We actually fix this race by mediating the object release/get race
(i.e. we destroy the scsi_disk object when its reference count goes
1->0, we use a semaphore to prevent something else trying to get a
reference after or during this).

The open/remove race is actually irrelevant because even if we open an
already removed object, all that will happen is that we get a
reference to a device that always returns EIO.

the SPI transport class DV should check the data we derive from the
inquiry to see if the device is capable of supporting wide/sync before
trying to validate the settings.

This fixes 

http://bugme.osdl.org/show_bug.cgi?id=2400

sym 2.1.18j:
 - Add SPI transport attributes (James Bottomley)
 - Use generic code to do Domain Validation (James Bottomley)
 - Stop using scsi_to_pci_dma_dir() (Christoph Hellwig)
 - Change some constants to their symbolic names (Grant Grundler)
 - Handle a race between a postponed command completing and the EH retrying
   it (James Bottomley)
 - If the auto request sense fails, issue a device reset (James Bottomley)

Modification of patch from SLES-9

some SCSI devices need more than 64bytes of sense buffer.
I know about one: The IBM MagStar tapes report the necessity to be
cleaned at bytes 70 and report 96 bytes in total.

Attached patch increases the sense buffer size to 96 bytes.

sd_remove calls sd_shutdown to finish I/O to the disc.  However, the
state model puts the device in SDEV_DEL before triggering the
sd_remove (which won't allow any further I/O at all).

Fix by making SDEV_CANCEL the intermediate state and only transitioning to 
SDEV_DEL after calling device_del().

From: Jens Axboe <axboe@suse.de>, Mike Christie <mikenc@us.ibm.com>

This still doesn't give us complete fast fail on
the eh path because we still get delayed to recover
the transport.

This small patch fixes the template size calculation in
{spi,fc}_attach_transport().

ith the completion of testing at Adaptec and IBM, the latest version of
the ServeRAID Linux device driver ( ips Version 7.00.15 )is now
available. Attached is the patches against the previous ServeRAID
releases.  Basically, this is a convergence to a common ips driver code
base for the 2.4 and 2.6 kernels.  

This driver is based on the driver currently in 2.6, so you while there
appears to be lots of changes to the 2.4 driver,  the changes to the
current 2.6 driver are trivial ( the only changes are  version number
changes and the addition of some new adapter ID's ).  For 2.4, this was
primarily a testing effort to ensure that the 2.6 driver code was stable
when built for a 2.4 kernel. 

From: "Moore, Eric Dean" <Emoore@lsil.com>

Changelog for mpt version 3.01.03:

(1)  Fix fifo memory allocation under 64bit systems
by merging 3 seperate memory allocations into one call.
Before this fix, it was possible that these seperate 
pci_alloc_consistent() calls were crossing the 4GB 
hardware boundary.
Delete the MPTBASE_MEM_ALLOC_FIFO_FIX logic.
Backout recent ak@muc.de workaround.

(2) Replace wrappers for CHIPREG_XXX with defines;
per request from Jeff Garzik [jgarzik@pobox.com]

(3) Remove support for FC909.

(4) Remove PortIo modules parameter.

(5) Move procmpt_destroy function before pci_unregister_driver,
as the memory allocated for the proc was not being deallocated.

(6) Remove mptscshi_reset_timeouts function. The timer was
already expired when mod_timer is called.

(7) Fix small bug in slave_destroy, which could prevent domain
validation on hidden drive in a RAID volume.

----------------------------------------------------------------------

Changelog for mpt version 3.01.02:

(1) Andi Kleen[ak@suse.de]
put warning "Device (0:0:0) reported QUEUE_FULL!" into debug messages

(2) Alexander Stohr[Alexander.Stohr@gmx.de]
fix warnings from mptscsih_setup when driver isn't compiled as module

(3) Randy.Dunlap[rddunlap@osdl.org]
Remove unnecessary min/max macros and change calls to 
use kernel.h macros instead.

Here is a patch that adds FC transport attributes to the qla2xxx driver.
The patch should apply cleanly to a recent 2.6 kernel with the FC
attributes update that I just sent to linux-scsi.

This patch brings the FC transport attributes up to date with the SPI
attributes API.

The only major change I made was to separate read-only attributes
from read-write attributes.

From: Arjan van de Ven <arjanv@redhat.com>

aic7xxx_old: constant array of PCI ID's can be static since it's read only

Following testing in more extreme situations, the following problems
turned up:

- The error handler can offline the device during DV (most particularly
true when transport parameters are undetectably mismatched).  Fixed by
modifying the state model to allow this and then having DV set the
device back online for the retry.

- DV needs to be serialised.  Fixed by introducing a per device
semaphore.

- Cosmetically, it's nice to trigger DV from userland, so added a
revalidate sysfs entry.

into mulgrave.(none):/home/jejb/BK/scsi-misc-2.6

fusion needs several separately allocated coherent regions and
requires that they all be in the same 4GB segment.  Obviously this may
fail.  The hack is to force the coherent_dma_mask to 0xffffffff thus
ensuring that all the allocations occur within the first 4GB.  This
hack breaks Altix entirely.

Work around the gcc-2.95 token pasting bug.

Attached is a patch to fix an oops in sg_cmd_done. Please apply.

remove dependency on kobj.name.

into ppc970.osdl.org:/home/torvalds/v2.6/linux

into ppc970.osdl.org:/home/torvalds/v2.6/linux

into kernel.bkbits.net:/home/davem/net-2.6

into ppc970.osdl.org:/home/torvalds/v2.6/linux

into ppc970.osdl.org:/home/torvalds/v2.6/linux

This rolls up Marc Zyngier's EISA correction (first two) and adds a missed
netdev_priv() conversion that was causing an oops on module removal.

Noticed by Jan Glauber, confirmed by Stephen Hemminger.

into ppc970.osdl.org:/home/torvalds/v2.6/linux

into kroah.com:/home/greg/linux/BK/pci-2.6

into ppc970.osdl.org:/home/torvalds/v2.6/linux

Call pci_claim_resources() so we can see what PCI resources are being used.

On ia64, the parent resources are not necessarily PCI resources and
so won't get found by pci_find_parent_resource.  Use the shiny new
insert_resource() function instead, which I think we would have used
here had it been available at the time.

If we start again, we can return an error even if we were successful.
Reset the result to 0 before beginning again.  Why don't we use a
tailcall here?