Peter M. Groen / redox

Fixes for Multi-Threading (bdallas)

From @bveldhoen, PR #25
4a964d09dbe1ce58ac0318ea719752721f174804

* Use lock_guard instead of unique_lock when possible
* Make a few types non-atomic since they are already guarded by a mutex
* Create setVariable helpers that lock and notify

…ed to and has no locking anyway so this will be the same functionality.

…ill make debugging easier if the error condition does occur.

…wise problems could occur where you increment, then fetch and get the wrong number.

Very similar to the fixes for Command and Subscriber, adds locking
around booleans used as flags for condition variables.

Similar to the fixes for subscriber, without holding the waiter_lock_,
updating the waiting_done_ flag may cause deadlock. In the deadlock
example below, Thread1 is waiting for the response, thread 2 is
processing the response and updating the waiter_lock_ variable:

----------------------------------------------------
|        Thread 1          |       Thread 2        |
----------------------------------------------------
| locks 'waiter_lock_'     |                       |
| check 'waiting_done_'    |                       |
|                          | Update 'waiting done' |
|                          | Send notification     |
| begin waiting for signal |                       |
| *deadlocked*             |                       |
----------------------------------------------------

The lock used with the condition variable should always be the same one
guarding changes to the condition itself, in these cases, the lock
guarding changes to the subscribe and psubscibe data structures.

This commit also limits the number of locks held by the subscribeBase
function at any one time.

Set CMAKE_INSTALL_PREFIX only when not externally set

In case of a connection failure (i.e. specify incorrect port number),
the main thread would be blocked. This is fixed by setting the condition
variables correctly in case of a connection error. The test
TestConnectionFailure will hang without this fix.

Add EVRUN_ONCE evloop call at the start of the event loop thread,
instead of just an EVRUN_NOWAIT call. We want to block until the
connection happens, or else we hang on the condition variable wait.

Fix logger include path

Fix the path for logger.hpp, otherwise make install fails.

Change-Id: I50aa9667e512c4fdf6de471dc16312d9eb3dbe30

Better timing statistics, better printing, adding pub sub test modes.

User can now enable no-wait mode, which chooses whether we use the
EVRUN_NOWAIT flag in ev_run. The default is off, so that we don't
use 100% CPU. Note added in tutorial to enable when performance
is critical. Added to the speed test examples.

Bump to 0.2.1. Remove patch number from HISTORY entry - that's what the
git log is for. Make note on minor release.