Use Message Queue as DMP and XCMP dispatch queue

[ggwpez]

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.

Final undeployment migrations are provided by cumulus_pallet_dmp_queue::UndeployDmpQueue and DeleteDmpQueue that can be configured with an aux config trait like:

parameter_types! {
	pub const DmpQueuePalletName: &'static str = "DmpQueue" < CHANGE ME;
	pub const RelayOrigin: AggregateMessageOrigin = AggregateMessageOrigin::Parent;
}

impl cumulus_pallet_dmp_queue::MigrationConfig for Runtime {
	type PalletName = DmpQueuePalletName;
	type DmpHandler = frame_support::traits::EnqueueWithOrigin<MessageQueue, RelayOrigin>;
	type DbWeight = <Runtime as frame_system::Config>::DbWeight;
}

// And adding them to your Migrations tuple:
pub type Migrations = (
	...
	cumulus_pallet_dmp_queue::UndeployDmpQueue<Runtime>,
	cumulus_pallet_dmp_queue::DeleteDmpQueue<Runtime>,
);

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.

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.

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
• optional: create TransformProcessMessageOrigin in Substrate and replace ProcessFromSibling

[ggwpez]

don’t we need to migrate the items from old pallet to new pallet? we can’t just delete them

Yes... i added cumulus_pallet_dmp_queue::UndeployDmp and cumulus_pallet_dmp_queue::DeleteDmp as described in the description (still need to be deployed everywhere).
I will check again, but something like this is probably also needed for the XCMP queue pallet.

[paritytech-cicd-pr]

The CI pipeline was cancelled due to failure one of the required jobs.
Job name: test-linux-stable
Logs: https://gitlab.parity.io/parity/mirrors/cumulus/-/jobs/3424752

[ggwpez]

@xlc the migration that i added for the DMP messages has the problem that it migrates them all at once.
This can cause overweight blocks and stall the chain.

Maybe we can do a multi-block migration in the DMP pallet without locking the runtime since the storage is inaccessible anyway.
Then the DMP pallet would migrate them one by one into the MQ pallet. Order would be wrong but XCM cannot rely on that anyway.