What I am trying to understand is why you are spinning in the handler. 
_With_ DMA you should hardly get into the serial handler under normal 
conditions. Running at 3.5MB/sec should give one byte every 2.8us and
48 Bytes every ~137us. This looks like plenty of time to get  out of
the handler. My *guess* is that serial8250_handle_irq() has IIR often
set to timeout and you end up fetching byte after byte. 

This patch should protocol when and why you got into the handler.

diff --git a/drivers/tty/serial/8250/8250_core.c b/drivers/tty/serial/8250/8250_core.c
index 7111b22de000..59852069e4a0 100644
--- a/drivers/tty/serial/8250/8250_core.c
+++ b/drivers/tty/serial/8250/8250_core.c

@@ -1583,6 +1583,7 @@ int serial8250_handle_irq(struct uart_port *port, unsigned int iir)
 	status = serial_port_in(port, UART_LSR);
 
 	DEBUG_INTR("status = %x...", status);
+	trace_printk("l%d IIR %x LSR %x\n", port->line, iir, status);
 
 	if (status & (UART_LSR_DR | UART_LSR_BI)) {
 		if (up->dma)

@@ -1707,6 +1708,7 @@ static irqreturn_t serial8250_interrupt(int irq, void *dev_id)
 
 	spin_unlock(&i->lock);
 
+	trace_printk("%d e\n", irq);
 	DEBUG_INTR("end.\n");
 
 	return IRQ_RETVAL(handled);