r=ccooper

Fix problem where none of the tests were checking for a successfull cleanup.
Error in cleanup when modules are still loaded.

- clean up output directory creation

- ext3-mballoc* patches export ext3_new_blocks()

- make the output directory if it doesn't exist
- add a forgotten item to the key

- simplify iozone file cleanup a little
- scale each thread's io contribution based on the number of threads
- fix sloppy shell arg passing that accidentally worked

- teach the iozone test to use multiple threads
- add a -e argument to specify the obdecho module by hand
- the echo_filter test performs reads and on the right number of pages
- use more representative io sizes
- use 'getconf PAGE_SIZE' to find the page size on the machine

NOTE: subsytem initialize order problem is to be solved yet

do just as well.

r=shaver

work with --mkfsoptions of ost in lmc and lconf

- run an iostat -x on each block device during each test
- reformat a bit to make room for the iostat reports

    . NOT bump the module refcount when a new peer is found

    . Clean up all open connections and autoconnect route table entries on
      module unload.

- more carefully cleanup to avoid error messages in early cleanup

  moving to the right side
- ext3_ext_find_goal() takes into account offset of a block to be allocated
  (thanks to Andreas for smart idea)

  cache-aliasing problems
- ext3_forget() should be used in conjuction with ext3_free_blocks()

NOTE: 2.4.24 + extents + batching block allocation passes data-consistency
      tests now (at least I haven't seen problems for many hours of testing)

r=shaver

rewrite new_uuid function for lmc

up for it.
r=coop

- they back-ported PDE from 2.6, which broke our compat macro
- they changed direct_IO method to pass struct file instead of struct inode

Both are now checked in portals/archdep.m4 directly instead of as a version
number check so we don't need to worry about other releases doing the same.