On 27.02.2018 10:40, David Hildenbrand wrote: > >> spin_lock_init(&kvm->arch.start_stop_lock); >> diff --git a/arch/s390/kvm/priv.c b/arch/s390/kvm/priv.c >> index 76a2380..d9bd147 100644 >> --- a/arch/s390/kvm/priv.c >> +++ b/arch/s390/kvm/priv.c >> @@ -208,19 +208,23 @@ int kvm_s390_skey_check_enable(struct kvm_vcpu *vcpu) >> struct kvm_s390_sie_block *sie_block = vcpu->arch.sie_block; >> >> trace_kvm_s390_skey_related_inst(vcpu); >> - if (!(sie_block->ictl & (ICTL_ISKE | ICTL_SSKE | ICTL_RRBE)) && >> + /* Already enabled? */ >> + if (vcpu->kvm->arch.use_skf && >> + !(sie_block->ictl & (ICTL_ISKE | ICTL_SSKE | ICTL_RRBE)) && >> !kvm_s390_test_cpuflags(vcpu, CPUSTAT_KSS)) >> return rc; > > Wonder if it is nicer to simply remember for each CPU if this function > has already been called. This way we can avoid calling > s390_enable_skey() in all configurations. > With the benefit being, that on a per cpu storage we wouldn't have to lock the call indication, as we are the thread of this vcpu and only we can alter it? Which would speed up hpage and VSIE cases a tad, as the down_write is quite heavy. We could also do a down_read, store enablement status, up_read, if !enabled -> 390_enable_skey(). I'm also wondering, if we have a clean way after migrating a skey using guest to disable interception right away when we get the skeys from QEMU. I don't think we do that yet.
Attachment:
signature.asc
Description: OpenPGP digital signature
