On Sat, Apr 03, 2010 at 10:44:22AM -0700, Eric W. Biederman wrote: > Joerg Roedel <joro at 8bytes.org> writes: > > Hmm, I think for this we need to change the gart code too and disable > > the gart before its initialization runs to not re-introduce issues fixed > > in commit bc2cea6a34fdb30f118ec75db39a46a191870607, no? > > That is a different code path with a different set of assumptions and > restrictions. On a normal kexec of course we want to do an orderly shutdown. Thats another problem with this patch. It introduces a difference between the panic-shutdown kexec and the ordinary kexec. > For the gart with a little luck we can just ignore it on kexec on > panic. The commit I mentioned above already proves this assumption wrong. > Unlike a virtualization capable iommu it doesn't prevent access > to devices, when it is enabled. Worst case is that we have to start > including iommu=off for gart systems. No no no. This is a maintenance nightmare for almost everybody. Where do you want to Document this special cases that 'if kernel uses gart then and only then boot the kexec kernel with iommu=off'. Always passing iommu=off to the kexec kernel doesn't work too for obvious reasons. > The best case is that we can figure out how to have the gart code > reinitialize itself sanely, starting from some arbitrary point. Yes, that is missing in this patch. But to keep changes small and don't bother with the gart code at all I suggest to remove the shutdown routine from the amd-iommu code only and not the whole shutdown call in the machine_crash_shutdown path. Joerg
