ahead of processing data arrived on the same socket, with EnqueueKill
that adds to same queue from which data's taken. So if device dies
immediately after sending data there won't be a race between closing
the cref (if this is the last open socket) and handling the data. I'm
still dying with assert fails when running 100 games at once, but much
less frequently
(e.g. socket closes) can also be queued; track what sockets are
currently being handled so no two threads are ever processing events
on same socket. This makes running 'NGAMES=50
./scripts/discon_ok2.sh' crash very infrequently rather than every
time, but the race condition in crefmgr remains. Queuing socket
closes should help.
lock. (This bug has always been there, but didn't matter before;
maybe pthreads changed.) Comment out self-spawning stuff until can
fix df-transfer problems. Other minor changes around logging.
kill crefs via state machine, and protect access to a cref so it can
die without another thread being in it; do timers via timeout to
poll() rather than interrupt (and integrate into state machine);
detect when all players are present and change state so new
connections on that cookie will get a new cref.
