Fixes several oopsen in the SA-1111 "fake" PCI support

Complete re-write of the SA-1111 DMA bug "bounce buffer" workaround.

Merge latest drivers/pci/pool.c into mach-sa1100/pcipool.c (pool
allocation debugging follows CONFIG_DEBUG_SLAB a la drivers/pci/pool.c)

/arch/arm/mach-sa1100/pcipool.h can be deleted
(unrelated: /arch/arm/mach-sa1100/sa1111-ohci.c can be deleted)

Applies to 2.5.30-rmk1.  Should be back ported to 2.4 as the existing
SA-1111 bounce buffer code is broken.

(This patch is *required* for the OHCI HCD driver that is part of 2.5.21+.)

We now have an input layer driver for the keyboard/mouse, so this
isn't required.

must be located at the top level of the interrupt handler.

XScale does not guarantee that CPU control register writes complete
their side effects immediately.  In fact, Intel give sample code to
demonstrate a way to ensure that the effect of the write has occurred.
This is specific to Xscale, but the fix should work on any ARM CPU.
We therefore jump through the loops in "spirit" but not necessarily
in exactly the same way as the sample code.

Since a bootp image contains a zImage and an initrd (or maybe later
an initramfs) image, and the zImage is now PIC, we don't have to
copy the zImage to a specific address before calling it.  We just
call it where it got loaded into memory and let it sort itself out.

Make was missing the need to rebuild the kernel decompressor when
files in arch/arm/boot/compressed are changed.  This is because
make believes that the object contained within is up to date
(because it doesn't know any better.)  This cset always forces us
to decend into these two directories if the objects inside these
are required.

This cset fixes a set of problems discovered while developing KLIBC
with Thumb support.  We now allow pure Thumb executables, and prevent
such executables from being run on non-Thumb code aware CPUs.
We also fix a bug in the fixup of Thumb late aborts which rather
messed things up.

On context switch, when the MM hasn't changed, we can avoid calling
the architecture specific area.

When CONFIG_DISCONTIGMEM=n, we define PHYS_TO_NID(x) to zero in each
architecture specific file.  This cset moves it into the generic
ARM code.  

Add support for Thumb stubs (for klibc.)
We drop a work-around for early gcc versions where register
variables didn't work as expected; this allows us to get rid
of several lines of ASM per syscall.

but some files still did (and got the wrong results on UP).

Since they didn't actually _use_ the BKL anyway, remove the
include.

since it was generic.

Remove all architecture-specific <asm/smplock.h> files.

 - Use req->rq_received to determine the message length instead of
   assuming that it goes to the end of the page.

 - If the server returned an illegal record so that we cannot make
   progress by retrying the request on a fresh page, truncate the
   entire listing and return a syslog error.

 - Fix the check for whether or not the received message length has
   somehow been truncated: we need to use req->rq_received rather
   than the receive buffer length (req->rq_rlen).

 - Ensure that we set res->eof correctly. In particular, we need to
   clear it if we find ourselves attempting to recover from a
   truncated READ.

 - Don't set PageUptodate() on those pages that are the victim of
   message truncation.

 - cleanup #defines: I do not follow the rationale behind the
   odd line-wrapped defines at the beginning of the file.  If
   we have to use multiple lines, then we might as well do so
   cleanly and according to normal practice...

 - Remove a level of indirection: do not have spin_lock_foo
   use spin_lock - just explicitly call what is needed.

 - we do not need to define the spin_lock functions twice, once
   for CONFIG_PREEMPT and once for !CONFIG_PREEMPT.  Defining
   them once with the preempt macros will optimize away fine.

 - cleanup preempt.h too

 - other misc. cleanup, improved comments, reordering, etc.

This fixes the ptrace wait4() anomaly that can be observed in any
previous Linux kernel i could get my hands at.

If the parent still has other children (that are being traced by
somebody), we wait for them or return immediately without an error in
case of WNOHANG.

Finally, this chunk removes the references to the UFS & ROMFS
entries in struct superblock, leaving just ext3 and hpfs as
the only remaining fs's to be fixed up.

This one from Brian Gerst seperates UFS from
the struct superblock union.

This patch from Christoph Hellwig divorces ROMFS
from the struct superblock union, as has been done
to various other filesystems during 2.5

try_to_unmap() dies if the to-be-unmapped page has a non-NULL ->mapping.

But the preceding logic in shrink_cache() forgot about the
rarely-occurring pages which have buffers and no ->mapping.

I have a test_and_set_bit(PG_chainlock, page->flags) in page reclaim.
Which works fine on SMP.  But on uniprocessor, we made
pte_chain_unlock() a no-op, so all pages end up with PG_chainlock set.
refill_inactive() cannot move any pages onto the inactive list and the
machine dies.

The patch removes the test_and_set_bit optimisation in there and just
uses pte_chain_lock().  If we want that (dubious) optimisation back
then let's do it right and create pte_chain_trylock().

__page_cache_release() needs to recheck the page count inside the LRU
lock, because shrink_cache() may have found the page on the LRU and
incremented its refcount again.

Which is carefully documented over __pagevec_release().  Duh.

Here's a patch from Andrea's tree that uses IRQ_BALANCE_INTERVAL to
define how often interrupts are balanced, staying independent from HZ.

  It also makes sure that there _is_ a change to the configuration
before it actually writes it.  It reminds me of the mod_timer
optimization.

the attached patch updates a number of items:

 - adds cleanups suggested by Christoph Hellwig: needed unlikely()
   statements, a superfluous #define and line length problems.

 - splits up the global ptrace list into per-task ptrace lists. This was
   pretty straightforward, and this makes the worst-case exit() latency
   O(nr_children).

the per-task ptrace lists unearthed a bug that the previous code did not
take care of: tasks on the ptrace list have to be correctly reparented as
well. This patch passed my stresstests as well.

oh, setup.S. nasty indeed, bogus GDT limit.

This updates the CLONE_CLEARTID case to use futexes to make it easier
to wait for a thread exit.

glibc/pthreads had been updated to use the TID-futex, this removes an
extra system-call and it also simplifies the pthread_join() code.  The
pthreads testcode works just fine with the new kernel and does not work
with a kernel that does not do the futex wakeup, so it's working fine.

and the call trace instead.

This is the second part of "broken cfb* support in the 2.5.31-bk".  I
needed fbcon-cfb2 on one of my systems, and so I went through all
fbcon-* drivers and fixed them.

line_length, type, type_aux and visual were moved from display to
fb_info in last James Simmon's fbdev update.  Unfortunately lowlevel
support modules were not updated.

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

I want to find a way to get all those PageReserved tests out of
there, because they are very expensive.  But now is not the
time.

Make sure that the VC tty pointer is cleared when the tty is free'd.

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

into samba.org:/scratch/anton/linux-2.5_ppc64drivers

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

into groveronline.com:/root/bk/linux-acpi

into groveronline.com:/root/bk/linux-acpi

into hostme.bitkeeper.com:/ua/repos/l/linux-acpi/linux-acpi

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