Re: BUG_ON drivers/char/random.c:986

[Hannes Frederic Sowa <hannes@stressinduktion.org>]

On Di, 2014-07-15 at 00:36 -0400, Dave Jones wrote:
> On Fri, May 16, 2014 at 10:18:40PM -0400, Theodore Ts'o wrote:
>  > On Fri, May 16, 2014 at 05:46:22PM -0700, Hannes Frederic Sowa wrote:
>  > > This should do the trick:
>  > > dd if=/dev/urandom of=/dev/zero bs=67108707
>  > > 
>  > > I suspect ee1de406ba6eb1 ("random: simplify accounting logic") as the
>  > > culprit.
>  > 
>  > Yep, that it's it.  Thanks for noticing this so quickly!  I'll push
>  > the following patch to Linus.
>  > 
>  >     	      	    			- Ted
>  > 
>  > commit 29fb0ca5b3922288fba3f4c975a55032a51df0f0
>  > Author: Theodore Ts'o <tytso@mit.edu>
>  > Date:   Fri May 16 21:40:41 2014 -0400
>  > 
>  >     random: fix BUG_ON caused by accounting simplification
> 
> I just hit this bug again on 3.16rc5, which has this patch, so
> there's still some corner case that isn't happy..
> 
> (old thread is here for reference: https://lkml.org/lkml/2014/5/16/724)

Looks like an overflow introduced by e33ba5fa7afce1a ("random: fix nasty
entropy accounting bug"). Ted, what do you think about the following
fix?

Bye,
Hannes

diff --git a/drivers/char/random.c b/drivers/char/random.c
index 0a7ac0a..72511e8 100644
--- a/drivers/char/random.c
+++ b/drivers/char/random.c
@@ -999,7 +999,9 @@ retry:
 	}
 	if (ibytes < min)
 		ibytes = 0;
-	if ((entropy_count -= ibytes << (ENTROPY_SHIFT + 3)) < 0)
+
+	if ((entropy_count -= ibytes << (ENTROPY_SHIFT + 3)) < 0 ||
+	    entropy_count > orig)
 		entropy_count = 0;
 
 	if (cmpxchg(&r->entropy_count, orig, entropy_count) != orig)