Jeff Moyer <jmoyer@xxxxxxxxxx> writes:
> Hi, Dave,
>
> Thanks for the review!
>
>> or better still, factor xfs_file_fsync() so that it calls a helper
>> that doesn't wait for data IO completion, and call that helper here
>> too. The semantics of fsync/fdatasync are too complex to have to
>> implement and maintain in multiple locations....
>
> I definitely agree with consolidating things. However, there are four
> blocking calls in xfs_file_fsync (filemap_write_and_wait_range,
> xfs_blkdev_issue_flush, _xfs_log_force_lsn, and another call to
> xfs_blkdev_issue_flush). How would you propose to make that
> non-blocking given that those steps have to happen in sequence?
OK, so re-reading your mail, I think you meant to just factor out
everything except the filemap_write_and_wait_range. Here are a couple
of patches which do that. Also, since we're not worried about blocking
in the endio processing, just making things synchronous makes the code a
lot simpler. Let me know what you think of the attached two patches
(which I've already run through xfstests).
Thanks!
Jeff
xfs-factor-out-xfs-file-fsync.patch
Description: Text document
xfs-flush-disk-cache-for-aio-dio-osync-writes.patch
Description: Text document
|