d_invalidate allowed a non-directory mountpoint to be invalidated,
which is bad, since the mountpoint becomes unreachable.
I know it's racy wrt attaching/detaching mount, but AFAICS so is
everything else that unhashes the dentry. This seems to be an
oversight when splitting out vfsmount_lock from dcache_lock. To be
fixed.
Signed-off-by: Miklos Szeredi <[email protected]>
---
Index: linux/fs/dcache.c
===================================================================
--- linux.orig/fs/dcache.c 2005-10-04 13:59:57.000000000 +0200
+++ linux/fs/dcache.c 2005-10-07 12:59:11.000000000 +0200
@@ -350,7 +350,8 @@ int d_invalidate(struct dentry * dentry)
*/
spin_lock(&dentry->d_lock);
if (atomic_read(&dentry->d_count) > 1) {
- if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
+ if (dentry->d_inode && (S_ISDIR(dentry->d_inode->i_mode) ||
+ d_mountpoint(dentry))) {
spin_unlock(&dentry->d_lock);
spin_unlock(&dcache_lock);
return -EBUSY;
On Fri, Oct 07, 2005 at 01:13:50PM +0200, Miklos Szeredi wrote:
> d_invalidate allowed a non-directory mountpoint to be invalidated,
> which is bad, since the mountpoint becomes unreachable.
>
> I know it's racy wrt attaching/detaching mount, but AFAICS so is
> everything else that unhashes the dentry. This seems to be an
> oversight when splitting out vfsmount_lock from dcache_lock. To be
> fixed.
NAK. That's a wrong way to deal with the problem and it's much older
than vfsmount_lock or dcache_lock (and affects directories too).
> > d_invalidate allowed a non-directory mountpoint to be invalidated,
> > which is bad, since the mountpoint becomes unreachable.
> >
> > I know it's racy wrt attaching/detaching mount, but AFAICS so is
> > everything else that unhashes the dentry. This seems to be an
> > oversight when splitting out vfsmount_lock from dcache_lock. To be
> > fixed.
>
> NAK. That's a wrong way to deal with the problem and it's much older
> than vfsmount_lock or dcache_lock (and affects directories too).
Sorry?
Directories are not invalidated if they have any other reference (like
a mount, or any subdirectories which may have mounts).
So how does it affect directories?
Miklos
On Fri, Oct 07, 2005 at 04:35:53PM +0200, Miklos Szeredi wrote:
> > > d_invalidate allowed a non-directory mountpoint to be invalidated,
> > > which is bad, since the mountpoint becomes unreachable.
> > >
> > > I know it's racy wrt attaching/detaching mount, but AFAICS so is
> > > everything else that unhashes the dentry. This seems to be an
> > > oversight when splitting out vfsmount_lock from dcache_lock. To be
> > > fixed.
> >
> > NAK. That's a wrong way to deal with the problem and it's much older
> > than vfsmount_lock or dcache_lock (and affects directories too).
>
> Sorry?
>
> Directories are not invalidated if they have any other reference (like
> a mount, or any subdirectories which may have mounts).
>
> So how does it affect directories?
The underlying problem is still there - parts of mount tree _can_ go
unreachable when remote object dies; trying to pin them down is hopeless
and the only sane way to deal with that is to dissolve the subtrees
of mount when that happens.
> > > > d_invalidate allowed a non-directory mountpoint to be invalidated,
> > > > which is bad, since the mountpoint becomes unreachable.
> > > >
> > > > I know it's racy wrt attaching/detaching mount, but AFAICS so is
> > > > everything else that unhashes the dentry. This seems to be an
> > > > oversight when splitting out vfsmount_lock from dcache_lock. To be
> > > > fixed.
> > >
> > > NAK. That's a wrong way to deal with the problem and it's much older
> > > than vfsmount_lock or dcache_lock (and affects directories too).
> >
> > Sorry?
> >
> > Directories are not invalidated if they have any other reference (like
> > a mount, or any subdirectories which may have mounts).
> >
> > So how does it affect directories?
>
> The underlying problem is still there - parts of mount tree _can_ go
> unreachable when remote object dies; trying to pin them down is hopeless
> and the only sane way to deal with that is to dissolve the subtrees
> of mount when that happens.
I get it.
Miklos