Re: [Qemu-devel] [PATCH] tests: set MALLOC_PERTURB_ to expose memory bugs

[Markus Armbruster]

"Daniel P. Berrange" <address@hidden> writes:

> On Fri, May 17, 2013 at 11:54:12AM +0200, Markus Armbruster wrote:
>> Stefan Hajnoczi <address@hidden> writes:
>> 
>> > glibc wipes malloc(3) memory when the MALLOC_PERTURB_ environment
>> > variable is set.  The value of the environment variable determines the
>> > bit pattern used to wipe memory.  For more information, see
>> > http://udrepper.livejournal.com/11429.html.
>> >
>> > Set MALLOC_PERTURB_ for gtester and qemu-iotests.  Note we always set
>> > the environment variable to 1 so the test is deterministic.  Setting a
>> > random variable might expose more bugs but would be harder to reproduce.
>> >
>> > Both make check and qemu-iotests pass with MALLOC_PERTURB_ enabled.
>> >
>> > Signed-off-by: Stefan Hajnoczi <address@hidden>
>> > ---
>> > Lucas noticed KVM autotest failures when enabling MALLOC_PERTURB_.  By 
>> > enabling
>> > it for in-tree test suites we can detect memory management errors earlier.
>> >
>> >  tests/Makefile           | 4 +++-
>> >  tests/qemu-iotests/check | 2 +-
>> >  2 files changed, 4 insertions(+), 2 deletions(-)
>> >
>> > diff --git a/tests/Makefile b/tests/Makefile
>> > index a307d5a..25f6d28 100644
>> > --- a/tests/Makefile
>> > +++ b/tests/Makefile
>> > @@ -171,6 +171,7 @@ GCOV_OPTIONS = -n $(if $(V),-f,)
>> >  $(patsubst %, check-qtest-%, $(QTEST_TARGETS)): check-qtest-%: 
>> > $(check-qtest-y)
>> >    $(if $(CONFIG_GCOV),@rm -f *.gcda */*.gcda */*/*.gcda */*/*/*.gcda,)
>> >    $(call quiet-command,QTEST_QEMU_BINARY=$*-softmmu/qemu-system-$* \
>> > +          MALLOC_PERTURB_=1 \
>> >            gtester $(GTESTER_OPTIONS) -m=$(SPEED) 
>> > $(check-qtest-$*-y),"GTESTER $@")
>> >    $(if $(CONFIG_GCOV),@for f in $(gcov-files-$*-y); do \
>> >      echo Gcov report for $$f:;\
>> 
>> If you want punishment, why not go for extra punishment?
>> 
>> MALLOC_PERTURB_=$(($RANDOM % 255 + 1))
>
> That could lead to non-reproducable failures though. I think it is better
> to use a fixed value so that you're more likely to be able to reproduce
> the issue every time you run the tests.

A fixed value misses failures.

A random value is just as reproducible as a fixed one, simply use the
same value.  Requires the value to be logged, of course.

> Rather than setting MALLOC_PERTURB_=1 unconditionally in the Makefile
> though, it ought to honour any existing MALLOC_PERTURB_ env variable
> the user has set. That could let automated test harness run repeatedly
> with random MALLOC_PERTURB_, while still giving a deterministic value
> for developers by default.

Yes, we should respect the user's MALLOC_PERTURB_.