diff --git a/arch/tile/Kconfig b/arch/tile/Kconfig
index a5302d319229754292bc87a8907b6ac25be0b097..0ad771f7a7e1144e3e69510d21e436c6d6dea140 100644
--- a/arch/tile/Kconfig
+++ b/arch/tile/Kconfig
@@ -212,6 +212,9 @@ config HIGHMEM
 
 	  If unsure, say "true".
 
+config ZONE_DMA
+	def_bool y
+
 # We do not currently support disabling NUMA.
 config NUMA
 	bool # "NUMA Memory Allocation and Scheduler Support"
diff --git a/arch/tile/kernel/pci-dma.c b/arch/tile/kernel/pci-dma.c
index 9814d7082f244a0df13b47d74af91806887fcd31..edd856a000c5ce6c837f54c9c41dd3fff248924e 100644
--- a/arch/tile/kernel/pci-dma.c
+++ b/arch/tile/kernel/pci-dma.c
@@ -45,14 +45,17 @@ void *dma_alloc_coherent(struct device *dev,
 	gfp |= __GFP_ZERO;
 
 	/*
-	 * By forcing NUMA node 0 for 32-bit masks we ensure that the
-	 * high 32 bits of the resulting PA will be zero.  If the mask
-	 * size is, e.g., 24, we may still not be able to guarantee a
-	 * suitable memory address, in which case we will return NULL.
-	 * But such devices are uncommon.
+	 * If the mask specifies that the memory be in the first 4 GB, then
+	 * we force the allocation to come from the DMA zone.  We also
+	 * force the node to 0 since that's the only node where the DMA
+	 * zone isn't empty.  If the mask size is smaller than 32 bits, we
+	 * may still not be able to guarantee a suitable memory address, in
+	 * which case we will return NULL.  But such devices are uncommon.
 	 */
-	if (dma_mask <= DMA_BIT_MASK(32))
+	if (dma_mask <= DMA_BIT_MASK(32)) {
+		gfp |= GFP_DMA;
 		node = 0;
+	}
 
 	pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
 	if (pg == NULL)
diff --git a/arch/tile/kernel/setup.c b/arch/tile/kernel/setup.c
index 6d179dfcc15e62e65e41330f2b8cc702988ee5c9..fdde3b6986e54f59b117fd61c25566dde5df0dc0 100644
--- a/arch/tile/kernel/setup.c
+++ b/arch/tile/kernel/setup.c
@@ -658,6 +658,8 @@ static void __init zone_sizes_init(void)
 	unsigned long zones_size[MAX_NR_ZONES] = { 0 };
 	int size = percpu_size();
 	int num_cpus = smp_height * smp_width;
+	const unsigned long dma_end = (1UL << (32 - PAGE_SHIFT));
+
 	int i;
 
 	for (i = 0; i < num_cpus; ++i)
@@ -729,6 +731,14 @@ static void __init zone_sizes_init(void)
 		zones_size[ZONE_NORMAL] = end - start;
 #endif
 
+		if (start < dma_end) {
+			zones_size[ZONE_DMA] = min(zones_size[ZONE_NORMAL],
+						   dma_end - start);
+			zones_size[ZONE_NORMAL] -= zones_size[ZONE_DMA];
+		} else {
+			zones_size[ZONE_DMA] = 0;
+		}
+
 		/* Take zone metadata from controller 0 if we're isolnode. */
 		if (node_isset(i, isolnodes))
 			NODE_DATA(i)->bdata = &bootmem_node_data[0];
@@ -738,7 +748,7 @@ static void __init zone_sizes_init(void)
 		       PFN_UP(node_percpu[i]));
 
 		/* Track the type of memory on each node */
-		if (zones_size[ZONE_NORMAL])
+		if (zones_size[ZONE_NORMAL] || zones_size[ZONE_DMA])
 			node_set_state(i, N_NORMAL_MEMORY);
 #ifdef CONFIG_HIGHMEM
 		if (end != start)
diff --git a/arch/tile/mm/init.c b/arch/tile/mm/init.c
index a2417a0a82226e0daaf0ceb220082b38e0ae2d30..ef29d6c5e10e9b76216b61e4cf7d2722c824d8e9 100644
--- a/arch/tile/mm/init.c
+++ b/arch/tile/mm/init.c
@@ -733,16 +733,15 @@ static void __init set_non_bootmem_pages_init(void)
 	for_each_zone(z) {
 		unsigned long start, end;
 		int nid = z->zone_pgdat->node_id;
+#ifdef CONFIG_HIGHMEM
 		int idx = zone_idx(z);
+#endif
 
 		start = z->zone_start_pfn;
-		if (start == 0)
-			continue;  /* bootmem */
 		end = start + z->spanned_pages;
-		if (idx == ZONE_NORMAL) {
-			BUG_ON(start != node_start_pfn[nid]);
-			start = node_free_pfn[nid];
-		}
+		start = max(start, node_free_pfn[nid]);
+		start = max(start, max_low_pfn);
+
 #ifdef CONFIG_HIGHMEM
 		if (idx == ZONE_HIGHMEM)
 			totalhigh_pages += z->spanned_pages;