Re: Tx queueing

[Jeff Garzik <jgarzik@xxxxxxxxxxxxxxxx>]

jamal wrote:
> Jeff, when you say some modern PCI hardware has
> problems with the described semantics: can you provide more details?

I was referring to PCI drivers, not PCI hardware.  What I meant was that
my experience has shown that some of the early softnet conversions
(example 'A') caused transmit timeouts quite easily, until updated to
look like example B.

Example A:

        drv_start_xmit() {
                netif_stop_queue()
                /* queue packet for xmit */
                if (!tx_full)
                        netif_start_queue()
        }
        interrupt() {
                /* Tx'd a packet */
                if (tx_full)
                        netif_stop_queue()
                else
                        netif_wake_queue()
        }

Example B:

        drv_start_xmit() {
                /* queue packet for xmit */
                if (tx_full)
                        netif_stop_queue()
        }
        interrupt() {
                /* Tx'd a packet */
                if (!tx_full)
                        netif_wake_queue()
        }

As a further note, since many of the PCI drivers do multiple iterations
of "work," I wonder if it would be useful to postpone the
netif_wake_queue() call until after the work loop in the interrupt
handler completes.

        Jeff



-- 
Jeff Garzik              | Liberty is always dangerous, but
Building 1024            | it is the safest thing we have.
MandrakeSoft, Inc.       |      -- Harry Emerson Fosdick