[Starpu-devel] Seek and destroy memory leaks.

[Cyril Roelandt <cyril.roelandt@inria.fr>]

Hello,

We often have to run Valgrind on our examples and our tests in order to 
track memory leaks. But Valgrind is quite slow, so it ends up being 
killed by buildbot after 600 seconds, or running for a very long time 
(hours ?) in our terminals, which is not very convenient.

I think it is Samuel who suggested that, when running Valgrind, we 
should skip the computations. It does make sense, since memory leaks 
probably happen in the core of StarPU, and not in the codelets.

Valgrind provides RUNNING_ON_VALGRIND, which is quite simple to use :

$ cat test.c
#include <stdio.h>
#include <stdlib.h>
#include <valgrind/valgrind.h>

int
main(void)
{
        if (RUNNING_ON_VALGRIND)
                (void) fprintf(stdout, "Running on valgrind.\n");
        else
                (void) fprintf(stdout, "Not running on valgrind.\n");

        return EXIT_SUCCESS;
}

$ ./test
Not running on valgrind.

$ valgrind ./test
// blah
Running on valgrind.
// blah


I think we could do something like this :

#if STARPU_HAVE_VALGRIND_H
#define STARPU_SKIP_IF_VALGRIND do { if(RUNNING_ON_VALGRIND) return; } 
while(0)
#else
#define STARPU_SKIP_IF_VALGRIND 0
#endif


And add STARPU_SKIP_IF_VALGRIND at the beginning of our codelets (which 
would be easy to do with a tool like Coccinelle) :

#if STARPU_HAVE_VALGRIND_H
#include <valgrind/valgrind.h>
#endif
...
void cpu_func_0(void *buffers[], void *arg)
{
        STARPU_SKIP_IF_VALGRIND;

        /* Computations */
}


I am afraid that every test will return EXIT_FAILURE, since the result 
will not be good, but is that really a problem ?

This way, Valgrind should take much less time to execute our tests, and 
we should be able to find memory leaks easily. Fixing them might be 
harder :) So, what do you think about this ?

Cyril.