Re: [PATCH mm 2/4] kasan: handle concurrent kasan_record_aux_stack calls

[Marco Elver <elver@google.com>]

On Wed, 13 Dec 2023 at 15:40, Andrey Konovalov <andreyknvl@gmail.com> wrote:
>
> On Tue, Dec 12, 2023 at 8:29 PM Marco Elver <elver@google.com> wrote:
> >
> > > -       stack_depot_put(alloc_meta->aux_stack[1]);
> > > +       new_handle = kasan_save_stack(0, depot_flags);
> > > +
> > > +       spin_lock_irqsave(&aux_lock, flags);
> >
> > This is a unnecessary global lock. What's the problem here? As far as
> > I can understand a race is possible where we may end up with
> > duplicated or lost stack handles.
>
> Yes, this is the problem. And this leads to refcount underflows in the
> stack depot code, as we fail to keep precise track of the stack
> traces.
>
> > Since storing this information is best effort anyway, and bugs are
> > rare, a global lock protecting this is overkill.
> >
> > I'd just accept the racyness and use READ_ONCE() / WRITE_ONCE() just
> > to make sure we don't tear any reads/writes and the depot handles are
> > valid.
>
> This will help with the potential tears but will not help with the
> refcount issues.
>
> > There are other more complex schemes [1], but I think they are
> > overkill as well.
> >
> > [1]: Since a depot stack handle is just an u32, we can have a
> >
> >  union {
> >    depot_stack_handle_t handles[2];
> >    atomic64_t atomic_handle;
> >   } aux_stack;
> > (BUILD_BUG_ON somewhere if sizeof handles and atomic_handle mismatch.)
> >
> > Then in the code here create the same union and load atomic_handle.
> > Swap handle[1] into handle[0] and write the new one in handles[1].
> > Then do a cmpxchg loop to store the new atomic_handle.
>
> This approach should work. If you prefer, I can do this instead of a spinlock.
>
> But we do need some kind of atomicity while rotating the aux handles
> to make sure nothing gets lost.

Yes, I think that'd be preferable. Although note that not all 32-bit
architectures have 64-bit atomics, so that may be an issue. Another
alternative is to have a spinlock next to the aux_stack (it needs to
be initialized properly). It'll use up a little more space, but that's
for KASAN configs only, so I think it's ok. Certainly better than a
global lock.