Re: Q: check_unsafe_exec() races (Was: [PATCH 2/4] fix setuid sometimes doesn't)

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

On Wed, Apr 01, 2009 at 01:28:01AM +0100, Hugh Dickins wrote:
> Minor bisectability issue: the third patch, which introduces
> int unshare_fs_struct(void), needs to return 0 when it succeeds:
> that gets corrected in the fourth patch.

ACK.

> Lockdep objects to how check_unsafe_exec nests write_lock(&p->fs_lock)
> inside lock_task_sighand(p, &flags).  It's right: we sometimes take
> sighand->siglock in interrupt, so if such an interrupt occurred just
> after you take fs_lock elsewhere, that could deadlock with this.  It
> seems happy with taking fs_lock just outside the lock_task_sighand.

Right you are, check_unsafe_exec() reordered.  Will push in a few.
 
> Otherwise it looks good to me, except I keep worrying about those
> EAGAINs.  The more so once I noticed current->cred_exec_mutex is
> already being used to handle a similar issue with ptrace.  What
> do you think of this rather smaller patch?  which I'd much rather
> send after having slept on it, since it may be embarrassingly and
> obviously wrong, but tomorrow may be too late ...
 
Eh...  I'm not particulary happy with fork() growing heavier and heavier.
Besides, there's a subtle problem avoided by another variant - think what
happens if past the point of no return execve() will unshare fs_struct
(e.g. by explicit unshare() from dynamic linker).

Frankly, -EAGAIN in situation when we have userland race is fine.  And
we *do* have a userland race here - execve() will kill -9 those threads
in case of success, so if they'd been doing something useful, they are
about to be suddenly screwed.

So I stand by my variant.  Note that if we have *other* tasks sharing
fs_struct, your variant will block their clone() for the duration of
execve() while mine will simply leave them alone (and accept that we
have unsafe sharing).