[PATCH net-next] selftests: drv-net: xdp: make the XDP qstats tests less flaky

[PATCH net-next] selftests: drv-net: xdp: make the XDP qstats tests less flaky

[Jakub Kicinski]

The XDP qstats tests send 2k packets over a single socket.
Looks like when netdev CI is busy running those tests in QEMU
occasionally flakes. The target doesn't get to run at all
before all 2000 packets are sent.

Lower the number of packets to 1000 and reopen the socket
every 50 packets, to give RSS a chance to spread the packets
to multiple queues.

For the netdev CI testing either lowering the count or using
multiple sockets is enough, but let's do both for extra resiliency.

Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
CC: shuah@kernel.org
CC: ast@kernel.org
CC: hawk@kernel.org
CC: john.fastabend@gmail.com
CC: sdf@fomichev.me
CC: linux-kselftest@vger.kernel.org
---
 tools/testing/selftests/drivers/net/xdp.py | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/tools/testing/selftests/drivers/net/xdp.py b/tools/testing/selftests/drivers/net/xdp.py
index a148004e1c36..834a37ae7d0d 100755
--- a/tools/testing/selftests/drivers/net/xdp.py
+++ b/tools/testing/selftests/drivers/net/xdp.py
@@ -687,9 +687,12 @@ from lib.py import ip, bpftool, defer
         "/dev/null"
     # Listener runs on "remote" in case of XDP_TX
     rx_host = cfg.remote if act == XDPAction.TX else None
-    # We want to spew 2000 packets quickly, bash seems to do a good enough job
-    tx_udp =  f"exec 5<>/dev/udp/{cfg.addr}/{port}; " \
-        "for i in `seq 2000`; do echo a >&5; done; exec 5>&-"
+    # We want to spew 1000 packets quickly, bash seems to do a good enough job
+    # Each reopening of the socket gives us a differenot local port (for RSS)
+    tx_udp = "for _ in `seq 20`; do " \
+        f"exec 5<>/dev/udp/{cfg.addr}/{port}; " \
+        "for i in `seq 50`; do echo a >&5; done; " \
+        "exec 5>&-; done"
 
     cfg.wait_hw_stats_settle()
     # Qstats have more clearly defined semantics than rtnetlink.
@@ -704,11 +707,11 @@ from lib.py import ip, bpftool, defer
     cfg.wait_hw_stats_settle()
     after = cfg.netnl.qstats_get({"ifindex": cfg.ifindex}, dump=True)[0]
 
-    ksft_ge(after['rx-packets'] - before['rx-packets'], 2000)
+    expected_pkts = 1000
+    ksft_ge(after['rx-packets'] - before['rx-packets'], expected_pkts)
     if act == XDPAction.TX:
-        ksft_ge(after['tx-packets'] - before['tx-packets'], 2000)
+        ksft_ge(after['tx-packets'] - before['tx-packets'], expected_pkts)
 
-    expected_pkts = 2000
     stats = _get_stats(prog_info["maps"]["map_xdp_stats"])
     ksft_eq(stats[XDPStats.RX.value], expected_pkts, "XDP RX stats mismatch")
     if act == XDPAction.TX:
-- 
2.51.1

Re: [PATCH net-next] selftests: drv-net: xdp: make the XDP qstats tests less flaky

[Stanislav Fomichev]

On 11/13, Jakub Kicinski wrote:
> The XDP qstats tests send 2k packets over a single socket.
> Looks like when netdev CI is busy running those tests in QEMU
> occasionally flakes. The target doesn't get to run at all
> before all 2000 packets are sent.
> 
> Lower the number of packets to 1000 and reopen the socket
> every 50 packets, to give RSS a chance to spread the packets
> to multiple queues.
> 
> For the netdev CI testing either lowering the count or using
> multiple sockets is enough, but let's do both for extra resiliency.
> 
> Signed-off-by: Jakub Kicinski <kuba@kernel.org>
> ---
> CC: shuah@kernel.org
> CC: ast@kernel.org
> CC: hawk@kernel.org
> CC: john.fastabend@gmail.com
> CC: sdf@fomichev.me
> CC: linux-kselftest@vger.kernel.org
> ---
>  tools/testing/selftests/drivers/net/xdp.py | 15 +++++++++------
>  1 file changed, 9 insertions(+), 6 deletions(-)
> 
> diff --git a/tools/testing/selftests/drivers/net/xdp.py b/tools/testing/selftests/drivers/net/xdp.py
> index a148004e1c36..834a37ae7d0d 100755
> --- a/tools/testing/selftests/drivers/net/xdp.py
> +++ b/tools/testing/selftests/drivers/net/xdp.py
> @@ -687,9 +687,12 @@ from lib.py import ip, bpftool, defer
>          "/dev/null"
>      # Listener runs on "remote" in case of XDP_TX
>      rx_host = cfg.remote if act == XDPAction.TX else None
> -    # We want to spew 2000 packets quickly, bash seems to do a good enough job
> -    tx_udp =  f"exec 5<>/dev/udp/{cfg.addr}/{port}; " \
> -        "for i in `seq 2000`; do echo a >&5; done; exec 5>&-"
> +    # We want to spew 1000 packets quickly, bash seems to do a good enough job
> +    # Each reopening of the socket gives us a differenot local port (for RSS)
> +    tx_udp = "for _ in `seq 20`; do " \
> +        f"exec 5<>/dev/udp/{cfg.addr}/{port}; " \
> +        "for i in `seq 50`; do echo a >&5; done; " \
> +        "exec 5>&-; done"

TIL about bash's /dev/udp, interesting..

Acked-by: Stanislav Fomichev <sdf@fomichev.me>

Re: [PATCH net-next] selftests: drv-net: xdp: make the XDP qstats tests less flaky

[patchwork-bot+netdevbpf]

Hello:

This patch was applied to netdev/net-next.git (main)
by Jakub Kicinski <kuba@kernel.org>:

On Thu, 13 Nov 2025 07:27:03 -0800 you wrote:
> The XDP qstats tests send 2k packets over a single socket.
> Looks like when netdev CI is busy running those tests in QEMU
> occasionally flakes. The target doesn't get to run at all
> before all 2000 packets are sent.
> 
> Lower the number of packets to 1000 and reopen the socket
> every 50 packets, to give RSS a chance to spread the packets
> to multiple queues.
> 
> [...]

Here is the summary with links:
  - [net-next] selftests: drv-net: xdp: make the XDP qstats tests less flaky
    https://git.kernel.org/netdev/net-next/c/eca8b8fc74cb

You are awesome, thank you!
-- 
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html