DOCA Comch – New

This guide provides instructions on building and developing applications that require communication channels between the x86 host and the BlueField Arm cores.

DOCA Comch provides a communication channel between client applications on the host and servers on the BlueField Arm.

Benefits of using DOCA Comch:

• Security – the communication channel is isolated from the network

• Network independent – the state of the communication channel does not depend on the state and configuration of the network

• Ease of use

DOCA Comch provides two different data path APIs:

• Basic DOCA Comch send/receive for control messages

• High bandwidth, low latency, zero-copy, multi-producer, multi-consumer API

The following table summarizes the differences between the two data path APIs:

This library follows the architecture of a DOCA Core Context, it is recommended to read the following sections before:

• DOCA Core Execution Model

• DOCA Core Device

• DOCA Core Memory Subsystem (fast path only)

DOCA Comch based applications can run either on the host machine or on the NVIDIA BlueField Arm.

Sending messages between the host and BlueField Arm can only be run with a BlueField configured with a mode as described in NVIDIA BlueField Modes of Operation.

For basic DOCA Comch send and receive, the following configuration is required:

• doca_comch_server context must run on the BlueField Arm cores

• doca_comch_client context must run on the host machine

DOCA Comch is comprised of four DOCA Core Contexts. All DOCA Comch contexts leverage DOCA Core architecture to expose asynchronous tasks/events that are offloaded to hardware.

A doca_comch_server context runs on the BlueField Arm and listens for incoming connections from the host side. Such host side connections are initiated by a doca_comch_client context.

Servers can receive connections from multiple clients in parallel, however, a client can only connect with one server. An established 1-to-1 connection between a client and a server is represented by a doca_comch_connection.

Once an established connection exists between a client and a server, the doca_comch_producer and doca_comch_consumer contexts can be used to run fast path channels.

The following diagram provides examples of the contexts use:

Objects

Security Considerations

• DOCA Comch guarantees:

  • The client is connected to the server by providing the exact server name on the client side

  • Only clients on the PF/VF/SF represented by the doca_dev_rep provided upon server creation can connect to the server

  • The connection requests and data path are isolated from the network

• DOCA Comch does not provide security at the application level:

  • It is up to the user to implement application-level security and verify the identity of the client application

  • A server handles applications from a single PF/VF/SF. If a server application detects a compromised client application, the server app should consider all clients (from that PF/VF/SF) compromised.

Initialization Flow

doca_comch_server Initialization Flow

1. A doca_comch_server is created on a specific doca_dev and a specific doca_dev_rep.

2. A doca_comch_server must have a unique name per doca_dev/doca_dev_rep (i.e., two servers on the same doca_dev and doca_dev_rep cannot have the same name).

3. Once doca_ctx_start() is called, the doca_comch_server can start receiving new connection requests.

4. For the doca_comch_server to process new connection requests and messages, the user must periodically call doca_pe_progress() .

5. When a new connection request arrives, doca_comch_server calls the connection request handler function and passes a doca_comch_connection object.

The server can now send and receive messages on the connection represented by doca_comch_connection .

doca_comch_client Initialization Flow

1. A doca_comch_client is created on a specific doca_dev is targeting a specific doca_comch_server.

2. Once doca_ctx_start() is called, doca_comch_client asynchronously tries to connect to the server.

3. To establish the connection and receive messages, the user must periodically call doca_pe_progress().

4. When the connection is established, doca_comch_client calls the state change callback indicating state change to “RUNNING”.

The client can now send a receive messages.

The following diagram describes the initialization of a basic client/server connection on DOCA Comch:

doca_comch_consumer Initialization Flow

1. A doca_comch_consumer is created on a specific doca_comch_connection.

2. doca_pe_progress()must be periodically called on the client/server PE to allow registration of the consumer.

3. After the doca_comch_consumer moves to “RUNNING” state:

   1. doca_comch_consumer notifies its existence to the peer (invoking a new consumer event).

   2. The application can start posting receive tasks.

   3. A doca_comch_producer on the peer side can start sending messages to that consumer.

The initialization flow is described in the following diagram:

Teardown Flow

The teardown flow must be executed in the following order, otherwise errors may occur.

Disconnecting Specific Connection

The proper disconnection process for a specific connection consists of the following steps:

1. Stop all consumers and producers linked to the connection.

2. Server/client:

   1. For server, a connection can be disconnected using doca_comch_server_disconnect() . If there are any active producers/consumers linked to the connection, the disconnect would fail. A disconnection notifies the client and initiates teardown on that side too.

   2. For client, since there is only one connection at any given time, the connection can be disconnected by calling doca_ctx_stop(). If there are any active producers/consumers, the command would fail. Stopping the client context notifies the server of the disconnection and causes a disconnection of the connection on it.

Tearing Down DOCA Comch

The proper teardown for a DOCA Comch context consists of the following:

1. Stop all consumers and producers linked to the context.

2. Call doca_ctx_stop(). If there are any active connections, they would all be disconnected. If there are any active consumers/producers, the command would fail. Disconnecting/stopping the context informs all active peers of the disconnection, and causes teardown (on clients) or disconnection (on server). Calling doca_ctx_stop() successfully moves the context to “stopping” state.

3. After moving to stopping state, doca_pe_progress() must be called until the context moves to idle state.

MsgQ (DPA Communication)

DOCA Comch MsgQ leverages the existing consumer/producer model to allow communication between host/BlueField and DPA.

Since communication between the host/BlueField and DPA is local, there is no need to create a server, client, or connection. Instead the user can create a MsgQ and use it to create producers and consumers directly.

When creating a consumer/producer using the MsgQ, it becomes possible to use them in the DPA application as well as the CPU application:

• The CPU application can utilize existing consumer/producer APIs for communication

• The DPA application has a different set of APIs that are usable within a DPA application

Communication Direction

Every instance of a MsgQ can only support a single communication direction as follows:

• Communication from host/BlueField to DPA
  

  • This direction may be specified using doca_comch_msgq_set_dpa_consumer

  • Consumers created from this MsgQ are referred to as DPA consumers, while producers are CPU producers

• Communication from DPA to host/BlueField
  

  • This direction may be specified using doca_comch_msgq_set_dpa_producer

  • Consumers created from this MsgQ are referred to as CPU consumers, while producers are DPA producers

To support bidirectional communication in an application, the user has to create 2 MsgQ instances, as shown in the above diagram.

To start using the library, users must go through a configuration phase as described in DOCA Core Context Configuration Phase.

This section describes how to configure and start the context to allow execution of tasks and retrieval of events.

Configurations

The context can be configured to match the application use case.

To find out if a certain configuration is supported, or what the min/max value for it is, refer to Device Support.

Mandatory Configurations

These configurations are mandatory and must be set by the application before attempting to start the context:

• For a basic send/receive client or server:

  • A send task callback

  • A receive event callback

  • A device with appropriate support must be provided on creation

  • A valid server name must be provided on creation (for clients this is the server to connect to)

  • A connection event callback (server only)

• For fast path producer or consumer:

  • A device with appropriate support must be provided on creation

  • An established client to server connection must be provided on creation

  • A doca_mmap with PCIe read/write permissions of where data should be received must be provided on creation (consumer only)

  • A post receive task callback (consumer only)

  • A send task callback (producer only)

  • A new consumer callback (triggered upon creation/destruction of a remove consumer)

• For MsgQ fast path producer or consumer:

  • A started MsgQ must be provided on creation

  • A DPA instance must be provided (DPA consumer/producer only)

  • A DPA consumer completion context must be connected (DPA consumer only)

  • A DPA completion context must be attached (DPA producer only)

  • A post receive task callback (CPU consumer only)

  • The number of receive operations (DPA consumer only)

  • A send task callback (CPU producer only)

  • The number of send operations (DPA producer only)

Optional Configurations

The following configurations are optional, if they are not set then a default value will be used:

For basic send/receive client:

• doca_comch_(server|client)_set_max_msg_size – set the maximum size of message that can be sent. If set, it must be matching between server and client.

• doca_comch_(server|client)_set_recv_queue_size – set the size of the queue to receive new messages on

For fast path consumers:

• doca_comch_consumer_set_imm_data_len – set the length of immediate data that a consumer can receive.

Device Support

DOCA Comch requires a device to operate. For instructions on picking a device, see DOCA Core Device Discovery.

As device capabilities are subject to change (see DOCA Core Device Support), it is recommended to select a device using the following methods:

• For basic client and server:

  • doca_comch_cap_server_is_supported

  • doca_comch_cap_client_is_supported

• For extended fast path functionality:

  • doca_comch_producer_cap_is_supported

  • doca_comch_consumer_cap_is_supported

Some devices can allow different capabilities as follows:

• The maximum length server name

• The maximum message size

• The maximum receive queue length

• The maximum clients that can connect to a server

• The maximum number of send tasks or post receive tasks

• The maximum buffer length for fast path

• The maximum immediate data supported by a fast path consumer

Buffer Support

Basic send and receive between a client and server does not use DOCA buffers and so has no restrictions on buffer type.

• For producers, supplied buffers need only be from a local mmap

• For consumers, post receive buffers are required to be from a PCIe export mmap

This section describes execution on CPU using DOCA Core Progress Engine. For additional execution environments, refer to section “Alternative Datapath Options”.

Tasks

DOCA Comch exposes asynchronous tasks that leverage the BlueField hardware according to DOCA Core architecture.

Control Channel Send Task

This task allows the sending of messages between connected client and server objects.

Configuration

Input

Common input as described in DOCA Core Task.

Output

Common output as described in DOCA Core Task.

Task Successful Completion

After the task completes successfully:

• The message is delivered to the connections remote client/server

• A receive event is triggered on the remote side

Task Failed Completion

If the task fails midway:

• The context may enter stopping state if a fatal error occurs

• The message is not delivered to the remote side

Limitations

• The operation is not atomic

• Once the task has been submitted, then the message should not be updated

• Other limitations are described in DOCA Core Task

Consumer Post Receive Task

This task allows consumer objects to publish buffers which are available for remote producers to write to.

Configuration

Input

Common input as described in DOCA Core Task.

Output

Common output as described in DOCA Core Task.

Task Successful Completion

The task only completes once a producer has written to the advertised buffer (or immediate data, or both), not when the post receive has completed.

Upon successful completion, the buffer contains the data written by the producer and its length is updated appropriately.

Task Failed Completion

Task failure occurs if a buffer has not been successfully posted to receive data.

If the task fails midway:

• The context may enter stopping state if a fatal error occurs

• Producers are not aware of the buffer so would not write to it

Limitations

• The operation is not atomic

• Once the task has been submitted, the buffer should not be read/written to

• Buffer must come from memory with PCIe read/write access

• Chained buffer lists are not supported

• MsgQ consumer does not support providing doca_buf, and can only receive immediate data

• Other limitations are described in DOCA Core Task

Producer Send Task

This task allows producer objects to copy buffers for use by remote consumers.

Configuration

Input

Common input as described in DOCA Core Task.

Output

Common output as described in DOCA Core Task.

Task Successful Completion

After the task is completed successfully:

• The data is copied form the buffer to the next free buffer posted by the given consumer

• Consumers process buffers from a given consumer in the order they are sent

Task Failed Completion

If the task fails midway:

• The context may enter stopping state if a fatal error occurs

• The source and destination doca_buf objects are not modified

• The destination memory may be modified

Limitations

• The operation is not atomic

• Once the task has been submitted, the buffer should not be read/written to

• The buffer length should not be greater than consumer post receive buffers (an invalid value is returned otherwise)

• MsgQ producer does not support providing doca_buf, and can only send immediate data

• All limitations described in DOCA Core Task

Events

DOCA Comch exposes asynchronous events to notify about changes that happen out of the blue, according to the DOCA Core architecture. See DOCA Core Event.

Common events as described in DOCA Core Event.

Control Channel Receive Event

This event triggers whenever a remote client/server has sent a message to the local client/server object.

Configuration

Trigger Condition

The event is triggered when a remote message is received on any currently active connection associated with the client or server.

Output

Upon event detection, the registered callback is triggered, passing the following parameters:

• A pointer to the message data

  Info

  The data is only valid in the context of the callback.

• The length in bytes of the message

• The active connection on which the message was received

Connection Status Changed Event (Server Only)

This event provides asynchronous updates on the state of any connections associated with a server.

Configuration

Trigger Condition

The event is triggered when a new connection is either established or a current connection disconnected on a server.

Output

Separate callbacks are registered for connection or disconnection events with the appropriate one triggered based on the specific event.

Both callbacks contain a Boolean indicating if the connection or disconnection was successful.

Consumer Event

This event indicates that a new consumer object has been created or an existing consumer object has been destroyed.

Configuration

Trigger Condition

The event is triggered whenever a new consumer is created or a current consumer destroyed on the remote side of an established DOCA Comch connection.

Output

The event hits a separate callback for either the creation or destruction of a consumer.

Callback parameters include:

• The established DOCA Comch connection on which the consumer is connected (on the remote side)

• The ID of the consumer (a unique value per Comch connection)

The DOCA Comch library follows the Context state machine described in DOCA Core Context State Machine.

The following section describes how to move to the state and what is allowed in each state.

Idle

In this state it is expected that the application either:

• Destroys the context

• Starts the context

Allowed operations:

• Configuring the context according to Configurations

• Starting the context

It is possible to reach this state as follows:

Starting

In this state it is expected that the application will:

• Call progress to allow transition to next state (e.g., when a connection attempt completes)

Allowed operations:

• Call progress

It is possible to reach this state as follows:

Running

In this state, it is expected that the application:

• Allocates and submit tasks

• Calls progress to complete tasks and/or receive events

Allowed operations:

• Allocate a previously configured task

• Submit an allocated task

• Call stop

It is possible to reach this state as follows:

Stopping

In this state, it is expected that the application will:

• Free any completed tasks

Allowed operations:

• Allocate previously configured task

• Submit an allocated task

• Call stop

It is possible to reach this state as follows:

DOCA Comch can be run on as part of DPA data path, using the MsgQ.

DPA

Using the MsgQ it is possible to create consumer/producer on the DPA. They follow the definition described in DOCA Core DPA.

Since these objects can be used in DPA, they have DPA APIs that can be used to perform the data path operations expanded on in the following subsections.

Consumer Ack

The doca_dpa_dev_comch_consumer_ack API prepares the DPA consumer to receive a number of immediate messages from CPU producers.

Configuration

Input

Completion

Whenever a message is received from the CPU producer a completion element is generatedand can be polled using doca_dpa_dev_comch_consumer_get_completion.

Using the generated completion, it is possible to get the following outputs:

Limitations

• The maximal immediate message size is 32 bytes

Producer Post Send Immediate Only

The doca_dpa_dev_comch_producer_post_send_imm_only API sends an immediate message to the CPU consumer. Once the message arrives at the CPU consumer side, the CPU consumer receive task completes.

The CPU producer must have posted a receive task prior to this. The user can verify if the consumer can receive the message using doca_dpa_dev_comch_producer_is_consumer_empty. Note, however, that this may add overhead.

Configuration

Input

Completion

Once the message arrives to the CPU consumer, a completion element is generated, indicating that the send is complete (this is separate from the completion the CPU consumer receives) and can be polled using doca_dpa_dev_get_completion.

Using the generated completion, it is possible to get the following outputs:

Limitations

• The maximal immediate message size is 32 bytes

Producer DMA Copy

The doca_dpa_dev_comch_producer_dma_copy API performs a DMA copy operation and, once the copy operation is done, sends an immediate message to the CPU consumer. Once the message arrives at the CPU consumer side, the CPU consumer receive task completes.

The CPU producer must have posted a receive task prior to this. The user can verify if the consumer can receive the message using doca_dpa_dev_comch_producer_is_consumer_empty. Note, however, that this may add overhead.

Configuration

Input

Completion

Once copy is complete and the message arrives to the CPU consumer, a completion element is generated, indicating that the copy is complete (this is separate from the completion the CPU consumer receives) and can be polled using doca_dpa_dev_get_completion.

Using the generated completion, it is possible to get the following outputs:

Limitations

• The maximal immediate message size is 32 bytes

This section describes DOCA Comch samples based on the DOCA Comch library.

The samples illustrates how to use the DOCA Comch API to do the following:

• Set up a client/server between host and BlueField Arm cores and use it to send text messages

• Configure fast path producers and consumers, and send messages between them

Running the Samples

1. Refer to the following documents:

   • NVIDIA DOCA Installation Guide for Linux for details on how to install BlueField-related software.

   • NVIDIA DOCA Troubleshooting Guide for any issue you may encounter with the installation, compilation, or execution of DOCA samples.

2. To build a given sample:

   Copy

   Copied!

               
   
               
   cd /opt/mellanox/doca/samples/doca_comch/<sample_name>
   meson /tmp/build
   ninja -C /tmp/build
           

   The binary doca_<sample_name> is created under /tmp/build/.

3. All DOCA Comch samples accept the same input arguments:

   Sample	Argument	Description
   doca_comch_ctrl_path_server doca_comch_ctrl_path_client doca_comch_data_path_high_speed_server doca_comch_data_path_high_speed_client	-p, --pci-addr	DOCA Comch device PCIe address
   	-r, --rep-pci	DOCA Comch device representor PCIe address (required only on BlueField Arm)
   	-t, --text	Text to be sent to the other side of channel (overwrites default)

4. For additional information per sample, use the -h option:

   Copy

   Copied!

               
   
               
   /tmp/build/<sample_name> -h
           

Samples

DOCA Comch Control Path Client/Server

This sample sets up a client server connection between the host and BlueField Arm cores.

The connection is used to pass two messages, the first sent by the client when the connection is established and the second by the server on receipt of the client’s message.

The sample logic includes:

1. Locating DOCA device.

2. Initializing the core DOCA structures.

3. Initializing and configuring client/server contexts.

4. Registering tasks and events for sending/receiving messages and tracking connection changes.

5. Allocating and submitting tasks for sending control path messages.

6. Handling event completions for receiving messages.

7. Stopping and destroying client/server objects.

References:

• /opt/mellanox/doca/samples/doca_comch/comch_ctrl_path_client/comch_ctrl_path_client_main.c

• /opt/mellanox/doca/samples/doca_comch/comch_ctrl_path_client/comch_ctrl_path_client_sample.c

• /opt/mellanox/doca/samples/doca_comch/comch_ctrl_path_server/comch_ctrl_path_server_main.c

• /opt/mellanox/doca/samples/doca_comch/comch_ctrl_path_server/comch_ctrl_path_server_sample.c

• /opt/mellanox/doca/samples/doca_comch/comch_ctrl_path_common.c

• /opt/mellanox/doca/samples/doca_comch/comch_ctrl_path_common.h

DOCA Comch Data Path Client/Server

This sample sets up a client server connection between host and BlueField Arm.

The connection is used to create a producer and consumer on both sides and pass a message across the two fastpath connections.

The sample logic includes:

1. Locating DOCA device.

2. Initializing the core DOCA structures.

3. Initializing and configuring client/server contexts.

4. Initializing and configuring producer/consumer contexts on top of an established connection.

5. Submitting post receive tasks for population by producers.

6. Submitting send tasks from producers to write to consumers.

7. Stopping and destroying producer/consumer objects.

8. Stopping and destroying client/server objects.

References:

• /opt/mellanox/doca/samples/doca_comch/comch_data_path_high_speed_client/comch_data_path_high_speed_client_main.c

• /opt/mellanox/doca/samples/doca_comch/comch_data_path_high_speedclient/comch_data_path_high_speed_client_sample.c

• /opt/mellanox/doca/samples/doca_comch/comch_data_path_high_speedserver/comch_data_path_high_speed_server_main.c

• /opt/mellanox/doca/samples/doca_comch/comch_data_path_high_speedserver/comch_data_path_high_speed_server_sample.c

• /opt/mellanox/doca/samples/doca_comch/comch_data_path_high_speed_common.c

• /opt/mellanox/doca/samples/doca_comch/comch_data_path_high_speed_common.h