On Thu, Jul 08, 2021 at 11:53:38AM +0200, David Hildenbrand wrote: > Postcopy never worked properly with 'free-page-hint=on', as there are > at least two issues: > > 1) With postcopy, the guest will never receive a VIRTIO_BALLOON_CMD_ID_DONE > and consequently won't release free pages back to the OS once > migration finishes. > > The issue is that for postcopy, we won't do a final bitmap sync while > the guest is stopped on the source and > virtio_balloon_free_page_hint_notify() will only call > virtio_balloon_free_page_done() on the source during > PRECOPY_NOTIFY_CLEANUP, after the VM state was already migrated to > the destination. > > 2) Once the VM touches a page on the destination that has been excluded > from migration on the source via qemu_guest_free_page_hint() while > postcopy is active, that thread will stall until postcopy finishes > and all threads are woken up. (with older Linux kernels that won't > retry faults when woken up via userfaultfd, we might actually get a > SEGFAULT) > > The issue is that the source will refuse to migrate any pages that > are not marked as dirty in the dirty bmap -- for example, because the > page might just have been sent. Consequently, the faulting thread will > stall, waiting for the page to be migrated -- which could take quite > a while and result in guest OS issues. > > While we could fix 1) comparatively easily, 2) is harder to get right and > might require more involved RAM migration changes on source and destination > [1]. > > As it never worked properly, let's not start free page hinting in the > precopy notifier if the postcopy migration capability was enabled to fix > it easily. Capabilities cannot be enabled once migration is already > running. > > Note 1: in the future we might either adjust migration code on the source > to track pages that have actually been sent or adjust > migration code on source and destination to eventually send > pages multiple times from the source and and deal with pages > that are sent multiple times on the destination. > > Note 2: virtio-mem has similar issues, however, access to "unplugged" > memory by the guest is very rare and we would have to be very > lucky for it to happen during migration. The spec states > "The driver SHOULD NOT read from unplugged memory blocks ..." > and "The driver MUST NOT write to unplugged memory blocks". > virtio-mem will move away from virtio_balloon_free_page_done() > soon and handle this case explicitly on the destination. > > [1] https://lkml.kernel.org/r/e79fd18c-aa62-c1d8-c7f3-ba3fc2c25...@redhat.com > > Fixes: c13c4153f76d ("virtio-balloon: VIRTIO_BALLOON_F_FREE_PAGE_HINT") > Cc: qemu-sta...@nongnu.org > Cc: Wei Wang <wei.w.w...@intel.com> > Cc: Michael S. Tsirkin <m...@redhat.com> > Cc: Philippe Mathieu-Daudé <phi...@redhat.com> > Cc: Alexander Duyck <alexander.du...@gmail.com> > Cc: Juan Quintela <quint...@redhat.com> > Cc: "Dr. David Alan Gilbert" <dgilb...@redhat.com> > Cc: Peter Xu <pet...@redhat.com> > Signed-off-by: David Hildenbrand <da...@redhat.com>
Reviewed-by: Peter Xu <pet...@redhat.com> Thanks, David. -- Peter Xu
