Re: [PATCH] tvec_bases too large for per-cpu data

["Jan Beulich" <JBeulich@novell.com>]

>> >Did you consider using alloc_percpu()?
>> 
>> I did, but I saw drawbacks with that (most notably the fact that all instances are allocated at
>> once, possibly wasting a lot of memory).
>
>It's 4k for each cpu which is in the possible_map but which will never be
>brought online.  I don't think that'll be a lot of memory - are there
>machines which have a lot of possible-but-not-really-there CPUs?

I would suppose so. Why wouldn't a machine supporting CPU hotplug not reasonably be able to double,
triple, etc the number of CPUs originally present?

>There _must_ be ordering issues.  Otherwise we'd just dynamically allocate
>all the structs up-front and be done with it.
>
>Presumably the ordering issue is that init_timers() is called before
>kmem_cache_init().  That's non-obvious and should be commented.

That I can easily do, sure.

>- The `#ifdef CONFIG_NUMA' in init_timers_cpu() seems to be unnecessary -
>  kmalloc_node() will use kmalloc() if !NUMA.

That is correct, but I wanted the fallback if kmalloc_node() fails (from briefly looking at that code it didn't
seem like it would do such fallback itself). And calling kmalloc() twice if !NUMA seemed pointless.

>- The likely()s in init_timers_cpu() seems fairly pointless - it's not a
>  fastpath.

OK, will change that.

>- We prefer to do this:
>
>	if (expr) {
>		...
>	} else {
>		...
>	}
>
>  and not
>
>	if (expr) {
>		...
>	}
>	else {
>		...
>	}

I can change that, too, but I don't see why this gets pointed out again and again when there really
is no consistency across the entire kernel...

Jan