Skip to content

Commit

Permalink
Merge pull request #898 from CosmWasm/document-execution-semantics
Browse files Browse the repository at this point in the history 
Document contract execution semantics
  • Loading branch information
@ethanfrey
ethanfrey authored Apr 27, 2021
2 parents c54fe43 + 2247d41 commit 7f73491
Showing 1 changed file with 370 additions and 0 deletions.
370 changes: 370 additions & 0 deletions SEMANTICS.md
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,370 @@
# Contract Semantics

This document aims to clarify the semantics of how a CosmWasm contract interacts
with its environment. There are two main types of actions: _mutating_ actions,
which receive `DepsMut` and are able to modify the state of the blockchain, and
_query_ actions, which are run on a single node with read-only access to the
data.

## Execution

In the section below, we will discuss how the `execute` call works, but the same
semantics apply to any other _mutating_ action - `instantiate`, `migrate`,
`sudo`, etc.

### SDK Context

Before looking at CosmWasm, we should look at the (somewhat under-documented)
semantics enforced by the blockchain framework we integrate with - the
[Cosmos SDK](https://v1.cosmos.network/sdk). It is based upon the
[Tendermint BFT](https://tendermint.com/core/) Consensus Engine. Let us first
look how they process transactions before they arrive in CosmWasm (and after
they leave).

First, the Tendermint engine will seek 2/3+ consensus on a list of transactions
to be included in the next block. This is done _without executing them_. They
are simply subjected to a minimal pre-filter by the Cosmos SDK module, to ensure
they are validly formatted transactions, with sufficient gas fees, and signed by
an account with sufficient fees to pay it. Notably, this means many transactions
that error may be included in a block.

Once a block is committed (typically every 5s or so), the transactions are then
fed to the Cosmos SDK sequentially in order to execute them. Each one returns a
result or error along with event logs, which are recorded in the `TxResults`
section of the next block. The `AppHash` (or merkle proof or blockchain state)
after executing the block is also included in the next block.

The Cosmos SDK `BaseApp` handles each transaction in an isolated context. It
first verifies all signatures and deducts the gas fees. It sets the "Gas Meter"
to limit the execution to the amount of gas paid for by the fees. Then it makes
an isolated context to run the transaction. This allows the code to read the
current state of the chain (after the last transaction finished), but it only
writes to a cache, which may be committed or rolled back on error.

A transaction may consist of multiple messages and each one is executed in turn
under the same context and same gas limit. If all messages succeed, the context
will be committed to the underlying blockchain state and the results of all
messages will be stored in the `TxResult`. If one message fails, all later
messages are skipped and all state changes are reverted. This is very important
for atomicity. That means Alice and Bob can both sign a transaction with 2
messages: Alice pays Bob 1000 ATOM, Bob pays Alice 50 ETH, and if Bob doesn't
have the funds in his account, Alice's payment will also be reverted. This is
just like a DB Transaction typically works.

[`x/wasm`](https://github.com/CosmWasm/wasmd/tree/master/x/wasm) is a custom
Cosmos SDK module, which processes certain messages and uses them to upload,
instantiate, and execute smart contracts. In particular, it accepts a properly
signed
[`MsgExecuteContract`](https://github.com/CosmWasm/wasmd/blob/master/proto/cosmwasm/wasm/v1beta1/tx.proto#L76-L89),
routes it to
[`Keeper.Execute`](https://github.com/CosmWasm/wasmd/blob/master/x/wasm/keeper/keeper.go#L311-L355),
which loads the proper smart contract and calls `execute` on it. Note that this
method may either return a success (with data and events) or an error. In the
case of an error here, it will revert the entire transaction in the block. This
is the context we find ourselves in when our contract receives the `execute`
call.

### Basic Execution

When we implement a contract, we provide the following entry point:

```rust
pub fn execute(
deps: DepsMut,
env: Env,
info: MessageInfo,
msg: ExecuteMsg,
) -> Result<Response, ContractError> { }
```

With `DepsMut`, this can read and write to the backing `Storage`, as well as use
the `Api` to validate addresses, and `Query` the state of other contracts or
native modules. Once it is done, it returns either `Ok(Response)` or
`Err(ContractError)`. Let's examine what happens next:

If it returns `Err`, this error is converted to a string representation
(`err.to_string()`), and this is returned to the SDK module. _All state changes
are reverted_ and `x/wasm` returns this error message, which will _generally_
(see submessage exception below) abort the transaction, and return this same
error message to the external caller.

If it returns `Ok`, the `Response` object is parsed and processed. Let's look at
the parts here:

```rust
pub struct Response<T = Empty>
where
T: Clone + fmt::Debug + PartialEq + JsonSchema,
{
/// Optional list of "subcalls" to make. These will be executed in order
/// (and this contract's subcall_response entry point invoked)
/// *before* any of the "fire and forget" messages get executed.
pub submessages: Vec<SubMsg<T>>,
/// After any submessages are processed, these are all dispatched in the host blockchain.
/// If they all succeed, then the transaction is committed. If any fail, then the transaction
/// and any local contract state changes are reverted.
pub messages: Vec<CosmosMsg<T>>,
/// The attributes that will be emitted as part of a "wasm" event
pub attributes: Vec<Attribute>,
pub data: Option<Binary>,
}
```

In the Cosmos SDK, a transaction returns a number of events to the user, along
with an optional data "result". This result is hashed into the next block hash
to be provable and can return some essential state (although in general client
apps rely on Events more). This result is more commonly used to pass results
between contracts or modules in the sdk. Note that the `ResultHash` includes
only the `Code` (non-zero meaning error) and `Result` (data) from the
transaction. Events and log are available via queries, but there are no
light-client proofs possible.

If the contract sets `data`, this will be returned in the `result` field.
`attributes` is a list of `{key, value}` pairs which will be
[appended to a default event](https://github.com/CosmWasm/wasmd/blob/master/x/wasm/types/types.go#L302-L321).
The final result looks like this to the client:

```json
{
"type": "wasm",
"attributes": [
{ "key": "contract_addr", "value": "cosmos1234567890qwerty" },
{ "key": "custom-key-1", "value": "custom-value-1" },
{ "key": "custom-key-2", "value": "custom-value-2" }
]
}
```

### Dispatching Messages

Now let's move onto the `messages` field. Some contracts are fine only talking
with themselves, such as a cw20 contract just adjusting its balances on
transfers. But many want to move tokens (native or cw20) or call into other
contracts for more complex actions. This is where messages come in. We return
[`CosmosMsg`, which is a serializable representation](https://github.com/CosmWasm/cosmwasm/blob/v0.14.0-beta4/packages/std/src/results/cosmos_msg.rs#L18-L40)
of any external call the contract can make. It looks something like this (with
`stargate` feature flag enabled):

```rust
pub enum CosmosMsg<T = Empty>
where
T: Clone + fmt::Debug + PartialEq + JsonSchema,
{
Bank(BankMsg),
/// This can be defined by each blockchain as a custom extension
Custom(T),
Staking(StakingMsg),
Distribution(DistributionMsg),
Stargate {
type_url: String,
value: Binary,
},
Ibc(IbcMsg),
Wasm(WasmMsg),
}
```

If a contract returns two messages - M1 and M2, these will both be parsed and
executed in `x/wasm` _with the permissions of the contract_ (meaning
`info.sender` will be the contract not the original caller). If they return
success, they will emit a new event with the custom attributes, the `data` field
will be ignored, and any messages they return will also be processed. If they
return an error, the parent call will return an error, thus rolling back state
of the whole transaction.

Note that the messages are executed _depth-first_. This means if contract A
returns M1 (`WasmMsg::Execute`) and M2 (`BankMsg::Send`), and contract B (from
the `WasmMsg::Execute`) returns N1 and N2 (eg. `StakingMsg` and
`DistributionMsg`), the order of execution would be **M1, N1, N2, M2**.

This may be hard to understand at first. "Why can't I just call another
contract?", you may ask. However, we do this to prevent one of most widespread
and hardest to detect security holes in Ethereum contracts - reentrancy. We do
this by following the actor model, which doesn't nest function calls, but
returns messages that will be executed later. This means all state that is
carried over between one call and the next happens in storage and not in memory.
For more information on this design, I recommend you read
[our docs on the Actor Model](https://docs.cosmwasm.com/0.13/architecture/actor.html).

### Submessages

As of CosmWasm 0.14 (April 2021), we have added yet one more way to dispatch
calls from the contract. A common request was the ability to get the result from
one of the messages you dispatched. For example, you want to create a new
contract with `WasmMsg::Instantiate`, but then you need to store the address of
the newly created contract in the caller. With `submessages`, this is now
possible. It also solves a similar use-case of capturing the error results, so
if you execute a message from eg. a cron contract, it can store the error
message and mark the message as run, rather than aborting the whole transaction.
It also allows for limiting the gas usage of the submessage (this is not
intended to be used for most cases, but is needed for eg. the cron job to
protect it from an infinite loop in the submessage burning all gas and aborting
the transaction).

This makes use of `CosmosMsg` as above, but it wraps it inside a `SubMsg`
envelope:

```rust
pub struct SubMsg<T = Empty>
where
T: Clone + fmt::Debug + PartialEq + JsonSchema,
{
pub id: u64,
pub msg: CosmosMsg<T>,
pub gas_limit: Option<u64>,
pub reply_on: ReplyOn,
}

pub enum ReplyOn {
/// Always perform a callback after SubMsg is processed
Always,
/// Only callback if SubMsg returned an error, no callback on success case
Error,
/// Only callback if SubMsg was successful, no callback on error case
Success,
}
```

What are the semantics of a submessage execution. First, we create a
sub-transaction context around the state, allowing it to read the latest state
written by the caller, but write to yet-another cache. If `gas_limit` is set, it
is sandboxed to how much gas it can use until it aborts with `OutOfGasError`.
This error is caught and returned to the caller like any other error returned
from contract execution (unless it burned the entire gas limit of the
transaction). What is more interesting is what happens on completion.

If it return success, the temporary state is committed (into the caller's
cache), and the `Response` is processed as normal (an event is added to the
current EventManager, messages and submessages are executed). Once the
`Response` is fully processed, this may then be intercepted by the calling
contract (for `ReplyOn::Always` and `ReplyOn::Success`). On an error, the
subcall will revert any partial state changes due to this message, but not
revert any state changes in the calling contract. The error may then be
intercepted by the calling contract (for `ReplyOn::Always` and
`ReplyOn::Error`). _In this case, the messages error doesn't abort the whole
transaction_

#### Handling the Reply

In order to make use of `submessages`, the calling contract must have an extra
entry point:

```rust
#[entry_point]
pub fn reply(deps: DepsMut, env: Env, msg: Reply) -> Result<Response, ContractError> { }

pub struct Reply {
pub id: u64,
/// ContractResult is just a nicely serializable version of `Result<SubcallResponse, String>`
pub result: ContractResult<SubcallResponse>,
}

pub struct SubcallResponse {
pub events: Vec<Event>,
pub data: Option<Binary>,
}
```

After the `submessage` is finished, the caller will get a chance to handle the
result. It will get the original `id` of the subcall so it can switch on how to
process this, and the `Result` of the execution, both success and error. Note
that it includes all events returned by the submessage, which applies to native
sdk modules as well (like Bank) as well as the data returned from below. This
and the original call id provide all context to continue processing it. If you
need more state, you must save some local context to the store (under the `id`)
before returning the `submessage` in the original `execute`, and load it in
`reply`. We explicitly prohibit passing information in contract memory, as that
is the key vector for reentrancy attacks, which are a large security surface
area in Ethereum.

The `reply` call may return `Err` itself, in which case it is treated like the
caller errored, and aborting the transaction. However, on successful processing,
`reply` may return a normal `Response`, which will be processed as normal -
events added to the EventManager, and all `messages` and `submessages`
dispatched as described above.

The one _critical difference_ with `reply`, is that we _do not drop data_. If
`reply` returns `data: Some(value)` in the `Response` object, we will overwrite
the `data` field returned by the caller. That is, if `execute` returns
`data: Some(b"first thought")` and the `reply` (with all the extra information
it is privy to) returns `data: Some(b"better idea")`, then this will be returned
to the caller of `execute` (either the client or another transaction), just as
if the original `execute` and returned `data: Some(b"better idea")`. If `reply`
returns `data: None`, it will not modify any previously set data state. If there
are multiple submessages all setting this, only the last one is used (they all
overwrite any previous `data` value).

#### Order and Rollback

Submessages (and their replies) are all executed before any `messages`. They
also follow the _depth first_ rules as with `messages`. Here is a simple
example. Contract A returns submessages S1 and S2, and message M1. Submessage S1
returns message N1. The order will be: **S1, N1, reply(S1), S2, reply(S2), M1**

Please keep in mind that submessage `execution` and `reply` can happen within
the context of another submessage. For example
`contract-A--submessage --> contract-B--submessage --> contract-C`. Then
`contract-B` can revert the state for `contract-C` and itself by returning `Err`
in the submessage `reply`, but not revert contract-A or the entire transaction.
It just ends up returning `Err` to contract-A's `reply` function.

Note that errors are not handled with `ReplyOn::Success`, meaning, in such a
case, an error will be treated just like a normal `message` returning an error.
This diagram may help explain. Imagine a contract returned two submesssage - (a)
with `ReplyOn::Success` and (b) with `ReplyOn::Error`:

| processing a) | processing b) | reply called | may overwrite result from reply | note |
| ------------- | ------------- | ------------ | ------------------------------- | ------------------------------------------------- |
| ok | ok | a) | a) | returns success |
| err | err | none | none | returns error (abort parent transaction) |
| err | ok | none | none | returns error (abort parent transaction) |
| ok | err | a)b) | a)b) | if both a) and b) overwrite, only b) will be used |

## Query Semantics

Until now, we have focused on the `Response` object, which allows us to execute
code in other contracts via the actor model. That is, each contract is run
sequentially, one after another, and no nested calls are possible. This is
essential to avoid reentrancy, which is when calling into another contract can
change my state while I am in the middle of a transaction.

However, there are many times we need access to information from other contracts
in the middle of processing, such as determining the contract's bank balance
before sending funds. To enable this, we have exposed the _read only_ `Querier`
to enable _synchronous_ calls in the middle of the execution. By making it
read-only (and enforcing that in the VM level), we can prevent the possibility
of reentrancy, as the query cannot modify any state or execute our contract.

When we "make a query", we serialize a
[`QueryRequest` struct](https://github.com/CosmWasm/cosmwasm/blob/v0.14.0-beta4/packages/std/src/query/mod.rs#L27-L48)
that represents all possible calls, and then pass that over FFI to the runtime,
where it is interpretted in the `x/wasm` SDK module. This is extensible with
blockchain-specific custom queries just like `CosmosMsg` accepts custom results.
Also note the ability to perform raw protobuf "Stargate" queries:

```rust
pub enum QueryRequest<C: CustomQuery> {
Bank(BankQuery),
Custom(C),
Staking(StakingQuery),
Stargate {
/// this is the fully qualified service path used for routing,
/// eg. custom/cosmos_sdk.x.bank.v1.Query/QueryBalance
path: String,
/// this is the expected protobuf message type (not any), binary encoded
data: Binary,
},
Ibc(IbcQuery),
Wasm(WasmQuery),
}
```

While this is flexible and needed encoding for the cross-language
representation, this is a bit of mouthful to generate and use when I just want
to find my bank balance. To help that, we often use
[`QuerierWrapper`](https://github.com/CosmWasm/cosmwasm/blob/v0.14.0-beta4/packages/std/src/traits.rs#L148-L314),
which wraps a `Querier` and exposes a lot of convenience methods that just use
`QueryRequest` and `Querier.raw_query` under the hood.

You can read a longer explanation of the
[`Querier` design in our docs](https://docs.cosmwasm.com/0.13/architecture/query.html).

0 comments on commit 7f73491

Please sign in to comment.