hacking.html.in

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE html>
<html xmlns="http://www.w3.org/1999/xhtml">
  <body>
    <h1>Contributor guidelines</h1>

    <ul id="toc"></ul>

    <h2><a id="patches">General tips for contributing patches</a></h2>
    <ol>
      <li>
        <p>Discuss any large changes on the mailing list first.  Post patches
        early and listen to feedback.</p>
      </li>

      <li>
        <p>Official upstream repository is kept in git
        (<code>https://libvirt.org/git/libvirt.git</code>) and is browsable
        along with other libvirt-related repositories
        (e.g. libvirt-python) <a href="https://libvirt.org/git/">online</a>.</p>
      </li>

      <li>
        <p>Patches to translations are maintained via
        the <a href="https://fedora.zanata.org/">zanata project</a>.
        If you want to fix a translation in a .po file, join the
        appropriate language team. The libvirt release process
        automatically pulls the latest version of each translation
        file from zanata.</p>
      </li>

      <li><p>Post patches using <code>git send-email</code>, with git
        rename detection enabled.  You need a one-time setup of:</p>
<pre>
  git config diff.renames true
</pre>
        <p>Also, for code motion patches, you may find that <code>git
        diff --patience</code> provides an easier-to-read patch.
        However, the usual workflow of libvirt developer is:</p>
<pre>
  git checkout master
  git pull
  git checkout -t origin -b workbranch
  Hack, committing any changes along the way
</pre>
        <p>More hints on compiling can be
        found <a href="compiling.html">here</a>.  When you want to
        post your patches:</p>
<pre>
  git pull --rebase
  (fix any conflicts)
  git send-email --cover-letter --no-chain-reply-to --annotate \
                 --confirm=always --to=libvir-list@redhat.com master
</pre>
        <p>For a single patch you can omit
        <code>--cover-letter</code>, but a series of two or more
        patches needs a cover letter.</p>
        <p>Note that the <code>git send-email</code> subcommand may not
        be in the main git package and using it may require installation
        of a separate package, for example the "git-email" package in
        Fedora and Debian.  If this is your first time using
        <code>git send-email</code>, you might need to configure it to
        point it to your SMTP server with something like:</p>
<pre>
  git config --global sendemail.smtpServer stmp.youremailprovider.net
</pre>
        <p>If you get tired of typing
        <code>--to=libvir-list@redhat.com</code> all the time, you can
        configure that to be automatically handled as well:</p>
<pre>
  git config sendemail.to libvir-list@redhat.com
</pre>
        <p>As a rule, patches should be sent to the mailing list only: all
        developers are subscribed to libvir-list and read it regularly, so
        please don't CC individual developers unless they've explicitly
        asked you to.</p>
        <p>Avoid using mail clients for sending patches, as most of them
        will mangle the messages in some way, making them unusable for our
        purposes.  Gmail and other Web-based mail clients are particularly
        bad at this.</p>
        <p>If everything went well, your patch should show up on the
        <a href="https://www.redhat.com/archives/libvir-list/">libvir-list
        archives</a> in a matter of minutes; if you still can't find it on
        there after an hour or so, you should double-check your setup.  Note
        that your very first post to the mailing list will be subject to
        moderation, and it's not uncommon for that to take around a day.</p>
        <p>Please follow this as close as you can, especially the rebase and
        <code>git send-email</code> part, as it makes life easier for other
        developers to review your patch set.</p>
        <p>One should avoid sending patches as attachments,
        but rather send them in email body along with commit message. If a
        developer is sending another version of the patch (e.g. to address
        review comments), they are advised to note differences to previous
        versions after the <code>---</code> line in the patch so that it helps
        reviewers but doesn't become part of git history. Moreover, such patch
        needs to be prefixed correctly with
        <code>--subject-prefix=PATCHv2</code> appended to <code>git
            send-email</code> (substitute <code>v2</code> with the correct
        version if needed though).</p>
      </li>

      <li><p>In your commit message, make the summary line reasonably
          short (60 characters is typical), followed by a blank line,
          followed by any longer description of why your patch makes
          sense.  If the patch fixes a regression, and you know what
          commit introduced the problem, mentioning that is useful.
          If the patch resolves a bugzilla report, mentioning the URL
          of the bug number is useful; but also summarize the issue
          rather than making all readers follow the link.  You can use
          'git shortlog -30' to get an idea of typical summary lines.
        </p>
      </li>

      <li><p>Contributors to libvirt projects <strong>must</strong>
          assert that they are in compliance with the
          <a href="https://developercertificate.org/">Developer
          Certificate of Origin 1.1</a>. This is achieved by adding
          a "Signed-off-by" line containing the contributor's name
          and e-mail to every commit message. The presence
          of this line attests that the contributor has read the
          above lined DCO and agrees with its statements.
      </p></li>

      <li><p>Split large changes into a series of smaller patches,
        self-contained if possible, with an explanation of each patch
        and an explanation of how the sequence of patches fits
        together. Moreover, please keep in mind that it's required to
        be able to compile cleanly (<b>including</b> <code>make
        check</code> and <code>make syntax-check</code>) after each
        patch.  A feature does not have to work until the end of a
        series, but intermediate patches must compile and not cause
        test-suite failures (this is to preserve the usefulness
        of <code>git bisect</code>, among other things).</p>
      </li>

      <li>
        <p>Make sure your patches apply against libvirt GIT.  Developers
        only follow GIT and don't care much about released versions.</p>
      </li>

      <li><p>Run the automated tests on your code before submitting any changes.
          In particular, configure with compile warnings set to
          -Werror.  This is done automatically for a git checkout; from a
          tarball, use:</p>
<pre>
  ./configure --enable-werror
</pre>
        <p>
          and run the tests:
        </p>
<pre>
  make check
  make syntax-check
  make -C tests valgrind
</pre>
        <p><a href="http://valgrind.org/">Valgrind</a> is a test that checks
        for memory management issues, such as leaks or use of uninitialized
        variables.
        </p>

        <p>
          Some tests are skipped by default in a development environment,
          based on the time they take in comparison to the likelihood
          that those tests will turn up problems during incremental builds.
          These tests default to being run when building from a
          tarball or with the configure option --enable-expensive-tests;
          you can also force a one-time toggle of these tests by
          setting VIR_TEST_EXPENSIVE to 0 or 1 at make time, as in:
        </p>
<pre>
  make check VIR_TEST_EXPENSIVE=1
</pre>
        <p>
          If you encounter any failing tests, the VIR_TEST_DEBUG
          environment variable may provide extra information to debug
          the failures. Larger values of VIR_TEST_DEBUG may provide
          larger amounts of information:
        </p>

<pre>
  VIR_TEST_DEBUG=1 make check    (or)
  VIR_TEST_DEBUG=2 make check
</pre>

        <p>
          When debugging failures during development, it is possible
          to focus in on just the failing subtests by using TESTS and
          VIR_TEST_RANGE:
        </p>

<pre>
  make check VIR_TEST_DEBUG=1 VIR_TEST_RANGE=3-5 TESTS=qemuxml2argvtest
</pre>

        <p>
          Also, individual tests can be run from inside the <code>tests/</code>
          directory, like:
        </p>
<pre>
  ./qemuxml2xmltest
</pre>

        <p>
          If you are adding new test cases, or making changes that alter
          existing test output, you can use the environment variable
          VIR_TEST_REGENERATE_OUTPUT to quickly update the saved test data.
          Of course you still need to review the changes VERY CAREFULLY to
          ensure they are correct.
        </p>
<pre>
  VIR_TEST_REGENERATE_OUTPUT=1 ./qemuxml2argvtest
</pre>

        <p>There is also a <code>./run</code> script at the top level,
          to make it easier to run programs that have not yet been
          installed, as well as to wrap invocations of various tests
          under gdb or Valgrind.
        </p>

        <p>When running our test suite it may happen that the test result is
        nondeterministic because of the test suite relying on a particular file
        in the system being accessible or having some specific value. To catch
        this kind of errors, the test suite has a module for that prints any
        path touched that fulfils constraints described above
        into a file. To enable it just set
        <code>VIR_TEST_FILE_ACCESS</code> environment variable.
        Then <code>VIR_TEST_FILE_ACCESS_OUTPUT</code> environment
        variable can alter location where the file is stored.</p>
<pre>
  VIR_TEST_FILE_ACCESS=1 VIR_TEST_FILE_ACCESS_OUTPUT="/tmp/file_access.txt" ./qemuxml2argvtest
</pre>

      </li>
      <li><p>The Valgrind test should produce similar output to
          <code>make check</code>. If the output has traces within libvirt
          API's, then investigation is required in order to determine the
          cause of the issue. Output such as the following indicates some
          sort of leak:
        </p>
<pre>
==5414== 4 bytes in 1 blocks are definitely lost in loss record 3 of 89
==5414==    at 0x4A0881C: malloc (vg_replace_malloc.c:270)
==5414==    by 0x34DE0AAB85: xmlStrndup (in /usr/lib64/libxml2.so.2.7.8)
==5414==    by 0x4CC97A6: virDomainVideoDefParseXML (domain_conf.c:7410)
==5414==    by 0x4CD581D: virDomainDefParseXML (domain_conf.c:10188)
==5414==    by 0x4CD8C73: virDomainDefParseNode (domain_conf.c:10640)
==5414==    by 0x4CD8DDB: virDomainDefParse (domain_conf.c:10590)
==5414==    by 0x41CB1D: testCompareXMLToArgvHelper (qemuxml2argvtest.c:100)
==5414==    by 0x41E20F: virtTestRun (testutils.c:161)
==5414==    by 0x41C7CB: mymain (qemuxml2argvtest.c:866)
==5414==    by 0x41E84A: virtTestMain (testutils.c:723)
==5414==    by 0x34D9021734: (below main) (in /usr/lib64/libc-2.15.so)
</pre>
        <p>In this example, the <code>virDomainDefParseXML()</code> had
           an error path where the <code>virDomainVideoDefPtr video</code>
           pointer was not properly disposed. By simply adding a
           <code>virDomainVideoDefFree(video);</code> in the error path,
           the issue was resolved.
        </p>

        <p>Another common mistake is calling a printing function, such as
           <code>VIR_DEBUG()</code> without initializing a variable to be
           printed.  The following example involved a call which could return
           an error, but not set variables passed by reference to the call.
           The solution was to initialize the variables prior to the call.
        </p>
<pre>
==4749== Use of uninitialised value of size 8
==4749==    at 0x34D904650B: _itoa_word (in /usr/lib64/libc-2.15.so)
==4749==    by 0x34D9049118: vfprintf (in /usr/lib64/libc-2.15.so)
==4749==    by 0x34D9108F60: __vasprintf_chk (in /usr/lib64/libc-2.15.so)
==4749==    by 0x4CAEEF7: virVasprintf (stdio2.h:199)
==4749==    by 0x4C8A55E: virLogVMessage (virlog.c:814)
==4749==    by 0x4C8AA96: virLogMessage (virlog.c:751)
==4749==    by 0x4DA0056: virNetTLSContextCheckCertKeyUsage (virnettlscontext.c:225)
==4749==    by 0x4DA06DB: virNetTLSContextCheckCert (virnettlscontext.c:439)
==4749==    by 0x4DA1620: virNetTLSContextNew (virnettlscontext.c:562)
==4749==    by 0x4DA26FC: virNetTLSContextNewServer (virnettlscontext.c:927)
==4749==    by 0x409C39: testTLSContextInit (virnettlscontexttest.c:467)
==4749==    by 0x40AB8F: virtTestRun (testutils.c:161)
</pre>
        <p>Valgrind will also find some false positives or code paths
           which cannot be resolved by making changes to the libvirt code.
           For these paths, it is possible to add a filter to avoid the
           errors. For example:
        </p>
<pre>
==4643== 7 bytes in 1 blocks are possibly lost in loss record 4 of 20
==4643==    at 0x4A0881C: malloc (vg_replace_malloc.c:270)
==4643==    by 0x34D90853F1: strdup (in /usr/lib64/libc-2.15.so)
==4643==    by 0x34EEC2C08A: ??? (in /usr/lib64/libnl.so.1.1)
==4643==    by 0x34EEC15B81: ??? (in /usr/lib64/libnl.so.1.1)
==4643==    by 0x34D8C0EE15: call_init.part.0 (in /usr/lib64/ld-2.15.so)
==4643==    by 0x34D8C0EECF: _dl_init (in /usr/lib64/ld-2.15.so)
==4643==    by 0x34D8C01569: ??? (in /usr/lib64/ld-2.15.so)

</pre>
        <p>In this instance, it is acceptable to modify the
           <code>tests/.valgrind.supp</code> file in order to add a
           suppression filter. The filter should be unique enough to
           not suppress real leaks, but it should be generic enough to
           cover multiple code paths. The format of the entry can be
           found in the documentation found at the
           <a href="http://valgrind.org/">Valgrind home page</a>.
           The following trace was added to <code>tests/.valgrind.supp</code>
           in order to suppress the warning:
        </p>
<pre>
{
    dlInitMemoryLeak1
    Memcheck:Leak
    fun:?alloc
    ...
    fun:call_init.part.0
    fun:_dl_init
    ...
    obj:*/lib*/ld-2.*so*
}
</pre>
      </li>

      <li>
        <p>Update tests and/or documentation, particularly if you are adding
        a new feature or changing the output of a program.</p>
      </li>

      <li>
        <p>Don't forget to update the <a href="news.html">release notes</a>
        by changing <code>docs/news.xml</code> if your changes are
        significant. All user-visible changes, such as adding new XML elements
        or fixing all but the most obscure bugs, must be (briefly) described
        in a release notes entry; changes that are only relevant to other
        libvirt developers, such as code refactoring, don't belong in the
        release notes. Note that <code>docs/news.xml</code> should be updated
        in its own commit not to get in the way of backports.</p>
      </li>
    </ol>

    <p>
      There is more on this subject, including lots of links to background
      reading on the subject, on
      <a href="http://people.redhat.com/rjones/how-to-supply-code-to-open-source-projects/">
        Richard Jones' guide to working with open source projects</a>.
    </p>

    <h2><a id="tooling">Tooling</a></h2>

    <p>
      libvirt includes support for some useful development tools right in its
      source repository, meaning users will be able to take advantage of them
      without little or no configuration. Examples include:
    </p>

    <ul>
      <li>
        <a href="https://github.com/jeaye/color_coded">color_coded</a>,
        a vim plugin for libclang-powered semantic syntax highlighting;
      </li>

      <li>
        <a href="http://valloric.github.io/YouCompleteMe/">YouCompleteMe</a>,
        a vim plugin for libclang-powered semantic code completion.
      </li>
    </ul>

    <h2><a id="naming">Naming conventions</a></h2>

    <p>
      When reading libvirt code, a number of different naming conventions will
      be evident due to various changes in thinking over the course of the
      project's lifetime. The conventions documented below should be followed
      when creating any entirely new files in libvirt. When working on existing
      files, while it is desirable to apply these conventions, keeping a
      consistent style with existing code in that particular file is generally
      more important. The overall guiding principal is that every file, enum,
      struct, function, macro and typedef name must have a 'vir' or 'VIR' prefix.
      All local scope variable names are exempt, and global variables are exempt,
      unless exported in a header file.
    </p>

    <dl>
      <dt>File names</dt>
      <dd>
        <p>
          File naming varies depending on the subdirectory. The preferred
          style is to have a 'vir' prefix, followed by a name which matches
          the name of the functions / objects inside the file. For example,
          a file containing an object  'virHashtable' is stored in files
          'virhashtable.c' and 'virhashtable.h'. Sometimes, methods which
          would otherwise be declared 'static' need to be exported for use
          by a test suite. For this purpose a second header file should be
          added with a suffix of 'priv', e.g. 'virhashtablepriv.h'. Use of
          underscores in file names is discouraged when using the 'vir'
          prefix style. The 'vir' prefix naming applies to src/util,
          src/rpc and tests/ directories. Most other directories do not
          follow this convention.
        </p>
      </dd>
      <dt>Enum type &amp; field names</dt>
      <dd>
        <p>
          All enums should have a 'vir' prefix in their typedef name,
          and each following word should have its first letter in
          uppercase. The enum name should match the typedef name with
          a leading underscore. The enum member names should be in all
          uppercase, and use an underscore to separate each word. The
          enum member name prefix should match the enum typedef name.
        </p>
        <pre>
    typedef enum _virSocketType virSocketType;
    enum _virSocketType {
        VIR_SOCKET_TYPE_IPV4,
        VIR_SOCKET_TYPE_IPV6,
    };</pre>
      </dd>
      <dt>Struct type names</dt>
      <dd>
        <p>
          All structs should have a 'vir' prefix in their typedef name,
          and each following word should have its first letter in
          uppercase. The struct name should be the same as the typedef
          name with a leading underscore. A second typedef should be
          given for a pointer to the struct with a 'Ptr' suffix.
        </p>
        <pre>
    typedef struct _virHashTable virHashTable;
    typedef virHashTable *virHashTablePtr;
    struct _virHashTable {
       ...
    };</pre>
      </dd>
      <dt>Function names</dt>
      <dd>
        <p>
          All functions should have a 'vir' prefix in their name,
          followed by one or more words with first letter of each
          word capitalized. Underscores should not be used in function
          names. If the function is operating on an object, then the
          function name prefix should match the object typedef name,
          otherwise it should match the filename. Following this
          comes the verb / action name, and finally an optional
          subject name. For example, given an object 'virHashTable',
          all functions should have a name 'virHashTable$VERB' or
          'virHashTable$VERB$SUBJECT", e.g. 'virHashTableLookup'
          or 'virHashTableGetValue'.
        </p>
      </dd>
      <dt>Macro names</dt>
      <dd>
        <p>
          All macros should have a "VIR" prefix in their name, followed
          by one or more uppercase words separated by underscores. The
          macro argument names should be in lowercase. Aside from having
          a "VIR" prefix there are no common practices for the rest of
          the macro name.
        </p>
      </dd>
    </dl>

    <h2><a id="indent">Code indentation</a></h2>
    <p>
      Libvirt's C source code generally adheres to some basic code-formatting
      conventions.  The existing code base is not totally consistent on this
      front, but we do prefer that contributed code be formatted similarly.
      In short, use spaces-not-TABs for indentation, use 4 spaces for each
      indentation level, and other than that, follow the K&amp;R style.
    </p>

    <p>
      If you use Emacs, the project includes a file .dir-locals.el
      that sets up the preferred indentation. If you use vim,
      append the following to your ~/.vimrc file:
    </p>
<pre>
  set nocompatible
  filetype on
  set autoindent
  set smartindent
  set cindent
  set tabstop=8
  set shiftwidth=4
  set expandtab
  set cinoptions=(0,:0,l1,t0,L3
  filetype plugin indent on
  au FileType make setlocal noexpandtab
  au BufRead,BufNewFile *.am setlocal noexpandtab
  match ErrorMsg /\s\+$\| \+\ze\t/
</pre>
    <p>
      Or if you don't want to mess your ~/.vimrc up, you can save the above
      into a file called .lvimrc (not .vimrc) located at the root of libvirt
      source, then install a vim script from
      http://www.vim.org/scripts/script.php?script_id=1408,
      which will load the .lvimrc only when you edit libvirt code.
    </p>

    <h2><a id="formatting">Code formatting (especially for new code)</a></h2>

    <p>
      With new code, we can be even more strict.
      Please apply the following function (using GNU indent) to any new code.
      Note that this also gives you an idea of the type of spacing we prefer
      around operators and keywords:
    </p>

<pre>
  indent-libvirt()
  {
    indent -bad -bap -bbb -bli4 -br -ce -brs -cs -i4 -l75 -lc75 \
      -sbi4 -psl -saf -sai -saw -sbi4 -ss -sc -cdw -cli4 -npcs -nbc \
      --no-tabs "$@"
  }
</pre>

    <p>
      Note that sometimes you'll have to post-process that output further, by
      piping it through <code>expand -i</code>, since some leading TABs can get through.
      Usually they're in macro definitions or strings, and should be converted
      anyhow.
    </p>

    <p>
      Libvirt requires a C99 compiler for various reasons.  However,
      most of the code base prefers to stick to C89 syntax unless
      there is a compelling reason otherwise.  For example, it is
      preferable to use <code>/* */</code> comments rather
      than <code>//</code>.  Also, when declaring local variables, the
      prevailing style has been to declare them at the beginning of a
      scope, rather than immediately before use.
    </p>


    <h2><a id="bracket_spacing">Bracket spacing</a></h2>

    <p>
      The keywords <code>if</code>, <code>for</code>, <code>while</code>,
      and <code>switch</code> must have a single space following them
      before the opening bracket. E.g.
    </p>
    <pre>
      if(foo)   // Bad
      if (foo)  // Good
</pre>

    <p>
      Function implementations must <strong>not</strong> have any whitespace
      between the function name and the opening bracket. E.g.
    </p>
    <pre>
      int foo (int wizz)  // Bad
      int foo(int wizz)   // Good
</pre>

    <p>
      Function calls must <strong>not</strong> have any whitespace
      between the function name and the opening bracket. E.g.
    </p>
    <pre>
      bar = foo (wizz);  // Bad
      bar = foo(wizz);   // Good
</pre>

    <p>
      Function typedefs must <strong>not</strong> have any whitespace
      between the closing bracket of the function name and opening
      bracket of the arg list. E.g.
    </p>
    <pre>
      typedef int (*foo) (int wizz);  // Bad
      typedef int (*foo)(int wizz);   // Good
</pre>

    <p>
      There must not be any whitespace immediately following any
      opening bracket, or immediately prior to any closing bracket. E.g.
    </p>
    <pre>
      int foo( int wizz );  // Bad
      int foo(int wizz);    // Good
</pre>

    <h2><a id="comma">Commas</a></h2>

    <p>
      Commas should always be followed by a space or end of line, and
      never have leading space; this is enforced during 'make
      syntax-check'.
    </p>
    <pre>
      call(a,b ,c);// Bad
      call(a, b, c); // Good
</pre>

    <p>
      When declaring an enum or using a struct initializer that
      occupies more than one line, use a trailing comma.  That way,
      future edits to extend the list only have to add a line, rather
      than modify an existing line to add the intermediate comma.  Any
      sentinel enumerator value with a name ending in _LAST is exempt,
      since you would extend such an enum before the _LAST element.
      Another reason to favor trailing commas is that it requires less
      effort to produce via code generators.  Note that the syntax
      checker is unable to enforce a style of trailing commas, so
      there are counterexamples in existing code which do not use it;
      also, while C99 allows trailing commas, remember that JSON and
      XDR do not.
    </p>
    <pre>
      enum {
          VALUE_ONE,
          VALUE_TWO // Bad
      };
      enum {
          VALUE_THREE,
          VALUE_FOUR, // Good
      };
</pre>

    <h2><a id="semicolon">Semicolons</a></h2>

    <p>
      Semicolons should never have a space beforehand.  Inside the
      condition of a <code>for</code> loop, there should always be a
      space or line break after each semicolon, except for the special
      case of an infinite loop (although more infinite loops
      use <code>while</code>).  While not enforced, loop counters
      generally use post-increment.
    </p>
    <pre>
      for (i = 0 ;i &lt; limit ; ++i) { // Bad
      for (i = 0; i &lt; limit; i++) { // Good
      for (;;) { // ok
      while (1) { // Better
</pre>
    <p>
      Empty loop bodies are better represented with curly braces and a
      comment, although use of a semicolon is not currently rejected.
    </p>
    <pre>
      while ((rc = waitpid(pid, &amp;st, 0) == -1) &amp;&amp;
             errno == EINTR); // ok
      while ((rc = waitpid(pid, &amp;st, 0) == -1) &amp;&amp;
             errno == EINTR) { // Better
          /* nothing */
      }
</pre>

    <h2><a id="curly_braces">Curly braces</a></h2>

    <p>
      Omit the curly braces around an <code>if</code>, <code>while</code>,
      <code>for</code> etc. body only when both that body and the condition
      itself occupy a single line.  In every other case we require
      the braces.  This ensures that it is trivially easy to identify a
      single-<i>statement</i> loop: each has only one <i>line</i> in its body.
    </p>

<pre>
  while (expr)             // single line body; {} is forbidden
      single_line_stmt();
</pre>

<pre>
  while (expr(arg1,
              arg2))      // indentation makes it obvious it is single line,
      single_line_stmt(); // {} is optional (not enforced either way)
</pre>

<pre>
  while (expr1 &amp;&amp;
         expr2) {         // multi-line, at same indentation, {} required
      single_line_stmt();
  }
</pre>

    <p>
      However, the moment your loop/if/else body extends on to a second
      line, for whatever reason (even if it's just an added comment), then
      you should add braces.  Otherwise, it would be too easy to insert a
      statement just before that comment (without adding braces), thinking
      it is already a multi-statement loop:
    </p>

<pre>
  while (true) // BAD! multi-line body with no braces
      /* comment... */
      single_line_stmt();
</pre>
    <p>
      Do this instead:
    </p>
<pre>
  while (true) { // Always put braces around a multi-line body.
      /* comment... */
      single_line_stmt();
  }
</pre>
    <p>
      There is one exception: when the second body line is not at the same
      indentation level as the first body line:
    </p>
<pre>
  if (expr)
      die("a diagnostic that would make this line"
          " extend past the 80-column limit"));
</pre>

    <p>
      It is safe to omit the braces in the code above, since the
      further-indented second body line makes it obvious that this is still
      a single-statement body.
    </p>

    <p>
      To reiterate, don't do this:
    </p>

<pre>
  if (expr)            // BAD: no braces around...
      while (expr_2) { // ... a multi-line body
          ...
      }
</pre>

    <p>
      Do this, instead:
    </p>

<pre>
  if (expr) {
      while (expr_2) {
          ...
      }
  }
</pre>

    <p>
      However, there is one exception in the other direction, when even a
      one-line block should have braces.  That occurs when that one-line,
      brace-less block is an <code>if</code> or <code>else</code>
      block, and the counterpart block <b>does</b> use braces.  In
      that case, put braces around both blocks.  Also, if
      the <code>else</code> block is much shorter than
      the <code>if</code> block, consider negating the
      <code>if</code>-condition and swapping the bodies, putting the
      short block first and making the longer, multi-line block be the
      <code>else</code> block.
    </p>

<pre>
  if (expr) {
      ...
      ...
  }
  else
      x = y;    // BAD: braceless "else" with braced "then",
                // and short block last

  if (expr)
      x = y;    // BAD: braceless "if" with braced "else"
  else {
      ...
      ...
  }
</pre>

    <p>
      Keeping braces consistent and putting the short block first is
      preferred, especially when the multi-line body is more than a
      few lines long, because it is easier to read and grasp the semantics of
      an if-then-else block when the simpler block occurs first, rather than
      after the more involved block:
    </p>

<pre>
  if (!expr) {
    x = y; // putting the smaller block first is more readable
  } else {
      ...
      ...
  }
</pre>

    <p>
      But if negating a complex condition is too ugly, then at least
      add braces:
    </p>

<pre>
  if (complex expr not worth negating) {
      ...
      ...
  } else {
      x = y;
  }
</pre>

    <p>Use hanging braces for compound statements: the opening brace
      of a compound statement should be on the same line as the
      condition being tested.  Only top-level function bodies, nested
      scopes, and compound structure declarations should ever have {
      on a line by itself.
    </p>

<pre>
  void
  foo(int a, int b)
  {                          // correct - function body
      int 2d[][] = {
        {                    // correct - complex initialization
          1, 2,
        },
      };
      if (a)
      {                      // BAD: compound brace on its own line
          do_stuff();
      }
      {                      // correct - nested scope
          int tmp;
          if (a &lt; b) {       // correct - hanging brace
              tmp = b;
              b = a;
              a = tmp;
          }
      }
  }
</pre>

    <h2><a id="preprocessor">Preprocessor</a></h2>

    <p>Macros defined with an ALL_CAPS name should generally be
      assumed to be unsafe with regards to arguments with side-effects
      (that is, MAX(a++, b--) might increment a or decrement b too
      many or too few times).  Exceptions to this rule are explicitly
      documented for macros in viralloc.h and virstring.h.
    </p>

    <p>
      For variadic macros, stick with C99 syntax:
    </p>
<pre>
  #define vshPrint(_ctl, ...) fprintf(stdout, __VA_ARGS__)
</pre>

    <p>Use parenthesis when checking if a macro is defined, and use
    indentation to track nesting:
    </p>
<pre>
  #if defined(HAVE_POSIX_FALLOCATE) &amp;&amp; !defined(HAVE_FALLOCATE)
  # define fallocate(a, ignored, b, c) posix_fallocate(a, b, c)
  #endif
</pre>

    <h2><a id="types">C types</a></h2>

    <p>
      Use the right type.
    </p>

    <h3>Scalars</h3>

    <ul>
      <li>If you're using <code>int</code> or <code>long</code>, odds are
          good that there's a better type.</li>
      <li>If a variable is counting something, be sure to declare it with an
        unsigned type.</li>
      <li>If it's memory-size-related, use <code>size_t</code> (use
        <code>ssize_t</code> only if required).</li>
      <li>If it's file-size related, use uintmax_t, or maybe <code>off_t</code>.</li>
      <li>If it's file-offset related (i.e., signed), use <code>off_t</code>.</li>
      <li>If it's just counting small numbers use <code>unsigned int</code>;
        (on all but oddball embedded systems, you can assume that that
        type is at least four bytes wide).</li>
      <li>If a variable has boolean semantics, give it the <code>bool</code> type
        and use the corresponding <code>true</code> and <code>false</code> macros.
         It's ok to include &lt;stdbool.h&gt;, since libvirt's use of gnulib ensures
          that it exists and is usable.</li>
      <li>In the unusual event that you require a specific width, use a
        standard type like <code>int32_t</code>, <code>uint32_t</code>,
        <code>uint64_t</code>, etc.</li>
      <li>While using <code>bool</code> is good for readability, it comes with
          minor caveats:
        <ul>
          <li>Don't use <code>bool</code> in places where the type size must be constant across
            all systems, like public interfaces and on-the-wire protocols.  Note
            that it would be possible (albeit wasteful) to use <code>bool</code> in libvirt's
            logical wire protocol, since XDR maps that to its lower-level <code>bool_t</code>
            type, which <b>is</b> fixed-size.</li>
          <li>Don't compare a bool variable against the literal, <code>true</code>,
            since a value with a logical non-false value need not be <code>1</code>.
            I.e., don't write <code>if (seen == true) ...</code>.  Rather,
            write <code>if (seen)...</code>.</li>
        </ul>
      </li>
    </ul>

    <p>
      Of course, take all of the above with a grain of salt.  If you're about
      to use some system interface that requires a type like <code>size_t</code>,
      <code>pid_t</code> or <code>off_t</code>, use matching types for any
      corresponding variables.
    </p>

    <p>
      Also, if you try to use e.g., <code>unsigned int</code> as a type, and that
      conflicts with the signedness of a related variable, sometimes
      it's best just to use the <b>wrong</b> type, if <i>pulling the thread</i>
      and fixing all related variables would be too invasive.
    </p>

    <p>
      Finally, while using descriptive types is important, be careful not to
      go overboard.  If whatever you're doing causes warnings, or requires
      casts, then reconsider or ask for help.
    </p>

    <h3>Pointers</h3>

    <p>
      Ensure that all of your pointers are <i>const-correct</i>.
      Unless a pointer is used to modify the pointed-to storage,
      give it the <code>const</code> attribute.  That way, the reader knows
      up-front that this is a read-only pointer.  Perhaps more
      importantly, if we're diligent about this, when you see a non-const
      pointer, you're guaranteed that it is used to modify the storage
      it points to, or it is aliased to another pointer that is.
    </p>

    <h2><a id="memalloc">Low level memory management</a></h2>

    <p>
      Use of the malloc/free/realloc/calloc APIs is deprecated in the libvirt
      codebase, because they encourage a number of serious coding bugs and do
      not enable compile time verification of checks for NULL. Instead of these
      routines, use the macros from viralloc.h.
    </p>

    <ul>
      <li><p>To allocate a single object:</p>

<pre>
  virDomainPtr domain;

  if (VIR_ALLOC(domain) &lt; 0)
      return NULL;
</pre>
      </li>

      <li><p>To allocate an array of objects:</p>
<pre>
  virDomainPtr domains;
  size_t ndomains = 10;

  if (VIR_ALLOC_N(domains, ndomains) &lt; 0)
      return NULL;
</pre>
      </li>

      <li><p>To allocate an array of object pointers:</p>
<pre>
  virDomainPtr *domains;
  size_t ndomains = 10;

  if (VIR_ALLOC_N(domains, ndomains) &lt; 0)
      return NULL;
</pre>
      </li>

      <li><p>To re-allocate the array of domains to be 1 element
      longer (however, note that repeatedly expanding an array by 1
      scales quadratically, so this is recommended only for smaller
      arrays):</p>
<pre>
  virDomainPtr domains;
  size_t ndomains = 0;

  if (VIR_EXPAND_N(domains, ndomains, 1) &lt; 0)
      return NULL;
  domains[ndomains - 1] = domain;
</pre></li>

      <li><p>To ensure an array has room to hold at least one more
      element (this approach scales better, but requires tracking
      allocation separately from usage)</p>

<pre>
  virDomainPtr domains;
  size_t ndomains = 0;
  size_t ndomains_max = 0;

  if (VIR_RESIZE_N(domains, ndomains_max, ndomains, 1) &lt; 0)
      return NULL;
  domains[ndomains++] = domain;
</pre>
      </li>

      <li><p>To trim an array of domains from its allocated size down
      to the actual used size:</p>

<pre>
  virDomainPtr domains;
  size_t ndomains = x;
  size_t ndomains_max = y;

  VIR_SHRINK_N(domains, ndomains_max, ndomains_max - ndomains);
</pre></li>

      <li><p>To free an array of domains:</p>
<pre>
  virDomainPtr domains;
  size_t ndomains = x;
  size_t ndomains_max = y;
  size_t i;

  for (i = 0; i &lt; ndomains; i++)
      VIR_FREE(domains[i]);
  VIR_FREE(domains);
  ndomains_max = ndomains = 0;
</pre>
      </li>
    </ul>

    <h2><a id="file_handling">File handling</a></h2>

    <p>
      Usage of the <code>fdopen()</code>, <code>close()</code>, <code>fclose()</code>
      APIs is deprecated in libvirt code base to help avoiding double-closing of files
      or file descriptors, which is particularly dangerous in a multi-threaded
      application. Instead of these APIs, use the macros from virfile.h
    </p>

   <ul>
      <li><p>Open a file from a file descriptor:</p>

<pre>
  if ((file = VIR_FDOPEN(fd, "r")) == NULL) {
      virReportSystemError(errno, "%s",
                           _("failed to open file from file descriptor"));
      return -1;
  }
  /* fd is now invalid; only access the file using file variable */
</pre></li>

      <li><p>Close a file descriptor:</p>
<pre>
  if (VIR_CLOSE(fd) &lt; 0) {
      virReportSystemError(errno, "%s", _("failed to close file"));
  }
</pre></li>

      <li><p>Close a file:</p>

<pre>
  if (VIR_FCLOSE(file) &lt; 0) {
      virReportSystemError(errno, "%s", _("failed to close file"));
  }
</pre></li>

      <li><p>Close a file or file descriptor in an error path, without losing
             the previous <code>errno</code> value:</p>

<pre>
  VIR_FORCE_CLOSE(fd);
  VIR_FORCE_FCLOSE(file);
</pre>
      </li>
    </ul>

    <h2><a id="string_comparision">String comparisons</a></h2>

    <p>
      Do not use the strcmp, strncmp, etc functions directly. Instead use
      one of the following semantically named macros
    </p>

    <ul>
      <li><p>For strict equality:</p>
<pre>
  STREQ(a,b)
  STRNEQ(a,b)
</pre>
      </li>

      <li><p>For case insensitive equality:</p>
<pre>
  STRCASEEQ(a,b)
  STRCASENEQ(a,b)
</pre>
      </li>

      <li><p>For strict equality of a substring:</p>
<pre>
  STREQLEN(a,b,n)
  STRNEQLEN(a,b,n)
</pre>
      </li>

      <li><p>For case insensitive equality of a substring:</p>
<pre>
  STRCASEEQLEN(a,b,n)
  STRCASENEQLEN(a,b,n)
</pre>
      </li>

      <li><p>For strict equality of a prefix:</p>
<pre>
  STRPREFIX(a,b)
</pre>
      </li>
      <li><p>To avoid having to check if a or b are NULL:</p>
<pre>
  STREQ_NULLABLE(a, b)
  STRNEQ_NULLABLE(a, b)
</pre>
      </li>
    </ul>


    <h2><a id="string_copying">String copying</a></h2>

    <p>
      Do not use the strncpy function.  According to the man page, it
      does <b>not</b> guarantee a NULL-terminated buffer, which makes
      it extremely dangerous to use.  Instead, use one of the replacement
      functions provided by libvirt:
    </p>

<pre>
  virStrncpy(char *dest, const char *src, size_t n, size_t destbytes)
</pre>
    <p>
      The first two arguments have the same meaning as for strncpy,
      namely the destination and source of the copy operation.  Unlike
      strncpy, the function will always copy exactly the number of bytes
      requested and make sure the destination is NULL-terminated, as the
      source is required to be; sanity checks are performed to ensure the
      size of the destination, as specified by the last argument, is
      sufficient for the operation to succeed.  On success, 0 is returned;
      on failure, a value &lt;0 is returned instead.
    </p>

<pre>
  virStrcpy(char *dest, const char *src, size_t destbytes)
</pre>
    <p>
      Use this variant if you know you want to copy the entire src
      string into dest.
    </p>

<pre>
  virStrcpyStatic(char *dest, const char *src)
</pre>
    <p>
      Use this variant if you know you want to copy the entire src
      string into dest <b>and</b> you know that your destination string is
      a static string (i.e. that sizeof(dest) returns something
      meaningful).  Note that this is a macro, so arguments could be
      evaluated more than once.
    </p>

<pre>
  VIR_STRDUP(char *dst, const char *src);
  VIR_STRNDUP(char *dst, const char *src, size_t n);
</pre>
    <p>
      You should avoid using strdup or strndup directly as they do not report
      out-of-memory error, and do not allow a NULL source. Use
      VIR_STRDUP or VIR_STRNDUP macros instead, which return 0 for
      NULL source, 1 for successful copy, and -1 for allocation
      failure with the error already reported. In very
      specific cases, when you don't want to report the out-of-memory error, you
      can use VIR_STRDUP_QUIET or VIR_STRNDUP_QUIET, but such usage is very rare
      and usually considered a flaw.
    </p>

    <h2><a id="strbuf">Variable length string buffer</a></h2>

    <p>
      If there is a need for complex string concatenations, avoid using
      the usual sequence of malloc/strcpy/strcat/snprintf functions and
      make use of the virBuffer API described in virbuffer.h
    </p>

    <p>Typical usage is as follows:</p>

<pre>
  char *
  somefunction(...)
  {
     virBuffer buf = VIR_BUFFER_INITIALIZER;

     ...

     virBufferAddLit(&amp;buf, "&lt;domain&gt;\n");
     virBufferAsprintf(&amp;buf, "  &lt;memory&gt;%d&lt;/memory&gt;\n", memory);
     ...
     virBufferAddLit(&amp;buf, "&lt;/domain&gt;\n");

     ...

     if (virBufferCheckError(&amp;buf) &lt; 0)
         return NULL;

     return virBufferContentAndReset(&amp;buf);
  }
</pre>


    <h2><a id="includes">Include files</a></h2>

    <p>
      There are now quite a large number of include files, both libvirt
      internal and external, and system includes.  To manage all this
      complexity it's best to stick to the following general plan for all
      *.c source files:
    </p>

<pre>
  /*
   * Copyright notice
   * ....
   * ....
   * ....
   *
   */

  #include &lt;config.h&gt;             Must come first in every file.

  #include &lt;stdio.h&gt;              Any system includes you need.
  #include &lt;string.h&gt;
  #include &lt;limits.h&gt;

  #if WITH_NUMACTL                Some system includes aren't supported
  # include &lt;numa.h&gt;              everywhere so need these #if guards.
  #endif

  #include "internal.h"           Include this first, after system includes.

  #include "util.h"               Any libvirt internal header files.
  #include "buf.h"

  static int
  myInternalFunc()                The actual code.
  {
      ...
</pre>

    <p>
      Of particular note: <b>Do not</b> include libvirt/libvirt.h,
      libvirt/virterror.h, libvirt/libvirt-qemu.h, or libvirt/libvirt-lxc.h.
      They are included by "internal.h" already and there are some special reasons
      why you cannot include these files explicitly. One of the special cases,
      "libvirt/libvirt.h" is included prior to "internal.h" in "remote_protocol.x",
      to avoid exposing *_LAST enum elements.
    </p>


    <h2><a id="printf">Printf-style functions</a></h2>

    <p>
      Whenever you add a new printf-style function, i.e., one with a format
      string argument and following "..." in its prototype, be sure to use
      gcc's printf attribute directive in the prototype.  For example, here's
      the one for virAsprintf, in util.h:
    </p>

<pre>
  int virAsprintf(char **strp, const char *fmt, ...)
      ATTRIBUTE_FORMAT(printf, 2, 3);
</pre>

    <p>
      This makes it so gcc's -Wformat and -Wformat-security options can do
      their jobs and cross-check format strings with the number and types
      of arguments.
    </p>

    <p>
      When printing to a string, consider using virBuffer for
      incremental allocations, virAsprintf for a one-shot allocation,
      and snprintf for fixed-width buffers.  Do not use sprintf, even
      if you can prove the buffer won't overflow, since gnulib does
      not provide the same portability guarantees for sprintf as it
      does for snprintf.
    </p>

    <h2><a id="goto">Use of goto</a></h2>

    <p>
      The use of goto is not forbidden, and goto is widely used
      throughout libvirt.  While the uncontrolled use of goto will
      quickly lead to unmaintainable code, there is a place for it in
      well structured code where its use increases readability and
      maintainability.  In general, if goto is used for error
      recovery, it's likely to be ok, otherwise, be cautious or avoid
      it all together.
    </p>

    <p>
      The typical use of goto is to jump to cleanup code in the case
      of a long list of actions, any of which may fail and cause the
      entire operation to fail.  In this case, a function will have a
      single label at the end of the function.  It's almost always ok
      to use this style.  In particular, if the cleanup code only
      involves free'ing memory, then having multiple labels is
      overkill.  VIR_FREE() and every function named XXXFree() in
      libvirt is required to handle NULL as its arg.  Thus you can
      safely call free on all the variables even if they were not yet
      allocated (yes they have to have been initialized to NULL).
      This is much simpler and clearer than having multiple labels.
    </p>

    <p>
      There are a couple of signs that a particular use of goto is not
      ok:
    </p>

    <ul>
      <li>You're using multiple labels.  If you find yourself using
      multiple labels, you're strongly encouraged to rework your code
      to eliminate all but one of them.</li>
      <li>The goto jumps back up to a point above the current line of
      code being executed.  Please use some combination of looping
      constructs to re-execute code instead; it's almost certainly
      going to be more understandable by others.  One well-known
      exception to this rule is restarting an i/o operation following
      EINTR.</li>
      <li>The goto jumps down to an arbitrary place in the middle of a
      function followed by further potentially failing calls.  You
      should almost certainly be using a conditional and a block
      instead of a goto.  Perhaps some of your function's logic would
      be better pulled out into a helper function.</li>
    </ul>

    <p>
      Although libvirt does not encourage the Linux kernel wind/unwind
      style of multiple labels, there's a good general discussion of
      the issue archived at
      <a href="http://kerneltrap.org/node/553/2131">KernelTrap</a>
    </p>

    <p>
      When using goto, please use one of these standard labels if it
      makes sense:
    </p>

<pre>
      error: A path only taken upon return with an error code
    cleanup: A path taken upon return with success code + optional error
  no_memory: A path only taken upon return with an OOM error code
      retry: If needing to jump upwards (e.g., retry on EINTR)
</pre>

    <p>
    Top-level labels should be indented by one space (putting them on
    the beginning of the line confuses function context detection in git):
    </p>

<pre>
int foo()
{
    /* ... do stuff ... */
 cleanup:
    /* ... do other stuff ... */
}
</pre>


    <h2><a id="committers">Libvirt committer guidelines</a></h2>

    <p>
      The AUTHORS files indicates the list of people with commit access right
      who can actually merge the patches.
    </p>

    <p>
      The general rule for committing a patch is to make sure
      it has been reviewed
      properly in the mailing-list first, usually if a couple of people gave an
      ACK or +1 to a patch and nobody raised an objection on the list it should
      be good to go. If the patch touches a part of the code where you're not
      the main maintainer, or where you do not have a very clear idea of
      how things work, it's better
      to wait for a more authoritative feedback though. Before committing, please
      also rebuild locally, run 'make check syntax-check', and make sure you
      don't raise errors. Try to look for warnings too; for example,
      configure with
    </p>
<pre>
  --enable-compile-warnings=error
</pre>
    <p>
      which adds -Werror to compile flags, so no warnings get missed
    </p>

    <p>
      An exception to 'review and approval on the list first' is fixing failures
      to build:
    </p>
    <ul>
      <li>if a recently committed patch breaks compilation on a platform
        or for a given driver, then it's fine to commit a minimal fix
        directly without getting the review feedback first</li>
      <li>if make check or make syntax-check breaks, if there is
        an obvious fix, it's fine to commit immediately.
        The patch should still be sent to the list (or tell what the fix was if
        trivial), and 'make check syntax-check' should pass too, before committing
        anything</li>
      <li>
        fixes for documentation and code comments can be managed
        in the same way, but still make sure they get reviewed if non-trivial.
      </li>
    </ul>
    <h2><a id="coverage">Code coverage reports</a></h2>

    <p>
      Code coverage HTML reports can be generated with:
    </p>

<pre>
  make coverage
</pre>

    <p>
      Reports will be generated in the <code>cov/</code> directory. Point a
      web browser at <code>cov/index.html</code> for the full report.
    </p>

    <p>
      The <code>make coverage</code> target is provided by <code>gnulib</code>.
      It is a convenience helper for calling the following 3 targets in order.
      It may be useful to occasionally call these directly.

    <ul>
      <li><code>make init-coverage</code>: run <code>make clean</code> and
          remove all code coverage counter files (*.gcno, etc.)</li>
      <li><code>make build-coverage</code>: run <code>make</code> and
          <code>make check</code> with <code>CFLAGS</code> filled in with
          necessary coverage flags</li>
      <li><code>make gen-coverage</code>: generate the HTML report</li>
    </ul>
    </p>
  </body>
</html>