Patch from Dave Hansen.

Fix simple bounding error found by some Stanford-checker type thing
to use the proper MAX_MP_BUSSES define instead of a constant.

This simple patch just makes sure the PIT code is available for NUMA-Q
(as its TSCs are not synced).

Has been tested in my tree for over a month on UP, SMP, and NUMA and
compile tested against a variety of different configs.

At the moment, there are two different switches used, irqbalance_disabled
and no_balance_irq ... each of which switches half the code off. This
patch harmonises them into one (irqbalance_disable), and uses the old
subarch stuff as a default value so that this is auto-disabled on boxes
like NUMA-Q that can't cope with it.

Also renamed no_balance_irq to NO_BALANCE_IRQ as it's really a static
defined number now, not pretending to be a switch variable any more.
Now off by default for NUMA-Q, on by default for others, but can be
disabled with the boot time flag if people desire.

Patch from Matthew Dobson

Provide a pcibus_to_cpumask function in the topology infrastructure to
access the pre-existing array.

Patch from William Lee Irwin

Inline and simplify pfn_to_nid - this is called heavily, it's a tiny
function, and makes a noticable difference in system time for kernel
compiles (sorry, lost the data). Is only used on NUMA machines.

Has been tested in my tree for over a month on UP, SMP, and NUMA and
compile tested against a variety of different configs.

translation_table could be unintialised on machines other than NUMA-Q.
You never see this, as the compiler happens to optimise it away, but
it's still a Really Bad Idea (tm). Thanks to someone whose name I have
unfortunately forgotten for pointing this out.

Has been tested in my tree for over two weeks on UP, SMP, and NUMA and
compile tested against a variety of different configs.

Patch from William Lee Irwin

This moves the node's pgdat into the node's own local memory area along
with the lmem_map.

It only touches arch/i386/mm/discontig.c - ie can only affect NUMA
machines.  Has been tested in my tree for over a month on UP, SMP, and
NUMA and compile tested against a variety of different configs.

This converts /proc/io{mem,ports} to the seq_file interface
(single_open).

This patch fixes the incomplete data return problem of /proc/ide/triflex.
When the number of consecutive read bytes are smaller than the total
data in triflex_get_info(), the second read() returns 0.

Fixes for the 31 bit compatability layer:
 * add compat function for jiffies_to_timeval
 * convert parameters of clone system call
 * incorporate changes from setup_arg_pages to setup_arg_pages32
 * incorporate changes from do_execve to do_execve32
 * take care of nanosecond field in struct timespec
 * remove functions of old module system calls
 * add TIF_31BIT thread flag and replace S390_FLAG_31BIT with it
 * add arch_get_unmapped_area
 * add wrapper for sys_set_tid_address

Add missing/fix existing s390 relocations in the kernel module loader.

trivial s390 fixes/typos.

updates for s390 scsi support

- remove bogus header file
- add a definition for free_dma() to make scsi.c work

updates for unified netdev config

- remove duplicate questions from drivers/s390/Kconfig
- some trivial fixes to make the s390 specific options work
- new default configurations

clean up the IUCV driver

updates for compiling with gcc-3.3pre

- add -finline-limit=10000 to make it build
- drop .eh_frame elf section from vmlinux
- fix common warnings inn asm headers
- make dasd compile
- Don't warn about signed/unsigned comparisions
- fix inline syscall macros

update extable support in s390 and s390x

this makes use of the unified extable code.
for 31 bit s390, this is slightly more complicated
than the other architectures, but as long as
no one outside /arch uses search_exception_tables,
everything should work nicely

minor changes to s390 documentation

updates for s390 dasd driver

Some problems have been found in the dasd error handling code, they
are fixed by the update to dasd_3990_erp.c.

Dasd is one of only two remaining drivers that use the null
elevator instead of iosched. Appearantly, the null elevator
has some bitrot and can result in random data loss. For now,
we just don't use it.

Bug fixes for the CTC driver

also removes the bogus intparm_t typedef

updates for the channel subsystem and qdio driver

This adds the missing support for chp machine checks, i.e.
enabling or disabling a set of devices from the service element.
Some minor bugs in the driver are fixed as well.

s390 arch changes/bug fixes:
 * add sys_fadvise64 system call
 * add initialization of init_sighand
 * add support for clone option CLONE_SETTLS
 * make use of ptrace_notify
 * sig -> sighand rename
 * move ptrace_signal_deliver to the correct place
 * make eieio a memory barrier
 * fix race condition on cpu_vm_mask in __flush_tlb_mm
 * add missing KM_SOFTIRQ0, KM_SOFTIRQ1 in asm-s390x/kmap_types.h
 * add inline assmelby for _raw_write_trylock

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

Make kmod force default handlers before executing the user process.

into kernel.bkbits.net:/home/davem/net-2.5

the lane spec says the following about path switching:

10.1.2.7 Transmitting Held Frames

Once the LE_FLUSH_RESPONSE is received, the LE Client MUST
transmit any held  data frames on the new data path before
transmitting any further frames on the new path.

10.1.2.8 Switching Over Paths Without Flush

Regardless of the provisions of the rest of Section 10.1, when
switching from the old path to a new path, if an LE Client has
not transmitted a data frame to a particular LAN Destination via
the old path for a period of time greater than or equal to the
C22 Path Switching Delay, then it MAY start using the new path
without employing the Flush protocol.

the lane client in linux-atm does do this but it doesnt send the
held frames until a frame is transmitted on the vcc after the flush
has completed.  often this doesnt happen until you get an ip retransmit.
the following patch sends the held frames after the flush completes
(or the switching delay expires)

Stelian Pop reported that all signals are blocked in processes
execed from the kernel as usermode helpers.

and gid in local inode (not just remote)

 - suspend & remove for pci_socket was broken -- thanks to Paul
   Mackerras for noting this
 - to correctly initialize multiple pci_socket devices, a sock_offset is
   needed.
 - s_info doesn't need to be an array.

 - switch the Intel x86 speedstep/Pentium 4/Xeon drivers and the
   Transmeta Crusoe Longrun driver to use the advanced cpufreq_driver
   registration process
 - cleanups
 - spelling fixes (Petri Koistinen) - thanks!

Move x86 CPU_FREQ config choices to extra file & menu. (Marc-Christian Petersen)