NWFPE performance improvements, part 6.  Builds on patch 1468/1.

Break NWFPE initialization into separate function and call it from
entry.S as appropriate.  Also avoid nRc variable usage in EmulateAll().
This buys 6% speedup in NWFPE.  (Note that this is over half of the
total speedup between original RMK and the netwinder cvs version).

(note that 2.5 already has the separate initialisation, and this
cset actually makes 2.5 conform to 2.4 function naming for this
part. --rmk)

NWFPE performance improvements, part 5.  Builds on patch 1467/1.

Inline the functions getTransferLength(), getRegisterCount(),
and getRoundingPrecision().  This buys 1.6% speed gain.

Fourth part of NWFPE performance improvements, building on patch 1466/1.
 
This patch makes 80-bit (extended precision) a compile-time option in NWFPE.
The effect of the extra code path on 32/64 bit operations is not as significant
as I had first thought - speedup is between 0.7 and 1.5% depending on which
tree (rmk or nw) you do the changes in.  However the reduction in code size,
particularly when softfloat has the same functions cut out, may be of value
on embedded systems.

NWFPE performance improvements, part three.

A minor simplification - as (opcode & MASK_ARITHMETIC_OPCODE) >> 20
is frequently used, compute that quantity once and store it.

Second part of NWFPE improvements.  Builds on top of patch 1464/1.

Remove unused code.  And fix a function prototype to match function.
No impact on generated code or NWFPE performance.

This is the first of a series of patches to NWFPE, which aim to bring
the performance improvements from the netwinder.org CVS tree into the
mainline ARM kernel.

This patch is merely a reformatting of the NWFPE sources.  All files
were processed with "indent -kr -i8 -ts8 -sob -l132 -ss" and a few
manual fixups. Exception: the softfloat files have not been touched.

This patch should be applied after the 5 patches previously published
by RMK on linux-arm-kernel have been applied.

Slightly better performance can be obtained by eliminating Fd and
calculating it after the math operation.

Eliminate getFd from {Single,Double,Extended}CPDO by passing a
pointer to the destination register into these functions.  Use this
same pointer when converting the destination register to the required
type.

Extending the idea in patch 2, we can eliminate the setting of fType
in {Single,Double,Extended}CPDO functions and always set it in
fpa11_cpdo.c to the ultimate destination size.

SingleCPDO, DoubleCPDO and ExtendedCPDO always set the destination
register type to their operating type (single, double, extended
respectively) when they complete successfully (nRc != 0).  This
means that we know what size result will be generated, and we don't
have to fetch it from the fType array.

In fact, since we know "nType" will be the type of the operand,
we don't even need to calculate value of Fd unless we need to
convert the result to the destination type.

This eliminates some unnecessary branches and instructions in the
TLB flush macros where configurations allow for this to occur.

Kill ide_* resource handling.  Add __ide_mm* macros to use our string
based mmio.

We include a header file in an indirect method, which mkdep didn't
understand.  Now that mkdep is gone, we don't need the hack to make
mkdep do the right thing.

This cuts down on the number of warnings during the build.

ecard.c used to use the reboot notifier to ensure that the expansion
cards are in the correct state for rebooting.  Use the driver model
shutdown method instead.

Acorn Ethernet drivers provided a duplicate set_mac_address
implementation which reflects the generic ethernet version in
drivers/net/net_init.c.  Use the generic version instead.

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

We had a dummy dependency on include/linux/compile.h, but it really caused
more trouble than benefits. It's not actually needed for the module
postprocessing, it was only put there to make sure we recognize when
gcc changed under us. However, we really can only do so much, and
the rest of kbuild won't notice a changed gcc either, so if the user replaces
gcc during a build, he just can't rely on the build doing the right thing.

The common cases are still covered, anyway. When the command to invoke
gcc changes ("CC=gcc32") we notice, and when the path to

	/usr/lib/gcc-lib/i386-redhat-linux/2.96/include/stdarg.h

changes (which luckily contains the "2.96", we'll notice and handle that,
too.

1) Remove dep from "make help", it is no longer useful
2) replace Generating with GEN when generating version.h

When a rule in the top-level Makefile includes scripts as one
of the prerequisites it inherits FORCE, and thus is always build.
include/linux/autoconf.h recently included scripts hereby forcing
split-include to be run for each compilation.

Fix all rules that lists scripts as a prerequisite but did not list FORCE.
Fixed by listing the executable needed direct.

We don't really have a nice way to say "compile this when CONFIG_FOO
is y, don't otherwise".

Alternatives are:

obj-$(subst m,,$(CONFIG_FOO)) := foo.o

or

obj-$(CONFIG_FOO)	:= foo.o
obj-m :=

or

obj-y			:= do_foo.o
do_foo-$(CONFIG_FOO)	:= foo.o

I chose the last one, though I'm not particularly happy with either.

This patch fixes what might have been a joke, but wasn't.

 Coverted -> Converted

One down, 285 to go.

Patch from Stephen Cameron <steve.cameron@hp.com>

Add new big passthrough ioctl to allow large buffers.  Used by e.g.  online
array controller firmware flash utility.

Patch from Stephen Cameron <steve.cameron@hp.com>

Add CCISS_GETLUNINFO ioctl.

This ioctl returns the LUNID, number of partitions, and current number of
opens on a logical volume.  Used by the array config utility or any app that
needs to send passthrough commands to a particular logical disk.

Patch from Stephen Cameron <steve.cameron@hp.com>

* add cmd_type to sendcmd parameters to allow scsi messages to be
  sent down.
* factor out duplicated code into fill_cmd function.

Patch from Stephen Cameron <steve.cameron@hp.com>

Make cciss driver retry 3rd party bus reset aborted commands up to 3 times.
(ported to 2.5 by me, original patch by Charles White)

This is needed for a multi port storage box that can have multiple hosts
connected to it, or be used in a multipath configuration.  In certain
configurations SCSI bus resets initiated by one host may affect another host.

Patch from Stephen Cameron <steve.cameron@hp.com>


* Make cciss driver wait longer for board to enter simple mode to
  handle an unlikely corner case.  (Hot replace of failed 144GB
  drive in RAID 5 set at just the wrong time prior to driver loading
  can make the board take a longer than usual time to go into "simple
  mode".  Without the patch, the driver gives up too early, and consequently
  doesn't work.  (A reboot will generally "fix" it.)  This patch avoids
  the problem.
* Fix a couple ioctls to return EAGAIN instead of inappropriate EFAULT.

Patch from Rohit Seth

ia64 reserves virtual address region 4 (any address which starts with 0x4)
for huge pages.

Apparently, for hardware reasons, we do not wish to allow mappings with
other page sizes to appear in that region.

This patch prevents the user from being able to place regular MAP_FIXED
mappings into region 4 on ia64.  It is a no-op for ia32.

Patch from Rohit Seth.  Updates the hugetlb page documentation.

Patch from Rohit Seth.  Remove a dead declaration.

Patch from Christoph Hellwig <hch@sgi.com>

There's a bunch of minor fixes needed to disable the swap code for systems
with mmu.

Patch from "Vitezslav Samel" <samel@mail.cz>

Obvious fix of printk level (wrong since 2.5.62).