Gcc should be smart enough to do this itself.

(Pavel Machek)

Noted by Viktor Radnai.

From Michael Still <mikal@stillhq.com>

(From Stelian Pop)
Upon investigation, the reason is that trying to set the performance
to 80% in longrun_determine_freqs leaves the performance to 100%.
The performance level, at least on this particular model, can be lowered
only in 33% steps. And in order to put the performance to 66%, the
code should try to set the barrier to 70%.
                                                                                                           
The following patch does even more, it tries every value from 80%
to 10% in 10% steps, until it succeeds in lowering the performance.
I'm not sure this is the best way to do it but in any case,
it works for me (and should continue to work for everybody else).

Introduce a per-cpu powernow_k8_data struct, and pass it around between all
the functions that need to know about it.

Move lots of code to the cpu init routine from the detection routine.
Do some SMP sanity checks in the detection routine.

Trying to get the diffsize down between me & Pavel.

into delerium.codemonkey.org.uk:/mnt/nfs/sepia/bar/src/kernel/2.6/trees/cpufreq

This cleans up a larger amount of superfluos ";;" statements in current
Linux kernel sources by converting them to the regular single ";"
statments.

It seems to be a common problem that at the end of a line the semicolon
key is producing an echo.

into ppc970.osdl.org:/home/torvalds/v2.6/linux

From: Anton Blanchard <anton@samba.org>
      Rusty Russell <rusty@rustcorp.com.au>

The iSeries has an arch-specific mapping from physical <-> absolute
addresses.  Fortunately this is only used in a few places.  However, the
following arch-specific macros/functions are provided in addition to the
standard macros:

	__a2p()
	__a2v()
	__p2a()
	__p2v()
	__v2a()
	__v2p()
	absolute_to_phys()
	phys_to_absolute()
	virt_to_absolute()
	absolute_to_virt()

Reduce them to these, with slightly shorter names, and taking either pointers
or unsigned long (as per __va and __pa) rather than making the caller cast:

	abs_to_phys()
	phys_to_abs()

And helper macros:

	virt_to_abs()
	abs_to_virt()

As is standard, virtual addresses are returned as void *, physical and
absolute as unsigned long.

Note that the change the iSeries_setup is a little subtle: ea is set to
__va(pa) above, so "phys_to_abs(pa)" is the same as "virt_to_abs(ea)".

Also, REALADDR is renamed to ISERIES_HV_ADDR and used in a couple of places
where appropriate.

From: Anton Blanchard <anton@samba.org>
      Stephen Rothwell <sfr@canb.auug.org.au>

This patch just ensures that the RI bit is set really early as it seems it
is not on iSeries Power4 machines (or maybe OS400 v5r3 does this).

au88x0 driver
Cleanups - removed duplicate PCI IDs

USB generic driver
add usb_device->speed wrapper for compiling with 2.2.x kernels

into suse.cz:/home/perex/bk/linux-sound/linux-sound

People were getting quite excited about this.

This backs out Maneesh's sysfs patch that was recently added to the
kernel.

In its defense, the original patch did solve some fixes that could be
duplicated on SMP machines, but the side affect of the patch caused lots
of problems.  Basically it caused kobjects to get their references
incremented when files that are not present in the kobject are asked for
(udev can easily trigger this when it looks for files call "dev" in
directories that do not have that file).  This can cause easy oopses
when the VFS later ages out those old dentries and the kobject has its
reference finally released (usually after the module that the kobject
lived in was removed.)

I will continue to work with Maneesh to try to solve the original bug,
but for now, this patch needs to be applied.

From: Ivan Kokshaysky <ink@jurassic.park.msu.ru>

The memory reserved for PCI probe is not freed properly in some cases,
for instance with a 3.5Gb of RAM.

Forward port from 2.4.

Oops, there was two different code path affected by this
bug (strace and normal) and I fixed only one. Here's the
other one:

For the same reason as ppc32, we need to ensure that all stores
done on a CPU has reached the coherency domain and are visible
to loads done by another CPU when context switching as the same
thread may be rescheduled almost right away there.

This updates the ppc32 PCI mmap facility to allow mmap'ing of space
outside of the actual devices, using the host bridge resources instead. 

This allow userland to map things like legacy IO space by either using
the bridge device itself, or simply any PCI device on the same bus
domain

On ppc32, CONFIG_PREEMPT wasn't settable along with CONFIG_SMP
for historical reasons (smp_processor_id() races). Those races have
been fixes since then (well, should have been at least) so it's now
safe to allow both options.

This fixes a few issues with context switch on ppc32:

 - Makes sure we properly flush out all stores to the coherency domain
   when switching out, since the same thread could be switched back in
   on another CPU right away, those stores must be visible to all other
   CPUs. 

 - Remove dssall in the assembly calls and do it now once in switch_mm
   (stop vmx streams).  Assume the G5 doesn't need a sync after dssall. 

 - Remove bogus isync in the loop setting the userland segment registers

 - Do not switch the userland segments when the mm stays the same

enable_kernel_fp is exported both in ppc_ksyms and near it's
definition in process.c, remove the former.

Add a warning if enable_kernel_{fp,altivec} is called with preempt
enabled since this is always an error, and make sure the alignement
exception handler properly disables preempt when doing FP operations.

We don't set the task state to TASK_INTERRUPTIBLE _before_ checking for
kthread_should_stop in do_stop.

into ppc970.osdl.org:/home/torvalds/v2.6/linux

From Dave Jones.

Oops.

   Check for valid application pointer inside api spinlock
   in diva_send_message().

Since pdflush was converted to be launched by the kthread infrastructure it
has inherited keventd's `nice -10' setting.  That hurts interactivity when
pdflush is doing lots of work writing back through the dm-crypt layer.

So set pdflush back to `nice 0'.

From: Mike Christie <michaelc@cs.wisc.edu>

A couple of drivers can sometimes fail the first segments in a bio then
requeue the rest of the request.  In this situation, if the last part of
the bio completes successfully bio_pair_end_* will miss that the beginging
of the bio had failed becuase they just return one when bi_size is not yet
zero.  The attached patch moves the error value test before the bi_size to
catch the above case.