> Subject: Re: [EXTERNAL] Re: [PATCH v2 0/2] vhost-vdpa: Add support for NO- > IOMMU mode > > On Wed, Oct 23, 2024 at 04: 19: 02AM -0400, Michael S. Tsirkin wrote: > On > Tue, Oct 22, 2024 at 11: 58: 19PM -0700, Christoph Hellwig wrote: > > On Sat, > Oct 19, 2024 at 08: 16: 44PM -0400, Michael S. Tsirkin wrote: > > > Because > > On Wed, Oct 23, 2024 at 04:19:02AM -0400, Michael S. Tsirkin wrote: > > On Tue, Oct 22, 2024 at 11:58:19PM -0700, Christoph Hellwig wrote: > > > On Sat, Oct 19, 2024 at 08:16:44PM -0400, Michael S. Tsirkin wrote: > > > > Because people want to move from some vendor specific solution > > > > with vfio to a standard vdpa compatible one with vdpa. > > > > > > So now you have a want for new use cases and you turn that into a > > > must for supporting completely insecure and dangerous crap. > > > > Nope. > > > > kernel is tainted -> unsupported > > > > whoever supports tainted kernels is already in dangerous waters. > > That's not a carte blanche for doing whatever crazy stuff you want. > > And if you don't trust me I'll add Greg who has a very clear opinion on > IOMMU-bypassing user I/O hooks in the style of the uio driver as well I think > :)
It is going in circles, let me give the summary, Issue: We need to address the lack of no-IOMMU support in the vhost vDPA driver for better performance. Measured Performance: On the machine "13th Gen Intel(R) Core(TM) i9-13900K, 32 Cores", we observed a performance improvement of 70 - 80% with intel_iommu=off when we run high-throughput network packet processing. Rationale for Fix: High-end machines which gives better performance with IOMMU are very expensive, and certain use cases, such as embedded environment and trusted applications, do not require the security features provided by IOMMU. Initial Approach: We initially considered a driver-based solution, specifically integrating no-IOMMU support into Marvell’s octep-vdpa driver. Initial Community Feedback: The community suggested adopting a VFIO-like scheme to make the solution more generic and widely applicable. Decision Point: Should we pursue a generic approach for no-IOMMU support in the vhost vDPA driver, or should we implement a driver-specific solution? Thanks, Srujana.
