Re: [PATCH liburing] tests: add cBPF filter tests for IORING_OP_CONNECT

Sure. Have sent v2 patch  — added a probe_connect_filter_support()
helper that fork-registers a CONNECT filter once up front and gates
connect subtests behind the result.

Regard,
Shouvik Kar


On Wed, May 13, 2026 at 10:36 PM Jens Axboe <[email protected]> wrote:
>
> On 5/13/26 6:10 AM, Shouvik Kar wrote:
> > Add subtests for IORING_OP_CONNECT to test/cbpf_filter.c, exercising
> > the io_connect_bpf_populate() helper added in the companion kernel
> > patch ("io_uring/net: allow filtering on IORING_OP_CONNECT").
> >
> > Coverage spans both blacklist and whitelist filters for each
> > connect-specific data field (family, v4 address, v6 address, port),
> > plus v4 and v6 subnet matching, and a test for the addr_len guard
> > in io_connect_bpf_populate that prevents stale io_async_msghdr
> > cache from leaking through to the filter on short connects.
>
> If you run this on a kernel that doesn't have your connect changes,
> then you get a lot of:
>
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
> Child: register failed: Message too long
>
> when the test is run. It's important that any liburing test cases
> handles older kernels appropriately. You get some of it for free with
> this test case, as previous tests will have already checked if cbpf
> filters are supported in the first place. But you still need to handle
> the case where cbpf filters are supported by io_uring, yet the kernel
> doesn't support your filter yet.
>
> It should just check for the error on the first case and skip testing
> the rest of them.
>
> --
> Jens Axboe