Re: [PATCH net 2/2] vsock/test: Test setting SO_ZEROCOPY on accept()ed socket

[Michal Luczaj <mhal@rbox.co>]

On 12/24/25 10:15, Stefano Garzarella wrote:
> On Tue, Dec 23, 2025 at 09:38:07PM +0100, Michal Luczaj wrote:
>> On 12/23/25 17:50, Stefano Garzarella wrote:
>>> On Tue, Dec 23, 2025 at 02:20:33PM +0100, Stefano Garzarella wrote:
>>>> On Tue, Dec 23, 2025 at 12:10:25PM +0100, Michal Luczaj wrote:
>>>>> On 12/23/25 11:27, Stefano Garzarella wrote:
>>>>>> On Tue, Dec 23, 2025 at 10:15:29AM +0100, Michal Luczaj wrote:
>>>>>>> Make sure setsockopt(SOL_SOCKET, SO_ZEROCOPY) on an accept()ed socket is
>>>>>>> handled by vsock's implementation.
>>>>>>>
>>>>>>> Signed-off-by: Michal Luczaj <mhal@rbox.co>
>>>>>>> ---
>>>>>>> tools/testing/vsock/vsock_test.c | 33 +++++++++++++++++++++++++++++++++
>>>>>>> 1 file changed, 33 insertions(+)
>>>>>>>
>>>>>>> diff --git a/tools/testing/vsock/vsock_test.c b/tools/testing/vsock/vsock_test.c
>>>>>>> index 9e1250790f33..8ec8f0844e22 100644
>>>>>>> --- a/tools/testing/vsock/vsock_test.c
>>>>>>> +++ b/tools/testing/vsock/vsock_test.c
>>>>>>> @@ -2192,6 +2192,34 @@ static void test_stream_nolinger_server(const struct test_opts *opts)
>>>>>>> 	close(fd);
>>>>>>> }
>>>>>>>
>>>>>>> +static void test_stream_accepted_setsockopt_client(const struct test_opts *opts)
>>>>>>> +{
>>>>>>> +	int fd;
>>>>>>> +
>>>>>>> +	fd = vsock_stream_connect(opts->peer_cid, opts->peer_port);
>>>>>>> +	if (fd < 0) {
>>>>>>> +		perror("connect");
>>>>>>> +		exit(EXIT_FAILURE);
>>>>>>> +	}
>>>>>>> +
>>>>>>> +	vsock_wait_remote_close(fd);
>>>
>>> On a second look, why we need to wait the remote close?
>>> can we just have a control message?
>>
>> I think we can. I've used vsock_wait_remote_close() simply as a sync
>> primitive. It's one line of code less.
>>
>>> I'm not sure even on that, I mean why this peer can't close the
>>> connection while the other is checking if it's able to set zerocopy?
>>
>> I was worried that without any sync, client-side close() may race
>> server-side accept(), but I've just checked and it doesn't seem to cause
>> any issues, at least for the virtio transports.
> 
> Okay, I see. Feel free to leave it, but if it's not really needed, I'd 
> prefer to keep the tests as simple as possible.

OK, dropping the sync here. It will be interesting to see if it ever blows up.

...
>>>> In my suite, I'm checking the client, and if the last test fails only
>>>> on the server, I'm missing it. I'd fix my suite, and maybe also
>>>> vsock_test adding another sync point.
>>>
>>> Added a full barrier here:
>>> https://lore.kernel.org/netdev/20251223162210.43976-1-sgarzare@redhat.com
>>
>> Which reminds me of discussion in
>> https://lore.kernel.org/netdev/151bf5fe-c9ca-4244-aa21-8d7b8ff2470f@rbox.co/
> 
> Oh, I forgot that we already discussed that.
> 
> My first attempt was exactly that, but then discovered that it didn't 
> add too much except for the last one since for the others we have 2 full 
> barriers back to back, so I preferred to move outside the loop. In that 
> way we can also be sure the 2 `vsock_tests` are in sync with the amount 
> of tests to run.

Might it be that we're solving different issues?

I was annoyed by the next test's name/prompt being printed when the
previous test is still running on the other side. Which happens e.g. when
one side takes longer than the other. Or when one of the sides is
unimplemented.

How about something like below; would that cover your case as well?

diff --git a/tools/testing/vsock/util.c b/tools/testing/vsock/util.c
index d843643ced6b..5d94ffd2fa82 100644
--- a/tools/testing/vsock/util.c
+++ b/tools/testing/vsock/util.c
@@ -495,7 +495,7 @@ void run_tests(const struct test_case *test_cases,
 			printf("skipped\n");

 			free(line);
-			continue;
+			goto sync;
 		}

 		control_cmpln(line, "NEXT", true);
@@ -510,6 +510,9 @@ void run_tests(const struct test_case *test_cases,
 			run(opts);

 		printf("ok\n");
+sync:
+		control_writeln("RUN_TESTS_SYNC");
+		control_expectln("RUN_TESTS_SYNC");
 	}
 }