On 02/05/2019 14:16, Jan Beulich wrote: >>>> On 02.05.19 at 14:59, <andrew.coop...@citrix.com> wrote: >> On 02/05/2019 13:29, Jan Beulich wrote: >>> --- a/xen/drivers/passthrough/pci.c >>> +++ b/xen/drivers/passthrough/pci.c >>> @@ -1450,17 +1450,36 @@ static int assign_device(struct domain * >>> if ( !iommu_enabled || !hd->platform_ops ) >>> return 0; >>> >>> - /* Prevent device assign if mem paging or mem sharing have been >>> - * enabled for this domain */ >>> - if ( unlikely(d->arch.hvm.mem_sharing_enabled || >>> - vm_event_check_ring(d->vm_event_paging) || >>> + domain_lock(d); >>> + >>> + /* >>> + * Prevent device assignment if any of >>> + * - mem paging >>> + * - mem sharing >>> + * - the p2m_ram_ro type >>> + * - global log-dirty mode >> XenServer has working live migration with GPUs, which this change would >> regress. >> >> Behind the scenes, we combine Xen's logdirty bitmap with one provided by >> the GPU, to ensure that all dirty updates are tracked. > But this says nothing for the patch here.
Yes it does. There is nothing inherent about global log-dirty mode which is incompatible with passthrough when the IOMMU pagetables are not shared with EPT. > As long as it doesn't work in the staging tree, it should be prevented. ... but it does work. > In XenServer > you could then patch out that check and comment line together > with whatever other changes you have to make for thins to > work. Everything is upstream in the various projects, other than the vendor's closed source library and kernel driver. ~Andrew _______________________________________________ Xen-devel mailing list Xen-devel@lists.xenproject.org https://lists.xenproject.org/mailman/listinfo/xen-devel
