Re: [PATCH] kmalloc_percpu

[David S. Miller]

On Tue, 2003-05-06 at 01:03, Rusty Russell wrote:
> And #3 is, in fact, the one I care about.  The extra memory reference
> is already probably cachehot (all static per-cpu use it), and might be
> in a register on some archs.
> 
> Doesn't break with sparce CPU #s, but yes, it is inefficient.

Maybe we should treat this stuff like architecture ABIs that
have a "small data" section?

Before elaborating, let me state that if we are going to use this
for things like disk stats and per-interface ipv6 ICMP stats, a
fixed size pool is absolutely unacceptable.  People are configuring up
thousands upon thousands of VLAN net devices today, and if IPV6 is
enabled each of those will get a per-cpu ICMP stats block allocated.
And as Andrew said, there can be thousands of block devices.

Therefore, for these per-cpu applications there must be no limits built
into the mechanism.

Now, what I propose is that kmalloc_percpu() keeps it's current
implementation.  Then we have a kmalloc_percpu_small() that must only
be invoked during module init and we limit the size to some reasonable
amount.  This kmalloc_percpu_small() uses the 32K pool or whatever, as
does DECLARE_PERCPU in a module.

How about this?

-- 
David S. Miller <[e-mail]>