Re: [PATCH] driver core: fix shutdown races with probe/remove

[Ming Lei <ming.lei@canonical.com>]

On Wed, Jun 6, 2012 at 10:44 PM, Alan Stern <stern@rowland.harvard.edu> wrote:

>> > On the whole, it might be easier just to hold the device lock during
>> > the shutdown call.
>>
>> Yes, it is easier, but it is not efficient because there are very less
>> drivers which implemented .shutdown callback. Suppose there are some
>> drivers which have no .shutdown but are holding device lock,  extra
>> time will be consumed in the lock acquiring, which may slow down 'power
>> down'.
>
> The main reasons for holding the device lock are probe and remove
> callbacks.  In your patch, synchronize_srcu has to wait for those too.
> So there won't be much extra time.

There are differences between holding dev->lock(parent->lock) and
using synchronize_srcu:

	- synchronize_srcu just wait for the ongoing probe/release
	to complete

	- holding dev->lock and parent->lock before shutdown means
	every dev->lock and parent->lock is to be acquired, so any
	drivers holding dev lock during device_shutdown will slow down
	'power down' or 'reset'.

	- if the device has not .shutdown implemented, it is not
	necessary to acquire the lock and parent->lock, either of
	which may has been held in other places already.

But after thinking about the problem further, the patch isn't good enough:

	- even probe/remove is disallowed during shutdown, device_add/
	device_del can be completed and only probe and release things
	are bypassed, so .shutdown may see a deleted device from view of
	driver model. For example, pci_device_shutdown need to access
	its devres which may have been deleted by device_del already.

	- violate driver model's implicit rule: once device_release_driver
	(in device_del path) is completed, other callbacks for the device(driver)
	won't be seen or called by others

So take the simply approach of holding lock to fix the races.


Thanks,
--
Ming Lei