A while ago some guy complained that his server connections / sec was
way down due to AS. The server would fork a process which would then read
some tiny bit of random data and hang around for a while doing nothing.
This defeated all AS's heuristics, and due to a new process being a good
anticipation candidate by default, his performance went bad.

The fix was just to treat new processes as being unsuitable for
anticipation. Fixed. Unfortunately ls, cat, grep, small programs
starting up, find | xargs, etc only submit a couple of requests, and they
really want anticipation to be on.

This patch keeps a simple goodness value per queue based on how successful
anticipation has been for new processes.

Clears the next_arq cache if a request is repositioned and added as an alias
behind another.  Should fix the BUGs which Mary Edie Meredith has been
reporting.

Also, fixes a few random (but harmless) cases where last_merge was being set
and causing warnings.

When 2 requests are merged, and one is thrown away, move its alias list onto
the merged request! This is likely to be the cause of Mary's hangs.

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

       grabbing of the pads, as well as simplifies the mousedev driver.

       The infoRot180 and infoPortrait bits are for information
       only. The touchpad uses the same X/Y coordinate system
       regardless of the orientation, so the software shouldn't
       care about these bits.

       untranslating to the AT keyboard driver as well. Lower
       PS/2 mouse default report rate. Fix repeat rate adjustment
       ioctls accordingly, and update other files to reflect the
       changes. This should fix most known keyboard problems in 2.6.

       support to mousedev.

       uses the variable.

This fixes a few x86-64 issues that have crept in and broke bootup.

 - Disable -funit-at-a-time. It breaks symbol exports with gcc 3.3.1-hammer.
 - Fix sched_clock to not access HPET.  The new scheduler uses it more
   extensively and it is not mapped early enough.  I opted to just
   disable the HPET access, because even a slightly non monotonous TSC
   should be accurate enough for scheduling purposes. 
 - Fixes for separate objdirs by Arnd Bergmann
 - Fix two warnings that have crept in

Previously, we called blk_queue_make_request(q,mddev->pers->make_request)
*before* calling mddev->pers->run(), but this left a hole in which the
device could be accessed before it was initialised.

So we moved blk_queue_make_request to *after* ->pers->run(), but now some
of the initialisation done in ->run is over-written by blk_queue_make_request(), 
particularly limits like ->max_sectors.

So now, we just open-code the one line of blk_queue_make_request that we
need at this point.

All the rest has been done by a separate called to blk_queue_make_request, 
either when the mddev was first allocated, or when a previous incarnation
of the device was stopped.

This fixes "bio too big" error that occured due to max_sectors being too large.

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

From:  Domen Puncer <root@coderock.org>

Don't know if the patch I released few days ago was still
applied. This is a minor revision of that patch which converts
saa9730 to spinlocks thus removing save_and_cli() and
restore_flags() calls.

Regards,
Angelo Dell'Aera

This does two things:

1) Fix alignment problem on PARISC64 (and maybe other 64bit archs.)
2) Add another couple of cards to the table.

Cheers,

Simon.

If baycom driver has never been opened, it will attempt to free
an IRQ that it never registered when removed.  The problem is that hdlcdrv
does not keep track of open/close state.

Update baycom drivers for 2.6.0-test5
	- get rid of MOD_INC/DEC
	    (looked into hdlcdrv_ops and don't need to have owner field because
	     baycom drivers unregister on  unload).
	- use alloc_netdev instead of static device structures.
	- hdlcdrv_register returns device instead of getting passed one.
	- put private data in space allocated at dev->priv in alloc_netdev
	- shorten name of hdlcdrv_register_hdlcdrv to hdlcdrv_register

I don't have actual baycom hardware, but driver builds and loads/unloads.
Real hardware initialization doesn't happen until open.

The first one is the important patch, the other three are just code review
type cleanups.

From:  Nicolas Kaiser <nikai@nikai.net>

From:  Nicolas Kaiser <nikai@nikai.net>

From:  Nicolas Kaiser <nikai@nikai.net>

From:  Nicolas Kaiser <nikai@nikai.net>

  Cheers, n.

From:  Nicolas Kaiser <nikai@nikai.net>

From:  Nicolas Kaiser <nikai@nikai.net>

From:  Nicolas Kaiser <nikai@nikai.net>

From:  Nicolas Kaiser <nikai@nikai.net>

Patch 3/5 (Update to version 6.17)
* Add: Removed SkNumber and SkDoDiv
* Add: Counter output as (unsigned long long)

Patch 1/5 (Update to version 6.17)
* Add: Better parameter check
* Add: UDP and TCP HW Csum changes
* Add: Interrupt Moderation infos

Add proper management of jiffies and time values, rather than punting
and doing long long arithmetic on usecs.

Still more driver data which was updated but never used.

Convert bogus test_and_set local wait, to a real spin_lock so it
has a chance of working on an SMP.  This also does the right thing
and locks out interrupts while giving commands on UP;  maybe the
comment in Kconfig was because there was never a proper mutex...

Don't have real hardware to try this, but it can't be worse than
the previous code.