On Thu, 11 Aug 2016 14:09:53 +0200
Vitaly Kuznetsov <vkuzn...@redhat.com> wrote:
> Yuval Mintz <yuval.mi...@qlogic.com> writes:
>
> >> +static void netvsc_inject_enable(struct net_device_context
> >> +*net_device_ctx) {
> >> + net_device_ctx->vf_inject = true;
> >> +}
> >> +
> >> +static void netvsc_inject_disable(struct net_device_context
> >> +*net_device_ctx) {
> >> + net_device_ctx->vf_inject = false;
> >> +
> >> + /* Wait for currently active users to drain out. */
> >> + while (atomic_read(&net_device_ctx->vf_use_cnt) != 0)
> >> + udelay(50);
> >> +}
> >
> > That was already the behavior before, but are you certain you
> > want to unconditionally block without any possible timeout?
>
> Yes, this is OK. After PATCH4 of this series there is only one place
> which takes the vf_use_cnt (netvsc_recv_callback()) and it is an
> interrupt handler, there are no sleepable operations there.
>
Since network devices are protected by RCU, it looks like the refcount
is not necessary. I think vf_inject flag and vf_use_cnt could just be replaced
by doing RCU on vf_netdev.
The callback is invoked from tasklet (softirq) context.
