Re: [PATCH 2/2] tracing/events: block: dev_t via driver core for plug and unplug events

[Bart Van Assche <Bart.VanAssche@wdc.com>]

On Tue, 2018-04-24 at 16:49 +0200, Steffen Maier wrote:
> The object life cycle seems to be:
> 
> (1) blk_alloc_queue() also creates gendisk kobj

I think this is a misinterpretation. blk_alloc_queue_node() initializes the
request queue kobj as follows:

	kobject_init(&q->kobj, &blk_queue_ktype);

register_disk() creates the /sys/class/block/<disk> node and /sys/block/<disk>
link as follows:

	if (device_add(ddev))
		return;
	if (!sysfs_deprecated) {
		err = sysfs_create_link(block_depr, &ddev->kobj,
					kobject_name(&ddev->kobj));
		if (err) {
			device_del(ddev);
			return;
		}
	}

> (1a) SCSI does LUN probing I/O
>      most of that is blktrace RWBS field value 'N' i.e. non-R/W with dev_t==0
>       since q->kobj.parent is still NULL we cannot get gendisk and thus dev_t
>      near the end is the first regular read block I/O with dev_t!=0
>       as bio!=NULL and bi_disk or rq_disk are non-NULL
> (2) blk_register_queue() also creates queue kobj with gendisk parent
>     now we can follow q->kobj.parent to gendisk to get dev_t,
>     e.g. for RWBS 'N' such as implicit SCSI test unit ready on blk dev close

Please have a look at the device_add_disk() call in sd_probe_async(). I think
that function triggers a call of blk_register_queue(). That last function
associates the request queue with the gendisk as follows:

	ret = kobject_add(&q->kobj, kobject_get(&dev->kobj), "%s", "queue");

> (3) optionally "regular" I/O
> (4) blk_unregister_queue() called by del_gendisk()
>     does kobject_del(&q->kobj) which NULLifies q->kobj.parent again
>     the last put of gendisk reference releases gendisk kobj
> (5) blk_cleanup_queue()
>     the last put of queue (/mq) reference releases queue (/mq) kobj

The /mq directory corresponds to q->mq_kobj. The block layer core drops its
reference to the queue kobject (q->kobj) by calling blk_put_queue() at the
end of blk_cleanup_queue().

> Since I cannot prove it's always like this, I followed Bart's suggestion
> and added another refcount get at (2) blk_register_queue().
> However, when I want to put that additional refcount at (5) blk_cleanup_queue(),
> we only get a queue pointer argument as context
> and q->kobj.parent==NULL since (4),
> so I don't know how to get to the gendisk object for the refcount put.

Have you tried to modify blk_register_queue() such that it stores the disk
pointer in a new struct request_queue member?

Thanks,

Bart.