Re: [PATCH] pci: Fixing drivers/pci/search.c compilation warning.

["Zhao, Yu" <yu.zhao@intel.com>]

Zhao, Yu wrote:
> Matthew Wilcox wrote:
>> On Mon, Oct 27, 2008 at 03:13:51PM +0800, Zhao, Yu wrote:
>>> Matthew Wilcox wrote:
>>>> On Mon, Oct 27, 2008 at 11:18:43AM +0800, Zhao, Yu wrote:
>>>>> Matthew Wilcox wrote:
>>>>>> Yes, that's why pci_find_device() is deprecated.  But it doesn't also
>>>>>> need to be buggy ;-)
>>>>> How about pci_get_bus_and_slot()? People would meet the problem with it 
>>>>> anyway.
>>>> What problem with it?  It's documented to return the device with an
>>>> increased refcount, and the implementation appears to do exactly that:
>>>>
>>> The 'dev' returned by pci_get_device() may be destroyed by PCI hotplug. 
>>> I suppose that passing this 'dev' to pci_get_device() in the next loop 
>>> would crash the system, right?
>> Erm, no, the 'dev' cannot be destroyed because the caller has a refcount
>> on it.  The physical device backing it might have gone away.  The dev
> 
> Why does the caller have a reference count? I don't see we increase the 
> reference count after the 'dev' is returned by following in 
> pci_get_dev_by_id():
> 
>          dev = bus_find_device(&pci_bus_type, dev_start, (void *)id,
>                                match_pci_dev_by_id);
> 
> And this 'dev' becomes the 'from' in the next loop, but it may be 
> destroyed before the 'pci_dev_get(from)', isn't it?

I checked the source code, there is no 'pci_dev_get(from)', the 
reference count is increased in bus_find_device().

         while ((dev = next_device(&i)))
                 if (match(dev, data) && get_device(dev))

But the essential problem is same: the reference count of 'dev' above 
may be decreased before the 'get_device(dev)', I guess.