On Tue, 19 Oct 2004 18:42:11 -0400
Lee Revell <rlrevell@xxxxxxxxxxx> wrote:
> On Tue, 2004-10-19 at 18:33, David S. Miller wrote:
> > On Tue, 19 Oct 2004 18:10:58 -0400
> > Lee Revell <rlrevell@xxxxxxxxxxx> wrote:
> > > /*
> > > * Since receiving is always initiated from a tasklet (in iucv.c),
> > > * we must use netif_rx_ni() instead of netif_rx()
> > > */
> > >
> > > This implies that the author thought it was a matter of correctness to
> > > use netif_rx_ni vs. netif_rx. But it looks like the only difference is
> > > that the former sacrifices preempt-safety for performance.
> > You can't really delete netif_rx_ni(), so if there is a preemptability
> > issue, just add the necessary preemption protection around the softirq
> > checks.
> Why not? AIUI the only valid reason to use preempt_disable/enable is in
> the case of per-CPU data. This is not "real" per-CPU data, it's a
> performance hack. Therefore it would be incorrect to add the preemption
> protection, the fix is not to manually call do_softirq but to let the
> softirq run by the normal mechanism.
> Am I missing something?
In code paths where netif_rx_ni() is called, there is not a softirq return
path check, which is why it is checked here.
Theoretically, if you remove the check, softirq processing can be deferred
What I'm saying, therefore, is that netif_rx_ni() it not just a performance
hack, it is necessary for correctness as well.