The build system

We are currently migrating from autotools to CMake as a build-system. This document currently describes how we intend to perform this migration, and will be updated after the migration to explain how the new CMake configuration works.

Stages during the build

  1. The netdata-installer.sh, take in arguments and environment settings to control the build.
  2. The configure step: autoreconf -ivf ; ./configure passing arguments into the configure script. This becomes generation-time in CMake. This includes package / system detection and configuration resulting in the config.h in the source root.
  3. The build step: recurse through the generated Makefiles and build the executable.
  4. The first install step: calls make install to handle all the install steps put into the Makefiles by the configure step (puts binaries / libraries / config into target tree structure).
  5. The second install step: the rest of the installer after the make install handles system-level configuration (privilege setting, user / groups, fetch/build/install go.d plugins, telemetry, installing service for startup, uninstaller, auto-updates.

The ideal migration result is to replace all of this with the following steps:

mkdir build ; cd build ; cmake .. -D... ; cmake --build . --target install

The -D... indicates where the command-line arguments for configuration are passed into CMake.

CMake generation time

At generation time we need to solve the following issues:

Feature flags

Every command-line switch on the installer and the configure script needs to becomes an argument to the CMake generation, we can do this with variables in the CMake cache:

CMakeLists.txt:

option(ENABLE_DBENGINE "Enable the dbengine storage" ON)
...
if(${ENABLE_DBENGINE})
...
endif()

Command-line interface

cmake -DENABLE_DBENGINE

Dependency detection

We have a mixture of soft- and hard-depedencies on libraries. For most of these we expect pkg-config information, for some we manually probe for libraries and include files. We should treat all of the external dependencies consistently:

  1. Default to autodetect using pkg-config (e.g. the standard jemalloc drops a .pc into the system but we do not check for it.
  2. If no .pc is found perform a manual search for libraries under known names, and check for accessible symbols inside them.
  3. Check that include paths work.
  4. Allow a command-line override (e.g. -DWITH_JEMALLOC=/...).
  5. If none of the above work then fail the install if the dependency is hard, otherwise indicate it is not present in the config.h.

Before doing any dependency detection we need to determine which search paths are really in use for the current compiler, after the project declaration we can use:

execute_process(COMMAND ${CMAKE_C_COMPILER} "--print-search-dirs"
COMMAND grep "^libraries:"
COMMAND sed "s/^libraries: =//"
COMMAND tr ":" " "
COMMAND tr -d "\n"
OUTPUT_VARIABLE CC_SEARCH_DIRS
RESULTS_VARIABLE CC_SEARCH_RES)
string(REGEX MATCH "^[0-9]+" CC_SEARCH_RES ${CC_SEARCH_RES})
#string(STRIP "${CC_SEARCH_RES}" CC_SEARCH_RES)
if(0 LESS ${CC_SEARCH_RES})
message(STATUS "Warning - cannot determine standard compiler library paths")
# Note: we will probably need a different method for Windows...
endif()

The output format for this switch works on both Clang and gcc, it also includes the include search path, which can be extracted in a similar way. Standard advice here is to list the ldconfig cache or use the -V flag to check, but this does not work consistently across platforms - in particular gcc will reconfigure ld when it is called to gcc's internal view of search paths. During experiments each of these alternative missed / added unused paths. Dumping the compiler's own estimate of the search paths seems to work consistently across clang/gcc/linux/freebsd configurations.

The default behaviour in CMake is to search across predefined paths (e.g. CMAKE_LIBRARY_PATH) that are based on heuristics about the current platform. Most projects using CMake seem to overwrite this with their own estimates.

We can use the extracted paths as a base, add our own heuristics based on OS and then set(CMAKE_LIBRARY_PATH ${OUR_OWN_LIB_SEARCH}) to get the best results. Roughly we do the following for each external dependency:

set(WITH_JSONC "Detect" CACHE STRING "Manually set the path to a json-c installation")
...
if(${WITH_JSONC} STREQUAL "Detect")
pkg_check_modules(JSONC json-c) # Don't set the REQUIRED flag
if(JSONC_FOUND)
message(STATUS "libjsonc found through .pc -> ${JSONC_CFLAGS_OTHER} ${JSONC_LIBRARIES}")
# ... setup using JSONC_CFLAGS_OTHER JSONC_LIBRARIES and JSONC_INCLUDE_DIRS
else()
find_library(LIB_JSONC
NAMES json-c libjson-c
PATHS ${CMAKE_LIBRARY_PATH}) # Includes our additions by this point
if(${LIB_JSONC} STREQUAL "LIB_JSONC-NOTFOUND")
message(STATUS "Library json-c not installed, disabling")
else()
check_library_exists(${LIB_JSONC} json_object_get_type "" HAVE_JSONC)
# ... setup using heuristics for CFLAGS and check include files are available
endif()
endif()
else()
# ... use explicit path as base to check for library and includes ...
endif()

For checking the include path we have two options, if we overwrite the CMAKE_... variables to change the internal search path we can use:

CHECK_INCLUDE_FILE(json/json.h HAVE_JSONC_H)

Or we can build a custom search path and then use:

find_file(HAVE_JSONC_H json/json.h PATHS ${OUR_INCLUDE_PATHS})

Note: we may have cases where there is no .pc but we have access to a .cmake (e.g. AWS SDK, mongodb,cmocka) - these need to be checked / pulled inside the repo while building a prototype.

Compiler compatibility checks

In CMakeLists.txt:

CHECK_INCLUDE_FILE(sys/prctl.h HAVE_PRCTL_H)
configure_file(cmake/config.in config.h)

In cmake/config.in:

#cmakedefine HAVE_PRCTL_H 1

If we want to check explicitly if something compiles (e.g. the accept4 check, or the strerror_r typing issue) then we set the CMAKE_... paths and then use:

check_c_source_compiles(
"
#include <string.h>
int main() { char x = *strerror_r(0, &x, sizeof(x)); return 0; }
"
STRERROR_R_CHAR_P)

This produces a bool that we can use inside CMake or propagate into the config.h.

We can handle the atomic checks with:

check_c_source_compiles(
"
int main (int argc, char **argv)
{
volatile unsigned long ul1 = 1, ul2 = 0, ul3 = 2;
__atomic_load_n(&ul1, __ATOMIC_SEQ_CST);
__atomic_compare_exchange(&ul1, &ul2, &ul3, 1, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
__atomic_fetch_add(&ul1, 1, __ATOMIC_SEQ_CST);
__atomic_fetch_sub(&ul3, 1, __ATOMIC_SEQ_CST);
__atomic_or_fetch(&ul1, ul2, __ATOMIC_SEQ_CST);
__atomic_and_fetch(&ul1, ul2, __ATOMIC_SEQ_CST);
volatile unsigned long long ull1 = 1, ull2 = 0, ull3 = 2;
__atomic_load_n(&ull1, __ATOMIC_SEQ_CST);
__atomic_compare_exchange(&ull1, &ull2, &ull3, 1, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
__atomic_fetch_add(&ull1, 1, __ATOMIC_SEQ_CST);
__atomic_fetch_sub(&ull3, 1, __ATOMIC_SEQ_CST);
__atomic_or_fetch(&ull1, ull2, __ATOMIC_SEQ_CST);
__atomic_and_fetch(&ull1, ull2, __ATOMIC_SEQ_CST);
return 0;
}
"
HAVE_C__ATOMIC)

For the specific problem of getting the correct type signature in log.c for the strerror_r calls we can replicate what we have now, or we can delete this code completely and use a better solution that is documented here. To replicate what we have now:

check_c_source_compiles(
"
#include <string.h>
int main() { char x = *strerror_r(0, &x, sizeof(x)); return 0; }
"
STRERROR_R_CHAR_P)
check_c_source_compiles(
"
#include <string.h>
int main() { int x = strerror_r(0, &x, sizeof(x)); return 0; }
"
STRERROR_R_INT)
if("${STRERROR_R_CHAR_P}" OR "${STRERROR_R_INT}")
set(HAVE_DECL_STRERROR_R 1)
endif()
message(STATUS "Result was ${HAVE_DECL_STRERROR_R}")

Note: I did not find an explicit way to select compiler when both clang and gcc are present. We might have an implicit way (like redirecting cc) but we should put one in.

Debugging problems in test compilations

Test compilations attempt to feed a test-input into the targeted compiler and result in a yes/no decision, this is similar to AC_LANG_SOURCE(.... if test $ac_... in .m4. We have two techniques to use in CMake:

cmake_minimum_required(VERSION 3.1.0)
include(CheckCCompilerFlag)
project(empty C)
check_c_source_compiles(
"
#include <string.h>
int main() { char x = *strerror_r(0, &x, sizeof(x)); return 0; }
"
STRERROR_R_CHAR_P)
try_compile(HAVE_JEMALLOC ${CMAKE_CURRENT_BINARY_DIR}
${CMAKE_CURRENT_SOURCE_DIR}/quickdemo.c
LINK_LIBRARIES jemalloc)

The check_c_source_compiles is light-weight:

  • Inline source for the test, easy to follow.
  • Build errors are reported in CMakeFiles/CMakeErrors.log

But we cannot alter the include-paths / library-paths / compiler-flags specifically for the test without overwriting the current CMake settings. The alternative approach is slightly more heavy-weight:

  • Can't inline source for try_compile - it requires a .c file in the tree.
  • Build errors are not shown, the recovery process for them is somewhat difficult.
rm -rf * && cmake .. --debug-trycompile
grep jemal CMakeFiles/CMakeTmp/CMakeFiles/*dir/*
cd CMakeFiles/CMakeTmp/CMakeFiles/cmTC_d6f0e.dir # for example
cmake --build ../..

This implies that we can do this to diagnose problems / develop test-programs, but we have to make them bullet-proof as we cannot expose this to end-users. This means that the results of the compilation must be crisp - exactly yes/no if the feature we are testing is supported.

System configuration checks

For any system configuration checks that fall outside of the above scope (includes, libraries, packages, test-compilation checks) we have a fall-back that we can use to glue any holes that we need, e.g. to pull out the packaging strings, inside the CMakeLists.h:

execute_process(COMMAND cat ${CMAKE_CURRENT_SOURCE_DIR}/packaging/version
COMMAND tr -d '\n'
OUTPUT_VARIABLE VERSION_FROM_FILE)
message(STATUS "Packaging version ${VERSION_FROM_FILE}")

and this in the config.h.in:

#define VERSION_FROM_FILE "@VERSION_FROM_FILE@"

CMake build time

We have a working definition of the targets that is in use with CLion and works on modern CMake (3.15). It breaks on older CMake version (e.g. 3.7) with an error message (issue#7091). No PoC yet to fix this, but it looks like changing the target properties should do it (in the worst case we can drop the separate object completely and merge the sources directly into the final target).

Steps needed for building a prototype:

  1. Pick a reasonable configuration.
  2. Use the PoC techniques above to do a full generation of CMAKE_ variables in the cache according to the feature options and dependencies.
  3. Push these into the project variables.
  4. Work on it until the build succeeds in at least one known configuration.
  5. Smoke-test that the output is valid (i.e. the executable loads and runs, and we can access the dashboard).
  6. Do a full comparison of the config.h generated by autotools against the CMake version and document / fix any deviations.

CMake install target

I've only looked at this superficially as we do not have a prototype yet, but each of the first-stage install steps (in make install) and the second-stage (in netdata-installer.sh) look feasible.

General issues

  • We need to choose a minimum CMake version that is an available package across all of our supported environments. There is currently a build issue #7091 that documents a problem in the compilation phase (we cannot link in libnetdata as an object on old CMake versions and need to find a different way to express this).

  • The default variable-expansion / comparisons in CMake are awkward, we need this to make it sane:

    cmake_policy(SET CMP0054 "NEW")
  • Default paths for libs / includes are not comprehensive on most environments, we still need some heuristics for common locations, e.g. /usr/local on FreeBSD.

Recommendations

We should follow these steps:

  1. Build a prototype.
  2. Build a test-environment to check the prototype against environments / configurations that the team uses.
  3. Perform an "internal" release - merge the new CMake into master, but not announce it or offer to support it.
  4. Check it works for the team internally.
  5. Do a soft-release: offer it externally as a replacement option for autotools.
  6. Gather feedback and usage reports on a wider range of configurations.
  7. Do a hard-release: switch over the preferred build-system in the installation instructions.
  8. Gather feedback and usage reports on a wider range of configurations (again).
  9. Deprecate / remove the autotools build-system completely (so that we can support a single build-system).

Some smaller miscellaeneous suggestions:

  1. Remove the _Generic / strerror_r config to make the system simpler (use the technique on the blog post to make the standard version re-enterant so that it is thread-safe).
  2. Pull in jemalloc by source into the repo if it is our preferred malloc implementation.

Background

Last updated on