RE: aic79xx U320 + e1000 Intel hangs on Idual Xeon 7505

["Justin T. Gibbs" <gibbs@scsiguy.com>]

> On Wed, 2003-04-09 at 21:27, Justin T. Gibbs wrote:
> 
> JG> The latest driver is 1.3.6:
> 
> I superimposed your 2.4-20030328 driver over my kernel tree and
> rebuilt.  It still locked up :-(

Do you have the nmi_watchdog enabled?  What bus speed are you running
for the aic7902 and the gig-E card?  Are the on the same physical PCI/PCI-X
bus?

> I tried lowering global tag depth to 4 (which I presume is a low number,
> but I don't really know what I'm doing).  And it still locks up.
> Moreover, according to /proc/scsi/aic79xx/[01], the driver has
> negotiated "Max Tagged Openings 0" with all the devices on this bus.

That seems really wierd - like you disabled disconnection.

> I noticed /proc/scsi/aic79xx/1 correctly refers to the controller as
> Channel B, but all the device info says Channel A.  Hope this doesn't
> mean it's getting scsi0 and scsi1 mixed up at a lower level.

Yes, that is a bit confusing.  The two channels are actually two independent,
single channel, controllers.  The user doesn't know that, and expects the
names to match those silk-screened on the card.  I'll review the code to
see if I can make it less confusing (perhaps just omit the channel identifier).

> My other theory is that all the devices on this bus are removable in one
> form or another, and hence are being polled for media changes.  The
> actions which cause the bus/driver to lock up are things which need a
> long period (several seconds) of data transfer - scanning in colour,
> writing a CD.  Could the disconnect logic be getting screwed up
> somewhere?  How could I test that?

A good start would be to send me privately (no need to spam the list) the
output of "cat /proc/scsi/aic79xx/*" and "cat /proc/scsi/scsi" as well
as a dmesg from the system.  From the last trace you sent, it did look
like we timed out while a command without the disconnection privledge was
out on the bus, but its not clear why yet.  If you compile the driver
with debugging enabled and a debug mask of 8, I can also see the content
of the serial eeprom to see if the settings are strange.

--
Justin