Re: [RFC PATCH 10/21] treewide: Drop function_nocfi

[Mark Rutland]

On Thu, May 05, 2022 at 09:51:39AM -0700, Sami Tolvanen wrote:
> On Thu, May 5, 2022 at 9:30 AM Mark Rutland <mark.rutland@xxxxxxx> wrote:
> > I also believe that in most cases we can drop the __nocfi annotation on callers
> > now that we can mark the called assembly function with SYM_TYPED_FUNC_START().
> 
> Good point, thanks for pointing that out. I'll add these to the next
> version of the series.

Also, I *think* we can drop __nocfi from __init, and always check calls to
functions in .init.text. IIUC we made those __nocfi because it leads to section
mismatches, and dangling entries in the jump tables after we discarded the init
text, neither of which should be a problem with kCFI.

Unfortuantely, that appears to be masking some existing type mismatches; e.g.
psci_dt_init() blows up because it uses the wrong type for its callees (a
mismatched `const`). With that fixed up, arm64 boots fine.

> > There' a latent bug here with the existing CFI scheme, since
> > `kpti_install_ng_mappings` isn't marked with __nocfi, and should explode when
> > calling `idmap_kpti_install_ng_mappings` via the idmap.
> 
> The CONFIG_UNMAP_KERNEL_AT_EL0 version of kpti_install_ng_mappings is
> marked __nocfi

Ah, so it is. Sorry for the noise!

> > There' a latent bug here with the existing CFI scheme, since
> > `machine_kexec` isn't marked with __nocfi, and should explode when calling
> > `cpu_soft_restart` via the idmap.
> 
> But it's indeed missing from this one.

Cool; I'll prep a patch that fixes just this, then.

Thanks,
Mark.