On 10/9/19 3:28 PM, Jan Beulich wrote: > On 09.10.2019 16:14, George Dunlap wrote: >> On 10/9/19 11:23 AM, Jürgen Groß wrote: >>>> Regardless of the merits of the change Andy wants to see, it's not a one >>>> that should be made during a feature freeze. >>> >>> Indeed. So either we take this patch or we have to revert the patch(es) >>> introducing the regression. >> >> Actually, just chatting with Ian -- the worse issue ATM, AFAICT, is that >> the IOMMU is enabled for a guest which has neither asked for PCI >> devices, nor explicitly enabled it; and he's currently working on a fix >> for that. Once that issue is fixed, then osstest should become >> unblocked again. >> >> It is, arguably, not ideal to refuse to migrate a VM with IOMMU enabled >> but no devices attached; but if it only affected guests who had >> specifically requested the IOMMU be enabled, that wouldn't be so >> terrible. (And in fact it has highlighted the other, more important issue.) >> >> In summary, this patch is not strictly needed to get a push to osstest. >> >> That said, the behavior in 4.12 was, as far as I can tell: >> >> 1. If a guest had never had a PCI device assigned, Xen will allow >> logdirty to be enabled. >> >> 2. If a guest has a PCI device assigned, Xen will not allow logdirty to >> be enabled (blocking migration). >> >> 3. If a guest had a PCI device assigned in the past but does not have >> one now, Xen will also not allow logdirty to be enabled (blocking >> migration). > > No - the connection previously was to whether IOMMU page tables > had been set up; these page tables would have been torn down > upon de-assignment of the last device, allowing migration again. > People actually use this behavior afaik, using a bond of a > passed through SR-IOV NIC and netfront provided device. To > migrate the VM, the SR-IOV NIC is taken away without the domain > losing network access, and a new one might then be assigned > again after migration. > > The "IOMMU page tables set up" property was previously identical > to "has devices assigned", i.e. even before we could have used > has_arch_pdevs() instead of is_iommu_enabled().
I didn't realize that. So yes, this patch fixes a regression over and above the toolstack fix Ian is working on. -George _______________________________________________ Xen-devel mailing list Xen-devel@lists.xenproject.org https://lists.xenproject.org/mailman/listinfo/xen-devel