Maxim Levitsky <mlevitsk@xxxxxxxxxx> writes: > On Mon, 2022-06-06 at 10:36 +0200, Vitaly Kuznetsov wrote: >> To allow flushing individual GVAs instead of always flushing the >> whole >> VPID a per-vCPU structure to pass the requests is needed. Use >> standard >> 'kfifo' to queue two types of entries: individual GVA (GFN + up to >> 4095 >> following GFNs in the lower 12 bits) and 'flush all'. > > Honestly I still don't think I understand why we can't just > raise KVM_REQ_TLB_FLUSH_GUEST when the guest uses this interface > to flush everthing, and then we won't need to touch the ring > at all. The main reason is that we need to know what to flush: L1 or L2. E.g. for VMX, KVM_REQ_TLB_FLUSH_GUEST is basically vpid_sync_context(vmx_get_current_vpid(vcpu)); which means that if the target vCPU transitions from L1 to L2 or vice versa before KVM_REQ_TLB_FLUSH_GUEST gets processed we will flush the wrong VPID. And actually the writer (the vCPU which processes the TLB flush hypercall) is not anyhow synchronized with the reader (the vCPU whose TLB needs to be flushed) here so we can't even know if the target vCPU is in guest more or not. With the newly added KVM_REQ_HV_TLB_FLUSH, we always look at the corresponding FIFO and process 'flush all' accordingly. In case the vCPU switches between modes, we always raise KVM_REQ_HV_TLB_FLUSH request to make sure we check. Note: we can't be raising KVM_REQ_TLB_FLUSH_GUEST instead as it always means 'full tlb flush' and we certainly don't want that. -- Vitaly
