Patch from "Theodore Ts'o" <tytso@mit.edu>

We now use 0x7ffffff as the EOF cookie, because Linux NFS stupidly interprets
the cookie (which is supposed to be a bag of bits without necessarily any
semantic value) as a signed 64 bit integer, and then converts it to a
unsigned integer, and then blows up if it cannot be expressed be expressed as
a 32-bit value!!

In order to do this, we have to fold the hash value 0x7ffffff into the hash
value 0x7ffffffe.  This is relatively safe; the only time we will lose if the
directory contains filenames that hash to both 0x7ffffffe and 0x7fffffff
(under the original hash), and the last directory entry which hashes to
0x7ffffffe is at the end of a leaf block, and the first directory entry which
hashes to 0x7fffffff is at the beginning of a leaf block.

Patch from "Theodore Ts'o" <tytso@mit.edu>

The following patch should (in theory) fix the htree/NFS readdir problems
that people have reported.  Specifically, it should fix the NFS looping on
EOF problem with readdir, as well as the problems caused by coverting a
directory to HTREE while an NFS readdir is in progress problem.

I'd appreciate it if people who can easily replicate these NFS/htree problems
could give this patch (against BK-recent / 2.5.63) a whirl.  Thanks!!

From: John M Flinchbaugh <glynis@butterfly.hjsoft.com>

this patch makes arch/i386/oprofile/init.c build.

From: Andreas Dilger <adilger@clusterfs.com>

Below are the patches which reserve the Lustre EA index.  The rest of the
code is part of the Lustre tree, which isn't working with 2.5 yet.

From: William Lee Irwin III <wli@holomorphy.com>

The new vmalloc() semantics from 2.5.32 had a race window.  As things stand,
the presence of a vm_area in the vmlist protects from allocators other than
the owner examining the ptes in that area.  This puts an ordering constraint
on unmapping, so that allocators are required to unmap areas before removing
them from the list or otherwise dropping the lock.

Currently, unmap_vm_area() is done outside the lock and after the area is
removed, which as we've seen from Felix von Leitner's test is oopsable.

The following patch folds calls to unmap_vm_area() into remove_vm_area() to
reinstate what are essentially the 2.4.x semantics of vfree().  This renders
a number of unmap_vm_area() calls unnecessary (and in fact oopsable since
they wipe ptes from later allocations).  It's an open question as to whether
this is sufficiently performant, but it is the minimally invasive approach.
The more performant alternative is to provide the right API hooks to wipe the
vmalloc() area clean before removing them from the list, using the ownership
of the area to eliminate holding the vmlist_lock for the duration of the
unmapping.  If it proves to be necessary wli is on standby to implement it.

From: Manfred Spraul <manfred@colorfullife.com>

de_thread calls list_del(&current->tasks), but current->tasks was never
added to the task list. The structure contains stale values from the parent.

switch_exec_pid() transforms a normal thread to a thread group leader.
Thread group leaders are included in the init_task.tasks linked list,
non-leaders are not in that list.  The patch adds the new thread group
leader to the linked list, otherwise de_thread corrupts the task list.

Rather than assuming that all the things which copy_process() calls want to
return -ENOMEM, correctly propagate the return values.

This turns out to be a no-op at present.

People like to see when the emergency sync and emergency remount operations
have completed.

From: Keith Mannthey <kmannth@us.ibm.com>

The following is a patch to fix inconsistent use of the function
set_ioapic_affinity.  In the current kernel it is unclear as to weather the
value being passed to the function is a cpu mask or valid apic id.  In
irq_affinity_write_proc the kernel passes on a cpu mask but the kirqd thread
passes on logical apic ids.  In flat apic mode this is not an issue because a
cpu mask represents the apic value.  However in clustered apic mode the cpu
mask is very different from the logical apic id.

This is an attempt to do the right thing for clustered apics.  I clarify that
the value being passed to set_ioapic_affinity is a cpu mask not a apicid.
Set_ioapic_affinity will do the conversion to logical apic ids.  Since many
cpu masks don't map to valid apicids in clustered apic mode TARGET_CPUS is
used as a default value when such a situation occurs.  I think this is a good
step in making irq_affinity clustered apic safe.

From: Andrey Panin <pazke@donpac.ru>

attached patch fixes penguin with sgi framebuffer logo for
visws subarch. It was broken in 2.5.68 IIRC.

From: Mingming Cao <cmm@us.ibm.com>

Basically, freeary() is called with the spinlock for that semaphore set
hold.  But after the semaphore set is removed from the ID array by
calling sem_rmid(), there is no lock to protect the waiting queue for
that semaphore set.  So, if a waiter is woken up by a signal (not by the
wakeup from freeary()), it will check the q->status and q->prev fields.
At that moment, freeary() may not have a chance to update those fields
yet.

static void freeary (int id)
{
	.......
        sma = sem_rmid(id);

	......
        /* Wake up all pending processes and let them fail with EIDRM.*/
        for (q = sma->sem_pending; q; q = q->next) {
                q->status = -EIDRM;
                q->prev = NULL;
                wake_up_process(q->sleeper); /* doesn't sleep */
        }
        sem_unlock(sma);
	......
}

So I propose move sem_rmid() after the loop of waking up every waiters.
That could gurantee that when the waiters are woke up, the updates for
q->status and q->prev have already done.  Similar thing in message queue
case.  The patch is attached below. Comments are very welcomed.

I have tested this patch on 2.5.68 kernel with LTP tests, seems fine to
me. Paul, could you test this on DOTS test again? Thanks!

exit_mmap() currently assumes that the exitting task used virtual address
span TASK_SIZE.

But on some platforms, TASK_SIZE is variable, based on current->mm.

But exit_mmap() can be called from (say) procfs's call to mmput.  In which
case current->mm has nothing to do with the mm which is being put in
mmput().

So rather than assuming that the mm which is being put is current->mm, we
need to calculate the virtual span of the mm.  Add a new per-arch macro
MM_VM_SIZE() for that.

Some platforms can currently go BUG over this (where?).  sparc64 is safe
because our TASK_SIZE is constant.

Platforms such as ia64 should stick the VM extent inside of mm_struct, I'd
suggest adding it to mm_context_t.

1) TASK_SIZE means what is valid for mmap()'s in the processes
   address space

2) MM_VM_SIZE means where things might be mapped for a MM, including
   private implementation-specific areas created by the kernel
   which the user cannot access

From: Rusty Russell <rusty@rustcorp.com.au>

__module_get is theoretically allowed on module inside init, since we
already hold an implicit reference.  Currently this BUG()s: make the
reference count explicit, which also simplifies delete path.  Also cleans
up unload path, such that it only drops semaphore when it's actually
sleeping for rmmod --wait.

From: Jan Kara <jack@suse.cz>

I'm sending a fix which fixes potential problems (dropping references which
were not acquired) when dquot_transfer() fails.

From: Jan Kara <jack@suse.cz>

I'm sending a patch which changes numbers of blocks reserved for quota writes
to more appropriate values (with current values ext3 asserts can be
triggered).

From: Zwane Mwaikambo <zwane@linuxpower.ca>

The proc interface has no way of telling wether there is an active cpufreq
driver or not. This means that if you don't have a cpufreq supported
processor, this will oops in various possible places.

From: Zwane Mwaikambo <zwane@linuxpower.ca>

Removes an extra initialisation and general nitpicking.

From: Zwane Mwaikambo <zwane@linuxpower.ca>

Commented out printk causes change in program flow in cpufreq/p4-clockmod.c

From: "Stephen C. Tweedie" <sct@redhat.com>

When a new vma can be merged simultaneously with its two immediate
neighbours in both directions, vma_merge() extends the predecessor vma and
deletes the successor.  However, if the vma maps a file, it fails to fput()
when doing the delete, leaving the file's refcount inconsistent.

From: Bjorn Stenberg <bjorn@haxx.se>
      OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

This simple patch makes the linux fat filesystem driver use the
next_cluster field in the fat_boot_fsinfo structure.  This field is a hint
where to start looking for free clusters.

Using this field makes a big difference for disks connected over slow links
such as USB 1.1.  Finding the first free cluster on a 40gig fat-formatted
usb disk can today take several minutes.  This patch cuts it down to a
fraction of a second.

Also, commit the next_cluster search hint toand from the superblock in
write_super/fill_super.

It looks like I guessed wrong on this one.

From: Christoph Hellwig <hch@lst.de>

Whee! devfs_register isn't used anymore in the whole tree and with
it some other devfs crap.  Kill it for good.

From: Rusty Russell <rusty@rustcorp.com.au>, David Mosberger

The patch below updates the other platforms with module_arch_cleanup().
Also, I added more debug output to kernel/module.c since I found it useful
to be able to see the final section layout.

From: george anzinger <george@mvista.com>

- Fix the sig_notify filtering code for the timer_create system call to
  properly check for the signal number being small enought, but only if
  SIG_NONE is not specified.

- Eliminate useless test of sig_notify.

From: Taral <taral@taral.net>

Looks like the recent access_ok fixes broke building of i386.
__verify_write is still referenced in a couple places.

Some of the fancier slab debugging options are disabled for caches whose
objects are larger than PAGE_SIZE/8, for speed/fragmentation reasons.

But this patch turns up bugs in the size-2048 slab, so it was a bad idea.
Enable the debugging for slabs up to 4096 bytes.

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

C preprocessor expressions work, guys!)

Here are three more fixes which I missed in the previous patch.

into nuts.ninka.net:/home/davem/src/BK/net-2.5

Don't depend on undefined preprocessor symbols evaluating to zero.

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

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