On Wed, 2024-05-29 at 16:07 -0700, Sean Christopherson wrote: > On Wed, May 29, 2024, Kai Huang wrote: > > On Wed, 2024-05-29 at 08:01 -0700, Sean Christopherson wrote: > > > Enabling virtualization should be entirely transparent to userspace, > > > at least from a functional perspective; if changing how KVM enables virtualization > > > breaks userspace then we likely have bigger problems. > > > > I am not sure how should I interpret this? > > > > "having a module param" doesn't necessarily mean "entirely transparent to > > userspace", right? :-) > > Ah, sorry, that was unclear. By "transparent to userspace" I meant the > functionality of userspace VMMs wouldn't be affected if we add (or delete) a > module param. E.g. QEMU should work exactly the same regardless of when KVM > enables virtualization. > > > > Performance is secondary for me, the primary motivation is simplifying the overall > > > KVM code base. Yes, we _could_ use on_each_cpu() and enable virtualization > > > on-demand for TDX, but as above, it's extra complexity without any meaningful > > > benefit, at least AFAICT. > > > > Either way works for me. > > > > I just think using a module param to resolve some problem while there can > > be solution completely in the kernel seems overkill :-) > > The module param doesn't solve the problem, e.g. we could solve this entirely > in-kernel simply by having KVM unconditionally enable virtualization during > initialization. The module param is mostly there to continue playing nice with > out-of-tree hypervisors, and to a lesser extent to give us a "break in case of > fire" knob. OK. Now I understand you want to make enabling virtualization at loading time as default behaviour, but make the module param as a way to defer to enable virt when creating the first VM.
