btrfs: make extent unpinning more efficient when committing transaction

btrfs: make extent unpinning more efficient when committing transaction

At btrfs_finish_extent_commit() we have this loop that keeps finding an
extent range to unpin in the transaction's pinned_extents io tree, caches
the extent state and then passes that cached extent state to
btrfs_clear_extent_dirty(), which will free that extent state since we
clear the only bit it can have set. So on each loop iteration we do a
full io tree search and the cached state is used only to avoid having
a tree search done by btrfs_clear_extent_dirty().

During the lifetime of a transaction we can pin many thousands of extents,
resulting in a large and deep rb tree that backs the io tree. For example,
for the following fs_mark run on a 12 cores boxes:

  $ cat test.sh
  #!/bin/bash

  DEV=/dev/nullb0
  MNT=/mnt/nullb0
  FILES=100000
  THREADS=$(nproc --all)

  echo "performance" | \
      tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

  mkfs.btrfs -f $DEV
  mount $DEV $MNT

  OPTS="-S 0 -L 8 -n $FILES -s 0 -t $THREADS -k"
  for ((i = 1; i <= $THREADS; i++)); do
      OPTS="$OPTS -d $MNT/d$i"
  done

  fs_mark $OPTS

  umount $MNT

an histogram for the number of ranges (elements) in the pinned extents
io tree of a transaction was the following:

  Count: 76
  Range: 5440.000 - 51088.000; Mean: 27354.368; Median: 28312.000; Stddev: 9800.767
  Percentiles:  90th: 40486.000; 95th: 43322.000; 99th: 51088.000
   5440.000 -  6805.809:     1 ###
   6805.809 - 10652.034:     1 ###
  10652.034 - 13326.178:     3 ########
  13326.178 - 16671.590:     8 ######################
  16671.590 - 20856.773:     7 ####################
  20856.773 - 26092.528:    13 ####################################
  26092.528 - 32642.571:    19 #####################################################
  32642.571 - 40836.818:    17 ###############################################
  40836.818 - 51088.000:     7 ####################

We can improve on this by grabbing the next state before calling
btrfs_clear_extent_dirty(), avoiding a full tree search on the next
iteration which always has an O(log n) complexity while grabbing the next
element (rb_next() rbtree operation) is in the worst case O(log n) too,
but very often much less than that, making it more efficient.

Here follow histograms for the execution times, in nanoseconds, of
btrfs_finish_extent_commit() before and after applying this patch and all
the other patches in the same patchset.

Before patchset:

  Count: 32
  Range: 3925691.000 - 269990635.000; Mean: 133814526.906; Median: 122758052.000; Stddev: 65776550.375
  Percentiles:  90th: 228672087.000; 95th: 265187000.000; 99th: 269990635.000
    3925691.000 -   5993208.660:     1 ####
    5993208.660 -  75878537.656:     4 ##################
   75878537.656 - 115840974.514:    12 #####################################################
  115840974.514 - 176850157.761:     6 ###########################
  176850157.761 - 269990635.000:     9 ########################################

After patchset:

  Count: 32
  Range: 1849393.000 - 231491064.000; Mean: 126978584.625; Median: 123732897.000; Stddev: 58007821.806
  Percentiles:  90th: 203055491.000; 95th: 219952699.000; 99th: 231491064.000
    1849393.000 -   2997642.092:     1 ####
    2997642.092 -  88111637.071:     9 #####################################
   88111637.071 - 142818264.414:     9 #####################################
  142818264.414 - 231491064.000:    13 #####################################################

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>