zone->lock and zone->lru_lock are two of the hottest locks in the
kernel.  Their usage patterns are quite independent.  And they have
just been put into the same structure.  It is essential that they not
fall into the same cacheline.

That could be fixed by padding with L1_CACHE_BYTES.  But the problem
with this is that a kernel which was configured for (say) a PIII will
perform poorly on SMP PIV.  This will cause problems for kernel
vendors.  For example, RH currently ship PII and Athlon binaries.  To
get best SMP performance they will end up needing to ship a lot of
differently configured kernels.

To solve this we need to know, at compile time, the maximum L1 size
which this kernel will ever run on.

This patch adds L1_CACHE_SHIFT_MAX to every architecture's cache.h.

Of course it'll break when newer chips come out with increased
cacheline sizes.   Better suggestions are welcome.

Now the LRUs are per-zone, make their lock per-zone as well.

In this patch the per-zone lock shares a cacheline with the zone's
buddy list lock, which is very bad.  Some groundwork is needed to fix
this well.

This change is expected to be a significant win on NUMA, where most
page allocation comes from the local node's zones.

For NUMA the `struct zone' itself should really be placed in that
node's memory, which is something the platform owners should look at.
However the internode cache will help here.

Per-node kswapd would make heaps of sense too.

Replace the global page LRUs with per-zone LRUs.

This fixes the failure described at
http://mail.nl.linux.org/linux-mm/2002-08/msg00049.html

It will also fixes the problem wherein a search for a reclaimable
ZONE_NORMAL page will undesirably move aged ZONE_HIGHMEM pages to the
head of the inactive list.  (I haven't tried to measure any benefit
from this aspect).

It will also reduces the amount of CPU spent scanning pages in page
reclaim. I haven't instrumented this either.

This is a minimal conversion - the aging and reclaim logic is left
unchanged, as far as is possible.

I was bitten by the "incremental min" logic in __alloc_pages again.
There's a state in which the sum-of-mins exceeds zone->pages_high.  So
we call into try_to_free_pages(), which does nothing at all (all zones
have free_pages > pages_high).  The incremental min is unchanged and
the VM locks up.

This was fixed in __alloc_pages: if zone->free_pages is greater than
zone->pages_high then just go and grab a page.

- Remove the zonelist_t typedef.  Rename struct zonelist_struct to
struct zonelist and use that everywhere.

- Remove the zone_t typedef.  Rename struct zone_struct to struct
zone and use that everywhere.

Don't align the buffer_head slab on hardware cacheline boundaries.
It's on the wrong side of the speed/space tradeoff, especially for
P4's.

It has been noticed that across a kernel build many calls to
tlb_flush_mmu() do not have anything to flush, apparently because glibc
is mmapping a file over a previously-mapped region which has no
faulted-in ptes.

This patch detects this case and optimises away a little over one third
of the tlb invalidations.

The functions which potentially cause an invalidate are
tlb_remove_tlb_entry(), pte_free_tlb() and pmd_free_tlb().  These have
been front-ended in asm-generic/tlb.h and the per-arch versions now
have leading double-underscores.  The generic versions tag the
mmu_gather_t as needing a flush and then call the arch-specific
version.

tlb_flush_mmu() looks at tlb->need_flush and if it sees that no real
activity has happened, the invalidation is avoided.

The success rate is displayed in /proc/meminfo for the while.  This
should be removed later.

- Kill duplicate debug check

- Add a test for pages which have a pte_chain coming off the buddy lists.

Some day gcc may drop support for __FUNCTION__ entirely, we'll
just add a #define __FUNCTION__ __func__ at that point. In the
meantime, gcc-3.x warns about pasting __FUNCTION__, so don't do it.

This came up in 2.5.30, apparently someone was fooled by names
which do not quite match actual functions.

add_gendisk()/del_gendisk() moved into ->reinit() and ->cleanup() of
ide-{disk,cd,floppy} - i.e.  moments when high-levle driver claims/gives
up a drive.

register_disk() also shifted into ->reinit().

consequently, revalidate_drives() is gone (it did messy postponed
rereading of partition tables; not needed anymore).  Ditto for
ide_geninit().

regular 2.5 changes in ->revalidate() and BLKRRPART handling - same as
all other block devices.

instead of messing with ide_module_t, we put ide_driver_t themselves on
a (cyclic) list - said list being the only use of ide_module_t for
high-level drivers.  ide_register_module()/ide_unregister_module() takes
ide_driver_t now (renamed to ide_register_driver()).  /proc/ide/drivers
switched to use of that cyclic list and uses seq_file instead of old
home-grown code.

->init() for high-level drivers is never called (other than as
module_init() when they are initialized).  Method removed, instances
cleaned up.

ide_reinit_drive() turned into ata_attach().  Said beast takes a drive,
tries to feed it to high-level drivers and drops it on the ata_unused if
nobody claims the sucker.  IOW, that's what ide_register_module() used
to do, but for a single drive.

ideprobe_init() calls ata_attach() instead of putting on ata_unused.

ide_register_module() eliminated.  Some of the callers do not need it
anymore, some (ide_replace_subdriver()) actually want ata_attach(drive).

ide_scan_devices() is gone.  There were two remaining callers - in
ide_register_module() and ide_unregister_module().  The former had been
turned into "put driver on the list, empty ata_unused into temporary
list and call ata_attach() on all drives there".  The latter is "remove
driver from the list, call ->cleanup() and ata_attach() for all drives"
(->cleanup() gives the drive up, ata_attach() gives the remaining
drivers a shot for that drive; if nobody claims it - it's put on
ata_unused).

->owner added to ide_driver_t.  MOD_INC_USE_COUNT/MOD_DEC_USE_COUNT
taken out of ->reinit().  ide_reinit_drive() turned into "call
->reinit() for all high-level drivers that are registered until somebody
claims the drive" (instead of open-coded variant in 2.5.32; cleaner and
works correctly for modular drivers).

loops in ide_cdrom_init()/ide_cdrom_exit(), etc. are pulled into
ide_register_module()/ide_unregister_module() resp.

Duplicate calls of ide_cdrom_init(), idedisk_init(), etc. are removed
from ide_init_builtin_drivers() (they were called both from there (i.e.
from ide_init()) and later as module_init() for high-level drivers).

Checks for media type, ->driver_req, etc.  are moved from the
ide_scan_devices() to ->reinit().  ide_scan_devices() had lost first two
arguments (it will completely disappear later).

This puts drives on cyclic lists - per-driver ones for drives that had
been claimed by high-level drivers and ata_unused for unclaimed drives.
We put drives on ata_unused in the very end of ideprobe_init() and then
move them to drivers' lists as they are claimed.

Finish introduction of ->reinit() - Jens had missed MOD_DEC_USE_COUNT on
several exits from ide-cd one and forgot to remove the loop from
ide-floppy ide-tape and ide-scsi ones ;-) (->reinit() is the body of
loop in ->init() - stuff that should be done one drive; in 2.5.32
ide-disk one is OK, ide-cd is OK modulo minor bugs and in the rest it's
a copy of ->init())

Move stuff from ide_register_subdriver() (associating drive with
high-level driver) to ide-probe.c, so that remaining stuff can be safely
called in parallel with IO on other drives [Andre]

This adds list_for_each_entry, which is the equivalent of list_for_each
and list_entry, except only one variable is needed.

Just pushes the BKL down a little, no big deal.

This optimisation is really noticeable as it avoids having to take the
BKL on every close().

 - remove elf_caddr_t. It's positively dangerous to #define this since
   elf_caddr_t foo, bar; creates variables of different types (foo is
   char *, bar is char).
 - rewrite large chunks of create_elf_tables(), it needed cleaning anyway.
 - add upwards-growing stack support to create_elf_tables.
 - redefine the ARCH_DLINFO stuff on powerpc -- it's tested, works.
 - add upwards-growing-stack support to exec.c too.

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

and one USB net driver:  3c574_cs, ibmtr_cs, pcnet_cs, ray_cs,
xirc2ps_cs, xircom_cb, and usb/net/kaweth

return EOPNOTSUPP if no eeprom present, in ETHTOOL_SEEPROM ioctl

(thanks for DaveM for advice and help)

* fix tx checksumming (it's still ifdef'd out by default)
* bump version to 0.2.0
* ifdef out dev->change_mtu support, it is reported broken.

o Feature: Merged NAPI support from Robert Olsson

o Bug fix: moved tx_timeout processing from interrupt context 
  to process context so h/w controller reset can busy-wait.

o Bug fix: ethtool SSET and NWAY we're broken when !netif_running.
o Cleanup: removed PPC code optimization - not needed.
o Bug fix: workaround for Dell Avalon system: change descriptor
  write-back policy to write back 16 bytes rather than 8 bytes.
o Bug fix: added mwb() to force memory transaction ordering on ia-64.
o Cleanup: misc. whitespace cleanup.