Re: [RESEND PATCH v2] media: mtk-jpeg: Fix use after free bug due to uncanceled work

[Zheng Hacker]

Dmitry Osipenko <dmitry.osipenko@xxxxxxxxxxxxx> 于2023年8月28日周一 10:04写道：
>
> On 8/24/23 11:20, Zheng Hacker wrote:
> > Dmitry Osipenko <dmitry.osipenko@xxxxxxxxxxxxx> 于2023年8月23日周三 02:51写道：
> >
> >>
> >> Hello Zheng,
> >>
> >> On 7/7/23 12:24, Zheng Wang wrote:
> >>> In mtk_jpeg_probe, &jpeg->job_timeout_work is bound with
> >>> mtk_jpeg_job_timeout_work. Then mtk_jpeg_dec_device_run
> >>> and mtk_jpeg_enc_device_run may be called to start the
> >>> work.
> >>> If we remove the module which will call mtk_jpeg_remove
> >>> to make cleanup, there may be a unfinished work. The
> >>> possible sequence is as follows, which will cause a
> >>> typical UAF bug.
> >>>
> >>> Fix it by canceling the work before cleanup in the mtk_jpeg_remove
> >>>
> >>> CPU0                  CPU1
> >>>
> >>>                     |mtk_jpeg_job_timeout_work
> >>> mtk_jpeg_remove     |
> >>>   v4l2_m2m_release  |
> >>>     kfree(m2m_dev); |
> >>>                     |
> >>>                     | v4l2_m2m_get_curr_priv
> >>>                     |   m2m_dev->curr_ctx //use
> >>> Fixes: b2f0d2724ba4 ("[media] vcodec: mediatek: Add Mediatek JPEG Decoder Driver")
> >>> Signed-off-by: Zheng Wang <zyytlz.wz@xxxxxxx>
> >>> ---
> >>> - v2: use cancel_delayed_work_sync instead of cancel_delayed_work suggested by Kyrie.
> >>> ---
> >>>  drivers/media/platform/mediatek/jpeg/mtk_jpeg_core.c | 1 +
> >>>  1 file changed, 1 insertion(+)
> >>>
> >>> diff --git a/drivers/media/platform/mediatek/jpeg/mtk_jpeg_core.c b/drivers/media/platform/mediatek/jpeg/mtk_jpeg_core.c
> >>> index 0051f372a66c..6069ecf420b0 100644
> >>> --- a/drivers/media/platform/mediatek/jpeg/mtk_jpeg_core.c
> >>> +++ b/drivers/media/platform/mediatek/jpeg/mtk_jpeg_core.c
> >>> @@ -1816,6 +1816,7 @@ static void mtk_jpeg_remove(struct platform_device *pdev)
> >>>  {
> >>>       struct mtk_jpeg_dev *jpeg = platform_get_drvdata(pdev);
> >>>
> >>> +     cancel_delayed_work_sync(&jpeg->job_timeout_work);
> >>>       pm_runtime_disable(&pdev->dev);
> >>>       video_unregister_device(jpeg->vdev);
> >>>       v4l2_m2m_release(jpeg->m2m_dev);
> >>
> >> AFAICS, there is a fundamental problem here. The job_timeout_work uses
> >> v4l2_m2m_get_curr_priv() and at the time when driver module is unloaded,
> >> all the v4l contexts must be closed and released. Hence the
> >> v4l2_m2m_get_curr_priv() shall return NULL and crash the kernel when
> >> work is executed before cancel_delayed_work_sync().
> >>
> >
> > Hi Dmitry,
> >
> > Thanks for your reply. I think you're right. As m2m_dev is freed in
> > v4l2_m2m_release,
> > the invoking in v4l2_m2m_get_curr_priv might cause either UAF or null
> > pointer dereference
> > bug. I am sure that context is closed when we invoke mtk_jpeg_remove.
> > But I'm not sure if
> > context is released when mtk_jpegdec_timeout_work running.
> >
> >> At the time when mtk_jpeg_remove() is invoked, there shall be no
> >> job_timeout_work running in background because all jobs should be
> >> completed before context is released. If you'll look at
> >> v4l2_m2m_cancel_job(), you can see that it waits for the task completion
> >> before closing context.
> >
> > Yes, so I think the better way is to put the cancel_delayed_work_sync
> > invoking into
> > v4l2_m2m_ctx_release function?
>
> The v4l2_m2m_ctx_release() already should wait for the job_timeout_work
> completion or for the interrupt fire. Apparently it doesn't work in
> yours case. You'll need to debug why v4l job or job_timeout_work is
> running after v4l2_m2m_ctx_release(), it shouldn't happen.
>

Yes, v4l2_m2m_cancel_job waits for m2m_ctx->job_flags to be  ~TRANS_RUNNING,
the mtk_jpeg_job_timeout_work will finally invoke v4l2_m2m_job_finish
to trigger that.

However, this is not the only path to call v4l2_m2m_job_finish. Here
is a invoking chain:
v4l_streamon
  ->v4l2_m2m_ioctl_streamon
    ->v4l2_m2m_streamon
      ->v4l2_m2m_try_schedule
        ->v4l2_m2m_try_run
          ->mtk_jpeg_dec_device_run
            ->schedule_delayed_work(&jpeg->job_timeout_work...
            ->error path goto dec_end
            ->v4l2_m2m_job_finish

In some specific situation, it starts the worker and also calls
v4l2_m2m_job_finish, which might
make v4l2_m2m_cancel_job continues.

> The interrupt handler cancels job_timeout_work, you shouldn't need to
> flush the work.

It will, but as I said, there might be an early invocation chain to
start the work.(Not very sure)

>
> Technically, interrupt handler may race with job_timeout_work, but the
> timeout is set to 1 second and in practice should be difficult to
> trigger the race. The interrupt handler needs to be threaded, it should
> use cancel_delayed_work_sync() and check the return value of this function.
>

Yes, it's better to use cancel_delayed_work_sync here.

> >>
> >> You shouldn't be able to remove driver module while it has active/opened
> >> v4l contexts. If you can do that, then this is yours bug that needs to
> >> be fixed.
> >>
> >> In addition to this all, the job_timeout_work is initialized only for
> >> the single-core JPEG device. I'd expect this patch should crash
> >> multi-core JPEG devices.
> >>
> >
> > I think that's true. As I'm not familiar with the code here. Could you
> > please give me some advice about the patch?
>
> We'll need to understand why v4l2_m2m_ctx_release() doesn't work as
> expected before thinking about the patch.
>
> --
> Best regards,
> Dmitry
>