[PATCH v4 1/2] dma-fence: allow signaling drivers to set fence timestamp

[PATCH v4 1/2] dma-fence: allow signaling drivers to set fence timestamp

[Veera Sundaram Sankaran]

Some drivers have hardware capability to get the precise HW timestamp
of certain events based on which the fences are triggered. The delta
between the event HW timestamp & current HW reference timestamp can
be used to calculate the timestamp in kernel's CLOCK_MONOTONIC time
domain. This allows it to set accurate timestamp factoring out any
software and IRQ latencies. Add a timestamp variant of fence signal
function, dma_fence_signal_timestamp to allow drivers to update the
precise timestamp for fences.

Changes in v2:
- Add a new fence signal variant instead of modifying fence struct

Changes in v3:
- Add timestamp domain information to commit-text and
dma_fence_signal_timestamp documentation

Signed-off-by: Veera Sundaram Sankaran <veeras@codeaurora.org>
---
 drivers/dma-buf/dma-fence.c | 70 ++++++++++++++++++++++++++++++++++++++++-----
 include/linux/dma-fence.h   |  3 ++
 2 files changed, 66 insertions(+), 7 deletions(-)

diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
index 7475e09..b83e9fa 100644
--- a/drivers/dma-buf/dma-fence.c
+++ b/drivers/dma-buf/dma-fence.c
@@ -312,22 +312,25 @@ void __dma_fence_might_wait(void)
 
 
 /**
- * dma_fence_signal_locked - signal completion of a fence
+ * dma_fence_signal_timestamp_locked - signal completion of a fence
  * @fence: the fence to signal
+ * @timestamp: fence signal timestamp in kernel's CLOCK_MONOTONIC time domain
  *
  * Signal completion for software callbacks on a fence, this will unblock
  * dma_fence_wait() calls and run all the callbacks added with
  * dma_fence_add_callback(). Can be called multiple times, but since a fence
  * can only go from the unsignaled to the signaled state and not back, it will
- * only be effective the first time.
+ * only be effective the first time. Set the timestamp provided as the fence
+ * signal timestamp.
  *
- * Unlike dma_fence_signal(), this function must be called with &dma_fence.lock
- * held.
+ * Unlike dma_fence_signal_timestamp(), this function must be called with
+ * &dma_fence.lock held.
  *
  * Returns 0 on success and a negative error value when @fence has been
  * signalled already.
  */
-int dma_fence_signal_locked(struct dma_fence *fence)
+int dma_fence_signal_timestamp_locked(struct dma_fence *fence,
+			ktime_t timestamp)
 {
 	struct dma_fence_cb *cur, *tmp;
 	struct list_head cb_list;
@@ -341,7 +344,7 @@ int dma_fence_signal_locked(struct dma_fence *fence)
 	/* Stash the cb_list before replacing it with the timestamp */
 	list_replace(&fence->cb_list, &cb_list);
 
-	fence->timestamp = ktime_get();
+	fence->timestamp = timestamp;
 	set_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &fence->flags);
 	trace_dma_fence_signaled(fence);
 
@@ -352,6 +355,59 @@ int dma_fence_signal_locked(struct dma_fence *fence)
 
 	return 0;
 }
+EXPORT_SYMBOL(dma_fence_signal_timestamp_locked);
+
+/**
+ * dma_fence_signal_timestamp - signal completion of a fence
+ * @fence: the fence to signal
+ * @timestamp: fence signal timestamp in kernel's CLOCK_MONOTONIC time domain
+ *
+ * Signal completion for software callbacks on a fence, this will unblock
+ * dma_fence_wait() calls and run all the callbacks added with
+ * dma_fence_add_callback(). Can be called multiple times, but since a fence
+ * can only go from the unsignaled to the signaled state and not back, it will
+ * only be effective the first time. Set the timestamp provided as the fence
+ * signal timestamp.
+ *
+ * Returns 0 on success and a negative error value when @fence has been
+ * signalled already.
+ */
+int dma_fence_signal_timestamp(struct dma_fence *fence, ktime_t timestamp)
+{
+	unsigned long flags;
+	int ret;
+
+	if (!fence)
+		return -EINVAL;
+
+	spin_lock_irqsave(fence->lock, flags);
+	ret = dma_fence_signal_timestamp_locked(fence, timestamp);
+	spin_unlock_irqrestore(fence->lock, flags);
+
+	return ret;
+}
+EXPORT_SYMBOL(dma_fence_signal_timestamp);
+
+/**
+ * dma_fence_signal_locked - signal completion of a fence
+ * @fence: the fence to signal
+ *
+ * Signal completion for software callbacks on a fence, this will unblock
+ * dma_fence_wait() calls and run all the callbacks added with
+ * dma_fence_add_callback(). Can be called multiple times, but since a fence
+ * can only go from the unsignaled to the signaled state and not back, it will
+ * only be effective the first time.
+ *
+ * Unlike dma_fence_signal(), this function must be called with &dma_fence.lock
+ * held.
+ *
+ * Returns 0 on success and a negative error value when @fence has been
+ * signalled already.
+ */
+int dma_fence_signal_locked(struct dma_fence *fence)
+{
+	return dma_fence_signal_timestamp_locked(fence, ktime_get());
+}
 EXPORT_SYMBOL(dma_fence_signal_locked);
 
 /**
@@ -379,7 +435,7 @@ int dma_fence_signal(struct dma_fence *fence)
 	tmp = dma_fence_begin_signalling();
 
 	spin_lock_irqsave(fence->lock, flags);
-	ret = dma_fence_signal_locked(fence);
+	ret = dma_fence_signal_timestamp_locked(fence, ktime_get());
 	spin_unlock_irqrestore(fence->lock, flags);
 
 	dma_fence_end_signalling(tmp);
diff --git a/include/linux/dma-fence.h b/include/linux/dma-fence.h
index 09e23ad..9f12efa 100644
--- a/include/linux/dma-fence.h
+++ b/include/linux/dma-fence.h
@@ -372,6 +372,9 @@ static inline void __dma_fence_might_wait(void) {}
 
 int dma_fence_signal(struct dma_fence *fence);
 int dma_fence_signal_locked(struct dma_fence *fence);
+int dma_fence_signal_timestamp(struct dma_fence *fence, ktime_t timestamp);
+int dma_fence_signal_timestamp_locked(struct dma_fence *fence,
+				      ktime_t timestamp);
 signed long dma_fence_default_wait(struct dma_fence *fence,
 				   bool intr, signed long timeout);
 int dma_fence_add_callback(struct dma_fence *fence,
-- 
2.7.4

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH v4 2/2] drm/drm_vblank: set the dma-fence timestamp during send_vblank_event

[Veera Sundaram Sankaran]

The explicit out-fences in crtc are signaled as part of vblank event,
indicating all framebuffers present on the Atomic Commit request are
scanned out on the screen. Though the fence signal and the vblank event
notification happens at the same time, triggered by the same hardware
vsync event, the timestamp set in both are different. With drivers
supporting precise vblank timestamp the difference between the two
timestamps would be even higher. This might have an impact on use-mode
frameworks using these fence timestamps for purposes other than simple
buffer usage. For instance, the Android framework [1] uses the
retire-fences as an alternative to vblank when frame-updates are in
progress. Set the fence timestamp during send vblank event using a new
drm_send_event_timestamp_locked variant to avoid discrepancies.

[1] https://android.googlesource.com/platform/frameworks/native/+/master/
services/surfaceflinger/Scheduler/Scheduler.cpp#397

Changes in v2:
- Use drm_send_event_timestamp_locked to update fence timestamp
- add more information to commit text

Changes in v3:
- use same backend helper function for variants of drm_send_event to
avoid code duplications

Changes in v4:
- remove WARN_ON from drm_send_event_timestamp_locked

Signed-off-by: Veera Sundaram Sankaran <veeras@codeaurora.org>
---
 drivers/gpu/drm/drm_file.c   | 69 ++++++++++++++++++++++++++++++++++++--------
 drivers/gpu/drm/drm_vblank.c |  9 +++++-
 include/drm/drm_file.h       |  3 ++
 3 files changed, 68 insertions(+), 13 deletions(-)

diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c
index 0ac4566..0f77f71 100644
--- a/drivers/gpu/drm/drm_file.c
+++ b/drivers/gpu/drm/drm_file.c
@@ -775,20 +775,19 @@ void drm_event_cancel_free(struct drm_device *dev,
 EXPORT_SYMBOL(drm_event_cancel_free);
 
 /**
- * drm_send_event_locked - send DRM event to file descriptor
+ * drm_send_event_helper - send DRM event to file descriptor
  * @dev: DRM device
  * @e: DRM event to deliver
+ * @timestamp: timestamp to set for the fence event in kernel's CLOCK_MONOTONIC
+ * time domain
  *
- * This function sends the event @e, initialized with drm_event_reserve_init(),
- * to its associated userspace DRM file. Callers must already hold
- * &drm_device.event_lock, see drm_send_event() for the unlocked version.
- *
- * Note that the core will take care of unlinking and disarming events when the
- * corresponding DRM file is closed. Drivers need not worry about whether the
- * DRM file for this event still exists and can call this function upon
- * completion of the asynchronous work unconditionally.
+ * This helper function sends the event @e, initialized with
+ * drm_event_reserve_init(), to its associated userspace DRM file.
+ * The timestamp variant of dma_fence_signal is used when the caller
+ * sends a valid timestamp.
  */
-void drm_send_event_locked(struct drm_device *dev, struct drm_pending_event *e)
+void drm_send_event_helper(struct drm_device *dev,
+			struct drm_pending_event *e, ktime_t timestamp)
 {
 	assert_spin_locked(&dev->event_lock);
 
@@ -799,7 +798,10 @@ void drm_send_event_locked(struct drm_device *dev, struct drm_pending_event *e)
 	}
 
 	if (e->fence) {
-		dma_fence_signal(e->fence);
+		if (timestamp)
+			dma_fence_signal_timestamp(e->fence, timestamp);
+		else
+			dma_fence_signal(e->fence);
 		dma_fence_put(e->fence);
 	}
 
@@ -814,6 +816,49 @@ void drm_send_event_locked(struct drm_device *dev, struct drm_pending_event *e)
 	wake_up_interruptible_poll(&e->file_priv->event_wait,
 		EPOLLIN | EPOLLRDNORM);
 }
+
+/**
+ * drm_send_event_timestamp_locked - send DRM event to file descriptor
+ * @dev: DRM device
+ * @e: DRM event to deliver
+ * @timestamp: timestamp to set for the fence event in kernel's CLOCK_MONOTONIC
+ * time domain
+ *
+ * This function sends the event @e, initialized with drm_event_reserve_init(),
+ * to its associated userspace DRM file. Callers must already hold
+ * &drm_device.event_lock.
+ *
+ * Note that the core will take care of unlinking and disarming events when the
+ * corresponding DRM file is closed. Drivers need not worry about whether the
+ * DRM file for this event still exists and can call this function upon
+ * completion of the asynchronous work unconditionally.
+ */
+void drm_send_event_timestamp_locked(struct drm_device *dev,
+			struct drm_pending_event *e, ktime_t timestamp)
+{
+	drm_send_event_helper(dev, e, timestamp);
+
+}
+EXPORT_SYMBOL(drm_send_event_timestamp_locked);
+
+/**
+ * drm_send_event_locked - send DRM event to file descriptor
+ * @dev: DRM device
+ * @e: DRM event to deliver
+ *
+ * This function sends the event @e, initialized with drm_event_reserve_init(),
+ * to its associated userspace DRM file. Callers must already hold
+ * &drm_device.event_lock, see drm_send_event() for the unlocked version.
+ *
+ * Note that the core will take care of unlinking and disarming events when the
+ * corresponding DRM file is closed. Drivers need not worry about whether the
+ * DRM file for this event still exists and can call this function upon
+ * completion of the asynchronous work unconditionally.
+ */
+void drm_send_event_locked(struct drm_device *dev, struct drm_pending_event *e)
+{
+	drm_send_event_helper(dev, e, 0);
+}
 EXPORT_SYMBOL(drm_send_event_locked);
 
 /**
@@ -836,7 +881,7 @@ void drm_send_event(struct drm_device *dev, struct drm_pending_event *e)
 	unsigned long irqflags;
 
 	spin_lock_irqsave(&dev->event_lock, irqflags);
-	drm_send_event_locked(dev, e);
+	drm_send_event_helper(dev, e, 0);
 	spin_unlock_irqrestore(&dev->event_lock, irqflags);
 }
 EXPORT_SYMBOL(drm_send_event);
diff --git a/drivers/gpu/drm/drm_vblank.c b/drivers/gpu/drm/drm_vblank.c
index f135b79..286edbe 100644
--- a/drivers/gpu/drm/drm_vblank.c
+++ b/drivers/gpu/drm/drm_vblank.c
@@ -1000,7 +1000,14 @@ static void send_vblank_event(struct drm_device *dev,
 		break;
 	}
 	trace_drm_vblank_event_delivered(e->base.file_priv, e->pipe, seq);
-	drm_send_event_locked(dev, &e->base);
+	/*
+	 * Use the same timestamp for any associated fence signal to avoid
+	 * mismatch in timestamps for vsync & fence events triggered by the
+	 * same HW event. Frameworks like SurfaceFlinger in Android expects the
+	 * retire-fence timestamp to match exactly with HW vsync as it uses it
+	 * for its software vsync modeling.
+	 */
+	drm_send_event_timestamp_locked(dev, &e->base, now);
 }
 
 /**
diff --git a/include/drm/drm_file.h b/include/drm/drm_file.h
index 716990b..b81b3bf 100644
--- a/include/drm/drm_file.h
+++ b/include/drm/drm_file.h
@@ -399,6 +399,9 @@ void drm_event_cancel_free(struct drm_device *dev,
 			   struct drm_pending_event *p);
 void drm_send_event_locked(struct drm_device *dev, struct drm_pending_event *e);
 void drm_send_event(struct drm_device *dev, struct drm_pending_event *e);
+void drm_send_event_timestamp_locked(struct drm_device *dev,
+				     struct drm_pending_event *e,
+				     ktime_t timestamp);
 
 struct file *mock_drm_getfile(struct drm_minor *minor, unsigned int flags);
 
-- 
2.7.4

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH v4 1/2] dma-fence: allow signaling drivers to set fence timestamp

[John Stultz]

On Fri, Jan 15, 2021 at 4:31 PM Veera Sundaram Sankaran
<veeras@codeaurora.org> wrote:
>
> Some drivers have hardware capability to get the precise HW timestamp
> of certain events based on which the fences are triggered. The delta
> between the event HW timestamp & current HW reference timestamp can
> be used to calculate the timestamp in kernel's CLOCK_MONOTONIC time
> domain. This allows it to set accurate timestamp factoring out any
> software and IRQ latencies. Add a timestamp variant of fence signal
> function, dma_fence_signal_timestamp to allow drivers to update the
> precise timestamp for fences.
>
> Changes in v2:
> - Add a new fence signal variant instead of modifying fence struct
>
> Changes in v3:
> - Add timestamp domain information to commit-text and
> dma_fence_signal_timestamp documentation
>
> Signed-off-by: Veera Sundaram Sankaran <veeras@codeaurora.org>
> ---
>  drivers/dma-buf/dma-fence.c | 70 ++++++++++++++++++++++++++++++++++++++++-----
>  include/linux/dma-fence.h   |  3 ++
>  2 files changed, 66 insertions(+), 7 deletions(-)

Thanks for respinning this!

Reviewed-by: John Stultz <john.stultz@linaro.org>

-john
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH v4 2/2] drm/drm_vblank: set the dma-fence timestamp during send_vblank_event

[John Stultz]

On Fri, Jan 15, 2021 at 4:31 PM Veera Sundaram Sankaran
<veeras@codeaurora.org> wrote:
>
> The explicit out-fences in crtc are signaled as part of vblank event,
> indicating all framebuffers present on the Atomic Commit request are
> scanned out on the screen. Though the fence signal and the vblank event
> notification happens at the same time, triggered by the same hardware
> vsync event, the timestamp set in both are different. With drivers
> supporting precise vblank timestamp the difference between the two
> timestamps would be even higher. This might have an impact on use-mode
> frameworks using these fence timestamps for purposes other than simple
> buffer usage. For instance, the Android framework [1] uses the
> retire-fences as an alternative to vblank when frame-updates are in
> progress. Set the fence timestamp during send vblank event using a new
> drm_send_event_timestamp_locked variant to avoid discrepancies.
>
> [1] https://android.googlesource.com/platform/frameworks/native/+/master/
> services/surfaceflinger/Scheduler/Scheduler.cpp#397
>
> Changes in v2:
> - Use drm_send_event_timestamp_locked to update fence timestamp
> - add more information to commit text
>
> Changes in v3:
> - use same backend helper function for variants of drm_send_event to
> avoid code duplications
>
> Changes in v4:
> - remove WARN_ON from drm_send_event_timestamp_locked
>
> Signed-off-by: Veera Sundaram Sankaran <veeras@codeaurora.org>
> ---

Looks good, as expected no longer seeing the warning.

Reviewed-by: John Stultz <john.stultz@linaro.org>

thanks
-john
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH v4 1/2] dma-fence: allow signaling drivers to set fence timestamp

[Sumit Semwal]

Hi Veera,

On Wed, 20 Jan 2021 at 02:00, John Stultz <john.stultz@linaro.org> wrote:
>
> On Fri, Jan 15, 2021 at 4:31 PM Veera Sundaram Sankaran
> <veeras@codeaurora.org> wrote:
> >
> > Some drivers have hardware capability to get the precise HW timestamp
> > of certain events based on which the fences are triggered. The delta
> > between the event HW timestamp & current HW reference timestamp can
> > be used to calculate the timestamp in kernel's CLOCK_MONOTONIC time
> > domain. This allows it to set accurate timestamp factoring out any
> > software and IRQ latencies. Add a timestamp variant of fence signal
> > function, dma_fence_signal_timestamp to allow drivers to update the
> > precise timestamp for fences.

Thank you for the patches!

Daniel, Christian, Gustavo: if there are no objections, I will queue
up this series via drm-misc-next.

> >
> > Changes in v2:
> > - Add a new fence signal variant instead of modifying fence struct
> >
> > Changes in v3:
> > - Add timestamp domain information to commit-text and
> > dma_fence_signal_timestamp documentation
> >
> > Signed-off-by: Veera Sundaram Sankaran <veeras@codeaurora.org>
> > ---
> >  drivers/dma-buf/dma-fence.c | 70 ++++++++++++++++++++++++++++++++++++++++-----
> >  include/linux/dma-fence.h   |  3 ++
> >  2 files changed, 66 insertions(+), 7 deletions(-)
>
> Thanks for respinning this!
>
> Reviewed-by: John Stultz <john.stultz@linaro.org>
>
> -john

Best,
Sumit.
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel