diff --git a/drivers/gpu/nvgpu/gk20a/hw_fb_gk20a.h b/drivers/gpu/nvgpu/gk20a/hw_fb_gk20a.h
index 46846b213..b9e124b75 100644
--- a/drivers/gpu/nvgpu/gk20a/hw_fb_gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/hw_fb_gk20a.h
@@ -250,4 +250,8 @@ static inline u32 fb_mmu_vpr_info_fetch_true_v(void)
 {
 	return 0x00000001;
 }
+static inline u32 fb_niso_flush_sysmem_addr_r(void)
+{
+	return 0x00100c10;
+}
 #endif
diff --git a/drivers/gpu/nvgpu/gk20a/mm_gk20a.c b/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
index b5ec5e25f..452d3e9c5 100644
--- a/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
@@ -424,6 +424,11 @@ static void gk20a_remove_mm_support(struct mm_gk20a *mm)
 	gk20a_vm_remove_support_nofree(&mm->cde.vm);
 }
 
+static int gk20a_alloc_sysmem_flush(struct gk20a *g)
+{
+	return gk20a_gmmu_alloc(g, SZ_4K, &g->mm.sysmem_flush);
+}
+
 int gk20a_init_mm_setup_sw(struct gk20a *g)
 {
 	struct mm_gk20a *mm = &g->mm;
@@ -447,6 +452,10 @@ int gk20a_init_mm_setup_sw(struct gk20a *g)
 		       (int)(mm->channel.user_size >> 20),
 		       (int)(mm->channel.kernel_size >> 20));
 
+	err = gk20a_alloc_sysmem_flush(g);
+	if (err)
+		return err;
+
 	err = gk20a_init_bar1_vm(mm);
 	if (err)
 		return err;
@@ -492,6 +501,9 @@ int gk20a_init_mm_setup_hw(struct gk20a *g)
 		mm->use_full_comp_tag_line =
 			g->ops.fb.set_use_full_comp_tag_line(g);
 
+	gk20a_writel(g, fb_niso_flush_sysmem_addr_r(),
+		     g->ops.mm.get_iova_addr(g, g->mm.sysmem_flush.sgt->sgl, 0)
+		     >> 8);
 
 	g->ops.mm.bar1_bind(g, inst_pa);
 
diff --git a/drivers/gpu/nvgpu/gk20a/mm_gk20a.h b/drivers/gpu/nvgpu/gk20a/mm_gk20a.h
index 2d137bbe1..2c2df1a9b 100644
--- a/drivers/gpu/nvgpu/gk20a/mm_gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/mm_gk20a.h
@@ -359,6 +359,8 @@ struct mm_gk20a {
 	bool has_physical_mode;
 	/* false if vidmem aperture actually points to sysmem */
 	bool vidmem_is_vidmem;
+
+	struct mem_desc sysmem_flush;
 };
 
 int gk20a_mm_init(struct mm_gk20a *mm);