list_del the starved_entry when sdev goes away.

Use GFP_ATOMIC when allocating starget, since we hold a lock.

Ensure that we cannot loop forever (however unlikely) when processing the
starved queues.

Get rid of the lock hierarchy for queue_lock and host_lock (even for the
single_lun case).

Change single_lun code to use a struct scsi_device *, so that we do not
need an sdev (or queue_lock) while checking if a single_lun target is in
use by a particular scsi_device.

James -

Patch against linux-scsi.bkbits.net/scsi-locking-2.5, per linux-scsi
discussion.

Rename scsi_check_sdev and scsi_check_shost to scsi_dev_queue_ready
and scsi_host_queue_ready, and change their return values.

===== drivers/scsi/scsi_lib.c 1.81 vs edited =====

Fix single_lun code for per-scsi_device queue_lock

Add and use a per scsi_device queue_lock.

Call scsi_alloc_queue each time we call scsi_alloc_sdev; call
scsi_free_queue each time we do not find backing LUN.

This code is less optimal, but leads to cleaner code, and the lock
split-up patch needs this change.

Cleanup and consolidate scsi_device and scsi_host checks in
scsi_request_fn.

Consolidate scsi single_lun code.

Add missing scsi_queue_next_request calls.

Add missing scsi_put_command and scsi_get_command exports.

Use a list_head per scsi_host to store a list of scsi request queues
that were "starved" (they were not able to send IO because of per host
limitations).

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

Fix more annoying compile problems due to wrong types
for comparing jiffies. This patch applies to alpha arch.

From Thomas Weyergraf.

Ivan.

- Check for parent PCI bus instead of bridge device to break the
  look in common_swizzle(). Functionally it's the same, but it's
  cleaner for PC-style host bridges (nautiluses).
- Generic PCI setup changes finally went in (thanks to rmk), so that
  FIXME in common_init_pci() can go.

Ivan.

It's possible to squeeze more bits out of lower half of pte,
but 32 seem to be a plenty...

Ivan.

into are.twiddle.net:/home/rth/BK/axp-2.5

into kroah.com:/home/greg/linux/BK/i2c-2.5

This is needed as we are still not using the driver core model for
matching up devices to drivers, but doing it by hand.  Once that is
changed, this will not be needed.

Still pretty ugly debug macro, but this at least
makes it do the right thing when used in if/else blocks

Fix from Joe Perches <joe@perches.com>

This patch adds selectable EdgeTriggered/LevelTriggered behaviour to
epoll.

It has been widely discussed on lkml about two weeks ago and everyone
very welcome the change.  It has been even more widely discussed through
private emails with application developers, that do not feel confortable
posting on lkml.  The great value of the patch is that selecting the LT
behaviour, applications using poll/select can be ported very easily to
epoll, making existing apps to benefit from epoll scalability with very
short ETA's.

The API remains the same with the addition of a EPOLLET event flag that
sets the LT/ET behaviour for that fd.

Many usb drivers use the usb_devfs_handle variable instead of just
adding the usb/ prefix directly to their devfs_register calls.  Fix
that and make usb_devfs_handle static and unexported.

.. by moving a bunch of devfs-related code from fs/partition/check.c
to fs/devfs/base.c.  Also has the nice sideffect of getting rid of
a bunch of ugly ifdefs.

[This is the new and improved, rediffed, applying and compilable
 version.  In short it's perfect]

From: Nikita Danilov, Maneesh Soni

d_lookup() is forgetting to release the lock when the rename race happens.

by having a cast followed by an initializer. It seems even gcc can't
do it right anyway in some versions (as reported by Jens Axboe).

+ C99 struct initialisers