diff --git a/Documentation/devicetree/bindings/display/apple,dcp.yaml b/Documentation/devicetree/bindings/display/apple,dcp.yaml
index 38be82bc7d1bd..dbb2234164d7e 100644
--- a/Documentation/devicetree/bindings/display/apple,dcp.yaml
+++ b/Documentation/devicetree/bindings/display/apple,dcp.yaml
@@ -72,6 +72,8 @@ properties:
   power-domains:
     maxItems: 1
 
+  dma-noncoherent: true
+
   resets:
     maxItems: 1
 
@@ -299,6 +301,7 @@ properties:
     properties:
       iommus:
         maxItems: 1
+      dma-noncoherent: true
       memory-region:
         minItems: 1
         maxItems: 64
diff --git a/Documentation/devicetree/bindings/gpu/apple,agx.yaml b/Documentation/devicetree/bindings/gpu/apple,agx.yaml
index f5b1afd7d7dd3..2004c43d42b1d 100644
--- a/Documentation/devicetree/bindings/gpu/apple,agx.yaml
+++ b/Documentation/devicetree/bindings/gpu/apple,agx.yaml
@@ -40,6 +40,8 @@ properties:
   power-domains:
     maxItems: 1
 
+  dma-noncoherent: true
+
   mboxes:
     minItems: 1
     maxItems: 2
diff --git a/arch/arm64/boot/dts/apple/t6040.dtsi b/arch/arm64/boot/dts/apple/t6040.dtsi
index 961af1b7b5bd4..b909d0c197f77 100644
--- a/arch/arm64/boot/dts/apple/t6040.dtsi
+++ b/arch/arm64/boot/dts/apple/t6040.dtsi
@@ -522,7 +522,6 @@
 	display: display-subsystem {
 		compatible = "apple,display-subsystem";
 		iommus = <&disp0_dart 0>;
-		dma-noncoherent;
 		status = "disabled";
 	};
 
diff --git a/arch/arm64/boot/dts/apple/t8132-j713-display.dtsi b/arch/arm64/boot/dts/apple/t8132-j713-display.dtsi
index d35e9fc89a2a3..e3cb6864c54d6 100644
--- a/arch/arm64/boot/dts/apple/t8132-j713-display.dtsi
+++ b/arch/arm64/boot/dts/apple/t8132-j713-display.dtsi
@@ -16,6 +16,8 @@
  * frontend mappings it adds memory-region and apple,j713-inherited-mappings
  * to piodma and display-subsystem and enables them and disp0_dart.
  * apple,j713-native-piodma and apple,j713-native-scanout opt in to that.
+ * The DCP, PIODMA and scanout paths use non-coherent DMA only, so those
+ * nodes are marked dma-noncoherent.
  * The DART driver also reads the ADT through the loader's phram "adt"
  * region. Without that handoff the DCP and frontend are refused.
  *
@@ -81,6 +83,7 @@
 		mbox-names = "mbox";
 		iommus = <&dcp_dart 23>;
 		power-domains = <&ps_disp_cpu>;
+		dma-noncoherent;
 		apple,j713-native-piodma;
 		apple,j713-native-scanout;
 		/* Full 2560x1664 panel including the notch rows, 224 ppi. */
@@ -91,6 +94,7 @@
 
 		disp0_piodma: piodma {
 			iommus = <&disp0_dart 16>;
+			dma-noncoherent;
 			status = "disabled"; /* Enabled by the loader. */
 		};
 	};
@@ -98,6 +102,7 @@
 	display: display-subsystem {
 		compatible = "apple,t8132-display-subsystem", "apple,t8132-display-diagnostics";
 		iommus = <&disp0_dart 0>;
+		dma-noncoherent;
 		status = "disabled"; /* Enabled by the loader. */
 	};
 };
diff --git a/arch/arm64/boot/dts/apple/t8132-j713-gpu.dtsi b/arch/arm64/boot/dts/apple/t8132-j713-gpu.dtsi
index 7cddbaf1e5b56..7db8d5a4c16b7 100644
--- a/arch/arm64/boot/dts/apple/t8132-j713-gpu.dtsi
+++ b/arch/arm64/boot/dts/apple/t8132-j713-gpu.dtsi
@@ -69,6 +69,8 @@
 		reg-names = "asc", "sgx";
 		mboxes = <&agx_mbox>;
 		power-domains = <&ps_gfx>;
+		/* G16 is admitted on non-coherent DMA only. */
+		dma-noncoherent;
 		apple,pmp = <&pmp>;
 		memory-region = <&gpu_ttbs>, <&gpu_pagetables>, <&gpu_handoff>,
 				<&gpu_shared_l2>, <&gpu_fw_text>, <&gpu_fw_data>;
diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
index e517dddb37499..a835ed06cba0d 100644
--- a/arch/arm64/mm/dma-mapping.c
+++ b/arch/arm64/mm/dma-mapping.c
@@ -7,6 +7,7 @@
 #include <linux/gfp.h>
 #include <linux/cache.h>
 #include <linux/dma-map-ops.h>
+#include <linux/property.h>
 #include <xen/xen.h>
 
 #include <asm/cacheflush.h>
@@ -42,12 +43,17 @@ void arch_setup_dma_ops(struct device *dev, bool coherent)
 {
 	int cls = cache_line_size_of_cpu();
 
-	if (!coherent && !CLIDR_LOC(read_sysreg(clidr_el1))) {
+	/*
+	 * A device explicitly marked "dma-noncoherent" keeps non-coherent DMA:
+	 * its driver may depend on the non-cacheable attributes of that path.
+	 */
+	if (!coherent && !CLIDR_LOC(read_sysreg(clidr_el1)) &&
+	    !device_property_present(dev, "dma-noncoherent")) {
 		/*
-		 * This holds for every device on the system, so say it once
+		 * This holds for every other device on the system, so say it once
 		 * and keep the per-device line for debugging.
 		 */
-		pr_info_once("CLIDR_EL1.LoC == 0, treating all devices as DMA coherent\n");
+		pr_info_once("CLIDR_EL1.LoC == 0, treating devices not marked dma-noncoherent as DMA coherent\n");
 		dev_dbg(dev, "CLIDR_EL1.LoC == 0, treating as coherent\n");
 		coherent = true;
 	}