From: Neil Brown <neilb@cse.unsw.edu.au>

If a single readdir entry needs to be split over two pages in the reply, we
first encode it into a new page, and then copy the bits into place.  When
we do this relocation, we have to modify the "offset" pointer to be either
in the first or the second page, as appropriate.

If the pointer should be at the start of the second page, it is currently
put past the end of the first page.

Note that as the offset and whole response is known to be 4byte-aligned,
the offset pointer will never be split over two pages.

From: David Mosberger <davidm@napali.hpl.hp.com>

These aren't syscalls, so rename them.  And make them static.

We keep on getting BUG()s from isofs_read_super() because it passes an insane
blocksize to bread().  See http://bugme.osdl.org/show_bug.cgi?id=2735 for
example.

I don't know what's up with isofs, but going BUG in there seems a bit rude.
Change it to drop a bunch of diagnostics and a backtrace then return a null
bh*.

Most callers of getblk() don't expect it to fail, so they'll oops anyway.  But
isofs does actually check for a NULL return.  This way, the machine stays up
and we get better debug diagnostics.

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

kallsyms contains only function names, but some debuggers (eg.  xmon on
PPC/PPC64) use it to lookup symbols: it'd be much nicer if it included data
symbols too.

Give appropriate and separate backing_dev_info's to both the ramdisk blockdev
inode and to the files which live atop the ramdisk.

Everything works now.

Implement an empty ->writepages() so that attempts to write back ramdisk pages
have less work to do.

When a filesystem does getblk() to get a buffer_head against the ramdisk the
VFS will allocate a new not-uptodate pagecache page and will attach buffers to
it.

The filesystem will then bring certain buffer_heads uptodate.  But not the
whole page.

Later, various ramdisk a_ops see the not-uptodate page and wipe the whole
thing out, including the parts to which the filesystem wrote!

Fix that up by only zapping those parts of the page which are covered by
non-uptodate buffers.

We don't actualy need to kmap the blockdev inode's pages at all, because
they're ~__GFP_HIGHMEM.  But it's future-safe, and cheap.

There's a race: one CPU writes a 1k block into a ramdisk page which isn't in
the blockdev pagecache yet.  It memsets the locked page to zeroes.

While this is happening, another CPU comes in and tries to write a different
1k block to the "disk".  But it doesn't lock the page so it races with the
memset and can have its data scribbled over.

Fix this up by locking the page even if it already existed in pagecache.

Locking a pagecache page in a make_request_fn sounds deadlocky but it is not,
because:

a) ramdisk_writepage() does nothing but a set_bit(), and cannot recur onto
   the same page.

b) Any higher-level code which holds a page lock is supposed to be
   allocating its memory with GFP_NOFS, and in 2.6 kernels that's equivalent
   to GFP_NOIO.

   (The distinction between GFP_NOIO and GFP_NOFS basically disappeared
   with the buffer_head LRU, although it was reused for writes to swap).

Allocating pagecache pages within the disk request_fn is deadlocky and prone
to page allocation failures, causing write I/O errors.

Attempt to improve things by fiddling with gfp masks.

- Remove the ramdisk special-case in fs-writeback.c - it will soon be
  unneeded.

- Fix rd_ioctl() to return -ENOTTY on invalid ioctl types, not -EINVAL.

- Make ramdisk Kconfig friendlier.

From: Paul Mackerras <paulus@samba.org>

We declare enter_rtas with a struct rtas_args * argument, though it is
supposed to be a physical address, and then every time we call it we cast the
unsigned long result from __pa() to a void *.  This patch changes the
declaration of enter_rtas to make it take an unsigned long argument, and
removes the cast from all the callers.  The actual enter_rtas() routine is in
assembler and doesn't need to be changed.

From: David Gibson <david@gibson.dropbear.id.au>

The ppc64 head.S contains an enable_32b_mode function which is used
nowhere.  This patch removes it.

From: Paul Mackerras <paulus@samba.org>

This patch fixes a whole pile of problems in the xmon kernel debugger for
ppc64.  This basically makes xmon SMP-safe.  Now, when we enter xmon it
sends an IPI to the other CPUs to get them into xmon too.  It also changes
the way we do single-stepping and breakpoints so that we don't have to
remove a breakpoint to proceed from it (instead we either emulate the
instruction where the breakpoint was, or execute it out of place).  With
this patch, if we get an exception inside xmon, it will just return to the
xmon command loop instead of hanging the system as at present.

The patch is quite large because it updates the disassembler to the latest
version from binutils (trimmed a bit), which is why I didn't cc lkml.

From: Paul Mackerras <paulus@samba.org>

When I added the out-of-line spinlocks on PPC64, I inadvertently introduced
a bug in the inline version of _raw_spin_trylock, where it returns the
opposite of what it should return.  The patch below fixes it.

From: Benjamin Herrenschmidt <benh@kernel.crashing.org>

This fixes busted asm constraints for readq & writeq implementation on
ppc64 that resulted in garbage beeing generated for writeq (plus an obvious
mistake in the prototype).

From: Anton Blanchard <anton@samba.org>

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

When hotplug cpu isn't enabled, cpu_is_offline is always false.  I had a stuck
cpu at boot that resulted in a lockup because we tried to start a migration
thread on it.  Instead of cpu_is_offline we can use !cpu_online which should
cover both the hotplug cpu enabled and disabled cases.

From: Bartlomiej Zolnierkiewicz <B.Zolnierkiewicz@elka.pw.edu.pl>

From: Rene Herman <rene.herman@keyaccess.nl>

Prevent the disks from spinning down across a reboot.

into ppc970.osdl.org:/home/torvalds/v2.6/linux

Remove a local 1k array.

nfs/read.c: dynamically allocate the big structs

Patch from akpm

    gss context is (duh!) in seconds, not jiffies!

Patch by Bruce Fields

    unwrap procedures; otherwise they're going to
    be ridiculously long after we add privacy support.

Patch by Bruce Fields

    are defined.

Patch by Bruce Fields

    Clean up the interface to the GSSAPI code.
                                                                                
Patch by Bruce Fields

    generally useful. This will help prepare for the spkm3
    and lipkey mechanisms.

Patch by Bruce Fields

     makes existing ones consistent with other
     trace messages in the RPC client.
                                                                                
Patch by Chuck Lever

    when mounting an NFS filesystem.

RPC: Make "major" timeouts be of fixed length "timeo<<retrans"
     rather than counting the number of retransmissions. The
     clock starts at the first attempt to send each request.
                                                                                
RPC: Ensure that we "slow start" the RTT estimation after a
     major timeout has occurred.

    no longer calls the generic VFS read and write routines.
    This allows all application read requests to pass through
    to the server, instead of just the ones that appear to be
    inside the file. this eliminates the requirement to use a
    GETATTR operation before each read or write to determine
    where the EOF is. This is a significant performance and
    scalability win.

    It also removes all requirements for holding the inode
    semaphore during NFS direct reads and writes, as the read
    and write logic no longer needs atomic access to the size
    of the file. this also helps client CPU scalability by
    reducing the serialization of writes against a single file.
                                                                                
    Patch by Chuck Lever

    where NFS_I(inode)->data_updates can get out of sync
    with reality, which will lead to a BUG() in nfs_clear_inode
    later on.

    Patch by Olaf Kirch.

     seconds in order to avoid flooding of servers.

into ppc970.osdl.org:/home/torvalds/v2.6/linux

into austin.ibm.com:/shaggy/bk/jfs-2.5

In several functions, d_instantiate is called before the transaction
is committed.  Under the rare condition that txCommit fails, the new
inode is released, but the dentry continues to point to it.  This
can lead to a seg fault.  The fix is to call d_instantiate after
txCommit has run successfully.

into ppc970.osdl.org:/home/torvalds/v2.6/linux

The jfs_log structure was never being freed at unmount time.

The NFS readlink() methods all take a buffer length argument.  Use that
instead of assuming PAGE_SIZE... 

We need to return ENAMETOOLONG rather than EIO.