Re: [PATCH v2 1/3] rcu/kasan: record and print call_rcu() call stack

[Dmitry Vyukov]

On Tue, May 12, 2020 at 6:14 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> > > > > > > This feature will record first and last call_rcu() call stack and
> > > > > > > print two call_rcu() call stack in KASAN report.
> > > > > >
> > > > > > Suppose that a given rcu_head structure is passed to call_rcu(), then
> > > > > > the grace period elapses, the callback is invoked, and the enclosing
> > > > > > data structure is freed.  But then that same region of memory is
> > > > > > immediately reallocated as the same type of structure and again
> > > > > > passed to call_rcu(), and that this cycle repeats several times.
> > > > > >
> > > > > > Would the first call stack forever be associated with the first
> > > > > > call_rcu() in this series?  If so, wouldn't the last two usually
> > > > > > be the most useful?  Or am I unclear on the use case?
> > > >
> > > > 2 points here:
> > > >
> > > > 1. With KASAN the object won't be immediately reallocated. KASAN has
> > > > 'quarantine' to delay reuse of heap objects. It is assumed that the
> > > > object is still in quarantine when we detect a use-after-free. In such
> > > > a case we will have proper call_rcu stacks as well.
> > > > It is possible that the object is not in quarantine already and was
> > > > reused several times (quarantine is not infinite), but then KASAN will
> > > > report non-sense stacks for allocation/free as well. So wrong call_rcu
> > > > stacks are less of a problem in such cases.
> > > >
> > > > 2. We would like to memorize 2 last call_rcu stacks regardless, but we
> > > > just don't have a good place for the index (bit which of the 2 is the
> > > > one to overwrite). Probably could shove it into some existing field,
> > > > but then will require atomic operations, etc.
> > > >
> > > > Nobody knows how well/bad it will work. I think we need to get the
> > > > first version in, deploy on syzbot, accumulate some base of example
> > > > reports and iterate from there.
> > >
> > > If I understood the stack-index point below, why not just move the
> > > previous stackm index to clobber the previous-to-previous stack index,
> > > then put the current stack index into the spot thus opened up?
> >
> > We don't have any index in this change (don't have memory for such index).
> > The pseudo code is"
> >
> > u32 aux_stacks[2]; // = {0,0}
> >
> > if (aux_stacks[0] != 0)
> >     aux_stacks[0] = stack;
> > else
> >    aux_stacks[1] = stack;
>
> I was thinking in terms of something like this:
>
> u32 aux_stacks[2]; // = {0,0}
>
> if (aux_stacks[0] != 0) {
>     aux_stacks[0] = stack;
> } else {
>    if (aux_stacks[1])
>         aux_stacks[0] = aux_stacks[1];
>    aux_stacks[1] = stack;
> }
>
> Whether this actually makes sense in real life, I have no idea.
> The theory is that you want the last two stacks.  However, if these
> elements get cleared at kfree() time, then I could easily believe that
> the approach you already have (first and last) is the way to go.
>
> Just asking the question, not arguing for a change!

Oh, this is so obvious... in hindsight! :)

Walter, what do you think?

I would do this. I think latter stacks are generally more interesting
wrt shedding light on a bug. The first stack may even be "statically
known" (e.g. if object is always queued into a workqueue for some lazy
initialization during construction).