land b_size on HEAD for 1.2

Adds support for client-side write cache accounting via OST write credits
(grants).  There is not yet a "correction" algorithm for grants if they
get out of sync between the client and OST (OST can handle clients who think
they have grant but do not).  Peter had previously suggested that grants
should decay over time (faster on the clients than on the OSTs) so idle
nodes do not consume space they aren't using.

Use /proc/fs/lustre/osc/OSC*/cur_grant_bytes (and existing cur_dirty_bytes)
to check client-side cache values.

Includes lustre-side changes for lfsck support (create files with
specific objids, create specific objids on OSTs).

Also includes some changes to the testing scripts (local.sh, lov.sh, acc-sm)
so that they include a "client nid '*'" client and can be used for sanityN.sh
testing.  This has been put into the main acceptance-small.sh loop so that
sanityN.sh will be run on both single and multi-OST setups.

b=974 b=2349
r=shaver

b=2603, b=2604

b=2443

b=2302

b=2642

b=2428

b=2648
r=phil

b=2636

page dirty in any lists (from 1.0.4).
b=2366

b=1792, b=2517

b=2370

b=2476
r=phil

be evicted if lock is revoked.
b=2451
r=phik

Break cyclic locking deadlock by dropping the MDC read lock before we
take the OSC read lock during getattr intents

Fix i_sem/journal inversion in mds_client_add, which was never updated
when we decided to re-order these a few months ago.  This became much
easier to hit after we fixed bug 2306.

Be more careful about the return codes from obd_match, lest we try to
cancel a lock which was never granted.

r=shaver
Print a much more meaningful error when a client is rejected because a
service node is waiting for recoverable clients.

r=shaver
This bug happens when a file is opened twice for write, then both close it
at the same time.  If they both drop the writecount, then race to
compare it against 0, one will free the fsdata and the other will assert.

This looks like a big patch, but it's mostly plumbing.  I had to do some
different argument passing, in order to keep everything protected under the
same lock.

I removed the writecount spinlock, and use the epoch semaphore for all three
things: management of the epoch, protection of the writecount, and atomicity of
writecount modifications which result in allocation or freeing of the fsdata.

In ldlm_process_{plain,extent}_lock, we used to remove and re-add the
lock to the waiting list after every -ERESTART loop.  But because of
the logic in the ldlm_*_compat_queue functions, in a very rare case,
this could lead to lock re-ordering and subsequent deadlock.

Andreas's patch to fix MDS lock inversions in getattr/reint paths.
I'm giving it one more day to bake on ALC before I commit to the 1.0.x
branch.

We don't need to re-check the atime/mtime set flags in fsfilt_ext3_setattr,
we do it already in mds_fix_attr().
b=56

 - dozens and dozens of fixes for working with 2.6 kernels
 - new 2.4 kernel APIs
 - uncountable bug fixes

- remove extN/ files

- client-side I/O cache
- O_DIRECT fixes for IA64
- liblustre improvements
- various small bug fixes