drm/i915/ringbuffer: Make IRQ refcnting atomic

In order to enforce the correct memory barriers for irq get/put, we need
to perform the actual counting using atomic operations.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

4 files changed, 56 insertions, 46 deletions

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 27fa2a1b26a5..726c2ccd674c 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -2000,17 +2000,19 @@ i915_do_wait_request(struct drm_device *dev, uint32_t seqno,
                 trace_i915_gem_request_wait_begin(dev, seqno);
 
                 ring->waiting_seqno = seqno;
-                ring->irq_get(ring);
-                if (interruptible)
-                        ret = wait_event_interruptible(ring->irq_queue,
-                                i915_seqno_passed(ring->get_seqno(ring), seqno)
-                                || atomic_read(&dev_priv->mm.wedged));
-                else
-                        wait_event(ring->irq_queue,
-                                i915_seqno_passed(ring->get_seqno(ring), seqno)
-                                || atomic_read(&dev_priv->mm.wedged));
+                ret = -ENODEV;
+                if (ring->irq_get(ring)) {
+                        if (interruptible)
+                                ret = wait_event_interruptible(ring->irq_queue,
+                                                               i915_seqno_passed(ring->get_seqno(ring), seqno)
+                                                               || atomic_read(&dev_priv->mm.wedged));
+                        else
+                                wait_event(ring->irq_queue,
+                                           i915_seqno_passed(ring->get_seqno(ring), seqno)
+                                           || atomic_read(&dev_priv->mm.wedged));
 
-                ring->irq_put(ring);
+                        ring->irq_put(ring);
+                }
                 ring->waiting_seqno = 0;
 
                 trace_i915_gem_request_wait_end(dev, seqno);
@@ -3157,14 +3159,15 @@ i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
                  * generation is designed to be run atomically and so is
                  * lockless.
                  */
-                ring->irq_get(ring);
-                ret = wait_event_interruptible(ring->irq_queue,
-                                               i915_seqno_passed(ring->get_seqno(ring), seqno)
-                                               || atomic_read(&dev_priv->mm.wedged));
-                ring->irq_put(ring);
+                if (ring->irq_get(ring)) {
+                        ret = wait_event_interruptible(ring->irq_queue,
+                                                       i915_seqno_passed(ring->get_seqno(ring), seqno)
+                                                       || atomic_read(&dev_priv->mm.wedged));
+                        ring->irq_put(ring);
 
-                if (ret == 0 && atomic_read(&dev_priv->mm.wedged))
-                        ret = -EIO;
+                        if (ret == 0 && atomic_read(&dev_priv->mm.wedged))
+                                ret = -EIO;
+                }
         }
 
         if (ret == 0)
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
index 02e4dd82f754..2ddb98b5c90f 100644
--- a/drivers/gpu/drm/i915/i915_irq.c
+++ b/drivers/gpu/drm/i915/i915_irq.c
@@ -1186,10 +1186,9 @@ void i915_trace_irq_get(struct drm_device *dev, u32 seqno)
         drm_i915_private_t *dev_priv = (drm_i915_private_t *) dev->dev_private;
         struct intel_ring_buffer *ring = LP_RING(dev_priv);
 
-        if (dev_priv->trace_irq_seqno == 0)
-                ring->irq_get(ring);
-
-        dev_priv->trace_irq_seqno = seqno;
+        if (dev_priv->trace_irq_seqno == 0 &&
+            ring->irq_get(ring))
+                dev_priv->trace_irq_seqno = seqno;
 }
 
 static int i915_wait_irq(struct drm_device * dev, int irq_nr)
@@ -1211,10 +1210,12 @@ static int i915_wait_irq(struct drm_device * dev, int irq_nr)
         if (master_priv->sarea_priv)
                 master_priv->sarea_priv->perf_boxes |= I915_BOX_WAIT;
 
-        ring->irq_get(ring);
-        DRM_WAIT_ON(ret, ring->irq_queue, 3 * DRM_HZ,
-                    READ_BREADCRUMB(dev_priv) >= irq_nr);
-        ring->irq_put(ring);
+        ret = -ENODEV;
+        if (ring->irq_get(ring)) {
+                DRM_WAIT_ON(ret, ring->irq_queue, 3 * DRM_HZ,
+                            READ_BREADCRUMB(dev_priv) >= irq_nr);
+                ring->irq_put(ring);
+        }
 
         if (ret == -EBUSY) {
                 DRM_ERROR("EBUSY -- rec: %d emitted: %d\n",
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
index 74b99718a1fb..a3fd993e0de0 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.c
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
@@ -327,25 +327,28 @@ ring_get_seqno(struct intel_ring_buffer *ring)
         return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
 }
 
-static void
+static bool
 render_ring_get_irq(struct intel_ring_buffer *ring)
 {
         struct drm_device *dev = ring->dev;
 
-        if (dev->irq_enabled && ++ring->irq_refcount == 1) {
+        if (!dev->irq_enabled)
+                return false;
+
+        if (atomic_inc_return(&ring->irq_refcount) == 1) {
                 drm_i915_private_t *dev_priv = dev->dev_private;
                 unsigned long irqflags;
 
                 spin_lock_irqsave(&dev_priv->irq_lock, irqflags);
-
                 if (HAS_PCH_SPLIT(dev))
                         ironlake_enable_graphics_irq(dev_priv,
                                                      GT_USER_INTERRUPT);
                 else
                         i915_enable_irq(dev_priv, I915_USER_INTERRUPT);
-
                 spin_unlock_irqrestore(&dev_priv->irq_lock, irqflags);
         }
+
+        return true;
 }
 
 static void
@@ -353,8 +356,7 @@ render_ring_put_irq(struct intel_ring_buffer *ring)
 {
         struct drm_device *dev = ring->dev;
 
-        BUG_ON(dev->irq_enabled && ring->irq_refcount == 0);
-        if (dev->irq_enabled && --ring->irq_refcount == 0) {
+        if (atomic_dec_and_test(&ring->irq_refcount)) {
                 drm_i915_private_t *dev_priv = dev->dev_private;
                 unsigned long irqflags;
 
@@ -417,12 +419,15 @@ ring_add_request(struct intel_ring_buffer *ring,
         return 0;
 }
 
-static void
+static bool
 ring_get_irq(struct intel_ring_buffer *ring, u32 flag)
 {
         struct drm_device *dev = ring->dev;
 
-        if (dev->irq_enabled && ++ring->irq_refcount == 1) {
+        if (!dev->irq_enabled)
+               return false;
+
+        if (atomic_inc_return(&ring->irq_refcount) == 1) {
                 drm_i915_private_t *dev_priv = dev->dev_private;
                 unsigned long irqflags;
 
@@ -430,6 +435,8 @@ ring_get_irq(struct intel_ring_buffer *ring, u32 flag)
                 ironlake_enable_graphics_irq(dev_priv, flag);
                 spin_unlock_irqrestore(&dev_priv->irq_lock, irqflags);
         }
+
+        return true;
 }
 
 static void
@@ -437,7 +444,7 @@ ring_put_irq(struct intel_ring_buffer *ring, u32 flag)
 {
         struct drm_device *dev = ring->dev;
 
-        if (dev->irq_enabled && --ring->irq_refcount == 0) {
+        if (atomic_dec_and_test(&ring->irq_refcount)) {
                 drm_i915_private_t *dev_priv = dev->dev_private;
                 unsigned long irqflags;
 
@@ -447,16 +454,15 @@ ring_put_irq(struct intel_ring_buffer *ring, u32 flag)
         }
 }
 
-
-static void
+static bool
 bsd_ring_get_irq(struct intel_ring_buffer *ring)
 {
-    ring_get_irq(ring, GT_BSD_USER_INTERRUPT);
+        return ring_get_irq(ring, GT_BSD_USER_INTERRUPT);
 }
 static void
 bsd_ring_put_irq(struct intel_ring_buffer *ring)
 {
-    ring_put_irq(ring, GT_BSD_USER_INTERRUPT);
+        ring_put_irq(ring, GT_BSD_USER_INTERRUPT);
 }
 
 static int
@@ -846,16 +852,16 @@ gen6_ring_dispatch_execbuffer(struct intel_ring_buffer *ring,
        return 0;
 }
 
-static void
+static bool
 gen6_bsd_ring_get_irq(struct intel_ring_buffer *ring)
 {
-    ring_get_irq(ring, GT_GEN6_BSD_USER_INTERRUPT);
+        return ring_get_irq(ring, GT_GEN6_BSD_USER_INTERRUPT);
 }
 
 static void
 gen6_bsd_ring_put_irq(struct intel_ring_buffer *ring)
 {
-    ring_put_irq(ring, GT_GEN6_BSD_USER_INTERRUPT);
+        ring_put_irq(ring, GT_GEN6_BSD_USER_INTERRUPT);
 }
 
 /* ring buffer for Video Codec for Gen6+ */
@@ -876,16 +882,16 @@ static const struct intel_ring_buffer gen6_bsd_ring = {
 
 /* Blitter support (SandyBridge+) */
 
-static void
+static bool
 blt_ring_get_irq(struct intel_ring_buffer *ring)
 {
-    ring_get_irq(ring, GT_BLT_USER_INTERRUPT);
+        return ring_get_irq(ring, GT_BLT_USER_INTERRUPT);
 }
 
 static void
 blt_ring_put_irq(struct intel_ring_buffer *ring)
 {
-    ring_put_irq(ring, GT_BLT_USER_INTERRUPT);
+        ring_put_irq(ring, GT_BLT_USER_INTERRUPT);
 }
 
 
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.h b/drivers/gpu/drm/i915/intel_ringbuffer.h
index 9652e4600b5e..8e2e357ad6ee 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.h
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.h
@@ -54,8 +54,8 @@ struct  intel_ring_buffer {
         u32             irq_seqno;              /* last seq seem at irq time */
         u32             waiting_seqno;
         u32             sync_seqno[I915_NUM_RINGS-1];
-        u32             irq_refcount;
-        void            (*irq_get)(struct intel_ring_buffer *ring);
+        atomic_t        irq_refcount;
+        bool __must_check (*irq_get)(struct intel_ring_buffer *ring);
         void            (*irq_put)(struct intel_ring_buffer *ring);
 
         int             (*init)(struct intel_ring_buffer *ring);