Re: [PATCH] CRED: Fix regression in cap_capable() as shown up by sys_faccessat() [ver #2]

[David Howells <dhowells@redhat.com>]

J. Bruce Fields <bfields@fieldses.org> wrote:

> > > > No.  I started bisecting, and it does appear to be a regression from the
> > > > cred patches, but at some point in the middle there it hangs on boot (a
> > > > softlockup report blames a spinlock in set_groups).
> > > 
> > > Do you remember which patch you were at?
> 
> More precisely:
> 	- The last working commit is b6dff3ec... "CRED: Separate task
> 	  security context from task_struct".
> 	- The first commit exhibiting the permissions problem is
> 	  a6f76f2... "CRED: Make execve() take advantage of
> 	  copy-on-write credentials".

I presume by 'first' you mean 'latest'.

> 	- The 9 commits in between (from f1752eec to d84f4f9) result in
> 	  a soft lookup on boot.

I think the problem may be that f1752eec removes the lock initialisation for
init_cred from the INIT_CRED() macro:

	-	.lock			= __SPIN_LOCK_UNLOCKED(p.lock),	\

but doesn't add it to the out of line init_cred:

	+struct cred init_cred = {
	+	.usage			= ATOMIC_INIT(3),
	+	.securebits		= SECUREBITS_DEFAULT,
	+	.cap_inheritable	= CAP_INIT_INH_SET,
	+	.cap_permitted		= CAP_FULL_SET,
	+	.cap_effective		= CAP_INIT_EFF_SET,
	+	.cap_bset		= CAP_INIT_BSET,
	+	.user			= INIT_USER,
	+	.group_info		= &init_groups,
	+};

Can you try adding:

		.lock			= __SPIN_LOCK_UNLOCKED(init_cred.lock),

to that and also add:

		spin_lock_init(&pcred->lock);

into copy_creds() see if the problem goes away?

I'm a bit surprised that lockdep doesn't bark at this one - I thought it
checked for lock initialisation.

David