diff --git a/drivers/gpu/nvgpu/gk20a/css_gr_gk20a.c b/drivers/gpu/nvgpu/gk20a/css_gr_gk20a.c
index 0fc39bf40..764ffe1a5 100644
--- a/drivers/gpu/nvgpu/gk20a/css_gr_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/css_gr_gk20a.c
@@ -22,9 +22,6 @@
  * DEALINGS IN THE SOFTWARE.
  */
 
-#include <linux/dma-mapping.h>
-#include <linux/dma-buf.h>
-
 #include <nvgpu/bitops.h>
 #include <nvgpu/kmem.h>
 #include <nvgpu/lock.h>
@@ -61,14 +58,14 @@
 #define CSS_MAX_PERFMON_IDS	256
 
 /* reports whether the hw queue overflowed */
-static inline bool css_hw_get_overflow_status(struct gk20a *g)
+bool css_hw_get_overflow_status(struct gk20a *g)
 {
 	const u32 st = perf_pmasys_control_membuf_status_overflowed_f();
 	return st == (gk20a_readl(g, perf_pmasys_control_r()) & st);
 }
 
 /* returns how many pending snapshot entries are pending */
-static inline u32 css_hw_get_pending_snapshots(struct gk20a *g)
+u32 css_hw_get_pending_snapshots(struct gk20a *g)
 {
 	return gk20a_readl(g, perf_pmasys_mem_bytes_r()) /
 			sizeof(struct gk20a_cs_snapshot_fifo_entry);
@@ -245,7 +242,7 @@ static void css_gr_free_shared_data(struct gr_gk20a *gr)
 }
 
 
-static struct gk20a_cs_snapshot_client*
+struct gk20a_cs_snapshot_client*
 css_gr_search_client(struct nvgpu_list_node *clients, u32 perfmon)
 {
 	struct gk20a_cs_snapshot_client *client;
diff --git a/drivers/gpu/nvgpu/gk20a/css_gr_gk20a.h b/drivers/gpu/nvgpu/gk20a/css_gr_gk20a.h
index bee8f4fbd..8aefe8487 100644
--- a/drivers/gpu/nvgpu/gk20a/css_gr_gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/css_gr_gk20a.h
@@ -129,6 +129,8 @@ struct gk20a_cs_snapshot {
 	struct gk20a_cs_snapshot_fifo_entry	*hw_get;
 };
 
+bool css_hw_get_overflow_status(struct gk20a *g);
+u32 css_hw_get_pending_snapshots(struct gk20a *g);
 void css_hw_set_handled_snapshots(struct gk20a *g, u32 done);
 int css_hw_enable_snapshot(struct channel_gk20a *ch,
 				struct gk20a_cs_snapshot_client *cs_client);
@@ -140,5 +142,7 @@ u32 css_gr_release_perfmon_ids(struct gk20a_cs_snapshot *data,
 				      u32 count);
 int css_hw_check_data_available(struct channel_gk20a *ch, u32 *pending,
 					bool *hw_overflow);
+struct gk20a_cs_snapshot_client*
+css_gr_search_client(struct nvgpu_list_node *clients, u32 perfmon);
 
 #endif /* CSS_GR_GK20A_H */
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.h b/drivers/gpu/nvgpu/gk20a/gk20a.h
index bc4bd6821..2605f9542 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.h
@@ -1185,6 +1185,8 @@ struct gpu_ops {
 				      u32 count);
 		int (*detach_snapshot)(struct channel_gk20a *ch,
 				struct gk20a_cs_snapshot_client *client);
+		bool (*get_overflow_status)(struct gk20a *g);
+		u32 (*get_pending_snapshots)(struct gk20a *g);
 	} css;
 #endif
 	struct {
diff --git a/drivers/gpu/nvgpu/gm20b/hal_gm20b.c b/drivers/gpu/nvgpu/gm20b/hal_gm20b.c
index 715946beb..6bc13a7fd 100644
--- a/drivers/gpu/nvgpu/gm20b/hal_gm20b.c
+++ b/drivers/gpu/nvgpu/gm20b/hal_gm20b.c
@@ -625,6 +625,8 @@ static const struct gpu_ops gm20b_ops = {
 		.set_handled_snapshots = css_hw_set_handled_snapshots,
 		.allocate_perfmon_ids = css_gr_allocate_perfmon_ids,
 		.release_perfmon_ids = css_gr_release_perfmon_ids,
+		.get_overflow_status = css_hw_get_overflow_status,
+		.get_pending_snapshots = css_hw_get_pending_snapshots,
 	},
 #endif
 	.falcon = {
diff --git a/drivers/gpu/nvgpu/gp106/hal_gp106.c b/drivers/gpu/nvgpu/gp106/hal_gp106.c
index 632f10636..f46f20efb 100644
--- a/drivers/gpu/nvgpu/gp106/hal_gp106.c
+++ b/drivers/gpu/nvgpu/gp106/hal_gp106.c
@@ -742,6 +742,8 @@ static const struct gpu_ops gp106_ops = {
 		.set_handled_snapshots = css_hw_set_handled_snapshots,
 		.allocate_perfmon_ids = css_gr_allocate_perfmon_ids,
 		.release_perfmon_ids = css_gr_release_perfmon_ids,
+		.get_overflow_status = css_hw_get_overflow_status,
+		.get_pending_snapshots = css_hw_get_pending_snapshots,
 	},
 #endif
 	.xve = {
diff --git a/drivers/gpu/nvgpu/gp10b/hal_gp10b.c b/drivers/gpu/nvgpu/gp10b/hal_gp10b.c
index 80e07b78c..d8e69dd80 100644
--- a/drivers/gpu/nvgpu/gp10b/hal_gp10b.c
+++ b/drivers/gpu/nvgpu/gp10b/hal_gp10b.c
@@ -672,6 +672,8 @@ static const struct gpu_ops gp10b_ops = {
 		.set_handled_snapshots = css_hw_set_handled_snapshots,
 		.allocate_perfmon_ids = css_gr_allocate_perfmon_ids,
 		.release_perfmon_ids = css_gr_release_perfmon_ids,
+		.get_overflow_status = css_hw_get_overflow_status,
+		.get_pending_snapshots = css_hw_get_pending_snapshots,
 	},
 #endif
 	.falcon = {
diff --git a/drivers/gpu/nvgpu/gv100/hal_gv100.c b/drivers/gpu/nvgpu/gv100/hal_gv100.c
index 4e1c3fb80..b881aaf7e 100644
--- a/drivers/gpu/nvgpu/gv100/hal_gv100.c
+++ b/drivers/gpu/nvgpu/gv100/hal_gv100.c
@@ -832,6 +832,8 @@ static const struct gpu_ops gv100_ops = {
 		.set_handled_snapshots = gv11b_css_hw_set_handled_snapshots,
 		.allocate_perfmon_ids = css_gr_allocate_perfmon_ids,
 		.release_perfmon_ids = css_gr_release_perfmon_ids,
+		.get_overflow_status = gv11b_css_hw_get_overflow_status,
+		.get_pending_snapshots = gv11b_css_hw_get_pending_snapshots,
 	},
 #endif
 	.xve = {
diff --git a/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.c b/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.c
index b4e2cb791..7f03be64f 100644
--- a/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.c
+++ b/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.c
@@ -22,9 +22,6 @@
  * DEALINGS IN THE SOFTWARE.
  */
 
-#include <linux/dma-mapping.h>
-#include <linux/dma-buf.h>
-
 #include <nvgpu/bitops.h>
 #include <nvgpu/kmem.h>
 #include <nvgpu/lock.h>
@@ -33,7 +30,9 @@
 #include <nvgpu/sizes.h>
 #include <nvgpu/enabled.h>
 #include <nvgpu/log.h>
+#include <nvgpu/nvgpu_mem.h>
 #include <nvgpu/bug.h>
+#include <nvgpu/dma.h>
 
 #include "gk20a/gk20a.h"
 #include "gk20a/css_gr_gk20a.h"
@@ -44,14 +43,14 @@
 
 
 /* reports whether the hw queue overflowed */
-static inline bool css_hw_get_overflow_status(struct gk20a *g)
+bool gv11b_css_hw_get_overflow_status(struct gk20a *g)
 {
 	const u32 st = perf_pmasys_control_membuf_status_overflowed_f();
 	return st == (gk20a_readl(g, perf_pmasys_control_r()) & st);
 }
 
 /* returns how many pending snapshot entries are pending */
-static inline u32 css_hw_get_pending_snapshots(struct gk20a *g)
+u32 gv11b_css_hw_get_pending_snapshots(struct gk20a *g)
 {
 	return gk20a_readl(g, perf_pmasys_mem_bytes_r()) /
 			sizeof(struct gk20a_cs_snapshot_fifo_entry);
@@ -74,7 +73,7 @@ static void gv11b_css_hw_reset_streaming(struct gk20a *g)
 			perf_pmasys_control_membuf_clear_status_doit_f());
 
 	/* pointing all pending snapshots as handled */
-	gv11b_css_hw_set_handled_snapshots(g, css_hw_get_pending_snapshots(g));
+	gv11b_css_hw_set_handled_snapshots(g, gv11b_css_hw_get_pending_snapshots(g));
 }
 
 /* informs hw how many snapshots have been processed (frees up fifo space) */
@@ -199,10 +198,10 @@ int gv11b_css_hw_check_data_available(struct channel_gk20a *ch, u32 *pending,
 	if (!css->hw_snapshot)
 		return -EINVAL;
 
-	*pending = css_hw_get_pending_snapshots(g);
+	*pending = gv11b_css_hw_get_pending_snapshots(g);
 	if (!*pending)
 		return 0;
 
-	*hw_overflow = css_hw_get_overflow_status(g);
+	*hw_overflow = gv11b_css_hw_get_overflow_status(g);
 	return 0;
 }
diff --git a/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.h b/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.h
index 186f5fcc6..26db4ee83 100644
--- a/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.h
+++ b/drivers/gpu/nvgpu/gv11b/css_gr_gv11b.h
@@ -31,5 +31,8 @@ void gv11b_css_hw_disable_snapshot(struct gr_gk20a *gr);
 int gv11b_css_hw_check_data_available(struct channel_gk20a *ch, u32 *pending,
 			bool *hw_overflow);
 void gv11b_css_hw_set_handled_snapshots(struct gk20a *g, u32 done);
+bool gv11b_css_hw_get_overflow_status(struct gk20a *g);
+u32 gv11b_css_hw_get_pending_snapshots(struct gk20a *g);
+
 
 #endif /* CSS_GR_GV11B_H */
diff --git a/drivers/gpu/nvgpu/gv11b/hal_gv11b.c b/drivers/gpu/nvgpu/gv11b/hal_gv11b.c
index ab9ff7076..84d8a811f 100644
--- a/drivers/gpu/nvgpu/gv11b/hal_gv11b.c
+++ b/drivers/gpu/nvgpu/gv11b/hal_gv11b.c
@@ -761,6 +761,8 @@ static const struct gpu_ops gv11b_ops = {
 		.set_handled_snapshots = gv11b_css_hw_set_handled_snapshots,
 		.allocate_perfmon_ids = css_gr_allocate_perfmon_ids,
 		.release_perfmon_ids = css_gr_release_perfmon_ids,
+		.get_overflow_status = gv11b_css_hw_get_overflow_status,
+		.get_pending_snapshots = gv11b_css_hw_get_pending_snapshots,
 	},
 #endif
 	.falcon = {