Re: How to cleanly shut down a block device

[Jens Axboe <jens.axboe@oracle.com>]

On Tue, Nov 14 2006, Pierre Ossman wrote:
> Hi Jens,
> 
> I've been trying to sort out some bugs in the MMC layer's block driver,
> but my knowledge about the block layer is severely lacking. So I was
> hoping you could educate me a bit. :)
> 
> Upon creation, the following happens:
> 
> alloc_disk()
> spin_lock_init()
> blk_init_queue()
> blk_queue_*() (Set up limits)
> disk->* = * (assign members)
> blk_queue_hardsect_size()
> set_capacity()
> add_disk()
> 
> And on a clean removal, where there are no users of a card when it is
> removed:
> 
> del_gendisk()
> put_disk()
> blk_cleanup_queue()
> 
> So far everything seems nice and peachy. The question is what to do when
> a card is removed when the device is open.
> 
> In that case, del_gendisk() will be called, which seems to be documented
> as blocking any new requests to be added to the queue. But there will be
> a lot of outstanding requests in the queue.
> 
> Is it up to each block device to iterate and fail these or can I tell
> the kernel "I'm broken, go away!"?
> 
> When the queue eventually drains (without too many oopses) and the user
> calls close(), then put_disk() and blk_cleanup_queue() will be called.

There is no helper to kill already queued requests when a device is
removed, if you look at SCSI you'll see that it handles this "manually"
as well in the request_fn handler. So you'll need a "device dead or
gone" check in your request_fn handler, and do it from there.

-- 
Jens Axboe