drm/i915: use a single intel_fbc_work struct
authorPaulo Zanoni <paulo.r.zanoni@intel.com>
Mon, 26 Oct 2015 18:27:49 +0000 (16:27 -0200)
committerPaulo Zanoni <paulo.r.zanoni@intel.com>
Thu, 3 Dec 2015 13:35:58 +0000 (11:35 -0200)
This was already on my TODO list, and was requested both by Chris and
Ville, for different reasons. The advantages are avoiding a frequent
malloc/free pair, and the locality of having the work structure
embedded in dev_priv. The maximum used memory is also smaller since
previously we could have multiple allocated intel_fbc_work structs at
the same time, and now we'll always have a single one - the one
embedded on dev_priv. Of course, we're now using a little more memory
on the cases where there's nothing scheduled.

The biggest challenge here is to keep everything synchronized the way
it was before.

Currently, when we try to activate FBC, we allocate a new
intel_fbc_work structure. Then later when we conclude we must delay
the FBC activation a little more, we allocate a new intel_fbc_work
struct, and then adjust dev_priv->fbc.fbc_work to point to the new
struct. So when the old work runs - at intel_fbc_work_fn() - it will
check that dev_priv->fbc.fbc_work points to something else, so it does
nothing. Everything is also protected by fbc.lock.

Just cancelling the old delayed work doesn't work because we might
just cancel it after the work function already started to run, but
while it is still waiting to grab fbc.lock. That's why we use the
"dev_priv->fbc.fbc_work == work" check described in the paragraph
above.

So now that we have a single work struct we have to introduce a new
way to synchronize everything. So we're making the work function a
normal work instead of a delayed work, and it will be responsible for
sleeping the appropriate amount of time itself. This way, after it
wakes up it can grab the lock, ask "were we delayed or cancelled?" and
then go back to sleep, enable FBC or give up.

v2:
  - Spelling fixes.
  - Rebase after changing the patch order.
  - Fix ms/jiffies confusion.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1)
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/intel_fbc.c

index 2493ad2f5c12d198159d304719f23c1669713538..fff165a35ce718ceee68e681572f19ed92d4162c 100644 (file)
@@ -919,9 +919,11 @@ struct i915_fbc {
        bool active;
 
        struct intel_fbc_work {
-               struct delayed_work work;
+               bool scheduled;
+               struct work_struct work;
                struct drm_framebuffer *fb;
-       } *fbc_work;
+               unsigned long enable_jiffies;
+       } work;
 
        const char *no_fbc_reason;
 
index 8460e3d72b98c4482567897886ad04aa06a8d198..07d3342b59de9d44124831d118395927955a2b19 100644 (file)
@@ -392,85 +392,72 @@ static void intel_fbc_activate(const struct drm_framebuffer *fb)
 
 static void intel_fbc_work_fn(struct work_struct *__work)
 {
-       struct intel_fbc_work *work =
-               container_of(to_delayed_work(__work),
-                            struct intel_fbc_work, work);
-       struct drm_i915_private *dev_priv = work->fb->dev->dev_private;
-       struct drm_framebuffer *crtc_fb = dev_priv->fbc.crtc->base.primary->fb;
+       struct drm_i915_private *dev_priv =
+               container_of(__work, struct drm_i915_private, fbc.work.work);
+       struct intel_fbc_work *work = &dev_priv->fbc.work;
+       struct intel_crtc *crtc = dev_priv->fbc.crtc;
+       int delay_ms = 50;
+
+retry:
+       /* Delay the actual enabling to let pageflipping cease and the
+        * display to settle before starting the compression. Note that
+        * this delay also serves a second purpose: it allows for a
+        * vblank to pass after disabling the FBC before we attempt
+        * to modify the control registers.
+        *
+        * A more complicated solution would involve tracking vblanks
+        * following the termination of the page-flipping sequence
+        * and indeed performing the enable as a co-routine and not
+        * waiting synchronously upon the vblank.
+        *
+        * WaFbcWaitForVBlankBeforeEnable:ilk,snb
+        */
+       wait_remaining_ms_from_jiffies(work->enable_jiffies, delay_ms);
 
        mutex_lock(&dev_priv->fbc.lock);
-       if (work == dev_priv->fbc.fbc_work) {
-               /* Double check that we haven't switched fb without cancelling
-                * the prior work.
-                */
-               if (crtc_fb == work->fb)
-                       intel_fbc_activate(work->fb);
 
-               dev_priv->fbc.fbc_work = NULL;
+       /* Were we cancelled? */
+       if (!work->scheduled)
+               goto out;
+
+       /* Were we delayed again while this function was sleeping? */
+       if (time_after(work->enable_jiffies + msecs_to_jiffies(delay_ms),
+                      jiffies)) {
+               mutex_unlock(&dev_priv->fbc.lock);
+               goto retry;
        }
-       mutex_unlock(&dev_priv->fbc.lock);
 
-       kfree(work);
+       if (crtc->base.primary->fb == work->fb)
+               intel_fbc_activate(work->fb);
+
+       work->scheduled = false;
+
+out:
+       mutex_unlock(&dev_priv->fbc.lock);
 }
 
 static void intel_fbc_cancel_work(struct drm_i915_private *dev_priv)
 {
        WARN_ON(!mutex_is_locked(&dev_priv->fbc.lock));
-
-       if (dev_priv->fbc.fbc_work == NULL)
-               return;
-
-       /* Synchronisation is provided by struct_mutex and checking of
-        * dev_priv->fbc.fbc_work, so we can perform the cancellation
-        * entirely asynchronously.
-        */
-       if (cancel_delayed_work(&dev_priv->fbc.fbc_work->work))
-               /* tasklet was killed before being run, clean up */
-               kfree(dev_priv->fbc.fbc_work);
-
-       /* Mark the work as no longer wanted so that if it does
-        * wake-up (because the work was already running and waiting
-        * for our mutex), it will discover that is no longer
-        * necessary to run.
-        */
-       dev_priv->fbc.fbc_work = NULL;
+       dev_priv->fbc.work.scheduled = false;
 }
 
 static void intel_fbc_schedule_activation(struct intel_crtc *crtc)
 {
-       struct intel_fbc_work *work;
        struct drm_i915_private *dev_priv = crtc->base.dev->dev_private;
+       struct intel_fbc_work *work = &dev_priv->fbc.work;
 
        WARN_ON(!mutex_is_locked(&dev_priv->fbc.lock));
 
-       intel_fbc_cancel_work(dev_priv);
-
-       work = kzalloc(sizeof(*work), GFP_KERNEL);
-       if (work == NULL) {
-               DRM_ERROR("Failed to allocate FBC work structure\n");
-               intel_fbc_activate(crtc->base.primary->fb);
-               return;
-       }
-
+       /* It is useless to call intel_fbc_cancel_work() in this function since
+        * we're not releasing fbc.lock, so it won't have an opportunity to grab
+        * it to discover that it was cancelled. So we just update the expected
+        * jiffy count. */
        work->fb = crtc->base.primary->fb;
-       INIT_DELAYED_WORK(&work->work, intel_fbc_work_fn);
+       work->scheduled = true;
+       work->enable_jiffies = jiffies;
 
-       dev_priv->fbc.fbc_work = work;
-
-       /* Delay the actual enabling to let pageflipping cease and the
-        * display to settle before starting the compression. Note that
-        * this delay also serves a second purpose: it allows for a
-        * vblank to pass after disabling the FBC before we attempt
-        * to modify the control registers.
-        *
-        * A more complicated solution would involve tracking vblanks
-        * following the termination of the page-flipping sequence
-        * and indeed performing the enable as a co-routine and not
-        * waiting synchronously upon the vblank.
-        *
-        * WaFbcWaitForVBlankBeforeEnable:ilk,snb
-        */
-       schedule_delayed_work(&work->work, msecs_to_jiffies(50));
+       schedule_work(&work->work);
 }
 
 static void __intel_fbc_deactivate(struct drm_i915_private *dev_priv)
@@ -1106,9 +1093,11 @@ void intel_fbc_init(struct drm_i915_private *dev_priv)
 {
        enum pipe pipe;
 
+       INIT_WORK(&dev_priv->fbc.work.work, intel_fbc_work_fn);
        mutex_init(&dev_priv->fbc.lock);
        dev_priv->fbc.enabled = false;
        dev_priv->fbc.active = false;
+       dev_priv->fbc.work.scheduled = false;
 
        if (!HAS_FBC(dev_priv)) {
                dev_priv->fbc.no_fbc_reason = "unsupported by this chipset";