Re: [PATCH v2 4/7] kvm/mips: rework guest entry logic

[Mark Rutland]

On Thu, Jan 20, 2022 at 06:29:25PM +0100, Paolo Bonzini wrote:
> On 1/20/22 18:15, Mark Rutland wrote:
> > As above, we'll also need the guest_state_{enter,exit}() calls
> > surrounding this (e.g. before that local_irq_enable() at the start of
> > kvm_mips_handle_exit(),
> 
> Oh, indeed.  And there is also an interrupt-enabled area similar to s390's,
> in both vcpu_run and the exception handler entry point (which falls through
> to the exit handler created by kvm_mips_build_exit).  For example:
> 
>         /* Setup status register for running guest in UM */
>         uasm_i_ori(&p, V1, V1, ST0_EXL | KSU_USER | ST0_IE);
>         UASM_i_LA(&p, AT, ~(ST0_CU0 | ST0_MX | ST0_SX | ST0_UX));
>         uasm_i_and(&p, V1, V1, AT);
>         uasm_i_mtc0(&p, V1, C0_STATUS);
>         uasm_i_ehb(&p);
> 
> I'd rather get rid altogether of the EQS for MIPS.

Ok; I'm not immediately sure how to do that without invasive changes around the
context tracking bits.

Did you have a specific approach in mind, or was that just a general statement?

> > and that needs to happen in noinstr code, etc.
> 
> There are bigger problems with instrumentation, because the
> runtime-generated code as far as I can tell is not noinstr.

The generated sequences themselves are not a problem -- they're not
compiler-instrumented, and kprobes will reject them since they live in a
kzalloc()'d buffer which is outside of kernel text.

Those call tlbmiss_handler_setup_pgd(), but that itself is runtime-generated,
and AFAICT doesn't call anything. It is placed within the kernel text, but it
could be blacklisted from kprobes.

Have I missed something there?

Thanks,
Mark.