Re: [git pull] mount API series

[Linus Torvalds]

On Thu, Nov 1, 2018 at 3:53 AM Steven Whitehouse <swhiteho@xxxxxxxxxx> wrote:
>
> When I look at the discussions I'm seeing two main issues (please
> correct me if you think I'm wrong about this) which are (a) whether the
> design is correct and (b) whether there are still bugs in the current
> patch set.
>
> Which of these are you most concerned about?

I'm most worried about bugs _due_ to the new design.

Exactly because it splits up what used to be an atomic sequence, I
worry about the intermediate states having issues (the refcounting
things we've already seen, for example), but I also worry about the
fact that it completely changes the model, and that that makes things
like security hooks fundamentally different.

The latter may not be a "bug" in the sense that it's all intentional,
but it does mean that I see *one* mount-time security hook now having
been replaced by *five* security hooks.

And that's ignoring the alloc/dup/free ones.

As far as I can tell, the patch-series simply added the hooks. It made
no attempt at making sure that previous hooks had sane semantics. Do
they?  So now a system that has an old mount hook can be bypassed by
simplky using the new model instead.

I dunno. The patches are illegible in this regard (and I don't blame
the fsmount ones, I blame the security subsystem that just is full of
random indirection to random sub-security systems, which in turn just
have hash lookups for data structures set up by other operations
entirely).

Eric was pointing out bugs as late as the weekend before the merge
window opened. That, to me, does not say "ready for the merge window".

                   Linus