Re: proper tty-ldisc refcounting (was Re: WARNING at:drivers/char/tty_ldisc.c)

[Linus Torvalds]

On Tue, 4 Aug 2009, Greg KH wrote:
> 
> Ok, but due to the lateness of the release cycle, is it worth it to add
> those 3 right now?  Or do we just take the BUG_ON() out as it's pretty
> harmless while shutting down in single user mode?
> 
> What do you think?

If the WARN_ON() happens, it's not just that we have a refcount being off, 
we'll also have memory corruption and a potential oops a bit later. 

Why? Simply because somebody will be touching that 'struct ldisc', even if 
it will be just a decrement of the word that contained the refcount. So we 
have a pretty much guaranteed use-after-free scenario.

So taking out the WARN_ON() is the wrong thing. In that case it would 
probably be better to just leave the WARN_ON(), and at least know "ok, bad 
things happened". 

So I think we have a few options:

 - leave things as-is, leave the WARN_ON(), and know that it's very rare, 
   but that bad things can happen when it triggers.

   The thing I really dislike about this one is that yes, it's rare, but 
   I could see it be user-triggerable. Users do have access to 
   /dev/console when they are logged in at the console.

 - Change the old code from

	WARN_ON(ld->refcount);
	kfree(ld);

   to

	if (!WARN_ON_ONCE(ld->refcount))
		kfree(ld);

   which at least doesn't free the ldisc if it is in use. So now you have 
   a memory leak, but at least hopefully no actual corruption and 
   use-after-free. It's still a bug, but it won't be causing other bugs 
   down the line (except for running out of memory if you can trigger it 
   really easily, but that's unlikely, and I think preferable to unknown 
   problems - even if the unknown problems are very unlikely)

 - Take the three patches now.

I suspect we should take the three now. All of the issues are due to 
totally rewritten code since 2.6.30 - I suspect the risk from new bugs 
from that refcounting series is _smaller_ that the risk of bugs from the 
original ldisc rewrite (commit c65c9bc3e), and if there are bugs, I 
suspect the three patches are more likely to help than to hurt.

			Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/