Re: CURRENT: WARNING! r273914 leaves filesystems in inconsistent/corrupted condition!

On Sat, 2014-11-01 at 15:59 +0100, Dag-Erling Smørgrav wrote:
> Ian Lepore <ian_at_FreeBSD.org> writes:
> > Dag-Erling Smørgrav <des_at_des.no> writes:
> > > That means we're not getting enough entropy during early boot, or
> > > we're underestimating the amount of entropy we're getting.  We added
> > > entropy harvesting to device_attach() about a year ago, which in
> > > most cases provides enough entropy to unblock /dev/random before we
> > > even run init(8).
> > And I vaguely remember being promised that things like that would
> > NEVER happen, even on systems with little or no entropy available
> > during early startup (which describes quite nicely the embedded
> > systems we build at work).
> 
> I think you misremember.  It is impossible to guarantee that the system
> will always have enough entropy right from the start.  Servers, desktops
> and laptops will be fine, but embedded systems and VMs might not be able
> to unblock until they've seen some network traffic or loaded a chunk of
> pre-generated entropy (which is what /etc/rc.d/random does).  This is
> especially true for embedded systems that don't have enumerable buses
> and rely on fdt(4) to create the device tree at boot time.
> 

And what about devices that are not connected to a network?  We've been
over this, I stressed again and again that we have an absolute
requirement to start up reliably when there is NO ENTROPY.  Saved
entropy is great, if you've got some saved.

Oh well, I'm sure I'll be able to find some hacks to undo whatever y'all
have done now, and we'll just have to carry them as local diffs forever.

-- Ian

> VMs have the additional problem of divergence between clones: if you
> clone a VM, all clones will start out with the exact same state and
> won't diverge until they've all reseeded after gathering entropy
> independently of eachother.  I don't really know how to solve this.  One
> possibility, assuming you have guest additions installed and that they
> can tell you that you've been suspended, is to block on resume.  It
> won't help VMs that were cloned while shut down, but they should diverge
> to some extent during boot.
> 
> DES