On Tue, 2015-09-22 at 22:15 +0200, Hannes Reinecke wrote:
> On 09/22/2015 09:49 PM, Ewan Milne wrote:
> > On Thu, 2015-08-27 at 14:41 +0200, Hannes Reinecke wrote:
> >> The current ALUA device_handler has two drawbacks:
> >> - We're sending a 'SET TARGET PORT GROUP' command to every LUN,
> >> disregarding the fact that several LUNs might be in a port group
> >> and will be automatically switched whenever _any_ LUN within
> >> that port group receives the command.
> >> - Whenever a LUN is in 'transitioning' mode we cannot block I/O
> >> to that LUN, instead the controller has to abort the command.
> >> This leads to increased traffic across the wire and heavy load
> >> on the controller during switchover.
> >
> > I'm not sure I understand what this means - why couldn't we block I/O?
> > and what does 'heavy load' mean? Aborting commands is 'heavy load'?
> >
> If we're getting a sense code indicating that the LUN is in
> transitioning _and_ we're blocking I/O we never ever send down I/Os to
> that driver anymore, so we cannot receive any sense codes indicating the
> transitioning is done.
> At the same time, every I/O we're sending down will be returned by the
> storage I/O with a sense code, requiring us to retry the command.
> Hence we're constantly retrying I/O.
Ah, OK. Perhaps including this explanation either in the comments with
patch 22/23 which adds the TEST UNIT READY commands to poll for the
status, or in the patch description somewhere would be helpful.
>
> [ .. ]
> >> @@ -811,10 +1088,17 @@ failed:
> >> static void alua_bus_detach(struct scsi_device *sdev)
> >> {
> >> struct alua_dh_data *h = sdev->handler_data;
> >> + struct alua_port_group *pg;
> >>
> >> - if (h->pg) {
> >> - kref_put(&h->pg->kref, release_port_group);
> >> - h->pg = NULL;
> >> + spin_lock(&h->pg_lock);
> >> + pg = h->pg;
> >> + rcu_assign_pointer(h->pg, NULL);
> >> + spin_unlock(&h->pg_lock);
> >> + synchronize_rcu();
> >> + if (pg) {
> >> + if (pg->rtpg_sdev)
> >> + flush_workqueue(pg->work_q);
> >> + kref_put(&pg->kref, release_port_group);
> >> }
> >> sdev->handler_data = NULL;
> >> kfree(h);
> >
> > So, you've already had a bit of discussion with Christoph about this,
> > the main portion of your ALUA rewrite, and I won't go over all of that,
> > except to say that I'd have to agree that having separate work queues
> > for the different RTPG/STPG functions and having them manipulate each
> > other's flags seems like we'd be better off having just one work
> > function that did everything. Less messy and easier to maintain.
> >
> > Also, it seems like wrong ordering of kref_get() vs. scsi_device_get(),
> > in alua_rtpg_queue() since they are released as kref_put() then
> > scsi_device_put()?
> >
> Yeah, I've reworked the reference counting.
> And reverted the workqueue handling to use the original model.
>
> Cheers,
>
> Hannes
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
