[PATCH] fs: erofs: remember if kobject_init_and_add was done

[Dongliang Mu]

On Tue, Mar 15, 2022 at 8:26 PM Gao Xiang <hsiangkao at linux.alibaba.com> wrote:
>
> Hi Dongliang,
>
> On Tue, Mar 15, 2022 at 07:59:26PM +0800, Dongliang Mu wrote:
> > On Tue, Mar 15, 2022 at 7:05 PM Huang Jianan <jnhuang95 at gmail.com> wrote:
> > >
> > > 在 2022/3/15 18:55, Gao Xiang 写道:
> > > > On Tue, Mar 15, 2022 at 06:43:01PM +0800, Huang Jianan wrote:
> > > >> 在 2022/3/15 15:51, Dongliang Mu 写道:
> > > >>> From: Dongliang Mu <mudongliangabcd at gmail.com>
> > > >>>
> > > >>> Syzkaller hit 'WARNING: kobject bug in erofs_unregister_sysfs'. This bug
> > > >>> is triggered by injecting fault in kobject_init_and_add of
> > > >>> erofs_unregister_sysfs.
> > > >>>
> > > >>> Fix this by remembering if kobject_init_and_add is successful.
> > > >>>
> > > >>> Note that I've tested the patch and the crash does not occur any more.
> > > >>>
> > > >>> Reported-by: syzkaller <syzkaller at googlegroups.com>
> > > >>> Signed-off-by: Dongliang Mu <mudongliangabcd at gmail.com>
> > > >>> ---
> > > >>>    fs/erofs/internal.h | 1 +
> > > >>>    fs/erofs/sysfs.c    | 9 ++++++---
> > > >>>    2 files changed, 7 insertions(+), 3 deletions(-)
> > > >>>
> > > >>> diff --git a/fs/erofs/internal.h b/fs/erofs/internal.h
> > > >>> index 5aa2cf2c2f80..9e20665e3f68 100644
> > > >>> --- a/fs/erofs/internal.h
> > > >>> +++ b/fs/erofs/internal.h
> > > >>> @@ -144,6 +144,7 @@ struct erofs_sb_info {
> > > >>>     u32 feature_incompat;
> > > >>>     /* sysfs support */
> > > >>> +   bool s_sysfs_inited;
> > > >> Hi Dongliang,
> > > >>
> > > >> How about using sbi->s_kobj.state_in_sysfs to avoid adding a extra member in
> > > >> sbi ?
> > > > Ok, I have no tendency of these (I'm fine with either ways).
> > > > I've seen some usage like:
> > > >
> > > > static inline int device_is_registered(struct device *dev)
> > > > {
> > > >          return dev->kobj.state_in_sysfs;
> > > > }
> > > >
> > > > But I'm still not sure if we need to rely on such internal
> > > > interface.. More thoughts?
> > >
> > > Yeah... It seems that it is better to use some of the interfaces
> > > provided by kobject,
> > > otherwise we should still maintain this state in sbi.
> > >
> >
> > I am fine with either way. Let me know if you reach to an agreement.
>
> If you have time, would you mind sending another patch by using
> state_in_sysfs? I'd like to know Chao's perference later, and
> apply one of them...

OK, let me test this patch in my local workspace. If it works, I will
send it later.

>
> Thanks,
> Gao Xiang
>