| blob | 732bde013f4f021a73b865ed77fc8caecd11c389 |
1 Relocatable binaries
2 ====================
4 |Gromacs| (mostly) implements the concept of relocatable binaries, i.e., that
5 after initial installation to ``CMAKE_INSTALL_PREFIX`` (or binary packaging
6 with CPack), the whole installation tree can be moved to a different folder and
7 |Gromacs| continues to work without further changes to the installation tree.
8 This page explains how this is implemented, and the known limitations in the
9 implementation. This information is mainly of interest to developers who need
10 to understand this or change the code, but it can also be useful for people
11 installing or packaging |Gromacs|.
13 A related feature that needs to be considered in all the code related to this
14 is that the executables should work directly when executed from the build tree,
15 before installation. In such a case, the data files should also be looked up
16 from the source tree to make development easy.
18 Finding shared libraries
19 ------------------------
21 If |Gromacs| is built with dynamic linking, the first part of making the
22 binaries relocatable is to make it possible for the executable to find
23 ``libgromacs``, no matter how it is executed. On platforms that support
24 a relative RPATH, this is used to make the |Gromacs| executables find the
25 ``libgromacs`` from the same installation prefix. This makes the executables
26 fully relocatable when it comes to linking, as long as the relative folder
27 structure between the executables and the library is kept the same.
29 If the RPATH mechanism does not work, ``GMXRC`` also adds the absolute path to
30 the ``libgromacs`` installed with it to ``LD_LIBRARY_PATH``. On platforms that
31 support this, this makes the linker search for the library here, but it is less
32 robust, e.g., when mixing calls to different versions of |Gromacs|. Note that
33 ``GMXRC`` is currently not relocatable, but hardcodes the absolute path.
35 On native Windows, DLLs are not fully supported; it is currently only possible
36 to compile a DLL with MinGW, not with Visual Studio or with Intel compilers.
37 In this case, the DLLs are placed in the ``bin/`` directory instead of
38 ``lib/`` (automatically by CMake, based on the generic binary type assignment
39 in ``CMakeLists.txt``). Windows automatically searches DLLs from the
40 executable directory, so the correct DLL should always be found.
42 For external libraries, standard CMake linking mechanisms are used and RPATH
43 for the external dependencies is included in the executable; on Windows,
44 dynamic linking may require extra effort to make the loader locate the correct
45 external libraries.
47 To support executing the built binaries from the build tree without
48 installation (critical for executing tests during development), standard CMake
49 mechanism is used: when the binaries are built, the RPATH is set to the build
50 tree, and during installation, the RPATH in the binaries is rewritten by CMake
51 to the final (relative) value. As an extra optimization, if the installation
52 tree has the same relative folder structure as the build tree, the final
53 relative RPATH is used already during the initial build.
55 The RPATH settings are in the root ``CMakeLists.txt``. It is possible to
56 disable the use of RPATH during installation with standard CMake variables,
57 such as setting ``CMAKE_SKIP_INSTALL_RPATH=ON``.
59 Finding data files
60 ------------------
62 The other, |Gromacs|-specific part, of making the binaries relocatable is
63 to make them able to find data files from the installation tree. Such data
64 files are used for multiple purposes, including showing the quotes at the end
65 of program execution. If the quote database is not found, the quotes are
66 simply not printed, but other files (mostly used by system preparation tools
67 like :ref:`gmx pdb2gmx` and :ref:`gmx grompp`, and by various analysis tools
68 for static data) will cause fatal errors if not found.
70 There are several considerations here:
72 * For relocation to work, finding the data files cannot rely on any hard-coded
73 absolute path, but it must find out the location of the executing code by
74 inspecting the system. As a fallback, environment variables or such set by
75 ``GMXRC`` or similar could be used (but currently are not).
76 * When running executables from the build tree, it is desirable that they will
77 automatically use the data files from the matching source tree to facilitate
78 easy testing. The data files are not copied into the build tree, and the
79 user is free to choose any relative locations for the source and build trees.
80 Also, the data files are not in the same relative path in the source tree and
81 in the installation tree (the source tree has ``share/top/``, the
82 installation tree ``share/gromacs/top/``; the latter is customizable during
83 CMake configuration).
84 * In addition to |Gromacs| executables, programs that link against
85 ``libgromacs`` need to be able to find the data files if they call certain
86 functions in the library. In this case, the executable may not be in the
87 same directory where |Gromacs| is. In case of static linking, no part of the
88 code is actually loaded from the |Gromacs| installation prefix, which makes
89 it impossible to find the data files without external information.
90 * The user can always use the ``GMXLIB`` environment variable to provide
91 alternative locations for the data files, but ideally this should never be
92 necessary for using the data files from the installation.
94 Not all the above considerations are fully addressed by the current
95 implementation, which works like this:
97 1. It finds the path to the current executable based on ``argv[0]``. If the
98 value contains a directory, this is interpreted as absolute or as relative
99 to the current working directory. If there is no directory, then a file by
100 that name is searched from the directories listed in ``PATH``. On Windows,
101 the current directory is also searched before ``PATH``. If a file with a
102 matching name is found, this is used without further checking.
103 2. If the executable is found and is a symbolic link, the symbolic links are
104 traversed until a real file is found. Note that links in the directory name
105 are not resolved, and if some of the links contain relative paths, the end
106 result may contain ``..`` components and such.
107 3. If an absolute path to the executable was found, the code checks whether the
108 executable is located in the build output directory (using ``stat()`` or
109 similar to account for possible symbolic links in the directory components).
110 If it is, then the hard-coded source tree location is returned.
111 4. If an absolute path to the executable was found and it was not in the build
112 tree, then all parent directories are checked. If a parent directory
113 contains :file:`share/{gromacs}/top/gurgle.dat`, this directory is returned
114 as the installation prefix. The file name ``gurgle.dat`` and the location
115 are considered unique enough to ensure that the correct directory has been
116 found. The installation directory for read-only architecture-independent
117 data files can be customized during CMake configuration by setting
118 ``CMAKE_INSTALL_DATADIR``, and the subdirectory under this that hosts the
119 GROMACS-specific data is set by ``GMX_INSTALL_DATASUBDIR``.
121 Note that this search does not resolve symbolic links or normalize the input
122 path beforehand: if there are ``..`` components *and* symbolic links in the
123 path, the search may proceed to unexpected directories, but this should not
124 be an issue as the correct installation prefix should be found before
125 encountering such symbolic links (as long as the ``bin/`` directory is not a
126 symbolic link).
127 5. If the data files have not been found yet, try a few hard-coded guesses
128 (like the original installation ``CMAKE_INSTALL_PREFIX`` and
129 ``/usr/local/``). The first guess that contains suitable files
130 (``gurgle.dat``) is returned.
131 6. If still nothing is found, return ``CMAKE_INSTALL_PREFIX`` and let the
132 subsequent data file opening fail.
134 The above logic to find the installation prefix is in
135 ``src/gromacs/commandline/cmdlineprogramcontext.cpp``. Note that code that
136 links to ``libgromacs`` can provide an alternative implementation for
137 ``gmx::IProgramContext`` for locating the data files, and is then fully
138 responsible of the above considerations.
140 Information about the used data directories is printed into the console output
141 (unless run with ``-quiet``), as well as to (some) error messages when locating
142 data files, to help diagnosing issues.
144 There is no mechanism to disable this probing search or affect the process
145 during compilation time, except for the CMake variables mentioned above.
147 Known issues
148 ------------
150 * ``GMXRC`` is not relocatable: it hardcodes the absolute installation path in
151 one assignment within the script, which no longer works after relocation.
152 Contributions to get rid of this on all the shells the ``GMXRC`` currently
153 supports are welcome.
154 * There is no version checking in the search for the data files; in case of
155 issues with the search, it may happen that the installation prefix from some
156 other installation of |Gromacs| is returned instead, and only cryptic errors
157 about missing or invalid files may reveal this.
158 * If the searching for the installation prefix is not successful, hard-coded
159 absolute guesses are used, and one of those returned. These guesses include
160 the absolute path in ``CMAKE_INSTALL_PREFIX`` used during compilation of
161 ``libgromacs``, which will be incorrect after relocation.
162 * The search for the installation prefix is based on the locating the
163 executable. This does not work for programs that link against
164 ``libgromacs``, but are not installed in the same prefix. For such cases,
165 the hard-coded guesses will be used, so the search will not find the correct
166 data files after relocation. The calling code can, however, programmatically
167 provide the |Gromacs| installation prefix, but ideally this would work
168 without offloading work to the calling code.
169 * One option to (partially) solve the two above issues would be to use the
170 ``GMXDATA`` environment variable set by ``GMXRC`` as the fallback (currently
171 this environment variable is set, but very rarely used).
172 * Installed ``pkg-config`` files are not relocatable: they hardcode the
173 absolute installation path.