aboutsummaryrefslogtreecommitdiff
path: root/content/en/blog
diff options
context:
space:
mode:
Diffstat (limited to 'content/en/blog')
-rw-r--r--content/en/blog/20211229-flow-vs-connection.md289
-rw-r--r--content/en/blog/20211229-oecho-1.pngbin0 -> 442563 bytes
-rw-r--r--content/en/blog/20211229-oecho-2.pngbin0 -> 543765 bytes
-rw-r--r--content/en/blog/20211229-oecho-3.pngbin0 -> 552489 bytes
-rw-r--r--content/en/blog/20211229-oecho-4.pngbin0 -> 352163 bytes
5 files changed, 289 insertions, 0 deletions
diff --git a/content/en/blog/20211229-flow-vs-connection.md b/content/en/blog/20211229-flow-vs-connection.md
new file mode 100644
index 0000000..89b246b
--- /dev/null
+++ b/content/en/blog/20211229-flow-vs-connection.md
@@ -0,0 +1,289 @@
+---
+date: 2021-12-29
+title: "Behaviour of Ouroboros flows vs UDP sockets and TCP connections/sockets"
+linkTitle: "Flows vs connections/sockets"
+author: Dimitri Staessens
+---
+
+A couple of days ago, I received a very good question from someone who
+was playing around around with Ouroboros/O7s. He started from the
+[_oecho_](https://ouroboros.rocks/cgit/ouroboros/tree/src/tools/oecho/oecho.c#n94) tool.
+
+_oecho_ is a very simple application. It establishes what we call a
+"raw" flow. Raw flows have no fancy features, they are the best-effort
+class of packet transport (a bit like UDP). Raw flows do not have an
+Flow-and-retransmission control protocol (FRCP) machine. This person
+changed oecho to use a _reliable_ flow, and slightly modified it, ran
+into some unexpected behaviour,and then asked: **is it possible to
+detect a half-closed connection?** Yes, it is, but it's not
+implemented (yet). But I think it's worth answering this in a fair bit
+of detail, as it highlights some differences between O7s flows and
+(TCP) connections.
+
+A bit of knowledge on the core protocols in Ouroboros is needed, and
+can be found [here](/docs/concepts/protocols/) and the flow allocator
+[here](/docs/concepts/fa/). If you haven't read these in while, it
+will be useful to first read them to make the most out of this post.
+
+## The oecho application
+
+The oecho server is waiting for a client to request a flow, reads the
+message from the client, sends it back, and deallocates the flow.
+
+The client will allocate a _raw_ flow, the QoS parameter for the flow
+is _NULL_. Then it will write a message, read the response and also
+deallocate the flow.
+
+In a schematic, the communication for this simple application looks
+like this[^1]:
+
+{{<figure width="90%" src="/blog/20211229-oecho-1.png">}}
+
+All the API calls used are inherently _blocking_ calls. They wait for
+some event to happen and do not always return immediately.
+
+First, the client will allocate a flow to the server. The server's
+_flow\_accept()_ call will return when it receives the request, the
+client's _flow\_alloc()_ call will return when the response message is
+received from the server. This exchange agrees on the Endpoint IDs and
+possibly the flow characteristics (QoS) that the application will
+use. For a raw flow, this will only set the Endpoint IDs that will be
+used in the DT protocol[^2]. On the server side, the _flow\_accept()_
+returns, and the server calls _flow\_read()_. While the _flow\_read()_
+is still waiting on the server side, the flow allocation response is
+underway to the client. The reception of the allocation response
+causes the _flow\_alloc()_ call on the client side to return and the
+(raw) flow is established[^3].
+
+Now the client writes a packet, the server reads it and sends it
+back. Immediately after sending that packet, the server _deallocates_
+the flow. The response, containing a copy of the client message, is
+still on its way to the client. After the client receives it, it also
+deallocates the flow. Flow deallocation destroys the state associated
+with the flow and will release the EIDs for reuse. In this case of
+raw, unreliable flows, _flow\_dealloc()_ will return almost
+immediately.
+
+## Flows vs connections
+
+The most important thing to notice from the diagram for _oecho_, is
+that flow deallocation _does not send any messages_! Suppose that the
+server would send a message to destroy the flow immediately after it
+sends the response. What if that message to destroy the flow arrives
+_before_ the response? When do we destroy the state associated with
+the flow? Flows are not connections. Raw flows like the one used in
+oecho behave like UDP. No guarantees. Now, let's have a look at
+_reliable_ flows, which behave more like TCP.
+
+## A modification to oecho with reliable flows
+
+{{<figure width="90%" src="/blog/20211229-oecho-2.png">}}
+
+To use a reliable flow, we call a _flow\_alloc()_ from the client with
+a different QoS spec (qos_data). The flow allocation works exactly as
+before. The flow allocation request now contains a data QoS request
+instead of a raw QoS request. Upon reception of this request, the
+server will create a protocol machine for FRCP, the protocol in O7s
+that is in charge of delivering packets reliably, in-order and without
+duplicates. FRCP also performs flow control to avoid sending more
+packets than the server can process. When the flow allocation arrives
+at the client, it will also create an FRCP protocol instance. When
+these FRCP instances are created, they are in an initial state where
+the Delta-t timers are _timed out_. This is the state that allows
+starting a new _run_. I will not explain every detail of FRCP here,
+these are explained in the
+[protocols](/docs/concepts/protocols/#flow-and-retransmission-control-protocol-frcp)
+section.
+
+Now, the client sends its first packet, with a randomly chosen
+sequence number (100) and the Data Run Flag (DRF) enabled. The meaning
+of the DRF is that there were no _previously unacknowledged_ packets
+in the currently tracked packet sequence, and it allows to avoid a
+3-way handshake.
+
+When that packet with sequence number 100 arrives in the FRCP protocol
+machine at the server, it will detect that DRF is set to 1, and that
+it is in an initial state where all timers are timed out. It will
+start accepting packets for this new run starting with sequence number
+100. The server almost immediately sends a response packet back. It
+has no active sending run, so a random sequence number is chosen (300)
+and the DRF is set to 1. This packet will contain an acknowledgment
+for the received packet. FRCP acknowledgements contain the lowest
+acceptable packet number (so 101). After sending the packet, the
+server calls _dealloc()_, which will block on FRCP still having
+unacknowledged packets.
+
+Now the client gets the return packet, it has no active incoming run,
+the receiver connection is set to initial timed out state, and like
+the server, it will see that the DRF is set to 1, and accept this new
+incoming run starting from sequence number 300. The client has no data
+packets anymore, so the deallocation will send a _bare_
+acknowledgement for 301 and exit. At the server side, the
+_flow\_dealloc()_ call will exit after it receives the
+acknowledgement. Not drawn in the figure, is that the flow identifiers
+(EIDs) will only time out internally after a full Delta-t timeout. TCP
+does something similar and will not reused closed connection state for
+2 * Maximum Segment Lifetime (MSL).
+
+## Unexpected behaviour
+
+{{<figure width="90%" src="/blog/20211229-oecho-3.png">}}
+
+While playing around with the prototype, a modification was made to
+oecho as above: another _flow_read()_ was added to the client. As you
+can see from the diagram, there will never be a packet sent, and, if
+no timeout is set on the read() operation, after the server has
+deallocated the flow (and re-entered the loop to accept a new flow),
+the client will remain in limbo, forever stuck on the
+_flow\_read()_. And so, I got the following question:
+
+```
+I would have expected the second call to abort with an error
+code. However, the client gets stuck while the server is waiting for a
+new request. Is this expected? If so, is it possible to detect a
+half-closed connection?
+```
+
+## A _"half-closed connection"_
+
+So, first things first: the observation is correct, and that second
+call should (and soon will) exit on an error, as the flow is now valid
+anymore. Now it will only exit if there was an error in the FRCP
+connection (packet retransmission fails to receive an acknowledgment
+within a certain timeout). It should also exit on a remotely
+deallocated flow. But how will Ouroboros detect it?
+
+Now, a "half closed connection" comes from TCP. TCP afficionados will
+probably think that I need to add something to FRCP, like
+[FIN](https://www.googlecloudcommunity.com/gc/Cloud-Product-Articles/TCP-states-explained/ta-p/78462)
+at the end of TCP to signal the end of a flow[^4]:
+
+```
+TCP A TCP B
+
+ 1. ESTABLISHED ESTABLISHED
+
+ 2. (Close)
+ FIN-WAIT-1 --> <SEQ=100><ACK=300><CTL=FIN,ACK> --> CLOSE-WAIT
+
+ 3. FIN-WAIT-2 <-- <SEQ=300><ACK=101><CTL=ACK> <-- CLOSE-WAIT
+
+ 4. (Close)
+ TIME-WAIT <-- <SEQ=300><ACK=101><CTL=FIN,ACK> <-- LAST-ACK
+
+ 5. TIME-WAIT --> <SEQ=101><ACK=301><CTL=ACK> --< CLOSED
+
+ 6. (2 MSL) CLOSED
+```
+
+While FRCP performs functions that are present in TCP, not everything
+is so readily transferable. Purely from a design perspective, it's
+just not FRCPs job to keep a flow alive or detect if the flow is
+alive. It's job is to deliver packets reliably, or and all it needs to
+do that job is present. But would adding FINs work?
+
+Well, the server can crash just before the dealloc() call, leaving it
+in the current situation (the client won't receive FINs). To resolve
+it, it would also need a keepalive mechanism. Yes, TCP also has a
+keepalive mechanism. And would adding that solve it? Not to my
+satisfaction. Because, Ouroboros flows are not connections, they don't
+always have an end-to-end protocol (FRCP) running[^5]. So if we add
+FIN and keepalive to FRCP, we would still need to add something
+_similar_ for flows that don't have FRCP. We would need to duplicate
+the keepalive functionality somewhere else. The main objective of O7s
+is to avoid functional duplication. So, can we kill all the birds with
+one stone? Detect flows that are down? Sure we can!
+
+## Flow liveness monitoring
+
+But we need to take a birds eye view of the flow first.
+
+On the server side, the allocated flow has a flow endpoint with
+internal Flow ID (FID 16), to which the oecho server writes using its
+flow descriptor, fd=71. On the client side, the client reads/writes
+from its fd=68, which behind the scenes is linking to the flow
+endpoint with ID 9. On the network side, the flow allocator in the
+IPCPs also reads and writes from these endpoints to transfer packets
+along the network. So, the flow endpoint marks the boundary between
+the "network".
+
+{{<figure width="80%" src="/blog/20211229-oecho-4.png">}}
+
+This is drawn in the figure above. I'll repeat it because it is
+important: the datastructure associated with a flow at the endpoints
+is this "flow endpoint". It forms the bridge between the application
+and the network layer. The role of the IRMd is to manage these
+endpoints and the associated datastructures.
+
+Flow deallocation is a two step process: both the IPCP and the
+application have a _dealloc()_ call. The endpoint is only destroyed if
+_both_ the application process and the IPCP signal they are done with
+it. So a _flow\_dealloc()_ from the application will kill only its use
+with the endpoint. This allows the IRMd to keep it alive until it
+sends an OK to the IPCP to also deallocate the flow and signal it is
+done with it. Usually, if all goes well, the application will
+deallocate the flow first.
+
+The IRMd also monitors all O7s processes. If it detects an application
+crashing, or an IPCP crashing, it will automatically perform that
+applications' half of the flow deallocation, but not the complete
+deallocation. If an IPCP crashes, applications still hold the FRCP
+state and can recover the connection over a different flow[^6].
+
+So, now it should be clear that the liveness of a flow has to be
+detected in the flow allocator of the IPCPs, not in the application
+(again, reminder: FRCP state is maintained inside the application).
+The IPCP will detect that its flow has been deallocated locally
+(either intentionally or because of a crash).It's paramount to do it
+here, because of the recursive nature of the network. Flows are
+everywhere, also between "router machines"! Routers usually restrict
+themselves to raw flows. No retransmissions, no flow control, no fuss,
+that's all too expensive to perform at high rates. But they need to be
+able to detect links going down. In IP networks, the OSPF protocol may
+use something like Bi-directional Forwarding Detection (BFD) to detect
+failed adjacencies. And then applications may use TCP keepalive and
+FIN. Or HTTP keepalive. All unneeded functional duplication, symptoms
+of a messy architecture, at least in my book. In Ouroboros, this flow
+liveness check is implemented once, in the flow allocator. It is the
+only place in the Ouroboros system where liveness checks are
+needed. Clean. Shipshape. Nice and tidy. Spick and span. We call it
+Flow Liveness Monitoring (FLM).
+
+If I recall correctly, we implemented an FLM in the RINA/IRATI flow
+allocator years ago when we were working on PRISTINE and were trying
+to get loop-free alternate (LFA) routes working. This needed to detect
+flows going down. In Ouroboros it is not implemented yet. Maybe I'll
+add it in the near future. Time is in short supply, the items on my
+todo list are not.
+
+Probably long enough for a blog post. Have yourselves a wonderful new
+year, and above all, stay curious!
+
+Dimitri
+
+[^1]: We are omiting the role of the Ouroboros daemons (IPCPd's and
+ IRMd) for now. There would be a name resolution step for "oecho"
+ to an address in the IPCPds. Also, the IRMd at the server side
+ brokers the flow allocation request to a valid oecho server. If
+ the server is not running when the flow allocation request
+ arrives at the IRMd, O7s can also start the oecho server
+ application _in response_ to a flow allocation request. But
+ going into those details are not needed for this discussion. We
+ focus solely on the application perspective here.
+
+[^2]: Flow allocation has no direct analogue in TCP or UDP, where the
+ protocol to be used and the destination port are known in
+ advance. In any case, flow allocation should not be confused
+ with a TCP 3-way handshake.
+
+[^3]: I will probably do another post on how flow allocation deals
+ with lost messages, as it is also an interesting subject.
+
+[^4]: Or even more bluntly tell me to "just use TCP instead of FRCP".
+
+[^5]: A UDP server that has clients exit or crash is also left to its
+ own devices to clean up the state associated with that UDP
+ socket.
+
+[^6]: This has not been implemented yet, and should make for a nice
+ demo. \ No newline at end of file
diff --git a/content/en/blog/20211229-oecho-1.png b/content/en/blog/20211229-oecho-1.png
new file mode 100644
index 0000000..b15ccd9
--- /dev/null
+++ b/content/en/blog/20211229-oecho-1.png
Binary files differ
diff --git a/content/en/blog/20211229-oecho-2.png b/content/en/blog/20211229-oecho-2.png
new file mode 100644
index 0000000..ea59987
--- /dev/null
+++ b/content/en/blog/20211229-oecho-2.png
Binary files differ
diff --git a/content/en/blog/20211229-oecho-3.png b/content/en/blog/20211229-oecho-3.png
new file mode 100644
index 0000000..9c9152e
--- /dev/null
+++ b/content/en/blog/20211229-oecho-3.png
Binary files differ
diff --git a/content/en/blog/20211229-oecho-4.png b/content/en/blog/20211229-oecho-4.png
new file mode 100644
index 0000000..07e30e8
--- /dev/null
+++ b/content/en/blog/20211229-oecho-4.png
Binary files differ