Re: memory leak in vq_meta_prefetch

[Dmitry Vyukov <dvyukov@google.com>]

On Fri, Jul 26, 2019 at 5:57 PM Catalin Marinas <catalin.marinas@arm.com> wrote:
>
> On Fri, Jul 26, 2019 at 05:20:55PM +0200, Dmitry Vyukov wrote:
> > On Fri, Jul 26, 2019 at 3:00 PM Catalin Marinas <catalin.marinas@arm.com> wrote:
> > > On Wed, Jul 24, 2019 at 12:18:07PM -0700, syzbot wrote:
> > > > syzbot found the following crash on:
> > > >
> > > > HEAD commit:    c6dd78fc Merge branch 'x86-urgent-for-linus' of git://git...
> > > > git tree:       upstream
> > > > console output: https://syzkaller.appspot.com/x/log.txt?x=15fffef4600000
> > > > kernel config:  https://syzkaller.appspot.com/x/.config?x=8de7d700ea5ac607
> > > > dashboard link: https://syzkaller.appspot.com/bug?extid=a871c1e6ea00685e73d7
> > > > compiler:       gcc (GCC) 9.0.0 20181231 (experimental)
> > > > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=127b0334600000
> > > > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=12609e94600000
> > > >
> > > > The bug was bisected to:
> > > >
> > > > commit 0e5f7d0b39e1f184dc25e3adb580c79e85332167
> > > > Author: Nicolas Ferre <nicolas.ferre@atmel.com>
> > > > Date:   Wed Mar 16 13:19:49 2016 +0000
> > > >
> > > >     ARM: dts: at91: shdwc binding: add new shutdown controller documentation
> > >
> > > That's another wrong commit identification (a documentation patch should
> > > not cause a memory leak).
> > >
> > > I don't really think kmemleak, with its relatively high rate of false
> > > positives, is suitable for automated testing like syzbot. You could
> >
> > Do you mean automated testing in general, or bisection only?
> > The wrong commit identification is related to bisection only, but you
> > generalized it to automated testing in general. So which exactly you
> > mean?
>
> I probably meant both. In terms of automated testing and reporting, if
> the false positives rate is high, people start ignoring the reports. So
> it requires some human checking first (or make the tool more robust).
>
> W.r.t. bisection, the false negatives (rather than positives) will cause
> the tool to miss the problematic commit and misreport. I'm not sure you
> can make the reporting deterministic on successive runs given that you
> changed the kernel HEAD (for bisection). But it may get better if you
> have a "stopscan" kmemleak option which freezes the machine during
> scanning (it has been discussed in the past but I really struggle to
> find time to work on it; any help appreciated ;)).


Do you have any data points wrt automated testing in general? This
disagrees with what I see.

For bisection, I agree. Need to look at the data we got over the past
days when it become enabled. But I suspect that, yes, false positives,
flakes, and other true leaks can make it infeasible.