diff --git a/drivers/gpu/nvgpu/common/mm/gmmu.c b/drivers/gpu/nvgpu/common/mm/gmmu.c
index 602dfb3b9..55fbcd3fc 100644
--- a/drivers/gpu/nvgpu/common/mm/gmmu.c
+++ b/drivers/gpu/nvgpu/common/mm/gmmu.c
@@ -747,6 +747,7 @@ u64 gk20a_locked_gmmu_map(struct vm_gk20a *vm,
 		.rw_flag   = rw_flag,
 		.sparse    = sparse,
 		.priv      = priv,
+		.coherent  = flags & NVGPU_AS_MAP_BUFFER_FLAGS_IO_COHERENT,
 		.valid     = !(flags & NVGPU_AS_MAP_BUFFER_FLAGS_UNMAPPED_PTE),
 		.aperture  = aperture
 	};
@@ -805,6 +806,7 @@ void gk20a_locked_gmmu_unmap(struct vm_gk20a *vm,
 		.rw_flag   = rw_flag,
 		.sparse    = sparse,
 		.priv      = 0,
+		.coherent  = 0,
 		.valid     = 0,
 		.aperture  = APERTURE_INVALID,
 	};
diff --git a/drivers/gpu/nvgpu/gp10b/mm_gp10b.c b/drivers/gpu/nvgpu/gp10b/mm_gp10b.c
index 590dd9605..bdc301433 100644
--- a/drivers/gpu/nvgpu/gp10b/mm_gp10b.c
+++ b/drivers/gpu/nvgpu/gp10b/mm_gp10b.c
@@ -230,9 +230,12 @@ static void __update_pte(struct vm_gk20a *vm,
 	u32 pte_addr = attrs->aperture == APERTURE_SYSMEM ?
 		gmmu_new_pte_address_sys_f(phys_shifted) :
 		gmmu_new_pte_address_vid_f(phys_shifted);
-	u32 pte_tgt = __nvgpu_aperture_mask(g, attrs->aperture,
-		gmmu_new_pte_aperture_sys_mem_ncoh_f(),
-		gmmu_new_pte_aperture_video_memory_f());
+	u32 pte_tgt = __nvgpu_aperture_mask(g,
+			attrs->aperture,
+			attrs->coherent ?
+				gmmu_new_pte_aperture_sys_mem_coh_f() :
+				gmmu_new_pte_aperture_sys_mem_ncoh_f(),
+			gmmu_new_pte_aperture_video_memory_f());
 
 	pte_w[0] = pte_valid | pte_addr | pte_tgt;
 
diff --git a/drivers/gpu/nvgpu/include/nvgpu/gmmu.h b/drivers/gpu/nvgpu/include/nvgpu/gmmu.h
index eff87c314..47af2f75f 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/gmmu.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/gmmu.h
@@ -149,6 +149,7 @@ struct nvgpu_gmmu_pd {
  *   rw_flag:     Flag from enum gk20a_mem_rw_flag
  *   sparse:      Set if the mapping should be sparse.
  *   priv:        Privilidged mapping.
+ *   coherent:    Set if the mapping should be IO coherent.
  *   valid:       Set if the PTE should be marked valid.
  *   aperture:    VIDMEM or SYSMEM.
  *   debug:       When set print debugging info.
@@ -166,6 +167,7 @@ struct nvgpu_gmmu_attrs {
 	int			 rw_flag;
 	bool			 sparse;
 	bool			 priv;
+	bool			 coherent;
 	bool			 valid;
 	enum nvgpu_aperture	 aperture;
 	bool			 debug;
diff --git a/include/uapi/linux/nvgpu-t18x.h b/include/uapi/linux/nvgpu-t18x.h
index 16cbd81ec..59db77c72 100644
--- a/include/uapi/linux/nvgpu-t18x.h
+++ b/include/uapi/linux/nvgpu-t18x.h
@@ -24,12 +24,6 @@
 #ifndef _UAPI__LINUX_NVGPU_T18X_IOCTL_H_
 #define _UAPI__LINUX_NVGPU_T18X_IOCTL_H_
 
-/*
- * this flag is used in struct nvgpu_as_map_buffer_ex_args
- * to specify IO coherence
- */
-#define NVGPU_AS_MAP_BUFFER_FLAGS_IO_COHERENT	    (1 << 4)
-
 /*
  * this flag is used in struct nvgpu_alloc_gpfifo_args
  * to enable re-playable faults for that channel
diff --git a/include/uapi/linux/nvgpu.h b/include/uapi/linux/nvgpu.h
index 22687eb5d..6c1137642 100644
--- a/include/uapi/linux/nvgpu.h
+++ b/include/uapi/linux/nvgpu.h
@@ -1744,6 +1744,7 @@ struct nvgpu_as_map_buffer_args {
 	__u32 flags;		/* in/out */
 #define NVGPU_AS_MAP_BUFFER_FLAGS_FIXED_OFFSET	    (1 << 0)
 #define NVGPU_AS_MAP_BUFFER_FLAGS_CACHEABLE	    (1 << 2)
+#define NVGPU_AS_MAP_BUFFER_FLAGS_IO_COHERENT	    (1 << 4)
 #define NVGPU_AS_MAP_BUFFER_FLAGS_UNMAPPED_PTE	    (1 << 5)
 #define NVGPU_AS_MAP_BUFFER_FLAGS_MAPPABLE_COMPBITS (1 << 6)
 	__u32 reserved;		/* in */