Geliang Tang wrote:
On Mon, 2024-05-27 at 21:36 +0200, Jakub Sitnicki wrote:
On Mon, May 27, 2024 at 10:12 AM -07, John Fastabend wrote:
Geliang Tang wrote:
From: Geliang Tang tanggeliang@kylinos.cn
Switch attachments to bpf_link using bpf_program__attach_sockmap() instead of bpf_prog_attach().
Sorry it took me a few days to get to this.
Is there a reason to push this to links vs just leave it as is? I had a plan to port all the test_sockmap tests into prog_tests anyways. I'll try to push some initial patch next week.
Great, I strongly agree with porting them into prog_tests. I am also willing to participate in implementing this plan together.
I have a first patch that starts to move things I'll dig it up here. Still a bit behind on everything as you see its Thr already.
The one advantage of test_sockmap is we can have it run for longer runs by pushing different options through so might be worth keeping just for that.
If you really want links here I'm OK with that I guess just asking.
It was me who suggested the switch to bpf_link in reaction to a series of cleanups to prog_type and prog_attach_type submitted by Geliang.
Yes, patches 3-5 address Jakub's suggestion: switching attachments to bpf_link.
OK. Lets just take them the series lgtm. Jakub any other comments?
Relevant threads:
https://lore.kernel.org/bpf/9c10d9f974f07fcb354a43a8eca67acb2fafc587.1715926... https://lore.kernel.org/bpf/20240522080936.2475833-1-jakub@cloudflare.com https://lore.kernel.org/bpf/e27d7d0c1e0e79b0acd22ac6ad5d8f9f00225303.1716372...
I thought bpf_links added more value than cleaning up "old style" attachments.
Other patches 1-2, 6-8 are small fixes which I found while trying to solve the NONBLOCK issue [1]. Yes, I haven't given up on solving this issue yet. I think it must be solved, since there is a bug somewhere. WDYT?
Yes I think this is an actual issue with the stream parser waking up sockets before the data is copied into the recv buffers.