Re: [PATCH] drivers: binderfs: fix memory leak in binderfs_fill_super

[Christian Brauner]

On Fri, Aug 12, 2022 at 04:09:23PM +0200, Greg Kroah-Hartman wrote:
> On Fri, Aug 12, 2022 at 09:56:46PM +0800, Dongliang Mu wrote:
> > On Fri, Aug 12, 2022 at 9:41 PM Greg Kroah-Hartman
> > <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, Aug 12, 2022 at 09:21:24PM +0800, Dongliang Mu wrote:
> > > > From: Dongliang Mu <mudongliangabcd@xxxxxxxxx>
> > > >
> > > > In binderfs_fill_super, if s_root is not successfully initialized by
> > > > d_make_root, the previous allocated s_sb_info will not be freed since
> > > > generic_shutdown_super first checks if sb->s_root and then does
> > > > put_super operation. The put_super operation calls binderfs_put_super
> > > > to deallocate s_sb_info and put ipc_ns. This will lead to memory leak
> > > > in binderfs_fill_super.
> > > >
> > > > Fix this by invoking binderfs_put_super at error sites before s_root
> > > > is successfully initialized.
> > > >
> > > > Fixes: 095cf502b31e ("binderfs: port to new mount api")
> > > > Reported-by: syzkaller <syzkaller@xxxxxxxxxxxxxxxx>
> > >
> > > Where is the specific syzkaller link for this report?  It would be good
> > > to reference it so it can be properly checked.
> > >
> > > Also, how did you test this change?
> > 
> > I found this memory leak in my local syzkaller, and there is no any
> > syzbot report about this crash, therefore I use such a Reported-by to
> > indicate.
> > 
> > Although my local syzkaller does generate any reproducer, this bug can
> > be triggered by injecting faults at new_inode and d_make_root (i.e.,
> > between s_sb_info allocation and code after d_make_root).
> > 
> > >
> > > > Signed-off-by: Dongliang Mu <mudongliangabcd@xxxxxxxxx>
> > > > ---
> > > >  drivers/android/binderfs.c | 8 ++++++--
> > > >  1 file changed, 6 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/drivers/android/binderfs.c b/drivers/android/binderfs.c
> > > > index 588d753a7a19..20f5bc77495f 100644
> > > > --- a/drivers/android/binderfs.c
> > > > +++ b/drivers/android/binderfs.c
> > > > @@ -710,8 +710,10 @@ static int binderfs_fill_super(struct super_block *sb, struct fs_context *fc)
> > > >       info->mount_opts.stats_mode = ctx->stats_mode;
> > > >
> > > >       inode = new_inode(sb);
> > > > -     if (!inode)
> > > > +     if (!inode) {
> > > > +             binderfs_put_super(sb);
> > > >               return -ENOMEM;
> > > > +     }
> > > >
> > > >       inode->i_ino = FIRST_INODE;
> > > >       inode->i_fop = &simple_dir_operations;
> > > > @@ -721,8 +723,10 @@ static int binderfs_fill_super(struct super_block *sb, struct fs_context *fc)
> > > >       set_nlink(inode, 2);
> > > >
> > > >       sb->s_root = d_make_root(inode);
> > > > -     if (!sb->s_root)
> > > > +     if (!sb->s_root) {
> > > > +             binderfs_put_super(sb);
> > > >               return -ENOMEM;
> > > > +     }
> > >
> > > How did you test this change to verify that you are not now just leaking
> > > memory?  It looks to me like you just changed one problem for another
> > > one :(
> > 
> > As mentioned above, I just tested my change by injecting faults at
> > new_inode and d_make_root.
> > 
> > Can you explain more about "changed one problem for another one"? I
> > don't quite understand this statement.
> 
> I think you are leaking memory in at least your second change here,
> possibly the first, I didn't look at the code very closely.

It's a bit tricky to follow but d_make_root() always consumes the inode.
On success via d_instantiate() and on failure via iput(). So when
d_make_root() has been called the inode is off limits. And as soon as
d_make_root() has returned successfully we're guaranteed that
sb->s_fs_info is cleaned up if a ->put_super() method has been defined.
Just fyi.