Messages in this thread | | | Date | Mon, 8 Jun 2009 09:00:14 +0200 (CEST) | From | Thomas Gleixner <> | Subject | Re: [PATCH] tick: add check for the existence of broadcast clock event device |
| |
On Mon, 8 Jun 2009, Feng Tang wrote: > On Mon, 8 Jun 2009 14:33:14 +0800 Thomas Gleixner <tglx@linutronix.de> wrote: > > On Mon, 8 Jun 2009, Feng Tang wrote: > > > Our apbt driver is pretty similar with HPET's, including its cpu > > > hotplug notifier. But our platform only has 2 available apbt to > > > use, otherwise we will configure it just like HPET, using one timer > > > as bc and others for per-cpu ones, then it won't hit this case > > > > > > There are 2 situations, one is for the normal boot, apbt0 will be > > > inited first and registered to OS as cpu0's timer, then tsc/lapic > > > is calculated based on it, and apbt1 is registered later in a > > > fs_initcall() (just like hpet.c does) after basic kernel core is > > > up. so the sequence is: apbt0 --> lapic0 --> lapic1 --> apbt1 > > > > Hmm, I do not like that at all. That explicitely relies on CPU0 doing > > some work which will kick CPU1. That's fragile as hell. > > I understand the concern. apbt0 is inited in a very early boot phase when > the cpu1 is not up yet, and os don't even know wether there is a cpu1, that's > why we registered apbt1 in fs_initcall(). If we explicitly setup apbt1 when > OS brings up cpu1, it is a little brutal and not generic as only our platform > has apbt, and I guess cpu hotplug maintainer won't like it :p
Why is that a problem ? You already have a special case for apbt0 in the early setup code. So where is the problem when you have an apbt1 init call on CPU1 _before_ the local APIC is initialized on CPU1.
That's definitely saner than relying on magic IPI wakeups.
Thanks,
tglx
| |