[Intel-gfx] [PATCH 1/2] drm: Mark up accesses of vblank->enabled outside of its spinlock

[Ville Syrjälä]

On Thu, Mar 16, 2017 at 11:47:48PM +0000, Chris Wilson wrote:
> Order the update to vblank->enabled after the timestamp is primed so
> that a concurrent unlocked reader will only see the vblank->enabled with
> the current timestamp.
> 
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Ville Syrjälä <ville.syrjala at linux.intel.com>
> Cc: Daniel Vetter <daniel.vetter at ffwll.ch>
> ---
>  drivers/gpu/drm/drm_irq.c | 30 ++++++++++++++++++------------
>  1 file changed, 18 insertions(+), 12 deletions(-)
> 
> diff --git a/drivers/gpu/drm/drm_irq.c b/drivers/gpu/drm/drm_irq.c
> index 53a526c7b24d..4cc9352ab6a8 100644
> --- a/drivers/gpu/drm/drm_irq.c
> +++ b/drivers/gpu/drm/drm_irq.c
> @@ -336,10 +336,8 @@ static void vblank_disable_and_save(struct drm_device *dev, unsigned int pipe)
>  	 * calling the ->disable_vblank() operation in atomic context with the
>  	 * hardware potentially runtime suspended.
>  	 */
> -	if (vblank->enabled) {
> +	if (cmpxchg_relaxed(&vblank->enabled, true, false))
>  		__disable_vblank(dev, pipe);
> -		vblank->enabled = false;
> -	}
>  
>  	/*
>  	 * Always update the count and timestamp to maintain the
> @@ -360,7 +358,7 @@ static void vblank_disable_fn(unsigned long arg)
>  	unsigned long irqflags;
>  
>  	spin_lock_irqsave(&dev->vbl_lock, irqflags);
> -	if (atomic_read(&vblank->refcount) == 0 && vblank->enabled) {
> +	if (atomic_read(&vblank->refcount) == 0 && READ_ONCE(vblank->enabled)) {

Hmm. Aren't most of these accesses inside the lock? Looks like you're
marking everything READ/WRITE_ONCE()?

>  		DRM_DEBUG("disabling vblank on crtc %u\n", pipe);
>  		vblank_disable_and_save(dev, pipe);
>  	}
> @@ -384,7 +382,7 @@ void drm_vblank_cleanup(struct drm_device *dev)
>  	for (pipe = 0; pipe < dev->num_crtcs; pipe++) {
>  		struct drm_vblank_crtc *vblank = &dev->vblank[pipe];
>  
> -		WARN_ON(vblank->enabled &&
> +		WARN_ON(READ_ONCE(vblank->enabled) &&
>  			drm_core_check_feature(dev, DRIVER_MODESET));
>  
>  		del_timer_sync(&vblank->disable_timer);
> @@ -564,7 +562,7 @@ int drm_irq_uninstall(struct drm_device *dev)
>  		for (i = 0; i < dev->num_crtcs; i++) {
>  			struct drm_vblank_crtc *vblank = &dev->vblank[i];
>  
> -			if (!vblank->enabled)
> +			if (!READ_ONCE(vblank->enabled))
>  				continue;
>  
>  			WARN_ON(drm_core_check_feature(dev, DRIVER_MODESET));
> @@ -1105,11 +1103,16 @@ static int drm_vblank_enable(struct drm_device *dev, unsigned int pipe)
>  		 */
>  		ret = __enable_vblank(dev, pipe);
>  		DRM_DEBUG("enabling vblank on crtc %u, ret: %d\n", pipe, ret);
> -		if (ret)
> +		if (ret) {
>  			atomic_dec(&vblank->refcount);
> -		else {
> -			vblank->enabled = true;
> +		} else {
>  			drm_update_vblank_count(dev, pipe, 0);
> +			/* drm_update_vblank_count() includes a wmb so we just
> +			 * need to ensure that the compiler emits the write
> +			 * to mark the vblank as enabled after the call
> +			 * to drm_update_vblank_count().
> +			 */
> +			WRITE_ONCE(vblank->enabled, true);

Ordering+barrier looks correct to me.

>  		}
>  	}
>  
> @@ -1148,7 +1151,7 @@ static int drm_vblank_get(struct drm_device *dev, unsigned int pipe)
>  	if (atomic_add_return(1, &vblank->refcount) == 1) {
>  		ret = drm_vblank_enable(dev, pipe);
>  	} else {
> -		if (!vblank->enabled) {
> +		if (!READ_ONCE(vblank->enabled)) {
>  			atomic_dec(&vblank->refcount);
>  			ret = -EINVAL;
>  		}
> @@ -1517,7 +1520,7 @@ static int drm_queue_vblank_event(struct drm_device *dev, unsigned int pipe,
>  	 * vblank disable, so no need for further locking.  The reference from
>  	 * drm_vblank_get() protects against vblank disable from another source.
>  	 */
> -	if (!vblank->enabled) {
> +	if (!READ_ONCE(vblank->enabled)) {
>  		ret = -EINVAL;
>  		goto err_unlock;
>  	}
> @@ -1644,7 +1647,7 @@ int drm_wait_vblank(struct drm_device *dev, void *data,
>  		DRM_WAIT_ON(ret, vblank->queue, 3 * HZ,
>  			    (((drm_vblank_count(dev, pipe) -
>  			       vblwait->request.sequence) <= (1 << 23)) ||
> -			     !vblank->enabled ||
> +			     !READ_ONCE(vblank->enabled) ||
>  			     !dev->irq_enabled));
>  	}
>  
> @@ -1714,6 +1717,9 @@ bool drm_handle_vblank(struct drm_device *dev, unsigned int pipe)
>  	if (WARN_ON(pipe >= dev->num_crtcs))
>  		return false;
>  
> +	if (!READ_ONCE(vblank->enabled))
> +		return false;

This to me looks like it could theoretically cause us to
miss an interrupt.

1. enable_irq()
2. drm_update_vblank_count()
3. irq fires
4. drm_handle_vblank() doesn't do anything
5. enabled=true

> +
>  	spin_lock_irqsave(&dev->event_lock, irqflags);
>  
>  	/* Need timestamp lock to prevent concurrent execution with
> -- 
> 2.11.0

-- 
Ville Syrjälä
Intel OTC