problems.xml

<?xml version="1.0" encoding="iso-8859-1" ?>
<!DOCTYPE chapter SYSTEM "chapter.dtd">
<chapter>
  <header>
    <title>Troubleshooting, Problems and 'Gotchas'</title>
    <prepared>Matthias Lang</prepared>
<docno></docno>
<date>2007-09-12</date>
    <rev>1.0</rev>
    <file>problems.xml</file>
  </header>

<p>

	This section looks at problems which trip up many beginners,
	starting with some minor syntactic irritations. If you
	bump into something confusing which isn't listed here,
	try posting a question to the <seealso marker="mailing-lists">
	mailing lists</seealso>.

</p>

<section><title>Why can't I write a = 3 (badmatch)?</title>
<p>

	In Erlang, all variables must start with a capital letter,
	so you can write <c>A = 3.</c> but not 
	<c>a = 3.</c> In the latter case Erlang 
	complains that this is a "bad match".
	</p><p>

	In the second version, the 'a' is an atom while 3 is a number.
	'a' cannot be <em>pattern matched</em> with 3. The use of 
	capital letters for variables comes from Prolog.
	</p><p>

	If the terms <em>atom</em> and <em>pattern match </em> are new to you,
	have a read of the first chapter or two of the
        <url href="http://www.erlang.org/doc.html"> Erlang book</url>.
	(<em>Very</em> roughly, an atom corresponds to 
	an enum in C and pattern matching in this case is like an 
	<c>assert</c> statement.

</p></section>
<section><title>Why can't I change the value of a variable?</title>
<p>

	Erlang only lets you assign a variable once in a function.
	Some simple examples of when you bump into this are:

	</p>
	<pre>
		1> A = 3.
		3
		2> A = 4.
		** exited: {{badmatch,4},[{erl_eval,expr,3}]} **
		3> A = A + 4.
		** exited: {{badmatch,7},[{erl_eval,expr,3}]} **
	</pre>
        <p>

	The usual "workaround" is to use a new variable:

	</p>
	<pre>
		4> B = A + 4.
		7
	</pre>

        <p>
	(This behaviour is called "single assignment", and it's
	considered to be a feature, not a bug. Most functional
	languages, including ML and Haskell, also behave this way.
	Among other things, It's nice being able to rely on a 
	certain variable always having the same value.)
        </p>
	
</section>
<section><title>Why do lists of numbers get printed incorrectly?</title>
<p>

	Sometimes the shell (or a program) prints a list of
	number in an unexpected way, for instance:

	</p>
	<pre>
	1> [65, 66, 67].
	"ABC"
	</pre>

<p>
	This happens because Erlang represents strings as lists of
	integers, so if you ask the shell to print a list of
	integers, the shell takes a guess as to whether you want
	to see it as a list of numbers or as a string. The shell bases
	its guess on checking whether or not the list contains all
	printable characters, so you can force a string to be unprintable:
	
	</p>
	<pre>
	5> [0, 65, 66, 67].
	[0,65,66,67]
	</pre>

<p>
	A similar problem occurs with <c>io:fwrite()</c>,
	but in that case you can take direct control by specifying the
	appropriate formatting character. "~s" always prints the 
	argument as a string, "~w" always prints it as a list.
</p>

</section>
<section><title> Why can't I call arbitrary functions in a guard?</title>
<p>

        If that was allowed, there would be no guarantee that guards
        were side-effect free.
        </p><p>

	Also, it is convenient to be able to program as though guards do not
	consume any significant amount of execution time. There's
	a list of BIFs which can be called from within guards in
	the Erlang book and the standard Erlang spec, some examples
	are <c>size(), length(), integer(), record()</c>.
	</p><p>

	The "problem" often crops up when using <c>if</c>:

	</p>
	<code>
	issue_warning() ->
	  if (os:type() == {win32, windows}) ->    %% illegal guard
	    ok = io:fwrite("you are using windows\\n");
	  true ->
	    ok = io:fwrite("no problem\\n")
	  end.
	</code>

        <p>
        The solution is usually to use <c>case</c> 
	instead. Case is used much more frequently than 
	<c>if</c> in most Erlang programs:
	
	</p>
	<code>
	issue_warning() ->
          case os:type() of
            {win32, windows} -> ok = io:fwrite("you are using windows\\n");
            _ -> ok = io:fwrite("no problem\\n")
	  end.
	</code>

</section>
<section><title> Why can't I use "or" in a guard?</title>
<p>

	You can. Since R6A, several guards separated with semicolons
	perform a logical 'or', for example:
	
	</p>
	<code>
	f(N) when (N - 1) > 3; atom(N) -> yes;
	f(N) -> no.
	</code>

<p>
	will return <c>yes</c> for 
	<c>f(5)</c> as well as 
	<c>f(blork)</c>. Similarly, the comma is 
	used for logical and:

	</p>
	<code>
	g(N) when integer(N), N > 5 -> yes;
	g(N) -> no.
	</code>

<p>
	Beware! The semicolon operator
	doesn't mean exactly the same thing as 'or', for instance
	<c>f(hello)</c> returns <c>yes </c>, whereas evaluating 
	<c>f(hello - 1) > 3 or atom(hello)</c> 
	returns a <c>badarith</c> error.
	The first example, f(), is <em>defined</em>
	 as having the same effect as writing:

	</p>
	<code>
	f(N) when (N - 1) > 3 -> yes;
	f(N) when atom(N) -> yes;
	f(N) -> no.
	</code>

</section>
<section><title> Why does 'catch' give me a syntax error?</title>
<p>

	It seems natural enough to write
</p>
	
	<code>
	2> A = catch 1/0.
	** 2: syntax error before: 'catch' **
	</code>

<p>
	But the parsing rules for Erlang are a bit unexpected here, 
	catch binds less tightly than you might expect it to.
	To make it parse the way you really want:
</p>

	<code>
	3> A = (catch 1/0).
	{'EXIT',{badarith,[{erl_eval,eval_op,3},
                   {erl_eval,expr,3},
                   {erl_eval,exprs,4},
                   {shell,eval_loop,2}]}}
	</code>

</section>
<section><title> What is a sticky directory?</title>
<p>

	This typically crops up when playing with examples from the
	old (last century) Erlang book, for instance:

</p>
	
	<code>
        1> c(sets.erl).
        {error,sticky_directory}
	</code>

<p>
	The problem is that there already is a standard library module 
	called 'sets'. The Erlang runtime system is protecting
	you.

	</p><p>

	The easiest solution is to rename your module, e.g. to
	<c>mysets.erl</c>. It is also possible to
	'un-stick' the directory containing the library module.

</p></section>
<section><title>Why won't my distributed Erlang nodes communicate?</title><p>

	For Erlang nodes to be able to communicate, you need 
</p>

	<list>
	<item><p>A working tcp network between the nodes. 
		On unix systems	you can check this by using telnet, 
		though a working telnet doesn't guarantee that
		enough of your network is working, e.g. DNS problems
		throw a spanner in Erlang's distribution mechanisms.
</p></item>

	<item><p>The nodes to use the same node naming scheme 
		(you cannot have a system where some nodes use fully 
		qualified names and others use short names).
</p></item>

	<item><p>The nodes must agree to use the same 
		"magic security cookie".
</p></item>
	</list>

<p>
	Here's an example of how to create two nodes on different machines 
	called <em>martell</em> and <em>grolsch</em>
	and verify that they're connected. On one machine:

	</p>
	<code>
	~ >rlogin martell
	Last login: Sat Feb 5 20:40:52 from super
	~ >erl -sname first_node 
	Eshell V4.9.1.1  (abort with ^G)
	(first_node@martell)1> erlang:set_cookie(first_node, nocookie).
	true
	</code>
<p>

	And on the other

	</p>
	<code>
	~ >rlogin grolsch
	Last login: Thu Feb 3 10:54:20 from :0
	~ >erl -sname second_node
	Eshell V4.9.1.1  (abort with ^G)
	(second_node@grolsch)1> erlang:set_cookie(second_node, nocookie).
	true
	(second_node@grolsch)2> net:ping(first_node@martell).
	pong
	(second_node@grolsch)3> rpc:call(first_node@martell, os, type, []).
	{unix,sunos}
	</code>
<p>

	The <c>pong</c> tells us that the 
	connection works, the result
	of <c>net:ping()</c> is <c>pang </c> when the connection isn't working.
	The <c>rpc:call()</c> command illustrates 
	executing a command on the other node.
	</p>

	<warning>
	<p>Cookies are a password for Erlang nodes. The example above
	sets the password to "nocookie", which removes almost
	all security. Anyone on your network can use your Erlang node
	for anything, including deleting all of your files. </p>
	</warning>

</section>
<section><title>When distribution won't work</title>

<p>
	One simple reason why distribution won't work is if you're
	combining different versions of Erlang.

	At the time of writing (R13B), the OTP group's strategy is to
	maintain backwards compatibility with the two previous major
	versions of Erlang/OTP, i.e. R13B works with all R13B minor
	releases as well as with R12B and R11B.
</p><p>
	
	Beyond that, you need to start digging. Erlang nodes communicate
	by connecting to the <c>epmd</c> daemon. 
	The daemon is started
	automatically the first time you start a distributed node. For
	debugging, it's useful to kill it by running <c>epmd -kill
	</c>) and
	then restart it with the debugging flag. You can find 
	<c>epmd</c> in the same directory as the erlang binary:
	
	</p>
	<code>
	~ >/otp/releases/otp_beam_sunos5_r6b/erts-4.9.1/bin/epmd -d -d
	epmd: Sat Feb 5 21:04:39 2000: epmd running - daemon = 0
	epmd: Sat Feb 5 21:04:39 2000: try to initiate listening port 4369
	epmd: Sat Feb 5 21:04:39 2000: starting
	epmd: Sat Feb 5 21:04:39 2000: entering the main select() loop
	</code>

<p>
	When you start a distributed node on the same system, you
	should see a message in the <c>epmd</c> window:

	</p>
	<code>
	epmd: Sat Feb 5 21:07:33 2000: registering 'first_node:1', port 53566
	</code>

<p>
	Similarly, when you use <c>net:ping</c> from a 
	node on another
	system, you should see some messages go by. If nothing seems
	to be happening on <c>epmd</c>, 
	it's useful to scan for
	packets on the ethernet to see if the information is really
	getting out on the net. On unix systems you can do this with
	tcpdump:

	</p>
	<code>
	# /usr/local/sbin/tcpdump port 4369
	tcpdump: listening on le0
	21:10:17.349286 grolsch.37558 > martell.4369: S 747683789:747683789(0) 
	</code>

</section>
<section><title>Avoiding DNS</title><p>

	One fairly common obstacle to getting distribution to work is
	a broken DNS setup. One way to avoid that, or test for it, is
	to start distributed Erlang with IP addresses instead of hostnames, 
	e.g. on one machine or window:
	</p>

        <code>
        ~ >erl -name first_node@127.0.0.1
        Erlang (BEAM) emulator version 5.5.5 [source] [64-bit]
        Eshell V5.5.5  (abort with ^G)
	(first_node@127.0.0.1)1> erlang:set_cookie('second_node@127.0.0.1', nocookie).
	true
	(first_node@127.0.0.1)2> net:ping('second_node@127.0.0.1').
	pong
        </code>

	<p>
        and in the other:
	</p>

        <code>
        ~ >erl -name second_node@127.0.0.1
        Erlang (BEAM) emulator version 5.5.5 [source] [64-bit] [async-threads:0]
        Eshell V5.5.5  (abort with ^G)
        (second_node@127.0.0.1)1> erlang:set_cookie('first_node@127.0.0.1', nocookie).
        </code>

</section>
<section><title>
Why does my application die every second time I load new code into it?
</title><p>

	Erlang's code replacement system is based around there being
	(up to) two copies of the code loaded at any time, these
	are called "old" and "new". When you load new code, the
	current version becomes "old" and the "old" code is thrown
	away. Any processes still running "old" code are killed.
	</p><p>

	You can check if there is any old code for a particular
	module still running:
	
	</p>
	<code>
	Eshell V4.9.1  (abort with ^G)
	1> l(erl).
	{module,erl}
	2> code:soft_purge(erl).
	false
	</code>

<p>
	In this case old code is still running. You can also check
	if a particular process is running old code, by using
	<c>erlang:check_process_code(Pid, Module)</c>.

</p></section>
<section><title>Why can't I open devices (e.g. a serial port) like normal files?
</title><p>

	Short answer: because the erlang runtime system was not
	designed to do that.

	The Erlang runtime system's internal file access system,
	efile, must avoid blocking, otherwise the whole Erlang
	system will block. This is not a good thing in a soft
	real-time system.

	When accessing regular files, it's generally a reasonable
	assumption that operations will not block. Devices, on the
	other hand, are quite likely to block. Some devices, such
	as serial ports, may block indefinitely.

        There are several possible ways to solve this. The Erlang
	runtime system could be altered, or an external port program
	could be used to access the device.

	Two mailing list discussions about the topic can be found
	<url href="http://erlang.org/pipermail/erlang-questions/2004-January/011351.html">here</url> and
	<url 
href="http://erlang.org/pipermail/erlang-questions/2000-March/001050.html">
here</url>.
	</p>

</section>
<section><title>
Why doesn't the stack backtrace show the right functions for
this code:
</title>
	<code>
	-module(erl).
	-export([a/0]).

	a() -> b().
	b() -> c().
	c() -> 3 = 4.           %% will cause badmatch
	</code>

<p>

	The stack backtrace only shows function c(), rather
	than a(), b() and c(). This is because of last-call-optimisation;
	the compiler knows it does not need to generate a stack frame
	for a() or b() because the last thing it did was call another
	function, hence the stack frame does not appear in the stack backtrace.

</p></section>
<section><title>What do I do if I get an error when compiling Open Source Erlang?
</title><p>

	Consider using a binary release (Debian GNU/Linux includes them in the
	woody and potato releases, windows binaries are on the 
	<url href="http://www.erlang.org/download.html"> 
	download page </url>).

	</p><p>

	Ask for help on the mailing list, providing as many details
	as you can, including
</p>

	<list>
	<item><p> your system details (OS, OS version, compiler version)
</p></item>
	<item><p> the error messages you get
</p></item>
	<item><p> any errors you got when you ran <c>configure
	</c> </p></item>
	</list>

</section>
<section><title>Why do errors in the shell kill other processes?</title>
<p>

	Because the other processes are linked to the shell process.

</p><p>

	Each time an error occurs in the erlang shell, the shell
        process exits and a new one is started:
</p>

	<pre><![CDATA[
	1> self().
	<0.23.0>
	2> x - y.
	** exited: {badarith,[{erl_eval,eval_op,3},
                      {erl_eval,exprs,4},
                      {shell,eval_loop,2}]} **
	3> self().
	<0.40.0>
	]]></pre>

<p>
	Thus all processes linked to &lt;0.23.0&gt; will receive exit
	signals. Unless they trap exits, they will exit too. One
	way to avoid this is avoid linking "background" processes
	to shell processes, for instance by using <c>spawn</c>
	instead of <c>spawn_link</c>.
</p>

</section>
</chapter>