drm/i915: Implement workaround for broken CS tlb on i830/845

Now that Chris Wilson demonstrated that the key for stability on early
gen 2 is to simple _never_ exchange the physical backing storage of
batch buffers I've tried a stab at a kernel solution. Doesn't look too
nefarious imho, now that I don't try to be too clever for my own good
any more.

v2: After discussing the various techniques, we've decided to always blit
batches on the suspect devices, but allow userspace to opt out of the
kernel workaround assume full responsibility for providing coherent
batches. The principal reason is that avoiding the blit does improve
performance in a few key microbenchmarks and also in cairo-trace
replays.
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet:
- Drop the hunk which uses HAS_BROKEN_CS_TLB to implement the ring
  wrap w/a. Suggested by Chris Wilson.
- Also add the ACTHD check from Chris Wilson for the error state
  dumping, so that we still catch batches when userspace opts out of
  the w/a.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drivers/gpu/drm/i915/i915_dma.c

@@ -989,6 +989,9 @@ static int i915_getparam(struct drm_device *dev, void *data,
 	case I915_PARAM_HAS_SECURE_BATCHES:
 		value = capable(CAP_SYS_ADMIN);
 		break;
+	case I915_PARAM_HAS_PINNED_BATCHES:
+		value = 1;
+		break;
 	default:
 		DRM_DEBUG_DRIVER("Unknown parameter %d\n",
 				 param->param);

drivers/gpu/drm/i915/i915_drv.h

@@ -1100,6 +1100,7 @@ struct drm_i915_gem_object {
 	 */
 	atomic_t pending_flip;
 };
+#define to_gem_object(obj) (&((struct drm_i915_gem_object *)(obj))->base)
 
 #define to_intel_bo(x) container_of(x, struct drm_i915_gem_object, base)
 
@@ -1199,6 +1200,9 @@ struct drm_i915_file_private {
 #define HAS_OVERLAY(dev)		(INTEL_INFO(dev)->has_overlay)
 #define OVERLAY_NEEDS_PHYSICAL(dev)	(INTEL_INFO(dev)->overlay_needs_physical)
 
+/* Early gen2 have a totally busted CS tlb and require pinned batches. */
+#define HAS_BROKEN_CS_TLB(dev)		(IS_I830(dev) || IS_845G(dev))
+
 /* With the 945 and later, Y tiling got adjusted so that it was 32 128-byte
  * rows, which changed the alignment requirements and fence programming.
  */

drivers/gpu/drm/i915/i915_gem_execbuffer.c

@@ -808,6 +808,8 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 
 		flags |= I915_DISPATCH_SECURE;
 	}
+	if (args->flags & I915_EXEC_IS_PINNED)
+		flags |= I915_DISPATCH_PINNED;
 
 	switch (args->flags & I915_EXEC_RING_MASK) {
 	case I915_EXEC_DEFAULT:

drivers/gpu/drm/i915/i915_irq.c

@@ -1087,6 +1087,18 @@ i915_error_first_batchbuffer(struct drm_i915_private *dev_priv,
 	if (!ring->get_seqno)
 		return NULL;
 
+	if (HAS_BROKEN_CS_TLB(dev_priv->dev)) {
+		u32 acthd = I915_READ(ACTHD);
+
+		if (WARN_ON(ring->id != RCS))
+			return NULL;
+
+		obj = ring->private;
+		if (acthd >= obj->gtt_offset &&
+		    acthd < obj->gtt_offset + obj->base.size)
+			return i915_error_object_create(dev_priv, obj);
+	}
+
 	seqno = ring->get_seqno(ring, false);
 	list_for_each_entry(obj, &dev_priv->mm.active_list, mm_list) {
 		if (obj->ring != ring)

drivers/gpu/drm/i915/intel_ringbuffer.c

@@ -547,9 +547,14 @@ static int init_render_ring(struct intel_ring_buffer *ring)
 
 static void render_ring_cleanup(struct intel_ring_buffer *ring)
 {
+	struct drm_device *dev = ring->dev;
+
 	if (!ring->private)
 		return;
 
+	if (HAS_BROKEN_CS_TLB(dev))
+		drm_gem_object_unreference(to_gem_object(ring->private));
+
 	cleanup_pipe_control(ring);
 }
 
@@ -969,6 +974,8 @@ i965_dispatch_execbuffer(struct intel_ring_buffer *ring,
 	return 0;
 }
 
+/* Just userspace ABI convention to limit the wa batch bo to a resonable size */
+#define I830_BATCH_LIMIT (256*1024)
 static int
 i830_dispatch_execbuffer(struct intel_ring_buffer *ring,
 				u32 offset, u32 len,
@@ -976,15 +983,47 @@ i830_dispatch_execbuffer(struct intel_ring_buffer *ring,
 {
 	int ret;
 
-	ret = intel_ring_begin(ring, 4);
-	if (ret)
-		return ret;
+	if (flags & I915_DISPATCH_PINNED) {
+		ret = intel_ring_begin(ring, 4);
+		if (ret)
+			return ret;
 
-	intel_ring_emit(ring, MI_BATCH_BUFFER);
-	intel_ring_emit(ring, offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
-	intel_ring_emit(ring, offset + len - 8);
-	intel_ring_emit(ring, 0);
-	intel_ring_advance(ring);
+		intel_ring_emit(ring, MI_BATCH_BUFFER);
+		intel_ring_emit(ring, offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
+		intel_ring_emit(ring, offset + len - 8);
+		intel_ring_emit(ring, MI_NOOP);
+		intel_ring_advance(ring);
+	} else {
+		struct drm_i915_gem_object *obj = ring->private;
+		u32 cs_offset = obj->gtt_offset;
+
+		if (len > I830_BATCH_LIMIT)
+			return -ENOSPC;
+
+		ret = intel_ring_begin(ring, 9+3);
+		if (ret)
+			return ret;
+		/* Blit the batch (which has now all relocs applied) to the stable batch
+		 * scratch bo area (so that the CS never stumbles over its tlb
+		 * invalidation bug) ... */
+		intel_ring_emit(ring, XY_SRC_COPY_BLT_CMD |
+				XY_SRC_COPY_BLT_WRITE_ALPHA |
+				XY_SRC_COPY_BLT_WRITE_RGB);
+		intel_ring_emit(ring, BLT_DEPTH_32 | BLT_ROP_GXCOPY | 4096);
+		intel_ring_emit(ring, 0);
+		intel_ring_emit(ring, (DIV_ROUND_UP(len, 4096) << 16) | 1024);
+		intel_ring_emit(ring, cs_offset);
+		intel_ring_emit(ring, 0);
+		intel_ring_emit(ring, 4096);
+		intel_ring_emit(ring, offset);
+		intel_ring_emit(ring, MI_FLUSH);
+
+		/* ... and execute it. */
+		intel_ring_emit(ring, MI_BATCH_BUFFER);
+		intel_ring_emit(ring, cs_offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
+		intel_ring_emit(ring, cs_offset + len - 8);
+		intel_ring_advance(ring);
+	}
 
 	return 0;
 }
@@ -1596,6 +1635,27 @@ int intel_init_render_ring_buffer(struct drm_device *dev)
 	ring->init = init_render_ring;
 	ring->cleanup = render_ring_cleanup;
 
+	/* Workaround batchbuffer to combat CS tlb bug. */
+	if (HAS_BROKEN_CS_TLB(dev)) {
+		struct drm_i915_gem_object *obj;
+		int ret;
+
+		obj = i915_gem_alloc_object(dev, I830_BATCH_LIMIT);
+		if (obj == NULL) {
+			DRM_ERROR("Failed to allocate batch bo\n");
+			return -ENOMEM;
+		}
+
+		ret = i915_gem_object_pin(obj, 0, true, false);
+		if (ret != 0) {
+			drm_gem_object_unreference(&obj->base);
+			DRM_ERROR("Failed to ping batch bo\n");
+			return ret;
+		}
+
+		ring->private = obj;
+	}
+
 	return intel_init_ring_buffer(dev, ring);
 }
 

drivers/gpu/drm/i915/intel_ringbuffer.h

@@ -94,6 +94,7 @@ struct  intel_ring_buffer {
 					       u32 offset, u32 length,
 					       unsigned flags);
 #define I915_DISPATCH_SECURE 0x1
+#define I915_DISPATCH_PINNED 0x2
 	void		(*cleanup)(struct intel_ring_buffer *ring);
 	int		(*sync_to)(struct intel_ring_buffer *ring,
 				   struct intel_ring_buffer *to,

include/uapi/drm/i915_drm.h

@@ -307,6 +307,7 @@ typedef struct drm_i915_irq_wait {
 #define I915_PARAM_HAS_PRIME_VMAP_FLUSH	 21
 #define I915_PARAM_RSVD_FOR_FUTURE_USE	 22
 #define I915_PARAM_HAS_SECURE_BATCHES	 23
+#define I915_PARAM_HAS_PINNED_BATCHES	 24
 
 typedef struct drm_i915_getparam {
 	int param;
@@ -677,6 +678,15 @@ struct drm_i915_gem_execbuffer2 {
  */
 #define I915_EXEC_SECURE		(1<<9)
 
+/** Inform the kernel that the batch is and will always be pinned. This
+ * negates the requirement for a workaround to be performed to avoid
+ * an incoherent CS (such as can be found on 830/845). If this flag is
+ * not passed, the kernel will endeavour to make sure the batch is
+ * coherent with the CS before execution. If this flag is passed,
+ * userspace assumes the responsibility for ensuring the same.
+ */
+#define I915_EXEC_IS_PINNED		(1<<10)
+
 #define I915_EXEC_CONTEXT_ID_MASK	(0xffffffff)
 #define i915_execbuffer2_set_context_id(eb2, context) \
 	(eb2).rsvd1 = context & I915_EXEC_CONTEXT_ID_MASK