Re: [RFC] readlink()-related oddities

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

On Fri, Nov 20, 2015 at 09:59:05AM +0000, David Howells wrote:
> Al Viro <viro@ZenIV.linux.org.uk> wrote:
> 
> > 3) normally, readlink(2) fails for non-symlinks.  Moreover, according to
> > POSIX it should do so (with -EINVAL).  There is a pathological case when
> > it succeeds for a directory, though.  Namely, one of the kinds of AFS
> > "mountpoints".
> 
> All AFS mountpoints are magic symlinks that are specially interpreted by the
> client as far as I'm aware.  I'm not sure why the designers didn't just select
> a different file type for them, but they didn't.

All of them?  I see two kinds there - one is magical symlink (recognized
by contents in afs_iget()), another is this autocell thing, the latter
having no ->readlink().  Both serve as automount points, don't they?

> > stat(2) reports those as directories, stepping into them leads to
> > automounting a directory there (why do we have ->open() for them, BTW?).
> 
> I think I put that in to make sure the open() syscall returned EREMOTE rather
> than another error if you tried to open it.  It can probably be removed
> because with the d_automount code you can't ever get there I think - unless
> you can pass AT_NO_AUTOMOUNT to openat().

Just how would openat() get the AT_... flags?  Only statat(2) accepts
AT_NO_AUTOMOUNT, sorry.
 
> > How the hell is userland supposed to guess to call readlink(2) on those
> > suckers to get the information of what'll get automounted there if we step
> > upon them?
> 
> There's an AFS userspace command that could be used to query a mountpoint that
> was going to use it.  However, I suspect readlink() will now always trigger
> the automount.  This is one of the things OpenAFS uses pioctl() for - but
> since I'm not allowed to add that to the kernel, I have to find some other way
> of doing it.

Well, pioctl() is a piec^H^Hle of shit interface; let's figure out what we'd
actually want to implement and do that.

One obvious thing is "here's a pathname, tell me what gets automounted here"
(with interesting question of what to do if the automount is being triggered
right now).  Another thing is locating those guys; if we had a separate file
type for them (i.e. could recognize them by st_mode _and_ d_type), we would
be fine (the usual tree-walkers would be able to spot such places and query
them for prospective automount targets), but without that... a syscall for
everything in a tree just to list those suckers?