I am working on a project and we are starting release binaries. We are using CMake to generate build files, and CPack to create binaries. Our binaries work, but we run into problem with shared objects. Essentially, many of the issues, arise from symbolic linking on the system, especially with shared objects with multiple links. So, let say the RPATH results from ldd (or otool) for some executable include libmpich.so.10 and I've linked /usr/lib/x86_64-linux-gnu/libmpich.so from cmake and these files are related like this:
/usr/lib/x86_64-linux-gnu/libmpich.so -> libmpich.so.10
/usr/lib/x86_64-linux-gnu/libmpich.so.10 -> libmpich.so.10.0.4
/usr/lib/x86_64-linux-gnu/libmpich.so.10.0.4
Now, for some reason the RPATH uses the intermediate link (so.10) but readlink on libmpich.so (or get_filename_component(... REALPATH)) returns libmpich.so.10.0.4. So if I install libmpich.so.10.0.4 under the name libmpich.so OR libmpich.so.10.0.4 (or create the symlink from one to the other), I've still missed the library asked for in the RPATH.
I've been playing whack-a-mole on when dealing with these and/or using a file glob to try to grab the intermediate link but I would like to do something more robust. Anyone use a good design pattern for this?
I have been looking into using functions like GET_PREREQUISITES, but those require the object to be built so I would need to add them into the install scripts somehow.... and it feels like their should be a better way.
-Jameson
P.S. I've also bee looking for a best practices guide for building binaries, either with cmake or in general. We are producing binaries on windows, linux and mac. If you know of some good links please post them as well.
I just recently dealt with this issue myself. The cmake command get_filename_component(... REALPATH) resovles ALL levels of a symlink in one call.
To resolve just a single level of a symlink you can call 'readlink' directly from cmake, since it's available on every symlink-enabled platform that you're likely to build on (Linux, Mac OS X, and *BSD).
So, if you want to reproduce the complete chain of symbolic links,
you'd code up something like this in your cmake script:
#If given the following library path:
set(lib "/usr/lib/x86_64-linux-gnu/libmpich.so")
#Make sure the initial path is absolute.
get_filename_component(lib "${lib}" ABSOLUTE)
#Store initial path as first element in list.
set(symlist "${lib}")
while(UNIX AND IS_SYMLINK "${lib}")
#Grab path to directory containing the current symlink.
get_filename_component(sym_path "${lib}" DIRECTORY)
#Resolve one level of symlink, store resolved path back in lib.
execute_process(COMMAND readlink "${lib}"
RESULT_VARIABLE errMsg
OUTPUT_VARIABLE lib
OUTPUT_STRIP_TRAILING_WHITESPACE)
#Check to make sure readlink executed correctly.
if(errMsg AND (NOT "${errMsg}" EQUAL "0"))
message(FATAL_ERROR "Error calling readlink on library.")
endif()
#Convert resolved path to an absolute path, if it isn't one already.
if(NOT IS_ABSOLUTE "${lib}")
set(lib "${sym_path}/${lib}")
endif()
#Append resolved path to symlink resolution list.
list(APPEND symlist "${lib}")
endwhile()
#Now symlist will contain the following:
# [...]/libmpich.so;[...]/libmpich.so.10;[...]/libmpich.so.10.0.4
Related
So I am working on a project that is intended to run on a remote server. I develop the program on a local pc, compile it, then upload it to the remote server. Both the local pc and the remote server are run on CentOS 7.7.
The program is developed using the CLion IDE, configured with CMake. The program depends a few shared libraries, which are supposed to link to the executable according to what I wrote in CMake. At my local PC, I can compile and run the program perfectly. However, after I scp the whole directory of the project to the remote server, the executable fails to run. It cannot find any of the ".so" files, according to what ldd says.
This is my CMakeList.txt, with every path being relative path, instead of absolute path.
cmake_minimum_required(VERSION 3.15)
project(YS_Test)
set(CMAKE_CXX_STANDARD 11)
set(SOURCE_PATH_ src)
file(GLOB SOURCE_FILES_ ${SOURCE_PATH_}/*.*)
set(PROJECT_LIBS_ libTapQuoteAPI.so libTapTradeAPI.so libTapDataCollectAPI.so)
include_directories(api/include)
link_directories(api/lib/linux)
add_executable(YS_Test ${SOURCE_FILES_})
target_link_libraries(YS_Test ${PROJECT_LIBS_})
Please do not tell me to set LD_LIBRARY_PATH to fix my issue. The program worked fine on my local pc without LD_LIBRARY_PATH, so I expect it to run on the remote server without LD_LIBRARY_PATH. I would like to know what is really going on here, instead of a work around. Thanks!
If I understand your problem correctly, you want to ship your compiled YS_Test program along with some dependencies and have it run on a remote server. By default an executable will only look in the directories configured in /etc/ld.so, which will not include the deploy path.
Note: Typically you do not deploy your entire build directory but only the compiled artifacts and dependencies. For this answer I will assume you deploy the binary and its dependencies to the same directory.
You have two options:
Require users of your program to set LD_LIBRARY_PATH, either by themselves or by a wrapper script. This variable will instruct the dynamic linker to look in the specified directories as well. Even if you do not like this solution, it is by far the most common approach.
Add -Wl,-rpath='$ORIGIN' to your linker options. This will add a DT_RUNPATH attribute to the executable's dynamic section. As you are using CMake you can also set this using BUILD_RPATH and/or INSTALL_RPATH target properties.
The ld.so manpage describes this attribute as follows:
If a shared object dependency does not contain a slash, then it is
searched for in the following order:
...
Using the directories specified in the DT_RUNPATH dynamic section
attribute of the binary if present.
The $ORIGIN part expands to the directory containing the program or shared
object.
If you really insist on shipping your build directory (eg during development), you can take a look at the CMake BUILD_RPATH_USE_ORIGIN property (and its usual global counterpart CMAKE_BUILD_RPATH_USE_ORIGIN), this will embed relative paths into binaries instead of absolute paths.
As you don't want a workaround (#Botje has given you two already), I will try an explanation instead. In your development machine, if you use this command:
ldd YS_Test
You will see all the shared libraries used by your program, with their corresponding paths. The libTapQuoteAPI.so libTapTradeAPI.so libTapDataCollectAPI.so are found at your 'api/lib/linux' directory, but resolved with full absolute paths. If you do the same at your server, some shared objects can't be resolved because they aren't at the same location.
If you use one of these commands (not sure which are available in Centos):
chrpath --list YS_Test
or
patchelf --print-rpath YS_Test
You will see the RPATH or RUNPATH tags embedded in your program. This is the path used by the Linux linker to locate dependencies that are outside the standard ld locations. You may find extended explanations on Internet about this, like this one or the Wikipedia article.
Breaking my promise, I give you a third workaround: use patchelf or chrpath at your server after scp to change the embedded RPATH tag, pointing it relative to $ORIGIN (which represents the program location).
I have SCons code in which I am using SConscripts to build different directories separately. In my Src directory, my SConscript builds a shared library, and then returns the resulting Node as the Python variable libMyLibrary. I typically use the install option to copy this library to a directory that is on my system's LD_LIBRARY_PATH (I'm using OpenSUSE).
So far, so good. Now, in another directory, Src/Test, another SConscript imports libMyLibrary and builds some Programs using code like this:
env.Program('myProgram', 'myProgram.cpp', LIBS=[env['LIBS'], libMyLibrary])
The program then gets installed to my local bin folder. This code does track the library dependency and build the program, but the problem is that since the library is in a sub-directory (Src), that sub-directory gets included in the linker command. Here is an abbreviated example of the linker command that SCons generates:
g++ -o Src/Test/myProgram Src/Test/myProgram.o Src/libMyLibrary.so
I believe this happens because the Node,libMyLibrary, is essentially a path. The problem is that when I try to run the program, it is not looking for libMyLibrary.so in my library folder, but rather Src/libMyLibrary.so, and of course it doesn't find it.
I do NOT want the libraries I build to be installed in sub-directories of my install folder.
I already add the Src folder to LIBPATH, so SCons adds the -LSrc option to the linker command, but that doesn't solve the problem. My preference would be that when I add a Node, the path should automatically get parsed out to add the appropriate -L and -l options.
I know that I can get around this problem by adding the string 'MyLibrary' to the LIBS variable instead of the libMyLibrary Node, but then I have to explicitly tell SCons that each Program Depends() on libMyLibrary. It seems very inefficient to short-circuit SCons's built-in dependency tracking this way. Does anyone know the correct, SCons-y way to do this?
I'm referring to your latest comment: It looks to me as if this is not really a SCons problem, but more a general linker question (XY problem). Are you perhaps simply searching for RPATH? Please also check this old SO question: scons executable + shared library in project directory
I'm zipping a pre-built (no source/object files) binary application for distribution. The binary application requires a couple of libraries not included by default. The only way I seem to be able to get the application to start on the end-user is by including a run.sh that sets the library path to the current directory:
export LD_LIBRARY_PATH=./:$LD_LIBRARY_PATH
./MyApp.out
However, I'd really like to allow the user to just unzip the zip and doubleclick MyApp.out (without the shell script). Can I edit MyApp.out to search the current directory for the library? I've done something similar on OSX using install_name_tool, but that tool isn't available here.
You want to set the rpath. See this answer. So link using
gcc yourobjects*.o -L/some/lib/dir/ -lsome -Wl,-rpath,.
But you might want even to use -Wl,-rpath,$PWD or perhaps -Wl,-rpath,'$ORIGIN'. See this.
You could also (and this should work for a pre-built executable) configure your /etc/ld.so.conf by adding a line there with an absolute path (of the directory containing the lib), then running ldconfig -v ... See ldconfig(8)
I would suggest adding /usr/local/lib into /etc/ld.so.conf and making a symlink from /usr/local/lib/libfoo.so to e.g. $HOME/libfoo.so etc... (then run ldconfig ...). I don't think adding a user specific directory to /etc/ld.so.conf is reasonable ...
PS. What you really want is to package your application (e.g. as a *.deb package for Debian or Ubuntu, or an *.rpm for Fedora or Redhat). Package management systems handle dependencies!
In short: This question is basically about telling Linux to load the development version of the .so file for executables in the dev directory and the installed .so file for others.
In long: Imagine a shared library, let's call it libasdf.so. And imagine the following directories:
/home/user/asdf/lib: libasdf.so
/home/user/asdf/test: ... perform_test
/opt/asdf/lib: libasdf.so
/home/user/jkl: ... use_asdf
In other words, you have a development directory for your library (/home/user/asdf) and you have an installed copy of its previous stable version (/opt/asdf) and some other programs using it (/home/user/jkl).
My question is, how can I tell Linux, to load /home/user/asdf/lib/libasdf.so when executing /home/user/asdf/test/perform_test and to load /opt/asdf/lib/libasdf.so when executing /home/user/jkl/use_asdf? Note that, even though I specify the directory by -L during link, Linux uses other methods (for example /ect/ld.so.conf and $LD_LIBRARY_PATH) to find the .so file.
The reason I need such a thing is that, of course the executables in the development directory need to link with the latest version of the library, while the other programs, would want to use the stable version.
Putting ../lib in the library path doesn't seem like a secure idea, not to mention not completely correct since you can't run the test from a different directory.
One solution I thought about is to have perform_test link with libasdf-dev.so and upon install, copy libasdf-dev.so as libasdf.so and have others link with that. This solution has one problem though. Imagine the following additional directory:
/home/user/asdf/tool: ... use_asdf_too
Which gets installed to:
/opt/asdf/bin: use_asdf_too
In my solution, it is unknown what use_asdf_too should be linked against. If linked against libasdf.so, it wouldn't work properly if invoked from the dev directory and if linked against libasdf-dev.so, it wouldn't work properly if invoked from the installed location.
What can I do? How is this managed by other people?
Installed shared objects usually don't just end with ".so". Usually they also include their soname, such as libadsf.so.42.1. The .so file for development is typically a symlink to a fully-versioned filename. The linker will look for the .so file and resolve it to the full filename, and the loader will then load the fully-versioned library instead.
I have an application that relies on Qt, GDCM, and VTK, with the main build environment being Qt. All of these libraries are cross-platform and compile on Windows, Mac, and Linux. I need to deploy the application to Linux after deploying on Windows. The versions of vtk and gdcm I'm using are trunk versions from git (about a month old), more recent than what I can get apt-get on Ubuntu 11.04, which is my current (and only) Linux deployment target.
What is the accepted method for deploying an application that relies on these kinds of libraries?
Should I be statically linking here, to avoid LD_LIBRARY_PATH? I see conflicting reports on LD_LIBRARY_PATH; tutorials like this one suggest that it's the 'right way' to modify the library path to use shared libraries through system reboots. Others suggest that I should never set LD_LIBRARY_PATH. In the default version of GDCM, the installation already puts libraries into the /usr/local/lib directory, so those libraries get seen when I run ldd <my program>. VTK, on the other hand, puts its libraries into /usr/local/lib/vtk-5.9, which is not part of the LD_LIBRARY_PATH on most user's machines, and so is not found unless some change is made to the system. Copying the VTK files into '/usr/local/lib' does not allow 'ldd' to see the files.
So, how can I make my application see VTK to use the libraries?
On windows, deploying the dlls is very straightforward, because I can just include them in the installer, and the application finds them because they are in the local directory. That approach does not work in Linux, so I was going to have the users install Qt, GDCM, and VTK from whatever appropriate source and use the default locations, and then have the application point to those default locations. However, since VTK is putting things into a non-standard location, should I also expect users to modify LD_LIBRARY_PATH? Should I include the specific versions of the libraries that I want and then figure out how to make the executable look in the local directory for those libraries and ignore the ones it finds in the library path?
Every "serious" commercial application I have ever seen uses LD_LIBRARY_PATH. They invariably include a shell script that looks something like this:
#!/bin/sh
here="${0%/*}" # or you can use `dirname "$0"`
LD_LIBRARY_PATH="$here"/lib:"$LD_LIBRARY_PATH"
export LD_LIBRARY_PATH
exec "$0".bin "$#"
They name this script something like .wrapper and create a directory tree that looks like this:
.wrapper
lib/ (directory full of .so files)
app1 -> .wrapper (symlink)
app1.bin (executable)
app2 -> .wrapper (symlink)
app2.bin (executable)
Now you can copy this whole tree to wherever you want, and you can run "/path/to/tree/app1" or "/path/to/tree/app2 --with --some --arguments" and it will work. So will putting /path/to/tree in your PATH.
Incidentally, this is also how Firefox and Chrome do it, more or less.
Whoever told you not to use LD_LIBRARY_PATH is full of it, IMHO.
Which system libraries you want to put in lib depends on which Linux versions you want to officially support.
Do not even think about static linking. The glibc developers do not like it, they do not care about supporting it, and they somehow manage to break it a little harder with every release.
Good luck.
In general, you're best off depending on the 'normal' versions of the libraries for whatever distribution you're targetting (and saying you don't support dists that don't support recent enough versions of the lib), but if you REALLY need to depend on a bleeding edge version of some shared lib, you can link your app with -Wl,-rpath,'$ORIGIN' and then install a copy of the exact version you want in the same directory as your executable.
Note that if you use make, you'll need $$ in the makefile to get a single $ into the argument that is actually sent to the linker. The single qutoes are needed so the shell doesn't munge things...
Well, there are two options for deploying Linux application.
The correct way:
make a package for your app and for the libraries, if they are so special, that they can't be installed from standard repositories
There are two major package formats. RPM and DEB.
The easy way:
make a self-extract file that will install the "windows way" into /opt.
You can have libraries in the same directory as the executable, it's just not the preferred way.