On Thu, Jul 05, 2018 at 08:05:03PM +0200, Peter Zijlstra wrote:
> On Mon, Jul 02, 2018 at 01:30:19AM +0800, Guo Ren wrote:
> > +static irqreturn_t handle_ipi(int irq, void *dev)
> > +{
> > + unsigned long *pending_ipis = &ipi_data[smp_processor_id()].bits;
> > +
> > + while (true) {
> > + unsigned long ops;
> > +
> > + /* Order bit clearing and data access. */
> > + mb();
> > +
> > + ops = xchg(pending_ipis, 0);
> > + if (ops == 0)
> > + return IRQ_HANDLED;
> > +
> > + if (ops & (1 << IPI_RESCHEDULE))
> > + scheduler_ipi();
> > +
> > + if (ops & (1 << IPI_CALL_FUNC))
> > + generic_smp_call_function_interrupt();
> > +
> > + BUG_ON((ops >> IPI_MAX) != 0);
> > +
> > + /* Order data access and bit testing. */
> > + mb();
> > + }
> > +
> > + return IRQ_HANDLED;
> > +}
> > +
> > +static void (*send_arch_ipi)(const unsigned long *mask, unsigned long irq) = NULL;
> > +
> > +void __init set_send_ipi(void (*func)(const unsigned long *, unsigned long))
> > +{
> > + if (send_arch_ipi)
> > + return;
> > +
> > + send_arch_ipi = func;
> > +}
> > +
> > +static void
> > +send_ipi_message(const struct cpumask *to_whom, enum ipi_message_type operation)
> > +{
> > + int i;
> > +
> > + mb();
> > + for_each_cpu(i, to_whom)
> > + set_bit(operation, &ipi_data[i].bits);
> > +
> > + mb();
> > + send_arch_ipi(cpumask_bits(to_whom), IPI_IRQ);
> > +}
>
>
> Please explain those mb()'s... I'm thinking you meant to use smp_mb().
Yes, smp_mb(). Current smp_mb()&mb() is the same: sync.is.
In next version patch, I'll seperate smp_mb() and mb() and use ld/st.barrier
instead of sync.is. Sync.is is expensive that it flush cpu's pipeline.
> But then for handle_ipi(), the xchg() should already imply all those.
Yes, approve.
> And the send_ipi_message() only needs the second.
Yes, approve.
Guo Ren
