Re: [PATCH 01/14] Introduce cpu_enabled_map and friends

From: Russell King
Date: Fri Jul 18 2008 - 17:56:16 EST


On Tue, Jul 15, 2008 at 01:15:15PM -0600, Alex Chiang wrote:
> * Russell King <rmk+lkml@xxxxxxxxxxxxxxxx>:
> > On Tue, Jul 15, 2008 at 12:16:32PM -0600, Matthew Wilcox wrote:
> > > On Tue, Jul 15, 2008 at 11:57:40AM -0600, Alex Chiang wrote:
> > > > My thought was that big SMP systems like ia64, possibly sparc and
> > > > ppc, and increasingly, x86, might find something like this
> > > > useful, as systems get larger and larger, and vendors are going
> > > > to want to do RAS-ish features, like the ability to keep CPUs in
> > > > firmware across reboots until told otherwise by the sysadmin.
> > > >
> > > > Right now, a 'present' CPU strongly implies 'online' as well,
> > > > since we're calling cpu_up() for all 'present' CPUs in
> > > > smp_init(). But this hurts if:
> > > >
> > > > - you don't actually want to bring up all 'present' CPUs
> > > > - you still want to interact with these weirdo zombie
> > > > CPUs that are 'present' but not 'online'
> > >
> > > Have you considered simply failing __cpu_up() for CPUs that are
> > > deconfigured by firmware?
> >
> > But what if you want to have a system boot with, say, 4 CPUs and
> > then decide at run time to bring up another 4 CPUs when required?
> >
> > How about having smp_init() call into arch code to query whether
> > it should bring up a not-already-online CPU? Architectures that
> > want to do something special can then make the decision there and
> > everyone else can define the test completely away.
>
> So this is exactly what I'm doing. The ia64 patch has this hunk:
>
> @@ -820,6 +824,9 @@ __cpu_up (unsigned int cpu)
> if (cpu_isset(cpu, cpu_callin_map))
> return -EINVAL;
>
> + if (!cpu_isset(cpu, cpu_enabled_map))
> + return -EINVAL;
> +
> per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
> /* Processor goes to start_secondary(), sets online flag */
> ret = do_boot_cpu(sapicid, cpu);
>
> That was the easiest, most-straightforward solution I could think
> of. If you have an idea for a version with lower taxes (doesn't
> touch all the archs or can be #define'd out), I'm happy to hear
> it.

I think I did make a suggestion in the bit you quote from me above.

Let me be more explicit:

static void __init smp_init(void)
{
unsigned int cpu;

/* FIXME: This should be done in userspace --RR */
for_each_present_cpu(cpu) {
if (num_online_cpus() >= setup_max_cpus)
break;
- if (!cpu_online(cpu))
+ if (smp_cpu_enabled(cpu) && !cpu_online(cpu))
cpu_up(cpu);
}

/* Any cleanup work */
printk(KERN_INFO "Brought up %ld CPUs\n", (long)num_online_cpus());
smp_cpus_done(setup_max_cpus);
}

and have architectures provide 'smp_cpu_enabled(cpu)' which can either
be a function, inline function or a macro (and therefore possible to be
completely eliminated.)

--
Russell King
Linux kernel 2.6 ARM Linux - http://www.arm.linux.org.uk/
maintainer of:
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/