Here's a patch which converts O_DIRECT to go direct-to-BIO, bypassing
the kiovec layer.  It's followed by a patch which converts the raw
driver to use the O_DIRECT engine.

CPU utilisation is about the same as the kiovec-based implementation.
Read and write bandwidth are the same too, for 128k chunks.   But with
one megabyte chunks, this implementation is 20% faster at writing.

I assume this is because the kiobuf-based implementation has to stop
and wait for each 128k chunk, whereas this code streams the entire
request, regardless of its size.

This is with a single (oldish) scsi disk on aic7xxx.  I'd expect the
margin to widen on higher-end hardware which likes to have more
requests in flight.

Question is: what do we want to do with this sucker?  These are the
remaining users of kiovecs:

	drivers/md/lvm-snap.c
	drivers/media/video/video-buf.c
	drivers/mtd/devices/blkmtd.c
	drivers/scsi/sg.c

the video and mtd drivers seems to be fairly easy to de-kiobufize.
I'm aware of one proprietary driver which uses kiobufs.  XFS uses
kiobufs a little bit - just to map the pages.

So with a bit of effort and maintainer-irritation, we can extract
the kiobuf layer from the kernel.

Synchronize with 2.5.25.

Incorporate IDE-94, as well as 95, 96, 97 and 98-pre as announced by Bartek and
unfortunately still not included in 2.5.25, which makes admittedly things
still fall appart:

Missing changelog for 98-pre by Bartlomiej Zolnierkiewicz (BTW.  Handling
Unicode should be essential at least to make proper crediting of many many
peoples possible!) follows here:

 - add missing channel->lock unlocking/locking and fix some comments
   in ide_timer_expiry()

 - allow PCI drivers to disable autodma in ->init_dma()
   (bug introduced in IDE 97, affects sl82c105.c only)

   noticed by Russell King

 - alim15x3.c, if revision is <= 0x20 disable autodma

 - remove unneeded checks (drive.dn > 3) from pdc202xx.c and sis5513.c

 - use block layer wrappers

And my additions follow:

 - Fix TCQ code. Patch based on work by Alexander Atanasov.

 - Use the FreeBSD derived request handler return values:

	ATA_OP_FINISHED
	ATA_OP_CONTINUES
	ATA_OP_RELEASED
	ATA_OP_READY	/* for status ready reporting during poll */

 - PMAC compilation fix by Paul Mackerras.

 - Simplify the ata_status_poll function significantly.

 - Fix logic used to prevent drive IRQ assertion from drive on channels sharing
   our interrupt.

NOTE: We will move it later to the time where a request is really finished
soon.

 - Don't use ata_busy_poll() use ata_status_poll() instead. This increases code
   unification.

NOTE: We should maybe invent some way to prevent the error recovery path to be
taken at all. In esp to prevent ata_error from trying to reissue commands.

"getpid()" returns "tgid" not "pid", so this is correct.

Fix from Saurabh Desai.

Eeek....

futex.c is seriously b0rken.

This is against 2.5.25. Lots of fixes. Brings things inline with 2.5.25.
Probably makes the subsystem actually work now (well, it will work, but
I bet without it, it doesn't). Merged in changes from the current 2.5.x
source.

about the "sendpage()" callback. Don't try to fall back on
a write with the page kmap'ed

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

- Move the directory index bitmap to use an attribute inode instead of
  having special fields for it inside the ntfs inode structure. This
  means that the index bitmaps now use the page cache for i/o, too,
  and also as a side effect we get support for non-resident index
  bitmaps for free.
- Simplify/cleanup error handling in fs/ntfs/dir.c::ntfs_readdir() and
  fix a page leak that manifested itself in some cases.
- Add fs/ntfs/inode.c::ntfs_put_inode(), which we need to release the
  index bitmap inode on the final iput().

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

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

- Apply block optimization added to fs/ntfs/aops.c::ntfs_read_block()
  to fs/ntfs/compress.c::ntfs_file_read_compressed_block() as well.
- Drop the "file" from ntfs_file_read_compressed_block().
- Rename fs/ntfs/aops.c::ntfs_enb_buffer_read_async() to
  ntfs_end_buffer_async_read() (more like the fs/buffer.c counterpart).
- Update ntfs_end_buffer_async_read() with the improved logic from
  its updated counterpart fs/buffer.c::end_buffer_async_read(). Apply
  further logic improvements to better determine when we set PageError.
- Update submission of buffers in fs/ntfs/aops.c::ntfs_read_block() to
  check for the buffers being uptodate first in line with the updated
  fs/buffer.c::block_read_full_page(). This plugs a small race
  condition.

- Define L1_CACHE_SHIFT for quota sake
- Fix FC4/PLUTO build, request is a pointer in scsi_cmnd now.
- Make binfmt_elf32 build properly wrt. jiffies_to_foo changes.

This one-liner does exactly $SUBJ.

Attached is a BK patch to add a new driver for X-Box gamepads. It's
fairly simple, and I think it can be put into the 2.5 tree.

> Hm, in building this, it looks like pid.c wants to be a separate module,
> yet the Makefile tries to build it into the hid.o module.  The linker
> complains horribly when this happens :)

Ok, I found the case. Patch attached.

This cset is update of the HID drivers to the latest version, as a part
of the Input merge. It finally includes ForceFeedback support by Johann
Deneux, enabling ForceFeedback on new Logitech and Microsoft devices.

- There was a narrow window between checking a buffer head for being
uptodate and locking it in ntfs_file_read_compressed_block(). We now
lock the buffer and then check whether it is uptodate or not.

- Modify fs/ntfs/inode.c::ntfs_read_locked_inode() to return an error
  code and update callers, i.e. ntfs_iget(), to pass that error code
  up instead of just using -EIO.
- Modifications to super.c to ensure that both mount and remount
  cannot set any write related options when the driver is compiled
  read-only.
- Optimize block resolution in fs/ntfs/aops.c::ntfs_read_block() to
  cache the current run list element. This should improve performance
  when reading very large and/or very fragmented data.

- Fix missed cli()/sti() conversions.
- Fix SA1100 sleep code.
- Update small_page.c for changes to struct page.
- Clean up assabet/neponset initialisation.
- Clean up compiler warnings in iop310 build.
- Add missed bits from previous xscale cset.

- Add sa1111_{dis,en}able_device()
- Add sa1111_{read,write}l() (always host endian)

- FSR "write" bit moved from bit 8 to bit 11.
- Handle bit 10 of FSR for xscale imprecise aborts.
- Allow Xscale CP0 and CP13 accesses.
- Move Xscale specific implementations to their own file.

The major points are:
- Move ptrace code from traps.c into ptrace.c so it's all in one place.
- Correct get_user_regs() to allow for the 8 byte offset.
- Fix missing ptrace_cancel_bpt() call which could miss some
  single stepping steps.
(other small pages; for details see individual file notes)

Some of the CPU implementations copied the ELF flags from the
StrongARM-110 implementation, without thinking whether they were
correct or not.  This changeset fixes up these CPU types.

- Fix ARM Makefile; we now build with -Os instead of -O2.
- Fix backtrace code to display more registers correctly.
- Fix various build errors.
- Fix PSR for architecture compliance.
- Update mach-types file.

Rather than handling the IRQ state each time we emulate each
instruction, we enable the IRQ before we enter the FP emulator.
We don't require any IRQ protection in the emulator anyway, so
it's safe to implement this trivial performance enhancement.