Re: [PATCH 14/39] ovl: stack file ops

[Al Viro]

On Tue, Jun 12, 2018 at 11:24:39AM +0200, Miklos Szeredi wrote:

> > Note that anything that uses file_dentry() anywhere near ->open(),
> > ->read_iter() or ->write_iter() is an instant trouble with your scheme.
> > Such as
> > int nfs_open(struct inode *inode, struct file *filp)
> > {
> >         struct nfs_open_context *ctx;
> >
> >         ctx = alloc_nfs_open_context(file_dentry(filp), filp->f_mode, filp);
> >         if (IS_ERR(ctx))
> >                 return PTR_ERR(ctx);
> >         nfs_file_set_open_context(filp, ctx);
> >         put_nfs_open_context(ctx);
> >         nfs_fscache_open_file(inode, filp);
> >         return 0;
> > }
> >
> > You do want to support NFS for lower layers, right?
> 
> There's no change regarding how file_dentry() works.  We've just
> pushed these weird files (f_path points to overlay, f_inode points to
> underlay) down into the guts of overlayfs and are not directly
> referenced from the file table anymore.  That shouldn't make *any*
> difference from the lower fs's pov.

*owwww*
I'd managed to push that particular nest of horrors out of mind ;-/
Having dug out my notes from back then and grepped around...  The real
mess is not even /proc/*/maps - it's /proc/*/map_files/* and yes, the
reasons for that kludge are still valid ;-/

Fuck.  OK, so we want to get rid of ->f_path.dentry accesses and see
that they don't come back.  Leaving them around due to "it won't come
anywhere near overlayfs" was a mistake of the same kind as leaving
d_add() in ->lookup() instances where we'd been certain that filesystem
would never get exported over NFS.  Just as we'd got open-by-handle for
e.g. NFS, we'd got nothing to prevent ecryptfs as lower layer in
overlayfs...

I hate it, but... consider path_open() objections withdrawn for now.
Uses of ->vm_file (and rules for those) are too convoluted to untangle
at the moment.  I still would love to get that straightened out, but
it's not this cycle fodder, more's the pity...