Re: [PATCH] arch/powerpc/kvm: kvmppc_core_vcpu_create_hv: check for kzalloc failure

[Kautuk Consul]

On 2023-03-28 23:02:09, Michael Ellerman wrote:
> Kautuk Consul <kconsul@xxxxxxxxxxxxxxxxxx> writes:
> > On 2023-03-28 15:44:02, Kautuk Consul wrote:
> >> On 2023-03-28 20:44:48, Michael Ellerman wrote:
> >> > Kautuk Consul <kconsul@xxxxxxxxxxxxxxxxxx> writes:
> >> > > kvmppc_vcore_create() might not be able to allocate memory through
> >> > > kzalloc. In that case the kvm->arch.online_vcores shouldn't be
> >> > > incremented.
> >> > 
> >> > I agree that looks wrong.
> >> > 
> >> > Have you tried to test what goes wrong if it fails? It looks like it
> >> > will break the LPCR update, which likely will cause the guest to crash
> >> > horribly.
> > Also, are you referring to the code in kvmppc_update_lpcr()?
> > That code will not crash as it checks for the vc before trying to
> > dereference it.
> 
> Yeah that's what I was looking at. I didn't mean it would crash, but
> that it would bail out early when it sees a NULL vcore, leaving other
> vcores with the wrong LPCR value.
> 
> But as you say it doesn't happen because qemu quits on the first ENOMEM.
> 
> And regardless if qemu does something that means the guest is broken
> that's just a qemu bug, no big deal as far as the kernel is concerned.
But there could be another user-mode application other than qemu that
actually tries to create a vcpu after it gets a -ENOMEM for another
vcpu. Shouldn't the kernel be independent of qemu?
> 
> > But the following 2 places that utilize the arch.online_vcores will have
> > problems in logic if the usermode test-case doesn't pull down the
> > kvm context after the -ENOMEM vcpu allocation failure:
> > book3s_hv.c:3030:       if (!kvm->arch.online_vcores) {
> > book3s_hv_rm_mmu.c:44:  if (kvm->arch.online_vcores == 1 && local_paca->kvm_hstate.kvm_vcpu)
> 
> OK. Both of those look harmless to the host.
Harmless to the host in terms of a crash, not in terms of behavior.
For example in the case of kvmhv_set_smt_mode:
If we got a kzalloc failure once (and online_vcores was wrongly incremented), 
then if kvmhv_set_smt_mode() is called after that then it would be
not be setting the arch.smt_mode and arch.emul_smt_mode correctly and it
would be wrongly returning with -EBUSY instead of 0.
Isn't that incorrect with respect to the intent of the code ?
I agree that applications like qemu might not do that but don't we need
to have some integrity with respect to the intent and value of variable
use ? What about good code and logic quality ?
> 
> If we find a case where a misbehaving qemu can crash the host then we
> need to be a bit more careful and treat it at least as a
> denial-of-service bug. But looks like this is not one of those.
> 
> cheers
beers