Re: [PATCH] random: Fix kernel panic due to system_wq use before init

[Linus Torvalds]

On Wed, Sep 14, 2016 at 12:34 PM, Waiman Long <waiman.long@xxxxxxx> wrote:
>
> I can try, but the 16-socket system that I have at the moment takes a long
> time (more than an hour) for one shutdown-reboot cycle. It may not be really
> more interrupts in 4.8, it may be that the random driver just somehow run
> very slow on my test machine as it seems to have a major rewrite in the 4.8
> cycle.

Looking at the random driver updates since 4.7, the only thing I see
is that .crng_fast_load() for the chacha20 randomness. And that should
trigger only until it's been initialized, so the cost looks like it
should be limited.

Is there some fundamental reason you think it's the random driver?
Other than the oops? Because I'd be more inclined to suspect just some
apic issue or something, where an actual interrupt line ends up
screaming or whatever. Is this UV? There's also the CPU hotplug state
machine changes etc.

But a few rounds of bisecting should hopefully cut down on the
suspects a lot. A *full* bisect might be 16-17 rounds, but if you can
do just four or five rounds of bisection, that should still cut it
down from 14k commits to "only" several hundred..

               Linus