input.h wants fs.h and timer.h, not sched.h

atmdev.h only wants wait.h, not sched.h

blkdev.h simply doesn't need sched.h

The inode reclaim is too aggressive at present - it is causing the
shootdown of lots of recently-used pagecache.  Simple testcase: run a
huge `dd' while running a concurrent `watch -n1 cat /proc/meminfo'.
The program text for `cat' gets loaded from disk once per second.

This is in fact because the dentry_unused reclaim is too aggressive.

(The general approach to inode reclaim is that it _not_ happen at the
inode level.  All the aging and lru activity happens at the dcache
level.)

The problem is partly due to a bug: shrink_dcache_memory() is returning
the *total* number of dentries to the VM, rather than the number of
unused dentries.

This patch fixes that, and goes a little further.

We do want to keep some unused dentries around.  Reclaiming the last
few thousand dentries is pretty pointless, and will allow reclaim of
the last few thousand inodes and their attached pagecache.

So the algorithm I have used is to not allow the number of unused
dentries to fall below the number of used ones.  This keeps a
reasonable number of dentries in cache while providing a level of
scaling to the system size and the current system activity.

(Magic number alert: why not pin nr_unused to seven times nr_used,
rather than one times??)

shrink_dcache_memory() has been changed to tell the VM that the number
of shrinkable dentries is:

	zero if (nr_unused < nr_used)
	otherwise (nr_unused - nr_used)

so when there is memory pressure the VM will prune the unused dentry
cache down to the size of the used dentry cache, but not below that.

The patch also arranges (awkwardly) for all modifications of
dentry_stat.nr_dentry to occur inside dcache_lock - it was racy.

Replace some kmaps in mpage.c with kmap_atomic.

Stephen Tweedie reports a 2.4.7 problem in which kswapd is chewing lots
of CPU trying to reclaim inodes which are pinned by buffer_heads at
i_dirty_buffers.

This can only happen when there's memory pressure on ZONE_HIGHMEM - the
2.4 kernel runs shrink_icache_memory in that case as well.  But there's
no reclaim pressure on ZONE_NORMAL so the VM is never running
try_to_free_buffers() against the ZONE_NORMAL buffers which are pinning
the inodes.

The 2.5 kernel also runs the slab shrinkers in response to ZONE_HIGHMEM
pressure.  This may be wrong - still thinking about that.

This patch arranges for prune_icache to try to remove the inode's buffers
when the inode is to be reclaimed.

It also changes inode_has_buffers() and the other inode-buffer-list
functions to look at inode->i_data, not inode->i_mapping.  The latter
was wrong.

slab does various consistency checks during `cat /proc/slabinfo'
processing.  If it finds one it stupidly goes BUG just before
displaying the info which is required to diagnose the bug.

Change it to not go BUG, but to emit some useful printks and continue.

The patch also removes an uninteresting printk from the boot process.

Patch from Rohit Seth, changelog from Bill Irwin:

(1) fixes failure to clear key->busy (yes, it's always under a lock)
(2) fixes key->count being unconditionally set to 1 in alloc_key()
(3) reduces search to key->size >> HPAGE_SHIFT from key->size
(4) actually uses vma->vm_private_data to release the key as intended

plus the cleanup:
(5) removes the int *new argument to alloc_key()

A rollup of Bill's 11-patch series which replaces hugetlb's custom
inode with a bare radix tree.  Reviewed and acked by Rohit.

- revert doublefreeing patch

- Put set_new_inode() and the inode allocation loop into an
  alloc_key(), and introduce a new opaque type "struct hugetlb_key".

- Wrap the release path in alloc_shared_hugetlb_pages() with a
  release_key() function that handles the release.

- Wrap hugetlb_prefault() with prefault_key() in order to isolate the
  dependency on inodes for prefaulting the hugetlb vma.

- Replaces the usage of inode->i_writecount as a flag marking keys
  busy with a boolean flag field in struct htlbpagekey, and removes the
  last dependency of alloc_shared_hugetlb_pages() on struct inode.

- Remove the last direct usage of struct inode within the hugetlb
  functions.

- Removes many direct usages of struct inode within the key
  manipulation API in exchange for passing references to the key
  structure itself.

- Expand out prefault_key() into its hugetlb_prefault() component,
  but substitute stubs to abstract out inode access.

- Move uid/gid/mode/size fields used in struct inode and the checks
  on them into key management code and structures.

- Substitute direct usage of radix trees for inodes, and removes the
  custom-allocated inodes.

- Wrap up the release path by adding proper refcounting of keys.

- add init_timer in bttv driver

- remove duplicated init_timer() in ncpfs.

- remove noisy printk's from 3c59x.c

- sparc64 compile fix with CONFIG_HUGETLBPAGE=y - htlbpage_max is now an
  int

Patch from Hugh Dickins <hugh@veritas.com>

tmpfs failed fsx+swapout tests after many hours, a page found zeroed.
Not a truncate problem, but mirror image of earlier truncate problems:
swap goes through mpage_writepages, which must therefore allow for a
sudden swizzle back to file identity.

Second time this caught us, so I've audited the tree for other places
which might be surprised by such swizzling.  The only others I found
were (perhaps) in the parisc and sparc64 flush_dcache_page called
from do_generic_mapping_read on a looped tmpfs file which is also
mmapped; but that's a very marginal case, I wanted to understand it
better before making any edit, and now realize that hch's sendfile
in loop eliminates it (now go through do_shmem_file_read instead:
similar but crucially this locks the page when raising its count,
which is enough to keep vmscan from interfering).

Patch from Hugh Dickins <hugh@veritas.com>

swap_readpage and swap_writepage forgot
to unlock_page if get_swap_bio failed.

Patch from Hugh Dickins <hugh@veritas.com>

On some architectures (cachetlb.txt gives HyperSparc as an example)
it is essential to flush_cache_page while pte is still valid: the
rmap VM diverged from the base 2.4 VM before that fix was made,
so this error has crept back into 2.5.

Patch below applies to 2.5.47 or 2.5.47-mm1 - needs more work over
shared pagetables, but they've silently fallen out of 2.5.47-mm1:
oversight?  I'll send Alan the equivalent for 2.4-ac shortly.

(I wonder, what happens if userspace now modifies the page
after the flush_cache_page, before the pte is invalidated?)

Patch from Andreas Gruenbacher <agruen@suse.de>

Add a gfp_mask parameter to the free() callback so that the callback can
safely do I/O, etc. The free callback can now also fail.  This will be
needed by reiserfs.

The order of entries on the cache entry lru is reversed so that
list_for_each_safe() can be used. Several helper functions that don't
make the code any better are removed. Finally, a couple of cosmetic
things.

From Badari.

bio_add_page returns zero on failure - we need to propagate that
to the dio_bio_add_page() caller.

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

This patch removes all the wait_queue handling code from sched.h and puts
it in wait.h with the rest of the wait_queue handling code.  Note that
sched.h must continue to include wait.h for the wait_queue_head_t embedded
in struct task.  However there may be files which only need wait.h now.

It really makes no sense to have request_irq and free_irq in sched.h
Let's move them to interrupt.h instead.  Note that I also remove sched.h
from interrupt.h since it's not needed.

A minor removal of 6 function definitions from sched.h.  They clearly
fit better in file.h.  All users of these functions already include file.h.
And none of them included sched.h directly...

This patch from William Lee Irwin III privatizes __d_path() to dcache.c,
uninlines d_path(), moves its declaration to dcache.h, moves it to
dcache.c, and exports d_path() instead of __d_path().

Grrr...  Two bugs in a patch that had moved md setup to late boot:

a) we need md_run_setup() run before parsing root name.
b) it's create_dev("/dev/md0",...), not create_dev("md0",...) ;-/

Still not safe (we use __MOD_INC_USE_COUNT in paride.c; old code has
MOD_INC_USE_COUNT in protocol drivers), but that takes crap in one
place.

	->owner added
	paride.c grabs/releases it if if present
	->proto_init() became empty for almost everything
	->proto_release() <<--->>
	->proto_init() returns int now (the only case where we do have a
	  non-empty ->proto_init() needed that all along).  paride.c
	  taught to deal with that.

dasd_proc.c : should be using ->owner instead of MOD_..._USE_COUNT in
->open()/->release().

s390/char/tape.c, s390/char/tapechar.c, usb/image/scanner.c,
intermezzo/psdev.c: ditto

intermezzo/super.c: they forgot to remove MOD_INC_USE_COUNT from the
..._fill_super()

binfmt_som.c: ->load_binary() and ->load_library() don't need
MOD_INC_USE_COUNT, since ->module is correctly set.

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

devfs_register_tape() returns the number it had assigned to tape.

new helper: devfs_unregister_tape(number) - removes symlink created by
devfs_register_tape()

devfs_register_tape() doesn't use devfs_auto_unregister() anymore.

devfs_register_tape() gets devfs entry of directory, instead of that of
a random file in that directory.

users updated

dm-ioctl.c does, er, interesting things to figure out the name of devfs
node it had just created.  Cleaned up.

dv1394.c piles amazing amounts of crap around its devfs entries.
Probably a result of times before devfs_find_and_unregister()...

In any case, code switched to use of devfs_find_and_unregister(),
crapectomy performed...

- Improved file cleanup code

- A more uniform poll queueing interface with tips from Manfred

- The f_op->poll() is done outside the irqlock to maintain compatibility
	with existing drivers that assume to be called with irq enabled

- Moved event mask setting inside ep_modify() with tips from John

- Fixed locking to fit the new "poll() outside the lock" approach

- Bufferd userspace event delivery to reduce irq_lock/irq_unlock switching
	rate and to reduce the number of __copy_to_user()

- Comments added

There's a long standing bug in blk_recount_segments(). Clustering means
physical segment colascing, not hardware segment colascing. This
basically means that we are mapping more segments than here than the bio
+ requests contains, and this causes a bug in the SCSI layer for host
adapters that have CLUSTERING enabled.

This patch makes sure that we are clustering physical segments
correctly, and correctly accounting hardware segments. Please apply.

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

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

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

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

instead of home-brew PCI allocations.