Re: [PATCH v3 net-next 09/15] selftests: xsk: run one test case per xskxceiver invocation

["Björn Töpel" <bjorn@kernel.org>]

Maciej Fijalkowski <maciej.fijalkowski@intel.com> writes:

> Replace test_xsk.sh and xsk_prereqs.sh with test_xsk.py, which runs
> xskxceiver once per mode and case, so that every case starts from a
> fresh process and a crash fails only its own case. A single case is
> also the unit that the next patch splits into RX and TX processes.
>
> xskxceiver now requires -m and -t, runs exactly that case, reports one
> KTAP result and exits with its verdict. An unsupported mode skips the
> case instead of exiting with XFAIL.
>
> test_xsk.py uses the networking selftests Python library, as does the
> hardware test added later, so every case is a KTAP case of its own:
> test_xsk runs each case that xskxceiver -l lists in skb and drv mode,
> e.g. test_xsk.drv_SEND_RECEIVE, and test_xsk_busy_poll runs them again
> with busy polling. The usual -t and -T options select cases in place of
> the script's -l, -m and -t, and -v and -d go away, as the output of a
> run that does not pass is logged. Each case runs on a veth pair of its
> own, so an MTU or an XDP program that a crashed case leaves behind
> cannot reach the next one. The pair stays in the current network
> namespace, as only the initial one has the net.core.max_skb_frags
> sysctl that xskxceiver reads. A case is killed and fails after two
> minutes, and the test skips when xskxceiver was not built.

This timeout might need to be easily changable from the cli...
follow-up!

> test_xsk.py reserves 32 huge pages for the UNALIGNED cases, restores
> the count when it ends and leaves those cases out when it cannot
> reserve them. ksft_run() exits on -l and -h before the count is
> restored, so test_xsk.py rejects both. The huge page, busy-poll and
> endpoint wait helpers, which the hardware test shares, go to
> net/lib/py/xsk.py, which joins the AF_XDP entry in MAINTAINERS.
>
> test_xsk.py leaves out two cases that test_progs skipped as flaky:
>
> - XDP_ADJUST_TAIL_SHRINK_MULTI_BUFF: bpf_xdp_adjust_tail() frees the
>   trimmed fragment as the memory type of the XDP RX queue, which veth
>   and generic XDP leave at MEM_TYPE_PAGE_SHARED for frames that they
>   copy into page_pool pages, so the page leaks from its pool and the
>   pool's shutdown stalls
> - XDP_ADJUST_TAIL_GROW: it skips in skb mode, and in drv mode veth
>   copies each frame into a page_pool fragment that leaves no room to
>   grow the 128-byte frames of the case unless it is the last one of its
>   page, so the XDP program drops most frames
>
> Drop the xskxceiver.c note on enabling packet dumps with -D in
> test_xsk.sh; xskxceiver has no such option and dumps packets on
> validation errors.
>
> Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>

[...]

>  tools/testing/selftests/net/test_xsk.py       | 140 +++++++++++
>  tools/testing/selftests/net/test_xsk.sh       | 230 ------------------

Yum! Nice!

Reviewed-by: Björn Töpel <bjorn@kernel.org>