Re: [RFC] write(2) semantics wrt return values and current position

[Al Viro <viro@ZenIV.linux.org.uk>]

On Mon, Apr 06, 2015 at 05:02:31PM +0100, Al Viro wrote:

7) commit 3a83b342c87e6d21290de8dc76ec20a67821261d
Author: Joseph Qi <joseph.qi@huawei.com>
Date:   Mon Feb 16 16:00:09 2015 -0800

    ocfs2: complete the rest request through buffer io

appears to be very odd.  Look:

                written = generic_file_direct_write(iocb, from, *ppos);
-               if (written < 0) {
+               if (written < 0 || written == count) {
                        ret = written;
                        goto out_dio;
                }
+
+               /*
+                * for completing the rest of the request.
+                */
+               *ppos += written;

ppos here is &iocb->ki_pos.  Now, take a look at the end of
generic_file_direct_write():
        if (written > 0) {
                pos += written;
                iov_iter_advance(from, written);
                if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
                        i_size_write(inode, pos);
                        mark_inode_dirty(inode);
                }
                iocb->ki_pos = pos;
        }
out:
        return written;

In other words, after short write done by ->direct_IO(), we end up incrementing
position by *twice* the amount written by it.  And if it's short, but not
empty, we appear to be buggered...

Unless I hear "Al, you idiot, it's doing the right thing, here's what you've
missed: ...", I'm going to take that increment in ocfs2_file_write_iter()
out, and send it to Linus for 4.0 - it's post-3.19 regression.