Same thing as patch #1165/1 but for 2.5.x

The kernel expects these to return values in unsigned long types.

- remove LAST_DESC in favour of passing the array size to iotable_init
- replace domain and permission bits/cache attributes with a generic
  "type", which can be easily converted to the required domain and
  permission bits/cache attributes at run time.  It also removes the
  possibility for getting such things wrong and (accidentally)
  allowing all user space to fiddle with devices directly.

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

explicitly.

 - Remove DEVICE_INTR and associated code from floppy driver.

 - Savlage s390 xpram code from kernel version dependant compilation disease.

 - Eliminate SET_INTR code from the places where it was used.

 - Eliminate bogous support for multiple sbpcd controllers. The driver didn't
   even compile right now before we could think about further supporting it at
   all we have to get rid of this hack first.
   Don't call invalidate_buffers in the release method there.
   Why should it be necessary?

 - Resurrect sonycd535 compilation.

 - Let CURRENT request macro use the same primitive at the remaining QUEUE macro
   in blk.h, which is still not quite right, but first things first :-).

Makes minixfs, sysvfs and ufs understand `mount -o dirsync'.

Fixes a pet peeve: the identifier "flushpage" implies "flush the page
to disk".  Which is very much not what the flushpage functions actually
do.

The patch renames block_flushpage and the flushpage
address_space_operation to "invalidatepage".

It also fixes a buglet in invalidate_this_page2(), which was calling
block_flushpage() directly - it needs to call do_flushpage() (now
do_invalidatepage()) so that the filesystem's ->flushpage (now
->invalidatepage) a_op gets a chance to relinquish any interest which
it has in the page's buffers.

A patch from Hugh Dickins which fixes a couple of error-path leaks
related to tmpfs (I think).

Also fixes a yield()-inside-spinlock bug.

It also includes code to clear the final page outside i_size on
truncate.  tmpfs should be returning zeroes when a truncated file is
later expanded and it currently is not.

Hugh is taking care of the 2.4 fix for this.

Replaces SetPageDirty() with set_page_dirty() in several places related
to in-memory filesystems.

SetPageDirty() is basically always the wrong thing to do.  Pages should
be moved to the ->dirty_pages list when dirtied so that writeback can
see them.

Without this change, dirty pages against in-memory filesystems would
churn around on the inactive list all the time, rather than getting
pushed away onto the active list.  A minor efficiency thing.

Fixes a race between unlink and writeback: on the sys_sync() and
pdflush paths the caller does not have a reference against the inode.

So run __iget prior to dropping inode_lock.

Oleg Drokin reported this and seems to believe that it fixes the
crashes he was observing.  But I was never able to reproduce them..

Fixes a few lock ranking bugs (and deadlocks) related to
swap_list_lock(), swap_device_lock(), mapping->page_lock and
mapping->private_lock.

- Cannot call block_flushpage->try_to_free_buffers() inside
  mapping->page_lock.  Because __set_page_dirty_buffers() takes
  ->page_lock inside ->private-lock.

- Cannot call swap_free->swap_list_lock/swap_device_lock inside
  mapping->page_lock because exclusive_swap_page() takes ->page_lock
  inside swap_info_get().


The patch also removes all the block_flushpage() calls from the swap
code in favour of a direct call to try_to_free_buffers().

The theory is that the page is locked, there is no I/O underway, nobody
else has access to the buffers so they MUST be freeable.  A bunch of
BUG() checks have been added, and unless someone manages to trigger
one, the "block_flushpage() inside spinlock" problem is fixed.

Give swapper_space a ->set_page_dirty() address_space_operation.

So swapcache pages do not need special-casing in
set_page_dirty_buffers().

Turn on direct-to-BIO writeback for ext3 in data=writeback mode.

block_symlink() is not a "block" function at all.  It is a pure
pagecache/address_space function.  Seeing driverfs calling it was
the last straw.

The patch renames it to `page_symlink()' and moves it into fs/namei.c

Remove i_wait from struct inode and hash it instead.

This is a pure space-saving exercise - 12 bytes from struct
inode on x86.

NFS was using i_wait for its own purposes.  Add a wait_queue_head_t to
the fs-private inode for that.  This change has been acked by Trond.

Implement buffer_boundary() for ext3.

buffer_boundary() is an I/O scheduling hint which the filesystem's
get_block() function passes up to the BIO assembly code.  It is
described in fs/mpage.c

The time to read 1,000 52 kbyte files goes from 8.6 seconds down to 2.9
seconds.  52 kbytes is the worst-case size.

Speeds up generic_file_write() by not calling mark_inode_dirty() when
the mtime and ctime didn't change.

There may be concerns over the fact that this restricts mtime and ctime
updates to one-second resolution.  But the interface doesn't support
that anyway - all the filesystem knows is that its dirty_inode()
superop was called.  It doesn't know why.

So filesystems which support high-resolution timestamps already need to
make their own arrangements.  We need an update_mtime i_op to support
those properly.

time to write a one megabyte file one-byte-at-a-time:

Before:
	ext3:		24.8 seconds
	ext2:		 4.9 seconds
	reiserfs:	17.0 seconds
After:
	ext3:		22.5 seconds
	ext2:		4.8  seconds
	reiserfs:	11.6 seconds

Not much improvement because we're also calling expensive
mark_inode_dirty() functions when i_size is expanded.  So compare the
overwrite case:

time dd if=/dev/zero of=foo bs=1 count=1M conv=notrunc

ext3 before:	20.0 seconds
ext3 after:	9.7  seconds

First some terminology: this patch introduces a kernel-wide `pgoff_t'
type.  It is the index of a page into the pagecache.  The thing at
page->index.  For most mappings it is also the offset of the page into
that mapping.  This type has a very distinct function in the kernel and
it needs a name.  I don't have any particular plans to go and migrate
everything so we can support 64-bit pagecache indices on x86, but this
would be the way to do it.

This patch improves the packing density of swapcache pages in the radix
tree.

A swapcache page is identified by the `swap type' (indexes the swap
device) and the `offset' (into that swap device).  These two numbers
are encoded into a `swp_entry_t' machine word in arch-specific code
because the resulting number is placed into pagetables in a form which
will generate a fault.

The kernel also need to generate a pgoff_t for that page to index it
into the swapper_space radix tree.  That pgoff_t is usually
bitwise-identical to the swp_entry_t.  That worked OK when the
pagecache was using a hash.  But with a radix tree, it produces
catastrophically bad results.

x86 (and many other architectures) place the `type' field into the
low-order bits of the swp_entry_t.  So *all* swapcache pages are
basically identical in the eight low-order bits.  This produces a very
sparse radix tree for swapcache.  I'm observing packing densities of 1%
to 2%: so the typical 128-slot radix tree node has only one or two
pages in it.

The end result is that the kernel needs to allocate approximately one
new radix-tree node for each page which is added to the swapcache.  So
no wonder we're having radix-tree node exhaustion during swapout!
(It's actually quite encouraging that the kernel works as well as it
does).

The patch changes the encoding of the swp_entry_t so that its
most-significant bits contain the `type' field and the
least-significant bits contain the `offset' field, right-aligned.

That is: the encoding in swp_entry_t is now arch-independent.  The new
file <linux/swapops.h> has conversion functions which convert the
swp_entry_t to and from its machine pte representation.

Packing density in the swapper_space mapping goes up to around 90%
(observed) and the kernel is tons happier under swap load.


An alternative approach would be to create new conversion functions
which convert an arch-specific swp_entry_t to and from a pgoff_t.  I
tried that.  It worked, but I liked it less.

Remove some unused PageSkip() macros.  Presumably leftovers from
PG_skip which isn't there any more.

A common and very subtle bug is to use list_heads which aren't on any
lists.  It causes kernel memory corruption which is observed long after
the offending code has executed.

The patch nulls out the dangling pointers so we get a nice oops at the
site of the buggy code.

I missed this one in the last patch I sent to you.

Too much copy'n paste between 2.4 and 2.5 code base, attached patch on
top of the previous block tag fixes makes it work/compile again. Sorry
about that.

A buglet and a few adjustments.

This should be the last of tq_disk, at least the trivial ones. md still
has some queue_task references, I'll let Ingo/Neil clean those up.
suspend is still broken, it was broken before too though. I guess Pavel
will want to fix that.

Also, I've documented the plug functions.

 - PPC compilation fix by Paul Mackerras.

 - Various fixes by Bartek:

   fix ata_irq_enable() and ata_reset() for legacy ATA-1 devices

   in start_request() for REQ_DRIVE_ACB
   a) don't run ->prehandler() twice
   b) return ata_taskfile() value

 - Don't use ata_taskfiles cmd field for drive status reporting,
   we can now simply use drive->status instead.

 - Unify command type parser entries which could be unified due to the
   unification of corresponding interrupt handlers.

 - Eliminate reading parameter from ata_do_udma(). We have this information
   already in the rq. This allows us to merge several methods.

 - Rename XXX_udma to udma_setup, since we have finally settled up on this
   semantics.

 - Simplify tons of host chip code by removing wrapper functions.

 - Sanitize the handling of the ioctl's and fix a bug on the way in dealing with
   the WIN_SMART command where arguments where exchanged.

 - Finally sanitize ioctl further until it turned out that we could get rid of
   the special request type REQ_DRIVE_CMD entierly. We are now using
   consistently REQ_DRIVE_ACB.

   One hidden code path less again!

 - Realize the ide_end_drive_cmd can be on the REQ_DRIVE_ACB only for ioctl() to
   a disk. Eliminate it's usage from device type driver modules.

 - Remove command member from struct  hd_drive_task_hdr and place it in strcut
   ata_taskfile. It is not common between the normal register file and HOB.

   We will have to introduce some helper functions for particular command types.

 - Fix typo in sparc_v9 code, in ns87415, just introduced.

 - Eliminate unnecessary struct hd_drive_hob_hdr those are
   in reality precisely the same registers as usual.

 - Eliminate control_t, nowhere used type.

 - Unfold ide_init_drive_cmd() at the places where it's used. This makes obvious
   that REQ_DRIVE_CMD gets only used on the ioctl command path.

 - Move ide_fixstring() from ide.c to probe.c, since this is the place, where it's
   most used.

 - Remove GET_STAT() - it's not used any longer.

 - Remove last parameter of ide_error. Rename it to ata_error().

 - Don't use ide_fixstring in qd65xx.c host chip driver. The model name is
   already fixed in probe.c.

 - Invent ata_irq_enable() for the handling of the trice nIEN bit of the
   control register.  Consistently use ch->intrproc method every time we toggle
   this bit.  This simply wasn't the case before!

 - Disable interrupts on a previous channel only when we share them indeed.

 - Eliminate simple drive command handling function drive_cmd.

 - Simplify the ioctl handler. Move it to ioctl, since that's the only place
   where it's actually used.

as it isn't even all that big.

  David Mosberger added argument page to clear_user_page, but
apparently did not scan whole tree to find callers. Please apply.
It's the only such call in the non-arch specific portion of tree.

The main Makefile includes .config - on the other hand, it also
generates .config when using make *config. This leads to recursion
problems when we make the build depend on .config.

(.config is generated by *config. So after running make *config,
 make notices one of its includes has been updated -> it restarts
 to read the new contents. Unfortunately, the restart runs *config
 again, which updates .config again. You get the picture)

Therefore, we're aiming at a two phase system: If no .config exists,
allow only make *config and make clean/mrproper. Afterwards, allow
all targets. (It's not much different from what we have today)

-DMODVERSIONS isn't used anymore, so it can go. Also, after cleaning
up include/linux/module.h, it's obvious that we don't need
include/linux/modversions.h at all if CONFIG_MODVERSIONS not set,
no need to generate it.

Rules.make explicitly lists files which depend on modversions.h,
since make dep cannot know about the
"-include include/linux/modversions.h" which gets added to
the command line. Now that we understand when it is needed, we can
even get that list right ;-)

Oh well, nice theory. .hdepend will touch module.h when modversions.h
changes, so we still get unnecessary recompiles. We really need to
switch to the new way of dependency generation, it gets all that
right without even thinking about these special cases.

We don't track dependencies for .ver files. In fact, we relied on that
checksum would only change if the corresponding exporting C source
changes. That's not true, of course, all of the included headers have
say as well. So we better force the hash to be checked
unconditionally every time "make dep" is run.

Observe that 
defined(MODVERSIONS) == defined(CONFIG_MODVERSIONS) && defined(MODULE)
and from there I step by step simplified the logic in
include/linux/module.h - staying logically equivalent, but it is much 
more understandable now, IMO.

Still added a huge comment trying to help other people understand
what kind of magic happens here.