Make x86-64 boot again after the INITIAL_JIFFIES changes
and some bug fixes. Also some work for NPTL.

 - Merge with i386/2.5.64-bk3
 - Fix memory leak in copy_thread
 - arch_prctl uses GDT for base if possible. Cleanup.
 - clone supports bases >32bit for SETTLS etc. %fs hardcoded now.
 - new ptrace support for 64bit TLS
 - Disable (set|get)_thread_* for 64bit processes.
 - Audit arch/x86_64 for jiffies wrap issues.
 - Fix initial jiffies problem (that caused hanging kernels)
 - FIx a few 32bit emulation bugs (sigaltstack, sigqueue)
 - Some cleanup from Pavel
 - Should compile again as UP
 - Shrink size a bit by not putting exception tables into object files.
 - Fix compilation with gcc 3.3 :- force inlining when needed
 - Work around 2.5.64-bk3 console init bug.
 - Fix some alignments in assembly code

Rationale:  devfs_only does nothing but disabling {un,}register_blkdev
and {un,}register_chrdev.  {un,}register_blkdev already do nothing but
adding it's name argument to a lookup table for the __bdevname and
/proc/device output so this use is already bogus.  The disabling of
the character device per-major arrays can work in practice but is
useless as any driver relying on it can't be used on non-devfs systems.

Cleanup the i2c procfs code a bit (less ifdef mess), partially based
on the lm_sensors CVS code.

This fixes a bunch of "src" versus "obj" confusions that apparently
existed in my bootsect removal patch.  How embarrassing...

This recent patch caused Nick Piggin's 2xPIII VIA686B chipset machine into an
interrupts-off lockup during IDE probing.

We don't really know why - it might be because an interrupt is delivered to a
secondary which doesn't expect it.

I have a second patch from Zwane which solves the same problem in a different
way, but until that's had some wider testing I suggest we just back off the
original.

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

 - Fix my email address to one which reaches me ;-)
 - Remove timer.h include as we don't use timers.
 - Add module.h.
 - Sort includes alphabetically.
 - Move EXPORT_SYMBOL from ksyms.c to locks.c.
 - Simplify locks_conflict().

entry routine load the real ESP0 off that per-cpu stack. Make this
even faster by putting the sysenter stack in the per-CPU TSS, so
that we can use the tss->esp0 value directly (which we have to
update on task switches anyway).

CAREFUL! This needs very subtle code for debug and NMI exceptions,
to make sure we don't run with the sysenter stack in any real kernel
code!

use the real ones instead.

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

This patch is a update for the video-buf mm helper module.  It has
some minor bugfixes and a number of signed/unsigned cleanups to make
gcc 3.3 happy.

This allows us to avoid having to use atomic updates for the lazy FP
status setting, since we don't have to worry about other CPU's racing
on the fields.

Also, fix x86 FP state after fork() by making sure the FP is unlazied
_before_ we copy the state information. Otherwise, if a process did a
fork() while holding the FP state lazily in the registers, the child
would incorrectly unlazy bogus state.

the otherwise unused cpl1 entry for SS), so that we can avoid
re-loading it on task switches if it doesn't change.

If we raced on a timer expire, we'd get a negative timeout and think
that is was a _huge_ positive timeout. 

to speed it up - use the proper bitmask for clearing "used-fpu"
state.

Following some changes on x86-64.

When cpu_has_fxsr is defined to 1 like in many kernels unlazy_fpu can
collapse to three instructions. For that inlining is a very good idea.
Otherwise it's 10 instructions or so, which can be still inlined.

We don't need the lock prefix to test our local thread flags state.
Unfortunately test_thread_flag currently always uses test_bit which
has a LOCK on SMP, but that's unnecessary. LOCK is costly on P4,
so it's a good idea to avoid it.

Work around this for now by testing directly. Better would be
probably to define __set_bit for all architectures to not guarantee
atomicity and then always use that for local thread_info accesses
in linux/thread_info.h

Without this we have a choice between dropping lots of counter events
for counters > 0, or getting dazed and confused.  This brings it inline
with the 2.4 module code.  Tested on my 2-way.

Also fix a typo from Steven Cole, and remove some unnecessary code

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

into raven.il.steeleye.com:/home/jejb/BK/scsi-for-linus-2.5

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

From: Ingo Molnar <mingo@elte.hu>

This cleans up the NUMA scheduler to have separate triggers for idle
and busy rebalance ratio, preventing some complex interactions where
idle cpus can cause busy ones to busy rebalance more often. The core
functionality of this patch is rebalace_tick().

Tested in -mjb tree for over 1 month ... no problems at all found -
retested on top of 64-bk3 with new scheduler updates.

Remove the unneeded forward declaration of eepro100_remove_one().

It is defeating the __attribute__(unused) tag on the definition of
eepro100_remove_one() and is causing a "defined but not used" warning.

With CONFIG_SMP=n, CONFIG_PREEMPT=y, CONFIG_DEBUG_SPINLOCK=y we get many
undefined refs to kernel_flag.

This is because spinlock debugging works on uniprocessor kernels now, and it
stores state inside the spinlock structure to do this.

Having working spinlock debugging on UP kernels seems more important than
saving four bytes, so...

Patch from: "Martin J. Bligh" <mbligh@aracnet.com>

This tiny patch removes the new compiler warning from my build - the new
console_initcall mechanism seems to require int from console_init ...  I made
serial8250_console_init look like con_init

Patch from Ulrich Weigand <weigand@immd1.informatik.uni-erlangen.de>

Fix a race wherein timer_pending() returns false, but the timer handler is
actually running on another CPU.  We need to call del_timer_sync()
regardless, so we wait for completion of the handler.

missing spin_unlock() in sysfs_remove_dir()

Patch from Andries.Brouwer@cwi.nl

The next patch in the dev_t series eliminates the last applied use
of MAX_BLKDEV - only the definition in major.h remains.

Sneaky as I am, I combine this patch with the fix for an Oops:
On open, raw_open does
	filp->f_dentry->d_inode->i_mapping =
		bdev->bd_inode->i_mapping;
storing a pointer to bdev stuff.
But on release this pointer stayed, the block device is not
referenced anymore and disappears, and the next open references
undefined stuff.
I checked, and this can actually cause an Oops - scenario:

  # raw /dev/raw/raw12 /dev/hdf
  # dd if=/dev/raw/raw12 of=/dev/null bs=512 count=1
  # raw /dev/raw/raw12 0 0
  # dd if=/dev/raw/raw12 of=/dev/null bs=512 count=1

Oops.

More precisely the problem is that dentry_open does
file_ra_state_init(&f->f_ra, inode->i_mapping);
And file_ra_state_init uses mapping->backing_dev_info->ra_pages.
Ugly, to use so much information about the inode even before
the inode has been opened.

In the patch below I reset i_mapping upon release of the raw device.

akpm: I fixed a typo and exported default_backing_dev_info to GPL modules for
this.

This message is the source of 90% of the traffic on linux-smp and is now
starting to fill up kernel bugzilla.

Nobody has ever actually done anything about it and the message is wasting
people's time.  Ingo acked this change.

Patch from Zwane Mwaikambo <zwane@linuxpower.ca>

Scale the buffer which is used for accumulating the /proc/interrupts output.
On large SMP it overflows.

Patch from Jason Lunz <lunz@falooley.org>

When support for the GSET and SSET ethtool ioctls was added to eepro100.c in
2.4.20, the tx lock was overloaded to serialize their use.  Unfortunately,
this lock is only initialized in dev->open(), causing ethtool to deadlock the
machine when used on an unconfigured eepro100 interface.

The fix is to initialize the spinlock at probe time.

show_interrupts() is accidentally modifying the incoming seq_file pointer.
It goes oops if seq_printf() returns non-zero.

Patch from Zwane Mwaikambo <zwane@linuxpower.ca>

On SMP or preemptible kernels, every instance of show_interrupts() is oopsily
racy wrt request_irq() and free_irq().

Fix that up by taking the irq_desc_t's lock while walking the action list.

Patch from Andreas Gruenbacher <agruen@suse.de>

Tony Dziedzic has found two bugs in the extended attributes code. Patches with
explanations are attached.

1: Extended attribute sharing on ext2/ext3 not working

   The mb_cache_entry_insert function constantly returns an -EBUSY error
   instead of 0, which causes the xattr cache that is needed by the xattr
   sharing mechanism on ext2/ext3 to not share anything.  This patch fixes
   the problem.  (It is possible that after applying this fix we will hit
   bugs in code that wasn't used before.)

2: Oops in one of the xattr debug statements: The old_bh variable is NULL
   if an inode that previously had no EA's assigned would share an EA block
   with another inode.  (This was hidden by the xattr sharing bug).

- Fix an error-path leak of memory and a dentry in load_elf_binary()
  (spotted by Oleg Drokin).

- Tidy up the handling of bad executable images: return -ENOEXEC rather
  than 0.  (We've committed suicide anyway, so probably nobody gets to see
  this).

- Fix up the logic in load_aout_interp().  It gets itself all set up to
  return -ENOEXEC but there is in fact no path by which it can do this.  Just
  return ~0UL on errors, like load_elf_interp().

Fix a double-kfree on an ext2 error path.  Reported by Oleg Drokin.

Patch from Kevin Pedretti <pedretti@ieee.org>

The unmap_vmas() logic is designed to chew away at all the pages without
holding off preemption for too long.  But with CONFIG_SMP=y and
CONFIG_PREEMPT=y the number of pages which we batch up between rescheduling
opportunities is not a multiple of HPAGE_SIZE.  So unmap_vmas() ends up
calling unmap_hugepage_range() with a poorly aligned&sized region, and it
goes BUG.

Fix that up by ensuring that we always work across hugepage regions in
HPAGE_SIZE chunks.

Fix up some unchecked copy_to_user()s

Patch from: Shmulik Hen <hshmulik@intel.com>

The following patch fixes a division by zero bug in the bonding module that
happens when transmitting through a bond with no slaves, in the XOR bonding
mode.