Skip to content

Latest commit

 

History

History
477 lines (367 loc) · 19.9 KB

fip-0031.md

File metadata and controls

477 lines (367 loc) · 19.9 KB
Raw
fip title authors discussions-to status type category created spec-sections requires replaces
0031
Atomic switch to non-programmable FVM
Raúl Kripalani (@raulk), Steven Allen (@stebalien)
Draft
Technical Core
Core
2022-02-03
TBD
FIP-0030 (Introducing the Filecoin Virtual Machine)
N/A

FIP-0031: Atomic switch to non-programmable FVM

Simple Summary

This FIP proposes the atomic activation of a non-programmable version of the FVM in Filecoin mainnet (and consequent deprecation of the legacy VM), along with a state migration and the introduction of a state object for the system actor (f00). It also establishes filecoin-project/builtin-actors as the canonical actor implemenetation for all clients to use.

Abstract

This FIP proposes an atomic switch from the legacy VM to a non-programmable version of the Filecoin Virtual Machine. A baseline specification of the Filecoin Virtual Machine is provided in FIP-0030.

The switch involves a state tree migration to replace synthetic actor Code CIDs with genuine content-addressed CID, computed over the respective Wasm bytecode.

We also introduce a state object for the system actor (up until now, stateless) to hold static network configuration, starting with the registry of built-in actor CodeCIDs.

Furthermore, we propose to canonicalize filecoin-project/builtin-actors as the standard built-in actor implementation.

This FIP introduces no functional actor behaviour changes; however, some exit codes have been adjusted.

Change Motivation

This step represents an important preparatory step in the trajectory towards full on-chain user-programmability. As a side effect, it improves overall network security by sandboxing execution and making it deterministic.

Specification

Canonicalization of built-in actors

We establish filecoin-project/builtin-actors 1 as the canonical implementation of Filecoin built-in actors (aka as System Actors in the Filecoin Spec), for all client implementations to use.

This FIP introduces actors v8, a version fundamentally equivalent to actors v7 (nv15), but reliant on the FVM as its execution environment.

Adoption of content-addressed Code CIDs

Every actor in the state tree specifies the CID of the executable code that backs it (CodeCID), doubling as an actor type designator.

Prior to the Wasm-based FVM, there was no universal, portable execution bytecode for the network to agree on. Therefore, the network used synthetic CIDs of the following form:

CodeCid = Cid(IPLD_RAW_CODEC, Mh(IDENTITY, bytes('fil/$actor_version/$actor_type')))

Where:

  • IPLD_RAW_CODEC=0x55
  • IDENTITY=0x00

This FIP transitions to true content-addressed CIDs for every actor, computed over their Wasm bytecode, in the following way:

CodeCid = Cid(IPLD_RAW_CODEC, Mh(BLAKE2B-256, $wasm_bytecode))

Sourcing built-in actors Wasm bytecode

The build process of filecoin-project/builtin-actors produces a CARv1 archive that bundles all Wasm bytecode for all actors. It has the following characteristics:

  • The CARv1 header is single-rooted.

  • The root CID resolves to a data structure that acts like a Manifest.

  • The Manifest type is a versioned struct, encoded as a CBOR tuple, defined as following:

    struct Manifest {
  version: u32
  data: CID
}

    The Manifest data for version 1 (the only version currently defined) is a CID linking to an IPLD Vector<(String, CID)>, ordered by explicit position, specified below.

    Each entry represents a built-in actor.

  • Keys (String) range over an enum that identifies the actor type. These strings match parts of the current synthetic Code CIDs, and therefore can eventually facilitate translations if necessary.

    • system (index 0)
    • init (index 1)
    • cron (index 2)
    • account (index 3)
    • storagepower (index 4)
    • storageminer (index 5)
    • storagemarket (index 6)
    • paymentchannel (index 7)
    • multisig (index 8)
    • reward (index 9)
    • verifiedregistry (index 10)

  • Values (CID) point to the Wasm bytecode of an actor as a single block.

Clients are free to ship the bundle as they see fit. The simplest approach is to embed the CARv1 into the client binary, but downloading it on-demand is also an option.

On node startup, the client is expected to:

  1. Import the CARv1 contents into its state blockstore, and retain the Manifest data link CID in memory.
  2. Instruct the FVM to preload all built-in actors and precompile Wasm modules, by supplying the manifest data Cid.

Point (2) is critical to avoid Wasm compilation costs at the upgrade epoch. If using [filecoin-project/ref-fvm], simply instantiating the Machine suppying the CID of the Manifest will trigger precompilation.

At this point, it is worthy to note that these block will be orphan, so garbage-collecting mechanisms may prune them, and walking the tree will not discover them (e.g. during snapshotting).

Upgrade procedure

At the upgrade epoch N, clients should do the following:

  1. Re-import the CARv1 bundle, to restore any bytecode blocks that may have been pruned.
  2. Perform a state migration.
  3. Atomically switch execution to the non-programmable FVM.
  4. Perform any message replacements needed to adjust gas limits on pending messages.

Points 2-4 are developed further in the next sections.

State migration

At epoch N, prior to executing the tipset assembled at epoch N-1, clients will run a migration over the state tree. The migration will perform two tasks:

  1. Patch the CodeCID of every actor will be by their corresponding content-addressed CIDs, according to this lookup table:
Old value (synthetic CID) New value (content-addressed CID)
fil/7/system TBD
fil/7/init TBD
fil/7/cron TBD
fil/7/account TBD
fil/7/storagepower TBD
fil/7/storageminer TBD
fil/7/storagemarket TBD
fil/7/paymentchannel TBD
fil/7/multisig TBD
fil/7/reward TBD
fil/7/verifiedregistry TBD
  1. Create and initialize a new state object for the System Actor (f00), and link it from the actor's entry in the state tree:
struct State {
  builtin_actor_registry: Cid, // Map<String, Cid>
}

The value of builtin_actor_registry is the CID of the Manifest imported from the bundle.

VM atomic switch

Exactly at the upgrade epoch N, message execution will switch from the legacy VM to the non-programmable FVM, for all system processes. That includes block validation, block production, gas estimation, and message publishing. JSON-RPC calls (like StateComputeTipset in the case of Lotus) will also begin using the new VM runtime.

In practice, this implies that blocks assembled by producers in epoch N-1 with the outgoing VM will be executed with the incoming VM at epoch N.

Following one finality (900 epochs), the decommissioning of the legacy VM will be considered final.

Message replacements

This FIP is likely to be bundled with FIP-0032 within the same network upgrade. FIP-0032 introduces gas changes by revising the gas model to accurately account for the execution costs under the FVM environment.

Pending messages awaiting in the mempool (whose gas limit was estimated with the legacy VM) are unlikely to be included with the FVM, since gas usage rules will have been altered.

To avoid disruption, clients must re-estimate gas for pending messages they are tracking, and re-submit them through the message replacement mechanism.

Optional: Legacy VM removal

After the upgrade epoch is over and one finality is attained, Filecoin clients may choose to complete erase the current VM from their respective codebases.

Because the canonical actors codebase supports only actors v6+, in doing so they will lose the ability to sync past portions of the chain. This may be an undesirable loss of functionality depending on the desired UX of every client, hence why this is entirely optional.

Alternatively, Filecoin clients may provide historical chain support by preserving the legacy VM in their codebases but activating it only through optional compilation (e.g. using Go build flags, or Rust Cargo features).

Design Rationale

Design rationale was integrated in the specification section.

Backwards Compatibility

Structure of Code CIDs

Altering the structure of the Code CID of built-in actors may have visible consequences for users:

  1. When constructing user-instantiable actors directly though the Init Actor (i.e. multisig or payment channel actors), users will have to specify unpredictable content-addressed Code CIDs instead of predictable synthetic CIDs.
  2. When querying an actor in the state tree (e.g. via the StateGetActor JSON-RPC API in Lotus), the Code field will no longer follow a structured pattern. This may break applications that parse Code CIDs.

A possible solution is to implement a JSON-RPC operation to query the CodeCID of a built-in actor. However, the Filecoin community was polled on this topic, and no concerns were raised. Thus, no action is proposed at this time.

Actor evolution

Pre-FVM, actor logic could change without its version being affected and, therefore, without the associated Code CIDs changing in the state tree. This is because the actor version (e.g. actors v6) represented the version of the ABI, not of the actor's logic.

However, with truly content-addressed CIDs, any and every change in actor logic will result in different bytecode. In practice this implies that state tree migrations will be more frequent.

There is no action to take in this FIP, but it is worth noting the difference in change management dynamics that this FIP will bring on.

Test Plan

This section offers a comprehensive test plan for this FIP. It harnesses multiple testing techniques to attain high confidence on this FIP across various facets. Some test efforts actually target the prerequisite FIP-0030 - Introducing the Filecoin Virtual Machine (which this FIP activates), and not direct aspects of this FIP.

  • Test vectors for mainnet equivalence up until nv15, inclusive

    • FVM-level conformance: FVM implementations should directly pass the test vector corpus in filecoin-project/fvm-test-vectors, when vectors are fed at the Machine level.
    • Client-level conformance: Clients integrating the FVM should pass the test vector corpus in filecoin-project/fvm-test-vectors, when vectors are fed at the client level.

  • Back-testing mainnet

    • Client implementations adopting the FVM should be capable of syncing mainnet the following mainnet chain range:
      • Start: epoch 1231620 (Chocolate upgrade, with activation of nv14 and actors v6)
      • End: most recent epoch following the OhSnap upgrade (nv15 / actors v7).
    • Implementers’ note: This test can be performed by:
      1. Obtaining the minimal snapshot for epoch 1233360, which includes the state trees and chain objects from 1802 epochs back (this number is the equivalent of 2 finalities + 2 epochs for buffer), thus covering the Chocolate upgrade at 1231620. Link to snapshot.
      2. Loading it into the client.
      3. Rewinding the client’s chain head to the cited start epoch.

  • Boosting Rust actors unit/integration test coverage

    • Compared to the outgoing canonical actors (Go specs-actors), Rust actors have poorer test coverage.
    • In order to not deteriorate the quality of software and enable the rapid development of changes, we must strengthen the test coverage of Rust actors to be on par with Go actors (at least), prior to appointing them as the canonical actors.

  • Live syncing mainnet as validators (mainnet shadow tests)

    • Client implementations adopting the FVM should have no problem keeping up with the live chain as it advances. Validation times should be within pre-FVM orders of magnitude.
    • Implementers’ note: this test can only be performed with a version of the FVM that does not implement this FIP, and therefore is equivalent to mainnet.

  • Servicing block producers

    • Client implementations adopting the FVM should have no problems servicing block producers. This entails serving the_correct_ mining base upon winning a round.
    • Implementers’ note: Verification may be performed in mainnet shadow tests, and/or in testnets (e.g. Calibrationnet).

  • Mainnet upgrade drills

    • Client implementations should perform the upgrade against live mainnet state. The migration must not disrupt block production at the upgrade epoch.
    • Implementers’ note: The filecoin-project/ent tool may come in handy.

  • Coordinated testnets

    • Client implementations wishing to join the network upgrade deploying this FIP should join the coordinated testnets plan outlined below.

Coordinated testnets

Testnets are a critical instrument to test and verify the behavior of implementations under different network conditions, some of which mimic mainnet.

Below is a tentative testnet deployment plan proposed by the maintainers of the reference implementation (Lotus). We encourage all implementations who have confirmed their ability to join the mainnet upgrade that activates this FIP to also join the testnet efforts.

FIP-0031 testnet plan

Per-implementation private testnets

All implementations should carry out private testing prior to joining the shared testnets below.

For illustrative purposes, here is the test plan of the reference implementation (Lotus + ref-fvm). Implementations may adopt a similar blueprint. This plan relies on the creation of a private network for rapid iteration named "Caterpillarnet", born with Lotus + mainnet-compatible FVM.

Test Phase 1 (runtime: 5 days) => Caterpillarnet with mainnet-compatible FVM:

  • Objective: rapid continuous sampling of FVM behavior in a brand new network.
  • Network parameters:
    • Block time: 5s.
    • Sector sizes: 512MiB, 32GiB, 64GiB.
    • Minimum number of reference implementation (Lotus) block producers/storage providers: 6.
    • Consensus: Expected Consensus with fake winning PoSt (in order to speed up block production).
  • Butterflynet keeps running, accumulating state in preparation for the Test Phase 2 migration.

Test Phase 2 (runtime: 2 days) => Caterpillarnet with FIP-0031 (this FIP):

  • Objective: validation of FIP-0031 upgrade procedure and basic migration.
  • Network parameters:
    • Block time: 30s.
    • Sector sizes: 512MiB, 32GiB, 64GiB.
    • Minimum number of reference implementation (Lotus) block producers/storage providers: 6.
    • Consensus: Expected Consensus.
  • Caterpillarnet goes through the FIP-0031 upgrade (switching to canonical actors and adopting content-addressed CodeCIDs).

Cross-implementation shared testnets

From here onwards, the test plan specifies collaborative testing phases.

Test Phase 3 (runtime: 1-2 weeks) => Butterflynet upgrade to network v16, including FIP-0031 (this FIP):

  • Objective: validation of FIP-0031 upgrade procedure, collecting observations such as time taken, physical state growth, IO workload, and more.
  • Butterflynet goes through the upgrade including the protocol changes introduced by FIP-0031 (switching to canonical actors and adopting content-addressed CodeCIDs).
  • Community members are encouraged to join Butterflynet prior to the network upgrade happening.
  • Based on the testing feedbacks and discoveries, the runtime may vary, however, for monitoring purpose we shall run nv16 on butterfly for at least one week before deploying nv16 to calibrationet.
  • Butterflynet might be reset if consensus bug is discovered during testing and auditing.

Test Phase 4 (runtime: 3 weeks) => Calibrationnet upgrade to network v16, release candidate testing and stabilization.

  • Butterflynet drills as we stabilize release candidates and release the final RC (1.16.0-rcN) for deployment on Calibrationnet.
  • Ongoing testing and monitoring.

Security Considerations

With the adoption of common codebases across all clients (ref-fvm and canonical actors), any bugs in them will threaten the stability of the entire network. These codebases also carry a novelty factor that should be mitigated via a collection of hardening and auditing actions.

Incentive Considerations

N/A.

Product Considerations

Besides the change in Code CIDs explained in previous sections, there are no other product considerations emerging from this FIP alone.

Implementation

This FIP will be implemented in the reference Filecoin client implementation: Lotus.

Credits

Thanks to @jennijuju for graphic work.

Copyright

Copyright and related rights waived via CC0.

Footnotes

  1. Originally forked from the Forest client implementation. ↩