On 05/02/2025 19:55, Darrick J. Wong wrote:
On Tue, Feb 04, 2025 at 12:01:24PM +0000, John Garry wrote:
Add xfs_file_dio_write_atomic() for dedicated handling of atomic writes.
In case of -EAGAIN being returned from iomap_dio_rw(), reissue the write
in CoW-based atomic write mode.
In the CoW-based atomic write mode, first unshare blocks so that we don't
have a cow fork for the data in the range which we are writing.
Signed-off-by: John Garry <john.g.garry@xxxxxxxxxx>
---
fs/xfs/xfs_file.c | 51 +++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 51 insertions(+)
diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index fd05b66aea3f..12af5cdc3094 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -619,6 +619,55 @@ xfs_file_dio_write_aligned(
return ret;
}
+static noinline ssize_t
+xfs_file_dio_write_atomic(
+ struct xfs_inode *ip,
+ struct kiocb *iocb,
+ struct iov_iter *from)
+{
+ unsigned int iolock = XFS_IOLOCK_SHARED;
+ bool use_cow = false;
+ unsigned int dio_flags;
+ ssize_t ret;
+
+retry:
+ ret = xfs_ilock_iocb_for_write(iocb, &iolock);
+ if (ret)
+ return ret;
+
+ ret = xfs_file_write_checks(iocb, from, &iolock);
+ if (ret)
+ goto out_unlock;
+
+ if (use_cow) {
+ ret = xfs_reflink_unshare(ip, iocb->ki_pos,
+ iov_iter_count(from));
Nit: continuation lines should be indented two tabs:
ret = xfs_reflink_unshare(ip, iocb->ki_pos,
iov_iter_count(from));
ok
+ if (ret)
+ goto out_unlock;
+ }
+
+ trace_xfs_file_direct_write(iocb, from);
+ if (use_cow)
+ dio_flags = IOMAP_DIO_ATOMIC_COW;
+ else
+ dio_flags = 0;
I also think you could eliminate use_cow by initializing dio_flags to
zero at the top, OR'ing in IOMAP_DIO_ATOMIC_COW in the retry clause
below, and using (dio_flags & IOMAP_DIO_ATOMIC_COW) to determine if you
should call unshare above.
ok, fine, if you think that it is better
Note: This serializes all the software untorn direct writes. I think
a more performant solution would allocate the cow staging blocks ondisk,
attach them to the directio ioend context, and alter ->iomap_begin and
the ioend remap to use the attached blocks, but that's a lot more
surgery.
sure, that does sound like it's quite intrusive. But whatever we do I
would like to keep the behaviour that racing reads and atomic writes
mean that a read sees all old or all new data. That is how SCSI and NVMe
behaves, even though it is not an advertised atomic write feature.
Thanks,
John
