Here is a patch that adds numa support for hugetlb page on ia64.  It is
taken from x86 numa implementation.  The low level hugetlb page
pre-allocation is done in round robin fashion on each numa node and
allocation for user level code is node local aware.

n platforms without an SRAT (e.g. zx1), the cpu_to_node_map will get
built incorrectly without this fix, making generic kernels fail when
they try to alloc_pages_node() from a nodeid of -1.

Small fix to build early_sn_setup even if early printk is turned off.

I didn't realize that any_online_cpu() wouldn't work for
is_headless_node(), so this patch reverts the function back to using the
node_to_cpu_mask[] array, fixing sn2 for systems with less than 64p.

Don't try to use the sn2 I/O code if we're not on sn2 or we may get into
trouble.  Only makes a difference for generic kernels.

I ran into some bugs testing epoll in ia32-emulation mode on ia64.
The attached patch fixes the problems and is well tested.
Here is a summary of the changes:

Changes to sys32_epoll_ctl() and sys32_epoll_wait():
- changed epoll_event32.data to an array (this is subjective, but seems
to make the code more readable)

Changes to sys32_epoll_wait():
- added call to __get_free_pages if kmalloc(epoll_event array) fails.  This
  provides needed scalability and fixes the -ENOMEM failure during
epoll-pipetest.
- deleted copy-in of epoll_event array since this is not a user-inputsyscall
- changed to check numevents > 0 as indicator of success on call to
  sys_epoll_wait.
- changed to loop on numevents not maxevents when copying out to userspace.

It's not like it would compile without PCI support anyway, and these
ifdefs are _really_ strange.

Same issues as the dma mapping code, but as SN2 interrupt handling is
totally static and all irqs are allocated on bootup an never freed it's
not a speed issue.

I really wonder what someone has to smoke to call a that expensive
dispatcher just to cast the data structure back to the lowlevel one
directly afterwards...

fed through Lindent and dead codepathes eliminated.

Currently the the SN2 dma mapping code uses a file systen lookup (cat /r/napali/tmp/p10 |diffstat -p1) to
find the operation vector for the underlying dma mapping implementation.

But not only there is exactly one of those implementation, also data
structures from this lowlevel driver are used in the actual Linux dma
mapping routines and quite a few assumption are made about it.  So let's
just call directly in the lowlevel code and give up this utter bullshit.

These strange mixups of Linux and IRIX PCI interfaces were used in SGI's
2.4 tree for their port of the IRIX scsi code.  They have no business to
stgay in 2.6.

More IRIX creepism

Just some crappy stubs that sneaked over from IRIX.

Keith shared my opionion that it's completly useless.  Also everyone
seems to favour keeping kdb stuff out of mainline.

	up the leftovers.

	by a module, even more the file doesn't compile when
	CONFIG_IA64_SGI_SN_DEBUG is set.

	the current tree.

	from kernel/setup.c into two separate files: mm/{dis,contig}.c
	to handle contiguous vs. discontiguous memory layouts.

This patch against 2.5 gets rid of the following annoying but
harmless messages when booting a generic kernel on a non-ZX1 box:

        No IOC for PCI Bus 0000:00 in ACPI
        No IOC for PCI Bus 0000:01 in ACPI
        ...

We just forgot to bail out of the SBA IOMMU init function if it didn't
find any SBA hardware.

I'm aligning the 2.4 and 2.5 versions of arch/ia64/hp/common/sba_iommu.c
and found a couple nits in the 2.5 version.  No functional change, just
whitespace, comment, and parameter name changes (and I made
one function static).

asm/acpi.h relies on struct pci_vector_struct which is defined in
asm/system.h.

Minor cleanup.

This patch removes a ton of pointless big endian defines for some
registers on SN2 and cleans up the #include hierachy making it include
some really big header files only when they are really needed.

This patch changes the kernel such that si_isr gets setup for hardware
breakpoints, single-step, and taken-branch traps.  This is useful, e.g., to
determine what kind of hw breakpoint triggered the signal.

Attached is a small cleanup patch for the sn2 header files which removes
some cases of excessive header file inclusion.

      SBA_PROC_FS off by default (it's too much of a scalability bottleneck).

	the #ifdef __KERNEL__ bracket.  Doing so pollutes the user-
	level namespace.  Bug report & proposed fix by GOTO Masanori.

This patch fixes a typo in pfm_write_pmcs() in the test for the default
value. The code was using the lop index instead of the register
index in the PMC_DFL_VAL() macro. This was causing valid values
for some PMCs to be rejected.

	by Jesse Barnes.

Having this print out for every CPU on a large system was a pain, so
protect the printk with EFI_DEBUG.

This code works pretty well now, so we don't need to dump all this stuff
at boot time (esp. on 128 node systems).

This patch coverts a few spots to use cpumask_t instead of unsigned
long.

If you try to compile 2.6.0-test[45] with spinlock debugging on, then
unwind.c won't compile, because it uses a #define magic that hides
the variable in the spinlock debugging code (also called magic).

MINSIGSTKSZ is defined differently for i386 and ia64. This patch improves
compatibility with apps which use sigaltstack(2) with sizes between
MINSIGSTKSZ_IA32 and MINSIGSTKSZ.

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