Re: drivers/char/random.c: more ruminations

["H. Peter Anvin" <hpa@linux.intel.com>]

On 06/11/2014 06:11 AM, Theodore Ts'o wrote:
> On Tue, Jun 10, 2014 at 11:58:06PM -0400, George Spelvin wrote:
>> You can forbid underflows, but the code doesn't forbid overflows.
>>
>> 1. Assume the entropy count starts at 512 bytes (input pool full)
>> 2. Random writer mixes in 20 bytes of entropy into the input pool.
>> 2a. Input pool entropy is, however, capped at 512 bytes.
>> 3. Random extractor extracts 32 bytes of entropy from the pool.
>>    Succeeds because 32 < 512.  Pool is left with 480 bytes of
>>    entropy.
>> 3a. Random extractor decrements pool entropy estimate to 480 bytes.
>>     This is accurate.
>> 4. Random writer credits pool with 20 bytes of entropy.
>> 5. Input pool entropy is now 480 bytes, estimate is 500 bytes.
> 
> Good point, that's a potential problem, although messing up the
> accounting betewen 480 and 500 bytes is not nearly as bad as messing
> up 0 and 20.
> 
> It's not something where if the changes required massive changes, that
> I'd necessarily feel the need to backport them to stable.  It's a
> certificational weakness, but it's a not disaster.
> 

Actually, with the new accounting code it will be even less serious,
because mixing into a nearly full pool is discounted heavily -- because
it is not like filling a queue; the mixing function will
probabilistically overwrite existing pool entropy.

So it is still a race condition, and still wrong, but it is a lot less
wrong.

	-hpa