Auto-generate FFF fake definitions for C API header files.
Incorporate this script into your normal build environment (like make) and automatically generate test-headers with faked function definitions ready-to-use with FFF's own gtest or some other unit-testing-framework.
Especially in the embedded world, compiling your (unit-)tests against the actual target platform isn't feasible, as the architecture you're executing the test on and your target platform you're writing the code for are generally not the same.
This is where faking a specific platform API might help you. Instead of calling the actual target platform API, you link your code-under-test (CuT) against a faked version of said API. Now, whenever your CuT tries to access the platform API,it is instead calling the fake-implementation which you can easily configure in your test-cases' setup phase.
FFF (fake-functions-framework) is a framework designed to easily create faked definitions of your API function-declarations, allowing you to configure return values and inspect call and argument histories that were called during the tests' runtime. Check out their awesome project on Github. AutoFFF utilizes the ability to easily create fake definitions provided by FFF, but could also be adapted to other mocking frameworks.
The problem with faking an API in embedded C is usually the infeasibility of using dynamic linking and C's lack of techniques like 'reflection' to manipulate your CuT during runtime. This makes the process of writing fake definitions a tedious, labor intensive and error prone matter.
Introducing AutoFFF, an attempt at automating the process of writing so called test-headers (headers which include the faked definitions).
When writing fakes you will notice that there are two approaches of laying out your fake.
- Banning the original API header
This strategy bans the original header by defining the API headers include guard, making it impossible to include the original function, variable and type declarations. This gives you ultimate freedom in the test-header, but also means that you will have to manually declare any types, functions and variables the API-user might expect. It also allows you to control the include hierarchy and maybe skip some headers which aren't compatible with your test-runner's architecture. In general this approach usually involves a lot of copy&pasting and is therefore more prone to "code rot". You also need to deep-dive any header you want to fake, understand its structure and inspect all the declarations and defines very closely. Not the optimal strategy if you're looking for an easy-to-maintain way of managing test-headers. - Wrapping the original API header
Conversely to the banning method, the wrapping strategy directly includes the original API header, and thereby imports any type, variable and function declarations. Also the include hierarchy is taken over from the original. The only thing to add into the test-header are the fake definitions. This method evidently grants you less freedom in the test-header, but is usually much shorter and slightly less prone to "rot" over time.
It should become obvious which method is better suited for automation. AutoFFF follows the wrapping approach of writing test-headers, which for most cases should be good enough.
Finally it must be stated, that these two philosophies seldomly mix well!
Use pip
to download and install AutoFFF from the PyPi repositories:
py -3.6 -m pip install autofff
Or install from source:
py -3.6 -m pip install .
Note that you'll most likely require the pycparser fake_libc_include
s header collection for AutoFFF to work. The pip
package does not ship with this external code. You may download the faked libc headers from pycparser
s Github, or check out the project as a submodule (when installing from source run git submodule update --init
).
py -3.6 -m autofff \
./examples/driver.h \
-O ./output/driver_th.h \
-I ./examples \
-F ./dependencies/pycparser/utils/fake_libc_include
To run build and run the tests, simply execute:
make run_tests
You can also let the makefile do the installation of AutoFFF for you.
make install_autofff
make -f examples/generate-via-makefile/generate_fakes.mk CRAWL_PATHS=examples
import autofff
import os.path
targetHeader = "examples/simple-headers/driver.h"
outputHeader = "output/driver_th.h"
fakes = './autofff/dependencies/pycparser/utils/fake_libc_include'
scnr = autofff.scanner.GCCScanner(targetHeader, fakes) # Create GCC code scanner
result = scnr.scan() # Scan for function declarations and definitions
gen = autofff.generator.SimpleFakeGenerator(os.path.splitext(os.path.basename(outputHeader))[0], targetHeader) # Create new generator with name output-header and path to target-header
if not os.path.exists(os.path.dirname(outputHeader)):
dirname = os.path.dirname(outputHeader)
os.makedirs(dirname)
with open(outputHeader, "w") as fs:
gen.generate(result, fs) # Generate fff-fakes from scanner-result
The format of the generated test-header obviously depends on the specifics of the FakeGenerator
being used.
- The
BareFakeGenerator
will only generate theFAKE_VALUE_
- andFAKE_VOID_FUNC
macros without any decorations, like include guards or header includes. Use this generator if you want to add your own (shell-based-)processing on top. - The
SimpleFakeGenerator
will generate a "minimum viable test header", meaning the result should be compilable without too much effort.
In some API headers functions may be defined within the header. This will cause issues when trying to fake this function, because by including the header the function definition is copied into each translation unit. If we try to apply a fake definition the usual way, we will end up with a "redefinition of function x" error.
AutoFFF implements a workaround to avoid this redefinition error and allowing to fake the original function. This workaround simply consists of some defines which will re-route any call to the original in-header definition to our faked one. For this to work it is required that the test-header is included (and thereby pre-processed) before any function call to the function under consideration is instructed, i.e. the test-header must be included before the CuT. Any function call that is processed before the workaround is being pre-processed will leave this function call targeted towards the original in-header definition.
In practice the workaround looks like this:
/* api.h */
#ifndef API_HEADER_H_
#define API_HEADER_H_
const char* foo(void)
{
return "Definitions inside headers are great!";
}
#endif
/* api_th.h */
#ifndef TEST_HEADER_H_
#define TEST_HEADER_H_
#include "fff.h"
#include "api.h"
/* Re-route any call to 'foo' to 'foo_fff' (our fake definition). */
#define foo foo_fff
/* By re-routing the '_fake' and '_reset' type and function the workaround becomes invisible in the test-case. */
#define foo_fake Foo_fff_fake
#define foo_reset Foo_fff_reset
/* Create the fake definition using the now re-routed 'foo'-symbol. */
FAKE_VALUE_FUNC(const char *, foo);
#endif
/* cut.c - code-under-test */
#include "api.h"
#include <stdio.h>
const char* bar(void)
{
const char* str = foo();
return str;
}
/* test.c */
#include "fff.h"
#include "api_th.h" /* fakes + workaround */
#include "cut.c"
setup(void)
{
RESET_FAKE(foo);
}
TEST_F(foo, ReturnBar_Success)
{
const char* expected_retval = "Definitions inside headers make faking difficult!";
foo_fake.return_val = expected_retval
const char* str = bar();
ASSERT_STREQ(expected_retval, str);
}
Some libraries like FreeRTOS or CMSIS require you to include their API headers in a very specific way. AutoFFF can't guess these policies (yet! ;P) from source-code alone. For cases where the include policy of your vendor lib does not allow each header to be preprocessed individually check out the -D
(--define
) and -i
(--includefile
) command line options. They may allow you to fix/trick the broken include chain.
As an example, for FreeRTOS' list.h
run:
py -3.6 -m autofff
[...]/include/list.h
-o [...]
-i [...]/include/FreeRTOS.h <<< inject FreeRTOS.h before preprocessing list.h
-F [...]