[PATCH v3] media: hantro: fix use after free bug in hantro_remove due to race condition

[PATCH v3] media: hantro: fix use after free bug in hantro_remove due to race condition

[Zheng Wang]

In hantro_probe, vpu->watchdog_work is bound with
hantro_watchdog. Then hantro_end_prepare_run may
be called to start the work.

If we close the file or remove the module which will
call hantro_release and hantro_remove to make cleanup,
there may be an unfinished work. The possible sequence
is as follows, which will cause a typical UAF bug.

The same thing will happen in hantro_release, and use
ctx after freeing it.

Fix it by canceling the work before cleanup in hantro_release.

CPU0                  CPU1

                    |hantro_watchdog
hantro_remove     |
  v4l2_m2m_release  |
    kfree(m2m_dev); |
                    |
                    | v4l2_m2m_get_curr_priv
                    |   m2m_dev->curr_ctx //use

Signed-off-by: Zheng Wang <zyytlz.wz@163.com>
---
v3:
- use cancel_delayed_work_sync instead of cancel_delayed_work and add it to
hantro_release suggested by Hans Verkuil

v2:
- move the cancel-work-related code to hantro_remove suggested by Hans Verkuil
---
 drivers/media/platform/verisilicon/hantro_drv.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/media/platform/verisilicon/hantro_drv.c b/drivers/media/platform/verisilicon/hantro_drv.c
index b0aeedae7b65..86a4c0fa8c7d 100644
--- a/drivers/media/platform/verisilicon/hantro_drv.c
+++ b/drivers/media/platform/verisilicon/hantro_drv.c
@@ -597,6 +597,7 @@ static int hantro_release(struct file *filp)
 	struct hantro_ctx *ctx =
 		container_of(filp->private_data, struct hantro_ctx, fh);
 
+	cancel_delayed_work_sync(&ctx->dev->watchdog_work);
 	/*
 	 * No need for extra locking because this was the last reference
 	 * to this file.
@@ -1099,6 +1100,7 @@ static int hantro_remove(struct platform_device *pdev)
 
 	v4l2_info(&vpu->v4l2_dev, "Removing %s\n", pdev->name);
 
+	cancel_delayed_work_sync(&vpu->watchdog_work);
 	media_device_unregister(&vpu->mdev);
 	hantro_remove_dec_func(vpu);
 	hantro_remove_enc_func(vpu);
-- 
2.25.1

Re: [PATCH v3] media: hantro: fix use after free bug in hantro_remove due to race condition

[Ezequiel Garcia]

Hi Zheng,

On Mon, Mar 13, 2023 at 12:42 PM Zheng Wang <zyytlz.wz@163.com> wrote:
>
> In hantro_probe, vpu->watchdog_work is bound with
> hantro_watchdog. Then hantro_end_prepare_run may
> be called to start the work.
>
> If we close the file or remove the module which will
> call hantro_release and hantro_remove to make cleanup,

It's not possible to close the file or remove the module while a watchdog is
scheduled.

That's because the watchdog is active only during a mem2mem job,
and the file won't be closed until the job is done.

v4l2_m2m_ctx_release calls v4l2_m2m_cancel_jobw
which waits until the job is done.

If you can confirm it's possible to remove or close the file
while a job is running, that would be a driver bug.

Thanks for the patch, but it's not needed.

Regards,
Ezequiel

> there may be an unfinished work. The possible sequence
> is as follows, which will cause a typical UAF bug.
>
> The same thing will happen in hantro_release, and use
> ctx after freeing it.
>
> Fix it by canceling the work before cleanup in hantro_release.
>
> CPU0                  CPU1
>
>                     |hantro_watchdog
> hantro_remove     |
>   v4l2_m2m_release  |
>     kfree(m2m_dev); |
>                     |
>                     | v4l2_m2m_get_curr_priv
>                     |   m2m_dev->curr_ctx //use
>
> Signed-off-by: Zheng Wang <zyytlz.wz@163.com>
> ---
> v3:
> - use cancel_delayed_work_sync instead of cancel_delayed_work and add it to
> hantro_release suggested by Hans Verkuil
>
> v2:
> - move the cancel-work-related code to hantro_remove suggested by Hans Verkuil
> ---
>  drivers/media/platform/verisilicon/hantro_drv.c | 2 ++
>  1 file changed, 2 insertions(+)
>
> diff --git a/drivers/media/platform/verisilicon/hantro_drv.c b/drivers/media/platform/verisilicon/hantro_drv.c
> index b0aeedae7b65..86a4c0fa8c7d 100644
> --- a/drivers/media/platform/verisilicon/hantro_drv.c
> +++ b/drivers/media/platform/verisilicon/hantro_drv.c
> @@ -597,6 +597,7 @@ static int hantro_release(struct file *filp)
>         struct hantro_ctx *ctx =
>                 container_of(filp->private_data, struct hantro_ctx, fh);
>
> +       cancel_delayed_work_sync(&ctx->dev->watchdog_work);
>         /*
>          * No need for extra locking because this was the last reference
>          * to this file.
> @@ -1099,6 +1100,7 @@ static int hantro_remove(struct platform_device *pdev)
>
>         v4l2_info(&vpu->v4l2_dev, "Removing %s\n", pdev->name);
>
> +       cancel_delayed_work_sync(&vpu->watchdog_work);
>         media_device_unregister(&vpu->mdev);
>         hantro_remove_dec_func(vpu);
>         hantro_remove_enc_func(vpu);
> --
> 2.25.1
>

Re: [PATCH v3] media: hantro: fix use after free bug in hantro_remove due to race condition

[Zheng Hacker]

Ezequiel Garcia <ezequiel@vanguardiasur.com.ar> 于2023年3月31日周五 10:38写道：
>
> Hi Zheng,
>
> On Mon, Mar 13, 2023 at 12:42 PM Zheng Wang <zyytlz.wz@163.com> wrote:
> >
> > In hantro_probe, vpu->watchdog_work is bound with
> > hantro_watchdog. Then hantro_end_prepare_run may
> > be called to start the work.
> >
> > If we close the file or remove the module which will
> > call hantro_release and hantro_remove to make cleanup,
>
> It's not possible to close the file or remove the module while a watchdog is
> scheduled.
>
> That's because the watchdog is active only during a mem2mem job,
> and the file won't be closed until the job is done.
>
> v4l2_m2m_ctx_release calls v4l2_m2m_cancel_jobw
> which waits until the job is done.
>
> If you can confirm it's possible to remove or close the file
> while a job is running, that would be a driver bug.
>
> Thanks for the patch, but it's not needed.
>

Hi Ezequiel,

Thanks for your detailed analysis. Got it :)

Best regards,
Zheng

> Regards,
> Ezequiel
>
> > there may be an unfinished work. The possible sequence
> > is as follows, which will cause a typical UAF bug.
> >
> > The same thing will happen in hantro_release, and use
> > ctx after freeing it.
> >
> > Fix it by canceling the work before cleanup in hantro_release.
> >
> > CPU0                  CPU1
> >
> >                     |hantro_watchdog
> > hantro_remove     |
> >   v4l2_m2m_release  |
> >     kfree(m2m_dev); |
> >                     |
> >                     | v4l2_m2m_get_curr_priv
> >                     |   m2m_dev->curr_ctx //use
> >
> > Signed-off-by: Zheng Wang <zyytlz.wz@163.com>
> > ---
> > v3:
> > - use cancel_delayed_work_sync instead of cancel_delayed_work and add it to
> > hantro_release suggested by Hans Verkuil
> >
> > v2:
> > - move the cancel-work-related code to hantro_remove suggested by Hans Verkuil
> > ---
> >  drivers/media/platform/verisilicon/hantro_drv.c | 2 ++
> >  1 file changed, 2 insertions(+)
> >
> > diff --git a/drivers/media/platform/verisilicon/hantro_drv.c b/drivers/media/platform/verisilicon/hantro_drv.c
> > index b0aeedae7b65..86a4c0fa8c7d 100644
> > --- a/drivers/media/platform/verisilicon/hantro_drv.c
> > +++ b/drivers/media/platform/verisilicon/hantro_drv.c
> > @@ -597,6 +597,7 @@ static int hantro_release(struct file *filp)
> >         struct hantro_ctx *ctx =
> >                 container_of(filp->private_data, struct hantro_ctx, fh);
> >
> > +       cancel_delayed_work_sync(&ctx->dev->watchdog_work);
> >         /*
> >          * No need for extra locking because this was the last reference
> >          * to this file.
> > @@ -1099,6 +1100,7 @@ static int hantro_remove(struct platform_device *pdev)
> >
> >         v4l2_info(&vpu->v4l2_dev, "Removing %s\n", pdev->name);
> >
> > +       cancel_delayed_work_sync(&vpu->watchdog_work);
> >         media_device_unregister(&vpu->mdev);
> >         hantro_remove_dec_func(vpu);
> >         hantro_remove_enc_func(vpu);
> > --
> > 2.25.1
> >