The current ip_fragment evictor kills the oldest entry of each hash bucket
starting with 0 instead of killing the oldest entry of all buckets. This 
leads
to unfair behaviour if one of the higher hash slots carries alot of 
fragments.
This patch holds the frag heads in a lru queue so we can kills the least
recently used first. Each arriving fragment counts as usage.

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

When compiling random.c without SYSCTL defined a warning is
issued about free_entropy_store being defined but not used.
Put ifdef's around the functions.

I could have moved the whole function, but it logically belongs
to this part of the file.

When compiling without PROC_FS enabled a warning is issued about
proc_dma_show defined but not used.

Move both versions of proc_dma_show inside the #ifdef CONFIG_PROC_FS

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

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

The 3 lines being deleted are repeated from above (in same file).
I pulled the 2 lines being added from an earlier version of the patch.

This is an update to the EISA/sysfs code :

 - Separate bus root code from generic code.
 - Add driver for PCI/EISA bridge.
 - Hacked parisc eisa driver so it can act as a root device.
 - Add driver for so-called virtual root (for bridge-less systems).
 - Allow multiple roots.
 - Moved quirk_eisa_bridge from alpha to generic code, since
   the EISA code needs it now (on x86, for example...).
 - 3c59x :
   Prevent driver from returning ENODEV in case it has registered with
   the EISA framework, but no device has been found yet (it happends
   when the driver is built into the kernel, and the EISA bus root has
   not been discovered yet).

It's been discussed on lkml, and all suggestions (mainly from Ivan
Kokshaysky <ink@jurassic.park.msu.ru>) have been taken into account.

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

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

jfs_sync_fs was implemented using the same code as the unmount code to flush
the journal and wait for the journal to quiesce.  Since jfs_sync_fs may  be
called while the volume is under heavy use, we can end up waiting
indefinately.  Code in jfs_flush_journal meant to detect a hang at unmount
time may be triggered in this case causing a trap.  This patch changes
jfs_sync_fs to only wait until the most recent transaction has been
commited to disk, rather than waiting until the commit queue is empty.

state changes due to execve() and exit(). We need to hold the tasklist
lock to guarantee stability of "task->sighand".

Update the ARM signal handling to use the generic
get_signal_to_deliver() implementation.  There is a caveat though -
ARM has no hardware support for single stepping, and thus needs the
kernel to help provide this.

We introduce a new ptrace flag for this, PT_SINGLESTEP, which
indicates to architecture specific code that it should take extra
actions in the signal handler to set and/or clear breakpoints as
appropriate.

From Geoffrey Lee <glee@gnupilgrims.org>.

Delay the marvel agp hose printk until after the check for whether a
valid AGP hose was found. Otherwise if none was found, hose->index in
the printk arg list dereferences NULL.

/jeff

into kanga.twiddle.net:/home/rth/linux/axp-2.5

  - AC'97 - added quirks for tuning of hardware configuration
  - PCM midlevel - improved ring buffer allocation
  - sa11xx-uda1341 - small fixes
  - ISA - fixed ring buffer pointer calculation
  - OSS PCM - fixed deadlock (rate plugin)
  - timer - fixed deadlock when user interface is used
  - CS4232 - added detection of plain CS4236
  - USB driver - added boot quirk for SoundBlaster Extigy

We really need fixdep before we can compile anything...

(Pointed out by wli, and others)

The checksums are generally 32-bit (unsigned int), though they are saved
in symbol values, which may 64-bit, so we need to convert.

Pointed out by Anton Blanchard.

I fixed scripts/Makefile.modpost to handle CONFIG_MODULES but no modules
selected, but forgot scripts/Makefile.modinst

(pointed out by akpm)

close succession. However, for this once we'll just call it "inspired".

But let's decide pair the lock with an unlock anyway, even if it is
boring and "square".

Urban Widmark points out that modprobe calls system() in many
configurations, which drops privs since request_module() doesn't
set ruid and rguid.

This gets rid of exec_usermodehelper and makes everyone use
call_usermodehelper, which has a new "wait" flag.

Fix compile errors after futex merge.

Patch from Guennadi Liakhovetski

An update of the partition-table and a typo-fix. Please, apply.

General updates to these two watchdog drivers, including:
- converting initialisers to C99 syntax
- correcting module descriptions
- ensuring private symbols are declared static
- ensuring correct options are reported in the watchdog_info structure
- ensuring we use the right clock value for calculating timeouts
- ensuring failure cleanup paths are properly implemented
- don't use MOD_{INC,DEC}_USE_COUNT

A compliment is an expression of esteem, respect, affection, or
admiration.  As far as I can tell, this word does not yet have a
place in the kernel source.

A complement on the other hand, is what is meant in the following
places.