If ext3_add_nondir() fails it will do an iput() of the inode.  But we
continue to run ext3_mark_inode_dirty() against the potentially-freed
inode.  This oopses when slab poisoning is enabled.

Fix it so that we only run ext3_mark_inode_dirty() if the inode was
successfully instantiated.

Renames the local variables `bh2', `i', `j', 'k', and `tmp' to
something meanigful.  This brings ext2_new_block() into line with
ext3_new_block().

The same thing, for ext2.

When an ext3 (or ext2) file is first created the filesystem has to
choose the initial starting block for its data allocations.  In the
usual (new-file) case, that initial goal block is the zeroeth block of
a particular blockgroup.

This is the worst possible choice.  Because it _guarantees_ that this
file's blocks will be pessimally intermingled with the blocks of
another file which is growing within the same blockgroup.

We've always had this problem with files in the same directory.  With
the introduction of the Orlov allocator we now have the problem with
files in different directories.  And it got noticed.  This is the cause
of the post-Orlov 50% slowdown in dbench throughput on ext3 on
write-through caching SCSI on SMP.  And 25% in ext2.

It doesn't happen on uniprocessor because a single CPU will not exhibit
sufficient concurrency in allocation against two or more files.

It will happen on uniprocessor if the files are growing slowly.

It has always happened if the files are in the same directory.

ext2 has the same problem but it is siginficantly less damaging there
because of ext2's eight-block per-inode preallocation window.

The patch largely solves this problem by not always starting the
allocation goal at the zeroeth block of the blockgroup.  We instead
chop the blockgroup into sixteen starting points and select one of those
based on the lower four bits of the calling process's PID.

The PID was chosen as the index because this will help to ensure that
related files have the same starting goal.  If one process is slowly
writing two files in the same directory, we still lose.


Using the PID in the heuristic is a bit weird.  As an alternative I
tried using the file's directory's i_ino.  That fixed the dbench
problem OK but caused a 15% slowdown in the fast-growth `untar a kernel
tree' workload.  Because this approach will cause files which are in
different directories to spread out more.  Suppressing that behaviour
when the files are all being created by the same process is a
reasonable heuristic.


I changed dbench to never unlink its files, and used e2fsck to
determine how many fragmented files were present after a `dbench 32'
run.  With this patch and the next couple, ext2's fragmentation went
from 22% to 13% and ext3's from 25% to 10.4%.

ext2 places non-directory objects into the same blockgroup as their
directory, as long as that directory has free inodes.  It does this
even if there are no free blocks in that blockgroup (!).

This means that if there are lots of files being created at a common
point in the tree, they _all_ have the same starting blockgroup.  For
each file we do a big search forwards for the first block and the
allocations end up getting intermingled.

So this patch will avoid placing new inodes in block groups which have
no free blocks.

So far so good.  But this means that if a lot of new files are being
created under a directory (or multiple directories) which are in the
same blockgroup, all the new inodes will overflow into the same
blockgroup.  No improvement at all.

So the patch arranges for the new inode locations to be "spread out"
across different blockgroups if they are not going to be placed in
their directory's block group.  This is done by adding parent->i_ino
into the starting point for the quadratic hash.  i_ino was chosen so
that files which are in the same directory will tend to all land in the
same new blockgroup.

- Add some (much-needed) commentary to the ext2/ext3 block allocator
  state fields.

- Remove the SEARCH_FROM_ZERO debug code.  I wrote that to trigger
  some race and it hasn't been used in a year.

The `low latency page reclaim' design works by preventing page
allocators from blocking on request queues (and by preventing them from
blocking against writeback of individual pages, but that is immaterial
here).

This has a problem under some situations.  pdflush (or a write(2)
caller) could be saturating the queue with highmem pages.  This
prevents anyone from writing back ZONE_NORMAL pages.  We end up doing
enormous amounts of scenning.

A test case is to mmap(MAP_SHARED) almost all of a 4G machine's memory,
then kill the mmapping applications.  The machine instantly goes from
0% of memory dirty to 95% or more.  pdflush kicks in and starts writing
the least-recently-dirtied pages, which are all highmem.  The queue is
congested so nobody will write back ZONE_NORMAL pages.  kswapd chews
50% of the CPU scanning past dirty ZONE_NORMAL pages and page reclaim
efficiency (pages_reclaimed/pages_scanned) falls to 2%.

So this patch changes the policy for kswapd.  kswapd may use all of a
request queue, and is prepared to block on request queues.

What will now happen in the above scenario is:

1: The page alloctor scans some pages, fails to reclaim enough
   memory and takes a nap in blk_congetion_wait().

2: kswapd() will scan the ZONE_NORMAL LRU and will start writing
   back pages.  (These pages will be rotated to the tail of the
   inactive list at IO-completion interrupt time).

   This writeback will saturate the queue with ZONE_NORMAL pages.
   Conveniently, pdflush will avoid the congested queues.  So we end up
   writing the correct pages.

In this test, kswapd CPU utilisation falls from 50% to 2%, page reclaim
efficiency rises from 2% to 40% and things are generally a lot happier.


The downside is that kswapd may now do a lot less page reclaim,
increasing page allocation latency, causing more direct reclaim,
increasing lock contention in the VM, etc.  But I have not been able to
demonstrate that in testing.


The other problem is that there is only one kswapd, and there are lots
of disks.  That is a generic problem - without being able to co-opt
user processes we don't have enough threads to keep lots of disks saturated.

One fix for this would be to add an additional "really congested"
threshold in the request queues, so kswapd can still perform
nonblocking writeout.  This gives kswapd priority over pdflush while
allowing kswapd to feed many disk queues.  I doubt if this will be
called for.

We keep getting in a mess with the current->flags setting and
unsetting.

Remove current->flags:PF_NOWARN and create __GFP_NOWARN instead.

- A C99 initialiser in drivers/char/mem.c

- Remove unneeded deref in madvise_willneed()

Add a generic_file_readonly_mmap() for !CONFIG_MMU.

slab poisons objects with 0x5a both when they are constructed and when
they are freed.  So it is not possible to tell whether a deref of
0x5a5a5a5a was a use-before-initialisation bug or a use-after-free bug.

The patch changes it so that

1) A deref of 0x5a5a5a5a means use-of-uninitialised-memory

2) A deref of 0x6b6b6b6b means use-of-freed-memory.

move_vma() calls do_munmap() and then uses the memory at *new_vma.

But when starting X11 it just happens that the memory which do_munmap
unmapped had the same start address and the range at *new_vma.  So new_vma
is freed by do_munmap().

This was never noticed before because (vm_flags & VM_LOCKED) evaluates
false when vm_flags is 0x5a5a5a5a.  But I just changed that to 0x6b6b6b6b
and boom - we call make_pages_present() with start == end == 0x6b6b6b6b and
it goes BUG.

So I think the right fix here is for move_vma() to not inspect the values
of any vma's after it has called do_munmap().

The patch does that, for `new_vma'.

The local variable `vma' is also being used after the call do do_munmap(),
and this may also be a bug.  Proving that this is not so, and adding a
comment to explain why is hereby added to Hugh's todo list ;)

There's a small window in which another CPU could dirty the page after
we've cleaned it, and before we've moved it to mapping->dirty_pages().
The end result is a dirty page on mapping->locked_pages, which is
wrong.

So take mapping->page_lock before clearing the dirty bit.

Running a `mount -o remount' against ext3 deadlocks if there is heavy
write activity.  It's a sort of AB/BA deadlock caused by calling
log_wait_commit() under lock_super().  The caller holds lock_super()
and is waiting for a commit, but the commit cannot complete because
lock_super() is also used in the block allocator.

The way we fixed this in tha past is to drop the superblock lock inside
ext3.  The way this patch fixes it is to arrange for lock_super() to
not be held around the ->sync_fs() call.

Also: sync_filesystems is on the sys_sync() path and is racy wrt
unmount.  Check sb->s_root after taking sb->s_umount.

 - set up kernel stack pointer for sysenter at each context switch.
 - disable sysenter while in vm86 mode.
 - clean up mtrr number defines and SEP feature testing

Don't disable PCI devices before changing the BARs, as discussed
recently.  Disabling PCI_COMMAND_MASTER bit is an obvious bug.

Further, pdev_enable_device() is a leftover from very old (2.0, I guess)
alpha PCI code.  It's used in pci_assign_unassigned_resources() to
enable *every* PCI device in the system.  So, if we have two graphic
cards on the same bus, both with legacy VGA IO...  oops.

Actually, only alpha relied on that due to the lack of
pcibios_enable_device (which has been already fixed).

This replaces the dynamically allocated two-level array in sys_poll with
a dynamically allocated linked list.  The current implementation causes
at least two alloc/free calls, even if only one or two descriptors are
polled.  This reduces that to one alloc/free, and the .text segment is
around 220 bytes shorter.  The microbenchmark that polls one pipe fd is
around 30% faster.  [1140 cycles instead of 1604 cycles, Celeron mobile
1.13 GHz]

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

into tetrachloride.(none):/mnt/stuff/kernel/2.5/agpgart

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

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

Description:
  everywhere the NFS client uses the req_offset() function today, it adds
  req->wb_offset to the result.  this patch simply makes "+req->wb_offset"
  a part of the req_offset() function.

Test status:
  Passes all Connectathon '02 tests with v2, v3, UDP and TCP.  Passes
  NFS torture tests on an x86 UP highmem system.

Description:
  The default set_page_dirty address space op is too heavyweight for NFS,
  which doesn't use buffers.

Description:
  andrew morton suggested there are places in the NFS client that could
  make use of kmap_atomic instead of vanilla kmap in order to improve
  scalability on 8-way and higher SMP systems.

Test status:
  Passes all Connectathon '02 tests with v2 and v3, UDP and TCP; passes
  NFS torture tests on a UP HIGHMEM x86 system.

This moves most of the duplicated text in the various v850 platform-
specific linker scripts (each of which was previously completely
standalone) into cpp macros in vmlinux.lds.S, which are then used by the
platform linker scripts as appropriate.  This should make the scripts a
lot easier to maintain.

Also, a number of linker-script bugs are fixed.

The old code seems completely wrong; I guess it was just left over from
whichever architecture this code was copied from.

These are used for the new in-kernel module loader (actually not all the
relocation types are used right now, but are included for completeness).

Only the EM_CYGNUS_V850 macro, which is in a global namespace, is added
to <linux/elf.h>; the relocation types, which are private to the v850,
are added to <asm-v850/elf.h>.  [Perhaps some other archs can do a
similar split, to reduce the bloat in <linux/elf.h>]

A few symbols are only defined when CONFIG_MMU=y, but are exported
(by kernel/ksyms.c) unconditionally.  This patch makes them conditional.

Adds extra includes needed because sched.h doesn't include them anymore,
and removes includes of sched.h where they're not really necessary.

Fix task->cpus_allowed bitmask truncations on 64.bit architectures.

Originally by Bjorn Helgaas for 2.4.x.

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