On 22-07-2007 09:05, David Miller wrote: > From: Stephen Hemminger <[EMAIL PROTECTED]> > Date: Thu, 19 Jul 2007 17:27:47 +0100 > >> Please revisit the requirements that netconsole needs and redesign >> it from scratch. The existing code is causing too much breakage. >> >> Can it be done without breaking the semantics of network devices, or >> should we rewrite the driver interface to take have a different >> interface like netdev_sync_send_skb() that is slow, synchronous and >> non-interrupt (ie polls for completion). Of course, then people >> will complain that netconsole traffic slows the machine down. for >> completion. > > I couldn't agree more. > > Since netpoll runs outside of all of the normal netdevice locking > rules, only the people using netpoll hit all the bugs. That means > most of us do not test out these code path, which is bad. > > So, if anything, a more integrated implementation is essential. >
But, IMHO, until this will be done some simpler measures could do be done to make poll_napi less dangerous (as a matter of fact I wonder why oopses observed & diagnosed by Olaf are so rare). Current locking with netpoll_poll_lock is mainly misleading. It seems somebody who planned napi didn't even think such helpers as poll_napi are possible on other cpus and somebody doing netpoll didn't want to show this all per cpu data needs full locking anyway. But since it's like this there is no reason to invent a wheel again and "normal" locking should be done: so global (not per device) spin_lock (#ifdef CONFIG_NETPOLL only) held during all net_rx_action and spin_trylock similarly in poll_napi (with STATE_RX_SCHED re-checking) should be minimum needed here. Regards, Jarek P. - To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
