* Tian, Kevin (kevin.t...@intel.com) wrote: > > From: Qemu-devel <qemu-devel-bounces+kevin.tian=intel....@nongnu.org> > > On Behalf Of Dr. David Alan Gilbert > > > > * Daniel P. Berrangé (berra...@redhat.com) wrote: > > > On Thu, Mar 11, 2021 at 12:50:09AM +0530, Tarun Gupta wrote: > > > > Document interfaces used for VFIO device migration. Added flow of state > > changes > > > > during live migration with VFIO device. Tested by building docs with the > > new > > > > vfio-migration.rst file. > > > > > > > > v2: > > > > - Included the new vfio-migration.rst file in index.rst > > > > - Updated dirty page tracking section, also added details about > > > > 'pre-copy-dirty-page-tracking' opt-out option. > > > > - Incorporated comments around wording of doc. > > > > > > > > Signed-off-by: Tarun Gupta <targu...@nvidia.com> > > > > Signed-off-by: Kirti Wankhede <kwankh...@nvidia.com> > > > > --- > > > > MAINTAINERS | 1 + > > > > docs/devel/index.rst | 1 + > > > > docs/devel/vfio-migration.rst | 135 > > ++++++++++++++++++++++++++++++++++ > > > > 3 files changed, 137 insertions(+) > > > > create mode 100644 docs/devel/vfio-migration.rst > > > > > > > > > > +Postcopy > > > > +======== > > > > + > > > > +Postcopy migration is not supported for VFIO devices. > > > > > > What is the problem here and is there any plan for how to address it ? > > > > There's no equivalent to userfaultfd for accesses to RAM made by a > > device. > > There's some potential for this to be doable with an IOMMU or the like, > > but: > > a) IOMMUs and devices aren't currently happy at recovering from > > failures > > b) the fragementation you get during a postcopy probably isn't pretty > > when you get to build IOMMU tables. > > To overcome such limitations one may adopt a prefault-and-pull scheme if > the vendor driver has the capability to track pending DMA buffers in the > migration process (with additional uAPI changes in VFIO or userfaultfd), > as discussed here: > > https://static.sched.com/hosted_files/kvmforum2019/7a/kvm-forum-postcopy-final.pdf
Did that get any further? I can imagine that might be tricikier for a GPU than a network card; the shaders in a GPU are pretty random as to what they go off and access, so I can't see how you could prefault. Dave > > > > > Postcopy is essentially the only migration mechanism that can reliably > > > complete, so it really should be considered the default approach to > > > migration for all mgmt apps wanting to do migration, except in special > > > cases. IOW, if we want VFIO migration to be viable, we need postcopy > > > support. > > > > There's lots of other things postcopy doesn't work with; so hmm. > > > > Agree. Also given the amount of work even for pre-copy migration, it makes > more sense to do things step-by-step. > > Thanks > Kevin > -- Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK
