Mainly sync w/ 2.4.x version.

Remove unused local `fattr' in nfs_readpage_result().

From: davej@codemonkey.org.uk

We currently run sync_blockdev() on every close of a blockdev.  This patch
arranges for the sync to happen only on the final close.

From: Brian Gerst <bgerst@didntduck.org>

- Reduce code duplication by putting the kmalloc cache sizes into a header
  file.

- Tidy up kmem_cache_sizes_init().

- If two CPUs run register_chrdev_region(major == 0) at the same time they
  can get the same major.

  Fix that by extending the lock coverage.

- local variable `cd' was leaky on an error path.

- Add some API commentary.

We just made bdevname() irq-safe.  This patch makes cdevname() callable from
interrupts as well.

From: Andi Kleen <ak@muc.de>

Fix a warning on x86_64 (and presumably ia64) which arises from casting an
int to a pointer.

From: Gerd Knorr <kraxel@bytesex.org>

2.5.x kernels don't look at the VM_DONTEXPAND flag when merging multiple vmas
into one.

From: William Lee Irwin III <wli@holomorphy.com>

inode_init() wants to deal with things in in units of the size of
struct hlist_head, not struct list_head.

From: William Lee Irwin III <wli@holomorphy.com>

__blk_queue_bounce() contains an open-coded page_to_pfn() for the
discontig, non-MAP_NR_DENSE() case (wherever MAP_NR_DENSE() went).
This converts it to use the standard page_to_pfn() abstraction.

Spotted by Oleg Drokin: remove a couple of local variables which aren't being
used for anything.

kmap() can sleep, but rarely does.  Add a check for kmap() being called from
inappropriate contexts.

There is never a need to write out b_assoc_buffers() in ext3_sync_file().

From: Andries.Brouwer@cwi.nl

Adding the unregister_chrdev_region call that is the counterpart to
register_chrdev_region, we get a nice cleanup of tty_io.c.

From: Patricia Gaughen <gone@us.ibm.com>

The boot-time ioermap code needs to invalidate the tlb entries after setting
up the new pte.

From: john stultz <johnstul@us.ibm.com>

This patch allows one to manually specify the i386 gettimeofday time-source
by passing clock=[pit|tsc|cyclone|...] as a boot argument.  The argument will
override the default probled selection, and in case the selected time-source
not be avalible the code defaults to using the PIT (printing a warning saying
so).

It is generally illegal to wait on an unpinned buffer - another CPU could
free it up even before __wait_on_buffer() has taken a ref against the buffer.

Maybe external locking rules will prevent this in specific cases, but that is
really subtle and fragile as locking rules are evolved.

The patch detects people calling wait_on_buffer() against an unpinned buffer
and issues a diagnostic.

Also remove the get_bh() from __wait_on_buffer().  It is too late.

ext3 is only permitting files of up to 32G.  It should be allowing files to
be up to 2TB.

This is because it is running ext3_max_size(1024) before the filesystem's
blocksize has been determined.

From: Hugh Dickins <hugh@veritas.com>

Stop vm_enough_memory double counting total_swapcache_pages: it dates
from the days when we didn't free swap when freeing swapcache page.

From: Hugh Dickins <hugh@veritas.com>

move_from_swap_cache and add_to_page_cache_lru are using GFP_ATOMIC,
which can easily fail in an intermittent way.  Rude if shmem_getpage
then fails with -ENOMEM: use blk_congestion_wait() to let kswapd in,
and repeat.

From: Hugh Dickins <hugh@veritas.com>

Recent testing has shown that swapoff can sneak a page back into the
tmpfs page cache after truncate_inode_pages has cleaned it, before
shmem_truncate resets next_index to stop that: BUG_ON(inode->i_blocks)
in shmem_delete_inode.  So call truncate_inode_pages again to be safe.

From: Hugh Dickins <hugh@veritas.com>

Make add_to_swap_cache static, it's only used by read_swap_cache_async;
and since that has just done a GFP_HIGHUSER allocation, surely it's
better for add_to_swap_cache to use GFP_KERNEL than GFP_ATOMIC.

From: Hugh Dickins <hugh@veritas.com>

Raised #endif CONFIG_SWAP in shrink_list, it was excluding
try_to_unmap of file pages.  Suspect !CONFIG_MMU relied on
that to suppress try_to_unmap, added SWAP_FAIL stub for it.

From: Hugh Dickins <hugh@veritas.com>

Delete unused SWAP_ERROR and non-existent page_over_rsslimit().

From: george anzinger <george@mvista.com>

Timer expirations are being reported twice.

install_page() needs to run flush_cache_page() prior to overwriting an
already-established pte.

From: Manfred Spraul <manfred@colorfullife.com>

Brian spotted a stupid bug in the slab initialization:

If multiple objects fit into one cacheline, then the allocator ignores
SLAB_HWCACHE_ALIGN and squeezes the objects into the same cacheline.  The
implementation contains an off by one error and thus doesn't work correctly:
For Athlon optimized kernels, the 32-byte slab uses 64 byte of memory.

From: Eric Piel <Eric.Piel@Bull.Net>

Fixes some long/int confusion on 64-bit machines which was causing failures
on ia64 - we end up trying to set bits in the 32-63 range on an int and the
kernel locks up.

Also cleans up idr.h.

George has acked this change.

The patch is designed to help locate where the kernel is dying during the
startup sequence.

- Boot parameter "initcall_debug" causes the kernel to print out the
  address of each initcall before calling it.

  The kallsyms tables do not cover __init sections, so printing the
  symbolic version of these symbols doesn't work.  They need to be looked up
  in System.map.

- Detect whether an initcall returns with interrupts disabled or with a
  locking imbalance.  If it does, complain and then try to fix it up.

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

Don't allow "bind_request" to be called before "register_pccard_driver".

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

This fixes ipmi compilation (with some documentation updates and another
minor fix, too).

It also fixes a problem with the state machine getting stuck in a
certain error condition.

into penguin.transmeta.com:/home/penguin/torvalds/repositories/kernel/linux