diff mbox

[10/11] drm: Use vblank timestamps to guesstimate how many vblanks were missed

Message ID 1442259832-23424-11-git-send-email-ville.syrjala@linux.intel.com (mailing list archive)
State New, archived
Headers show

Commit Message

Ville Syrjala Sept. 14, 2015, 7:43 p.m. UTC
From: Ville Syrjälä <ville.syrjala@linux.intel.com>

When lacking am accurate hardware frame counter, we can fall back to
using the vblank timestamps to guesstimagte how many vblanks have
elapsed since the last time the vblank counter was updated.

Take the oppostunity to unify the vblank_disable_and_save() and
drm_handle_vblank_events() to call the same function
(drm_update_vblank_count()) to perform the vblank updates.

If the hardware/driver has an accurate frame counter use it instead of
the timestamp based guesstimate. If the hardware/driver has neither
a frame counter nor acurate vblank timestamps, we fall back to assuming
that each drm_handle_vblank_events() should increment the vblank count
by one.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
---
 drivers/gpu/drm/drm_irq.c | 206 ++++++++++++++++++++--------------------------
 1 file changed, 91 insertions(+), 115 deletions(-)

Comments

Michel Dänzer Sept. 28, 2015, 2:57 a.m. UTC | #1
On 15.09.2015 04:43, ville.syrjala@linux.intel.com wrote:
> From: Ville Syrjälä <ville.syrjala@linux.intel.com>
> 
> When lacking am accurate hardware frame counter, we can fall back to
> using the vblank timestamps to guesstimagte how many vblanks have
> elapsed since the last time the vblank counter was updated.
> 
> Take the oppostunity to unify the vblank_disable_and_save() and
> drm_handle_vblank_events() to call the same function
> (drm_update_vblank_count()) to perform the vblank updates.

It would be nice to keep the drm_update_vblank_count unification
separate. As it is, it's very hard to keep track of which parts of the
patch are for each logical change.


BTW, I think the fact that I was hitting the problem fixed by 209e4dbc
("drm/vblank: Use u32 consistently for vblank counters") within a few
days indicates that there's another bug which causes the counter to jump
forward with drm_vblank_on/off(). It may not manifest itself with
current Intel hardware because that has a full 32-bit hardware frame
counter, turning the related calculations into no-ops. I haven't had
time to investigate this further yet.
Thierry Reding Sept. 30, 2015, 2:42 p.m. UTC | #2
On Mon, Sep 14, 2015 at 10:43:51PM +0300, ville.syrjala@linux.intel.com wrote:
[...]
> @@ -167,7 +238,7 @@ static void drm_update_vblank_count(struct drm_device *dev, unsigned int pipe,
>  	if (!rc)
>  		t_vblank = (struct timeval) {0, 0};
>  
> -	store_vblank(dev, pipe, diff, &t_vblank);
> +	store_vblank(dev, pipe, diff, &t_vblank, cur_vblank);

I think clearing the t_vblank is now wrong here. This breaks weston on
Tegra (and I think all other drivers that don't implement hardware
VBLANK timestamps). The reason is that if ->get_vblank_timestamp() isn't
implemented, rc will always be false at this point, hence resulting in a
zero VBLANK timestamp being reported to userspace.

In fact the comment, reproduced here because it's not in the patch
context:

	/*
	 * Only reinitialize corresponding vblank timestamp if high-precision query
	 * available and didn't fail. Otherwise reinitialize delayed at next vblank
	 * interrupt and assign 0 for now, to mark the vblanktimestamp as invalid.
	 */

is no longer accurate because this function is now called at the VBLANK
interrupt. If you want to keep invalidating timestamps for drivers that
implement VBLANK timestamp queries I think the condition needs to be
changed to something like:

	if (dev->get_vblank_timestamp && !rc)

and the comment updated to reflect the truth. Alternatively this code
can be removed, though I guess having the monotonic timestamp fallback
would be kind of missing the point.

I've tested either change and they both restore weston on Tegra.

Thierry
diff mbox

Patch

diff --git a/drivers/gpu/drm/drm_irq.c b/drivers/gpu/drm/drm_irq.c
index 07b0cb1..88fbee4 100644
--- a/drivers/gpu/drm/drm_irq.c
+++ b/drivers/gpu/drm/drm_irq.c
@@ -76,13 +76,15 @@  module_param_named(timestamp_monotonic, drm_timestamp_monotonic, int, 0600);
 
 static void store_vblank(struct drm_device *dev, unsigned int pipe,
 			 u32 vblank_count_inc,
-			 struct timeval *t_vblank)
+			 struct timeval *t_vblank, u32 last)
 {
 	struct drm_vblank_crtc *vblank = &dev->vblank[pipe];
 	u32 tslot;
 
 	assert_spin_locked(&dev->vblank_time_lock);
 
+	vblank->last = last;
+
 	/* All writers hold the spinlock, but readers are serialized by
 	 * the latching of vblank->count below.
 	 */
@@ -103,6 +105,54 @@  static void store_vblank(struct drm_device *dev, unsigned int pipe,
 }
 
 /**
+ * drm_reset_vblank_timestamp - reset the last timestamp to the last vblank
+ * @dev: DRM device
+ * @pipe: index of CRTC for which to reset the timestamp
+ *
+ * Reset the stored timestamp for the current vblank count to correspond
+ * to the last vblank occurred.
+ *
+ * Only to be called from drm_vblank_on().
+ *
+ * Note: caller must hold dev->vbl_lock since this reads & writes
+ * device vblank fields.
+ */
+static void drm_reset_vblank_timestamp(struct drm_device *dev, unsigned int pipe)
+{
+	u32 cur_vblank;
+	bool rc;
+	struct timeval t_vblank;
+	int count = DRM_TIMESTAMP_MAXRETRIES;
+
+	spin_lock(&dev->vblank_time_lock);
+
+	/*
+	 * sample the current counter to avoid random jumps
+	 * when drm_vblank_enable() applies the diff
+	 */
+	do {
+		cur_vblank = dev->driver->get_vblank_counter(dev, pipe);
+		rc = drm_get_last_vbltimestamp(dev, pipe, &t_vblank, 0);
+	} while (cur_vblank != dev->driver->get_vblank_counter(dev, pipe) && --count > 0);
+
+	/*
+	 * Only reinitialize corresponding vblank timestamp if high-precision query
+	 * available and didn't fail. Otherwise reinitialize delayed at next vblank
+	 * interrupt and assign 0 for now, to mark the vblanktimestamp as invalid.
+	 */
+	if (!rc)
+		t_vblank = (struct timeval) {0, 0};
+
+	/*
+	 * +1 to make sure user will never see the same
+	 * vblank counter value before and after a modeset
+	 */
+	store_vblank(dev, pipe, 1, &t_vblank, cur_vblank);
+
+	spin_unlock(&dev->vblank_time_lock);
+}
+
+/**
  * drm_update_vblank_count - update the master vblank counter
  * @dev: DRM device
  * @pipe: counter to update
@@ -126,6 +176,7 @@  static void drm_update_vblank_count(struct drm_device *dev, unsigned int pipe,
 	bool rc;
 	struct timeval t_vblank;
 	int count = DRM_TIMESTAMP_MAXRETRIES;
+	int framedur_ns = vblank->framedur_ns;
 
 	/*
 	 * Interrupts were disabled prior to this call, so deal with counter
@@ -144,20 +195,40 @@  static void drm_update_vblank_count(struct drm_device *dev, unsigned int pipe,
 		rc = drm_get_last_vbltimestamp(dev, pipe, &t_vblank, flags);
 	} while (cur_vblank != dev->driver->get_vblank_counter(dev, pipe) && --count > 0);
 
-	/* Deal with counter wrap */
-	diff = cur_vblank - vblank->last;
-	if (cur_vblank < vblank->last) {
-		diff += dev->max_vblank_count + 1;
+	if (dev->max_vblank_count != 0) {
+		/* trust the hw counter when it's around */
+		diff = (cur_vblank - vblank->last) & dev->max_vblank_count;
+	} else if (rc && framedur_ns) {
+		const struct timeval *t_old;
+		u64 diff_ns;
+
+		t_old = &vblanktimestamp(dev, pipe, vblank->count);
+		diff_ns = timeval_to_ns(&t_vblank) - timeval_to_ns(t_old);
 
-		DRM_DEBUG("last_vblank[%u]=0x%x, cur_vblank=0x%x => diff=0x%x\n",
-			  pipe, vblank->last, cur_vblank, diff);
+		/*
+		 * Figure out how many vblanks we've missed based
+		 * on the difference in the timestamps and the
+		 * frame/field duration.
+		 */
+		diff = DIV_ROUND_CLOSEST_ULL(diff_ns, framedur_ns);
+
+		if (diff == 0 && flags & DRM_CALLED_FROM_VBLIRQ)
+			DRM_DEBUG("crtc %u: Redundant vblirq ignored."
+				  " diff_ns = %lld, framedur_ns = %d)\n",
+				  pipe, (long long) diff_ns, framedur_ns);
+	} else {
+		/* some kind of default for drivers w/o accurate vbl timestamping */
+		diff = (flags & DRM_CALLED_FROM_VBLIRQ) != 0;
 	}
 
-	DRM_DEBUG("updating vblank count on crtc %u, missed %d\n",
-		  pipe, diff);
+	DRM_DEBUG("updating vblank count on crtc %u:"
+		  " current=%u, diff=%u, hw=%u hw_last=%u\n",
+		  pipe, vblank->count, diff, cur_vblank, vblank->last);
 
-	if (diff == 0)
+	if (diff == 0) {
+		WARN_ON_ONCE(cur_vblank != vblank->last);
 		return;
+	}
 
 	/*
 	 * Only reinitialize corresponding vblank timestamp if high-precision query
@@ -167,7 +238,7 @@  static void drm_update_vblank_count(struct drm_device *dev, unsigned int pipe,
 	if (!rc)
 		t_vblank = (struct timeval) {0, 0};
 
-	store_vblank(dev, pipe, diff, &t_vblank);
+	store_vblank(dev, pipe, diff, &t_vblank, cur_vblank);
 }
 
 /*
@@ -180,11 +251,6 @@  static void vblank_disable_and_save(struct drm_device *dev, unsigned int pipe)
 {
 	struct drm_vblank_crtc *vblank = &dev->vblank[pipe];
 	unsigned long irqflags;
-	u32 vblcount;
-	s64 diff_ns;
-	bool vblrc;
-	struct timeval tvblank;
-	int count = DRM_TIMESTAMP_MAXRETRIES;
 
 	/* Prevent vblank irq processing while disabling vblank irqs,
 	 * so no updates of timestamps or count can happen after we've
@@ -193,26 +259,6 @@  static void vblank_disable_and_save(struct drm_device *dev, unsigned int pipe)
 	spin_lock_irqsave(&dev->vblank_time_lock, irqflags);
 
 	/*
-	 * If the vblank interrupt was already disabled update the count
-	 * and timestamp to maintain the appearance that the counter
-	 * has been ticking all along until this time. This makes the
-	 * count account for the entire time between drm_vblank_on() and
-	 * drm_vblank_off().
-	 *
-	 * But only do this if precise vblank timestamps are available.
-	 * Otherwise we might read a totally bogus timestamp since drivers
-	 * lacking precise timestamp support rely upon sampling the system clock
-	 * at vblank interrupt time. Which obviously won't work out well if the
-	 * vblank interrupt is disabled.
-	 */
-	if (!vblank->enabled &&
-	    drm_get_last_vbltimestamp(dev, pipe, &tvblank, 0)) {
-		drm_update_vblank_count(dev, pipe, 0);
-		spin_unlock_irqrestore(&dev->vblank_time_lock, irqflags);
-		return;
-	}
-
-	/*
 	 * Only disable vblank interrupts if they're enabled. This avoids
 	 * calling the ->disable_vblank() operation in atomic context with the
 	 * hardware potentially runtime suspended.
@@ -222,47 +268,13 @@  static void vblank_disable_and_save(struct drm_device *dev, unsigned int pipe)
 		vblank->enabled = false;
 	}
 
-	/* No further vblank irq's will be processed after
-	 * this point. Get current hardware vblank count and
-	 * vblank timestamp, repeat until they are consistent.
-	 *
-	 * FIXME: There is still a race condition here and in
-	 * drm_update_vblank_count() which can cause off-by-one
-	 * reinitialization of software vblank counter. If gpu
-	 * vblank counter doesn't increment exactly at the leading
-	 * edge of a vblank interval, then we can lose 1 count if
-	 * we happen to execute between start of vblank and the
-	 * delayed gpu counter increment.
-	 */
-	do {
-		vblank->last = dev->driver->get_vblank_counter(dev, pipe);
-		vblrc = drm_get_last_vbltimestamp(dev, pipe, &tvblank, 0);
-	} while (vblank->last != dev->driver->get_vblank_counter(dev, pipe) && (--count) && vblrc);
-
-	if (!count)
-		vblrc = 0;
-
-	/* Compute time difference to stored timestamp of last vblank
-	 * as updated by last invocation of drm_handle_vblank() in vblank irq.
-	 */
-	vblcount = vblank->count;
-	diff_ns = timeval_to_ns(&tvblank) -
-		  timeval_to_ns(&vblanktimestamp(dev, pipe, vblcount));
-
-	/* If there is at least 1 msec difference between the last stored
-	 * timestamp and tvblank, then we are currently executing our
-	 * disable inside a new vblank interval, the tvblank timestamp
-	 * corresponds to this new vblank interval and the irq handler
-	 * for this vblank didn't run yet and won't run due to our disable.
-	 * Therefore we need to do the job of drm_handle_vblank() and
-	 * increment the vblank counter by one to account for this vblank.
-	 *
-	 * Skip this step if there isn't any high precision timestamp
-	 * available. In that case we can't account for this and just
-	 * hope for the best.
+	/*
+	 * Always update the count and timestamp to maintain the
+	 * appearance that the counter has been ticking all along until
+	 * this time. This makes the count account for the entire time
+	 * between drm_vblank_on() and drm_vblank_off().
 	 */
-	if (vblrc && (abs64(diff_ns) > 1000000))
-		store_vblank(dev, pipe, 1, &tvblank);
+	drm_update_vblank_count(dev, pipe, 0);
 
 	spin_unlock_irqrestore(&dev->vblank_time_lock, irqflags);
 }
@@ -1325,16 +1337,8 @@  void drm_vblank_on(struct drm_device *dev, unsigned int pipe)
 		vblank->inmodeset = 0;
 	}
 
-	/*
-	 * sample the current counter to avoid random jumps
-	 * when drm_vblank_enable() applies the diff
-	 *
-	 * -1 to make sure user will never see the same
-	 * vblank counter value before and after a modeset
-	 */
-	vblank->last =
-		(dev->driver->get_vblank_counter(dev, pipe) - 1) &
-		dev->max_vblank_count;
+	drm_reset_vblank_timestamp(dev, pipe);
+
 	/*
 	 * re-enable interrupts if there are users left, or the
 	 * user wishes vblank interrupts to be enabled all the time.
@@ -1717,9 +1721,6 @@  static void drm_handle_vblank_events(struct drm_device *dev, unsigned int pipe)
 bool drm_handle_vblank(struct drm_device *dev, unsigned int pipe)
 {
 	struct drm_vblank_crtc *vblank = &dev->vblank[pipe];
-	u32 vblcount;
-	s64 diff_ns;
-	struct timeval tvblank;
 	unsigned long irqflags;
 
 	if (WARN_ON_ONCE(!dev->num_crtcs))
@@ -1743,32 +1744,7 @@  bool drm_handle_vblank(struct drm_device *dev, unsigned int pipe)
 		return false;
 	}
 
-	/* Fetch corresponding timestamp for this vblank interval from
-	 * driver and store it in proper slot of timestamp ringbuffer.
-	 */
-
-	/* Get current timestamp and count. */
-	vblcount = vblank->count;
-	drm_get_last_vbltimestamp(dev, pipe, &tvblank, DRM_CALLED_FROM_VBLIRQ);
-
-	/* Compute time difference to timestamp of last vblank */
-	diff_ns = timeval_to_ns(&tvblank) -
-		  timeval_to_ns(&vblanktimestamp(dev, pipe, vblcount));
-
-	/* Update vblank timestamp and count if at least
-	 * DRM_REDUNDANT_VBLIRQ_THRESH_NS nanoseconds
-	 * difference between last stored timestamp and current
-	 * timestamp. A smaller difference means basically
-	 * identical timestamps. Happens if this vblank has
-	 * been already processed and this is a redundant call,
-	 * e.g., due to spurious vblank interrupts. We need to
-	 * ignore those for accounting.
-	 */
-	if (abs64(diff_ns) > DRM_REDUNDANT_VBLIRQ_THRESH_NS)
-		store_vblank(dev, pipe, 1, &tvblank);
-	else
-		DRM_DEBUG("crtc %u: Redundant vblirq ignored. diff_ns = %d\n",
-			  pipe, (int) diff_ns);
+	drm_update_vblank_count(dev, pipe, DRM_CALLED_FROM_VBLIRQ);
 
 	spin_unlock(&dev->vblank_time_lock);