Index: head/sys/arm/arm/busdma_machdep-v6.c
===================================================================
--- head/sys/arm/arm/busdma_machdep-v6.c
+++ head/sys/arm/arm/busdma_machdep-v6.c
@@ -491,6 +491,7 @@
 		newtag->highaddr = MAX(parent->highaddr, newtag->highaddr);
 		newtag->alignment = MAX(parent->alignment, newtag->alignment);
 		newtag->flags |= parent->flags & BUS_DMA_COULD_BOUNCE;
+		newtag->flags |= parent->flags & BUS_DMA_COHERENT;
 		if (newtag->boundary == 0)
 			newtag->boundary = parent->boundary;
 		else if (parent->boundary != 0)
@@ -755,11 +756,19 @@
 	}
 	map->flags = DMAMAP_DMAMEM_ALLOC;
 
-	/* Choose a busdma buffer allocator based on memory type flags. */
-	if (flags & BUS_DMA_COHERENT) {
+	/* For coherent memory, set the map flag that disables sync ops. */
+	if (flags & BUS_DMA_COHERENT)
+		map->flags |= DMAMAP_COHERENT;
+
+	/*
+	 * Choose a busdma buffer allocator based on memory type flags.
+	 * If the tag's COHERENT flag is set, that means normal memory
+	 * is already coherent, use the normal allocator.
+	 */
+	if ((flags & BUS_DMA_COHERENT) &&
+	    ((dmat->flags & BUS_DMA_COHERENT) == 0)) {
 		memattr = VM_MEMATTR_UNCACHEABLE;
 		ba = coherent_allocator;
-		map->flags |= DMAMAP_COHERENT;
 	} else {
 		memattr = VM_MEMATTR_DEFAULT;
 		ba = standard_allocator;
@@ -829,7 +838,8 @@
 	struct busdma_bufzone *bufzone;
 	busdma_bufalloc_t ba;
 
-	if (map->flags & DMAMAP_COHERENT)
+	if ((map->flags & DMAMAP_COHERENT) &&
+	    ((dmat->flags & BUS_DMA_COHERENT) == 0))
 		ba = coherent_allocator;
 	else
 		ba = standard_allocator;
@@ -1030,7 +1040,7 @@
 			sgsize = MIN(sgsize, PAGE_SIZE - (curaddr & PAGE_MASK));
 			curaddr = add_bounce_page(dmat, map, 0, curaddr,
 			    sgsize);
-		} else {
+		} else if ((dmat->flags & BUS_DMA_COHERENT) == 0) {
 			if (map->sync_count > 0)
 				sl_end = sl->paddr + sl->datacount;
 
@@ -1144,7 +1154,7 @@
 		    sgsize)) {
 			curaddr = add_bounce_page(dmat, map, kvaddr, curaddr,
 			    sgsize);
-		} else {
+		} else if ((dmat->flags & BUS_DMA_COHERENT) == 0) {
 			if (map->sync_count > 0) {
 				sl_pend = sl->paddr + sl->datacount;
 				sl_vend = sl->vaddr + sl->datacount;
@@ -1353,8 +1363,9 @@
 				    bpage->datacount);
 				if (tempvaddr != 0)
 					pmap_quick_remove_page(tempvaddr);
-				dcache_wb_poc(bpage->vaddr, bpage->busaddr,
-				    bpage->datacount);
+				if ((dmat->flags & BUS_DMA_COHERENT) == 0)
+					dcache_wb_poc(bpage->vaddr,
+					    bpage->busaddr, bpage->datacount);
 				bpage = STAILQ_NEXT(bpage, links);
 			}
 			dmat->bounce_zone->total_bounced++;
@@ -1374,8 +1385,9 @@
 		if ((op & BUS_DMASYNC_PREREAD) && !(op & BUS_DMASYNC_PREWRITE)) {
 			bpage = STAILQ_FIRST(&map->bpages);
 			while (bpage != NULL) {
-				dcache_inv_poc_dma(bpage->vaddr, bpage->busaddr,
-				    bpage->datacount);
+				if ((dmat->flags & BUS_DMA_COHERENT) == 0)
+					dcache_inv_poc_dma(bpage->vaddr,
+					    bpage->busaddr, bpage->datacount);
 				bpage = STAILQ_NEXT(bpage, links);
 			}
 		}
@@ -1391,8 +1403,9 @@
 		 */
 		if (op & BUS_DMASYNC_POSTREAD) {
 			while (bpage != NULL) {
-				dcache_inv_poc(bpage->vaddr, bpage->busaddr,
-				    bpage->datacount);
+				if ((dmat->flags & BUS_DMA_COHERENT) == 0)
+					dcache_inv_poc(bpage->vaddr,
+					    bpage->busaddr, bpage->datacount);
 				tempvaddr = 0;
 				datavaddr = bpage->datavaddr;
 				if (datavaddr == 0) {
@@ -1421,7 +1434,8 @@
 	if (map->flags & DMAMAP_COHERENT) {
 		if (op & BUS_DMASYNC_PREWRITE) {
 			dsb();
-			cpu_l2cache_drain_writebuf();
+			if ((dmat->flags & BUS_DMA_COHERENT) == 0)
+				cpu_l2cache_drain_writebuf();
 		}
 		return;
 	}