summaryrefslogtreecommitdiff
path: root/fs/btrfs
diff options
context:
space:
mode:
authorFilipe Manana <fdmanana@suse.com>2020-07-22 12:29:01 +0100
committerDavid Sterba <dsterba@suse.com>2020-07-27 12:55:48 +0200
commit5e548b32018d96c377fda4bdac2bf511a448ca67 (patch)
tree19b50ad05564bf26e39b17a605376b4167723691 /fs/btrfs
parentfbc2bd7e7ab95704a96dade6a7d00ea99a68b015 (diff)
btrfs: do not set the full sync flag on the inode during page release
When removing an extent map at try_release_extent_mapping(), called through the page release callback (btrfs_releasepage()), we always set the full sync flag on the inode, which forces the next fsync to use a slower code path. This hurts performance for workloads that dirty an amount of data that exceeds or is very close to the system's RAM memory and do frequent fsync operations (like database servers can for example). In particular if there are concurrent fsyncs against different files, by falling back to a full fsync we do a lot more checksum lookups in the checksums btree, as we do it for all the extents created in the current transaction, instead of only the new ones since the last fsync. These checksums lookups not only take some time but, more importantly, they also cause contention on the checksums btree locks due to the concurrency with checksum insertions in the btree by ordered extents from other inodes. We actually don't need to set the full sync flag on the inode, because we only remove extent maps that are in the list of modified extents if they were created in a past transaction, in which case an fsync skips them as it's pointless to log them. So stop setting the full fsync flag on the inode whenever we remove an extent map. This patch is part of a patchset that consists of 3 patches, which have the following subjects: 1/3 btrfs: fix race between page release and a fast fsync 2/3 btrfs: release old extent maps during page release 3/3 btrfs: do not set the full sync flag on the inode during page release Performance tests were ran against a branch (misc-next) containing the whole patchset. The test exercises a workload where there are multiple processes writing to files and fsyncing them (each writing and fsyncing its own file), and in total the amount of data dirtied ranges from 2x to 4x the system's RAM memory (16GiB), so that the page release callback is invoked frequently. The following script, using fio, was used to perform the tests: $ cat test-fsync.sh #!/bin/bash DEV=/dev/sdk MNT=/mnt/sdk MOUNT_OPTIONS="-o ssd" MKFS_OPTIONS="-d single -m single" if [ $# -ne 3 ]; then echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ" exit 1 fi NUM_JOBS=$1 FILE_SIZE=$2 FSYNC_FREQ=$3 cat <<EOF > /tmp/fio-job.ini [writers] rw=write fsync=$FSYNC_FREQ fallocate=none group_reporting=1 direct=0 bs=64k ioengine=sync size=$FILE_SIZE directory=$MNT numjobs=$NUM_JOBS thread EOF echo "Using config:" echo cat /tmp/fio-job.ini echo mkfs.btrfs -f $MKFS_OPTIONS $DEV &> /dev/null mount $MOUNT_OPTIONS $DEV $MNT fio /tmp/fio-job.ini umount $MNT The tests were performed for different numbers of jobs, file sizes and fsync frequency. A qemu VM using kvm was used, with 8 cores (the host has 12 cores, with cpu governance set to performance mode on all cores), 16GiB of ram (the host has 64GiB) and using a NVMe device directly (without an intermediary filesystem in the host). While running the tests, the host was not used for anything else, to avoid disturbing the tests. The obtained results were the following, and the last line printed by fio is pasted (includes aggregated throughput and test run time). ***************************************************** **** 1 job, 32GiB file, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=29.1MiB/s (30.5MB/s), 29.1MiB/s-29.1MiB/s (30.5MB/s-30.5MB/s), io=32.0GiB (34.4GB), run=1127557-1127557msec After patchset: WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=32.0GiB (34.4GB), run=1119042-1119042msec (+0.7% throughput, -0.8% run time) ***************************************************** **** 2 jobs, 16GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=33.5MiB/s (35.1MB/s), 33.5MiB/s-33.5MiB/s (35.1MB/s-35.1MB/s), io=32.0GiB (34.4GB), run=979000-979000msec After patchset: WRITE: bw=39.9MiB/s (41.8MB/s), 39.9MiB/s-39.9MiB/s (41.8MB/s-41.8MB/s), io=32.0GiB (34.4GB), run=821283-821283msec (+19.1% throughput, -16.1% runtime) ***************************************************** **** 4 jobs, 8GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=52.1MiB/s (54.6MB/s), 52.1MiB/s-52.1MiB/s (54.6MB/s-54.6MB/s), io=32.0GiB (34.4GB), run=629130-629130msec After patchset: WRITE: bw=71.8MiB/s (75.3MB/s), 71.8MiB/s-71.8MiB/s (75.3MB/s-75.3MB/s), io=32.0GiB (34.4GB), run=456357-456357msec (+37.8% throughput, -27.5% runtime) ***************************************************** **** 8 jobs, 4GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=76.1MiB/s (79.8MB/s), 76.1MiB/s-76.1MiB/s (79.8MB/s-79.8MB/s), io=32.0GiB (34.4GB), run=430708-430708msec After patchset: WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=245458-245458msec (+74.7% throughput, -43.0% run time) ***************************************************** **** 16 jobs, 2GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=74.7MiB/s (78.3MB/s), 74.7MiB/s-74.7MiB/s (78.3MB/s-78.3MB/s), io=32.0GiB (34.4GB), run=438625-438625msec After patchset: WRITE: bw=184MiB/s (193MB/s), 184MiB/s-184MiB/s (193MB/s-193MB/s), io=32.0GiB (34.4GB), run=177864-177864msec (+146.3% throughput, -59.5% run time) ***************************************************** **** 32 jobs, 2GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=72.6MiB/s (76.1MB/s), 72.6MiB/s-72.6MiB/s (76.1MB/s-76.1MB/s), io=64.0GiB (68.7GB), run=902615-902615msec After patchset: WRITE: bw=227MiB/s (238MB/s), 227MiB/s-227MiB/s (238MB/s-238MB/s), io=64.0GiB (68.7GB), run=288936-288936msec (+212.7% throughput, -68.0% run time) ***************************************************** **** 64 jobs, 1GiB files, fsync frequency 1 **** ***************************************************** Before patchset: WRITE: bw=98.8MiB/s (104MB/s), 98.8MiB/s-98.8MiB/s (104MB/s-104MB/s), io=64.0GiB (68.7GB), run=663126-663126msec After patchset: WRITE: bw=294MiB/s (308MB/s), 294MiB/s-294MiB/s (308MB/s-308MB/s), io=64.0GiB (68.7GB), run=222940-222940msec (+197.6% throughput, -66.4% run time) Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
Diffstat (limited to 'fs/btrfs')
-rw-r--r--fs/btrfs/extent_io.c10
1 files changed, 8 insertions, 2 deletions
diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index d619babede0f..617ea38e6fd7 100644
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -4524,8 +4524,14 @@ int try_release_extent_mapping(struct page *page, gfp_t mask)
if (em->generation >= cur_gen)
goto next;
remove_em:
- set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
- &btrfs_inode->runtime_flags);
+ /*
+ * We only remove extent maps that are not in the list of
+ * modified extents or that are in the list but with a
+ * generation lower then the current generation, so there
+ * is no need to set the full fsync flag on the inode (it
+ * hurts the fsync performance for workloads with a data
+ * size that exceeds or is close to the system's memory).
+ */
remove_extent_mapping(map, em);
/* once for the rb tree */
free_extent_map(em);