David Howells wrote:

Willem de Bruijn [off-list ref] wrote:

quoted

Is it intentional to add MSG_MORE here in this patch?

I do see that patch 3 removes this branch:

Yeah.  I think I may have tcp_bpf a bit wrong with regard to handling
MSG_MORE.

How about the attached version of tcp_bpf_push()?

I wonder if it's save to move the setting of MSG_SENDPAGE_NOPOLICY out of the
loop as I've done here.  The caller holds the socket lock.

Also, I'm not sure whether to take account of apply/apply_bytes when setting
MSG_MORE mid-message, or whether to just go on whether we've reached
sge->length yet.  (I'm not sure exactly how tcp_bpf works).

I'm not very familiar with it either.

Instead of inferring whether MSG_MORE is safe to set, as below, sufficient to
rely on the caller to pass it when appropriate?

size = min(apply_bytes, sge->length). I doubt that size < apply_bytes is
ever intended.

And instead of this former branch

                if (flags & MSG_SENDPAGE_NOTLAST)
                        msghdr.msg_flags |= MSG_MORE;

update any caller to pass MSG_MORE instead of MSG_SENDPAGE_NOTLAST, if not yet
done so.

		msghdr.msg_flags = flags;

		/* Determine if we need to set MSG_MORE. */
		if (!(msghdr.msg_flags & MSG_MORE)) {
			if (apply && size < apply_bytes)
				msghdr.msg_flags |= MSG_MORE;
			else if (!apply && size < sge->length &&
				 msg->sg.start != msg->sg.end)
				msghdr.msg_flags |= MSG_MORE;
		}