From: Neil Horman <nhorman@redhat.com>
Date: Wed, 13 Aug 2008 16:09:05 -0400
Subject: [scsi] aic79xx: reset HBA on kdump kernel boot
Message-id: 20080813200905.GF439@devserv.devel.redhat.com
O-Subject: [RHEL 5.3] PATCH: aic79xx: reset HBA on kdump kernel boot
Bugzilla: 458620
RH-Acked-by: Vivek Goyal <vgoyal@redhat.com>

Hey all-
	We have a problem on some aic 79xx HBAs.  Its possible for them to be in
a paused state during a crash (for various and sundry reasons), when a crash
occurs.  Since the driver doesn't fully reset the HBA hardware on module load,
its possible for this state to not get cleared, and as a result the driver can't
issue any requests to attached devices during the kdump kernel boot, and as a
result we can't capture the vmcore, mount the rootfs, or do any of that other
good stuff that kdump would like to be able to do.

Upstream has (inadvertently) fixed this with a series of patches that has
reworked how the device detects the need for resets, and how it executes those
There was no directed effort toward this goal, but there is a series of commits
over the last several months to the aic79xx driver which changes the reset
behavior, causing the most recent version to "just work".

Given that its a series of 3 or 4 commits spread
over a long period with lots of interviening change, the proximity of the 5.3
kernel freeze deadline, and my profound lack of knoweldge in the scsi HBA arena,
I think it would be best if for 5.3 we instead fixed aic79xx with the patch
below. It isolates the change to only kdump kernel boots, and while it can cause
loss of in flight data we are guaranteed to not be accessing any devices on the
HBA's bus while this additional reset introduced by this patch is taking place.
We can do a proper backport for 5.4 from upstream when we have time to suck in
the actually commits, its dependencies and manage to verify that we're not going
to do anything catasrophic to the attached disks.

Regards
Neil

diff --git a/drivers/scsi/aic7xxx/aic79xx_pci.c b/drivers/scsi/aic7xxx/aic79xx_pci.c
index 328e38f..1285716 100644
--- a/drivers/scsi/aic7xxx/aic79xx_pci.c
+++ b/drivers/scsi/aic7xxx/aic79xx_pci.c
@@ -385,6 +385,18 @@ ahd_pci_config(struct ahd_softc *ahd, struct ahd_pci_identity *entry)
 	error = ahd_pci_map_int(ahd);
 	if (!error)
 		ahd->init_level++;
+
+	/*
+	 * If we are a kdump kernel rebooting the box, this controller was not
+	 * shut down properly, and as a result its possible for I/O operations
+	 * to be left in flight that can cause the device to stop responding
+	 * specifically the card can be in a paused state, and requests can
+	 * be queued to it prior to it being unpaused in ahd_resume, leading to
+	 * panic.  Handle this by resetting the card here, as we do in shutdown
+	 */
+	if (reset_devices)
+		error = ahd_reset(ahd, TRUE);
+
 	return error;
 }