On Mon, Sep 12, 2016 at 09:29:02PM +0200, Miklos Szeredi wrote:
> The first patch is actually a bug fix, but I put it into this bunch for
> simplicity...
>
> The rest are really cleanups as well as minor bugfixes that are byproducts
> of the cleanups.
>
> This series builds on the fact that i_op.readlink is already set to
> generic_readlink() in 43/50 of the cases. And of those 7 only 4 are doing
> something special. So more than 90% of readlinks are/could actually just
> call back into get_link.
>
> The interesting cases are:
>
> - AFS, which has readlink but not get_link
> - proc, that allow jumping while following symlinks
>
> The first is handled by setting IOP_NOFOLLOW on the inode by the fs.
>
> The second one is handled by introducing is_following_link() which returns
> a bool depending on whether current->nameidata is NULL or not. If it
> returns false ->get_link() should behave as ->readlink() did. Otherwise it
> should behave as id did previously.
>
> Builds and boots. Can even read symlinks.
I have no problem with "let's get rid of generic_readlink" - not that
it bought us much, but sure, if you want to have decision made based upon
the combination of flags, let's do it. Just make NULL ->readlink + non-NULL
->get_link() mean generic_readlink(), and we are done. Especially if you
do the usual "set the flag on inode the first time we need to check".
I also have no problem with overlayfs and ecryptfs assuming that we only deal
with normal symlinks.
Overloading ->get_link() for procfs-style ones is just plain wrong,
though. Your current->nameidata != NULL thing is bloody brittle - what
happens if some code triggers those readlinks when called by something
during pathname resolution? Sure, right now existing callers won't.
But it doesn't take much to grow such a place _and_ have the implications
go unnoticed for quite a while.
