[PATCH 10/10] cpu bulk removal interface

[PATCH 10/10] cpu bulk removal interface

[Shaohua Li]

Interface for bulk cpu removal. It's /sys/devices/system/cpu/cpu_bulk_remove

Signed-off-by: Shaohua Li <shaohua.li@intel.com>
---

 linux-2.6.17-rc3-root/drivers/base/cpu.c  |   47 +++++++++++++++++++++++++++++-
 linux-2.6.17-rc3-root/include/linux/cpu.h |    3 +
 2 files changed, 49 insertions(+), 1 deletion(-)

diff -puN drivers/base/cpu.c~bulk_cpu_remove_interface drivers/base/cpu.c
--- linux-2.6.17-rc3/drivers/base/cpu.c~bulk_cpu_remove_interface	2006-05-07 07:47:02.000000000 +0800
+++ linux-2.6.17-rc3-root/drivers/base/cpu.c	2006-05-07 09:29:54.000000000 +0800
@@ -76,6 +76,46 @@ static inline void register_cpu_control(
 }
 #endif /* CONFIG_HOTPLUG_CPU */
 
+#ifdef CONFIG_BULK_CPU_REMOVE
+static ssize_t cpu_bulk_remove_show(struct sysdev_class *c, char *buf)
+{
+	int len;
+
+	len = cpulist_scnprintf(buf, PAGE_SIZE-1, cpu_online_map);
+	len += sprintf(buf + len, "\n");
+	return len;
+}
+
+static ssize_t cpu_bulk_remove_store(struct sysdev_class *c,
+	const char *buf, size_t count)
+{
+	int err;
+	cpumask_t removed_cpus;
+
+	if ((err = lock_cpu_hotplug_interruptible() != 0))
+		return err;
+	err = cpulist_parse(buf, removed_cpus);
+	if (err) {
+		unlock_cpu_hotplug();
+		return err;
+	}
+
+	unlock_cpu_hotplug();
+	cpu_down_mask(removed_cpus);
+	return count;
+}
+
+static SYSDEV_CLASS_ATTR(cpu_bulk_remove, 0600, cpu_bulk_remove_show,
+	cpu_bulk_remove_store);
+
+void __init cpu_bulk_remove_sysfs_init(void)
+{
+	sysdev_class_create_file(&cpu_sysdev_class, &attr_cpu_bulk_remove);
+}
+#else
+#define cpu_bulk_remove_sysfs_init()
+#endif
+
 #ifdef CONFIG_KEXEC
 #include <linux/kexec.h>
 
@@ -145,5 +185,10 @@ EXPORT_SYMBOL_GPL(get_cpu_sysdev);
 
 int __init cpu_dev_init(void)
 {
-	return sysdev_class_register(&cpu_sysdev_class);
+	int ret;
+
+	ret = sysdev_class_register(&cpu_sysdev_class);
+	if (!ret)
+		cpu_bulk_remove_sysfs_init();
+	return ret;
 }
diff -puN include/linux/cpu.h~bulk_cpu_remove_interface include/linux/cpu.h
--- linux-2.6.17-rc3/include/linux/cpu.h~bulk_cpu_remove_interface	2006-05-07 07:47:02.000000000 +0800
+++ linux-2.6.17-rc3-root/include/linux/cpu.h	2006-05-07 07:47:02.000000000 +0800
@@ -74,6 +74,9 @@ extern int lock_cpu_hotplug_interruptibl
 	register_cpu_notifier(&fn##_nb);			\
 }
 int cpu_down(unsigned int cpu);
+#ifdef CONFIG_BULK_CPU_REMOVE
+int cpu_down_mask(cpumask_t remove_mask);
+#endif
 #define cpu_is_offline(cpu) unlikely(!cpu_online(cpu))
 #else
 #define lock_cpu_hotplug()	do { } while (0)
_

Re: [PATCH 10/10] cpu bulk removal interface

[Nathan Lynch]

Shaohua Li wrote:
> 
> Interface for bulk cpu removal. It's /sys/devices/system/cpu/cpu_bulk_remove
> 
> Signed-off-by: Shaohua Li <shaohua.li@intel.com>
> ---
> 
>  linux-2.6.17-rc3-root/drivers/base/cpu.c  |   47 +++++++++++++++++++++++++++++-
>  linux-2.6.17-rc3-root/include/linux/cpu.h |    3 +
>  2 files changed, 49 insertions(+), 1 deletion(-)
> 
> diff -puN drivers/base/cpu.c~bulk_cpu_remove_interface drivers/base/cpu.c
> --- linux-2.6.17-rc3/drivers/base/cpu.c~bulk_cpu_remove_interface	2006-05-07 07:47:02.000000000 +0800
> +++ linux-2.6.17-rc3-root/drivers/base/cpu.c	2006-05-07 09:29:54.000000000 +0800
> @@ -76,6 +76,46 @@ static inline void register_cpu_control(
>  }
>  #endif /* CONFIG_HOTPLUG_CPU */
>  
> +#ifdef CONFIG_BULK_CPU_REMOVE
> +static ssize_t cpu_bulk_remove_show(struct sysdev_class *c, char *buf)
> +{
> +	int len;
> +
> +	len = cpulist_scnprintf(buf, PAGE_SIZE-1, cpu_online_map);
> +	len += sprintf(buf + len, "\n");
> +	return len;
> +}

This doesn't really seem meaningful.  I'd say the attribute could do
without a show method.

> +static ssize_t cpu_bulk_remove_store(struct sysdev_class *c,
> +	const char *buf, size_t count)
> +{
> +	int err;
> +	cpumask_t removed_cpus;
> +
> +	if ((err = lock_cpu_hotplug_interruptible() != 0))
> +		return err;
> +	err = cpulist_parse(buf, removed_cpus);
> +	if (err) {
> +		unlock_cpu_hotplug();
> +		return err;
> +	}
> +
> +	unlock_cpu_hotplug();
> +	cpu_down_mask(removed_cpus);
> +	return count;
> +}

Shouldn't this make sure that we don't offline all cpus?

Why are you using lock_cpu_hotplug_interruptible instead of
lock_cpu_hotplug?

Why is only the parsing of the cpumask buffer protected by the cpu
hotplug lock?  Shouldn't cpu_down_mask be called with the lock held?

Can cpu_down_mask fail, and if so, shouldn't we be reporting the
error?

Re: [PATCH 10/10] cpu bulk removal interface

[Shaohua Li]

On Mon, 2006-05-08 at 01:31 -0500, Nathan Lynch wrote:
> Shaohua Li wrote:
> > 
> > Interface for bulk cpu removal. It's /sys/devices/system/cpu/cpu_bulk_remove
> > 
> > Signed-off-by: Shaohua Li <shaohua.li@intel.com>
> > ---
> > 
> >  linux-2.6.17-rc3-root/drivers/base/cpu.c  |   47 +++++++++++++++++++++++++++++-
> >  linux-2.6.17-rc3-root/include/linux/cpu.h |    3 +
> >  2 files changed, 49 insertions(+), 1 deletion(-)
> > 
> > diff -puN drivers/base/cpu.c~bulk_cpu_remove_interface drivers/base/cpu.c
> > --- linux-2.6.17-rc3/drivers/base/cpu.c~bulk_cpu_remove_interface	2006-05-07 07:47:02.000000000 +0800
> > +++ linux-2.6.17-rc3-root/drivers/base/cpu.c	2006-05-07 09:29:54.000000000 +0800
> > @@ -76,6 +76,46 @@ static inline void register_cpu_control(
> >  }
> >  #endif /* CONFIG_HOTPLUG_CPU */
> >  
> > +#ifdef CONFIG_BULK_CPU_REMOVE
> > +static ssize_t cpu_bulk_remove_show(struct sysdev_class *c, char *buf)
> > +{
> > +	int len;
> > +
> > +	len = cpulist_scnprintf(buf, PAGE_SIZE-1, cpu_online_map);
> > +	len += sprintf(buf + len, "\n");
> > +	return len;
> > +}
> 
> This doesn't really seem meaningful.  I'd say the attribute could do
> without a show method.
Ok, I'll remove it.

> > +static ssize_t cpu_bulk_remove_store(struct sysdev_class *c,
> > +	const char *buf, size_t count)
> > +{
> > +	int err;
> > +	cpumask_t removed_cpus;
> > +
> > +	if ((err = lock_cpu_hotplug_interruptible() != 0))
> > +		return err;
> > +	err = cpulist_parse(buf, removed_cpus);
> > +	if (err) {
> > +		unlock_cpu_hotplug();
> > +		return err;
> > +	}
> > +
> > +	unlock_cpu_hotplug();
> > +	cpu_down_mask(removed_cpus);
> > +	return count;
> > +}
> 
> Shouldn't this make sure that we don't offline all cpus?
cpu_down_mask will check it.

> Why are you using lock_cpu_hotplug_interruptible instead of
> lock_cpu_hotplug?
> 
> Why is only the parsing of the cpumask buffer protected by the cpu
> hotplug lock?  Shouldn't cpu_down_mask be called with the lock held?
we actually don't need lock here. I'll remove it.

> Can cpu_down_mask fail, and if so, shouldn't we be reporting the
> error?
Makes sense.

Thanks,
Shaohua