# ADR 3: Dynamic Capability Store
# Changelog
- 12 December 2019: Initial version
- 02 April 2020: Memory Store Revisions
# Context
Full implementation of the IBC specification (opens new window) requires the ability to create and authenticate object-capability keys at runtime (i.e., during transaction execution), as described in ICS 5 (opens new window). In the IBC specification, capability keys are created for each newly initialised port & channel, and are used to authenticate future usage of the port or channel. Since channels and potentially ports can be initialised during transaction execution, the state machine must be able to create object-capability keys at this time.
At present, the Cosmos SDK does not have the ability to do this. Object-capability keys are currently pointers (memory addresses) of StoreKey
structs created at application initialisation in app.go
(example (opens new window))
and passed to Keepers as fixed arguments (example (opens new window)). Keepers cannot create or store capability keys during transaction execution — although they could call NewKVStoreKey
and take the memory address
of the returned struct, storing this in the Merklised store would result in a consensus fault, since the memory address will be different on each machine (this is intentional — were this not the case, the keys would be predictable and couldn't serve as object capabilities).
Keepers need a way to keep a private map of store keys which can be altered during transaction execution, along with a suitable mechanism for regenerating the unique memory addresses (capability keys) in this map whenever the application is started or restarted, along with a mechanism to revert capability creation on tx failure. This ADR proposes such an interface & mechanism.
# Decision
The Cosmos SDK will include a new CapabilityKeeper
abstraction, which is responsible for provisioning,
tracking, and authenticating capabilities at runtime. During application initialisation in app.go
,
the CapabilityKeeper
will be hooked up to modules through unique function references
(by calling ScopeToModule
, defined below) so that it can identify the calling module when later
invoked.
When the initial state is loaded from disk, the CapabilityKeeper
's Initialise
function will create
new capability keys for all previously allocated capability identifiers (allocated during execution of
past transactions and assigned to particular modes), and keep them in a memory-only store while the
chain is running.
The CapabilityKeeper
will include a persistent KVStore
, a MemoryStore
, and an in-memory map.
The persistent KVStore
tracks which capability is owned by which modules.
The MemoryStore
stores a forward mapping that map from module name, capability tuples to capability names and
a reverse mapping that map from module name, capability name to the capability index.
Since we cannot marshal the capability into a KVStore
and unmarshal without changing the memory location of the capability,
the reverse mapping in the KVStore will simply map to an index. This index can then be used as a key in the ephemeral
go-map to retrieve the capability at the original memory location.
The CapabilityKeeper
will define the following types & functions:
The Capability
is similar to StoreKey
, but has a globally unique Index()
instead of
a name. A String()
method is provided for debugging.
A Capability
is simply a struct, the address of which is taken for the actual capability.
A CapabilityKeeper
contains a persistent store key, memory store key, and mapping of allocated module names.
The CapabilityKeeper
provides the ability to create scoped sub-keepers which are tied to a
particular module name. These ScopedCapabilityKeeper
s must be created at application initialisation
and passed to modules, which can then use them to claim capabilities they receive and retrieve
capabilities which they own by name, in addition to creating new capabilities & authenticating capabilities
passed by other modules.
ScopeToModule
is used to create a scoped sub-keeper with a particular name, which must be unique.
It MUST be called before InitialiseAndSeal
.
InitialiseAndSeal
MUST be called exactly once, after loading the initial state and creating all
necessary ScopedCapabilityKeeper
s, in order to populate the memory store with newly-created
capability keys in accordance with the keys previously claimed by particular modules and prevent the
creation of any new ScopedCapabilityKeeper
s.
NewCapability
can be called by any module to create a new unique, unforgeable object-capability
reference. The newly created capability is automatically persisted; the calling module need not
call ClaimCapability
.
AuthenticateCapability
can be called by any module to check that a capability
does in fact correspond to a particular name (the name can be untrusted user input)
with which the calling module previously associated it.
ClaimCapability
allows a module to claim a capability key which it has received from another module
so that future GetCapability
calls will succeed.
ClaimCapability
MUST be called if a module which receives a capability wishes to access it by name
in the future. Capabilities are multi-owner, so if multiple modules have a single Capability
reference,
they will all own it.
GetCapability
allows a module to fetch a capability which it has previously claimed by name.
The module is not allowed to retrieve capabilities which it does not own.
ReleaseCapability
allows a module to release a capability which it had previously claimed. If no
more owners exist, the capability will be deleted globally.
# Usage patterns
# Initialisation
Any modules which use dynamic capabilities must be provided a ScopedCapabilityKeeper
in app.go
:
# Creating, passing, claiming and using capabilities
Consider the case where mod1
wants to create a capability, associate it with a resource (e.g. an IBC channel) by name, then pass it to mod2
which will use it later:
Module 1 would have the following code:
SomeFunction
, running in module 2, could then claim the capability:
Later on, module 2 can retrieve that capability by name and pass it to module 1, which will authenticate it against the resource:
Module 1 will then check that this capability key is authenticated to use the resource before allowing module 2 to use it:
If module 2 passed the capability key to module 3, module 3 could then claim it and call module 1 just like module 2 did (in which case module 1, module 2, and module 3 would all be able to use this capability).
# Status
Proposed.
# Consequences
# Positive
- Dynamic capability support.
- Allows CapabilityKeeper to return same capability pointer from go-map while reverting any writes to the persistent
KVStore
and in-memoryMemoryStore
on tx failure.
# Negative
- Requires an additional keeper.
- Some overlap with existing
StoreKey
system (in the future they could be combined, since this is a superset functionality-wise). - Requires an extra level of indirection in the reverse mapping, since MemoryStore must map to index which must then be used as key in a go map to retrieve the actual capability
# Neutral
(none known)