Re: Question about commit f9a67b1182e5 ("md/bitmap: clear bitmap if bitmap_create failed").

[Marion & Christophe JAILLET]

Le 14/09/2016 à 10:25, Guoqing Jiang a écrit :
> On 09/13/2016 01:24 PM, Shaohua Li wrote:
> > On Mon, Sep 12, 2016 at 09:09:48PM +0200, Christophe JAILLET wrote:
> > > Hi,
> > >
> > > I'm puzzled by commit f9a67b1182e5 ("md/bitmap: clear bitmap if
> > > bitmap_create failed").
> > Hi Christophe,
> > Thank you very much to help check this!
> >
> > > Part of the commit is:
> > >
> > > @@ -1865,8 +1866,10 @@ int bitmap_copy_from_slot(struct mddev 
> > > *mddev, int
> > > slot,
> > >       struct bitmap_counts *counts;
> > >       struct bitmap *bitmap = bitmap_create(mddev, slot);
> > >
> > > -    if (IS_ERR(bitmap))
> > > +    if (IS_ERR(bitmap)) {
> > > +        bitmap_free(bitmap);
> > >           return PTR_ERR(bitmap);
> > > +    }
> > >
> > > but if 'bitmap' is an error, I think that bad things will happen in
> > > 'bitmap_free()' when, at the beginning of the function, we will 
> > > execute:
> > >
> > >      if (bitmap->sysfs_can_clear) <-----------------
> > >          sysfs_put(bitmap->sysfs_can_clear);
>
> I guess it is safe, since below part is at the beginning of bitmap_free.
>
>         if (!bitmap) /* there was no bitmap */
>                 return;

I don't share your feeling.
bitmap_create() can return ERR_PTR(-ENOMEM) or ERR_PTR(-EINVAL).

In such cases 'if (!bitmap)' will not be helpful.

Maybe it should be turned into 'if (IS_ERR_OR_NULL(bitmap))' to handle 
errors returned by bitmap_create.
Maybe just removing the call to 'bitmap_free(bitmap)' is enough.

In any case, I think that the current logic is somehow broken.

Best regards,
CJ