[PATCH] drm/syncobj: make lockdep complain on WAIT_FOR_SUBMIT v3

[Daniel Vetter]

On Tue, Jan 19, 2021 at 12:48 PM Christian König
<ckoenig.leichtzumerken at gmail.com> wrote:
>
> DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT can't be used when we hold locks
> since we are basically waiting for userspace to do something.
>
> Holding a lock while doing so can trivial deadlock with page faults
> etc...
>
> So make lockdep complain when a driver tries to do this.
>
> v2: Add lockdep_assert_none_held() macro.
> v3: Add might_sleep() and also use lockdep_assert_none_held() in the
>     IOCTL path.
>
> Signed-off-by: Christian König <christian.koenig at amd.com>
> ---
>  drivers/gpu/drm/drm_syncobj.c | 12 ++++++++++++
>  include/linux/lockdep.h       |  5 +++++
>  2 files changed, 17 insertions(+)
>
> diff --git a/drivers/gpu/drm/drm_syncobj.c b/drivers/gpu/drm/drm_syncobj.c
> index 6e74e6745eca..9ecfec1e5567 100644
> --- a/drivers/gpu/drm/drm_syncobj.c
> +++ b/drivers/gpu/drm/drm_syncobj.c
> @@ -387,6 +387,15 @@ int drm_syncobj_find_fence(struct drm_file *file_private,
>         if (!syncobj)
>                 return -ENOENT;
>
> +       /* Waiting for userspace with locks help is illegal cause that can
> +        * trivial deadlock with page faults for example. Make lockdep complain
> +        * about it early on.
> +        */
> +       if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
> +               might_sleep();
> +               lockdep_assert_none_held_once();
> +       }
> +
>         *fence = drm_syncobj_fence_get(syncobj);
>         drm_syncobj_put(syncobj);

I noticed a use-after-free here that goes back to 5.2, sending out a
patch for that.

>
> @@ -940,6 +949,9 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
>         uint64_t *points;
>         uint32_t signaled_count, i;
>
> +       if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT)
> +               lockdep_assert_none_held_once();
> +
>         points = kmalloc_array(count, sizeof(*points), GFP_KERNEL);
>         if (points == NULL)
>                 return -ENOMEM;
> diff --git a/include/linux/lockdep.h b/include/linux/lockdep.h
> index b9e9adec73e8..6eb117c0d0f3 100644
> --- a/include/linux/lockdep.h
> +++ b/include/linux/lockdep.h
> @@ -310,6 +310,10 @@ extern void lock_unpin_lock(struct lockdep_map *lock, struct pin_cookie);
>                 WARN_ON_ONCE(debug_locks && !lockdep_is_held(l));       \
>         } while (0)
>
> +#define lockdep_assert_none_held_once()        do {                            \
> +               WARN_ON_ONCE(debug_locks && current->lockdep_depth);    \
> +       } while (0)
> +
>  #define lockdep_recursing(tsk) ((tsk)->lockdep_recursion)
>
>  #define lockdep_pin_lock(l)    lock_pin_lock(&(l)->dep_map)
> @@ -387,6 +391,7 @@ extern int lockdep_is_held(const void *);
>  #define lockdep_assert_held_write(l)   do { (void)(l); } while (0)
>  #define lockdep_assert_held_read(l)            do { (void)(l); } while (0)
>  #define lockdep_assert_held_once(l)            do { (void)(l); } while (0)
> +#define lockdep_assert_none_held_once()        do { } while (0)
>
>  #define lockdep_recursing(tsk)                 (0)

ofc needs ack from Peter, but drm parts look all good to me.

Reviewed-by: Daniel Vetter <daniel.vetter at ffwll.ch>

>
> --
> 2.25.1
>


--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch