Make a bit more effort to clean up test cruft.

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

r = braam

make portals_debug_dumpstack(void) calling panic("LBUG") in uml

r=shaver

create true uuid in lmc

with lots of clients.  Tested at LLNL for a long time already.
Reduce max per-OSC RPCs in flight to reduce OST file fragmentation with
out-of-order writes to the filesystem.
r=Zach

b: 2332
r: Shaver

Only create config logs for the MDS on the local node.

Use kmalloc() instead of slab to allocate the MD and MSG portal descriptors.
On x86_64 the MD and MSG structs got quite large, into the 2k range, though
pages are still 4k there.  This tricked slab into trying to back their slabs
with multi-page allocations.  These allocations failed under high load which
lead to LBUG()s and timeouts.

While we're in here, properly initialize the in_use counts and stop using
sleeping allocations while holding the state lock.

  now run in a local, single node mode by default.

- Add some tests for old bugs to recovery-small.sh. These tests were
  in the ltest/recovery scripts, but that is now unused.

- Move the OBD_FAIL funcs to test-framework.

- add the -v command line option for test scripts; it will print out
  the commands its executing on the nodes.

r=shaver

Prevent a race between mds_client_add and mds_client_free by clearing the client
index bit after the data has been zeroed. Also make config messages consistent
(with each other and with the obdfilter messages).

r=Adilger

Log:
    remove the '\n's not at the end of log lines

Approved by Andreas Dilger.

that less log space is wasted, and it's easier to visually scan.

Jacob reported that when MDS/OST recovery requires new objects to be
created, the OST throws an assertion.

Bug 2425 remains open to track the creation of many more tests for
missing MDS/OST recovery cases.

r=shaver

Delete IOC_CONNECT,DISCONNECT and use obd_self_export instead
of creating connections for lctl. Also delete the IOC_DEVICE comamnd
and make the ioctl interface stateless.  The lctl probe command is now
a noop, and lctl device is still used to set the device, although the
current device state is only saved in lctl now, and not the kernel.

        grab the dchild directly and sample the data. Fixes recovery-small.sh.
r=phik,buffalo

Stop taking a PR lock in mds_readpage; a PR is already held by the
client, so if there is a PW in the queue, deadlock will result.  Just
assume that the client has a lock.

r=adilger
(didn't see regressions in buffalo, confirmed read throughput increases
with sf and fpp multi-node IOR)

This cleans up llite's readpage path and implements our own read-ahead window
that hangs off of ll_file_data.  The broad goal is to keep a fair amount of
read-ahead pages issued and queued which can be fired off into read rpcs as
read-ahead rpcs are completed.

  of their stay in the page cache.  This lets us display the contents
  of the page cache via llite/*/dump_pgcache file.  This was done as part
  of b=2252 and is being committed seperately from the read-ahead work.

because we exit early on error, so the mds_destroy_mfd() is bogus.  I left
RETURN(rc) in case things change in the future though.

We don't use request_body() anywhere inside mds_put_write_access(), but
since all of that code is just commented out I didn't do a real cleanup.
Just a bogus compiler warning fixed.

- export brw histograms from the filter that record discontiguous offsets
  in the brw request and discontigous blocks that satisfy the request
  (seen as /proc/fs/lustre/obdfilter/$name/brw_stats)

Fix for bug 974, Also adds a test to check for OOM (modified script from
bug 1135), fixes to sanity.sh's test 45 to obtain a grant (closes 2387).