[PATCH 1/1] Fix segfault in DTC

[David Gibson]

On Wed, Oct 03, 2012 at 03:33:30PM -0600, Stephen Warren wrote:
> On 10/01/2012 06:08 PM, David Gibson wrote:
> > On Mon, Oct 01, 2012 at 10:41:09AM -0600, Stephen Warren wrote:
> >> On 10/01/2012 12:46 AM, David Gibson wrote:
> >>> On Sun, Sep 30, 2012 at 11:34:50PM -0600, Stephen Warren wrote:
> >>>> On 09/29/2012 05:53 PM, David Gibson wrote:
> >>>>> On Fri, Sep 28, 2012 at 01:05:33PM -0600, Stephen Warren wrote:
> >>>>>> On 09/28/2012 12:53 PM, Jon Loeliger wrote:
> >>>>>>>>>
> >>>>>>>>> Yeah, seems like the kernel DTC is quite old.
> >>>>>>>>
> >>>>>>>> FYI, I'm working on a patch to the kernel to bring in the latest dtc.
> >>>>>>>
> >>>>>>> Awesome.  Thank you.
> >>>>>>>
> >>>>>>>> I've run a regression test vs. the old dtc in the kernel ...
> >>>>>>>
> >>>>>>> Which is the icky step.  Again, thank you.
> >>>>>>>
> >>>>>>>> ... and found that
> >>>>>>>> some of the PowerPC .dts files don't compile with the new dtc (but did
> >>>>>>>> with the old), all due to non-existent labels/paths being referenced.
> >>>>>>>> I'll try and track down whether this is a regression in dtc, or simply
> >>>>>>>> buggy .dts files that weren't noticed before.
> >>>>>>>
> >>>>>>> I think you should just smack the PowerPC guys. :-)
> >>>>>>
> >>>>>> For the record in this thread, it was a regression I introduced into dtc
> >>>>>> - the patch I just sent was for this.
> >>>>>
> >>>>> I would be nice to add a testcase for this regression into dtc.
> >>>>
> >>>> The issue here was caused by uninitialized memory, so it would, I think,
> >>>> be basically impossible to create a test-case that would be guaranteed
> >>>> to fail because of this; it'd depend on the internal details of the
> >>>> malloc library and how/when it re-used previously free()d memory blocks.
> >>>
> >>> It doesn't have to be guaranteed to fail to be useful.  Plus, we
> >>> already have the infrastructure to run the tests under valgrind, which
> >>> would catch it.
> >>
> >> I certainly disagree here; the absolute worst kind of test is one which
> >> gives different results each time it's run, or statically gives
> >> different results to different people. People will either ignore the
> >> test because it's flaky, or it'll end up blaming the wrong person due to
> >> some entirely unrelated and correct change just happening to tickle the
> >> test.
> > 
> > I'd agree 100% if the test could give false failures.  But in this
> > case it can only give false passes.
> 
> That is true. I still dislike flaky tests irrespective of
> false-{negative,positive} though.

I don't love it, but I think a test which can generate false-negatives
is better than no test.

> > If the test fails there is a bug
> > *somewhere*, even if it's not actually in whatever changed last.  The
> > test framework actually has a "PASS (inconclusive)" result for exactly
> > this sort of case.
> > 
> >> If we were to force any such new test to always run under valgrind, then
> >> hopefully the test would always fail (assuming the test harness triggers
> >> failure if valgrind finds problems).
> > 
> > Aside: it's supposed to;  if it doesn't, that's a bug.  You can try it
> > easily enough with "make checkm".
> 
> Ah, I didn't know about that make target (or even "make check"; I'd
> always run run_tests.sh manually).
> 
> Incidentally, before commit 317a5d9 "dtc: zero out new label objects"
> the following two tests fail under make checkm:
> 
> dtc -I dts -O dtb -o multilabel.test.dtb multilabel.dts:        FAIL
> Returned error code 126
> 
> dtc -I dts -O dtb -o multilabel_merge.test.dtb multilabel_merge.dts:
> FAIL Returned error code 126
> 
> However, they pass at/after that commit.

Ah, interesting.  Goes to show that neither Jon nor I runs the
valgrind check as often as we probably should.

> Equally, those failures were introduced with commit 45013d8 "dtc: Add
> ability to delete nodes and properties", which is exactly what I'd
> expect given the fix was a fix for that commit.
> 
> So, it seems like we already have tests that catch this problem. Do we
> need to do anything given that?

Hm, I guess not.

> I am slightly surprised that the problem didn't cause all tests to fail
> make checkm though (I'd expect any usage of a label to trigger the
> problem); I'll have to think about why some more...

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson