On Thu, Mar 08, 2012 at 11:12:58AM -0800, Casey Barker wrote: > I'm experimenting with running multiple ofproto instances from within a > single process, where Open vSwitch is linked as a library, and the > datapaths operate outside the kernel. These ofproto instances are > completely independent of each other, each representing a separate logical > switching entity. (The reasons for putting them in the same process are... > complex.) > > Since they're independent, there's now a desire to use some common port > names on every switching entity -- including those ofprotos that happen to > be managed by the same process. That's fine for the ofprotos, but netdev.c > tracks all created devices in a static 'netdev_dev_shash,' keyed only by > netdev_dev.name. That means all netdevs across all ofprotos in a process > share a global namespace (as is customary for sane use of netdevs). > > My proposed departure from sanity is to assign the ports an > ofproto-specific netdev provider 'type' to differentiate the devs from each > other, similar to how type/name works for datapaths. Does anyone see any > issues with keying the netdev_dev_shash by the concatenation of netdev_dev > type + name? It causes some minor ripples in the netdev API, but nothing > insurmountable. > > I imagine this isn't a desirable change for the mainline code, so I intend > to keep this hack to myself. I'm more looking for any issues that I'm not > seeing, or that could crop up in future iterations. (However, I did notice > that netdev_open() doesn't compare the 'type' before re-opening an existing > dev. Might be worth adding a sanity check as is done for args.)
Around here, we've been talking about support for Linux kernel network namespaces, where a single ovs-vswitchd would be able to manage devices in multiple namespaces. This has the same issue, in that each network namespace has, um, a separate namespace, so that you can have "eth0" in multiple network namespaces. So I think that a clean solution to the problem would help in both cases. I've thought of a few different solutions (but I haven't implemented anything yet). One idea is the same as your solution. Another is to do the qualification in the "name" part, not the type. For example, eth0 in the xyzzy namespace would be something like "xyzzy:eth0". I haven't worked the details in either case. If you work out something clean, then posting it here as a proposal might be helpful for mainline after all. The netdev code always seems to have a bigger ripple effect than I expect. I think that the API must still be badly designed, even after the number of rewrites we've done after the years. I hope we can figure out the Right Way(tm) someday. Thanks, Ben. _______________________________________________ discuss mailing list [email protected] http://openvswitch.org/mailman/listinfo/discuss
