Re: [PATCH -v3 8/8] dnotify: reimplement dnotify using fsnotify

[Eric Paris <eparis@redhat.com>]

On Fri, 2008-11-28 at 05:14 +0000, Al Viro wrote:
> On Tue, Nov 25, 2008 at 12:21:33PM -0500, Eric Paris wrote:
> 
> > +	.mark_clear_inode = clear_mark_dir_notify,
> 
> ... called under a spinlock
> 
> > +static void clear_mark_dir_notify(struct fsnotify_mark_entry *entry, struct inode *inode, unsigned long mask __attribute__ ((unused)), unsigned int flags)
> > +{
> ...
> > +	fsnotify_put_group(dnotify_group);
> 
> ... which grabs a mutex.

You're right, I should drop and retake the spinlock.  But in reality I
shouldn't ever get here and plan to replace all of this code with a
BUG() rather than the WARN() I have today since I know I can safely
recover.

> Incidentally, why the hell do you bother with refcounting on groups here?
> dnotify is not something that's going to be unloaded, for fsck sake...

Well, I do unregister dnotify if you stop watching any files.  I also
plan to implement inotify as one inotify_init() per group.  And fsnotify
groups exist only as long as there is a fsnotify socket bound....

-Eric