gpu: nvgpu: Add busy/idle ref counting

Add reference counting for gk20a_busy() and gk20a_idle() so that
the driver can keep track of whether the driver is active.

Bug 1816516
Bug 1807277

Signed-off-by: Alex Waterman <alexw@nvidia.com>
Reviewed-on: http://git-master/r/1250019
(cherry picked from commit 7f558019735bb34cf00dd1ec17df1797501cff60)
Change-Id: I64c2ff1719673912ae127707e58ee557966c4d4d
Reviewed-on: http://git-master/r/1261922
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

drivers/gpu/nvgpu/gk20a/gk20a.c

@@ -1732,16 +1732,67 @@ void gk20a_busy_noresume(struct device *dev)
 	pm_runtime_get_noresume(dev);
 }
 
+/*
+ * Start the process for unloading the driver. Set g->driver_is_dying.
+ */
+void gk20a_driver_start_unload(struct gk20a *g)
+{
+	g->driver_is_dying = 1;
+}
+
+int gk20a_wait_for_idle(struct device *dev)
+{
+	struct gk20a *g = get_gk20a(dev);
+	int wait_length = 150; /* 3 second overall max wait. */
+
+	if (!g)
+		return -ENODEV;
+
+	while (atomic_read(&g->usage_count) && wait_length-- >= 0)
+		msleep(20);
+
+	if (wait_length < 0) {
+		pr_warn("%s: Timed out waiting for idle (%d)!\n",
+			__func__, atomic_read(&g->usage_count));
+		return -ETIMEDOUT;
+	}
+
+	return 0;
+}
+
+/*
+ * Check if the device can go busy. Basically if the driver is currently
+ * in the process of dying then do not let new places make the driver busy.
+ */
+static int gk20a_can_busy(struct gk20a *g)
+{
+	if (g->driver_is_dying)
+		return 0;
+	return 1;
+}
+
 int gk20a_busy(struct device *dev)
 {
 	int ret = 0;
 	struct gk20a *g = get_gk20a(dev);
+	struct gk20a_platform *platform;
+
+	if (!dev)
+		return -ENODEV;
+
+	platform = gk20a_get_platform(dev);
+
+	if (!g || !gk20a_can_busy(g))
+		return -ENODEV;
+
+	atomic_inc(&g->usage_count);
 
 	down_read(&g->busy_lock);
 	if (pm_runtime_enabled(dev)) {
 		ret = pm_runtime_get_sync(dev);
 		if (ret < 0) {
 			pm_runtime_put_noidle(dev);
+			atomic_dec(&g->usage_count);
 			goto fail;
 		}
 	} else {
@@ -1749,8 +1800,10 @@ int gk20a_busy(struct device *dev)
 			ret = gk20a_gpu_is_virtual(dev) ?
 				vgpu_pm_finalize_poweron(dev)
 				: gk20a_pm_finalize_poweron(dev);
-			if (ret)
+			if (ret) {
+				atomic_dec(&g->usage_count);
 				goto fail;
+			}
 		}
 	}
 
@@ -1769,6 +1822,17 @@ void gk20a_idle_nosuspend(struct device *dev)
 
 void gk20a_idle(struct device *dev)
 {
+	struct gk20a_platform *platform;
+	struct gk20a *g;
+
+	if (!dev)
+		return;
+
+	g = get_gk20a(dev);
+	platform = gk20a_get_platform(dev);
+
+	atomic_dec(&g->usage_count);
+
 	if (pm_runtime_enabled(dev)) {
 #ifdef CONFIG_PM
 		if (atomic_read(&dev->power.usage_count) == 1)

drivers/gpu/nvgpu/gk20a/gk20a.h

@@ -34,6 +34,7 @@ struct acr_desc;
 #include <linux/irqreturn.h>
 #include <linux/tegra-soc.h>
 #include <linux/version.h>
+#include <linux/atomic.h>
 
 #include "../../../arch/arm/mach-tegra/iomap.h"
 
@@ -784,6 +785,8 @@ struct gk20a {
 	struct device *dev;
 	struct platform_device *host1x_dev;
 
+	atomic_t usage_count;
+	int driver_is_dying;
 
 	struct resource *reg_mem;
 	void __iomem *regs;
@@ -1238,6 +1241,9 @@ int gk20a_do_unidle(void);
 int __gk20a_do_idle(struct device *dev, bool force_reset);
 int __gk20a_do_unidle(struct device *dev);
 
+void gk20a_driver_start_unload(struct gk20a *g);
+int gk20a_wait_for_idle(struct device *dev);
+
 #define NVGPU_GPU_ARCHITECTURE_SHIFT 4
 
 /* constructs unique and compact GPUID from nvgpu_gpu_characteristics