Currently we index free space on offset only, because usually we have a
hint from the allocator that we want to honor for locality reasons.
However if we fail to use this hint we have to go back to a brute force
search through the free space entries to find a large enough extent.

With sufficiently fragmented free space this becomes quite expensive, as
we have to linearly search all of the free space entries to find if we
have a part that's long enough.

To fix this add a cached rb tree to index based on free space entry
bytes.  This will allow us to quickly look up the largest chunk in the
free space tree for this block group, and stop searching once we've
found an entry that is too small to satisfy our allocation.  We simply
choose to use this tree if we're searching from the beginning of the
block group, as we know we do not care about locality at that point.

I wrote an allocator test that creates a 10TiB ram backed null block
device and then fallocates random files until the file system is full.
I think go through and delete all of the odd files.  Then I spawn 8
threads that fallocate 64MiB files (1/2 our extent size cap) until the
file system is full again.  I use bcc's funclatency to measure the
latency of find_free_extent.  The baseline results are

     nsecs               : count     distribution
         0 -> 1          : 0        |                                        |
         2 -> 3          : 0        |                                        |
         4 -> 7          : 0        |                                        |
         8 -> 15         : 0        |                                        |
        16 -> 31         : 0        |                                        |
        32 -> 63         : 0        |                                        |
        64 -> 127        : 0        |                                        |
       128 -> 255        : 0        |                                        |
       256 -> 511        : 10356    |****                                    |
       512 -> 1023       : 58242    |*************************               |
      1024 -> 2047       : 74418    |********************************        |
      2048 -> 4095       : 90393    |****************************************|
      4096 -> 8191       : 79119    |***********************************     |
      8192 -> 16383      : 35614    |***************                         |
     16384 -> 32767      : 13418    |*****                                   |
     32768 -> 65535      : 12811    |*****                                   |
     65536 -> 131071     : 17090    |*******                                 |
    131072 -> 262143     : 26465    |***********                             |
    262144 -> 524287     : 40179    |*****************                       |
    524288 -> 1048575    : 55469    |************************                |
   1048576 -> 2097151    : 48807    |*********************                   |
   2097152 -> 4194303    : 26744    |***********                             |
   4194304 -> 8388607    : 35351    |***************                         |
   8388608 -> 16777215   : 13918    |******                                  |
  16777216 -> 33554431   : 21       |                                        |

avg = 908079 nsecs, total: 580889071441 nsecs, count: 639690

And the patch results are

     nsecs               : count     distribution
         0 -> 1          : 0        |                                        |
         2 -> 3          : 0        |                                        |
         4 -> 7          : 0        |                                        |
         8 -> 15         : 0        |                                        |
        16 -> 31         : 0        |                                        |
        32 -> 63         : 0        |                                        |
        64 -> 127        : 0        |                                        |
       128 -> 255        : 0        |                                        |
       256 -> 511        : 6883     |**                                      |
       512 -> 1023       : 54346    |*********************                   |
      1024 -> 2047       : 79170    |********************************        |
      2048 -> 4095       : 98890    |****************************************|
      4096 -> 8191       : 81911    |*********************************       |
      8192 -> 16383      : 27075    |**********                              |
     16384 -> 32767      : 14668    |*****                                   |
     32768 -> 65535      : 13251    |*****                                   |
     65536 -> 131071     : 15340    |******                                  |
    131072 -> 262143     : 26715    |**********                              |
    262144 -> 524287     : 43274    |*****************                       |
    524288 -> 1048575    : 53870    |*********************                   |
   1048576 -> 2097151    : 55368    |**********************                  |
   2097152 -> 4194303    : 41036    |****************                        |
   4194304 -> 8388607    : 24927    |**********                              |
   8388608 -> 16777215   : 33       |                                        |
  16777216 -> 33554431   : 9        |                                        |

avg = 623599 nsecs, total: 397259314759 nsecs, count: 637042

There's a little variation in the amount of calls done because of timing
of the threads with metadata requirements, but the avg, total, and
count's are relatively consistent between runs (usually within 2-5% of
each other).  As you can see here we have around a 30% decrease in
average latency with a 30% decrease in overall time spent in
find_free_extent.
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>