Re: general protection fault in qdisc_put

[Dmitry Vyukov]

On Sun, Sep 8, 2019 at 6:19 PM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Sat, Sep 7, 2019 at 11:08 PM syzbot
> <syzbot+d5870a903591faaca4ae@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > The bug was bisected to:
> >
> > commit e41d58185f1444368873d4d7422f7664a68be61d
> > Author: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
> > Date:   Wed Jul 12 21:34:35 2017 +0000
> >
> >      fault-inject: support systematic fault injection
>
> That commit does seem a bit questionable, but not the cause of this
> problem (just the trigger).
>
> I think the questionable part is that the new code doesn't honor the
> task filtering, and will fail even for protected tasks. Dmitry?

That commit added a new fault injection mode with a new API that is
used by syzkaller to inject faults. Before that commit the fault
inject is not working for syzkaller at all. I think this bisection
result simply means "the GPF is related to an earlier failure".

> > kasan: GPF could be caused by NULL-ptr deref or user memory access
> > general protection fault: 0000 [#1] PREEMPT SMP KASAN
> > CPU: 1 PID: 9699 Comm: syz-executor169 Not tainted 5.3.0-rc7+ #0
> > Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS
> > Google 01/01/2011
> > RIP: 0010:qdisc_put+0x25/0x90 net/sched/sch_generic.c:983
>
> Yes, looks like 'qdisc' is NULL.
>
> This is the
>
>         qdisc_put(q->qdisc);
>
> in sfb_destroy(), called from qdisc_create().
>
> I think what is happening is this (in qdisc_create()):
>
>         if (ops->init) {
>                 err = ops->init(sch, tca[TCA_OPTIONS], extack);
>                 if (err != 0)
>                         goto err_out5;
>         }
>         ...
> err_out5:
>         /* ops->init() failed, we call ->destroy() like qdisc_create_dflt() */
>         if (ops->destroy)
>                 ops->destroy(sch);
>
> and "ops->init" is sfb_init(), which will not initialize q->qdisc if
> tcf_block_get() fails.
>
> I see two solutions:
>
>  (a) move the
>
>         q->qdisc = &noop_qdisc;
>
>      up earlier in sfb_init(), so that qdisc is always initialized
> after sfb_init(), even on failure.
>
>  (b) just make qdisc_put(NULL) just silently work as a no-op.
>
>  (c) change all the semantics to not call ->destroy if ->init failed.
>
> Honestly, (a) seems very fragile - do all the other init routines do
> this? And (c) sounds like a big change, and very fragile too.
>
> So I'd suggest that qdisc_put() be made to just ignore a NULL pointer
> (and maybe an error pointer too?).
>
> But I'll leave it to the maintainers to sort out the proper fix.
> Maybe people prefer (a)?
>
>                    Linus