Re: [PATCH] nilfs: Avoid use-after-free caused by nilfs->ns_writer

[Ryusuke Konishi]

On Thu, Nov 3, 2022 at 11:18 PM Shigeru Yoshida wrote:
>
> syzbot reported use-after-free in nilfs_segctor_sync() [1].
>
> The use-after-free occurs with nilfs->ns_writer.  The scenario for the
> issue is as follows:
>
> Task1                                   Task2
> ----------------------------------------------------------------------
> nilfs_construct_segment
>   nilfs_segctor_sync
>     init_wait
>     init_waitqueue_entry
>     add_wait_queue
>     schedule
>                                         nilfs_detach_log_writer
>                                           nilfs_segctor_destroy
>                                             kfree
>     finish_wait
>       _raw_spin_lock_irqsave
>         __raw_spin_lock_irqsave
>           do_raw_spin_lock
>             debug_spin_lock_before  <-- use-after-free
>
> While Task1 is sleeping, nilfs->ns_writer is freed by Task2.  After
> Task1 waked up, Task1 accesses nilfs->ns_writer which is already
> freed.
>
> This patch fixes the issue by taking nilfs->ns_segctor_sem in
> nilfs_construct_segment() so that nilfs->ns_segctor_sem cannot be
> freed while nilfs_segctor_sync() is sleeping.
>
> Link: https://syzkaller.appspot.com/bug?id=79a4c002e960419ca173d55e863bd09e8112df8b [1]
> Reported-by: syzbot+f816fa82f8783f7a02bb@xxxxxxxxxxxxxxxxxxxxxxxxx
> Signed-off-by: Shigeru Yoshida <syoshida@xxxxxxxxxx>
> ---
>  fs/nilfs2/segment.c | 14 +++++++++++---
>  1 file changed, 11 insertions(+), 3 deletions(-)
>
> diff --git a/fs/nilfs2/segment.c b/fs/nilfs2/segment.c
> index b4cebad21b48..d4f10d82664d 100644
> --- a/fs/nilfs2/segment.c
> +++ b/fs/nilfs2/segment.c
> @@ -2239,16 +2239,24 @@ static void nilfs_segctor_wakeup(struct nilfs_sc_info *sci, int err)
>  int nilfs_construct_segment(struct super_block *sb)
>  {
>         struct the_nilfs *nilfs = sb->s_fs_info;
> -       struct nilfs_sc_info *sci = nilfs->ns_writer;
> +       struct nilfs_sc_info *sci;
>         struct nilfs_transaction_info *ti;
> +       int ret;
>
> -       if (!sci)
> +       down_write(&nilfs->ns_segctor_sem);
> +       sci = nilfs->ns_writer;
> +       if (!sci) {
> +               up_write(&nilfs->ns_segctor_sem);
>                 return -EROFS;
> +       }
>
>         /* A call inside transactions causes a deadlock. */
>         BUG_ON((ti = current->journal_info) && ti->ti_magic == NILFS_TI_MAGIC);
>
> -       return nilfs_segctor_sync(sci);
> +       ret = nilfs_segctor_sync(sci);
> +       up_write(&nilfs->ns_segctor_sem);
> +
> +       return ret;
>  }
>
>  /**
> --
> 2.37.3
>

Thank you for your help Yoshida-san.

Your analysis of the bug cause is quite correct.

However, this patch caused deadlock for tests with fsync() or mkcp
command (i.e. nilfs_ioctl_sync).
They both call nilfs_construct_segment().
The approach implemented in this patch may not avoid the deadlock regression.

I'd like to consider other approaches.
My current thoughts are either to stop detaching ns_writer on remount
which is the root cause of the UAF races, or to introduce a refcount
on struct nilfs_sc_info.

Regards,
Ryusuke Konishi