Re: [RFC PATCH v5 2/2] Use kernfs_break_active_protection() for device online store callbacks

From: "Rafael J. Wysocki" <rafael.j.wysocki@intel.com>
To: Tejun Heo <tj@kernel.org>
Cc: Li Zhong <zhong@linux.vnet.ibm.com>,
	LKML <linux-kernel@vger.kernel.org>,
	gregkh@linuxfoundation.org, toshi.kani@hp.com
Subject: Re: [RFC PATCH v5 2/2] Use kernfs_break_active_protection() for device online store callbacks
Date: Wed, 23 Apr 2014 00:21:33 +0200	[thread overview]
Message-ID: <5356EB6D.3010102@intel.com> (raw)
In-Reply-To: <20140422204455.GB3615@mtj.dyndns.org>

On 4/22/2014 10:44 PM, Tejun Heo wrote:
> Hello,
>
> On Tue, Apr 22, 2014 at 11:34:39AM +0800, Li Zhong wrote:
>>> Is this assumption true?  If so, can we add lockdep assertions in
>>> places to verify and enforce this?  If not, aren't we just feeling
>>> good when the reality is broken?
>> It seems not true ... I think there are devices that don't have the
>> online/offline concept, we just need to add it, remove it, like ethernet
>> cards.
>>
>> Maybe we could change the comments above, like:
>> 	/* We assume device_hotplug_lock must be acquired before
>> 	 * removing devices, which have online/offline sysfs knob,
>> 	 * and some locks are needed to serialize the online/offline
>> 	 * callbacks and device removing. ...
>> ?
>>
>> And we could add lockdep assertions in cpu and memory related code? e.g.
>> remove_memory(), unregister_cpu()
>>
>> Currently, remove_memory() has comments for the function:
>>
>>   * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
>>   * and online/offline operations before this call, as required by
>>   * try_offline_node().
>>   */
>>      
>> maybe it could be removed with the lockdep assertion.
> I'm confused about the overall locking scheme.  What's the role of
> device_hotplug_lock?  Is that solely to prevent the sysfs deadlock
> issue?  Or does it serve other synchronization purposes depending on
> the specific subsystem?  If the former, the lock no longer needs to
> exist.  The only thing necessary would be synchronization between
> device_del() deleting the sysfs file and the unbreak helper invoking
> device-specific callback.  If the latter, we probably should change
> that.  Sharing hotplug lock across multiple subsystems through driver
> core sounds like a pretty bad idea.

Can you please elaborate a bit?

It is there to protect hotplug operations involving multiple devices (in 
different subsystems) from racing with each other.  Why exactly is it bad?

Rafael