Ok, I found it.  The problem was in the decompression code.

In inflate.c, in the routine zlib_inflate(), it reuses the passed-in "f"
parameter as a return value, but the PPP packet compression stuff needs
that return value later.

This patch adds two intermediate general cache sizes, 96 and 192 bytes.
On my system this saves about 34k.

size-256              63    180    256    5   12    1
size-192              95    120    192    5    6    1
size-128             213    240    128    8    8    1
size-96             1080   1120     96   28   28    1

During 2.5.5-pre schedule_tail was optimized away under UP.  We need it
for preempt-kernel, too, since it decrements the preempt_count to 0
coming off do_fork.

Without this patch, CONFIG_PREEMPT + !CONFIG_SMP does not gain one very
much.  

	Robert Love

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

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

	Below is a followup to bd_claim patch - it is the last one
from the current series.  Removes SWP_BLOCKDEV flag - it isn't needed
anymore.

*	new functions - bd_claim(bdev, holder) and bd_release(bdev).
	bd_claim(bdev, holder) fails is device is already claimed by
	somebody else; bd_release(bdev) gives device up.

*	get_sb_bdev() claims device for fs_type; it means that we don't need
	to look through entire least of superblocks anymore - just through
	the list of superblocks belonging to that type (i.e. the same thing
	we do for non-block filesystems; that will allow to merge quite a
	bit of code afterwards).

*	sys_swapon claims device for itself;  free exclusion with mounting,
	end of problems with bogus set_blocksize().

*	is_mounted() and is_swap_partition() are gone - what we actually
	wanted was "try to claim device for ourselves".  Which we can do
	now - without races inherent to is_mounted()/is_swap_partition().

*	RAID lock_rdev() claims device for itself.  I.e. we get rid of
	is_mounted() in there (BTW, is_swap_partition() was missing) and
	we get protection both ways - not only RAID won't take an already
	mounted device, but mount won't stomp on a device claimed by RAID.

There are other places that would benefit from the same (e.g. ext3 with
external journal almost definitely wants to claim device for itself).

Notice that it's a cooperative thing - neither open() nor raw device stuff
claim the block device, so they don't care if device is mounted, etc.  So
we don't break fsck and friends - exclusion is between those who know that
they want that exclusion.

	1.	moves a bunch of generic ioctls from sr_dev_ioctl() to
		cdrom_ioctl().
	2.	switches blk_ioctl() to struct block_device *
	3.	ditto for blkpg_ioctl()

	Missing export for dparent_lock.

ir256_usb_cow_urballoc.diff :
 ---------------------------
	o [FEATURE] Don't use skb_cow() unless we really need to
	o [CORRECT] Reorder URB init to avoid races
	o [CORRECT] USB dealy adds processing time, not removes it
        <Following patch from Greg KH <greg@kroah.com> himself !!!>
	o [CRITICA] Use dynamically allocated URBs (instead of statically)

ir256_lap_icmd_fix-4.diff :
 -------------------------
	o [CORRECT] Fix Tx queue handling (remove race, keep packets in order)
	o [CORRECT] Synchronise window_size & line_capacity and make sure
	  we never forget to increase them (would stall Tx queue)
	o [FEATURE] Group common code out of if-then-else
	o [FEATURE] Don't harcode LAP header size, use proper constant
	o [FEATURE] Inline irlap_next_state() to decrease bloat

ir256_irnet_disc_ind.diff :
 -------------------------
	o [CORRECT] Fix IrNET disconnection to not reconnect but
	  instead to hangup pppd

ir256_sock_connect_cli.diff :
 ---------------------------
	o [CRITICA] Fix socket connect to remove dangerous cli()
	<Tested on SMP>

ir256_bus_to_virt.diff :
 ----------------------
	o [CRITICA] Fix ISA FIR drivers for new DMA API
	<PCI FIR drivers are still broken and need fixing>

into flint.arm.linux.org.uk:/usr/src/linux-bk-2.5/linux-2.5-rmk

This patch adds PCMCIA/CF support for the
PT Digital Board (CONFIG_SA1100_PT_SYSTEM3).

2nd try .... :)

Changed files:
linux/drivers/pcmcia/Makefile
linux/drivers/pcmcia/sa1100.h
linux/drivers/pcmcia/sa1100_generic.c
Added files:
linux/drivers/pcmcia/sa1100_system3.c

	- bug fix for lockup on SMP machines

into flint.arm.linux.org.uk:/usr/src/linux-bk-2.5/linux-2.5-rmk

into flint.arm.linux.org.uk:/usr/src/linux-bk-2.5/linux-2.5-rmk-patchtest

This patch allow for no MSDOS partition manager for the ARM. There are many arm
system with no hard disk partitions forcing a partition manager is not required.
 
Files affected:
 
fs/partitions/Config.in

2.5.x IRQ changes for PT Digital Board.
Also some code cleanups.

Changed files:
linux/arch/arm/mach-sa1100/system3.c
linux/include/asm-arm/arch-sa1100/irqs.h
linux/include/asm-arm/arch-sa1100/system3.h

This is a minimal "get it working again" patch; there are plans a
foot to re-jig the page table code to work better with Ingo Molnar's
changes.  These same plans also allow the ARM page tabkes to fit
into Rik van Riel's rmap significantly better.

(We're currently abusing the struct page * returned from pte_alloc_one,
treating it as if it were the same as a pte_t *)

Major SA1100 generic DMA cleanup.
Fix suspend/resume bugs.
Provide and use new SA1111 generic driver for SA1111-based devices.

Thought I should add myself for the
"ehci-hcd" and "usbnet" drivers.

This is against 2.5.5 ...

Here is a batch of accumulated bugfixes and cleanups for the Coda kernel
module. Patch is against 2.5.6-pre2, I could also send these as separate
patches.

bugfix: Fix coda_dentry_revalidate bug

    Due to a bad test, coda_dentry_revalidate was forcing revalidation
    of cacheable inodes, and allowed caching of non-cacheable inodes.

bugfix: Corrected i_mtime/i_ctime setting

    i_mtime and i_ctime were not always updated when writing to a file,
    or when modifying inode attributes.

cleanup/optimization: Avoid getattr upcalls

    We can use coda_iget directly instead of coda_cnode_make when an
    upcall returns attributes and avoid the getattr upcall altogether.

cleanup: Removed debugging messages

    CDEBUG macros haven't been useful ever since the initial development
    when they were introduced. They are too verbose for debugging
    purposes. Removing these saves about a third of the compiled size of
    the module.

    Removed print_entry variable that was used by ENTRY/EXIT macros
    which are already gone.

cleanup/optimization: Readdir simplification

    Relying on the fact that the pagecache is already buffering far more
    efficiently, simplified coda_readdir implementation. We can now fill
    the complete userbuffer instead of returning after reading only 2KB.

    Passing dir entry types that are present in the venus_dirent
    structure to the user as well.

cleanup: Removed redundant permissions statistics counters.

    The permission check count is about identical to the 'permission'
    field in the VFS stats, and the permission hit counter can trivially
    be derived from upcall_stats.access - vfs_stats.permission. Removed
    these redundant counters.

cleanup: Removed useless test for c_flags in coda_revalidate_inode.

    We already know c_flags is set due to earlier tests.

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

as intended.

This implements ptes-in-highmem for PPC, removes the quicklist
and zero-page stuff.  PTEs in highmem on SMP turned out to need
some significant changes to avoid deadlocks on the hash_table_lock
(now renamed to mmu_hash_lock).  The PMDs now contain the physical
address of the PTE page rather than the virtual address.
Anything that takes the mmu_hash_lock now operates with the DMMU
off to avoid MMU hash-table misses.

into samba.org:/home/paulus/kernel/for-linus-ppc

These are all in the 2.4.19-pre2.

"Jahn Veach" <V64@V64.net>: [PATCH 2.5.4] Typo corrections.

Pavel Machek <pavel@ucw.cz>: Whitespace PCI cleanups (fwd).

quinlan@transmeta.com: [PATCH] take COPYING into new century:
	These mods are directly from the FSF version.

	This set of one-liners are the ones which have already gone
into 2.2.19-pre2.

Andrey Panin <pazke@orbita1.ru>: [PATCH] arch_i386_kernel_smpboot.c missing __init:
  this patch adds missing __init directive for div64() function
  in arch/i386/kernel/smpboot.c. div64() is only used
  synchronize_tsc_bp() function which is marked __init.

Andrey Panin <pazke@orbita1.ru>: [PATCH] remove annoying ISAPNP message:
  This patch removes useless messages like:
  "isapnp: Calling quirk for 02:00".

John Fremlin <john@fremlin.de>: Patch to fix off by one in ide-scsi:
  Patch to fix off by one in ide-scsi

  To see it is correct compare with ide-cd behaviour

René Scharfe <l.s.r@web.de>: [PATCH] compiler warnings in scripts_tkgen.c:
  this patch fixes two compiler warnings during make xconfig which
  turn up if one uses -Wshadow

From: Dave Jones <davej@suse.de>
  kernel builds and boots with patch (with CONFIG_BLK_DEV_INITRD disabled).

(Trivial because CONFIG_BLK_DEV_RAM is Y if CONFIG_BLK_DEV_INITRD is Y)

	New helper:
path_lookup(name, flags, nd)
{
	int err = 0;
	if (path_init(name, flags, nd))
		err = path_walk(name, nd);
	return err;
}

Places doing that by hand converted to calling it.

Actually, quite a few of them were doing equivalent of __user_walk()
(getname() and if it was successful - call path_lookup() and putname()).
Converted to calling __user_walk().

	LOOKUP_POSITIVE is not needed anymore.  All callers of path_walk()
treat -ENOENT and negative dentry the same way.  If you want a proof of
correctness - I'll send it, but it's a couple of pages of induction, basically
boiling down to "let's show that for any N we can replace the
               if (lookup_flags & (LOOKUP_POSITIVE|LOOKUP_DIRECTORY))
                       break;
in link_path_walk() with
               if ((lookup_flags & (LOOKUP_POSITIVE|LOOKUP_DIRECTORY)) ||
		   current->link_count <= N)
                       break;
without changing behaviour of the system".  Pretty straightforward for
N = 0, then we look for places that can lead to call link_path_walk()
with current->link_count equal to N and show that if result of the test
changes, behaviour of callers doesn't.  Since the depth of recursion is
limited, we had shown that test in question can be replaced with if (1).
And that's the only place in tree the ever checks for LOOKUP_POSITIVE.

	The real reason behind that is very simple - indeed, suppose
we get a negative dentry out of path_walk().  What the hell could we
do with it?  Its parent isn't locked, so both the name and parent can
change at any moment (could have changed already).  There used to be
places that tried to play "let's get a negative dentry, lock its parent
and start doing something".  All of them racy and all of them fixed
in 2.3.  Fixed by switching to LOOKUP_PARENT...

	More cleanups (getblk() use, etc.) in reiserfs (again, cleaning
up after merge).