Currently when dropping extent maps for a file range, through
btrfs_drop_extent_map_range(), we do the following non-optimal things:
1) We lookup for extent maps one by one, always starting the search from
the root of the extent map tree. This is not efficient if we have
multiple extent maps in the range;
2) We check on every iteration if we have the 'split' and 'split2' spare
extent maps in case we need to split an extent map that intersects our
range but also crosses its boundaries (to the left, to the right or
both cases). If our target range is for example:
[2M, 8M)
And we have 3 extents maps in the range:
[1M, 3M) [3M, 6M) [6M, 10M[
The on the first iteration we allocate two extent maps for 'split' and
'split2', and use the 'split' to split the first extent map, so after
the split we set 'split' to 'split2' and then set 'split2' to NULL.
On the second iteration, we don't need to split the second extent map,
but because 'split2' is now NULL, we allocate a new extent map for
'split2'.
On the third iteration we need to split the third extent map, so we
use the extent map pointed by 'split'.
So we ended up allocating 3 extent maps for splitting, but all we
needed was 2 extent maps. We never need to allocate more than 2,
because extent maps that need to be split are always the first one
and the last one in the target range.
Improve on this by:
1) Using rb_next() to move on to the next extent map. This results in
iterating over less nodes of the tree and it does not require comparing
the ranges of nodes to our start/end offset;
2) Allocate the 2 extent maps for splitting before entering the loop and
never allocate more than 2. In practice it's very rare to have the
combination of both extent map allocations fail, since we have a
dedicated slab for extent maps, and also have the need to split two
extent maps.
This patch is part of a patchset comprised of the following patches:
btrfs: fix missed extent on fsync after dropping extent maps
btrfs: move btrfs_drop_extent_cache() to extent_map.c
btrfs: use extent_map_end() at btrfs_drop_extent_map_range()
btrfs: use cond_resched_rwlock_write() during inode eviction
btrfs: move open coded extent map tree deletion out of inode eviction
btrfs: add helper to replace extent map range with a new extent map
btrfs: remove the refcount warning/check at free_extent_map()
btrfs: remove unnecessary extent map initializations
btrfs: assert tree is locked when clearing extent map from logging
btrfs: remove unnecessary NULL pointer checks when searching extent maps
btrfs: remove unnecessary next extent map search
btrfs: avoid pointless extent map tree search when flushing delalloc
btrfs: drop extent map range more efficiently
And the following fio test was done before and after applying the whole
patchset, on a non-debug kernel (Debian's default kernel config) on a 12
cores Intel box with 64G of ram:
$ cat test.sh
#!/bin/bash
DEV=/dev/nvme0n1
MNT=/mnt/nvme0n1
MOUNT_OPTIONS="-o ssd"
MKFS_OPTIONS="-R free-space-tree -O no-holes"
cat <<EOF > /tmp/fio-job.ini
[writers]
rw=randwrite
fsync=8
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=psync
filesize=2G
runtime=300
time_based
directory=$MNT
numjobs=8
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
Result before applying the patchset:
WRITE: bw=197MiB/s (206MB/s), 197MiB/s-197MiB/s (206MB/s-206MB/s), io=57.7GiB (61.9GB), run=300188-300188msec
Result after applying the patchset:
WRITE: bw=203MiB/s (213MB/s), 203MiB/s-203MiB/s (213MB/s-213MB/s), io=59.5GiB (63.9GB), run=300019-300019msec
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
db21370bff