Re: [RFC v2-fix 1/1] x86/tdx: Handle in-kernel MMIO

[Sean Christopherson]

On Tue, May 18, 2021, Andi Kleen wrote:
> 
> On 5/18/2021 11:22 AM, Sean Christopherson wrote:
> > On Tue, May 18, 2021, Andi Kleen wrote:
> > > > The extra bytes for .altinstructions is very different than the extra bytes for
> > > > the code itself.  The .altinstructions section is freed after init, so yes it
> > > > bloats the kernel size a bit, but the runtime footprint is unaffected by the
> > > > patching metadata.
> > > > 
> > > > IIRC, patching read/write{b,w,l,q}() can be done with 3 bytes of .text overhead.
> > > > 
> > > > The other option to explore is to hook/patch IO_COND(), which can be done with
> > > > neglible overhead because the helpers that use IO_COND() are not inlined.  In a
> > > > TDX guest, redirecting IO_COND() to a paravirt helper would likely cover the
> > > > majority of IO/MMIO since virtio-pci exclusively uses the IO_COND() wrappers.
> > > > And if there are TDX VMMs that want to deploy virtio-mmio, hooking
> > > > drivers/virtio/virtio_mmio.c directly would be a viable option.
> > > Yes but what's the point of all that?
> > Patching IO_COND() is relatively low effort.  With some clever refactoring, I
> > suspect the net lines of code added would be less than 10.  That seems like a
> > worthwhile effort to avoid millions of faults over the lifetime of the guest.
> 
> AFAIK IO_COND is only for iomap users. But most drivers don't even use
> iomap. virtio doesn't for example, and that's really the only case we
> currently care about.

virtio-pci, which is going to used by pretty much all traditional VMs, uses iomap.
See vp_get(), vp_set(), and all the vp_io{read,write}*() wrappers.