Re: [Xen-devel] [PATCH] xen: point xen_start_info to a dummy struct for PV on HVM guests

[Ian Campbell <Ian.Campbell@citrix.com>]

On Wed, 2012-10-03 at 17:05 +0100, Stefano Stabellini wrote:
> On Wed, 3 Oct 2012, Ian Campbell wrote:
> > On Wed, 2012-10-03 at 16:48 +0100, Stefano Stabellini wrote:
> > > On Wed, 3 Oct 2012, Ian Campbell wrote:
> > > > On Wed, 2012-10-03 at 15:11 +0100, Konrad Rzeszutek Wilk wrote:
> > > > > On Wed, Oct 03, 2012 at 02:54:42PM +0100, Ian Campbell wrote:
> > > > > > On Wed, 2012-10-03 at 14:51 +0100, Stefano Stabellini wrote:
> > > > > > > On Wed, 3 Oct 2012, Ian Campbell wrote:
> > > > > > > > On Wed, 2012-10-03 at 14:37 +0100, Stefano Stabellini wrote:
> > > > > > > > > PV on HVM guests don't have a start_info page mapped by Xen, so
> > > > > > > > > xen_start_info is just NULL for them.
> > > > > > > > > That is problem because other parts of the code expect xen_start_info to
> > > > > > > > > point to something valid, for example xen_initial_domain() is defined as
> > > > > > > > > follow:
> > > > > > > > > 
> > > > > > > > > #define xen_initial_domain()    (xen_domain() && \
> > > > > > > > >                  xen_start_info->flags & SIF_INITDOMAIN)
> > > > > > > > 
> > > > > > > > But anyone who calls this before xen_start_info is setup is going to get
> > > > > > > > a bogus result, specifically in this case they will think they are domU
> > > > > > > > when in reality they are dom0 -- wouldn't it be better to fix those
> > > > > > > > callsites?
> > > > > > > 
> > > > > > > That cannot be the case because setting up xen_start_info is the very
> > > > > > > first thing that is done, before even calling to C.
> > > > > > 
> > > > > > On PV, yes, but you are trying to fix PVHVM here, no?
> > > > > > 
> > > > > > Otherwise if this is always set before calling into C then what is the
> > > > > > purpose of this patch?
> > > > > 
> > > > > to fix this - as PVHVM has it set to NULL and we end up de-referencing
> > > > > the xen_start_info and crashing. As so::
> > > > > 
> > > > 
> > > > Right, so returning to my original point: The caller here is calling
> > > > xen_initial_domain() *before* start info is setup. This is bogus and is
> > > > your actual bug, all this patch does is hide that real issue.
> > > 
> > > That is because xen_start_info wasn't setup at all for PV on HVM guests.
> > > 
> > > The real reason is that PV on HVM guests don't have one, but that is
> > > another matter. Until we get rid of all the references to xen_start_info
> > > outside of PV specific code, we should just assume that there is one,
> > > and that is already setup.
> > > 
> > > One day not too far from now, we might refactor the code to never
> > > reference xen_start_info directly, but I don't think that now is the
> > > time for that. Also consider that this is the same thing we do on ARM.
> > 
> > We actual fill in the dummy start info with valid information on ARM
> > though, we don't just leave it full of zeroes.
> > 
> > If we do start out with start_info pointing to an uninitialised
> > start_info on ARM too then I would argue that this is also a mistake.
> 
> Yes, we do point xen_start_info to an uninitialised start_info on ARM
> too (I don't think is a mistake). Then when and if we have more
> information we write them to start_info.

So callers of xen_initial_domain in dom0 before xen_guest_init is called
get the wrong result?

That sounds like a mistake to me.

> > We
> > should leave the NULL pointer in place until we setup the content of the
> > dummy start info -- exactly because the resulting crash indicates to us
> > that someone has accessed the si before we've initialised it.
> 
> I don't think so. It is initialized to zero, that is the right thing to
> do.

Except it isn't in the dom0 case...

> > > > With this "fix" the caller of xen_initial_domain shown in this trace now
> > > > gets a rubbish result based on the content of a dummy shared info
> > > > instead of the real answer from that actual shared info.
> > > 
> > > That is not true. The caller gets a zero result, that is completely
> > > appropriate in this case, given that a PV on HVM guest doesn't have a
> > > start_info.
> > 
> > It's just a side effect of Linux zeroing its bss though and zero
> > happening to be the right answer for a PVHVM guest in this case.
> 
> well, I would call that "by design" ;-)

Well, in that case it should be documented not just implicit!

> > Is it true that zero is an appropriate result for all uses of fields in
> > start_info on PVHVM?
> 
> I think so. In fact, if we wanted to, we could have the dummy struct
> initialized to something different, but I don't think that we should.