On Monday, August 6, 2018 9:29 PM, Tetsuo Handa wrote:
> On 2018/08/06 21:44, Wang, Wei W wrote:
> > On Monday, August 6, 2018 6:29 PM, Tetsuo Handa wrote:
> >> On 2018/08/06 18:56, Wei Wang wrote:
> >>> On 08/03/2018 08:11 PM, Tetsuo Handa wrote:
> >>>> On 2018/08/03 17:32, Wei Wang wrote:
> >>>>> +static int virtio_balloon_register_shrinker(struct virtio_balloon
> >>>>> +*vb) {
> >>>>> + vb->shrinker.scan_objects = virtio_balloon_shrinker_scan;
> >>>>> + vb->shrinker.count_objects = virtio_balloon_shrinker_count;
> >>>>> + vb->shrinker.batch = 0;
> >>>>> + vb->shrinker.seeks = DEFAULT_SEEKS;
> >>>> Why flags field is not set? If vb is allocated by
> >>>> kmalloc(GFP_KERNEL) and is nowhere zero-cleared, KASAN would
> complain it.
> >>>
> >>> Could you point where in the code that would complain it?
> >>> I only see two shrinker flags (NUMA_AWARE and MEMCG_AWARE), and
> >> they seem not related to that.
> >>
> >> Where is vb->shrinker.flags initialized?
> >
> > Is that mandatory to be initialized?
>
> Of course. ;-)
>
> > I find it's not initialized in most shrinkers (e.g. zs_register_shrinker,
> huge_zero_page_shrinker).
>
> Because most shrinkers are "statically initialized (which means that
> unspecified fields are implicitly zero-cleared)" or "dynamically allocated with
> __GFP_ZERO or zero-cleared using
> memset() (which means that all fields are once zero-cleared)".
>
> And if you once zero-clear vb at allocation time, you will get a bonus that
> calling unregister_shrinker() without corresponding register_shrinker() is safe
> (which will simplify initialization failure path).
Oh, I see, thanks. So it sounds better to directly kzalloc vb.
Best,
Wei
