We audited the latticexyz/mud repository at commit 12a9eb1.
In scope were the following contracts:
packages
├── store
│ └── src
│ ├── Bytes.sol
│ ├── FieldLayout.sol
│ ├── Hook.sol
│ ├── Memory.sol
│ ├── PackedCounter.sol
│ ├── ResourceId.sol
│ ├── Schema.sol
│ ├── Slice.sol
│ ├── Storage.sol
│ ├── StoreCore.sol
│ ├── StoreData.sol
│ ├── StoreHook.sol
│ ├── StoreRead.sol
│ ├── StoreSwitch.sol
│ ├── constants.sol
│ ├── leftMask.sol
│ ├── storeHookTypes.sol
│ ├── storeResourceTypes.sol
│ └── tightcoder
│ └── TightCoder.sol
└── world
└── src
├── AccessControl.sol
├── Create2.sol
├── Create2Factory.sol
├── Delegation.sol
├── DelegationControl.sol
├── Module.sol
├── SystemCall.sol
├── SystemHook.sol
├── World.sol
├── WorldContext.sol
├── WorldFactory.sol
├── WorldResourceId.sol
├── constants.sol
├── modules
│ └── core
│ ├── CoreModule.sol
│ ├── constants.sol
│ └── implementations
│ ├── AccessManagementSystem.sol
│ ├── BalanceTransferSystem.sol
│ ├── BatchCallSystem.sol
│ ├── ModuleInstallationSystem.sol
│ ├── StoreRegistrationSystem.sol
│ └── WorldRegistrationSystem.sol
├── requireInterface.sol
├── revertWithBytes.sol
├── systemHookTypes.sol
├── version.sol
└── worldResourceTypes.sol
The MUD system provides a cohesive range of standard and extensible functionality to quickly develop blockchain applications.
The central component of the MUD system is the Store, which is a contract that behaves like a database. All persistent storage is presented as a set of tables, and the Store contract takes care of mapping this structure to the linear EVM storage layout. Importantly, the structure is not fixed at deployment time, which means developers can easily add new tables or any other functionality (described below) as their application evolves.
In addition, the Store contract allows multiple applications to provide different overlapping views of the same underlying data. This is achieved by standardizing the events that describe all state changes, so off-chain tooling can represent the storage in an application-agnostic way. In fact, some tables can be designated as off-chain tables, where they will emit the standard event stream on storage changes but the actual data will not be saved on-chain and must be reconstructed from these events.
Tables can also include hooks. These are arbitrary code snippets that can be executed whenever a record is added, removed or modified. This would typically be used for data validation, or to ensure consistency within a database record.
Although the Store contract provides the basic database functionality, it will typically be extended to provide higher-level abstractions for users and developers. One such extension is the World contract (in conjunction with the CoreSystem contract) which provides generic and flexible mechanisms for many standard features.
For example, the World contract registers tables in the Store contract to track access control relationships for all the other tables. Specifically, the database is segmented into different namespaces each of which has an owner address. The namespace owner can create new tables and manage which addresses can modify which tables in the namespace.
In addition, the namespace owner can register systems which are arbitrary contracts that are granted access to the namespace. These would typically be used for application-specific logic, to modify multiple records and tables in the same transaction, and to do so in a consistent way. They can even register system hooks (arbitrary code snippets) to run before or after the system is invoked. Moreover, there are some convenience functions which can be used to invoke systems more easily or to make multiple calls in a single transaction. The namespace owner also controls which addresses can invoke which systems.
The World contract also provides a delegation mechanism whereby any address can allow other addresses to perform actions on its behalf. If desired, the delegation can be managed by an arbitrary contract, so approval can be granted or rejected based on the details of the action and the particular delegatee. The namespace owner can also provide a fallback namespace delegation contract that manages all delegations for systems within the namespace.
The ETH held by the World contract is partitioned among the namespaces. This provides the systems in the namespace with a convenient way to share the same balance. Naturally, each individual system contract can also maintain a private ETH balance if that is desired.
Lastly, there is a root namespace which can be used to register privileged systems with complete control over the storage. In this way, the behavior of any deployed World contract is upgradeable and customizable.
The codebase is intended to be deployed by application developers. Depending on the use case, an application might require its own World or it could use an existing World instance, possibly with its own namespace for encapsulation. In either case, all Worlds are independent and the Lattice team does not retain any privileged access over them.
In the standard use case, the World deployer will initialize the system to become the owner of the root namespace, the world namespace, and the store namespace. They will then create the tables and records required to implement the features described above.
Anyone can register a new namespace and become its owner. This will let them manage the access control requirements (including transferring and renouncing ownership) and create new tables, systems, and hooks within the namespace.
However, it is worth noting that the root namespace owner can install new privileged systems or replace the CoreSystem functionality which implicitly gives them control over all the resources. Therefore, unless they renounce their power, they should be trusted by all the World users.
In the interest of efficiency, the system intentionally bypasses many data integrity checks. For example, users can set a table field to a value that is larger than what the type supports, which would overflow into subsequent fields, possibly beyond the allocated space. Similarly, users do not need to follow the expected key format when referencing records which could lead to them storing values at "invalid" pseudorandom locations. Users who have been granted write access to a table are implicitly trusted not to corrupt the data in this way.
It is worth noting that the system already relies on the assumption that pseudorandom locations (for all records in all tables in all namespaces) are sufficiently separated to prevent data collisions between different records. The ability for users to overwrite locations near their legitimate records, or at some other pseudorandom locations, does not undermine the basic assumption. Although users can corrupt their own tables, it should still be impossible to manipulate records in other tables and namespaces.
The WorldRegistrationSystem contract takes care of registering namespaces, among other things. A namespace sets the context for a user to deploy systems and register tables. Ownership of or access to this namespace allows for the performance of critical operations such as setting Store values or transferring balance. Access to a namespace also gives access to all resource-specific systems or tables.
However, there is a problem when it comes to the arbitrary options of registering namespaces. For instance, consider the following attack. It assumes that the victim initially registers something for a new namespace (so the namespace does not yet exist):
Store writes.Consider making the design more self-contained by removing the automatic namespace registration from the aforementioned functions. Instead, make it mandatory for a user to register the namespace beforehand such that a front-run would make the user's call fail due to the existence check.
Update: Resolved in pull request #2007.
The CoreSystem contract is registered in the ROOT_NAMESPACE in the World contract. This means that it is called using delegatecall.
As a pure implementation contract, there are no access control protections when its functions are invoked directly (bypassing the World). In particular, an attacker can register a system for the CORE_SYSTEM_ID and pass in any _msgSender value in the calldata to bypass the access control check. This is not an attack in itself because the contract storage of the CoreSystem contract is not relevant to the World contract. However, when the attacker invokes the registerTable function directly on the CoreSystem, it will make a delegatecall to the attacker-controlled core system, which could execute the selfdestruct opcode. This will remove the CoreSystem, effectively disabling it for the World contract.
Consider removing the namespace registration branch so that the CoreSystem will not execute a delegatecall. Alternatively (or in addition), consider protecting all CoreSystem functions with an onlyProxy modifier.
Update: Resolved in pull request #2111, pull request #2007, and pull request #2180.
Any storage hook that implements the onBeforeSpliceDynamicData function should receive the state of the record before the update. However, it will receive the updated encoded lengths instead, which may cause the hook to respond incorrectly.
Consider passing the original encoded lengths.
Update: Resolved in pull request #2020.
requireInterface Is Incorrectly SpecifiedThe requireInterface function is incorrectly specified in two ways.
Firstly, it uses a try-catch block to revert with a meaningful error if the contract does not implement the expected interface. However, this will only catch errors that occur in the context of the target contract. If the target contract has no code, or returns a value that cannot be decoded as a boolean, the error will occur in the caller's context and will revert outside the try-catch block.
Secondly, it discards some requirements of the EIP specification. To be compliant, it should confirm that the supportsInterface function returns false for the invalid interface 0xffffffff, and that it does not consume more than 30000 gas. This ensures that unrelated contracts that happen to have a fallback function that returns at least 32 bytes will not mistakenly pass the validation.
Consider using the OpenZeppelin ERC165Checker contract, or correcting these deviations.
Update: Resolved in pull request #2016.
In the Bytes library, the slice4 function intends to return a bytes4 value from a bytes32 value at a given index. While the output is correctly returned as a bytes4 value, the actual output variable is declared as a bytes2 value. This leads to only two correct bytes being returned from the index position followed by two zero bytes. The wrong data could escalate into more severe security issues depending on the context.
Consider correcting the output variable type to bytes4.
Update: Resolved in pull request #2031.
In the Storage library, the load function can load data from a storage location into a memory pointer location. The loaded data can further be specified by an offset in the storage and a length to enable loading data that spans over multiple slots.
However, there is an edge case in the parameter input set that causes memory corruption. This is when data is loaded from an offset position, but the length is less than the remainder of the slot:
storage slot: [ 0 ---------- 10 ---------- 20 ----------- 31 ]
^ offset
<-- length -->
<------- word remainder ------->
The problem occurs because the bitmask is based on the word remainder instead of accounting for the length. This leads to extra bytes being written into memory (e.g., bytes 20 to 31 in the example above).
There are no significant consequences in the current version of the codebase. This is because all memory loads either use a zero offset or invoke the three-parameter version which reserves sufficient space to cover and ignore the unwanted bytes. Nevertheless, the library is intended to support external codebases and the inconsistency may lead to arbitrarily severe memory corruption.
Consider respecting the length when constructing the bitmask as seen in the store function.
Update: Resolved in pull request #1978.
registerFunctionSelector Can Be Front-Run and DoS'edThe registerFunctionSelector function is used to register system function selectors for the World context. This unique bytes4 worldFunctionSelector is based on the user-provided namespaceString, nameString, and systemFunctionSignature. In the fallback function of the World contract, the worldFunctionSelector is taken as a look-up table value to forward the call to the right system and function. This mechanism exists in parallel to the call and callFrom functions that allow the user to specify the system to address through a function parameter.
The problem is that function selectors are only four bytes, so collisions (technically, second-preimages) can be realistically generated. This is particularly true if the worldFunctionSelector is known ahead of time (e.g., in the context of DAO voting). If a worldFunctionSelector is already taken, subsequent registrations will revert. This allows a malicious actor to brute-force a worldFunctionSelector that is identical to that of the victim and front-run the transaction. Hence, the legitimate victim transaction would fail.
Further, the worldFunctionSignature is constructed by concatenating the namespace, name, and system function signature with underscores. If the function name contains underscores itself, a selector collision can be trivially constructed. For example, the worldFunctionSignature of MyNS_MySystem_do_things() can be broken down into:
| actor | namespace | name | function signature |
|---|---|---|---|
| victim | MyNS |
MySystem |
do_things() |
| attacker | MyNS_MySystem |
do |
things() |
The impact of this attack is Denial of Service with the intention of griefing. The victim would have to re-write their code, redeploy it, and try to register it again. Alternatively, they will be unable to use this selector mapping feature.
Another collision that may occur is between registered worldFunctionSignatures and any of the external/public functions in the World contract. Since the external/public function is prioritized over the fallback function, the intended system call could lead to two problems:
Consider changing the world function selector delimiter from an underscore to a character that is invalid for function names (e.g., a colon). Also, consider registering the World's functions so that they can't be registered for the fallback function mechanism. However, the front-running issue will persist without a larger redesign. Hence, consider reducing the attack surface and code complexity by removing this feature entirely, or possibly restricting it to the root namespace and enforcing the usage of the call and callFrom functions that address systems specifically.
Update: Resolved in pull request #2160, pull request #2169, and pull request #2182. The Lattice Labs team stated:
The fixes address the potential conflicts between function selectors of different namespaces. It does not address the possibility of front-running namespace/function selector registration which we've decided to punt on for now and later address with an optional module. This module will allow "committing" to a hashed namespace and then "revealing/registering" the namespace in a second step. Since C-01 is fixed, this is not a security issue anymore but just a potential griefing vector.
We have also decided to punt on addressing the possibility of brute-force calculating a conflicting function selector - since the value space is only 4 bytes it seems impossible to prevent this - but the feature is too useful to remove it altogether. The only issue this could lead to is a griefing attack if a function selector is known long enough before it is registered. This seems like a relatively small issue compared to the value custom function selectors provide in the vast majority of situations.
Throughout the codebase, there are multiple instances of misleading documentation:
loadField function of the Storage library, the documentation suggests that the bytes beyond the length parameter are zero. However, considering a storage slot as [ aa..aa bb..bb cc..cc ] and bb..bb being the desired field, then despite its respective length, the result would return cc..cc as part of the bytes response. As such, the documentation is misleading and can cause consecutive memory corruption. Consider cleaning up the memory space by applying a bitmask or updating the documentation.zero function of the Storage library, the length parameter's documentation suggests it should be specified in words. However, in the code, the value is expected to be in bytes. This can lead to a shortcoming in overwrites to zero and therefore in preserving old data, which can lead to additional security issues. Furthermore, the length is rounded up to the nearest 32 bytes multiple which is not clear either. Consider clarifying the length documentation.Storage.load function refer to the masking as "middle part" and "surrounding parts" whereas it is actually the left and right parts.pack functions of the PackedCounterLib library read "Packs a single value into a PackedCounter", but most of them pack multiple values.leftPaddingBits parameter in the TightCoder.encode function is ambiguous. It says, "The number of bits to pad on the left for each element", but the code actually shifts the values by that count to the left. For instance, this means when encoding a uint120 value, the leftPaddingBits count would be 136 as seen in the EncodeArray library.Consider correcting the documentation to align with the code's behavior. This will help improve the clarity and readability of the codebase.
Update: Resolved in pull request #2100.
This issue was independently identified and fixed by the Lattice team during the audit.
The Core module registers the tables it uses, but it excludes the FunctionSignatures table. This prevents the indexers from decoding the FunctionSignatures events. Consider registering this table as well.
Update: Resolved in pull request #1841 at commit f96d8b3.
The StoreCore contract allows for the manipulation of static and dynamic data in the tables. This includes setting a record of table elements, splicing static and dynamic data, and deleting a record. Hooks before and after the respective action enable reactivity to these storage changes. Whenever one of these functions is called, an event is emitted to notify off-chain indexers about the latest storage changes. It is expected that the indexers can recreate the on-chain storage from these events.
A problem can arise due to the order of events and calls within any of the aforementioned functions which can cause an indexer to perceive a different state outcome than what happened on-chain. The flow for those functions is generally the following:
However, a malicious or mistaken namespace owner can set up a hook that can reenter any of those functions to change the flow (simplified) to the following:
A.B.B.A.As such, although event B was emitted last, storage change A is in fact true. For instance, this could cause major confusion when the same record is first deleted and then created which would be perceived in a reversed order off-chain.
Consider placing the event emissions together with the actual state change to prevent this type of confusion for off-chain indexers.
Update: Resolved in pull request #2017.
The BalanceTransferSystem is used to send a namespace's accounted ETH value to either another namespace or to an address. However, balances can be sent to a non-existent namespace. This could occur due to human error and would likely lead to a loss of funds. Although the non-existent namespace could be registered afterwards, any other user could front-run this registration to get control over the funds.
Consider checking whether the recipient namespace exists before proceeding with the transaction.
Update: Resolved in pull request #2095.
The WorldRegistrationSystem allows users to register a delegation. This will enable the delegatee to make a call on behalf of the delegator, provided it meets the delegator's criteria.
When setting up delegation control with a system, it is possible to register an invalid delegation control without it being immediately noticeable. This would be the case when initCallData is not required (has length zero), which causes the registration function to skip the interface conformance check of the delegation control system. An attacker can take advantage of this mistake, if the delegation is not account-specific, by registering this non-existent system ID to thereby gain control over the system.
Consider performing the interface checks on the delegation control system in all cases, whether or not initCallData is required. It should be noted that this would prevent clearing the delegation (i.e., resetting the delegationControlId to zero). Thus, consider introducing a new function to unregister delegations.
Update: Resolved in pull request #2096.
The World contract's initialize function sets up the CoreSystem functionality. To make it only callable once, it checks if there is no existing "core" module installed.
However, the core module is installed under the name that the module responds with. If it responds with a different name, initialize will still be callable. This should not happen if the World creator is honest. A malicious owner can utilize this function as a backdoor through the delegate call. As the initialize function's access is bound to the creator role, this attack vector is active despite potentially revoking any root namespace ownership. Also, it allows overwriting any other record in the table of installed modules.
It is worth noting that the initial validation reads a record from the InstalledModules table before it is registered, and the subsequent update assumes the core module registered the table. This still works whether or not the core module actually registers the InstalledModules table because both operations directly access the contract storage location where the relevant table record would be stored. Nevertheless, it undermines the expected table abstraction.
Consider explicitly installing the core module under the CORE_MODULE_NAME to ensure the initialize function is single-use. Alternatively, to avoid accessing an unregistered table, consider introducing a new initialization flag.
Update: Resolved in pull request #2170. There is a new CoreModuleAddress table to save the address.
The following interface IDs include the ERC165_INTERFACE_ID constant:
STORE_HOOK_INTERFACE_IDMODULE_INTERFACE_IDSYSTEM_HOOK_INTERFACE_IDWORLD_CONTEXT_CONSUMER_INTERFACE_IDThis is inconsistent with the EIP (as can be seen by the Simpson example). Consider excluding the ERC165_INTERFACE_ID constant from the custom interface IDs.
Update: Resolved in pull request #2014.
There are several consistency checks when registering a new table, but they are incomplete. In particular:
Inconsistent table specifications may interfere with saving, retrieving and interpreting the database records. Consider including these additional validations.
Update: Resolved in pull request #2046.
The ModuleInstallationSystem allows anyone to install any module. Since this makes an external call to the module, it does not provide any additional functionality. Instead, it is simply a convenient way to execute several related operations atomically. However, it can overwrite any record in the InstalledModules table, which doesn't have any subsequent effects in the current codebase.
Consider whether the InstalledModules table should use the module address (along with the arguments hash) instead of the module name as the record key. This is so that it identifies the action that was taken more directly and is more resistant to being overwritten. Alternatively, consider whether the InstalledModules table should be removed entirely.
Update: Resolved in pull request #2168.
Resource IDs in the World are expected to encode three components (the type, namespace and name) that are used to ensure consistency between different table records. However, there are several instances of incomplete consistency checks:
transferOwnership on any resource ID, provided the caller owns the corresponding namespace. This includes a table, system, an unknown resource ID, or an incorrectly specified resource ID.transferBalanceToNamespace with a toNamespaceId that has a non-zero "name" component. This will effectively transfer the funds to a new account in the destination namespace, which needs to be spent individually (i.e., it cannot be consumed by systems within the namespace unless they are explicitly designed to handle it).registerFunctionSelector with any resource ID, provided the caller owns the corresponding namespace. In practice, if it does not correspond to a valid system, it will be unusable.grantAccess to an incorrectly specified resource ID, or a resource that does not exist.registerNamespace or registerNamespaceDelegation with a namespaceId that has a non-zero "name" component.registerSystemHook with any resource ID, provided the caller owns the corresponding namespace.registerStoreHook with a non-existent tableId, provided the caller owns the corresponding namespace.In the interest of predictability and limiting the attack surface, consider validating all three components of user-provided resource IDs and validating the existence of a resource wherever relevant throughout the codebase.
Update: Resolved in pull request #2142 and pull request #2195.
In the getDynamicFieldSlice function of the StoreCore library, a bytes range specified by start and end can be read from a dynamic field.
While the range is checked to be within the field's length, the order of start and end is not checked. If start were to be greater than end, the subtraction for the length would underflow and the transaction would revert implicitly without providing any contextual information.
Consider adding a range check to fail more explicitly.
Update: Resolved in pull request #2034.
The WorldResourceIdInstance library has helper functions to get information from an encoded World resource ID such as type, namespace and name. These fields can also be returned as a string using the toString function.
However, while the resource namespace is expected to occupy 14 bytes, the ROOT_NAMESPACE_STRING is actually a bytes16 value. The string output could, therefore, have an unexpected length.
Consider changing the constant's type from bytes16 to bytes14.
Update: Resolved in pull request #1976.
The Module contract overrides the supportsInterface function and no longer supports the WORLD_CONTEXT_CONSUMER_INTERFACE_ID. This is inconsistent with the equivalent DelegationControl override, which retains support for the WORLD_CONTEXT_CONSUMER_INTERFACE_ID.
Consider ensuring ERC-165 support for all implemented interfaces.
Update: Resolved in pull request #2032.
Several functions in the codebase have side effects or handle multiple use cases:
AccessManagementSystem does not provide a mechanism to renounce the ownership of a namespace, so users will need to transfer ownership to an uncontrolled address (like the zero address). However, this will create an unnecessary record in the ResourceAccess table.registerTable function might register a namespace as well.registerSystem function might register a namespace or delete a system. In fact, this is the only way to delete a system using the World functionality.registerSystem function also accepts an existing system and system ID pair, which will delete the system records only to recreate them again, possibly changing the publicAccess flag in the process.UserDelegationControl table.The stray records could potentially be addressed by directly updating the tables, but this is fragile and undermines the abstractions provided by the World contract. Similarly, functions with side effects or multiple use cases can be fragile, and the most significant issues in this report are a result of imprecise handling at the boundaries of different functionality.
To reduce the attack surface and increase predictability, consider limiting each function to a single use case with a linear code path where possible.
Update: Resolved in pull request #2157.
The Storage library contains functions to store and load data according to the storage pointer, offset, length, and memory pointer. However, both functions take these arguments in different orders which can be non-intuitive for a developer.
Consider aligning the order of function parameters when the set of arguments overlaps.
Update: Resolved in pull request #2033.
Throughout the codebase, there are code element names that are ambiguous:
_staticFields parameter of the encode function could be called _staticFieldLengths.requireNoCallback modifier could be called prohibitDirectCallback.byteCode parameter in the deploy function could be called creationCode as referred to in the documentation.tableWithHooks parameter of the filterListByAddress function could be called elementWithHooks to account for the possibility that it refers to a system.Update: Resolved in pull request #2091.
Throughout the codebase there are multiple instances of unused functions and variables:
Bytes library, the functions toBytes32, equals, setBytes1, setBytes2, setBytes4, setBytes5, setBytes7, as well as some of the slice functionsNAME_BITS, TYPE_MASK, RESOURCE_MODULE and MASK_PTR constantsConsider removing this code to improve the overall clarity and readability of the codebase.
Update: Resolved in pull request #2090.
Throughout the codebase, the following imports are unused and could be removed:
STORE_VERSION of StoreRead.solSliceLib of TightCoder.solFunctionSelectors of SystemCall.solIStore, Schema, System, ROOT_NAMESPACE, ROOT_NAME, revertWithBytes, NamespaceOwner, IDelegationControl, Systems, and SystemHooks of World.solROOT_NAMESPACE, IBaseWorld, ResourceId, WorldResourceIdLib, WorldResourceIdInstance, RESOURCE_SYSTEM, AccessManagementSystem, BalanceTransferSystem, BatchCallSystem, ModuleInstallationSystem, and StoreRegistrationSystem of CoreModule.solIModule, WorldResourceIdLib, and InstalledModules of AccessManagementSystem.solWorldResourceIdLib of BalanceTransferSystem.solAccessControl and ResourceAccess of ModuleInstallationSystem.solROOT_NAMESPACE, WorldContextProviderLib, NamespaceOwner, ResourceAccess, SystemHooks, SystemRegistry, and FunctionSelectors of StoreRegistrationSystem.solIDelegationControl of WorldRegistrationSystem.solDecodeSlice of Slice.solConsider removing unused imports to improve the overall clarity and readability of the codebase.
Update: Resolved in pull request #2028.
There is a duplicate import of the WorldContextProviderLib library in the SystemCall.sol file.
Consider removing duplicate imports to improve the readability of the codebase.
Update: Resolved in pull request #2024.
There are some instances of missing explicitly declared visibility:
coreSystem state variable_toBool functionrequireInterface functionConsider always explicitly declaring the visibility of variables and functions, even when the default visibility matches the intended visibility.
Update: Resolved in pull request #2029.
Throughout the codebase, there are multiple instances where the code could be simplified:
Schema library can be simplified. Consider looping over the _numStaticFields schema types first to ensure they all have a non-zero static byte length. Then, loop over the _numDynamicFields schema types to ensure they have a zero static byte length. This way, it is not necessary to count the fields (and validate the count later) or branch inside the loop.Memory.copy function is redundant to the identity precompile. Consider replacing the function's logic with a call to the precompile for simplicity and efficiency.leftMask function is constructed by first creating a right mask (of size 32 - byteLength) and then inverting it. However, the actual usage involves using both left and right masks [1, 2, 3, 4, 5]. Hence, the left mask is inverted back to a right mask. Consider returning a right mask in the first place.FieldLayout, PackedCounter and Schema, the ResourceId does not have a using statement in the ResourceId.sol file. Instead, it is duplicated several times throughout the codebase.Slice library uses pointer(self) first and self.pointer() later.supportsInterface functions. Instead, consider making use of the interface ID provided by the interface type like type(Interface).interfaceId.FieldLayout.numFields function could make use of the numStaticFields and numDynamicFields functions instead of reimplementing them.Slice.sol and Schema.sol files, there is an inconsistency when it comes to the leading underscores for stack variables.StoreCore.getFieldLayout function, instead of getting the field layout through Storage.loadField, consider using Tables._getFieldLayout, similar to how it is done for the key schema and value schema.WorldContextConsumer._world, the call to StoreSwitch could invoke WorldContextConsumerLib._world for consistency with the rest of the contract.coreModule state variable of the WorldFactory is set in the constructor and never changed. Consider making it immutable.WorldRegistrationSystem.registerSystem function accepts a WorldContextConsumer type argument, whereas it should accept a System type argument.slice functions in the Bytes library can be simplified by declaring the output variable as a named return to save on the extra declaration and return in the function body.getSubslice function could be replaced with start > end || end > data.length.extcodesize opcode is not necessary to check whether the Create2.deploy call was successful since the address will also be zero in that case.Consider applying the code changes as outlined above to improve the overall clarity and readability of the codebase.
Update: Resolved in pull request #2140. The Lattice Labs team stated:
We chose not to implement three suggestions due to gas use and readability concerns.
The codebase defines the BYTES_TO_BITS constant, but then fails to use it in several places [1, 2, 3, 4, 5]. Consider using the constant wherever it applies.
Update: Resolved in pull request #2015.
The following typographical errors were identified in the codebase:
DYNMAIC_DATA_SLOT should be called DYNAMIC_DATA_SLOTConsider fixing all instances of typographical errors throughout the codebase.
Update: Resolved in pull request #2023.
There are multiple cases [1, 2, 3] where the full record in the Systems table is retrieved but only the system address is retained. Consider using the more specialized _getSystem function instead.
Update: Resolved in pull request #2022.
When the CoreModule is installed, the root namespace and world namespace are both registered with the creator as the owner. However, the store namespace is not registered. If another user registers it later, they would control the store tables (i.e., Tables, StoreHooks and ResourceIds), which essentially gives them control over the entire World contract.
Update: Resolved in pull request #1712 at commit c14d140.
The MUD system is a comprehensive framework for rapidly developing blockchain applications. It revolves around a central component called the Store, which acts as a flexible database contract allowing developers to dynamically add tables and functionalities. The World extends this functionality, enabling higher-level abstractions, access control, system registration, and delegation mechanisms, providing a customizable and upgradable structure for blockchain applications.
Throughout the audit period, the Lattice team was very responsive and provided us with useful explanations and clarifications. We also appreciated the code's organization and documentation. As noted in the report, our main suggestions involved limiting the number of branching code paths to limit the attack surface. The Lattice team implemented all suggested improvements. Overall, we believe the system is well-designed and implemented.