Running and deploying packages

Executables and applications including shared libraries can also be distributed, deployed and run with Conan. This might have some advantages compared to deploying with other systems:

  • A unified development and distribution tool, for all systems and platforms.

  • Manage any number of different deployment configurations in the same way you manage them for development.

  • Use a Conan server remote to store all your applications and runtimes for all Operating Systems, platforms and targets.

There are different approaches:

Using virtual environments

We can create a package that contains an executable, for example from the default package template created by conan new:

$ conan new Hello/0.1

The source code used contains an executable called greet, but it is not packaged by default. Let’s modify the recipe package() method to also package the executable:

def package(self):
    self.copy("*greet*", src="bin", dst="bin", keep_path=False)

Now we create the package as usual, but if we try to run the executable it won’t be found:

$ conan create . user/testing
...
Hello/0.1@user/testing package(): Copied 1 '.h' files: hello.h
Hello/0.1@user/testing package(): Copied 1 '.exe' files: greet.exe
Hello/0.1@user/testing package(): Copied 1 '.lib' files: hello.lib

$ greet
> ... not found...

By default, Conan does not modify the environment, it will just create the package in the local cache, and that is not in the system PATH, so the greet executable is not found.

The virtualrunenv generator generates files that add the package’s default binary locations to the necessary paths:

  • It adds the dependencies lib subfolder to the DYLD_LIBRARY_PATH environment variable (for OSX shared libraries)

  • It adds the dependencies lib subfolder to the LD_LIBRARY_PATH environment variable (for Linux shared libraries)

  • It adds the dependencies bin subfolder to the PATH environment variable (for executables)

So if we install the package, specifying such virtualrunenv like:

$ conan install Hello/0.1@user/testing -g virtualrunenv

This will generate a few files that can be called to activate and deactivate the required environment variables

$ activate_run.sh # $ source activate_run.sh in Unix/Linux
$ greet
> Hello World!
$ deactivate_run.sh # $ source deactivate_run.sh in Unix/Linux

Imports

It is possible to define a custom conanfile (either .txt or .py), with an imports section, that can retrieve from local cache the desired files. This approach requires a user conanfile. For more details see example below runtime packages

Deployable packages

With the deploy() method, a package can specify which files and artifacts to copy to user space or to other locations in the system. Let’s modify the example recipe adding the deploy() method:

def deploy(self):
    self.copy("*", dst="bin", src="bin")

And run conan create

$ conan create . user/testing

With that method in our package recipe, it will copy the executable when installed directly:

$ conan install Hello/0.1@user/testing
...
> Hello/0.1@user/testing deploy(): Copied 1 '.exe' files: greet.exe
$ bin\greet.exe
> Hello World!

The deploy will create a deploy_manifest.txt file with the files that have been deployed.

Sometimes it is useful to adjust the package ID of the deployable package in order to deploy it regardless of the compiler it was compiled with:

def package_id(self):
    del self.info.settings.compiler

See also

Read more about the deploy() method.

Using the deploy generator

The deploy generator is used to have all the dependencies of an application copied into a single place. Then all the files can be repackaged into the distribution format of choice. For instance, if the application depends on boost, we may not know that it also requires many other 3rt-party libraries, such as zlib, bzip2, lzma, zstd, iconv, etc.

$ conan install . -g deploy

This helps to collect all the dependencies into a single place, moving them out of the Conan cache.

Using the json generator

A more advanced approach is to use the json generator: This generator works in a similar fashion as the deploy one, although it doesn’t copy the files to a directory. Instead, it generates a JSON file with all the information about the dependencies including the location of the files in the Conan cache.

$ conan install . -g json

The conanbuildinfo.json file produced is fully machine-readable and could be used by scripts to prepare the files and recreate the appropriate format for distribution. The following code shows how to read the library and binary directories from the conanbuildinfo.json:

import os
import json

data = json.load(open("conanbuildinfo.json"))

dep_lib_dirs = dict()
dep_bin_dirs = dict()

for dep in data["dependencies"]:
    root = dep["rootpath"]
    lib_paths = dep["lib_paths"]
    bin_paths = dep["bin_paths"]

    for lib_path in lib_paths:
        if os.listdir(lib_path):
            lib_dir = os.path.relpath(lib_path, root)
            dep_lib_dirs[lib_path] = lib_dir
    for bin_path in bin_paths:
        if os.listdir(bin_path):
            bin_dir = os.path.relpath(bin_path, root)
            dep_bin_dirs[bin_path] = bin_dir

While with the deploy generator all the files were copied into a folder, the advantage with the json one is that you have fine-grained control over the files and those can be directly copied to the desired layout. In that sense, the script above could be easily modified to apply some sort of filtering (e.g. to copy only shared libraries, and omit any static libraries or auxiliary files such as pkg-config .pc files).

Additionally, you could also write a simple startup script for your application with the extracted information like this:

executable = "MyApp"  # just an example
varname = "$APPDIR"

def _format_dirs(dirs):
    return ":".join(["%s/%s" % (varname, d) for d in dirs])

path = _format_dirs(set(dep_bin_dirs.values()))
ld_library_path = _format_dirs(set(dep_lib_dirs.values()))
exe = varname + "/" + executable

content = """#!/usr/bin/env bash
set -ex
export PATH=$PATH:{path}
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:{ld_library_path}
pushd $(dirname {exe})
$(basename {exe})
popd
""".format(path=path,
       ld_library_path=ld_library_path,
       exe=exe)

Note

The full example might be found on GitHub.

Running from packages

If a dependency has an executable that we want to run in the conanfile, it can be done directly in code using the run_environment=True argument. It internally uses a RunEnvironment() helper. For example, if we want to execute the greet app while building the Consumer package:

from conans import ConanFile, tools, RunEnvironment

class ConsumerConan(ConanFile):
    name = "Consumer"
    version = "0.1"
    settings = "os", "compiler", "build_type", "arch"
    requires = "Hello/0.1@user/testing"

    def build(self):
        self.run("greet", run_environment=True)

Now run conan install and conan build for this consumer recipe:

$ conan install . && conan build .
...
Project: Running build()
Hello World!

Instead of using the environment, it is also possible to explicitly access the path of the dependencies:

def build(self):
    path = os.path.join(self.deps_cpp_info["Hello"].rootpath, "bin")
    self.run("%s/greet" % path)

Note that this might not be enough if shared libraries exist. Using the run_environment=True helper above is a more complete solution.

Finally, there is another approach: the package containing the executable can add its bin folder directly to the PATH. In this case the Hello package conanfile would contain:

def package_info(self):
    self.cpp_info.libs = ["hello"]
    self.env_info.PATH = os.path.join(self.package_folder, "bin")

We may also define DYLD_LIBRARY_PATH and LD_LIBRARY_PATH if they are required for the executable.

The consumer package is simple, as the PATH environment variable contains the greet executable:

def build(self):
    self.run("greet")

Runtime packages and re-packaging

It is possible to create packages that contain only runtime binaries, getting rid of all build-time dependencies. If we want to create a package from the above “Hello” one, but only containing the executable (remember that the above package also contains a library, and the headers), we could do:

from conans import ConanFile

class HellorunConan(ConanFile):
    name = "HelloRun"
    version = "0.1"
    build_requires = "Hello/0.1@user/testing"
    keep_imports = True

    def imports(self):
        self.copy("greet*", src="bin", dst="bin")

    def package(self):
        self.copy("*")

This recipe has the following characteristics:

  • It includes the Hello/0.1@user/testing package as build_requires. That means that it will be used to build this HelloRun package, but once the HelloRun package is built, it will not be necessary to retrieve it.

  • It is using imports() to copy from the dependencies, in this case, the executable

  • It is using the keep_imports attribute to define that imported artifacts during the build() step (which is not define, then using the default empty one), are kept and not removed after build

  • The package() method packages the imported artifacts that will be created in the build folder.

To create and upload this package to a remote:

$ conan create . user/testing
$ conan upload HelloRun* --all -r=my-remote

Installing and running this package can be done using any of the methods presented above. For example:

$ conan install HelloRun/0.1@user/testing -g virtualrunenv
# You can specify the remote with -r=my-remote
# It will not install Hello/0.1@...
$ activate_run.sh # $ source activate_run.sh in Unix/Linux
$ greet
> Hello World!
$ deactivate_run.sh # $ source deactivate_run.sh in Unix/Linux

Deployment challenges

When deploying a C/C++ application there are some specific challenges that have to be solved when distributing your application. Here you will find the most usual ones and some recommendations to overcome them.

The C standard library

A common challenge for all the applications no matter if they are written in pure C or in C++ is the dependency on C standard library. The most wide-spread variant of this library is GNU C library or just glibc.

Glibc is not a just C standard library, as it provides:

  • C functions (like malloc(), sin(), etc.) for various language standards, including C99.

  • POSIX functions (like posix threads in the pthread library).

  • BSD functions (like BSD sockets).

  • Wrappers for OS-specific APIs (like Linux system calls)

Even if your application doesn’t use directly any of these functions, they are often used by other libraries, so, in practice, it’s almost always in actual use.

There are other implementations of the C standard library that present the same challenge, such as newlib or musl, used for embedded development.

To illustrate the problem, a simple hello-world application compiled in a modern Ubuntu distribution will give the following error when it is run in a Centos 6 one:

$ /hello
/hello: /lib64/libc.so.6: version `GLIBC_2.14' not found (required by /hello)

This is because the versions of the glibc are different between those Linux distributions.

There are several solutions to this problem:

Some people also advice to use static of glibc, but it’s strongly discouraged. One of the reasons is that newer glibc might be using syscalls that are not available in the previous versions, so it will randomly fail in runtime, which is much harder to debug (the situation about system calls is described below).

It’s possible to model either glibc version or Linux distribution name in Conan by defining custom Conan sub-setting in the settings.yml file (check out sections Adding new settings and Adding new sub-settings). The process will be similar to:

  • Define new sub-setting, for instance os.distro, as explained in the section Adding new sub-settings.

  • Define compatibility mode, as explained by sections package_id() and build_id() (e.g. you may consider some Ubuntu and Debian packages to be compatible with each other)

  • Generate different packages for each distribution.

  • Generate deployable artifacts for each distribution.

C++ standard library

Usually, the default C++ standard library is libstdc++, but libc++ and stlport are other well-known implementations.

Similarly to the standard C library glibc, running the application linked with libstdc++ in the older system may result in an error:

$ /hello
/hello: /usr/lib64/libstdc++.so.6: version `GLIBCXX_3.4.21' not found (required by /hello)
/hello: /usr/lib64/libstdc++.so.6: version `GLIBCXX_3.4.26' not found (required by /hello)

Fortunately, this is much easier to address by just adding -static-libstdc++ compiler flag. Unlike C runtime, C++ runtime can be linked statically safely, because it doesn’t use system calls directly, but instead relies on libc to provide required wrappers.

Compiler runtime

Besides C and C++ runtime libraries, the compiler runtime libraries are also used by applications. Those libraries usually provide lower-level functions, such as compiler intrinsics or support for exception handling. Functions from these runtime libraries are rarely referenced directly in code and are mostly implicitly inserted by the compiler itself.

$ ldd ./a.out
libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007f6626aee000)

you can avoid this kind of dependency by the using of the -static-libgcc compiler flag. However, it’s not always sane thing to do, as there are certain situations when applications should use shared runtime. The most common is when the application wishes to throw and catch exceptions across different shared libraries. Check out the GCC manual for the detailed information.

System API (system calls)

New system calls are often introduced with new releases of Linux kernel. If the application, or 3rd-party libraries, want to take advantage of these new features, they sometimes directly refer to such system calls (instead of using wrappers provided by glibc).

As a result, if the application was compiled on a machine with a newer kernel and build system used to auto-detect available system calls, it may fail to execute properly on machines with older kernels.

The solution is to either use a build machine with lowest supported kernel, or model supported operation system (just like in case of glibc). Check out sections Adding new settings and Adding new sub-settings to get a piece of information on how to model distribution in conan settings.