Re: [RFC] Move the memory_notifier out of the memory_hotplug lock

[David Rientjes]

On Wed, 5 Feb 2014, Nathan Zimmer wrote:

> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> index 62a0cd1..a3cbd14 100644
> --- a/mm/memory_hotplug.c
> +++ b/mm/memory_hotplug.c
> @@ -985,12 +985,12 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages, int online_typ
>  		if (need_zonelists_rebuild)
>  			zone_pcp_reset(zone);
>  		mutex_unlock(&zonelists_mutex);
> +		unlock_memory_hotplug();
>  		printk(KERN_DEBUG "online_pages [mem %#010llx-%#010llx] failed\n",
>  		       (unsigned long long) pfn << PAGE_SHIFT,
>  		       (((unsigned long long) pfn + nr_pages)
>  			    << PAGE_SHIFT) - 1);
>  		memory_notify(MEM_CANCEL_ONLINE, &arg);
> -		unlock_memory_hotplug();
>  		return ret;
>  	}
>  
> @@ -1016,9 +1016,10 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages, int online_typ
>  
>  	writeback_set_ratelimit();
>  
> +	unlock_memory_hotplug();
> +
>  	if (onlined_pages)
>  		memory_notify(MEM_ONLINE, &arg);
> -	unlock_memory_hotplug();
>  
>  	return 0;
>  }

That looks a little problematic, what happens if a nid is being brought 
online and a registered callback does something like allocate resources 
for the arg->status_change_nid and the above two hunks of this patch end 
up racing?

Before, a registered callback would be guaranteed to see either a 
MEMORY_CANCEL_ONLINE or MEMORY_ONLINE after it has already done 
MEMORY_GOING_ONLINE.

With your patch, we could race and see one cpu doing MEMORY_GOING_ONLINE, 
another cpu doing MEMORY_GOING_ONLINE, and then MEMORY_ONLINE and 
MEMORY_CANCEL_ONLINE in either order.

So I think this patch will break most registered callbacks that actually 
depend on lock_memory_hotplug(), it's a coarse lock for that reason.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/