btrfs: prepare extents to be logged before locking a log tree path

btrfs: prepare extents to be logged before locking a log tree path

When we want to log an extent, in the fast fsync path, we obtain a path
to the leaf that will hold the file extent item either through a deletion
search, via btrfs_drop_extents(), or through an insertion search using
btrfs_insert_empty_item(). After that we fill the file extent item's
fields one by one directly on the leaf.

Instead of doing that, we could prepare the file extent item before
obtaining a btree path, and then copy the prepared extent item with a
single operation once we get the path. This helps avoid some contention
on the log tree, since we are holding write locks for longer than
necessary, especially in the case where the path is obtained via
btrfs_drop_extents() through a deletion search, which always keeps a
write lock on the nodes at levels 1 and 2 (besides the leaf).

This change does that, we prepare the file extent item that is going to
be inserted before acquiring a path, and then copy it into a leaf using
a single copy operation once we get a path.

This change if part of a patchset that is comprised of the following
patches:

  1/6 btrfs: remove unnecessary leaf free space checks when pushing items
  2/6 btrfs: avoid unnecessary COW of leaves when deleting items from a leaf
  3/6 btrfs: avoid unnecessary computation when deleting items from a leaf
  4/6 btrfs: remove constraint on number of visited leaves when replacing extents
  5/6 btrfs: remove useless path release in the fast fsync path
  6/6 btrfs: prepare extents to be logged before locking a log tree path

The following test was run to measure the impact of the whole patchset:

  $ cat test.sh
  #!/bin/bash

  DEV=/dev/sdi
  MNT=/mnt/sdi
  MOUNT_OPTIONS="-o ssd"
  MKFS_OPTIONS="-R free-space-tree -O no-holes"

  NUM_JOBS=8
  FILE_SIZE=128M
  RUN_TIME=200

  cat <<EOF > /tmp/fio-job.ini
  [writers]
  rw=randwrite
  fsync=1
  fallocate=none
  group_reporting=1
  direct=0
  bssplit=4k/20:8k/20:16k/20:32k/10:64k/10:128k/5:256k/5:512k/5:1m/5
  ioengine=sync
  filesize=$FILE_SIZE
  runtime=$RUN_TIME
  time_based
  directory=$MNT
  numjobs=$NUM_JOBS
  thread
  EOF

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

  echo
  echo "Using config:"
  echo
  cat /tmp/fio-job.ini
  echo

  umount $MNT &> /dev/null
  mkfs.btrfs -f $MKFS_OPTIONS $DEV
  mount $MOUNT_OPTIONS $DEV $MNT

  fio /tmp/fio-job.ini

  umount $MNT

The test ran inside a VM (8 cores, 32G of RAM) with the target disk
mapping to a raw NVMe device, and using a non-debug kernel config
(Debian's default config).

Before the patchset:

WRITE: bw=116MiB/s (122MB/s), 116MiB/s-116MiB/s (122MB/s-122MB/s), io=22.7GiB (24.4GB), run=200013-200013msec

After the patchset:

WRITE: bw=125MiB/s (131MB/s), 125MiB/s-125MiB/s (131MB/s-131MB/s), io=24.3GiB (26.1GB), run=200007-200007msec

A 7.8% gain on throughput and +7.0% more IO done in the same period of
time (200 seconds).

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