On Thu, Mar 23, 2017 at 12:35:46PM +0800, Xin Long wrote:
> On Thu, Mar 23, 2017 at 1:33 AM, Marcelo Ricardo Leitner
> <marcelo.leit...@gmail.com> wrote:
> > On Wed, Mar 22, 2017 at 02:07:37PM +0000, David Laight wrote:
> >> Regardless of the MSG_MORE flags associated with any specific send()
> >> request there will always be protocol effects (like retransmissions
> >> or flow control 'on') that will generate different 'chunking'.
> >
> > Yes, those are the ones that may lead to some confusion on how it
> > actually works, and mangling them is not really desired for the
> > sideeffects that it might have.
> >
> > Sooner or later we could have bug reports like "hey this chunk shouldn't
> > have been packed with that." if we stick with the initial proposition,
> > while with David's view, we are only promising to not send packets with
> > a single chunk and as long as the application send more data fast enough.
> >
> > David, are we on the same page now? ;-)
> >
> > Xin, what do you think?
> If we insist that MSG_MORE means not to send  ANY data, I compromise.
> does ANY include retransmission DATA? should MSG_MORE block
> retransmission ?

That's not really what he meant by that, I think. That "ANY" in there is
a way to refer to the entire buf and not that msg sendmsg is sending.
Later I explained what I got from his explanation, which should be more
like:
"If MSG_MORE was used, and there are no packets in flight, do not send a
packet right away because the application is going to send more data."
Would have to handle the (Not-)Nagle situation too:
"If not using Nagle and using MSG_MORE, try to not generate a packet
right away." (because this may send packets with a single chunk even if
in_flight != 0)
In both cases, if the flush is generated by other triggers, it's okay.

Because if there are chunks already queued, they will be sent as soon as
in_flight reaches 0 or some other break is lifted (flow control).
Holding the chunk that was queued with MSG_MORE and sending a partial
packet in this case because of MSG_MORE is not good, it's possibly not
saving anything.

  Marcelo

Reply via email to