drm/vc4: Allow using more than 256MB of CMA memory.

Until now, we've had to limit Raspberry Pi to 256MB of CMA memory to
keep from triggering the hardware addressing bug between the tile
binner and the tile alloc memory (where the top 4 bits come from the
tile state data array's address).

To work around that and allow more memory to be reserved for graphics,
allocate a single BO to store tile state data arrays and tile
alloc/overflow memory while the GPU is active, and make sure that that
one BO doesn't happen to cross a 256MB boundary.  With that in place,
we can allocate textures and shaders anywhere in system memory (still
contiguous, of course).

Signed-off-by: Eric Anholt <eric@anholt.net>
Link: http://patchwork.freedesktop.org/patch/msgid/20170327231025.19391-1-eric@anholt.net
Reviewed-by: Boris Brezillon <boris.brezillon@free-electrons.com>

1 files changed, 28 insertions, 33 deletions

diff --git a/drivers/gpu/drm/vc4/vc4_irq.c b/drivers/gpu/drm/vc4/vc4_irq.c
index 1384af9fc987..7d7af3a93d94 100644
--- a/drivers/gpu/drm/vc4/vc4_irq.c
+++ b/drivers/gpu/drm/vc4/vc4_irq.c
@@ -59,50 +59,45 @@ vc4_overflow_mem_work(struct work_struct *work)
 {
 	struct vc4_dev *vc4 =
 		container_of(work, struct vc4_dev, overflow_mem_work);
-	struct drm_device *dev = vc4->dev;
-	struct vc4_bo *bo;
+	struct vc4_bo *bo = vc4->bin_bo;
+	int bin_bo_slot;
+	struct vc4_exec_info *exec;
+	unsigned long irqflags;
 
-	bo = vc4_bo_create(dev, 256 * 1024, true);
-	if (IS_ERR(bo)) {
+	bin_bo_slot = vc4_v3d_get_bin_slot(vc4);
+	if (bin_bo_slot < 0) {
 		DRM_ERROR("Couldn't allocate binner overflow mem\n");
 		return;
 	}
 
-	/* If there's a job executing currently, then our previous
-	 * overflow allocation is getting used in that job and we need
-	 * to queue it to be released when the job is done.  But if no
-	 * job is executing at all, then we can free the old overflow
-	 * object direcctly.
-	 *
-	 * No lock necessary for this pointer since we're the only
-	 * ones that update the pointer, and our workqueue won't
-	 * reenter.
-	 */
-	if (vc4->overflow_mem) {
-		struct vc4_exec_info *current_exec;
-		unsigned long irqflags;
-
-		spin_lock_irqsave(&vc4->job_lock, irqflags);
-		current_exec = vc4_first_bin_job(vc4);
-		if (!current_exec)
-			current_exec = vc4_last_render_job(vc4);
-		if (current_exec) {
-			vc4->overflow_mem->seqno = current_exec->seqno;
-			list_add_tail(&vc4->overflow_mem->unref_head,
-				      &current_exec->unref_list);
-			vc4->overflow_mem = NULL;
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+
+	if (vc4->bin_alloc_overflow) {
+		/* If we had overflow memory allocated previously,
+		 * then that chunk will free when the current bin job
+		 * is done.  If we don't have a bin job running, then
+		 * the chunk will be done whenever the list of render
+		 * jobs has drained.
+		 */
+		exec = vc4_first_bin_job(vc4);
+		if (!exec)
+			exec = vc4_last_render_job(vc4);
+		if (exec) {
+			exec->bin_slots |= vc4->bin_alloc_overflow;
+		} else {
+			/* There's nothing queued in the hardware, so
+			 * the old slot is free immediately.
+			 */
+			vc4->bin_alloc_used &= ~vc4->bin_alloc_overflow;
 		}
-		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 	}
+	vc4->bin_alloc_overflow = BIT(bin_bo_slot);
 
-	if (vc4->overflow_mem)
-		drm_gem_object_unreference_unlocked(&vc4->overflow_mem->base.base);
-	vc4->overflow_mem = bo;
-
-	V3D_WRITE(V3D_BPOA, bo->base.paddr);
+	V3D_WRITE(V3D_BPOA, bo->base.paddr + bin_bo_slot * vc4->bin_alloc_size);
 	V3D_WRITE(V3D_BPOS, bo->base.base.size);
 	V3D_WRITE(V3D_INTCTL, V3D_INT_OUTOMEM);
 	V3D_WRITE(V3D_INTENA, V3D_INT_OUTOMEM);
+	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 }
 
 static void