Kautuk Consul <kcon...@linux.vnet.ibm.com> writes:
> On 2023-03-28 23:02:09, Michael Ellerman wrote:
>> Kautuk Consul <kcon...@linux.vnet.ibm.com> writes:
>> > On 2023-03-28 15:44:02, Kautuk Consul wrote:
>> >> On 2023-03-28 20:44:48, Michael Ellerman wrote:
>> >> > Kautuk Consul <kcon...@linux.vnet.ibm.com> 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?
Yes, the kernel is independent of qemu.
On P8 we had kvmtool, and there's several other VMMs these days, though
most don't support Power.
I didn't mean qemu specifically above. If any VMM continues blindly
after getting ENOMEM back from the KVM API then that's a bug in that
VMM.
>> > 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.
But again that bug only affects that VM, which the VMM should have
terminated when it got the ENOMEM back.
It's definitely a bug that we increment online_vcores incorrectly, but
it only affects that VM, and a correctly operating VMM will terminate
the VM anyway because of the ENOMEM.
cheers
