<?xml version="1.0"?> <!-- -*- sgml -*- -->
<!DOCTYPE chapter PUBLIC "-//OASIS//DTD DocBook XML V4.2//EN"
  "http://www.oasis-open.org/docbook/xml/4.2/docbookx.dtd"
[ <!ENTITY % vg-entities SYSTEM "vg-entities.xml"> %vg-entities; ]>


<chapter id="manual-core" xreflabel="Valgrind's core">
<title>Using and understanding the Valgrind core</title>

<para>This section describes the Valgrind core services, flags and
behaviours.  That means it is relevant regardless of what particular
tool you are using.  A point of terminology: most references to
"valgrind" in the rest of this section (Section 2) refer to the Valgrind
core services.</para>

<sect1 id="manual-core.whatdoes" 
       xreflabel="What Valgrind does with your program">
<title>What Valgrind does with your program</title>

<para>Valgrind is designed to be as non-intrusive as possible. It works
directly with existing executables. You don't need to recompile, relink,
or otherwise modify, the program to be checked.</para>

<para>Simply put 
<computeroutput>valgrind --tool=tool_name</computeroutput> 
at the start of the command line normally used to run the program.  For
example, if want to run the command 
<computeroutput>ls -l</computeroutput> using the heavyweight
memory-checking tool Memcheck, issue the command:</para>

<programlisting><![CDATA[
valgrind --tool=memcheck ls -l]]></programlisting>

<para>(Memcheck is the default, so if you want to use it you can
actually omit the <option>--tool</option> flag.</para>

<para>Regardless of which tool is in use, Valgrind takes control of your
program before it starts.  Debugging information is read from the
executable and associated libraries, so that error messages and other
outputs can be phrased in terms of source code locations (if that is
appropriate).</para>

<para>Your program is then run on a synthetic CPU provided by the
Valgrind core.  As new code is executed for the first time, the core
hands the code to the selected tool.  The tool adds its own
instrumentation code to this and hands the result back to the core,
which coordinates the continued execution of this instrumented
code.</para>

<para>The amount of instrumentation code added varies widely between
tools.  At one end of the scale, Memcheck adds code to check every
memory access and every value computed, increasing the size of the code
at least 12 times, and making it run 25-50 times slower than natively.
At the other end of the spectrum, the ultra-trivial "none" tool
(a.k.a. Nulgrind) adds no instrumentation at all and causes in total
"only" about a 4 times slowdown.</para>

<para>Valgrind simulates every single instruction your program executes.
Because of this, the active tool checks, or profiles, not only the code
in your application but also in all supporting dynamically-linked
(<computeroutput>.so</computeroutput>-format) libraries, including the
GNU C library, the X client libraries, Qt, if you work with KDE, and so
on.</para>

<para>If you're using one of the error-detection tools, Valgrind will
often detect errors in libraries, for example the GNU C or X11
libraries, which you have to use.  You might not be interested in these
errors, since you probably have no control over that code.  Therefore,
Valgrind allows you to selectively suppress errors, by recording them in
a suppressions file which is read when Valgrind starts up.  The build
mechanism attempts to select suppressions which give reasonable
behaviour for the libc and XFree86 versions detected on your machine.
To make it easier to write suppressions, you can use the
<option>--gen-suppressions=yes</option> option which tells Valgrind to
print out a suppression for each error that appears, which you can then
copy into a suppressions file.</para>

<para>Different error-checking tools report different kinds of errors.
The suppression mechanism therefore allows you to say which tool or
tool(s) each suppression applies to.</para>

</sect1>


<sect1 id="manual-core.started" xreflabel="Getting started">
<title>Getting started</title>

<para>First off, consider whether it might be beneficial to recompile
your application and supporting libraries with debugging info enabled
(the <option>-g</option> flag).  Without debugging info, the best
Valgrind tools will be able to do is guess which function a particular
piece of code belongs to, which makes both error messages and profiling
output nearly useless.  With <option>-g</option>, you'll hopefully get
messages which point directly to the relevant source code lines.</para>

<para>Another flag you might like to consider, if you are working with
C++, is <option>-fno-inline</option>.  That makes it easier to see the
function-call chain, which can help reduce confusion when navigating
around large C++ apps.  For whatever it's worth, debugging
OpenOffice.org with Memcheck is a bit easier when using this flag.  You
don't have to do this, but doing so helps Valgrind produce more accurate
and less confusing error reports.  Chances are you're set up like this
already, if you intended to debug your program with GNU gdb, or some
other debugger.</para>

<para>This paragraph applies only if you plan to use Memcheck: On rare
occasions, optimisation levels at <computeroutput>-O2</computeroutput>
and above have been observed to generate code which fools Memcheck into
wrongly reporting uninitialised value errors.  We have looked in detail
into fixing this, and unfortunately the result is that doing so would
give a further significant slowdown in what is already a slow tool.  So
the best solution is to turn off optimisation altogether.  Since this
often makes things unmanagably slow, a plausible compromise is to use
<computeroutput>-O</computeroutput>.  This gets you the majority of the
benefits of higher optimisation levels whilst keeping relatively small
the chances of false complaints from Memcheck.  All other tools (as far
as we know) are unaffected by optimisation level.</para>

<para>Valgrind understands both the older "stabs" debugging format, used
by gcc versions prior to 3.1, and the newer DWARF2 format used by gcc
3.1 and later.  We continue to refine and debug our debug-info readers,
although the majority of effort will naturally enough go into the newer
DWARF2 reader.</para>

<para>When you're ready to roll, just run your application as you
would normally, but place 
<computeroutput>valgrind --tool=tool_name</computeroutput> in front of
your usual command-line invocation.  Note that you should run the real
(machine-code) executable here.  If your application is started by, for
example, a shell or perl script, you'll need to modify it to invoke
Valgrind on the real executables.  Running such scripts directly under
Valgrind will result in you getting error reports pertaining to
<computeroutput>/bin/sh</computeroutput>,
<computeroutput>/usr/bin/perl</computeroutput>, or whatever interpreter
you're using.  This may not be what you want and can be confusing.  You
can force the issue by giving the flag
<option>--trace-children=yes</option>, but confusion is still
likely.</para>

</sect1>


<sect1 id="manual-core.comment" xreflabel="The Commentary">
<title>The Commentary</title>

<para>Valgrind tools write a commentary, a stream of text, detailing
error reports and other significant events.  All lines in the commentary
have following form:

<programlisting><![CDATA[
==12345== some-message-from-Valgrind]]></programlisting>
</para>

<para>The <computeroutput>12345</computeroutput> is the process ID.
This scheme makes it easy to distinguish program output from Valgrind
commentary, and also easy to differentiate commentaries from different
processes which have become merged together, for whatever reason.</para>

<para>By default, Valgrind tools write only essential messages to the
commentary, so as to avoid flooding you with information of secondary
importance.  If you want more information about what is happening,
re-run, passing the <option>-v</option> flag to Valgrind.  A second
<option>-v</option> gives yet more detail.
</para>

<para>You can direct the commentary to three different places:</para>

<orderedlist>

  <listitem id="manual-core.out2fd" xreflabel="Directing output to fd">
    <para>The default: send it to a file descriptor, which is by default
    2 (stderr).  So, if you give the core no options, it will write
    commentary to the standard error stream.  If you want to send it to
    some other file descriptor, for example number 9, you can specify
    <option>--log-fd=9</option>.</para>

    <para>This is the simplest and most common arrangement, but can
    cause problems when valgrinding entire trees of processes which
    expect specific file descriptors, particularly stdin/stdout/stderr,
    to be available for their own use.</para>
  </listitem>

  <listitem id="manual-core.out2file" 
            xreflabel="Directing output to file"> <para>A less intrusive
    option is to write the commentary to a file, which you specify by
    <option>--log-file=filename</option>.  Note carefully that the
    commentary is <command>not</command> written to the file you
    specify, but instead to one called
    <filename>filename.12345</filename>, if for example the pid of the
    traced process is 12345.  This is helpful when valgrinding a whole
    tree of processes at once, since it means that each process writes
    to its own logfile, rather than the result being jumbled up in one
    big logfile.  If <filename>filename.12345</filename> already exists,
    then it will name new files <filename>filename.12345.1</filename>
    and so on.</para>

    <para>If you want to specify precisely the file name to use, without
    the trailing <computeroutput>.12345</computeroutput> part, you can
    instead use <option>--log-file-exactly=filename</option>.</para>

    <para>You can also use the
    <option>--log-file-qualifier=&lt;VAR&gt;</option> option to modify
    the filename via according to the environment variable
    <varname>VAR</varname>.  This is rarely needed, but very useful in
    certain circumstances (eg. when running MPI programs).  In this
    case, the trailing <computeroutput>.12345</computeroutput> part is
    replaced by the contents of <varname>$VAR</varname>.  The idea is
    that you specify a variable which will be set differently for each
    process in the job, for example
    <computeroutput>BPROC_RANK</computeroutput> or whatever is
    applicable in your MPI setup.</para>
  </listitem>

  <listitem id="manual-core.out2socket" 
            xreflabel="Directing output to network socket"> <para>The
    least intrusive option is to send the commentary to a network
    socket.  The socket is specified as an IP address and port number
    pair, like this: <option>--log-socket=192.168.0.1:12345</option> if
    you want to send the output to host IP 192.168.0.1 port 12345 (I
    have no idea if 12345 is a port of pre-existing significance).  You
    can also omit the port number:
    <option>--log-socket=192.168.0.1</option>, in which case a default
    port of 1500 is used.  This default is defined by the constant
    <computeroutput>VG_CLO_DEFAULT_LOGPORT</computeroutput> in the
    sources.</para>

    <para>Note, unfortunately, that you have to use an IP address here,
    rather than a hostname.</para>

    <para>Writing to a network socket is pretty useless if you don't
    have something listening at the other end.  We provide a simple
    listener program,
    <computeroutput>valgrind-listener</computeroutput>, which accepts
    connections on the specified port and copies whatever it is sent to
    stdout.  Probably someone will tell us this is a horrible security
    risk.  It seems likely that people will write more sophisticated
    listeners in the fullness of time.</para>

    <para>valgrind-listener can accept simultaneous connections from up
    to 50 valgrinded processes.  In front of each line of output it
    prints the current number of active connections in round
    brackets.</para>

    <para>valgrind-listener accepts two command-line flags:</para>
    <itemizedlist>
       <listitem>
         <para><option>-e</option> or <option>--exit-at-zero</option>: 
         when the number of connected processes falls back to zero,
         exit.  Without this, it will run forever, that is, until you
         send it Control-C.</para>
       </listitem>
       <listitem>
        <para><option>portnumber</option>: changes the port it listens
        on from the default (1500).  The specified port must be in the
        range 1024 to 65535.  The same restriction applies to port
        numbers specified by a <option>--log-socket</option> to
        Valgrind itself.</para>
      </listitem>
    </itemizedlist>

    <para>If a valgrinded process fails to connect to a listener, for
    whatever reason (the listener isn't running, invalid or unreachable
    host or port, etc), Valgrind switches back to writing the commentary
    to stderr.  The same goes for any process which loses an established
    connection to a listener.  In other words, killing the listener
    doesn't kill the processes sending data to it.</para>
  </listitem>

</orderedlist>

<para>Here is an important point about the relationship between the
commentary and profiling output from tools.  The commentary contains a
mix of messages from the Valgrind core and the selected tool.  If the
tool reports errors, it will report them to the commentary.  However, if
the tool does profiling, the profile data will be written to a file of
some kind, depending on the tool, and independent of what
<option>--log-*</option> options are in force.  The commentary is
intended to be a low-bandwidth, human-readable channel.  Profiling data,
on the other hand, is usually voluminous and not meaningful without
further processing, which is why we have chosen this arrangement.</para>

</sect1>


<sect1 id="manual-core.report" xreflabel="Reporting of errors">
<title>Reporting of errors</title>

<para>When one of the error-checking tools (Memcheck, Addrcheck,
Helgrind) detects something bad happening in the program, an error
message is written to the commentary.  For example:</para>

<programlisting><![CDATA[
==25832== Invalid read of size 4
==25832==    at 0x8048724: BandMatrix::ReSize(int, int, int) (bogon.cpp:45)
==25832==    by 0x80487AF: main (bogon.cpp:66)
==25832==  Address 0xBFFFF74C is not stack'd, malloc'd or free'd]]></programlisting>

<para>This message says that the program did an illegal 4-byte read of
address 0xBFFFF74C, which, as far as Memcheck can tell, is not a valid
stack address, nor corresponds to any currently malloc'd or free'd
blocks.  The read is happening at line 45 of
<filename>bogon.cpp</filename>, called from line 66 of the same file,
etc.  For errors associated with an identified malloc'd/free'd block,
for example reading free'd memory, Valgrind reports not only the
location where the error happened, but also where the associated block
was malloc'd/free'd.</para>

<para>Valgrind remembers all error reports.  When an error is detected,
it is compared against old reports, to see if it is a duplicate.  If so,
the error is noted, but no further commentary is emitted.  This avoids
you being swamped with bazillions of duplicate error reports.</para>

<para>If you want to know how many times each error occurred, run with
the <option>-v</option> option.  When execution finishes, all the
reports are printed out, along with, and sorted by, their occurrence
counts.  This makes it easy to see which errors have occurred most
frequently.</para>

<para>Errors are reported before the associated operation actually
happens.  If you're using a tool (Memcheck, Addrcheck) which does
address checking, and your program attempts to read from address zero,
the tool will emit a message to this effect, and the program will then
duly die with a segmentation fault.</para>

<para>In general, you should try and fix errors in the order that they
are reported.  Not doing so can be confusing.  For example, a program
which copies uninitialised values to several memory locations, and later
uses them, will generate several error messages, when run on Memcheck.
The first such error message may well give the most direct clue to the
root cause of the problem.</para>

<para>The process of detecting duplicate errors is quite an
expensive one and can become a significant performance overhead
if your program generates huge quantities of errors.  To avoid
serious problems, Valgrind will simply stop collecting
errors after 1000 different errors have been seen, or 100000 errors
in total have been seen.  In this situation you might as well
stop your program and fix it, because Valgrind won't tell you
anything else useful after this.  Note that the 1000/100000 limits
apply after suppressed errors are removed.  These limits are
defined in <filename>m_errormgr.c</filename> and can be increased
if necessary.</para>

<para>To avoid this cutoff you can use the
<option>--error-limit=no</option> flag.  Then Valgrind will always show
errors, regardless of how many there are.  Use this flag carefully,
since it may have a bad effect on performance.</para>

</sect1>


<sect1 id="manual-core.suppress" xreflabel="Suppressing errors">
<title>Suppressing errors</title>

<para>The error-checking tools detect numerous problems in the base
libraries, such as the GNU C library, and the XFree86 client libraries,
which come pre-installed on your GNU/Linux system.  You can't easily fix
these, but you don't want to see these errors (and yes, there are many!)
So Valgrind reads a list of errors to suppress at startup.  A default
suppression file is cooked up by the
<computeroutput>./configure</computeroutput> script when the system is
built.</para>

<para>You can modify and add to the suppressions file at your leisure,
or, better, write your own.  Multiple suppression files are allowed.
This is useful if part of your project contains errors you can't or
don't want to fix, yet you don't want to continuously be reminded of
them.</para>

<formalpara><title>Note:</title> <para>By far the easiest way to add
suppressions is to use the <option>--gen-suppressions=yes</option> flag
described in <xref linkend="manual-core.flags"/>.</para>
</formalpara>

<para>Each error to be suppressed is described very specifically, to
minimise the possibility that a suppression-directive inadvertantly
suppresses a bunch of similar errors which you did want to see.  The
suppression mechanism is designed to allow precise yet flexible
specification of errors to suppress.</para>

<para>If you use the <option>-v</option> flag, at the end of execution,
Valgrind prints out one line for each used suppression, giving its name
and the number of times it got used.  Here's the suppressions used by a
run of <computeroutput>valgrind --tool=memcheck ls l</computeroutput>:</para>

<programlisting><![CDATA[
--27579-- supp: 1 socketcall.connect(serv_addr)/__libc_connect/__nscd_getgrgid_r
--27579-- supp: 1 socketcall.connect(serv_addr)/__libc_connect/__nscd_getpwuid_r
--27579-- supp: 6 strrchr/_dl_map_object_from_fd/_dl_map_object]]></programlisting>

<para>Multiple suppressions files are allowed.  By default, Valgrind
uses <filename>$PREFIX/lib/valgrind/default.supp</filename>.  You can
ask to add suppressions from another file, by specifying
<option>--suppressions=/path/to/file.supp</option>.
</para>

<para>If you want to understand more about suppressions, look at an
existing suppressions file whilst reading the following documentation.
The file <filename>glibc-2.2.supp</filename>, in the source
distribution, provides some good examples.</para>

<para>Each suppression has the following components:</para>

<itemizedlist>

  <listitem>
    <para>First line: its name.  This merely gives a handy name to the
    suppression, by which it is referred to in the summary of used
    suppressions printed out when a program finishes.  It's not
    important what the name is; any identifying string will do.</para>
  </listitem>

  <listitem>
    <para>Second line: name of the tool(s) that the suppression is for
    (if more than one, comma-separated), and the name of the suppression
    itself, separated by a colon (Nb: no spaces are allowed), eg:</para>
<programlisting><![CDATA[
tool_name1,tool_name2:suppression_name]]></programlisting>

    <para>Recall that Valgrind is a modular system, in which
    different instrumentation tools can observe your program whilst it
    is running.  Since different tools detect different kinds of errors,
    it is necessary to say which tool(s) the suppression is meaningful
    to.</para>

    <para>Tools will complain, at startup, if a tool does not understand
    any suppression directed to it.  Tools ignore suppressions which are
    not directed to them.  As a result, it is quite practical to put
    suppressions for all tools into the same suppression file.</para>

    <para>Valgrind's core can detect certain PThreads API errors, for
    which this line reads:</para>

<programlisting><![CDATA[
core:PThread]]></programlisting>
  </listitem>

  <listitem>
    <para>Next line: a small number of suppression types have extra
    information after the second line (eg. the <varname>Param</varname>
    suppression for Memcheck)</para>
  </listitem>

  <listitem>
    <para>Remaining lines: This is the calling context for the error --
    the chain of function calls that led to it.  There can be up to four
    of these lines.</para>

    <para>Locations may be either names of shared objects/executables or
    wildcards matching function names.  They begin
    <computeroutput>obj:</computeroutput> and
    <computeroutput>fun:</computeroutput> respectively.  Function and
    object names to match against may use the wildcard characters
    <computeroutput>*</computeroutput> and
    <computeroutput>?</computeroutput>.</para>

    <para><command>Important note: </command> C++ function names must be
    <command>mangled</command>.  If you are writing suppressions by
    hand, use the <option>--demangle=no</option> option to get the
    mangled names in your error messages.</para>
  </listitem>

  <listitem>
    <para>Finally, the entire suppression must be between curly
    braces. Each brace must be the first character on its own
    line.</para>
  </listitem>

 </itemizedlist>

<para>A suppression only suppresses an error when the error matches all
the details in the suppression.  Here's an example:</para>

<programlisting><![CDATA[
{
  __gconv_transform_ascii_internal/__mbrtowc/mbtowc
  Memcheck:Value4
  fun:__gconv_transform_ascii_internal
  fun:__mbr*toc
  fun:mbtowc
}]]></programlisting>


<para>What it means is: for Memcheck only, suppress a
use-of-uninitialised-value error, when the data size is 4, when it
occurs in the function
<computeroutput>__gconv_transform_ascii_internal</computeroutput>, when
that is called from any function of name matching
<computeroutput>__mbr*toc</computeroutput>, when that is called from
<computeroutput>mbtowc</computeroutput>.  It doesn't apply under any
other circumstances.  The string by which this suppression is identified
to the user is
<computeroutput>__gconv_transform_ascii_internal/__mbrtowc/mbtowc</computeroutput>.</para>

<para>(See <xref linkend="mc-manual.suppfiles"/> for more details
on the specifics of Memcheck's suppression kinds.)</para>

<para>Another example, again for the Memcheck tool:</para>

<programlisting><![CDATA[
{
  libX11.so.6.2/libX11.so.6.2/libXaw.so.7.0
  Memcheck:Value4
  obj:/usr/X11R6/lib/libX11.so.6.2
  obj:/usr/X11R6/lib/libX11.so.6.2
  obj:/usr/X11R6/lib/libXaw.so.7.0
}]]></programlisting>

<para>Suppress any size 4 uninitialised-value error which occurs
anywhere in <filename>libX11.so.6.2</filename>, when called from
anywhere in the same library, when called from anywhere in
<filename>libXaw.so.7.0</filename>.  The inexact specification of
locations is regrettable, but is about all you can hope for, given that
the X11 libraries shipped with Red Hat 7.2 have had their symbol tables
removed.</para>

<para>Note: since the above two examples did not make it clear, you can
freely mix the <computeroutput>obj:</computeroutput> and
<computeroutput>fun:</computeroutput> styles of description within a
single suppression record.</para>

</sect1>


<sect1 id="manual-core.flags" 
       xreflabel="Command-line flags for the Valgrind core">
<title>Command-line flags for the Valgrind core</title>

<para>As mentioned above, Valgrind's core accepts a common set of flags.
The tools also accept tool-specific flags, which are documented
seperately for each tool.</para>

<para>You invoke Valgrind like this:</para>

<programlisting><![CDATA[
valgrind [valgrind-options] your-prog [your-prog-options]]]></programlisting>

<para>Valgrind's default settings succeed in giving reasonable behaviour
in most cases.  We group the available options by rough
categories.</para>

<sect2 id="manual-core.toolopts" xreflabel="Tool-selection option">
<title>Tool-selection option</title>

<para>The single most important option.</para>

<itemizedlist>
  <listitem id="tool_name">
    <para><option>--tool=&lt;name&gt;</option> [default=memcheck]</para>
    <para>Run the Valgrind tool called <emphasis>name</emphasis>,
    e.g. Memcheck, Addrcheck, Cachegrind, etc.</para>
  </listitem>
</itemizedlist>

</sect2>


<sect2 id="manual-core.basicopts" xreflabel="Basic Options">
<title>Basic Options</title>

<!-- start of xi:include in the manpage -->
<para id="basic.opts.para">These options work with all tools.</para>

<variablelist id="basic.opts.list">

  <varlistentry id="opt.help" xreflabel="--help">
    <term><option>-h --help</option></term>
    <listitem>
      <para>Show help for all options, both for the core and for the
      selected tool.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.help-debug" xreflabel="--help-debug">
    <term><option>--help-debug</option></term>
    <listitem>
      <para>Same as <option>--help</option>, but also lists debugging
      options which usually are only of use to Valgrind's
      developers.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.version" xreflabel="--version">
    <term><option>--version</option></term>
    <listitem>
      <para>Show the version number of the Valgrind core. Tools can have
      their own version numbers. There is a scheme in place to ensure
      that tools only execute when the core version is one they are
      known to work with. This was done to minimise the chances of
      strange problems arising from tool-vs-core version
      incompatibilities.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.quiet" xreflabel="--quiet">
    <term><option>-q --quiet</option></term>
    <listitem>
      <para>Run silently, and only print error messages. Useful if you
      are running regression tests or have some other automated test
      machinery.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.verbose" xreflabel="--verbose">
    <term><option>-v --verbose</option></term>
    <listitem>
      <para>Be more verbose. Gives extra information on various aspects
      of your program, such as: the shared objects loaded, the
      suppressions used, the progress of the instrumentation and
      execution engines, and warnings about unusual behaviour. Repeating
      the flag increases the verbosity level.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.d" xreflabel="-d">
    <term><option>-d</option></term>
    <listitem>
      <para>Emit information for debugging Valgrind itself.  This is
      usually only of interest to the Valgrind developers.  Repeating
      the flag produces more detailed output.  If you want to send us a
      bug report, a log of the output generated by 
      <option>-v -v -d -d</option> will make your report more
      useful.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.tool" xreflabel="--tool">
    <term>
      <option><![CDATA[--tool=<toolname> [default: memcheck] ]]></option>
    </term>
    <listitem>
      <para>Run the Valgrind tool called <varname>toolname</varname>,
      e.g. Memcheck, Addrcheck, Cachegrind, etc.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.trace-children" xreflabel="--trace-children">
    <term>
      <option><![CDATA[--trace-children=<yes|no> [default: no] ]]></option>
    </term>
    <listitem>
      <para>When enabled, Valgrind will trace into child processes.
      This can be confusing and isn't usually what you want, so it is
      disabled by default.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.track-fds" xreflabel="--track-fds">
    <term>
      <option><![CDATA[--track-fds=<yes|no> [default: no] ]]></option>
    </term>
    <listitem>
      <para>When enabled, Valgrind will print out a list of open file
      descriptors on exit.  Along with each file descriptor is printed a
      stack backtrace of where the file was opened and any details
      relating to the file descriptor such as the file name or socket
      details.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.time-stamp" xreflabel="--time-stamp">
    <term>
      <option><![CDATA[--time-stamp=<yes|no> [default: no] ]]></option>
    </term>
    <listitem>
      <para>When enabled, each message is preceded with an indication of
      the elapsed wallclock time since startup, expressed as days,
      hours, minutes, seconds and milliseconds.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.log-fd" xreflabel="--log-fd">
    <term>
      <option><![CDATA[--log-fd=<number> [default: 2, stderr] ]]></option>
    </term>
    <listitem>
      <para>Specifies that Valgrind should send all of its messages to
      the specified file descriptor.  The default, 2, is the standard
      error channel (stderr).  Note that this may interfere with the
      client's own use of stderr, as Valgrind's output will be
      interleaved with any output that the client sends to
      stderr.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.log-file" xreflabel="--log-file">
    <term>
      <option><![CDATA[--log-file=<filename> ]]></option>
    </term>
    <listitem>
      <para>Specifies that Valgrind should send all of its messages to
      the specified file. In fact, the file name used is created by
      concatenating the text <filename>filename</filename>, "." and the
      process ID, (ie. <![CDATA[<filename>.<pid>]]>), so as to create a
      file per process.  The specified file name may not be the empty
      string.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.log-file-exactly" xreflabel="--log-file-exactly">
    <term>
      <option><![CDATA[--log-file-exactly=<filename> ]]></option>
    </term>
    <listitem>
      <para>Just like <option>--log-file</option>, but the suffix
      <computeroutput>".pid"</computeroutput> is not added.  If you
      trace multiple processes with Valgrind when using this option the
      log file may get all messed up.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.log-file-qualifier" xreflabel="--log-file-qualifier">
    <term>
      <option><![CDATA[--log-file-qualifier=<VAR> ]]></option>
    </term>
    <listitem>
      <para>When used in conjunction with <option>--log-file</option>,
      causes the log file name to be qualified using the contents of the
      environment variable <computeroutput>$VAR</computeroutput>.  This
      is useful when running MPI programs.  For further details, see
      <link linkend="manual-core.comment">Section 2.3 "The Commentary"</link>
      in the manual.
      </para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.log-socket" xreflabel="--log-socket">
    <term>
      <option><![CDATA[--log-socket=<ip-address:port-number> ]]></option>
    </term>
    <listitem>
      <para>Specifies that Valgrind should send all of its messages to
      the specified port at the specified IP address.  The port may be
      omitted, in which case port 1500 is used.  If a connection cannot
      be made to the specified socket, Valgrind falls back to writing
      output to the standard error (stderr).  This option is intended to
      be used in conjunction with the
      <computeroutput>valgrind-listener</computeroutput> program.  For
      further details, see 
      <link linkend="manual-core.comment">Section 2.3 "The Commentary"</link>
      in the manual.</para>
    </listitem>
  </varlistentry>

</variablelist>
<!-- end of xi:include in the manpage -->

</sect2>


<sect2 id="manual-core.erropts" xreflabel="Error-related Options">
<title>Error-related options</title>

<!-- start of xi:include in the manpage -->
<para id="error-related.opts.para">These options are used by all tools
that can report errors, e.g. Memcheck, but not Cachegrind.</para>

<variablelist id="error-related.opts.list">

  <varlistentry id="opt.xml" xreflabel="--xml">
    <term>
      <option><![CDATA[--xml=<yes|no> [default: no] ]]></option>
    </term>
    <listitem>
      <para>When enabled, output will be in XML format.  This is aimed
      at making life easier for tools that consume Valgrind's output as
      input, such as GUI front ends.  Currently this option only works
      with Memcheck and Nulgrind.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.xml-user-comment" xreflabel="--xml-user-comment">
    <term>
      <option><![CDATA[--xml-user-comment=<string> ]]></option>
    </term>
    <listitem>
      <para>Embeds an extra user comment string at the start of the XML
      output.  Only works when <option>--xml=yes</option> is specified;
      ignored otherwise.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.demangle" xreflabel="--demangle">
    <term>
      <option><![CDATA[--demangle=<yes|no> [default: yes] ]]></option>
    </term>
    <listitem>
      <para>Enable/disable automatic demangling (decoding) of C++ names.
      Enabled by default.  When enabled, Valgrind will attempt to
      translate encoded C++ names back to something approaching the
      original.  The demangler handles symbols mangled by g++ versions
      2.X, 3.X and 4.X.</para>

      <para>An important fact about demangling is that function names
      mentioned in suppressions files should be in their mangled form.
      Valgrind does not demangle function names when searching for
      applicable suppressions, because to do otherwise would make
      suppressions file contents dependent on the state of Valgrind's
      demangling machinery, and would also be slow and pointless.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.num-callers" xreflabel="--num-callers">
    <term>
      <option><![CDATA[--num-callers=<number> [default: 12] ]]></option>
    </term>
    <listitem>
      <para>By default, Valgrind shows twelve levels of function call
      names to help you identify program locations.  You can change that
      number with this option.  This can help in determining the
      program's location in deeply-nested call chains.  Note that errors
      are commoned up using only the top four function locations (the
      place in the current function, and that of its three immediate
      callers).  So this doesn't affect the total number of errors
      reported.</para>

      <para>The maximum value for this is 50. Note that higher settings
      will make Valgrind run a bit more slowly and take a bit more
      memory, but can be useful when working with programs with
      deeply-nested call chains.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.error-limit" xreflabel="--error-limit">
    <term>
      <option><![CDATA[--error-limit=<yes|no> [default: yes] ]]></option>
    </term>
    <listitem>
      <para>When enabled, Valgrind stops reporting errors after 100,000
      in total, or 1,000 different ones, have been seen.  This is to
      stop the error tracking machinery from becoming a huge performance
      overhead in programs with many errors.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.stack-traces" xreflabel="--show-below-main">
    <term>
      <option><![CDATA[--show-below-main=<yes|no> [default: no] ]]></option>
    </term>
    <listitem>
      <para>By default, stack traces for errors do not show any
      functions that appear beneath <function>main()</function> 
      (or similar functions such as glibc's
      <function>__libc_start_main()</function>, if
      <function>main()</function> is not present in the stack trace);
      most of the time it's uninteresting C library stuff.  If this
      option is enabled, those entries below <function>main()</function>
      will be shown.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.suppressions" xreflabel="--suppressions">
    <term>
      <option><![CDATA[--suppressions=<filename> [default: $PREFIX/lib/valgrind/default.supp] ]]></option>
    </term>
    <listitem>
      <para>Specifies an extra file from which to read descriptions of
      errors to suppress.  You may use as many extra suppressions files
      as you like.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.gen-suppressions" xreflabel="--gen-supressions">
    <term>
      <option><![CDATA[--gen-suppressions=<yes|no|all> [default: no] ]]></option>
    </term>
    <listitem>
      <para>When set to <varname>yes</varname>, Valgrind will pause
      after every error shown and print the line:
      <literallayout>    ---- Print suppression ? --- [Return/N/n/Y/y/C/c] ----</literallayout>

      The prompt's behaviour is the same as for the
      <option>--db-attach</option> option (see below).</para>

      <para>If you choose to, Valgrind will print out a suppression for
      this error.  You can then cut and paste it into a suppression file
      if you don't want to hear about the error in the future.</para>

      <para>When set to <varname>all</varname>, Valgrind will print a
      suppression for every reported error, without querying the
      user.</para>

      <para>This option is particularly useful with C++ programs, as it
      prints out the suppressions with mangled names, as
      required.</para>

      <para>Note that the suppressions printed are as specific as
      possible.  You may want to common up similar ones, eg. by adding
      wildcards to function names.  Also, sometimes two different errors
      are suppressed by the same suppression, in which case Valgrind
      will output the suppression more than once, but you only need to
      have one copy in your suppression file (but having more than one
      won't cause problems).  Also, the suppression name is given as
      <computeroutput>&lt;insert a suppression name
      here&gt;</computeroutput>; the name doesn't really matter, it's
      only used with the <option>-v</option> option which prints out all
      used suppression records.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.db-attach" xreflabel="--db-attach">
    <term>
      <option><![CDATA[--db-attach=<yes|no> [default: no] ]]></option>
    </term>
    <listitem>
      <para>When enabled, Valgrind will pause after every error shown
      and print the line:
      <literallayout>    ---- Attach to debugger ? --- [Return/N/n/Y/y/C/c] ----</literallayout>

      Pressing <varname>Ret</varname>, or <varname>N Ret</varname> or
      <varname>n Ret</varname>, causes Valgrind not to start a debugger
      for this error.</para>

      <para>Pressing <varname>Y Ret</varname> or
      <varname>y Ret</varname> causes Valgrind to start a debugger for
      the program at this point. When you have finished with the
      debugger, quit from it, and the program will continue. Trying to
      continue from inside the debugger doesn't work.</para>

      <para><varname>C Ret</varname> or <varname>c Ret</varname> causes
      Valgrind not to start a debugger, and not to ask again.</para>

      <para><command>Note:</command> <option>--db-attach=yes</option>
      conflicts with <option>--trace-children=yes</option>.  You can't
      use them together.  Valgrind refuses to start up in this
      situation.</para>

      <para>May 2002: this is a historical relic which could be easily
      fixed if it gets in your way.  Mail us and complain if this is a
      problem for you.</para>
      <para>Nov 2002: if you're sending output to a logfile or to a
      network socket, I guess this option doesn't make any sense.
      Caveat emptor.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.db-command" xreflabel="--db-command">
    <term>
      <option><![CDATA[--db-command=<command> [default: gdb -nw %f %p] ]]></option>
    </term>
    <listitem>
      <para>Specify the debugger to use with the
      <option>--db-attach</option> command. The default debugger is
      gdb. This option is a template that is expanded by Valgrind at
      runtime.  <literal>%f</literal> is replaced with the executable's
      file name and <literal>%p</literal> is replaced by the process ID
      of the executable.</para>

      <para>This specifies how Valgrind will invoke the debugger.  By
      default it will use whatever GDB is detected at build time, which
      is usually <computeroutput>/usr/bin/gdb</computeroutput>.  Using
      this command, you can specify some alternative command to invoke
      the debugger you want to use.</para>

      <para>The command string given can include one or instances of the
      <literal>%p</literal> and <literal>%f</literal> expansions. Each
      instance of <literal>%p</literal> expands to the PID of the
      process to be debugged and each instance of <literal>%f</literal>
      expands to the path to the executable for the process to be
      debugged.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.input-fd" xreflabel="--input-fd">
    <term>
      <option><![CDATA[--input-fd=<number> [default: 0, stdin] ]]></option>
    </term>
    <listitem>
      <para>When using <option>--db-attach=yes</option> and
      <option>--gen-suppressions=yes</option>, Valgrind will stop so as
      to read keyboard input from you, when each error occurs.  By
      default it reads from the standard input (stdin), which is
      problematic for programs which close stdin.  This option allows
      you to specify an alternative file descriptor from which to read
      input.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.max-stackframe" xreflabel="--max-stackframe">
    <term>
      <option><![CDATA[--max-stackframe=<number> [default: 2000000] ]]></option>
    </term>
    <listitem>
      <para>The maximum size of a stack frame - if the stack pointer moves by
      more than this amount then Valgrind will assume that
      the program is switching to a different stack.</para>

      <para>You may need to use this option if your program has large
      stack-allocated arrays.  Valgrind keeps track of your program's
      stack pointer.  If it changes by more than the threshold amount,
      Valgrind assumes your program is switching to a different stack,
      and Memcheck behaves differently than it would for a stack pointer
      change smaller than the threshold.  Usually this heuristic works
      well.  However, if your program allocates large structures on the
      stack, this heuristic will be fooled, and Memcheck will
      subsequently report large numbers of invalid stack accesses.  This
      option allows you to change the threshold to a different
      value.</para>

      <para>You should only consider use of this flag if Valgrind's
      debug output directs you to do so.  In that case it will tell you
      the new threshold you should specify.</para>

      <para>In general, allocating large structures on the stack is a
      bad idea, because (1) you can easily run out of stack space,
      especially on systems with limited memory or which expect to
      support large numbers of threads each with a small stack, and (2)
      because the error checking performed by Memcheck is more effective
      for heap-allocated data than for stack-allocated data.  If you
      have to use this flag, you may wish to consider rewriting your
      code to allocate on the heap rather than on the stack.</para>
    </listitem>
  </varlistentry>

</variablelist>
<!-- end of xi:include in the manpage -->

</sect2>


<sect2 id="manual-core.mallocopts" xreflabel="malloc()-related Options">
<title><computeroutput>malloc()</computeroutput>-related Options</title>

<!-- start of xi:include in the manpage -->
<para id="malloc-related.opts.para">For tools that use their own version of
<computeroutput>malloc()</computeroutput> (e.g. Memcheck and
Addrcheck), the following options apply.</para>

<variablelist id="malloc-related.opts.list">

  <varlistentry id="opt.alignment" xreflabel="--alignment">
    <term>
      <option><![CDATA[--alignment=<number> [default: 8] ]]></option>
    </term>
    <listitem>
      <para>By default Valgrind's <function>malloc()</function>,
      <function>realloc()</function>, etc, return 8-byte aligned
      addresses.  This is standard for most processors.  However, some
      programs might assume that <function>malloc()</function> et al
      return 16-byte or more aligned memory.  The supplied value must be
      between 8 and 4096 inclusive, and must be a power of two.</para>
    </listitem>
  </varlistentry>

</variablelist>
<!-- end of xi:include in the manpage -->

</sect2>


<sect2 id="manual-core.rareopts" xreflabel="Uncommon Options">
<title>Uncommon Options</title>

<!-- start of xi:include in the manpage -->
<para id="uncommon.opts.para">These options apply to all tools, as they
affect certain obscure workings of the Valgrind core.  Most people won't
need to use these.</para>

<variablelist id="uncommon.opts.list">

  <varlistentry id="opt.run-libc-freeres" xreflabel="--run-libc-freeres">
    <term>
      <option><![CDATA[--run-libc-freeres=<yes|no> [default: yes] ]]></option>
    </term>
    <listitem>
      <para>The GNU C library (<function>libc.so</function>), which is
      used by all programs, may allocate memory for its own uses.
      Usually it doesn't bother to free that memory when the program
      ends - there would be no point, since the Linux kernel reclaims
      all process resources when a process exits anyway, so it would
      just slow things down.</para>

      <para>The glibc authors realised that this behaviour causes leak
      checkers, such as Valgrind, to falsely report leaks in glibc, when
      a leak check is done at exit.  In order to avoid this, they
      provided a routine called <function>__libc_freeres</function>
      specifically to make glibc release all memory it has allocated.
      Memcheck and Addrcheck therefore try and run
      <function>__libc_freeres</function> at exit.</para>

      <para>Unfortunately, in some versions of glibc,
      <function>__libc_freeres</function> is sufficiently buggy to cause
      segmentation faults.  This is particularly noticeable on Red Hat
      7.1.  So this flag is provided in order to inhibit the run of
      <function>__libc_freeres</function>.  If your program seems to run
      fine on Valgrind, but segfaults at exit, you may find that
      <option>--run-libc-freeres=no</option> fixes that, although at the
      cost of possibly falsely reporting space leaks in
      <filename>libc.so</filename>.</para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.sim-hints" xreflabel="--sim-hints">
    <term>
      <option><![CDATA[--sim-hints=hint1,hint2,... ]]></option>
    </term>
    <listitem>
      <para>Pass miscellaneous hints to Valgrind which slightly modify
      the simulated behaviour in nonstandard or dangerous ways, possibly
      to help the simulation of strange features.  By default no hints
      are enabled.  Use with caution!  Currently known hints are:</para>
      <itemizedlist>
        <listitem>
          <para><option>lax-ioctls: </option> Be very lax about ioctl
          handling; the only assumption is that the size is
          correct. Doesn't require the full buffer to be initialized
          when writing.  Without this, using some device drivers with a
          large number of strange ioctl commands becomes very
          tiresome.</para>
        </listitem>
        <listitem>
          <para><option>enable-inner: </option> Enable some special
          magic needed when the program being run is itself
          Valgrind.</para>
        </listitem>
      </itemizedlist>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.kernel-variant" xreflabel="--kernel-variant">
    <term>
      <option>--kernel-variant=variant1,variant2,...</option>
    </term>
    <listitem>
      <para>Handle system calls and ioctls arising from minor variants
      of the default kernel for this platform.  This is useful for
      running on hacked kernels or with kernel modules which support
      nonstandard ioctls, for example.  Use with caution.  If you don't
      understand what this option does then you almost certainly don't
      need it.  Currently known variants are:</para>
      <itemizedlist>
        <listitem>
          <para><option>bproc: </option> Support the sys_broc system
          call on x86.  This is for running on BProc, which is a minor
          variant of standard Linux which is sometimes used for building
          clusters.</para>
        </listitem>
      </itemizedlist>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.show-emwarns" xreflabel="--show-emwarns">
    <term>
      <option><![CDATA[--show-emwarns=<yes|no> [default: no] ]]></option>
    </term>
    <listitem>
      <para>When enabled, Valgrind will emit warnings about its CPU
      emulation in certain cases.  These are usually not
      interesting.</para>
   </listitem>
  </varlistentry>

  <varlistentry id="opt.smc-check" xreflabel="--smc-check">
    <term>
      <option><![CDATA[--smc-check=<none|stack|all> [default: stack] ]]></option>
    </term>
    <listitem>
      <para>This option controls Valgrind's detection of self-modifying
      code.  Valgrind can do no detection, detect self-modifying code on
      the stack, or detect self-modifying code anywhere.  Note that the
      default option will catch the vast majority of cases, as far as we
      know.  Running with <varname>all</varname> will slow Valgrind down
      greatly (but running with <varname>none</varname> will rarely
      speed things up, since very little code gets put on the stack for
      most programs).</para>
    </listitem>
  </varlistentry>

</variablelist>
<!-- end of xi:include in the manpage -->

</sect2>


<sect2 id="manual-core.debugopts" xreflabel="Debugging Valgrind Options">
<title>Debugging Valgrind Options</title>

<!-- start of xi:include in the manpage -->
<para id="debug.opts.para">There are also some options for debugging
Valgrind itself.  You shouldn't need to use them in the normal run of
things.  If you wish to see the list, use the
<option>--help-debug</option> option.</para>
<!-- end of xi:include in the manpage -->

</sect2>


<sect2 id="manual-core.defopts" xreflabel="Setting default options">
<title>Setting default Options</title>

<para>Note that Valgrind also reads options from three places:</para>

  <orderedlist>
   <listitem>
    <para>The file <computeroutput>~/.valgrindrc</computeroutput></para>
   </listitem>

   <listitem>
    <para>The environment variable
    <computeroutput>$VALGRIND_OPTS</computeroutput></para>
   </listitem>

   <listitem>
    <para>The file <computeroutput>./.valgrindrc</computeroutput></para>
   </listitem>
  </orderedlist>

<para>These are processed in the given order, before the
command-line options.  Options processed later override those
processed earlier; for example, options in
<computeroutput>./.valgrindrc</computeroutput> will take
precedence over those in
<computeroutput>~/.valgrindrc</computeroutput>.  The first two
are particularly useful for setting the default tool to
use.</para>

<para>Any tool-specific options put in
<computeroutput>$VALGRIND_OPTS</computeroutput> or the
<computeroutput>.valgrindrc</computeroutput> files should be
prefixed with the tool name and a colon.  For example, if you
want Memcheck to always do leak checking, you can put the
following entry in <literal>~/.valgrindrc</literal>:</para>

<programlisting><![CDATA[
--memcheck:leak-check=yes]]></programlisting>

<para>This will be ignored if any tool other than Memcheck is
run.  Without the <computeroutput>memcheck:</computeroutput>
part, this will cause problems if you select other tools that
don't understand
<computeroutput>--leak-check=yes</computeroutput>.</para>

</sect2>

</sect1>


<sect1 id="manual-core.clientreq" 
       xreflabel="The Client Request mechanism">
<title>The Client Request mechanism</title>

<para>Valgrind has a trapdoor mechanism via which the client
program can pass all manner of requests and queries to Valgrind
and the current tool.  Internally, this is used extensively to
make malloc, free, etc, work, although you don't see that.</para>

<para>For your convenience, a subset of these so-called client
requests is provided to allow you to tell Valgrind facts about
the behaviour of your program, and also to make queries.
In particular, your program can tell Valgrind about changes in
memory range permissions that Valgrind would not otherwise know
about, and so allows clients to get Valgrind to do arbitrary
custom checks.</para>

<para>Clients need to include a header file to make this work.
Which header file depends on which client requests you use.  Some
client requests are handled by the core, and are defined in the
header file <filename>valgrind/valgrind.h</filename>.  Tool-specific
header files are named after the tool, e.g.
<filename>valgrind/memcheck.h</filename>.  All header files can be found
in the <literal>include/valgrind</literal> directory of wherever Valgrind
was installed.</para>

<para>The macros in these header files have the magical property
that they generate code in-line which Valgrind can spot.
However, the code does nothing when not run on Valgrind, so you
are not forced to run your program under Valgrind just because you
use the macros in this file.  Also, you are not required to link your
program with any extra supporting libraries.</para>

<para>The code left in your binary has negligible performance impact:
on x86, amd64 and ppc32, the overhead is 6 simple integer instructions
and is probably undetectable except in tight loops.
However, if you really wish to compile out the client requests, you can
compile with <computeroutput>-DNVALGRIND</computeroutput> (analogous to
<computeroutput>-DNDEBUG</computeroutput>'s effect on
<computeroutput>assert()</computeroutput>).
</para>

<para>You are encouraged to copy the <filename>valgrind/*.h</filename> headers
into your project's include directory, so your program doesn't have a
compile-time dependency on Valgrind being installed.  The Valgrind headers,
unlike the rest of the code, are under a BSD-style license so you may include
them without worrying about license incompatibility.</para>

<para>Here is a brief description of the macros available in
<filename>valgrind.h</filename>, which work with more than one
tool (see the tool-specific documentation for explanations of the
tool-specific macros).</para>

 <variablelist>

  <varlistentry>
   <term><command><computeroutput>RUNNING_ON_VALGRIND</computeroutput></command>:</term>
   <listitem>
    <para>returns 1 if running on Valgrind, 0 if running on the
    real CPU.  If you are running Valgrind on itself, it will return the
    number of layers of Valgrind emulation we're running on.
    </para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_DISCARD_TRANSLATIONS</computeroutput>:</command></term>
   <listitem>
    <para>discard translations of code in the specified address
    range.  Useful if you are debugging a JITter or some other
    dynamic code generation system.  After this call, attempts to
    execute code in the invalidated address range will cause
    Valgrind to make new translations of that code, which is
    probably the semantics you want.  Note that code invalidations
    are expensive because finding all the relevant translations
    quickly is very difficult.  So try not to call it often.
    Note that you can be clever about
    this: you only need to call it when an area which previously
    contained code is overwritten with new code.  You can choose
    to write code into fresh memory, and just call this
    occasionally to discard large chunks of old code all at
    once.</para>
    <para>
    Alternatively, for transparent self-modifying-code support,
    use<computeroutput>--smc-check=all</computeroutput>.
    </para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_COUNT_ERRORS</computeroutput>:</command></term>
   <listitem>
    <para>returns the number of errors found so far by Valgrind.  Can be
    useful in test harness code when combined with the
    <option>--log-fd=-1</option> option; this runs Valgrind silently,
    but the client program can detect when errors occur.  Only useful
    for tools that report errors, e.g. it's useful for Memcheck, but for
    Cachegrind it will always return zero because Cachegrind doesn't
    report errors.</para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_MALLOCLIKE_BLOCK</computeroutput>:</command></term>
   <listitem>
    <para>If your program manages its own memory instead of using
    the standard <computeroutput>malloc()</computeroutput> /
    <computeroutput>new</computeroutput> /
    <computeroutput>new[]</computeroutput>, tools that track
    information about heap blocks will not do nearly as good a
    job.  For example, Memcheck won't detect nearly as many
    errors, and the error messages won't be as informative.  To
    improve this situation, use this macro just after your custom
    allocator allocates some new memory.  See the comments in
    <filename>valgrind.h</filename> for information on how to use
    it.</para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_FREELIKE_BLOCK</computeroutput>:</command></term>
   <listitem>
    <para>This should be used in conjunction with
    <computeroutput>VALGRIND_MALLOCLIKE_BLOCK</computeroutput>.
    Again, see <filename>memcheck/memcheck.h</filename> for
    information on how to use it.</para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_CREATE_MEMPOOL</computeroutput>:</command></term>
   <listitem>
    <para>This is similar to
    <computeroutput>VALGRIND_MALLOCLIKE_BLOCK</computeroutput>,
    but is tailored towards code that uses memory pools.  See the
    comments in <filename>valgrind.h</filename> for information
    on how to use it.</para>
   </listitem>
  </varlistentry>
  
  <varlistentry>
  <term><command><computeroutput>VALGRIND_DESTROY_MEMPOOL</computeroutput>:</command></term>
   <listitem>
    <para>This should be used in conjunction with
    <computeroutput>VALGRIND_CREATE_MEMPOOL</computeroutput>
    Again, see the comments in <filename>valgrind.h</filename> for
    information on how to use it.</para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_MEMPOOL_ALLOC</computeroutput>:</command></term>
   <listitem>
    <para>This should be used in conjunction with
    <computeroutput>VALGRIND_CREATE_MEMPOOL</computeroutput>
    Again, see the comments in <filename>valgrind.h</filename> for
    information on how to use it.</para>
   </listitem>
  </varlistentry>
   
  <varlistentry>
   <term><command><computeroutput>VALGRIND_MEMPOOL_FREE</computeroutput>:</command></term>
   <listitem>
    <para>This should be used in conjunction with
    <computeroutput>VALGRIND_CREATE_MEMPOOL</computeroutput>
    Again, see the comments in <filename>valgrind.h</filename> for
    information on how to use it.</para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_NON_SIMD_CALL[0123]</computeroutput>:</command></term>
   <listitem>
    <para>executes a function of 0, 1, 2 or 3 args in the client
    program on the <emphasis>real</emphasis> CPU, not the virtual
    CPU that Valgrind normally runs code on.  These are used in
    various ways internally to Valgrind.  They might be useful to
    client programs.</para> 

    <para><command>Warning:</command> Only use these if you
    <emphasis>really</emphasis> know what you are doing.</para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_PRINTF(format, ...)</computeroutput>:</command></term>
   <listitem>
    <para>printf a message to the log file when running under
    Valgrind.  Nothing is output if not running under Valgrind.
    Returns the number of characters output.</para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_PRINTF_BACKTRACE(format, ...)</computeroutput>:</command></term>
   <listitem>
    <para>printf a message to the log file along with a stack
    backtrace when running under Valgrind.  Nothing is output if
    not running under Valgrind.  Returns the number of characters
    output.</para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_STACK_REGISTER(start, end)</computeroutput>:</command></term>
   <listitem>
    <para>Register a new stack.  Informs Valgrind that the memory range
    between start and end is a unique stack.  Returns a stack identifier
    that can be used with other
    <computeroutput>VALGRIND_STACK_*</computeroutput> calls.</para>
    <para>Valgrind will use this information to determine if a change to
    the stack pointer is an item pushed onto the stack or a change over
    to a new stack.  Use this if you're using a user-level thread package
    and are noticing spurious errors from Valgrind about uninitialized
    memory reads.</para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_STACK_DEREGISTER(id)</computeroutput>:</command></term>
   <listitem>
    <para>Deregister a previously registered stack.  Informs
    Valgrind that previously registered memory range with stack id
    <computeroutput>id</computeroutput> is no longer a stack.</para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term><command><computeroutput>VALGRIND_STACK_CHANGE(id, start, end)</computeroutput>:</command></term>
   <listitem>
    <para>Change a previously registered stack.  Informs
    Valgrind that the previously registerer stack with stack id
    <computeroutput>id</computeroutput> has changed it's start and end
    values.  Use this if your user-level thread package implements
    stack growth.</para>
   </listitem>
  </varlistentry>

 </variablelist>

<para>Note that <filename>valgrind.h</filename> is included by
all the tool-specific header files (such as
<filename>memcheck.h</filename>), so you don't need to include it
in your client if you include a tool-specific header.</para>

</sect1>


<sect1 id="manual-core.pthreads" xreflabel="Support for Threads">
<title>Support for Threads</title>

<para>Valgrind supports programs which use POSIX pthreads.
Getting this to work was technically challenging but it all works
well enough for significant threaded applications to work.</para>

<para>The main thing to point out is that although Valgrind works
with the built-in threads system (eg. NPTL or LinuxThreads), it
serialises execution so that only one thread is running at a time.  This
approach avoids the horrible implementation problems of implementing a
truly multiprocessor version of Valgrind, but it does mean that threaded
apps run only on one CPU, even if you have a multiprocessor
machine.</para>

<para>Valgrind schedules your program's threads in a round-robin fashion,
with all threads having equal priority.  It switches threads
every 50000 basic blocks (on x86, typically around 300000
instructions), which means you'll get a much finer interleaving
of thread executions than when run natively.  This in itself may
cause your program to behave differently if you have some kind of
concurrency, critical race, locking, or similar, bugs.</para>

<para>Your program will use the native
<computeroutput>libpthread</computeroutput>, but not all of its facilities
will work.  In particular, synchonisation of processes via shared-memory
segments will not work.  This relies on special atomic instruction sequences 
which Valgrind does not emulate in a way which works between processes.
Unfortunately there's no way for Valgrind to warn when this is happening,
and such calls will mostly work; it's only when there's a race that
it will fail.
</para>

<para>Valgrind also supports direct use of the
<computeroutput>clone()</computeroutput> system call,
<computeroutput>futex()</computeroutput> and so on.
<computeroutput>clone()</computeroutput> is supported where either
everything is shared (a thread) or nothing is shared (fork-like); partial
sharing will fail.  Again, any use of atomic instruction sequences in shared
memory between processes will not work reliably.
</para>


</sect1>

<sect1 id="manual-core.signals" xreflabel="Handling of Signals">
<title>Handling of Signals</title>

<para>Valgrind has a fairly complete signal implementation.  It should be
able to cope with any valid use of signals.</para>
 
<para>If you're using signals in clever ways (for example, catching
SIGSEGV, modifying page state and restarting the instruction), you're
probably relying on precise exceptions.  In this case, you will need
to use <computeroutput>--vex-iropt-precise-memory-exns=yes</computeroutput>.
</para>

<para>If your program dies as a result of a fatal core-dumping signal,
Valgrind will generate its own core file
(<computeroutput>vgcore.NNNNN</computeroutput>) containing your program's
state.  You may use this core file for post-mortem debugging with gdb or
similar.  (Note: it will not generate a core if your core dump size limit is
0.)  At the time of writing the core dumps do not include all the floating
point register information.</para>

<para>If Valgrind itself crashes (hopefully not) the operating system
will create a core dump in the usual way.</para>

</sect1>


<sect1 id="manual-core.wrapping" xreflabel="Function Wrapping">
<title>Function wrapping</title>

<para>
Valgrind versions 3.2.0 and above and can do function wrapping on all
supported targets.  In function wrapping, calls to some specified
function are intercepted and rerouted to a different, user-supplied
function.  This can do whatever it likes, typically examining the
arguments, calling onwards to the original, and possibly examining the
result.  Any number of different functions may be wrapped.</para>

<para>
Function wrapping is useful for instrumenting an API in some way.  For
example, wrapping functions in the POSIX pthreads API makes it
possible to notify Valgrind of thread status changes, and wrapping
functions in the MPI (message-passing) API allows notifying Valgrind
of memory status changes associated with message arrival/departure.
Such information is usually passed to Valgrind by using client
requests in the wrapper functions, although that is not of relevance
here.</para>

<sect2 id="manual-core.wrapping.example" xreflabel="A Simple Example">
<title>A Simple Example</title>

<para>Supposing we want to wrap some function</para>

<programlisting><![CDATA[
int foo ( int x, int y ) { return x + y; }]]></programlisting>

<para>A wrapper is a function of identical type, but with a special name
which identifies it as the wrapper for <computeroutput>foo</computeroutput>.
Wrappers need to include
supporting macros from <computeroutput>valgrind.h</computeroutput>.
Here is a simple wrapper which prints the arguments and return value:</para>

<programlisting><![CDATA[
#include <stdio.h>
#include "valgrind.h"
int I_WRAP_SONAME_FNNAME_ZU(NONE,foo)( int x, int y )
{
   int    result;
   OrigFn fn;
   VALGRIND_GET_ORIG_FN(fn);
   printf("foo's wrapper: args %d %d\n", x, y);
   CALL_FN_W_WW(result, fn, x,y);
   printf("foo's wrapper: result %d\n", result);
   return result;
}
]]></programlisting>

<para>To become active, the wrapper merely needs to be present in a text
section somewhere in the same process' address space as the function
it wraps, and for its ELF symbol name to be visible to Valgrind.  In
practice, this means either compiling to a 
<computeroutput>.o</computeroutput> and linking it in, or
compiling to a <computeroutput>.so</computeroutput> and 
<computeroutput>LD_PRELOAD</computeroutput>ing it in.  The latter is more
convenient in that it doesn't require relinking.</para>

<para>All wrappers have approximately the above form.  There are three
crucial macros:</para>

<para><computeroutput>I_WRAP_SONAME_FNNAME_ZU</computeroutput>: 
this generates the real name of the wrapper.
This is an encoded name which Valgrind notices when reading symbol
table information.  What it says is: I am the wrapper for any function
named <computeroutput>foo</computeroutput> which is found in 
an ELF shared object with an empty
("<computeroutput>NONE</computeroutput>") soname field.  The specification 
mechanism is powerful in
that wildcards are allowed for both sonames and function names.  
The fine details are discussed below.</para>

<para><computeroutput>VALGRIND_GET_ORIG_FN</computeroutput>: 
once in the the wrapper, the first priority is
to get hold of the address of the original (and any other supporting
information needed).  This is stored in a value of opaque 
type <computeroutput>OrigFn</computeroutput>.
The information is acquired using 
<computeroutput>VALGRIND_GET_ORIG_FN</computeroutput>.  It is crucial
to make this macro call before calling any other wrapped function
in the same thread.</para>

<para><computeroutput>CALL_FN_W_WW</computeroutput>: eventually we will
want to call the function being
wrapped.  Calling it directly does not work, since that just gets us
back to the wrapper and tends to kill the program in short order by
stack overflow.  Instead, the result lvalue, 
<computeroutput>OrigFn</computeroutput> and arguments are
handed to one of a family of macros of the form 
<computeroutput>CALL_FN_*</computeroutput>.  These
cause Valgrind to call the original and avoid recursion back to the
wrapper.</para>
</sect2>

<sect2 id="manual-core.wrapping.specs" xreflabel="Wrapping Specifications">
<title>Wrapping Specifications</title>

<para>This scheme has the advantage of being self-contained.  A library of
wrappers can be compiled to object code in the normal way, and does
not rely on an external script telling Valgrind which wrappers pertain
to which originals.</para>

<para>Each wrapper has a name which, in the most general case says: I am the
wrapper for any function whose name matches FNPATT and whose ELF
"soname" matches SOPATT.  Both FNPATT and SOPATT may contain wildcards
(asterisks) and other characters (spaces, dots, @, etc) which are not 
generally regarded as valid C identifier names.</para> 

<para>This flexibility is needed to write robust wrappers for POSIX pthread
functions, where typically we are not completely sure of either the
function name or the soname, or alternatively we want to wrap a whole
bunch of functions at once.</para> 

<para>For example, <computeroutput>pthread_create</computeroutput> 
in GNU libpthread is usually a
versioned symbol - one whose name ends in, eg, 
<computeroutput>@GLIBC_2.3</computeroutput>.  Hence we
are not sure what its real name is.  We also want to cover any soname
of the form <computeroutput>libpthread.so*</computeroutput>.
So the header of the wrapper will be</para>

<programlisting><![CDATA[
int I_WRAP_SONAME_FNNAME_ZZ(libpthreadZdsoZd0,pthreadZucreateZAZa)
  ( ... formals ... )
  { ... body ... }
]]></programlisting>

<para>In order to write unusual characters as valid C function names, a
Z-encoding scheme is used.  Names are written literally, except that
a capital Z acts as an escape character, with the following encoding:</para>

<programlisting><![CDATA[
     Za   encodes    *
     Zp              +
     Zc              :
     Zd              .
     Zu              _
     Zh              -
     Zs              (space)
     ZA              @
     ZZ              Z
]]></programlisting>

<para>Hence <computeroutput>libpthreadZdsoZd0</computeroutput> is an 
encoding of the soname <computeroutput>libpthread.so.0</computeroutput>
and <computeroutput>pthreadZucreateZAZa</computeroutput> is an encoding 
of the function name <computeroutput>pthread_create@*</computeroutput>.
</para>

<para>The macro <computeroutput>I_WRAP_SONAME_FNNAME_ZZ</computeroutput> 
constructs a wrapper name in which
both the soname (first component) and function name (second component)
are Z-encoded.  Encoding the function name can be tiresome and is
often unnecessary, so a second macro,
<computeroutput>I_WRAP_SONAME_FNNAME_ZU</computeroutput>, can be
used instead.  The <computeroutput>_ZU</computeroutput> variant is 
also useful for writing wrappers for
C++ functions, in which the function name is usually already mangled
using some other convention in which Z plays an important role; having
to encode a second time quickly becomes confusing.</para>

<para>Since the function name field may contain wildcards, it can be
anything, including just <computeroutput>*</computeroutput>.
The same is true for the soname.
However, some ELF objects - specifically, main executables - do not
have sonames.  Any object lacking a soname is treated as if its soname
was <computeroutput>NONE</computeroutput>, which is why the original 
example above had a name
<computeroutput>I_WRAP_SONAME_FNNAME_ZU(NONE,foo)</computeroutput>.</para>
</sect2>

<sect2 id="manual-core.wrapping.semantics" xreflabel="Wrapping Semantics">
<title>Wrapping Semantics</title>

<para>The ability for a wrapper to replace an infinite family of functions
is powerful but brings complications in situations where ELF objects
appear and disappear (are dlopen'd and dlclose'd) on the fly.
Valgrind tries to maintain sensible behaviour in such situations.</para>

<para>For example, suppose a process has dlopened (an ELF object with
soname) <computeroutput>object1.so</computeroutput>, which contains 
<computeroutput>function1</computeroutput>.  It starts to use
<computeroutput>function1</computeroutput> immediately.</para>

<para>After a while it dlopens <computeroutput>wrappers.so</computeroutput>,
which contains a wrapper
for <computeroutput>function1</computeroutput> in (soname) 
<computeroutput>object1.so</computeroutput>.  All subsequent calls to 
<computeroutput>function1</computeroutput> are rerouted to the wrapper.</para>

<para>If <computeroutput>wrappers.so</computeroutput> is 
later dlclose'd, calls to <computeroutput>function1</computeroutput> are 
naturally routed back to the original.</para>

<para>Alternatively, if <computeroutput>object1.so</computeroutput>
is dlclose'd but wrappers.so remains,
then the wrapper exported by <computeroutput>wrapper.so</computeroutput>
becomes inactive, since there
is no way to get to it - there is no original to call any more.  However,
Valgrind remembers that the wrapper is still present.  If 
<computeroutput>object1.so</computeroutput> is
eventually dlopen'd again, the wrapper will become active again.</para>

<para>In short, valgrind inspects all code loading/unloading events to
ensure that the set of currently active wrappers remains consistent.</para>

<para>A second possible problem is that of conflicting wrappers.  It is 
easily possible to load two or more wrappers, both of which claim
to be wrappers for some third function.  In such cases Valgrind will
complain about conflicting wrappers when the second one appears, and
will honour only the first one.</para>
</sect2>

<sect2 id="manual-core.wrapping.debugging" xreflabel="Debugging">
<title>Debugging</title>

<para>Figuring out what's going on given the dynamic nature of wrapping
can be difficult.  The 
<computeroutput>--trace-redir=yes</computeroutput> flag makes 
this possible
by showing the complete state of the redirection subsystem after
every
<computeroutput>mmap</computeroutput>/<computeroutput>munmap</computeroutput>
event affecting code (text).</para>

<para>There are two central concepts:</para>

<itemizedlist>

  <listitem><para>A "redirection specification" is a binding of 
  a (soname pattern, fnname pattern) pair to a code address.
  These bindings are created by writing functions with names
  made with the 
  <computeroutput>I_WRAP_SONAME_FNNAME_{ZZ,_ZU}</computeroutput>
  macros.</para></listitem>

  <listitem><para>An "active redirection" is code-address to 
  code-address binding currently in effect.</para></listitem>

</itemizedlist>

<para>The state of the wrapping-and-redirection subsystem comprises a set of
specifications and a set of active bindings.  The specifications are
acquired/discarded by watching all 
<computeroutput>mmap</computeroutput>/<computeroutput>munmap</computeroutput>
events on code (text)
sections.  The active binding set is (conceptually) recomputed from
the specifications, and all known symbol names, following any change
to the specification set.</para>

<para><computeroutput>--trace-redir=yes</computeroutput> shows the contents 
of both sets following any such event.</para>

<para><computeroutput>-v</computeroutput> prints a line of text each 
time an active specification is used for the first time.</para>

<para>Hence for maximum debugging effectiveness you will need to use both
flags.</para>

<para>One final comment.  The function-wrapping facility is closely
tied to Valgrind's ability to replace (redirect) specified
functions, for example to redirect calls to 
<computeroutput>malloc</computeroutput> to its
own implementation.  Indeed, a replacement function can be
regarded as a wrapper function which does not call the original.
However, to make the implementation more robust, the two kinds
of interception (wrapping vs replacement) are treated differently.
</para>

<para><computeroutput>--trace-redir=yes</computeroutput> shows 
specifications and bindings for both
replacement and wrapper functions.  To differentiate the 
two, replacement bindings are printed using 
<computeroutput>R-></computeroutput> whereas 
wraps are printed using <computeroutput>W-></computeroutput>.
</para>
</sect2>


<sect2 id="manual-core.wrapping.limitations-cf" 
       xreflabel="Limitations - control flow">
<title>Limitations - control flow</title>

<para>For the most part, the function wrapping implementation is robust.
The only important caveat is: in a wrapper, get hold of
the <computeroutput>OrigFn</computeroutput> information using 
<computeroutput>VALGRIND_GET_ORIG_FN</computeroutput> before calling any
other wrapped function.  Once you have the 
<computeroutput>OrigFn</computeroutput>, arbitrary
intercalling, recursion between, and longjumping out of wrappers
should work correctly.  There is never any interaction between wrapped
functions and merely replaced functions 
(eg <computeroutput>malloc</computeroutput>), so you can call
<computeroutput>malloc</computeroutput> etc safely from within wrappers.
</para>

<para>The above comments are true for {x86,amd64,ppc32}-linux.  On
ppc64-linux function wrapping is more fragile due to the (arguably
poorly designed) ppc64-linux ABI.  This mandates the use of a shadow
stack which tracks entries/exits of both wrapper and replacement
functions.  This gives two limitations: firstly, longjumping out of
wrappers will rapidly lead to disaster, since the shadow stack will
not get correctly cleared.  Secondly, since the shadow stack has
finite size, recursion between wrapper/replacement functions is only
possible to a limited depth, beyond which Valgrind has to abort the
run.  This depth is currently 16 calls.</para>

<para>For all platforms ({x86,amd64,ppc32,ppc64}-linux) all the above
comments apply on a per-thread basis.  In other words, wrapping is
thread-safe: each thread must individually observe the above
restrictions, but there is no need for any kind of inter-thread
cooperation.</para>
</sect2>


<sect2 id="manual-core.wrapping.limitations-sigs" 
       xreflabel="Limitations - original function signatures">
<title>Limitations - original function signatures</title>

<para>As shown in the above example, to call the original you must use a
macro of the form <computeroutput>CALL_FN_*</computeroutput>.  
For technical reasons it is impossible
to create a single macro to deal with all argument types and numbers,
so a family of macros covering the most common cases is supplied.  In
what follows, 'W' denotes a machine-word-typed value (a pointer or a
C <computeroutput>long</computeroutput>), 
and 'v' denotes C's <computeroutput>void</computeroutput> type.
The currently available macros are:</para>

<programlisting><![CDATA[
CALL_FN_v_v       -- call an original of type  void fn ( void )
CALL_FN_W_v       -- call an original of type  long fn ( void )

CALL_FN_v_W       -- void fn ( long )
CALL_FN_W_W       -- long fn ( long )

CALL_FN_v_WW      -- void fn ( long, long )
CALL_FN_W_WW      -- long fn ( long, long )

CALL_FN_v_WWW     -- void fn ( long, long, long )
CALL_FN_W_WWW     -- long fn ( long, long, long )

CALL_FN_W_WWWW    -- long fn ( long, long, long, long )
CALL_FN_W_5W      -- long fn ( long, long, long, long, long )
CALL_FN_W_6W      -- long fn ( long, long, long, long, long, long )
and so on, up to 
CALL_FN_W_12W
]]></programlisting>

<para>The set of supported types can be expanded as needed.  It is
regrettable that this limitation exists.  Function wrapping has proven
difficult to implement, with a certain apparently unavoidable level of
ickyness.  After several implementation attempts, the present
arrangement appears to be the least-worst tradeoff.  At least it works
reliably in the presence of dynamic linking and dynamic code
loading/unloading.</para>

<para>You should not attempt to wrap a function of one type signature with a
wrapper of a different type signature.  Such trickery will surely lead
to crashes or strange behaviour.  This is not of course a limitation
of the function wrapping implementation, merely a reflection of the
fact that it gives you sweeping powers to shoot yourself in the foot
if you are not careful.  Imagine the instant havoc you could wreak by
writing a wrapper which matched any function name in any soname - in
effect, one which claimed to be a wrapper for all functions in the
process.</para>
</sect2>

<sect2 id="manual-core.wrapping.examples" xreflabel="Examples">
<title>Examples</title>

<para>In the source tree, 
<computeroutput>memcheck/tests/wrap[1-8].c</computeroutput> provide a series of
examples, ranging from very simple to quite advanced.</para>

<para><computeroutput>auxprogs/libmpiwrap.c</computeroutput> is an example 
of wrapping a big, complex API (the MPI-2 interface).  This file defines 
almost 300 different wrappers.</para>
</sect2>

</sect1>


<sect1 id="manual-core.install" xreflabel="Building and Installing">
<title>Building and Installing</title>

<para>We use the standard Unix
<computeroutput>./configure</computeroutput>,
<computeroutput>make</computeroutput>, <computeroutput>make
install</computeroutput> mechanism, and we have attempted to
ensure that it works on machines with kernel 2.4 or 2.6 and glibc
2.2.X or 2.3.X.  You may then want to run the regression tests
with <computeroutput>make regtest</computeroutput>.
</para>

<para>There are five options (in addition to the usual
<option>--prefix=</option> which affect how Valgrind is built:
<itemizedlist>

  <listitem>
    <para><option>--enable-inner</option></para>
    <para>This builds Valgrind with some special magic hacks which make
     it possible to run it on a standard build of Valgrind (what the
     developers call "self-hosting").  Ordinarily you should not use
     this flag as various kinds of safety checks are disabled.
   </para>
  </listitem>

  <listitem>
    <para><option>--enable-tls</option></para>
    <para>TLS (Thread Local Storage) is a relatively new mechanism which
    requires compiler, linker and kernel support.  Valgrind tries to
    automatically test if TLS is supported and if so enables this option.
    Sometimes it cannot test for TLS, so this option allows you to
    override the automatic test.</para>
  </listitem>

  <listitem>
    <para><option>--with-vex=</option></para>
    <para>Specifies the path to the underlying VEX dynamic-translation
     library.  By default this is taken to be in the VEX directory off
     the root of the source tree.
   </para>
  </listitem>

  <listitem>
    <para><option>--enable-only64bit</option></para>
    <para><option>--enable-only32bit</option></para>
    <para>On 64-bit
     platforms (amd64-linux, ppc64-linux), Valgrind is by default built
     in such a way that both 32-bit and 64-bit executables can be run.
     Sometimes this cleverness is a problem for a variety of reasons.
     These two flags allow for single-target builds in this situation.
     If you issue both, the configure script will complain.  Note they
     are ignored on 32-bit-only platforms (x86-linux, ppc32-linux).
   </para>
  </listitem>

</itemizedlist>
</para>

<para>The <computeroutput>configure</computeroutput> script tests
the version of the X server currently indicated by the current
<computeroutput>$DISPLAY</computeroutput>.  This is a known bug.
The intention was to detect the version of the current XFree86
client libraries, so that correct suppressions could be selected
for them, but instead the test checks the server version.  This
is just plain wrong.</para>

<para>If you are building a binary package of Valgrind for
distribution, please read <literal>README_PACKAGERS</literal>
<xref linkend="dist.readme-packagers"/>.  It contains some
important information.</para>

<para>Apart from that, there's not much excitement here.  Let us
know if you have build problems.</para>

</sect1>


<sect1 id="manual-core.problems" xreflabel="If You Have Problems">
<title>If You Have Problems</title>

<para>Contact us at <ulink url="&vg-url;">&vg-url;</ulink>.</para>

<para>See <xref linkend="manual-core.limits"/> for the known
limitations of Valgrind, and for a list of programs which are
known not to work on it.</para>

<para>All parts of the system make heavy use of assertions and 
internal self-checks.  They are permanently enabled, and we have no 
plans to disable them.  If one of them breaks, please mail us!</para>

<para>If you get an assertion failure on the expression
<computeroutput>blockSane(ch)</computeroutput> in
<computeroutput>VG_(free)()</computeroutput> in
<filename>m_mallocfree.c</filename>, this may have happened because
your program wrote off the end of a malloc'd block, or before its
beginning.  Valgrind hopefully will have emitted a proper message to that
effect before dying in this way.  This is a known problem which
we should fix.</para>

<para>Read the <xref linkend="FAQ"/> for more advice about common problems, 
crashes, etc.</para>

</sect1>


<sect1 id="manual-core.limits" xreflabel="Limitations">
<title>Limitations</title>

<para>The following list of limitations seems long.  However, most
programs actually work fine.</para>

<para>Valgrind will run Linux ELF binaries, on a kernel 2.4.X or 2.6.X
system, on the x86, amd64, ppc32 and ppc64 architectures, subject to the
following constraints:</para>

 <itemizedlist>
  <listitem>
   <para>On x86 and amd64, there is no support for 3DNow! instructions.
   If the translator encounters these, Valgrind will generate a SIGILL
   when the instruction is executed.  Apart from that, on x86 and amd64,
   essentially all instructions are supported, up to and including SSE2.
   Version 3.1.0 includes limited support for SSE3 on x86.  This could
   be improved if necessary.</para>

   <para>On ppc32 and ppc64, almost all integer, floating point and Altivec
   instructions are supported.  Specifically: integer and FP insns that are
   mandatory for PowerPC, the "General-purpose optional" group (fsqrt, fsqrts,
   stfiwx), the "Graphics optional" group (fre, fres, frsqrte, frsqrtes), and
   the Altivec (also known as VMX) SIMD instruction set, are supported.</para>
  </listitem>

  <listitem>
   <para>Atomic instruction sequences are not properly supported, in the
   sense that their atomicity is not preserved.  This will affect any
   use of synchronization via memory shared between processes.  They
   will appear to work, but fail sporadically.</para>
  </listitem>

  <listitem>
   <para>If your program does its own memory management, rather than
   using malloc/new/free/delete, it should still work, but Valgrind's
   error checking won't be so effective.  If you describe your program's
   memory management scheme using "client requests" 
   (see <xref linkend="manual-core.clientreq"/>), Memcheck can do
   better.  Nevertheless, using malloc/new and free/delete is still the
   best approach.</para>
  </listitem>

  <listitem>
   <para>Valgrind's signal simulation is not as robust as it could be.
   Basic POSIX-compliant sigaction and sigprocmask functionality is
   supplied, but it's conceivable that things could go badly awry if you
   do weird things with signals.  Workaround: don't.  Programs that do
   non-POSIX signal tricks are in any case inherently unportable, so
   should be avoided if possible.</para>
  </listitem>

  <listitem>
   <para>Machine instructions, and system calls, have been implemented
   on demand.  So it's possible, although unlikely, that a program will
   fall over with a message to that effect.  If this happens, please
   report ALL the details printed out, so we can try and implement the
   missing feature.</para>
  </listitem>

  <listitem>
   <para>Memory consumption of your program is majorly increased whilst
   running under Valgrind.  This is due to the large amount of
   administrative information maintained behind the scenes.  Another
   cause is that Valgrind dynamically translates the original
   executable.  Translated, instrumented code is 12-18 times larger than
   the original so you can easily end up with 50+ MB of translations
   when running (eg) a web browser.</para>
  </listitem>

  <listitem>
   <para>Valgrind can handle dynamically-generated code just fine.  If
   you regenerate code over the top of old code (ie. at the same memory
   addresses), if the code is on the stack Valgrind will realise the
   code has changed, and work correctly.  This is necessary to handle
   the trampolines GCC uses to implemented nested functions.  If you
   regenerate code somewhere other than the stack, you will need to use
   the <option>--smc-check=all</option> flag, and Valgrind will run more
   slowly than normal.</para>
  </listitem>

  <listitem>
   <para>As of version 3.0.0, Valgrind has the following limitations
   in its implementation of x86/AMD64 floating point relative to 
   IEEE754.</para>

   <para>Precision: There is no support for 80 bit arithmetic.
   Internally, Valgrind represents all such "long double" numbers in 64
   bits, and so there may be some differences in results.  Whether or
   not this is critical remains to be seen.  Note, the x86/amd64
   fldt/fstpt instructions (read/write 80-bit numbers) are correctly
   simulated, using conversions to/from 64 bits, so that in-memory
   images of 80-bit numbers look correct if anyone wants to see.</para>

   <para>The impression observed from many FP regression tests is that
   the accuracy differences aren't significant.  Generally speaking, if
   a program relies on 80-bit precision, there may be difficulties
   porting it to non x86/amd64 platforms which only support 64-bit FP
   precision.  Even on x86/amd64, the program may get different results
   depending on whether it is compiled to use SSE2 instructions (64-bits
   only), or x87 instructions (80-bit).  The net effect is to make FP
   programs behave as if they had been run on a machine with 64-bit IEEE
   floats, for example PowerPC.  On amd64 FP arithmetic is done by
   default on SSE2, so amd64 looks more like PowerPC than x86 from an FP
   perspective, and there are far fewer noticable accuracy differences
   than with x86.</para>

   <para>Rounding: Valgrind does observe the 4 IEEE-mandated rounding
   modes (to nearest, to +infinity, to -infinity, to zero) for the
   following conversions: float to integer, integer to float where
   there is a possibility of loss of precision, and float-to-float
   rounding.  For all other FP operations, only the IEEE default mode
   (round to nearest) is supported.</para>

   <para>Numeric exceptions in FP code: IEEE754 defines five types of
   numeric exception that can happen: invalid operation (sqrt of
   negative number, etc), division by zero, overflow, underflow,
   inexact (loss of precision).</para>

   <para>For each exception, two courses of action are defined by 754:
   either (1) a user-defined exception handler may be called, or (2) a
   default action is defined, which "fixes things up" and allows the
   computation to proceed without throwing an exception.</para>

   <para>Currently Valgrind only supports the default fixup actions.
   Again, feedback on the importance of exception support would be
   appreciated.</para>

   <para>When Valgrind detects that the program is trying to exceed any
   of these limitations (setting exception handlers, rounding mode, or
   precision control), it can print a message giving a traceback of
   where this has happened, and continue execution.  This behaviour used
   to be the default, but the messages are annoying and so showing them
   is now optional.  Use <option>--show-emwarns=yes</option> to see
   them.</para>

   <para>The above limitations define precisely the IEEE754 'default'
   behaviour: default fixup on all exceptions, round-to-nearest
   operations, and 64-bit precision.</para>
  </listitem>
   
  <listitem>
   <para>As of version 3.0.0, Valgrind has the following limitations in
   its implementation of x86/AMD64 SSE2 FP arithmetic, relative to 
   IEEE754.</para>

   <para>Essentially the same: no exceptions, and limited observance of
   rounding mode.  Also, SSE2 has control bits which make it treat
   denormalised numbers as zero (DAZ) and a related action, flush
   denormals to zero (FTZ).  Both of these cause SSE2 arithmetic to be
   less accurate than IEEE requires.  Valgrind detects, ignores, and can
   warn about, attempts to enable either mode.</para>
  </listitem>

  <listitem>
   <para>As of version 3.2.0, Valgrind has the following limitations
   in its implementation of PPC32 and PPC64 floating point 
   arithmetic, relative to IEEE754.</para>

   <para>Scalar (non-Altivec): Valgrind provides a bit-exact emulation of
   all floating point instructions, except for "fre" and "fres", which are
   done more precisely than required by the PowerPC architecture specification.
   All floating point operations observe the current rounding mode.
   </para>

   <para>However, fpscr[FPRF] is not set after each operation.  That could
   be done but would give measurable performance overheads, and so far
   no need for it has been found.</para>

   <para>As on x86/AMD64, IEEE754 exceptions are not supported: all floating
   point exceptions are handled using the default IEEE fixup actions.
   Valgrind detects, ignores, and can warn about, attempts to unmask 
   the 5 IEEE FP exception kinds by writing to the floating-point status 
   and control register (fpscr).
   </para>

   <para>Vector (Altivec, VMX): essentially as with x86/AMD64 SSE/SSE2: 
   no exceptions, and limited observance of rounding mode.  
   For Altivec, FP arithmetic
   is done in IEEE/Java mode, which is more accurate than the Linux default
   setting.  "More accurate" means that denormals are handled properly, 
   rather than simply being flushed to zero.</para>
  </listitem>
 </itemizedlist>

 <para>Programs which are known not to work are:</para>
 <itemizedlist>
  <listitem>
   <para>emacs starts up but immediately concludes it is out of
   memory and aborts.  It may be that Memcheck does not provide
   a good enough emulation of the 
   <computeroutput>mallinfo</computeroutput> function.
   Emacs works fine if you build it to use
   the standard malloc/free routines.</para>
  </listitem>
 </itemizedlist>

</sect1>


<sect1 id="manual-core.example" xreflabel="An Example Run">
<title>An Example Run</title>

<para>This is the log for a run of a small program using Memcheck
The program is in fact correct, and the reported error is as the
result of a potentially serious code generation bug in GNU g++
(snapshot 20010527).</para>

<programlisting><![CDATA[
sewardj@phoenix:~/newmat10$
~/Valgrind-6/valgrind -v ./bogon 
==25832== Valgrind 0.10, a memory error detector for x86 RedHat 7.1.
==25832== Copyright (C) 2000-2001, and GNU GPL'd, by Julian Seward.
==25832== Startup, with flags:
==25832== --suppressions=/home/sewardj/Valgrind/redhat71.supp
==25832== reading syms from /lib/ld-linux.so.2
==25832== reading syms from /lib/libc.so.6
==25832== reading syms from /mnt/pima/jrs/Inst/lib/libgcc_s.so.0
==25832== reading syms from /lib/libm.so.6
==25832== reading syms from /mnt/pima/jrs/Inst/lib/libstdc++.so.3
==25832== reading syms from /home/sewardj/Valgrind/valgrind.so
==25832== reading syms from /proc/self/exe
==25832== 
==25832== Invalid read of size 4
==25832==    at 0x8048724: _ZN10BandMatrix6ReSizeEiii (bogon.cpp:45)
==25832==    by 0x80487AF: main (bogon.cpp:66)
==25832==  Address 0xBFFFF74C is not stack'd, malloc'd or free'd
==25832==
==25832== ERROR SUMMARY: 1 errors from 1 contexts (suppressed: 0 from 0)
==25832== malloc/free: in use at exit: 0 bytes in 0 blocks.
==25832== malloc/free: 0 allocs, 0 frees, 0 bytes allocated.
==25832== For a detailed leak analysis, rerun with: --leak-check=yes
==25832==
==25832== exiting, did 1881 basic blocks, 0 misses.
==25832== 223 translations, 3626 bytes in, 56801 bytes out.]]></programlisting>

<para>The GCC folks fixed this about a week before gcc-3.0
shipped.</para>

</sect1>


<sect1 id="manual-core.warnings" xreflabel="Warning Messages">
<title>Warning Messages You Might See</title>

<para>Most of these only appear if you run in verbose mode
(enabled by <computeroutput>-v</computeroutput>):</para>

 <itemizedlist>

  <listitem>
    <para><computeroutput>More than 100 errors detected.  Subsequent
    errors will still be recorded, but in less detail than
    before.</computeroutput></para>

    <para>After 100 different errors have been shown, Valgrind becomes
    more conservative about collecting them.  It then requires only the
    program counters in the top two stack frames to match when deciding
    whether or not two errors are really the same one.  Prior to this
    point, the PCs in the top four frames are required to match.  This
    hack has the effect of slowing down the appearance of new errors
    after the first 100.  The 100 constant can be changed by recompiling
    Valgrind.</para>
  </listitem>

  <listitem>
    <para><computeroutput>More than 1000 errors detected.  I'm not
    reporting any more.  Final error counts may be inaccurate.  Go fix
    your program!</computeroutput></para>

    <para>After 1000 different errors have been detected, Valgrind
    ignores any more.  It seems unlikely that collecting even more
    different ones would be of practical help to anybody, and it avoids
    the danger that Valgrind spends more and more of its time comparing
    new errors against an ever-growing collection.  As above, the 1000
    number is a compile-time constant.</para>
  </listitem>

  <listitem>
    <para><computeroutput>Warning: client switching stacks?</computeroutput></para>

    <para>Valgrind spotted such a large change in the stack pointer,
    <literal>%esp</literal>, that it guesses the client is switching to
    a different stack.  At this point it makes a kludgey guess where the
    base of the new stack is, and sets memory permissions accordingly.
    You may get many bogus error messages following this, if Valgrind
    guesses wrong.  At the moment "large change" is defined as a change
    of more that 2000000 in the value of the <literal>%esp</literal>
    (stack pointer) register.</para>
  </listitem>

  <listitem>
    <para><computeroutput>Warning: client attempted to close Valgrind's
    logfile fd &lt;number&gt;</computeroutput></para>

    <para>Valgrind doesn't allow the client to close the logfile,
    because you'd never see any diagnostic information after that point.
    If you see this message, you may want to use the
    <option>--log-fd=&lt;number&gt;</option> option to specify a
    different logfile file-descriptor number.</para>
  </listitem>

  <listitem>
    <para><computeroutput>Warning: noted but unhandled ioctl
    &lt;number&gt;</computeroutput></para>

    <para>Valgrind observed a call to one of the vast family of
    <computeroutput>ioctl</computeroutput> system calls, but did not
    modify its memory status info (because I have not yet got round to
    it).  The call will still have gone through, but you may get
    spurious errors after this as a result of the non-update of the
    memory info.</para>
  </listitem>

  <listitem>
    <para><computeroutput>Warning: set address range perms: large range
    &lt;number></computeroutput></para>

    <para>Diagnostic message, mostly for benefit of the Valgrind
    developers, to do with memory permissions.</para>
  </listitem>

 </itemizedlist>

</sect1>


<sect1 id="manual-core.mpiwrap" xreflabel="MPI Wrappers">
<title>Debugging MPI Parallel Programs with Valgrind</title>

<para> Valgrind supports debugging of distributed-memory applications
which use the MPI message passing standard.  This support consists of a
library of wrapper functions for the
<computeroutput>PMPI_*</computeroutput> interface.  When incorporated
into the application's address space, either by direct linking or by
<computeroutput>LD_PRELOAD</computeroutput>, the wrappers intercept
calls to <computeroutput>PMPI_Send</computeroutput>,
<computeroutput>PMPI_Recv</computeroutput>, etc.  They then
use client requests to inform Valgrind of memory state changes caused
by the function being wrapped.  This reduces the number of false
positives that Memcheck otherwise typically reports for MPI
applications.</para>

<para>The wrappers also take the opportunity to carefully check
size and definedness of buffers passed as arguments to MPI functions, hence
detecting errors such as passing undefined data to
<computeroutput>PMPI_Send</computeroutput>, or receiving data into a
buffer which is too small.</para>

<para>Unlike the rest of Valgrind, the wrapper library is subject to a
BSD-style license, so you can link it into any code base you like.
See the top of <computeroutput>auxprogs/libmpiwrap.c</computeroutput>
for details.</para>


<sect2 id="manual-core.mpiwrap.build" xreflabel="Building MPI Wrappers">
<title>Building and installing the wrappers</title>

<para> The wrapper library will be built automatically if possible.
Valgrind's configure script will look for a suitable
<computeroutput>mpicc</computeroutput> to build it with.  This must be
the same <computeroutput>mpicc</computeroutput> you use to build the
MPI application you want to debug.  By default, Valgrind tries
<computeroutput>mpicc</computeroutput>, but you can specify a
different one by using the configure-time flag
<computeroutput>--with-mpicc=</computeroutput>.  Currently the
wrappers are only buildable with
<computeroutput>mpicc</computeroutput>s which are based on GNU
<computeroutput>gcc</computeroutput> or Intel's
<computeroutput>icc</computeroutput>.</para>

<para>Check that the configure script prints a line like this:</para>

<programlisting><![CDATA[
checking for usable MPI2-compliant mpicc and mpi.h... yes, mpicc
]]></programlisting>

<para>If it says <computeroutput>... no</computeroutput>, your
<computeroutput>mpicc</computeroutput> has failed to compile and link
a test MPI2 program.</para>

<para>If the configure test succeeds, continue in the usual way with
<computeroutput>make</computeroutput> and <computeroutput>make
install</computeroutput>.  The final install tree should then contain
<computeroutput>libmpiwrap.so</computeroutput>.
</para>

<para>Compile up a test MPI program (eg, MPI hello-world) and try
this:</para>

<programlisting><![CDATA[
LD_PRELOAD=$prefix/lib/valgrind/<platform>/libmpiwrap.so   \
           mpirun [args] $prefix/bin/valgrind ./hello
]]></programlisting>

<para>You should see something similar to the following</para>

<programlisting><![CDATA[
valgrind MPI wrappers 31901: Active for pid 31901
valgrind MPI wrappers 31901: Try MPIWRAP_DEBUG=help for possible options
]]></programlisting>

<para>repeated for every process in the group.  If you do not see
these, there is an build/installation problem of some kind.</para>

<para> The MPI functions to be wrapped are assumed to be in an ELF
shared object with soname matching
<computeroutput>libmpi.so*</computeroutput>.  This is known to be
correct at least for Open MPI and Quadrics MPI, and can easily be
changed if required.</para> 
</sect2>


<sect2 id="manual-core.mpiwrap.gettingstarted" 
       xreflabel="Getting started with MPI Wrappers">
<title>Getting started</title>

<para>Compile your MPI application as usual, taking care to link it
using the same <computeroutput>mpicc</computeroutput> that your
Valgrind build was configured with.</para>

<para>
Use the following basic scheme to run your application on Valgrind with
the wrappers engaged:</para>

<programlisting><![CDATA[
MPIWRAP_DEBUG=[wrapper-args]                                  \
   LD_PRELOAD=$prefix/lib/valgrind/<platform>/libmpiwrap.so   \
   mpirun [mpirun-args]                                       \
   $prefix/bin/valgrind [valgrind-args]                       \
   [application] [app-args]
]]></programlisting>

<para>As an alternative to
<computeroutput>LD_PRELOAD</computeroutput>ing
<computeroutput>libmpiwrap.so</computeroutput>, you can simply link it
to your application if desired.  This should not disturb native
behaviour of your application in any way.</para>
</sect2>


<sect2 id="manual-core.mpiwrap.controlling" 
       xreflabel="Controlling the MPI Wrappers">
<title>Controlling the wrapper library</title>

<para>Environment variable
<computeroutput>MPIWRAP_DEBUG</computeroutput> is consulted at
startup.  The default behaviour is to print a starting banner</para>

<programlisting><![CDATA[
valgrind MPI wrappers 16386: Active for pid 16386
valgrind MPI wrappers 16386: Try MPIWRAP_DEBUG=help for possible options
]]></programlisting>

<para> and then be relatively quiet.</para>

<para>You can give a list of comma-separated options in
<computeroutput>MPIWRAP_DEBUG</computeroutput>.  These are</para>

<itemizedlist>
  <listitem>
    <para><computeroutput>verbose</computeroutput>:
    show entries/exits of all wrappers.  Also show extra
    debugging info, such as the status of outstanding 
    <computeroutput>MPI_Request</computeroutput>s resulting
    from uncompleted <computeroutput>MPI_Irecv</computeroutput>s.</para>
  </listitem>
  <listitem>
    <para><computeroutput>quiet</computeroutput>: 
    opposite of <computeroutput>verbose</computeroutput>, only print 
    anything when the wrappers want
    to report a detected programming error, or in case of catastrophic
    failure of the wrappers.</para>
  </listitem>
  <listitem>
    <para><computeroutput>warn</computeroutput>: 
    by default, functions which lack proper wrappers
    are not commented on, just silently
    ignored.  This causes a warning to be printed for each unwrapped
    function used, up to a maximum of three warnings per function.</para>
  </listitem>
  <listitem>
    <para><computeroutput>strict</computeroutput>: 
    print an error message and abort the program if 
    a function lacking a wrapper is used.</para>
  </listitem>
</itemizedlist>

<para> If you want to use Valgrind's XML output facility
(<computeroutput>--xml=yes</computeroutput>), you should pass
<computeroutput>quiet</computeroutput> in
<computeroutput>MPIWRAP_DEBUG</computeroutput> so as to get rid of any
extraneous printing from the wrappers.</para>

</sect2>


<sect2 id="manual-core.mpiwrap.limitations" 
       xreflabel="Abilities and Limitations of MPI Wrappers">
<title>Abilities and limitations</title>

<sect3>
<title>Functions</title>

<para>All MPI2 functions except
<computeroutput>MPI_Wtick</computeroutput>,
<computeroutput>MPI_Wtime</computeroutput> and
<computeroutput>MPI_Pcontrol</computeroutput> have wrappers.  The
first two are not wrapped because they return a 
<computeroutput>double</computeroutput>, and Valgrind's
function-wrap mechanism cannot handle that (it could easily enough be
extended to).  <computeroutput>MPI_Pcontrol</computeroutput> cannot be
wrapped as it has variable arity: 
<computeroutput>int MPI_Pcontrol(const int level, ...)</computeroutput></para>

<para>Most functions are wrapped with a default wrapper which does
nothing except complain or abort if it is called, depending on
settings in <computeroutput>MPIWRAP_DEBUG</computeroutput> listed
above.  The following functions have "real", do-something-useful
wrappers:</para>

<programlisting><![CDATA[
PMPI_Send PMPI_Bsend PMPI_Ssend PMPI_Rsend

PMPI_Recv PMPI_Get_count

PMPI_Isend PMPI_Ibsend PMPI_Issend PMPI_Irsend

PMPI_Irecv
PMPI_Wait PMPI_Waitall
PMPI_Test PMPI_Testall

PMPI_Iprobe PMPI_Probe

PMPI_Cancel

PMPI_Sendrecv

PMPI_Type_commit PMPI_Type_free

PMPI_Bcast PMPI_Gather PMPI_Scatter PMPI_Alltoall
PMPI_Reduce PMPI_Allreduce PMPI_Op_create

PMPI_Comm_create PMPI_Comm_dup PMPI_Comm_free PMPI_Comm_rank PMPI_Comm_size

PMPI_Error_string
PMPI_Init PMPI_Initialized PMPI_Finalize
]]></programlisting>

<para> A few functions such as
<computeroutput>PMPI_Address</computeroutput> are listed as
<computeroutput>HAS_NO_WRAPPER</computeroutput>.  They have no wrapper
at all as there is nothing worth checking, and giving a no-op wrapper
would reduce performance for no reason.</para>

<para> Note that the wrapper library itself can itself generate large
numbers of calls to the MPI implementation, especially when walking
complex types.  The most common functions called are
<computeroutput>PMPI_Extent</computeroutput>,
<computeroutput>PMPI_Type_get_envelope</computeroutput>,
<computeroutput>PMPI_Type_get_contents</computeroutput>, and
<computeroutput>PMPI_Type_free</computeroutput>.  </para>
</sect3>

<sect3>
<title>Types</title>

<para> MPI-1.1 structured types are supported, and walked exactly.
The currently supported combiners are
<computeroutput>MPI_COMBINER_NAMED</computeroutput>,
<computeroutput>MPI_COMBINER_CONTIGUOUS</computeroutput>,
<computeroutput>MPI_COMBINER_VECTOR</computeroutput>,
<computeroutput>MPI_COMBINER_HVECTOR</computeroutput>
<computeroutput>MPI_COMBINER_INDEXED</computeroutput>,
<computeroutput>MPI_COMBINER_HINDEXED</computeroutput> and
<computeroutput>MPI_COMBINER_STRUCT</computeroutput>.  This should
cover all MPI-1.1 types.  The mechanism (function
<computeroutput>walk_type</computeroutput>) should extend easily to
cover MPI2 combiners.</para>

<para>MPI defines some named structured types
(<computeroutput>MPI_FLOAT_INT</computeroutput>,
<computeroutput>MPI_DOUBLE_INT</computeroutput>,
<computeroutput>MPI_LONG_INT</computeroutput>,
<computeroutput>MPI_2INT</computeroutput>,
<computeroutput>MPI_SHORT_INT</computeroutput>,
<computeroutput>MPI_LONG_DOUBLE_INT</computeroutput>) which are pairs
of some basic type and a C <computeroutput>int</computeroutput>.
Unfortunately the MPI specification makes it impossible to look inside
these types and see where the fields are.  Therefore these wrappers
assume the types are laid out as <computeroutput>struct { float val;
int loc; }</computeroutput> (for
<computeroutput>MPI_FLOAT_INT</computeroutput>), etc, and act
accordingly.  This appears to be correct at least for Open MPI 1.0.2
and for Quadrics MPI.</para>

<para>If <computeroutput>strict</computeroutput> is an option specified 
in <computeroutput>MPIWRAP_DEBUG</computeroutput>, the application
will abort if an unhandled type is encountered.  Otherwise, the 
application will print a warning message and continue.</para>

<para>Some effort is made to mark/check memory ranges corresponding to
arrays of values in a single pass.  This is important for performance
since asking Valgrind to mark/check any range, no matter how small,
carries quite a large constant cost.  This optimisation is applied to
arrays of primitive types (<computeroutput>double</computeroutput>,
<computeroutput>float</computeroutput>,
<computeroutput>int</computeroutput>,
<computeroutput>long</computeroutput>, <computeroutput>long
long</computeroutput>, <computeroutput>short</computeroutput>,
<computeroutput>char</computeroutput>, and <computeroutput>long
double</computeroutput> on platforms where <computeroutput>sizeof(long
double) == 8</computeroutput>).  For arrays of all other types, the
wrappers handle each element individually and so there can be a very
large performance cost.</para>

</sect3>

</sect2>


<sect2 id="manual-core.mpiwrap.writingwrappers" 
       xreflabel="Writing new MPI Wrappers">
<title>Writing new wrappers</title>

<para>
For the most part the wrappers are straightforward.  The only
significant complexity arises with nonblocking receives.</para>

<para>The issue is that <computeroutput>MPI_Irecv</computeroutput>
states the recv buffer and returns immediately, giving a handle
(<computeroutput>MPI_Request</computeroutput>) for the transaction.
Later the user will have to poll for completion with
<computeroutput>MPI_Wait</computeroutput> etc, and when the
transaction completes successfully, the wrappers have to paint the
recv buffer.  But the recv buffer details are not presented to
<computeroutput>MPI_Wait</computeroutput> -- only the handle is.  The
library therefore maintains a shadow table which associates
uncompleted <computeroutput>MPI_Request</computeroutput>s with the
corresponding buffer address/count/type.  When an operation completes,
the table is searched for the associated address/count/type info, and
memory is marked accordingly.</para>

<para>Access to the table is guarded by a (POSIX pthreads) lock, so as
to make the library thread-safe.</para>

<para>The table is allocated with
<computeroutput>malloc</computeroutput> and never
<computeroutput>free</computeroutput>d, so it will show up in leak
checks.</para>

<para>Writing new wrappers should be fairly easy.  The source file is
<computeroutput>auxprogs/libmpiwrap.c</computeroutput>.  If possible,
find an existing wrapper for a function of similar behaviour to the
one you want to wrap, and use it as a starting point.  The wrappers
are organised in sections in the same order as the MPI 1.1 spec, to
aid navigation.  When adding a wrapper, remember to comment out the
definition of the default wrapper in the long list of defaults at the
bottom of the file (do not remove it, just comment it out).</para>
</sect2>

<sect2 id="manual-core.mpiwrap.whattoexpect" 
       xreflabel="What to expect with MPI Wrappers">
<title>What to expect when using the wrappers</title>

<para>The wrappers should reduce Memcheck's false-error rate on MPI
applications.  Because the wrapping is done at the MPI interface,
there will still potentially be a large number of errors reported in
the MPI implementation below the interface.  The best you can do is
try to suppress them.</para>

<para>You may also find that the input-side (buffer
length/definedness) checks find errors in your MPI use, for example
passing too short a buffer to
<computeroutput>MPI_Recv</computeroutput>.</para>

<para>Functions which are not wrapped may increase the false
error rate.  A possible approach is to run with
<computeroutput>MPI_DEBUG</computeroutput> containing
<computeroutput>warn</computeroutput>.  This will show you functions
which lack proper wrappers but which are nevertheless used.  You can
then write wrappers for them.
</para>

</sect2>

</sect1>

</chapter>