Re: [PATCH] e1000 - get rid of tx_lock

[Francois Romieu <romieu@xxxxxxxxxxxxx>]

Stephen Hemminger <shemminger@xxxxxxxx> :
> The e1000 driver has an tx_lock which unneeded.  It is only used to protect
> the start/stop queue flags which are already handled by doing atomic bit
> operations.

I am not terribly used to this code but as far as I can read it, the lock
avoids that the condition which leads to netif_stop_queue() in the xmit
thread changes (due to an irq) "just before" netif_stop_queue() is actually
called. 

So, if the lock is removed, I would be tempted to have the xmit thread
issue an smp_rmb() and revalidate that it was fine to netif_stop_queue().

Otherwise, one could have:

[e1000_xmit_frame]
        if (E1000_DESC_UNUSED(&adapter->tx_ring) < count + 2 ) {
        <- at the same time on a different CPU ->
        [...]
        e1000_clean_tx_irq::netif_wake_queue() // Plenty of room has been made
        [...]
        <- back to e1000_xmit_frame() ->
                netif_stop_queue
                return 1; 
        }


--
Ueimor