tcp_filters: Remove filter when efrm_filter_redirect fails

[zhuyifei1999]

When efrm_filter_redirect fails with ENOENT or ENODEV, the code path would reach efrm_filter_insert, which would try to insert a filter rule with the same filter but different flow steering target queue.

In the case of AF_XDP, efrm_filter_redirect is not implemented (always ENODEV), and some netdevs return an error for af_xdp_filter_insert (via set_rxnfc) when it detects duplicate flow steering rules, even if the target queue is different, requiring the caller to remove the prior filter first. This patch addresses this.

Without the patch, if filter_insert errors, oofilter->filter_id[hwport] is left as the error code from efrm_filter_redirect and the old rule is forgotten by onload and left dangling.

[sianj-xilinx]

Thank you for this suggestion. I'm actually actively working in this area because I'm looking at filter redirect support for one of our other hardware platforms, and had come across this exact issue. I'd like to complete the batch of changes I'm currently looking at and then will come back to this to see whether it still applies as is, or is resolved by the rest of the changes I'm doing.

[mina]

Thank you very much. Glad to see we're seeing the same issue. If it's at all possible to give us a preview of the changes you have in mind via a github branch or something, please do. We can help you with the testing.

[mina]

Hi sianj-xilinx,

With regards to this bit from your earlier comment:

I'd like to complete the batch of changes I'm currently looking at and then will come back to this to see whether it still applies as is, or is resolved by the rest of the changes I'm doing.

I was wondering if you had a chance to complete your work, and if so, should I retest on a particular onload release where your changes are done/merged? I could refresh this pull request with an updated fix if I still see the issue.

[mina]

Hi folks,

This is a very small fix to a critical issue we're encountering with Onload running on AF_XDP. Any update on review merge? What can we do to push this forward?