Re: [RFC 00/15] x86_64: Optimize percpu accesses

[Mike Travis]

H. Peter Anvin wrote:
> Mike Travis wrote:
>>
>> The biggest growth came from moving all the xxx[NR_CPUS] arrays into
>> the per cpu area.  So you free up a huge amount of unused memory when
>> the NR_CPUS count starts getting into the ozone layer.  4k now, 16k
>> real soon now, ??? future?
>>
> 
> Even (or perhaps especially) so, allocating the percpu area in 2 MB
> increments is a total nonstarter.  It hurts the small, common
> configurations way too much.  For SGI, it's probably fine.
> 
>     -hpa

Yes, "right-sizing" the kernel for systems from 512M laptops to 4k cpu
systems with memory totally maxed out, using only a binary distribution
has proven tricky (at best... ;-)

One alternative was to only allocate a chunk similar in size to
PERCPU_ENOUGH_ROOM and allow for startup options to create a bigger
space if needed.  Though the realloc idea has some merit if we can
validate the non-use of pointers to specific cpu's percpu vars.
(CPU_ALLOC contains a function to dereference a percpu offset.)

The discussion started at:

	http://marc.info/?l=linux-kernel&m=121212026716085&w=4

Thanks,
Mike
--
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/