gpu: nvgpu: Add CBC clean and invalidate

Bug 1409151

Change-Id: I232af159d402f818cf972498d721c3b57846ce74
Signed-off-by: Arto Merilainen <amerilainen@nvidia.com>

drivers/gpu/nvgpu/gk20a/gk20a.h

@@ -56,12 +56,19 @@ struct cooling_device_gk20a {
 	struct gk20a *g;
 };
 
+enum gk20a_cbc_op {
+	gk20a_cbc_op_clear,
+	gk20a_cbc_op_clean,
+	gk20a_cbc_op_invalidate,
+};
+
 struct gpu_ops {
 	struct {
 		int (*determine_L2_size_bytes)(struct gk20a *gk20a);
 		void (*set_max_ways_evict_last)(struct gk20a *g, u32 max_ways);
 		int (*init_comptags)(struct gk20a *g, struct gr_gk20a *gr);
-		int (*clear_comptags)(struct gk20a *g, u32 min, u32 max);
+		int (*cbc_ctrl)(struct gk20a *g, enum gk20a_cbc_op op,
+				u32 min, u32 max);
 		void (*set_zbc_color_entry)(struct gk20a *g,
 					    struct zbc_entry *color_val,
 					    u32 index);

drivers/gpu/nvgpu/gk20a/hw_ltc_gk20a.h

@@ -82,6 +82,14 @@ static inline u32 ltc_ltcs_ltss_cbc_ctrl1_r(void)
 {
 	return 0x0017e8c8;
 }
+static inline u32 ltc_ltcs_ltss_cbc_ctrl1_clean_active_f(void)
+{
+	return 0x1;
+}
+static inline u32 ltc_ltcs_ltss_cbc_ctrl1_invalidate_active_f(void)
+{
+	return 0x2;
+}
 static inline u32 ltc_ltcs_ltss_cbc_ctrl1_clear_v(u32 r)
 {
 	return (r >> 2) & 0x1;

drivers/gpu/nvgpu/gk20a/ltc_gk20a.c

@@ -108,10 +108,12 @@ static int gk20a_ltc_init_comptags(struct gk20a *g, struct gr_gk20a *gr)
 	return 0;
 }
 
-static int gk20a_ltc_clear_comptags(struct gk20a *g, u32 min, u32 max)
+static int gk20a_ltc_cbc_ctrl(struct gk20a *g, enum gk20a_cbc_op op,
+			      u32 min, u32 max)
 {
+	int err = 0;
 	struct gr_gk20a *gr = &g->gr;
-	u32 fbp, slice, ctrl1, val;
+	u32 fbp, slice, ctrl1, val, hw_op = 0;
 	unsigned long end_jiffies = jiffies +
 		msecs_to_jiffies(gk20a_get_gr_idle_timeout(g));
 	u32 delay = GR_IDLE_CHECK_DEFAULT;
@@ -124,13 +126,24 @@ static int gk20a_ltc_clear_comptags(struct gk20a *g, u32 min, u32 max)
 	if (gr->compbit_store.size == 0)
 		return 0;
 
+	mutex_lock(&g->mm.l2_op_lock);
+
+	if (op == gk20a_cbc_op_clear) {
 		gk20a_writel(g, ltc_ltcs_ltss_cbc_ctrl2_r(),
 			     ltc_ltcs_ltss_cbc_ctrl2_clear_lower_bound_f(min));
 		gk20a_writel(g, ltc_ltcs_ltss_cbc_ctrl3_r(),
 			     ltc_ltcs_ltss_cbc_ctrl3_clear_upper_bound_f(max));
+		hw_op = ltc_ltcs_ltss_cbc_ctrl1_clear_active_f();
+	} else if (op == gk20a_cbc_op_clean) {
+		hw_op = ltc_ltcs_ltss_cbc_ctrl1_clean_active_f();
+	} else if (op == gk20a_cbc_op_invalidate) {
+		hw_op = ltc_ltcs_ltss_cbc_ctrl1_invalidate_active_f();
+	} else {
+		BUG_ON(1);
+	}
+
 	gk20a_writel(g, ltc_ltcs_ltss_cbc_ctrl1_r(),
-		     gk20a_readl(g, ltc_ltcs_ltss_cbc_ctrl1_r()) |
+		     gk20a_readl(g, ltc_ltcs_ltss_cbc_ctrl1_r()) | hw_op);
-		     ltc_ltcs_ltss_cbc_ctrl1_clear_active_f());
 
 	for (fbp = 0; fbp < gr->num_fbps; fbp++) {
 		for (slice = 0; slice < slices_per_fbp; slice++) {
@@ -143,8 +156,7 @@ static int gk20a_ltc_clear_comptags(struct gk20a *g, u32 min, u32 max)
 
 			do {
 				val = gk20a_readl(g, ctrl1);
-				if (ltc_ltcs_ltss_cbc_ctrl1_clear_v(val) !=
+				if (!(val & hw_op))
-				    ltc_ltcs_ltss_cbc_ctrl1_clear_active_v())
 					break;
 
 				usleep_range(delay, delay * 2);
@@ -157,11 +169,13 @@ static int gk20a_ltc_clear_comptags(struct gk20a *g, u32 min, u32 max)
 			if (!time_before(jiffies, end_jiffies)) {
 				gk20a_err(dev_from_gk20a(g),
 					   "comp tag clear timeout\n");
-				return -EBUSY;
+				err = -EBUSY;
+				goto out;
 			}
 		}
 	}
-
+out:
+	mutex_unlock(&g->mm.l2_op_lock);
 	return 0;
 }
 
@@ -200,7 +214,7 @@ void gk20a_init_ltc(struct gpu_ops *gops)
 	gops->ltc.determine_L2_size_bytes = gk20a_determine_L2_size_bytes;
 	gops->ltc.set_max_ways_evict_last = gk20a_ltc_set_max_ways_evict_last;
 	gops->ltc.init_comptags = gk20a_ltc_init_comptags;
-	gops->ltc.clear_comptags = gk20a_ltc_clear_comptags;
+	gops->ltc.cbc_ctrl = gk20a_ltc_cbc_ctrl;
 	gops->ltc.set_zbc_color_entry = gk20a_ltc_set_zbc_color_entry;
 	gops->ltc.set_zbc_depth_entry = gk20a_ltc_set_zbc_depth_entry;
 	gops->ltc.clear_zbc_color_entry = gk20a_ltc_clear_zbc_color_entry;

drivers/gpu/nvgpu/gk20a/mm_gk20a.c

@@ -1380,7 +1380,7 @@ u64 gk20a_vm_map(struct vm_gk20a *vm,
 			gk20a_get_comptags(d, dmabuf, &comptags);
 
 			/* init/clear the ctag buffer */
-			g->ops.ltc.clear_comptags(g,
+			g->ops.ltc.cbc_ctrl(g, gk20a_cbc_op_clear,
 					  comptags.offset,
 					  comptags.offset + comptags.lines - 1);
 		}