Re: [PATCH] block: generic_unplug_device implicitly makes irq enable

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

On Tue, May 08 2007, Konstantin Baydarov wrote:
> Hi,
> 
> while working on LKCD I found out common block device issue.
> LKCD uses generic_unplug_device() to finish disk IO in case of blockdev kernel core dump.
> I found out that after calling of generic_unplug_device() IRQs become implicitly enabled (it is supposed that IRQs are disabled).
> I looked at generic_unplug_device() code and found out that function lose state of irq flags, it use spin_lock_irq()/spin_unlock_irq() instead of spin_lock_irqsave()/spin_unlock_irqrestore():
> void generic_unplug_device(request_queue_t *q)
> {
> 	spin_lock_irq(q->queue_lock);
> 	__generic_unplug_device(q);
> 	spin_unlock_irq(q->queue_lock);
> }
> 
> generic_unplug_device() - is a method to unplug the block device, named unplug_fn in struct request_queue.
> I've compared generic_unplug_device() with other methods:
> file drivers/md/raid5.c ...
> static void raid5_unplug_device(request_queue_t *q)
> {
> 	mddev_t *mddev = q->queuedata;
> 	raid5_conf_t *conf = mddev_to_conf(mddev);
> 	unsigned long flags;
> 
> 	spin_lock_irqsave(&conf->device_lock, flags);
> 
> 	if (blk_remove_plug(q)) {
> 		conf->seq_flush++;
> 		raid5_activate_delayed(conf);
> 	}
> 	md_wakeup_thread(mddev->thread);
> 
> 	spin_unlock_irqrestore(&conf->device_lock, flags);
> 
> 	unplug_slaves(mddev);
> }
> 
> file drivers/block/umem.c ...
> static void mm_unplug_device(request_queue_t *q)
> {
> 	struct cardinfo *card = q->queuedata;
> 	unsigned long flags;
> 
> 	spin_lock_irqsave(&card->lock, flags);
> 	if (blk_remove_plug(q))
> 		activate(card);
> 	spin_unlock_irqrestore(&card->lock, flags);
> }
> 
> As you can see raid5_unplug_device() and mm_unplug_device() are using
> spin_lock_irqsave()/spin_unlock_irqrestore(), instead of spin_lock_irq()/spin_unlock_irq().
> I found out that generic_unplug_device() is used not only in LKCD, so I suggest to switch generic_unplug_device() to spin_lock_irqsave()/spin_unlock_irqrestore() to prevent implicitly losing of IRQ flags.
> Here is patch against kernel 2.6.21.1.
> Please CC me to answers/comments.
> Thanks.
> 
> Signed-off-by: Konstantin Baydarov <kbaidarov@ru.mvista.com>
> 
>  block/ll_rw_blk.c |    6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
> 
> Index: linux-2.6.21.1/block/ll_rw_blk.c
> ===================================================================
> --- linux-2.6.21.1.orig/block/ll_rw_blk.c
> +++ linux-2.6.21.1/block/ll_rw_blk.c
> @@ -1602,9 +1602,11 @@ EXPORT_SYMBOL(__generic_unplug_device);
>   **/
>  void generic_unplug_device(request_queue_t *q)
>  {
> -	spin_lock_irq(q->queue_lock);
> +	unsigned long flags;
> +
> +	spin_lock_irqsave(q->queue_lock, flags);
>  	__generic_unplug_device(q);
> -	spin_unlock_irq(q->queue_lock);
> +	spin_unlock_irqrestore(q->queue_lock, flags);
>  }
>  EXPORT_SYMBOL(generic_unplug_device);

The patch is not correct, some ->request_fn() functions may sleep. So
it's illegal to call generic_unlug_device() with interrupts disabled in
the first place, which is why spin_lock_irq() is used instead of the
flags saving variant.

So it looks like you need to fix your caller instead.

-- 
Jens Axboe