On 11/30/20 4:48 PM, David Woodhouse wrote: > On Mon, 2020-11-30 at 15:08 +0000, Joao Martins wrote: >> On 11/30/20 12:55 PM, David Woodhouse wrote: >>> On Mon, 2020-11-30 at 12:17 +0000, Joao Martins wrote: >>>> On 11/30/20 9:41 AM, David Woodhouse wrote: >>>>> On Wed, 2019-02-20 at 20:15 +0000, Joao Martins wrote: >>>> One thing I didn't quite do at the time, is the whitelisting of unregistered >>>> ports to userspace. Right now, it's a blacklist i.e. if it's not handled in >>>> the kernel (IPIs, timer vIRQ, etc) it goes back to userspace. When the only >>>> ones which go to userspace should be explicitly requested as such >>>> and otherwise return -ENOENT in the hypercall. >>> >>> Hm, why would -ENOENT be a fast path which needs to be handled in the >>> kernel? >>> >> >> It's not that it's a fast path. >> >> Like sending an event channel to an unbound vector, now becomes an possible vector to >> worry about in userspace VMM e.g. should that port lookup logic be fragile. >> >> So it's more along the lines of Nack-ing the invalid port earlier to rather go >> to go userspace to invalidate it, provided we do the lookup anyway in the kernel. > > If the port lookup logic is fragile, I *want* it in the sandboxed > userspace VMM and not in the kernel :) > Yes definitely -- I think we are on the same page on that. But it's just that we do the lookup *anyways* to check if the kernel has a given evtchn port registered. That's the lookup I am talking about here, with just an extra bit to tell that's a userspace handled port. > And unless we're going to do *all* of the EVTCHNOP_bind*, EVTCHN_close, > etc. handling in the kernel, doesn't userspace have to have all that > logic for managing the port space anyway? > Indeed. > I think it's better to let userspace own it outright, and use the > kernel bypass purely for the fast paths. The VMM can even implement > IPI/VIRQ support in userspace, then use the kernel bypass if/when it's > available. > True, and it's pretty much how it's implemented today. But felt it was still worth having this discussion ... should this be considered or discarded. I suppose we stick with the later for now. >>>> Perhaps eventfd could be a way to express this? Like if you register >>>> without an eventfd it's offloaded, otherwise it's assigned to userspace, >>>> or if neither it's then returned an error without bothering the VMM. >>> >>> I much prefer the simple model where the *only* event channels that the >>> kernel knows about are the ones it's expected to handle. >>> >>> For any others, the bypass doesn't kick in, and userspace gets the >>> KVM_EXIT_HYPERCALL exit. >>> >> >> /me nods >> >> I should comment on your other patch but: if we're going to make it generic for >> the userspace hypercall handling, might as well move hyper-v there too. In this series, >> I added KVM_EXIT_XEN, much like it exists KVM_EXIT_HYPERV -- but with a generic version >> I wonder if a capability could gate KVM_EXIT_HYPERCALL to handle both guest types, while >> disabling KVM_EXIT_HYPERV. But this is probably subject of its own separate patch :) > > There's a limit to how much consolidation we can do because the ABI is > different; the args are in different registers. > Yes. It would be optionally enabled of course and VMM would have to adjust to the new ABI -- surely wouldn't want to break current users of KVM_EXIT_HYPERV. > I do suspect Hyper-V should have marshalled its arguments into the > existing kvm_run->arch.hypercall and used KVM_EXIT_HYPERCALL but I > don't think it makes sense to change it now since it's a user-facing > ABI. I don't want to follow its lead by inventing *another* gratuitous > exit type for Xen though. > I definitely like the KVM_EXIT_HYPERCALL better than a KVM_EXIT_XEN userspace exit type ;) But I guess you still need to co-relate a type of hypercall (Xen guest cap enabled?) to tell it's Xen or KVM to specially enlighten certain opcodes (EVTCHNOP_send). Joao
