[R-pkg-devel] Failing Valgrind Passing R CMD check

[Dirk Eddelbuettel]

On 8 March 2022 at 18:46, Ivan Krylov wrote:
| On Sun, 6 Mar 2022 11:00:43 -0600
| Dirk Eddelbuettel <edd using debian.org> wrote:
| 
| > ==1485886== LEAK SUMMARY:
| > ==1485886==    definitely lost: 32 bytes in 1 blocks
| 
| > ==1485886== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0
| > from 0)
| 
| R CMD check only looks at the exit code of the process.
| 
| Valgrind crashes the process on invalid memory access, but doesn't
| consider leaks to be errors, unless you turn on --leak-check=full, and

That is a related problem I did not stress. We can pass debugger arguments to
the debugger used (here: valgrind) when we call R, but we cannot for R CMD check.
That should be added.

| even then still returns the exit code of the child process by default,
| i.e. zero:
| 
| $ cat leak.c 
| #include <stdlib.h>
| 
| int main(void) {
|         malloc(42);
|         return 0;
| }
| $ cc -o leak -g -O0 leak.c
| $ valgrind ./leak; echo exitcode=$?
| ...
| ==14675==    definitely lost: 42 bytes in 1 blocks
| ...
| ==14675== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)
| exitcode=0
| $ valgrind --leak-check=full ./leak; echo exitcode=$?
| ...
| 
| ==14688==    definitely lost: 42 bytes in 1 blocks
| ...
| ==14688== ERROR SUMMARY: 1 errors from 1 contexts (suppressed: 0 from 0)
| exitcode=0
| $ valgrind --leak-check=full --error-exitcode=42 ./leak; \
|  echo exitcode=$?
| ...
| ==14729==    definitely lost: 42 bytes in 1 blocks
| ...
| ==14729== ERROR SUMMARY: 1 errors from 1 contexts (suppressed: 0 from 0)
| exitcode=42
| 
| R CMD check could be patched to run 
| R -d 'valgrind --leak-check=full --error-exitcode=255' instead of just
| R -d valgrind to produce warnings on Valgrind-detected leaks.
| 
| The problem with this approach is the need for Valgrind suppression
| files. I've run an interactive example of a pure R package under R -d
| 'valgrind --leak-check=full --error-exitcode=255' (a relatively fresh
| SVN build) and got "definitely lost: 15,744 bytes in 41 blocks", mostly
| in plot-related code, allocations originating in Pango and GObject.

I have setup a CI setting using valgrind via R CMD check (and hence lacking
'--leak-check=full' though one could add a wrapper...) and it is a little
frustrating to not have it fail.

But of course it looks like it is a missing feature and we all could work
towards adding it.  All always that would require cooperation by R Core.

Dirk

-- 
https://dirk.eddelbuettel.com | @eddelbuettel | edd using debian.org