Add workaround for McKinley Erratum 7.

	accelerate normal system calls, but it puts the infrastructure in place
	and lets you write fsyscall-handlers to your hearts content.  A null system-
	call (such as getpid()) can now run in as little as 35 cycles!

	etc.

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

The linear target was getting the start sector wrong when doing a
dm_get_device(). [Kevin Corry]

The block layer does not honour bio->bi_size when issuing io, instead
it performs io to the complete bvecs.  This means we have to change
the bio splitting code slightly.

Given a bio we repeatedly apply one of the following three operations
until there is no more io left in the bio:

1) The remaining io does not cross an io/target boundary, so just
   create a clone and issue all of the io.

2) There are some bvecs at the start of the bio that are not split by
   a target boundary.  Create a clone for these bvecs only.

3) The first bvec needs splitting, use bio_alloc() to create *two*
   bios, one for the first half of the bvec, the other for the second
   half.  A bvec can never contain more than one boundary.

o  If there's an error you still need to call bio_endio with bio->bi_size
   as the 'done' param.

o  Simplify clone_endio.

[Kevin Corry]

Remove verbose debug message 'Splitting page'.

Remove some paranoia in highmem.c

Some fields in the duplicated bio weren't being set up properly in
__split_page(). [Kevin Corry]

md->pending was being incremented for each clone rather than just
once. [Kevin Corry]

dec_pending(): only bother spin locking if io->error is going to be
updated. [Kevin Corry]

Add a blk_run_queues() call to encourage pending io to flush
when we're doing a dm_suspend().

dm_suspend(): Stop holding the read lock around the while loop that
waits for pending io to complete.

queue_io() was checking the DMF_SUSPENDED flag rather than the new
DMF_BLOCK_IO flag.  This meant suspend could deadlock under load.

Give each device its own io mempool to avoid a potential
deadlock with stacked devices.  [HM + EJT]

There's a bug in the dm-stripe.c constructor failing top check if enough
destinations are handed in. [Heinz Mauelshagen]

minor change for dm-stripe.c. Tests for correct chunksize before it allocates
the stripe context. [Heinz Mauelshagen]

check_device_area was comparing the bytes with sectors.
[Stefan Lauterbach]

No need to validate the parameters if we are doing a
REMOVE_ALL command.

Return -ENOTBLK if lookup_device() finds the inode, but it
is not a block device. [Cristoph Hellwig]

An error value was not being checked correctly in open_dev().
[Kevin Corry]

Four constants:
   DM_DIR,
   DM_MAX_TYPE_NAME,
   DM_NAME_LEN,
   DM_UUID_LEN

Were being declared in device-mapper.h, these are all specific to
the ioctl interface, so they've been moved to dm-ioctl.h.  Nobody
in userland should ever include <linux/device-mapper.h> so remove
ifdef __KERNEL guards.

The writepage-removal patch broke filesystems which do not want to
support writeable mappings.

Fix that up by making those filesystems point their mmap vector at the
new generic_file_readonly_mmap().

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

This patch addes the new restart_block field support to m68knommu
thread_info.h.

This patch adds a definition for TASK_UNMAPPED_BASE in m68knommu
process.h.  Recent changes need a definition for this, although its
value is unsed for nommu targets.

The patch include cache.h in m68knommu hardirq.h.

Cleans up compile problems, and make its consistent with
all other architecture hardirq.h files.