* [parisc-linux] [PATCH] Remove spinlock_t from cpu_data (and thus from processor.h)
@ 2006-11-25 16:53 Kyle McMartin
2006-11-26 7:33 ` Grant Grundler
2006-11-26 18:38 ` Matthew Wilcox
0 siblings, 2 replies; 5+ messages in thread
From: Kyle McMartin @ 2006-11-25 16:53 UTC (permalink / raw)
To: parisc-linux
This should enable us to clean up our header files nicely.
Signed-off-by: Kyle McMartin <kyle@parisc-linux.org>
---
diff --git a/arch/parisc/kernel/processor.c b/arch/parisc/kernel/processor.c
index fb81e56..0af1fad 100644
--- a/arch/parisc/kernel/processor.c
+++ b/arch/parisc/kernel/processor.c
@@ -153,8 +153,6 @@ #endif
p->cpuid = cpuid; /* save CPU id */
p->txn_addr = txn_addr; /* save CPU IRQ address */
#ifdef CONFIG_SMP
- spin_lock_init(&p->lock);
-
/*
** FIXME: review if any other initialization is clobbered
** for boot_cpu by the above memset().
diff --git a/arch/parisc/kernel/smp.c b/arch/parisc/kernel/smp.c
index 4a23a97..5301600 100644
--- a/arch/parisc/kernel/smp.c
+++ b/arch/parisc/kernel/smp.c
@@ -76,6 +76,9 @@ cpumask_t cpu_possible_map __read_mostly
EXPORT_SYMBOL(cpu_online_map);
EXPORT_SYMBOL(cpu_possible_map);
+spinlock_t fencepost[NR_CPUS] = {
+ [0 ... NR_CPUS-1] = SPIN_LOCK_UNLOCKED
+};
struct smp_call_struct {
void (*func) (void *info);
@@ -167,10 +170,11 @@ ipi_interrupt(int irq, void *dev_id)
mb(); /* Order interrupt and bit testing. */
for (;;) {
- spin_lock_irqsave(&(p->lock),flags);
+ spinlock_t *lock = &fencepost[this_cpu];
+ spin_lock_irqsave(lock, flags);
ops = p->pending_ipi;
p->pending_ipi = 0;
- spin_unlock_irqrestore(&(p->lock),flags);
+ spin_unlock_irqrestore(lock, flags);
mb(); /* Order bit clearing and data access. */
@@ -275,12 +279,13 @@ static inline void
ipi_send(int cpu, enum ipi_message_type op)
{
struct cpuinfo_parisc *p = &cpu_data[cpu];
+ spinlock_t *lock = &fencepost[cpu];
unsigned long flags;
- spin_lock_irqsave(&(p->lock),flags);
+ spin_lock_irqsave(lock, flags);
p->pending_ipi |= 1 << op;
gsc_writel(IPI_IRQ - CPU_IRQ_BASE, cpu_data[cpu].hpa);
- spin_unlock_irqrestore(&(p->lock),flags);
+ spin_unlock_irqrestore(lock, flags);
}
diff --git a/include/asm-parisc/processor.h b/include/asm-parisc/processor.h
index fd7866d..435afe5 100644
--- a/include/asm-parisc/processor.h
+++ b/include/asm-parisc/processor.h
@@ -87,7 +87,6 @@ struct cpuinfo_parisc {
unsigned long hpa; /* Host Physical address */
unsigned long txn_addr; /* MMIO addr of EIR or id_eid */
#ifdef CONFIG_SMP
- spinlock_t lock; /* synchronization for ipi's */
unsigned long pending_ipi; /* bitmap of type ipi_message_type */
unsigned long ipi_count; /* number ipi Interrupts */
#endif
_______________________________________________
parisc-linux mailing list
parisc-linux@lists.parisc-linux.org
http://lists.parisc-linux.org/mailman/listinfo/parisc-linux
^ permalink raw reply related [flat|nested] 5+ messages in thread* Re: [parisc-linux] [PATCH] Remove spinlock_t from cpu_data (and thus from processor.h)
2006-11-25 16:53 [parisc-linux] [PATCH] Remove spinlock_t from cpu_data (and thus from processor.h) Kyle McMartin
@ 2006-11-26 7:33 ` Grant Grundler
2006-11-26 18:37 ` Kyle McMartin
2006-11-26 18:38 ` Matthew Wilcox
1 sibling, 1 reply; 5+ messages in thread
From: Grant Grundler @ 2006-11-26 7:33 UTC (permalink / raw)
To: Kyle McMartin; +Cc: parisc-linux
On Sat, Nov 25, 2006 at 11:53:44AM -0500, Kyle McMartin wrote:
> This should enable us to clean up our header files nicely.
...
Yes, but...
> +spinlock_t fencepost[NR_CPUS] = {
> + [0 ... NR_CPUS-1] = SPIN_LOCK_UNLOCKED
> +};
Unless each CPU gets it's own cacheline, I believe this is
going to ping-pong across CPU caches like hell. Thus
the "per-CPU" initialization. If the fencepost[] locks
are seldom used, then it doesn't matter and we could just use
one cacheline.
but I think you are on the right track to unraveling our
header dependencies a bit more...
thanks,
grant
_______________________________________________
parisc-linux mailing list
parisc-linux@lists.parisc-linux.org
http://lists.parisc-linux.org/mailman/listinfo/parisc-linux
^ permalink raw reply [flat|nested] 5+ messages in thread* Re: [parisc-linux] [PATCH] Remove spinlock_t from cpu_data (and thus from processor.h)
2006-11-26 7:33 ` Grant Grundler
@ 2006-11-26 18:37 ` Kyle McMartin
2006-11-26 18:56 ` Grant Grundler
0 siblings, 1 reply; 5+ messages in thread
From: Kyle McMartin @ 2006-11-26 18:37 UTC (permalink / raw)
To: parisc-linux
On Sun, Nov 26, 2006 at 12:33:06AM -0700, Grant Grundler wrote:
>
diff --git a/arch/parisc/kernel/smp.c b/arch/parisc/kernel/smp.c
index 4a23a97..5b6bc6e 100644
--- a/arch/parisc/kernel/smp.c
+++ b/arch/parisc/kernel/smp.c
@@ -76,6 +76,7 @@ cpumask_t cpu_possible_map __read_mostly
EXPORT_SYMBOL(cpu_online_map);
EXPORT_SYMBOL(cpu_possible_map);
+DEFINE_PER_CPU(spinlock_t, ipi_lock) = SPIN_LOCK_UNLOCKED;
struct smp_call_struct {
void (*func) (void *info);
@@ -167,10 +168,11 @@ ipi_interrupt(int irq, void *dev_id)
mb(); /* Order interrupt and bit testing. */
for (;;) {
- spin_lock_irqsave(&(p->lock),flags);
+ spinlock_t *lock = &per_cpu(ipi_lock, this_cpu);
+ spin_lock_irqsave(lock, flags);
ops = p->pending_ipi;
p->pending_ipi = 0;
- spin_unlock_irqrestore(&(p->lock),flags);
+ spin_unlock_irqrestore(lock, flags);
mb(); /* Order bit clearing and data access. */
@@ -275,12 +277,13 @@ static inline void
ipi_send(int cpu, enum ipi_message_type op)
{
struct cpuinfo_parisc *p = &cpu_data[cpu];
+ spinlock_t *lock = &per_cpu(ipi_lock, cpu);
unsigned long flags;
- spin_lock_irqsave(&(p->lock),flags);
+ spin_lock_irqsave(lock, flags);
p->pending_ipi |= 1 << op;
gsc_writel(IPI_IRQ - CPU_IRQ_BASE, cpu_data[cpu].hpa);
- spin_unlock_irqrestore(&(p->lock),flags);
+ spin_unlock_irqrestore(lock, flags);
}
diff --git a/include/asm-parisc/processor.h b/include/asm-parisc/processor.h
index fd7866d..435afe5 100644
--- a/include/asm-parisc/processor.h
+++ b/include/asm-parisc/processor.h
@@ -87,7 +87,6 @@ struct cpuinfo_parisc {
unsigned long hpa; /* Host Physical address */
unsigned long txn_addr; /* MMIO addr of EIR or id_eid */
#ifdef CONFIG_SMP
- spinlock_t lock; /* synchronization for ipi's */
unsigned long pending_ipi; /* bitmap of type ipi_message_type */
unsigned long ipi_count; /* number ipi Interrupts */
#endif
_______________________________________________
parisc-linux mailing list
parisc-linux@lists.parisc-linux.org
http://lists.parisc-linux.org/mailman/listinfo/parisc-linux
^ permalink raw reply related [flat|nested] 5+ messages in thread* Re: [parisc-linux] [PATCH] Remove spinlock_t from cpu_data (and thus from processor.h)
2006-11-26 18:37 ` Kyle McMartin
@ 2006-11-26 18:56 ` Grant Grundler
0 siblings, 0 replies; 5+ messages in thread
From: Grant Grundler @ 2006-11-26 18:56 UTC (permalink / raw)
To: Kyle McMartin; +Cc: parisc-linux
On Sun, Nov 26, 2006 at 01:37:42PM -0500, Kyle McMartin wrote:
> diff --git a/arch/parisc/kernel/smp.c b/arch/parisc/kernel/smp.c
> index 4a23a97..5b6bc6e 100644
> --- a/arch/parisc/kernel/smp.c
> +++ b/arch/parisc/kernel/smp.c
> @@ -76,6 +76,7 @@ cpumask_t cpu_possible_map __read_mostly
> EXPORT_SYMBOL(cpu_online_map);
> EXPORT_SYMBOL(cpu_possible_map);
>
> +DEFINE_PER_CPU(spinlock_t, ipi_lock) = SPIN_LOCK_UNLOCKED;
If this works - this is the right way to do it.
You rock!
thanks!
grant
>
> struct smp_call_struct {
> void (*func) (void *info);
> @@ -167,10 +168,11 @@ ipi_interrupt(int irq, void *dev_id)
> mb(); /* Order interrupt and bit testing. */
>
> for (;;) {
> - spin_lock_irqsave(&(p->lock),flags);
> + spinlock_t *lock = &per_cpu(ipi_lock, this_cpu);
> + spin_lock_irqsave(lock, flags);
> ops = p->pending_ipi;
> p->pending_ipi = 0;
> - spin_unlock_irqrestore(&(p->lock),flags);
> + spin_unlock_irqrestore(lock, flags);
>
> mb(); /* Order bit clearing and data access. */
>
> @@ -275,12 +277,13 @@ static inline void
> ipi_send(int cpu, enum ipi_message_type op)
> {
> struct cpuinfo_parisc *p = &cpu_data[cpu];
> + spinlock_t *lock = &per_cpu(ipi_lock, cpu);
> unsigned long flags;
>
> - spin_lock_irqsave(&(p->lock),flags);
> + spin_lock_irqsave(lock, flags);
> p->pending_ipi |= 1 << op;
> gsc_writel(IPI_IRQ - CPU_IRQ_BASE, cpu_data[cpu].hpa);
> - spin_unlock_irqrestore(&(p->lock),flags);
> + spin_unlock_irqrestore(lock, flags);
> }
>
>
> diff --git a/include/asm-parisc/processor.h b/include/asm-parisc/processor.h
> index fd7866d..435afe5 100644
> --- a/include/asm-parisc/processor.h
> +++ b/include/asm-parisc/processor.h
> @@ -87,7 +87,6 @@ struct cpuinfo_parisc {
> unsigned long hpa; /* Host Physical address */
> unsigned long txn_addr; /* MMIO addr of EIR or id_eid */
> #ifdef CONFIG_SMP
> - spinlock_t lock; /* synchronization for ipi's */
> unsigned long pending_ipi; /* bitmap of type ipi_message_type */
> unsigned long ipi_count; /* number ipi Interrupts */
> #endif
> _______________________________________________
> parisc-linux mailing list
> parisc-linux@lists.parisc-linux.org
> http://lists.parisc-linux.org/mailman/listinfo/parisc-linux
_______________________________________________
parisc-linux mailing list
parisc-linux@lists.parisc-linux.org
http://lists.parisc-linux.org/mailman/listinfo/parisc-linux
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [parisc-linux] [PATCH] Remove spinlock_t from cpu_data (and thus from processor.h)
2006-11-25 16:53 [parisc-linux] [PATCH] Remove spinlock_t from cpu_data (and thus from processor.h) Kyle McMartin
2006-11-26 7:33 ` Grant Grundler
@ 2006-11-26 18:38 ` Matthew Wilcox
1 sibling, 0 replies; 5+ messages in thread
From: Matthew Wilcox @ 2006-11-26 18:38 UTC (permalink / raw)
To: Kyle McMartin; +Cc: parisc-linux
On Sat, Nov 25, 2006 at 11:53:44AM -0500, Kyle McMartin wrote:
> This should enable us to clean up our header files nicely.
Thanks for looking at this.
My first thought was "Surely there's a better way to do this, maybe
DEFINE_PER_CPU".
Then I looked at what you'd had to touch. Essentially, we're
multiplexing 6 different ops onto one interrupt. This would make sense
if we were short of interrupts, but on the vast majority of systems, we
aren't.
We define NOP, RESCHEDULE, CALL_FUNC, CPU_START, CPU_STOP and CPU_TEST.
CPU_TEST is never sent, down to 5. CPU_NOP and CPU_RESCHEDULE are the
same other than debug statements, down to 4. So by taking 3 additional
CPU interrupts, we can eliminate the spinlock entirely.
We could do away with RESCHEDULE if we pass a nop function to CALL_FUNC.
I don't know if we would suffer a performance hit from doing that
though.
We could also only enable CPU_STOP if CPU_HOTPLUG was set. And
fold CPU_START/CPU_STOP into the CALL_FUNC bucket. Or deregister
CPU_START after the CPU's come up.
In any case, I think this is a whole lot preferable to the spinlock that
we're currently using. I have some other ideas (like using set_bit()
and xchg() which would take the atomic_hash_spinlock), but I think
redoing our IPI support in the way I've outlined above is cleaner.
_______________________________________________
parisc-linux mailing list
parisc-linux@lists.parisc-linux.org
http://lists.parisc-linux.org/mailman/listinfo/parisc-linux
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2006-11-26 18:56 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2006-11-25 16:53 [parisc-linux] [PATCH] Remove spinlock_t from cpu_data (and thus from processor.h) Kyle McMartin
2006-11-26 7:33 ` Grant Grundler
2006-11-26 18:37 ` Kyle McMartin
2006-11-26 18:56 ` Grant Grundler
2006-11-26 18:38 ` Matthew Wilcox
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox