* [PATCH] powerpc: Reduce footprint of irq_stat
@ 2010-01-12 10:56 Anton Blanchard
2010-01-12 10:58 ` [PATCH] powerpc: Reduce footprint of xics_ipi_struct Anton Blanchard
2010-02-01 5:21 ` [PATCH] powerpc: Reduce footprint of irq_stat Benjamin Herrenschmidt
0 siblings, 2 replies; 6+ messages in thread
From: Anton Blanchard @ 2010-01-12 10:56 UTC (permalink / raw)
To: benh; +Cc: linuxppc-dev
PowerPC is currently using asm-generic/hardirq.h which statically allocates an
NR_CPUS irq_stat array. Switch to an arch specific implementation which uses
per cpu data:
On a kernel with NR_CPUS=1024, this saves quite a lot of memory:
text data bss dec hex filename
8767938 2944132 1636796 13348866 cbb002 vmlinux.baseline
8767779 2944260 1505724 13217763 c9afe3 vmlinux.irq_cpustat
A saving of around 128kB.
Signed-off-by: Anton Blanchard <anton@samba.org>
---
Index: linux-cpumask/arch/powerpc/include/asm/hardirq.h
===================================================================
--- linux-cpumask.orig/arch/powerpc/include/asm/hardirq.h 2010-01-06 16:24:31.220084536 +1100
+++ linux-cpumask/arch/powerpc/include/asm/hardirq.h 2010-01-06 17:45:25.219563704 +1100
@@ -1 +1,22 @@
-#include <asm-generic/hardirq.h>
+#ifndef _ASM_POWERPC_HARDIRQ_H
+#define _ASM_POWERPC_HARDIRQ_H
+
+#include <linux/threads.h>
+#include <linux/irq.h>
+
+typedef struct {
+ unsigned int __softirq_pending;
+} ____cacheline_aligned irq_cpustat_t;
+
+DECLARE_PER_CPU_SHARED_ALIGNED(irq_cpustat_t, irq_stat);
+
+#define __ARCH_IRQ_STAT
+
+#define local_softirq_pending() __get_cpu_var(irq_stat).__softirq_pending
+
+static inline void ack_bad_irq(unsigned int irq)
+{
+ printk(KERN_CRIT "unexpected IRQ trap at vector %02x\n", irq);
+}
+
+#endif /* _ASM_POWERPC_HARDIRQ_H */
Index: linux-cpumask/arch/powerpc/kernel/irq.c
===================================================================
--- linux-cpumask.orig/arch/powerpc/kernel/irq.c 2010-01-06 16:29:31.370083802 +1100
+++ linux-cpumask/arch/powerpc/kernel/irq.c 2010-01-06 16:48:04.210086075 +1100
@@ -73,6 +73,9 @@
#define CREATE_TRACE_POINTS
#include <asm/trace.h>
+DEFINE_PER_CPU_SHARED_ALIGNED(irq_cpustat_t, irq_stat);
+EXPORT_PER_CPU_SYMBOL(irq_stat);
+
int __irq_offset_value;
static int ppc_spurious_interrupts;
^ permalink raw reply [flat|nested] 6+ messages in thread* [PATCH] powerpc: Reduce footprint of xics_ipi_struct 2010-01-12 10:56 [PATCH] powerpc: Reduce footprint of irq_stat Anton Blanchard @ 2010-01-12 10:58 ` Anton Blanchard 2010-02-01 5:18 ` Benjamin Herrenschmidt 2010-02-01 5:21 ` [PATCH] powerpc: Reduce footprint of irq_stat Benjamin Herrenschmidt 1 sibling, 1 reply; 6+ messages in thread From: Anton Blanchard @ 2010-01-12 10:58 UTC (permalink / raw) To: benh; +Cc: linuxppc-dev 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> --- 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; 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 ^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] powerpc: Reduce footprint of xics_ipi_struct 2010-01-12 10:58 ` [PATCH] powerpc: Reduce footprint of xics_ipi_struct Anton Blanchard @ 2010-02-01 5:18 ` Benjamin Herrenschmidt 2010-02-01 6:26 ` Anton Blanchard 0 siblings, 1 reply; 6+ messages in thread From: Benjamin Herrenschmidt @ 2010-02-01 5:18 UTC (permalink / raw) To: Anton Blanchard; +Cc: linuxppc-dev 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 ^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] powerpc: Reduce footprint of xics_ipi_struct 2010-02-01 5:18 ` Benjamin Herrenschmidt @ 2010-02-01 6:26 ` Anton Blanchard 0 siblings, 0 replies; 6+ messages in thread From: Anton Blanchard @ 2010-02-01 6:26 UTC (permalink / raw) To: Benjamin Herrenschmidt; +Cc: linuxppc-dev 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 ^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] powerpc: Reduce footprint of irq_stat 2010-01-12 10:56 [PATCH] powerpc: Reduce footprint of irq_stat Anton Blanchard 2010-01-12 10:58 ` [PATCH] powerpc: Reduce footprint of xics_ipi_struct Anton Blanchard @ 2010-02-01 5:21 ` Benjamin Herrenschmidt 2010-02-01 5:44 ` Anton Blanchard 1 sibling, 1 reply; 6+ messages in thread From: Benjamin Herrenschmidt @ 2010-02-01 5:21 UTC (permalink / raw) To: Anton Blanchard; +Cc: linuxppc-dev > +typedef struct { > + unsigned int __softirq_pending; > +} ____cacheline_aligned irq_cpustat_t; > + > +DECLARE_PER_CPU_SHARED_ALIGNED(irq_cpustat_t, irq_stat); > + > +#define __ARCH_IRQ_STAT > + > +#define local_softirq_pending() __get_cpu_var(irq_stat).__softirq_pending > + > +static inline void ack_bad_irq(unsigned int irq) > +{ > + printk(KERN_CRIT "unexpected IRQ trap at vector %02x\n", irq); > +} Looks like some unrelated gunk slipped into this patch :-) Cheers, Ben. ^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH] powerpc: Reduce footprint of irq_stat 2010-02-01 5:21 ` [PATCH] powerpc: Reduce footprint of irq_stat Benjamin Herrenschmidt @ 2010-02-01 5:44 ` Anton Blanchard 0 siblings, 0 replies; 6+ messages in thread From: Anton Blanchard @ 2010-02-01 5:44 UTC (permalink / raw) To: Benjamin Herrenschmidt; +Cc: linuxppc-dev Hi, > > +static inline void ack_bad_irq(unsigned int irq) > > +{ > > + printk(KERN_CRIT "unexpected IRQ trap at vector %02x\n", irq); > > +} > > Looks like some unrelated gunk slipped into this patch :-) We need that to link :) kernel/irq/handle.c: In function ‘handle_bad_irq’: kernel/irq/handle.c:43: error: implicit declaration of function ‘ack_bad_irq’ It's basically a copy of the fallback handler in the asm-generic hardirq.h I'm happy to change it print "IRQ on fire" :) Anton ^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2010-02-01 6:26 UTC | newest] Thread overview: 6+ messages (download: mbox.gz follow: Atom feed -- links below jump to the message on this page -- 2010-01-12 10:56 [PATCH] powerpc: Reduce footprint of irq_stat Anton Blanchard 2010-01-12 10:58 ` [PATCH] powerpc: Reduce footprint of xics_ipi_struct Anton Blanchard 2010-02-01 5:18 ` Benjamin Herrenschmidt 2010-02-01 6:26 ` Anton Blanchard 2010-02-01 5:21 ` [PATCH] powerpc: Reduce footprint of irq_stat Benjamin Herrenschmidt 2010-02-01 5:44 ` Anton Blanchard
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for NNTP newsgroup(s).