When we insert the delayed items of an inode, which corresponds to the
directory index keys for a directory (key type BTRFS_DIR_INDEX_KEY), we
do the following:

1) Pick the first delayed item from the rbtree and insert it into the
   fs/subvolume btree, using btrfs_insert_empty_item() for that;

2) Without releasing the path returned by btrfs_insert_empty_item(),
   keep collecting as many consecutive delayed items from the rbtree
   as possible, as long as each one's BTRFS_DIR_INDEX_KEY key is the
   immediate successor of the previously picked item and as long as
   they fit in the available space of the leaf the path points to;

3) Then insert all the collected items into the leaf;

4) Release the reserve metadata space for each collected item and
   release each item (implies deleting from the rbtree);

5) Unlock the path.

While this is much better than inserting items one by one, it can be
improved in a few aspects:

1) Instead of adding items based on the remaining free space of the
   leaf, collect as many items that can fit in a leaf and bulk insert
   them. This results in less and larger batches, reducing the total
   amount of time to insert the delayed items. For example when adding
   100K files to a directory, we ended up creating 1658 batches with
   very variable sizes ranging from 1 item to 118 items, on a filesystem
   with a node/leaf size of 16K. After this change, we end up with 839
   batches, with the vast majority of them having exactly 120 items;

2) We do the search for more items to batch, by iterating the rbtree,
   while holding a write lock on the leaf;

3) While still holding the leaf locked, we are releasing the reserved
   metadata for each item and then deleting each item, keeping a write
   lock on the leaf for longer than necessary. Releasing the delayed items
   one by one can take a significant amount of time, because deleting
   them from the rbtree can often be a bit slow when the deletion results
   in rebalancing the rbtree.

So change this so that we try to create larger batches, with a total
item size up to the maximum a leaf can support, and by unlocking the leaf
immediately after inserting the items, releasing the reserved metadata
space of each item and releasing each item without holding the write lock
on the leaf.

The following script that runs fs_mark was used to test this change:

  $ cat test.sh
  #!/bin/bash

  DEV=/dev/nvme0n1
  MNT=/mnt/nvme0n1
  MOUNT_OPTIONS="-o ssd"
  MKFS_OPTIONS="-m single -d single"
  FILES=1000000
  THREADS=16
  FILE_SIZE=0

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

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

  OPTS="-S 0 -L 5 -n $FILES -s $FILE_SIZE -t 16"
  for ((i = 1; i <= $THREADS; i++)); do
      OPTS="$OPTS -d $MNT/d$i"
  done

  fs_mark $OPTS

  umount $MNT

It was run on machine with 12 cores, 64G of ram, using a NVMe device and
using a non-debug kernel config (Debian's default config).

Results before this change:

FSUse%        Count         Size    Files/sec         App Overhead
     1     16000000            0      76182.1             72223046
     3     32000000            0      62746.9             80776528
     5     48000000            0      77029.0             93022381
     6     64000000            0      73691.6             95251075
     8     80000000            0      66288.0             85089634

Results after this change:

FSUse%        Count         Size    Files/sec         App Overhead
     1     16000000            0      79049.5 (+3.7%)     69700824
     3     32000000            0      65248.9 (+3.9%)     80583693
     5     48000000            0      77991.4 (+1.2%)     90040908
     6     64000000            0      75096.8 (+1.9%)     89862241
     8     80000000            0      66926.8 (+1.0%)     84429169
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>