Re: [PATCH v2 22/44] metag: Time keeping

On 04/01/13 10:05, Vineet Gupta wrote:

Hi James,

On Wednesday 05 December 2012 09:38 PM, James Hogan wrote:

quoted

+static unsigned int hwtimer_freq = HARDWARE_FREQ;
+static DEFINE_PER_CPU(struct clock_event_device, local_clockevent);
+static DEFINE_PER_CPU(char [11], local_clockevent_name);

quoted

+void __cpuinit local_timer_setup(unsigned int cpu)
+{
+	unsigned int txdivtime;
+	struct clock_event_device *clk = &per_cpu(local_clockevent, cpu);
+	char *name = per_cpu(local_clockevent_name, cpu);
+
+	txdivtime = TBI_GETREG(TXDIVTIME);
+
+	txdivtime &= ~TXDIVTIME_DIV_BITS;
+	txdivtime |= (HARDWARE_DIV & TXDIVTIME_DIV_BITS);
+
+	TBI_SETREG(TXDIVTIME, txdivtime);
+
+	sprintf(name, "META %d", cpu);
+	clk->name = name;
+	clk->features = CLOCK_EVT_FEAT_ONESHOT,
+
+	clk->rating = 200,
+	clk->shift = 12,
+	clk->irq = TBID_SIGNUM_TRT,
+	clk->set_mode = metag_timer_set_mode,
+	clk->set_next_event = metag_timer_set_next_event,
+
+	clk->mult = div_sc(hwtimer_freq, NSEC_PER_SEC, clk->shift);
+	clk->max_delta_ns = clockevent_delta2ns(0x7fffffff, clk);
+	clk->min_delta_ns = clockevent_delta2ns(0xf, clk);
+	clk->cpumask = cpumask_of(cpu);
+
+	clockevents_register_device(clk);
+
+	/*
+	 * For all non-boot CPUs we need to synchronize our free
+	 * running clock (TXTIMER) with the boot CPU's clock.
+	 *
+	 * While this won't be accurate, it should be close enough.
+	 */
+	if (cpu) {
+		unsigned int thread0 = cpu_2_hwthread_id[0];
+		unsigned long val;
+
+		val = core_reg_read(TXUCT_ID, TXTIMER_REGNUM, thread0);
+
+		asm volatile("MOV TXTIMER, %0\n" : : "r" (val));
+	}
+}
+
+void __init time_init(void)
+{
+	/*
+	 * On Meta 2 SoCs, the actual frequency of the timer is based on the
+	 * Meta core clock speed divided by an integer, so it is only
+	 * approximately 1MHz. Calculating the real frequency here drastically
+	 * reduces clock skew on these SoCs.
+	 */
+#ifdef CONFIG_METAG_META21
+	hwtimer_freq = get_coreclock() / (metag_in32(EXPAND_TIMER_DIV) + 1);
+#endif
+	clocksource_register_hz(&clocksource_metag, hwtimer_freq);
+
+	setup_irq(TBID_SIGNUM_TRT, &metag_timer_irq);
+
+	local_timer_setup(smp_processor_id());
+}

I have a kludge in ARC port in this subsystem - which I hope you could help clear.

ARC also has a local timer device used for clockevent on each CPU. A one-time
setup_irq() with IRQF_PERCPU - would indeed setup the generic IRQ subsystem - for
making registration effective for all CPUs. However don't you need some per-cpu
magic - say enabling the IRQ at cpu or embedded interrupt controller level -
assuming you starts off with all IRQs disabled (which ARC Linux does).

Hi Vineet,

For Meta this is done in secondary_start_kernel in
arch/metag/kernel/smp.c (see
https://github.com/jahogan/metag-linux/blob/metag-core/arch/metag/kernel/smp.c#L276).
It uses tbi_startup_interrupt which is also called by the irq_startup
callback for the root irq_chip.

So we end up using different APIs - request_percpu_irq() as equivalent of
setup_irq() on boot-cpu only and then each CPU calling enable_percpu_irq() to do
the local magic. request_percpu_irq() in turn requires an apriori call to
irq_set_percpu_devid().

For Meta we could probably do something similar if we moved it's root
irq_startup irq_chip callback code into irq_enable or irq_unmask.

Cheers
James

https://lkml.org/lkml/2012/11/7/128

Do don't seem to be requiring all of this hence I'm wondering how it works for you!

-Vineet