Bart did this patch, I changed it a bit. Basically it cleans the mapping
interface up a bit, and adds a little helper to set up the request from
the bio.

In addition, it fixes a long standing bug where bio_map_user() would
call blk_queue_bounce() without the direction bit being set, auch.

 - Abstract out bio request preparation
 - Have bio_map_user() set data direction (fixes bug where blk_queue_bounce()
   is called without it set)
 - Split bio_map_user() in two

So we don't have to lookup the private data everytime.

Kill one user of mem_map_reserve/mem_map_unreserve I missed the
last time and it can go away.

	PPC logics for choice of default boot device number switched to
dev_t.  Rationale: it belongs to userland; we notice partitions with
certain properties and choose one of them as the best candidate for
being a root fs; then we put the resulting device number into ROOT_DEV,
which has only one use - it's passed to mknod(2) to create a device node
on rootfs and allow mount(2) get us the final root.

	IOW, all that code has nothing whatsoever with kernel internals
of any description - the value we are generating will be passed to
mknod(2) anyway.  Switched to dev_t.

	Note: that code should eventually be moved to early userland.

	* bogus calls of invalidate_buffers() gone from floppy_open()
	* invalidate_buffers() killed.

	* new helper - __invalidate_device(bdev, do_sync).  invalidate_device()
is calling it.

	* fixed races between floppy_open()/floppy_open and
floppy_open()/set_geometry():
	a) floppy_open()/floppy_release() is done under a semaphore.  That
closes the races between simultaneous open() on /dev/fd0foo and /dev/fd0bar.
	b) pointer to struct block_device is kept as long as floppy is
opened (per-drive, non-NULL when number of openers is non-zero, does not
contribute to block_device refcount).
	c) set_geometry() grabs the same semaphore and invalidates the
devices directly instead of messing with setting fake "it had changed"
and calling __check_disk_change().

	* __check_disk_change() killed - no remaining callers
	* full_check_disk_change() killed - ditto.

	capifs switched to ramfs-style tree and cleaned up; it's the same
changes that had been done its prototype (devpts) + stuff needed to
deal with unload (devpts is non-modular and always internally mounted).

	fbdev.node converted from kdev_t to int - all of its users
have register_framebuffer() which sets .node to mk_kdev(FB_MAJOR, index)
already called and all of them start with applying minor().  IOW, what
they actually want is framebuffer number.

	* type of ->node changed to int
	* register_framebuffer() sets it to index instead of mk_kdev(...)
	* users converted from minor(foo.node) to foo.node
	* useless assignments (typically to NODEV) removed - we never
look at that field before register_framebuffer() overwrites it and thus
any assignments prior to register_framebuffer() call are dead code.

	Console drivers cleanup.  In current tree interaction between
console and tty layer sits in the ->device() method of struct console.
It takes a pointer to console and returns device number of its tty
device.  open(2) on /dev/console goes through the list of registered
consoles, picks the first one that has ->device() and remaps the device
number to console->device(console).  Then it proceeds with normal
opening of tty.  This is the only caller of ->device().

	Cleanup: let ->device() return a pair (pointer to tty_driver, index
of tty in question) instead of device number.  Note that
	a) the first thing tty_open() does with remapped device number is
conversion to such pair.
	b) console driver _knows_ which tty_driver we want - one that
implements tty interface to the same physical device (i.e. the part of
the same driver).
	c) current code expects the result of ->device() to be a device
number of tty device - anything else is immediate -ENODEV from tty_open();
might as well have NULL ->device in that driver.

	Console drivers converted, (the only) caller updated.

	Preparation to console->device() cleanup: serial drivers converted
to common helper for their ->device() methods.

Don't disable the Northbridge Machine Check.

Use the unrolled "INTEL_USERCOPY" too.

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

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

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

into nuts.ninka.net:/home/davem/src/BK/net-2.5

Linus, could you please apply this patch so all thos poor devfs
users get their disks back?

The patch is obviously correct and has been floating on lkml for some time.

From Manfred Spraul:
 "My init_irq cleanup introduced a bug: on error, the function must return 1."

Looks like my finger fell off the control key..

Fix some warnings from the new code-patching stuff.

- Fix printk bug in the diagnostic code.

- If an error occurs, print the address of all the offending action
  handlers.  Also the symbol name if CONFIG_KALLSYMS.

The invalidate_buffers in ->release is wrong, get a reference to the
discipline in the beginning of ->open.

This one probably crept in during an ide merge from 2.4..

It looks like no one even tried to use it on 2.5..

Return an error code instead of a devfs_handle_t.  The handle isn't
useful for anything and the !CONFIG_DEVFS_FS stub in fact returned
NULL which made it entirely useless.  Thus only one driver is actually
checking the retval in the current tree..

Now that devfs_handle_t is gone from all structs there is no
reason to include it in headers.

Fix the fallout by including previously implicit headers and fixing
the drivers that didn't include devfs_fs_kernel.h explicitly.

They are obsfucating aliases for SetPageReserved/ClearPageReserved.
And once they're gone we can nuke <linux/wrapper.h>

And uninline end_request - it's calling to many functions to be useful
inline.

They don't have any code in common, so the initrd support can
go into a separate file and not require ramdisk support.

Lots of ifdefs gone and smaller kernel images for initrd users.

not be marked __init.

Noted by Petr Vandrovec