Re: [PATCH] per-cpu areas for 2.5.3-pre6

[Andrew Morton <akpm@zip.com.au>]

Rusty Russell wrote:
> 
> In message <3C5669D6.B81E0B4@zip.com.au> you write:
> > Rusty Russell wrote:
> > >
> > > This patch introduces the __per_cpu_data tag for data, and the
> > > per_cpu() & this_cpu() macros to go with it.
> > >
> > > This allows us to get rid of all those special case structures
> > > springing up all over the place for CPU-local data.
> >
> > Am I missing something? smp_init() is called quite late in
> > the boot process, and if any code touches per-cpu data before
> > this, it'll get a null pointer deref, won't it?
> 
> Yes.  But for a large amount of code it doesn't matter, and most
> architectures don't know how many CPUs there are before smp_init().
> Of course, we could make it NR_CPUS...

I don't think there's a need.  You can use .data.percpu directly
for the boot CPU, dynamically allocate storage for the others.
This actually saves one CPU's worth of memory :)

> Do you have an example where you want to use this before
> smp_boot_cpus()?  If so, we can bite the bullet.  Otherwise I'd prefer
> not to waste memory.

Well, the slab allocator uses per-CPU data, so with the current patch,
slab.c wouldn't be able to use per_cpu().

But if you use:

 unsigned long __per_cpu_offset[NR_CPUS] = { (unsigned long *)&__per_cpu_start, };

Then each CPU has, at all times, a valid personal __per_cpu_offset[]
entry.   The only restriction is that the boot CPU cannot
touch other CPU's per-cpu data before those CPUs are brought
up.

-