Latest premption patch.

Please find attached a small hugetlb page support update for 2.5.60
IA-64 kernel (+ my previous hugetlb page patch).  This patch adds the
HAVE_ARCH_HUGETLB_UNMAPPED_AREA define for IA-64 arch.
hugetlb_get_unmapped_area function is added in
arch/ia64/mm/hugetlbpage.c

Please find attached a hugetlb page patch for IA-64 2.5.60 kernel .
Changes in generic files are mostly backported from 2.5.62 (to get
hugetlb support functioning properly).

And here's the SN specific part of the update.  This should get an SN2
compile all the way to the link stage, where I still have some devfs
stuff to cleanup.

Here's an update to the ia64 mmzone.h file that adds some SN2 specific
defines (which will probably have to be changed later, but it's a
start).

IA64 ABI specifies that globals > 8 bytes need to be aligned to 16 bytes.
gcc doesn't follow this convention. Current kernel code will fail to work with
a compiler which follows the ABI.

size of structure ia64_machine_vector is > 8 bytes and not multiple of 16 bytes.
When we have CONFIG_IA64_GENERIC, each machine specific vector from different
object files gets linked into a user defined section(forming array of
structures).

Now with a compiler conforming to ABI, there will be holes in this array
resulting in the code failure when it goes through this array of structures.

Following patch will make the size of the structure to be multiple of 16 bytes
making both ABI confirming/non-conforming compilers happy.

Current 2.5 kernels are broken with CONFIG_IA64_GENERIC option. Two unused
members are still remaining in the structure, resulting in corruption of
structure elements during the MACHVEC_INIT initialization. Attached patch
fixes this problem.

Rationalize __init/__devinit attributes.  The noteworthy changes are
that
iosapic_system_init(),
iosapic_init(),
iosapic_register_platform_intr(), and
iosapic_override_isa_irq()
are __init (only called from ACPI __init functions), but
iosapic_lists[],
num_iosapic,
find_iosapic(),
register_intr(),
iosapic_register_intr(), and
acpi_register_irq()
are not because they may be used after init-time by modules.

More detailed analysis:

        iosapic_lists[], num_iosapic: normal, referenced by
                find_iosapic (normal)
                register_intr (normal)
                iosapic_init (__init)

        pcat_compat: __initdata, referenced by
                iosapic_system_init (__init)
                iosapic_init (__init)
                iosapic_parse_prt (__init)

        find_iosapic: normal, called by
                register_intr (normal)

        register_intr: normal, called by
                iosapic_register_intr (normal), called by
                        acpi_register_irq (normal), called by
                                modules (=> can't be __init or __devinit)
                iosapic_register_platform_intr (__init)
                iosapic_override_isa_irq (__init)
                iosapic_parse_prt (__init)

        iosapic_reassign_vector: __init, called by
                iosapic_register_platform_intr (__init), called by
                        acpi_parse_plat_int_src (__init)

        iosapic_system_init: __init, called by
                acpi_parse_madt (__init)

        iosapic_init: __init, called by
                acpi_parse_iosapic (__init)

        iosapic_register_platform_intr: __init, called by
                acpi_parse_plat_int_src (__init)

        iosapic_override_isa_irq: __init, called by
                acpi_parse_int_src_ovr (__init)
                iosapic_init (__init)

        fixup_vector: __init, called by
                iosapic_parse_prt (__init), called by
                        acpi_pci_irq_init (__init)

Make interrupt registration functions take named constants for
polarity and trigger mode.  Old -> new magic decoder ring:
polarity 0 -> IOSAPIC_POL_LOW(#defined to 1)
polarity 1 -> IOSAPIC_POL_HIGH(#defined to 0)
trigger 0  -> IOSAPIC_LEVEL(#defined to 1)
trigger 1  -> IOSAPIC_EDGE(#defined to 0)

Simplify ISA IRQ init by taking advantage of iosapic_override_isa_irq(),
which already does what we need.

emove IOSAPIC address and GSI base from external interrupt
registration interfaces.  This lets us remove acpi_find_iosapic(),
which is functionally similar to find_iosapic().

Make pcat_compat a system property, not a per-IOSAPIC property.

	and reformat to make it fit in 100 columns.

	Keith Owens.

This patch has been running inside SGI for 2 months.  It
handles kernel stacks with multiple struct pt_regs, as found
while debugging the kernel.

I had to apply the patch below to build with CONFIG_NUMA.

	fsys_gettimeofday() patch.

This version executes in around 300 cycles on Itanium I (down from 900
or so for the original version), and so can be said to have
microsecond precision.

	invalid values.

Here is the patch that matches the changes I made for RHAS w.r.t
perfmon and the O(1) scheduler. This fixes the deadlocks we were
seing in RHAS and therefore on 2.5 as well.

The key change is that SIGPROF notifications are now delivered
in the perfmon kernel exit handler (called from entry.S).
At that point, we know none of the runqueue locks are held.

I have run the same test suite I used on RHAS overnight and no
sign of deadlocks.

into tiger.hpl.hp.com:/data1/bk/lia64/to-linus-2.5

	incomplete (as a result of a faulting mandatory RSE load).

Part of the bounds checking bugs found by the Stanford checker.

The 4 fixes below have been acked by their maintainers.

Move the definition of version[] down to where __initdata has been defined.

Patch from Joel Becker <Joel.Becker@oracle.com>

The task_struct->sig -> task_struct->signal bits.

Mikulas Patocka <mikulas@artax.karlin.mff.cuni.cz> points out a bug in
ll_rw_block() usage.

Typical usage is:

	mark_buffer_dirty(bh);
	ll_rw_block(WRITE, 1, &bh);
	wait_on_buffer(bh);

the problem is that if the buffer was locked on entry to this code sequence
(due to in-progress I/O), ll_rw_block() will not wait, and start new I/O.  So
this code will wait on the _old_ I/O, and will then continue execution,
leaving the buffer dirty.

It turns out that all callers were only writing one buffer, and they were all
waiting on that writeout.  So I added a new sync_dirty_buffer() function:

	void sync_dirty_buffer(struct buffer_head *bh)
	{
		lock_buffer(bh);
		if (test_clear_buffer_dirty(bh)) {
			get_bh(bh);
			bh->b_end_io = end_buffer_io_sync;
			submit_bh(WRITE, bh);
		} else {
			unlock_buffer(bh);
		}
	}

which allowed a fair amount of code to be removed, while adding the desired
data-integrity guarantees.

UFS has its own wrappers around ll_rw_block() which got in the way, so this
operation was open-coded in that case.

into tetrachloride.(none):/mnt/stuff/kernel/2.5/agpgart-respin

This removes lots of annoying problems trying to prevent both modules
from being loaded, and also shares quite a bit of code.
CONFIG_AGP3 will disable AGP3 mode operation of KT400s.