Re: [PATCH/RFC] cpuhp code and data usage on UP systems

From: Thomas Gleixner
Date: Mon Oct 02 2017 - 14:52:32 EST


On Mon, 2 Oct 2017, Nicolas Pitre wrote:
>
> On !SMP systems, we end up with the following array definition:
>
> static struct cpuhp_step cpuhp_ap_states[] = {
> [CPUHP_ONLINE] = {
> .name = "online",
> .startup.single = NULL,
> .teardown.single = NULL,
> },
> };
>
> where CPUHP_ONLINE = 187. That means up to 99.5% of this array is unused
> but still allocated in the kernel's .data section i.e. 3760 bytes.
>
> The same issue exists with cpuhp_bp_states being 1720 bytes.
>
> The goal of this patch is mainly about illustrating the issue and
> reducing .data usage. I have no clear idea when this stuff is really
> needed besides standard CPU hotplug. Not knowing exactly what I'm doing
> here, I made it conditional on CONFIG_SMP for now. There might be a case
> for moving that code to a separate file (cpuhp.c maybe?) and omitting it
> from the kernel when unneeded.
>
> Comments?

Sure.

> +#ifdef CONFIG_CPUHP
> +#define ___P(proto, def_retcode) extern proto;
> +#else
> +#define ___P(proto, def_retcode) static inline proto { return def_retcode }
> +#endif
> +
> +___P(
> int __cpuhp_setup_state(enum cpuhp_state state, const char *name, bool invoke,
> int (*startup)(unsigned int cpu),
> - int (*teardown)(unsigned int cpu), bool multi_instance);
> -
> + int (*teardown)(unsigned int cpu), bool multi_instance), 0; )

What exactly is invoking the callbacks of states in the proper context and
manages instances on registration? Ditto for teardwon in case of modules.

Thanks,

tglx