Since it gets 1 device right it wasnt hard to fix 8)

I'm looking into the other IDE problem from the merge - several people
see hangs. Bartolomiej has found one suspicious looking candidate. I'll
try and pin it down ASAP.

- Redo the timing stuff using jiffies properly
- Clean up the exit paths
- Make the ioctl use a semaphore
- Fix broken locking on the AEN list

Two problems with the original change

1. We should return bytes actually processed on an error according to
SuS/POSIX. Technically the EFAULT path is outside the spec but its best
we follow

2. We need to fix most of this anyway because the final section of the
change was wrong. If retval was set we retried and got an efault again
in some cases

I think this way of doing it is right but it could do with further
review

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

* Increase default Rx descriptors from 80 to 256 to give
  better Rx buffering capability in the case of heavy
  Rx load with small packets.

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

Fix incorrect PCI cache line size assumptions on i386 and thus
avoid potential memory corruption with Memory Write-and-Invalidate.

Comments from rmk:

Make setup-bus.c properly aware of cardbus bridges.  We treat the
bus behind a cardbus bridge more or less like any other bus, except
we don't explicitly descend below.  We do, however, explicitly
reserve IO and memory space as we have done in the past.  Memory
space is doubed to 32MB as a measure to allow the Mobility
cardbus-pci stuff to work.  The amount of space reserved is now
specified by a couple of #defines at the top of the file.

This allows pci_bus_assign_resources() and pci_bus_size_bridges()
to be called for both root buses as well as cardbus secondary buses.

Comments from Ivan follows:

This patch combines your(rmk) cardbus changes (formerly pci-11)
and my "arbitrary resource layout" stuff. This + current bk works
on nautilus.

Most interesting feature: this can be used on partially
allocated PCI tree. For instance, i386 PCI code has always been
absolutely helpless wrt incorrectly initialized p2p bridges.
Now it can just call pci_assign_unassigned_resources() in the
end of PCI init and it would fix following problems:
- completely uninitialized bridge windows (with base and limit 0);
- erroneously "closed" windows;
- windows overlapping with something else.

  
Minor cleanup: don't call pci_update_resource() for bridges,
get rid of bogus "trying to set non-standard region" messages thus.

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

head-netwinder.S was a work around for old Netwinder NeTTrom
firmware.  It is no longer required.

This reduces stack usage in drivers/net/wireless/airo.c by
dynamically allocating 2KB buffers in 2 places.

This reduces stack usage in drivers/cdrom/optcd.c by
dynamically allocating a large (> 2 KB) buffer.

RPCSEC_GSS options are related to (only) nfs/nfsd, so it is
more logical to ask RPCSEC_GSS questions immediately after
nfs/nfsd, not after half of screen more questions.

NCPFS related options should appear immediately below
ncpfs question, not after Coda and RPCSEC...

As the netfilter folks don't seem to have any interest in 2.5 currently
I decided to fix their last uses of those old module interfaces myself.
The implementation (get a reference first and release it again when
not actually needed) might be slightly suboptimial but the netfilter
team should just fix it if/when they get any interest in Linux 2.5/2.6.

Also fix the MOD_INC_USE_COUNT/MOD_DEC_USE_COUNT to give more accurate
deprecation warnings.

Devfs tries to be super smart and rereads partition tables at all
kinds of wierd points.  This breaks a bunch of stuff were you
can't get the right disk changed information (i.e. CompactFlash).

If people actually need this kind of stuff they should just call partx
from devfsd instead of relying on the kernel doing something like this.

Cleans up the devfs code significatnly (aka removes tons of junk)

Not use anywhere else nor should it.

Adrian Bunk noticed that there's an exposed config option to use the
now gone DEVFS_FL_AUTO_DEVNUM in the dvb code - remove it.

devfs_d_revalidate_wait adds to a waitqueue but never removes from it
again so we there's one entry full of reused stack space added on
each call (I wonder how this ever worked).

The function has a few more bugs (it effectivly does a sleep_on instead
of checking for the actual even and can't deal with negative dentries
at all), but I just had breakfast and don't want to poke into devfs
internals deeper - I still hope Adam's smalldevfs will get merged
anyway..

fix yet another manual reschedule point

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

Also add abort functionality