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

Patch from Rohit Seth: allow hugetlb pages to be allocated from the
highmem zone.

From Marcus Alanen <maalanen@ra.abo.fi>

Don't retake the zone lock after spilling a batch of pages into the
buddy.

Instead, just clear local variable `zone' to indicate that no lock is
held.

This is actually a common case - whenever release_pages() is called
with exactly 16 pages (truncate, page reclaim..) Marcus' patch will
save a lock and an unlock.

Also, remove some lock-avoidance heuristics in
pagevec_deactivate_inactive(): the caller has already made these
checks, and the chance of the check here actually doing anything useful
is negligible.

- Spell Jeremy's name correctly.

- Fix compile warning in raw.c

- Do a waitqueue_active() test before waking klogd in printk.

  Not only is is negligibly faster, but the wake_up() in there causes
  deadlocks when you try to print debug info out from inside scheduler
  code.

  This patch gives a delightfully obscure way of avoiding the
  deadlock: kill off klogd.

- Fix a couple of compile warnings in the mtrr code.

From Christoph Hellwig, acked by Jens.

- remove some unneeded runtime initializers.

- remove the explicit call to hd_init() - it already goes through
module_init(), so we're currently running hd_init() twice.

From Christoph Hellwig, acked by Rohit.

- fix config.in description: we know we're on i386 and we also know
that a feature can only be enabled if the hw supports it, the code
alone is not enough

- the sysctl is VM-releated, so move it from /proc/sys/kernel tp
/proc/sys/vm

- adopt to standard sysctl names

From Christoph Hellwig.

There are no lock_kernel() calls in mm/

From Hubertus Franke.

The MAP_LOCKED flag to mmap() currently does nothing.  Hubertus' patch
fixes it so that the relevant mapping is locked into memory, if the
called has CAP_IPC_LOCK.

Somebody somewhere is stomping on PF_NOWARN, and page allocation
failure warnings are coming out of the wrong places.

So change the handling of current->flags to be:

int pf_flags = current->flags;

current->flags |= PF_NOWARN;
...
current->flags = pf_flags;

which is a generally more robust approach.

- writev currently returns -EFAULT if _any_ of the segments has an
invalid address.  We should only return -EFAULT if the first segment
has a bad address.

If some of the first segments have valid addresses we need to write
them and return a partial result.

- The current code only checks if the sum-of-lengths is negative.  If
individual segments have a negative length but the result is positive
we miss that.

So rework the code to detect this, and to be immune to odd wrapping
situations.

As a bonus, we save one pass across the iovec.

- ditto for readv.

The check for "does any segment have a negative length" has already
been performed in do_readv_writev(), but it's basically free here, and
we need to do it for generic_file_read/write anyway.

This all means that the iov_length() function is unsafe because of
wrap/overflow isues.  It should only be used after the
generic_file_read/write or do_readv_writev() checking has been
performed.  Its callers have been reviewed and they are OK.

The code now passes LTP testing and has been QA'd by Janet's team.

A patch from Hirokazu Takahashi to speed up the new sped-up writev
code.

Instead of running ->prepare_write/->commit_write for each individual
segment, we walk the segments between prepage and commit.  So
potentially much larger amounts of data are passed to commit_write(),
and prepare_write() is called much less often.

Added bonus: the segment walk happens inside the kmap_atomic(), so we
run kmap_atomic() once per page, not once per segment.

We've demonstrated a speedup of over 3x.  This is writing 1024-segment
iovecs where the individual segments have an average length of 24
bytes, which is a favourable case for this patch.

Silly bug which was halving swapout bandwidth: we've taken a copy of
page->mapping into a local convenience variable, but forgot to update
that local after adding the page to swapcache.

This was designed to be a really sterm throttling threshold: if dirty
memory reaches this level then perform writeback and actually wait on
it.

It doesn't work.  Because memory dirtiers are required to perform
writeback if the amount of dirty AND writeback memory exceeds
dirty_async_ratio.

So kill it, and rely just on the request queues being appropriately
scaled to the machine size (they are).

This is basically what 2.4 does.

Bill Irwin's patch to avoid having to walk pagetables while generating
/proc/stat output.

It can significantly overstate the size of various mappings because it
assumes that all VMAs are fully populated.

But spending 100% of one of my four CPUs running top(1) is a bug.

Bill says this fixes a bug, too.  The `SIZE' parameter is supposed to
display the amount of memory which the process would consume if it
faulted everything in.  But "before it only showed instantiated
3rd-level pagetables, so if something within a 4MB aligned range hadn't
been faulted in it would slip past the old one".

Patch from Martin Bligh.  It should only affect machines using
discontigmem.

"This patch is was originally from Andrea's tree (from SGI??), and has
been tweaked since by both Christoph (who cleaned up all the code),
and myself (who just hit it until it worked).

It removes _alloc_pages, and adds all nodes to the zonelists
directly, which also changes the fallback zone order to something more
sensible ...  instead of: "foreach (node) { foreach (zone) }" we now
do something more like "foreach (zone_type) { foreach (node) }"

Christoph has a more recent version that's fancier and does a couple
more cleanups, but it seems to have a bug in it that I can't track
down easily, so I propose we do the simple thing for now, and take the
rest of the cleanups when it works ...  it seems to build nicely on
top of this seperately to me.

Tested on 16-way NUMA-Q with discontigmem + NUMA support."

Patch from Martin Bligh.  It should only affect machines using
discontigmem.

"This patch cleans up free_area_init stuff, and undefines mem_map and
max_mapnr for discontigmem, where they were horrible kludges anyway
...  We just use the lmem_maps instead, which makes much more sense.
It also kills pgdat->node_start_mapnr, which is tarred with the same
brush.

It breaks free_area_init_core into a couple of sections, pulls the
allocation of the lmem_map back into the next higher function, and
passes more things via the pgdat.  But that's not very interesting,
the objective was to kill mem_map for discontigmem, which seems to
attract bugs like flypaper.  This brings any misuses to obvious
compile-time errors rather than wierd oopses, which I can't help but
feel is a good thing.

It does break other discontigmem architectures, but in a very obvious
way (they won't compile) and it's easy to fix.  I think that's a small
price to pay ...  ;-) At some point soon I will follow up with a patch
to remove free_area_init_node for the contig mem case, or at the very
least rename it to something more sensible, like __free_area_init.

Christoph has grander plans to kill mem_map more extensively in
addition to the attatched, but I've heard nobody disagree that it
should die for the discontigmem case at least.

Oh, and I renamed mem_map in drivers/pcmcia/sa1100 to pc_mem_map
because my tiny little brain (and cscope) find it confusing like that.

Tested on 16-way NUMA-Q with discontigmem + NUMA support and on a
standard PC (well, boots and appears functional).  On top of
2.5.33-mm4"

The writeback code paths which walk the superblocks and inodes are
getting an increasing arguments passed to them.

The patch wraps those args into the new `struct writeback_control',
and uses that instead.  There is no functional change.

The new writeback_control structure is passed down through the
writeback paths in the place where the old `nr_to_write' pointer used
to be.

writeback_control will be used to pass new information up and down the
writeback paths.  Such as whether the writeback should be non-blocking,
and whether queue congestion was encountered.

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 mandrakesoft.com:/home/jgarzik/repo/net-drivers-2.5

when implementing ethtool media ioctls.

No behavior should change with this patch (except the ethtool media
ioctls now work, of course)

Also, re-format comments to the right of the private struct to line up.

to take 'struct net_device *' not 'long' as their first
argument.  This makes eepro100 compatible with the standard
MII ethtool API, preparing it for that support.

No functional changes should occur with this patch, if anything
changes at all it is a bug.  (and testing shows no changes...)

(kill more e100 compiler warnings)

* PCI write posting fixes, remove pa-specific code
* support 2.5.x kernels (synchronize_irq, cli/sti cleanups)

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

(no change in function).

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

The partition code registers a generic device for disks
which have a dev->driver non-NULL but whose dev->driver->remove
points into outer space.  So when reboot happens --> OOPS
in drivers/base/power.c:device_shutdown()

Ok, amusingly in my case dev->driver == &scsi_done(), hehe :-)

Two cases of uninitialized memory spotted, here is the patch.

(suggested by Donald Becker)

into mandrakesoft.com:/home/jgarzik/repo/net-drivers-2.5

* default to PIO (fixes bugs in some chips), but add CONFIG_xxx option
  for MMIO
* proper support for variable MTU sizes
* add missing unregister_netdev in an error path
(with a kudos to Jason Lunz for merging most of this)