Re: [syzbot] WARNING in __kfence_free

[Marco Elver <elver@google.com>]

On Fri, 22 Apr 2022 at 07:09, Muchun Song <songmuchun@bytedance.com> wrote:
>
> On Thu, Apr 21, 2022 at 11:12:17AM +0200, Marco Elver wrote:
> > On Thu, Apr 21, 2022 at 01:58AM -0700, syzbot wrote:
> > > Hello,
> > >
> > > syzbot found the following issue on:
> > >
> > > HEAD commit:    559089e0a93d vmalloc: replace VM_NO_HUGE_VMAP with VM_ALLO..
> > > git tree:       upstream
> > > console output: https://syzkaller.appspot.com/x/log.txt?x=10853220f00000
> > > kernel config:  https://syzkaller.appspot.com/x/.config?x=2e1f9b9947966f42
> > > dashboard link: https://syzkaller.appspot.com/bug?extid=ffe71f1ff7f8061bcc98
> > > compiler:       aarch64-linux-gnu-gcc (Debian 10.2.1-6) 10.2.1 20210110, GNU ld (GNU Binutils for Debian) 2.35.2
> > > userspace arch: arm64
> > >
> > > Unfortunately, I don't have any reproducer for this issue yet.
> > >
> > > IMPORTANT: if you fix the issue, please add the following tag to the commit:
> > > Reported-by: syzbot+ffe71f1ff7f8061bcc98@syzkaller.appspotmail.com
> > >
> > > ------------[ cut here ]------------
> > > WARNING: CPU: 0 PID: 2216 at mm/kfence/core.c:1022 __kfence_free+0x84/0xc0 mm/kfence/core.c:1022
> >
> > That's this warning in __kfence_free:
> >
> >       #ifdef CONFIG_MEMCG
> >               KFENCE_WARN_ON(meta->objcg);
> >       #endif
> >
> > introduced in 8f0b36497303 ("mm: kfence: fix objcgs vector allocation").
> >
> > Muchun, are there any circumstances where the assumption may be broken?
> > Or a new bug elsewhere?
>
> meta->objcg always should be NULL when reaching __kfence_free().
> In theory, meta->objcg should be cleared via memcg_slab_free_hook().
>
> I found the following code snippet in do_slab_free().
>
>   /* memcg_slab_free_hook() is already called for bulk free. */
>   if (!tail)
>         memcg_slab_free_hook(s, &head, 1);
>
> The only posibility is @tail is not NULL, which is the case of
> kmem_cache_free_bulk(). However, here the call trace is kfree(),
> it seems to be impossible that missing call memcg_slab_free_hook().

Fair enough - we can probably wait for the bug to reoccur on another
instance, and until then assume something else wrong. What is slightly
suspicious is that it only occurred once on a QEMU TCG arm64 MTE
instance.

Thanks,
-- Marco