signals2/doc/thread_safety.xml

<?xml version="1.0" encoding="utf-8"?>
<!DOCTYPE section PUBLIC "-//Boost//DTD BoostBook XML V1.0//EN"
  "http://www.boost.org/tools/boostbook/dtd/boostbook.dtd">
<!--
Copyright Frank Mori Hess 2009

Distributed under the Boost Software License, Version 1.0. (See accompanying
file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
-->
<section last-revision="$Date: 2007-06-12 14:01:23 -0400 (Tue, 12 Jun 2007) $" id="signals2.thread-safety">
  <title>Thread-Safety</title>

  <using-namespace name="boost::signals2"/>
  <using-namespace name="boost"/>

  <section>
    <title>Introduction</title>
    <para>
      The primary motivation for Boost.Signals2 is to provide a version of
      the original Boost.Signals library which can be used safely in a
      multi-threaded environment.
      This is achieved primarily through two changes from the original Boost.Signals
      API.  One is the introduction of a new automatic connection management scheme
      relying on <classname>shared_ptr</classname> and <classname>weak_ptr</classname>,
      as described in the <link linkend="signals2.tutorial.connection-management">tutorial</link>.
      The second change was the introduction of a <code>Mutex</code> template type
      parameter to the <classname alt="signals2::signal">signal</classname> class.  This section details how
      the library employs these changes to provide thread-safety, and
      the limits of the provided thread-safety.
    </para>
  </section>
  <section>
    <title>Signals and combiners</title>
    <para>
      Each signal object default-constructs a <code>Mutex</code> object to protect
      its internal state.  Furthermore, a <code>Mutex</code> is created
      each time a new slot is connected to the signal, to protect the
      associated signal-slot connection.
    </para>
    <para>
      A signal's mutex is automatically locked whenever any of the
      signal's methods are called.  The mutex is usually held until the
      method completes, however there is one major exception to this rule.  When
      a signal is invoked by calling
      <methodname alt="signal::operator()">signal::operator()</methodname>,
      the invocation first acquires a lock on the signal's mutex.  Then
      it obtains a handle to the signal's slot list and combiner.  Next
      it releases the signal's mutex, before invoking the combiner to
      iterate through the slot list.  Thus no mutexes are held by the
      signal while a slot is executing.  This design choice
      makes it impossible for user code running in a slot
      to deadlock against any of the
      mutexes used internally by the Boost.Signals2 library.
      It also prevents slots from accidentally causing
      recursive locking attempts on any of the library's internal mutexes.
      Therefore, if you invoke a signal concurrently from multiple threads,
      it is possible for the signal's combiner to be invoked concurrently
      and thus the slots to execute concurrently.
    </para>
    <para>
      During a combiner invocation, the following steps are performed in order to
      find the next callable slot while iterating through the signal's
      slot list.
    </para>
    <itemizedlist>
      <listitem>
        <para>The <code>Mutex</code> associated with the connection to the
          slot is locked.</para>
      </listitem>
      <listitem>
        <para>All the tracked <classname>weak_ptr</classname> associated with the
          slot are copied into temporary <classname>shared_ptr</classname> which
          will be kept alive until the invocation is done with the slot.  If this fails due
          to any of the
          <classname>weak_ptr</classname> being expired, the connection is
          automatically disconnected.  Therefore a slot will never be run
          if any of its tracked <classname>weak_ptr</classname> have expired,
          and none of its tracked <classname>weak_ptr</classname> will
          expire while the slot is running.
        </para>
      </listitem>
      <listitem>
        <para>
          The slot's connection is checked to see if it is blocked
          or disconnected, and then the connection's mutex is unlocked.  If the connection
          was either blocked or disconnected, we
          start again from the beginning with the next slot in the slot list.
          Otherwise, we commit to executing the slot when the combiner next
          dereferences the slot call iterator (unless the combiner should increment
          the iterator without ever dereferencing it).
        </para>
      </listitem>
    </itemizedlist>
    <para>
      Note that since we unlock the connection's mutex before executing
      its associated slot, it is possible a slot will still be executing
      after it has been disconnected by a
      <code><methodname>connection::disconnect</methodname>()</code>, if
      the disconnect was called concurrently with signal invocation.
    </para>
    <para>
      You may have noticed above that during signal invocation, the invocation only
      obtains handles to the signal's slot list and combiner while holding the
      signal's mutex.  Thus concurrent signal invocations may still wind up
      accessing the
      same slot list and combiner concurrently.  So what happens if the slot list is modified,
      for example by connecting a new slot, while a signal
      invocation is in progress concurrently?  If the slot list is already in use,
      the signal performs a deep copy of the slot list before modifying it.
      Thus the a concurrent signal invocation will continue to use the old unmodified slot list,
      undisturbed by modifications made to the newly created deep copy of the slot list.
      Future signal invocations will receive a handle to the newly created deep
      copy of the slot list, and the old slot list will be destroyed once it
      is no longer in use.  Similarly, if you change a signal's combiner with
      <methodname alt="signal::set_combiner">signal::set_combiner</methodname>
      while a signal invocation is running concurrently, the concurrent
      signal invocation will continue to use the old combiner undisturbed,
      while future signal invocations will receive a handle to the new combiner.
    </para>
    <para>
      The fact that concurrent signal invocations use the same combiner object
      means you need to insure any custom combiner you write is thread-safe.
      So if your combiner maintains state which is modified when the combiner
      is invoked, you
      may need to protect that state with a mutex.  Be aware, if you hold
      a mutex in your combiner while dereferencing slot call iterators,
      you run the risk of deadlocks and recursive locking if any of
      the slots cause additional mutex locking to occur.  One way to avoid
      these perils is for your combiner to release any locks before
      dereferencing a slot call iterator.  The combiner classes provided by
      the Boost.Signals2 library are all thread-safe, since they do not maintain
      any state across invocations.
    </para>
    <para>
      Suppose a user writes a slot which connects another slot to the invoking signal.
      Will the newly connected slot be run during the same signal invocation in
      which the new connection was made?  The answer is no.  Connecting a new slot
      modifies the signal's slot list, and as explained above, a signal invocation
      already in progress will not see any modifications made to the slot list.
    </para>
    <para>
      Suppose a user writes a slot which disconnects another slot from the invoking signal.
      Will the disconnected slot be prevented from running during the same signal invocation,
      if it appears later in the slot list than the slot which disconnected it?
      This time the answer is yes.  Even if the disconnected slot is still
      present in the signal's slot list, each slot is checked to see if it is
      disconnected or blocked immediately before it is executed (or not executed as
      the case may be), as was described in more detail above.
    </para>
  </section>
  <section>
    <title>Connections and other classes</title>
    <para>
      The methods of the <classname>signals2::connection</classname> class are thread-safe,
      with the exception of assignment and swap.  This is achived via locking the mutex
      associated with the object's underlying signal-slot connection.  Assignment and
      swap are not thread-safe because the mutex protects the underlying connection
      which a <classname>signals2::connection</classname> object references, not
      the <classname>signals2::connection</classname> object itself.  That is,
      there may be many copies of a <classname>signals2::connection</classname> object,
      all of which reference the same underlying connection.  There is not a mutex
      for each <classname>signals2::connection</classname> object, there is only
      a single mutex protecting the underlying connection they reference.
    </para>
    <para>The <classname>shared_connection_block</classname> class obtains some thread-safety
      from the <code>Mutex</code> protecting the underlying connection which is blocked
      and unblocked.  The internal reference counting which is used to keep track of
      how many <classname>shared_connection_block</classname> objects are asserting
      blocks on their underlying connection is also thread-safe (the implementation
      relies on <classname>shared_ptr</classname> for the reference counting).
      However, individual <classname>shared_connection_block</classname> objects
      should not be accessed concurrently by multiple threads.  As long as two
      threads each have their own <classname>shared_connection_block</classname> object,
      then they may use them in safety, even if both <classname>shared_connection_block</classname>
      objects are copies and refer to the same underlying connection.
    </para>
    <para>
      The <classname>signals2::slot</classname> class has no internal mutex locking
      built into it.  It is expected that slot objects will be created then
      connected to a signal in a single thread.  Once they have been copied into
      a signal's slot list, they are protected by the mutex associated with
      each signal-slot connection.
    </para>
    <para>The <classname>signals2::trackable</classname> class does NOT provide
      thread-safe automatic connection management.  In particular, it leaves open the
      possibility of a signal invocation calling into a partially destructed object
      if the trackable-derived object is destroyed in a different thread from the
      one invoking the signal.
      <classname>signals2::trackable</classname> is only provided as a convenience
      for porting single-threaded code from Boost.Signals to Boost.Signals2.
    </para>
  </section>
</section>