The change to allow allocating ICM chunks from coherent memory did not
increment the count of sg entries properly, so a chunk that required
more than allocation would not be mapped properly by the HCA.
Fix this by adding the missing increment of chunk->nsg.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
                if (!ret) {
                        ++chunk->npages;
 
-                       if (!coherent && chunk->npages == MTHCA_ICM_CHUNK_LEN) {
+                       if (coherent)
+                               ++chunk->nsg;
+                       else if (chunk->npages == MTHCA_ICM_CHUNK_LEN) {
                                chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
                                                        chunk->npages,
                                                        PCI_DMA_BIDIRECTIONAL);