[openssl-project] Random devices in chroot environments revisited

[Kurt Roeckx]

On Tue, Oct 30, 2018 at 07:23:52PM +0000, Dr. Matthias St. Pierre wrote:
> Hi,
> 
> I'd like to recall that the following issue
> 
>     #7419 - RAND_keep_random_devices_open not working
> 
> still needs to be fixed until 1.1.1a and that currently there are two
> alternative approaches for doing it:
> 
>     #7429 - Conditionally open random devices on initialization
>     #7437 - rand_unix.c: open random devices on first use only 
> 
> A short recall of the background story: In pull request
> 
>     #6432 - Keep /dev/random open for seeding
> 
> a regression was fixed that affected applications in chroot environments.
> It compensated the fact that the new OpenSSL CSPRNG now reseeds
> periodically, which the previous one didn't.
> 
> The solution was to open all random devices early and keep them
> open.  An API call (RAND_keep_random_devices_open()) was added
> for the application to opt-out from this behaviour.
> 
> In issue #7419 it was reported that this opt-out did not work as expected.
> 
> *  Pull request #7429 fixes the opt-out issue but remains along the lines of
>     the initial solution #7432. This approach has the side effect that the
>     random devices are always opened, even if they are never used
>     (for example, because getrandom() is available). Even more, if the
>     application forks and execs, these handles will be left open and unused
>     in the child process.
> 
>     An application which does not want this behavior, could explicitly opt-out,
>     but this would require a recompilation, which is somewhat contrary to
>     the assumptions of the initial chroot problem.
> 
> *  Pull request #7429 works without explicit opt-out, because it opens the

#7437 I guess

>     random devices on first use only (and then keeps them open, unless
>     the opt-out was called). The advantage of this approach: If getrandom()
>     is available and working, the opening will never happen.
> 
>     The lazy opening does not add a regression for applications compiled
>     against 1.1.0, because the old SSLEAY CSPRNG used to be initialized 
>     on first use, too. So also with 1.1.0 it was necessary to initialize the
>     CSPRNG properly before chrooting (unless the random device was
>     mounted into the chroot jail).
> 
> Your thoughts?

I prefer the option where it doesn't get opened if getrandom() is
avialable.


Kurt