(Ralf said to send this straight to you)

Rusty Russell <rusty@rustcorp.com.au>: Trivial MIPS baget patch to remove minmax macros:
  In favour of kernel.h one:

Roger Luethi <rl@hellgate.ch>: DMA-mapping.txt: sg_dma_length() -> sg_dma_len():

[UTS_MACHINE is now defined in the (generated) compile.h]

David Gibson <david@gibson.dropbear.id.au>: [TRIVIAL PATCH] Fix order of #includes in init_version.c:
  Linus, please apply.  compile.h must be #included before uts.h, or
  uts.h will define UTS_MACHINE (incorrectly) which is then redefined in
  compile.h.

David Gibson <david@gibson.dropbear.id.au>: TRIVIAL: Remove warning in fs_nfs_nfsroot.c:
  Linus, please apply.  The patch below removes a warning in
  fs/nfs/nfsroot.c by including a header file providing a prototype for
  in_aton().

Andrey Panin <pazke@orbita1.ru>: printk() cleanup in ide-pnp.c:
  Attached patch adds a few missing printk levels in ide-pnp.c file.

Andrey Panin <pazke@orbita1.ru>: drivers_net_sundance.c: missing __devinit:
  this patch adds missing __devinit modifier to sundance.c network driver.
  Patch against 2.5.15. Compiles, but untested.

"Maciej W. Rozycki" <macro@ds2.pg.gda.pl>: [patch] 2.4.19-pre7: A few declance multicast updates:
  Hello,

   It seems all Ethernet device drivers were bulk-converted to use the new
  common CRC functions.  I discovered declance used incorrect endianness to
  calculate the sum for its multicast filter and had a few alignment
  problems there.  I fixed these bugs in the MIPS/Linux CVS tree which is at
  2.4.18 now.  Here is the respective update for the official kernel.

   The bugs make the filter non-functional.  Please apply.

    Maciej

Andrey Panin <pazke@orbita1.ru>: drivers_net_epic100.c: missing __devinit:
  This patch adds missing __devinit modifier for read_eeprom() function.
  Patch against 2.5.9. Compiles, but untested.

	nfsd_acceptable() calls permission() on a a dentry that is
not guaranteed to stay around (its child is pinned down, but there
is no promises that child won't move and nothing pins dentry itself).
Fixed, cleaned up.

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

into tp1.ruhr-uni-bochum.de:/home/kai/kernel/v2.5/linux-2.5.make-built-in

Use the just introduced Rules.make extension to clean up 
drivers/net/*/Makefile.

New-style Makefiles have a nice way of declaring objects
which need to be built either built-in or as modules:

        obj-$(CONFIG_EEPRO100) += eepro100.o

However, handling objects in subdirectories, which need to be
built and linked is not as nice:

        subdir-$(CONFIG_E100) += e100

        ifeq ($(CONFIG_E100),y)
          obj-y += e100/built-in.o
        endif

This means we descend into the subdirectory when building
vmlinux / modules, depending on CONFIG_XXX. When we are building
vmlinux we also need to link whatever has been built in the
subdirectory, so we add it to $(obj-y) at the appropriate place
(link order is important).

Now, the extension below allows to rewrite the second case into

         obj-$(CONFIG_E100) += e100/

which looks much nicer ;-) Existing behavior is not changed, and the
only prerequisite to using the extension above is that the O_TARGET in
the subdir is named "built-in.o".

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

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

Currently fs.h is full of unrelated declarations and included in almost
any source file. Thus it makes sense to spilt certain aspects out that are
only used by few users.

This patch starts with the namei/path lookup interface and splits it into
<linux/namei.h> which is now directly included by the 24 files that actually
need it.

Transform new quota code to use sysctl instead of /proc/fs.

	default encode_fh() doesn't take dparent_lock around messing
with dereferences of ->d_parent->d_inode->i_....  Fixed.

	OK, it turned out that since the last time I've done that audit
a couple new users of d_parent had appeared.  First, more crapectomy (strictly
speaking not necessary, but just look at that code - where the hell does IBM
find them?)

 - Apply small host chip driver cosmetics by Andrej xxx Panin
   and Vojtech Pavlik.

 - Remove support for "disc recovery time". It could only supposedly
   help with really simplistic broken devices from the past,
   which didn't have moderately sophisticated controllers.
   And finally Vojtech voted for it as well... so I just trust him.

 - Apply icside host chip driver and other ARM related updates by Russell King,
   which finally settle the "portability" work a bit, well hopefully.

 - Make the different ATAPI device type drivers use a unified packet command
   structure. We have to start to push them together.

This patch is rather trivial in itself, but the plentora of code duplication it
is trying to fight against is making it unfortunately rather big...

 - Nuke COMMERIAL and similar spurious configuration options...
   The fact that every single default configuration option contained
   those bits makes this trivial patch appear rather big.

 - Move ll_10byte_cmd_build to the only place where it's used: ide-cd.  The SCSI
   layer does have it's own implementation which additionally it's messing
   around with the hard_nr_sectors struct request value.  One should *not*
   provide "infrastructure" until its really used as such.

   If anywhere this should reside in a file called ATAPI.

 - Unfold the INIT_REQUEST macro from blk.h. This showed up plenty of duplicate
   checks for QUEUE_EMPTY. Clean them as well. Remove the over cautious
   major(CURRENT->rq_dev != MAJOR_NR) checks. During the last several years I
   never saw any report about it. Looking at the !CURRENT->bio it is clear that
   dereferencing NULL will provide the same kind of panic as the check.  Some
   comments around the code in question show nicely that indeed INIT_REQUEST
   was a good example of code obfuscation.

 - A short look at RQ_INACTIVE shows that it is only used inside the scsi.c file
   and during the removal of devices. This shows that the many checks for
   RQ_INACTIVE are not necessary. Looking closer even shows that some of them
   did happen before checks for an empty queue. Plenty of drivers didn't care
   about it and the CD-ROM ones should be handled properly, because the
   most common drivers would fail as well. Comments indicate that this
   was an leftover from 1.3 days...

 - Apply cleanup of host chip drivers by Bartomiej Zonierkiewicz:

     affected drivers - aec62xx.c, alim15x3.c, cmd64x.c, hpt34x.c, sis5513.c

     new tuning scheme (wip) part 1:
	- introduce ratemask()
	- use ata_timing_mode()
	- use ide_config_drive_speed() return value

     forward port from convert.10:
	- support for AEC6280, AEC6280R
	- misc cleanups

     I had to fix a small typo in sis5513 code...

 - Add a new entry for an old VIA cell hiding as something new.
   (Pointed out by Kees Bakker.)

 - Make the synchronization token active resident on the same level as the
   spin lock. They interact with each other.

 - Synchronize with linux-2.5.17.

 - HPT366 driver typo fix by Andries Brouwer.

 - Export udma_tcq_enable() symbol right now. The blk_get_request() is undefined
   as well.

	Unlike other drivers, rd.c wants block size to be set once an
forever (for everybody else setting block size as high as possible and
letting filesystems change it with set_blocksize() is OK; rd.c treats
invalidate_buffers() as "kill the ramdisk contents".

new helpers for seq_file - for cases where we don't have a non-trivial
iterator and just want to use seq_{printf,putc,...}.

sane readdir() for ramfs-style filesystems

kill ->i_op->revalidate()

massive crapectomy in s390 code - removed procfs abuses

 add proper ->getattr(), clean up do_getattr() and friends.

- TLB infrastructure changes
- Make flush_tlb_pgtables not need to cook up
a dummy vma
- Update for do_fork return value change
- Update defconfig

into nuts.ninka.net:/home/davem/src/BK/sparc-2.5

People relied on the build generation number to find out if they actually
booted the right kernel, only incrementing it on config changes isn't
enough for them. So this patch goes back to the old behavior, only done
right this time:

We now don't increment the generation number on every invocation of make, 
but on every link of vmlinux. At the same time, this patch fixes the
Makefile to only relink vmlinux if any of its prequisite object files
(or the command line) changed. Incrementing the version number
happens *after* we decided to relink vmlinux.

Before overwriting version.h, check if it actually changed - otherwise
we will unnecessarily rebuild a lot of files, as e.g. module.h depends
on version.h, and many files include module.h

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

It seems by the comments the statements were only there to prevent the
thread from finding itself as TASK_RUNNING after a kernel preemption.
Since we reverted to the original preemption behavior (which does not do
that) it should be safe.  Looking over the code confirms - it seems to
not be doing anything evil and is properly locked otherwise.

As andrew pointed out, it is bad idea to run_task_queue() when
spinlock is held. This fixes it (only could be triggered by user
pressing Magic-D). There's minor problem left where Magic-D could be
delayed/ignored under high load. I'll either fix that or kill Magic-D
support (with acpi, it is very easy to trigger swsusp by echo 4 >
/proc/acpi/sleep, so magic key is probably not neccessary any more).

This implements a get_cpu() and matching put_cpu() to safely hand out
the current CPU to avoid preempt races.  Andrew and I have been bitching
about the need for such a method.

I also went ahead and replaced an example of current explicit
preempt-off with the new methods, as a case in point.