Fix: sessiond: size-based notification occasionally not triggered

Fix: sessiond: size-based notification occasionally not triggered

Issue observed
==============

When tracing to multiple sessions with scheduled size-based rotations
that occur simultaneously (typically because they trace the same events
and use the same rotation schedule configuration), the start of some
rotations seems to be delayed indefinitely.

Cause
=====

The size-based rotations are implemented by piggy-backing onto the
channel monitoring facilities.

Essentially, a per-channel timer samples a number of statistics on the
consumer daemon end, which transmits them to the session daemon.

The session daemon's notification subsystem evaluates the statistics
against the various registered triggers bound to the channels being
monitored when a statistics sample is received.

To implement size-based rotations, internal triggers are registered with
the "consumed size" condition set to a given threshold. A session
rotation management thread (which also performs other tasks) uses a
notification channel to wait for sessions to reach their target size,
starts rotations as needed, and sets a new threshold according to the
sessions' configured rotation schedule.

The rotation thread uses liblttng-ctl's API to consume notifications
from a notification channel.

At any time, a notification channel may have multiple notifications
queued-up internally in its buffers. This is because a notification
channel multiplexes command replies and notifications over the same UNIX
socket. The current protocol specifies that multiple notifications can
be received before the reply to a command.

In such cases, the notification channel client implementation internally
queues them and provides them on the next calls to
lttng_notification_channel_get_next_notification().

This is correct with respect to the public API, which is intended to be
used in "blocking mode". However, this internal user uses the
notification channel's raw file descriptor to wake-up when a
notification is available.

This is problematic because notifications may be queued by the
notification channel (and thus removed from the socket) while waiting
for command replies (subscribing and unsubscribing from notification
conditions). In such a case, a notification is available but the
rotation thread does not wake-up to consume it as nothing is available
in the socket's buffer.

When this happens, a session that is supposed to rotate automatically
appears to grow indefinitely. It will typically eventually rotate as new
notifications become available and cause the rotation thread to wake-up.
However, a "lag" builds up as the notification that caused the wake-up
is not consumed. Instead, the last buffered notification is provided to
the rotation thread.

Solution
========

Use an event_fd to wake-up the rotation thread whenever a command
completes on the notification channel. This ensures that any
notification that was queued while waiting for a reply to the command is
eventually consumed.

Known drawbacks
===============

None.

Note
====

The use of C++ features is kept to a minimum in this patch in order to
make it easier to backport to the stable releases. A clean-up patch
follows and makes the code conform to the coding standards.

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>
Change-Id: I8974b10124704d1e66e8da32d495fee738e3d43f