into maxwell.earthlink.net:/usr/src/linus-2.5

The Documentation/Changes in the summary has been updated to require
make 3.78 but the other references were not updated.  And 3.78 really
is required.  This patch updates the other locations.

Dirty memory thresholds are currently set by /proc/sys/vm/dirty_ratio.

Background writeout levels are controlled by
/proc/sys/vm/dirty_background_ratio.

Problem is that these levels are hard to get right - they are too
static.  If there is a lot of mapped memory around then the 40%
clamping level causes too much dirty data.  We do lots of scanning in
page reclaim, and the VM generally starts getting into distress.  Extra
swapping, extra page unmapping.

It would be much better to simply tell the caller of write(2) to slow
down - to write out their dirty data sooner, to make those written
pages trivially reclaimable.  Penalise the offender, not the innocent
page allocators.

This patch changes the writer throttling code so that we clamp down
much harder on writers if there is a lot of mapped memory in the
machine.  We only permit memory dirtiers to dirty up to 50% of unmapped
memory before forcing them to clean their own pagecache.

Since /proc/sys/vm/dirty_sync_ratio went away, the name
"dirty_async_ratio" makes no sense.

So rename it to just /proc/sys/vm/dirty_ratio.

We're currently adding anon pages to the inactive list.  But they're
all referenced, so when they reach the tail of the inactive list the
kernel will always then bump them up to the active list.

Not only does this waste CPU, but it leads to inactive/active
imbalance.  We end up with enormous sequences of unreclaimable,
to-be-activated pages hitting the tail of the LRU and large amounts of
scanning need to be done.  Which upsets the VM, making it think that it
is "under distress".

So just start them out on the active list.

/proc/sys/vm/swappiness controls the VM's tendency to unmap pages and to
swap things out.

100 -> basically current 2.5 behaviour
0 -> not very swappy at all

The mechanism which is used to control swappiness is: to be reluctant
to bring mapped pages onto the inactive list.  Prefer to reclaim
pagecache instead.

The control for that mechanism is as follows:

- If there is a large amount of mapped memory in the machine, we
  prefer to bring mapped pages onto the inactive list.

- If page reclaim is under distress (more scanning is happening) then
  prefer to bring mapped pages onto the inactive list.  This is
  basically the 2.4 algorithm, really.

- If the /proc/sys/vm/swappiness control is high then prefer to bring
  mapped pages onto the inactive list.

The implementation is simple: calculate the above three things as
percentages and add them up.  If that's over 100% then start reclaiming
mapped pages.

The `proportion of mapped memory' is downgraded so that we don't swap
just because a lot of memory is mapped into pagetables - we still need
some VM distress before starting to swap that memory out.

For a while I was adding a little bias so that we prefer to unmap
file-backed memory before swapping out anon memory.  Because usually
file backed memory can be evicted and reestablished with one I/O, not
two.  It was unmapping executable text too easily, so here I just treat
them equally.

zap_pte_range() is currently just dropping the pte.  Change it to mark
the page referenced if the pte says it was.  This has the effect of
delaying the eviction of recently-mapped pagecache.

This means that we're currently marking the page accessed when it is
first faulted in as well as when we drop it from pagetables.  Which
matches up with the (strange) behaviour of the VM: it reclaims
PageReferenced pagecache pages off the inactive list.

Probably, it makes sense to remove the mark_page_accessed() from
filemap_nopage() and just use the pte bits everywhere.  Reviewing all
the PageReferenced()/mark_page_accessed() usage is on my todo list.

The current writer throttling in balance_dirty_pages() assumes that the
writer will be effectively throttled on request queues.

That works fine when the amount of data which can be placed into a
queue is "much less than" total memory.

But if the machine has a small amount of memory, or many disks, or has
large request queues, or large requests, it can go wrong.

For example, with mem=96m and dirty_async_ratio=15, we want to be able
to clamp dirty+writeback memory at 15 megabytes.  But it doesn't work,
because a single SCSI request queue can hold 40 megs or more.  The
heavy writer keeps on dirtying memory until that queue fills up.

So add a test for that - if we did some writeback, and we're *still*
over the dirty+writeback threshold then make the caller take an
explicit nap on some writes terminating.  And keep on doing that until
the dirty+writeback memory subsides.

Software suspend needs a way of forcing page reclaim, up to the point
where 50% of memory is free.

This patch implements a function to do that:

	int shrink_all_memory(int nr_pages);

Will attempt to reclaim `nr_pages' pages and return them to the free
pages pool.  It returns the number of pages which it actually freed.

If called with a "large" number of pages it will only free up to a few
hundred, so the caller needs to loop on it.

If it returns zero then there is no point in calling it again.

Patch from Badari Pulavarty

There was a corner case in the conversion of direct-io to use
bio_add_page() where we would start a new page out at the wrong sector
number.

Fix that by explicitly passing in the current page's starting sector,
and use that in the new BIO if we have to open a new one.

Fix an error-path page->count leak in dio_bio_add_page().

The patch from Rohit and David M-T changes the hugetlb page info in
/proc/meminfo slightly.

It makes the identifiers a little clearer while ensuring that we don't
add any identifiers which have whitespace.  glibc is/shall be parsing
this information to determine the size and alignment requirements of
the hugetlb pages.

This basically means that procfs is a requirement for successful
hugetlb page usage.  Not very nice, but I suspect real-world userspace
fails without procfs anyway.

- drivers/char/n_r3964.c does not compile.  r3964_open() is doing an
  INIT_LIST_HEAD() on a timer->list.  But timer's don't have a `list'
  any more.

  Do an init_timer() instead.

I'm getting a compile failure in scsi_syms.c because it doesn't know
about the new scsi_set_medium_removal().

I'm suprised that this didn't cause errors for more people -- a
MODE_SENSE request for 128 bytes with a stated buffer length of 24
bytes.

Fix: Make the buffer length match the size of the request.

This patch fixes a typo introduced by my last patch to dv1394.

This adds the dmask option. Yes, the dmask option is the permission
bitmask for directory.

This adds fat_show_options() to fat. And instead, this doesn't output
the charset name in fat_fill_super().

This removes the posix option of vfat. The current posix options works
only as an alias of name_check=s.

This merges parse_options() of fat and parse_options() of vfat.
And this doesn't recognize the unknown options.

This fixes the error code which fat_fill_super() returns.

 - Cleanup (purge) some of our old compat code (never thouched)
 - Fix dv1394 compilation warnings without devfs
 - Added new config-rom handling features. Allows for on-the-fly
   config-rom generation for dynamic functionality of the host nodes.
 - Convert to workqueue from taskqueue interfaces. This is actually
   abstracted compatibility code between tqueue/workqueue.

To my suprise a lot of big site/beowulf type people all really want this
diff, which I'd otherwise filed as 'interesting but not important'

(someone really needs to indent the cpqfc driver!)