Re: [Xen-devel] PCI Pass-through in Xen ARM - Draft 2.

[Ian Campbell <ian.campbell@xxxxxxxxxx>]

On Fri, 2015-07-31 at 18:20 +0530, Manish Jaggi wrote:
> 
> On 31/07/15 4:49 pm, Ian Campbell wrote:
> > On Fri, 2015-07-31 at 16:37 +0530, Manish Jaggi wrote:
> > > On Friday 31 July 2015 01:35 PM, Ian Campbell wrote:
> > > > On Fri, 2015-07-31 at 13:16 +0530, Manish Jaggi wrote:
> > > > > > > Secondly, the vdev-X entry is created async by dom0 watching 
> > > > > > > on
> > > > > > > event.
> > > > > > > So how the tools could read back and call assign device 
> > > > > > > again.
> > > > > > Perhaps by using a xenstore watch on that node to wait for the
> > > > > > assignment
> > > > > > from pciback to occur.
> > > > > As per the flow in the do_pci_add function, assign_device is 
> > > > > called
> > > > > first and based on the success xenstore entry is created.
> > > > > Are you suggesting to change the sequence.
> > > > Perhaps that is what it would take, yes, or maybe some other
> > > > refactoring
> > > > (e.g. splitting assign_device into two stages) might be the answer.
> > > The hypercall from xenpciback (what I implemented) is actually making
> > > the assign device in 2 stages.
> > > I think the point of contention is the second stage should be from
> > > toolstack.
> > > 
> > > I think calling xc_assign_device after xenstore from the watch 
> > > callback
> > > is the only option.
> > Only if you ignore the other option I proposed.
> > 
> > > One question is how to split the code for ARM and x86 as this is the
> > > common code.
> > > Would #ifdef CONFIG_ARM64 ok with maintainers.
> > No. arch hooks in libxl_$ARCH.c (with nop implementations where 
> > necessary)
> > would be the way to approach this. However I still am not convinced 
> > this is
> > the approach we should be taking.
> > 
> > > > My current preference is for the suggestion below which is to let 
> > > > the
> > > > toolstack pick the vdevfn and have pciback honour it.
> > > That would duplicate code for dev-fn generation into toolstack from
> > > __xen_pcibk_add_pci_dev.
> > IMHO the toolstack is the correct place for this code, at least for ARM
> > guests. The toolstack is, in general, responsible for all aspects of 
> > the
> > guest layout. I don't think delegating the PCI bus parts of that to the
> > dom0 kernel makes sense.
> Ok, i will implement the same from pciback to toolstack. I am not sure 
> about the complexity but will give it a try.

Thank you.

> With this xen-pciback will not create the vdev-X entry at all.


Uh, where did you get that idea? That node communicates from b.e. to f.e.,
surely it is still needed for that purpose?

The flow seems like it should be:

toolstack -> xenstore:vdevfn-X -> backend -> xenstore:vdev-X -> frontend

where "backend" is codei n pciback which does something like (maybe not in
one function, but logically):

  foo = read(vdevfn-X)
  if (!foo)
     allocate SBDF as code today, setting foo to the result

  write(vdev-X, foo)


IOW if vdevfn-X is unset the backend does the same as today, if it is set
(by the toolstack) then pciback honours it.

Ian.


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel