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, 150 insertions, 0 deletions

diff --git a/drivers/gpu/drm/vc4/vc4_v3d.c b/drivers/gpu/drm/vc4/vc4_v3d.c
index 7cc346ad9b0b..a88078d7c9d1 100644
--- a/drivers/gpu/drm/vc4/vc4_v3d.c
+++ b/drivers/gpu/drm/vc4/vc4_v3d.c
@@ -156,6 +156,144 @@ static void vc4_v3d_init_hw(struct drm_device *dev)
 	V3D_WRITE(V3D_VPMBASE, 0);
 }
 
+int vc4_v3d_get_bin_slot(struct vc4_dev *vc4)
+{
+	struct drm_device *dev = vc4->dev;
+	unsigned long irqflags;
+	int slot;
+	uint64_t seqno = 0;
+	struct vc4_exec_info *exec;
+
+try_again:
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+	slot = ffs(~vc4->bin_alloc_used);
+	if (slot != 0) {
+		/* Switch from ffs() bit index to a 0-based index. */
+		slot--;
+		vc4->bin_alloc_used |= BIT(slot);
+		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+		return slot;
+	}
+
+	/* Couldn't find an open slot.  Wait for render to complete
+	 * and try again.
+	 */
+	exec = vc4_last_render_job(vc4);
+	if (exec)
+		seqno = exec->seqno;
+	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+
+	if (seqno) {
+		int ret = vc4_wait_for_seqno(dev, seqno, ~0ull, true);
+
+		if (ret == 0)
+			goto try_again;
+
+		return ret;
+	}
+
+	return -ENOMEM;
+}
+
+/**
+ * vc4_allocate_bin_bo() - allocates the memory that will be used for
+ * tile binning.
+ *
+ * The binner has a limitation that the addresses in the tile state
+ * buffer that point into the tile alloc buffer or binner overflow
+ * memory only have 28 bits (256MB), and the top 4 on the bus for
+ * tile alloc references end up coming from the tile state buffer's
+ * address.
+ *
+ * To work around this, we allocate a single large buffer while V3D is
+ * in use, make sure that it has the top 4 bits constant across its
+ * entire extent, and then put the tile state, tile alloc, and binner
+ * overflow memory inside that buffer.
+ *
+ * This creates a limitation where we may not be able to execute a job
+ * if it doesn't fit within the buffer that we allocated up front.
+ * However, it turns out that 16MB is "enough for anybody", and
+ * real-world applications run into allocation failures from the
+ * overall CMA pool before they make scenes complicated enough to run
+ * out of bin space.
+ */
+int
+vc4_allocate_bin_bo(struct drm_device *drm)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(drm);
+	struct vc4_v3d *v3d = vc4->v3d;
+	uint32_t size = 16 * 1024 * 1024;
+	int ret = 0;
+	struct list_head list;
+
+	/* We may need to try allocating more than once to get a BO
+	 * that doesn't cross 256MB.  Track the ones we've allocated
+	 * that failed so far, so that we can free them when we've got
+	 * one that succeeded (if we freed them right away, our next
+	 * allocation would probably be the same chunk of memory).
+	 */
+	INIT_LIST_HEAD(&list);
+
+	while (true) {
+		struct vc4_bo *bo = vc4_bo_create(drm, size, true);
+
+		if (IS_ERR(bo)) {
+			ret = PTR_ERR(bo);
+
+			dev_err(&v3d->pdev->dev,
+				"Failed to allocate memory for tile binning: "
+				"%d. You may need to enable CMA or give it "
+				"more memory.",
+				ret);
+			break;
+		}
+
+		/* Check if this BO won't trigger the addressing bug. */
+		if ((bo->base.paddr & 0xf0000000) ==
+		    ((bo->base.paddr + bo->base.base.size - 1) & 0xf0000000)) {
+			vc4->bin_bo = bo;
+
+			/* Set up for allocating 512KB chunks of
+			 * binner memory.  The biggest allocation we
+			 * need to do is for the initial tile alloc +
+			 * tile state buffer.  We can render to a
+			 * maximum of ((2048*2048) / (32*32) = 4096
+			 * tiles in a frame (until we do floating
+			 * point rendering, at which point it would be
+			 * 8192).  Tile state is 48b/tile (rounded to
+			 * a page), and tile alloc is 32b/tile
+			 * (rounded to a page), plus a page of extra,
+			 * for a total of 320kb for our worst-case.
+			 * We choose 512kb so that it divides evenly
+			 * into our 16MB, and the rest of the 512kb
+			 * will be used as storage for the overflow
+			 * from the initial 32b CL per bin.
+			 */
+			vc4->bin_alloc_size = 512 * 1024;
+			vc4->bin_alloc_used = 0;
+			vc4->bin_alloc_overflow = 0;
+			WARN_ON_ONCE(sizeof(vc4->bin_alloc_used) * 8 !=
+				     bo->base.base.size / vc4->bin_alloc_size);
+
+			break;
+		}
+
+		/* Put it on the list to free later, and try again. */
+		list_add(&bo->unref_head, &list);
+	}
+
+	/* Free all the BOs we allocated but didn't choose. */
+	while (!list_empty(&list)) {
+		struct vc4_bo *bo = list_last_entry(&list,
+						    struct vc4_bo, unref_head);
+
+		list_del(&bo->unref_head);
+		drm_gem_object_put_unlocked(&bo->base.base);
+	}
+
+	return ret;
+}
+
 #ifdef CONFIG_PM
 static int vc4_v3d_runtime_suspend(struct device *dev)
 {
@@ -164,6 +302,9 @@ static int vc4_v3d_runtime_suspend(struct device *dev)
 
 	vc4_irq_uninstall(vc4->dev);
 
+	drm_gem_object_put_unlocked(&vc4->bin_bo->base.base);
+	vc4->bin_bo = NULL;
+
 	return 0;
 }
 
@@ -171,6 +312,11 @@ static int vc4_v3d_runtime_resume(struct device *dev)
 {
 	struct vc4_v3d *v3d = dev_get_drvdata(dev);
 	struct vc4_dev *vc4 = v3d->vc4;
+	int ret;
+
+	ret = vc4_allocate_bin_bo(vc4->dev);
+	if (ret)
+		return ret;
 
 	vc4_v3d_init_hw(vc4->dev);
 	vc4_irq_postinstall(vc4->dev);
@@ -208,6 +354,10 @@ static int vc4_v3d_bind(struct device *dev, struct device *master, void *data)
 		return -EINVAL;
 	}
 
+	ret = vc4_allocate_bin_bo(drm);
+	if (ret)
+		return ret;
+
 	/* Reset the binner overflow address/size at setup, to be sure
 	 * we don't reuse an old one.
 	 */