| Message ID | 20180906133228.118282-1-liran.alon@oracle.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | : KVM: nVMX: Consider TLB are tagged with different EPTP if L1 uses EPT | expand |
On 06/09/2018 15:32, Liran Alon wrote: > Hi, > > This patch series aim to enhance L1<->L2 transitions to avoid > unnecessary TLB flushes in case L1 uses VPID and EPT. > > The 1st patch fixes a bug that INVEPT was executed on L1<->L2 > transitions to flush TLB but with the wrong EPTP because it hasn't been > loaded yet with the destination EPTP. The patch handles this issue > easily by just delaying vmx_flush_tlb() to be called after EPTP has been > set to destination by using KVM_REQ_TLB_FLUSH. > > The 2nd patch avoids unnecessary flushing TLB on L2->L1 VMExit if L1 > uses VPID and EPT. This is because TLB entries of L1 and L2 are > separated by the fact they are tagged with different EPTP. > > The 3rd patch modifies L1->L2 VMEntry code such that it will set > vmcs02->vpid to vmcs12->vpid in case L1 uses EPT. This leverages the > fact TLB entries of L1 and L2 are separated by being tagged with > different EPTPs and therefore we don't need to use the special vpid02. > > The 4rd patch then modifies L1->L2 VMEntry code to avoid unnecessary > flushing TLB in case L1 uses both VPID and EPT as TLB entries of L1 and > L2 are separated by different EPTP tags. Queued patches 1-2-4, thanks. Paolo
> On 1 Oct 2018, at 15:56, Paolo Bonzini <pbonzini@redhat.com> wrote: > > On 06/09/2018 15:32, Liran Alon wrote: >> Hi, >> >> This patch series aim to enhance L1<->L2 transitions to avoid >> unnecessary TLB flushes in case L1 uses VPID and EPT. >> >> The 1st patch fixes a bug that INVEPT was executed on L1<->L2 >> transitions to flush TLB but with the wrong EPTP because it hasn't been >> loaded yet with the destination EPTP. The patch handles this issue >> easily by just delaying vmx_flush_tlb() to be called after EPTP has been >> set to destination by using KVM_REQ_TLB_FLUSH. >> >> The 2nd patch avoids unnecessary flushing TLB on L2->L1 VMExit if L1 >> uses VPID and EPT. This is because TLB entries of L1 and L2 are >> separated by the fact they are tagged with different EPTP. >> >> The 3rd patch modifies L1->L2 VMEntry code such that it will set >> vmcs02->vpid to vmcs12->vpid in case L1 uses EPT. This leverages the >> fact TLB entries of L1 and L2 are separated by being tagged with >> different EPTPs and therefore we don't need to use the special vpid02. >> >> The 4rd patch then modifies L1->L2 VMEntry code to avoid unnecessary >> flushing TLB in case L1 uses both VPID and EPT as TLB entries of L1 and >> L2 are separated by different EPTP tags. > > Queued patches 1-2-4, thanks. > > Paolo Thanks. However, because we haven’t taken patch 3 of the series, I think we should also remove patch 4 from kvm.git/queue. Otherwise, it introduce a bug. This dependency is documented on commit message :) We can replace patch 4 with a simpler one that in case vmx->nested.vpid02 isn’t allocated, we need to vmx_flush_tlb() only in case (!nested_cpu_has_ept(vmcs12) || (vmcs12->virtual_processor_id != vmx->nested.last_vpid)). I will create such a patch and submit it separately. -Liran