[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [dm-devel] [PATCH v2] dm mpath: maintain reference count for underlying devices

Hi James,

On 09/20/11 16:13, James Bottomley wrote:
>> James's suggestion was to fix b) by not freeing elevator
>> until blk_release_queue() is called:
>>   https://lkml.org/lkml/2011/8/10/421
>> But it would hit the same issue that there's no guarantee
>> of q->queue_lock validity after blk_cleanup_queue().
>> James's other suggestion was to add a callback mechanism
>> for driver to free q->queue_lock.
> Actually, there isn't a problem with this patch:  Any driver that
> expects blk_cleanup_queue to guarantee last use of the lock has to call
> it as the last releaser.  If it doesn't, it would oops (or would have
> oopsed before we started putting the block guards in).

'q->sysfs_lock' seems to protect the queue access via sysfs.
Are you sure it never delay the last put?

The separation of blk_cleanup_queue() and blk_release_queue()
was introduced with the following commit:

  commit 483f4afc421435b7cfe5e88f74eea0b73a476d75
  Author: Al Viro <viro zeniv linux org uk>
  Date:   Sat Mar 18 18:34:37 2006 -0500
  [PATCH] fix sysfs interaction and lifetime rules handling for queues

I added Al to Cc hoping he knows why elevator_exit() was
placed in blk_cleanup_queue() from the first time.

Other than the above concern, I have no objection to your patch.

>> I was trying to fix a) in the following thread:
>>   https://lkml.org/lkml/2011/8/18/103
>> but haven't gotten response from James so it seems rejected.
> I think I said quite a few times that this would reintroduce the oops I
> was trying to fix.  Alan seems to think that there are sufficient guards
> just to move the blk_cleanup_queue(), but I'd prefer to get
> blk_cleanup_queue() working properly like the del method it's supposed
> to be.

I think you haven't mentioned what oops would be reintroduced.

Jun'ichi Nomura, NEC Corporation

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]