Implement Everything #267

seliopou · 2014-03-01T16:00:17Z

This pull request subsumes #253 and #256, and implements additional functionality. Specifically, it:

defines an Frenetic application API
implements an async-based controller against that API
enabled writing policies with pipes, i.e., named controller locations
implements packet-in dispatch to applications, which involves running part of the policy on the controller
provides baked-in topology discovery as a frenetic application (can be easily removed)
provides a learning switch demo app

portId is currently a synonym for a VInt.t but that will likely change at some point to be a synonym for int64. Use the type consistently here, which may save time later.

This commit includes two important type definitions: * event - describes the types of events that frenetic applications will be exposed to, and will be expected to handle in some way. * result - the ultimate result of a frenetic application. A list of packet_outs to send to various switches, as well as the new policy that should be installed on those switches. It's on this type that sequence and union will be defined, and then lifted to a function type.

For all intents and purposes, Async_NetKAT.mli is a specification of the Frenetic application API. Closes #244.

The name did not imply that there was anything else being sequenced. Now it implies that after the action, the type of the module the function belongs to is also being sequenced.

matches determines whether or not some pattern-like structure matches a given packet.

from_pipe takes a Local.t, which a controller may cache per switch, and a packet, and determines the pipe that the packet belongs to. In general this is not always possible. For a brief interval during switch setup, packets may be sent to the controller not due to a policy. If a policy makes use of different pipe name in non-disjoint parallel compositions, then it may come from either, in which case the pipe the packet belongs to is ambiguous.

The user may have composed two applications that are interested in packets that match identical filters, but are sent to different pipes. from_pipe must account for this, and therefore return all pipes that match the packet. It is the job of the runtime to then deliver the packets to the proper pipe within the application, when possible.

(cherry picked from commit 434b1f8)

The from_pipes function now assumes that packets are not modified on the switch before being sent to the controller. That means that any modification of a field preceding a "port := <pipe>" is ignored for the purposes of compilation, and implemented at the controller.

Not sure why the code was clearing it before.

Any action that involves "port := <pipe>" should skip all other modifications, and send the packet straight to the controller. The skipped modifications will be applied there.

Skip filters that match on pipes.

When the controller receives a packet, it uses a cached representation of the policy installed on the switch to determine which pipe it should be sent to.

Previously called from_pipes. This function will evaluate a packet on a switch's policy, returning two lists. The first contains pairs of pipe names and modified packets. These will be sent to listening applications. The second is a list of modified packets. These will be used to generate packet out messages.

When a packet arrives at the controller, it's eval'd given the policy currently installed on the switch. If that produces packets located at physical locations, the controller will generate packet outs that will perform the appropriate modifications in the form of actions. Previously, the packet was not modified and no actions were inclued with the packet_out.

Was actually setting vlanPcp

Due to the flattening of seqs that involve Controller actions, this function may now produce duplicates. So, deduplicate. This approach does not preserve the order of the pars.

Right now, the compiler will not generates rules to modify packets if they'll ultimately be sent to the controller. It's up to the controller to implement those changes, which must happen before passing packets to the application.

The port optional port in a OFPT_PACKET_OUT is only used when the action list contains an action that would send the packet to the TABLE pseudoport. As of now, Frenetic does not allow this.

(cherry picked from commit fbfcc09)

If a branch of a union has side-effects, the user may want to ensure their ordering by using `Sequential, which will evaluate the left application to completion before running the right one. `Parallel provides no guarantee about evaluation order.

This reverts commit 8ba2bd0. Conflicts: lib/NetKAT_Types.ml

Fill out the rest of the topology events. These have some overlap. For example, the user will see a PortUp event as well as a LinkUp event containing the same switch_endpoint. It's up to each application to decide what level of detail they want to observe, and ignore the rest.

Use a custom instance of Network and stop using the topology disocvery stage from ocaml-openflow.

Asserts seem to not play nicely with async.

This reverts commit 2d4f046.

The unit argument indicates a partial application point. Anything computed in the partial application not be recomputed on applications of subsequent arguments.

Conflicts: _tags lib/LocalCompiler.ml lib/NetKAT_Types.ml lib/Semantics.ml lib/netcore.mllib setup.ml

tweaks to discovery

Move the creation of the event pipes to Ctl, and expose a functions to use it, as well as query to state of pending ports.

The Discovery.Host app will copy all arp traffic and send it to the controller. It assumes that the entire network is under the management of the controller.

The events written to the pipe will eventually get fed back into the app, so blocking on pushback was causing a deadlock.

jnfoster · 2014-03-01T16:27:14Z

Officially an "epic" pull request.

arjunguha · 2014-03-01T20:44:30Z

I ran $ sudo mn --controller=remote --topo=tree,2,2 and in another terminal got this error:

$ ./katnetic.byte run local examples/tree-2-2.kat 
(lib/typed_tcp.ml.Make.Connection_closed
  lib/typed_tcp.ml.Make.Eof_from_client)
 [INFO] [topology.switch] ↑ { switch = 1 }
[ERROR] switch 1: Failed to update table
 [INFO] [topology.switch] ↑ { switch = 1; port = 2 }
[ERROR] switch 1: Failed to send packet_out
 [INFO] [topology.switch] ↑ { switch = 1; port = 1 }
[ERROR] switch 1: Failed to send packet_out
[ERROR] switch 1: Failed to send packet_out
[ERROR] switch 1: Failed to send packet_out
[ERROR] switch 1: Failed to send packet_out
(lib/typed_tcp.ml.Make.Connection_closed
  lib/typed_tcp.ml.Make.Eof_from_client)
 [INFO] [learning] switch 1: learn 42:3c:a5:a9:17:5e => 1@1
 [INFO] [learning] switch 1: flood dlSrc=42:3c:a5:a9:17:5e,dlDst=33:33:ff:a9:17:5e,frameType=0x86DD
[ERROR] switch 1: Failed to send packet_out
[ERROR] switch 1: Failed to send packet_out
("unhandled exception"
 ((lib/monitor.ml.Error_
   ((exn (lib/typed_tcp.ml.Make.Missing_client_id 2))
    (backtrace
     ("Raised at file \"async/Async_OpenFlowChunk.ml\", line 50, characters 27-30"
      "Called from file \"lib/deferred.ml\", line 12, characters 64-67"
      "Called from file \"lib/jobs.ml\", line 216, characters 10-13" ""))
    (monitor
     (((name Monitor.protect) (here ()) (id 8) (has_seen_error true)
       (someone_is_listening true) (kill_index 0))))))
  (Pid 2071)))

It seems to be reproducible. Any ideas?

seliopou · 2014-03-01T21:02:54Z

A few things to try:

Make sure that mininet's properly cleaned up and you're starting fresh. Run sudo mn -c and check to make sure there are no zombie processes.
Are you using the current HEAD of ocaml-openflow?
While it print out errors, it may still be in working order. Did you check for connectivity?
Instead of using examples/tree-2-2.kat can you try with a static policy consisting of drop and nothing else?

seliopou · 2014-03-01T21:11:59Z

Also, note that in order for the policy to work you shoud run mn with the --mac option as well.

arjunguha · 2014-03-01T22:03:03Z

Ah. I think it has to do with the sizes we use for ports and switches. The
static policy uses older sizes, I think. Investigating.

On Sat, Mar 1, 2014 at 4:11 PM, Spiros Eliopoulos
[email protected]:

Also, note that in order for the policy to work you shoud run mn with the
--mac option as well.

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/267#issuecomment-36436804
.

arjunguha · 2014-03-01T23:01:42Z

I'm getting the error here:

https://github.com/frenetic-lang/ocaml-openflow/blob/master/lib/VInt.ml#L23

With a little more debugging output, it prints the following:

[ERROR] switch 3: Failed to send packet_out ((lib/monitor.ml.Error_
  ((exn (Invalid_argument "get_int (Int32 2)"))
    (backtrace
      ("Raised at file \"lib/VInt.ml\", line 23, characters 15-65"
        "Called from file \"async/Async_NetKAT_Controller.ml\", line 140, characters 23-38"
...

From my console:

$ ocaml
        OCaml version 4.01.0

# 2l > Int32.of_int max_int;;
- : bool = true
# max_int;;
- : int = 4611686018427387903
# Int32.of_int max_int;;
- : int32 = -1l
#

Can anyone else reproduce this?

arjunguha · 2014-03-01T23:03:55Z

Wait, this is obviously bogus. I'm on a 64-bit platform.

Besides, all these overflow checks should just use Int32.max_int, Int64.max_int, etc.

arjunguha · 2014-03-01T23:04:41Z

How has this not affected anyone else? Is everyone else using a 32-bit platform?

seliopou · 2014-03-01T23:17:58Z

Yes, this is a known issue. See frenetic-lang/ocaml-openflow#105. It needs to be fixed for the next release. Maybe everybody's still developing on the frenetic vm, which was 32-bit?

jnfoster · 2014-03-02T02:32:24Z

Huzzah!

seliopou and others added 30 commits February 17, 2014 17:17

apps: Use existing type synonyms consistently

f310a05

portId is currently a synonym for a VInt.t but that will likely change at some point to be a synonym for int64. Use the type consistently here, which may save time later.

apps: async interface and combinators

573c8ae

For all intents and purposes, Async_NetKAT.mli is a specification of the Frenetic application API. Closes #244.

apps: implementation of netkat async controller

9ef1ae0

from_pipe: rename seq_act to seq_act_t

0bf4f5c

The name did not imply that there was anything else being sequenced. Now it implies that after the action, the type of the module the function belongs to is also being sequenced.

from_pipe: Add matches operation to relevant modules

e04fc98

matches determines whether or not some pattern-like structure matches a given packet.

core_map: Use Core_map in Semantics

1e6e496

(cherry picked from commit 434b1f8)

from_pipe: Preserve location in returned packets

e3728c8

Not sure why the code was clearing it before.

from_pipe: Properly compile pipe location assignment

9c2f676

Any action that involves "port := <pipe>" should skip all other modifications, and send the packet straight to the controller. The skipped modifications will be applied there.

from_pipe: Properly compile location filter

f18cfaa

Skip filters that match on pipes.

from_pipes: apply semantics in Controller

f48ff72

When the controller receives a packet, it uses a cached representation of the policy installed on the switch to determine which pipe it should be sent to.

Fix bug in mk_ethType

da5eae4

Was actually setting vlanPcp

from_pipe: fix bug in LocalCompiler.set_to_action

d4f14f3

Due to the flattening of seqs that involve Controller actions, this function may now produce duplicates. So, deduplicate. This approach does not preserve the order of the pars.

from_pipe: Implement packet modification in controller

549f900

Right now, the compiler will not generates rules to modify packets if they'll ultimately be sent to the controller. It's up to the controller to implement those changes, which must happen before passing packets to the application.

from_pipe: remove unused port in packet_out

8ba2bd0

The port optional port in a OFPT_PACKET_OUT is only used when the action list contains an action that would send the packet to the TABLE pseudoport. As of now, Frenetic does not allow this.

from_pipe: move result type from NetKAT_Types to Async_NetKAT

2dd3429

from_pipe: Document Async_NetKAT

2f3b161

Really remove PolicyGenerator for now

96b3d0d

(cherry picked from commit fbfcc09)

from_pipe: port to latest topology API

2613fc3

Revert "from_pipe: remove unused port in packet_out"

08363a7

This reverts commit 8ba2bd0. Conflicts: lib/NetKAT_Types.ml

topo: don't use ocaml-openflow topology discovery

3092efb

Use a custom instance of Network and stop using the topology disocvery stage from ocaml-openflow.

topo: pass around a reference to Topology.t

caa4cb0

topo: replace assertions with raises

3dc5c93

Asserts seem to not play nicely with async.

seliopou added 16 commits February 24, 2014 23:52

Revert "topo: Remove LinkDown event"

25886e4

This reverts commit 2d4f046.

topo: Support LinkDown event

3157326

topo: a comment for posterity

897b549

topo: remove dead code related to seq_act

06650c1

topo: rename switch to learn, as switch is a reserved word

eb0ce8c

topo: have to_event use pipe for packet_outs

797d4bf

apps: Add unit argument to handler type

c264b4a

The unit argument indicates a partial application point. Anything computed in the partial application not be recomputed on applications of subsequent arguments.

Merge branch 'master' into topo

d71e5fe

Conflicts: _tags lib/LocalCompiler.ml lib/NetKAT_Types.ml lib/Semantics.ml lib/netcore.mllib setup.ml

topo: refactor switch discovery, send probes periodically

c805d8b

tweaks to discovery

topo: refactor switch discovery and ctl

7ccca85

Move the creation of the event pipes to Ctl, and expose a functions to use it, as well as query to state of pending ports.

topo: arp-based host discovery

ca8ff6e

The Discovery.Host app will copy all arp traffic and send it to the controller. It assumes that the entire network is under the management of the controller.

topo: add logging to learning app

b11a2e9

topo: improve logging for discovery app

ea2dac9

topo: print more logging tags in controller

c63d372

topo: Write without pushback to event pipe in Discovery

1c5f02a

The events written to the pipe will eventually get fed back into the app, so blocking on pushback was causing a deadlock.

topo: Add error logging to controller loops

a6e6268

seliopou mentioned this pull request Mar 1, 2014

Implement pipe locations #256

Closed

seliopou added this to the Open Networking Summit - Demo milestone Mar 1, 2014

arjunguha merged commit a6e6268 into master Mar 2, 2014

seliopou deleted the topo branch March 2, 2014 06:47

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement Everything #267

Implement Everything #267

seliopou commented Mar 1, 2014

jnfoster commented Mar 1, 2014

arjunguha commented Mar 1, 2014

seliopou commented Mar 1, 2014

seliopou commented Mar 1, 2014

arjunguha commented Mar 1, 2014

arjunguha commented Mar 1, 2014

arjunguha commented Mar 1, 2014

arjunguha commented Mar 1, 2014

seliopou commented Mar 1, 2014

jnfoster commented Mar 2, 2014

Implement Everything #267

Implement Everything #267

Conversation

seliopou commented Mar 1, 2014

jnfoster commented Mar 1, 2014

arjunguha commented Mar 1, 2014

seliopou commented Mar 1, 2014

seliopou commented Mar 1, 2014

arjunguha commented Mar 1, 2014

arjunguha commented Mar 1, 2014

arjunguha commented Mar 1, 2014

arjunguha commented Mar 1, 2014

seliopou commented Mar 1, 2014

jnfoster commented Mar 2, 2014