drm/vblank: Add support for precise vblank timestamping.

The DRI2 swap & sync implementation needs precise
vblank counts and precise timestamps corresponding
to those vblank counts. For conformance to the OpenML
OML_sync_control extension specification the DRM
timestamp associated with a vblank count should
correspond to the start of video scanout of the first
scanline of the video frame following the vblank
interval for that vblank count.

Therefore we need to carry around precise timestamps
for vblanks. Currently the DRM and KMS drivers generate
timestamps ad-hoc via do_gettimeofday() in some
places. The resulting timestamps are sometimes not
very precise due to interrupt handling delays, they
don't conform to OML_sync_control and some are wrong,
as they aren't taken synchronized to the vblank.

This patch implements support inside the drm core
for precise and robust timestamping. It consists
of the following interrelated pieces.

1. Vblank timestamp caching:

A per-crtc ringbuffer stores the most recent vblank
timestamps corresponding to vblank counts.

The ringbuffer can be read out lock-free via the
accessor function:

struct timeval timestamp;
vblankcount = drm_vblank_count_and_time(dev, crtcid, &timestamp).

The function returns the current vblank count and
the corresponding timestamp for start of video
scanout following the vblank interval. It can be
used anywhere between enclosing drm_vblank_get(dev, crtcid)
and drm_vblank_put(dev,crtcid) statements. It is used
inside the drmWaitVblank ioctl and in the vblank event
queueing and handling. It should be used by kms drivers for
timestamping of bufferswap completion.

The timestamp ringbuffer is reinitialized each time
vblank irq's get reenabled in drm_vblank_get()/
drm_update_vblank_count(). It is invalidated when
vblank irq's get disabled.

The ringbuffer is updated inside drm_handle_vblank()
at each vblank irq.

2. Calculation of precise vblank timestamps:

drm_get_last_vbltimestamp() is used to compute the
timestamp for the end of the most recent vblank (if
inside active scanout), or the expected end of the
current vblank interval (if called inside a vblank
interval). The function calls into a new optional kms
driver entry point dev->driver->get_vblank_timestamp()
which is supposed to provide the precise timestamp.
If a kms driver doesn't implement the entry point or
if the call fails, a simple do_gettimeofday() timestamp
is returned as crude approximation of the true vblank time.

A new drm module parameter drm.timestamp_precision_usec
allows to disable high precision timestamps (if set to
zero) or to specify the maximum acceptable error in
the timestamps in microseconds.

Kms drivers could implement their get_vblank_timestamp()
function in a gpu specific way, as long as returned
timestamps conform to OML_sync_control, e.g., by use
of gpu specific hardware timestamps.

Optionally, kms drivers can simply wrap and use the new
utility function drm_calc_vbltimestamp_from_scanoutpos().
This function calls a new optional kms driver function
dev->driver->get_scanout_position() which returns the
current horizontal and vertical video scanout position
of the crtc. The scanout position together with the
drm_display_timing of the current video mode is used
to calculate elapsed time relative to start of active scanout
for the current video frame. This elapsed time is subtracted
from the current do_gettimeofday() time to get the timestamp
corresponding to start of video scanout. Currently
non-interlaced, non-doublescan video modes, with or
without panel scaling are handled correctly. Interlaced/
doublescan modes are tbd in a future patch.

3. Filtering of redundant vblank irq's and removal of
some race-conditions in the vblank irq enable/disable path:

Some gpu's (e.g., Radeon R500/R600) send spurious vblank
irq's outside the vblank if vblank irq's get reenabled.
These get detected by use of the vblank timestamps and
filtered out to avoid miscounting of vblanks.

Some race-conditions between the vblank irq enable/disable
functions, the vblank irq handler and the gpu itself (updating
its hardware vblank counter in the "wrong" moment) are
fixed inside vblank_disable_and_save() and
drm_update_vblank_count() by use of the vblank timestamps and
a new spinlock dev->vblank_time_lock.

The time until vblank irq disable is now configurable via
a new drm module parameter drm.vblankoffdelay to allow
experimentation with timeouts that are much shorter than
the current 5 seconds and should allow longer vblank off
periods for better power savings.

Followup patches will use these new functions to
implement precise timestamping for the intel and radeon
kms drivers.

Signed-off-by: Mario Kleiner <mario.kleiner@tuebingen.mpg.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>

5 files changed, 662 insertions, 26 deletions

diff --git a/drivers/gpu/drm/drm_crtc_helper.c b/drivers/gpu/drm/drm_crtc_helper.c
index f7af91cb273d..4c200931a6bc 100644
--- a/drivers/gpu/drm/drm_crtc_helper.c
+++ b/drivers/gpu/drm/drm_crtc_helper.c
@@ -336,7 +336,7 @@ bool drm_crtc_helper_set_mode(struct drm_crtc *crtc,
                               struct drm_framebuffer *old_fb)
 {
         struct drm_device *dev = crtc->dev;
-        struct drm_display_mode *adjusted_mode, saved_mode;
+        struct drm_display_mode *adjusted_mode, saved_mode, saved_hwmode;
         struct drm_crtc_helper_funcs *crtc_funcs = crtc->helper_private;
         struct drm_encoder_helper_funcs *encoder_funcs;
         int saved_x, saved_y;
@@ -350,6 +350,7 @@ bool drm_crtc_helper_set_mode(struct drm_crtc *crtc,
         if (!crtc->enabled)
                 return true;
 
+        saved_hwmode = crtc->hwmode;
         saved_mode = crtc->mode;
         saved_x = crtc->x;
         saved_y = crtc->y;
@@ -427,11 +428,21 @@ bool drm_crtc_helper_set_mode(struct drm_crtc *crtc,
 
         }
 
+        /* Store real post-adjustment hardware mode. */
+        crtc->hwmode = *adjusted_mode;
+
+        /* Calculate and store various constants which
+         * are later needed by vblank and swap-completion
+         * timestamping. They are derived from true hwmode.
+         */
+        drm_calc_timestamping_constants(crtc);
+
         /* XXX free adjustedmode */
         drm_mode_destroy(dev, adjusted_mode);
         /* FIXME: add subpixel order */
 done:
         if (!ret) {
+                crtc->hwmode = saved_hwmode;
                 crtc->mode = saved_mode;
                 crtc->x = saved_x;
                 crtc->y = saved_y;
diff --git a/drivers/gpu/drm/drm_irq.c b/drivers/gpu/drm/drm_irq.c
index 9d3a5030b6e1..4e82d0d3c378 100644
--- a/drivers/gpu/drm/drm_irq.c
+++ b/drivers/gpu/drm/drm_irq.c
@@ -40,6 +40,22 @@
 #include <linux/slab.h>
 
 #include <linux/vgaarb.h>
+
+/* Access macro for slots in vblank timestamp ringbuffer. */
+#define vblanktimestamp(dev, crtc, count) ( \
+        (dev)->_vblank_time[(crtc) * DRM_VBLANKTIME_RBSIZE + \
+        ((count) % DRM_VBLANKTIME_RBSIZE)])
+
+/* Retry timestamp calculation up to 3 times to satisfy
+ * drm_timestamp_precision before giving up.
+ */
+#define DRM_TIMESTAMP_MAXRETRIES 3
+
+/* Threshold in nanoseconds for detection of redundant
+ * vblank irq in drm_handle_vblank(). 1 msec should be ok.
+ */
+#define DRM_REDUNDANT_VBLIRQ_THRESH_NS 1000000
+
 /**
  * Get interrupt from bus id.
  *
@@ -77,6 +93,87 @@ int drm_irq_by_busid(struct drm_device *dev, void *data,
         return 0;
 }
 
+/*
+ * Clear vblank timestamp buffer for a crtc.
+ */
+static void clear_vblank_timestamps(struct drm_device *dev, int crtc)
+{
+        memset(&dev->_vblank_time[crtc * DRM_VBLANKTIME_RBSIZE], 0,
+                DRM_VBLANKTIME_RBSIZE * sizeof(struct timeval));
+}
+
+/*
+ * Disable vblank irq's on crtc, make sure that last vblank count
+ * of hardware and corresponding consistent software vblank counter
+ * are preserved, even if there are any spurious vblank irq's after
+ * disable.
+ */
+static void vblank_disable_and_save(struct drm_device *dev, int crtc)
+{
+        unsigned long irqflags;
+        u32 vblcount;
+        s64 diff_ns;
+        int vblrc;
+        struct timeval tvblank;
+
+        /* Prevent vblank irq processing while disabling vblank irqs,
+         * so no updates of timestamps or count can happen after we've
+         * disabled. Needed to prevent races in case of delayed irq's.
+         * Disable preemption, so vblank_time_lock is held as short as
+         * possible, even under a kernel with PREEMPT_RT patches.
+         */
+        preempt_disable();
+        spin_lock_irqsave(&dev->vblank_time_lock, irqflags);
+
+        dev->driver->disable_vblank(dev, crtc);
+        dev->vblank_enabled[crtc] = 0;
+
+        /* 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 {
+                dev->last_vblank[crtc] = dev->driver->get_vblank_counter(dev, crtc);
+                vblrc = drm_get_last_vbltimestamp(dev, crtc, &tvblank, 0);
+        } while (dev->last_vblank[crtc] != dev->driver->get_vblank_counter(dev, crtc));
+
+        /* Compute time difference to stored timestamp of last vblank
+         * as updated by last invocation of drm_handle_vblank() in vblank irq.
+         */
+        vblcount = atomic_read(&dev->_vblank_count[crtc]);
+        diff_ns = timeval_to_ns(&tvblank) -
+                  timeval_to_ns(&vblanktimestamp(dev, crtc, 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.
+         */
+        if ((vblrc > 0) && (abs(diff_ns) > 1000000))
+                atomic_inc(&dev->_vblank_count[crtc]);
+
+        /* Invalidate all timestamps while vblank irq's are off. */
+        clear_vblank_timestamps(dev, crtc);
+
+        spin_unlock_irqrestore(&dev->vblank_time_lock, irqflags);
+        preempt_enable();
+}
+
 static void vblank_disable_fn(unsigned long arg)
 {
         struct drm_device *dev = (struct drm_device *)arg;
@@ -91,10 +188,7 @@ static void vblank_disable_fn(unsigned long arg)
                 if (atomic_read(&dev->vblank_refcount[i]) == 0 &&
                     dev->vblank_enabled[i]) {
                         DRM_DEBUG("disabling vblank on crtc %d\n", i);
-                        dev->last_vblank[i] =
-                                dev->driver->get_vblank_counter(dev, i);
-                        dev->driver->disable_vblank(dev, i);
-                        dev->vblank_enabled[i] = 0;
+                        vblank_disable_and_save(dev, i);
                 }
                 spin_unlock_irqrestore(&dev->vbl_lock, irqflags);
         }
@@ -117,6 +211,7 @@ void drm_vblank_cleanup(struct drm_device *dev)
         kfree(dev->last_vblank);
         kfree(dev->last_vblank_wait);
         kfree(dev->vblank_inmodeset);
+        kfree(dev->_vblank_time);
 
         dev->num_crtcs = 0;
 }
@@ -129,6 +224,8 @@ int drm_vblank_init(struct drm_device *dev, int num_crtcs)
         setup_timer(&dev->vblank_disable_timer, vblank_disable_fn,
                     (unsigned long)dev);
         spin_lock_init(&dev->vbl_lock);
+        spin_lock_init(&dev->vblank_time_lock);
+
         dev->num_crtcs = num_crtcs;
 
         dev->vbl_queue = kmalloc(sizeof(wait_queue_head_t) * num_crtcs,
@@ -161,6 +258,19 @@ int drm_vblank_init(struct drm_device *dev, int num_crtcs)
         if (!dev->vblank_inmodeset)
                 goto err;
 
+        dev->_vblank_time = kcalloc(num_crtcs * DRM_VBLANKTIME_RBSIZE,
+                                    sizeof(struct timeval), GFP_KERNEL);
+        if (!dev->_vblank_time)
+                goto err;
+
+        DRM_INFO("Supports vblank timestamp caching Rev 1 (10.10.2010).\n");
+
+        /* Driver specific high-precision vblank timestamping supported? */
+        if (dev->driver->get_vblank_timestamp)
+                DRM_INFO("Driver supports precise vblank timestamp query.\n");
+        else
+                DRM_INFO("No driver support for vblank timestamp query.\n");
+
         /* Zero per-crtc vblank stuff */
         for (i = 0; i < num_crtcs; i++) {
                 init_waitqueue_head(&dev->vbl_queue[i]);
@@ -279,7 +389,7 @@ EXPORT_SYMBOL(drm_irq_install);
  *
  * Calls the driver's \c drm_driver_irq_uninstall() function, and stops the irq.
  */
-int drm_irq_uninstall(struct drm_device * dev)
+int drm_irq_uninstall(struct drm_device *dev)
 {
         unsigned long irqflags;
         int irq_enabled, i;
@@ -335,7 +445,9 @@ int drm_control(struct drm_device *dev, void *data,
 {
         struct drm_control *ctl = data;
 
-        /* if we haven't irq we fallback for compatibility reasons - this used to be a separate function in drm_dma.h */
+        /* if we haven't irq we fallback for compatibility reasons -
+         * this used to be a separate function in drm_dma.h
+         */
 
 
         switch (ctl->func) {
@@ -360,6 +472,287 @@ int drm_control(struct drm_device *dev, void *data,
 }
 
 /**
+ * drm_calc_timestamping_constants - Calculate and
+ * store various constants which are later needed by
+ * vblank and swap-completion timestamping, e.g, by
+ * drm_calc_vbltimestamp_from_scanoutpos().
+ * They are derived from crtc's true scanout timing,
+ * so they take things like panel scaling or other
+ * adjustments into account.
+ *
+ * @crtc drm_crtc whose timestamp constants should be updated.
+ *
+ */
+void drm_calc_timestamping_constants(struct drm_crtc *crtc)
+{
+        s64 linedur_ns = 0, pixeldur_ns = 0, framedur_ns = 0;
+        u64 dotclock;
+
+        /* Dot clock in Hz: */
+        dotclock = (u64) crtc->hwmode.clock * 1000;
+
+        /* Valid dotclock? */
+        if (dotclock > 0) {
+                /* Convert scanline length in pixels and video dot clock to
+                 * line duration, frame duration and pixel duration in
+                 * nanoseconds:
+                 */
+                pixeldur_ns = (s64) div64_u64(1000000000, dotclock);
+                linedur_ns  = (s64) div64_u64(((u64) crtc->hwmode.crtc_htotal *
+                                              1000000000), dotclock);
+                framedur_ns = (s64) crtc->hwmode.crtc_vtotal * linedur_ns;
+        } else
+                DRM_ERROR("crtc %d: Can't calculate constants, dotclock = 0!\n",
+                          crtc->base.id);
+
+        crtc->pixeldur_ns = pixeldur_ns;
+        crtc->linedur_ns  = linedur_ns;
+        crtc->framedur_ns = framedur_ns;
+
+        DRM_DEBUG("crtc %d: hwmode: htotal %d, vtotal %d, vdisplay %d\n",
+                  crtc->base.id, crtc->hwmode.crtc_htotal,
+                  crtc->hwmode.crtc_vtotal, crtc->hwmode.crtc_vdisplay);
+        DRM_DEBUG("crtc %d: clock %d kHz framedur %d linedur %d, pixeldur %d\n",
+                  crtc->base.id, (int) dotclock/1000, (int) framedur_ns,
+                  (int) linedur_ns, (int) pixeldur_ns);
+}
+EXPORT_SYMBOL(drm_calc_timestamping_constants);
+
+/**
+ * drm_calc_vbltimestamp_from_scanoutpos - helper routine for kms
+ * drivers. Implements calculation of exact vblank timestamps from
+ * given drm_display_mode timings and current video scanout position
+ * of a crtc. This can be called from within get_vblank_timestamp()
+ * implementation of a kms driver to implement the actual timestamping.
+ *
+ * Should return timestamps conforming to the OML_sync_control OpenML
+ * extension specification. The timestamp corresponds to the end of
+ * the vblank interval, aka start of scanout of topmost-leftmost display
+ * pixel in the following video frame.
+ *
+ * Requires support for optional dev->driver->get_scanout_position()
+ * in kms driver, plus a bit of setup code to provide a drm_display_mode
+ * that corresponds to the true scanout timing.
+ *
+ * The current implementation only handles standard video modes. It
+ * returns as no operation if a doublescan or interlaced video mode is
+ * active. Higher level code is expected to handle this.
+ *
+ * @dev: DRM device.
+ * @crtc: Which crtc's vblank timestamp to retrieve.
+ * @max_error: Desired maximum allowable error in timestamps (nanosecs).
+ *             On return contains true maximum error of timestamp.
+ * @vblank_time: Pointer to struct timeval which should receive the timestamp.
+ * @flags: Flags to pass to driver:
+ *         0 = Default.
+ *         DRM_CALLED_FROM_VBLIRQ = If function is called from vbl irq handler.
+ * @refcrtc: drm_crtc* of crtc which defines scanout timing.
+ *
+ * Returns negative value on error, failure or if not supported in current
+ * video mode:
+ *
+ * -EINVAL   - Invalid crtc.
+ * -EAGAIN   - Temporary unavailable, e.g., called before initial modeset.
+ * -ENOTSUPP - Function not supported in current display mode.
+ * -EIO      - Failed, e.g., due to failed scanout position query.
+ *
+ * Returns or'ed positive status flags on success:
+ *
+ * DRM_VBLANKTIME_SCANOUTPOS_METHOD - Signal this method used for timestamping.
+ * DRM_VBLANKTIME_INVBL - Timestamp taken while scanout was in vblank interval.
+ *
+ */
+int drm_calc_vbltimestamp_from_scanoutpos(struct drm_device *dev, int crtc,
+                                          int *max_error,
+                                          struct timeval *vblank_time,
+                                          unsigned flags,
+                                          struct drm_crtc *refcrtc)
+{
+        struct timeval stime, raw_time;
+        struct drm_display_mode *mode;
+        int vbl_status, vtotal, vdisplay;
+        int vpos, hpos, i;
+        s64 framedur_ns, linedur_ns, pixeldur_ns, delta_ns, duration_ns;
+        bool invbl;
+
+        if (crtc < 0 || crtc >= dev->num_crtcs) {
+                DRM_ERROR("Invalid crtc %d\n", crtc);
+                return -EINVAL;
+        }
+
+        /* Scanout position query not supported? Should not happen. */
+        if (!dev->driver->get_scanout_position) {
+                DRM_ERROR("Called from driver w/o get_scanout_position()!?\n");
+                return -EIO;
+        }
+
+        mode = &refcrtc->hwmode;
+        vtotal = mode->crtc_vtotal;
+        vdisplay = mode->crtc_vdisplay;
+
+        /* Durations of frames, lines, pixels in nanoseconds. */
+        framedur_ns = refcrtc->framedur_ns;
+        linedur_ns  = refcrtc->linedur_ns;
+        pixeldur_ns = refcrtc->pixeldur_ns;
+
+        /* If mode timing undefined, just return as no-op:
+         * Happens during initial modesetting of a crtc.
+         */
+        if (vtotal <= 0 || vdisplay <= 0 || framedur_ns == 0) {
+                DRM_DEBUG("crtc %d: Noop due to uninitialized mode.\n", crtc);
+                return -EAGAIN;
+        }
+
+        /* Don't know yet how to handle interlaced or
+         * double scan modes. Just no-op for now.
+         */
+        if (mode->flags & (DRM_MODE_FLAG_INTERLACE | DRM_MODE_FLAG_DBLSCAN)) {
+                DRM_DEBUG("crtc %d: Noop due to unsupported mode.\n", crtc);
+                return -ENOTSUPP;
+        }
+
+        /* Get current scanout position with system timestamp.
+         * Repeat query up to DRM_TIMESTAMP_MAXRETRIES times
+         * if single query takes longer than max_error nanoseconds.
+         *
+         * This guarantees a tight bound on maximum error if
+         * code gets preempted or delayed for some reason.
+         */
+        for (i = 0; i < DRM_TIMESTAMP_MAXRETRIES; i++) {
+                /* Disable preemption to make it very likely to
+                 * succeed in the first iteration even on PREEMPT_RT kernel.
+                 */
+                preempt_disable();
+
+                /* Get system timestamp before query. */
+                do_gettimeofday(&stime);
+
+                /* Get vertical and horizontal scanout pos. vpos, hpos. */
+                vbl_status = dev->driver->get_scanout_position(dev, crtc, &vpos, &hpos);
+
+                /* Get system timestamp after query. */
+                do_gettimeofday(&raw_time);
+
+                preempt_enable();
+
+                /* Return as no-op if scanout query unsupported or failed. */
+                if (!(vbl_status & DRM_SCANOUTPOS_VALID)) {
+                        DRM_DEBUG("crtc %d : scanoutpos query failed [%d].\n",
+                                  crtc, vbl_status);
+                        return -EIO;
+                }
+
+                duration_ns = timeval_to_ns(&raw_time) - timeval_to_ns(&stime);
+
+                /* Accept result with <  max_error nsecs timing uncertainty. */
+                if (duration_ns <= (s64) *max_error)
+                        break;
+        }
+
+        /* Noisy system timing? */
+        if (i == DRM_TIMESTAMP_MAXRETRIES) {
+                DRM_DEBUG("crtc %d: Noisy timestamp %d us > %d us [%d reps].\n",
+                          crtc, (int) duration_ns/1000, *max_error/1000, i);
+        }
+
+        /* Return upper bound of timestamp precision error. */
+        *max_error = (int) duration_ns;
+
+        /* Check if in vblank area:
+         * vpos is >=0 in video scanout area, but negative
+         * within vblank area, counting down the number of lines until
+         * start of scanout.
+         */
+        invbl = vbl_status & DRM_SCANOUTPOS_INVBL;
+
+        /* Convert scanout position into elapsed time at raw_time query
+         * since start of scanout at first display scanline. delta_ns
+         * can be negative if start of scanout hasn't happened yet.
+         */
+        delta_ns = (s64) vpos * linedur_ns + (s64) hpos * pixeldur_ns;
+
+        /* Is vpos outside nominal vblank area, but less than
+         * 1/100 of a frame height away from start of vblank?
+         * If so, assume this isn't a massively delayed vblank
+         * interrupt, but a vblank interrupt that fired a few
+         * microseconds before true start of vblank. Compensate
+         * by adding a full frame duration to the final timestamp.
+         * Happens, e.g., on ATI R500, R600.
+         *
+         * We only do this if DRM_CALLED_FROM_VBLIRQ.
+         */
+        if ((flags & DRM_CALLED_FROM_VBLIRQ) && !invbl &&
+            ((vdisplay - vpos) < vtotal / 100)) {
+                delta_ns = delta_ns - framedur_ns;
+
+                /* Signal this correction as "applied". */
+                vbl_status |= 0x8;
+        }
+
+        /* Subtract time delta from raw timestamp to get final
+         * vblank_time timestamp for end of vblank.
+         */
+        *vblank_time = ns_to_timeval(timeval_to_ns(&raw_time) - delta_ns);
+
+        DRM_DEBUG("crtc %d : v %d p(%d,%d)@ %d.%d -> %d.%d [e %d us, %d rep]\n",
+                  crtc, (int) vbl_status, hpos, vpos, raw_time.tv_sec,
+                  raw_time.tv_usec, vblank_time->tv_sec, vblank_time->tv_usec,
+                  (int) duration_ns/1000, i);
+
+        vbl_status = DRM_VBLANKTIME_SCANOUTPOS_METHOD;
+        if (invbl)
+                vbl_status |= DRM_VBLANKTIME_INVBL;
+
+        return vbl_status;
+}
+EXPORT_SYMBOL(drm_calc_vbltimestamp_from_scanoutpos);
+
+/**
+ * drm_get_last_vbltimestamp - retrieve raw timestamp for the most recent
+ * vblank interval.
+ *
+ * @dev: DRM device
+ * @crtc: which crtc's vblank timestamp to retrieve
+ * @tvblank: Pointer to target struct timeval which should receive the timestamp
+ * @flags: Flags to pass to driver:
+ *         0 = Default.
+ *         DRM_CALLED_FROM_VBLIRQ = If function is called from vbl irq handler.
+ *
+ * Fetches the system timestamp corresponding to the time of the most recent
+ * vblank interval on specified crtc. May call into kms-driver to
+ * compute the timestamp with a high-precision GPU specific method.
+ *
+ * Returns zero if timestamp originates from uncorrected do_gettimeofday()
+ * call, i.e., it isn't very precisely locked to the true vblank.
+ *
+ * Returns non-zero if timestamp is considered to be very precise.
+ */
+u32 drm_get_last_vbltimestamp(struct drm_device *dev, int crtc,
+                              struct timeval *tvblank, unsigned flags)
+{
+        int ret = 0;
+
+        /* Define requested maximum error on timestamps (nanoseconds). */
+        int max_error = (int) drm_timestamp_precision * 1000;
+
+        /* Query driver if possible and precision timestamping enabled. */
+        if (dev->driver->get_vblank_timestamp && (max_error > 0)) {
+                ret = dev->driver->get_vblank_timestamp(dev, crtc, &max_error,
+                                                        tvblank, flags);
+                if (ret > 0)
+                        return (u32) ret;
+        }
+
+        /* GPU high precision timestamp query unsupported or failed.
+         * Return gettimeofday timestamp as best estimate.
+         */
+        do_gettimeofday(tvblank);
+
+        return 0;
+}
+EXPORT_SYMBOL(drm_get_last_vbltimestamp);
+
+/**
  * drm_vblank_count - retrieve "cooked" vblank counter value
  * @dev: DRM device
  * @crtc: which counter to retrieve
@@ -375,6 +768,40 @@ u32 drm_vblank_count(struct drm_device *dev, int crtc)
 EXPORT_SYMBOL(drm_vblank_count);
 
 /**
+ * drm_vblank_count_and_time - retrieve "cooked" vblank counter value
+ * and the system timestamp corresponding to that vblank counter value.
+ *
+ * @dev: DRM device
+ * @crtc: which counter to retrieve
+ * @vblanktime: Pointer to struct timeval to receive the vblank timestamp.
+ *
+ * Fetches the "cooked" vblank count value that represents the number of
+ * vblank events since the system was booted, including lost events due to
+ * modesetting activity. Returns corresponding system timestamp of the time
+ * of the vblank interval that corresponds to the current value vblank counter
+ * value.
+ */
+u32 drm_vblank_count_and_time(struct drm_device *dev, int crtc,
+                              struct timeval *vblanktime)
+{
+        u32 cur_vblank;
+
+        /* Read timestamp from slot of _vblank_time ringbuffer
+         * that corresponds to current vblank count. Retry if
+         * count has incremented during readout. This works like
+         * a seqlock.
+         */
+        do {
+                cur_vblank = atomic_read(&dev->_vblank_count[crtc]);
+                *vblanktime = vblanktimestamp(dev, crtc, cur_vblank);
+                smp_rmb();
+        } while (cur_vblank != atomic_read(&dev->_vblank_count[crtc]));
+
+        return cur_vblank;
+}
+EXPORT_SYMBOL(drm_vblank_count_and_time);
+
+/**
  * drm_update_vblank_count - update the master vblank counter
  * @dev: DRM device
  * @crtc: counter to update
@@ -392,7 +819,8 @@ EXPORT_SYMBOL(drm_vblank_count);
  */
 static void drm_update_vblank_count(struct drm_device *dev, int crtc)
 {
-        u32 cur_vblank, diff;
+        u32 cur_vblank, diff, tslot, rc;
+        struct timeval t_vblank;
 
         /*
          * Interrupts were disabled prior to this call, so deal with counter
@@ -400,8 +828,18 @@ static void drm_update_vblank_count(struct drm_device *dev, int crtc)
          * NOTE!  It's possible we lost a full dev->max_vblank_count events
          * here if the register is small or we had vblank interrupts off for
          * a long time.
+         *
+         * We repeat the hardware vblank counter & timestamp query until
+         * we get consistent results. This to prevent races between gpu
+         * updating its hardware counter while we are retrieving the
+         * corresponding vblank timestamp.
          */
-        cur_vblank = dev->driver->get_vblank_counter(dev, crtc);
+        do {
+                cur_vblank = dev->driver->get_vblank_counter(dev, crtc);
+                rc = drm_get_last_vbltimestamp(dev, crtc, &t_vblank, 0);
+        } while (cur_vblank != dev->driver->get_vblank_counter(dev, crtc));
+
+        /* Deal with counter wrap */
         diff = cur_vblank - dev->last_vblank[crtc];
         if (cur_vblank < dev->last_vblank[crtc]) {
                 diff += dev->max_vblank_count;
@@ -413,6 +851,16 @@ static void drm_update_vblank_count(struct drm_device *dev, int crtc)
         DRM_DEBUG("enabling vblank interrupts on crtc %d, missed %d\n",
                   crtc, diff);
 
+        /* Reinitialize corresponding vblank timestamp if high-precision query
+         * available. Skip this step if query unsupported or failed. Will
+         * reinitialize delayed at next vblank interrupt in that case.
+         */
+        if (rc) {
+                tslot = atomic_read(&dev->_vblank_count[crtc]) + diff;
+                vblanktimestamp(dev, crtc, tslot) = t_vblank;
+                smp_wmb();
+        }
+
         atomic_add(diff, &dev->_vblank_count[crtc]);
 }
 
@@ -429,15 +877,27 @@ static void drm_update_vblank_count(struct drm_device *dev, int crtc)
  */
 int drm_vblank_get(struct drm_device *dev, int crtc)
 {
-        unsigned long irqflags;
+        unsigned long irqflags, irqflags2;
         int ret = 0;
 
         spin_lock_irqsave(&dev->vbl_lock, irqflags);
         /* Going from 0->1 means we have to enable interrupts again */
         if (atomic_add_return(1, &dev->vblank_refcount[crtc]) == 1) {
+                /* Disable preemption while holding vblank_time_lock. Do
+                 * it explicitely to guard against PREEMPT_RT kernel.
+                 */
+                preempt_disable();
+                spin_lock_irqsave(&dev->vblank_time_lock, irqflags2);
                 if (!dev->vblank_enabled[crtc]) {
+                        /* Enable vblank irqs under vblank_time_lock protection.
+                         * All vblank count & timestamp updates are held off
+                         * until we are done reinitializing master counter and
+                         * timestamps. Filtercode in drm_handle_vblank() will
+                         * prevent double-accounting of same vblank interval.
+                         */
                         ret = dev->driver->enable_vblank(dev, crtc);
-                        DRM_DEBUG("enabling vblank on crtc %d, ret: %d\n", crtc, ret);
+                        DRM_DEBUG("enabling vblank on crtc %d, ret: %d\n",
+                                  crtc, ret);
                         if (ret)
                                 atomic_dec(&dev->vblank_refcount[crtc]);
                         else {
@@ -445,6 +905,8 @@ int drm_vblank_get(struct drm_device *dev, int crtc)
                                 drm_update_vblank_count(dev, crtc);
                         }
                 }
+                spin_unlock_irqrestore(&dev->vblank_time_lock, irqflags2);
+                preempt_enable();
         } else {
                 if (!dev->vblank_enabled[crtc]) {
                         atomic_dec(&dev->vblank_refcount[crtc]);
@@ -463,15 +925,17 @@ EXPORT_SYMBOL(drm_vblank_get);
  * @crtc: which counter to give up
  *
  * Release ownership of a given vblank counter, turning off interrupts
- * if possible.
+ * if possible. Disable interrupts after drm_vblank_offdelay milliseconds.
  */
 void drm_vblank_put(struct drm_device *dev, int crtc)
 {
-        BUG_ON (atomic_read (&dev->vblank_refcount[crtc]) == 0);
+        BUG_ON(atomic_read(&dev->vblank_refcount[crtc]) == 0);
 
         /* Last user schedules interrupt disable */
-        if (atomic_dec_and_test(&dev->vblank_refcount[crtc]))
-                mod_timer(&dev->vblank_disable_timer, jiffies + 5*DRM_HZ);
+        if (atomic_dec_and_test(&dev->vblank_refcount[crtc]) &&
+            (drm_vblank_offdelay > 0))
+                mod_timer(&dev->vblank_disable_timer,
+                          jiffies + ((drm_vblank_offdelay * DRM_HZ)/1000));
 }
 EXPORT_SYMBOL(drm_vblank_put);
 
@@ -480,10 +944,8 @@ void drm_vblank_off(struct drm_device *dev, int crtc)
         unsigned long irqflags;
 
         spin_lock_irqsave(&dev->vbl_lock, irqflags);
-        dev->driver->disable_vblank(dev, crtc);
+        vblank_disable_and_save(dev, crtc);
         DRM_WAKEUP(&dev->vbl_queue[crtc]);
-        dev->vblank_enabled[crtc] = 0;
-        dev->last_vblank[crtc] = dev->driver->get_vblank_counter(dev, crtc);
         spin_unlock_irqrestore(&dev->vbl_lock, irqflags);
 }
 EXPORT_SYMBOL(drm_vblank_off);
@@ -599,7 +1061,6 @@ static int drm_queue_vblank_event(struct drm_device *dev, int pipe,
         e->base.file_priv = file_priv;
         e->base.destroy = (void (*) (struct drm_pending_event *)) kfree;
 
-        do_gettimeofday(&now);
         spin_lock_irqsave(&dev->event_lock, flags);
 
         if (file_priv->event_space < sizeof e->event) {
@@ -609,7 +1070,8 @@ static int drm_queue_vblank_event(struct drm_device *dev, int pipe,
         }
 
         file_priv->event_space -= sizeof e->event;
-        seq = drm_vblank_count(dev, pipe);
+        seq = drm_vblank_count_and_time(dev, pipe, &now);
+
         if ((vblwait->request.type & _DRM_VBLANK_NEXTONMISS) &&
             (seq - vblwait->request.sequence) <= (1 << 23)) {
                 vblwait->request.sequence = seq + 1;
@@ -718,11 +1180,10 @@ int drm_wait_vblank(struct drm_device *dev, void *data,
         if (ret != -EINTR) {
                 struct timeval now;
 
-                do_gettimeofday(&now);
-
+                vblwait->reply.sequence = drm_vblank_count_and_time(dev, crtc, &now);
                 vblwait->reply.tval_sec = now.tv_sec;
                 vblwait->reply.tval_usec = now.tv_usec;
-                vblwait->reply.sequence = drm_vblank_count(dev, crtc);
+
                 DRM_DEBUG("returning %d to client\n",
                           vblwait->reply.sequence);
         } else {
@@ -741,8 +1202,7 @@ void drm_handle_vblank_events(struct drm_device *dev, int crtc)
         unsigned long flags;
         unsigned int seq;
 
-        do_gettimeofday(&now);
-        seq = drm_vblank_count(dev, crtc);
+        seq = drm_vblank_count_and_time(dev, crtc, &now);
 
         spin_lock_irqsave(&dev->event_lock, flags);
 
@@ -780,11 +1240,64 @@ void drm_handle_vblank_events(struct drm_device *dev, int crtc)
  */
 void drm_handle_vblank(struct drm_device *dev, int crtc)
 {
+        u32 vblcount;
+        s64 diff_ns;
+        struct timeval tvblank;
+        unsigned long irqflags;
+
         if (!dev->num_crtcs)
                 return;
 
-        atomic_inc(&dev->_vblank_count[crtc]);
+        /* Need timestamp lock to prevent concurrent execution with
+         * vblank enable/disable, as this would cause inconsistent
+         * or corrupted timestamps and vblank counts.
+         */
+        spin_lock_irqsave(&dev->vblank_time_lock, irqflags);
+
+        /* Vblank irq handling disabled. Nothing to do. */
+        if (!dev->vblank_enabled[crtc]) {
+                spin_unlock_irqrestore(&dev->vblank_time_lock, irqflags);
+                return;
+        }
+
+        /* Fetch corresponding timestamp for this vblank interval from
+         * driver and store it in proper slot of timestamp ringbuffer.
+         */
+
+        /* Get current timestamp and count. */
+        vblcount = atomic_read(&dev->_vblank_count[crtc]);
+        drm_get_last_vbltimestamp(dev, crtc, &tvblank, DRM_CALLED_FROM_VBLIRQ);
+
+        /* Compute time difference to timestamp of last vblank */
+        diff_ns = timeval_to_ns(&tvblank) -
+                  timeval_to_ns(&vblanktimestamp(dev, crtc, 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 (abs(diff_ns) > DRM_REDUNDANT_VBLIRQ_THRESH_NS) {
+                /* Store new timestamp in ringbuffer. */
+                vblanktimestamp(dev, crtc, vblcount + 1) = tvblank;
+                smp_wmb();
+
+                /* Increment cooked vblank count. This also atomically commits
+                 * the timestamp computed above.
+                 */
+                atomic_inc(&dev->_vblank_count[crtc]);
+        } else {
+                DRM_DEBUG("crtc %d: Redundant vblirq ignored. diff_ns = %d\n",
+                          crtc, (int) diff_ns);
+        }
+
         DRM_WAKEUP(&dev->vbl_queue[crtc]);
         drm_handle_vblank_events(dev, crtc);
+
+        spin_unlock_irqrestore(&dev->vblank_time_lock, irqflags);
 }
 EXPORT_SYMBOL(drm_handle_vblank);
diff --git a/drivers/gpu/drm/drm_stub.c b/drivers/gpu/drm/drm_stub.c
index cdc89ee042cc..d59edc18301f 100644
--- a/drivers/gpu/drm/drm_stub.c
+++ b/drivers/gpu/drm/drm_stub.c
@@ -40,12 +40,22 @@
 unsigned int drm_debug = 0;     /* 1 to enable debug output */
 EXPORT_SYMBOL(drm_debug);
 
+unsigned int drm_vblank_offdelay = 5000;    /* Default to 5000 msecs. */
+EXPORT_SYMBOL(drm_vblank_offdelay);
+
+unsigned int drm_timestamp_precision = 20;  /* Default to 20 usecs. */
+EXPORT_SYMBOL(drm_timestamp_precision);
+
 MODULE_AUTHOR(CORE_AUTHOR);
 MODULE_DESCRIPTION(CORE_DESC);
 MODULE_LICENSE("GPL and additional rights");
 MODULE_PARM_DESC(debug, "Enable debug output");
+MODULE_PARM_DESC(vblankoffdelay, "Delay until vblank irq auto-disable [msecs]");
+MODULE_PARM_DESC(timestamp_precision_usec, "Max. error on timestamps [usecs]");
 
 module_param_named(debug, drm_debug, int, 0600);
+module_param_named(vblankoffdelay, drm_vblank_offdelay, int, 0600);
+module_param_named(timestamp_precision_usec, drm_timestamp_precision, int, 0600);
 
 struct idr drm_minors_idr;
 
diff --git a/include/drm/drmP.h b/include/drm/drmP.h
index 274eaaa15c36..2b3398004aa5 100644
--- a/include/drm/drmP.h
+++ b/include/drm/drmP.h
@@ -683,6 +683,21 @@ struct drm_master {
         void *driver_priv; /**< Private structure for driver to use */
 };
 
+/* Size of ringbuffer for vblank timestamps. Just double-buffer
+ * in initial implementation.
+ */
+#define DRM_VBLANKTIME_RBSIZE 2
+
+/* Flags and return codes for get_vblank_timestamp() driver function. */
+#define DRM_CALLED_FROM_VBLIRQ 1
+#define DRM_VBLANKTIME_SCANOUTPOS_METHOD (1 << 0)
+#define DRM_VBLANKTIME_INVBL             (1 << 1)
+
+/* get_scanout_position() return flags */
+#define DRM_SCANOUTPOS_VALID        (1 << 0)
+#define DRM_SCANOUTPOS_INVBL        (1 << 1)
+#define DRM_SCANOUTPOS_ACCURATE     (1 << 2)
+
 /**
  * DRM driver structure. This structure represent the common code for
  * a family of cards. There will one drm_device for each card present
@@ -760,6 +775,68 @@ struct drm_driver {
          */
         int (*device_is_agp) (struct drm_device *dev);
 
+        /**
+         * Called by vblank timestamping code.
+         *
+         * Return the current display scanout position from a crtc.
+         *
+         * \param dev  DRM device.
+         * \param crtc Id of the crtc to query.
+         * \param *vpos Target location for current vertical scanout position.
+         * \param *hpos Target location for current horizontal scanout position.
+         *
+         * Returns vpos as a positive number while in active scanout area.
+         * Returns vpos as a negative number inside vblank, counting the number
+         * of scanlines to go until end of vblank, e.g., -1 means "one scanline
+         * until start of active scanout / end of vblank."
+         *
+         * \return Flags, or'ed together as follows:
+         *
+         * DRM_SCANOUTPOS_VALID = Query successfull.
+         * DRM_SCANOUTPOS_INVBL = Inside vblank.
+         * DRM_SCANOUTPOS_ACCURATE = Returned position is accurate. A lack of
+         * this flag means that returned position may be offset by a constant
+         * but unknown small number of scanlines wrt. real scanout position.
+         *
+         */
+        int (*get_scanout_position) (struct drm_device *dev, int crtc,
+                                     int *vpos, int *hpos);
+
+        /**
+         * Called by \c drm_get_last_vbltimestamp. Should return a precise
+         * timestamp when the most recent VBLANK interval ended or will end.
+         *
+         * Specifically, the timestamp in @vblank_time should correspond as
+         * closely as possible to the time when the first video scanline of
+         * the video frame after the end of VBLANK will start scanning out,
+         * the time immmediately after end of the VBLANK interval. If the
+         * @crtc is currently inside VBLANK, this will be a time in the future.
+         * If the @crtc is currently scanning out a frame, this will be the
+         * past start time of the current scanout. This is meant to adhere
+         * to the OpenML OML_sync_control extension specification.
+         *
+         * \param dev dev DRM device handle.
+         * \param crtc crtc for which timestamp should be returned.
+         * \param *max_error Maximum allowable timestamp error in nanoseconds.
+         *                   Implementation should strive to provide timestamp
+         *                   with an error of at most *max_error nanoseconds.
+         *                   Returns true upper bound on error for timestamp.
+         * \param *vblank_time Target location for returned vblank timestamp.
+         * \param flags 0 = Defaults, no special treatment needed.
+         * \param       DRM_CALLED_FROM_VBLIRQ = Function is called from vblank
+         *              irq handler. Some drivers need to apply some workarounds
+         *              for gpu-specific vblank irq quirks if flag is set.
+         *
+         * \returns
+         * Zero if timestamping isn't supported in current display mode or a
+         * negative number on failure. A positive status code on success,
+         * which describes how the vblank_time timestamp was computed.
+         */
+        int (*get_vblank_timestamp) (struct drm_device *dev, int crtc,
+                                     int *max_error,
+                                     struct timeval *vblank_time,
+                                     unsigned flags);
+
         /* these have to be filled in */
 
         irqreturn_t(*irq_handler) (DRM_IRQ_ARGS);
@@ -983,6 +1060,8 @@ struct drm_device {
 
         wait_queue_head_t *vbl_queue;   /**< VBLANK wait queue */
         atomic_t *_vblank_count;        /**< number of VBLANK interrupts (driver must alloc the right number of counters) */
+        struct timeval *_vblank_time;   /**< timestamp of current vblank_count (drivers must alloc right number of fields) */
+        spinlock_t vblank_time_lock;    /**< Protects vblank count and time updates during vblank enable/disable */
         spinlock_t vbl_lock;
         atomic_t *vblank_refcount;      /* number of users of vblank interruptsper crtc */
         u32 *last_vblank;               /* protected by dev->vbl_lock, used */
@@ -1284,11 +1363,22 @@ extern int drm_wait_vblank(struct drm_device *dev, void *data,
                            struct drm_file *filp);
 extern int drm_vblank_wait(struct drm_device *dev, unsigned int *vbl_seq);
 extern u32 drm_vblank_count(struct drm_device *dev, int crtc);
+extern u32 drm_vblank_count_and_time(struct drm_device *dev, int crtc,
+                                     struct timeval *vblanktime);
 extern void drm_handle_vblank(struct drm_device *dev, int crtc);
 extern int drm_vblank_get(struct drm_device *dev, int crtc);
 extern void drm_vblank_put(struct drm_device *dev, int crtc);
 extern void drm_vblank_off(struct drm_device *dev, int crtc);
 extern void drm_vblank_cleanup(struct drm_device *dev);
+extern u32 drm_get_last_vbltimestamp(struct drm_device *dev, int crtc,
+                                     struct timeval *tvblank, unsigned flags);
+extern int drm_calc_vbltimestamp_from_scanoutpos(struct drm_device *dev,
+                                                 int crtc, int *max_error,
+                                                 struct timeval *vblank_time,
+                                                 unsigned flags,
+                                                 struct drm_crtc *refcrtc);
+extern void drm_calc_timestamping_constants(struct drm_crtc *crtc);
+
 /* Modesetting support */
 extern void drm_vblank_pre_modeset(struct drm_device *dev, int crtc);
 extern void drm_vblank_post_modeset(struct drm_device *dev, int crtc);
@@ -1340,6 +1430,9 @@ extern void drm_put_dev(struct drm_device *dev);
 extern int drm_put_minor(struct drm_minor **minor);
 extern unsigned int drm_debug;
 
+extern unsigned int drm_vblank_offdelay;
+extern unsigned int drm_timestamp_precision;
+
 extern struct class *drm_class;
 extern struct proc_dir_entry *drm_proc_root;
 extern struct dentry *drm_debugfs_root;
diff --git a/include/drm/drm_crtc.h b/include/drm/drm_crtc.h
index 029aa688e787..acd7fade160d 100644
--- a/include/drm/drm_crtc.h
+++ b/include/drm/drm_crtc.h
@@ -351,8 +351,14 @@ struct drm_crtc {
 
         bool enabled;
 
+        /* Requested mode from modesetting. */
         struct drm_display_mode mode;
 
+        /* Programmed mode in hw, after adjustments for encoders,
+         * crtc, panel scaling etc. Needed for timestamping etc.
+         */
+        struct drm_display_mode hwmode;
+
         int x, y;
         const struct drm_crtc_funcs *funcs;
 
@@ -360,6 +366,9 @@ struct drm_crtc {
         uint32_t gamma_size;
         uint16_t *gamma_store;
 
+        /* Constants needed for precise vblank and swap timestamping. */
+        s64 framedur_ns, linedur_ns, pixeldur_ns;
+
         /* if you are using the helper */
         void *helper_private;
 };