Commit 1ae399382512 ("Btrfs: do not use async submit for small DIO io's")
benefits small DIO io's.
However, if we're owning the SINGLE profile, this also affects large DIO io's
since in that case, map_length is (chunk_length - bio's offset_in_chunk),
it's farily large so that it's very likely to be larger than a large bio's
size, which avoids asynchronous submit.
For instance, if we have a 512k bio, the efforts of calculating (512k/4k=128)
checksums will be taken by the DIO task.
Test results with fio (tested on a hard disk, not tested on ssd, 4cpu, 8g memory)
bs async sync async sync
bw bw(KB/S) iops iop
4k 115312 115480 28827.6 28869.6
8k 114381 115586 14297.4 14447.6
16k 115393 116290 7211.4 7267.6
32k 114268 116589 3570.4 3643
64k 115421 113417 1803 1771.8 <-----ASYNC wins here
128k 115545 112585 902 879
256k 115178 111521 449.2 435
512k 115874 111620 226 217.6
This adds a limit 'BTRFS_STRIPE_LEN(64k)' to decide if it's small enough to avoid
asynchronous submit.
Still, in this case we don't need to split the bio and can submit it directly.
Signed-off-by: Liu Bo <bo.li.liu@xxxxxxxxxx>
---
The job in the test,
[global]
rw=write
ioengine=libaio
direct=1
iodepth=64
iodepth_batch=64
iodepth_batch_complete=64
iodepth_low=64
bs=4k
size=8g
sync=0
group_reporting
fallocate=posix
invalidate=1
runtime=30
[dio]
directory=/mnt/btrfs
filename=foobar
fs/btrfs/inode.c | 16 ++++++++++++++--
1 file changed, 14 insertions(+), 2 deletions(-)
diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index e687bb0..c640d7e 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -7792,6 +7792,7 @@ static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
int nr_pages = 0;
int ret;
int async_submit = 0;
+ u64 alloc_profile;
map_length = orig_bio->bi_iter.bi_size;
ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
@@ -7799,15 +7800,26 @@ static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
if (ret)
return -EIO;
+ alloc_profile = btrfs_get_alloc_profile(root, 1);
+
if (map_length >= orig_bio->bi_iter.bi_size) {
bio = orig_bio;
dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
+
+ /*
+ * In the case of 'single' profile, the above check is very
+ * likely to be true as map_length is (chunk_length - offset),
+ * so checking BTRFS_STRIPE_LEN here.
+ */
+ if ((alloc_profile & BTRFS_BLOCK_GROUP_PROFILE_MASK) == 0 &&
+ orig_bio->bi_iter.bi_size >= BTRFS_STRIPE_LEN)
+ async_submit = 1;
+
goto submit;
}
/* async crcs make it difficult to collect full stripe writes. */
- if (btrfs_get_alloc_profile(root, 1) &
- (BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6))
+ if (alloc_profile & (BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6))
async_submit = 0;
else
async_submit = 1;
--
1.8.1.4
--
To unsubscribe from this list: send the line "unsubscribe fstests" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
