*     - end     - virtual end address of region
  */
 v6_dma_inv_range:
+#ifdef CONFIG_DMA_CACHE_RWFO
+       ldrb    r2, [r0]                        @ read for ownership
+       strb    r2, [r0]                        @ write for ownership
+#endif
        tst     r0, #D_CACHE_LINE_SIZE - 1
        bic     r0, r0, #D_CACHE_LINE_SIZE - 1
 #ifdef HARVARD_CACHE
        mcrne   p15, 0, r0, c7, c11, 1          @ clean unified line
 #endif
        tst     r1, #D_CACHE_LINE_SIZE - 1
+#ifdef CONFIG_DMA_CACHE_RWFO
+       ldrneb  r2, [r1, #-1]                   @ read for ownership
+       strneb  r2, [r1, #-1]                   @ write for ownership
+#endif
        bic     r1, r1, #D_CACHE_LINE_SIZE - 1
 #ifdef HARVARD_CACHE
        mcrne   p15, 0, r1, c7, c14, 1          @ clean & invalidate D line
        mcrne   p15, 0, r1, c7, c15, 1          @ clean & invalidate unified line
 #endif
 1:
-#ifdef CONFIG_DMA_CACHE_RWFO
-       ldr     r2, [r0]                        @ read for ownership
-       str     r2, [r0]                        @ write for ownership
-#endif
 #ifdef HARVARD_CACHE
        mcr     p15, 0, r0, c7, c6, 1           @ invalidate D line
 #else
 #endif
        add     r0, r0, #D_CACHE_LINE_SIZE
        cmp     r0, r1
+#ifdef CONFIG_DMA_CACHE_RWFO
+       ldrlo   r2, [r0]                        @ read for ownership
+       strlo   r2, [r0]                        @ write for ownership
+#endif
        blo     1b
        mov     r0, #0
        mcr     p15, 0, r0, c7, c10, 4          @ drain write buffer
  *     - end     - virtual end address of region
  */
 ENTRY(v6_dma_flush_range)
-       bic     r0, r0, #D_CACHE_LINE_SIZE - 1
-1:
 #ifdef CONFIG_DMA_CACHE_RWFO
-       ldr     r2, [r0]                        @ read for ownership
-       str     r2, [r0]                        @ write for ownership
+       ldrb    r2, [r0]                @ read for ownership
+       strb    r2, [r0]                @ write for ownership
 #endif
+       bic     r0, r0, #D_CACHE_LINE_SIZE - 1
+1:
 #ifdef HARVARD_CACHE
        mcr     p15, 0, r0, c7, c14, 1          @ clean & invalidate D line
 #else
 #endif
        add     r0, r0, #D_CACHE_LINE_SIZE
        cmp     r0, r1
+#ifdef CONFIG_DMA_CACHE_RWFO
+       ldrlob  r2, [r0]                        @ read for ownership
+       strlob  r2, [r0]                        @ write for ownership
+#endif
        blo     1b
        mov     r0, #0
        mcr     p15, 0, r0, c7, c10, 4          @ drain write buffer