Re: [libvirt] [PATCH 2/2] 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.

+static int +testSocketAccept(const void *opaque) { virNetSocketPtr *lsock = NULL; /* Listen socket */ size_t nlsock = 0, i; virNetSocketPtr ssock = NULL; /* Server socket */ - virNetSocketPtr csock = NULL; /* Client socket */ + virNetSocketPtr rsock = NULL; /* Server from poll */

+ virThread th; + struct testClientData cdata = { 0 }; + bool goodsock = false; + char a = 'a', b = '\0';

+ if (virNetSocketWrite(ssock, &a, 1) < 0 || + virNetSocketRead(ssock, &b, 1) < 0) + goto join;

+ if (a != b) { + virReportError(VIR_ERR_INTERNAL_ERROR, + "Bad data received '%x' != '%x'", a, b); + goto join; + }