drm/i915: Resurrect ring kicking for semaphores, selectively
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 14 Mar 2013 15:52:05 +0000 (17:52 +0200)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Sun, 17 Mar 2013 21:52:51 +0000 (22:52 +0100)
Once we thought we got semaphores working, we disabled kicking the ring
if hangcheck fired whilst waiting upon a ring as it was doing more harm
than good:

commit 4e0e90dcb8a7df1229c69e30abebb59b0b3c2a1f
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date:   Wed Dec 14 13:56:58 2011 +0100

    drm/i915: kicking rings stuck on semaphores considered harmful

However, life is never that easy and semaphores are still causing
problems whereby the value written by one ring (bcs) is not being
propagated to the waiter (rcs). Thus the waiter never wakes up and we
declare the GPU hung, which often has unfortunate consequences, even if
we successfully reset the GPU.

But the GPU is idle as it has completed the work, just didn't notify its
clients. So we can detect the incomplete wait during hang check and
probe the target ring to see if has indeed emitted the breadcrumb seqno
following the work and then and only then kick the waiter.

Based on a suggestion by Ben Widawsky.

v2: cross-check wait with iphdr. fix signaller calculation.

References: https://bugs.freedesktop.org/show_bug.cgi?id=54226
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Ben Widawsky <ben@bwidawsk.net>
Acked-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_irq.c

index 39621e5..63abf2f 100644 (file)
@@ -1787,6 +1787,37 @@ static bool i915_hangcheck_ring_idle(struct intel_ring_buffer *ring, bool *err)
        return false;
 }
 
+static bool semaphore_passed(struct intel_ring_buffer *ring)
+{
+       struct drm_i915_private *dev_priv = ring->dev->dev_private;
+       u32 acthd = intel_ring_get_active_head(ring) & HEAD_ADDR;
+       struct intel_ring_buffer *signaller;
+       u32 cmd, ipehr, acthd_min;
+
+       ipehr = I915_READ(RING_IPEHR(ring->mmio_base));
+       if ((ipehr & ~(0x3 << 16)) !=
+           (MI_SEMAPHORE_MBOX | MI_SEMAPHORE_COMPARE | MI_SEMAPHORE_REGISTER))
+               return false;
+
+       /* ACTHD is likely pointing to the dword after the actual command,
+        * so scan backwards until we find the MBOX.
+        */
+       acthd_min = max((int)acthd - 3 * 4, 0);
+       do {
+               cmd = ioread32(ring->virtual_start + acthd);
+               if (cmd == ipehr)
+                       break;
+
+               acthd -= 4;
+               if (acthd < acthd_min)
+                       return false;
+       } while (1);
+
+       signaller = &dev_priv->ring[(ring->id + (((ipehr >> 17) & 1) + 1)) % 3];
+       return i915_seqno_passed(signaller->get_seqno(signaller, false),
+                                ioread32(ring->virtual_start+acthd+4)+1);
+}
+
 static bool kick_ring(struct intel_ring_buffer *ring)
 {
        struct drm_device *dev = ring->dev;
@@ -1798,6 +1829,15 @@ static bool kick_ring(struct intel_ring_buffer *ring)
                I915_WRITE_CTL(ring, tmp);
                return true;
        }
+
+       if (INTEL_INFO(dev)->gen >= 6 &&
+           tmp & RING_WAIT_SEMAPHORE &&
+           semaphore_passed(ring)) {
+               DRM_ERROR("Kicking stuck semaphore on %s\n",
+                         ring->name);
+               I915_WRITE_CTL(ring, tmp);
+               return true;
+       }
        return false;
 }