Valgrind

From OPeNDAP Documentation
Revision as of 22:57, 22 April 2013 by Jimg (talk | contribs) (Created page with "== Why == Why use ''valgrind'' with unit tests? Aside from why not, because it can catch errors like memory corruption, uninitialized value use and leaks. This addition drops i...")
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
⧼opendap2-jumptonavigation⧽

Why

Why use valgrind with unit tests? Aside from why not, because it can catch errors like memory corruption, uninitialized value use and leaks.

This addition drops into autotools very easily and works well with CppUnit.

How

Modify configure.ac and Makefile.am

Changes to configure.ac

First, modify configure.ac so that it takes an option to turn this feature on. The m4 file gcov_valgrind.m4 from libdap defines a very simple macro that will do this so only one line needs to be added to configure.ac. Copy and add the macro and add the line:

DODS_GCOV_VALGRIND

to configure.ac. This macro add the --enable-valgrind (and --enable-gcov; see below) options and defines an automake variable that can be used in a Makefile.am file.

Note that gcov is the GNU compile coverage analysis tool. To use it, you must configure using ./configure --enable-coverage (assuming you use the macro described above). Once configured for coverage analysis, run the tests and coverage data will be written to disk which can be analyzed using the gcov and gprof tools. The gcov program provides coverage data while the gprof tool provides profiling information.

Changes to Makefile.am

Add a conditional line to the unit tests Makefile.am that will set the Makefile variable TEST_ENVIRONMENT to the following: valgrind --quiet --trace-children=yes --error-exitcode=1 --dsymutil=yes --leak-check=yes --log-file="valgrind-%p.log". Here's what the options do:

quiet
suppress the usual (non-error) output of valgrind
trace-children
when a test starts a subprocess, it's trace too (not the default). This is important when starting programs made using libtool since they are scripts by default that fork/exec to start the 'real' program.
error-exitcode
valgrind does not usually return a non-zero code when it finds a leak, etc., but instead returns the program's exit code. Setting this to 1 means that leaks will show up as errors when the tests are run by make.
dsymutil
needed for OS/X; does nothing on Linux
leak-check
Using yes gets leak details (the default is summary)
log-file
Write any output detail errors to a file with the PID.

To add this in a Makefile.am use:

if USE_VALGRIND
TESTS_ENVIRONMENT=valgrind --quiet --trace-children=yes --error-exitcode=1 \
--dsymutil=yes --leak-check=yes --log-file="valgrind-%p.log"
endif

Other valgrind options that might be useful:

show-reachable
this will show all memory in use when the process exits
track-origins
the memcheck tool in valgrind shows returns errors for uninitialized use, but not where the uninitialized memory/variable was defined. setting this to yes wil get that information, but at the cost of about doubling the runtime of the test/program.