set frag_off field of IPv4 header in hot paths.
Use __constant_htons as appropriate.

tcp_timewait_kill.  This is what made the following
bug harder to find:
Put new timewait buckets into the bind hash _FIRST_
before they appear into the established hash to
kill some races with socket creation/lookup.

out of struct sock.
Fix -EFAULT handling in TCP direct user copy handling.
Use struct initializers in IPV6 ndisc code.

into penguin.transmeta.com:/home/penguin/torvalds/repositories/kernel/linux

* Make matroxfb compilable with Mystique, but without G450 support.

			Petr Vandrovec

into penguin.transmeta.com:/home/penguin/torvalds/repositories/kernel/linux

Documentation/filesystems/cramfs.txt
  - remove comment about ROM size limit
  - fix up magic
  - update location of tools

fs/cramfs/README
  - add note about sorted directory entries
  - mkcramfs compile-time DO_HOLES option replaced by run-time -z option
  - update tools section
  - add note about PAGE_CACHE_SIZE possibly changing on arm and ia64

fs/cramfs/inode.c
  - statfs->f_namelen = CRAMFS_MAXPATHLEN

include/linux/cramfs_fs.h
  - add CRAMFS_MAXPATHLEN (252)
  - clarify CRAMFS_SUPPORTED_FLAGS definition

scripts/cramfs
  - directory removed, the cramfs user-space tools are now located at
    <http://sourceforge.net/projects/cramfs/>.

Here are two small compile fixes for x86-64 in 2.5.6pre2.
- Remove call to schedule_tail.
- Fix inline assembly of semaphores to work with latest gcc 3.1.

This patch adds unicode support and wants to be applied on top of the LFS
one. It uses a fake nls module to do the (little endian) unicode
translation.

This patch adds LFS and moves some smb operations into per-protocol level
structs. It wants the nls patch to applied already.

Fixes smbfs oopsing on failed nls translations and maps unknown chars to
:#### strings. Also PATHLEN vs NAMELEN mixups.

This patch enables PCI buses on nodes above node 0 for
the NUMA-Q architecture. It also enables node-directed
port/IO, and cleans up a couple of tiny things that only
affect CONFIG_MULTIQUAD.

  
the patch is against 2.5.6-pre3 and contains:
        - ethtool support;
        - using mii.h for the MII registers and constants;
        - 2 more device/vendor IDs added;

This is one of the very early steps on cleaning up the SCSI cdrom
driver.  It gets rid of directly accessing the scsi_CDs array in favour
of using the handle we get from the generic cdrom layer.  Also uses
local vars instead of many grouped scsi_CDs accesses in other places.

The gain is to get rid of the global, static array of CDROMS
in the end.

ehci-0306, iso, philips, speedups
  
      - adds preliminary highspeed ISO support
      - tweaks the driver to support the Philips EHCI
      - does less in the IRQ handler
      - avoids accessing one immutable PCI register
  
The ISO support should be enough to start writing
drivers, not that I know of any ISO devices that are
really available yet, but it's not fully cooked yet.
  
As a functional milestone, this means Linux now
handles all kinds of highspeed device I/O.  (But it
doesn't yet handle split periodic transactions, to
full or low speed devices through USB 2.0 hubs.)
    
Thanks to Rory Bolt for the non-ISO bits here!

kill all subsidiary threads in a thread group when the main thread
exits.

Features:

 - It sends the subsidiary threads SIGKILL with SI_DETHREAD.

 - Subsidiary threads doing an execve() just leave the thread group (rather
   than forcing the master thread to do an execve() which would be more POSIX
   like).

hcd-0305, periodic and pci fixup
  
      - removes the pci dependency you mentioned in the
        rh_string code (friendlier to non-PCI HCs)
      - makes code match doc (8859-1 chars, not just ascii)
      - adds sanity checking for the periodic transfer interval,
        and forces it to a power-of-two (code can leave HCDs)
      - facilitates better IRQ sharing

	- removed dependancy on net/irda header files from the driver.

 - changed printer.c to use dynamic urbs, as that is now necessary.

These are the small subset which were obviously correct.

Evgeniy Polyakov <johnpol@2ka.mipt.ru>:
	Patches check return values for request_region() and misc_register().
	This patches make janitorial project TODO list a bit smaller.

Stephen Rothwell <sfr@canb.auug.org.au>: [PATCH] APM idleing fix:

	This bug slipped back in with the need_resched() macro substitution.

atomic ops are *not* barriers any more.

Sebastian Wilhelmi <wilhelmi@ira.uka.de>: Re: Question on your "Unreliable Guide To Locking":
  > Yes, this is no longer true.  The modern assumptions are that they are
  > not barriers.

set.

Declare pnpbios_init as returning int, as __initcalls are supposed to.

The compiler warns about about crd_load being defined but never used with
my config. The appended patch avoids compiling in the unused code in this
case.

Several people (including Alan Cox) on lkml claimed that the BIOS
returns CPU addresses, so using phys_to_virt is actually correct -
and it makes my kernel compile again.

As vesafb only compiles on i386, using a more portable API isn't useful,
anyway.

irtty.c includes irqueue.h which includes linux/cache.h (via
asm/processor.h <- asm/thread_info.h <- linux/thread_info.h <-
linux/spinlock.h)

both irqueue.h and cache.h defines a ALIGN (for different
purposes).

This patch renames ALIGN in irqueue.h to IRDA_ALIGN.

into penguin.transmeta.com:/home/penguin/torvalds/repositories/kernel/linux

makefile rules..

inside the kernel on x86 (TF is writable)

into penguin.transmeta.com:/home/penguin/torvalds/repositories/kernel/linux