Use Message Queue as DMP and XCMP dispatch queue

[ggwpez]

(imported from paritytech/cumulus#2157)

Changes

This MR refactores the XCMP, Parachains System and DMP pallets to use the MessageQueue for delayed execution of incoming messages. The DMP pallet is entirely replaced by the MQ and thereby removed. This allows for PoV-bounded execution and resolves a number of issues that stem from the current work-around.

All System Parachains adopt this change.
The most important changes are in primitives/core/src/lib.rs, parachains/common/src/process_xcm_message.rs, pallets/parachain-system/src/lib.rs, pallets/xcmp-queue/src/lib.rs and the runtime configs.

DMP Queue Pallet

The pallet got removed and its logic refactored into parachain-system. Overweight message management can be done directly through the MQ pallet.

The undeployment of the pallet commences through the on_idle hook step-by step and moves all the remaining messages into the DmpSink.

impl cumulus_pallet_dmp_queue::Config for Runtime {
	type DmpSink = MessageQueue;
	...

The pallet can be removed after the lazy migration is done and event Completed is emitted.

XCMP Queue pallet

Removed all dispatch queue functionality. Incoming XCMP messages are now either: Immediately handled if they are Signals, enqueued into the MQ pallet otherwise.

New config items for the XCMP queue pallet:

/// The actual queue implementation that retains the messages for later processing.
type XcmpQueue: EnqueueMessage<ParaId>;

/// How a XCM over HRMP from a sibling parachain should be processed.
type XcmpProcessor: ProcessMessage<Origin = ParaId>;

/// The maximal number of suspended XCMP channels at the same time.
#[pallet::constant]
type MaxInboundSuspended: Get<u32>;

How to configure those:

// Use the MessageQueue pallet to store messages for later processing. The `TransformOrigin` is needed since
// the MQ pallet itself operators on `AggregateMessageOrigin` but we want to enqueue `ParaId`s.
type XcmpQueue = TransformOrigin<MessageQueue, AggregateMessageOrigin, ParaId, ParaIdToSibling>;

// Process XCMP messages from siblings. This is type-safe to only accept `ParaId`s. They will be dispatched
// with origin `Junction::Sibling(…)`.
type XcmpProcessor = ProcessFromSibling<
	ProcessXcmMessage<
		AggregateMessageOrigin,
		xcm_executor::XcmExecutor<xcm_config::XcmConfig>,
		RuntimeCall,
	>,
>;

// Not really important what to choose here. Just something larger than the maximal number of channels.
type MaxInboundSuspended = sp_core::ConstU32<1_000>;

The InboundXcmpStatus storage item was replaced by InboundXcmpSuspended since it now only tracks inbound queue suspension and no message indices anymore.

Now only sends the most recent channel Signals, as all prio ones are out-dated anyway.

Parachain System pallet

For DMP messages instead of forwarding them to the DMP pallet, it now pushes them to the configured DmpQueue. The message processing which was triggered in set_validation_data is now being done by the MQ pallet on_initialize.

XCMP messages are still handed off to the XcmpMessageHandler (XCMP-Queue pallet) - no change here.

New config items for the parachain system pallet:

/// Queues inbound downward messages for delayed processing. 
///
/// Analogous to the `XcmpQueue` of the XCMP queue pallet.
type DmpQueue: EnqueueMessage<AggregateMessageOrigin>;

How to configure:

/// Use the MQ pallet to store DMP messages for delayed processing.
type DmpQueue = MessageQueue;

Message Flow

The flow of messages on the parachain side. Messages come in from the left via the Validation Data and finally end up at the Xcm Executor on the right.

Further changes

• Bumped the default suspension, drop and resume thresholds in QueueConfigData::default().
• XcmpQueue::{suspend_xcm_execution, resume_xcm_execution} errors when they would be a noop.
• Properly validate the QueueConfigData before setting it.
• Marked weight files as auto-generated so they wont auto-expand in the MR files view.
• Move the hypothetical asserts to frame_support under the name experimental_hypothetically

Questions:

• What about the ugly #[cfg(feature = \"runtime-benchmarks\")] in the runtimes? Not sure how to best fix. Just having them like this makes tests fail that rely on the real message processor when the feature is enabled.
• Need a good weight for MessageQueueServiceWeight. The scheduler already takes 80% so I put it to 10% but that is quite low.

TODO:

• Remove c&p code after Use Message Queue pallet for UMP dispatch polkadot#6271
• Use HandleMessage once it is public in Substrate
• fix runtime-benchmarks feature Propagate runtime-benchmarks feature polkadot#6966
• Benchmarks
• Tests
• Migrate InboundXcmpStatus to InboundXcmpSuspended
• Possibly cleanup Migrations (DMP+XCMP)
• optional: create TransformProcessMessageOrigin in Substrate and replace ProcessFromSibling
• Rerun weights on ref HW

[xlc]

There is no UndeployDmpQueue. So I guess the PR description is outdated? @ggwpez

[ggwpez]

There is no UndeployDmpQueue. So I guess the PR description is outdated?

Yes updated it. The migrations are now running automatic in on_idle and it can be removed afterwards.