Device Keepalive State for Local Live Migration and VMM Fast Restart
summary
Adds a VFIO-managed device keepalive state so PCI pass-through devices survive a host kexec and enable local live migration and VMM fast restart. Implementation preserves posted interrupt descriptors and IRTEs, saves IOMMU domain and context tables, uses a vfio token for ownership, and puts guest RAM into DAX persistent memory so qemu skips copying. Adds qemu commands to set/clear keepalive. PoC on Haswell/Broadwell restores YouTube streaming and SCP workloads after kexec.
AI-written summary · about
chapters (16)
- 0:00KVM FORUM
- 0:42Recap of Problem
- 1:55Solution Spectrum & Where is VMM Fast Restart intel.
- 3:50Overview
- 4:07Device Keepalive State
- 5:20Two Incremental Stages Stage-1 Keepalive States
- 6:43Example Commands for VMM Fast Restart intel.
- 8:24What to Preserve, What to Re-create
- 10:36Keep Alive States
- 10:50Keep IRQ Alive
- 13:06Preserve/Restore pi_desc
- 14:09Preserve/Restore IRTE
- 15:46Keep DMA Alive
- 17:03VFIO Device Ownership
- 18:59PCI Enumeration
- 21:13Status & Future Plan
related talks
33mPreserving IOMMU States During Kexec Reboot
24mSupport for Fast and Reliable VMM Live Upgrades in Libvirt
14mDeep Optimization of VMM Live Upgrade
10mTowards a More Efficiently Synchronization in KVM
22mQEMU/KVM Upgrade Test (Stable Guest ABI and in Place Upgrade)
29mPostcopy Preemption
25mThe Challenges of Asynchronous Teardown of KVM Guests
31mLive Update for Xenstore Stubdom