Re: [PATCH] fscrypt: cache decrypted symlink target in ->i_link

[Al Viro <viro@zeniv.linux.org.uk>]

On Tue, Apr 09, 2019 at 06:22:49PM -0700, Eric Biggers wrote:

> > Non-NULL ->get_link() => DCACHE_SYMLINK_TYPE in ->d_flags =>
> > d_is_symlink() true => step_into() progresses to pick_link().
> > 
> > IOW, non-NULL ->get_link() is what tells you that we have
> > a symlink there.
> 
> I think that's pretty unintuitive.  The fact that multiple filesystems including
> ext4 set ->i_link on fast symlinks, then set ->get_link() to a function that
> returns ->i_link, made me assume that's the mechanism by which such symlink
> targets are returned to the VFS.  When in fact fs/namei.c just uses ->i_link,
> and never calls ->get_link().
> 
> Is there any reason why d_flags_for_inode() doesn't check S_ISLNK() instead, and
> then fs/namei.c would call ->get_link() if non-NULL, otherwise use ->i_link?

Extra check and dereference on hot path with no visible benefits of doing it
that way, for starters.  Really, what _is_ the benefit of pessimizing that?  
Most of the symlinks we run into will have ->i_link set; checking ->i_op->get_link
first is extra work for no good reason...

What's more, ->get_link is visible in inode_operations; ->i_link (let alone ->i_mode)
isn't.  As it is, we can easily tell symlink inode_operations from everything else
on the source level.  With your scheme we won't.