[net] protecting interfaces from races between control and data ?
Jack Vogel
jfvogel at gmail.com
Mon Aug 5 17:49:16 UTC 2013
Sigh, this ends up being ugly I'm afraid. I need some time to look at code
and think about it.
Jack
On Mon, Aug 5, 2013 at 10:36 AM, Luigi Rizzo <rizzo at iet.unipi.it> wrote:
> On Mon, Aug 5, 2013 at 7:17 PM, Adrian Chadd <adrian at freebsd.org> wrote:
>
> > I'm travelling back to San Jose today; poke me tomorrow and I'll brain
> > dump what I did in ath(4) and the lessons learnt.
> >
> > The TL;DR version - you don't want to grab an extra lock in the
> > read/write paths as that slows things down. Reuse the same per-queue
> > TX/RX lock and have:
> >
> > * a reset flag that is set when something is resetting; that says to
> > the queue "don't bother processing anything, just dive out";
> > * 'i am doing Tx / Rx' flags per queue that is set at the start of
> > TX/RX servicing and finishes at the end; that way the reset code knows
> > if there's something pending;
> > * have the reset path grab each lock, set the 'reset' flag on each,
> > then walk each queue again and make sure they're all marked as 'not
> > doing TX/RX'. At that point the reset can occur, then the flag cna be
> > cleared, then TX/RX can resume.
> >
>
> so this is slightly different from what Bryan suggested (and you endorsed)
> before, as in that case there was a single 'reset' flag IFF_DRV_RUNNING
> protected by the 'core' lock, then a nested round on all tx and rx locks
> to make sure that all customers have seen it.
> In both cases the tx and rx paths only need the per-queue lock.
>
> As i see it, having a per-queue reset flag removes the need for nesting
> core + queue locks, but since this is only in the control path perhaps
> it is not a big deal (and is better to have a single place to look at to
> tell whether or not we should bail out).
>
> cheers
> luigi
> _______________________________________________
> freebsd-net at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-net
> To unsubscribe, send any mail to "freebsd-net-unsubscribe at freebsd.org"
>
More information about the freebsd-current
mailing list