Skip to contentSkip to conversation

question about periodic clocks

DISCUSSION Jeremy Fitzhardinge 2 participants 5 messages Mar 9, 2007 – Mar 10, 2007

lore
Share & tools
lore
#0 lore raw

How does the clock period get set on periodic timers? In my clock driver, I'm seeing a call to ->set_mode(CLOCK_EVT_MODE_PERIODIC, evt), but then... nothing. I was expecting a call to set_next_event to set the timer period.

The calltrace is:

#0 xen_new_set_mode (mode=CLOCK_EVT_MODE_PERIODIC, evt=0xc10a2ac0)

    at arch/i386/xen/time.c:275

#1 0xc01323da in clockevents_set_mode (dev=0xc10a2ac0,
mode=CLOCK_EVT_MODE_PERIODIC) at kernel/time/clockevents.c:64
#2 0xc0132854 in tick_setup_periodic (dev=0xc10a2ac0,
broadcast=<value optimized out>) at kernel/time/tick-common.c:111

and tick_setup_periodic does just call clockevents_set_mode, but nothing to set a period.

Am I supposed to assume some default period? HZ? (That's what hpet seems to do.)

Is set_next_event only ever called if the timer is in ONESHOT mode?

Thanks,
J

#1 lore raw
Quoted 4 lines · Jeremy Fitzhardinge

How does the clock period get set on periodic timers? In my clock driver, I'm seeing a call to ->set_mode(CLOCK_EVT_MODE_PERIODIC, evt), but then... nothing. I was expecting a call to set_next_event to set the timer period.

Good point. I never thought about that and we set the period in the clock event device itself. You are right, the clockevents layer should hand over the period either with the set_mode call or seperately. Probably with the set_mode call, as it is needed exactly there and we don't want to have a "if (dev->mode == XXX)" check in set_next_event().

I look into this.

Thanks,

tglx

#2 lore raw
Quoted 9 lines · Thomas Gleixner

Good point. I never thought about that and we set the period in the clock event device itself. You are right, the clockevents layer should hand over the period either with the set_mode call or seperately. Probably with the set_mode call, as it is needed exactly there and we don't want to have a "if (dev->mode == XXX)" check in set_next_event().

I look into this.

So, in the meantime, the period is 1/HZ?

I also have a question about clockevent cpumasks. I was using the lapic clockevent as a model, but as I understand it there's a lapic per CPU, which explains why it registers a clockevent per cpu with that cpu alone in the cpumask.

The Xen timer is a bit different; I guess more like hpet. There's a single (virtual-)machine-wide timer, which is "owned" by the last cpu with programmed it; ie, that cpu is the one which gets the resulting event interrupt. Does this mean I should register a single clockevent device with a cpumask of CPU_MASK_ALL? Or should I constrain it to a single cpu?

There's a comment in hpet.c saying

* Start hpet with the boot cpu mask and make it
* global after the IO_APIC has been initialized.

but I don't see any place where the hpet cpumask is updated.

Thanks,
J

#3 lore raw
Quoted 10 lines · Jeremy Fitzhardinge
Quoted 8 lines · Thomas Gleixner

Good point. I never thought about that and we set the period in the clock event device itself. You are right, the clockevents layer should hand over the period either with the set_mode call or seperately. Probably with the set_mode call, as it is needed exactly there and we don't want to have a "if (dev->mode == XXX)" check in set_next_event().

I look into this.

So, in the meantime, the period is 1/HZ?

Yep.

Quoted 11 lines

I also have a question about clockevent cpumasks. I was using the lapic clockevent as a model, but as I understand it there's a lapic per CPU, which explains why it registers a clockevent per cpu with that cpu alone in the cpumask.

The Xen timer is a bit different; I guess more like hpet. There's a single (virtual-)machine-wide timer, which is "owned" by the last cpu with programmed it; ie, that cpu is the one which gets the resulting event interrupt. Does this mean I should register a single clockevent device with a cpumask of CPU_MASK_ALL? Or should I constrain it to a single cpu?

Uuurg. That's ugly. clockevents expect a per CPU timer especially for dynamic ticks. If you cannot provide a per cpu timer, then you probably need to use the broadcast trick.

Register a primary clocksource (as PIT/HPET) and register per CPU dummy clocksources with CLOCK_EVT_FEAT_DUMMY set - we use the same trick, when the lapic timer is broken. The clockevents code then uses PIT/HPET as the primary tick source and broadcasts the periodic tick to the other CPUs. In that case the dyntick / highres features are disabled.

We did some experiments to support multiple CPUs with one timer for hres/dyntick but it does not scale and it is so ugly that it is not worth the trouble. It works for the lapic stops in C3 case, as we have a well defined point (right before going into the deep power state) where we can rearm the global clock event device. As we are idle at that point anyway there is not much penalty, but I really dont want to do that in an active system.

Quoted 6 lines

There's a comment in hpet.c saying

* Start hpet with the boot cpu mask and make it
* global after the IO_APIC has been initialized.

but I don't see any place where the hpet cpumask is updated.

I wanted to do that in the first place, but never bothered. In an UP environment it does not matter. On a sane SMP box (where we do not have the local APIC stops in C3 problem) the HPET (analogous PIT) is switched off for ever. In the case of LAPIC stops in C3 the HPET(PIT) is used as a broadcast fallback. That means before we go into C3 we arm the HPET/PIT for the earliest to expire lapic event of all CPUs. In that case it does not matter, whether HPET/PIT is pinned to CPU#0 or anything else.

tglx

#4 lore raw
Quoted 4 lines · Thomas Gleixner

Uuurg. That's ugly. clockevents expect a per CPU timer especially for dynamic ticks. If you cannot provide a per cpu timer, then you probably need to use the broadcast trick.

Ah, apologies, I'm wrong about this. I misread the Xen code; the timers are per-vcpu, but as an implementation detail within the hypervisor it keeps track of which real cpu is managing that timer. But logically the timers are per-vcpu.

J