blob: e2a0a11dc8bf2c6221927156f07f9e94fe5766cb [file] [log] [blame]
.. _module-pw_transfer:
``pw_transfer`` is a reliable data transfer protocol which runs on top of
Pigweed RPC.
.. attention::
``pw_transfer`` is under construction and so is its documentation.
Transfer thread
To run transfers as either a client or server (or both), a dedicated thread is
required. The transfer thread is used to process all transfer-related events
safely. The same transfer thread can be shared by a transfer client and service
running on the same system.
.. note::
All user-defined transfer callbacks (i.e. the virtual interface of a
``Handler`` or completion function in a transfer client) will be
invoked from the transfer thread's context.
In order to operate, a transfer thread requires two buffers:
- The first is a *chunk buffer*. This is used to stage transfer packets received
by the RPC system to be processed by the transfer thread. It must be large
enough to store the largest possible chunk the system supports.
- The second is an *encode buffer*. This is used by the transfer thread to
encode outgoing RPC packets. It is necessarily larger than the chunk buffer.
Typically, this is sized to the system's maximum transmission unit at the
transport layer.
A transfer thread is created by instantiating a ``pw::transfer::Thread``. This
class derives from ``pw::thread::ThreadCore``, allowing it to directly be used
when creating a system thread. Refer to :ref:`module-pw_thread-thread-creation`
for additional information.
**Example thread configuration**
.. code-block:: cpp
#include "pw_transfer/transfer_thread.h"
namespace {
// The maximum number of concurrent transfers the thread should support as
// either a client or a server. These can be set to 0 (if only using one or
// the other).
constexpr size_t kMaxConcurrentClientTransfers = 5;
constexpr size_t kMaxConcurrentServerTransfers = 3;
// The maximum payload size that can be transmitted by the system's
// transport stack. This would typically be defined within some transport
// header.
constexpr size_t kMaxTransmissionUnit = 512;
// The maximum amount of data that should be sent within a single transfer
// packet. By necessity, this should be less than the max transmission unit.
// pw_transfer requires some additional per-packet overhead, so the actual
// amount of data it sends may be lower than this.
constexpr size_t kMaxTransferChunkSizeBytes = 480;
// Buffers for storing and encoding chunks (see documentation above).
std::array<std::byte, kMaxTransferChunkSizeBytes> chunk_buffer;
std::array<std::byte, kMaxTransmissionUnit> encode_buffer;
transfer_thread(chunk_buffer, encode_buffer);
} // namespace
// pw::transfer::TransferThread is the generic, non-templated version of the
// Thread class. A Thread can implicitly convert to a TransferThread.
pw::transfer::TransferThread& GetSystemTransferThread() {
return transfer_thread;
Transfer server
``pw_transfer`` provides an RPC service for running transfers through an RPC
To know how to read data from or write data to device, a ``Handler`` interface
is defined (``pw_transfer/public/pw_transfer/handler.h``). Transfer handlers
represent a transferable resource, wrapping a stream reader and/or writer with
initialization and completion code. Custom transfer handler implementations
should derive from ``ReadOnlyHandler``, ``WriteOnlyHandler``, or
``ReadWriteHandler`` as appropriate and override Prepare and Finalize methods
if necessary.
A transfer handler should be implemented and instantiated for each unique
resource that can be transferred to or from a device. Each instantiated handler
must have a globally-unique integer ID used to identify the resource.
Handlers are registered with the transfer service. This may be done during
system initialization (for static resources), or dynamically at runtime to
support ephemeral transfer resources.
**Example transfer handler implementation**
.. code-block:: cpp
#include "pw_stream/memory_stream.h"
#include "pw_transfer/transfer.h"
// A simple transfer handler which reads data from an in-memory buffer.
class SimpleBufferReadHandler : public pw::transfer::ReadOnlyHandler {
SimpleReadTransfer(uint32_t resource_id, pw::ConstByteSpan data)
: ReadOnlyHandler(resource_id), reader_(data) {
pw::stream::MemoryReader reader_;
The transfer service is instantiated with a reference to the system's transfer
thread and registered with the system's RPC server.
**Example transfer service initialization**
.. code-block:: cpp
#include "pw_transfer/transfer.h"
namespace {
// In a write transfer, the maximum number of bytes to receive at one time
// (potentially across multiple chunks), unless specified otherwise by the
// transfer handler's stream::Writer.
constexpr size_t kDefaultMaxBytesToReceive = 1024;
pw::transfer::TransferService transfer_service(
GetSystemTransferThread(), kDefaultMaxBytesToReceive);
// Instantiate a handler for the data to be transferred. The resource ID will
// be used by the transfer client and server to identify the handler.
constexpr uint32_t kMagicBufferResourceId = 1;
char magic_buffer_to_transfer[256] = { /* ... */ };
SimpleBufferReadHandler magic_buffer_handler(
kMagicBufferResourceId, magic_buffer_to_transfer);
} // namespace
void InitTransferService() {
// Register the handler with the transfer service, then the transfer service
// with an RPC server.
Transfer client
``pw_transfer`` provides a transfer client capable of running transfers through
an RPC client.
.. note::
Currently, a transfer client is only capable of running transfers on a single
RPC channel. This may be expanded in the future.
The transfer client provides the following two APIs for starting data transfers:
.. cpp:function:: pw::Status pw::transfer::Client::Read(uint32_t resource_id, pw::stream::Writer& output, CompletionFunc&& on_completion, pw::chrono::SystemClock::duration timeout = cfg::kDefaultChunkTimeout, pw::transfer::ProtocolVersion version = kDefaultProtocolVersion)
Reads data from a transfer server to the specified ``pw::stream::Writer``.
Invokes the provided callback function with the overall status of the
Due to the asynchronous nature of transfer operations, this function will only
return a non-OK status if it is called with bad arguments. Otherwise, it will
return OK and errors will be reported through the completion callback.
.. cpp:function:: pw::Status pw::transfer::Client::Write(uint32_t resource_id, pw::stream::Reader& input, CompletionFunc&& on_completion, pw::chrono::SystemClock::duration timeout = cfg::kDefaultChunkTimeout, pw::transfer::ProtocolVersion version = kDefaultProtocolVersion)
Writes data from a source ``pw::stream::Reader`` to a transfer server.
Invokes the provided callback function with the overall status of the
Due to the asynchronous nature of transfer operations, this function will only
return a non-OK status if it is called with bad arguments. Otherwise, it will
return OK and errors will be reported through the completion callback.
**Example client setup**
.. code-block:: cpp
#include "pw_transfer/client.h"
namespace {
// RPC channel on which transfers should be run.
constexpr uint32_t kChannelId = 42;
pw::transfer::Client transfer_client(
GetSystemRpcClient(), kChannelId, GetSystemTransferThread());
} // namespace
Status ReadMagicBufferSync(pw::ByteSpan sink) {
pw::stream::Writer writer(sink);
struct {
pw::sync::ThreadNotification notification;
pw::Status status;
} transfer_state;
[&transfer_state](pw::Status status) {
transfer_state.status = status;
// Block until the transfer completes.
return transfer_state.status;
Atomic File Transfer Handler
Transfers are handled using the generic `Handler` interface. A specialized
`Handler`, `AtomicFileTransferHandler` is available to handle file transfers
with atomic semantics. It guarantees that the target file of the transfer is
always in a correct state. A temporary file is written to prior to updating the
target file. If any transfer failure occurs, the transfer is aborted and the
target file is either not created or not updated.
Module Configuration Options
The following configurations can be adjusted via compile-time configuration of
this module, see the
:ref:`module documentation <module-structure-compile-time-configuration>` for
more details.
The default maximum number of times a transfer should retry sending a chunk
when no response is received. This can later be configured per-transfer.
The default amount of time, in milliseconds, to wait for a chunk to arrive
before retrying. This can later be configured per-transfer.
The fractional position within a window at which a receive transfer should
extend its window size to minimize the amount of time the transmitter
spends blocked.
For example, a divisor of 2 will extend the window when half of the
requested data has been received, a divisor of three will extend at a third
of the window, and so on.
.. automodule:: pw_transfer
:members: ProgressStats, Manager, Error
.. code-block:: python
import pw_transfer
# Initialize a Pigweed RPC client; see pw_rpc docs for more info.
rpc_client = CustomRpcClient()
rpcs =
transfer_service =
transfer_manager = pw_transfer.Manager(transfer_service)
# Read the transfer resource with ID 3 from the server.
data =
except pw_transfer.Error as err:
print('Failed to read:', err.status)
# Send some data to the server. The transfer manager does not have to be
# reinitialized.
transfer_manager.write(2, b'hello, world')
except pw_transfer.Error as err:
print('Failed to write:', err.status)
Provides a simple interface for transferring bulk data over pw_rpc.
.. code-block:: typescript
import { pw_transfer } from 'pigweedjs';
const { Manager } from pw_transfer;
const client = new CustomRpcClient();
service =!.service('pw.transfer.Transfer')!;
const manager = new Manager(service, DEFAULT_TIMEOUT_S);, (stats: ProgressStats) => {
console.log(`Progress Update: ${stats}`);
}).then((data: Uint8Array) => {
console.log(`Completed read: ${data}`);
}).catch(error => {
console.log(`Failed to read: ${error.status}`);
manager.write(2, textEncoder.encode('hello world'))
.catch(error => {
console.log(`Failed to read: ${error.status}`);
Transfers run as a series of *chunks* exchanged over an RPC stream. Chunks can
contain transferable data, metadata, and control parameters. Each chunk has an
associated type, which determines what information it holds and the semantics of
its fields.
The chunk is a protobuf message, whose definition can be found
:ref:`here <module-pw_transfer-proto-definition>`.
Resources and sessions
Transfers are run for a specific *resource* --- a stream of data which can be
read from or written to. Resources have a system-specific integral identifier
defined by the implementers of the server-side transfer node.
The series of chunks exchanged in an individual transfer operation for a
resource constitute a transfer *session*. The session runs from its opening
chunk until either a terminating chunk is received or the transfer times out.
Sessions are assigned unique IDs by the transfer server in response to an
initiating chunk from the client.
``pw_transfer`` attempts to be a reliable data transfer protocol.
As Pigweed RPC is considered an unreliable communications system,
``pw_transfer`` implements its own mechanisms for reliability. These include
timeouts, data retransmissions, and handshakes.
.. note::
A transfer can only be reliable if its underlying data stream is seekable.
A non-seekable stream could prematurely terminate a transfer following a
packet drop.
Opening handshake
Transfers begin with a three-way handshake, whose purpose is to identify the
resource being transferred, assign a session ID, and synchronize the protocol
version to use.
A read or write transfer for a resource is initiated by a transfer client. The
client sends the ID of the resource to the server in a ``START`` chunk,
indicating that it wishes to begin a new transfer. This chunk additionally
encodes the protocol version which the client is configured to use.
Upon receiving a ``START`` chunk, the transfer server checks whether the
requested resource is available. If so, it prepares the resource for the
operation, which typically involves opening a data stream, alongside any
additional user-specified setup. The server generates a session ID, then
responds to the client with a ``START_ACK`` chunk containing the resource,
session, and configured protocol version for the transfer.
Transfer completion
Either side of a transfer can terminate the operation at any time by sending a
``COMPLETION`` chunk containing the final status of the transfer. When a
``COMPLETION`` chunk is sent, the terminator of the transfer performs local
cleanup, then waits for its peer to acknowledge the completion.
Upon receving a ``COMPLETION`` chunk, the transfer peer cancels any pending
operations, runs its set of cleanups, and responds with a ``COMPLETION_ACK``,
fully ending the session from the peer's side.
The terminator's session remains active waiting for a ``COMPLETION_ACK``. If not
received after a timeout, it re-sends its ``COMPLETION`` chunk. The session ends
either following receipt of the acknowledgement or if a maximum number of
retries is hit.
.. _module-pw_transfer-proto-definition:
Server to client transfer (read)
.. image:: read.svg
Client to server transfer (write)
.. image:: write.svg
Protocol buffer definition
.. literalinclude:: transfer.proto
:language: protobuf
:lines: 14-
Protocol errors
The following table describes the meaning of each status code when sent by the
sender or the receiver (see `Transfer roles`_).
.. cpp:namespace-push:: pw::stream
| Status | Sent by sender | Sent by receiver |
| ``OK`` | (not sent) | All data was received |
| | | and handled |
| | | successfully. |
| ``ABORTED`` | The service aborted the transfer because the |
| | client restarted it. This status is passed to the |
| | transfer handler, but not sent to the client |
| | because it restarted the transfer. |
| ``CANCELLED`` | The client cancelled the transfer. |
| ``DATA_LOSS`` | Failed to read the data | Failed to write the |
| | to send. The | received data. The |
| | :cpp:class:`Reader` | :cpp:class:`Writer` |
| | returned an error. | returned an error. |
| ``FAILED_PRECONDITION`` | Received chunk for transfer that is not active. |
| ``INVALID_ARGUMENT`` | Received a malformed packet. |
| ``INTERNAL`` | An assumption of the protocol was violated. |
| | Encountering ``INTERNAL`` indicates that there is |
| | a bug in the service or client implementation. |
| ``PERMISSION_DENIED`` | The transfer does not support the requested |
| | operation (either reading or writing). |
| ``RESOURCE_EXHAUSTED`` | The receiver requested | Storage is full. |
| | zero bytes, indicating | |
| | their storage is full, | |
| | but there is still data | |
| | to send. | |
| ``UNAVAILABLE`` | The service is busy with other transfers and |
| | cannot begin a new transfer at this time. |
| ``UNIMPLEMENTED`` | Out-of-order chunk was | (not sent) |
| | requested, but seeking | |
| | is not supported. | |
.. cpp:namespace-pop::
Transfer roles
Every transfer has two participants: the sender and the receiver. The sender
transmits data to the receiver. The receiver controls how the data is
transferred and sends the final status when the transfer is complete.
In read transfers, the client is the receiver and the service is the sender. In
write transfers, the client is the sender and the service is the receiver.
Sender flow
.. mermaid::
graph TD
start([Client initiates<br>transfer]) -->data_request
data_request[Receive transfer<br>parameters]-->send_chunk
send_chunk[Send chunk]-->sent_all
sent_all{Sent final<br>chunk?} -->|yes|wait
sent_requested{Sent all<br>pending?}-->|yes|data_request
wait[Wait for receiver]-->is_done
is_done{Received<br>final chunk?}-->|yes|done
done([Transfer complete])
Receiver flow
.. mermaid::
graph TD
start([Client initiates<br>transfer]) -->request_bytes
request_bytes[Set transfer<br>parameters]-->wait
wait[Wait for chunk]-->received_chunk
received_chunk{Received<br>chunk by<br>deadline?}-->|no|request_bytes
check_chunk{Correct<br>offset?} -->|yes|process_chunk
check_chunk --> |no|request_bytes
process_chunk[Process chunk]-->final_chunk
received_requested{Received all<br>pending?}-->|yes|request_bytes
signal_completion[Signal completion]-->done
done([Transfer complete])
Legacy protocol
``pw_transfer`` was initially released into production prior to several of the
reliability improvements of its modern protocol. As a result of this, transfer
implementations support a "legacy" protocol mode, in which transfers run without
utilizing these features.
The primary differences between the legacy and modern protocols are listed
- There is no distinction between a transfer resource and session --- a single
``transfer_id`` field represents both. Only one transfer for a given resource
can run at a time, and it is not possible to determine where one transfer for
a resource ends and the next begins.
- The legacy protocol has no opening handshake phase. The client initiates with
a transfer ID and starting transfer parameters (during a read), and the data
transfer phase begins immediately.
- The legacy protocol has no terminating handshake phase. When either end
completes a transfer by sending a status chunk, it does not wait for the peer
to acknowledge. Resources used by the transfer are immediately freed, and
there is no guarantee that the peer is notified of completion.
Modern transfer server and client implementations will detect if their transfer
peer is running the legacy protocol and automatically switch to it if required.
It is **strongly** unadvised to use the legacy protocol in new code.
Integration tests
The ``pw_transfer`` module has a set of integration tests that verify the
correctness of implementations in different languages.
`Test source code <>`_.
To run the tests on your machine, run
.. code:: bash
$ bazel run pw_transfer/integration_test:cross_language_integration_test
The integration tests permit injection of client/server/proxy binaries to use
when running the tests. This allows manual testing of older versions of
pw_transfer against newer versions.
.. code:: bash
# Test a newer version of pw_transfer against an old C++ client that was
# backed up to another directory.
$ bazel run pw_transfer/integration_test:cross_language_integration_test -- \
--cpp-client ../old_pw_transfer_version/cpp_client
CI/CQ integration
`Current status of the test in CI <>`_.
By default, these tests are not run in CQ (on presubmit) because they are too
slow. However, you can request that the tests be run in presubmit on your
change by adding to following line to the commit message footer:
.. code::
Cq-Include-Trybots: luci.pigweed.try:pigweed-integration-transfer