Re: [patch] kmalloc_percpu -- 2 of 2

[Andrew Morton <akpm@digeo.com>]

Dipankar Sarma wrote:
> 
> On Thu, Dec 05, 2002 at 09:10:16PM +0100, Andrew Morton wrote:
> >
> > I'd suggest that you drop the new allocator until a compelling
> > need for it (in real, live 2.5/2.6 code) has been demonstrated.
> 
> Fine with me since atleast one workaround for fragmentation with small
> allocations is known. I can't see anything in 2.5 timeframe
> requiring small per-cpu allocations.
> 
> Would you like me to resubmit a simple kmalloc-only version ?
> 

I think that would be best.

BTW, looking at the snmp application of this work:

+#define ICMP_INC_STATS_USER_FIELD(offt)                                \
+       (*((unsigned long *) ((void *)                                  \
+                            per_cpu_ptr(icmp_statistics[1],            \
+                                        smp_processor_id())) + offt))++;

This guy is racy on preempt.  Just a little bit.  It is effectively:

	ptr = per_cpu_ptr(...);
	(*ptr)++;

On some architectures, `(*ptr)++' is not atomic wrt interrupts.  The
CPU could be preempted midway through the increment.

Surely it's not an issue for SNMP stats, but for some applications
such as struct page_state, such a permanent off-by-a-little-bit would
be a showstopper.

So some big loud comments which describe the worthiness of get_cpu_ptr(),
and the potential inaccuracy of per_cpu_ptr would be useful.

And as this is the first application of the kmalloc_precpu infrastructure,
it may be best to convert it to use get_cpu_ptr/put_cpu_ptr.