* Bug fix: 82544 hang with PCI-X: if outgoing Tx buffers terminate
  within evenly-aligned dwords, and the device is sharing the bus
  segment with another PCI-X device, 82544 can hang the bus on a
  split-completion transaction.  Fix is to split buffer into two
  buffers with the first one not terminating within evenly-aligned
  dword address, and the second one being 4-bytes, which goes as a
  non-split-conpletion PCI-X transaction.
* 8254x controllers that support TSO do an internal calculation to
  make sure there is enough FIFO space to handle the overhead
  of each TSO segment before DMA'ing TSO data from host memory.  The
  internal calculation is dependent on the mss of the TSO (defines
  the number of segments), but the reserved space is a constant, so
  we need to adjust the maximum size of each buffer queued to the
  hardware to hold the equation and not overrun the FIFO.  This is
  per TSO because the mss can change from one send to the next.

* Added 82545 (rev3), 82546 (rev3), and 82541/7 (rev2) support
	- new device IDs
	- internal SERDES support for 82545/6 (rev3)
	- don't apply MMRBC workaround for 82545/6 (rev3)
	- don't use IO mapping for reset for 82545/6 (rev3)

Thish fixes Suparna's fsx and aio-stress and Dave Olien's concurrent
mke2fs problems.

It moves the hash adding into as_add_request.  This fixes a small bug
where a request could be added to the merge hash after being put on the
dispatch list. 

It also moves responsibility for handling rbtree aliases (requests with
the same start sector) out of as_add_arq_rb and into its caller.  This
cleans up a layering problem which was causing the state machine to go
silly.  Now instead of forcing the alias out onto the dispatch list, we
chain the new request behind it, and they get sorted out at dispatch
time.

The PS/2 ESDI driver has some problems: a couple of typos on the modular
side and misuse of module_init():

A colleague reported that the keyboard of his brandnew laptop
is dead under 2.6 while 2.4 works.

Now I once wrote

  In order to avoid interference between scancode sequences
  or mouse packets and the reponses given to commands,
  the keyboard or mouse should always be disabled before
  giving a command that requires a response, and probably
  enabled afterwards. Some keyboards or mice do the disable
  automatically in this situation, but still require an
  explicit enable afterwards. 

(http://www.win.tue.nl/~aeb/linux/kbd/scancodes-9.html)

This is what happens on this laptop. The routine atkbd_probe()
probes for a keyboard, and after detecting it, enables it.
But immediately afterwards the routine atkbd_set_3() reads
the current scancode set and sets the desired set, and as a
side effect of these commands, the keyboard gets disabled again.

Thus, the keyboard enable must be moved after all command sending
has been done.

Now that I patch this area anyway: we are almost always in
scancode set 2 but send the ATKBD_CMD_SETALL_MB command
that only works in scancode set 3. At best this is useless.
At worst it confuses the keyboard. So, I put this command
in a separate routine and call that only when we really
are in scancode set 3.

into home.osdl.org:/home/torvalds/v2.5/linux

Combine the two initialisation functions, allow SA1111 to be built
as a module, and remove a redundant SA1111 function prototype.

Patch from Holger Freyther

Make it work ;)

Patch from Holger Freyther

This fixes initialization of PCMCIA for the SIMpad



unsigned long now user for the flag

Patch from Holger Freyther

This the the resubmit of the flash partition changes for SIMpad.



Two versions one for cramfs one for all jffs2 and also the SIM has two banks instead of one even if they're next to each other...

into home.osdl.org:/home/torvalds/v2.5/linux

into flint.arm.linux.org.uk:/usr/src/bk/linux-2.6-pcmcia

devfs_handle is not used any more.

My fault.  Fix for broken aarp.c which got an extra closing parenthesis.

<linux/interrupt.h> uses barrier() but does not include <linux/kernel.h>.

Whitespace cleanup (mostly deleting trailing whitespace).

This patch introduces a 1GB stack size limit for stack-grows-up (ie
PA-RISC), as discussed previously.

Dont print the contents of the initramfs, for any decent sized cpio it will
overflow the kernel ring buffer. 

Also relax permissions on /dev (755 not 700).

Set correct permissions on initramfs directories and special files. We dont
want to obey the umask here, so do the same thing we do on normal files -
call sys_chmod.

tty_register_driver already calls tty_register_device so there is no need
to do it in hvc_console. Besides, it oopses when we do that.

Thanks to Stephen Hemminger for pointing out how obsolete modules.txt is.

modules.txt contains mainly ancient information which is replicated
in the kconfig help message, README, makefile.txt or the modprobe manual
page.  The only part which is not covered elsewhere is the "building
external modules" which is still being debated (and belongs under the
kbuild docs).  kmod.txt reference removed from index, too.

pci_device_id can not be marked __devinitdata,
was re-added with last update by accident.

This fills in the ELF EI_OSABI field.  This doesn't matter for most
architectures, but PA-RISC uses the Linux flavour of the ABI (since HPUX
uses the None flavour).

Patch by Randolph Chung.

This is a lot better than what is there know.

From: Felipe W Damasio <felipewd@terra.com.br>

 - cli-sti removal
 - blk API update
 - set_current_state
 - Remove 'panic' line.

.. and we can now remove the BROKEN_ON_SMP Kconfig annotation.

From: Stephen Hemminger <shemminger@osdl.org>

Compiler warning due to missing equal sign.

From: Andrew Theurer <habanero@us.ibm.com>

This change:
http://linux.bkbits.net:8080/linux-2.5/diffs/kernel/sched.c@1.202

does not seem to make sense:

#define CAN_MIGRATE_TASK(p,rq,this_cpu)					\
	((!idle || (jiffies - (p)->last_run > cache_decay_ticks)) &&    \
		!task_running(rq, p) &&					\
			cpu_isset(this_cpu, (p)->cpus_allowed))

It should be just the opposite; an idle cpu should be able to have a more
aggressive steal, and a busy cpu should not.

From: Matt Mackall <mpm@selenic.com>

The new locking in the random driver is consuming 60% of CPU resources in
Anton's monster power5 boxes.


Basically, when the primary pool is 7/8th full, we shut off the firehose
and go into a trickle mode to keep the pool fresh.  Saves CPU for everyone
and should make the contention drop off the charts too (though the trickle
factor might need adjusting again for Origin-class machines).

From: Dave Hansen <haveblue@us.ibm.com>

I was compiling for my plain 'ol PC, and was getting unresolved symbols
for get_memcfg_from_srat() and get_zholes_size().  The CONFIG_NUMA
definition right now allows it to be turned on for plain old X86_PC. 
Does anyone know why this is? 

depends on SMP && HIGHMEM64G &&
          (X86_PC || X86_NUMAQ || X86_GENERICARCH ||
                  (X86_SUMMIT && ACPI && !ACPI_HT_ONLY))

In any case, the summit code incorrectly assumes in at least 2 places
that NUMA && !NUMAQ means summit.  Someone was evidently trying to cover
the generic subarch case, but that's already taken care of by the lovely
config system and CONFIG_ACPI_SRAT.  This patch fixes those assumptions
and adds a nice little warning for people that try to #include srat.h
without having srat support turned on.

From: Manfred Spraul <manfred@colorfullife.com>

The mwave driver uses a user space daemon for some modem operations. The
user space daemon calls ioctl(,IOCTL_MW_GET_IPC), and the driver returns
after an interrupt arrived. The actual wait used
interruptible_sleep_on(), which can lead to lost wakeups. A local 
spinlock on the stack is used to close that race, but this is broken on 
SMP, perhaps even with preempt.

The attached patch fixes that by switching to the normal 
add_wait_queue/test_if_race_occured/schedule/remove_wait_queue sequence.