diff --git a/arch/ia64/hp/common/sba_iommu.c b/arch/ia64/hp/common/sba_iommu.c
index 45bf04e..c412fe6 100644
--- a/arch/ia64/hp/common/sba_iommu.c
+++ b/arch/ia64/hp/common/sba_iommu.c
@@ -1265,7 +1265,7 @@
  * the sglist do both.
  */
 static SBA_INLINE int
-sba_coalesce_chunks( struct ioc *ioc,
+sba_coalesce_chunks(struct ioc *ioc, struct device *dev,
 	struct scatterlist *startsg,
 	int nents)
 {
@@ -1275,6 +1275,7 @@
 	struct scatterlist *dma_sg;        /* next DMA stream head */
 	unsigned long dma_offset, dma_len; /* start/len of DMA stream */
 	int n_mappings = 0;
+	unsigned int max_seg_size = dma_get_max_seg_size(dev);
 
 	while (nents > 0) {
 		unsigned long vaddr = (unsigned long) sba_sg_address(startsg);
@@ -1314,6 +1315,9 @@
 			    > DMA_CHUNK_SIZE)
 				break;
 
+			if (dma_len + startsg->length > max_seg_size)
+				break;
+
 			/*
 			** Then look for virtually contiguous blocks.
 			**
@@ -1441,7 +1445,7 @@
 	** w/o this association, we wouldn't have coherent DMA!
 	** Access to the virtual address is what forces a two pass algorithm.
 	*/
-	coalesced = sba_coalesce_chunks(ioc, sglist, nents);
+	coalesced = sba_coalesce_chunks(ioc, dev, sglist, nents);
 
 	/*
 	** Program the I/O Pdir