Re: [PATCH] random: mix hwgenerator randomness before sleeping

[Dominik Brodowski <linux@dominikbrodowski.net>]

Hi Jason,

Am Wed, May 04, 2022 at 02:45:46AM +0200 schrieb Jason A. Donenfeld:
> On Tue, May 3, 2022 at 9:56 PM Dominik Brodowski
> <linux@dominikbrodowski.net> wrote:
> >
> > Am Tue, May 03, 2022 at 09:51:41PM +0200 schrieb Jason A. Donenfeld:
> > > The add_hwgenerator_randomness() function is called in a loop from a
> > > kthread by the hwgenerator core. It's supposed to sleep when there's
> > > nothing to do, and wake up periodically for more entropy. Right now it
> > > receives entropy, sleeps, and then mixes it in. This commit reverses the
> > > order, so that it always mixes in entropy sooner and sleeps after. This
> > > way the entropy is more fresh.
> >
> > ... however, the hwgenerator may take quite some time to accumulate entropy
> > after wakeup. So now we might have a delay between a wakeup ("we need more
> > entropy!") and that entropy becoming available. Beforehand, the thread only
> > went to sleep when there is no current need for "fresh" entropy.
> 
> Huh, interesting consideration. I didn't think about that. You wrote,
> "hwgenerator may take quite some time to accumulate entropy" -- any
> idea how long in the worst case? A second? A minute?

For TPM or virtio-rng, this shouldn't take long. But, for example, the
Kconfig entry for HW_RANDOM_TIMERIOMEM states that

	  This driver provides kernel-side support for a generic Random
	  Number Generator used by reading a 'dumb' iomem address that
	  is to be read no faster than, for example, once a second;
	  the default FPGA bitstream on the TS-7800 has such functionality.

Then, optee-rng.c contains a value "data_rate" in random bytes per second,
and may sleep for a considerable time:

	msleep((1000 * (max - read)) / pvt_data->data_rate);

Thanks,
	Dominik