direct-io: only inc/dec inode->i_dio_count for file systems
do_blockdev_direct_IO() increments and decrements the inode
->i_dio_count for each IO operation. It does this to protect against
truncate of a file. Block devices don't need this sort of protection.
For a capable multiqueue setup, this atomic int is the only shared
state between applications accessing the device for O_DIRECT, and it
presents a scaling wall for that. In my testing, as much as 30% of
system time is spent incrementing and decrementing this value. A mixed
read/write workload improved from ~2.5M IOPS to ~9.6M IOPS, with
better latencies too. Before:
clat percentiles (usec):
| 1.00th=[ 33], 5.00th=[ 34], 10.00th=[ 34], 20.00th=[ 34],
| 30.00th=[ 34], 40.00th=[ 34], 50.00th=[ 35], 60.00th=[ 35],
| 70.00th=[ 35], 80.00th=[ 35], 90.00th=[ 37], 95.00th=[ 80],
| 99.00th=[ 98], 99.50th=[ 151], 99.90th=[ 155], 99.95th=[ 155],
| 99.99th=[ 165]
After:
clat percentiles (usec):
| 1.00th=[ 95], 5.00th=[ 108], 10.00th=[ 129], 20.00th=[ 149],
| 30.00th=[ 155], 40.00th=[ 161], 50.00th=[ 167], 60.00th=[ 171],
| 70.00th=[ 177], 80.00th=[ 185], 90.00th=[ 201], 95.00th=[ 270],
| 99.00th=[ 390], 99.50th=[ 398], 99.90th=[ 418], 99.95th=[ 422],
| 99.99th=[ 438]
In other setups, Robert Elliott reported seeing good performance
improvements:
https://lkml.org/lkml/2015/4/3/557
The more applications accessing the device, the worse it gets.
Add a new direct-io flags, DIO_SKIP_DIO_COUNT, which tells
do_blockdev_direct_IO() that it need not worry about incrementing
or decrementing the inode i_dio_count for this caller.
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Elliott, Robert (Server Storage) <elliott@hp.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jens Axboe <axboe@fb.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff --git a/fs/ext4/indirect.c b/fs/ext4/indirect.c
index 3580629e..9588240 100644
--- a/fs/ext4/indirect.c
+++ b/fs/ext4/indirect.c
@@ -682,11 +682,11 @@
* via ext4_inode_block_unlocked_dio(). Check inode's state
* while holding extra i_dio_count ref.
*/
- atomic_inc(&inode->i_dio_count);
+ inode_dio_begin(inode);
smp_mb();
if (unlikely(ext4_test_inode_state(inode,
EXT4_STATE_DIOREAD_LOCK))) {
- inode_dio_done(inode);
+ inode_dio_end(inode);
goto locked;
}
if (IS_DAX(inode))
@@ -697,7 +697,7 @@
inode->i_sb->s_bdev, iter,
offset, ext4_get_block, NULL,
NULL, 0);
- inode_dio_done(inode);
+ inode_dio_end(inode);
} else {
locked:
if (IS_DAX(inode))
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 063052e..bccec41 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -2977,7 +2977,7 @@
* overwrite DIO as i_dio_count needs to be incremented under i_mutex.
*/
if (iov_iter_rw(iter) == WRITE)
- atomic_inc(&inode->i_dio_count);
+ inode_dio_begin(inode);
/* If we do a overwrite dio, i_mutex locking can be released */
overwrite = *((int *)iocb->private);
@@ -3079,7 +3079,7 @@
retake_lock:
if (iov_iter_rw(iter) == WRITE)
- inode_dio_done(inode);
+ inode_dio_end(inode);
/* take i_mutex locking again if we do a ovewrite dio */
if (overwrite) {
up_read(&EXT4_I(inode)->i_data_sem);