Re: I cannot get any message from domU by console / pv_ops domU kernel crashes with xen_create_contiguous_region failed

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

On Tue, Dec 22, 2009 at 02:58:47PM +0000, Ian Campbell wrote:
> On Tue, 2009-12-22 at 14:32 +0000, Konrad Rzeszutek Wilk wrote:
> > > >    (early) [    0.000000] Kernel panic - not syncing:
> > > >    <3>xen_create_contiguous_region failed
> > 
> > This implies you did not have enough DMA32 memory in the hypervisor for the guest.
> > Meaning at least 64MB.
> 
> Does this mean that every domU is allocating 64M for swiotlb by default?

Unfortunately yes. We can remove the:

ommit 4bc6b1a9dd5d7447af8c3d27c1449f73f5f764ec
Author: root <root@localhost.localdomain>
Date:   Thu Nov 5 16:33:10 2009 -0500

    Enable Xen-SWIOTLB if running in [non-]privileged and disable the Xen-IOMMU if an IOMMU is detected.
    
    For PCI passthrough to work correctly, we need the Xen-SWIOTLB.
    Otherwise PCI devices in the non-privileged domains might not be
    able to do DMA.
    
And revert to the old-style behaviour.

> That seems like an awful lot of wastage when the vast majority of guests
> have no I/O device and therefore no need for swiotlb, especially given
> that this is memory from a limited pool (OK, so 64G isn't that low a
> limit).
> 
> If we are unable to automatically determine early enough whether a guest
> needs swiotlb or not (which I suspect is the case without additional

That is possible. Which is why I am working on the new SWIOTLB that
can be turned on _after_ xen-pcifront has been activated.

> tools support) then I think the old-style behaviour of only enabling
> swiotlb in domU on explicit request matches the common case better.
> 
> Ian.
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel