[Drbd-dev] Protocol C

[Lars Ellenberg]

On Wed, May 28, 2014 at 05:25:28PM +0200, raulhp wrote:
> 
> 
> El 2014-05-28 14:49, Lars Ellenberg escribió:
> >On Wed, May 28, 2014 at 02:32:05PM +0200, raulhp wrote:
> >>Hi Lars,
> >>
> >>Thanks so much, I focus only the replication link, the first develop
> >>is implement the replication to n servers with reliable multicast,
> >>ensuring the write in the remote disk.
> >>
> >>Can you explain me this ? Where? ---> You would need to find the
> >>"relevant" requests on the resource->transfer_log()
> >
> >See in drbd_send_and_submit(), and helpers, e.g.
> >drbd_process_write_request,
> >where the requests are put on the transfer log in incoming order.
> >
> >The drbd sender threads (wait_for_sender_todo, process_one_request)
> >find them there, and send them.
> >
> >The drbd asender thread receives the acks (or, in your case
> >the retransmit requests), needs to validate thes acks (or neg acks,
> >or retrsnamit requests), and change the state of the respective
> >request
> >object.
> >
> >That's what happens in got_BlockAck and friends.
> All this it´s ok!
> 
> >Now you want to implement some mcast.
> >I don't know what information you envison to have,
> >what would make you ask for retransmit,
> >what exactly you would ask for,
> >what information you could use in your got_retransmit_request
> >function.
> >
> I want to implement a mcast with NACK, where the receiver send NACK
> message to the sender when the receiver detects a loss packet, and
> the sender send the packet in question, with this is possible
> guarantee all writes in the remote node and reduce the traffic in
> the network.
> 
> >But the drbd_request will stay linked on the resource->transfer_log
> >until it is "acked", or "cleaned up" due to connection loss.
> 
> It´s ok
> 
> >Depending on the information you have in that restransmit request,
> >you will be able to find the relevant requests on that transfer log.
> >I don't know what information you have, so I cannot tell how to best
> >find them.
> >
> >Non-relevant are:
> >  those that have not been sent yet (to that peer)
> >  those that have already been ACKed (by that peer)
> >Which means that relevant are
> >  those that have been sent, but not acked,
> >  and fall in whatever "range" you request to be restransmitted.
> This is ok, but with Multicast communication is necesary to use UDP
> Protocol with Reliability.

Of course.
So?

We already have a "NACK", in the DRBD protocol,
which carries the meaning "failed to serve this request"
(the IO subsystem of the peer reported failure).

Lets not overload the term "nack",
and use the term "retransmit request",
because that is basically what a "nack"
in any reliable multicast protocol is supposed to trigger: 
retransmit of not received messages.

The receiver notices (by timeouts and sequence numbers and other
conditions) that it missed messages, and sends back that it'd like
to please have "these" messages again.

Maybe it is useful to have both:
the "multicast" channel,
and still the tcp mesh channels for retransmits,
and maybe for read requests or other directed messages.
(just an idea)

The node that originally sent the requests
now sees this retransmit request,
and needs to process it.

To do that, it needs to find the original request,
and resend it.

Luckily we have all requests
  * on the transfer log
     (struct drbd_resource: transfer_log)
  * and in the interval tree
     (see validate_req_change_req_state
      and find_request).

So based on what information you put in your retransmit request
you should be able to find the original either on the transfer log
or in the interval tree, and resend it.

Am I misunderstanding your question?

-- 
: Lars Ellenberg
: LINBIT | Your Way to High Availability
: DRBD/HA support and consulting http://www.linbit.com