[Pkg-xen-devel] [Xen-devel] Crash during boot in Debian lenny default dom0 kernel (2.6.26-2-xen-686) / bugfix patches

[Pasi Kärkkäinen]

Hello,

The email (thread) below from xen-devel mailinglist indicates the changesets 
that should fix the bug that causes the lenny 2.6.26 xen kernel to crash on 
newer xen hypervisors.

The workaround for this bug has been to use pci=nomsi on the dom0 kernel cmdline,
but it would be better to apply the actual fix to the xen kernel in lenny.

More information on xen-devel (archives).

-- Pasi

----- Forwarded message from Jan Beulich <JBeulich at novell.com> -----

From: Jan Beulich <JBeulich at novell.com>
To: George Dunlap <George.Dunlap at eu.citrix.com>, pasik at iki.fi
Cc: Sander Eikelenboom <linux at eikelenboom.it>,
	xen-devel at lists.xensource.com
Date: Thu, 25 Feb 2010 09:16:01 +0000
Subject: Re: [Xen-devel] Crash during boot in Debian lenny default dom0
	kernel (2.6.26-2-xen-686)

>>> George Dunlap <George.Dunlap at eu.citrix.com> 25.02.10 00:57 >>>
>I realize dom0 is a privileged guest, but it still seems like we
>should try not to crash Xen as a result on guest input. :-)

While generally I agree, I think in the given case this is unavoidable -
Xen could apply some sanity check, but the passing of a machine
address from Dom0 to Xen implies that Dom0 knows what it does,
and Xen trusts it. Specifically, struct physdev_map_pirq has this
contents according to the trace

.domid = 00007ff0
.type = 00000000
.index = ffffffff
.pirq = ffffffff
.bus = 00000000
.devfn = 00000008
.entry_nr = 00000000
.table_base = f5861e4a00000001

table_base would seem like not having been initialized at all. I
would guess that they use the structure definition from before
c/s 18323 (which had, instead of a table_base member, an
int field indicating MSI vs. MSI-X. The original definition was
added with c/s 17534 and 17535, but all of those changes
happened during 3.3 development, so no-one should be using
the old definition in released code..

Jan


----- End forwarded message -----