We noticed that recently, reboot of a 1024 cpu machine takes approx 16
minutes of just stopping the cpus. The slowdown was tracked to commit
f96972f which went into v3.7 and then to the stable trees.
x86 does not need to be running the boot cpu to pull reset and I don't
think it is really needed for shutdown either.
I decided to go the "simple" way and make this a config option that is
selected by the x86 arch. I don't know which other arch's would also
benefit, if any.
Signed-off-by: Robin Holt <[email protected]>
To: Andrew Morton <[email protected]>
Cc: Russ Anderson <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: "H. Peter Anvin" <[email protected]>
Cc: Shawn Guo <[email protected]>
Cc: <[email protected]>
---
arch/x86/Kconfig | 3 +++
kernel/Kconfig.shutdown | 3 +++
kernel/sys.c | 4 ++++
3 files changed, 10 insertions(+)
create mode 100644 kernel/Kconfig.shutdown
diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index 70c0f3d..9611942 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -120,6 +120,7 @@ config X86
select OLD_SIGSUSPEND3 if X86_32 || IA32_EMULATION
select OLD_SIGACTION if X86_32
select COMPAT_OLD_SIGACTION if IA32_EMULATION
+ select ARCH_SHUTDOWN_TO_ANY_CPU
config INSTRUCTION_DECODER
def_bool y
@@ -839,6 +840,8 @@ config SCHED_MC
making when dealing with multi-core CPU chips at a cost of slightly
increased overhead in some places. If unsure say N here.
+source "kernel/Kconfig.shutdown"
+
source "kernel/Kconfig.preempt"
config X86_UP_APIC
diff --git a/kernel/Kconfig.shutdown b/kernel/Kconfig.shutdown
new file mode 100644
index 0000000..d79fc04
--- /dev/null
+++ b/kernel/Kconfig.shutdown
@@ -0,0 +1,3 @@
+
+config ARCH_SHUTDOWN_TO_ANY_CPU
+ bool
diff --git a/kernel/sys.c b/kernel/sys.c
index 39c9c4a..c0b8880 100644
--- a/kernel/sys.c
+++ b/kernel/sys.c
@@ -369,7 +369,9 @@ EXPORT_SYMBOL(unregister_reboot_notifier);
void kernel_restart(char *cmd)
{
kernel_restart_prepare(cmd);
+#ifndef CONFIG_ARCH_SHUTDOWN_TO_ANY_CPU
disable_nonboot_cpus();
+#endif
if (!cmd)
printk(KERN_EMERG "Restarting system.\n");
else
@@ -413,7 +415,9 @@ void kernel_power_off(void)
kernel_shutdown_prepare(SYSTEM_POWER_OFF);
if (pm_power_off_prepare)
pm_power_off_prepare();
+#ifndef CONFIG_ARCH_SHUTDOWN_TO_ANY_CPU
disable_nonboot_cpus();
+#endif
syscore_shutdown();
printk(KERN_EMERG "Power down.\n");
kmsg_dump(KMSG_DUMP_POWEROFF);
--
1.8.1.2
* Robin Holt <[email protected]> wrote:
> We noticed that recently, reboot of a 1024 cpu machine takes approx 16
> minutes of just stopping the cpus. The slowdown was tracked to commit
> f96972f which went into v3.7 and then to the stable trees.
>
> x86 does not need to be running the boot cpu to pull reset and I don't
> think it is really needed for shutdown either.
>
> I decided to go the "simple" way and make this a config option that is
> selected by the x86 arch. I don't know which other arch's would also
> benefit, if any.
>
> Signed-off-by: Robin Holt <[email protected]>
> To: Andrew Morton <[email protected]>
> Cc: Russ Anderson <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: Ingo Molnar <[email protected]>
> Cc: "H. Peter Anvin" <[email protected]>
> Cc: Shawn Guo <[email protected]>
> Cc: <[email protected]>
>
> ---
> arch/x86/Kconfig | 3 +++
> kernel/Kconfig.shutdown | 3 +++
> kernel/sys.c | 4 ++++
> 3 files changed, 10 insertions(+)
> create mode 100644 kernel/Kconfig.shutdown
>
> diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> index 70c0f3d..9611942 100644
> --- a/arch/x86/Kconfig
> +++ b/arch/x86/Kconfig
> @@ -120,6 +120,7 @@ config X86
> select OLD_SIGSUSPEND3 if X86_32 || IA32_EMULATION
> select OLD_SIGACTION if X86_32
> select COMPAT_OLD_SIGACTION if IA32_EMULATION
> + select ARCH_SHUTDOWN_TO_ANY_CPU
>
> config INSTRUCTION_DECODER
> def_bool y
> @@ -839,6 +840,8 @@ config SCHED_MC
> making when dealing with multi-core CPU chips at a cost of slightly
> increased overhead in some places. If unsure say N here.
>
> +source "kernel/Kconfig.shutdown"
> +
> source "kernel/Kconfig.preempt"
>
> config X86_UP_APIC
> diff --git a/kernel/Kconfig.shutdown b/kernel/Kconfig.shutdown
> new file mode 100644
> index 0000000..d79fc04
> --- /dev/null
> +++ b/kernel/Kconfig.shutdown
> @@ -0,0 +1,3 @@
> +
> +config ARCH_SHUTDOWN_TO_ANY_CPU
> + bool
> diff --git a/kernel/sys.c b/kernel/sys.c
> index 39c9c4a..c0b8880 100644
> --- a/kernel/sys.c
> +++ b/kernel/sys.c
> @@ -369,7 +369,9 @@ EXPORT_SYMBOL(unregister_reboot_notifier);
> void kernel_restart(char *cmd)
> {
> kernel_restart_prepare(cmd);
> +#ifndef CONFIG_ARCH_SHUTDOWN_TO_ANY_CPU
> disable_nonboot_cpus();
> +#endif
> if (!cmd)
> printk(KERN_EMERG "Restarting system.\n");
> else
> @@ -413,7 +415,9 @@ void kernel_power_off(void)
> kernel_shutdown_prepare(SYSTEM_POWER_OFF);
> if (pm_power_off_prepare)
> pm_power_off_prepare();
> +#ifndef CONFIG_ARCH_SHUTDOWN_TO_ANY_CPU
> disable_nonboot_cpus();
> +#endif
> syscore_shutdown();
> printk(KERN_EMERG "Power down.\n");
> kmsg_dump(KMSG_DUMP_POWEROFF);
Hm, the 'fix' is a pretty ugly workaround that does not fix much IMHO.
I think the original commit:
f96972f2dc63 kernel/sys.c: call disable_nonboot_cpus() in kernel_restart()
actually regressed your 1024 CPU systems, and should possibly be reverted or fixed
in some other fashion - such as by migrating to the primary CPU (on architectures
that require that), instead of hotplug offlining every secondary CPU on every
architecture!
Alternatively, disable_nonboot_cpus() could perhaps be improved to down CPUs in
parallel: issue the CPU-down requests to every CPU, then wait for them to complete
- instead of the loop over every CPU?
This would be the conceptual counter part to parallel boot up of CPUs - something
SGI might be interested in as well?
Thanks,
Ingo
On 04/08/2013 08:57 AM, Ingo Molnar wrote:
>
> I think the original commit:
>
> f96972f2dc63 kernel/sys.c: call disable_nonboot_cpus() in kernel_restart()
>
> actually regressed your 1024 CPU systems, and should possibly be reverted or fixed
> in some other fashion - such as by migrating to the primary CPU (on architectures
> that require that), instead of hotplug offlining every secondary CPU on every
> architecture!
>
> Alternatively, disable_nonboot_cpus() could perhaps be improved to down CPUs in
> parallel: issue the CPU-down requests to every CPU, then wait for them to complete
> - instead of the loop over every CPU?
>
> This would be the conceptual counter part to parallel boot up of CPUs - something
> SGI might be interested in as well?
>
Migrating to the boot processor and then calling stop_machine() to
defang any other processors should be sufficient, no?
I don't know if there is any reason to deschedule all tasks?
-hpa
--
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel. I don't speak on their behalf.
On Mon, Apr 08, 2013 at 05:57:01PM +0200, Ingo Molnar wrote:
>
> * Robin Holt <[email protected]> wrote:
>
> > We noticed that recently, reboot of a 1024 cpu machine takes approx 16
> > minutes of just stopping the cpus. The slowdown was tracked to commit
> > f96972f which went into v3.7 and then to the stable trees.
> >
> > x86 does not need to be running the boot cpu to pull reset and I don't
> > think it is really needed for shutdown either.
> >
> > I decided to go the "simple" way and make this a config option that is
> > selected by the x86 arch. I don't know which other arch's would also
> > benefit, if any.
> >
> > Signed-off-by: Robin Holt <[email protected]>
> > To: Andrew Morton <[email protected]>
> > Cc: Russ Anderson <[email protected]>
> > Cc: Thomas Gleixner <[email protected]>
> > Cc: Ingo Molnar <[email protected]>
> > Cc: "H. Peter Anvin" <[email protected]>
> > Cc: Shawn Guo <[email protected]>
> > Cc: <[email protected]>
> >
> > ---
> > arch/x86/Kconfig | 3 +++
> > kernel/Kconfig.shutdown | 3 +++
> > kernel/sys.c | 4 ++++
> > 3 files changed, 10 insertions(+)
> > create mode 100644 kernel/Kconfig.shutdown
> >
> > diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> > index 70c0f3d..9611942 100644
> > --- a/arch/x86/Kconfig
> > +++ b/arch/x86/Kconfig
> > @@ -120,6 +120,7 @@ config X86
> > select OLD_SIGSUSPEND3 if X86_32 || IA32_EMULATION
> > select OLD_SIGACTION if X86_32
> > select COMPAT_OLD_SIGACTION if IA32_EMULATION
> > + select ARCH_SHUTDOWN_TO_ANY_CPU
> >
> > config INSTRUCTION_DECODER
> > def_bool y
> > @@ -839,6 +840,8 @@ config SCHED_MC
> > making when dealing with multi-core CPU chips at a cost of slightly
> > increased overhead in some places. If unsure say N here.
> >
> > +source "kernel/Kconfig.shutdown"
> > +
> > source "kernel/Kconfig.preempt"
> >
> > config X86_UP_APIC
> > diff --git a/kernel/Kconfig.shutdown b/kernel/Kconfig.shutdown
> > new file mode 100644
> > index 0000000..d79fc04
> > --- /dev/null
> > +++ b/kernel/Kconfig.shutdown
> > @@ -0,0 +1,3 @@
> > +
> > +config ARCH_SHUTDOWN_TO_ANY_CPU
> > + bool
> > diff --git a/kernel/sys.c b/kernel/sys.c
> > index 39c9c4a..c0b8880 100644
> > --- a/kernel/sys.c
> > +++ b/kernel/sys.c
> > @@ -369,7 +369,9 @@ EXPORT_SYMBOL(unregister_reboot_notifier);
> > void kernel_restart(char *cmd)
> > {
> > kernel_restart_prepare(cmd);
> > +#ifndef CONFIG_ARCH_SHUTDOWN_TO_ANY_CPU
> > disable_nonboot_cpus();
> > +#endif
> > if (!cmd)
> > printk(KERN_EMERG "Restarting system.\n");
> > else
> > @@ -413,7 +415,9 @@ void kernel_power_off(void)
> > kernel_shutdown_prepare(SYSTEM_POWER_OFF);
> > if (pm_power_off_prepare)
> > pm_power_off_prepare();
> > +#ifndef CONFIG_ARCH_SHUTDOWN_TO_ANY_CPU
> > disable_nonboot_cpus();
> > +#endif
> > syscore_shutdown();
> > printk(KERN_EMERG "Power down.\n");
> > kmsg_dump(KMSG_DUMP_POWEROFF);
>
> Hm, the 'fix' is a pretty ugly workaround that does not fix much IMHO.
>
> I think the original commit:
>
> f96972f2dc63 kernel/sys.c: call disable_nonboot_cpus() in kernel_restart()
>
> actually regressed your 1024 CPU systems, and should possibly be reverted or fixed
> in some other fashion - such as by migrating to the primary CPU (on architectures
> that require that), instead of hotplug offlining every secondary CPU on every
> architecture!
>
> Alternatively, disable_nonboot_cpus() could perhaps be improved to down CPUs in
> parallel: issue the CPU-down requests to every CPU, then wait for them to complete
> - instead of the loop over every CPU?
>
> This would be the conceptual counter part to parallel boot up of CPUs - something
> SGI might be interested in as well?
Interested, but even more so interested in parellelizing memory setup. ;)
How can we proceed with this?
Robin
On Mon, Apr 08, 2013 at 09:11:06AM -0700, H. Peter Anvin wrote:
> On 04/08/2013 08:57 AM, Ingo Molnar wrote:
> >
> > I think the original commit:
> >
> > f96972f2dc63 kernel/sys.c: call disable_nonboot_cpus() in kernel_restart()
> >
> > actually regressed your 1024 CPU systems, and should possibly be reverted or fixed
> > in some other fashion - such as by migrating to the primary CPU (on architectures
> > that require that), instead of hotplug offlining every secondary CPU on every
> > architecture!
> >
> > Alternatively, disable_nonboot_cpus() could perhaps be improved to down CPUs in
> > parallel: issue the CPU-down requests to every CPU, then wait for them to complete
> > - instead of the loop over every CPU?
> >
> > This would be the conceptual counter part to parallel boot up of CPUs - something
> > SGI might be interested in as well?
> >
>
> Migrating to the boot processor and then calling stop_machine() to
> defang any other processors should be sufficient, no?
>
> I don't know if there is any reason to deschedule all tasks?
My reading of the original commit indicated that some architecture's
firmware needs the boot cpu to be the one initiating reboot.
If that is correct, then I can not see why a stop_machine() implementation
will not work.
Since this is in generic kernel code, how can I proceed?
Robin
On Mon, Apr 08, 2013 at 05:57:01PM +0200, Ingo Molnar wrote:
>
> * Robin Holt <[email protected]> wrote:
>
> > We noticed that recently, reboot of a 1024 cpu machine takes approx 16
> > minutes of just stopping the cpus. The slowdown was tracked to commit
> > f96972f which went into v3.7 and then to the stable trees.
> >
> > x86 does not need to be running the boot cpu to pull reset and I don't
> > think it is really needed for shutdown either.
> >
> > I decided to go the "simple" way and make this a config option that is
> > selected by the x86 arch. I don't know which other arch's would also
> > benefit, if any.
> >
> > Signed-off-by: Robin Holt <[email protected]>
> > To: Andrew Morton <[email protected]>
> > Cc: Russ Anderson <[email protected]>
> > Cc: Thomas Gleixner <[email protected]>
> > Cc: Ingo Molnar <[email protected]>
> > Cc: "H. Peter Anvin" <[email protected]>
> > Cc: Shawn Guo <[email protected]>
> > Cc: <[email protected]>
> >
> > ---
> > arch/x86/Kconfig | 3 +++
> > kernel/Kconfig.shutdown | 3 +++
> > kernel/sys.c | 4 ++++
> > 3 files changed, 10 insertions(+)
> > create mode 100644 kernel/Kconfig.shutdown
> >
> > diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> > index 70c0f3d..9611942 100644
> > --- a/arch/x86/Kconfig
> > +++ b/arch/x86/Kconfig
> > @@ -120,6 +120,7 @@ config X86
> > select OLD_SIGSUSPEND3 if X86_32 || IA32_EMULATION
> > select OLD_SIGACTION if X86_32
> > select COMPAT_OLD_SIGACTION if IA32_EMULATION
> > + select ARCH_SHUTDOWN_TO_ANY_CPU
> >
> > config INSTRUCTION_DECODER
> > def_bool y
> > @@ -839,6 +840,8 @@ config SCHED_MC
> > making when dealing with multi-core CPU chips at a cost of slightly
> > increased overhead in some places. If unsure say N here.
> >
> > +source "kernel/Kconfig.shutdown"
> > +
> > source "kernel/Kconfig.preempt"
> >
> > config X86_UP_APIC
> > diff --git a/kernel/Kconfig.shutdown b/kernel/Kconfig.shutdown
> > new file mode 100644
> > index 0000000..d79fc04
> > --- /dev/null
> > +++ b/kernel/Kconfig.shutdown
> > @@ -0,0 +1,3 @@
> > +
> > +config ARCH_SHUTDOWN_TO_ANY_CPU
> > + bool
> > diff --git a/kernel/sys.c b/kernel/sys.c
> > index 39c9c4a..c0b8880 100644
> > --- a/kernel/sys.c
> > +++ b/kernel/sys.c
> > @@ -369,7 +369,9 @@ EXPORT_SYMBOL(unregister_reboot_notifier);
> > void kernel_restart(char *cmd)
> > {
> > kernel_restart_prepare(cmd);
> > +#ifndef CONFIG_ARCH_SHUTDOWN_TO_ANY_CPU
> > disable_nonboot_cpus();
> > +#endif
> > if (!cmd)
> > printk(KERN_EMERG "Restarting system.\n");
> > else
> > @@ -413,7 +415,9 @@ void kernel_power_off(void)
> > kernel_shutdown_prepare(SYSTEM_POWER_OFF);
> > if (pm_power_off_prepare)
> > pm_power_off_prepare();
> > +#ifndef CONFIG_ARCH_SHUTDOWN_TO_ANY_CPU
> > disable_nonboot_cpus();
> > +#endif
> > syscore_shutdown();
> > printk(KERN_EMERG "Power down.\n");
> > kmsg_dump(KMSG_DUMP_POWEROFF);
>
> Hm, the 'fix' is a pretty ugly workaround that does not fix much IMHO.
>
> I think the original commit:
>
> f96972f2dc63 kernel/sys.c: call disable_nonboot_cpus() in kernel_restart()
>
> actually regressed your 1024 CPU systems, and should possibly be reverted or fixed
> in some other fashion - such as by migrating to the primary CPU (on architectures
> that require that), instead of hotplug offlining every secondary CPU on every
> architecture!
Sure. There are multiple ways to fix this.
> Alternatively, disable_nonboot_cpus() could perhaps be improved to down CPUs in
> parallel: issue the CPU-down requests to every CPU, then wait for them to complete
> - instead of the loop over every CPU?
I took a look at this. disable_nonboot_cpus() loops through all online cpus,
shutting them down one cpu thread at a time. More frustrating, it ends up
calling __stop_machine() to stop all the cpus, then loops back up to stop
the next thread. The underlying code takes a cpu bitmask, so changing
disable_nonboot_cpus() to pass in a cpu bitmask and changing _cpu_down()
to accept it allows __stop_machine() to be called just once. This change
reduced the shutdown time on a 1024 cpus system from 16 minutes down to 4.
A significant improvement, but not good enough.
The next significant bottleneck is __cpu_notify(). Tried creating worker
threads to parallelize the shutdown, but the problem is __cpu_notify() is
not thread safe. Putting a lock around it caused all the worker threads
to fight over the lock.
Wondered if __cpu_notify() needed to be called for all cpus being shut down,
and it does because the cpu_chain notifier call chain has cpu as a parameter.
So the delema is that cpu_chain notifiers need to be called on all cpus, but
cannot be done in parallel due to __cpu_notify() not being thread safe.
Spinning through the notifier chain sequentially for all cpus just takes a
long time.
The real fix would be to make the &cpu_chain notifier per cpu, or at
least thread safe, so that all the cpus being shut down could do so
in parallel. That is a significant change with ramifications on
other code.
> This would be the conceptual counter part to parallel boot up of CPUs - something
> SGI might be interested in as well?
Yes, which is why I spent some time digging into this. I can clean
up my patch for the first part. The second part needs more discussion.
> Thanks,
>
> Ingo
--
Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc [email protected]
* Robin Holt <[email protected]> wrote:
> On Mon, Apr 08, 2013 at 09:11:06AM -0700, H. Peter Anvin wrote:
> > On 04/08/2013 08:57 AM, Ingo Molnar wrote:
> > >
> > > I think the original commit:
> > >
> > > f96972f2dc63 kernel/sys.c: call disable_nonboot_cpus() in kernel_restart()
> > >
> > > actually regressed your 1024 CPU systems, and should possibly be reverted or fixed
> > > in some other fashion - such as by migrating to the primary CPU (on architectures
> > > that require that), instead of hotplug offlining every secondary CPU on every
> > > architecture!
> > >
> > > Alternatively, disable_nonboot_cpus() could perhaps be improved to down CPUs in
> > > parallel: issue the CPU-down requests to every CPU, then wait for them to complete
> > > - instead of the loop over every CPU?
> > >
> > > This would be the conceptual counter part to parallel boot up of CPUs - something
> > > SGI might be interested in as well?
> > >
> >
> > Migrating to the boot processor and then calling stop_machine() to
> > defang any other processors should be sufficient, no?
> >
> > I don't know if there is any reason to deschedule all tasks?
>
> My reading of the original commit indicated that some architecture's
> firmware needs the boot cpu to be the one initiating reboot.
>
> If that is correct, then I can not see why a stop_machine() implementation
> will not work.
>
> Since this is in generic kernel code, how can I proceed?
I think rebooting on the same CPU where we booted up is something worth having in
general, as a firmware robustness feature. (assuming the CPU in question is still
online)
We have similar constraints in the suspend code for example - some x86 firmware
breaks if suspend related ACPI calls are not done on the boot CPU ...
So how about restoring the old "just reboot, don't shut down the others" behavior,
extended with a "reboot on the CPU that booted up" reboot affinity logic?
That should fix the 1024 CPUs regression, and it should also keep those ARM
systems working - without any special casing.
Of course I'd also be entirely happy about having true parallel shutdown...
It does not have to be entirely threaded: I bet most of the shutdown latency is in
a few paranoia udelay()s or so, where some simple global lock could be dropped.
Thanks,
Ingo
On Wed, Apr 10, 2013 at 01:16:20PM +0200, Ingo Molnar wrote:
>
> * Robin Holt <[email protected]> wrote:
>
> > On Mon, Apr 08, 2013 at 09:11:06AM -0700, H. Peter Anvin wrote:
> > > On 04/08/2013 08:57 AM, Ingo Molnar wrote:
> > > >
> > > > I think the original commit:
> > > >
> > > > f96972f2dc63 kernel/sys.c: call disable_nonboot_cpus() in kernel_restart()
> > > >
> > > > actually regressed your 1024 CPU systems, and should possibly be reverted or fixed
> > > > in some other fashion - such as by migrating to the primary CPU (on architectures
> > > > that require that), instead of hotplug offlining every secondary CPU on every
> > > > architecture!
> > > >
> > > > Alternatively, disable_nonboot_cpus() could perhaps be improved to down CPUs in
> > > > parallel: issue the CPU-down requests to every CPU, then wait for them to complete
> > > > - instead of the loop over every CPU?
> > > >
> > > > This would be the conceptual counter part to parallel boot up of CPUs - something
> > > > SGI might be interested in as well?
> > > >
> > >
> > > Migrating to the boot processor and then calling stop_machine() to
> > > defang any other processors should be sufficient, no?
> > >
> > > I don't know if there is any reason to deschedule all tasks?
> >
> > My reading of the original commit indicated that some architecture's
> > firmware needs the boot cpu to be the one initiating reboot.
> >
> > If that is correct, then I can not see why a stop_machine() implementation
> > will not work.
> >
> > Since this is in generic kernel code, how can I proceed?
>
> I think rebooting on the same CPU where we booted up is something worth having in
> general, as a firmware robustness feature. (assuming the CPU in question is still
> online)
>
> We have similar constraints in the suspend code for example - some x86 firmware
> breaks if suspend related ACPI calls are not done on the boot CPU ...
>
> So how about restoring the old "just reboot, don't shut down the others" behavior,
> extended with a "reboot on the CPU that booted up" reboot affinity logic?
Just want to be sure I am going the write direction, but in the shutdown and
reboot case, you would support something like:
diff --git a/kernel/sys.c b/kernel/sys.c
index 39c9c4a..35845c5 100644
--- a/kernel/sys.c
+++ b/kernel/sys.c
@@ -358,6 +358,18 @@ int unregister_reboot_notifier(struct notifier_block *nb)
}
EXPORT_SYMBOL(unregister_reboot_notifier);
+void migrate_to_boot_cpu(void)
+{
+ cpumask_t *shutdown_cpu_mask;
+
+ shutdown_cpu_mask = kzalloc(sizeof(cpumask_t), GFP_KERNEL);
+ if (shutdown_cpu_mask) {
+ cpumask_set_cpu(0, shutdown_cpu_mask);
+ cpumask_and(shutdown_cpu_mask, shutdown_cpu_mask, cpu_online_mask);
+ set_cpus_allowed_ptr(current, shutdown_cpu_mask);
+ }
+}
+
/**
* kernel_restart - reboot the system
* @cmd: pointer to buffer containing command to execute for restart
@@ -369,7 +381,7 @@ EXPORT_SYMBOL(unregister_reboot_notifier);
void kernel_restart(char *cmd)
{
kernel_restart_prepare(cmd);
- disable_nonboot_cpus();
+ migrate_to_boot_cpu();
if (!cmd)
printk(KERN_EMERG "Restarting system.\n");
else
@@ -413,7 +425,7 @@ void kernel_power_off(void)
kernel_shutdown_prepare(SYSTEM_POWER_OFF);
if (pm_power_off_prepare)
pm_power_off_prepare();
- disable_nonboot_cpus();
+ migrate_to_boot_cpu();
syscore_shutdown();
printk(KERN_EMERG "Power down.\n");
kmsg_dump(KMSG_DUMP_POWEROFF);
On Wed, Apr 10, 2013 at 4:16 AM, Ingo Molnar <[email protected]> wrote:
>
> I think rebooting on the same CPU where we booted up is something worth having in
> general, as a firmware robustness feature. (assuming the CPU in question is still
> online)
Yeah, we've had issues with ACPI in the past, so I do think we should
always reboot using the BP. Even if it almost certainly works on 99+%
of all machines on any random CPU.
The optimal solution would be to just speed up the
disable_nonboot_cpus() code so much that it isn't an issue. That would
be good for suspending too, although I guess suspend isn't a big issue
if you have a thousand CPU's.
Has anybody checked whether we could do the cpu_down() on non-boot
CPU's in parallel? Right now we serialize the thing completely, with
one single
for_each_online_cpu(cpu) {
...
loop that does a synchrinous _cpu_down() for each CPU. No wonder it
takes forever. We do __stop_machine() over and over and over again:
the whole thing is basically O(n**2) in CPU's.
Linus
On Wed, Apr 10, 2013 at 08:10:05AM -0700, Linus Torvalds wrote:
> On Wed, Apr 10, 2013 at 4:16 AM, Ingo Molnar <[email protected]> wrote:
> >
> > I think rebooting on the same CPU where we booted up is something worth having in
> > general, as a firmware robustness feature. (assuming the CPU in question is still
> > online)
>
> Yeah, we've had issues with ACPI in the past, so I do think we should
> always reboot using the BP. Even if it almost certainly works on 99+%
> of all machines on any random CPU.
>
> The optimal solution would be to just speed up the
> disable_nonboot_cpus() code so much that it isn't an issue. That would
> be good for suspending too, although I guess suspend isn't a big issue
> if you have a thousand CPU's.
>
> Has anybody checked whether we could do the cpu_down() on non-boot
> CPU's in parallel? Right now we serialize the thing completely, with
> one single
>
> for_each_online_cpu(cpu) {
> ...
>
> loop that does a synchrinous _cpu_down() for each CPU. No wonder it
> takes forever. We do __stop_machine() over and over and over again:
> the whole thing is basically O(n**2) in CPU's.
Yes, I have a test patch that replaces for_each_online_cpu(cpu)
with a cpu bitmask in disable_nonboot_cpus(). The lower level
routines already take a bitmask. It allows __stop_machine() to
be called just once. That change reduces shutdown time on a
1024 cpu machine from 16 minutes 4 minutes. Significant improvement,
but not good enough.
The next significant bottleneck is __cpu_notify(). Tried creating
worker threads to parallelize the shutdown, but the problem is
__cpu_notify() is not thread safe. Putting a lock around it
caused all the worker threads to fight over the lock.
Note that __cpu_notify() has to be called for all cpus being
shut down because the cpu_chain notifier call chain has cpu as a
parameter. The delema is that cpu_chain notifiers need to be called on
all cpus, but cannot be done in parallel due to __cpu_notify() not being
thread safe. Spinning through the notifier chain sequentially for all
cpus just takes a long time.
The real fix would be to make the &cpu_chain notifier per cpu, or at
least thread safe, so that all the cpus being shut down could do so
in parallel. That is a significant change with ramifications on
other code.
I will post a patch shortly with the cpu bitmask change. Changing
__cpu_notify() will take more discussion.
> Linus
--
Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc [email protected]
* Russ Anderson <[email protected]> wrote:
> Yes, I have a test patch that replaces for_each_online_cpu(cpu) with a cpu
> bitmask in disable_nonboot_cpus(). The lower level routines already take a
> bitmask. It allows __stop_machine() to be called just once. That change
> reduces shutdown time on a 1024 cpu machine from 16 minutes 4 minutes.
> Significant improvement, but not good enough.
>
> The next significant bottleneck is __cpu_notify(). Tried creating worker
> threads to parallelize the shutdown, but the problem is __cpu_notify() is not
> thread safe. Putting a lock around it caused all the worker threads to fight
> over the lock.
4 minutes bootup is 240 seconds, with 1024 CPUs that's about 240 msecs per CPU.
That sounds a lot, given that unlike bootup there's not much real work to be done
during shutdown - we don't initialize anything, etc.
Maybe much of those 240 msecs are spent in some stupid udelay loop or so, which
could be made parallel?
Would it be possible to create a 'reboot but stop at the end and reactivate all
CPUs again' reboot flag, so that it can all be NMI-profiled, to see where the true
bottleneck is? A naked disable_nonboot_cpus() call in essence.
Thanks,
Ingo
On Wed, Apr 10, 2013 at 06:59:34PM +0200, Ingo Molnar wrote:
>
> * Russ Anderson <[email protected]> wrote:
>
> > Yes, I have a test patch that replaces for_each_online_cpu(cpu) with a cpu
> > bitmask in disable_nonboot_cpus(). The lower level routines already take a
> > bitmask. It allows __stop_machine() to be called just once. That change
> > reduces shutdown time on a 1024 cpu machine from 16 minutes 4 minutes.
> > Significant improvement, but not good enough.
> >
> > The next significant bottleneck is __cpu_notify(). Tried creating worker
> > threads to parallelize the shutdown, but the problem is __cpu_notify() is not
> > thread safe. Putting a lock around it caused all the worker threads to fight
> > over the lock.
>
> 4 minutes bootup is 240 seconds, with 1024 CPUs that's about 240 msecs per CPU.
>
> That sounds a lot, given that unlike bootup there's not much real work to be done
> during shutdown - we don't initialize anything, etc.
>
> Maybe much of those 240 msecs are spent in some stupid udelay loop or so, which
> could be made parallel?
>
> Would it be possible to create a 'reboot but stop at the end and reactivate all
> CPUs again' reboot flag, so that it can all be NMI-profiled, to see where the true
> bottleneck is? A naked disable_nonboot_cpus() call in essence.
What, exactly, are you proposing with the NMI profiling? Currently,
if I NMI the system, I get dump_stack() output for all cpus. Without
introducing a lock to serialize those, they stacks are really just a
jumbled mess. With a lock, things are fairly slow.
Are you proposing something other than looking at stack dumps?
Thanks,
Robin
* Robin Holt <[email protected]> wrote:
> On Wed, Apr 10, 2013 at 06:59:34PM +0200, Ingo Molnar wrote:
> >
> > * Russ Anderson <[email protected]> wrote:
> >
> > > Yes, I have a test patch that replaces for_each_online_cpu(cpu) with a cpu
> > > bitmask in disable_nonboot_cpus(). The lower level routines already take a
> > > bitmask. It allows __stop_machine() to be called just once. That change
> > > reduces shutdown time on a 1024 cpu machine from 16 minutes 4 minutes.
> > > Significant improvement, but not good enough.
> > >
> > > The next significant bottleneck is __cpu_notify(). Tried creating worker
> > > threads to parallelize the shutdown, but the problem is __cpu_notify() is not
> > > thread safe. Putting a lock around it caused all the worker threads to fight
> > > over the lock.
> >
> > 4 minutes bootup is 240 seconds, with 1024 CPUs that's about 240 msecs per CPU.
> >
> > That sounds a lot, given that unlike bootup there's not much real work to be done
> > during shutdown - we don't initialize anything, etc.
> >
> > Maybe much of those 240 msecs are spent in some stupid udelay loop or so, which
> > could be made parallel?
> >
> > Would it be possible to create a 'reboot but stop at the end and reactivate all
> > CPUs again' reboot flag, so that it can all be NMI-profiled, to see where the true
> > bottleneck is? A naked disable_nonboot_cpus() call in essence.
>
> What, exactly, are you proposing with the NMI profiling? [...]
I'm proposing to make 'reboot' overhead profilable, via a debug hack:
echo 1 > /proc/sys/kernel/magic_dont_fully_reboot_flag
perf record reboot
perf is using NMIs to profile - and since much of cpu_down() is with irqs
disabled, NMI profiling would be needed to see inside the overhead.
(Assuming the 240 msecs is CPU overhead, not waiting for some IRQ/IPI event.)
Thanks,
Ingo
On Wed, Apr 10, 2013 at 07:22:36PM +0200, Ingo Molnar wrote:
>
> * Robin Holt <[email protected]> wrote:
>
> > On Wed, Apr 10, 2013 at 06:59:34PM +0200, Ingo Molnar wrote:
> > >
> > > * Russ Anderson <[email protected]> wrote:
> > >
> > > > Yes, I have a test patch that replaces for_each_online_cpu(cpu) with a cpu
> > > > bitmask in disable_nonboot_cpus(). The lower level routines already take a
> > > > bitmask. It allows __stop_machine() to be called just once. That change
> > > > reduces shutdown time on a 1024 cpu machine from 16 minutes 4 minutes.
> > > > Significant improvement, but not good enough.
> > > >
> > > > The next significant bottleneck is __cpu_notify(). Tried creating worker
> > > > threads to parallelize the shutdown, but the problem is __cpu_notify() is not
> > > > thread safe. Putting a lock around it caused all the worker threads to fight
> > > > over the lock.
> > >
> > > 4 minutes bootup is 240 seconds, with 1024 CPUs that's about 240 msecs per CPU.
> > >
> > > That sounds a lot, given that unlike bootup there's not much real work to be done
> > > during shutdown - we don't initialize anything, etc.
> > >
> > > Maybe much of those 240 msecs are spent in some stupid udelay loop or so, which
> > > could be made parallel?
> > >
> > > Would it be possible to create a 'reboot but stop at the end and reactivate all
> > > CPUs again' reboot flag, so that it can all be NMI-profiled, to see where the true
> > > bottleneck is? A naked disable_nonboot_cpus() call in essence.
> >
> > What, exactly, are you proposing with the NMI profiling? [...]
>
> I'm proposing to make 'reboot' overhead profilable, via a debug hack:
>
> echo 1 > /proc/sys/kernel/magic_dont_fully_reboot_flag
>
> perf record reboot
>
> perf is using NMIs to profile - and since much of cpu_down() is with irqs
> disabled, NMI profiling would be needed to see inside the overhead.
>
> (Assuming the 240 msecs is CPU overhead, not waiting for some IRQ/IPI event.)
Let me give it a try.
Robin
On 04/10/2013 09:59 AM, Ingo Molnar wrote:
>
> 4 minutes bootup is 240 seconds, with 1024 CPUs that's about 240 msecs per CPU.
>
> That sounds a lot, given that unlike bootup there's not much real work to be done
> during shutdown - we don't initialize anything, etc.
>
> Maybe much of those 240 msecs are spent in some stupid udelay loop or so, which
> could be made parallel?
>
I wonder if it isn't calling stop_machine() 1024 times, each time
rendezvousing all the still-active CPUs.
-hpa
> > I'm proposing to make 'reboot' overhead profilable, via a debug hack:
> >
> > echo 1 > /proc/sys/kernel/magic_dont_fully_reboot_flag
> >
> > perf record reboot
> >
> > perf is using NMIs to profile - and since much of cpu_down() is with irqs
> > disabled, NMI profiling would be needed to see inside the overhead.
> >
> > (Assuming the 240 msecs is CPU overhead, not waiting for some IRQ/IPI event.)
I had the machine booted as 512 cpus.
I tweaked the kernel like this:
diff --git a/kernel/sys.c b/kernel/sys.c
index 39c9c4a..b42bd4f 100644
--- a/kernel/sys.c
+++ b/kernel/sys.c
@@ -368,8 +368,10 @@ EXPORT_SYMBOL(unregister_reboot_notifier);
*/
void kernel_restart(char *cmd)
{
- kernel_restart_prepare(cmd);
+ // kernel_restart_prepare(cmd);
disable_nonboot_cpus();
+ enable_nonboot_cpus();
+ return;
if (!cmd)
printk(KERN_EMERG "Restarting system.\n");
else
perf record -a /sbin/reboot -d -f -n
The top of 'perf report' has:
Events: 14M cycles
22.58% swapper [kernel.kallsyms] [k] update_cfs_rq_blocked_load
10.52% swapper [kernel.kallsyms] [k] load_balance
4.96% swapper [kernel.kallsyms] [k] ktime_get
4.12% swapper [kernel.kallsyms] [k] update_blocked_averages
3.55% swapper [kernel.kallsyms] [k] idle_cpu
1.97% swapper [kernel.kallsyms] [k] uv_read_rtc
0.98% swapper [kernel.kallsyms] [k] rcu_process_gp_end
0.84% swapper [kernel.kallsyms] [k] apic_timer_interrupt
0.84% swapper [kernel.kallsyms] [k] __lock_text_start
0.84% swapper [kernel.kallsyms] [k] _raw_spin_lock_irqsave
0.73% swapper [kernel.kallsyms] [k] native_safe_halt
0.56% swapper [kernel.kallsyms] [k] rcu_check_callbacks
0.56% swapper [kernel.kallsyms] [k] native_write_msr_safe
0.44% swapper [kernel.kallsyms] [k] cpumask_next_and
0.42% reboot [kernel.kallsyms] [k] kmem_cache_alloc_node
The perf data is 676 MB. I don't know how well it compresses, but the
lzma task has been running for a while.
Thanks,
Robin
On Wed, Apr 10, 2013 at 10:29:12AM -0500, Russ Anderson wrote:
> On Wed, Apr 10, 2013 at 08:10:05AM -0700, Linus Torvalds wrote:
> >
> > Yeah, we've had issues with ACPI in the past, so I do think we should
> > always reboot using the BP. Even if it almost certainly works on 99+%
> > of all machines on any random CPU.
> >
> > The optimal solution would be to just speed up the
> > disable_nonboot_cpus() code so much that it isn't an issue. That would
> > be good for suspending too, although I guess suspend isn't a big issue
> > if you have a thousand CPU's.
> >
> > Has anybody checked whether we could do the cpu_down() on non-boot
> > CPU's in parallel? Right now we serialize the thing completely, with
> > one single
> >
> > for_each_online_cpu(cpu) {
> > ...
> >
> > loop that does a synchrinous _cpu_down() for each CPU. No wonder it
> > takes forever. We do __stop_machine() over and over and over again:
> > the whole thing is basically O(n**2) in CPU's.
>
> Yes, I have a test patch that replaces for_each_online_cpu(cpu)
> with a cpu bitmask in disable_nonboot_cpus(). The lower level
> routines already take a bitmask. It allows __stop_machine() to
> be called just once. That change reduces shutdown time on a
> 1024 cpu machine from 16 minutes 4 minutes. Significant improvement,
> but not good enough.
>
> The next significant bottleneck is __cpu_notify(). Tried creating
> worker threads to parallelize the shutdown, but the problem is
> __cpu_notify() is not thread safe. Putting a lock around it
> caused all the worker threads to fight over the lock.
>
> Note that __cpu_notify() has to be called for all cpus being
> shut down because the cpu_chain notifier call chain has cpu as a
> parameter. The delema is that cpu_chain notifiers need to be called on
> all cpus, but cannot be done in parallel due to __cpu_notify() not being
> thread safe. Spinning through the notifier chain sequentially for all
> cpus just takes a long time.
>
> The real fix would be to make the &cpu_chain notifier per cpu, or at
> least thread safe, so that all the cpus being shut down could do so
> in parallel. That is a significant change with ramifications on
> other code.
>
> I will post a patch shortly with the cpu bitmask change. Changing
> __cpu_notify() will take more discussion.
Here is the test patch with the cpu bitmask change. It results
in calling __stop_machine() just once.
After making feedback changes I'll formally submit the patch.
---
kernel/cpu.c | 94 +++++++++++++++++++++++++++++++++++------------------------
1 file changed, 56 insertions(+), 38 deletions(-)
Index: linux/kernel/cpu.c
===================================================================
--- linux.orig/kernel/cpu.c 2013-04-10 17:16:40.084960495 -0500
+++ linux/kernel/cpu.c 2013-04-10 17:17:27.988245160 -0500
@@ -262,10 +262,11 @@ static int __ref take_cpu_down(void *_pa
}
/* Requires cpu_add_remove_lock to be held */
-static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
+static int __ref _cpu_down(const cpumask_t *cpus_to_offline, int tasks_frozen)
{
- int err, nr_calls = 0;
+ int cpu = 0, err = 0, nr_calls = 0;
void *hcpu = (void *)(long)cpu;
+ cpumask_var_t cpus_offlined;
unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
struct take_cpu_down_param tcd_param = {
.mod = mod,
@@ -278,46 +279,65 @@ static int __ref _cpu_down(unsigned int
if (!cpu_online(cpu))
return -EINVAL;
+ if (!alloc_cpumask_var(&cpus_offlined, GFP_KERNEL))
+ return -ENOMEM;
+
cpu_hotplug_begin();
+ cpumask_clear(cpus_offlined);
+ cpumask_copy(cpus_offlined, cpus_to_offline);
- err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
- if (err) {
- nr_calls--;
- __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
- printk("%s: attempt to take down CPU %u failed\n",
+ for_each_cpu_mask(cpu, *cpus_to_offline) {
+ hcpu = (void *)(long)cpu;
+ if (!cpu_online(cpu))
+ continue;
+ tcd_param.hcpu = hcpu;
+ err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
+ if (err) {
+ nr_calls--;
+ __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
+ pr_err("%s: attempt to take down CPU %u failed\n",
__func__, cpu);
- goto out_release;
+ goto out_release;
+ }
+ smpboot_park_threads(cpu);
}
- smpboot_park_threads(cpu);
- err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
+ err = __stop_machine(take_cpu_down, &tcd_param, cpus_to_offline);
if (err) {
/* CPU didn't die: tell everyone. Can't complain. */
- smpboot_unpark_threads(cpu);
- cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
+ for_each_cpu_mask(cpu, *cpus_to_offline) {
+ hcpu = (void *)(long)cpu;
+ smpboot_unpark_threads(cpu);
+ cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
+ }
goto out_release;
}
- BUG_ON(cpu_online(cpu));
/*
* The migration_call() CPU_DYING callback will have removed all
* runnable tasks from the cpu, there's only the idle task left now
* that the migration thread is done doing the stop_machine thing.
- *
- * Wait for the stop thread to go away.
*/
- while (!idle_cpu(cpu))
- cpu_relax();
+ for_each_cpu_mask(cpu, *cpus_offlined) {
+ BUG_ON(cpu_online(cpu));
- /* This actually kills the CPU. */
- __cpu_die(cpu);
-
- /* CPU is completely dead: tell everyone. Too late to complain. */
- cpu_notify_nofail(CPU_DEAD | mod, hcpu);
-
- check_for_tasks(cpu);
+ /*
+ * Wait for the stop thread to go away.
+ */
+ while (!idle_cpu(cpu))
+ cpu_relax();
+
+ /* This actually kills the CPU. */
+ __cpu_die(cpu);
+
+ /* CPU is completely dead: tell everyone. Too late to complain. */
+ hcpu = (void *)(long)cpu;
+ cpu_notify_nofail(CPU_DEAD | mod, hcpu);
+ check_for_tasks(cpu);
+ }
out_release:
+ free_cpumask_var(cpus_offlined);
cpu_hotplug_done();
if (!err)
cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
@@ -327,6 +347,7 @@ out_release:
int __ref cpu_down(unsigned int cpu)
{
int err;
+ cpumask_var_t cpumask;
cpu_maps_update_begin();
@@ -335,7 +356,11 @@ int __ref cpu_down(unsigned int cpu)
goto out;
}
- err = _cpu_down(cpu, 0);
+ if (!alloc_cpumask_var(&cpumask, GFP_KERNEL))
+ return -ENOMEM;
+ cpumask_set_cpu(cpu, cpumask);
+ err = _cpu_down(cpumask, 0);
+ free_cpumask_var(cpumask);
out:
cpu_maps_update_done();
@@ -459,7 +484,7 @@ static cpumask_var_t frozen_cpus;
int disable_nonboot_cpus(void)
{
- int cpu, first_cpu, error = 0;
+ int first_cpu, error = 0;
cpu_maps_update_begin();
first_cpu = cpumask_first(cpu_online_mask);
@@ -470,18 +495,11 @@ int disable_nonboot_cpus(void)
cpumask_clear(frozen_cpus);
printk("Disabling non-boot CPUs ...\n");
- for_each_online_cpu(cpu) {
- if (cpu == first_cpu)
- continue;
- error = _cpu_down(cpu, 1);
- if (!error)
- cpumask_set_cpu(cpu, frozen_cpus);
- else {
- printk(KERN_ERR "Error taking CPU%d down: %d\n",
- cpu, error);
- break;
- }
- }
+ cpumask_copy(frozen_cpus, cpu_online_mask);
+ cpumask_clear_cpu(first_cpu, frozen_cpus); /* all but one cpu*/
+ error = _cpu_down(frozen_cpus, 1);
+ if (error)
+ pr_err("Error %d stopping cpus\n", error);
if (!error) {
BUG_ON(num_online_cpus() > 1);
--
Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc [email protected]
On Wed, Apr 10, 2013 at 06:59:34PM +0200, Ingo Molnar wrote:
> * Russ Anderson <[email protected]> wrote:
>
> > Yes, I have a test patch that replaces for_each_online_cpu(cpu) with a cpu
> > bitmask in disable_nonboot_cpus(). The lower level routines already take a
> > bitmask. It allows __stop_machine() to be called just once. That change
> > reduces shutdown time on a 1024 cpu machine from 16 minutes 4 minutes.
> > Significant improvement, but not good enough.
> >
> > The next significant bottleneck is __cpu_notify(). Tried creating worker
> > threads to parallelize the shutdown, but the problem is __cpu_notify() is not
> > thread safe. Putting a lock around it caused all the worker threads to fight
> > over the lock.
>
> 4 minutes bootup is 240 seconds, with 1024 CPUs that's about 240 msecs per CPU.
>
> That sounds a lot, given that unlike bootup there's not much real work to be done
> during shutdown - we don't initialize anything, etc.
>
> Maybe much of those 240 msecs are spent in some stupid udelay loop or so, which
> could be made parallel?
I was hoping for a stupid udelay when I first started looking
at this code, but found nothing obvious.
The bulk of the time (after making the cpu bitmask change) is
spent in __cpu_notify(), as explained above.
> Would it be possible to create a 'reboot but stop at the end and reactivate all
> CPUs again' reboot flag, so that it can all be NMI-profiled, to see where the true
> bottleneck is? A naked disable_nonboot_cpus() call in essence.
My testing was similar. I hacked a kernel module to call
disable_nonboot_cpus() and enable_nonboot_cpus() and used
printks to narrow down the slow functions. That points
at the cpu notifier call chain. It's not clear if any
of the functions on the call chain take a long time, or
just going sequentially through the list for all cpus just
takes a long time.
--
Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc [email protected]
On Wed, Apr 10, 2013 at 08:10:05AM -0700, Linus Torvalds wrote:
> The optimal solution would be to just speed up the
> disable_nonboot_cpus() code so much that it isn't an issue. That would
> be good for suspending too, although I guess suspend isn't a big issue
> if you have a thousand CPU's.
>
> Has anybody checked whether we could do the cpu_down() on non-boot
> CPU's in parallel? Right now we serialize the thing completely, with
I thought Srivatsa S. Bhat had a patchset that did exactly that.
Srivatsa?
Paul.
* Robin Holt <[email protected]> wrote:
> I had the machine booted as 512 cpus.
> I tweaked the kernel like this:
>
> diff --git a/kernel/sys.c b/kernel/sys.c
> index 39c9c4a..b42bd4f 100644
> --- a/kernel/sys.c
> +++ b/kernel/sys.c
> @@ -368,8 +368,10 @@ EXPORT_SYMBOL(unregister_reboot_notifier);
> */
> void kernel_restart(char *cmd)
> {
> - kernel_restart_prepare(cmd);
> + // kernel_restart_prepare(cmd);
> disable_nonboot_cpus();
> + enable_nonboot_cpus();
> + return;
> if (!cmd)
> printk(KERN_EMERG "Restarting system.\n");
> else
>
> perf record -a /sbin/reboot -d -f -n
>
> The top of 'perf report' has:
> Events: 14M cycles
> 22.58% swapper [kernel.kallsyms] [k] update_cfs_rq_blocked_load
> 10.52% swapper [kernel.kallsyms] [k] load_balance
> 4.96% swapper [kernel.kallsyms] [k] ktime_get
> 4.12% swapper [kernel.kallsyms] [k] update_blocked_averages
> 3.55% swapper [kernel.kallsyms] [k] idle_cpu
> 1.97% swapper [kernel.kallsyms] [k] uv_read_rtc
> 0.98% swapper [kernel.kallsyms] [k] rcu_process_gp_end
> 0.84% swapper [kernel.kallsyms] [k] apic_timer_interrupt
> 0.84% swapper [kernel.kallsyms] [k] __lock_text_start
> 0.84% swapper [kernel.kallsyms] [k] _raw_spin_lock_irqsave
> 0.73% swapper [kernel.kallsyms] [k] native_safe_halt
> 0.56% swapper [kernel.kallsyms] [k] rcu_check_callbacks
> 0.56% swapper [kernel.kallsyms] [k] native_write_msr_safe
> 0.44% swapper [kernel.kallsyms] [k] cpumask_next_and
> 0.42% reboot [kernel.kallsyms] [k] kmem_cache_alloc_node
Ok, so it looks profilable.
The result above is not surprising: most CPUs sit in idle and don't do anything,
while the loop goes on, right?
The interesting thing to profile would be the parallel bring-down, with the
simplest global lock solution you mentioned. In that case most CPUs should be
doing 'something' all the time - maybe spinning on the lock, maybe something else,
right?
Thanks,
Ingo
> Ok, so it looks profilable.
>
> The result above is not surprising: most CPUs sit in idle and don't do anything,
> while the loop goes on, right?
>
> The interesting thing to profile would be the parallel bring-down, with the
> simplest global lock solution you mentioned. In that case most CPUs should be
> doing 'something' all the time - maybe spinning on the lock, maybe something else,
> right?
Again, mostly looks idle.
Events: 5M cycles
31.69% swapper [kernel.kallsyms] [k] update_cfs_rq_blocked_load
14.22% swapper [kernel.kallsyms] [k] load_balance
12.95% swapper [kernel.kallsyms] [k] ktime_get
4.64% swapper [kernel.kallsyms] [k] idle_cpu
3.46% swapper [kernel.kallsyms] [k] uv_read_rtc
2.26% swapper [kernel.kallsyms] [k] ktime_get_update_offsets
2.25% swapper [kernel.kallsyms] [k] rcu_check_callbacks
1.72% swapper [kernel.kallsyms] [k] _raw_spin_lock_irqsave
1.57% swapper [kernel.kallsyms] [k] native_write_msr_safe
1.53% swapper [kernel.kallsyms] [k] native_safe_halt
1.52% swapper [kernel.kallsyms] [k] apic_timer_interrupt
1.52% swapper [kernel.kallsyms] [k] update_blocked_averages
1.51% swapper [kernel.kallsyms] [k] __lock_text_start
1.48% swapper [kernel.kallsyms] [k] rcu_process_gp_end
1.40% swapper [kernel.kallsyms] [k] rcu_process_callbacks
1.19% reboot [kernel.kallsyms] [k] kmem_cache_alloc_node
0.63% swapper [kernel.kallsyms] [k] check_for_new_grace_period
0.58% swapper [kernel.kallsyms] [k] rebalance_domains
0.55% swapper [kernel.kallsyms] [k] cpumask_next_and
0.54% swapper [kernel.kallsyms] [k] __tick_nohz_idle_enter
0.53% swapper [kernel.kallsyms] [k] perf_adjust_freq_unthr_context
0.49% swapper [kernel.kallsyms] [k] _raw_spin_lock
Robin
* Robin Holt <[email protected]> wrote:
> > Ok, so it looks profilable.
> >
> > The result above is not surprising: most CPUs sit in idle and don't do anything,
> > while the loop goes on, right?
> >
> > The interesting thing to profile would be the parallel bring-down, with the
> > simplest global lock solution you mentioned. In that case most CPUs should be
> > doing 'something' all the time - maybe spinning on the lock, maybe something else,
> > right?
>
> Again, mostly looks idle.
Forgot to suggest:
perf record -a /sbin/reboot
... to capture remote CPU activity too.
> Events: 5M cycles
> 31.69% swapper [kernel.kallsyms] [k] update_cfs_rq_blocked_load
> 14.22% swapper [kernel.kallsyms] [k] load_balance
> 12.95% swapper [kernel.kallsyms] [k] ktime_get
> 4.64% swapper [kernel.kallsyms] [k] idle_cpu
> 3.46% swapper [kernel.kallsyms] [k] uv_read_rtc
> 2.26% swapper [kernel.kallsyms] [k] ktime_get_update_offsets
> 2.25% swapper [kernel.kallsyms] [k] rcu_check_callbacks
> 1.72% swapper [kernel.kallsyms] [k] _raw_spin_lock_irqsave
> 1.57% swapper [kernel.kallsyms] [k] native_write_msr_safe
> 1.53% swapper [kernel.kallsyms] [k] native_safe_halt
> 1.52% swapper [kernel.kallsyms] [k] apic_timer_interrupt
> 1.52% swapper [kernel.kallsyms] [k] update_blocked_averages
> 1.51% swapper [kernel.kallsyms] [k] __lock_text_start
> 1.48% swapper [kernel.kallsyms] [k] rcu_process_gp_end
> 1.40% swapper [kernel.kallsyms] [k] rcu_process_callbacks
> 1.19% reboot [kernel.kallsyms] [k] kmem_cache_alloc_node
> 0.63% swapper [kernel.kallsyms] [k] check_for_new_grace_period
> 0.58% swapper [kernel.kallsyms] [k] rebalance_domains
> 0.55% swapper [kernel.kallsyms] [k] cpumask_next_and
> 0.54% swapper [kernel.kallsyms] [k] __tick_nohz_idle_enter
> 0.53% swapper [kernel.kallsyms] [k] perf_adjust_freq_unthr_context
> 0.49% swapper [kernel.kallsyms] [k] _raw_spin_lock
If even perf record -a shows a mostly idle system, then the overhead must be in
sleep/wakeup latencies - for that the next step would be to figure out where all
the waiting happens, for example via call-graph context-switch profiling:
perf stat --null perf record -a -g -e sched:sched_switch /sbin/reboot
(the perf stat --null will tell us the runtime of the whole operation.)
Thanks,
Ingo
On Thu, Apr 11, 2013 at 02:00:27PM +0200, Ingo Molnar wrote:
>
> * Robin Holt <[email protected]> wrote:
>
> > > Ok, so it looks profilable.
> > >
> > > The result above is not surprising: most CPUs sit in idle and don't do anything,
> > > while the loop goes on, right?
> > >
> > > The interesting thing to profile would be the parallel bring-down, with the
> > > simplest global lock solution you mentioned. In that case most CPUs should be
> > > doing 'something' all the time - maybe spinning on the lock, maybe something else,
> > > right?
> >
> > Again, mostly looks idle.
>
> Forgot to suggest:
>
> perf record -a /sbin/reboot
I used perf record -a /sbin/reboot -f -d -n
Robin
>
> ... to capture remote CPU activity too.
>
> > Events: 5M cycles
> > 31.69% swapper [kernel.kallsyms] [k] update_cfs_rq_blocked_load
> > 14.22% swapper [kernel.kallsyms] [k] load_balance
> > 12.95% swapper [kernel.kallsyms] [k] ktime_get
> > 4.64% swapper [kernel.kallsyms] [k] idle_cpu
> > 3.46% swapper [kernel.kallsyms] [k] uv_read_rtc
> > 2.26% swapper [kernel.kallsyms] [k] ktime_get_update_offsets
> > 2.25% swapper [kernel.kallsyms] [k] rcu_check_callbacks
> > 1.72% swapper [kernel.kallsyms] [k] _raw_spin_lock_irqsave
> > 1.57% swapper [kernel.kallsyms] [k] native_write_msr_safe
> > 1.53% swapper [kernel.kallsyms] [k] native_safe_halt
> > 1.52% swapper [kernel.kallsyms] [k] apic_timer_interrupt
> > 1.52% swapper [kernel.kallsyms] [k] update_blocked_averages
> > 1.51% swapper [kernel.kallsyms] [k] __lock_text_start
> > 1.48% swapper [kernel.kallsyms] [k] rcu_process_gp_end
> > 1.40% swapper [kernel.kallsyms] [k] rcu_process_callbacks
> > 1.19% reboot [kernel.kallsyms] [k] kmem_cache_alloc_node
> > 0.63% swapper [kernel.kallsyms] [k] check_for_new_grace_period
> > 0.58% swapper [kernel.kallsyms] [k] rebalance_domains
> > 0.55% swapper [kernel.kallsyms] [k] cpumask_next_and
> > 0.54% swapper [kernel.kallsyms] [k] __tick_nohz_idle_enter
> > 0.53% swapper [kernel.kallsyms] [k] perf_adjust_freq_unthr_context
> > 0.49% swapper [kernel.kallsyms] [k] _raw_spin_lock
>
> If even perf record -a shows a mostly idle system, then the overhead must be in
> sleep/wakeup latencies - for that the next step would be to figure out where all
> the waiting happens, for example via call-graph context-switch profiling:
>
> perf stat --null perf record -a -g -e sched:sched_switch /sbin/reboot
>
> (the perf stat --null will tell us the runtime of the whole operation.)
>
> Thanks,
>
> Ingo
On Thu, Apr 11, 2013 at 07:03:58AM -0500, Robin Holt wrote:
> On Thu, Apr 11, 2013 at 02:00:27PM +0200, Ingo Molnar wrote:
> >
> > * Robin Holt <[email protected]> wrote:
> >
> > > > Ok, so it looks profilable.
> > > >
> > > > The result above is not surprising: most CPUs sit in idle and don't do anything,
> > > > while the loop goes on, right?
> > > >
> > > > The interesting thing to profile would be the parallel bring-down, with the
> > > > simplest global lock solution you mentioned. In that case most CPUs should be
> > > > doing 'something' all the time - maybe spinning on the lock, maybe something else,
> > > > right?
> > >
> > > Again, mostly looks idle.
> >
> > Forgot to suggest:
> >
> > perf record -a /sbin/reboot
>
> I used perf record -a /sbin/reboot -f -d -n
OK. Looking at Russ' patch, I understand now why it is looking idle.
We are still serially doing the DOWN_PREPARE, etc. All those other cpus
are still sitting idle.
Can we call the __cpu_down functions from an smp_call_function()?
Robin
>
> Robin
> >
> > ... to capture remote CPU activity too.
> >
> > > Events: 5M cycles
> > > 31.69% swapper [kernel.kallsyms] [k] update_cfs_rq_blocked_load
> > > 14.22% swapper [kernel.kallsyms] [k] load_balance
> > > 12.95% swapper [kernel.kallsyms] [k] ktime_get
> > > 4.64% swapper [kernel.kallsyms] [k] idle_cpu
> > > 3.46% swapper [kernel.kallsyms] [k] uv_read_rtc
> > > 2.26% swapper [kernel.kallsyms] [k] ktime_get_update_offsets
> > > 2.25% swapper [kernel.kallsyms] [k] rcu_check_callbacks
> > > 1.72% swapper [kernel.kallsyms] [k] _raw_spin_lock_irqsave
> > > 1.57% swapper [kernel.kallsyms] [k] native_write_msr_safe
> > > 1.53% swapper [kernel.kallsyms] [k] native_safe_halt
> > > 1.52% swapper [kernel.kallsyms] [k] apic_timer_interrupt
> > > 1.52% swapper [kernel.kallsyms] [k] update_blocked_averages
> > > 1.51% swapper [kernel.kallsyms] [k] __lock_text_start
> > > 1.48% swapper [kernel.kallsyms] [k] rcu_process_gp_end
> > > 1.40% swapper [kernel.kallsyms] [k] rcu_process_callbacks
> > > 1.19% reboot [kernel.kallsyms] [k] kmem_cache_alloc_node
> > > 0.63% swapper [kernel.kallsyms] [k] check_for_new_grace_period
> > > 0.58% swapper [kernel.kallsyms] [k] rebalance_domains
> > > 0.55% swapper [kernel.kallsyms] [k] cpumask_next_and
> > > 0.54% swapper [kernel.kallsyms] [k] __tick_nohz_idle_enter
> > > 0.53% swapper [kernel.kallsyms] [k] perf_adjust_freq_unthr_context
> > > 0.49% swapper [kernel.kallsyms] [k] _raw_spin_lock
> >
> > If even perf record -a shows a mostly idle system, then the overhead must be in
> > sleep/wakeup latencies - for that the next step would be to figure out where all
> > the waiting happens, for example via call-graph context-switch profiling:
> >
> > perf stat --null perf record -a -g -e sched:sched_switch /sbin/reboot
> >
> > (the perf stat --null will tell us the runtime of the whole operation.)
> >
> > Thanks,
> >
> > Ingo
* Robin Holt <[email protected]> wrote:
> On Thu, Apr 11, 2013 at 07:03:58AM -0500, Robin Holt wrote:
> > On Thu, Apr 11, 2013 at 02:00:27PM +0200, Ingo Molnar wrote:
> > >
> > > * Robin Holt <[email protected]> wrote:
> > >
> > > > > Ok, so it looks profilable.
> > > > >
> > > > > The result above is not surprising: most CPUs sit in idle and don't do anything,
> > > > > while the loop goes on, right?
> > > > >
> > > > > The interesting thing to profile would be the parallel bring-down, with the
> > > > > simplest global lock solution you mentioned. In that case most CPUs should be
> > > > > doing 'something' all the time - maybe spinning on the lock, maybe something else,
> > > > > right?
> > > >
> > > > Again, mostly looks idle.
> > >
> > > Forgot to suggest:
> > >
> > > perf record -a /sbin/reboot
> >
> > I used perf record -a /sbin/reboot -f -d -n
>
> OK. Looking at Russ' patch, I understand now why it is looking idle.
> We are still serially doing the DOWN_PREPARE, etc. All those other cpus
> are still sitting idle.
>
> Can we call the __cpu_down functions from an smp_call_function()?
I think the kthread_park() will generally schedule.
But ... whether it's an IPI or a wakeup should matter little: wakeups are IPI
based (sometimes faster, mwait based).
So the main overhead is the serial loop - if that's done in parallel, and then all
CPUs are waited for in a second loop, then much of the work can go on in parallel.
Thanks,
Ingo
On 04/11/2013 11:01 AM, Paul Mackerras wrote:
> On Wed, Apr 10, 2013 at 08:10:05AM -0700, Linus Torvalds wrote:
>> The optimal solution would be to just speed up the
>> disable_nonboot_cpus() code so much that it isn't an issue. That would
>> be good for suspending too, although I guess suspend isn't a big issue
>> if you have a thousand CPU's.
>>
>> Has anybody checked whether we could do the cpu_down() on non-boot
>> CPU's in parallel? Right now we serialize the thing completely, with
>
> I thought Srivatsa S. Bhat had a patchset that did exactly that.
> Srivatsa?
>
Thanks for the CC, Paul! Adding some more people to CC.
Actually, my patchset was about removing stop_machine() from the CPU
offline path.
http://lwn.net/Articles/538819/
And here is the performance improvement I had measured in the version
prior to that:
http://article.gmane.org/gmane.linux.kernel/1435249
I'm planning to revive this patchset after the 3.10 merge window closes,
because it depends on doing a tree-wide sweep, and I think its a little
late to do it in time for the upcoming 3.10 merge window itself.
Anyway, that's about removing stop_machine from CPU hotplug.
Coming to bulk CPU hotplug, yes, I had ideas similar to what Russ suggested.
But I believe we can do more than that.
As Russ pointed out, the notifiers are not thread-safe, so calling them
in parallel with different CPUs as arguments isn't going to work.
So, first, we can convert all the CPU hotplug notifiers to take a cpumask
instead of a single CPU. So assuming that there are 'n' notifiers in total,
the number of function calls would become n, instead of n*1024.
But that itself most likely won't give us much benefit over the for-loop
that Russ has done in his patch, because it'll simply do longer processing
in each of those 'n' notifiers, by iterating over the cpumask inside each
notifier.
Now comes the interesting thing:
Consider a notifier chain that looks like this:
Priority 0: A->B->C->D
We can't invoke say notifier callback A simultaneously on 2 CPUs with 2
different hotcpus as argument. *However*, since A, B, C, D all (more or less)
belong to different subsystems, we can call A, B, C and D in parallel on
different CPUs. They won't even serialize amongst themselves because they
take locks (if any) of different subsystems. And since they are of same
priority, the ordering (A after B or B after A) doesn't matter as well.
So with this, if we combine the idea I wrote above about giving a cpumask
to each of these notifiers to work with, we end up in this:
CPU 0 CPU 1 CPU2 ....
A(cpumask) B(cpumask) C(cpumask) ....
So, for example, the CPU_DOWN_PREPARE notification can be processed in parallel
on multiple CPUs at a time, for a given cpumask! That should definitely
give us a good speed-up.
One more thing we have to note is that, there are 4 notifiers for taking a
CPU offline:
CPU_DOWN_PREPARE
CPU_DYING
CPU_DEAD
CPU_POST_DEAD
The first can be run in parallel as mentioned above. The second is run in
parallel in the stop_machine() phase as shown in Russ' patch. But the third
and fourth set of notifications all end up running only on CPU0, which will
again slow down things.
So I suggest taking down the 1024 CPUs in multiple phases, like a binary search.
First, take 512 CPUs down, then 256 CPUs, then 128 CPUs etc. So at every bulk
CPU hotplug, we have enough online CPUs to handle the notifier load, and that
helps speed things up. Moreover, a handful of calls to stop_machine() is OK
because, stop_machine() takes progressively lesser and lesser time because
lesser CPUs are online on each iteration (and hence it reduces the
synchronization overhead of the stop-machine phase).
The only downside to this whole idea of running the notifiers of a given
priority in parallel, is error handling - if a notifier fails, it would be
troublesome to rollback I guess. But if we forget that for a moment, we can
give this idea a try!
Regards,
Srivatsa S. Bhat
For the v3.9 release, can we consider my awful patch?
Thanks,
Robin
On Thu, Apr 11, 2013 at 06:15:18PM +0530, Srivatsa S. Bhat wrote:
> On 04/11/2013 11:01 AM, Paul Mackerras wrote:
> > On Wed, Apr 10, 2013 at 08:10:05AM -0700, Linus Torvalds wrote:
> >> The optimal solution would be to just speed up the
> >> disable_nonboot_cpus() code so much that it isn't an issue. That would
> >> be good for suspending too, although I guess suspend isn't a big issue
> >> if you have a thousand CPU's.
> >>
> >> Has anybody checked whether we could do the cpu_down() on non-boot
> >> CPU's in parallel? Right now we serialize the thing completely, with
> >
> > I thought Srivatsa S. Bhat had a patchset that did exactly that.
> > Srivatsa?
> >
>
> Thanks for the CC, Paul! Adding some more people to CC.
>
> Actually, my patchset was about removing stop_machine() from the CPU
> offline path.
> http://lwn.net/Articles/538819/
>
> And here is the performance improvement I had measured in the version
> prior to that:
> http://article.gmane.org/gmane.linux.kernel/1435249
>
> I'm planning to revive this patchset after the 3.10 merge window closes,
> because it depends on doing a tree-wide sweep, and I think its a little
> late to do it in time for the upcoming 3.10 merge window itself.
>
> Anyway, that's about removing stop_machine from CPU hotplug.
>
> Coming to bulk CPU hotplug, yes, I had ideas similar to what Russ suggested.
> But I believe we can do more than that.
>
> As Russ pointed out, the notifiers are not thread-safe, so calling them
> in parallel with different CPUs as arguments isn't going to work.
>
> So, first, we can convert all the CPU hotplug notifiers to take a cpumask
> instead of a single CPU. So assuming that there are 'n' notifiers in total,
> the number of function calls would become n, instead of n*1024.
> But that itself most likely won't give us much benefit over the for-loop
> that Russ has done in his patch, because it'll simply do longer processing
> in each of those 'n' notifiers, by iterating over the cpumask inside each
> notifier.
>
> Now comes the interesting thing:
>
> Consider a notifier chain that looks like this:
> Priority 0: A->B->C->D
>
> We can't invoke say notifier callback A simultaneously on 2 CPUs with 2
> different hotcpus as argument. *However*, since A, B, C, D all (more or less)
> belong to different subsystems, we can call A, B, C and D in parallel on
> different CPUs. They won't even serialize amongst themselves because they
> take locks (if any) of different subsystems. And since they are of same
> priority, the ordering (A after B or B after A) doesn't matter as well.
>
> So with this, if we combine the idea I wrote above about giving a cpumask
> to each of these notifiers to work with, we end up in this:
>
> CPU 0 CPU 1 CPU2 ....
> A(cpumask) B(cpumask) C(cpumask) ....
>
> So, for example, the CPU_DOWN_PREPARE notification can be processed in parallel
> on multiple CPUs at a time, for a given cpumask! That should definitely
> give us a good speed-up.
>
> One more thing we have to note is that, there are 4 notifiers for taking a
> CPU offline:
>
> CPU_DOWN_PREPARE
> CPU_DYING
> CPU_DEAD
> CPU_POST_DEAD
>
> The first can be run in parallel as mentioned above. The second is run in
> parallel in the stop_machine() phase as shown in Russ' patch. But the third
> and fourth set of notifications all end up running only on CPU0, which will
> again slow down things.
>
> So I suggest taking down the 1024 CPUs in multiple phases, like a binary search.
> First, take 512 CPUs down, then 256 CPUs, then 128 CPUs etc. So at every bulk
> CPU hotplug, we have enough online CPUs to handle the notifier load, and that
> helps speed things up. Moreover, a handful of calls to stop_machine() is OK
> because, stop_machine() takes progressively lesser and lesser time because
> lesser CPUs are online on each iteration (and hence it reduces the
> synchronization overhead of the stop-machine phase).
>
> The only downside to this whole idea of running the notifiers of a given
> priority in parallel, is error handling - if a notifier fails, it would be
> troublesome to rollback I guess. But if we forget that for a moment, we can
> give this idea a try!
>
> Regards,
> Srivatsa S. Bhat
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
On Thu, Apr 11, 2013 at 06:15:18PM +0530, Srivatsa S. Bhat wrote:
> On 04/11/2013 11:01 AM, Paul Mackerras wrote:
> > On Wed, Apr 10, 2013 at 08:10:05AM -0700, Linus Torvalds wrote:
> >> The optimal solution would be to just speed up the
> >> disable_nonboot_cpus() code so much that it isn't an issue. That would
> >> be good for suspending too, although I guess suspend isn't a big issue
> >> if you have a thousand CPU's.
> >>
> >> Has anybody checked whether we could do the cpu_down() on non-boot
> >> CPU's in parallel? Right now we serialize the thing completely, with
> >
> > I thought Srivatsa S. Bhat had a patchset that did exactly that.
> > Srivatsa?
> >
>
> Thanks for the CC, Paul! Adding some more people to CC.
>
> Actually, my patchset was about removing stop_machine() from the CPU
> offline path.
> http://lwn.net/Articles/538819/
I certainly agree with the intent.
> And here is the performance improvement I had measured in the version
> prior to that:
> http://article.gmane.org/gmane.linux.kernel/1435249
>
> I'm planning to revive this patchset after the 3.10 merge window closes,
> because it depends on doing a tree-wide sweep, and I think its a little
> late to do it in time for the upcoming 3.10 merge window itself.
>
> Anyway, that's about removing stop_machine from CPU hotplug.
>
> Coming to bulk CPU hotplug, yes, I had ideas similar to what Russ suggested.
> But I believe we can do more than that.
>
> As Russ pointed out, the notifiers are not thread-safe, so calling them
> in parallel with different CPUs as arguments isn't going to work.
>
> So, first, we can convert all the CPU hotplug notifiers to take a cpumask
> instead of a single CPU. So assuming that there are 'n' notifiers in total,
> the number of function calls would become n, instead of n*1024.
> But that itself most likely won't give us much benefit over the for-loop
> that Russ has done in his patch, because it'll simply do longer processing
> in each of those 'n' notifiers, by iterating over the cpumask inside each
> notifier.
As an alternative, how about each cpu have their own notifier list?
Then one task per cpu can spin through that cpu's notifier list,
allowing them to run in parallel.
I don't know if that would be a faster solution than adding cpumask
to notifiers, but it my guess is it may.
> Now comes the interesting thing:
>
> Consider a notifier chain that looks like this:
> Priority 0: A->B->C->D
>
> We can't invoke say notifier callback A simultaneously on 2 CPUs with 2
> different hotcpus as argument. *However*, since A, B, C, D all (more or less)
> belong to different subsystems, we can call A, B, C and D in parallel on
> different CPUs. They won't even serialize amongst themselves because they
> take locks (if any) of different subsystems. And since they are of same
> priority, the ordering (A after B or B after A) doesn't matter as well.
>
> So with this, if we combine the idea I wrote above about giving a cpumask
> to each of these notifiers to work with, we end up in this:
>
> CPU 0 CPU 1 CPU2 ....
> A(cpumask) B(cpumask) C(cpumask) ....
>
> So, for example, the CPU_DOWN_PREPARE notification can be processed in parallel
> on multiple CPUs at a time, for a given cpumask! That should definitely
> give us a good speed-up.
>
> One more thing we have to note is that, there are 4 notifiers for taking a
> CPU offline:
>
> CPU_DOWN_PREPARE
> CPU_DYING
> CPU_DEAD
> CPU_POST_DEAD
>
> The first can be run in parallel as mentioned above. The second is run in
> parallel in the stop_machine() phase as shown in Russ' patch. But the third
> and fourth set of notifications all end up running only on CPU0, which will
> again slow down things.
In my testing the third and fourth set were a small part of the overall
time. Less than 10%, with cpu notifiers 90+% of the time. So you may
not need the added complexity, or at least fix the cpu notifier part
first.
> So I suggest taking down the 1024 CPUs in multiple phases, like a binary search.
> First, take 512 CPUs down, then 256 CPUs, then 128 CPUs etc. So at every bulk
> CPU hotplug, we have enough online CPUs to handle the notifier load, and that
> helps speed things up. Moreover, a handful of calls to stop_machine() is OK
> because, stop_machine() takes progressively lesser and lesser time because
> lesser CPUs are online on each iteration (and hence it reduces the
> synchronization overhead of the stop-machine phase).
>
> The only downside to this whole idea of running the notifiers of a given
> priority in parallel, is error handling - if a notifier fails, it would be
> troublesome to rollback I guess. But if we forget that for a moment, we can
> give this idea a try!
Yes.
> Regards,
> Srivatsa S. Bhat
--
Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc [email protected]
On 04/11/2013 07:53 PM, Russ Anderson wrote:
> On Thu, Apr 11, 2013 at 06:15:18PM +0530, Srivatsa S. Bhat wrote:
>> On 04/11/2013 11:01 AM, Paul Mackerras wrote:
>>> On Wed, Apr 10, 2013 at 08:10:05AM -0700, Linus Torvalds wrote:
>>>> The optimal solution would be to just speed up the
>>>> disable_nonboot_cpus() code so much that it isn't an issue. That would
>>>> be good for suspending too, although I guess suspend isn't a big issue
>>>> if you have a thousand CPU's.
>>>>
>>>> Has anybody checked whether we could do the cpu_down() on non-boot
>>>> CPU's in parallel? Right now we serialize the thing completely, with
>>>
>>> I thought Srivatsa S. Bhat had a patchset that did exactly that.
>>> Srivatsa?
>>>
>>
>> Thanks for the CC, Paul! Adding some more people to CC.
>>
>> Actually, my patchset was about removing stop_machine() from the CPU
>> offline path.
>> http://lwn.net/Articles/538819/
>
> I certainly agree with the intent.
>
Thank you!
>> And here is the performance improvement I had measured in the version
>> prior to that:
>> http://article.gmane.org/gmane.linux.kernel/1435249
>>
>> I'm planning to revive this patchset after the 3.10 merge window closes,
>> because it depends on doing a tree-wide sweep, and I think its a little
>> late to do it in time for the upcoming 3.10 merge window itself.
>>
>> Anyway, that's about removing stop_machine from CPU hotplug.
>>
>> Coming to bulk CPU hotplug, yes, I had ideas similar to what Russ suggested.
>> But I believe we can do more than that.
>>
>> As Russ pointed out, the notifiers are not thread-safe, so calling them
>> in parallel with different CPUs as arguments isn't going to work.
>>
>> So, first, we can convert all the CPU hotplug notifiers to take a cpumask
>> instead of a single CPU. So assuming that there are 'n' notifiers in total,
>> the number of function calls would become n, instead of n*1024.
>> But that itself most likely won't give us much benefit over the for-loop
>> that Russ has done in his patch, because it'll simply do longer processing
>> in each of those 'n' notifiers, by iterating over the cpumask inside each
>> notifier.
>
> As an alternative, how about each cpu have their own notifier list?
> Then one task per cpu can spin through that cpu's notifier list,
> allowing them to run in parallel.
>
> I don't know if that would be a faster solution than adding cpumask
> to notifiers, but it my guess is it may.
>
That might not work out well because those notifiers will have to lock
against each other. That is, notifier callback A cannot run as A(cpuX)
and A(cpuY) in parallel. They will have to serialize themselves, which
will make the whole effort useless. But, as I mentioned earlier, A(cpuX)
and B(cpuX) can run in parallel without additional serialization, if A
and B are completely different callbacks (ie., belonging to different
subsystems).
>> Now comes the interesting thing:
>>
>> Consider a notifier chain that looks like this:
>> Priority 0: A->B->C->D
>>
>> We can't invoke say notifier callback A simultaneously on 2 CPUs with 2
>> different hotcpus as argument. *However*, since A, B, C, D all (more or less)
>> belong to different subsystems, we can call A, B, C and D in parallel on
>> different CPUs. They won't even serialize amongst themselves because they
>> take locks (if any) of different subsystems. And since they are of same
>> priority, the ordering (A after B or B after A) doesn't matter as well.
>>
>> So with this, if we combine the idea I wrote above about giving a cpumask
>> to each of these notifiers to work with, we end up in this:
>>
>> CPU 0 CPU 1 CPU2 ....
>> A(cpumask) B(cpumask) C(cpumask) ....
>>
>> So, for example, the CPU_DOWN_PREPARE notification can be processed in parallel
>> on multiple CPUs at a time, for a given cpumask! That should definitely
>> give us a good speed-up.
>>
>> One more thing we have to note is that, there are 4 notifiers for taking a
>> CPU offline:
>>
>> CPU_DOWN_PREPARE
>> CPU_DYING
>> CPU_DEAD
>> CPU_POST_DEAD
>>
>> The first can be run in parallel as mentioned above. The second is run in
>> parallel in the stop_machine() phase as shown in Russ' patch. But the third
>> and fourth set of notifications all end up running only on CPU0, which will
>> again slow down things.
>
> In my testing the third and fourth set were a small part of the overall
> time. Less than 10%, with cpu notifiers 90+% of the time.
*All* of them are cpu notifiers! All of them invoke __cpu_notify() internally.
So how did you differentiate between them and find out that the third and
fourth sets take less time?
> So you may
> not need the added complexity, or at least fix the cpu notifier part
> first.
>
To make the 3rd and 4th run fast, the only thing we need to do is take CPUs
offline in smaller steps, like 512, 256 etc.. It doesn't add any extra
complexity over and above what is necessary to make the cpu notifiers run
in parallel in the first place.
Regards,
Srivatsa S. Bhat
On Thu, Apr 11, 2013 at 08:15:27PM +0530, Srivatsa S. Bhat wrote:
> On 04/11/2013 07:53 PM, Russ Anderson wrote:
> > On Thu, Apr 11, 2013 at 06:15:18PM +0530, Srivatsa S. Bhat wrote:
> >>
> >> One more thing we have to note is that, there are 4 notifiers for taking a
> >> CPU offline:
> >>
> >> CPU_DOWN_PREPARE
> >> CPU_DYING
> >> CPU_DEAD
> >> CPU_POST_DEAD
> >>
> >> The first can be run in parallel as mentioned above. The second is run in
> >> parallel in the stop_machine() phase as shown in Russ' patch. But the third
> >> and fourth set of notifications all end up running only on CPU0, which will
> >> again slow down things.
> >
> > In my testing the third and fourth set were a small part of the overall
> > time. Less than 10%, with cpu notifiers 90+% of the time.
>
> *All* of them are cpu notifiers! All of them invoke __cpu_notify() internally.
> So how did you differentiate between them and find out that the third and
> fourth sets take less time?
I reran a test on a 1024 cpu system, using my test patch to only call
__stop_machine() once. Added printks to show the kernel timestamp
at various points.
When calling disable_nonboot_cpus() and enable_nonboot_cpus() just after
booting the system:
The loop calling __cpu_notify(CPU_DOWN_PREPARE) took 376.6 seconds.
The loop calling cpu_notify_nofail(CPU_DEAD) took 8.1 seconds.
My guess is that notifiers do more work in the CPU_DOWN_PREPARE case.
I also added a loop calling a new notifier (CPU_TEST) which none of
notifiers would recognize, to measure the time it took to spin through
the call chain without the notifiers doing any work. It took
0.0067 seconds.
On the actual reboot, as the system was shutting down:
The loop calling __cpu_notify(CPU_DOWN_PREPARE) took 333.8 seconds.
The loop calling cpu_notify_nofail(CPU_DEAD) took 2.7 seconds.
I don't know how many notifiers are on the chain, or if there is
one heavy hitter accounting for much of the time in the
CPU_DOWN_PREPARE case.
FWIW, the overall cpu stop times are somewhat longer than what I
measured before. Not sure if the difference is due to changes in
my test patch, other kernel changes pulled in, or some difference
on the test system.
> > So you may
> > not need the added complexity, or at least fix the cpu notifier part
> > first.
> >
>
> To make the 3rd and 4th run fast, the only thing we need to do is take CPUs
> offline in smaller steps, like 512, 256 etc.. It doesn't add any extra
> complexity over and above what is necessary to make the cpu notifiers run
> in parallel in the first place.
>
> Regards,
> Srivatsa S. Bhat
--
Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc [email protected]
On 04/12/2013 01:38 AM, Russ Anderson wrote:
> On Thu, Apr 11, 2013 at 08:15:27PM +0530, Srivatsa S. Bhat wrote:
>> On 04/11/2013 07:53 PM, Russ Anderson wrote:
>>> On Thu, Apr 11, 2013 at 06:15:18PM +0530, Srivatsa S. Bhat wrote:
>>>>
>>>> One more thing we have to note is that, there are 4 notifiers for taking a
>>>> CPU offline:
>>>>
>>>> CPU_DOWN_PREPARE
>>>> CPU_DYING
>>>> CPU_DEAD
>>>> CPU_POST_DEAD
>>>>
>>>> The first can be run in parallel as mentioned above. The second is run in
>>>> parallel in the stop_machine() phase as shown in Russ' patch. But the third
>>>> and fourth set of notifications all end up running only on CPU0, which will
>>>> again slow down things.
>>>
>>> In my testing the third and fourth set were a small part of the overall
>>> time. Less than 10%, with cpu notifiers 90+% of the time.
>>
>> *All* of them are cpu notifiers! All of them invoke __cpu_notify() internally.
>> So how did you differentiate between them and find out that the third and
>> fourth sets take less time?
>
> I reran a test on a 1024 cpu system, using my test patch to only call
> __stop_machine() once. Added printks to show the kernel timestamp
> at various points.
>
> When calling disable_nonboot_cpus() and enable_nonboot_cpus() just after
> booting the system:
> The loop calling __cpu_notify(CPU_DOWN_PREPARE) took 376.6 seconds.
> The loop calling cpu_notify_nofail(CPU_DEAD) took 8.1 seconds.
>
> My guess is that notifiers do more work in the CPU_DOWN_PREPARE case.
>
> I also added a loop calling a new notifier (CPU_TEST) which none of
> notifiers would recognize, to measure the time it took to spin through
> the call chain without the notifiers doing any work. It took
> 0.0067 seconds.
>
> On the actual reboot, as the system was shutting down:
> The loop calling __cpu_notify(CPU_DOWN_PREPARE) took 333.8 seconds.
> The loop calling cpu_notify_nofail(CPU_DEAD) took 2.7 seconds.
>
> I don't know how many notifiers are on the chain, or if there is
> one heavy hitter accounting for much of the time in the
> CPU_DOWN_PREPARE case.
>
>
> FWIW, the overall cpu stop times are somewhat longer than what I
> measured before. Not sure if the difference is due to changes in
> my test patch, other kernel changes pulled in, or some difference
> on the test system.
>
>
Thanks a lot for reporting the time taken at each stage. Its extremely
useful. So, we can drop the idea of taking CPUs down in multiple rounds
like 512, 256 etc. And, like you mentioned earlier, just running the
CPU_DOWN_PREPARE notifiers in parallel (like we discussed earlier) should
give us all the performance improvement. Or perhaps, we can instrument
the code in kernel/notifier.c (notifier_call_chain) to find out if there
is a rogue notifier which contributes most to the ~300 seconds.
Regards,
Srivatsa S. Bhat
On Thu, Apr 11, 2013 at 03:08:20PM -0500, Russ Anderson wrote:
> On Thu, Apr 11, 2013 at 08:15:27PM +0530, Srivatsa S. Bhat wrote:
> > On 04/11/2013 07:53 PM, Russ Anderson wrote:
> > > On Thu, Apr 11, 2013 at 06:15:18PM +0530, Srivatsa S. Bhat wrote:
> > >>
> > >> One more thing we have to note is that, there are 4 notifiers for taking a
> > >> CPU offline:
> > >>
> > >> CPU_DOWN_PREPARE
> > >> CPU_DYING
> > >> CPU_DEAD
> > >> CPU_POST_DEAD
> > >>
> > >> The first can be run in parallel as mentioned above. The second is run in
> > >> parallel in the stop_machine() phase as shown in Russ' patch. But the third
> > >> and fourth set of notifications all end up running only on CPU0, which will
> > >> again slow down things.
> > >
> > > In my testing the third and fourth set were a small part of the overall
> > > time. Less than 10%, with cpu notifiers 90+% of the time.
> >
> > *All* of them are cpu notifiers! All of them invoke __cpu_notify() internally.
> > So how did you differentiate between them and find out that the third and
> > fourth sets take less time?
>
> I reran a test on a 1024 cpu system, using my test patch to only call
> __stop_machine() once. Added printks to show the kernel timestamp
> at various points.
>
> When calling disable_nonboot_cpus() and enable_nonboot_cpus() just after
> booting the system:
> The loop calling __cpu_notify(CPU_DOWN_PREPARE) took 376.6 seconds.
> The loop calling cpu_notify_nofail(CPU_DEAD) took 8.1 seconds.
>
> My guess is that notifiers do more work in the CPU_DOWN_PREPARE case.
>
> I also added a loop calling a new notifier (CPU_TEST) which none of
> notifiers would recognize, to measure the time it took to spin through
> the call chain without the notifiers doing any work. It took
> 0.0067 seconds.
>
> On the actual reboot, as the system was shutting down:
> The loop calling __cpu_notify(CPU_DOWN_PREPARE) took 333.8 seconds.
> The loop calling cpu_notify_nofail(CPU_DEAD) took 2.7 seconds.
How about if you take the notifier_call_chain function copy it
to kernel/sys.c, and time each notifier_call() callout individually.
Robin
* Robin Holt <[email protected]> wrote:
> For the v3.9 release, can we consider my awful patch?
How about trying what I suggested, to make reboot affine to the boot CPU
explicitly, not by shutting down all the other CPUs, but by set_cpus_allowed() or
so?
That should solve the regression, without the ugly special-casing - while giving
time to address the hot-unplug performance bottleneck.
Once that is done disable_nonboot_cpus() can be used again for reboot.
(But no strong feelings either way - both solutions are a workaround in a sense.)
Thanks,
Ingo
On 04/12/2013 11:07 AM, Ingo Molnar wrote:
>
> * Robin Holt <[email protected]> wrote:
>
>> For the v3.9 release, can we consider my awful patch?
>
> How about trying what I suggested, to make reboot affine to the boot CPU
> explicitly, not by shutting down all the other CPUs, but by set_cpus_allowed() or
> so?
>
I agree, that sounds like the right thing to do for 3.9. Of course, it would be
nice if Shawn could verify that doing that doesn't break his platform due to
some unknown corner case.
> That should solve the regression, without the ugly special-casing - while giving
> time to address the hot-unplug performance bottleneck.
>
> Once that is done disable_nonboot_cpus() can be used again for reboot.
>
> (But no strong feelings either way - both solutions are a workaround in a sense.)
>
Regards,
Srivatsa S. Bhat
On Fri, Apr 12, 2013 at 11:39:51AM +0530, Srivatsa S. Bhat wrote:
> On 04/12/2013 11:07 AM, Ingo Molnar wrote:
> >
> > * Robin Holt <[email protected]> wrote:
> >
> >> For the v3.9 release, can we consider my awful patch?
> >
> > How about trying what I suggested, to make reboot affine to the boot CPU
> > explicitly, not by shutting down all the other CPUs, but by set_cpus_allowed() or
> > so?
> >
>
> I agree, that sounds like the right thing to do for 3.9. Of course, it would be
> nice if Shawn could verify that doing that doesn't break his platform due to
> some unknown corner case.
>
> > That should solve the regression, without the ugly special-casing - while giving
> > time to address the hot-unplug performance bottleneck.
> >
> > Once that is done disable_nonboot_cpus() can be used again for reboot.
> >
> > (But no strong feelings either way - both solutions are a workaround in a sense.)
>From 1767003c943325e52ac78cac6fdbaf2ab638888d Mon Sep 17 00:00:00 2001
From: Robin Holt <[email protected]>
Date: Wed, 3 Apr 2013 13:52:00 -0500
Subject: [PATCH] Migrate shutdown/reboot to boot cpu.
We recently noticed that reboot of a 1024 cpu machine takes approx 16
minutes of just stopping the cpus. The slowdown was tracked to commit
f96972f.
The current implementation does all the work of hot removing the cpus
before halting the system. We are switching to just migrating to the
boot cpu and then calling continuing with shutdown/reboot.
This also has the effect of not breaking x86's command line parameter for
specifying the reboot cpu. Note, this code was shamelessly copied from
arch/x86/kernel/reboot.c with bits removed pertaining to the reboot_cpu
command line parameter.
Signed-off-by: Robin Holt <[email protected]>
To: Shawn Guo <[email protected]>
To: Ingo Molnar <[email protected]>
To: Russ Anderson <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: "H. Peter Anvin" <[email protected]>
Cc: Lai Jiangshan <[email protected]>
Cc: Linus Torvalds <[email protected]>
Cc: Linux Kernel Mailing List <[email protected]>
Cc: Michel Lespinasse <[email protected]>
Cc: Oleg Nesterov <[email protected]>
Cc: "Paul E. McKenney" <[email protected]>
Cc: Paul Mackerras <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Robin Holt <[email protected]>
Cc: "[email protected]" <[email protected]>
Cc: Tejun Heo <[email protected]>
Cc: the arch/x86 maintainers <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: <[email protected]>
---
kernel/sys.c | 17 +++++++++++++++--
1 file changed, 15 insertions(+), 2 deletions(-)
diff --git a/kernel/sys.c b/kernel/sys.c
index 0da73cf..4d1047d 100644
--- a/kernel/sys.c
+++ b/kernel/sys.c
@@ -357,6 +357,19 @@ int unregister_reboot_notifier(struct notifier_block *nb)
}
EXPORT_SYMBOL(unregister_reboot_notifier);
+void migrate_to_boot_cpu(void)
+{
+ /* The boot cpu is always logical cpu 0 */
+ int reboot_cpu_id = 0;
+
+ /* Make certain the cpu I'm about to reboot on is online */
+ if (!cpu_online(reboot_cpu_id))
+ reboot_cpu_id = smp_processor_id();
+
+ /* Make certain I only run on the appropriate processor */
+ set_cpus_allowed_ptr(current, cpumask_of(reboot_cpu_id));
+}
+
/**
* kernel_restart - reboot the system
* @cmd: pointer to buffer containing command to execute for restart
@@ -368,7 +381,7 @@ EXPORT_SYMBOL(unregister_reboot_notifier);
void kernel_restart(char *cmd)
{
kernel_restart_prepare(cmd);
- disable_nonboot_cpus();
+ migrate_to_boot_cpu();
syscore_shutdown();
if (!cmd)
printk(KERN_EMERG "Restarting system.\n");
@@ -414,7 +427,7 @@ void kernel_power_off(void)
kernel_shutdown_prepare(SYSTEM_POWER_OFF);
if (pm_power_off_prepare)
pm_power_off_prepare();
- disable_nonboot_cpus();
+ migrate_to_boot_cpu();
syscore_shutdown();
printk(KERN_EMERG "Power down.\n");
kmsg_dump(KMSG_DUMP_POWEROFF);
--
1.8.1.2
Meant to send this to Shawn. Too early in the morning.
Robin
On Fri, Apr 12, 2013 at 04:31:49AM -0500, Robin Holt wrote:
> On Fri, Apr 12, 2013 at 11:39:51AM +0530, Srivatsa S. Bhat wrote:
> > On 04/12/2013 11:07 AM, Ingo Molnar wrote:
> > >
> > > * Robin Holt <[email protected]> wrote:
> > >
> > >> For the v3.9 release, can we consider my awful patch?
> > >
> > > How about trying what I suggested, to make reboot affine to the boot CPU
> > > explicitly, not by shutting down all the other CPUs, but by set_cpus_allowed() or
> > > so?
> > >
> >
> > I agree, that sounds like the right thing to do for 3.9. Of course, it would be
> > nice if Shawn could verify that doing that doesn't break his platform due to
> > some unknown corner case.
> >
> > > That should solve the regression, without the ugly special-casing - while giving
> > > time to address the hot-unplug performance bottleneck.
> > >
> > > Once that is done disable_nonboot_cpus() can be used again for reboot.
> > >
> > > (But no strong feelings either way - both solutions are a workaround in a sense.)
>
>
> >From 1767003c943325e52ac78cac6fdbaf2ab638888d Mon Sep 17 00:00:00 2001
> From: Robin Holt <[email protected]>
> Date: Wed, 3 Apr 2013 13:52:00 -0500
> Subject: [PATCH] Migrate shutdown/reboot to boot cpu.
>
> We recently noticed that reboot of a 1024 cpu machine takes approx 16
> minutes of just stopping the cpus. The slowdown was tracked to commit
> f96972f.
>
> The current implementation does all the work of hot removing the cpus
> before halting the system. We are switching to just migrating to the
> boot cpu and then calling continuing with shutdown/reboot.
>
> This also has the effect of not breaking x86's command line parameter for
> specifying the reboot cpu. Note, this code was shamelessly copied from
> arch/x86/kernel/reboot.c with bits removed pertaining to the reboot_cpu
> command line parameter.
>
> Signed-off-by: Robin Holt <[email protected]>
> To: Shawn Guo <[email protected]>
> To: Ingo Molnar <[email protected]>
> To: Russ Anderson <[email protected]>
> Cc: Andrew Morton <[email protected]>
> Cc: "H. Peter Anvin" <[email protected]>
> Cc: Lai Jiangshan <[email protected]>
> Cc: Linus Torvalds <[email protected]>
> Cc: Linux Kernel Mailing List <[email protected]>
> Cc: Michel Lespinasse <[email protected]>
> Cc: Oleg Nesterov <[email protected]>
> Cc: "Paul E. McKenney" <[email protected]>
> Cc: Paul Mackerras <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Cc: Robin Holt <[email protected]>
> Cc: "[email protected]" <[email protected]>
> Cc: Tejun Heo <[email protected]>
> Cc: the arch/x86 maintainers <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: <[email protected]>
> ---
> kernel/sys.c | 17 +++++++++++++++--
> 1 file changed, 15 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sys.c b/kernel/sys.c
> index 0da73cf..4d1047d 100644
> --- a/kernel/sys.c
> +++ b/kernel/sys.c
> @@ -357,6 +357,19 @@ int unregister_reboot_notifier(struct notifier_block *nb)
> }
> EXPORT_SYMBOL(unregister_reboot_notifier);
>
> +void migrate_to_boot_cpu(void)
> +{
> + /* The boot cpu is always logical cpu 0 */
> + int reboot_cpu_id = 0;
> +
> + /* Make certain the cpu I'm about to reboot on is online */
> + if (!cpu_online(reboot_cpu_id))
> + reboot_cpu_id = smp_processor_id();
> +
> + /* Make certain I only run on the appropriate processor */
> + set_cpus_allowed_ptr(current, cpumask_of(reboot_cpu_id));
> +}
> +
> /**
> * kernel_restart - reboot the system
> * @cmd: pointer to buffer containing command to execute for restart
> @@ -368,7 +381,7 @@ EXPORT_SYMBOL(unregister_reboot_notifier);
> void kernel_restart(char *cmd)
> {
> kernel_restart_prepare(cmd);
> - disable_nonboot_cpus();
> + migrate_to_boot_cpu();
> syscore_shutdown();
> if (!cmd)
> printk(KERN_EMERG "Restarting system.\n");
> @@ -414,7 +427,7 @@ void kernel_power_off(void)
> kernel_shutdown_prepare(SYSTEM_POWER_OFF);
> if (pm_power_off_prepare)
> pm_power_off_prepare();
> - disable_nonboot_cpus();
> + migrate_to_boot_cpu();
> syscore_shutdown();
> printk(KERN_EMERG "Power down.\n");
> kmsg_dump(KMSG_DUMP_POWEROFF);
> --
> 1.8.1.2
On 04/12, Robin Holt wrote:
>
> +void migrate_to_boot_cpu(void)
> +{
> + /* The boot cpu is always logical cpu 0 */
> + int reboot_cpu_id = 0;
> +
> + /* Make certain the cpu I'm about to reboot on is online */
> + if (!cpu_online(reboot_cpu_id))
> + reboot_cpu_id = smp_processor_id();
> +
> + /* Make certain I only run on the appropriate processor */
> + set_cpus_allowed_ptr(current, cpumask_of(reboot_cpu_id));
This is only theoretical, but perhaps it makes sense to set
PF_THREAD_BOUND before set_cpus_allowed_ptr() ? To prevent the
race with another thread doing sched_setaffinity().
Oleg.
On 04/12/2013 03:01 PM, Robin Holt wrote:
> kernel/sys.c | 17 +++++++++++++++--
> 1 file changed, 15 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sys.c b/kernel/sys.c
> index 0da73cf..4d1047d 100644
> --- a/kernel/sys.c
> +++ b/kernel/sys.c
> @@ -357,6 +357,19 @@ int unregister_reboot_notifier(struct notifier_block *nb)
> }
> EXPORT_SYMBOL(unregister_reboot_notifier);
>
> +void migrate_to_boot_cpu(void)
> +{
> + /* The boot cpu is always logical cpu 0 */
> + int reboot_cpu_id = 0;
> +
> + /* Make certain the cpu I'm about to reboot on is online */
> + if (!cpu_online(reboot_cpu_id))
> + reboot_cpu_id = smp_processor_id();
> +
If CPU 0 is offline, there is no point in binding, right?
[Fenghua (in CC) added the support to offline CPU0 on x86 Intel platforms.
So its possible that CPU0 is offline when you try a reboot.]
> + /* Make certain I only run on the appropriate processor */
> + set_cpus_allowed_ptr(current, cpumask_of(reboot_cpu_id));
> +}
> +
> /**
> * kernel_restart - reboot the system
> * @cmd: pointer to buffer containing command to execute for restart
> @@ -368,7 +381,7 @@ EXPORT_SYMBOL(unregister_reboot_notifier);
> void kernel_restart(char *cmd)
> {
> kernel_restart_prepare(cmd);
> - disable_nonboot_cpus();
> + migrate_to_boot_cpu();
> syscore_shutdown();
> if (!cmd)
> printk(KERN_EMERG "Restarting system.\n");
> @@ -414,7 +427,7 @@ void kernel_power_off(void)
> kernel_shutdown_prepare(SYSTEM_POWER_OFF);
> if (pm_power_off_prepare)
> pm_power_off_prepare();
> - disable_nonboot_cpus();
> + migrate_to_boot_cpu();
Okay, so you are touching poweroff also. Restart was only recently altered
by Shawn, so we can assume that his fix was necessary only to his platform.
However, for poweroff, I see the commit below in the git log, which added
the disable_nonboot_cpus() call.
commit 4047727e5ae33f9b8d2b7766d1994ea6e5ec2991
Author: Mark Lord <[email protected]>
Date: Mon Oct 1 01:20:10 2007 -0700
Fix SMP poweroff hangs
Its an old commit, so perhaps the issue no longer holds good, but I thought
I should bring this to notice, just in case.
> syscore_shutdown();
> printk(KERN_EMERG "Power down.\n");
> kmsg_dump(KMSG_DUMP_POWEROFF);
>
Regards,
Srivatsa S. Bhat
* Robin Holt <[email protected]> wrote:
> From 1767003c943325e52ac78cac6fdbaf2ab638888d Mon Sep 17 00:00:00 2001
> From: Robin Holt <[email protected]>
> Date: Wed, 3 Apr 2013 13:52:00 -0500
> Subject: [PATCH] Migrate shutdown/reboot to boot cpu.
>
> We recently noticed that reboot of a 1024 cpu machine takes approx 16
> minutes of just stopping the cpus. The slowdown was tracked to commit
> f96972f.
>
> The current implementation does all the work of hot removing the cpus
> before halting the system. We are switching to just migrating to the
> boot cpu and then calling continuing with shutdown/reboot.
>
> This also has the effect of not breaking x86's command line parameter for
> specifying the reboot cpu. Note, this code was shamelessly copied from
> arch/x86/kernel/reboot.c with bits removed pertaining to the reboot_cpu
> command line parameter.
>
> Signed-off-by: Robin Holt <[email protected]>
> To: Shawn Guo <[email protected]>
> To: Ingo Molnar <[email protected]>
> To: Russ Anderson <[email protected]>
> Cc: Andrew Morton <[email protected]>
> Cc: "H. Peter Anvin" <[email protected]>
> Cc: Lai Jiangshan <[email protected]>
> Cc: Linus Torvalds <[email protected]>
> Cc: Linux Kernel Mailing List <[email protected]>
> Cc: Michel Lespinasse <[email protected]>
> Cc: Oleg Nesterov <[email protected]>
> Cc: "Paul E. McKenney" <[email protected]>
> Cc: Paul Mackerras <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Cc: Robin Holt <[email protected]>
> Cc: "[email protected]" <[email protected]>
> Cc: Tejun Heo <[email protected]>
> Cc: the arch/x86 maintainers <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: <[email protected]>
> ---
> kernel/sys.c | 17 +++++++++++++++--
> 1 file changed, 15 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sys.c b/kernel/sys.c
> index 0da73cf..4d1047d 100644
> --- a/kernel/sys.c
> +++ b/kernel/sys.c
> @@ -357,6 +357,19 @@ int unregister_reboot_notifier(struct notifier_block *nb)
> }
> EXPORT_SYMBOL(unregister_reboot_notifier);
>
> +void migrate_to_boot_cpu(void)
> +{
> + /* The boot cpu is always logical cpu 0 */
> + int reboot_cpu_id = 0;
> +
> + /* Make certain the cpu I'm about to reboot on is online */
> + if (!cpu_online(reboot_cpu_id))
> + reboot_cpu_id = smp_processor_id();
> +
> + /* Make certain I only run on the appropriate processor */
> + set_cpus_allowed_ptr(current, cpumask_of(reboot_cpu_id));
> +}
I guess you suspect what I'm going to suggest next? :-)
While I think something like this commit would be acceptable as a minimal
regression fix, it would be really lovely to add a second patch as well, which
removes the same code from arch/x86/kernel/reboot.c:native_machine_shutdown() and
merged it with the kernel/sys.c version? That way all platforms gained a
reboot_cpu command line, and we'd have less code duplication as well. Win-win.
( While at it, it might make sense to move the reboot/shutdown related bits from
kernel/sys.c to a new kernel/shutdown.c or kernel/reboot.c file or so. )
Thanks,
Ingo
On Mon, Apr 15, 2013 at 12:16:44PM +0200, Ingo Molnar wrote:
>
> * Robin Holt <[email protected]> wrote:
>
> > From 1767003c943325e52ac78cac6fdbaf2ab638888d Mon Sep 17 00:00:00 2001
> > From: Robin Holt <[email protected]>
> > Date: Wed, 3 Apr 2013 13:52:00 -0500
> > Subject: [PATCH] Migrate shutdown/reboot to boot cpu.
> >
> > We recently noticed that reboot of a 1024 cpu machine takes approx 16
> > minutes of just stopping the cpus. The slowdown was tracked to commit
> > f96972f.
> >
> > The current implementation does all the work of hot removing the cpus
> > before halting the system. We are switching to just migrating to the
> > boot cpu and then calling continuing with shutdown/reboot.
> >
> > This also has the effect of not breaking x86's command line parameter for
> > specifying the reboot cpu. Note, this code was shamelessly copied from
> > arch/x86/kernel/reboot.c with bits removed pertaining to the reboot_cpu
> > command line parameter.
> >
> > Signed-off-by: Robin Holt <[email protected]>
> > To: Shawn Guo <[email protected]>
> > To: Ingo Molnar <[email protected]>
> > To: Russ Anderson <[email protected]>
> > Cc: Andrew Morton <[email protected]>
> > Cc: "H. Peter Anvin" <[email protected]>
> > Cc: Lai Jiangshan <[email protected]>
> > Cc: Linus Torvalds <[email protected]>
> > Cc: Linux Kernel Mailing List <[email protected]>
> > Cc: Michel Lespinasse <[email protected]>
> > Cc: Oleg Nesterov <[email protected]>
> > Cc: "Paul E. McKenney" <[email protected]>
> > Cc: Paul Mackerras <[email protected]>
> > Cc: Peter Zijlstra <[email protected]>
> > Cc: Robin Holt <[email protected]>
> > Cc: "[email protected]" <[email protected]>
> > Cc: Tejun Heo <[email protected]>
> > Cc: the arch/x86 maintainers <[email protected]>
> > Cc: Thomas Gleixner <[email protected]>
> > Cc: <[email protected]>
> > ---
> > kernel/sys.c | 17 +++++++++++++++--
> > 1 file changed, 15 insertions(+), 2 deletions(-)
> >
> > diff --git a/kernel/sys.c b/kernel/sys.c
> > index 0da73cf..4d1047d 100644
> > --- a/kernel/sys.c
> > +++ b/kernel/sys.c
> > @@ -357,6 +357,19 @@ int unregister_reboot_notifier(struct notifier_block *nb)
> > }
> > EXPORT_SYMBOL(unregister_reboot_notifier);
> >
> > +void migrate_to_boot_cpu(void)
> > +{
> > + /* The boot cpu is always logical cpu 0 */
> > + int reboot_cpu_id = 0;
> > +
> > + /* Make certain the cpu I'm about to reboot on is online */
> > + if (!cpu_online(reboot_cpu_id))
> > + reboot_cpu_id = smp_processor_id();
> > +
> > + /* Make certain I only run on the appropriate processor */
> > + set_cpus_allowed_ptr(current, cpumask_of(reboot_cpu_id));
> > +}
>
> I guess you suspect what I'm going to suggest next? :-)
>
> While I think something like this commit would be acceptable as a minimal
> regression fix, it would be really lovely to add a second patch as well, which
> removes the same code from arch/x86/kernel/reboot.c:native_machine_shutdown() and
> merged it with the kernel/sys.c version? That way all platforms gained a
> reboot_cpu command line, and we'd have less code duplication as well. Win-win.
>
> ( While at it, it might make sense to move the reboot/shutdown related bits from
> kernel/sys.c to a new kernel/shutdown.c or kernel/reboot.c file or so. )
That all sounds good. I will work on that this morning.
Robin
I have the patches sort-of finished. The patch set starts by
moving the halt/shutdown/reboot functions over to a new
kernel/reboot.c, next patch gets a checkpatch.pl cleanup to
work, third patch is essentially the below patch against the
new file, and the fourth patch introduces a kernel boot parameter.
That said, I don't like them because of the 'stable' marking for
these patches. I think I am going submit them with the
existing patch first in the series, then introduce the kernel parameter,
then move them to kernel/reboot.c, and finally pass checkpatch.pl.
Does that sound alright?
Robin
On Mon, Apr 15, 2013 at 07:02:16AM -0500, Robin Holt wrote:
> On Mon, Apr 15, 2013 at 12:16:44PM +0200, Ingo Molnar wrote:
> >
> > * Robin Holt <[email protected]> wrote:
> >
> > > From 1767003c943325e52ac78cac6fdbaf2ab638888d Mon Sep 17 00:00:00 2001
> > > From: Robin Holt <[email protected]>
> > > Date: Wed, 3 Apr 2013 13:52:00 -0500
> > > Subject: [PATCH] Migrate shutdown/reboot to boot cpu.
> > >
> > > We recently noticed that reboot of a 1024 cpu machine takes approx 16
> > > minutes of just stopping the cpus. The slowdown was tracked to commit
> > > f96972f.
> > >
> > > The current implementation does all the work of hot removing the cpus
> > > before halting the system. We are switching to just migrating to the
> > > boot cpu and then calling continuing with shutdown/reboot.
> > >
> > > This also has the effect of not breaking x86's command line parameter for
> > > specifying the reboot cpu. Note, this code was shamelessly copied from
> > > arch/x86/kernel/reboot.c with bits removed pertaining to the reboot_cpu
> > > command line parameter.
> > >
> > > Signed-off-by: Robin Holt <[email protected]>
> > > To: Shawn Guo <[email protected]>
> > > To: Ingo Molnar <[email protected]>
> > > To: Russ Anderson <[email protected]>
> > > Cc: Andrew Morton <[email protected]>
> > > Cc: "H. Peter Anvin" <[email protected]>
> > > Cc: Lai Jiangshan <[email protected]>
> > > Cc: Linus Torvalds <[email protected]>
> > > Cc: Linux Kernel Mailing List <[email protected]>
> > > Cc: Michel Lespinasse <[email protected]>
> > > Cc: Oleg Nesterov <[email protected]>
> > > Cc: "Paul E. McKenney" <[email protected]>
> > > Cc: Paul Mackerras <[email protected]>
> > > Cc: Peter Zijlstra <[email protected]>
> > > Cc: Robin Holt <[email protected]>
> > > Cc: "[email protected]" <[email protected]>
> > > Cc: Tejun Heo <[email protected]>
> > > Cc: the arch/x86 maintainers <[email protected]>
> > > Cc: Thomas Gleixner <[email protected]>
> > > Cc: <[email protected]>
> > > ---
> > > kernel/sys.c | 17 +++++++++++++++--
> > > 1 file changed, 15 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/kernel/sys.c b/kernel/sys.c
> > > index 0da73cf..4d1047d 100644
> > > --- a/kernel/sys.c
> > > +++ b/kernel/sys.c
> > > @@ -357,6 +357,19 @@ int unregister_reboot_notifier(struct notifier_block *nb)
> > > }
> > > EXPORT_SYMBOL(unregister_reboot_notifier);
> > >
> > > +void migrate_to_boot_cpu(void)
> > > +{
> > > + /* The boot cpu is always logical cpu 0 */
> > > + int reboot_cpu_id = 0;
> > > +
> > > + /* Make certain the cpu I'm about to reboot on is online */
> > > + if (!cpu_online(reboot_cpu_id))
> > > + reboot_cpu_id = smp_processor_id();
> > > +
> > > + /* Make certain I only run on the appropriate processor */
> > > + set_cpus_allowed_ptr(current, cpumask_of(reboot_cpu_id));
> > > +}
> >
> > I guess you suspect what I'm going to suggest next? :-)
> >
> > While I think something like this commit would be acceptable as a minimal
> > regression fix, it would be really lovely to add a second patch as well, which
> > removes the same code from arch/x86/kernel/reboot.c:native_machine_shutdown() and
> > merged it with the kernel/sys.c version? That way all platforms gained a
> > reboot_cpu command line, and we'd have less code duplication as well. Win-win.
> >
> > ( While at it, it might make sense to move the reboot/shutdown related bits from
> > kernel/sys.c to a new kernel/shutdown.c or kernel/reboot.c file or so. )
>
> That all sounds good. I will work on that this morning.
>
> Robin
On Sat, Apr 13, 2013 at 06:30:22PM +0200, Oleg Nesterov wrote:
> On 04/12, Robin Holt wrote:
> >
> > +void migrate_to_boot_cpu(void)
> > +{
> > + /* The boot cpu is always logical cpu 0 */
> > + int reboot_cpu_id = 0;
> > +
> > + /* Make certain the cpu I'm about to reboot on is online */
> > + if (!cpu_online(reboot_cpu_id))
> > + reboot_cpu_id = smp_processor_id();
> > +
> > + /* Make certain I only run on the appropriate processor */
> > + set_cpus_allowed_ptr(current, cpumask_of(reboot_cpu_id));
>
> This is only theoretical, but perhaps it makes sense to set
> PF_THREAD_BOUND before set_cpus_allowed_ptr() ? To prevent the
> race with another thread doing sched_setaffinity().
I don't quite understand this comment. We are migrating our own thread.
How does setting PF_THREAD_BOUND have any effect?
Robin
On Mon, Apr 15, 2013 at 11:04:08AM -0500, Robin Holt wrote:
> On Sat, Apr 13, 2013 at 06:30:22PM +0200, Oleg Nesterov wrote:
> > On 04/12, Robin Holt wrote:
> > >
> > > +void migrate_to_boot_cpu(void)
> > > +{
> > > + /* The boot cpu is always logical cpu 0 */
> > > + int reboot_cpu_id = 0;
> > > +
> > > + /* Make certain the cpu I'm about to reboot on is online */
> > > + if (!cpu_online(reboot_cpu_id))
> > > + reboot_cpu_id = smp_processor_id();
> > > +
> > > + /* Make certain I only run on the appropriate processor */
> > > + set_cpus_allowed_ptr(current, cpumask_of(reboot_cpu_id));
> >
> > This is only theoretical, but perhaps it makes sense to set
> > PF_THREAD_BOUND before set_cpus_allowed_ptr() ? To prevent the
> > race with another thread doing sched_setaffinity().
>
> I don't quite understand this comment. We are migrating our own thread.
> How does setting PF_THREAD_BOUND have any effect?
I should have taken a few more minutes. I understand now. Will
do.
Robin
On 04/15, Robin Holt wrote:
>
> On Sat, Apr 13, 2013 at 06:30:22PM +0200, Oleg Nesterov wrote:
> > On 04/12, Robin Holt wrote:
> > >
> > > +void migrate_to_boot_cpu(void)
> > > +{
> > > + /* The boot cpu is always logical cpu 0 */
> > > + int reboot_cpu_id = 0;
> > > +
> > > + /* Make certain the cpu I'm about to reboot on is online */
> > > + if (!cpu_online(reboot_cpu_id))
> > > + reboot_cpu_id = smp_processor_id();
> > > +
> > > + /* Make certain I only run on the appropriate processor */
> > > + set_cpus_allowed_ptr(current, cpumask_of(reboot_cpu_id));
> >
> > This is only theoretical, but perhaps it makes sense to set
> > PF_THREAD_BOUND before set_cpus_allowed_ptr() ? To prevent the
> > race with another thread doing sched_setaffinity().
>
> I don't quite understand this comment. We are migrating our own thread.
> How does setting PF_THREAD_BOUND have any effect?
Suppose that another thread does, say, sys_sched_setaffinity(our_pid)
right after set_cpus_allowed_ptr(). If we set PF_THREAD_BOUND the
next set_cpus_allowed() will fail.
Oleg.
* Robin Holt <[email protected]> wrote:
> I have the patches sort-of finished. The patch set starts by
> moving the halt/shutdown/reboot functions over to a new
> kernel/reboot.c, next patch gets a checkpatch.pl cleanup to
> work, third patch is essentially the below patch against the
> new file, and the fourth patch introduces a kernel boot parameter.
>
> That said, I don't like them because of the 'stable' marking for
> these patches. I think I am going submit them with the
> existing patch first in the series, then introduce the kernel parameter,
> then move them to kernel/reboot.c, and finally pass checkpatch.pl.
>
> Does that sound alright?
Yeah, that ordering sounds right.
If there are no objections from others I'll first apply the first patch (with a
-stable tag), test it for a day, then apply the rest.
Even patch #1 probably won't make it for v3.9-final [there's too many potential
downsides IMHO], but this could be one of the cases where marking a patch for
-stable and merging it in the merge window is legit.
Thanks,
Ingo