Some drivers have hardware capability to get the precise timestamp of certain events based on which the fences are triggered. 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
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..8887c5a 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 * * 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 + * + * 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,
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
Signed-off-by: Veera Sundaram Sankaran veeras@codeaurora.org --- drivers/gpu/drm/drm_file.c | 43 +++++++++++++++++++++++++++++++++++++++++++ drivers/gpu/drm/drm_vblank.c | 9 ++++++++- include/drm/drm_file.h | 3 +++ 3 files changed, 54 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c index 0ac4566..5944bb9 100644 --- a/drivers/gpu/drm/drm_file.c +++ b/drivers/gpu/drm/drm_file.c @@ -775,6 +775,49 @@ void drm_event_cancel_free(struct drm_device *dev, EXPORT_SYMBOL(drm_event_cancel_free);
/** + * 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 + * + * 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_timestamp_locked(struct drm_device *dev, + struct drm_pending_event *e, ktime_t timestamp) +{ + assert_spin_locked(&dev->event_lock); + + if (e->completion) { + complete_all(e->completion); + e->completion_release(e->completion); + e->completion = NULL; + } + + if (e->fence) { + dma_fence_signal_timestamp(e->fence, timestamp); + dma_fence_put(e->fence); + } + + if (!e->file_priv) { + kfree(e); + return; + } + + list_del(&e->pending_link); + list_add_tail(&e->link, + &e->file_priv->event_list); + wake_up_interruptible(&e->file_priv->event_wait); +} +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 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);
On Thu, Jan 7, 2021 at 12:54 AM Veera Sundaram Sankaran veeras@codeaurora.org wrote:
Thanks for sending this out! One small note:
This seems to duplicate much of drm_send_event_locked(). Should a common backend function be used between them?
thanks -john
On Thu, Jan 7, 2021 at 12:53 AM Veera Sundaram Sankaran veeras@codeaurora.org wrote:
So, on quick review, this seems mostly sane. Though, it might be good to add some more detail about how the hardware timestamping fits into the kernel's CLOCK_MONOTONIC time domain.
I just want to make sure this interface isn't abused to jam raw hardware-domain timestamps into the fence->timestamp, causing the meaning or time-domain of the fence->timestamp to be unclear or inconsistent.
It may be useful to add an additional patch to the documentation around the dma_fence structure to make the timestamp field semantics more explicit and avoid confusion?
thanks -john
On 2021-01-08 11:55, John Stultz wrote:
Thanks for the comments. Sure, let me add more information in the commit-text about the HW timestamp conversion to kernel time-domain. Will explicitly mention the timestamp domain expected as part of the new dma_fence_signal_timestamp api documentation, since that would be the only place the timestamp would be set externally from drivers. On top of it, do suggest if still documentation on dma_fence struct would be required.
Thanks, Veera
thanks -john
dri-devel@lists.freedesktop.org
-
John Stultz -
Veera Sundaram Sankaran -
veeras＠codeaurora.org