On 2022/5/24 下午9:33, Willem de Bruijn wrote:
On Tue, May 24, 2022 at 6:04 AM Yun Lu <luyun_611@xxxxxxx> wrote:
On 2022/5/23 下午9:32, Willem de Bruijn wrote:
On Mon, May 23, 2022 at 5:25 AM Yun Lu <luyun_611@xxxxxxx> wrote:
On 2022/5/20 下午9:52, Willem de Bruijn wrote:
On Fri, May 20, 2022 at 2:40 AM Yun Lu <luyun_611@xxxxxxx> wrote:
From: luyun <luyun@xxxxxxxxxx>
The psock_snd test sends and recievs packets over loopback, but the
parameter lo.accept_local is disabled by default, this test will
fail with Resource temporarily unavailable:
sudo ./psock_snd.sh
dgram
tx: 128
rx: 142
./psock_snd: recv: Resource temporarily unavailable
I cannot reproduce this failure.
Passes on a machine with accept_local 0.
accept_local is defined as
"
accept_local - BOOLEAN
Accept packets with local source addresses. In combination
with suitable routing, this can be used to direct packets
between two local interfaces over the wire and have them
accepted properly.
"
I did this test on my system(Centos 8.3 X86_64):
[root@localhost net]# sysctl net.ipv4.conf.lo.accept_local
net.ipv4.conf.lo.accept_local = 0
[root@localhost net]# ./psock_snd -d
tx: 128
rx: 142
./psock_snd: recv: Resource temporarily unavailable
[root@localhost net]# sysctl -w net.ipv4.conf.lo.accept_local=1
net.ipv4.conf.lo.accept_local = 1
[root@localhost net]# ./psock_snd -d
tx: 128
rx: 142
rx: 100
OK
This failure does seem to be related to accept_local.
Also, it's reported on Ubuntu:
https://bugs.launchpad.net/ubuntu-kernel-tests/+bug/1812618
That is an old kernel, 4.18 derived.
I simply am unable to reproduce this on an upstream v4.18 or v5.18.
Likely something with either the distro kernel release, or another
distro feature that interacts with this. Can you try v5.18 or another
clean upstream kernel?
Else it requires instrumenting IN_DEV_ACCEPT_LOCAL tests to understand
where this sysctl makes a meaningful change for you when running this
test.
I found another parameter rp_filter which interacts with this test:
Set rp_filter = 0, the psock_snd test OK.
Or set rp_filter = 1 and accept_local=1, the psock_snd test OK.
I get the same test results on kernel v5.10 or v5.15. Analysis from
source code, this two parameters
will change the result of fib_validate_source when running this test.
For most distro kernel releases,
rp_filter is enabled by default, so this test will fail when
accept_local is kept to be zero.
That explains.
Since this test runs inside a netns, changing a sysctl is fine.
Can you resubmit with a more detailed description now that the exact
check is more clear, as well as interplay with rp_filter, and the
default config of these two parameters configured by distros? Thanks.
Please double check typos (s/recievs/receives).
OK, thanks for your reply and suggestion. I will send v2 version of this
patch soon.
