2.6.28-rc2: (mtd)block/partitions BUG with kobject reference count

[Bryan Wu]

On Fri, Oct 31, 2008 at 7:47 AM, Kay Sievers <kay.sievers at vrfy.org> wrote:
> On Thu, Oct 30, 2008 at 22:51, Kay Sievers <kay.sievers at vrfy.org> wrote:
>> On Thu, Oct 30, 2008 at 00:28, Peter Korsgaard <jacmet at sunsite.dk> wrote:
>>>>>>>> "Rafael" == Rafael J Wysocki <rjw at sisk.pl> writes:
>>>
>>>  Rafael> On Wednesday, 29 of October 2008, Peter Korsgaard wrote:
>>>  >> Hi,
>>>  >>
>>>  >> I'm seing what looks like a kobject reference count issue with
>>>  >> mtdblock_ro + mtd_dataflash + mtd partitions and repeated unbind/bind.
>>>  >> I'm on 2.6.28-rc2, but I can reproduce the problem on 2.6.27 as well.
>>>
>>>  Rafael> Is it reproducible with 2.6.26 too?
>>>
>>> Sorry, I haven't backported my platform support code to such "old"
>>> kernel. I can do it though, if you think it will help pinpoint the
>>> issue.
>>
>> This sounds like a possible reason for the problem:
>> "After digging into the mtd code, this bug is not related to our driver. It
>> should be a subtle bug in mtd core code.
>>
>> In add_mtd_partition, for 2 partitions, 2 gendisk structures will be
>> allocated. But these 2 gendisk->queue will be set to the same
>> request_queue. Then when unregistering the 1st partition, from the
>> same request_queue->backing_dev_info, the bdi struct will be set to
>> NULL. So for the 2nd partition (bdi == NULL), the sysfs dir of 2nd
>> partition will not be removed. Finally, when modprobe the module
>> again, the 2nd partition won't be added"
>> https://blackfin.uclinux.org/gf/tracker/4463
>
> Looks like a bdi issue:
>  http://lkml.org/lkml/2008/10/30/519
>
> Peter, if I do this (whitespace mangled, just pasted in here), the
> error goes away for me. Can you try this?
>
> Thanks,
> Kay
>
> --- a/block/genhd.c
> +++ b/block/genhd.c
> @@ -515,7 +515,8 @@ void add_disk(struct gendisk *disk)
>        blk_register_queue(disk);
>
>        bdi = &disk->queue->backing_dev_info;
> -       bdi_register_dev(bdi, disk_devt(disk));
> +       if (!bdi->dev)
> +               bdi_register_dev(bdi, disk_devt(disk));
>        retval = sysfs_create_link(&disk_to_dev(disk)->kobj, &bdi->dev->kobj,
>                                   "bdi");
>        WARN_ON(retval);

IMHO, this is a workaround, right? I think the final solution should
provide every 'gendisk' a dedicated 'bdi'.
So they won't mess up and overwrite.

-Bryan