From Herbert Xu

The following patches against 2.4 and 2.5 makes cciss unregister properly
if initialisation fails.

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

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

  - created snd-page-alloc module
    - moved all page allocation code there
    - preserves preallocated DMA buffers for devices
  - USB audio driver updated
  - AC'97 - better modem initialization code
  - timer API - enhanced (added pause and more event notifications)
  - splitted ice1724 code from ice1712 to own module
  - general
    - timerstamp cleanups (timeval -> timespec)
    - C99-like cleanups
  - trident driver
    - more workaround for wrong IRQ acks
  - OSS sequencer emulation
    - fixed OOPS (wrong free order)
    - more compatible with level 1 of sequencer (/dev/sequencer)
  - CS46xx driver updated
  - intel8x0 driver updated
  - emu10k1 driver updated

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

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

From: Angus Sawyer <angus.sawyer@dsl.pipex.com>

Prevent OOPS on removing floppy driver with "rmmod -f floppy".

floppy.c would attempt to unregister resources for nonexistent device.

Patch stops the driver attempting to register and unregister the nonexistent
device by removing the drive from the allowed drives mask (defaults to
present).

From: Manfred Spraul <manfred@colorfullife.com>

Now that slab supports monster objects (up to 32MB) for !CONFIG_MMU we really
don't want to keep spare instances of them in the slab head arrays.

- limit head array sizes for huge slab caches to one object per cpu.

- round the batch count up for default head array sizing - batch count 0 is
  illegal.

- There was an unchecked bdget().  bdget can fail due to ENOMEM.

- rework the error handling implementation in raw_ctl_ioctl().

- Replace MOD_INC_USE_COUNT with try_module_get(THIS_MODULE).  This allows
  the raw module to be unloaded again.

  The core kernel has already taken a ref on the module prior to entering
  the ioctl, so try_module_get() cannot fail.

The recent fix to the raw driver wasn't quite right: it rewrites the
character-special inode's i_mapping to point back to itself on each close.
So any other currently-open handles against /dev/raw/rawN get a nasty
surprise.

Change it to only rewrite i_mapping on the final close.

Also, change it so that it only redirects its i_mapping on the initial open.
This is not necessary, but is neater.

From: Zwane Mwaikambo <zwane@holomorphy.com>

This patch makes us bail out in case we may have an interrupt which couldn't
be associated with an interrupt controller.  Without this we allow
unconfigured interrupts to be assigned and then later on we get "unexpected
IRQ trap at vector xx" during the ack phase.

scenario: This can occur if we fail irq setup during setup_IO_APIC_irqs for
some reason or other and then miss getting assigned a vector.  Later on we
then get assigned no_irq_type as our handler.

Patch for i386 and x86_64

From: Steven Rostedt <rostedt@goodmis.org>

The n_tty driver is missing some put_user checks.

- jiffies signedness fix

- Fix compile warning

From: Stephen Hemminger <shemminger@osdl.org>

The CPU frequency change detection code can change the values used to compute
time of day with TSC; but there was no locking around it.

From: Tim Schmielau <tim@physik3.uni-rostock.de>

Fixes the problem wherein nanosleep() is sleeping for the wrong duration.

When starting out with timer_jiffies=0, the timer cascade is (unneccessarily)
triggered on the first timer interrupt, incrementing all the higher indices.
When starting with any other initial jiffies value, we miss that and end up
with all higher indices being off by one.

From: William Lee Irwin III <wli@holomorphy.com>

The NMI watchdog has two different "modes": NMI_IO_APIC, which delivers NMI's
through the IO-APIC, and NMI_LOCAL_APIC, which uses the local APIC vector
table (LVT) to deliver the periodic NMI's.

Only NMI_IO_APIC requires being able to set up the PIT so it can deliver
NMI's through the IO-APIC, and so NMI_LOCAL_APIC has no dependency on the
timer being set up through the IO-APIC and is unjustifiably disabled by
check_timer() when the PIT cannot deliver interrupts through the IO-APIC.

This is important because one of the most important uses of NMI_LOCAL_APIC is
to get the NMI watchdog going when NMI_IO_APIC doesn't work.

So what this patch does to repair the situation is instead of checking to see
if the NMI watchdog is enabled at all, it instead checks whether the NMI
watchdog is being used in NMI_IO_APIC mode when a failure to set up the NMI
timer interrupt through the IO-APIC occurs.

The current nanosleep implementation has a signedness problem which can cause
it to sleep for 0x7ffffffe jiffies if a clock interrupt happens at the wrong
time.

The patch fixes that up, and also fixes an wrapping-unsafe 64-bit time
comparison.  Also uninline tstojiffie(), which has three call sites.

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

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

happens exactly on the sysenter entry path before the kernel stacks
have been switched to the proper ones.

Also, bump version to 1.5.

Leak fix contributed by Don Fry @ IBM

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

* Previous patch wiped NAPI support, adding it back here.  But,
  with a twist: this one doesn't disable/enable interrupts each
  time we enter/leave polling.  (It's EXPERIMENTAL).

* Corrected indentation from previous patches

* Fixed controller wake-up through ASF

* Added Interrupt Throttle Rate tuning support

* Added method to flush Tx FIFO after link disconnect; the hardware
  hangs on to Tx skb's that were in flight prior to link loss

* Miscellaneous whitespace changes

* Changed "Compaq" branding to "HP"

* Optimized E1000_*_REG macros

* Completely clean Tx descriptor to avoid potential dirty descriptor
  fetching (rare, but possible)

* ISR cleanup; performing single PCI read

* Removed modulus math; decreases CPU utilization, especially on PPC64
  [anton@samba.org]

* Added MII support

* Added support for 82541 and 82547 gigabit ethernet adapters

* Fixed IRQ registration bug; IRQ now registered after resources are
  acquired

* Setting speed/duplex is now it's own routine
* Update ETHTOOL_GEEPROM routine to use new eeprom size variable

* Version, copyright, changelog and MAINTAINERS updates