Re: [PATCH v2] random: set fast pool count to zero in cpuhp teardown

[Sebastian Andrzej Siewior <bigeasy@linutronix.de>]

On 2022-02-14 15:52:34 [+0100], Jason A. Donenfeld wrote:
> On Mon, Feb 14, 2022 at 3:49 PM Sebastian Andrzej Siewior
> <bigeasy@linutronix.de> wrote:
> >
> > On 2022-02-14 15:42:50 [+0100], Jason A. Donenfeld wrote:
> > > Hi Sebastian,
> > Hi Jason,
> >
> > > If we move this to startup, is there a phase during which no interrupt
> > > will arrive? That is, can this happen very very early in startup, so
> > > that zeroing out count happens *before* ++count?
> >
> > Interrupts will arrive starting with CPUHP_AP_ONLINE_IDLE from the CPU
> > HP point of view. My suggestion had a check for upper most bit and only
> > clear count if that bit was seen. So we wouldn't clear the counter if we
> > wouldn't suspect one of the rare corner cases.
> 
> That doesn't work for the other use cases I have for this (see the other patch).

But you acked my question regarding boot-up? So the teardown callback
won't happen during boot-up.

> So I think it seems better to keep it before CPUHP_TIMERS_PREPARE, but
> do it on startup rather than teardown. Seem reasonable? Would that
> mean we zero out before IRQs are enabled?
I would only zero it if the upper-most bit is there.
If you need (want) to reset the get_random_uXX() pools and such, there
is nothing wrong with having an early notifier at CPU up time before the
CPU gets active (say CPUHP_RANDOM_PREPARE) where you make sure that the
pools will re-fill during first usage.
And then have another one after CPUHP_AP_WORKQUEUE_ONLINE to ensure that
a possible scheduled worker between CPUHP_AP_ONLINE_IDLE and
CPUHP_AP_WORKQUEUE_ONLINE runs on the correct CPU. And this covers also
the rollback problem.

Sebastian