Re: 3.0.2-testing: pci_set_dma_mask, pci_set_consistent_dma_mask(pci, 0x0fffffff) returns < 0 (ICE1712)

["Stephen C. Tweedie" <sct@redhat.com>]

Hi,

On Thu, 2006-04-13 at 08:46 +0100, Keir Fraser wrote:

> > If the Xen HV runs out of MEMZONE_DMADOM pages, aren't we basically out
> > of luck right now?  Xen guests can't see that shortage, nor does the
> > vmscan.c code have any code to target pages for stealing based on MFN
> > rather than PFN.

> Yes, if that happens then we could be in trouble. Although mostly low 
> memory is allocated for devices only at start of day. 

DomU creation too, if we're using driver domains, I guess.  But that's a
minority case right now.

> The main fly in 
> the ointment is PAE pgds. We get around that right now by reserving a 
> lowmem pool in Xen that normal allocations cannot fall back to.

Right, although that pool defaults to empty unless you manually
configure it.  It would be good if we could avoid the fragility of
requiring custom intervention to set this up.

>  That 
> sorts out most kinds of bad behaviour. We would need the guest kernels 
> to help out with reclamation (at least for guests not on shadow page 
> tables). If Linux had support for hotplug memory (add and remove) I 
> suspect we could make use of that to help out, if we were careful.

Or simply get the balloon driver to target the appropriate pages.  The
problem with using hotplug memory is that that generally works on the
basis of ranges of physical addresses, not machine addresses.  

--Stephen