This problem was reported against 2.4 by Eddie.Williams@SteelEye.com

There's a problem in the sd spinup code in that if the unit returns NOT
READY, we begin to spin it up, but thereafter if it returns anything
other than NOT READY or success, the while loop in the spinup code will
be executed *without* the 1s delay that's in the NOT READY case.

The problem was seen with a real device: Compaq multi-path storage
arrays return NOT READY to probes down inactive paths, but when the
start unit is sent to activate the path, they can then respond back with
error conditions.

The fix is to terminate the while loop for any unexpected return.

(1)  second arg to scsi_queue_next_request() is only ever non-NULL
     inside scsi_lib.c and only used in the first conditional inside
     that function - ripped out of scsi_queue_next_request() and
     put into a new helper scsi_requeue_command().
(2)  Most remaining callers of are in the form
	scsi_put_command(cmd);
	scsi_queue_next_request(q, NULL);
     add a new helper, scsi_next_command() for them.
(2b) many callers of that still contain a repeated codepath, namely
     everything from scsi_release_request except the final kfree.
     New helper __scsi_release_request() for those.
(3)  All remaining callers loop over the devices of a host and call
     scsi_queue_next_request() on them - new helper
     scsi_run_host_queues().
(4)  scsi_queue_next_request() renamed to scsi_run_queue(), second
     arg is gone and it's static to scsi_lib.c now.

I have the new aacraid driver working on my system now.  The patch is
against the 2.5.66 updates that you gave me.  I made the following
changes:

aachba.c aac_scsi_cmd()
There was a race accessing the scsicmd pointer accessing the host_lock.
I made a local pointer to the Scsi_Host so the spin_lock_irq after
aac_read wouldn't panic.  I think that sometimes the I/O would be done
and the memory freed before returning invalidating the scsicmd pointer.
I made the same change in aac_io_done in case scsi_done had freed the
scsicmd memory before returning.

comminit.c aac_alloc_comm()
AdapterFibsVirtualAddress was set to the virtual address of base.  I
changed it to set it to the phys address.  I compared this to code
pointed to by matt domsch on the aacraid devel list on the 5th.  Its
aac_alloc_comm sets this variable to the phys address.  This fixed the
probelem where the entry->addr was bad.  Another was to fix it I guess
would be to leave this change alone and not try to convert the address
in aac_command_normal.

dpcsup.c aac_response_normal()
Changed the bus_to_virt to the calculation we talked about last month.
dpcsup.c aac_command_normal()
Changed the bus_to_virt to the calculation.

- stanford checker fixes (randy.dunlap)
- updated io_request_lock to correct 2.5 lock
- spelling fixes
- torvalds daemonize changes
- updated templates etc
- update scsicmd-> to scsicmd->device-> for new command alloc code
- update biosparam and add slave_configure
- gendisk name changes
- fix compile warnings

(a) scsi_check_device_busy() is unused now, kill it.  Btw, although I
    love to see this this really means we need to imply a
    scsi_set_device_offline (or even better scsi_set_host_offline)
    in scsi_remove_host now..
(b) make shost_class static to scsi_sysfs.c, with the new device model
    changes no LLDD needs this anymore
(c) move private prototypes where they belong.

BTW, Mike, did I miss something or will your changes make every driver
converted to scsi_add_host & co OOPS on removal now?

I added those two to factor out common code from the upper drivers
a long time ago, but after Doug & Lubens nice work there's nothing
left but incrementing/decrementing a counter in struct scsi_device
that's never used except in the case were we not it must be NULL
because we just walked the chain of drivers to detach every single
one..

Use the new dev_printk.

Thanks go to Mike Christie who originally
created this patch.

 ips.c |  135 +++++++++++++++++++++++++++---------------------------------------
 ips.h |    5 ++
 2 files changed, 62 insertions(+), 78 deletions(-)

This patch checks the return code of
pci_set_dma_mask for a 32bit mask,  adds a break
to quiet some compilers, and adds some 2.4 compat.
code.

 ips.c |    8 ++++++--
 ips.h |    4 +++-
 2 files changed, 9 insertions(+), 3 deletions(-)

This patch adds a missing kfree on an error path
and removes some cases where static variables
were being explicitly initiated to 0.

 ips.c |   13 ++++++-------
 1 files changed, 6 insertions(+), 7 deletions(-)

This is the proper way to report if the interrupt
was from a serveraid or not.

David Jeffery

 ips.c |   29 ++++++++++++++++-------------
 ips.h |    6 +++++-
 2 files changed, 21 insertions(+), 14 deletions(-)

into raven.il.steeleye.com:/home/jejb/BK/scsi-misc-2.5

Lost in the hch/andmike merge

-andmike
--
Michael Anderson
andmike@us.ibm.com


DESC
Change scsi sysfs to support scsi host class device and call release
functions when ref count goes to zero.
EDESC


 drivers/scsi/scsi_sysfs.c |  111 +++++++++++++++++++++++++++++++++++++++-------
 1 files changed, 96 insertions(+), 15 deletions(-)

-andmike
--
Michael Anderson
andmike@us.ibm.com


DESC
Change scsi host to class device model. Change scsi host and scsi device
to release when ref count goes to zero.
EDESC


 drivers/scsi/hosts.c     |   23 +++++------------------
 drivers/scsi/hosts.h     |   20 ++++++++++++++------
 drivers/scsi/scsi_scan.c |    4 +---
 3 files changed, 20 insertions(+), 27 deletions(-)

-andmike
--
Michael Anderson
andmike@us.ibm.com


DESC
This patch changes the structure of sdebug_host_info and changes the
method / order of driver model cleanup.
EDESC


 drivers/scsi/scsi_debug.c |  215 ++++++++++++++++++++++++----------------------
 drivers/scsi/scsi_debug.h |    2
 2 files changed, 115 insertions(+), 102 deletions(-)

-andmike
--
Michael Anderson
andmike@us.ibm.com

DESC
This patch removes the shost_devclass device class support that was
previously added, but incomplete.
EDESC


 drivers/acorn/scsi/acornscsi.c |    1 -
 drivers/acorn/scsi/arxescsi.c  |    1 -
 drivers/acorn/scsi/cumana_1.c  |    1 -
 drivers/acorn/scsi/cumana_2.c  |    1 -
 drivers/acorn/scsi/eesox.c     |    1 -
 drivers/acorn/scsi/oak.c       |    1 -
 drivers/acorn/scsi/powertec.c  |    1 -
 drivers/scsi/scsi_sysfs.c      |   23 -----------------------
 8 files changed, 30 deletions(-)

Oookay - I really got upset by that tmeplates in headers crap when
grepping for certain methods.  The patch below moves all templates
from the headers into the actual implementation files and removes
the ifdef mess for unmaintained drivers - for maintained support
gazillion kernel releases drivers like gdth I've kept them.

This means a driver works fine without any problems for all modular
builds and builtin kernel >= 2.4.0.  If you want certain drivers
to work with 2.0/2.2 statyic builds too I can hack something up for
you, but I'd prefer not supporting stuff like that anymore.

Tested by compiling all drivers with make -k and not getting more
warnings than before :)

into raven.il.steeleye.com:/home/jejb/BK/scsi-misc-2.5

into home.transmeta.com:/home/torvalds/v2.5/linux

into samba.org:/scratch/anton/linux-2.5_ppc64drivers

into samba.org:/scratch/anton/linux-2.5_ppc64

into samba.org:/scratch/anton/tmp3

- Remove some 2.4 compatibility macros
- Fix userspace pointer misuse in video1394 ioctl. Caught by the
  Stanford Checker. Gotta love the automated systems.
- Move our hotplug stuff around, to make the transition to putting our
  module dev table in mod_devicetable.h.
- Fix ohci1394 for possible crash with async stream packets. Cleaned up
  some debug messages.

This fixes two bugs introduced by some 2.5 changes:

- O_NDELAY handling typo in floppy_open()

- handling of failed transfers in floppy_end_request()
  (do equivalent of what 2.4 does)

Without first fix I was getting "floppy0: disk absent or changed during
operation" infinite loop on opening and without second fix, infinite loop
on error retry.

Now floppy driver seems to be (somehow) working :-).

This file was _the_ header for block-device related stuff in earlier
Linux versions, but nowdays there's just a few prototypes left that
really belong into blkdev.h or genhd.h (and in one case elevator.h).

This patch moves them over and removes everything but including
blkdev.h from blk.h  Note that blkdev.h gets all the headers that
were included in blk.h inmplicitly too.  Now we can start removing
all references to it an maybe kill it off before 2.6.  *sniff*

This adds ieee1394 for module table registration.

This fixes the compat_ioctl interface for the case where a NULL handler
is registered. This should produce a "compatible" as opposed to
"translated" interface for the specified ioctl. The patch was sent to
linux-kernel and no one complained (atleast with this second rev).

Currently __bdevname walks the obsolete list of block majors to
find a name for the given dev_t and falls back to unknown-block(%u,%u)
if that's not possible.  Replace this with an attempted get_gendisk() +
disk_name.  This means __bdevname can't be called from irq context
anymore, but as all old irq context callers are using bdevname() now
that fine (and I've added a big comment).

 - both rpciod_up and rpciod_down do a gratious inc/dec of the
   use count - but we can't ever be inside those function unless
   it's called from an other module -> totally useless
 - rpciod() (the kernel thread) also bumps the refcount when starting
   and decrements it when exiting.  but as a different module must
   initiate this using rpciod_up/rpciod_down this is again not needed.
   (except when a module does rpciod_up without a matching rpciod_down -
   but that a big bug anyway and we don't need to partially handle that
   using module refcounts).

into home.transmeta.com:/home/torvalds/v2.5/linux