Synchronize all walks of the device and driver lists of a bus with an rwsem wrapped
around the entire iterator, instead of using device_lock and dropping it after we 
grabbed each node. 

Note this also prevents deadlock when walking the list of drivers and calling 
get_driver(), since get_driver() tries to take device_lock while we already have it
held.

device classes join the club of devices, buses, and drivers. They get a 
->present flag, which is set on registration and cleared on unregistration.
They also get get_devclass() and put_devclass(), which, you guessed it, 
bump the reference count. get_...() of course checks the present flag and
returns NULL if clear. 

In the spirit of devices and buses, change driver refcounting model to 
match the way that devices and buses are done.

struct device_driver gets a ->present field, which is set on registration 
and cleared in driver_unregister(). get_device() checks the state of this
flag and returns NULL if cleared. 

Note that the horribly wrong remove_driver() is deprecated and simply BUG()s
when called. Please convert callers to use driver_unregister(). Updates to 
callers will be coming soon.

Note also that this still doesn't fix the race in which a driver module can
be removed while the refcount on a driver > 1. Near future work should help
to remedy it, but no solutions are guaranteed..

Based on recent changes to device refcounting/unregistration.

struct bus_type gets a ->present field, which is set when the bus is 
registered. bus_unregister() is added, which only clears ->present and 
calls put_bus(). get_bus() checks the ->present and returns NULL if clear.

This allows all current references to buses to keep valid structures, while
preventing any new references to the bus structure.

into osdl.org:/home/mochel/src/kernel/devel/linux-2.5-core

There are races WRT walking the lists of the buses and classes and the 
device being unregistered while it's the current node we're accessing. If 
the node was deleted, ->next would point to itself, and we'd sit in an
inifinite loop (until that memory was freed from underneath us and we Oops'd).

This changes device_unregister() to only mark the device as not present. put_device()
deletes the device from all the lists it belongs to, while holding device_lock,
then drops it to call device_del().

This should guarantee that everything in the device is valid until the last reference
goes away, regardless of whether it's still physically present or not. 

If you compile the kernel with -ffunction-sections, each function gets
put in a section .text.function_name.  This collides with our current use
of .text.init.  So here's a patch which converts x86 to use .init.text
instead.

I've tested it on x86 and it still frees 120k of ram, so it seems to work.
Other architectures will need to change their vmlinux.lds appropriately,
and may need other changes (arm, m68k seem to use .text.init verbatim).

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

into samba.org:/scratch/anton/linux-2.5_ppc64_work

into samba.org:/scratch/anton/linux-2.5_ppc64_work

[PATCH] md - 1 of 1 - Register mergeable function for linear so requests don't cross device boundries

We abstract out "which_dev" which finds which device a sector is on
and us it to make sure both ends of a to-be-merged request will
be on the same device.

As the expkey cache references entries in export cache,
it should be flushed first.
Also, the auth_unix caches should go last.

This brings the linux-2.5.42 tree up to spec WRT setclientid and
setclientid_confirm

Switch knfsd to vfs_read/vfs_write to work on aio-only filesystems.
This also gets stuff like the LSM checks and mandatory lock checking
for free.

o don't define __NO_VERSION__ (no use in 2.4/2.5)
o don't include <linux/version.h>

... as generic names probably aren't such a good idea.
Also sort the related EXPORT_SYMBOLS.

A 'qword' is a 'quoted word' and can be added, added as hex,
or got (getted??).

When releasing a cache item we always release anything
that is part of the key, but we ONLY release the content
is if it is valid.  i.e. if CACHE_VALID but not CACHE_NEGATIVE.

The new caching and auth code needs symbols to be exported.

An rpc service like NFSD knows about a number of
versions (2,3,4) and needs to know the max and min valid
version numbers, and also the maximum xdr buffer size used
by any version.

These max/mins are currently computed at compile time which means
we need to include all the xdr*.h headers into nfssvc.c just for
computing a couple of numbers.

With this patch, these max/mins are calculated at run time from
information in the svc_version structure which has been extended to
include per-version xdrsize.

Also the nfsd_version2, nfsd_version3, and nfsd_version4
structures are moved from nfssvc.c to their more correct home
of nfs*proc.c

Code gets neated, xdr*.h are no-longer included in
nfssvc.c, and the ugly ifdef can be removed fom xdr4.h

With gcc-2.95.3, text size goes from 2,399,190 to 2,390,004 - nearly
10 kbytes.

- Don't inline mpage_writepage().

  Even though it had but a single call site, uninlining this function
  took mpage.o from 4529 bytes of text to 3985.  gcc-2.95.3.

- Various buffer.c tweaks.  13233 bytes down to 12977

I had some code in page reclaim which is designed to catch the caller
of write(2) in page reclaim and force him to wait on writeout of his
own spindle's pages.

It works too well - there are large stalls in things like `cp
large-file foo' on small-memory machines.

So remove it, and rely on the other, more predictable throttling
schemes.

From Marcus Alanen <marcus@infa.abo.fi>

vmalloc is allocating one more page than necessary.

From Christpoh Hellwig.

Make filemap_sync() static, and not exported to modules

From Hugh Dickins

simple_rename, like shmem_rename until I fixed that a few weeks ago,
got the parent directory's link count wrong in the exceptional case
of renaming a directory in place of an existing empty directory.

Patch from Hugh Dickins.

There's a loop in there which goes infinite if the active list is full
of zero-ref pages.

Leftovers from the kmap_atomic conversion

From Peter Chubb

OK, make it X86 only.  64-bit architectures now get it for free;
if/when PPC32 (the only other 32-bit architecture I can seriously imagine
needing it, except possibly s390) fixes do_div() and needs large block
support, then enable it back there again.

There isn't as far as I know a flag I can test in a config.in file
that distinguishes 64 from 32-bit architectures.  The intention was to
enable the choice of CONFIG_LBD where it made sense;  it doesn't make
sense for 64-bit architectures; and rather than enumerate all the
32-bit architectures, I just chose two that I knew were used in
servers currently.

Use 'p', not 'j'

Fix a couple of things for no-io-apic builds:

- mpparse.c wants BAD_APICID

- mpparse.c (an perhaps others) want Dprintk()

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

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

This fixes a recent dyslexicism in cpqarray.

Use cond_resched() instead, and add a comment to explain.

Thanks again to davem...

The following patch fixes 2 obvious typos. Thanks to davem and George
Anzinger for pointing them out.