For some unknown reasons Nvidia NForce3 doesn't use the standard Hammer AGP architecture,
but requires set up of some shadow registers. This patch adds that to the K8 AGP driver.

Based on an old 2.4 patch from someone at Nvidia.

Also includes another bug fix for the K8 AGP handler, from Brad House.
We should not assume that there is only one northbridge in a Uniprocessor system.
Always flush all.

Also some minor cleanup.

Give 32bit emulation ioctl handlers the same locking rules as normal ioctl handlers.
This will avoid surprises in driver code.

Most call sys_ioctl who would take it anyways.

This adds a new completely rewritten machine check handler for x86-64.
The old one never worked on 2.6.

The new handler has many improvements. It closely follows the Intel and AMD
recommendations on MCE handlers now (the old one had many violations). It handles
unrecoverable errors in user space better now - it will only kill the process now
if possible instead of panicing.

This one is CPU independent now - it should work on any CPU that supports the standard
x86 MCA architecture.

This new handler only logs fatal errors that lead to kernel panic to the console.
Non fatal errors are logged race free into a new (non ring) buffer now
and supplied to the user using a new character device.  The old one could
deadlock on console and printk locks. This also separates machine check errors
from real kernel errors better. The new buffer has been also designed to
be easily accessible from external debugging tools: it has a signature
and could be even recovered after reboot. It is not organized as a ring buffer -
this means the first errors are kept unless explicitely cleared.

The new error formats can be parsed using ftp://ftp.suse.com/pub/people/ak/x86-64/mcelog.c
The new character device for it can be created with mknod /dev/mcelog c 10 227

There is a new sysfs interface to configure the machine check handler.
It has a "tolerant" parameter that defines the aggressiveness of the machine check:

0: always panic
1: panic if deadlock possible (e.g. MCE happened in the kernel)
2: try to avoid panic

Default is 2

Despite of having more features the new handler is shorter.

Bring the x86-64 port up to date. Lots of smaller bug fixes that have accumulated.
Also fixes another nasty bug introduced by the IA32e changes that causes BUGs at
boot for some people.

Only changes x86-64 specific files. There are some other changes that I'm sending
separately.

 - Some cleanup in NMI watchdog code
 - Fix HyperThreading CPU setup race (Suresh B. Siddha)
 - Update defconfig
 - Add a comment on why iommu_fullflush is disabled.
 - Export sys_ioctl again
 - Fix build with IA32_EMULATION=y and SYSVIPC=n
 - Remove noisy boot printks in the mptable scan.
 - Implement automatic NMI watchdog switching for real now
 - Remove redundant 32bit ioctl handlers for autofs
 - Remove CONFIG ifdefs around rtc 32bit ioctl handlers
 - Remove useless nfsctl ifdef in syscall.c (Al Viro)
 - Increase padding for prefetchw alternative
 - Check for NX bit early before setting up memory maps (Suresh B. Siddha)
 - Change Intel IA32e config description and fix help texts (Jun Nakajima)
 - Fix microcode driver build really now (Dave Jones)
 - Add nohpet option to disable HPET timer
 - Fix double semicolon in aperture.c
 - Add cmpxchg16b cpuid entry
 - Fix return value of read_pci_config_16 (Paul Menage)
 - Fix __KERNEL_COMPAT32_CS (Zachary Amsden)
 - Disable the infamous 30 minutes check in CMOS time setting
 - Update URLs in Kconfig (Petri T. Koistinen)
 - Fix ACPI interrupt source parsing for Nforce3 (Maciej W. Rozycki)
 - Fix 32bit ipc version parsing.
 - Run local APIC NMI watchdog only once a second (or less often on idle boxes)
 - Merge ACPI APIC SCI functions from i386
 - Add i8254 timer suspend code from i386
 - Merge with 2.6.2-rc3 + minor changes from i386
 - Fix empty_zero_page declaration (Greg Johnson)
 - Readd sysctls for exception/page fault trace and vsyscall32
 - Fix WCHAN
 - Fix STACK_TOP usage. Stack for 64bit processes should be at the
   top of memory now again. Also set it correctly for LINUX32_3GB.
 - Add warning fixes for gcc 3.4 and -Wdeclaration-after-statement

into ppc970.osdl.org:/home/torvalds/v2.5/linux

From: Andreas Gruenbacher <agruen@suse.de>,
      and me.

Latest gcc cvs is able to detect mismatches between functions which are
tagged asmlinkage and declarations which are missing asmlinkage.  Or vice
versa.

Fix up the fallout from an x86 allyesconfig build.

From: "Randy.Dunlap" <rddunlap@osdl.org>

Add syscalls.h, which contains prototypes for the kernel's system calls.
Replace open-coded declarations all over the place.  This patch found a
couple of prior bugs.  It appears to be more important with -mregparm=3 as we
discover more asmlinkage mismatches.

Some syscalls have arch-dependent arguments, so their prototypes are in the
arch-specific unistd.h.  Maybe it should have been asm/syscalls.h, but there
were already arch-specific syscall prototypes in asm/unistd.h...

Tested on x86, ia64, x86_64, ppc64, s390 and sparc64.  May cause
trivial-to-fix build breakage on other architectures.

into ppc970.osdl.org:/home/torvalds/v2.5/linux

into ppc970.osdl.org:/home/torvalds/v2.5/linux

into intel.com:/home/lenb/src/linux-acpi-test-2.6.4

drivers/acpi/sleep/proc.c:359: warning: initialization from incompatible pointer type
drivers/acpi/sleep/proc.c:367: warning: initialization from incompatible pointer type

drivers/acpi/utils.c: In function `acpi_evaluate_reference':
drivers/acpi/utils.c:353: warning: unsigned int format, different type arg (arg 5)

into intel.com:/home/lenb/src/linux-acpi-test-2.6.4

into nuts.davemloft.net:/disk1/BK/sparc-2.6

into intel.com:/home/lenb/src/linux-acpi-test-2.6.4

From: Pavel Machek <pavel@suse.cz>

swsusp/s3 assembly parts, and parts called from assembly are not properly
marked asmlinkage; that leads to double fault on resume when someone
compiles kernel with regparm.  Thanks go to Stefan Seyfried for discovering
this.

From: James Morris <jmorris@redhat.com>

ENOTSUPP is the wrong value, and should not be returned to userspace.

From: Armin <armin@melware.de>

Pointers to __devexit functions must be wrapped with the __devexit_p()
macro.

From: Ian Wienand <ianw@gelato.unsw.edu.au>

- Fix inline function declarations

- Use #ifdef for CONFIG_*, not #if

From: David Woodhouse <dwmw2@infradead.org>

Don't jump between contexts.

(don't write comprehensible changelogs, either).

From: Albert Cahalan <albert@users.sourceforge.net>

Now that /dev/pts is using the 12:20 dev_t, a new procps is required.

From: Martine Silbermann <Martine.Silbermann@hp.com>

Having spent a non trivial amount of time trying to pull in the code to
enable MSI, I would suggest that a clear indication in Kconfig that MSI
requires CONFIG_PCI_USE_VECTOR would be very helpful.  Also since the MSI
code was integrated into 2.6.1 I've updated the comment that called for
installing the MSI patch.

That check I just added to sys_sysctl() is not needed: do_sysctl() checks as
well.

From: Arjan van de Ven <arjanv@redhat.com>

One of my machines prints the following in dmesg during boot;

CPU: Trace cache: 4K uops<6>CPU: L2 cache: 256K

cause is a missing \n being printed; fix below.

From: Arjan van de Ven <arjanv@redhat.com>

some more hardcoded THREAD_SIZE cleanups.

From: James Bottomley <James.Bottomley@SteelEye.com>

The alder has an intel Extended Express System Support Controller which
presents apparently spurious BARs.  When the pci resource code tries to
reassign these BARs, the second IO-APIC gets disabled (with disastrous
consequences).

The first BAR is the actual IO-APIC, the remaining five bars seem to be
spurious resources, so we forcibly insert the first one into the resource
tree and clear all the others.

From: James Bottomley <James.Bottomley@SteelEye.com>

This is a necessary precursor patch for getting the Intel Alder motherboard
working (it has a PCI device corresponding to the IO-APIC which has to be
forcibly inserted into the machine's reserved memory region).

Eric Biederman was going to come up with a more comprehensive fix, but in
the meantime, this is the minimum necessary to get insert_resource to work
when the covering region is larger than the resource being inserted.

From: Tim Hockin <thockin@sun.com>

Attached is a simple patch to expose NGROUPS_MAX via sysctl.  Nothing
fancy, just a read-only variable.  glibc can use this to sysconf() the
value properly, so apps will stop relying on NGROUPS_MAX as a real
constant.