[PATCH] add polling support for ns16550

[PATCH] add polling support for ns16550

[Alex Williamson]

   The patch below adds ac_timer based polling to the ns16550 UART
driver.  This is useful when the interrupt line is not connected in
hardware or the mechanism to enable it is not readily available in the
hypervisor.  Polling is only enabled when the UART IRQ is set to zero.
Thanks,

	Alex

Signed-off-by: Alex Williamson <alex.williamson@hp.com>
---

diff -r 378e1c58bcd2 xen/drivers/char/ns16550.c
--- a/xen/drivers/char/ns16550.c	Fri Nov 18 16:54:23 2005
+++ b/xen/drivers/char/ns16550.c	Wed Nov 23 09:09:58 2005
@@ -30,6 +30,8 @@
     unsigned long io_base;   /* I/O port or memory-mapped I/O address. */
     char *remapped_io_base;  /* Remapped virtual address of mmap I/O.  */ 
     struct irqaction irqaction;
+    struct ac_timer timer;
+    unsigned int timeout;
 } ns16550_com[2] = { { 0 } };
 
 /* Register offsets */
@@ -121,6 +123,34 @@
     }
 }
 
+static void ns16550_timeout(void *data)
+{
+    struct serial_port *port = (struct serial_port *)data;
+    struct ns16550 *uart = port->uart;
+    struct cpu_user_regs *regs = guest_cpu_user_regs();
+    char lsr, ier;
+    int rx_cnt = port->tx_fifo_size;
+
+    /* interrupts must be disabled, this is likely to generate some */
+    ier = ns_read_reg(uart, IER);
+    if ( ier )
+        ns_write_reg(uart, IER, 0);
+
+    lsr = ns_read_reg(uart, LSR);
+
+    while ( (lsr & LSR_DR) && rx_cnt-- ) {
+        serial_rx_interrupt(port, regs);
+        lsr = ns_read_reg(uart, LSR);
+    }
+    if ( lsr & LSR_THRE )
+        serial_tx_interrupt(port, regs);
+
+    if ( ier )
+        ns_write_reg(uart, IER, ier);
+
+    set_ac_timer(&uart->timer, NOW() + MILLISECS(uart->timeout));
+}
+
 static int ns16550_tx_empty(struct serial_port *port)
 {
     struct ns16550 *uart = port->uart;
@@ -178,15 +208,38 @@
         port->tx_fifo_size = 16;
 }
 
+static void ns16550_set_timeout(struct serial_port *port)
+{
+    struct ns16550 *uart = port->uart;
+    unsigned int bits;
+
+    bits = uart->data_bits + uart->stop_bits;
+
+    if ( uart->parity != PARITY_NONE )
+        bits++;
+
+    bits = bits * port->tx_fifo_size;
+
+    /* uart timeout in ms */
+    uart->timeout = (1000 * bits) / uart->baud;
+}
+
 static void ns16550_init_postirq(struct serial_port *port)
 {
     struct ns16550 *uart = port->uart;
     int rc;
 
-    if ( uart->irq <= 0 )
+    if ( uart->irq < 0 )
         return;
 
     serial_async_transmit(port);
+
+    if ( uart->irq == 0 ) {
+        ns16550_set_timeout(port);
+        init_ac_timer(&uart->timer, ns16550_timeout, port, smp_processor_id());
+        set_ac_timer(&uart->timer, NOW() + MILLISECS(uart->timeout));
+        return;
+    }
 
     uart->irqaction.handler = ns16550_interrupt;
     uart->irqaction.name    = "ns16550";

Re: [PATCH] add polling support for ns16550

[Keir Fraser]

On 23 Nov 2005, at 16:40, Alex Williamson wrote:

>   The patch below adds ac_timer based polling to the ns16550 UART
> driver.  This is useful when the interrupt line is not connected in
> hardware or the mechanism to enable it is not readily available in the
> hypervisor.  Polling is only enabled when the UART IRQ is set to zero.
> Thanks,

I reworked the patch a little and checked it in, thanks.

One thing I removed was the clear-and-reset of IER in the timeout 
handler. Given we write zero to IER in the preirq steup function, and 
also we do not set the master-enable bit in the MCR, I cannot see how 
you could end up accidentally fielding interrupts in the timeout 
handler. Did I miss something?

  -- Keir

Re: [PATCH] add polling support for ns16550

[Alex Williamson]

On Thu, 2005-11-24 at 11:13 +0000, Keir Fraser wrote:
> On 23 Nov 2005, at 16:40, Alex Williamson wrote:
> 
> >   The patch below adds ac_timer based polling to the ns16550 UART
> > driver.  This is useful when the interrupt line is not connected in
> > hardware or the mechanism to enable it is not readily available in the
> > hypervisor.  Polling is only enabled when the UART IRQ is set to zero.
> > Thanks,
> 
> I reworked the patch a little and checked it in, thanks.

   Thanks!

> One thing I removed was the clear-and-reset of IER in the timeout 
> handler. Given we write zero to IER in the preirq steup function, and 
> also we do not set the master-enable bit in the MCR, I cannot see how 
> you could end up accidentally fielding interrupts in the timeout 
> handler. Did I miss something?

   It was mainly paranoia, but on ia64 we don't have a good way to hide
UARTs from the guest domains, so it's possible that a dom0 with 8250
support will enable those bits on the UART.  All sorts of strange things
happen when the hypervisor and a guest compete for access to a UART, but
I thought I would at least avoid the hypervisor polling triggering the
real interrupt.  I'm ok with dropping it, the UART isn't really
functional no matter what paranoia checks are added when multiple
drivers are competing to use it.  Thanks,

	Alex