Re: target/ppc: sPAPR invalid function calls when compiling without TCG

[David Gibson]

On Tue, Apr 20, 2021 at 06:27:38PM +0000, Lucas Mateus Martins Araujo e Castro 
wrote:
> 
> 
> 
> ________________________________
> 
> >> spapr_hcall.c:
> >>     function h_enter call ppc_hash64_hpte_page_shift_noslb, 
> >> ppc_hash64_map_hptes and ppc_hash64_unmap_hptes
> >>     function remove_hpte call ppc_hash64_map_hptes, ppc_hash64_unmap_hptes 
> >> and ppc_hash64_tlb_flush_hpte
> >>     function h_protect call ppc_hash64_map_hptes, ppc_hash64_unmap_hptes 
> >> and ppc_hash64_tlb_flush_hpte
> >>     function h_read call ppc_hash64_map_hptes and ppc_hash64_unmap_hptes
> >>     function rehash_hpte call ppc_hash64_hpte_page_shift_noslb
> >>     function rehash_hpt call ppc_hash64_map_hptes and 
> >> ppc_hash64_unmap_hptes
> >
> >h_enter, remove_hpte, h_protect, h_read and rehash_hpte should never
> >be called for a KVM guest (KVM will intercept the relevant hcalls
> >before they reach qemu).  I think it would make sense to split of a
> >TCG only portion of spapr_hcall.c into a new file.  You'll then
> >probably need !TCG stubs for those hypercalls, which should trip a
> >fatal error (if KVM hasn't handled the calls, there's nothing we can
> >do that will salvage the situation).
> 
> Besides the aforementioned functions, will all spapr hypercalls
> beginning with H_* be intercepted (and therefore, should trip a
> fatal error in the KVM-only file) or only a subset of these
> functions will be intercepted?

No, only a few hcalls will be KVM intercepted.  There are basically
two categories:
  1) hcalls which *must* be implemented in KVM, because qemu doesn't
     have access to the KVM data structures to implement them.  I
     think this is pretty much just the MMU related hypercalls.
  2) performance critical hcalls where we have a KVM accelerated
     implementation, but *could* be implemented in qemu (and in a few
     cases the kernel fast path will fall back to the qemu slow path
     sometimes)

> If the latter is there an interval of
> opcodes that'll be intercepted?

No.  The opcode numbers aren't defined by KVM or qemu, they're from
from PAPR/PowerVM and have no bearing on how concerns are divided in
the KVM/qemu world.  I'd start by only separating hypercall
implementations which are actively causing you problems (calling into
TCG code is a strong hint that its in category (1)).

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature