Scalable Work Submission in Device Virtualization
summary
Claim: using the EQ command lets guest applications submit work directly to shared device queues while hardware and VMX translate guest PASIDs to host PASIDs so DMA uses correct context. It requires device support for deferred memory writes (DSA is first), IA32_PASID MSR, and a PSID translation table managed by KVM (updates force all vCPUs into root mode). Native kernel EQ support is merged, the driver and PASID virtualization patches are under review, and live migration support is not yet implemented.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:30Agenda
- 1:33Scalability In Device Virtualization
- 4:16Shared Work Queue In Device Virtualization
- 5:25Challenge In Device Virtualization
- 6:34ENQCMD Instruction - Overview
- 9:35ENOCMD Instruction - Device Requirement
- 10:42ENOCMD Virtualization - Non-Root Mode Operation
- 11:55ENQCMD Virtualization - PASID Translation Support
- 13:21Manage PASID Translation Table in KVM
- 17:31Example - SVA Work Submission In Guest • Prepare a Work Descriptor
- 19:56References
- 20:26Status
- 21:28Summary
related talks
20mPassthrough/Headless GPU Gets Ahead
21mNVMe Emulation Performance Optimization
14mAIE-ML dev PMD for VRAN and UPF Datapath Acceleration
26mAnalysis of AMD HW-assisted vIOMMU Implementation and Performance
40mAchieving the Ultimate Performance with KVM
20mMulti-Queue Fair Queuing
33mGPU Para-Virtualization on Xen
25mExploring Intel Sub-Page Protection in Context of Live Migration