Re: Races in net_rx_action vs netpoll?

[Olaf Kirch <okir@lst.de>]

On Thursday 12 July 2007 04:33, David Miller wrote:
> I'll add merge your patch with a target of 2.6.23
> 
> If you really want, after this patch has sat in 2.6.23 for a while
> and got some good testing, we can consider a submission for -stable.

Okay, those of you who followed the discussion on lkml will have
read why this patch breaks on e1000.

Short summary: some NIC drivers expect that there is a one-to-one
relation between calls to net_rx_schedule (where we put the device
on the poll list) and netif_rx_complete (where it's supposed to be
taken off the list). The e1000 is such a beast. Not sure if other
drivers make the same assumption re NAPI.

So: should a driver be allowed to rely on this behavior? Or should
I go and look for another fix to the poll_napi issue?

I keep coming back to the question Jarek asked - why does netpoll
want to call dev->poll() anyway? I dug around a little and it
seems the original idea was to do this only if netpoll_poll was
running on the CPU the netdevice was scheduled to.

So one way to fix the problem is to add a dev->poll_cpu field
that tells us on which CPU's poll list it has been added - and
check for this in poll_napi.

Comments?

David, should I submit an updated patch for 2.6.23, or do you
prefer to yank the patch now and try again for 2.6.24?

Olaf
-- 
Olaf Kirch  |  --- o --- Nous sommes du soleil we love when we play
okir@lst.de |    / | \   sol.dhoop.naytheet.ah kin.ir.samse.qurax