P4Runtime Specification
version 1.0.0-rc3
The P4.org API Working Group

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 or program defined 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 [14] 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/v1.0.0-rc3/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 1.1.0 [1].

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

Refers to the process through which P4Runtime ensures that at any given time, there is a single master (i.e. a client with write access) for a given role. Also referred to as “master-slave arbitration”.
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).
Class of Service.
Synonymous with target, although device usually connotes a physical appliance or other hardware, whereas target can signify hardware or software.
An instantiated P4 program object such as a table or an extern (from PSA or any other architecture).
gRPC Remote Procedure Calls, an open-source client-server RPC framework. See [8].
High-Availability. Refers to a redundancy architecture.
The part of the P4Runtime server which implements the calls to the device or target native “SDK” or backend.
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 which comprises the Forwarding Pipeline Configuration. This is opaque, architecture- and target-specific binary data which can be loaded onto the device to change its “program.”
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.
Abbreviation for P4Runtime.
Protobuf (Protocol Buffers)
The wire serialization format for P4Runtime. Protobuf version 3 (proto3) is used to define the P4Runtime interface. See [15].
Portable Switch Architecture; a target architecture that describes common capabilities of network switch devices that process and forward packets across multiple interface ports.
Remote Procedure Call.
Round-trip time.
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 probrammable, 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.
The gRPC server which accepts P4Runtime requests on the device or target. It uses instrumentation to translate P4Runtime API calls into target-specific actions.
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 master-slave arbitration, among other things.
switch config
Refers to non-forwarding config (different from P4 forwarding 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.
The hardware or software entity which “executes” the P4 pipeline and hosts the P4Runtime Service; often used interchangeably with “device”.
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. A multi-master protocol allows more than one controller to participate, and a role-based arbitration scheme ensures only one controller has write access to each r/w 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 [12]. 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 a P4 Target supporting P4Runtime, as well as sample clients, may be available on the p4lang/PI GitHub repository [13]. 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) and installing the associated P4Info metadata. Furthermore, the controller can query the target for the ForwardingPipelineConfig to retrieve the device config and the P4Info.


Figure 1. P4Runtime Reference Architecture.

3.1. 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 P4 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.

Presumably, a controller can access a library of P4 “packages” consisting 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.

3.2. 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. Furthermore, it is not necessary to have a P4 source program to begin with, since the controller does not use it. From the standpoint of controller (not pipeline) implementers, the P4 source code is just helpful documentation. Some parties may wish to keep their P4 source code private. The minimum requirement is a P4Info file which can be loaded by a controller in order to render the correct P4Runtime API. As long as the target supports the operations implied by the P4Info file, the underlying implementation is moot.

This leads to the notion that a good P4Info file should be complete and self-sufficient in terms of documentation, specifically the metadata in the PkgInfo message as well as the embedded doc messages. Nevertheless, a P4 program which describes the pipeline and produces the device config via a compiler is ideally available. The contents of the P4Info file will be described in later sections.

3.3. Alternative Workflows

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

3.3.1. 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.

3.3.2. 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 “contract” specified in the P4Info.

3.3.3. 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.

3.3.4. 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.

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.


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.


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.


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.


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 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 multi-master arbitration scheme supports it.


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

5. Master-Slave Arbitration and Controller Replication

The P4Runtime interface allows multiple 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 entitites simultaneously. The control plane can be partitioned into multiple roles and each role will have a set of controllers, one of which is the master and the rest are slaves. 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 slave controllers, which take over controlling the devices in case the master controller goes offline.

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

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 0 (default). If using a default role, all RPCs from the controller (e.g. Write) must set the role_id to 0.

5.2. Role Config

The role.config field in the MasterArbitrationUpdate message sent by the controller describes the role configuration, i.e. which operations, P4 entities, behaviors, etc. are in the scope of a given role. An unset role.config implies “full pipeline access” as well (similar to the default role explained above). In order to support different role definition schemes, role.config is defined as an Any Protobuf message [19]. 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.

5.3. Rules for Handling MasterArbitrationUpdate Messages Received from Controllers

  1. If the MasterArbitrationUpdate message is received for the first time (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 FAILED_PRECONDITION error.

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

    3. If the max number of clients for the given (device_id, role_id) exceeds the supported limit, the P4Runtime server shall terminate the stream by returning a RESOURCE_EXHAUSTED error.

    4. Otherwise, the controller is added to list of connected controllers for the given (device_id, role_id) and the controller is notified by sending a StreamMessageResponse message back to it, as explained earlier.

  2. If the MasterArbitrationUpdate message is received from an already connected 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. Otherwise, if the role.id matches the current role_id assigned to this stream:

      1. If the election_id also matches the one assigned to this stream, the server will accept the (new) role.config only if this controller is the current master. If the controller is not a master, the operation is a no-op.

      2. If the election_id is already assigned to another controller stream for the same (device_id, role_id), the P4Runtime server shall terminate the stream by returning an INVALID_ARGUMENT error.

      3. Otherwise, the P4Runtime server updates the election_id for this controller. If this makes the client the new master, the server will also accept the given role.config and follow the “mastership change rules” described in the following section.

    3. Otherwise (i.e. role.id is different from current role_id assigned to this stream), the P4Runtime server moves the controller to the new role. This controller will then be treated as a new controller for the new (device_id, role_id). The server accepts the given role.config only if the client becomes master, in which case the server also follows the “mastership change rules” described in the following section.

5.4. Mastership Change

“Mastership change” refers to either one of these cases:

  1. A new MasterArbitrationUpdate is received from an already connected controller for a given (device_id, role_id), which changes the controller mastership status (the controller becomes master or slave).

  2. A streaming channel for a given master controller breaks, forcing a new master to be elected.

In case of a mastership change, P4Runtime server shall send the election_id of the master to all the connected controllers for a given (device_id, role_id). The StreamMessageResponse sent back to all the connected controllers has a MasterArbitrationUpdate message populated with device_id, role_id, and election_id of the master, as well as an OK status for the master and non-OK status (with ALREADY_EXISTS error code) for slaves.

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 the Documentation 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 for the P4 object, probably shorter than its 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. The P4 programmer may also override the
  // default alias for any P4 object (TBD). When resolving a P4 object id, an
  // application should be able to indiscriminately use the name or the alias.
  string alias = 3;
  repeated string annotations = 4;
  // Documentation of the entity
  Documentation doc = 5;

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.

Note, the known P4 compilers as of this writing don't emit PkgInfo as part of the P4Info output. Until compiler support is added, a utility to post-process and insert PkgInfo can be used [11].

// 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;
  // 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;
}  // A more verbose description of something.
   // Multiline is accepted. Markup format (if any) is TBD.
  string description = 2;

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 is 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 resultings 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.

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 actA { } 0x0112ab34

Table 3. Example of statically-assigned P4Info object IDs

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:

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.

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 PSA extern is attached.

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 packet-out the other. 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.

header PacketOut_t {
  bit<9> egress_port; /* suggested port where the packet
                         should be sent */
  bit<8> queue_id;    /* suggested queue ID */

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 [25].

The ValueSet message defines the following fields:

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 architetcure. 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 FowardingPipelineConfig 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. Set / Unset Protobuf Field

In Protobuf version 3 (proto3), the default value for a message field is “unset” [4]. An application, such as the P4Runtime client or server, is able to distinguish between an unset field and a field set to its default value. We use this distinction quite a lot 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 {
  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;
      // 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;
    • 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. 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 the doc tries to offer canonical representations for various data types, but this principle should be thought of where it falls short. Ensuring this will allow a 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.

8.3. Zero as Reserved Value

p4runtime.proto uses proto3 syntax, and so it does not allow not specifying a scalar data type, such as a uint32. Therefore, we usually reserve value 0 for those 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 switch, such as in a WriteRequest or a SetForwardingPipelineConfigRequest.

8.4. 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:

The shortest possible binary string for an integer value with high_bit as the most-significant non-zero bit in the value is computed as:

if (value != 0)
  encoded_string_size = (high_bit + 8) / 8;
  encoded_string_size = 1;

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

Upon receiving a binary string, the P4Runtime server does not impose any restrictions on the length of the string itself. Instead, the server verifies that the highest-order non-zero bit position in the string is within the bounds established by the P4Info bitwidth. If p4_bitwidth is the P4Info- specified bitwidth for a P4 object value, then the server computes the number of binary string bytes required to represent the maximum integer of p4_bitwidth bits as:

binary_string_bytes = (p4_bitwidth + 7) / 8;

The server's decision to accept or reject the binary string depends on the string's content and its actual length in the P4Runtime request, as follows:

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) \x00\x9d no
int<12> -739 (-0x2e3) \x0d\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> \xff\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 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 INVALID_ARGUMENT error code otherwise.

8.5. Representation of Arbitrary P4 Types

8.5.1. Problem Statement

The P416 language includes more complex types than just binary strings [3]. 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.

value_set<tuple<bit<16>, bit<8> > >(16) pvs_complex;
state parse_ipv4 {
   transition select({ hdr.ipv4.version, hdr.ipv4.protocol }){
       pvs_complex: parse_inner;
       default: accept;
// ...
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.

8.5.2. 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, everywhere where the P4Runtime client can provide a value for a P416 expression. P4DataTypeSpec describes the compile-time type of the expression 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.

8.5.3. P4Data in p4runtime.proto

P4Runtime uses the P4Data message to represent values with 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.

The order of members in P4StructLike, the order of bitstrings in P4Header, and the order of entries in P4HeaderStack and P4HeaderUnionStack must match the order in the corresponding p4info.proto type specification and hence the order in the corresponding P416 type declaration.

8.5.4. 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: # ...

8.5.5. 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) [5]. 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>). Integer values must be assigned to each member entry by the P4 programmer. P4TypeInfo includes the mapping between entry name and entry value. When providing serializable enum values through P4Data, one can either use the enum entry's name (enum human-readable string field) or its assigned value (enum_value bytestring field).

8.5.6. User-defined types

P416 enables programmers to introduce new types [9]. 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 dataplane 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>) and the type exposed to the control-plane will also be a fixed-width unsigned bitstring, with a potentially different bitwidth. It 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 bitwidth of the bitstring type exposed to the control-plane. It is recommended that the URI includes at least the P4 architecture name and the type name.

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:

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

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

type_info {
  new_types {
    key: "PortId_t"
    value {  # P4NewTypeSpec
      translated_type {  # P4NewTypeTranslation
        uri: "p4.org/psa/v1/PortId_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.

8.5.7. Trade-off for v1.0 Release

For the v1.0 release of 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-release implementations of P4Runtime. Similarly it has been decided to keep using bytes to provide action parameter 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 and p4.config.v1.Action.Param.

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 a 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).

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 = {

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.


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

trailing_zeros = countTrailingZeros(match.lpm().value())
assert(trailing_zeros >= field_bits - pLen)
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)

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

assert(low <= high)

assert(low != min_field_value && high != max_field_value)

9.1.2. Action Specification

The TableEntry action field must be set for every INSERT and MODIFY update, except when resetting the default entry. Based on the implementation property value of the P4 table, the oneof in the TableAction message will either be:

If the action field is not set (and if is_default_action is false) or 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, which is used for direct tables only, 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 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. 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

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

9.1.5. 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 and 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).

9.1.6. 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 must generate a stream notification - using the IdleTimeoutNotification message - to the master 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 accomodate 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.

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. 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.

               /*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 = {
  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 uint32 identifier that uniquely identifies a member or group programmed in the action selector as.

If a table entry in an indirect table with 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 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 profle member may be inserted, modified or deleted as per the following semantics.

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.

An ActionProfileGroup entity update message has the following fields:

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

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, which in turn have 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: 1
  match { /* lpm match */ }
  action {
    action_profile_action_set {
      action_profile_actions {
        action { /* set nexthop 1 */ }
        weight: 1
        watch: 1
      action_profile_actions {
        action { /* set nexthop 2 */ }
        weight: 2
        watch: 2
      action_profile_actions {
        action { /* set nexthop 3 */ }
        weight: 3
        watch: 3

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

table_entry {
  table_id: 1
  match { /* lpm match */ }
  action { action_profile_group_id: GROUP_ID }
action_profile_group {
  action_profile_id: 1
  group_id: GROUP_ID
  members {
    member_id: MEMBER_ID_1
    weight: 1
    watch: 1
  members {
    member_id: MEMBER_ID_2
    weight: 2
    watch: 2
  members {
    member_id: MEMBER_ID_3
    weight: 3
    watch: 3
action_profile_member {
  action_profile_id: 1
  member_id: MEMBER_ID_1
  action { /* set nexthop 1 */ }
action_profile_member {
  action_profile_id: 1
  member_id: MEMBER_ID_2
  action { /* set nexthop 2 */ }
action_profile_member {
  action_profile_id: 1
  member_id: MEMBER_ID_3
  action {  /* set nexthop 3 */ }

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.

9.3. 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 to be 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 [2]. 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.

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

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;

A WriteRequest may only include an Update message of type MODIFY with a DirectMeterEntry, whose fields are to be 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;

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:

9.5. 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 dataplane. P4Runtime defines an API to program the PRE with multicast groups and clone sessions to allow replication of dataplane 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 dataplane behavior of multicasting ARP packets in the ingress. Note that the dataplane 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 {
      muticast_group_id : 1
      replicas { egress_port : 5 instance: 1 }
      replicas { egress_port : 12 instance: 2 }
      replicas { egress_port : 18 instance: 3 }
      replicas { egress_port : 24 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 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.

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 dataplane 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 dataplane behavior of sending clones of low TTL packets to the CPU for monitoring. Note that the dataplane 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 and 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 { egress_port : 0xFFFFFFFD 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 dataplane. 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 must be a 32-bit SDN port number and must refer to a singleton port.

Furthermore, even though the Protobuf representation for clone session entry allows multiple clones to be specified (by the repeated replicas message field), PSA version 1.0 allows creating only 1 clone of a packet in the ingress and egress. Therefore, a target may reject a clone session entry update that carries more than one replica. Cloning does not impact the original packet. If the clone_session_id 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:

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 ethtypes is used to transition the parser state machine to the parse_trill_types state.

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

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

type: INSERT
entity {
  value_set_entry {
    value_set_id : 1
    match { exact { value: 0x22F3 } } }
    match { 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 be inserted, modified or deleted as per the following semantics:

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 must generate a DigestList stream message with the buffer contents and send it to the master 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 master 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 master client, up-to ack_timeout_ns in the past. The server must delete all cache entries for a given digest list when they are at least ack_timeout_ns old or when a matching DigestListAck message (i.e. with the same digest_id and list_id fields as the DigestList message) is received.

The acknowledgement mechanism described above is not used to implement some sort of reliable transport for digest messages. The loss of digest messages or acknowledgement messages is considered non-critical. The P4Runtime server may drop digest messages if they are generated from the data plane faster than the server software, the channel or the client can handle. P4Runtime does not impose a limit on the number of in-flight, unacknowledged DigestList messages.

When max_timeout_ns is set to 0 and / or max_list_size is set to 1, the server must generate a DigestList message for every digest message generated by the data plane which is not already in the cache. If ack_timeout_ns is set to 0, the cache must always be an empty set. If max_list_size is set to 0, there is no limit on the maximum size of digest lists: the server can use any non-zero value as long as it honors the max_timeout_ns configuration parameter.

The P4Runtime server may empty the digest message cache in case of a client mastership change.

Here is some pseudo-code implementing the handling of digest messages in the P4Runtime server:

DigestStream stream;
DigestCache cache;
DigestBuffer buffers;

// sends digest list when it is ready
send_buffer(Id digest_id) {
  buffer = buffers[digest_id];
  cache.merge(buffer);  // updates cache with new digest list

// callback which handles data plane digest messages from device
handle_dataplane_digest(Digest msg) {
  digest_id = msg.digest_id();
  buffer = buffers[digest_id];
  if (msg in cache OR msg in buffer) return;
  if (buffer.length() < max_list_size(digest_id)) return;

// callback which handles ack messages received on the stream
handle_stream_ack(DigestListAck ack) {
  // clear all cache entries matching the tuple (digest_id, list_id)
  cache.erase( (ack.digest_id(), ack.list_id() )

// loop to enforce timeouts
while (true) {
  now = now();
  // check for buffers that need to be sent
  for ((digest_id, buffer) in buffers) {
    if (now - buffer.first_enq_time() >= max_timeout_ns(digest_id))
  // check for expired entries in cache
  for ((digest_id, list_id, sent_time) in cache) {
    if (now - sent_time >= ack_timeout_ns(digest_id))
      cache.erase( (digest_id, list_id) );

9.9. ExternEntry

This is used to support a P4 extern entity that is not part of PSA. It is defined as:

message ExternEntry {
  uint32 extern_type_id = 1;
  uint32 extern_id = 2;
  google.protobuf.Any entry = 3;

Each ExternEntry entity maps to an Extern message in the P4Info and an ExternInstance message within that message. The extern_type_id field must be equal to the one in ExternEntry. The extern_id field must be equal to the ID included in the preamble of the corresponding ExternInstance message.

entry itself is embedded as an Any Protobuf message [19] to keep the protocol extensible. It includes the extern-specific parameters required by the P4Runtime server to perform the read or write operation. The underlying Protobuf message should be defined in a separate architecture-specific Protobuf file. See section on Extending P4Runtime for non-PSA Architectures for more information.

10. Error Reporting Messages

P4Runtime is based on gRPC and all RPCs return a status to indicate success or failure. gRPC supports multiple language bindings; we use C++ binding below to explain how error reporting works in the failure case.

gRPC uses grpc::Status [20] to represent the status returned by an RPC. It has 3 attributes:

StatusCode code_;
grpc::string error_message_;
grpc::string binary_error_details_;

The code_ represents a canonical error [22] and describes the overall RPC status. The error_message_ is a developer-facing error message, which should be in English. The binary_error_details_ carries a serialized google.rpc.Status message [17] message, which has 3 fields:

int32 code = 1;  // see code.proto
string message = 2;
repeated google.protobuf.Any details = 3;

The code and message fields must be the same as code_ and error_message_ fields from grpc::Status above. The details field is a list that consists of p4.Error messages that carry error details for individual elements inside batch-request RPCs (e.g. Write and Read). The p4.Error also enables different target vendors to additionally express their own error codes in their chosen error-space.

Figure 7 illustrates how these messages fit together.


Figure 7. P4Runtime Error Report Mesage Format

gRPC provides utility functions ExtractErrorDetails() and SetErrorDetails() [21] to easily convert between grpc::Status and google.rpc.Status.

Please see sections on individual P4Runtime RPCs for details on how grpc::Status is populated for reporting errors.

11. Write RPC

The Write RPC updates one or more P4 entities on the target. The request is defined as:

message WriteRequest {
  uint64 device_id = 1;
  uint64 role_id = 2;
  Uint128 election_id = 3;
  repeated Update updates = 4;
  enum Atomicity {
  Atomicity atomicity = 5;

The device_id uniquely identifies the target P4 device. The role_id and election_id define the client role and election-id as described in the Master-Slave Arbitration and Controller Replication section.

The updates field is a list of P4 entity updates to be applied. Each update is defined as:

message Update {
  enum Type {
    INSERT = 1;
    MODIFY = 2;
    DELETE = 3;
  Type type = 1;
  Entity entity = 2;

This is modeled as performing an update operation on the given entity against its entity container. The entity container is either a logical table (e.g. CounterEntry) or an actual table (e.g. TableEntry) in the P4 data plane. Each entity in the container is uniquely identified by its key. Please refer to the P4 Entity Messages section for details on what parts of the entity specification make up the key for each P4 entity.

An update can be one of the following types:

The Write RPC is idempotent, i.e. multiple invocations of the same RPC do not have any side effects. The end result (modified end state on P4Runtime server and P4 device) is always the same as the result of the initial invocation, even if the response differs.

11.1. Batching and Ordering of Updates

P4Runtime supports batching of Write operations. The list of updates in a WriteRequest is referred to as a batch. A batch can consist of arbitrary updates on an arbitrary set of P4 entities. It is not restricted to a particular entity or table (in the case of TableEntry entities).

The P4Runtime server may arbitrarily reorder message within a batch to maximize performance, and clients should not depend on a specific processing order (e.g. FIFO or inferring implicit dependencies within a batch). In particular, P4 entities (e.g. table entries) may be inserted in the data plane in an order different than what is received in the WriteRequest.

The Write RPC demarcates the batch boundary, and can be used to ensure ordering between dependent updates. When the Write RPC returns, it is required that all operations in the batch have been committed to hardware (P4 data plane). If two updates from the client depend on each other (e.g. inserting an ActionProfileMember followed by pointing a TableEntry to it), they should be separated across two batches (and therefore two Write RPCs). In other words, the client must wait until the dependent Write RPC is acknowledged before invoking a Write RPC that depends on it.

P4Runtime is based on gRPC which provides a concurrent server design. A target implementation may support concurrent execution of a gi