On 06/04/2022 16:12, Konstantin Meskhidze wrote:
4/4/2022 12:44 PM, Mickaël Salaün пишет:
On 04/04/2022 10:28, Konstantin Meskhidze wrote:
4/1/2022 7:52 PM, Mickaël Salaün пишет:
[...]
+static int create_socket(struct __test_metadata *const _metadata)
+{
+
+ int sockfd;
+
+ sockfd = socket(AF_INET, SOCK_STREAM | SOCK_CLOEXEC, 0);
+ ASSERT_LE(0, sockfd);
+ /* Allows to reuse of local address */
+ ASSERT_EQ(0, setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR,
&one, sizeof(one)));
Why is it required?
Without SO_REUSEADDR there is an error that a socket's port is in
use.
I'm sure there is, but why is this port reused? I think this means
that there is an issue in the tests and that could hide potential
issue with the tests (and then with the kernel code). Could you
investigate and find the problem? This would make these tests reliable.
The next scenario is possible here:
"In order for a network connection to close, both ends have to send
FIN (final) packets, which indicate they will not send any additional
data, and both ends must ACK (acknowledge) each other's FIN packets. The
FIN packets are initiated by the application performing a close(), a
shutdown(), or an exit(). The ACKs are handled by the kernel after the
close() has completed. Because of this, it is possible for the process
to complete before the kernel has released the associated network
resource, and this port cannot be bound to another process until the
kernel has decided that it is done."
https://hea-www.harvard.edu/~fine/Tech/addrinuse.html.
So in this case we have busy port in network selfttest and one of the
solution is to set SO_REUSEADDR socket option, "which explicitly allows
a process to bind to a port which remains in TIME_WAIT (it still only
allows a single process to be bound to that port). This is the both the
simplest and the most effective option for reducing the "address already
in use" error".
In know what this option does, but I'm wondering what do you need it for
these tests: which specific line requires it and why? Isn't it a side
effect of running partial tests? I'm worried that this hides some issues
in the tests that may make them flaky.
Without removing the need to find this issue, the next series should
use a network namespace per test, which will confine such issue from
other tests and the host.
So there are 2 options here:
1. Using SO_REUSEADDR option
2. Using network namespace.
I prefer the first option - "the simplest and the most effective one"
If SO_REUSEADDR is really required (and justified), then it should be
used. Either it is required or not, we should use a dedicated network
namespace for each test anyway. This enables to not mess with the host
and not be impacted by it neither (e.g. if some process already use such
ports).
[...]