This adds the async buffered write support to ext4,
the following is the relevant test data.
iodepth | 1 | 2 | 4 | 8 | 16 | 32 |
before(M/s) | 842 | 1047 | 1098 | 1167 | 1172 | 1247 |
after(M/s) | 1293 | 1674 | 1794 | 1862 | 1897 | 1833 |
The following is the fio configuration:
[global]
ioengine=io_uring
sqthread_poll=1
threads=1
iodepth=16
hipri=0
direct=0
fixedbufs=0
uncached=0
nowait=0
force_async=0
randrepeat=0
time_based=0
size=256M
filename=/data/test/local/io_uring_test
group_reporting
[read256B-rand]
bs=4096
rw=randwrite
numjobs=1
Signed-off-by: Lu Hongfei <[email protected]>
Signed-off-by: Yangtao Li <[email protected]>
---
fs/ext4/file.c | 5 +----
1 file changed, 1 insertion(+), 4 deletions(-)
diff --git a/fs/ext4/file.c b/fs/ext4/file.c
index 6a16d07965f9..5da7c5612324
--- a/fs/ext4/file.c
+++ b/fs/ext4/file.c
@@ -288,9 +288,6 @@ static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
ssize_t ret;
struct inode *inode = file_inode(iocb->ki_filp);
- if (iocb->ki_flags & IOCB_NOWAIT)
- return -EOPNOTSUPP;
-
inode_lock(inode);
ret = ext4_write_checks(iocb, from);
if (ret <= 0)
@@ -905,7 +902,7 @@ static int ext4_file_open(struct inode *inode, struct file *filp)
return ret;
}
- filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC |
+ filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC |
FMODE_DIO_PARALLEL_WRITE;
return dquot_file_open(inode, filp);
}
--
2.39.0
On Wed, Jun 14, 2023 at 04:31:04PM +0800, Lu Hongfei wrote:
> This adds the async buffered write support to ext4,
> the following is the relevant test data.
>
> iodepth | 1 | 2 | 4 | 8 | 16 | 32 |
> before(M/s) | 842 | 1047 | 1098 | 1167 | 1172 | 1247 |
> after(M/s) | 1293 | 1674 | 1794 | 1862 | 1897 | 1833 |
>
> The following is the fio configuration:
>
> [global]
> ioengine=io_uring
> sqthread_poll=1
> threads=1
> iodepth=16
> hipri=0
> direct=0
> fixedbufs=0
> uncached=0
> nowait=0
> force_async=0
> randrepeat=0
> time_based=0
> size=256M
> filename=/data/test/local/io_uring_test
> group_reporting
> [read256B-rand]
> bs=4096
> rw=randwrite
> numjobs=1
>
> Signed-off-by: Lu Hongfei <[email protected]>
> Signed-off-by: Yangtao Li <[email protected]>
> ---
> fs/ext4/file.c | 5 +----
> 1 file changed, 1 insertion(+), 4 deletions(-)
>
> diff --git a/fs/ext4/file.c b/fs/ext4/file.c
> index 6a16d07965f9..5da7c5612324
> --- a/fs/ext4/file.c
> +++ b/fs/ext4/file.c
> @@ -288,9 +288,6 @@ static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
> ssize_t ret;
> struct inode *inode = file_inode(iocb->ki_filp);
>
> - if (iocb->ki_flags & IOCB_NOWAIT)
> - return -EOPNOTSUPP;
> -
> inode_lock(inode);
Ah, no, that's not how we enable IOCB_NOWAIT functionality. Just
because removing the check makes it go a little bit faster, it
doesn't mean that it is actually doing the right thing.
i.e. before remove IOCB_NOWAIT checks like this, we actually have to
make the functionality operate in a non-blocking manner when
IOCB_NOWAIT is set. Clearly, this patch does not do that, because
the very first thing that occurs after the removed NOWAIT check is a
blocking inode_lock() call....
Please go and look at the series that added NOWAIT buffered write
support to XFS for io_uring an example of sorts of changes that need
to be made to ext4 to support IOCB_NOWAIT buffered writes.
https://lore.kernel.org/linux-xfs/[email protected]/
Cheers,
Dave.
--
Dave Chinner
[email protected]