tests: rewrite socket to do something sensible and reliable
The current socket test is rather crazy in that it sets up a server listening for sockets and then runs a client connect call, relying on the fact that the kernel will accept this despite the application not having called accept() yet. It then closes the client socket and calls accept() on the server. On Linux accept() will always see that the client has gone and so skip the rest of the code. On FreeBSD, however, the accept sometimes succeeds, causing us to then go into code that attempts to read and write to the client which will fail aborting the test. The accept() never succeeds on FreeBSD guests with a single CPU, but as you add more CPUs, accept() becomes more and more likely to succeed, giving a 100% failure rate for the test when using 8 CPUs. This completely rewrites the test so that it is avoids this designed in race condition. We simply spawn a background thread to act as the client, which will read a byte from the server and write it back again. The main thread can now properly listen and accept the client in a synchronous manner avoiding any races. Reviewed-by: NAndrea Bolognani <abologna@redhat.com> Signed-off-by: NDaniel P. Berrangé <berrange@redhat.com>
Showing
想要评论请 注册 或 登录