Re: [PATCHv2 13/22] scsi_dh_alua: Use workqueue for RTPG

[Christoph Hellwig <hch@lst.de>]

On Wed, Jan 13, 2016 at 08:02:02AM +0100, Hannes Reinecke wrote:
>>> +	/* Check for existing port_group references */
>>> +	spin_lock(&h->pg_lock);
>>> +	if (h->pg) {
>>> +		old_pg = pg;
>>> +		/* port_group has changed. Update to new port group */
>>> +		if (h->pg != pg) {
>>> +			old_pg = h->pg;
>>> +			rcu_assign_pointer(h->pg, pg);
>>> +			pg_updated = true;
>>> +		}
>>> +	} else {
>>> +		rcu_assign_pointer(h->pg, pg);
>>> +		pg_updated = true;
>>> +	}
>>> +	alua_rtpg_queue(h->pg, sdev, NULL);
>>> +	spin_unlock(&h->pg_lock);
>>> +
>>> +	if (pg_updated)
>>> +		synchronize_rcu();
>>> +	if (old_pg) {
>>> +		if (old_pg->rtpg_sdev)
>>> +			flush_delayed_work(&old_pg->rtpg_work);
>>> +		kref_put(&old_pg->kref, release_port_group);
>>> +	}
>>
>> The synchronize_rcu() needs to be done in release_port_group, or even
>> better be replaced by doing a kfree_rcu there instead of a kfree.
>>
> Point is that we don't necessarily have an old_pg to call 
> release_port_group() on, but we still need to call synchronize_rcu()
> to inform everyong that h->pg now has a new value.

No, you don't.  synchronize_rcu() just waits for all previously
started RCU grace periods to finish.  If old_pg doesn't get freed
you don't need to wait for the grace period.  synchronize_rcu does
not notify anyone about anything, it just waits.

>> And unless I'm mistaken the flush_delayed_work should probably be
>> done in release_port_group as well.
>>
> _Actually_ we only need to call flush_delayed_work if sdev == rtgp_sdev. 
> Otherwise the workqueue item is running off a different device and won't be 
> affected.

I'm actually pretty sure I suggested that in an earlier iteratation,
but you said that's not the case.  So please only run it if
sdev == rtgp_sdev only then.