powerpc: Reduce footprint of xics_ipi_struct

Right now we allocate a cacheline sized NR_CPUS array for xics IPI
communication. Since irq_stat is now PowerPC specific and using
DECLARE_PER_CPU_SHARED_ALIGNED (which should mean remote writes to
this should not conflict with other per cpu data), we can put it in there.

On a kernel with NR_CPUS=1024, this saves quite a lot of memory:

   text    data     bss      dec         hex    filename
8767779 2944260 1505724 13217763         c9afe3 vmlinux.irq_cpustat
8767555 2813444 1505724 13086723         c7b003 vmlinux.xics

A saving of around 128kB.

Signed-off-by: Anton Blanchard <anton@samba.org>
---

On Tue, 2010-01-12 at 21:58 +1100, Anton Blanchard wrote:

> Index: linux-cpumask/arch/powerpc/include/asm/hardirq.h
> ===================================================================
> --- linux-cpumask.orig/arch/powerpc/include/asm/hardirq.h	2010-01-12 12:36:47.174226189 +1100
> +++ linux-cpumask/arch/powerpc/include/asm/hardirq.h	2010-01-12 12:36:47.432976459 +1100
> @@ -6,6 +6,9 @@
>  
>  typedef struct {
>  	unsigned int __softirq_pending;
> +#if defined(CONFIG_XICS) && defined(CONFIG_SMP)
> +	unsigned long xics_ipi;
> +#endif
>  } ____cacheline_aligned irq_cpustat_t;

This is still a gross abuse of irq_cpustat_t ... Can't we do
a separate DECLARE_PER_CPU_SHARED_ALIGNED(unsigned long, xics_ipi)
inside xics.c instead ?

Cheers,
Ben.

>  DECLARE_PER_CPU_SHARED_ALIGNED(irq_cpustat_t, irq_stat);
> Index: linux-cpumask/arch/powerpc/platforms/pseries/xics.c
> ===================================================================
> --- linux-cpumask.orig/arch/powerpc/platforms/pseries/xics.c	2010-01-12 12:36:46.905477650 +1100
> +++ linux-cpumask/arch/powerpc/platforms/pseries/xics.c	2010-01-12 12:40:54.782975198 +1100
> @@ -514,15 +514,12 @@ static void __init xics_init_host(void)
>  /*
>   * XICS only has a single IPI, so encode the messages per CPU
>   */
> -struct xics_ipi_struct {
> -        unsigned long value;
> -	} ____cacheline_aligned;
> -
> -static struct xics_ipi_struct xics_ipi_message[NR_CPUS] __cacheline_aligned;
>  
>  static inline void smp_xics_do_message(int cpu, int msg)
>  {
> -	set_bit(msg, &xics_ipi_message[cpu].value);
> +	unsigned long *tgt = &(per_cpu(irq_stat, cpu).xics_ipi);
> +
> +	set_bit(msg, tgt);
>  	mb();
>  	if (firmware_has_feature(FW_FEATURE_LPAR))
>  		lpar_qirr_info(cpu, IPI_PRIORITY);
> @@ -548,25 +545,23 @@ void smp_xics_message_pass(int target, i
>  
>  static irqreturn_t xics_ipi_dispatch(int cpu)
>  {
> +	unsigned long *tgt = &(per_cpu(irq_stat, cpu).xics_ipi);
> +
>  	WARN_ON(cpu_is_offline(cpu));
>  
>  	mb();	/* order mmio clearing qirr */
> -	while (xics_ipi_message[cpu].value) {
> -		if (test_and_clear_bit(PPC_MSG_CALL_FUNCTION,
> -				       &xics_ipi_message[cpu].value)) {
> +	while (*tgt) {
> +		if (test_and_clear_bit(PPC_MSG_CALL_FUNCTION, tgt)) {
>  			smp_message_recv(PPC_MSG_CALL_FUNCTION);
>  		}
> -		if (test_and_clear_bit(PPC_MSG_RESCHEDULE,
> -				       &xics_ipi_message[cpu].value)) {
> +		if (test_and_clear_bit(PPC_MSG_RESCHEDULE, tgt)) {
>  			smp_message_recv(PPC_MSG_RESCHEDULE);
>  		}
> -		if (test_and_clear_bit(PPC_MSG_CALL_FUNC_SINGLE,
> -				       &xics_ipi_message[cpu].value)) {
> +		if (test_and_clear_bit(PPC_MSG_CALL_FUNC_SINGLE, tgt)) {
>  			smp_message_recv(PPC_MSG_CALL_FUNC_SINGLE);
>  		}
>  #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
> -		if (test_and_clear_bit(PPC_MSG_DEBUGGER_BREAK,
> -				       &xics_ipi_message[cpu].value)) {
> +		if (test_and_clear_bit(PPC_MSG_DEBUGGER_BREAK, tgt)) {
>  			smp_message_recv(PPC_MSG_DEBUGGER_BREAK);
>  		}
>  #endif

Hi,

> > +#if defined(CONFIG_XICS) && defined(CONFIG_SMP)
> > +	unsigned long xics_ipi;
> > +#endif
> >  } ____cacheline_aligned irq_cpustat_t;
> 
> This is still a gross abuse of irq_cpustat_t ... Can't we do
> a separate DECLARE_PER_CPU_SHARED_ALIGNED(unsigned long, xics_ipi)
> inside xics.c instead ?

Yeah it was a bit tasteless :) Respinning...

Anton