Re: mmotm 2008-12-01-19-41: early exception (page fault -- deref of 0x20)

[Matt Mackall <mpm@selenic.com>]

On Tue, 2008-12-02 at 15:50 -0800, Arjan van de Ven wrote: 
> Matt Mackall wrote:
> > 
> > If we oops or warn while picking a timesource, we'll have lots of fun?
> > 
> 
> we really only need to mix in the tsc; ktime_get() is just an arch friendly way to get that
> I supposed (wrongly).
> 
> but yes we need to do a few things
> 1) seed on demand with a platform time source

Currently we use jiffies + get_cycles(). That's going to have somewhere
between, oh, 3 bits of entropy (very stable boot with only jiffies) and
25 bits of entropy (TSC with lots of waiting for hardware) at boot. 
Ideally, we'd have access to a wall clock of some sort as well. But
that's also a fairly limited source - wall clocks are both low
resolution and predictable/collision-prone.

> 2) have a way where arch init can just hand semi random data during the boot process to
>     increase the randomness (even if it doesn't count as entropy)

A simple wrapper around mix_pool_bytes probably fits the ticket.

But I don't think this will solve the general problem of 'large numbers
of practically identical machines booting up with the same pre-init
random number pools'. Beyond things like MAC addresses and serial
numbers (predictable/observable but at least not collision-prone), we
have no way to differentiate some boxes. We may need to forcibly
generate some timing entropy. Perhaps something like this:

http://markmail.org/message/xwsbywr6ziil2qu2

(which is way too slow in its current form)

There's a related problem of systems with no way to store a seed across
boots.

-- 
Mathematics is the supreme nostalgia of our time.