Re: [PATCH, v4] hwmon: coretemp: use list instead of fixed size array for temp data

[Guenter Roeck <guenter.roeck@ericsson.com>]

On Wed, May 09, 2012 at 06:16:34AM -0400, Kirill A. Shutemov wrote:
> On Wed, May 09, 2012 at 02:56:17AM -0700, Guenter Roeck wrote:
> > On Wed, May 09, 2012 at 03:23:39AM -0400, Kirill A. Shutemov wrote:
> > > On Wed, May 09, 2012 at 10:09:06AM +0300, Kirill A. Shutemov wrote:
> > > > On Tue, May 08, 2012 at 09:39:40AM -0700, Guenter Roeck wrote:
> > > > > On Tue, 2012-05-08 at 06:49 -0400, Kirill A. Shutemov wrote:
> > > > > > From: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
> > > > > > 
> > > > > > Let's rework code to allow arbitrary number of cores on a CPU, not
> > > > > > limited by hardcoded array size.
> > > > > > 
> > > > > > Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> > > > > > ---
> > > > > >  v4:
> > > > > >    - address issues pointed by Guenter Roeck;
> > > > > >  v3:
> > > > > >    - drop redundant refcounting and checks;
> > > > > >  v2:
> > > > > >    - fix NULL pointer dereference. Thanks to R, Durgadoss;
> > > > > >    - use mutex instead of spinlock for list locking.
> > > > > > ---
> > > > > 
> > > > > Hi Kirill,
> > > > > 
> > > > > unfortunately now we have another race condition :(. See below ...
> > > > 
> > > > Ughh..
> > > > 
> > > > > > @@ -557,11 +579,22 @@ exit_free:
> > > > > >  static int __devexit coretemp_remove(struct platform_device *pdev)
> > > > > >  {
> > > > > >  	struct platform_data *pdata = platform_get_drvdata(pdev);
> > > > > > -	int i;
> > > > > > +	struct temp_data *tdata;
> > > > > >  
> > > > > > -	for (i = MAX_CORE_DATA - 1; i >= 0; --i)
> > > > > > -		if (pdata->core_data[i])
> > > > > > -			coretemp_remove_core(pdata, &pdev->dev, i);
> > > > > > +	for (;;) {
> > > > > > +		mutex_lock(&pdata->temp_data_lock);
> > > > > > +		if (!list_empty(&pdata->temp_data_list)) {
> > > > > > +			tdata = list_first_entry(&pdata->temp_data_list,
> > > > > > +					struct temp_data, list);
> > > > > > +			list_del(&tdata->list);
> > > > > > +		} else
> > > > > > +			tdata = NULL;
> > > > > > +		mutex_unlock(&pdata->temp_data_lock);
> > > > > > +
> > > > > > +		if (!tdata)
> > > > > > +			break;
> > > > > > +		coretemp_remove_core(tdata, &pdev->dev);
> > > > > > +	}
> > > > > >  
> > > > > Unfortunately, that results in a race condition, since the tdata list
> > > > > entry is gone before the attribute file is deleted.
> > > > > 
> > > > > I think you can still use list_for_each_entry_safe, only outside the
> > > > > mutex, and remove the list entry at the end of coretemp_remove_core()
> > > 
> > > I haven't got how list_for_each_entry_safe() will be really safe without
> > > the lock.
> > > 
> > We know that it by itself won't be called multiple times. So the only question is 
> > if the functions to add/remove a core can be called while coretemp_remove is called,
> > or if that is mutually exclusive (not that the current code handles this case).
> > 
> > Fortunately, there is a function to block CPU removal/insertion: get_online_cpus()
> > and put_online_cpus(). I have no idea if it is necessary to protect coretemp_remove()
> > with it, but it might be on the safe side anyway.
> > 
> > > > > after deleting the attribute file. Just keep the code as it was, and
> > > > > remove the list entry (mutex-protected) where core_data[] was set to
> > > > > NULL.
> > > > 
> > > > I think
> > > > 
> > > > if (tdata)
> > > > 	return -ENODEV;
> > > > 
> > > > in show methods will fix the issue. Right?
> > > 
> > > It won't. Stupid me.
> > > 
> > > But the check + kref seems will work...
> > > 
> > Yes, but would be way too complicated.
> 
> More code, yes, but complicated? What you propose looks like a trick. It
> has too many assumptions on context.
> 

There is an even better solution: unregistering the hotplug notifier
before removing the driver. And, as you will notice, that is already done.
So list_for_each_entry_safe() is safe after all, since no other remove/add
activity will occur at the same time.

> I personally prefer kref since it's straight forward and more friendly for
> future changes.
> 
Guess we have to agree to disagree on that one.

Thanks,
Guenter