drm/vc4: Cache LBM allocations to avoid double-buffering

LBM is only relevant for each active dlist, so there is
no need to double-buffer the allocations.

Cache the allocations per plane so that we can ensure the
allocations are possible.

Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.com>

Author: 6by9

drivers/gpu/drm/vc4/tests/vc4_test_lbm_size.c

@@ -188,6 +188,7 @@ static void drm_vc4_test_vc4_lbm_size(struct kunit *test)
 	struct drm_framebuffer *fb;
 	struct drm_plane *plane;
 	struct drm_crtc *crtc;
+	struct vc4_dev *vc4;
 	unsigned int i;
 	int ret;
 
@@ -248,7 +249,12 @@ static void drm_vc4_test_vc4_lbm_size(struct kunit *test)
 	ret = drm_atomic_check_only(state);
 	KUNIT_ASSERT_EQ(test, ret, 0);
 
-	KUNIT_EXPECT_EQ(test, vc4_plane_state->lbm.size, params->expected_lbm_size);
+	vc4 = to_vc4_dev(state->dev);
+	KUNIT_ASSERT_NOT_NULL(test, vc4);
+	KUNIT_ASSERT_NOT_NULL(test, vc4->hvs);
+	KUNIT_EXPECT_EQ(test,
+			vc4->hvs->lbm_refcounts[vc4_plane_state->lbm_handle].size,
+			params->expected_lbm_size);
 
 	for (i = 0; i < 2; i++) {
 		KUNIT_EXPECT_EQ(test,

drivers/gpu/drm/vc4/vc4_drv.h

@@ -322,6 +322,19 @@ struct vc4_v3d {
 	struct debugfs_regset32 regset;
 };
 
+#define VC4_NUM_LBM_HANDLES 64
+struct vc4_lbm_refcounts {
+	refcount_t refcount;
+
+	/* Allocation size */
+	size_t size;
+	/* Our allocation in LBM. */
+	struct drm_mm_node lbm;
+
+	/* Pointer back to the HVS structure */
+	struct vc4_hvs *hvs;
+};
+
 #define VC4_NUM_UPM_HANDLES 32
 struct vc4_upm_refcounts {
 	refcount_t refcount;
@@ -363,6 +376,8 @@ struct vc4_hvs {
 
 	/* Memory manager for the LBM memory used by HVS scaling. */
 	struct drm_mm lbm_mm;
+	struct ida lbm_handles;
+	struct vc4_lbm_refcounts lbm_refcounts[VC4_NUM_LBM_HANDLES + 1];
 
 	/* Memory manager for the UPM memory used for prefetching. */
 	struct drm_mm upm_mm;
@@ -462,7 +477,7 @@ struct vc4_plane_state {
 	bool is_yuv;
 
 	/* Our allocation in LBM for temporary storage during scaling. */
-	struct drm_mm_node lbm;
+	unsigned int lbm_handle;
 
 	/* The Unified Pre-Fetcher Handle */
 	unsigned int upm_handle[DRM_FORMAT_MAX_PLANES];

drivers/gpu/drm/vc4/vc4_hvs.c

@@ -379,6 +379,27 @@ static int vc4_hvs_debugfs_dlist_allocs(struct seq_file *m, void *data)
 	return 0;
 }
 
+static int vc4_hvs_debugfs_lbm_allocs(struct seq_file *m, void *data)
+{
+	struct drm_debugfs_entry *entry = m->private;
+	struct drm_device *dev = entry->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_hvs *hvs = vc4->hvs;
+	struct drm_printer p = drm_seq_file_printer(m);
+	struct vc4_lbm_refcounts *refcount;
+	unsigned int i;
+
+	drm_printf(&p, "LBM Handles:\n");
+	for (i = 0; i < VC4_NUM_LBM_HANDLES; i++) {
+		refcount = &hvs->lbm_refcounts[i];
+		drm_printf(&p, "handle %u: refcount %u, size %zu [%08llx + %08llx]\n",
+			   i, refcount_read(&refcount->refcount), refcount->size,
+			   refcount->lbm.start, refcount->lbm.size);
+	}
+
+	return 0;
+}
+
 /* The filter kernel is composed of dwords each containing 3 9-bit
  * signed integers packed next to each other.
  */
@@ -1511,6 +1532,8 @@ int vc4_hvs_debugfs_init(struct drm_minor *minor)
 		drm_debugfs_add_file(drm, "hvs_dlists", vc4_hvs_debugfs_dlist, NULL);
 	}
 
+	drm_debugfs_add_file(drm, "hvs_lbm", vc4_hvs_debugfs_lbm_allocs, NULL);
+
 	drm_debugfs_add_file(drm, "hvs_underrun", vc4_hvs_debugfs_underrun, NULL);
 
 	drm_debugfs_add_file(drm, "hvs_dlist_allocs", vc4_hvs_debugfs_dlist_allocs, NULL);
@@ -1627,6 +1650,7 @@ struct vc4_hvs *__vc4_hvs_alloc(struct vc4_dev *vc4,
 	}
 
 	drm_mm_init(&hvs->lbm_mm, 0, lbm_size);
+	ida_init(&hvs->lbm_handles);
 
 	if (vc4->gen >= VC4_GEN_6_C) {
 		ida_init(&hvs->upm_handles);

drivers/gpu/drm/vc4/vc4_plane.c

@@ -294,12 +294,12 @@ struct drm_plane_state *vc4_plane_duplicate_state(struct drm_plane *plane)
 	if (!vc4_state)
 		return NULL;
 
-	memset(&vc4_state->lbm, 0, sizeof(vc4_state->lbm));
-
 	for (i = 0; i < DRM_FORMAT_MAX_PLANES; i++) {
 		if (vc4_state->upm_handle[i])
 			refcount_inc(&hvs->upm_refcounts[vc4_state->upm_handle[i]].refcount);
 	}
+	if (vc4_state->lbm_handle)
+		refcount_inc(&hvs->lbm_refcounts[vc4_state->lbm_handle].refcount);
 
 	vc4_state->dlist_initialized = 0;
 
@@ -319,6 +319,21 @@ struct drm_plane_state *vc4_plane_duplicate_state(struct drm_plane *plane)
 	return &vc4_state->base;
 }
 
+static void vc4_plane_release_lbm_ida(struct vc4_hvs *hvs, unsigned int lbm_handle)
+{
+	struct vc4_lbm_refcounts *refcount = &hvs->lbm_refcounts[lbm_handle];
+	unsigned long irqflags;
+
+	spin_lock_irqsave(&hvs->mm_lock, irqflags);
+	drm_mm_remove_node(&refcount->lbm);
+	spin_unlock_irqrestore(&hvs->mm_lock, irqflags);
+	refcount->lbm.start = 0;
+	refcount->lbm.size = 0;
+	refcount->size = 0;
+
+	ida_free(&hvs->lbm_handles, lbm_handle);
+}
+
 static void vc4_plane_release_upm_ida(struct vc4_hvs *hvs, unsigned int upm_handle)
 {
 	struct vc4_upm_refcounts *refcount = &hvs->upm_refcounts[upm_handle];
@@ -342,12 +357,13 @@ void vc4_plane_destroy_state(struct drm_plane *plane,
 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
 	unsigned int i;
 
-	if (drm_mm_node_allocated(&vc4_state->lbm)) {
-		unsigned long irqflags;
+	if (vc4_state->lbm_handle) {
+		struct vc4_lbm_refcounts *refcount;
 
-		spin_lock_irqsave(&hvs->mm_lock, irqflags);
-		drm_mm_remove_node(&vc4_state->lbm);
-		spin_unlock_irqrestore(&hvs->mm_lock, irqflags);
+		refcount = &hvs->lbm_refcounts[vc4_state->lbm_handle];
+
+		if (refcount_dec_and_test(&refcount->refcount))
+			vc4_plane_release_lbm_ida(hvs, vc4_state->lbm_handle);
 	}
 
 	for (i = 0; i < DRM_FORMAT_MAX_PLANES; i++) {
@@ -939,10 +955,14 @@ static int vc4_plane_allocate_lbm(struct drm_plane_state *state)
 {
 	struct drm_device *drm = state->plane->dev;
 	struct vc4_dev *vc4 = to_vc4_dev(drm);
+	struct vc4_hvs *hvs = vc4->hvs;
 	struct drm_plane *plane = state->plane;
 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
+	struct vc4_lbm_refcounts *refcount;
 	unsigned long irqflags;
+	int lbm_handle;
 	u32 lbm_size;
+	int ret;
 
 	lbm_size = vc4_lbm_size(state);
 	if (!lbm_size)
@@ -966,29 +986,71 @@ static int vc4_plane_allocate_lbm(struct drm_plane_state *state)
 	/* Allocate the LBM memory that the HVS will use for temporary
 	 * storage due to our scaling/format conversion.
 	 */
-	if (!drm_mm_node_allocated(&vc4_state->lbm)) {
-		int ret;
+	lbm_handle = vc4_state->lbm_handle;
+	if (lbm_handle &&
+	    hvs->lbm_refcounts[lbm_handle].size == lbm_size) {
+		/* Allocation is the same size as the previous user of
+		 * the plane. Keep the allocation.
+		 */
+		vc4_state->lbm_handle = lbm_handle;
+	} else {
+		if (lbm_handle &&
+		    refcount_dec_and_test(&hvs->lbm_refcounts[lbm_handle].refcount)) {
+			vc4_plane_release_lbm_ida(hvs, lbm_handle);
+			vc4_state->lbm_handle = 0;
+		}
+
+		lbm_handle = ida_alloc_range(&hvs->lbm_handles, 1,
+					     VC4_NUM_LBM_HANDLES,
+					     GFP_KERNEL);
+		if (lbm_handle < 0) {
+			drm_err(drm, "Out of lbm_handles\n");
+			return lbm_handle;
+		}
+		vc4_state->lbm_handle = lbm_handle;
+
+		refcount = &hvs->lbm_refcounts[lbm_handle];
+		refcount_set(&refcount->refcount, 1);
+		refcount->size = lbm_size;
 
-		spin_lock_irqsave(&vc4->hvs->mm_lock, irqflags);
+		spin_lock_irqsave(&hvs->mm_lock, irqflags);
 		ret = drm_mm_insert_node_generic(&vc4->hvs->lbm_mm,
-						 &vc4_state->lbm,
+						 &refcount->lbm,
 						 lbm_size, 1,
 						 0, 0);
-		spin_unlock_irqrestore(&vc4->hvs->mm_lock, irqflags);
+		spin_unlock_irqrestore(&hvs->mm_lock, irqflags);
 
 		if (ret) {
 			drm_err(drm, "Failed to allocate LBM entry: %d\n", ret);
+			refcount_set(&refcount->refcount, 0);
+			ida_free(&hvs->lbm_handles, lbm_handle);
+			vc4_state->lbm_handle = 0;
 			return ret;
 		}
-	} else {
-		WARN_ON_ONCE(lbm_size != vc4_state->lbm.size);
 	}
 
-	vc4_state->dlist[vc4_state->lbm_offset] = vc4_state->lbm.start;
+	vc4_state->dlist[vc4_state->lbm_offset] = hvs->lbm_refcounts[lbm_handle].lbm.start;
 
 	return 0;
 }
 
+static void vc4_plane_free_lbm(struct drm_plane_state *state)
+{
+	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
+	struct drm_device *drm = state->plane->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(drm);
+	struct vc4_hvs *hvs = vc4->hvs;
+	unsigned int lbm_handle;
+
+	lbm_handle = vc4_state->lbm_handle;
+	if (!lbm_handle)
+		return;
+
+	if (refcount_dec_and_test(&hvs->lbm_refcounts[lbm_handle].refcount))
+		vc4_plane_release_lbm_ida(hvs, lbm_handle);
+	vc4_state->lbm_handle = 0;
+}
+
 static int vc6_plane_allocate_upm(struct drm_plane_state *state)
 {
 	const struct drm_format_info *info = state->fb->format;
@@ -2174,9 +2236,10 @@ int vc4_plane_atomic_check(struct drm_plane *plane,
 		struct drm_plane_state *old_plane_state =
 				drm_atomic_get_old_plane_state(state, plane);
 
-		if (vc4->gen >= VC4_GEN_6_C && old_plane_state &&
-		    plane_enabled(old_plane_state)) {
-			vc6_plane_free_upm(new_plane_state);
+		if (old_plane_state && plane_enabled(old_plane_state)) {
+			if (vc4->gen >= VC4_GEN_6_C)
+				vc6_plane_free_upm(new_plane_state);
+			vc4_plane_free_lbm(new_plane_state);
 		}
 		return 0;
 	}