Few problems in mtd system

[Michael Trimarchi]

Hi

Maxim Levitsky wrote:
> On Tue, 2009-12-22 at 21:32 +0100, Michael Trimarchi wrote: 
>   
>> Maxim Levitsky wrote:
>>     
>>> On Mon, 2009-12-21 at 15:00 +0100, Jörn Engel wrote: 
>>>   
>>>       
>>>> On Sun, 20 December 2009 22:59:55 +0200, Maxim Levitsky wrote:
>>>>     
>>>>         
>>>>> I suspect that older non type M cards didn't have such emulation, but
>>>>> were real nand chips, bacause there are some references on the web about
>>>>> using XD card as a nand chip replacement.
>>>>> Such card as I have will really will make very poor nand replacement...
>>>>>       
>>>>>           
>>>> If you want to know whether the cards or the readers are to blame, you
>>>> can try to buy an old alauda reader on ebay.  It is too slow to be
>>>> useful for most purposes, but I believe it did give me full access with
>>>> my cards.
>>>>
>>>>     
>>>>         
>>>>> Folks, could you review my other questions about bugs in mtd core, and
>>>>> tell your opinion?
>>>>>       
>>>>>           
>>> I have several problems, more correctly bugs in mtd system I have to fix
>>> to make my driver work.
>>>
>>> Lets start from  the problem I face now.
>>>
>>> Problem is that add_mtd_blktrans_dev is called with mtd_table_mutex
>>> locked, but it calls add_disk which opens the block device if you
>>> specify that you need partitions on the disk. Open routine 'looks' at
>>> mtd table using get_mtd_device, and thus deadlocks.
>>>
>>> Do you have a clue how to fix that so it won't break anything?
>>>   
>>>       
>> I try to follow the call, can you send the block stack trace?
>> just change the lock on get_mtd_device with a try_lock and BUG on
>> if it is taken
>>     
> I don't want to crash the system now, but I know exactly the trace:
> (this is created manually)
>
> get_mtd_device
> blktrans_open
> __blkdev_get
> blkdev_get
> register_disk
> add_disk
> add_mtd_blktrans_dev
> ssfdcr_add_mtd
> blktrans_notify_add
>   
The problem can be introduced by this commit
8022c13c27b822cf22f13df10b42aae89cd56bf0

 mtd: blkdevs: do not forget to get MTD devices
   
    Nowadays MTD devices have to be "get" before they can be
    used. This has to be done with 'get_mtd_device()'. The
    'blktrans_open()' function did not do this and instead
    used 'try_module_get()'. Fix this.
   
    Since 'get_mtd_device()' already gets the module, extra
    'try_module_get()' is not needed.
   
    This fixes oops when one tries to use mtdblock on top of
    gluebi.
   
    Reported-by: Holger Brunck <holger.brunck at keymile.com>
    Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy at nokia.com>
    Signed-off-by: David Woodhouse <David.Woodhouse at intel.com>

That call the

-       if (!try_module_get(dev->mtd->owner))
+       if (!get_mtd_device(NULL, dev->mtd->index))
                goto out;

Regards Michael
> add_mtd_device - takes the lock
>   
How do you define partition? cmdline partition or pdata?
Look of the probe of other device it calls the add_mtd_partions and no
the add_mtd_device.
> rsc_register_nand_device - my driver function
>
>
> The point is that that ssfdc specifies:
>
> .part_bits = SSFDCR_PARTN_BITS
>
> This triggers the open of the block device by block core.
>
> Tomorrow I will post full backtrace.
>
> Best regards,
> Maxim Levitsky
>
>
>