Subject: Re: port-xen/29775: Cannot use a vnode disk for domU
To: None <tech-kern@NetBSD.org>
From: Manuel Bouyer <bouyer@antioche.eu.org>
List: tech-kern
Date: 03/23/2005 22:12:45
[ moving to tech-kern ]

On Wed, Mar 23, 2005 at 05:42:00PM +0000, juan@xtraeme.nopcode.org wrote:
> >Description:
> 	With -current Xen sources, I want to create a domU using
> 	a vnode disk in a file:
> 
> 	$ dd if=/dev/zero of=mydisk.img bs=1k seek=20480k count=1
> 	$ vnconfig vnd0 mydisk.img
> 	
> 	$ grep vnd0 /etc/xen/netbsd
> 	disk = [ 'phy:/dev/cd0a,cd0a,r', 'phy:/dev/vnd0a,wd0d,w' ]
> 	$
> 
> 	The system will crash.

With:
panic: lockmgr: no context
It is in vndstrategy:
		vn_lock(vnd->sc_vp, LK_EXCLUSIVE | LK_RETRY | LK_CANRECURSE);

This means that I/O on the vnd block device expect to be called from
a process context, while the xbd backend calls it from interrupt context.
And obvious way to fix this would be to add a kernel thread to the xbd backend
to do this. But maybe this needs a closer look:
- Can we safely say that vnd will always be called from process context ?
  What if we stack something else on vnd (fss or cgd for example) ?
- Always taking a kernel thread in the xbd backend means an extra context
  switch, which we may want to avoid if possible. Is there a way to know if
  the underlying device can handle interrupt context or not ?

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
     NetBSD: 26 ans d'experience feront toujours la difference
--