On Sat, Dec 09, 2017 at 04:23:17PM +0000, Wang, Wei W wrote: > On Friday, December 8, 2017 4:34 PM, Stefan Hajnoczi wrote: > > On Fri, Dec 8, 2017 at 6:43 AM, Wei Wang <wei.w.w...@intel.com> wrote: > > > On 12/08/2017 07:54 AM, Michael S. Tsirkin wrote: > > >> > > >> On Thu, Dec 07, 2017 at 06:28:19PM +0000, Stefan Hajnoczi wrote: > > >>> > > >>> On Thu, Dec 7, 2017 at 5:38 PM, Michael S. Tsirkin <m...@redhat.com> > > > Thanks Stefan and Michael for the sharing and discussion. I think > > > above 3 and 4 are debatable (e.g. whether it is simpler really > > > depends). 1 and 2 are implementations, I think both approaches could > > > implement the device that way. We originally thought about one device > > > and driver to support all types (called it transformer sometimes :-) > > > ), that would look interesting from research point of view, but from > > > real usage point of view, I think it would be better to have them > > > separated, > > because: > > > - different device types have different driver logic, mixing them > > > together would cause the driver to look messy. Imagine that a > > > networking driver developer has to go over the block related code to > > > debug, that also increases the difficulty. > > > > I'm not sure I understand where things get messy because: > > 1. The vhost-pci device implementation in QEMU relays messages but has no > > device logic, so device-specific messages like VHOST_USER_NET_SET_MTU are > > trivial at this layer. > > 2. vhost-user slaves only handle certain vhost-user protocol messages. > > They handle device-specific messages for their device type only. This is > > like > > vhost drivers today where the ioctl() function returns an error if the > > ioctl is > > not supported by the device. It's not messy. > > > > Where are you worried about messy driver logic? > > Probably I didn’t explain well, please let me summarize my thought a little > bit, from the perspective of the control path and data path. > > Control path: the vhost-user messages - I would prefer just have the > interaction between QEMUs, instead of relaying to the GuestSlave, because > 1) I think the claimed advantage (easier to debug and develop) doesn’t seem > very convincing
You are defining a mapping from the vhost-user protocol to a custom virtio device interface. Every time the vhost-user protocol (feature bits, messages, etc) is extended it will be necessary to map this new extension to the virtio device interface. That's non-trivial. Mistakes are possible when designing the mapping. Using the vhost-user protocol as the device interface minimizes the effort and risk of mistakes because most messages are relayed 1:1. > 2) some messages can be directly answered by QemuSlave , and some messages > are not useful to give to the GuestSlave (inside the VM), e.g. fds, > VhostUserMemoryRegion from SET_MEM_TABLE msg (the device first maps the > master memory and gives the offset (in terms of the bar, i.e., where does it > sit in the bar) of the mapped gpa to the guest. if we give the raw > VhostUserMemoryRegion to the guest, that wouldn’t be usable). I agree that QEMU has to handle some of messages, but it should still relay all (possibly modified) messages to the guest. The point of using the vhost-user protocol is not just to use a familiar binary encoding, it's to match the semantics of vhost-user 100%. That way the vhost-user software stack can work either in host userspace or with vhost-pci without significant changes. Using the vhost-user protocol as the device interface doesn't seem any harder than defining a completely new virtio device interface. It has the advantages that I've pointed out: 1. Simple 1:1 mapping for most that is easy to maintain as the vhost-user protocol grows. 2. Compatible with vhost-user so slaves can run in host userspace or the guest. I don't see why it makes sense to define new device interfaces for each device type and create a software stack that is incompatible with vhost-user. > > > Data path: that's the discussion we had about one driver or separate driver > for different device types, and this is not related to the control path. > I meant if we have one driver for all the types, that driver would look > messy, because each type has its own data sending/receiving logic. For > example, net type deals with a pair of tx and rx, and transmission is skb > based (e.g. xmit_skb), while block type deals with a request queue. If we > have one driver, then the driver will include all the things together. I don't understand this. Why would we have to put all devices (net, scsi, etc) into just one driver? The device drivers sit on top of the vhost-pci driver. For example, imagine a libvhost-user application that handles the net device. The vhost-pci vfio driver would be part of libvhost-user and the application would only emulate the net device (RX and TX queues). Stefan
signature.asc
Description: PGP signature