event based initramfs

[Clint Byrum]

Excerpts from Phillip Susi's message of Fri May 11 07:25:48 -0700 2012:
> On 5/11/2012 4:04 AM, Robert Collins wrote:
> > So a related issue is dmraid, which is very similar to mdadm. My last
> > two dmraid machines fail - every time - to activate on boot. The
> > reason being that the devices they depend upon don't come up fast
> > enough, and the configuration process isn't event-driven *enough* to
> 
> Event driven *enough*?  What do you mean?  How can it be more or less 
> event driven ( in other words, either it is or is not, there isn't 
> really an in between )?  When the drive add event comes in, udev runs 
> dmraid to activate the array.  As long as that happens before 
> wait-for-root gives up, all is well.
> 
> If you're getting timeouts, then either the timeout is too short, or you 
> have some VERY slow initializing hardware.  How would upstart help this?
> 

So the timeout would be the one in 'udevadm settle'. I think it probably
is long enough for the general case. The real issue is that we are failing
a little too fast when we fail to find the root FS that we expect to
find. I'm not so sure upstart is really the answer for this either.
If we trust in upstart to wait until this event comes, we are in the
same boat as we'd be with udevadm settle --timeout=9999999999. Perhaps
the right thing is for dmraid's initramfs scripts to wait longer than
the standard amount of time, however, at some point we do need to fail
and give the user a chance to inspect things.