Re: [syzbot] linux-next boot error: KASAN: slab-out-of-bounds Read in _find_next_bit

[Dmitry Vyukov <dvyukov@google.com>]

On Wed, 7 Sept 2022 at 07:53, Andrew Morton <akpm@linux-foundation.org> wrote:
> > > > The buggy address belongs to the object at ffff888017576600
> > > >  which belongs to the cache kmalloc-192 of size 192
> > > > The buggy address is located 184 bytes inside of
> > > >  192-byte region [ffff888017576600, ffff8880175766c0)
> > >
> > > At offset 184 of a 192-byte region.
> > >
> > > So what's wrong with doing that?  Does KASAN have an off-by-one?
> >
> > Hi Andrew, all,
> >
> > This is a bug in FIND_NEXT_BIT(). It should be
> >   if (idx >= sz / BITS_PER_LONG)                                   \
> >           goto out;                                                \
> >
> > instead of
> >   if (idx > sz / BITS_PER_LONG)                                    \
> >           goto out;                                                \
> >
> > The fix is in bitmap-for-next, expected to be in -next by tomorrow.
> > Sorry for the noise.
>
> OK... but why is KASAN reporting a bad access from an area
> which appears to be OK?

Hi Andrew,

Good point. Filed https://bugzilla.kernel.org/show_bug.cgi?id=216457 for this.
Copy-pasting description below:

KASAN says:

==================================================================
BUG: KASAN: slab-out-of-bounds in _find_next_bit+0x143/0x160 lib/find_bit.c:109
Read of size 8 at addr ffff8880175766b8 by task kworker/1:1/26
...
The buggy address belongs to the object at ffff888017576600
 which belongs to the cache kmalloc-192 of size 192
The buggy address is located 184 bytes inside of
 192-byte region [ffff888017576600, ffff8880175766c0)
...
Memory state around the buggy address:
>ffff888017576680: 00 00 00 00 00 00 00 fc fc fc fc fc fc fc fc fc
                                        ^
==================================================================

This "address is located 184 bytes inside of 192-byte region" is
confusing b/c it does not look like an out-of-bounds access.

What happens here is that the allocation request was for 184 bytes, so
the last 8 bytes in the 192-byte slab are poisoned. But KASAN does not
store the requested size in the object header, so it just prints the
full slab size everywhere.

User-space ASAN does store 48-bit requested size in the header. But
KASAN uses additional bytes in the header for:

struct kasan_alloc_meta {
depot_stack_handle_t aux_stack[2];

So we don't have space for requested size w/o increasing header size
(currently should be 16 bytes).

We could either try to infer requested size from the shadow (count
poisoned bytes at the end); or improve wording of the message at least
to make it clear that 192 is just full slab size.

For context see:
https://lore.kernel.org/all/20220906173154.6f2664c8fc6b83470c5dfea1@linux-foundation.org/