When foo is unlinked, nfs_unlink() does a sillyrename, this puts the
dentry on nfs_delete_queue, and (in the VFS) unhashes it from the
dcache.  This causes problems, since any later access to the
silly-renamed new .nfs file will create a NEW dentry that aliases the
one we originally created, but unhashed.

This causes various confusion, especially if we want to try to delete it
again later.

So fix this by not unhash the dentry after silly-renaming.  In 2.2, each
fs was responsible for doing a d_delete(), in 2.4 and later it happens
in the VFS layer and I think it was just an oversight that the 2.4 VFS
doesn't consider sillyrename (considering the code and comments that are
cruft).

Also fixed up some comments while debugging this.

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

[From Patrick Caulfield]

Pull dm_div_up() out of dm-table.c into dm.h

Some targets may perform io of their own volition, eg. a mirror
performing recovery, a cache target pulling in different chunks.  We
cannot let them perform this io while the device is suspended.  This
patch adds 2 new methods to the target type, which instruct the target
to suspend/resume itself.  All targets start in the suspended state,
so should expect an initial resume call.  Simple targets do not need
to implement these functions.

Replace __HIGH() and __LOW() with max() and min_not_zero().

This removes the last sources of "non-ANSI parameter list" warnings for
zlib_deflate.

Reduce MAX_MEM_LEVEL to 8.  This reduces zlib memory consumption by
128k (from ~400k to ~270k) at the theoretical cost of worse
compression.  No code currently in the kernel actually uses the better
compression, so the practical cost is zero.

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

chr_dev_init() should be a subsys_initcall(), since it needs to
initialize before any drivers that use the character device
infrastructure.

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

into mars.ravnborg.org:/home/sam/src/linux/kernel/bk/docbook

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

from the MAINTAINERS list. How they're HP, and maintained
by Stephen Cameron.

More merging from zlib-1.1.4

force windowBits > 8 to avoid a bug in the encoder for a window size
of 256 bytes. (A complete fix will be available in 1.1.5).

James Carlson:

        The problem is that s->strstart gets set to a very large
        positive integer when wsize (local copy of s->w_size) is
        subtracted in deflate.c:fill_window().  This happens because
        MAX_DIST(s) resolves as a negative number when the window size
        is 8 -- MAX_DIST(s) is defined as s->w_size-MIN_LOOKAHEAD in
        deflate.h.  MIN_LOOKAHEAD is MAX_MATCH+MIN_MATCH+1, and that
        is 258+3+1 or 262.  Since a window size of 8 gives s->w_size
        256, MAX_DIST(s) is 256-262 or -6.

        This results in read_buf() writing over memory outside of
        s->window, and a crash.

Don't think anyone actually bothers to check specific error codes, but
it shouldn't hurt either.

This removes unnecessary NULL casting.

s/Z_NULL/NULL/g.

Simple s/local/static/.

They may have some new HP address, unknown for now.

Most of it is reformatting, but the functional bits should fix real
problems.  A loop is introduced, just like in the turboc patch and one
of the three condition bodies has been expanded.

This is the first bit of the missing merge towards 1.1.4.  Applies on
top of the previous cleanups.

This one rips out an ugly #ifdef and seems to catch a theoretical
error possibility.  Always thought that they fixed more than they
officially admitted.

We don't use any cplusplus in the kernel.

The old 26-bit ARM support was long since dropped out of the regular ARM
support, since it was different enough to not make sense to maintain as
one port.

This re-introduces arm26 as an architecture of its own.

lib-y is the new way to define what objects belongs to a library.  The
implementation was not made backwards compatible and therefore an update
to all architectures are needed.

This is a simple replacement of obj-* to lib-* and deletion of L_TARGET.
The new mechanish where lib.a can be mixed with built-in.o is not
utilised.