侯敏熙 <[email protected]> writes: > Hello Aaron > > Thanks for the suggestion, Aaron. I dug into this a bit and I think > it's worth doing. > > The good news is that 3 of 4 OVS families already have ynl specs > in-tree (ovs_datapath, ovs_flow, ovs_vport). The flow spec is pretty > thorough -- covers all the key/action/tunnel nested attrs. pyynl > handles genetlink, dump, and nested encode/decode, so the library side > looks solid. There's even an OVS C sample under tools/net/ynl/samples/. > > Gaps I found: > > - No ovs_packet.yaml spec. We'd need to write one for upcall > handling (MISS/ACTION/EXECUTE). Shouldn't be too bad, maybe > 100 lines based on the other specs. > > - The existing specs are missing a few operation defs (flow DEL/SET, > vport SET, datapath SET). Small additions. > > - ovs-dpctl.py also uses pyroute2.iproute for tunnel interface > creation and NDB for interface lookup. Those aren't OVS netlink, > so we'd either keep a minimal pyroute2 dep for just that or > shell out to `ip`.
We could also add support for using just a direct netlink socket to query for interfaces. This means we would not need to shell out to iproute2 binary. But I think if iproute2 isn't available then the selftests would also not pass, so subprocess is probably okay. > - The ODP string parser (~1500 lines for flow key/action encoding) > is orthogonal to the netlink wire format, so that stays regardless. > > I think the natural order would be: start with datapath ops (simplest, > spec already complete), then vport, then flow (bulk of the work), then > packet/upcall after writing the spec. The f-string fixes from my > current series still apply. > > Minxi > > Aaron Conole <[email protected]> 于2026年5月15日周五 14:53写道： >> >> Minxi Hou <[email protected]> writes: >> >> > This series cleans up all pylint warnings in ovs-dpctl.py, >> > bringing the score from 7.62/10 to 10.00/10. >> > >> > This series applies on top of: >> > [PATCH net-next v10 1/2] selftests: openvswitch: add vlan() and >> > encap() flow string parsing >> > >> > https://lore.kernel.org/netdev/[email protected]/ >> > >> > Patch 1 converts 86 %-format strings to f-strings. >> > Patch 2 fixes misc warnings (unused import, bare except, unused >> > variables, redundant expressions). >> > Patch 3 renames classes to PascalCase and variables to snake_case. >> > Patch 4 adds one-line docstrings to all definitions. >> > Patch 5 suppresses complexity warnings from pyroute2 constraints. >> > >> > Tested with vng on x86_64, all OVS selftests pass. >> > >> > Minxi Hou (5): >> > selftests: openvswitch: convert %-formatting to f-strings >> > selftests: openvswitch: fix misc pylint warnings in ovs-dpctl.py >> > selftests: openvswitch: rename classes and variables in ovs-dpctl.py >> > selftests: openvswitch: add missing docstrings in ovs-dpctl.py >> > selftests: openvswitch: suppress pylint complexity warnings >> > >> > .../selftests/net/openvswitch/ovs-dpctl.py | 794 ++++++++++-------- >> > 1 file changed, 428 insertions(+), 366 deletions(-) >> >> Thinking about this series, it might be better to go a bit further and >> just drop the pyroute2 in favor of using ynl to generate the netlink >> encode/decode. The value that ovs-dpctl.py brings is the ability to >> interoperate with the ovs-vswitchd odp encoding. That we are using >> pyroute2 to do the actual 'wire' format of netlink is incidental. >> >> Some of the work would still be needed (I think some of the f-string >> adjustments) but that would allow trimming a good chunk of the code and >> let us just rely on the in-tree ynl rather than pyroute2 dependency. >> >> WDYT? Ilya and I can work with you offline if you are interested in >> pursuing this approach. I think it should make the overall code much >> better to work with, too. >> _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
