Before, it was only done for 'logfile'.

- Review error handling. Only 2 exceptions remain in connector.py:

  - Drop useless exception handling for EAGAIN since it should not happen
    if the kernel says the socket is ready.
  - Do not distinguish other socket errors. Just close and log in a generic way.
  - No need to raise a specific exception for EOF.
  - Make 'connect' return a boolean instead of raising an exception.
  - Raise appropriate exception when answer/ask/notify is called on a closed
    non-MT connection.

- Add support for more complex connectors, which may need to write for a read
  operation, or to read when there's pending data to send. This will be
  required for SSL support (more exactly, the handshake will be done in
  a transparent way):

  - Move write buffer to connector.
  - Make 'receive' fill the read buffer, instead of returning the read data.
  - Make 'receive' & 'send' return a boolean to switch polling for writing.
  - Tolerate that sockets return 0 as number of bytes sent.

- In testConnection, simply delete all failing tests, as announced
  in commit 71e30fb9.

Application will hold SSL parameters.

This frees a reference to the last handler and there's no need to make the
instance reusable.

This follows the behaviour of FileStorage.

There remain only one leak in ClientApplicationTests.test_connectToPrimaryNode
because of Mock objects.

admin: do not reset the list of known masters from configuration (or command line) when reconnecting

This is questionable but a lot of NodeManager must be reviewed if we want to do
differently. At least, admin nodes now behave like clients.

It's been a long time that the polling thread never ends and don't need to be
restarted. On the other side, there will be a need for the admin to define a
different polling loop, hence the move from threaded_poll to threaded_app.

It was still used to stop a cluster.

This makes tests easier to write, with more determinism.
If only I had the idea to monkey-patch SimpleQueue several years ago.

The previous implementation was built around a 'pending' global variable that
was set by a few monkey-patches when some network activity was pending between
nodes. All this is replaced by an extra epoll object is used to wait for nodes
that have pending network events: this is simpler, and faster since it
significantly reduces the number of context switches.

This is a regression caused by commit eef52c27
("Tickless poll loop, for lowest latency and cpu usage"), affecting:
- admins
- storages
- primary masters of backup clusters

Recent Python already catches exceptions due to garbage collection on exit.

This fixes a random failure in testClientReconnection:

Traceback (most recent call last):
  File "neo/tests/threaded/test.py", line 754, in testClientReconnection
    self.assertTrue(cluster.client.history(x1._p_oid))
failureException: None is not true

Traceback (most recent call last):
  File "neo/tests/threaded/test.py", line 838, in testRecycledClientUUID
    x = client.load(ZERO_TID)
  [...]
  File "neo/tests/threaded/test.py", line 822, in notReady
    m2s.remove(delayNotifyInformation)
  File "neo/tests/threaded/__init__.py", line 482, in remove
    del self.filter_dict[filter]
KeyError: <function delayNotifyInformation at 0x7f511063a578>

NEOCluster.tic() gets a new 'slave' parameter that must be True when a client
node is in 'master' mode (i.e. setPoll(True)). In this case, tic() will wait
that all nodes finish their work and the client polls with a non-zero timeout.

Here, tic(slave=1) is used to wait for the storage to process
NotifyUnlockInformation notification from the master.

Traceback (most recent call last):
  File "neo/tests/threaded/test.py", line 80, in testBasicStore
    self.assertEqual(data_info, cluster.storage.getDataLockInfo())
  File "neo/tests/__init__.py", line 170, in assertEqual
    return super(NeoTestBase, self).assertEqual(first, second, msg=msg)
failureException: {('\x0b\xee\xc7\xb5\xea?\x0f\xdb\xc9]\r\xd4\x7f<[\xc2u\xda\x8a3', 0): 0} != {('\x0b\xee\xc7\xb5\xea?\x0f\xdb\xc9]\r\xd4\x7f<[\xc2u\xda\x8a3', 0): 1}

All these changes were useful to debug deadlocks in threaded tests:
- New verbose Semaphore.
- Logs with numerical 'ident' were too annoying to read so revert to thread
  name (before commit 5b69d553), with an
  exception for threaded tests. There remains one case where the result is not
  unique: when several client apps are instantiated.
- Make deadlock detection optional.
- Make it possible to name locks.
- Make output more compact.
- Remove useless 'debug_lock' option.
- Add timing information.
- Make exception more verbose when an un-acquired lock is released.

Here is how I used 'locking':

--- a/neo/tests/threaded/__init__.py
+++ b/neo/tests/threaded/__init__.py
@@ -37,0 +38 @@
+from neo.lib.locking import VerboseSemaphore
@@ -71 +72,2 @@ def init(cls):
-        cls._global_lock = threading.Semaphore(0)
+        cls._global_lock = VerboseSemaphore(0, check_owner=False,
+                                            name="Serialized._global_lock")
@@ -265 +267,2 @@ def start(self):
-        self.em._lock = l = threading.Semaphore(0)
+        self.em._lock = l = VerboseSemaphore(0, check_owner=False,
+                                             name=self.node_name)
@@ -346 +349,2 @@ def __init__(self, master_nodes, name, **kw):
-        self.em._lock = threading.Semaphore(0)
+        self.em._lock = VerboseSemaphore(0, check_owner=False,
+                                         name=repr(self))

deadlocks mainly happened while stopping a cluster, hence the complete review
of NEOCluster.stop()

A major change is to make the client node handle its lock like other nodes
(i.e. in the polling thread itself) to better know when to call
Serialized.background() (there was a race condition with the test of
'self.poll_thread.isAlive()' in ClientApplication.close).

For example, a backup storage node that was rejected because the upstream
cluster was not ready could reconnect in loop without delay, using 100% CPU
and flooding logs.

A new 'setReconnectionNoDelay' method on Connection can be used for cases where
it's legitimate to quickly reconnect.

With this new delayed reconnection, it's possible to remove the remaining
time.sleep().

Such kind of test has never helped to detect regressions and any bug in
EpollEventManager would be quickly reported by other tests.

testConnection may go the same way if it keeps annoying me too much.

This is currently not an issue because the 'time.sleep(1)' in iterateForObject
(storage) and _connectToPrimaryNode (master) leave enough time. What could
happen is a new connection attempt for a node that already has a connection
(causing a failure assertion in Node.setConnection).

This could happen if a file descriptor was reallocated by the kernel.

With this patch, the epolling object is not awoken every second to check
if a timeout has expired. The API of Connection is changed to get the smallest
timeout.

This is a prerequisite for tickless poll loops.