Re: [PATCH 1/2] kasan, mm: reset tag when access metadata

[Marco Elver <elver@google.com>]

On Wed, 28 Jul 2021 at 13:05, Kuan-Ying Lee <Kuan-Ying.Lee@mediatek.com> wrote:
>
> On Tue, 2021-07-27 at 20:22 +0100, Catalin Marinas wrote:
> > On Tue, Jul 27, 2021 at 04:32:02PM +0800, Kuan-Ying Lee wrote:
> > > On Tue, 2021-07-27 at 09:10 +0200, Marco Elver wrote:
> > > > +Cc Catalin
> > > >
> > > > On Tue, 27 Jul 2021 at 06:00, Kuan-Ying Lee <
> > > > Kuan-Ying.Lee@mediatek.com> wrote:
> > > > >
> > > > > Hardware tag-based KASAN doesn't use compiler instrumentation,
> > > > > we
> > > > > can not use kasan_disable_current() to ignore tag check.
> > > > >
> > > > > Thus, we need to reset tags when accessing metadata.
> > > > >
> > > > > Signed-off-by: Kuan-Ying Lee <Kuan-Ying.Lee@mediatek.com>
> > > >
> > > > This looks reasonable, but the patch title is not saying this is
> > > > kmemleak, nor does the description say what the problem is. What
> > > > problem did you encounter? Was it a false positive?
> > >
> > > kmemleak would scan kernel memory to check memory leak.
> > > When it scans on the invalid slab and dereference, the issue
> > > will occur like below.
> > >
> > > So I think we should reset the tag before scanning.
> > >
> > > # echo scan > /sys/kernel/debug/kmemleak
> > > [  151.905804]
> > > ==================================================================
> > > [  151.907120] BUG: KASAN: out-of-bounds in scan_block+0x58/0x170
> > > [  151.908773] Read at addr f7ff0000c0074eb0 by task kmemleak/138
> > > [  151.909656] Pointer tag: [f7], memory tag: [fe]
> >
> > It would be interesting to find out why the tag doesn't match.
> > Kmemleak
> > should in principle only scan valid objects that have been allocated
> > and
> > the pointer can be safely dereferenced. 0xfe is KASAN_TAG_INVALID, so
> > it
> > either goes past the size of the object (into the red zone) or it
> > still
> > accesses the object after it was marked as freed but before being
> > released from kmemleak.
> >
> > With slab, looking at __cache_free(), it calls kasan_slab_free()
> > before
> > ___cache_free() -> kmemleak_free_recursive(), so the second scenario
> > is
> > possible. With slub, however, slab_free_hook() first releases the
> > object
> > from kmemleak before poisoning it. Based on the stack dump, you are
> > using slub, so it may be that kmemleak goes into the object red
> > zones.
> >
> > I'd like this clarified before blindly resetting the tag.
>
> This kasan issue only happened on hardware tag-based kasan mode.
> Because kasan_disable_current() works for generic and sw tag-based
> kasan.
>
> HW tag-based kasan depends on slub so slab will not hit this
> issue.
> I think we can just check if HW tag-based kasan is enabled or not
> and decide to reset the tag as below.
>
> if (kasan_has_integrated_init()) // slub case, hw-tag kasan
>         pointer = *(unsigned long *)kasan_reset_tag((void *)ptr);
> else
>         pointer = *ptr; // slab

This is redundant. kasan_reset_tag() is a noop if
!IS_ENABLED(CONFIG_KASAN_HW_TAGS).

> Is this better or any other suggestions?
> Any suggestion is appreciated.

The current version is fine. But I think Catalin's point about why
kmemleak accesses the data in the first place still deserves some
investigation. Could it be a race between free and kmemleak scan?