vsock/test: Fix occasional failure in SIOCOUTQ tests
authorKonstantin Shkolnyy <kshk@linux.ibm.com>
Wed, 7 May 2025 15:14:56 +0000 (10:14 -0500)
committerJakub Kicinski <kuba@kernel.org>
Tue, 13 May 2025 22:01:50 +0000 (15:01 -0700)
These tests:
    "SOCK_STREAM ioctl(SIOCOUTQ) 0 unsent bytes"
    "SOCK_SEQPACKET ioctl(SIOCOUTQ) 0 unsent bytes"
output: "Unexpected 'SIOCOUTQ' value, expected 0, got 64 (CLIENT)".

They test that the SIOCOUTQ ioctl reports 0 unsent bytes after the data
have been received by the other side. However, sometimes there is a delay
in updating this "unsent bytes" counter, and the test fails even though
the counter properly goes to 0 several milliseconds later.

The delay occurs in the kernel because the used buffer notification
callback virtio_vsock_tx_done(), called upon receipt of the data by the
other side, doesn't update the counter itself. It delegates that to
a kernel thread (via vsock->tx_work). Sometimes that thread is delayed
more than the test expects.

Change the test to poll SIOCOUTQ until it returns 0 or a timeout occurs.

Signed-off-by: Konstantin Shkolnyy <kshk@linux.ibm.com>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
Fixes: 18ee44ce97c1 ("test/vsock: add ioctl unsent bytes test")
Link: https://patch.msgid.link/20250507151456.2577061-1-kshk@linux.ibm.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
tools/testing/vsock/vsock_test.c

index d0f6d25..6135511 100644 (file)
@@ -1264,21 +1264,25 @@ static void test_unsent_bytes_client(const struct test_opts *opts, int type)
        send_buf(fd, buf, sizeof(buf), 0, sizeof(buf));
        control_expectln("RECEIVED");
 
-       ret = ioctl(fd, SIOCOUTQ, &sock_bytes_unsent);
-       if (ret < 0) {
-               if (errno == EOPNOTSUPP) {
-                       fprintf(stderr, "Test skipped, SIOCOUTQ not supported.\n");
-               } else {
+       /* SIOCOUTQ isn't guaranteed to instantly track sent data. Even though
+        * the "RECEIVED" message means that the other side has received the
+        * data, there can be a delay in our kernel before updating the "unsent
+        * bytes" counter. Repeat SIOCOUTQ until it returns 0.
+        */
+       timeout_begin(TIMEOUT);
+       do {
+               ret = ioctl(fd, SIOCOUTQ, &sock_bytes_unsent);
+               if (ret < 0) {
+                       if (errno == EOPNOTSUPP) {
+                               fprintf(stderr, "Test skipped, SIOCOUTQ not supported.\n");
+                               break;
+                       }
                        perror("ioctl");
                        exit(EXIT_FAILURE);
                }
-       } else if (ret == 0 && sock_bytes_unsent != 0) {
-               fprintf(stderr,
-                       "Unexpected 'SIOCOUTQ' value, expected 0, got %i\n",
-                       sock_bytes_unsent);
-               exit(EXIT_FAILURE);
-       }
-
+               timeout_check("SIOCOUTQ");
+       } while (sock_bytes_unsent != 0);
+       timeout_end();
        close(fd);
 }