No reason to link the netfilter subdirs from net/ instead of
net/ipv[46], the result looks just cleaner ;)

In both preempt_schedule in sched.c and resume_kernel in entry.S, it is
possible to return with need_resched set and thus a pending preemption
but not service that preemption for some time.

Consider:

	- return from schedule() to preempt_schedule
	- interrupt occurs, sets need_resched
	- we cannot preempt since preempt_count = PREEMPT_ACTIVE
	- back in preempt_schedule, set preempt_count = 0

Now we again can preempt, but we will not.  Instead we return and
continue executing.  On the next interrupt, we will redo the whole
fiasco which is a waste since we could of reentered schedule while we
were there.  Worse, if we acquire a lock before the next interrupt we
can potentially delay the pending reschedule a very long time.  This is
not acceptable.

The solution is to check for and loop on need_resched on resume_kernel
and preempt_schedule like schedule itself does.

Two big ouches in x86 entry.S:

(1) Up until 2.5.20, we were not properly reading the irq_stat
    and bh_count values from the right CPU.  Brian Gerst sent you
    a patch to fix this.  This raises the question: why was this not
    a problem?  Seems we do not need this check at all as having a
    nonzero irq_star or bh_count implies having a nonzero preempt_count,
    which we test for above.  Thus this patch removes those tests and
    the related defines.

(2) What if it is possible to preempt even with interrupts disabled?
    Turns out it is.  Consider if we take an exception (say, for a
    TLB miss) and enter resume_kernel and preempt.  Even if interrupts
    are disabled, an exception can occur and end up in resume_kernel.
    We need to check to make sure interrupts are not off, to ensure
    we are not coming off an unmasked exception.

Even with the added check from issue #2, we have less code after #1 so
we can walk away with a bugfix and an optimization here. ;)

Thanks to George Anzinger, with whom I actually had #2 bite me, and who
helped with these issues.

 - Realize that the only place where ata_do_taskfile gets used is ide-disk.c
   move it and its "friends' over there.

 - Unify the do_request method for disk devices. This saves quite a lot of code.

 - Make task_muin_intr and task_in_intr use the same busy status checks on
   entry.

 - Unfold get_command at the single only place where it's used.

 - Add missing __ata_end_request on kill_rq path.

 - Rename udma_tcq_taskfile() to udma_tcq_init to make the code look like to
   normal udma_init. Revert the logics of udma_init and it's
   implementations to mirror that of udma_tcq_init().

 - Fix a tiny bug in pmac_udma_init() where it was reporting the wrong value up
   on failure.

 - Revert the logics of udma_start(). It's called from udma_init context.
   Realize that it is always returning ide_started. Make it self and the
   implementations of it return void.

 - Implement the assertion that the lock is already held, if we run
   ata_status_poll.

 - The usual host chip code load from Bartomiej onierkiewicz.  Fortunately
   I'm able to check the sis5513 fixes on a life system.

   More serious stuff... for IDE 88:

	- implement ata_best_pio_mode() in ata-timing.c
	  (play safer than driver's config_chipset_for_pio())

	- replace config_chipset_for_pio() by ata_best_pio_mode()
	  in host chips drivers
	  (trivial for hpt34x.c, hpt366.c and serverworks.c,
	   non-trivial for cmd64x.c and sis5513.c)

	- set PIO also for (U)DMA modes in cmd64x_tune_chipset()
	  (cmd64x.c)

	- fix bug in setting PIO0-2 for devices also supporting PIO3/4
	  (sis5513.c)

	- misc cleanups

 - Rename XXX_do_request back do do_request. This was just tmporary tagging.

 - Move ata_taskfile() call down to the ide-disk request handler.c, which is the
   only place needing it. Rename ata_taskfile to ata_do_taskfile().  This
   allowed us to implement the assertion that ata_do_taskfile() will be only
   called with the channel lock already held.

 - fix lock initialization  for device tree handling. Well I would be lucky if
   it would be just a device tree like in Solaris.  But after studying the code
   which goes in right now it appears to develop quite fast in to a "just in
   case" and "ad-hoc" inferface mess we have already in /proc. Shit!

   Look for example at the initialization of a device struct. First we have a
   name to the directory encompassing the device and the immediately we fill in
   the georgeous name field - which is enterly unnecessary.

 - Pull locking out from the device type drivers do_request handlers.  This
   allowed us to remove the draddy unlocking on entry to start_request.

   Much more of host controller rigister acces on crutial code paths is
   now covered by the queue access lock.

Linus, betcha cannot find a more trivial (but correct) patch...

This showstopper is against 2.5.21 - please apply.

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

Todd R. Eigenschink <todd@tekinteractive.com> reported 2.4 swapoff
kernel BUG at filemap.c:122 to LKML 24 May.  Other problems on that
system may have contributed, but yes, despite __delete_from_swap_cache
doing ClearPageDirty before __remove_inode_page to avoid the BUG(),
a concurrent zap_pte_range might race to set_page_dirty.  So skip
that oops in PageSwapCache case.  Remove the prior ClearPageDirty?
maybe but not without deeper thought: let stay for now.

Recent testing has shown that BUG() check in try_to_unuse is unsafe.
delete_from_swap_cache does final swap_free just after removing page
from swap cache, and add_to_swap_cache does swap_duplicate just before
putting page into swap cache, therefore swapin_readahead may resurrect
a dying swap entry and assign a new page to it.  That's fine, there's
no need to change this ordering; but it does mean that try_to_unuse's
page may have left the swap cache yet its swap_map count still be set.
That BUG() has done good service for swapoff sanity, but now abandon it.

Burton Windle <bwindle@fint.org> reported Kernel memory leak with
swapon/swapoff? LKML 31 May.  swapon uses rw_swap_page_nolock to read
swap_header page (peculiar! should probably rework that sometime),
nothing freed the buffers from the page, thus page also never freed.

Remove /* SMP-safe */ comments before shmem_truncate and shmem_mknod:
don't know who or why put there, but they seem to imply that the rest
of shmem.c is unsafe.

Several simple speedups to tmpfs swapoff: without patch, swapoff of a
kernel tree in tmpfs might take take 2 minutes, with patch 4 seconds.
Inline search go no further than necessary; only search inode when it
has swapped pages; start next search from same inode; list in order.

(There's a "list_move_tail" in this patch: not available in 2.5.21,
but I believe you now have it in your ongoing tree.)

shmem_mknod should not update directory times if it cannot get an inode.

shmem_symlink was letting a long symlink overwrite its vfs_inode, now
included within struct shmem_inode_info - oops! and failed long symlink
inodes were freed but still left on the shmem_inodes list, causing oops
in swapoff at shutdown (if not earlier).

shmem_rename was not maintaining the correct link count on the
parent directories when a directory was moved from one to another.
This patch from Christoph Rohland <cr@sap.com>, already in 2.5-dj.

 - Bunch of cleanups by Bartlomiej (accounts for over a half of the patch):

    cmd64x.c:
	- kill SPLIT_BYTE() macro
	- kill wrappers for cmd64x_config_drive_for_dma()
	- misc cleanups

    cy82c693.c:
	- kill obsolete comments
	- clean cy82c693_tune_drive() and calc_clk()
	- misc cleanups

    hpt34x.c:
	- kill obsolete comment
	- kill SPLIT_BYTE()
	- kill hpt34x_clear_chipset()
	- simplify hpt34x_tune_drive()

    hpt366.c:
	- kill hpt_min_rev()
	- kill redundant hpt368_tune_chipset() and hpt374_tune_chipset()
	- fix badlists checking in config_chipset_for_dma()
	- misc cleanups

    pdc202xx.c:
	- clean registers decoding
	- clean pdc202xx_tune_chipset()
	- kill pdc202xx_udma_irq_status(), use generic udma_pci_irq_status()
	- rationalize pdc202xx_reset()
	- kill UDMA_SPEED_FLAG() and PDC_CLOCK() macros,
	  do it right by defining constants PDC_UDMA and PDC_CLK
	- kill init_high_16() inline, no need to hide internals
	- clean pdc202xx_init_chipset()
	- split ata66_pdc202xx() and pdc202xx_init_chipset()
	- clean config_chipset_for_dma()
	- misc cleanups

 - Fix plug in of CF cards. The previously used sub device driver attach method
   lookup was entirely hosed.

 - Enforce indentation style on ide-cs.c. Enable debugging there. (Makes the
   patch quite big...)

This makes the IO bitmap allocations a separately allocated structure,
shrinking the default task size.

We alloc it in sys_ioperm() and copy_thread() and frees in
exit_thread().  It also gets rid of the IO_BITMAP_SIZE+1 crap, as only
the tss actually needs the tail long, and we weren't copying it into the
bitmap anyways.

The utime and utimes should do exactly the smae permission check
according to SUSv3.
	"The effective user ID of the process shall match the owner of the file,
or has write access to the file or appropriate privileges to use this call
in this manner."

utimes when passed a NULL second argument would fail on a read only
file even if the file is owned by the caller.

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

The following comes from James Bottemley, as he updated the SCSI layer
to use the generic tagging.

Basically just put blk_queue_tag_request() into ll_rw_blk.c and name it
something more sane, blk_queue_find_tag(). In addition, remove REQ_CMD
requirement in blk_queue_start_tag() -- this is actually IDE specific
(we can only tag read/write requests there...), SCSI actually requires
everything to be tagged once enabled. It's replaced with a safety check
for an already tagged request.

This is a merge of the two things that are needed for the plugging as it
stands in 2.5.21. The first is fixing the locking to be both clearer and
safe (Andrew repeatedly broke the old version). The second is a few
changes that allow make_request_fn drivers to utilize plugging. This is
needed for umem and raid, for instance, that have their private
plugging.

- i've extended the scheduler context-switch mechanism with the following
  per-arch defines:

        prepare_arch_schedule(prev_task);
        finish_arch_schedule(prev_task);
        prepare_arch_switch(rq);
        finish_arch_switch(rq);

- plus switch_to() takes 3 parameters again:

        switch_to(prev, next, last);

- schedule_tail() has the 'prev' task parameter again, it must be passed
  over in switch_to() and passed in to the fork() startup path.

architectures that need to unlock the runqueue before doing the switch can
do the following:

 #define prepare_arch_schedule(prev)             task_lock(prev)
 #define finish_arch_schedule(prev)              task_unlock(prev)
 #define prepare_arch_switch(rq)                 spin_unlock(&(rq)->lock)
 #define finish_arch_switch(rq)                  __sti()

this way the task-lock makes sure that a task is not scheduled on some
other CPU before the switch-out finishes, but the runqueue lock is
dropped. (Local interrupts are kept disabled in this variant, just to
exclude things like TLB flushes - if that matters.)

architectures that can hold the runqueue lock during context-switch can do
the following simplification:

 #define prepare_arch_schedule(prev)             do { } while(0)
 #define finish_arch_schedule(prev)              do { } while(0)
 #define prepare_arch_switch(rq)                 do { } while(0)
 #define finish_arch_switch(rq)                  spin_unlock_irq(&(rq)->lock)

further optimizations possible in the 'simple' variant:

- an architecture does not have to handle the 'last' parameter in
  switch_to() if the 'prev' parameter is unused in finish_arch_schedule().
  This way the inlined return value of context_switch() too gets optimized
  away at compile-time.

- an architecture does not have to pass the 'prev' pointer to
  schedule_tail(), if the 'prev' parameter is unused in
  finish_arch_schedule().

the x86 architecture makes use of these optimizations.

Via this solution we have a reasonably flexible context-switch setup which
falls back to the current (faster) code on x86, but on other platforms the
runqueue lock can be dropped before doing the context-switch as well.

        Ingo

NOTE: i have coded and tested the 'complex' variant on x86 as well to make
      sure it works for you on Sparc64 - but since x86's switch_mm() is
      not too subtle it can use the simpler variant. [ The following
      things had to be done to make x86 arch use the complex variant: the
      4 complex macros have to be used in system.h, entry.S has to
      'pushl %ebx' and 'addl $4, %esp' around the call to schedule_tail(),
      and switch_to() had to be reverted to the 3-parameter variant
      present in the 2.4 kernels.

NOTE2: prepare_to_switch() functionality has been moved into the
       prepare_arch_switch() macro.

NOTE3: please use macros for prepare|finish_arch_switch() so that we can
       keep the scheduler data structures internal to sched.c.

Silly oversight - read_pages needs to pass the file *
down to ->readpage().

- If we're PF_MEMALLOC and BIO allocation failed, fall back to try
  to allocate a smaller BIO (from a different mempool)

- When writepages() gets confused, call a_ops->writepage()
  instead of going directly to block_write_full_page().  It
  gives the fs more flexibility, and XFS may want this.

It will need swapops.h for swp_entry().

Fix the order of inodes being marked dirty in a couple of corner cases.

The only impact of this bug is that the on-disk copies of i_version
might got out of sync for a directory, or that an error inserting an
inode into a directory might leave its i_nlinks incorrect on disk for a
short interval.  Neither problem will cause trouble for ext3 during
normal operation, but the nlink problem might cause e2fsck to emit
unnecessary warnings if we crash while the incorrect version of the
inode is in the journal.

Removes the put_iode optimisation.  It's racy, as
Chris pointed out.

ext3 will remount the fs readonly if it runs out of inode
space.  It shouldn't do that - it should just return -ENOSPC.

Cset exclude: bcrl@redhat.com|ChangeSet|20020429021546|12619

This allows compilation without ACPI_DEBUG defined.

This adds an export so that processor.o and thermal.o can be compiled as
modules

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

into tp1.ruhr-uni-bochum.de:/home/kai/kernel/v2.5/linux-2.5.make

A couple of targets were still in the main part of the Makefile,
even though they did not need .config to exist for execution.

Move them into the noconfig section of the Makefile, allowing them
to be executed before make *config.

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