Date:	Fri, 12 Aug 2016 15:11:25 +0200
From:	Johannes Thumshirn <jthumshirn@...e.de>
To:	James Bottomley <jejb@...ux.vnet.ibm.com>
Cc:	"Martin K . Petersen" <martin.petersen@...cle.com>,
	Hannes Reinecke <hare@...e.de>,
	Linux SCSI Mailinglist <linux-scsi@...r.kernel.org>,
	Linux Kernel Mailinglist <linux-kernel@...r.kernel.org>,
	stable@...r.kernel.org, #@...e.de, v4.5+@...e.de
Subject: Re: [PATCH] SAS: use sas_rphy instead of sas_end_device to obtain
 address.

On Fri, Aug 12, 2016 at 12:08:54PM +0200, Johannes Thumshirn wrote:
> On Thu, Aug 11, 2016 at 11:00:07AM -0700, James Bottomley wrote:
> > On Thu, 2016-08-11 at 18:43 +0200, Johannes Thumshirn wrote:
> > > On Thu, Aug 11, 2016 at 08:09:35AM -0700, James Bottomley wrote:
> > > > On Thu, 2016-08-11 at 09:59 +0200, Johannes Thumshirn wrote:
> > > > > Since commit 3f8d6f2a0 ('ses: fix discovery of SATA devices in
> > > > > SAS
> > > > > enclosures') ses_match_to_enclosure() is calling
> > > > > sas_get_address(),
> > > > > which is coming from commit bcf508c13385 ('scsi_transport_sas:
> > > > > add
> > > > > function to get SAS endpoint address'). sas_get_address() itself 
> > > > > calls sas_sdev_to_rdev() which BUG_ON()s if a given scsi_device's
> > > > > rphy is not a SAS_END_DEVICE.
> > > > 
> > > > Is the BUG_ON the problem?  you're supposed to gate this call with
> > > > is_sas_attached().
> > > > 
> > > > > As SAS Enclosure is a SAS expander device,
> > > > 
> > > > This isn't necessarily true.  There are several separated enclosure
> > > > chips even in the SAS world (although most of the new ones are
> > > > integrated).
> > > 
> > > Maybe change it to "As a SAS enclosure can be a SAS expander device,
> > > [..]"
> > > 
> > > > 
> > > > >  we really shouldn't tie the lookup of a SAS address to the SAS
> > > > > End
> > > > > Device but the sas_rphy, which holds the address information.
> > > > 
> > > > This is conceptually wrong.  A wide end device may have many rphys
> > > > forming a port.  In that case the end device address is likely to
> > > > be
> > > > only one of the rphy addresses ... how do you know this code picks
> > > > the
> > > > right one?
> > > > 
> > > > > Fixes: 3f8d6f2a0 ('ses: fix discovery of SATA devices in SAS
> > > > > enclosures')
> > > > > Cc: stable@...r.kernel.org # v4.5+
> > > > 
> > > > What's the actual bug being fixed here?
> > > 
> > > This is the callchain I have:
> > > 
> > > ses_init()
> > > `-> scsi_register_interface()
> > >     `-> class_interface_register()
> > >         `-> ses_intf_add()
> > > 	    `-> ses_match_to_enclosure()
> > > 	        `-> sas_get_address()
> > > 		    `-> sas_sdev_to_rdev()
> > >                         `-> BUG_ON(rphy->identify.device_type !=
> > > SAS_END_DEVICE);
> > > 		  	    `-> KABOOM, Bug report, yelling release
> > > manager, etc..
> > 
> > So what is at the end?  is_sas_attached() indicates the transport class
> > attached to something and that something generated an sdev ... that can
> > only happen I think if that something is an end device.
> 
> sas_get_address() in ses_match_to_enclosure() _is_ guarded by
> is_sas_attached(), so it probably can not only happen if that
> something is an end device or the BUG_ON() wouldn't fire.
> 
> > 
> > > Unfortunately I do not have a SAS enclosure here so all I could do 
> > > was read the spec and code and have the customer test the patch.
> > > 
> > > But from my git archeology:
> > > Since 3f8d6f2a0 sas_match_to_enclosure() is calling
> > > sas_get_address(). Which in turn is calling sas_sdev_to_rdev() and so
> > > on...
> > > 	
> > > The thing is, in sas_get_address() we want to get the address of a 
> > > sas device, don't we? And looking at the UML diagram in the SAS spec, 
> > > we see an enclosure as well as an end device do have a rphy attached 
> > > to it, which in turn has a SAS address.
> > 
> > Because, as I said in the previous reply, the rphy is only the sas
> > address for non-wide ports.  They phy layer is just above the physical
> > layer.  The end device sits at the application layer, which is four
> > layers above that. Whatever diagram you're looking at is probably
> > showing port layer connections.  The way the transport class works is
> > that each rphy, even when part of a formed wide port, is individually
> > addressable, so we can send SMP phy controls to it, but the device is
> > separately addressable by its own SAS address.
> 
> Ok, we can't use the rphy because of wide-ports. We can't fix it to an
> end device either, as this makes some peoples systems unbootable. Now
> let's find a third option satisfying the needs of SAS wide-ports and
> my customers (and others running 4.5+ with a SAS enclosure).
> 
> I'm digging...


To answer myself, Hannes suggested doing it like this:

diff --git a/drivers/scsi/ses.c b/drivers/scsi/ses.c
index 53ef1cb6..1d82053 100644
--- a/drivers/scsi/ses.c
+++ b/drivers/scsi/ses.c
@@ -587,7 +587,7 @@ static void ses_match_to_enclosure(struct enclosure_device *edev,

        ses_enclosure_data_process(edev, to_scsi_device(edev->edev.parent), 0);

-       if (is_sas_attached(sdev))
+       if (scsi_is_sas_rphy(&sdev->sdev_gendev))
                efd.addr = sas_get_address(sdev);

		if (efd.addr) {


The reasoning behind this is, we only read the address if we have an
actual sas_rphy.

Would this be OK for you?

      Johannes

-- 
Johannes Thumshirn                                          Storage
jthumshirn@...e.de                                +49 911 74053 689
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: Felix Imendörffer, Jane Smithard, Graham Norton
HRB 21284 (AG Nürnberg)
Key fingerprint = EC38 9CAB C2C4 F25D 8600 D0D0 0393 969D 2D76 0850

Powered by blists - more mailing lists