On Mon, 2017-08-07 at 15:49 +0200, Paolo Bonzini wrote: > On 07/08/2017 15:25, Mihai Donțu wrote: > > > "Pause all VCPUs and stop all DMA" would definitely be a layering > > > violation, so it cannot be added. > > > > > > "Pause all VCPUs" is basically a shortcut for many "pause the VCPU with > > > a given id" commands. I lean towards omitting it. > > > > The case where the introspector wants to scan the guest memory needs a > > KVMI_PAUSE_VM, which as discussed in a previous email, can be the > > actual qemu 'pause' command. > > Do you mean it needs to stop DMA as well? No, DMA can proceed normally. I remain of the opinion that KVMI users must know what guest memory ranges are OK to access by looking at MTRR- s, PAT or guest kernel structures, or a combination of all three. > > However, we would like to limit the > > communication channels we have with the host and not use qmp (or > > libvirt/etc. if qmp is not exposed). Instead, have a command that > > triggers a KVM_RUN exit to qemu which in turn will call the underlying > > pause function used by qmp. Would that be OK with you? > > You would have to send back something on completion, and then I am > worried of races and deadlocks. Plus, pausing a VM at the QEMU level is > a really expensive operation, so I don't think it's a good idea to let > the introspector do this. You can pause all VCPUs, or use memory page > permissions. Pausing all vCPU-s was my first thought, I was just trying to follow your statement: "I lean towards omitting it". :-) It will take a bit of user-space-fu, in that after issuing N vCPU pause commands in a row we will have to wait for N events, which might race with other events (MSR, CRx etc.) which need handling otherwise the pause ones will not arrive ... I wonder if there's a way to do this cleanly in kernel (piggyback on the code for pausing a single vCPU and then somehow 'coalesce' all pause events into a single KVMI_EVENT_VCPUS_PAUSED). > > > However, now that I'm thinking of it, we need a new event for "new VCPU > > > created". When the event is enabled, newly-created VCPUs should be in > > > paused mode. > > > > I assume you are thinking about vCPU hotplug here. If so, yes, an event > > that gives the introspector the chance to update its internal > > bookkeeping would be useful. > > Yes, exactly. OK. We will update the design document. -- Mihai Donțu
