Re: How to cleanly shut down a block device

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

On Tue, Nov 14 2006, Pierre Ossman wrote:
> Russell King wrote:
> > Just arrange for the mmc_queue_thread() to empty the queue when
> > MMC_QUEUE_EXIT is set, and then exit.  I thought this was something
> > that the block layer looked after (Jens must have missed this in his
> > original review of the MMC code.)
> >   
> 
> mmc_queue_thread() will empty the thread when MMC_QUEUE_EXIT is set. The
> problem is that we do not set that bit until the last person closes the
> device. In order to avoid problems we need to empty the queue before
> mmc_blk_remove() exits (after which the card structure is no longer valid).
> 
> > The handling of userspace keeping the device open despite the hardware
> > having been removed is already in place.
> >
> >   
> 
> Ok, that's one less problem for me to worry about. :)
> 
> Jens Axboe wrote:
> > What do you mean by "killing off the queue"? As long as the queue can be
> > gotten at, it needs to remain valid. That is what the references are
> > for.
> >   
> 
> I do:
> 
> del_gendisk();
> (wait for queue to become empty, i.e. elv_next_request() == NULL)
> blk_cleanup_queue();

elv_next_request() returning NULL means nothing wrt the queue being
empty.

> and then assume that the request function will no longer be called for
> this queue.
> 
> Suggested patch:

I think you are making this way too complicated, it's actually pretty
simple: you call blk_put_queue() or blk_cleanup_queue() (same thing)
when _you_ drop your reference to the queue. That's just normal cleanup.
When a device goes away, you make sure that you know about this. I said
that SCSI clears q->queuedata, so it knows that when ->request_fn is
invoked with a NULL q->queuedata (where it stores the device pointer),
the device is not there and the request should just be flushed to
heaven.

Don't make any assumptions about when request_fn will be called or not.
That's bound to be racy anyway.

-- 
Jens Axboe