While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.

As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.

Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).

For example, if we have the following in the cache:

* extent entry representing free space range: [128Mb - 256Kb, 128Mb[

* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
  representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
  space in this 128Mb area is marked as free

An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.

Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.

A test was added to the sanity tests that explains in detail the issue too.

Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.

Test: compilebench -D /mnt -i 30 -r 1000 --makej

Before this change:

   intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
   compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
   read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
   delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)

After this change:

   intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
   compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
   read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
   delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>