On 01/31/2013 09:44 PM, Eric Blake wrote: > On 01/31/2013 12:00 AM, Jason Wang wrote: >> On 01/31/2013 02:29 AM, Eric Blake wrote: >>> On 01/30/2013 04:12 AM, Jason Wang wrote: >>> >>>> With this changes, user could start a multiqueue virtio-net device through >>>> >>>> ./qemu -netdev tap,id=hn0,queues=2,vhost=on -device virtio-net-pci,netdev=hn0 >>>> >>>> Management tools such as libvirt can pass multiple pre-created fds/vhostfds through >>>> >>>> ./qemu -netdev tap,id=hn0,fds=X:Y,vhostfds=M:N -device virtio-net-pci,netdev=hn0 >>> Do we really need specific fds= parsing, or can we reuse the existing >>> -add-fd command line option to our advantage? I guess what I'm asking >>> is how hotplug will work; and if hotplug takes a file name, shouldn't >>> the command line also take a name; and if the command line takes a name, >>> what's wrong with: >>> >>> ./qemu -add-fd fdset=1,fd=X -add-fd fdset=2,fd=Y -add-fd fdset=3,fd=M >>> -add-fd fdset=4,fd=N -netdev >>> tap,id=hn0,fds=/dev/fdset/1:/dev/fdset/2,vhostfds=/dev/fdset/3:/dev/fdset/4 >>> -device virtio-net-pci,netdev=hn0 >>> >> AFAIK, tap does not support fdset now, so this requirement is beyond the >> scope of multiqueue itself. We can do this in the future. Btw does >> libvirt support add-fd now? > Anything that uses qemu_open() supports fdset now. The question I'm > asking is whether the command line has a way to pass /path/to/name > (which can be presented as /dev/fdset/nnn for add-fd usage) now, or > whether it only supports fds=integers. Nothing special with 'fds' and 'vhostfds', it just split the params by ':' and pass them one by one through monitor_handle_fd_param() just like "fd" and "vhostfd". So if 'fd' and 'vhostfd' supports /path/to/name, so do 'fds' and 'vhostfds'. So for command line, you do can pass /path/to/name to fd/vhostfd but it won't work since monitor_handle_fd_param() can not handle it because 1) it's not an integer 2) it was not named before. But for hotplug, non-integers works since it has already named by getfd, so does fds and vhostfds. For management such as libvirt, what's needed is just to connect the fdname with ':'. >> For hotplug, it just work if you pass multiple file descriptors one by >> one through getfd and then use fds=X:Y,vhostfds=M:N. > For hotplug, you can't pass integers; you have to name the fds either > way. Either you name it with getfd, or you name it with add-fd. But > getfd is not as nice as add-fd when it comes to ensuring that fds are > not leaked in qemu, even when the management app such as libvirt > restarts. Furthermore, if it is possible to specify taps by pathname > instead of by fd inheritance, then using getfd means you have to support > two different approaches in QMP to distinguish which string is being > supplied, while supporting add-fd means you only have to support > qemu_open() which handles both direct names and fd passing in a single > string interface. > > As for libvirt support of add-fd, I'm currently working with Stefan > Berger to get patches applied; the goal is tha libvirt 1.0.3 (end of > February) will support add-fd. > Thanks, I know there are advantages of add-fd, but current tap does not use qemu_open() which means it can't support fdset. We can add this in the future. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
