27 KiB

Raw Blame History

layout	page_title	sidebar_current	description
docs	ACL System	docs-internals-acl	Consul provides an optional Access Control List (ACL) system which can be used to control access to data and APIs. The ACL system is a Capability-based system that relies on tokens which can have fine grained rules applied to them. It is very similar to AWS IAM in many ways.

ACL System

Consul provides an optional Access Control List (ACL) system which can be used to control access to data and APIs. The ACL is Capability-based, relying on tokens to which fine grained rules can be applied. It is very similar to AWS IAM in many ways.

Scope

When the ACL system was launched in Consul 0.4, it was only possible to specify policies for the KV store. In Consul 0.5, ACL policies were extended to service registrations. In Consul 0.6, ACL's were further extended to restrict service discovery mechanisms, user events, and encryption keyring operations.

ACL Design

The ACL system is designed to be easy to use, fast to enforce, and flexible to new policies, all while providing administrative insight.

Every token has an ID, name, type, and rule set. The ID is a randomly generated UUID, making it unfeasible to guess. The name is opaque to Consul and human readable. The type is either "client" (meaning the token cannot modify ACL rules) or "management" (meaning the token is allowed to perform all actions).

The token ID is passed along with each RPC request to the servers. Agents can be configured with an acl_token property to provide a default token, but the token can also be specified by a client on a per-request basis. ACLs were added in Consul 0.4, meaning prior versions do not provide a token. This is handled by the special "anonymous" token. If no token is provided, the rules associated with the anonymous token are automatically applied: this allows policy to be enforced on legacy clients.

ACLs can also act in either a whitelist or blacklist mode depending on the configuration of acl_default_policy. If the default policy is to deny all actions, then token rules can be set to whitelist specific actions. In the inverse, the allow all default behavior is a blacklist where rules are used to prohibit actions. By default, Consul will allow all actions.

ACL Datacenter

Enforcement is always done by the server nodes. All servers must be configured to provide an acl_datacenter which enables ACL enforcement but also specifies the authoritative datacenter. Consul relies on RPC forwarding to support Multi-Datacenter configurations. However, because requests can be made across datacenter boundaries, ACL tokens must be valid globally. To avoid consistency issues, a single datacenter is considered authoritative and stores the canonical set of tokens.

When a request is made to a server in a non-authoritative datacenter server, it must be resolved into the appropriate policy. This is done by reading the token from the authoritative server and caching the result for a configurable acl_ttl. The implication of caching is that the cache TTL is an upper bound on the staleness of policy that is enforced. It is possible to set a zero TTL, but this has adverse performance impacts, as every request requires refreshing the policy via a cross-datacenter WAN RPC call.

Outages and ACL Replication

The Consul ACL system is designed with flexible rules to accommodate for an outage of the acl_datacenter or networking issues preventing access to it. In this case, it may be impossible for servers in non-authoritative datacenters to resolve tokens. Consul provides a number of configurable acl_down_policy choices to tune behavior. It is possible to deny or permit all actions or to ignore cache TTLs and enter a fail-safe mode. The default is to ignore cache TTLs for any previously resolved tokens and to deny any uncached tokens.

Consul 0.7 added an ACL Replication capability that can allow non-authoritative datacenter servers to resolve even uncached tokens. This is enabled by setting an acl_replication_token in the configuration on the servers in the non-authoritative datacenters. With replication enabled, the servers will maintain a replica of the authoritative datacenter's full set of ACLs on the non-authoritative servers. The ACL replication token needs to be a valid ACL token with management privileges, it can also be the same as the master ACL token.

Replication occurs with a background process that looks for new ACLs approximately every 30 seconds. Replicated changes are written at a rate that's throttled to 100 updates/second, so it may take several minutes to perform the initial sync of a large set of ACLs.

If there's a partition or other outage affecting the authoritative datacenter, and the acl_down_policy is set to "extend-cache", tokens will be resolved during the outage using the replicated set of ACLs. An ACL replication status endpoint is available to monitor the health of the replication process.

Locally-resolved ACLs will be cached using the acl_ttl setting of the non-authoritative datacenter, so these entries may persist in the cache for up to the TTL, even after the authoritative datacenter comes back online.

ACL replication can also be used to migrate ACLs from one datacenter to another using a process like this:

Enable ACL replication in all datacenters to allow continuation of service during the migration, and to populate the target datacenter. Verify replication is healthy and caught up to the current ACL index in the target datacenter using the ACL replication status endpoint.
Turn down the old authoritative datacenter servers.
Rolling restart the servers in the target datacenter and change the acl_datacenter configuration to itself. This will automatically turn off replication and will enable the datacenter to start acting as the authoritative datacenter, using its replicated ACLs from before.
Rolling restart the servers in other datacenters and change their acl_datacenter configuration to the target datacenter.

Bootstrapping ACLs

Bootstrapping the ACL system is done by providing an initial acl_master_token configuration which will be created as a "management" type token if it does not exist. The acl_master_token is only installed when a server acquires cluster leadership. If you would like to install or change the acl_master_token, set the new value for acl_master_token in the configuration for all servers. Once this is done, restart the current leader to force a leader election.

Rule Specification

A core part of the ACL system is a rule language which is used to describe the policy that must be enforced.

Key policies are defined by coupling a prefix with a policy. The rules are enforced using a longest-prefix match policy: Consul picks the most specific policy possible. The policy is either "read", "write", or "deny". A "write" policy implies "read", and there is no way to specify write-only. If there is no applicable rule, the acl_default_policy is applied.

Service policies are defined by coupling a service name and a policy. The rules are enforced using an longest-prefix match policy (this was an exact match in 0.5, but changed in 0.5.1). The default rule, applied to any service that doesn't have a matching policy, is provided using the empty string. A service policy is either "read", "write", or "deny". A "write" policy implies "read", and there is no way to specify write-only. If there is no applicable rule, the acl_default_policy is applied. The "read" policy in a service ACL rule allows restricting access to the discovery of that service prefix. More information about service discovery and ACLs can be found below.

The policy for the "consul" service is always "write" as it is managed internally by Consul.

User event policies are defined by coupling an event name prefix with a policy. The rules are enforced using a longest-prefix match policy. The default rule, applied to any user event without a matching policy, is provided by an empty string. An event policy is one of "read", "write", or "deny". Currently, only the "write" level is enforced during event firing. Events can always be read.

Prepared query policies control access to create, update, and delete prepared queries. Service policies are used when executing prepared queries. See below for more details.

We make use of the HashiCorp Configuration Language (HCL) to specify policy. This language is human readable and interoperable with JSON making it easy to machine-generate.

Specification in the HCL format looks like:

# Default all keys to read-only
key "" {
  policy = "read"
}
key "foo/" {
  policy = "write"
}
key "foo/private/" {
  # Deny access to the dir "foo/private"
  policy = "deny"
}

# Default all services to allow registration. Also permits all
# services to be discovered.
service "" {
    policy = "write"
}

# Deny registration access to services prefixed "secure-".
# Discovery of the service is still allowed in read mode.
service "secure-" {
    policy = "read"
}

# Allow firing any user event by default.
event "" {
    policy = "write"
}

# Deny firing events prefixed with "destroy-".
event "destroy-" {
    policy = "deny"
}

# Default prepared queries to read-only.
query "" {
    policy = "read"
}

# Read-only mode for the encryption keyring by default (list only)
keyring = "read"

# Read-only mode for Consul operator interfaces (list only)
operator = "read"

This is equivalent to the following JSON input:

{
  "key": {
    "": {
      "policy": "read"
    },
    "foo/": {
      "policy": "write"
    },
    "foo/private": {
      "policy": "deny"
    }
  },
  "service": {
      "": {
          "policy": "write"
      },
      "secure-": {
          "policy": "read"
      }
  },
  "event": {
    "": {
      "policy": "write"
    },
    "destroy-": {
      "policy": "deny"
    }
  },
  "query": {
    "": {
      "policy": "read"
    }
  },
  "keyring": "read",
  "operator": "read"
}

Building ACL Policies

Blacklist Mode and `consul exec`

If you set acl_default_policy to deny, the anonymous token won't have permission to read the default _rexec prefix; therefore, Consul agents using the anonymous token won't be able to perform consul exec actions.

Here's why: the agents need read/write permission to the _rexec prefix for consul exec to work properly. They use that prefix as the transport for most data.

You can enable consul exec from agents that are not configured with a token by allowing the anonymous token to access that prefix. This can be done by giving this rule to the anonymous token:

key "_rexec/" {
    policy = "write"
}

Alternatively, you can, of course, add an explicit acl_token to each agent, giving it access to that prefix.

Blacklist Mode and Service Discovery

If your acl_default_policy is set to deny, the anonymous token will be unable to read any service information. This will cause the service discovery mechanisms in the REST API and the DNS interface to return no results for any service queries. This is because internally the API's and DNS interface consume the RPC interface, which will filter results for services the token has no access to.

You can allow all services to be discovered, mimicing the behavior of pre-0.6.0 releases, by configuring this ACL rule for the anonymous token:

service "" {
    policy = "read"
}

The above will allow access for reading service information only. This level of access allows discovering other services in the system, but is not enough to allow the agent to sync its services and checks into the global catalog during anti-entropy.

The most secure way of handling service registration and discovery is to run Consul 0.6+ and issue tokens with explicit access for the services or service prefixes which are expected to run on each agent.

Blacklist mode and Events

Similar to the above, if your acl_default_policy is set to deny, the anonymous token will have no access to allow firing user events. This deviates from pre-0.6.0 builds, where user events were completely unrestricted.

Events have their own first-class expression in the ACL syntax. To restore access to user events from arbitrary agents, configure an ACL rule like the following for the anonymous token:

event "" {
    policy = "write"
}

As always, the more secure way to handle user events is to explicitly grant access to each API token based on the events they should be able to fire.

Blacklist Mode and Prepared Queries

After Consul 0.6.3, significant changes were made to ACLs for prepared queries, including a new query ACL policy. See Prepared Query ACLs below for more details.

Blacklist Mode and Keyring Operations

Consul 0.6 and later supports securing the encryption keyring operations using ACL's. Encryption is an optional component of the gossip layer. More information about Consul's keyring operations can be found on the keyring command documentation page.

If your acl_default_policy is set to deny, then the anonymous token will not have access to read or write to the encryption keyring. The keyring policy is yet another first-class citizen in the ACL syntax. You can configure the anonymous token to have free reign over the keyring using a policy like the following:

keyring = "write"

Encryption keyring operations are sensitive and should be properly secured. It is recommended that instead of configuring a wide-open policy like above, a per-token policy is applied to maximize security.

Blacklist Mode and Consul Operator Actions

Consul 0.7 added special Consul operator actions which are protected by a new operator ACL policy. The operator actions cover:

If your acl_default_policy is set to deny, then the anonymous token will not have access to Consul operator actions. Granting read access allows reading information for diagnostic purposes without making any changes to state. Granting write access allows reading information and changing state. Here's an example policy:

operator = "write"

~> Grant write access to operator actions with extreme caution, as improper use could lead to a Consul outage and even loss of data.

Services and Checks with ACLs

Consul allows configuring ACL policies which may control access to service and check registration. In order to successfully register a service or check with these types of policies in place, a token with sufficient privileges must be provided to perform the registration into the global catalog. Consul also performs periodic anti-entropy syncs, which may require an ACL token to complete. To accommodate this, Consul provides two methods of configuring ACL tokens to use for registration events:

Using the acl_token configuration directive. This allows a single token to be configured globally and used during all service and check registration operations.
Providing an ACL token with service and check definitions at registration time. This allows for greater flexibility and enables the use of multiple tokens on the same agent. Examples of what this looks like are available for both services and checks. Tokens may also be passed to the HTTP API for operations that require them.

Restricting service discovery with ACLs

In Consul 0.6, the ACL system was extended to support restricting read access to service registrations. This allows tighter access control and limits the ability of a compromised token to discover other services running in a cluster.

The ACL system permits a user to discover services using the REST API or UI if the token used during requests has "read"-level access or greater. Consul will filter out all services which the token has no access to in all API queries, making it appear as though the restricted services do not exist.

Consul's DNS interface is also affected by restrictions to service registrations. If the token used by the agent does not have access to a given service, then the DNS interface will return no records when queried for it.

Prepared Query ACLs

As we've gotten feedback from Consul users, we've evolved how prepared queries use ACLs. In this section we first cover the current implementation, and then we follow that with details about what's changed between specific versions of Consul.

Managing Prepared Queries

Managing prepared queries includes creating, reading, updating, and deleting queries. There are a few variations, each of which uses ACLs in one of two ways: open, protected by unguessable IDs or closed, managed by ACL policies. These variations are covered here, with examples:

Static queries with no Name defined are not controlled by any ACL policies. These types of queries are meant to be ephemeral and not shared to untrusted clients, and they are only reachable if the prepared query ID is known. Since these IDs are generated using the same random ID scheme as ACL Tokens, it is infeasible to guess them. When listing all prepared queries, only a management token will be able to see these types, though clients can read instances for which they have an ID. An example use for this type is a query built by a startup script, tied to a session, and written to a configuration file for a process to use via DNS.
Static queries with a Name defined are controlled by the query ACL policy. Clients are required to have an ACL token with a prefix sufficient to cover the name they are trying to manage, with a longest prefix match providing a way to define more specific policies. Clients can list or read queries for which they have "read" access based on their prefix, and similar they can update any queries for which they have "write" access. An example use for this type is a query with a well-known name (eg. prod-master-customer-db) that is used and known by many clients to provide geo-failover behavior for a database.
Template queries queries work like static queries with a Name defined, except that a catch-all template with an empty Name requires an ACL token that can write to any query prefix.

Executing Pepared Queries

When prepared queries are executed via DNS lookups or HTTP requests, the ACL checks are run against the service being queried, similar to how ACLs work with other service lookups. There are several ways the ACL token is selected for this check:

If an ACL Token was captured when the prepared query was defined, it will be used to perform the service lookup. This allows queries to be executed by clients with lesser or even no ACL Token, so this should be used with care.
If no ACL Token was captured, then the client's ACL Token will be used to perform the service lookup.
If no ACL Token was captured and the client has no ACL Token, then the anonymous token will be used to perform the service lookup.

In the common case, the ACL Token of the invoker is used to test the ability to look up a service. If a Token was specified when the prepared query was created, the behavior changes and now the captured ACL Token set by the definer of the query is used when looking up a service.

Capturing ACL Tokens is analogous to PostgreSQL’s SECURITY DEFINER attribute which can be set on functions, and using the client's ACL Token is similar to the complementary SECURITY INVOKER attribute.

ACL Implementation Changes for Prepared Queries

Prepared queries were originally introduced in Consul 0.6.0, and ACL behavior remained unchanged through version 0.6.3, but was then changed to allow better management of the prepared query namespace.

These differences are outlined in the table below:

Operation	Version <= 0.6.3	Version > 0.6.3
Create static query without `Name`	The ACL Token used to create the prepared query is checked to make sure it can access the service being queried. This token is captured as the `Token` to use when executing the prepared query.	No ACL policies are used as long as no `Name` is defined. No `Token` is captured by default unless specifically supplied by the client when creating the query.
Create static query with `Name`	The ACL Token used to create the prepared query is checked to make sure it can access the service being queried. This token is captured as the `Token` to use when executing the prepared query.	The client token's `query` ACL policy is used to determine if the client is allowed to register a query for the given `Name`. No `Token` is captured by default unless specifically supplied by the client when creating the query.
Manage static query without `Name`	The ACL Token used to create the query, or a management token must be supplied in order to perform these operations.	Any client with the ID of the query can perform these operations.
Manage static query with a `Name`	The ACL token used to create the query, or a management token must be supplied in order to perform these operations.	Similar to create, the client token's `query` ACL policy is used to determine if these operations are allowed.
List queries	A management token is required to list any queries.	The client token's `query` ACL policy is used to determine which queries they can see. Only management tokens can see prepared queries without `Name`.
Execute query	Since a `Token` is always captured when a query is created, that is used to check access to the service being queried. Any token supplied by the client is ignored.	The captured token, client's token, or anonymous token is used to filter the results, as described above.

ACL Changes Coming in Consul 0.8

Consul 0.8 will feature complete ACL coverage for all of Consul. To ease the transition to the new policies, a beta version of complete ACL support is available starting in Consul 0.7.2.

Here's a summary of the upcoming changes:

Agents now check node and service ACL policies for catalog-related operations in /v1/agent endpoints, such as service and check registration and health check updates.
Agents enforce a new agent ACL policy for utility operations in /v1/agent endpoints, such as joins and leaves.
A new node ACL policy is enforced throughout Consul, providing a mechanism to restrict registration and discovery of nodes by name. This also applies to service discovery, so provides an additional dimension for controlling access to services.
A new session ACL policy controls the ability to create session objects by node name.
Anonymous prepared queries (non-templates without a Name) now require a valid session, which ties their creation to the new session ACL policy.
The existing event ACL policy has been applied to the /v1/event/list endpoint.

New Configuration Options

To enable beta support for complete ACL coverage, set the acl_enforce_version_8 configuration option to true on Consul clients and servers.

Two new configuration options are used once complete ACLs are enabled:

acl_agent_master_token is used as a special access token that has agent ACL policy write privileges on each agent where it is configured. This token should only be used by operators during outages when Consul servers aren't available to resolve ACL tokens. Applications should use regular ACL tokens during normal operation.
acl_agent_token is used internally by Consul agents to perform operations to the service catalog when registering themselves or sending network coordinates to the servers.

For clients, this token must at least have node ACL policy write access to the node name it will register as. For servers, this must have node ACL policy write access to all nodes that are expected to join the cluster, as well as service ACL policy write access to the consul service, which will be registered automatically on its behalf.

Since clients now resolve ACLs locally, the acl_down_policy now applies to Consul clients as well as Consul servers. This will determine what the client will do in the event that the servers are down.

Consul clients do not need to have the acl_master_token or the acl_datacenter configured. They will contact the Consul servers to determine if ACLs are enabled. If they detect that ACLs are not enabled, they will check at most every 2 minutes to see if they have become enabled, and will start enforcing ACLs automatically.

New ACL Policies

The new agent ACL policy looks like this:

agent "<node name prefix>" {
    policy = "<read|write|deny>"
}

This affects utility-related agent endpoints, such as /v1/agent/self and /v1/agent/join.

The new node ACL policy looks like this:

node "<node name prefix>" {
  policy = "<read|write|deny>"
}

This affects node registration, node discovery, service discovery, and endpoints like /v1/agent/members.

The new session ACL policy looks like this:

session "<node name prefix>" {
    policy = "<read|write|deny>"
}

This affects all the of /v1/session endpoints.

27 KiB Raw Blame History Unescape Escape