On Fri, Jan 16, 2015 at 3:14 PM, Yidao Liu <zeroless.mvk@xxxxxxxxx> wrote: > 2015-01-16 19:00 GMT+08:00 Stefan Hajnoczi <stefanha@xxxxxxxxx>: >> On Fri, Jan 16, 2015 at 12:14:17PM +0800, Yidao Liu wrote: >>> Hi, I want to use a dedicated guest VM to handle I/O request just as >>> I/O service domain used in xen. >>> >>> Specifically, using network I/O as an example, I should directly >>> assign the NIC to one guest VM (using pci-assign option), after that >>> all other guest VMs should perform network I/O through that VM rather >>> than the host OS. >>> >>> Is there currently any viable approach to do this? >>> >>> If not, I want to implement one. Currently I'm thinking of combining >>> nahanni shared memory and vhost architecture to implement it, is there >>> any other suggestions? >> >> There are security features like SELinux, seccomp, and file descriptor >> passing that allow the QEMU userspace process to run unprivileged. >> Why does it matter if the I/O is happening in an isolated userspace >> process or another VM? >> > Although we use unprivileged Qemu to emulate the I/O operation, > current I/O mechanism still need to use Host to do the last part of I/O, so > if there's any bug in this phase, the whole system may crash due to > the Host OS down, is it? > So if using I/O service domain, there's no such problem, only service > domain will crash rather than the whole system. Sure, the same reliability properties as Linux on physical machines. This is the monolithic kernels vs microkernels argument. >> You can of course have appliance VMs that do network I/O on behalf of >> other guests. Assign the physical NIC to the appliance VM and then use >> a private bridge on the host so guests can only communicate through the >> appliance VM. >> > I'm curious why use a private bridge on the host, but not in the appliance VM? > since the physical NIC is already assigned to the VM. The physical NIC is assigned to the appliance VM. Now regular VMs need to talk to the networking appliance VM. Without modify code, the easiest way is for regular VMs and the networking appliance VM to talk to each other via networking. Regular VM (virtio-net) <-> (virtio-net) Networking appliance VM (PCI NIC driver) <-> physical network This is how you can set up VMs to do networking in a guest without introducing custom inter-VM communication channels. Stefan -- 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
