The basic and random test cases checked c->sbytes immediately after draining all data from SSL_read, without waiting for the server thread to return from SSL_sendfile().
On a loopback connection, the client can consume all data from the kernel receive buffer before SSL_sendfile() returns on the server side, leaving c->sbytes unset at the point of the ATF_REQUIRE check.

Add wait_server() which holds the mutex and waits on the condvar until the server thread transitions out of RUNNING state, ensuring c->sbytes is stable before it is read.

This fixes a flaky failure appearing roughly once every 40 runs with error:

tests/sys/kern/ssl_sendfile.c:327: c.sbytes == (ssize_t)expect not met

Initialy I’ve increased the failure rate with a small patch by adding usleep(1) temporarily in the test:

ATF_REQUIRE(nread == expect);
usleep(1);  /* widen the race window */
ATF_REQUIRE(c.sbytes == (ssize_t)expect);

With this patch applied, I’ve ran stress-ng to load all my core stress-ng --cpu $(nproc) while running the test in loop thousand of time while sudo cpuset -l 0 kyua test sys/kern/ssl_sendfile:random; do :; done