Skip to content

spacemeshos/go-svm

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

go-svm

https://img.shields.io/badge/godoc-go-svm-5272B4.svg

The Golang client for SVM. Its primary goal is supplying an ergonomic API for go-spacemesh

Installation

git clone https://github.com/spacemeshos/go-svm
cd go-svm
make
make test # Optional. Rerun if you want to test any changes to `go-svm`.
make install

Structs

Transaction

Each binary (over-the-wire) transaction will always have two parts:

  • Envelope - A Transaction agnostic content.
  • Message - Transaction-specific content.

Each Message will be preceded by the Envelope - together; both make a complete binary Transaction.

In addition to the data sent over the wire, there will be implicit fields inferred from it. One such example is the Transaction Id. That field is part of the Context described later. The computation of the TransactionId will be done externally to go-svm(i.e., go-spacemesh).

There are in total three types of transactions under SVM:

  • Deploy - For deploying Templates (see Deploying a Template later).
  • Spawn - For spawning new accounts out of existing Templates (see Spawning an Account later).
  • Call - For calling an existing account (see Calling an Account later).

Envelope

The Envelope contains pieces of data that are part of any transaction. When the Full-Node (e.g., go-spacemesh) receives a binary transaction from the network, it needs to decode the Envelope part into a Golang struct.

The other part of the Transaction, a.ka the Message, should be kept as []byte It's the job of SVM to decode the transaction Message. More information about each Message type appears later on this document.

An Envelope will contain the following fields:

  • Type - The transaction type (Deploy / Spawn / Call)
  • Principal - The Address of the Account paying for the Gas.
  • Amount - For funding the target account (relevant only for Spawn/Call transactions).
  • TxNonce - The Transaction's nonce.
  • GasLimit - Maximum units of Gas to be paid.
  • GasFee - Fee per Unit of Gas.

And this is the corresponding Golang struct:

type Envelope struct {
	Type      TxType    // Alias for `uint8`
	Principal Address   // Alias for `[20]byte`
	Amount    Amount    // Alias for `uint64`
	TxNonce   TxNonce   // A struct holding a pair of `uint64` (Golang has no `uint128` primitive)
	GasLimit  Gas       // Alias for `uint64`
	GasFee    GasFee    // Alias for `int`
}

To create a new Envelope, use this helper function:

func NewEnvelope(principal Address, amount Amount, txNonce TxNonce, gasLimit Gas, gasFee GasFee) *Envelope

Message

A Message is essentially a blob of bytes. Each go-svm API expecting a Message will ask for it in its binary form (i.e. [] byte). It's the job of SVM itself to decode a binary Message and figure out what's inside.

There are in total three types of transactions under SVM - each with its corresponding Message:

  • Deploy Message - The Message of a Deploy transaction.
  • Spawn Message - The Message of a Spawn transaction.
  • Call Message - The Message of a Call transaction.

Each Transaction is detailed later in this document.

Context

In addition to the Transaction, there is the Execution Context (or simply Context). The Context structure will contain additional data (alongside the Transaction) to be used by SVM when executing a transaction.

It will contain data relevant to the currently executing Context within the Full-Node (i.e., go-spacemesh) and properties computed from the Transaction itself. (such as the Transaction Id). It's the role of the Full-Node (e.g., go-spacemesh) to create a Context instance and pass it forward to go-svm.

Here is the current declaration of a Context

type Context struct {
	Layer Layer  // The `Layer` (alias for `uint64`) we're about to execute the Transaction in.
	TxId  TxId   // The computed `Transaction Id` out of the `Transaction` data (`TxId` is an alias for `[32]byte`).
}

For Creating a new Context, use this helper:

func NewContext(layer Layer, txId TxId) *Context

Deploy Message

A Deploy Message will be generated using the Template Toolchain

By saying a Template Toolchain, we mean the process of:

  • Compiling the Template code and emitting binary Wasm and Metadata files. Currently, the only way to generate such Wasm is by writing Rust code using the [SVM SDK](https://github.com/spacemeshos/svm/tree/master/crates/sdk) crate.

Here is a link for such an example Template: (execute the build.sh for compiling into Wasm) https://github.com/spacemeshos/svm/tree/master/crates/runtime/tests/wasm/calldata

  • Utilizing the SVM CLI for crafting a binary Deploy transaction.

Here is CLI usage for the generation of a binary Spawn message:

svm-cli craft-deploy --smwasm Template.wasm --meta Template-meta.json --output template.svm

In the future, there might be other alternatives to achieve the above. If Spacemesh has its Smart-Contracts programming language in the future, it'll make sense to let that language compiler take care of everything. In such a case, the output will be a Deploy Message. From here, filling in the missing parts (Envelope and signing the Transaction) should be the same solution used today for the SVM SDK and SVM CLI.

Spawn Message

Each Spawn Message contains the following fields:

  • template (Template Address) - The Template we'll spawn an account of.
  • name (String) - The name of the Account (optional).
  • ctor (String) - The constructor identifier to execute.
  • calldata (Blob) - The input for the constructor to run.

Generating a binary Spawn Message can be achieved in two ways using the SVM CLI and SVM Codec

SVM CLI

Here is an example for a Spawn JSON

{
  "version": 0,
  "template": "b5eba98957e6a93173ffb50207cceeedfddb1a72",
  "name": "My Account",
  "ctor_name": "initialize",
  "calldata": {
    "abi": ["address", "bool"],
    "data": ["8f20ed1a0e342c2a75b1b3f8014545dd3d886078", true]
  }
}

In order to turn it into a binary Spawn Message using the CLI execute:

svm-cli tx --tx-type=spawn --input=tx.json --output=tx.bin

Using the CLI is very useful for tests inputs generation.

SVM Codec

The SVM project ships with an artifact called svm_codec.wasm. That Wasm package could be used for encoding a transaction Message. There has been implemented an npm package for interfacing against that Wasm package.

svm-codec-npm:

https://github.com/spacemeshos/svm-codec-npm

This npm package will be consumed by smapp or the Process Explorer. Similarly, new clients could be added in the future (for example, a Golang client to be used by smrepl)

Call Message

Each Call Message contains the following fields:

  • target (Account Address) - The Address of the Account which we're calling.
  • function (String) - The function's name to execute.
  • verifydata (Blob) - The input for the svm_verify function.
  • calldata (Blob) - The input for the function to run.

In a very similar manner to the Spawn - we can generate a binary Call Message given a JSON. And the same information about the svm_codec.wasm applies here as well.

Here is an example for a Call Message given as a JSON:

{
  "version": 0,
  "target": "066818abe361dd44f425da19e17c45babc40e232",
  "func_name": "store_addr",
  "verifydata": {
    "abi": [],
    "data": []
  },
  "calldata": {
    "abi": ["address"],
    "data": ["102030405060708090102030405060708090AABB"]
  }
}

To turn it into a binary Call Message using the CLI execute:

svm-cli tx --tx-type=call --input=tx.json --output=tx.bin

High-level API

Init

Init is the entry point for interacting with SVM in any way. It runs internal initialization logic; it is fully thread-safe and idempotent.

func Init() (*API, error)

Creating a Runtime

Creates a new SVM Runtime. You can think of it as opening a connection to SVM. Please make sure to call Init (see above) first.


Here is the `Create Runtime` API:
func (*API) NewRuntime() (*Runtime, error)

Destroying a Runtime

When the usage of a Runtime is over, we need to release its resources. You can think of it as closing a connection.


And here is the `Destroy Runtime` API:
func (rt *Runtime) Destroy()

Verifying a Transaction

Performs the verify stage as dictated by the Account Unification design. Since the verify flow involves the running Wasm function as done when running a Call transaction, the output will also be of type CallReceipt.


This is the relevant API to be used:
func (rt *Runtime) Verify(env *Envelope, msg []byte, ctx *Context) (*CallReceipt, error)

Starting a new Layer

Signaling SVM that we are about to start playing a list of transactions under the input layer Layer. The value of the Layer is expected to equal the last known committed/rewinded Layer plus one. Any other layer given as input will result in an error.


For starting a new `Layer`, use the following:
func (rt *Runtime) Open(layer Layer) error

Committing Layer changes

Commits SVM dirty changes. It also signals the termination of the current Layer. In other words, after finishing executing the layer transactions, we should call a Commit.


The matching API:
func (rt *Runtime) Commit() (Layer, State, error)

If the Commit went out fine, it would return a tuple consisting of:

  • The Layer we have just committed.
  • The newly computed Global State Root Hash
  • Setting nil under the error

If the Commit errored, then the output will be:

  • The Layer we have just tried to commit (but have failed)
  • nil under the State position.
  • The error that occurred.

Rewinding State

Rewinds SVM Global State to the given Layer. This capability is necessary for self-healing.


Here is the API for rewind:
func (rt *Runtime) Rewind(layer Layer) (State, error)

If the rewind succeeds, it returns the Global-State Root Hash at that given point. (the error returned will be assigned with nil) Otherwise, a nil will be placed under the State position, and the 2nd tuple element will contain the error that occurred.

Retrieving an Account

Given an Account Address - retrieves its most basic information encapsulated within an Account struct.


Here is the API to be used for retrieving an account:
func (rt *Runtime) GetAccount(addr Address) (Account, error)

And this is the definition of an Account at go-svm:

type Account struct {
	Addr    Address  // The `Address` of the account
	Balance Amount   // The account's balance (`Amount` is an alias for `uint64`)
	Counter TxNonce // The account's counter. It's a struct holding a pair of `uint64` (Golang has no `uint128` primitive)
}

Increasing an Account's Balance

Increases an account's balance. The motivation for that API was supporting Rewards


The API for increasing an account's balance:
func (rt *Runtime) IncreaseBalance(addr Address, amount Amount)

TODO: What should be the behavior of go-svm when there is no account with the given Address?

Deploying a Template

Deploying a Template exposes two dedicated APIs: ValidateDeploy and Deploy.

Validate Deploy

Syntactically validates the Deploy Message given in a binary form and returns whether it's valid or not.


The API for validation:
func (rt *Runtime) ValidateDeploy(msg []byte) (bool, error)

Deploy

Performs the actual deployment of a Template and returns a DeployReceipt.


The Deploy API:

func (rt *Runtime) Deploy(env *Envelope, msg []byte, ctx *Context) (*DeployReceipt, error)

That is the DeployReceipt definition:

type DeployReceipt struct {
	Success      bool            // Whether the Transaction succeeded or not
	Error        *RuntimeError   // Returns `nil` when `Success` is true and otherwise the runtime error that occurred
	TemplateAddr TemplateAddr    // The `Template Address` for the newly deployed template
	GasUsed      Gas             // The amount of `Gas` used during the transaction execution (in units of Gas)
	Logs         []Log           // Logs created as part of transaction execution
}

Spawning an Account

Performs the spawning of a new Account out of the existing Template. Similarly to Deploy - spawning a new Account exposes two dedicated APIs: ValidateSpawn and Spawn.

Validate Spawn

Syntactically validates the Spawn Message given in a binary form and returns whether it's valid or not.


The validation API:
func (rt *Runtime) ValidateSpawn(msg []byte) (bool, error)

Spawn

Performs the spawning of a new Account out of the existing Template and returns a SpawnReceipt.


The Spawn API:

func (rt *Runtime) Spawn(env *Envelope, msg []byte, ctx *Context) (*SpawnReceipt, error)

Here is the SpawnReceipt definition:

type SpawnReceipt struct {
	Success         bool            // Whether the Transaction succeeded or not
	Error           *RuntimeError   // Returns `nil` when `Success` is true and otherwise the runtime error that occurred
	AccountAddr     Address         // The `Address` for the newly spawned Account
	InitState       State           // The newly computed `Global-State Root Hash` after spawning the Account []byte          // The data returned by the constructor running during spawning the account
	GasUsed         Gas             // The amount of `Gas` used during the transaction execution (in units of Gas)
	Logs            []Log           // Logs created as part of transaction execution
	TouchedAccounts []Address       // A list of `Account Addresses` engaged in any at least a single coins-transfer during transaction execution
}

Calling an Account

Validate Call

Syntactically validates the Call Message given in a binary form and returns whether it's valid or not.


The validation API:
func (rt *Runtime) ValidateCall(msg []byte) (bool, error)

Call

Performs the actual calling an Account and returns a CallReceipt.


The Call API:

func (rt *Runtime) Call(env *Envelope, msg []byte, ctx *Context) (*CallReceipt, error)

Here is the CallReceipt definition:

type CallReceipt struct {
	Success         bool            // Whether the Transaction succeeded or not
	Error           *RuntimeError   // Returns `nil` when `Success` is true and otherwise the runtime error that occurred
	NewState        State           // The newly computed `Global-State Root Hash` after calling the Account []byte          // The data returned by calling the account
	GasUsed         Gas             // The amount of `Gas` used during the transaction execution (in units of Gas)
	Logs            []Log           // Logs created as part of transaction execution
	TouchedAccounts []Address       // A list of `Account Addresses` engaged in any at least a single coins-transfer during transaction execution
}

Tests helpers:

Runtimes Count

Returns the number of living SVM Runtimes

SVM was designed to execute transactions sequentially. It means that the number of existing Runtime instances should not exceed one. There are functions of SVM that could have been called in parallel (for example, validation) - it's not recommended at this stage to take extra caution and not do that. This helper function is intended to be used for testing purposes. However, it could be used for telemetry/tracing/debugging as well.

The API:

func (*API) RuntimesCount() int

Receipts Count

Returns the number of living Receipts returned by SVM It's the job of the go-svm internals to release binary Receipts returned by SVM

If there're no bugs, the reported living Receipt count should be zero after each transaction execution. The helper should be applied for testing purposes. However, the production code can log (with a fatal severity level) if this number somehow stops being zero.

The API:

func (*API) ReceiptsCount() int

Errors Count

Returns the number of internal errors returned by SVM. First, it's important to stress what we mean by saying an Error. When a transaction has failed due to panic or running out-of-gas - SVM needs to return a valid Receipt setting Success to false An error should be returned in the case that SVM itself panicked - this is undefined behavior.

We, of course, hope never to reach such a point since an internal error might occur only on one Operating-System. However, this will break the consensus. If we, unfortunately, did hit an internal error, we need to make sure the error data returned by SVM will be freed.

This helper function should be used for testing. It's up to the go-svm client to decide what to do in case an internal error is being returned. One way is to crash to process. Another alternative is to convert that error to Golang Receipt Struct and hope for the best. Turning the internal error to Receipt can ease debugging since

the Process Explorer will display that Receipt as well.

About

The Golang client for SVM

Resources

License

Code of conduct

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published