P4Runtime Specification
version 1.4.1
The P4.org API Working Group
2024-10-21

Abstract. P4 is a language for programming the data plane of network devices. The P4Runtime API is a control plane specification for controlling the data plane elements of a device defined or described by a P4 program. This document provides a precise definition of the P4Runtime API. The target audience for this document includes developers who want to write controller applications for P4 devices or switches.

1. Introduction and Scope

This document is published by the P4.org API Working Group, which was chartered [21] to design and standardize vendor-independent, protocol-independent runtime APIs for P4-defined or P4-described data planes. This document specifies one such API, called P4Runtime. It is meant to disambiguate and augment the programmatic API definition expressed in Protobuf format and available at https://github.com/p4lang/p4runtime/tree/main/proto.

1.1. P4 Language Version Applicability

P4Runtime is designed to be implemented in conjunction with the P416 language version or later. P414 programs should be translated into P416 to be made compatible with P4Runtime. This version of P4Runtime utilizes features which are not in P416 1.0, but were introduced in P416 1.2.4 [35]. For this version of P4Runtime, we recommend using P416 1.2.4 [35].

This version of the P4Runtime specification does not yet explicitly address compatibility with the following P416 language features introduced in versions 1.2.2 or 1.2.4 of the language specification:

1.2. In Scope

This specification document defines the semantics of P4Runtime messages, whose syntax is defined in Protobuf format. The following are in scope of P4Runtime:

The following are in the scope of this specification document:

1.3. Not In Scope

The following are not in scope of P4Runtime:

The following are not in scope of this specification document:

2. Terms and Definitions

arbitration
Refers to the process through which P4Runtime ensures that at any given time, there is a single primary controller (i.e. a client with write access) for a given role. Also referred to as “client arbitration”.
client
The gRPC client is the software entity which controls the P4 target or device by communicating with the gRPC agent or server. The client may be local (within the device) or remote (for example, an SDN controller).
COS
Class of Service.
device
Synonymous with target, although device usually connotes a physical appliance or other hardware, whereas target can signify hardware or software.
entity
An instantiated P4 program object such as a table or an extern (from PSA or any other architecture).
gRPC
gRPC Remote Procedure Calls, an open-source client-server RPC framework. See [11].
HA
High-Availability. Refers to a redundancy architecture.
Instrumentation
The part of the P4Runtime server which implements the calls to the device or target native “SDK” or backend.
IPC
Inter-Process Communication.
P4 Blob
A more colloquial term for P4 Device Config (Blob = Binary Large Object).
P4 Device Config
The output of the P4 compiler backend, which is included in the Forwarding Pipeline Config. This is opaque, architecture- and target-specific binary data which can be loaded onto the device to change its “program.”
P4Info
Metadata which specifies the P4 entities which can be accessed via P4Runtime. These entities have a one-for-one correspondence with instantiated objects in the P4 source code.
P4RT
Abbreviation for P4Runtime.
Protobuf (Protocol Buffers)
The wire serialization format for P4Runtime. Protobuf version 3 (proto3) is used to define the P4Runtime interface. See [26].
PSA
Portable Switch Architecture [24]; a target architecture that describes common capabilities of network switch devices that process and forward packets across multiple interface ports.
RPC
Remote Procedure Call.
RTT
Round-trip time.
SDN
Software-Defined Networking, an approach to networking that advocates the separation of the control and forwarding planes, as well as the abstraction of the networking infrastructure, in order to promote programmability of the network control. SDN is often associated with OpenFlow, a communications protocol that enables remote control of the network infrastructure through a programmable, centralized network controller.
SDN port
A 32-bit port number defined by a remote Software-Defined Network (SDN) controller. The SDN port number maps to a unique device port id, which may be in a different number space.
server
The gRPC server which accepts P4Runtime requests on the device or target. It uses instrumentation to translate P4Runtime API calls into target-specific actions.
stream
Refers to a gRPC Stream, which is a RPC on which several messages can be sent and received. P4Runtime defines one Stream RPC (StreamChannel), which is a bidirectional stream (both the client and the server can send messages) which is used for packet I/O and client arbitration, among other things.
switch config
Refers to the non-forwarding config (different from the P4 Forwarding Pipeline Config) that is delivered to the switch via a different interface. For example, the switch config may be captured using OpenConfig models and delivered through a gNMI interface.
target
The hardware or software entity which “executes” the P4 pipeline and hosts the P4Runtime Service; often used interchangeably with “device”.
URI
Uniform Resource Identifier; a string of characters designed for unambiguous identification of resources.

3. Reference Architecture

Figure 1 represents the P4Runtime Reference Architecture. The device or target to be controlled is at the bottom, and one or more controllers is shown at the top. P4Runtime only grants write access to a single primary controller for each read/write entity. A role defines a grouping of P4 entities. P4Runtime allows for a primary controller for each role, and a role-based client arbitration scheme ensures only one controller has write access to each read/write entity, or the pipeline config itself. Any controller may perform read access to any entity or the pipeline config. Later sections describe this in detail. For the sake of brevity, the term controller may refer to one or more controllers.

The P4Runtime API defines the messages and semantics of the interface between the client(s) and the server. The API is specified by the p4runtime.proto Protobuf file, which is available on GitHub as part of the standard [19]. It may be compiled via protoc the Protobuf compiler to produce both client and server implementation stubs in a variety of languages. It is the responsibility of target implementers to instrument the server.

Reference implementations of P4 targets supporting P4Runtime, as well as sample clients, may be available on the p4lang/PI GitHub repository [20]. A future goal may be to produce a reference gRPC server which can be instrumented in a generic way, e.g. via callbacks, thus reducing the burden of implementing P4Runtime.

The controller can access the P4 entities which are declared in the P4Info metadata. The P4Info structure is defined by p4info.proto, another Protobuf file available as part of the standard.

The controller can also set the ForwardingPipelineConfig, which amounts to installing and running the compiled P4 program output, which is included in the p4_device_config Protobuf message field, and installing the associated P4Info metadata. Furthermore, the controller can query the target for the ForwardingPipelineConfig to retrieve the device config and the P4Info.

reference-architecture


Figure 1. P4Runtime Reference Architecture.

3.1. P4Runtime Service Implementation

The P4Runtime API is implemented by a program that runs a gRPC server which binds an implementation of auto-generated P4Runtime Service interface. This program is called the “P4Runtime server.” The server must listen on TCP port 9559 by default, which is the port that has been allocated by IANA for the P4Runtime service. Servers should allow users to override the default port using a configuration file or flag when starting the server. Uses of other port numbers as the default should be discontinued.

3.1.1. Security concerns

Appropriate measures and security best practices must be in place to protect the P4Runtime server and client, and the communication channel between the two. For example, firewalling and authenticating the incoming connections to the P4Runtime server can prevent a malicious actor from taking over the switch. Similarly, using TLS to authenticate and encrypt the gRPC channel can prevent man-in-the-middle attacks between the server and client. Mutual TLS (mTLS) may be used to facilitate the authentication of the client by the server and vice-versa.

3.2. Idealized Workflow

In the idealized workflow, a P4 source program is compiled to produce both a P4 device config and P4Info metadata. These comprise the ForwardingPipelineConfig message. A P4Runtime controller chooses a configuration appropriate to a particular target and installs it via a SetForwardingPipelineConfig RPC. Metadata in the P4Info describes both the overall program itself (PkgInfo) as well as all entity instances derived from the P4 program tables and extern instances. Each entity instance has an associated numeric ID assigned by the P4 compiler which serves as a concise “handle” used in API calls.

In this workflow, P4 compiler backends are developed for each unique type of target and produce P4Info and a target-specific device config. The P4Info schema is designed to be target and architecture-independent, although the specific contents are likely to be architecture-dependent. The compiler ensures the code is compatible with the specific target and rejects code which is incompatible.

In some use cases, it is expected that a controller will store a collection of multiple P4 “packages”, where each package consists of the P4 device config and P4Info, and install them at will onto the target. A controller can also query the ForwardingPipelineConfig from the target via the GetForwardingPipelineRequest RPC. This can be useful to obtain the pipeline configuration from a running device to synchronize the controller to its current state.

3.3. P4 as a Behavioral Description Language

P4 can be considered a behavioral description of a switching device which may or may not execute “P4” natively. There is no requirement that a P4 compiler be used in the production of either the P4 device config or the P4Info. There is no absolute requirement that the target accept a SetForwardingPipelineRequest to change its pipeline “program”, as some devices may be fixed in function, or configured via means other than P4 programs. Furthermore, a controller can run without a P4 source program, since the P4Info file provides all of the information necessary to describe the P4Runtime API messages needed to configure such a device.

While a P4 program does provide a precise description of the data plane behavior, and this can prove invaluable in writing correct control plane software, in some cases it is enough for a control plane software developer to have the control plane API, plus good documentation of the data plane behavior. Some device vendors may wish to keep their P4 source code private. The minimum requirement for the controller and device to communicate properly is a P4Info file that can be loaded by a controller in order to render the correct P4Runtime API.

In such scenarios, it is crucial to have detailed documentation, perhaps included in the P4Info file itself, specifically the metadata in the PkgInfo message as well as the embedded doc fields. Nevertheless, a P4 program which describes the pipeline is ideally available. The contents of the P4Info file will be described in later sections.

3.4. Alternative Workflows

Given the notions above concerning P4 code as behavioral description and P4Info as API metadata, some other workflows are possible. The scenarios below are just examples and actual situations may vary.

3.4.1. P4 Source Available, Compiled into P4Info but not Compiled into P4 Device Config

In this situation, P4 source code is available mainly as a behavioral model and compiled to produce P4Info, but it is not compiled to produce the p4_device_config. The device's configuration might be derived via some other means to implement the P4 source code's intentions. The P4 code, if available, can be studied to understand the pipeline, and the P4Info can be used to implement the control plane.

3.4.2. No P4 Source Available, P4Info Available

In this situation, P4Info is available but no P4 source is available for any number of reasons, the most likely of which are:

  1. The vendor or organization does not wish to divulge the P4 source code, to protect intellectual property or maintain security.

  2. The target was not implemented using P4 code to begin with, although it still obeys the control plane API specified in the P4Info.

As discussed in Section 3.3, in the absence of a P4 program describing the data plane behavior, the detailed knowledge required to write correct control plane code must come from other sources, e.g. documentation.

3.4.3. Partial P4Info and P4 Source are Available

In this situation, a subset of the target's pipeline configuration is exposed as P4 source code and P4Info. The complete device behavior might be expressed as a larger P4 program and P4Info, but these are not exposed to everybody. This limits API access to only certain functions and behaviors. The hidden functions and APIs might be available to select users who would have access to the complete P4Info and possibly P4 source code.

3.4.4. P4Info Role-Based Subsets

In this situation, P4Info is selectively packaged into role-based subsets to allow some controllers access to just the functionality required. For example, a controller may only need read access to statistics counters and nothing more.

3.5. P4Runtime State Across Restarts

All targets support full restarts, where all forwarding state is reset and the P4Runtime server starts with a clean state. Some targets may also support In-Service Software Upgrade (ISSU), where the software on the target can be restarted while traffic is being forwarded. In this case, the P4Runtime server may have the ability to access information from memory before the upgrade.

4. Controller Use-cases

P4Runtime allows for more than one controller. The mechanisms and semantics are described in a later section. Here we present a number of use-cases. Each use-case highlights a particular aspect of P4Runtime's flexibility and is not intended to be exhaustive. Real-world use-cases may combine various techniques and be more complex.

4.1. Single Embedded Controller

Figure 2 shows perhaps the simplest use-case. A device or target has an embedded controller which communicates to an on-board switch via P4Runtime. This might be appropriate for an embedded appliance which is not intended for SDN use-cases.

P4Runtime was designed to be a viable embedded API. Complex controller architectures typically feature multiple processes communicating with some sort of IPC (Inter-Process Communications). P4Runtime is thus both an ideal RPC and an IPC.

single-embedded-controller


Figure 2. Use-Case: Single Embedded Controller

4.2. Single Remote Controller

Figure 3 shows a single remote Controller in charge of the P4 target. In this use-case, the device has no control of the pipeline, it just hosts the server. While this is possible, it is probably more practical to have a hybrid use-case as described in subsequent sections.

single-remote-controller


Figure 3. Use-Case: Single Remote Controller

4.3. Embedded + Single Remote Controller

Figure 4 illustrates the use-case of an embedded controller plus a single remote controller. Both controllers are clients of the single server. The embedded controller is in charge of one set of P4 entities plus the pipeline configuration. The remote controller is in charge of the remainder of the P4 entities. An equally-valid, alternative use-case, could assign the pipeline configuration to the remote controller.

For example, to minimize round-trip times (RTT) it might make sense for the embedded controller to manage the contents of a fast-failover table. The remote controller might manage the contents of routing tables.

embedded-plus-single-remote-controller


Figure 4. Use-Case: Embedded Plus Single Remote Controller

4.4. Embedded + Two Remote Controllers

Figure 5 illustrates the case of an embedded controller similar to the previous use-case, and two remote controllers. One of the remote controllers is responsible for some entities, e.g. routing tables, and the other remote controller is responsible for other entities, perhaps statistics tables. Role-based access divides the ownership.

embedded-plus-two-remote-controllers


Figure 5. Use-Case: Embedded Plus Two Remote Controllers

4.5. Embedded Controller + Two High-Availability Remote Controllers

Figure 6 illustrates a single embedded controller plus two remote controllers in an active-standby (i.e. primary-backup) HA (High-Availability) configuration. Controller #1 is the active controller and is in charge of some entities. If it fails, Controller #2 takes over and manages the tables formerly owned by Controller #1. The mechanics of HA architectures are beyond the scope of this document, but the P4Runtime role-based client arbitration scheme supports it.

embedded-plus-two-remote-ha-controllers


Figure 6. Use-Case: Embedded Plus Two Remote High-Availability Controllers

5. Client Arbitration and Controller Replication

The P4Runtime interface allows multiple clients (i.e. controllers) to be connected to the P4Runtime server running on the device at the same time for the following reasons:

  1. Partitioning of the control plane: Multiple controllers may have orthogonal, non-overlapping, “roles” (or “realms”) and should be able to push forwarding entities simultaneously. The control plane can be partitioned into multiple roles and each role will have a set of controllers, one of which is the primary and the rest are backups. Role definition, i.e. how P4 entities get assigned to each role, is out-of-scope of this document.

  2. Redundancy and fault tolerance: Supporting multiple controllers allows having one or more standby backup controllers. These can already have a connection open, which can help them become primary more quickly, especially in the case where the control-plane traffic is in-band and connection setup might be more involved.

To support multiple controllers, P4Runtime uses the streaming channel (available via StreamChannel RPC) for session management. The workflow is described as follows:

gRPC enables the server to identify which client originated each message in the StreamChannel stream. For example, the C++ gRPC library [12] in synchronous mode enables a server process to cause a function to be called when a new client creates a StreamChannel stream. This function should not return until the stream is closed and the server has completed any cleanup required when a StreamChannel is closed normally (or broken, e.g. because a client process unexpectedly terminated). Thus the server can easily associate all StreamChannel messages received from the same client, because they are processed within the context of the same function call.

A P4Runtime implementation need not rely on the gRPC library providing information with unary RPC messages that identify which client they came from. Unary RPC messages include requests to write table entries in the data plane, or read state from the data plane, among others described later. P4Runtime relies on clients identifying themselves in every write request, by including the values device_id, role, and election_id in all write requests. The server trusts clients not to use a triple of values other than their own in their write requests. gRPC provides authentication methods [10] that should be deployed to prevent untrusted clients from creating channels, and thus from making changes or even reading the state of the server.

5.1. Default Role

A controller can omit the role message in MasterArbitrationUpdate. This implies the “default role”, which corresponds to “full pipeline access”. This also implies that a default role has a role_id of "" (default). If using a default role, all RPCs from the controller (e.g. Write) must leave the role unset.

5.2. Role Config

The role.config field in the MasterArbitrationUpdate message sent by the controller describes the role configuration, i.e. which operations are in the scope of a given role. In particular, the definition of a role may include the following:

An unset role.config implies “full pipeline access” (similar to the default role explained above). In order to support different role definition schemes, role.config is defined as an Any Protobuf message [37]. Such schemes are out-of-scope of this document. When partitioning of the control plane is desired, the P4Runtime client(s) and server need to agree on a role definition scheme in an out-of-band fashion.

It is the job of the P4Runtime server to remember the role.config for every device_id and role pair.

5.3. Rules for Handling MasterArbitrationUpdate Messages Received from Controllers

  1. If the MasterArbitrationUpdate message is received for the first time on this particular channel (i.e. for a newly connected controller):

    1. If device_id does not match any of the devices known to the P4Runtime server, the server shall terminate the stream by returning a NOT_FOUND error.

    2. If the election_id is set and is already used by another live controller for the same (device_id, role), the P4Runtime server shall terminate the stream by returning an INVALID_ARGUMENT error.

    3. If role.config does not match the “out-of-band” scheme previously agreed upon, the server must return an INVALID_ARGUMENT error.

    4. If the number of open streams for the given (device_id, role) exceeds the supported limit, the P4Runtime server shall terminate the stream by returning a RESOURCE_EXHAUSTED error.

    5. Otherwise, the controller is added to a list of live controllers for the given (device_id, role) and the server remembers the controllers device_id, role and election_id for this gRPC channel. See below for the rules to determine if this controller becomes a primary or backup, and what notifications are sent as a consequence.

  2. Otherwise, if the MasterArbitrationUpdate message is received from an already live controller:

    1. If the device_id does not match the one already assigned to this stream, the P4Runtime server shall terminate the stream by returning a FAILED_PRECONDITION error.

    2. If the role does not match the current role assigned to this stream, the P4Runtime server shall terminate the stream by returning a FAILED_PRECONDITION error. If the controller wishes to change its role, it must close the current stream channel and open a new one.

    3. If role.config does not match the “out-of-band” scheme previously agreed upon, the server must return an INVALID_ARGUMENT error.

    4. If the election_id is set and is already used by another live controller (excluding the controller making the request) for the same (device_id, role), the P4Runtime server shall terminate the stream by returning an INVALID_ARGUMENT error.

    5. Otherwise, the server updates the election_id it has stored for this controller. This change might cause a change in the primary client (this controller might become primary, or the controller might have downgraded itself to a backup, see below), as well as notifications being sent to one or more controllers.

If the MasterArbitrationUpdate is accepted by either of the two steps above (cases 1.5. and 2.5. above), then the server determines if there are changes in the primary client. Let election_id_past be the highest election ID the server has ever seen for the given device_id and role (including the one of the current primary if there is one).

  1. If election_id is greater than or equal to election_id_past, then the controller becomes, or stays, primary. The server updates the role configuration to role.config for the given role. Furthermore:

    1. If there was no primary for this device_id and role before and there are no Write requests still processing from a previous primary, then the server immediately sends an advisory notification to all controllers for this device_id and role. See the following section for the format of the advisory message.

    2. If there was a previous primary, including this controller, or Write requests in flight, then the server carries out the following steps (in this order):

      1. The server stops accepting Write requests from the previous primary (if there is one). At this point, the server will reject all Write requests with PERMISSION_DENIED.

      2. The server notifies all controllers other than the new primary client of the change by sending the advisory notification described in the following section.

      3. The server will finish processing any Write requests that have already started. If there are errors, they are reported as usual to the previous primary. If the previous primary has already disconnected, any possible errors are dropped and not reported.

      4. The server now accepts the current controller as the new primary, thus accepting Write requests from this controller. The server updates the highest election ID (i.e. election_id_past) it has seen for this device_id and role to election_id.

      5. The server notifies the new primary by sending the advisory message described in the following section.

  2. Otherwise, the controller becomes a backup. If the controller was previously a primary (and downgraded itself), then an advisory message is sent to all controllers for this device_id and role. Otherwise, the advisory message is only sent to the controller that sent the initial MasterArbitrationUpdate. See the following section for the format of the advisory message.

5.4. Client Arbitration Notifications

For any given device_id and role, any time a new primary is chosen, a primary downgrades its status to a backup, a primary disconnects, or the role.config is updated by the primary, all controllers for that (device_id, role) are informed of this by sending a StreamMessageResponse. The MasterArbitrationUpdate is populated as follows:

Note that on primary client changes with outstanding Write request, some notifications might be delayed, see the previous section for details.

6. The P4Info Message

The purpose of P4Info was described under Reference Architecture. Here we describe the various components.

6.1. Common Messages

These messages appear nested within many other messages.

6.1.1. Documentation Message

Documentation is used to carry both brief and long descriptions of something. Good content within a documentation field is extremely helpful to P4Runtime application developers.

message Documentation {
  // A brief description of something, e.g. one sentence
  string brief = 1;
  // A more verbose description of something.
  // Multiline is accepted. Markup format (if any) is TBD.
  string description = 2;
}

6.1.2. Preamble Message

The preamble serves as the “descriptor” for each entity and contains the unique instance ID, name, alias, annotations and documentation.

message Preamble {
  // ids share the same number-space; e.g. table ids cannot overlap with counter
  // ids. Even though this is irrelevant to this proto definition, the ids are
  // allocated in such a way that it is possible based on an id to deduce the
  // resource type (e.g. table, action, counter, ...). This means that code
  // using these ids can detect if the wrong resource type is used
  // somewhere. This also means that ids of different types can be mixed
  // (e.g. direct resource list for a table) without ambiguity. Note that id 0
  // is reserved and means "invalid id".
  uint32 id = 1;
  // fully qualified name of the P4 object, e.g. c1.c2.ipv4_lpm
  string name = 2;
  // an alias (alternative name) for the P4 object, probably shorter than its
  // fully qualified name. The only constraint is for it to be unique with
  // respect to other P4 objects of the same type. By default, the compiler uses
  // the shortest suffix of the name that uniquely identifies the object. For
  // example if the P4 program contains two tables with names s.c1.t and s.c2.t,
  // the default aliases will respectively be c1.t and c2.t. In the future, the
  // P4 programmer may also be able to override the default alias for any P4
  // object (TBD).
  string alias = 3;
  repeated string annotations = 4;
  // Optional. If present, the location of `annotations[i]` is given by
 // `annotation_locations[i]`.
  repeated SourceLocation annotation_locations = 7;
  // Documentation of the entity
  Documentation doc = 5;
  repeated StructuredAnnotation structured_annotations = 6;
}

6.1.3. Annotating P4 Entities with Documentation

P4 entities may be annotated using the following annotations:

@brief(string...)
@description(string...)

Attaching either or both of these annotations to an entity will generate a P4Info Documentation Message, which in turn will appear in the Preamble Message for the entity.

The P4 compiler should not emit annotation messages in the P4Info for these specific cases; instead, it should generate the Documentation messages as described.

The following example shows documentation annotations for a table entity:

@brief("Match IPv4 addresses to next-hop MAC and port")
@description("Match IPv4 addresses to next-hop MAC and port. \
Uses LPM match type.")
table my_ipv4_lkup {
  ...
}

6.1.4. Structured Annotations

P4 supports both unstructured and structured annotations [17]. Unstructured annotations of the form MyAnno1 or MyAnno2(body-content) can either be empty, or contain free-form content; anything between the pair of matched parentheses is legal. Conversely, structured annotations of the form MyAnno3[] or MyAnno4[kvList|expressionList] have a more prescribed syntax, which allows declaring key-value lists or expression lists. Both unstructured and structured annotations may be used simultaneously on a P4 element and P4Info supports this.

The annotations described up to this point, e.g. @brief(), have all been unstructured annotations, or simply annotations. These are represented in P4Info as repeated string annotations fields in the various messages. Similarly, structured annotations are represented in repeated StructuredAnnotation structured_annotations fields which are siblings to the unstructured annotations. The structured_annotations contain parsed representations of the original annotation source. This parsing includes expression-evaluation, so the resulting P4Info may contain a simplified replica of the original structured annotations.

The structured annotation messages are defined in p4types.proto.

message KeyValuePair {
  string key = 1;
  Expression value = 2;
}

message KeyValuePairList {
  repeated KeyValuePair kv_pairs = 1;
}

message Expression {
  oneof value {
    string string_value = 1;
    int64 int64_value = 2;
    bool bool_value = 3;
  }
}

message ExpressionList {
  repeated Expression expressions = 1;
}

message StructuredAnnotation {
  string name = 1;
  oneof body {
    ExpressionList expression_list = 2;
    KeyValuePairList kv_pair_list = 3;
  }
  // Optional. Location of the '@' symbol of this annotation in the source code.
  SourceLocation source_location = 4;
}

The StructuredAnnotation message can represent either a KeyValuePairList or an ExpressionList.

The type of an expression is intentionally limited to one of the following base types: string literal, 64-bit signed integer, or boolean. The p4c compiler frontend which generates P4Info will evaluate all expressions and simplify them to one of the valid types. Any expressions which don't match one of the valid types will generate an error. For integers exceeding 64 bits, besides issuing an error, the compiler may print a suggestion to use a string representation, and the P4Info consumer may perform any necessary conversions.

The following invariants hold:

  1. For any P4 entity, there are no two StructuredAnnotations that have the same name.

  2. Within a KeyValuePairList, there are no two KeyValuePairs that have the same key.

6.1.4.1. Structured Annotation Examples

We omit the source_location field in the following examples.

Empty Expression List

@Empty[]
table t {
    ...
}

The generated P4Info will contain the following.

structured_annotations {
  name: "Empty"
}

Mixed Expression List

#define TEXT_CONST "hello"
#define NUM_CONST 6
@MixedExprList[1,TEXT_CONST,true,1==2,5+NUM_CONST]
table t {
    ...
}

The generated P4Info will contain:

structured_annotations {
  name: "MixedExprList"
  expression_list {
    expressions {
      int64_value: 1
    }
    expressions {
      string_value: "hello"
    }
    expressions {
      bool_value: true
    }
    expressions {
      bool_value: false
    }
    expressions {
      int64_value: 11
    }
  }
}

kvList of Mixed Expressions

@MixedKV[label="text", my_bool=true, int_val=2*3]
table t {
    ...
}

The generated P4Info will contain:

structured_annotations {
  name: "MixedKV"
  kv_pair_list {
    kv_pairs {
      key: "label"
      value {
        string_value: "text"
      }
    }
    kv_pairs {
      key: "my_bool"
      value {
        bool_value: true
      }
    }
    kv_pairs {
      key: "int_val"
      value {
        int64_value: 6
      }
    }
  }
}

6.1.5. SourceLocation Message

A source location describes a location within a .p4-source file. The SourceLocation message is defined in p4types.proto as follows:

// Location of code relative to a given source file.
message SourceLocation {
  // Path to the source file (absolute or relative to the working directory).
  string file = 1;
  // Line and column numbers within the source file, 1-based.
  int32 line = 2;
  int32 column = 3;
}

We provide source locations for structured and unstructured annotations. This information may be useful when annotations require further parsing or processing, as it allows tools to point out the precise source of errors for invalid annotations.

The SourceLocation message associated with an annotation holds the location of the @ symbol introducing the annotation in the P4 source code; the message can be found in the following place:

6.2. PkgInfo Message

The PkgInfo message contains package-level metadata which describes the overall P4 program itself, as opposed to P4 entities. PkgInfo can be extracted and used to facilitate “browsing” of available P4 programs from a library. Although all fields are technically “optional,” every implementation should include as a minimum the name, version, doc and arch fields. The other fields are recommended to be included.

// Can be used to manage multiple P4 packages.
message PkgInfo {
  // a definitive name for this configuration, e.g. switch.p4_v1.0
  string name = 1;
  // configuration version, free-format string
  string version = 2;
  // brief and detailed descriptions
  Documentation doc = 3;
  // Miscellaneous metadata, free-form; a way to extend PkgInfo
  repeated string annotations = 4;
  // Optional. If present, the location of `annotations[i]` is given by
  // `annotation_locations[i]`.
  repeated SourceLocation annotation_locations = 10;
  // the target architecture, e.g. "psa"
  string arch = 5;
  // organization which produced the configuration, e.g. "p4.org"
  string organization = 6;
  // contact info for support,e.g. "tech-support@acme.org"
  string contact = 7;
  // url for more information, e.g. "http://support.p4.org/ref/p4/switch.p4_v1.0"
  string url = 8;
  // Miscellaneous metadata, structured; a way to extend PkgInfo
  repeated StructuredAnnotation structured_annotations = 9;
  // If set, specifies the properties that the underlying platform should have.
  // If the platform does not conform to these properties, the server should
  // reject the P4Info when used with a SetForwardingPipelineConfigRequest.
  PlatformProperties platform_properties = 11;
}

where the PlatformProperties message looks as follows:

// Used to describe the required properties of the underlying platform.
message PlatformProperties {
  // The minimum number of multicast entries (i.e. multicast groups) that the
  // platform is required to support. If 0, there are no requirements.
  int32 multicast_group_table_size = 1;
  // The minimum number of replicas that the platform is required to support
  // across all  groups. If 0, there are no requirements.
  int32 multicast_group_table_total_replicas = 2;
  // The number of replicas that the platform is required to support per
  // group/entry. If 0, `multicast_group_table_total_replicas` should be used.
  // Must be no larger than `multicast_group_table_total_replicas`.
  int32 multicast_group_table_max_replicas_per_entry = 3;
}

6.2.1. Annotating P4 code with PkgInfo

A P4 program's PkgInfo may be declared using one or more of the following annotations, attached to the main block only:

@pkginfo(key=value)
@pkginfo(key=value[,key=value,...])
@brief("A brief description")
@description("A longer\
description")
@custom_annotation(...)
@another_custom_annotation(...)
@platform_property(key=value)
@platform_property(key=value[,key=value,...])

Above we see several different types of annotations:

Declaring one or more of these annotations on main will generate a single corresponding PkgInfo message in the P4Info as described in PkgInfo Message.

The following example shows @pkginfo annotations using a mixture of single and multiple key-value pairs. It also shows @brief and @description annotations, plus some additional custom annotations. The well-known annotations will produce corresponding fields inside the PkgInfo message. The custom annotations will be appended to the PkgInfo.annotations list.

@pkginfo(name="switch.p4",version="2")
@pkginfo(organization="p4.org")
@pkginfo(contact="info@p4.org")
@pkginfo(url="www.p4.org")
@brief("L2/L3 switch")
@description("L2/L3 switch.\
Built for data-center profile.")
@my_annotation1(...) // Not well-known, this will appear in PkgInfo annotations
@my_annotation2(...) // Not well-known, this will appear in PkgInfo annotations
PSA_Switch(IgPipeline, PacketReplicationEngine(), EgPipeline,
           BufferingQueueingEngine()) main;

6.3. ID Allocation for P4Info Objects

P4Info objects receive a unique ID, which is used to identify the object in P4Runtime messages. IDs are 32-bit unsigned integers which are assigned by the compiler during the P4Info generation process. IDs are assigned in such a way that it is possible based on the ID value alone to deduce the type of the object (e.g. table, action, counter, ). The most significant 8 bits of the ID encodes the object type (as per Table 1). The p4info.proto file includes a mapping from object type to 8-bit prefix value, encoded as an enum definition (p4.config.v1.P4Ids.Prefix). These values must be used (e.g. by the compiler) when allocating IDs. The remaining 24 bits must be generated in such a way that the resulting IDs must be globally unique in the scope of the P4Info message. Table 2 shows the ID layout.

8-bit prefix value P4 object type
0x00 Reserved (unspecified)
0x01 Action
0x02 Table
0x03 Value-set
0x04 Controller header (header type with @controller_header annotation)
0x050x0f Reserved (for future P4 built-in objects)
0x10 Reserved (start of PSA extern types)
0x11 PSA Action profiles / selectors
0x12 PSA Counter
0x13 PSA Direct counter
0x14 PSA Meter
0x15 PSA Direct meter
0x16 PSA Register
0x17 PSA Digest
0x180x7f Reserved (for future PSA extern types)
0x80 Reserved (start of vendor-specific extern types)
0x810xfe Vendor-specific extern types
0xff Reserved (max prefix value)

Table 1. Mapping of P4Info object type to 8-bit ID prefix value
MSB bit 31 .. bit 24 bit 23 .. bit 0 LSB
Object type prefix Generated suffix (e.g. by the compiler)

Table 2. Format of P4Info object IDs

It is possible to statically set the least-significant 24 bits of the ID in the P4 program source by annotating the object with @id (see Table 3). The compiler must honor the @id annotations when generating the P4Info message and must fail the compilation if statically-assigned ID suffixes lead to non-unique IDs (i.e. if the P4 programmer tries to assign the same ID suffix to two different P4 objects of the same type by annotating them with the same @id value). Note that it is not possible for the P4 programmer to change the value of the 8-bit ID prefix, which encodes the object type. The programmer is free to leave the 8-bit prefix as 0, in which case the compiler will replace the 0 with the correct value for the kind of object the annotation is annotating. The programmer may also fill in the 8-bit prefix with a non-zero value, in which case the compiler will give an error if the 8-bit prefix does not contain the correct value, or leave it as is if it is correct.

P4 declaration(s) Compiler-allocated ID(s)
@id(0x12ab34) table tA... 0x0212ab34
@id(0x12ab34) table tA... Error(same ID suffixes for 2 objects of the same type)
@id(0x12ab34) table tB...
@id(0x12ab34) table tA... 0x0212ab34
@id(0x12ab34) action act1... 0x0112ab34

Table 3. Example of statically-assigned P4Info object IDs

The @id annotation can also be used to choose the ID for match fields, action parameters, and packet metadata. In this case, there is no 8-bit prefix and the programmer is free to choose any 32-bit number. The compiler must fail if the IDs chosen by the programmer are not unique (within a table, action, or header, respectively).

6.4. P4Info Objects

6.4.1. Table

Table messages are used to specify all possible match-action tables exposed to a control plane. This message contains the following fields:

6.4.2. Action

Action messages are used to specify all possible actions of all match-action tables.

The Action message defines the following fields:

6.4.3. ActionProfile

ActionProfile messages are used to specify all available instances of Action Profile and Action Selector PSA externs.

PSA Action Profiles are used to describe implementations of match-action tables where multiple table entries can share the same action instance. Indeed, differently from a regular match-action table where each entry contains the action specification, when using Action Profile-based tables, the control plane can insert entries pointing to an Action Profile member, where each member then points to an action instance. The control plane is responsible for creating, modifying, or deleting members at runtime.

PSA Action Selectors extend Action Profiles with the capability of bundling together multiple members into groups. Match-action table entries can point to a member or group. When processing a packet, if the table entry points to a group, a dynamic selection algorithm is used to select a member from the group and apply the corresponding action to the packet. The dynamic selection algorithm is typically specified in the P4 program when instantiating the Action Selector, however it is not specified in the P4Info. The control plane is responsible for creating, modifying, or deleting both members and groups at runtime.

While PSA defines Action Profile and Action Selector as two different externs, P4Info uses the same ActionProfile message to describe both.

The ActionProfile message includes the following fields:

PSA programs can use the @selector_size_semantics annotation with one of sum_of_weights or sum_of_members to specify this value for Action Selectors. In the sum_of_members case, the @max_member_weight annotation can be used to specify max_member_weight. Unless otherwise specified, the value of selector_size_semantics should default to sum_of_weights. However, an unset selector_size_semantics should also be treated as sum_of_weights for backwards compatibility in Action Selectors. In Action Profiles, this value must be unset.

6.4.4. Counter & DirectCounter

Counter and DirectCounter messages are used to specify all possible instances of Counter and Direct Counter PSA externs respectively. Both externs are used to represent data plane counters that keep statistics such as the number of packets or bytes. The main difference between (indexed) counters and direct counters is:

Both Counter and DirectCounter messages share the following fields:

For indexed counters, the Counter message contains also a size field, an int64 representing the maximum number of independent values that can be held by this counter array. Conversely, the DirectCounter message contains a direct_table_id field that carries the unit32 identifier of the table to which this direct counter is attached.

For indexed counters, the Counter message contains also an index_type_name field, which indicates whether the index has a user-defined type. This is useful for translation. The underlying built-in type must be a fixed-width unsigned bitstring (bit<W>).

6.4.5. Meter & DirectMeter

Meter and DirectMeter messages are used to specify all possible instances of Meter and Direct Meter PSA externs. Both externs provide mechanism to keep data plane statistics typically used to mark or drop packets that exceed a given packet or bit rate. Similarly to counters, the main difference between (indexed) meters and direct meters is:

Both Meter and DirectMeter messages share the following fields:

For indexed meters, the Meter message contains also a size field, an int64 representing the maximum number of independent cells that can be held by this meter. Conversely, the DirectMeter message contains a direct_table_id field that carries the uint32 identifier of the table to which this direct meter is attached.

For indexed meters, the Meter message contains also an index_type_name field, which indicates whether the index has a user-defined type. This is useful for translation. The underlying built-in type must be a fixed-width unsigned bitstring (bit<W>).

6.4.6. ControllerPacketMetadata

ControllerPacketMetadata messages are used to describe any metadata associated with controller packet-in and packet-out. A packet-in is defined as a data plane packet that is sent by the P4Runtime server to the control plane for further inspection. Similarly, a packet-out is defined as a data packet generated by the control plane and injected in the data plane via the P4Runtime server.

When inspecting a packet-in, the control plane might need to have access to additional information such as the original data plane port where the packet was received, the timestamp when the packet was received, if the packet is a clone, etc. Similarly, when sending a packet-out, the control plane might need to specify additional information used by the device to process the data packet.

Such additional information for packet-in and packet-out can be expressed by means of P4 headers carrying P4 standard annotations @controller_header("packet_in") and @controller_header("packet_out"), respectively. ControllerPacketMetadata messages capture the information contained within these special headers and are needed by the P4Runtime server to process packet-in and packet-out stream messages (see section on Packet I/O stream messages).

A P4Info message can contain at most two ControllerPacketMetadata messages, one describing the packet-in header, and the other the packet-out header. Each message contains the following fields:

As an example, consider the following snippet of a P4 program where controller headers are specified and we show the corresponding ControllerPacketMetadata messages.

@controller_header("packet_out")
header PacketOut_t {
  bit<9> egress_port; /* suggested port where the packet
                         should be sent */
  bit<8> queue_id;    /* suggested queue ID */
}

@controller_header("packet_in")
header PacketIn_t {
  bit<9> ingress_port; /* data plane port ID where
                          the original packet was received */
  bit<1> is_clone;     /* 1 if this is a clone of the
                          original packet */
}
controller_packet_metadata {
  preamble {
    id: 2868916615
    name: "packet_out"
    annotations: "@controller_header(\"packet_out\")"
  }
  metadata {
    id: 1
    name: "egress_port"
    bitwidth: 9
  }
  metadata {
    id: 2
    name: "queue_id"
    bitwidth: 8
  }
}

controller_packet_metadata {
  preamble {
    id: 2868941301
    name: "packet_in"
    annotations: "@controller_header(\"packet_in\")"
  }
  metadata {
    id: 1
    name: "ingress_port"
    bitwidth: 9
  }
  metadata {
    id: 2
    name: "is_clone"
    bitwidth: 1
  }
}

Note that the use of @controller_header is optional for Packet I/O. The P4 program may define controller headers without this annotation and use them to encapsulate controller packets. However, in this case the client will be responsible for extracting the metadata from the serialized header in packet-in messages and for serializing the metadata when generating packet-out messages.

6.4.7. ValueSet

ValueSet messages are used to specify all possible P4 Parser Value Sets. Parser Value Sets can be used by the control plane to specify runtime matches used by the P4 parser to determine transitions from one state to another. For more information on Parser Value Sets, refer to the P416 specification [45].

The ValueSet message defines the following fields:

According to the P4 specification, the type parameter of a Value Set, which defines the type of the expression that can be matched against the Value Set in a parser transition, and therefore determines the format of the members that can be inserted into the Value Set by the control plane, must be one of bit<W>, tuple, or struct [31]. The rest of this section looks at all 3 of these cases and gives an example ValueSet message when appropriate.

  1. If the type parameter is bit<W>, match will include exactly one MatchField message, with the following fields (if a field is omitted here, it means the default Protobuf value should be used):

    • id: set to 1
    • bitwidth: set to the value of W
    • match_type: set to EXACT
@id(1) value_set<bit<8> >(4) pvs;
select (hdr.f8) { /* ... */ }
value_sets {
  preamble {
    id: 0x03000001
    name: "pvs"
  }
  match {
    id: 1
    bitwidth: 8
    match_type: EXACT
  }
  size: 4
}
  1. If the type parameter is a tuple, this version of P4Runtime does not support runtime programming of the Value Set. If the P4Info message is generated by a compiler, and the P4 program includes such a Value Set, the compiler must reject the program.

  2. If the type parameter is a struct, this version of P4Runtime requires that all the fields of the struct be of type bit<W> (where W can be different for each field). Otherwise, if the P4Info message is generated by a compiler, the compiler must reject the program. If the Value Set is supported, the match field will include one MatchField message for each field in the struct, with the following fields:

    • id: must be unique with respect to the other match entries. If the P4Info message was generated from a P4 compiler, we recommend that the IDs be assigned incrementally, starting from 1, in the same order as the fields in the P4 struct declaration. The P4 programmer can choose the IDs using the @id annotation, or let the compiler choose them.
    • name: set to the name of the corresponding struct field.
    • annotations: set to the list of P4 annotations associated with the struct field, except for the @match annotation, if present (see the match field below).
    • bitwidth: set to the value of W for the corresponding struct field.
    • type_name, which indicates whether the struct field has a user-defined type; this is useful for translation.
    • match: by default match_type is set to EXACT; the P4 programmer can specify a different match type by using the @match annotation [31].
    • doc: documentation associated with the struct field.
struct match_t {
  @id(1) bit<8> f8;
  @id(2) @match(ternary) bit<16> f16;
  @id(3) @match(custom) bit<32> f32;
}
@id(1) value_set<match_t>(4) pvs;
select ({ hdr.f8, hdr.f16, hdr.f32 }) { /* ... */ }
value_sets {
  preamble {
    id: 0x03000001
    name: "pvs"
  }
  match {
    id: 1
    name: "f8"
    bitwidth: 8
    match_type: EXACT
  }
  match {
    id: 2
    name: "f16"
    bitwidth: 16
    match_type: TERNARY
  }
  match {
    id: 3
    name: "f32"
    bitwidth: 32
    other_match_type: "custom"
  }
  size: 4
}

In the above example, the @id annotations on the P4 struct fields are optional. When omitted, the compiler will choose appropriate IDs.

Although not mentioned in the P4 specification, P4Runtime also supports the cases where the Value Set type parameter is a user-defined type that resolves to a bit<W>, or a struct where one or more fields is a user-defined type that resolves to a bit<W>. For each MatchField that corresponds to a user-defined type, the type_name field must be set to the appropriate value (i.e. the name of the type).

6.4.8. Register

Register messages are used to specify all possible instances of Register PSA externs.

Registers are stateful memories that can be read and written by data plane during packet forwarding. The control plane can also access registers at runtime.

The Register message defines the following fields:

6.4.9. Digest

Digest messages are used to specify all possible instances of Packet Digest PSA externs.

A packet digest is a mechanism to efficiently send notifications from the data plane to the control plane. This mechanism differs from packet-in which is generally used to send entire packets (headers plus payload), each one as a separate P4Runtime stream message. A digest for a packet has a size typically much smaller than the packet itself, as it can be used to send only a subset of the headers or P4 metadata associated with the packet. To reduce the rate of messages sent to the control plane, a P4Runtime server can combine digests for multiple packets into larger messages.

The Digest message defines the following fields:

6.4.10. Extern

Extern messages are used to specify all extern instances across all extern types for a non-PSA architecture. This is useful when extending P4Runtime to support a new architecture. Each architecture-specific extern type corresponds to at most one Extern message instance in P4Info. The Extern message defines the following fields:

If the P4 program does not include any instance of a given extern type, the Extern message instance for that type should be omitted from the P4Info.

6.5. Support for Arbitrary P4 Types with P4TypeInfo

See section on Representation of Arbitrary P4 Types.

7. P4 Forwarding-Pipeline Configuration

The ForwardingPipelineConfig captures data needed to realize a P4 forwarding-pipeline and map various IDs passed in P4Runtime entity messages. It is formally called the “Device Configuration” and sometimes also referred to as the “P4 Blob”. It is defined as:

message ForwardingPipelineConfig {
  config.P4Info p4info = 1;
  bytes p4_device_config = 2;
  message Cookie {
    uint64 cookie = 1;
  }
  Cookie cookie = 3;
}

The p4info field captures the P4 program metadata as described by the P4Info. This message is the output of the P4 compiler and is target-agnostic.

The p4_device_config is opaque binary data which contains the target-specific configuration to realize the P4 program. The P4 program running on a target is changed by loading a new ForwardingPipelineConfig on that target.

The cookie field is opaque data which may be used by a control plane to uniquely identify a forwarding-pipeline configuration among others managed by the same control plane. For example, a controller can compute its value using a hash function over the P4Info and/or target-specific binary data. However, there are no restrictions on how such value is computed, or where this is stored on the target, as long as it is returned with a GetForwardingPipelineConfig RPC. When writing the config via a SetForwardingPipelineConfig RPC, the cookie field is optional. For this reason, the actual value is wrapped in its own message to clearly identify cases where a cookie is not present.

8. General Principles for Message Formatting

8.1. Default-valued Fields

There is a subtle distinction between the treatment of default-valued scalar fields vs default-valued message fields in P4Runtime.

8.1.1. Set / Unset Scalar Fields

In Protobuf version 3 (proto3), the default value of scalar fields is 0 for numeric types such as int32, and the empty string "" for string types (string and bytes). An application, such as the P4Runtime client or server, is unable to distinguish between an unset scalar field and a scalar field set to its default value. Therefore, we usually reserve the default values 0 and “” of scalar fields to mean “unset”.

In particular, 0 is not a valid P4 object ID and it is an error to specify 0 for any P4 object ID in a non-read request towards the server, such as in a WriteRequest or a SetForwardingPipelineConfigRequest.

In contrast to scalar fields, note that for message fields, we often do make a distinction between an unset message field vs a message field set to its default value, see the next section.

8.1.2. Set / Unset Message Fields

In Protobuf version 3 (proto3), the default value for a message field is “unset” [6]. An application, such as the P4Runtime client or server, is able to distinguish between an unset message field and a message field set to its default value. We often use this distinction in P4Runtime and the meaning of a message can vary based on which of its message fields are set. For example, when reading values from an indirect PSA counter using the CounterEntry message, an “unset” index field means that all entries in the counter array should be read and returned to the P4Runtime client (we refer to this as a wildcard read). On the other hand, if the index message field is set, a single entry will be read.

Let's look at the counter example in more details. Based on this specification document, the C++ server code which processes CounterEntry messages may look like this:

auto *counter_entry = ...
if (counter_entry->has_index()) {
  auto index = counter_entry->index().index();
  read_one_entry(counter_entry->id(), index);
} else {
  read_all_entries(counter_entry->id());
}
  1. Reading a single counter entry at index 0 in the counter array with id <id>:

    • Here is the C++ client code:
      p4::v1::CounterEntry entry;
      entry.set_counter_id(<id>);
      entry.mutable_index();
      // The above line sets the index field; it is equivalent to:
      // auto *index = entry.mutable_index();
      // index->set_index(0);
    • Here is the corresponding Protobuf message in text format:
      counter_id: <id>
      index {}
    • Expected behavior: Counter entry at index 0 is read. Notice that the index subfield is missing under the index field message of CounterEntry in the text dump of the message. This is because the subfield is a scalar numeric type and 0 is therefore its default value. Scalar fields with default values are omitted from the textual representation of Protobuf messages.
  2. Reading all counter entries by leaving the index field unset

    • Here is the C++ client code:
      p4::v1::CounterEntry entry;
      entry.set_counter_id(<id>);
    • Here is the corresponding Protobuf message in text format:
      counter_id: <id>
    • Expected behavior: All counter entries for the provided counter instance are read. Notice that the index message field is unset (default value) and is therefore omitted from the textual representation of the message.

8.2. Read-Write Symmetry

The reads and writes a client issues towards a server should be symmetrical and unambiguous. More specifically, if a client writes a P4 entity and then reads it back then the client should expect that the message it wrote and the message it read should match if the RPCs finished successfully (with the exception of parts of the response known to be data plane volatile, as explained in section on Data plane volatile objects). Consider the following pseudocode as an example:

intended_value = value

status = server.write(intended_value, p4_entity)
observed_value = server.read(p4_entity)

assert(intended_value == observed_value)

To ensure read-write symmetry, the rest of this document tries to offer canonical representations for various data types, but this principle should be thought of where it falls short. Ensuring this will allow client software to recover programmatically from failures that can affect the switch stack software, communication channel, or the client replicas. If Read RPC returns a semantically-same but syntactically-different response then the client would have to canonicalize the read values to check its internal state, which only pushes the protocol's complexities to the client implementations.

In order to avoid placing too much burden on the P4Runtime server implementation, we do not in general mandate that the order of values in a Protobuf repeated field be preserved. For example, the server is not required to preserve the order of the match fields in a TableEntry message. If there is a specific case for which the order is significant and / or needs to be preserved, it will be explicitly stated in this document. The MessageDifferencer class [42] included in the Protobuf C++ API supports comparing messages while treating repeated fields as sets, so that different orderings of the same elements are considered equal. This method of comparing Protobuf messages may come at a cost in performance.

8.2.1. Data plane volatile objects

An exception to read-write symmetry are objects whose contents or fields can change by the action of the data plane alone, even if no controller modifies them. These objects are called data plane volatile.

The following sections describe all possible values of an Entity message, since these are the messages that a controller can use to modify objects in the data plane via an Update message. For each, a description is given of the parts of that entity that are data plane volatile.

8.2.1.1. ExternEntry

Data plane volatility depends upon the definition of the extern and its control plane API.

8.2.1.2. TableEntry

For a table with a direct counter associated with it, the counter_data field of a TableEntry can be modified by the data plane when packets match the entry.

For a table with a direct meter associated with it, the meter_counter_data field of a TableEntry can be modified by the data plane when packets match the entry.

For a PSA [24] table with property psa_idle_timeout equal to PSA_IdleTimeout_t.NOTIFY_CONTROL, the data plane can modify the elapsed_ns field of a TableEntry when no packets match the entry for an implementation-specific amount of time.

For a PNA [23] table with property pna_idle_timeout equal to PNA_IdleTimeout_t.NOTIFY_CONTROL or PNA_IdleTimeout_t.AUTO_DELETE the data plane can modify the elapsed_ns field of a TableEntry when no packets cause the extern function restart_expire_timer to be called for an implementation-specific amount of time (nor any other extern function defined to also have the same effect as restart_expire_timer).

Similarly, for a table in PNA with any of the values of pna_idle_timeout listed above, the data plane can modify the idle_timeout_ns field of a TableEntry when packets match the entry and the action calls the set_entry_expire_time extern function (or any of the other extern functions defined to have an effect similar to calling set_entry_expire_time).

For a PNA [23] table with the property add_on_miss equal to true the data plane can insert new entries into the table without any controller's involvement.

For a PNA [23] table with the property pna_idle_timeout equal to PNA_IdleTimeout_t.AUTO_DELETE, the data plane can delete existing entries from the table without any controller's involvement.

8.2.1.3. ActionProfileMember

Not data plane volatile in any architectures defined by P4.org specifications.

8.2.1.4. ActionProfileGroup

Not data plane volatile in any architectures defined by P4.org specifications. The watch_port feature does affect how action selectors behave while processing packets, but this feature does not affect what a P4Runtime client sees when it reads the configuration.

8.2.1.5. MeterEntry

The field counter_data is modified by the data plane when the corresponding meter is updated in the data plane.

8.2.1.6. DirectMeterEntry

The field counter_data is modified by the data plane when the corresponding meter is updated in the data plane.

8.2.1.7. CounterEntry

The field data is modified by the data plane when the corresponding counter is updated in the data plane.

8.2.1.8. DirectCounterEntry

The field data is modified by the data plane when the corresponding counter is updated in the data plane.

8.2.1.9. PacketReplicationEngineEntry

Not data plane volatile in any architectures defined by P4.org specifications.

8.2.1.10. ValueSetEntry

Not data plane volatile in any architectures defined by P4.org specifications.

8.2.1.11. RegisterEntry

The field data can be modified by the data plane when the corresponding register entry is updated in the data plane.

8.2.1.12. DigestEntry

Not data plane volatile in any architectures defined by P4.org specifications.

8.3. Bytestrings

P4Runtime integer values may be too large to fit in Protobuf primitive data types (32-bit and 64-bit words). The P4 language does not put any limit on the size of integer values, whether unsigned (bit<W>) or signed (int<W>), and it is up to the P4 programmer to choose the appropriate sizes. Because of this flexibility, P4Runtime represents P4 integer values as binary strings, using the bytes Protobuf type. The correct bitwidth as per the P4 program of each integer variable exposed through P4Runtime is specified in the P4Info message.

The canonical binary string representation uses the shortest string that fits the encoded integer value. This representation achieves three goals:

In particular, the kinds of P4 program updates that this representation facilitates are those where a P4Runtime server and client can continue to transmit P4Runtime messages between them when one has a P4Info file for version A of a P4 program, at the same time that the other has a P4Info file for version B of a P4 program, and those P4 programs differ in the bitwidths of some values of type bit<W> and/or int<W>.

Note that this representation does not make it possible to seamlessly change the type of a value from signed to unsigned, or vice versa. If you attempt to do so, this mechanism can quietly change negative signed values to positive unsigned values, or vice versa. It also limits the magnitude of the values transmitted to those that fit within the smaller of the bitwidths supported by either end of the message transmission. If a message sender attempts to send a value larger than the receiver expects, the receiver will detect it as out of range.

In the P4Runtime API version 1.0 (including minor version revisions), values of table key fields, action parameters, and fields in packet-in and packet-out headers between a device and the controller (see 6.4.6), may not be of type int<W>. The rules for encoding signed values thus only apply to messages of type P4Data (see 8.4.3).

For a value of type bit<W>, the fewest number of bits required to represent the integer value $V > 0$ is the smallest integer $A$ such that $V \leq 2^A -1$.

For a value of type int<W>, the fewest number of bits required to represent the integer value $V \neq 0$ in 2's complement form is the smallest integer $A$ such that $-2^{A-1} \leq V \leq 2^{A-1} - 1$.

As a special case, define that the value $V=0$ requires at least $A=1$ bit to represent, regardless of whether it is signed or unsigned.

The shortest possible binary string for an integer $V$ that needs $A$ bits to represent it is computed as:

minimum_string_size = floor((A + 7) / 8)

Binary strings with the byte length computed as minimum_string_size promote P4Runtime read-write symmetry in both client-to-server requests and server-to-client replies.

Any additional bits in the bytes sent for an unsigned integer value (type bit<W>) must be 0. If additional bytes are transmitted above the minimum_string_size minimum required, they must be filled with 0.

Any additional bits in the bytes sent for a signed integer value (type int<W>) must be copies of the sign bit, i.e. 0 for non-negative values, or 1 for negative values. If additional bytes are transmitted above the minimum_string_size minimum required, they must be filled with copies of the sign bit, i.e. 0 for non-negative values, or 0xff for negative values. In 2's complement representation, this is called “sign extension”, and leaves the numeric value represented unchanged.

Upon receiving a binary string, the P4Runtime receiver (whether the server or the client) does not impose any restrictions on the maximum length of the string itself. Instead, the receiver verifies that the value encoded by the string fits within the expected type (signed or unsigned) and P4Info-specified bitwidth for the P4 object value.

For a received bitstring expected to fit within a bit<W> type, the value it represents is in range if, after removing all most significant 0 bits, the remaining bitstring's width is W bits or less.

For a received bitstring expected to fit within an int<W> type, the value it represents is in range if, after “undoing sign extension”, the remaining bit string's width is W bits or less. To undo sign extension, start by eliminating the most significant bit, but only if it is equal to the bit that follows it (otherwise removing the most significant bit would change the sign of the value). Repeat that process until either only a single bit remains, or until the two most significant bits are different from each other.

If the string's byte length is zero, the server always rejects the string.

When the server rejects a binary string due to any of the previous criteria, it returns an OUT_OF_RANGE error.

For all binary strings, P4Runtime uses big-endian (i.e. network) byte-order. For signed integer values (int<W> P4 type), P4Runtime uses the same two's complement bitwise representation as P4. Table 4 shows various examples of integer values that the server accepts as valid P4Runtime binary strings according to the criteria in the list above.

P4 type Integer value P4Runtime binary string Read-write symmetry
bit<8> 99 (0x63) \x63 yes
bit<16> 99 (0x63) \x00\x63 no
bit<16> 99 (0x63) \x63 yes
bit<16> 12388 (0x3064) \x30\x64 yes
bit<16> 12388 (0x3064) \x00\x30\x64 no
bit<12> 99 (0x63) \x00\x63 no
bit<12> 99 (0x63) \x63 yes
bit<12> 99 (0x63) \x00\x00\x63 no
int<8> 99 (0x63) \x63 yes
int<8> -99 (-0x63) \x9d yes
int<8> -99 (-0x63) \xff\x9d no
int<12> -739 (-0x2e3) \xfd\x1d yes
int<16> 0 (0x0) \x00\x00 no
int<16> 0 (0x0) \x00 yes

Table 4. Examples of Valid Bytestring Encoding

Table 5 shows some examples of invalid P4Runtime binary strings:

P4 type P4Runtime binary string
bit<8> \x01\x63
bit<8> empty string
bit<16> \x01\x00\x63
bit<12> \x10\x63
bit<12> \x01\x00\x63
bit<12> \x00\x40\x63
int<8> \x00\x9d
int<12> \x8d\x1d
int<16> empty string

Table 5. Examples of Invalid Bytestring Encoding

As the preceding examples illustrate, a P4Runtime server must accept a wide assortment of possible binary string encodings for the same integer value. This requirement addresses P4 program upgrade scenarios where binary string widths can expand or contract. In some P4Runtime environments, the changes cannot be deployed simultaneously to all P4Runtime clients and servers. Given a hypothetical match field type change from bit<8> to bit<9>, a server running the bit<9> version of the P4 program will accept requests from clients that remain on the bit<8> P4Runtime version.

Despite the server's binary string flexibility for P4 program update support, the client and server must both remain aware of the read-write symmetry requirements. As described earlier, read-write symmetry requires that the encoder of a P4Runtime request or reply uses the shortest strings that fit the encoded integer values.

Representation of variable-length integer values (varbit<W> P4 type) is similar to the representation of fixed-width unsigned integers. We use a binary string, whose length is the dynamic-length of the expression. When the value is provided by the P4Runtime client, the server must verify that the length of the binary string is less than the maximum length specified in the P4 program, and return an OUT_OF_RANGE error code otherwise.

8.4. Representation of Arbitrary P4 Types

8.4.1. Problem Statement

The P416 language includes more complex types than just binary strings [5]. Most of these complex data types can be exposed to the control plane through table key expressions, Value Set lookup expressions, Register (PSA extern type) value types, etc. Not supporting these more complex types can be very limiting. Table 6 shows the different P416 types and how they are allowed to be used, as per the P416 specification.

Container type
Element type header header_union struct or tuple
bit<W> allowed error allowed
int<W> allowed error allowed
varbit<W> allowed error allowed
int error error error
void error error error
error error error allowed
match_kind error error error
bool error error allowed
enum allowed1 error allowed
header error allowed allowed
header stack error error allowed
header_union error error allowed
struct error error allowed
tuple error error allowed

Table 6. P4 Type Usage

For example, the following P416 objects involve complex types that need to be exposed in P4Runtime in order to support runtime operations on these objects.

Digest<tuple<bit<4>, bit<8> > >() digest_complex;
digest_complex.pack({ hdr.ipv4.version, hdr.ipv4.protocol });
// ...
header_union ip_t {
   ipv4_t ipv4;
   ipv6_t ipv6;
}
Register<ip_t, bit<32> >(128) register_ip;

One solution would be to use only binary string (bytes type) in p4runtime.proto and to define a custom serialization format for complex P416 types. The serialization would maybe be trivial for header types but would require some work for header unions, header stacks, etc. For example, in the case of a PSA Register storing header unions, a client reading from that Register would need to receive information about which member header is valid, in addition to the binary contents of this header. Rather than coming-up with a serialization format from scratch, we decided to use a Protobuf representation for all P416 types.

8.4.2. P4 Type Specifications in p4info.proto

In order for the P4Runtime client to generate correctly-formatted messages and for the P4Runtime service implementation to validate them, P4Info needs to specify the type of each P4 expression which is exposed to the control plane. In the Register example above, client and server need to know that each element of the register has type ip_t, which is a header union with 2 possible headers: ipv4 with type ipv4_t and ipv6 with type ipv6_t. Similarly, they need to know the field layout for both of these header types.

To achieve this we introduce 2 main Protobuf messages: P4TypeInfo and P4DataTypeSpec.

P4TypeInfo is a top-level member of P4Info and includes Protobuf maps storing the type specification for all the named types in the P416 program. These named types are struct, header, header_union, enum and serializable_enum; for each of these we have a type specification message, respectively P4StructTypeSpec, P4HeaderTypeSpec, P4HeaderUnionTypeSpec, P4EnumTypeSpec and P4SerializableEnumTypeSpec. We preserve P4 annotations for named types, which is useful to identify well-known headers, such as IPv4 or IPv6. P4TypeInfo also includes the list of parser errors for the program, as a P4ErrorTypeSpec message.

P4DataTypeSpec is meant to be used in P4Info, to specify the expected format of the P4-dependent values being exchanged between the P4Runtime client and server. Each P4DataTypeSpec message corresponds to a compile-time type in the original P416 program (e.g. the type parameter of an extern). This compile-time type is represented as a Protobuf oneof, which can be:

For all P416 compound types (tuple, struct, header, and header_union), the order of members in the repeated field of the Protobuf type specification is guaranteed to be the same as the order of the members in the corresponding P416 declaration. The same goes for the order of members of an enum (serializable or not) or members of error.

8.4.3. P4Data in p4runtime.proto

P4Runtime uses the P4Data message to represent values of arbitrary types. The P4Runtime client must generate correct P4Data messages based on the type specification information included in P4Info. The P4Data message was designed to introduce little overhead compared to using binary strings in the most common case (P416 bit<W> type).

Just like its P4Info counterpart - P4DataTypeSpec -, P4Data uses a Protobuf oneof to represent all possible values.

We define a canonical representation for P4Data messages therefore guaranteeing read-write symmetry by introducing the following requirements:

8.4.4. Example

Let's look at the Register example again:

header_union ip_t {
   ipv4_t ipv4;
   ipv6_t ipv6;
}
Register<ip_t, bit<32> >(128) register_ip;

Here's the corresponding entry in the P4Info message:

registers {
  preamble {
    id: 369119267
    name: "register_ip"
    alias: "register_ip"
  }
  type_spec {
    header_union {
      name: "ip_t"
    }
  }
  size: 128
}
type_info {
  headers {
    key: "ipv4_t"
    value {
      members {
        name: "version"
        type_spec {

          bit {
            bitwidth: 4
          }
        }
      } # ...
  headers {
    key: "ipv6_t"
    value {
      members {
        name: "version"
        type_spec {
          bit {
            bitwidth: 4
          }
        }
      } # ...
  header_unions {
    key: "ip_t"
    value {
      members {
        name: "ipv4"
        header {
          name: "ipv4_t"
        }
      }
      members {
        name: "ipv6"
        header {
          name: "ipv6_t"
        }
      }
    }
  }
}

Here's a p4.WriteRequest to set the value of register_ip[12]:

update {
  type: INSERT
  entity {
    register_entry {
      register_id: 369119267
      index {
        index: 12
      }
      data {
        header_union {
          valid_header_name: "ipv4"
          valid_header {
            is_valid: true
            bitstrings: "\x04"
            bitstrings: # ...
          }
        }
      }
    }
  }
}

8.4.5. enum, serializable enum and error

P416 supports 2 different classes of enumeration types: without underlying type (safe enum) and with underlying type (serializable enum or “unsafe” enum) [7]. For enum types with no underlying type as well as error there is no integer value associated with each symbolic member entry (whether assigned automatically by the compiler or directly in the P4 source). We therefore use a human-readable string in P4Data to represent enum and error values.

Serializable enum types have an underlying fixed-width unsigned integer representation (bit<W>). All named enum members must be assigned an integer value by the P4 programmer, but not all valid numeric values for the underlying type need to have a corresponding name. P4TypeInfo includes the mapping between entry name and entry value. When providing serializable enum values through P4Data, one must use the assigned integer value (enum_value bytestring field). P4Runtime does not provide a way for the client to use the name even when the enum member has one instead of the value, as it makes it easier for the server to respect the read-write symmetry principle.

8.4.6. User-defined types

P416 enables programmers to introduce new types [13]. While similar to typedef, this mechanism introduces in fact a new type, which is not a strict synonym of the original type. It is important to preserve this distinction in the P4Info message, in particular for the purposes of translation. When introducing a new type, the declaration can be annotated with @p4runtime_translation to indicate that the type exposed to the P4Runtime client is different from the original P4 type. One important use-case is for port numbers, whose underlying data plane representation may vary on different targets, but for which it may be convenient to present a unified representation and numbering scheme to the control plane. The @p4runtime_translation annotation can only be used if the underlying P4 built-in type is a fixed-width unsigned bitstring type (bit<W>). The type exposed to the control plane (referred to as the controller_type) can be either a fixed-width unsigned bitstring, with a potentially different bitwidth, or a string. The annotation takes two parameters: a URI (Uniform Resource Identifier) which uniquely identifies the translation being performed on entities of the new type to the P4Runtime server and the controller_type of the values exposed to the control plane. The controller_type can be either bit<W> where W is any positive integer, or string, or a positive integer W which has the same meaning as bit<W>. Specifying just an integer is deprecated.

It is recommended that the URI includes at least the P4 architecture name and the type name.

A @p4runtime_translation annotation need not have any effect if it is used to annotate anything in a P4 program other than a type declaration. It is recommended that P4 development tools have an option to issue a warning if such an annotation is used in a part of a P4 program where it has no effect.

User-defined types are specified using the P4NewTypeSpec message, which has the following fields:

For example, an architecture in this case PSA may introduce a new type for port numbers:

// Controller refers to ports as a string, e.g. "eth0".
@p4runtime_translation("p4.org/psa/v1/PortId_String_t", string)
type bit<9> PortId_String_t;

// Controller refers to ports as a 32-bit integer, e.g. 0xffffffff.
@p4runtime_translation("p4.org/psa/v1/PortId_Bit32_t", bit<32>)
type bit<9> PortId_Bit32_t;

// Same as above.
@p4runtime_translation("p4.org/psa/v1/PortId_32_t", 32)
type bit<9> PortId_32_t;

In this case, the P4Info message would include the following P4TypeInfo messages:

type_info {
  new_types {
    key: "PortId_String_t"
    value {  # P4NewTypeSpec
      translated_type {  # P4NewTypeTranslation
        uri: "p4.org/psa/v1/PortId_String_t"
        sdn_string: {}
      }
    }
  }
}

type_info {
  new_types {
    key: "PortId_Bit32_t"
    value {  # P4NewTypeSpec
      translated_type {  # P4NewTypeTranslation
        uri: "p4.org/psa/v1/PortId_Bit32_t"
        sdn_bitwidth: 32
      }
    }
  }
}

type_info {
  new_types {
    key: "PortId_32_t"
    value {  # P4NewTypeSpec
      translated_type {  # P4NewTypeTranslation
        uri: "p4.org/psa/v1/PortId_32_t"
        sdn_bitwidth: 32
      }
    }
  }
}

Note that a P4 compiler may provide a mechanism external to the language to specify if and how a user-defined type is to be translated (e.g. through some configuration file passed on the command-line when invoking the compiler). This mechanism should take precedence over @p4runtime_translation to enable users to overwrite annotations included as part of the P4 architecture definition.

8.4.7. Trade-off for v1.x Releases

For the v1.x release ofs P4Runtime, it was decided not to replace occurrences of bytes with P4Data in the p4.v1.FieldMatch message, which is used to represent table and Value Set entries. This is to avoid breaking pre-v1.0 implementations of P4Runtime. Similarly it has been decided to keep using bytes to provide action parameter values and controller metadata values. However P4Data is used whenever appropriate for PSA externs and we encourage the use of P4Data in architecture-specific extensions.

In order to support translation for action parameters and match fields, we include a type_name field in p4.config.v1.MatchField, p4.config.v1.Action.Param and p4.config.v1.ControllerPacketMetadata.Metadata. In addition, the bitwidth field for all of these message types must abide by the following rule when type_name names a translated user-defined type:

For example, consider the following P4 snippet, which declares a P4 table matching on two expressions with translated user-defined types:

@p4runtime_translation("p4.org/psa/v1/PortId_String_t", string)
type bit<9> PortId_String_t;

@p4runtime_translation("p4.org/psa/v1/PortId_Bit32_t", bit<32>)
type bit<9> PortId_Bit32_t;

@name(".t")
table t {
  key = {
    meta.port1: exact;  // meta.port1 has type PortId_String_t
    meta.port2: exact;  // meta.port2 has type PortId_Bit32_t
  }
  actions = {
    drop;
  }
}

This table would have the following representation in the generated P4Info message:

tables {
  preamble {
    id: 34728461
    name: "t"
    alias: "t"
  }
  match_fields {
    id: 1
    name: "meta.port1"
    # notice that bitwidth is unset
    match_type: EXACT
    type_name {
      name: "PortId_String_t"
    }
  }
  match_fields {
    id: 2
    name: "meta.port2"
    bitwidth: 32
    match_type: EXACT
    type_name {
      name: "PortId_Bit32_t"
    }
  }
  # ...
}

9. P4 Entity Messages

P4Runtime covers P4 entities that are either part of the P416 language, or defined as PSA externs. The sections below describe the messages for each supported entity.

9.1. TableEntry

The match-action table is the core packet-processing construct of the P4 language. It consists of a collection of table entries, or flow rules, each mapping a key value to a P4 action along with input values for the action's parameters. Packets are looked-up in the table by matching them against the flow rules. In case of a match, the corresponding action is applied on the packet, otherwise, a default action is applied. The exact behavior of P4 tables is described in the P4 specification.

P4Runtime supports inserting, modifying, deleting and reading table entries with the TableEntry entity, which has the following fields:

The priority field must be set to a non-zero value if the match key includes a ternary match (i.e. in the case of PSA if the P4Info entry for the table indicates that one or more of its match fields has an OPTIONAL, TERNARY or RANGE match type) or to zero otherwise. A higher priority number indicates that the entry must be given higher priority when performing a table lookup. Clients must allow multiple entries to be added with the same priority value. If a packet can match multiple entries with the same priority, it is not deterministic in the data plane which entry a packet will match. If a client wishes to make the matching behavior deterministic, it must use different priority values for any pair of table entries that the same packet matches.

The match and priority fields are used to uniquely identify an entry within a table. Therefore, these fields cannot be modified after the entry has been inserted and must be provided for MODIFY and DELETE updates. When deleting an entry, these key fields (along with is_default_action) are the only fields considered by the server. All other fields must be ignored, even if they have nonsensical values (such as an invalid action field). In the case of a keyless table (the table has an empty match key), the server must reject all attempts to INSERT a match entry and return an INVALID_ARGUMENT error.

The number of match entries that a table should support is indicated in P4Info (size field of Table message). The guarantees provided to the P4Runtime client are the same as the ones described in the P416 specification for the size property [36]. In particular, some implementations may not be able to always accommodate an arbitrary set of entries up to the requested size, and other implementations may provide the P4Runtime client with more entries than requested. The P4Runtime server must return RESOURCE_EXHAUSTED when a table entry cannot be inserted because of a size limitation. It is recommended that, for the sake of portability, P4Runtime clients do not try to insert additional entries once the size indicated in P4Info has been reached.

The is_const field must be false in any INSERT, MODIFY, or DELETE write request of a table entry. If it is true, the server must reject the operation and return an INVALID_ARGUMENT error.

9.1.1. Match Format

The bytes fields in the FieldMatch message follow the format described in Bytestrings.

For “don't care” matches, the P4Runtime client must omit the field's entire FieldMatch entry when building the match repeated field of the TableEntry message. This requirement leads to smaller Protobuf messages overall, while enabling a canonical representation for “don't care” matches, which is needed to ensure read-write symmetry. For PSA match types, a “don't care” match for a specific match key field is defined as follows:

Note that there is no “don't care” value for EXACT matches and therefore exact match fields can never be omitted from the TableEntry message.

The following example shows a P4Runtime message that treats a TERNARY field as a “don't care” match. The P4 program defines table t with TERNARY and EXACT fields in its match key:

table t {
  key = {
    hdr.ipv4.dip: ternary;
    istd.ingress_port: exact;
  }
  actions = {
    drop;
  }
}

In this P4Runtime request, the client omits the table's TERNARY field from the repeated match field to indicate a “don't care” match. As shown below, the match specifies only the EXACT field given by field_id: 2.

device_id: 3
entities {
  table_entry {
    table_id: 33554439  # Table t's ID.
    match {
      # field_id 1 is not present to use the don't care ternary value.
      field_id: 2
      exact {
        value: "\x20"
      }
    }
    action {
      # Action selection goes here.
    }
  }
}

For every member of the TableEntry repeated match field, field_id must be a valid id for the table, as per the P4Info, and one of the fields in field_match_type must be set. We summarize additional constraints which depend on the match-type in the following list. If any one of them is violated, the P4Runtime server must return an INVALID_ARGUMENT error code.

assert(BytestringValid(match.exact().value()))
assert(BytestringValid(match.lpm().value()))

pLen = match.lpm().prefix_len()
assert(pLen > 0)

trailing_zeros = countTrailingZeros(match.lpm().value())
assert(trailing_zeros >= field_bits - pLen)
assert(BytestringValid(match.ternary().value()))
assert(BytestringValid(match.ternary().mask()))
assert(match.ternary().value().size() <= match.ternary().mask().size());

value = parseInteger(match.ternary().value())
mask = parseInteger(match.ternary().mask())

assert(mask != 0)

assert(value & mask == value)
assert(BytestringValid(match.range().low()))
assert(BytestringValid(match.range().high()))

low = parseInteger(match.range().low())
high = parseInteger(match.range().high())

assert(low <= high)

assert(low != min_field_value || high != max_field_value)
assert(BytestringValid(match.optional().value()))

9.1.2. Action Specification

The TableEntry action field must be set for every INSERT update but may be left unset for MODIFY updates, in which case the action specification for the table entry will not be modified (if a MODIFY update has an unset action field and does not modify any direct resource attached to the table then the MODIFY update is a no-op). Based on the implementation property value of the P4 table, the oneof in the TableAction message will either be:

If the oneof does not match the table description in the P4Info (e.g. the oneof is action_profile_member_id for a direct table), the server must return an INVALID_ARGUMENT error code.

The Action Protobuf message has the following fields:

For indirect tables, if the P4Runtime client provides a member or group id which has not been inserted in the corresponding action profile instance yet, the P4Runtime server must return a NOT_FOUND error code.

9.1.3. Default Entry

According to the P4 specification, the default entry for a table is always set. It can be set at compile-time by the P4 programmer or defaults to NoAction (which is a no-op) otherwise and assuming it is not declared as const, can be modified by the P4Runtime client. Because the default entry is always set, we do not allow INSERT and DELETE updates on the default entry and the P4Runtime server must return an INVALID_ARGUMENT error code if the client attempts one.

The default entry is identified by setting the is_default_action boolean field to true. When this flag is set to true, the repeated match field must be empty and the priority field must be set to zero, otherwise the P4Runtime server must return an INVALID_ARGUMENT error code. When performing a MODIFY update on the default entry, the client can either provide a valid action for the table or leave the action field unset, in which case the default entry will be reset to its original value, as defined in the P4 program. When resetting the default entry, its controller_metadata and metadata value as well as the configurations for its direct resources will be reset to their defaults. If the default entry is constant (as indicated by the P4 program and the P4Info message), the server must return a PERMISSION_DENIED error code if the client attempts to modify it.

Apart from the above restrictions, the default entry is treated like a regular entry, including with regards to direct resources.

In this P4Runtime release, we have decided to restrict the default entry for indirect tables tables with an ActionProfile or ActionSelector implementation property to a constant NoAction action entry, with the hope that it would simplify the implementation of the P4Runtime service.

9.1.4. Constant Tables

Constant tables are defined as tables whose match entries are immutable. They are identified by the table property const entries in the P416 source code. In the P4Info, such tables have is_const_table equal to true, and if the list of entries in the source code has at least one entry in it, they also have has_initial_entries flag equal to true. For tables declared with the entries property, without const before entries see Section 9.1.5.

The only write updates which are allowed for constant tables are MODIFY operations on direct resources, and the default action (assuming the default action itself is not constant). If the P4Runtime client attempts to perform any other kind of write update on a constant table the server must return a PERMISSION_DENIED error. Just like any table entry MODIFY request, the request must specify the match fields and priority to identify the table entry. Because the action of a const entry cannot be modified, the request may not specify an action, even if the action is equal to the existing action. If an action is specified, the switch will return a PERMISSION_DENIED error.

The contents of const tables can be queried by the client through a ReadRequest. When reading static (immutable) entries from a constant table, the following fields must be set by the server: table_id, match, action, is_default_action, and priority (if required). This is in addition to any direct resources that are being queried. Idle timeouts are not supported for static entries.

When a priority value is required (e.g. for tables including RANGE, TERNARY or OPTIONAL matches), it is inferred based on the order in which entries appear in the table declaration. As of August 2023, the open source p4c compiler always assigns entry priority values in a constant table with N entries starting at N for the first entry and decrementing the value by 1 for each successive entry. The P416 language specification does not preclude the P4 developer from explicitly specifying priorities for entries in constant tables, but p4c does not yet support this.

9.1.5. Preinitialized tables

Preinitialized tables are those defined with an entries table property in the P416 source code, with no const qualifier before entries, and at least one entry in that list. In the P4Info, such tables have has_initial_entries flag equal to true, but is_const_table is false. For tables declared with const entries, see section on Constant Tables.

Every P4 table falls into one of three categories:

A preinitialized table is allowed to have a mix of some entries marked const, and other entries not marked const.

Entries not marked const may be modified or deleted, just as a client may do for any entry in a normal table.

Entries marked const behave like entries in a constant table, i.e. only MODIFY operations on direct resources are allowed.

Unlike a table with is_const_table = true, a client may insert entries into a table with has_initial_entries = true and is_const_table = false, subject to capacity constraints on the number of entries supported by the target for the table.

The contents of preinitialized tables can be queried by the client through a ReadRequest. The server fills in the same fields in the response as it does for constant tables, as described in section on Constant Tables, and with the same restrictions on table features supported.

If the table requires a priority value for entries, the priorities of the initial entries are determined according to the P416 language specification. After the P4 program is initially loaded, the entries not marked const can be modified at run time just as table entries in a normal table can.

The contents of all table entries within the entries table properties in a P4 program can be written to a separate output file by the open source p4c compiler. See Section A.5 for details.

9.1.6. Wildcard Reads

When performing a ReadRequest, the P4Runtime client can select all entries from one or all tables on the target and use several of the TableEntry fields to filter the results, much like when performing a SQL request. For each field that can be used to filter the result, the client may use the default value for the field to act as a wildcard. This default value is zero for scalar fields such as priority and “unset” for message fields such as match. The following fields may be used to select and filter results:

For example, in order to read all entries from all tables from device 3, the client can use the following ReadRequest message.

device_id: 3
entities {
  table_entry {
    table_id: 0
    priority: 0
    controller_metadata: 0
    metadata: ""
  }
}

In order to read all entries with priority 11 from a specific table (with id 0x0212ab34) from device 3, the client can use the following ReadRequest message:

device_id: 3
entities {
  table_entry {
    table_id: 0x0212ab34
    priority: 11
    controller_metadata: 0
    metadata: ""
  }
}

The canonical representation of “don't care” matches, combined with the ability to do a wildcard read on all table entries by leaving the match field unset, means that there exists a specific ambiguous case in which the same message could be used to either read a single “don't care” entry or to do a wildcard read. If a table has no fields with match kind EXACT, it is possible via P4Runtime to add an entry that is “don't care” for all fields (i.e. has an empty match field) but is not the default entry (i.e. is_default_action is false). When reading this entry from the table, there is no way to read only that entry from the table, because it would require providing an unset match field in the request, which in turn indicates that the client wishes to perform a wildcard read on all non-default entries. Consider the following example which uses a table with a single LPM match:

table t {
  key = {
    hdr.ipv4.dip: lpm;
  }
  actions = {
    drop; fwd;
  }
}

The following WriteRequest message can be used to add 2 entries:

device_id: 3
entities {
  table_entry { # don't care entry
    table_id: 0x0212ab34
    # ...
  }
  table entry {
    table_id: 0x0212ab34
    match {
      field_id: 1
      lpm {
        value: 0x0a000000
        prefix_len: 8
      }
    # ...
  }
}

The first entry is a “don't care” entry, while the second one matches all 10.0.0.0/8 addresses. The second entry has higher priority than the first one.

The following ReadRequest message will return all entries in the table, not just the “don't care” entry.

device_id: 3
entities {
  table_entry {
    table_id: 0x0212ab34
  }
}

This issue also exists for tables with TERNARY, RANGE, and OPTIONAL matches. However, in this case the priority is also taken into account for wildcard reads, and because a priority of 0 is not valid, in practice only the entries with the same priority as the “don't care” entry will be returned to the client. If the client uses distinct priority values for all entries which is strongly recommended to achieve deterministic behavior , then there is no ambiguity because the wildcard read will actually return a single entry (the “don't care” entry) as long as the priority field is set to the correct value.

9.1.7. Direct Resources

In addition to the DirectCounterEntry and DirectMeterEntry entities, P4Runtime support reading and writing direct resources as part of the TableEntry message. This is convenient for two reasons:

Once the table entry has been inserted, the P4Runtime client is free to use the DirectCounterEntry and DirectMeterEntry messages for read and write operations on DirectCounter and DirectMeter instances. For example, it is usually more convenient as well as more efficient to use DirectCounterEntry to query a counter entry value rather than use TableEntry, assuming the client is not interested in reading other table entry properties as well, such as the controller metadata cookie or the action entry.

The PSA specification states that when a table is assigned a direct resource (meter or counter), this direct resource does not need to be “executed” in every action bound to the table. It is an error to provide a direct resource configuration in a TableEntry message when programming an action that does not execute the direct resource, and the server must return an INVALID_ARGUMENT error code.

We leverage Protobuf's ability to differentiate between set and unset fields to give the P4Runtime client fined-grained control over how direct resources are read and written through the TableEntry message. The list below describes how the server must handle the meter_config, counter_data and meter_counter_data fields for read and write requests, based on whether the fields are set or not. We do not cover error cases in the list, i.e. we assume that we are dealing with a table which is assigned a direct counter / a direct meter, and that the action being used for the table entry “executes” the direct resource appropriately.

In its default configuration, a meter returns the GREEN color for every packet when it is executed. This default configuration can be achieved by leaving the meter_config field unset when inserting or modifying a table entry. When modifying a table entry, if the P4Runtime client wishes to maintain the same meter configuration, it needs to be provided again in the TableEntry message (i.e. the meter_config field must be set to match the existing configuration).

9.1.8. Idle-timeout

P4Runtime supports idle timeout for table entries. When adding a table entry, the client can specify a Time-To-Live (TTL) value. If at any time during its lifetime, the data plane entry is not “hit” (i.e. not selected by any packet lookup) for a lapse of time greater or equal to its TTL, the P4Runtime should, with best effort, generate a stream notification using the IdleTimeoutNotification message to the primary client, which can then take action, such as remove the idle table entry.

Two fields of the TableEntry Protobuf message are used to implement idle timeout:

These fields can only be set if idle timeout is supported for the table, as per the P4Info message. If idle timeout is not supported by the table, the P4Runtime server must return an INVALID_ARGUMENT error code if at least one of these conditions is met:

The target should do its best to approximate the idle_timeout_ns value provided by the client. For example, most targets may not be able to accommodate arbitrarily small values of TTL, in which case they should use the smallest value they can support, rather than reject the TableEntry write with an error code. Similarly, each target should do its best to provide reasonably-accurate values for time_since_last_hit.

P4Runtime does not support idle timeout for default entries. When the is_default_action flag is set in a TableEntry message, idle_timeout_ns must be set to 0 (default) and time_since_last_hit must be unset. If the server receives a TableEntry message which violates this, it must return an INVALID_ARGUMENT error.

For more information about idle timeout, in particular regarding IdleTimeoutNotification, please refer to the Table idle timeout notifications section.

9.2. ActionProfileMember & ActionProfileGroup

P4Runtime defines an API for programming a PSA ActionProfile extern using ActionProfileMember messages. A PSA ActionSelector extern can be programmed using both ActionProfileMember and ActionProfileGroup messages. PSA supports tables that can be implemented with an action profile or selector instance. Such tables are referred to as indirect tables, in contrast to direct tables, whose entries are directly bound to an action instance. The following P4 snippet illustrates an indirect table t for L3 routing, implemented with an action selector as.

ActionSelector(HashAlgorithm.crc32,
               /*size = */ 32w1024,
               /*output_width = */ 32w10) as;

action set_nhop(PortId_t p, EthAddr smac, EthAddr dmac) {
  istd.egress_port = p;
  hdr.ethernet.smac = smac;
  hdr.ethernet.dmac = dmac;
}

table t {
  key = {
    hdr.ipv4.dip: lpm;  // LPM on destination IP address
  }
  actions = {
    set_nhop;
  }
  implementation = as;
}

When programming table t in the example above, a P4Runtime client should specify the TableAction in the TableEntry to be a reference to either an action profile member or group. The reference is a non-zero uint32 identifier that uniquely identifies a member or group programmed in the action selector as.

If a table entry in an indirect table with an ActionProfile implementation is hit, then the corresponding table action gives a member id. The member table is looked up with the member id, and the corresponding action specification is used to modify the packet or its metadata.

If a table entry in an indirect table with an ActionSelector implementation is hit, then the corresponding table action gives either a member id or a group id. For a member id, the member table in the selector is looked up, and the corresponding action specification is used to modify the packet or its metadata. For a group id, a hash algorithm, defined in the P4 ActionSelector specification is used to obtain a member id from the set of members in the group. For example, the hash algorithm in the P4 example above is 32-bit CRC. The obtained member id is used to look up the member table in the selector and obtain the action specification, which is then used to modify the packet or its metadata.

9.2.1. Action Profile Member Programming

Action profile members are entries in the ActionProfile or ActionSelector and are referenced by a uint32 identifier that is bound to an action specification. An action profile member for an ActionProfile or ActionSelector extern instance may be bound only to the actions that appear in the actions attribute of the table implemented using the extern instance. If multiple table implementations share an extern instance, then the actions attributes of the tables must have an identical list of P4 actions. The IDs of the tables implemented with a selector will appear in P4Info as part of the ActionProfile message for the selector.

An ActionProfileMember entity update message has the following fields:

An action profile member may be inserted, modified or deleted as per the following semantics.

When reading, an ActionProfileMember message with action_profile_id and member_id equal to 0 will read all members of all ActionProfile and ActionSelector objects. A message with action_profile_id equal to the id of an existing ActionProfile or ActionSelector object, and a member_id equal to 0, will read all members of that specified object.

9.2.2. Action Profile Group Programming

Action profile groups are entries in an ActionSelector and are referenced by a uint32 identifier that is bound to a set of action profile members already programmed in the selector. The action profile members in a group must be bound to actions of the same type.

Within a single ActionSelector object, the uint32 values used to identify its members are in a separate ‘scope’ from the uint32 values used to identify its groups. That is, the id 5 can simultaneously be used within a single ActionSelector object to identify a member 5, and a group 5.

There is not a separate scope within each group for member ids. For example, if at the same time both groups 5 and 10 contain member 6, the action associated with member 6 is the action for all groups containing member 6. Modifying the action associated with member 6 updates the behavior of all groups containing it.

An ActionProfileGroup entity update message has the following fields:

An action profile group may be inserted, modified or deleted as per the following semantics.

When setting the group membership with INSERT or MODIFY, the members repeated field must not include duplicates, i.e. members with the same member_id. The weight field is used instead to logically “repeat” the member inside the group.

It is explicitly allowed for the same member to be present in multiple groups at the same time. If, as a result, an implementation “stores” the action id and parameters in the target in multiple locations, the server must update all of those locations when a request to modify such a member is made.

When reading, an ActionProfileGroup message with action_profile_id and group_id equal to 0 will read all groups of all ActionSelector objects. A message with action_profile_id equal to the id of an existing ActionSelector object, and a group_id equal to 0, will read all groups of that one specified object.

9.2.2.1. Rules on Setting max_size

The valid values for max_size depend on the static max_group_size included in the P4Info message:

9.2.3. One Shot Action Selector Programming

P4Runtime supports syntactic sugar to program a table, which is implemented with an action selector, in one shot. One shot means that a table entry, an action profile group, and a set of action profile members can be programmed with a single update message. Using one shots has the advantage that the controller does not need to keep track of group ids and member ids.

One shots are programmed by choosing the ActionProfileActionSet message as the TableAction. The ActionProfileActionSet message consists of a set of ActionProfileAction messages. This set should have cardinality no greater than max_group_size (if max_group_size is nonzero) if selector_size_semantics is sum_of_members, or else the server must return INVALID_ARGUMENT. Each ActionProfileAction message has the following fields:

Semantically, one shots are equivalent to programming the table entry, group, and members individually; with the necessary group id and member ids bound to unused ids. An implementation is free to implement one shots in other ways, as long as the implementation matches the above semantics.

To preserve read-write symmetry, an implementation must answer ReadRequests with the original one shot messages. It may not return a desugared version of the one shot message.

For example, consider the action selector table defined here. This table could be programmed with the following one shot update:

table_entry {
  table_id: 0x0212ab34
  match { /* lpm match */ }
  action {
    action_profile_action_set {
      action_profile_actions {
        action { /* set nexthop 1 */ }
        weight: 1
        watch_port: "\x01"
      }
      action_profile_actions {
        action { /* set nexthop 2 */ }
        weight: 2
        watch_port: "\x02"
      }
      action_profile_actions {
        action { /* set nexthop 3 */ }
        weight: 3
        watch_port: "\x03"
      }
    }
  }
}

Which would be equivalent to the following updates, where GROUP_ID, MEMBER_ID_1, MEMBER_ID_2, and MEMBER_ID_3 are unused ids:

action_profile_member {
  action_profile_id: 0x11ab12cd
  member_id: MEMBER_ID_1
  action { /* set nexthop 1 */ }
}
action_profile_member {
  action_profile_id: 0x11ab12cd
  member_id: MEMBER_ID_2
  action { /* set nexthop 2 */ }
}
action_profile_member {
  action_profile_id: 0x11ab12cd
  member_id: MEMBER_ID_3
  action {  /* set nexthop 3 */ }
}
action_profile_group {
  action_profile_id: 0x11ab12cd
  group_id: GROUP_ID
  members {
    member_id: MEMBER_ID_1
    weight: 1
    watch_port: "\x01"
  }
  members {
    member_id: MEMBER_ID_2
    weight: 2
    watch_port: "\x02"
  }
  members {
    member_id: MEMBER_ID_3
    weight: 3
    watch_port: "\x03"
  }
}
table_entry {
  table_id: 0x0212ab34
  match { /* lpm match */ }
  action { action_profile_group_id: GROUP_ID }
}

Note that when using the above method (members and groups), the client also needs to use multiple messages to ensure correct ordering between the dependent updates. Members need to be inserted first, then the group needs to be created, and finally the match entry can be inserted. Therefore, 3 distinct WriteRequest batches are required.

It is possible to include several ActionProfileAction messages with the same exact action specification in one ActionProfileActionSet message. However, the P4Runtime client is encouraged not to do so, as the same can be achieved by using the weight field. Note that to preserve read-write symmetry, the server must not coalesce multiple ActionProfileAction messages with the same action specification into one. Additionally, each action specification would count as a separate member for the purposes of e.g. the sum_of_members group size calculation for Action Selectors.

All the tables associated with an action selector may either be programmed exclusively with one shots, or exclusively with ActionProfileMember and ActionProfileGroup messages. Programming some entries with one shots, and other entries with ActionProfileMember and ActionProfileGroup messages is not allowed, and the server must return the error code INVALID_ARGUMENT in that case.

A P4Runtime server must support the one shot style of programming tables with an action selector implementation. Support for the ActionProfileMember and ActionProfileGroup style is optional. If ActionProfileMember and ActionProfileGroup are not supported by a server, it must return an UNIMPLEMENTED error for every ActionProfileMember or ActionProfileGroup message that it receives.

9.2.4. Constraints on action selector programming

The PSA specification states that the following features are optional in action selector implementations [27]:

  1. Support for non-empty groups where in the same group, different members are bound to different actions.
  2. Predictable data plane behavior when a matched table entry points to an empty group.

For 1., if a client tries to INSERT or MODIFY a group with members bound to different actions, the server should return UNIMPLEMENTED if not supported by the target. This applies to the one shot style of programming as well. We recommend that control plane implementations take into account this possible limitation and be designed so as not to rely on this feature for the sake of portability. A target with this restriction is also not expected to support modifying the action function of a member which is part of one or more groups and should return UNIMPLEMENTED (modifying the action parameter values must be supported, however).

PSA 1.1 introduces the psa_empty_group_action table property in order to enable the P4 programmer to specify the action to perform on the packet when the matched table entry points to an empty action selector group. This action may be different from the default action, which is performed in case of table miss. psa_empty_group_action is one possible way to achieve property 2. in the list above. We recommend that all P4Runtime implementations support this property. Note that this version of P4Runtime does not provide any mechanism to modify the value of psa_empty_group_action at runtime, so the value will be constant and will either be provided by the P4 programmer or will default to NoAction. Even when psa_empty_group_action is not implemented by the target, P4Runtime does not require the server to return an error code when the client performs an operation which results in an empty group, despite the possibility for undeterministic or target-specific behavior. It is likely that future PSA versions will make the implementation of psa_empty_group_action mandatory and that future P4Runtime versions will provide a mechanism to change the property value dynamically. Note that the discussion above also applies to the one shot style of programming.

The PSA specification includes a discussion on how to implement psa_empty_group_action in software in the P4Runtime server [30].

If a P4Runtime implementation does support psa_empty_group_action, that action should be executed when an action selector group is selected that uses the watch port feature, and every member of the group has a watch port that is down.

If a P4Runtime implementation does not support psa_empty_group_action, and does support the watch port feature, we recommend that its developers document its behavior when a group effectively becomes empty because the watch ports of all members of a group are down.

9.3. CounterEntry & DirectCounterEntry

PSA defines Counters as a mechanism for keeping statistics of bytes and packets. Statistics may be updated as a result of an action associated with a table entry, or a direct invocation such as from a P4 control. The CounterData P4Runtime message can be used for all three types of PSA counters PACKETS, BYTES and PACKETS_AND_BYTES and consists of the following fields:

message CounterData {
  int64 byte_count = 1;
  int64 packet_count = 2;
}

P4Runtime does not distinguish between the different PSA counter types, and allows for simultaneous updates of byte_count and packet_count fields, which is equivalent to specifying the counter type PACKETS_AND_BYTES. Counters may be defined as direct or indirect (indexed) instances.

9.3.1. DirectCounterEntry

A direct counter is a direct resource associated with a TableEntry (see Direct Resources). The counter_data field of the TableEntry message can be used to initialize the counter value at the same time as the table entry is inserted. Once the table entry has been created, the P4Runtime client may modify the associated direct counter entry using the DirectCounterEntry message. Once the table entry is deleted the associated direct counter entry can no longer be accessed.

message DirectCounterEntry {
  TableEntry table_entry = 1;
  CounterData data = 2;
}

A WriteRequest may only include an Update message of type MODIFY with a DirectCounterEntry, whose fields are specified by the client as follows:

Specifying DirectCounterEntry in an Update message of type INSERT or DELETE is not allowed, and the server must return the error code INVALID_ARGUMENT in that case.

A client may use ReadRequest in two ways to read the contents of a DirectCounter:

9.3.2. CounterEntry

An indirect or indexed counter is not associated with a specific TableEntry and may be updated independently of any action. It may be read or written using the P4Runtime CounterEntry message whose fields are defined as follows:

message CounterEntry {
  uint32 counter_id = 1;
  Index index = 2;
  CounterData data = 3;
}

The CounterEntry can only be used in a WriteRequest with the MODIFY update type. The P4Runtime server must return an INVALID_ARGUMENT error code for update types INSERT and DELETE. By default all the counter entries in the array have default value 0.

A P4Runtime client may request to read the counter values of one or more indirect counter instances with a ReadRequest by including a CounterEntry entity for each of the instances, specifying the counter_id and index. Wildcard reads are also supported as follows:

9.4. MeterEntry & DirectMeterEntry

Meters are an advanced mechanism for keeping statistics, involving stateful “marking” and usually “throttling” of packets based on configured rates of traffic. The PSA metering function is based on the Two Rate Three Color Marker (trTCM) defined in RFC 2698 [3]. P4Runtime clients may additionally restrict meter usage on a table to RFC 2697's [2] Single Rate Three Color Marker (srTCM) or a simplified version of it that we call Single Rate Two Color Marker. The type of a table's meter is set by the MeterSpec.Type as described in the Meter & DirectMeter section.

The trTCM meters an arbitrary packet stream using two configured rates the Peak Information Rate (PIR) and Committed Information Rate (CIR), and their associated burst sizes and “marks” its packets as GREEN, YELLOW or RED based on the observed rate.

The srTCM meters an arbitrary packet stream using a single configured rate the Committed Information Rate (CIR) and its associated burst size as well as the Excess Burst Size (EBS) and “marks” its packets as GREEN, YELLOW or RED based on the observed rate.

The Single Rate Two Color Marker meters an arbitary packet stream using a single configured rate the Committed Information Rate (CIR) and its associated burst size and “marks” its packets as GREEN or RED based on the observed rate.

MeterEntry & DirectMeterEntry have an additional field counter_data that may hold per color counter data for targets that support it, and that must always be unset for targets that do not support it. If set in a request and unsupported by a target, an UNIMPLEMENTED error code should be returned. These counters provide a granular list of the counters applicable to the possible meter colors GREEN, YELLOW and RED. The counter associated with a specific color is incremented when a packet is “marked” with that color. The primary purpose of the color counters is for debugging purposes.

A meter may be configured as a direct or indirect instance, similar to a counter. The MeterConfig P4Runtime message represents meter configuration.

message MeterConfig {
  int64 cir = 1;  // Committed Information Rate
  int64 cburst = 2;  // Committed Burst Size
  int64 pir = 3;  // Peak Information Rate
  int64 pburst = 4;  // Peak Burst Size
  int64 eburst = 5; // Excess burst size (only used for srTCM)
}

A MeterConfig for a trTCM typed meter must only be accepted if eburst is unset. Otherwise, the server should return an INVALID_ARGUMENT error.

A MeterConfig for a srTCM typed meter must only be accepted if pir and pburst are equal to cir and cburst respectively. Otherwise, the server should return an INVALID_ARGUMENT error.

A MeterConfig for a Single Rate Two Color Marker typed meter must only be accepted if pir and pburst are equal to cir and cburst respectively and eburst is unset. Otherwise, the server should return an INVALID_ARGUMENT error.

Note: Any acceptable MeterConfig for a Single Rate Two Color Marker typed meter is also acceptable for either of the other two meter types and will mark packets identically in all three cases. Thus, changing the meter type from Single Rate Two Color Marker to Single Rate Three Color Marker or Two Rate Three Color Marker is a non-breaking change.

9.4.1. DirectMeterEntry

A direct meter is a direct resource associated with a TableEntry (see Direct resources). The meter_config field of the TableEntry message can be used to initialize the meter configuration at the same time as the table entry is inserted. Once the table entry has been created, the P4Runtime client may modify the associated direct meter entry using the DirectMeterEntry message. Once the table entry is deleted the associated direct meter entry can no longer be accessed.

message DirectMeterEntry {
  TableEntry table_entry = 1;
  MeterConfig config = 2;
  MeterCounterData counter_data = 3;
}

A WriteRequest may only include an Update message of type MODIFY with a DirectMeterEntry, whose fields are specified by the client as follows:

Specifying DirectMeterEntry in an Update message of type INSERT or DELETE is not allowed, and the server must return the error code INVALID_ARGUMENT in that case.

A client may use ReadRequest in two ways to read a DirectMeter config.

9.4.2. MeterEntry

An indirect or indexed meter is not associated with a specific TableEntry and may be executed independently of any action. Its configuration may be read or written using the P4Runtime MeterEntry message whose fields are defined as follows:

message MeterEntry {
  uint32 meter_id = 1;
  Index index = 2;
  MeterConfig config = 3;
  MeterCounterData counter_data = 4;
}

The MeterEntry can only be used in a WriteRequest with the MODIFY update type. The P4Runtime server must return an INVALID_ARGUMENT error code for update types INSERT and DELETE. By default all the meter entries in the array have a default configuration (GREEN for all packets).

A P4Runtime client may request to read the configuration of one or more indirect meter instances with a ReadRequest by including a MeterEntry entity for each of the instances, specifying the meter_id and index. Wildcard reads are also supported as follows:

9.4.3. MeterCounterData

The MeterCounterData P4Runtime message represents the per color counters associated with a meter entry. Whenever a meter is executed and returns a color, the corresponding color counter GREEN, YELLOW or RED is updated.

As seen above, these counters can be associated with a DirectMeterEntry or MeterEntry. Targets not capable of supporting these counters should return UNIMPLEMENTED if a MeterCounterData field was set in a read or write request.

message MeterCounterData {
  CounterData green = 1;
  CounterData yellow = 2;
  CounterData red = 3;
}

9.5. PacketReplicationEngineEntry

The PSA Packet Replication Engine (PRE) is an extern that is implicitly instantiated in all PSA programs. The PRE is responsible for implementing multicasting and cloning functionality in the data plane. P4Runtime defines an API to program the PRE with multicast groups and clone sessions to allow replication of data plane packets.

9.5.1. MulticastGroupEntry

Multicasting is achieved in PSA programs by setting the multicast_group ingress output metadata to a non-zero identifier. The number of replicas and their egress ports for the multicast group is programmed at runtime by the client using the MulticastGroupEntry API in P4Runtime. The following P4 program illustrates a possible data plane behavior of multicasting ARP packets in the ingress. Note that the data plane type of the multicast group metadata is 10 bits on the PSA device in this example.

control arp_multicast(inout H hdr, inout M smeta) {
  apply {
    if (hdr.ethernet.isValid() &&
        hdr.ethernet.eth_type == ETH_TYPE_ARP) {
      smeta.multicast_group = (MulticastGroup_t) 1;
    }
  }
}

At runtime, the client writes the following update in the target (shown in Protobuf text format).

type: INSERT
entity {
  packet_replication_engine_entry {
    multicast_group_entry {
      multicast_group_id: 1
      replicas { port: "\x05" instance: 1 }
      replicas { port: "\x0c" instance: 2 }
      replicas { port: "\x12" instance: 3 }
      replicas { port: "\x18" instance: 4 }
    }
  }
}

As a result of the above P4Runtime programming, the target device will create four replicas of an ARP packet. These replicas will appear in the egress pipeline as independent packets with egress port set to PSA device port numbers corresponding to SDN port numbers 5, 12, 18 and 24. For more discussion on the translation between SDN ports and PSA device ports, refer to the PSA Metadata Translation section.

The egress packets may be distinguished for further processing in the egress using the instance metadata. Note that a packet may not be both unicast and multicast; if the multicast group is set, it will override the unicast egress port. If the P4 multicast_group metadata is set to a value that is not programmed in the PRE, then the packet is dropped.

A multicast group may be inserted, modified or deleted as per the following semantics.

When reading a multicast group, only multicast_group_id is considered. All other fields in MulticastGroupEntry are ignored. To perform a wildcard Read on all configured multicast group entries, the multicast_group_id field must be set to 0, its default value.

9.5.1.1. Valid Values for multicast_group_id

The PSA specification states that the valid data plane values for multicast group ids (MulticastGroup_t) range from 1 (0 is a special value that indicates no multicast replication is to be performed for a packet) to the maximum value supported by the target [29]. This means that, in the absence of translation, the client must set the multicast_group_id field to a value in this range when inserting a multicast group. However, because P4Runtime reserves 0 as a special wildcard value which is used to read all the multicast groups configured in the target, the multicast_group_id field must never be set to 0 when performing a Write RPC, even when numerical translation is enabled for multicast group ids. In other words, it is not possible to map (using translation) a zero SDN multicast group id value to a non-zero data plane multicast group id value.

9.5.2. CloneSessionEntry

PSA supports cloning of packets in both the ingress and egress pipeline. Ingress cloning creates a mirror of the packet as seen in the beginning of the ingress pipeline, while egress cloning creates a mirror of the packet as seen at the end of the egress pipeline. A packet is cloned in the data plane by setting a clone_session_id identifier and a boolean flag clone in the packet metadata. The clone_session_id serves as a handle to the clone attributes, namely a set replicas of (egress port, instance) pairs to which cloned packets should be sent, a packet length, and class of service. These are programmed at runtime via the P4Runtime CloneSessionEntry API.

The following P4 program illustrates a possible data plane behavior of sending clones of low TTL packets to the CPU for monitoring. Note that the data plane type of the clone session metadata is 10 bits on the PSA device in this example. We assume that the clone_low_ttl control block is applied in the ingress pipeline to create an ingress-to-egress clone.

control clone_low_ttl(inout H hdr, inout M smeta) {
  apply {
    if (hdr.ipv4.isValid() &&
        hdr.ipv4.ttl <= LOW_TTL_THRESHOLD) {
      smeta.clone_session_id = 10w100;
      smeta.clone = true;
    }
  }
}

At runtime, the client writes the following update in the target (shown in Protobuf text format).

type: INSERT
entity {
  packet_replication_engine_entry {
    clone_session_entry {
      session_id: 100
      replicas { port: "\xff\xff\xff\xfd" instance: 1 } # to CPU
      class_of_service: 2
      packet_length_bytes: 4096
    }
  }
}

As a result of the above P4Runtime programming, the target device will create one replica of a low TTL packet from the ingress to the egress. Note that the clone session ID of the programmed PRE entry is identical to the value used in the data plane in this example (no numerical translation, which is the default for values of type CloneSessionId_t [29]). The clone will be treated for scheduling in the PRE with a class of service value of 2. If the packet is larger than 4096 bytes, it will be truncated to carry at most 4096 bytes.

The cloned replica will appear in the egress pipeline as an independent packet with egress port set to CPU (corresponding to SDN port 0xfffffffd; see Translation of Port Numbers). Note that the egress port (port field) must be an SDN port and must refer to a singleton port.

If the clone_session_id data plane metadata is set to a value that is not programmed in the PRE, then no clones are created.

A clone session may be inserted, modified or deleted as per the following semantics:

When reading a clone session, only session_id is considered. All other fields in CloneSessionEntry are ignored. To perform a wildcard Read on all configured clone session entries, the session_id field must be set to 0, its default value. The session_id field can never be equal to 0 in a Write RPC. If it does, the server must return an INVALID_ARGUMENT error.

9.5.2.1. Valid Values for session_id

The PSA specification states that the valid data plane values for clone session ids (CloneSessionId_t) range from 0 to the maximum value supported by the target [29]. Note that unlike for multicast group ids, 0 is a valid data plane value for clone session ids. However, just like for multicast group ids, P4Runtime reserves 0 as a special wildcard value which is used to read all the clone sessions configured in the target. This means that 0 can never be used as a session_id value when inserting a clone session, whether or not numeric translation is enabled for clone session ids. If translation is not enabled, we effectively “lose” one clone session, assuming the target supports 0 as mandated by the PSA specification. If this is an issue (e.g. because the target supports a very limited number of clone sessions), one can enable translation on CloneSessionId_t and map any non-zero SDN session id to the data plane clone session with id 0, then insert a clone session with the chosen SDN session id.

9.6. ValueSetEntry

Parser Value Set is a construct in P4 that is used to support programmability of parser state transitions. A transition select statement in P4 can use a parser Value Set to define a runtime programmable state transition as shown in the example below. A runtime programmable set of TRILL Ethertypes is used to transition the parser state machine to the parse_trill_types state.

state parse_l2 {
  @id(1) value_set<ETH_TYPE_BITWIDTH>(MAX_TRILL_TYPES) trill_types;
  extract(hdr.ethernet);
  select (hdr.ethernet.eth_type) {
    ETH_TYPE_IPV4: parse_ipv4;
    ETH_TYPE_IPV6: parse_ipv6;
    trill_types:   parse_trill_types;
    _:             reject;
  }
}

The corresponding entry in the P4Info for this Value Set is:

value_sets {
  preamble {
    id: 0x03000001
    name: "trill_types"
  }
  match {
    id: 1
    bitwidth: <ETH_TYPE_BITWIDTH>
    match_type: EXACT
  }
  size: <MAX_TRILL_TYPES>
}

At runtime, the client writes the following update in the target (shown in Protobuf text format).

type: MODIFY
entity {
  value_set_entry {
    value_set_id: 0x03000001
    members {
      match {
        field_id: 1
        exact { value: 0x22f3 }
      }
      match {
        field_id: 1
        exact { value: 0x893b }
      }
    }
  }
}

As a result of the above P4Runtime programming, all packets with EtherType values of 0x22f3 and 0x893b will be parsed as per the state machine starting at the parse_trill_types state.

A ValueSetEntry entity update message has the following fields:

A ValueSetEntry may only be modified. If the update type is INSERT or DELETE, the server must return an INVALID_ARGUMENT error. If the update type is MODIFY, the server writes the members given in the repeated field to the Value Set entry indexed by the given value_set_id. The maximum number of matches must not exceed the maximum size given by the size field in P4Info of the Value Set, otherwise the server must return a RESOURCE_EXHAUSTED error. To empty a Value Set (i.e. restore it to its initial state), the P4Runtime client can perform a MODIFY update with an empty members repeated field.

To facilitate read-write symmetry, the server must return an ALREADY_EXISTS error in case of duplicate members. Unlike for match tables, a priority value is not required for ternary, range and optional matches: overlapping entries do not need to be ordered and the parse state transition is determined by whether or not the packet matches at least one entry in the set.

See Appendix A.3 for a more complex Value Set example.

9.7. RegisterEntry

The PSA Register extern is a stateful memory array that can be read and written during packet forwarding. The RegisterEntry P4Runtime entity is used by the client to read and write the contents of a Register instance as part of control plane operations.

RegisterEntry has the following fields:

9.8. DigestEntry

A digest is one mechanism to send a message from the data plane to the control plane. It is traditionally used for MAC address learning: when a packet with an unknown source MAC address is received by the device, the control plane is notified and can populate the L2 forwarding tables accordingly.

The DigestEntry P4Runtime entity is used to configure how the device must generate digest messages. The DigestEntry Protobuf message is not used to carry digest data, which is done on the StreamChannel bidirectional stream using the DigestList (digest data sent by the target to the client) and DigestListAck (digest data acknowledgments sent by the client to the target) Protobuf messages.

In this section, we refer to the data learned by a single data plane call to Digest<T>::pack as a “digest message” and we use “digest list” to designate the list of digest messages bundled by the P4Runtime service in a single DigestList stream message. Note that all the digest messages in a single digest list correspond to the same P4 Digest extern instance. We say that 2 digest messages are “duplicate” if the data emitted by the data plane is exactly the same as per P4 equality rules. We say that 2 digest messages are “distinct” if they are not duplicate.

DigestEntry has the following fields:

Here is the significance of the different Update types for DigestEntry:

A server should buffer digest messages until either:

At which point the server should, with best effort, generate a DigestList stream message with the buffer contents and send it to the primary client. All the messages in a digest list must be distinct, which means that duplicates must either be filtered-out directly by the device or in the P4Runtime server software.

To avoid sending duplicate digest messages across different DigestList messages, which could make the channel busy, we define an acknowledgement mechanism through which the primary client indicates that it has received the digest list and acted on it. The server must keep a “cache” containing the set of all digest messages that have been sent, but not acknowledged yet by the primary client, up-to