lttng-load(1): fix synopsis and style

[lttng-tools.git] / doc / relayd-architecture.txt

LTTng Relay Daemon Architecture
Mathieu Desnoyers, August 2015

This document describes the object model and architecture of the relay
daemon, after the refactoring done within the commit "Fix: Relay daemon
ownership and reference counting".

We have the following object composition hierarchy:

relay connection (main.c, for sessiond/consumer)
  |
  \-> 0 or 1 session
       |
       \-> 0 or many ctf-trace
                 |
                 \-> 0 or many stream
                      |   |
                      |   \-> 0 or many index
                      |
                      \-------> 0 or 1 viewer stream

live connection (live.c, for client)
  |
  \-> 1 viewer session
       |
       \-> 0 or many session (actually a reference to session as created
             |                by the relay connection)
             |
             \-> ..... (ctf-trace, stream, index, viewer stream)

There are global tables declared in lttng-relayd.h for sessions
(sessions_ht, indexed by session id), streams (relay_streams_ht, indexed
by stream handle), and viewer streams (viewer_streams_ht, indexed by
stream handle). The purpose of those tables is to allow fast lookup of
those objects using the IDs received in the communication protocols.

There is also one connection hash table per worker thread. There is one
worker thread to receive data (main.c), and one worker thread to
interact with viewer clients (live.c). Those tables are indexed by
socket file descriptor.

A RCU lookup+refcounting scheme has been introduced for all objects
(except viewer session which is still an exception at the moment). This
scheme allows looking up the objects or doing a traversal on the RCU
linked list or hash table in combination with a getter on the object.
This getter validates that there is still at least one reference to the
object, else the lookup acts just as if the object does not exist. This
scheme is protected by a "reflock" mutex in each object. "reflock"
mutexes can be nested from the innermost object to the outermost object.
IOW, the session reflock can nest within the ctf-trace reflock.

The relay_connection (connection between the sessiond/consumer and the
relayd) is the outermost object of its hierarchy.

The live connection (connection between a live client and the relayd)
is the outermost object of its hierarchy.

There is also a "lock" mutex in each object. Those are used to
synchronize between threads (currently the main.c relay thread and
live.c client thread) when objects are shared. Locks can be nested from
the outermost object to the innermost object. IOW, the ctf-trace lock can
nest within the session lock.

A "lock" should never nest within a "reflock".

RCU linked lists are used to iterate using RCU, and are protected by
their own mutex for modifications. Iterations should be confirmed using
the object "getter" to ensure its refcount is not 0 (except in cases
where the caller actually owns the objects and therefore can assume its
refcount is not 0).

RCU hash tables are used to iterate using RCU. Iteration should be
confirmed using the object "getter" to ensure its refcount is not 0
(except again if we have ownership and can assume the object refcount is
not 0).

Object creation has a refcount of 1. Each getter increments the
refcount, and needs to be paired with a "put" to decrement it. A final
put on "self" (ownership) will allow refcount to reach 0, therefore
triggering release, and thus free through call_rcu.

In the composition scheme, we find back references from each composite
to its container. Therefore, each composite holds a reference (refcount)
on its container. This allows following pointers from e.g. viewer stream
to stream to ctf-trace to session without performing any validation,
due to transitive refcounting of those back-references.

In addition to those back references, there are a few key ownership
references held. The connection in the relay worker thread (main.c)
holds ownership on the session, and on each stream it contains. The
connection in the live worker thread (live.c) holds ownership on each
viewer stream it creates. The rest is ensured by back references from
composite to container objects. When a connection is closed, it puts all
the ownership references it is holding. This will then eventually
trigger destruction of the session, streams, and viewer streams
associated with the connection when all the back references reach 0.

RCU read-side locks are now only held during iteration on RCU lists and
hash tables, and within the internals of the get (lookup) and put
functions. Those functions then use refcounting to ensure existence of
the object when returned to their caller.