path: root/lib/kernel/doc/src/seq_trace.xml

<?xml version="1.0" encoding="utf-8" ?>
<!DOCTYPE erlref SYSTEM "erlref.dtd">

<erlref>
  <header>
    <copyright>
      <year>1998</year><year>2016</year>
      <holder>Ericsson AB. All Rights Reserved.</holder>
    </copyright>
    <legalnotice>
      Licensed under the Apache License, Version 2.0 (the "License");
      you may not use this file except in compliance with the License.
      You may obtain a copy of the License at
 
          http://www.apache.org/licenses/LICENSE-2.0

      Unless required by applicable law or agreed to in writing, software
      distributed under the License is distributed on an "AS IS" BASIS,
      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
      See the License for the specific language governing permissions and
      limitations under the License.
    
    </legalnotice>

    <title>seq_trace</title>
    <prepared>[email protected]</prepared>
    <docno></docno>
    <date>1998-04-16</date>
    <rev>A</rev>
  </header>
  <module>seq_trace</module>
  <modulesummary>Sequential tracing of messages.</modulesummary>
  <description>
    <p>Sequential tracing makes it possible to trace all messages
      resulting from one initial message. Sequential tracing is
      independent of the ordinary tracing in Erlang, which
      is controlled by the <c>erlang:trace/3</c> BIF. For more information
      about what sequential tracing is and how it can be used, see section
      <seealso marker="#whatis">Sequential Tracing</seealso>.</p>
    <p><c>seq_trace</c> provides functions that control all aspects of
      sequential tracing. There are functions for activation,
      deactivation, inspection, and for collection of the trace output.</p>
  </description>
  <datatypes>
    <datatype>
      <name name="token"/>
      <desc>
        <p>An opaque term (a tuple) representing a trace token.</p>
      </desc>
    </datatype>
  </datatypes>
  <funcs>
    <func>
      <name name="set_token" arity="1"/>
      <fsummary>Set the trace token</fsummary>
      <desc>
        <p>Sets the trace token for the calling process to <c><anno>Token</anno></c>.
          If <c><anno>Token</anno> == []</c> then tracing is disabled, otherwise
          <c><anno>Token</anno></c> should be an Erlang term returned from
          <c>get_token/0</c> or <c>set_token/1</c>. <c>set_token/1</c>
          can be used to temporarily exclude message passing from
          the trace by setting the trace token to empty like this:</p>
        <code type="none">
OldToken = seq_trace:set_token([]), % set to empty and save 
                                    % old value
% do something that should not be part of the trace
io:format("Exclude the signalling caused by this~n"),
seq_trace:set_token(OldToken), % activate the trace token again
...  </code>
        <p>Returns the previous value of the trace token.</p>
      </desc>
    </func>
    <func>
      <name name="set_token" arity="2"/>
      <fsummary>Set a component of the trace token</fsummary>
      <type name="component"/>
      <type name="flag"/>
      <type name="value"/>
      <desc>
        <p>Sets the individual <c><anno>Component</anno></c> of the trace token to
          <c><anno>Val</anno></c>. Returns the previous value of the component.</p>
        <taglist>
          <tag><c>set_token(label, <anno>Integer</anno>)</c></tag>
          <item>
            <p>The <c>label</c> component is an integer which
              identifies all events belonging to the same sequential
              trace. If several sequential traces can be active
              simultaneously, <c>label</c> is used to identify
              the separate traces. Default is 0.</p>
          </item>
          <tag><c>set_token(serial, SerialValue)</c></tag>
          <item>
            <p><c>SerialValue = {<anno>Previous</anno>, <anno>Current</anno>}</c>.
              The <c>serial</c> component contains counters which
              enables the traced messages to be sorted, should never be
              set explicitly by the user as these counters are updated
              automatically. Default is <c>{0, 0}</c>.</p>
          </item>
          <tag><c>set_token(send, <anno>Bool</anno>)</c></tag>
          <item>
            <p>A trace token flag (<c>true | false</c>) which
              enables/disables tracing on message sending. Default is
              <c>false</c>.</p>
          </item>
          <tag><c>set_token('receive', <anno>Bool</anno>)</c></tag>
          <item>
            <p>A trace token flag (<c>true | false</c>) which
              enables/disables tracing on message reception. Default is
              <c>false</c>.</p>
          </item>
          <tag><c>set_token(print, <anno>Bool</anno>)</c></tag>
          <item>
            <p>A trace token flag (<c>true | false</c>) which
              enables/disables tracing on explicit calls to
              <c>seq_trace:print/1</c>. Default is <c>false</c>.</p>
          </item>
          <tag><c>set_token(timestamp, <anno>Bool</anno>)</c></tag>
          <item>
            <p>A trace token flag (<c>true | false</c>) which
              enables/disables a timestamp to be generated for each
              traced event. Default is <c>false</c>.</p>
          </item>
          <tag><c>set_token(strict_monotonic_timestamp, <anno>Bool</anno>)</c></tag>
          <item>
            <p>A trace token flag (<c>true | false</c>) which
            enables/disables a strict monotonic timestamp to be generated
	    for each traced event. Default is <c>false</c>. Timestamps will
	    consist of
	    <seealso marker="erts:time_correction#Erlang_Monotonic_Time">Erlang
	    monotonic time</seealso> and a monotonically increasing
	    integer. The time-stamp has the same format and value
	    as produced by <c>{erlang:monotonic_time(nanosecond),
	    erlang:unique_integer([monotonic])}</c>.</p>
          </item>
          <tag><c>set_token(monotonic_timestamp, <anno>Bool</anno>)</c></tag>
          <item>
            <p>A trace token flag (<c>true | false</c>) which
            enables/disables a strict monotonic timestamp to be generated
	    for each traced event. Default is <c>false</c>. Timestamps
	    will use
	    <seealso marker="erts:time_correction#Erlang_Monotonic_Time">Erlang
	    monotonic time</seealso>. The time-stamp has the same
	    format and value as produced by
	    <c>erlang:monotonic_time(nanosecond)</c>.</p>
          </item>
        </taglist>
	  <p>If multiple timestamp flags are passed, <c>timestamp</c> has
	  precedence over <c>strict_monotonic_timestamp</c> which
	  in turn has precedence over <c>monotonic_timestamp</c>. All
	  timestamp flags are remembered, so if two are passed
	  and the one with highest precedence later is disabled
	  the other one will become active.</p>
      </desc>
    </func>
    <func>
      <name name="get_token" arity="0"/>
      <fsummary>Return the value of the trace token</fsummary>
      <desc>
        <p>Returns the value of the trace token for the calling process.
          If <c>[]</c> is returned, it means that tracing is not active.
          Any other value returned is the value of an active trace
          token. The value returned can be used as input to
          the <c>set_token/1</c> function.</p>
      </desc>
    </func>
    <func>
      <name name="get_token" arity="1"/>
      <fsummary>Return the value of a trace token component</fsummary>
      <type name="component"/>
      <type name="flag"/>
      <type name="value"/>
      <desc>
        <p>Returns the value of the trace token component
          <c>Component</c>. See
          <seealso marker="#set_token/2">set_token/2</seealso> for
          possible values of <c>Component</c> and <c>Val</c>.</p>
      </desc>
    </func>
    <func>
      <name name="print" arity="1"/>
      <fsummary>Put the Erlang term <c>TraceInfo</c>into the sequential trace output</fsummary>
      <desc>
        <p>Puts the Erlang term <c><anno>TraceInfo</anno></c> into the sequential
          trace output if the calling process currently is executing
          within a sequential trace and the <c>print</c> flag of
          the trace token is set.</p>
      </desc>
    </func>
    <func>
      <name name="print" arity="2"/>
      <fsummary>Put the Erlang term <c>TraceInfo</c>into the sequential trace output</fsummary>
      <desc>
        <p>Same as <c>print/1</c> with the additional condition that
          <c><anno>TraceInfo</anno></c> is output only if <c>Label</c> is equal to
          the label component of the trace token.</p>
      </desc>
    </func>
    <func>
      <name name="reset_trace" arity="0"/>
      <fsummary>Stop all sequential tracing on the local node</fsummary>
      <desc>
        <p>Sets the trace token to empty for all processes on the
          local node. The process internal counters used to create
          the serial of the trace token is set to 0. The trace token is
          set to empty for all messages in message queues. Together
          this will effectively stop all ongoing sequential tracing in
          the local node.</p>
      </desc>
    </func>
    <func>
      <name name="set_system_tracer" arity="1"/>
      <fsummary>Set the system tracer</fsummary>
      <type name="tracer"/>
      <desc>
        <p>Sets the system tracer. The system tracer can be either a
          process, port or <seealso marker="erts:erl_tracer">tracer module</seealso>
          denoted by <c><anno>Tracer</anno></c>.
          Returns the previous value (which can be <c>false</c> if no system
          tracer is active).</p>
        <p>Failure: <c>{badarg, Info}}</c> if <c><anno>Pid</anno></c> is not an
          existing local pid.</p>
      </desc>
    </func>
    <func>
      <name name="get_system_tracer" arity="0"/>
      <fsummary>Return the pid() or port() of the current system tracer.</fsummary>
      <type name="tracer"/>
      <desc>
        <p>Returns the pid, port identifier or tracer module of the current system
          tracer or <c>false</c> if no system tracer is activated.</p>
      </desc>
    </func>
  </funcs>

  <section>
    <title>Trace Messages Sent to the System Tracer</title>
    <p>The format of the messages is one of the following, depending on if
      flag <c>timestamp</c> of the trace token is set to <c>true</c> or
      <c>false</c>:</p>
    <code type="none">
{seq_trace, Label, SeqTraceInfo, TimeStamp}</code>
    <p>or</p>
    <code type="none">
{seq_trace, Label, SeqTraceInfo}</code>
    <p>Where:</p>
    <code type="none">
Label = int()
TimeStamp = {Seconds, Milliseconds, Microseconds}  
  Seconds = Milliseconds = Microseconds = int()</code>
    <p><c>SeqTraceInfo</c> can have the following formats:</p>
    <taglist>
      <tag><c>{send, Serial, From, To, Message}</c></tag>
      <item>
        <p>Used when a process <c>From</c> with its trace token flag
          <c>print</c> set to <c>true</c> has sent a message.</p>
      </item>
      <tag><c>{'receive', Serial, From, To, Message}</c></tag>
      <item>
        <p>Used when a process <c>To</c> receives a message with a
          trace token that has flag <c>'receive'</c> set to <c>true</c>.</p>
      </item>
      <tag><c>{print, Serial, From, _, Info}</c></tag>
      <item>
        <p>Used when a process <c>From</c> has called
          <c>seq_trace:print(Label, TraceInfo)</c> and has a trace
          token with flag <c>print</c> set to <c>true</c>, and
          <c>label</c> set to <c>Label</c>.</p>
      </item>
    </taglist>
    <p><c>Serial</c> is a tuple <c>{PreviousSerial, ThisSerial}</c>,
      where:</p>
    <list type="bulleted">
      <item><p>Integer <c>PreviousSerial</c> denotes the serial
        counter passed in the last received message that carried a trace
        token. If the process is the first in a new sequential trace,
        <c>PreviousSerial</c> is set to the value of the process internal
        "trace clock".</p></item>
      <item><p>Integer <c>ThisSerial</c> is the serial
        counter that a process sets on outgoing messages. It is based
        on the process internal "trace clock", which is incremented by one
        before it is attached to the trace token in the message.</p></item>
    </list>
  </section>

  <section>
    <marker id="whatis"></marker>
    <title>Sequential Tracing</title>
    <p>Sequential tracing is a way to trace a sequence of messages sent
      between different local or remote processes, where the sequence
      is initiated by a single message. In short, it works as follows:</p>
    <p>Each process has a <em>trace token</em>, which can be empty or
      not empty. When not empty, the trace token can be seen as
      the tuple <c>{Label, Flags, Serial, From}</c>. The trace token is
      passed invisibly with each message.</p>
    <p>To start a sequential trace, the user must explicitly set
      the trace token in the process that will send the first message
      in a sequence.</p>
    <p>The trace token of a process is set each time the process
      matches a message in a receive statement, according to the trace
      token carried by the received message, empty or not.</p>
    <p>On each Erlang node, a process can be set as the <em>system tracer</em>.
      This process will receive trace messages each time
      a message with a trace token is sent or received (if the trace
      token flag <c>send</c> or <c>'receive'</c> is set). The system
      tracer can then print each trace event, write it to a file, or
      whatever suitable.</p>
    <note>
      <p>The system tracer only receives those trace events that
        occur locally within the Erlang node. To get the whole picture
        of a sequential trace, involving processes on many Erlang
        nodes, the output from the system tracer on each involved node
        must be merged (offline).</p>
    </note>
    <p>The following sections describe sequential tracing and its most
      fundamental concepts.</p>
  </section>

  <section>
    <title>Trace Token</title>
    <p>Each process has a current trace token. Initially, the token is
      empty. When the process sends a message to another process, a
      copy of the current token is sent "invisibly" along with
      the message.</p>
    <p>The current token of a process is set in one of the following two
      ways:</p>
    <list type="bulleted">
      <item>
        <p>Explicitly by the process itself, through a call to
          <c>seq_trace:set_token/1,2</c></p>
      </item>
      <item>
        <p>When a message is received</p>
      </item>
    </list>
    <p>In both cases, the current token is set. In particular, if
      the token of a received message is empty, the current token of
      the process is set to empty.</p>
    <p>A trace token contains a label and a set of flags. Both
      the label and the flags are set in both alternatives above.</p>
  </section>

  <section>
    <title>Serial</title>
    <p>The trace token contains a component called
      <c>serial</c>. It consists of two integers, <c>Previous</c> and
      <c>Current</c>. The purpose is to uniquely identify each traced
      event within a trace sequence, as well as to order the messages
      chronologically and in the different branches, if any.</p>
    <p>The algorithm for updating <c>Serial</c> can be described as
      follows:</p>
    <p>Let each process have two counters, <c>prev_cnt</c> and
      <c>curr_cnt</c>, both are set to <c>0</c> when a process is created.
      The counters are updated at the following occasions:</p>
    <list type="bulleted">
      <item>
        <p><em>When the process is about to send a message and the trace token
          is not empty.</em></p>
        <p>Let the serial of the trace token be <c>tprev</c> and
          <c>tcurr</c>.</p>
        <pre>
curr_cnt := curr_cnt + 1
tprev := prev_cnt
tcurr := curr_cnt</pre>
        <p>The trace token with <c>tprev</c> and <c>tcurr</c> is then
          passed along with the message.</p>
      </item>
      <item>
        <p><em>When the process calls</em> <c>seq_trace:print(Label, Info)</c>,
          <c>Label</c> <em>matches the label part of the trace token and the
          trace token print flag is <c>true</c>.</em></p>
        <p>The algorithm is the same as for send above.</p>
      </item>
      <item>
        <p><em>When a message is received and contains a non-empty trace
          token.</em></p>
        <p>The process trace token is set to the trace token from
          the message.</p>
        <p>Let the serial of the trace token be <c>tprev</c> and
          <c>tcurr</c>.</p>
        <code>
<![CDATA[if (curr_cnt < tcurr )]]>
   curr_cnt := tcurr
prev_cnt := tcurr</code>
      </item>
    </list>
    <p><c>curr_cnt</c> of a process is incremented each time
      the process is involved in a sequential trace. The counter can
      reach its limit (27 bits) if a process is very long-lived and is
      involved in much sequential tracing. If the counter overflows, the
      serial for ordering of the trace events cannot be used. To prevent
      the counter from overflowing in the middle of
      a sequential trace, function <c>seq_trace:reset_trace/0</c>
      can be called to reset <c>prev_cnt</c> and <c>curr_cnt</c> of
      all processes in the Erlang node. This function also sets all
      trace tokens in processes and their message queues to empty, and
      thus stops all ongoing sequential tracing.</p>
  </section>

  <section>
    <title>Performance Considerations</title>
    <p>The performance degradation for a system that is enabled for
      sequential tracing is negligible as long as no tracing is
      activated. When tracing is activated, there is an
      extra cost for each traced message, but all other messages are
      unaffected.</p>
  </section>

  <section>
    <title>Ports</title>
    <p>Sequential tracing is not performed across ports.</p>
    <p>If the user for some reason wants to pass the trace token to a
      port, this must be done manually in the code of the port
      controlling process. The port controlling processes have to check
      the appropriate sequential trace settings (as obtained from
      <c>seq_trace:get_token/1</c>) and include trace information in
      the message data sent to their respective ports.</p>
    <p>Similarly, for messages received from a port, a port controller
      has to retrieve trace-specific information, and set appropriate
      sequential trace flags through calls to
      <c>seq_trace:set_token/2</c>.</p>
  </section>

  <section>
    <title>Distribution</title>
    <p>Sequential tracing between nodes is performed transparently.
      This applies to C-nodes built with <c>Erl_Interface</c> too. A C-node
      built with <c>Erl_Interface</c> only maintains one trace token, which
      means that the C-node appears as one process from
      the sequential tracing point of view.</p>
  </section>

  <section>
    <title>Example of Use</title>
    <p>This example gives a rough idea of how the new
      primitives can be used and what kind of output it produces.</p>
    <p>Assume that you have an initiating process with
      <c><![CDATA[Pid == <0.30.0>]]></c> like this:</p>
    <code type="none">
-module(seqex).
-compile(export_all).

loop(Port) ->
    receive 
        {Port,Message} ->
            seq_trace:set_token(label,17),
            seq_trace:set_token('receive',true),
            seq_trace:set_token(print,true),
            seq_trace:print(17,"**** Trace Started ****"),
            call_server ! {self(),the_message};
        {ack,Ack} ->
            ok
    end,
    loop(Port).</code>
    <p>And a registered process <c>call_server</c> with
      <c><![CDATA[Pid == <0.31.0>]]></c> like this:</p>
    <code type="none">
loop() ->
    receive
        {PortController,Message} ->
            Ack = {received, Message},
            seq_trace:print(17,"We are here now"),
            PortController ! {ack,Ack}
    end,
    loop().</code>
    <p>A possible output from the system's <c>sequential_tracer</c> can be
      like this:</p>
    <pre>
17:&lt;0.30.0> Info {0,1} WITH
"**** Trace Started ****"
17:&lt;0.31.0> Received {0,2} FROM &lt;0.30.0> WITH
{&lt;0.30.0>,the_message}
17:&lt;0.31.0> Info {2,3} WITH
"We are here now"
17:&lt;0.30.0> Received {2,4} FROM &lt;0.31.0> WITH
{ack,{received,the_message}}</pre>
    <p>The implementation of a system tracer process that produces
      this printout can look like this:</p>
    <code type="none">
tracer() ->
    receive
        {seq_trace,Label,TraceInfo} ->
           print_trace(Label,TraceInfo,false);
        {seq_trace,Label,TraceInfo,Ts} ->
           print_trace(Label,TraceInfo,Ts);
        Other -> ignore
    end,
    tracer().        

print_trace(Label,TraceInfo,false) ->
    io:format("~p:",[Label]),
    print_trace(TraceInfo);
print_trace(Label,TraceInfo,Ts) ->
    io:format("~p ~p:",[Label,Ts]),
    print_trace(TraceInfo).

print_trace({print,Serial,From,_,Info}) ->
    io:format("~p Info ~p WITH~n~p~n", [From,Serial,Info]);
print_trace({'receive',Serial,From,To,Message}) ->
    io:format("~p Received ~p FROM ~p WITH~n~p~n", 
              [To,Serial,From,Message]);
print_trace({send,Serial,From,To,Message}) ->
    io:format("~p Sent ~p TO ~p WITH~n~p~n", 
              [From,Serial,To,Message]).</code>
    <p>The code that creates a process that runs this tracer function
      and sets that process as the system tracer can look like this:</p>
    <code type="none">
start() ->
    Pid = spawn(?MODULE,tracer,[]),
    seq_trace:set_system_tracer(Pid), % set Pid as the system tracer 
    ok.</code>
    <p>With a function like <c>test/0</c>, the whole example can be
      started:</p>
    <code type="none">
test() ->
    P = spawn(?MODULE, loop, [port]),
    register(call_server, spawn(?MODULE, loop, [])),
    start(),
    P ! {port,message}.</code>
  </section>
</erlref>