Reduce the "rain forest" of pointers a bit.

- Use the channel info from the drive instead of from hwgroup where possible.
   Since we have now a direct association between a channel and irq, we can fix
   the handler for unexpected interrupts to don't have to iterate over multiple
   possibilities.

- Finally it was just possible to remove the circular list of channels
   associated with a hwif. The hwif member of struct hwgroup_s as well as
   next member if struct ata_channel are gone.

- Unify ide_build_sglist() and raw_build_sglist() in to one function.  Use the
   queue associated with a request in the case of a CMD request there instead of
   referencing the device minor number. This will be pushed even further later.

Some of the following is a bit of sensitive... but anyway it has to be done.

- Scarp "portability" macros ide_request_irq() and ide_free_irq().  What a
   great of a surprise in the face of a kernel nowadays ported to nearly any
   major CPU out there, but apparently no platform needs something special here.
   All of them are just using the normal request_irq() and free_irq(). If some
   unusual platform (read obsolete door blocker) I have overlooked intentionally
   needs something special - please just hack the platform specific
   request_irq() and free_irq()-  they are *far off* from being performance
   relevant. (Hint: You could use cooked default irq values to detect there
   whatever the caller was the ATA driver to decide what to do about it.)

- Kill ide_check_region() ide_request_region() and ide_release_region()
   altogether. Add comments explaining the code there.  Same arguments as above
   apply. Additionally we want to be able at some time to do for example mmapped
   PCI address space access on IA32 instead of port IO.

- Rewrite the initialization code to don't use check_region at all. Add some
   comments there, which explain what's actually happening.

- Fix creeping in of a c99 construct.

- Make the main interrupt handler take a channel as data instead of the
   hwgroup. This is the entity an interrupt is physically associated with. PCI
   interrupt sharing is handled in a generically anyway. Analogously the main
   data entity of a device request queue (q->queuedata) is the device and not
   the hwgroup.  However right now we are only able to push it down to the
   channel level. The intention is to advance the choose_drive() function to
   just check for equal interrupt values for serialization of request and
   thus to get rid of the hwgroup altogether.

- Streamline the usage of sector_t over the strategy routines in question a
   bit. Streamline the do_request code in ide-disk.c.

- Improve the readability of start_request in ide.c.

- Remove obsolete/nowhere used stuff from hdreg.h.

- Splitup special_t into classical flag field.

- Use only a single field to determine the capacity of a drive.  Make this
   field and the code paths it follows as far as possible use the sector_t
   instead of a hard coded integer types.  This increases the chances that at
   some distant point in time we will indeed be able to use 64 bit wide sector_t
   entities. (Disks are getting huge those times now...)

- Initialise the per-request_queue readahead parameter properly,
  rather than the dopey "if it's zero you get the deafult"
  approach.

- Permit zero-length readahead.

- 80-columnify mm/readahead.c

into penguin.transmeta.com:/home/penguin/torvalds/repositories/kernel/linux

into penguin.transmeta.com:/home/penguin/torvalds/repositories/kernel/linux

into tp1.ruhr-uni-bochum.de:/home/kai/kernel/v2.5/linux-2.5.make

Get rid of traces of the old-style ALL_MOBJS variable.

This also fixes the following issue, which could cause a warning with
certain configs:

Makefiles (legally) do:

obj-$(CONFIG_SND_AD1848) += snd-pcm.o snd-timer.o snd.o
obj-$(CONFIG_SND_CS4231) += snd-pcm.o snd-timer.o snd.o snd-rawmidi.o

so when the first option is set to y, the second to m, we have e.g.
snd-pcm.o on both $(obj-y) and $(obj-m). We correctly don't build
the modular version in this case.

However, if snd-pcm.o was a multi-part object, we did automatically
generate a link rule for both the modular and the built-in version, 
which caused a warning - that's now fixed.

 - switch md_error() to struct block_device * - just cleaning up
   after ->bi_dev switch.

 - bdevname() switched to struct block_device *.  Old variant (taking
   kdev_t) renamed to __bdevname() (very few callers remain).  This
   allow to drop ->b_dev conveniently - it's duplicated by ->b_bdev
   and most of remaining users were bdevname(bh->b_dev) in various
   places.

 - *NOW* all places that (re)assign ->bi_dev have relevant struct
   block_device *.  ->bi_bdev (struct block_device * equivalent of
   ->bi_dev) introduced, ->bi_dev removed, users updated.

 - block/loop.c - switch ->lo_device to struct block_device *.

 - (JFS) removes ->dev from struct jfs_log - it's duplicated by
   ->bdev.  Old users of ->dev updated.

 - md/linear.c - bring struct block_device * into private data.

 - this one is interesting and will play in the next series as well;
   affected place is fs/block_dev.c::do_open().  We check if bdev is
   a partition (same way it is done in generic_make_request()) and
   if it is - open entire disk and put pointer to its bdev into a
   new field bdev->bd_contains.  Otherwise (non-partition) we
   set bdev->bd_contains to bdev.  Corresponding cleanup done in
   blkdev_put() (and failure path in do_open()) - when the last opener
   goes away we close bdev->bd_contains if bdev is a partition (i.e.
   not equal to its ->bd_contains) and set it to NULL.
   Immediate effect is that we can get from bdev of partition to
   bdev of disk when submitting a bio, but it also opens a way
   to handle partition-parsing in a sane way.  That will be done
   in the next series.

 - md/raid1.c - bring struct block_device * into private data.

 - md/multipath.c convert to bio, compile fixes, bring struct
   block_device * into private data.

 - switch loop_set_fd() to struct block_device

 - switch blk_{get,set}_readahead() to struct block_device *

 - switch block_ioctl() to struct block_device *

 - new helpers: queue_hardsect_size() and bdev_hardsect_size().  Analogous
   to get_hardsect_size() for queue and struct block_device * respectively
   Most of get_hardsect_size() callers converted to these.

 - remove abuses of blksize_size[] and (late) hardsect_size[].  Mostly
   drivers/s390/char/*, BTW...

 - remove abuses of blksize_size[] and (late) hardsect_size[].  Mostly
   drivers/s390/char/*, BTW...

Remove unnecessary linux/bitops.h from fat/inode.c.

This patch cleans up fat_clusters_flush().

This patch does misc cleanup.

OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

This patch makes detection of a FAT filesystem more exact by checking
the first FAT entry. 

Currently fatfs disregards most wrong cluster chains, and only continues
processing. 

This fixes it.  If a wrong cluster chain is detected, change filesystem
to the read only, and return the error code. 

 - Revoke the TCQ stuff. Well having it for some time showed just nicely what
   has to be done before it can be included cleanly. But it's just not ready
   yet.

   For more explanations please simply track the usage of the special field of
   struct request - *both* in the generic request handling code and in overall
   driver code.

The ABC extensions are not in the official kernel, so remove references
to them.

No idea why that didn't hit me during testing ;-(

Forward port changes from 2.4.19-pre, originally by Keith Owens.
This driver used to build a module target called "aic7xxx_mod.o" and
introduced the MOD_TARGET variable to rename it to "aic7xxx.o" at
modules_install time.
    
Now, we just build "aic7xxx.o" directly, which becomes possible after
renaming aic7xxx.c to aic7xxx_core.c, as done in 2.4.19-pre.

Reduce cacheline bouncing when a dentry is in the cache.

Specifically, the d_count reference counter is not incremented and
decremented for every dentry in a path during path walking if the dentry
is in the dcache.  Execcisve atomic inc/dec's are expensive on SMP
systems due to the cachline bouncing.

into penguin.transmeta.com:/home/penguin/torvalds/repositories/kernel/linux