify.c	2005-10-20 02:23:05.000000000 -0400
+++ linux-2.6.14-rc5/fs/inotify.c	2005-12-06 19:01:22.000000000 -0500
@@ -363,11 +363,18 @@
 /*
  * find_inode - resolve a user-given path to a specific inode and return a nd
  */
-static int find_inode(const char __user *dirname, struct nameidata *nd)
+static int find_inode(const char __user *dirname, struct nameidata *nd,
+		      int only_dir, int dont_follow_symlink)
 {
 	int error;
+	unsigned flags = 0;
 
-	error = __user_walk(dirname, LOOKUP_FOLLOW, nd);
+	if (!dont_follow_symlink)
+		flags |= LOOKUP_FOLLOW;
+	if (only_dir)
+		flags |= LOOKUP_DIRECTORY;
+
+	error = __user_walk(dirname, flags, nd);
 	if (error)
 		return error;
 	/* you can only watch an inode if you have read permissions on it */
@@ -943,7 +950,7 @@
 		goto fput_and_out;
 	}
 
-	ret = find_inode(path, &nd);
+	ret = find_inode(path, &nd, mask & IN_ONLYDIR, mask & IN_DONT_FOLLOW);
 	if (unlikely(ret))
 		goto fput_and_out;
 
Index: linux-2.6.14-rc5/include/linux/inotify.h
===================================================================
--- linux-2.6.14-rc5.orig/include/linux/inotify.h	2005-10-20 02:23:05.000000000 -0400
+++ linux-2.6.14-rc5/include/linux/inotify.h	2005-12-06 18:57:11.000000000 -0500
@@ -47,6 +47,8 @@
 #define IN_MOVE			(IN_MOVED_FROM | IN_MOVED_TO) /* moves */
 
 /* special flags */
+#define IN_ONLYDIR		0x01000000	/* only watch the path if it is a directory */
+#define IN_DONT_FOLLOW		0x02000000	/* don't follow a sym link */
 #define IN_MASK_ADD		0x20000000	/* add to the mask of an already existing watch */
 #define IN_ISDIR		0x40000000	/* event occurred against dir */
 #define IN_ONESHOT		0x80000000	/* only send event once */

Re: [patch] add two inotify_add_watch flags

[Robert Love]

On Tue, 2005-12-06 at 22:54 -0500, John McCutchan wrote:

> The below patch lets user space have more control over the inodes that
> inotify will watch. It introduces two new flags. 
> 
> 	IN_ONLYDIR -- only watch the inode if it is a directory
> 	This is needed to avoid the race that can occur when we want to be sure
> that we are watching a directory.
> 
> 	IN_DONT_FOLLOW -- don't follow a symlink. In combination with
> IN_ONLYDIR we can make sure that we don't watch the target of symlinks.
> 
> The issues the flags fix came up when writing the gnome-vfs inotify
> backend. Default behaviour is unchanged.

Looks good to me, and I confirm we hit these issues in real-life with
gnome-vfs.

Acked-by: Robert Love <rml@novell.com>

	Robert Love

Re: [patch] add two inotify_add_watch flags

[Andrew Morton]

John McCutchan <ttb@tentacle.dhs.org> wrote:
>
> -static int find_inode(const char __user *dirname, struct nameidata *nd)
> +static int find_inode(const char __user *dirname, struct nameidata *nd,
> +		      int only_dir, int dont_follow_symlink)
>  {
>  	int error;
> +	unsigned flags = 0;
>  
> -	error = __user_walk(dirname, LOOKUP_FOLLOW, nd);
> +	if (!dont_follow_symlink)
> +		flags |= LOOKUP_FOLLOW;
> +	if (only_dir)
> +		flags |= LOOKUP_DIRECTORY;
> +
> +	error = __user_walk(dirname, flags, nd);
>  	if (error)
>  		return error;
>  	/* you can only watch an inode if you have read permissions on it */
> @@ -943,7 +950,7 @@
>  		goto fput_and_out;
>  	}
>  
> -	ret = find_inode(path, &nd);
> +	ret = find_inode(path, &nd, mask & IN_ONLYDIR, mask & IN_DONT_FOLLOW);
>  	if (unlikely(ret))
>  		goto fput_and_out;

I'd have thought that it'd be more general to pass the __user_walk flags
into find_inode(), rather than calculating them in the callee.  Slightly
more efficient too.  

Re: [patch] add two inotify_add_watch flags

[John McCutchan]

On Fri, 2005-09-12 at 14:51 -0800, Andrew Morton wrote:
> John McCutchan <ttb@tentacle.dhs.org> wrote:
> >
> > -static int find_inode(const char __user *dirname, struct nameidata *nd)
> > +static int find_inode(const char __user *dirname, struct nameidata *nd,
> > +		      int only_dir, int dont_follow_symlink)
> >  {
> >  	int error;
> > +	unsigned flags = 0;
> >  
> > -	error = __user_walk(dirname, LOOKUP_FOLLOW, nd);
> > +	if (!dont_follow_symlink)
> > +		flags |= LOOKUP_FOLLOW;
> > +	if (only_dir)
> > +		flags |= LOOKUP_DIRECTORY;
> > +
> > +	error = __user_walk(dirname, flags, nd);
> >  	if (error)
> >  		return error;
> >  	/* you can only watch an inode if you have read permissions on it */
> > @@ -943,7 +950,7 @@
> >  		goto fput_and_out;
> >  	}
> >  
> > -	ret = find_inode(path, &nd);
> > +	ret = find_inode(path, &nd, mask & IN_ONLYDIR, mask & IN_DONT_FOLLOW);
> >  	if (unlikely(ret))
> >  		goto fput_and_out;
> 
> I'd have thought that it'd be more general to pass the __user_walk flags
> into find_inode(), rather than calculating them in the callee.  Slightly
> more efficient too.  

Makes sense.

Signed-off-by: John McCutchan <ttb@tentacle.dhs.org>

Index: linux-2.6.14-rc5/fs/inotify.c
===================================================================
--- linux-2.6.14-rc5.orig/fs/inotify.c	2005-12-09 17:59:10.000000000 -0500
+++ linux-2.6.14-rc5/fs/inotify.c	2005-12-09 18:04:22.000000000 -0500
@@ -363,11 +363,12 @@
 /*
  * find_inode - resolve a user-given path to a specific inode and return a nd
  */
-static int find_inode(const char __user *dirname, struct nameidata *nd)
+static int find_inode(const char __user *dirname, struct nameidata *nd, 
+		      unsigned flags)
 {
 	int error;
 
-	error = __user_walk(dirname, LOOKUP_FOLLOW, nd);
+	error = __user_walk(dirname, flags, nd);
 	if (error)
 		return error;
 	/* you can only watch an inode if you have read permissions on it */
@@ -932,6 +933,7 @@
 	struct file *filp;
 	int ret, fput_needed;
 	int mask_add = 0;
+	unsigned flags = 0;
 
 	filp = fget_light(fd, &fput_needed);
 	if (unlikely(!filp))
@@ -943,7 +945,12 @@
 		goto fput_and_out;
 	}
 
-	ret = find_inode(path, &nd);
+	if (!(mask & IN_DONT_FOLLOW))
+		flags |= LOOKUP_FOLLOW;
+	if (mask & IN_ONLYDIR)
+		flags |= LOOKUP_DIRECTORY;
+
+	ret = find_inode(path, &nd, flags);
 	if (unlikely(ret))
 		goto fput_and_out;
 
Index: linux-2.6.14-rc5/include/linux/inotify.h
===================================================================
--- linux-2.6.14-rc5.orig/include/linux/inotify.h	2005-12-09 17:59:10.000000000 -0500
+++ linux-2.6.14-rc5/include/linux/inotify.h	2005-12-09 18:00:49.000000000 -0500
@@ -47,6 +47,8 @@
 #define IN_MOVE			(IN_MOVED_FROM | IN_MOVED_TO) /* moves */
 
 /* special flags */
+#define IN_ONLYDIR		0x01000000	/* only watch the path if it is a directory */
+#define IN_DONT_FOLLOW		0x02000000	/* don't follow a sym link */
 #define IN_MASK_ADD		0x20000000	/* add to the mask of an already existing watch */
 #define IN_ISDIR		0x40000000	/* event occurred against dir */
 #define IN_ONESHOT		0x80000000	/* only send event once */