lkml.org 
[lkml]   [2018]   [Jun]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH 1/3] drm/v3d: Take a lock across GPU scheduler job creation and queuing.
From
Date
Am Dienstag, den 05.06.2018, 12:03 -0700 schrieb Eric Anholt:
> Between creation and queueing of a job, you need to prevent any other
> job from being created and queued.  Otherwise the scheduler's fences
> may be signaled out of seqno order.
>
> > Signed-off-by: Eric Anholt <eric@anholt.net>
> Fixes: 57692c94dcbe ("drm/v3d: Introduce a new DRM driver for Broadcom V3D V3.x+")
> ---
>
> ccing amd-gfx due to interaction of this series with the scheduler.
>
>  drivers/gpu/drm/v3d/v3d_drv.h |  5 +++++
>  drivers/gpu/drm/v3d/v3d_gem.c | 11 +++++++++--
>  2 files changed, 14 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/v3d/v3d_drv.h b/drivers/gpu/drm/v3d/v3d_drv.h
> index a043ac3aae98..26005abd9c5d 100644
> --- a/drivers/gpu/drm/v3d/v3d_drv.h
> +++ b/drivers/gpu/drm/v3d/v3d_drv.h
> @@ -85,6 +85,11 @@ struct v3d_dev {
> >    */
> >   struct mutex reset_lock;
>  
> > + /* Lock taken when creating and pushing the GPU scheduler
> > +  * jobs, to keep the sched-fence seqnos in order.
> > +  */
> > + struct mutex sched_lock;
> +
> >   struct {
> >   u32 num_allocated;
> >   u32 pages_allocated;
> diff --git a/drivers/gpu/drm/v3d/v3d_gem.c b/drivers/gpu/drm/v3d/v3d_gem.c
> index b513f9189caf..9ea83bdb9a30 100644
> --- a/drivers/gpu/drm/v3d/v3d_gem.c
> +++ b/drivers/gpu/drm/v3d/v3d_gem.c
> @@ -550,13 +550,16 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
> >   if (ret)
> >   goto fail;
>  
> > + mutex_lock(&v3d->sched_lock);
> >   if (exec->bin.start != exec->bin.end) {
> >   ret = drm_sched_job_init(&exec->bin.base,
> >    &v3d->queue[V3D_BIN].sched,
> >    &v3d_priv->sched_entity[V3D_BIN],
> >    v3d_priv);
> > - if (ret)
> > + if (ret) {
> > + mutex_unlock(&v3d->sched_lock);
>   goto fail_unreserve;

I don't see any path where you would go to fail_unreserve with the
mutex not yet locked, so you could just fold the mutex_unlock into this
error path for a bit less code duplication.

Otherwise this looks fine.

Regards,
Lucas

> + }
>  
> >   exec->bin_done_fence =
> >   dma_fence_get(&exec->bin.base.s_fence->finished);
> @@ -570,12 +573,15 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
> >    &v3d->queue[V3D_RENDER].sched,
> >    &v3d_priv->sched_entity[V3D_RENDER],
> >    v3d_priv);
> > - if (ret)
> > + if (ret) {
> > + mutex_unlock(&v3d->sched_lock);
> >   goto fail_unreserve;
> > + }
>  
> >   kref_get(&exec->refcount); /* put by scheduler job completion */
> >   drm_sched_entity_push_job(&exec->render.base,
> >     &v3d_priv->sched_entity[V3D_RENDER]);
> > + mutex_unlock(&v3d->sched_lock);
>  
> >   v3d_attach_object_fences(exec);
>  
> @@ -615,6 +621,7 @@ v3d_gem_init(struct drm_device *dev)
> >   spin_lock_init(&v3d->job_lock);
> >   mutex_init(&v3d->bo_lock);
> >   mutex_init(&v3d->reset_lock);
> > + mutex_init(&v3d->sched_lock);
>  
> >   /* Note: We don't allocate address 0.  Various bits of HW
> >    * treat 0 as special, such as the occlusion query counters

\
 
 \ /
  Last update: 2018-06-06 10:47    [W:0.039 / U:29.960 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site