Re: [PATCH] test-skeleton.c: Do not enable M_PERTURB

[Florian Weimer <fweimer at redhat dot com>]

On 06/23/2016 07:16 PM, Siddhesh Poyarekar wrote:
> On Thu, Jun 23, 2016 at 12:57:17PM -0400, Zack Weinberg wrote:
> > Is there a well-documented way to run the tests under valgrind?
>
> I don't think so, but it ought to be sufficient to run valgrind
> --trace-children=yes for all of make check.  Dead slow, but I don't
> see why it shouldn't work.
>
> Maybe this is a good candidate for enhancement in glibc to allow
> running individual tests under valgrind, something like
> 'make valgrind-check'.

Running tests under valgrind is one thing.  Interpreting the errors is 
another.

On some architectures, running valgrind on a non-installed libc gives 
rather poor results.  Some versions do not recognize malloc as such, and 
on some architectures, the suppression files are inactive because the 
libc paths do not match the hard-coded expectations of valgrind.

With certain tests, we have the problem that we have memory leaks when 
subprocesses call exit, without freeing data, which is marked as a leak 
in the error log.

Thread stack deallocation is generally racy, and stacks may be reported 
by valgrind as a memory leak.

These are just the issues I have seen so far when running valgrind 
manually.  I expect there are more.

Thanks,
Florian