The existing logic incorrectly maps this buffer list:
    0: addr 0x10001000, size 0x1000
    1: addr 0x10002000, size 0x1000
To this bogus page list:
    0: 0x10000000
    1: 0x10002000
The shift calculation must also take into account the address of the
first entry masked by the page_mask as well as the last address+size
rounded up to the next page size.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
                *total_size += buffer_list[i].size;
                if (i > 0)
                        mask |= buffer_list[i].addr;
+               else
+                       mask |= buffer_list[i].addr & PAGE_MASK;
+               if (i != num_phys_buf - 1)
+                       mask |= buffer_list[i].addr + buffer_list[i].size;
+               else
+                       mask |= (buffer_list[i].addr + buffer_list[i].size +
+                               PAGE_SIZE - 1) & PAGE_MASK;
        }
 
        if (*total_size > 0xFFFFFFFFULL)