diff --git a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h
index 063a5457e..451b207bd 100644
--- a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h
@@ -103,6 +103,13 @@ struct gk20a_channel_sync {
 void gk20a_channel_sync_destroy(struct gk20a_channel_sync *sync);
 struct gk20a_channel_sync *gk20a_channel_sync_create(struct channel_gk20a *c);
 bool gk20a_channel_sync_needs_sync_framework(struct channel_gk20a *c);
+
+#ifdef CONFIG_SYNC
 void gk20a_channel_cancel_pending_sema_waits(struct gk20a *g);
+#else
+static inline void gk20a_channel_cancel_pending_sema_waits(struct gk20a *g)
+{
+}
+#endif
 
 #endif
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.c b/drivers/gpu/nvgpu/gk20a/gk20a.c
index a65ca6c24..bed728cf4 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.c
@@ -47,6 +47,8 @@
 #include "nvgpu_common.h"
 #include "debug_gk20a.h"
 #include "ctrl_gk20a.h"
+#include "channel_sync_gk20a.h"
+
 #include "hw_mc_gk20a.h"
 #include "hw_timer_gk20a.h"
 #include "hw_bus_gk20a.h"
@@ -66,6 +68,7 @@
 #include "pstate/pstate.h"
 #endif
 
+
 #define CREATE_TRACE_POINTS
 #include <trace/events/gk20a.h>
 
@@ -708,6 +711,10 @@ void gk20a_remove_support(struct device *dev)
 	if (g->dbg_regops_tmp_buf)
 		kfree(g->dbg_regops_tmp_buf);
 
+	nvgpu_wait_for_deferred_interrupts(g);
+
+	gk20a_channel_cancel_pending_sema_waits(g);
+
 	if (g->pmu.remove_support)
 		g->pmu.remove_support(&g->pmu);
 
@@ -1740,6 +1747,7 @@ void gk20a_busy_noresume(struct device *dev)
  */
 void gk20a_driver_start_unload(struct gk20a *g)
 {
+	gk20a_dbg(gpu_dbg_shutdown, "Driver is now going down!\n");
 	g->driver_is_dying = 1;
 }
 
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.h b/drivers/gpu/nvgpu/gk20a/gk20a.h
index e8c1b30f5..dc130e339 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.h
@@ -1081,6 +1081,7 @@ enum gk20a_dbg_categories {
 	gpu_dbg_sema_v	= BIT(16), /* verbose semaphore debugging */
 	gpu_dbg_pmu_pstate = BIT(17), /* p state controlled by pmu */
 	gpu_dbg_xv      = BIT(18), /* XVE debugging */
+	gpu_dbg_shutdown = BIT(19), /* GPU shutdown tracing */
 	gpu_dbg_mem     = BIT(31), /* memory accesses, very verbose */
 };
 
diff --git a/drivers/gpu/nvgpu/pci.c b/drivers/gpu/nvgpu/pci.c
index 0ba36a30f..7cc058e40 100644
--- a/drivers/gpu/nvgpu/pci.c
+++ b/drivers/gpu/nvgpu/pci.c
@@ -356,10 +356,25 @@ static void nvgpu_pci_remove(struct pci_dev *pdev)
 	struct gk20a_platform *platform = gk20a_get_platform(&pdev->dev);
 	struct gk20a *g = get_gk20a(&pdev->dev);
 
-	if (g->remove_support)
-		g->remove_support(g->dev);
+	gk20a_dbg(gpu_dbg_shutdown, "Removing nvgpu driver!\n");
+	gk20a_driver_start_unload(g);
+
+	disable_irq(g->irq_stall);
+	devm_free_irq(&pdev->dev, g->irq_stall, g);
+	gk20a_dbg(gpu_dbg_shutdown, "IRQs disabled.\n");
+
+	/*
+	 * Wait for the driver to finish up all the IOCTLs it's working on
+	 * before cleaning up the driver's data structures.
+	 */
+	gk20a_wait_for_idle(&pdev->dev);
+	gk20a_dbg(gpu_dbg_shutdown, "Driver idle.\n");
 
 	gk20a_user_deinit(g->dev, &nvgpu_pci_class);
+	gk20a_dbg(gpu_dbg_shutdown, "User de-init done.\b");
+
+	if (g->remove_support)
+		g->remove_support(g->dev);
 
 	debugfs_remove_recursive(platform->debugfs);
 	debugfs_remove_recursive(platform->debugfs_alias);
@@ -368,6 +383,7 @@ static void nvgpu_pci_remove(struct pci_dev *pdev)
 
 	if (platform->remove)
 		platform->remove(g->dev);
+	gk20a_dbg(gpu_dbg_shutdown, "Platform remove done.\b");
 
 	kfree(g);
 }