bmistree / pronghorn Goto Github PK

Can push an update to hardware, but then have to back it out, if, for instance another switch that is part of transaction fails. In this case, should keep track of info needed to undo change and actually apply it to hardware.

Currently, fairness test reads then writes to contended switch variables. This means that can have cases where after an event completes all other events that were waiting on the read lock get scheduled. However, only one acquires the write lock and all others get preempted. This can cause some slipping in fairness results. Instead, for test all should assume write lock before reading.

Receive an error code from switch. Other match fields seem to work okay. Likely a problem with floodlight or pronghorn (not with switch because tested inserting matches using ovs-ofctl).

In FTableUpdate.java, string_to_actions_list does not actually take user's action and turn it into a list of actions to apply. It's currently just a placeholder. See FIXME in the code.

Commit 618e3ba changes rtable in example to prevent the problem in examples and experiments. But does not actually clear the rtable.

Have a single controller test, should add a multi-controller version.

In ralph_field_type_to_floodlight_field_type, left a fixme and assert in code for case where user enters and invalid/unrecognized field type as part of set_field action. Do something more appropriate. Either in Java or do check in Ralph.

All entries in routing table should be assumed active, not just the ones that are tagged active.

Currently, not actually sending http request to update routing table when logical routing table changes.

If a programmer adds a flow table entry and then removes it as part of the same transaction, should filter this duplicate in wrapped_switch.rph. Reason to filter is that because switch does not ensure order of flow table changes, could get a case where remove and add are reversed.

Ensure that they are properly formatted in ralph and throw exception there if they are not. Currently, have an issue where will submit incorrectly formatted match, do not push incorrect match to switches (see FIXME in match_from_internal_match in DeltListStateSupplier.java), but incorrect match persists in software state.

See fixme in code.

Daniel reported seeing a backout exception in the coarse locking tests
https://gist.github.com/jackowayed/8645395/raw/4d7e8b360c6eebc402e2a18e5bd80d6082e856a7/output2.txt

for sha id: 9fce3be.

Right now, not doing anything intelligent to parse for ipv6 addresses. For actions, users must enter the ipv6 address as a string of bytes. This is really, really ugly and bad.

Have it shutdown cleanly after printing values to stdout.

I think this happens because it assumes a single floodlight controller per host, and both controllers run on a single host. Consider how to fix.

Currently, just setting a string for actions. But want to allow script to goto table, etc. Do so by creating a new struct, etc. for instructions and adding those to Struct FlowTableEntry in place of current actions field.

Previously, had assumed that identifiers needed to be unique per switch. Can easily fix by creating identifiers as the concatenation of the switch id, some separator, and current switch-local unique identifiers.

Previous submission only had distributed controller test.

In commit 135fe38, correctly taking into account potential backouts on state by short-circuiting with continues in produce_ftable_updates of DeltaListStateSupplier. Note, can probably do this faster by returning out with an empty list of updates when backout has been detected, rather than using continue. Could potentially save a flow mod and barrier to switches for a multi-update.

In DeltaListStateSupplier, when running get_state_to_push, makes unprotected access to internal_ft_deltas_list and its contents. What happens if those objects receive backouts while we're making this access?

Figure out a way to shut down floodlight+pronghorn cleanly.

Currently, only reset flow_table_deltas list to be empty in speculate listener. This means that deltas_list grows and grows when speculation is turned off. Should fix.

When we force speculation, we overwrite dirty values for all existing state. This is bad, because our state supplier eventually needs to be able to read internal dirty state. Just moving force_speculate call into DeltaListStateSupplier breaks, because then we never get speculation on read only-s. Should fix.