The loop thread is getting permanently stuck in balance_dirty_pages()
(nr_writeback is exceeded) because the loop thread itself is responsible for
completing writeback on behalf of higher layers.

So we need to take that out: don't throttle the loop thread.  Throttle the
tasks which are generating all the dirty data instead.

From: Matthew Dobson <colpatch@us.ibm.com>

This resurrects the old /proc/sys/vm/free_pages functionality: the ability to
tell page reclaim how much free memory to maintain.

This may be needed for specialised networking applications, and it provides
an interesting way to stress the kernel: set it very low so atomic
allocations can easily fail.

Also, a 16G ppc64 box currently cruises along at 1M free memory, which is
surely too little to supporthigh-speed networking.  We have not changed that
setting here, but it is now possible to do so.

The patch also reduces the amount of free memory which the VM will maintain
in ZONE_HIGHMEM, as it is almost always wasted memory.

From: Dave Olien <dmo@osdl.org>

The change makes the transfer rate numbers come out right for the fibre
channel version of this controller.

For 1G FC, the NegotiatedSynchronousMegaTransfers is 1000, and the
NegotiatedDataWidthBIts is 1.  The old code assumed NegotiatedDataWidthBits
was always either 8 or 16.  The new code is simpler, and does the
calculation correctly for all cases.

From: Matthew Dobson <colpatch@us.ibm.com>

This patch implements a generic version of the nr_cpus_node(node) macro
implemented for ppc64 by the previous patch.

The generic version simply computes an hweight of the bitmask returned by
node_to_cpumask(node) topology macro.

This patch also adds a generic_hweight64() function and an hweight_long()
function which are used as helpers for the generic nr_cpus_node() macro.

This patch also adds a for_each_node_with_cpus() macro, which is used in
sched_best_cpu() in kernel/sched.c to fix the original problem of
scheduling processes on CPU-less nodes.  This macro should also be used in
the future to avoid similar problems.

Test compiled and booted by Andrew Theurer (habanero@us.ibm.com) on both
x440 and ppc64.

From: Matthew Dobson <colpatch@us.ibm.com>

This patch defines a topology macro for ppc64, nr_cpus_node(node) which
returns the number of CPUs on 'node'.  This patch also adds code to compute
and store these values in an array for quick lookup.

Test compiled and booted by Andrew Theurer (habanero@us.ibm.com) on both
x440 and ppc64.

From: Matthew Dobson <colpatch@us.ibm.com>

sched_best_cpu schedules processes on nodes based on node_nr_running.  For
CPU-less nodes, this is always 0, and thus sched_best_cpu tends to migrate
tasks to these nodes, which eventually get remigrated elsewhere.

This patch adds include/linux/topology.h, and modifies all includes of
asm/topology.h to linux/topology.h.  A subsequent patch in this series adds
helper functions to linux/topology.h to ensure processes are only migrated
to nodes with CPUs.

Test compiled and booted by Andrew Theurer (habanero@us.ibm.com) on both
x440 and ppc64.

From: Ingo Molnar <mingo@elte.hu>

Apparently this BUG is triggerable due to correct and expected events.

As "Dmitry A.  Fedorov" <D.A.Fedorov@inp.nsk.su> points out,
get_current_user() has a local variable __user which conflicts with the
sparse tagging.  But get_current_user() has no callers.

From: Adam Belay <ambx1@neo.rr.com>

- cs4236 doesn't check if the memory for the resource table was
  successfully allocated.

- Add comment about slab ctor behaviour (Ingo Oeser)

- mm/slab.c:fprob() shows up in profiles a lot.  Rename it to something more
  meaningful.

- fatfs printk warning fix (Randy Dunlap)

- give the the time interpolator list and lock file-static scope (hch)

From: Christoph Hellwig <hch@lst.de>

Currently the tty code abuses tty_driver.name as the prefix for the devfs
names of the ttys.  This is a very bad idea because it means the tty name
changes depending on whether devfs is enabled or not, leading to different
names in /proc/tty/ depending on whether we have devfs or not (and not
whether it actually is mounted!) and a huge amount of ifdefs.

The patch below adds a .devfs_name member instead, similar to the block
device changes a few weeks ago.

From: Samuel Thibault <Samuel.Thibault@ens-lyon.fr>

Still working on kernel facilities for braille devices, the need for being
able to force blanking and unblanking raised: even when a key is pressed, the
screen must remain blank, for privacy of the blind user who is typing on the
keyboard and reading on its braille terminal.

I merely added an ignore_poke variable which is set, and the screen blanked.
Then, poke_blanked_console returns immediatly.  Upon real unblank (because of
an Oops or an explicit tioclinux), ignore_poke is reset to get back to normal
operation mode.

I had to remove the (unnecessary ?) call to unblank_screen from set_selection
to prevent mouse selection unblanking the screen.

I also added a way for processes to know whether the screen is blanked (the
blind user might hence know whether people can read the screen).

From: Samuel Thibault <Samuel.Thibault@ens-lyon.fr>

Some fixes for console blanking: on some laptops, doing VESA blanking after
the bios did an apm blanking because of a screen closure thrashes the
recovery (the video board doesn't seem to have synchronisation registers
correctly initialized, since the LCD panel progressively turns white, maybe
damaging it ?).

I hence moved the schedule for vesa powerdown after the apm blank hook
call, so that if it succeeds, it won't be called.  I also moved the apm
unblank & palette restoration after the vesa unblank, to have a more lifo
scheme (also required, or the screen remains black).

Btw, why del_timer_sync was called twice in timer_do_blank_screen when
vesa_off_interval==0 ?

From: john stultz <johnstul@us.ibm.com>

Since jiffies didn't necessarily start incrementing at a second boundary,
jiffies/HZ doesn't increment at the same moment as xtime.tv_sec.  This
causes one second wobbles in the calculation of btime (xtime.tv_sec -
jiffies/HZ).

This fix increases the precision of the calculation so the usec component
of xtime is used as well.  Additionally it fixes some of the non-atomic
reading of time values.

From: David Mosberger <davidm@napali.hpl.hp.com>, Christoph Hellwig

I believe this is the last outstanding piece that prevents ia64 from being
fully in sync with Linus' tree (yes, there are some minor ACPI changes
outstanding and a toolchain bug that's left to fix, but other than that, I
think we're clean).

Many architectures (alpha, ia64, ppc, ppc64, sparc, and sparc64 at least)
use a syscall convention which provides for a return value and a separate
error flag.  On those architectures, it can be beneficial if the kernel
provides a mechanism to signal that a syscall call has completed
successfully, even when the returned value is potentially a (small)
negative number.  The patch below provides a hook for such a mechanism via
a macro called force_successful_syscall_return().  On x86, this would be
simply a no-op (because on x86, user-level has to be hacked to handle such
cases).  On Alpha, it would be something along the lines of:

 #define force_successful_syscall_return()  ptregs->r0 = 0

where "ptregs" is a pointer to the user's ptregs structure of the current
task.  On ia64, we have been using this for a long time:

 static inline void force_successful_syscall_return (void) {
	ia64_task_regs(current)->r8 = 0;
 }

The other architectures (ppc, ppc64, sparc, and sparc64) currently have no
mechanism to force a syscall return to be successful.  But since the
syscall convention already provide for a separate error flag, the arch
maintainers could change this if they wanted to.

There are only 3 places in the platform-independent portion of the kernel
that need this macro:

 - memory_lseek() in drivers/char/mem.c
 - fs/fcntl.c for F_GETOWN
 - lseek for /proc/mem in fs/proc/array.c

Ideally, there are a couple of other places that could benefit from this
macro:

 - sys_getpriority()
 - sys_shmat()
 - sys_brk()
 - do_mmap2()
 - do_mremap()

but these are not so critical, because the can be worked around in
platform-specific code (e.g., see arch/ia64/kernel/sys_ia64.c).

Note that for the above 3 cases, handling them in user level is rather
suboptimal:

 - it would affect all lseek() syscalls, even though only /proc/mem and
   /dev/mem need the special treatment (at least until there are
   filesystems that can handle files >= 2^63 bytes)

 - all fcntl() calls would be affected, even though only F_GETOWN needs
   the special treatment

so I think handling these in the kernel for the platforms that can makes
tons of sense.

The only limitation of force_successful_syscall_return() is that it doesn't
help with system calls performed by the kernel.  But the kernel does that
so rarely and for such a limited set of syscalls that this is not a real
problem.

The addition of more fields to irq_desc_t may have broken compilation of
other architectures.  Go through and C99ify them (was needed anyway).

Attempt to do something intelligent with IRQ handlers which don't return
IRQ_HANDLED.

- If they return neither IRQ_HANDLED nor IRQ_NONE, complain.

- If they return IRQ_NONE more than 99900 times in 100000 interrupts, complain
  and disable the IRQ.

  I did have it at 750-in-1000, but someone had an otherwise-functioning
  system which triggered it.

  The 99.9% ratio is designed to address the problem wherein the babbling
  device shares an IRQ with a good device.  We don't want the good device's
  trickle of IRQ_HANDLED callouts to defeat the lockup detector.  (fat chance
  os this working right).

- Add a kernel boot parameter `noirqdebug' to turn the whole thing off.

From: Rusty Russell <rusty@rustcorp.com.au>

OK, this does the *minimum* required to support DEFINE_PER_CPU inside
modules.  If we decide to change kmalloc_percpu later, great, we can turf
this out.

Basically, overallocates the amount of per-cpu data at boot to at least
PERCPU_ENOUGH_ROOM if CONFIG_MODULES=y (arch-specific by default 32k: I have
only 7744 bytes of percpu data in my kernel here, so makes sense), and a
special allocator in module.c dishes it out.

From: Rusty Russell <rusty@rustcorp.com.au>

Several tweaks to the kmalloc_percpu()/kfree_percpu() interface, to
allow future implementations to be more flexible, and make easier to
use now we can see how it's actually being used.

1) No flags argument: GFP_ATOMIC doesn't make much sense,

2) Explicit alignment argument, so we don't have to give SMP_CACHE_BYTES
   alignment always,

3) Zeros memory, since most callers want that and it's not entirely
   trivial,

4) Convenient type-safe wrapper which takes a typename, and

5) Rename to alloc_percpu/__alloc_percpu, since usage no longer matches
   kmalloc.

This fixes a problem which can show up with the new select facility, e.g.
a symbol is forced to 'y', so we should never even try to change such
symbols.

This is an important fix to allow changing boolean symbols, whose
dependency is 'm'. All internal symbol states must be converted from
the tristate into boolean the state.

I missed this change while adding expression support for defaults,
please apply.

into redhat.com:/garz/repo/net-drivers-2.5

dev_ioctl already checks capable(CAP_NET_ADMIN) for SOICETHTOOL, so
privileged reference are not necessary.

Add 10GbE support for ethtool.

This one should be safe as we're protected by the xmit_lock in all instances

Attached is a driver for Atmel at76c50x WiFi cards. This code started
out as a GPL release from Atmel of pretty horrible quality and I've
extensively re-worked it with the aim of making it acceptable in the
kernel. Please could you take a look and either pass it into the patch
stream or let me know what's wrong with it?

The code has been tested on at least three different brand cards by
different people. Jean Tourrilhes took a look at an earlier version an
was positive. He's put incorporating this into 2.6 as a priority 1.
The patch works fine on 2.5.70.

The firmware issue has been addressed now. The only firmware in the
driver is a small stub which reads the MAC address from NVRAM on the
card. The source for that is included so there are no GPL issues. The
main firmware is loaded from userspace using Manuel Estrada Sainz's
sysfs firmware class. I know that the  patch for that has been
accepted but it hasn't turned up anywhere I can see yet. The 
driver compiles fine even without the firmware class. I've made a
package of the firmware images which is available from my website.

The remaining issues with the driver are migrating PCMCIA to the new
driver model and PCI support. I'm happy to produce followup patches as
the PCMCIA system gets evolved to the new driver model: the timing on
that is controlled by others. This set of chips includes a PCI version
and the driver should support that, but AFAIK there is no PCI hardware
available anywhere. If Atmel can provide me with some it will be
simple to add PCI support.

The driver uses the CRC32 library module and the firmware loader. I've
not put in dependencies on those, but when the lastest set of patches
go into Kconfig I'll set it up so that selecting the Atmel driver
selects CRC32 and FW_LOADER too.

Inspecting the sb1000 driver showed some interesting bugs:
	- net device pointer is used before the device is allocated; gcc
	  does catch this.
	- unregister is called even though device not registered successfully
	- net device is not freed on remove.

Compiles but don't have hardware to test.  Don't know how it ever worked though.

Two ioctl functions in sound/oss/awe_wave.c were directly dereferencing
a user-supplied pointer in a few places.

Jörn missed a few places of FAR conversion in inflate

Remove the stale support for K&R function declarations through the OF()
macro.

This is the last patch to clean up zconf.h, at least for now.

This nice macro must have been one of the good intentions on the road
to hell.  Completely unused. :)

Just a simple s/ZEXPORT//.

This one was just simple s/ZEXTERN/extern/g.

Remove unused __32BIT__ and STDC macros

This removes FAR, the typedefs using FAR (Bytef and friends) and the
function prototypes for zalloc and zfree that should have gone earlier
already.

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

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