2.6.38 (FC15) with PCI passthrough fails mysteriously with iommu=soft..

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Wed, Jun 01, 2011 at 07:32:54PM +0200, Laszlo Ersek wrote:
> On 06/01/11 16:31, Konrad Rzeszutek Wilk wrote:
> >On Wed, Jun 01, 2011 at 01:02:02PM +0200, Laszlo Ersek wrote:
> >>(Perhaps it's best to write a separate mail about this.)
> >
> >Yes. You did use 'iommu=soft' on your bootup line right?
> 
> I thought I did; turns out I was wrong. Sorry for the noise.
> 
> Now I tried with "iommu=soft swiotlb=force", but this way the guest

You don't need the swiotlb=force. I wonder what page talks about that?
The Wiki mentions it is only required for older kernels - not the new ones.


> immediately disappears; it doesn't even start to produce console

It probably was panicing b/c it couldn't swizzle out 64MB
of DMA32 memory. You can find that out if you do 'earlyprintk=xenboot' and
that should print out the bootlog in your Xen debug console (if you have
configured guest_loglvl=all).

You can also do 'swiotlb=1024' to lower the amount. .. which should have worked,
except that I found it does not work - so try this patch:

https://lkml.org/lkml/2011/6/1/554

Also, you can set this parameter in you guest file to analyze its stack:

on_crash="preserve"

using xenctx.

> output.
> 
> # xm create -c f15-64bit-pv
> Using config file "/etc/xen/f15-64bit-pv".
> Using <class 'grub.GrubConf.GrubConfigFile'> to parse /grub/menu.lst
> Started domain f15-64bit-pv
> #
> 
> http://wiki.xen.org/xenwiki/XenPCIpassthrough says though:
> 
>     Bugs:
> 
>     # Starting the DomU using pvgrub with 'iommu=soft swiotlb=force'
>     breaks pvgrub.
> 
> Perhaps that could be the reason. Removing "swiotlb=force" and
> keeping only "iommu=soft" crashes (?) the same way.

That was .. a bug that Daniel Kiper fixed at some point. It was all in the
MiniOS. Are you using pygrub or pvgrub?