ovs_dp_change() records the requested user features before it calls the update that can fail, so a request rejected for an empty per-cpu-pids leaves the datapath claiming a feature it never enabled. The tc skb extension key is a reference count shared with cls_flower, and the code that would keep it in step is skipped on that path.
Patch 1 fixes the ordering, and drops the reference on the two other paths that take one and never give it back. Patch 2 adds a set-dp command to the openvswitch selftest tool, and two tests that send such a request and check the features the datapath reports afterwards. They fail without the fix and pass with it. The problem turned up while the YNL spec for this family was being written: describing the ovs_datapath set command meant reading ovs_dp_change(), where the ordering stands out. --- Changes in v3: - Patch 1 drops the tc skb extension reference on the other two paths that take one and lose it: a datapath that fails after ovs_dp_change() enabled the key, and the reset that runs for an outdated user space instance. The description separates the users that hold a reference from the ones that only read the key. - Patch 1 records the features last, in a setter that takes and drops the reference along with them, so the rule has one home instead of four copies, and no caller holds a snapshot of the old value across the update. The datapath.h entry says what the bit means. - The recirc round is a test of its own, so a kernel without the option reports one passing test and one skip instead of hiding both, and the skip reason goes to stdout so a run that skipped it is not silent. The EOPNOTSUPP number is read from python, because it differs between architectures outside the base list. - The tests also cover the reset that an outdated client triggers, check that the datapath keeps working across it, and use datapath names of their own. The reference count itself cannot be observed from user space, which the commit message says. - The selftest cleans up its sandbox before it reports a skip, the way the other mid-test skips in the file do. - The config fragment selects NET_SCHED and NET_CLS_ACT along with NET_TC_SKB_EXT, since the tc symbols sit under them. - set-dp refuses to send a set when the datapath reports no features, and reports a bad pid list instead of a traceback. - Three prints use f-strings, which keeps the lint delta at zero. Changes in v2: - Patch 1 limits its closing claim to OVS_DP_CMD_SET. - Patch 1 names the reachability and the effect: only CAP_NET_ADMIN is needed, datapath_policy does not validate the per-cpu-pids attribute, and a disable on a count of zero warns in static_key_dec_not_one(). - Patch 2 treats only EOPNOTSUPP as the missing key and skips with the kselftest skip code; any other failure fails the test. - Patch 2 selects CONFIG_NET_CLS_ACT and CONFIG_NET_TC_SKB_EXT in the selftest config fragment, so the recirc round runs in CI. - Patch 2 sends the features the datapath reports when --features is left out, instead of clearing them. v1: https://lore.kernel.org/netdev/[email protected]/ v2: https://lore.kernel.org/netdev/[email protected]/ Signed-off-by: Minxi Hou <[email protected]> Minxi Hou (2): net: openvswitch: do not record user features when the change fails selftests/net/openvswitch: check that a failed set keeps the features net/openvswitch/datapath.c | 34 +++-- net/openvswitch/datapath.h | 5 +- .../testing/selftests/net/openvswitch/config | 3 + .../selftests/net/openvswitch/openvswitch.sh | 126 ++++++++++++++++++ .../selftests/net/openvswitch/ovs-dpctl.py | 60 +++++++++ 5 files changed, 214 insertions(+), 14 deletions(-) base-commit: 8373c4e8863ac5f5120d200527aca9880330f343 _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
